Structural Design Optimization Considering Uncertainties

Structural design optimization considering uncertainties

Structures and Infrastructures Series ISSN 1747-7735

Book Series Editor:

Dan M. Frangopol Professor of Civil Engineering and Fazlur R. Khan Endowed Chair of Structural Engineering and Architecture Department of Civil and Environmental Engineering Center for Advanced Technology for Large Structural Systems (ATLSS Center) Lehigh University Bethlehem, PA, USA

Volume 1

Structural design optimization considering uncertainties

Edit ed by

Yiannis Tsompanakis1, Nikos D. Lagaros2 & Manolis Papadrakakis3 1

Department of Applied Sciences, Technical University of Crete, University Campus, Chania, Crete, Greece 2,3 Institute of Structural Analysis & Seismic Research, Faculty of Civil Engineering, National Technical University of Athens, Zografou Campus, Athens, Greece

LONDON / LEIDEN / NEW YORK / PHILADELPHIA / SINGAPORE

Colophon Book Series Editor : Dan M. Frangopol Volume Editors: Yiannis Tsompanakis, Nikos D. Lagaros and Manolis Papadrakakis Cover illustration: Objective space of the M-3OU multi-criteria optimization problem Nikos D. Lagaros September 2007 This edition published in the Taylor & Francis e-Library, 2008. “To purchase your own copy of this or any of Taylor & Francis or Routledge’s collection of thousands of eBooks please go to www.eBookstore.tandf.co.uk.” Taylor & Francis is an imprint of the Taylor & Francis Group, an informa business ©2008 Taylor & Francis Group, London, UK All rights reserved. No part of this publication or the information contained herein may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, by photocopying, recording or otherwise, without written prior permission from the publishers. Although all care is taken to ensure integrity and the quality of this publication and the information herein, no responsibility is assumed by the publishers nor the author for any damage to the property or persons as a result of operation or use of this publication and/or the information contained herein. British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library Library of Congress Cataloging-in-Publication Data Structural design optimization considering uncertainties / Edited by Yiannis Tsompanakis, Nikos D. Lagaros & Manolis Papadrakakis. p. cm. – (Structures and infrastructures series ; 1747-7735) Includes bibliographical references and index. ISBN 978-0-415-45260-1 (hardcover : alk. paper) ISBN 978-0-203-93852-2 (e-book) 1. Structural optimization. I. Tsompanakis, Yiannis. 1969– II. Lagaros, Nikos D. 1970– III. Papadrakakis, Manolis. 1948– TA658.8.S73 2007 624.1
7713–dc22 2007040343 Published by: Taylor & Francis/Balkema P.O. Box 447, 2300 AK Leiden, The Netherlands e-mail: [email protected] www.balkema.nl, www.taylorandfrancis.co.uk, www.crcpress.com ISBN 0-203-93852-6 Master e-book ISBN

ISBN13 978-0-415-45260-1(Hbk) ISBN13 978-0-203-93852-2(eBook) Structures and Infrastructures Series: ISSN 1747-7735 Volume 1

Table of Contents

Editorial About the Book Series Editor Foreword Preface Brief Curriculum Vitae of the Editors List of Contributors Author Data

IX XI XIII XV XXI XXIII XXV

PART 1

Reliability-Based Design Optimization (RBDO) 1

2

Principles of reliability-based design optimization Alaa Chateauneuf , University Blaise Pascal, France Reliability-based optimization of engineering structures

3

31

John D. Sørensen, Aalborg University, Aalborg, Denmark 3

4

Reliability analysis and reliability-based design optimization using moment methods Sang Hoon Lee, Northwestern University, Evanston, IL, USA Byung Man Kwak, Korea Advanced Institute of Science and Technology, Daejeon, Korea Jae Sung Huh, Korea Aerospace Research Institute, Daejeon, Korea Efficient approaches for system reliability-based design optimization Efstratios Nikolaidis, University of Toledo, Toledo, OH, USA Zissimos P. Mourelatos, Oakland University, Rochester, MI, USA Jinghong Liang, Oakland University, Rochester, MI, USA

57

87

VI

5

6

Contents

Nondeterministic formulations of analytical target cascading for decomposition-based design optimization under uncertainty Michael Kokkolaras, University of Michigan, Ann Arbor, MI, USA Panos Y. Papalambros, University of Michigan, Ann Arbor, MI, USA Design optimization of stochastic dynamic systems by algebraic reduced order models Gary Weickum, University of Colorado at Boulder, Boulder, CO, USA Matt Allen, University of Colorado at Boulder, Boulder, CO, USA Kurt Maute, University of Colorado at Boulder, Boulder, CO, USA Dan M. Frangopol, Lehigh University, Bethlehem, PA, USA

7

Stochastic system design optimization using stochastic simulation Alexandros A. Taflanidis, California Institute of Technology, CA, USA James L. Beck, California Institute of Technology, CA, USA

8

Numerical and semi-numerical methods for reliability-based design optimization Ghias Kharmanda, Aleppo University, Aleppo, Syria

9

Advances in solution methods for reliability-based design optimization Alaa Chateauneuf , University Blaise Pascal, France Younes Aoues, University Blaise Pascal, France

10

Non-probabilistic design optimization with insufficient data using possibility and evidence theories Zissimos P. Mourelatos, Oakland University, Rochester, MI, USA Jun Zhou, Oakland University, Rochester, MI, USA

115

135

155

189

217

247

11

A decoupled approach to reliability-based topology optimization for structural synthesis 281 Neal M. Patel, University of Notre Dame, Notre Dame, IN, USA John E. Renaud, University of Notre Dame, Notre Dame, IN, USA Donald Tillotson, University of Notre Dame, Notre Dame, IN, USA Harish Agarwal, General Electric Global Research, Niskayuna, NY, USA Andrés Tovar, National University of Colombia, Bogota, Colombia

12

Sample average approximations in reliability-based structural optimization: Theory and applications Johannes O. Royset, Naval Postgraduate School, Monterey, CA, USA Elijah Polak, University of California, Berkeley, CA, USA

13

Cost-benefit optimization for maintained structures

307

335

Rüdiger Rackwitz, Technical University of Munich, Munich, Germany Andreas E. Joanni, Technical University of Munich, Munich, Germany 14

A reliability-based maintenance optimization methodology Wu Y.-T., Applied Research Associates Inc., Raleigh, NC, USA

369

C o n t e n t s VII

15

Overview of reliability analysis and design capabilities in DAKOTA with Application to shape optimization of MEMS Michael S. Eldred, Sandia National Laboratories, Albuquerque, NM, USA Barron J. Bichon, Vanderbilt University, Nashville, TN, USA Brian M. Adams, Sandia National Laboratories, Albuquerque, NM, USA Sankaran Mahadevan, Vanderbilt University, Nashville, TN, USA

401

PART 2

Robust Design Optimization (RDO) 16

Structural robustness and its relationship to reliability Jorge E. Hurtado, National University of Colombia, Manizales, Colombia

17

Maximum robustness design of trusses via semidefinite programming Yoshihiro Kanno, University of Tokyo, Tokyo, Japan Izuru Takewaki, Kyoto University, Kyoto, Japan

18

19

20

21

Design optimization and robustness of structures against uncertainties based on Taylor series expansion Ioannis Doltsinis, University of Stuttgart, Stuttgart, Germany Info-gap robust design of passively controlled structures with load and model uncertainties Izuru Takewaki, Kyoto University, Kyoto, Japan Yakov Ben-Haim, Technion, Haifa, Israel Genetic algorithms in structural optimum design using convex models of uncertainty Sara Ganzerli, Gonzaga University, Spokane, WA, USA Paul De Palma, Gonzaga University, Spokane, WA, USA Metamodel-based computational techniques for solving structural optimization problems considering uncertainties Nikos D. Lagaros, National Technical University of Athens, Athens, Greece Yiannis Tsompanakis, Technical University of Crete, Chania, Greece Michalis Fragiadakis, University of Thessaly, Volos, Greece Vagelis Plevris, National Technical University of Athens, Athens, Greece Manolis Papadrakakis, National Technical University of Athens, Athens, Greece

References Author index Subject index

435

471

499

531

549

567

599 631 633

Editorial

Welcome to the New Book Series Structures and Infrastructures. Our knowledge to model, analyze, design, maintain, manage and predict the lifecycle performance of structures and infrastructures is continually growing. However, the complexity of these systems continues to increase and an integrated approach is necessary to understand the effect of technological, environmental, economical, social and political interactions on the life-cycle performance of engineering structures and infrastructures. In order to accomplish this, methods have to be developed to systematically analyze structure and infrastructure systems, and models have to be formulated for evaluating and comparing the risks and benefits associated with various alternatives. We must maximize the life-cycle benefits of these systems to serve the needs of our society by selecting the best balance of the safety, economy and sustainability requirements despite imperfect information and knowledge. In recognition of the need for such methods and models, the aim of this Book Series is to present research, developments, and applications written by experts on the most advanced technologies for analyzing, predicting and optimizing the performance of structures and infrastructures such as buildings, bridges, dams, underground construction, offshore platforms, pipelines, naval vessels, ocean structures, nuclear power plants, and also airplanes, aerospace and automotive structures. The scope of this Book Series covers the entire spectrum of structures and infrastructures. Thus it includes, but is not restricted to, mathematical modeling, computer and experimental methods, practical applications in the areas of assessment and evaluation, construction and design for durability, decision making, deterioration modeling and aging, failure analysis, field testing, structural health monitoring, financial planning, inspection and diagnostics, life-cycle analysis and prediction, loads, maintenance strategies, management systems, nondestructive testing, optimization of maintenance and management, specifications and codes, structural safety and reliability, system analysis, time-dependent performance, rehabilitation, repair, replacement, reliability and risk management, service life prediction, strengthening and whole life costing. This Book Series is intended for an audience of researchers, practitioners, and students world-wide with a background in civil, aerospace, mechanical, marine and automotive engineering, as well as people working in infrastructure maintenance, monitoring, management and cost analysis of structures and infrastructures. Some volumes are monographs defining the current state of the art and/or practice in the field, and some are textbooks to be used in undergraduate (mostly seniors), graduate and

X

Editorial

postgraduate courses. This Book Series is affiliated to Structure and Infrastructure Engineering (http://www.informaworld.com/sie), an international peer-reviewed journal which is included in the Science Citation Index. If you like to contribute to this Book Series as an author or editor, please contact the Book Series Editor ([email protected]) or the Publisher ([email protected]). A book proposal form can be downloaded at www.balkema.nl. It is now up to you, authors, editors, and readers, to make Structures and Infrastructures a success. Dan M. Frangopol Book Series Editor

About the Book Series Editor

Dr. Dan M. Frangopol is the first holder of the Fazlur R. Khan Endowed Chair of Structural Engineering and Architecture at Lehigh University, Bethlehem, Pennsylvania, USA, and a Professor in the Department of Civil and Environmental Engineering at Lehigh University. He is also an Emeritus Professor of Civil Engineering at the University of Colorado at Boulder, USA, where he taught for more than two decades (1983–2006). Before joining the University of Colorado, he worked for four years (1979–1983) in structural design with A. Lipski Consulting Engineers in Brussels, Belgium. In 1976, he received his doctorate in Applied Sciences from the University of Liège, Belgium, and holds an honorary doctorate degree (Doctor Honoris Causa) and a B.S. degree from the Technical University of Civil Engineering in Bucharest, Romania. He is a Fellow of the American Society of Civil Engineers (ASCE), American Concrete Institute (ACI), and International Association for Bridge and Structural Engineering (IABSE). He is also an Honorary Member of both the Romanian Academy of Technical Sciences and the Portuguese Association for Bridge Maintenance and Safety. He is the initiator and organizer of the Fazlur R. Khan Lecture Series (www.lehigh.edu/frkseries) at Lehigh University. Dan Frangopol is an experienced researcher and consultant to industry and government agencies, both nationally and abroad. His main areas of expertise are structural reliability, structural optimization, bridge engineering, and life-cycle analysis, design, maintenance, monitoring, and management of structures and infrastructures. He is the Founding President of the International Association for Bridge Maintenance and Safety (IABMAS, www.iabmas.org) and of the International Association for Life-Cycle Civil Engineering (IALCCE, www.ialcce.org), and Past Director of the Consortium on Advanced Life-Cycle Engineering for Sustainable Civil Environments (COALESCE). He is also the Chair of the Executive Board of the International Association for Structural Safety and Reliability (IASSAR, www.columbia.edu/cu/civileng/iassar) and the Vice-President of the International Society for Health Monitoring of Intelligent Infrastructures (ISHMII, www.ishmii.org). Dan Frangopol is the recipient of several prestigious awards including the 2007 ASCE Ernest Howard Award, the 2006 IABSE OPAC Award, the 2006 Elsevier Munro Prize, the 2006 T. Y. Lin Medal, the 2005 ASCE Nathan M. Newmark Medal, the 2004 Kajima Research Award, the 2003

XII A b o u t t h e B o o k S e r i e s E d i t o r

ASCE Moisseiff Award, the 2002 JSPS Fellowship Award for Research in Japan, the 2001 ASCE J. James R. Croes Medal, the 2001 IASSAR Research Prize, the 1998 and 2004 ASCE State-of-the-Art of Civil Engineering Award, and the 1996 Distinguished Probabilistic Methods Educator Award of the Society of Automotive Engineers (SAE). Dan Frangopol is the Founding Editor-in-Chief of Structure and Infrastructure Engineering (Taylor & Francis, www.informaworld.com/sie) an international peerreviewed journal, which is included in the Science Citation Index. This journal is dedicated to recent advances in maintenance, management, and life-cycle performance of a wide range of structures and infrastructures. He is the author or co-author of over 400 refereed publications, and co-author, editor or co-editor of more than 20 books published by ASCE, Balkema, CIMNE, CRC Press, Elsevier, McGraw-Hill, Taylor & Francis, and Thomas Telford and an editorial board member of several international journals. Additionally, he has chaired and organized several national and international structural engineering conferences and workshops. Dan Frangopol has supervised over 70 Ph.D. and M.Sc. students. Many of his former students are professors at major universities in the United States, Asia, Europe, and South America, and several are prominent in professional practice and research laboratories. For additional information on Dan M. Frangopol’s activities, please visit www.lehigh.edu/∼dmf206/

Foreword

The aim of structural optimization is to achieve the best possible design by maximizing benefits under conflicting criteria. Uncertainties are unavoidable in the structural optimization process. Therefore, a realistic optimal design process should definitely consider uncertainties. Two broad types of uncertainty have to be considered: (a) uncertainty associated with randomness, the so-called aleatory uncertainty, and (b) uncertainty associated with imperfect modeling, the so-called epistemic uncertainty. It has been clearly demonstrated that both aleatory and epistemic uncertainties can be treated, separately or combined, and analyzed using the principles of probability and statistics. Structural reliability theory has been developed during the past decades to handle problems considering such uncertainties. This continuous development has had considerable impact in recent years on structural optimization. The purpose of this book is to present the latest research findings in the field of structural optimization considering uncertainties. A wide variety of topics are covered by leading researchers. The first part (Chapters 1 to 15) is devoted to reliability-based design optimization, and the second part (Chapters 16 to 21) deals with robust design optimization. To provide the reader with a good overview of pertinent literature, all cited papers and additional references on the topics discussed, are collected in a comprehensive list of references. The Book Series Editor would like to express his appreciation to the Editors and all Authors who contributed to this book. It is his hope that this first volume in the Structures and Infrastructures Book Series will generate a lot of interest and help engineers to design the best structural systems under uncertainty.

Dan M. Frangopol Book Series Editor Bethlehem, Pennsylvania November 2, 2007

Preface

Uncertainties are inherent in engineering problems and the scatter of structural parameters from their nominal ideal values is unavoidable. The response of structural systems can sometimes be very sensitive to uncertainties encountered in the material properties, manufacturing conditions, external loading conditions and analytical and/or numerical modelling. In recent years, probabilistic-based formulations of optimization problems have been developed to account for uncertainties through stochastic simulation and probabilistic analysis. Stochastic analysis methods have been developed significantly over the last two decades and have stimulated the interest for the probabilistic optimum design of structures. There are mainly two design formulations that account for probabilistic systems response: Reliability-Based Design Optimization (RBDO) and Robust Design Optimization (RDO). The main goal of RBDO methods is to design for safety with respect to extreme events. RDO methods primarily seek to minimize the influence of stochastic variations on the mean design. The selected contributions of this book deal with the use of probabilistic methods for optimal design of different types of structures and various considerations of uncertainties. This volume is a collective book of twenty-one self-contained chapters, which present state-of-the-art theoretical advances and applications in various fields of probabilistic computational mechanics. The first fifteen chapters of the book are focused on RBDO theory and applications, while the rest of the chapters deal with advances in RDO and combined RBDO-RDO theory and applications. Apart from the reference list that is given separately for each chapter, a complete list of references is also provided for the reader. In order to obtain contributions that cover a wide spectrum of engineering problems, the problem of optimum design is considered in a broad sense. The probabilistic framework allows for a consistent treatment of both cost and safety. In what follows a short description of the book content is presented. In the introductory chapter by Chateauneuf, the fundamental theoretical and computational issues related to RBDO are described and the advantages of RBDO compared to conventional deterministic optimization approaches are outlined. This chapter emphasizes the role of uncertainties in deriving a “true’’ optimal solution, defined as the best compromise between cost minimization and safety assurance. The presented RBDO formulations cover various important probabilistic issues (theoretical, computational and practical), such as multi-component reliability analysis, safety factor calibration, multi-objective applications, as well as a great variety of engineering applications, such as topology, maintenance and time-variant problems.

XVI

Preface

The theoretical basis for reliability-based structural optimization is described by Sørensen within the framework of Bayesian statistical decision theory. This contribution presents the latest findings in RBDO with respect to three major types of decision problems with increased degree of complexity and uncertainty: a) decisions with given information (e.g. planning of new structures), b) decisions when new information is provided (e.g. for re-assessment and retrofitting of existing structures), c) decisions involving planning of experiments/inspections to obtain new information (e.g. for inspection planning). Furthermore, RBDO issues related to decisions with systematic reconstruction are also discussed. Reliability-based, cost-benefit problems are formulated and exemplified with structural optimization. Illustrative examples are presented including a simple introductory example, a decision problem related to bridge re-assessment and a reliability-based decision problem for offshore wind turbines. Lee, Kwak and Huh deal with reliability analysis and reliability-based design optimization using moment methods. By using this approach, a finite number of statistical moments of a system response function are calculated and the probability density function (PDF) of the system response is identified by empirical distribution systems, such as the Pearson or the Johnson system. In this chapter, a full factorial moment method (FFMM) procedure is introduced for reliability analysis calculations. A response surface augmented moment method (RSMM) is developed to construct a series of approximate response surface for enhancing the efficiency of FFMM. The probability of failure is calculated using an empirical distribution system and the first four statistical moments of system’s performance function are calculated from appropriate design simulations. The design sensitivity of the probability of failure, required during RBDO process, is calculated in a semi-analytic way using moment methods. As stated in the chapter by Nikolaidis, Mourelatos and Liang, a designer faces many challenges when applying RBDO to engineering systems. The high computational cost required for RBDO and the efficient computation of the system failure probability are the two principal challenges. As a result, most RBDO studies are restricted to the safety levels of the individual failure modes. In order to overcome this deficiency, two efficient approaches for RBDO are presented in this chapter. Both approaches apportion optimally the system reliability among the failure modes by considering the target values of the failure probabilities of the modes as design variables. The first approach uses a sequential optimization and reliability assessment (SORA) approach, while the second system RBDO approach uses a single-loop method where the searches for the optimum design and for the most probable failure points proceed simultaneously. The two approaches are illustrated and compared on characteristic design examples. Moreover, it is shown that the single-loop approach, enhanced with an active set strategy, is considerably more efficient than the SORA approach. According to the work of Kokkolaras and Papalambros, design subproblems are formulated and solved so that their solution can be integrated to represent the optimal design of the decomposed system. This approach requires appropriate problem formulation and coordination of the distributed, multilevel system design problem. The presented analytical target cascading (ATC) is a methodology suitable for multilevel optimal design problems. Design targets are cascaded to lower levels using the modelbased, hierarchical decomposition of the original design problem. An optimization problem is posed and solved for each design subproblem to minimize deviations from

Preface

XVII

propagated targets. Solving the subproblems and using an appropriate coordination strategy the overall system compatibility is preserved. The required computational effort motivated Weickum, Allen, Maute and Frangopol to address the need for developing efficient numerical probabilistic techniques for the reliability analysis and design optimization of stochastic dynamic systems. This work seeks to alleviate the computational costs for optimizing dynamic systems by employing reduced order models. The key to utilize reduced order models in stochastic analysis and optimization lies in making them adaptable to design changes and variations of the random parameters. For this purpose, an extended reduced order model (EROM) method, which is a reduced order model accounting for parameter changes, is integrated into stochastic analysis and design optimization. The application of the proposed EROM is tested both for deterministic and probabilistic optimization of the characteristic connecting rod example. Taflanidis and Beck consider a two stage framework for efficient implementation of RBDO of dynamical systems under stochastic excitation (e.g. earthquake, wind or wave loading), where uncertainties are assumed for both the excitation characteristics and the structural model adopted. In the first stage a novel approach, the so called stochastic subset optimization (SSO), is implemented for iteratively identifying a subset of the original design space that has high probability of containing the optimal design variables. The second stage adopts a stochastic optimization algorithm to pinpoint, if needed, the optimal design variables within that subset. Topics related to the combination of the two different stages, in order to enhance the overall efficiency of the presented methodology, are also discussed. An illustrative example for the seismic retrofitting, via viscous dampers, is presented. The minimization of the expected lifecycle cost is adopted as the design objective, in which the cost associated with damage caused by future earthquakes is calculated by stochastic simulation via a realistic probabilistic model for the structure and the ground motion that involves the formulation of an effective loss function model. Kharmanda discusses in his contribution issues related to RBDO formulation and solution procedures. The RBDO formulation is defined as a nonlinear mathematical programming problem in which the mean values of uncertain system parameters are used as design variables and its weight or cost is optimized subjected to prescribed probabilistic constraints. In this chapter, recent developments for the efficient RBDO problem solving using semi-numerical and numerical techniques are presented. Following a detailed description of the proposed methods, their efficiency is demonstrated in computationally demanding dynamic applications. The obtained results as well as the computational implications of the methods are compared and their advantages and disadvantages are highlighted in a comprehensive manner. In the contribution by Chateauneuf and Aoues, the main objective is to apply appropriate numerical methods in order to solve RBDO problems more efficiently. A comprehensive description of the most commonly used RBDO formulations and the corresponding numerical methods is provided. A good RBDO algorithm should satisfy the conditions of efficiency (computation time), precision (accuracy of finding the optimum), generality (capability to deal with different kinds of problems) and robustness (stability of the convergence for any admissible initial point, local or global convergence criteria, etc). All these aspects are discussed in detail, and effective solutions are proposed via characteristic test examples.

XVIII

Preface

In the chapter by Mourelatos and Zhou, evidence theories are used to account for uncertainty in structural design with incomplete and/or fuzzy information. A sequential possibility-based design optimization (SPDO) method is presented which decouples the design loop and the reliability assessment of each constraint and is also capable of handling both random and possibilistic design variables. Furthermore, a computationally efficient optimum design formulation using evidence theory is presented, which can handle a mixture of epistemic and aleatory uncertainties. Numerical examples demonstrate the application of possibility and evidence theories in probabilistic optimum design and highlight the trade-offs among reliability-based, possibility-based and evidence-based design approaches. In the chapter by Patel, Renaud, Tillotson, Agarwal, and Tovar, the mode of failure is considered to be the maximum deflection of the structure in reliability-based topology optimization (RBTO). A decoupled approach is employed in which the topology optimization stage is separate from the reliability analysis. The proposed decoupled reliability-based design optimization methodology is an approximate technique to obtain consistent reliable designs at lower computational expense. An efficient nongradient Hybrid Cellular Automaton (HCA) method has been implemented in the proposed decoupled approach for evaluating density changes, while the strain energy for every new design is evaluated via finite element structural analyses. The chapter by Royset and Polak presents recent advances in combining Monte Carlo sampling and nonlinear programming algorithms for RBDO problems utilizing effective approximation techniques that can lead to the reduction of the excessive computational cost. More specifically, they present an approach where the reliability term in the problem formulation is replaced by a statistical estimate of the reliability obtained by means of Monte Carlo sampling. The authors emphasize on the calculation of “adaptive optimal’’ sample size which is achieved using sample-adjustment rules by solving auxiliary optimization tasks during the evolution of RBDO. The efficiency of the methods is verified in a number of numerical examples arising in design of various types of structures having a single or multiple limit-state functions, in which reliability terms are included in both objective and constraint functions. Rackwitz and Joanni describe theoretical and practical issues leading to cost-efficient optimization formulations for existing aging structures. In order to establish an efficient methodology for optimizing maintenance, an elaborate model, based on renewal theory that uses systematic reconstruction or repair schemes after suitable inspection, is formulated in which life-cycle cost perspective is used. The presented implementation shows the impact of the choice of the objective function, the risk acceptability and the transient behaviour of the failure rate. The emphasis is given on concrete structures, but the described methodology can be applied to any material and any type of engineering structures. In particular, minimal age-dependent block repairs and maintenance by inspection and repair have been studied via an illustrative example. Wu describes in his contribution a reliability-based damage tolerance (RBDT) approach that employs a systematic approach for probabilistic fracture-mechanics damage tolerance analysis with maintenance planning under various uncertainties. Moreover, he presents the successful integration of RBDT in the proposed reliabilitybased maintenance optimization (RBMO) methodology, focusing on efficient sampling and other computational strategies for handling the uncertainties related to structural maintenance issues (fatigue, failure, inspection, repair, etc). A comparison of different

Preface

XIX

versions of the proposed RBMO for analytical benchmark examples as well as for realistic test cases is presented. Eldred, Bichon, Adams and Mahadevan present an overview of recent research related to first and second-order reliability methods. They outline both the forward reliability analysis of computing probabilities for specified response levels (using the so-called RIA, i.e. the reliability index approach) and the inverse reliability analysis of computing response levels for specified probabilities (the performance measure approach or PMA). A number of algorithmic variations are described and the effect of different limit state approximations, probability integrations, warm starting, most probable point search algorithms, and Hessian approximations is discussed. Relative performance of these reliability analysis and design algorithms is presented for several benchmark test problems as well as for real-world applications related to the probabilistic analysis and design of micro-electro-mechanical systems (MEMS) using the DAKOTA software. Hurtado aims at exploiting the complementary nature of RDO and RBDO probabilistic optimization approaches, using effective expansion techniques. Under this viewpoint, an efficient approximate methodology that integrates RDO and RBDO is proposed, in an effort to allow the designer to foresee the implications of adopting RDO or RBDO in the optimization process of probabilistic applications and to combine them in an optimum manner. On this basis, the concept of “robustness assurance’’ in structural design is introduced, in a similar manner to the “quality assurance’’ in the construction phase. For this purpose, a practical method for robust optimal design interpreted as entropy minimization is presented. Illustrative examples are presented to elucidate the advantages of the proposed approach. The robustness function is a measure of the performance of structural systems and expresses the greatest level of non-probabilistic uncertainty at which any constraint on structural performance cannot be violated. Kanno and Takewaki propose an efficient scheme for robust design optimization of trusses under various uncertainties. The structural optimization problem is formulated in the framework of an info-gap decision theory, aiming at maximizing the robustness function and is solved using semi-definite programming methods. Characteristic truss examples are used to demonstrate the efficiency of the proposed methodology. In his chapter, Doltsinis advocates the importance of an elaborate consideration of random scatter in industrial engineering with regard to reliability, and for securing standards of operation performance (robustness). For this purpose, synthetic Monte Carlo sampling and analytic Taylor series expansion that offer alternatives of stochastic analysis and design improvement are described. The robust optimum design problem is formulated as a two-criteria task that involves minimization of mean value and standard deviation of the objective function, while randomness of the constraints is also considered. Numerical applications justify the efficiency of the proposed approach are presented with linear and nonlinear structural response. Takewaki and Ben-Haim present a robust design concept, capable of incorporating uncertainties for both demand (loads) and capacity (various structural design parameters) of a dynamically loaded structure. Since uncertainties are prevalent in many cases, it is necessary to satisfy critical performance requirements, rather than to optimize performance, and to maximize the robustness to uncertainty. In the proposed implementation, the so called, “info-gap models of uncertainty’’ are used to represent

XX

Preface

uncertainties in the Fourier amplitude spectrum of the dynamic loading and the basic structural parameters related the vibration model of the structure. Furthermore, earthquake input energy is introduced as a new measure of structural performance for passively controlled structures and uncertainties of damping coefficients of control devices are also considered. Ganzerli and De Palma focus on the use of convex models of uncertainty with genetic algorithms for optimal structural design. Convex models theory together with probability and fuzzy sets, convex models can be considered part of the so-called “uncertainty triangle’’. Following, a literature review on convex models and their applications a description of convex models theory as an efficient alternative way to deal with problems having severe structural uncertainties, is presented. Subsequently, applications including the use of convex models of uncertainty combined with genetic algorithms for optimal structural design of trusses are demonstrated, and directions for further research in this area are given. In the last chapter, Lagaros, Tsompanakis, Fragiadakis, Plevris and Papadrakakis present efficient methodologies for performing standard RBDO and combined reliability-based and robust design optimization (RRDO) of stochastic structural systems in a multi-objective optimization framework. The proposed methodologies incorporate computationally efficient structural optimization and probabilistic analysis procedures. The optimization part is performed with evolutionary methods while the probabilistic analysis is carried out with the Monte Carlo Simulation (MCS) method with the Latin Hypercube Sampling (LHS) technique for the reduction of the sample size. In order to reduce the excessive computational cost and make the whole procedure feasible for real-world engineering problems the use of Neural Networks (NN) based metamodels is incorporated in the proposed methodology. The use of NN is motivated by the time-consuming repeated FE solutions required in the reliability analysis phase and by the evolutionary optimization algorithm during the optimization process. The editors of this book would like to express their deep gratitude to all the contributors for their most valuable support during the preparation of this volume and for their time and effort devoted to the completion of their contributions. In addition, we are most appreciative to the Book Series Editor, Professor Dan M. Frangopol, for his kind invitation to edit this volume, for preparing the foreword of this book, and for his constructive comments and suggestions offered during the publication process. Finally, the editors would like to thank all the personnel of Taylor and Francis Publishers, especially Germaine Seijger, Richard Gundel, Lukas Goosen, Tessa Halm, Maartje Kuipers and Janjaap Blom, for their most valuable support for the publication of this book.

Yiannis Tsompanakis Nikos D. Lagaros Manolis Papadrakakis September 2007

Brief Curriculum Vitae of the Editors

Yiannis Tsompanakis is Assistant Professor in the Department of Applied Sciences of the Technical University of Crete, Greece, where he teaches structural and computational mechanics as well as earthquake engineering lessons. His scientific interests include computational methods in structural and geotechnical earthquake engineering, structural optimization, probabilistic mechanics, structural assessment and the application of artificial intelligence methods in engineering. Dr. Tsompanakis has published many scientific papers and is the co-editor of several books in computational mechanics. He is involved in the organization of minisymposia and special sessions in international conferences as well as special issues of scientific journals as guest editor. He serves as a board member in various conferences, organized the COMPDYN-2007 conference together with the other editors of this book and acts as a co-editor of the resulting selected papers volume.

Nikos D. Lagaros is Lecturer of structural dynamics and computational mechanics in the School of Civil Engineering of the National Technical University of Athens, Greece. His research activity is focused on the development and the application of novel computational methods and information technology to structural and earthquake engineering analysis and design. In addition, Dr. Lagaros has provided consulting and expert-witness services to private companies and federal government agencies in Greece. He also serves as a member of the editorial board and reviewer of various international scientific journals. He has published numerous scientific papers, and is the co-editor of a number of forthcoming books, one of which is dealing with innovative soft computing applications in earthquake engineering. Nikos Lagaros is co-organizer of COMPDYN 2007 and co-editor of its selected papers volume.

XXII

Brief Curriculum Vitae of the Editors

Manolis Papadrakakis is Professor of Computational Structural Mechanics in the School of Civil Engineering at the National Technical University of Athens, Greece. His main fields of interest are: large-scale, stochastic and adaptive finite element applications, nonlinear dynamics, structural optimization, soil-fluid-structure interaction and soft computing applications in structural engineering. He is co-Editor-in-chief of the Computer Methods in Applied Mechanics and Engineering Journal, an Honorary Editor of the International Journal of Computational Methods, and an Editorial Board member of a number of international scientific journals. He is also a member of both the Executive and the General Council of the International Association for Computational Mechanics, Chairman of the European Committee on Computational Solid and Structural Mechanics and Vice President of the John Argyris Foundation. Professor Papadrakakis has chaired many international conferences and presented numerous invited lectures. He has written and edited various books and published a large variety of scientific articles in refereed journals and book chapters.

List of Contributors

Adams, B.M., Sandia National Laboratories, Albuquerque, NM, USA Agarwal, H., General Electric Global Research, Niskayuna, NY, USA Allen, M., University of Colorado at Boulder, Boulder, CO, USA Aoues, Y., University Blaise Pascal, France Beck, J.L., California Institute of Technology, CA, USA Ben-Haim, Y., Technion, Haifa, Israel Bichon, B.J., Vanderbilt University, Nashville, TN, USA Chateauneuf, A., University Blaise Pascal, France De Palma, P., Gonzaga University, Spokane, WA, USA Doltsinis, I., University of Stuttgart, Stuttgart, Germany Eldred, M.S., Sandia National Laboratories, Albuquerque, NM, USA Fragiadakis, M., University of Thessaly, Volos, Greece Frangopol, D.M., Lehigh University, Bethlehem, PA, USA Ganzerli, S., Gonzaga University, Spokane, WA, USA Huh, J.S., Korea Aerospace Research Institute, Daejeon, Korea Hurtado, J.E., National University of Colombia, Manizales, Colombia Joanni, A.E., Technical University of Munich, Munich, Germany Kanno, Y., University of Tokyo, Tokyo, Japan Kharmanda, G., Aleppo University, Aleppo, Syria Kokkolaras, M., University of Michigan, Ann Arbor, MI, USA Kwak, B.M., Korea Advanced Institute of Science and Technology, Daejeon, Korea Lagaros, N.D., National Technical University of Athens, Athens, Greece Lee, S.H., Northwestern University, Evanston, IL, USA Liang, J., Oakland University, Rochester, MI, USA Mahadevan, S., Vanderbilt University, Nashville, TN, USA Maute, K., University of Colorado at Boulder, Boulder, CO, USA Mourelatos, Z.P., Oakland University, Rochester, MI, USA Nikolaidis, E., University of Toledo, Toledo, OH, USA Papadrakakis, M., National Technical University of Athens, Athens, Greece Papalambros, P.Y., University of Michigan, Ann Arbor, MI, USA Patel, N.M., University of Notre Dame, Notre Dame, IN, USA Plevris, V., National Technical University of Athens, Athens, Greece Polak, E., University of California, Berkeley, CA, USA Rackwitz, R., Technical University of Munich, Munich, Germany

XXIV

List of Contributors

Renaud, J.E., University of Notre Dame, Notre Dame, IN, USA Royset, J.O., Naval Postgraduate School, Monterey, CA, USA Sørensen, J.D., Aalborg University, Aalborg, Denmark Taflanidis, A.A., California Institute of Technology, CA, USA Takewaki, I., Kyoto University, Kyoto, Japan Tillotson, D., University of Notre Dame, Notre Dame, IN, USA Tovar, A., National University of Colombia, Bogota, Colombia Tsompanakis, Y., Technical University of Crete, Chania, Greece Weickum, G., University of Colorado at Boulder, Boulder, CO, USA Wu, Y.-T., Applied Research Associates Inc., Raleigh, NC, USA Zhou, J., Oakland University, Rochester, MI, USA

Author Data

Adams, B.M. Sandia National Laboratories PO Box 5800, MS 1318 Albuquerque, NM 87185-1318 USA Phone: (505)284-8845 Fax: (505)284-2518 Email: [email protected] Agarwal, H. General Electric Global Research Niskayuna, New York, 12309 USA Phone: (574) 631-9052 Fax: (574) 631-8341 Email: [email protected] Allen, M. Research Assistant Center for Aerospace Structures Department of Aerospace Engineering Sciences University of Colorado at Boulder Boulder, CO 80309-0429, USA Phone: (303) 492 0619 Fax: (303) 492 4990 Email: [email protected] Aoues, Y. Laboratory of Civil Engineering University Blaise Pascal Complexe Universitaire des Cézeaux, BP 206 63174 Aubière Cedex, France Phone: +33(0)473407532 Fax: +33(0)473407494 Email: [email protected]

XXVI A u t h o r D a t a

Beck, J.L. Professor Engineering and Applied Science Division California Institute of Technology Pasadena, CA 91125 USA Phone: (626) 395-4139 Fax: (626) 568-2719 Email: [email protected] Ben-Haim, Y. Professor Faculty of Mechanical Engineering Technion – Israel Institute of Technology Haifa 32000, Israel Phone: 972-4-829-3262 Fax: 972-4-829-5711 Email: [email protected] Bichon, B.J. PhD Student Civil and Environmental Engineering Vanderbilt University VU Station B 351831 Nashville, TN 37235 USA Phone: 615-322-3040 Fax: 615-322-3365 Email: [email protected] Chateauneuf, A. Professor Polytech’Clermont-Ferrand Department of Civil Engineering University Blaise Pascal Complexe Universitaire des Cézeaux, BP 206 63174 Aubière Cedex, France Phone: +33(0)473407526 Fax: +33(0)473407494 Email: [email protected] De Palma, P. Professor Department of Computer Science School of Engineering and Applied Science Gonzaga University Spokane, WA 99258-0026 USA

Author Data

Phone: 509-323-3908 Email: [email protected] Doltsinis, I. Professor Institute for Statics and Dynamics of Aerospace Structures Faculty of Aerospace Engineering and Geodesy University of Stuttgart Pfaffenwaldring 27 D-70569 Stuttgart, Germany Phone: 0711-685-67788 Fax: 0711-685-63644 Email: [email protected] Eldred, M.S. Sandia National Laboratories P.O. Box 5800, Mail Stop 1318 Albuquerque, NM 87185-1318 USA Phone: (505)844-6479 Fax: (505)284-2518 Email: [email protected] Fragiadakis, M. Lecturer Faculty of Civil Engineering University of Thessaly Pedion Areos, Volos 383 34, Greece Phone: +30-210-748 9191 Fax: +30-210-772 1693 Email: [email protected] Frangopol, D.M. Professor of Civil Engineering and Fazlur R. Khan Endowed Chair of Structural Engineering and Architecture Department of Civil and Environmental Engineering Center for Advanced Technology for Large Structural Systems (ATLSS Center) Lehigh University 117 ATLSS Drive, Imbt Labs Bethlehem, PA 18015-4729, USA Phone: 610-758-6103 or 610-758-6123 Fax: 610-758-4115 or 610-758-5553 Email: [email protected] Ganzerli, S. Associate Professor Department of Civil Engineering School of Engineering

XXVII

XXVIII A u t h o r D a t a

Gonzaga University Spokane, WA 99258-0026 USA Phone: 509-323-3533 Fax: 509-323-5871 Email: [email protected] Huh, J.S. Senior Researcher Engine Department/KHP Development Division Korea Aerospace Research Institute 45 Eoeun-Dong, Yuseong-Gu Daejeon 305-330, Republic of Korea Phone: +82-42-860-2334 Fax: +82-42-860-2626 Email: [email protected] Hurtado, J.E. Professor Universidad Nacional de Colombia Apartado 127 Manizales Colombia Phone: +57-68863990 Fax: +57-68863220 Email: [email protected] Joanni, A.E. Research Engineer Institute for Materials and Design Technical Univerisity of Munich D-80290 München, Germany Phone: +49 89 289-25038 Fax: +49 89 289-23096 Email: [email protected] Kanno, Y. Assistant Professor Department of Mathematical Informatics Graduate School of Information Science and Technology University of Tokyo, Tokyo 113-8656, Japan Phone & Fax: +81-3-5841-6906 Email: [email protected] Kharmanda, G. Dr Eng Faculty of Mechanical Engineering

Author Data

University of Aleppo Aleppo – Syria Phone: +963-21-5112 319 Fax: +963-21-3313 910 Email: [email protected] Kokkolaras, M. Associate Research Scientist, Research Fellow Optimal Design (ODE) Laboratory Mechanical Engineering Department University of Michigan 2250 G.G. Brown Bldg. 2350 Hayward Ann Arbor, MI 48109-2125, USA Phone: (734) 615-8991 Fax: (734) 647-8403 Email: [email protected] Kwak, B.M. Samsung Chair Professor Center for Concurrent Engineering Design Department of Mechanical Engineering Korea Advanced Institute of Science and Technology 373-1 Guseong-dong, Yuseong-gu Daejeon 305-701 Republic of Korea Phone: +82-42-869-3011 Fax: +82-42-869-8270 Email: [email protected] Lagaros, N.D. Lecturer Institute of Structural Analysis & Seismic Research Faculty of Civil Engineering National Technical University of Athens Zografou Campus Athens 157 80, Greece Phone: +30-210-772 2625 Fax: +30-210-772 1693 Email: [email protected] Lee, S.H. Postdoctoral Research Fellow Department of Mechanical Engineering Northwestern University 2145 Sheridan Road Tech B224 Evanston IL 60201, USA Phone: +1-847-491-5066

XXIX

XXX A u t h o r D a t a

Fax: +1-847-491-3915 Email: [email protected] Liang, J. Graduate Research Assistant Department of Mechanical Engineering Oakland University Rochester, MI 48309-4478 USA Phone: (248) 370-4185 Fax: (248) 370-4416 Email: [email protected] Mahadevan, S. Professor Civil and Environmental Engineering Vanderbilt University VU Station B 351831 Nashville, TN 37235, USA Phone: 615-322-3040 Fax: 615-322-3365 Email: [email protected] Maute, K. Associate Professor Center for Aerospace Structures Department of Aerospace Engineering Sciences University of Colorado at Boulder Room ECAE 183, Campus Box 429 Boulder, Colorado 80309-0429, USA Phone: (303) 735 2103 Fax: (303) 492 4990 Email: [email protected] Mourelatos, Z.P. Professor Department of Mechanical Engineering Oakland University Rochester, MI 48309-4478 USA Phone: (248) 370-2686 Fax: (248) 370-4416 Email: [email protected] Nikolaidis, E. Professor Mechanical Industrial and Manufacturing Engineering Department

Author Data

4035 Nitschke Hall The University of Toledo Toledo, OH 43606 USA Phone: (419) 530-8216 Fax: (419) 530-8206 Email: [email protected] Papadrakakis, M. Professor Institute of Structural Analysis & Seismic Research Faculty of Civil Engineering National Technical University of Athens Zografou Campus Athens 157 80, Greece Phone: +30-210-772 1692 & 4 Fax: +30-210-772 1693 Email: [email protected] Papalambros, P.Y. Professor Director, Optimal Design (ODE) Laboratory University of Michigan 2250 GG Brown Building Ann Arbor, Michigan 48104-2125 USA Phone: (734) 647-8401 Fax: (734) 647-8403 Email: [email protected] Patel, N.M. Graduate Research Assistant Design Automation Laboratory Aerospace and Mechanical Engineering 365 Fitzpatrick Hall of Engineering University of Notre Dame Notre Dame, Indiana 46556-5637 USA Phone: (574) 631-9052 Fax: (574) 631-8341 Email: [email protected] Plevris, V. PhD Candidate Institute of Structural Analysis & Seismic Research Faculty of Civil Engineering National Technical University of Athens

XXXI

XXXII A u t h o r D a t a

Zografou Campus Athens 157 80, Greece Phone: +30-210-772-2625 Fax: +30-210-772-1693 Email: [email protected] Polak, E. Professor Emeritus, Professor in the Graduate School Department of Electrical Engineering and Computer Sciences University of California at Berkeley 255M Cory Hall 94720-1770 Berkeley, CA USA Phone: 510-642-2644 Fax: 510-841-4546 Email: [email protected] Rackwitz, R. Professor Institute for Materials and Design Technical Univerisity of Munich D-80290 München, Germany Phone: +49 89 289-23050 Fax: +49 89 289-23096 Email: [email protected] Renaud, J.E. Professor Design Automation Laboratory Aerospace and Mechanical Engineering 365 Fitzpatrick Hall of Engineering University of Notre Dame Notre Dame, Indiana 46556-5637 USA Phone: (574) 631-8616 Fax: (574) 631-8341 Email: [email protected] Royset, J.O. Assistant Professor Operations Research Department Naval Postgraduate School Monterey, California 93943 USA Phone: 1-831-656-2578 Fax: 1-831-656-2595 Email: [email protected]

Author Data

Sørensen, J.D. Professor Department of Civil Engineering Aalborg University Sohngardsholmsvej 57 9000 Aalborg, Denmark Phone: +45 9635 8581 Fax: +45 9814 8243 Email: [email protected] Taflanidis, A.A. Ph.D Candidate Engineering and Applied Science Division California Institute of Technology Pasadena, CA 91125 USA Phone: (626) 379-3570 Fax: (626) 568-2719 Email: [email protected] Takewaki, I. Professor Department of Urban and Environmental Engineering Graduate School of Engineering Kyoto University Kyotodaigaku-Katsura, Nishikyo-ku, Kyoto 615-8540 Japan Phone: +81-75-383-3294 Fax: +81-75-383-3297 Email: [email protected] Tillotson, D. Research Assistant Design Automation Laboratory Aerospace and Mechanical Engineering 365 Fitzpatrick Hall of Engineering University of Notre Dame Notre Dame, Indiana 46556-5637 USA Phone: (574) 631-8616 Fax: (574) 631-8341 Email: [email protected] Tovar, A. Assistant Professor Department of Mechanical and Mechatronic Engineering Universidad Nacional de Colombia

XXXIII

XXXIV A u t h o r D a t a

Cr. 30 45-03, Of. 453-401 Bogota, Colombia Phone: +57-3165320 - 3165000 ext. 14062 Fax: +57-3165333 - 3165000 ext. 14065 Email: [email protected] Tsompanakis, Y. Assistant Professor Department of Applied Sciences Technical University of Crete University Campus Chania 73100, Crete, Greece Phone: +30 28210 37 634 Fax: +30 28210 37 843 Email: [email protected] Weickum, G. Graduate Research Assistant Center for Aerospace Structures Department of Aerospace Engineering Sciences University of Colorado at Boulder Room ECAE 188, Campus Box 429 Boulder, Colorado 80309-0429 USA Phone: (303) 492 0619 Fax: (303) 492 4990 Email: [email protected] Wu, Y.-T. Fellow, Applied Research Associates, Inc. 8540 Colonnade Center Dr., Ste 301 Raleigh, NC 27615 USA Phone: 919-582-3335 or 919-810-1788 Email: [email protected] Zhou, J. Graduate Research Assistant Department of Mechanical Engineering Oakland University Rochester, MI 48309-4478 USA Phone: (248) 370-4185 Fax: (248) 370-4416 Email: [email protected]

Part 1

Reliability-Based Design Optimization (RBDO)

Chapter 1

Principles of reliability-based design optimization Alaa Chateauneuf University Blaise Pascal, France

ABSTRACT: Reliability-Based Design Optimization (RBDO) aims at searching for the best compromise between cost reduction and safety assurance, by controlling the structural uncertainties allover the design process, which cannot be achieved by deterministic optimization. This chapter describes the fundamental concepts in RBDO. It aims to explain the role of uncertainties in deriving the optimal solution, where emphasis is put on the comparison with conventional deterministic optimization. The interest of RBDO formulation can also be extended to cover different design aspects, such as multi-component reliability analysis, safety factor calibration, multi-objective applications and time-variant problems.

1 Introduction The design of structures must fulfill a number of different criteria, such as cost, safety, performance and durability, leading to conflicting requirements to be simultaneously considered by the engineer. Therefore, the challenge in the design process is how to define the best compromise between contradictory design requirements. Moreover, the complexity of the design process does not allow for simultaneous optimization of all the design criteria with respect to all the parameters. Traditionally, this complexity is reduced by dividing the process into simpler sub-processes where each requirement can be handled separately. The designer can hence concentrate his effort on only one goal, generally the cost, and then checks if the other requirements can be, more or less, fulfilled. If necessary, further adjustments are introduced in order to improve the obtained solution. However, this procedure cannot assure performance-based optimal design. In structural engineering, the deterministic optimization procedures have been successfully applied to systematically reduce the structure cost and to improve the performance. However, uncertainties related to design, construction and loading, lead to structural behavior which does not correspond to the expected optimal performance. The gap between expected and obtained performances is even larger when the structure is optimized, as the remaining margins are reduced to their lower bounds; in other terms, the optimal structure is usually sensitive to uncertainties. In deterministic design, the propagation of uncertainties is usually hidden by the use of the well-known “safety factors’’, without direct connection with reliability specifications. Traditionally, the optimal cost is looked for by iterative search procedures, while the required reliability level is assumed to be ensured by the applied safety factors, as described by the design codes of practice. As a matter of fact, these safety factors are calibrated for average

4

Structural design optimization considering uncertainties

design situations and cannot ensure consistent reliability levels for specific design conditions. They may even lead to poor design, as the optimization procedure will search for the weakest region in the domain covered by the code of practice. This weakest region often presents not only the lowest cost but also the lowest safety. The deterministic optimal design is pushed to the admissible domain boundaries, leaving very little space for safety margins in design, manufacturing and operating processes. Moreover, the optimization process leads to a redistribution of the roles of uncertainties which can only be controlled by reliability assessment on the basis of the sensitivity measures. For these reasons, the Deterministic Design Optimization (DDO) cannot ensure appropriate reliability levels. If the DDO solution is more reliable than required, the losses can be avoided in construction and manufacturing costs; however, if the reliability is lower than required, the economic solution is not really achieved, because of the increase of the failure rate, leading to failure losses higher than the expected money saving. In this sense, the Reliability-Based Design Optimization (RBDO) becomes a very powerful tool for robust and cost-effective designs (Frangopol 1995). The RBDO aims to find a balanced design by reducing the expected total cost, which is defined in terms of the initial cost (i.e. including design, manufacturing, transport and construction costs), the failure cost, the operation cost and the maintenance costs. In addition, the RBDO takes the benefit of driving the search procedure by the wellcontrolled variables having great impact on the total cost. On the other side, the variables with high uncertainties are penalized independently of their mechanical role. In this sense, the system robustness is achieved as the role of highly uncertain and fluctuating variables is diminished during the optimization process. Contrary to the DDO, the solution does not lie in the weakest domain of the design code of practice, but a better compromise is defined by satisfying the target reliability levels. The RBDO can also be applied for robust design purposes, where the mean values of random variables are used as nominal design parameters, and the cost is minimized under a prescribed probability. Therefore, the solution of RBDO provides not only an improved design but also a higher level of confidence in the design. From the practical point of view, solving the RBDO problems is a heavy task because of the nested nonlinear procedures: optimization procedure, reliability analysis and numerical simulation of structural systems. Several methods have been developed for solving efficiently this problem, in order to allow for complex industrial applications; this topic will be discussed in a subsequent chapter by Chateauneuf and Aoues. This chapter aims at describing the RBDO principles, in order to give a clear vision of the links between classical deterministic approach and the reliability-based one. It emphasizes the fact that the deterministic optimization, based on safety factor considerations, is not anymore sufficient for safety control and assurance. The Reliability-Based Design Optimization has the advantage of ensuring a minimum cost without affecting the target safety level. At the end of the present chapter, the use of the RBDO in different kinds of engineering problems is briefly discussed in order to show how large can be the application spectrum.

2 Historical background Since the beginning of the twentieth century, the need for rational way to consider structural safety motivated a number of researchers, such as Forsell (1924), Wierzbicki

P r i n c i p l e s o f r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

5

(1936) and Lévi (1948). In the conference on structural safety, held in Liège 1948, by the Association Internationale des Ponts et Charpentes, Torroja stated, probably for the first time, that the reduction of the total cost, have to include not only the construction cost, but also the expected failure cost. CT = CI + CF

(1)

where CT is the expected total cost, CI is the initial cost (i.e. design and construction cost) and CF is the expected failure cost. This expression has been easily approved, as the increase of construction cost should lead to higher safety margin and so decreasing the failure probability. Even that the formulation of the RBDO is known since 1948 (and even earlier), the direct application was impossible because of the difficulties related to the failure probability computation for realistic structures. With the development of the reliability theory starting in the 1950s, the solution procedures became available in 1970s and improved in the 1980s, in order to allow for the analysis of practical engineering structures. However, till now, the difficulty to estimate the failure cost is still remain a main problem, especially when dealing with human lives and environmental deterioration. On the basis of the target reliability index, the RBDO is really born in the second half of 1980s and developed along 1990s. Nowadays, the industrial applications of RBDO still face many difficulties due to the very high computational effort required to solve large-scale systems. Most of practical applications of structural optimization requires at least three conflicting goals (Kuschel and Rackwitz 1997): – – –

Low structural cost, including or not the expected failure cost. High reliability levels for components and systems. Good structural performance under various operating conditions.

Actually, the new trend is to include the inspection, maintenance, repair and operating costs in the definition of the expected total cost CT , in order to reach a performance-based design on the basis of multi-criteria considerations (Frangopol 2000). A comprehensive overview of these approaches is given by Frangopol and Maute (2003).

3 Reliability analysis The design of structures requires the verification of a certain number of rules resulting from the knowledge of physics and mechanics, combined with the experience of designers and constructors. These rules come from the necessity to limit the loading effects such as stresses and displacements. Each rule represents an elementary event and the occurrence of several events leads to a failure scenario. In addition to the deterministic variables d to be used in the system control and optimization, the uncertainties are modeled by stochastic variables affecting the failure scenario. The knowledge of these variables is not, at best, more than statistical information and we admit a representation in the form of random variables X, whose realizations are noted x. For a given design rule, the basic random variables are defined by their probability distribution with some statistical parameters (generally, the mean and the standard deviation).

6

Structural design optimization considering uncertainties

Safety domain

Joint distribution

x2

G>0

Pf

Failure domain

G<0 x1

Figure 1.1 Joint distribution and failure probability.

The safety is defined as the state where the structure is able to fulfil all the operating requirements: mechanical and serviceability, for which it is designed, during the whole lifetime. To evaluate the failure probability with respect to a given failure scenario, a performance function G(x, d) is defined by the condition of good operation of the structure. The limit between the state of failure G(x, d) ≤ 0 and the state of safety G(x, d) > 0 is known as the limit state surface G(x, d) = 0. Having the performance function G(x, d), known also as the limit state function or the safety margin, it is possible to evaluate the probability of failure by integrating the joint probability density over the failure domain (Figure 1.1):
Pf (d) = fX (x, d) dx (2) G(x,d)≤0

It is to be noted that the joint density function fX (x, d) depends on the design parameters d only when the distribution parameters belong to the design variables; this is especially the case when the mean value is considered as a design variable in the optimization process. There is a special case when the performance function is simply written by the margin between the resistance R and the load effect S, where both variables are independent normal random variables. The performance function and the failure probability are simply given by: G(X, d) = R − S Pf (d) =
(−β(d))

with:

m R − mS β(d) =  σR2 + σS2

(3)

where
( · ) is the standard gaussian cumulated distribution function, β(d) is the reliability index, mR , mS , σR and σS are respectively the means and standard deviations of the resistance and load effect. For this simple configuration, the optimization variable could be the mean design strength, and probably, in some cases, the mean load effect.

P r i n c i p l e s o f r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

7

It to note that also standard deviations can be taken as optimization variables if the relationship between the quality control and the structural cost can be established. In practice, the performance function cannot be written in a simple linear form of normal variables, and equation 3 can rarely be applied. It is thus necessary to evaluate, more or less precisely, the failure probability as given in equation 2. Direct integration is impossible even for small structures due to: 1) the high-required precision, 2) the computation cost of the mechanical response, and 3) the multidimensional space. Numerical methods have to be applied to give an approximation of the failure probability. Three methods are commonly used for this purpose: –

–

–

Monte Carlo simulations allowing to estimate the failure probability for any general problem. It has two main advantages: 1) the possibility to deal with practically any mechanical or physical model (linear, nonlinear, continuous, discrete, . . .) and 2) the simple implementation without any modification of the mechanical model (e.g. finite element software) which is considered as a blackbox. However, the two main drawbacks are: 1) the very high computational time, especially for realistic structures with low failure probability and 2) the numerical noise due to random sampling, leading to non-monotonic estimates during simulations, and consequently, it becomes impossible to get accurate and stable evaluation of the response gradient. Although computation time can be reduced by using importance sampling and other variance reduction techniques, the numerical noise still remains a serious difficulty for practical applications in RBDO. First- and Second-Order Reliability Methods, known as FORM/SORM, which are based on the approximation of the performance function in the standard gaussian space by using polynomial series. An optimization algorithm is applied to search for the design point, called also the most probable failure point or β-point, which is the nearest failure point to the origin in the normal space. Then, linear (FORM) or quadratic (SORM) approximations are adopted for the performance function in order to get an asymptotic approximation of the failure probability. It is approved that FORM is usually sufficient for the majority of practical engineering systems. In RBDO context, FORM/SORM techniques have the advantages of: 1) high numerical efficiency; and 2) direct computation of the gradients of the reliability index, and consequently of the failure probability. The main drawbacks are: 1) the limited precision and convergence problems in some cases, especially for highly nonlinear limit states; and 2) the computation time for large number of random variables. Response Surface Methods (RSM), which are commonly used to approximate the mechanical response of the structure, by building what is called a meta-model. Quadratic polynomials are shown to be suitable for localized approximation of structural systems. The large part of the computational cost lies in the evaluation of the polynomial coefficients. Then, the failure probability can be simply evaluated by using the response surface which is an analytical expression, instead of the mechanical model itself (generally, complex finite element model). The advantages are mainly: 1) the reduction of the computation time for moderate number of random variables; and 2) the possibility of coupling reliability and optimization algorithms to achieve high efficiency. The most common drawback lies in the large number of mechanical calls for moderate and high number of variables.

8

Structural design optimization considering uncertainties

x2

u2

Physical space

Normalized space Failure domain Gu(u, d)  0 P* MPP

Failure domain G(x, d)  0 mX

2

Safe domain mX

G(x, d)
0

b x1

1

u*2

a

Gu(u, d)
0

u*1

u1

Figure 1.2 Reliability index and the Most Probable Failure Point (MPP).

In First Order Reliability Method, the failure probability Pf is approximated in terms of the reliability index β according to the expression: Pf (d) = Pr[G(X, d) ≤ 0] ≈
(−β(d))

(4)

where Pr[·] is the probability operator and
( · ) is the standard Gaussian cumulated function. The invariant reliability index β, introduced by Hasofer and Lind (1974), is evaluated by solving the constrained optimization problem (Figure 1.2):

β = min

u =



(Ti (x))2

i

(5)

under the constraint: G(T(x), d) ≤ 0 where u is the distance between the median point (corresponding to the space origin) and the failure subspace in the normalized space u and T(x) is an appropriate probabilistic transformation: i.e. ui = Ti (x). The image of the performance function G(x) in the normalized space is noted: Gu (u, d) = G(T(x), d). The solution of this problem is called the Most Probable Failure Point, the design point or the β-point; it is noted P∗ , or either x* or u*, whether physical or normalized space is considered, respectively. At this point, the following relationship holds: β = u. For the case of two random variables, Figure 1.3 illustrates the important points involved in structural design: the mean point represents the average stress and strength at operation, the characteristic values are loading and resistance values that can be guaranteed in the design process (they correspond to small probability to find higher loading level or to find lower strengths; percentiles of 95% or 5% are commonly adopted) and finally the Most Probable Failure Point (MPP) where the failure configuration has the highest joint probability density. While the reliability analysis aims at finding the Most Probable failure Point, the design procedure aims at setting the characteristic and mean values of strength and dimensions, according to economical considerations.

P r i n c i p l e s o f r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

s

Strength R

fS(s), fR(r) Load effect S

9

P*

Limit state

s*

sk

xk G(x, d) = 0

mS

mX r

mS

sk

rk mR s* = r*

s, r

r* rk

mR

Figure 1.3 Mean, characteristic and design points.

Alternatively to equation 2, the reliability level of a structure can also be characterized by the performance function Pp defined as:
Pp (d, p) = fX (x, d) dx (6) G(x,d)≤p

where the subscript p is the performance measure (in standard reliability, p is set to zero). This formulation can be useful for specific RBDO formulations (see chapter by Chateauneuf and Aoues). 3.1

S ystem reliability analys is

Due to optimization, the structural components are strongly stretched close to the limit state, and their contribution in the overall safety becomes significant. That is why structural reliability cannot be correctly computed unless the complete system is considered, by taking into consideration the contributions of all the failure modes through appropriate modeling of system configurations, material behavior, load variability, strength uncertainty and statistical correlation. As structures are made of the assembly of several members, the overall ultimate capacity is highly conditioned by the redundancy degree. For many structures, several components can reach their ultimate capacity much before reaching the overall structural failure load. On the other hand, the structure could contain a number of critical members, leading to the overall failure if any one of them fails. In this context, the system reliability can be quite different from the reliability of its components. In the last decades, many research works have been dedicated to compute the system reliability, especially for series and parallel systems. A series system, representing a “weakest-link’’ chain, fails if any link fails; superstructures and building foundations are generally good examples of a series system. A parallel system implies that each component contributes more or less in the structural good-standing; the system failure takes place if all components fail.

10

Structural design optimization considering uncertainties

Practical expressions for system reliability include lower and upper bounds for both series and parallel systems, some of these bounds consider the correlation between pairs of potential failure modes. Also, more complex system models involving mixed series-parallel systems can be used (Ditlevsen and Madsen 1996). For series and parallel systems, the first order approximation of the failure probabilities can be computed as following: 



Pf = Pr

Gj (X, d) ≤ 0 ≈ 1 −
m (β(d), ρ)

j

 Pf = Pr





for series system



(7)

Gj (X, d) ≤ 0 ≈
m (−β(d), ρ)

for parallel system

j

where
m (β(d), ρ) is the multi-dimensional standard normal distribution, β(d) is the vector of the reliability indices for the different modes and ρ is the matrix of correlations between the failure modes. For practical RBDO analysis, the failure probability can be estimated by Ditlevsen bounds (Ditlevsen 1979), which is written for series system as:

Pf1 +

m  j=2

⎡ max ⎣Pfj −

j−1  k=1

⎤ Pfjk , 0⎦ ≤ Pfs ≤

m  j=1

Pfj −

m  j=2

max Pfjk k<j

(8)

where m is the number of dominant failure modes, Pfj is the failure probability of mode j and Pfjk is the probability of the intersection of modes j and k; in this expression, the failure probabilities follow a decreasing order.

4 Formulation of Reliability-Based Design Optimization The Reliability-Based Design Optimization (RBDO) aims at finding the optimal solution that fulfills the prescribed reliability requirements. The fluctuation of loads, the variability of material properties and the uncertainties regarding the analysis models, contribute to make the performance of the optimal design different from the expected one. In this sense, the optimization process has a large effect on the structural reliability. It is today well recognized that the safety factor approach cannot ensure the required safety levels, as they do not explicitly consider the probability of failure regarding some performance criteria. In other words, the optimal design resulting from deterministic optimization procedures does not necessarily meet the reliability targets. The RBDO allows us to consider the safety margin evolution, leading to the settlement of the best compromise between the life-cycle cost and the required reliability. This task is rather complicated due to the inherent non-deterministic nature of the input information. For this reason, many analysis methods have been developed to deal with the statistical nature of data. The process efficiency is mandatory to deal with realistic engineering problems (Kharmanda et al. 2002). The solution based on reliability concepts is rather robust, as the uncertain parameters are penalized during the design process, compared to a greater commitment of the well-controlled parameters.

P r i n c i p l e s o f r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

4.1

11

Insuf f iciency of Determinis tic Des ig n O p ti mi zati o n

In Deterministic Design Optimization (DDO), it is aimed to reduce the initial structural cost CI (d) under a number of constraints gj (d, γ), j = 1, 2, . . . , ng ; where d is the vector of design parameters and γ is the vector of partial safety factors. The optimization problem is thus written: min CI (d) d

subject to

gj (d, γ) ≤ 0 for

(9)

j = 1, 2, . . . , ng

In this problem, the structural safety is assumed to be ensured by the introduction of the safety factors within the constraint equations. A typical constraint for stress can be written as: g = σ − fy /γ where σ is the applied stress, fy is the yield stress and γ is the safety factor. Usually the strength fy is defined either by the mean value or by the characteristic value; the former is common in mechanical engineering and the latter is common in civil engineering. In the DDO, it is assumed that the safety factors are appropriate whatever the chosen optimal configuration. For most of systems, it can be shown that the safety level is not independent of the selected optimal design parameters. Figure 1.4a illustrates how the deterministic optimal design is defined on a constraint which is simply described by shifting the limit state. It can be said that the failure limit state g(d) is transformed to a safe limit state g(d, γ), by introducing the safety factor to take account for uncertainties. Starting from the initial point x0 , the DDO is based on the use of classical optimization algorithms to find the optimal design d∗ , which is generally located on the boundary of the reduced design space, including the safety factor. The Reliability-Based Design Optimization aims at finding the optimal solution, such that the failure limit state is kept sufficiently far from the operating point. In other words, the failure surface must lie on the iso-reliability level corresponding to the prescribed safety target (Figure 1.4b). It is clear that even for simple cases, the solution can be quite different from the deterministic optimization where homothetic Safe design constraint

s

Limit state

s

g(d )
0

Optimum design

γ

Optimum design Iso-reliability contours

Limit state n

ctio

g(d) = 0

d*

s*

s

Co

g(d, γ) = 0 x0 s

Co

(a)

d*

s* x0

n

du t re

r*

du t re

ctio

r

r (b)

r*

Figure 1.4 Comparison of optimal points corresponding to DDO and RBDO.

12

Structural design optimization considering uncertainties

fG(g)

Safety factor g1

1

g

g

Figure 1.5 Distribution of the global safety factor.

reduction of the design space is applied. In this sense, the RBDO can really ensure optimal cost, without compromising the structural safety. As a matter of fact, the uncertainties related to structural geometry, material properties and loading lead to stochastic cost, strength and stress in the structure, which automatically leads to random safety factors. When the optimal design is defined, it can be possible to see the effect of uncertainties on the global safety factor. This can be carried out by plotting the distribution of the strength-stress ratio, as illustrated in figure 1.5. In this case, the structural failure is observed when the safety factor realizations become less than unity. The failure probability can thus be computed by evaluating either Pr [γ ≤ 1] or Pr [G(d∗ , X) ≤ 0]. It is also to emphasize that the system uncertainties may lead to random total cost, which can be considered in one of the two following ways: –

–

If the optimal configuration is specified, the structure realization involves random variations in material properties, geometrical parameters, material unit cost, construction costs and failure costs. These viabilities and fluctuations produce a random total cost, where the probabilistic distribution depends on the inherent random variables. The goal of RBDO is usually to minimize the expected total cost. If the structural realizations are considered for cost optimization, a scatter of the optimal solutions is obtained, as different optimum is found for each structural realization. In other words, even if random variables are not involved in the cost, the optimal deterministic solution changes according to structure and loading fluctuations. Therefore, the total cost becomes random as solutions differ in terms of input uncertainties. This leads to what can be seen as a lack of robustness.

Deterministic optimization is even worst when multi-constraints or multicomponents are considered. The difficulty lies in the way to set the safety factors in order to ensure simultaneously safe and optimal design. As illustrated in figure 1.6, while the deterministic optimum leads to uncontrolled safety levels with respect to various limit states (due to the application of either the same safety factor or an inconsistent set of safety factors), the reliability-based design optimization looks for the situation where the safety levels can be simultaneously controlled for all the limit states. In this case, the optimum design is oriented such that safety requirements are fulfilled with respect to the uncertainty degrees; practically, greater margins are taken for largely scattered variables, while small margins are considered for well controlled variables.

P r i n c i p l e s o f r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

13

Deterministic optimum

x2

Limit state with safety factors Real limit state

Iso-reliability contours

Reliability-based optimum

x1

Figure 1.6 Comparison between DDO and RBDO solutions.

P

1

y

S1

v

2 S2

600 mm

u

x 800 mm

450 mm

Figure 1.7 Two-link structure under vertical force.

In other words, the design is driven by the variables with small uncertainties. That is why the Reliability-Based Design Optimization (RBDO) aims at searching for the best compromise between cost reduction and reliability assurance, by taking the system uncertainties into consideration; therefore, the RBDO ensures economical and safe design. It offers a good alternative to the safety factor approach, which is based on deterministic considerations and cannot take account for reduction of safety margins during the optimization procedure. In order to illustrate this idea, let us consider the two-bar system shown in figure 1.7. The system is supported at the end nodes and a vertical load P is applied at the internal node. The bar cross-sections are noted S1 and S2 for bars 1 and 2, respectively. The

14

Structural design optimization considering uncertainties

design criteria are related to member strengths and nodal displacements; buckling is assumed to be neglected. Under stress and deflection constraints, the deterministic optimization problem is written: min V = 1000 S1 + 750 S2 S1 ,S2

subject to g1 = F1 −

fY S1 ≤ 0 γσ (10)

fY S2 ≤ 0 γσ vL ≤0 g3 = v − γv

g2 = F2 − and with: F1 = 0.6P;

F2 = 0.8P;

P v= E

480 360 + S1 S2

 (11)

where E is the Young’s modulus, fY is the yield stress, vL is the limit displacement, γσ is the load safety factor and γv is the displacement safety factor. They are calculated by: 

fY S1 fY S2 ; γσ = min 0.6P 0.8P

 (12)

E S1 S2 vL γv = P (480 S1 + 360 S2 )

For deterministic optimization, these safety factors are taken as γσ = 1.5 and γv = 1.2. Considering only the two resistance limit states, Figure 1.8 shows the

S2

G1(d) ⴝ 0

G1(d, γ) ⴝ 0

Deterministic optimum 0.9P/fY

Iso-reliability contours G2(d, γ) ⴝ 0

P1*

0.6P/fY

Safety design domain G1(d, γ)  0 and G2(d, γ )  0

P*2

G2(d) ⴝ 0

Failure domain S1 0.8P/fY

1.2P/fY

Figure 1.8 Deterministic design and inconsistent reliability levels.

P r i n c i p l e s o f r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

15

deterministic optimum solution in the space of the design variables S1 and S2 . This optimum is obtained at the intersection of the shifted limit states, due to the application of the safety factor of 1.5. If the uncertainties on the cross-sections are considered by using independent normal variables with the same standard deviation, we get directly the iso-reliability contours in this superposed design/random space. It is clearly observed that the Most probable failure points P1∗ and P2∗ do not lie on the same reliability level. As the system represents a series combination, the bar 2 reduces the overall structural reliability and the target safety is not met. A more rational approach implies that the two MPP should be at the same reliability contour; for more complex systems, this should take into account the failure costs and system reliability models for effective setting of the MPP locations. It has to be stressed that even for this simple problem, the task is not that easy, it can be imagined how complex is to set appropriate safety factors for nonlinear limit states (e.g. the displacement constraint) with non-gaussian variables, including statistical correlations. For this reason, the DDO approach cannot clearly give convenient solution with consistent uncertainty considerations.

4.2

Reliability-bas ed des ign optimizatio n

Although the total cost includes all costs from construction till destruction and recycling, including the in-service costs, the high complexity of engineering systems leads to difficulties in dealing with both aspects: design and maintenance uncertainties. A common procedure consists in separating the design into two steps. In the first step, the structure is designed in order to avoid “failure’’ configurations, with respect to the limit states (ultimate, serviceability, fatigue, . . .). At this design stage, the optimization is applied to assure the structural survival (or good-standing) with the lowest cost. In the second step, the maintenance planning is optimized for the structure designed in the first step. In this way, the design optimization is first carried out to define the optimal structural configuration for which Reliability-Based Maintenance Optimization (RBMO) is performed to define the maintenance-inspection-replacement policy. In this sense, the total cost minimization is carried out in two times: minimizing the initial and failure costs at the first time and then minimizing the maintenance cost at the second time. Of course, this implies an approximation, as some design variables can play different roles in the cost components. For some engineering systems, the decoupling of design and maintenance costs may not lead to globally optimal costs, due to interaction between design decisions and deterioration rates and time-dependent failure probability. In other words, the variation of some variables may increase failure cost and decrease maintenance costs, and vice-versa. However, this approach is widely accepted in engineering practice. It has also a practical advantage, as the optimal design becomes independent of the maintenance policy, where the operating conditions (loading, environment, deterioration rates, costs, . . .) may strongly varies over the structure lifespan. The total cost CT depends on two kinds of variables (Kharmanda et al. 2002b): –

Design variables, noted d, which are the deterministic control parameters. They should be optimized for cost reduction. They can be mechanical parameters

16

Structural design optimization considering uncertainties

Cost

Expected total cost CT Expected failure cost CF
Cf Pf

Minimum expected total cost

Initial construction cost CI Failure probability Optimum reliability level

Pf

Figure 1.9 Evolution of the costs in function of the failure probability.

–

(e.g. geometrical dimensions, material properties) or probabilistic parameters (e.g. means of random distributions). Random variables, noted X, whose realizations are x, representing the uncertainties and the fluctuations in the system configuration. Each of the random variables is defined by a probabilistic distribution. They usually represent geometrical, material or loading uncertainties.

Basically, the RBDO aims at minimizing the total expected cost CT (Figure 1.9) which is given in terms of initial cost CI (including design, manufacturing, transport and construction costs) and direct failure cost Cf (Torroja 1948, Ditlevsen and Madsen 1996). min CT (d) = CI (d) + Cf Pf (d) d

subject to gj (d) ≤ 0

(13)

A more rigorous mathematical notation consists in writing E[CT (d, X)] instead of CT (d), as what is optimized is the expectation, not the cost itself (which is a random function); however, for simplicity, the notation CT (d) is maintained to indicate the expectation. This problem can also be written in terms of the utility function U(d) as following (Frangopol 1995): max U(d) = B(d) − CI (d) − L(d) x

subject to

gj (d) ≤ 0

(14)

where B is the benefit derived from the system operation, C I is the initial construction cost and L is the expected loss due to inspection, maintenance and failure. This total cost expression indicates that the possible increase of initial cost should be balanced by a decrease in the risk C F (i.e. product: Cf Pf ). The minimization is carried out for the design parameters such as member sizes, structural configuration

P r i n c i p l e s o f r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

17

and material parameters. These design parameters may correspond to probabilistic distribution parameters: cost is related to the mean value when it represents the nominal design value and to the standard deviation when it represents the quality control and the dispersion reduction aspects. Usually, the cost of consequences is taken as fixed, but in fact it should be a function of the failure probability. This means that the failure cost Cf (e.g. reconstruction cost, direct damage cost and pollution) is independent of the failure probability Pf , and consequently, the expected failure cost can be written: CF = Cf × Pf . This expression holds as long as the failure rate remains below a commonly accepted level. However, with abnormal failure rates, the failure cost Cf becomes a function of the failure probability according to indirect damage (propaganda, market losses, public opinion on company/authority, accelerated effects, . . .). For example in automotive industry, if a defect is observed for few cars, the failure cost for each car can be assumed equal to the repair of that car (added to some indemnity for the car owner). But if the defect is observed for a large number of cars, the company should repair the whole produced cars, beside the social damage to the company itself which can be traduced by significant selling losses. In other domains, such as nuclear energy, the failure cost can be a jump function as only one accident in a nuclear power plant leads to very high economic, social and political consequences. To take account for this failure rate dependence, it could be appropriate to estimate the failure cost by nonlinear functions, such as: CF = E[Cf ] ≈ Cf (Pf ) × Pf

(15)

where Cf (Pf ) may take one of the following forms: Polynomial Exponential Sigmoidal

Cf (Pf ) = Cf0 (1 + Pfα )  Cf0 for Pf ≤ Pf0 Cf (Pf ) = Cf0 exp(µ(Pf − Pf0 )α ) for Pf > Pf0 Cf 1 Cf (Pf ) = Cf0 + 1 + exp (−µ(Pf − Pf0 ))

where Cf0 and Cf1 are respectively the basic and the extra failure costs, Pf0 is the probability threshold, and α, µ are parameters to be estimated in terms of failure consequences. More generally, the expected total cost CT can be expressed in terms of all the costs involved in the structural system, from birth to death. It thus includes inspection, maintenance, repair and operating costs (Frangopol 2003), leading to: CT = CI + CF + CM + CS + CR + CU + CD

(16)

where CI is the initial construction cost, CF is the expected failure cost, usually defined as: CF = Cf × Pf , CM is the expected preventive maintenance cost, CS is the expected inspection cost, CR is the expected repair cost, CU is the expected use cost and CD is the expected recycling and destruction cost, which is particularly important for sensitive structures, such as nuclear powerplants.

18

Structural design optimization considering uncertainties

In practice, the design objective of only minimizing the expected total cost is not yet applicable, and is somehow dangerous from human point of view. For example, if the designer underestimates the failure consequences with respect to the initial cost, the optimal solution will allow for high failure rates, leading to accept the use of low-reliable structures. The extrapolation to rich and poor countries or cities, leads also to low reliability levels in poor countries (or cities) because of the lower failure costs, as human lives and constructions have statistically lower monetary values. One can imagine the political consequences of such a strategy. At least theoretically, the correct estimation of the failure cost should lead to coherent results. The problem of cost estimation is even more complicated when talking about the whole lifetime management, because the failure cost may change along the structure lifetime due to socio-economical considerations (e.g. life quality of the society). In all cases, special care is strongly required when minimizing the expected total cost, even when other reliability constraints are considered. Due to difficulties in estimating the failure cost Cf (especially when dealing with human lives and environmental deterioration, political consequences, . . .), the direct use of the above equation is not that easy. For design purpose, an alternative to the expect total cost formulation is usually to minimize the initial cost under a prescribed reliability constraint (Moses 1977): min CI (d) d

subject to Pf (d) ≤ Pft d ≤d≤d L

(17) U

where dL and dU are respectively the lower and upper bounds of the design variables and Pft is the admissible failure probability, which is set on the basis of engineering state-of-knowledge and experience. An equivalent formulation is defined in terms of the target reliability index βt : min C(d) d

subject to

β(d) ≥ βt d ≤d≤d L

(18) U

This formulation has the advantage of avoiding the failure cost computation. Nevertheless, the failure consequences can be indirectly included by selecting suitable target safety levels. In civil engineering, it is common to use an admissible failure probabilities of 10−4 for the ultimate limit state and of 10−2 for the serviceability limit state. More refined target values are given in the Eurocodes, in terms of the economical gravity and the number of exposed persons. In principle, the target system reliability should be determined by social and economical considerations. There is no general rule, so far, to select the target value of the system-reliability index. Furthermore, the designer’s experience and preferences still play an important role in the process. A reasonable choice consists in taking the reliability of old design codes as a target for the new codes. Nevertheless, the choice of the target value is still very important in system reliability-based optimization, because it is the regulator of the reliability indexes of the failure modes.

P r i n c i p l e s o f r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

19

The above formulation represents two embedded optimization problems (Enevoldsen and Sørensen 1994; Enevoldsen 1994). The outer one concerns the search for optimal design variables to minimize the cost and the inner one concerns the evaluation of the reliability index in the space of random variables. The coupling between the optimization and reliability problems is a complex task and leads to a very high calculation cost. The major difficulty lies in the evaluation of the structural reliability, which is carried out by a particular optimization procedure. In the random variable space, the reliability analysis implies a large number of mechanical calls, where in the design variable space, the search procedure modifies the structural configuration and hence requires the re-evaluation of the reliability level at each iteration. For this reason, the solution of these two problems (optimization and reliability) requires very large computation resources that seriously reduces the applicability of this approach. This topic will be intensively discussed later on in the chapter by Chateauneuf and Aoues. In general, the RBDO can be formulated according to one of the following forms: –

RBDO1: Minimize the design cost under reliability and structural constraints: min: CI (d) d

subject to: β(d) ≥ βt and: gj (d) ≤ 0

min: CI (d) or

d

subject to: Pf (d) ≤ Pft and: gj (d) ≤ 0

where βt is the target reliability index and Pft is the maximum allowable failure probability. When first order approximation is applied, the relationship between these two forms is given by: Pf =
(−β) or β = −
−1 (Pf ). –

RBDO2: Maximize the reliability under cost and structural constraints: max: β(d)

min: Pf (d)

d

subject to: CI (d) ≤ CIt and: gj (d) ≤ 0 –

d

or

subject to: CI (d) ≤ CIt and: gj (d) ≤ 0

RBDO3: Maximize the reliability per unit cost under structural constraints: max: β(d)/CI (d) d

subject to: gj (d) ≤ 0

or

max: 1/Pf (d)/CI (d) d

subject to: gj (d) ≤ 0

which is equivalent to minimize the ratio cost/reliability: min: CI (d)/β(d) d

subject to: gj (d) ≤ 0

or

min: CI (d) · Pf (d) d

subject to: gj (d) ≤ 0

This kind of formulation is particularly useful when there is no limitation on the total cost in RBDO2.

20

Structural design optimization considering uncertainties

P

h

mR

A

L

Figure 1.10 Perforated beam subjected to uniform load.

–

RBDO4: Minimize the total expected cost under reliability and structural constraints: min: CI (d) + Cf Pf (d) d

subject to: β(d) ≥ βt and: gj (d) ≤ 0

min: CI (d) + Cf Pf (d) or

d

subject to: Pf (d) ≤ Pft and: gj (d) ≤ 0

These formulations are considered as the basic forms of reliability-based design optimization, where the goal is to better redistribute the material within the structure by taking into account the effects of uncertainties and fluctuations. 4.3

Il l ustrati o n o n pe r fo r at e d s imple be a m

A simply supported beam, with length L = 2 m and height h = 0.3 m, is perforated by 5 holes of mean radius mR . The beam is subjected to uniformly distributed load P with mean value 1 MN/m and coefficient of variation of 15%. The maximum stress is located at point A in Figure 1.10. Under the effect of geometrical uncertainties, the nominal hole radius mR has to be designed according to the RBDO basis. In Figure 1.11, the initial, failure and total costs are plotted in function of the mean hole radius. The minimum cost corresponds to mR = 7.5 cm, and to the failure probability of 1.07 × 10−4 . Figure 1.12 shows the expected total cost for different values of consequence severity. It is observed that the hole radius should be decreased with higher consequence costs, in order to reduce the probability of failure and therefore the risk. The optimal solutions are found with respect to each failure cost case: Low: mR = 7.9 cm (Pf = 3.4 × 10−3 ), Moderate: mR = 7.5 cm (Pf = 1.1 × 10−4 ), High: mR = 7.1 cm (Pf = 4.3 × 10−6 ) and Very High: mR = 6.7 cm (Pf = 3.7 × 10−7 ). It can be observed that the failure probability levels are very sensitive to the failure consequences, showing that special care should be considered in estimating these consequences, as they changes drastically the optimal solution.

5 Multi-component RBDO In practical structural systems, the overall failure is generally dependent on a certain number of components where each one may have several failure modes, arranged in

P r i n c i p l e s o f r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

21

Expected costs (Euros)

4000 3500

Initial cost Failure cost Total cost

3000 2500 2000 1500 1000 500 0 6,500

7,000 7,500 Mean hole radius (cm)

8,000

Figure 1.11 Initial, failure and total costs of the perforated beam.

Expected costs (Euros)

4,00E03 3,50E03 3,00E03 2,50E03 2,00E03 1,50E03 1,00E03 5,00E02 0,00E00 6,500

Low failure cost Moderate failure cost High failure cost Very high failure cost 7,000

7,500

8,000

Mean hole radius (cm)

Figure 1.12 Expected total costs in function of the failure consequence costs.

series and/or parallel systems. During the optimization of redundant structures, the contribution of various members is highly redistributed and the prediction of the most important components is not easy. Some insignificant components at the beginning of the RBDO procedure can become very important in the neighborhood of the optimal point. That is why structural reliability cannot be correctly computed without the whole system consideration, by taking account for all the failure modes. In this case, the constraint on system reliability becomes a computational challenge because of the different levels of embedded optimization loops. So, the system RBDO has common limitations due to system reliability computation and the necessity to make some approximations in practical cases (e.g. bounds, reduction of failure paths, . . .). This is probably the main reason why the system approach is less popular than the component approach. Another difficulty arises from the fact that the component assembly is rather a logical combination (i.e. union and intersection of events) than just algebraic operation, which is hard to deal with in system optimization, as sensitivity computation is not easy for

22

Structural design optimization considering uncertainties

logical operators. For example, the derivation of the union of two events is not simple to handle when one of them is totally or partially included inside the other, as the derivative operator can only capture the dominant event sensitivity. This difficulty is emphasized by the fact that failure mode combination is strongly related to the few significant failure modes at a given instance of the computing process. However, as the design variable values change in each iteration, the significant failure modes are not always the same, which greatly influences the convergence of the optimization procedure. Fortunately, in practice the significant failure modes identified in the system reliability analysis tend to be stabilized after few iterations. 5.1 Sy stem RB DO fo r mulat io n The system RBDO can be formulated either at the component level or at the system level (Enevoldsen 1994). At the component level, the RBDO can be written by specifying the target reliability for each one of the structural components, leading to: min CI (d) x

subject to βi (d) ≥ βti

and

gj (d) ≤ 0

(19)

where βi (d) and βti are respectively the reliability index and the target index for the ith component. Each one of the component reliability constraints includes a minimum reliability requirement for a specific failure mode at a specific location in the structure. For example, a member has several critical cross-sections which may fail according to several modes, such as yielding, cracking and excessive deformation, in addition to member buckling failure and structural instability. At the system level, the RBDO is formulated by only specifying the target system reliability for the whole structure: min CI (d) d

subject to βsys (d) ≥ βt

and

gj (d) ≤ 0

(20)

where βsys (d) and βt are respectively the reliability index and the target index for the whole system. The system reliability is generally evaluated by the use of upper and lower bounds. Some authors combined the constraints on component and system reliabilities, but this approach could lead to either redundant or inconsistent constraints. Aoues and Chateauneuf (2007) proposed a scheme for consistent RBDO of structural systems. The basic idea consists in updating the component target safety levels in order to fulfill the overall system target. In the main optimization loop, the cost function is minimized under the constraints that component reliability indexes must satisfy the updated target values. min C(d) subject to

Updated

βj (d) ≥ βtj

d ≤d≤d L

Updated

(21)

U

where βtj is the updated target reliability index for the jth failure mode and βj (d) is the reliability index for the considered design configuration. In the updating procedure,

P r i n c i p l e s o f r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

P

q

23

P

q q/8

d2

d2 d1

Lc
3 m

Lc
3 m

L
8m M1

M2 M2
M1/8

M1

M2

Bending moment diagram

Figure 1.13 Overhanged beam with variable cantilever depth.

the target indexes are adjusted to meet the system reliability requirement. This can be performed by solving the problem: min

Updated j

βt

mp 

Updated

(βtj

− β j )2

(22)

i=1

subjected to

Updated βsys (βtj , ρjk )

≥ βt Updated

which is solved for the updated target indexes βtj 5.2

.

Overhanged reinforced concrete be am

In order to show the interest of system analysis, an overhanged beam with variable cantilever depth is considered, as shown in Figure 1.13 (Aoues and Chateauneuf 2007). With a constant breadth of 20 cm, the beam is defined by the middle-span depth d1 and the cantilever end depth d2 . The span is L = 8 m and the cantilever length is Lc = 3 m. The beam is subjected to uniformly distributed loads q and q/8 as illustrated in Figure 1.13. In order to reduce the negative moments, two tension rods are acting at the cantilever ends, modeled by the tensile force P. The concrete strength is taken as fcu = 25 MPa and the steel yield strength is fY = 200 MPa. An extreme loading case is considered where q = 40 kN/m and P = 30 kN; leading to the maximum moments M(x = 0.75) = 11.25 kNm and M(x = 3) = −90 kNm. The considered random variables are the applied loads and the effective depth of RC cross-sections, which are considered as normally distributed to allow for easy graphical illustrations. For a given cross-section, the design equation is written by:

Gi = fY Asi

f Y As i di − 2(0.85fcu b)

 − Mi

(23)

24

Structural design optimization considering uncertainties Table 1.1 Statistical data for random variables. Random variable

Mean

St-deviation

Middle span depth d1 Cantilever end depth d2 Reference moment M

md1 md2 mM = 90 kNm

σd1 = 5 cm σd2 = 2.5 cm σM = 18 kNm

The reinforcement is chosen as As1 =12 cm2 and As2 = 6 cm2 , leading to the limit states: G1 = 0.24(d1 − 0.02824) − M1 G2 = 0.12(d2 − 0.01412) − M2

(24)

which can be written in the normalized space by probabilistic transformation: H1 = 0.24(md1 + σd1 ud1 − 0.02824) − (mM + σM uM )  2 H2 = 0.12(md2 + ρσd2 ud1 + 1 − ρ σd2 ud2 − 0.01412) −0.125(mM + σM uM )

(25)

where M is a reference moment (equal to M1 ), ui are the normalized variables and ρ is the correlation between d1 and d2 . The distribution parameters are given in Table 1.1. The correlation between d1 and d2 is taken as ρ = −0.6. As this situation is considered as extreme one, the allowable failure probability of the system is set to: Pf _system = 0.05 (naturally, this is a conditional probability as it assumes that extreme situation occurs). The reliability solution leads to the direction cosines: αd1 = 0.55 and αM = −0.83 for the limit state H1 , and αd1 = −0.48, αd2 = 0.64 and αM = −0.60 for the limit state H2 . Thus, the correlation between these two limit states is equal to 0.233. The overall RC volume in this beam is computed by V = 0.2(11 d1 + 3 d2 ). To take account for the workmanship in the cost calculation, the depths are set to the power 3. The final cost of RC is estimated by 150a/m3 . The system RBDO is applied to the structure, by adopting two considerations: 1) the target reliability index is the same for all the limit states, and 2) the target reliability indexes are adapted to find a better solution, under the satisfaction of the system target. In the first case, the target system failure probability of 0.05 is reached when both components have reliability indexes of 1.943, knowing the correlation of 0.233. In the second case, the target of 0.05 is searched, where the cost is to be set as low as possible. The interest of the adaptive target strategy is shown by comparing these two RBDO formulations, as indicated in Table 1.2. For the same system reliability level, the adaptive target methodology allows us to significantly decrease the structural cost, by better distributing the material within the structure. Figure 1.14 compares the failure domains for both solutions (the 2D graph is given for the limit states projected on the plane uM = 0). As the system failure probability is the same for both formulations, the decrease of the margin for H1 implies the increase of the margin for H2 . For the same system reliability, the adaptive approach allows us to reach a cost reduction of 12.4%. Figure 1.14 shows also the beam profile obtained

P r i n c i p l e s o f r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

25

Table 1.2 RBDO formulation and solutions. Considered aspect

Component-based formulation

System-based formulation

Formulation

Minimize: 300(11m3d1 + 3m3d2 ) under: β1 ≥ 1.9434 and: β2 ≥ 1.9434

Minimize: 300(11m3d1 + 3m3d2 ) under: βsys ≥ 1.64485

Failure point U ∗ :

u∗d1 = −1.07; u∗d2 = 0; u∗M = 1.61 ∗ ∗ ud1 = 0.93; ud2 = −1.24; u∗M = 1.17

u∗d1 = −0.91; u∗d2 = 0; u∗M = 1.37 ∗ ∗ ud1 = −1.58; ud2 = −2.11; u∗M = 1.98

Reliability levels at optimum:

β1 = 1.9434 β2 = 1.9434 Pfsys = 0.05

β1 = 1.6487 β2 = 3.2959 Pfsys = 0.05

Optimum design:

m∗d1 = 57.8 cm m∗d2 = 16.9 cm CT = 64.3a

m∗d1 = 55.2 cm m∗d2 = 21.1 cm CT = 56.3a

H1

ud

2

H2 ud

1

Limit states for identical component reliabilities

Limit states for adaptive targets

16.9 cm 21.1 cm 55.2 cm 57.8 cm Design with identical component reliabilities

Design with adaptive targets

Figure 1.14 Failure domains and optimum design for identical and adaptive formulations.

by the two approaches. It is clear that the adaptive target approach tries to decrease the depth where cost is widely involved, without decreasing the overall system safety.

6 RBDO issues The interest of RBDO is not limited to the design of new structures, but it also offers a powerful tool to solve a large class of structural problems. The RBDO is applied to various levels of reliability assessment, design, maintenance and codification. Some of these issues are briefly presented in this section. 6.1

Multicriteria approach for RBDO

As a matter of fact, the RBDO is a multicriteria optimization problem where the objective is to minimize the costs and to maximize the safety (Kuschel and Rackwitz 1997). It is generally acceptable that reliability and economy have conflicting requirements

26

Structural design optimization considering uncertainties

which must be considered simultaneously in the optimization process. The usual formulations aims either to combine these two objectives in only one weighted objective or to deal with one of these objectives as an optimization constraint. A more rational formulation can be stated as real multicriteria problem where the designer can get the Pareto optimal configurations in order to make consistent choices in the design process. As an example, Frangopol (2003) proposed a four-objective vector for bridge structures: f(d, x) = [V(d), PfCOL (d, x), Pf YLD (d, x), Pf DFM (d, x)]

(26)

where V(d) is the material volume, Pf COL (d, x) is the collapse probability, Pf YLD (d, x) is the first yield probability and Pf DFM (d, x) is the excessive deformation probability. This problem can be solved by any general multi-criteria technique. 6.2

C o d e c al i b r at io n b y R B DO

The design codes of practice must fit a certain objective for the whole applicable domain. Many actual design codes derive from a reliability-based calibration procedure to determine the partial safety factors to be applied in design (Sørensen et al. 1994). The objective of these codes is generally to keep the structural reliability above the specified target level (Ang and De Leon 1997). The problem of defining the safety factors is solved by the minimization of a penalization function for all the design situations covered by the design code (Gayton et al. 2004); the optimization problem is thus (Ditlevsen & Madsen 1996): min f (γi ) = γi

L 

W(ωj , βj (γi ), βt )

(27)

j=1

where W( · ) is a penalty function, γi are the partial safety factors, βj (γi ) is the safety index for the j-th situation and βt is the target reliability. Several kinds of penalty function have been proposed in the literature. The simplest one is defined by the weighted least square function: W1 (γi ) = ωj (βj (γi ) − βt )2

(28)

This function has the advantage of being very simple and the solution of the optimization problem (equation 27) can be greatly simplified if βj (γi ) has a simple explicit expression. Nevertheless, this function is symmetrical with respect to βt , i.e. it only depends on the difference βj − βt , and structures with a reliability index smaller than the target are not more penalized than structures with higher reliability index. Another function can take the following form (Lind 1977): W2 (γi ) = ωj (k(βj (γi ) − βt ) + exp(−k(βj (γi ) − βt )) − 1)

(29)

where k > 0 is the curvature parameter. This function penalizes the reliability indexes which are smaller than the target, compared to those higher than the target. When the parameter k increases this function becomes more penalizing for βj < βt than the least square function. For large values of k, the penalty goes to infinity for βj < βt , and so

P r i n c i p l e s o f r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

27

reliability indexes lower than the target become forbidden. Other penalty functions can be proposed on the basis of socio-economic measures of the gap between the code and its objective. In such a case, the relationship between cost and target reliability index must be known. Classically, the goal of the design codes is to minimize the expression (27) for the whole spectrum of the design situations. Nevertheless, new evolutions of some codes of practice tend to homogenize the risk (i.e. product of failure probability by the consequences) instead of the reliability level (or failure probability). As an example, the RBDO calibration could take the form: min CT (d, γ) = d,γ

subject to

L 

L 

CTi (d, γ)

i=1

W(ωj , βj (d, γ), βt ) ≤ ε

(30)

j=1

gj (d) ≤ 0 dL ≤ d ≤ dU where ε is an acceptable tolerance for target fitting. 6.3 Topology bas ed RBDO Knowing that the RBDO concerns mostly shape optimization, the application to topology optimization is a new research field (Kharmanda et al. 2004). The basic idea concerns the use of uncertainties as a control parameter for topology selection. In fact, the reliability constraint allows us to get a robust structural topology. Figure 1.15 illustrates the fact that different topologies can be suitable for the same ground structure. Usually, the comparison in deterministic topology optimization is only related to minimized mean compliance, without observing the solution dispersion. The principle of reliability-based topology robustness consists in defining the topology which is less sensitive to system uncertainties. The main difficulty in dealing with topology lies in the fact that topology optimization is a qualitative approach, while the reliability-based design is a quantitative

Robust topology Compliance

Large dispersion

Optimization procedure iterations

Figure 1.15 RBTO and principle of reliability-based topology robustness.

28

Structural design optimization considering uncertainties

approach. The coupling of the two methods requires special developments to overcome formulation and efficiency problems. 6.4 T i m e-v a ri ant R B DO Every designer knows well that system information are not perfect and their validity is limited under system aging. In fact, most of phenomena involved in the total cost function are time-variant. It can be mentioned, for example, the loading fluctuation over the structural lifetime, the deterioration of material properties with time, the variation of operating and maintenance costs, and the monetary fluctuation of failure costs. All these time-variant phenomena lead to time-variant optimal solution. However, the designer must take decisions in a given stage of the project (largely before the construction or the manufacturing of the system), in function of the available data at that stage. The resulting solution is optimal only in the first part of the structure lifetime, as it does not account for aging and long-term exposure. In time-variant RBDO (Kuschel and Rackwitz 1998), the ideal scheme consists in designing the system for best optimal solution, considering the whole lifetime of the system. In this case, the utility function takes the form: max x

U(p, d, T) = B(d, t) − CI (d) − L(p, d, T)

subject to

gj (d) ≤ 0

(31)

with
T B(d, T) =

b(t) d(t) (1 − Pf (p, d, t))dt 0

(32)


T L(p, d, x, T) =

Cf (p, d) f (p, d, t) d(t)dt 0

where b(t) is the benefit derived from the existence of the system, Cf is the failure cost, f (p, d, t) is the probability density of the time to failure, d(t) is the discount function (or capitalization function) and T is the system age. 6.5

C o upl ed r e liab ilit y-b as e d d es ig n an d m a i n t e n a n ce p l a n n i n g

Although in design practice and due to system complexity, the maintenance planning is often considered as an independent step, the Reliability-based optimization can also be applied to a coupled set of design and maintenance parameters. In this case, the problem is formulated as: min x

CT (d) = CI (d) + CF (d) + CM (d)

subject to gj (d) ≤ 0

(33)

At the design stage, the maintenance cost is minimized by selecting the best set of parameters. At this stage, there is no available site information (as the system is

P r i n c i p l e s o f r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

29

not constructed yet) and a priori hypotheses have to be formulated. Generally, regular maintenance intervals are chosen at this stage. The maintenance cost is usually a function of the type of inspection method mS , the number of inspections in the remaining lifetime nS , and the time for different inspections t. The maintenance cost can be described by (Enevoldsen and Sørensen 1994): CM (d, p) = CPM (d, p) + CINS (d, p) + CREP (d, p)

(34)

where CM is the expected maintenance cost, CPM is the preventive maintenance cost, CINS is the expected inspection cost, CREP is the expected cost of repairs, and p is the vector of maintenance parameters. Enevoldsen and Sørensen (1994) suggested to use the following expressions to evaluate inspection and repair costs: CINS (d, mS , nS , t) =

nS  i=1

CREP (d, nS , t) =

nS  i=1

CSi (mS )(1 − Pf (d, ti ))

1 (1 + r)ti

1 CRi (x)PRi (d, ti ) (1 + r)ti

(35)

where CSi is the ith inspection cost, Pf is the failure probability in the time interval [0, ti ], r is the discount rate, CRi is the cost of a repair at the ith inspection and PRi is the probability of performing a repair after the ith inspection for surviving components.

7 Conclusions RBDO is a powerful tool for robust design of structural systems. The explicit consideration of safety level allows us to optimize the total cost where the solution becomes less sensitive to system uncertainties. Contrary to traditional deterministic design optimization, the RBDO allows us to modulate the safety margins in function of the uncertainty effects for each variable, in order to reach economic, safe, efficient and robust design. In this sense, the safety factors are optimally defined within the system, compared to deterministic design where the safety factors are set before undergoing the optimization process. RBDO is still an active research field in order to extend the possibilities for new applications. Design, topology and time-variant reliability-based optimizations are very interesting field to reach performance-based design for cost-effective, durability and lifetime management of structural systems.

References Ang, A.H.-S. & De Leon, D. 1997. Determination of optimal target reliabilities for design and upgrading of structures. Structural Safety 19:91–103. Aoues, Y. & Chateauneuf, A. 2007. Reliability-based optimization of structural systems by adaptive target safety application to RC frames. Structural Safety. Article in Press. Ditlevsen, O. 1979. Narrow reliability bounds of structural systems. Journal of Structural Mechanics 7:435–451. Ditlevsen, O. & Madsen, H. 1996. Structural Reliability Methods. John Wiley & Sons.

30

Structural design optimization considering uncertainties

Enevoldsen, I. 1994. Reliability-based optimization as an information tool. Mech. Struct. & Mach. 22:117–135. Enevoldsen, I. & Sørensen, J.D. 1994. Reliability-based optimization in structural engineering. Structural Safety 15:169–196. Frangopol, D.M. 1995. Reliability-based optimum structural design. In: Probabilistic structural mechanics handbook, edited by C. Raj Sundararajan, Chapman Hall, USA, 352–387. Frangopol, D.M. 1999. Life-cycle cost analysis for bridges. In: Bridge safety and reliability. ASCE, Reston, Virginia, 210–236. Frangopol, D.M. 2000. Advances in life-cycle reliability-based technology for design an maintenance of structural systems. In: Computational mechanics for the twenty-first century. Edinburgh: Saxe-Coburg Publishers, 257–270. Frangopol, D.M. & Maute K. 2003. Life-cycle reliability-based optimization of civil and aerospace structures. Computers & Structures 81:397–410. Gayton, N., Mohamed-Chateauneuf, A., Sørensen, J.D., Pendola, M. & Lemaire, M. 2004. Calibration methods for reliability-based design codes. Structural Safety 26(1):91–121. Hasofer, A.M. & Lind, N.C. 1974. An Exact and Invariant First Order Reliability Format. J. Eng. Mech., ASCE, 100, EM1:11–121. Kharmanda, G., Mohamed-Chateauneuf, A. & Lemaire, M. 2002. Efficient reliability-based design optimization using a hybrid space with application to finite element analysis. Journal of Structural and Multidisciplinary Optimization 24(3):233–245. Kharmanda, G., Mohamed-Chateauneuf, A. & Lemaire, M. 2002. CAROD: Computer-Aided Reliable and Optimal Design as a concurrent system for real structures. Journal of Computer Aided Design and Computer Aided Manufacturing CAD/CAM 1(1):1–12. Kharmanda, G., Olhoff, N., Mohamed-Chateauneuf, A. & Lemaire, M. 2004. Reliability-based topology optimization. Struct. Multidisc. Optim. 26:295–307. Kuschel, N. & Rackwitz, R. 1997. Two basic problems in reliability-based structural optimization. Mathematical Methods of Operations Research 46:309–333. Kuschel, N. & Rackwitz, R. 1998. Structural optimization under time-variant reliability constraints. Proceeding of the eighth IFIP WG 7.5 Working conference on Reliability and Optimization of Structural Systems, edited by Nowak, University of Michigan, Ann Arbor, Michigan, USA, 27–38. Kuschel, N. & Rackwitz, R. 2000. A new approach for structural optimization of series system. In: R.E. Melchers & M.G. Stewart (eds). Proceedings of the 8th International conference on applications of statistics and probability (ICASP) in Civil engineering reliability and risk analysis, Sydney, Australia, December 1999, Vol. 2. pp. 987–994. Lemaire, M., in collaboration with Chateauneuf, A. & Mitteau, J.C. 2006. Structural reliability. ISTE, UK. Lind, N.C. 1977. Reliability based structural codes, practical calibration. Safety of structures under dynamic loading, Trondheim, Norway, 149–160. Madsen, H.O. & Friis Hansen, P. 1991. Comparison of some algorithms for reliabilitybased structural optimization and sensitivity analysis. In: C.A. Brebbia & S.A. Orszag (eds): Reliability and Optimization of Structural Systems, Springer-Verlag, Germany, 443–451. Moses, F. 1977. Structural System Reliability and Optimization. Comput. Struct. 7:283–290. Moses, F. 1997. Problems and prospects of reliability based optimization. Engineering Structures 19(4):293–301. Rackwitz, R. 2001. Reliability analysis, overview and some perspectives. Structural Safety 23:366–395. Sørensen, J.D., Kroon, I.B. & Faber, M.H. 1994. Optimal reliability-based code calibration. Structural Safety 15:197–208.

Chapter 2

Reliability-based optimization of engineering structures John D. Sørensen Aalborg University, Aalborg, Denmark

ABSTRACT: The theoretical basis for reliability-based structural optimization within the framework of Bayesian statistical decision theory is briefly described. Reliability-based cost benefit problems are formulated and exemplified with structural optimization. The basic reliability-based optimization problems are generalized to the following extensions: interactive optimization, inspection and repair costs, systematic reconstruction, re-assessment of existing structures. Illustrative examples are presented including a simple introductory example, a decision problem related to bridge re-assessment and a reliability-based decision problem for offshore wind turbines.

1 Introduction The theoretical basis for reliability-based structural optimization can be formulated within the framework of Bayesian statistical decision theory mainly developed and described in the period 1940–60, see for example (Raiffa & Schlaifer 1961), (Aitchison & Dunsmore 1975), (Benjamin & Cornell 1970) and (Ang & Tang 1975). By statistical decision theory it is possible to solve a large number of decision problems where some of the parameters are modeled as uncertain. The uncertain parameters are modeled by stochastic variables or stochastic processes. Uncertain costs and benefits can thus be accounted for in a rational way. A large number of “simple’’ examples for application of statistical decision theory within structural and civil engineering are given in e.g. (Benjamin & Cornell 1070), (Rosenbleuth & Mendoza 1971) and (Ang & Tang 1975). During the last decades significant achievements have been obtained in development of efficient numerical techniques which can be used in solving problems formulated by statistical decision theory. Especially the development of FORM (First Order Reliability Methods), SORM (Second Order Reliability Methods) and simulation techniques to evaluate the reliability of components and systems has been important, see e.g. (Madsen et al. 1986). In the same period efficient methods to solve non-linear optimization problems have also been developed, e.g. the sequential quadratic optimization algorithms (Schittkowski 1986) and (Powell 1982). These developments have made it possible to solve problems formulated in a decision theoretical framework. Examples are: •

Reliability-based inspection and repair planning for offshore structures and concrete structures, formulated as a preposterior decision problem, see e.g. (Kroon 1994), (Engelund 1997), (Skjong 1985), (Thoft-Christensen & Sørensen 1987),

32

•

Structural design optimization considering uncertainties

(Fujita et al. 1989), (Madsen et al. 1989), (Madsen & Sørensen 1990), (Fujimoto et al. 1989), (Sørensen & Thoft-Christensen 1988) and (Faber et al. 2000). Reliability-based structural optimization problems and associated techniques for sensitivity analysis and numerical solution. Basic formulations of reliability-based structural optimization are given in e.g. (Murotsu et al. 1984), (Frangopol 1985), (Sørensen & Thoft-Christensen 1985) and (Enevoldsen & Sørensen 1994). System aspects are considered in e.g. (Enevoldsen & Sørensen 1993), interactive reliabilitybased optimization in (Sørensen et al. 1995) and optimization with time-variant reliability in e.g. (Kuschel & Rackwitz 1998). Further it is noted that a one-level approach for reliability-based optimization is described in (Streicher & Rackwitz 2002) based on an idea in (Madsen & Hansen 1992).

In section 2 a short description of Bayesian decision theory for engineering decisions is given and in section 3 reliability-based structural optimization problems are formulated. Only time-invariant reliability problems are considered. Three levels of decision problems with increasing degree of complexity can be identified: (1) decisions with given information (e.g. for new structures), (2) decisions with given new information (e.g. for existing structures), (3) decisions involving planning of experiments/inspections to obtain new information (e.g. for inspection planning). Further, interactive optimization aspects are discussed. In order to solve reliability-based optimization problems it is important to have accurate and numerically effective methods to evaluate probabilities of different events and of expectations. In section 4 some probabilistic methods, such as FORM/SORM, are briefly mentioned. Also techniques are described for sensitivity analyses to be used in numerical solution of the optimization problems using general optimization algorithms. In section 5 illustrative examples are presented, including applications with re-assessment of a concrete bridge and with reliability-based design of support structure for wind turbines.

2 Decision theory for engineering decisions Engineers are often in the situation to take decisions on design of a new structure, on repair/maintenance of existing structures where statistical information is available. In the following it is shown how Bayesian statistical decision theory can be used for making such decisions in a rational way, see (Raiffa & Schlaifer 1961) and (Benjamin & Cornell 1970) for a detailed description. An important difficulty in Bayesian statistical decision theory when applied in civil and structural engineering is that it can be difficult to assign values to cost of failure, or not acceptable behavior, especially when loss of human lives is involved. One solution is to calibrate the cost models to existing structures or to base the decisions on comparisons with alternative solutions. Further, organizational factors can have a rather significant influence in the decision process. These factors often have an influence, which is not rational from a cost-benefit point of view. Examples are the influence of the organizational structure, personal preferences and organizational culture. The first problem to consider is that of making rational decisions when some of the parameters defining the model are uncertain, but a statistical description of the

R e l i a b i l i t y-b a s e d o p t i m i z a t i o n o f e n g i n e e r i n g s t r u c t u r e s

33

Cost C(z, X) Design decision z

State of nature X

Figure 2.1 Decisions with given information.

parameters is available, i.e. the statistical information is given. The uncertain parameters are modeled by n stochastic variables X = (X1 , X2 , . . . , Xn ). The density function of the stochastic variables is fX (x, θ) where θ are statistical parameters, for example mean values, standard deviations and correlation coefficients. Further, it is assumed that a decision has to be taken between a number of alternatives which can be modeled by design/decision variables z = (z1 , z2 , . . . , zN ). In figure 2.1 a decision model with one discretized variable z is shown. The decision is taken before the realization by nature of the stochastic variables is known. Besides the decision variables z and the uncertain variables X also a cost function C(z, X) is introduced in the decision model in figure 2.1. When a decision z has been taken and a realization x of the stochastic variables appears then the value obtained is denoted C(z, X) and represents a numerical measure of the consequences of the decision and the realization obtained. C(z, X) is assumed to be related to money and represents in general costs minus benefits, if relevant. As an example the design parameters z could be the geometrical parameters of a structural system (cross-sectional dimensions and topology), the stochastic variables X could be loads and material strengths and objective function C could be the cost of the structure. In some decision problems it can be difficult to specify the cost function, especially if the consequences not directly measurable in money are involved, for example personal preferences. However, as described in von (von Neumann & Morgenstern 1943) rational decisions can be taken if the cost function is made such that the expected value of the cost function is consistent with the personal preferences. Thus, if the decisionmaker wants to act rationally the strategy z, which minimizes the expected cost, has to be chosen as
C ∗ = min EX [C(z, X)] = C(z, X)fX (x) dx (1) z

EX [−] is the expectation with respect to the joint density function of the stochastic variables X is the minimum cost corresponding to the optimal decision z∗ . The optimization problem can be generalized to include benefits B(z) such that the total expected benefits minus costs, Z are maximized. (1) is then written
(2) Z∗ = max Z(z) = B(z) − EX [C(z, X)] = B(z) − C(z, X)fX (x) dx z

where it is assumed that the benefits are not dependent on the stochastic variables X.

34

Structural design optimization considering uncertainties

3 Reliability-based structural optimization The formulations given above can be used in a number of cases related to design of structures. As mentioned in section 2 they can e.g. be used in a design situation where z models the design variables (size and shape variables in a structural system), X models uncertain loads and material parameters, B models the benefits and C models the total expected costs to design and possible failure. As mentioned only time-invariant reliability problems are considered.

3.1 Ba si c rel i ab ilit y-b as e d o pt imizat io n f o r m u l a t i o n s First, it is assumed that • •

There is no systematic reconstruction of the structure in case of failure Discounting can be ignored The total expected cost-benefits can then be written Z(z) = B(z) − C(z) = B(z) − CI (z) − Cf PF (z)

(3)

where CI (z) and Cf model the costs due to construction and failure, B(z) models the benefits and PF (z) is the probability of failure. Failure/no failure should here be considered in a general sense as satisfactory/not satisfactory behavior. The optimal design z∗ is obtained from the optimization problem: max Z(z) = max {B(z) − CI (z) − Cf PF (z)} z

z

(4)

(4) can equivalently be formulated as a reliability-constrained optimization problem max B(z) − CI (z) z

(5)

subject to β(z) ≥ βmin where the generalized reliability index is defined by β(z) = −
−1 (PF (z))

(6)


is the standard normal distribution function. βmin can be a code specified minimum acceptable reliability level related to annual or lifetime reference time intervals. Other design constraints can be added to (5) if needed. (4) and (5) give the same optimal decision if βmin is chosen as the reliability level corresponding to the optimal solution z∗ of (4): βmin = β(z∗ ), i.e. there is a close connection between βmin and Cf /CI . This can easily be seen considering the Kuhn-Tucker optimality conditions for (4) and (5). (5) is a two-level optimization problem, sine the calculation of the reliability index β by FORM requires an optimization problem to be solved, see section 4.

R e l i a b i l i t y-b a s e d o p t i m i z a t i o n o f e n g i n e e r i n g s t r u c t u r e s

35

The optimization problem in (5) can be generalised to the following element reliability-based structural optimization problem: m  mP D   max Z(z) = B(z) − C(z) = B(z) − CIi Vi (z) + Cfi
(−βi (z)) z

subject to

i=1

βi (z) ≥ BI,i (z) ≥ 0, BE,i (z) = 0, z1i ≤ zi ≤ ziu , βimin ,

i=1

i = 1, . . . , M i = 1, . . . , mI i = 1, . . . , mE i = 1, . . . , N

(7)

where z = (z1 , . . . , zN ) are the design (or optimization) variables. The optimization variables are assumed to be related to parameters defining the geometry of the structure (for example diameter and thickness of tubular elements) and coordinates (or related parameters) defining the geometry (shape) of the structural system. The objective function C consists of a deterministic and a probabilistic part with mD and mP terms, respectively. Vi is e.g. a volume in the ith deterministic term and Vi is the cost per volume of the ith term modelling the construction costs. Vi is assumed to be deterministic. If stochastic variables influence Vi then design values, see below, are assumed to be used to calculate Vi . In the probabilistic part Cfi is the cost due to failure of failure mode i. βi , i = 1, . . . , mP are reliability indices for the mP failure modes. The general formulation of (7) allows the objective function to model both the structural weight and the total expected costs of construction and failure. The constraints in (7) are based on the reliability indices βi , i = 1, . . . , M for M failure modes. βimin , i = 1, . . . , M are the corresponding lower limits on the reliabilities. BI,i , i = 1, . . . , mI and BE,i , i = 1, . . . , mE define the deterministic inequality and equality constraints in (7) which can ensure that response characteristics such as displacements and stresses do not exceed codified critical values. Determination of the inequality constraints usually includes finite element analyses of the structural system. The inequality constraints can also include general design requirements for the design variables. Finally also simple bounds are included as constraints. The variables (parameters) used to model the structure are characterized as stochastic or deterministic if the variable can be modelled as stochastic or deterministic and design or fixed if the variable is a design (optimization) variable or a fixed constant. The optimization problem in (5) can further be generalised to the following system reliability-based structural optimization problem: m  mP D  CIi Vi (z) + Cfi
(−βi (z)) max Z(z) = B(z) − C(z) = B(z) − z

subject to

βS (z) ≥ βmin , BI,i (z) ≥ 0, BE,i (z) = 0, z1i ≤ zi ≤ ziu ,

i=1

i = 1, . . . , mI i = 1, . . . , mE i = 1, . . . , N

i=1

(8)

where βS is the system reliability index. If failure of the structure can be modelled as by a series/parallel system then βS can be obtained from: βS (z) = −
−1 (Pf (z))

(9)

36

Structural design optimization considering uncertainties

where Pf (z) is the probability of failure of the system, e.g. obtained by FORM/SORM techniques. 3.2

Intera c ti v e o pt imizat io n

In practical solution of an optimization problem it will often be very relevant to be able to make different types of interaction between the user and the numerical formulation/ solution of the design problem. The basic types of interactive optimization which influences the formulation of the optimization problems are, see (Haftka & Kamat 1985) and (Sørensen et al. 1995): • • • •

include (delete) a design (optimization) variable include (delete) a constraint modify a constraint or modify (change) the objective function.

In order to investigate the effect of interactive optimization on the optimality criteria, (9) is restated as the following general optimization problem: min C(z) z

subject to

ci (z) = 0, ci (z) = 0,

i = 1, . . . , mE i = mE + 1, . . . , m

(10)

First order necessary conditions that have to be satisfied at a (local) optimum point z∗ are given by the Kuhn-Tucker conditions. If the optimization process has almost converged, a good guess on the optimal design is available. A modification of the optimization problem is then specified by the user. In (Sørensen et al. 1995) the details are described. Figure 2.2 illustrates the data flow in interactive structural optimization. The modules used are: • • • • • 3.3

User interface OPT: general optimization algorithm REL: module for reliability analysis, e.g. FORM, incl. optimization FEA: finite element program module DSA: module for calculating design sensitivity coefficients. General i z a t io n: inc lud e ins pec t io n a n d r e p a i r co s t s

The basic decision problems considered in section 2 can as mentioned be generalized to be used in reliability-based experiment and inspection planning, see figure 2.3. If e model the inspection times and qualities, and d models the repair decision given uncertain inspection result S, the optimization problem can be written: 0 (e) + CR0 (e, d)PR (e, d) + Cf0 PF (e, d)} max Z(e, d) = B0 − {CIN e,d

(11)

0 where B0 models the benefits, CIN models the inspection costs, CR0 models the repair costs, PR is the probability of repair and PF is the probability of failure, both obtained using stochastic models for S and X.

R e l i a b i l i t y-b a s e d o p t i m i z a t i o n o f e n g i n e e r i n g s t r u c t u r e s

37

CARBOS: User interface

OPT

DSA

Modifications ? N Final design

Y

Reliability analysis

FEA

Sensitivity analysis

DSA

FEA

Interactive optimization Optimization (probabilistic) Optimization (deterministic) Sensitivity analysis

REL

CARBOS: Modify variables, constraints and obj. function

Figure 2.2 Data flow interactive optimization, from (Sørensen et al. 1995).

Inspection plan e

Inspection result S

Repair decision d

State of nature X

Cost-benefit Z(e, S, d, X)

Figure 2.3 Decisions for with given information.

(11) can be further generalised if the total expected costs are divided into construction, inspection, repair and failure costs and a constraint related to a maximum annual (or accumulated) failure probability PFmax is added. If the inspections performed at times T1 , T2 , . . . , TN are part of e the optimization problem can be written max Z(e, d) = B(e, d) − {CI (e, d) + CIN (e, d) + CR (e, d) + CF (z, e)} e,d

subject to

ei1 ≤ ei ≤ eiu ,

i = 1, . . . , N

Pt (t, e, d) ≤ PFmax ,

t = 1, 2, . . . , TL

(12)

where, B is the expected benefits, CI is the initial costs, CIN is the expected inspection costs, CR is the expected costs of repair and CF is the expected failure costs. The annual probability of failure in year t is PF,t . The N inspections are assumed performed at times 0 ≤ T1 ≤ T2 ≤ · · · ≤ TN ≤ TL .

38

Structural design optimization considering uncertainties

The total capitalised benefits are written B(e, d) =

N 

Bi (1 − PF (Ti ))

i=1

1 (1 + r)Ti

(13)

The ith term represents the capitalized benefits in year i given that failure has not occurred earlier, Bi is the benefits in year i, PF (Ti ) is the probability of failure in the time interval [0, Ti ] and r is the real rate of interest. The total capitalised expected inspection costs are written CIN (e, d) =

N 

CIN,i (e)(1 − PF (Ti ))

i=1

1 (1 + r)Ti

(14)

The ith term represents the capitalized inspection costs at the ith inspection when failure has not occurred earlier, CIN,i is the inspection cost of the ith inspection, PF (Ti ) is the probability of failure in the time interval [0, Ti ] and r is the real rate of interest. The total capitalised expected repair costs are CR (e, d) =

N 

CR,i PRi (e, d)

i=1

1 (1 + r)Ti

(15)

CR,i is the cost of a repair at the ith inspection and PRi is the probability of performing a repair after the ith inspection when failure has not occurred earlier and no earlier repair has been performed. The total capitalised expected costs due to failure are estimated from CF (e, d) =

TL 

CF (t) PF,t PCOL|FAT

t=1

1 (1 + r)t

(16)

where CF (t) is the cost of failure at the time t. PCOL|FAT is the conditional probability of collapse of the structure given failure of the considered component. 3.4

G en eral i z a t io n: inc lud e s y s t emat ic r e co n s t r u ct i o n

The following assumptions are made: (1) the structure is assumed to be systematically rebuild in case of failure, (2) only initial costs, CI (z) and direct failure costs, CF are included, (3) the benefits per year are b and (4) failure events are assumed to be modeled by a Poisson process with rate λ. The probability of failure is PF (z). The optimal design is determined from the following optimization problem, see e.g. (Rackwitz 2001): 

CI (z) CF λPF (z) b Ci (z) − − + max Z(z) = z rC0 C0 C0 C0 r + λPF (z) (17) subject to zl ≤ z ≤ zu , i = 1, . . . , N i

i

PF (z) ≤

i

PFmax

R e l i a b i l i t y-b a s e d o p t i m i z a t i o n o f e n g i n e e r i n g s t r u c t u r e s

39

where zl and zu are lower and upper bounds on the design variables. PFmax is the maximum acceptable probability of failure e.g. with a reference time of one year. This type of constraint is typically required by regulators. The optimal design z∗ is determined by solution of (17). If the constraint on the maximum acceptable probability of failure is omitted, then the corresponding value PF (z∗ ) can be considered as the optimal probability of failure related to the failure event and the actual cost-benefit ratios used. The failure rate λ and probability of failure can be estimated for the considered failure event, if a limit state equation, g(X1 , . . . , Xn , z) and a stochastic model for the stochastic variables, (X1 , . . . , Xn ) are established. If more than one failure event is critical, then a series-parallel system model of the relevant failure modes can be used. 3.5

Generalisation: optimal re-as s es s me nt o f e xi s ti ng s truc ture s

In re-assessment of structures and engineering systems, engineers are often in the situation to be involved in decisions on repair and/or strengthening of an existing system/structure where some statistical information is available. In the following it is shown how Bayesian statistical decision theory can be used for making such decisions in a rational way. The theoretical basis is detailed described in e.g. (Raiffa & Schlaifer 1961) and (Benjamin & Cornell 1970). It is assumed that the decision is taken on behalf of the owner of the structure, and that a cost-benefit approach is used with constraints related to minimum safety requirements specified by national/international codes of practice and/or the society. The same principles can be applied in case of other decision makers. It is noted that the optimal solution from the cost-benefit problem should be used as one input to the decision process. The decision problem on possible repair and/or strengthening in a re-assessment situation is illustrated in figure 2.4. It is assumed that the design variables in the initial design situation are denoted z. After the initial design information about the uncertain variables influencing the behaviour of the structure is collected, and are denoted S. Often this information will be collected in connection with the re-assessment. The decision variables at the time TR of re-assessment are denoted d. The uncertain variables describing the state of nature are denoted X.

Time TR

Design decision z

Information S

Repair/re-design decision d

State of nature X

Cost-benefit Z(z, S, d, X)

Figure 2.4 Decisions in re-assessment with given information. The vertical line illustrates the time of re-assessment.

40

Structural design optimization considering uncertainties

The decision is taken before the realization by nature of the stochastic variables is known. Besides the decision variables d and the uncertain variables X also a costbenefit function Z(z, S, d, X) is introduced in the decision model. When a decision d in the re-assessment problem has been taken and a realisation x of the stochastic variables appears then the value obtained is denoted Z(z, S, d, x) and represents a numerical measure of the consequences of the re-assessment decision and the realisation obtained. Z(z, S, d, x) is assumed to be measured in monetary units and represents in general costs minus benefits, if relevant. Illustrative examples of the decision variables z and d, and the stochastic variables S and X are: • • • •

z: design parameters, e.g. geometrical parameters of a structural system (crosssectional dimensions and topology). The design parameters are already chosen at the initial design, and are therefore fixed at the time of re-assessment. S: information collected, e.g. concrete compression strengths obtained from samples taken from the structure, measured wave heights, non-failure of the structure, no-find of defects by an inspection. d: design parameters in the re-assessment, e.g. geometrical parameters of a repair (cross-sectional dimensions and topology). X : stochastic variables, representing e.g. loads and material strengths.

In some decision problems it can be difficult to specify the cost function, especially if the consequences not directly measurable in money are involved, for example personal preferences. However, as described in (von Neumann, J. and Morgenstern 1943) rational decisions can be taken if the cost function is made such that the expected value of the cost function is consistent with the personal preferences. If the information S is related the stochastic variables X then a predictive density function (updated density function) fX

(x|s) of the stochastic variables X taking into account a realization s can be obtained using Bayesian statistical theory, see (Lindley 1976) and (Aitchison & Dunsmore 1975). If the decision-maker wants to act rationally, taking into account the information s the strategy d, which maximizes the expected cost-benefits, has to be chosen from Z∗ = max E

X|s [Z(z, s, d, X)] d

(18)

E

X|s [−] is the expectation with respect to the predictive (updated) density function fX

(x|s). In the following the initial design variables z are not written explicitly. Z∗ is the maximum cost-benefit corresponding to the optimal decision. If the benefits are not dependent on the stochastic variables then the optimization problem can be written: Z∗ = max Z(d) = max {B(d) − E

X|s [C(s, d, X)]} d

d

(19)

where the future benefits are denoted B and the future costs are denoted C. Both benefits and costs should be discounted to the time of the re-assessment. The optimization formulation can also be generalised to include decision variables related to experiment planning.

R e l i a b i l i t y-b a s e d o p t i m i z a t i o n o f e n g i n e e r i n g s t r u c t u r e s

41

In the following time-invariant reliability problems are considered. It is assumed that there is no systematic reconstruction of the structure in case of failure and discounting can be ignored. The total expected cost-benefits can then be written Z(d) = B(d) − C(d) = B(d) − CS (d) − Cf Pf

(d)

(20)

where CS (d) and Cf models the costs due to repair/strengthening after the re-assessment and due to failure, B(d) models the benefits and Pf

(d) is the probability of failure updated with the information s. Failure/no failure should here be considered in a general sense as satisfactory/not satisfactory behaviour. In the case the information S models (one or more) events modelled by an event margin {h(d, X) ≤ 0}, and failure is modelled by a limit state function g(d, X), the updated probability of failure is obtained from: Pf

(d) = P(g(d, X) ≤ 0|h(d, X) ≤ 0)

(21)

In the case the information S is related to the measurements of the stochastic variables X then the (updated) density function fX

(x|s) is used. The optimal design d∗ is obtained from the optimization problem max Z(d) = max {B(d) − CS (d) − Cf Pf

(d)} d

d

(22)

(22) can equivalently be formulated as a reliability-constrained optimization problem max B(d) − CS (d), d

subject to

β

(d) ≥ βmin

(23)

where the generalised reliability index is defined by β

(d) = −
−1 (Pf

(d)). βmin is a code specified minimum acceptable reliability level related to annual or lifetime reference time intervals. Other design constraints can be added to (23) if needed. (22) and (23) give the same optimal decision if βmin is chosen as the reliability level corresponding to the optimal solution d∗ of (22): βmin = β

(d∗ ), i.e. there is a close connection between βmin and Cf /CS . This can easily be seen considering the Kuhn-Tucker optimality conditions for (22) and (23). The basic decision problems considered above can be generalized to be used in reliability-based experiment and inspection planning as described in section 3.3. 3.6

Numeric al s olution of decis ion pro bl e ms

Numerical solution of the decision problems requires solution of one or more optimization problems. Since the optimization problems formulated are generally continuous with continuous derivatives sequential quadratic optimization algorithms such as (Schittkowski 1986) and (Powell 1982) can be expected to be the most effective, see (Gill et al. 1981). These algorithms require that values of the objective function and the constraints be evaluated together with gradients with respect to the decision variables. The probabilities in the optimization problems can be solved using FORM techniques, see (Madsen et al. 1986). Associated with the FORM estimates of the

42

Structural design optimization considering uncertainties

probabilities also sensitivities with respect to parameters are obtained. If the decision problem includes analysis of a structural system the finite element method in combination with sensitivity analyses can be used. The sensitivity analyses can be based on the direct or adjoint load method in combination with the discrete quasi-analytical method or with the continuum method.

4 Reliability analysis and sensitivity analysis As mentioned in the previous section the evaluation of the probability of failure events is an integral part of decision analysis and reliability-based structural optimization problems. Further, the decision analysis involves the evaluation of expected values of the costs. Both the relevant failure probabilities and expected values can be determined using modern reliability analysis techniques. If all variables in the reliability problem can be modelled as time-invariant random variables, the failure probability, PF (z), for a given limit state equation, g(x, z) can be evaluated as
PF (z) = P(g(X, z) ≤ 0) = fX (x, z) dx (24) g(x,z)≤0

where fX (x, z) is the joint density function of the stochastic variables X. The integral in (24) plays a central role in the reliability analysis and has therefore been devoted special attention over the last decades. As the integral in general has no analytical solution it is easily realised that its solution or numerical approximation becomes a major task for integral dimension larger than say 6 and for small probabilities. Sufficiently accurate approximations have been developed which are based on asymptotic integral expansions. These FORM/SORM methods are standard in reliability analysis and commercial software, see e.g. (Madsen et al. 1986). Also simulation methods can in many cases be very effective alternatives to FORM/SORM methods. By FORM analysis the failure surface is approximated by its tangent at the design point. On the basis of the linearised failure surface the probability of failure can be approximated by, see (9): PF (z) ≈
(−β(z))

(25)

Most optimization algorithms for solution of the reliability-based optimization problems formulated in section 3 require that the sensitivities with respect to objective functions and reliability estimates can be determined efficiently. By a FORM analysis these derivatives can be computed numerically by the finite difference method. However, it is more efficient to use a semi-analytical expression. For an element analysis the derivative of the first order reliability index, β, with respect to a parameter p, which may be a design variable z, is ∂β ∂ g(u∗ ; p) 1  = ∇u g(u∗ ; p) ∂p ∂p

(26)

If a gradient-based algorithm is used in order to locate the design point the gradient vector ∇u g(u∗ ; p) is already available and it is only necessary to determine the derivative

R e l i a b i l i t y-b a s e d o p t i m i z a t i o n o f e n g i n e e r i n g s t r u c t u r e s

43

of the failure function with respect to the parameter p. The derivative of the first order estimate of the probability of failure with respect to p is ∂Pf ∂β = −ϕ(−β) ∂p ∂p

(27)

where ϕ denotes the density function of a standard normally distributed variable. Also for series and parallel systems semi-analytical expressions for the derivatives of the first order reliability index can be derived. The following optimization problem corresponding to the general optimization problems defined in section 3, is considered.  min CI (z, p) = C0 (z, p) + Cj (z, p)Pj (z, p) z j (28) subject to Pf (z, p) ≤ Pfmax where z are decision/design variables, p are quantities defining the costs and/or the stochastic model. Pj denotes a probability (failure or repair), Pf denotes a failure probability and Pfmax is the maximum accepted failure probability. The sensitivity of the total expected costs C with respect to the elements in p is obtained from, see (Haftka & Kamat 1985) and (Enevoldsen 1994)  dPj dPf dC = Cj +λ dpi dpi dpi

(29)

j

where λ is the Lagrangian multiplier associated with the constraint in (25). The sensitivity of the decision variables z with respect to pi can be calculated using the formulas given below which are obtained from a sensitivity analysis of the Kuhn-Tucker conditions related to the optimization problem defined in (28). dz/dpi is obtained from ⎡

⎤ dz ⎡ ⎤ A B ⎢ C ⎥ dp i ⎥=⎣ ⎦ ⎣ ⎦⎢ ⎣ ⎦ dλ 0 BT 0 dpi ⎡

⎤

The elements in the matrix A and the vectors B and C are   ∂2 Cj ∂2 Pf ∂Pj ∂Cj ∂2 Pj ∂ 2 C0 Ars = Pj + +2 + Cj + λ ∂zr ∂zs ∂zr ∂zs ∂zr ∂zs ∂zr ∂zs ∂zr ∂zs

(30)

(31)

j

Br =

∂Pf ∂zr

Cr = −

(32)

  ∂2 Cj ∂Pj ∂Cj ∂ 2 C0 − + ∂zr ∂pi ∂zr ∂pi ∂zr ∂pi

(33)

j

It is seen that the sensitivity of the objective function (the total expected cost) with respect to some parameters can be determined on the basis of the first order sensitivity

44

Structural design optimization considering uncertainties

coefficients of the probabilities and of the cost functions, see (29). However, calculation of the sensitivities of the decision parameters is much more complicated because it involves estimation of the second order sensitivity coefficients of the probabilities, see e.g. (Enevoldsen 1994).

5 Examples 5.1 Ex am pl e 1 – S imple c o s t-b e nefit an a l ys i s In this section a simple, introductory example is presented. A structural component is considered. It is assumed to have strength R and load S, which for simplicity both are Normal distributed: Load S: expected value µS = 20 kN and Coefficient of Variation = 25% Strength R: expected value µR = 50 kN/m2 and Coefficient of Variation = 10% The design variable z represents the cross-sectional area. The limit state equation is written: g = zR − S

(34)

In the initial design situationz = z0 = 1 m2 is chosen. The corresponding reliability index is β = (1 · 50 − 20)/ (1 · 5)2 + 52 = 4.24 and the probability of failure Pf =
(−4.24) = 1.1 · 10−5 . The benefits and cost of failure are B0 = 10 and CF = 107 . New information has been collected. It consists of n = 5 tests with samples of similar components with the following results: 51, 53, 56, 57 and 58 kN/m2 . The mean value of the test results is X = 55 kN/m2 . For updating Bayesian statistics is used. It is assumed that the strength has a known standard deviation σR = 4 kN/m2 . The expected value is assumed to have a prior which is Normal distributed with expected value µ0 = 50 kN/m2 and standard deviation σ0 = 3 kN/m2 . It is noted that these assumptions are consistent with the initial model for the strength (µR = 50 kN/m2 and COV = 10%). The (updated) posterior for the expected value becomes Normal distributed with (nXσ02 + µ0 σR2 )/(nσ02 + σR2 ) = 53.7 kN/m2 and expected value of µR equal to µ

= 

standard deviation of µR equal to σ

= (σ02 σR2 )/(nσ02 + σR2 ) = 1.5 kN/m2 . The predictive (updated) distribution for the strength becomes Normal distributed with expected  value of R equal to µ


= µ

= 53.7 kN/m2 and standard deviation of R

equal to σ


= σ 2 + (σ02 σR2 )/(nσ02 + σR2 ) = 4.3 kN/m2 .  The updated reliability index and probability of failure becomes β = (1 · 53.7 − 20)/ (1 · 4.3)2 + 52 = 5.12 and the probability of failure Pf = 1.56 · 10−7 . At time TR the following two alternatives re-design situations are considered: 1)

continue with existing design The cost-benefits becomes: Z = B0 − CF Pf

= 10 − 107 · 1.5610−7 = 8.44

R e l i a b i l i t y-b a s e d o p t i m i z a t i o n o f e n g i n e e r i n g s t r u c t u r e s

45

9.0

Z(z)

8.5

8.0

7.5

7.0 1.00

1.05

1.10

1.15

1.20

z

Figure 2.5 Cost-benefit as function of design variable z.

2)

use a modified design with increased benefits The design variable is chosen to be z = 1.1 m2 . The benefits are assumed to be changed to: B
(z) = B0 + (z − z0 ) · 0.5. The cost of the design change is assumed to be: CI (z) = 1 + (z − z0 ) · 2. The updated reliability index and probability of failure becomes:  β = (1.1 · 53.7 − 20)/ (1.1 · 4.3)2 + 52 = 5.68 and the probability of failure Pf = 6.60 · 10−9 . The cost-benefits become: Z = B0 + (z − z0 ) · 0.5 − (1 + (z − z0 ) · 2) − CF Pf

= 10 + (1.1 − 1) · 0.5 − (1 + (1.1 − 1) · 2) − 107 · 6.6010−9 = 8.78

Since the cost-benefits are larger for the modified design than continuing with the existing design, the modified design should be chosen. In figure 2.5 the cost-benefits are shown as function of z. It is seen that the optimal decision is to chose a modified design with z = 1.12. It is noted that the known information also could be in the form of an event, e.g. an inspection, and that there could be many more decision alternatives. 5.2 Example2 – Repair decis ion for conc re te bri dg e A road bridge with concrete columns is considered. The total expected lifetime is assumed to be TL . The concrete columns are exposed to chloride ingress due to spread of de-icing salts on and below the bridge. There are some indications that

46

Structural design optimization considering uncertainties

chloride has penetrated the concrete and that corrosion of the reinforcement could be expected within the next few years. Therefore a re-assessment is performed at time TR as illustrated in figure 2.4. Chloride ingress is one of the most common destructive mechanisms for this type of structures. The most typical type of chloride initiated corrosion is pitting corrosion which may locally cause a substantial reduction of the cross-sectional area and cause maintenance and repair actions which can be very costly. Further, the corrosion may make the reinforcement brittle, implying that failure of the structure might occur without warning. The probabilistic analysis of the time to initiation of corrosion in concrete structures is in this example based on models described in (Engelund & Sørensen 1998). At the time of re-assessment it is assumed that chloride profiles are taken from representative parts of the concrete columns. The estimation of the time to initiation of corrosion is based on these chloride profiles combined with prior knowledge. A chloride profile consists of a number of measurements of the chloride concentration as a function of the distance to the surface, y. Using the chloride profiles, the surface concentration and the diffusion coefficient can be estimated. It is assumed that diffusion (transportation) of chlorides into the concrete can be described by a one-dimensional diffusion model where C(y, t) is the content of chloride at time t in the depth y, D(y, t) is the coefficient of diffusion (transportation) at time t in the depth y, CS is the surface concentration and Cinit is the initial chloride concentration. It is assumed that the diffusion coefficients can be written:

a t0 (35) D(y, t) = D0 (y) t where D0 (y) is the reference diffusion coefficient at the reference time t0 and a is an age coefficient (0 < a < 1). Models for the diffusion coefficient can include different diffusion coefficients in different depths. Based on n measurements in one chloride profile the surface concentration cS , the coefficient of diffusion D0 and the age coefficient a can be estimated using the Maximum Likelihood method, see (Engelund & Sørensen 1998). Next using Bayesian statistics a predictive (updated) distribution for the stochastic variables X can be obtained. On the basis of the available information described above the decision maker has to decide which repair/maintenance strategy should be applied. As an example, three different strategies are described below based on the models in (Engelund & Sørensen 1998). All the costs given below are in some monetary unit. It is assumed that the repair is carried out before the probability of any critical event such as total collapse of the bridge. Therefore, in the following the optimization problem is solved without any restriction on the probability of some critical event. Strategy 1: consists of a cathodic protection. This strategy is implemented when corrosion has been initiated at some point. In order to determine when corrosion is initiated, inspections are carried out each year, beginning five years before the expected time of initiation of corrosion. The cost of these inspections is 25 each year except for the last year before expected initiation of corrosion where the cost is 100. The cost of

R e l i a b i l i t y-b a s e d o p t i m i z a t i o n o f e n g i n e e r i n g s t r u c t u r e s

47

the cathodic protection is 1000 and the cost of running the cathodic protection is 20 each year. Strategy 2: is implemented when 5% of the surface of the bridge columns shows minor signs of corrosion, e.g. small cracks and discolouring of the surface. The repair consists of repairing the minor damages and applying a cathodic protection. As for strategy 1 the costs related to this strategy are the costs of the repair and the costs of an extended inspection programme which starts three years before the expected time of repair. However, by this strategy, also the costs related to running the cathodic protection must be taken into account. The cost of repair is 2000, the cost of inspection for three years before the repair is 100 each year and the cost of running the cathodic protection is 30 each year. Strategy 3: repair is performed as a complete exchange of concrete and reinforcement in the corroded areas. The strategy is implemented when 30% of the surface at the bridge columns shows distinct signs of corrosion, such as cracking and spalling of the cover. The cost related to this strategy are the cost of the repair and the cost of an extended inspection programme which starts three years before the expected time of repair. The cost of repair is 3000 and the cost of inspection in the three years before repair is 200 each year. Traffic restrictions in the year of repair he bridge decrease the benefits with 1000. The total expected costs for maintenance/repair is determined from CS (z1 , z2 , z3 ) =

TL 

Pi (z)Ci (z)

(36)

i=TR

where z = (z1 , z2 , z3 ) is the three repair/maintenance options, Pi (z) is the probability that repair/maintenance is performed in year i and Ci (z) is the total costs of the repair strategy if the repair is performed in year i: Ci (z) =

TL  j=TR

Ci,j (z)

1 (1 + r)j−TR

(37)

Ci,j (z) is the repair/maintenance cost in year j if the repair is performed in year i. These costs can be found in the descriptions of the repair strategies. The costs are discounted to the time of re-assessment TR using the real rate of interest r. The expected benefits in the remaining lifetime are determined from B(z) =

TL  i=TR

TL  1 B0 − Pi (z)Bi (z) (1 + r)i−TR

(38)

i=TR

where Bi (z) =

TL  j=TR

Bi,j (z)

1 (1 + r)j−TR

(39)

B0 is the basic annual benefit from use of the bridge and Bi,j (z) is the loss of benefits in year j due to repair in year i, e.g. due to traffic restrictions.

48

Structural design optimization considering uncertainties

The optimal repair strategy is obtained solving the optimization problem max B(z) − CS (z) z

(40)

The expected costs are determined using the predictive stochastic model for the surface concentration cS , the coefficient of diffusion D0 and the age coefficient a obtained using the available information. 5.3

Ex am pl e 3 – O pt imal d es ig n o f o ffsh o r e w i n d t u r b i n e s

Wind turbines for electricity production are increasing drastically these years both in production capability and in size. Offshore wind turbines with an electricity production of 2–5 MW are now being produced. The main failure modes are fatigue failure of wings, hub, shaft and main tower, local buckling of main tower, and failure of the foundation. This example considers reliability-based optimization of the tower and foundation, see (Sørensen & Tarp-Johansen 2005a) and (Sørensen & Tarp-Johansen 2005b). 5.3.1 F ormu l a tio n o f r e lia b ilit y-b a s e d o p t im i z at i o n pr o bl e ms f or w in d tu r b in e s Reliability based optimization problems can be formulated in different ways, e.g. with or without systematic reconstruction. In this example it is assumed that the control system is performing as expected, one single wind turbine is considered and the wind turbine is systematically reconstructed in case of failure. It is noted that it is assumed that the probability of loss of human lives is negligible. The the main design variables are denoted z = (z1 , . . . , zN ), e.g. diameter and thickness of tower and main dimension of wings. The initial (building) costs are CI (z), the direct failure costs are CF , the benefits per year are b and the real rate of interest is γ. Failure events are assumed to be modelled by a Poisson process with rate λ. The probability of failure is PF (z). The optimal design can thus be determined from the following optimization problem, see section 3.4:

 CI (z) CF λPF (z) b CI (z) − − + max W(z) = z γC0 C0 C0 C0 γ (41) l u subject to zi ≤ zi ≤ zi , i = 1, . . . , N PF (z) ≤ PFmax where zl and zu are lower and upper bounds on the design variables. C0 is the reference initial cost of corresponding to a reference design z0 . PFmax is the maximum acceptable probability of failure e.g. with a reference time of one year. This type of constraint is typically required by regulators. The optimal design z∗ is determined by solution of (41). If the constraint on the maximum acceptable probability of failure is omitted, then the corresponding value PF (z∗ ) can be considered as the optimal probability of failure related to the failure event and the actual cost-benefit ratios used. The failure rate λ and probability of failure can be estimated for the considered failure event, if a limit state equation, g(X1 , . . . , Xn , z) and a stochastic model for the

R e l i a b i l i t y-b a s e d o p t i m i z a t i o n o f e n g i n e e r i n g s t r u c t u r e s

49

DT

t1

H/3

t2

H/3

t3

H/3

H

hw d tP

HP

D DP

Figure 2.6 Design variables in wind turbine example (not in scale).

stochastic variables, (X1 , . . . , Xn ) are established. If more than one failure event is critical, then a series-parallel system model of the relevant failure modes can be used. An offshore 2 MW wind turbine with monopile foundation is considered, see figure 2.6. The wind turbine tower has height h = 63 m and a diameter which h increases linearly from D at bottom to DT at the top. The tower is divided in three sections each with height h/3 and each with the same thickness: t1 in top section, t2 in middle and t3 in bottom section. Diameter and thickness of monopile are constant: DP and tP . Tower and monopile are made of structural steel. The distance from bottom of the tower to the water surface is hw = 7 m and the distance from the water surface to the sea bed (the water depth) is d = 9 m. Wind and wave loads on the tower itself are neglected. The following failure modes are included: (a) yielding in cross sections in tower just above and below changes in thickness, (b) local stability in cross sections in tower just above and below changes in thickness, (c) fatigue in cross sections just above and below changes in thickness, and (d) yielding in monopile in cross-section with maximum bending moment. The stochastic model for the extreme loading at the top of the tower is described in (Sørensen & Tarp-Johansen 2005a) and (Sørensen & Tarp-Johansen 2005b). For the failure mode yielding of cross-section the limit state function is written: σ=

M N + ≥ Fy A W

(42)

where the cross-sectional forces in the cross-section is the normal force N, a shear force Q and a bending moment M. Further A is the cross-sectional area (= πt(D − t)), W is the cross-sectional section modulus and Fy is the yield stress.

50

Structural design optimization considering uncertainties

The cross-sectional forces are calculated from the stochastic variables HT , MT , and NT . The yield stress, Fy , is modelled as a LogNormal variable with coefficient of variation (COV) = 0.05 and characteristic values (5 percentile) equal to 235 MPa and 340 MPa for the tower and the mono-pile, respectively. For the failure mode local buckling of cross-section the limit state function is written: σ=

N M + ≥ Fyc A W

(43)

where the local buckling strength is estimated by the model in (ISO 19902 2001). The cross-sectional forces are calculated from the stochastic variables HT and MT . The yield stress, Fy is modelled as for yielding failure. Model uncertainty is introduced through a factor XB multiplied to Fyc . XB is assumed LogNormal distributed with expected value 1 and COV = 0.10. For the failure mode fatigue failure SN-curves and linear damage accumulation by the Miner rule are used. It is assumed that the SN-curve is bilinear and can be described by: N = K1 ( s)−m1

for N ≤ NC

(44)

N = K2 ( s)−m2

for N > NC

(45)

where s is the stress range, N is the number of cycles to failure, K1 , m1 are the material parameters for N ≤ NC , K2 , m2 are the material parameters for N > NC , sC is the stress range corresponding to NC . Further it is assumed that the total number of stress ranges for a given fatigue critical detail can be grouped in nσ groups/bins such that the number of stress ranges in group i is ni per year. In a deterministic design check the design equation can be written: 

ni TF

si ≥ sC

1 K1C s−m i

+



ni TF

si < sC

2 K2C s−m i

≥1

(46)

where si = Mi /z is the stress range in group i, Mi is the bending moment range z is a design parameter, KiC is the characteristic value of Ki (log KiC mean of log Ki minus two standard deviations of log Ki ), TF = FDF TL is the fatigue life time, TL is the service life and FDF is the Fatigue Design Factor which can be considered as a fatigue safety factor. In a reliability analysis the reliability index (or the probability of failure) is calculated using the limit state function associated with (46). This limit state equation can be written: g =1−



ni TL

si ≥ sC

1 K1 s−m i

−



ni TL

si < sC

2 K2 s−m i

(47)

where si = XS Mi /p is the stress range in group i, XS is a stochastic variable modelling model uncertainty related to the fatigue wind load and to calculation of the relevant fatigue stresses with given wind load. XS is assumed LogNormal distributed with

R e l i a b i l i t y-b a s e d o p t i m i z a t i o n o f e n g i n e e r i n g s t r u c t u r e s

51

Table 2.1 Stochastic model. D: Deterministic; N: Normal; LN: LogNormal. Variable

Distribution

Expected value

Standard deviation

X stress X wind TL m1 log K 1 m2 log K 2

LN LN D D N D N

1 1 T F /FDF 3 12.151 + 2 · 0.20 5 15.786 + 2 · 0.25

COV stress = 0.05 COV wind = 0.15 20 years 0.20 0.25

log K 1 and log K 2 are fully correlated.

 2 2 mean value = 1 and COV = COVwind + COVstress . log Ki is modelled by a Normal distributed stochastic variable according to a specific SN-curve. Representative statistical parameters are shown in Table 2.1. The basic SN curve used correspond to the SN 90 curve in (EC 3 2003). The optimal design is determined from the following optimization problem: b CI (z) max W(z) = − − z rC0 C0 subject to

pli ≤ pi ≤ pui , PF (z) ≤

CI (z) CF + C0 C0



λPF (z) γ

i = 1, . . . , N

(48)

PFmax

ω1 (z) ≥ ωL where PFmax is the maximum acceptable annual probability of failure. ω1 is the lowest natural frequency of the wind turbine structure and ωL is a minimum acceptable eigen frequency.  The probability of failure is estimated by the simple upper bound: PF ≈ N i=1
(−βi ) where βi is the annual reliability index in failure element i of the N failure elements/failure modes. The following design/optimization variables related to the tower and pile model are used: DT is the diameter at tower top, D is the diameter of tower at bottom, t1 , t2 and t3 are thickness of tower sections, DP is the diameter of the monopile, tP is the thickness of monopile, HP is the length of the monopile. The initial costs is modelled by:

 1 1 Vmono + CI = C0,foundation 2 2 Vmono,0

 1 3 Vtower + CI,blades + CI,powertrain + CI,others + C0,tower + 4 4 Vtower,0    turbine

(49)

52

Structural design optimization considering uncertainties

where Vmono,0 and Vtower,0 are reference cross-sectional areas for the mono-pile foundation and the tower, respectively. Thus, the model is a linear model that gives the initial costs for designs that deviate from a given reference. The term CI,others accounts for initial costs connected to external and internal grid connections that are of course independent of the extreme load. Because, in current practice, the design of the blades and the power train are driven by fatigue and operation loads respectively, the dependence of the initial costs of these main parts of the turbine on the extreme load is assumed negligible in this model. The following model is used for the normalised initial costs at the considered site  

 CI 1 1 1 3 Vtower 1 1 1 1 1 Vmono 1 + + + + + (50) = + C0 6 2 2 Vmono,0 2 3 4 4 Vtower,0 3 3 3    turbine The ratios appearing in this formula will be site specific. For a far off offshore site the grid connection will become a larger part of the total costs. Likewise the foundation costs will depend on water depth. For other sites the cost ratios may e.g. be: 1 5 , , and 13 for the foundation, the turbine, and the other costs, respectively. For the 4 12 reference turbine Vmono,0 = 25.5 m3 and Vtower,0 = 14.0 m3 , which have been derived from the following reference values: h = 63 m, hw = 7 m, DT = 2.43 m, tT = 17 mm, DB = 3.90 m, tB = 29 mm, hP = 41 m, tp = 49.5 mm, and DP = 4.1 m. Thus 1 Vmono 1 Vtower CI 1 1 1 1 + + = + + + C0 12 12 14.0 m3 24 8 23.2 m3 3 3    turbine

(51)

It is noted that out of the total initial costs only a minor part depends on the loads because the study is restricted to the support structure. For a gravity foundation the normalised failure costs are estimated to be: CF,foundation 1 = C0,foundation 6

(52)

Compared to this the failure costs for the turbine are negligible. The turbine failure costs could be virtually zero if one just leaves the turbine at the bottom of the sea like a shipwreck, a solution that may hardly be accepted by environmentalists. It is noted that, at least in Denmark, it is for aesthetic reasons not accepted to rebuild the turbine a little away from the collapsed turbine, whereby the failure costs could otherwise practically vanish. Indeed Danish building licences demand that a new turbine, which replaces a collapsed turbine, must be situated at the exact same spot. That is, the space cannot even be left unused. Assuming that the damage to the grid is small the failure costs become: CF 1 = C0 36

(53)

For the considered site and turbine, and Assumption: Given site-i.e. climate (A = 10.8, k = 2.4), specified rated power (2 MW) and turbine height and rotor

R e l i a b i l i t y-b a s e d o p t i m i z a t i o n o f e n g i n e e r i n g s t r u c t u r e s

53

Table 2.2 Optimal values of design variables, objective function and natural frequency. γ

0.03

0.05

0.10

0.05

0.05

b/C 0 C F /C 0 DT D t1 t2 t3 DP tP HP W ω1

1/8 1/36 2.92 m 4.00 m 20 mm 28 mm 35 mm 5.41 m 21 mm 34.7 m 3.264 2.71

1/8 1/36 2.89 m 4.00 m 20 mm 29 mm 33 mm 5.40 m 20 mm 34.7 m 1.602 2.67

1/8 1/36 2.81 m 4.00 m 20 mm 25 mm 32 mm 4.93 m 20 mm 34.7 m 0.359 2.43

1/10 1/36 2.89 m 4.00 m 20 mm 29 mm 33 mm 5.40 m 20 mm 34.7 m 1.102 2.67

1/8 1/360 2.77 m 4.00 m 20 mm 28 mm 33 mm 5.31 m 20 mm 34.7 m 1.603 2.63

diameter. The average power is 1095 kW which with an assumption of 2% down time the annual average production may be computed. In the Danish community subsidising currently ensures that the market price for 1 kWh wind turbine generated electric power is 0.43 DKK/kWh. From this, one should subtract, as a lifetime average, 0.1 DKK/kWh for operation and maintenance expenses. The normalised average benefits per year becomes approximately b 1 = C0 8

(54)

The real rate of interest r is assumed to be 5% because, as argued, a purely monetary reliability optimization is considered. Assuming a lower tower frequency of 0.33 Hz a frequency constraint becomes ω1 ≥ 2π 0.33 Hz = 2.07 s−1 . The optimal design is determined from the optimization problem (5.9). The following bounds on the design variables are used: Thicknesses: 20 mm and 50 mm Diameter tower: 2m and 4 m Diameter monopile: 2 m and 6 m The optimal values of the design variables are shown in Table 2.2, including cases where the real rate of interest r is 3%, 5% and 10%, b/C0 is 1/8 and 1/10, and CF /C0 is 1/36 and 1/360. In Table 2.3 reliability indices for the different failure modes and for the system are shown. It is seen that • •

For increasing rate γ the dimensions and the value of the objective function as expected decreases. Further also the corresponding system reliability indices and eigenfrequencies decrease slightly. The optimal dimensions are not influenced by a change in the benefits – only the value of the objective function decreases with decreasing benefits per year.

54

Structural design optimization considering uncertainties

Table 2.3 Optimal values reliability indices for failure modes and system – first value is for local buckling/yielding and second value is for fatigue. γ

0.03

0.05

0.10

0.05

0.05

b/C 0 C F /C 0 Top section

1/8 1/36 13.6/4.90 5.63/4.25 7.96/5.62 5.12/3.67 6.37/4.32 5.08/3.60 7.09/4.09 3.41

1/8 1/36 13.4/4.79 5.55/4.20 8.02/5.64 5.22/3.72 6.08/4.09 4.85/3.49 7.09/3.98 3.34

1/8 1/36 12.9/4.52 5.37/4.08 6.91/5.01 4.35/3.50 5.79/3.95 4.66/3.26 7.09/3.47 3.06

1/10 1/36 13.4/4.79 5.55/4.20 8.02/5.64 5.22/3.72 6.08/4.09 4.85/3.49 7.09/3.98 3.34

1/8 1/360 12.7/4.52 5.28/4.01 7.39/5.20 4.83/3.54 5.95/4.01 4.86/3.49 7.09/3.86 3.26

Middle section Bottom section

Top Bottom Top Bottom Top Bottom

Pile System

• • • •

For decreasing failure costs the optimal dimensions, the objective function, the system reliability level and the eigenfrequency decrease slightly. The system reliability index β is 3.1–3.4. In this example the fatigue failure mode has the smallest reliability indices (largest probabilities of failure). The frequency constraint is not active.

The example shows that the optimal reliability level related to structural failure of offshore wind turbines is of the order of a probability per year equal to 2 · 10−4 − 10−3 corresponding to an annual reliability index equal to 3.1–3.4. This reliability level is significantly lower than for civil engineering structures in general.

6 Conclusions The theoretical basis for reliability-based structural optimization within the framework of Bayesian statistical decision theory is briefly described. Reliability-based cost benefit problems are formulated and exemplified with structural optimization. The basic reliability-based optimization problems are generalized to the following extensions: interactive optimization, inspection and repair costs, systematic reconstruction, re-assessment of existing structures. Illustrative examples are presented including a simple introductory example, a decision problem related to bridge re-assessment and a reliability-based decision problem for offshore wind turbines.

References Aitchison, J. & Dunsmore, I.R. 1975. Statistical Prediction Analysis. Cambridge University Press, Cambridge. Ang, H.-S.A. & Tang, W.H. 1975. Probabilistic concepts in engineering planning and design, Vol. I and II, Wiley.

R e l i a b i l i t y-b a s e d o p t i m i z a t i o n o f e n g i n e e r i n g s t r u c t u r e s

55

Benjamin, J.R. & Cornell, C.A. 1970. Probability, Statistics and Decision for Civil Engineers. Mc-Graw-Hill. EN 1993-1-9 2003. Eurocode 3: Design of steel structures – Part 1–9: Fatigue. Enevoldsen, I. & Sørensen, J.D. 1993. Reliability-Based Optimization of Series Systems of Parallel Systems. ASCE Journal of Structural Engineering, Vol. 119, No. 4, pp. 1069–1084. Enevoldsen, I. 1994. Sensitivity Analysis of a Reliability-Based Optimal Solution. ASCE, Journal of Engineering Mechanics. Enevoldsen, I. & Sørensen, J.D. 1994. Reliability-based optimization in structural engineering. Structural Safety, Vol. 15, pp. 169–196. Enevoldsen, S. & Sørensen, J.D. 1998. A Probabilistic Model for Chloride-Ingress and Initiation of Corrosion in Reinforced Concrete Structures. Structural Safety, Vol. 20, pp. 69–89. Engelund, S. 1997. Probabilistic models and computational methods for chloride ingress in concrete. Ph.D. thesis, Department of Building Technology and Structural Engineering, Aalborg University. Faber, M.H., Engelund, S. Sørensen, J.D. & Bloch, A. 1989. Simplified and generic risk based inspection planning. Proc. OMAE2000, New Orleans. Frangopol, D.M. 1985. Sensitivity of reliability-based optimum design. ASCE, Journal of Structural Engineering, Vol. 111, No. 8, pp. 1703–1721. Fujimoto, Y., Itagaki, H., Itoh, S., Asada, H. & Shinozuka, M. 1989. Bayesian Reliability Analysis of Structures with Multiple Components. Proceedings ICOSSAR 89, pp. 2143–2146. Fujita, M., Schall, G. & Rackwitz, R. 1989. Adaptive Reliability Based Inspection Strategies for Structures Subject to Fatigue. Proceedings ICOSSAR 89, pp. 1619–1626. Gill, P.E., Murray, E.W. & Wright, M.H. 1981. Practical Optimization. Academic Press. Haftka, R.T. & Kamat, M.P. 1985. Elements of Structural Optimization. Martinus Nijhoff, The Hague. ISO 19902 2001. Petroleum and natural gas industries – Fixed steel offshore structures. Kroon, I.B. 1994. Decision Theory Applied to Structural Engineering Problems. Ph.D. thesis, Department of Building Technology and Structural Engineering, Aalborg University. Kuschel, N. & Rackwitz, R. 1998. Structural optimization under time-variant reliability constraints. Proc. 8th IFIP WG 7.5 Conf. On Reliability and optimization of structural systems, University of Ann Arbor, pp. 27–38. Lindley, D.V. 1976. Introduction to Probability and Statistics from a Bayesian Viewpoint, Vol. 1 + 2. Cambridge University Press, Cambridge. Madsen, H.O. & Friis-Hansen, P. 1992. A comparison of some algorithms for reliabilitybased structural optimization and sensitivity analysis. Proc. IFIP WG7.5 Workshop, Munich, Springer-Verlag, pp. 443–451. Madsen, H.O. & Sørensen, J.D. 1990. Probability-Based Optimization of Fatigue Design Inspection and Maintenance. Presented at Int. Symp. On Offshore Structures, University of Glasgow. Madsen, H.O., Krenk, S. & Lind, N.C. 1986. Methods of Structural Safety. Prentice-Hall. Murotsu, Y., Kishi, M., Okada, H., Yonezawa, M. & Taguchi, K. 1984. Probabilistically optimum design of frame structures. Proc. 11th IFIP Conf. On System modeling and optimization, Springer-Verlag, pp. 545–554. Madsen, H.O., Sørensen, J.D. & Olesen, R. 1989. Optimal Inspection Planning for Fatigue Damage of Offshore Structures. Proceedings ICOSSAR 89, pp. 2099–2106. Powell, M.J.D. 1982. VMCWD: A FORTRAN Subroutine for Constrained Optimization. Report DAMTP 1982/NA4, Cambridge University, England. Rackwitz, R. 2001. Risk control and optimization for structural facilities. Proc. 20th IFIP TC7 Conf. On System modeling and optimization, Trier, Germany. Raiffa, H. & Schlaifer, R. 1961. Applied Statistical Decision Theory. Harward University Press, Cambridge, Mass.

56

Structural design optimization considering uncertainties

Rosenblueth, E. & Mendoza, E. 1971. Reliability optimization in isostatic structures. J. Eng. Mech. Div. ASCE, pp. 1625–1642. Schittkowski, K. 1986. NLPQL: A FORTRAN Subroutine Solving Non-Linear Programming Problems. Annals of Operations Research. Skjong, R. 1985. Reliability-Based Optimization of Inspection Strategies. Proc. ICOSSAR’85 Vol. III. pp. 614–618. Streicher, H. & Rackwitz, R. 2002. Structural optimization – a one level approach. Proc. Workshop on Reliability-based Design and Optimization – rbo02, IPPT, Warsaw. Sørensen, J.D. & Thoft-Christensen, P. 1985. Structural optimization with reliability constraints. Proc. 12th IFIP Conf. On System modeling and optimization, Springer-Verlag, pp. 876–885. Sørensen, J.D. & Thoft-Christensen, P. 1988. Inspection Strategies for Concrete Bridges. Proc. IFIP WG 7.5, Springer Verlag, Vol. 48, pp. 325–335. Sørensen, J.D., Thoft-Christensen, P., Siemaszko, A., Cardoso, J.M.B. & Santos, J.L.T. 1995. Interactive reliability-based optimal design. Proc. 6th IFIP WG 7.5 Conf. On Reliability and Optimization of Structural Systems, Chapman & Hall, pp. 249–256. Sørensen, J.D. & Tarp-Johansen, N.J. 2005a. Reliability-based optimization and optimal reliability level of offshore wind turbines. International Journal of Offshore and Polar Engineering (IJOPE), Vol. 15, No. 2, pp. 1–6. Sørensen, J.D. & Tarp-Johansen, N.J. 2005b. Optimal Structural Reliability of Offshore Wind Turbines. CD-rom Proc. ICOSSAR’2005, Rome. Thoft-Christensen, P. & Sørensen, J.D. 1987. Optimal Strategies for Inspection and Repair of Structural Systems. Civil Engineering Systems, Vol. 4, pp. 94–100. von Neumann, J. & Morgenstern, O. 1943. Theory of Games and Economical Behavior. Princeton University Press.

Chapter 3

Reliability analysis and reliabilitybased design optimization using moment methods Sang Hoon Lee Northwestern University, Evanston, IL, USA

Byung Man Kwak Korea Advanced Institute of Science and Technology, Daejeon, Korea

Jae Sung Huh Korea Aerospace Research Institute, Daejeon, Korea

ABSTRACT: Reliability analysis methods using the design of experiments (DOE) are introduced and integrated into reliability based design optimization (RBDO) frame work with a semi-analytic design sensitivity analysis (DSA) for the reliability measure. A procedure using the full factorial DOE with optimal levels and weighs is introduced and named as full factorial moment method (FFMM) for reliability analysis. The probability of failure is calculated using an empirical distribution system and the first four statistical moments of system performance function calculated from DOE. To enhance the efficiency of FFMM, a response surface augmented moment method (RSMM) is developed to construct a series of approximate response surface approaching to that of FFMM. A semi-analytic design sensitivity analysis for the probability of failure is proposed in combination with FFMM and RSMM. It is shown that the proposed methods are accurate and effective especially when the inputs are non-normal.

1 Introduction One of the fundamental problems in the structural reliability theory is the calculation of the probability of failure which is defined as a multifold probability integral of the joint probability density function of random variables over the domain of structural failure. Because the analytic calculation of this integral is practically impossible, many approximate methods and simulation methods are developed so far (Madsen et al. 1986, Kiureghian 1996, Bjerager 1991). Among the methods, the first order reliability method (FORM) (Hasofer & Lind 1974, Rackwitz & Fiessler 1978) is considered to be one of the most efficient computational methods and over the past three decades, contributions from numerous studies have made FORM the most popular reliability method. The reliability based design approaches (Lee & Kwak 1987–1988, Enevoldsen & Sørensen 1994, Frangopol & Corotis 1996, Tu et al. 1999, Youn et al. 2003) have adopted FORM as their main analysis tool for reliability due to its efficiency. The difficulties in FORM such as numerical difficulty in finding the most probable failure point (MPFP), errors involved in the nonlinear failure surface including the possibility of multiple design points (Kiureghian & Dakessian 1998), and errors caused

58

Structural design optimization considering uncertainties

by non-normality of variables (Hohenbichler & Rackwitz 1981) are well recognized and efforts to overcome these are also made. They include the second order reliability method (SORM) (Fiessler et al. 1979, Breitung 1984, Koyluoglu & Nielsen 1994, Kiureghian et al. 1987), the advanced Monte Carlo simulation (MCS) such as importance sampling (Bucher 1988, Mori & Ellingwood 1993, Melchers 1989), directional sampling (Bjerager 1988, Nie & Ellingwood 2005) and the response surface based approaches (Faravelli 1989, Bucher & Bourgund 1990, Rajashekhar & Ellingwood 1993). However, finding MPFP is a still numerically difficult task in FORM and often the error involved degrades the accuracy of final results. In this chapter, we investigate another route for structural reliability, the moment method. The moment method calculates the probability of failure by computing the statistical moments of the performance function and fitting the moments with some empirical distribution systems such as the Pearson system, Johnson system, Gram-Charlier series, and so on (Johnson et al. 1995). For this purpose, the performance function must be computed for a set of well-designed calculation points, often called quadrature points or designed experimental points. Compared with FORM, the moment method has advantages that it dose not involve the difficulties of searching for the MPFP and the information of cumulative distribution function (CDF) is readily available. Not so many attempts are reported about the reliability analysis using the moment method. For statistical moment estimation, Evans (1972) proposed a quadrature formula which uses 2n2 + 1 nodes and weights for a system with n random variables and applied it to tolerance analysis problems. Li & Lumb (1985) adopted Evans’ quadrature formula in structural reliability analysis in combination with the Pearson system. Rosenblueth (1981) devised a 2n point estimate method and (Hong 1996) proposed a nonlinear system of equations for point estimate of probability in combination with the Johnson distribution system and Gram-Charlier series. Zhao & Ono (2001) proposed a point estimate method using Rosenblatt transformation and kn point concentration where k is the number of quadrature points for each random variable. Taguchi (1978) proposed a design of experiment (DOE) technique which uses three level experiments for each random variable to calculate the mean and standard deviation of performance function for tolerance design. Taguchi’s method was improved by (D’Errico & Zaino 1988). These methods can treat only normally distributed random variables and the DOE becomes a 3n full factorial design when n random variables are under consideration. Actually, the levels and weights proposed by D’Errico & Zaino are equivalent to the nodes and weights in Gauss-Hermite quadrature formula (Abramowitz & Stegun 1972, Engels 1980). Seo & Kwak (2002) extended D’Errico & Zaino’s method to treat non-normal distributions by deriving an explicit formula of three levels and weights for general distributions. In addition to the strong points of moment method mentioned above, the moment method using DOE has several good aspects. It is very easy and simple to use and does not involve any deterioration of accuracy or additional efforts for treating non-normal random variables. However, the common problem of the moment based methods is the numerical efficiency. The methods often tend to become very expensive as the number of random variables increases. To overcome this shortcoming, (Lee & Kwak 2006) developed a new moment method integrating the response surface method with the 3n full factorial DOE. (Huh et al. 2006) developed a response surface approximation scheme based on the moment method and applied it to the design study of a precision nano-positioning system.

Reliability analysis and optimization using moment methods

59

In this chapter, we present our previous developments of moment methods which utilize DOE for statistical moment estimation and propose a RBDO framework with a semi-analytic design sensitivity analysis in combination with the moment methods. In section 2, the full factorial moment method (FFMM) is introduced with an explanation on the selection of optimal DOE. In section 3, the response surface augmented moment method (RSMM) is introduced and the accuracy and efficiency of RSMM are compared with other methods via several examples. In section 4, a RBDO procedure is proposed using FFMM and RSMM with a semi-analytic design sensitivity analysis. Section 5 provides some discussions on moment methods and the proposed RBDO procedure and concluding remarks.

2 Reliability analysis using full factorial moment method The probability of failure of a system is defined by a multifold probability integral as
Pf = Pr [g(X) < 0] =

fX (x)dx

(1)

g(x)<0

where X is the vector of input random variables, g(x) is the system performance function whose negative value indicates the failure state, and fX (x) is the joint probability density function (PDF) of X. In moment method, the probability of failure is calculated from the PDF of g(X) which is found by fitting the first four statistical moments of system performance function with empirical distribution systems as in Equation 2:





Pf =

fX (x)dx = g(x)<0

0 −∞

fg(X) (g(x))dg(x)

(2)

where fg means the PDF of g(X). One essential procedure for this calculation is the accurate estimation of the statistical moments of g(X). Since the empirical distribution systems require high order moments usually up to fourth order, approximate methods using perturbation or Taylor series expansion are not adequate in most cases. In this section, we introduce a moment estimation scheme based on the design of experiments (DOE) which is applicable to general non-normal random variables. And a reliability analysis procedure using the Pearson system is introduced with examples.

2.1

Design of experiments for s tatis tical mo me nt e s ti mati o n

For a random variable X, the k-th order statistical moments of a one-dimensional function g(X) can be calculated using a quadrature formula with m nodes as follows:
E{g k } =

∞

−∞

[g(x)]k fX (x)dx

≈ w1 [g(µ + α1 σ)]k + w2 [g(µ + α2 σ)]k + · · · + wm [g(µ + αm σ)]k

(3)

60

Structural design optimization considering uncertainties

where fX (x) is the probability density function of X, µ and σ denote the mean and standard deviation of X, respectively. To estimate accurately up to the fourth moment, which is often required by the empirical distribution systems such as the Pearson system, at least three node quadrature rule is necessary and the parameters, αi and wi can be found ideally by solving the following moment matching equations (Engels 1980):
∞ µk = (x − µ)k fX (x)dx −∞

= w1 (α1 σ)k + w2 (α2 σ)k + · · · + wm (αm σ)k , (k = 0, 1, . . . , 2m − 1)

(4)

where µk is the k-th statistical moment of random variable X, which can be calculated from the PDF of X, and 2m − 1 is the polynomial order of the quadrature rule. By introducing levels, li = µ + αi σ, Equation 4 can be rewritten in terms of li and wi from the point of view of an experimental design. In the equations, there are 2m unknowns and the number of equations is also 2m. Thus, if we provide the values of µk , li and wi are uniquely determined. It is not a simple task to find the solution of Equation 4 algebraically. For 3 level DOE, (Seo & Kwak 2002) derived a simple explicit formula of li and wi , which will be discussed in section 2.2. For general cases, Equation 4 can be solved with a numerical method but from the experience of the authors, it is found that solving Equation 4 for cases m ≥ 7 is very difficult. When there are n random variables in the system and if we use the same number of levels, m for each random variable, the DOE becomes a mn full factorial design from the product quadrature rule and the first four statistical moments of system response function g(X) can be calculated as follows: µg =

m 

w1·i1 · · ·

i1 =1

⎡ σg = ⎣

m 

wn·in g(l1·i1 , . . . , ln·in )

(5)

in =1

m 

w1·i1 · · ·

i1 =1

m 

⎤1/2 wn·in (g(l1·i1 , . . . , ln·in ) − µg )2 ⎦

⎡ ⎤ m m    β1g = ⎣ w1·i1 · · · wn·in (g(l1·i1 , . . . , ln·in ) − µg )3 ⎦ σg3 i1 =1

⎡ β2g = ⎣

m 

i1 =1

(6)

in =1

w1·i1 · · ·

(7)

in =1 m 

⎤ wn·in (g(l1·i1 , . . . , ln·in ) − µg )4 ⎦ σg4

(8)

in =1

 where wi·j and li·j mean the j-th weight and level of i-th variable and µg , σg , β1g and β2g denote the mean, standard deviation, skewness and kurtosis of g(X), respectively. If we know a priori that g(X) shows more nonlinear dependence on some of the random variables, we can use more levels for those variables. In general, we can use different number of levels for each random variable, e.g. m1 , . . . , mn instead of m in Equations

Reliability analysis and optimization using moment methods

61

5–8, and the total number of experiments then becomes m1 · m2 · · · mn . The extension to the general case is straightforward. 2.2

Determination of optimal levels and we i g hts

In this section, we look further inside how we can find the optimal levels and weights for statistical moment estimation. As mentioned in section 2.1, they can be found by solving the moment matching equation, Equation 4. When we use three levels, that is, m is chosen as 3, the system of equations can be written as follows: w1 + w2 + w3 = 1

(9)

w1 l1 + w2 l2 + w3 l3 = µ

(10)

w1 (l1 − µ)2 + w2 (l2 − µ)2 + w3 (l3 − µ)2 = σ 2

(11)

w1 (l1 − µ)3 + w2 (l2 − µ)3 + w3 (l3 − µ)3 =

 β1 σ 3

(12)

w1 (l1 − µ)4 + w2 (l2 − µ)4 + w3 (l3 − µ)4 = β2 σ 4

(13)

w1 (l1 − µ)5 + w2 (l2 − µ)5 + w3 (l3 − µ)5 = µ5

(14)

Since it is difficult to solve these equations algebraically, (Seo & Kwak 2002) replaced the condition on the fifth moment (Eq. 14) with l2 = µ and obtained an explicit formula of li and wi in terms of first four statistical moments of input random variable as follows: √

⎤T σ β1 σ ⎢ µ + 2 − 2 4β2 − 3β1 ⎥ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ µ {l1 , l2 , l3 } = ⎢ ⎥ ⎥ ⎢ √ ⎦ ⎣ σ β1 σ + µ+ 4β2 − 3β1 2 2 ⎡

⎡ ⎢ ⎢ ⎢ ⎢ {w1 , w2 , w3 } = ⎢ ⎢ ⎢ ⎢ ⎣

√  (4β2 − 3β1 ) + β1 4β2 − 3β1 2(4β2 − 3β1 )(β2 − β1 ) β2 − β1 − 1 β2 − β 1 √  (4β2 − 3β1 ) − β1 4β2 − 3β1 2(4β2 − 3β1 )(β2 − β1 )

(15)

⎤T ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎦

(16)

The levels and weights in Equations 15 and 16 can be calculated very easily, but applicable to only symmetric distributions in their exactness. With dissymmetric distributions such as lognormal, exponential distribution, the first level l1 can be located outside the domain where the distribution is defined. For example, for an exponential

62

Structural design optimization considering uncertainties

1.0 0.9 0.8 Proposed 3 level Seo & Kwak's 3 level P.D.F. Outside

Weight

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 1

0

1

2

3 Level

4

5

6

7

Figure 3.1 Levels and weights from the moment matching equation and by Seo & kwak’s formula for an exponential distribution (λ = 1).

distribution defined in (x ≥ 0) with distribution parameter λ (Hahn & Shapiro 1967), the four statistical moments are calculated as, µ=

1 , λ

σ=

1 , λ

 β1 = 2,

β2 = 9

(17)

and the levels by Equation 15 are given as follows: l1 =

√ √ 1 1 1 (2 − 6) < 0, l2 = , l3 = (2 + 6) λ λ λ

(18)

It is seen that the first level is located outside the domain of distribution and this may cause severe numerical problems in applying DOE for moment estimation. The other problem of the levels and weights by Equations 15 and 16 is that since the requirement on the fifth moment is replaced with the requirement that the midlevel should be the mean value of the random variable, the accuracy of high moment estimation degrades for a nonlinear performance function. It will be illustrated in the examples and the levels and weights by Equations 15 and 16 may not be optimal in terms of accuracy, unless all the distributions have symmetry. With this reason, it is preferable to solve Equation 4 directly with numerical methods and numerical equation solving algorithms such as modified Powell’s hybrid method (More et al. 1980) can be used for this purpose. Figure 3.1 compares the levels and weights obtained by solving Equation 4 with numerical methods with those obtained by Equations 15 and 16 for the exponential distribution example mentioned above. It is found that the levels and weight obtained directly from Equation 4 are free from the

Reliability analysis and optimization using moment methods

63

0.7 0.6

0.6

0.5

0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1 0

0 0

1

2

3

4

5

6

7

8

9

10

0

1

2

3

4

5

6

7

8

9

10

(b) Uniform distribution 0  x  10

(a) Normal distribution m
5, s
1 0.7

7

0.6

6

0.5

5

0.4

4

0.3

3

0.2

2

0.1

1 0

0 0

1

2

3

4

5

6

7

8

(c) Rayleigh distribution ^ s
2

9

10

0

0.1 0.2 0.3 0.4

0.5 0.6 0.7 0.8 0.9

1

(d) Beta distribution h
0.3, g
0.6

Figure 3.2 Three level DOE for different distributions. Notations of the distribution parameters are from (Hahn & Shapiro 1967).

problems found in Seo & Kwak’s levels and weights. In Figure 3.2, levels and weights for different distributions are depicted for a three level case. The extension of the procedure for finding levels and weights to cases with more than three levels is straightforward, but there are two difficulties in calculation. Firstly, the calculation of high order moments of input random variable can be complicated and tedious depending on the PDF of the input variable. For example, when calculating the levels and weights for a five level DOE (m = 5), the moments of input X should be provided up to ninth order, which might be very complicated for some distributions. Secondly, the solution of Equation 4 becomes very difficult or sometimes impossible. From our experience, the calculation up to five level DOE turns out to be manageable, but the calculation for more levels than five was not successful in many cases. When the PDF of a distribution has the same form with the weight function in a Gaussian quadrature formula (Abramowitz & Stegun 1972), the levels and weights can be directly derived from the nodes and weights of the Gaussian quadrature rule instead of solving Equation 4. Among the well known distributions, the normal distribution matches with the Gauss-Hermite quadrature, the exponential distribution matches with the Gauss-Laguerre quadrature, and the uniform distribution matches with the Gauss-Legendre quadrature formula.

64

Structural design optimization considering uncertainties

2.3 Em pi ri c al d is t r ib ut io n s y s t ems Once the four statistical moments of g(X) are obtained, the PDF of g(X) can be approximated by the empirical distribution systems, such as Johnson system and Pearson system (Johnson et al. 1995, Hahn & Shapiro 1967). In our approach, the Pearson system of distribution is adopted, which approximates the PDF of a random variable X as a solution of a differential equation which follows: 1 df (x) x+a =− f (x) dx c0 + c1 x + c2 x2

(19)

where f (x) is the PDF to be found, x is x − µ and c0 , c1 , c2 and a are the coefficients determined from the four statistical moments of X. The relation is given as, c0 = (4β2 − 3β1 )(10β2 − 12β1 − 18)−1 σ  c1 = a = β1 (β2 + 3)(10β2 − 12β1 − 18)−1 σ 2 c2 = (2β2 − 3β1 − 6)(10β2 − 12β1 − 18)−1

(20)

The shape of f (x) changes considerably with the characteristics of roots of the following equation: c0 + c1 x + c2 x2 = 0

(21)

Pearson classified the types of distribution into seven groups according to the types of roots of Equation 21. It is summarized in Table 3.1. It is notable that the type of a distribution is determined solely by the skewness and kurtosis. As a special case, β1 = 0, β2 = 3 corresponds to the normal distribution. The Pearson system is a convenient tool that enables to find a PDF from the first four statistical moments of a random variable, however, one thing that should be noted is that the PDF found by the Pearson system is not the unique solution corresponding to the moments. It is important with this reason, to understand the mathematical background and assumptions in the Pearson system: For example, it can represent Table 3.1 Pearson system of distributions (classifications and corresponding distributions). Types Type I: Type II: Type III: Type IV: Type V: Type VI: Type VII:

Case β1 (β2 + 3)2 <0 4(2β2 − 3β1 − 6)(4β2 − 3β1 ) β1 = 0,β2 < 3 2β2 − 3β1 − 6 = 0 0<κ<1 κ=1 κ>1 β1 = 0,β2 > 3 κ=

Distributions Beta Beta (Symmetric) Gamma No match Inverse Gaussian Beta prime Student’s t

Reliability analysis and optimization using moment methods

65

PDF which has only one mode. More detailed explanations about the Pearson system can be found in (Johnson et al. 1995). 2.4

Proc edure of reliability analys is us i ng f ul l f ac to ri al mo me nt method

The procedure for reliability analysis using the full factorial DOE and the Pearson system is summarized in Figure 3.3. This procedure calculating the probability of failure using the full factorial experimental set is named in our work as full factorial moment method (FFMM). 2.5

Examples

Two examples are provided to demonstrate the accuracy of FFMM in statistical moment estimation and reliability calculation. The first example is a simple linear polynomial function (Eq. 22) whose statistical moments can be calculated analytically. FFMM is applied with several different input distributions and compared with exact solution to verify its accuracy. The input setting of the problem is summarized in Table 3.2. g(x1 , x2 , x3 ) = 3x1 − 2x2 + 5x3

(22)

The 3n full factorial DOE has been applied and the results of moment estimation are summarized in Table 3.3. It is seen that FFMM provides very accurate results regardless of the types of input distributions. The second example of FFMM is an application to the overrunning clutch assembly known as Fortini’s clutch (Fig. 3.4). This problem has been discussed by several authors including (Greenwood & Chase 1990, Creveling 1997) and so on. The contact angle

Given input distribution Calculate first 2m 1 moments of Xi

mX , sX , b1Xi , b2Xi , m5.Xi ,…, m2m1.Xi i

i

Find levels and weights: section 2.2 {li .1,…,li .m} and {wi .1,…,wi .m}

Run full factorial DOE: Eqs. (5)~(8) mg , sg , b1g , b2g

Pearson system: Eqs. (19), (20) fg(X) (g(x)) and Pr[ g(X)  0]

Figure 3.3 Overall procedure of full factorial moment method.

66

Structural design optimization considering uncertainties

Table 3.2 Four different settings of input distributions and their parameters. Case

(a) (b) (c) (d)

Parameters of distribution∗

Distribution

Exponential Gamma Lognormal Mixed

x1

x2

x3

λ1 = 1.0 η1 = 1.0, λ1 = 1.5 µ ˆ 1 = 0.1, σˆ 1 = 0.1 λ1 = 4.0 (exponential)

λ2 = 2.0 η2 = 3.0, λ2 = 5.0 µ ˆ 2 = 0.5, σˆ 2 = 0.1 µ ˆ 2 = 0.5, σˆ 2 = 0.1 (lognormal)

λ3 = 3.0 η3 = 3.0, λ3 = 1.0 µ ˆ 3 = 1.0, σˆ 3 = 0.1 a3 = 4.0, b3 = 1.0 (Weibull)

∗ The notations of the distribution parameters are from (Hahn & Shapiro 1967).

Table 3.3 Results of moment estimation using FFMM for linear performance function.  Case Method µg σg β1g β2g (a) (b) (c) (d)

Exact Proposed Exact Proposed Exact Proposed Exact Proposed

3.6667 3.6667 15.800 15.800 13.678 13.678 1.9680 1.9680

3.5746 3.5746 8.9152 8.9152 1.4482 1.4482 1.5131 1.5131

1.3412 1.3412 1.0805 1.0805 0.25520 0.25520 0.18862 0.18862

6.2969 6.2969 4.7962 4.7962 3.1306 3.1306 3.2369 3.2369

µ5g 1.3944e4 1.3944e4 8.3428e5 8.3428e5 16.871 16.871 21.712 21.712

y Cage

Hub x2

x4

x1

x3 Roller bearing

Figure 3.4 The overrunning clutch assembly (Fortini’s clutch).

Y is given in terms of the independent component variables, X1 , X2 , X3 and X4 as follows:

 X1 + 0.5(X2 + X3 ) Y = arccos (23) X4 − 0.5(X2 + X3 ) The design requirement for this mechanism is that the contact angle Y must lie between 0.087264 radian (5◦ ) and 0.157075 radian (9◦ ). The distribution types and parameters of random variables are listed in Table 3.4.

Reliability analysis and optimization using moment methods

67

Table 3.4 Distribution types and parameters of input random variables in clutch example. Component

Distribution

Mean

Standard deviation

Parameters for non-normal variables

X1 X2 X3 X4

Beta Normal Normal Rayleigh

55.29 mm 22.86 mm 22.86 mm 101.60 mm

0.0793 mm 0.0043 mm 0.0043 mm 0.0793 mm

γ1 = η1 = 5.0 (55.0269 ≤ x1 ≤ 55.5531) σˆ 4 = 0.1211 (x4 ≥ 101.45)

Table 3.5 Results of moment estimation and probability calculation for Fortini’s clutch example. Moment

3n (S&K ∗ )

3n (MM∗∗ )

5n (MM∗∗ )

FORM

MCS

Mean STD Skewness Kurtosis Pr [y < 5◦ ] Pr [y < 6◦ ] Pr [y < 7◦ ] Pr [y < 8◦ ] Pr [y < 9◦ ] Pr [5◦ < y < 9◦ ] Function call

0.1219 0.0117 −0.0578 2.9216 0.00159 0.07266 0.50430 0.93617 0.99925 0.99767 81

0.1219 0.0117 −0.0497 2.8488 0.00124 0.07288 0.50467 0.93570 0.99943 0.99819 81

0.1219 0.0117 −0.0530 2.8827 0.00140 0.07272 0.50452 0.93595 0.99934 0.99794 625

• • • • diverge 0.08777 (40) 0.52037 (25) 0.93564 (15) 0.99921 (15) • No. in ( )

0.1219 0.0117 −0.0523 2.8822 0.00122 0.07388 0.50265 0.93666 0.99926 0.99804 1,000 k

∗ Seo & Kwak’s three level formula (Eqs. 15, 16) ∗∗ Levels and weights from Equation 4.

For comparison, FORM with HL-RF algorithm (Hasofer & Lind 1974, Rackwitz & Fiesseler 1978) and Monte Carlo simulation (MCS) are applied together with FFMM to calculate the probability of the contact angle being outside the allowable range. And to check the difference in the accuracy made by the selection of levels and weights, the Seo & Kwak’s three level DOE (Eqs. 15 and 16) is also tried. The results are summarized in Table 3.5. In this example, FORM has some numerical difficulty in treating the non-normal distributions when the probability is small. During the calculation of the probability Pr(y < 5◦ ), the MPFP search point goes far outside the domain of the non-normal random variables and this results in the divergence of HL-RF algorithm. On the contrary, FFMM shows good accuracy throughout the range of y values. It is seen that there is a subtle difference in the results by Seo & Kwak’s 3 level DOE and DOE obtained by directly solving the moment matching equation (Eq. 4). This difference becomes more significant as the system performance function becomes more nonlinear. The PDF found by the Pearson system is plotted in Figure 3.5 with the histogram obtained by MCS. It is shown that the Pearson system gives very accurate PDF estimation in this problem. The number of function calls required for calculating the probability is also listed in Table 3.5. The number of function evaluations in FFMM seems significantly bigger than that of FORM, and it increases very rapidly as the number of random variables increases. This becomes the weak point of FFMM, which hinders its application to more practical engineering problems.

68

Structural design optimization considering uncertainties

PDF by MCS PDF by Pearson system

0.7 0.6 0.5 0.4 f (y)

Failure region

0.3

Failure region

0.2 0.1 0.0 3

4

5

6

7

8

9

10

y (degree)

Figure 3.5 Probability density function of contact angle y.

3 Response surface augmented moment method The FFMM introduced in section 2 provides good accuracy and ability to treat nonnormal distributions as shown in section 2.5. However, the number of function evaluations required by FFMM increases exponentially with the number of random variables and it is often prohibitive for applications to practical engineering problems where the evaluation of system performance function requires considerable time and computational resources. To tackle this problem, a novel way to integrate the response surface approximation with the 3n FFMM was developed and named as response surface augmented moment method (RSMM) (Lee & Kwak 2006). In RSMM, instead of performing expensive full factorial experiments, experiments are selectively performed at the points with bigger weight and the rest of the data are approximated by a second order response surface. This response surface is updated with addition of experiments one by one until a convergence in the probability of failure is achieved. In this section, the overall procedure of RSMM is introduced and some important concepts utilized in the method are explained together with examples of reliability analysis. 3.1

Ov era l l pr o c e d ur e

Two strategies are taken in developing RSMM. Firstly, to reduce the number of function evaluations, the experimental data are used not only for moment estimation but also for function approximation. Experiments important in probability calculation are selectively performed and the rest of data in the set of full factorial design are approximated using a response surface. Secondly, the initially simple response surface is updated progressively with the addition of experiments by introducing new cross product terms into the approximation model. The overall procedure of RSMM is as follows: (a)

Establish 3n full factorial DOE with levels and weights obtained by solving Equation 4.

Reliability analysis and optimization using moment methods

69

x2.2 1 36

1 9

1 36

1 9

4 9

1 9

x1.2

1 36

1 9

1 36

Figure 3.6 Experiment layout of RSMM at the initial approximation stage.

(b)

Calculate performance function g(x) at the 2n + 1 experimental points located on the axis of mid-level. Usually, the weight on the mid-level is much larger than the rest. Figure 3.6 is the example of a 2 normal variable case. The numbers in the figure are the weights imposed on the experimental points calculated by, wi = w1i · w2i · · · · · wni =

n 

wji

(24)

j=1

(c)

where wi means the overall weight imposed on the i-th experimental point and wji is the weight of the j-th variable at the i-th experimental point. The circled experimental points in the figure are those at which experiments for the initial approximation are performed. With the 2n + 1 data obtained in step (b), build a quadratic response surface using the least square estimation (Myers & Montgomery 1995) without cross product terms as, g(x) ˜ =a+

n  i=1

(d)

(e)

bi xi +

n 

ci x2i

(25)

i=1

Using g(x) ˜ complement data at points where experiments are not performed and calculate the first four statistical moments of g(x) with Equations 5–8. Then obtain the probability of failure Pr [g(x) < 0] using the Pearson system just as done in FFMM. Calculate the influence index at the points where experiments are not performed. The influence index, κi , at the i-th experimental point is defined as follows:    dPf   (26) κi =  d g(x ˜ i) 

70

(f) (g)

Structural design optimization considering uncertainties

where xi is the vector x at the i-th experimental point. The influence index is a measure devised to figure out the relative importance of experiments in calculating the probability. Detailed explanations about the influence index are given in the subsequent section. Perform one additional experiment at the point with the biggest κi . With the data obtained in step (f), update g(x). ˜ A new cross product term may be added into g(x) ˜ as in Equation 27, g(x) ˜ =a+

n 

bi xi +

i=1

(h) (i)

3.2

n  i=1

ci x2i +

nmix 

dk xi(k) xj(k)

(27)

k=1

where nmix is the number of cross product terms included in the formulation and i(k) and j(k) are the indices of the first and second variable in the k-th cross product term respectively where i(k) < j(k). The way of updating response surface approximation is discussed in the subsequent section. With the updated g(x), ˜ calculate the probability of failure as in step (d). Repeat the steps from (e) to (h) until the value of the probability of failure converges. Inf l uen c e ind ex

RSMM tries to find a reliable solution by taking the converged value of probability as its final solution after sufficient updates of the approximation with successive additions of experiments. One important procedure in RSMM is then the arrangement of the order of experiment addition. It is obvious that to reduce the number of experiments effectively a higher priority should be given to the experimental point which can bring in the greatest change of probability when the approximation at that point is replaced with the real experimental data. Influence index is devised to compare the magnitudes of the expected change of probability at the points where experiments are not performed. The change of probability can be roughly estimated as follows:

Pf

dPf · (g(xi ) − g(x ˜ i )) d g(x ˜ i)

(28)

where the term ˜ i ) means the approximation error at the i-th experimental  g(xi ) − g(x ˜ i ) is the derivative of Pf with respect to g(x) ˜ at the i-th experimental point and dPf d g(x point, which implies the importance of the point in the calculation of Pf in the cur˜ i ) cannot be estimated unless the value rently approximated system. Since g(xi ) − g(x of g(xi ) is calculated with an additional experiment,  the influence index, κi , is defined ˜ i ). This derivative denotes the as the absolute value of the coefficient term, dPf d g(x ˜ at xi . sensitivity of Pf due to the change of the estimated value g(x) The influence index κi can be calculated very effectively without additional g(x) evaluations. Pf is a function of four statistical moments of g(X) and can be expressed as follows:  Pf = Pf (µg , σg , β1g , β2g ) (29)

Reliability analysis and optimization using moment methods

Equations 5–8 about µg , σg , 

71

 β1g , β2g can be rewritten as follows:

n exp

µg

wi g(x ˆ i)

(30)

i=1

σg

n exp 

 12 w (g(x ˆ i ) − µg ) i

2

(31)

i=1 n exp

 β1g

n exp

β2g

wi (g(x ˆ i ) − µ g )3

i=1

σg3

(32)

wi (g(x ˆ i ) − µ g )4

i=1

σg4

(33)

where n exp is the number of total experimental points, which is equal to 3n and wi is the weight imposed on the i-th experimental point calculated by Equation 24. g(x ˆ i ) is defined in RSMM as follows:  g(xi ) if experiment at xi is performed g(x ˆ i) = (34) g(x ˜ i ) if experiment at xi is not performed yet The derivative of Pf can be written as follows:  d β1g dPf ∂Pf ∂Pf ∂Pf ∂Pf dµg dσg dβ2g = + +  + · · · · d g(x ˜ i) ∂µg d g(x ˜ i ) ∂σg d g(x ˜ i ) ∂ β1g d g(x ˜ i) ∂β2g d g(x ˜ i)  d β1g

Pf

Pf

Pf

Pf dµg dσg dβ2g + +  + (35) · · · ·

µg d g(x ˜ i ) σg d g(x ˜ i ) β1g d g(x ˜ i)

β2g d g(x ˜ i)  The terms, Pf / µg , Pf / σg , Pf / β1g and Pf / β2g can be calculated using the finite difference method from the Pearson system program and the rest of the terms can be obtained by differentiating Equations 30–33 as follows: dµg = wi d g(x ˜ i) dσg wi = (g(x ˜ i ) − µg ) d g(x ˜ i) σg 

 d β1g dσg  3wi 3 wi + = 3 · (g(x · β1g ˜ i ) − µg )2 − d g(x ˜ i) σg σg d g(x ˜ i)

  dβ2g dσg 4 4wi wi β1g + ˜ i ) − µg )3 − = 4 · (g(x · β2g σg σg d g(x ˜ i) d g(x ˜ i)

(36) (37) (38) (39)

72

Structural design optimization considering uncertainties

3.3 U pd ate o f r es po ns e s ur fac e appr ox i m a t i o n Once an additional experiment is performed, we can update the response surface with the newly obtained data. In RSMM, not only the regression coefficients but also the regression bases are updated to improve the accuracy of the approximation for the given set of observations. At the initial stage of RSMM, a total of 2n + 1 polynomial terms are used in the response surface of up to quadratic terms. The cross product terms are introduced into the approximation model during the update so that we can account for the interactions among the random variables. As we may use at most the same number of terms in the approximation model with the number of observations and experiments are added one by one, one cross product term may be added in one updating step. To select the appropriate cross product term to be added, a simple procedure is established as follows. Let the coordinate of the experimental point where experiment has been newly performed denoted by xN = {x1·N , x2·N , . . . , xn·N } where xi·N is the value of variable xi at xN which takes a value among {li·1 , li·2 , li·3 }. Suppose the response surface before update is given as,

g(x) ˜ = a+

n 

bi xi +

i=1

= a+

n 

n 

ci x2i +

i=1

bi ξi +

i=1

n  i=1

nmix 

dk xi(k) xj(k)

k=1

ci ξi2 +

nmix 

d k ξi(k) ξj(k)

(40)

k=1

where ξi is the coded variable defined as follows: ξi =

xi − (li·1 + li·3 )/2 (li·3 − li·1 )/2

(41)

and a, bi , ci and d k are the regression coefficients when g(x) ˜ is expressed in the coded variables. (a) (b)

(c)

List up the variables which do not have mid-level at xN , say xi·N = li·2 . Among all possible combinations with the variables listed in step (a), select the cross product terms that are not included in the former approximation, Equation 40 and add them into C, the set of candidate cross product terms. At the initial stage of RSMM, C is a null set. If C is null, that is, all addible cross product terms are already used in the former approximation, no term is added at this updating step. If C has only one element, it is added to the regression bases. If C has more than one element, build g(x) ˜ including each cross product term in C and compare the residual sum of squares of each model. If there is a cross product term which makes the regression matrix singular, it is discarded. Choose the model corresponding to the minimum residual sum of squares. If all the RS models have the same residual sum of squares, then go to step (d). Else, go to step (e).

Reliability analysis and optimization using moment methods

(d)

Calculate coefficient sum CSij defined as Equation 42 for all members of C, CSij = |bi | + |ci | + |bj | + |cj |

(e)

73

(42)

where bi and ci are the coefficients in Equation 40. Choose xi xj that has the greatest value of CSij and add it to the regression bases. If a term is added, then remove it from the candidate set C and go to the next stage of RSMM.

The reason why cross product terms which consist only of variables off the mean-axis at currently or formerly executed experiments are added is to prevent the singularity or ill-conditioning during the least square estimation. With this adaptive update of response surface approximation, the number of terms in the response surface model is kept as small as possible while including the interaction effect that might exist among variables.

3.4

Examples

Two examples are taken to check accuracy and efficiency of RSMM. For comparison, the results by FFMM, MCS, FORM and method proposed by (Zhao & Ono 2001) are provided. In FORM, the HL-RF algorithm (Rackwitz & Fiessler 1978) is used to find MPFP. The method of Zhao & Ono is a recently reported moment based reliability method which uses samples solely on the mean axis of each variable while in RSMM non-axial samples are also utilized. The first example is taken from (Masen et al. 1986, Kiureghian et al. 1987 and Zhao & Ono 2001). The performance function representing the failure in one plastic collapse mechanism of a one bay frame is given as, g(x) = x1 + 2x2 + 2x3 + x4 − 5x5 − 5x6

(43)

where the variables are statistically independent and log-normally distributed with the means µ1 = . . . µ4 = 120, µ5 = 50, µ6 = 40 and standard deviations σ1 = . . . σ4 = 12, σ5 = 15, σ6 = 12. The probabilities Pr [g(x) < 0] calculated are summarized in Table 3.6. It is observed that the RSMM and FFMM give equally accurate results for present example while

Table 3.6 Results of moment estimation and reliability analysis of the first example.

Mean STD Skewness Kurtosis Pr[g(x) < 0] Function calls

FORM

Zhao & Ono

FFMM

RSMM

MCS

• • • • 9.430e-3 50

270.000 103.271 −0.528 3.650 1.219e-2 42

270.000 103.271 −0.523 3.612 1.212e-2 729

270.000 103.271 −0.523 3.612 1.212e-2 16

269.990 103.174 −0.530 3.623 1.213e-2 1,000 k

74

Structural design optimization considering uncertainties

P1

P2

P3

P4

P5

P6

E1, A1

2m E2, A2 E1, A1

PNT1

E2, A2

DISP1 6@4 m

Figure 3.7 Truss structure with 23 members.

Table 3.7 Input random variables for truss example. Variables

(unit)

Distribution type

Mean

Standard deviation

1 2 3 4 5 6 7 8 9 10

E1 (N/m2 ) E2 (N/m2 ) A1 (m2 ) A2 (m2 ) P1 (N) P2 (N) P3 (N) P4 (N) P5 (N) P6 (N)

lognormal lognormal lognormal lognormal gumbel gumbel gumbel gumbel gumbel gumbel

2.1 × 1011 2.1 × 1011 2.0 × 10−3 1.0 × 10−3 5.0 × 104 5.0 × 104 5.0 × 104 5.0 × 104 5.0 × 104 5.0 × 104

2.1 × 1010 2.1 × 1010 2.0 × 10−4 1.0 × 10−4 7.5 × 103 7.5 × 103 7.5 × 103 7.5 × 103 7.5 × 103 7.5 × 103

FORM shows rather erroneous result due to the non-normality of variables. Zhao & Ono’s method also gives good results for the present example. Since the performance function is approximated exactly with the quadratic response surface without cross product terms, convergence is achieved just after the initial approximation in RSMM. The second example is a truss structure with 23 members, as shown in Figure 3.7. Ten random variables are considered which are summarized in Table 3.7. It is assumed that all the horizontal members have perfectly correlated Young’s modulus and cross sectional areas with each other and so is the case with the diagonal members. The requirement of this problem is that the displacement at PNT1 in Figure 3.7 should not exceed 0.11 m. The performance function is defined as, g(x) = 0.11 − DISP1

(44)

The displacement at PNT1 is calculated using the commercial software, ANSYS 6.0.

Reliability analysis and optimization using moment methods

75

0.009

Pf

0.008 0.007 0.006 0.005 0.004 25

30 35 40 Number of function call

45

Figure 3.8 Convergence history of probability of failure in truss example.

RSMM finds a solution with 24 additional experiments after initial approximation with 21 experiments. At the final stage, the response surface model of g(x) is obtained with 17 cross product terms as follows: g(ξ(x)) ˜ = 2.8070 + 1.2598ξ1 + 0.2147ξ2 + 1.2559ξ3 + 0.2133ξ4 − 0.1510ξ5 − 0.4238ξ6 − 0.6100ξ7 − 0.6100ξ8 − 0.4238ξ9 − 0.1510ξ10 − 0.1978ξ12 − 0.0362ξ22 − 0.2016ξ32 − 0.0346ξ42 + 0.0023ξ52 2 + 0.0008ξ62 + 0.0036ξ72 + 0.0036ξ82 + 0.0008ξ92 + 0.0023ξ10

− 0.0042ξ1 ξ2 − 0.3022ξ1 ξ3 − 0.0110ξ1 ξ4 + 0.0381ξ1 ξ5 + 0.0871ξ1 ξ6 + 0.1232ξ1 ξ7 + 0.1232ξ1 ξ8 + 0.0871ξ1 ξ9 + 0.0346ξ1 ξ10 + 0.0041ξ2 ξ3 + 0.0110ξ3 ξ4 + 0.0261ξ3 ξ5 + 0.0831ξ3 ξ6 + 0.1172ξ3 ξ7 + 0.1172ξ3 ξ8 + 0.0832ξ3 ξ9 + 0.0296ξ3 ξ10

(45)

The history of probability is depicted in Figure 3.8. At the first approximation, the probability is calculated as 0.00451 and the converged value is 0.00880. The finally obtained distribution of g(x) is Pearson type I, a beta distribution as in Figure 3.9. The results are compared in Table 3.8 with the other methods. As in the previous examples, the result of RSMM shows good agreement with those of FFMM and MCS with 100,000 samples but the results of FORM and Zhao & Ono’s method have some discrepancy with the other results. When checking the numerical efficiency, RSMM shows very good performance compared to the other methods. FORM and Zhao & Ono’s method also show good numerical efficiency but the accuracies of those methods are not satisfactory for the present example. The total elapsed time for RSMM is 90 seconds with a Pentium 4 machine while FORM spends 151 seconds.

76

Structural design optimization considering uncertainties

PDF by MCS PDF by Pearson system

0.40

Probability density function

0.35 0.30 0.25

Failure region

0.20 0.15 0.10 0.05 0.00 2

0

2

4

6

g(x)

Figure 3.9 Probability density function of g(x) in truss example. Table 3.8 Results of moment estimation and reliability analysis of truss example. FORM

Mean STD Skewness Kurtosis Pr[g(x) < 0] Function calls

Zhao & Ono’s

• • • • 5.019 × 10−3 77

5n

7n

0.0306 0.0111 −0.4989 3.4289 4.356 × 10−3 50

0.0307 0.0110 −0.5708 3.4120 4.357 × 10−3 70

FFMM

RSMM

MCS

0.0306 0.0109 −0.2009 3.0786 8.821 × 10−3 59,049

0.0306 0.0109 −0.2008 3.0785 8.804 × 10−3 45

0.0307 0.0111 −0.4789 3.3826 8.330 × 10−3 100,000

4 Reliability-based design optimization using moment methods The reliability-based design optimization (RBDO) can generally be formulated as follows: Minimize W(d, z) subject to Pr [g  im(d, z, x) < 0] ≤ pi i = 1, . . . , m  {gi (d, z, x) < 0} ≤ p0 Pr

(46)

i=1

where W and gi are the objective function and limit state function and d, z, x are the vector of design variables, state variables and random variables respectively. The design

Reliability analysis and optimization using moment methods

77

Initial design, Distribution of random variables

Optimization engine (SQP, MMFD, etc) EvaluateW, Pr[gi < 0]

Evaluate

dW dP f , dd dd

N

RSMM or FFMM

Converge? Feasible? Y STOP

Figure 3.10 Flowchart of RBDO using RSMM or FFMM.

variable can be a deterministic variable or a mean value of a random variable existing in the system. The first constraint is imposed on the component failure probability and the second one is imposed on the system failure probability which can be evaluated by the reliability bounds concept (Ditlevsen 1979). The formulation in Equation 46 can be applied also to the tolerance synthesis problems (Creveling 1997, Lee & Woo 1990), and in this case, the tolerance, which is usually defined as some multiple of standard deviation of a dimension becomes a design variable. In this section, a RBDO procedure using FFMM and RSMM is introduced with an approximate design sensitivity analysis for probabilistic constraints. 4.1

Proc edure of RBDO us ing moment me tho ds

FFMM and RSMM can be combined with a mathematical programming for RBDO. The overall procedure is depicted in Figure 3.10. The optimization engine calls RSMM or FFMM whenever it needs to evaluate the probabilistic constraints. The procedure is close to the double loop strategy where constraint feasibility is checked at every design point during the optimization. The efficiency of the procedure will be discussed in section 5. 4.2

Design sens itivity analys is

Since one evaluation of a probabilistic constraint usually takes considerable number of performance function evaluations, it is important to provide the design sensitivity of the probabilistic constraint in an analytic or semi-analytic way during the RBDO process. In this section, a semi-analytic design sensitivity analysis with moment method is provided (Lee & Kwak 2005). Both RSMM and FFMM can be adopted in this procedure and the calculation is done without any additional g(x) evaluations. Instead,

78

Structural design optimization considering uncertainties

the experimental data or response surface model previously obtained in the reliability analysis is utilized. The procedure is as follows.  Since the probability of failure is a function of four statistical moment, µg , σg , β1g and β2g , the design sensitivity of Pf can be written as follows:  d β1g ∂Pf dµg ∂Pf dσg ∂Pf ∂Pf dβ2g dPf = + +  + · · · · dd ∂µg dd ∂σg dd dd ∂β dd ∂ β1g 2g  d β1g

Pf dσg

Pf

Pf dβ2g

Pf dµg + +  + · · · ·

µg dd

σg dd dd

β2g dd

β1g

(47)

 The terms Pf / µg , Pf / σg , Pf / β1g and Pf / β2g can be calculated using the finite difference method with the Pearson system program and the rest of terms can be obtained from Equations 5–8, as follows: 3 3  dµg ∂µg dlk·ik  ∂µg dwk·ik = + ddk ∂lk,ik ddk ∂wk,ik ddk

(48)

3 3  dσg ∂σg dlk·ik  ∂σg dwk·ik = + ddk ∂lk,ik ddk ∂wk,ik ddk

(49)

ik =1

ik =1

ik =1

ik =1

   3 3  ∂ β1g dlk·ik  ∂ β1g dwk·ik d β1g = + ddk ∂lk,ik ddk ∂wk,ik ddk ik =1

(50)

ik =1

3 3  dβ2g ∂β2g dlk·ik  ∂β2g dwk·ik = + ddk ∂lk,ik ddk ∂wk,ik ddk ik =1

(51)

ik =1

where dk is the design variable related with the k-th random variable. The partial derivatives in Equations 48–51 can be calculated by directly differentiating Equations 5–8 as follows: 3 3 3 3     ∂µg ∂g(l1·i1 , . . . , ln·in ) = w1·i1 · · · wk−1·ik−1 wk+1·ik+1 · · · wn·in · wk·ik ∂lk·ik ∂lk·ik i1 =1

ik−1 =1

3 

3 

ik+1 =1

in =1

(52) ∂µg = ∂wk·ik

i1 =1

w1·i1 · · ·

ik−1 =1

wk−1·ik−1

3  ik+1 =1

wk+1·ik+1 · · ·

3 

" ! wn·in g l1·i1 , . . . , ln·in (53)

in =1

For the lack ofspace, the derivation only for the mean is presented. The partial derivatives of σg , β1g , β2g with respect to lk·ik and wk·ik can be obtained in the same way. It is noted that for calculating the partial derivatives of moments with respect to lk·ik , we have to calculate the partial derivative of performance function g with respect to lk·ik . When FFMM is used, it is calculated using a backward or forward difference

Reliability analysis and optimization using moment methods

79

schemes on the previously obtained experimental data. For a three level case, it can be formulated as follows: ∂g(l1·i1 , . . . , ln·in ) ∼ g(l1·i1 , . . . , lk·3 , . . . , ln·in )h1 g(l1·i1 , . . . , lk·2 , . . . , ln·in )(h1 + h2 ) + =− ∂lk·1 h2 (h1 + h2 ) h1 h2 −

g(l1·i1 , . . . , lk·1 , . . . , ln·in )(2h1 + h2 ) h1 (h1 + h2 )

(54)

∂g(l1·i1 , . . . , ln·in ) ∼ g(l1·i1 , . . . , lk·3 , . . . , ln·in )h1 g(l1·i1 , . . . , lk·2 , . . . , ln·in )(h2 − h1 ) + = ∂lk·2 h2 (h1 + h2 ) h1 h2 −

g(l1·i1 , . . . , lk·1 , . . . , ln·in )h2 h1 (h1 + h2 )

(55)

∂g(l1·i1 , . . . , ln·in ) ∼ g(l1·i1 , . . . , lk·3 , . . . , ln·in )(h1 + 2h2 ) = ∂lk·3 h2 (h1 + h2 ) −

g(l1·i1 , . . . , lk·2 , . . . , ln·in )(h1 + h2 ) g(l1·i1 , . . . , lk·1 , . . . , ln·in )h2 + h1 h2 h1 (h1 + h2 ) (56)

where h1 and h2 are lk·2 − lk·1 and lk·3 − lk·2 respectively. This finite difference scheme can be extended to cases where more than 3 levels are used in DOE. When RSMM is used, the partial derivative can be calculated using the previously obtained response surface model g˜ as follows: ⎧ ⎪ 0 if ik = m m = 1, 2, 3   ∂g(l1,i1 , . . . , ln,in ) ⎨ ∂g(x) ˜  ∂g(x)  = (57) ⎪ ∂lk,m  ⎩ ∂x  ∂xk x=(l1,i ,...,lk,m ,...,ln,in ) k x=(l1,i ,...,lk,m ,...,ln,in ) 1

1

The derivatives dlk·ik /ddk , dwk·ik /ddk in Equations 48–51 can be obtained from the relationship between lk·ik , wk·ik and dk . Since lk·ik , and wk·ik are determined from the four statistical moments of xk (Eq. 4), the derivatives can be written as follows:  dlk·ik ∂lk·ik dµxk ∂lk·ik dσxk ∂lk·ik d β1xk ∂lk·ik dβ2xk = + +  + ddk ∂µxk ddk ∂σxk ddk ∂β2xk ddk ∂ β1xk ddk  ∂wk·ik dµxk ∂wk·ik dσxk ∂wk·ik d β1xk ∂wk·ik dβ2xk dwk·ik = + +  + ddk ∂µxk ddk ∂σxk ddk ∂β2xk ddk ∂ β1xk ddk

(58) (59)

When we use the explicit formula of levels and weights derived by Seo & Kwak (2002), the partial derivatives in Equations 58 and 59 can be obtained by directly differentiating the formula. If we solve Equation 4 numerically in order to obtain lk·ik and wk·ik , then they can be calculated with finite difference method.  The derivatives, dµxk /ddk , dσxk /ddk , d β1xk /ddk and dβ2xk /ddk , are determined from the definition of the optimization problem. In RBDO, the design variable dk becomes the mean value of xk , that is, µxk , and it is usually assumed that the distribution characteristics are not changed during the optimization, so dµxk /ddk becomes 1

80

Structural design optimization considering uncertainties

and the others become 0. In tolerance synthesis problem, dk becomes the tolerance of dimension xk , which is 3σk for the usual definition of tolerance, and the other moments are assumed invariant, so dσxk /ddk become 1/3 with the other derivatives become 0. With these, all the derivatives and partial derivatives necessary to calculate Equations 48–51 have been derived and the design sensitivity of failure probability can be calculated from Equation 47. The whole procedure described in this section seems a little bit tedious and complex, but it is easy to implement and the computational efficiency is also good. 4.3

Ex am pl es

In this section, we present two examples of RBDO performed by RSMM and the design sensitivity analysis explained in the earlier section. The first example is a mathematical problem introduced in (Xiao et al. 1999). The problem is stated as follows: Minimize subject to

W(d) ≡ πd12 + d2 Pf 1 ≡ Pr [g1 ≡ X13 X2 − 95.5 ≤ 0] ≤ 0.0010 Pf 2 ≡ Pr [g2 ≡ X12 X2 − 70.7 ≤ 0] ≤ 0.0010 1.0 ≤ d1 ≤ 2.0 20.0 ≤ d2 ≤ 50.0

(60)

where the design variables, d1 and d2 are the mean values of random variable X1 , X2 respectively and X1 and X2 follow normal distribution with standard deviation 0.1 and 3.0. Initial design is (1.5, 35.0) and the optimization is performed with the sequential quadratic programming (SQP). For comparison, the reliability index approach (RIA) is also applied and the results are summarized in Table 3.9. It is seen that the two methods converged to similar solutions with the second constraint activated. However, slight differences in the values of d2 and the objective function are noticed. When we apply the MCS to the final designs found by RSMM and RIA for verification, the actual value of Pf 2 is calculated as 0.001012 for RSMM Table 3.9 Optimization results of first example. RSMM

RIA

W

d1

d2

W

d1

d2

Initial Final

42.069 41.564

1.500 2.000

35.000 28.998

42.069 41.459

1.500 2.000

35.000 28.893

Probability

Initial

Final

Final by MCS∗

Initial

Final

Final by MCS∗

P f1 P f2

1.7917e-1 2.6086e-1

2.5110e-6 9.9975e-4

8.5000e-6 1.0120e-3

1.7097e-1 2.5203e-1

7.6580e-6 9.9958e-4

1.0000e-5 1.1060e-3

Function calls

W

g1

g2

W

g1

g2

27

342

342

49

1011

972

∗ Sample size: 1,000,000.

R e l i a b i l i t y a n a l y s i s a n d o p t i m i z a t i o n u s i n g m o m e n t m e t h o d s 81

and 0.001106 for RIA. The result of FORM contains rather larger error than that of RSMM and actually, RIA violates the constraint slightly more than RSMM. RSMM provides more accurate estimate of probability of failure than FORM and this is the reason why the two methods converge to different solutions. The numbers of function calls are also listed in the table. It should be noted here that during the optimization with RIA, the design sensitivity is calculated with finite difference method, so judging the efficiency of the method via the number of function calls may not appropriate in this case. However, even considering this fact, it is seen that the RBDO by RSMM shows good performance in terms of efficiency. The second example is the optimization of the truss structure introduced in section 3. In this example, a symmetry condition is imposed on the geometry and load condition with respect to the axis of symmetry (Fig. 3.11). Three new random variables, X1 , X2 , X3 are introduced which determine the shape of the truss structure so there are 10 random variables in total. X1 , X2 , X3 are distributed normally with standard deviation of 2 cm and the distribution parameters of the other variables are the same as in Table 3.7. The system requirement is that the displacement at the center point should not exceed 11.5 cm. The optimization problem is formulated as follows: Minimize Weight(d) Subject to Pf ≡ Pr [G(X) < 0] ≤ 0.05 where G(X) = 11.5 − |disp1|, 100 < di < 400

(61) i = 1, 2, 3

where di is the mean value of random variable Xi with initial value of 200 cm. We apply SQP to solve this problem and optimization with RIA is also tried for comparison. The results are summarized in Table 3.9. It is seen that RIA and RSMM find somewhat different solutions. RIA converges to a solution activating the constraint, but in the result of RSMM, the constraint is not activated. We perform MCS with 100,000 samples to verify the reliability analysis at the final design found by the two methods and it is listed in Table 3.10 as Pf _MCS . It is seen that greater error in the probability is involved in the result of RIA and the final design found by RIA actually violates the constraint by a relatively large amount. RSMM finds the solution with 903 function evaluations of G(x) and RIA spends 4004. As in the first example of this section, design sensitivity is calculated using finite difference method in case of RIA, so the direct comparison of the numerical efficiency

P1

P2

P3

P3

P2 E1, A1

E2, A2 x1

x2

x3 E2, A2

E1, A1 400

Disp1

Figure 3.11 Truss structure wit 23 members (RBDO example).

P1

82

Structural design optimization considering uncertainties

Table 3.10 Optimization result of truss structure with RSMM and RIA. RSMM

RIA

d

weight/w 0

Pf

P ∗f_MCS

d

weight/w 0

Pf

P ∗f_MCS

115.143 155.181 220.897

0.9806

0.0456

0.0465

110.142 166.210 204.371

0.9775

0.0500

0.0632

903 G(x) evaluations

4004 G(x) evaluations

∗ Sample size: 100,000.

via the number of function evaluations is not appropriate. However, even considering that fact, the numerical efficiency of RSMM seems satisfactory in this problem.

5 Conclusions So far, moment based reliability analysis methods, FFMM and RSMM have been introduced and illustrated by examples. The following strong points of moment methods could be figured out. Firstly, it does not involve the difficulties of searching for the most probable failure point (MPFP) as in FORM/SORM. Especially, the procedure of FFMM is very simple and easy to use. Secondly, not only the probability value but also the information of PDF and cumulative distribution function (CDF) of a system response function is made available, which can give a deeper insight about the statistical characteristics of the engineering system. Thirdly, they do not use any transformation to deal with non-normal distributions and is free from the deterioration of accuracy and efficiency, which is suffered by other existing methods. Meanwhile, the moment methods have certain drawbacks and limitations and those should also be well recognized and reminded before the applications. The information of MPFP is very important in calculating the small probability in tail region, but moment methods do not make use of it. Also the approximation using only finite number of moments has put some limitation in the accuracy. For this reason, moment methods are often considered more suitable for high probability problems. In most of our test cases, probability calculation using the Pearson system gives reliable results up to about 4 sigma level, corresponding to a probability of 10−5 order. However, at probability levels less than 10−5 , the failure probability found by the Pearson system might not be reliable. The non-uniqueness of the PDF mentioned in section 2.3 also should be carefully reminded. The accuracy of moment estimation is determined from the integration order provided by the method and the degree of nonlinearity of performance function in the high probability region that is defined in terms of coefficients of variation of variables. It is noted that large system nonlinearity and coefficient of variation of variables can degrade the accuracy of moment estimation. The accuracy of moment calculation can be extended by introducing more levels into DOE, which may cost much more computational cost. In RSMM, the probability converges to the value which is expected to be found by FFMM since all the samples are taken from the set of full factorial design. The convergence is expedited since important samples in probability are selectively taken in

R e l i a b i l i t y a n a l y s i s a n d o p t i m i z a t i o n u s i n g m o m e n t m e t h o d s 83

the early stage of RSMM, while the rest of samples are approximated with a response surface. The initial approximation is made using samples with high weights and additional samples are selected considering the impacts they will bring to the probability. Up to now, we covered examples with uni-modal distributions, and hence the levels of the highest weight are mid-levels. In the case of non-uni-modal shape, the selection of the initial set of experiments must be changed accordingly. One difficulty in RSMM is to determine stopping criterion. A tight criterion is necessary but there should be a compromise between accuracy and computational cost. This is a topic of more study in the future. Since the current version of RSMM is based on the 3n FFMM, in case more accurate DOE is necessary, an extension of the method should be made accordingly. A semi-analytic design sensitivity analysis is proposed in combination with FFMM and RSMM, which is shown to be robust and accurate through several tests. The proposed procedure of RBDO is applied successfully to simple RBDO problems and the efficiency of the procedure turns out to be comparable to or even better than conventional RIA. However, it should be noted that the RIA is a classical approach and these days, much more efficient algorithms and strategies are available in the field of RBDO, such as the performance measure approach (Lee & Kwak 1987–88, Tu et al. 1999, Youn & Choi 2003) and the single loop optimization strategy like sequential optimization and reliability assessment (SORA) method (Du & Chen 2003). The comparison with those methods are not provided in this chapter, however in our comparative study, it was recognized that the efficiency of the RBDO procedure proposed in this chapter cannot match that of a single loop method like SORA due to its double loop nature. It is not easy to adopt the single loop optimization strategy into RBDO with the moment methods. However, better accuracy of moment methods than the first order reliability approximation is still a strong point compared to the RBDO method based on FORM.

References Abramowitz, M. & Stegun, I.A. 1972. Handbook of mathematical functions. 10th ed. New York: Dover. Bjerager, P. 1988. Probability integration by directional simulation. ASCE Journal of Engineering Mechanics 114(8):1285–1302. Bjerager, P. 1991. Methods for structural reliability computation. In: F. Casciati (ed.), Reliability problems: general principles and applications in mechanics of solid and structures. New York: Springer Verlag. Breitung, K. 1984. Asymptotic approximation for multi-normal integrals. ASCE Journal of Engineering Mechanics 10(3):357–366. Bucher, C.G. 1988. Adaptive sampling: an iterative fast Monte-Carlo procedure. Structural Safety 5(2):119–126. Bucher, C.G. & Bourgund, U. 1990. A fast and efficient response surface approach for structural reliability problems. Structural Safety 7:57–66. Creveling, C.M. 1997. Tolerance design: A handbook for developing optimal specification. Cambridge, MA: Addison-Wesley. D’Errico, J.R. & Zaino, N.A. 1988. Statistical tolerancing using a modification of Taguchi’s method. Technometrics 30(4):397–405. Ditlevsen, O. 1979. Narrow reliability bounds for structural systems. Journal of Structural Mechanics 7:453–472.

84

Structural design optimization considering uncertainties

Du, X. & Chen, W. 2004. Sequential optimization and reliability assessment for efficient probabilistic design. ASME Journal of Mechanical Design 126(2):225–233. Enevoldsen, I. & Sørensen, J.D. 1994. Reliability-based optimization in structural Engineering. Structural Safety 15:169–196. Engels, H. 1980. Numerical quadrature and qubature. New York: Academic Press. Evans, D.H. 1972. An application of numerical integration techniques to statistical tolerancing, III – general distributions. Technometrics 14:23–35. Faravelli, L. 1989. Response surface approach for reliability analysis. ASCE Journal of Engineering Mechanics 115(12):2763–2781. Fiessler, B. et al. 1979. Quadratic limit states in structural reliability. ASCE Journal of Engineering Mechanics 105(4):661–676. Frangopol, D.M. & Corotis, R.B. 1996. Reliability-Based Structural System Optimization: State-of-the-Art versus State-of the-Practice. In F.Y. Cheng (ed.), Analysis and Computation: Proceedings of the Twelfth Conference held in Conjunction with Structures Congress XIV pp. 67–78. Greenwood, W.H. & Chase, K.W. 1990. Root sum squares tolerance analysis with nonlinear problems. ASME Journal of Engineering for Industry 112:382–384. Hahn, G.J. & Shapiro, S.S. 1967. Statistical models in engineering. John Wiley & Sons: New York. Hasofer, A.M. & Lind, N.C. 1974. Exact and invariant second order code format. ASCE Journal of Engineering Mechanics 100(1):111–121. Hohenbichler, M. & Rackwitz, R. 1981. Nonnormal dependent vectors in structural reliability. ASCE Journal of Engineering Mechanics 107:1127–1238. Hong, H.P. 1996. Point-estimate moment-based reliability analysis. Civil Engineering Systems 13(4):281–294. Huh, J.S. et al. 2006. Performance evaluation of precision nanopositioning devices caused by uncertainties due to tolerances using function approximation moment method. Review of Scientific Instrument 77:15–103. Johnson, N.L. et al. 1995. Continuous univariate distributions. New York: John Wiley & Sons. Kiureghian, A.D. et al. 1987. Second-order reliability approximations. ASCE Journal of Engineering Mechanics 113(8):1208–1225. Kiureghian, A.D. 1996. Structural reliability methods for seismic safety assessment: a review. Engineering Structures 95:412–24. Kiureghian, A.D. & Dakessian, T. 1998. Multiple design points in first and second-order reliability. Structural Safety 20:37–49. Koyluoglu, H.U. & Nielsen, S.R.K. 1994. New approximations for SORM integrals. Structural Safety 13:235–246. Lee, S.H. & Kwak, B.M. 2005. Reliability based design optimization using response surface augmented moment method. Proceedings of 6th World Congress on Structural and Multidisciplinary Optimization, Rio de Janeiro, Brazil. Lee, S.H. & Kwak, B.M. 2006. Response surface augmented moment method for efficient reliability analysis. Structural Safety 28:261–272. Lee, T.W. & Kwak, B.M. 1987–88. A reliability-based optimal design using advanced first order second moment method. Mechanics of Structures and Machines 15(4):523–542. Lee, W.J. & Woo, T.C. 1990. Tolerances: their analysis and synthesis. ASME Journal of Engineering for Industry 112:113–121. Li, K.S. & Lumb, P. 1985. Reliability analysis by numerical integration and curve fitting. Structural Safety 3:29–36. Madsen, H.O. et al. 1986. Methods of structural safety. Englewood Cliffs: Prentice-Hall. Melchers, R.E. 1989. Importance sampling in structural systems. Structural Safety 6(1):3–10.

R e l i a b i l i t y a n a l y s i s a n d o p t i m i z a t i o n u s i n g m o m e n t m e t h o d s 85 Moré, J.J. et al. 1980. User guide for MINPACK-1. Argonne National Labs Report ANL-80-74. Argonne. Illinois. Mori, Y. & Ellingwood, B.R. 1993. Time-dependent system reliability analysis by adaptive importance sampling. Structural Safety 12(1):59–73. Myers, R.H. & Montgomery, D.C. 1995. Response surface methodology: process and product optimization using designed experiments. New York: John-Wiley & Sons. Nie, J. & Ellingwood, B.R. 2005. Finite element-based structural reliability assessment using efficient directional simulation. ASCE Journal of Engineering Mechanics 131(3):259–267. Rackwitz, R. & Fiessler, B. 1978. Structural reliability under combined random load sequences. Computers and Structures 9:489–494. Rajashekhar, M.R. & Ellingwood, B.R. 1993. A new look at the response surface approach for reliability analysis. Structural Safety 12:205–220. Rosenblueth, E. 1981. Two-point estimate in probabilities. Applied Mathematical Modeling 5:329–335. Seo, H.S. & Kwak, B.M. 2002. Efficient statistical tolerance analysis for general distributions using three-point information. International Journal of Production Research 40(4):931–44. Taguchi, G. 1978. Performance analysis design. International Journal of Production Research 16:521–530. Tu, J. et al. 1999. A new study on reliability-based design optimization. ASME Journal of Mechanical Design 121(4):557–564. Xiao, Q. et al. 1999. Computational strategies for reliability based Multidisciplinary optimization, Proceedings of the 13th ASCE EMD Conference. Youn, B.D. & Choi, K.K. 2003. Hybrid Analysis Method for Reliability-Based Design Optimization. ASME Journal of Mechanical Design 125(2):221–232. Zhao, Y.G. & Ono, T. 2001. Moment methods for structural reliability. Structural Safety 23: 47–75.

Chapter 4

Efficient approaches for system reliability-based design optimization Efstratios Nikolaidis University of Toledo, Toledo, OH, USA

Zissimos P. Mourelatos & Jinghong Liang Oakland University, Rochester, MI, USA

ABSTRACT: Two efficient approaches for series system reliability-based design optimization (RBDO) are presented. Both approaches apportion optimally the system reliability among the failure modes by considering the target values of the failure probabilities of the modes as design variables. The first approach uses a sequential optimization and reliability assessment (SORA) approach. It approximates the coordinates of the most probable points of the failure modes as the design changes through linear extrapolation. The second system RBDO approach uses a single-loop method where the searches for the optimum design and for the most probable failure points proceed simultaneously. The efficiency and robustness of the single-loop based approach is enhanced through an easy to implement active set strategy. The two approaches are illustrated and compared on design examples. It is shown that both approaches yield more efficient designs than a conventional component RBDO formulation. Moreover, it is shown that the single-loop approach is considerably more efficient than the SORA approach.

1 Introduction This section presents an overview of reliability-based design optimization (RBDO). Two groups of methods for performing component RBDO efficiently are explained. The section concludes with an outline of this chapter. 1.1

RBDO: Benefits and challenges

Competitive pressures compel manufacturers to minimize weight and cost while maintaining an acceptable safety level. Designers face significant uncertainty such as uncertainty in the operating conditions, material properties and the geometry of a design. In this chapter, the term “uncertainty’’ refers to both random and epistemic uncertainties. In deterministic design, uncertainty is traditionally accounted for by using safety factors and conservative design (characteristic) values for strength and loads. This empirical approach often leads to overdesign or occasionally to unsafe designs. Moreover, deterministic design is not adequate for novel designs involving new materials and geometries. RBDO provides safer and more efficient designs than deterministic design optimization because it explicitly accounts for uncertainty using probability theory. As a result, RBDO is being increasingly accepted as an effective design tool for aerospace, automotive, civil and ocean engineering structures. An overview of RBDO is given in

88

Structural design optimization considering uncertainties

(Frangopol & Maute 2004) with applications to aerospace, civil and micoelectromechanical system design. Studies have demonstrated that RBDO can produce a more efficient design than a deterministic approach without sacrificing safety, or alternatively, RBDO can yield a safer design than a deterministic approach for a given maximum allowable cost. Here the safety of a design is measured by its system reliability. A designer faces many challenges when applying RBDO to practical problems. The high computational cost and the consideration of the system failure probability in RBDO are two principal challenges that the study presented in this chapter tries to address. Finding the probability of failure of a design requires repeated structural analyses for different sets of values of the random variables, which may be computationally expensive, especially when finite element analysis is required (Madsen et al. 1986, Moses 1995, Melchers 2001). Also, the computational cost is grossly compounded when calculating the probability of failure of different designs during the search for the optimum reliability-based design. For example, if a second moment method is employed to calculate the probability of failure, then two nested optimizations must be performed. This approach, called double loop (DLP) approach, requires one optimization for finding the optimum values of the design variables (outer loop) and a second optimization (inner loop) for finding the most probable point (MPP), which is needed for estimating the probabilities of the failure modes. Another principal challenge in RBDO is to consider the system reliability. Calculating system reliability is expensive, especially if one wants to account for the probability of the intersection of the failure modes. As a result, most RBDO studies constrain only the safety levels of the individual failure modes. Thus, the user must decide the minimum allowable safety level (or equivalently safety index) of each failure mode (constraint) instead of specifying only an acceptable system safety level. This approach has several shortcomings. First, the required safety levels of the failure modes are usually not optimal. Second it does not allow consideration of the cost required to achieve a certain reliability level for each failure mode. Finally, it does not account for the interactions of the failure modes (e.g., the probability of these modes occurring simultaneously). We believe that we can obtain considerably better designs by enabling the user to specify the minimum acceptable safety level of the system only, and allowing the optimizer to determine the required safety levels the failure modes, instead of asking the user to select them. 1.2 Pro g ress i n r e d uc ing t he c o mput at i o n a l co s t o f R B D O To address the problem of the high computational cost required by DLP approaches, two new classes of RBDO formulations have been recently proposed. The first class decouples the RBDO process into a sequence of cycles consisting of a deterministic design optimization followed by a reliability assessment of the found optimum (Du & Chen 2004, Royset et al. 2001). The constraints in the deterministic optimization dictate that the design does not fail at a checking point, which is some approximation of the MPP. The Sequential Optimization and Reliability Assessment (SORA) method uses the reliability information from the previous cycle to shift the violated deterministic constraints in the feasible domain. SORA appears to be similar to the safety-factor approach reported in (Wu et al. 2001). Another decoupled approach has also been

E f f i c i e n t a p p r o a c h e s f o r s y s t e m r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n 89

proposed in (Royset et al. 2001). However, it is restricted to deterministic design variables in the design optimization loop. The second class of RBDO methods converts the problem into an equivalent, singleloop deterministic optimization by integrating the two optimization loops into one. The approach in (Thanedar & Kodiyalam 1992) uses a mean value first-order reliability method. However, it is numerically inaccurate or unstable due to a wrong estimation of the probabilistic constraints. The single-loop, single-vector (SLSV) approach (Chen et al. 1997) is the first attempt in a truly single-loop approach. It improves the RBDO computational efficiency by eliminating the inner reliability estimation loops. However, it requires a probabilistic active set strategy for identifying the active constraints, which may hinder its practicality. A single-level RBDO approach has also been reported in (Kuschel & Rackwitz 2000, Streicher & Rackwitz 2004, and Agarwal et al. 2004). It integrates the nested optimization loops into one by enforcing the Karush-Kuhn-Tucker (KKT) optimality conditions of the inner loop as equality constraints in the outer design optimization loop. In doing so however, it increases the number of design variables because it uses the standard normal variants for each constraint as additional design variables. This can increase the computational cost substantially, especially for practical problems with many design variables and constraints. Furthermore, the approach requires second-order derivatives that are computationally costly and difficult to calculate accurately. One of the system RBDO approaches presented in this chapter uses the single-loop RBDO of Liang et al. (2004). This single-loop system RBDO approach is summarized in section 2.3.

1.3 Approaches for s ys tem RBDO This chapter presents two RBDO approaches for series systems. The probability of failure of a series system is approximated using the first-order or the second-order Ditlevsen upper bounds (Ditlevsen 1979). In the first-order bound, the system failure probability is approximated by the sum of the failure probabilities of all failure modes. This approximation is accurate for most systems whose probability of failure is low (e.g., less than 10−5 ). The second-order bound provides a more accurate approximation of the system failure probability than the first-order bound by accounting for the joint probability of pairs of failures modes. The first approach, which was first reported in (Ba-abbad et al. 2006), uses a sequence of deterministic optimization problems. The MPPs of the failure modes are approximated using linear extrapolation. This approach is a modified formulation of the SORA method in which the reliabilities of all failure modes are considered as design variables. As a result, the approach allows for an optimal apportionment of the reliability of a system among its failure modes. The second approach for system RBDO utilizes the single-loop RBDO approach of (Liang et al. 2004) to determine the optimum design. This approach approximates the MPP’s in each design iteration, using a relation representing the KKT optimality conditions instead of using linear extrapolation (Ba-abbad et al. 2006). To facilitate convergence, an active set strategy is used for identifying the critical failure modes whose failure probabilities affect significantly the system failure probability, in each

90

Structural design optimization considering uncertainties

iteration. The failure probabilities of the remaining non critical failure modes are assumed zero. Three major developments are involved in the above two approaches for system RBDO. The fundamental development is the approach that allows for an optimal apportionment of the reliability of a system among its components. Although this approach increases the number of design variables, it does not affect significantly the algorithmic efficiency because the objective function is not a function of the additional design variables. Furthermore, the inclusion of the failure probabilities of the modes in the design variable set does not increase significantly the cost of calculating the constraints. The second major development is the use of approximations (SORA or a single-loop approach) to solve efficiently the RBDO problem. The single-loop approach is more efficient than the SORA approach because the single-loop approach eliminates the sequence of solutions of deterministic optimization problems. Moreover, the single-loop approach does not approximate the MPP using extrapolation. Finally, the third development is an active set strategy used by the single-loop system RBDO approach. This strategy updates the active constraint set in each iteration to ensure algorithmic stability. Both system RBDO approaches presented in this chapter provide more efficient designs than component RBDO approaches because the system RBDO approaches account for the relation between the cost (or weight) and the reliability of each failure mode. Specifically, these approaches optimize the reliabilities of the failure modes using information about the sensitivity derivatives of the reliabilities of the modes and the sensitivity derivatives of the cost with respect to the design variables. An optimality condition for the reliabilities of the modes that involves these sensitivities is presented in section 4. The details of the proposed system RBDO approaches for series systems including algorithms for these approaches are described in section 2. The efficiency and accuracy of the two approaches are demonstrated and compared in section 3, using two examples that involve a cantilever beam and an internal combustion engine. Section 4 presents an optimality condition for a general, system RBDO problem and shows that the optimum cantilever beam design satisfies this condition. Section 5 presents the conclusions.

2 System RBDO methods The two system RBDO approaches are presented in this section. First, a general form for a series system RBDO problem is presented in subsection 2.1. An equivalent performance measure approach formulation, in which the failure probabilities of the constraints are considered as design variables, is derived. This formulation is used as a basis to develop the SORA and single-loop based approaches in sections 2.2 and 2.3. 2.1 F o rm ul a ti o n o f s y s t e m R B DO pr o bl e m The general system RBDO problem seeks the most efficient design whose system prob. The formulation of ability of failure does not exceed a maximum allowable value pall f

E f f i c i e n t a p p r o a c h e s f o r s y s t e m r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

91

the system RBDO problem is as follows, min f (d, µX , µP )

(1)

d,µx

 subject to

psys = P

n '

 Gi (d, X, P) ≤ 0 ≤ pall f

i=1

dL ≤ d ≤ dU , µLX ≤ µX ≤ µU X where d ∈ Rk is the vector of deterministic design variables, X ∈ Rm is the vector of random design variables and P ∈ Rq is the vector of random parameters. Symbols µX and µP denote the mean values of the random variables and random parameters respectively. A bold letter indicates a vector, an upper case letter indicates a random variable or parameter and a lower case letter indicates a realization of a random variable or parameter. The probability of failure of a system with n failure modes is denoted by psys and it is equal to P[∪ni=1 Gi (d, X, P) ≤ 0], where Gi (d, X, P) is the performance function of the ith failure mode. In subsequent formulations of the RBDO problem, the side constraints on the design variables will be omitted for simplicity. The system failure probability is a function of the probabilities pfi of the failure modes and their joint probabilities, psys =

n 

pfi −

i=1

n n  

pfij + · · · (−1)(n−1) pf1,2,...,n

(2)

i=2 j
Let us constrain the probabilities of failure of the modes to be no greater than some bounds ptfi , called target probabilities, and consider these probabilities as design variables. Then the system RBDO formulation in Eq. (1) becomes, min f d,µx ,ptf ,...,ptf 1

(3)

n

subject to psys ≈

(d, µX , µP )

n  i=1

pfi ≤ ptfi ptfi −

i = 1, 2, . . . , n

n  n 

pfij + · · · (−1)(n−1) pf1,2,...,n ≤ pall f

i=2 j
In this formulation, the optimizer should determine the optimum values of the target failure probabilities of the modes, besides the values of the original design variables, d and µX . Consider a performance measure approach (PMA) formulation (Tu et al. 1999, Youn et al. 2003) of problem (2) in which the constraints on the failure probabilities of the modes are written in terms of their safety indices, βi . In the PMA formulation, instead of checking if the minimum distance of the MPP from the origin is no less than

92

Structural design optimization considering uncertainties

the target value of the safety index of each failure mode, we check if the performance function is nonnegative at the MPPs, XMPP (βit ), PMPP (βit ). This formulation is, min f (d, µX , µP )

(4)

d,µx ,β1t ,...,βnt

Gi [d, XMPP (βit ), PMPP (βit )] ≥ 0,

subject to psys ≈

n 


(−βit ) −

i=1

n n  

i = 1, 2, . . . , n

pfij + · · · (−1)(n−1) pf1,2,...,n ≤ pall f

i=2 j
where
is the cumulative probability distribution of a standard normal random variable (zero mean, unit standard deviation) and
(−βit ) = ptfi and XMPP and PMPP are the values of the random design variables and parameters at the MPP for each constraint. Symbol βit denotes the target value of the safety index of the ith failure mode. As Eq. (3) indicates, for each mode, XMPP and PMPP are functions of the target value of its safety index. The above RBDO problem is too expensive to solve for most practical problems because the probabilities of failure of the modes need to be calculated each time a design changes. Therefore, two efficient approaches that use approximations of the system failure probability in terms for the failure probabilities of the modes and approximations of the MPP are presented below. The first approach uses linear extrapolation to find the MPP of a design as a function of its safety index, while the second solves the optimality conditions for the MPP.

2.2

SORA-b a se d s y s t em R B DO appr o ac h

2.2.1 F ormu l a tio n Assume that the system failure probability is equal to the sum of the probabilities of  the failure modes psys ≈ ni=1 pfi ≈ ni=1
(−βit ). This is a conservative approximation and it is accurate for small failure probabilities, i.e., 10−5 (Liang et al. 2007). The key idea of the SORA approach is to approximate the coordinates of the most probable point as a function of the value of the safety index. Let UMPP (βi ) = (β ) X T −1 MPP i denote the vector of the reduced values of random variables at the PMPP (βi ) MPP, and T the transformation from the space of the reduced variables to the space of the original variables. The MPP, UMPP (βi ), is approximated as a function of the value of the safety index, βi , given the MPP, UMPP (βi0 ), for a baseline value of the safety index, βi0 , as follows (Fig. 4.1), UMPP (βi ) =

βi UMPP (βi0 ) βi0

(5)

E f f i c i e n t a p p r o a c h e s f o r s y s t e m r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

U MPP2 (b)

93

G
const.

b

U MPP2

(b0) b0 G
const. U MPP1 (b0)

U MPP1 (b)

Figure 4.1 Approximation of MPP as function of the safety index.

This approximation allows recasting the system RDBO problem formulation in Eq. (4) as a deterministic optimization problem as follows; min f (d, µX , µP )

(6)

d,µx ,β1t ,...,βnt

 subject to

psys ≈

n 

Gi



βt d, T 0i UMPP (βi0 ) βi

 ≥ 0,

i = 1, 2, . . . , n


(−βit ) ≤ pall f

i=1

In the above formulation, the target failure probabilities of the modes have been replaced by the corresponding target values of the safety indices. Since there is a oneto-one relation between the probability of failure and the safety index of a mode, this substitution of the design variables does not change the solution of the optimization problem. The solution of Eq. (6) is a design whose failure modes have safety indices approximately equal to βit . These approximate values are called herein “projected values of the safety indices’’. The approximation of the design point as a function of the safety index in Eq. (5) is only valid in a trust region around the baseline value of the safety index. The progress of the optimization should be monitored in each iteration and the change in the value of the safety index should be constrained within some move limit to remain in the trust region. Available methods for optimization using trust regions can be used for this purpose (Moré & Sørensen 1983, Steihaug 1983, Sørensen 1994). 2.2.2 A lgorithm Figure 4.2 describes the system RBDO algorithm. Each cycle of this algorithm consists of three operations: (a) reliability analysis using a First-Order Reliability Method

94

Structural design optimization considering uncertainties

Select initial design

Perform PMA analysis to find MPP for minimum acceptable value of safety index

Approximate MPP as function of safety index

Solve approximate deterministic optimization problem

N Convergence? Y Stop

Figure 4.2 Algorithm of a SORA based system RBDO method.

(FORM) of the initial design or of the design obtained from the previous cycle to check if this design has acceptable reliability, (b) PMA reliability analysis of the design to determine the MPPs of the failure modes, UMPP (βi0 ), and (c) approximate deterministic optimization to update the optimum design and find the target values of the probabilities of the failure modes, Pfti . First, we perform a deterministic optimization using a factor of safety to find an initial design. Then we perform FORM reliability analysis of the deterministic optimum. At this stage we can identify the non-critical failure modes, that is, those failure modes of the deterministic optimum design that do not affect significantly the system reliability because they have negligible probability of failure compared to the failure probabilities of the remaining critical modes. The target values of the safety indices of these modes are removed from the set of design variables to facilitate convergence. Then we perform an inverse reliability analysis (PMA) of the deterministic optimum assuming equal probabilities of failure, which are obtained by dividing the allowable /nc . failure probability of the system by the number of critical failure modes; ptfi = pall f Finally, we perform approximate deterministic optimization considering the target values of the safety indices of the nc failure modes, β1t , . . . , βnt c , as design variables in addition to the original design variables. Now the optimizer seeks both the optimum values of the design variables and the optimum target values of the safety indices to minimize the objective function (i.e. cost or weight), and the optimization problem is given by Eq. (6.b). In this case, the optimization problem formulation (6) becomes, min f d,µx ,β1t ,...,βnt c

(d, µX , µP ) i = 1, 2, . . . , nc

E f f i c i e n t a p p r o a c h e s f o r s y s t e m r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

subject to

   βi Gi d, T 0 UMPP (βi0 ) ≥ 0, βi psys ≈

nc 

95

i = 1, 2, . . . , nc (6.b)


(−βit ) ≤ pall f

i=1

Once we have found the optimum, we check the failure probabilities of all failure modes (including the non critical ones) using FORM. At this step, we may remove (or add) the target values of the safety indices of some failure modes with low (high) failure probabilities from the set of design variables. Then we perform PMA analysis for the optimum values of the target values of the safety indices found from the deterministic optimization. Finally, we solve the approximate deterministic optimization problem again. We repeat this process until convergence. 2.3

Single-loop RBDO approach

The proposed approach is based on the single-loop RBDO algorithm of Liang et al. (2004), which is referred herein as a component, single-loop algorithm. It is based on an equivalent deterministic optimization formulation, which eliminates the need for inner reliability loops without increasing the number of design variables. For completeness, a brief overview of the component, single-loop RBDO algorithm is given below. 2.3.1 O verview of a component si ngl e-l oop RB D O Designers replace the constraint on the system reliability with constraints dictating that the reliabilities of the components are greater than or equal to some target values, in order to circumvent the calculation of the system reliability. These target values are often chosen based on judgment and experience. A typical component RBDO problem is formulated as, min f (d, µX , µP ) d,µX

subject to

(7) Ri = P[Gi (d, X, P) ≥ 0] ≥ Rti ,

i = 1, 2, . . . , n

where Ri = 1 − pfi is the actual reliability level of the ith constraint (or failure mode) and Rti is the corresponding minimum allowable reliability. A method solving directly the optimization problem (7) constitutes the double-loop RBDO method. This method employs two nested optimization loops; the design optimization loop (outer) and the reliability assessment loop (inner). The latter is needed for the evaluation of each probabilistic constraint. If the probability of failure is estimated using FORM, then every time the design optimization loop calls for a constraint evaluation, a reliability assessment loop is executed that searches for the MPP in the standard normal space. If the random variables are not normal, a nonlinear transformation maps the original space to the standard (or reduced) normal space. Using an R-percentile formulation (Du & Chen 2004), the RBDO problem (7) can be expressed as, min f (d, µX , µP ) d,µX

subject to

(8) Gi (d, X, P) ≥ 0,

i = 1, 2, . . . , n

96

Structural design optimization considering uncertainties

where the vectors X and P are evaluated at the MPP; i.e. X = XMPP and P = PMPP for each constraint. The objective function is minimized subject to constraints that are evaluated in the X space. It is therefore, necessary to have a consistent relationship between vectors d, µX , µP , for which the objective function is evaluated, and vectors d, X, P for which the constraints are evaluated. This is done by solving the Karush-KuhnTacker (KKT) optimality conditions (Papalambros & Wilde 2000) of the reliability loops in the design optimization loop for X and P. Using the PMA method, the performance measure Gp = minG(U) is minimized on U

the beta-circle H(U) = U − βt = 0 in the standard normal space U. At the optimal point, according to the KKT optimality condition, the gradient ∇G(U) of the limit state and the gradient ∇H(U) of the beta-circle are collinear and point in opposite directions. This condition yields, U = −βall · α

(9)

where α = ∇GU (d, X, P)/∇GU (d, X, P)

(10)

is the normalized gradient of the constraint in U-space. Based on Eq. (9), the following relations between X, P and µX , µP hold for normal random variables, X = µX − σ · βt · α,

P = µP − σ · βt · α

(11)

where α = σ · ∇GX,P (d, X, P)/σ · ∇GX,P (d, X, P). Using Eqs (11), the double-loop RBDO problem (8) is transformed to the following single-loop, equivalent deterministic optimization problem, min f (d, µX , µP ) d,µx subject to

Gi (d, Xi , Pi ) ≥ 0,

(12) i = 1, 2, . . . , n

where Xi = µX − σ · βit · αi ; Pi = µP − σ · βit · αi ; αi = σ · ∇GiX,P (d, Xi , Pi )/σ · ∇GiX,P (d, Xi , Pi ) Symbol αi represents the normalized gradient of the ith constraint and σ is the standard deviation vector of random variables X and random parameters P. In the single-loop RBDO problem (12), the objective function is evaluated at the mean point d, µX , µP and the constraints are calculated at point d, X, P. The relationship of Eq. (11) is used to evaluate the constraints consistently with the values of the design variables. The single-loop method does not search for the MPP of each constraint in each iteration. Instead, in each iteration, an approximation of the MPP’s of the active constraints is used for each constraint. The sequence of the MPP approximations convergences to the correct MPP. This dramatically improves the efficiency of the single-loop method without compromising the accuracy. Figures 4.3a and 4.3b show a combined flowchart of the component single-loop method and the proposed system, single-loop RBDO method (see section 2.3.2). The

E f f i c i e n t a p p r o a c h e s f o r s y s t e m r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

97

Initialize d0, ␮0X, µp, p t,0 f , ␴, lb, ub

Assign X0
␮0X, P0
␮P

Assign CF = 1 Calculate bit
Φ1(1 − pfit)

Calculate ␣i0
␴•∇Gi(X,P)(d0, X0, P0)/||␴•∇Gi(X,P) (d0, X0, P0)||

• • • • • •

Repeat the following steps until convergence: Update counter, k Determine active constraint set Calculate objective function Approximate MPPs of random variables Check active constraint set Perform new iteration of problem (12) for component RBDO or problem (19) for system RBDO See Fig. 3b for detailed calculations in this part of the algorithm

Figure 4.3a Overview of the single-loop RBDO algorithm.

component, single-loop algorithm consists only of operations in boxes with solid line border. Operations in boxes with dashed line border belong to the system, singleloop algorithm. For the component single-loop method, the initial point d0 , µ0X , µP and the target safety index vector βt for all constraints are first specified. Also, the user specifies the standard deviation vector σ for the random variables and random parameters and the upper and lower bound vectors (ub and lb, respectively) for all deterministic and probabilistic design variables. The initial point d0 , X0 , P0 that is needed to evaluate the constraints is assumed equal to d0 , µ0X , µP ; i.e. X0 = µ0x and P0 = µP . At this point, the initial normalized gradient vector α for the ith constraint is calculated as, α0i = σ · ∇Gi(X,P) (d0 , X0 , P0 )/σ · ∇Gi(X,P) (d0 , X0 , P0 )

(13)

At the kth iteration of the optimization loop, the objective function is calculated at point dk , µkX , µP . For the evaluation of the constraints, the algorithm checks if the optimizer has changed the design vector µkX compared with the previous iteration;

98

Structural design optimization considering uncertainties

dk, ␮kX Change? Yes, k
k1

No

Calculate ␣ik
␴⋅∇Gi(X,P)(dk1,Xk1 ,Pk1 )/|| ␴•∇Gi(X,P) (dk,Xk1 ,Pk1 )|| i i i i

Yes ptf ≤ ε ? i

No

Assign CF(i)
0; bti
0

Assign CF(i)
1 Calculate bti
Φ−1(1ptf ) i

Assign MaxBeta
max(bti ) CF (i)
0?

Yes

βi t
MaxBeta

Perform new iteration

Calculate Xki
␮Xk  βit•␣ki •␴, Pik
␮Pbti •␣ki •␴ Yes

CF (i)
0?

Calculate Gi(dk,Xki ,Pki ) No

No

Gi  0? Yes Update: C F (i)
1

Calculate Gi(dk,Xki ,Pik) for active constraints n n Calculate ∑ ptfi  ∑ max pfij i
1 i=2 j
Is f minimized?

Figure 4.3b Calculations performed in one iteration of single-loop algorithm.

if not, the current gradient vector αk is used to calculate Xk = µkX − σ · βt · αk and Pk = µP − σ · βt · αk for each constraint. If µkX has changed from the previous iteration, the normalized gradient vector αk is updated before it is used to calculate Xk and Pk , which are needed for the constraint evaluation.

E f f i c i e n t a p p r o a c h e s f o r s y s t e m r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

99

This is an essential step for keeping the design variable vector µX and the X, P vectors consistent, resulting in a robust and stable algorithm. Furthermore, it greatly improves the efficiency since the algorithm avoids unnecessary gradient evaluations. When non-normal variables are used, equivalent σXN and µN X (Haldar & Mahadevan 2000) are calculated every time the optimizer updates the design variables and design parameters. The main advantage of the component single-loop method is that it eliminates the repeated reliability analysis loops without increasing the number of design variables or adding equality constraints. Instead of performing nested design optimization and reliability loops, it solves an equivalent single-loop deterministic optimization problem. The consistency between the design variable vector d, µX , µP and vector d, X, P needed to evaluate the constraints makes the single-loop algorithm robust. It should be noted that the component single-loop RBDO algorithm does not require an active constraint set as is the case with the SLSV algorithm (Chen et al. 1997). The active constraint set is simply identified by the algorithm. This is a significant advantage, which simplifies the implementation of the single-loop algorithm and enhances its robustness and efficiency.

2.3.2 Single-loop approach for seri es sy stem RB D O The component single-loop RBDO approach of the previous subsection handles RBDO problems in which each critical failure mode of a series system has a predetermined minimum target safety index βit . Thus the user arbitrarily assigns a minimum safety level for each failure mode instead of letting the optimizer determine this level in order to achieve some required system reliability. A single-loop RBDO approach for series systems is proposed in this subsection. The optimizer determines the optimal values of the target failure probabilities of all failure modes besides the original design variables d, µX and µP . The user specifies a system reliability level and the optimizer allocates optimally the specified system reliability among its failure modes. The optimal reliability allocation and the optimal reliable design are determined simultaneously using an equivalent single-loop RBDO formulation. The proposed series system approach is a modification of the component singleloop approach of the previous section. According to Eq. (12), a target safety index βit = −
−1 (ptfi ) is needed for each constraint (failure mode). However, the optimizer must determine the target failure probability ptfi of each failure mode by apportioning the allowable system probability of failure pall among all failure modes. A natural way f to do this is to include all the target values of the failure probabilities of the constraints, ptfi , into the design variable set. In each iteration, the optimizer determines each ptfi and the corresponding target safety index βit = −
−1 (ptfi ) is calculated so that transformations Xi = µX − σ · βit · αi and Pi = µP − σ · βit · αi in Eq. (12) hold. Simultaneously, we must make sure that the system failure probability does not exceed the maximum allowable its value psys ≤ pall . The system failure probability is approximated by the f   upper second order Ditlevsen bound, psys = ni=1 ptfi − ni=2 max (pfij ), where Pfij is the j
joint probability between the ith and jth failure modes.

100

Structural design optimization considering uncertainties

Based on the first-order reliability analysis (FORM), the failure set is approximated by a polyhedral bounded by the tangent hyperplanes at the MPP points. In this case, ptfi =
(−βit ), where βit is the safety index for the ith failure mode. Similarly pfij is obtained by approximating the joint failure set using the tangent hyperplanes at the MPP points of the two failure modes,
pfij =
(−βi , −βj ; ρij ) =

−βi

−∞




−βj

−∞

ϕ(x, y; ρij ) dxdy

(14)

where ϕ(, ; ρ) is the PDF of a bivariate normal vector with zero means, unit variances, and a correlation coefficient ρ given by, 

2 2 1 βi + βj − 2ρβi βj ϕ(−βi , −βj ; ρ) = exp −  2 1 − ρ2 2π 1 − ρ2

1

 (15)

In Eq. (14),
(, ; ρ) is the bivariate normal CDF, which has the property, ∂ϕ ∂2
= ∂x∂y ∂ρ

(16)

Combining Eqs (14) to (16) yields,
pfij =
(−βi , −βj ; 0) +

ρij 0




ρij

=
(−βi )
(−βj ) +

∂
(−βi , −βj ; z) dz ∂ρ

(17)

ϕ(−βi , −βj ; z) dz

0

Note that pfij is directly related to the degree of correlation between the failure modes, expressed by the correlation coefficient ρij , which varies between −1 (fully, negatively correlated) and +1 (fully, positively correlated). In this work, pfij is obtained by approximating the joint failure set by the tangent hyperplanes at the corresponding MPP points of the two failure modes. If U denotes the MPP point in standard normal are then replaced by the linear safety margins space, the  safety margins Gi (U) and G j (U) m Mi = βi − m α U and M = β − α ir r j j r=1 s=1 is Us , so that the correlation coefficient ρij is given by, ρij = ρ[Mi , Mj ] =

n 

αir αjr = cos (υij )

(18)

r=1

After the correlation coefficient ρij between two active failure modes is calculated, Eq. provides the joint probability of failure needed in evaluating constraint n n (17) t p − max ptfij ≤ pall . i=1 fi i=2 f j
E f f i c i e n t a p p r o a c h e s f o r s y s t e m r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

101

The component, single-loop RBDO algorithm of Eq. (12) is therefore, modified as follows, (d, µX , µP )

min f d,µx ,ptf ,...,ptf 1

(19)

n

subject to Gi (d, Xi , Pi ) ≥ 0, i = 1, 2, . . . , n  n  ' Gi (d, Xi , Pi ) ≤ 0 psys = P i=1

≈

n  i=1

ptfi −

n  i=2

max (pfij ) ≤ pall f j
where Xi = µX − σ · βit · αi ;

Pi = µP − σ · βit · αi ,

αi = σ · ∇GiX,P (d, Xi , Pi )/σ · ∇GiX,P (d, Xi , Pi )

i = 1, 2, . . . , n

If the target probability of the ith failure mode ptfi is very small and the corresponding safety index βit = −
−1 (ptfi ) is very large, the optimization algorithm of problem (2.3.2) becomes computationally inefficient and in most cases, it fails to converge because the constraints become insensitive to the values of design variables ptfi . This is also true for any DLP RBDO algorithm based on the PMA approach if a very large target safety index is used. That is why the PMA approach with a “small’’ target safety index is superior to the RIA approach where the safety index of inactive constraints is very large (Youn et al. 2003). To avoid this problem, the proposed single-loop system RBDO algorithm uses an active set strategy that is very easy to implement because all target failure probabilities are included in the design variable set. In every iteration of the optimization process, ptfi is compared with a small predefined threshold value ε. If ptfi ≤ ε, the ith constraint is probabilistically inactive and its probability of failure is assumed zero (i.e. ptfi = 0 for all inactive constraints). Therefore, in each iteration, an active constraint is easily identified. This set is updated in each iteration. The predefined threshold value ε depends on the problem and can be set equal to a fixed small value or equal to small percentage of the largest failure probability of the modes. In this study the threshold was set equal to 10−7 , which corresponds to a safety index of 5.1993. The flowchart of the proposed single-loop, series system RBDO algorithm is shown in Figs. 3a and 3b. First, the algorithm initializes all design variables and parameters (including probabilities of failure of each failure mode), and specifies lower and upper bounds of design variables and standard deviations. Subsequently, the initial gradients for all constraints are calculated. After initialization, the flowchart is similar to that of the component single-loop method. The only difference is the implementation of an active set strategy, which is explained in detail below. The following two points regarding the efficiency and stability of the proposed system single-loop algorithm are important.

102

•

•

Structural design optimization considering uncertainties

The increased number of design variables increases the number of iterations of the optimizer but it does not affect appreciably the computational cost in each iteration. The new approach adds the target failure probabilities of the constraints, ptfi , to the design variable set, thereby increasing the number of design variables by the number of constraints. This increases the number of iterations of the optimizer needed to achieve convergence. However, the computational cost in each iteration does not increase appreciably because the biggest part of this cost is due to gradient calculation, which is easy to perform. The reasons are that the objective function is not a function of the target failure probabilities of the modes and each constraint is a linear function of its safety index. A probabilistic active set strategy is used to increase the efficiency and stability of the algorithm. It has been mentioned that in any iteration, if the probability of failure of the ith constraint is smaller than a threshold value ε the constraint is assumed probabilistically inactive. It is very easy to check if a constraint is smaller than the threshold because all failure probabilities are design variables. As is indicated in Fig. 3b, a constraint flag CF(i), which is set equal to zero for inactive constraints and one for active constraints, is used to identify inactive constraints. The safety indices of all inactive constraints are set equal to the maximum safety index of the active constraints. Fig. 3b shows the details of the procedure identifying inactive constraints.

After a safety index is calculated or assigned for each constraint, the algorithm calculates quantities Xi = µX − σ · βit · αi and Pi = µP − σ · βit · αi for the ith constraint, relating X, P and µX , µP at the current MPP approximation using the KKT conditions (see section 2.1). At this point, the feasibility of all inactive constraints is checked by calculating the value of each inactive constraint at (Xi , Pi ) allowing us to update the active constraint set (see Fig. 3b) in each iteration. Subsequently, the value of all active constraints  at (Xi , Pi ) is obtained as well as the system probability of failure psys = ni=1 ptfi − ni=2 max (pfij ). In calculating the joint failure probability j
pfij , we first check the value of the correlation coefficient ρij . If ρij is equal to 1, the two failure modes are positively fully correlated and their joint failure probability can be approximated by min (ptfi , ptfj ). If ρij is equal to −1 the two failure modes are negatively fully correlated and their joint failure probability can be assumed zero. If ρij ≈ 0, then the failure modes are independent and ptfij =
(−βi ) ·
(−βj ). In any other (ρ case, ptfij =
(−βi )
(−βj ) + 0 ij ϕ(−βi , −βj ; z) dz is used in problem (2.3.2), and the system single-loop RBDO algorithm continues similarly with the component single-loop version.

3 Applications This section demonstrates and the accuracy and efficiency of the proposed SORA and single-loop approaches for series system RBDO problems using a beam example, and an internal combustion engine design example. Deterministic optimization,

E f f i c i e n t a p p r o a c h e s f o r s y s t e m r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

103

Y

Z

t L
100 in

w

Figure 4.4 Cantilever beam under vertical and lateral bending.

a component single-loop method, and the proposed series system, single-loop RBDO approaches are compared. In all cases, the same initial point and similar convergence criteria are used. MATLAB was used for deterministic optimization, and the single-loop approaches. The add-in tool “Solver’’ in MS-Excel was used for constrained optimization in the SORA system RBDO approach. 3.1 A cantilev er beam example Consider a cantilever beam in vertical and lateral bending (Wu et al. 2001) (see Fig. 4.4). The beam is loaded at its tip by vertical and lateral loads Y and Z, respectively. Its length L is equal to 100 in. The width w and thickness t of the cross-section are deterministic design variables. The objective is to minimize the weight of the beam. This is equivalent to minimizing the cross sectional area f = w · t, assuming that the material density and the beam length are constant. Two non-linear failure modes are considered. The first failure mode is yielding at the fixed end of the beam; the other failure mode is that the tip displacement exceeds the allowable value D0 = 2.2535 in. In the single-loop system RBDO approach the problem is formulated as, min f = w · t

w,t,ptf ,ptf 1

2

subject to

P[Gi (X) ≥ 0] ≥ 1 − ptfi

i = 1, 2

 600 600 · Z · Y + wt 2 w2 t 

 2 2 Z Y 4L3 G2 (E, Z, Y, w, t) = D0 − + Ewt t2 w2

G1 (SY , Z, Y, w, t) = SY −

Gsys = 0.0027 − psys = 0.0027 − (ptf1 + ptf2 − pf12 ) ≥ 0 0 ≤ w, t ≤ 5 where G1 and G2 are the limit state functions corresponding to the two failure modes. In the SORA approach, the problem formulation is identical but the system probability

104

Structural design optimization considering uncertainties Table 4.1 Comparison of RBDO methods for the beam example.

w = x1 t = x2 ptf1 ptf2 pf12 psys Objective f (X) G 1 (X) G 2 (X) No. of F.E.

Deterministic optimization

Component single-loop

System single-loop (SORA)

2.3520 3.3263

2.4484 3.8884 0.00135 0.00135 Neglected 0.00270 9.5202 0 0 115

2.6093 (2.6209) 3.6126 (3.6001) 0.002412 (0.002328) 0.000426 (0.000372) 0.0001389 (Neglected) 0.00270 9.4263 (9.4356) 0 0 624∗

7.8235 640.3600 0 83

∗ This number refers to the number of function evaluations of the Single-Loop approach. The number of function evaluations for the SORA approach could not be determined.

10

Objective values

9.5 9 8.5 8 7.5 Deterministic optimization Component single-loop System single-loop/JPDF

7 6.5 6 0

2

4

6

8

10

12

14

16

Iteration numbers

Figure 4.5 Optimization history for the beam example.

of failure is approximated by the sum of the failure probabilities of the two modes. Design variables w and t are deterministic while Y, Z, SY and E are normally distributed random parameters with Y ∼ N (1000, 100) lb, Z ∼ N (500, 100) lb, SY ∼ N (40 000, 2000) psi and E ∼ N (29 · 106 , 1.45 · 106 )psi; SY is the random yield strength, Z and Y are mutually independent random loads in the vertical and lateral directions respectively, and E is the Young’s modulus. A safety index βit = 3, which corresponds to a maximum allowable probability of failure of 0.00135, is used for both constraints in the component single-loop approach (see Eq. 7). Table 4.1 and Fig. 4.5 compare the efficiency of the deterministic, component single-loop and system single-loop optimizations and the resulting optimum designs. A common initial point (w = 2.5, t = 2.5) and common convergence conditions were used for all three optimizations. Both the component and system RBDO problems have the same allowable system failure probability pall = 0.0027, which corresponds f

E f f i c i e n t a p p r o a c h e s f o r s y s t e m r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

105

all to βsys = 2.78215 – the difference is that in component RBDO the probabilities of both failure modes are were constrained to be less than or equal to 0.00135, whereas in system RBDO the maximum allowable failure probabilities of the modes were determined by the optimizer. The component and system single-loop methods converged in six and nine iterations, respectively (see Fig. 4.5). The deterministic optimization converged in eight iterations. The SORA system RBDO approach required four cycles each involving a deterministic optimization and reliability analysis. A total of 25 iterations were required by the optimizer (seven in each of the first three cycles and four in the fourth). Due to the optimal apportionment of the allowable system reliability, both system RBDO approaches resulted in an optimum cross sectional area of approximately 9.43 in2 , which is better than the component optimum of 9.5202 in2 . The optimizer allocated a much lower probability of failure to the second failure mode than the first, indicating that the reliability of this mode can be increased at a much lower cost (cross sectional area) than the reliability of the first mode. Note that for the component approach, the system failure probability is 0.0027 (0.00135 + 0.00135). Both constraints are active in the component and system approaches while only the second constraint is active in the deterministic optimization. The single-loop system approach yielded a more efficient design than SORA because the former approximates more accurately the system failure probability than the latter, but the difference between the two optimum designs is small. The last row of Table 4.1 shows the number of function evaluations for the deterministic, component single loop and system single-loop optimizations. The number of function evaluations of the SORA system RBDO approach could not be determined in MS-Excel solver. Each call of the objective function or any constraint, counts as a function evaluation. The system RBDO uses more function evaluations than the component RBDO due to the active constraint set strategy. Also, the component RBDO uses more function evaluations than the deterministic optimization. However, both the component and single-loop system approaches are efficient, as evidenced by the low number of function evaluations. The optimal apportionment of the system probability of failure among the failure modes indicates the significance of each mode in the overall system probability of failure. In this beam example, the first failure mode is much more significant than the second because the failure probability of the first mode is about seven times bigger than the second. Therefore, if we want to change the problem formulation in order to reduce the system probability of failure and/or the area of the optimum design, we must allocate more resources to reduce the probability of failure of the first mode. For example if we can select a different material for the beam it is better to increase the yield stress than the Young’s modulus of the beam, because yielding is the most important mode. An important advantage of a system RBDO approach is that it allocates the system reliability among the failure modes by accounting for the cost of increasing reliability. An optimality condition relating the sensitivity derivatives of the reliabilities of the failure modes and the sensitivity derivatives of the cost (objective) function will be presented in section 4. This condition indicates that high reliability is allocated to failure modes whose reliability can be increased at low cost, that is, a small increase in the objective function is required in order to increase the reliability of these modes.

106

Structural design optimization considering uncertainties

3.2 Intern al c o mb us t io n e ng ine examp l e This example addresses a flat head design of an internal combustion engine from a thermodynamic viewpoint (Papalambros & Wilde 2000, McAllister & Simpson 2003). Design variables are the cylinder bore b, compression ratio cr , exhaust valve diameter dE , intake valve diameter dI , and the revolutions per minute (RPM) at peak power divided by 1000, ω. The goal is to obtain preliminary values for these variables that maximize the power output per unit displacement while meeting specific fuel economy and packaging constraints. The problem is stated as, Find : b, dI , dE , cr , ω, ptf1 , ptf2 , . . . , ptf9 max f = ω[3688ηt (cr , b)ηv (ω, dI ) − FMEP(cr , ω, b)]/120 FMEP = 4.826(cr − 9.2) + 7.97 + 0.253 · [8V/(πNc )] · ωb(−2) + 9.7 · 10−6 {[8V/(πNc )] · ωb(−2) }2 ηt = 0.8595(1 − cr−0.33 ) − Sv (1.5/ω)0.5 Sv = 0.83 · [8 + 4cr + 1.5(cr − 1)b3 πNc /V]/[(2 + cr ) · b] ηv = ηvb (1 + 5.96 · 10−3 ω2 )/{1 + [9.428 · 10−5 · 4V/(πNc Cs ) · (ω/dI2 )]2 }  1.067 − 0.038 e(ω−5.25) for ω ≥ 5.25 ηvb = 0.637 + 0.13 ω − 0.014 ω2 + 0.00066 ω3 for ω ≤ 5.25 subject to: P[Gi (X) ≥ 0] ≥ 1 − ptfi  Gsys = 0.006539 −

9  i=1

i = 1, 2, . . . , 9 ptfi

−

9  i=2

 max (ptfij ) j
≥0

where: (min. bore wall thickness), G1 = 400 − 1.2Nc b G2 = b − [8V/(200πNc )]0.5 (max. engine height), (valve geometry & structure), G3 = 0.82b − dI − dE (min. valve diameter ratio), G4 = dE − 0.83dI (max. valve diameter ratio), G5 = 0.89dI − dE (max. Mech Index), G6 = 0.6Cs − 9.428 · 10−5 (4V/πNc )(ω/dI2 ) (knock-limit compression ratio), G7 = − 0.045 · b − cr + 13.2 (max. torque converter RPM), G8 = 6.5 − ω G9 = 230.5Qηtw − 3.6 · 106 (max. fuel economy), ηtw = 0.8595 · (1 − cr−0.33 ) − Sv V = 1.859 · 106 (mm3 ) Q = 43,958 kJ/kg Cs = 0.44 Nc = 4

E f f i c i e n t a p p r o a c h e s f o r s y s t e m r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

107

Table 4.2 Distribution parameters and bounds of design variables.

Cylinder bore, b, mm Intake valve diameter, d I , mm Exhaust valve diameter, d E , mm Compression ratio, c r RPM at peak power/1000, ω

Standard deviation

Lower bound

Upper bound

0.4 0.15 0.15 0.05 0.25

70 25 25 6 5

90 50 50 12 12

Many of the above expressions are valid only within the limited range of bore-tostroke ratio of 0.7 ≤ b/s ≤ 1.3. More information on the problem definition can be found in (Papalambros & Wilde 2000). All design variables are assumed normally distributed with standard deviation and bounds as shown in Table 4.2. First, the deterministic optimization problem was solved. For the component singleloop approach, a target safety index of 3 (which corresponds to a failure probability of 0.00135) was assumed for each failure mode. The system probability of failure was therefore, equal to 0.00675 (=5 × 0.00135), assuming that all modes are disjoint. The assumed system probability of failure of 0.00675 was checked using Monte Carlo (MC) simulation with importance sampling. It was found that the actual probability of failure is 0.006539, instead. The latter was used as the maximum allowable system probability of failure for both the single-loop and SORA system RBDO approaches. It was also verified using MC simulation that the joint probabilities of failure for all pairs of active constraints are negligible for this example. The same initial point of (80, 35, 40, 11, 6) and convergence conditions were used in all optimizations. Table 4.3 compares the results from deterministic optimization, component singleloop RBDO, system single-loop RBDO and SORA system RBDO. In the deterministic optimization, the constraint values are given at the optimum point. For the component and system approaches, the constraint values are given at their corresponding MPP’s. Finally, in the system approaches, the active constraint values are given at their MPP points corresponding to different β values as calculated by the algorithm, while the inactive constraint values are given at their approximate MPP’s (point on β – circle closest to the limit state) corresponding to the assigned maximum β (see section 2.3.2). As shown in Table 4.3, the system single-loop problem has the same active constraint set with the component single-loop problem. Also, the deterministic optimization has the same active set excluding the sixth constraint. Table 4.3 shows that the single loop and SORA based approaches yielded practically identical designs. Table 4.3 shows the apportionment of the specified 0.006539 system probability of failure among all failure modes. The most critical mode is the sixth one followed by the third and first, with probabilities of failure of 0.002370, 0.001665 and 0.001448, respectively. The deterministic optimum has highest output power but the least reliability. The component and the system optima are very similar (see values of design variables) and have almost the same output power (50.9713 and 51.1023, respectively), and system reliability. However, the system reliability approach is superior to the component

108

Structural design optimization considering uncertainties

Table 4.3 Comparison of results for the engine design example. Design variables

Deterministic optimization

Component single-loop

System single-loop

System SORA

b dI dE cr ω pf1 pf3 pf4 pf6 pf7 psys Objective f(X) G 1 (X) G 2 (X) G 3 (X) G 4 (X) G 5 (X) G 6 (X) G 7 (X) G 8 (X) G 9 (X) No. of F.E.

83.3333 37.3406 30.9927 9.4500 6.0720

82.1333 35.8430 30.3345 9.3446 5.3141 0.00135 0.00135 0.00135 0.00135 0.00135 0.006539 50.9713 0 4.0088 0 0 0.9633 0 0 0.4359 0.0892 471

82.1419 35.8456 30.3641 9.3174 5.3598 0.001448 0.001665 0.000811 0.002370 0.000232 0.006539 51.1023 0 3.548016 0 0 0.700382 0 0 0.0968 0.0771 1290

82.1413 35.8356 30.3645 9.3170 5.3639 0.001441 0.001544 0.000722 0.002595 0.000223 0.006539 51.1013 0 3.14∗ 0 0 0.63∗ 0 0 0.01∗ 0.0476∗ N/A

55.6677 0 6.4088 0 0 2.2404 0.0211 0 0.4280 0.1179 309

∗ The values of the performance function correspond to a safety index of 4.5 for modes 2, 5, 8, and 9 for the SORA

system RBDO. Therefore, they should not be compared to the results of the single loop RBDO.

reliability approach for the following two reasons. First, the system approach allows the designer to control directly the system reliability, whereas the component reliability approach only allows the designer to control the reliabilities of the failure modes. For example, if the designer did not know that modes 2, 5, 8 and 9 were insignificant and distributed the minimum allowable reliability equally among the modes then he/ she would obtain a design with lower power output than the maximum achievable one and unnecessarily high reliability. Second, the system reliability approach helps the designer identify the significant failure modes that contribute the most to the overall system reliability. This is very important considering that if we wanted to change the problem formulation then resources should be targeted towards improving the reliability of those “critical’’ failure modes. Figure 4.6 shows the optimization histories and the last row of Table 4.3 shows the number of function evaluations for all approaches, but SORA. The computational cost of the single-loop system approach is higher than that of the component approach due to the active set strategy. In general, the more failure modes in a problem, the higher the system computational cost is expected to be. The SORA based RBDO was considerably slower than its single-loop counterpart. The SORA approach converged in two cycles involving a total of 22 iterations; the first cycle involved 15 iterations while the second involved 7 iterations.

E f f i c i e n t a p p r o a c h e s f o r s y s t e m r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

109

57 56

Objective value

55 54 53 52 51 50 49

Deterministic optimization Component single-loop System single-loop/JPDF

48 47 0

1

2

3

4

5

6

7

Iteration number

Figure 4.6 Optimization histories for the engine design example.

4 Optimality condition for series system RBDO and validation of the optimum of the beam example Consider the general system RBDO problem formulation in Eq. (1). The Lagrangian, L, of this constrained optimization problem is, L = f + λ(pall f − psys )

(20)

At the optimum, the constraint on the system safety index is usually active, that is the system failure probability assumes its maximum allowable value to minimize the objective function, f . At the optimum, the gradient of the Lagrangian is zero. Therefore, the optimality conditions become, ∂psys ∂f ∂L =0⇒ =λ ∂yk ∂yk ∂yk

(21)

where yk is the kth design variable. From Eq. (21) we obtain the following optimality criterion for the series system RBDO problem, ∂f ∂yk ∂psys ∂yk

= λ = constant

(22)

Equation (21) says that at the optimum, the iso-cost surface (loci of all designs with constant objective function, f ) and iso-reliability surface (loci of all designs with same system reliability, or equivalently system failure probability) have same slope, that is they are tangent to each other. This equation can be used to validate the optimum design obtained by the SORA or single-loop system RBDO approaches. Note that, there is no reason that the failure probabilities of the modes be equal at the optimum.

110

Structural design optimization considering uncertainties

Any deviation from system reliability optimum results in a heavier or less reliable design

4 System failure probability
0.0027

3.9

Area
9.43 in2 System reliability optimum

3.8

Component RBDO optimum

t (in)

3.7 3.6 3.5 3.4 3.3 2.35

2.45

2.55

2.65

2.75

w (in)

Figure 4.7 Optimality condition for beam design.

We will check if the optimum beam design in the first example (Table 1) satisfies the above optimality condition. Figure 4.7 compares the optimum designs obtained by the proposed system RBDO approach and the design obtained by component RBDO. It is observed that the system reliability optimum has the smallest area among all the designs with same system reliability (Pf = 0.0027, system reliability = 1 − 0.0027, system safety index = 2.782). Iso-reliability curves (curves representing designs with system reliability 1 − 0.0027) and iso-cost curves (designs with area = 9.43 in2 ) are tangent at the point representing the optimum reliability-based design. Any deviation from this optimum will result in a design with larger area or an infeasible design (violation of the minimum system reliability constraint). The design obtained by component RBDO lies on the same iso-reliability curve as the RBDO optimum but it has higher area. This shows that the proposed approach saves area by apportioning reliability in an optimal way among the failure modes of a system.

5 Conclusions Two approaches for system RBDO were presented that allow for an optimal apportionment of the reliability of a series system among its failure modes (constraints). These approaches use SORA and single-loop RBDO algorithms to determine the optimum design. The target values of the failure probabilities of the failure modes are considered as design variables. An active set strategy is used for algorithmic stability. The active constraint set is updated in each iteration during the optimization process. The first-order and second-order Ditlevsen upper bounds are used to approximate conservatively the probability of failure of a series system. The proposed algorithms ensure

E f f i c i e n t a p p r o a c h e s f o r s y s t e m r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

111

overall system reliability rather than an arbitrary reliability for each failure mode, as is the case with component RBDO methods. The user specifies an acceptable system reliability level instead of deciding arbitrarily on a minimum reliability level for each failure mode, which is usually not optimal. The efficiency and robustness of the two approaches was demonstrated on two design examples involving a beam, and an internal combustion engine. The results were compared with deterministic optimization and the conventional RBDO formulation. It was shown that both system RBDO approaches identified identical optimal designs that have the specified system reliability and provide an optimal reliability for each failure mode. In doing so, the algorithms for the two system RBDO approaches identified the critical failure modes that contributed the most to the system reliability. The single-loop system RBDO approach was found considerably more efficient than its SORA counterpart because the former approach performs only one deterministic optimization, while the latter approach performs a sequence of optimizations. Moreover, the single-loop approach avoids the extrapolation of the MPP’s of the SORA approach. The results from the examples also showed that the number of function evaluations is higher for the system approaches compared with the component approach due to the active set strategy. In general, the more failure modes are in a problem, the higher the system computational cost is expected to be.

Acknowledgments This study was performed with funding for the last two authors from the General Motors Research and Development Center and the Automotive Research Center (ARC), a U.S. Army Center of Excellence in Modeling and Simulation of Ground Vehicles at the University of Michigan. The support is gratefully acknowledged. Such support does not however, constitute an endorsement by the funding agencies of the opinions expressed in the chapter.

Nomenclature Latin symbols d: deterministic design variables f : objective function Gi (d, X, P): ith deterministic constraint (performance function of the ith failure mode of a system) P: random parameters PMPP : values of the random parameters at the Most Probable Point psys : actual system failure probability pall : maximum allowable failure probability of a system f pfi : actual failure probability of ith mode of a system ptfi : target failure probability of ith mode of a system T: Transformation from space of reduced variables (U or Z) to the space of the space of the original variables (X) UMPP : values of the vector of the random variables and the random parameters at the Most Probable Point in the reduced space (Z- or U-space)

112

Structural design optimization considering uncertainties

X: random variables XMPP : values of the random variables at the Most Probable Point in the space of the original variables (X-space) y: set of all design variables (both deterministic and random) Greek symbols α: normalized gradient of the performance function of a failure mode all βsys : minimum allowable value of the safety index of a system βi : actual value of the safety index of the ith mode of a system βit : target value of the safety index of the ith mode of a system µX : mean values of random variables µP : mean values of random parameters ∇: gradient operator. A b b rev i a ti o ns CF: Constraint Flag DLP: Double Loop FORM: First Order Reliability Method IC: Internal Combustion KKT: Karush-Kuhn-Tucker MPP: Most Probable Point PMA: Performance Measure Approach RBDO: Reliability-Based Design Optimization SLSV: Single-Loop, Single-Vector SORA: Sequential Optimization and Reliability Assessment.

References Agarwal, H., Renaud, J.E., Lee, J.C. & Watson, L.T. 2004. A Unilevel Method for Reliability Based Design Optimization. 45th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference. Palm Springs, CA. Ba-abbad, M., Nikolaidis, E. & Kapania, R. 2006. A New Approach for System ReliabilityBased Design Optimization. AIAA Journal, Vol. 44, No. 5, pp. 1087–1096. Chen, X., Hasselman, T.K. & Neill, D.J. 1997. Reliability Based Structural Design Optimization for Practical Applications. Proceedings of the 38th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference. Ditlevsen, O. 1979. Narrow Reliability Bounds for Structural Systems. Journal of Structural Mechanics 3:453–472. Du, X. & Chen, W. 2004. Sequential Optimization and Reliability Assessment Method for Efficient Probabilistic Design. ASME Journal of Mechanical Design 126(2):225–233. Frangopol, D.M. & Maute, C. 2004. Reliability-Based Optimization of Civil and Aerospace Structural Systems. Engineering Design Reliability Handbook. Chapter 24, CRC Press, Boca Raton, Florida. Haldar, A. & Mahadevan, S. 2000. Probability, Reliability and Statistical Methods in Engineering Design. John Wiley & Sons, Inc. Kuschel, N. & Rackwitz, R. 2000. Optimal Design under Time-Variant Reliability Constraints. Structural Safety 22(2):113–128.

E f f i c i e n t a p p r o a c h e s f o r s y s t e m r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

113

Liang, J., Mourelatos, Z.P. & Tu, J. 2004. A Single-Loop Method for Reliability-Based Design Optimization. International Journal of Product Development. Interscience Enterprises Limited, Great Britain (in press). Liang, J., Mourelatos, Z.P. & Nikolaidis, E. 2007. A Single-Loop Approach for System Reliability-Based Design Optimization. ASME Journal of Mechanical Design (accepted). Madsen, H.O., Krenk, S. & Lind, N.C. 1986. Methods of Structural Safety. Prentice-Hall, Inc. McAllister, C.D. & Simpson, T.W. 2003. Multidisciplinary Robust Design Optimization of an Internal Combustion Engine. ASME Journal of Mechanical Design 125(1):124–130. Melchers, R.E. 2001. Structural Reliability Analysis and Prediction. John Wiley & Sons. Moré, J.J. & Sorensen, D.C. 1983. Computing a Trust Region Step. SIAM Journal on Scientific and Statistical Computing, Vol. 3, pp. 553–572. Moses, F. 1995. Probabilistic Analysis of Structural Systems. Probabilistic Structural Mechanics Handbook: Theory and Industrial Applications, edited by C. Raj Sundararajan, Chapman & Hall, 166–187. Papalambros, P.Y. & Wilde, D.J. 2000. Principles of Optimal Design; Modeling and Computation. 2nd Edition, Cambridge University Press. Royset, J.O., Der Kiureghian, A. & Polak, E. 2001. Reliability-based Optimal Design of Series Structural Systems. Journal of Engineering Mechanics 607–614. Sørensen, D.C. 1994. Minimization of a Large Scale Quadratic Function Subject to an Ellipsoidal Constraint. Department of Computational and Applied Mathematics, Rice University. Technical Report TR94-27. Steihaug, T. 1983. The Conjugate Gradient Method and Trust Regions in Large Scale Optimization. SIAM Journal on Numerical Analysis, Vol. 20, pp. 626–637. Streicher, H. & Rackwitz, R. 2004. Time-Variant Reliability-Oriented Structural Optimization and a Renewal Model for Life-cycle Costing. Probabilistic Engineering Mechanics 19(1–2): 171–183. Thanedar, P.B. & Kodiyalam, S. 1992. Structural Optimization Using Probabilistic Constraints. Structural Optimization 4:236–240. Tu, J., Choi, K.K. & Park, Y.H. 1999. A New Study on Reliability-Based Design Optimization. ASME Journal of Mechanical Design 121:557–564. Wu, Y.-T., Shin, Y., Sues, R. & Cesare, M. 2001. Safety – Factor Based Approach for Probabilistic – Based Design Optimization. 42nd AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference. Seattle, WA. Youn, B.D., Choi, K.K. & Park, Y.H. 2003. Hybrid Analysis Method for Reliability-Based Design Optimization. ASME Journal of Mechanical Design 125:221–232.

Chapter 5

Nondeterministic formulations of analytical target cascading for decomposition-based design optimization under uncertainty Michael Kokkolaras & Panos Y. Papalambros University of Michigan, Ann Arbor, MI, USA

ABSTRACT: Analytical target cascading (ATC) is a methodology for optimal design of hierarchically decomposed systems. In this chapter, we present non-deterministic formulations of ATC to account for uncertainties in decomposition-based optimal system design. Depending on the amount of available information, we adopt either a probabilistic or a robust optimization approach to formulate the multilevel design optimization problem, and use appropriate techniques to estimate uncertainty propagation. We demonstrate the application of all presented ATC formulations using an engine optimal design example, and discuss the obtained results.

1 Introduction The dictionary definition of a system is “an organized integrated whole made up of diverse but interrelated and interdependent parts,’’ and complex is one of its synonyms (Merriam-Webster 2007). It is not surprising then that developing large engineering systems is accomplished by assigning the task of designing the diverse but interrelated parts to different teams, and that the challenge is to organize these activities so that the parts can be integrated successfully to form the whole. Accordingly, large engineering systems are typically decomposed into subsystems, subsystems are decomposed into components, components are decomposed into parts, and so on. This results in a multilevel hierarchy, an example of which is shown in Figure 5.1. Different teams (or individuals) are then assigned with the optimal design Level i = 0

Level i = 1

System j = 1

Subsystem j = 1

Level i = 2 Component j = 1

Subsystem j = 2

Component j = 2

Subsystem j = n

Component j = m

Figure 5.1 Example of a hierarchically decomposed multilevel system.

116

Structural design optimization considering uncertainties

problem of each element in this hierarchy. If these design teams are not given exact specifications, they focus on their own objectives without taking into consideration interactions with other elements. This situation is compounded when design variables are shared among elements; if the obtained values of shared design variables are not equal in all elements, the system design is inconsistent and cannot be realized. Hierarchical decomposition facilitates the use of decentralized optimization approaches that aid systems engineers to identify interactions among elements at lower levels and to transfer this information to higher levels, and has become standard design practice, as evidenced by the organizational structure of engineering companies (Haimes et al. 1990). Analytical target cascading (ATC) is a methodology for solving such hierarchical multilevel optimal design problems. Design targets are cascaded to lower levels using the model-based hierarchy. An optimization problem is posed and solved for each design subproblem to minimize deviations from propagated targets. Solving the subproblems according to an appropriate coordination strategy yields overall system compatibility. The deterministic formulation of the ATC methodology assumes that complete information of the system design problem is available, and that design decisions can be implemented precisely. These assumptions imply that optimization results are as good (and therefore useful) as the design and simulation/analysis models used to obtain them, and that they are meaningful only if they can be realized exactly. In reality, these assumptions do not hold. We are rarely in a position to represent a physical system without using approximations, have complete knowledge on all of its parameters, or control the design variables with high accuracy. Therefore, uncertainty is inherently present in simulation-based design of complex engineering systems. The analysis models used for the simulation depend on assumptions and include many approximations and empirical constants. Also, advanced yet relatively immature technologies are often associated with uncertainty. The designer is not sure about the validity of the decisions he/she has made, and would like to be able to perform optimization studies under uncertainty. It is therefore imperative to represent uncertainties and take them into account during the early design assessment process. Uncertainty identification, representation, and quantification are the cornerstones of design optimization under uncertainty. Given the design model and the necessary analysis/simulation models, the designer must first identify all possible sources of uncertainty. Then, she/he must choose an appropriate means to represent and quantify them. A popular approach is to represent them as random variables, and quantify them by means of some probability distribution utilizing expertise and data. This approach is useful when there are sufficient data to infer probability distributions for the considered random variables. It should be adopted since a plethora of techniques exists for solving probabilistically formulated optimal design problems. However, in many situations the designer does not have the necessary information available. In this case, he/she must assume that the uncertain quantities can take any value within intervals that are used to quantify uncertainty. In this chapter, we review the ATC methodology, and we extend its deterministic formulation using both probabilistic and interval analysis approaches. We address the issue of representing uncertain quantities as optimization variables, formulate the

Nondeterministic formulations of analytical target cascading

117

associated nondeterministic design problems appropriately, and present techniques for estimating uncertainty propagation through the multilevel hierarchy of decomposed systems. The proposed methodologies are applied to a simple engine design example to illustrate the introduced concepts.

2 Analytical target cascading Analytical target cascading (ATC) is a mathematical methodology for translating (“cascading’’) overall system design targets to element specifications based on a hierarchical multilevel decomposition (Michelena et al. 1999; Papalambros 2001; Kim 2001; Kim et al. 2003). The objective is to assess interactions and identify possible tradeoffs among elements early in the design development process, and to determine specifications that yield consistent system design with minimized deviation from system design targets. For an engineering corporation, ATC provides a means to dictate technical objectives to different design teams, knowing a-priori that these goals can be achieved without conflicting with those of other teams. Consistent system design can then be accomplished with minimum communication overhead, i.e., maximum efficiency, avoiding costly iterations late in the process. ATC operates by formulating and solving a minimum deviation optimization problem for each element in the hierarchy. Assuming that responses of higher level elements are functions of responses of lower-level elements, it aims at minimizing the gap between what upper-level elements “want’’ and what lower-level elements “can.’’ Similarly, if design variables are shared among some elements at the same level, their consistency is coordinated by their common parent element at the level above. The ATC process is proven to be convergent when using a specific class of coordination strategies (Michelena et al. 2003), and has been successfully applied to a variety of optimal design problems, e.g., (Kim et al. 2002; Kokkolaras et al. 2002; Kim et al. 2003). We refer the reader to the above references for a detailed description of ATC. Here, we present the concept and the general mathematical formulation. The key assumption of the ATC methodology is that there is a functional dependency in the hierarchical, multilevel system decomposition. Assuming that element j at level i has nij children, this functional dependency is expressed as rij = fij (r(i+1)1 , . . . , r(i+1)nij , xij , yij )

(1)

where rij are element responses, r(i+1)1 , . . . , r(i+1)nij denote children responses, xij represent local design variables, and yij denote local shared design variables (i.e., design variables that this element shares with other elements at the same level). The mathematical formulation of problem pij for element j at level i is min rij (r(i+1)1 , . . . , r(i+1)nij , xij , yij ) − riju 22 + yij − yiju 22 + nij nij l l r(i+1)k − r(i+1)k 22 + k=1 y(i+1)k − y(i+1)k 22 k=1

(2)

with respect to r(i+1)1 , . . . , r(i+1)nij , xij , yij , y(i+1)1 , . . . , y(i+1)nij subject to gij (rij , xij , yij ) ≤ 0 where coordinating variables for the shared design variables of the children are denoted by y(i+1)1 , . . . , y(i+1)nij , and superscripts u (l) are used to denote response and shared

118

Structural design optimization considering uncertainties

Optimization inputs Response and shared variable values cascaded down from the parent

Optimization outputs

ruij

rlij

yuij

ylij

Response and shared variable values passed up to the parent

Element optimization problem pij, where rij is provided by the analysis/simulation model rij
fij (r(i1)k1,..., r(i1)kc ,xij, yij) ij

Response and shared r1(i1)k ,..., r1(i1)k 1 cij variable values passed 1 1 up from the children y(i1)k ,..., y(i1)k 1

cij

ru(i1)k1,..., ru(i1)kcij yu(i1)k1,..., yu(i1)kc

ij

Response and shared variable values cascaded down to the children

Figure 5.2 ATC information flow at element j of level i.

variable values that have been obtained at the parent (children) problem(s), and have been cascaded down (passed up) as design targets (consistency parameters). Shared design variables are restricted to exist only among elements at the same level having the same parent. The top-level problem of the hierarchy is a special case: at this level (i = 0), there is only one element (j = 1 – the system), and responses cascaded from above are the given system design targets T = Ru01 (there is no parent element); also, since this is the sole element of the level, there exist no shared variables. The bottom-level problems are also a special case since they have no children. Finally, note that although communication among levels, i.e., updating parameter values associated with the ATC process, is bi-directional, functional dependency is strictly hierarchical. Figure 5.2 illustrates the information flow of the ATC process at element j in level i. Assuming that all the parameters have been updated using the solutions obtained at the parent- and children-problems, Problem (2) is solved to update the parameters of the parent- and children-problems. This process is repeated until the variables in all optimization problems do not change significantly after consecutive iterations. The sequence in which the subproblems are solved is called a coordination strategy. As in any distributed multidisciplinary optimization methodology, the choice of coordination strategy among the many available alternatives is critical. In contrast to other methodologies for multilevel system design, global convergence properties have been proven for a specific class of coordination strategies under standard smoothness and convexity assumptions (Michelena et al. 2003). Nevertheless, case studies have also demonstrated that the ATC process may terminate successfully in practice when other coordination schemes are used (Michelena et al. 2001; Kim et al. 2002; Louca et al. 2002; Kokkolaras et al. 2004). It is emphasized that ATC should not be viewed either solely or merely as a design optimization methodology. ATC addresses the early part of the product development process (cf. Figure 5.3). Its purpose is to account for the interrelations of the system

119

ifica tion

s

Nondeterministic formulations of analytical target cascading

Par ts s pec

Analytical target cascading process Part 1

Enterprise target setting

Design targets

Part 2 Part specifications feasible?

System

Part n

Yes Design targets feasible?

No

No

Part 1 design embodiment Part 2 design embodiment

Yes

Final system design

Part n design embodiment

Figure 5.3 ATC in the product development process. Brake-specific fuel consumption

Engine simulation Power loss due to friction Piston-ring/cylinder-liner subassembly

Ring and liner surface roughness

Oil consumption blow-by liner wear rate

Liner material properties (Young’s modulus and hardness)

Figure 5.4 Hierarchical bi-level system.

parts, identify possible tradeoffs, and determine optimal and consistent design specifications to match design targets as close as possible (i.e., it can also be used to check whether the given design targets can be achieved by the available means). Once this is accomplished, the design embodiment for each part can be carried out concurrently or outsourced.

3 Application to engine design In this section, we apply the ATC methodology to a simple yet illustrative simulationbased optimal design example to demonstrate the introduced concepts. Specifically, we consider a V6 gasoline engine as the system, which is decomposed into six subsystems, each of which represents the piston-ring/cylinder-liner subassembly of a single cylinder. The system simulation predicts engine performance in terms of brake-specific fuel consumption. Although the engine has six cylinders, they are all designed to be identical. For this reason, we can actually consider only one subsystem. The associated bi-level hierarchy, shown in Figure 5.4, includes the engine as a system at the top level and the piston-ring/cylinder-liner subassembly as a subsystem at

120

Structural design optimization considering uncertainties

the bottom level. The ring/liner subassembly simulation takes as inputs the surface roughness of the ring and the liner and the Young’s modulus and hardness and computes power loss due to friction, oil consumption, blow-by, and liner wear rate. The engine simulation takes then as input the power loss and computes brake-specific fuel consumption of the engine. Commercial software packages were used to perform the simulations. A detailed description of the problem can be found in (Chan et al. 2004). Due to the simplicity of the given problem structure, we use a simplified version of the notation introduced earlier. Since there are only two levels with only one element in each, we skip element indices and denote the upper-level element with subscript 0 and the lower-level element with subscript 1. We use second indices to denote the components of the design variable vector of the lower-level element optimization problem. The design problem is to find optimal values for the piston-ring and cylinder-liner surface roughness design optimization variables x11 and x12 , respectively, and optimal values for the design optimization variables representing the material properties (Young’s modulus x13 and hardness x14 ) of the liner that yield minimized brake-specific fuel consumption, i.e., system response R0 . The optimal design problem includes constraints on liner wear rate, oil consumption, and blow-by. The power loss due to friction, i.e., subsystem response R1 , links the two levels. The top- and bottom-level ATC problems are formulated as min (R0 (R1 ) − T)2 + (R1 − Rl1 )2 R1

(3)

and min

x11 ,x12 ,x13 ,x14

subject to

(R1 (x11 , x12 , x13 , x14 ) − Ru1 )2 liner wear rate = g11 (x11 , x12 , x13 , x14 ) ≤ 2.4 × 10−12 m3 /s blow-by = g12 (x11 , x12 , x13 , x14 ) ≤ 4.25 × 10−5 kg/s oil consumption = g13 (x11 , x12 , x13 , x14 ) ≤ 15.3 × 10−3 kg/hr 1 µm ≤ x11 , x12 ≤ 10 µm 80 GPa ≤ x13 ≤ 340 GPa 150 BHV ≤ x14 ≤ 240 BHV

(4)

respectively. The fuel consumption target T was set to zero to achieve the best fuel economy possible. 3.1

D eterm i nis t ic d es ig n o pt imizat io n r e s u l t s

It is desired to minimize power loss due to friction in order to optimize engine operation and thus maximize fuel economy. Therefore, it was anticipated that the bottom-level optimization problem would yield a design with as smooth surfaces (low surface roughnesses) as possible without violating the bounds or the nonlinear design constraints. The ATC process of solving Problems (4) and (3) iteratively converged after two iterations. The obtained deterministic optimal ring/liner subassembly design is shown in Table 5.1. The ring surface roughness and the liner’s Young’s modulus optimal values are at their lower bounds; the liner surface roughness and hardness have interior optimal values. Figure 5.5 shows the two-dimensional projection of the design space

Nondeterministic formulations of analytical target cascading

121

Table 5.1 Deterministic optimal ring/liner subassembly design. Variable

Description

Value

µX 11 µX 12 x 13 x 14

Ring surface roughness, [µm] Liner surface roughness, [µm] Liner Young’s modulus, [GPa] Liner hardness, [BHV]

1.0 3.5 80 175

10 0.486

0.48

0.475

9

Liner wear rate
2.4  1012

SFC

asing B

Decre

7

0.485

0.48

6

49

0.475

Liner surface roughness [m]

8

5 Optimal design 4 Oil consumption
15.3 3 0.48

2

1

BowBy
4.5  105

1

2

3

5 7 4 6 Ring surface roughness [m]

8

9

10

Figure 5.5 Two-dimensional projection of the design space.

spanned by the two surface roughness variables when the liner Young’s modulus and the liner hardness are kept fixed at 80 GPa and 175 BHV, respectively. The liner surface roughness is not at its lower bound because the oil consumption constraint is active: increased liner surface roughness is required to maintain an optimal oil film thickness in order to avoid excessive oil consumption.

4 The probabilistic ATC formulation In this section, the ATC formulation is extended to account for uncertainties. Adopting a probabilistic framework, we model uncertain quantities as random variables

122

Structural design optimization considering uncertainties

(denoted by upper case Latin symbols). In general, we use the terms random design optimization variable and random design optimization parameter to differentiate between random variables that are design optimization variables and random variables that are design optimization parameters in the optimization problems. Here, to avoid confusion, and without loss of generality, we assume that all design optimization parameters are deterministic, and we omit them in the mathematical formulations. We use the means of random design variables as optimization variables and assume that their standard deviation is known or has been estimated with sufficient accuracy. The objective and the constraints must be reformulated. We replace the objective function with its expectation, and we now require that the probability of violating a constraint is less than some pre-specified probability of failure. The probabilistic formulation of Problem (2) is (Kokkolaras et al. 2006) min E[Rij ] − µuRij 22 + µYij − µuYij 22 + nij nij µR(i+1)k − µlR(i+1)k 22 + k=1 µY(i+1)k − µlY(i+1)k 22 k=1 with respect to subject to with

µR(i+1)1 , . . . , µR(i+1)nij , µXij , µYij , µY(i+1)1 , . . . , µY(i+1)nij P[gijk (Rij , Xij , Yij ) > 0] ≤ Pfk , k = 1, 2, . . . , Mij Rij = fij (R(i+1)1 , . . . , R(i+1)nij , Xij , Yij )

(5)

where Mij is the number of design constraints, P[ · ] denotes probability measure, and Pfk is a pre-specified probability of failure for design constraint k. Liu et al. considered more than one moments to represent random variables in the ATC optimization problems (Liu et al. 2006). 4.1

U n c erta i n t y pr o pag at io n

In a multilevel hierarchy, responses (outputs) of lower-level elements are inputs to higher-level elements. This is an issue of utmost importance in design optimization of hierarchically decomposed systems under uncertainty, since the solution of probabilistic optimization problems requires moment estimation of high-level random optimization variables that are functions of low-level random optimization variables. In other words, we need appropriate techniques for uncertainty propagation. Consider element j at level i. By solving Problem (5), we obtain optimal values µ∗R(i+1)1 , . . . , µ∗R(i+1)n , µ∗Xij , and µ∗Yij . Using the functional dependency relation ij

Rij = fij (R(i+1)1 , . . . , R(i+1)nij , Xij , Yij ), we must now estimate the moments (typically the first two, mean and standard deviation) of the responses Rij since the latter constitute random optimization variables of the parent probabilistic optimal design problem. This needs to be done for all problems at all levels of the hierarchy. An efficient and accurate technique is therefore required for propagating uncertainties through the multilevel hierarchy. We assume that all element responses in the multilevel hierarchy are uncorrelated. Many probabilistic design methods and software packages use a first-order Taylor expansion about the current mean design to estimate the mean and standard deviation of propagated random responses. We have found that while the mean values can be estimated relatively accurately, standard deviation estimates are unacceptably inaccurate in may cases (Youn et al. 2004; Kokkolaras et al. 2004). Thus, we propose

Nondeterministic formulations of analytical target cascading

123

an uncertainty propagation technique we developed based on the highly efficient and accurate Advanced Mean Value (AMV) method (Wu et al. 1990). The AMV method has been originally proposed as a computationally efficient method for generating the cumulative distribution function (CDF) of a response R = f (X) that is a random variable (Wu et al. 1990). It uses a simple correction to compensate for errors introduced by a utilized Taylor series approximation. Based on the CDF definition, we have the following first-order relation between the CDF value of R at a particular value f0 and the reliability index β: P[f ≤ f0 ] = P[g ≤ 0] =
(−β)

(6)

where g(X) = f (X) − f0 and
is the standard normal cumulative distribution function. According to the AMV method, if the random variables X are uncorrelated and normally distributed with means µX and standard deviations σX , the most probable point (MPP) of failure (or design point) in the standard normal space can be computed by U∗ = −βX

∇glin (µX ) ∇f (µX ) = −βX |∇glin (µX )| |∇f (µX )|

(7)

where glin (X) is a linear approximation of g(X) at µX and X is a diagonal matrix, whose diagonal is the vector σX . In the original space the MPP coordinates are X ∗ = x U ∗ + µ x

(8)

Note that for random variables that are not normally distributed, a nonlinear transformation is needed according to the Rackwitz-Fiessler method (Haldar and Mahadevan 2000). The AMV method corrects the CDF value of R in Equation (6) with P[f ≤ f (X∗ )] =
(−β)

(9)

by replacing the f0 value corresponding to the reliability index β with f (X∗ ). The process of Equations (6) through (9) is repeated for a few (different) β values, so that a region of the CDF of R is constructed. The derivative of that CDF region provides the corresponding probability density function (PDF) value. The obtained CDF and PDF values are finally used to compute equivalent mean and standard deviation at the current design point. This AMV-based technique is used to estimate the mean and standard deviation of each response for all the elements of the multilevel hierarchy according to the discussion in Section 4.1. The technique is computationally efficient since it requires only a single linearization of the performance function at the mean value and an additional function evaluation at each required CDF level. Reference (Wu 1994) provides more details regarding the accuracy and efficiency of the AMV method on several applications. 4.1.1 Illustra tive exampl es The linearization (or MVFOSM-based or method of moments) and AMV-based techniques were used to estimate the first two moments of several nonlinear functions. All

124

Structural design optimization considering uncertainties

random variables were assumed to be normal. Test functions and input statistics are presented in Table 5.2 and results are summarized in Table 5.3. One million samples were used for the Monte Carlo simulations. By inspecting Table 5.3, it can be seen that while the mean-related errors of the linearization approach are within acceptable limits, standard deviation errors can be quite large. The AMV-based moment estimation method performs always better, and never exhibits unacceptable errors. 4.2 Pro b ab i l i st ic e ng ine d es ig n We now apply the probabilistic ATC methodology to our bi-level engine design problem. Here, the root mean square (RMS) of asperity height is used to represent asperity roughness, which is assumed to be normally distributed. Thus, the surface roughness design variables are now normal random design optimization variables. The probabilistic formulation of the top- and bottom-level ATC problems are min (E[R0 ] − T)2 + (µR1 − E[R1 ]l )2 µR1

with

(10)

R0 = f0 (R1 )

Table 5.2 Test functions and input statistics. #

Function

Input statistics

1 2

X 21 + X 22 −exp(X 1 − 7) − X 2 + 10

X 1 ∼ N(10,2), X 2 ∼ N(10,1) X 1,2 ∼ N(6,0.8)

3

1−

4

1−

5

1−

X21 X2 20 (X1 + X2 − 5)2 30 80 2 X1 + 8X2 + 5

X 1,2 ∼ N(5,0.3) −

(X1 − X2 − 12)2 30

X 1,2 ∼ N(5,0.3) X 1,2 ∼ N(5,0.3)

Table 5.3 Estimated moments and errors relative to Monte Carlo simulation (MCS) results. #

1

µlin µAMV µMCS lin [%] AMV [%]

200.0 203.4 205.0 −2.44 −0.78

σlin σAMV σMCS lin [%] AMV [%]

44.72 45.20 45.10 −0.84 0.22

2

3

4

5

3.6321 3.6029 3.4921 4.00 3.17

−5.25 −5.3495 −5.3114 −1.15 0.71

−1.0333 −1.0380 −1.0404 −0.68 −0.23

−0.1428 −0.1454 −0.1448 −1.30 0.41

1.9386 0.9013 0.9327 107.85 −3.36

0.8385 0.8423 0.8407 −0.26 0.19

0.1166 0.1653 0.1653 29.46 0

0.00627 0.00631 0.00630 −0.47 0.15

Nondeterministic formulations of analytical target cascading

125

and min

µX11 ,µX12 ,x13 ,x14

subject to

with

(E[R1 ] − µuR1 )2

P[liner wear rate = G11 (X11 , X12 , x13 , x14 ) > 2.4 × 10−12 m3 /s] ≤ Pf P[blow-by = G12 (X11 , X12 , x13 , x14 ) > 4.25 × 10−5 kg/s] ≤ Pf P[oil consumption = G13 (X11 , X12 , x13 , x14 ) > 15.3 × 10−3 kg/hr] ≤ Pf P[X11 < 1 µm] ≤ Pf , P[X11 > 10 µm] ≤ Pf P[X12 < 1 µm] ≤ Pf , P[X12 > 10 µm] ≤ Pf 80 GPa ≤ x13 ≤ 340 GPa 150 BHV ≤ x14 ≤ 240 BHV R1 = f1 (X11 , X12 , x13 , x14 )

(11)

respectively. The standard deviation of the surface roughnesses was assumed to be 1.0 µm, and remained constant throughout the ATC process. The assigned probability of failure Pf was 0.13%, which corresponds to the target reliability index β = 3. The fuel consumption target T was simply set to zero to achieve the best fuel economy possible. Note that since the random variables are normally distributed, the associated linear probabilistic bound constraints are reformulated as deterministic. For example, P[X11 < 1 µm] ≤ Pf ⇔ P[X11 − 1 µm < 0] ≤ Pf ⇔ 

µX − 1 µm µX11 − 1 µm ≤
( − β) ⇒ − 11
0− ≤ −β ⇔ σX11 σX11 µX11 − 1 µm ≥ β ⇔ µX11 − 1 µm ≥ βσX11 ⇔ σX11 µX11 ≥ 1 µm + βσX11 ⇔ µX11 ≥ 4 µm Similarly, the other three probabilistic bound constraints in Problem (4) are reformulated as µX11 ≤ 7 µm;

µX12 ≥ 4 µm;

µX12 ≤ 7 µm

The obtained probabilistic optimal ring/liner subassembly design is shown in Table 5.4. The ring surface roughness optimal value is at its probabilistic lower Table 5.4 Probabilistic optimal ring/liner subassembly design. Variable

Description

Value

µX11 µX12 x 13 x 14

Ring surface roughness, [µm] Liner surface roughness, [µm] Liner Young’s modulus, [GPa] Liner hardness, [BHV]

4.00 6.15 80 240

126

Structural design optimization considering uncertainties Table 5.5 Reliability analysis results. Constraint

Active

Pf

MCS Pf

Liner wear rate Blow-by Oil consumption

No No Yes

<0.13% <0.13% 0.13%

0% 0% 0.16%

Table 5.6 Estimated moments and errors relative to Monte Carlo simulation (MCS). Response

Power loss [kW]

Fuel consumption [kg/kWhr]

µlin µAMV µMCS lin [%] AMV [%] σlin σAMV σMCS lin [%] AMV [%]

0.3950 0.3922 0.3932 0.45 −0.25 0.0481 0.0309 0.0311 54.6 −0.64

0.5341 0.5431 0.5432 −0.01 −0.01 0.00757 0.00760 0.00759 −0.25 0.13

minimum, while the liner’s Young’s modulus and hardness optimal values are at their deterministic lower and upper bounds, respectively. The liner surface roughness variable has an interior optimal value because the oil consumption constraint is probabilistically active. Constraint activity in probabilistic design optimization indicates that the constraint’s MPP lies on the target reliability circle. The probabilistic optimal values of the surface roughness optimization variables have changed relative to their deterministic counterparts to accommodate the uncertainty, i.e., the optimum shown in the two-dimensional projection of the design space (Figure 5.5) moved to the inside (we cannot show the location of the probabilistic optimum in the same figure because it lies in a different two-dimensional projection of the design space due to the change in the liner hardness optimal value). A Monte Carlo simulation was performed to assess the accuracy of the reliability analyses of the probabilistic constraints. One million samples were generated using the mean and standard deviation values of the design variables, and the constraints were evaluated using these samples to calculate the probability of failure. Results are summarized in Table 5.5. The obtained design is 0.03% less reliable than found for the active probabilistic constraint. This error is due to the first-order reliability approximation used in the probabilistic optimization problem. Propagation of uncertainty was modeled using the AMV-based technique described in Section 4.1. Table 5.6 summarizes the estimated moments for the two responses of the bi-level hierarchy. Results obtained using the first-order approximation approach (linearization) are included to illustrate the large error that may be introduced. Specifically, it can be seen that the standard deviation estimate of the power loss (necessary for solving the top-level probabilistic optimization problem) is 0.0481 kW

Nondeterministic formulations of analytical target cascading

20

20

18

18

16

16

14

14

12

12

10

10

8

8

6

6

4

4

2

2 0 0.15

0

0.2

0.25

0.3

0.35 (a)

0.4

0.45

0.5

0.2

0.25

0.3 0.35 (b)

0.4

0.45

127

0.5

Figure 5.6 Power loss uncertainty: (a) PDF obtained using the AMV-based technique and (b) frequency diagram obtained using Monte Carlo simulation.

when using a first-order approximation. This value is 54.6% larger than the Monte Carlo simulation estimate of 0.0311 kW. Such large errors will be propagated during the ATC process and yield useless design results. Using the AMV-based approach, we obtained an estimate of 0.0309 kW, which is only 0.64% smaller than the Monte Carlo estimate. Using the AMV-based technique is advantageous because CDFs and PDFs can be generated with high efficiency. In our example, power loss (the subsystem response) is a highly nonlinear function of the subsystem’s inputs. In fact, its PDF is multimodal, as shown in Figure 5.6. This figure depicts a) the PDF obtained using the AMV-based technique and b) the frequency diagram generated from a histogram that was obtained using Monte Carlo simulation with one million samples. The agreement is quite satisfactory and illustrates the usefulness of the AMV-based approach to propagate uncertainty for highly nonlinear functions.

5 The ATC formulation for interval uncertainty quantification The probabilistic approach is very useful and should be adopted when the designer has sufficient data to model uncertain quantities as random variables with appropriate probability distributions. When this is not the case, it is imperative to assume that the uncertain quantities can take any value within a range. Note that this not equivalent to assuming a uniform distribution as it does not imply that the probability of taking a specific value in a range is equal to any other value within that range. We view the interval analysis approach as a special case of possibility theory (Dubois and Prade 1988), where information availability is limited to a minimum. Designs obtained using possibility-based design optimization (PBDO) methods are typically conservative compared to the ones obtained using probabilistic design optimization, also known as reliability-based optimization (RBDO), methods. Possibility-based designs sacrifice

128

Structural design optimization considering uncertainties

additional optimality compared to RBDO designs to account for lack of uncertainty information and avoid constraint violation. According to possibility theory, the possibility π(A) of event A occurring provides an upper bound on the probability P(A) of that event occurring, i.e., P(A) ≤ π(A). From the design point of view, we can conclude that what is possible may not be probable, and what is impossible is also improbable. If the possibility of violating a constraint is zero, then the probability of violating the same constraint will also be zero. If feasibility of a constraint g is formulated in negative null form (g ≤ 0), the constraint is always satisfied if π(g > 0) = 0. By introducing the notion of membership functions and α-cuts, we can relax this requirement as π(g > 0) ≤ α, provided that 0 < α  1 (Zadeh 1978). It can be shown that if the maximum possibly attainable value of the constraint g α at the corresponding α-cut is less than or equal to zero, i.e., gmax ≤ 0, the possibility of violating this constraint is less than α (Mourelatos and Zhou 2005). In general, membership functions express how ranges of values that bound the uncertainty quantities are decreased with increasing amount of information. The α-cuts denote levels of information, starting at the lowest (α = 0), where the range is largest, and increasing to the highest (α = 1), where the range is the smallest (possibly a crisp value). In this work, we will assume that the lowest level of information is available, where α is equal to zero. Therefore, we do not have to consider membership functions and higher α-cuts, eliminating thus ad-hoc selections, but also maximizing the conservative nature of the obtained designs. Given an interval uncertainty in a design variable X, the process of identifying the maximum attainable value gmax of a constraint g(X) requires the solution of an optimization problem. Given a nominal value XN for the design variable X, we first identify the uncertainty interval [(1 − δX )XN , (1 + δX )XN ], where δX denotes the relative deviation from the nominal value XN . Then, we solve the simple bound-constrained problem max g(x) x

subject to

(1 − δX )XN ≤ x ≤ (1 + δX )XN

(12)

to compute gmax . In a design optimization problem with many constraints where design variables are subject to interval uncertainty, finding the optimal design involves a nested optimization process known as robust optimization. An outer-loop optimization generates a sequence of iterates of nominal value vectors XN for the uncertain design variables X. For each iterate XN , an inner-loop optimization problem like the one formulated in Equation (12) is solved for each constraint. These worst-case optimization problems (also referred to as “anti-optimization’’ problems (Elishakoff et al. 1994)) may involve a larger number of optimization variables, but are only bound-constrained. The primary purpose of solving these problems is to obtain the maximal (worst) value of each constraint g that may be attained due to the uncertainty in X. These constraint values are used in the outer-loop optimization, where the worst objective value is maximized and the worst constraint value must be feasible. Nevertheless, the ∗ inner-loop optimal values XN can be used to attempt to control uncertainty, i.e., what values to strive for and what values to avoid, if possible. The ATC formulation for design variables and parameters that are subject to interval uncertainty is a straightforward application of the robust optimization problem

Nondeterministic formulations of analytical target cascading

129

formulation. The implication of dealing with intervals is that two values must be matched for each uncertain quantity that links two elements: the “worst-case’’ value (computed solving a maximization problem of the form presented in Equation (12)), and the “best-case’’ value (computed by solving a minimization problem). The ATC formulation for interval uncertainties is min Rijw − Rijuw 22 + Rijb − Rijub 22 + Yijw − Yijuw 22 + Yijb − Yijub 22 + nij nij l l R(i+1)kw − R(i+1)k 22 + k=1 R(i+1)kb − R(i+1)k 2 + k=1 w b 2  nij n ij l l Y(i+1)kw − Y(i+1)k 2 + k=1 Y(i+1)kb − Y(i+1)k 2 k=1 w 2 b 2 with respect to R(i+1)1N , . . . , R(i+1)nij N XijN , YijN , Y(i+1)1N , . . . , Y(i+1)nij N subject to gijmax (R(i+1)1N , . . . , R(i+1)nij N XijN , YijN ) ≤ 0 with {Rijw , Rijb } = fij (R(i+1)1N , . . . , R(i+1)nij N XijN , YijN )

(13)

5.1 ATC-based optimization res ults The ATC process for design optimization problems with interval uncertainty variables is illustrated in this section using the same engine design problem (Kokkolaras et al. 2006). As in the probabilistic case, the considered uncertain quantities are ring and liner surface roughnesses; root mean square (RMS) of asperity height is used to represent and quantify surface roughness. Here, let us assume that we do not have sufficient data to infer that surface roughness is normally distributed. Instead, we assume that it exhibits deviations from nominal values that can be quantified by an interval. This surface roughness interval uncertainty is propagated through the simulation hierarchy to estimate intervals for power loss and fuel consumption. Since uncertainty information is available at the bottom-level we first formulate and solve the bottom-level problem min

X11N ,X12N ,x13 ,x14

subject to

(R1w − Ru1w )2 + (R1b − Ru1b )2

(14)

max. liner wear rate = G11max (X11N , X12N , x13 , x14 ) ≤ 2.4 × 10−12 m3 /s max. blow-by = G12max (X11N , X12N , x13 , x14 ) ≤ 4.25 × 10−5 kg/s max. oil consumption = G13max (X11N , X12N , x13 , x14 ) ≤ 15.3 × 10−3 kg/hr 2 µm ≤ X11N ≤ 9 µm 2 µm ≤ X12N ≤ 9 µm 80 GPa ≤ x13 ≤ 340 GPa 150 BHV ≤ x14 ≤ 240 BHV

where X11 and X12 are (uncertain) ring and liner surface roughness design variables, respectively, x13 and x14 are (deterministic) liner Young’s modulus and hardness design variables, respectively, and R1 is power loss due to friction (subscripts w and b denote worst and best possible values due to interval uncertainty, respectively, while superscript u denotes target value from the upper level). According to the interval analysis approach, at the outer-loop optimization we determine nominal values X11N and X12N (as well as optimal values for x13 and x14 ), while solving five inner-loop optimization

130

Structural design optimization considering uncertainties

problems given the (assumed invariant) surface roughness interval uncertainty: one best-case scenario for the power loss, one worst-case scenario for the power loss, and one worst-case scenario each for oil consumption, blow-by, and wear rate. Since we do not have information from the top-level problem yet, i.e., target values for Ru1w and Ru1b , we assume these to be equal to zero. Once the power loss uncertainty interval [R1b , R1w ] has been obtained, we compute the midpoint and the percentage deviation from the endpoints to pass this uncertainty information to the top-level problem, which is formulated as min (R0w − Tw )2 + (R0w − Tw )2 + ω((R1w − Rl1w )2 + (R1b − Rl1b )2 )

(15)

R1N

where R0 denotes fuel consumption. The symbol T denotes fixed engine design target values, while the superscript “l’’ denotes interval target values from the lower level, so that the top-level problem does not consider solutions that are too far from what the bottom-level can provide. The weight ω can be adjusted to emphasize consistency rather than fuel consumption optimality. At the outer-loop optimization of this problem we determine nominal values of power loss while solving two inner-loop optimization problems given the quantified (at the lower level) power loss interval uncertainty: one best-case scenario for the fuel consumption and one worst-case scenario for the fuel consumption. After the top-level problem is solved (note that the desired fuel consumption interval target values may not be achieved), the power loss interval and the corresponding uncertainty is updated, passed down to the bottom-level problem, which is then solved again and so on. We assume that the ATC coordination process is converged when all quantities do not change significantly anymore. Table 5.7 reports the results obtained assuming δX = 0.1 (10%) for both the ring and the liner surface roughness uncertainty. The power loss links the two problems. In order to achieve the best (minimal) fuel consumption possible, we set the top-level problem target values for both the worst and the best fuel consumption equal to zero. Of course, these target values are unattainable. Therefore, the power loss interval computed by solving the bottom-level problem ([0.277, 0.369]) cannot be matched exactly when solving the top-level problem. By increasing the values of the weight ω, we increase consistency, i.e., interval matching for the power loss ([0.263, 0.356] for ω = 1000). It is interesting that while the power loss uncertainty is invariantly quantified at 15% around the interval midpoint, the fuel consumption uncertainty Table 5.7 Results of the ring/liner problem using the interval ATC formulation. Bottom level

X 11N [µm] 2.06

X 12N [µm] 5.87

x 13 [GPa] 80

x 14 [BHV] 40

R 1b [kW] 0.277

R 1w [kW] 0.369

Top-level

R 1b [kW] 0.176 0.253 0.263

R 1w [kW] 0.238 0.343 0.356

δR1 [%] 15 15 15

R 0b [kg/kWhr] 0.486 0.502 0.504

R 0w [kg/kWhr] 0.499 0.522 0.525

δR0 [%] 1.3 2 2

ω=1 ω = 10 ω = 1000

δR1 % 15

Nondeterministic formulations of analytical target cascading

131

changes for different weight values (from 1.3% to 2% around the interval midpoint). This implies that uncertainty is not always invariant with respect to the design point, as assumed in many design under uncertainty methodologies.

6 Conclusions We presented how analytical target cascading (ATC), a methodology for design optimization of hierarchically decomposed multilevel systems, can account for uncertainties. We first assumed that we have sufficient information available to model the uncertain quantities as random variables and used the popular and powerful probabilistic framework to reformulate the ATC problems as reliability-based design optimization (RBDO) problems. We used the moments of the random variables as optimization variables. Recognizing that first-order approximations may yield inaccurate estimates of standard deviations of propagated random variables, we developed an uncertainty propagation technique that is based on the advanced mean value (AMV) method. This technique can be used to generate approximate CDFs and PDFs that yield sufficiently accurate estimations of means and standard deviations of propagated random variables. A simple yet illustrative bi-level example was used to demonstrate the probabilistic ATC methodology. The results showed that the probabilistic formulation of the ATC process can be applied successfully using a bottom-up coordination. The computationally efficient AMV-based technique for the required propagation of uncertainties produced standard deviation estimates that were much more accurate relative to the ones obtained using first-order approximations, ensuring the meaningfulness of the ATC results. We then considered the case where we have incomplete uncertainty information available, and we assumed ranges for the uncertain quantities, adopting an interval analysis approach to formulate and solve robust optimization ATC problems (also known as worst-case optimization or anti-optimization). The interval analysis approach yields design solutions that are conservative relative to the ones obtained using a probabilistic design approach, especially as interval uncertainty increases. However, the interval analysis approach ensures feasibility at all times. In terms of computational cost, the nested optimization of the interval analysis approach seems to be less expensive than the required reliability analysis (analytical or simulation-based) in the probabilistic approach. It is also less challenging numerically since the inner-loop optimization problems are simple bound-constrained problems. The main challenge is that the inner-loop problems require global solutions to ensure consideration of the worst-case scenario. One of the advantages of the interval analysis approach is that the solution of the inner-loop problems provides information to the designer with respect to the beneficial or adversary effects of uncertainty so that, if possible, resources can be allocated to control critical uncertainty quantities. A significant finding is that interval uncertainty does not necessarily propagate symmetrically or invariantly.

References Chan, K.Y., Kokkolaras, M., Papalambros, P.Y., Skerlos, S.J. & Mourelatos, Z. 2004. Propagation of uncertainty in optimal design of multilevel systems: Piston-ring/cylinder-liner case study. In Proceedings of the SAE World Congress, Detroit, Michigan, Paper No. 2004-01-1559.

132

Structural design optimization considering uncertainties

Dubois, D. & Prade, H. 1988. Possibility Theory. New York: Plenum Press. Elishakoff, I., Haftka, R.T. & Fang, J.J. 1994. Structural design under bounded uncertainty – optimization with anti-optimization. International Journal of Computers and Structures 53(6):1401–1405. Haimes, Y.Y., Tarvainen, K., Shima, T. & Thadathil, J. 1990. Hierarchical Multiobjective Analysis of Large-Scale Systems. Hemisphere Publishing Corporation, pages 41–42. Haldar, A. & Mahadevan, S. 2000. Probability, Reliability, and Statistical Methods in Engineering Design. John Wiley & Sons, p. 205. Kim, H.M. 2001. Target Cascading in Optimal System Design. PhD thesis, University of Michigan. Kim, H.M., Kokkolaras, M., Louca, L.S., Delagrammatikas, G.J., Michelena, N.F., Filipi, Z.S., Papalambros, P.Y., Stein, J.L. & Assanis, D.N. 2002. Target cascading in vehicle redesign: A class VI truck study. International Journal of Vehicle Design 29(3):1–27. Kim, H.M., Michelena, N.F., Papalambros, P.Y. & Jiang, T. 2003. Target cascading in optimal system design. ASME Journal of Mechanical Design 125(3):474–480. Kim, H.M., Rideout, D.G., Papalambros, P.Y. & Stein, J.L. 2003. Analytical target cascading in automotive vehicle design. ASME Journal of Mechanical Design 125(3):481–489. Kokkolaras, M., Fellini, R., Kim, H.M., Michelena, N.F. & Papalambros, P.Y. 2002. Extension of the target cascading formulation to the design of product families. Structural and Multidisciplinary Optimization 24(4):293–301. Kokkolaras, M., Louca, L.S., Delagrammatikas, G.J., Michelena, N.F., Filipi, Z.S., Papalambros, P.Y., Stein, J.L. & Assanis, D.N. 2004. Simulation-based optimal design of heavy trucks by model-based decomposition: An extensive analytical target cascading case study. International Journal of Heavy Vehicle Systems 11(3-4):402–432. Kokkolaras, M., Mourelatos, Z.P. & Papalambros, P.Y. 2004. Design optimization of hierarchically decomposed multilevel systems under uncertainty. In Proceedings of the ASME Design Engineering Technical Conferences, Salt Lake City, Utah, Paper No. DETC2004/DAC-57357. Kokkolaras, M., Mourelatos, Z.P. & Papalambros, P.Y. 2006. Design optimization of hierarchically decomposed multilevel systems under uncertainty. ASME Journal of Mechanical Design 128(2):503–508. Kokkolaras, M., Mourelatos, Z.P. & Papalambros, P.Y. 2006. Impact of uncertainty quantification on design decisions for a hydraulic-hybrid powertrain engine. In Proceedings of the 47th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, Newport, Rhode Island, Paper No. AIAA-2006-2001. Liu, H., Chen, W., Kokkolaras, M., Papalambros, P.Y. & Kimk H.M. 2006. Probabilistic analytical target cascading – a moment matching formulation for multilevel optimization under uncertainty. ASME Journal of Mechanical Design 128(4):991–1000. Louca, S., Kokkolaras, M., Delagrammatikas, G.J., Michelena, N.F., Filipi, Z.S., Papalambros, P.Y. & Assanis, D.N. 2002. Analytical target cascading for the design of an advanced technology heavy truck. In Proceedings of the 2002 ASME International Mechanical Engineering Congress and Exposition, New Orleans, LA. Paper No. IMECE-2002-32860. Merriam-Webster on-line (www.m-w.com), accessed April 2007. Michelena, N.F., Kim, H.M. & Papalambros, P.Y. 1999. A system partitioning and optimization approach to target cascading. In Proceedings of the 12th International Conference on Engineering Design, Munich, Germany. Michelena, N.F., Louca, L., Kokkolaras, M., Lin, C.-C., Jung, D., Filipi Z., Assanis, D., Papalambros, P.Y., Peng, H., Stein, J. & Feury, M. 2001. Design of an advanced heavy tactical truck: A target cascading case study. SAE 2001 Transactions – Journal of Commercial Vehicles. Also appeared in the Proceedings of the 2001 SAE International Truck and Bus Meeting and Exhibition, Chicago, IL, Paper No. 2001-01-2793. Michelena, N.F., Park, H. & Papalambros, P.Y. 2003. Convergence properties of analytical target cascading. AIAA Journal 41(5):897–905.

Nondeterministic formulations of analytical target cascading

133

Mourelatos, Z.P. & Zhou, J. 2005. Reliability estimation and design with insufficient data based on possibility theory. AIAA Journal 43(8):1696–1705. Papalambros, P.Y. 2001. Analytical target cascading in product development. In Proceedings of the 3rd ASMO UK/ISSMO Conference on Engineering Design Optimization, Harrogate, North Yorkshire, England. Wu, Y.T. 1994. Computational methods for efficient structural reliability and reliability sensitivity analysis. AIAA Journal 32(8):1717–1723. Wu, Y.T., Millwater, H.R. & Cruse, T.A. 1990. Advanced probabilistic structural analysis method of implicit performance functions. AIAA Journal 28(9):1663–1669. Youn, B.D., Kokkolaras, M., Mourelatos, Z.P., Papalambros, P.Y., Choi, K.K. & Gorsich, D. 2004. Uncertainty propagation techniques for probabilistic design of multilevel systems. In Proceedings of the 10th AIAA/ISSMO Symposium on Multidisciplinary Analysis and Optimization, Albany, New York, Paper No. AIAA-2004-4470. Zadeh, L.A. 1978. Fuzzy sets as a basis for a theory of possibility. Fuzzy sets and systems 1:3–28.

Chapter 6

Design optimization of stochastic dynamic systems by algebraic reduced order models Gary Weickum, Matt Allen & Kurt Maute University of Colorado at Boulder, Boulder, CO, USA

Dan M. Frangopol Lehigh University, Bethlehem, PA, USA

ABSTRACT: This chapter addresses the need for efficient numerical stochastic techniques in the analysis and design optimization of dynamic systems. Most stochastic analysis techniques result in a heavy computational burden, the cost of which is amplified if embedded into a design optimization framework. This work seeks to alleviate the computational costs of analyzing dynamic systems by reduced order modeling techniques. The key to utilizing reduced order models for stochastic analysis and optimization lies in making them adaptable to design changes and variations in random parameters. This chapter presents an extended reduced order modeling method approximating the response of a dynamic system in the space of design and random parameters. The extended reduced order modeling technique is embedded into a stochastic analysis and design optimization framework. The accuracy and computational efficiency of extended reduced order models are verified with the stochastic analysis and design optimization of a linear structural dynamic system. Stochastic analyses are performed using Monte Carlo simulation, the first-order reliability method, and polynomial chaos expansion. The utility of the extended reduced order modeling method for design optimization purposes is illustrated by solving deterministic and reliability-based design optimization problems. Comparing the stochastic analyses and design optimization results using full and reduced order models show that the overall computational costs can be significantly diminished by the extended reduced order modeling method presented.

1 Introduction Stochastic and reliability analyses of static structures are well explored, and mature computational procedures have been developed (Bjerager 1990, Ghanem and Spanos 1991; Schuëller 1997; Schuëller 2001). The integration of these analysis tools into design optimization processes has been widely accomplished in the design of static structural systems, as shown by (Enevoldsen and Sørensen 1994a,b; Chandu and Grandhi 1995; Yu, Choi, and Chang 1997a,b, 1998; Grandhi and Wang 1998; Luo and Grandhi 1997), among others. However, limited work has been done on formal methods to include reliability analyses in the design of dynamic systems. There has been a considerable amount of work done on optimization of the harmonic response of a dynamic system, such as optimization with eigenvalue criteria (Haug and Choi 1982; Masur 1984; Diaz and Kikuchi 1992). The dynamic system of interest in this

136

Structural design optimization considering uncertainties

work is those that require a time integration to find the transient response of the system. The computational costs associated with dynamic analyses of realistic models with a large number of degrees of freedom hinder their inclusion into stochasticbased optimization methods. Stochastic analysis techniques, such as Monte Carlo simulation and the first-order reliability method, attempt to characterize a system’s probabilistic response due to random or uncertain inputs. A deterministic analysis of a system, assuming no input uncertainty, requires one system analysis. In contrast, the common factor to all stochastic-based analysis methods is the requirement of multiple deterministic analyses of the system at various points in the uncertain or random variable space. Design optimization seeks to find the optimal system within a design space satisfying a set of constraints. The common thread of all optimization algorithms is the necessity of multiple system analyses within the design variable space. One link between stochastic analysis and optimization is the requirement of multiple analyses of altered system configurations in a parameter space. This requirement is emphasized in a stochastic-based design optimization framework, as design criteria in the optimization procedure are now stochastic in nature. Therefore, the key to incorporating any computationally expensive system into a stochastic design framework is to decrease the expense of analyzing systems altered in a parameter space. Surrogate models have been developed allowing the approximation of the system response as a function of the design parameters based on performance predictions from high-fidelity simulation models. Surrogate models may be broadly characterized as data fit (local, multi-point, or global approximations), multi-fidelity (omitted physics, coarsened discretization or tolerances), or reduced order model (ROM) surrogates. A ROM mathematically reduces the system modeled, while still capturing the physics of the governing partial differential equations (PDEs), by projecting the original system response using a computed set of basis functions. For example, the projection reduces the number of degrees of freedom (DOFs) in a large finite element or finite volume model (O(104 to 109 ) DOF) down to a handful of basis coordinates (typically O(100 to 102 )). Thus, the ROM case is distinguished from the data fit case in that it is still intimately tied to the original PDEs and retains the physics, and is distinguished from the multi-fidelity case in that it is derived directly from the original high fidelity model and does not require multiple models of differing fidelity. ROM models have proven a successful means of reducing the computational costs of a system’s response in time (Ravindran 1999; Thomas, Dowell, and Hall 2001; Legresley and Alonso 2001; Willcox and Peraire 2001). However, ROMs typically approximate the response of only one particular configuration and are therefore of limited use for design optimization and stochastic analysis purposes. The utility of these ROMs lies in a particular system’s time integration, and any changes in the design may render the ROM inaccurate. The key missing component for the application of a ROM implemented into a reliability-based design optimization (RBDO) framework and, the focus of this work, is the extension of ROMs into the space of the design and uncertainty parameters. To date, most approximation methods used in this field consider the physical analysis as a black-box tool, and build a response surface on the results. In contrast, the objective of this work is to build an approximation technique capturing the physical nature of a system through the inclusion of

Design optimization of stochastic dynamic systems

137

the partial differential equations governing the system response. The feasibility and differences, in terms of both computational cost and accuracy, of the approximation methods will be studied herein. 1.1

Reduced order models

The construction of the ROM surrogate model from the dynamic analysis of the finite element model is discussed. The governing equation of interest is a structural dynamic response ˙ = f ext (u, u, ˙ t) M u¨ + F int (u, u)

(1)

where M is the mass matrix, F int is the internal force, u is the displacement, and ˙ t). The dynamic response is either linear or nonlinear the external force is f ext (u, u, ˙ t), depending on the internal forces F int and the external forcing function, f ext (u, u, ˙ which depend on the time, t, as well as the displacements, u, and velocities, u. For large systems, the calculation of the linear dynamic response is costly and the cost is further increased for nonlinear systems. The cost of the dynamic response is reduced using a reduced order model, which is a low dimensional approximation. Following a Galerkin type projection scheme, the displacements of the system response are approximated by k basis vectors (
) and generalized variables η as follows: u(t) =

k 

ηj (t)φj =
η(t)

(2)

j=1

The reduction of the dynamic response is performed using the approximation of (2) in (1) and premultiplying by
T as shown below. ˙ t)
T M
¨η(t) +
T F int
(η(t), η˙ (t)) =
T f ext (u, u,

(3)

The system, originally n × n, is reduced to a k × k system, where k < n. The force vectors are reduced from n × 1 for the full system to k × 1 for the reduced system. The reduced system may be written as: MR η¨ + FRint (η, η˙ ) = fRext (η, η˙ , t)

(4)

where MR , FRint , and fRext are dependent upon the basis vectors and the system matrices for a particular design. For linear systems, FRint (η, η˙ ) and fRext (η, t) are only linear functions of η and η˙ are calculated as follows: FRint (η, η˙ ) =
T K int
η(t) +
T D
η˙ (t) fRext (η, t)

=
f

T ext

(t) +
K T

ext


η(t)

(5) (6)

where K int and K ext are the stiffness matrices associated with the internal and external forces, and D is the damping matrix accounting for viscous damping.

138

Structural design optimization considering uncertainties

In nonlinear systems, the internal and external forces are dependent on the displacements and the reduced force vectors are updated by either one of the following two methods. “On-line’’ approximation evaluates the forces in the full order model using the approximation of the displacements from (2). The second method approximates the forces FRint (η, η˙ ) and fRext (η, η˙ , t) by explicit functions of η and t apriori, that is “offline’’, and does not require any computations involving the full order model when using the ROM. The advantage of using an on-line approximation is that the system response is calculated more accurately. Using an off-line approximation, the CPU time is decreased at the cost of accuracy. In the following, only linear dynamic problems will be considered and both off and on-line approaches studied. The key to building an effective ROM is to identify a set of basis vectors capturing the physics of a system. Two of the most common methods of identifying appropriate vectors are eigen analyses and snapshot methods, such as the proper orthogonal decomposition (POD). This chapter uses eigenmodes as the basis. 1.2

Ei g en m od e s

In structural dynamics, eigenmodes are the most common means of reducing a system. The eigenmodes φ and eigenvalues ω2 of an undamped system are the results of the following eigenvalue problem: ! " (7) K − ωi2 M φi = 0 where ωi is the frequency (rad/sec) of the eigenmode φi . The greatest benefit of using eigenmodes as a reduced basis is the un-coupling of equation (3) due to the orthogonality with respect to the system matrices. The projected mass matrix is φiT Mφj = δij where δij is the Kronecker symbol. The projected stiffness matrix is φiT Kφj = δij ωi2 . Using all the modes yields the same response as the full order model. In most cases, higher modes tend not to contribute much to the system response so only low modes are considered. For large-scale numerical models, computing eigenmodes can be prohibitively costly. In this case as well as for nonlinear problems, other approaches for constructing basis vectors must be used, such as proper orthogonal decomposition.

2 Extended ROM in structural dynamics The goal of this work is to use an “extended reduced order model’’ (E-ROM) to approximate the system response with respect to changes in the design and uncertainty parameters. Using a full order model tends to be computational expensive for optimization and more so for an uncertainty analysis. The proposed method involves building a ROM, using eigenmodes as a basis, for the original design (in optimization cases) or mean design (in uncertainty analyses). The ROM is only applicable to the original design. Performing an uncertainty analysis or optimization, the E-ROM must be capable of capturing the system response in the spaces of the uncertainty parameters (r) and design variables (p). An extension of a ROM into a physical parameter space has been considered by (Kirsch 2002) as a reanalysis approach. This work is an extension of these ideas into the field of stochastic

Design optimization of stochastic dynamic systems

Matrix approximation

Ini

TS1

TS1

TS2

TS2

Full

Full

Computational cost

Full

FCA

SCA

Ini

Ini

TS1

Modal approximation

Full

FCA

SCA

TS1

Ini

Modal approximation

139

Accuracy

Figure 6.1 Comparison of matrix and eigenmode approximation methods. “Ini’’ is the initial design, “TS1 ’’ and“TS2 ’’ are the first and second orderTaylor Series approximation respectively, “SCA’’ and“FCA’’ a single and full CA and“Full’’ is a full recomputed basis and matrices.

analysis and design optimization. To formulate an effective E-ROM, the set of basis functions (φ) and system matrices (M, C and K) are considered. Five methods are studied to deal with the changes in the eigenmodes and four for updating the system matrices. Using the initial modes or recomputing the modes are obvious options. Following classical perturbation methods, the eigenmodes at the altered design can be approximated by Taylor Series expansion (Kleiber and Hien 1992; Nieuwenhof and Coyotte 2002) and a combined approximation (CA) method (Kirsch 2002, Kirsch 2003; Kirsch 2001; Kirsch 1999; Kirsch 2000). There are two methods analyzed within CA, a single and full CA. The difference between single and full CA is explained in Section 2.4. Figure 6.1 shows the approximate computational cost and accuracy of each combination of updating the system matrices and eigenmodes. The reader may note utilizing updated eigenmodes is not a practical option due to the computational costs of an eigen analysis. The main costs for approximating the system matrices and the eigenmodes are due to the gradient computations. The first and second order sensitivities of M and K can be evaluated either based on the analytically derived finite element formulations or by finite differencing at comparable costs. The sensitivities of the eigenmodes are discussed in the next section. 2.1

Eigenmode s ens itivity analys is

In an effort to approximate the change of a system, the derivatives of the eigenmodes of the initial system with respect to the design variables are needed to build a first order approximation of the new design. Since finite difference methods would drastically increase the computational costs, by requiring the solution of additional eigensystems for each design variable, an analytical approach is used (Adelman and Haftka 1986; Mills-Curran 1988; Dailey 1989).

140

Structural design optimization considering uncertainties

The derivative of the eigensystem (7) with respect to a system parameter pj , expanded using the product rule results in: ) * ! " dφi dωi d T ∂K 2 ∂M − 2ωi M − ωi = 0; (φ Mφi ) = 0 (8) φi + K − ωi2 M ∂pj dpj ∂pj dpj dpj i Premultiplying (8) by φiT eliminates the second term and the gradients of the eigenvalues with respect to the system parameter are found,

 ∂M ∂K φi φiT − ωi2 ∂pj ∂pj dωi = (9) dpj 2ωi φiT Mφi The solution for the sensitivities of the eigenvalues is substituted into dωi /dpj of the first term in (8), resulting in the following system: ) * ! " ∂M ∂K dφ dω d T i i K − ωi2 M =− − 2ωi M − ωi2 (φ Mφi ) = 0 (10) φi ; dpj ∂pj dpj ∂pj dpj i This system is solved for dφi /dpj considering the norm of the eigen vector by ) * dφi d φ˜ i d φ˜ i 1 T ∂M T = − φi M + φi φi φi ; dpj dpj dpj 2 ∂pj

 ∂K dωi d φ˜ i + 2 ∂M ˜ = −K − 2ωi M − ωi φi dpj ∂pj dpj ∂pj

(11)

where K˜ + is the generalized inverse (Adelman and Haftka 1986; Dailey 1989) of the singular matrix K˜ = (K − ωi2 M). 2.2

Num eri c al mo d e l: c o nne c t ing r o d

The application of the aforementioned methodology is tested on a rod, shown in Figure 6.2, used in various past studies (Bennet and Botkin 1985; Zhang, Beckers, and Fleury 1995). The rod is clamped at the inner circumference of the left hole, and a transient force is applied to the inner circumference of the right hole. The rod is lightly damped using Raleigh damping, with α = 10−5 and β = 10−5 . The beam is modeled using 400 isoparametric 4 node elements, resulting in a total of 936 degrees of freedom. All computations are performed within MATLAB utilizing the CALFEM finite

Applied force p3 p1

Design parameters (p)

p4 p2 0

p5

Thickness = 3 mm

Figure 6.2 Finite element model of connecting rod.

50

Times [s] 1200

Design optimization of stochastic dynamic systems

141

element toolbox (Austrell, Dahblom, Lindemann, Olsson, Olsson, Persson, Petersson, Ristinmaa, Sandberg, and Wernbergk 1999). Two geometric parameters, p1 and p2 , control the horizontal positions of the center points of the left and right circular segments of the center hole, as depicted in Figure 6.2. The radii of the circular segments are kept constant. The rod has an overall length of 51 mm, a thickness of 3 mm, a Poisson ratio of 0.3, and a Young’s modulus of E = 7.2 × 105 N/mm2 . Modal and sensitivity analyses are performed on the initial system, and are used to approximate the response associated with different design changes. The ROM is based on four eigenmodes, resulting in a decrease from 936 degrees of freedom to 4. The greatest benefit of this reduction lies within the decreased computational cost of the time integration reduced by a factor of 110. 2.3

First order Taylor s eries approxima ti o n o f e i g e nmo de s

Once the geometry of the rod is altered, for optimization or uncertainty analysis, the eigenmodes themselves change. In an effort to track the change, a first order Taylor Series approximation is used about the original system to describe the basis at any set of system parameters p.
(p) ≈
0 +

∂
T0 (p − p0 ) ∂p

(12)

Displacement (mm)

The derivatives of the basis are found following the method described in Section 2.1. The utility of the approximation is studied in the following example. The reduced system is built at the original system. The parameter p1 in Figure 6.2, is altered in the design of the rod, and the eigenvalues are approximated at the design change. The design change represents a shift in the left circular segment of the center hole by 0.75 mm, full range −4 ≤ p2 ≤ 4. To isolate the effects of the approximated eigenvalues on the design change, the actual mass and stiffness matrices of the altered system are used. The plot on the left of Figure 6.3 uses the eigenmodes from the initial design to approximate the system response at the design change. The plot on the right of Figure 6.3 0.5

0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0

0

0.1

0.1

0.2

0.2

0.3

0.3

0.4

0.4

0.5

0

1

2

3

Time (ms)

4

5

6

0.5

Initial design – FA Altered design – E-ROM Altered design – FA 0

1

2

3

4

5

6

Time (ms)

Figure 6.3 Actual and approximated responses: vertical displacement at right end of the rod over time for initial and altered design. “FA’’ and “E-ROM’’ is an analysis through a full model analysis and a E-ROM using the no update (left) and update (right) of the basis.

142

Structural design optimization considering uncertainties

demonstrates the accuracy in the first order Taylor Series approximation of the eigenmodes. Therefore, using the eigenmodes from the initial design is not sufficiently accurate, but using a Taylor Series approximation leads to acceptable approximation errors. 2.4 C o m b i ned appr o ximat io n o f eig e n m o d e s Combined approximation (CA) (Kirsch 2002; Kirsch 2003; Kirsch 2001; Kirsch 1999; Kirsch 2000) is a reanalysis method used to approximate the basis vectors due to a change in system parameters. The new basis is approximated as a linear combination of another basis: φ˜ i (p) = y1 r1 + y2 r2 + · · · + yn rn

(13)

where yi are constants and r is the basis used for CA. A binomial series expansion about the original design is often chosen as the reduced basis (Kirsch 2002; Kirsch 2003). In this study, two different methods are used both of which require eigenmodes and derivatives with respect to the design/random variables. The first method approximates the ith eigenmode φ˜ i through the corresponding mode φi and its derivatives ∂φi /∂pj : φ˜ i (p) ≈ y1 φi + y2

∂φi ∂φi + · · · + yn+1 ∂p1 ∂pn

(14)

This approach is labeled “single CA’’ and is equivalent to a first order Taylor series expansion if y1 = 1 and yi > 1 = pi−1 . The second method uses all modes and its derivatives for approximating ith eigenmode φ˜ i . φ˜ i (p) ≈ y1 φi + y2

∂φi ∂φi ∂φm ∂φm + · · · + yn+1 + · · · + ym φm + ym+1 + · · · + yk ∂p1 ∂pn ∂p1 ∂pn

(15)

This approach is labeled “full CA’’. To find the coefficients y, the newly assembled system matrices are reduced by the basis r. Once the reduced matrix MCA and KCA are found, where MCA = rT Mr and KCA = rT Kr, the following eigenvalue problem is solved to find y: KCA y = λ MCA y

(16)

where y are the eigenmodes of (16). In a single CA, only the first eigenmodes from (16) is used in (14) to approximate the modes at the design change. This is done for each of the i modes, needing i separate eigen analyses. A full CA, requires only one eigen analysis, but is as large as there are modes and derivatives. The modes obtained from a single CA returns the same number of modes (i) where a full CA will return the same number of modes as there are modes and derivatives. A study is performed to see the approximation (Taylor Series, Single and Full CA) technique yielding a better approximation of the eigenmodes with respect to the full order response. To do this, the E-ROM is calibrated at an initial design and then the

Design optimization of stochastic dynamic systems

1.66

 104

2 Full

1.68

TS

1.5

1.7

SCA 1

FCA

1.72 0.5

1.9

1

1.5

2

 104

Percent error

Dissipation energy [mJ]

1.74 1.76

0

2

2

1

2.05

0

1

∆p

1.5

2

1

1.5

2

1.5

2

3

1.95

2.1

143

1

1

∆p

Figure 6.4 Dissipation energy approximating eigenmodes by Taylor Series and CA.

system parameter p1 from Figure 6.2 is varied to a p of 1 mm. The energy dissipation from t = 0.4 ms to t = 1.0 ms is used as the performance metric. The results of the study are illustrated in Figure 6.4. The top and bottom figures represent two different studies, started at different values of p0 . The figures on the left represent the recorded energy dissipation values for the three different approximation methods compared with the full order updated model. The figures on the right represent the percent error of the three approximation methods with respect to the full system analysis. In the top two graphs of Figure 6.4, both CA techniques are able to approximate the dissipation energy better for the entire p. In the bottom two figures, Taylor Series captures the response well up to a p of approximately 0.85 but then diverges rapidly, where both CA techniques are more consistent in approximating the response. In this example, the full CA leads to better approximations over the parameter intervals considered than the single CA. For the remainder of this study, both CA approximation techniques will be used. The reader may note a linear approximation of the eigenmodes captures well the nonlinear behavior of the system response with respect to system parameters. This illustrates the idea that an approximation before the physical solution of the system, i.e. the eigenmodes, is better than a similar approximation of the output of the system.

144

Structural design optimization considering uncertainties

2.5 Appro x i m a t io n o f mas s and s t iffne s s m a t r i ce s While the computation of eigenmodes after a parameter change is not feasible, calculating the altered mass and stiffness matrices is a possibility, as the computational costs of doing so does not increase as drastically with respect to degrees of freedom as do the time integration and eigen-analysis. Therefore, a first-order and secondorder approximation of the system matrices are compared against the actual mass and stiffness matrices in the analysis of an altered system. In comparing these, the approximated eigenvalues will be used for all altered responses, as their utility has already been illustrated. The positive effect of using the actual M and K is illustrated in Figure 6.5. Therefore, the difference between not updating M and K, a first order approximation, and a second order approximation is established. The negative effect of using the initial matrices for the analysis of an altered system is illustrated in the top left of Figure 6.5, where the poor approximation has drastically smaller displacements. The attempt to approximate M and K with a first-order approximation results in an even worse description of the altered system, as shown in the top right of Figure 6.5. The reader may note the first-order approximation results in a drastic reduction in stiffness and a large erroneous displacement. A linear approximation for the stiffness matrix is therefore not an option. The efforts to use a second order approximation for M and K are shown in bottom left of Figure 6.5. The approximated model fails to deviate significantly from the original model and capture the change of the structural response due to the design change. However, some utility in the second order approximation is demonstrated if the design change between the altered and initial system is small. The results for a system change 1/5 the size used in the other examples (original change of 0.75 mm) are illustrated in the bottom right of Figure 6.5. Although the change in the response of the system is significantly smaller, the second order approximation effectively captures the change. The study of approximating the mass and stiffness matrices illustrates updating the system matrices is versatile and an effective method. However, there is still utility in a second-order estimation of the matrices for small system changes, due to the computational savings in forgoing additional assembly computations. The E-ROM approximation recommended is a full CA and recalculation of M and K, with the acceptance of a second order approximation of M and K for small design changes. Further application of the suggested E-ROM will be discussed in the subsequent sections.

3

E-ROMs for uncertainty analysis

To illustrate the utility of the E-ROM in uncertainty analysis, the rod in Figure 6.2 is used, and the horizontal locations of the centers of the circular segments of the center hole are modeled as a manufacturing uncertainty by the two uncertainty variables r1 and r2 . The radii of the circular segments are kept constant. A normal distribution is assigned to the horizontal positions of the centers of both circular segments with a standard deviation of 0.2 mm. As a performance measure, the amount of energy dissipated from the system between 0.4 ms and 1.0 ms is measured and utilized to evaluate altered design configurations. The E-ROM is used in three different uncertainty analysis methods in the subsequent sections, and the results compared against using the full order system.

Displacement (mm)

Design optimization of stochastic dynamic systems

Original matrices (dp
0.75)

0.5 0.4 0.3 0.2 0.1 0 0.1 0.2 0.3 0.4 0.5

First order (dp
0.75) 0.5 0.4 0.3 0.2 0.1 0 0.1 0.2 0.3 0.4 0.5

0

Displacement (mm)

145

1

2

3 4 Time (ms)

5

6

Second order (dp
0.75)

0.5 0.4 0.3 0.2 0.1 0 0.1 0.2 0.3 0.4 0.5

0

1

2

3 4 Time (ms)

5

6

Second order (dp
0.15)

0.2

0.1

0 0

1

2

3

4

5

6

3

Time (ms) Initial design – FA

3.5

4

4.5

5

5.5

6

Time (ms) Altered design E-ROM

Altered design – FA

Figure 6.5 Approximated response for first and second order approximations of the system matrices M and K.

3.1

Monte Carlo Simulation

The most general uncertainty analysis is Monte Carlo Simulation (MCS). In general, MCS is impossible for most realistic dynamic systems due to the computational costs of each simulation, and the high number of computations required for an accurate solution. This cost is significantly reduced by utilizing the E-ROM due to the reduced cost of time integration. However, the E-ROM proposed still requires assembly of the altered system matrices, which is expensive for a large number of samples. MCS is performed here not as a proposed solution of alleviating the computational burden, but as a means of demonstrating the effectiveness of the E-ROM in the uncertainty space. A Monte Carlo analysis is performed on both the full order and the E-ROM system. 10,000 samples are taken in all, and the same sample points are used for both the full

146

Structural design optimization considering uncertainties

order model and E-ROM. Each sample point represents one particular realization of the uncertainty parameters, for which a dynamic analysis is carried out and the energy dissipated is recorded. To test the framework in terms of an uncertainty analysis, a failure surface is created by picking a critical energy dissipation level of −16.5 mJ. If a design failed to dissipate at least 16.5 mJ of energy, that is Ed ≥ −16.5 mJ, then the design is considered unsafe. The E-ROM, calibrated at the initial design, was first tested within the range of possible sample points, the mean design plus or minus 4 standard deviations, and the results were sufficiently accurate. The MCS using the full order system analysis returned a probability of failure of 19.72%, and the MCS utilizing the E-ROM returned a probability of failure of 19.4% for a single CA and 19.8% for full CA. The difference between these predictions is 0.3% for a single CA and 0.05% for the full CA. This is the error between the full order model and E-ROM analyses since the same samples points were used. 3.2

F i rst Ord er R e liab ilit y Me t ho d

To address the computational burden of Monte Carlo analysis for uncertainty quantification, the First Order Reliability Method (FORM) is studied with the E-ROM approximation. FORM employs an approximation of the limit state function at the most probable point (MPP) of failure (Hasofer and Lind 1974). FORM requires the first-order derivatives to linearize the failure surface at the MPP, and therefore it is considered accurate as long as the curvature of the failure surface in the space of the random variables is not too large at the MPP. The MPP is determined by solving an optimization process in the standard normal space of the random parameters. FORM based RBDO methods are often used within the structural design community (Enevoldsen and Sørensen 1994b; Enevoldsen and Sørensen 1994a; Frangopol and Corotis 1996; Yu, Choi, and Chang 1997a; Grandhi and Wang 1998; Luo and Grandhi 1997). However, FORM based RBDO methodologies are still in their infancy for multi-physics and dynamic systems due to the additional cost of RBDO being magnified by high analysis costs (Allen, Raulli, Maute, and Frangopol 2004; Allen and Maute 2004). The major expense of FORM lies in the MPP search. The MPP search has certain characteristics suitable for the E-ROMs. The search is conducted in the standard normal space of the random parameters centered about the mean design. Therefore, the E-ROM is calibrated at the mean design. Also, the objective of the optimization process is to minimize the distance to the origin, aiming to keep the process close to the calibration point. For low reliability requirements the MPP lies close to the origin and no recalibration is required. However, for high reliability requirements a recalibration and trust region framework are typically required. FORM analysis were performed using both the full order system and E-ROM. The analyses were performed on the initial configuration of the rod, and the failure criteria was the dissipation energy below −16.5 mJ. The quantitative results of the FORM analyses are summarized in Table 6.1. The two optimization problems converged to slightly different MPPs leading in turn to a change in the reliability index and the calculated probability of failure. However, the difference between the two model results is small. To improve the convergence of the E-ROM based MPP search towards the solution of the full order model, the E-ROM could be recalibrated at the MPP initially found and the MPP search continued.

Design optimization of stochastic dynamic systems

147

Table 6.1 FORM results of full and E-ROM model analysis. Analysis method

MPP s1

MPP s2

Beta

Pf

Full model E-ROM Single CA Full CA

0.837

0.194

0.859

19.52%

0.848 0.839

0.179 0.192

0.867 0.861

19.30% 19.46%

Table 6.2 PCE based MCS results of full and E-ROM model analysis. Analysis method

Full model E-ROM Single CA Full CA

3.3

MSC (%)

FORM (%)

PCE 1st order (%)

2nd order (%)

3rd order (%)

19.7

20.38

20.8

20.1

19.8

19.4 19.8

20.42 20.32

21.8 20.7

19.8 20.1

19.5 19.7

Polynomial chaos expans ion bas ed M o nte Carl o Si mul ati o n

The obvious drawback of running MCS is the number of simulations to obtain an error low enough to accurately represent the PDF of the model. The major drawbacks of FORM are the inability to capture nonlinearities in the failure surface, and the limitation of having only one particular probability measure instead of a complete PDF. The third uncertainty analysis technique discussed utilizes a polynomial chaos expansion (PCE) on the system output to allow a MCS to be used at a relatively low computational cost compared to a full MCS (Xiu, Lucor, Su, and Karniadakis 2002; Field, Red-Horse, and Paez 2000; Field 2002; Nurdin 2002). PCE uses system analyses at collocation points to build a polynomial approximation of the model response. Depending on the accuracy of the PCE needed, a different number of collocation points are used. The more collocation points, the better the PCE to accurately represent the model response. The computational cost of this method lies within the system analyses at the collocation points. There, the E-ROM is utilized to reduce the computational costs of these analyses. Once the model is sampled and the PCE is built, a MCS can be conducted on the PCE approximation to get the PDF of the model at a low computation cost. If a sufficiently large number of samples are taken for the MCS, the majority of the error is due to the PCE approximation, and not the MCS. Considering the same limit state function as used previously, a first, second, and third order PCE is built and the probability of failure is calculated. The results are shown in Table 6.2. The probability of failure for MCS, FORM, and PCE are all within one percent of each other. Between the two different methods of E-ROM, the full CA is able to predict the stochastic response more accurately, where the error between each of the methods is at most 0.1%.

148

Structural design optimization considering uncertainties

4 Deterministic optimization with E-ROMs While the intended use of the proposed method is to alleviate the computational costs of stochastic-based optimization, its utility is demonstrated first with a deterministic optimization problem. The E-ROM approach sufficiently models a system within a certain region around the design point for which the E-ROM was calibrated. Most design optimization problems, unlike FORM optimization problems, contain variables with bounds larger than the trust region of the E-ROM. Therefore, an adaptation strategy for updating the trust region is used to incorporate the E-ROM into optimization problems with large bounds. In this study, the trust region framework of (Giunta and Eldred 2000 and Eldred, Giunta, and Collis 2004) is used. The initial bounds of the trust region are from −2 to 2 for both design variables while the global bound range from −4 to 4 for both design variables. For optimization, the connecting rod’s energy dissipation is used as the objective to maximize in the design problem. The two design variables are the horizontal positions of the center points of the left and right circular segments of the center hole, as depicted in their initial configuration in Figure 6.2. The deterministic optimization problem is as follows: min (−Ediss ) s

(17)

subject to −4 ≤ si ≤ 4

where Ediss is the energy dissipated and si are the optimization variables describing the center hole geometry. The contour of the dissipation energy is seen on the right of Figure 6.6. The contour lines in the figure are for illustrative purposes only, and obtained by sampling the design space to compare the optimization results. The results from the deterministic optimum are shown in Table 6.3. All three models converged to the same deterministic optimum. Each recalibration is more expensive to obtain than a function evaluation of the full model which is twice the cost of the full model. This is due to obtaining the basis which includes derivatives with respect to the design variables. The function evaluations of single and full CA also accrue computational costs, but not as much as a function evaluation of the full model. For large systems, the function evaluations of the full order model would be significantly more costly than a function evaluation of the E-ROM. Comparing Single and Full CA, each recalibration requires the same costs. Performing a full CA is more costly than a single CA because more basis vectors are used with full CA.

Table 6.3 Full and E-ROM model deterministic optimization results. Analysis method

s1

s2

Function evaluations

Full model E-ROM Single CA Full CA

1.548

4.0

27

1.551 1.549

4.0 4.0

122 144

Equivalent recalibration cost

16 16

Design optimization of stochastic dynamic systems

149

5 RBDO with E-ROMs In general, the solution of optimization problems with stochastic criteria is significantly more expensive than the solution of problems with purely deterministic criteria. Virtually all stochastic analysis procedures require additional analyses of the system for various points within the uncertainty space around the mean design. These analyses are required for each evaluation of a stochastic criterion within each iteration of the design optimization. If these analyses are solved using an E-ROM instead of a full system model, with similar convergence results, then significant computational savings are realized. The framework is tested on a RBDO of the rod in Figure 6.2. The energy dissipated is again used as the objective to be maximized in the design problem. However, a reliability-based constraint is imposed on the system. The constraint limits the standard deviation of the dissipation energy less than 300 µJ, making the RBDO problem as follows: min (−Ediss ) s

subject to

σE − 300 ≤ 0 and −4 ≤ si ≤ 4

(18)

where Ediss is the energy dissipated and σE is the standard deviation of the energy dissipated. The constraint on the standard deviation forces the optimization to a more robust design, limiting the sensitivity of the system performance to uncertainties. The standard deviation is found by a Monte Carlo simulation based on PCE, as highlighted in Section 3. This method was chosen due to its computational efficiency and its ability to obtain the entire PDF of the output response. The derivatives of the standard deviation are obtained by finite differencing. The positions of the centers of the circular segments of the center hole are now treated as both the design variables and the random variables. The design variables represent the mean or intended design, and a normal distribution is assigned to the horizontal positions of the center points of the left and right circular segments of the center hole, each with a standard deviation of 0.2 mm. The radii of the circular segments are kept constant. To illustrate this academic example problem, the constraint is explored in the design and uncertainty space by sampling uniformly throughout. The results of the sampling are illustrated in the contour plot in Figure 6.6. The plot on the left of Figure 6.6 represents the contour plots for the standard deviation of the energy dissipated in µJ. The reader may note the constraint boundary has been highlighted in the figure, and the feasible and infeasible regions identified. The initial design is feasible, but the deterministic design is infeasible. Figure 6.6 on the right overlays the constraint boundaries onto the contour plot of the objective, to give the reader the general idea of the RBDO problem. 5.1

Results

The RBDO problem is solved using the full order system analyses and two E-ROM approximation methods, single and full CA. The optimization results are summarized in Table 6.4. Each of the methods has similar search directions, step sizes and

Structural design optimization considering uncertainties

Standard Deviation of Energy

Energy and Standard Deviation of Energy

0 30

0 1600

750

30 0

0

300

1

120

00

30 0

300 450

150

14000

15 0

15

0

3

300

3

300

0

2

1600

300

150

300

2

14000

15

0

600 450

1

14000

14 00 0

450

0

0

30

16000

1

0

45

0

1400 0

12 00 0

140

900

750

60 0

0 60

2

0

30

1

200 00

600

450

3

18000

0 450

1

450

300

2

50

0

45

30

0

90 150

3

75

p2

4

900

00

4

30 0

150

4 4

p1 3

2

1

0

1

2

p

3

4 4

4

3

2

1

0

1

2

3

4

1

Deterministic optimum

RBDO optimum

Infeasible solution

Initial design (starting point)

Figure 6.6 Contour plots of objective and stochastic constraint of RBDO problem in the space of the optimization variables.

Table 6.4 Full and E-ROM model RBDO results. Analysis method

s1

s2

Iterations

Time (min)

Full model E-ROM Single CA Full CA

1.31

2.81

16

130

1.35 1.27

2.84 2.86

25 13

50.8 21.5

converged solutions. The E-ROM is recalibrated within each trust region step. Within each trust region the stochastic analyses and function evaluation are computed using the E-ROM. The E-ROM framework converges quickly to the general solution of the RBDO problem as did the deterministic optimum. However, it is recommend imposing a weak convergence criterion on the E-ROM optimization and switching to the full model to fine tune the optimization variables in the vicinity of the solution. The benefit of reliability-based design optimization is demonstrated through the standard deviation of the dissipation energy. The standard deviation at the deterministic optimum is 579 mJ, 299 mJ at the RBDO optimum and 171 mJ at the initial design. The stochastic response is characterized by a probability density function with a mean and standard deviation, as opposed to a deterministic formulation where the output is characterized by a single value. The objective of the optimization is to maximize the energy dissipated by the system, or move the mean of the design to a greater standard deviation in dissipation energy. With no stochastic constraint, the deterministic

Design optimization of stochastic dynamic systems

151

Table 6.5 Computational costs of E-ROM and full model analysis.

Assembly Time integration

Full model

E-ROM

1.11 sec 5.59 sec

1.11 sec 0.051 sec

optimization achieves a large level of energy dissipation, but results in a design whose performance is highly susceptible to uncertainties. 5.2

Computational cos t

The computational cost measured by CPU time of the above examples are reported in Tables 6.4 and 6.5. Since the overall analysis and optimization times are proportional to the computational time required to analyze one altered design configuration, the computational costs associated with each individual analysis are reported in Table 6.5. The cost for each analysis consists of two main components: the cost to assemble the mass and stiffness matrices for the new design and the cost to perform the time integration for the transient analysis of the new design. The computational savings of the reduced time integration over the full time integration is a factor a 110. The overall computational savings per analysis is significantly smaller, approximately 43%, due to the relatively large assembly time. As the number of degrees of freedom of a system increase, the full integration time generally increases at a higher rate than the required assemble time, thus increasing the savings with the E-ROM approach. The overall computational savings of the E-ROM in an optimization framework is dependent upon the costs of recalibration, and the frequency of recalibration required. Again, the E-ROM is most beneficial in the RBDO framework requiring many analyses about a mean design used as the recalibration point. Table 6.4 demonstrates the effectiveness of the E-ROM to save CPU time of the RBDO problem. The single CA E-ROM saves approximately 39% of the time it takes the full order model to run and the full CA E-ROM saved 71%. When moving to larger models, the time saved running an E-ROM will be more significant.

6 Conclusions A computational framework has been presented allowing reliability-based design optimization of dynamic systems by reducing the associated computational costs. The framework utilizes reduced order models extended into the parameter space of the design and random variables. This extension allows for the analysis of altered designs by the E-ROM at a significant reduction in computational cost. Various approaches for constructing an E-ROM were studied. Numerical studies showed the system matrices need recomputing at each altered design and can not be approximated by a Taylor series expansion. In contrast, the reduced basis can be well approximated with a first order Taylor series expansion, using the full combined approximation approach. The utility of the approach was tested on a linear elastic rod subjected to a time-varying load. The E-ROM was used for various stochastic analysis techniques compared against the full model analyses. The E-ROMs ability to capture the essential characteristics of

152

Structural design optimization considering uncertainties

the system was demonstrated by both deterministic and reliability-based design optimization examples. The E-ROM framework proved to converge to the solution with significantly less computational effort than the full system model.

Acknowledgments The authors acknowledge the support by the National Science Foundation under grant DMI-0300539. The opinions and conclusions presented are those of the authors and do not necessarily reflect the views of the sponsoring organizations.

References Adelman, H. & Haftka, R. 1986. Sensitivity analysis of discrete structural systems. AIAA Journal 24:823–832. Allen, M. & Maute, K. 2004. Reliability-based design optimization of aeroelastic structures. Structural and Multidisciplinary Optimization 27(4):228–242. Allen, M., Raulli, M., Maute, K. & Frangopol, D. 2004. Reliability-based analysis and design optimization of electrostatically actuated MEMS. Computers and Structures 82(13–14): 1007–1020. Austrell, P.-E., Dahblom, O., Lindemann, J., Olsson, A., Olsson, K.-G., Persson, K., Petersson, H., Ristinmaa, M., Sandberg, G. & Wernbergk, P.-A. CALFEM: A finite element toolbox to MATLAB, Version 3.3., Structural Mechanics, LTH, Sweden, 1999. http://www.byggmek.lth.se/Calfem/index.htm. Bennet, J. & Botkin, M. 1985. Structural shape optimization with geometric description and adaptive mesh refinement. AIAA Journal 23:458–464. Bjerager, P. 1990. On computational methods for structural reliability analysis. Structural Safety 9(2):79–96. Chandu, S. & Grandhi, R. 1995. General purpose procedure for reliability based structural optimization under parametric uncertainties. Advances in Engineering Software 23:7–14. Dailey, R. 1989. Eigenvector derivatives with repeated eigenvalues. AIAA Journal 27: 486–491. Diaz, A. & Kikuchi, N. 1992. Solution to shape and topology eigenvalue optimization problems using a homogenization method. International Journal for Numerical Methods in Engineering 35:1487–1502. Eldred, M.S., Giunta, A.A. & Collis, S. 2004. Second-order corrections for surrogate-based optimization with model hierarchies. In AIAA 2004-44570, 10th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, 30 August – 1 September 2004, Albany, NY. Enevoldsen, I. & Sørensen, J.D. 1994a. Reliability-based optimization as an information tool. Mechanics of Structures & Machines 22(1):117–135. Enevoldsen, I. & Sørensen, J.D. 1994b. Reliability-based optimization in structural engineering. Structural Safety 15:169–196. Field, R. 2002. Numerical methods to estimate the coefficients of the polynomial chaos expansion. In 15th Engineering Mechanics Conference, Columbia University, NY. ASCE. Field, R., Red-Horse, J. & Paez, T. 2000. A nondeterministic shock and vibration application using polynomial chaos expansions. In 8th ASCE Joint Specialty Conference on Probabilistic Mechanics and Structural Reliability, South Bend, IN. Frangopol, D. & Corotis, R. 1996. Reliability-based structural system optimization: Stateof-the-art verse state-of-the-practice. In Cheng (ed.), Analysis and Computation: Proceedings of the Twelfth Conference held in Conjunction with Structures Congress XIV, pp. 67–78.

Design optimization of stochastic dynamic systems

153

Ghanem, R.G. & Spanos, P.D. 1991. Stochastic finite element: a spectral approach, Springer. Giunta, A.A. & Eldred, M.S. 2000. Implementation of a trust region model management strategy in the dakota optimization toolkit. In AIAA/USA/NASA/ISSMO Symposium on Multidisciplinary Analysis and Optimization, Long Beach, CA. Grandhi, R. & Wang, L. 1998. Reliability-based structural optimization using improved two-point adaptive nonlinear approximations. Finite Elements in Analysis and Design, 35–48. Hasofer, A. & Lind, N. 1974. Exact and invariant second-moment code format. J. of Engineering Mechanics 100:111–121. Haug, E.J. & Choi, K.K. 1982. Systematic occurrence of repeated eigenvalues in structural optimization. Journal of Optimization Theory and Applications 38:251–274. Kirsch, U. 1999. Efficient, accurate reanalysis for structural optimization. AIAA Journal 37(12):1663–1669. Kirsch, U. 2000. Combined approximations – a general reanalysis approach for structural optimization. Structural and Multidisciplinary Optimization 20(2):97–106. Kirsch, U. 2001. Exact and accurate solutions in the approximate reanalysis of structures. AIAA Journal 39(11):2198–2205. Kirsch, U. 2002. A unified reanalysis approach for structural analysis, design, and optimization. Structural and Multidisciplinary Optimization 25(1):67–85. Kirsch, U. 2003. Approximate vibration reanalysis of structures. AIAA Journal 41(3): 504–511. Kleiber, M. & Hien, T. 1992. The Stochastic Finite Element Method, Basic Perturbation Technique and Computer Implementation. Wiley. Legresley, P. & Alonso, J. 2001. Investigation of nonlinear projection for POD based reduced order models for aerodynamics. In AIAA 2001-16737, 39th Aerospace Sciences Meeting & Exhibit, January 8–11, 2001, Reno, NV. Luo, X. & Grandhi, R. 1997. Astros for reliability-based multidisciplinary structural analysis and optimization. Computers and Structures 62:737–745. Masur, E. 1984. Optimal structural design under multiple eigenvalue constraints. International Journal of Solids and Structures 20:117–120. Mills-Curran, W. 1988. Calculation of eigenvector derivatives for structures with repeated eigenvalues. AIAA Journal 26(7):867–871. Nieuwenhof, B. & Coyotte, J. 2002. A perturbation stochastic finite element method for the time-harmonic analysis of structures with random mechanical properties. In 5th World Congress on Computational Mechanics, Vienna, Austria. WCCM. Nurdin, H. 2002. Mathematical modeling of bias and uncertainty in accident risk assessment. Mathematical Sciences, University of Twente, The Netherlands. Ravindran, S. 1999. Proper orthogonal decomposition in optimal control of fluids. Technical report, NASA TM-1999-209113. Schuëller, G. 1997. A state-of-the-art report on computational stochastic mechanics. Probabilistic Engineering Mechanics 12:197–321. Schuëller, G. 2001. Computational stochastic mechanics – recent advances. Computers & Structures 79:2225–2234. Thomas, J., Dowell, E. & Hall, K. 2001. Three-dimensional transonic aeroelasticity using proper orthogonal decomposition based reduced order models. In 42nd AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials (SDM) Conference, April 2001, Seattle, WA, AIAA Paper 2001-1526. Willcox, K. & Peraire, J. 2001. Balanced model reduction via the proper orthogonal decomposition. In 15th AIAA Computational Fluid Dynamics Conference, June 11–14, Anaheim, CA, AIAA 2001-2611.

154

Structural design optimization considering uncertainties

Xiu, D., Lucor, D., Su, C.-H. & Karniadakis, G. 2002. Stochastic modeling of flow-structure interactions using generalized polynomial chaos. Journal of Fluids Engineering 124:51–59. Yu, X., Chang, K. & Choi, K. 1998. Probabilistic structural durability prediction. AIAA Journal 36(4):628–637. Yu, X., Choi, K. & Chang, K. 1997a. A mixed design approach for probabilistic structural durability. Journal of Structural Optimization 14(2–3):81–90. Yu, X., Choi, K. & Chang, K. 1997b, January. Reliability and durability based design sensitivity analysis and optimization. Technical Report R97-01, Center for Computer Aided Design University of Iowa. Zhang, W.-H., Beckers, P. & Fleury, C. 1995. A unified parametric design approach to structural shape optimization. International Journal for Numerical Methods in Engineering 38:2283– 2292.

Chapter 7

Stochastic system design optimization using stochastic simulation Alexandros A. Taflanidis & James L. Beck California Institute of Technology, CA, USA

ABSTRACT: Engineering design in the presence of uncertainties often involves optimization problems that include as objective function the expected value of a system performance measure, such as expected life-cycle cost or failure probability. For complex systems, this expected value can rarely be evaluated analytically. In this study, it is calculated using stochastic simulation techniques which allow explicit consideration of nonlinear characteristics of the system and excitation models, as well as complex failure modes. At the same time, though, these techniques involve an unavoidable estimation error and significant computational cost which make the associated optimization challenging. An efficient framework, consisting of two stages, is presented here for such optimal system design problems. The first stage implements a novel approach, called Stochastic Subset Optimization, for iteratively identifying a subset of the original design space that has high plausibility of containing the optimal design variables. The second stage adopts some stochastic optimization algorithm to pinpoint, if needed, the optimal design variables within that subset. Topics related to the combination of the two different stages for overall enhanced efficiency are discussed. An illustrative example is presented that shows the efficiency of the proposed methodology; it considers the retrofitting of a four-story structure with viscous dampers. The minimization of the expected lifetime cost is adopted as the design objective. The expected cost associated with damage caused by future earthquakes is calculated by stochastic simulation using a realistic probabilistic model for the structure and the ground motion.

1 Introduction In engineering design, the knowledge about a planned system is never complete. First it is not known in advance which design will lead to the best system performance in terms of the specified metric; it is therefore desirable to optimize the performance measure over the space of design variables that define the set of acceptable designs. Second, modeling uncertainty arises because no mathematical model can capture perfectly the behavior of a real system and its future excitation. In practice, the designer chooses a model that he or she feels will adequately represent the behavior of the system when built; however, there is always uncertainty about which values of the model parameters will give the best representation of the system, so this parameter uncertainty should be quantified. Furthermore, whatever model is chosen, there will always be an uncertain prediction error between the model and system responses. For an efficient engineering design, all uncertainties, involving future events as well as the modeling of the system, must be explicitly accounted for. A probability logic approach provides a rational and

156

Structural design optimization considering uncertainties

consistent framework for this purpose (Jaynes 2003). In this case, this process is often called stochastic system design. In this context, consider some controllable parameters that define the system design, referred to herein as design variables, ϕ = [ϕ1 , ϕ2 , . . . , ϕnϕ ] ∈ Φ ⊂ Rnϕ , where Φ denotes the bounded admissible design space. Also consider a model class that is chosen to represent a system design and its future excitation, where each model in the class is specified by a nθ -dimensional vector θ = [θ1 , θ2 , . . . , θnθ ] lying in Θ ⊂ Rnθ , the set of possible values for the model parameters. Because there is uncertainty in which model best represents the system behavior, a PDF (probability density function) p(θ|ϕ), which incorporates available knowledge about the system, is assigned to these parameters. The objective function for a robust-to-uncertainties design is, then, expressed as:
Eθ [h(ϕ, θ)] =

Θ

h(ϕ, θ)p(θ|ϕ)dθ

(1)

where Eθ [·] denotes expectation with respect to the PDF for θ and h(ϕ, θ) : Rnϕ × Rnθ → R denotes the performance measure of the system, referred to also as the loss function; possible examples for h(ϕ, θ) are the life-cycle cost (see (37) later) or the indicator function for system failure so that (1) gives the failure probability (see (6) later). We then have the optimal stochastic design problem: Minimize Eθ [h(ϕ, θ)] subject to f c (ϕ) ≥ 0

(2)

where f c (ϕ) corresponds to a vector of constraints. Such optimization problems, arising in decision making under uncertainty, are typically referred to as stochastic optimization problems (e.g. Ruszczynski & Shapiro 2003, Spall 2003). In structural engineering, stochastic design problems are usually related to the expected life-cycle cost of a structure (e.g. Ang & Lee 2001) or to its reliability, quantified in terms of the probability of failure P(F|ϕ) (e.g. Enevoldsen & Sørensen 1994, Gasser & Schuëller 1997). Many variants of such problems have been posed, typically expressed in one of three following forms: (a) optimization of the system reliability given deterministic constraints, (related, for example, to construction cost), (b) optimization of the cost of the structure given reliability constraints, or (c) optimization of the expected life-cycle cost of the structure. Approaches have been suggested for transforming the latter problem to one of the former two. This is established by approximating the cost related to future damages to the structure in terms of its failure probability (Sørensen et al. 1994). In this setting, Reliability-Based Design Optimization (RBDO), i.e. design considering reliability measures in the objective function or the design constraints, has emerged as one of the standard tools for robust and costeffective design of structural systems. An alternative design methodology that also accounts for probabilistic system response is the Robust Design Optimization (RDO). RDO primarily seeks to minimize the influence of stochastic variations on the mean design; as such, it focuses on reduction of the mean performance rather than looking at optimizing the response that exceeds some acceptable thresholds, as RBDO does. Still, RBDO and RDO represent only a portion of the potential problems encountered in robust-to-uncertainties system design optimization.

Stochastic system design optimization using stochastic simulation

157

In this study we discuss general stochastic system design problems that involve as objective function the expected value of a system performance measure. Some special attention is given to problems with reliability objectives, i.e. when that expected value corresponds to a failure probability. This class of problems, which belongs to RBDO, will be referred to herein as ROP (reliability objective problems). We focus on analysis methods that are applicable to complex systems, involving, for example, nonlinear models with high-dimensional uncertainties. These types of problems appear often in the study of dynamic systems when the excitation is modeled as a stochastic process, for example, in the field of dynamic reliability (Au & Beck 2003b). An efficient framework for analysis and optimization is presented here for such design problems using stochastic simulation techniques to evaluate the system performance.

2 Optimal stochastic system design using stochastic simulation 2.1

G eneral cas e

We consider the optimization described by (2), which may be equivalently formulated as: ϕ∗ = arg min Eθ [h(ϕ, θ)] ϕ∈Φ

(3)

where the deterministic constraints are taken into account by appropriate definition of the admissible design space Φ. In the probabilistic setting described earlier, model uncertainties may be incorporated in the system description as a model prediction error, i.e. an error between the response of the actual system and that of the assumed model. This error can be model probabilistically as a random variable (Beck & Katafygiotis (1998)) and augmented into θ to form an uncertain parameter vector, comprised of both the uncertain model parameters and the model prediction error. For optimization (3), the integral in (1) must be evaluated. A particular source of difficulty for structural design when complex systems are considered is the high computational cost associated with this evaluation. To reduce this computational effort, many specialized approximate approaches have been proposed for structural optimizations (e.g. Enevoldsen & Sørensen 1994, Gasser & Schuëller 1997, Jensen 2005). These approaches include using response surface methods to approximate the structural behavior or using some proxy for the structural reliability in ROP (for example, a reliability index obtained through FORM or SORM). These specialized approaches may work satisfactorily under certain conditions, but are not proved to converge to the solution of the original design problem. For this reason such approaches are avoided in this study. Instead, evaluation of the integral in (1) through stochastic simulation techniques is considered. In this setting, an unbiased estimate of the expected value in (1) can be obtained using a finite number, N, of random samples of θ, drawn from p(θ|ϕ):  ˆ θ,N [h(ϕ, ΩN )] = 1 h(ϕ, θi ) E N N

i=1

(4)

158

Structural design optimization considering uncertainties

where ΩN = [θ1 . . . θN ], and vector θi denotes the sample of the uncertain parameters used in the ith simulation. This estimate involves an unavoidable error eN (ϕ, ΩN ). The optimization in (3) is, then, only approximately equivalent to: ϕ∗N = arg min Eˆ θ [h(ϕ, ΩN )]

(5)

ϕ∈Φ

However, if the stochastic simulation procedure is a consistent one, then as N → ∞, Eˆ θ,N [h(ϕ, ΩN )] → Eθ [h(ϕ,θ)] and ϕ∗N → ϕ∗ . The existence of the estimation error eN (ϕ, ΩN ), which may be considered as noise in the objective function, contrasts with classical deterministic optimization where it is assumed that one has perfect information. Figure 7.1a illustrates the difficulties associated with eN (ΩN , ϕ). The curves corresponding to simulation-based evaluation of the objective function have non-smooth characteristics, a feature which makes application of gradient-based algorithms challenging. Also, the estimated optimum depends on the exact influence of the estimation error, which is not the same for all evaluations; different runs of the algorithm converge to different solutions, which do not necessarily correspond to the true optimum. An efficient framework, consisting of two stages, is discussed in the following sections for such optimizations. The first stage implements a novel approach, called Stochastic Subset Optimization (SSO) (Taflanidis & Beck 2007a, Taflanidis & Beck 2007b) for efficiently exploring the sensitivity of the objective function to the design variables and iteratively identifying a subset of the original design space that has high plausibility of containing the optimal design variables. The second stage adopts some appropriate stochastic optimization algorithm to pinpoint the optimal design variables using information from the first stage. Topics related to the combination of the two different stages for enhanced overall efficiency are discussed. Before presenting this framework some special characteristics of ROP are considered.

(a) 0.09

(b)

0.08 0.07

0.6

0.06

0.4

0.05 0.04 10

~ ␪)) Pε(g(␸,

0.8 h(␸, ␪)

E [h(␸, ␪)]

1

analytical sim N
1000 sim N
4000

0.2 12

14

␸

16

18

20

0 0.1

IF(␸, ␪) 0.05

0 ~ g(␸, ␪)

0.05

0.1

Figure 7.1 (a) Analytical and simulation-based (sim) evaluation of objective function and (b) comparison between the two candidate loss functions for reliability objective problems.

Stochastic system design optimization using stochastic simulation

2.2

159

Reliability objective problems

In a reliability context, the robust probability of failure (Papadimitriou et al. 2001) can be employed to include probabilistic model uncertainties when evaluating the performance of a system. This probability quantifies the performance by giving a measure of the plausibility of the occurrence of system failure, based on all available information, and it is expressed as:
P(F|ϕ) = Eθ [IF (ϕ, θ)] = IF (ϕ, θ)p(θ|ϕ)dθ (6) Θ

where IF (ϕ, θ) is the indicator function of failure, which is 1 if the system that corresponds to (ϕ, θ) fails, i.e. its response departs from the acceptable performance set, and 0 if it does not. An equivalent expression can also be used for the robust failure probability when a model prediction error, ε(ϕ, θ), is assumed. Let g(ϕ) > 0 and g(ϕ, ˜ θ) > 0 be the limit state quantities defining the system’s and model’s failure respectively, and let the model prediction error be defined in such a way that the relationship ε(ϕ, θ) = g(ϕ, ˜ θ) − g(ϕ) holds; then, if Pε (·) is the conditional on (ϕ, θ) cumulative distribution function for the model prediction error ε(ϕ, θ) and noting that g(ϕ) > 0 is equal to ε(ϕ, θ) < g(ϕ, ˜ θ), the robust failure probability can be equivalently expressed as (Taflanidis & Beck 2007a):
P(F|ϕ) = Pε (g(ϕ, ˜ θ))p(θ|ϕ)dθ (7) Θ

where in this case the vector θ corresponds solely to the uncertain parameters for the system and excitation model, i.e. excluding the prediction error. Thus, the loss function in ROP corresponds either to (a) h(ϕ, θ) = IF (ϕ, θ) or (b) h(ϕ, θ) = Pε (g(ϕ, ˜ θ)), depending on which formulation is adopted, (6) or (7). Both of these formulations are used in the two stages of the framework suggested. In Figure 7.1b these two loss functions are compared when ε is Normally distributed with mean 0 and standard deviation 0.01.

3 Stochastic subset optimization SSO is an efficient algorithm for exploring the sensitivity of stochastic design optimization problems using a small number of system analyses (Taflanidis & Beck 2007a, Taflanidis & Beck 2007b). 3.1 Augmented problem Consider, initially, the modified positive loss function hs (ϕ, θ) : Rnϕ × Rnθ → R+ defined for a constant s as: hs (ϕ, θ) = h(ϕ, θ) − s

where s < min h(ϕ, θ) ϕ,θ

(8)

and note that Eθ [hs (ϕ, θ)] = Eθ [h(ϕ, θ)] − s. Since the two expected values differ only by a constant, optimization of the expected value of h(·) is equivalent, in terms of the

160

Structural design optimization considering uncertainties

optimal design choice, to optimization of the expected value of hs (·). In the SSO setting we focus on the latter optimization. The basic idea in SSO is the formulation of an augmented problem where the design variables are artificially considered as uncertain with distribution p(ϕ) over the design space Φ. In the setting of this augmented stochastic design problem, define the auxiliary PDF π(ϕ, θ) as: π(ϕ, θ) =

hs (ϕ, θ)p(ϕ, θ) Eϕ,θ [hs (ϕ, θ)]

(9)

where p(ϕ, θ) = p(ϕ)p(θ|ϕ) and the normalizing integral in the denominator corresponds to the expected value in the augmented uncertain space:

Eϕ,θ [hs (ϕ, θ)] = hs (ϕ, θ)p(ϕ, θ)dθdϕ (10) Φ

Θ

This expected value will not be explicitly needed, but it can be obtained though stochastic simulation, which leads to an expression similar to (4) but with the pair [ϕ, θ] defining the uncertain parameters. The transformation of the loss function in (8) may be necessary to ensure that π(ϕ, θ) ≥ 0. For most structural design problems h(ϕ, θ) ≥ 0 and the transformation in (8) is usually unnecessary, which is always the case for ROP. However, in some cases it may be advantageous to choose s near the minimum of h(ϕ, θ) to increase efficiency of SSO (see later). In terms of the auxiliary PDF, the objective function Eθ [hs (ϕ, θ)] can be expressed as: Eθ [hs (ϕ, θ)] =

π(ϕ) Eϕ,θ [hs (ϕ, θ)] p(ϕ)

where the marginal PDF π(ϕ) is equal to:
π(ϕ) = π(ϕ, θ)dθ Θ

(11)

(12)

Define, now, J(ϕ) as: J(ϕ) =

π(ϕ) Eθ [hs (ϕ, θ)] = Eϕ,θ [hs (ϕ, θ)] p(ϕ)

(13)

Since Eϕ,θ [hs (ϕ, θ)] can be viewed simply as a normalizing constant, minimization of Eθ [hs (ϕ, θ)] is equivalent to the minimization of the quotient J(ϕ) = π(ϕ)/p(ϕ). For this purpose the marginal PDF π(ϕ) in the numerator must be evaluated. Samples of this PDF can be obtained through stochastic sampling/simulation techniques (Robert & Casella (2004)). These techniques will give sample pairs [ϕ, θ] that are distributed according to π(ϕ, θ). Their ϕ component corresponds to samples from the marginal distribution π(ϕ). Appendix A briefly discusses two appropriate sampling algorithms, one using a direct approach to Monte Carlo (MC) simulation and one using Markov Chain Monte Carlo (MCMC) simulation. SSO is based on exploiting the information in these samples.

Stochastic system design optimization using stochastic simulation

3.2

161

Subset analys is

An analytical approximation for π(ϕ) based on these samples for ϕ can be established using, for example, the maximum entropy method (Ching & Hsieh 2007), histograms (Au 2005) or kernel density estimators. Experience indicates that for challenging problems, including, for example, cases where the dimension nϕ is not small (e.g. larger than two) or the sensitivity for a design variable is complex, such methods may be problematic and are generally unreliable as means of approximating π(ϕ) (Taflanidis & Beck 2007a). In the SSO framework, such an approximation for π(ϕ) is avoided. The sensitivity analysis is performed by looking at the average value of J(ϕ) over I, H(I), which for any subset of the design space I ⊂ Φ with volume VI is defined as: ( (
J(ϕ)dϕ 1 1 π(ϕ) I Eθ [hs (ϕ, θ)]dϕ dϕ (14) = I = H(I) = Eϕ,θ [hs (ϕ, θ)] VI VI VI I p(ϕ) To simplify the evaluation of H(I), a uniform distribution is chosen for p(ϕ). Note that p(ϕ) does not reflect the uncertainty in ϕ but is simply a device for formulating the augmented problem and thus can be selected according to convenience. Finally H(I) and an estimate of it based on the samples from π(ϕ), obtained as described previously, are given, respectively, by: VΦ H(I) = VI




NI /VI ˆ π(ϕ)dϕ and H(I) = NΦ /VΦ I

(15)

where NI and NΦ denote the number of samples from π(ϕ) belonging to the sets I and Φ, respectively, and VΦ is the volume of the design space Φ. The estimate for H(I) is equal to the ratio of the volume density of samples from π(ϕ) in sets I and Φ. The coefficient of variation (c.o.v.) for this estimate depends on the simulation technique used for obtaining the samples from π(ϕ). For a broad class of sampling algorithms this c.o.v. may be expressed as:  ˆ c.o.v. H(I) =

1 − P(ϕ ∈ I) ≈ N · P(ϕ ∈ I)



1 − NI /NΦ , N · NI /NΦ






P(ϕ ∈ I) =

π(ϕ)dϕ

(16)

I

where N = NΦ /(1 + γ), γ ≥ 0, is the effective number of independent samples. If direct MC techniques are used then γ = 0, but if MCMC sampling is selected then γ > 0 because of the correlation of the generated samples. Ultimately, the value of γ depends on the characteristics of the algorithm used. See (Au & Beck 2003b) for a formula for γ when the Metropolis-Hasting algorithm is used. For the uniform PDF for p(ϕ), note that H(I) is equal to the ratio: ( Eθ [hs (ϕ, θ)]dϕ/VI H(I) = ( I (17) Φ Eθ [hs (ϕ, θ)]dϕ/VΦ where the integrals in the numerator and denominator could be considered as the “average set content’’ in I and Φ respectively. Thus H(I) expresses the average sensitivity of Eθ [hs (ϕ, θ)] to ϕ within the set I ⊂ Φ.

162

Structural design optimization considering uncertainties

3.3 Sub set o pt imizat io n Consider a set of admissible subsets A in Φ that have some predetermined shape and some size constraint, for example related to the set volume, and define optimization: I ∗ = arg min H(I) I∈A

(18)

This definition is motivated by the fact that, as explained above, minimization of Eθ [hs (ϕ, θ)] is equivalent to minimization of J(ϕ) and that H(I) corresponds to the volume average integral of J(ϕ) over subset I. Based on the estimate in (15), optimization (18) is approximately equal to identification of the subset that contains the smallest volume density NI /VI of samples: NI ˆ Iˆ = arg min H(I) = arg min I∈A I∈A VI

(19)

Note that the relationship of the position in the design space of a set I ∈ A and the number of sample points in it is non-differentiable. Thus, methods appropriate for non-smooth optimization problems, such as genetic algorithms, should be chosen for optimization (19). The evaluation of the objective function for this problem involves small computational effort. Thus, the optimization can be efficiently solved if an appropriate algorithm is chosen. If set A is properly chosen, for example if its shape is “close’’ to the contours of Eθ [hs (ϕ, θ)] in the vicinity of ϕ∗ , then ϕ∗ ∈ I ∗ for the optimization in (18). This argument is not necessarily true for the optimization in (19) because only estimates of H(I) are used. Iˆ is simply the set, within the admissible subsets A, that has the largest likelihood, in terms of the information available through the obtained samples, of including ϕ∗ . This likelihood defines the quality of the identification and ultimately depends (Taflanidis & Beck 2007b) on the ratio of average set content, given by H(I) (see (17)). Taking into account the fact that the set content in the neighborhood of ˆ the optimal solution is the smallest in Φ, it is evident that smaller values of H(I) ∗ ˆ correspond to greater plausibility for the set I to include ϕ . Since only estimates of ˆ are available in the stochastic identification, the quality depends, ultimately, on H(I) ˆ I) ˆ and (b) its coefficient of variation (defining the accuracy both: (a) the estimate H( of that estimate). Smaller values for these parameters correspond to better quality of identification. Both of them should be used as a measure of this quality. 3.4

D eta i l s f or R O P

When SSO is implemented for ROP, selection of IF (ϕ, θ) as the loss function is beneficial because it simplifies the task of simulating samples from π(ϕ, θ). In this case these samples correspond simply to failed samples, i.e. samples that lead to failure of the system (IF (ϕ, θ) = 1), and the auxiliary PDF π(ϕ, θ) is simply the PDF for the augmented uncertain parameter vector conditioned on failure of the system, i.e. p(ϕ, θ|F). Similarly, the marginal π(ϕ) corresponds to p(ϕ|F). Monte Carlo simulation can then be used for simulating samples from p(ϕ|F). For design problems that involve small failure probabilities this approach may be inefficient because 1/PF trials are needed on

Stochastic system design optimization using stochastic simulation

163

the average in order to simulate one failed sample, where PF is the failure probability in the augmented design problem, defined, similarly to (10), as:

PF = I(ϕ, θ)p(ϕ, θ) dθdθϕ (20) Φ

Θ

Other stochastic simulation methods, such as Subset Simulation (Au & Beck 2001) should be preferred in such cases. 3.5

Implementation is s ues

3.5.1

Resolu tion for desi gn v ari abl es and i te ra t i v e i d e n t i fi ca t i o n

The size of the admissible subsets I define (a) the resolution of ϕ∗ and (b) the informaˆ tion about the accuracy of H(I) that is extracted from the samples from π(ϕ). Selecting smaller size for the admissible sets leads to better resolution for ϕ∗ . At the same time, though, this selection leads to smaller values for the ratio NI /NΦ (since smaller number of samples are included in smaller sets) and thus it increases the c.o.v. (reduces accuracy) of the estimation, as seen from (16). In order to maintain the same quality for the estimate, the effective number of independent samples must be increased, which means that more simulations must be performed. Since we are interested in subsets in Φ with small average set content, the required number of simulations to gather accurate information for subsets with small size is large. To account for this characteristic and to increase the efficiency of the identification process, an iterative approach can be adopted. At iteration k, additional samples in Iˆk−1 (where I0 = Φ) that are distributed according to π(ϕ) are obtained. A region Iˆk ⊂ Iˆk−1 for the optimal design parameters is then identified as above. The quality of the identification is improved by applying such an iterative scheme, since the ratio of the samples in Iˆk−1 to the one in Iˆk is larger ˆ Iˆk ) smaller) than the equivalent ratio when comparing Iˆk and (and thus the c.o.v. for H( the original design space Φ. The samples [ϕ, θ] available from the previous iteration, whose ϕ component lies inside set Iˆk−1 , can be exploited for improving the efficiency of the sampling process. In terms of the algorithms described in Appendix A this may be established for example by (a) forming better proposal PDFs and/or (b) using the samples already available as seeds for MCMC simulation. Some guidelines for the MCMC simulation are given later on, in the context of the example considered. Another way to improve the efficiency in this iterative process is to continually update hs (ϕ, θ) in (8) by re-defining s: hs,k (ϕ, θ) = h(ϕ, θ) − sk

where sk = min h(ϕ, θ) ϕ∈Iˆk ,θ

(21)

Figure 7.2 illustrates this concept. For choice hs,2 (ϕ, θ), which corresponds to a larger value of s, the sensitivity of the objective function, in the SSO setting, is larger and a candidate region for the optimal choice is more easily discernible (better quality is established) based on samples from π(ϕ). If hs (ϕ, θ) is reformulated, though, the ancillary density π(ϕ, θ) changes and the samples from the previous iteration cannot provide useful information for the next iteration unless the previous and the next loss

Structural design optimization considering uncertainties

0.3

0.1

0.1 15 ␸

0 10

20

100

80

80

60

60

N1

100

(b)

40

40

20

20

0 10

s
0.10 hs,2(␸, ␪)
h(␸, ␪)0.10

0.2

0.2

0 10

(a)

s
0 hs,1(␸, ␪)
h(␸, ␪)

E [hs(␸,␪)]

E [hs(␸, ␪)]

0.3

N1

164

15 ␸

20

0 10

15 ␸

20

15 ␸

20

Figure 7.2 Influence of selection of s in SSO (a) objective function Eθ [hs (ϕ, θ)] and (b) histograms of samples from π(ϕ) obtained through MC simulation.

functions hs (ϕ, θ) are similar. For cases where the sensitivity of the objective function is small, our experience indicates that the re-formulation of the loss function can be beneficial (assuming that s can be set to a larger value). When the sensitivity is quite high, though, it is preferable to keep the same loss function and use the samples available to improve the efficiency when generating new samples. 3.5.2 S el ec tion o f a d m is s ib le s u b s e t s Proper selection of the geometrical shape and size of the admissible sets is important for the efficiency of SSO. The geometrical shape should be chosen so that the challenging, non-smooth optimization (19) can be efficiently solved and still the sensitivity of the objective function to each design variable is fully explored. For example, a hyper-rectangle or a hyperellipse can be appropriate choices, with the latter expected to be closer to the objective function contours but requiring more computational complexity, especially in high dimensions. The size of admissible subsets is related to the quality of identification as discussed earlier. Selection of size for the admissible subsets can be determined by incorporating a constraint for either (i) the volume ratio δ = VI /VΦ or (ii) the number of samples ratio ρ = NI /NΦ . The first choice cannot be directly related to any of the measures of quality of identification; thus proper selection of δ is not straightforward. The second choice allows for directly controlling the coefficient of variation (see (16)) and thus one of the parameters influencing the quality of identification. This characterization for

Stochastic system design optimization using stochastic simulation

165

the admissible subsets is adopted here. The subset optimization in (19) corresponds, finally, to identification of a subset that has smallest estimated average value, within the class of subsets that guarantee a specific c.o.v. for that estimate: Iˆ = arg min NI /VI = arg max VI , I∈Aρ

I∈Aρ

Aρ = {I ⊂ Φ : ρ = NI /NΦ }

(22)

The same comment as for the optimization in (19) applies in this case; an algorithm appropriate for non-smooth optimizations should be selected. The volume (size) of the admissible subsets in this scheme is adaptively chosen so that the ratio of samples in the identified set is equals to ρ. The choice of the value for ρ affects the efficiency of the identification. If ρ is large, fewer number of samples is required for the same level of accuracy (c.o.v. in (16)). However, a large value of ρ means that the size of the identified subsets will decreases slowly (larger size sets are identified). A slow sequence requires more steps to converge to the optimal solution. It can thus be seen that the choice of the constraint ρ is a trade-off between the number of samples required in each step and the number of steps required to converge to the optimal design choice. In the applications we have investigated so far it was found that choosing ρ = 0.1–0.2 yields good efficiency. 3.5.3

Influence of di mensi on of desi gn v ari a b l e s v e ct o r

For a specific reduction of the volume of the search space in some step of the set identification, δk = VIk /VIk−1 , the mean reduction of the length for each design variable √ is nϕ δk . The mean total length reduction over all variables in niter iterations is: + , niter / . , n iter nϕ = (δmean )niter /nϕ δk ≈ ϕ δnmean

(23)

k=1

where δmean is the geometric mean of the volume reductions over all of the iterations. It is evident from (23) that for the same mean total length reduction, the number of iterations is proportional to the dimension of the design space. This proportionality relationship has been verified in the examples considered in (Taflanidis & Beck 2007a; Taflanidis & Beck 2007b). Assuming that the mean total length reduction over all variables describes adequately the computational efficiency of SSO, this argument shows that this efficiency decreases linearly with the dimension of the design space, so SSO should be considered appropriate for problems that involve a large number of design variables. 3.5.4

SSO algori thm and sampl e i mpl ement a t i o n e x a m p l e

The SSO algorithm is summarized as follows (Figure 7.3 illustrates some of the key steps of the algorithm). Initialization: Define the desired geometrical shape for the subsets I. Decide on the constraint value ρ in (22) and the desired number of samples N. The latter should be

166

Structural design optimization considering uncertainties

(b) Identification of set Î1

(a) Initial samples through MC w2

8 6

8 6 4 2

4 2 100

w2

200

300

400

500

600

200

300

400

500

600

(d) Magnification of previous plot

(c) Retained samples in set Î1

8 6

4

4

3

2

2 100

200

300

400

500

600

(e) New samples through MCMC and set Î2 w2

100

4 3 2 100

200

300 w1

400

500

100

200

300

400

500

(f) Last step, stopping criteria are satisfied 4.5 4 3.5 3 2.5 150

200 w1

250

Figure 7.3 Illustration of some key steps in SSO for an example with a two-dimensional design space. The X in the plots corresponds to the optimal solution.

ˆ chosen such that the c.o.v. for H(I) in (16) is equal to some pre-specified value for the given value of ρ. For example, for c.o.v. 5% and choice of ρ = 0.2, N should be 1600 if direct MC techniques are used. Step 1: Simulate, for example using MC simulation, NΦ = N samples from π(ϕ, θ). Identify subset Iˆ1 as the solution of the optimization problem (22) and keep only the NIˆ1 samples whose ϕ component belongs to the subset Iˆ1 . Step k: Use some sampling technique, such as Metropolis-Hastings algorithm, to obtain in total N samples from π(ϕ, θ) inside the subset Iˆk−1 . Identify subset Iˆk : Iˆk = arg max VIk , I∈Aρ,k

Aρ,k = {I ⊂ Iˆk−1 : ρ = NI /N}

(24)

Keep only the NIˆk samples whose ϕ component belongs to the subset Iˆk and exploit them in the next iteration. Stopping criteria: At each step, estimate ratio ˆ Iˆk ) = H(

NIˆk VIˆk−1 NIˆk−1 VIˆk

(25)

and its coefficient of variation according to the appropriate expression (depending on the algorithm used). Based on these two quantities and the desired quality of the identification (see next section), decide on whether to stop or to proceed to step k + 1. Table 7.1 presents the results for a sample run of the SSO algorithm for the design problems considered in (Taflanidis & Beck 2007a). Problem D1 involves 2 design

Stochastic system design optimization using stochastic simulation

167

Table 7.1 Results from a sample run of the SSO algorithm for two design problems. Problem D1 (nϕ = 2) δk =VˆIk /VˆIk−1 √ nϕ δk ˆ ˆIk ) H(  nϕ V /V ˆI3 Φ

Iteration of SSO

Problem D2 (nϕ = 4)

1

2

3

0.370 0.608 0.541

0.314 0.561 0.636 0.167

0.270 0.489 0.835

δk =VˆIk /VˆIk−1 √ nϕ δk ˆ ˆIk ) H(  nϕ V /V ˆI6 Φ

Iteration of SSO 1

2

3

4

0.442 0.815 0.453

0.378 0.784 0.529

0.346 0.323 0.767 0.754 0.577 0.619 0.183

5

6

0.269 0.720 0.743

0.223 0.687 0.896

variables, nϕ = 2, whereas problem D2 involves 2 more design variables, in total nϕ = 4. Figure 7.3 presents graphically the results for the sample run in problem D1 . In these examples the shape of the subsets I was selected as hyper-rectangles and ρ was chosen equal to 0.2. Figure 7.3 clearly illustrates the dependence of the quality of the identification on ˆ Iˆk ) which expresses (see (15)) the difference in volume density of samples from π(ϕ) H( inside and outside of the identified set Iˆk (corresponding to the interior rectangle in these plots). In the first iteration, this difference is clearly visible. As SSO evolves, the difference becomes smaller and by the last iteration (Figure 7.3f), it is difficult to visually discriminate which region in the set has smaller volume density of samples from π(ϕ). This corresponds to a decrease in the quality of the identification. In order to maintain plausibility for the identified set to contain the optimal solution, the iterative process stops. This figure also shows the capability of SSO to explore the sensitivity of the objective function with respect to each design variable. Within the initial design space (Figure 7.3a) the sensitivity with respect to design variable ϕ2 appears to be significantly higher, based on the density of failed samples. The set identified by SSO corresponds to larger size reduction for that variable (Figure 7.3b), thus it efficiently captures the difference in sensitivities. Note that in order to take advantage of this capability, no proportionality relationship should be enforced for the dimensions of the admissible subsets in different directions. Looking now at the results in Table 7.1, it is clear that as the identification process in SSO evolves, the reduction in the size of the identified subsets, expressed by δk , becomes larger and δk approaches the value for ρ (selected here as 0.2). Also the value ˆ Iˆk ) increases, which corresponds to reduction in the quality of identification. of H( All these patterns can be theoretically justified (Taflanidis & Beck 2007a) assuming that as the SSO identification progresses, regions of the design space with smaller sensitivity to the objective function are approached. The influence of the number of design variables in the efficiency of√ SSO is also evident from the results in Table 7.1. As mentioned earlier, the quantity nϕ δk corresponds to the mean length reduction per design variable. For D2 that reduction is much smaller, even though the values for δk are of the same level for the two design problems, which leads to more iterations until a set with small sensitivity to the objective function is identified. The average total length reduction is similar (look at the last row in Table 7.1 which is equivalent to the expression (23)) but the number of required iterations for D2 is double. Since design

168

Structural design optimization considering uncertainties

case D2 has double the number of design variables, this verifies the proportionality dependence (argued in Section 3.5.3) between required number of iterations and the number of design variables to establish the same average reduction per design variable. Note also that the mean reduction in size per design variable (last row of Table 7.1) is significant. This means that the size of the subset identified by SSO is considerably smaller than the original design space. Since this identification requires a small number of iterations it verifies the efficiency of the SSO algorithm. 3.6 C o nv ergen c e t o o pt imal s o lut io n The algorithm described above can adaptively identify a relatively small sub-region for the optimal design variables ϕ∗ within the original design space. The efficiency of convergence to ϕ∗ depends a lot on the sensitivity of the objective function around the optimal point. If that sensitivity is large then SSO will ultimately converge to a “small’’ ˆ satisfying at the same time the accuracy requirements that make it highly likely set I, ˆ The center point of this set, denoted herein as ϕSSO , gives the estimate that ϕ∗ is in I. for the optimal design variables. In cases that this sensitivity is not large enough, such convergence will be problematic and will require increasing the number of samples in order to satisfy the requirement for the quality of identification. Another important topic related to the identification in such cases is that there is no measure of the quality of the identified solution, i.e. how close ϕSSO is to ϕ∗ , that can be directly established through the SSO results. If the identification is performed multiple times and a sequence {ϕSSO,i } is obtained, the c.o.v. of {Eˆ θ [h(ϕSSO,i ,θ)]} could be considered a good candidate for characterizing this quality. This might not be always a good measure though. For example, if the choice for admissible subsets is inappropriate for the problem considered, it could be the case that consistent results are obtained for ϕSSO (small c.o.v.) that are far from the optimal design choice ϕ∗ . Also, this approach involves higher computational cost because of the need to perform the identification multiple times. For such cases, it would be more computationally efficient (instead of increasing N in SSO and performing the identification multiple times) and more accurate (in terms of identifying the true optimum), to combine SSO with some other optimization algorithm for pinpointing ϕ∗ . A discussion of topics related to such algorithms is presented next.

4 Stochastic optimization algorithms We go back to the original formulation of the objective function, i.e. (1). In principle, though, the techniques discussed here are applicable to the case that the loss function h(θ, ϕ) is replaced by hs (θ, ϕ) used in the SSO setting (given by (8)). 4.1

C om m o n r and o m numb e r s

The efficiency of stochastic optimizations such as (5) can be enhanced by the reduction of the absolute and/or relative importance of the estimation error eN (ϕ, ΩN ). The absolute importance may be reduced by obtaining more accurate estimates of the objective function, i.e. by reducing the variance of the estimates. This can be established in various ways, for example by using importance sampling or by selecting a larger sample size N in (4), but these typically involve extra computational effort. It is, thus,

Stochastic system design optimization using stochastic simulation

169

more efficient to seek a reduction in the relative importance of the estimation error. ˆ θ,N [h(ϕ1 , Ω1 )] This means reducing the variance of the difference of the estimates E N 2 2 ˆ θ,N [h(ϕ , Ω )] that correspond to two different design choices ϕ1 and ϕ2 . This and E N variance can be decomposed as: var(Eˆ θ,N [h(ϕ1 , Ω1N )] − Eˆ θ,N [h(ϕ2 , Ω2N )]) = var(Eˆ θ,N [h(ϕ1 , Ω1N )]) ˆ θ,N [h(ϕ1 , Ω1 )], E ˆ θ,N [h(ϕ2 , Ω2 )]) + var(Eˆ θ,N [h(ϕ2 , Ω2N )]) − 2cov(E N N

(26)

ˆ θ,N [h(ϕ2 , Ω2 )] are evaluated independently their covariance If Eˆ θ,N [h(ϕ1 , Ω1N )] and E N is zero; deliberately introducing dependence, increases the covariance (i.e. increases their correlation) and thus decreases their variability (the variance on the left). This decrease in the variance improves the efficiency of the comparison of the two estimates; it may be considered as creating a consistent estimation error. In a simulation-based context this task is achieved by adopting common random number streams (CRN), i.e. Ω1N = Ω2N , in the simulations generating the two different estimates. Figure 7.4a shows the influence of such a selection: the curves that correspond to CRN are characterized by consistent estimation error and are smoother. Also note that the absolute influence of the estimation error for the case that corresponds to larger N (curve (iii)) is, as expected, smaller. Two important questions regarding the use of CRN are: will the variance be reduced (efficiency)? Is this the best one can do (optimality)? The answer to both these questions depends on the way the random sample θ (input) influences the sample value of the loss function h(ϕ, θ) (output) in each simulation. Optimality can be proved only in special cases but efficiency can be guaranteed under mild conditions (Glasserman & Yao 1992). Continuity and monotonicity of the output with respect to the random number input are key issues for establishing efficiency. If h(ϕ, θ) is sufficiently smooth then the two aforementioned requirements for CRN-based comparisons can be guaranteed,

(a) 0.09 0.08 0.07

0.1

0.06

0.08

0.05

0.06

0.04 10

12

14

␸

16

18

(i) analytical (ii) CRN sim using IF (iii) CRN sim using Pε

0.12

P(F|␸)

E [h(␸,␪)]

(b) 0.14

(i) analytical (ii) sim N
4000 (iii) CRN sim N
4000 (iv) CRN sim N
1000

20

0.04 10

12

14

␸

16

18

20

Figure 7.4 Illustration of some points in CRN-based evaluation of the objective function for (a) general stochastic design problems and (b) reliability objective problem.

170

Structural design optimization considering uncertainties

as long as the design choices compared are not too far apart in the design variable space. In such cases it is expected that use of CRN will at least be advantageous (if not optimal). If the systems compared are significantly different, i.e. correspond to ϕ that are not close, then CRN does not necessarily guarantee a consistent estimation error. This might occur if the regions of Θ that contribute most to the integral of the expected value for the two systems are drastically different and the CRN streams selected do not efficiently represent both of these regions. This feature is also indicated in Figure 7.4a; the estimation error is not consistent along the whole range of ϕ for the CRN curves (compare the objective function for curve (iv) for large and small values of ϕ) but for local comparisons it appears to be consistent. For ROP, CRN does not necessarily have a similar effect on the calculated output if formulation (6) is adopted since the indicator function IF (ϕ, θ) is discontinuous. Thus the aforementioned requirements for establishing efficiency of CRN cannot be guaranteed. It is thus beneficial to use the formulation (7) for the probability of failure in CRN-based optimizations. For design problems where no prediction error in the model response is actually assumed, a small fictitious error should be chosen so that the optimization problems with and without the model prediction error are practically equivalent, i.e. correspond to the same optimum. Figure 7.4b illustrates this concept; the influence on P(F|ϕ) of the two different loss functions in Figure 7.1b and the advantage of selecting Pε (g(ϕ, ˜ θ)) is clearly demonstrated. 4.2

Ex teri or sampling appr o ximat io n

Solution approaches to optimization problems using stochastic simulation are based on either interior or exterior sampling techniques (Ruszczynski & Shapiro 2003). Interior sampling methods resample ΩN at each iteration of the optimization algorithm. On the other hand, exterior sampling approximations (ESA) adopt the same stream of random numbers throughout all iterations in the optimization process, thus transforming problem (5) into a deterministic one, which can be solved by any appropriate routine. These methods are also commonly referred to as sample average approximations (Royset & Polak 2004) and they are closely related to CRN. The CRN cases in Figure 7.4 correspond actually to ESA. Several asymptotic results are available for ESA and their rate of convergence under weak assumptions. For finite-dimensional sample sizes, the optimal solution depends on the sample ΩN selected. Figure 7.4a clearly demonstrates this issue (compare the optimum values in the CRN curves (iii) and (iv)). Usually ESA is implemented by selecting N “large enough’’, typically much higher than it would be for interior sampling methods, in order to get better quality estimates for the objective function and thus more accurate solutions to the optimization problem. See (Ruszczynski & Shapiro 2003) for more details and (Royset & Polak 2007) for a computationally efficient iterative approach that adaptively implements higher accuracy estimates as the algorithm converges to the optimal solution. The quality of the solution obtained through ESA is commonly assessed by solving the optimization problem multiple times, for different independent random sample streams. Even though the computational cost for the ESA deterministic optimization is typically smaller than that of the original stochastic search problem, the overall efficiency may be worse because of the requirement to perform the optimization multiple times.

Stochastic system design optimization using stochastic simulation

171

4.3 Appropriate s tochas tic optimizatio n al g o ri thms Both gradient-based and gradient-free algorithms can be used in conjunction with CRN or ESA and can be appropriate for stochastic optimizations. Gradient-based algorithms use derivative information to iterate in the direction of steepest descent for the objective function. Only local designs are compared in each iteration, which makes the implementation of CRN efficient and allows for application of stochastic approximation which can significantly improve the computational efficiency of stochastic search methods (Kushner & Yin 2003). The latter can be established by applying an equivalent averaging across the iterations of the algorithm instead of establishing higher accuracy estimates at each iteration. In simple examples, the loss function h(ϕ, θ) (or even the limit state function g(ϕ, ˜ θ) in ROP) are such that the gradient of the objective function with respect to ϕ can be obtained through a single stochastic simulation analysis (Royset & Polak 2004, Spall 2003). In many structural design problems though, the models used are generally complex, and it is difficult, or impractical, to develop an analytical relationship between the design variables and the objective function. Finite difference numerical differentiation is often the only possibility for obtaining information about the gradient vector but this involves computational cost which increases linearly with the dimension of the design parameters. Simultaneous-perturbation stochastic approximation (SPSA) (Kleinmann et al. 1999, Spall 2003) is an efficient alternative search method. It is based on the observation that one properly chosen simultaneous random perturbation in all components of ϕ provides as much information for optimization purposes in the long run as a full set of one at a time changes of each component. Thus, it uses only two evaluations of the objective function, in a direction randomly chosen at each iteration, to form an approximation to the gradient vector. Gradient-free optimization methods, such as evolutionary algorithms, direct search and objective function approximation methods are based on comparisons of design choices that are distributed in large regions of the design space. They require information only for the objective function which makes them highly appropriate for stochastic optimizations (Beck et al. 1999, Lagaros et al. 2002) because they avoid the difficulty of obtaining derivative information. They involve, though, significant computational effort if the dimension of the design variables is high. Use of CRN in these algorithms may only improve the efficiency of the comparisons in special cases; for example, if the size (volume) of the design space is “relatively small’’ and thus the design variables being compared are always close to each other. More detailed discussion of algorithms for stochastic optimization can be found in (Spall 2003; Ruszczynski & Shapiro 2003). Only SPSA is briefly summarized here. 4.3.1

Sim u lt a n eous-perturbati on stochasti c a p p ro x i m a t i o n u s i n g co m m o n ra n d om numbers

The implementation of SPSA using CRN takes the iterative form: ϕk+1 = ϕk − αk gk (ϕk , ΩN ) ϕk+1 ∈ Φ

(27)

172

Structural design optimization considering uncertainties

where ϕ1 is the chosen point to initiate the algorithm and the jth component for the CRN simultaneous perturbation approximation to the gradient vector in the kth iteration, gk (ϕ, ΩkN ), is given by: gk,j =

ˆ θ,N (ϕk − ck ∆k , Ωk ) ˆ θ,N (ϕk + ck ∆k , Ωk ) − E E N N 2ck κ,j

(28)

where ∆k ∈ Rnϕ is a vector of mutually independent random variables that defines the random direction of simultaneous perturbation for ϕk and that satisfies the statistical properties given in (Spall 2003). A symmetric Bernoulli ±1 distribution is typically chosen for the components of ∆k . The selection for the sequences {ck } and {αk } is discussed in detail in (Kleinmann et al. 1999). A choice that guarantees asymptotic convergence to ϕ∗ is αk = α/(k + w)β and ck = c1 /kζ , where 4ζ − β > 0, 2β − 2ζ > 1, with w, ζ > 0 and 0 < β < 1. This selection leads to a rate of convergence that asymptotically approaches k−β/2 when CRN is used (Kleinmann et al. 1999). The asymptotically optimal choice for β is, thus, 1. In applications where efficiency using a small number of iterations is sought after, use of smaller values for β are suggested in (Spall 2003). For complex structural design optimizations, where the computational cost for each iteration of the algorithm is high, the latter suggestion should be adopted. Implementation of CRN contributes to reducing the variance of the gradient approximation in (28) and thus the variability in estimating ϕκ ; for example, the rate of convergence is k−β/3 when CRN is not used. Regarding the rest of the parameters for the sequences {ck } and {αk }: w is typically set to 10% of the number of iterations selected for the algorithm and the initial step c1 is chosen “close’’ to the standard deviation of the measurement error eN (ΩN , ϕ1 ). This last selection prevents the finite difference gradient from getting excessively large in magnitude but might be inefficient if the standard deviation of the error changes dramatically with ϕ. The value of α can be determined based on the estimate of g1 and the desired step size for the first iteration. Some initial trials are generally needed in order to make a good selection for α, especially when little prior insight is available for the sensitivity of the objective function to each of the design variables. Typically SPSA is implemented adopting interior sampling techniques. Convergence of the iterative process is judged based on the value ϕk+1 − ϕk  in the last few steps, for an appropriate selected vector norm. Note that since the progress of the algorithm at each step depends on the sample ΩkN and the randomly chosen perturbation direction, converˆ θ,N [h(ϕk+1 , Ωk+1 )] − Eˆ θ,N [h(ϕk , Ωk )]| gence cannot be judged based on the value of |E N N (because the two estimates are evaluated using different steams of random samples and thus include different estimation error) or the value of ϕk+1 − ϕk  at the last step only (because this value depends on the random search direction chosen). This notion of convergence, though, depends on the selection of the sequence {αk }; for example, selection of small step sizes might in some cases give a false impression that convergence has been established, even though this is not true. Such problems can be avoided by restarting the SPSA algorithm at the converged optimal solution to monitor the behavior for some small number of iterations. Blocking rules can also be applied in order to avoid potential divergence of the algorithm, especially in the first iterations (Spall 2003).

Stochastic system design optimization using stochastic simulation

173

5 Framework for stochastic optimization using stochastic simulation 5.1

Outline of the framework

As already mentioned, a two-stage framework for stochastic system design may be established by combining the algorithms presented in the previous two sections. In the first stage, SSO is implemented in order to efficiently explore the sensitivity of the objective function and adaptively identify a subset ISSO ⊂ Φ containing the optimal design variables. In the second stage, any appropriate stochastic optimization algorithm is implemented in order to pinpoint the optimal solution within ISSO . The specific algorithm selected for the second stage determines the level of quality that should be established in the SSO identification. If a method is chosen that is restricted to search only within ISSO (typically characteristic of gradient-free methods), then better quality is needed. The iterations of SSO should stop at a larger size set, and establish greater plausibility that the identified set includes the optimal design point. If, on the other hand, a method is selected that allows for convergence to points outside the identified set, lower quality may be acceptable in the identification. Our experience ˆ Iˆk ) with a c.o.v. of 4% for that estimate, indicates that a value around 0.75–0.80 for H( ˆ indicates high certainty that Ik includes the optimal solution. Of course, this depends on the characteristics of the problem too and particularly on the selection of the shape of admissible subsets, but this guideline has proved to be robust in the applications we have considered so far. The efficiency of the stochastic optimization considered in the second stage is influenced by (a) the size of the design space Φ defined by its volume VΦ , and, depending on the characteristics of the algorithm chosen, by (b) the initial point ϕ1 at which the algorithm is initiated, and (c) the knowledge about the local behavior of the objective function in Φ. For example, topic (b) is important for gradient-based algorithms whereas topic (c) is relevant for iterative algorithms that require user insight for selecting appropriate step sizes (like SPSA). The SSO stage gives valuable insight for all these topics and can, therefore, contribute to increasing the efficiency of convergence to the optimal solution ϕ∗ . The set ISSO has smaller size (volume) than the original design space Φ. Also, it is established that the sensitivity of the objective function with respect to all components of ϕ is small. This allows for efficient normalization of the design space (in selecting step sizes) and choice of approximating functions. In Taflanidis & Beck (2007a), 60% reduction of the overall computational cost for convergence to the optimal solution was reported when comparing the combined framework discussed here to SPSA optimization (without the SSO stage). In that study, the following guidelines were suggested for tuning of the SPSA parameters using information from SSO: ϕ1 should be selected as the center of the set ISSO and parameter α chosen so that the initial step for each component of ϕ is smaller than a certain fraction (chosen as 1/10) of the respective size of ISSO , based on the estimate for g1 from (28). This estimate should be averaged over ng (chosen as 6) evaluations because of its importance in the efficiency of the algorithm. Also, no movement in any direction should be allowed that is greater than a quarter of the size of the respective dimension of ISSO (blocking rule).

174

Structural design optimization considering uncertainties

The information from the SSO stage can also be exploited in order to reduce the variance of the estimate Eθ [h(ϕ, ΩN )] by using importance sampling. This choice is discussed next. 5.2

Im po rtanc e s ampling

Importance sampling (IS) is an efficient variance reduction technique. It is based on choosing an importance sampling density pis (θ|ϕ) to generate samples in regions of Θ that contribute more to the integral of Eθ [h(ϕ, θ)]. The estimate for Eθ [h(ϕ, θ)] is given in this case by:  ˆ θ,N [h(ϕ, ΩN )] = 1 E h(ϕ, θi )R(θi |ϕ) N N

(29)

i=1

where the samples θi are simulated according to pis (θ|ϕ) and R(θi |ϕ) =

p(θi |ϕ) pis (θi |ϕ)

(30)

is the importance sampling quotient. The main problem is how to choose a good IS density. The optimal density is simply the PDF that is proportional to the absolute value of the integrand of (1) |h(ϕ, θ)|p(θ|ϕ) (Robert & Casella (2004)) leading to a selection: pis,opt (θ|ϕ) =

|h(ϕ, θ)|p(θ|ϕ) Eθ [|h(ϕ, θ)|]

(31)

Samples for θ that are distributed proportional to hs (ϕ, θ)p(θ|ϕ) when ϕ ∈ ISSO are available from the last iteration of the SSO stage. They simply correspond to the θ component of the available sample pairs [ϕ, θ]. Re-sampling can be performed within these samples, using weighting factors |h(ϕi , θi )|/hs (ϕi , θi ) for each sample, in order to approximately simulate samples proportional to |h(ϕ, θ)|p(θ|ϕ) when ϕ ∈ ISSO . The efficiency of this re-sampling procedure depends on how different hs (ϕi , θi ) and h(ϕi , θi ) are. In most cases the difference will not be significant and good efficiency can be established. Alternatively, hs (ϕi , θi ) can be used as loss function in the second stage of the optimization. In this case there is no need to modify the samples from SSO. This choice would be inappropriate if s was negative because it makes the loss function less sensitive to the uncertain parameters θ, thus possibly reducing the efficiency of IS. In such design problems it is better to use the original loss function h(ϕ, θ). The samples simulated proportional to |h(ϕ, θ)|p(θ|ϕ) can be finally used to create an importance sampling density pis (θ|ϕ) to use in (30), since the set ISSO is small. Various strategies have been discussed in the literature for such an adaptive importance sampling (see for example (Au & Beck 1999)). For problems with high-dimensional vector θ, the efficiency of IS can be guaranteed only under strict conditions (Au & Beck (2003a)). An alternative approach can be applied for such cases: the uncertain parameter vector is partitioned into two sets, Θ1 and Θ2 . Θ1 is comprised of parameters that individually do not significantly influence the loss function (they have significant influence only when viewed as a group), for

Stochastic system design optimization using stochastic simulation

175

example, the white noise sequence modeling the stochastic excitation in dynamic reliability problems, while Θ2 is comprised of parameters that have individually a strong influence on h(ϕ, θ). The latter set typically corresponds to a low-dimensional vector. IS is applied for the elements of Θ2 only. This approach is similar to the one discussed in (Pradlwater et al. 2007) and circumvents the problems that may appear when applying IS to design problems involving a large number of uncertain parameters.

6 Illustrative example: optimization of the life-cycle cost of an office building The retrofitting of a symmetric, four-story, office building with linear viscous dampers is considered. The building is a non-ductile reinforced concrete, perimeter momentframe structure. The dimension of the building is 45 m × 45 m and the height of each story is 3.9 m. The perimeter frames in the two building directions are separated from each other, which allows structural analysis in each direction to be done separately. Because of the symmetry of the building, analysis of only one of the directions is necessary. 6.1

Probabilistic s tructural model

A class of shear-frame models (illustrated in Figure 7.5) with hysteretic behavior and deteriorating stiffness and strength is assumed (using a distributed element model assumption for the deteriorating part (Iwan & Cifuentes 1986). The lumped mass of the top story is 935 ton while it is 1215 ton for the bottom three. The initial inter-story stiffnesses ki of all the stories are parameterized by ki = kˆ i θk,i , i = 1, 2, 3, where [kˆ i ] = [700.0, 616.1, 463.6, 281.8] MN/m are the most probable values and θk,i are nondimensional uncertain parameters, assumed to be correlated Gaussian variables with mean value one and covariance matrix with elements ij = (0.1)2 exp[−(i − j)2 /22 ]. For each story, the post-yield stiffness coefficient αi , stiffness deterioration coefficient βi , over-strength factor γi , yield displacement δy,i and

Fi m4

viscous damper

Fy,i

d4

m3

ki dy,i

m3

dp,i

biki retrofitting scheme

m1

dy,i

dp,i

du,i –Fr

d1

di du,i

i-th story restoring force

–Fu,i

Fi

m2

Fu, i
(1gi )Fy,i

aiki

Fu,i Fr

i du,i
dp,ihidy,i

Fu,i Illustration of deteriorating stiffness and strength characteristics

Figure 7.5 Structural model assumed in the study.

176

Structural design optimization considering uncertainties

displacement coefficient ηi have mean values 0.1, 0.2 0.3, 0.22% of story height and 2, respectively (see Figure 7.5 for proper definition of some of these parameters). All these parameters are treated as independent Gaussian variables with c.o.v. 10%. The structure is assumed to be modally damped. The damping ratios for all modes are treated similarly as Gaussian variables with mean values 5% and coefficients of variation 10%.

6.2

Prob ab i l i s t ic s it e s eis mic hazar d a n d gr o u n d m o t i o n m o d e l

In order to estimate the earthquake losses, probability models are established for the seismic hazard at the structural site and for the ground motion, as in (Au & Beck 2003b). Seismic events are assumed to occur following a Poisson distribution and so are independent of previous occurrences. The uncertainty in moment magnitude M is modeled by the Gutenberg-Richter relationship (Kramer 2003) truncated on the interval [Mmin , Mmax ] = [5.5, 8], leading to a PDF: p(M) =

b exp( − b · M) exp( − b · Mmin ) − exp( − b · Mmax )

(32)

and expected number of events per year v = exp (a − bMmin ) − exp (a − bMmax )

(33)

The regional seismicity factors are selected as b = 0.9 loge (10) and a = 4.35 loge (10), leading to v = 0.25. For the uncertainty in the event location, the epicentral distance, r, for the earthquake events is assumed to follow a lognormal distribution with median 20 km and logarithmic standard deviation 0.4. Figure 7.7a illustrates the PDFs for M and r. For modeling the ground motion, the methodology described in Boore (2003) is adopted (also characterized as the “stochastic method’’). This methodology, which was initially developed for generating synthetic ground motions, is reinterpreted here to form a probabilistic model for the earthquake excitation. According to this model, the time-history (output) for a specific event magnitude, M, and source distance, r, is obtained by modulating a white-noise sequence Zw (input) through the following steps: (i) the sequence Zw is multiplied by an envelope function e(t; M, r); (ii) this modified sequence is then transformed to the frequency domain; (iii) it is normalized by the square root of the mean square of the amplitude spectrum; (iv) the normalized sequence is multiplied by a radiation spectrum A(f ; M, r); and finally (v) it is transformed back to the time domain to yield the desired acceleration time history. The characteristics for A(f ; M, r) and e(t; M, r) are presented in Appendix B. Figure 7.6a shows functions A(f ; M, r) and e(t; M, r) for different values of M and r = 15 km. It can be seen that as the moment magnitude increases, the duration of the envelope function also increases and the spectral amplitude becomes larger at all frequencies with a shift of dominant frequency content towards the lower-frequency regime. Figure 7.6b shows a sample ground motion for M = 6.7 and r = 15 km.

Stochastic system design optimization using stochastic simulation

1

M
6 (b) M
6.7 200 M
7.5 100

102 101 100

M
6 M
6.7 M
7.5

101 102 103

cm/sec2

0.8 e(t; M,r)

A(f; M,r) (cm/sec)

(a)

0.6 0.4

0

0

100

0.2

101 101 100 f(rad/sec)

177

200 0

10

20

30

t(sec)

0

5

10 t(sec)

15

Figure 7.6 (a) Radiation spectrum and envelope function for various M and r = 15 km and (b) sample ground motion for M = 6.7, r = 15 km.

6.3

Expected life-cycle cos t

The objective function in the stochastic design problem is the expected life-cycle cost of the structure for a life-time of tlife = 60 years after the retrofit. This cost, C(ϕ), as a function of the design variables is given by (Porter et al. 2004):  1 − e−rd tlife C(ϕ) = Cd (ϕ) + L(ϕ, θ) vtlife p(θ)dθ rd tlife Θ


(34)

where Cd (ϕ) is the cost of the viscous dampers, rd equals the discount rate (taken here as 2.5%) and L(ϕ, θ) is the expected cost given the earthquake event and the system specified by the pair [ϕ, θ]. The uncertain parameter vector in this design problem consists of the structural model parameters, θs , the seismological parameter θg = [M, r] and the white noise sequence, Zw , so θ = [θs , θg , Zw ]. The term in the brackets in (34) is the present worth factor, which is used in order to calculate the present value of the expected future earthquake losses (Porter et al. (2004)). The earthquake damage and loss are calculated assuming that after each event the building is quickly restored to its undamaged state. The cost of the dampers at each floor is estimated based on their maximum force capacity Fud,i as Cd,i (ϕ) = $80(Fud,i )0.8 . This simplified relationship comes from fitting a curve to the cost of some commercially-available dampers. The viscosity of the dampers is selected assuming that the maximum force capacity is established at a velocity of 0.2 m/sec. The earthquake losses are estimated adopting the methodology described in (Goulet et al. 2007, Porter et al. 2004). The components of the structure are grouped into nas damageable assemblies. For each assembly j, nd,j different damage states are designated and a fragility function is established for each damage state dk,j . These functions quantify the probability that the component has reached or exceeded that damage state conditional on some engineering demand parameter (EDPj ). Damage state 0 always corresponds to an undamaged condition. Each fragility function is a conditional cumulative log-normal distribution with median xm and logarithm standard deviation bm , as

178

Structural design optimization considering uncertainties

Table 7.2 Characteristics of fragility functions and expected repair costs for each story. d k,j

xm

1 (light) 2 (moderate) 3 (significant) 4 (severe) 5 (collapse) 1 (damage)

bm

nel

Structural Components 1.4δy,i 0.2 22 (δy,i + δp,i )/2 0.35 22 δp,i 0.4 22 δu,i 0.4 22 3% 0.5 22 Contents 0.6g 0.3 100

$/nel

d k,j

xm

2000 9625 18200 21600 34300

1 (patch) 2 (replace)

Partitions 0.33% 0.2 0.7% 0.25

3000

1 (damage)

1 (damage)

bm

nel

$/nel

500 500

180 800

Acoustical Ceiling 1g 0.7 103 m2 Paint 0.33% 0.2 3500 m2

25 25

presented in Table 7.2. Indirect losses because of (a) fatalities and (b) building downtime, i.e. loss of revenue while the building is being repaired, are ignored in this study. The expected losses in the event of the earthquake are given by: n

L(ϕ, θ) =

d,j nas  

P[dk,j |ϕ, θ]Ck,j

(35)

j=1 k=1

where P[dk,j |ϕ, θ] is the probability that the assembly j will be in its kth damage state and Ck,j is the corresponding expected repair cost. Table 7.2 summarizes the characteristics for the fragility functions (xm , βm ) and the expected cost $/nel . The nel in this table corresponds to the number of elements that belong to each damageable assembly in each direction of each floor. For the structural contents and the acoustical ceiling, the maximum story absolute acceleration is used as EDP and for all other assemblies the maximum inter-story drift ratio is used. For estimating the total wall area requiring a fresh coat of paint, the simplified formula developed in (Goulet et al. 2007) is adopted. According to this formula a fraction of the undamaged wall area is also repainted, considering the desire of the owner to achieve a uniform appearance. This fraction depends on the extent of the damaged area and is chosen here based on a lognormal distribution with median 0.25 and logarithmic standard deviation 0.5. 6.4

Opti m a l d ampe r d e s ig n

The maximum force capacities of the dampers in each floor are the four design variables ϕ = [Fud,i : i = 1, . . . , 4]. The initial design space for each variable is set to [0, 13000] kN for Fud,1 and Fud,2 and [0, 8000] kN for Fud,3 and Fud,4 . Results for a sample run of the optimization algorithm are presented in Table 7.3. For the SSO stage the sets I3 and ISSO are reported here only. Also lIi denotes the length of set I in the direction of the ith design variable. 6.4.1 S toc h a stic s u b s e t o p t im iza t io n The objective function (34) can be written as:  
1 − e−rd tlife C(ϕ) = Eθ [hs (ϕ, θ)] = Cd (ϕ) + L(ϕ, θ) vtlife p(θ)dθ rd tlife Θ

(36)

Stochastic system design optimization using stochastic simulation

179

Table 7.3 Results from the optimization (sample run).

ϕ

I3 (kN)

F ud,1 F ud,2 F ud,3 F ud,4

[3610, 7657] [3557, 7756] [4034, 7095] [1566, 4751]

ISSO (kN) [5857, 6980] [4539, 6045] [4085, 5517] [1841, 2959]

ϕSSO

ϕ∗

(kN)

(kN)

6418 5292 4801 2400

6420 5195 4481 2060

Eˆ θ [h(ϕ∗ , θ)] Eˆ θ [h(ϕSSO , θ)]

l iSSO /liΦ

0.430 × 106 $

0.094 0.126 0.179 0.139

0.438 × 106 $

Thus, the loss function used in the SSO stage of the optimization is:  1 − e−rd tlife hs (ϕ, θ) = Cd (ϕ) + L(ϕ, θ) vtlife rd tlife

 VISSO /VΦ

nϕ

0.131

(37)

The parameter selections for SSO are: ρ = 0.2, N = 2000, s = 0. The shape for the sets I is selected as a hyper-rectangle and the adaptive identification is stopped when ˆ Iˆk ) becomes larger than 0.80. The optimization in (24) is performed using a genetic H( algorithm. In total, 6 iterations of the SSO algorithm are performed. After 3 iterations the loss functions hs (ϕ, θ) is reformulated by choosing s = $200.000. Algorithm 1 (Appendix A) is used for sampling in the 1st and 4th iterations and Algorithm 2 in all others. For the MCMC simulation (Algorithm 2) a global proposal PDF equal to p(Zw ) is chosen for the white noise sequence, to avoid the problems with the highdimensionality of the uncertain parameter vector, and local random walk proposal PDFs for all other parameters. A uniform PDF centered at the current sample, with wide spread (covering 0.7 of the current subset Iˆk−1 at each iteration k), is chosen for the proposal PDF for ϕ. This is a proposal PDF that is easy to sample from and still approximates the form of π(ϕ), which is expected to look like a convex function with small sensitivity as the identification converges to a set near the optimal design variables. A global uniform proposal PDF could also be chosen for ϕ, as regions with small sensitivity are approached. Such a global proposal PDF avoids rejecting samples due to their ϕ component, in the candidate sampling step, falling outside the given subset Iˆk−1 at iteration k, which can occur with a local uniform PDF and which increases the correlation in the generated Markov Chain. For the rest of the uncertain parameters, θs and θg , independent conditional Gaussian distributions are chosen, centered at the current sample with standard deviation equal to ½ the standard deviation of the samples retained from the previous step. Ultimately, the efficiency of the MCMC simulation (Algorithm 2) depends a lot on the quality of the selected proposal PDFs. In cases that such PDFs cannot be easily chosen then MC simulation can be used. The results in Table 7.3 show that SSO efficiently identifies a subset for the optimal design variables and leads to a significant reduction of the size (volume) of the search space (look at the last two columns of Table 7.3). The converged optimal solution in the second stage, ϕ∗ , is close to the center ϕSSO of the set that is identified by SSO; also the objective function at that center point Eθ [h(ϕSSO , θ)] is not significantly different from the optimal value Eθ [h(ϕ∗ , θ)]. Thus, selection of ϕSSO as the design choice leads to a sub-optimal design that is, though, close to the true optimum in terms of both the design vector selection and its corresponding performance. This agrees with the

180

Structural design optimization considering uncertainties

findings of all of our other studies and indicates that the sole use of SSO might be adequate for many problems (see Taflanidis & Beck 2007a) for a more thorough comparison and discussion).

6.4.2 S i m u lta n eo u s-p e r t u r b a t io n s t o ch a s t ic a ppr o x i mat i o n wi t h c ommon ran d o m n u m b e r s For the second stage of the optimization framework the formulation of the objective function in (34) is adopted. Stochastic simulation is used in order to estimate only the second part, since the cost of the dampers can be deterministically evaluated, so:  1 − e−rd tlife h(ϕ, θ) = L(ϕ, θ) vtlife rd tlife

(38)

Following the discussion in Section 5.2, importance sampling densities are established for the structural model parameters and the seismological parameters, M and r, but not for the high-dimensional white-noise sequence. Figure 7.7b illustrates this concept for M and r. A truncated lognormal distribution is selected for the IS PDF for M (with median 7 and logarithmic standard deviation 0.1) and a lognormal for r (with median 15 and logarithmic standard deviation 0.4). Note that the IS PDF for M is significantly different from its initial distribution; since M is expected to have a strong influence on h(θ, ϕ), the difference between the distributions is expected to have a big effect on the accuracy of the estimation. The respective difference between the PDFs for r is much smaller. For the structural model parameters this difference is negligible, and the IS PDFs were approximated to be Normal distributions, like p(θs ), with a slightly ˆ θ,N [h(ϕ, ΩN )] for a sample shifted mean value but the same variance. The c.o.v. for E

(a) 1500

1500

2 p(M) 1

1000 N 500

0 8

0

1000 N 500 0 5.5

6

6.5

7

7.5

0.04 p(r) 0.02

0

20

M

r

40

60

0

(b) 100

80 60 N 40

0.4

20

0.2

0 5.5

0.06

0.6

0 6

6.5

7 M

7.5

8

pis(M)

0.04

50 N

pis(r)

0.02 0

0 0

20

r

40

60

Figure 7.7 Details about importance sampling densities formulation for M and r (a) Initial PDF and samples and (b) samples from SSO stage and IS PDF.

Stochastic system design optimization using stochastic simulation

181

size N = 1000 is 16% without using IS and 4% when IS is used. This c.o.v. is of the same level for all values of ϕ√ ∈ ISSO , since the ISSO set is relatively small. Note that the c.o.v. varies according to 1/ N (Robert & Casella 2004); thus, the sample size for direct estimation (i.e. without use of IS) of the objective function with the same level of accuracy as in the case when IS is applied would be approximately 16 times larger. This illustrates the efficiency increase that can be established by the IS scheme discussed earlier. The converged optimal solution in the second stage is included in Table 7.3. Forty iterations were needed in the second stage of the framework using a sample size of N = 1000. This computational cost can be characterized as small. Convergence is judged by looking at the norm ϕk+1 − ϕk ∞ for each of the 5 last iterations. If that norm is less than 0.2% (normalized by the dimension of the initial design space), then we assume that convergence to the optimal solution has been established. 6.4.3 Effic ien c y of the tw o-stage opti mi zati on fra m e wo rk To evaluate the efficiency of the optimization framework, the same optimization was performed without the use of SSO in the first stage. In this case the starting point for SPSA was selected as the center of the design space Φ and α was chosen so that in the first iteration the movement for any design variable is not larger than 5% of the respective dimension of the design space. In this case IS is not implemented; since search inside the whole design space Φ is considered, it is unclear how samples of θ can be obtained to form the IS densities and separately establishing an IS density for each design choice ϕ is too computationally expensive. The larger variability of the estimates caused the gradient-based algorithm to diverge in the first couple of iterations. Thus a larger value for the sample size N = 3000 used. The required number of iterations for convergence of the algorithm in a sample run (and the total number of system simulations) was 102 (612,000). When the combined framework was used the corresponding numbers were 40 (80,000). This comparison illustrates the efficiency of the proposed two-stage optimization framework. The better starting point of the algorithm, as well as the smaller size of the search space, which allows for better normalization, that the SSO subset identification can provide, are the features that contribute to this improvement of efficiency.

6.5

Ef f icienc y of s eis mic protection s ys te m

The expected lifetime cost for the structure in each direction without the dampers is estimated as $1.1 million. The expected lifetime cost of the retrofitted system is $430,000, so the addition of the viscous dampers improves significantly the performance of the structural system. Of this amount, $267,000 corresponds to the cost for the installation of the viscous dampers and $163,00 to the present worth of the expected repair cost for damage from future earthquakes. Figure 7.8 shows the decomposition of the expected lifetime repair cost into its different components for both the initial structure and the retrofitted structure. Only minor changes occur in the distribution of the total cost over its different components. Note that the relative importance of the repair cost for acceleration-sensitive assemblies increases by the addition of the dampers, as expected, but still the importance of this cost remains small, practically negligible.

182

Structural design optimization considering uncertainties

(b) Structure with dampers

(a) Structure with no dampers

structural 33%

structural 36%

paint 29%

paint 33%

Contents 2% partitions 32% Ceiling 1%

Ceiling 1%

partitions 30%

Contents 3%

Figure 7.8 Breakdown of expected lifetime repair costs for (a) initial and (b) retrofitted structures.

7 Conclusions The robust-to-uncertainties design of engineering systems is of great importance. In this study, we discussed stochastic optimization problems that entail as objective function the expected value of a general system performance measure. We focused on problems that involve complex models and high-dimensional uncertain parameter vectors. Stochastic simulation was considered for evaluation of the system performance. This simulation-based approach allows explicit consideration of (a) nonlinearities in the models assumed for the system and its future excitation and (b) complex failure modes. The only constraint in the complexity of the system description stems from the accessible computational power, since a large number of simulations of the system response is needed. The constant advances in computer technology (hardware and software related) are continuously reducing the significance of this constraint. A two-stage framework for the associated optimization problem was discussed. The first stage implements an innovative approach, called Stochastic Subset Optimization (SSO), for efficient exploring the sensitivity of the objective function to the design variables and adaptively identifying sub-regions within the original design space that (a) have high likelihood of including the optimal design variables and (b) are characterized by small sensitivity with respect to each design variable. SSO is combined in the second stage with some other stochastic optimization algorithms for overall enhanced efficiency and accuracy of the optimization process. Simultaneous perturbation stochastic approximation was considered for this purpose in this study and suggestions for enhanced efficiency of the overall framework were given. With respect to SSO, guidelines for establishing good quality in the identification and stopping criteria for the iterative process were suggested. Topics related to the use of common random numbers for the second stage of the optimization framework were extensively discussed. Implementation of importance sampling for this stage was also considered by using the information available in the last iteration of SSO. In all discussions, special attention was given to optimization problems that involve the reliability of a system as the objective function.

Stochastic system design optimization using stochastic simulation

183

An example was presented that shows the efficiency of the proposed methodology and illustrates a systematic way to design structural systems under stochastic earthquake excitation considering all important probabilistic information. The example considered the retrofitting of a four-story non-ductile reinforced concrete office building with viscous dampers. The minimization of the expected lifetime cost was adopted as the design objective. A realistic probabilistic model was presented for representing future ground motions. An efficient and accurate methodology for estimating the damages caused by earthquake events was adopted. The structural performance was evaluated by nonlinear simulation that incorporates all important characteristics of the behavior of the structural system and all available information about the structural model and the expected future earthquakes. In this example, SSO was shown to efficiently identify a set that contains the optimal design variables and to improve the efficiency of SPSA when combined in the context of the suggested optimization framework. The center of the set identified by SSO was found to be close to the true optimal values in terms of both the design variables and the corresponding performance. Thus, use of SSO solely would lead to a sub-optimal design that is close, though, to the optimal one. For better resolution and accuracy the combined two-stage framework should be preferred.

Appendix A Two algorithms that can be used for simulating samples from π(ϕ, θ) are discussed here. Algorithm 1: Accept-reject method, which can be considered a direct Monte Carlo approach. First, choose an appropriate proposal PDF f (ϕ, θ) and then generate a sequence of independent samples as follows: (1) (2)

Randomly simulate candidate sample [ϕc , θc ] from f (ϕ, θ) and u from uniform (0,1). Accept [ϕ, θ] = [ϕc , θc ] if hs (ϕc , θc )

(3)

p(ϕc , θc ) p(ϕ, θ) > u, where M > max hs (ϕ, θ) ϕ,θ Mf (ϕc , θc ) f (ϕ, θ)

(39)

Return to (1) otherwise.

Algorithm 2: Metropolis-Hastings algorithm, which belongs to Markov Chain Monte Carlo methods (MCMC) and is expressed through the iterative form: (1) (2)

Randomly simulate a candidate sample [ϕ˜ k+1 , θ˜ k+1 ] from a proposal PDF q(ϕ˜ k+1 , θ˜ k+1 |ϕk , θk ). Compute acceptance ratio: rk+1 =

hs (ϕ˜ k+1 , θ˜ k+1 )p(ϕ˜ k+1 , θ˜ k+1 )q(ϕk , θk |ϕ˜ k+1 , θ˜ k+1 ) hs (ϕk+1 , θk+1 )p(ϕk+1 , θk+1 )q(ϕ˜ k+1 , θ˜ k+1 |ϕk , θk )

(40)

184

Structural design optimization considering uncertainties

(3)

Simulate u from uniform (0,1) and set  [ϕ˜ k+1 , θ˜ k+1 ] if rk+1 ≥ u [ϕk+1 , θk+1 ] = otherwise [ϕk , θk ]

(41)

In this case the samples are correlated (the next sample depends on the previous one) but follow the target distribution after a burn-in period, i.e. after the Markov chain reaches stationarity. The algorithm is particularly efficient when samples that follow the target distribution are already available since then no burn-in period is needed. Assume, in this setting, that there are Na samples [ϕ, θ] and a total N > Na are desired. Starting from each of the Na original samples, [N/Na ] samples are generated by the above process. Since the initial samples are distributed according to π(ϕ, θ) the Markov Chain generated in this way is always in its stationary state and all samples simulated follow the target distribution. Note that knowledge of the normalizing constant in the denominator of π(ϕ, θ) is not needed for any of the two algorithms. The efficiency of both these sampling algorithms depends on the proposal PDFs f (ϕ, θ) and q(ϕ, θ). These PDFs should be chosen to closely resemble hs (ϕ, θ)p(ϕ, θ) and still be easy to sample from. If the first feature is established then most samples are accepted and the efficiency of the algorithm is high. For Metropolis-Hastings the proposal PDFs can either be global (independent), i.e. q(·) = q(ϕ˜ k+1 , θ˜ k+1 ), or establish a local random walk, i.e. q = q(ϕ˜ k+1 , θ˜ k+1 |ϕk , θk ). In the latter case, the spread of the proposal PDFs is particularly important because it affects the size of the region covered by the Markov Chain samples. Excessively large spread may reduce the acceptance rate, increasing the number of repeated samples and thus slowing down convergence and creating correlation between samples. Small spread does not allow for efficient investigation of the whole region of the uncertain parameters and creates correlation between samples because of their proximity. If the dimension of the uncertain parameter vector is high, a typical characteristic for dynamic problems where the excitation is modeled using a white-noise sequence, the efficiency of the MCMC simulation process might be reduced (Au & Beck (2001)) because high correlation might exist between the current and the next chain state. For ROP the modified Metropolis-Hastings algorithm, discussed in detail in (Au & Beck 2003b) can be used in these cases (assuming that the loss function is described by the indicator function of failure). For general stochastic design problems, a global PDF should be chosen for parameters that individually do not significantly influence the objective function, but have significant influence only when viewed as a group. The white-noise sequence in dynamic problems typically belongs to this category.

Appendix B According to the stochastic method (Boore 2003), the total amplitude spectrum A(f ; M, r) for the acceleration time history is expressed as a product of the source, path and site contributions: A(f ; M, r) = (2πf )2 S(f ; M)

1 exp (−πko f ) exp [−πfR/(Q(f )βs )]  8 A m

R f 1+ fmax

(42)

Stochastic system design optimization using stochastic simulation

185

Here S(f ; M) is the “equivalent two point-source spectrum’’ given by (Atkinson & Silva 2000):  S(f ; M) = CMw

1−e e + 2 1 + (f /fa ) 1 + (f /fb )2

(43)

where Mw is the seismic moment (in dyn-cm) which is connected to the moment magis given nitude, M, by the relationship log10 Mw = 1.5(M + 10.7) and the constant C √ by C = RΦ VF/(4πRo ρs βs ); R
= 0.55 is the average radiation pattern, V = 1/ 2 represents the partition of total shear-wave velocity into horizontal components, F = 2 is the free surface amplification, ρs = 2.8 g/cm3 and βs = 3.5 km/s are the density and shear-wave velocity in the vicinity of the source, and Ro = 1 is a reference distance. The frequencies fa and fb in (43) are given by log10 fa = 2.181 − 0.496 M and log10 fb = 2.41 − 0.408 M, respectively, and e is a weighting parameter described by the expression log10 e = 0.605 − 0.255 M. For the rest of the parameters in (42) the term 1/R is the geometric spread factor, where R = [h2 + r2 ]1/2 is the radial distance from the earthquake source to the site, with log10 h = 0.15M − 0.05 representing a moment dependent, nominal “pseudo-depth’’. The term exp [−πfR/(Q(f )βs )] accounts for elastic attenuation through the earth’s crust with Q(f ) = 180f 0.45 a regional quality factor. The quotient factor in (42) is related to near-surface attenuation with parameters fmax = 10 and ko = 0.03. Finally Am is a near-surface amplification factor which is described through the empirical curves for generic rock sites given by (Boore & Joyner 1997). An alternative approach suggested by (Au & Beck 2003b) (instead of using the empirical curves) would be to set Am to an average constant value equal to 2. The envelope function for the earthquake excitation is represented by (Boore 2003): e(t; M, R) = a(t/tn )b exp (−c(t/tn ))

(44)

where b = −λ ln (η)/[1 + λ( ln (λ) − 1)], c = b/λ, a = [ exp (1)/λ]b and tn = 0.1R + 1/fa with λ = 0.2, η = 0.05.

References Ang, H.-S.A. & Lee, J.-C. 2001. Cost optimal design of R/C buildings. Reliability Engineering and System Safety 73:233–238. Atkinson, G.M. & Silva, W. 2000. Stochastic modeling of California ground motions. Bulletin of the Seismological Society of America 90(2):255–274. Au, S.K. 2005. Reliability-based design sensitivity by efficient simulation. Computers and Structures 83:1048–1061. Au, S.K. & Beck, J.L. 1999. A new adaptive importance sampling scheme. Structural Safety 21:135–158. Au, S.K. & Beck, J.L. 2001. Estimation of small failure probabilities in high dimensions by subset simulation. Probabilistic Engineering Mechanics 16:263–277. Au, S.K. & Beck, J.L. 2003a. Importance sampling in high dimensions. Structural Safety 25(2): 139–163. Au, S.K. & Beck, J.L. 2003b. Subset simulation and its applications to seismic risk based on dynamic analysis. Journal of Engineering Mechanics 129(8):901–917.

186

Structural design optimization considering uncertainties

Beck, J.L., Chan, E., Irfanoglu, A. & Papadimitriou, C. 1999. Multi-criteria optimal structural design under uncertainty. Earthquake Engineering and Structural Dynamics 28(7):741–761. Beck, J.L. & Katafygiotis, L.S. 1998. Updating models and their uncertainties. I: Bayesian statistical framework. Journal of Engineering Mechanics 124(4):455–461. Boore, D.M. 2003. Simulation of ground motion using the stochastic method. Pure applied Geophysics 160:635–676. Boore, D.M. & Joyner, W.B. 1997. Site amplifications for generic rock sites. Bulletin of the Seismological Society of America 87:327–341. Ching, J. & Hsieh, Y.-H. 2007. Local estimation of failure probability function and its confidence interval with maximum entropy principle. Probabilistic Engineering Mechanics 22:39–49. Enevoldsen, I. & Sørensen, J.D. 1994. Reliability-based optimization in structural engineering. Structural Safety 15(3):169–196. Gasser, M. & Schuëller, G.I. 1997. Reliability-based optimization of structural systems. Mathematical Methods of Operations Research 46:287–307. Glasserman, P. & Yao, D.D. 1992. Some guidelines and guarantees for common random numbers. Management Science 38:884–908. Goulet, C.A., Haselton, C.B., Mitrani-Reiser, J., Beck, J.L., Deierlein, G., Porter, K.A. & Stewart, J.P. 2007. Evaluation of the seismic performance of code-conforming reinforcedconcrete frame building-From seismic hazard to collapse safety and economic losses. Earthquake Engineering and Structural Dynamics 36(13):1973–1997. Iwan, W.D. & Cifuentes, A.O. 1986. A model for system identification of degrading structures. Earthquake Engineering and Structural Dynamics 14:877–890. Jaynes, E.T. 2003. Probability theory: the logic of science. Cambridge, UK: Cambridge University Press. Jensen, H.A. 2005. Structural optimization of linear dynamical systems under stochastic excitation: a moving reliability database approach. Computer Methods in Applied Mechanics and Engineering 194:1757–1778. Kleinmann, N.L., Spall, J.C. & Naiman, D.C. 1999. Simulation-based optimization with stochastic approximation using common random numbers. Management Science 45(11): 1570–1578. Kramer, S.L. 2003. Geotechnical earthquake engineering. New Jersey: Prentice Hall. Kushner, H.J. & Yin, G.G. 2003. Stochastic approximation and recursive algorithms and applications. New York: Springer. Lagaros, N.D., Papadrakakis, M. & Kokossalakis, G. 2002. Structural optimization using evolutionary algorithms. Computers and Structures 80(7–8):571–589. Papadimitriou, C., Beck, J.L. & Katafygiotis, L.S. 2001. Updating robust reliability using structural test data. Probabilistic Engineering Mechanics 16:103–113. Porter, K.A., Beck, J.L., Shaikhutdinov, R.V., Au, S.K., Mizukoshi, K., Miyamura, M., Ishida, H., Moroi, T., Tsukada, Y. & Masuda, M. 2004. Effect of seismic risk on lifetime property values. Earthquake Spectra 20:1211–1237. Pradlwater, H.J., Schuëller, G.I., Koutsourelakis, P.S. & Champris, D.C. 2007. Application of line sampling simulation method to reliability benchmark problems. Structural Safety 29(3):208–221. Robert, C.P. & Casella, G. 2004. Monte Carlo Statistical Methods. New York, NY: Springer. Royset, J.O. & Polak, E. 2004. Reliability-based optimal design using sample average approximations. Probabilistic Engineering Mechanics 19:331–343. Royset, J.O. & Polak, E. 2007. Efficient sample size in stochastic nonlinear programming. Journal of Computational and Applied Mathematics (in press). Ruszczynski, A. & Shapiro, A. 2003. Stochastic Programming. New York: Elsevier. Sørensen, J.D., Kroon, I.B. & Faber, M.H. 1994. Optimal reliability-based code calibration. Structural Safety 15:197–208.

Stochastic system design optimization using stochastic simulation

187

Spall, J.C. 2003. Introduction to stochastic search and optimization. New York: WileyInterscience. Taflanidis, A.A. & Beck, J.L. 2007a. Stochastic subset optimization for optimal reliability problems. Journal of Probabilistic Engineering Mechanics (In press). Taflanidis, A.A. & Beck, J.L. 2007b. Stochastic subset optimization for stochastic design. ECCOMAS Thematic Conference on Computational Methods in Structural Dynamics and Earthquake Engineering, Rethymno, Greece, 13–16 June.

Chapter 8

Numerical and semi-numerical methods for reliability-based design optimization Ghias Kharmanda Aleppo University, Aleppo, Syria

ABSTRACT: In the Reliability-Based Design Optimization (RBDO) model for robust system design, the mean values of uncertain system variables are usually used as design variables, and the cost is optimized subject to prescribed probabilistic constraints as defined by a nonlinear mathematical programming problem. Therefore, a RBDO solution that reduces the structural weight in uncritical regions does not only provide an improved design but also a higher level of confidence in the design. In this work, we present recent developments for the RBDO model relative to two points of view: reliability and optimization. Next, we present our recent developments for reliability-based design optimization model. Finally, we demonstrate the efficiency of our methods on different applications.

1 Introduction When Deterministic Design Optimization (DDO) methods are used, deterministic optimum designs are frequently pushed to the design constraint boundary, leaving little or no room for tolerances (or uncertainties) in design, manufacture, and operating processes. So deterministic optimum designs obtained without consideration of uncertainties could lead to unreliable designs, therefore calling for Reliability-Based Design Optimization (RBDO). It is the objective of Reliability-Based Design Optimization (RBDO) to design structures that should be both economic and reliable. However, the coupling between the mechanical modeling, the reliability analyses and the optimization methods leads to very high computational cost and weak convergence stability (Feng & Moses 1986). To overcome these difficulties, two points of view have been considered. From a reliability view point, RBDO involves the evaluation of probabilistic constraints, which can be executed in two different ways: either using the Reliability Index Approach (RIA) or the Performance Measure Approach (PMA) (see Tu et al. 1999, Youn et al. 2003). The major difficulty lies in the evaluation of the probabilistic constraints, which is prohibitively expensive and even diverges for many applications. However, from an optimization view point, we have two categories of methods: numerical and semi-numerical methods. For the first category, a hybrid method based on simultaneous solution of the reliability and the optimization problem has successfully reduced the computational time problem (Kharmanda et al. 2002). Next, an improved hybrid method has been recently proposed to improve the optimum value of the objective function more than the resulting value by the hybrid method (Mohsine et al. 2005). However, the hybrid and improved hybrid RBDO problems are more complex than

190

Structural design optimization considering uncertainties

that of deterministic design and may not lead to local optima. For the second category, an Optimum Safety Factor (OSF) method has been proposed to compute safety factors satisfying a required reliability level without demanding additional computing cost for the reliability evaluation (Kharmanda et al. 2004b). However, the OSF method cannot be used for all cases such as modal analysis. So a safest point method has been proposed to deal with these problems (Kharmanda et al. 2006). We finally note that the developments based on the reliability view point is less efficient than those based on the optimization view point because the latter provides us with reliability-based optimum designs without additional computing cost for probabilistic (reliability) constraints and leads, at least, to local optima. The numerical methods need a much higher computing time than the semi-numerical ones but to improve the optimum values, we need to use the numerical method with very expensive operations.

2 Two points of view for developing the RBDO model 2.1 Rel i a b i l i ty vie w po int The work of (Tu et al. 1999, Youn et al. 2003) depends on the development of several approaches based on a reliability view point. Here, two design requirements are coupled for each probabilistic constraint: the performance requirement is described implicitly by the performance measure, and the reliability requirement is approximated explicitly by the first- or second-order reliability index. The conventional Reliability Index Approach (RIA) for RBDO has been developed and applied to design for against fatigue crack initiation of a road arm of the M1A1 tank to successfully obtain an optimum shape design of the component, see (Youn et al. 2003). However, it was found that the computational requirement of RIA is extremely intensive because the evaluation of each probabilistic constraint during an overall RBDO iteration is quite expensive. To alleviate this computational burden, it was proposed to develop a Performance Measure Approach (PMA) for RBDO. In this approach, the reliability constraint is defined from the design perspective (rather than from the reliability analysis perspective) to measure the design constraint violation. The prescribed reliability requirement (such as six-sigma design, see Koch et al. 2004) was assumed to be satisfied, and the probabilistic performance measure (the value of the limit-state function) that satisfies this prescribed reliability requirement was used to measure the degree of the reliability constraint violation. Using PMA, an inverse reliability analysis problem was associated with the evaluation of the reliability constraint and a nonlinear ball constraint optimization problem was proposed for this inverse reliability analysis problem. The inverse reliability analysis problem in the proposed PMA was solved in a far more efficient and stable way than the conventional RIA. From a broader perspective, it was shown that the probabilistic constraints can be evaluated using either RIA or the PMA. However, there are two major advantages in PMA compared to RIA, see (Youn et al. 2004). First, it is found that the performance measure approach is inherently robust and is more effective when the reliability constraint is inactive. This fact is not surprising, since it is easier to minimize a complicated cost function subject to a simple constraint function than minimizing a simple cost function subject to a complicated constraint function. The inverse reliability analysis problem of PMA provides

N u m e r i c a l a n d s e m i-n u m e r i c a l m e t h o d s

191

this benefit. Secondly, which is more significant, the PMA always yields a solution, whereas RIA does not yield solutions for certain types of distributions. The major difficulty is associated with the reliability evaluation. So, we found that it is more efficient to select the optimality criterion as a point of view for the developments. 2.2

Optimization view point

Not surprisingly, efforts were directed towards the development of efficient techniques and general purpose programs to perform the reliability analysis. These programs and procedures compute the reliability index of a structure for a defined failure mode, but do not provide an optimum set of design parameters for improving the reliability of a structure for defined reliability information. Since the reliability index is computed iteratively, an enormous amount of computer time is involved in the whole design process. Two categories of methods have been developed. For the first category, called numerical methods, a hybrid method based on simultaneous solution of the reliability and the optimization problem has successfully reduced the computational time problem (Kharmanda et al. 2002). Next, an improved hybrid method has been recently proposed to improve the optimum value of the objective function more than the resulting value by the hybrid method (Mohsine et al. 2005). However, the hybrid and improved hybrid RBDO problems are more complex than that of deterministic design and may not lead to local optima. For the second category, called semi-numerical methods, an optimum safety factor (OSF) method has been proposed to compute safety factors satisfying a required reliability level without demanding additional computing cost for the reliability evaluation (Kharmanda et al. 2004a). However, the OSF method cannot be used for all cases such as modal analysis. So a safest point method has been proposed to deal with these problems (Kharmanda et al. 2006). In the next sections, we present our developments and some applications.

3 Numerical RBDO methods 3.1

Classical method (CM)

3.1.1

Basic formul ati on

The classical reliability-based optimization is performed by nesting the following two problems: 1.

Optimization problem: min f (x) x

subject to

gk (x) ≤ 0,

k = 1, . . . , K

(1)

β(x, u) ≥ βt where f (x) is the objective function, gk (x) ≤ 0 are the associated constraints, β(x, u) is the reliability index of the structure, and βt is the target reliability.

192

Structural design optimization considering uncertainties

x2

Failure domain

Failure domain

u2

H(x, u)0

G(x, y)0 G(x, y)
0

P* 

x1

0

Physical space

0

H(x, u)
0



u1

Normalized space

Figure 8.1 (a) Physical space or X-Space and (b) normalized space or U-Space.

Reliability analysis: the reliability index β(x, u) is equal to the minimum distance between the limit state function H(x, u) and the origin, see Figure 8.1b. This index is determined by solving the minimization problem:

2.

min d(u) = u



u2i

(2)

subject to H(x, u) ≤ 0 where d(u) is the distance in the normalized random space, defined as above, and H(x, u) is the performance function (or limit state function) in the normalized space, defined such that H(x, u) ≤ 0 implies failure, see Figure 8.1b. In the physical space, the image of H(x, u) is the limit state function G(x, y), see Figure 8.1a.

3.1.2

F u rth er d e v e lo p m e n t

According to the sub-problems (1) and (2), the classical solution consists in minimizing two Lagrangians: min L1 (x, u, λk , λβ ) = f (x) + λβ [βt − β(x, u)] + min L2 (u, λH ) = d(u) + λH H(x, u)

 k

λk gk (x)

(3a) (3b)

where λk , λβ and λH are, respectively, the Lagrangian multipliers for the constraints, the reliability index and the limit state function; (λk ≥ 0, λβ ≥ 0 and λH ≥ 0). The optimality

N u m e r i c a l a n d s e m i-n u m e r i c a l m e t h o d s

193

conditions of these two Lagrangians are, respectively,  ∂L1 ∂β ∂gk ∂f = − λβ + λk =0 k ∂xi ∂xi ∂xi ∂xi ∂L1 = βt − β(x, u) = 0 ∂λβ

(4a) (4b)

∂L1 = gk (x) = 0 ∂λk

(4c)

∂H ∂d ∂L2 = + λH =0 ∂uj ∂uj ∂uj

(5a)

∂L2 = H(x, u) = 0 ∂λH

(5b)

and

It has been demonstrated that the classical approach needs a high computational time and may lead to weak convergence stability. Furthermore, it is very difficult to implement of the machine (see Kharmanda et al. 2001, 2002). 3.2

Hybrid method (HM)

3.2.1

Basic formul ati on

The solution procedure in two separate spaces requires large computational time, especially for large-scale structures (Feng & Moses 1986). In order to improve the numerical performance, the hybrid approach consists in minimizing a new form of the objective function F(x, y) subject to a limit state as well as deterministic and reliability constraints, i.e., min F(x, y) = f (x) · dβ (x, y) x,y

subject to

G(x, y) ≤ 0 gk (x) ≤ 0,

k = 1, . . . , K

(6)

dβ (x, y) ≥ βt The minimization of the function F(x, y) is carried out in the Hybrid Design Space (HDS) of deterministic variables x and random variables y. Here, dβ (x, y) is the distance in the hybrid space between the optimum point and the design point, dβ (x, y) = d(u). Since the random variables and the deterministic ones are treated in the same space (HDS), it is very important to know the types of the used random variables (continuous and/or discrete) and the distribution law that has been used. The normalized variable u is used to evaluate the reliability index (2). However, the reliability index can be obtained in terms of the probability of failure as: β = −
−1 (Pf )

(7)

Structural design optimization considering uncertainties

Hybrid Design Space

Hybrid Design Space

db  bt db
bt P*x

P*x

ng decr easi

P*y

Failure domain

f(x)

f(x)

P*y

db bt db
bt

db

Objective function levels

decr easi

ng

db

Safe domain

Objective function levels

G(x, y)0

G(x, y)0

G(x, y)
0

Limit state decreasing G(x, y)0

x2, y2

Limit state decreasing G(x, y)0

x2, y2

G(x, y)
0

194

x1, y1

x1, y1

(a)

(b)

Hybrid Design Space G(x, y)0

Safe domain db bt db
bt p*x

decr easi

ng

db

Objective function levels

G(x, y)0

G(x, y)
0

x2, y2 Limit state decreasing

P*y

f(x)

Failure domain

x1, y1 (c)

Figure 8.2 Hybrid design spaces: (a) normal law, (b) lognormal law and (c) uniform law.

where
is the cumulative density function and Pf is the probability of failure. In many engineering applications, the evaluation of the failure probability can be carried out in several ways (Ditlevsen & Madsen 1996). 3.2.2 F u rth er d e v e lo p m e n t The hybrid Lagrangian is written as LH (x, y, λ) = f (x) · dβ (x, y) + λβ [βt − dβ (x, y)] + λG G(x, y) +

 k

λk gk (x)

(8)

N u m e r i c a l a n d s e m i-n u m e r i c a l m e t h o d s

195

The optimality conditions of this Lagrangian are ∂G  ∂gk ∂LH ∂f ∂dβ = dβ (x, y) + [f (x) − λβ ] + λG + λk =0 ∂xi ∂xi ∂xi ∂xi ∂xi

(9a)

∂G ∂LH ∂dβ = [f (x) − λβ ] + λG =0 ∂yi ∂yi ∂yi

(9b)

∂LH = βt − dβ (x, y) = 0 ∂λβ

(9c)

k

∂LH = G(x, y) = 0 ∂λG ∂LH = gk (x) = 0 ∂λk

(9d) (9e)

The optimality conditions (9) represent the optimal solution by a linear combination of different gradients of f , dβ , G and gk . At the convergence, the distance dβ stretches toward the reliability index β, which next stretches toward βt when the associated constraint is active. By comparing the conditions (9) with the optimality conditions of the classical formulation (see (4 and 5)); we can note that the only difference in the search direction lies in the coupled term: ∂G/∂xi . In fact, two cases may occur in function of the type of the optimization variables xi . Case 1: xi is a deterministic mechanical parameter (e.g. xi is a parameter of the limit state). In this case, the limit state sensitivity takes the form (Ditlevsen & Madsen 1996) ∂dβ ∂G =η ∂xi ∂xi

(10)

with the norm η 0 0 0 0 ∂Tj−1 (x, u) 0 0 ∂H 0 0 ∂G 0 0 0 η=0 0 0 ∂u 0 = 0 0 0 ∂y ∂u j j j

(11)

Case 2: xi is a probability distribution parameter of the random variable yi (e.g. xi is the mean of yi ). In this case, xi is a pure probability variable and has no effect on the limit state function, leading to: ∂G/∂xi = 0. In this case, we obtain ⎧ ⎨ ∂G ∂H = ∂yj ⎩ ∂xi 0

for

i=j

for

i = j

(12a)

where ∂dβ ∂G =η ∂yi ∂yi

(12b)

196

Structural design optimization considering uncertainties

From (10) and (12), we can see that the gradient vectors of G and dβ are co-directional. It means that there is no modification of the search direction. The introduction of this result in the first optimality condition of the hybrid Lagrangian (9a) leads to ∂LH ∂f ∂dβ  ∂gk = dβ (x, y) + [f (x) − λβ + ηλG ] + λk =0 ∂xi ∂xi ∂xi ∂xi

(13)

k

The comparison of the optimality conditions for classical and hybrid approaches gives the relationships between the Lagrangian multipliers in the two formulations: λβ =

λβ − f (x) − ηλG dβ (x, y)

(14)

λH =

λG f (x) − λβ

(15)

and

These developments show that the solution of problem (8) respects exactly the optimality conditions of the initial problem, given by (4a) and (5b), where the two phenomena were separated. Otherwise, the hybrid Lagrangian definition does not introduce any modification in the optimality conditions. In the literature, the hybrid method has been successfully applied for several examples (Kharmanda et al. 2001–2003). An industrial application of a lorry brake system design (for KNORR-BREMSE Company) has been successfully carried out during the PhD thesis of (Mohsine 2006). 3.3

Im prov ed hy b r id me t ho d (I H M)

3.3.1 Ba si c f ormu la t io n Using the hybrid method, we can obtain local optima and designer may then select the best optimum. In the improved hybrid method, we introduce the design point and the optimum solution in the objective function and the constraints at the design point and at the optimum solution as follows: min x,y

F(x, y) = f (x) · dβ (x, y) · f (my )

subject to G(x, y) = 0 gk (x) ≤ 0 gj (my ) ≤ 0 and dβ (x, y) ≥ βt

(16)

The random vector y has mean values my and standard-deviations σy · f (my ) is the optimal objective function and gj (my ) is the constraint at which we can control the optimal configuration. The solution of this problem depends on two important points. It can be carried out simultaneously in the hybrid design space (HDS).

N u m e r i c a l a n d s e m i-n u m e r i c a l m e t h o d s

3.3.2

197

F u rt h er dev el opment

We show the equivalence of the improved method and the classical (initial) one. The improved hybrid Lagrangian is written as LI (x, y, λ) = f (x) · dβ (x, y) · f (my ) + λβ βt − dβ (x, y)   + λG G(x, y) + λk gk (x) + λj gj (my ) k

(17)

j

In order to write the optimality conditions of the improved hybrid Lagrangian, we note that the derivatives of f (my ) and of g(my ) with respect to y are nil: ! " ∂f my   ∂y 

=0

(18)

 ∂g(my )  =0 ∂y y∗

(19)

y∗

and

Because the my value coincide the optimal solution of the objective function and we derivate with respect to random distribution for which we propose a function Q that we can write my = Q(y). So it gives:  ∂f ◦ Q  (y) = 0 ∂y y∗

(20)

So the optimality conditions of the improved hybrid Lagrangian are: ∂LI ∂f ∂dβ = dβ (x, y) · f (my ) + [f (x) · f (my ) − λβ ] ∂xi ∂xi ∂xi ∂G  ∂gk + λG + λk =0 ∂xi ∂xi

(21a)

k

∂G ∂LI ∂dβ = [f (x) · f (my ) − λβ ] + λG =0 ∂yi ∂yi ∂yi

(21b)

∂LI = βt − dβ (x, y) = 0 ∂λβ

(21c)

∂LI = G(x, y) = 0 ∂λG ∂LI = gk (x) = 0 ∂λk ∂LI = gj (my ) = 0 ∂λj

(21d) (21e) (21f)

198

Structural design optimization considering uncertainties

The optimality conditions (21) represent the optimal solution by a linear combination of different gradients of f , dβ , G and gk . At the convergence, the distance dβ stretches toward the reliability index β, which next stretches toward βt when the associated constraint is active. By comparing the conditions (21) with the optimality conditions of the classical formulation (see (4 and 5)); we can note that the only difference in the search direction lies in the coupled term: ∂G/∂xi . In fact, two cases may occur in function of the type of the optimization variables xi . Case 1: xi is a deterministic mechanical parameter (e.g. xi is a parameter of the limit state). In this case, the limit state sensitivity takes the form (Ditlevsen & Madsen 1996) ∂dβ ∂G =η ∂xi ∂xi

(22)

with the norm η 0 0 0 0 ∂Tj−1 (x, u) 0 0 ∂H 0 0 ∂G 0 0 0 η=0 0 0 ∂u 0 = 0 0 0 ∂yj ∂uj j

(23)

Case 2: xi is a probability distribution parameter of the random variable yi (e.g. xi is the mean of yi ). In this case, xi is a pure probability variable and has no effect on the limit state function, leading to: ∂G/∂xi = 0. In this case, we obtain ⎧ ⎨ ∂G ∂H = ∂yj ⎩ ∂xi 0

for

i=j

for

i = j

(24)

where ∂dβ ∂G =η ∂yi ∂yi

(25)

From (22) and (24), we can see that the gradient vectors of G and dβ are co-directional. It means that there is no modification of the search direction. The introduction of this result in the first optimality condition of the improved hybrid Lagrangian (21a) leads to ∂f ∂dβ  ∂gk ∂LI = dβ (x, y) · f (my ) + [f (x · f (my ) − λβ + ηλG ] + λk =0 ∂xi ∂xi ∂xi ∂xi

(26)

k

The comparison of the optimality conditions for classical and improved hybrid approaches gives the relationships between the Lagrangian multipliers in the two formulations: λβ =

λβ − f (x) · f (my ) − ηλG dβ (x, y) · f (my )

(27)

N u m e r i c a l a n d s e m i-n u m e r i c a l m e t h o d s

199

and λG =

λG f (x) · f (my ) − λβ

(28)

These developments show that the solution of problem (17) respects exactly the optimality conditions of the initial problem, given by (4) and (5), where the two phenomena were separated. Otherwise, the improved hybrid Lagrangian definition does not introduce any modification in the optimality conditions. Applications of transient analysis have demonstrated that the main benefit of the improved hybrid method. Here, we improve the structure performance by much more minimizing the objective function than the hybrid method (Mohsine et al. 2005). To conclude this section, we can compare between the three kinds of numerical methods: classical, hybrid and improved hybrid RBDO methods. The classical method leads to very high computational cost and weak convergence. The hybrid method has successfully reduced the computing time relative to the classical one (Kharmanda et al. 2001–2003). The improved hybrid method can improve the optimum value of the objective function more than the hybrid method (Mohsine et al. 2005). In these presented numerical method, the reliability-based design optimization problem has two kinds of variables: random and deterministic that still leads to expensive procedures and may not yield local optima. In the next section, we present two semi-analytical methods that reduce the optimization problem scale.

4 Semi-numerical RBDO methods 4.1 4.1.1

Optimum s afety factor method (O SF) Basic formul ati on

It is our aim that the safety factors should be independent of the engineering experience. In fact the engineering experience is based on experimental work, design knowledge, etc. However, when designing a new type of structure, we usually need some experimental background for proposing suitable safety factors. When applying safety factors the initial cost will increase, and this increase should not be too large. Given that sensitivity analysis plays a very important role and can provide us with the influence of the parameters on the structure studied, we will use this concept in the proper direction and combine it with the reliability analysis. The main disadvantage of the Deterministic Design Optimization (DDO) procedure is that it may not satisfy an appropriate required reliability level. Although we improve the reliability level of the structure when using the hybrid RBDO, this approach leads to a saving of computational time (which may be then available for the reliability analysis). Thus, our OSF approach consists in using both sensitivity analysis and reliability analysis to overcome the disadvantages of DDO and RBDO by numerical methods (Kharmanda & Olhoff 2003, Kharmanda et al. 2003, 2004c, Kharmanda & Olhoff 2007). Table 8.1 shows the different formulations of optimum safety factors for normal, lognormal and uniform distributions.

200

Structural design optimization considering uncertainties Table 8.1 Optimum safety factors for normal, lognormal and uniform distributions. Law

OSF

Normal

Sfi = 1 + γi ·u∗i

1 2 1 S fi =  exp ln(1 + γi2 ) · u∗i 2 1 + γi √ Sfi = 1 − 3γi (1 − 2
(u∗i ))

Lognormal Uniform

u2

u2

Infeasible domain

f

b

PD

P*

u1

bt H(u)
0

a u1

H(u)
0

POp Feasible domain

(b) (a)

Figure 8.3 (a) Design point modeling and (b) Optimum solution modeling.

4.1.2 F u rth er d e v e lo p m e n t Let us consider an example of only two normalized variables u1 and u2 (see Figure 8.3a). For an assumed failure scenario H(u) ≤ 0, the design point P∗ is calculated by min d 2 = u21 + u22 u

subject to H(u1 , u2 ) ≤ 0

(29)

The Lagrangian function for the problem (29) can be written as: L(u, λ, s) = d 2 (u) + λ · [H(u) + s2 ]

(30)

where the inequality constraint in (29) is adjoined by the Lagrange multiplier λ, after having converted the inequality constraint into equality H(u) + s2 = 0, by introducing

N u m e r i c a l a n d s e m i-n u m e r i c a l m e t h o d s

201

the real slack variable s. The optimality conditions for this Lagrangian are: ∂L ∂d 2 ∂H = +λ = 0, ∂ui ∂ui ∂ui ∂L = H(u) + s2 = 0 ∂λ ∂L = 2sλ = 0 ∂s

i = 1, 2

(31a) (31b) (31c)

The optimality condition for L with respect to s, yields the so-called switching condition sλ = 0, and the necessary condition ∂2 L/∂s2 ≥ 0 for a minimum of L implies that the Lagrangian multiplier λ must be non-negative, i.e., λ ≥ 0. So due to the condition (31c), we distinguish between two cases: Case 1: If the real slack variable is non-zero (s = 0), the Lagrangian multiplier has to be zero (λ = 0) and the limit state constraint must be less than zero (H(u) < 0), which correspond to the case of failure. Case 2: If the real slack variable is zero (s = 0), the Lagrangian multiplier is nonnegative (λ ≥ 0) and the limit state is defined by the equality constraint H(u) = 0. The solution here is found on the limit state function and represents the Design Point. The first case is not suitable to our reliability-based study whereas the second one is basic for our approach. Since we have only two normalized variables u1 and u2 , equation (31a) can be written as: ∂d 2 ∂H ∂L = +λ =0 ∂u1 ∂u1 ∂u1

(32a)

∂L ∂d 2 ∂H = +λ =0 ∂u2 ∂u2 ∂u2

(32b)

Using the square distance d 2 in equation (29), we get: ∂H λ ∂H = 0 ⇔ u1 = − ∂u1 2 ∂u1 ∂H λ ∂H = 0 ⇔ u2 = − 2u2 + λ ∂u2 2 ∂u2

2u1 + λ

(33a) (33b)

From Figure 8.3, at the design point P∗ , the tangent of α is given by: tan α = u2 /u1 and using equations (33a) and (33b), we get: ∂H u2 ∂u2 = tan α = ∂H u1 ∂u1

(34)

Equation (34) shows the relationship between the distribution of the normalized vector components and the sensitivity of the limit state function. Problem (29) gives us

202

Structural design optimization considering uncertainties

the reliability index β as the minimum distance between the limit state function and the origin (Hasofer & Lind 1974). This way the resulting reliability index may be lower or higher than the target reliability index βt . As we wish to satisfy a required target reliability level, we now write: βt2 = u21 + u22

(35)

Using equations (34) and (35), we get:

 ∂H 2 ∂u1 2 βt2 = u22  + u2 ∂H 2 ∂u2 or

⎞  ∂H 2 ⎟ ⎜ ∂u ⎟ ⎜ 1 2 u22 ⎜ 2 + 1⎟ = βt ⎠ ⎝ ∂H

(36)

⎛

(37)

∂u2 Here, the value of normalized vector components principally depends on the percentage of the limit state gradient. So u2 is written as: +

 , , ∂H 2 , , ∂u2 (38) u2 = βt , 2  , - ∂H ∂H 2 + ∂u1 ∂u2 In general, when considering the normal distribution law, the normalized variable ui is given by: ui =

yi − m i , σi

i = 1, . . . , n

(39)

The standard deviation σi can be related to the mean value mi by: σi = γi · mi ,

i = 1, . . . , n

(40)

This way we introduce the safety factors Sfi corresponding to the design variables xi . The design point can be expressed by: yi = Sfi · mi ,

i = 1, . . . , n

(41)

By (40) and (41), we replace σi and xi in equation (39) and get: ui =

Sfi − 1 , γi

i = 1, . . . , n

(42)

N u m e r i c a l a n d s e m i-n u m e r i c a l m e t h o d s

203

Using equation (42), we can write (38) in the following form: +

 , , ∂H 2 , , Sf2 − 1 ∂u2 = βt , 2  , γ2 - ∂H ∂H 2 + ∂u1 ∂u2

(43)

or in terms of Sf2 : +

 , , ∂H 2 , , ∂u2 Sf2 = 1 + γ2 · βt ,  

, - ∂H 2 ∂H 2 + ∂u1 ∂u2

(44)

The calculation of the normalized gradient ∂H/∂u is not directly accessible because the mechanical analysis is carried out in the physical space, not in the standard space. The computation of the normalized gradient is carried out by applying the chain rule on the physical gradient ∂G/∂x: ∂H ∂G ∂Tk−1 (u, y) = , ∂ui ∂yk ∂ui

i = 1, . . . , n, k = 1, . . . , K

(45)

where T −1 (y, u) is the probabilistic transformation function. After some algebra, the normalized gradient can be written as: ∂H = ∂ui

   ∂G     ∂y , i

i = 1, . . . , n

(46)

The distribution of the components of the vector u can be measured by the sensitivity analysis of the limit state function with respect to the design point vector y. +   ,  ∂G  ,   ,  ∂y  , 2    Sf2 = 1 ± γ2 · βt ,  -  ∂G   ∂G   +   ∂y   ∂y  1

(47)

2

For a single limit state problem of n design variables, and sum from j = 1 to n, equation (47) can thus be written in the following form: +   ,  ,  ∂G  ,  , ∂yi   , Sfi = 1 ± γi · βt , , n -  ∂G    j=1 ∂yj

i = 1, . . . , n

(48)

204

Structural design optimization considering uncertainties

with the optimum values u∗i of the normalized vector: +   ,  ,  ∂G  ,  , ∂yi   , u∗i = ±βt , , n -  ∂G    j=1 ∂yj

i = 1, . . . , n

Here, the sign ± depends on the sign of the derivative, i.e. ∂G > 0 ⇔ Sfi > 1, ∂yi

i = 1, . . . , n

(49)

∂G < 0 ⇔ Sfi < 1, ∂yi

i = 1, . . . , n

(50)

Using these safety factors, we can satisfy the required reliability level and avoid the complexity of the problem. In the literature, the OSF method has been successfully applied for several static examples (Kharmanda et al. 2003–2004c). For the transient analysis, (Yang et al. 2005) from Ford Motor Company compared the results and efficiencies of different RBDO methods on an exhaust system. The objective was to minimize the weight of the system subject to constraints that the reliability of the resultant forces in each frequency region should be less than specified values. All in all 144 constraints were imposed, but many of them were inactive. (Yang et al. 2005) tested several RBDO methods. They concluded that: ‘(Kharmanda et al. 2004c) also used structural safety factors, based on the sensitivity of the limit-state function, for RBDO. In addition to its simplified computational framework to completely decouple the optimization and the reliability analyses, the method has two advantages: 1.

2.

It incorporates the partial safety-factor concept with which most designers are familiar. And, theoretically, safety factors do not have to be tied to the individual random variables and thus the MPPs (Most Probable Points). It produces progressively improved reliable designs in the initial steps that help designers keep track of their designs.’

According to the experience of Ford Motor Company, our method is considered as a very good active constraint strategy (for problems with many constraints). For modal analysis, it has been applied for a special case (Kharmanda et al. 2004d), where the reliability-based optimum solution was determined subject to a prescribed eigen-frequency fn . But if the failure interval [fa , fb ] is given, it is also very difficult to determine the safest solution using the OSF method. So we have to develop an efficient method to find the best point correspond to the eigen-frequency for a given frequency interval.

N u m e r i c a l a n d s e m i-n u m e r i c a l m e t h o d s

205

Table 8.2 Mean values by safest point method for normal, lognormal and uniform distributions. Law

Mean values

Normal

mi =

yai + ybi , i = 1, … , n 2

a b  ln(yi yi ) , i = 1, … , n mi = 1 + γi2 exp 2 ya + ybi , i = 1, … , n mi = i 2

Lognormal Uniform



Failure domain Safety domain

ba fa

bb fn

f(Hz) fb

Figure 8.4 The safest point at frequency fn .

4.2

Safest point method (SP)

4.2.1

Basic formul ati on

The safest structure under free vibrations for a given interval of eigen-frequency is found at the safest position of this interval where the safest point has the same reliability index relative to both sides of the interval. The use of the hybrid method here needs a multiple procedures and high computing time (Kharmanda et al. 2007). Thus, we present efficient formulations of the safest point method for normal, lognormal and uniform distributions (see Table 8.2). 4.2.2

F u rt h er dev el opment

Let consider a given interval [fa , fb ]. For the first shape mode, to get the reliabilitybased optimum solution for a given interval, we consider the equality of the reliability indices: βa = βb

(51a)

206

Structural design optimization considering uncertainties

with + , n , βa = (uai )2

and

+ , n , βa = (ubi )2

i=1

i = 1, . . . , n

(51b)

i=1

To verify the equality (51a), we propose the equality of each term. So we have: uai = −ubi ,

i = 1, . . . , n

(52)

According to the normal distribution law, the normalized variable ui is given by (39) and (52), we get: y b − mi yia − mi =− i , σi σi

i = 1, . . . , n

(53)

To obtain equality between the reliability indices (see equation 51a), the mean value of variable corresponds to the structure at fn . So the mean values of safest solution are located in the middle of the variable interval [yia , yib ] as follows: xi = mi =

yia + yib , 2

i = 1, . . . , n

(54)

In a recent publication (Kharmanda et al. 2006, 2007), we found that the safest point method is suitable for modal analysis more than the other methods that are complex to implement and to converge in this kind of study. To conclude this section, the OSF is simple to implement, can satisfy required reliability levels, has only a single type of variable y, and only needs a single simple optimization process to determine the design point. This method can be successfully used for static and transient analysis problems but for a modal analysis where the aim is to optimize a structure for a given eigen-frequencies interval, the safest point method is very suitable to find the best structure. Finally, to compare between the numerical and semi-numerical methods, we can note that the computational time when using the numerical methods is very high relative to the semi-numerical methods because we deal with two kinds of variables for numerical methods but with only one kind for semi-numerical methods. As result, the numerical methods can solve both optimization and reliability problems by numerical procedures but the semi-numerical methods solve the reliability problem by analytical form and the optimization problem by numerical procedure that leads to a reduction of the computing time. In the next section, we present some numerical examples to compare for different methods with object of showing their advantages depending on the studied cases.

5 Numerical applications We study three examples: static, modal and transient cases in order to provide the designer with the suitable method for each case.

N u m e r i c a l a n d s e m i-n u m e r i c a l m e t h o d s

207

H3

H2

H1 L (a)

(b)

Figure 8.5 Layout of tri-material cantilever beam.

5.1

Static analys is: A cantilever tri-mat e ri al be am

The objectives of the following static analysis are to demonstrate: 1. 2. 3.

DDO procedure cannot satisfy the required reliability level, Semi-numerical methods such as OSF are simple to implement relative to numerical ones such as hybrid method, Semi-numerical methods such as OSF reduce efficiently the problem scale relative to numerical ones such as hybrid method that leads to a reduction of the computing time.

The design problem under consideration pertains to a short tri-material cantilever beam of length L = 100 mm, height H = 50 mm and width T = 20 mm, which is loaded by a distributed pressure q = 15 N/mm2 . The beam structure is composed of three layers of material (Figure 8.5) of different Young’s moduli E1 = 200 GPa, E2 = 100 GPa and E3 = 150 GPa, Poisson’s ratios ν1 = 0.3, ν2 = 0.1 and ν3 = 0.2, and yield stresses y y y σ1 = 48 MPa, σ2 = 18 MPa and σ3 = 42 MPa. The heights of the three layers are: H1 = 10 mm, H2 = 30 mm, and H3 = 10 mm. To optimize the tri-material beam structure, the mean values mH1 , mH2 and mH3 of the heights H1, H2 and H3 are the design variables. The physical heights H1, H2 and H3 are elements of the vector of random variables. The target reliability index is taken to be: βt = 3, and the standard-deviations are given by σH1 = 0.1mH1 , σH2 = 0.1mH2 and σH3 = 0.1mH3 . During the subsequent design optimization processes, we consider all variables to be bounded by upper and lower limits. 5.1.1

O ptimizat ion procedures

DDO procedure: The objective is to minimize the volume subject to the design constraints and consider a safety factor Sf that is applied to the stress and based on engineering experience. The structure has to be designed by considering the maximum y allowable values σjw = σj /Sf , j = 1, 2, 3 for the von Mises stresses σjmax , j = 1, 2, 3 in the

208

Structural design optimization considering uncertainties

most critical points in each of the three layers of different material. Thus, the structural optimization problem with the safety factor taken into account, can be written as min

H1,H2,H3

Volume(H1, H2, H3) y

subject to σ1max (H1, H2, H3) ≤ σ1w = σ1 /Sf y

σ2max (H1, H2, H3) ≤ σ2w = σ2 /Sf

(55)

y

σ3max (H1, H2, H3) ≤ σ3w = σ3 /Sf The associated reliability evaluation without consideration of the safety factor can be written in the form min

uH1 ,uH2 ,uH3

d(uH1 , uH2 , uH3 )

subject to

y

σ1 − σ1max (H1, H2, H3; uH1 , uH2 , uH3 ) ≤ 0 y

σ2 − σ2max (H1, H2, H3; uH1 , uH2 , uH3 ) ≤ 0

(56)

y

σ3 − σ3max (H1, H2, H3; uH1 , uH2 , uH3 ) ≤ 0 Here, we take the value of the global safety factor applied to the yield stresses to be Sf = 1.5. This way the allowable stresses will be: σ1w = 32, σ2w = 12 and σ3w = 28 MPa. After having optimized the structure according to (55), the resulting volume is found to be VDDO = 43 252 mm3 . The reliability index depends on the distribution law, and optimum values of the reliability index for the three different types of distribution are found to be: βDDO = 3.5127. Using DDO we cannot control a required reliability level. However, by integrating the reliability concept into the design optimization process (thereby performing RBDO), we can satisfy the reliability constraint. Hybrid procedure: The classical method implies very high computational cost and exhibits weak convergence stability. So we use the hybrid method to satisfy the required reliability level (within admissible tolerances of 1%). In the hybrid procedure of RBDO, we minimize the product of the volume and the reliability index subject to the limit state functions and the required reliability level. The hybrid RBDO problem is written as min

mH1 ,mH2 ,mH3 ,H1,H2,H3

subject to

Volume(H1, H2, H3) · dβ (mH1 , mH2 , mH3 , H1, H2, H3) y

σ1max (H1, H2, H3) ≤ σ1

y

σ2max (H1, H2, H3) ≤ σ2

(57)

y

σ3max (H1, H2, H3) ≤ σ3

dβ (mH1 , mH2 , mH3 , H1, H2, H3) ≥ βt This optimization process is carried out in a hybrid design space. The resulting optimal values of the reliability index are found to be: dβ = 3.0001 ≈ βt (i.e., 0.03% higher than the target reliability index). The resulting optimum volumes are determined as: Volhybrid = 41 782 mm3 . The experience of the designer on finite element software

N u m e r i c a l a n d s e m i-n u m e r i c a l m e t h o d s

209

Table 8.3 Safety factor values. β=3

Variables

∂σ1 /∂yi

∂σ2 /∂yi

∂σ3 /∂yi

Sf

H1 H2 H3

−1.0520 −0.7452 −0.8432

−0.2160 −0.2041 −0.6796

−0.7318 −0.6119 −0.8271

0.8255 0.8458 0.8108

plays a very important role in improving the objective function and controlling the convergence. Although the method yields results that satisfy the required reliability level within admissible tolerances, the problem is a complex optimization problem and needs a large number of iterations to converge and improve the value of objective function. OSF procedure: This method includes three main steps: 1.

The first step is to obtain the design point (the Most Probable Point). Here, we minimize the volume subject to the design constraints without consideration of the safety factors. This way the optimization problem is simply written as: min

H1,H2,H3

Volume(H1, H2, H3)

subject to

y

σ1 (H1, H2, H3) ≤ σ1

y

σ2 (H1, H2, H3) ≤ σ2

(58)

y

σ3 (H1, H2, H3) ≤ σ3

2.

3.

The design point is found to correspond to the maximum von Mises stresses σ1max = 47.335 MPa, σ2max = 17.177 MPa and σ3max = 41.999 MPa, that are almost y y y equivalent to the yield stresses σ1 , σ2 and σ3 . The second step is to compute the optimum safety factors for normal distribution. In this example, the number of the deterministic variables is equal to that of the random ones. During the optimization process, we obtain the sensitivity values of the limit state with respect to all variables. So there is no need for additional computational cost. Table 8.3 shows the results leading to the values of the safety factors, namely the sensitivity results for the different limit state functions. The third step is to calculate the optimum solution. This encompasses inclusion of the values of the safety factors in the values of the design variables in order evaluate the optimum solution.

5.1.2 D iscussion Table 8.4 presents the different results of the DDO and RBDO procedures. Both RBDO procedures can satisfy the required reliability level βt = 3 but the DDO cannot. The DDO may lead to high or low reliability levels because it does not control the reliability. In order to demonstrate the efficiency of the OSF (semi-numerical) method relative to the hybrid (numerical) procedure, we discuss below the results obtained by these procedures. The resulting design obtained by the OSF method is the best solution relative to the design obtained by hybrid RBDO procedure as the objective is to provide the

210

Structural design optimization considering uncertainties Table 8.4 Results for DDO and RBDO procedures. Variables

mH1 mH2 mH3 σ1max σ2max σ3max H1 H2 H3 y σ1 y σ2 y σ3 β Volume

DDO

8.6285 25.232 9.3910 31.152 11.026 27.999 7.5034 18.961 7.4072 47.009 17.007 41.598 3.5127 43 252

RBDO Hybrid method

OSF method

8.3992 24.753 8.6298 33.096 12.059 29.718 7.4942 18.726 7.4368 47.488

17.075 41.997 3.0001 41 782

9.3974 21.851 9.1176 34.347 12.134 30.915 7.7576 18.482 7.3930 47.335 17.177 41.999 3.0000 40 366

best compromise between cost and safety. The OSF methodology satisfies the required reliability level βt = 3 and gives a smaller structural volume than the hybrid method for the reliability level. In order to improve the resulting structure by the hybrid method, the designer can obtain several local optima and then select the best solution. The resulting optimum volume obtained by OSF (VOSF = 40 366 mm3 ) is smaller than the resulting volume determined by the hybrid method by 3.39%. In general the DDO is simple to implement but it has two kinds of optimization variables x and u and also needs two optimization procedures: the first determines the optimal solution using safety factor, and the second yields the value of the reliability index. Note that DDO cannot perform design subject to a required reliability level. The hybrid method as a numerical method, can generally satisfy the required reliability level but it has two types of optimization variables x and y and needs also to solve a single, complex optimization problem. This means that the designer needs more iteration to get several local optima in order to improve the objective function, and the hybrid method is complex to implement exactly. The OSF as a semi-numerical method, is simple to implement, can satisfy required reliability levels, has only a single type of variable y, and only needs a single, simple optimization process to determine the design point. It is demonstrated that the OSF method possesses several advantages: a smaller number of optimization variables, good convergence stability, lower computing time, and satisfaction of required reliability levels (see also Kharmanda et al. 2002, 2004c). 5.2

Mod al analy s is: An air c r aft wing

The objectives of the following modal analysis are to demonstrate that the safest point method is the most suitable to use for the modal cases because of its simple

N u m e r i c a l a n d s e m i-n u m e r i c a l m e t h o d s

A

(a)

B

C

211

D

(b)

Figure 8.6 Aircraft wing.

implementation and its computing time reduction relative to the other methods. The wing is uniform along its length with cross sectional area as illustrated in Figure 8.6a. It is firmly attached to the body of the airplane at one end. The chord of the airfoil has dimensions and orientation as shown in Figure 8.6b. The wing is made of low density polyethylene with a Young’s modulus of 38e3 psi, Poisson’s ration of 0.3, and a density of 8.3E-5 lbf-sec2/in4. Assume the side of the wing connected to the plane is completely fixed in all degrees of freedom. The wing is solid and material properties are constant and isotropic. Here, we can consider thee structures: The first structure must be optimized subject to the first frequency value of the given fa , the second one must be optimized at the end frequency value of the interval fb , and the third structure must be optimized subject to a frequency value fn that verifies the equality of reliability indices relative to both sides of the given interval (see Figure 8.5). Let consider the interval [16,18] Hz and a given interval to design the beam structure. This way, we consider that the frequency values as follows: fa = 16 Hz, fb = 18 Hz and fn = ? Hz, where fn must verify the equality of reliability indices:βa = βb . Table 8.5 shows that the safest point method provides the solution with a good computational time relative to the HM. 5.3 Transient analys is: A triangular plate The objective of the following transient analysis is to demonstrate that the improved hybrid method can provide the designer with a better optimum value than the hybrid method. A triangular plate structure being illustrated in Figure 8.7 is submitted to pressure 200 Mpa. The Young’s modulus is: 207 GPa and Poisson’s ratio is: 0.3. The thickness of this plate is: R0 = 10 mm and T1 = 30 mm. The radius of fillet is: FIL = 10 rad. The yield stress is: σy = 235 Mpa. The optimization problem is to find the optimum value of the structural volume subject the maximum stress (transient response). This hybrid RBDO problem can be expressed as: min x,y

Volume(x) · dβ (x, y)

subject to σmax (y) − σy = 0 σk (y) − σy ≤ 0 and

dβ (x, y) ≥ βt

(59a)

212

Structural design optimization considering uncertainties Table 8.5 Results for Hybrid and SP procedures when βa = βb . Parameter

Fn

A B C D A1 B1 C1 D1 A2 B2 C2 D2 Fn Fa Fb Volume Time(S)

Fa

Fb

Initial

0.13295 0.24112 0.30834 0.26316 0.11295 0.20112 0.24834 0.18316 0.15295 0.28112 0.36834 0.34316 17.9030 14.3580 21.8460 6.18645 —

Optimum solutions Hybrid method

SP method

0.13391 0.20138 0.29656 0.20562 0.12331 0.24105 0.28214 0.26306 0.14441 0.24120 0.31071 0.26330 17.1080 16.0990 17.9530 5.55177 25

0.12300 0.22838 0.29963 0.22668 0.11301 0.21578 0.27162 0.23855 0.13320 0.24121 0.30939 0.26406 16.9790 16.0000 17.9510 5.83910 151

and the improved hybrid RBDO problem can be presented as follows: min x,y

Volume(x) · dβ (x, y) · Volume(my )

subject to σmax (y) − σy = 0 σk (y) − σy ≤ 0 and

(59b)

dβ (x, y) ≥ βt

Here, we can regroup T1, R0 and FIL in a random vector y but to optimize the design, the means mT1 , mR0 and mFIL are regrouped in a deterministic vector x, and their fix standard-deviation equals to 0.1 mx . Here, the normalized variable ui is given by: ui =

yi − m i , σi

i = 1, . . . , n

(60)

where the mean mi and the standard deviation σi are two parameters of the distribution, usually estimated from the available data. Table 8.6 shows the hybrid and improved hybrid results. The improved and the hybrid RBDO satisfy the required reliability level βt . However, the optimal volume obtained by the improved hybrid method is less than that obtained by the hybrid method. This way the volume value reduction is almost 26% that leads to economic structures.

N u m e r i c a l a n d s e m i-n u m e r i c a l m e t h o d s

T1

30°

213

30°

Dimension: mm

R
Fil z x

30° 30°

40

INRAD

200

30° 30°

Figure 8.7 Geometry and Boundary conditions of triangular plate structure.

Table 8.6 Results for RBDO procedures by HM and IHM. Parameter

HM

IHM

T1 FIL R0 σy mT1 mFIL mR0 σw Volume β

24.985 8.5833 7.3251 234.92 29.678 10.600 7.6991 204.51 105 874 3.8096

24.058 9.1013 9.8216 235.04 26.092 9.1062 6.0869 216.42 78 250 3.8

In this example, we demonstrate that the improved hybrid method can improve the structural performance relative to the hybrid method but it needs a more complex model (complex implementation) than the hybrid method.

6 Conclusions For the static analysis, we first demonstrate that the DDO procedure may lead to low or high reliability levels because it necessitates a proposition of a global safety factor depending on the engineering experience (cannot control the reliability levels). However, all methods of RBDO (Reliability-Based Design Optimization) respect the required reliability level. Comparing the RBDO methods, it has been demonstrated that the classical approach needs a high computing time relative to the hybrid method and has weak convergence stability (see Kharmanda et al. 2001, 2002). The improved

214

Structural design optimization considering uncertainties

Table 8.7 Advantages and disadvantages of DDO and RBDO procedures. Models

Advantages

Disadvantages

DDO

– Simple to implement

– – – –

No satisfaction of reliability requirements Two optimization processes Two types of optimization variables x, u May lead to local optima

– – – – – – – – – – – – – – –

Weak convergence stability High computing time Very complicated to implement Two types of optimization variables x, u May lead to local optima Single, complex optimization process Two types of optimization variables x, y Complicated to implement May lead to local optima Improvement of the objective function Single, complex optimization process Two types of optimization variables x, y Very complicated to implement More iteration to improve the objective May lead to local optima

RBDO 1. Numerical methods CM – Satisfaction of reliability requirements

HM

– Satisfaction of reliability requirements

IHM

– Satisfaction of reliability requirements

2. Semi-numerical methods OSF – Simple to implement – Satisfaction of reliability requirements – Single, simple optimization process – Reduction of computing time – Single type of variables y SP – Simple to implement – Satisfaction of reliability requirements – Double simple optimization processes – Reduction of computing time – Single type of variables y

– Leads, at least, to local optima

– Used only for modal analysis

hybrid method needs a complex model to improve the optimum value of the objective function relative to the hybrid method. The hybrid method has a good convergence stability that makes it suitable for RBDO problems as a numerical method. However, the hybrid RBDO problem is more complex than that of deterministic design and may not lead to local optima. To overcome both drawbacks, an Optimum Safety Factor (OSF) method has been proposed to provide us with reliability-based optimum designs without additional computing cost for probabilistic (reliability) constraints and leads, at least, to local optima (Kharmanda et al. 2004c). As result, the OSF being a semi-numerical method is very efficient for the RBDO problems in static cases because of its simple implementation and the reduction of number of optimization variables. If the designer needs to improve the objective function, the hybrid and improved hybrid methods are suitable by testing several starting points and next selecting the best solution. The improved hybrid method provides us with a better

N u m e r i c a l a n d s e m i-n u m e r i c a l m e t h o d s

215

solution than the hybrid method but needs a complex implementation (Mohsine et al. 2005, Kharmanda & Olhoff 2007). For modal analysis, the hybrid method has been applied for a special case of a structure performing free vibrations (Kharmanda et al. 2003), where the reliabilitybased optimum solution was determined subject to a prescribed eigen-frequency fn . The optimum safety factor method has been also applied for a special case of a structure performing free vibrations (Kharmanda et al. 2004a), where the reliability-based optimum solution was determined subject to a prescribed eigen-frequency fn . But if the failure interval [fa , fb ] is given, we cannot determine the reliability-based optimum solution using optimum safety factor method and the hybrid necessitates a complex procedure to optimize three structures simultaneously to get the equality between reliability indices. The semi-numerical method called Safest Point (SP) method is very suitable for the modal cases because of its simple implementation and small computing time (Kharmanda et al. 2006, 2007). For transient analysis, the hybrid and improved hybrid methods (numerical methods) and the OSF method (semi-numerical method) are suitable to be used. When saving the computational time or/and needing simple implementation, the OSF method is the best approach to be used. However, for getting several solutions and improving the optimum value of the objective function, we use the hybrid and the improved hybrid methods (Mohsine et al. 2006). The improved hybrid method provides the designer with a local optimum better than the hybrid one but the hybrid method is simpler to implement than the improved hybrid method. As a general conclusion, the DDO is simple to implement but it has two kinds of optimization variables x and u and also needs two optimization procedures: the first determines the optimal solution using safety factor, and the second yields the value of the reliability index. Note that DDO cannot perform design subject to a required reliability level. All numerical and semi-numerical methods RBDO satisfy the required reliability level but they are different at computing time, convergence stability, simplicity implementation, improvement of objective function value, kind of variables, suitable uses.

References Ditlevsen, O. & Madsen, H. 1996. Structural Reliability Methods. John Wiley & Sons. Feng, Y.S. & Moses, F. 1986. A method of structural optimization based on structural system reliability. J. Struct. Mech. 14:437–453. Kharmanda, G., Mohamed, A. & Lemaire, M. 2001. New hybrid formulation for reliabilitybased optimization of structures. The Fourth World Congress of Structural and Multidisciplinary Optimization, WCSMO-4, Dalian, China, 4–8 June 2001. Kharmanda, G., Mohamed, A. & Lemaire, M. 2002. Efficient reliability-based design optimization using hybrid space with application to finite element analysis. Structural and Multidisciplinary Optimization 24:233–245. Kharmanda, G., Mohamed, A. & Lemaire, M. 2003. Integration of reliability-based design optimization within CAD and FE models. In: Recent Advances in Integrated Design and Manufacturing in Mechanical Engineering, Kluwer Academic Publishers. Kharmanda, G., El-Hami, A. & Olhoff, N. 2004a. Global Reliability–Based Design Optimization. In: Frontiers on Global Optimization, C.A. Floudas (ed.), 255 (20), Kluwer Academic Publishers.

216

Structural design optimization considering uncertainties

Kharmanda, G. 2004b. Two points of view for developing reliability-based design optimization. NT2F4 (New Trends in Fatigue and Fracture IV), Aleppo, Syria, 10–12 May 2004. Kharmanda, G., Olhoff, N. & El-Hami, A. 2004c. Optimum values of structural safety factors for a predefined reliability level with extension to multiple limit states. Structural and Multidisciplinary Optimization, 27:421–434. Kharmanda, G., Olhoff, N. & El-Hami, A. 2004d. Recent Developments in Reliability-Based Design Optimization (Keynote Lecture). In: Computational Mechanics, Proc. Sixth World Congress of Computational Mechanics (WCCM VI in conjunction with APCOM’04), Sept. 5–10, 2004, Beijing, China. Tsinghua University Press & Springer-Verlag. Kharmanda, G. & Olhoff, N. 2007. Extension of optimum safety factor method to nonlinear reliability-based design optimization. Journal of Structural and Multidisciplinary Optimization, to appear. Kharmanda, G., Altonji, A. & El-Hami, A. 2006. Safest point method for reliability-based design optimization of freely vibrating structures. 1st International Francophone Congress for Advanced Mechanics, IFCAM01, Aleppo, Syria, 02–04 May 2006. Kharmanda, G., Makhloufi, A. & Elhami, A. 2007. Efficient computing time reduction for reliability-based design optimization. Qualita2007, 20–22 March 2007, Tanger, Maroc. Koch, P.N., Yang, R.J. & Gu, L. 2004. Design for six sigma through robust optimization. Struct. and Multidisc. Optim. 26:235–248. Mohsine, A., Kharmanda, G. & El-hami, A. 2006. Improved hybrid method as a robust tool for reliability-based design optimization. Structural and Multidisciplinary Optimization 32:203–213. Mohsine, A. 2006. Contribution à l’optimization fiabiliste en dynamique des structures mécaniques. Thèse de doctorat, INSA de Rouen, France (French version). Tu, J., Choi, K.K. & Park, Y.H. 1999. A new study on reliability-based design optimization. Journal of Mechanical Design, ASME 121(4):557–564. Youn, B.D. & Choi, K.K. 2004. Selecting Probabilistic Approaches for Reliability-Based Design Optimization. AIAA Journal 42(1):124–131. Yang, R.J., Chuang C., Gu, L. & Li, G. 2005. Experience with approximate reliabilitybased optimization methods II: an exhaust system problem. Structural and Multidisciplinary Optimization 29:488–497.

Chapter 9

Advances in solution methods for reliability-based design optimization Alaa Chateauneuf & Younes Aoues University Blaise Pascal, France

ABSTRACT: The solution of Reliability-Based Design Optimization implies high computational efforts due to the coupling of reliability and optimization problems. The probabilistic constraint is the key constraint in RBDO, which requires considerable computational effort and reveals the classical iterative problems of numerical efficiency, accuracy and stability. To solve the RBDO systems, three approaches are commonly used: the two-level approach, the one-level approach and the decoupled approach. A good algorithm should satisfy the conditions of efficiency, precision, generality and robustness. This chapter describes the recent advances in numerical methods for RBDO solution, in order to give a comprehensive overview of the basis and characteristics of the different approaches. The numerical applications on simple structures allow us to compare the efficiency of the RBDO approaches.

1 Introduction Reliability-Based Design Optimization (RBDO) aims at searching for the best compromise between cost reduction and reliability assurance, by considering system uncertainties. Although the basic RBDO ideas have been established more than thirty years ago, the solution is not yet easy, even for simple structures. The difficulty lies in the consideration of the reliability constraints, which require a large computational effort and involves classical numerical problems, such as convergence, accuracy and stability. The situation becomes worst when finite element and CAD models are involved, especially when material and geometrical nonlinearities are considered. While the optimization process is carried out in the space of the design variables, the reliability analysis is performed in the space of the random variables, where a lot of numerical calculations are required to evaluate the failure probability. Consequently, in order to search for the optimal structural configuration, the design variables are repeatedly changed, and each set of design variables corresponds to a new random variable space which then needs to be manipulated to evaluate the structural reliability at that point (Murotsu et al. 1994). Because of the too many repeated searches needed in the above two spaces, the computational time for such an optimization becomes the main problem. Figure 9.1 shows the main involved models in the RBDO modeling of engineering structures. The nested optimization, reliability, CAD and finite element models involve nonlinear iterative numerical procedures, where the problems of convergence, precision and computation time are omnipresent. For practical design, the cost of simple

218

Structural design optimization considering uncertainties

Optimization problem (design space) Reliability problem (random variable space) CAD model (geometrical variables) Finite Element Model (nodal variables) Nonlinearities and transient behavior (mechanical variables)

Figure 9.1 Nested models in the RBDO of engineering structures.

finite element (which is already a large time consuming procedure) is multiplied by a factor between ten and several thousands, which cannot be afforded in the design process. The computation scheme is thus a big problem that researchers should overcome in order to allow for practical applications. Generally speaking, a good solver should satisfy the conditions of efficiency (computation time), precision (accuracy of finding the optimum), generality (capability to deal with different kinds of problems with or without large number of variables) and robustness (stability of the convergence for any admissible initial point, local or global convergence criteria, etc). In the last decade, many advanced methods and techniques have been intensively developed in both fields: optimization and reliability. This chapter aims to describe the most common numerical methods to solve RBDO problems. After describing the basic formulation, the two-level, the one-level and the decoupled approaches are presented and discussed. Numerical applications are then presented for illustration and comparative purposes. For more details about the presented approaches, the reader is encouraged to review the original works referenced at the end of this chapter.

2 Basic RBDO formulation Basically, the RBDO problem is defined as the minimization of either the initial cost or the expected total cost (i.e. initial and expected failure costs), subjected to the constraint of an admissible failure probability Pft , in addition to the other structural constraints. As mentioned above, the particularity in RBDO lies in the computation of the reliability constraint, which involves additional computational effort and convergence difficulties. This constraint can be evaluated by one of the reliability methods, such as FORM/SORM, RSM or even Monte Carlo (Ditlevsen et Madsen 1996, Rackwitz 2001, Lemaire 2006). The RBDO is formulated as: min f (d) d

subject to Pf (d) ≤ Pft and gj (d) ≤ 0

(1)

A d v a n c e s i n s o l u t i o n m e t h o d s f o r r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

x2

Normalized space

u2

Physical space

Failure domain Gu (u, d)  0

Failure domain G(x, d)  0

P* u*2

G(x, d)
0

mx2

Safe domain mx

x1

1

219

β

MPP Gu(u, d)
0

α

u* 1

u1

Figure 9.2 Reliability index solution and probabilistic transformation.

where d is the vector of design variables, f (d) is the objective function, gj (d) are the structural deterministic constraints, Pf (d) is the failure probability of the structure and Pft is the admissible failure probability. In the First Order Reliability Method (FORM), the failure probability Pf is given as a function of the reliability index β: Pf (d) =
(−β(d)) ≈ Pr[G(X, d) ≤ 0]

(2)

where X is the vector of random variables (whose realization is noted x), Pr[·] is the probability operator and
(·) is the standard gaussian cumulated function. It is to be noted that the design variables d may be either independent deterministic variables or distribution parameters, especially the mean values, of some of the random variables. These two cases should be carefully taken into account when computing the gradient vectors. The reliability level is defined by an invariant reliability index β, as defined by (Hasofer and Lind 1974), which is evaluated by solving the constrained optimization problem: β = min u =



(Ti (x))2

i

(3)

under: G(T(x), d) ≤ 0 where u is the distance between the median point and the failure subspace in the normalized space ui and Ti (·) is an appropriate probabilistic transformation: i.e. ui = Ti (x). The image of the performance function G(x, d) in the normalized space is written: Gu (u,d) = G(x, d) (Figure 9.2). The solution of this problem is called the Most Probable Failure Point (MPP), the design point or the β-point, where β = u∗i ; it is noted P∗ or either x* or u*, whether physical or normalized space is considered, respectively. In fact, the term Most Probable Point is not rigorous from the probabilistic point of view, P∗ is just the point corresponding to the maximum joint density in the failure domain. However, in RBDO, the term MPP is preferred to the term design point, as it avoids confusion between design optimization and design for reliability.

220

Structural design optimization considering uncertainties

For the reliability problem describe in equation 3, the Kuhn-Tucker optimality conditions are written: ∇u u + λ∇u Gu (u∗ , d) = 0 Gu (u∗ , d) = 0

(4)

where ∇u is the gradient operator in the normalized space and λ is the Lagrange multiplier. The solution of the above equations leads to: λ = 1/∇u Gu (u∗ , d), and hence the reliability problem must satisfy the conditions: Gu (u∗ , d) = 0 ∇u Gu (u∗ , d) · u∗ + ∇u Gu (u∗ , d) · u = 0

(5)

The optimization process is carried out in the space of the design parameters d, which are deterministic. In parallel, the solution of the reliability problem is performed in the space of the random variables by solving the optimization problem in equation 3. Traditional reliability-based design optimization requires a double loop iteration procedure, where reliability analysis is carried out in the inner loop for each change in the design parameters, in order to evaluate the reliability constraints. The computational time for this procedure is extremely high due to the multiplication of the number of iterations in both optimization and reliability problems, involving a very high number of mechanical analyses. Recent developments in the literature aims at solving the numerical difficulties, through three main approaches: – –

–

Two-level approaches, which are based on the improvement of the traditional double-loop approach by increasing the efficiency of the reliability analysis. Mono-level approaches which aim at solving simultaneously the optimization and the reliability problems within a single loop dealing with both design and random variables. Decoupled approaches, where the reliability constraint is replaced by an equivalent deterministic (or pseudo-deterministic) constraint, involving some additional simplifications.

In the following sections, the basic ideas behind these approaches will be briefly described.

3 Two-level approaches A straight forward approach to solve RBDO problems is a two-level approach, where the outer loop aims to solve the optimization problem by improving the design variables d and the inner loop aims to solve the reliability problem by dealing with the random variables x. In order to reduce the computational effort in the two-level formulation, two RBDO approaches have been proposed to deal with probabilistic constraints: –

Reliability Index Approach (RIA) considers the cost reduction under the reliability index constraint.

A d v a n c e s i n s o l u t i o n m e t h o d s f o r r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

–

3.1

221

Performance Measure Approach (PMA) involves an inverse reliability problem as an alternative constraint. Reliability Index Approach (RIA)

Traditionally, the RBDO procedure is solved in the two spaces: the space of design variables, corresponding to deterministic physical space and the space of Gaussian random variables, obtained by probabilistic transformation of the random physical variables. In the classical approach, the RBDO is calculated by nesting the two following problems: •

optimization problem under reliability constraints: min f (d) x

subject to and

•

β(d) ≥ βt gj (d) ≤ 0

(6)

where f (d) is the objective function, gj (d) are the associated deterministic constraints, β(d) is the reliability index of the structure and βt is the target reliability. calculation of the reliability index β(d): min u = x

subject to



[Ti (x, d)]2

i

(7)

G(x, d) ≤ 0

where u is the distance between the origin and the considered point in the normalized random space, G(x,d) is the limit state function and Ti (·) is the probabilistic transformation to the normalized space. The solution of this RBDO problem consists in solving the two nested optimization problems. For each new set of the design parameters, the reliability analysis is performed in order to get the new MPP, corresponding to a given reliability level. As illustrated in Figure 9.3, this procedure leads to slow convergence scheme and zigzagging due to the sequential changes of the optimal point and the Most Probable Point. The method is somehow similar to relaxation procedures, known as a low convergence scheme. Actually, it is well established that RIA converges slowly or even fails to converge for a number of problems (Choi and Youn 2002). 3.2 Perf ormance Meas ure Approach (P MA) This method is based on an inverse reliability analysis, where the performance function level is specified as a constraint, instead of the reliability index itself (Tu et al. 1999, 2000). The performance measure is written: −1 Gp (d) = FG [
(−βt ); x, d]

(8)

222

Structural design optimization considering uncertainties

Target reliability index location

u2

u2

MPP

MPP

Limit state G (d) = 0

Limit state G (d)
0

G (d) > 0

βt

G (d)  0

x0

I nc

re

o as e

fG

(a)

x0 u1

(b)

In

a cre

se

of

G

u1

Figure 9.3 Illustration of (a) RIA and (b) PMA searches.

where
(·) is the standard gaussian cumulated distribution function and FG (·) is the −1 (·) its inverse. In the standard gaussian CDF of the performance function G(·) and FG space, the performance measure is directly evaluated at the Most Probable Failure Point P*, such that the target reliability can be satisfied. Gp (d) = G(x∗ , d | u∗  = βt )

(9)

The RBDO is then formulated as: min f (d) dk

subject to Gp (d) ≤ 0 and gj (x) ≤ 0

(10)

where Gp (·) is obtained by solving the problem defined in equation 9. PMA is shown to be efficient and robust, since it is easier to minimize a complicated objective function subjected to a simple constraint than to minimize a simple objective function subjected to a complicated constraint. However, several numerical examples using PMA show inefficiency and instability in the assessment of probabilistic constraints during the RBDO process, even with Advanced Mean Value AMV or Hybrid Mean Value HMV methods (Youn and Choi 2004a). For this reason, (Youn and Choi 2004b) proposed a coupling of HMV with Response Surface Method, specifically developed for reliability and optimization analyses. In Figure 9.3, the search scheme of PMA is compared with RIA. While RIA is zigzagging, PMA goes first to the hyper-sphere with a radius equal to the target reliability index, then iterations are carried out on this hyper-sphere. This is the reason why convergence is faster and more stable in the case of PMA.

A d v a n c e s i n s o l u t i o n m e t h o d s f o r r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

223

Although that many applications, such as those given by Frangopol (1995) and Nikolaidis and Burdisso (1988), are based on RIA algorithms, the PMA is increasingly used for large-scale problems. Lee et al. (2002) have conducted a comparative study between RIA and PMA, where RIA has shown to be less efficient for high reliability levels. They analyzed several examples and concluded that conventional RIA is not computationally attractive, compared with recently introduced target performance based approaches.

4 Mono-level approaches The mono-level methods are aimed at improving the efficiency of the RBDO procedures, by introducing the reliability at the same loop as optimization. The basis of the one-level approaches consists in solving both optimization and reliability problems without nesting the two problems. In this way, parallel convergence can be reached in both design and random spaces, and the computational cost may be saved. Among the mono-level approaches in the literature, one can indicate the well-known works of (Madsen and Friis Hansen 1992; Kuschel and Rackwitz 1997), which are based on reformulating the RBDO problem. The solution can then be obtained by traditional nonlinear optimization algorithms. 4.1 Total cost formulation The work of (Madsen and Friis Hansen 1992) belongs to the earlier efforts in this topic. They proposed a combined method integrating the expected failure cost in the objective function. The proposed mono-level formulation is written as: min CT (d) = CI (d) + Cf
(−u) d

subject to

Gu (u, d) = 0

and

∇u Gu (u, d) u =− u ∇u Gu (u, d)

(11)

The last condition can also be written: ∇u Gu (u, d) · u + ∇u Gu (u, d) · u = 0

(12)

This formulation has the advantage of being solved by standard optimization algorithms, but requires the explicit implementation of the probabilistic transformation, as well as the computation of the second order derivatives. The numerical examples carried out by the authors showed very large number of mechanical calls, compared to two-level RBDO models. Despite the high computational cost, further improvements of the combined method were still possible to make it an attractive alternative to the classical nested RBDO. 4.2

Formulation with optimality conditi o ns

Kuschel and Rackwitz (1997) have developed two formulations for RBDO: either by minimizing the expected total cost, or by maximizing the structural reliability for a

224

Structural design optimization considering uncertainties

given cost. In this mono-level approach, the reliability constraints are replaced by Karush-Kuhn-Tucker conditions for the first order reliability problem. These optimality conditions are then introduced as new constraints in the mono-level optimization problem. The total cost formulation is written as: min f (d, u) = CI (d) + Cf (d)
(−u) d

subject to

Gu (u, d) = 0 ∇u Gu (u, d) · u + ∇u Gu (u, d) · u = 0
(−u) ≤ Pft

(13)

u = T(x, d) and

gj (d) ≤ 0

The maximum reliability formulation is written as: max u d

subject to

Gu (u, d) = 0 ∇u Gu (u, d) · u + ∇u Gu (u, d) · u = 0 CI (d) + Cf (d)
(−u) ≤ Ct

and

(14)

u = T(x, d) gj (d) ≤ 0

To allow for efficient solution of both problems, the sensitivity operators are provided in the algorithm. The reliability index sensitivity is given by (Enevoldsen and Sørensen 1994): ∂β ∂Gu (u∗ , d) 1 = ∂dk ∇u Gu (u∗ , d) ∂dk

(15)

and the expected cost sensitivities are computed as following: ∂Cf (d) ∂CT (u, d) ∂CI (d) = +
(−u) ∂dk ∂dk ∂dk ∂CT (u, d) ui = −Cf (d) φ(u) ∂ui u

(16)

The authors have applied this approach on several examples and showed the efficiency of the approach. 4.3

Hyb ri d f or mulat io n

A mono-level approach has also been introduced by the hybrid formulation, proposed by (Kharmanda et al. 2002), allowing to combine deterministic and random variables.

A d v a n c e s i n s o l u t i o n m e t h o d s f o r r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

225

The RBDO formulation is based on defining a new objective function F(x,d) which integrates cost and reliability aspects as following: min F(d, x) = f (d) · Tβ (x, d) d,x

subject to

gj (d) ≤ 0

(17)

Tβ (x, d)) ≥ βt and

G(x, d) ≤ 0

where Tβ (x, d) is the image of u(x, d) in the physical space (while u(x, d) is a straight line, Tβ is generally a curve). The minimization of the function F(x,d) is carried out in the hybrid space of deterministic and random variables. An example of this hybrid design space is given in Figure 9.4, where the reliability levels Tβ are represented by ellipses (case of normal joint distribution), the objective function levels are given by solid curves and the limit state function is represented by dashed lines. Two important points can be observed: the optimal solution Pd∗ and the reliability solution Px∗ (i.e. the design point found on the curves G(x, d) = 0 and Tβ = βt ). This hybrid space contains all information about the RBDO model (e.g. optimal points, sensitivities, reliability levels, objective function iso-values and constraints). The optimality conditions for this hybrid formulation are: ∇x F(d, x) − λ∇x Tβ (x, d) + ∇x G(x, d) = 0 ∇d F(d, x) + λj ∇d gj (d) − λ∇d Tβ (x, d) + ∇d G(x, d) = 0 λj gj (d) = 0

(18)

λ(βt − Tβ (x, d)) = 0 G(x, d) = 0

Hybrid Design Space

Px*

f (x→)

Pd*

decr easin

g

Tb

Tb bt Tb
bt

Objective function levels

G(x→, y→)
0

→ →

G(x→, →y )0

Limit state decreasing G(x , y ) 0

x2, d2

x1, d1

Figure 9.4 Hybrid design space.

226

Structural design optimization considering uncertainties

It can be shown that these optimality conditions satisfy the initial two-level RBDO formulation (Kharmanda et al. 2003). While the method is theoretically attractive, numerical applications have shown that special care should be considered in the implementation of such a procedure, in order to ensure efficiency and convergence. Kaymaz and Marti (2006) have developed a specific formulation to apply two- and one-level approaches to elastoplastic structural behavior. In this study, the one-level approach required a more complex formulation and a larger number of optimization variables, but no difficulties have been observed for convergence. According to Royset et al. (2001), the mono-level approach may have several disadvantages: 1) even with first order optimization algorithms, the mono-level approach requires second-order derivatives, 2) an explicit formulation of the probabilistic transformation is required, 3) the mono-level approach is not suitable for system reliability constraints. Nevertheless, the mono-level approach seems to be very attractive, but still requires specific developments.

5 Decoupled approaches The idea of decoupling optimization and reliability problems seems to be very attractive as nested loops can be avoided and a lot of reliability analyses can be saved. This is generally carried out by defining a specific approximation and an equivalent deterministic parameter. However, the main challenge lies in the specification of the equivalent RBDO problem allowing to reach accurate precision. A basic idea consists in defining an equivalent deterministic constraint in terms of the standard deviation of the performance function. The optimal design is then searched for under approximated percentile of the performance function; this method is known as the Approximate Moment Approach (AMA). The updating of the equivalent deterministic constraint can be carried out by performing a reliability analysis after each convergence to new optimal points. Starting from the initial point, an alternative solution consists in performing reliability analysis to determine the Most Probable Failure Point (MPP) and hence the reliability index and the safety factors. The new limit state equation at the MPP is then used as a constraint in the deterministic optimization analysis, where the output are the new design parameters. These two steps can be solved in sequence until convergence (Torng an Yan 1993; Zou et al. 2004). Der Kiureghian and Polak (1988), Kirjner-Neto et al. (1998) and Royset et al. (2001) developed decoupled approaches by reformulating the RBDO problem as a deterministic semi-infinite optimization problem, where outer approximation method allows to solve the reliability problem independently of the optimization scheme. In this approach, the reliability constraint is firstly transformed into an infinite number of deterministic limit state constraints, and then the application of an outer programming algorithm allows to solve the RBDO problem. In a recent work, (Ching and Hsu 2006) proposed a method to transform the reliability constraint into deterministic constraint, by introducing the so-called limit state factor, multiplying the nominal limit state. When the equivalent deterministic constraint is defined, the RBDO can be solved as a classical deterministic optimization

A d v a n c e s i n s o l u t i o n m e t h o d s f o r r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

227

problem. However, it is to be proved whether this approach can be applied for real engineering structures, with several limit states involving large number of design and random variables.

5.1 Approximate Moment Approach (AMA) This approach transforms the probabilistic constraints to an approximated deterministic constraint, given by a percentile of the performance function (similar to the characteristic value approach in classical design). The RBDO problem is written as (Koch et al. 2007): min f (d) d

(19)

subject to

mG (d) + k σG (x, d) ≤ 0

where mG and σG are respectively the mean and the standard deviation of the performance function and k is a coefficient to be specified for a given safety level. Unlike other methods, the AMA does not require reliability analysis, as the required information are only the first and second moments of the performance function. While the mean is approximately computed in terms of the mean values of the random variables, the variance is based on first order development of the performance function, which can be written for independent variables as:

2 σG

=

  ∂G(x, d) i

∂xi

2

σX i

(20) x=mX

The method is efficient, as practically no extra cost is required, with respect to standard deterministic optimization. However, this approach implies many simplifications and cannot lead to accurate reliability results. Consequently, the error in the reliability estimation does not allow for convenient RBDO procedure, and in many cases leads to meaningless results. The main defect lies in the assumption that the random variables and the performance function are normally distributed, which is far from being appropriate for most of engineering structures. The other strong assumption lies in the computation of the variance of G, which assumes linear combination of random variables, leading to very limited application field.

5.2

Sequential Optimization with Reliabi l i ty A s s e s s me nt (SORA)

The Sequential Optimization with Reliability Assessment SORA is based on a single loop strategy composed by a sequence of deterministic optimization and reliability analyses. For each loop, the deterministic optimization is carried out, then the performance measure is checked and updated. The new value of the performance measure is then used in the next loop as a constraint limit.

228

Structural design optimization considering uncertainties

Three major ideas are introduced in the SORA method (Du & Chen 2002): –

A reliability percentile formulation is used to evaluate the design feasibility at the desired reliability level, An equivalent deterministic optimization is applied in order to reduce the number of reliability analyses, Efficient MPP search algorithm (the Modified Advance Mean Value MethodMAMV) is used for inverse reliability evaluation.

– –

The use of the reliability percentile instead of full reliability analysis leads to computational time reduction. This percentile allows for the identification of the feasible domain in design optimization. For a given reliability level R = 1 − Pf , the percentile reliability performance is given by: Gp = G(x∗ , d)

such as: Pr [G(X, d) ≥ Gp ] = R

(21)

The RBDO model can thus be written as: min f (d) d

subject to G(x∗ , d) ≥ 0

(22)

This formulation has the advantage of being fully deterministic, which can be solved by any classical optimization algorithm. However, the solution of equation 21 requires several calls to the structural model, which reduces the efficiency of the approach.

5.3

Seq uen ti al appr o ximat e pr o g r amm i n g (SAP)

In this method (Chen et al. 2006; Yi et al. 2006), a sequence of approximate programming is performed until the identification of the optimum point. In each subprogramming problem, the reliability analysis is approximated at the current MPP. By using suitable linearization, a recurrence formula derived from the optimality conditions at the MPP is developed in order to approximate the reliability index and its derivatives. At each step, the previously found MPP is taken as the linearization point. The use of response sensitivities improves the efficiency of the proposed algorithm. This procedure enables concurrent convergence of design optimization and reliability calculations. The optimization problem is written: min f (d) d

˜ d) ≥ βt subject to β(x, and

gj (d) ≤ 0

(23)

A d v a n c e s i n s o l u t i o n m e t h o d s f o r r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

229

where β˜ is the approximated reliability index, obtained through the recurrence formula: ˜ d β(x,

(r+1)

(r)  ∂β(x, ˜ d)  (r+1) (r) ˜ d )+ ) = β(x, − dl )  · (dl ∂dl  (r)

l

˜ d(r) ) = β(x, k

G(r) −

 i

∇u

(r)

∂G  ∂ui 

G(r) 

(r)

· ui

=

 (r) (r) G(r) − αi · ui (r) ∇u G 

(24)

i

˜ d(r) ) u∗i = −αi · β(x, (r)

where the subscripts (r) and (r + 1) indicate the iteration numbers and αi is the direction cosine for the variable ui . 5.4

Probabilistic s ufficiency factor

Wu et al. (2001) and Qu and Haftka (2003) introduced the probabilistic sufficiency factors in order to replace the RBDO with a series of deterministic optimizations by converting reliability constraints into equivalent deterministic constraints. For a prescribed failure probability Pft , the probabilistic sufficiency factor Psf is given by solving: Pr [γ ≤ Psf ] = Pft

(25)

where γ is the global safety factor, define by the random ratio between strength and stress. This means that the probabilistic sufficiency factor is simply a percentile of the safety factor that corresponds to the target failure probability. Qu and Haftka (2003) proposed to compute Pft by Monte Carlo simulations. When Pft is defined, the RBDO problem can be written as: min f (d) d

subject to

1 − Psf ≤ 0

and

gj (d) ≤ 0

(26)

It can be seen that the sufficiency factor constraint is equivalent to the target reliability constraint. The drawback of the method lies in the use of Monte Carlo simulations, which is generally large time consuming and presents significant numerical noise. 5.5

S ingle-loop double vector (SLDV)

In the Single-loop double-vector method (SLDV), there are two variable vectors: one for the mean values (design parameters) and one for the MPP values. This method has been improved by (Chen et al. 1997) who proposed to work with only one vector, leading to Single-loop single-vector approach (SLSV), on the basis of first order approximation of the limit state.

230

Structural design optimization considering uncertainties

6 System reliability optimization The progress in System Reliability-Based Design Optimization SRBDO is relatively slow because it depends on the system reliability analysis where the computational time, and the numerical instability lead to many difficulties in the SRBDO formulation. The solution of the relevant failure modes is a time consuming process, which is mainly due to design variable changes at each iteration of the optimization procedure. Consequently, the relevant failure modes also change during optimization iterations (e.g. a failure mode which is the most important within a given iteration may become negligible in the following iterations, due to design variable changes). The redundancy in the system reliability must be taken into account in the SRBDO Process (Murotsu et al. 1994), but it was found that a non redundant structure would need a higher safety margin than redundant one in order to achieve the same acceptable level of damage tolerance. Moses (1997) indicated that although many efforts have been made to compute the system reliability, the fundamental idea of system reliability problem is to extrapolate the analysis of the component reliability and performance to an overall structural risk assessment. Feng and Moses (1986) proposed an algorithm to identify the failure modes through incremental loading models, in order to be introduced in the system reliability constraint in the formulation of the reliability-based optimization. Different frameworks have presented many methodologies for SRBDO: the system reliability may be considered as a single probabilistic constraint (Moses 1997), the system reliability is replaced by the reliability indexes of the significant failure modes (Rackwitz 2001; Enevoldsen and Sørensen 1993), which is an alternative to the original formulation, and finally the system reliability constraint and component reliability constraints were simultaneously taken into account; An alternative approach of SRBDO can be based on multi-criteria optimization (Frangopol 1995, Kuschel and Rackwitz 2000). In an early work, (Enevoldsen and Sørensen 1993) proposed a sequential strategy to solve the RBDO of structural systems. The use of sensitivity operators for cost and reliability index, allows the authors to ensure stable convergence of the RBDO algorithm under system reliability constraints. More recently, Kuschel and Rackwitz (2000) proposed a mono-level approach for the reliability-based optimization of series systems. From another point of view, (Fu and Frangopol 1990) proposed to deal with RBDO of structural systems as a multi-objective optimization problem. This leads to a consistent decision making procedure for structural design and assessment. Although system optimization is usually considered either as a macro-component or as an independently acting components where safety constraints are specified separately, (Aoues and Chateauneuf 2007) proposed a scheme for consistent RBDO of structural systems. The basic idea consists in updating the component target safety levels in order to meet the overall system target and to avoid over-designed components. In the main optimization loop, the cost function is minimized under the constraints that component reliability indexes must satisfy the adapted target values. In the inner updating procedure, the target indexes are adjusted to

A d v a n c e s i n s o l u t i o n m e t h o d s f o r r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

231

meet the system reliability requirement. The proposed formulation is written in the form: ⎧ C(d) ⎪ ⎪min ⎪ ⎨ d Updated (27) subject to βj (d) ≥ βtj ⎪ ⎪ ⎪ ⎩ dL ≤ d ≤ dU Updated

where βtj is the updated target reliability index for the jth failure mode and βj (d) is the reliability index for the considered design configuration. The system reliability depends on its component reliabilities as well as the correlation ρjk between the different failure modes, it can be expressed as: βsys = f (β, ρ)

(28)

where β is the reliability index vector and ρ is the matrix of correlations between the failure modes. The embedded updating procedure is expressed by least square minimization for the difference between the updated targets and the actual indexes under the constraint of satisfying the required system safety. The procedure aims to solve the system: ⎧ mp  ⎪ ⎪ Updated ⎨ min (βtj − β j )2 Updated (29) βt i=1 j ⎪ ⎪ Updated ⎩ subjected to βsys (βtj , ρjk ) ≥ βt_sys Updated

where the updated targets βtj are themselves the optimization parameters. The optimal solution corresponds to the best quadratic fitting between the component Updated indexes βj and the corresponding target indexes βtj , under the constraint of satisfying the system target; this constraint is always active at the optimal solution: Updated βsys (βtj , ρjk ) = βt_sys . The updating procedure plays a key-role as it searches for the best values of the target indexes which pull down the reliability indexes for structural components.

7 Numerical applications In this section, three examples are presented in order to illustrate the application of RBDO methods. In the first example, a steel hook is optimized by a mono-level approach. The second example concerns a bracket truss, where different methods are compared for high nonlinear performance function. Finally, statically determinate and redundant trusses are optimized to show the numerical efficiency of the applied algorithms. 7.1

Steel hook

The RBDO is applied to the design of the steel hook shown in Figure 9.5 (Kharmanda et al. 2002). The hook is loaded by a shaft in contact with the circular surface of radius

232

Structural design optimization considering uncertainties

a

a

R2

R3

R3 t3

b

t3 L

c d

f

t2

t1

e t1

t2

Figure 9.5 Hook configuration and finite element mesh.

Table 9.1 Random and design variables. Variable

Mean

Std-deviation

a b c d e f t1 t2 t3 F

ma mb mc md me mf mt1 mt2 mt3 400

3 2 4 4 4 4 1 1 1 20

R1 and supported by an axis through the upper hole of radius R2 . While the upper part has uniform thickness, trapezoidal cross-section is chosen for the curved part in order to better redistribute the stresses. The following dimensions are fixed: the loading radius R1 = 190 mm, the hanging hole R2 = 100 mm, the filet radius R3 = 100 mm and the hook height L = 1200 mm. The used material is construction steel with Young’s modulus E = 200 GPa and yield stress fY = 235 MPa.The hook is modeled by 1602 solid 20-node elements, with 18,600 degrees of freedom, the applied load F = 400 kN is distributed over 30 elements on the contact surface. In this study, the optimal design is to be found under reliability considerations. The mean values of dimensional properties (ma , mb , mc , md , me , mf , mt1 , mt2 , mt3 ) are considered as design parameters d, while the applied force F and the geometric variables (a, b, c, d, e, f , t1, t2, t3) are taken as random variables X, as given in Table 9.1. For this problem, the target reliability index is set to βt = 3.35, corresponding to a failure probability of 4 × 10−4 .

A d v a n c e s i n s o l u t i o n m e t h o d s f o r r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

7.1.1

233

D D O soluti on

In Deterministic Design Optimization, the structural volume is minimized under the constraint of allowable stress corresponding to the yield stress divided by suitable safety factor. min V(d) d

subject to

SF σmax ≤ fY

(30)

where SF is the global safety factor, related to loading force F, is set to 1.5 according to common practice. The optimal volume is found to be: VDDO−1.5 = 0.2927 × 108 mm3 and the optimal design is given in Table 9.2. For this solution, the reliability analysis is carried out, leading to the reliability index: β = 7.49 which is much higher than the target value: βt = 3.35. Following this result, a cost reduction has been decided by decreasing the global safety factor to: SF = 1.25. In this case, the optimal volume is decreased to: VDDO−1.25 = 0.2508 × 108 mm3 and the corresponding reliability index is: β = 3.64 > βt . Three disadvantages can be observed in DDO approach: the first one is that safety factors given in recommendation are not always suitable for structural systems, the second one is the difficulty of reasonable choice of the safety factor because of their critical role on manufacturing cost and structural reliability, and the third one is the bad distribution of safety margins for different variables due to global scaling of the safety level. For these reasons, it is very important to integrate the reliability analysis in the optimization process. 7.1.2

RBD O soluti on

The Reliability-Based Design Optimization is formulated by introducing explicitly the reliability constraint: min V(d) d

subject to

β(d) ≥ βt

(31)

where the reliability index is calculated by the solution of: min u(x, d) x

subject to

σmax (x, d) ≥ fY

(32)

In this formulation, the stress becomes a random function. The hybrid formulation is applied to solve the RBDO problem, leading to the optimal design parameters and partial safety factors indicated in Table 9.2. While DDO is based on global safety factor for loading F (SF = 1.25), RBDO shows that some parameters of the structure such as dimensions, can also play a very important role on the structural safety (γt2 = 1.306 and γF = 1.068). Therefore, RBDO satisfies the required reliability level by adding or removing material where it is necessary, and hence improves the structural performance

234

Structural design optimization considering uncertainties Table 9.2 Optimal solutions obtained by DDO and RBDO. Variable

a b c d e f t1 t2 t3 F Volume Reliability

DDO SF = 1.5

SF = 1.25

125.7 74.5 187.1 216.5 185.8 173.5 39.4 10.4 13.2 – 0.2927 × 108 7.49

135.4 78.1 191.1 219.3 191.4 181.2 30.8 10.0 10.5 – 0.2508 × 108 3.64

DDO stress distribution

RBDO optimum

Safety factors

110.7 80.0 198.2 198.2 198.1 151.6 27.8 13.1 10.1 – 0.235 × 108 3.36

1.005 1.006 1.001 1.001 1.002 1.000 1.007 1.306 1.006 1.068

RBDO stress distribution

Figure 9.6 Optimal solutions for DDO and RBDO.

by reducing the structural volume in uncritical regions. This can be understood as a better distribution of the safety factors. Figure 9.6 shows the stress distributions resulting from DDO and RBDO procedures. It can be seen that stress field is more homogeneous for Reliability-Based Design Optimization than the distribution in the Deterministic Design Optimization. 7.1.3 Bra c k et st r u ct u r e Figure 9.7 shows a two-member bracket supporting a vertical load P applied at a distance L from the wall hinge. The member AB, with 60◦ of inclination, is linked to member CD through a pin-joint at B. Both members have rectangular cross-sections: wAB × t for AB and wCD × t for CD, w stands for width and t for thickness. It is aimed

A d v a n c e s i n s o l u t i o n m e t h o d s f o r r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

235

L L /3

C

D

B PW

E 60˚ E

w

t Cross section E-E A

Figure 9.7 The parameterization of the bracket structure.

to optimally define the values of the parameters: t, wAB and wCD , by considering the uncertainties in material and geometrical properties. The two design constraints are: •

the maximum bending stress σb in member CD must be less than the yield stress fY , taken as 225 MPa for the used steel. The maximum bending stress σb is located at point B and is given by the following formulas: σb =

•

6 MB wCD t 2

with: MB =

PL ρgwCD tL2 + 3 18

(33)

the compression force FAB in member AB must be less than the buckling load Fb . The normal force in member AB is given by: 

1 3P 3ρgwCD tL + (34) FAB = 2 4 cos θ

and the buckling load for member AB is written as: Fb =

3 π2 E t wAB π2 EI = ! 2L "2 2 LAB 12 3 sin θ

(35)

Therefore, it is aimed to minimize the structural weight under the two limit states: G1 = fY − σb (wCD , t, L, P) G2 = Fb (wAB , t, L) − FAB (wCD , t, L, P)

(36)

The deterministic optimization is performed by using the partial safety factors, corresponding to live and dead load factors: γs = 1.5 and γp = 1.35, respectively. For bending stress, the partial factor is γr = 1.5; hence, in DDO, the admissible stress bending is fy /γr . The random variables are given in Table 9.3, where the design variables are considered as the distribution means of the geometrical properties.

236

Structural design optimization considering uncertainties Table 9.3 Statistical data of the random variables. Parameter

Symbol

Mean

C.O.V

Distribution

Applied load Young’s modulus Yield stress Unit mass Length Width of member AB Width of member CD Thickness

P (kN) E (GPa) f y (MPa) ρ (kg/m3 ) L (m) wAB (m) w CD (m) t (m)

100 200 225 7860 5 wAB w CD t

0.15 0.08 0.08 0.10 0.05 0.05 0.05 0.05

Gumbel Gumbel Lognormal Weibull Normal Normal Normal Normal

Table 9.4 Summary of the numerical results in the design of bracket structure. Design method

weight (kg)

βG1

βG2

Iteration

CPU

G-eval

wAB (cm)

w CD (cm)

t (cm)

DDO RIA PMA SORA

787.17 678.18 678.88 678.88

4.86 1.99 2.00 2.00

2.94 2.00 2.01 2.01

9 5 7 22

0.07 0.45 0.57 0.39

40 2340 2736 1340

6.13 6.08 6.08 6.08

20.21 15.68 15.69 15.69

26.94 20.91 20.91 20.91

For a target reliability βt = 2 (corresponding to a failure probability of 1%), the reliability-based optimization problem is written: ) √ * 4 3 wAB + wCD min W = ρgtL wAB ,wCD ,t 9 subject to β1 ≥ 2 and β2 ≥ 2

(37)

where β1 and β2 are the reliability indexes related to G1 and G2 , respectively. Table 9.4 compares the optimization results and computational effort for different methods. It can be first seen that deterministic optimization often leads to high cost and reliability, while RBDO approaches lead to better fit of the target safety. The Reliability Index Approach (RIA), the Performance Measure Approach (PMA) and the Sequential Optimization with Reliability Assessment (SORA) lead to the same design point, corresponding to 12.7% of weight reduction. However, the computational cost in SORA is much less than for the other RBDO methods: 1340 evaluations of the performance function instead of 2340 and 2736 evaluations for RIA and RBDO-PMA, respectively. At the optimal RBDO point, Table 9.5 indicates the Most Probable Failure Point and the corresponding partial safety factors. The load factor is lower for the buckling limit state, as the reliability is more affected by the other random variables, especially by the width wAB . Compared to DDO, these results show the advantage of RBDO in adjusting the partial safety factors in terms of the reliability sensitivity with respect to the uncertain variables, which have different influences on the different failure modes.

A d v a n c e s i n s o l u t i o n m e t h o d s f o r r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

237

Table 9.5 The most probable point and partial safety factors for RBDO solution.

G1 G2 Safety factors γ G1 γ G2

P* (kN)

E* (GPa)

f Y∗ (MPa)

ρ∗ (kg/m3 )

L* (m)

∗ wAB (cm)

w ∗CD (cm)

t* (cm)

126.74 117.45

197 191

212.48 224.28

7739 7754

5.10 5.17

6.08 5.67

15.37 15.70

20.06 20.56

1.26 1.17

1.04 1.05

1.06 1.00

1.01 1.01

1.03 1.03

1.00 1.07

1.02 1.00

1.04 1.02

L 3

2L 3 C B α

D

θ t

h

A

w

W

Figure 9.8 Inclined bracket structure.

The bracket structure is now considered by introducing the inclination of the bar AB as an additional design parameter (Figure 9.8). This inclination is defined by the angle α. The normal force FAB in member AB takes the form: FAB =

 L ρgwCD tL P+ h sin θ 2 cos α

(38)

and the maximum bending moment is: MB =

PL ρgwCD tL2 + 3 18 cos α

(39)

The angle α introduces a high degree of nonlinearity in the limit state functions, allowing to test the stability of the RBDO methods. Even with many initial trials, the Reliability Index Approach (RIA) could not converged, because of the limit state nonlinearity. The Performance Measure Approach (PMA) approach did not converge when the AMV algorithm (Advanced Mean Value Method) was applied to perform the inverse reliability analysis and to estimate the performance measure. However, the use of HMV algorithm (Hybrid Mean Value) allows the PMA to converge. This result confirms that HMV algorithm is more convenient for highly nonlinear limit states. The optimal inclination of the bracket is 25.4◦ for DDO and 24.5◦ for RBDO. Figure 9.9 shows the convergence of the performance measure and the reliability index

238

Structural design optimization considering uncertainties

1000

12 βG

1

10

0

βG

2

GPMA

Reliability index

Performance measures

8 –1000

1

GPMA

–2000

2

GSORA

1

–3000

GSORA

2

6 4 2 0

–4000 –2 –5000 –6000 0 (a)

–4 1

2

3

5 4 Iterations

6

7

8

–6

0 (b)

0.5

1

1.5

2

2.5 3 Iterations

3.5

4

4.5

5

Figure 9.9 (a) Performance measure in PMA and SORA. (b) Reliability index during PMA iterations.

Table 9.6 Numerical results in the design of bracket structure with inclination. βG1

Design method

weight (kg)

DDO RIA PMA SORA

716.96 4.87 Not converged 556.44 2.07 556.45 2.07

G-eval

CPU (s)

wAB (cm)

w CD (cm)

t (cm)

α(◦ )

22

147

0.16

5.36

20.24

27.00

25.43

13 30

6790 1744

1.42 0.51

5.37 5.37

15.69 15.69

20.93 20.93

24.64 24.53

βG2

Iteration

2.77 2.00 2.00

during optimization iterations. For both limit states, the reliability indexes converge to the target value βt = 2. It can be generally observed that PMA converges slower than SORA for this kind of problem. Table 9.6 confirms these results by indicating 6790 mechanical calls for PMA, against only 1744 calls for SORA. Once again, SORA has proven to be more efficient and robust than RIA and PMA. 7.2 T i m b er trus s The design of timber trusses is usually carried out by checking the ultimate crosssection capacities with respect to the ultimate limit state. However, as these structures are made of the assembly of several members, the overall ultimate capacity is highly conditioned by the redundancy degree. In many structures, several components can reach their ultimate capacity largely before reaching the overall structural failure load. On the other hand, the structure could contain a number of critical members, that produces the overall failure if any one of them fails, even for redundant structures. In this context, the system reliability can be greatly different from the reliability of its components. The numerical applications are carried out for two roof trusses (Chateauneuf and Noret 2005), where the depth and the breadth of the horizontal, bracing and upper

A d v a n c e s i n s o l u t i o n m e t h o d s f o r r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

239

Table 9.7 Model parameter data. Parameter

Value

Young’s Modulus Poisson’s ratio Timber density Distance between trusses Truss span Truss height Beam breadth

11 GPa 0.25 420 kg/m3 5m 20 m 5.77 m 0.10 m

Table 9.8 Statistical data of the random variables. Parameter

Characteristic value

Mean

C.O.V

Distribution

Permanent load (Roof Load) (kN/m2 ) Variables load (Concentrated Load) (kN) Snow (kN/m2 ) Wind (kN/m2 ) Bending strength (MPa) Tension strength (MPa) Compression strength (MPa) Young’s Modulus (MPa)

479.2 1422.3 932.5 400.8 14.16 8.26 12.39 8654.8

384.6 1071.4 625 301.8 24 14 21 11 000

0.15 0.25 0.30 0.20 0.25 0.25 0.25 0.13

Normal Gumbel Normal Weibull Lognormal Lognormal Lognormal Lognormal

roof member are considered as design variables. The two trusses correspond to statically determinate and indeterminate structures, respectively. In the RBDO analysis, the uncertainties of the strength and the applied loads are considered random variables, as detailed in Table 9.8. The characteristic values correspond to a percentile of 95% for loading and 5% for timber strength. The target failure probability is set to 10−4 which corresponds to βc = 3.7. The RBDO algorithms are implemented in Matlab environment (Mathworks Inc. 2007), where the optimization toolbox is applied to solve the system. The mechanical computation is carried out by the Finite element method, using CALFEM library (CALFEM 2007). The comparative study is performed for different RBDO methods. The limit state functions considered in this application are: ⎧

 ⎪ σc,d 2 σm,d ⎪ ⎪ + ≤ 1 in compression ⎨G = fc,d fm,d ⎪ σt,d σm,d ⎪ ⎪ + ≤1 in tension ⎩G = ft,d fm,d

(40)

where σc,d , σt,d , σm,d , fc,d , ft,d and fm,d are respectively the design values of stress and strength in compression, tension and bending.

240

Structural design optimization considering uncertainties

b.L

Z

X L

a.L

Figure 9.10 Truss with rigid joints. Table 9.9 Initial design values and bounds. Design variable

Initial design

Lower bound

Upper bound

D1 (member 1,2,3) (cm) H1 (member 1,2,3) (cm) D2 (member 3,4,5,6) (cm) H2 (member 3,4,5,6) (cm) D3 (member 7,8,9,10,11) (cm) H3 (member 7,8,9,10,11) (cm)

30 10 40 10 10 10

2 2 2 2 2 2

100 100 100 100 100 100

Table 9.10 Optimal results according to different RBDO methods. Design method

Optimal weight (kg)

min{β1 , β2 ,…,β11 }

Optimization iterations

Reliability iterations

FEA calls

CPU (s)

DDO RIA PMA SORA

928.51 802.61 802.61 802.49

4.37 3.70 3.70 3.70

6 12 12 15

– 1095 807 171

49 103 521 70 323 2610

0.36 749 510 20

In the Deterministic Design Optimization, the partial safety factors are drawn from the Eurocodes; the strength modification factor is also introduced in DDO and in RBDO, to account for the humidity and the duration of loading. 7.2.1 S ta ti c a l ly d e t e r m in a t e t r u s s The truss, illustrated in Figure 9.10, is formed by 11 rectangular timber members. The cross-sections are rectangular with breadth b and depth d, where the initial values of the six design variables are given in Table 9.9. It is to mention that this truss presents 11 performance functions, involving 11 reliability constraints in the RBDO procedure. It is thus aimed to keep the lowest reliability index above the target level of 3.7. Table 9.10 compares the results obtained by the different methods: DDO, RIA, PMA and SORA. All the RBDO approaches converge to the optimal weight of 802.6 kg, which is 14% lower than the deterministic result. While the number of optimization

A d v a n c e s i n s o l u t i o n m e t h o d s f o r r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

241

Table 9.11 Iteration of the reliability analysis. Method

0

1

2

3

4

5

6

7

8

9

10

11

12

Total

RIA PMA SORA

88 63 69

71 62 51

89 62 51

89 62

89 62

89 62

89 62

90 62

89 62

78 62

78 62

78 62

78 62

1095 807 171

Table 9.12 Numerical comparison of the optimal design. Method

d1

b1

d2

b2

d3

b3

DDO RIA PMA SORA

14.64 13.11 13.11 13.11

13.91 12.45 12.45 12.45

36.53 34.39 34.39 34.39

18.26 17.19 17.19 17.19

11.74 10.72 10.72 10.72

11.15 10.20 10.20 10.18

iterations is comparable for different RBDO methods, the number of reliability iterations is much higher for RIA and even for PMA. The number of the Finite Element Analyses (FEA) is huge for these methods: more than 1,00,000 runs for RIA and 70,000 runs for PMA. It is to note that these FEA include those necessary to compute the constraint gradients by finite difference techniques. The fifth column in Table 9.10 gives the number of iterations realized either to perform the reliability analyses in RIA or the inverse reliability analysis in PMA and SORA. Table 9.11 shows how these reliability iterations (inner loop) are distributed over the optimization iterations (outer loop). In the decoupled method (i.e. SORA), this number corresponds to the number of reliability iterations at each cycle of the equivalent deterministic design. The optimal designs from these optimization methods are detailed in Table 9.12. All the RBDO methods converge to the same optimal design, where all the dimensions are lower than those from deterministic optimization. The iteration history is illustrated in Figure 9.11, where the decoupled approach (SORA) can be easily distinguished. Figure 9.12 compares the characteristics of the RBDO methods on the basis of the evaluation criteria: Cost, Safety, number of iterations, FEA calls and CPU time.

7.2.2

Bra c ed t russ

Let us consider the same truss layout with additional members forming an X bracing system; the truss has now 25 members. The cross-section depths and breadths are noted d1 , b1 for members 1 to 6, d2 , b2 for members 7 to 12, d3 , b3 for members 13 to 25. The redundant truss configuration implies a large amount of internal force redistribution during the optimization process. Among the 25 limit states, the critical failure modes change along the optimization iterations. The structural response becomes strongly nonlinear due to interdependence of design and random variables.

242

Structural design optimization considering uncertainties

1000 DDO RBDO-RIA RBDO-PMA SORA

Structural weight

900 800 700 600 500 400 0

2

4

6

8 10 Iterations

12

14

16

Figure 9.11 History of the structural weight.

MEF-eval

RBDO-RIA RBDO-PMA RBDO-SORA DDO

CPU

Optimal cost

Reliability index

Iterations

Figure 9.12 Numerical performance of the design optimization methods.

The Reliability Index Approach could not converge in this example because of the limit state non-linearity and of the probabilistic transformations. In addition, the bracing members in this example have large reliability indexes, their evaluation implies very high time consumption and leads to the divergence of the optimization algorithm. The low number of finite element calls in SORA explains why the CPU time is so low for this method, compared to PMA. This advantage is even larger for more complex structural models. Although that PMA and SORA lead to almost the same structural

A d v a n c e s i n s o l u t i o n m e t h o d s f o r r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

L3

L8

L22

L18 Y

L7

Z

X L1

L13

L15

L10

L17

L23

L25 L24

L11 L16

L19

L21 L20

L3

L2

243

L4

L14

L5

L12 L6

Figure 9.13 Braced truss with 25 members.

Table 9.13 Optimization results for the truss with X bracing. min{β1 , β2 ,…,β25 }

Design method

Optimal weight (kg)

(DDO RIA PMA SORA

1080.54 5.234 No convergence 682.24 3.700 677.87 3.699

Optimization iterations

Reliability iterations

5

FEA calls

–

6 17

989 327

42 76 153 5962

CPU (s) 1.67 1175 97

Table 9.14 Optimal designs of the truss with X bracing.

DDO PMA SORA

d1

b1

d2

b2

d3

b3

30.23 23.02 16.51

15.11 11.51 15.69

30.18 24.45 24.49

15.09 12.22 12.24

10.43 8.51 8.49

9.90 8.08 8.07

weight and reliability index, the optimal design parameters are quite different in both approaches, especially for the dimensions b1 and d1 , as indicated in Table 9.14. Figure 9.14 shows the iteration history for the three methods: DDO, PMA and SORA. Although that SORA requires more iterations than PMA, it involves lower number of reliability analyses, leading to a global reduction of the computation cost. It proves, once more, its capacity to deal with engineering structures, by ensuring convergence stability and efficiency.

8 Conclusions As briefly described in the previous sections, a very intensive research activity is performed in the field of RBDO solution methods. Three approaches are usually adopted: two-level, mono-level and decoupled approaches. Although significant progress is performed in developing efficient numerical methods, the application to practical engineering structures is still a challenge, knowing the complexity of realistic industrial systems. In order to select a method, the designer has to search for a reasonable compromise

244

Structural design optimization considering uncertainties

1100

DDO RBDO-PMA SORA

1000 bopt = 5.23

Structural weight

900 800 bopt = 3.70 700 600 500 400 300 0

2

4

6

8 10 Iterations

12

14

16

Figure 9.14 History of the optimal design.

between the accuracy, the efficiency and the robustness of the applied RBDO algorithm. As a basic choice, the two-level approach requires less development effort to carry out RBDO. In this category, the performance measure approach leads to robust and efficient scheme, with respect to conventional reliability index approach. Globally, the decoupled approaches, such as the Sequential Optimization with Reliability Assessment, are very interesting, as they are stable and highly efficient, as many reliability analyses can be avoided. In all cases, the RBDO algorithms should be considered with special care and the results should be well validated by the designer, especially for complex structural systems where several failure points and local optima often co-exist.

References Aoues, Y. & Chateauneuf, A. 2007. Reliability-based optimization of structural systems by adaptive target safety application to RC frames. Structural Safety. Article in Press. CALFEM, A finite element toolbox to MATLAB, Version 3.3, Division of Structural Mechanics and Division of Solid Mechanics, Lund University, Sweden, http://www.civeng.ucl.ac.uk/ Chateauneuf, A. & Noret, E. 2005. System reliability-based optimization of redundant timber trusses. In: J.D. Sørensen (ed.). Reliability and optimization of structural systems, Proceedings of the IFIP WG7.5 Working Conference on reliability and optimization of structural systems, Aalborg, Denmark, May. Chen, X., Hasselman, T.K. & Neill, D.J. 1997. Reliability based structural design optimization for practical applications. Proceedings of the 38th AIAA/ASME/ASCE/AHS/ASC structures, structural dynamics and material conference, Kissimmee, Florida, AIAA-97-1403. Cheng, G., Xu, L. & Jiang, L. 2006. A sequential approximate programming strategy for reliability-based structural optimization. Computers and Structures. Article in Press.

A d v a n c e s i n s o l u t i o n m e t h o d s f o r r e l i a b i l i t y-b a s e d d e s i g n o p t i m i z a t i o n

245

Ching, J. & Hsu, W.-C. 2006. Transforming reliability limit state constraints into deterministic limit state constraints. Structural Safety. In Press. Choi, K.K. & Youn, B.D. 2002. On probabilistic approaches for reliability-based design optimization. In: 9th AIAA/NASA/USA/ISSMO symposium on Multidisciplinary Analysis and Optimization, September 4–6, Atlanta, GA, USA. Der Kiureghian, A. & Polak, E. 1988. Reliability-based optimal design: a decoupled approach. In: A.S. Nowak (ed.). Reliability and optimization of structural systems, Proceedings of the 8th IFIP WG7.5 Working Conference on reliability and optimization of structural systems, Chelsea, MI, USA: Book Crafters. pp. 197–205. Ditlevsen, O. & Madsen, H.O. 1996. Structural reliability methods. John Wiley & Sons. Du, X. & Chen, W. 2002. Sequential optimization and reliability assessment method for efficient probabilistic design. ASME, design engineering technical conferences and computers and information in engineering conference, DETC2002/DAC-34127, Montreal, Canada. EN 1995-1-1, Eurocode 5: Design of timber structures; Part 1-1: General rules and rules for buildings. Comité Européen de Normalisation, 2005. Enevoldsen, I. & Sørensen, J.D. 1993. Reliability-based optimization of series systems of parallel systems. Journal of Structural Engineering 119(4):1069–1084. Enevoldsen, I. & Sørensen, J.D. 1994. Reliability-based optimization in structural engineering. Structural Safety 15:169–196. Feng, Y.S. & Moses, F. 1986. A method of structural optimization based on structural system reliability. J. Struct. Mech. 14(4):437–453. Fu, G. & Frangopol, D.M. 1990. Reliability-based Vector optimization of structural systems, J. of Struct. Engrg. ASCE 116(8):2143–2161. Hasofer, A.M. & Lind, N.C. 1974. An Exact and Invariant First Order Reliability Format. J. Eng. Mech. ASCE 100, EM1:11–121. Kaymaz, I. & Marti, K. 2006. Reliability-based design optimization for elastoplastic mechanical structures. Computers and Structures. Article In Press. Kharmanda, G., Mohamed-Chateauneuf, A. & Lemaire, M. 2002. Efficient reliability-based optimization using a hybrid space with application to finite element analysis. Journal of Structural and Multidisciplinary Optimization 24(3):233–245. Kirjner-Neto, C., Polak, E. & Der Kiureghian, A. 1998. An outer approximations approach to reliability-based optimal design of structures. Journal Optim. Theory Appl. 98(1):1–16. Koch, P.N., Yang, R.J. & Gu, L. Design for six sigma through robust optimization. Structural and Multidisciplinary optimization. In Press. Kuschel, N. & Rackwitz, R. 1997. Two basic problems in reliability-based structural optimization. Mathematical Methods of Operations Research 46:309–333. Kuschel, N. & Rackwitz, R. 2000. A new approach for structural optimization of series system. In: R.E. Melchers, M.G. Stewart (ed.). Proceedings of the 8th International conference on applications of statistics and probability (ICASP) in Civil engineering reliability and risk analysis, Sydney, Australia, December 1999, Vol. 2, pp. 987–994. Lee, J.O., Yang, Y.S. & Ruy, W.S. 2002. A comparative study on reliability index and target performance based probabilistic structural design optimization. Computers and Structures (80):257–269. Lemaire, M., in collaboration with Chateauneuf, A. & Mitteau, J.C. 2006. Structural reliability. ISTE, UK. Madsen, H.O. & Friis Hansen, P. 1992. Comparison of some algorithms for reliability-based structural optimization and sensitivity analysis. In: R. Rackwitz & P. Thoft-Christensen (eds): Reliability and optimization of structural systems, Proceedings of the 4th IFIP WG7.5 Working conference on Reliability and Optimization of Structural Systems, Munich, Germany, September 1991. Berlin: Springer. pp. 443–451. Mathworks Inc. www.mathworks.com, 2007.

246

Structural design optimization considering uncertainties

Moses, F. 1997. Problems and prospects of reliability based optimization. Engineering Structures 19(4):293–301. Murotsu, Y., Shao, S. & Watanabe, A. 1994. An approach to reliability-based optimization of redundant structures. Structural Safety 16:133–143. Nikolaidis, E. & Burdisso, R. 1988. Reliability-based optimization: a safety index approach. Computer and Structures 28(6):781–788. Qu, X. & Haftka, R.T. 2003. Design under uncertainty using Monte Carlo simulation and probabilistic sufficiency factor. In: Proceedings of DET’03 conference, Chicago, IL,USA. Rackwitz, R. 2001. Reliability analysis, overview and some perspectives. Structural Safety 23:366–395. Royset, J.O., Der Kiureghian, A. & Polak, E. 2001. Reliability-based optimal structural design by the decoupling approach. Reliability Engineering and System Safety 73:213–221. Torng, T.Y. & Yan, R.J. 1993. Robust structural system design using a system reliabilitybased design optimization method. In: P.D. Spanos & Y.T. Wu (ed.), Probabilistic Mechanics: Advances in structural reliability methods, Springer-Verlag, NY:534–549. Tu, J., Choi, K.K. & Park, Y.H. 1999. A new study on reliability-based design optimization. Journal of Mechanical Design 121:557–564. Tu, J., Choi, K.K. & Park, Y.H. 2000. Design potential method for robust system parameter design. AIAA Journal 39(4):667–677. Youn, B.D. & Choi, K.K. 2004. Selecting probabilistic approaches for reliability-based design optimization. AIAA Journal 42(1):124–131. Youn, B.D. & Choi, K.K. 2004. A new response surface methodology for reliability-based design optimization. Computeres and Structures 82:241–256. Yi, P., Cheng, G. & Jiang, L. 2006. A sequential approximate programming strategy for performance-measure based probabilistic structural design optimization. Structural Safety. Article in Press. Wu, Y.T., Shin, Y., Sues, R. & Cesare, M. 2001. Safety factor based approach for probabilitybased design optimization. In: Proceedings of the 42nd AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, Seattle, WA, USA, Paper n◦ AIAA 2001-1522. Zou, T., Mahadevan, S. & Sopory, A. 2004. A reliability-based design method using simulation techniques and efficient optimization approach. ASME Design Engineering technical conferences, Salt Lake City, Utah, DETC2004/DAC-57457.

Chapter 10

Non-probabilistic design optimization with insufficient data using possibility and evidence theories Zissimos P. Mourelatos & Jun Zhou Oakland University, Rochester, MI, USA

ABSTRACT: Early in the engineering design cycle, it is difficult to quantify product reliability due to insufficient data or information for modeling the uncertainties. Design decisions are therefore, based on fuzzy information that is vague, imprecise, qualitative, linguistic or incomplete. The uncertain information is usually available as intervals with lower and upper limits. In this chapter, the possibility and evidence theories are used to account for uncertainty in design with incomplete information. Possibility-based and evidence-based design optimization methods are presented which handle a combination of probabilistic and non-probabilistic design variables. Also, a computationally efficient sequential possibility-based design optimization (SPDO) method is implemented, which decouples the design loop and the reliability assessment of each constraint. Two numerical examples demonstrate the application of possibility and evidence theories in design and highlight the trade-offs among reliability-based, possibility-based and evidence-based designs.

1 Introduction Engineering design under uncertainty has recently gained a lot of attention. Uncertainties are usually modeled using probability theory. In Reliability-Based Design Optimization (RBDO), variations are represented by standard deviations which are typically assumed constant, and a mean performance is optimized subject to probabilistic constraints (Lee et al. 2002, Liang et al. 2007, Tu et al. 1999, Wu et al. 2001, Youn et al. 2001). In general, probability theory is very effective when sufficient data is available to quantify uncertainty using probability distributions. However, when sufficient data is not available or there is lack of information due to ignorance, the classical probability methodology may not be appropriate. For example, during the early stages of product development, quantification of the product’s reliability or compliance to performance targets is practically very difficult due to insufficient data for modeling the uncertainties. A similar problem exists when the reliability of a complex system is assessed in the presence of incomplete information on the variability of certain design variables, parameters, operating conditions, boundary conditions etc. Uncertainties can be classified in two general types; aleatory (stochastic or random) and epistemic (subjective) (Klir & Filger 1988, Klir & Yuan 1995, Oberkampf et al. 2001, Sentz & Ferson 2002, Yager et al. 1994) Aleatory or irreducible uncertainty is related to inherent variability and is efficiently modeled using probability theory. However, when data is scarce or there is lack of information, the probability theory is not

248

Structural design optimization considering uncertainties

useful because the needed probability distributions cannot be accurately constructed. In this case, epistemic uncertainty, which describes subjectivity, ignorance or lack of information, can be used. Epistemic uncertainty is also called reducible because it can be reduced with increased state of knowledge or collection of more data. Formal theories to handle uncertainty have been proposed in the literature including evidence theory or Dempster – Shafer theory (Klir & Filger 1988, Yager et al. 1994), possibility theory (Dubois & Prade 1988) and interval analysis (Moore 1966). Two large classes of fuzzy measures, called belief and plausibility measures, respectively, characterize the mathematical theory of evidence. They are mutually dual in the sense that one of them can be uniquely determined from the other. Evidence theory uses plausibility and belief (upper and lower bounds of probability) to measure the likelihood of events. When the plausibility and belief measures are equal, the general evidence theory reduces to the classical probability theory. Therefore, the classical probability theory is a special case of evidence theory. Possibility theory handles epistemic uncertainty if there is no conflicting evidence among experts (Klir & Filger 1988). It uses a special subclass of dual plausibility and belief measures, called possibility and necessity measures, respectively. In possibility theory, a fuzzy set approach is common, where membership functions characterize the input uncertainty (Zadeh 1965). Even if a probability distribution is not available due to limited information, lower and upper bounds (intervals) on uncertain design variables are usually known. In this case, interval analysis (Moore 1966, Muhanna & Mullen 1999, Muhanna & Mullen 2001) and fuzzy set theory (Zadeh 1965) have been extensively used to characterize and propagate input uncertainty in order to calculate the interval of the uncertain output. An efficient method for reliability estimation with a combination of random and interval variables is presented in (Penmetsa & Grandhi 2002). However, it is not implemented in a design optimization framework. A few design optimization studies have been also reported, where some or all of the uncertain design variables are in interval form (Du, Sudjianto & Huang 2005, Gu et al. 1998, Rao & Cao 2002). Optimization with input ranges has also been studied under the term antioptimization (Elishakoff et al. 1994, Lombardi & Haftka 1998). Anti-optimization is used to describe the task of finding the “worst-case’’ scenario for a given problem. It solves a two-level (usually nested) optimization problem. The outer level performs the design optimization while the inner level performs the anti-optimization. The latter seeks the worst condition under the interval uncertainty. A decoupled approach is suggested in (Lombardi & Haftka 1998) where the design optimization alternates with the anti-optimization rather than nesting the two. It was mentioned that this method takes longer to converge and may not even converge at all if there is strong coupling between the interval design variables and the rest of the design variables. A “worst-case’’ scenario approach using interval variables has also been considered in multidisciplinary systems design (Du & Chen 2000, Gu et al. 1998). Very recently, possibility-based design algorithms have been proposed (Choi et al. 2004, Mourelatos & Zhou 2005) where a mean performance is optimized subject to possibilistic constraints. It was shown that more conservative results are obtained compared with the probability-based RBDO. A comprehensive comparison of probability and possibility theories is given in (Nikolaidis et al. 2004) for design under uncertainty.

N o n-p r o b a b i l i s t i c o p t i m i z a t i o n u s i n g p o s s i b i l i t y a n d e v i d e n c e t h e o r i e s

249

Evidence theory is more general than probability and possibility theories, even though the methodologies of uncertainty propagation are completely different (Bae et al. 2004, Oberkampf & Helton 2002). It can be used in design under uncertainty if limited, and even conflicting, information is provided from experts. Furthermore, the basic axioms of evidence theory allow to combine aleatory (random) and epistemic uncertainty in a straightforward way without any assumptions (Bae et al. 2004). Evidence theory however, has been barely explored in engineering design. One of the reasons may be its high computational cost due mainly to the discontinuous nature of uncertainty quantification. Evidence-based methods have been only recently used to propagate epistemic uncertainty (Bae et al. 2004) in large-scale engineering systems. Although a computationally efficient method is proposed in (Bae et al. 2004), the design issue is not addressed. We are aware of only one study which propagates epistemic uncertainty using evidence theory and also performs a design optimization (Agarwal et al. 2004). The optimum design is calculated for multidisciplinary systems under uncertainty using a trust region sequential approximate optimization method with surrogate models representing the uncertain measures as continuous functions. In this chapter, the possibility and evidence theories are used to account for uncertainty in design with incomplete information. The formal theories to handle uncertainty are first introduced using the theoretical fundamentals of fuzzy measures. The chapter highlights how the possibility theory can be used in design. A computationally efficient and accurate hybrid (global-local) optimization approach is presented for calculating the confidence level of “fuzzy’’ response, combining the advantages of the commonly used vertex and discretization methods. A possibility-based design optimization method is subsequently described where all design constraints are expressed possibilistically. The method gives a conservative solution compared with all conventional reliability-based designs obtained with different probability distributions. A general possibility-based design optimization method is also presented which handles a combination of random and possibilistic design variables. Furthermore, a sequential algorithm for possibility-based design optimization (SPDO) is presented. It decouples a double-loop PBDO process into a sequence of cycles composed of a deterministic design optimization followed by a set of worst-case possibility evaluation loops. The series of deterministic and possibility loops is repeated until convergence is achieved. A computationally efficient design optimization method is also described based on evidence theory, which can handle a mixture of epistemic and random uncertainties. The method can be used when limited and often conflicting, information is available from “expert’’ opinions. The algorithm quickly identifies the vicinity of the optimal point and the active constraints by moving a hyper-ellipse in the original design space, using an RBDO algorithm. Subsequently, a derivative-free optimizer calculates the evidence-based optimum, starting from the close-by RBDO optimum, considering only the identified active constraints. The computational cost is kept low by first moving to the vicinity of the optimum quickly and subsequently using local surrogate models of the active constraints only. The chapter is organized as follows. Section 2 gives an introduction to fuzzy measures. Section 3 describes the fundamentals of possibility theory based on fuzzy measures as well as some numerical methods for propagating non-probabilistic

250

Structural design optimization considering uncertainties

uncertainty, which are essential in possibility-based design. A detailed formulation of Possibility-Based Design Optimization (PBDO) where design constraints are satisfied possibilistically, is presented in section 4. Section 5 presents a detailed formulation of an Evidence-Based Design Optimization (EBDO) method and its implementation. Section 6 introduces the sequential algorithm for possibility-based design optimization. All principles are demonstrated with two examples in section 7. Results are compared among deterministic optimization, RBDO, PBDO, EBDO and SPDO. Finally, a summary and conclusions are given in section 8.

2 Fuzzy measures The evidence and possibility theories are based on the mathematical foundation of fuzzy measures which provide the foundation of fuzzy set theory. Before we introduce the basics of fuzzy measures, it is helpful to review the used notation on set representation. A universe X represents the entire collection of elements having the same characteristics. The individual elements in the universe X are denoted by x, which are usually called singletons. A set A is a collection of some elements of X. All possible sets of X constitute a special set called the power set ℘(X). A fuzzy measure is defined by a function g: ℘(X) → [0, 1] which assigns to each crisp (Ross 1995) subset of X a number in the unit interval [0, 1]. The assigned number in the unit interval for a subset A ∈ ℘(X), denoted by g(A), represents the degree of available evidence or belief that a given element of X belongs to the subset A. In order to qualify as a fuzzy measure, the function g must obey the following three axioms: Axiom 1 (boundary conditions): g(Ø) = 0 and g(X) = 1. Axiom 2 (monotonicity): For every A, B ∈ ℘(X), if A ⊆ B, then g(A) ≤ g(B). Axiom 3 (continuity): For every sequence (Ai ∈ ℘(X), i = 1, 2,…) of subsets of ℘(X), if either A1 ⊆ Ai ⊆…or A1 ⊇ A2 ⊇ …(i.e., the sequence is monotonic), then lim g(Ai ) = g( lim Ai ). i→∞

i→∞

A belief measure is a function Bel: ℘(X) → [0, 1] which satisfies the three axioms of fuzzy measures and the following additional axiom (Klir & Filger 1988): Bel(A1 ∪ A2 ) ≥ Bel(A1 ) + Bel(A2 ) − Bel(A1 ∩ A2 )

(1)

The axiom (1) can be expanded for more than two sets. For A ∈ ℘(X), Bel(A) is interpreted as the degree of belief, based on available evidence, that a given element of X belongs to the set A. A plausibility measure is a function Pl : ℘(X) ⇒ [0, 1]

(2)

which satisfies the three axioms of fuzzy measures and the following additional axiom (Klir & Filger 1988) Pl(A1 ∩ A2 ) ≤ Pl(A1 ) + Pl(A2 ) − Pl(A1 ∪ A2 )

(3)

N o n-p r o b a b i l i s t i c o p t i m i z a t i o n u s i n g p o s s i b i l i t y a n d e v i d e n c e t h e o r i e s

251

Every belief measure and its dual plausibility measure can be expressed with respect to the non-negative function m : ℘(X) ⇒ [0, 1]

(4)

such that m(Ø) = 0 and  m(A) = 1

(5)

A∈℘(X)

The function m is called Basic Probability Assignment (BPA) due to the resemblance of Eq. (5) with a similar equation for probability distributions. The basic probability assignment m(A) is interpreted either as the degree of evidence supporting the claim that a specific element of X belongs to the set A or as the degree to which we believe that such a claim is warranted. At this point, it should be noted that the BPA is very different from the probability distribution function. Basic probability assignments are defined on sets of the power set (i.e., on A ∈ ℘ (X)), whereas the probability distribution functions are defined on the singletons x of the power set (i.e., on x ∈ ℘ (X)). Every set A ∈ ℘ (X) for which m(A) > 0 is called a focal element of m. Focal elements are subsets of X on which the available evidence focuses; i.e. available evidence exists. Given a BPA m, a belief measure and a plausibility measure are uniquely determined by  Bel(A) = m(B) (6) B⊆A

Pl(A) =



m(B)

(7)

B∩A =0

which are applicable for all A ∈ ℘(X). In Eq. (6), Bel(A) represents the total evidence or belief that the element belongs to A as well as to various subsets of A. The Pl(A) in Eq. (7) represents not only the total evidence or belief that the element in question belongs to set A or to any of its subsets but also the additional evidence or belief associated with sets that overlap with A. Therefore, we have Pl(A) ≥ Bel(A)

(8)

Probability theory is a subset of evidence theory. When the additional axiom of belief measures (see Eq. (1)) is replaced with the stronger axiom Bel(A ∪ B) = Bel(A) + Bel(B) where A ∩ B = Ø

(9)

we obtain a special type of belief measures which are the classical probability measures. In this case, the right hand sides of Eqs (6) and (7) become equal and therefore,   Bel(A) = Pl(A) = m(x) = p(x) (10) x∈A

x∈A

252

Structural design optimization considering uncertainties

for all A ∈ ℘(X), where p(x) is the classical probability distribution function (PDF). Note that the BPA m(x) is equal to p(x). Therefore with evidence theory, we can simultaneously handle a mixture of input parameters. Some of the inputs can be described probabilistically (random uncertainty) and some can be described through expert opinions (epistemic uncertainty with incomplete data). In the second case, the range of each input parameter will be discretized using a finite number of intervals. The BPA value for each interval must be equal to the PDF area within the interval. It should be noted that according to evidence theory, the Bel(A) and Pl(A) bracket the true probability P(A) (Klir & Filger 1988), i.e. Bel(A) ≤ P(A) ≤ Pl(A)

(11)

Evidence obtained from independent sources or experts must be combined. If the BPA’s m1 and m2 express evidence from two experts, the combined evidence m can be calculated by the following Dempster’s rule of combining (Sentz & Ferson 2002)  m(A) =

m1 (B)m2 (C)

B∩C=A

1−K

for A = 0

(12)

where K=



m1 (B)m2 (C)

(13)

B∩C=0

represents the conflict between the two independent experts. Dempster’s rule filters out any conflict, or contradiction among the provided evidence, by normalizing with the complementary degree of conflict. It is usually appropriate for relatively small amounts of conflict where there is some consistency or sufficient agreement among the opinions of the experts. Yager (Yager et al. 1994) has proposed an alternative rule of combination where all degrees of contradiction are attributed to total ignorance. Other rules of combining can be found in (Sentz & Ferson 2002). The possibility theory is a subcase of the general evidence theory. It can be used to characterize epistemic uncertainty, when incomplete data is available. It applies only when there is no conflict in the provided body of evidence. In such a case, the focal elements of the body of evidence are nested and the associated belief and plausibility measures are called consonant. In contrary, when there is conflicting evidence, the belief and plausibility measures are dissonant. A family of subsets of the universal set is nested if they can be ordered in such a way that each is contained within the next. Thus, A1 ⊂ A2 ⊂ · · · ⊂ An are nested sets. Consonant belief and plausibility measures are usually known as necessity measures n and possibility measures π, respectively. Therefore, if there is no conflicting information, n(A) = Bel(A) and π(A) = Pl(A). The necessity and possibility are dual measures, related by n(A) = 1 − π(A) where A is the complement of set A.

(14)

N o n-p r o b a b i l i s t i c o p t i m i z a t i o n u s i n g p o s s i b i l i t y a n d e v i d e n c e t h e o r i e s

253

3 Fundamentals of possibility theory This section highlights the fundamentals of possibility theory as it was originally introduced in the context of fuzzy set theory (Zadeh 1978). In the fuzzy set approach to possibility theory, focal elements are represented by a-cuts of the associated fuzzy set. Focal elements are subsets that are assigned nonzero degrees of evidence. The possibility theory can be used to bracket the true probability based on the fuzzy set approach at various confidence intervals (a-cuts). The advantage of this is that as the design progresses and the confidence level on the input parameter bounds increases, the design need not be reevaluated to obtain the new bounds of the response. Similarly to the probability measures, which are represented by the probability distribution functions, the possibility measures can be represented by the possibility distribution function r : X ⇒ [0, 1] such that π(A) = max r(x)

(15)

x∈A

It can be shown that possibility measures are formally equivalent to fuzzy sets. In this equivalence, the membership grade of an element x corresponds to the plausibility of the singleton consisting of that x. Therefore, a consonant belief structure is equivalent to a fuzzy set of X. A fuzzy set is an imprecisely defined set that does not have a crisp boundary. It provides instead, a gradual transition from “belonging’’ to “not belonging’’ to the set. A function can be defined such that the values assigned to the elements of the set are within a specified range and indicate the membership grade of these elements in the set. Larger values denote higher degrees of set membership. Such a function is called a membership function and the set defined by it a fuzzy set. The membership function µA by which a fuzzy set A is usually defined has the form µA : X → [0, 1] where [0, 1] denotes the interval of real numbers from 0 to 1, inclusive. Given a fuzzy subset A of X with membership function µA , Zadeh (Zadeh 1978) defines a possibility distribution function r associated with A as numerically equal to µA , i.e. r(x) = µA (x) for all x ∈ X. Then, he defines the corresponding possibility measure π as π(A) = sup r(x)

for each A ∈ ℘(X)

(16)

x∈A

Eq. (16) is equivalent to Eq. (15) when X is finite. In the fuzzy set approach to possibility theory, focal elements are represented by a-cuts of the associated fuzzy set. For the remaining of this discussion, we will follow the fuzzy set approach to possibility theory. Eq. (11) states that the true probability is bracketed by the belief and plausibility measures. If we know the possibility distribution function µY (y) of the response Y, then the true probability P(Y) can be also bracketed as n(Y) ≤ P(Y) ≤ π(Y)

(17)

where the necessity n(Y) and possibility π(Y) measures are calculated from Eqs (14) and (16), respectively. The “extension principle’’ (Klir & Filger 1988, Ross 1995,

254

Structural design optimization considering uncertainties

x

(x)

1.0 ␣

0.0

XaL

XN

dL (a)

XaU

X

dU (a)

Figure 10.1 Triangular possibility distribution for a fuzzy variable.

Yager et al. 1994) is used to calculate the possibility distribution function µY (y) of the response. 3.1

F uz z i f i c a ti o n pr o c es s and e xt e ns io n p r i n ci p l e

The process of quantifying a fuzzy variable is known as fuzzification. If any of the input variables is imprecise, it is considered fuzzy and must be therefore, fuzzified in order for the uncertainty to be propagated using fuzzy calculus. The fuzzification is done by constructing a possibility distribution, or membership function, for each imprecise (fuzzy) variable. Details can be found in (Ross 1995). The membership function takes values in the [0, 1] interval. Here, we use convex normal possibility distributions to characterize the fuzzy variables. An example of a convex normal triangular possibility distribution is shown in Fig. 10.1. The point for which the possibility is equal to one is called normal point. The possibility distribution is convex since it is strictly decreasing to the left and right of the normal point. At each confidence level, or a-cut, a set Xa is defined as Xa = {x : xaL ≤ x ≤ xaR , a ∈ [0, 1]}

(18a)

which is a monotonically decreasing function of a; i.e. a1 > a2 ⇒ Xa1 ⊂ Xa2

for every a1 , a2 ∈ [0, 1]

(18b)

Due to the convexity of the possibility distribution function, all sets generated at different a-cuts are nested according to Eq. (18b). Therefore, the convexity and normality of the possibility distribution function satisfies the basic requirement of nested sets (no conflicting evidence) in possibility theory. After the fuzzification of the imprecise input variables, the “extension principle’’ is used to propagate the epistemic uncertainty through the transfer function in order to calculate the fuzzy response. The “extension principle’’ calculates the possibility distribution of the fuzzy response from the possibility distributions of the fuzzy input variables. In particular, given the transfer function y = f (x), where the output y depends

N o n-p r o b a b i l i s t i c o p t i m i z a t i o n u s i n g p o s s i b i l i t y a n d e v i d e n c e t h e o r i e s

255

on the N independent fuzzy inputs x = {x1 , . . . , xN }, the “extension principle’’ states that the possibility distribution µY of the output is given by µY [y = f (x)] = sup {min [µX (f (xj ))]} j

y

(19)

where “sup’’ denotes the suprenum operator that gives the least upper bound. The above equation can be interpreted as follows. For a crisp value of the output y, there may exist more than one combination of crisp values of input variables x resulting in the same output. The possibility of each combination is given by the smallest possibility value for all fuzzy input variables. The possibility that y = f (x), is given by the maximum possibility for all these combinations. Note that in probability theory, the probability of an outcome is equal to the product of the probabilities of the constituent events. In fuzzy set theory however, the possibility of an outcome is equal to the minimum possibility of the constituent events. If the outcome can be reached in many ways, then the outcome probability, in probability theory, is given by the sum of the probabilities of all the ways. In fuzzy theory, the possibility of the outcome is given by the maximum possibility of all the possibilities (Ross 1995). The direct (“brute force’’) solution of Eq. (19) is practically intractable except for simple cases involving one or two fuzzy variables. The computational effort increases exponentially with increasing number of fuzzy input variables. For this reason, approximate numerical techniques have been proposed, among which the discretization method (Akpan et al. 2002) and the vertex method (Penmetsa & Grandhi 2002) are the most popular ones. In the discretization method, the domain of each fuzzy variable i; 1 ≤ i ≤ N is discretized with Mi discrete values at each a-cut. Then the output y is evaluated at all N 9 Mi for each a-cut. Subsequently, Eq. (19) is used to calculate possible combinations i=1

the possibility distribution of the output. The range of the output is defined by the minimum and maximum response from all combinations. Although this method can be very accurate, the associated computational cost is practically prohibitive. In the vertex method, all the binary combinations of only the extreme values of the fuzzy variables at an a-cut are fed into the deterministic transfer function. The bounds of the fuzzy response are then obtained at the a-cut, by choosing the maximum and minimum responses. The procedure is repeated for all a-cuts of interest. The method has the potential to give accurate bounds of the response based on the bounded input. However, when the transfer function exhibits minima or maxima within the domain defined by the extreme values of the input variables, the vertex method is inaccurate. This is due to the fact that the function is evaluated only at the binary combinations of the input variable bounds. For a problem with N fuzzy input variables, the required number of function evaluations for the vertex method is A ∗ 2N , where A is the number of a-cuts. In general, the vertex method is computationally more efficient compared with the discretization method. However, the required computational effort grows exponentially with the number of input fuzzy variables (Ross 1995). For this reason, most of the reported applications are restricted to very few fuzzy variables (Chen & Rao 1997, Mullen & Muhanna 1999, Rao & Sawyer 1995).

256

Structural design optimization considering uncertainties

A hybrid (global-local) optimization method has been reported in (Mourelatos & Zhou 2005), which ensures computational efficiency without loss of accuracy. An optimization algorithm is used to calculate the minimum and maximum values of the response at each a-cut. Because the global minimum and maximum values of the response are needed, a derivative free, global optimizer called DIRECT (DIvisions of RECTangles), is used in order to avoid being trapped at a local optimum and obtain therefore, an inaccurate solution. DIRECT is a modification of the standard Lipschitzian approach that eliminates the need to specify a Lipschitz constant (Jones et al. 1993). Although global optimizers may get close to the global optimum quickly, it takes them longer to achieve a high degree of accuracy because they usually have a slow rate of convergence. This suggests that the best performance can be obtained by combining DIRECT with a gradient-based local optimizer in a hybrid approach. In this work, DIRECT is first used, followed by a local optimizer based on Sequential Quadratic Programming (SQP). DIRECT provides a converged global optimum based on “loose’’ convergence criteria. Subsequently, the DIRECT solution is used as starting point for SQP, which identifies the optimum accurately and efficiently. 3.2 A m a th em at ic al e xample The following two-variable, six-hump camel function (Wang 2003) is used y(x1 , x2 ) = 4x21 − 2.1x41 +

1 6 x + x1 x2 − 4x22 + 4x42 , 3 1

x1,2 ∈ [−2, 2]

to illustrate the accuracy and efficiency of the hybrid optimization method of the previous section and compare it with the vertex and discretization methods. For demonstration reasons, the following simple triangular membership functions are used for the two input variables x1 and x2 ⎧ 0 ≤ xi ≤ 2 ⎪ ⎨ − xi + 1, 2 µXi (xi ) = ⎪ ⎩ xi + 1, −2 ≤ x ≤ 0 i 2

i = 1, 2

Fig. 10.2 shows the contour plot of the six hump camel function. The H’s indicate all extreme points. Points H2 and H5 with coordinates (0.0898, −0.7127) and (−0.0898, 0.7127) respectively, are two global optima with an equal function value of ymin = −1.0316. The calculated membership functions of the response y using the vertex, discretization and hybrid optimization methods are plotted in Fig. 10.3. Ten a-cuts are used for all three methods. For the discretization method, the range of each input fuzzy variable, at each a-cut, is equally split in 15 divisions. It is known that if the input membership functions are convex normal, the response membership function must also be convex normal. The justification is that when the input uncertainty increases (low a-cut values), the uncertainty of the response must remain the same or increase. As shown in Fig. 10.3, the response membership function obtained by the vertex method is not convex and therefore, it is wrong.

N o n-p r o b a b i l i s t i c o p t i m i z a t i o n u s i n g p o s s i b i l i t y a n d e v i d e n c e t h e o r i e s

257

2 1.5 1

X2

0.5 0

0.5 1 1.5 2 2

1.5

1

0.5

0 X1

1

0.5

1.5

2

Figure 10.2 Contour plot for mathematical example. 1 0.9

Number of alpha-cuts
10

0.8 0.7

Vertex method

0.6 ␮Y (y)

Discretization method 0.5 Optimization method

0.4 0.3 0.2 0.1 0 10

0

10

20

30

40

50

60

y

Figure 10.3 Response membership function for mathematical example.

As explained in section 3.1, the discretization method evaluates the function not only at the upper and lower limits of the input variables at each alpha cut but also between the bounds. Thus, it can capture the extreme points that might be present in between the upper and lower bounds. At each alpha cut, all combinations are obtained and the minimum and maximum response values are calculated in order to get the response membership function. It is clear that the response becomes more accurate as the number of divisions per alpha cut increases. As shown in Fig. 10.3,

258

Structural design optimization considering uncertainties Table 10.1 Accuracy and efficiency comparison of vertex, discretization and hybrid optimization methods.

Lower Bound Upper Bound No. of F.E.

Vertex

Discretization

Hybrid Optimization

47.73 55.73 4

−1.01 55.73 256

−1.03 55.73 140

the response membership function calculated with the discretization method, is convex and normal. The uncertainty decreases as the level of confidence increases (increasing a-cut values). The major disadvantage of this method is that as the number of design variables increases and the number of divisions per a-cut also increases, the method becomes computationally very expensive. In this example, the number of a-cuts is 10 and the number of divisions per a-cut is 15. Therefore, the number of function evaluations is 10 ∗ (15 + 1)2 = 2560. The response membership function of the six hump camel function is also calculated using the proposed hybrid optimization method. The result is identical with that obtained with the discretization method (see Fig. 10.3). Table 10.1 summarizes the lower and upper bound values of the response at the zero a-cut, as calculated by the vertex, discretization and hybrid optimization methods. The vertex method is very efficient but inaccurate. The hybrid optimization method however, has the same accuracy with the “brute force’’ discretization method but it is much more efficient.

4 Possibility-based design optimization In deterministic design optimization, an objective function is minimized subject to satisfying a set of constraints. In Reliability-Based Design Optimization (RBDO), where all design variables are characterized probabilistically, an objective function is usually minimized subject to the probability of satisfying each constraint being greater than a specified high reliability level. In this section, a methodology is presented on how to use possibility theory in design. We will show that the possibility-based design is conservative compared with all RBDO designs obtained with different probability distributions. In RBDO, some optimality is usually sacrificed in order to accommodate the random uncertainty. The possibility-based design sacrifices a little more optimality in order to accommodate the lack of probability distribution information. It therefore, encompasses all RDBO designs obtained with different distributions. According to Eq. (11), the probability P(A) of event A is bracketed by the belief Bel(A) and plausibility Pl(A); i.e. Bel(A) ≤ P(A) ≤ Pl(A). We have also mentioned that for consonant (no conflicting evidence) belief structures, the plausibility measures are equal to the possibility measures, resulting in η(A) ≤ P(A) ≤ π(A), where η and π are the necessity and possibility measures, respectively (see Eq. 17). This means that the possibility π(A) provides an upper bound to the probability P(A). From the design point of view, we can thus conclude (Klir & Filger 1988, Ross 1995, Zadeh 1978) that what is possible may not be probable, and what is impossible is also improbable.

N o n-p r o b a b i l i s t i c o p t i m i z a t i o n u s i n g p o s s i b i l i t y a n d e v i d e n c e t h e o r i e s

259

mG(g)

1

a a1

gmin

a gmin

gN

a gmax

gmax

g

g0

Figure 10.4 Used notation in possibility-based design optimization.

Note that for an impossible event A, the possibility π(A) is zero. If we therefore, make sure that the possibility of violating a constraint is zero, then the probability of violating the same constraint will be also zero. If feasibility of a constraint g is expressed with the positive null form g ≥ 0, the constraint is always satisfied if π(g ≤ 0) = 0

(20)

The possibility π in Eq. (20) is calculated using Eq. (16). Fig. 10.4 shows the membership function µG (g) of constraint g. The possibility of set A = {g: gmin ≤ α α α , α ∈ [0, 1]} is π(A) = α and the possibility of set B = {g: gmin ≤ g ≤ gmax , g ≤ gmin α ∈ [0, 1]} is π(B) = 1. Similarly, the possibility of constraint violation is π(g ≤ 0) = α1 . Eq. (20) can be relaxed as π(g ≤ 0) ≤ α

(21)

where the a-cut level is small; i.e. α << 1. Based on Fig. 10.4, the relation (21) is satisfied if α gmin ≥0

(22)

α where gmin is the global minimum of g at the a-cut. Eq. (22) is analogous to the R-percentile formulation (Tu et al. 1999) of a probabilistic constraint in RBDO. The α = 0. possibilistic constraint of Eqs (21) or (22) becomes active if gmax Based on this discussion, a possibility-based design optimization (PBDO) problem can be formulated as

min f (d, xN , pN ) d,xN

subject to

π(gi (d, X, P) ≤ 0) ≤ α,

i = 1, . . . , n

dL ≤ d ≤ dU , xL ≤ xN ≤ xU

(23)

260

Structural design optimization considering uncertainties

where d ∈ Rk is the vector of deterministic design variables, X ∈ Rm is the vector of possibilistic design variables, P ∈ Rq is the vector of possibilistic design parameters and xN and pN are the normal point vectors for the possibilistic design variables and parameters, respectively. According to the used notation, a bold letter indicates a vector, an upper case letter indicates a possibilistic variable or parameter and a lower case letter indicates a deterministic variable or a realization of a possibilistic variable or parameter. Feasibility of the ith deterministic constraint is expressed with the positive null form gi ≥ 0. The possibilistic design variables are represented with convex normal possibility distributions (membership functions). Note that they may not be necessarily triangular. The superscript N denotes the normal point of each distribution where the membership function value is equal to one. Subscripts L and U denote lower and upper bounds, respectively. In PBDO, we will assume that the membership functions of the possibilistic design variables have a constant shape and that their normal points are design variables moving within predetermined bounds. This is analogous to RBDO where the PDF of each random design variable stays constant while its mean value is a design variable. Based on Eq. (22), the PBDO formulation (23) is equivalent to min f (d, xN , pN ) d,xN

subject to giαmin ≥ 0

i = 1, . . . , n

(24)

dL ≤ d ≤ dU , xL ≤ xN ≤ xU The PBDO formulation (23) or (24) is a double-loop optimization problem where an optimization is performed (inner loop) when the design optimization (outer loop) calls for a possibilistic constraint evaluation. It should be noted that the PBDO optimum at a = 1 coincides with the deterministic optimum.

4.1

PBD O wi t h a c o mb inat io n o f r and o m a n d possi b i l i st ic var iab les

Reliability-based design optimization (RBDO) provides optimum designs in the presence of only random (or aleatory) uncertainty (Liang et al. 2007, Tu et al. 1999, Wu et al. 2001). A typical RBDO problem is formulated as (Liang et al. 2007) min f (d, µY , µZ ) d,µY

subject to P(gi (d, Y, Z) ≥ 0) ≥ Ri = 1 − pfi ,

i = 1, . . . , n

(25)

dL ≤ d ≤ dU , µLY ≤ µY ≤ µU Y where Y ∈ R is the vector of random design variables and Z ∈ Rr is the vector of random design parameters. For a variety of practical applications however, there may not be enough information to characterize all design variables and parameters probabilistically. A subset

N o n-p r o b a b i l i s t i c o p t i m i z a t i o n u s i n g p o s s i b i l i t y a n d e v i d e n c e t h e o r i e s

261

of them can be therefore, characterized possibilistically using membership functions. A possibility-based design optimization problem with a combination of random and possibilistic (or fuzzy) variables can be formulated as min f (d, µY , µZ , xN , pN )

d,xN ,µY

subject to

giαmin ≥ 0,

i = 1, . . . , n

dL ≤ d ≤ dU

(26)

N µLY ≤ µY ≤ µU Y , xL ≤ x ≤ xU

with giαmin = min gi (d, U, x, p), x,U,p

subject to

i = 1, . . . , n,

βi = βti xLα (xN ) ≤ x ≤ xUα (xN ), pαL ≤ p ≤ pαU

where βt is the target reliability index. Note that xLα and xUα are the lower and upper limits of X at an a-cut. Problem (26) represents a double-loop performance measure approach (PMA) optimization sequence. The design optimization of the outer loop calls a series of possibilistic constraints in the inner loop. Each possibilistic constraint is in general, a global optimization problem. The inner loop calculates the minimum value of each limit state function considering that the realizations of possibilistic variables X vary between xLα and xUα , the realizations of possibilistic parameters P vary between pαL and pαU and the reliability requirement βi = βti is satisfied. It therefore, calculates the worst-case scenario.

5 Evidence-based design optimization (EBDO) In this section, a methodology is presented on how to use evidence theory in design. We will show that the evidence theory-based design is more conservative compared with all RBDO designs obtained with different probability distributions and less conservative compared with the PBDO design. If feasibility of a constraint g is expressed with the non-negative null form g ≥ 0, we have shown that Bel(g ≥ 0) ≤ P(g ≥ 0) ≤ Pl(g ≥ 0) where P(g ≥ 0) is the probability of constraint satisfaction. Therefore, P(g < 0) ≤ pf is satisfied if Pl(g < 0) ≤ pf

(27)

where pf is the probability of failure which is usually a small prescribed value. The above statement is equivalent to P(g ≥ 0) ≥ R is satisfied if Bel(g ≥ 0) ≥ R where R = 1 − pf is the corresponding reliability level.

(28)

262

Structural design optimization considering uncertainties

Hence, an evidence theory-based design optimization (EBDO) problem can be formulated as min f (d, xN , pN ) d,xN

subject to Pl(gi (d, X, P) < 0) ≤ pfi ,

i = 1, . . . , n

(29)

N dL ≤ d ≤ dU , xLN ≤ xN ≤ xU

where X ∈ Rm and P ∈ Rq are the vectors of uncertain design variables and parameters. The superscript “N’’ indicates nominal value of uncertain variables or parameters. The uncertainty is provided by expert opinions. It should be noted that the plausibility measure is used instead of the equivalent belief measure, in Problem (29). The reason is that at the optimum, the failure domain for each active constraint is usually much smaller than the safe domain over the frame of discernment (FD) (domain of all focal elements with nonzero combined BPA; see next section). As a result, the computation of the plausibility of failure is much more efficient than the computation of the belief of safe region. 5.1 Assessi ng B el and Pl wit h d emps t e r-s h a f e r t h e o r y Evidence theory can quantify epistemic uncertainty, even when the experts provide conflicting evidence. This section shows how to propagate epistemic uncertainty through a given model (transfer function) which is necessary in calculating the plausibility of constraint violation in Problem (29). The uncertainty propagation will be illustrated using the following simple transfer function y = f (a, b)

(30)

where a ∈ A, b ∈ B are two independent input parameters and y is the output. The combined BPA’s for both a and b are obtained from Dempster’s rule of combining of Eq. (12) if multiple experts have provided evidence for either a or b. With combined information for each input parameter, we define a vector c = aci , bcj , needed to calculate the output y as C = A × B = {c = [aci , bcj ], aci ∈ A, bcj ∈ B}

(31)

where subscript c stands for “combined’’ and i, j indicate focal elements. Taking advantage of assumed parameter independency, the BPA for c is mc (hij ) = m(aci )m(bcj )

(32)

where hij = [aci , bcj ] and aci , bcj denote intervals such that a ∈ aci and b ∈ bcj . Equation (32) can be used to calculate the combined BPA structure for the entire domain C. For every (a, b) ∈ c |c ∈ C, needed to evaluate the output y, the combined BPA mc is used. A representative combined BPA structure is shown in Fig. 10.5. The Cartesian product C of Eq. (31) is also called frame of discernment (FD) in the literature. It consists of all focal elements (rectangles in Fig. 10.5 with nonzero combined BPA) and can be viewed as the finite sample space in probability theory.

N o n-p r o b a b i l i s t i c o p t i m i z a t i o n u s i n g p o s s i b i l i t y a n d e v i d e n c e t h e o r i e s

263

BPA 0.8 0.6

b

0.4

0.6 0.4

0.2 0 0

0.2

0.2 a

0.4 0.6

Figure 10.5 Representative BPA structure for two parameters a and b.

gmax

g0 gmin g0

g0

g 0 g0

g
0 g min

gmax

g
0

gmax g
0

gmin g0

Figure 10.6 Schematic illustration of focal element contribution to belief and plausibility measures.

Let a domain F being defined as F = {g : g = f (a, b) − y0 > 0, (a, b) ∈ c, c = [ac , bc ] ⊂ C}

(33)

where y0 is a specified value. According to evidence theory, Bel(F) ≤ pf ≤ Pl(F)

(34)

where pf = P(g > 0) is the true probability. The Bel (F) and Pl (F) are calculated using Eqs (6) and (7) where set A is equal to set F of Eq. (33) and B is a rectangular domain (focal element) such that B ⊆ A for Eq. (6) and B ∩ A = 0 for Eq. (7). B ⊆ A means that the focal element must be entirely within the domain g > 0 and B ∩ A = 0 means that the focal element must be entirely or partially within the domain g > 0 (see Fig. 10.6). In order to identify if a focal element B satisfies B ⊆ A or B ∩ A = 0, the following minimum and maximum values of g must be calculated [gmin , gmax ] = [min g(x), max g(x)] x

x

(35)

for xL ≤ x ≤ xU where (xL , xU ) defines the focal element domain. For monotonic functions, the vertex method (Penmetsa & Grandhi 2002) can be used to calculate the minimum and maximum values in Eq. (35) by simply identifying the minimum and

264

Structural design optimization considering uncertainties

Hyper-ellipse

X1

Initial design point g1 (x1, x2)
0 Feasible region Frame of discernment g2(x1, x2)
0 Objective reduces

B

MPP for g1
0

EBDO optimum Deterministic optimum X2

Figure 10.7 Geometrical interpretation of the EBDO algorithm.

maximum values among all vertices of the focal element domain. For non-monotonic functions, a global optimizer is needed. If for a focal element, gmin and gmax are both positive, the focal element will contribute to the calculation of belief and plausibility. On the other hand, if gmin and gmax are both negative, the focal element will not contribute to the calculation of belief or plausibility. If however, gmin is negative and gmax is positive, the focal element will not contribute to the belief but it will contribute to the plausibility calculation. This is shown schematically in Fig. 10.6. In summary the following tasks are performed in order to calculate the belief and plausibility of the failure region: 1.

2.

3. 4.

For each input parameter, combine the evidence from the experts by combining the individual BPA’s from each expert using Dempster’s rule of combining (Eq. (12)). Construct the BPA structure for the m-dimensional frame of discernment, where m is the number of input parameters. Assuming independent input parameters, Eq. (32) is used. Identify the failure region space (set F of Eq. (33)). Use Eqs (6) and (7) to calculate the belief and plausibility measures of the failure region. The failure region must be identified only within the frame of discernment. The true probability of failure is bracketed according to Eq. (34).

5.2 Im pl em ent at io n o f t he EB DO alg o r i t h m A computationally efficient solution of Problem (29) is presented here. As a geometrical interpretation of it, we can view the design point (d, x) moving within the feasible domain so that the objective f is minimized (see Fig. 10.7). If the entire FD is in the

N o n-p r o b a b i l i s t i c o p t i m i z a t i o n u s i n g p o s s i b i l i t y a n d e v i d e n c e t h e o r i e s

265

feasible domain, the constraints are satisfied and are inactive. A constraint becomes active if part of the FD is in the “failure’’ region so that the plausibility of constraint violation is equal to pf . In general, Problem (29) represents movement of a hyper-cube (FD) within the feasible domain. In order to save computational effort, the bulk of the FD movement, from the initial design point to the vicinity of the optimal point (point B of Fig. 10.7), can be achieved by moving a hyper-ellipse which contains the FD. The center of the hyper-ellipse is the “approximate’’ design point and each axis is arbitrarily taken equal to three times the standard deviation of a hypothetical normal distribution. This assumes that each dimension of the FD hyper-cube is equal to six times the standard deviation of the hypothetical normal distribution. The hyper-ellipse can be easily moved in the design space by solving a RBDO problem. The RBDO optimum (point B of Fig. 10.7) is in the vicinity of the solution of Problem (29) (EBDO optimum). The RBDO solution also identifies all active constraints and their corresponding most probable points (MPP’s). The maximal possibility search algorithm (Choi et al. 2004) can also be used to move the FD hyper-cube in the feasible domain. It should be noted that the 3-sigma axes hyper-ellipse is arbitrary. The size of the hyper-ellipse is not however, crucial because it is only used to calculate the initial point (point B of Fig. 10.7) of the EBDO algorithm. The latter calculates the true EBDO optimum accurately. From our experience, a 3 to 4-σ size works fine. At this point, we generate a local response surface of each active constraint around its MPP. In this work, the Cross-Validated Moving Least Squares (CVMLS) (Tu & Jones 2003) method is used based on an Optimum Symmetric Latin Hypercube (OSLH) (Ye et al. 2000) “space-filling’’ sampling. A derivative-free optimizer calculates the EBDO optimum. It uses as initial point the previously calculated RBDO optimum which is close to the EBDO optimum. Problem (29) is solved, considering only the identified active constraints. For the calculation of the plausibility of failure Pl(g < 0) of each active constraint, an algorithm presented in (Mourelatos & Zhou 2005) is used. It identifies all focal elements which contribute to the plausibility of failure. The computational effort is significantly reduced because accurate local response surfaces are used for the active constraints. The cost can be much higher if the optimization algorithm evaluates the actual active constraints instead of their efficient surrogates (response surfaces). It should be noted that a derivative-free optimizer is needed due to the discontinuous nature of the combined BPA structure. The DIRECT derivative-free, global optimizer is used (Jones et al. 1993).

6 A sequential algorithm for possibility-based design optimization (SPDO) The computational effort of the double-loop approach of Problem (26) may be prohibitive especially for large-scale applications. For this reason, a Sequential algorithm for Possibility-based Design Optimization (SPDO) method is proposed in this section. It decouples the double-loop PBDO process of Problem (26) by using successive cycles composed of a deterministic design optimization followed by a set of possibilistic evaluation loops. In each cycle, the deterministic optimization and the possibilistic

266

Structural design optimization considering uncertainties

xN2, X2

Deterministic constraint Possibilistic constraint g(xN1, xN2)
0

p[g(x1, x2 )  0]  a

cut (xN1, xN2) Shift to:

SP2

g(xN1

– SP1, xN2 – SP2)

SP1 (x1,worst , x2,worst )

xN1,X1

Figure 10.8 Shifting of feasible domain for only uncertain variables.

evaluations are decoupled. The latter are conducted after the deterministic optimization. If at the deterministic optimum of a cycle, a particular possibilistic constraint is violated, a “shifting’’ vector is determined which moves the constraint boundary within the deterministic feasible domain. The “shifted’’ constraints are then used to perform a new deterministic design optimization. The series of deterministic and possibilistic evaluation loops continues until convergence is achieved; i.e. the objective function is minimized without violating any possibilistic constraint. At convergence the magnitude of the “shifting’’ vector is zero. The idea of using a “shifting’’ vector has been originally proposed in (Du & Chen 2004).

6.1 SPD O wi th o nly po s s ib ilis t ic var iabl e s Before we present the proposed SPDO algorithm for a combination of possibilistic and random variables, we will introduce the approach when there are not any random variables. For comprehension purposes, we initially assume without loss of generality, that there are not any deterministic design variables or possibilistic design parameters. In this case, Problem (24) reduces to min f (xN ) xN

subject to giαmin ≥ 0

i = 1, . . . , n

(36)

xL ≤ xN ≤ xU For illustration purposes, Fig. 10.8 gives a geometrical interpretation using only two possibilistic variables X1 and X2 .

N o n-p r o b a b i l i s t i c o p t i m i z a t i o n u s i n g p o s s i b i l i t y a n d e v i d e n c e t h e o r i e s

267

N The normal points of the two possibilistic variables are denoted by xN 1 and x2 , respectively. The deterministic and possibilistic constraint boundaries are denoted N by g(xN 1 , x2 ) = 0 and π[g(x1 , x2 ) ≤ 0] ≤ α, respectively. Because the possibility-based design is more conservative than the deterministic design, its feasible region is reduced compared with that of the deterministic design. Problem (36) is solved using a sequence of cycles. Each cycle is composed of a deterministic optimization followed by a possibilistic evaluation. During the first cycle, the following deterministic problem is solved

min f (xN ) xN

subject to

gi (xN ) ≥ 0 i = 1, . . . , n xL ≤ xN ≤ xU

N The optimal point xN = (xN 1 , x2 ) is on the boundary of the active deterministic constraints. A possibilistic evaluation at the desired a-cut, is then implemented N for each constraint at xN = (xN 1 , x2 ) in order to determine the worst-case point xworst = (x1,worst , x2,worst ). The following problem is solved for the ith constraint,

min gi (x) x

subject to

xN − δL (α) ≤ x ≤ xN + δU (α)

where δL (α) and δU (α) are the lower and upper bounds of x at the desired a-cut (see Fig. 10.1). If the solution xi,worst (worst-case point) of the above problem is deterministically infeasible, we must force it at least onto the deterministic constraint in order to ensure feasibility of the ith possibilistic constraint. This can be achieved by using a “shifting’’ vector SP = (SP1 , SP2 ) similarly to (Du & Chen 2004). In this case, the deterministic optimization of the next cycle is formulated as min f (xN ) xN

subject to

N gi (xN 1 − SP1 , x2 − SP2 ) ≥ 0,

i = 1, . . . , n

xL ≤ xN ≤ xU For multiple possibilistic constraints, the boundary of each constraint is shifted inside the deterministic feasible region by the distance between the deterministic optimal point and the worse case point. The new feasible region is smaller in comparison with that of the previous cycle. In general, the deterministic optimization problem for the kth cycle is min f (k d,k xN , pN )

k d,k xN

subject to

gi (k d,k xN −k SP,k−1 pworst ) ≥ 0,

i = 1, . . . , n

(37)

dL ≤k d ≤ dU , xL ≤k xN ≤ xU where the left superscript indicates the cycle number. The “shifting’’ vector for the possibilistic design variables is k SP = k−1 xN − k−1 xworst . Because the “shifting’’ vector

268

Structural design optimization considering uncertainties

xN2, X2

Deterministic constraint Possibilistic constraint g(mY1, xN2)
0

p[g(mY1, x2 )  0]  a

␣ cut (mY1, xN2) Shift to: g(mY1 – SS, xN2 – SP)

SP SS (Y1,MPP , x2,worst )

mY1, Y1

Figure 10.9 Shifting of feasible domain for a combination of uncertain and random variables.

idea cannot be used for the possibilistic design parameters P, the worst-case vector k−1 pworst from the previous cycle, is used. For the first cycle, the worst-case vector pworst is assumed equal to the nominal point pN . Subsequently, n possibilistic evaluation problems are solved (one for each possibilistic constraint). The following problem is solved for the ith constraint min gi (k d, x, p) x,p

subject to

x − δL (α) ≤ x ≤ k xN + δU (α)

k N

(38)

pL (α) ≤ p ≤ pU (α) and its solution determines k xworst and k pworst which are used in the next cycle. Problems (37) and (38) are repeated for a few cycles until convergence is achieved. 6.2

SPD O wi th b o t h po s s ib ilis t ic and r a n d o m v a r i a b l e s

A sequential approach is described in this section for a mixture of possibilistic and random variables. For demonstration purposes, Fig. 10.9 shows a geometric interpretation for a hypothetical problem with one random design variable (Y1 ), one possibilistic design variable (X2 ), and one deterministic constraint g(µY1 , xN 2 ) = 0 and one possibilistic constraint π[g(µY1 , x2 ) ≤ 0] ≤ α. For this general case, Problem (26) is solved. In the first cycle, the following deterministic optimization is performed min f (d, µY , µZ , xN , pN )

d,xN ,µY

subject to

gi (d, µY , µZ , xN , pN ) ≥ 0,

i = 1, . . . , n

dL ≤ d ≤ dU , µLY ≤ µY ≤ µU Y xL ≤ xN ≤ xU

(39)

N o n-p r o b a b i l i s t i c o p t i m i z a t i o n u s i n g p o s s i b i l i t y a n d e v i d e n c e t h e o r i e s

269

A possibilistic evaluation is then implemented for each constraint at the optimal point (d, µY , xN ) of Problem (39) in order to identify the worst-case point (xworst , pworst ), at the desired a-cut. The following problem is solved for the ith constraint, min gi (d, U, x, p)

U,x,p

subject to

U = βti

(40)

xN − δL (α) ≤ x ≤ xN + δU (α) pL (α) ≤ p ≤ pU (α)

The solution of the above problem is the worst-case point (d, Yi,MPP , Zi,MPP , xi,worst , pi,worst ) where (Yi,MPP , Zi,MPP ) is the worst-case most probable point (MPP) for the ith constraint. If point (d, Yi,MPP , Zi,MPP , xi,worst , pi,worst ) is deterministically infeasible, we must force it at least onto the deterministic constraint in order to ensure feasibility. This is achieved by using a “shifting’’ vector SS = {SS1 , . . . , SS } for the  random variables and a “shifting’’ vector SP = {SP1 , . . . , SPm } for the m possibilistic variables. In this case, the deterministic optimization of the next cycle is min f (d, µY , µZ , xN , pN )

d,xN ,µY

subject to

gi (d, µY − SS,1 ZMPP , xN − SP,1 pworst ) ≥ 0,

i = 1, . . . , n

(41)

xL ≤ xN ≤ xU In summary, the deterministic optimization problem for the kth cycle is min

k d,k xN ,k µ Y

f (k d,k µY , µZ , k xN , pN )

subject to

gi (k d,k µY −k SS,k−1 ZMPP , k xN − k SP,k−1 pworst ) ≥ 0, dL ≤k d ≤ dU , xL ≤ k xN ≤ xU

i = 1, . . . , n

(42)

µLY ≤ k µY ≤ µU Y where the left superscript indicates the cycle number and the “shifting’’ vectors for the random design variables and the possibilistic design variables are k SS = k−1 µY − k−1 YMPP and k SP = k−1 xN − k−1 xworst , respectively. Each constraint has its own “shifting’’ vectors. After the deterministic optimization, n possibilistic evaluation problems are solved (one for each constraint). The possibilistic assessment problem for the ith constraint is, min gi (k d, U, x, p)

U,x,p

subject to

U = βti x − δL (α) ≤ x ≤ k xN + δU (α)

k N

(43)

pL (α) ≤ p ≤ pU (α) The solution of Problem (43) determines the worst-case point (k d, Yi,MPP , Zi,MPP , xi,worst , pi,worst ) which is used in the next cycle. The sequence of Problems (42) and (43) is repeated until convergence.

270

Structural design optimization considering uncertainties

Starting

k
1,0 YMPP
0␮Y,0ZMPP
␮Z, 0xworst
0xN,0pworst
pN

kSS


k1␮Y –k1YMPP

kSP


k1xN –k1xworst

Det. Optimization min f(kd, k ␮Y, ␮Z, kxN, pN)

k ,k ,k N d ␮Y X

k
k1

Subject to g(kd,k␮Y – kSS, k1ZMPP,kxN – kSP,k1pworst ) 0 k

d,k␮Y ,kXN

Possibilistic Evaluation Find kYMPP,kZMPP ,kxworst and kpworst min g(kd, kU , kX, kp)

kU,kx,kp

Subject to

U
bt

k

x ␦L (a)  kx  kxN  ␦U(a)

k N

pL(a)k p  pU(a) kY k k k MPP, ZMPP , xworst , pworst

N

f converget? feasible solution? Y End

Figure 10.10 Flowchart for the SPDO algorithm.

Figure 10.10 shows the flowchart of the proposed SPDO algorithm for a combination of possibilistic and random variables. The details of the algorithm have been already provided in this section. More information is provided in (Zhou & Mourelatos 2007).

7 Examples In this section, the possibility-based and evidence-based design algorithms as well as the SPDO are demonstrated with a cantilever beam example and a pressure vessel example. In both examples, comparisons are made with deterministic design and reliability-based design results. It should be noted that theoretically, the possibility and reliability-based results cannot be compared because the possibility and reliability theories are based on different axioms. However for practical purposes, we attempt to compare them by arbitrarily using membership functions which “resemble’’ the probability density functions used in the reliability-based results.

N o n-p r o b a b i l i s t i c o p t i m i z a t i o n u s i n g p o s s i b i l i t y a n d e v i d e n c e t h e o r i e s

271

Y

Z

t L
100 in

w

Figure 10.11 Cantilever beam under vertical and lateral bending.

7.1 A cantilev er beam example In this example, a cantilever beam in vertical and lateral bending (Wu et al. 2001) is used (see Fig. 10.11). The beam is loaded at its tip by the vertical and lateral loads Y and Z, respectively. Its length L is equal to 100 in. The width w and thickness t of the cross-section are deterministic design variables. The objective is to minimize the weight of the beam. This is equivalent to minimizing f = w ∗ t, assuming that the material density and the beam length are constant. Two non-linear failure modes are used. The first failure mode is yielding at the fixed end of the cantilever; the other failure mode is that the tip displacement exceeds the allowable value of D0 = 2.5

. The PBDO problem is formulated as, min f = w ∗ t w,t

subject to

gjαmin ≥ 0 j = 1, 2  600 600 g1 (y, Z, Y, w, t) = y − ∗Y + 2 ∗Z wt 2  w t

 2 2 Z Y 4L3 g2 (E, Z, Y, w, t) = D0 − + 2 Ewt t w2 0 ≤ w, t ≤ 5

(44)

where g1 and g2 are the limit states corresponding to the two failure modes. The design variables w and t are deterministic. In the RBDO study of (Liang et al. 2007), Y, Z, y and E are normally distributed random parameters with Y ∼ N(1000, 100) lb, Z ∼ N (500, 100) lb, y ∼ N (40 000, 2000) psi and E ∼ N(29 ∗ 106 , 1.45 ∗ 106 ) psi; y is the random yield strength, Z and Y are mutually independent random loads in the vertical and lateral directions respectively, and E is the Young modulus. A reliability index β = 3 has been used in (Liang et al. 2007) for both constraints. For the PBDO case, Y, Z, y and E are possibilistic parameters described with the triangular membership functions (xN − 3 ∗ σ, xN , xN + 3 ∗ σ) where xN is the normal point of each variable is and σ is the used standard deviation in the RBDO study. The frame of discernment defined by the (xN − 3 ∗ σ, xN + 3 ∗ σ) coordinates is also used in EBDO. Table 10.2 compares the deterministic optimization, RBDO, PBDO and EBDO results. The PBDO optimum (objective function) with a = 0 is higher than the RBDO optimum. Because it represents the worst case design, it provides an upper bound of

272

Structural design optimization considering uncertainties

Table 10.2 Comparison of PBDO, EBDO and RBDO optima for the cantilever beam example. Design variables

w t Objective f(w,t) Constraints g 1 (x)/ y g 2 (x)/D0

Determ. optimum

Reliability optimum

Possibility optimum

Evidence optimum

α = 0.1

pf = 0.1

pf = 0.0013

2.0470 3.7459

2.4781 3.8421

2.4534 3.6162

2.5028 3.9902

7.6679

9.5212

10.556

10.901

8.8721

9.9868

0 0

0 0.1436

0 0.15

0 0.168

0 0.00428

0.0032 0.0835

2.5298 4.1726

α=0 2.5901 4.210

all RBDO optima obtained with different distributions, as long as these distributions have similar variability ranges (e.g. different beta distributions defined over the same range). For a higher a-cut (a = 0.1), the PBDO optimum reduces. It should be noted that the PBDO optimum at a = 1 coincides with the deterministic optimum. The last two rows of Table 10.2 show the normalized values of the two constraints at the optimum. The first constraint is normalized by the mean yield strength y = 40 000 and the second constraint is normalized by the allowable tip displacement D0 = 2.5. Although both constraints are active at the deterministic optimum, only the first constraint is active for both the RBDO and PBDO optima. The EBDO problem formulation is the same with Problem (44) but with different constraints. The new constraints are Pl(gi < 0) ≤ pf , i = 1, 2. The uncertain parameters P = [Y, Z, y, E] have the BPA structure of Table 10.3. The BPA for each interval of an uncertain parameter is assumed to be equal to the area under the PDF used in RBDO, in order to compare the EBDO design with the corresponding RBDO design. This is not how the BPA is obtained in general. As it has been mentioned, expert opinions are used to construct the BPA structure. If however, a random variable or parameter is described probabilistically, equivalent BPA values within specified intervals are calculated as equal to the area under the PDF. In doing so, the evidence theory can be used to handle a mixture of probabilistic and non-probabilistic variables. The last two columns of Table 10.2 show the EBDO results for pf = 0.1 and 0.0013 (β = 3). As expected, the deterministic optimum of 7.6679 is less than the RBDO optimum of 9.5212 which in turn, is less than the EBDO optimum of 9.9868 at pf = 0.0013 (β = 3). For pf = 0.1, the EBDO optimum reduces. Furthermore, the EBDO optimum of 9.9868 at pf = 0.0013 is better than the worst case PBDO optimum of 10.901 (a = 0). Although only the first constraint is active for the RBDO and PBDO optima, both constraints are active for the EBDO optima, similarly to the deterministic case. 7.2 A pressu re ve s s e l e xample This example considers the design of a thin-walled pressure vessel (Lewis & Mistree 1997) which has hemispherical ends as shown in Fig. 10.12. The design objective is to

N o n-p r o b a b i l i s t i c o p t i m i z a t i o n u s i n g p o s s i b i l i t y a n d e v i d e n c e t h e o r i e s

273

Table 10.3 BPA structure for y,Y, Z and E. y (×103 )

Z Interval

BPA (%)

Interval

BPA (%)

[200 300] [300 400] [400 450] [450 500] [500 550] [550 600] [600 700] [700 800]

2.2 13.6 15 19.2 19.2 15 13.6 2.2

[35 37] [37 38] [38 39] [39 40] [40 41] [41 42] [42 43] [43 45]

6.1 9.2 15 19.2 19.2 15 9.2 7.1

E (×106 )

Y Interval

BPA (%)

Interval

BPA (%)

[700 800] [800 900] [900 1 000] [1000 1 100] [1100 1 200] [1200 1 300]

2.2 13.6 34.1 34.1 13.6 2.4

[26.5 27.5] [27.5 28.5] [28.5 29] [29 29.5] [29.5 30.5] [30.5 31.3]

10 21 13.5 13.5 21 21

T

R

R

L

R R L R

Figure 10.12 Thin-walled pressure vessel.

calculate the radius R, mid-section length L and wall thickness t in order to maximize the volume while avoiding yielding of the material in both the circumferential and radial directions under an internal pressure P. Geometric constraints are also considered. The material yield strength is Y. A safety factor SF = 2 is used.

274

Structural design optimization considering uncertainties

Table 10.4 BPA structure for R, L, t, P and Y. R

L

t

BPA (%)

[RN − 6.0 RN − 4.5] [RN − 4.5 RN − 3.0] [RN − 3.0 RN ] [RN RN + 3.0] [RN + 3.0 RN + 4.5] [RN + 4.5 RN + 6.0]

[LN − 12 LN − 9] [LN − 9 LN − 6] [LN − 6 LN ] [LN LN + 6] [LN + 6 LN + 9] [LN + 9 LN + 12]

[tN − 0.4 tN − 0.3] [tN − 0.3 tN − 0.2] [tN − 0.2 tN ] [tN tN + 0.2] [tN + 0.2 tN + 0.3] [tN + 0.3 tN + 0.4]

0.13 2.15 47.72 47.72 2.15 0.13

P

Y

BPA (%)

[800 850] [850 900] [900 1000] [1000 1100] [1100 1150] [1150 1200]

[208000 221000] [221000 234000] [234000 260000] [260000 286000] [286000 299000] [299000 312000]

0.13 2.15 47.72 47.72 2.15 0.13

The PBDO problem is stated as 4 3 πR + πR2N LN 3 N subject to gjα ≥ 0 j = 1, . . . , 5 max f =

RN ,LN ,tN

min

where, P(R + 0.5t)SF 2tY P(2R2 + 2Rt + t 2 )SF g2 (X) = 1.0 − (2Rt + t 2 )Y L + 2R + 2t g3 (X) = 1.0 − 60 R+t g4 (X) = 1.0 − 12 5t g5 (X) = 1.0 − R 0.25 ≤ tN ≤ 2.0 6.0 ≤ RN ≤ 24 10 ≤ LN ≤ 48 g1 (X) = 1.0 −

The EBDO problem formulation is the same but with constraints Pl(gj (X) < 0) ≤ pf j = 1, . . . , 5. For the EBDO case, the uncertainty in design variables R, L, and t and design parameters P and Y are represented with the combined BPA structure of Table 10.4. To compare results with RBDO, the BPA values of R, L, t, P and Y are taken equal to the area under the PDF of a normal distribution for the intervals shown in Table 10.4. The normal distributions for R, L, t, P and Y have standard deviations

N o n-p r o b a b i l i s t i c o p t i m i z a t i o n u s i n g p o s s i b i l i t y a n d e v i d e n c e t h e o r i e s

275

Table 10.5 Comparison of deterministic, RBDO, and EBDO optima for vessel example. Design variables

RN LN tN Objective −f (RN , LN )

Determ. optimum

Reliability optimum

Evidence optimum pf = 0.2

pf = 0.0228

11.750 36.000 0.250

8.7244 33.5186 0.269

8.333 30.407 0.347

8.1111 26.1852 0.3472

22 400

10 791

9053

7644

Table 10.6 Convergence history for the pressure vessel example. Cycle #

Design variables (RN , µL , µt )

Obj.

g1 (X)

g2 (X)

g3 (X)

g4 (X)

g5 (X)

α=0 1 2 3

(11.75, 36.0, 0.25) (7.0108, 30.3867, 0.2892) (7.0107, 30.3867, 0.2893)

22 398 6132 6132

−0.2551 0.4996 0.5

−1.5101 0.0 0.0

−0.2502 0.0 0.0

−0.3917 0.0 0.0

0.6897 0.0258 0.0256

α = 0.2 1 2 3

(11.75, 36.0, 0.25) (7.9108, 30.3867, 0.2892) (7.9107, 30.3867, 0.2893)

22 398 8044 8044

−0.1857 0.4996 0.5

−1.3713 0.0 0.0

−0.2202 0.0 0.0

−0.3167 0.0 0.0

0.7239 0.4326 0.4325

equal to 1.5, 3, 0.1, 50 and 13 000, respectively. The mean values for parameters P and Y are taken equal to 1000 and 260 000. The intervals for R, L, t, P and Y extend four standard deviations from each side of the normal point, in an attempt to use a similar variation with the RBDO study. Finally, EBDO and PBDO use the same frame of discernment. Table 10.5 compares the deterministic optimization, RBDO and EBDO results. Similar conclusions with the previous example are drawn. A reliability index β = 2.0 (pf = 0.0228) has been used in the RBDO study for all constraints. The EBDO maximum volume for pf = 0.0228 is lower than the corresponding RBDO volume. For comparison purposes, the EBDO optima for pf = 0.2 and pf = 0.0228 have also been calculated. As shown in Table 10.5, the EBDO maximum volume increases with increasing pf , as expected. In this example, the third and fourth constraints are active for the deterministic, RBDO and EBDO optima. Table 10.6 gives the convergence history of the SPDO method for α = 0 and a = 0.2. It lists the values of design variables, objective function and the five constraints for each cycle. For both a-cuts, the algorithm converges in three cycles.

276

Structural design optimization considering uncertainties Table 10.7 SPDO results and comparisons for the pressure vessel example. Design variables

Determ. opt.

Double-loop RBDO

Double-loop PBDO

SPDO

a = 0.2

a=0

a = 0.2

a=0

RN µL µt

11.750 36.000 0.250

8.7244 33.5186 0.269

7.9108 30.483 0.2894

7.000 30.660 0.2997

7.9107 30.3867 0.2893

7.0107 30.3867 0.2893

Objective f (RN , µL )

22 400

10 791

8062

6150

8044

6132

Constraints g 1 (X) g 2 (X) g 3 (X) g 4 (X) g 5 (X) No. of F.E.

0.8173 0.6346 0 0 0.8936 96

0.5003 0 0 0 0.6891 5904

0.5 0 0 0 0.4323 9470

0.55 0.1 0 0 0 10 534

0.5 0 0 0 0.4325 1832

0.5 0 0 0 0.0256 2121

Table 10.7 compares the deterministic optimization, RBDO, double-loop PBDO and SPDO results. Two a-cuts (α = 0 and α = 0.2) are used for the possibilistic design. Similarly to the previous example, the proposed SPDO approach gives the same results with the double-loop PBDO with a much better efficiency. For example, the number of function evaluations for a = 0 is 2121 and 10 534 for SPDO and double-loop PBDO, respectively. As expected, the deterministic optimum of 22 400 is higher than the RBDO optimum of 10 791 which is in turn, higher than the worst-case (a = 0) PBDO optimum of 6150. Note that in this example the objective is maximized. Also, the PBDO optimum of 8062 (a = 0.2) is higher than the worst-case optimum of 6150 (a = 0). At the deterministic optimum, only the third and fourth constraints are active. However at the RBDO and PBDO optima, the second constraint is also active. It should be also noted that the computational cost of the double-loop PBDO is usually higher than that of the double-loop RBDO (see Table 10.7) due to the different problem formulation between the two.

8 Conclusions In this chapter, the possibility and evidence theories were used to assess design reliability with incomplete information. The possibility theory was viewed as a variant of fuzzy set theory. The different types of uncertainty and formal uncertainty theories were first introduced using the fundamentals of fuzzy measures. Subsequently, the commonly used vertex and discretization methods which are used for propagating non-probabilistic uncertainty were reviewed and compared with a hybrid (globallocal) optimization method. It was showed that the hybrid optimization method is very efficient and has the same accuracy with the “brute force’’ discretization method. The possibility theory was also used in design. A possibility-based design optimization method was proposed where all design constraints are expressed possibilistically.

N o n-p r o b a b i l i s t i c o p t i m i z a t i o n u s i n g p o s s i b i l i t y a n d e v i d e n c e t h e o r i e s

277

It was shown that the method gives a conservative solution compared with all conventional reliability-based designs obtained with different probability distributions. A general possibility-based design optimization method was also presented which handles a combination of random and possibilistic design variables. Furthermore, a sequential algorithm for possibility-based design optimization (SPDO) was introduced. It decouples a double-loop PBDO process into a sequence of cycles composed of a deterministic design optimization followed by a set of worst-case reliability evaluation loops. The computational cost is kept low by first using the performance measure approach in reliability analysis and second by decoupling the deterministic design optimization from the worst-case reliability evaluation. A computationally efficient evidenced-based design optimization method was also described, which can handle a mixture of epistemic and random uncertainties. A mean performance is optimized subject to the plausibility of constraint violation being small. Uncertainty is quantified using “expert’’ opinions. Two examples demonstrated the proposed possibility-based and evidence-based design optimization methods. It was shown that both the PBDO and EBDO designs are more conservative compared with the RBDO design. However, the EBDO design is usually less conservative compared with the PBDO design.

References Agarwal, H., Renaud, J.E., Preston, E.L. & Padmanabhan, D. 2004. Uncertainty Quantification Using Evidence Theory in Multidisciplinary Design Optimization. Reliability Engineering and System Safety 85:281–294. Akpan, U.O., Rushton, P.A. & Koko, T.S. 2002. Fuzzy Probabilistic Assessment of the Impact of Corrosion on Fatigue of Aircraft Structures. Paper AIAA-2002-1640. Bae, H.-R., Grandhi, R.V. & Canfield, R.A. 2004. An Approximation Approach for Uncertainty Quantification Using Evidence Theory. Reliability Engineering and System Safety 86: 215–225. Bae, H.-R., Grandhi, R.V. & Canfield, R.A. 2004. Epistemic Uncertainty Quantification Techniques Including Evidence Theory for Large-Scale Structures. Computers and Structures 82: 1101–1112. Chen, L. & Rao, S.S. 1997. Fuzzy Finite Element Approach for the Vibration Analysis of Imprecisely Defined Systems. Finite Elements in Analysis and Design 27:69–83. Choi, K.K., Du, L. & Youn, B.D. 2004. A New Fuzzy Analysis Method for PossibilityBased Design Optimization. 10th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, AIAA 2004-4585, Albany, NY. Du, X. & Chen, W. 2000. An Integrated Methodology for Uncertainty Propagation and Management in Simulation-Based Systems Design. AIAA Journal 38(8):1471–1478. Du, X. & Chen, W. 2004. Sequential Optimization and Reliability Assessment Method for Efficient Probabilistic Design. ASME Journal of Mechanical Design 126:225–233. Du, X., Sudjianto, A. & Huang, B. 2005. Reliability-Based Design with a Mixture of Random and Interval Variables. ASME Journal of Mechanical Design 127:1068–1076. Dubois, D. & Prade, H. 1988. Possibility Theory. New York: Plenum Press. Elishakoff, I.E., Haftka, R.T. & Fang, J. 1994. Structural Design under Bounded Uncertainty – Optimization with Anti-Optimization. Computers and Structures 53:1401–1405. Gu, X., Renaud, J.E. & Batill, S.M. 1998. An Investigation of Multidisciplinary Design Subject to Uncertainties. 7th AIAA/USAF/NASA/ISSMO Multidisciplinary Analysis and Optimization Symposium, St. Louis, Missouri.

278

Structural design optimization considering uncertainties

Jones, D.R., Perttunen, C.D. & Stuckman, B.E. 1993. Lipschitzian Optimization Without the Lipschitz Constant. Journal of Optimization Theory and Applications 73(1):157–181. Klir, G.J. & Filger, T.A. 1988. Fuzzy Sets, Uncertainty, and Information. Prentice Hall. Klir, G.J. & Yuan, B. 1995. Fuzzy Sets and Fuzzy Logic: Theory and Applications. Prentice Hall. Lee, J.O., Yang, Y.O. & Ruy, W.S. 2002. A Comparative Study on Reliability Index and Target Performance Based Probabilistic Structural Design Optimization. Computers and Structures 80:257–269. Lewis, K. & Mistree, F. 1997. Collaborative, Sequential and Isolated Decisions in Design. Proceedings of ASME Design Engineering Technical Conferences, Paper# DETC1997/ DTM-3883. Liang, J., Mourelatos, Z.P. & Tu, J. 2007. A Single-Loop Method for Reliability-Based Design Optimization. In press International Journal of Product Development. Also, Proceedings of ASME Design Engineering Technical Conferences, 2004, Paper# DETC2004/DAC-57255. Lombardi, M. & Haftka, R.T. 1998. Anti-Optimization Technique for Structural Design under Load Uncertainties. Computer Methods in Applied Mechanics and Engineering 157:19–31. Moore, R.E. 1966. Interval Analysis. Prentice-Hall. Mourelatos, Z.P. & Zhou, J. 2006. A Design Optimization Method using Evidence Theory. ASME Journal of Mechanical Design 128(4):901–908. Mourelatos, Z.P. & Zhou, J. 2005. Reliability Estimation with Insufficient Data Based on Possibility theory. AIAA Journal 43(8):1696–1705. Muhanna, R.L. & Mullen, R.L. 2001. Uncertainty in Mechanics Problems – Interval-Based Approach. Journal of Engineering Mechanics 127(6):557–566. Mullen, R.L. & Muhanna, R.L. 1999. Bounds of Structural Response for all Possible Loadings. ASCE Journal of Structural Engineering 125(1):98–106. Nikolaidis, E., Chen, S., Cudney, H., Haftka, R.T. & Rosca, R. 2004. Comparison of Probability and Possibility for Design Against Catastrophic Failure Under Uncertainty. ASME Journal of Mechanical Design 126:386–394. Oberkampf, W.L. & Helton, J. 2002. Investigation of Evidence Theory for Engineering Applications. AIAA Non-Deterministic Approaches Forum, AIAA 2002-1569, Denver, CO. Oberkampf, W., Helton, J. & Sentz, K. 2001. Mathematical Representations of Uncertainty. AIAA Non-Deterministic Approaches Forum, AIAA 2001-1645, Seattle, WA, April 16–19. Penmetsa, R.C. & Grandhi, R.V. 2002. Efficient Estimation of Structural Reliability for Problems with Uncertain Intervals. Computers and Structures 80:1103–1112. Penmetsa, R.C. & Grandhi, R.V. 2002. Estimating Membership Response Function using Surrogate Models. Paper AIAA 2002-1234. Rao, S.S. & Cao, L. 2002. Optimum Design of Mechanical Systems Involving Interval Parameters. ASME Journal of Mechanical Design 124:465–472. Rao, S.S. & Sawyer, J.P. 1995. A Fuzzy Finite Element Approach for the Analysis of Imprecisely Defined Systems. AIAA Journal 33:2264–2370. Ross, T.J. 1995. Fuzzy Logic with Engineering Applications. McGraw Hill. Sentz, K. & Ferson, S. 2002. Combination of Evidence in Dempster – Shafer Theory. Sandia National Laboratories Report SAND2002-0835. Tu, J., Choi, K.K. & Park, Y.H. 1999. A New Study on Reliability-Based Design Optimization. ASME Journal of Mechanical Design 121:557–564. Tu, J. & Jones, D.R. 2003. Variable Screening in Metamodel Design by Cross-Validated Moving Least Squares Method. Proceedings 44th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference, AIAA-2003-1669, Norfolk, VA. Wang, G. 2003. Adaptive Response Surface Method Using Inherited Latin Hypercube Design Points. ASME Journal of Mechanical Design 125:1–11.

N o n-p r o b a b i l i s t i c o p t i m i z a t i o n u s i n g p o s s i b i l i t y a n d e v i d e n c e t h e o r i e s

279

Wu, Y.-T., Shin, Y., Sues, R. & Cesare, M. 2001. Safety – Factor Based Approach for Probabilistic – Based Design Optimization. 42nd AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference. Seattle, WA. Yager, R.R., Fedrizzi, M. & Kacprzyk, J. (eds) 1994. Advances in the Dempster – Shafer Theory of Evidence. John Wiley & Sons, Inc. Ye, K.Q., Li, W. & Sudjianto, A. 2000. Algorithmic Construction of Optimal Symmetric Latin Hypercube Designs. Journal of Statistical Planning and Inference 90:145–159. Youn, B.D., Choi, K.K. & Park, Y.H. 2001. Hybrid Analysis Method for Reliability-Based Design Optimization. ASME Journal of Mechanical Design 125(2):221–232. Zadeh, L.A. 1965. Fuzzy Sets. Information and Control 8:338–353. Zadeh, L.A. 1978. Fuzzy Sets as a Basis for a Theory of Possibility. Fuzzy Sets and Systems 1:3–28. Zhou, J. & Mourelatos, Z.P. 2007. A Sequential Algorithm for Possibility-Based Design Optimization. In press ASME Journal of Mechanical Design. Also, Proceedings of ASME Design Engineering Technical Conferences, 2006, Paper# DETC2006-99232.

Chapter 11

A decoupled approach to reliabilitybased topology optimization for structural synthesis Neal M. Patel, John E. Renaud & Donald Tillotson University of Notre Dame, Notre Dame, IN, USA

Harish Agarwal General Electric Global Research, Niskayuna, NY, USA

Andrés Tovar National University of Colombia, Bogota, Colombia

ABSTRACT: Conceptual designs of structures have been generated using topology optimization over the past two decades. However, traditional topology optimization techniques neglect uncertainties that exist in the real-world. In this chapter, this problem is addressed by including the notion of reliability into the design process. A reliability-based topology optimization (RBTO) framework for structural synthesis is proposed using a decoupled reliability-based design optimization (RBDO) approach, so that the topology synthesis is separate from the reliability analysis. In the algorithm presented, a maximum allowable displacement failure mode is considered. Starting from a continuum design space of uniform material distribution and initial uncertain variable values, a deterministic topology optimization is followed by a reliability analysis of the resulting structure to determine the most probable point of failure (MPP) for the current structure. The MPP is determined with respect to the maximum allowable deflection of the structure for a given applied loading. The non-gradient Hybrid Cellular Automaton (HCA) method used for topology optimization is combined with the decoupled approach for RBDO to develop a new continuum-based approach to RBTO. The objective of this chapter is to present the background behind the methods employed and demonstrate capabilities of the RBTO framework using examples.

1 Introduction Concept designs for minimum compliance structures can be synthesized using topology optimization (Bendsoe and Kikuchi 1988). Traditional techniques neglect variabilities that occur over the life and use of a structure. For example, the structure of a bridge can incur vastly different loading depending on the traffic pattern for a given time of the day. Uncertainties may exist in certain material properties as well. Reliability-based design optimization (RBDO) is a probabilistic optimization method that has been used in design problems to account for variation and uncertainty. The objective of RBDO is to mediate between cost and safety. In deterministic optimization, designs are often driven to the limits of the design constraints, neglecting tolerances in modeling and simulation uncertainties. Therefore, the resulting optimized designs can be unreliable with a high probability of failure when in use. Factor of safety techniques have been

282

Structural design optimization considering uncertainties

employed as a popular method for accounting for uncertainties and off-design operation, but these designs are typically over-engineered, leading to higher cost since the uncertainties are not necessarily quantified. The probabilistic RBDO approach facilitates the design to a specific risk and target reliability level accounting for the various sources of uncertainty. In probabilistic optimization methods, these variational uncertainties are modeled as random variables. In this respect, the deterministic analysis can be viewed as an extension of the probabilistic analysis, where the deterministic quantities are a trivial instance of the random variables. Reliability-based topology optimization (RBTO) extends the notion of reliability to the area of topology optimization. In this chapter, we consider a discretized continuum design domain, where the density of each element is used as a design variable. Traditional topology optimization methods drive the topology of a structure to an optimum design based on a single constraint on mass. However, nothing can be said about the reliability of the resulting topology since it does not account for uncertainties and modes of failure that the structure realistically would require. Because of the large number of design variables associated with topology optimization problems, the inclusion of RBDO methods could be computationally time prohibitive for largescale problems because of gradient calculations required in the sensitivity analysis. Therefore, research in this area is on concentrated in developing efficient reliability based topology optimization techniques. Kharmanda et al. (Kharmanda, Olhoff, Mohamed, and Lemaire 2004) proposed a reliability-based methodology for topology optimization using a heuristic strategy that aims to reduce mass while improving the reliability level of the structure without increasing its weight. However, in this approach, the failure mode is purely a linear combination of the random variables and does not have any physical meaning. Mogami et al. (Mogami, Nishiwaki, Izui, Yoshimura, and Kogiso 2006) incorporated reliability-based constraints in the topology optimization method using discrete frame elements and the traditional double-loop approach. Maute and Frangopol (Maute and Frangopol 1998) extended the notion of reliability to Micro-Electro-Mechanical Systems (MEMS) design using topology optimization. In the RBTO framework presented here, a decoupled approach is employed such that the topology optimization is separate from the reliability analysis (Agarwal and Renaud 2006). The decoupled reliability-based design optimization methodology is an approximate technique to obtain consistent reliable designs at a lower computational expense. Starting from an initial design domain of full material and uncertain parameters, such as loads, a complete topology optimization is followed by a reliability analysis of the structure; because the main optimization and the reliability analysis phases are detached, we refer to this as a decoupled approach. Although the RBTO framework can be generalized for use with any topology optimization method, in this work the Hybrid Cellular Automaton (HCA) method is utilized for deterministic continuum structural synthesis of minimum compliance structures (Tovar, Quevedo, Patel, and Renaud 2006). It is assumed that the structural deformation is elastic and loading is static. The change in density is evaluated locally using a CA rule, while the compliance is evaluated using a global structural analysis via the finite element method (FEM). In the presented methodology, RBTO has the same objective as the deterministic topology optimization: minimize compliance. Typically maximum deflection and stress are of concern when deigning a structure for

A d e c o u p l e d a p p r o a c h t o r e l i a b i l i t y-b a s e d t o p o l o g y o p t i m i z a t i o n

283

maximum stiffness. Here we consider the mode of failure to be the maximum deflection of the structure when loads are applied. Therefore, a constraint on maximum allowable displacement of the structure is implemented as well as a similar displacement constraint formulation for the limit-state function. The utilization of the gradient free HCA method in the RBTO framework adds efficiency to the methodology. In the topology optimization problem, the design variables are the densities of the material elements that make up the design domain. Characteristics of the problem that may have some associated uncertainty are identified as uncertain parameters. The reliability subproblem is applied to the topology generated. A new topology optimization is executed using the uncertain parameter values at the most probable point of failure (MPP), as determined in this subproblem. This process is repeated until convergence. The RBTO framework is applied to two design problems and the final designs are validated using the Monte Carlo simulation. In these problems, the elastic modulus and applied loading are considered as the uncertain parameters, characterized by a normal distribution, and a first-order estimate is used to approximate the failure surface.

2 Reliability-based design optimization Optimized designs based on a deterministic formulation are usually associated with a high probability of failure due to inherent uncertainties associated with the imposed design constraints. In today’s competitive marketplace, it is very important that the resulting designs are both optimum and at the same time reliable. Optimized designs without considering the variability of design variables and parameters can be prone to failure in service. In order to achieve the objective of obtaining reliable optimum designs, a designer must replace a deterministic optimization with a reliability-based design optimization (RBDO), where the critical probabilistic constraints are replaced with reliability constraints, as shown below min f (x, p) x

subject to

gD (V) ≥ 0 gR (x, p) ≥ 0

(1)

xl ≤ x ≤ xu where x and p represent the design variables and fixed parameters, respectively, and gR and gD denote reliability and deterministic constraints. The reliability constraints are either constraints on probabilities of failure corresponding to each probabilistic constraint, or a single constraint on the overall system probability of failure. The reliability constraints (gR ) can be formulated as giR = Pallowi − Pi

i = 1, . . . , k

gR = Pallowsys − Psys

(2) (3)

for k constraints where Pi is the failure probability of the probabilistic constraint giR at a given design and Pallowi is the allowable probability of failure for that failure mode.

284

Structural design optimization considering uncertainties

The parameter Psys is the system failure probability at a given design and Pallowsys is the allowable system probability of failure. These probabilities of failure are usually estimated by employing standard reliability techniques. The reliability analysis is a tool used to compute the reliability index or the probability of failure corresponding to a given failure mode or for the entire system (Haldar and Mahadevan 2001). The reliability analysis involves a probability distribution transformation, the search for the MPP, and the evaluation of the cumulative Gaussian distribution function. The uncertainties are modeled as continuous random variables V = (V1 , V2 , . . . , Vn )T , with known (or assumed) continuously differentiable distribution functions, FV (v). The ith random probabilistic constraint is denoted as giR (V, η), where η refers to deterministic parameters, also called limit state parameters. In the following, v denotes a realization of the random variables V. Letting giR (V, η) ≤ 0 represent the failure domain and giR (V, η) = 0 be the so-called limit state function, then the time-invariant probability of failure for the ith probabilistic constraint is given by
Pi (η) =

fV (v)dx

(4)

giR (x,η)≤0

where fV (v) is the joint probability density of V. It is almost impossible to find an analytical solution to the above integral. In standard reliability techniques, a probability distribution transformation T is usually employed, as illustrated in Fig. 11.1. An arbitrary n-dimensional random vector V = (V1 , V2 , . . . , Vn )T is mapped into an independent standard normal vector U = (U1 , U2 , . . . , Un )T . The standard normal random variables are characterized by zero mean and unit variance. The limit state function in U-space can be obtained as giR (x, η) = giR (T(u), η) = GR i (u, η) = 0. The failure domain in U-space is GR i (u, η) ≤ 0.

u2

x2

GR (u, h)  0 (fail) gR (x, h)  0 (safe)

u
T (x)

u*(MPP) b

FORM u1

x1

SORM gR

(x, h)  0 (fail) GR (u, h)  0 (safe)

Original space

Standard space

Figure 11.1 Transformation from the original space to the standard space.

A d e c o u p l e d a p p r o a c h t o r e l i a b i l i t y-b a s e d t o p o l o g y o p t i m i z a t i o n

285

Equation (4) thus transforms to
Pi (η) =

φU (u) du

(5)

GR i (u,η)≤0

where φU (u) is the standard normal density of u. If the limit state function in U-space is affine, i.e., if GR (u, η) = T u + β, then an exact result for the probability of failure is Pf =
(−β), where
() is the cumulative Gaussian distribution function. If the limit state function is close to being linear, i.e., if GR (u, η) ≈ T u + β with β = −T u∗ , where u∗ is the solution of the following optimization problem min ||u|| u

subject to

GR (u, η) = 0

(6)

then the first-order estimate of the probability of failure is Pf =
(−βp ), where  represents the vector of direction cosines at the solution point. The solution u∗ of the above optimization problem, the so called design point, β-point or the MMP of failure, defines the reliability index βp = ||u∗ ||. This method of estimating the probability of failure is known as the First-Order Reliability Method (FORM) (Haldar and Mahadevan 2001). In the second-order reliability method (SORM), the limit state function is approximated as a quadratic surface (Breitung 1984). However, first order approximations, Pf (η) ≈
(−βp ), are usually sufficient for most practical cases and, therefore, are used in this chapter. Using the FORM estimate, the reliability constraints in Eq. (2) can be written in terms of reliability indices as follows girc = βi − βreqdi

(7)

where βi is the calculated reliability index and βreqdi = −
−1 (Pallowi ) is the desired reliability index for the ith probabilistic constraint. This is referred to the reliability index approach (RIA). RIA can be solved as an optimization problem to solve for the constraint in Eq. (2). The reliability index corresponding to a failure mode requires the solution of the optimization problem in (6). Various algorithms have been reported in the literature (P. Lui 1991) to solve for the solution, which typically requires many system analysis evaluations. Moreover, RIA may fail to provide a solution to the FORM problem, especially when the limit state surface is far away from the origin in U-space or when the case GR (u, η ) = 0 never occurs at a particular design variable setting. Thus, the most challenging task is the search for the MPP. To overcome these difficulties in RIA, Choi et al. (Choi, Youn, and Yang 2001) provides an improved formulation to solve the RBDO problem. In this method, known as the performance measure approach (PMA), the reliability constraints are stated by an inverse formulation girc = Gi (ui∗ , η ) R,∗

i = 1, . . . , k

(8)

286

Structural design optimization considering uncertainties R,∗

where Gi is the solution to an inverse reliability analysis (IRA). This optimization problem is stated as min GR i (u, η ) u

subject to

u = βreqdi

(9)

where ui∗ is the optimum (corresponding MPP in IRA) of the ith reliability constraint. Solving RBDO by the PMA formulation is usually more efficient and robust than the RIA formulation, where the reliability is evaluated directly. PMA is, therefore, used in the proposed methodology. The efficiency lies in the fact that the search for the MPP of an inverse reliability problem is easier to solve than the search of the MPP corresponding to an actual reliability. 2.1

Rel i a b i l i ty in s t r uc t ur al o pt imizat i o n

Reliability in structural design has been developed considerably since the 1970’s (Moses 1973). Haldar and Mahadevan (Haldar and Mahadevan 2001), Haftka et al. (Haftka, Gürdal, and Kamat 1990), among others (Frangopol 1998), present a comprehensive background in structural reliability. Murotsu and Shao (Murotsu and Shao 1989) applied the notion of reliability to shape optimization of truss structures, where nodal coordinates are used as shape design variables with sizing design variables, such as the cross-sectional areas of the truss members. Papadrakakis and Lagaros utilized neural networks and the Monte Carlo simulation to perform reliability-based structural optimization of large-scale structural systems. Royset et al. (Royset, Kiureghian, and Polak 2001) developed a decoupled technique for reliability-based structural optimization where the structural optimization and reliability analysis were separated. In that methodology, a semi-definite optimization algorithm was utilized for the structural optimization. Frangopol and Maute (Frangopol and Maute 2003) reviewed the state of the art in reliability based design in both civil and aerospace structures. The inclusion of reliability was then extended to the design of aeroelastic structures by Allen and Maute (Allen and Maute 2004). In the current work, reliability is explored in the area of topology optimization.

3 Topology optimization The roots of topology optimization date back to the late 1980’s. This computational technique for the optimal distribution of material of continuum structures was first introduced by Bendsøe and Kikuchi (Bendsoe and Kikuchi 1988). Topology optimization can be viewed as a method for developing an initial, or concept, design. The optimization process systematically eliminates and redistributes material throughout the domain to minimize or maximize a specified objective. Early work in topology optimization generally dealt with simple problems that used the assumptions of elastic material properties, linear deformations, and static loading conditions. A comprehensive review of topology optimization can be found in literature by Bendsøe and Sigmund (Bendsøe and Sigmund 1989), Rozvany (Rozvany 1997), and Eschenaueuer and Olhoff (Eschenaueuer and Olhoff 2001).

A d e c o u p l e d a p p r o a c h t o r e l i a b i l i t y-b a s e d t o p o l o g y o p t i m i z a t i o n

287

The objective and constraints considered are a global structural response such as mean compliance, von Mises stresses, eigenfrequencies, or geometrical parameters such as volume (or mass) or perimeter. This can be extended to multiple loading conditions. Traditionally, using the static-elastic assumption, the objective of a structural optimization problem is to achieve minimum compliance or strain energy with a constraint on the mass, or volume V. This can be expressed formally as min f (ρ) ρi

subject to

N i=1

ρi v i ≤ V

(10)

ρmin ≤ ρi ≤ ρmax where ρ are the elemental densities. The compliance of a structure due to loading can be expressed as c(x) = dT K(x)d

(11)

where K is the global stiffness matrix, d is the vector of global displacements, and F is the vector of external global forces. The vector x is the set of design variables related to the material state of the elements in the design domain, such as ρ. 3.1

M aterial parametrization

Ultimately, the goal of topology optimization is to determine a material distribution within the design domain to achieve a specified objective. One can utilize discrete structural elements, known as the ground structure approach, to described a structure. In this work, continuum elements are used. For continuum structures, the homogenization and density approaches are the two most popular material parameterizations. 3.1.1 The hom og eni zati on approach The initial work in topology optimization of continuum structures was based on composite material models to describe the material properties in all dimensions. This technique, presented in the seminal work of Bendsøe and Kikuchi (Bendsoe and Kikuchi 1988), is referred to as the homogenization approach, which uses composite materials as the basis for describing varying material properties where each element is a microstructure. The homogenization can be viewed as an interpolation model for void and full material. In the homogenization approach, the design domain consists of square cells. Each cell has a rectangular hole at the centroid defined by lengths a and b, as shown in Fig. 11.2. The rectangle is orientated at an angle θ. Ultimately, the density of an element is a function of the variables ai , bi , and θi . Thus, each element has three variables associated with it. The relationship between the size of the cavity and the material properties is obtained using homogenization method. This method is typically employed in two stages (Duysinx 1997). In the first stage, the microstructure orientation θ is varied,

288

Structural design optimization considering uncertainties



a

u b

Figure 11.2 A unit cell of a microstructure parameterized using the homogenization method.

x



Figure 11.3 An illustration of the density approach to material parametrization in topology optimization.

based on the principal strains. In the second stage, the microstructure parameters a and b are updated. This approach to material parametrization is typically utilized for linear-elastic material assumptions (Bendsoe and Kikuchi 1988), but has been applied to problems with both material and geometric nonlinearities as well (Yuge and Kikuchi 1995; Yuge, Iwai, and Kikuchi 1999). 3.1.2 Th e d en si t y a p p r o a ch A second technique for material parametrization deals with the more direct approach of associating just one design variable with each individual material element, as illustrated in Fig. 11.3. This is called the density approach. The material model is defined to allow the material to assume intermediate property values by utilizing an interpolation function. The design variables are the relative densities (xi ) of the elements where

A d e c o u p l e d a p p r o a c h t o r e l i a b i l i t y-b a s e d t o p o l o g y o p t i m i z a t i o n

289

0 represents a void and 1 is full density. The density of a material element can be expressed as ρi (xi ) = xi ρ0

(0 < xi ≤ 1)

(12)

where ρ0 is the density of the base material. In utilizing the finite element method (FEM), the design variable is mapped to the global stiffness matrix by relating the relative density of an element to its elastic modulus. The solid isotropic material with penalization (SIMP) model (Bendsoe 1989; Zhou and Rozvany 1991) is a commonly utilized interpolation scheme that heuristically relates the relative density to the elastic modulus of each element using the following expression p

Ei (xi ) = xi E0

(13)

where p is the penalization parameter (p ≥ 1) and E0 is the elastic modulus of the base isotropic material. Therefore, we can view the elements of differing relative densities in the design domain as unique isotropic material elements. The power p is used to penalize intermediate densities to drive the elemental densities within in the design domain to have either full density (x = 1) or no density (x = 0). Most optimizers require this penalization to generate 0–1 topologies. Although the density approach is used in gradient-based optimization methods because it is a continuous function, this material parametrization can be utilized with non-gradient methods so that material is distributed in a continuous manner from one iteration to the next. This allows the topology to evolve in a smooth, efficient manner. In this RBTO framework, a linear interpolation model (p = 1) is utilized to relate the design variable xi to the elastic modulus of a material element with an intermediate density, as expressed by Ei (xi ) = xi E0

3.2

(14)

Optimization techniques

Various methodologies have been developed for topology optimizations over the past two decades. Topology optimization algorithms fall into the categories of mathematical programming (MP), optimality criteria (OC), and evolutionary programming methods (Bendsøe and Sigmund 1989). Mathematical programming techniques are mathematically based methods for optimization. OC methods are derived from the Karush-Kuhn-Tucker (KKT) optimality conditions. Evolutionary methods are heuristic, or intuition-based, approaches that use mechanisms inspired by biological evolution, such as reproduction, mutation, and survival of the fittest, to find an optimal solution to a problem. An important distinction between classes of methods is that MP and OC methods utilize continuous design variables whereas evolutionary methods use discrete representations as design variables. Sequential Convex Programming (SCP) is an example of a MP approach used for solving topology design problems. The Method of Moving Asymptotes (MMA), developed by Svanberg (Svanberg 1987), is the most popular SCP algorithm used for structural optimization because of its efficiency. In this method, a strictly convex subproblem is approximated at each iteration based on sensitivity information

290

Structural design optimization considering uncertainties

at the current design and then solved. The roots of OC methods in continuum-based topology optimization date back to the pioneering work of Bendsøe and Kikuchi. OC methods are primarily suited for problems containing a small number of constraints as compared to the number of design variables. In general, the OC methods are more computationally efficient than conventional MP methods (Rozvany, Bendsoe, and Kirsh 1995). Since the material volume constraint is the only active constraint, an OC method can be used to provide more rapid convergence compared to other optimization schemes. The aforementioned algorithms require gradient information in obtaining the final solution. In contrast, numerous topology optimization methodologies have been developed using evolutionary strategies that do not require gradients. An often used but inefficient approach is to utilize genetic algorithms (GAs) or semi-stochastic techniques. These methods may be more likely to find global solutions, but they require thousands of function calls. Another non-gradient based methodology developed by Xie and Stevens (Xie and Stevens 1997) is called Evolutionary Structural Optimization (ESO). It is based on the concept of progressively removing inefficient material from a structure so that it evolves into an optimal design. Another approach that requires no gradient information and utilizes cellular automata (CA) is the Hybrid Cellular Automaton (HCA) method. Since there is no randomness in the HCA formulation, this method is considered to be a MP method. 3.2.1 Topology s y n t h e s is u s in g h y b r id ce llu lar aut o mat a A cellular automaton (CA) is a discrete model studied in computability theory and mathematics (Wolfram 2002). It consists of an regular grid of cells, or lattice, where each cell is characterized by a finite number of states. The state of each cell at a given time, or generation, is a function of the states of a finite number of neighboring cells, called the neighborhood. Every cell has the same set of rules, which are applied based on information in its neighborhood. These rules are applied to the entire CA lattice each generation. The notion of cellular automata was initially conceived by John von Neumann in the late 1940’s. According to Burks (Burks 1970), the first CA proposed by von Neumann was a two-dimensional square lattice comprised of several thousands cells. Each of these cells had up to 29 possible states. The CA rule required the state of each cell plus its four nearest neighbors, located directly north, south, east, west. This CA model was so complex that it has only been partially implemented on a computer. The von Neumann rule has the so-called property of universal computation, meaning that there exists an initial configuration of the CA which leads to the solution of any computer algorithm. Accordingly, any universal computer circuit (i.e., logical gate) can be simulated by the rule of the automaton. This illustrates that complex and unexpected behavior can emerge from a CA rule. Cellular automata rules are applied over a number of discrete time steps on each CA element based on information collected in its neighborhood. The rules operate on the set of the states of neighboring cells. The rules are applied iteratively for as many time steps as required. Therefore, the global behavior of the CAs is governed by the set of local rules. These rules operate according to local information collected in the neighborhood of each cellular automaton. The final state of a CA is defined by

A d e c o u p l e d a p p r o a c h t o r e l i a b i l i t y-b a s e d t o p o l o g y o p t i m i z a t i o n

291

the state of itself and states of the CAs within the neighborhood. For example, the information collected from a neighborhood can be expressed as S¯ i =

1

ˆ N 

ˆ +1 N

j=0

Sj

(15)

ˆ is the number of elements in its neighborwhere S0 is the field state of the ith CA and N hood. This can be viewed as a filtering technique that prevents numerical instabilities of checkerboarding and mesh dependency. In practice, the size of the neighborhood is often limited to the adjacent cells but can also be extended. Figure 11.4 depicts some common two-dimensional neighborhood layouts. In the cellular automata paradigm, the same neighborhood is applied for all CA in the lattice. In the context of structural optimization, no state information exists outside of the design domain. Therefore, the neighborhood is modified for the boundary elements to only include neighbors within the design region. One of the first applications of cellular automata to structural design was presented by Inou et al. (Inou, Shimotai, and Uesugi 1994; Inou, Uesugi, Iwasaki, and Ujihashi 1998). CAs have been applied for both discrete and continuous structures. Gürdal and Tatting (Gürdal and Tatting 2000) and Slotta et al. (Slotta, Tatting, Watson, Gürdal, and Missoum 2002) applied cellular automata to truss structures. In that application, a rectangular design domain was composed of an array of truss elements. Each cell was composed of a node and the eight trusses from neighboring nodes in an forty five degree arrangement. Kita and Toyoda (Kita and Toyoda 2000) presented a methodology that is similar to the HCA method developed by Tovar for structural synthesis in that it utilizes the finite element method for structural analysis. The local update rule was based on the minimization of both the weight of the structure and the deviation between the yield stress and the von Mises equivalent stress for each cell. Furthermore, a two-dimensional isotropic material was considered where the thickness of each CA was the design variable. Hajela and Kim (Hajela and Kim 2001) used a genetic algorithm (GA) based on energy minimization to determine an appropriate CA rule for a two-dimensional continuum. The Hybrid Cellular Automaton (HCA) method is a computational technique that has demonstrated the ability to act as an optimization tool for the synthesis of optimal topologies. This approach is inspired by the biological process of bone remodeling and was first presented by Tovar (Tovar 2004). As done in other topology optimization methods, the design domain is discretized into material elements. To use the finite element method for structural analysis, the design domain is represented using a finite

(a) Empty (N
0)

(b) Von Neumann (N
4)

(c) Moore (N
8) (d) Extended (N
24)

ˆ is the number of neighboring CAs. Figure 11.4 Typical 2-D neighborhoods for CAs. N

292

Structural design optimization considering uncertainties

element model that is discretized using continuum finite elements (FE). The states of the material elements in the design domain are represented using a lattice of CAs, where a one-to-one correspondence between CA and FE generally exists, although this is not a requirement. However, uniformity in the CA discritization is required. A set of local rules is used to determine material distribution. These rules are applied to the local information collected in the neighborhood of each CA. At a discrete position i and time/iteration k, a CA is defined by a set of states that are operated on by a set of rules belonging to a given neighborhood of the CA. The state of each CA αi , is defined by the associated design variables xi (e.g., density, thickness) and field variables Si (e.g., compliance). The field variables are computed by a finite element analysis; hence this is a hybrid approach since each CA is provided global information. The complete state of each cell is expressed by 

(k) αi

(k)

Si = (k) xi

 (16)

where k denotes the state applies to a specific iteration. The HCA method has been shown to be an efficient non-gradient based technique for the design of stiff, or minimum compliance, structures. For the traditional linear-static problem, the algorithm synthesizes or evolves a structure that is equivalent to solving the following problem min x

N i=0

|Si (xi ) − Si∗ |

subject to Kd = F 0<x≤1

(17)

where the field variable state S being operated on is compliance. The idea is to drive the state of each CA to a specified target. In the HCA method for minimum compliance design, the density state of each cellular automaton is modified so that the elements in the design domain have uniform compliance. The rules used in this chapter that govern material distribution are control-based. In the case of the design for maximum stiffness, a monotonically decreasing relationship occurs between mass and compliance (Tovar 2004). An inversely proportional relationship exists between elastic modulus and compliance, i.e., when a load is applied to an elastic structure, as its modulus decreases, the compliance increases. Therefore, in the design of stiff structures, mass must be added to reduce the compliance of an element; to increase compliance mass is removed. The setpoint directly controls the total mass distributed within the design domain, as there is a one-to-one correspondence between the compliance of the structure under a given load and the total mass of the structure. Adapting the principles of fully stressed design (Haftka, Gürdal, and Kamat 1990), HCA is utilized to allocate material based on the compliance of each element. Although numerous distribution rules can be used (Tovar, Patel, Kaushik, and Renaud 2007), a simple proportional error material update is used here. The change in relative density of element i at the kth iteration can be expressed as (k) (k)

xi = KP (S¯ i − Si∗ (k) )

(18)

A d e c o u p l e d a p p r o a c h t o r e l i a b i l i t y-b a s e d t o p o l o g y o p t i m i z a t i o n

293

(k) where KP is a scaling parameters where S¯ i is the effective field state of a CA, which reflects the average field state of itself and its neighborhood as expressed in Eq. (15). When designing for minimum compliance, Si = ci . Note that the setpoint is not necessarily static and can change from one iteration to the next as explained in the following section.

3.2.1.1

M U LT I P L E L OA D I N G C O N D IT I O N S

When a design problem is posed such that loading can existed in multiple, independent scenarios, an analysis must be performed for each loading condition, or load case. In traditional topology optimization of static-elastic problems, a weighted sum of the compliance or strain energy from each load case for each element is often used to represent the final value (Bendsøe and Sigmund 1989). Thus, the final compliance state for the ith element in the design domain is represented by a weighted sum of the compliance for each load case ci =

NL 

αL ciL

(19)

L=1

where ciL is the compliance of the element for load case L and NL is the total number of load cases. A smaller load can have more of an influence in the final structure by give more weight to that load case through the weight parameter αL . 3.2.1.2

MA S S C O NT R O L

A mass control scheme utilized to generate topologies of a specified mass. To accommodate mass control, the appropriate setpoint must be determined. It has been shown by Tovar (Tovar 2004) that for structural optimization, there exists a direct relationship between the compliance in a structure when loaded and the final mass of the structure. The higher the setpoint, the lower the mass and vice versa. The error in the current field state of a CA and setpoint directly affects the material distribution within the design domain. Therefore, to design for a specific mass, the corresponding setpoint must be must determined. A scheme for finding this target can be accomplished by simply iterating on the HCA update rules and updating the setpoint, as shown in Fig. 11.5, until the correct mass results after applying the design rule expressed in Eq. (18). The setpoint for the k + 1 HCA iteration is found by iterating on the update ⎛ S∗ (j+1) = S∗ (j) ⎝

(k+1)

Mf

Mf∗

⎞ ⎠

(20)

where j is an iterator for the sub-loop on the HCA rules in Eq. (18) and Mf∗ is the mass fraction target. The mass fraction of a design domain is defined as N Mf =

i=1

N

xi

(21)

294

Structural design optimization considering uncertainties

Current design S*(0)(x(k)) Apply HCA material update rules

x(k), S*( j1)

x(k1) Convergence test |M(k1)  Mf*|  ε f

no S*(j)

Update global setpoint

yes New design

Figure 11.5 Illustration of HCA material update for mass control using a setpoint update strategy.

where N is the number of elements in the design domain. When mass of the structure at the kth iteration satisfies the mass target, the resulting material update control loop is terminated and the dynamic analysis is performed on the resulting material distribution for the k + 1 iteration unless the topology has converged based on the stopping criterion. Thus, the mass constraint is enforced at each HCA iteration. 3.2.1.3

D I S P LA C E M E NT C O N ST RA I NT

Using the ability to control mass we can include constraints that are related to mass. For use in the RBTO framework, the HCA method must incorporate a constraint on which we can consider a mode of failure, i.e., a limit state function. Here, we will consider the design of structures that are reliable with respect to the maximum allowable displacement. Koˇcvara (Koˇcvara 1997) developed a linear constraint on displacement using a minimum compliance formulation for the optimization of a truss structure. A bi-level approach was proposed, where the primal goal was to satisfy a displacement constraint and the secondary goal was to minimize compliance. A displacement constraint is formulated for continuum-based topology optimization problems by Deqing et al. (Deqing, Yunkang, Zhengxing, and Huanchun 2000) using dual programming approach. The maximum displacement of a structure is a global behavior. The total mass of structure, a global property, is used to control the displacement of the structure. First, the relationship between displacement and mass must be quantified. Since HCA operates on local information, a single element is studied. Making the assumption of linearly elastic material behavior and constant boundary conditions (e.g., time independent, etc.), we can solve for the nodal displacements using the finite element method formulation that assumes linear elastic material behavior and static loading. The term linear refers to the relationship exists between the strains and stresses. To find the solution

A d e c o u p l e d a p p r o a c h t o r e l i a b i l i t y-b a s e d t o p o l o g y o p t i m i z a t i o n

295

for the displacements, a system must be in equilibrium where the potential energy is at an extremum as 9 stated by the principle of stationary potential energy. The total potential energy ( ) for a structure that has been discretized into finite elements can be expressed in terms of the internal or strain energy (U) and the work done by the external forces (W)


=U−W =

1 T d Kd − FT d 2

(22)

where K is the global stiffness matrix, d is the global vector of nodal displacements, and F is the vector of external global forces applied at each node. The extrema of the total potential energy of the deformable body is expressed by, 9 ∂ =0 ∂d

(23)

From this condition, the resulting equilibrium equation to be solved is, Kd − F = 0

or

Kd = F

(24)

The static loading assumption requires that an independent relationship among the global stiffness matrix, K, the force vector, F, and the displacements, d, exists. Constructing K using the material parametrization described by Eq. (14), the displacements can be solved for in Eq. (24). The relationship for the maximum displacement of all nodal degrees of freedom for a two-dimensional four-node element, as a function of the relative density (or elastic modulus), is shown in Fig. 11.6. For a linear-elastic analysis,

600

500 400 d 300 200 100

0.1

0.2

0.3

0.4

0.5 x

0.6

0.7

0.8

0.9

1

Figure 11.6 The uniaxial relationship between displacement (d) of an element and it relative density (x) by plotting the compression of an single element with unit height and width (E = F = 1).

296

Structural design optimization considering uncertainties

the relationship has the form δmax (x) = C

1 x

(25)

where C is a constant. This inversely proportional relationship is same as that for the “compliance-elastic modulus’’ relationship mentioned previously. Characterizing this local relationship, it is observed that displacement can be controlled through mass. However, since the maximum displacement of a structure is a global behavior, the displacement constraint developed in this work is applied globally by penalizing, or reducing, the mass constraint, which is described above, until the displacement equality constraint is satisfied. The HCA method is used to drive the topology to the minimum mass that satisfies the displacement constraint.

4 Decoupled RBTO formulation In the reliability-based topology optimization (RBTO) method developed in this work, the structural synthesis is performed separately from the gradient-based reliability analysis. The HCA topology optimization occurs in sequence with the reliability analysis. Following a topology optimization using the HCA method, a reliability analysis is performed to find the most probable point of failure (MPP), u∗ , using the performance measure approach (PMA) described in Eq. (9). In the optimization subproblem, the design variable in the topology optimization, i.e., the relative density of each element, is fixed and the uncertain parameters, or random variables, are the design variables. The MPP is returned and used as fixed input parameters for in the topology optimization. Convergence is achieved when a target reliability index is reached. The general form of the RBTO, a minimum compliance problem can be expressed as min c(x) x

subject to Pf (V) = P(G(x, v) < 0) ≤ Pt Kd = F 0≤x≤1 i = 1, . . . , n and j = 1, . . . , m where

(26)

G = − + max

for i density elements and j uncertain variables where Pt is a tolerance on the probability of failure. The limit state function G states that if the performance parameter  is larger than the limit value max then the system fails. To find the reliability index, the limit state function is approximated at each iteration. In this chapter, PMA is employed with a first-order approximation (FORM) of the limit state function. The optimization algorithm is described in Fig. 11.7, where a deterministic HCA update is executed based on the uncertain variables calculated from a reliability analysis performed for each iteration. This process continues until convergence. 4.1

RBT O m et ho d o lo g y

In this methodology, a single limit state constraint on the maximum allowable displacement of the structure is considered. For the example problems considered, the

A d e c o u p l e d a p p r o a c h t o r e l i a b i l i t y-b a s e d t o p o l o g y o p t i m i z a t i o n

297

Start x(0), u(0) Initial density x(t), u(t) Topology optimization min x s.t.

c(x) Ni
1xivi  V 0  x 1

x(t1) u*(t1) min u s.t.

Reliability analysis ⌿ (u, x, h)  ⌿max

||u||
breqd

i

u*(t1)

Convergence test

no

yes End

Figure 11.7 A flowchart of the decoupled approach to reliability-based topology algorithm. For the example problems considered, the HCA method is used for the topology optimization to synthesize the structure and then a reliability anaylsis is performed using a maximum displacement constraint ( ≡ δmax ).

performance parameter  and the limit value max in the formulation (26) are the displacements δmax and δ∗max , respectively. Two sets of random variables are considered on this work: the elastic modulus of the material E0 and the loads Fi on the structure. These uncertainties account for material/manufacturing uncertainties and operational uncertainties. Other uncertain parameters may be included. The specific RBTO formulation solved for the design problems presented can be expressed as min c(x) x

subject to g D : Kd = F g R : −δmax + δ∗max ≤ 0 0<x≤1

(27)

298

Structural design optimization considering uncertainties

Starting with a fully dense material in the design domain, i.e., all density variables x are at their upper bound and initial values of the uncertain variables are set to their mean values, a topology optimization is performed subject to a maximum allowable displacement, followed by a reliability analysis. In the analysis, the optimization subproblem in (9) is solved, where the density design parameters are fixed, using uncertain variables to determine the MPP with respect to the constraint imposed. In this implementation, a sequential quadratic programming (SQP) algorithm is utilized to solve for the values of random variables v at the MPP. For the optimization subproblem in the reliability analysis, a warm-start approach is included as an improvement over previous investigations (Patel, Agarwal, Tovar, and Renaud 2005). Therefore, the sub-optimization starts from the MPP of the previous iteration. Fixing the resulting set of uncertain variables, a new topology optimization is executed and the process is repeated until convergence. The algorithm is described below. Let t denote the iteration counter for the global RBTO and k used in the previous sections represent a local counter for the topology optimization. Step 1.

Step 2. Step 3. Step 4. Step 5.

Define the design domain, deterministic material properties, constraint g R , and initial design, x(0) (full density). Define the random material properties and random loading conditions. Initialize the design domain densities and set random variable values V as the fixed design parameters. Perform topology optimization using the HCA method. Perform the reliability analysis to obtain random variable values at the MPP. ∗(t + 1) ∗(t) Check for convergence, | g g∗(t)− g | ≤ ε1 and |(u(t + 1) − u(t) )T (u(t + 1) − u(t) )| ≤ ε2 for the tolerance parameters ε1 and ε2 . If the convergence criteria are satisfied, the final topology is obtained; otherwise, go to Step 2.

5 Numerical examples The RBTO framework presented is applied to two example problems. The first problem is a Michell-type structure that considers a single concentrated load. The second example considers the loading conditions of a three bar truss. A normal distribution for each random variable vi in the reliability analysis, which is expressed by uj =

vj − mvj σvj

(28)

where mvj and σvj denote the mean value and the standard deviation about the jth random variable, respectively. The parameter uj represents the random variable transformed into the standard space. For these examples, the Poisson’s ratio is ν = 0.33. The mean values of the random parameters are used to generate the structural topology for the first RBTO cycle. Furthermore, for these problems the range of uncertainty about the mean is assumed to be 5% of the mean. This may be an unrealistically large uncertainty for the elastic modulus of the material, but the large uncertainty is used to demonstrate the methodology. The design rule scale parameter in Eq. (18) is KP = 0.2 for the topology synthesis.

A d e c o u p l e d a p p r o a c h t o r e l i a b i l i t y-b a s e d t o p o l o g y o p t i m i z a t i o n

5.1

299

Mic hell-type s tructure problem

This example considers a two-dimensional 2 m × 1 m beam structure that is fully constrained at one end and loaded at the free end. The elastic modulus of the material used is E = 200 GPa and a concentrated load of F = 100 N is applied at the midpoint of the lower boundary of the beam as shown in Fig. 11.8. These values represent the mean values that are varied in the reliability analysis component of the methodology to find the MPP. The design domain is discretized into 5000 elements. For this example, a constraint on maximum allowable displacement δmax = 1 × 10−5 m is imposed. The resulting topologies for each algorithm cycle is shown in Fig. 11.9. Around 30 FEA analyses are required for the topology synthesis and 23 analyses are required for the reliability analysis during each RBTO cycle. A summary of the HCA performance with the deflection constraint and the subsequent reliability analyses is tabulated in Table 11.1. As expected, the elastic modulus E is driven to a lower value and the load F is increased to a higher value. In this example, the resulting change in structural characteristics is quite significant. A mass increase of 33% is required to obtain a reliable design as compared to the structure synthesized using the mean values of the random

F Figure 11.8 The 2 m × 1 m design domain with a single load.

(a) Cycle 1: Mf
0.359

(b) Cycle 2: Mf
0.478

Figure 11.9 The resulting topologies for the Michell-type structure after each RBTO cycle for β = 3. Table 11.1 Summary of FEA evaluations and uncertain parameter values for the Michell-type structure. RBTO Cycle

HCA Iters

Reliability FEA evals

E (GPa)

F (kN)

1 2

33 29

23 23

178.1 178.1

−110.2 −110.2

300

Structural design optimization considering uncertainties

b

Mass Mf

b
0 (50%)†

0.359

b
0.5 (69.15%)

0.388

b
1 (84.13%)

0.392

b
2 (97.72%)

0.431

b
3 (99.87%)

0.478

Topology

†

Deterministic (no uncertainty)

Figure 11.10 Comparison of Michell-type structures for varying prescribed levels of reliability with respect to a constraint on the maximum allowable displacement δmax = 1 × 10−5 m. The baseline run (β = 0) is 50% reliable. Table 11.2 Validation of the Michell topologies using a 10,000 sample Monte Carlo simulation. Reliability

Expected

Actual

β = 0.5 β=1 β=2 β=3

69.15% 84.13% 97.72% 99.87%

69.11% 82.94% 97.43% 99.87%

variables. Figure 11.10 shows an increase in mass required to satisfy an increase in prescribed reliability. An excellent correlation between the prescribed level of reliability and the actual reliability of the structures is shown in Table 11.2. This confirms that FORM accurately predicted the failure surface in each case. 5.2 T russ struc t ur e pr o b lem In this example, we will consider the three-bar truss problem subject to three loading conditions presented by Duysinx (Duysinx 1997), shown in Fig. 11.11. The elastic

A d e c o u p l e d a p p r o a c h t o r e l i a b i l i t y-b a s e d t o p o l o g y o p t i m i z a t i o n

301

F1 F2

F3

Figure 11.11 The 2 m × 1 m design domain with three load cases.

(a) Cycle 1: Mf
0.269

(b) Cycle 2: Mf
0.338

Figure 11.12 The resulting topologies for the “three bar’’ structure after each RBTO cycle for β = 3. Table 11.3 Summary of FEA evaluations and uncertain parameter values for the three bar problem. RBTO Cycle

HCA Iters

Reliability FEA evals

E (GPa)

F1 (kN)

F2 (kN)

F3 (kN)

1 2

28 27

47 47

177.0 177.0

−100.0 −200.0

−299.9 −300.0

−438.4 −438.5

modulus of the material used is E = 200 GPa. The mean values for the three load cases are as follows: −10 kN, −30 kN, and −40 kN. The maximum displacement constraint is considered with respect to all load cases, i.e., the constraint is violated if the maximum displacement exceeds the allowable value for any loading. The design domain is discretized into 5000 elements. For this example, a constraint on maximum allowable displacement δmax = 1 × 10−5 m is imposed. The resulting topologies generated at each algorithm cycle is shown in Fig. 11.12, where we see that the RBTO algorithm converges after 2 cycles. Fewer than 30 FEA analyses are required for the topology synthesis and fewer than 40 analyses are required for the reliability analysis during each RBTO cycle. A summary of the HCA performance with the deflection constraint and the reliability analyses is tabulated in Table 11.3. It is observed that the load cases F1 and F2 do not change after each

302

Structural design optimization considering uncertainties

Reliability

Mass Mf

b
0 (50%)†

0.269

b
0.5 (69.15%)

0.278

b
1 (84.13%)

0.288

b
2 (97.72%)

0.309

b
3 (99.87%)

0.338

Topology

†

Deterministic (no uncertainty)

Figure 11.13 Comparison of “three bar’’ structures for varying levels of reliability with respect to a constraint on the maximum allowable displacement δmax = 1 × 10−5 m. The baseline run ( β = 0) is 50% reliable.

reliability analysis since the third load case F3 dominates. Although the evolution of the structure is more subtle in this problem, the topologies following the initial reliability analyses distribute approximately 26% more mass compared to the initial structure synthesized using the mean values of the random variables. Again, mass is the driver to generating a more reliable design, as illustrated in Fig. 11.13. Using the Monte Carlo simulation to validate the reliability of each structure, it was determined the first-order approximation used in the reliability analysis accurately predicts the failure surface in this problem. These results are tabulated in Table 11.4.

6 Conclusions In this chapter, a new methodology for reliability-based topology optimization (RBTO) is presented. A decoupled approach for reliability-based design optimization is combined with the HCA method for structural synthesis. The objective could be generalized

A d e c o u p l e d a p p r o a c h t o r e l i a b i l i t y-b a s e d t o p o l o g y o p t i m i z a t i o n

303

Table 11.4 Validation of the “three bar’’ topologies using a 10,000 sample Monte Carlo simulation. Reliability

Expected

Actual

β = 0.5 β=1 β=2 β=3

69.15% 84.13% 97.72% 99.87%

69.11% 82.94% 97.43% 99.87%

and applied to different types of problems. The reliability formulation used in this investigation is known as the performance measure approach (PMA) where the reliability index β is included as a constraint in this subproblem and the random variables are driven to the most probable point of failure (MPP) for the current structural design with respect to a displacement constraint. The MPP is required satisfy the specified reliability index β. This RBTO methodology facilitates structural designs that are reliable with respect to a specified performance parameter. Using a maximum allowable displacement constraint as a failure mode, we observe that the reliable topology requires more mass for each example problem, compared to the initial deterministic topology. For both example cases, where uncertainties in the elastic modulus and applied loads were considered, only two algorithm cycles are required for convergence to a design. The excellent correlation shown between the prescribed and actual radiabilities demonstrates the First-Order Reliability Method (FORM) is sufficient for use with problem that use the static-elastic assumptions. The inclusion of the Hybrid Cellular Automata (HCA) method adds to the efficiency of the proposed methodology for the design of structures using the static-elastic assumptions as previous investigations. The use of the HCA method in the RBTO framework could show great benefit other design problems, such as aeroelastic design, where gradients are not easily computed for the topology synthesis. However, the FORM approximation must be investigated for use with other design problems.

References Agarwal, H. & Renaud, J.E. 2006. A new decoupled framework for reliability based design optimization. AIAA Journal 44(7):1524–1531. Allen, M. & Maute, K. 2004. Reliability-based optimization of aeroelastic structures. Struct. Multidisc. Optim. 27:228–242. Bendsøe, M.P. 1989. Optimal shape design as a material distribution problem. Comp. Mth. Appl. Mech. Engrg. 1:193–202. Bendsøe, M.P. & Kikuchi, N. 1988. Generating optimal topologies in optimal design using a homogenization method. Comp. Mth. Appl. Mech. Engrg. 71:197–224. Bendsøe, M.P. & Sigmund, O. 1989. Topology Optimization: Theory, Methods and Applications. Springer-Verlag, Berlin. Breitung, K. 1984. Asymptotic approximations for multinormal integral. Journal of Engineering Mechanics 110(3):357–366.

304

Structural design optimization considering uncertainties

Burks, A. 1970. Essays on Cellular Automata, Chapter Von Neumann’s self-reproducing automata, pp. 3–64. University of Illinois Press. Choi, K.K., Youn, B.D. & Yang, R. 2001. Moving least square method for reliability-based design optimization. In The Fourth World Congress of Structural and Multidisciplinary Optimization (WCSMO-4). Deqing, Y., Yunkang, S., Zhengxing, L. & Huanchun, S. 2000. Topology optimization design of continuum structures under stress and displacement constraints. Applied Mathematic and Mechanics 21:1–26. Duysinx, P. 1997. Layout optimization: A mathematical programming approach. Technical Report DCAMM report No. 540, University of Liege. Eschenaueuer, H.A. & Olhoff, N. 2001. Topology optimization of continuum structures: A review. Applied Mechanics Reviews 54:331–390. Frangopol, D.M. 1998. Probabilistic structural optimization. Progress in Structural Engineering and Materials 1(2):223–230. Frangopol, D.M. & Maute, K. 2003. Life-cycle reliability-based optimization of civil and aerospace structures. Computers and Structures 81:397–410. Gurdal, Z. & Tatting, B. 2000. Cellular automata for design of truss structures with linear and nonlinear response. In Proceedings of the 41st AIAA Structures, Strucutural Dynamics, and Materials Conference, Number 2000-1580, 2000, April 3–6, Atlanta, Georgia. Haftka, R.T., Gurdal, Z. & Kamat, M.P. 1990. Elements of Structural Optimization. Kluwer Academic Publishers, Dordrecht, The Netherlands, 2nd ed. Hajela, P. & Kim, B. 2001. On the use if energy minimization for ca based analysis in elasticity. Struct. Multidisc. Optim. 23:24–33. Haldar, A. & Mahadevan, S. 2001. Probability, Reliability and Statistical Methods in Engineering Design. New York: Wiley. Inou, N., Shimotai, N. & Uesugi, T. 1994. Cellular automaton generating topological structures. In 2nd European Conference on Smart Structures and Materials, Number 2361-08, 1994, October Glasgow, United Kingdom, pp. 47–50. Inou, N., Uesugi, T., Iwasaki, A. & Ujihashi, S. 1998. Self-organization of mechanical structure by cellular automata. Fracture and Strength of Solids 145(9):1115–1120. Kharmanda, G., Olhoff, N., Mohamed, A. & Lemaire, M. 2004. Reliability-based topology optimization. Struct. Multidisc. Optim. 26:295–307. Kita, E. & Toyoda, T. 2000. Structural design using cellular automata. Struct. Multidisc. Optim. 19:64–73. Kocvara, M. 1997. Topology optimization with displacement constraints: a bilevel programming approach. Struct. Optim. 4:256–263. Liu, P.-L. & Kiureghian, A.D. 1991. Optimization algorithms for structural reliability. Structural Stafety 9:161–177. Maute, K. & Frangopol, D.M. 1998. Reliability-based design of mems mechanisms by topology optimization. Computers and Structures 81:813–824. Mogami, K., Nishiwaki, S., Izui, K., Yoshimura, M. & Kogiso, N. 2006. Reliability-based structural optimization of frame structures for multiple failure criteria using topology optimization techniques. Struct. Multidisc. Optim. 32(4):299–311. Moses, F. 1973. Desing for reliability: concepts and applications. Wiley. Murotsu, Y. & Shao, S. 1989. Optimum shape design of truss structures based on reliability. Struct. Multidisc. Optim. 2(2):65–76. Patel, N.M., Agarwal, H., Tovar, A. & Renaud, J.E. 2005. Reliability based topology optimization using the hybrid cellular automaton method. In 1st AIAA Multidisciplinary Design Optimization Specialist Conference, 2005, April 18–21, Austin, Texas. Royset, J.O., Kiureghian, A.D. & Polak, E. 2001. Reliability-based optimial structural design by the decoupled approach. Reliability Engineering and Systems Safety 73:213–221.

A d e c o u p l e d a p p r o a c h t o r e l i a b i l i t y-b a s e d t o p o l o g y o p t i m i z a t i o n

305

Rozvany, G.I.N. 1997. Topology Optimization in Structural Mechanics. Springer. Rozvany, G.I.N., Bendsoe, M.P. & Kirsh, U. 1995. Optimality criteria: a basis for multidisciplinary optimization. Appl. Mech. Rev. 48:41–119. Slotta, D., Tatting, B., Watson, L., Gurdal, Z. & Missoum S. 2002. Convergence analysis for cellular automata applied to truss design. Engineering Computations 19(8):953–969. Svanberg, K. 1987. The method of moving asymptotes a new method for structural optimization. Int. J. Numer. Meth. Engrg. 24:359–373. Tovar, A. 2004. Bone Remodeling as a Hybrid Cellular Automaton Optimization Process. Ph.D. thesis, University of Notre Dame. Tovar, A., Patel, N.M., Kaushik, A.K. & Renaud, J.E. 2007. Optimality conditions of the hybrid cellular automata for structural optimization. AIAA Journal. Tovar, A., Quevedo, W., Patel, N. & Renaud, J.E. 2006. Topology optimization with stress and displacement constraints using the hybrid cellular automaton method. In Proceedings of the 3rd European Conference on Computational Mechanics, 2006, June 5–8, Lisbon, Portugal. Wolfram, S. 2002. A New Kind of Science. Wolfram Media. Xie, Y.M. & Stevens, G. 1997. Evolutionary Structural Optimization. Springer-Verlag, London. Yuge, K. & Kikuchi, N. 1995. Optimization of a frame structure subjected to a plastic deformation. Struct. Optim. 10:197–208. Yuge, K., Iwai, N. & Kikuchi, N. 1999. Optimization of 2-d structures subjected to nonlinear deformations using the homogenization method. Struct. Optim. 17(4):286–299. Zhou, M. & Rozvany, G.I.N. 1991. The COC algorithm, part II: Topological, geometrical and generalized shape optimization. Comp. Meth. Appl. Mech. Engrg. 89:309–336.

Chapter 12

Sample average approximations in reliability-based structural optimization:Theory and applications Johannes O. Royset Naval Postgraduate School, Monterey, CA, USA

Elijah Polak University of California, Berkeley, CA, USA

ABSTRACT: This chapter describes recent advances in combining Monte Carlo sampling and nonlinear programming algorithms for reliability-based structural optimization. Specifically, we present an approach where the reliability term in the problem formulation is replaced by a statistical estimate of the reliability obtained by means of Monte Carlo sampling. This replacement introduces a sampling error and gives rise to sample average approximations. The chapter presents rules for adjusting the sample size effectively.

1 Introduction Cost efficient bridges, building frames, aircraft wings, and other mechanical structures can be achieved by formulating and solving nonlinear optimization problems. However, such problems become significantly harder to solve when a structure’s reliability is accounted for in the problem formulation. This difficulty is caused by the fact that the failure probability of most structures, as well as the corresponding gradient with respect to design variables, cannot be computed exactly, but must be approximated. The difficulty is further aggravated by the challenge of deriving suitable expressions for the gradient of the failure probability. One possible approach to overcome these difficulties is to estimate the failure probability and its gradient using Monte Carlo sampling. This chapter describes recent advances in combining Monte Carlo sampling and nonlinear programming algorithms for reliability-based structural optimization. Other approaches for such optimization include (successive) first-order approximations (Madsen and Friis Hansen 1992; Enevoldsen and Sørensen 1994; Kuschel and Rackwitz 2000; Royset et al. 2006), gradient-free heuristics, (Itoh and Liu 1999; Nakamura et al. 2000; Beck et al. 1999), response surfaces (Gasser and Schuëller 1998; Igusa and Wan 2003), and surrogate functions (Torczon and Trosset 1998; Eldred et al. 2002). However, a review of these approaches is beyond the scope of this chapter. Here, we present an approach where the failure probability in the problem formulation is replaced by a statistical estimate obtained by means of Monte Carlo sampling. This replacement introduces a sampling error and gives rise to approximate optimization problems. Such approximate problems are referred to as sample average approximations.

308

Structural design optimization considering uncertainties

Even if a sample average approximation is solvable by some nonlinear programming algorithm to sufficient accuracy, the design obtained might be far from the optimal design of the original problem due to the induced sampling error. This deficiency can be overcome by constructing a sample average approximation with a large sample size, which tends to have a small sample error and hence tends to have optimal designs near the optimal designs of the original problem. Unfortunately, a large sample size implies high computational cost. For example, each sample point may involve a finite element analysis of the structure at hand. Hence, applying a nonlinear programming algorithm to a sample average approximation with a large sample size is usually impractical. On the other hand, as we have already mentioned, a small sample size is computationally less expensive, but may lead to designs far from an optimal one. Intuition and empirical evidence indicate that the following adaptive approach is efficient. Initially, consider a sample average approximation with a small sample size, i.e., a coarse, but inexpensive approximation, and apply some optimization algorithm to achieve a certain amount of design improvement. When this design improvement is achieved, refine the approximation by increasing the sample size and apply the optimization algorithm to this refined approximation. Initiate the optimization from the improved design achieved at the coarser approximation level, i.e., the calculations on the refined approximation is “warm started.’’ Repeat the process until an acceptable design is achieved. This adaptive approach avoids spending excessive computational effort on estimating the failure probability of the relatively poor designs produced by the early iterations of the optimization algorithm. Increasingly large efforts are expended only as better and better designs are achieved and accurate estimates of the failure probability are needed to ensure further design improvements. This approach tends to be efficient because coarse estimates of the failure probability (and its gradient) are usually sufficient to steer an optimization algorithm towards better designs in the early stages of the calculations. We have derived a theory for the described adaptive increase in sample size (Royset and Polak 2007; Polak and Royset 2007). In this chapter, we review some of these theoretical results and show their application to reliability-based structural optimization. Section 2 formally defines the reliability-based structural optimization problem. Section 3 discusses the properties of the failure probability as a function of the design variables and derives an expression for its gradient. The gradient is derived for general structural systems consisting of an arbitrary number of unions and intersections of failure events. A Monte Carlo estimate of this gradient is used to direct the calculations towards better designs. Section 4 describes the basic algorithmic approach, which involves approximately solving a sequence of sample average approximations with increasing sample size. Section 5 presents sample-adjustment rules that ensure computational efficiency and theoretical convergence. Clearly, a rapid increase in sample size may result in many algorithmic iterations on computationally costly sample average approximations. In fact, too rapid increase in sample size may prevent convergence to a solution. By contrast, a slow increase in sample size may lead to unnecessarily many iterations with coarse approximations. Hence, it is important to balance the increase of sample size with the progress of the optimization algorithm towards an optimal design. We present two different sample-adjustment rules: (i) a feedback rule specifying an increase in sample size whenever the optimization algorithm’s progress falls below a threshold value

S a m p l e a v e r a g e a p p r o x i m a t i o n s i n r e l i a b i l i t y-b a s e d s t r u c t u r a l o p t i m i z a t i o n

309

and (ii) the solution of an auxiliary optimization problem that determines the “optimal’’ sample size at every iteration using estimated values of rate of convergence, computational cost, distance to optimal design, and sampling error. In Section 6, we illustrate the theoretical results with three numerical examples arising in design of various mechanical structures. Structures with both a single and with multiple limit-state functions are considered and reliability terms are included in both objective and constraint functions. We also present an example with a nontraditional objective: determine several “good’’ designs that are significantly different. This objective is useful when qualitative factors such as practical, esthetic, social, and political requirements are especially important. In such situations, the designer may seek to generate several “good’’ designs with respect to quantitative factors (e.g., cost and reliability) using some optimization algorithm and then select among these designs using his or her judgment regarding the other, qualitative factors. Finally, the concluding remarks of this study are presented in Section 7.

2 Problem formulation Consider the design of a mechanical structure such as a bridge, a building frame, or an aircraft wing. Let x be an n-dimensional vector of design variables, for example related to the size and form of the structure, and let c0 (x) and c(x) be the initial and failure costs, respectively, of the structure given design x. Furthermore, let p(x) be the failure probability of the structure, given design x, to be defined precisely below. Then, the reliability-based design optimization problem takes the form min{c0 (x) + c(x)p(x)|p(x) ≤ q, x ∈ X} x

(1)

where q is a bound on the failure probability, X is a constraint set for x defined in terms of J constraint functions fj (x), j ∈ J = {1, 2, . . . , J}, i.e., X = {x|fj (x) ≤ 0, j ∈ J}

(2)

The objective function in (1) consists of the initial cost plus the expected cost of failure. The constraint functions represent restrictions on shape and form of the structure, amount and location of steel reinforcement, as well as other factors. We assume that there are no integer restrictions on the design variables x. We also assume that the constraint and cost functions are fairly simple functions, e.g., analytic expressions, that can easily be evaluated. Hence, the challenge is associated with the failure probability p(x). When the failure cost is positive, i.e., c(x) > 0 for all x ∈ X, (1) is equivalent to the following problem min{c0 (x) + c(x)x0 |p(x) ≤ x0 , 0 ≤ x0 ≤ q, x ∈ X} x0 ,x

(3)

where x0 is an auxiliary design variable (Royset et al., 2006). The transformation from (1) to (3) is beneficial for numerical reasons; the multiplication of a presumably large failure cost c(x) with a presumably small, inaccurately estimated failure probability p(x) in (1) may cause numerical difficulties. Hence, we always recommend solving (3) instead of (1). Consequently, we focus primarily on problems with a deterministic

310

Structural design optimization considering uncertainties

objective function and a failure probability constraint as in (3). To simplify the notation and without loss of generality, we consequently consider the problem: P : min{c(x)|p(x) ≤ q, x ∈ X} x

(4)

where c(x) is some deterministic objective (cost) function. Mechanical structures are assessed using one or more performance measures, e.g., displacement and stress levels at various locations in the structure. In this chapter, we consider the general case of “system failure’’ where the (system) failure probability is defined by a collection of performance measures. Specifically, failure occurs when certain combinations of the performance measures are unsatisfactory. Let gk (x, u), k ∈ K = {1, 2, . . . , K}, be a collection of K limit-state functions describing the relevant performance measures. The functions gk (x, u) depend on the design x and the realization u of a standard normal random m-vector U. This random vector incorporates the uncertainty in the structure and its environment. Note that a limit-state function given in terms of multivariate normal (possibly with correlation) and lognormal random vectors can be transformed into one defined in terms of a standard normal vector using a smooth bijective mapping. A limit-state function given in terms of random vectors governed by other distributions can also be transformed, possibly by introducing an approximation. Hence, the limitation to a multivariate standard normal distribution is in many applications not restrictive (see e.g. Chapter 7 of (Ditlevsen and Madsen 1996) and (Liu and Kuo 2003; Akgul and Frangopol 2003; Holicky and Markova 2003)). By convention, gk (x, u) ≤ 0 represents unsatisfactory performance of the k-th measure. Hence, we define the failure probability of the structure as p(x) = P[F(x)], where the failure domain ': {gk (x, U) ≤ 0} (5) F(x) = i∈I k∈Ci

with Ci ⊂ K and I = {1, 2, . . . , I} defining the combinations of performance measures that lead to structural failure. For example, suppose that a structure is defined with three limit-state functions, i.e., K = 3, representing stress level (g1 (x, u)), displacement at location A ((g2 (x, u)), and displacement at location B (g3 (x, u)). Also, suppose that the structure is defined to have failed if the first performance measure (stress level) is unsatisfactory, regardless of the displacement levels, and it is also defined to have failed if both of the displacement measures are unsatisfactory, regardless of the stress level. In this case, I = {1, 2}, C1 = {1}, and C2 = {2, 3}, i.e., F(x) = {g1 (x, U) ≤ 0} ∪ ({g2 (x, U) ≤ 0} ∩ {g3 (x, U) ≤ 0})

(6)

3 Failure probability and gradient Since P, see (4), is a nonlinear optimization problem it would be natural to apply a standard nonlinear programming algorithm to this problem. However, such an approach requires two assumptions to be satisfied. First, all the functions in P must be at least once differentiable (with respect to the design variables x) with continuous gradients. We refer to this assumption as the smoothness assumption. Second, we must be able to

S a m p l e a v e r a g e a p p r o x i m a t i o n s i n r e l i a b i l i t y-b a s e d s t r u c t u r a l o p t i m i z a t i o n

311

compute, relatively easily, all the functions and their gradients for any given design x. We refer to this assumption as the computability assumption. Since we assume that the cost function c(x) and constraint functions fj (x) are all analytic functions or in some other form satisfying our two assumptions (smoothness and computability), the challenge is associated with the failure probability p(x). Due to the complicated form of p(x) it is not clear whether the assumptions are satisfied. In fact, it appears unlikely that the computability assumption is satisfied due to the m-dimensional integral in the definition of p(x). It is also difficult to perceive situations under which the smoothness assumption is satisfies when the limit-state functions are not differentiable. Hence, we assume throughout this chapter that the limit-state functions gk (x, u) are differentiable with respect to both arguments and have continuous gradients. If the limit-state functions are not differentiable and/or the design variables are restricted to integers, then the theory and algorithms derived in this chapter are not applicable. This section rewrites the expression for the failure probability in a form that is equivalent, under weak assumptions, to the original definition. As seen in the following, this effort results in an expression that satisfies the smoothness assumption and that lends itself to estimation of both the failure probability and its gradient. This effort would have been unnecessary if we were only interested in computing the failure probability and not in optimization. Standard Monte Carlo sampling (possibly with importance sampling) would have sufficed in such a situation (see, e.g., (Ditlevsen and Madsen 1996)). However, within an optimization algorithm we also need the gradient of the failure probability and the gradient is not easily available from the definition of the failure probability. In (Uryasev 1995), we find a theoretical expression for the gradient of the failure probability. However, this expression may involve surface integrals, which are difficult to estimate in practice. In (Marti 1996) (see also (Marti 2005)), an integral transformation is presented, which, when it exists, leads to a simple expression for the gradient of the failure probability. However, it is not clear under what conditions this transformation exists. As in (Uryasev 1995), (Tretiakov 2002) assumes that the failure domain F(x) is bounded and given by a union of events. With this restriction, an expression for the gradient of the failure probability involving integration over a simplex is derived. In principle, this integral can be evaluated by Monte Carlo sampling. However, to the authors’ knowledge, there is no computational experience with estimation of failure probabilities for highly reliable mechanical structures using this expression. In (Royset and Polak, 2004a; Royset and Polak, 2004b) we find expressions for the failure probability and its gradient that can be estimated by Monte Carlo and importance sampling. However, the expressions are limited to the case with one performance measure (i.e., K = 1). In Section 9.2 of (Ditlevsen and Madsen 1996), an expression for the gradient of the failure probability is suggested, without a complete proof, for the case with one performance measure. This expression is based on a form of p(x) that has been found computationally efficient in applications. In (Royset and Polak 2007) a generalization of this special-case formula was derived and formally proven. We proceed by describing the expression for the failure probability given in (Royset and Polak 2007). It can be shown that the failure domain is equivalently expressed as   (7) F(x) = min max gk (x, U) ≤ 0 i∈I

k∈Ci

312

Structural design optimization considering uncertainties

As in (Deak 1980) (see alternatively (Ditlevsen et al., 1987; Bjerager 1988), and Section 9.2 of (Ditlevsen and Madsen 1996)), we observe the following fact: If the standard normal random vector U = RW and R2 is Chi-square distributed with m degrees of freedom, then W is a random vector, independent of R, uniformly distributed over the surface of the m-dimensional unit hypersphere. Note that W represents a direction and R a positive length. Hence, we obtain from the total probability rule that      (8) p(x) = E P min max gk (x, RW) ≤ 0W i∈I

k∈Ci

Here, P[{mini∈I maxk∈Ci gk (x, RW) ≤ 0|W}] is the conditional probability of a failure event in the random direction W for a given x. This conditional probability takes a particular simple form if the safe domain, i.e., the complement of the failure domain F(x)c , is “star-shaped.’’ A safe domain is star-shaped if in any direction w one passes from the safe to the failure region only once when moving from the origin in the u-space1 in the direction of w; see (Royset and Polak 2007) for a mathematically precise definition. When the safe domain is star-shaped, the expression inside the expectation 2 2 2 (r (x, W)), where χm ( · ) is the Chi-square cumulative distribution in (8) equals 1 − χm function with m degrees of freedom and r(x, W) is the minimum distance in direction W from the origin of the u-space to the surface of F(x). This distance can expressed in terms of the minimum distances in direction W from the origin to the surface of {gk (x, RW) ≤ 0}, k ∈ K. Let rk (x, W) denote this distance. Then, r(x, W) = min max rk (x, W) i∈I

k∈Ci

(9)

2 Since χm ( · ) and the square function (positive domain) are strictly increasing, we find that

p(x) = E[φ(x, W)]

(10)

where 2 2 (rk (x, W))} φ(x, W) = max min{1 − χm i∈I

k∈Ci

(11)

This is the new expression for the failure probability we will use in the following. As noted earlier, this expression is equivalent to the original definition of the failure probability under the assumption of a star-shaped safe domain, see (Royset and Polak 2007) for a proof. We observe that when the safe domain is not star-shaped, (10) may overestimate the failure probability. Hence, it is conservative to assume a star-shaped safe domain. For a given design x, it is possible to obtain an indication whether the star-shape assumption is satisfied by computing an estimate N j=1 IF (x) (uj )/N of p(x), where u1 , u2 , . . . , uN are realizations of independent, identically distributed standard normal vectors and IF (x) (uj ) = 1 if uj ∈ F(x), and zero otherwise. If this estimate is significantly smaller than 1 We refer to the m-dimensional space of realizations of U as the “u-space.’’

S a m p l e a v e r a g e a p p r o x i m a t i o n s i n r e l i a b i l i t y-b a s e d s t r u c t u r a l o p t i m i z a t i o n

313

the one of (10), then the star-shape assumption is violated. We also note that equivalent assumptions were adopted by (Tretiakov 2002; Ditlevsen et al., 1987; Bjerager 1988) and Section 9.2 of (Ditlevsen and Madsen 1996). The main advantage of the new expression for the failure probability (10) over the original expression p(x) = P[F(x)] is that a useful expression for the gradient of the failure probability can be derived. At first glance, it appears that the gradient of p(x) in (10) is simply the expectation of the gradient of φ(x, W) with respect to x. However, closer examination shows that φ(x, W) is not differentiable with respect to x due to ˆ its max-min form. Hence, we define the set of active limit-state functions K(x, W) as those limit-state functions that define the surface of the failure domain F(x) in the direction W. More precisely, ˆ ˆ i (x, W), i ∈ I(x, ˆ W)} K(x, W) = {k ∈ K|k ∈ C

(12)

where ˆ W) = I(x, ˆ i (x, W) = C

   max r (x, W) = max r (x, W) i ∈ I min k k


(13)

   k ∈ Ci  max rk
(x, W) = rk (x, W)

(14)

 

i ∈I k∈Ci


k∈Ci

k
∈Ci

ˆ Using the definition of the set of active limit-state functions K(x, W), we derive the subgradient of φ(x, W) as  ∂φ(x, W) = conv

ˆ k∈K(x,W)

2fχm2 (r2k (x, W))rk (x, W)

∇x gk (x, rk (x, W)W) ∇u gk (x, rk (x, W)W)T W

 (15)

where conv{·} denotes the convex hull, fχm2 ( · ) is the Chi-square probability density function with m degrees of freedom, and ∇x gk (x, u) and ∇u gk (x, u) the gradient of gk (x, u) with respect to x and u, respectively. Informally, the expression in the brackets 2 2 (rk (x, W)) obtained using implicit of (15) is the gradient with respect to x of 1 − χm differentiation. This leads to the following expression for the gradient of the failure probability (see (Royset and Polak 2007) for a proof) ∇p(x) = E[dφ (x, W)]

(16)

where dφ (x, W) is any element of the subgradient ∂φ(x, W). We note that (16) is only valid if the safe domain is bounded, i.e., the minimum distance in every direction w from the origin of the u-space to the surface of F(x) is bounded from above by some (large) number. This may not always be the case in applications. However, it is always possible to define an artificial limit-state function gK+1 (x, U) = ρ − U, with a sufficiently large ρ > 0, replace I by I + 1, and set CI = {K + 1}. Then, F(x) satisfies the assumption about a bounded safe domain. This is equivalent to enlarging the failure domain. The probability associated with the enlarged failure domain is slightly larger than the one associated with the original failure domain. The difference, however, is no

314

Structural design optimization considering uncertainties

2 greater than 1 − χm (ρ2 ) and therefore negligible for sufficiently large ρ. Consequently, this boundedness assumption is not restrictive in practice. From the above derivation we see that the failure probability is differentiable with a continuous gradient given by (16), i.e., the failure probability satisfies the required smoothness assumption for nonlinear optimization. However, for this to have any practical value, we also need to be able to compute the failure probability and its gradient, i.e., we need the computability assumption to be satisfied. Clearly, (10) and (16) cannot, in general, be evaluated analytically, but must be estimated by Monte Carlo sampling. Let w1 , w2 , . . . , wN be a set of N sample points, each generated by independent sampling from the uniform distribution on the m-dimensional unit hypersphere. Given this sample, we define the estimate of (10):

pN (x) =

N 

φ(x, wj )/N

(17)

j=1

Since W corresponds to a direction, this type of Monte Carlo simulation is referred to as directional sampling (Bjerager 1988). It is well-known (see, e.g., (Rubinstein and Shapiro 1993) for a proof) that pN (x) converges to p(x) uniformly over any closed and bounded set, as N → ∞. Hence, at least in principle, we can obtain an accurate estimate of the failure probability by computing (17) with a large N. (Of course, a large sample size may be prohibitive computationally.) We now consider an estimate of the gradient (16). Since φ(x, w) is not differentiable with respect to x, we see that pN (x) is generally not differentiable either. However, since φ(x, w) has a subgradient, see (15), it can be shown that pN (x) also has a subgradient denoted by ∂pN (x), see (Royset and Polak 2007). This subgradient is given by ∂pN (x) =

N 

∂φ(x, wj )/N

(18)

j=1

see (15) for the expression for ∂φ(x, wj ). It is shown in (Royset and Polak 2007) that the subgradient ∂pN (x) converges (shrinks) to ∇p(x) uniformly over any closed and bounded set, as N → ∞. We note that there is typically no need to estimate the subgradient ∂pN (x), but only one of its elements. To generate such an element, proceed as follows: (i) obtain N sample points w1 , w2 , . . . , wN , (ii) for each sample point wj deterˆ mine one active limit-state function, i.e., find one element in K(x, wj ), and compute the numerical value of the vector within the brackets of (15) for the active limit-state function, and (iii) average the numerical values over all the sample points.

4 Algorithm based on sample average approximations In this section, we follow (Royset and Polak 2007; Polak and Royset 2007) and present an algorithm that utilizes the sample average estimates of the failure probability and its gradient derived in the previous section, see (17) and (18). The algorithm carries out nonlinear optimization iterations on a sequence of sample average approximations for

S a m p l e a v e r a g e a p p r o x i m a t i o n s i n r e l i a b i l i t y-b a s e d s t r u c t u r a l o p t i m i z a t i o n

315

the original problem P. Given the sample points w1 , w2 , . . . , wN , we define the sample average approximation of P as the following optimization problem: PN : min{c(x)|pN (x) ≤ q, x ∈ X} x

(19)

It is noted that the only difference between P and PN is that p(x) has been replaced by its sample average. Intuitively, PN becomes a better approximation to P as N increases. In fact, under weak assumptions, a global minimum of PN converges to a global minimum of P, as N → ∞, see (Royset and Polak 2007) and more generally Chapter 6 of (Ruszczynski and Shapiro 2003) and references therein. Since we can evaluate pN (x) for a given sample, PN satisfies our computability assumption. However, PN does not satisfy our smoothness assumption since (17) is generally not differentiable – it only has a subgradient (18). Hence, standard nonlinear programming algorithms may perform poorly when applied to PN . As seen in Subsection 5.1 below, we are able to overcome this difficulty by utilizing the fact that P satisfies the smoothness assumption. In this section, we proceed under the assumption that there is some optimization algorithm that can effectively be applied to PN . As discussed in Section 1, the simplest scheme for approximately solving P would be to select some sample size N and apply some optimization algorithm to PN for a number of iterations. The obtained design would be an estimate of the optimal design of P. However, this may be a poor estimate if the sample size is small, and if the sample size is large, the computational cost may be prohibitive. In (Royset and Polak 2007; Polak and Royset 2007), the following adaptive scheme is proposed. Conceptual Algorithm for Solving P. Step 0. Select an initial design x0 , an initial sample size N, and sample w1 , w2 , . . . , wN . Set iteration counter j = 0. Step 1. Consider the sample average approximation PN and compute a new design xj+1 by carrying out one iteration of some optimization algorithm applied to PN . This iteration is initialized by the current design xj . Step 2. Use some sample-adjustment rule and determine if the sample size should be augmented. If the sample size should be augmented, replace N by some larger N and generate additional sample points to complement the existing sample points. Step 3. Replace j by j + 1, and go to Step 1. The conceptual algorithm describes an adaptive scheme, but does not specify how Steps 1 and 2 can be implemented. What optimization algorithm can be used in Step 1? What sample-adjustment rule should be used in Step 2? At first glance, the first question appears easier. However, as discussed above, PN may not satisfy the smoothness assumption and standard nonlinear programming algorithms may perform poorly. In fact, as we will see in Subsection 5.1 below, care must be taken when selecting the optimization algorithm in Step 1 to ensure convergence of the overall algorithm. The second question appears to be difficult and embodies the following fundamental trade-off. A rapid increase in sample size may result in many iterations with large N and hence high computational cost. As we see in Subsection 5.1 below, there is also a theoretical concern; a rapid increase in sample size may prevent convergence to an optimal design.

316

Structural design optimization considering uncertainties

On the contrary, a slow increase in sample size may lead to unnecessarily many iterations on coarse sample average approximations. The next section discusses two approaches for implementing Step 2. We also briefly discuss the implementation of Step 1. There is also a third question that is not addressed in the conceptual algorithm: when to stop the calculations? As in all nonlinear programming, this is a fundamentally difficult questions that is substantially aggravated by the presence of sample averages. A simple approach would be to augment the sample size until it reaches a “sufficiently large’’ number, e.g., an N that results in a coefficient of variation for pN (x) of less than 5%. Then, keep that sample size for a number of iterations until the optimization algorithm in Step 1 ceases to make substantial progress from iteration to iteration. Another approach is to simply run the algorithm until the dedicated time is consumed. Techniques for checking whether a given design is close to optimal includes statistical testing, see e.g. Section 6.4 of (Ruszczynski and Shapiro 2003). A further discussion of stopping criteria and solution quality is beyond the scope of this chapter.

5 Selection of sample sizes The conceptual algorithm presented in the previous section needs a sample-adjustment rule (see Step 2). There are two main concerns when constructing a sample-adjustment rule: (i) theoretical convergence and (ii) computational efficiency. This section presents two different rules. The first rule satisfies (i), but its efficiency is sensitive to input parameters. The second rule has weaker convergence properties, but allocates samples optimally in some sense. 5.1 F eed b ac k r ule The first sample-adjustment rule for Step 2 of the conceptual algorithm augments the sample size when the progress of the optimization algorithm in Step 1 is sufficiently small. This rule is motivated by the following observation: when the optimization algorithm in Step 1 is making small progress towards an optimal design of the current sample average approximation PN , the current design is probably near that optimal design. Hence, there is little to be gained from computing even better designs for PN ; it is better to increase the sample size N and start to calculate with a more accurate sample average approximation. In (Royset and Polak 2007), the progress of the optimization algorithm in Step 1 is measured in terms of a function FN (x
, x

) defined by FN (x
, x

) = max{c(x

) − c(x
) − γψN (x
)+ , ψN (x

) − ψN (x
)+ }

(20)

where

  ψN (x) = max pN (x) − q, max fj (x) j∈J

(21)

ψN (x)+ = max{0, ψN (x)}, and the parameter γ > 0. The function FN (x
, x

) measures how much “better’’ the design x

is compared to design x
. Suppose that x
is a feasible design for PN . Then, ψN (x
) ≤ 0 and ψN (x
)+ = 0 and, hence, FN (x
, x

) = max{c(x

) − c(x
), ψN (x

}

(22)

S a m p l e a v e r a g e a p p r o x i m a t i o n s i n r e l i a b i l i t y-b a s e d s t r u c t u r a l o p t i m i z a t i o n

317

We see that if FN (x
, x

) ≤ −ω, with ω being some positive number, then the objective function in PN for design x

is reduced with at least the amount ω compared to the value for design x
. Additionally, x

is feasible for PN because ψN (x

) ≤ −ω. Suppose that x
is not a feasible design for PN . Then, ψN (x
) > 0. When FN (x
, x

) ≤ −ω, the constraint violation for PN at x

is reduced with at least the amount ω compared to the value at x
because ψN (x

) − ψN (x
) ≤ −ω. The above observation leads to the following sample-adjustment rule: If FN (xj , xj+1 ) is no larger than a threshold, then the progress is sufficient and the current sample size is kept. (Note that FN (xj , xj+1 ) is a negative number and that it measures the decrease in cost or constraint violation. Hence, a large negative number corresponds to a large progress towards an optimal design.) If FN (xj , xj+1 ) is larger than the threshold, then the progress is too small and the sample size is increased. The challenge with this rule is to determine an appropriate threshold. In (Royset and Polak 2007), we find a sample-size dependent threshold that results in the following sample-adjustment rule: If

 τ FN (xj , xj+1 ) > −η ( log log N)/N

(23)

then augment the sample size. Otherwise, keep the current sample size in the next iteration. Here, η is a positive parameter and τ is a parameter strictly between 0 and 1. Since the threshold is increasing (approaches zero from below) with increasing sample size, the rule becomes successively more stringent. For large N, the sample size is only increased if the optimization algorithm in Step 1 of the conceptual algorithm makes a tiny progress (FN (xj , xj+1 ) is close to zero). This means that for large N, it is necessary to solve the sample average approximation to near optimality before the sample size is increased. On the other hand, for small N, the sample size is increased even if the optimization algorithm in Step 1 is making a relatively large progress. Hence, the rule avoids having to solve low-precision sample average approximations to high accuracy before switching to a larger sample size. But, the rule eventually forces the algorithm to solve high-precision sample average approximations to high accuracy. The double logarithmic form of the threshold in (23) relates to the Law of the Iterated Logarithm (see (Royset and Polak 2007) and references therein). It is shown in (Royset and Polak 2007) that this exact form of the sample-adjustment rule guarantees convergence of the conceptual algorithm when implemented with a specific optimization algorithm in Step 1. This specific optimization algorithm is motivated by the Polak-He algorithm (see Section 2.6 of (Polak 1997)) and takes the following form. For any current design xj and current sample size N, the next iteration xj+1 = xj + λN (xj , d)hN (xj , d)

(24)

where d is any element in the subgradient ∂pN (xj ), see (18) and its subsequent paragraph, and the stepsize λN (xj , d) is given by Armijo’s rule: λN (xj , d) =

max {βk |FN (xj , xj + βk hN (xj , d)) ≤ βk αθN (xj , d)}

k∈{0,1,2,...,}

(25)

318

Structural design optimization considering uncertainties

Here, α ∈ (0, 1] and β ∈ (0, 1) are parameters, and θN (x, d) = − min{zT bN (x) + zT BN (x, d)T BN (x, d)z/(2δ)} z∈Z

with parameter δ > 0, the J + 2-dimensional unit simplex Z given by ⎫ ⎧  J+2 ⎨  ⎬ Z = z  zj = 1, zj ≥ 0, ∀j ⎭ ⎩  j=1

(26)

(27)

the (J + 2)-dimensional vector (γ as in (20)) bN (x) = (γ ψN (x)+ , ψN (x)+ − pN (x) + q, ψN (x)+ − f1 (x), . . . , ψN (x)+ − fJ (x))T (28) and the n × (J + 2)-matrix BN (x, d) = (∇c(x), d, ∇f1 (x), . . . , ∇fJ (x))

(29)

Finally, the search direction hN (xj , d) = −BN (xj , d)ˆz/δ

(30)

where zˆ is any optimal solution of (26). The problem in (26) is quadratic and can be solved in a finite number of iterations by a standard QP-solver (e.g. Quadprog (Mathworks, Inc. 2004)). Usually, the one-dimensional root finding problems in the evaluation of rk (x, w), needed in (15), cannot be solved exactly in finite computing time. One possibility is to introduce a precision parameter that ensures a gradually better accuracy in the root finding as the algorithm progresses. Alternatively, we can prescribe a rule saying that the root finding algorithm should terminate after CN iterations, with C being some constant. For simplicity, we have not discussed the issue of root finding. In fact, this issue is not problematic in practice. The root finding problems can be solved in a few iterations with high accuracy using standard algorithms. Hence, the root finding problems are solved with fixed precision for all iterations in the algorithm giving a negligible error. The feedback rule (23) requires the user to determine the parameters η and τ as well as the amount of sample size increase. To avoid a quick increase in sample size and corresponding high computational costs, the parameter τ is typically set to 0.9999. However, it is nontrivial to determine an efficient value for the parameter η. If η is large, then the sample size tends to be augmented frequently. Hence, η should be small to avoid costly sample average approximations in the early iterations. However, too small η may result in an excessive number of iterations for each sample average approximation. Overall, in Section 6 we see empirically that the numerical value of η may influence computing times significantly. Furthermore, neither the conceptual algorithm nor the feedback rule specify how much the sample size should be increased – only when to increase it. Typically, the user specifies a rule of the form: replace N by ξN, with ξ > 1, whenever the sample size needs to be increased. Naturally, the

S a m p l e a v e r a g e a p p r o x i m a t i o n s i n r e l i a b i l i t y-b a s e d s t r u c t u r a l o p t i m i z a t i o n

319

computationally efficiency may vary with the amount increased each time. We note that (Royset and Polak 2007) proves that the conceptual algorithm with the sampleadjustment rule (23) and the optimization algorithm (24) is guaranteed to converge to a solution for any τ ∈ (0, 1), η > 0, and sample size increase. Hence, the above discussion only relates to how fast the algorithm will converge. As indicated in the previous paragraph, it can be difficult to select efficient values for the parameter η as well as an efficient sample size increase every time the algorithm is prompted by the sample-adjustment rule. Typically, some numerical experimentation and parameter tuning for the problem at hand is needed. In the next subsection, we describe an alternative, more complex sample-adjustment rule that avoids such tuning. 5.2

Ef f icient s cheme

In this subsection, we present the sample-adjustment scheme given in (Polak and Royset 2007), which modifies a methodology originally developed in (He and Polak 1990). Instead of having a simple sample-adjustment rule as in Subsection 5.1 for Step 2 of the conceptual algorithm, the scheme in (Polak and Royset 2007) consists of a precalculation step that determines the “optimized’’ sample size for subsequent iterations. In the pre-calculation step, the user selects a required accuracy of the final design (e.g., a feasible design with cost within 5% of the minimum cost) and solves an auxiliary optimization problem that determines the sample size for each iteration (e.g., 100, 100, 100, 200, 200, 300, etc., sample points, for iterations 1, 2, 3, 4, 5, 6, etc., respectively). Hence, whenever the conceptual algorithm reaches Step 2, it simply looks up the prescribed sample size from the output of the auxiliary optimization problem. The objective function of the auxiliary optimization problem, to be derived below, is the total computational work needed to obtain a solution of required accuracy, and the constraint is that the required cost reduction be achieved. Let a stage be a number of iterations carried out by the conceptual algorithm for a constant sample size. The decision variables in the auxiliary problem are (i) the number of stages, s, to be used, (ii) the sample size Ni to be used in stage i, i = 1, 2, . . . , s, and (iii) the number of iterations ni to be carried out in stage i. For example, 100, 100, 100, 200, 200, and 300 sample points, for iterations 1, 2, 3, 4, 5, and 6 respectively, correspond to three stages, with stage 1 consisting of three iterations (n1 = 3) and sample size 100 (N1 = 100), stage 2 consisting of two iterations (n2 = 2) and sample size 200 (N2 = 200), and stage 3 consisting of one iteration (n3 = 1) and sample size 300 (N3 = 300). While the number of stages s has to be treated as an integer variable, the variables Ni and ni can be treated as continuous variables and rounded at the end of their optimization. In practice, it turns out that the optimal number of stages s∗ hardly ever exceeds 10, with 3–7 being a most likely range for s∗ . Incidentally, if one assigns the number of stages to be s > s∗ , and then solves the reduced auxiliary optimization problem for the Ni and ni , the optimal solution will consist of several Ni being equal, so that the total number of distinct stages is s∗ . The auxiliary problem depends on a sampling-error bound, on the initial distance to the optimal value, and on the rate of convergence of the optimization algorithm applied in Step 1 of the conceptual algorithm. All of these may have to be estimated. As a result, it may be presumptuous to call the solution of the auxiliary optimization

320

Structural design optimization considering uncertainties

problem an “optimal strategy,’’ and hence we will call it an “efficient strategy.’’ As we will see from our numerical results, despite the use of estimated quantities, the efficient strategy is considerably more effective than the obvious alternatives. 5.2.1 Au xi li a ry o p t im iza t io n p r o b le m We begin by deriving the auxiliary optimization problem. First we penalize the constraint in P to convert it into an equivalent, unconstrained min-max problem. This simplifies the derivation since it avoids distinguishing between feasible and infeasible design. For a given parameter π > 0, we define c˜ (x) = c(x) + π max{0, p(x) − q, f1 (x), f2 (x), . . . , fJ (x)}

(31)

c˜ N (x) = c(x) + π max{0, pN (x) − q, f1 (x), f2 (x), . . . , fJ (x)}

(32)

and the unconstrained problem P˜ : min c˜ (x)

(33)

x

We refer to π as a penalty since it adds a positive number to the objective functions c(x) and cN (x) for any infeasible design x. If P is calm (see, e.g., (Burke 1991; Clarke 1983)) and π is sufficiently large, then the design x is a local minimizer of P˜ if and only if it is a local minimizer of P. Similarly, the unconstrained problem P˜ N : min c˜ N (x)

(34)

x

is equivalent to PN for sufficiently large π. An appropriate penalty π can be selected using well-known techniques such as the one in Section 2.7.3 of (Polak 1997). The implementation of such techniques is beyond the scope of this chapter and we assume in the following that a sufficiently large penalty π > 0 has been determined so that optimal solutions of P˜ and P˜ N are feasible for P and PN , respectively. As above, we assume that each sample point is independently generated and that sample points are reused at later stages, i.e., for all stages i = 2, 3, . . . , s, the sample at stage i consists of the Ni−1 sample points at stage i − 1 and of Ni − Ni−1 new, independent sample points. To construct an auxiliary optimization model for determining the number of stages, the sample size at each stage, and the number of iterations to be performed at each stage, we introduce the following assumptions. Suppose that the optimization algorithm in Step 1 of the conceptual algorithm is linearly convergent with a rate of convergence coefficient independent of the sample size in the sample average approximations. That i is, for any stage i and iteration j, the costs of the design at the next iteration, xj+1 , and i ∗ the current design, x , relate to the cost of the optimal design x of P˜ N as follows: j

i ∗ ∗ c˜ Ni (xj+1 ) − c˜ Ni (xN ) ≤ θ(˜cNi (xji ) − c˜ Ni (xN )) i i

Ni

i

(35)

where θ ∈ (0, 1) is the rate of convergence coefficient. Hence, every iteration of the optimization algorithm reduces the remaining distance to the optimal value by a factor θ.

S a m p l e a v e r a g e a p p r o x i m a t i o n s i n r e l i a b i l i t y-b a s e d s t r u c t u r a l o p t i m i z a t i o n

321

Many optimization algorithms including the Pshenichnyi-Pironneau-Polak Min-Max Algorithm (see Section 2.4.1 of (Polak 1997)) are linearly convergent. Next, we assume that for any design x the sampling error is given by |˜cN (x) − c˜ (x)| ≤ (N)

(36)

where (N) is a strictly decreasing positive function with (N) → 0, as N → ∞. We return to the form of (N) below, but for now we only assume that such a function exists. To simplify the notation, we deviate from the numbering scheme of the conceptual algorithm and let j note the iteration number of the current stage (and not from the beginning). Then, xji is the design at iteration j of the i-th stage. Hence, we plan to compute the designs x01 , x11 , . . . , xn11 on stage 1, x02 , x12 , . . . , xn22 on stage 2, . . . , and , i.e., designs x0s , x1s , . . . , xns s on stage s. To make use of “warm’’ starts, we set x0i = xni−1 i−1 the last design of the current stage is taken as the initial design of the next stage. ∗ Let x∗ and xN be optimal designs for P˜ and P˜ N , respectively. Then, in view of (36) we have that ∗ ∗ c˜ (x∗ ) ≤ c˜ (xN ) ≤ c˜ N (xN ) + (N) ∗ ) c˜ N (xN

∗

∗

≤ c˜ N (x ) ≤ c˜ (x ) + (N)

(37) (38)

We refer to the distance between the cost c˜ (x) of some design x and the cost c˜ (x∗ ) of an optimal design x∗ for P˜ as the cost error of design x. Here, the term “error’’ refers to the discrepancy between x and x∗ . For any stage i = 1, 2, . . . , s, we define the cost error after the last iteration of the i-th stage by ei = c˜ (xni i ) − c˜ (x∗ )

(39)

Also let e0 = c˜ (x01 ) − c˜ (x∗ ). Using (36)–(38) and (35), we obtain that for all i = 1, 2, . . . , s, ∗ ei ≤ c˜ Ni (xni i ) − c˜ Ni (xN ) + 2 (Ni ) i ∗ )] + 2 (Ni ) ≤ θ ni [˜cNi (x0i ) − c˜ Ni (xN i

≤θ

ni

[˜c(xni−1 ) i−1

∗

− c˜ (x )] + 4 (Ni )

≤ θ ni ei−1 + 4 (Ni )

(40) (41) (42) (43)

Hence, es ≤ e0 θ k0 (s) + 4

s 

θ ki (s) (Ni )

(44)

i=1

 where ki (s) = sl=i+1 nl if i < s and ki (s) = 0 if i = s. We observe that (44) gives an upper bound on the cost error after completing s stages with ni iterations and Ni sample points at stage i. As shown in (Polak and Royset 2007), the cost error is guarantee to vanish as the number of stages s increases to infinity. This shows that such gradual sample

322

Structural design optimization considering uncertainties

size increase can lead to asymptotic convergence. This is a valuable result, but in this subsection we aim to determine efficient sample-adjustment schemes, i.e., schemes that minimize the computing time to reach a specific reduction in cost error from an initial value. To be able to construct efficient sample-adjustment schemes we need to quantify the computational effort associated with one iteration of the optimization algorithm used in Step 1 of the conceptual algorithm as a function of the sample size N. Suppose that this computational effort is given by the positive function w(N) for any design x. We are now ready to present the auxiliary optimization problem. Given an initial cost error e0 > 0 and a required fractional reduction in cost error  ∈ (0, 1), we seek to determine the number of stages s as well as sample sizes Ni and numbers of iterations ni at each stage i, i = 1, 2, . . . , s, such that the computational effort to reach a cost error of e0 is minimized. We note that the cost error is the discrepancy between the cost of the current design and the cost of the optimal design ˜ In view of (44), this optimization problem takes the following form of P. min

s,ni ,Ni

 s  i=1

s    ni w(Ni )e0 θ k0 (s) + 4 θ ki (s) (Ni ) ≤ e0 i=1

Ni+1 ≥ Ni ,

i = 1, 2, . . . , s − 1

s, ni , Ni integer,



(45)

i = 1, 2, . . . , s

The objective function in D(e0 , ) represents the total computational effort needed to carry out the planned iterations. The first constraint ensures that the cost error has at least been reduced to the required level e0 and the second set of constraints ensures that the sample size is nondecreasing. The estimation of the parameters defining problem D(e0 , ) is discussed in the next section. 5.2.2 Im p lem en t a t io n o f a u x ilia r y o p t im iza t i o n pr o bl e m The auxiliary optimization problem D(e0 , ) involves the work and sampling-error functions w(N) and (N) as well as the rate of convergence parameter θ and the initial cost error e0 = c˜ (x01 ) − c˜ (x∗ ). All these quantities must be determined before D(e0 , ) can be solved. We deal with these issues one at a time. In view of (17) and (18), the computing effort required to evaluate pN (x) and an element of the subgradient grows linearly in N. Hence, the work associated with one iteration of the optimization algorithm used in Step 1 of the conceptual algorithm is proportional to N and we set the work function w(N) = N. The (almost sure) sampling error (N) can be determined  using the Law of the Iterated Logarithm, see (Royset and Polak 2007). However, ( log log N)/N is a pessimistic estimate of the sampling error “typically’’ experienced. Since our goal is to determine efficient number of stages, sample sizes, and numbers of iterations, it appears √ to be more reasonable to assume that the sampling error is proportional to 1/ N as proposed by classical estimation theory: For a given design x, it follows under weak assumption from the Central Limit Theorem that pN (x) is approximately normally

S a m p l e a v e r a g e a p p r o x i m a t i o n s i n r e l i a b i l i t y-b a s e d s t r u c t u r a l o p t i m i z a t i o n

323

distributed with mean p(x) and variance σ(x)2 /N for large N, where σ(x)2 = Var[φ(x, W)]. Hence, for sufficiently large N, √ P[|pN (x) − p(x)| ≤ 1.96σ(x)/ N] ≥ 0.95 (46) However, we are primarily interested in the difference between c˜ N (x) and c˜ (x). Since the max-function in (32) only makes the variance less, it follows that P[|˜cN (x) − c˜ (x)| ≤ (N)] ≥ 0.95 (47) √ when (N) = 1.96πσ(x)/ N. This error expression appears to be appropriate for our auxiliary optimization problem, and we set √ (48)

(N) = 1.96π max σ(x)/ N where the maximization is over all designs examined in a preliminary calculation described below. We determine σ(x), θ, and e0 in an estimation phase consisting of n0 iterations of the optimization algorithm in Step 1 of the conceptual algorithm applied to P˜ N0 , with 0 N0 being a small sample size. Let {xj0 }nj=0 be the iterates computed in this estimation phase. Each time pN0 (x) is computed, the corresponding variance σ(x)2 is estimated by σ(x) = 2

N0 

(φ(x, wj ) − pN0 (x))2 /(N0 − 1)

(49)

j=1

We always retain the largest σ(x)-value computed and use that in the calculation of

(N), see (48). The rate of convergence parameter θ is estimated by the solution of the following ∗ least-squares problem, where the optimal value c˜ N0 (xN ) of P˜ N0 is also estimated: 0 min θˆ ,ˆc

n0 

[(ˆc + (˜cN0 (x00 ) − cˆ )θˆ j ) − c˜ N0 (xj0 )]2

(50)

j=0

This least-square problem minimizes the squared error between the calculated cost at each iteration c˜ N0 (xj0 ) and the nonlinear model cˆ + (˜cN0 (x00 ) − cˆ )θˆ j . The nonlinear model estimates that the cost of the design at iteration j is the optimal cost cˆ plus the initial cost error c˜ N0 (x00 ) − cˆ reduced by a factor. The factor is simply the rate of convergence coefficient raised to the power of the number of iterations. Using the results of the ∗ ) by cˆ . Finally, we (coarsly) least square calculations, we estimate θ by θˆ and c˜ N0 (xN 0 1 ∗ estimate the initial cost error e0 = c˜ (x0 ) − c˜ (x ) by eˆ 0 = c˜ N0 (x00 ) − cˆ . We have now established procedures for estimating all the unknown quantities in D(e0 , ). D(e0 , ) is a nonlinear integer program that appears difficult to solve directly, but this fact can be circumvented by the following observations. First, the restriction of D(e0 , ) obtained by fixing s to a number in the range 5–10 tends to be insignificant since more than 5–10 stages is rarely advantageous and fewer than 5–10 stages is still effectively allowed in the model by setting Ni = Ni+1 for some i. Second, Ni , and to some extent also ni , tend to be large integers. Hence, a continuous relaxation

324

Structural design optimization considering uncertainties

with rounding of the optimal solutions to the nearest integers is justified. In view of these observations, D(e0 , ) can be solved approximately using a standard nonlinear programming algorithm. 5.2.3 Overa ll a lg o r it h m w it h e f f icie n t s a m p l e-adj us t me nt s c he me We now summarize our approach and discuss how the auxiliary optimization problem can be integrated in an algorithm for solving P. As indicated above, the process of solving the auxiliary optimization problem must be preceded by an estimation phase where parameters are determined. This leads to the following overall algorithm for solving P approximately. Algorithm with Efficient Sample-Adjustment Scheme. Parameters. Number of iterations in estimation phase n0 , sample size in estimation phase N0 , maximum number of stages s, and constraint penalty π > 0. Data. Required fractional reduction in cost error  > 0, initial design x00 , and independent sample points w1 , w2 , . . . . Step 0. Compute variance estimate σ(x00 )2 using (49). Step 1. For j = 0 to n0 − 1, perform: 0 by starting from xj0 and carrying out Sub-step 1.1. Compute the next design xj+1 one iteration of some optimization algorithm applied to PN0 . 0 Sub-step 1.2. Compute the variance estimate σ(xj+1 )2 using (49). Step 2. Set σˆ equal to the largest variance estimate encountered in Steps 0 and 1. Step 3. Determine θˆ and cˆ as√the optimal solution of (50). ˆ Step 4. Set (N) = 1.96πσ/ ˆ N, and determine ni and Ni by solving  s s     ˆ i ) ≤ ˆe0 min θˆ ki (s) (N ni Ni ˆe0 θˆ k0 (s) + 4 ni ,Ni

i=1

i=1

Ni+1 ≥ Ni ,

i = 1, 2, . . . , s − 1 

ni , Ni ≥ 1,

i = 1, 2, . . . , s

(51)

Step 5. For i = 1 to s, perform: Sub-step 5.1. Set the first design of the current stage equal to the last design of the previous stage, i.e., x0i = xni−1 . i−1 i Sub-step 5.2. For j = 0 to ni − 1, compute the next design xj+1 by starting from xji and carrying out one iteration of some optimization algorithm applied to PNi . We note that the optimization algorithm used in Sub-Step 1.1 should be identical to the one used in Sub-Step 5.2 since the former sub-step is used to estimate the behavior of the latter. However, any nonlinear programming algorithm can be used in Steps 3 and 4. The proposed algorithm consists of three phases: estimation of parameters (Steps 0–3), solution of auxiliary optimization problem (Step 4), and main iterations (Step 5). This represents the simplest implementation of our idea. Alternatively, we can adopt a moving-horizon approach, where Step 5 is completed only for i = 1, followed by Step 4, then by Step 5 for i = 1 again, followed by Step 4, etc. Hence,

S a m p l e a v e r a g e a p p r o x i m a t i o n s i n r e l i a b i l i t y-b a s e d s t r u c t u r a l o p t i m i z a t i o n

325

the sample-adjustment plan is re-optimized after each stage, which may lead to an improved plan. With re-optimization, it is also possible to re-compute σ, ˆ using all previous iterates, as well as θˆ and cˆ . Other implementations can also be imagined. In the following numerical study, we adopt the simple implementation described above.

6 Numerical examples We illustrate our sample-adjustment approaches using three numerical examples. The examples are implemented in Matlab 7.0 (Mathworks, Inc. 2004) on a 2.8 GHz PC running Microsoft Windows 2000. 6.1

Feedbac k rule and efficient s chem e

This subsection presents a comparative study of the two sample-adjustment approaches given in Section 5. The numerical results of this subsection were reported in (Polak and Royset 2007). Ex a mple 1 The first example arises in the optimal design of a short structural column with a rectangular cross section of dimensions x1 × x2 . Hence, x = (x1 , x2 ) is the design vector. The column is subjected to bi-axial bending moments V1 and V2 , which, together with the yield strength V3 of the material, are considered to be independent, lognormally distributed random variables. The column is also subject to a deterministic axial force af . This gives rise to a failure probability p(x) = P[{G(x, V) ≤ 0}]

(52)

where the random vector V = (V1 , V2 , V3 ) and G(x, V) is a limit-state function defined by G(x, V) = 1 −

4V1 4V2 − 2 − 2 x1 x2 V3 x1 x2 V3

af x1 x2 V3

2 (53)

As discussed in Section 2, this limit-state function can be transformed into one give in terms of a standard normal vector U. Let g1 (x, U) be this transformed limit-state function. Since the resulting safe domain is not bounded, we introduce an auxiliary limit-state function g2 (x, U) = ρ − U, where ρ = 6.5 in this example. (This introduces negligible error.) Then, we redefine the failure probability of the structure as p(x) = P[{g1 (x, U) ≤ 0} ∪ {g2 (x, U) ≤ 0}]

(54)

which is in the form considered in this chapter. We seek a design of the column which satisfies the constraints defined by f1 (x) = −x1 , f2 (x) = −x2 , f3 (x) = x1 /x2 − 2, f4 (x) = 0.5 − x1 /x2 , f5 (x) = x1 x2 − 0.175, and minimize p(x). This is problem (1) with c0 (x) = 0, c(x) = 1, and J = 5. As discussed above, pN (x) does not satisfy the smoothness assumption. Hence, care must be taken when selecting an optimization algorithm for Step 1 in the conceptual algorithm or in Sub-Steps 1.1 and 5.2 in the algorithm with efficient sample-adjustment scheme. For simplicity in

326

Structural design optimization considering uncertainties

these numerical tests, we ignore the fact that the smoothness assumption may be violated and use the Pshenichnyi-Pironneau-Polak Min-Max Algorithm (see Section 2.4.1 of (Polak 1997)) as the optimization algorithm for solving PN . No detrimental behavior of the Pshenichnyi-Pironneau-Polak Min-Max Algorithm was observed because of this simplification. (Note that since p(x) is smooth, pN (x) is, for practical purposes, effectively smooth for large N.) The parameters for the algorithm with efficient sample-adjustment scheme were selected to be n0 = 25, N0 = 50, s = 5, and π = 2. We note that π = 2 suffices to ensure feasibility. Finally, the required fractional reduction in cost error was  = 0.01 and the √ √ initial point was chosen to be x00 = ( 0.175, 0.175). The auxiliary optimization problem yielded a sample-adjustment strategy of three stages with 25, 8, and 8 iterations, with sample sizes 50, 251, and 1621, respectively, which was executed in 458 seconds. Note that this computing time includes the estimation phase (30 seconds) and the solution time of the auxiliary optimization problem (3 seconds). For comparison, we also solve the problem using the feedback rule of Subsection 5.1 to adjust the sample size. We experiment with the thresholds −η

τ

 ( log log N)/N

(55)

and √ −η/ N

(56)

for determining if the progress is “small’’ in Step 1 of the conceptual algorithm. We note that (55) is the same as in (23). This threshold formula guarantees convergence as proven in (Royset and Polak 2007). The threshold in (56) leads to a heuristic algorithm, but offers the advantage that the threshold tends to zero faster for increasing N as compared to (55). In the numerical tests, we set τ = 0.9999. As mentioned above, it is difficult to select and effective value of η, so we experiment with a range of values. Furthermore, we must determine how much the sample size should increase when prompted by the sample-adjustment rule. In this example, we selected five stages with sample sizes equally spaced between the minimum and maximum sample sizes given by the auxiliary optimization problem, i.e., 50, 443, 836, 1228, and 1621. We used the same random seed in both algorithms. We ran the algorithm with the feedback rule until c˜ 1621 ( · ) was equal to the cost achieved in the last iteration of the algorithm with the efficient scheme. We did not augment the sample size beyond 1621, but continued computing iterates at that stage until the target costvalue was achieved. This is a somewhat favorable stopping criterion for the algorithm with the feedback rule because this algorithm might augment, prematurely, the sample size beyond 1621 resulting in long computing times. The computing times for the algorithm with the feedback rule are summarized in Table 12.1 for various values of the parameter η and for the two threshold formulae (55) and (56). In Table 12.1, the row with η = ∞ gives the computing time for a fixed sample size equal to the largest sample size 1621 for all iterations.

S a m p l e a v e r a g e a p p r o x i m a t i o n s i n r e l i a b i l i t y-b a s e d s t r u c t u r a l o p t i m i z a t i o n

327

Table 12.1 Computing times [seconds] for the algorithm with feedback rule for sample adjustment as applied to Example 1. The algorithm with efficient sample-adjustment scheme computes the same design in 458 seconds. η

∞ 10−1 10−2 10−3 10−4 10−5 10−6 10−7 10−8 10−9

Threshold (55)

(56)

980 1044 1084 678 675 682 476 574 603 898

980 1036 654 675 677 676 477 554 601 901

As seen from Table 12.1, a fixed sample size can result in poor computing times compared to an adaptive scheme using a feedback rule. However, in the adaptive schemes there is a trade-off between solving the approximating problems accurately at an early stage (i.e., using small η), potentially wasting time, and solving the early approximations too coarsely (i.e., using large η), leading to many iterations at stages with high computational cost. In the efficient sample-adjustment scheme of Sub-Section 5.2, the trade-off is balanced by solving the auxiliary optimization problem. In the feedback rule, the user needs to consider the trade-off manually by selecting a value for the parameter η. If the right balance is found, i.e., a good η, then the feedback rule can be efficient. In fact, the feedback rule with η = 10−6 is only marginally slower than the efficient scheme. Of course, it is difficult to select η a priori. To illustrate this difficulty, we repeated the example for the higher accuracy  = 0.005. Then, the efficient scheme increased the sample size up to 6473 and solved the problem in 1461 seconds. From Table 12.1 it appears that η = 10−6 is a good choice. We selected this value and re-solved the problem using the feedback rule with five stages equally spaced in the range [50, 6473] as above. The computing time turned out to be 4729 seconds. Hence, η = 10−6 was not efficient in this case. Exa mple 2 The second example considers the design of a simply supported reinforced concrete T-girder for minimum cost according to the specifications in (American Association of State Highway and Transportation Officials 1992), using the nine design variables x = (As , b, hf , bw , hw , Av , S1 , S2 , S3 ), where As is the area of the tension steel reinforcement, b is the width of the flange, hf is the thickness of the flange, bw is the width of the web, hw is the height of the web, Av is the area of the shear reinforcement (twice the cross-section area of a stirrup), and S1 , S2 and S3 are the spacings of

328

Structural design optimization considering uncertainties Table 12.2 Computing times [seconds] for the algorithm with feedback rule for sample adjustment as applied to Example 2. The algorithm with efficient sample-adjustment computes the same design in 1001 seconds. η

∞ 10−2 10−3 10−4 10−5 10−6 10−7

Threshold (55)

(56)

>36000 >12600 2004 2256 6721 1209 11108

>36000 7416 1990 2342 2327 1608 >7200

shear reinforcements in the high, medium, and low shear force zones of the girder, respectively. We model uncertainty using eight independent random variables collected in a vector V. We assumed that the girder can fail in four different modes corresponding to bending stress in mid-span and shear stress in the high, medium, and low shear force zones. Structural failure occurs if any of the four failure modes occur. This gives rise to four nonlinear, smooth limit-state functions Gk (x, V), k = 1, 2, 3, 4, whose exact form is rather complicated and is given in (Royset et al. 2006). This results in a failure probability p(x) = P[ ∪4k=1 {Gk (x, V) ≤ 0}]. As above, these limit-state functions can be transformed into ones given in terms of a standard normal vector U. Let gk (x, U) be these transformed limit-state functions. Since the resulting safe domain is not bounded, we introduce an auxiliary limit stage function g5 (x, U) = ρ − U, where ρ = 10 in this example. (This introduces negligible error.) Then, we redefine the failure probability of the structure as   5 ' p(x) = P {gk (x, U) ≤ 0} (57) k=1

which is in the form considered in this chapter. We also imposed 24 deterministic, nonlinear constraints as described in (Royset et al. 2006). Algorithm parameters were selected to be n0 = 50, N0 = 50, s = 5, and π = 1. Finally, the required fractional reduction in cost error  = 0.0001 and the initial point x00 = (0.01, 0.5, 0.5, 0.5, 0.5, 0.0005, 0.5, 0.5, 0.5) were chosen. The algorithm with the efficient sample-adjustment scheme gave three stages with 65, 20, and 20 iterations, with sample sizes 50, 373, and 2545, respectively. The total computing time was 1001 seconds. Again we compared this result with that obtained using the algorithm with the feedback rule. Here, we use five stages of equally spaced sample sizes between 50 and 2545. Using the same stopping criterion as for the first example, we obtained the computing times in Table 12.2. We observe that the computing times using the feedback rule can be significantly longer than those achieved using the efficient scheme. We also

S a m p l e a v e r a g e a p p r o x i m a t i o n s i n r e l i a b i l i t y-b a s e d s t r u c t u r a l o p t i m i z a t i o n

329

4

3

6

5

8.66 m

7

2

1 L 10 m

10 m

Figure 12.1 Truss for Example 3.

note that an approach with a fixed sample size of 2545 for all iterations takes more than 10 hours (see the first row in Table 12.2). 6.2 Alternative objective functions We conclude this chapter by demonstrating how our solution methodology can also solve other problems than P (and (1) and (3)). Typically, engineers need to account for not only quantitative factors such as cost and reliability, but also esthetic, social, and political requirements. Most esthetic, social, and political requirements are qualitative in nature and cannot easily be incorporated into numerical models. Even quantitative factors may not fully represent reality due to imprecise models and lack of data. In this subsection, we show how multiple optimization models can be formulated and solved to account for this situation. We adopt an approach originally proposed in (Brill Jr. 1979) for public sector planning: determine a small set of design alternatives that satisfy the stated requirements, are “good’’ with respect to the stated objective, and are also dispersed in the design space. Instead of searching for one optimal design or an efficient frontier, as in singleand multi-objective objective optimization, respectively, this approach seeks several design alternatives (e.g., 3–12) that the engineer and the decision maker can further assess using qualitative objectives. As pointed out in (Brill Jr. 1979), the best design from the perspective of the decision maker may not be located on the efficient frontier, as assumed by a multi-objective optimization formulation, due to the fact that not all objectives are included in the multi-objective formulation. Furthermore, by seeking a dispersed set of design alternatives, the engineer and decision maker are presented with a wide range of alternatives which may stimulate new considerations and ideas about designs, objectives, and constraints. See also (White 1996; Drezner and Erkut 1995) for similar approaches. We illustrate this approach with an example. Ex am ple 3 Consider the simply supported truss in Figure 12.1. The truss is subject to a random load L in its mid-span. L is lognormally distributed with mean 1000 kN and standard

330

Structural design optimization considering uncertainties

deviation 400 kN. Let Sk be the yield stress of member k. Members 1 and 2 have lognormally distributed yield stresses with mean 100 N/mm2 and standard deviation 20 N/mm2 . The other members have lognormally distributed yield stresses with mean 200 N/mm2 and standard deviation 40 N/mm2 . The yield stresses of members 1 and 2 are correlated with correlation coefficients 0.8. However, their correlation coefficients with the other yield stresses are 0.5. Similarly, the yield stresses of members 3–7 are correlated with correlation coefficients 0.8, but their correlation coefficients with the yield stresses of members 1 and 2 are 0.5. The load L is independent of the yield stresses. Let V = (S1 , S2 , . . . , S7 , L). The design vector x = (x1 , x2 , . . . , x7 ), where xk is the cross-section area (in 1000 mm2 ) of member k. The truss fails if any of the members exceed their yield stress. (We ignore the possibility of buckling.) This gives rise to seven limit state functions: Gk (x, V) = Sk xk − L/ζk ,

k = 1, 2, . . . , 7

(58)

where ζk is factor given by √ the geometry and loading of√the truss. From Figure 12.1, we determine that ζk = 1/(2 3) for k = 1, 2, and ζk = 1/ 3 for k = 3, 4, . . . , 7. Using a Nataf distribution (see (Ditlevsen and Madsen 1996), Section 7.2), we transform these limit-state functions into limit-state functions given in terms of a standard normal random vector U. Let gk (x, U) denote these transformed limit-state functions. Since the resulting safe domain is not bounded, we introduce an auxiliary limit state function g8 (x, U) = ρ − U, where ρ = 20 in this example. (This introduces negligible error.) Then, we redefine the failure probability of the structure as  p(x) = P

8 '

 {gk (x, U) ≤ 0}

(59)

k=1

which is in the form considered in this chapter. We impose the constraint that the failure probability should be no larger than 0.001350, i.e., p(x) ≤ q = 0.001350. We also impose the 14 deterministic constraints 0.5 ≤ xk ≤ 2, k = 1, 2, . . . , 7, that limit the allowable area of each member to be between 500 mm2 and 2000 mm2 . We initially seek a design of the truss that minimizes the cost of  the truss, i.e., we aim to solve P. Since all members are equally long, the cost c(x) = 7k=1 xk . We use the conceptual algorithm implemented with the feedback rule (23) for sample-adjustment, with parameters η = 0.002 and τ = 0.9999, and optimization algorithm (24) for Step 1, with parameters α = 0.5, β = 0.8, γ = 2, and δ = 1. The sample size is initially 375 and is increased by a factor of 4 every time it is prompted by the sample-adjustment rule. However, the sample size is not increased beyond 24000. We start the calculations with initial design x0 = (1.000, 1.000, . . . , 1.000) and stop when a feasible solution for P24000 is found. The resulting design is given in the first row of Table 12.3. With the motivation that a decision maker may want to be presented with a small set of good designs, from which he or she may select, we formulate an optimization model that generates substantially different designs. Specifically, suppose that we have a set of existing design alternatives xd , d ∈ D. Let cˆ be the smallest cost over all existing design alternatives, i.e., cˆ = mind∈D c(xd ). Then, the following optimization model provides a

S a m p l e a v e r a g e a p p r o x i m a t i o n s i n r e l i a b i l i t y-b a s e d s t r u c t u r a l o p t i m i z a t i o n

331

Table 12.3 Alternative designs for Example 3. The first row gives the optimal design, but the subsequent rows are at most 10% more costly. Design

Dispersion

x1

x2

x3

x4

x5

x6

x7

1.138 1.169 1.982 1.124 1.121 1.123 1.122 1.123 1.087

1.156 2.000 1.164 1.146 1.145 1.147 1.146 1.146 1.595

1.118 1.089 1.100 1.110 1.113 1.107 1.944 1.106 1.536

1.107 1.096 1.100 1.946 1.108 1.109 1.109 1.108 1.104

1.119 1.096 1.102 1.109 1.109 1.947 1.109 1.110 1.107

1.113 1.103 1.104 1.111 1.949 1.111 1.110 1.110 1.119

1.108 1.091 1.092 1.100 1.100 1.101 1.104 1.941 1.098

– 0.8451 0.8449 0.8393 0.8367 0.8286 0.8269 0.8331 0.6085

design that is no more costly than aˆc, with a > 1, and that is as “different’’ compared to the existing designs xd as possible: max{x0 |p(x) ≤ q, x ∈ X, c(x) ≤ aˆc, x − xd  ≥ x0 , d ∈ D} x0 ,x

(60)

Here, x0 is an auxiliary design variable that we seek to maximize. The last set of constraints in (60) ensures that the difference (measured in the Euclidean distance) between the new design x and the existing designs xd are all no smaller than x0 . Hence, (60) maximizes the smallest difference between a new design and the existing designs, while ensuring that the new design is feasible and no more costly than aˆc. We note that (60) is in the form P (after redefining the cost and constraint functions) and, hence, it can be solved by the conceptual algorithm described in Section 4. Using the same algorithm parameters as in the beginning of this example, we obtain the designs reported in Table 12.3. In this table, the first row reports the optimal design. The second row is obtained by solving (60) with a = 1.1 and D consisting only of the design in the first row. We observe that the design in the second row is substantially different than the one in the first row, even though it is no more than 10% more costly. The last column of Table 12.3 shows that the second design lies 0.8451 “away’’ from the first design measured in the Euclidean distance. The remaining rows in Table 12.3 are computed in a similar manner, but with D now consisting of all the designs in the rows above. We note that all the designs cost no more than 10% more than the minimum cost. It is seen from Table 12.3 that the minimum cost design (row 1) distributes the material evenly between the different members. However, good designs can also be achieved by selecting one of the members to have cross-section area close to 2 (rows 2–8). Moreover, good designs can be found by setting two members to approximately 1.5 (last row). Naturally, it becomes harder and harder to find a “different’’ design as the set of existing designs D grows, i.e., the last column of Table 12.3 tends to decrease for later designs. Hence, after some solutions of (60) with steadily increasing D, the designs we generate will not be substantially different

332

Structural design optimization considering uncertainties

compared to the ones already computed. This is an interactive process, which should be ended whenever a useful set of designs have been generated and further calculations will provide only limited insight.

7 Conclusions We have presented an approach for solving reliability-based optimal structural design problems using Monte Carlo sampling and nonlinear programming. The approach replaces failure probabilities in the problems by Monte Carlo estimates with increasing sample sizes, and solves the resulting approximate problems with increasing precision. We have also described rules for adjusting the sample sizes, which ensure theoretical convergence and computational efficiency. The numerical examples show empirically that the sample-adjustment rules can reduce computing times substantially compared with an implementation using a fixed sample size. The approach in this chapter is directed towards reliability-based structural optimization problems where the design variables are not restricted to be integers and the relevant limit-state functions are differentiable with continuous gradients. Furthermore, the approach requires many limit-state function evaluations, which (currently) prevent its application to problems involving, e.g., computationally intensive finite element analysis. We note, however, that the sample-adjustment rules described in this chapter dramatically reduce the number of limit-state function evaluations compared to an approach with a fixed sample size. Consequently, the results of this chapter open the possibility for solving, to high accuracy, many previously intractable reliability-based structural optimization problems.

References Akgul, F. & Frangopol, D.M. 2003. Probabilistic analysis of bridge networks based on system reliability and Monte Carlo simulation. In A. Der Kiureghian, S. Madanat & J.M. Pestana (eds), Applications of Statistics and Probability in Civil Engineering, Rotterdam, Netherlands, pp. 1633–1637. Millpress. American Association of State Highway and Transportation Officials (1992). Standard specifications for highway bridges. Washington, D.C.: American Association of State Highway and Transportation Officials. 15th edition. Beck, J.L., Chan, E., Irfanoglu, A. & Papadimitriou, C. 1999. Multi-criteria optimal structural design under uncertainty. Earthquake Engineering & Structural Dynamics 28(7):741–761. Bjerager, P. 1988. Probability integration by directional simulation. Journal of Engineering Mechanics 114(8):1288–1302. Brill Jr., E.D. 1979. The use of optimization models in public-sector planning. Management Science 25(5):413–422. Burke, J.V. 1991. Calmness and exact penalization. SIAM J. Control and Optimization 29(2):493–497. Clarke, F. 1983. Optimization and nonsmooth analysis. New York, New York: Wiley. Deak, I. 1980. Three digit accurate multiple normal probabilities. Numerische Mathematik 35:369–380. Ditlevsen, O. & Madsen, H.O. 1996. Structural reliability methods. New York, New York: Wiley.

S a m p l e a v e r a g e a p p r o x i m a t i o n s i n r e l i a b i l i t y-b a s e d s t r u c t u r a l o p t i m i z a t i o n

333

Ditlevsen, O., Oleson, R. & Mohr, G. 1987. Solution of a class of load combination problems by directional simulation. Structural Safety 4:95–109. Drezner, Z. & Erkut, E. 1995. Solving the continuous p-dispersion problem using nonlinear programming. The Journal of the Operational Research Society 46(4):516–520. Eldred, M.S., Giunta, A.A., Wojtkiewicz, S.F. & Trucano, T.G. 2002. Formulations for surrogate-based optimization under uncertainty. In Proceedings of the 9th AIAA/ISSMO Symposium on Multidisciplinary Analysis and Optimization, Paper AIAA-2002-5585, Atlanta, Georgia. Enevoldsen, I. & Sørensen, J.D. 1994. Reliability-based optimization in structural engineering. Structural Safety 15(3):169–196. Gasser, M. & Schuëller, G.I. 1998. Some basic principles in reliability-based optimization (RBO) of structures and mechanical components. In Stochastic programming methods and technical applications, K. Marti & P. Kall (eds), Lecture Notes in Economics and Mathematical Systems 458, Springer-Verlag, Berlin, Germany. He, L. & Polak, E. 1990. Effective diagonalization strategies for the solution of a class of optimal design problems. IEEE Transactions on Automatic Control 35(3):258–267. Holicky, M. & Markova, J. 2003. Reliability analysis of impacts due to road vehicles. In A. Der Kiureghian, S. Madanat & J.M. Pestana (eds), Applications of Statistics and Probability in Civil Engineering, Rotterdam, Netherlands, pp. 1645–1650. Millpress. Igusa, T. & Wan, Z. 2003. Response surface methods for optimization under uncertainty. In Proceedings of the 9th International Conference on Application of Statistics and Probability, A. Der Kiureghian, S. Madanat & J. Pestana (eds), San Francisco, California. Itoh, Y. & Liu, C. 1999. Multiobjective optimization of bridge deck maintenance. In Case Studies in Optimal Design and Maintenance Planning if Civil Infrastructure Systems, D.M. Frangopol (ed.), ASCE, Reston, Virginia. Kuschel, N. & Rackwitz, R. 2000. Optimal design under time-variant reliability constraints. Structural Safety 22(2):113–127. Liu, P.-L. & Kuo, C.-Y. 2003. Safety evaluation of the upper structure of bridge based on concrete nondestructive tests. In A. Der Kiureghian, S. Madanat & J.M. Pestana (eds), Applications of Statistics and Probability in Civil Engineering, Rotterdam, Netherlands, pp. 1683–1688. Millpress. Madsen, H.O. & Friis Hansen, P. 1992. A comparison of some algorithms for reliability-based structural optimization and sensitivity analysis. In Reliability and Optimization of Structural Systems, Proceedings IFIP WG 7.5, R. Rackwitz & P. Thoft-Christensen (eds), SpringerVerlag, Berlin, Germany. Marti, K. 1996. Differentiation formulas for probability functions: the transformation method. Mathematical Programming 75:201–220. Marti, K. 2005. Stochastic Optimization Methods. Berlin: Springer. Mathworks, Inc. 2004. Matlab reference manual, Version 7.0. Natick, Massachusetts: Mathworks, Inc. Nakamura, H., Miyamoto, A. & Kawamura, K. 2000. Optimization of bridge maintenance strategies using GA and IA techniques. In Reliability and Optimization of Structural Systems, Proceedings IFIP WG 7.5, A.S. Nowak & M.M. Szerszen (eds), Ann Arbor, Michigan. Polak, E. 1997. Optimization. Algorithms and consistent approximations. New York, New York: Springer-Verlag. Polak, E. & Royset, J.O. 2007. Efficient sample sizes in stochastic nonlinear programming. J. Computational and Applied Mathematics. To appear. Royset, J.O., Der Kiureghian, A. & Polak, E. 2006. Optimal design with probabilistic objective and constraints. J. Engineering Mechanics 132(1):107–118. Royset, J.O. & Polak, E. 2004a. Implementable algorithm for stochastic programs using sample average approximations. J. Optimization. Theory and Application 122(1):157–184.

334

Structural design optimization considering uncertainties

Royset, J.O. & Polak, E. 2004b. Reliability-based optimal design using sample average approximations. J. Probabilistic Engineering Mechanics 19(4):331–343. Royset, J.O. & Polak, E. 2007. Extensions of stochastic optimization results from problems with simple to problems with complex failure probability functions. J. Optimization. Theory and Application 133(1):1–18. Rubinstein, R. & Shapiro, A. 1993. Discrete Event Systems: Sensitivity Analysis and Stochastic Optimization by the Score Function Method. New York, NY: Wiley. Ruszczynski, A. & Shapiro, A. 2003. Stochastic Programming. New York, New York: Elsevier. Torczon, V. & Trosset, M.W. 1998. Using approximations to accelerate engineering design optimization. In Proceedings of the 7th AIAA/USAF/NASA/ISSMO Symp. on Multidisciplinary Analysis and Optimization, AIAA Paper 98-4800, St. Louis, Missouri. Tretiakov, G. 2002. Stochastic quasi-gradient algorithms for maximization of the probability function. A new formula for the gradient of the probability function. In Stochastic Optimization Techniques, New York, pp. 117–142. Springer. Uryasev, S. 1995. Derivatives of probability functions and some applications. Annals of Operations Research 56:287–311. White, D.J. 1996. A heuristic approach to a weighted maxmin dispersion problem. IMA Journal of Mathematics Applied in Business and Industry 7:219–231.

Chapter 13

Cost-benefit optimization for maintained structures Rüdiger Rackwitz & Andreas E. Joanni Technical University of Munich, Munich, Germany

ABSTRACT: In this chapter the theoretical and practical issues for setting up effective costbenefit optimization formulations for existing aging structures are presented. These formulations include deterioration and failure models as well as inspection and repair models. An elaborate optimization methodology, based on renewal theory that uses systematic reconstruction or repair schemes after suitable inspection is formulated, in which life-cycle cost perspectives are used is implemented for maintained concrete structures.

1 Introduction Many civil engineering structures are exposed not only to loads but also to the technical or natural environment. They are aging because of wear, corrosion, fatigue and other phenomena. At a certain age they need to be inspected and, possibly, repaired or replaced. Many aging phenomena are rather complex and all but fully understood in their physical and chemical context. For concrete structures the most important aging phenomena in temperate climates are corrosion due to carbonation and/or chloride attack, for steel structures it is rusting and fatigue. Moreover, the concepts for costbenefit optimization of such structures are not very well developed, although it is known that the cost for maintenance can be considerable and, in the long term, can even exceed the cost of the initial investment. It should be clear that only a rigorous lifecycle consideration can fully account for all cost involved, and that design rules and maintenance strongly interact. While the techniques for design optimization appear sufficiently developed, no clear concepts exist for optimizing maintenance. In this contribution suitable failure models for physically based deterioration phenomena are first reviewed. Their computation is essentially based on FORM/SORM (see, for example, (Rackwitz 2001)) which can be shown to be accurate enough for the purpose under discussion. Several schemes for computing first passage time distributions are discussed. Failure time models for series systems are also given. This is followed by some remarks about classical renewal theory, Bayesian updating, inspection and repair models. Then, the well-known renewal theory (Rosenblueth and Mendoza 1971; Rackwitz 2000) for cost-benefit optimization of structures is outlined. It is extended and generalized to optimal and integrated inspection and maintenance strategies. When setting up suitable maintenance strategies we follow closely the concepts developed in classical reliability theory as described, for example, in (Barlow and Proschan 1965; Barlow and Proschan 1975) which we find still very valid and which, to our knowledge,

336

Structural design optimization considering uncertainties

have not been applied to structures so far (see, however, (Van Noortwijk 2001)). In particular, we study minimal, age-dependent and block repairs and maintenance by inspection and repair. The models are generalized for maintenance optimization of series systems. Some special optimization techniques are briefly reviewed. An example illustrates aspects of the theory. Clearly, the considerations are no more valid if other than economic reasons exist to repair and/or retrofit an existing structure.

2 Preliminaries 2.1 F ai l ure m od e ls wit ho ut d e t er io r at i o n As a matter of fact, there are very few exact, time-variant failure models available which are amenable to practicable computation. In some cases consideration of (stationary or non-stationary) time-variant actions and time-variant structural state function is necessary. Let G(X(t), t) be the structural state function such that G(X(t), t) ≤ 0 denotes failure states and X(t) a random process. Examples of such processes are the Gaussian and related processes and the rectangular wave renewal processes. But X(t) can also include simple random variables. Then, the failure time distribution can be computed numerically by the outcrossing approach. A well-known upper bound is


t

F(t) ≤

ν(τ)dτ ≤ 1

(1)

0

with the outcrossing rate (more specifically, the downcrossing rate) 1 P({G(X(t), t) > 0} ∩ {G(X(t + ), t + ) ≤ 0})

→0

ν(τ) = lim

(2)

This upper bound is only tight for small probabilities. Frequently, an asymptotic result is used (Cramér and Leadbetter 1967) 
t F(t) ≈ 1 − exp − ν(τ)dτ (3) 0

with 
t f (t) ≈ ν(t) exp − ν(τ)dτ

(4)

0

Equation (3) implies a non-homogeneous Poisson process of failure events with intensity ν(t). For stationary failure processes Equation (3) reduces to a homogeneous Poisson process and simplifies somewhat. In general, computations are done by first transforming the original process and/or random variables into the so-called standard space of uncorrelated standard normal variates (Hohenbichler and Rackwitz 1981) which enables to use FORM/SORM (see, for example, (Rackwitz 2001)) provided that the dependence structure of the two events {G(X(t), t) > 0} and {G(X(t + ), t + ) ≤ 0} can be determined in terms of correlations coefficients. Some computational details are given in (Streicher and Rackwitz 2004). However, the relevant conditions must be fulfilled, i.e. the outcrossing events must become independent

C o s t-b e n e f i t o p t i m i z a t i o n f o r m a i n t a i n e d s t r u c t u r e s

337

and rare asymptotically. For example, the independence property is lost if X(t) contains not only (mixing) random processes but also simple random variables. Therefore, in many cases this approach yields only crude approximations. An alternative approach will be discussed in the next subsection. 2.2

Failure mo dels for deterioration

Obviously, the outcrossing approach can also be applied if there is deterioration. It appears as if it performs better if the outcrossing rate is increasing with time. For aging structures a closed-form failure time (first passage time) distribution is hardly available except for some special, usually oversimplifying cases. The log-normal, inverse Gaussian or Weibull distribution function with a suitable deterioration mechanism for the mean (or other parameters) has been used. They, at most, can serve as approximations. Realistic failure models must be derived from physical multi-variable deterioration models (cumulative wear, corrosion, fatigue, etc.). For (monotonically and continuously) deteriorating structures a widely used failure model is as follows. Let G(X, t) = g(U, t) be the (differentiable) structural state function of a structural component with G(X, t) = g(U, t) ≤ 0 the failure domain. X is a vector of random variables and time t is a parameter. Transition from X to U denotes the usual probability transformation from the original into the standard space of variables (Hohenbichler and Rackwitz 1981). Within FORM/SORM the probability of the time to first failure is F(t) = P(T ≤ t) = P(g(U, t) ≤ 0) ≈
(−β(t))C(t)

(5)

for t ≥ 0 and the failure density is ∂F(t) ∂β(t) ∂C(t) ≈ −ϕ(β(t)) C(t) +
(−β(t)) ∂t ∂t ∂t ) * − ∂t∂ g(u∗ ,t) ∂C(t) = −ϕ(β(t)) C(t) +
(−β(t)) ∇u g(u∗ , t) ∂t

f (t) =

(6)

T is the time to first entrance into a failure state.
( · ) and ϕ( · ) denote the univariate standard normal distribution function and corresponding density, respectively. β(t) is the (geometrical) reliability index. C(t) is a correction factor evaluated according to SORM and/or importance sampling which can be neglected in many cases. In Equation (6) it frequently can be assumed that C(t) does not vary with t. Clearly, this model does not take account of the randomness in the deterioration process caused by a (large) number of small disturbances which, however, is small to negligible for cumulative deterioration phenomena, at least for larger t. A numerical computation scheme for first-passage time distributions under less restrictive conditions than the outcrossing approach can also be given. It is based on the following lower bound formula F(t) = P(T ≤ t) ≥ P

) n ' i=1

* P(G(X(ti ), ti ) ≤ 0)

(7)

338

Structural design optimization considering uncertainties

with t = tn and ti < t denoting a narrow not necessarily regular time spacing of the interval [0, t]. As demonstrated by examples in (Au and Beck 2001), the lower bound F(t) = P(T ≤ t) = 1 − P(G(X(θ), θ) > 0 for all θ in [0, t]) ) n * ) n * ' ' ≥ P {g(U(θi ), θi ) ≤ 0} ≈ P {α(θi )T U(θi ) + β(θi ) ≤ 0} i=0

)

= 1−P

*

n :

i=0

{Zi ≤ β(θi )} = 1 −
n+1 (β; R)

(8)

i=0

to the first-passage time distribution turns out to be surprisingly accurate for all values of F(t), if the time-spacing τ = θi − θi−1 is chosen sufficiently close and where θi = iτ and t = θn . Here again, a probability distribution transformation from the original space into the standard space is performed and the boundaries of each failure domain are linearized. The last line represents a first order approximation (Hohenbichler and Rackwitz 1983) where
n (·; ·) is the n-dimensional standard normal integral with β = {β(θi )} the vector of reliability indices of the various components in the union and the dependence structure of the events is determined in terms of correlation coefficients R = {ρij = α(θi )T α(θj )}. Suitable computation schemes for the multinormal integral even for high dimensions and arbitrary probability levels have been proposed, for example in (Hohenbichler and Rackwitz 1983; Gollwitzer and Rackwitz 1988; Pandey 1998; Ambartzumian et al. 1998; Genz 1992). It would appear that slight improvements can be achieved if the probabilities for the individual events are determined by SORM (or any other suitable improvement) and an equivalent value of βe (θ) is computed from βe (θ) = −
−1 (
(−β(θ))CSORM ). This computation scheme is approximate but quite general if the correlation structure of the state functions in the different points in time can be established. In (Au and Back 2001) a Monte Carlo method is used to compute Equation (7) which can be recommended if high accuracy requirements are imposed – at the expense of in part considerable numerical effort. The special case of equi-dependent (equi-correlated) components is worth mentioning. In this case we simply have (see, for example (Dunnett and Sobel 1955))
Fe (t) = 1 −

∞ −∞

ϕ(τ)

t  i=1




√  βi − ρτ dτ √ 1−ρ

(9)

For equi-reliable components (no variation of resistance quantities with time) this result simplifies further. The corresponding values of the density function needed when taking Laplace transforms as required later are most easily calculated by f (θi ) = (F(θi ) − F(θi−1 ))/τ or a higher order differentiation rule. For equi-reliable components Equation (9) has a decreasing risk function.

C o s t-b e n e f i t o p t i m i z a t i o n f o r m a i n t a i n e d s t r u c t u r e s

339

The results obtained so far carry over to systems without any further conceptual difficulty. Only the numerical computations become more involved. Any system can be reduced to a minimal cut set system so that its failure probability is represented as ⎛ ⎞ mi s : ' Pf (t) = P(T ≤ t) = F(t) = P ⎝ {Tij ≤ t}⎠

(10)

i=1 j=1

Assume that the failure times of the parallel systems can be determined which, in general, can involve quite some numerical effort. The remaining series system then is computed as ) s * ) s * s ' :  Pf (t) = P(T ≤ t) = F(t) = P {Ti ≤ t} = 1 − P {Ti > t} ≤ P(Ti ≤ t) i=1 i=1 i=1 (11) where usually the failure and survival events are dependent. The upper bound in Equation (11) is less useful for larger, low reliability systems. Equation (8) can be combined with Equation (11), especially if the parallel systems can be represented sufficiently well by equivalent, linearly bounded failure domains of the components (Gollwitzer and Rackwitz 1983). Some specific results for the computation of series systems are given in (Streicher and Rackwitz 2004). The failure densities are obtained by differentiation. Note that, by definition, a series system fails if any of its components fails. In passing it is also noted that the formulation in Equation (11) also includes failure due to extreme disturbances. And it should be clear that the series system model must be applied if several hazards are present. Deterioration of structural resistance is frequently preceded by an initiation phase. In this phase failure is dominated by normal (extreme-value) failure. Structural resistance is virtually unaffected. Only in the succeeding phase resistances degrade. Examples are crack initiation and crack propagation or chloride penetration into concrete up to the reinforcement and subsequent reduction of the reinforcement cross-section by corrosion and, similarly, for initial carbonation and subsequent corrosion. In many cases the initiation phase is much longer than the actual degradation phase. Let Ti denote the random time of initiation, Te the random time to normal (first-passage extreme-value) failure and Td the random time from the end of the initiation phase to deterioration failure with degraded resistance. Then, F(t) = P(T ≤ t) = P[({Ti > t} ∩ {Te ≤ t}) ∪ ({Ti ≤ t} ∩ {Te < Ti }) ∪({Ti ≤ t} ∩ {Te > Ti } ∩ {Ti + Td ≤ t})]

(12)

= P[{Ti > t} ∩ {Te ≤ t}] + P[{Ti ≤ t} ∩ {Te < Ti }] + P[{Te > Ti } ∩ {Ti + Td ≤ t}] Note, extreme-value failure during the initiation phase and failure in the deterioration phase are mutually exclusive. Assume that Ti is independent of the other two variables.

340

Structural design optimization considering uncertainties

If the variables Te and Td can also be assumed independent, the following formula can be used
t F(t) = Fe (t)F i (t) + fi (τ)[Fe (τ) + (1 − Fe (τ))Fd (t − τ)]dτ (13) 0

2.3 T h e renew al mo d el A sufficiently general setting is to assume that the structure fails at a random time in the future. After failure or serious deterioration it is systematically renewed by reconstruction or retrofit/repair. Reconstruction, repair or retrofit reestablish all (stochastic) structural properties. The times between failure (renewal) events have identical distribution functions F(t), t ≥ 0 with probability densities f (t) and are independent. The sequence of failures and renewals then forms an ordinary renewal process. Renewal theory allows for a useful refinement which will be found to be important for the problem under discussion, namely the distribution of the time to the first event can have distribution function F1 (t) = F(t), t ≥ 0 (see (Cox 1962) for details). The process of renewals is then denoted by modified or delayed renewal process. The independence assumption between failure times needs to be verified carefully. In particular, one has to assume that loads and resistances in the system are independent for consecutive renewal periods and there is no change in the design rules after the first and all subsequent failures (renewals). Even if designs change failure time distributions must remain the same. But the model allows for a different design rule for the initial design which can be one of the reasons for F1 (t) = F(t). Throughout the chapter the point process of renewals is an orderly point process, that is multiple occurrences of renewals in a small time interval are excluded (Cox and Isham 1980). The renewal function for a modified renewal process which will be used extensively later on is (Cox 1962) E[N(t)] = M1 (t) =

∞ 

np(N(t) = n) =

n=1

= F1 (t) +

∞
 n=1

∞ 

n(Fn (t) − Fn+1 (t)) =

n=1 t

Fn (t)

n=1




t

Fn (t − u)dF(u) = F1 (t) +

0

∞ 

M1 (t − u)dF(u)

(14)

0

with N(t) the random number of renewals and Fn (t) = P(N(t) ≥ n) = P(Tn ≤ t) the distribution function of the time to the n-th renewal. The renewal intensity (or, if applied to failure processes, the unconditional failure rate) is obtained upon differentiation ∞

dM1 (t)  P(one renewal in [t, t + dt]) = = fn (t) m1 (t) = lim dt dt dt→0

(15)

n=1

For ordinary processes the index ‘1’ is omitted. The last expression in Equation (14) is called ‘renewal equation’. As pointed out in (Cox 1962), m(t) (or m1 (t)) has a limit m(t → ∞) = lim m(t) = t→∞

1 E[Tf ]

(16)

C o s t-b e n e f i t o p t i m i z a t i o n f o r m a i n t a i n e d s t r u c t u r e s

341

for f (t) → 0 if t → ∞. In approaching the limit m(t) can be strictly increasing, strictly decreasing or oscillate in a damped manner around 1/E[Tf ]. Ordinary renewal processes then tend to be large around t = E[Tf ], 2E[Tf ], . . . and small around t = 0, 32 E[Tf ], 53 E[Tf ], . . .. For a Poisson process with parameter λ it is constant, i.e. m(t) = λ. If there are oscillations they die out more rapidly for larger dispersions of the failure time distribution. In many examples oscillations have been found when the risk function is increasing. Also, in many cases the failure rate is increasing for small t. Only for some special models, especially those with very large coefficient of variation of failures times, m(t) is decreasing. The transient behavior of m(t) will later be of interest. Unfortunately, Equation (14) has closed-form solutions for only very few special mathematical failure models (see (Streicher et al. 2006) for a list of relevant references) and otherwise can be computed directly only with extreme numerical effort. In general, Equation (14) or Equation (15) have to be determined numerically. A particularly suitable numerical method is proposed in (Ayhan et al. 1999). It makes use of the upper and lower sum in Riemann-Stieltjes integration for the discrete version of Equation (14). ( t Because M(t) is non-decreasing, we have the following bounds for M(t) = F(t) + 0 M(t − s)dF(s)

MLB (kτ) = F(kτ) +

k 

MLB ((k − i)τ) F(iτ)

i=1

≤ M(kτ) ≤ F(kτ) +

k 

MUB ((k − i + 1)τ) F(iτ) = MUB (kτ)

(17)

i=1

for equal partitions of length τ in [0, t] with F(iτ) = F(iτ) − F((i − 1)τ) and nτ = t. The resulting system of linear equations is solved easily. If the first failure time distribution is different from the others one obtains by one additional convolution


t

M1 (t) = F1 (t) +

F1 (t − s)dM(s)

(18)

0

which, in turn, is bounded by

M1,LB = F1 (t) +

k  i=1

inf

(i−1)τ≤x≤iτ

≤ M1,UB ≤ F1 (t) +

F1 (t − x)(MLB (iτ) − MLB ((i − 1)τ)) ≤ M1 (kτ)

k 

sup

i=1 (i−1)τ≤x≤iτ

F1 (t − x)(MUB (iτ) − MUB ((i − 1)τ)) (19)

m1 (t) is obtained by numerical differentiation. The computation methods in Equations (17) and (19) are useful whenever interest lies in the (unconditional) failure rate or risk acceptance questions. Other approximation methods have also been proposed.

342

Structural design optimization considering uncertainties

For aging components with increasing risk function the following bounds on the renewal function are given in (Barlow and Proschan 1965, p. 54) t tF(t) t t − 1 ≤ M(t) ≤ ( t ≤ −1 ≤ (t E[Tf ] E[T f] 0 (1 − F(τ))dτ 0 (1 − F(τ))dτ

(20)

The sharper upper bound in Equation (20) turns out to be remarkably close to the exact result for small t. Under suitable conditions one also has m(t) =

d d tF(t) M(t) ≤ (t dt dt 0 (1 − F(τ))dτ

(21)

Again, the upper bound for Equation (21) is found to be very close to the exact result up to approximately E[T]. It approaches the limit 1/E[T] for large t. The lower bound obtainable from Equation (20) by differentiation is generally less useful. Equation (21) can be used with advantage in Sections 4.4 and 4.5. 2.4

U pd a ti ng t he pr o b ab ilis t ic mo d el

There are many types of updating of a probabilistic model depending on the type of information collected during the experimental and numerical investigations. In general, one can distinguish between variable updating and event updating. In a Bayesian context information is collected about a variable by taking (independent) samples and testing them. This leads to an improved estimate of the parameters of the distribution of a variable. Let xn be values of a sample of size n and θ a parameter (vector), then an improved posterior distribution is






f (θ | xn ) = (

L(xn | θ)f (θ)
L(x n | θ)f (θ)dθ θ

(22)


where L(xn | θ) is the likelihood function and f (θ) the prior density. The Bayesian or predictive density function is


f (x | xn ) = f (x | θ)f (θ | xn )dθ (23) θ

For many important distributions analytical results are available (Aitchison and Dunsmore 1975). Updating by events is generally more difficult. We  show this for the model from Equation (5) and previous informative events B = i=1 Bi . For example, such events could be the knowledge about the maximum load in the past, some measured damage indicator or just the knowledge that the structure has survived up to the present time. Then, we have two types of observations, namely equalities and inequalities which  require different treatment. For B = i=1 bi (X, t0 ) ≤ 0 it is F(t | B) =

 P({g(X, t) ≤ 0} ∩ i=1 {bi (X, t0 ) ≤ 0})  P( i=1 bi (X, t0 ) ≤ 0)

(24)

C o s t-b e n e f i t o p t i m i z a t i o n f o r m a i n t a i n e d s t r u c t u r e s

343

It is assumed that the observation events B can always be written in the form given. In most cases the observation and decision point is t0 = 0. Within FORM one can write for one observation event F(t | B) =


2 (−βg (t), −βb (t0 ), ρ)
(−βb (t0 ))

(25)

where
2 (x, y, ρ) is the two-dimensional normal integral and ρ = αTg αb with αg , αb the two normalized gradients of the limit state functions. This scheme applies analogously if more than one event has to be considered. For B = {b(X, t0 ) = 0} we have F(t | B) =

∂ ∂βb

( βb

−∞

P(Zg ≤ βg | Zb (t0 ) = z)ϕ(z)dz ϕ(βb (t0 ))

)

−βg (t) + ρ(t, t0 )βb (t0 ) =
 1 − ρ(t, t0 )2

*

(26)

3 Cost-benefit optimization 3.1

G eneral

It is generally accepted that the ultimate target to be achieved in structural design including proper maintenance is to maximize the net benefit derived from the structure over its lifetime, subject to constraints related to safety and serviceability. For technical facilities the following objective has been proposed by (Rosenblueth and Mendoza 1971) based on earlier proposals in economics for cost benefit analysis: Z(p) = B(p) − C(p) − D(p)

(27)

A facility is financially optimal if Equation (27) is maximized. It is assumed that all quantities in Equation (27) can be measured in monetary units. p is the vector of all safety relevant parameters. B(p) is the (expected) benefit derived from the existence of the facility, C(p) is the cost of design and construction and D(p) is the (expected) cost in case of failure. Later we will also include all expenses for maintenance in D(p). Statistical decision theory dictates that expected values are to be taken. In the following it is assumed that C(p) and D(p) are differentiable in each component of p. The facility has to be optimized during design and construction at the decision point which is taken as t = 0. Now it is a well-established principle of cost-benefit analysis that future costs and benefits must be discounted, using a compound interest formula. A continuous discounting function is assumed for analytical convenience which is accurate enough for all practical purposes. δ(t) = exp [−γt]

(28)

γ is a time-independent, time-averaged interest rate. In most cost-benefit analyses a tax and inflation-free discount rate should be taken. If a discrete discount rate γ
is given, one converts with γ = ln (1 + γ
). The principles of choosing appropriate discount rates are thoroughly discussed in (Rackwitz et al. 2005).

344

Structural design optimization considering uncertainties

Cost and benefits may differ for the different parties involved having different economic objectives, e.g. the owner, the builder, the user and society. Also, the discount rate may vary among the different parties in their cost-benefit analysis. A facility makes sense only if Z(p) is positive within certain parameter ranges for all parties involved. 3.2 D eri v ati o ns A complete cost-benefit analysis must include not only the direct and indirect cost for possible failure and for maintenance of the structure to be built, but also the cost for all future realizations if the concepts of sustainability are applied (Rackwitz et al. 2005). But this is just the situation for the application of renewal theory. It is assumed that structures will be systematically reconstructed after failure and/or maintained. This rebuilding strategy is in agreement with the principles of life cycle engineering and also fulfills the demand for sustainability (Rackwitz et al. 2005). Clearly, it rests on the assumption that future preferences are the same as the present preferences. For regular renewal processes some objective functions based on the renewal model are already derived in (Rosenblueth and Mendoza 1971; Rackwitz 2000; Streicher and Rackwitz 2004) and elsewhere. For existing structures the time to first failure is generally different from the other failure times due to additional experimental and numerical investigations and subsequent updating of the structural state and/or due to repair or retrofit of the existing structure. But there can also be other reasons for assuming f1 (t, p) = f (t, p). Therefore, we derive our model for cost-benefit optimization in full generality. The objective function is given by Equation (27). The expected damage cost D(p) are derived as follows. The discrete cost associated with failure including the reconstruction or repair cost are denoted as CV,1 at the first renewal and CV = CV,n at subsequent renewals. Let θi = ti − ti−1 be the times between renewals with density f (t, p) whereas θ1 = t1 has density f1 (t, p). The time to the n-th renewal is Tn = ni=1 θi . Systematic reconstruction is assumed. The discounted expected damage cost are then ∞   n   D(p) = E CV,n exp −γ θk n=1

k=1



= E[CV,1 exp[−γ θ1 ]] + E  = E[CV,1 exp[−γ θ1 ]] + E

∞  n=2 ∞ 

CV,n

n 

∞ 

exp[−γ θk ]

k=1

CV,n exp[−γ θ1 ]

n=2

= E[CV,1 exp[−γ θ1 ]] +



n−1 

 exp[−γ θk ] exp[−γ θn ]

k=2

E[ exp[−γ θ1 ]]E[ exp(−γ θ)]n−2 E[CV,n exp[−γ θn ]]

n=2

= E[CV,1 exp[−γ θ1 ]] + E[ exp[−γ θ1 ]] =

E[CV exp[−γ θ]] 1 − E[ exp[−γ θ]]

CV,1 E[exp[−γ θ1 ]] 1 − E[exp[−γ θ]] −CV,1 E[exp[−γ θ1 ]]E[exp[−γ θ]] + E[exp[−γ θ1 ]]CV E[exp[−γ θ]] + (29) 1 − E[exp[−γ θ]]

C o s t-b e n e f i t o p t i m i z a t i o n f o r m a i n t a i n e d s t r u c t u r e s

345

 a n−k where we have made use of the relation s = ∞ = 1−q for k < ∞. n=k aq (∞ (∞ ∗ E[exp[−γ θ1 ]] = 0 exp[−γt]f1 (t, p)dt = f1 (t, p) and E[exp[−γ θ]] = 0 exp[−γt] f (t, p)dt = f ∗ (t, p) is also denoted as Laplace transform of f1 (t, p) and f (t, p). If f (t, p) is a probability density it is f ∗ (0, p) = 1 and 0 < f ∗ (γ, p) ≤ 1 for all γ ≥ 0. Equation (27) can be rewritten in case of systematic reconstruction after failure with CV,1 = (C1 (p) + L) as well as CV = (C(p) + L) as Z(p) = B(p) − C(p) − +

(C1 (p) + L)f1∗ (γ, p) 1 − f ∗ (γ, p)

(C1 (p) + L)f ∗ (γ, p)f1∗ (γ, p) − (C(p) + L)f1∗ (γ, p)f ∗ (γ, p) 1 − f ∗ (γ, p)

(30)

for the modified renewal process. L is the monetary loss in case of failure including direct failure cost, loss of business and, possibly, the cost to reduce the risk to human life and health (or, better, the compensation cost). If only C1 (p) = C(p) the two terms in the numerator of the forth term cancel. This is usually the case for existing and systematically renewed structures and, therefore Z(p) = B(p) − Cini (p) − (C(p) + L)

f1∗ (γ, p) 1 − f ∗ (γ, p)

(31)

It has to be mentioned that the design parameters p can be different after the first renewal compared to the initial design. Also, the cost for the initial design Cini (p) can be different from the reconstruction cost C(p). The term m∗1 (γ, p) =

f1∗ (γ, p) 1 − f ∗ (γ, p)

(32)

is also denoted by the Laplace transform of the renewal intensity. If f1 (t, p) = f (t, p), f1∗ (t, p) in Equation (31) must be replaced by f ∗ (t, p). The benefit B(p) is also discounted down to the decision point. For a benefit rate b(t) unaffected by possible renewals and negligibly short times of reconstruction (retrofitting) one finds
∞ B= b(t) exp[−γt]dt (33) 0

Clearly, the integral must converge imposing some restriction on the form of b(t). If the benefit rate b = b(t) is constant one can integrate to obtain


∞

B= 0

b exp[−γt]dt =

b γ

(34)

The upper integration limit is extended to infinity because the full sequence of life cycle benefits is considered. A model which represents realistically the observation that with increasing age of a component its suitability for use diminishes according to b(t) has been established in (Hasofer and Rackwitz 2000). Decreasing benefit was associated with obsolescence in

346

Structural design optimization considering uncertainties

this reference. But b(t) can have any form. At each renewal the benefit rate starts again at b(0) for systematic reconstruction. The total benefit is already given in (Streicher 2004) and is repeated here in full generality.  B(p) = E

∞ 

)

 exp −γ

i=1







θi

θk

* exp[−γ τ]b(τ)dτ

0

k=1

θ1

=E

i−1 

exp[−γ τ]b(τ)dτ

0

 + E exp[−γ θ1 ]

∞  i−1 


exp[−γ θk ]




θ1

 exp[−γ τ]b(τ)dτ

0

i=2 k=2

=E

θi

exp[−γ τ]b(τ)dτ + E[exp[−γ θ1 ]]

0

∞ 

E[exp[−γ θ]]i−2

i=2




θ

×E

exp[−γ τ]b(τ)dτ

0




θ1

= 0

(θ E[exp[−γ θ1 ]]E[ 0 exp[−γ τ]b(τ)dτ] exp[−γ τ]b(τ)dτ + 1 − E[ exp[−γ θ]]

(35)

Equation (35) can be simplified for the case of systematic reconstruction after failure to
∞
t B(p) = exp[−γτ]b(τ)dτ f1 (t, p)dt 0

0

f1∗ (γ, p) 1 − f ∗ (γ, p)

+


∞

=


∞


t

exp[−γτ]b(τ)dτ f (t, p)dt 0

0

BD (t)f1 (t, p)dt +

0

f1∗ (γ, p) 1 − f ∗ (γ, p)




∞

BD (t)f (t, p)dt

(36)

0

with


t

BD (t) =

exp[−γτ]b(τ)dτ

(37)

0

For f1 (t, p) = f (t, p) Equation (36) simplifies to: 1 B(p) = 1 − f ∗ (γ, p)




∞

BD (t)f (t, p)dt

(38)

0

For completeness, the objective function is also given for the case where the component is given up after failure or a finite service time ts
Z(p) = 0

ts




ts

BD (t)f1 (t, p)dt − C(p) − L

exp[−γt]f1 (t, p)dt 0

(39)

C o s t-b e n e f i t o p t i m i z a t i o n f o r m a i n t a i n e d s t r u c t u r e s

347

Because the failure densities, in general, are known only numerically and pointwise, the corresponding Laplace transforms have to be taken numerically. Suitable techniques are presented in (Streicher and Rackwitz 2004) and Section 5. The formulae are easily extended for systems with several components and/or multiple failure modes in series as demonstrated by Equation (11) (see also (Streicher and Rackwitz 2004)). In particular, one component of the system can model replacement due to obsolescence. Non-constant discounting is discussed in (Rackwitz et al. 2005). Optimization of Equation (31) with respect to the design parameter p can be performed by one of the available algorithms (see Section 5). Application to existing, aging but maintained structures requires a few more remarks. It is assumed that the structure is already in use for some time. At a special point in time it will be decided to inspect and possibly repair or retrofit the structure. The cost which occur at this decision point are CR (p). Clearly, all cost incurred before that point are irrelevant if the decision is to keep the structure rather than demolishing and rebuilding it. The value of CR (p) can be zero if the structure is left as is but the probabilistic model for the time to first failure f1 (t, p) possibly is updated. Then, renewal of the structure is a question as to when the possibly updated failure rate is no more acceptable. The modified density f1 (t, p) of the time to first failure has to be determined depending on the repair/retrofitting actions and the information collected about the actual state of the structure. CR (p) generally differs from C(p), the reconstruction cost after failure, or even exceeds it if retrofitting is more expensive than reconstruction. A maintenance plan for the existing structure has to be designed. After the first renewal due to future failure the regular failure time density f (t, p) is valid.

3.3 Applicatio n to s tationary Pois s onia n di s turbanc e s Unfortunately, analytic Laplace transforms are available only for a few analytic failure models, for instance the exponential, uniform, gamma, normal and inverse normal distribution. The important exponential distribution with parameter λ corresponding to a Poisson process has f1∗ (γ) = f ∗ (γ) = λ/γ + λ and, therefore, m∗ (γ) = λ/γ. A very useful generalization is when a modified renewal process models disturbance (loading) events (Hasofer 1974; Rosenblueth 1976). Such disturbances generally are extreme events like shocks, explosions, earthquakes, storms or floods. The distribution functions between events are G1 (t) and G(t), respectively. If such an event occurs the failure probability is Pf (p). By definition, the occurrence of disturbance events and the failure events are independent. The density function of the time to the first failure event then is f1 (t, p) =

∞ 

gn (t)Pf (p)Rf (p)n−1

(40)

n=1

i.e. the first failure event can occur after the first, second, third, . . . disturbance event and where Rf (p) = 1 − Pf (p). The density of the n-th event can be obtained by recursive convolution so that in terms of Laplace transforms ∗ gn∗ (γ, p) = gn−1 (γ, p)g ∗ (γ, p) = g1∗ (γ, p)[g ∗ (γ, p)]n−1

(41)

348

Structural design optimization considering uncertainties

Application to the renewal intensity yields m∗1 (γ, p) =

∞ 

∗ g1∗ (γ)gn−1 (γ)Pf (p)Rf (p)n−1

n=1

=

∞ 

g1∗ (γ)[g ∗ (γ)]n−1 Pf (p)Rf (p)n−1 =

n=1

Pf (p)g1∗ (γ) 1 − Rf (p)g ∗ (γ)

(42)

For the regular renewal process m∗1 (γ, p) has to be replaced by m∗ (γ, p). Let reconstruction and damage cost be C(p) and L, respectively. Also, as a special case, let the times between disturbances be the (exponential failure time distributions with (failure) rate ∞ λPf (p). Therefore, E(e−γt ) = −∞ e−γt (λPf (p))e−λPf (p)t dt = λPf (p)/γ + λPf (p). Then, if only failures due to such disturbances are considered it is (Rackwitz 2000) Z(p) = B − Cini (p) − (C(p) + L)

λPf (p) γ

(43)

  For a series system it is Pf (p) = P( sk=1 P(Fk (p))) = 1 − P( sj=1 F j (p)) in Equation (11) where Fk (p) is the failure event in the k-th mode and F k (p) its complement. Then, the following generalization is possible ⎡ ⎛ ⎞⎤ s : λ Z(p) = B − Cini (p) − (C(p) + L) ⎣1 − P ⎝ F j (p)⎠⎦ (44) γ j=1

The benefit B and the initial cost Cini (p) as well as the damage cost are related to the whole system. If there are n different, independent hazards each with rate λi one derives ⎡ ⎛ ⎞⎤ n s : λi Z(p) = B − Cini (p) − i=1 (Ci (p) + Li ) ⎣1 − P ⎝ F ij (p)⎠⎦ (45) γ j=1

These generalizations also apply analogously for the more complicated cases discussed below.

4 Preventive maintenance 4.1 Ma i n tenanc e s t r at eg ie s Repair after failure is but the simplest maintenance strategy. For aging components, i.e. components with increasing risk function (conditional failure rate) r(t) = f (t)/1 − F(t), i.e. r
(t) > 0, the risk of failure with potentially large consequences increases with age and alternative maintenance strategies have been proposed in order to reduce expected failure consequences. The most important alternative is called preventive maintenance at random or fixed times. Preventive maintenance actions can be replacements or

C o s t-b e n e f i t o p t i m i z a t i o n f o r m a i n t a i n e d s t r u c t u r e s

349

(perfect) repairs. Preventive repairs occur only if corrective renewals have not occurred before due to failure or obsolescence. Note that preventive maintenance is usually suboptimal for non-aging components, i.e. with constant or decreasing risk function. A first strategy repairs a system (component) at age a or after failure, whichever comes first. In (Barlow and Proschan 1965) this strategy is denoted by age replacement. It requires knowledge of the age a of a component. (Barlow and Proschan 1965) also investigate so-called block repairs. In this maintenance strategy the components in a system are repaired either after failure or all at once at a given time d irrespective of their actual age. It is clear for increasing risk functions and, in fact, is shown in (Barlow and Proschan 1965) that the total number of repairs is smaller for age repairs than for block repairs. However, the number of failures (with large consequences) is larger in the first strategy and so, possibly, the total cost. Block repairs also may be organizationally easier. Sometimes they are necessary, i.e. whenever a single repair of a component prevents the whole system from functioning. While knowledge about the actual deterioration state of a component is irrelevant for the block repair strategy, this may be vital for the age repair strategy. An improvement is when repairs are only performed if inspections indicate that they are necessary. Otherwise further inspections and possible repairs are postponed to a later time. A strategy where repairs are preceded by inspections is also denoted as condition-based strategy. In practice, mixtures of these maintenance strategies will also be found. 4.2

Inspections

Inspections should determine the actual state of a component in order to decide on repair or leave it as is. But inspections can rarely be perfect. A decision about repair can only be reached with certain probability depending on the inspection method used. The repair probability depends on the magnitude of one or more suitable damage indicators (chloride penetration depth, crack length, abrasion depth, etc.) measured during inspection. For cumulative damage phenomena the damage indicators increase with time and so does the repair probability PR (t). The parameter t is the time elapsed since the beginning of the deterioration process. For example, the repair probability may be presented as PR (t) = P(S(t, X) > sc ) = P(sc − S(t, X) ≤ 0)

(46)

with S(t, X) a suitable, monotonically increasing damage indicator, X a random vector taking into account of all uncertainties during inspection and sc a given threshold level. If this is exceeded a decision for repair is taken. The vector X usually also includes a random variable modeling the measurement error. Frequently, the damage indicator function S(t, X) reflects the damage progression and has a similar form as the failure function. It involves, at least in part, the same random variables. In this case failure and no repair/repair events become dependent events. It is, of course, possible to consider multidimensional damage indicators and derive repair decisions from an arbitrary combination thereof. A discussion of the details of the efficiency of various inspection methods and the corresponding repair probabilities is beyond the scope of this chapter. They depend on the particular deterioration phenomenon under consideration.

350

Structural design optimization considering uncertainties

4.3 Repai r m od el After failure of a system or component it is repaired unless it is given up after failure or it is repaired systematically in the age-dependent maintenance strategy or it is repaired after an indicative inspection in the condition-based maintenance strategy. The name repair is used synonymously for renewal, replacement or reconstruction. Repairs, if undertaken, restore the properties of a component to its original (stochastic) state, i.e. repairs are equivalent to renewals (AGAN = As Good As New) so that the life time distribution of the repaired component is again F(t). The repair times can either be assumed negligibly short or have finite length. The model is a somewhat idealized model. It rests on a number of assumptions the most important of which is probably that repairs fully restore the (stochastic) properties of the component. Imperfect repairs cannot be handled because the renewal argument repeatedly used in the following breaks down. In the literature several models for imperfect repairs are discussed which only partially reflect the situations met in the structures area. An important case is when minimal repairs not essentially changing the initial lifetime are done right after an inspection. If one generalizes this model to a model where a renewal (perfect repair) occurs with probability π but minimal repair with probability 1−π, one has essentially the model proposed in (Brown and Proschan 1983). This model, in fact, resembles the one studied herein with π = PR (t). Negligibly short times of inspection and repair are most often only a more or less good approximation. Consideration of finite, random renewal times in the age-repair strategy appears possible but is complicated because inspections and probably also failures cannot occur during repairs. No benefit can be earned during repair times. Another important case is when repairs are delayed, for example due to budget restrictions. It appears possible to handle this case by adding a random delay time to the random repair time. During a delay time the component can still degrade or fail while this is unlikely to happen during repair. Finite renewal times are not considered in this chapter. Some more but still first results are given in (Joanni and Rackwitz 2006). It turns out that for realistic repair times their influence is very small. Inspection/repair at strictly regular time intervals as assumed below is also not very realistic. However, as will be shown in the examples, the objective function is rather flat in the vicinity of the optimal value so that small variations will not noticeably change the results. Repair operations necessarily lead to discontinuities (drops) in the risk function, and similarly in the renewal intensity. They can substantially reduce the number of failures and, thus, corrective renewals. In an effective maintenance scheme the majority of renewals will, in fact, be preventive renewals. 4.4 Ag e-d epend ent r e pair s It is convenient to start with the general case of replacements (repairs, renewals) at random times Tr with distribution Fr (t) or after failure at random times Tf with distribution Ff (t, p). The renewal time then is the minimum of these times with distribution function F(t, p) = 1 − (1 − Ff (t, p))(1 − Fr (t)) = 1 − F f (t, p)F r (t)

(47)

C o s t-b e n e f i t o p t i m i z a t i o n f o r m a i n t a i n e d s t r u c t u r e s

351

for independent times Tf and Tr with density f (t, p) = ff (t, p)F r (t) + fr (t)F f (t, p)

(48)

and where the notation F(x) = 1 − F(x) is used. Application of Equation (29) then gives for the damage term of an ordinary renewal process D(p) =

(C(p) + L)fF∗ (γ, p) + R(p)fF∗ (γ, p) r

1−

(fF∗ (γ, p) f

f

(49)

+ fF∗ (γ, p)) r

and, similarly, for the benefit term with the model in Equation (35) (∞ B(p) =

0

(∞ BD (t)ff (t, p)F r (t)dt + 0 BD (t)fr (t)F f (t, p)dt 1 − (fF∗ (γ, p) + fF∗ (γ, p)) f

where fF∗ (γ, p) = r

(∞ 0

(50)

r

exp[−γt]ff (t, p)F r (t)dt and fF∗ (γ, p) =

(∞

f

0

exp[−γt]fr (t)F f (t, p)dt

are the modified complete Laplace transforms of ff (t, p)F r (t) and fr (t)F f (t, p), respectively. R(p) is the cost of repair and BD (t) is as in Equation (37). The case of random maintenance actions has hardly any practical application except if there is continuous monitoring of the system state. Then, the time until intervention by repair is random and can be defined as the first passage time of a given threshold by the continuous observation process. Alternatively, assume maintenance actions at (almost) fixed intervals a, 2a, 3a, . . . so that fr (t) = δe (a) and Fr (t) = He (a) (δe (x) = Dirac’s delta function, He (a) = Heavyside’s unit step function. Equation (49) then specializes to DM (p,a) =

(C(p) + L)f ∗∗ (γ, p,a) + R(p) exp[−γa]F(p,a) 1 − (f ∗∗ (γ, p,a) + exp[−γa]F(p,a))

(51)

and similarly Equation (50) to (a BM (p,a) =

0

BD (t)f (t, p)dt + BD (a)F(p,a)

(52) 1 − (f ∗∗ (γ, p,a) + exp[−γa]F(p,a)) (a with f ∗∗ (γ, p,a) = 0 exp[−γt]f (t, p)dt the incomplete Laplace transform of f (t, p) and F(p,a) the probability of survival up to a. The quantity BD (t) is given in Equation (37). Note that the Laplace transform of a deterministic repair time fr (t) = δe (a) is f ∗ (γ) = exp[−γa]. The repair cost R(p) should be substantially smaller than C(p) + L so that it is worth making preventive repairs and, thus, avoiding the large failure and reconstruction cost in case of failure. Equation (51) goes back to some early work in (Cox 1962; Barlow and Proschan 1965; Fox 1966). In (Van Noortwijk 2001) parallel results are developed for discrete failure models and discrete discounting schemes. Next, assume that the structure is already in use. At a special decision point in time inspection, retrofit or repair at cost CR (p) takes place. Depending on the state of the structure and the action which is done, an updated failure time density f1 (t, p) for

352

Structural design optimization considering uncertainties

the time to the first renewal is calculated. Therefore, a new cost benefit optimization is necessary in order to find optimal replacement intervals and design variables. The first replacement interval a1 with f1 (t, p) is different from the subsequent intervals a with ordinary failure time density f (t, p). It will further be assumed that for the first renewal the optimized parameter p, which is also valid for all subsequent renewals, is calculated without having regard to the special parameters realized in the existing structure. If the structure undergoes a complete renewal at the decision point it is even possible to introduce the design variables p already in that structure. Then, the existing design variables p have to be augmented by the additional variables. The expected damage cost are then determined according to Equation (49) as DMa1−a (p, a1 , a = (C(p) + L)f1∗∗ (γ, p, a1 ) + R(p) exp[−γa1 ]F 1 (p, a1 ) +

f1∗∗ (γ, p, a1 ) + exp[−γa1 ]F 1 (p, a1 ) 1 − (f ∗∗ (γ, p, a) + exp [−γa]F(p, a))

× ((C(p) + L)f ∗∗ (γ, p, a)

+ R(p) exp[−γa]F(p, a))

(53)

For constant benefit rate b(t) = b the benefit is as in Equation (34). The expected benefit for a non-constant rate b(t) as in Equation (35) is (Streicher 2004)
a1 BMa1−a (p, a1 , a) = BD (t)f1 (t, p)dt + BD (a1 )F 1 (p, a1 ) 0

+

f1∗∗ (γ, p, a1 ) + exp[−γa1 ]F 1 (p, a1 )

1 − (f ∗∗ (γ, p, a) + exp[−γa]F(p, a))


a  × BD (t)f (t, p)dt + BD (a)F(p, a)

(54)

0

with BD (t) from Equation (37). The cost for continuous monitoring and/or maintenance could alternatively also be taken into account in the benefit term by replacing b(t) with b(t) − c(t). The objective function then is ZMa1−a (p, a1 , a) = BMa1−a (p, a1 , a) − CR (p) − DMa1−a (p, a1 , a)

(55)

Repair is interpreted as preventive renewal (replacement of an aging component after a finite time of use a). Renewal after failure is called corrective renewal. Equation (55) can be subject to optimization not only with respect to the design parameter p but also with respect to the inspection/repair intervals a1 and a, respectively. Optimal inspection/ repair intervals do not always exist, as pointed out already in (Fox 1966). They exist for failure models with increasing risk function (Fox 1966). If they do not exist, then it is preferable to wait with renewal until failure unless the failure rate exceeds a given value. When optimizing Equation (55) it is, of course, important that the owner, builder or other party does not only enjoy the benefits but also carries the cost of construction, the cost of failures and the cost for preventive maintenance. Only then, a joint optimization of design and maintenance makes sense. If one is only interested in optimal maintenance it is still possible to optimize the cost for preventive and corrective repairs with respect to the repair intervals keeping the design parameter p constant.

C o s t-b e n e f i t o p t i m i z a t i o n f o r m a i n t a i n e d s t r u c t u r e s

4.5

353

Bloc k repairs

The damage cost for block repairs are composed of the (discounted) cost of planned systematic renewals at time d (or d1 for the first interval, where the time to the first failure has the updated failure time density f1 (t, p)) plus the (discounted) cost of failure(s) before d (or d1 ). Therefore, DB (p, d1 , d) = R(p)e−γd1 + (C(p) + L)[f1∗∗ (γ, p, d1 ) + m∗∗ 1 (γ, p, d1 )] +

e−γd1 [R(p)e−γd + (C(p) + L)f ∗∗ (γ, p, d)[1 + m∗∗ (γ, p, d)]] 1 − e−γd (56)

(d ( d1 −γt ∗∗ (γ, p, d1 ) = 0 (1) e−γt f(1) (t, p)dt, m∗∗ m(1) (t, p)dt with where f(1) (1) (γ, p, d1 ) = 0 e m1 (t, p) for the updated failure rate to the first renewal until d1 and m(t, p) in subsequent intervals until d as the renewal (Cox 1962). m1 (t, p) intensities in Equation (15) ∞ f (t, p) and m(t, p) = and m(t, p) are given by m1 (t, p) = ∞ n=1 1,n n=1 fn (t, p), respectively (see Equation (15)). Here and in the following the notation x(1) means either x or x1 whatever is relevant. Remember, integration of m(1) (t) is simply the mean number of renewals in [0, d(1) ] but here discounting is introduced additionally. The computation of m(1) (t) is the numerically expensive part (see Equation (17), Equation (19) or Equation (21)). Note that all components are repaired at time d(1) with certainty and cost R(p) but some components are already renewed earlier because they failed. For f1 (t, p) = f (t, p) and d1 = d Equation (56) simplifies to DB (p, d) =

R(p)e−γd + (C(p) + L)f ∗∗ (γ, p, d)[1 + m∗∗ (γ, p, d)] 1 − e−γd

(57)

For benefit rates unaffected by renewals one simply has the results in Equation (34) or (33). The benefit term for the case in Equation (35) is for finite integration intervals [0, d].
BB (p, d1 , d) =

d1

BD (t)f1 (t, p)dt +

0

e−γd1 + 1 − e−γd




d

m∗∗ 1 (γ, p, d1 )




d1

BD (t)f (t, p)dt 0 ∗∗




BD (t)f (t, p)dt + m (γ, p, d)

0



d

BD (t)f (t, p)dt 0

(58) with BD (t) in Equation (35). For f1 (t, p) = f (t, p) and d1 = d Equation (58) simplifies to (d BB (p, d) =

0

BD (t)f (t, p)dt + m∗∗ (γ, p, d) 1 − e−γd

(d 0

BD (t)f (t, p)dt

(59)

The length d (and/or d1 ) of a replacement interval can also be subject to optimization with respect to benefits and cost. In general, there is little difference between agedependent and block repairs unless the failure cost are very large.

354

Structural design optimization considering uncertainties

4.6 Inspec ti o n and r epair In the structures and many other areas any expensive maintenance operation is preceded by inspections involving cost I if damage progression and/or changes in system performance are observable. We understand that the inspections are essential inspections leading eventually to decisions about repair or no repair. If there are inspections at times a(1) , 2a(1) , 3a(1) , . . . there is not necessarily a repair because aging processes and inspections are uncertain or the signs of deterioration are vague. Repairs occur only with a certain probability PR (t), for example according to Equation (46). For cumulative damage phenomena this probability should increase with time as in Equation (46) and should depend on the actual observed damage state. As mentioned before the same (physical or chemical) damage process determines an (observable) damage state but also failure. For this reason inspection results and thus repair events and failure events are dependent. In fact, only if inspections address the same damage process, specifically the same realization, can we expect to make reasonable decisions about repair or no repair. Such dependencies makes an analytical and numerical treatment complicated but still computationally manageable. The objective is ZIR (p, a1 , a) = BIR (p, a1 , a) − CR (p) − DIR (p, a1 , a)

(60)

where in generalizing Equation (53) DIR (p, a1 , a) = N1 +

N2 N3 D

(61)

with:

N1 = (C(p) + L)

n=1

+I

∞ 

⎛ ⎞ n−1 d ⎝: exp[−γt] P {R(ja1 )} ∩ {T1 ≤ θ}⎠ dθ (n−1)a1

∞


na1

j=0

⎛ exp[−γ(na1 )]P⎝{R(na1 )} ∩

n=1

+ (I + R(p))

∞ 

n−1 :

exp[−γna1 ]P⎝{R(na1 )} ∩

n=1

+

⎞

n−1 :

{R(ja1 )} ∩ {T1 > na1 }⎠ (62a)

j=0

⎛ ⎞ n−1 d ⎝: exp[−γt] P {R(ja1 )} ∩ {T1 ≤ θ}⎠ dθ (n−1)a1

∞


∞  n=1

|θ=t

{R(ja1 )} ∩ {T1 > na1 }⎠

j=0

⎛

n=1

N2 =

⎞

dt

na1

j=0

⎛ exp[−γna1 ]P⎝{R(na1 )} ∩

|θ=t

n−1 :

dt ⎞

{R(ja1 )} ∩ {T1 > na1 }⎠

j=0

(62b)

C o s t-b e n e f i t o p t i m i z a t i o n f o r m a i n t a i n e d s t r u c t u r e s

355

⎛ ⎞ n−1 d ⎝: N3 = (C(p) + L) exp[−γt] P {R(ja)} ∩ {T ≤ θ}⎠ dt dθ n=1 (n−1)a j=0 |θ=t ⎛ ⎞ ∞ n−1  : exp[−γ(na)]P⎝{R(na)} ∩ {R(ja)} ∩ {T > na}⎠ +I ∞


na

n=1

+ (I + R(p))

∞ 

⎛

j=0

exp[−γna]P⎝{R(na)} ∩

n=1

⎞

n−1 :

{R(ja)} ∩ {T > na}⎠

⎛ ⎞ n−1 d ⎝: exp[−γt] P {R(ja)} ∩ {T ≤ θ}⎠ dt D=1− dθ n=1 (n−1)a j=0 |θ=t ⎛ ⎞ ∞ n−1  : exp[−γna]P⎝{R(na)} ∩ − {R(ja)} ∩ {T > na}⎠ ∞


(62c)

j=0

na

n=1

(62d)

j=0

and CR (p) = cost of investigating and/or retrofitting an existing structure C(p) = reconstruction cost after failure L = direct damage cost after failure R(na(1) ) = repair event at the j-th inspection R(ja(1) ) = no-repair event at the j-th inspection PR (ja(1) ) = probability of repair after the j-th inspection PR (ja(1) ) = (1 − PR (ja(1) )) = probability of no repair after the j-th inspection a(1) = deterministic inspection interval I = cost per inspection R(p) = repair cost for preventive maintenance. The first term N1 in Equation (61) is the replacement cost after first failure or repair, N3 the replacement cost for subsequent renewal cycles. In both cases the replacement cost include the cost of failure, the cost of inspections given that no failure and no repairs have occurred before and the third term accounts for the cost of inspection and repair given that no failure occurred before. Here, one has to extend the renewal interval to 2a(1) , 3a(1) , . . . if an inspection is not followed by repair and no failure occurred. Since PR (a(1) ) < 1 it is usually sufficient to consider only a few terms in the sums. The higher order terms vanish for PR (a(1) ) → 1 and are significant only for relatively small a(1) . As concerns numerical computions consider the fractional Laplace transform of the failure density given dependencies between no repair and failure events, that is (Joanni and Rackwitz 2006). ∗∗∗ f(1) (γ, p, (n − 1)a(1) ≤ t ≤ na(1) ) ⎛ ⎞
na(1) n−1 d ⎝: = exp[−γt] P {R(ja(1) )} ∩ {T(1) ≤ θ}⎠ dθ (n−1)a(1) j=0

|θ=t

dt

(63)

356

Structural design optimization considering uncertainties

where T(1) is the random time to failure. Here again, the intersection probabilities can be determined by FORM/SORM but alternative methods such as Monte Carlo n−1 simulation can also be used. Remember that a typical intersection event j=0 {R(ja(1) )} ∩ {T(1) ≤ t} after the probability distribution transformation into  standard space is given by n−1 j=0 {sc − S(ja(1) , UR ) > 0} ∩ {g(1) (UF , t) ≤ 0} according to Equations (5) and (46), for example. UR and UF denote the variables in the random vector defining the damage indicator (including measurement error) and the variables defining failure, respectively. Because UR and UF have some components in common the events are dependent. Within FORM/SORM the event boundaries are now linearized in the most likely failure point(s) and the correlation coefficients between the respective state functions are computed. The dependencies can be taken into account by evaluating the corresponding multivariate normal integrals. The differentiation under the integral that is necessary for evaluation of Equation (63) is best done numerically, but can also be performed analytically under certain conditions. For F1 (t) = F(t) the damage term in Equation (61) simplifies to DIR =

N3 . D

(64)

The benefit is given by Equation (33) or (34) if it is unaffected by renewals. It has a similar structure as Equation (61). Generalizing Equation (54) for the model in Equation (35) one obtains BIR (p, a1 , a) = B1 +

B2 B3 D

(65)

and B1 =

n=1

B2 =

⎡

∞


na1 (n−1)a1

d ⎣: P P({R(ja1 )} ∩ {T1 ≤ θ})|θ=t dt dθ j=0 ⎛ ⎞⎤ ∞ n−1  : + B∗D (na1 )P ⎝ {R(ja1 )} ∩ {T1 > na1 }⎠⎦ (66a)

⎡

∞
 n=1

B∗D (t)

na1 (n−1)a1

n−1

n=1

j=0

d ⎣: P P({R(ja1 )} ∩ {T1 ≤ θ})|θ=t dt + exp[−γ(na1 )] dθ j=0 ⎛ ⎞⎤ n−1 : {R(ja1 )} ∩ {T1 > na1 }⎠⎦ × P⎝{R(na1 )} ∩ n−1

(66b)

j=0

B3 =

⎛ ⎞ n−1 : d B∗D (t) P⎝ {R(ja)} ∩ {T ≤ t}⎠ dθ (n−1)a

∞
 n=1

+

∞  n=1

na

⎛ B∗D (na)P⎝

j=0

n−1 : j=0

⎞

{R(ja)} ∩ {T > na}⎠

dt |θ=t

(66c)

C o s t-b e n e f i t o p t i m i z a t i o n f o r m a i n t a i n e d s t r u c t u r e s

where BD (t) is given in Equation (37) and
t ∗ BD (t) = exp [−γ τ]b(τ)dτ.

357

(67)

(n−1)a(1)

For F1 = F(t) an analogous simplification as in Equation (64) is possible. For independent repair and failure events the intersection signs must simply be replaced by product signs simplifying the numerical computations considerably. The question is when the independence assumption becomes at least approximately true. This must depend on the case under consideration. Dependencies become weaker for larger measurement errors during inspections and for smaller dependencies between damage indicators and failure processes. 4.7

Preventive maintenance for s eries s y s te ms

By definition, a series system fails if any of its components fails. Consequently, all of its components have to renewed. This requires only a few modifications of the theory developed in Section 4.6. For a system with s components we have ⎛ ⎞ ∞
na1 n−1 s  : dP ⎝ : N1s = (C(p) + L) exp[−γt] × (−1) {R(ja1 )} ∩ {Tm1 > θ}⎠ dt dθ (n−1)a1 n=1

+I

∞ 

j=1

⎛

exp[−γ(na1 )]P⎝{R(na1 )} ∩

n=1

+ (I + R(p))

∞ 

n−1 :

∞
 n=1

+

exp[−γt] × (−1)

s :

{R(ja1 )} ∩

N3s = (C(p) + L)

n=1 ∞ 

m=1

+ (I + R(p))

∞  n=1

m=1

n−1 :

s :

j=0

m=1

{R(ja1 )} ∩

(68a) dt

θ=t

⎞

{Tm1 > na1 }⎠

(68b)

⎞ ⎛ n−1 s : dP ⎝ : exp[−γt] × (−1) {R(ja)} ∩ {Tm > θ}⎠ dθ (n−1)a na

j=1

⎛

exp[−γ(na)]P⎝{R(na)} ∩

n=1

{Tm1 > na1 }⎠ ⎞

j=1

exp[−γna1 ]P⎝{R(na1 )} ∩ ∞


⎞

n−1 s : dP ⎝ : {R(ja1 )} ∩ {Tm1 > θ}⎠ dθ

⎛

n=1

+I

n−1 : j=0

(n−1)a

∞ 

{Tm1 > na1 }⎠

⎛

na

θ=t

⎞

m=1

exp[−γna1 ]P⎝{R(na1 )} ∩

n=1

N2s =

s :

{R(ja1 )} ∩

j=0

⎛

m=1

⎛

m=1

n−1 :

s :

j=0

m=1

{R(ja)} ∩

exp[−γna]P⎝{R(na)} ∩

n−1 : j=0

dt

θ=t

⎞

{Tm1 > na1 }⎠

{R(ja)} ∩

s :

⎞ {Tm > na}⎠

m=1

(68c)

358

Structural design optimization considering uncertainties

Ds = 1 −

∞
 n=1

+

∞ 

⎞ ⎛ n−1 s : dP ⎝ : exp [−γt] × (−1) {R(ja)} ∩ {Tm > θ}⎠ dθ (n−1)a na

j=1

⎛

exp[−γna]P ⎝R(na) ∩

n=1

n−1 :

s :

R(ja) ∩

j=0

m=1

⎞

dt θ=t

{Tm > na}⎠

(68d)

m=1

in Equation (61) with N1 , N2 , N3 and Dd replaced by N1s , N2s , N3s and Ds , respectively. Similar modifications have to made for the benefit term.

B1s =

⎛ ⎞ n−1 s : : dP ⎝ {R(ja1 )} ∩ B∗D (t) × (−1) {Tm1 > θ}⎠ dθ (n−1)a1

∞
 n=1

+

∞ 

na1

j=1

⎛ B∗D (na1 )P ⎝{R(na1 )} ∩

n=1

B2s =

+

n−1 :

s :

j=0

m=1

{R(ja1 )} ∩

+

(69a)

⎛

exp [−γ(na1 )]P ⎝{R(na1 )} ∩

n−1 :

s :

j=0

m=1

{R(ja1 )} ∩

dt θ=t

⎞

{Tm1 > na1 }⎠

(69b)

⎞ ⎛ n−1 s : : dP ⎝ {R(ja)} ∩ B∗D (t) × (−1) {Tm > θ}⎠ dθ (n−1)a

∞
 n=1

{Tm1 > na1 }⎠

m=1

n=1

B3s =

θ=t

na1

j=1

∞ 

dt

⎞

⎛ ⎞ n−1 s : dP ⎝ : ×(−1) {R(ja1 )} ∩ {Tm1 > θ}⎠ dθ (n−1)a1

∞
 n=1

m=1

∞  n=1

na

⎛ B∗D (na)P⎝{R(na)} ∩

j=1

m=1

n−1 :

s :

j=0

m=1

{R(ja)} ∩

⎞

{Tm > na}⎠

θ=t

(69c)

  Here, we have used again P( s=1 E ) = 1 − P( sm=1 Em ) in order to retain operations solely on intersections. The series system is a realistic assumption for many but not for all civil engineering systems. For example, if one bridge of several bridges in a road connection between A and B fails or a river dam breaks at a certain point the infrastructure or flood protection system fails but only the failed bridge or dam section must be restored in order to make the system functional again. This may require certain modifications in the models outlined so far. If the block maintenance regime is chosen, all components in the systems will be restored. But if the age-dependent regime with inspection and repair is chosen any repair action may also be associated to a specific component. An analytical treatment will then become rather difficult and complex because the components in the system will have different ages. More complex systems can involve considerable numerical effort.

C o s t-b e n e f i t o p t i m i z a t i o n f o r m a i n t a i n e d s t r u c t u r e s

359

5 Some remarks about suitable optimization methods 5.1

G eneral

It is necessary to speak a little about the technical aspects of optimization. When designing and applying appropriate optimization techniques to the objective functions derived in the foregoing sections one faces the problem that, in fact, two optimization tasks have to be solved: (i) Optimization with respect to the design parameter p and (ii) Optimization with respect to the standard vector u to find the (local) reliability index, at least if FORM/SORM methods are applied. More specifically, the reliability optimization has to be solved for each step in the design parameter optimization. Even if one assumes differentiability of the objective and in the stochastic model as well as in the structural state function and uniqueness of the solution point(s), overall optimization still is a formidable task requiring quite some numerical effort. In the recent literature one distinguishes between one-level and bi-level optimization methods. For the bi-level method one optimizer solves the cost benefit optimization and another, possibly different optimizer solves the reliability optimization. In the one-level approach both optimization tasks are solved simultaneously by a suitable optimizer. In the following we shall briefly comment on both concepts. Both usually work and it is a matter of taste to select one or the other. If the abovementioned smoothness properties do not hold, then other optimization procedures are in order. 5.2

Bi-level optimization

In order to obtain the set of parameters for which the objective function Z(p) becomes optimal, the so-called bi-level approach can be chosen. Here, the optimization task in standard normal space for computation of the required reliability statistics corresponding to a fixed parameter set p is carried out separately using one of the sequential quadratic programming or similar methods. The results, in turn, serve as input to the main optimization loop for the parameter set p for which any of the available optimization methods can be employed. Alternatively, a direct search optimization method developed by (Powell 1994) can be applied. It does not require derivatives. This approach proved to be robust and reliable and only slightly more expensive than other methods. For the main optimization loop, lower and upper bounds should be imposed on the parameters, and it usually turns out to be advantageous to scale the optimization domain such that its shape becomes a hypercube. 5.3

One-level optimization

Let p be a parameter vector which enters in both the cost function and the limit state function g(u, p) = 0. Benefit, construction and damage function as well as the limit state function(s) are differentiable in p and u. The conditions for the application of FORM/SORM hold. In the so-called β-point u∗ the optimality conditions (Kuhn-Tucker conditions) are (Kuschel and Rackwitz 1997): g(u, p) = 0 u ∇u g(u, p) = − u ∇u g(u, p)

(70)

360

Structural design optimization considering uncertainties

The geometrical meaning of Equation (70) is that the gradient of g(u, p) = 0 is perpendicular to the vector of direction cosines of u∗ . The basic idea mentioned first in (Madsen and FriisHansen 1992) and elaborated in (Kuschel and Rackwitz 1997) now is to use these conditions as constraints in the cost optimization problem thus avoiding a bi-level optimization. It will turn out that this concept is crucial for further numerical analysis. For example, for the model in Equation (43) this leads to: Z(p) = B − C(p) − (C(p) + L)

subject to

λPf (p) γ

(71)

g(u, p) = 0 ui ∇u g(u, p) + ∇u g(u, p)i u = 0; i = 1, . . . , n − 1 hk (p) ≤ 0, k = 1, . . . , q

where hk (p) ≤ 0, k = 1, . . . , q are some constraints on the admissible parameter range. One may also add a constraint on the failure rate λPf (p). It is important to reduce the set of the gradient conditions in the Kuhn-Tucker conditions by one. Otherwise the system of Kuhn-Tucker conditions is overdetermined. It is also important that the remaining Kuhn-Tucker conditions are retained under all circumstances, for example, if one or more gradient Kuhn-Tucker conditions become co-linear with one or more of the other constraints. Otherwise the β-point conditions are not fulfilled. λP (p) must simply be replaced by If there are multiple failure modes (C(p) + L) γf  s λ (C(p) + L)(1 − P( j=1 F j (p))) (see Equation (44)). In this case γ ⎞⎤ ⎡ ⎛ s : λ⎣ Z(p) ≤ B − C(p) − (C(p) + L) 1 − P⎝ F j (p)⎠⎦ γ

(72)

j=1

subject to gk (uk , p) = 0; k = 1, . . . , s ui,k ∇u gk (uk , p) + ∇u gk (uk , p)i uk  = 0; i = 1, . . . , nk − 1; k = 1, . . . , s h (p) ≤ 0,  = 1, . . . , q where the Kuhn-Tucker conditions have to be fulfilled separately for each failure mode. Note that there are s distinct independent vectors uk . It may be noted that all failure mode equations are fulfilled simultaneously. For (locally) non-stationary problems, especially aging problems and for problems with non-Poissonian failures, it is possible to propose a numerical solution. More precisely, the Laplace transform is taken numerically and each value of the failure density is computed by FORM/SORM. The same scheme, however, applies to the full

C o s t-b e n e f i t o p t i m i z a t i o n f o r m a i n t a i n e d s t r u c t u r e s

361

Laplace transform of non-stationary problems as well. Z(p) ≈ B − C(p) − (C(p) + H) g(uj , p,tj ) = 0

f ∗ (γ, p) 1 − f ∗ (γ, p)

(73)

for j = 0, 1, . . . , m

ui,j ∇u g(uj , p,tj ) + ∇u g(uj , p,tj )i uj  = 0

i = 1, . . . , n − 1; j = 0, . . . , m

h (p) ≤ 0,  = 1, . . . , q where f ∗ (γ, p) ≈

m 

wj exp[−γtj ]fT (tj , p)

(74)

j=0

with m the number of time steps and wj the weights for numerical integration of Equation (74). In order to solve the optimization problem a suitable optimization algorithm is required. Unfortunately, off-shelf sequential quadratic programming methods turned out to have problems, possibly due to the many equality constraints. Based on sequential linear programming methods a new optimization algorithm JOINT 5 (Pshenichnyj 1994) has been developed from an earlier algorithm proposed by Enevoldsen and Sørensen (Enevoldsen and Sørensen 1992). This turned out necessary because the tasks in Equations (71), (72) and (73) require special precautions which are not necessarily available in most of the off-shelf algorithms. For example, the algorithm includes a reliable and robust slow down strategy to improve stability of the algorithm instead of an exact (or approximate) line search which too often is the reason for non-convergence. A special ‘extended’ equation system is solved in case of failure in the quadratic subalgorithm, e.g. due to linear dependence of the linearized constraints. In addition, the algorithm contains a careful active set strategy (for further details see (Streicher 2004)). As in the bi-level method a suitable scaling of the objective is advantageous. Gradient-based methods need first derivatives of the objective and all active constraints. In case of cost optimization under reliability constraints first order KuhnTucker optimality conditions for a design point are restrictions to the optimization problem. These equations are given in terms of the first derivatives of the limit state function. The gradients of these conditions involve second derivatives. Thus, the solution of the quadratic subproblem needs second derivatives, i.e. the complete Hessian of g(u, p). The determination of the Hessian in each iteration step is laborious and can be numerically inexact. In order to avoid this, an approximation by iteration is proposed. The Hessian is first preset with zeros. Note that linear limit state functions always have a zero Hessian matrix. This implies loss of efficiency, but the overall numerical effort needs not to rise, because calculation of the Hessian is no more necessary. In order to improve the results in case of strongly nonlinear limit state functions, it is possible to evaluate the Hessian after the first optimization run and restart the algorithm. For the improved solution the starting point is the solution of the previous run and the Hessian matrix is fixed for the whole run. This iterative improvement with subsequent restarts continues until the results differ only with respect to a given precision which is usually after very few steps. The results can be simultaneously improved by including

362

Structural design optimization considering uncertainties

second-order corrections during reiteration (see Kuschel and Rackwitz 2000). Any other more exact improvement can be taken into account in a similar manner. The techniques proposed enable the solution of quite general problems. They are based on a one-level optimization but rather strong requirements on differentiability of the objective, limit state functions and other restrictions must be made. Also, a possibly substantial increase of the problem dimension must be expected in extreme cases and, hence, much computing time will be necessary. In passing it is worthwhile to remark that for the bi-level approach a proof of convergence is not yet available whereas it is available for the one-level approach.

6 Illustrating example – Chloride attack in an existing building The following, slightly academic example shows several interesting features and is an appropriate test case. Chloride attack due to salting and subsequent corrosion, for example, in the entrance area of a parking house or in a concrete bridge is considered. A simplified, approximate model for chloride concentration in concrete is C(x, t) = Cs (1 − erf( 2√xDt )) where Cs = surface chloride content (measured ≈ 0.5 cm below surface and extrapolated), x = depth and D = diffusion parameter. A suitable criterion for the time to the start of chloride corrosion of the reinforcement is: 

 c Ccr − Cs 1 − erf √ ≤0 (75) 2 Dt where Ccr = critical chloride content and c = concrete cover. Inversion gives the initiation time

  Ccr −2 c2 −1 Ti = 1− erf (76) 4D Cs The stochastic model is Variable

[unit]

Distr. function

Parameters

C cr Cs c D

% % cm

Uniform Uniform Log-normal Uniform

0.4, 0.6 0.8, 1.2 mc , 0.8 0.5, 0.8

cm2 year

The planned concrete cover is mc = 5.0 cm. By drilling small holes and analyzing chemically the drill dust one has determined a chloride concentration of 0.4 at a depth of 3 cm with measurement error 0.05. Applying Equation (26) and truncation at t = 12 years gives an updated distribution function of the time to the start of corrosion as shown in Figure 13.1, where it is compared with the initiation time distribution. It is seen that chloride penetration occurred slightly more rapid than expected in renewed structures. During initiation time the structure can fail due to time-variant, stationary extreme loading. It is assumed that each year there is an independent extreme realization of the load. Load effects are normally distributed with mean 2.0 and coefficient of

C o s t-b e n e f i t o p t i m i z a t i o n f o r m a i n t a i n e d s t r u c t u r e s

363

1.0

0.8

0.6

0.4

0.2 Regular distribution Updated distribution 24

48

72 96 Time [years]

120

144

Figure 13.1 Updated distribution for first failure time and subsequent failure time distributions.

variation of 40%. Structural resistance is also distributed normally with mean 3-times as large as the mean load effect and coefficient of variation 30% (p = 6.0). Once corrosion has started mean resistance deteriorates with rate δ(t) = 1 − 0.07t + 0.00002t 2 . The distribution and density functions of the time to first failure are computed using SORM in Equation (13) with the failure time distributions in the initiation phase and in the deterioration phase determined by Equation (7). The structural states in two arbitrary time steps have constant correlation coefficient of ρ = ρij = 0.973. First, the mean times of the various distributions in Equation (13) are determined. One finds E[Ti ] = 51 and E[Td ] = 9.4. The mean of Te does not exist. Using the distribution in Equation (13) one determines E[T] = 61. These mean times indicate that virtually no failures occur during initiation. The risk functions for both distributions assuming the repair probabilities in Figure 13.2 are first increasing but decrease slightly for larger t. Visual inspections, inspections by half-cell potential measurements and chemical analyses are performed at regular intervals a(1) . They are followed by renewals (repairs) with probability   ) *  mc,(1) ≤0 (77) PA (a(1) ) = P r(1 + 0.05UR ) − Cs,R 1 − erf  2 DR a(1) shown in Figure 13.2 if a chloride concentration of r at the reinforcement was observed. The term (1 + 0.2UR ) models the measurement error with UR a standard normal variable. Repair times are assumed negligibly short. Remember, the existing structure is already 12 years old and has suffered from chloride attack during the whole period. The first inspection is undertaken after 5 years. For all subsequent renewed structures the first inspection is after 8 years. Erection cost are C(mc , mr ) = C0 + C1 m2c + C2 mr , inspection cost are I = 0.02C0 , and we have C0 = 106 , C1 = C2 = 104 , L = 10C0 , γ = 0.03. For preventive repairs the cost are R(mc , mr ) = 0.6C(mc , mr ) · mr is the safety

364

Structural design optimization considering uncertainties

1.0

Probability of repair

0.8

0.6

0.4

0.2 Regular probability, r
0.42 Updated probability, r
0.43 16

32 48 Time [years]

64

80

Figure 13.2 Repair probabilities.

Expected maintenance cost [106 MU]

8 1 unit 2 units 5 units

7 6 5 4 3 2 1

6

12 Replacement age a [a]

18

24

Figure 13.3 Age replacement.

factor separating the means of load effect and resistance. The benefit is determined by using a decaying rate b(t) = b exp [−0.0001t 2 ], b = 0.15C0 , in the model in Equation (35). All cost are in appropriate currency units. It is noted that the physical and cost parameters are somewhat extreme but not yet unrealistic. When optimizing with respect to the inspection interval the Laplace transforms are taken numerically using Simpson’s integration formula. We first show the total cost (preventive and corrective) for the case of systematic age-dependent repairs and system sizes of s = 1, 2 and 5 (Figure 13.3). As expected, the total cost are higher for larger systems and the optimum replacement interval decreases.

C o s t-b e n e f i t o p t i m i z a t i o n f o r m a i n t a i n e d s t r u c t u r e s

365

Expected maintenance cost in [106 MU]

8 1 unit, r
0.41 2 units, r
0.36 5 units, r
0.33

7 6 5 First inspection after 8 years 4 3 2 1

6

12 18 Inspection interval a [a]

24

Figure 13.4 Total cost for inspection and repair.

6 12 18 24 Inspection interval a [a]

18 12 6

4 4.4

1

Inspection interval a [a]

1

Inspection interval a [a]

1.9

*

1

n
5, r
0.35, r
0.32, D
2.30×106 24

3.6 3.2 2.8

12 6

6

2.4

6 12 18 24 Inspection interval a [a]

18

2.9 2.7

1.6

6

*

3.1

2

12

1

n
2, r
0.40, r
0.36, D
1.84 ×10 24

2.5 2.3 1 2.

1.8

2.2

18

6

1

*

Inspection interval a [a]

1

n
1, r
0.43, r
0.42, D
1.55×10 24

6 12 18 24 Inspection interval a [a]

Figure 13.5 Expected maintenance cost of an existing n-unit structure in [106 MU], with periodic inspections at an interval of a1 and a beginning after 5 and 8 years, respectively.

Figure 13.4 shows the results for the inspection/repair strategy. Here, we have also optimized the repair thresholds r. They become more stringent for larger systems. Also, the optimum inspection/replacement intervals are much smaller than in the simple agedependent case. The differences in cost between systematic age-dependent repairs and repairs after inspections are not large in this example. By parameter changes it is, however, easy to make them larger. The result of an optimization with respect to a1 and a is shown in Figure 13.5 for mc = 5 and mr = 6. One sees that the contour lines are spaced more narrow for a1 than for a. The optima with respect to a and a1 are rather flat. If, however, the repair probabilities are much smaller than given in (2) no optimum would be found. The inspection intervals depend strongly on the system size.

366

Structural design optimization considering uncertainties

7 Conclusions The theory developed earlier for optimal design and maintenance of aging structural components and systems is extended to optimal repair and retrofit of existing structures. It is assumed that structures are maintained (inspected and repaired with certain probability) at regular time intervals and systematically reconstructed after failure. Age-dependent and block repairs are studied assuming negligibly short repair times. Three models for the benefit are discussed. Due to updating by additional investigations, the time to first failure usually has different probabilistic characteristics than all other times. Appropriate objective functions for cost-benefit optimization are derived. It is pointed out that inspections and possible repair events and failure events must address the same realization of the damage process if preventive maintenance makes at all sense. Even if the risk function initially was increasing, maintenance operations will let the risk function drop. Perfect inspections and repairs will reduce the risk function down to zero. For imperfect inspections the risk function will drop down to finite values. This generally requires the numerical computation of the renewal intensity by differentiating the renewal function for which tight bounds can be given.

References Aitchison, J. & Dunsmore, I.R. 1975. Statistical Prediction Analysis. New York: Cambridge University Press. Ambartzumian, R., Der Kiureghian, A., Ohaniana, V. & Sukiasiana, H. 1998. Multinormal probability by sequential conditioned importance sampling: Theory and application. Probabilistic Engineering Mechanics 13(4):299–308. Au, S.-K. & Beck, J.L. 2001. First excursion probabilities for linear systems by very efficient importance sampling. Probabilistic Engineering Mechanics 16(3):193–207. Ayhan, H., Limón-Robles, J. & Wortman, M.A. 1999. An approach for computing tight numerical bounds on renewal functions. IEEE Transactions on Reliability 48(2): 182–188. Barlow, R.E. & Proschan, F. 1965. Mathematical Theory of Reliability. New York: John Wiley & Sons. Barlow, R.E. & Proschan, F. 1975. Statistical Theory of Reliability and Life Testing: Probabilistic Models. New York: Holt, Rinehart & Winston. Brown, M. & Proschan, F. 1983. Imperfect repair. Journal of Applied Probability 20: 851–859. Cox, D.R. 1962. Renewal Theory. Monographs on Applied Probability and Statistics. London: Chapman & Hall. Cox, D.R. & Isham, V. 1980. Point Processes. Monographs on Applied Probability and Statistics. London: Chapman & Hall. Cramér, H. & Leadbetter, M.R. 1967. Stationary and Related Stochastic Processes. New York: John Wiley & Sons. Dunnett, C.W. & Sobel, M. 1955. Approximations to the probability integral and certain percentage points of multivariate analogue of Student’s t-distribution. Biometrika 42: 258–260. Enevoldsen, I. & Sørensen, J. 1992. Optimization algorithms for calculation of the joint design point in parallel systems. Structural and Multidisciplinary Optimization 4(2):121–127. Fox, B. 1966. Age replacement with discounting. Operations Research 14(3):533–537.

C o s t-b e n e f i t o p t i m i z a t i o n f o r m a i n t a i n e d s t r u c t u r e s

367

Genz, A. 1992. Numerical computation of multivariate normal probabilities. Journal of Computational and Graphical Statistics 1:141–149. Gollwitzer, S. & Rackwitz, R. 1983. Equivalent components in first-order system reliability. Reliability Engineering 5:99–115. Gollwitzer, S. & Rackwitz, R. 1988. An efficient numerical solution to the multinormal integral. Probabilistic Engineering Mechanics 3(2):98–101. Hasofer, A. 1974. Design for infrequent overloads. Earthquake Engineering and Structural Dynamics 2(4). Hasofer, A.M. & Rackwitz, R. 2000. Time-dependent models for code optimization. In R.E. Melchers & M.G. Stewart (eds), Proceedings of the 8th International Conference on Applications of Statistics and Probability (ICASP8), Sydney, Australia, December 1999, Vol. 1, Rotterdam/Brookfield, pp. 151–158. CERRA: A.A. Balkema. Hohenbichler, M. & Rackwitz, R. 1981. Non-normal dependent vectors in structural safety. ASCE Journal of the Engineering Mechanics Division 107(6):1227–1249. Hohenbichler, M. & Rackwitz, R. 1983. First-order concepts in system reliability. Structural Safety 1(3):177–188. Joanni, A.E. & Rackwitz, R. 2006. Stochastic dependencies in inspection, repair and failure models. In C. Guedes Soares & E. Zio (eds), Proceedings of the European Safety and Reliability Conference, Estoril, Portugal, September 2006, London, pp. 531–537. Taylor & Francis. Kuschel, N. & Rackwitz, R. 1997. Two basic problems in reliability-based structural optimization. Mathematical Methods of Operations Research (ZOR) 46(3):309–333. Kuschel, N. & Rackwitz, R. 2000. Time-variant reliability-based structural optimization using SORM. Optimization 47(3/4):349–368. Madsen, H.O. & Friis-Hansen, P. 1992. A comparison of some algorithms for reliabilitybased structural optimization and sensitivity analysis. In R. Rackwitz & P. Thoft-Cristensen (eds), Proceedings of the 4th IFIP WG 7.5 Working Conference on Reliability and Optimization of Structural Systems, Munich, Germany, September 1991, Berlin, pp. 443–451. IFIP: Springer. Pandey, M.D. 1998. An effective approximation to evaluate multinormal integrals. Structural Safety 20(1):51–67. Powell, M.J.D. 1994. A direct search optimization method that models the objective and constraint functions by linear interpolation. In S. Gomez & J.-P. Hennart (eds), Proceedings of the 6th Workshop on Optimization and Numerical Analysis, Oaxaca, Mexico, January 1992, Dordrecht, pp. 51–67. Kluwer Academic Publishers. Pshenichnyj, B.N. 1994. The Linearization Method for Constrained Optimization, Volume 22 of Computational Mathematics. Berlin: Springer. Rackwitz, R. 2000. Optimization – the basis of code making and reliability verification. Structural Safety 22(1):27–60. Rackwitz, R. 2001. Reliability analysis – a review and some perspectives. Structural Safety 23(4):365–395. Rackwitz, R., Lentz, A. & Faber, M.H. 2005. Socio-economically sustainable civil engineering infrastructures by optimization. Structural Safety 27(3):187–229. Rosenblueth, E. 1976. Optimum Design for Infrequent Disturbances. ASCE Journal of the Structural Division 102(9):1807–1825. Rosenblueth, E. & Mendoza, E. 1971. Reliability optimization in isostatic structures. ASCE Journal of the Engineering Mechanics Division 97(6):1625–1642. Streicher, H. 2004. Zeitvariante zuverlässigkeitsorientierte Kosten-Nutzen-Optimierung für Strukturen unter Berücksichtigung neuer Ansätze für Erneuerungs- und Instandhaltungsmodelle. PhD dissertation, Technische Universität München, Munich, Germany. In German.

368

Structural design optimization considering uncertainties

Streicher, H., Joanni, A. & Rackwitz, R. 2006. Cost-benefit optimization and target reliability levels for existing, aging and maintained structures. Structural Safety. Accepted for publication. Streicher, H. & Rackwitz, R. 2004. Time-variant reliability-oriented structural optimization and a renewal model for life-cycle costing. Probabilistic Engineering Mechanics 19(1–2):171–183. Van Noortwijk, J.M. (2001). Cost-based criteria for obtaining optimal design decisions. In Proceedings of the 8th International Conference on Structural Safety and Reliability, Newport Beach, CA, U.S.A., June 2001. CD-ROM.

Chapter 14

A reliability-based maintenance optimization methodology Wu Y.-T. Applied Research Associates Inc., Raleigh, NC, USA

ABSTRACT: Many mechanical and structural systems, including aircraft, ship, car, oil and gas pipeline, utilize a structural integrity program to monitor and sustain structural integrity throughout the service life. Developing optimal maintenance plans under various uncertainties requires probabilistic analyses of damage accumulations, damage detections, and mitigation actions. Given the wide spectrum of the options and the complexities in modeling, the most practical way to conduct maintenance optimization is by random simulations, preferably efficient sampling methods. This chapter presents a reliability-based maintenance optimization (RBMO) methodology with a focus on computational strategies that involve physics-based models. In particular, a two-stage importance sampling (TIS) approach that drastically reduces computational time is described. Stage 1 computes failure probability and systematically generates failure samples, given no inspections. The failure samples are then repeatedly used in Stage 2 for inspection optimization. The RBMO methodology is demonstrated using analytical examples as well as applications related to aircraft and helicopter structural components.

1 Introduction For economical and reliability/safety reasons, many mechanical and structural systems apply maintenance practices to sustain structural integrity and reliability over the design life or extend the life beyond the original design for un-anticipated reasons. Since fatigue and fracture is one of the main failure modes for such systems, this chapter will focus on fracture failure analysis even though the methodology is applicable to more general damage accumulation models including corrosion. Most existing computational fracture mechanics methods and tools used in the design of structures apply safety margins to deterministic models. With the realization that many design parameters including defect or flaw characteristics, crack growth law, crack detection, loads, and usages are uncertain, various conservative assumptions are often employed to help ensure structural integrity. As an example, a comparison between deterministic and probabilistic damage tolerance analyses is shown in Table 14.1. The safety-factor based approach applies bounds, either explicitly or implicitly, to key design variables. The probabilistic approach, on the other hand, requires relatively more precise characterizations of the input uncertainties based on data and expert knowledge. In the more traditional safe-life design approach (Palmberg et al. 1987), the fatigue and facture life of a structure is assumed to be governed by crack initiation time, and

370

Structural design optimization considering uncertainties

Table 14.1 Example of deterministic vs. probabilistic damage tolerance.

Reliability principals Flaw/defect size Existence Inspection schedule Safety measure Other variables

Deterministic

Probabilistic

Bounds, safety factors A given crack size Probability = 1 Life/N Safety margin Bounds, safety factors

Probability & confidence Distribution of crack size 0<= Probability<= 1 Max. risk reduction Reliability = 1 − Pf Distributions

the products are designed to survive their design life with a safety margin. The safe-life approach is typically used for structures that are either very difficult to repair or, if failed, may cause severe consequences. These products are designed and built to work without the requirements of any repairs. One drawback is that the over-conservatism can cause high cost and poor performance (e.g., due to weight increase), and there are no provisions to extend the service life even though the product may still have a considerable remaining life after the design life. In addition, safe-design has been proven to be un-conservative in cases where there are initial defects that cannot be detected. To overcome these drawbacks, an alternative approach is to use damage tolerant design which recognizes and allows defects, with a provision that a plan needs to be in place to monitor damage and implement mitigation actions such as repair, replacement, load reduction, corrosion rate control, and other methods. Damage tolerant design is technically more challenging because it requires using capable detection devices to catch damage just in time – not too early when damage cannot be detected effectively, and not too late when the damage has grown to such a dangerous size that a failure is likely before the next inspection. Scheduled inspections have been applied for structures such as aircraft wings, engine blades and disks, and oil/gas transmission pipelines for safety, performance or economical reasons (Berens et al. 1991; Wu et al. 2002; Cunha et al. 2006.). However, because of lack of models and computational tools, inspection schedules are usually over-simplified. For example, the easiest scheduling approach is based on equally dividing the total service life by a number of inspections. This approach, which is simple, is clearly illinformed, especially for aging structures where the system deteriorates more rapidly toward the end of the service life. An optimal scheduling requires a reliability-based approach which is the focus of this chapter. Note that in this chapter “reliability-based’’ and “risk-based’’ are often used interchangeably even though “risk’’ usually involves consequences of failures such as risk = (failure probability) × (cost of failure). Using this simplified definition, and assuming the same cost of failure, reducing risk means reducing probability-of-failure.

2 Reliability-based methodology 2.1 Rel i a b i l i ty-b as e d d amag e t o ler anc e f r a m e w o r k Building on an earlier NASA Probabilistic Structural Analysis Methods program (Millwater et al. 1996) and recent FAA research work for risk assessment of aircraft

A r e l i a b i l i t y-b a s e d m a i n t e n a n c e o p t i m i z a t i o n m e t h o d o l o g y

371

turbine engines (Leverant et al. 1997; Wu et al. 2002; Millwater et al. 2000), a framework for reliability based damage tolerance (RBDT) has been developed in a feasibility study (Wu et al. 2004; Wu et al. 2005), summarized in Figure 14.1, that considers a wide range of uncertainties including: • • • • • • •

Random or uncertain parameters in material (e.g., threshold of the stress-intensity factor, modulus of elasticity) Defect or flaw (including size, shape, and location, and the frequency of occurrence) Loading, type of usage (with frequency of occurrence) Finite element model (including modeling error) Crack growth model (including modeling error) Random inspection time Probability of detection (POD).

A typical maintenance program includes inspection schedules, probability of detection curves, repairs, replacements, and other mitigation methods that are useful for slowing down or stopping the damage accumulation process. For brevity, the term “inspection optimization’’ is loosely used to mean optimizing the schedules of inspections associated with a maintenance program that defines repair and replacement requirements. Unless specifically mentioned, “repair’’ is a term loosely used to include replacement as a special case. RBDT models initial defects and other time-independent uncertain variables as random variables, tracks the distributions of the growing defects, and simulates detections and repairs of the defects by modifying defect probability distributions. A failure is assumed to occur when the defect size exceeds the critical size for fracture. Optimization of inspection schedules is based on minimizing probability of failure or, more generally, risk. In practical applications, constructing and updating crack size distribution may be computationally difficult. Alternatively, it is relatively straightforward to conduct RBDT using random simulation such as Monte Carlo. However, the downside is that standard Monte Carlo can be highly time-consuming. Therefore, approximate methods and more efficient sampling methods are of significant interest. To demonstrate RBDT for practical applications, a prototype RBDT software was developed that integrated a finite element stress analysis module, a fracture mechanics module, and a probabilistic module. The modular design is illustrated in Figure 14.2.

3 Probabilistic analysis methods 3.1

Frac ture failure limit-s tate functio n

Under cyclic loading, an initially small flaw may grow and cause a fracture when the stress-intensity factor K reaches the fracture toughness Kc . The fracture limit state is: K(X1 , . . . , Xn , Ns ) = Kc

(1)

Modeling error

1

0

Time

Time

Initial

Crack size

At insp.

POD

2nd insp.

Critical size

Inspection planning

1st insp.

Updated distribution after maintenance

Cumulative probability

Initial flaw

Crack

Crack growth

Figure 14.1 A reliability based damage tolerance framework.

Inspection time

Material

Flaw or FOD Stress

Usage load spectra

– FM life model

– FE stress model

Principal structural Element DT model

Integrated FE stress, FM life, and Probabilistic Analyses

Prob. of fracture

Failure samples

Flight hours

Max. risk reduction

Failure samples

Most likely failure point

Without inspection With inspection

Joint probability density

Two-Stage failure sampling

A r e l i a b i l i t y-b a s e d m a i n t e n a n c e o p t i m i z a t i o n m e t h o d o l o g y

373

User command FE stress module

Prob. analysis module

Generic function interface FM life module

Plug-in modules

Figure 14.2 RBDT software modular design.

The stress-intensity factor is a function of service life, Ns , and the random variable vector X that includes all but the inspection-related random parameters. Alternatively, the limit state can be expressed as: g(X , NS ) = Nf (X ) − NS

(2)

where Nf is cycles-to-failure. 3.2

Probability of failure without ins pe c ti o ns

The probability-of-failure without inspection, pof , can be formulated as:
pof = Pr [Nf (X ) ≤ Ns ] =


···

fX (X )dX

(3)

Nf ≤Ns

in which fX (X ) is the joint probability density function of X . Many methods originating from Structural Reliability could be used to compute pof (see e.g., books by Ang & Tang 1984; Madsen et al. 1986; Thoft-Christensen & Murotsu 1986; Melcher 1987; Ditlevsen & Madsen 1996; Nikolaidis et al. 2005). Relative to with-inspection, this integral is relatively easier to compute as many approximation methods are suitable and effective. For well-behaved functions, some of the more widely used approximation methods include First-Order Reliability Method and Second-Order Reliability Method (Schuëller 1998; Rackwitz 2001; Der Kiureghian 2005) which are based on approximation after transforming the original random variables to the standard normal variables, u. For well-behaved but implicitly defined g-functions (such as finite element models) requiring extensive computations, the Advanced Mean Value (AMV) method and its extension, AMV+, provide a quick estimate of the response CDF using n + 1 + number of CDF levels (Wu et al. 1990). Denoting the original CDF of X as FX (x) and the standard normal CDF of u as
(u), −1 the transformation between X and u is done by using xi = FX (
(ui )) for independent i

374

Structural design optimization considering uncertainties

random variables and Rosenblatt and other transformations for correlated random variables (Rosenblatt 1952; Ang & Tang 1984; Liu & Der Kiureghian 1986). In the u-space, it is more convenient to estimate the probability using linear or quadratic approximate g-functions. In the independent Gaussian space, Equation 3 becomes:
pof = Pr [Nf (X ) ≤ Ns ] =


···

φu (u)du

(4)

Nf ≤Ns

in which φu (u) is the standardized-normal Joint PDF (JPDF). Figure 14.3 shows the JPDF of a bivariate Gaussian after removing the JPDF from the failure region that represents the density volume of pf . When the g-function is well-behaved, the pf volume may be represented well by a volume-cut centered at the maximum JPDF point, or Most Probable Point, MPP (in structural reliability literatures, this point has been called Design Point, Minimum Distance Point, or Most Likely Failure Point). The MPP-based results can be represented as: pof =
( − β) · r1

(5)

In which β is the distance from the origin to the MPP, and r1 is an error factor (larger or smaller than 1). The FORM solution assumes r1 = 1. Equation 4 can be recast into Equation 6, a form suitable for using Monte Carlo simulation.

o (6) pf = · · · I(u)φu (u)du = E[I(u)] where E[.] is the expectation operator and φu (u) is the JPDF of standardized normal distribution. By sampling u-vector K times and taking the sampling average of the indicator function defined as I(u) = 1 for g <= 0 and 0 for g > 0, the probability integral can

Joint PDF

Max. JPDF | g <= 0 Limit state: g = 0

b

u1 Failure (g ⴝ 0) random samples

u2

pf
 (b)•r1

Figure 14.3 MPP-based method for computing pf .

A r e l i a b i l i t y-b a s e d m a i n t e n a n c e o p t i m i z a t i o n m e t h o d o l o g y

be estimated using the sampling average. By assuming the estimator pof = follows a binomial distribution, sampling error can be estimated using:  γ(%) = 100 · (−
−1 (α/2)) ·

1 − pf pf · K

375

K

k=1 Ik (u)/K

(7)

where γ(%) is the estimated error bound√at a confidence level of 1 − α. Equation 7 shows that the error is proportional to 1/ K. The MPP-based methods are suitable if the g-function can be well represented by a linear or quadratic surface, at least around the MPP. The fact that FORM/SORM cannot consistently provide accurate results has prompted the development of efficient sampling approaches to supplement the FORM/SORM methods (Harbitz 1986; Bucher 1988; Ditlevsen et al. 1989; Hohenbichler & Rackwitz 1988; Karamchandani & Cornell 1991; Kale et al. 2005). Typically these hybrid methods require the MPP solution to guide the selection of a sampling density to put more weight to where the failure is more likely. The selected importance sampling density h(u) must first satisfy the basic PDF requirement:



hu (u)du = 1

···

(8)

Using h(u) as the sampling density, Equation 4 can be written as:



pof =

··· Nf ≤Ns

 φu (u) φu (u) · hu (u)du = E hu (u) hu (u)

(9)

To achieve high efficiency, the shape of h(u) is usually selected to resemble the original failure density function and the ideal support is one that is slightly greater than the failure domain. An extreme case is h(u) = φu (u) for g <= 0 and h(u) = 0 for g > 0, resulting in:



···



φu (u) I φu (u)/pof

 ·

φu (u) du = E[pof ] = pof pof

(10)

which is a constant with zero variance. In practice, approaching the above limiting condition means one needs to somehow know pof already, which is not the case. Nevertheless, the limiting case suggests that the sampling variance is approaching zero when h(u) is approaching φu (u) and therefore one should select h(u) such that, while satisfying Equation 8, its shape mimics the shape of φu (u) in the failure domain and is zero otherwise. Examples of the importance sampling densities include those using the same φu (u) but with a slightly conservative limit state. While many researchers suggest that the MPP methods, including the hybrids, are effective for many practical problems, it should be emphasized that with these methods, though featuring excellent “local’’ approximations, there is no guarantee that the failure domain is sufficiently explored and modeled properly. This lack-of-assurance

376

Structural design optimization considering uncertainties

issue is somewhat similar to the situation of optimizing a function where there are multiple local minimums, and where optimizers could settle on a local minimum without a mechanism to jump out and search for a global minimum. However, in optimization, a local minimum is at least better than other solutions in an explored region, whereas in probabilistic analysis, a local MPP-based solution could have a large error in pof . Therefore, at the expense of losing efficiency, a more reliable h(u) should be selected that is less local and has the ability to self-correct the error in case the initially found MPP is not representative of the failure density.

3.3

Pro b ab i l i ty o f failur e wit h ins pec t i o n s

3.3.1 Prob a b i li ty o f d e t e ct io n m o d e l Optimal inspection schedules depend on how effective detection devices are and how fast defects grow. The effectiveness of a detection device is typically characterized using a POD(a) curve which is defined as the probability of detecting a defect with a size greater than or equal to a. The simplest POD model is a step function: POD is one if the defect is greater than a threshold and zero otherwise. The complementary function of POD is probability of non-detection, PND, i.e., PND(a) = 1 − PND(a). POD depends on the physics of the detection method (e.g., eddy current, magnetic flux, ultrasonic, vibrations) and other factors such as the damage mechanism, the geometry of the component, defect shape, location and accessibility, presence of coating or insulation, equipment sensitivity setting, measurement error, etc. Creating PODs by experiments using manufactured products are often difficult due to cost and time constraints. As an example of the attempts to overcome the issue, recently a number of industries in the Netherlands have initiated a joint industry project to develop a physics based POD model and corresponding software tool (Volker et al. 2004) for pipeline and chemical plants, among other applications. Sometimes damage data are available to correlate the detector-measured and the actual defect sizes. One example is that the dimensions of pipeline metal loss (e.g., depth, length, and width) can be estimated using an in-line inspection machine such as an MFL (magnetic flux leakage) device; when larger defects are found and some pipeline sections excavated, actual defect sizes may become available. Such verification data can be used to develop a statistical model that expresses the actual size as a function of the measured size with a sizing error term. Such an error term can be included as an additional random variable in RBMO. Alternatively, Non-Destructive Examination (NDE) equipment manufactures may provide a sizing-error bound at a specified confidence level. For example, MFL providers might specify the accuracy as: +/−15% depth/wall-thickness at 80 confidence level for general corrosion (e.g., Rosen 2004). This information can be used to develop a random error term by using, e.g., a Gaussian distribution. In the remaining portions of the chapter, we will focus on a simplified defect detection-and-repair model that assumes a binary result, detected or not-detected, and further assumes that a repair is always needed when a defect is detected. This is a reasonable model for safety-critical structures such as aircraft engines but could be too conservative for other types of structures, such as oil pipelines, where if a

A r e l i a b i l i t y-b a s e d m a i n t e n a n c e o p t i m i z a t i o n m e t h o d o l o g y

377

detected corrosion depth is a small fraction of the pipe wall thickness, the repair decision can wait until the next inspection. In the latter case, a repair threshold can be selected that corresponds to a tolerable risk and an “equivalent POD’’ can be developed by truncating the left “tail’’ such that POD = 0 for defects less than the threshold. 3.3.2

MPP-Based methods

Using the MPP-Based methods, the first step is to compute pof for Nf ≤ NI , where NI is the first inspection time. Immediately after the survived defected parts have been inspected, detected, measured, and repaired or replaced, the defect distribution will need to be updated to reflect a “healthier’’ population, and then the time is reset for a new probabilistic analysis for the next duration before reaching the next inspection time or the end of the service life, whichever comes first. The summation of the probabilityof-failure over the time increments is the cumulative probability of failure (White et al. 2002). The effectiveness of a maintenance plan can be measured by the difference between the cumulative probability of failure and pof . Updating fX (X ) after each maintenance is a difficult task using analytical methods. One approach is to make an additional approximation in FORM calculation (Harkness et al. 1994). Other approaches include FORM-based pf conditioned on inspection results where the detected-and-measured crack size is modeled as a random variable (Madsen et al. 1987). Based on the conditional-probability formulation, the FORM solution for system reliability updating were developed and demonstrated. While these methods can be used in certain applications, they are difficult to generalize especially to deal with more complex maintenance situations such as when the inspection schedule is random and when different detection devices (with different PODs) are to be applied to different defect locations. In contrast, the random simulation approach only requires tracking the defect population, and the process to simulate maintenance can be implemented easily without sophisticated computations. The only drawback is the need for a large number of samples to achieve accurate results, particularly for small pf problems. This provided the motivation for developing efficient sampling methods including TIS. 3.3.3 Two-st a ge i mportance sampl i ng (TIS) In the two-stage importance sampling approach, the first stage computes pof and generates failure samples without inspections. Stage 2 simulates inspection and repair using failure samples from Stage 1. The approach is built on the assumption that any random sample from a standard Monte Carlo (MC) method that is safe in Stage 1 will be safe in Stage 2. This assumption is valid as long as the maintenance program is sound and well executed, so that what is safe in Stage 1 part would not become unsafe after inspection and repair. The concept of TIS is illustrated in Figure 14.4. Using the TIS approach, the probability-of-failure with inspection, pW , is: f o pW f = pf ·

K − NRepaired K

(11)

378

Structural design optimization considering uncertainties

Defect size Critical crack size

Crack growth curves

Samples removed if detected based on POD

Defect population that would fail before service life, assuming no inspections

2nd inspection Defect PDF 1st inspection Service life N

Time (flight hours)

Figure 14.4 Illustration of two-stage importance sampling method.

Equation 11 shows that the risk is reduced by repairing (including replacing) those weaker parts before they fail. The reason TIS is efficient is because the computation of the indicator functions in the safe region in a full MC simulation is skipped completely. The computational challenge is to systematically generate samples directly from the failure domain. Three methods will be discussed. 3.3.3.1

E XA CT FA I L U R E-R E G I O N SA M P L I N G BY N U M E R I CA L I NT E G RAT I O N

The first method, most suitable for a small number of random variables in addition to inspection-related variables (such as inspection time and POD), is to generate random samples using conditional distributions constrained by the limit state. To illustrate, consider three random variables with a joint PDF of f (x1 , x2 , x3 ). The probability of failure can be formulated using a three-dimensional integral:


pf =

f (x1 , x2 , x3 )dx1 dx2 dx3

(12)

g(x)<=0

Now define a “truncated’’ distribution by letting f (x) = 0 in the region where g(x) > 0. In the failure region ! = [g(x) ≤ 0], the joint PDF becomes: f! =

f (x1 , x2 , x3 ) pf

(13)

A r e l i a b i l i t y-b a s e d m a i n t e n a n c e o p t i m i z a t i o n m e t h o d o l o g y

which can be used to derive the following marginal and conditional PDFs:

fX1 (x1 ) = f! (x1 , x2 , x3 )dx2 dx3 (( f! (x1 , x2 , x3 )dx1 dx3 fX2 (x2 ) = fX2 |x1 (x2 ) = fX1 (x1 ) fX1 (x1 ) (( f! (x1 , x2 , x3 )dx1 dx2 fX3 (x3 ) fX3 |x1 ,x2 (x3 ) = = f (x1 , x2 ) fX2 |x1 (x2 ) · fX1 (x1 )

379

(14)

The above PDFs lead to: FX1 (x1 ) = pC f (x1 ) FX2 |x1 (x2 ) = pC f (x2 |x1 )

(15)

FX3 |x1 ,x2 (x3 ) = pC f (x3 |x1 , x2 ) Finally, the inverses of the above marginal and conditional CDFs are used to generate failure samples in the following sequence (See, e.g., Ang & Tang 1984): −1 x1 = FX (U1 ) 1 −1 (U2 |x1 ) x2 = F X 2

x3 =

(16)

−1 FX (U3 |x1 , x2 ) 3

where Ui (i = 1 : 3)are uniform random variables. Note that the computation of Equation 15 can be done in the transformed u-space without explicitly using Equation 14. The above random number generation process involves n-fold integrations in the failure region. Thus, the increased complexity in higher-dimensional integration limits the method to simple g-functions with a few random variables. Nevertheless, when the implementation is practical, this approach provides effective generation of failure samples for TIS. The method has been used in an application involving aircraft engine disk design (Wu et al. 2002). 3.3.3.2

M P P-BA S E D I M P O RTA N C E SA M P L I N G W IT H C O N S E RVAT IV E L I M IT STAT E

The second method is MPP based, but with a conservative limit state designed to help correct MPP and FORM errors. The advantage relative to Method 1 is the capability to generate independent failure samples easily regardless of the number of random variables. The method is efficient and is recommended for users familiar with the MPP limitations. Using MPP with a conservative limit state, the TIS approach consists of four steps: Step 1. Using FORM or an equivalent method, compute the inspection-free risk pof . The FORM solution is pof =
(−β). Step 2. A second FORM analysis is conducted using an adjusted limit state defined as: g(X , NS ) = Nf (X ) − A · NS

(17)

380

Structural design optimization considering uncertainties

where A is a “safety factor’’ introduced to generate conservative samples in Step 3 to check the solution from Step 1. A is always greater than 1 so that the adjusted , failure region contains the true failure region and the adjusted failure probability, pA f is greater than pof . The value of A can be created based on the FORM result to anticipate a larger pf , say 30% larger, by parallel-shifting the MPP tangent surface towards the origin to reduce β. For example, let βFORM = 3, pof =
( − 3) = 0.00135. To increase pf by 30% requires the adjusted β to be βA = −
−1 (0.00135 ∗ 1.3) = 2.919. Assume that the service life is NS = 20 000, and at the shifted MPP the calculated life is N = 22 000, then the safety factor A ≈ 22 000/20 000 = 1.1. To speed up Step 2, we can use the first MPP as the initial guess to search for the second MPP. Step 3. Generate failure samples in the adjusted failure region using the (hyper-) tangent surface at the second MPP. The generation of the u-samples can be made more conveniently by an MPP-plane rotation so that the vector from the origin to the MPP coincides with the nth coordinate. The rotation matrix can be constructed using a plane rotation procedure such as the Gram-Schmidt process. The sample-generation procedure is as follows: (i) generate a “tail’’ sample of un from a one-dimensional normal pdf φ(u) truncated at u = βA , (ii) generate independent u1 to un−1 from φ(u), (iii) rotate the u sample back to the original u-space. The fracture lives of all the generated samples are computed. These samples will in general include failures samples as well as typically a small fraction of safe samples, as illustrated in Figure 14.5. The crack growth histories of the failure samples should be saved for Stage 2 analysis. The larger the safety factor A is, the more conservative the limit state is, and the larger the number of safe samples will be. In addition, the

A = 1 (FORM) Short life domain

A=1 (Exact)

u2

MPP

Long life domain

Nf < Ns Nf < A * Ns

bA

A >1 u1, Crack size

Figure 14.5 Compute pof using conservative limit state and 2-stage importance sampling.

A r e l i a b i l i t y-b a s e d m a i n t e n a n c e o p t i m i z a t i o n m e t h o d o l o g y

381

distances of the failed samples can be sorted and the sample with the smallest distance is the sample-based MPP. Step 4. Re-compute pof using the samples. The simulated samples with lives shorter than NS are used to compute a new pof using: pof =
(−βA ) ·

No. of failures without inspections Total number of samples

(18)

and the result is compared with the result from Step 1. If the new result is similar to the first result, it would provide an increased confidence that FORM gives a good estimate. In general, with a sufficient number of samples, the result using Equation 18 is better than the Step 1 FORM result because Equation 18 is capable of handling nonlinear gfunctions. In addition, the sample-based MPP can be compared with the FORM-based MPP from Step 1 to help accept or reject the MPP from Step 1. On the other hand, if the two results are significantly different, it would suggest the rejection of the first FORM solution and one might try Step 2 with a larger A value and more samples. If results do not converge after using more conservative A values, other methods should be used. Step 5. Compute pW using random simulations. f For each sample, a random number, U, between 0 and 1 is generated. For a simulated defect size a, POD(a) is compared with U; the defect is detected if POD(a) > U, and the inspected part is either repaired, replaced, or passed (if the defect size is considered safe). If the part is repaired or replaced, a new defect size is randomly drawn from a proper distribution and the simulation repeats until the end of the service life is reached. The purpose of the simulations is to compute conditional failure probability, pcf , i.e., the probability-of-failure with inspections conditioned on the total number of simulation samples. The final probability-of-failure with inspections is: c A pW f = pf · pf =

3.3.3.3

No. of failures with inspections A · pf Total number of TIS samples

(19)

E XA CT FA I L U R E-R E G I O N SA M P L I N G BY M C M C

A more recent approach to generating failure samples, practically without the constraints in the number of variables and the non-linearity of the limit states, is the Markov Chain Monte Carlo, or MCMC, method (Gamerman 1997; Gentle 1998; Robert & Casella 2004). The unique feature of MCMC is that the samples are generated sequentially using the ratio of the PDFs of the current and the next candidate states. This feature allows the generation of samples using f (x1 , x2 , . . . , xn ) without integrations. For example, using the Metropolis-Hasting algorithm, the procedure for generating failure samples can be designed as follows: (1) (2) (3)

Explore the space to find a starting point in the failure region. A “proposal’’ distribution, q(xNew |xCurrent ), is selected to generate a random move to a candidate point xNew . Compute f (xNew ), f (xCurrent ), q(xCurrent |xNew ), and q(xNew |xCurrent ).

382

Structural design optimization considering uncertainties

(4) (5)

Compute g(xNew ) and let f (xNew ) = 0 if g(xNew ) > 0.1 f (xNew ) Accept the new point with a probability of ρ = min f (x · Current )

(6)

Reject the candidate point and save the current point as the “next’’ point with a probability of 1 − ρ.

2

q(xCurrent |xNew ) ,1 q(xNew |xCurrent )

.

Step 4 ensures that candidate points in the safe domain will be rejected with a probability of one; consequently all the generated points will be in the failure domain. Steps 5 and 6 can be executed using a uniform random number generator. The effectiveness of the algorithm depends on the proposal distribution, which defines how to randomly move around in the failure region and the acceptance rate. A simple proposal distribution is a uniform distribution centered at the current point. In this case the proposal distribution is symmetrical and q(xCurrent |xNew )/q(xNew |xCurrent ) = 1; as a result, the ratio of the PDFs of the current and candidate points drives the random movement in a way that ensures that the frequency of visits at any point will be asymptotically proportional to the JPDF in the failure region. The selection of the range of the proposal distribution, which characterizes the step sizes of the random moves, can significantly affect the rejection rate and the convergence rate towards the target distribution, and therefore needs to be tuned. Additionally, a “burn-in’’ period (in which the samples would be thrown away) may be needed to improve the quality of the samples, especially if the number of samples used is relatively small. For higher dimension problems, the use of f (xNew )/f (xCurrent ) could decrease the acceptance rate drastically and slow down the converging process. To address the issue, a modified M-H algorithm has been proposed (Au and Beck, 2001) in which a one-dimensional symmetrical proposal distribution is used in combination with individual ratio f (xi_New )/f (xi_Current ) to allow for random movements in some of the variables and increase the acceptance rate. The M-H algorithm has been used in one of the RBMO example presented below. Note that the M-H algorithm itself does not provide an answer to the pf calculation, but the generated samples can be used with a pf computed from other methods such as importance sampling. 3.3.4 F a il u re-sa m p le b a s e d a n a ly s is f o r s in g l e i ns pe c t i o n Before applying the above methods to general inspection optimization applications, we first analyze a simple but practical case where only one inspection is feasible due to economical and other constraints. The objective of inspection optimization is to find the best inspection time that minimizes pf . Define the inspection time as t1 and the service time t2 . The probability of failure at t1 , prior to inspection, is pf (t1 ), which is the lower bound of pof (t2 ). The defect population includes the stronger parts that would not fail by t2 and the weaker parts, defined here as the “critical parts,’’ that would fail by t2 . The probability of the parts that would survive t1 is pof (t2 ) − pf (t1 ). Given the critical parts, which have a probability of pof (t2 ), there are three failure paths as shown in Figure 14.6, an event diagram. In summary, a defect of size a from the critical parts can (1) fail by t1 with a probability of pf (t1 )/pof , (2) survive t1 , escape inspection with PND(a), and fail by t2 with a probability of [pof − pf (t1 )] · E[PND(a(t1 ))]/pof , and (3) survive t1 , be detected with POD(a), be replaced by a part from the original population, and fail by t2 with a probability of

A r e l i a b i l i t y-b a s e d m a i n t e n a n c e o p t i m i z a t i o n m e t h o d o l o g y

a(t1) > a* (F)

383

- > Fail path 1

pf (t1) a(0) > a*(0)

Reach Insp. Time t1

- > Fail path 2

[pof (t2)

a(t2) > a* (F) Missed Detection

a(t1) < a* a*: Critical defect size a(t): Defect size of critical part at t F: Fail by t2 t1: Time of inspection t2: End of service life

(t2)

[pof pf (t1)]•E[PND(a)]

NDE at t1

Detected/Replaced

(t2 - t1) a (t2-t1) > a * (F )

- > Fail path 3

[pof  pf (t1)]•E[POD(a)]•pf (t2  t1)

Figure 14.6 Probabilities of failure events for one inspection.

[pof − pf (t1 )] · E[POD(a(t1 ))] · pf (t2 − t1 )/pof . Therefore, the total pf with inspection can be summarized in Equation 20. o pW f = pf (t1 ) + [pf − pf (t 1 )] · E[PND(a(t1 ))]

+ [pof

(20)

− pf (t 1 )] · E[POD(a(t1 ))] · pf (t 2 − t 1 )

and pof is the amount of risk reduction, pr , which is: The difference between pW f pr = [pof − pf (t1 )] · E[POD(a(t1 ))] · [1 − pf (t2 − t1 )]

(21)

In Equations 20 and 21, pof is computed from Stage 1 of TIS; the failure samples are used to compute pf (t 1 ) and pf (t 2 − t 1 ), and E[POD(a(t1 ))] can be computed for any t1 using the defect growth history data. Therefore, the Stage 1 failure samples, including defect growth histories, should be saved to calculate risk reduction for any inspection time without additional stress or life analysis. Since pf (t 2 − t 1 ) < pof , the last product term in Equation 21 is 1 − pf (t 2 − t 1 ) > 1 − pof , which is approximately 1 for small pof . This suggests that for small pof , a replacement using an original part can be approximated by assuming a “perfect repair,’’ meaning the part is “fail-proof,’’ and therefore pf (t 2 − t 1 ) = 0. In practice, this condition can be achieved if the flawed parts can be detected in time and can be either mitigated to eliminate the re-occurrence of failure (e.g., a corroded pipe section is wrapped with a corrosion-free composite sleeve) or replaced with new or better-grade parts that are guaranteed to survive the remaining service life. In the undesirable scenario where a bad repair is likely, the impact of the repair can be simulated using a worse-than-new distribution, and a fresh analysis is needed to compute pf (t 2 − t 1 ). If pf (t 2 − t 1 ) can be neglected, risk reduction becomes a product of two timedependent terms, where the first term, [pof − pf (t 1 )], is the risk-reduction potential, a monotonically decreasing function of time, and the second term, E[POD(a)], is a monotonically increasing function. This suggests that in practice, optimal inspection

384

Structural design optimization considering uncertainties

time is neither at time zero (when the risk reduction potential is at the highest but E[POD(a)] is relatively small because of small initial defects), nor at the end of service life (when the risk-reduction opportunity is approaching zero). Equation 21 also implies that better POD(a) will create more risk reduction and the best inspection time is earlier with a better POD(a). An example of single-inspection analysis using an event-tree analysis is shown in Figure 14.7. In this example, the initiating event is the critical parts with pof = 0.01. There are four failure paths. The first two are the same as mentioned above. Given the critical parts, 30% of the parts fail by t1 . At inspection, 10% of the survived critical parts is missed and subsequent failures by t2 cannot be avoided. The last two are the results of two types of repairs, each with a 50% chance. The first is replacement by original part and the second is repair with worse-than-new part. Both the replacement/repair parts only need to survive a time of (t2 − t1 ). The result (see the right-hand side column in Figure 14.7 showing pf contribution) demonstrates that the risk contribution related to pf (t 2 − t 1 ) is insignificant (less than 0.5%) as expected. The use of Equation 21 will be demonstrated further in the RBDT examples described below. 3.3.5 TIS error a n a ly s is s in g le in s p e ct io n A typically small error is inherent in the TIS approach due to ignoring the samples that are originally in the safe region (Wu & Shin 2005). Such an error would arise if an originally safe part is unfortunately repaired to a worse condition. This scenario could be the result of poor workmanship and lack of a quality assurance process. The probability of failure due to ignoring the “safe’’ parts is a product of the probability of safe parts and the conditional probabilities of a sequence of events (detected, repaired if detected, bad repair that causes failure if repaired, and failure before service life) that lead to a failure: p∗f (t2 ) = P(Safe) · POD · P(Repair|Detected) · P(Bad Repair) · pf (t2 − t1 )

(22)

Using the relations P(Safe) ≤ 1(which is close to 1 for high reliability parts) and pf (t2 − t1 ) ≤ pof , and also assuming the worst case that POD = 1 (worse in the sense that more chances are created for bad repairs), the TIS error with respect to pof is dominated by two factors: TIS Error =

p∗f pof

≤ P(Repair|Detected) · P(Bad Repair)

(23)

The first factor, P(Repair|Detection), is expected to be small (at least for high reliability products) because it is unlikely that a large percentage of products will be repaired regardless of the detected defect sizes and knowing that repairs may produce negative (un-safe) results. The second term is also expected to be small assuming a good quality control procedure is in place. In the unlikely worst case scenario, the error is 100% if every defect is detected, every detection leads to a decision to repair, and every repair is a bad repair. In summary, the above error analysis suggests that the TIS error is small if the safe parts are ignored, which is the basis for the high efficiency of TIS compared with standard Monte Carlo sampling.

1%

a(t1) < a*

70%

Reach insp. time t1

30 %

10 %

PND

30%

Detected

Figure 14.7 An event tree analysis example for one inspection.

90%

POD

Inspection at t1

Missed

--> Fail path 1

a*: Critical defect size a(t): Defect size of critical part at t t1: Time of inspection t2: End of service life F: Fail by t2 Replacement: Original parts Repair: Better than before but worse than replacement

Critical parts

a (0) > a*(0)

pf = 1% if no inspection

a(t 1 ) > a* (F)

10 0%

50 %

0%

Repair

50%

Repair/Replacement

Replacement

a(t2) < a*

Reach sevice time t2

a (t2 ) > a* (F)

a( t2 t1 ) < a*

Reach t2

a( t2 t1) > a* (F)

a( t2 t1) < a*

Reach t2

a( t2 t1) > a* (F)

--> Fail path 2

99.75%

0.25%

99.85%

0.15%

7%

Total risk Risk reduced

0.079%

--> Fail path 4

0.047%

--> Fail path 3

0.00371 0.00629

7.88E-06

4.73E-06

100% 63%

0.2%

0.1%

18.9%

80.8%

0.003

0.0007

% Pf

Pf

386

Structural design optimization considering uncertainties

3.3.6 G en era l re p a ir a n d m u lt ip le in s p e ct io ns The above analysis procedure for a single inspection can, in principle, be easily extended for multiple inspections, assuming replacement by original-part (i.e., same defect distribution) or ideal repair. However, when the effect of bad-repair needs to be carefully studied, a full set of MC samples is recommended. When there are multiple inspections, a brute-force MC for inspection optimization would require a set of MC runs for every selected option of inspection schedules. Clearly this is computationally very challenging. Recently, a recursive-probability-integration (RPI) procedure has been developed (Shiao 2006; Shiao & Wu 2004) to more rapidly calculate probability of failure for any number of inspections and types of distributions where the bookkeeping of the risk contributions from every inspection result becomes very tedious. In the RPI approach, the sum of the probabilities of failures from a potentially very large number of failure paths (created by multiple inspections and repairs) is formulated using a condensed formula that involves recursive calculations at every branch (with sub-branches and sub-sub branches, etc.). The formulation provides a systematic way to manage failure paths. Similar to the TIS approach, RPI also uses saved Monte Carlo crack growth histories to compute all the probabilities after each inspection and repair. RPI requires a baseline MC for the original defect distribution and an additional MC for each new repair distribution, the number of which is typically very small. The computational efficiency can be improved further by integrating RPI with the conditional expectation method (CEM), where the random variables are separated into two groups, X 1 and X 2 , and the failure probability is formulated as:

(24) pf = . . . P[g(X 1 , X 2 ) ≤ 0]fX1 fX2 dX 1 dX 2 = E[P[g c (X 2 |x1 ) ≤ 0]] To compute Equation 24, a set of realizations of X 1 is randomly generated and the corresponding P[g c (X2 |x1 ) < 0] values are computed using numerical integration or fast probability integrators. As demonstrated, with proper grouping of X , E[P[g c (X2 |X1 ) ≤ 0]] can be estimated using a relatively small number of samples (Shiao 2006). 3.3.7 S a mpli n g b a s e d r is k s e n s it iv it y a n a ly s i s As a by-product, the failure samples from Stage 1 and Stage 2 can be used to conduct risk sensitivities. The sensitivity of pf with respect to changes in the distribution parameters (mean or standard deviation) θi of a random variable Xi can be evaluated from: 

∂pf > ∂θi σi ∂fx σi ∂fx Sθi = = ··· fx dx = E (25) pf σi pf fx ∂θi pf fx ∂θi !

!

where Sθi are the sensitivity coefficients. Equation 25 leads to the following two non-dimensional sensitivities that can be computed easily using TIS samples. Sµi =

∂p/p = E[ui ]! ∂µui /σui

(26)

A r e l i a b i l i t y-b a s e d m a i n t e n a n c e o p t i m i z a t i o n m e t h o d o l o g y

Sσi =

∂p/p = E[u2i ]! − 1 ∂σui /σui

387

(27)

The expectations in Equations 26 and 27 are over the failure region !; µui is the mean of ui with a nominal value of zero and σui is the standard deviation of ui with a nominal value of one. These two sensitivities have been found to be useful for identifying and ranking important random variables (Karamchandani 1990; Enright & Wu 1999; Wu & Mohanty 2006).

4 Examples 4.1

Rotor disk

Consider a rotor disk subject to fracture failure due to rare manufacturing anomalies such as the alpha defect (Leverant et al. 1997). The potential random variables in such an application include defect size and location, stress, material property, and the time and effectiveness of the inspections. The following numerical example was one of the several test examples developed for TIS methodology development. The example represents the analysis at a highly stressed zone assuming a circular imbedded crack with a specified probability of occurrence. For the entire disk, the risk can be integrated using a zone-based risk integration approach (Wu et al. 2002). In the example, all the units are MKS-based. The analyses were conducted using a computer program written in Matlab language. The fracture mechanics model is: da = C( K)m dN

(28)

√ where m = 3.0, C = 1.021E-11, a = crack radius, K = Yσ πa, σ = 414, and the crack geometry factor is Y = 0.636. Simplified stress and life random variables are used. Stress uncertainty is modeled as σ = X1 · σmodel where X1 is a random variable accounting for the errors in geometry and numerical (such as finite element) modeling. Similarly, a simplified stochastic life model is defined as N = X2 · Nmodel where Nmodel is the life model, and X2 is a life scatter random variable. Both X1 and X2 are modeled using log-normally distributed random variables with a median value of one and a specified coefficient of variation (COV). Assume the defect occurrence rate is 0.00348 per disk and the initial crack size follows a three-parameter Weibull distribution with a location parameter of 0.0028 m, a scale parameter of 0.00043 m, and a shape parameter of = 0.41. The critical crack size for fracture is:

 1 KC 2 (29) ac = π YS where KC = 60. The time of inspection is assumed to be normally distributed with a specified COV. The inspection has a POD(a) of:

 ln (a) − 2.996 POD(a) =
(30) 0.4724

388

Structural design optimization considering uncertainties

1 0.9 0.8

PDF or POD

0.7 PDF(a)/1000 POD(a) POD(a)*PDF(a)/10 Crit. defect size

0.6 0.5 0.4 0.3 0.2 0.1 0 0

0.01

0.02

0.03 0.04 Defect size (m)

0.05

0.06

Figure 14.8 Initial defect PDF and POD, and their product.

The normalized initial-defect PDF and the POD(a) are plotted in Figure 14.8. Also plotted are the nominal critical defect size (diameter) and PDF(a)∗POD(a). The later curve, when integrated, is the percentage of defects that can be detected. Given all the disks with a defect, the percentage is 1.45. Assuming there are 1000 disks, the number of disks that have a defect is 3.48, and the number of defected disks that can be detected at time zero is near zero (3.48*0.0145 = 0.05). Thus, the best inspection time should be after the defect population has grown much bigger. After integration, Equation 28 becomes: 1−m/2

N = X2 ·

1−m/2

− ac ao (m/2 − 1) · C · Y m · (X1 S)m/2 · πm/2

(31)

For zero stress and life scatters, leaving defect as the only random variable, pof can be analyzed by using Equation 27 and setting N = NService . Figure 14.9 compares the pW ’s f using TIS and Monte Carlo and the analytical solution (solid curve). Given a defect, the conditional pof calculated analytically is 0.109. The unconditional pof , plotted in Figure 14.9, is 0.109∗0.00348 = 3.807e-04, which results in an average of 1.90e-08 per flight cycle. With inspection at 10 000 cycles, the unconditional pW is 1.29e-04, f o which is approximately one third of pf or 67% risk reduction. For this example with a relatively high pof (0.109), the TIS (500 samples) is ten times as efficient as Monte Carlo (5000 samples). In general, the efficiency of TIS will be higher with smaller pof . Figure 14.10 shows the increase of risk after adding uncertainties to inspection time, stress scatter, and life scatter. 500 TIS samples proved to be sufficient for the analysis.

A r e l i a b i l i t y-b a s e d m a i n t e n a n c e o p t i m i z a t i o n m e t h o d o l o g y

5

389

104

No inspection (Exact) With inspection (MC 500) With inspection (MC 5000) With inspection (IS 500)

4.5 4

Prob. of failure

3.5 3 2.5 2 1.5 1 0.5 0 0

0.2

0.4

0.6

0.8 1 1.2 Flight cycles

1.4

1.6

1.8

2 104

Figure 14.9 Risk with and without inspection (at 10 000 cycles) fixed stress and life scatter.

5

104 No insp. fixed stress and life With insp. random stress and life

4.5 4

Prob. of failure

3.5 3 2.5 2 1.5 1 0.5 0

0

0.2

0.4

0.6

0.8 1 1.2 Flight cycles

1.4

1.6

1.8

2 104

Figure 14.10 Random inspection, stress, and life scatter (All with COV = 0.1); 500 IS samples.

We will now demonstrate the use of failure samples for inspection optimization. Figure 14.11 shows the failure samples (i.e., for N < 20 000 cycles) in the threedimensional u-space of defect, stress scatter, and life scatter. These samples were created using FPA, a Fast Probability Analyzer software that integrates the Metropolis-Hasting

390

Structural design optimization considering uncertainties

u3 (Life scatter)

3

0

3

3 0

0

u1 (Defect size)

3

u2 (Life scatter)

Figure 14.11 Failure samples in the three-dimensional u-space.

1.2

104

1

Pf reduction

0.8 0.6 0.4 Optimal time ⴝ 13 300 cycles

0.2 0

0

0.2

0.4

0.6

0.8 1 1.2 Inspection time

1.4

1.6

1.8

2 104

Figure 14.12 Risk reduction plot for inspection optimization.

algorithm and importance sampling method. In the example, the selected proposal distribution was a uniform distribution with a range of 1. The samples were generated in the failure region that has a probability of 0.195 with approximately +/−20% error at 90% confidence. Using the failure samples and applying Equation 21, the risk-reduction versus inspection can be computed easily to create figure 14.12. The optimal time of 13 300 cycles is approximate due to the relatively small sample size. Figure 14.13 shows pf versus time for three inspection times: 10 000, 13 300, and 16 000 cycles. Clearly, inspection

A r e l i a b i l i t y-b a s e d m a i n t e n a n c e o p t i m i z a t i o n m e t h o d o l o g y

391

104

5

Insp. at N = 10 000 Insp. at N = 13 300 Insp. at N = 16 000

4.5 4

Prob. of failure

3.5 3 2.5 2 1.5 1 0.5 0

0

0.2

0.4

0.6

0.8 1 1.2 Flight cycles

1.4

1.6

1.8

2 104

Figure 14.13 Probability of failure curves for one inspection at 1000, 13 300, and 16 000 cycles.

1.2 1 0.8

S

0.6 0.4 0.2 0 0.2 0.4

1 Defect

2 Stress scatter

3 Life scatter

Figure 14.14 Probability sensitivities for three random variables.

at 10 000 is too early, at 16 000 is too late, and at 13 300 is significantly better. Using Equation 21, the failure samples were used to compute the mean sensitivities, as displayed in Figure 14.14, which shows that the initial defect size is the most influential random variable, followed by stress scatter and life scatter.

392

Structural design optimization considering uncertainties

6R

t P
140 KN

2R r

P 2R

4R

P

R

Reference R
0.25 m, Thickness
67 mm, Initial flaw size
0.4 mm

Figure 14.15 Spindle lug (Forth et al. 2002).

CC03 P/Wt

W

P S3
— Dt c

P

a

t

D c

Figure 14.16 NASGRO model for the lug example.

4.2 L u g ex a m p le A helicopter spindle lug model is shown in Figure 14.15 (Forth et al. 2002) with its fracture mechanics model (using NASGRO software) shown in Figure 14.16. Figure 14.17 shows a one-hour load spectra, FELIX/28 based on the main rotor blade of a military helicopter with four mission types and 140 flights (Everett et al. 2002). This study was conducted using an RBDT software that integrated a probabilistic function evaluation system (Wu and Shin 2005; Wu et al. 2006), a finite element software (ANSYS), and a fracture mechanics software, NASGRO. The random variables are listed in Table 14.2, where the load random variable represents the point load applied to the center of the pin, P, in Figure 14.16. The initial flaw size distribution, shown in Figure 14.18, is based on an equivalent initial flaw size distribution (Forth et al. 2002), EIFS, derived from stress-life experiments. We will compare the performance of three PODs shown in Figure 14.19 representing poor, fair, and good NDE devices.

A r e l i a b i l i t y-b a s e d m a i n t e n a n c e o p t i m i z a t i o n m e t h o d o l o g y

393

Percent of max. load in spectrum

400 300 200 100 0 Cycles : 2755 Flight hours ⴝ 3.26 hr

100 200

0

500

1000 1500 2000 Number of cycles

2500

3000

Figure 14.17 Felix/28 Helicopter load spectra (Everett et al. 2002). Table 14.2 Random variables for the Lug model.

Thickness, t (mm) Max. load (N) Initial flaw size (mm) Delta Kth Life scatter

Distribution

Mean

Std. Dev.

COV(%)

LN LN User-defined LN LN

28 145 000 0.074 48 1

0.14 10 000 0.0224 4 0.1

0.50 6.9 30.2 8.33 10.0

Equivalent initial flaw size 1 0.8

CDF

0.6 0.4 0.2 0

0

0.05

0.1 0.15 0.2 Flaw size (mm)

0.25

0.3

Figure 14.18 Equivalent initial flaw size distribution.

We will use conservative limit states to illustrate the TIS steps discussed in Section 3.3.3.2. Table 14.3 shows the probabilistic analysis results using three A values: 1, 1.07, and 1.33, which are associated with three target lives. Because the conditional probability of failure given a flaw is relatively high (about 7%), 1000 Monte Carlo samples are sufficient for illustration purposes.

394

Structural design optimization considering uncertainties

Three POD curves (Log-logistic function)

Probability of detection

1 POD I POD II POD III

0.8 0.6

0.4 0.2

0

2

4 6 Defect size (mm)

8

10

Figure 14.19 Three POD curves for the lug example.

Table 14.3 Lug example results.

A 1.00 1.07 1.33 MCS

N = Nf 750 800 1000 750

No. of samp. in samp. region

FORM and samp. Beta

Angle from FORM MPP

No. of failures in samp. region

Prob. in samp. region (Ps)

P f in samp. region (Pc)

Pf

n

Beta 1.495 1.709 1.682 1.587

Deg. 0.0 33.6 22.1 17.3

[N < 750] – 216 222 208

[N < Nf ] – 0.081 0.157 1.000

[N < 750] – 0.864 0.444 0.069

[N < 750] 0.0675 0.0699 0.0696 0.0693

– 250 500 3000

Based on A = 1.067, the probability in the IS (importance sampling) region is 0.081. Of the 250 samples generated in this region, there are 216 failure samples with a conditional pof of 0.864. Therefore pof is 0.0699, which is close to the FORM solution, 0.0675. The agreement, and the fact that the angles between the FORM and the sampling MPPs are reasonably small, suggest that the IS region covers the failure region. Now consider using A = 1.33. The probability in the IS region is 0.151, about twice larger than for A = 1.067. Of the 500 samples generated in this region, there are 222 failure samples with a conditional pof of 0.444, about half smaller than for A = 0.167. The resulting pof is still 0.0699, as before. This means that by doubling the IS region, no additional failure region has been found. This further suggests that the IS region is sufficient, provided that the MPP-based model is reasonably good (which was true as determined using an independent check). The above results are very close to the Monte Carlo result (0.0693) that took 60 hours of CPU time. Note that 500 IS samples are equivalent to 500/0.151 = 3311

A r e l i a b i l i t y-b a s e d m a i n t e n a n c e o p t i m i z a t i o n m e t h o d o l o g y

395

Reducible risk (Pr) given a flaw POD 1 POD II POD III

0.03 0.025

Pr

0.02 0.015 0.01 0.005 0 0

100

200

300 400 500 600 Inspection time (flight hours)

700

800

Figure 14.20 Lug inspection optimization.

Monte Carlo samples, and therefore, for this example, IS provides a slightly better accuracy. However, unlike MCS, IS could miss failure regions. In general, a sufficient number of Monte Carlo samples should be used to ensure that IS has not missed any significant failure region. For complex applications with possible multiple MPPs, more robust error-checking should be considered, including Markov Chain Monte Carlo methods. Using the 222 saved simulated crack growth histories and applying Equation 21, the risk reduction curves for three PODs are obtained as shown in Figure 14.20. The unsmooth curves are due to lack of samples, but are still reasonable for illustration purposes. The results show that POD I is superior and that there is an “effective inspection window,’’ roughly between 300 to 650 hours, with an optimal inspection time at about 550 flight hours. When POD II and III are used, the best inspection time is around 650 hours with a narrower inspection window. These results confirm the earlier suggestions that (1) the best inspection time is earlier for a better POD capability, and (2) a better POD capability will always produce a better optimal risk reduction. Figure 14.21 displays the simulation results of pf versus flight hours with and without inspection for POD I. Clearly the slope of pf changes more drastically at about 550 hours, coinciding with the best inspection time.

5 Conclusions A reliability-based maintenance optimization (RBMO) methodology was presented, with a focus on computational strategies. The RBMO methodology was demonstrated using examples related to aircraft and helicopter structures. The examples suggest that the RBDT methodology is well suited for inspection planning, and appears to be applicable to other structures such as ships, cars, and oil and gas pipelines, to more

396

Structural design optimization considering uncertainties

Pf given a flaw for various PODs 0.08

No inspection POD III at 650 hours POD II at 650 hours POD I at 550 hours

0.07 0.06

Pf

0.05 0.04 0.03 0.02 0.01 0 0

100

200

300

400 500 Flight hours

600

700

800

Figure 14.21 Probability of failure for one inspection at three different times using POD-I.

systematically design reliable and economical structures with associated maintenance programs to sustain structural integrity and reliability. RBMO involves time-dependent damage accumulation models, NDE detections, repairs, replacements and other risk control measures, and an optimal maintenance plan must consider a potentially large number of options including inspection schedules, mitigation options, and selection of NDE devices. In addition, the planning must factor in uncertainties. Given the wide spectrum of the options and the complexities in modeling, the best practical way to conduct RBMO is through the use of random simulations, preferably efficient sampling methods. The TIS approach has been developed to face the challenge. At its core, TIS is a type of Monte Carlo method that uses the power of random simulation. However, drastic efficiency improvement can be achieved by systematically generating samples in the failure domain. When mitigation effects can be reasonably modeled using ideal repairs or replacements with original parts, additional speed improvement can be realized by reusing crack growth histories for various maintenance options. Methods for generating failure-only samples were discussed, including one built on the MPP-based linear surface of a conservative limit state and another based on the Metropolis-Hasting algorithm. It is emphasized that the MPP methods, while widely known and used, are limited to well-behaved functions. For TIS, MPP offers an easy way to generate independent failure samples. M-H, on the other hand, can handle more difficult (non-smooth and nonlinear) functions, but the generated Markov-chain failure samples are correlated and therefore more samples are needed to reach the target distribution. Thus, both methods provide useful tools for RBMO with different strengths and limitations. The disk and lug examples represented the feasibility of the RBMO method to physics-based modeling applications. The software used for the lug example integrated a probabilistic analysis module, a finite element module, and a fracture mechanics module. As an illustration of the analysis CPU time needed for RBMO, in one lug analysis

A r e l i a b i l i t y-b a s e d m a i n t e n a n c e o p t i m i z a t i o n m e t h o d o l o g y

397

using a 2 GHz desktop PC, it took several hours, including time for a FE analysis, to carry out a thousand NSAGRO analyses with the rotorcraft load spectra, which took the most time, and a probabilistic analysis, which took the least time. The CPU time would increase if larger FE models were used or more failure samples were generated. This example suggests that, even with the efficient TIS method, for complex problems involving physics-based models, RBDT can still be very time-consuming unless further model approximations are made. Potential approximation methods for RBMO analysis include kriging (Sacks et al. 1989; Martin & Simpson 2005) and moving least squares (Krishnamurthy 2003) with error-checking procedures.

References Ang, A.H.-S. & Tang, W.H. 1984. Probability Concepts in Engineering Planning and Design, Volume II; Decision, Risk, and Reliability, New York, John Wiley & Sons. Au, S.K. & Beck, J.L. 2001. Estimation of small failure probabilities in high dimensions by subset simulation, Probabilistic Engineering Mechanics, Vol. 16, No. 4, pp. 263–277. Berens, A.P., Hovey, P.W. & Skinn, D.A. 1991. Risk Analysis for Aging Aircraft Fleets, Air Force Wright Lab Report, WL-TR-91-3066, Vol. 1. Bucher, C.G. 1988. Adaptive Sampling – An Iterative Fast Monte Carlo Procedure, Structural Safety, Vol. 5, pp. 119–26. Cunha, S.B., De Souza, A.P.F., Nicolleti, E.S.M. & Aguiar, L.D. 2006. A Risk-Based Inspection Methodology to Optimize Pipeline In-Line Inspection Programs, Journal of Pipeline Integrity, Q3. Ditlevsen, O., Bjerager, Olesen, R. & Hasofer, A.M. 1989. Directional Simulation in Gaussian Space, Probabilistic Engineering Mechanics, Vol. 3, No. 4, pp. 207–217. Ditlevsen, O. & Madsen, H.O. 1996, Structural Reliability Methods. J. Wiley & Sons, New York, 384 pp. Der Kiureghian, A. & Dakessian, T. 1998. Multiple Design Points in First and Second-order Reliability, Structural Safety, Vol. 20, pp. 37–50. Der Kiureghian, A. 2005. First- and Second-Order Reliability Methods, Chapter 14 in Engineering Design Reliability Handbook, E. Nikolaidis, D.M. Ghiocel & S. Singhal, (eds), CRC Press, Boca Raton, FL. Everett, Jr. R.A. 2002. Crack-Growth Characteristics of Fixed and Rotary Wing Aircraft, 6th Joint FAA/DoD/NASA Aging Aircraft Conference. Enright, M.P. & Wu, Y.-T. 1999. Probabilistic Fatigue Life Sensitivity Analysis of Titanium Rotors, Proceedings of the AIAA 41st SDM Conference, Atlanta, GA. Forth, S.C., Everett, Jr. R.A. & Newman, J.A. 2002. A Novel Approach to Rotorcraft Damage Tolerance, 6th Joint FAA/DoD/NASA Aging Aircraft Conference. Gamerman, D. 1997. Markov Chain Monte Carlo, Chapman & Hall. Gentle, J.E. 1998. Random Number Generation and Monte Carlo Methods, Springer-Verlag New York. Harbitz, A. 1986. An Efficient Sampling Method for Probability of Failure Calculation. Structural Safety, Vol. 3, pp. 109–115. Harkness, H.H., Fleming, M., Moran, B. & Belytschko, T. 1994. Fatigue Reliability With In-Service Inspections, FAA/NASA International Symposium on Advanced Structural Integrity Methods for Airframe Durability and Damage Tolerance. Hohenbichler, R. & Rackwitz, R. 1988. Improvement of Second-order Reliability Estimates by Importance Sampling. J. Eng. Mech. ASCE, Vol. 114, No. 12, pp. 2195–2199. Kale, A., Haftka, R.T. & Sankar, B.V. 2007. Efficient Reliability Based Design and Inspection of Stiffened Panels Against Fatigue. Journal of Aircraft.

398

Structural design optimization considering uncertainties

Karamchandani, A. 1990. New Methods in Systems Reliability, Ph.D. dissertation, Stanford University. Karamchandani, A. & Cornell, C.A. 1991. Adaptive Hybrid Conditional Expectation Approaches for Reliability Estimation, Structural Safety, Vol. 11, pp. 59–74. Krishnamurthy, T. 2003. Response Surface Approximation with Augmented and Compactly Supported Radial Basis Functions, Proceedings of the AIAA 44th SDM Conference. Leverant, G.R., Littlefield, D.L., McClung, R.C., Millwater, H.R. & Wu, Y.-T. 1997. A Probabilistic Approach to Aircraft Turbine Rotor Material Design, The International Gas Turbine & Aeroengine Congress & Exhibition, Paper No. 97-GT-22, Orlando, FL. Liu, P.-L. & Der Kiureghian, A. 1986. Multivariate Distribution Models with Prescribed Marginals and Covariances, Probabilistic Engineering Mechanics, Vol. 1, No. 2, pp. 105–112. Martin, J.D. & Simpson, T.W. 2005. Use of Kriging Models to Approximate Deterministic Computer Models, AIAA Journal, Vol. 43, No. 4. Madsen, H.O., Krenk, S. & Lind, N.C. 1986. Methods of Structural Safety, Englewood Cliffs, New Jersey; Prentice Hall. Madsen, H.O., Skjong, R.K., Talin, A.G. & Kirkemo, F. 1987. Probabilistic Fatigue Crack Growth Analysis of Offshore Structures, with Reliability Updating Through Inspection, SNAME, Arlington, VA. Melchers, R.E. 1987. Structural Reliability: Analysis and Prediction, Wiley. Millwater, H.R., Wu, Y.-T., Cardinal, J.W. & Chell, G.G. 1996. Application of Advanced Probabilistic Fracture Mechanics to Life Evaluation of Turbine Rotor Blade Attachments, Journal of Engineering for Gas Turbines and Power, Vol. 118, pp. 394–398. Millwater, H.R., Fitch, S., Riha, D.S., Enright, M.P., Leverant, G.R., McClung, R.C., Kuhlman, C.J., Chell, G.G. & Lee, Y.-D. 2000. A Probabilistically-Based Damage Tolerance Analysis Computer Program for Hard Alpha Anomalies In Titanium Rotors, Proceedings, 45th ASME International Gas Turbine & Aeroengine Technical Congress, Munich, Germany. Nikolaidis, E., Ghiocel, D.M. & Singhal, S. (eds). 2005. Engineering Design Reliability Handbook, CRC Press, Boca Raton, FL. Palmberg, B., Blom, A.F. & Eggwertz. 1987. Probabilistic Damage Tolerance Analysis of Aircraft Structures, In Probabilistic Fracture Mechanics and Reliability, J.W. Provan (ed.). Martinus Nijhoff Publishers. Rackwitz, R. 2001. Reliability Analysis – A Review and Some Perspectives, Structural Safety, Vol. 23, pp. 365–395. Robert, C.P. & Casella, G. 2004. Monte Carlo Statistical Methods. New York: Springer. Rosen Group, www.Roseninspection.net, 2004. Metal Loss Inspection Performance Specifications, Standard_CDP_POFspec_56_rev3.62.doc. Rosenblatt, M. 1952. Remarks on a Multivariate Transformation, The Annals of Mathematical Statistics 23(3), pp. 470–472. Sacks, J., Schiller, S.B. & Welch, W.J. 1989. Design for Computer Experiments, Technometrics, Vol. 31, No. 1. Schuëller, G.I. 1998. Structural Reliability – Recent Advances, Proc. 7th ICOSSAR’97, pp. 3–33. Shiao, M.C. & Wu, Y.-T. 2004. An Efficient Simulation-Based Method for Probabilistic Damage Tolerance Analysis With Maintenance Planning, Proceedings of the ASCE Specialty Conference on Probabilistic Mechanics and Reliability. Shiao, M.C. 2006. Risk-Based Maintenance Optimization, Proceedings of the International Conference on Structural Safety and Reliability. Thoft-Christensen, P. & Murotsu, Y. 1986. Application of Structural Systems Theory, Springer. Volker, A.W.F., Dijkstra, F.H., Heerings, J.H.A.M. & Terpstra, S. 2004. Modeling of NDE Reliability; Development of A POD-Generator, 16th WCNDT 2004 – World Conference on NDT.

A r e l i a b i l i t y-b a s e d m a i n t e n a n c e o p t i m i z a t i o n m e t h o d o l o g y

399

White, P., Barter, S. & Molent, L. 2002. Probabilistic Fracture Prediction Based On Aircraft Specific Fatigue Test Data, 6th Joint FAA/DoD/NASA Aging Aircraft Conference. Wu, Y.-T., Millwater, H.R. & Cruse, T.A. 1990. An Advanced Probabilistic Structural Analysis Method for Implicit Performance Functions, AIAA Journal, Vol. 28, No. 9, pp. 1663–1669. Wu, Y-T., Enright, M.P. & Millwater, H.R. 2002. Probabilistic Methods for Design Assessment of Reliability With Inspection, AIAA Journal, Vol. 40, No. 5, pp. 937–946. Wu, Y.-T. & Shin, Y. 2004. Probabilistic Damage Tolerance Methodology For Reliability Design And Inspection Optimization, Proceedings of the AIAA 45th SDM Conference. Wu, Y.-T., Shiao, M., Shin, Y. & Stroud, W.J. 2005. Reliability-Based Damage Tolerance Methodology for Rotorcraft Structures, Transactions Journal of Materials and Manufacturing. Wu, Y.-T. & Shin, Y. 2005. Probabilistic Function Evaluation System for Maintenance Optimization, Proceedings of the AIAA 46th SDM Conference. Wu, Y.-T., Shin, Y., Sues, R. & Cesare, M. 2006. Probabilistic Function Evaluation System (ProFES) for Reliability-Based Design, Journal of Structural Safety, Vol. 28, Issues 1–2, pp. 164–195. Wu, Y.-T. & Mohanty, S. 2006. Variable Screening and Ranking Using Several Sampling Based Sensitivity Measures, Journal of Reliability Engineering and System Safety, Vol. 91, Issue 6, pp. 634–647.

Chapter 15

Overview of reliability analysis and design capabilities in DAKOTA with application to shape optimization of MEMS Michael S. Eldred Sandia National Laboratories, Albuquerque, NM, USA∗

Barron J. Bichon Vanderbilt University, Nashville, TN, USA

Brian M. Adams Sandia National Laboratories, Albuquerque, NM, USA

Sankaran Mahadevan Vanderbilt University, Nashville, TN, USA

ABSTRACT: Reliability methods are probabilistic algorithms for quantifying the effect of uncertainties in simulation input on response metrics of interest. In particular, they compute approximate response function distribution statistics (such as response mean, variance, and cumulative probability) based on specified probability distributions for input random variables. In this chapter, recent algorithm research in first and second-order local reliability methods is overviewed for both the forward reliability analysis of computing probabilities for specified response levels (the reliability index approach (RIA)) and the inverse reliability analysis of computing response levels for specified probabilities (the performance measure approach (PMA)). A number of algorithmic variations have been explored, and the effect of different limit state approximations, probability integrations, warm starting, most probable point search algorithms, and Hessian approximations is discussed. In addition, global reliability methods are presented for performing reliability analysis in the presence of nonsmooth, multimodal limit state functions. This set of reliability analysis capabilities is then used as the algorithmic foundation for reliability-based design optimization (RBDO) methods, and bi-level and sequential formulations are presented. These RBDO formulations may employ analytic sensitivities of reliability metrics with respect to design variables that either augment or define distribution parameters for the uncertain variables. Relative performance of these reliability analysis and design algorithms is presented for a number of benchmark test problems using the DAKOTA software, and algorithm recommendations are given. These recommended algorithms are subsequently applied to real-world applications in the probabilistic analysis and design of microelectromechanical systems (MEMS), and the calculation of robust and reliable MEMS designs is demonstrated.

∗ Sandia is a multiprogram laboratory operated by Sandia Corporation, a Lockheed Martin Company,

for the United States Department of Energy’s National Nuclear Security Administration under Contract DE-AC04-94AL85000.

402

Structural design optimization considering uncertainties

1 Introduction Uncertainty quantification (UQ) is the process of determining the effect of input uncertainties on response metrics of interest. These input uncertainties may be characterized as either aleatory uncertainties, which are irreducible variabilities inherent in nature, or epistemic uncertainties, which are reducible uncertainties resulting from a lack of knowledge. Since sufficient data is generally available for aleatory uncertainties, probabilistic methods are commonly used for computing response distribution statistics based on input probability distribution specifications. Conversely, for epistemic uncertainties, data is generally sparse, making the use of probability theory questionable and leading to nonprobabilistic methods based on interval specifications. Reliability methods are probabilistic algorithms for quantifying the effect of aleatory input uncertainties on response metrics of interest. In particular, they perform UQ by computing approximate response function distribution statistics based on specified probability distributions for input random variables. These response statistics include response mean, response standard deviation, and cumulative or complementary cumulative distribution function (CDF/CCDF) response level and probability level pairings. These methods are often more efficient at computing statistics in the tails of the response distributions (events with low probability) than sampling-based approaches since the number of samples required to resolve a low probability can be prohibitive. Thus, these methods, as their name implies, are often used in a reliability context for assessing the probability of failure of a system when confronted with an uncertain environment. A number of classical reliability analysis methods are discussed in (Haldar and Mahadevan 2000), including Mean-Value First-Order Second-Moment (MVFOSM), First-Order Reliability Method (FORM), and Second-Order Reliability Method (SORM). More recent methods which seek to improve the efficiency of FORM analysis through limit state approximations include the use of local and multipoint approximations in Advanced Mean Value methods (AMV/AMV+ (Wu, Millwater, and Cruse 1990)) and Two-point Adaptive Nonlinearity Approximation-based methods (TANA (Wang and Grandhi 1994; Xu and Grandhi 1998)), respectively. Each of the FORM-based methods can be employed for “forward’’ or “inverse’’ reliability analysis through the reliability index approach (RIA) or performance measure approach (PMA), respectively, as described in (Tu, Choi, and Park 1999). The capability to assess reliability is broadly useful within a design optimization context, and reliability-based design optimization (RBDO) methods are popular approaches for designing systems while accounting for uncertainty. RBDO approaches may be broadly characterized as bi-level (in which the reliability analysis is nested within the optimization, e.g. (Allen and Maute 2004)), sequential (in which iteration occurs between optimization and reliability analysis, e.g. (Wu, Shin, Sues, and Cesare 2001; Du and Chen 2004)), or unilevel (in which the design and reliability searches are combined into a single optimization, e.g. (Agarwal, Renaud, Lee, and Watson 2004)). Bi-level RBDO methods are simple and general-purpose, but can be computationally demanding. Sequential and unilevel methods seek to reduce computational expense by breaking the nested relationship through the use of iterated or simultaneous approaches, respectively. In order to provide access to a variety of uncertainty quantification capabilities for analysis of large-scale engineering applications on high-performance parallel

O v e r v i e w o f r e l i a b i l i t y a n a l y s i s a n d d e s i g n c a p a b i l i t i e s i n DA K OTA

403

computers, the DAKOTA project (Eldred, Brown, Adams, Dunlavy, Gay, Swiler, Giunta, Hart, Watson, Eddy, Griffin, Hough, Kolda, Martinez-Canales, and Williams 2006) at Sandia National Laboratories has developed a suite of algorithmic capabilities known as DAKOTA/UQ (Wojtkiewicz, Jr., Eldred, Field, Jr., Urbina, and Red-Horse 2001). This package contains the reliability analysis capabilities described in this chapter and provides the foundation for the RBDO approaches. DAKOTA is freely available for download worldwide through an open source license. This chapter overviews recent algorithm research activities that have explored a variety of approaches for performing reliability analysis. In particular, forward and inverse local reliability analyses have been explored using multiple limit state approximation, probability integration, warm starting, Hessian approximation, and optimization algorithm selections. New global reliability analysis methods based on Gaussian process surrogate models have also been explored for handling response functions which may be nonsmooth or multimodal. Finally, these reliability analysis capabilities are used to provide a foundation for exploring bi-level and sequential RBDO formulations. Sections 2 and 3 describe these algorithmic components, Section 4 summarizes computational results for several analytic benchmark test problems, Section 5 presents deployment of these methodologies to the probabilistic analysis and design of MEMS, and Section 6 provides concluding remarks.

2 Reliability method formulations 2.1

Mean Value method

The Mean Value method (MV, also known as MVFOSM in (Haldar and Mahadevan 2000)) is the simplest, least-expensive reliability method because it estimates the response means, response standard deviations, and all CDF/CCDF responseprobability-reliability levels from a single evaluation of response functions and their gradients at the uncertain variable means. This approximation can have acceptable accuracy when the response functions are nearly linear and their distributions are approximately Gaussian, but can have poor accuracy in other situations. The expressions for approximate response mean µg , approximate response standard deviation σg , response target to approximate probability/reliability level mapping (z → p, β), and probability/reliability target to approximate response level mapping (p, β → z) are µg = g(µx ) σg =

 i

j

(1) Cov(i, j)

dg dg (µx ) (µx ) dxi dxj

(2)

βcdf =

µg − z σg

(3)

βccdf =

z − µg σg

(4)

404

Structural design optimization considering uncertainties

z = µg − σg βcdf

(5)

z = µg + σg βccdf

(6)

respectively, where x are the uncertain values in the space of the original uncertain variables (“x-space’’), g(x) is the limit state function (the response function for which probability-response level pairs are needed), and βcdf and βccdf are the CDF and CCDF reliability indices, respectively. With the introduction of second-order limit state information, MVSOSM calculates a second-order mean as µg = g(µx ) +

1  d2g Cov(i, j) (µx ) 2 dxi dxj i

(7)

j

This is commonly combined with a first-order variance (Eq. 2), since second-order variance involves higher order distribution moments (skewness, kurtosis) (Haldar and Mahadevan 2000) which are often unavailable. The first-order CDF probability p(g ≤ z), first-order CCDF probability p(g > z), βcdf , and βccdf are related to one another through p(g ≤ z) =
(−βcdf )

(8)

p(g > z) =
(−βccdf )

(9)

βcdf = −
−1 (p(g ≤ z))

(10)

βccdf = −
−1 (p(g > z))

(11)

βcdf = −βccdf p(g ≤ z) = 1 − p(g > z)

(12) (13)

where
() is the standard normal cumulative distribution function. A common convention in the literature is to define g in such a way that the CDF probability for a response level z of zero (i.e., p(g ≤ 0)) is the response metric of interest. The formulations in this chapter are not restricted to this convention and are designed to support CDF or CCDF mappings for general response, probability, and reliability level sequences. 2.2

L oc al MPP s e ar c h me t ho d s

Other local reliability methods solve a nonlinear optimization problem to compute a most probable point (MPP) and then integrate about this point to compute probabilities. Regardless of specified input probability distributions, the MPP search is performed in uncorrelated standard normal space (“u-space’’) since it simplifies the probability integration: the distance of the MPP from the origin has the meaning of the number of input standard deviations separating the median response from a particular response threshold. The transformation from correlated non-normal distributions (x-space) to uncorrelated standard normal distributions (u-space) is denoted as u = T(x) with the reverse transformation denoted as x = T −1 (u).

O v e r v i e w o f r e l i a b i l i t y a n a l y s i s a n d d e s i g n c a p a b i l i t i e s i n DA K OTA

405

These transformations are nonlinear in general, and possible approaches include the Rosenblatt (Rosenblatt 1952), Nataf (Der Kiureghian and Liu 1986), and Box-Cox (Box and Cox 1964) transformations. The nonlinear transformations may also be linearized, and common approaches for this include the RackwitzFiessler (Rackwitz and Fiessler 1978) two-parameter equivalent normal and the Chen-Lind (Chen and Lind 1983) and Wu-Wirsching (Wu and Wirsching 1987) threeparameter equivalent normals. The results in this chapter employ the Nataf nonlinear transformation which occurs in the following two steps. To transform between the original correlated x-space variables and correlated standard normals (“z-space’’), the CDF matching condition is used:
(zi ) = F(xi )

(14)

where F( ) is the cumulative distribution function of the original probability distribution. Then, to transform between correlated z-space variables and uncorrelated u-space variables, the Cholesky factor L of a modified correlation matrix is used: z = Lu

(15)

where the original correlation matrix for non-normals in x-space has been modified to represent the corresponding correlation in z-space (Der Kiureghian and Liu 1986). The forward reliability analysis algorithm of computing CDF/CCDF probability/reliability levels for specified response levels is called the reliability index approach (RIA), and the inverse reliability analysis algorithm of computing response levels for specified CDF/CCDF probability/reliability levels is called the performance measure approach (PMA) (Tu, Choi, and Park 1999). The differences between the RIA and PMA formulations appear in the objective function and equality constraint formulations used in the MPP searches. For RIA, the MPP search for achieving the specified response level z is formulated as minimize uT u subject to G(u) = z

(16)

and for PMA, the MPP search for achieving the specified reliability/probability level β, p is formulated as minimize subject to

±G(u) 2 uT u = β

(17)

where u is a vector centered at the origin in u-space and g(x) ≡ G(u) by definition. In the RIA case, the optimal MPP solution u∗ defines the reliability index from β = ±u∗ 2 , which in turn defines the CDF/CCDF probabilities (using Eqs. 8–9 in the case of first-order integration). The sign of β is defined by G(u∗ ) > G(0): βcdf < 0, βccdf > 0 G(u∗ ) < G(0): βcdf > 0, βccdf < 0

(18)

406

Structural design optimization considering uncertainties

where G(0) is the median limit state response computed at the origin in u-space (where βcdf = βccdf = 0 and first-order p(g ≤ z) = p(g > z) = 0.5). In the PMA case, the sign applied to G(u) (equivalent to minimizing or maximizing G(u)) is similarly defined by β βcdf < 0, βccdf > 0: maximize G(u) βcdf > 0, βccdf < 0: minimize G(u)

(19)

and the limit state at the MPP (G(u∗ )) defines the desired response level result. When performing PMA with specified p, one must compute β to include in Eq. 17. While this is a straightforward one-time calculation for first-order integrations (Eqs. 10–11), the use of second-order integrations complicates matters since the β corresponding to the prescribed p is a function of the Hessian of G (see Eq. 36), which in turn is a function of location in u-space. The β target must therefore be updated in Eq. 17 as the minimization progresses (e.g., using Newton’s method to solve Eq. 36 for β given p and κi ). This works best when β can be fixed during the course of an approximate optimization, such as for the AMV2 + and TANA methods described in Section 2.2.1. For second-order PMA without limit state approximation cycles (i.e., PMA SORM), the constraint must be continually updated and the constraint derivative should include ∇u β, which would require third-order information for the limit state to compute derivatives of the principal curvatures. This is impractical, so the PMA SORM constraint derivatives are only approximated analytically or estimated numerically. Potentially for this reason, PMA SORM has not been widely explored in the literature.

2.2.1 L i m it sta te a p p r o x im a t io n s There are a variety of algorithmic variations that can be explored within RIA/PMA reliability analysis. First, one may select among several different limit state approximations that can be used to reduce computational expense during the MPP searches. Local, multipoint, and global approximations of the limit state are possible. (Eldred, Agarwal, Perez, Wojtkiewicz, Jr., and Renaud 2007) investigated local first-order limit state approximations, and (Eldred and Bichon 2006) investigated local second-order and multipoint approximations. These techniques include: 1.

a single Taylor series per response/reliability/probability level in x-space centered at the uncertain variable means. The first-order approach is commonly known as the Advanced Mean Value (AMV) method: g(x) ∼ = g(µx ) + ∇x g(µx )T (x − µx )

(20)

and the second-order approach has been named AMV2 : g(x) ∼ = g(µx ) + ∇x g(µx )T (x − µx ) +

1 (x − µx )T ∇x2 g(µx )(x − µx ) 2

(21)

O v e r v i e w o f r e l i a b i l i t y a n a l y s i s a n d d e s i g n c a p a b i l i t i e s i n DA K OTA

2.

407

same as AMV/AMV2 , except that the Taylor series is expanded in u-space. The first-order option has been termed the u-space AMV method: G(u) ∼ = G(µu ) + ∇u G(µu )T (u − µu )

(22)

where µu = T(µx ) and is nonzero in general, and the second-order option has been named the u-space AMV2 method: G(u) ∼ = G(µu ) + ∇u G(µu )T (u − µu ) + 3.

1 (u − µu )T ∇u2 G(µu )(u − µu ) 2

(23)

an initial Taylor series approximation in x-space at the uncertain variable means, with iterative expansion updates at each MPP estimate (x∗ ) until the MPP converges. The first-order option is commonly known as AMV+: g(x) ∼ = g(x∗ ) + ∇x g(x∗ )T (x − x∗ )

(24)

and the second-order option has been named AMV2 +: g(x) ∼ = g(x∗ ) + ∇x g(x∗ )T (x − x∗ ) + 4.

1 (x − x∗ )T ∇x2 g(x∗ )(x − x∗ ) 2

(25)

same as AMV+/AMV2 +, except that the expansions are performed in u-space. The first-order option has been termed the u-space AMV+ method. G(u) ∼ = G(u∗ ) + ∇u G(u∗ )T (u − u∗ )

(26)

and the second-order option has been named the u-space AMV2 + method: G(u) ∼ = G(u∗ ) + ∇u G(u∗ )T (u − u∗ ) + 5.

1 (u − u∗ )T ∇u2 G(u∗ )(u − u∗ ) 2

(27)

a multipoint approximation in x-space. This approach involves a Taylor series approximation in intermediate variables where the powers used for the intermediate variables are selected to match information at the current and previous expansion points. Based on the two-point exponential approximation concept (TPEA, (Fadel, Riley, and Barthelemy 1990)), the two-point adaptive nonlinearity approximation (TANA-3, (Xu and Grandhi 1998)) approximates the limit state as: g(x) ∼ = g(x2 ) +

1−p n n   xi,2 i pi ∂g 1 p p p (x2 ) (xi − xi,2i ) + (x) (xi i − xi,2i )2 ∂xi pi 2 i=1

i=1

(28)

408

Structural design optimization considering uncertainties

where n is the number of uncertain variables and: ?   ∂g (x1 ) xi,1 ∂xi ln pi = 1 + ln ∂g xi,2 (x2 ) ∂x

(29)

i

H n pi pi 2 pi pi 2 (x − x ) i=1 i=1 (xi − xi,2 ) i i,1 +   1−p n  xi,2 i pi ∂g pi (x2 ) (xi,1 − xi,2 ) H = 2 g(x1 ) − g(x2 ) − ∂xi pi

(x) = n

(30)

(31)

i=1

6.

and x2 and x1 are the current and previous MPP estimates in x-space, respectively. Prior to the availability of two MPP estimates, x-space AMV+ is used. a multipoint approximation in u-space. The u-space TANA-3 approximates the limit state as: G(u) ∼ = G(u2 ) +

n  ∂G i=1

∂ui

1−pi

(u2 )

ui,2

pi

 p 1 p (u) (ui i − ui,2i )2 2 n

p

p

(ui i − ui,2i ) +

(32)

i=1

where:  ∂G pi = 1 + ln (u) = n

(u1 ) ∂ui ∂G (u2 ) ∂ui

pi i=1 (ui



?

 ln

H −

p ui,1i )2

+

H = 2 G(u1 ) − G(u2 ) −

ui,1 ui,2

n

7.

p

i=1

(34)

p

(ui i − ui,2i )2

n  ∂G i=1

(33)

∂ui



1−pi

(u2 )

ui,2

pi

p (ui,1i

−

p ui,2i )

(35)

and u2 and u1 are the current and previous MPP estimates in u-space, respectively. Prior to the availability of two MPP estimates, u-space AMV+ is used. the MPP search on the original response functions without the use of any approximations. Combining this option with first-order and second-order integration approaches results in the traditional first-order and second-order reliability methods (FORM and SORM).

The Hessian matrices in AMV2 and AMV2 + may be available analytically, estimated numerically, or approximated through quasi-Newton updates (see Section 2.2.3). The quasi-Newton variant of AMV2 + is conceptually similar to TANA in that both approximate curvature based on a sequence of gradient evaluations. TANA estimates curvature by matching values and gradients at two points and includes it through the use of exponential intermediate variables and a single-valued diagonal Hessian approximation. Quasi-Newton AMV2 + accumulates curvature over a sequence of points and then uses it directly in a second-order series expansion. Therefore, these methods may be expected to exhibit similar performance.

O v e r v i e w o f r e l i a b i l i t y a n a l y s i s a n d d e s i g n c a p a b i l i t i e s i n DA K OTA

409

The selection between x-space or u-space for performing approximations depends on where the approximation will be more accurate, since this will result in more accurate MPP estimates (AMV, AMV2 ) or faster convergence (AMV+, AMV2 +, TANA). Since this relative accuracy depends on the forms of the limit state g(x) and the transformation T(x) and is therefore application dependent in general, DAKOTA/UQ supports both options. A concern with approximation-based iterative search methods (i.e., AMV+, AMV2 + and TANA) is the robustness of their convergence to the MPP. It is possible for the MPP iterates to oscillate or even diverge. DAKOTA/UQ contains checks that monitor for this behavior; however, implementation of a robust model management approach (Giunta and Eldred 2000; Eldred and Dunlavy 2006) is an important area for future work. Another concern with TANA is numerical safeguarding. First, there is the possibility of raising negative xi or ui values to nonintegral pi exponents in Eqs. 30–32, and 34–35. This is particularly likely for u-space. Safeguarding techniques include the use of linear bounds scaling for each xi or ui , offseting negative xi or ui , or promotion of pi to integral values for negative xi or ui . In numerical experimentation, the offset approach has been the most effective in retaining the desired data matches without overly inflating the pi exponents. Second, there are a number of potential numerical difficulties with the logarithm ratios in Eqs. 29 and 33. In this case, a safeguarding strategy is to revert to either the linear (pi = 1) or reciprocal (pi = −1) ∂g ∂G (x1 ) or ∂u approximation based on which approximation has lower error in ∂x (u1 ). i i 2.2.2

Prob a b ilit y i ntegrati ons

The second algorithmic variation involves the integration approach for computing probabilities at the MPP, which can be selected to be first-order (Eqs. 8–9) or second-order integration. Second-order integration involves applying a curvature correction (Breitung 1984; Hohenbichler and Rackwitz 1988; Hong 1999). Breitung applies a correction based on asymptotic analysis (Breitung 1984): p =
(−βp )

n−1  i=1

1  1 + β p κi

(36)

where κi are the principal curvatures of the limit state function (the eigenvalues of an orthonormal transformation of ∇u2 G, taken positive for a convex limit state) and βp ≥ 0 (select CDF or CCDF probability correction to obtain correct sign for βp ). An alternate correction in (Hohenbichler and Rackwitz 1988) is consistent in the asymptotic regime (βp → ∞) but does not collapse to first-order integration for βp = 0: p =
(−βp )

n−1  i=1

1  1 + ψ(−βp )κi

(37)

φ() and φ() is the standard normal density function. (Hong 1999) applies where ψ() =
() further corrections to Eq. 37 based on point concentration methods. To invert a second-order integration and compute βp given p and κi (e.g., for second-order PMA as described in Section 2.2), Newton’s method can be applied as described in (Eldred and Bichon 2006). Additional probability integration approaches

410

Structural design optimization considering uncertainties

can involve importance sampling in the vicinity of the MPP (Hohenbichler and Rackwitz 1988; Wu 1994), but are outside the scope of this chapter. While secondorder integrations could be performed anywhere a limit state Hessian has been computed, the additional computational effort is most warranted for fully converged MPPs from AMV+, AMV2 +, TANA, FORM, and SORM, and is of reduced value for MVFOSM, MVSOSM, AMV, or AMV2 . 2.2.3 Hessi a n a p p r o x im a t io n s To use a second-order Taylor series or a second-order integration when second-order information (∇x2 g, ∇u2 G, and/or κ) is not directly available, one can estimate the missing information using finite differences or approximate it through use of quasiNewton approximations. These procedures will often be needed to make second-order approaches practical for engineering applications. In the finite difference case, numerical Hessians are commonly computed using either first-order forward differences of gradients using ∇ 2 g(x) ∼ =

∇g(x + hei ) − ∇g(x) h

(38)

to estimate the ith Hessian column when gradients are analytically available, or secondorder differences of function values using ∇ 2 g(x) ∼ =

g(x + hei + hej ) − g(x + hei − hej ) − g(x − hei + hej ) + g(x − hei − hej ) 4h2 (39)

to estimate the ijth Hessian term when gradients are not directly available. This approach has the advantage of locally-accurate Hessians for each point of interest (which can lead to quadratic convergence rates in discrete Newton methods), but has the disadvantage that numerically estimating each of the matrix terms can be expensive. Quasi-Newton approximations, on the other hand, do not reevaluate all of the second-order information for every point of interest. Rather, they accumulate approximate curvature information over time using secant updates. Since they utilize the existing gradient evaluations, they do not require any additional function evaluations for evaluating the Hessian terms. The quasi-Newton approximations of interest include the Broyden-Fletcher-Goldfarb-Shanno (BFGS) update Bk+1 = Bk −

Bk sk sTk Bk sTk Bk sk

+

yk ykT ykT sk

(40)

which yields a sequence of symmetric positive definite Hessian approximations, and the Symmetric Rank 1 (SR1) update Bk+1 = Bk +

(yk − Bk sk )(yk − Bk sk )T (yk − Bk sk )T sk

(41)

O v e r v i e w o f r e l i a b i l i t y a n a l y s i s a n d d e s i g n c a p a b i l i t i e s i n DA K OTA

411

which yields a sequence of symmetric, potentially indefinite, Hessian approximations. Bk is the kth approximation to the Hessian ∇ 2 g, sk = xk+1 − xk is the step and yk = ∇gk+1 − ∇gk is the corresponding yield in the gradients. The selection of BFGS versus SR1 involves the importance of retaining positive definiteness in the Hessian approximations; if the procedure does not require it, then the SR1 update can be more accurate if the true Hessian is not positive definite. Initial scalings for B0 and numerical safeguarding techniques (damped BFGS, update skipping) are described in (Eldred and Bichon 2006). 2.2.4 O ptimizat ion al gori thms The next algorithmic variation involves the optimization algorithm selection for solving Eqs. 16 and 17. The Hasofer-Lind Rackwitz-Fissler (HL-RF) algorithm (Haldar and Mahadevan 2000) is a classical approach that has been broadly applied. It is a Newton-based approach lacking line search/trust region globalization, and is generally regarded as computationally efficient but occasionally unreliable. DAKOTA/UQ takes the approach of employing robust, general-purpose optimization algorithms with provable convergence properties. This chapter employs the sequential quadratic programming (SQP) and nonlinear interior-point (NIP) optimization algorithms from the NPSOL (Gill, Murray, Saunders, and Wright 1998) and OPT++ (Meza 1994) libraries, respectively. 2.2.5 Wa rm starti ng of MPP searches The final algorithmic variation for local reliability methods involves the use of warm starting approaches for improving computational efficiency. (Eldred, Agarwal, Perez, Wojtkiewicz, Jr., and Renaud 2007) describes the acceleration of MPP searches through warm starting with approximate iteration increment, with z/p/β level increment, and with design variable increment. Warm started data includes the expansion point and associated response values and the MPP optimizer initial guess. Projections are used when an increment in z/p/β level or design variables occurs. Warm starts were consistently effective in (Eldred, Agarwal, Perez, Wojtkiewicz, Jr., and Renaud 2007), with greater effectiveness for smaller parameter changes, and are used for all computational experiments presented in this chapter.

2.3 G lobal reliability methods Local reliability methods, while computationally efficient, have well-known failure mechanisms. When confronted with a limit state function that is nonsmooth, local gradient-based optimizers may stall due to gradient inaccuracy and fail to converge to an MPP. Moreover, if the limit state is multimodal (multiple MPPs), then a gradientbased local method can, at best, locate only one local MPP solution. Finally, a linear (Eqs. 8–9) or parabolic (Eqs. 36–37) approximation to the limit state at this MPP may fail to adequately capture the contour of a highly nonlinear limit state. For these reasons, efficient global reliability analysis (EGRA) is investigated in (Bichon, Eldred, Swiler, Mahadevan and McFarland 2007).

412

Structural design optimization considering uncertainties

In this approach, ideas from Efficient Global Optimization (EGO) (Jones, Shonlau, and Welch 1998) are adapted for use in reliability analysis. This approach employs a Gaussian process (GP) model to approximate the true response function ˆ G(u) = h(u)T β + Z(u)

(42)

where h( ) is the trend of the model, β is the vector of trend coefficients, and Z() is a stationary Gaussian process with zero mean and covariance defined from a squaredexponential correlation function that describes the departure of the model from its underlying trend. Gaussian process (GP) models are set apart from other surrogate models because they provide not just a predicted value at an unsampled point, but a full Gaussian distribution with an expected value and a predicted variance. This variance gives an indication of the uncertainty in the model, which results from the construction of the covariance function. This function is based on the idea that when input points are near one another, the correlation between their corresponding outputs will be high. As a result, the uncertainty associated with the model’s predictions will be small for input points which are near the points used to train the model, and will increase as one moves further from the training points. In EGO, the mean and variance estimates from the GP are used to form an expected improvement function (EIF), which calculates the expectation that any point in the search space will provide a better solution than the current best solution. An important feature of the EIF is that it provides a balance between exploiting areas of the design space where good solutions have been found, and exploring areas of the design space where the uncertainty is high. To adapt this concept to forward reliability analysis (z → p), an expected feasibility function (EFF) is used to provide an indication of how well the true value of the response is expected to satisfy the equality constraint G(u) = z by integrating over a region in the immediate vicinity of the threshold value z ± :
z+ ˆ ˆ EF(G(u)) = dG (43) [ − |z − G|] G(u) z−

where  is proportional to the standard deviation predicted by the GP at the point u. This integral can be evaluated analytically, as described in (Bichon, Eldred, Swiler, Mahadevan, and McFarland 2007), to create a simple GP-based function to maximize with a global optimization algorithm. Once a new point or points are computed which maximize the EFF, the GP is updated and the process is continued until the maximum EFF value falls below a tolerance. With a converged GP representation of the limit state, multimodal adaptive importance sampling is then applied to the GP to evaluate an approximation to the probabilities of interest.

3 Reliability-based design optimization Reliability-based design optimization (RBDO) methods are used to perform design optimization accounting for reliability metrics. The reliability analysis capabilities described in Section 2 provide a rich foundation for exploring a variety of RBDO formulations. (Eldred, Agarwal, Perez, Wojtkiewics, Jr., and Renaud 2007) investigated bi-level, fully-analytic bi-level, and first-order sequential RBDO approaches employing

O v e r v i e w o f r e l i a b i l i t y a n a l y s i s a n d d e s i g n c a p a b i l i t i e s i n DA K OTA

413

underlying first-order reliability assessments. (Eldred and Bichon 2006) investigated fully-analytic bi-level and second-order sequential RBDO approaches employing underlying second-order reliability assessments. These methods are overviewed in the following sections. 3.1

Bi-level RBDO

The simplest and most direct RBDO approach is the bi-level approach in which a full reliability analysis is performed for every optimization function evaluation. This involves a nesting of two distinct levels of optimization within each other, one at the design level and one at the MPP search level. Since an RBDO problem will typically specify both the z level and the p/β level, one can use either the RIA or the PMA formulation for the UQ portion and then constrain the result in the design optimization portion. In particular, RIA reliability analysis maps z to p/β, so RIA RBDO constrains p/β: minimize f subject to β ≥ β or p ≤ p

(44)

And PMA reliability analysis maps p/β to z, so PMA RBDO constrains z: minimize subject to

f z≥z

(45)

where z ≥ z is used as the RBDO constraint for a cumulative failure probability (failure defined as z ≤ z) but z ≤ z would be used as the RBDO constraint for a complementary cumulative failure probability (failure defined as z ≥ z). Note that many other objective and constraint formulations are possible (see (Eldred, Giunta, Wojtkiewicz Jr., and Trucano 2002) for general mappings which allow flexible use of statistics within multiple objectives, inequality constraints, and equality constraints); formulations with a deterministic objective and a single probabilistic inequality constraint are just convenient examples. An important performance enhancement for bi-level methods is the use of sensitivity analysis to analytically compute the gradients of probability, reliability, and response levels with respect to the design variables. When design variables are separate from the uncertain variables (i.e., they are not distribution parameters), then the following firstorder expressions may be used (Hohenbichler and Rackwitz 1986; Karamchandani and Cornell 1992; Allen and Maute 2004): ∇d z = ∇d g 1 ∇d g  ∇u G  = −φ( − βcdf )∇d βcdf

(46)

∇d βcdf =

(47)

∇d pcdf

(48)

where it is evident from Eqs. 12–13 that ∇d βccdf = −∇d βcdf and ∇d pccdf = −∇d pcdf . In the case of second-order integrations, Eq. 48 must be expanded to include the curvature

414

Structural design optimization considering uncertainties

correction. For Breitung’s correction (Eq. 36), ⎡

⎛

⎢ ∇d pcdf = ⎢ ⎣
(−βp )

n−1 

⎜ −κi ⎜ ⎝ 2(1 + β κ ) 32 p i i=1

⎞ n−1  j=1 j =i

⎤ n−1 

⎟ ⎥ 1 1 ⎟ − φ(−βp ) ⎥ ∇d βcdf   ⎠ 1 + βp κj 1 + βp κi ⎦ i=1

(49) where ∇d κi has been neglected and βp ≥ 0 (see Section 2.2.2). Other approaches assume the curvature correction is nearly independent of the design variables (Rackwitz 2002), which is equivalent to neglecting the first term in Eq. 49. To capture second-order probability estimates within an RIA RBDO formulation using well-behaved β constraints, a generalized reliability index can be introduced where, similar to Eq. 10, ∗ βcdf = −
−1 (pcdf )

(50)

for second-order pcdf . This reliability index is no longer equivalent to the magnitude of u, but rather is a convenience metric for capturing the effect of more accurate probability estimates. Since reliability levels behave more linearly under design variable change than probability levels, replacing a second-order probability constraint with a generalized reliability constraint can improve optimization performance. The corresponding generalized reliability index sensitivity, similar to Eq. 48, is avoid numerical differencing across full reliability analyses, which can be costly or (worse) inaccurate. ∗ =− ∇d βcdf

1 ∇d pcdf ∗ φ(−βcdf )

(51)

where ∇d pcdf is defined from Eq. 49. Even when ∇d g is estimated numerically, Eqs. 46–51 can be used to avoid numerical differencing across full reliability analyses. When the design variables are distribution parameters of the uncertain variables, ∇d g is expanded with the chain rule and Eqs. 46 and 47 become ∇d z = ∇d x∇x g ∇d βcdf =

1 ∇d x∇x g  ∇u G 

(52) (53)

where the design Jacobian of the transformation (∇d x) may be obtained analytically for uncorrelated x or semi-analytically for correlated x (∇d L is evaluated numerically) by differentiating Eqs. 14 and 15 with respect to the distribution parameters. Eqs. 48–51 remain the same as before. For this design variable case, all required information for the sensitivities is available from the MPP search. Since Eqs. 46–53 are derived using the Karush-Kuhn-Tucker optimality conditions for a converged MPP, they are appropriate for RBDO using AMV+, AMV2 +, TANA, FORM, and SORM, but not for RBDO using MVFOSM, MVSOSM, AMV, or AMV2 .

O v e r v i e w o f r e l i a b i l i t y a n a l y s i s a n d d e s i g n c a p a b i l i t i e s i n DA K OTA

3.2

415

Sequential/Surrogate-bas ed RBDO

An alternative RBDO approach is the sequential approach, in which additional efficiency is sought through breaking the nested relationship of the MPP and design searches. The general concept is to iterate between optimization and uncertainty quantification, updating the optimization goals based on the most recent probabilistic assessment results. This update may be based on safety factors (Wu, Shin, Sues, and Cesare 2001) or other approximations (Du and Chen 2004). A particularly effective approach for updating the optimization goals is to use the p/β/z sensitivity analysis of Eqs. 46–53 in combination with local surrogate models (Zou, Mahadevan, and Rebba 2004). In (Eldred, Agarwal, Perez, Wojtkiewicz, Jr., and Renaud 2007) and (Eldred and Bichon 2006), first-order and second-order Taylor series approximations were employed within a trust-region model management framework (Giunta and Eldred 2000; Eldred and Dunlavy 2006) in order to adaptively manage the extent of the approximations and ensure convergence of the RBDO process. Surrogate models were used for both the objective function and the constraints, although the use of constraint surrogates alone is sufficient to remove the nesting. In particular, RIA trust-region surrogate-based RBDO employs surrogate models of f and p/β within a trust region k centered at dc . For first-order local surrogates: minimize f (dc ) + ∇d f (dc )T (d − dc ) subject to β(dc ) + ∇d β(dc )T (d − dc ) ≥ β or p(dc ) + ∇d p(dc )T (d − dc ) ≤ p  d − dc ∞ ≤ k

(54)

and for second-order local surrogates: minimize f (dc ) + ∇d f (dc )T (d − dc ) + 12 (d − dc )T ∇d2 f (dc )(d − dc ) subject to β(dc ) + ∇d β(dc )T (d − dc ) + 12 (d − dc )T ∇d2 β(dc )(d − dc ) ≥ β or p(dc ) + ∇d p(dc )T (d − dc ) + 12 (d − dc )T ∇d2 p(dc )(d − dc ) ≤ p  d − dc ∞ ≤ k

(55)

For PMA trust-region surrogate-based RBDO, surrogate models of f and z are employed within a trust region k centered at dc . For first-order surrogates: minimize f (dc ) + ∇d f (dc )T (d − dc ) subject to z(dc ) + ∇d z(dc )T (d − dc ) ≥ z  d − dc ∞ ≤ k

(56)

and for second-order surrogates: minimize f (dc ) + ∇d f (dc )T (d − dc ) + 12 (d − dc )T ∇d2 f (dc )(d − dc ) subject to z(dc ) + ∇d z(dc )T (d − dc ) + 12 (d − dc )T ∇d2 z(dc )(d − dc ) ≥ z  d − dc ∞ ≤ k

(57)

416

Structural design optimization considering uncertainties

where the sense of the z constraint may vary as described previously. The second-order information in Eqs. 55 and 57 will typically be approximated with quasi-Newton updates. 3.3 Pro b l em f or mulat io n is s ues When performing RBDO in practice, a number of formulation issues arise. In particular, a flexible set of design parameterizations are needed for the input random variables and a rich set of output statistical metrics are needed for the optimization objectives and constraints. 3.3.1 In pu t p a ram e t e r iza t io n As described in Section 3.1, design variables in RBDO may be separate from the uncertain variables or they may define distribution parameters for the random variables. In the latter case, an implementation should first support design variable insertion into any of the native distribution parameters (e.g., mean, standard deviation, lower and upper bounds) for the supported probability distributions. While this supplies sufficient design authority for many distributions (e.g., normal, lognormal, extreme value distributions), other distributions (e.g., uniform, loguniform, triangular) do not directly support location and scale control within the native parameters. In these cases, location and scale are derived quantities and the native distribution parameters may be insufficient for design purposes, depending on the application. For example, the distribution parameters for a triangular distribution are lower bound, mode, and upper bound. Design control of any one of these three parameters independent of the other two is useful in some applications, but it will be insufficient to arbitrarily translate or scale the distribution in other applications. To provide additional design control in these cases, supporting the ability to design derived distribution parameters (from which the native parameters are updated) is an important extension. When gross distribution control (location, scale) and fine distribution control (native parameters) can both be provided, a broad range of design scenarios can be supported. 3.3.2 Ou tpu t m e t r ics Similar to the input parameterization, output metric characterization requires careful thought when developing optimization under uncertainty capabilities. In particular, a rich, expressive set of metrics is needed for arbitrary control of the shape of output distributions. Generally speaking, designing for robustness involves the control of moments; for example, minimizing an output variance statistic. On the other hand, design for reliability requires the control of tail statistics; for example, constraining a probability of failure statistic. Reliability methods are better suited for computing tail statistics, as MPP methods do not directly calculate moments. To control output variance, a PMAbased response interval, e.g. |zβ=3 − zβ = −3 |, may be substituted for the variance in order to achieve similar goals. To control both robustness and reliability, a multiobjective formulation can provide for general trade-off analysis. If, however, a particular reliability goal is known (e.g., β > 3), then formulations such as the two shown in Section 5 can be effective in reducing output variance while achieving prescribed reliability.

O v e r v i e w o f r e l i a b i l i t y a n a l y s i s a n d d e s i g n c a p a b i l i t i e s i n DA K OTA

417

Finally, model calibration under uncertainty studies typically involve the estimation of random variable distribution parameters which result in the best match in statistics between simulation and experiment. When using reliability methods, a convenient formulation is a nonlinear least squares objective which sums the discrepancies in CDF values (e.g., probabilities in an RIA RBDO formulation or response levels in a PMA RBDO formulation). In this case, all of the same analytic machinery applies (i.e., the sensitivity analysis of Section 3.1), only a broader set of distribution parameters may be of interest and a more complete set of CDF points may be required.

4 Benchmark problems (Eldred, Agarwal, Perez, Wojtkiewicz, Jr., and Renaud 2007) and (Eldred and Bichon 2006) have examined the performance of first and second-order local reliability analysis and design methods for four analytic benchmark test problems: lognormal ratio, short column, cantilever beam, and steel column. (Bichon, Eldred, Swiler, Mahadevan, and McFarland, 2007) has examined the performance of global reliability analysis methods for two additional analytic benchmark test problems that cause problems for local methods. 4.1

Local reliability analys is res ults

Within the reliability analysis algorithms, various limit state approximation (MVFOSM, MVSOSM, x-/u-space AMV, x-/u-space AMV2 , x-/u-space AMV+, x-/u-space AMV2 +, x-/u-space TANA, FORM, and SORM), probability integration (first-order or second-order), warm starting, Hessian approximation (finite difference, BFGS, or SR1), and MPP optimization algorithm (SQP or NIP) selections have been investigated. A sample comparison of reliability analysis performance, taken from the short column example, is shown in Tables 15.1 and 15.2 for RIA and PMA analysis, respectively, where “*’’ indicates that one or more levels failed to converge. Consistent with the employed probability integrations, the error norms are measured with respect Table 15.1 RIA results for short column problem. RIA Approach

SQP Function evaluations

NIP Function evaluations

CDF p Error norm

Target z Offset norm

MVFOSM MVSOSM x-space AMV u-space AMV x-space AMV2 u-space AMV2 x-space AMV+ u-space AMV+ x-space AMV2 + u-space AMV2 + x-space TANA u-space TANA FORM SORM

1 1 45 45 45 45 192 207 125 122 245 296* 626 669

1 1 45 45 45 45 192 207 131 130 246 278* 176 219

0.1548 0.1127 0.009275 0.006408 0.002063 0.001410 0.0 0.0 0.0 0.0 0.0 6.982e-5 0.0 0.0

0.0 0.0 18.28 18.81 2.482 2.031 0.0 0.0 0.0 0.0 0.0 0.08014 0.0 0.0

418

Structural design optimization considering uncertainties

Table 15.2 PMA results for short column problem. PMA Approach

SQP Function evaluations

NIP Function evaluations

CDF z Error norm

Target p Offset norm

MVFOSM MVSOSM x-space AMV u-space AMV x-space AMV2 u-space AMV2 x-space AMV+ u-space AMV+ x-space AMV2 + u-space AMV2 + x-space TANA u-space TANA FORM SORM

1 1 45 45 45 45 171 205 135 132 293* 325* 720 535

1 1 45 45 45 45 179 205 142 139 272 311* 192 191*

7.454 6.823 0.9420 0.5828 2.730 2.828 0.0 0.0 0.0 0.0 0.04259 2.208 0.0 2.410

0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.598e-4 5.600e-4 0.0 6.522e-4

to fully-converged first-order results for MV, AMV, AMV2 , AMV+, and FORM methods, and with respect to fully-converged second-order results for AMV2 +, TANA, and SORM methods. Also, it is important to note that the simple metric of “function evaluations’’ is imperfect, and (Eldred and Bichon 2006) provides more detailed reporting of individual response value, gradient, and Hessian evaluations. Overall, reliability analysis results for the lognormal ratio, short column, and cantilever test problems indicate several trends. MVFOSM, MVSOSM, AMV, and AMV2 are significantly less expensive than the fully-converged MPP methods, but come with corresponding reductions in accuracy. In combination, these methods provide a useful spectrum of accuracy and expense that allow the computational effort to be balanced with the statistical precision required for particular applications. In addition, support for forward and inverse mappings (RIA and PMA) provide the flexibility to support different UQ analysis needs. Relative to FORM and SORM, AMV+ and AMV2 + has been shown to have equal accuracy and consistent computational savings. For second-order PMA analysis with prescribed probability levels, AMV2 + has additionally been shown to be more robust due to its ability to better manage β udpates. Analytic Hessians were highly effective in AMV2 +, but since they are often unavailable in practical applications, finite-difference numerical Hessians and quasi-Newton Hessian approximations were also demonstrated, with SR1 quasi-Newton updates being shown to be sufficiently accurate and competitive with analytic Hessian performance. Relative to first-order AMV+ performance, AMV2 + with analytic Hessians had consistently superior efficiency, and AMV2 + with quasi-Newton Hessians had improved performance in most cases (it was more expensive than first-order AMV+ only when a more challenging second-order p problem was being solved). In general, second-order reliability analyses appear to serve multiple synergistic needs. The same Hessian information that allows for more accurate probability integrations can also be applied to making MPP

O v e r v i e w o f r e l i a b i l i t y a n a l y s i s a n d d e s i g n c a p a b i l i t i e s i n DA K OTA

419

solutions more efficient and more robust. Conversely, limit state curvature information accumulated during an MPP search can be reused to improve the accuracy of probability estimates. For nonapproximated limit states (FORM and SORM), NIP optimizers have shown promise in being less susceptible to PMA u-space excursions and in being more efficient than SQP optimizers in most cases. Warm starting with projections has been shown to be consistently effective for reliability analyses, with typical savings on the order of 25%. The x-space and u-space approximations for AMV, AMV2 , AMV+, AMV2 +, and TANA were both effective, and the relative performance was strongly problemdependent (u-space was more efficient for lognormal ratio, x-space was more efficient for short column, and x-space and u-space were equivalent for cantilever). Among all combinations tested, AMV2 + (with analytic Hessians if available, or SR1 Hessians if not) is the recommended approach. An important question is how Taylor-series based limit state approximations (such as AMV+ and AMV2 +) can frequently outperform the best general-purpose optimizers (such as SQP and NIP) which may employ similar internal approximations. The answer likely lies in the exploitation of the structure of the RIA and PMA MPP problems. By approximating the limit state but retaining uT u explicitly in Eqs. 16 and 17, specific problem structure knowledge is utilized in formulating a mixed surrogate/direct approach. 4.2

G lobal reliability analys is res ults

Our test problem for demonstrating global reliability analysis is taken from (Bichon, Eldred, Swiler, Mahadevan, and McFarland 2007). It has a highly nonlinear response defined by: g(x) =

(x21 + 4)(x2 − 1) 5x1 − sin −2 20 2

(58)

The distribution of x1 is Normal(1.5, 1), x2 is Normal(2.5, 1), and the variables are uncorrelated. The response level of interest for this study is z = 0 with failure defined by g > z. Figure 15.1(a) shows a plot in u-space of the limit state throughout the ±5 standard deviation search space. This problem has several local optima to the forwardreliability MPP search problem (see Eq. 16). Figure 15.1(b) shows an example of an EGRA execution, with the total set of truth model evaluations performed from building the initial surrogate model and then repeatedly maximizing the expected feasibility function derived from the GP model. It is evident that the algorithm preferentially selects the data points needed to accurately resolve the limit state contour of interest. This multimodal problem was also solved using a number of local reliability methods for comparison purposes. Two approximation-based methods (AMV2 + and TANA) were investigated in x-space and u-space as well as the no approximation case (FORM/SORM). To produce results consistent with an implicit response function, numerical gradients and quasi-Newton Hessians from Symmetric Rank 1 updates were used. For each method, at the converged MPP, both first-order and second-order integration (using Eqs. 9 and 37) were used to calculate the probability.

420

Structural design optimization considering uncertainties

5 4 3 2 1 0 1 2 3 4 5 5 4 3 2 1

0

1

2

3

4

(a) Contour of the true limit state function

5

5 4 3 2 1 0 1 2 3 4 5 5 4 3 2 1 0 1 2 3 (b) Gaussian process approximation with data points generated by EGRA

4

5

Figure 15.1 Multimodal test problem.

Table 15.3 Results for the multimodal test problem. Reliability method

Function evaluations

First-order pf (% Error)

Second-order pf (% Error)

Sampling pf (% Error, Avg. Error)

No Approximation x-space AMV2 + u-space AMV2 + x-space TANA u-space TANA x-space EGRA u-space EGRA True LHS solution

66 26 26 506 131 50.4 49.4 1M

0.11798 (276.3%) 0.11798 (276.3%) 0.11798 (276.3%) 0.08642 (175.7%) 0.11798 (276.3%) — — —

0.02516 (−19.7%) 0.02516 (−19.7%) 0.02516 (−19.7%) 0.08716 (178.0%) 0.02516 (−19.7%) — — —

— — — — — 0.03127 (0.233%, 0.929%) 0.03136 (0.033%, 0.787%) 0.03135 (0.000%, 0.328%)

Table 15.3 gives a summary of the results from all methods. To establish an accurate estimate of the true solution, 20 independent studies were performed using 106 Latin hypercube samples per study. The average probability from these studies is reported as the “true’’ solution. Because the EGRA method is stochastic, it was also run 20 times and the average probability is reported. To measure the accuracy of the methods, two errors are reported for the EGRA results: the error in the average probability, and the average of the absolute errors from the 20 studies. For comparison, the same errors are given for the 20 LHS studies. Most of the MPP search methods converge to the same MPP (in the vicinity of (0.5, 1) in u-space) and thus report the same probability. These probabilities are more accurate when second-order integration is used, but still have significant errors. However, x-space TANA converges to a secondary MPP, which lies in a relatively flat region of the limit state (in the vicinity of (2, 1) in u-space). This local lack of curvature means that first-order and second-order integration produce approximately the same probability. In isolation, this second-order result could be viewed as a verification of the first-order probability and thus provide a misguided confidence in the local reliability

O v e r v i e w o f r e l i a b i l i t y a n a l y s i s a n d d e s i g n c a p a b i l i t i e s i n DA K OTA

421

Table 15.4 Analytic bi-level RBDO results, short column test problem. RBDO Approach

Function evaluations

Objective function

Constraint violation

RIA z → p x-space AMV+ RIA z → p x-space AMV2 + RIA z → p FORM RIA z → p SORM RIA z → β x-space AMV+ RIA z → β x-space AMV2 + RIA z → β FORM RIA z → β SORM PMA p, β → z x-space AMV+ PMA p → z x-space AMV2 + PMA β → z x-space AMV2 + PMA p, β → z FORM PMA p → z SORM PMA β → z SORM

149 129 911 1204 72 67 612 601 100 98 98 285 306 329

217.1 217.1 217.1 217.1 216.7 216.7 216.7 216.7 216.8 216.8 216.8 216.8 217.2 216.8

0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

analysis. For this problem, the new EGRA method is more expensive than AMV2 +, but cheaper than all the other methods, and provides much more accurate results. Thus, global reliability analysis can provide accuracy similar to that of exhaustive sampling with expense comparable to local reliability. It handles both multimodal and nonsmooth limit states and does not require any derivative information from the response function. The primary limitation of the technique is dimensionality. For larger scale uncertainty quantification problems, the expense of building a global approximation grows quickly with dimension, although this can be mitigated to some extent by requiring accuracy only along a single contour and only in the highest probability regions. 4.3 RBDO results These reliability analysis capabilities provide a substantial foundation for RBDO formulations, and bi-level and sequential RBDO approaches based on local reliability analyses have been investigated. Both approaches have utilized analytic gradients for z, β, and p with respect to augmented and inserted design variables, and sequential RBDO has additionally utilized a trust-region surrogate-based approach to manage the extent of the Taylor-series approximations. A sample comparison of RBDO performance, taken again from the short column example, is shown in Tables 15.4 and 15.5 for bi-level and sequential surogate-based RBDO, respectively. Overall, RBDO results for the short column, cantilever, and steel column test problems build on the reliability analysis trends. Basic first-order bi-level RBDO has been evaluated with up to 18 variants (RIA/PMA with different p/β/z mappings for MV, x-/u-space AMV, x-/u-space AMV+, and FORM), and fully-analytic bi-level and sequential RBDO have each been evaluated with up to 21 variants (RIA/PMA with different p/β/z mappings for x-/u-space AMV+, x-/u-space AMV2 +, FORM, and SORM). Bi-level RBDO with MV and AMV are inexpensive but give only approximate

422

Structural design optimization considering uncertainties

Table 15.5 Surrogate-based RBDO results, short column test problem. RBDO Approach

Function evaluations

Objective function

Constraint violation

RIA z → p x-space AMV+ RIA z → p x-space AMV2 + RIA z → p FORM RIA z → p SORM RIA z → β x-space AMV+ RIA z → β x-space AMV2 + RIA z → β FORM RIA z → β SORM PMA p, β → z x-space AMV+ PMA p → z x-space AMV2 + PMA β → z x-space AMV2 + PMA p, β → z FORM PMA p → z SORM PMA β → z SORM

75 86 577 718 65 51 561 560 76 58 79 228 128 171

216.9 218.7 216.9 216.5 216.7 216.7 216.7 216.7 216.7 216.8 216.8 216.7 217.2 216.8

0.0 0.0 0.0 1.110e-4 0.0 0.0 0.0 0.0 2.1e-4 0.0 0.0 2.1e-4 0.0 0.0

optima. These approaches may be useful for preliminary design or for warm-starting other RBDO methods. Bi-level RBDO with AMV+ was shown to have equal accuracy and robustness to bi-level FORM-based approaches and be significantly less expensive on average. In addition, usage of β in RIA RBDO constraints was preferred due to it being more well-behaved and more well-scaled than constraints on p. Warm starts in RBDO were most effective when the design changes were small, with the most benefit for basic bi-level RBDO (with numerical differencing at the design level), decreasing to marginal effectiveness for fully-analytic bi-level RBDO and to relative ineffectiveness for sequential RBDO. However, large design changes were desirable for overall RBDO efficiency and, compared to basic bi-level RBDO, fully-analytic RBDO and sequential RBDO were clearly superior. In second-order bi-level and sequential RBDO, the AMV2 + approaches were consistently more efficient than the SORM-based approaches. In general, sequential RBDO approaches demonstrated consistent computational savings over the corresponding bi-level RBDO approaches, and the combination of sequential RBDO using AMV2 + was the most effective of all of the approaches. With initial trust region size tuning, sequential RBDO computational expense for these test problems was shown to be as low as approximately 40 function evaluations per limit state (35 for a single limit state in short column, 75 for two limit states in cantilever, and 45 for a single limit state in steel column). At this level of expense, probabilistic design studies can become tractable for expensive engineering applications.

5 Application to MEMS In this section, we consider the application of DAKOTA’s reliability algorithms to the design of micro-electro-mechanical systems (MEMS). In particular, we summarize results for the MEMS application described in (Adams, Eldred, and Wittwer 2006).

O v e r v i e w o f r e l i a b i l i t y a n a l y s i s a n d d e s i g n c a p a b i l i t i e s i n DA K OTA

423

Force

Actuation force

Switch contact Fmax

Anchors

Shuttle E2

E3

E1 vernier Fmin

(a) Scanning electron micrograph of a MEMS bistable mechanism in its second stable position. The attached vernier provides position measurements.

Displacement

(b) Schematic of force–displacement curve for bistable MEMS mechanism. The arrows indicate stability of equilibria E1 and E3 and instability of E2.

Figure 15.2 Bi-stable MEMS mechanism.

These types of application studies provide essential feedback on the performance of algorithms for real-world design applications, which may contain computational challenges not well-represented in analytically defined test problems. The reliability analysis and design results in (Adams, Eldred, and Wittwer 2006) are extended to include parameter-adaptive solution verification through the use of finite element a posteriori error estimation in (Adams, Bichon, Eldred, Carnes, Copps, Neckels, Hopkins, Notz, Subia, and Wittwer 2006; Eldred, Adams, Copps, Carnes, Notz, Hopkins, and Wittwer 2007). Pre-fabrication design optimization of microelectromechanical systems (MEMS) is an important emerging application of uncertainty quantification and reliability-based design optimization. Typically crafted of silicon, polymers, metals, or a combination thereof, MEMS serve as micro-scale sensors, actuators, switches, and machines with applications including robotics, biology and medicine, automobiles, RF electronics, and optical displays (Allen 2005). Design optimization of these devices is crucial due to high cost and long fabrication timelines. Uncertainty in the micromachining and etching processes used to manufacture MEMS can lead to large uncertainty in the behavior of the finished products, resulting in low part yield and poor durability. RBDO, coupled with computational mechanics models of MEMS, offers a means to quantify this uncertainty and determine a priori the most reliable and robust designs that meet performance criteria. Of particular interest is the design of MEMS bistable mechanisms which toggle between two stable positions, making them useful as micro switches, relays, and nonvolatile memory. We focus on shape optimization of compliant bistable mechanisms, where instead of mechanical joints, material elasticity enables the bistability of the mechanism (Kemeny, Howell, and Magleby 2002; Ananthasuresh, Kota, and Gianchandani 1994; Jensen, Parkinson, Kurabayashi, Howell, and Baker 2001). Figure 15.2(a) contains an electron micrograph of a MEMS compliant bistable mechanism in its second stable position. The first stable position is the as-fabricated position. One achieves transfer between stable states by applying force to the center shuttle via a thermal actuator, electrostatic actuator, or other means to move the shuttle past an unstable equilibrium.

424

Structural design optimization considering uncertainties

Tapered beam

Anchor Shuttle

Actuation force

(a) Schematic of a tapered beam bistable mechanism in as fabricated position (not to scale).

(b) Scale rendering of tapered beam leg for bistable mechanism.

Figure 15.3 Tapered beams for bistable MEMS mechanism.

Bistable switch actuation characteristics depend on the relationship between actuation force and shuttle displacement for the manufactured switch. Figure 15.2(b) contains a schematic of a typical force–displacement curve for a bistable mechanism. The switch characterized by this curve has three equilibria: E1 and E3 are stable equilibria whereas E2 is an unstable equilibrium (arrows indicate stability). A device with such a force–displacement curve could be used as a switch or actuator by setting the shuttle to position E3 as shown in Figure 15.2(a) (requiring large actuator force Fmax ) and then actuating by applying the comparably small force Fmin in the opposite direction to transfer back through E2 toward the equilibrium E1 . One could utilize this force profile to complete a circuit by placing a switch contact near the displaced position corresponding to maximum (closure) force as illustrated. Repeated actuation of the switch relies on being able to reset it with actuation force Fmax . The device design considered in this chapter is similar to that in the electron micrograph in Figure 15.2(a), for which design optimization has been previously considered (Jensen, Parkinson, Kurabayashi, Howell, and Baker 2001), as has robust design under uncertainty with mean value methods (Wittwer, Baker, and Howell 2006). The primary structural difference in the present design is the tapering of the legs, shown schematically in Figure 15.3(a). Figure 15.3(b) shows a scale drawing of one tapered beam leg (one quarter of the full switch system). A single leg of the device is approximately 100 µm wide and 5–10 µm tall. This topology is a cross between the fully compliant bistable mechanism reported in (Jensen, Parkinson, Kurabayashi, Howell, and Baker 2001) and the thickness-modulated curved beam in (Qiu and Slocum 2004). As described in the optimization problem below, this tapered geometry offers many degrees of freedom for design. The tapered beam legs of the bistable MEMS mechanism are parameterized by the 13 design variables shown in Figure 15.4, including widths and lengths of beam segments as well as angles between segments. For simulation, a symmetry boundary condition allowing only displacement in the negative y direction is applied to the right surface (x = 0) and a fixed displacement condition is applied to the left surface. With appropriate scaling, this allows the quarter model to reasonably represent the full four-leg switch system.

O v e r v i e w o f r e l i a b i l i t y a n a l y s i s a n d d e s i g n c a p a b i l i t i e s i n DA K OTA

425

1 0.5

W4 W3

0 0.5

W2

1 y (m)

<

4

3

1.5 2 2.5

2

1

3.5 4 4.5

W1

W0

3

L1 100

80

L2 60

L3 40 x (m)

L4 20

0

Figure 15.4 Design parameters for the tapered-beam fully-compliant bistable mechanism (geometry not to scale). Displacement is applied in the negative y direction at the right face (x = 0), while at the left face, a fixed displacement condition is enforced.

Table 15.6 Uncertain variables x = [ W, Sr ] used in reliability analysis. Variable

Mean (µ)

Std. dev.

Distribution

W (width bias) Sr (residual stress)

−0.2 µm −11 Mpa

0.08 4.13

Normal Normal

Due to manufacturing processes, fabricated geometry can deviate significantly from design-specified beam geometry. As a consequence of photo lithography and etching processes, fabricated in-plane geometry edges (contributing to widths and lengths) can be 0.1 ± 0.08 µm less than specified. This uncertainty in the manufactured geometry leads to substantial uncertainty in the positions of the stable equilibria and in the maximum and minimum force on the force–displacement curve. The manufactured thickness of the device is also uncertain, though this does not contribute as much to variability in the force–displacement behavior. Uncertain material properties such as Young’s modulus and residual stress also influence the characteristics of the fabricated beam. For this application two key uncertain variables are considered: W (edge bias on beam widths, which yields effective manufactured widths of Wi + W, i = 0, . . . , 4) and Sr (residual stress in the manufactured device), with distributions shown in Table 15.6. Given the 13 geometric design variables d = [L1 , L2 , L3 , L4 , θ1 , θ2 , θ3 , θ4 , W0 , W1 , W2 , W3 , W4 ] and the specified uncertain variables x = [ W, Sr ] we formulate a

426

Structural design optimization considering uncertainties

m

5.0

Fmin

b
2 (a) Response PDF control of mean and right tail

zb
2 m

5.0

Fmin

b
2 (b) Response PDF control of both tails

Figure 15.5 Schematic representation of design formulations for output response PDF control.

reliability-based design optimization problem to achieve a design that actuates reliably with at least 5 µN force. The RBDO formulation uses the limit state g(x) = Fmin (x)

(59)

and failure is defined to be actuation force with magnitude less than 5.0 µN (Fmin > −5.0). Reliability index βccdf ≥ 2 is required. The RBDO problem utilizes the RIA z → β approach (Eq. 44) with z = −5.0: max subject to

2 ≤ 50 ≤

A @ E Fmin (d, x) @ βccdf (d) A E F @ max (d, x)A ≤ 150 E@ E2 (d, x) A ≤ 8 E Smax (d, x) ≤ 3000

(60)

although the PMA β → z approach (Eq. 45) could also be used. The use of the Fmin metric in both the objective function and the reliability constraint results in a powerful problem formulation, because in addition to yielding a design with specified reliability, it also produces a robust design. By forcing the expected value of Fmin toward the −5.0 target while requiring two input standard deviations of surety, the optimization problem favors designs with less variability in Fmin . This renders the design performance less sensitive to uncertainties. The response PDF control is depicted in Figure 15.5(a), where the mean is maximized subject to a reliability constraint on the right tail. Alternatively, the response PDF control depicted in Figure 15.5(b) could be employed by maximizing the PMA z level corresponding to β = −2. This has the advantage of controlling both sides of the response PDF, but it is more computationally expensive since it requires the solution of two MPP optimization problems per design cycle instead of one. For this reason, the RIA RBDO formulation in Eq. 60 is used for all results in this section. Results using the MVFOSM, AMV2 +, and FORM methods are presented in Table 15.7 and the optimal force–displacement curves are shown in Figure 15.6. Optimization with MVFOSM reliability analysis offers substantial improvement over the initial design, yielding a design with actuation force Fmin nearer the −5.0 target and

O v e r v i e w o f r e l i a b i l i t y a n a l y s i s a n d d e s i g n c a p a b i l i t i e s i n DA K OTA

427

Table 15.7 Reliability formulation RBDO: design variable bounds and optimal designs from MVFOSM, AMV2 +, and FORM methods for MEMS bistable mechanism. Variable/metric l.b.

Name

AMV 2 + Optimal

FORM Optimal

−26.29 5.376 68.69 4.010 470 3.771 3.771

−5.896 2.000 50.01 5.804 1563 1.804 1.707

−6.188 1.998 57.67 5.990 1333 – 1.784

−6.292 1.999 57.33 6.008 1329 – –

150 8 1200

70

2

60

2.5

50

3

30

AMV2 Target force

4

4.5

20

5

10

5.5

0

6

10

MVFOSM

3.5

40

Force (µN)

Force (µN)

MVFOSM Optimal

u.b.

E[F min ] (µN) β E[F max ] (µN) E[E2 ] (µm) E[Smax ] (MPa) AMV2 + verified β FORM verified β

2 50

Initial

0

2

4 6 8 Displacement (µm)

10

6.5

6

6.5

7 7.5 Displacement (µm)

8

Figure 15.6 Optimal force–displacement curves resulting from RBDO of MEMS bistable mechanism with mean value and AMV2 + methods. The right plot shows the area near the minimum force. Two input standard deviations (as measured by the method used during optimization) separate F min from the target Fmin = −5.0.

tight reliability constraint β = 2. However, since mean value analyses estimate reliability based solely on evaluations at the means of the uncertain variables, they can yield inaccurate reliability metrics in cases of nonlinearity or nonnormality. In this example, the actual reliability (verified with MPP-based methods) of the optimal MVFOSMbased design is only 1.804 (AMV2 +) or 1.707 (FORM); both less than the prescribed reliability β ≥ 2. In this example, the additional computational expense incurred when using MPP-based reliability methods appears to be justified. Reliability-based design optimization with either the AMV2 + or FORM methods for reliability analysis yield constraint-respecting optimal beam designs with significantly different geometries than MVFOSM. The MPP-based methods yield a more conservative value of Fmin due to the improved estimation of β. Each of the three methods

428

Structural design optimization considering uncertainties

Fmin(∆W, Sr)

1.5 2

2.74

2.5

Residual stress Sr (MPa)

3 6.87

3.5 4

11

4.5 5

15.13

5.5 6

19.26

6.5 0.36

0.28

0.2

0.12

0.04

Width bias ∆W (µm)

Figure 15.7 Contour plot of F min (d, x) as a function of uncertain variables x (design variables d fixed at MVFOSM optimum). Dashed line: limit state g(x) = F min (x) = −5.0; plus sign: MPP from AMV2 +; circle: MPP from FORM; triangle indicates contour corresponding to F min = −6.2 (optimal expected value from MPP-based RBDO runs).

yields an improved design that respects the reliability constraint. The variability in Fmin has been reduced from approximately 5.6 (initial) to 0.52 (MVFOSM design), 0.67 (AMV2 + .design), or /0.65 (FORM design) µN per (FORM verified) input standard deviation E[Fminβ]−Fmin , resulting in designs that are less sensitive to input uncertainties. For the MVFOSM optimal design, the verified values of β calculated by AMV2 + and FORM differ by 6%, illustrating a typical challenge engineering design problems pose to reliability analysis methods. Figure 15.7 displays the results of a parameter study for the metric Fmin (d, x) as a function of the uncertain variables x for design variables fixed at the optimum from MVFOSM RBDO. Since the uncertain variables are both normal, the transformation to u-space used by AMV2 + and FORM is linear, so the contour plot is scaled to a ±3 standard deviation range in the native x-space. The relevant limit state for MPP searches, g(x) = Fmin (x) = −5.0, is indicated by the dashed line. For some design variable sets d (not depicted), the limit state is relatively well-behaved in the range of interest and first-order probability integrations would be sufficiently accurate. For the design variable set used to generate Figure 15.7, the limit state has significant nonlinearity, and thus demands more sophisticated probability integrations. The most probable points converged to by the AMV2 + and FORM methods are denoted in Figure 15.7 by the plus sign and circle, respectively. While the distance from each point

O v e r v i e w o f r e l i a b i l i t y a n a l y s i s a n d d e s i g n c a p a b i l i t i e s i n DA K OTA

429

to the origin differs slightly (see verified β values in Table 15.7), there clearly exist multiple candidates for the most probable point u satisfying Eq. 16. This appears to be a common occurrence when using RBDO methods: the tendency of the optimizer to push the design into a corner where the mean response is encircled by the failure domain. Unfortunately, even second-order probability integration does a poor job in these situations due to exception handling requirements for negative principal curvatures in Eqs. 36–37. This motivates future use of global reliability methods within RBDO to properly estimate the probabilities in these situations. Another computational difficulty observed during design optimization of an earlier bistable mechanism design is simulation failure resulting from model evaluation at extreme values of physical and/or geometric parameters. For example, during an MPP search, edge bias W might grow in magnitude into its left tail causing the effective width of the beam to shrink, possibly resulting in too flimsy a structure to simulate. In summary, highly nonlinear limit states, nonsmooth and multimodal limit states, and simulation failures caused by, e.g., evaluations in the tails of input distributions, pose challenges for RBDO in engineering applications, and must be mitigated through development of algorithms hardened against these challenges, careful attention to problem formulation, and ongoing simulation refinement.

6 Conclusions This chapter has overviewed recent algorithm research in first and second-order local reliability methods. A number of algorithmic variations have been presented, and the effect of different limit state approximations, probability integrations, warm starting, most probable point search algorithms, and Hessian approximations has been discussed. These local reliability analysis capabilities have provided the foundation for reliability-based design optimization (RBDO) methods, and bi-level and sequential formulations have been presented. The RBDO formulations employ analytic sensitivities of reliability metrics with respect to design variables that either augment or define distribution parameters for the uncertain variables. An emerging algorithmic capability is global reliability analysis methodologies which address the common limitations of local methods. In particular, nonsmooth limit states can cause convergence problems with gradient-based optimizers and probability integrations for highly nonlinear or multimodal limit states cannot be performed accurately using a low order polynomial representation from a single MPP solution. Efficient global reliability analysis (EGRA), on the other hand, can handle highly nonlinear and multimodal limit states and is insensitive to nonsmoothness since it does not require any derivative information from the response function. Relative performance of these reliability analysis and design algorithms has been measured for a number of benchmark test problems using the DAKOTA software. The most effective local techniques in these computational experiments have been AMV2 + for reliability analysis and second-order sequential/surrogate-based approaches for RBDO. In a low-dimensional multimodal example problem, global reliability analysis has been shown to provide accuracy similar to that of exhaustive sampling with expense comparable to local reliability. Continuing efforts in algorithm research will build on these successful methods through investigation of trust-region model management for approximation-based local reliability analysis, sequential RBDO with mixed surrogate

430

Structural design optimization considering uncertainties

and direct models (for probabilistic and deterministic components, respectively) and RBDO formulations based on global reliability assessment. These reliability analysis and design algorithms have been applied to real-world applications in the shape optimization of micro-electro-mechanical systems, and experiences with this deployment have been presented. Issues identified in deploying reliability methods to complex engineering applications include highly nonlinear, nonsmooth/noisy, and multimodal limit states, and potential simulation failures when evaluating parameter sets in the tails of input distributions. In addition, RBDO methods tend to exacerbate the reliability analysis challenges by exhibiting the tendency to push the design into a corner where the mean response is encircled by the failure domain. To mitigate these challenges, continuing development of new algorithms that have been hardened for engineering design applications, careful attention to design under uncertainty problem formulations, and refinements to modeling and simulation capabilities are recommended.

Acknowledgments The authors would like to express their thanks to the Sandia Computer Science Research Institute (CSRI) for support of this collaborative work between Sandia National Laboratories and Vanderbilt University.

References Adams, B.M., Bichon, B.J., Eldred, M.S., Carnes, B., Copps, K.D., Neckels, D.C., Hopkins, M.M., Notz, P.K., Subia, S.R. & Wittwer, J.W. 2006. Solution-verified reliability analysis and design of bistable mems using error estimation and adaptivity. Technical Report SAND2006-6286, Sandia National Laboratories, 2006, October, Albuquerque, NM. Adams, B.M., Eldred, M.S. & Wittwer J.W. 2006. Reliability-based design optimization for shape design of compliant micro-electro-mechanical systems. In Proceedings of the 11th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, Number AIAA2006-7000, 2006, September 6–8, Portsmouth, VA. Agarwal, H., Renaud, J.E., Lee, J.C. & Watson, L.T. 2004. A unilevel method for reliabilitybased design optimization. In Proceedings of the 45th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, Number AIAA-2004-2029, 2004, April 19–22, Palm Springs, CA. Allen, J.J. 2005. Micro Electro Mechanical System Design. Boca Raton: Taylor and Francis. Allen, M. & Maute, K. 2004. Reliability-based design optimization of aeroelastic structures. Struct. Multidiscip. Optim. 27:228–242. Ananthasuresh, G.K., Kota, S. & Gianchandani, Y. 1994. A methodical approach to the design of compliant micromechanisms. In Proc. IEEE Solid-State Sensor and Actuator Workshop, Hilton Head Island, SC, pp. 189–192. Bichon, B.J., Eldred, M.S., Swiler, L.P., Mahadevan, S. & McFarland, J.M. 2007. Multimodal reliability assessment for complex engineering applications using efficient global optimization. In Proceedings of the 9th AIAA Non-Deterministic Approaches Conference, Number AIAA2007-1946, 2007, April 23–26, Honolulu, HI. Box, G.E.P. & Cox, D.R. 1964. An analysis of transformations. J. Royal Stat. Soc. 26: 211–252. Breitung, K. 1984. Asymptotic approximation for multinormal integrals. J. Eng. Mech., ASCE 110(3):357–366.

O v e r v i e w o f r e l i a b i l i t y a n a l y s i s a n d d e s i g n c a p a b i l i t i e s i n DA K OTA

431

Chen, X. & Lind, N.C. 1983. Fast probability integration by three-parameter normal tail approximation. Struct. Saf. 1:269–276. Der Kiureghian, A. & Liu, P.L. 1986. Structural reliability under incomplete probability information. J. Eng. Mech. ASCE 112(1):85–104. Du, X. & Chen, W. 2004. Sequential optimization and reliability assessment method for efficient probabilistic design. J. Mech. Design 126:225–233. Eldred, M.S. & Bichon, B.J. 2006. Second-order reliability formulations in DAKOTA/UQ. In Proceedings of the 47th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference, Number AIAA-2006-1828, 2006, May 1–4, Newport, RI. Eldred, M.S. & Dunlavy, D.M. 2006. Formulations for surrogate-based optimization with data fit, multifidelity, and reduced-order models. In Proceedings of the 11th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, Number AIAA-2006-7117, 2006, September 6–8, Portsmouth, VA. Eldred, M.S., Adams, B.M., Copps, K.D., Carnes, B., Notz, P.K., Hopkins, M.M. & Wittwer, J.W. 2007. Solution-verified reliability analysis and design of compliant microelectro-mechanical systems. In Proceedings of the 9th AIAA Non-Deterministic Approaches Conference, Number AIAA-2007-1934, 2007, April 23–26, Honolulu, HI. Eldred, M.S., Agarwal, H., Perez, V.M., Wojtkiewicz, S.F. Jr. & Renaud, J.E. 2007. Investigation of reliability method formulations in DAKOTA/UQ. Structure & Infrastructure Engineering: Maintenance, Management, Life-Cycle Design & Performance, Vol. 3, No. 3, Sept. 2007, pp. 199–213. Eldred, M.S., Brown, S.L., Adams, B.M., Dunlavy, D.M., Gay, D.M., Swiler, L.P., Giunta, A.A., Hart, W.E., Watson, J.-P., Eddy, J.P., Griffin, J.D., Hough, P.D., Kolda, T.G., MartinezCanales, M.L. & Williams, P.J. 2006. DAKOTA, a multilevel parallel object-oriented framework for design optimization, parameter estimation, uncertainty quantification, and sensitivity analysis: Version 4.0 users manual. Technical Report SAND2006-6337, Sandia National Laboratories, Albuquerque, NM. See http://www.cs.sandia.gov/DAKOTA/software. html. Accessed October, 2006. Eldred, M.S., Giunta, A.A., Wojtkiewicz, S.F. Jr. & Trucano, T.G. 2002. Formulations for surrogate-based optimization under uncertainty. In Proceedings of the 9th AIAA/ISSMO Symposium on Multidisciplinary Analysis and Optimization, Number AIAA-2002-5585, 2002, September 4–6, Atlanta, GA. Fadel, G.M., Riley, M.F. & Barthelemy, J.-F.M. 1990. Two point exponential approximation method for structural optimization. Structural Optimization 2(2):117–124. Gill, P.E., Murray, E.W., Saunders, M.A. & Wright, M.H. 1998. User’s guide for npsol 5.0: A fortran package for nonlinear programming. Technical Report SOL 86-1, System Optimization Laboratory, Stanford University, Stanford, CA. Giunta, A.A. & Eldred, M.S. 2000. Implementation of a trust region model management strategy in the DAKOTA optimization toolkit. In Proceedings of the 8th AIAA/USAF/NASA/ISSMO Symposium on Multidisciplinary Analysis and Optimization, Number AIAA-2000-4935, 2000, September 6–8, Long Beach, CA. Haldar, A. & Mahadevan, S. 2000. Probability, Reliability, and Statistical Methods in Engineering Design. New York: Wiley. Hohenbichler, M. & Rackwitz, R. 1986. Sensitivity and importance measures in structural reliability. Civil Eng. Syst. 3:203–209. Hohenbichler, M. & Rackwitz, R. 1988. Improvement of second-order reliability estimates by importance sampling. J. Eng. Mech., ASCE 114(12):2195–2199. Hong, H.P. 1999. Simple approximations for improving second-order reliability estimates. J. Eng. Mech. ASCE 125(5):592–595. Jensen, B.D., Parkinson, M.B., Kurabayashi, K., Howell, L.L. & Baker, M.S. 2001. Design optimization of a fully-compliant bistable micro-mechanism. In Proc. 2001 ASME Intl. Mech. Eng. Congress and Exposition, 2001, November 11–16, New York, NY.

432

Structural design optimization considering uncertainties

Jones, D.R., Shonlau, M. & Welch, W. 1998. Efficient global optimization of expensive blackbox functions. INFORMS J. Comp. 12:272–283. Karamchandani, A. & Cornell, C.A. 1992. Sensitivity estimation within first and second order reliability methods. Struct. Saf. 11:95–107. Kemeny, D.C., Howell, L.L. & Magleby, S.P. 2002. Using compliant mechanisms to improve manufacturability in MEMS. In Proc. 2002 ASME DETC, Number DETC2002/DFM-34178. Meza, J.C. 1994. OPT++: An object-oriented class library for nonlinear optimization. Technical Report SAND94-8225, Sandia National Laboratories, 1994, March, Albuquerque, NM. Qiu, J. & Slocum, A.H. 2004. A curved-beam bistable mechanism. J. Microelectromech. Syst. 13(2):137–146. Rackwitz, R. 2002. Optimization and risk acceptability based on the Life Quality Index. Struct. Saf. 24:297–331. Rackwitz, R. & Fiessler, B. 1978. Structural reliability under combined random load sequences. Comput. Struct. 9:489–494. Rosenblatt, M. 1952. Remarks on a multivariate transformation. Ann. Math. Stat. 23(3): 470–472. Tu, J., Choi, K.K. & Park, Y.H. 1999. A new study on reliability-based design optimization. J. Mech. Design 121:557–564. Wang, L. & Grandhi, R.V. 1994. Efficient safety index calculation for structural reliability analysis. Comput. Struct. 52(1):103–111. Wittwer, J.W., Baker, M.S. & Howell, L.L. 2006. Robust design and model validation of nonlinear compliant micromechanisms. J. Microelectromechanical Sys. 15(1). To appear. Wojtkiewicz, Jr. S.F., Eldred, M.S., Field, Jr. R.V., Urbina, A. & Red-Horse, J.R. 2001. A toolkit for uncertainty quantification in large computational engineering models. In Proceedings of the 42rd AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, Number AIAA-2001-1455, 2001, April 16–19, Seattle, WA. Wu, Y.-T. 1994. Computational methods for efficient structural reliability and reliability sensitivity analysis. AIAA J. 32(8):1717–1723. Wu, Y.-T. & Wirsching, P.H. 1987. A new algorithm for structural reliability estimation. J. Eng. Mech. ASCE 113:1319–1336. Wu, Y.-T., Millwater, H.R., & Cruse, T.A. 1990. Advanced probabilistic structural analysis method for implicit performance functions. AIAA J. 28(9):1663–1669. Wu, Y.-T., Shin, Y., Sues, R. & Cesare, M. 2001. Safety-factor based approach for probabilitybased design optimization. In Proceedings of the 42rd AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, Number AIAA-2001-1522, 2001, April 16–19, Seattle, WA. Xu, S. & Grandhi, R.V. 1998. Effective two-point function approximation for design optimization. AIAA J. 36(12):2269–2275. Zou, T., Mahadevan, S. & Rebba, R. 2004. Computational efficiency in reliability-based optimization. In Proceedings of the 9th ASCE Specialty Conference on Probabilistic Mechanics and Structural Reliability, 2004, July 26–28, Albuquerque, NM.

Part 2

Robust Design Optimization (RDO)

Chapter 16

Structural robustness and its relationship to reliability Jorge E. Hurtado National University of Colombia, Manizales, Colombia

ABSTRACT: Two main ways of incorporating structural uncertainties in design optimization under a probabilistic point of view have been proposed in the international literature: robust and reliability-based design options. While the former is oriented to a reduction of the spread of critical responses, the latter aims to control the probabilities of failure. However, since the reduction of response spread does not preclude a regard to extreme cases, both methods can be considered as complementary. This makes desirable to have at a disposal methods yielding a design satisfying reliability and robustness criteria. In this chapter, methods allowing a simultaneous calculation of the leading probabilistic quantities used in these approaches are examined. It is shown that the the combination of the saddlepoint expansion of probability density together with the method of point estimates for approximating the response statistical moments yields good results. The chapter also deals with a rigorous definition of robustness, which is somewhat loose in the literature. It is shown that the entropy concept is highly appealing to such a purpose, as it shows similarities with that of controllability and stability concepts in dynamic systems theory. On this basis the concept of Robustness Assurance in structural design is introduced, paralleling that of Quality Assurance in the construction phase. A practical method for robust optimal design interpreted as entropy minimization is presented for the common case of linear structures.

1 Introduction 1.1 Theoretic al and practical approache s to unc e rtai nty In an essay that deserves attention (Sexsmith 1999), R. G. Sexsmith remarks that, despite the rapid development of structural reliability theories and methods, their inclusion into the design practice by structural engineers has been little. This rejection is attributed by the author mainly to educational problems. In fact, modern natural science, which arose from mechanical sciences developed in the sixteenth century on the basis of a mathematical interpretation of nature, showed in its beginnings a trend to interpret the random results of experiments as a deficiency of mathematical models rather than as a property of nature itself. In another interesting essay, G. F. Klir quotes the explicit objection common in the XIX century uncertainty jargon (Klir 1997). In those times, uncertainty was rejected as a natural phenomenon because the illusion of a science providing exact answers was still alive and enthusiastic. However, the introduction of the mathematical models for probability and randomness became absolutely

436

Structural design optimization considering uncertainties

necessary to explain phenomena in thermodynamics and quantum mechanics. From that time on, the old paradigm of a exact science was abandoned in those areas where the evidence and the magnitude of randomness could no longer be ignored. Nowadays, as a consequence of the need of considering complex systems, we assist to the development of proposals intended to enhance or overcome the modeling of randomness and uncertainty offered by probability theory, such as possibility theory (Dubois and Prade 1988), fuzzy set theory (Kosko 1992, e.g.), interval analysis (Kharitonov 1997; Hansen and Walster 2004), clustering analysis (Ben-Haim 1985; Ben-Haim 1996; Elishakoff 1999), ellipsoidal modeling (Chernousko 1999), etc. Structural and mechanical engineering continue the tradition of classical mechanics as developed by Galileo and Newton. This fact may be invoked to explain the above mentioned reluctance to include uncertainty models in structural design, at a difference to the well established central consideration of randomness in quantum mechanics and other, later branches of physics. In the author’s opinion, the explanation of this fact cannot entirely be attributed to this historical reason. In addition to this, there is the smaller randomness present in most structural and mechanical situations (with the exemption of earthquake loads and others), as compared to that present in quantum mechanics. But more important is the nature of the challenge posed to the structural engineer, namely, the design of an object. In Kant’s terms, modern science is oriented by the two-way approach of analysis (a priori mathematical principles, which are exact) and synthesis (a posteriori empirical facts, to which mathematical models must accommodate), and it offers a knowledge that remains valid for some time until it is falsified, according to Popper’s theory of science. Engineering, on the contrary, aims to offer not a knowledge but a product, which is defended not by arguments but by its quality, whose cost must be a minimum and whose design must be produced in most cases with resort to simplifying rules. This implies taking decisions, which is a challenge not hanging over knowledge discovering. All this may perhaps explain the somewhat paradoxical fact that structural engineers, on the one hand, do not include probabilities into their calculations, but, on the other, have for long recognized the importance of uncertainties in the design practice, as expressed in the use of safety factors of several kinds and of statistical analysis of experiments for fixing their code values. From the practical design viewpoint it is not realistic to expect that in actual structural designs failure probabilities will ever be calculated as a part of conventional design process of most structures. Besides the educational and computational problems involved, there is the lack of sufficient probabilistic information on load and material parameters, the difficulties for interpreting such probabilities, the high sensitivity of these values to probabilistic models, the randomness of the results conveyed by the most universal method (Monte Carlo simulation), etc. But the main reason is and will be the pragmatism of design. Thus, it can be said that randomness is in fact considered in structural design, but in a manner that results quite unsatisfactory from the analytical, argumentative, mathematically-oriented point of view. In fact, the close examination of the relationship between safety factors and failure probabilities conducted recently by I. Elishakoff (Elishakoff 2005) shows that a link between them strongly depends on the probabilistic structure of the random variables x in hand1 . But this is just the information from 1 In this chapter an underlined letter indicates a random variable or a vector, if it is written boldface.

Non-underlined letters are used to denote either their realizations, their deterministic counterparts or deterministic variables in general.

Structural robustness and its relationship to reliability

437

whose need safety factors intend to dispense with. The contradiction lies in the fact that reliability requires a scientifically-oriented calculation, whereas safety factors are a mere practical tool for producing a qualified product. The requirement of practical, sometimes simplifying approaches can be considered as the implicit but dominating rule in engineering design. This has fostered the development of the concept of robustness, meaning a product that exhibits strength with respect to variations or fluctuations of parameters, sometimes random, sometimes uncontrollable and sometimes unknown. A robust product assures the engineer that it can absorb such fluctuation without compromising its quality, which is its main feature. Such, in fact, is the rationale behind the concept of BIBO (bounded input, bounded output) stability of dynamic systems (Szidarovszky and Bahill 1992, e.g.). In that field, the certainty that under an action of a bounded input the system response will be also bounded is considered sufficient by the engineer to be relieved from the need of tracing the exact trajectory of the response in particular situations. In general, robust design orientation aims at overcoming the need of considering particular uncertain situations and to assure the designer the imperturbability of the system under the presence of unknown, unpredictable or random parameters. Notice, however, that the question for a quantitative measure of the uncertainty in extreme situations is not solved by the robustness approach. 1.2

Optimal structural des ign under un c e rtai nty

Arising from the deterministic orientation of modern engineering mentioned above, structural optimization is normally performed without regard to random fluctuations of the parameters. It consists in minimizing a cost function C(y) subject to deterministic constraints posed upon responses (displacements, stresses, etc) and geometrical quantities. Formally, this problem is expressed as (Haftka et al. 1990; Kirsch 1993, e.g.) Problem Deterministic optimization : find : y minimizing : C(y) subject to : fi (y) < Fi , i = 1, 2, . . . y− ≤ y ≤ y+

(1)

In this equation y is the vector of design variables, fi (y) are system responses depending on them and Fi are their limiting values. These constitute the so-called behavioral constraints. On the other hand y− and y+ are bounds imposed to the design variables, normally constituting geometric constraints. It is evident that the uncertainties present in loads, materials and elements are not explicitly taken into account. Thus the result may be a fragile structure with respect to random changes in the design parameters. Notice, however, that in the definition of the upper bounds in behavioral and geometric constraints there is an implicit recognition of the risk associated to values excessively low. The decision on these bounds is normally taken on the basis of safety factors, which simply express a caution with respect to randomness and uncertainty. Anyhow, the robustness and the reliability of a structure designed using safety factors without regard to the probabilistic definition of the random variables present in it remain rather uncertain.

438

Structural design optimization considering uncertainties

The explicit consideration of uncertainties in structural design optimization is a challenging task, as it demands the minimization of cost functions in a noisy environment generated by the presence of certain random variables. However, it can be considered an analysis of maximum importance, because it yields the solution to the ideal of producing a structural model that is both economical and safe. Two main families of methods have been proposed to this end: 1) Robust Design Optimization (RDO), which is oriented to minimizing the spread of the structural responses, as measured by low-order statistical moments. 2) Reliability-Based Design Optimization (RBDO), which minimizes the cost function with probabilistic constraints (Rosenblueth and Mendoza 1971; Gasser and Schuëller 1997; Frangopol 1995; Royset et al. 2001; Royset and Polak 2004). A common formulation of RBDO can be formally presented as follows: Problem Reliability-based optimization : find : y minimizing : C(y) (2) subject to : P[fi (x, y) > Fi ] ≤ Pi , i = 1, 2, . . . y− ≤ y ≤ y+ where x is a set of random variables, P[A] the probability of the random event A and Pi its limiting value. Function gi (x, y) = Fi − fi (x, y) is known in structural reliability as the limit state function. Other formulations than Eq. (2) are, however, possible. The following are some criticisms that have been addressed to the application of structural reliability for performing a structural optimization under uncertainty: • • •

•

The lack of information about actual probability models for materials and loads and the concern about the applicability of the published ones in every case. The sensitivity of the tails of the input probability density models to their parameters (Elishakoff 1991; Ben-Haim 1996; Elishakoff 1999). This undoubtedly affects the value of the failure probability. The difficulty of interpretation among the engineering community of the meaning of the failure probability. Despite structural safety researchers stress that it is a nominal failure indicator, there is a natural tendency to interpret it in the frequentist sense. On the other hand the Bayesian interpretation as a belief measure has gained favor but in the limited field of health monitoring and other tasks associated to existing structures. It is difficult to accommodate such an interpretation for new projects. These and other phenomena explain the little use of structural reliability concepts in design practice (Sexsmith 1999). The limitations of some popular methods of calculating such probabilities. For instance, FORM presents problems of accuracy for non-linear limit state functions and convergence problems (Schuëller and Stix 1987); the Response Surface Method exhibits problems of instability (Guan and Melchers 2001); Monte Carlo simulation requires high computational efforts, etc. There is a continuous effort among researchers for improving these techniques and developing new ones (Au and Beck 2001, e.g.), but there is no agreement about a method that satisfies both the requirements of generality, accuracy and low computational cost. An updated, general benchmark study is presently lacking.

Structural robustness and its relationship to reliability

•

439

The need of calculating one or more failure probabilities, which in some cases is a time-consuming task, for each trial model, increasing enormously the computational effort with respect to conventional, deterministic optimization and to reliability analysis as well. If Monte Carlo simulation is used this problem can be greatly alleviated if use is made of convenient to apply solver surrogate techniques such as neural networks (Papadrakakis et al. 1996; Hurtado and Alvarez 2001; Hurtado 2001) or support vector machines (Hurtado 2004a; Hurtado 2004b; Hurtado 2007). For performing the optimization, these methods can be combined with optimization techniques with biological optimization such as genetic algorithms, evolutionary strategies (Papadrakakis et al. 1998; Lagaros et al. 2002; Lagaros and Papadrakakis 2003), particle swarm optimization (Hurtado 2006), etc.

Some structural designers tend to favor the concept of robustness, understood as safety against unpredictable variations of the design parameters, over the concept of failure probability, which is normally a very low value lacking significant meaning in practice. This may be explained by the production-oriented approach of design discussed above. Robustness can be defined in several forms, depending on whether use is made of the clustering (Ben-Haim 1985; Ben-Haim 1996) or conventional, frequentist interpretation of uncertainty. In this chapter the second interpretation is adopted. The following formulation of robust design optimization corresponds to the proposal in (Doltsinis and Kang 2004; Doltsinis et al. 2005): Problem Robust optimization : find : y minimizing : subject to :

 C(y) = (1 − α)E[f (y)]/µ∗ + α Var[f (y)]/σ ∗ (3)  E[gi (y)] + βi Var[gi (y)] ≤ 0, i = 1, 2, . . .  Var[hj (y)] ≤ σj+ , j = 1, 2, . . . y− ≤ y ≤ y+

where f (y) is a performance function, 0 < α < 1 is a factor weighting the minimization of its mean and standard deviation, βi > 0 is a factor defining the control of the response gi (y) in the tail of its distribution, σj+ an upper bound to the standard deviation of response hj (y) and µ∗ , σ ∗ are normalizing factors. Many other formulations are, however, possible. Globally speaking, the essential of robustness optimization is the control of low order statistical moments of the response. The nature of these two alternative methods can be explained with the help of Fig. 16.1, which shows three alternative probability density functions of a structural response. While RDO aims to reduce the spread, RBDO is intended to bound the probability of surpassing the critical threshold. Notice that in applying RDO the effect pursued by RBDO is indirectly obtained, because the reduction of the spread implies a reduction of the failure probability. The reliability (or its complement, the failure probability) refers to the occurrence of extreme events, whereas the robustness refers to the low spread of the structural responses under large variation of the input parameters.

440

Structural design optimization considering uncertainties

pz(z)

z

Figure 16.1 Robust and reliability-based design options.While the first aims at reducing the spread of the response function, the second attempts to control the probability of surpassing a critical threshold (dashed line). However, low failure probabilities may correspond to large spreads (dotted line).

This is assumed to assure a narrow response density function, which in turn assures a low failure probability, if it is unimodal, as is common case. However, this is not necessarily true: to a significant spread of the structural response may correspond a low failure probability because the definition of the limit state can be such that the possibility of surpassing it is very rare, as the situation it describes is rather extreme (See Fig. 16.1). In applying the RDO other possibilities exist, such as moving the probability density function away from a critical threshold or a combination of both approaches. 1.3 Ai m s and s c o pe The above discussion means that a comparison between RDO and RBDO on the basis of some examples cannot be conclusive, because, as Fig. 16.1 shows, it all depends on the critical thresholds selected for reliability estimations. Besides, the relationship between statistical moments and probabilities is severely nonlinear. For these reasons, to a good consideration of the uncertainties in structural design both approaches are valuable and complementary. This justifies the search of techniques that allow establishing a link between them, which is the purpose of the research reported herein. In fact, since both kinds of designs correspond to a different way of incorporating the uncertainties and to different goals, a method allowing a joint monitoring both the moments (mean and variance), on the one hand, and the failure probabilities, on the other, at a low computational cost, would be of avail. A simple link between RDO and RBDO is given by inequalities involving low order moments and the probability of exceeding a certain threshold. Two of them are the following (Abramowitz and Stegun 1972): •

Bienaymé – Markov inequality: P[x > ω] ≤

E(x) , ω

ω>0

(4)

Structural robustness and its relationship to reliability

•

441

Chebyshev inequality: P[|x − E(x)| ≥ tσx ] ≤

1 , t2

t>0

(5)

where σx is the standard deviation of the random variable x. These bounds, however, are reputed to be not tight when the probabilities are very low, as is common case in structural safety. Bounds such as those expressed by the above inequalities are employed when the probabilistic information is not sufficient to calculate exceedance probabilities. In probability theory the concept of entropy associates information and uncertainty in a clear, positive manner. For this reason, a second aim of the chapter is to discuss the concept of robustness from this point of view. From the production-oriented, boundassuring approach of engineering design discussed above, robustness with respect to uncontrollable external actions can be controlled in a similar fashion as the randomness of the material properties can be subjected to quality control. For this reason the concept of Robustness Assurance is introduced, referring to the control of the response spread under the influence of the uncertainty of external actions, in a similar manner as Quality Assurance in construction industry subjects to control the randomness of material properties and structural member dimensions. It is shown that robustness assurance defined in this manner can easily be incorporated into conventional deterministic optimization. The chapter is organized as follows. First, the methods for estimating the failure probability upon this information are discussed: They can be grouped into (a) global expansions and (b) local expansions. It is shown that the later offers significant advantages for the purpose in hand. However, one of the global expansion techniques, namely the maximum entropy method, is useful for linking robustness and reliability and therefore it is discussed in some detail. Next the methods allowing estimation of high order moments of the response are briefly presented, with an emphasis on the point estimates technique. The application of this method to robust design is then discussed. An example illustrates the accuracy and the low computational cost of the joint computation of moments and probability estimates by the proposed procedures. Then, the definition of robustness in terms of entropy and the ensuing derivations for optimization of linear structures is introduced. It is shown that Robustness Assurance can easily be incorporated into conventional deterministic optimization. The practical application of this concept is developed and illustrated for the case of linear structures. The chapter ends with some conclusions. Since the information on concepts and methods involved in the exposition is disperse in journals and books published along five decades, the chapter is as self-contained as possible.

2 Probability estimation based on moments In this section the estimation of probability density function based on the information provided by statistical moments is reviewed, as it offers a general link between robust and reliability-based design methods. This problem can be approached by means of Pearson, Johnson or other families of distributions (Johnson et al. 1994). However,

442

Structural design optimization considering uncertainties

the discussion herein limits to polynomial and maximum entropy families of methods. A digression into the latter is instructive as it sheds light on the nature of robustness discussed at the end of the chapter. The above mentioned proposals for density estimation are global, i.e. they are valid for all values of the random variable. For estimating the reliability it would be more interesting to expand the density about a critical threshold. This is the purpose of the saddlepoint expansion, explained in the last paragraph of the present section.

2.1 Pol yno m i al expans io ns Classical probability theory provides two expansions of the probability density function based on moments. They are the Gram-Charlier and the Edgeworth expansions given respectively by (Muscolino 1993, e.g.):

pz (z) =

1 κ3 1 κ4 10 κ32 H (z) + H (z) + H6 (z) 3 4 3! σz3 4! σz4 6! σz6  280 κ33 35 κ3 κ4 H7 (z) + H9 (z) φ(z) + 7! σz7 9! σz9 1+

(6)

and

pz (z) =

1+ +

1 κ3 1 κ4 10 κ32 H (z) + H (z) + H6 (z) 3 4 3! σz3 4! σz4 6! σz6

 35 κ3 κ4 280 κ33 1 κ52 H (z) + H (z) + H (z) φ(z) 5 7 9 5! σz5 7! σz7 9! σz9

(7)

where φ(z) is the standard Gaussian density 

1 1 φ(z) = √ exp − z2 2 2π

(8)

κi and Hi (z) are respectively the cumulant and the Hermite polynomial of order i. As is well known the following relationships hold for the first cumulants and moments µj = E[zj ] (Kolassa 1997): κ1 = µ1 κ2 = µ2 − µ21 κ3 = µ3 − 3µ1 µ2 + 2µ31 κ4 = µ4 − 4µ1 µ3 − 3µ22 + 12µ2 µ21 − 6µ41

(9)

Structural robustness and its relationship to reliability

443

The coefficients are those of the Pascal triangle. The first Hermite polynomials are given by (Abramowitz and Stegun 1972, e.g.) H1 (z) = z H2 (z) = z2 − 1 H3 (z) = z3 − 3z H4 (z) = z4 − 6z2 + 3 H5 (z) = z5 + 10z3 + 15z H6 (z) = z6 − 15z4 + 45z2 − 15

(10)

Despite the similarities between Gram-Charlier and Edgeworth expansions it is worth noticing that they emerge from rather different approaches: The first from an orthogonal expansion of the probability density function and the second from the Fourier transform of a non-Normal characteristic function. In practice the use of Edgeworth expansion is more often recommended. However, notice first that they use polynomials, whose behavior is more oscillating as their order increase. Hermite polynomials, in particular, are negative for intervals that are the larger, the higher their order (Abramowitz and Stegun 1972, e.g.). As a consequence, the probability density estimate may not be strictly positive in certain intervals and, in compensation, may exhibit undesirable multimodality in other ones. A discussion on the use of polynomial approximations to the density function based on moments can be found in (Kennedy and Lennox 2000; Kennedy and Lennox 2001), where the authors propose a method based on non-classical orthogonal polynomials. As far as the examples presented in the mentioned chapters may be conclusive, the method seems to overcome the deficiencies of the classical approaches. It consists in an approximation of the form pz (z) = w(z)

r 

ai Qi (z)

(11)

i=0

where w(z) is a weighting function selected upon judgement of the moments in hand, Qi (z) are orthogonal polynomials and ai coefficients to be determined. Notice, however, that the possibility of having a negative value of the density remains. This problem is corrected in Er’s method (Er 1998) with an approximation of the form pz (z) = C exp(Q(z))

(12)

due to the strict nonnegativity of the exponential function. Here C is a normalizing constant and Q(z) =

r  i=1

ai zi

(13)

444

Structural design optimization considering uncertainties

The coefficients are obtained by solving the following algebraic problem: ⎞⎛ ⎞ ⎛ ⎞ a1 0 1 2µ1 . . . rµr−1 ⎟ ⎜ µ1 2µ2 . . . rµr ⎟ ⎜ a2 ⎟ ⎜ −1 ⎟⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ .. ⎟ = ⎜ ⎟ ⎜ .. .. .. . . .. ⎝ ⎝ ⎠ ⎠ ⎠ ⎝ . . . . . . −(r − 1)µr−2 µr−1 2µr . . . rµ2(r−1) ar ⎛

(14)

Notice that for calculating r coefficients the number of moments needed equals 2(r − 1).

2.2

Ma x i m u m e nt r o py me t ho d

In probability theory, entropy is a simultaneous measure of the information and uncertainty given by the present samples (Shannon 1948; Jaynes 1957). In fact, a deterministic event offers no information at all, while a purely random event (having uniform distribution) offers the maximum. Therefore, entropy establishes a connection between information and uncertainty. The random samples of an event A can be expressed by means of many possible partitions U, i.e. collections of mutually exclusive subsets Ai , i = 1, 2, . . . of A in which the random occurrences are allocated. The entropy of the partition is defined by Shannon (Shannon 1948) as H(U) = −



pi ln pi

(15)

i

where pi is the probability associated to subset Ai . Empirically, if there are N samples of the event and Ni are located in subset Ai , then pi ≈ Ni /N. A continuous expression of a partition is a probability density function, in terms of which entropy is defined as
Hx = −

px (x) ln px (x)dx

(16)

There is an important difference between entropy definitions for discrete and continuous cases: It is an absolute measure of uncertainty in the former case, while a relative one in the latter, as it changes with the coordinate system (Shannon 1948). (See Eq. (76)). This remark is important for the development of a robustness assurance method proposed in the final section of present chapter. The principle of maximum entropy states that the most unbiased estimate of the probability density function of a random event is that maximizing Eq. (16). The principle determines a method for estimating the density function upon the availability of knowledge about the random event, such as e.g. statistical moments. If, for instance, such a knowledge consists of ordinary moments µk , k = 1, 2, . . . , the

Structural robustness and its relationship to reliability

445

method of maximum entropy (MEM) consists in solving the following optimization problem: Problem Maximum entropy : find : px (x)
maximizing : H = − px (x) ln px (x)dx
subject to : gk (x)px (x)dx = θk , k = 1, 2, . . .

(17)

where gk (x), θk are known functions and their expected values, respectively. When these correspond to ordinary moments, i.e. gk (x) = xk , θk = µk the result is px (x) = exp(−λ0 − λ1 x − λ2 x2 − λ3 x3 · · · ) where λk are Lagrange multipliers, with λ0 acting as a normalizing constant: 
λ0 = ln exp(−λ1 x − λ2 x2 − λ3 x3 − · · · )

(18)

(19)

Upon replacing these equations into the definition (16) the following important result is obtained: Hx = λ0 + λ1 µ1 + λ2 µ2 + λ3 µ3 + · · ·

(20)

It is worth noticing that the maximum entropy method is not limited to moment information but it applies to expected values of function in general. In (Shore and Johnson 1980) it is proved that this is the uniquely correct method that satisfies all consistency axioms. Two families of methods have been proposed to find the Lagrange multipliers. The first consists in solving the set of nonlinear equations by means of Newton methods (Mead and Papanicolau 1984; Sobczyk and Tr¸ebicki 1990; Tr¸ebicki and Sobczyk 1996; Hurtado and Barbat 1998; Ching and Hsieh 2007). The other consists in the unconstrained minimization of the concave functional (Agmon et al. 1979; Pandey and Ariaratman 1996) F(ζ0 , ζ1 , . . . ) = ζ0 + ζ1 µ1 + ζ2 µ2 + ζ3 µ3 + · · ·

(21)

because its minimum is the entropy given by Eq. (20). According to the author’s experience, the second approach is much faster and numerically more stable. Notice that Er’s method mentioned above (Er 1998) is based on the same functional form of the density as that resulting from applying the MEM to the case when the information is given by ordinary moments. This is shown by a simple comparison of Eqs. (12) with (13), on the one hand, and (18) with (19), on the other, indicating that λ0 is equivalent to − ln C. However, Er’s method requires a larger number of moments as said in the preceding and, therefore, the results are not coincident.

446

Structural design optimization considering uncertainties

2.3 Sad d l epoint e xpans io n The saddlepoint approximation to an ordinate of a density or distribution function was originally proposed by Daniels (Daniels 1954). In contrast to the classical Gram-Charlier or Edgeworth expansions, it has the advantage of producing good approximations far into the tails, because it is a local rather than a global approximation method. In other words, it is aimed at estimating the functions at a single point only. For a detailed exposition see (Barndorff-Nielsen and Cox 1979; Reid 1988; Cheah et al. 1993; Kolassa 1997). The saddlepoint approximation is based on the idea of embedding the target function within a family of parameterized functions and to select one member of the family for the approximation. Let us first approximate the density function pg (g) by the family rg (g, η) = exp(ηg − Kf (η))pg (g)

(22)

where Kf (η) is the cumulant generating function of the density pg (g), Kf (η) = ln Mf (η) = ln E[ exp(ηg)]
= ln exp(ηg)pg (g)dg

(23)

and η is a parameter. In the preceding equation Mf (η) is the moment generating function of pg (g). Notice that function rg (g, η) satisfies the normalization condition for a density since
exp(ηg − Kf (η))pg (g)dg = exp(−Kf (η))Mf (η) = 1 (24) The parameter η is selected such that the mean of the family of functions equals the ordinate at which the density is to be estimated, g, ¯ which in structural reliability is normally zero. The mean of the family is
g exp(ηx − Kf (η))pg (g)dg


d (exp(ηg))pg (g)dg dη
d = exp(−Kf (η)) (exp(ηg))pg (g)dg dη = exp(−Kf (η))

=

1 d Mf (η) Mf (η) dη

=

d ( ln Mf (η)) dη

= Kf
(η)

(25)

Structural robustness and its relationship to reliability

447

It can also be easily shown that the variance of the saddlepoint density Kf

(η). Hence the parameter is the solution of Kf
(¯η) = g¯

(26)

A convenient choice for the family of approximating functions is the standard Normal (8). Using ¯  this density implies standardizing variable g in the form q = (g − g)/σ, where σ = Kf

(¯η) is the standard deviation. The density of the standardized variable is σ pg (σq + g, ¯ η¯ ) according to the probability transformation rules. Hence we have 

. 1 2 / . / Kf

(¯η) exp η¯ Kf

(¯η)q + g¯ − Kf (¯η) pg Kf

(¯η)q + g¯ = φ(q)

(27)

Setting q ≡ 0 and solving for pg ( · ) yields 1 pg (g) ¯ = ¯ exp(Kf (¯η) − η¯ g) 2πKf

(¯η)

(28)

which is the sought-after saddlepoint approximation for the density at g. ¯ The computation of a probability Q = P[G ≥ g] ¯ eventually requires the calculation of the integral
∞ Q=

 g¯

1 2πKf

(¯η(u))

exp(Kf (¯η(u)) − η¯ u)du

(29)

where η¯ (u) is the solution of Kf
(¯η(u)) = u. Since this should be solved at each integration point, the computational demands and the accumulation of errors in approximating this integral can be large. As an alternative, direct formulas for computing Q have been proposed (Robinson 1982; Lugannani and Rice 1980). In this chapter use will be made of the proposal in (Lugannani and Rice 1980), since in the comparisons made in (Kolassa 1997) it yields the best performance. It is given by

1 1 − Q = 1 −
(ω) ¯ + φ(ω) ¯ ν¯ ω¯

 (30)

  where ν¯ = η¯ Kf

(¯η) and ω¯ = 2(¯ηg¯ − Kf (¯η)). The saddlepoint approximation is commonly applied in Statistics for estimating the density or the distribution at a given ordinate for sums of variables with widely different properties for which the Central Limit Theorem does not give good results (Lange 1999). This implies the solution of Eq. (26) using the derivative of the actual cumulant generating function by means of Newton methods. In our case such a function is not

448

Structural design optimization considering uncertainties

known and resort must be made to an approximation in terms of the cumulants using the series Kf (η) =

∞  κj ηj j=1

(31)

j!

Upon deriving Eq. (31) with respect to η and equating to the threshold, according to Eq. (26), one obtains a polynomial whose lowest real positive root yields the value η¯ . The probability of failure Pf can then be readily estimated with Eq. (30). To conclude the present exposition of the saddlepoint expansion, mention should be done to the use of Monte Carlo simulation for approximating the integral (29). To this end, random numbers are generated from the saddlepoint density and the probability is estimated as the average of the values of the indicator function located on the threshold g, ¯ as is usual in Monte Carlo integration. A method for doing this, using the Metropolis-Hastings simulation method has been proposed (Robert and Casella 1999). The method uses the alternative formulation of the integral, given by 1


2 Q= η¯



Kf

(ϑ) 2π

exp(Kf (ϑ) − ϑKf
(ϑ))dϑ

(32)

which can be obtained from (29) with the change of variable u = Kf
(ϑ). In the numerical experiments reported in the quoted reference the method gives quite similar results to the exact integration. However, notice that the value of the integral hinges upon the lower limit, which in the method proposed herein is known only approximately via the point estimate technique. Hence very small differences can be expected from this simulation approach in comparison to the simple application of the LugannaniRice formula. In addition, the randomness of the failure probability, common to all simulation-based methods, appears.

3 Structural response moment estimation 3.1 Perturb a ti o n appr o ac h Perturbation methods in structural analysis (Hisada and Nakagiri 1981; Liu et al. 1995; Kleiber and Hien 1992) are based on a basic result of the probability theory concerning the approximation of the mean vector and covariance matrix of a function h(x) of a set of r basic variables x = {x1 , xj , . . . , xr }. Function h( · ) can be expanded in Taylor series about the mean vector µx as r r r  ∂h 1   ∂2 h h(x)=h(µ ˙ (µ )[x −µxk ]+ (µ )(x −µxk )(xl −µxl ) x )+ ∂xk x k 2 ∂xk xl x k k=1

k=1 l=1

(33)

Structural robustness and its relationship to reliability

449

Applying the expectation operator to this equations we obtain 1   ∂2 h (µ )Ckl (x) 2 ∂xk xl x r

E[h(x)]=h(µ ˙ x) +

r

(34)

k=1 l=1

where it has been taken into account that E[xk − µxk ] = 0. Here Ckl (x) denotes the (k, l) element of the covariance matrix of the vector x. This equation is known as the second order approximation of the mean of function h( · ). Let us now derive a first order approximation to the covariance of two functions hi (x) and hj (x). To this end multiply the Taylor expansion of the two functions up to the first order derivative terms, i.e. r B C  ∂hi . (µx )[xk − µxk ] hi (x)hj (x) = hi (µx ) + ∂xk k=1

r B C  ∂hj hj (µx ) + (µx )[xl − µxl ] ∂xl

(35)

l=1

Arranging terms yields  ∂hi . hi (x)hj (x) = hi (µx )hj (µx ) + hj (µx ) (µ )[x − µxk ] ∂xk x k r

k=1

+ hi (µx )

r  l=1

+

r 

∂hj (µ )[x − µxl ] ∂xl x l

 ∂hj ∂hi (µx )[xk − µxk ] (µ )[x − µxl ] ∂xk ∂xl x l r

k=1

(36)

l=1

Moving the product hi (µx )hj (µx ) to the left-hand side and taking expectations at both sides of this equation leads to the final result: ∂hj .   ∂hi cov(hi (x)hj (x)) = (µx ) (µ )Ckl (x) ∂xk ∂xl x r

r

(37)

k=1 l=1

The variance of either of the two functions is but a particular case of this equation: ∂hi .   ∂hi (µ ) (µ )Ckl (x), var(hi (x)) = ∂xk x ∂xl x r

r

i = 1, 2

(38)

k=1 l=1

At least three objections can be addressed to perturbation methods for the purpose of present chapter. First, they are reputed to be accurate for low coefficient of variation of the basic variables x (Elishakoff and Ren 2003). Second, they require special computational codes to their application (Kleiber and Hien 1992). Third, as is evident, they do not yield equations for estimating moments of order higher than two, which

450

Structural design optimization considering uncertainties

are needed for applying local or global expansions of probability distributions in order to estimate the probability of failure. The method of point estimates summarized next overcomes these deficiencies.

3.2 Poi n t esti m at e me t ho d The method of Point Estimates (Rosenblueth 1975; Ordaz 1988; Christian and Baecher 1998; Harr 1989; Hong 1998) is a valuable tool for estimating the low order statistical moments of a system response with good accuracy. The reason explaining this property is that the method imposes the annihilation of some order terms in the Taylor expansion of the response and the concentration of their information in some weights located around the mean vector of the basic variables. This is the main difference with perturbation approaches based in the Taylor expansion, which are built over the assumption that the high order terms of the expansion are negligible. Besides, the method of point estimates has the additional advantage over perturbation schemes that it can be easily applied to the estimation of moments of order higher than the second. Last but not least, the method does not need special finite element codes for its application, as required by the perturbation approach. As a consequence, it can be used in connection to practically any structural problem using available numerical tools. In the basic formulation of the method, the total number of finite element solver calls is only twice the number of independent random variables, which in a problem determined by a few basic variables implies a low computational effort. These features make the method an accurate and practical technique for the stochastic performance analysis of mechanical systems. In the following lines the proposal in (Hong 1998) is summarized, because in the experience reported in (Hong et al. 1998) using actual structural models it offers by far the best approximation over the other point estimate alternatives cited above. In addition, the applicability of the method for higher order moment evaluation is discussed. Let us consider a structural function g(x) that is a function of a single variable x. The Taylor expansion of a power function g j (x) about the mean value of x is g j (x) = b(x) = b(µx ) +

∞  1 (l) b (µx )(x − µx )l l!

(39)

l=1

Taking expectations on both sides of the above equation one obtains ∞  1 ∂b (µ )E[(x − µx )l ] E[g (x)] = b(µx ) + l! ∂xl x j

(40)

l=1

which can be put in the form E[g j (x)] = b(µx ) +

∞  1 ∂b (µ )γ σ l l! ∂xl x x,l x l=1

(41)

Structural robustness and its relationship to reliability

451

where σx is the standard deviation of x and γx,l is a normalized central moment defined as
1 ∞ γx,l = l (x − µx )l px (x)dx (42) σx −∞ Multiplying successively equation (39) by two weights wi , l = 1, 2 assigned to the concentration points xl and summing up the result yields w1 b(x1 ) + w2 b(x2 ) = b(µx )(w1 + w2 ) +

∞  1 ∂b (µ )(w1 ξ1l + w2 ξ2l )σxl l! ∂xl x

(43)

l=1

where ξi , i = 1, 2 is the standardized random variable ξi =

xi − µx

(44)

σx

Solving equation (43) for b(µx ), imposing the condition w1 + w2 ≡ 1

(45)

and substituting back the result into equation (41) yields E[g j (x)] = w1 b(x1 ) + w2 b(x2 ) +

∞  1 ∂b (µ )[γ − (w1 ξ1l + w2 ξ2l )]σxl l! ∂xl x x,l

(46)

l=1

This equation suggests the approximation . E[g j (x)] = w1 b(x1 ) + w2 b(x2 ) = w1 g(x1 )j + w2 g(x2 )j

(47)

in which function g( · ) is evaluated at points xi = µx + ξi σx , i = 1, 2. Implicit in the above approximation is the condition that the concentration parameters must also satisfy the following constraint: w1 ξ1i + w2 ξ2i = γx,i

(48)

for an adequate number of normalized moments γx,i allowing the determination of ξl and wl . For determining two weights and concentration points three moment equations of the type (48) are necessary. These, appended to Eq. (45), yield the values of the four parameters. In this case the system has the following closed-form solution: ξi =

γx,3 2

 + ( − 1)

wi = (−1)i

ξ3−i ζ

3−i

1+

γx,3

2

2 (49)

452

Structural design optimization considering uncertainties

 2 /4. For the more general case of a function g(x) of n mutually where ζ = 2 1 + γx,3 uncorrelated random variables xk , k = 1, . . . , n, collected in vector x, it is possible to apply the same strategy as above by setting all variables at their means and applying the Taylor expansion about the mean of each xk in turn. The derivation of the equations for the weights and concentration points can be performed in the same way as for the one dimensional case. As a result, the approximation of the ordinary moment of order j of g(x) with m points per variable is given by .  E[g j (x)] = wk,i g(µ1 , . . . , xk,i , µk+1 , . . . )j m

n

(50)

k=1 i=1

The total number of solver calls will then be S = mn. In (Hong 1998) several alternatives for calculating the weights and point locations are offered, according to the number of concentration points. These are the following: •

S = 2n scheme: ξk,i =

γxk ,3 2

 + ( − 1)

3−i

n+

γ

xk ,3

2

2

ξk,3−i 1 (−1)i (51) n ζk  2 /4. Notice that, in this case, the approximation (50) is = 2 n + γx,3

wk,i = with ζk

•

accurate to the third order of the Taylor expansion, as determined by the number of normalized central moments used in the calculation of the weights and concentration points. S = 2n + 1 scheme: 

γ 2 γxk ,3 xk ,3 3−i + ( − 1) γxk ,4 − 3 ξk,i = 2 2 wk,i = ( − 1)3−i

1 ξk,i (ξk,1 − ξk,2 )

(52)

for i = 1, 2 and ξk,3 = 0 wk,3 =

1 − wk,1 − wk,2 n

(53)

Note that the repetition of the point ξk,3 = 0 makes this three-point scheme equivalent to a 2n + 1-point scheme  . E[g (x)] = w0 g(µ1 , . . . , µk , µk+1 , . . . )j + wk,i g(µ1 , . . . , xk,i , µk+1 , . . . )j (54) n

2

j

k=1 i=1

Structural robustness and its relationship to reliability

•

453

S = 3n scheme: ξk,j ξk,l , (ξk,j ξk,i )(ξk,l ξk,i )

wk,i =

i, j, l = 1, 2, 3;

i = j = l = i

(55)

The locations ξk,i , i = 1, 2, 3 are the roots of the polynomial ω0 + ω1 q1 + ω2 q22 + ω3 q33 = 0

(56)

in which . / ω0 = γxk ,5 − γxk ,3 2γxk ,4 − γx2 ,3

ω3 =

•



k

γx ,4  − γxk ,3 + γxk ,4 1 − k n n

 γ 1 x ,5 − k ω2 = γxk ,3 γxk ,4 + n n ω1 = γxk ,3

γ

xk ,5

γxk ,4 − (1 + γx2 ,3 ) k

(57)

n

The approximation obtained with these points is accurate to the fifth order because it supposes the cancelling of 2m − 1 = 5 terms of the Taylor expansion. S > 3n scheme: In the general case the size of the nonlinear system for determining the weights and location points of becomes 2m. This implies the solution of the following system of nonlinear equations for each variable k: m 

wk,i =

i=1 m 

1 n

j

wk,i ξk,i = γxk ,j

(58)

i=1

A system like this can be solved by an algorithm described in (Hamming 1973; Miller and Rice 1983). Let us expand the system of equations (58) + w2

w1 ξ1

+ w2 ξ2

+ · · · + wm + · · · + wm ξm

w1 ξ12 .. .

+ w2 ξ22 .. .

+ · · · + wm ξm2 .. .. . .

w1

= b0 = n1 = b1 = γ1 = b2 = γ2 .. .

w1 ξ12m−1 + w2 ξ22m−1 + · · · + wm ξm2m−1 = b2m−1 = γ2m−1

(59)

454

Structural design optimization considering uncertainties

where the subindex denoting the random variable has been dropped for clarity. Define a polynomial p(ξ) =

m 

ωl ξ l

(60)

l=0

whose roots are the desired values ξ1 , ξ2 , . . . , ξm , i.e p(ξ) = (ξ − ξ1 )(ξ − ξ2 ) · · · (ξ − ξm )

(61)

From this equation follows that ωm = 0 and that p(ξi ) = 0 for all i. Now, take the first m equations from system (59) and multiply the first by ω0 , the second by ω1 , etc., and add them to obtain: m 

ws p(ξs ) =

s=0

m 

ωl bl

(62)

l=0

Then take the groups made up by the r to the m + r −1 equations, for r = 1, 2, . . . and apply the same multiplications and sums. The result is the following linear set b0 ω0 + b1 ω1 + · · · + bm−1 ωm−1 = −γm b1 ω0 + b2 ω1 + · · · + bm ωm−1 = −γm+1 .. .. .. .. .. . . . . . bm−1 ω0 + bm ω1 + · · · + b2m−2 ωm−1 = −γ2m−1

(63)

The solution of this system gives the values of the coefficients ωl , l = 0, 1, . . . , m. Substituting them into the definition of the polynomial p(ξ) (Eqs. 60 and 61) yields the value of the roots ξi . Finally, the weights wi can be computed from Eq. (59) which now becomes a linear system. The treatment of correlated variables in this method can be consulted in (Hong 1998). It consists in rotating the basic variable space to a new one in which no correlation exists, using well-known spectral techniques. Let us now examine the limitations of the point estimate method using some simple examples. A first limitation concerns the simple 2n scheme. In fact, as noted in (Christian and Baecher 1998), when the number of input variables is very large the locations of the concentration points may be very far from the mean value thus making the concentration points meaningless from engineering viewpoint. With respect to the 2n + 1 scheme, it has been observed (Hong 1998) that, eventually, the following condition for applying Eq. (52) is not satisfied by some density functions:

γ

γxk ,4 − 3

xk ,3

2

2 >0

(64)

Further, in the 3n plot some roots of the polynomial (56) may be complex, rendering impossible the application of the method. Finally, in the S > 3n scheme, a solution may not exist.

Structural robustness and its relationship to reliability

455

As an example of this latter case, let us examine the application of the above numerical procedure for obtaining the location and weights of m = 4 concentration points for Normal variables. For a single variable x ∼ N(0, σx2 ) the moments are  µx,j =

1 · 3 · · · (n − 1)σxj , j even 0, j odd

(65)

Consequently, the right hand vector in Eq. (63) is [−3 0

− 15 0]T

and the solution of the system is α = [3 0

− 6 0]T

Hence, the locations of the weights are the roots of the polynomial 3 − 6ξ 2 + ξ 4 = 0 which are −2.334, −0.742, 0.742 and 2.334. The weights are calculated by Eq. (59), yielding 0.459, 0.454, 0.454, 0.459. Let us now turn to the case n = 2 for which b0 = 0.5 in Eq. (59). In this case the problem has a solution ξ = [ − 2.715, −1.275, 1.275, 2.715]. However, for n = 3 the matrix of coefficients in Eq. (63) becomes ill-conditioned, so there is no stable solution. And for n = 4 the linear system has a solution but two roots of the polynomial are complex.

4 Robust analysis with point estimates Before describing the linkage between RDO and RBDO approaches, it is useful to make some remarks about the use of point estimates for robust optimization. Notice that the method gives estimates of the ordinary moments. Since robust optimization is oriented towards spread control, it is necessary to compute the variance of the response, given by ! "2 Var[g(x)] = E[g 2 (x)] − E[g(x)]

(66)

Evidently, the minimum of the variance corresponds to the maximum of the mean and the minimum of the mean square. However, since a requirement of the robust optimization is also a minimization of the mean (see Eq. (3)), it results that in using point estimates we eventually have to minimize a weighted cost function of the form C = ω1 E[g(x)] − ω2 E[g(x)] + ω3

!

" E[g 2 (x)]

(67)

456

Structural design optimization considering uncertainties

P

Q 10

12 9

2

14 11

4 1

16 13

6 3

15 [email protected] m

8 5

7

4@1 m

Figure 16.2 Finite element mesh for the numerical example.

where the dependence on the design variables y has been removed for clarity of notation. However, since ω1 and ω3 are both related to spread, they can be made equal, ω1 = ω3 ≡ ω, with the result ! " (68) C = (1 − ω)E[g(x)] + ω E[g 2 (x)] in which the basic requirement ω1 + ω2 + ω3 = 1 has been taken into account.

5 Uniting RDO and RBDO The proposed approach for performing simultaneously a robust and reliability-based design consists in the following steps: (a) To estimate the moments of the response by means of the point estimate method, as this requires a minimal number of solver calls and no other solver than that used for deterministic computations. (b) To estimate the failure probabilities using either the saddlepoint expansion at the critical point. Notice that several responses defining an equal number of limit states can be calculated simultaneously. 5.1 Ex am pl e In this example the method of point estimates will be applied to the estimation of the failure probabilities corresponding to surpassing a threshold by the von Mises yield stress τm in all the finite elements forming the elastic beam shown in Fig. 16.2: gi (x) = τ¯m,i − τm,i (x) = 0, i = 1, 2, . . . , 16

(69)

For a plane problem, the von Mises stress is  τm =

τ12 − τ1 τ2 + τ22 3

where τi , i = 1, 2 are the principal stresses.

(70)

Structural robustness and its relationship to reliability

457

Table 16.1 Random variable definition. Variable

Type

Mean

Standard deviation

P Q E

Lognormal Lognormal Normal

500 50 20,000,000

75 5 3,000,000

Table 16.2 Samples for the 2N scheme. Sample

P

Q

E

1 2 3 4 5 6

648.01 385.99 500.00 500.00 500.00 500.00

50.00 50.00 59.44 42.06 50.00 50.00

2.0e07 2.0e07 2.0e07 2.0e07 3.1374e07 1.7606e07

The elements are constant strain triangles. The beam is subject to two random loads. The elasticity modulus is also random and the Poisson modulus is fixed at 0.2. The stochastic properties of these independent random variables are shown in Table 16.1. A different von Mises stress threshold τ¯m was assigned to each element in order to assure a probability of failure around 10−3 for all elements. In order to have an idea of the estimation errors, 50,000 Monte Carlo simulations were calculated. Er’s method for density estimation mentioned above was also computed for comparison. Notice that the relationship between the limit state function and the input random variables is highly nonlinear. The attempts for calculating three and four concentration points per variable failed in that no real solutions were found for the roots. For this reason the 2n point estimate strategy was applied. Table 16.2 shows the coordinates of the point estimate samples. Only four cumulants were used for the estimation of the failure probabilities. In spite of the reduced number of concentration points and cumulants, the results given by both methods are reasonably good as shown by Table 16.3. The second column of the Table informs the threshold values for each element. Notice that in general the saddlepoint method exhibits better accuracy than Er’s technique. This is especially so for element No. 10, in which case the moment structure of the stress in the element implied that there were three real roots for Eq. (26), at a difference with the rest of elements, for which one real and two complex roots were found. Notice that the saddlepoint method somewhat underestimates the failure probability with respect to Monte Carlo simulation. Figure 16.3 compares a histogram density obtained with a subset of 10,000 Monte Carlo samples and Er’s estimation for element No. 1. Figure 16.4 depicts the standard deviations of von Mises element stresses and the failure probabilities multiplied by 104 for each finite element, as given by Monte Carlo simulation with 50,000 samples. It can be noticed that there is no clear-cut relationship between the two uncertainty

458

Structural design optimization considering uncertainties Table 16.3 Estimates of the failure probabilities. τ¯m,i

Finite element i

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

600 550 750 300 325 700 625 425 425 350 750 950 225 1100 275 450

8

Pf,i Monte Carlo

Pˆ f,i 2N scheme + saddlepoint

Pˆ f,i 2N scheme + Er’s method

(50,000 samples)

(6 samples)

(6 samples)

0.0031 0.0042 0.0049 0.0041 0.0041 0.0063 0.0041 0.0048 0.0053 0.0029 0.0054 0.0033 0.0144 0.0005 0.0121 0.0054

0.0024 0.0034 0.0043 0.0033 0.0034 0.0057 0.0034 0.0041 0.0047 0.0019 0.0047 0.0027 0.0142 0.0004 0.0116 0.0048

0.0056 0.0072 0.0078 0.0069 0.0069 0.0094 0.0069 0.0076 0.0084 0.0107 0.0083 0.0060 0.0180 0.0021 0.0153 0.0084

 103 Er method (4 moments) Monte Carlo (10,000 samples)

7

Probability density

6 5 4 3 2 1 0 200

300

400 500 600 Von Mises stress in element 1

700

800

Figure 16.3 Comparison of Er’s method of density estimation and Monte Carlo histogram.

measures. In fact, in some cases to two similar deviations there correspond rather different probabilities; also, the correlation coefficient between the two measures is rather poor (0.542). Finally, notice that the highest probability there corresponds the lowest standard deviation (element No. 13), which contradicts the naive intuition

Structural robustness and its relationship to reliability

459

s, 104 Pf 180 160

s

140

104 Pf

120 100 80 60 40 20 0

Element No. 0

2

4

6

8

10

12

14

Figure 16.4 Comparison of standard deviation and amplified failure probability for each element.

expressed by Figure 16.1. Similar conclusions arise from the comparison of the failure probability and the coefficient of variation of the von Mises stress. In this case, the correlation is even poorer: −0.109. All this means that optimizing with respect to the statistical moments may yield rather different results than when optimizing with respect to the failure probability and that it is important to consider both kinds of approaches in designing safe structures in the noisy environment of random loads and structural material parameters.

6 Robustness as entropy minimization The main goal of robust design is to control the spread of the structural response. This can be regarded as a minimization of the entropy of the response. A simple illustration of this is given by the fact that the entropy of a Normal density function 

1 (x − µ)2 φx (x) = √ (71) exp − 2σ 2 2πσ increases along with the standard deviation: . √ / Hx = ln σ 2πe

(72)

In order to more rigorously define robustness as entropy minimization, let us consider a set of random variables x = [x1 , x2 , x3 , . . . , xn ] which is transformed to a set z = [z1 , z2 , z3 , . . . , zn ] by functions of the form zj = gj (x)

(73)

460

Structural design optimization considering uncertainties

According to the laws of probability transformation, the density function of vector z is given by (Papoulis 1991) pz (z) =

1 px (x) |J(x)|

(74)

where J(x) is the Jacobian of the transformation:    ∂g1 . . . ∂g1   ∂x1 ∂xn    J(x) =  ... ... ...   ∂gn ∂gn    ∂x . . . ∂x 1

(75)

n

Upon substituting this result in Hz = −

(

pz (z) ln pz (z)dz one obtains (Papoulis 1991)

@ A Hz ≤ Hx + E ln|J(x, y)|

(76)

In this result the dependence of the Jacobian on the vector of design variables y has been made explicit in order to emphasize the relevance of the design in entropy transformation. If the set of equations (73) has a unique inverse, as is the case for linear structures, then equality holds. Equation (76) is useful for understanding why entropy is relevant for defining structural robustness. In fact, assuming that x is the set of input random variables of the structural system and z that of observed random responses, Eq. (76) states that the system is an entropy dissipating system, i.e. it can reduce the scatter of the input variables if @ A E ln|J(x, y)| ≤ 0

(77)

because, in that case, @ A Hz − Hx ≤ E ln|J(x, y)| ≤ 0

(78)

implying Hz ≤ Hx

(79)

Recalling the remark made above about the relativity of the entropy measure of uncertainty in the case of continuous @ Adistributions, our focus will not be the terms Hx and Hz but on the term E ln|J(x, y)| . Let us delve into its expression for the common case of linear structures. 6.1 Rob ustn ess o f linear s t r uc t ur e s Consider a linear structure modeled with the finite element method, so that it is described with the classical equation Ku = q

(80)

Structural robustness and its relationship to reliability

461

where u is the displacement vector, K is the stiffness matrix and q the external force vector, respectively. Assume that the external loads are random. The structure may have some random properties of the materials, but since their dispersion can be controlled with Quality Assurance (QA), we are interested only in reducing the sensitivity of the response to random changes in external loads. By analogy with QA applied in the construction phase, we may call Robustness Assurance (RA) the control of randomness applied in the design phase. It may seem surprising that the RA analysis just proposed ignores the randomness of the material properties and of other structural variables such as geometrical dimensions, etc. The underlying reason for leaving it to the Quality Assurance in the construction phase is that the robustness approach stems from the product-oriented nature of engineering design and construction, that is not interested in establishing the actual risk of the structure, as in the knowledge-oriented reliability approach, but only in assuring a product capable of dissipating the randomness imposed by external actions as much as possible. Thus, the proposed approach to robustness separates design from the determination of the actual reliability, considering this latter as a specialized task whose need arises in certain situations. In addition, notice that the robustness defined in terms of entropy incorporates the available information on external actions in a positive manner, as indicated by the maximum entropy principle (see Eq. 17) and the quotation of Jaynes’ classical chapter above. Thus, beyond the second order analysis on which some proposals for robust design are based, which may invoke the lack of full probabilistic information to proceed in that way, the entropy approach to robustness incorporates such a deficiency as an element of design calculations. In fact, the maximum entropy principle establishes the distribution that accords with several situations of available information. If, for instance, both the mean and variance are prescribed and the variable can be either negative or positive, the principle indicates that the distribution is Gaussian. However, if it is known that the variable is strictly positive and the mean alone is prescribed, the solution is the exponential distribution. And so on. (See (Kapur 1989) for a detailed exposition). Formulating the problem as u = K −1 q,

(81)

a typical element of vector u is ui =



(K−1 )ij qj

(82)

The sensitivity of the i-th response with respect to the j-th random variable is, therefore, ∂ui = (K−1 )ij ∂qj

(83)

which does not depend on any element of vector x. Other responses, such as end forces and stresses can be expressed as other linear combinations of the displacements and,

462

Structural design optimization considering uncertainties

Hz

r r

2

r

3





r

r

2

1

1

y

Figure 16.5 Variation of response entropy of structures as a function of a structural dimension (y) and the coefficient of variation of an external load (ρ).

therefore, of the external loads. For this reason the ensuing development will be done in terms of the displacements. The expectation in Eq. (76) is     @ A (84) E ln|J(x, y)| = ln detK−1 (y) = −ln detK(y) Since K is a positive definite matrix, |detK(y)| = detK(y) > 0 yielding @ A E ln|J(x, y)| = −ln detK(y)

(85)

The term R(y) = ln detK(y)

(86)

will be simply called robustness. Figure 16.5 illustrates the behavior of Eq. (76) for a linear structure as a function of the single cross section dimension subject to design and the coefficient of variation of a single random load. According to this exposition a RA-design of linear structures with random external loads can be involved in a Deterministic Optimization program (see Eq. (1)) as follows: Problem Robustness and Cost Optimization : find :

y

minimizing : Z(y) = −αR(y)/R∗ + (1 − α)C(y)/C ∗ subject to :

fi (y) < Fi , i = 1, 2, . . . y− ≤ y ≤ y+

(87)

Here R∗ and C ∗ are normalizing factors. The meaning of this equation is that cost C(y) is minimized while robustness R(y) is maximized. Thus, the solution will be a saddlepoint instead of the global minimum of the cost. Notice that the robustness term prevents from a one-sided search of the minimum cost in a global manner not

Structural robustness and its relationship to reliability

463

Table 16.4 Comparison of Shannon (entropy) and Lyapounov (stability) functionals. Shannon

Lyapounov

H(U ) is a continuous function of pi If all pi , i = 1, . . . , n are equal, H(U ) is an increasing function of n H(U ) has a unique global maximum

V(x(t)) is a continuous function of t V(x(t)) is a non-increasing function of t V(x(t)) has a unique global minimum

provided by usual behavioral or geometric constraints, some of which could eventually be removed from optimization programs. Factor α, 0 ≤ α ≤ 1, should weight the relative importance of cost and robustness with regard to external loads and, therefore, it must be selected judiciously. The relative weight of robustness should consider the departure of the external loads from determinism, in such a way that the larger their spread, the higher α.

6.2 Analogy to s ys tem dynamics It is interesting to make some remarks on the analogy of entropy dissipation and the theory of dynamic systems and control (See, e.g., (Szidarovszky and Bahill 1992, e.g.)). In fact, one of the basic concerns in dynamic systems is that of controllability, meaning that given an initial state there exists an input capable of leading the system to another state. In dynamic system theory the control force is the result of a trade-off between its cost and the reduction of the responses. Similarly, in robust design under uncertainties the engineer is interested in obtaining a product such that, given an uncertain input, the uncertainty of the response can be controlled to a given value without excessive cost. Another relevant analogy is that with stability. While in the designing dynamical systems the engineer is not interested in particular trajectories of the system but only in assuring its overall stability, in robust design the designer is interested in assuring that the system will not be seriously perturbed by random fluctuations of the input parameters, without detailed stochastic characterizations of the paths of randomness inside the structural model. This is an important, practical reason motivating the development of robust alternatives to the more theoretical, argumentative approach of establishing failure probabilities, according to the exposition made in the Introduction. The theory of stable systems makes use of a Lyapounov functional V(x(t)) of the dynamic system state x(t), whose characteristics have close resemblance to those of the entropy function, as illustrated by Table 16.4. While the Lyapounov functional has its global minimum at the equilibrium state of the system, the entropy functional has its maximum at a density function which “may be asserted for the positive reason that it is uniquely determined as the one which is maximally noncommittal with regard to missing information, instead of the negative one that there was no reason to think otherwise’’ as stated by Jaynes in his classical chapter (Jaynes 1957).

464

Structural design optimization considering uncertainties

6.3 Ex am pl e 1: A s imple b ar in t e ns io n Let us consider the simplest structural model, i.e. an elastic bar of cross section A, length l, elasticity modulus E subject to a tension force P, as shown in Fig. 16.6. Let as assume that the set of random variables is E D x = P, E

(88)

of which it is known that both are positive. On the other hand, the set of responses is D E z = u, τ

(89)

where τ is the random tension in the bar. The single design variable is y=A

(90)

For the transformation Pl EA P τ = A

u =

(91)

the Jacobian is J(x, y) =

Pl A2 E2

(92)

Since all quantities are positive, ln|J(x, y)| = lnJ(x, y) and @ A E ln|J(x, y)| = E[lnP − 2lnE] + lnl − 2lnA

(93)

Thus, the bar is entropy dissipating if 2lnA > E[lnP − 2lnE] + lnl. Now, if A is constrained to lie in the range A1 ≤ A ≤ A2 and no reference is made to cost, the robust design consists simply in assigning A = A2

(94)

since this value minimizes the expectation in Eq. (93). It is evident that no probabilistic information is necessary to arrive to this result. However, without such an information it is not possible to ascertain whether the structure is entropy dissipating or not. 6.4

Ex am pl e 2: A t r us s

Consider the three-bar structure shown in Fig. 16.6. The random variables are x = (P, E), the design variables are y = (A1 , A2 ) and the the observed responses are

Structural robustness and its relationship to reliability

465

E, A

P

l

Figure 16.6 Simple bar in tension.

z = (u, τ) which are the horizontal component of the displacement of load point and the tension in the left bar, respectively. They are given by u = τ =

Pl EA1 A2 +

√

2A1 P √ 2A1 A2 + 2A21

(95)

The Jacobian of the transformation is √ Pl A2 + 2A1 J(x, y) = 2 √ E 2A21 A2 + 2A31

(96)

which has the separable form J(x, y) = Q(x)R(y). Therefore, without reference to cost, a robust design can be obtained with no probabilistic information of P and E by simply finding the values of (A1 , A2 ) that maximize ln

A2 + 2A21 A2

√

2A1 √ + 2A31

(97)

within the specific bounds assigned to each cross section area. If cost is considered, the solution must be a trade-off between cost minimization and robustness maximization. 6.5

Example 3: A clamped beam

Consider finally the clamped beam of variable shape shown in Fig. 16.8. The only random variable is the external load P. The cross section is a square tube of external dimension yi , i = 1, 2 and thickness t, so that the moment of inertia is Ii =

2 3 ty 3 i

(98)

The vertical displacement of the end point is u=

7Pl 3 Pl 3 + 3EI1 3EI2

(99)

466

Structural design optimization considering uncertainties

P

45°

E, A2

E, A1

l

45°

E, A1

Figure 16.7 A simple truss. P E, I1

E, I2

l

l

Figure 16.8 A clamped beam with variable section.

so that l3 J= 2tE

1 7 + 3 3 y1 y2

 (100)

Let us use E = 2100 t/cm2 , t = 1 cm and l = 200 cm. A beam with y1 = 51.2 cm, y2 = 31.5 cm minimizes the cost subject to the constraint that the end displacement is less than or equal to l/250 (Hernández 1990). This solution is entropy dissipating since ln |J| < 0. On the contrary, for y1 = 30 cm, y2 = 15 cm the structure increases entropy.

7 Conclusions The following conclusions stem from the research reported in this chapter: •

The concept of entropy is useful for clarifying the meaning of robustness in structural systems. In fact, the entropy of the structural responses decreases as the stiffness increases. Accordingly, in parallel to Quality Assurance operating on the spread control of structural material properties in the construction phase, one may define Robustness Assurance as the control of the entropy of response variables such as displacements and stresses due to the uncertainty in random external loads in the design phase. Such Robustness Assurance can easily be incorporated

Structural robustness and its relationship to reliability

•

•

•

•

467

into conventional Deterministic Optimization programs. A proposal in this regard has been exposed. It has also been shown that structural robustness defined in these terms exhibits similarities to the theory of controllability and stability studied with the assistance of Lyapounov functions in the context of dynamic systems. Both subjects are instances of the production-oriented approach that is characteristic of engineering design process at a difference to the knowledge-oriented process of scientific discovery. For optimizing a structure under uncertainty both robust and reliability-based approaches are valuable and therefore complementary. The first aims at a control of response spread whereas the second to reducing the probability of extreme undesirable situations. For the above reason, methods allowing simultaneous monitoring of the basic statistical quantities implied by both approaches (namely statistical moments and failure probabilities) are of importance. In this chapter, this has been sought by means of the method of saddlepoint local expansion of the density function about the critical threshold and the method of maximum entropy. While the former exhibits better accuracy for estimating the failure probability, the latter is highly useful for the assessment of the degree of robustness of the structural system as commented above. The method of point estimates allows a fast and simple estimation of response moments using the finite element solver employed for deterministic calculations. For these reasons it is highly for both robust and reliability-based design optimization.

Further research is needed to develop the entropy approach to structural robustness as well as on computational methods for the complex task of structural optimization granting the accomplishment of both reliability and robustness requirements.

References Abramowitz, M. & Stegun, I.A. 1972. Handbook of mathematical functions. New York: Dover Publications. Agmon, N., Alhassid, Y. & Levine, R.D. 1979. An algorithm for finding the distribution of maximal entropy. Journal of Computational Physics 30:250–258. Au, S.K. & Beck, J.L. 2001. Estimation of small failure probabilites in high dimensions by subset simulation. Probabilistic Engineering Mechanics 16:263–277. Barndorff-Nielsen, O. & Cox, D.R. 1979. Edgeworth and saddle-point approximations with statistical applications. Journal of the Royal Statistical Society 41:279–312. Ben-Haim, Y. 1985. The Assay of Spatially Random Material. Dordrecht: D. Reidel Publishing Company. Ben-Haim, Y. 1996. Robust Reliability in the Mechanical Sciences. Berlin: Springer-Verlag. Cheah, P.K., Fraser, D.A.S. & Reid, N. 1993. Some alternatives to edgeworth. Canadian Journal of Statistics 21:131–138. Chernousko, F.L. 1999. What is ellipsoidal modelling and how to use it for control and state estimation? In I. Elishakoff (ed.), Whys and Hows in Uncertainty Modelling, pp. 127–188. Wien: Springer-Verlag.

468

Structural design optimization considering uncertainties

Ching, J. & Hsieh, Y.H. 2007. Local estimation of failure probability function and its confidence interval with maximum entropy principle. Probabilistic Engineering Mechanics 22: 39–49. Christian, J.T. & Baecher, G.B. 1998. Point-estimate method and numerical quadrature. Journal of Geotechnical and Geoenvironmental Engineering 125:779–786. Daniels, H.E. 1954. Saddlepoint approximations in statistics. Annals of Mathematical Statistics 25:631–650. Doltsinis, I. & Kang, A. 2004. Robust design of structures using optimization methods. Computer Methods in Applied Mechanics and Engineering 193:2221–2237. Doltsinis, I., Kang, A. & Cheng, G. 2005. Robust design of non-linear structures using optimization methods. Computer Methods in Applied Mechanics and Engineering 194: 1779–1795. Dubois, D. & Prade, H. 1988. Possibility Theory. New York: Plenum Press. Elishakoff, I. 1991. Essay on reliability index, probabilistic interpetation of safety factor and convex models of uncertainty. In F. Casciati & J.B. Roberts (eds), Reliability Problems: General principles and Applications in Mechanics of Solids and Structures, pp. 237–271. Wien: Springer-Verlag. Elishakoff, I. 1999. Are probabilistic and anti-optimization approaches compatible? In I. Elishakoff (ed.), Whys and Hows in Uncertainty Modelling, pp. 263–355. Wien: Springer-Verlag. Elishakoff, I. 2005. Safety Factors and Reliability: Friends or Foes? New York: Kluwer. Elishakoff, I. & Ren, Y. 2003. Finite Element Methods for Structures with Large Stochastic Variations. Oxford: Oxford University Press. Er, G.K. 1998. A method for multi-parameter PDF estimation of random variables. Structural Safety 20:25–36. Frangopol, D.M. 1995. Reliability-based structural design. In C.R. Sundararajan (ed.), Probabilistic Structural Mechanics Handbook, pp. 352–387. New York: Chapman & Hall. Gasser, M. & Schuëller, G.I. 1997. Reliability-based optimization of structural systems. Mathematical Methods of Operations Research 46:287–307. Guan, X.L. & Melchers, R. 2001. Effect of response surface parameter variation on structural reliability estimates. Structural Safety 23:429–444. Haftka, R.T., Gurdal, Z. & Kamat, M.P. 1990. Elements of Structural Optimization. Dordrecht: Kluwer Academic Publishers. Hamming, R.W. 1973. Numerical Methods for Scientists and Engineers. New York: Dover Publications. Hansen, E. & Walster, G.W. 2004. Global Optimization using Interval Analysis. New York: Marcel Dekker, Inc. Harr, M. 1989. Probabilistic estimates for multivariate analysis. Applied Mathematical Modelling 13:313–318. Hernández, S. 1990. Métodos de diseño óptimo de estructuras. Madrid: Colegio de Ingenieros de Caminos, Canales y Puertos. Hisada, T. & Nakagiri, S. 1981. Stochastic finite element method developed for structural safety and reliability. In Proceedings of the Third International Conference on Structural Safety and Reliability, pp. 395–408. Rotterdam: Elsevier. Hong, H.P. 1998. An efficient point estimate method for probabilistic analysis. Reliability Engineering and System Safety 59:261–267. Hong, H.P., Escobar, J.A. & Gómez, R. 1998. Probabilistic assessment of the in seismic response of structural asymmetric models. In Proceedings of the Tenth European Conference on Earthquake Engineering, Paris, 1998, Rotterdam. Balkema. Hurtado, J.E. 2001. Neural networks in stochastic mechanics. Archives of Computational Methods in Engineering 8:303–342.

Structural robustness and its relationship to reliability

469

Hurtado, J.E. 2004a. An examination of methods for approximating implicit limit state functions from the viewpoint of statistical learning theory. Structural Safety 26:271–293. Hurtado, J.E. 2004b. Structural Reliability. Statistical Learning Perspectives. Heidelberg: Springer. Hurtado, J.E. 2006. Optimal reliability-based design using support vector machines and artificial life algorithms. In Y. Tsompanakis & N.D. Lagaros (eds), Intelligent Computational Paradigms in Earthquake Engineering. Hershey: Idea Group Inc. Hurtado, J.E. 2007. Filtered importance sampling with support vector margin: a powerful method for structural reliability analysis. Structural Safety 29:2–15. Hurtado, J.E. & Alvarez, D.A. 2001. Neural network-based reliability analysis: A comparative study. Computer Methods in Applied Mechanics and Engineering 191: 113–132. Hurtado, J.E. & Barbat, A. 1998. Fourier-based maximum entropy method in stochastic dynamics. Structural Safety 20:221–235. Jaynes, E.T. 1957. Information Theory and Statistical Mechanics. The Physical Review 106: 620–630. Johnson, N.L., Kotz, S. & Balakrishnan, N. 1994. Continuous Univariate Distributions, Vol. 1. New York: John Wiley and Sons. Kapur, J.N. 1989. Maximum Entropy Models in Science and Engineering. New York: John Wiley and Sons. Kennedy, C.A. & Lennox, W.C. 2000. Solution to the practical problem of moments using nonclassical orthogonal polynomials with applications for probabilistic analysis. Probabilistic Engineering Mechanics 15:371–379. Kennedy, C.A. & Lennox, W.C. 2001. Moment operations on random variables, with applications for probabilistic analysis. Probabilistic Engineering Mechanics 16:253–259. Kharitonov, V. 1997. Interval uncertainty structure: Conservative but simple. In H. GüntherNaske & Y. Ben-Haim (eds), Uncertainty: Models and Measures, pp. 231–243. Berlin: Akademie Verlag. Kirsch, U. 1993. Structural Optimization. Fundamentals and Applications. Heidelberg: Springer Verlag. Kleiber, M. & Hien, T.D. 1992. The Stochastic Finite Element Method. Chichester: John Wiley and Sons. Klir, G.J. 1997. Uncertainty theories, models and principles: An overview of personal views and contributions. In H. Günther-Natke & Y. Ben-Haim (eds), Uncertainty: Models and Measures, pp. 27–43. Berlin: Akademie Verlag. Kolassa, J.E. 1997. Series Approximation Methods in Statistics. New York: Springer Verlag. Kosko, B. 1992. Neural Networks and Fuzzy Systems. Englewood Cliffs: Prentice Hall. Lagaros, N., Papadrakakis, M. & Kokossalakis, G. 2002. Structural optimization using evolutionary algorithms. Computers and Structures 80:571–579. Lagaros, N. & Papadrakakis, M. 2003. Soft computing methodologies for structural optimization. Applied Soft Computing 3:283–300. Lange, K. 1999. Numerical Analysis for Statisticians. New York: Springer Verlag. Liu, W.K., Belytschko, T. & Lua, Y.J. 1995. Probabilistic finite element method. In C.R. Sundararajan (ed.), Probabilistic Structural Mechanics Handbook, pp. 70–105. New York: Chapman & Hall. Lugannani, R. & Rice, S. 1980. Saddle point approximation for the distribution of sums of random variables. Advances in Applied Probability 12:475–490. Mead, L.R. & Papanicolau, N. 1984. Maximum entropy in the problem of moments. Journal of Mathematical Physics 25:2404–2417. Miller, A.C. & Rice, T.R. 1983. Discrete approximations of probability distributions. Management Science 29:352–362.

470

Structural design optimization considering uncertainties

Muscolino, G. 1993. Response of linear and non-linear structural systems under gaussian or non-gaussian filtered input. In F. Casciati (ed.), Dynamic Motion: Chaotic and Stochastic Behaviour, pp. 203–299. Wien: Springer-Verlag. Ordaz, M. 1988. On the use of probability concentrations. Structural Safety 5:317–318. Pandey, M.D. & Ariaratman, S.T. 1996. Crossing rate analysis of non gaussian response of linear systems. Journal of Engineering Mechanics 122:507–511. Papadrakakis, M., Lagaros, N. & Tsompanakis, Y. 1998. Structural optimization using evolution strategies and neural networks. Computer Methods in Applied Mechanics and Engineering 156:309–333. Papadrakakis, M., Papadopoulos, V. & Lagaros, N. 1996. Structural reliability analysis of elastic-plastic structures using neural networks and Monte Carlo simulation. Computer Methods in Applied Mechanics and Engineering 136:145–163. Papoulis, A. 1991. Probability, Random Variables and Stochastic Processes. New York: McGraw-Hill. Reid, N. 1988. Saddlepoint methods and statistical inference. Statistical Science 3:213–238. Robert, C.P. & Casella, G. 1999. Monte Carlo Statistical Methods. New York: Springer. Robinson, J. 1982. Saddlepoint approximations for permutation tests and confidence intervals. Journal of the Royal Statistical Society Series B 44:91–101. Rosenblueth, E. 1975. Point estimates for probability moments. Proceedings of the National Academy of Sciences of the USA 72:3812–3814. Rosenblueth, E. & Mendoza, E. 1971. Reliability optimization in isostatic structures. Journal of the Engineering Mechanics Division ASCE 97:1625–1640. Royset, J.O., Kiureghian, A.D. & Polak, E. 2001. Reliability-based optimal structural design by the decoupling approach. Reliability Engineering and System Safety 73:213–221. Royset, J.O. & Polak, E. 2004. Reliability-based optimal design using sample average approximations. Probabilistic Engineering Mechanics 19:331–343. Schuëller, G.I. & Stix, R. 1987. A critical appraisal of methods to determine failure probabilities. Structural Safety 4:293–309. Sexsmith, R.G. 1999. Probability-based safety analysis–value and drawbacks. Structural Safety 21:303–310. Shannon, C.E. 1948. A Mathematical Theory of Communication. The Bell System Technical Journal 27:379–423. Shore, J.E. & Johnson, R.W. 1980. Axiomatic derivation of the principle of maximum entropy and the principle of minimum cross-entropy. IEEE Transactions on Information Theory 26(1):26–37. Sobczyk, K. & Tr¸ebicki, J. 1990. Maximum entropy principle in stochastic dynamics. Probabilistic Engineering Mechanics 5:102–110. Szidarovszky, F. & Bahill, A.T. 1992. Linear Systems Theory. Boca Ratón: CRC Press. Tr¸ebicki, J. & Sobczyk, K. 1996. Maximum entropy principle and non-stationary distributions of stochastic systems. Probabilistic Engineering Mechanics 11:169–178.

Chapter 17

Maximum robustness design of trusses via semidefinite programming Yoshihiro Kanno University of Tokyo, Tokyo, Japan

Izuru Takewaki Kyoto University, Kyoto, Japan

ABSTRACT: This chapter discusses evaluation and maximization of the robustness function of trusses, which is regarded as one of measures of structural robustness under the uncertainties of member stiffnesses and external forces. By using quadratic embedding of the uncertainty and the S-procedure, we formulate a quasiconvex optimization problem which provides a lower bound of the robustness function. We next formulate the maximization problem of the robustness function as a robust structural optimization scheme. An algorithm based on the semidefinite program is proposed to obtain the optimal truss design. Numerical examples are shown to demonstrate the validity of the algorithms presented.

1 Introduction Recently, the info-gap decision theory has been proposed as a non-probabilistic decision theory under uncertainties (Ben-Haim 2006), and has been applied to wide fields including neural networks (Pierce et al. 2006), biological conservation (Moilanen & Wintle 2006), financial economics (Ben-Haim 2005), etc. In the info-gap decision theory, the robustness function plays a key role as a measure of robustness of systems having uncertainties (Ben-Haim 2006). The robustness function is regarded to represent the immunity against failure, and is defined as the greatest level of uncertainty at which any failure cannot occur. In structural engineering, the robustness function represents the greatest level of uncertainty, caused by manufacture errors, limitation of knowledge of input disturbance, observation errors, etc., at which any constraint on mechanical performance cannot be violated. The constraints on mechanical performance can be violated only at great level of uncertainty in a structure with a large robustness function, while they can be violated at small level of uncertainty in a structure with a small robustness function. Thus, we can compare robustness of structures quantitatively in terms of robustness functions. Takewaki & Ben-Haim (2005) computed the robustness function of structures in a particular case where the worst case can be obtained analytically. Unfortunately it is difficult to compute exactly the robustness function of structures exactly, and no efficient method has ever been proposed to the authors’ knowledge. The first contribution of the work in this chapter is to propose a numerically tractable optimization problem to obtain a lower bound of the robustness function of trusses considering various constraints and circumstances of uncertainties. The solution of the problem presented can

472

Structural design optimization considering uncertainties

be obtained efficiently by solving some semidefinite programming (SDP) (Wolkowicz et al. 2000) problems. Note that a lower bound is regarded as a conservative estimate of the robustness function, i.e. a level of uncertainty at which the satisfaction of the constraints on mechanical performance is guaranteed. Hence, finding a lower bound, not an upper bound, is meaningful when it is difficult to find the exact value of the robustness function. Secondly, we consider a structural optimization problem in which we seek for the truss design maximizing the robustness function. Based on the stochastic uncertainty model of mechanical parameters, various methods were proposed for reliability-based optimization (see the other chapters and the references therein). One of the motivations for the info-gap theory is an awareness of the limitations of the probabilistic approaches as discussed by Ben-Haim (2004, Section 7). As a non-probabilistic uncertainty model, Ben-Haim & Elishakoff (1990) developed the so-called convex model, where the uncertainty of a system is expressed in terms of unknown-but-bounded parameters. Pantelides & Ganzerli (1998) proposed a robust truss optimization method based on the convex model. The mathematical programming problems including uncertain data have also been investigated extensively. For various classes of convex optimization problems, a unified methodology of robust counterpart was developed by Ben-Tal & Nemirovski (2002), where the data in optimization problems are assumed to be unknown but bounded. Calafiore & El Ghaoui (2004) proposed a method for finding the ellipsoidal bounds of the solution set of uncertain linear equations by using SDP. In this chapter, we deal with the robustness function of trusses that consist of members with uncertain stiffness and/or are subjected to uncertain external forces. The non-probabilistic uncertain parameters are assumed not to be known precisely but to be bounded. The details of background of decision strategies based on the info-gap theory may be consulted to the basic textbook (Ben-Haim 2006). To overcome the difficulty of computing the robustness function, we utilize the framework of SDP. For mathematical backgrounds and algorithms for SDP, the readers may refer to the review chapters (Helmberg 2002; Vandenberghe & Boyd 1996) and the handbook (Wolkowicz et al. 2000).

2 Preliminary results Some useful technical results used in this chapter are listed in Appendix A. Throughout this chapter, all vectors are assumed to be column vectors. However, for vectors u ∈ Rn and v ∈ Rm , we often simplify the notation (uT , v T )T as (u, v). The standard Euclidean norm (pT p)1/2 of a vector p ∈ Rn is denoted by p2 . The l∞ -norm of p, denoted by p∞ , is defined as p∞ = maxi∈{1,...,n} |pi |. Define Rn+ ⊂ Rn by Rn+ = {p ∈ Rn | p ≥ 0} For p = (pi ) ∈ Rn and q = (qi ) ∈ Rn , we write p ≥ 0 and p ≥ q, respectively, if p ∈ Rn+ and pi ≥ qi (i = 1, . . . , n). Let S n ⊂ Rn×n denote the set of all n × n real symmetric matrices. We write A  O if A ∈ S n is positive semidefinite, i.e. if all the eigenvalues of the matrix A are nonnegative.

Maximum robustness design of trusses via semidefinite programming

473

For A ∈ S n and B ∈ S n , we write A  B if the matrix A − B is positive semidefinite. The Moore–Penrose pseudo-inverse of C ∈ Rm×n is denoted by C† ∈ Rn×m . 2.1

Semidefinite program

The semidefinite program (SDP) is classified as a convex and nonlinear mathematical program. The SDP problem refers to the optimization problem having the form of (Wolkowica et al. 2000)   m  T max b y: C − A i yi  O (1) i=1

Here y ∈ Rm is a variable vector, b ∈ Rm is a constant vector, and Ai ∈ S n (i = 1, . . . , m) and C ∈ S n are constant symmetric matrices. Recently, SDP has received increasing attention for its wide fields of application (Ohsaki et al. 1999; Ben-Tal & Nemirovski 2001). It is well known that the linear program and the second-order cone program are included in SDP as particular cases. The primal-dual interior-point method, which has been first developed for LP, has been naturally extended to SDP. It is theoretically guaranteed that the primal-dual interiorpoint method converges to the global optimal solution of the SDP problem (1) within the number of arithmetic operations bounded by a polynomial of m and n (Ben-Tal & Nemirovski 2001; Wolkowicz et al. 2000). 2.2

Quasic onvex optimization problem

The α-sublevel set of a function f : Rn → R is defined as Lf (α) = {x ∈ Rn | f (x) ≤ α} A function f is called quasiconvex if its domain and all its sublevel sets Lf (α) for α ∈ R are convex. Let f0 : Rn → R be quasiconvex, and let f1 , . . . , fm : Rn → R be convex. The quasiconvex optimization problem refers to the optimization problem having the form of (Boyd & Vandenberghe 2004, Section 4.2.5) min{f0 (x): fi (x) ≤ 0 (i = 1, . . . , m), Ax = b}

(2)

where A ∈ Rm×n and b ∈ Rm . The difference between convex and quasiconvex optimization problems is that a quasiconvex optimization problem can have locally optimal solutions that are not globally optimal. It is known that the global optimal solution of a quasiconvex optimization problem can be obtained by using the bisection method in which some convex optimization problems are solved (Boyd & Vandenberghe 2004).

3 Modeling uncertain trusses and mechanical constraints Consider a linear elastic truss in the two- or three-dimensional space. Small rotations and small strains are assumed. Letting nd denote the number of degrees of freedom

474

Structural design optimization considering uncertainties d

d

of displacements, u ∈ Rn and f ∈ Rn denote the vectors of nodal displacements and external forces, respectively. The system of equilibrium equations can be written as Ku = f

(3) d

where K ∈ S n denotes the stiffness matrix of the truss. In (3) we explicitly deal with the model- and data-uncertainties of K and f , that shall be rigorously defined below. m Let a = (ai ) ∈ Rn denote the vector of member cross-sectional areas, where nm denotes the number of members. For trusses, the stiffness matrix K is a function of a, and can be decomposed as m

K(a) =

n 

m

n 

ai K i =

i=1

ai bi bTi

(4)

i=1

d

d

where K i ∈ S n and bi = (bij ) ∈ Rn (i = 1, . . . , nm ) are constant matrices and constant vectors, respectively. 3.1

U n c erta i n t y mo d el

Assume that the uncertainty of K is caused only by the uncertainties of stiffness of members, while the locations of nodes are assumed to be certain. We represent the uncertainties of stiffness of members through the uncertainties of member cross-sectional areas a. m d f = (F fj ) ∈ Rn denote the nominal values (or the best estimates) Let F a = (F ai ) ∈ Rn and F m d of a and f , respectively. Let ζ a = (ζai ) ∈ Rn and ζ f = (ζfj ) ∈ Rn denote the parameter vectors that are considered to be unknown but bounded. We describe the uncertainties of a and f by using ζ a and ζ f , respectively. Suppose that a and f depend on ζ a and ζ f affinely, i.e. ai + a0i ζai , ai = F

i = 1, . . . , nm

(5)

fj + f 0 ζfj , fj = F

j = 1, . . . , nd

(6)

m

Here, a0 = (a0i ) ∈ Rn+ and f 0 ∈ R+ are constant coefficients satisfying F ai > a0i 0 m 0 (i = 1, . . . , n ). Note that ai and f represent the relative magnitude of uncertainties of ai and f , respectively. Moreover, a0 and f 0 make ζ a and ζ f have no dimensions. d For p = 1, . . . , nt , let mp ∈ {1, . . . , nd } and let T p ∈ Rmp ×n be a constant matrix. For m d a fixed α ∈ R+ , define two sets Za (α) ⊂ Rn and Zf (α) ⊂ Rn by m

Za (α) = {ζ a ∈ Rn | α ≥ ζ a ∞ } nd

(7)

Zf (α) = {ζ f ∈ R | α ≥ T p ζ f 2 (p = 1, . . . , n )} t

(8)

Here, we choose T 1 , . . . , T nt so that Zf (α) becomes bounded for any α ∈ R+ . It is obvious that Za (α) is bounded. Since a truss is an assemblage of nodes connected by some independent members, the perturbation of stiffness of a member from its nominal value does not affect those

Maximum robustness design of trusses via semidefinite programming

475

of the other members. Hence, in (7) we choose the l∞ -norm which represents the independent uncertainties of scalars ζa1 , . . . , ζanm . On the other hand, the definition (8) permits us to suppose that there exist correlation among some components of ζ f by choosing T p appropriately. Moreover, these matrices allow to represent the difference of magnitudes of uncertainties among some components of ζ f . For examples of T p , see Example 3.1 and Section 6.3. The uncertain parameters ζ a and ζ f in (5) and (6), respectively, are assumed to be running through the uncertain sets Za (α) and Zf (α) defined by (7) and (8), i.e. ζ a ∈ Za (α),

ζ f ∈ Zf (α)

(9)

For simplicity, we often write ζ = (ζ Ta , ζ Tf )T ,

Z(α) = Za (α) × Zf (α)

so that (9) is simplified as ζ ∈ Z(α) Roughly speaking, ζ a and ζ f perturb around the origin with the “width’’ of α. Then a and f , respectively, vary around the center-points F a and F f . The greater the value of α, the greater the range of possible variations of a and f , and hence α is called the uncertainty parameter (Ben-Haim 2004). Note that the value of α is usually unknown in structures actually built. Throughout the following robustness analysis based on the info-gap theory, we do not use any knowledge of the actual range of uncertainty of a truss, that is regarded as one of advantages of using the robustness function. It is easy to check that the uncertainty model of a and f defined by (5)–(8) obey a, and F f , let the info-gap model (Ben-Haim 2006) of uncertainty. For given α ∈ R+ , F m d T A(α,F a, F f ) ⊆ Rn × Rn be the set of all vectors (aT , f )T satisfying (5)–(8). Then A(α) satisfies the two basic axioms of the info-gap model: (i) (ii)

Nesting: 0 ≤ α1 < α2 implies A(α1 ,F a, F f ) ⊂ A(α2 ,F a, F f ); F Contraction: the info-gap model A(0,F a, f ) coincides with a singleton set T containing its center point, i.e. A(0,F a, F f ) = {(F a T,F f )T }.

From the nesting axiom we see that the uncertainty set A(α,F a, F f ) becomes more inclusive as α becomes larger. The contraction axiom guarantees that the estimates F a and F f are correct at α = 0. Example 3.1 (interval uncertainty of external load). The interval uncertainty model of the external load f is conventionally used in the so-called interval analysis of uncertain structures; see, e.g., Chen et al. (2002). We show in this example that the uncertainty model of f defined by (6), (8), and (9) includes the interval uncertainty

476

Structural design optimization considering uncertainties d

model as a particular case. For each p = 1, . . . , nd , let ep ∈ Rn denote the pth row d vector of the identity matrix I ∈ S n , and let δp be a positive constant. Then, by putting Tp =

1 T e , δp p

p = 1, . . . , nd

with m1 = · · · = mnt = 1 and nt = nd , Zf (α) defined by (8) is reduced to       ζfj  nd  d   Zf (α) = ζ f ∈ R  α ≥   , j = 1, . . . , n δj

(10)

Consequently, the uncertainty of f obeying (6), (9), and (10) can be alternatively written as fj ∈ [F fj − αf 0 δj , F fj + αf 0 δj ],

j = 1, . . . , nd

which coincides with the conventional interval uncertainty model. 3.2 C o nstrai n t s o n mec hanic al per fo r m a n ce Consider the mechanical performance of trusses that can be expressed by the cond d straints in terms of displacements. Let Ql ∈ S n , ql ∈ Rn , and γl ∈ R. Suppose that the constraints on mechanical performance can be written in the following quadratic inequalities in terms of u: uT Ql u + 2qTl u + γl ≤ 0,

l = 1, . . . , nc

(11)

where nc denotes the number of constraints. Suppose that Ql , ql , and γl are functions r of r c ∈ Rn . Here, r c is regarded as the vector of parameters representing the level of r d+1 performance, and nr denotes the number of these parameters. Define H l : Rn → S n by ) * c c Q (r ) q (r ) l l H l (r c ) = − ql (r c )T γl (r c ) r

d

For a given vector r c ∈ Rn , define a set F ⊆ Rn as    

  T u c nd  u c c F(r ) = u ∈ R  H l (r ) ≥ 0 (l = 1, . . . , n ) 1  1

(12)

Then the constraint (11) is equivalently rewritten as u ∈ F(r c )

(13)

Note that we have restricted ourselves to cases in which the constraints on the truss can be represented by a finite number of quadratic inequalities. However, there exist various constraints that can be described via (12) and (13) from a practical point of

Maximum robustness design of trusses via semidefinite programming

477

view, because it is known that any single polynomial inequality can be converted into a system of (a finite number of) quadratic inequalities (Kojima & Tunçel 2000). Example 3.2 (stress constraints). We show the explicit reformulation of the stress constraints into (13). Let σi (u) denote the stress of the ith member compatible with u, m and let σ c = (σic ) ∈ Rn+ . Then the stress constraints may be written in the form of |σi (u)| ≤ σic ,

i = 1, . . . , nm

(14)

Here, we assume for simplicity that the lower and the upper bounds of stress of each member have the common absolute value σic . Let E denote the elastic modulus of truss members; let i denote the initial unstressed length of the ith member. From (4) we see  E T σi (u) = b u (15) i i From (15) it follows that (14) is equivalently rewritten as u ∈ F(σ c ) with        T u −(E/i )bi bTi 0 c nd  u m F(σ ) = u ∈ R  ≥ 0 (i = 1, . . . , n ) 1 0T −(σic )2  1 Thus, the stress constraints (14) can be embedded into the form of (13) with nc = nm . The parameters σ c determine the level of performance required. Hence, we have r c = σ c with nr = nm in (13).

4 Definition of robustness function for truss In this section, we show that the robustness function (Ben-Haim 2006) of trusses is obtained as the optimal objective value of a mathematical programming problem with infinitely many constraint conditions. d F = K(F For simplicity, we often write K a). By introducing auxiliary variables η ∈ Rn and from (5) and (6), the system (3) of uncertain equilibrium equations is reduced to m

F + Ku

n 

a0i ζai K i u = η,

ζ a ∈ Za (α)

(16)

i=1

F f + f 0 ζ f = η,

ζ f ∈ Zf (α)

(17) d

a) ⊂ Rn denote the set of all possible solutions to (16) and For a given α ∈ R+ , let U(α,F (17), that is defined by B C d (18) U(α,F a) = u ∈ Rn  (16), (17) Recall that the (nominal) constraint has been introduced in (13). We next consider the robust counterpart of (13). Let α ∈ R+ be fixed. Since the equilibrium equations (16) and (17) include the unknown parameters ζ = (ζ a , ζ t ), the nodal displacement u is regarded as a function of ζ, namely, we may write u(ζ) for ζ ∈ Z(α). To define the robustness function, we require that the constraint (13) should be satisfied by

478

Structural design optimization considering uncertainties

all possible realization of u(ζ) when ζ takes any vector satisfying ζ ∈ Z(α). This requirement can be written as u(ζ) ∈ F(r c ),

∀ζ ∈ Z(α)

(19)

By using the set U introduced in (18), the condition (19) is equivalently rewritten as u ∈ F(r c ),

∀u ∈ U(α,F a)

(20)

m

r

For a given F a ∈ Rn and r c ∈ Rn , the robustness function G α(F a, r c ) represents the largest α with which the robust constraint (20) is satisfied. Rigorously, the robustness m r function G α: Rn × Rn → (−∞, +∞] associated with the constraints (11) is defined by (Ben-Haim 2006, Chapter 3)  ∗ α , if Problem (22) is feasible G α(F a, r c ) = (21) 0, if Problem (22) is infeasible where α∗ = max{α: u ∈ F(r c ), ∀u ∈ U(α,F a)}

(22)

Problem (22) is classified to the semi-infinite programming. By semi-infinite we mean an optimization problem having a finite number of scalar variables and infinitely many inequality constraints. Note that α∗ defined by (22) depends on the level r c of constraints on mechanical performance as well as the nominal cross-sectional areas F a. Throughout the chapter, we assume U(0,F a) ⊆ F(r c ) for simplicity, and hence α or G α(F a). Problem (22) is feasible. In what follows, G α(F a, r c ) is often abbreviated by G m m a2 ∈ Rn , we say For the two different vectors of design variables F a1 ∈ Rn and F a2 if G α(F a1 , r c ) >G α(F a2 , r c ). Let ζ 1 ∈ Z(G α). If there exists an that F a1 is more robust than F c l ∈ {1, . . . , n } such that (11) becomes active at a given ζ 1 , then we say that ζ 1 is the worst case. Note that there exists typically more than a single worst case. Especially, optimum truss designs maximizing the robustness function or for specified robustness function often have many worst cases, as will be illustrated in Section 8.2. a), and G α with Figure 17.1 illustrates the schematic relations among F(r c ), U(α,F various values of α. Here, Figure 17.1(a) and 17.1(b), respectively, correspond to α and αb =G α, where we see that the constraint u ∈ F(uc ) is satisfied for all possible αa
u2

u2

(aa)

Worst case u2 (ab)

u1 ^ (a) aa  a

(ac)

u1 ^ (b) ab
a

Figure 17.1 Relation among F (r c ), U (α,F a), and G α for various α.

u1 (c) ac  a^

Maximum robustness design of trusses via semidefinite programming

479

F(r c ) in Figure 17.1(b). It is observed in Figure 17.1(c) that some solutions u ∈ U(αc ,F a) α. to the equilibrium equations violate the constraint u ∈ F(r c ), which implies αc >G

5 Illustrative example of robustness analysis As an illustrative example, consider a two-bar truss shown in Figure 17.2. The nodes (b) and (c) are pin-supported at (x, y) = (0, 100.0) and (0, 0) in cm, respectively, while the node (a) is free, i.e.√nd = nm = 2. The lengths of members (1) and (2), respectively, are 100.0 cm and 100 2 cm. The elastic modulus of each member is 200 GPa. Let f = (f1 , f2 )T denote the external force vector applied at the node (a). The nominal value F f of f is given as F f = (1000.0, 0)T kN The vector of nominal cross-sectional areas is denoted by F a = (F a1 ,F a2 ), and is given by F a = (20.0, 30.0)T cm2 Consider the uncertainty model introduced in section 3.1. In accordance with (5) and (6), define the uncertainties of a and f as ai = F ai + a0i ζai ,

i = 1, 2;

ζ a ∈ Za (α)

(23)

fj + f 0 ζfj , fj = F

j = 1, 2;

ζ f ∈ Zf (α)

(24)

where the coefficients of uncertainty are a0i = 5.0 cm2 ,

i = 1, 2;

f 0 = 200.0 kN

(25)

For a given α, the uncertain sets Za (α) and Zf (α) are defined as Za (α) = {ζ a ∈ R2 | α ≥ |ζai |, i = 1, 2}

(26)

Zf (α) = {ζ f ∈ R2 | α ≥ ζ f 2 }

(27)

y

u2, f2 (1) (a)

(b)

u1, f1

(2)

0 (c)

Figure 17.2 2-bar truss.

x

480

Structural design optimization considering uncertainties

Here, we have put nt = 1, T 1 = I, and m1 = 2 in (8). For simplicity, we often write ζ ∈ Z(α) if ζ a ∈ Za (α) and ζ f ∈ Zf (α). Let σ1 and σ2 denote the stresses of members (1) and (2), respectively. Consider the stress constraints of all members defined by (14) with σ1c = σ2c = 1.0 GPa, i.e. the conditions |σi (u)| ≤ σic ,

i = 1, 2

(28)

should be satisfied for any ζ ∈ Z(α). As an example, putting α = 1.0, we randomly generate a number of ζ satisfying ζ ∈ Z(α) with (23) and (24). The corresponding generated a and f defined by (23) and (24) are shown in Figures 17.3 and 17.4, respectively.

36 34

a2 (cm2)

32 30 28 26 24

14

16

18

20 22 a1 (cm2)

24

26

Figure 17.3 The cross-sectional areas a for randomly generated ζ a ∈ Za (α) with α = 1.0.

200

f2 (kN)

100

0

100 200 700

800

900

1000

1100

1200

1300

f1 (kN)

Figure 17.4 The external forces f for randomly generated ζ f ∈ Zf (α) with α = 1.0.

Maximum robustness design of trusses via semidefinite programming

481

The axial forces q1 and q2 of the members (1) and (2), respectively, are written as q1 = f1 − f2 ,

q2 =

√ 2f2

(29)

Note that q1 and q2 are independent of a, because the truss is statitiscally determinate. From (24), (27), and (29), the maximum value of q1 under the uncertain external force f is obtained as max{q1 (ζ): ζ ∈ Z(α)} = F f1 +

√

2f 0 α

(30)

The minimum value of q1 and the maximum and minimum values of q2 are obtained similarly. Figure 17.5 depicts the variations of (q1 , q2 ) for randomly generated ζ ∈ Z(α) with α = 1.0, and Figure 17.6 shows the corresponding variation of (u1 , u2 ). Figure 17.7 shows the stress states (σ1 , σ2 ) computed from randomly generated ζ ∈ Z(α) with α = 1.0. By using (23), (26), and (30), the maximum value σ1max of σ1 among possible realization of uncertain parameters ζ can be computed analytically as σ1max (α):

√ f1 + 2f 0 α max{q1 (ζ): ζ ∈ Z(α)} F = = max{σ1 (ζ): ζ ∈ Z(α)} = min{a1 (ζ): ζ ∈ Z(α)} F a1 − a01 α

(31)

Similarly, we obtain σ1min (α):

√ F f1 − 2f 0 α = min{σ1 (ζ): ζ ∈ Z(α)} = F a1 + a01 α

(32)

300 200

q2 (kN)

100 0 100 200 300

700

800

900

1000 1100 q1 (kN)

1200

1300

Figure 17.5 The axial forces q for randomly generated (ζ a , ζ f ) ∈ Za (α) × Zf (α) with α = 1.0.

482

Structural design optimization considering uncertainties

0 0.1

u2 (cm)

0.2 0.3 0.4 0.5 0

0.1

0.2

0.3 u1 (cm)

0.4

0.5

0.6

Figure 17.6 The nodal displacements u for randomly generated (ζ a ,ζ f ) ∈ Za (α) × Zf (α) with α = 1.0.

0.2 0.15

s2 (GPa)

0.1 0.05 0

0.05 0.1 0.15 0.2 0.2

0.3

0.4

0.5

0.6 0.7 s1 (GPa)

0.8

0.9

1

Figure 17.7 Stress states σ of the 2-bar truss with F a =F a1 for randomly generated (ζ a ,ζ f ) ∈ Za (α) × Zf (α) with α = 1.0.

√

2f 0 α F a2 − a02 α

(33)

σ2min (α): = min{σ2 (ζ): ζ ∈ Z(α)} = −σ2max (α)

(34)

σ2max (α):

= max{σ2 (ζ): ζ ∈ Z(α)} =

Substitution α = 1.0 and F a =F a1 into (31)–(34) results in σ1max = 855.2 MPa,

σ1min = 286.9 MPa,

σ2max = −σ2min = 113.1 MPa

(35)

Maximum robustness design of trusses via semidefinite programming

483

0.2 0.15

s2 (GPa)

0.1 0.05 0

0.05 0.1 0.15 0.2 0.2

0.3

0.4

0.5 0.6 0.7 s1 (GPa)

0.8

0.9

1

Figure 17.8 Stress states σ of the 2-bar truss with F a =F a1 for randomly generated (ζ a ,ζ f ) ∈ 1 Za (α) × Zf (α) with α =G α(F a ) = 1.277.

It is verified by Figure 17.7 and (35) that the stress constraints (28) are always inactive for any ζ ∈ Z(α) with α = 1.0. This implies that the robustness function G α(F a1 , σ c ) is greater than 1.0. Observe that the definition (21) (with (22)) of the robustness function is alternatively rewritten as G α(F a1 , σ c ) = max{α: σimax (α) ≤ σic , σimin (α) ≥ −σic (i = 1, 2)}

(36)

By substituting F a =F a1 into (31)–(34), we see that σ1max (α) > σ2max (α) and σ1max (α) > min α(F a1 , σ c ) satisfies the condition |σ1 (α)| hold for any α ≥ 0. Hence, (36) implies that G α) = σ1c σ1max (G from which we obtain G α(F a1 , σ c ) =

σ1cF f1 a1 − F √ c 0 σ1 a1 + 2f 0

= 1.277

α(F a1 , σ c ) The stress states (σ1 , σ2 ) computed from randomly generated ζ ∈ Z(α) with α =G is shown in Figure 17.8. It is observed from Figure 17.8 that the stress constraints 17.8 are always satisfied for the generated ζ, and that the worst case corresponds to the case in which σ1 (ζ) = σ1c holds. The stress constraints on the member (2) are always inactive. We next consider the nominal cross-sectional areas F a2 = (31.7, 21.7)T cm2

484

Structural design optimization considering uncertainties

as an alternative truss design. Note that F a1 and F a2 share the same structural volume, 2 and at F a =F a the condition σ1max (α) = σ2max (α)

(37)

is satisfied. Thus, the robustness function G α(F a2 , σ c ) now satisfies the condition σ1max (G α) = σ1c ,

σ2max (G α) = σ2c

from which we obtain G α(F a2 , σ c ) = 2.774 For the truss defined by F a =F a2 , Figure 17.9 depicts the stress states (σ1 , σ2 ) computed from randomly generated ζ ∈ Z(α) with α =G α(F a2 , σ c ). From Figure 17.9 it is seen that c c c the constraints σ1 ≤ σ1 , σ2 ≤ σ2 , and σ2 ≥ −σ2 become active in the worst cases, i.e. the constraints on both members can happen to be active. It is of interest to note that the robustness function of the truss design F a2 is larger 1 1 2 a and F a have the same structural than twice of that of F a in spite of the fact that F volume. This implies that the truss defined by F a2 violates the constraints only at larger ambient uncertainty compared with F a1 . Thus, we may naturally conclude that the truss 2 1 a . design F a is more robust than F Unfortunately, if a truss has moderately many degrees of freedom and/or the uncertainty set has a complicated structure, it is difficult to find the worst case parameters and the corresponding active constraint conditions. This is the crucial difficulty in evaluating the robustness function. This motivates us to propose a numerically tractable formulation for finding a lower bound of the robustness function in the following section.

1 0.8 0.6

s2 (GPa)

0.4 0.2 0 0.2 0.4 0.6 0.8 1 0

0.2

0.4

0.6 s1 (GPa)

0.8

1

Figure 17.9 Stress states σ of the 2-bar truss with F a =F a2 for randomly generated (ζ a ,ζ f ) ∈ 2 Za (α) × Zf (α) with α =G α(F a ) = 2.774.

Maximum robustness design of trusses via semidefinite programming

485

6 Computation of robustness function In this section, we propose an approximation algorithm for Problem (22), which provides a lower bound on the robustness function G α(F a, r c ). We also show that the exact value of the robustness function can be obtained by solving an SDP problem if a is certain. 6.1

Lower bounds of robus tnes s functio n

We start with embedding (16) and (17) into a finite number of quadratic inequalities. d m Define the matrix ∈ Rn ×n by

= (b1 , . . . , bnm ) where bi has been introduced in (4). In what follows, we assume nd < nm , which is usually satisfied for moderately large trusses. Define nn by nn = nm − rank

(38)

where rank denotes the row rank of . Then we see nn > 0. m d m n Let † ∈ Rn ×n denote the pseudo-inverse of . We denote by ⊥ ∈ Rn ×n a basis nm for the nullspace of , where the nullspace of is the set of all vectors β ∈ R satisfying n n d n d

β = 0. Letting ν ∈ Rn , define ξ ∈ Rn +2n +1 and H
l (r c ) ∈ S n +2n +1 (l = 1, . . . , nc ) by ξ = (ν, η, u, 1)

 O O H
l (r c ) = O H l (r c ) so that H
l ξ

 u = Hl 1

holds, where H l (r c ) has been introduced in (12). Let †i,· and ⊥ i,· denote the ith row of

the matrices † and ⊥ , respectively. Note that †i,· and ⊥ i,· are row vectors. Define n +2nd +1

n

d

(i = 1, . . . , nm ) and p (α2 ) ∈ S n +2n +1 (p = 1, . . . , nt ) as ⎞ ⎛ T −( ⊥ ⎛ ⎞ i,· ) 0 ⎟ ⎜ ⎜ ( † )T ⎟ ⎟ ⎜ ⎟ ⎜ i,· † † F 2 2⎜ 0 ⎟ T 0 i (α ) = α ⎝ 0 ⎠ (0T 0T ai bi 0) − ⎜ ⎟ (− ⊥ i,· i,· − i,· K 0) ai bi ⎜−( † K) F T⎟ ⎝ ⎠ i,· 0 0

i (α2 ) ∈ S n

⎛ ⎞ 0 ⎜ ⎟ 2 2⎜0⎟ p (α ) = α ⎝ ⎠ (0T 0 f0

⎞ O T ⎜ Tp ⎟ ⎟ f 0) − ⎜ ⎝ O ⎠ (O T −F f TT ⎛

0T

0T

p

Tp

O

f) −T pF

486

Structural design optimization considering uncertainties

Proposition 6.1.

The conditions (16) and (17) hold if and only if ξ satisfies

ξ T i (α2 )ξ ≥ 0,

i = 1, . . . , nm

(39)

ξ T p (α2 )ξ

p = 1, . . . , n

(40)

Proof.

≥ 0,

t

m

By introducing w = (wi ) ∈ Rn , we see that (16) is equivalently rewritten as

F − η,

w = Ku

(41)

wi = ζai (−a0i bTi u),

α ≥ |ζai |,

i = 1, . . . , nm

(42)

⊥

From the definition of and , we see that any solution to (41) can be written as †

F − η) + ⊥ ν w = † (Ku

(43)

nn

with ν ∈ R . On the other hand, the condition (42) is equivalent to wi2 ≤ (a0i α)2 (bTi u)2 ,

i = 1, . . . , nm

(44)

Consequently, by using (43) and (44), we see that (41) and (42) are equivalent to F − η) + ⊥ ν]2 ≥ 0, (a0i α)2 (bTi u)2 − [ †i,· (Ku i,·

i = 1, . . . , nm

Thus, the condition (16) is equivalent to (39). From the definition (8) of Zf it follows that ζ f ∈ Zf if and only if ζ f satisfies α2 ≥ ζ Tf T Tp T p ζ f ,

p = 1, . . . , nt

Hence, the condition (17) can be equivalently embedded into the following quadratic inequalities in terms of η: f )T T Tp T p (η − F f ), (f 0 α)2 ≥ (η − F

p = 1, . . . , nt

which can be rewritten as (17). Consequently, (17) is equivalent to (40). t c

Let ρ ∈ Rn n and τ ∈ Rn

m nc

be

ρ = (ρ11 , . . . , ρnt 1 , . . . , ρ1nc , . . . , ρnt nc )T τ = (τ11 , . . . , τnm 1 , . . . , τ1nc , . . . , τnm nc )T The following proposition, which plays a key role in constructing an approximation of Problem (22), shows a relaxation of infinitely many constraints by using a finite number of constraints: Proposition 6.2. u ∈ U(α,F a)

The implication =⇒

u ∈ F(r c )

(45)

holds if there exist ρ and τ satisfying t

H
l (rc )

−

n  p=1

m

ρpl p (α ) − 2

n  i=1

τil i (α2 )  O,

l = 1, . . . , nc

(46)

Maximum robustness design of trusses via semidefinite programming

ρ ≥ 0,

τ≥0

487

(47)

Proof. From Proposition 6.1 it follows that u ∈ U(α,F a) if and only if (39) and (40) are satisfied. Observe that the constraint (13) is reduced to ξ T H
l ξ ≥ 0,

l = 1, . . . , nc

Consequently, the implication (45) holds if and only if the implication ξ T p ξ ≥ 0,

p = 1, . . . , nt

ξ T i ξ ≥ 0 i = 1, . . . , nm

=⇒

ξ T H
l ξ ≥ 0

(48)

holds for each l = 1, . . . , nc . The assertion of this proposition is obtained by applying Lemmas A.1 and A.2 (ii) to (48). Proposition 6.2 implies that the set of a finite number of constraints (46) and (47) in terms of a finite number of variables corresponds to a sufficient condition for the infinitely many constraints of Problem (22). A lower bound of Problem (22) is then naturally constructed as follows: t c

Consider the following problem in variables (t, ρ, τ) ∈ R × Rn n × Rn ⎧ nt nm ⎨  
c ∗ t : = max ρpl p (t) − τil i (t)  O (l = 1, . . . , nc ), t : Hl (r ) − t,ρ,τ ⎩

Lemma 6.3.

p=1



m nc

:

i=1

ρ ≥ 0, τ ≥ 0

(49)

Then G α(F a, r c )2 ≥ t ∗ Proof. Recall that the robustness function G α is defined by (21) with Problem (22). It follows from Proposition 6.2 that the constraints of Problem (22) are satisfied if the constraints of Problem (49) are satisfied. This completes the proof. 6.2 Algorithm for computing lower bounds Lemma 6.4.

Problem (49) is a quasiconvex programming problem.

For a given t, define a set T by ⎧   nt nm ⎨   t m c  T (−t) = (ρ, τ) ∈ R(n +n )n  H
l − ρpl p (t) − τil i (t)  O (l = 1, . . . , nc ), ⎩  p=1 i=1 

Proof.

ρ ≥ 0, τ ≥ 0

488

Structural design optimization considering uncertainties

By regarding t ∈ R as an auxiliary variable, Problem (49) is equivalently rewritten as min{t : (ρ, τ) ∈ T (t)}

(50)

t,ρ,τ

Observe that T (t) is defined by nc linear matrix inequalities and (nt + nm )nc linear inequalities. Hence, T (t) is convex for any given t ∈ R. This implies that Problem (50) is a quasiconvex optimization problem. Let I denote the identity matrix with an appropriate size. For a fixed t, consider the t c m c following problem in the variables (s, ρ, τ) ∈ R × Rn n × Rn n :

∗

s : = min s,ρ,τ

⎧ ⎨

t

⎩

s : H
l (r c )

−

n 

m

ρpl p (t) −

p=1

n 

τil i (t) + sI  O (l = 1, . . . , nc ) ,

i=1

 ρ ≥ 0, τ ≥ 0

(51)

Problem (51) corresponds to a convex feasibility problem of Problem (49) at the given level t. Lemma 6.4 guarantees that the following bisection method solves Problem (49): Algorithm 6.5 (bisection method for Problem (49)). Step 0: Step 1: Step 2: Step 3: Step 4:

0

0

Choose t 0 and t satisfying 0 ≤ t 0 ≤ t ∗ ≤ t , and the small tolerance  > 0. Set k = 0. k k If t − t k ≤ , then stop. Otherwise, set t = (t k + t )/2. ∗ ∗ ∗ Find an optimal solution (s , ρ , τ ) to the SDP problem (51). k+1 k k+1 If s∗ ≤ 0, then set t k+1 = t and t = t . Otherwise, set t = t and t k+1 = t k . Set k ← k + 1, and go to Step 1.

Algorithm 6.5 finds a global optimal value t ∗ of Problem (49) by solving some SDP 0 problems, where exactly #log2 ((t − t 0 )/)$ iterations are required before the algorithm terminates. Here, we denote by #γ$ the minimum integer that is not smaller than γ ∈ R. From Lemma 6.3 it follows that (t ∗ )1/2 corresponds to a lower bound of the robustness function G α(F a, r c ). At Step 0, we may simply choose t 0 = 0, and a 0 sufficiently large t . At Step 2 of each iteration, we solve Problem (51), which can be embedded into the standard form of SDP problem (1) with m = nc (nt + nm + 1) + 1 and n = nc (nn + 2nd + nm + nt + 1). It should be emphasized that a global optimal solution to an SDP problem (51) can be obtained by using the primal-dual interior-point method, where the number of arithmetic operations is bounded by a polynomial of m and n (Wolkowicz et al. 2000).

Maximum robustness design of trusses via semidefinite programming

6.3

489

Spec ial cas e

The remainder of this section is devoted to investigating the case in which a is certain. The following result shows that, under some assumptions on the uncertainty set, the robustness function G α can be obtained by solving an SDP problem: By putting nt = 1 in (8), let Zf be

Proposition 6.6.

d

Zf (α) = {ζ ∈ Rn |α ≥ T1 ζ2 }

(52)

and let a0i = 0,

i = 1, . . . , nm

(53)

a) ∈ S n in (5). Assume that Hl  O (l = 1, . . . , nc ). Define 0 (α2 ,F ) *

 (T1F K)T 0 ! T 0" 2 2 0 f − (T1F K −F f ). 0 (α ,F a) = α T f0 −F f

d +1

by

Then the robustness function G α(F a, r c ) is obtained by solving the following SDP problem nc in the variables (t, µ) ∈ R × R with µ = (µl ) ∈ nc : D E G α(F a, r c )2 = max t : µl Hl (r c ) − 0 (t,F a)  O (l = 1, . . . , nc ), µ ≥ 0 (54) t,µ

Proof. From (52) and (53), the uncertain equilibrium equations (16) and (17) are reduced to F =F Ku f + f 0 ζ f , α ≥ T 1 ζ f 2 Hence, u ∈ U(α,F a) if and only if 1 2T 1 2 F −F F −F T 1 (Ku T 1 (Ku f) f ) ≤ α2 from which we obtain u ∈ U(α,F a)

⇐⇒

T

 u u 0 (α2 ,F a) ≥0 1 1

(55)

By using (55) and Lemmas A.2 (i) and A.1, we see that the implication (45) holds if and only if ∃ρl ≥ 0

subject to

H l  ρl 0 (α2 ,F a),

l = 1, . . . , nc

(56)

Note that H l  O implies that ρl = 0 does not satisfy (56). Hence, by putting µl = 1/ρl , l = 1, . . . , nc , the implication (56) is reduced to ∃µl ≥ 0

subject to µl H l − 0 (α2 ,F a)  O,

l = 1, . . . , nc

Consequently, Problem (22) is reduced to D E a)  O (l = 1, . . . , nc ), µ ≥ 0 G α(F a, r c ) = max α : µl H l (r c ) − 0 (α2 ,F α,µ

(57)

We see in Problem (57) that maximizing α is equivalent to maximizing α2 , which concludes the proof.

490

Structural design optimization considering uncertainties

7 Maximization of robustness function Throughout this section, we assume that the assumptions in Proposition 6.6 hold, i.e. only f possesses the uncertainty defined by (6) and (52) and a =F a is always satisfied. In Section 5, we have observed through an analytical example that the truss with the larger robustness function is considered to be more robust. We attempt in this section to find F a which maximizes the robustness function G α(F a, r c ). We call this structural optimization problem the maximization problem of robustness function. Consider the conventional constraints on F a which are dealt with in the usual structural optimization problems, e.g. the upper and lower-bound constraints of F a and m g the upper-bound constraint of structural volume. Letting g : Rn → Rn be a smooth function, we assume that these constraints can be written in the form of g(F a) ≥ 0

(58)

Note that g(F a) involves neither u nor f . For the given r c and g, the maximization problem of robustness function is formulated as max{G α(F a, r c ) : g(F a) ≥ 0} F a

(59)

In what follows, the argument r c is often omitted for brevity. m α(F a, r c ) > 0 and g(F a) ≥ 0. Then the objecAssume that there exists F a ∈ Rn satisfying G tive function of Problem (59) can be replaced by G α(F a, r c )2 without changing the optimal solution. From this observation and Proposition 6.6 it follows that Problem (59) is equivalent to the following problem: max{t : µl H l − 0 (t,F a)  O (l = 1, . . . , nc ), µ ≥ 0, g(F a) ≥ 0} t,µ,F a

(60)

Problem (60) is sometimes referred to as nonlinear semidefinite programming problem (Kanzow et al. 2005). To solve Problem (60), we next propose a sequential SDP method, which is an extension of the successive linearization method for standard nonlinear programming problems. Let DG(x
) denote the derivative of the smooth mapping G : Rm → S n at x
= (x
i ) ∈ Rm defined such that DG(x
)h is a linear function of h = (hi ) ∈ Rm given by DGl (x
)h =

 m  ∂Gl (x)  hi ∂xi x=x
i=1

The following is the sequential SDP method solving Problem (60) based on the successive linearization method: Algorithm 7.1 (Sequential SDP method for Problem (60)). Step 0: Step 1:

Choose F a0 satisfying g(F a0 ) ≥ 0 and G α(F a0 , r c ) > 0; choose cmax ≥ cmin > 0, 0 c ∈ [cmin , cmax ], and the small tolerance  > 0. Set k = 0. Find an optimal solution (t k , µk ) of Problem (54) by setting F a =F ak .

Maximum robustness design of trusses via semidefinite programming

Step 2:

c

491

m

Find the (unique) optimal solution ( t k , µk , F ak ) ∈ R × Rn × Rn of the SDP problem ⎫ 1 k ⎪ ⎪ c ( t, µ, F a)22 ⎪ ⎪

t, µ, F a 2 ⎬ k c subject to F l ( t, µ, F a)  O, l = 1, . . . , n , ⎪ ⎪

µ + µk ≥ 0, ⎪ ⎪ ⎭ k T k ∇g(F a ) F a + g(F a )≥0 max

t −

(61)

where a) = ( µl + µkl )H l − D 0 (t k ,F ak )( t, F a T )T − 0 (t k ,F ak ). F kl ( t, µ, F Step 3: Step 4:

If ( t k , µk , F ak )2 ≤ , then stop. k+1 k =F a + F ak . Set F a k+1 ∈ [cmin , cmax ]. Set k ← k + 1, and go to Step 1. Choose c

Essentially, Algorithm 7.1 solves the nonlinear SDP problem (60) by successively approximating it as the SDP problems. In Steps 1 and 2, we solve the SDP problems (54) and (61) by using the primal-dual interior-point method (Wolkowicz et al. 2000). The following proposition shows the global convergence property of Algorithm 7.1: Proposition 7.2. (Kanno & Takewaki 2006b). SupposeF f = 0 and that Problem (61) is strictly feasible at each iteration. Let {(t k , µk ,F ak )} be a sequence generated by Algorithm ak )} is a stationary point of Problem (60). 7.1. Then any accumulation point of {(t k , µk ,F

8 Numerical examples The lower bounds on the robustness functions are computed for various trusses by using Algorithm 6.5. Moreover, the optimal designs with the maximal robustness functions are computed for various trusses by using Algorithm 7.1 in the case where only the external forces possess uncertainties. In these algorithms, the SDP problems are solved by using SeDuMi Ver. 1.05 (Sturm 1999), which implements the primaldual interior-point method for the linear programming problems over symmetric cones. Computation has been carried out with MATLAB Ver. 6.5.1 (The MathWorks, Inc. 2002). 8.1

20-bar trus s

Consider a plane truss illustrated in Figure 17.10, where nd = 16 and nm = 20. Nodes (a) and (b) are pin-supported. The lengths of members in the x- and y-directions, respectively, are 100 cm and 50 cm. The elastic modulus of each member is 200 GPa. We assume that the cross-sectional areas of members (1)–(5) have uncertainty, whereas those of members (6)–(20) are certain. The external loads applied at nodes (e)–(j) have uncertainty, whereas those applied at nodes (c) and (d) are certain. No external loads are applied at nodes (c) and (d). The nominal cross-sectional areas are

492

Structural design optimization considering uncertainties

y

(11)

(19)

(16)

(c)

(18)

(3) (4)

0

(12)

(d)

(1)

(2) (a)

(13)

(f)

(6) (15)

(14)

(h)

(7)

(e) (9)

(20)

(17)

(g) (10)

(j)

(8)

(i)

(5) (b)

兵

兵

Uncertain loads & certain stiffness

Uncertain stiffness & certain loads

x

Figure 17.10 20-bar truss.

F ai = 20.0 cm2 (i = 1, . . . , 20). As the nominal external loads, we consider the following two cases: (Case 1): (Case 2):

(200.0, 0) kN, (500.0, 0) kN, (700.0, −400.0) kN, and (0, −400.0) kN are applied at the nodes (e), (g), (i), and (j), respectively; (200.0, 0) kN, (500.0, 0) kN, (700.0, −700.0) kN, and (0, −700.0) kN are applied at the nodes (e), (g), (i), and (j), respectively.

The coefficients of uncertainty in (5) and (6) are a0i = 2.5 cm2 (i = 1, . . . , 5) and (j) (j) f 0 = 50.0 kN. The uncertainty set for ζ f is given by (52) with T 1 = I. Let u(j) = (ux , uy )T denote the nodal displacement vector of the node (j). As the constraint (13) we consider the following conditions: |u(j) x | ≤ 5.0 cm,

(62)

|u(j) y |

(63)

≤ 2.0 cm.

The lower bound of the robustness function G α(F a, uc ) is computed by using Algorithm 0 0 6.5 for each case. We set t = 0, t = 10.0, and  = 10−4 . The lower bounds (t ∗ )1/2 are obtained as 2.672 and 2.412 for (Case 1) and (Case 2), respectively, after 17 SDP problems are solved. Thus, the robustness functions depend on the nominal external loads.

Maximum robustness design of trusses via semidefinite programming

493

0.8 1

u(j)y

1.2 1.4 1.6 1.8 2 2

2.5

3

3.5 u(j) x

4

4.5

5

Figure 17.11 Nodal displacements of the node ( j ) in (Case 1) for randomly generated ζ ∈ Z(α) with α = 2.6717. 0.8 1 1.2

u(j) y

1.4 1.6 1.8 2 2

2.5

3

3.5 u(j) x

4

4.5

5

Figure 17.12 Nodal displacements of the node ( j ) in (Case 2) for randomly generated ζ ∈ Z(α) with α = 2.4124.

We next randomly generate a number of ζ a and ζ f satisfying (7) and (8), respectively, by putting α = (t ∗ )1/2 , and compute the corresponding nodal displacements. Figures 17.11 and 17.12 depict the obtained displacement of the node (j) for (Case 1) and (Case 2), respectively. It is observed from Figures 17.11 and 17.12 that the

494

Structural design optimization considering uncertainties

constraints (62) and (63) are satisfied for all generated (ζ a , ζ f ), which verifies that α. In (Case 1), from Figthe obtained values (t ∗ )1/2 are certainly the lower bounds of G ure 17.11 we may conjecture that the worst case corresponds to the case in which the constraint (62) becomes active. On the other hand, in (Case 2), Figure 17.12 shows that the worst case corresponds to the case in which the constraint (63) becomes active. In (j) (j) both cases, at least one of |ux | and |uy | possibly becomes very close to its bound. This implies that Algorithm 6.5 provides sufficiently tight lower bounds, i.e. the obtained α in each case. value (t ∗ )1/2 is very close to the exact value of G 8.2

2 9 - b ar trus s

Consider a truss illustrated in Figure 17.13. The nodes (a) and (b) are pin-supported, where nd = 20 and nm = 29. The lengths of members both in x- and y-directions are 50.0 cm. The elastic modulus of each member is 200 GPa. Suppose that the force of (0, −10.0) kN is applied at the nodes (c) and (d) as the nominal external load F f . The uncertainty set for ζ f is given by (52) with T 1 = I. We put f 0 = 1.0 kN in (6). The member cross-sectional areas a are assumed to be certain. Hence, we can compute the exact value of the robustness function by using Proposition 6.6. Consider the stress constraints (14) with σic = 500 MPa for each member. The maximization problem (60) of the robustness function is solved by using Algorithm 7.1. As the constraints (58) in Problem (59), we consider the conventional constraint on structural volume as well as nonnegative constraints of F a, namely, g is defined as

g(F a) =

 F a V − V(F a)

y (a)

(9)

(10)

(18) (1)

(15)

0

(b)

(14)

(22) (4)

(23) (6)

(28)

(8)

(29) (17)

(16)

(c)

(d) ~ ƒ

Figure 17.13 29-bar truss.

(7)

(26)

(13)

(21) (27)

(5)

(25)

(12)

(2)

(20)

(19) (3)

(24)

(11)

x ~ ƒ

Maximum robustness design of trusses via semidefinite programming

495

Here, V(F a) denotes the total structural volume of a truss, which is a linear function a0i = 20.0 cm2 (i = 1, . . . , nm ). We of a, and ng = nm + 1. The initial solution is given as F first compute the robustness function at the initial solution F a =F a0 by using Proposition 6.6. Since only the external load possesses the uncertainty, the robustness function is obtained as G α(F a0 ) = 0.7261 by solving only one SDP problem. In Algorithm 7.1 we set  = 0.1, cmax = cmin = 10−5 , and V = 3.3971 × 104 cm3 so aopt found by that the volume constraint becomes active at F a =F a0 . The optimal design F Algorithm 7.1 after 53 iterations is shown in Figure 17.14, where the width of each member is proportional to its cross-sectional area. The corresponding robustness function is G α(F aopt ) = 11.0710. We compute the optimal designs for various V. Figure 17.15 depicts the relation between V and the robustness function at the optimal design. For comparison, we compute the robustness function for the cross-sectional areas that

Figure 17.14 Optimal design of the 29-bar truss. 12

Robustness function ^ a

10 8 6 4 2 0 0

0.5

1

1.5 2 Volume V (cm3)

2.5

3

3.5 104

Figure 17.15 Relation between V and G α of the optimal trusses (×: initial solution; •: optimal solution; ∗: solution obtained by scaling aopt ).

496

Structural design optimization considering uncertainties

are obtained by scaling F aopt . It is observed from Figure 17.15 that the optimal design cannot be obtained only by scaling F aopt . It is of interest to note that, from the definition of the robustness function, all truss designs are plotted in (or on the boundary of) the domain D in Figure 17.15. Thus, engineers may be able to make decisions incorporating the trade-off between the robustness and the structural volume.

9 Conclu sions Based on the info-gap theory (Ben-Haim 2006), the robustness function of trusses has been investigated extensively as a measure of robustness of a truss under load and structural uncertainties. We have proposed an approximation algorithm for computing the robustness functions of trusses under the load and structural uncertainties. A global convergent algorithm has been proposed for the maximization problem of the robustness function. We have introduced an uncertainty model of trusses, where the external forces as well as the member stiffness include uncertainties. We assume that the constraints on mechanical performance can be expressed by using some quadratic inequalities in terms of displacements. In fact, we can deal with the polynomial inequality constraints in terms of displacements by converting them into a finite number of quadratic inequalities. Then we have formulated a quasiconvex optimization problem, which provides a lower bound, i.e. a conservative estimation, of the robustness function. In order to obtain a global optimal solution to the present quasiconvex optimization problem, a bisection method has been proposed, where a finite number of SDP problems are successively solved by the primal-dual interior-point method. In order to solve the maximization problem of the robustness function for variable member cross-sectional areas, a sequential SDP approach has been presented, where the SDP problems are successively solved by the primal-dual interior-point method to obtain the optimal truss designs. The method has been shown to be globally convergent under certain assumptions.

Technical lemmas Lemma A.1 (homogenization). two conditions are equivalent: (a) (b)

 

T x x Q p ≥ 0, 1 1 pT r 

Q p O pT r

Let Q ∈ S n , p ∈ Rn , and r ∈ R. Then the following

∀ x ∈ Rn

Proof. The implication from (b) to (a) is trivial. We show that (a) implies (b) by the contradiction. Suppose that there exist x
∈ Rn and η ∈ R satisfying


T 
 x Q p x <0 pT r η η

(64)

Maximum robustness design of trusses via semidefinite programming

497

for the contradiction to the assertion (b). If η = 0, then (64) is reduced to

x
/η 1

T


 x /η Q p <0 1 pT r

which contradicts the assertion (a). Alternatively, if η = 0, then (64) is reduced to x
T Qx
< 0

(65)

Letting x = ζx
, the left-hand side of (a) is reduced to (xT Qx
)ζ 2 + 2(pT x
)ζ + r

(66)

which is regarded as a function of ζ. (65) implies that (66) is not bounded below, from which it follows that there exists a ζ such that (66) becomes negative. Thus, we see the contradiction to (a). Lemma A.2 (S-lemma). Then, (i) the implication f1 (x) ≥ 0

=⇒

Let f0 (x), f1 (x), . . . , fm (x) be quadratic functions of x ∈ Rn .

f0 (x) ≥ 0

holds if and only if there exists τ1 ( ≥ 0) such that f0 (x) ≥ τ1 f1 (x),

∀ x ∈ Rn

(ii) the implication f1 (x) ≥ 0, . . . , fm (x) ≥ 0

=⇒

f0 (x) ≥ 0

holds if there exist τ1 , . . . , τm ≥ 0 such that f0 (x) ≥

m 

τi fi (x),

∀ x ∈ Rn

i=1

Proof.

See Ben-Tal & Nemirovski (2001, Theorem 4.3.3).

References Ben-Haim, Y. 2004. Uncertainty, probability and information-gaps. Reliability Engineering and System Safety 85:249–266. Ben-Haim, Y. 2005. Value at risk with Info-gap uncertainty. Journal of Risk Finance 6:388–403. Ben-Haim, Y. 2006. Information-gap Decision Theory: Decisions Under Severe Uncertainty. 2nd edition. London: Academic Press. Ben-Haim, Y. & Elishakoff, I. 1990. Convex Models of Uncertainty in Applied Mechanics. New York: Elsevier. Ben-Tal, A. & Nemirovski, A. 2001. Lectures on Modern Convex Optimization: Analysis, Algorithms, and Engineering Applications. Philadelphia: SIAM.

498

Structural design optimization considering uncertainties

Ben-Tal, A. & Nemirovski, A. 2002. Robust optimization – methodology and applications. Mathematical Programming B92:453–480. Boyd, S. & Vandenberghe, L. 2004. Convex Optimization. Cambridge: Cambridge University Press. Calafiore, G. & El Ghaoui, L. 2004. Ellipsoidal bounds for uncertain linear equations and dynamical systems. Automatica 40:773–787. Chen, S., Lian, H. & Yang, X. 2002. Interval static displacement analysis for structures with interval parameters. International Journal for Numerical Methods in Engineering 53: 393–407. Helmberg, C. 2002. Semidefinite programming. European Journal of Operational Research 137:461–482. Kanno, Y. & Takewaki, I. 2006a. Robustness analysis of trusses with separable load and structural uncertainties. International Journal of Solids and Structures 43:2646–2669. Kanno, Y. & Takewaki, I. 2006b. Sequential semidefinite program for maximum robustness design of structures under load uncertainties. Journal of Optimization Theory and Applications 130:265–287. Kanzow, C., Nagel, C., Kato, H. & Fukushima, M. 2005. Successive linearization methods for nonlinear semidefinite programs. Computational Optimization and Applications 31: 251–273. Kojima, M. & Tunçel, L. 2000. Cones of matrices and successive convex relaxations of nonconvex sets. SIAM Journal on Optimization 10:750–778. Moilanen, A. & Wintle, B.A. 2006. Uncertainty analysis favors selection of spatially aggregated reserve structures. Biological Conservation 129:427–434. Ohsaki, M., Fujisawa, K., Katoh, N. & Kanno, Y. 1999. Semi-definite programming for topology optimization of truss under multiple eigenvalue constraints. Computer Methods in Applied Mechanics and Engineering 180:203–217. Pantelides, C.P. & Ganzerli, S. 1989. Design of trusses under uncertain loads using convex models. Journal of Structural Engineering (ASCE) 124:318–329. Pierce, S.G., Worden, K. & Manson, G. 2006. A novel information-gap technique to assess reliability of neural network-based damage detection. Journal of Sound and Vibration 293: 96–111. Sturm, J.F. 1999. Using SeDuMi 1.02, a MATLAB toolbox for optimization over symmetric cones. Optimization Methods and Software 11/12:625–653. Takewaki, I. & Ben-Haim, Y. 2005. Info-gap robust design with load and model uncertainties. Journal of Sound and Vibration 288:551–570. Vandenberghe, L. & Boyd, S. 1996. Semidefinite programming. SIAM Review 38:49–95. Wolkowicz, H., Saigal, R. & Vandenberghe, L. (eds) 2000. Handbook of Semidefinite Programming – Theory, Algorithms, and Applications. Dordrecht: Kluwer. Using MATLAB, 2002. Natick: The MathWorks.

Chapter 18

Design optimization and robustness of structures against uncertainties based on Taylor series expansion Ioannis Doltsinis University of Stuttgart, Stuttgart, Germany

ABSTRACT: Synthetic Monte Carlo sampling and analytic Taylor series expansion offer two different techniques for the treatment of random input scatter. The present chapter expounds on the Taylor series approximation as applied to the stochastic analysis and design optimization of structures including robustness against uncertainties. A unified approach is presented that encompasses linear and nonlinear elastic structures as well as path dependent elastic–plastic response. The methodology refers to finite element systems, and assumes that the response as a function of the input varies continuously; representation of the probability distribution is restricted to mean and variance. The approach is applicable to input scatter of practical relevance and is computationally efficient; its analytic nature allows utilization of optimization algorithms.

1 Introduction Consideration of random scatter in the analysis and design of structures is important with regard to reliability, and for securing standards of operation performance which implies robustness against uncertainties. The synthetic Monte Carlo sampling and the analytic Taylor series expansion offer alternatives of stochastic analysis and design improvement (Doltsinis 2003). This chapter expounds on the Taylor series approximation as applied to the stochastic analysis and design optimization of structures including nonlinear elastic and path dependent elastic–plastic response. The Taylor series expansion supplies a formalism suitable for developing the robust design of structures using optimization algorithms. The robustness problem is stated as a two criteria task involving minimization of mean value and standard deviation of the objective function, randomness of the constraints included. The quantities requested from stochastic analysis are mean vector and covariance matrix of the response variables; gradient based optimization demands design sensitivity expressions as well. The approximation is by a second-order Taylor series expansion of the finite element equations with respect to the random input. A uniform methodology is presented starting at elastic structural systems with randomness and extending to nonlinear elastic as well as path dependent elastic–plastic response. In elastoplasticity the procedure is adapted to the incrementation of the associated deterministic analysis. The proposed stochastic analysis is applied in conjunction with standard optimization to the robust design of linear and nonlinear structures. Monte Carlo sampling verifies the results. Merits and deficiencies of the Taylor expansion approach are discussed.

500

Structural design optimization considering uncertainties

The text is organized as follows. After Section 1 which introduces into the theme treated and summarizes the contents, Section 2 deals with design optimization and robustness. The task of structural optimization is briefly posed, the implications of randomness are discussed, and the issue of design robustness is addressed. The formal statement of stochastic optimization is extended to robust design that raises the problem of the concurrent observance of two criteria: mean and standard deviation of the objective function that characterize level and variability, respectively. The compound desirability function is introduced as the single requirement compromising the two criteria. The subject of robustness is illustrated by an example, the difference between performance robustness and structural reliability is pointed out. Section 3 is concerned with the random response of deformable systems as represented by finite elements. A Taylor series expansion to the second order about the mean values of the random input is employed to approximate the fluctuating response displacements of the system. Therefrom the expressions for the mean vector and the covariance matrix of the structural response are deduced in general terms; the interaction with the finite element equations is indicated. Section 4 specifies the formalism for the stochastic analysis of elastic structures. First, the linear finite element equations are developed in order to deduce the second-order approximate of the mean response and the first-order transform of the response covariance matrix. In view of an application of gradient based optimization algorithms, the response sensitivity analysis is carried out and the design sensitivity expressions are presented. The stochastic analysis is then extended to large displacement problems of elastic structures taking full account of the kinematic nonlinearity. The methodology is exposed for the quasistatic conditions underlying the equilibrium equations; application to transient response problems governed by the dynamic equations of motion is seen to be straightforward. Section 5 addresses the stochastic analysis of path dependent elastic–plastic response. In difference to the description of linear or nonlinear elastic structures, elastic–plastic systems imply incremental linearization of the process as a consequence of the path dependent constitutive nature of the problem. Following the incrementation of the primary (deterministic) algorithm of the elastic–plastic structural analysis, the stochastic formalism for means and covariances is developed essentially along the lines of the previous issues. The sensitivity expressions for the response displacements also have to be advanced incrementally during the course of the loading programme. A simplification introduced by discarding in the incremental step past scatter in geometry and stress is contrasted to the rigorous formalism, and the possible impact on the results is commented by a subsequent numerical demonstration. Of course, the incremental technique is equally applicable to the nonlinear elastic problem, and although not necessary from the theoretical point of view it is employed for structural analysis in practice. Section 6 illustrates the theory by numerical applications to nonlinear and to path dependent problems. For this purpose, the proposed methodology of stochastic analysis is combined with standard optimization to perform tasks of structural design optimization and robustness. To be specific, the stochastic analysis supplies means and variances of the response displacements that enter the evaluation of the objective, resp. of the desirability function in the context of robustness, design sensitivities support utilization of gradient based optimization algorithms. Apart from the instructive case of a planar cantilever truss that verifies the algorithm, selected problems include

Taylor approach to structural optimization and robustness

501

the structural compliance optimization of a space truss structure under material and geometrical nonlinearity, and the robust design of an antenna structure undergoing large displacements. Results are compared to synthetic Monte Carlo sampling. Section 7 concludes the chapter with a summary of the presented methodology and with a critical appraisal of the Taylor expansion approach as regarding the inherent assumptions, the quality of the approximation, the range of applicability and the computational efficiency. The present account refers to the work reported in (Doltsinis and Kang 2004) and (Doltsinis et al. 2005) as inspired by (Kleiber and Hien 1992); Zhan Kang contributed to the advancement of the subject in the framework of his doctoral dissertation (Kang 2005). The theoretical development here uses the explicit formalism of (Doltsinis 2003). The proposed methodology of stochastic structural design allows solution of the associated optimization problem by a software package available in the Internet (Lawrence et al. 2007). To this end, the optimization algorithm based on sequential quadratic programming is operated in conjunction with the input supplied by the developed stochastic finite element analysis.

2 Design optimization, robustness 2.1 The optimization tas k The purpose of structural design is the achievement of a certain performance of the system. If design variations are possible, optimization can be envisaged with respect to a specified objective. The following deals with structures as deformable systems represented by finite elements, the terminology is as in (Doltsinis and Rodiˇc 1999). The response to applied actions is defined by the N displacements of the mesh nodal points of the discretized object, collected in the N × 1 vector u(z). It depends on the p × 1 vector z = {z1 . . . zp } which comprises the set of p design parameters that are not fixed and are therefore disposable in optimization. Optimum design is attempted by minimizing a scalar objective function fo (z) = f [u(z), z]

(1)

which defines the performance. The minimum of fo (z) specifies the values of the design parameters z. The mathematical statement of the optimization problem is: find minimizing subject to and

z fo (z) gci (z) ≤ 0, i = 1, . . . , k zL ≤ z ≤ zU

(2)

In the above gci (z) = gi [u(z), z] are constraint functions, zL and zU lower and upper bounds, respectively, that restrict the value of the design variables. 2.2 Implicatio n of randomnes s When randomness is present in the system the problem is no longer deterministic but stochastic in nature. Scatter may be due to the external actions, such as fluctuating loads. Besides, shape geometry and part dimensions have to tolerate imperfections to a

502

Structural design optimization considering uncertainties

Load

Displacement

Figure 18.1 Response to fluctuating action and random deviation from nominal behaviour due to parameter scatter.

O

P, u

l

O

A, E

P

Figure 18.2 Bar subjected to tension.

certain extent, material properties exhibit random deviations from nominal behaviour. Both, fluctuating external actions and randomness of inherent parameters contribute to the scatter of the response of a structure (Figure 18.1). For the purpose of illustration let pay attention to the bar element (length l, crosssection A) depicted in Figure 18.2. Tension of the bar by the force P applied at the upper end induces there a displacement u while the other end is pin-jointed. Assuming linear elastic behaviour with modulus E: u=

l 1 P= P EA k

(3)

Taylor approach to structural optimization and robustness

503

where k = EA/l denotes the stiffness of the bar element, the quantity 1/k is known as the flexibility. Random fluctuations of the response displacement u can be due to the flexibility 1/k, and the force P. Mean value µu and variance σu2 are obtained by the expectation operations (for probabilistic terminology see (Breipohl 1970)) µu = E(u),

σu2 = Var(u) = E[(u − µu )2 ]

(4)

with u from Eq. (3). For instance, if P is random with mean µP and covariance σP2 while 1/k fixed, the mean value and the variance of the ensuing displacement u are given by µu =

1 µP , k

σu2 =

1 2 σ k2 P

(5)

Analogously, if the flexibility 1/k varies among bar elements in a series while P is fixed, one has µu = Pµ1/k ,

2 σu2 = P2 σ1/k

(6)

In either case the coefficient of variation (COV = σ/µ) of the random input is transferred to the output without alteration by the system. If both 1/k and P exhibit randomness but vary independently, mean value and variance of the displacement are obtained after substitution of Eq. (3) in the expectation operations of Eq. (4) as µu = µ1/k µP

(7)

2 2 ∼ 2 2 σu2 = σ1/k σP2 + µ21/k σP2 + µ2P σ1/k = µ1/k σP2 + µ2P σ1/k

(8)

and

2 The last, linearized expression in Eq. (8) presumes that the variances σ1/k , σP2 are small. Alternatively, a Taylor series expansion of Eq. (3) to the first order about the mean values of the input quantities 1/k, P gives

u

1 ,P k



∂u ∂P



∂u ∂(1/k) µ

 1 − µ1/k = µ1/k µP + µ1/k (P − µP ) + µP k

∼ = u(µ1/k , µP ) +

(P − µP ) +

 µ

1 − µ1/k k



(9)

Application to the above of the variance operator of Eq. (4) while bearing in mind the independence of 1/k and P confirms the linearized expression for the variance σu2 in Eq. (8). The approximation is not of relevance for Eq. (7), the mean value of u. Assuming deterministic input except of the applied force P, and taking the weight W of the bar with mass density # as the design objective, randomness is introduced in the

504

Structural design optimization considering uncertainties

problem by constraining the response displacement: u ≤ u0 . With the bar cross-section A as the design variable, the simple task of dimensioning is formally presented as find minimizing subject to

A W = #lA µu + βσu ≤ u0

(10)

On account of Eq. (5) the displacement constraint determines the cross-section A by µu + βσu =

l (µP + βσP ) ≤ u0 EA

(11)

The value of the parameter β controls the impact of randomness. For the choice β = 0 only the mean displacement µu complies with the constraint, for β = 1 scatter is covered up to a displacement exceeding the mean by the standard deviation σu . The random variability of the system in general reflects on the objective function, Eq. (1). In the following a direct dependence on the design parameters and the implicit dependence via the response variables will be discussed separately. The first case refers to an objective function fo = f (z), and demonstrates dealing with design parameters z that are random variables. The Taylor series expansion about the mean values µz = {µz1 . . . µzp } gives to the second order

2   df d f 1 t (z − µz ) + (z − µz ) (z − µz ) = (12) dz µ 2 dzdzt µ   p p  ∂f ∂2 f 1  (zk − µzk ) + (zk − µzk )(zl − µzl ) = f (µz ) + ∂zk µ 2 ∂zk ∂zl µ

f (z) ∼ = f (µz ) +

k=1

k,l=1

Note that df/dz is a row matrix, df/dzt = (df/dz)t represents a vector (column matrix). The last expression is helpful in obtaining a second order approximation to the mean value of the objective function:  p ∂2 f 1  µf ∼ σz z = f (µz ) + 2 ∂zk ∂zl µ k l k,l=1

) * p 1  ∂2 f = f (µz ) + σz2k 2 ∂zk2 k=1

µ

(13)

Taylor approach to structural optimization and robustness

505

The first order approximate of the variance of the objective function reads σf2 ∼ =

df dz

 µ

z

df dz

t µ

=

  p  ∂f ∂f σz z ∂zk µ ∂zl µ k l

k,l=1

2 p  ∂f = σ2 ∂zk µ zk

(14)

k=1

In the above, z = [σzk zl ] k, l = 1, . . . , p, denotes the symmetric p × p covariance matrix of the design variables. It comprises the mutual covariances σzk zl of the parameters zk , zl . The entries on the diagonal are the individual variances of the p parameters: σz2k = σzk zk with σzk being the standard deviation. The last expressions in Eqs. (13) and (14) are applicable only if the random variables are independent such that the population covariances σzk zl (k = l) vanish. The terminology referring to several random variables follows (Rencher 1995; Doltsinis 1999). In the approximation of µf by Eq. (13) independent random design variables are represented by the mean values µzk and the variances resp. the standard deviations σzk . If the second-order terms are discarded such that µf ∼ = f (µz ), only the mean values are design parameters: z ⇐ {µzk } k = 1, . . . , p. The standard deviations essentially determine the variance of the objective function, (Eq. (14)), but they may constitute additional design variables in higher order mean value approximation as well: z ⇐ {µzk ; σzk } k = 1, . . . , p. The general case of statistically dependent variables in z involves all the relevant entities of the covariance matrix. With an objective function fo = f (z) the random response affects merely the constraints of the optimization problem. This becomes different if the objective function is defined in terms of response variables determined by the displacement u(z): fo = f [u(z)] Keeping the design parameters hidden, the expansion to the second order about the mean displacement µu = {µu1 . . . µuN } gives here, f (u) ∼ = f (µu ) +

df du

1 + (u − µu )t 2

 µ

(u − µu )

d2 f dudut

 µ

(u − µu )

(15)

Therefrom, the second order approximate of the mean value of the objective function is deduced as  N ∂2 f 1 µf ∼ f (µ ) + σu u = u 2 ∂ui ∂uj µ i j i,j=1

(16)

Structural design optimization considering uncertainties

1.0

1.0

1

2

3

(2)

(1)

(3) 4

P

1.0

506

(4) 5

u

Figure 18.3 Four-bar truss.

and the first order approximation to the variance is

σf2

∼ =

df du

 µ

u

df du

t µ

  N  ∂f ∂f = σu u ∂ui µ ∂uj µ i j

(17)

i,j=1

Unlike the design parameters in Eqs. (13), (14) independence can hardly be assumed between the response variables ui , uj . The mean and variance of the objective function as by Eqs. (16) and (17) require knowledge of the mean vector µu and the covariance matrix u = [σui uj ] i, j = 1, . . . , N of the response displacements, which still are to be obtained. The stochastic counterpart of the deterministic optimization problem in Eq. (2) might be stated as to obtain best mean performance. There remains the issue of the variability caused by the fluctuating factors, however, and its diminution by specifying appropriate values for the design variables can be equally of importance. To illustrate the argument of random fluctuations, consider the displacement minimization problem for the four-bar truss shown in Figure 18.3. The plane structure is loaded by a static horizontal force P = 1 acting on node no. 4. The elastic modulus of bars 1, 3 is E1 , that of bars 2, 4 is E2 . The elastic moduli are considered independent random variables with mean and standard deviation µE1 = 210.0, σE1 = 21.0, and µE2 = 100.0, σE2 = 5.0. The cross-section areas A1 (bars 1, 3) and A2 (bars 2, 4) are disposable as the design variables. The design objective is to minimize at node no. 4 the horizontal displacement u under constraint structural weight W: µW ≤ 5.0. The mass density of the material is uniquely taken to # = 1.0. In this design problem the structural weight is an active constraint such that only one of the design parameters is independent and needs to be determined. Figure 18.4 shows the variation of mean and standard deviation of the observed displacement u as a function of the parameter A1 , the selected independent variable. From the left frame, the design that minimizes the mean value µu = E(u) is for A
1 = 1.7678, the right limit of the diagram. The appertaining value of the other parameter is A
2 = 0, which eliminates bars 2, 4. From the right frame of Figure 18.4, the above solution exhibits the largest standard deviation. The solution appertaining to minimum

Taylor approach to structural optimization and robustness

4.04

 103

4

507

 104 +

3.5

4 +

3 σ



3.96 3.92

2.5

3.88

2 +

3.84 0

+

A*1 0.5

1

A1

1.5

A1'

2

1.5 0

A*1 0.5

1

A1

1.5 A'1

2

Figure 18.4 Mean value µ (left) and standard deviation σ (right) of displacement u versus crosssection area A1 .

Figure 18.5 Probability (relative frequency) density distribution for two alternatively optimized designs: minimum mean displacement u (left), least variance (right).

standard deviation is A$1 = 0.3531, A$2 = 2.0006. The associated mean displacement is higher than before: 3.969 × 10−3 to 3.848 × 10−3 while 1.7 × 10−4 to 3.97 × 10−4 for the standard deviation in the two alternative solutions. The effect on level and scatter of u is demonstrated in Figure 18.5 where the probability (relative frequency) density of u has been approached by synthetic sampling (Monte Carlo simulation) with 106 realizations from normal distributions of the random input variables.

2.3

Robust des ign

Apart from best performance in the mean, robustness of the design requires the variability of the performance to be possibly low. Usually the option of reducing input

Structural design optimization considering uncertainties

Frequency of occurrence

508

2nd design 1st design

µ1

µ2

f

Figure 18.6 Concept of robust design.

scatter is excluded and robustness against fluctuating factors is attempted by adjusting the values of the design parameters. The concept of robust design is illustrated in Figure 18.6, which refers to the probability distribution of a fluctuating objective function f , the structural performance measure of interest to be minimized. The first design marked in the figure exhibits a smaller mean value of the performance function, but a larger variability than the second design. The second design is less sensitive to the scatter of input data, and is said to be more robust in this respect. At this point it is important to contrast robust design as pursued here and reliability based optimization. Reliability refers to the probability of failure PF in extreme situations where the defined failure criterion h exceeds a specified critical value c:
PF = P(h ≥ c) =

∞

ph dh

(18)

c

with ph denoting the probability density function for h. Optimization minimizes then the objective function under the constraint of a prescribed probability P0 of survival (Luo and Grandhi 1997) PS = 1 − PF ≥ P0

(19)

Methods to obtain the failure probability are described in (Hurtado 2004). There may be several probabilistic safety constraints like the above in addition to other, ordinary ones. The purpose here is to avoid, with a certain probability, system catastrophe in the presence of random parameters. Robust design, on the other hand, addresses the regular employment of the system. It aims at the reduction of performance variability due to fluctuating input. Robustness is assessed by the scatter of the performance function, most frequently measured by its standard deviation. Figure 18.7 distinguishes between robustness and reliability. For the purpose of illustration the performance function f in the robustness issue serves here also as the failure criterion for assessing reliability (h = f ).

Taylor approach to structural optimization and robustness

509

Probability density s

Limit state

Reliability

Robustness µ

Performance f

Figure 18.7 Distinguishing between robustness and reliability of design.

The task of robust optimum design involves both the mean value µfo of the objective function fo (z) and its variance σf2o resp. the standard deviation σfo . In mathematical terms: find minimizing subject to and

z



fo (z) =

µfo (z)



σfo (z) µgci (z) + βi σgc i (z) ≤ 0, zL ≤ z ≤ zU

(20) i = 1, . . . , k

The design parameters in the vector z can be deterministic quantities or mean and standard deviations of random variables. The constraint functions gci enter the optimization problem by the respective mean value µgci and the standard deviation σgci . The coefficient βi can be interpreted as the feasibility index for the individual constraint condition. Its value controls the impact of scatter on the constraint, and has to be specified according to the requirements. As a rule, the larger the value of βi the closer the constraint is met under fluctuating conditions. Equation (20) states a problem of multicriteria optimization associated with the simultaneous minimization of mean and standard deviation of the fluctuating objective function. The two criteria have been assembled to the vector valued objective function fo (z); involvement of the standard deviation rather than the variance makes the two quantities commensurable. A survey of vector or multicriteria optimization is found in (Stadler 1984). The definition of a vector optimum is not unique; one option is the Pareto optimum: it is achieved when further improvement in any one of the criteria values implies worsening of at least one other criterion. The vector optimization problem will be converted in the following to a scalar one by introducing the so-called desirability function (Myers and Montgomery 1995). Thereby the single criteria, which can be conflicting, are merged to a compound objective that compromises the requirements. Among various possibilities, a simple proposal defines the desirability function as the weighted sum of the two objectives (Lee and Park 2001): Fo (z, ξ) = (1 − ξ)

µfo (z) σf (z) +ξ o , µopt σopt

0<ξ<1

(21)

510

Structural design optimization considering uncertainties

The denominators serve as standardization, here taken equal to the separate minima of the mean value µopt and the standard deviation σopt of the objective function fo (z). These quantities will be associated to the limits ξ = 1 and ξ = 0 of the weighting factor, not included in Eq. (21). In the multicriteria task, the value of 0 < ξ < 1 must be selected according to the importance put on mean and variance of fo (z). With the desirability function Fo (z, ξ) for ξ = const. instead of the vector fo (z) in Eq. (20), the minimization task can be carried out using the standard optimization algorithms applicable to the deterministic problem of Eq. (2) under observance of the due constraints and restrictions. The computed optimum appertains to the desiraibility function Fo (z, ξ) defined by Eq. (21) and the value selected for ξ; the associated µfo (z) and σfo (z) are inherent in the system. Effecting the optimization with different values of the parameter ξ supplies a set of points in the µfo , σfo –plane, which enables the analyst to decide on the selection of the design variables. The stochastic optimization problem involves the calculation of the mean µ and the standard deviation σ of the objective function fo (z) = f [u(z), z], and of the constraints in terms of the design variables z. Gradient based optimization techniques require in addition availability of sensitivity expressions for the above quantities with respect to the design variables z. Focusing on the implicit dependence via the displacement u and neglecting the second-order terms in Eq. (16), the design sensitivity of the approximated mean objective follows to ∂µf = ∂zk

df du

 µ

∂µu ∂zk

and from Eq. (17) for the variance, 

t

2 

  ∂σf2 d f ∂µu df ∂u df = + 2u ∂zk du µ ∂zk du µ dudut µ ∂zk

(22)

(23)

3 Random response Execution of the optimization task demands eventually the mean and the covariances of the response quantities possibly also their design sensitivities, that is the derivatives with respect to the design variables. To this end, let the displacement u of the deformable system be stated as a function of deterministic parameters arranged in the vector b and of random parameters constituting the vector α: u = u(b, α) The design variables may be part of b and/or α, the generalized input to the system. Expanding u to the second order about the mean of α such that α = µα + α, while b is silent gives u2 = u0 + u +

1 2

u 2

(24)

Taylor approach to structural optimization and robustness

511

The above distinguishes between zeroth-order, first-order and second-order terms: u0 = u(b, µα ), 

 q  ∂u ∂u

u = (α − µα ) = (αi − µαi ) ∂α µ ∂αi µ  q

2 u =

i,j=1

∂ u ∂αi ∂αj 2

(25)

i=1

 µ

(αi − µαi )(αj − µαj )

The second-order approximate of the mean displacement thus becomes µu2

 q ∂2 u 1 = u(b, µα ) + σα α 2 ∂αi ∂αj µ i j i,j=1

= µu0 +

1 2

µu 2

(26)

and the first-order approximation to the covariance matrix is

u1 =

∂u ∂α

 µ

α

∂u ∂α

t (27) µ

If α = {y z} is composed of two sets y and z where y comprises random input parameters other than the design ones z, and the two sets are statistically independent, then the second-order term in Eq. (26) reads   p r   ∂2 u ∂2 u

µu = σy y + σz z ∂yi ∂yj µ i j ∂zk ∂zl µ k l 2

i,j=1

(28)

k,l=1

and Eq. (27),

u1 =

∂u ∂y

 µ

y

∂u ∂y

t µ

+

∂u ∂z

 µ

z

∂u ∂z

t (29) µ

The design sensitivity of the mean response vector, Eq. (26), is obtained by differentiation with respect to the design variables zk . Approximately: ∂µu ∼ ∂µu0 ∂u(b, µα ) = = ∂zk ∂zk ∂zk

(30)

For the covariance matrix of the response, Eq. (27): ∂u1 = ∂zk

∂u ∂α

 µ

∂α ∂zk

∂u ∂α



t µ

+2

∂u ∂α



∂ α ∂z k µ

∂u ∂α

t  (31) µ S

512

Structural design optimization considering uncertainties

On the right, the subscript S refers to the symmetric part of the matrix expression in the parentheses. Recall that in the expressions for µu and u random design variables are represented by mean value and variance resp. covariance. Usually, a displacement function is not available explicitly. In quasistatic deformation the response of the system is governed by the condition of equilibrium between the applied forces and the induced stresses. In finite element terms S(σ, X) = P(t, X)

(32)

The vector X = o X + u comprises the coordinates of the mesh nodal points in the displaced state, o X defines the original position. The stress in the interior is symbolized by σ, the stress resultants at the nodal points are collected in the vector S(σ, X). They are in equilibrium with the forces applied at the homologous positions, arranged in the vector P(t, X), a function of time t and possibly depending on the actual geometry. The stress σ as a consequence of the deformation depends on the material law applicable under the circumstances. This completes the sequence of operations that defines the response displacement u as an implicit function of the input parameters.

4 Elastic structures 4.1 L i nea r res po ns e In addition to the nodal point coordinates, structural systems require the input of member dimensions, which will be assumed collected in the array A in what follows. When the displacement does not change appreciably the geometry, the static equilibrium of the structure is established for the initial configuration o X. The stress resultants in the elastic system are S(σ, o A, o X) = K(κ, o A, o X)u

(33)

with K the stiffness matrix of the structure and κ that of the elastic material. In elasticity the stress σ is a function of the elastic constants in the matrix κ and of the strain that derives from the displacements in u. The entities in o X and κ may be deterministic, classified as elements of the vector b, or random, allocated to the vector α. The applied forces in P(t, o X) are taken at fixed level, but they may be fluctuating depending on parameters in α. For the elastic structure the equilibrium condition, Eq. (32), becomes K(b, α)u = P(b, α)

(34)

and defines the displacement u as a function of the parameter sets b and α. Then, the approximation to the mean vector and covariance matrix of the response displacement are obtained as follows. Eq. (34) written in the form Kµ µu0 = Pµ

(35)

Taylor approach to structural optimization and robustness

513

determines the zeroth-order approximate µu0 of the mean displacement. The quantities Kµ = K(b, µα ),

Pµ = P(b, µα )

(36)

refer to mean input. Differentiation of Eq. (34) with respect to the random variables αj yields the first partial derivatives of the displacement

Kµ



∂u ∂αj

= µ

∂P ∂K − u ∂αj ∂αj

 ,

j = 1, . . . , q

(37)

µ

All terms are evaluated at mean input and u = µu0 on the right-hand side. The solution of Eq. (37) extends over q right-hand vectors, each associated with a single random variable, to furnish the transformation operator in Eq. (27) for the displacement covariance matrix

∂u ∂α



 µ

=

∂u ∂u ∂u ··· ∂α1 ∂α2 ∂αq

µ

Differentiating once more Eq. (34) gives the second partial derivatives of u with respect to the variables in α. The second-order term to the mean displacement in Eq. (26) follows then from   q q   ∂2 u ∂2 P ∂2 K ∂K ∂u Kµ µu = Kµ σα α = − u−2 σα α ∂αi ∂αj µ i j ∂αi ∂αj ∂αi ∂αj ∂αi ∂αj µ i j 2

i,j=1

i,j=1

(38) The coefficient matrix in Eqs. (35) to (38) is uniquely the stiffness matrix Kµ of the elastic structure for mean (nominal) input. The analysis encompasses two matrix solutions with single vectors on the right-hand side, Eqs. (35) and (38), and one solution with q vectors on the right-hand side, equal to the number of the random input variables, Eq. (37). The design sensitivities of the displacement mean vector µu and of the covariance matrix u , Eq. (30) and Eq. (31), require differentiation of Eqs. (35) and (37) with respect to the design parameters zk . This gives for the mean Kµ

∂µu0 ∂Pµ ∂Kµ = − µ ∂zk ∂zk ∂zk u0

(39)

and for the covariance transformation Kµ

∂ ∂zk

∂u ∂αj

 = µ

∂ ∂zk

∂P ∂K − u ∂αj ∂αj

 − µ

∂Kµ ∂zk

∂u ∂αj

 , µ

j = 1, . . . , q

(40)

514

Structural design optimization considering uncertainties

The latter equation is solved for q right-hand side vectors to furnish the entire matrix operator needed in Eq. (31):

  ∂u ∂u ∂u ∂ ∂ ∂u = ··· ∂zk ∂α µ ∂zk ∂α1 ∂α2 ∂αq µ 4.2 L a rge d i sp lac ement s If the structure undergoes large displacements the equilibrium condition must be stated at the displaced geometry X = o X + u; changes in member dimensions are negligible under the small strain assumption (Doltsinis 2003). So Eq. (32) assumes here the nonlinear form S(σ, o A, o X + u) = P(t, o X + u)

(41)

Numerical solution of nonlinear systems relies on iterative techniques applied to the residual vector R = P − S which is a function of the displacement u and depends on the input parameters. When the displacement complies with equilibrium the residual function vanishes. At fixed time, in terms of random parameters α and deterministic parameters b for the input R(b, α, u) = P − S = 0

(42)

This furnishes the zeroth-order equation for the mean displacement µu as R(b, µα , µu0 ) = 0

(43)

The solution for the mean displacement µu0 is frequently obtained by the NewtonRaphson iteration scheme −(Gµ )n [(µu0 )n+1 − (µu0 )n ] = R[b, µα , (µu0 )n ]

(44)

where the indices n, n + 1 point on consecutive iterations and the coefficient matrix −G is the tangent operator −G = −

∂R ∂S ∂P = − = KT − KL ∂u ∂u ∂u

(45)

The contribution KT , the tangential stiffness matrix emanates from straining and kinematic effects on the stress resultants S(σ, X). Symbolically, KT =

∂S dσ ∂S ∂S = + = KM + KG ∂u ∂σ du ∂X

(46)

The matrix KM projects the material constitutive properties onto the structural level, the matrix KG is known as the geometric stiffness matrix of the structure. The load correction matrix KL accounts for geometry dependent applied forces: KL =

∂P ∂P = ∂u ∂X

(47)

Taylor approach to structural optimization and robustness

515

The tangent operator is taken here at the mean values of α and u. It replaces the elastic stiffness matrix of the linear problem. When the convergent solution of the zeroth-order equation is achieved, the operator is available for the first-order and the second-order approximations to follow. The first-order equations are obtained from Eq. (42) by differentiation:

−Gµ

∂u ∂αj



∂R ∂αj

= µ

 j = 1, . . . , q

,

(48)

µ

The transformation operator for the displacement covariance matrix in Eq. (27) requires solution of Eq. (48) for the q right-hand vectors associated with the single random variables. The equation for the second-order term to the mean displacement relies on the second derivatives of u with respect to the random variables in α. With reference to Eq. (26)  q  ∂2 u −Gµ µu = −Gµ σα α ∂αi ∂αj µ i j 2

i,j=1

  ∂2 R dG ∂u + σα α ∂αi ∂αj dαi ∂αj µ i j q

=

(49)

i,j=1

The total derivative on the right-hand side of the equation includes the implicit dependence of G(b, α, u) on the random parameters α via the displacements u; it is computed by finite differences. The equations for the nonlinear case are structured as Eqs. (35), (37) and (38) appertaining to linearity, the tangent operator G replacing here the stiffness matrix K of the linear elastic system. The derivatives on the structural level are actually assembled from individual finite elements contributions. Regarding the sensitivity with respect to the design variables z, Eq. (43) gives for the mean displacement −Gµ

∂µu0 = ∂zk

∂R ∂zk

 (50) u

the differentiation of the residual vector R = P − S on the right-hand side to be performed at u = const. The design sensitivity of the covariance matrix of the response displacement, Eq. (31), needs the derivative of the operator ∂u/∂α with respect to z. From Eq. (48), ∂ −Gµ ∂zk

∂u ∂αj

 µ

∂ = ∂zk

∂R ∂αj



∂Gµ + ∂zk µ

∂u ∂αj

 j = 1, . . . , q µ

Obviously, Eqs. (50) and (51) represent generalized forms of Eqs. (39) and (40).

(51)

516

Structural design optimization considering uncertainties

4.3 Nonlinear dynamics Incorporation of inertia effects complements Eq. (32) for the quasistatic deformation of the finite element system to the equation of motion Mu¨ + S(u) = P(t, u)

(52)

The vector u¨ = ∂2 u/∂t 2 comprises the acceleration at the mesh nodal points, M denotes the mass matrix. Damping forces induced by the velocity u˙ = ∂u/∂t are not included in Eq. (52). Approximate integration (Doltsinis 1999, 2003) links the kinematic quantities, acceleration u, ¨ velocity u˙ and displacement u, such that at the end of an increment in time τ = b t − a t: b

u˙ = a u˙ + τ(c1 a u¨ + c2 b u) ¨

b

u = a u + τ a u˙ + τ 2 (c3 a u¨ + c4 b u) ¨

(53)

The parameters c1 , c2 ; c3 , c4 control the performance of the numerical integration. They can be specified as to reproduce time stepping schemes known in the literature (Tamma et al. 2000). Utilization of Eq. (53) in Eq. (52) stated at current t = b t leaves an equation ¨ all kinematic quantities at past stage t = a t known. Explicit for the acceleration u¨ = b u, schemes (c2 = c4 = 0) ensure linearity of the incremental problem. The dynamic counterpart of the stochastic Eq. (42) at instant t = b t reads R(b, α, b u) − M b u¨ = 0

(54)

The random parameters α affect the acceleration and induce fluctuations in velocity and displacement, resp. in the deformed geometry. For the mean values Eq. (53) gives 1 2 µu˙ = a µu˙ + τ c1 (a µu¨ ) + c2 (b µu¨ ) 1 2 b µu = a µu + τ(a µu˙ ) + τ 2 c3 (a µu¨ ) + c4 (b µu¨ ) b

(55)

The respective operators that enter the first-order covariance transformation, Eq. (27), derive to ) * ∂a u˙ ∂a u¨ ∂b u¨ ∂b u˙ = + τ c1 + c2 ∂α ∂α ∂α ∂α ) * a b ∂a u ∂a u˙ ∂ u ¨ ∂ u ¨ ∂b u = +τ + τ 2 c3 + c4 (56) ∂α ∂α ∂α ∂α ∂α Thus the task reduces to the stochastic analysis of Eq. (54) for the acceleration. Solving the zeroth-order equation R(b, µα , b µu0 ) − Mb µu0 ¨ =0

(57)

Taylor approach to structural optimization and robustness

517

for b µu0 ¨ by the iteration scheme of Eq. (44) introduces the tangent operator −GD =

2 ∂ 1 b b M u ¨ − R(b, α, u) = M − τ 2 c4 G ∂b u¨

(58)

with the matrix G defined by Eq. (45). Differentiation of Eq. (54) with respect to the random parameters αj furnishes M

∂b u¨ ∂M b ∂R ∂R ∂b u + u¨ = + ∂αj ∂αj ∂αj ∂u ∂αj

(59)

Therefrom the first-order stochastic equations for b u¨ are deduced in conjunction with Eq. (56) for ∂b u/∂αj as ) −(GD )µ

∂b u¨ ∂αj



* = µ

)  * ∂R ∂b u ∂M b +G − u¨ ∂αj ∂αj ∂αj H

j = 1, . . . , q

(60)

µ

On the right-hand side of the equation, the subscript H refers to the historical contributions (t = a t) to the partial derivative ∂b u/∂α included on the left-hand side of the equation. The derivation of higher-order stochastic terms and of sensitivity expressions is straightforward. Their reproduction is suppressed here for typographical brevity.

5 Path dependence – elastoplasticity If the association of stress and strain is not unique but depends on the loading sequence, the material constitutive law is incremental in nature – as in elastoplasticity – specified by a momentary stiffness matrix κ. On the structural level this requests incrementation of the loading process and accumulation of the ensuing displacement (Doltsinis 1999). For the time step a t → b t: b

u = a u + u,

b

S = S(a σ + σ, a X + u, o A)

(61)

The displacement increment u is governed by the equilibrium condition, Eq. (32), stated at instant t = b t, the end of the current interval b

S = S(a σ, a X, o A, κ, u) = P(b t, a X + u) = b P

(62)

which may be contrasted to Eq. (41), the equilibrium condition for the elastic case. Usually the incremental approach is employed in elasticity as well for various reasons, comprising the algorithmic treatment of nonlinearities and the interest in the response throughout the course of the loading programme (Doltsinis 2003). The incremental counterpart of Eq. (42) in terms of the random parameters α and the deterministic parameters b is b

R(b, α, u) = b P − b S = 0

(63)

518

Structural design optimization considering uncertainties

The equation for the zeroth-order approximation µ u0 to the mean increment b

R(b, µα , µ u0 ) = 0

(64)

may be solved by an application of the iteration scheme of Eq. (44). The appertaining matrix operator −G = −

∂S ∂P ∂R = − = K T − KL ∂( u) ∂( u) ∂( u)

(65)

is set up for the state at instant t = b t. The tangent stiffness b

KT =

dS(a σ + σ, a X + u) db S ∂b S(b, α, u) = = b ∂( u) d( u) d u

(66)

is structured as the matrix KT of Eq. (46), but with the momentary elastoplastic material stiffness matrix κ in place of the elastic κ. Differentiation of Eq. (63) supplies the equations for the first-order terms:  −Gµ

∂( u) ∂αj

) = µ

∂b R ∂αj

* j = 1, . . . , q

(67)

µ

The equation for the second-order approximation 2 µ u to the mean incremental displacement follows analogously to Eq. (49): −Gµ 2 µ u = −Gµ

q  2  ∂ ( u) i,j=1

∂αi ∂αj

µ

σαi αj

  q  ∂2 (b R) dG ∂( u) = + σαi αj ∂αi ∂αj dαi ∂αj i,j=1

(68)

µ

The total derivative on the right-hand side of the equation accounts for the additional dependence of G(b, α, u) on the random parameters α via the incremental displacements u. It is evaluated by a finite difference scheme. The mean value of the displacement increment approximated to the second order µ u2 = µ u0 +

1 2

µ u 2

(69)

requires solution of Eqs. (64) and (68). The accumulated mean displacement at instant t = b t follows to b

µu = a µu + µ u

(70)

Taylor approach to structural optimization and robustness

519

The first-order approximate of the displacement covariance matrix for b u = a u + u reads 



b

u1

∂a u ∂( u) + = ∂α ∂α

µ

α

∂a u ∂( u) + ∂α ∂α

t

(71) µ

It requires q solutions of Eq. (67) which determine the respective columns of the derivative matrix ∂( u)/∂α. The historical term ∂a u/∂α results from past incremental accumulation. Analysis of Eq. (71) reveals the composition of the covariance matrix:  b

u1 = u1 + 2 a

∂a u ∂α

 µ

α

∂( u) ∂α

t  µ S

+  u1

(72)

The index S points on the symmetric part of the matrix expression within the parentheses. The sensitivity of the mean displacement with respect to the design variable zk is ∂ b µu ∂a µu ∂µ u = + ∂zk ∂zk ∂zk

(73)

From Eq. (64) by differentiation ∂µ −Gµ u0 = ∂zk

)

∂b Rµ ∂zk

* (74)

u

which furnishes an approximation to the advancement of the sensitivity in Eq. (73) within the increment. The design sensitivity of the covariance matrix of the ensuing displacements as from Eq. (31) reads here )

∂ u1 = ∂zk b

∂ u ∂α b

* µ

∂α ∂zk

)

∂ u ∂α b

⎡)

*t

+ 2⎣ µ

∂ u ∂α b

* α µ

∂ ∂zk

)

*t ⎤ ∂ u ⎦ ∂α b

µ

(75) S

Evaluation of the above expression requires determination of the derivative matrix ∂ ∂zk

)

∂b u ∂α

* µ

∂ = ∂zk

∂a u ∂α



∂ + ∂z k µ



∂( u) ∂α

(76) µ

specifically of the incremental change on the right-hand side. From Eq. (67) one obtains for the respective columns by differentiation, ∂ −Gµ ∂zk



∂( u) ∂αj

µ

∂ = ∂zk

)

∂b R ∂αj

* µ

∂Gµ + ∂zk



∂( u) ∂αj

j = 1, . . . , q µ

(77)

520

Structural design optimization considering uncertainties

Computation requires formation of the q right-hand side vectors and solution with the coefficient matrix −GT set up uniquelly with mean values at t = b t. That far fluctuations of the incremental response due to the appearance of a σ and a X in Eq. (62) have been disregarded. These quantities are represented by their mean values, which simplifies the stochastic analysis. For a rigorous treatment the argument of the function for the residual vector in Eq. (63) is complemented to read b

R(b, α, a σ, a X, u) = 0

(78)

This does not modify the determination of the mean displacement increment to the zeroth order by Eq. (64) except of the explicit appearance here of a µσ and a µX . But, the first-order equation for the covariance operator, Eq. (67), now becomes   



 ∂( u) ∂b R b ∂S ∂a σ b ∂R ∂a X −Gµ = − + j = 1, . . . , q (79) ∂αj µ ∂αj ∂σ ∂αj ∂X ∂αj µ

On the right-hand side, the stress term in the middle can be computed by utilizing the stress resultants module S(σ, X) with modified arguments: S(∂a σ/∂αj ,b X). The operator to the last term on the right is identified as ∂P ∂S ∂R = − = KL − KG ∂X ∂X ∂X

(80)

with the matrices KL and KG defined by Eqs. (46), (47). The incremental computation requires, in addition, the updates ∂a σ ∂a σ ∂( σ) ⇐ + , ∂αj ∂αj ∂αj

∂a X ∂a X ∂( u) ⇐ + ∂αj ∂αj ∂αj

(81)

6 Applications 6.1 Perf o rm anc e o f t he inc r ement al ap p r o a ch The incrementation procedure is exemplified for the planar 10-bar cantilever truss depicted in Figure 18.8. The problem exhibits geometrical and material nonlinearity. The two equally increasing forces P3y , P5y acting vertically at nodes no.3 and no.5 of the truss induce large displacements. The bar members are assumed of elastic–material with yield stress 250.0, the relationship between the stress σ and the strain ε being  dσ =

1.0 × 104 dε (250.0/|ε|)dε

elastic range elastic–plastic

The mass density of the material is taken as # = 1.0. The horizontal position x4 of node no.4, the elastic moduli grouped as EI (bars 1,2), EII (bars 3,4), EIII (bars 5,6), EIV (bars 7,8,9,10) and the cross-section areas A of the bar members are considered as random variables characterized by the mean value

Taylor approach to structural optimization and robustness

360

360 (2)

4

(6)

(5) (8)

1

(7)

(10)

3

(3)

6

360

(1)

2

521

(9)

(4)

P3y

5 P5y

Figure 18.8 Planar 10-bar cantilever truss.

Mean (present) Mean (MCS) Std. dev. (present) Std. dev. (MCS)

Mean displacement

150

30

100

20

50

10

0

0

500

1000 Load magnitude

1500

Std. dev. of displacement

40

200

0

Figure 18.9 Mean and standard deviation of nodal displacement v 5 .

µ and variance σ 2 or coefficient of variation σ/µ. The mean values and variances for the distinct parameters are specified as µx4 = 360.0, σx24 = 400.0, µE = 1.0 × 104 for all groups, σE2 = 9.0 × 104 for groups I, II, III, σE2 = 1.0 × 106 for group IV, µAi = 5.0 (i = 1, 2, . . . , 10), σAi /µAi = 0.05 (i = 1, 2, . . . , 6), σAi /µAi = 0.1 (i = 7, . . . , 10). The performance of the incremental approximation to the nonlinear stochastic problem is commented with reference to Figures 18.9–18.11. Figure 18.9 shows the second-order mean and the first-order standard deviation of the vertical displacement v5 of node no.5 as a function of the magnitude of the loading, which is increased by 10 equidistant incremental steps. The results of the incremental Taylor-series approximation are compared with results from synthetic Monte Carlo sampling with 3000 realizations of the random input. It is seen from the figure that the mean values agree well; the difference in the standard deviation may arise because the present, simplified

Structural design optimization considering uncertainties

700

70 Mean (present) Mean (MCS) Std. dev. (present) Std. dev. (MCS)

Mean max. member stress

600 500

60 50

400

40

300

30

200

20

100

10

0

0

500

1000 Load magnitude

Std. dev. of max. member stress

522

0

1500

Sensitivity of mean

0

0

2

2

4

4

6

6

8

Sensitivity of mean (MCS) Sensitivity of mean (present) Sensitivity of Standard Deviation (MCS) Sensitivity of Standard Deviation (present)

10 12

0

500

1000 Load magnitude

1500

8 10 12

Sensitivity of Standard Deviation

Figure 18.10 Mean and standard deviation of maximum member stress.

Figure 18.11 Sensitiviy of mean and standard deviation of displacement v 5 .

approach neglects past scatter in stress and geometry. Figure 18.10 refers to the evolution of the maximum member stress during the course of loading. The figure shows that both techniques yield here practically the same results for mean value as well as for standard deviation. The design pays attention to the vertical displacement v5 of node no.5. The sensitivity of mean value and standard deviation of this quantity with respect to the cross-section area of bar no.5 are plotted in Figure 18.11 versus the magnitude of the loading. The sensitivities obtained from Monte Carlo simulations base on finite difference approximations. From both, Figure 18.11 and Figure 18.9, moderate differences to the simplified incremental stochastic computation appear after a number of incremental steps and are seen to accumulate slowly. The design objective in this example problem is to minimize the vertical displacement v5 at node no.5, at loading P3y = P5y = 1000.0. The design variables are the mean cross-section areas of the bars, restricted by 0.05 ≤ µAi ≤ 10.0 (i = 1, 2, . . . , 10).

Taylor approach to structural optimization and robustness

523

Table 18.1 Optimal results for 10-bar truss.

A1 A2 A3 A4 A5 A6 A7 A8 A9 A10 µf σf µW × 104

Init.

(Determ.)

ξ=0

ξ = .25

ξ = .5

ξ = .75

ξ = 1.0

4.00 4.00 4.00 4.00 5.00 5.00 4.00 4.00 4.00 4.00 92.35 6.13 2.16

(9.14) (0.05) (8.83) (6.05) (0.05) (0.05) (4.07) (7.61) (6.54) (0.05) (56.30) (3.53) (1.80)

9.54 0.05 8.02 4.92 0.31 0.05 4.80 6.90 7.42 0.05 59.30 3.76 1.80

8.34 0.05 7.83 4.82 0.27 0.05 4.93 7.30 7.96 0.06 59.62 3.55 1.80

7.15 0.58 7.98 4.21 0.05 0.54 5.54 6.85 7.46 1.01 61.68 3.38 1.80

6.73 0.60 7.89 3.61 0.05 0.61 5.65 7.26 7.51 1.15 62.66 3.30 1.80

6.19 0.81 7.76 3.13 0.05 0.78 5.75 7.21 7.45 1.70 64.51 3.27 1.80

The structural weight constraint µW ≤ 1.8 × 104 and the member stress constraints µ|σi | + 3σσi ≤ 400 (i = 1, 2, . . . , 10) are to be observed. The optimal solutions obtained with different values of the weighting factor ξ in the design desirability function are listed in Table 18.1. The value ξ = 0 is associated with the stochastic mean value minimization, ξ = 1.0 appertains to the pure variance minimization problem. The stochastic mean value design (ξ = 0) exhibits a standard deviation of the objective σf = σv5 = 3.76 which reduces to σv5 = 3.27 for the most robust design (ξ = 1). At the same time the mean value increases from µv5 = 59.30 to µv5 = 64.51, respectively. Increasing the weighting factor ξ is found to diminish the variability of the observed performance measure at the penalty of an increasing mean value. Inspection of the optimum solutions for ξ = 0 and ξ = 1 in Table 18.1 may suggest different design topologies if the bar members with minor cross-section area are discarded. In fact, according to mean value optimization node no.6 (upper corner on the right) and the adjanced bar elements could be eliminated; the robust design suggests that only the vertical bar 5 (middle of the truss) is unnecessary. Thereby it becomes evident that the more robust design requires additional structural members. The deterministic optimum (nominal i.e. mean input values; no randomness) is equivalent to a mean value optimization with the zeroth-order approximation, only if the constraints are free of variances. In the present case, the deterministic result violates the stress constraints. For instance, the mean and standard deviation of bar 7 are 375.3 and 40.0, respectively, such that the stochastic stress constraint 375.3 + 3 × 40.0 > 400 exceeds the allowed value.

6.2

Structural compliance optimizatio n o f a 25-bar s p ac e trus s

The space truss under investigation resembles a power transmission power with topology as shown in Figure 18.12. The geometry of the truss is defined by the nodal point coordinates in Table 18.2. The task to be performed is the minimization of the structural compliance, the work performed by the applied forces on the displacements which

524

Structural design optimization considering uncertainties

9

6

22 14

1

1

8

54 32

10 3

15 23

2

12 5

13

7

6

11 4

18

20

19 25 16

21

17 24

7 10

8

Z Y

9

X

Figure 18.12 Twentyfive-bar truss. Table 18.2 Nodal coordinates of 25-bar truss. Node

x

y

z

1 2 3 4 5 6 7 8 9 10

−37.5 37.5 −37.5 37.5 37.5 −37.5 −100.0 100.0 100.0 −100.0

0.0 0.0 37.5 37.5 −37.5 −37.5 100.0 100.0 −100.0 −100.0

200.0 200.0 100.0 100.0 100.0 100.0 0.0 0.0 0.0 0.0

defines the objective, the performance function f = Pt u. The design variables are the cross-section areas of the 25 bars which are reduced to six independent variables by allocating equivalent members to the groups shown in Table 18.3. For each of the six groups, the common cross-section and the elastic modulus are random quantities with mean value µ and standard deviation σ resp. coefficient of variation (COV = σ/µ) as given in Table 18.4. The elastic range of the structural material is limited by the yield stress σs = Eεs with εs = 0.003. The stress-strain relationship is  Edε elastic range dσ = E(εs /|ε|)dε elastic–plastic. The mass density of the material is # = 0.1.

Taylor approach to structural optimization and robustness

525

Table 18.3 Group membership for 25-bar truss. Group number

Bar members

I II III IV V VI

1 2,3,4,5 6,7,8,9 10,11,12,13 14,15,16,17,18,19,20,21 22,23,24,25

Table 18.4 Random variables for the nonlinear 25-bar truss. No.

Variable

Mean

Std. deviation

1–5 6 7 8 9–14

EI −EV EVI P 3x P 6x AI −AVI

1.0 × 107 1.0 × 107 5.0 × 103 5.0 × 103

2.0 × 105 1.5 × 106 5.0 × 102 5.0 × 102

COV

0.05

Table 18.5 Optimal solution for the nonlinear 25-bar truss.

AI AII AIII AIV AV AVI µf (× 105 ) σf (× 105 ) µW (×102 )

Init.

(Determ.)

ξ=0

ξ = .25

ξ = .5

ξ = .75

ξ=1

2.500 2.500 2.500 2.500 2.500 2.500 6.460 0.639 8.268

(1.956) (2.594) (3.766) (1.247) (1.138) (6.286) (4.344) (0.293) (8.489)

0.050 3.608 3.692 0.794 1.315 5.409 4.322 0.316 8.500

1.879 2.093 3.468 0.920 1.312 6.758 4.374 0.272 8.500

1.822 1.798 3.374 0.979 1.366 6.948 4.424 0.269 8.500

1.483 1.546 3.463 1.060 1.368 7.122 4.430 0.265 8.500

2.116 1.929 3.333 1.409 1.282 6.800 4.435 0.261 8.500

The loading induces geometrical nonlinearities. It consists of forces P1x = P1y = P2x = P2y = −1.0 × 105 with fixed magnitude imposed at nodes no.1, no.2 along the x- and the y-direction, and of randomly fluctuating forces P3x , P6x acting at nodes no.3, no.6 along the x-direction with mean value 5.0 × 103 (Table 18.4). The structural weight is constrained by µW ≤ 850, the member stress by µ|σi | + 3σσi ≤ 4.0 × 104 (i = 1, 2, . . . , 25). The bounds for the design variables are 0.05 ≤ Aj ≤ 10.0 (j = I, . . . , VI). Design optimization based on the desirability function F(µf , σf , ξ) is employed in conjunction with the incremental stochastic analysis. The initial design, the deterministic optimum and the optimal solutions for different values of the weighting factor ξ are listed in Table 18.5. The deterministic optimum refers to nominal, or mean input. It is equivalent to mean value optimization (ξ = 0) when the mean is approximated to the zeroth-order. An augmented factor ξ puts weight to the minimization of the standard deviation of the objective while the significance of the mean value diminishes. It

526

Structural design optimization considering uncertainties

0.33

σ (105)

0.31

0.29

0.27

0.25 4.3

4.35

4.4

4.45

 (105)

Figure 18.13 Pareto set for the non-linear 25-bar truss. Pairs of standard deviation σ and mean µ of the objective that minimize the desirability function.

Figure 18.14 Antenna structure (example).

is seen from Table 18.5 that the robust solutions (ξ > 0) reduce the standard deviation of the objective by 14–18% as compared to ξ = 0. The constraints posed are satisfied throughout. Mean values and standard deviations of the objective appertaining to the optimum solutions that minimize the desirability function at distinct values of the weighting factor ξ, form the Pareto set depicted in Figure 18.13. 6.3 Antenna st r uc t ur e und er lar g e d is p l a ce m e n t s The antenna (radius 2.8 m, height 2.5 m, Figure 18.14) is considered under quasi-static wind loading. Thereby, the elastic structure undergoes large displacements which cause shape distortion and affect the pointing accuracy of the reflection surface. The aim of the design is to minimize displacements and to ensure robustness against fluctuating input.

Taylor approach to structural optimization and robustness

527

q = 6000 N/m2

Figure 18.15 Layout of the antenna structure model and sideload. Table 18.6 Optimal solution for the antenna structure. Design variable −6

A1 (×10 m2 ) A2 (×10−6 m2 ) A3 (×10−6 m2 ) t (×10−3 m) µW (kg) µf (×10−3 m) σf (×10−3 m)

Lower

Upper

Initial

ξ = 0.0

ξ = 0.5

ξ = 1.0

10.00 10.00 10.00 2.00 / / /

25.00 25.00 25.00 3.50 350.0 / /

15.00 15.00 15.00 2.80 331.5 102.7 7.4

13.30 13.62 12.53 3.02 350.0 87.28 6.9

25.00 18.47 18.63 2.68 350.0 90.56 5.9

25.00 19.15 18.08 2.65 350.0 91.27 5.8

The finite element model of the simplified antenna structure consists of a complex truss framework in three-dimensional space, which is covered by a membrane skin (Figure 18.15). The wind loading is modelled as distributed surface force that acts sidewards with a magnitude of 6000 N per square meter of projection area. The bars forming the truss structure are collected into five groups according to their position and orientation: the circumferential, radial and skew members supporting the skin on the reflection surface, short and long bars at the bottom part. In the model the skin cover consists of four peripheral regions. The uncertain parameters accounted for are those having major effects. They are considered as uncorrelated random variables that comprise: the five elastic moduli of the grouped bar members, the elastic modulus of the skin material, the skin thickness in each of the four regions of the cover. The mean value of the elastic moduli is µE = 7.1 × 1010 N/m2 , the coefficient of variation is µ/σ = 0.1 for all random quantities. The design variables are the skin thickness t and the cross-section areas A1 , A2 , A3 of the bars belonging to the circumferential, the radial and the skew group at the reflection part of the antenna, respectively. The design objective is to minimize the largest radial displacement at the upper edge of the antenna structure under the constraint of total structural weight µW ≤ 350.0 kg and restriction of the design variables (Table 18.6).

528

Structural design optimization considering uncertainties

The optimum solutions obtained for different values of the weighting factor ξ are listed in Table 18.6. Optimization with ξ = 0.5 reduces the variability of the initial design and enhances the robustness markedly beyond the result of mean value minimization (ξ = 0). The further improvement with ξ = 1.0 is not significant, however.

7 Conclusions Structural analysis and design accounting for random input data necessitate employment of computational procedures which are different from deterministic methods but base thereon. The prevalent approaches to stochastic structural analysis can be classified either as synthetic, statistical techniques like the Monte Carlo sampling, or as nonstatistical, analytical methods based on Taylor series expansion. The former involve sampling of the input statistics, function evaluation and synthesis of the output statistics. The latter methods use functional expansion presuming smoothness and differentiability. By its nature, the Taylor approach is restricted to the analysis of systems with moderate parameter variability; it provides the analyst with a powerful tool, however. The computational effort is determined by the number of random input variables in contrast to the number of sampling units to be evaluated until statistical convergence in the Monte Carlo technique. In turn, the latter supplies the complete output statistics for the sample, while Taylor expansion describes random input and output up to the second statistical moment regardless of the actual distribution. Methods of stochastic analysis are surveyed in (Schuëller 2001). Robust performance in service as considered here has to be contrasted to the reliability task which concerns failure. In the design issue the Taylor approach allows for the derivation of sensitivity expressions useful in utilizing gradient based optimization algorithms. Optimizing for mean performance does not differ from the analogous deterministic task, in principle, but robustness against uncertain factors observes also the random variability of the performance as measured by the standard deviation. Thereby a two-objective problem is posed, but standard algorithms are applicable in conjunction with the desirability function, a scalar substitute that compromises the demand for both best mean performance and least variability. While the desirability function can be defined in various ways, adjustment of robustness is a property inherent in the system. Given the input scatter, the variability of the output can be influenced only by the input level. If the input level has no effect, there is no issue of robustness. The argument is elucidated by considering a performance function as follows f (z, α) = a + bt z + zt Bz + ct α + zt Cα

(82)

with design variables in the vector z and random input in the vector α. This gives for the variance

t

 ∂f ∂f 2 ∼ α = (c + Cz)t α (c + Cz) (83) σf = ∂α µ ∂α µ the matrix C assumed symmetric. It turns out that the last term in Eq. (82) is decisive for robustness. Accordingly, empirical approximations of the performance as a function

Taylor approach to structural optimization and robustness

529

of the design variables should explore interaction with the random input. In general terms, designing for robustness requires that the variance of the performance function is sensitive to the design variables. In this chapter, stochastic analysis has been based on second-order expansion about the mean of the random input variables. The mean response to the second order is of relevance when considering a series of products. For assessing single events the response for nominal input, the zeroth-order mean, appears a suitable reference in conjunction with the scatter about this level. Scatter is characterized by the covariance matrix, represented by the first-order approximation. The unified methodology worked out encompasses linear elastic structures, geometric and material nonlinearity; path dependence implies incrementation of the loading process. The associated algorithm extends standard finite element procedures. It supplies mean vector and convariance matrix of the response variables, the displacements, and evaluates design sensitivity expressions. The results serve as input to conventional optimization that minimizes the desirability function, a compromise between mean value and variability of the performance measure. The importance put on each criterion has to be specified in advance by the analyst. The designer will benefit from a diversity of importance settings, however. Numerical results obtained with the proposed analytic approach agree well with synthetic (Monte Carlo) sampling. Apart from the applicability of the technique the comparisons indicate superiority with respect to the computational effort. This should not divert from the particularities of the method, however, which is local in nature. Large input scatter as well as non-smooth response, for instance, necessitate employment of alternative techniques, like the statistical sampling. The latter proves useful in exploring a wider range of parameter variation and is not restricted by discontinuities. At an early stage, exploration of the design space will be more important than absolute optimization to follow in the subsequent specification of details. Referring to the exposition in (Doltsinis 2003) a design package should allow access to both the global Monte Carlo technique and the local Taylor approach. The development of the presented method relies on the availability of discrete random variables. Continuous fluctuations within the domain like material properties or geometrical dimensions have not been a subject. Such conditions define random fields which can be described as α(x) = µα (x) + β(x)

(84)

and raise the task of discretized representation in the context of finite element simulations. In Eq. (84) the spatially fluctuating part β(x) = α(x) − µα (x) with zero mean defines the randomness of α(x). This description proves particularly useful for random fields with a constant mean µα = µα (x) like the linear elastic properties of homogeneous materials. Incorporation of random fields in finite element expressions can be effected in various manners. The mean point method for instance bases on the field value at the mid-point of the finite element, the local averaging technique represents the field by its element average, the weighted integral method employs weighted integrals over the individual element domains. For a farther reaching brief outline refer to (Der Kiureghian and Zhang 1999).

530

Structural design optimization considering uncertainties

Acknowledgment The author is indebted to Mrs. Knapp Christiansen for linguistic assistance and for the preparation of the manuscript in LaTeX.

References Breipohl, A.M. 1970. Probabilistic Systems Analysis. Wiley, New York. Der Kiureghian, A. & Zhang, Y. 1999. Space-variant finite element reliability analysis, Comput. Methods Appl. Mech. Engrg. 168:173–183. Doltsinis, I. 1999. Elements of Plasticity – Theory and Computation, WIT Press, Southampton. Doltsinis, I. 2003. Inelastic deformation processes with random parameters – methods of analysis and design. Comput. Methods Appl. Mech. Engrg. 192:2405–2423. Doltsinis, I. 2003. Large Deformation Processes of Solids – From Fundamentals to Computer Simulation and Engineering Applications, WIT Press, Southampton. Doltsinis, I. & Kang, Z. 2004. Robust design of structures using optimization methods. Comput. Methods Appl. Mech. Engrg. 193:2221–2237. Doltsinis, I. & Rodic, T. 1999. Process design and sensitivity analysis in metal forming. Int. J. Numer. Meth. Engrg. 45:661–692. Doltsinis, I. (ed.) 1999. Stochastic Analysis of Multivariate Systems in Computational Mechanics and Engineering, CIMNE, Barcelona. Doltsinis, I., Kang, Z. & Cheng, G. 2005. Robust design of non-linear structures using optimization methods. Comput. Methods Appl. Mech. Engrg. 194:1779–1795. Hurtado, J.E. 2004. Structural reliability – Statistical learning perspectives, Springer-Verlag, Berlin Heidelberg. Kang Z. 2005. Robust Design Optimization of Structures under Uncertainties, Doctoral Thesis, University of Stuttgart. Kleiber, M. & Hien, T.D. 1992. The Stochastic Finite Element Method, Wiley, Chichester. Lawrence, C., Zhou, J.L. & Tits, A.L. 2007. User’s guide for CFSQP version 2.5. Available from http://www.aemdesign.com. Lee, K.H. & Park, G.J. 2001. Robust optimization considering tolerances of design variables, Computers and Structures 7:77–86. Luo, X. & Grandhi, R.V. 1997. ASTROS for reliability-based multidisciplinary structural analysis and optimization. Computers and Structures 62:737–745. Myers, H. & Montgomery, D.C. 1995. Response Surface Methodology, Wiley, New York. Rencher, A.C. 1995. Methods of Multivariate Analysis, Wiley, New York. Schuëller, G.I. 2001. Computational stochastic mechanics–recent advances, Computers and Structures 79:2225–2234. Stadler, W. 1984. Multicriteria optimization in mechanics (A Survey). Appl. Mech. Rev. 37: 277–286. Tamma, K.K., Zhou, X. & Sha, D. 2000. The time dimension: a theory towards the evolution, classification, characterization and design of computational algorithms for transient/dynamic applications. Archives of Computational Methods in Engineering 7:67–290.

Chapter 19

Info-gap robust design of passively controlled structures with load and model uncertainties Izuru Takewaki Kyoto University, Kyoto, Japan

Yakov Ben-Haim Technion, Haifa, Israel

ABSTRACT: A new structural design concept is developed which incorporates uncertainties in both the load and the structural parameters. Info-gap models of uncertainty (non-probabilistic uncertainty models) are used to represent uncertainties in the Fourier amplitude spectrum of the load (input ground acceleration) and in parameters of the vibration model of the structure. Since non-probabilistic uncertainties are prevalent in many situations, this chapter shows that it is necessary to satisfy critical performance requirements (rather than to optimize performance), and to maximize the robustness to uncertainty. Earthquake input energy to passively controlled structures is introduced as a new measure of structural performance. It is usual that, while structural properties and performances of ordinary structural systems are well recognized through vast experiences and extensive databases, those of control devices added to those ordinary structural systems are not necessarily reliable. It may therefore be reasonable to consider uncertainties of damping coefficients of control devices. The design implications of the robust-satisficing approach are demonstrated with these passively controlled structures.

1 Introduction Load and structural model uncertainties are two major sources of actual uncertainties in the design of civil, mechanical and aerospace structures. While simultaneous consideration of both the load and structural model uncertainties is very important and challenging, only a limited number of publications can be found on this subject (Cherng and Wen 1994, Ghanem and Spanos 1991, Igusa and Der Kiureghian 1988, Jensen and Iwan 1992, Jensen 2000, Katafygiotis and Papadimitriou 1996). Because civil engineering structures are not mass-produced and the occurrence rate of large earthquakes and other severe disturbances is very low, the probabilistic representation of the effect of these disturbances on structural systems seems to be difficult in most cases. The critical excitation method is one of the promising strategies for overcoming difficulties in modeling the non-probabilistic load uncertainty (Drenick 1970, Shinozuka 1970, Ben-Haim and Elishakoff 1990, Takewaki 2001a, b, 2002a, b, 2004, 2006, Westermo 1985). In most of these critical excitation methods except Westermo (1985) and Takewaki (2004, 2006), deformation or displacement parameters were treated as response performance functions defining the criticality of the loads. In this chapter, the earthquake input energy to passively controlled structures is introduced as a new measure of structural performance. This is because some control devices have limitation on

532

Structural design optimization considering uncertainties

energy dissipation capacity and modeling. It is usual that, while structural properties and performance of ordinary structural systems are well recognized through a lot of experiences and database, those of control devices added to those ordinary structural systems are not necessarily reliable. In this situation, it may be reasonable to consider uncertainties of damping coefficients of control devices and use the earthquake input energy to those passively controlled structures as a new measure of structural performance. The purpose of this chapter is to propose a new structural design concept which incorporates uncertainties in both the load and the structural parameters. For that purpose, it is necessary to identify the critical load (excitation) and the corresponding critical set of structural model parameters. It is clear that the critical load (excitation) depends on the structural model parameters and it is extremely difficult to deal with load and structural model parameter uncertainties simultaneously. Info-gap models of uncertainty (non-probabilistic uncertainty models) by Ben-Haim (1996, 2001, 2005, 2006) are used to represent uncertainties in the Fourier amplitude spectrum of the load (input ground acceleration) and in parameters of the vibration model of the structure.

2 Info-gap uncertainty analysis As an example, let us consider that damping coefficients ci are very uncertain and can be expressed in terms of the nominal values c˜ i and the unknown uncertainty level α as shown in Figure 19.1 (Takewaki and Ben-Haim, 2005).      ci − c˜ i    C(α, c˜ ) = c :  ≤ α, i = 1, . . . , N , α ≥ 0 (1) c˜ i  The info-gap model C(α, c˜ ) is not a single set of damping coefficients, but rather an unbounded family of nested sets of coefficients. Let F(ω, c, k) denote the ‘energy transfer function’ defined in the following section. The energy transfer function is a function of the damping coefficients ci and the following info-gap model, which is an unbounded family of nested sets of functions, may be introduced in terms of the nominal function F˜ corresponding to the nominal damping coefficients c˜ i .       ˜ = F(ω, c, k) :  ci − c˜ i  ≤ α, i = 1, . . . , N , α ≥ 0 F(α, F) (2)  c˜  i

Realizable region of damping coefficient ac~i 0

ac~i ~ ci

ci Nominal value of damping coefficient

Figure 19.1 Uncertain damping coefficient with unknown horizon of uncertainty α.

I n f o-g a p r o b u s t d e s i g n o f p a s s i v e l y c o n t r o l l e d s t r u c t u r e s

533

The following family of sets of functions may also be considered for the definition of the info-gap uncertainty model.

2.1

˜ = {F(ω, k) : |F(ω, k) − F(ω, ˜ F∗ (α, F) k)| ≤ α}, α ≥ 0

(3)

˜ = {F(ω, k) : |F(ω, k) − F(ω, ˜ F∗∗ (α, F) k)| ≤ αψ(ω)}, α ≥ 0

(4)

Inf o-gap robus tnes s function

The info-gap robustness is the greatest horizon of uncertainty, α, up to which the performance function f (c, k) does not exceed a critical value, fC . Let us define the following info-gap robustness function corresponding to the info-gap uncertainty model represented by equation (1). 

 α(k, ˆ fC ) = max α : { max f (c, k)} ≤ fC

(5)

c∈C(α,˜c)

Another info-gap robustness function corresponding to the info-gap uncertainty model by equation (2) may be introduced by 



α(k, ˆ fC ) = max α : { max f (c, k)} ≤ fC

(6)

˜ F∈F(α,F)

Info-gap robustness function

Let fC0 = f (˜c, k). Then one can show that α(k, ˆ fC0 ) = 0, as shown in Figure 19.2. We define α(k, ˆ fC ) = 0 if fC ≤ fC0 (see Figure 19.2). The definitions in equations (5) and (6) imply that the robustness is the maximum level of the structural model parameter uncertainty, α, satisfying the performance requirement f (c, k) ≤ fC for all admissible variation of the structural model parameter represented by equation (1) or (2).

Large robustness

aˆ 2

aˆ 1

0

Small robustness

fC0

fC1

fC2

Design requirement

Figure 19.2 Info-gap robustness function αˆ with respect to design requirement f C .

534

Structural design optimization considering uncertainties

m(üg  x)

m(üg  x) ug dt

Figure 19.3 Free-body diagram for defining input energy.

3 Earthquake input energy to SDOF system Much work has accumulated on the topics of the earthquake input energy since the work by Housner (1959). In contrast to most of the previous works (Akiyama 1985), the earthquake input energy is formulated here in the frequency domain (Page 1952; Lyon 1975; Ordaz et al. 2003) to facilitate the derivation of a bound of the earthquake input energy. Consider a damped linear single-degree-of-freedom (SDOF) system of mass m, stiff ness k and damping coefficient c. Let ! = k/m, h = c/(2!m) and x denote the undamped natural circular frequency, the damping ratio and the displacement of the mass relative to the ground, respectively. The time derivative is denoted by overdot. The input energy to the SDOF system by a uni-directional ground acceleration u¨ g (t) = a(t) from t = 0 to t = t0 (end of input) can be defined by the work of the ground on the structural system and is expressed by
t0 EI = m(u¨ g + x) ¨ u˙ g dt (7) 0

The term −m(u¨ g + x) ¨ with minus sign indicates the inertial force on the mass and is equal to the sum of the restoring force kx and the damping force cx˙ in the system as shown in Figure 19.3. Integration by parts of equation (7) provides
t0
t0 2t0 1 EI = m(x¨ + u¨ g )u˙ g dt = mx¨ u˙ g dt + (1/2)mu˙ 2g 0 0 0 (8)
t0 2 1 t @ At0 2 0 = mx˙ u˙ g 0 − mx˙ u¨ g dt + (1/2)mu˙ g 0

0

If x˙ = 0 at t = 0 and u˙ g = 0 at t = 0 and t = t0 , the input energy can be reduced to the following form.
t0 mu¨ g x˙ dt (9) EI = − 0

For example, consider the recorded ground motion of El Centro NS 1940 (Imperial Valley) shown in Figure 19.4. The time history of the input energy per unit mass is shown in Figure 19.5. This was computed using equation (9) by regarding t0 as t.

I n f o-g a p r o b u s t d e s i g n o f p a s s i v e l y c o n t r o l l e d s t r u c t u r e s

535

Acceleration (g)

0.4 0.2 0 0.2 0.4 0

10

20

30

40

50

Time (s)

Input energy per unit mass (J/kg)

Figure 19.4 Ground motion of El Centro NS 1940 (Imperial Valley).

1.2 1 0.8 0.6 0.4 0.2 0

0

10

20

30 Time (s)

40

50

Figure 19.5 Time history of input energy under El Centro NS 1940 (Imperial Valley).

It is known (Page 1952; Lyon 1975; Ordaz et al. 2003; Takewaki 2004, 2006) that the input energy expressed by equation (9) can also be expressed in the frequency domain.

EI /m = = = =




∞ 1 ˙ iω t dω a dt π − xa ˙ dt = − Xe −∞ −∞ 2 −∞ 
∞
∞ D E 1 iω t ae dt HV (ω; !, h)A(ω) dω − π 2 −∞ −∞
∞ D E 1 − π A(−ω) HV (ω; !, h)A(ω) dω 2 −∞
∞ D E |A(ω)|2 −Re[HV (ω; !, h)]/π dω


∞




∞

(10)

0


≡

∞

|A(ω)|2 F(ω) dω

0

˙ where HV (ω; !, h) is the velocity transfer function defined by X(ω) = HV (ω; !, h)A(ω) ˙ and F(ω) = −Re[HV (ω; !, h)]/π. The functions X and A(ω) are the Fourier transforms

536

Structural design optimization considering uncertainties

Fourier amplitude of acceleration (m/s)

3 2.5 2 1.5 1 0.5 0

0

20

40

60 80 100 120 Frequency (rad/s)

140

160

Figure 19.6 Fourier amplitude spectrum of ground acceleration of El Centro NS 1940 (Imperial Valley).

of x˙ and u¨ g (t) = a(t), respectively. The symbol i denotes the imaginary unit. HV (ω; !, h) can be expressed by HV (ω; !, h) = −iω/(!2 − ω2 + 2ih!ω)

(11)

Equation (10) indicates that the earthquake input energy to damped linear elastic SDOF systems does not depend on the phase of input motions and this fact is well known (Page 1952, Lyon 1975, Ordaz et al. 2003; Takewaki 2004). Figure 19.6 shows the Fourier amplitude spectrum |A(ω)| of El Centro NS 1940.

4 Earthquake input energy to MDOF system Consider next a damped linear elastic multi-degree-of-freedom (MDOF) shear building model of mass matrix [M] subjected to a uni-directional horizontal ground acceleration u¨ g (t) = a(t). The present method can be applied to both proportionally damped and non-proportionally damped structures. Let {x} denote a set of the nodal horizontal displacements relative to the ground. The time derivative is denoted by over-dot. The input energy to the MDOF system by the ground motion from t = 0 to t = t0 (end of input) can be defined by the work of the ground on the MDOF system and is expressed by


t0

EI =

{1}T [M]({1}u¨ g + {x}) ¨ u˙ g dt

(12)

0

where {1} = {1 · · · 1}T . The term {1}T [M]({1}u¨ g + {x}) ¨ indicates the minus sign of the sum of the horizontal inertial forces acting on the system as shown in Figure 19.7. Integration by parts of equation (12) provides 1

EI = (1/2){1}

T

[M]{1}u˙ 2g

2t0 0

@

+ {x} ˙ [M]{1}u˙ g T

At0 0




t0

− 0

{x} ˙ T [M]{1}u¨ g dt

(13)

I n f o-g a p r o b u s t d e s i g n o f p a s s i v e l y c o n t r o l l e d s t r u c t u r e s

537

Sum of inertial forces

 {1} T [ M ]({1} u g { x})

dug = ug dt : base displacement during dt

ug

Figure 19.7 Free-body diagram for defining input energy to MDOF model.

If {x} ˙ = {0} at t = 0 and u˙ g = 0 at t = 0 and t = t0 , the input energy can be reduced to the following form:


t0

EI = −

{x} ˙ T [M]{1}u¨ g dt

(14)

0

The input energy can also be expressed in the frequency domain as in the SDOF ˙ denote the Fourier transform of {x}. system. Let {X} ˙ Application of the Fourier inverse transformation of the relative nodal velocities {x} ˙ to equation (14) leads to
EI = −

∞ −∞

= −

1 2π

= −

1 2π






1 2π ∞

−∞
∞ −∞




∞ −∞

˙ T eiωt dω [M]{1}u¨ g dt {X} 


˙ T [M]{1} {X}

∞ −∞

u¨ g eiωt dt dω

(15)

˙ T [M]{1}A(−ω)dω {X}

A(ω) is the Fourier transform of ground acceleration u¨ g (t) = a(t) and the symbol i denotes the imaginary unit. ˙ From the Fourier transform of the equations of motion, the Fourier transform X(ω) of the nodal velocities can be expressed by ˙ {X(ω)} = −iω(−ω2 [M] + iω[C] + [K])−1 [M]{1}A(ω)

(16)

538

Structural design optimization considering uncertainties

After the substitution of equation (16) into equation (15), the input energy may be computed by
∞ EI = FM (ω)|A(ω)|2 dω (17) 0

where FM (ω) = Re[iω{1}T [M]T [Y(ω)][M]{1}]/π

(18a)

[Y(ω)] = (−ω2 [M] + iω[C] + [K])−1

(18b)

5 Critical excitation problem It is shown in this section that a critical excitation method for the earthquake input energy can provide upper bounds on earthquake input energy. Westermo (1985) has discussed a similar problem for the maximum input energy to an SDOF system subjected to external forces. His solution is restrictive because it is of the form including the velocity response quantity containing the solution itself implicitly. A more general solution procedure will be presented here. The capacity of ground motions is often defined in terms of the time integral of squared ground acceleration (Arias 1970; Housner and Jennings 1975). This quantity is well known as the Arias intensity measure except for difference in the coefficient. The constraint on this quantity can be expressed by
∞
∞ |A(ω)|2 dω = C A a(t)2 dt = (1/π) (19) −∞

0

where C A is the specified value of the time integral of squared ground acceleration. It is also clear that the maximum value of the Fourier amplitude spectrum of input ground acceleration is finite. The infinite spectrum may correspond to a perfect harmonic function or that multiplied by an exponential function (Drenick 1970) which is unrealistic as an actual ground motion. The constraint on this property may be described by |A(ω)| ≤ A

(A : specified value)

(20)

The critical excitation problem for the MDOF system may be stated as follows: Find |A(ω)| that maximizes the earthquake input energy, equation (17), subject to the constraints (19) and (20) on ground acceleration. It is clear from the work (Takewaki 2001a, b, 2002b) on power spectral density functions that, if A is infinite, |A(ω)|2 turns out to be the Dirac delta function which has a non-zero value at the point maximizing F(ω). On the other hand, if A is finite, |A(ω)|2 2 yields a rectangular function attaining A in a certain range. The band-width of the 2 frequency can be obtained as ω = πC A /A . The position of the rectangular function, 2( ω i.e. the lower and upper limits, can be computed by maximizing A ωLU F(ω)dω. It is noted that ωU − ωL = ω. It can be shown that a good and simple approximation can be made by (ωU + ωL )/2 = !. The essential feature of the solution procedure presented

I n f o-g a p r o b u s t d e s i g n o f p a s s i v e l y c o n t r o l l e d s t r u c t u r e s

A→ ∞

539

Dirac delta function

Function F(ω)

A : finite A(ω)

2

Function F(ω)

ω L Ω ωU

Frequency ω

Figure 19.8 Schematic diagram of solution procedure for critical excitation problem.

in this section is shown in Figure 19.8. It is interesting to note that Westermo’s periodic solution (Westermo 1985) may correspond to the case of infinite A.

6 Info-gap robust design for load and model uncertainties 6.1

Info-gap models of load uncertaint y

Consider an uncertainty model of load which is expressed in terms of a Fourier ampli˜ and αs denote the nominal Fourier tude spectrum of the input acceleration. Let A ˜ for amplitude spectrum and its uncertainty level. An info-gap model of load A(αs , A) αs ≥ 0 is introduced to represent uncertainty in the Fourier amplitude spectrum of the input acceleration. The info-gap model of load may be defined by  ˜ 2 (ω; ω, CA )] = |A2 (ω)| = s∗ A ˜ 2 (ω ; ω/s∗ , CA ) : s∗ A[αs , A     s − s˜   ≤ αs , αs ≥ 0 = s/˜s,  (21) s˜  In the past Takewaki and Ben-Haim (2005) proposed a similar info-gap model for a power spectral density function. ˜ 2 ) can be found in Figure 19.9. Note that the The graphical expression of A(αs , A quantity CA is related to the power or intensity of the input acceleration and is assumed to be constant. This leads to the constant area of the critical rectangular function of the squared Fourier amplitude spectrum. As the amplitude changes uncertainly, the band-width varies correspondingly. In order to explain the physical meaning of variation of the squared Fourier amplitude spectrum shown in Figure 19.9, let us consider two waves as shown in Figures 19.10 and 19.11. These two waves have the same acceleration power CA . While Figure 19.10 represents a short-duration ground motion (near-field ground motion), Figure 19.11 presents a long-duration ground motion and simulates approximately a far-field ground motion. Figure 19.12 is the Fourier amplitude spectrum of the wave shown in Figure 19.10 and Figure 19.13 is the Fourier amplitude spectrum of the wave shown in Figure 19.11. From these figures, it may be said that a smaller-level and wider-range

540

Structural design optimization considering uncertainties

A2(ω)

~

s1*  s2*  s3*

s3*A2

Nominal squared Fourier amplitude spectrum ~

~ s1*A2

~ A2

s2*A2

ω

~ ∆ω

Figure 19.9 Variation of critical rectangular function of the squared Fourier amplitude spectrum of input acceleration.

Acceleration (g)

2 1 0

1 2

0

5

10 Time (s)

15

20

15

20

Figure 19.10 Short-duration motion.

Acceleration (g)

2 1 0

1 2

0

5

10 Time (s)

Figure 19.11 Long-duration motion.

squared Fourier amplitude spectrum represents a variation to a short-duration ground motion (near-field ground motion) and a larger-level and narrower-range squared Fourier amplitude spectrum assumes a variation to a long-duration ground motion (far-field ground motion). 6.2 Ro b ustn ess func t io n ˜ for αm ≥ 0 in The info-gap model for uncertainty in the dynamic model is F(αm , F) compliance with the definition (2). It should be noted that two different uncertainty

I n f o-g a p r o b u s t d e s i g n o f p a s s i v e l y c o n t r o l l e d s t r u c t u r e s

541

Fourier amplitude

5 4 3 2 1 0

0

2

4 6 8 10 Circular frequency (rad/s)

12

14

Figure 19.12 Fourier amplitude spectrum of short-duration motion.

Fourier amplitude

5 4 3 2 1 0

0

2

4 6 8 10 Circular frequency (rad/s)

12

14

Figure 19.13 Fourier amplitude spectrum of long-duration motion.

parameters αm and αs are used. The parameter αs for load uncertainty has been introduced just above and αm for model uncertainty is defined here. f (A, F, k) is the performance requirement based on the earthquake input energy for the Fourier amplitude spectrum A based on energy transfer function F and design k. As in the definition of the robustness in equations (5) and (6), the performance requirement may be expressed by f (A, F, k) ≤ fC

(22)

The info-gap robustness function can then be introduced as a measure of robustness for model uncertainty for a given load spectral uncertainty level αs . ⎧ ⎧ ⎫ ⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ ⎨ ⎬ ⎬ max (23) αˆ m (k, fC , αs ) = max αm : f (A, F, k) ≤ fC ⎪ ⎪ ⎪ ⎪ ˜ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ F ∈ F(αm , F) ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ ⎩ ⎭ ⎭ ˜ 2) A ∈ A(αs , A

7 Numerical examples Consider a six-degree-of-freedom mass-spring-dashpot system, equivalent to a sixstory shear building model as shown in Figure 19.14(a). This system has a uniform

542

Structural design optimization considering uncertainties

ug (t) (a)

(b)

(c)

(d)

Figure 19.14 Six-story shear building model; (a) Bare frame, (b) frame with an added damper in the first story, (c) frame with an added damper in the third story, (d) frame with an added damper in the sixth story.

6

Story number

5 4 3 2 1 0

1  106 2  106 3  106 4  106 Structural damping coefficient (Ns/m)

Figure 19.15 Structural damping coefficient.

structural damping, 3.76 × 105 (N · s/m), as shown in Figure 19.15. This structural damping corresponds to the damping ratio = 0.04 in the fundamental vibration mode. Each element has the same mass mi = 32 × 103 (kg) and every spring has the same stiffness 3.76 × 107 (N/m). The undamped fundamental natural period of the model is T1 = 0.72(s). An added viscous damper for passive control is installed in the first, third or sixth story as shown in Figures 19.14(b)–(d). The magnitude of the added viscous damper is shown in Figure 19.16. The nominal damping coefficient c˜ d of the added viscous damper is ten times the damping coefficient of the structural damping in the same story. The uncertain damping coefficient cd of the added viscous damper

6

6

5

5

4

Story number

Story number

I n f o-g a p r o b u s t d e s i g n o f p a s s i v e l y c o n t r o l l e d s t r u c t u r e s

am
1.0 am
0.5

3

am
0.0

2

543

am
1.0 am
0.5 am
0.0

4 3 2 1

1 0

1  106 2  106 3  106 4  106

1  106 2  106 3  106 4  106

0

Added damping coefficient (Ns/m)

Added damping coefficient (Ns/m)

(a)

(b) 6

Story number

5

am
0.0

4

am
0.5 am
1.0

3 2 1 0

1  106 2  106 3  106 4  106

Added damping coefficient (Ns/m) (c)

Figure 19.16 Added viscous damping coefficient; (a) first-story allocation, (b) third-story allocation, (c) sixth-story allocation.

is expressed by cd = c˜ d (1 ± 0.5αm ) where αm is the unknown horizon of uncertainty in the model coefficients. The uncertainty of the load is assumed to be expressed by the variation s = s˜(1 ± αS ) of the squared rectangular Fourier amplitude spectrum of the input ground acceleration where αS is the unknown horizon of uncertainty in the load. The power of the input defined by Eq.(19) does not vary and is given by C A = 11.4(m2 /s3 ) . The nominal level of the rectangular Fourier amplitude spectrum ˜ = 2.91(m/s) and its nominal band-width is ω˜ = 4.21(rad/s). is A The worst case (critical case), up to uncertainties αm and αS , can be obtained from cd = c˜ d (1 − 0.5αm ) and s = s˜(1 + αS ). Although the problem of finding the worst case is very complicated in general (Kanno and Takewaki 2007), the present case is almost self-evident. This enables one to discuss the info-gap robustness function directly with respect to the input energy performance.

544

Structural design optimization considering uncertainties

6  104 5  104

1st-story damping 3rd-story damping 6th-story damping

Function F(ω)

4  104 3  104 2  104 1  104 0 1  104 0

5 10 15 Circular frequency (rad/s)

20

Figure 19.17 Energy transfer functions F M (ω) defined by equation (17) for three models, one with an added damper in the first story, one in the third story and the other in the sixth story.

Figure 19.17 shows the energy transfer functions FM (ω) defined by equation (17) for three models, one with an added damper in the first story, one in the third story and the other in the sixth story. It can be observed that the energy transfer functions FM (ω) of a passively controlled mass-spring-damper system with an added damper near the fixed support (base) is smaller than that of the system with an added damper near its tip. This means that the allocation of passive dampers into lower stories is effective in reducing the earthquake input energy. Figure 19.18(a) illustrates the plot of the info-gap robustness function αˆ m versus the specified limit value of the earthquake input energy for the load spectral uncertainty αs = 0.0. Figures (b), (c) and (d) illustrate the plots of αˆ m for the load spectral uncertainties αs = 0.1, 0.3, 0.5. Comparing these figures we see that robustness to model-uncertainty, αˆ m , decreases as load-uncertainty, αS , increases. It can be observed that a passively controlled mass-spring-damper system with an added damper near the fixed support is ‘more robust’ than that with an added damper near its tip in all the cases of the load spectral uncertainties. Figure 19.19(a) shows the plot of the info-gap robustness function αˆ m , of the model with an added damper in the first story, versus the specified limit value of the earthquake input energy for various levels of load uncertainties. It can be understood that, as the level of load uncertainty increases, the info-gap robustness function αˆ m gets smaller, i.e. less robust for variation of the structural parameter. Figures 19.19(b) and (c) illustrate the plots of info-gap robustness function αˆ m of the models with an added damper in the third and sixth stories, respectively, with respect to the specified value of the earthquake input energy for various levels of load uncertainties. A similar tendency to Figure 19.19(a) can be observed.

I n f o-g a p r o b u s t d e s i g n o f p a s s i v e l y c o n t r o l l e d s t r u c t u r e s

1

1

as
0.0

0.8

0.6

0.6

τm

am

0.8

0.4

as
0.1

0.4

0.2 0

0

5  106 1  107 Input energy (Nm)

1st-story damping 3rd-story damping 6th-story damping

0.2

1st-story damping 3rd-story damping 6th-story damping

1.5  107

0

0

5  106 1  107 Input energy (Nm)

(a)

1.5  107

(b) 1

as
0.3

0.8

0.8

0.6

0.6

am

am

1

0.4

as
0.5

0.4

0.2 0

545

1st-story damping 3rd-story damping 6th-story damping

0

5  106 1  107 Input energy (Nm)

(c)

1.5  107

0.2 0

1st-story damping 3rd-story damping 6th-story damping

0

5  106 1  107 Input energy (Nm)

1.5  107

(d)

Figure 19.18 Plot of the info-gap robustness function αˆ m versus the specified limit value of the earthquake input energy for various load spectral uncertainties; (a) αs = 0.0, (b) αs = 0.1, (c) αs = 0.3, (d) αs = 0.5.

Figure 19.20 presents the plot of the info-gap robustness function αˆ m with respect to the level of the load spectral uncertainty αs for the model with an added damper in the first story. From this figure, the designer can understand the effect of the load spectral uncertainty αs on the info-gap robustness function. It is also interesting to note that the info-gap robustness function αˆ m and the level of the load spectral uncertainty αs introduce a new trade-off relationship.

8 Conclusions This chapter has developed a new methodology for design of passively controlled mass-spring-damper systems subject to severe uncertainties in both the loads and the structural models. The following are the main results of this chapter.

546

Structural design optimization considering uncertainties

1

1

3rd-story damping

1st-story damping 0. 8

0.6

0. 6

am

as = 0.0

0.4

am

0.8

as = 0.1 as = 0.3

0.2

as = 0.0

0. 4

as = 0.1 as = 0.3

0. 2

as = 0.5

as = 0.5 0

0

5106 1107 Input energy (Nm)

0 0

1.5107

(a)

5106 1107 Input energy (Nm) (b)

1.5107

1

6th-story damping 0. 8

am

0. 6 as = 0.0

0. 4

as = 0.1 as = 0.3

0. 2

as = 0.5 0

0

5106 1107 Input energy (Nm)

1.5107

(c)

Figure 19.19 Plot of the info-gap robustness function αˆ m versus the specified limit value of the earthquake input energy for various levels of load uncertainties; (a) model with an added damper in the first story, (b) model with an added damper in the third story, (c) model with an added damper in the sixth story.

(1)

(2)

The earthquake input energy is an appropriate measure for evaluating the performance level of passively controlled structures. A critical excitation problem can be stated for the earthquake input energy as a criticality measure. The critical excitations depend upon the dynamic properties of the passively controlled mass-spring-damper systems and it is necessary to deal with load and structural model uncertainties simultaneously. Info-gap uncertainty models are very useful in describing both the load and structural model uncertainties. Determination of the critical states in the load and structural parameters is an essential step to the investigation on the robustness of the passively controlled mass-spring-damper systems.

I n f o-g a p r o b u s t d e s i g n o f p a s s i v e l y c o n t r o l l e d s t r u c t u r e s

547

0.8

Model robustness function am

0.7 EI
8106 (Nm) 0.6 EI
6106 (Nm)

0.5 0.4

EI
4106 (Nm) 0.3 0.2 0.1 0

0

0.1

0.2 0.3 0.4 Spectral uncertainty as

0.5

Figure 19.20 Plot of the info-gap robustness function αˆ m with respect to the level of the load spectral uncertainty αs for various requirements of earthquake input energies EI = 4.0 × 106 ,6.0 × 106 ,8.0 × 106 (Nm) (first-story damping model).

(3)

(4)

A passively controlled mass-spring-damper system with an added damper near the fixed support is more robust than that with an added damper near its tip. The added robustness is evaluated quantitatively. The simultaneous consideration of the load and structural model uncertainties introduces a new trade-off. The robustness to structural model uncertainty increases as the uncertainty level of the load gets smaller.

Acknowledgements Part of this chapter was written while one author (YBH) was a fellow of the Japan Society for the Promotion of Science, at the University of Tokyo and Kyoto University. The support of the JSPS is gratefully acknowledged. Part of this chapter is supported by the Kajima Foundation and JSPS (2006). This support is also gratefully acknowledged.

References Akiyama, H. 1985. Earthquake Resistant Limit-State Design for Buildings. University of Tokyo Press, Tokyo, Japan. Arias, A. 1970. A measure of earthquake intensity. In Seismic Design for Nuclear Power Plants, R.J. Hansen (ed.), The MIT Press, Cambridge, MA, 438–469. Ben-Haim, Y. 1996. Robust reliability in the mechanical sciences, Springer-Verlag, Berlin. Ben-Haim, Y. 2005. Info-gap decision theory for engineering design, in Engineering Design Reliability Handbook, E. Nikolaide, D. Ghiocel & S. Singhal (eds), CRC Press. Ben-Haim, Y. 2006. Info-gap decision theory: decisions under severe uncertainty, 2nd edition, Academic Press, London. Ben-Haim, Y. & Elishakoff, I. 1990. Convex Models of Uncertainty in Applied Mechanics. Elsevier Science Publishers, Amsterdam.

548

Structural design optimization considering uncertainties

Cherng, R.H. & Wen, Y.K. 1994. Reliability of uncertain nonlinear trusses under random excitation, I., II. J. of Engineering Mechanics, ASCE 120(4):733–757. Drenick, R.F. 1970. Model-free design of aseismic structures. J. Engrg. Mech. Div., ASCE 96(EM4):483–493. Ghanem, R.G. & Spanos, P.D. 1991. Stochastic Finite Elements: A spectral approach. Springer, Berlin. Housner, G.W. 1959. Behavior of structures during earthquakes. Journal of the Engineering Mechanics Division, ASCE 85(4):109–129. Housner, G.W. & Jennings, P.C. 1975. The capacity of extreme earthquake motions to damage structures. Structural and geotechnical mechanics: A volume honoring N.M. Newmark edited by W.J. Hall, 102–116, Prentice-Hall Englewood Cliff, NJ. Igusa, T. & Der Kiureghian, A. 1988. Response of uncertain systems to stochastic excitations. J. Eng. Mech., ASCE 114(5):812–832. Jensen, H. 2000. On the structural synthesis of uncertain systems subjected to environmental loads. Structural and Multidisciplinary Optimization 20:37–48. Jensen, H. & Iwan, W.D. 1992. Response of systems with uncertain parameters to stochastic excitations. J. Eng. Mech., ASCE 114:1012–1025. Kanno, Y. & Takewaki, I. 2007. Worst-case plastic limit analysis of trusses under uncertain loads via mixed 0–1 programming. Journal of Mechanics of Materials and Structures 2(2): 247–273. Koyluoglu, H.U., Cakmak, A.S. & Nielsen, S.R.K. 1995. Interval algebra to deal with pattern loading and structural uncertainties. J. Eng. Mech., ASCE 121(11):1149–1157. Lyon, R.H. 1975. Statistical energy analysis of dynamical systems, The MIT Press, Cambridge, MA. Ordaz, M., Huerta, B. & Reinoso, E. 2003. Exact computation of input-energy spectra from Fourier amplitude spectra Earthquake Engineering and Structural Dynamics 32: 597–605. Page, C.H. 1952. Instantaneous power spectra. Journal of Applied Physics 23(1):103–106. Qiu, Z. & Wang, X. 2003. Comparison of dynamic response of structures with uncertainbut-bounded parameters using non-probabilistic interval analysis method and probabilistic approach. Int. J. Solids and Structures 40:5423–5439. Shinozuka, M. 1970. Maximum structural response to seismic excitations. J. Engrg. Mech. Div., ASCE 96(EM5):729–738. Takewaki, I. 2001a. A new method for nonstationary random critical excitation. Earthquake Engineering and Structural Dynamics 30(4):519–535. Takewaki, I. 2001b. Probabilistic critical excitation for MDOF elastic-plastic structures on compliant ground. Earthquake Engineering and Structural Dynamics 30(9):1345–1360. Takewaki, I. 2002a. Critical excitation method for robust design: A review. Journal of Structural Engineering, ASCE 128(5):665–672. Takewaki, I. 2002b. Robust building stiffness design for variable critical excitations. Journal of Structural Engineering, ASCE 128(12):1565–1574. Takewaki, I. 2004. Bound of earthquake input energy. Journal of Structural Engineering, ASCE 130(9):1289–1297. Takewaki, I. & Ben-Haim, Y. 2005. Info-gap robust design with load and model uncertainties. Journal of Sound and Vibration 288(3):551–570. Takewaki, I. 2006. Critical Excitation Methods in Earthquake Engineering. Elsevier Science Publishers, Amsterdam. Uang, C.M. & Bertero, V.V. 1990. Evaluation of seismic energy in structures. Earthquake Engineering and Structural Dynamics 19:77–90. Westermo, B.D. 1985. The critical excitation and response of simple dynamic systems. Journal of Sound and Vibration 100(2):233–242.

Chapter 20

Genetic algorithms in structural optimum design using convex models of uncertainty Sara Ganzerli & Paul De Palma Gonzaga University, Spokane,WA, USA

ABSTRACT: This chapter focuses on the use of convex models of uncertainty with genetic algorithms for optimal structural design. The chapter is comprised of five sections. Section 1 is a literature review of convex models and their application to optimal structural design and other engineering fields. Section 2 explores the use of convex models to deal with uncertainties as an alternative to the more traditional probabilistic approach. In this section the superposition method to implement the uniform bound convex model is illustrated. Section 3 underlines the benefits of incorporating genetic algorithms in optimal structural design. Section 4 presents applications of convex models to optimal structural design and Section 5 suggests new avenues for research and application of convex models. This chapter grows from research conducted since 2000 by the Gonzaga University Center for Evolutionary Algorithms. A preliminary version was published by Millpress as (Ganzerli et al. 2003).

1 Literature survey on convex models of uncertainty applied to optimal structural design The book, “Convex Models of Uncertainty in Applied Mechanics’’ by Ben-Haim and Elishakoff, established the foundation for the convex model theory and application in 1990. Since then, much work has been done on convex models. Together with probability and fuzzy sets, convex models can be considered part of the uncertainty triangle (Elishakoff 1995). Convex models are especially useful for problems where data on the uncertain parameters is scarce, as in the case of severe uncertainties. One of the most widely used of the convex models is the uniform bound model. Here the convex set that encloses the uncertain parameters has the shape of a rectangle in two dimensions. The uniform bound convex model has been implemented in structural design using a technique called “anti-optimization’’ (Elishakoff et al. 1994). This technique requires that a complete optimization routine be invoked for each cycle of the algorithm that minimizes the cost function. The nested routine maximizes the effects due to the uncertain parameters on the structure. That is, it generates the worst-case scenario. This process is computationally quite expensive. In addition, it requires that the constraints, i.e. stresses and displacements, be written as an explicit function of the design variables, i.e. the cross-sectional areas. Another approach, called the superposition method, has been proposed by (Ganzerli and Pantelides 2000). This method allows one to account for a large number of uncertainties in structures with many members. It eliminates the nested optimization

550

Structural design optimization considering uncertainties

and does not require that the structural response be written as a function of the uncertain parameters. The superposition method is in the applications presented in Section 4. Ben-Haim et al. (1996) proved the efficiency of convex models in identifying the worst-case scenario due to multiple uncertainties. Perhaps surprisingly, for structures subjected to uncertain static loads, the maximum structural response due to the uncertain parameters cannot be identified simply by increasing each of the loads to their maximum value. However, the convex model design is robust against constraint violations for any load condition within the established bounds. Convex models have efficiently been applied in the study of thin-walled stiffened composite panels that are highly sensitive to geometrical imperfections. In (Elseifi et al. 1998), a comparison between the convex models and a Monte Carlo simulation led to similar results but with an effort and cost reduction favoring the convex models. Recently, convex models have been employed in several fields of engineering. Tonon et al. (2001) illustrate hybrid systems that combine the three methods available to deal with uncertainties: probability, fuzzy sets, and convex models. Optimization of laminated composites considering uncertainty using convex models is the focus of three studies: (Kim and Sin 2001; Cho and Rhee 2003; Cho and Rhee 2004). Attoh-Okine (2002) examines the convex model method for pavement design where layer coefficients are uncertain. Spletzer and Taylor (2003) suggest convex models as a new approach to the multi-robot localization problem. Recently, attention has been given to convex models and interval analysis as comparable methods to deal with uncertainty (Qiu 2003 and Qiu and Wang 2006). An extension of convex models to decision-making theory has been recently presented in two books by (Ben-Haim 1996 and 2001). Ben-Haim showed that for the sake of robustness, the uncertainty must be allowed to vary with no imposition of fixed bounds. In addition, he presents design curves that, plotting robustness against uncertainty and structural cost, illustrate the tradeoffs. Design curves include an array of possible solutions to a design problem and constitute a useful tool for the designer, an area explored in (Ganzerli et al. 2005). Robustness for trusses is presented in (Au et al. 2003) and (Kanno and Takewaki 2006a and 2006b). Traditional optimization techniques encompass: (1) mathematical programming, (2) optimality criteria, (3) approximation methods, and (4) steepest descent methods. Optimal structural design has been implemented for many years (Kirsch 1981). In traditional optimization, the domain is searched using the gradient of the function to be optimized, the objective function. A problem with this method is that in order to calculate the gradient, the function must be continuous. Genetic algorithms (GA), loosely based on Darwin’s theory of natural selection, are not subject to this limitation. First proposed by John Holland of the University of Michigan in 1975, they were not widely used until one of his students (Goldberg 1989) showed that they could help solve difficult problems. Since then, genetic algorithms have been employed in many science and engineering fields to successfully solve optimization problems. The literature on the theory of practice of GA is extensive. A focal point is the annual Genetic and Evolutionary Computation Conference (GECCO). GA represent a step forward in the area of optimization, because they do not require the gradient of the function to be optimized. Thus, they are effective in solving complex problems with multiple objective functions that are discontinuous.

Genetic algorithms in structural optimum design using convex models

551

In structural design, genetic algorithms have been used extensively in the past decades to study minimum volume problems. In particular, many chapters deal with the optimization of trusses. Rajeev and Krishnamoorthy (1997) examined the use of GA for discrete optimization of truss structures. Keeping in mind that practical problems often involve discrete variables, they have developed a simple GA. Here the algorithm assigns a penalty to structures that violate constraints, say, the known structural response of a member of a given material to a specified static load. Rajeev and Krishnamoorthy present the complete optimization history of a three bar truss. The method is then applied to larger trusses composed of up to 160 members. Optimization of large trusses using traditional algorithms was presented by (Schemit and Lai 1994). Ghasemi et al. (1999) have considered the optimization of trusses with both continuous and discrete variables. They demonstrate the efficiency of GA in solving large two-dimensional trusses. They present a solved example for a 940member truss. The literature includes very few examples of chapters that combine convex models with genetic algorithms. Cho and Rhee (2003 and 2004) have studied the layup optimization for free edge strength using GA for the optimal design and convex models to deal with uncertainties. Ganzerli et al. (2002 and 2003) have implemented GACON, a GA-based optimization routine, combined with the uniform bound convex model to deal with load uncertainties. Combining genetic algorithms with convex models shows considerable promise. It is an area open to investigation.

2 Uniform bound convex model Convex models of uncertainty are a non probabilistic method. Convex models are appealing because they are easy to use and do not depend on knowledge of the statistical distribution of the uncertain parameter values. They are especially useful when an info-gap situation arises, i.e. when severe uncertainties must be handled. In this section convex models are employed in the structural design of trusses subject to an uncertain static load condition. The uniform bound convex model was chosen because of its easy implementation. The variation of uncertain parameters from their nominal values is required to be bounded by a convex set, also called the convex domain. The uniform bound convex set is represented by a rectangle in the plane, when only two uncertain parameters are present. The convex domain can be generalized to three dimensions when it assumes the shape of a parallelepiped, also known as “box’’. Furthermore, if a generic n-number of uncertain parameters is present, the convex domain is a multidimensional “box’’. To illustrate the implementation of the uniform bound convex model, let us consider a three bar truss represented in Figure 20.1. For the sake of simplicity and without loss of generality, only two uncertain parameters are handled. These are the static loads P1 and P2 . The three bar truss has two degrees of freedom x1 and x2 . Superscripts U, L, and 0 will denote the upper, lower, and nominal values of the loads respectively. Hereafter, the subscript i will indicate the number of degrees of freedom, j the number of truss members, and n the number of uncertain parameters. The loads might vary from their nominal values Pn0 by a percent value βn. So, for example, load 1 at its minimum is designated

552

Structural design optimization considering uncertainties

x2 P1 x1

P2

Figure 20.1 Three-bar truss.

( P1L, P2U )

P2

(P1U, P2U )

( P1n, P2n ) (P1L, P2L )

P1

(P1U, P2L)

Figure 20.2 The convex domain.

as P1L . Therefore, the convex set can be defined as follows for the three-bar truss of Figure 20.1:  CP =

P10 − β1 ≤ P10 ≤ P10 + β1



P20 − β2 ≤ P20 ≤ P20 + β2

(1)

In general terms it is possible to express the domain as Pn0 − β1 ≤ Pn0 ≤ Pn0 + β1

(2)

The representation of this convex set is given in Figure 20.2. P1 is the x coordinate, P2 the y. The nominal values, the points of no uncertainty, are at the origin. It is clear that the range of possible values that can be assumed by either P1 or P2 is within the box. Nevertheless, it has been demonstrated by (Elishakoff et al. 1994) that the worst effect due to the uncertainties needs to be searched on the convex hull, i.e. the rectangle’s perimeter. For the particular case of a linear problem, the search can

Genetic algorithms in structural optimum design using convex models

553

actually be limited to the vertices of the convex set, where the uncertain parameters assume their maximum and minimum values. The convex structural response can be conveniently found through a superposition method first proposed in (Ganzerli and Pantelides 2000). This method is simple and can be used to handle many uncertain parameters and large structures whenever superposition applies. In the structural analysis convex values for the displacements and internal forces are used instead of the nominal ones. The convex portion is added to the nominal values of the structural response (displacements, forces and stresses) calculated when all the given nominal loads are acting on the structure. That is, the convex values are superimposed on the nominal values. The latter consists of summing the nominal structural responses calculated with one load at a time and multiplying them by the percent of uncertainty for that load. The resulting expressions for the convex displacement and force vectors are:

xi,con = xi (P10 , P20 ) ± {β1 |xi (P10 )| + β2 | xi (P20 )|} Fj,con =

Fj (P10 , P20 )

±

{β1 |Fj (P10 )|

+

β2 |Fj (P20 )|}

(3) (4)

where i = number of degrees of freedom (varies from one to eight for the 10-bar truss), j = number of members (varies from one to ten for the 10-bar truss), xi,con and Fj,con are the convex displacements and internal forces, xi (P10 , P20 ) and Fj (P10 , P20 ) are the nominal displacements and internal forces calculated loading the structure with both P10 and P20 , |xi (P10 )| and |Fj (P10 )| are the absolute values of the nominal displacements and internal forces calculated loading the structure with only P10 (P2 = 0), |xi (P20 )| and |Fj (P20 )| are the absolute values of the nominal displacements and internal forces calculated loading the structure with only P20 (P1 = 0), β1 and β2 are the percents of uncertainty for P1 and P2 respectively. In Eqs. (3) and (4) the ± sign is in agreement with the sign of the first term. In other words, if xi (P10 , P20 ) and Fj (P10 , P20 ) are positive, the plus/minus sign will turn into a plus sign and vice versa. This guarantees that the nominal displacements and forces are always increased when uncertainty is present. The worst-case scenario, due to the uncertain parameters, is captured by the equations. Convex stresses can be obtained directly from the convex internal forces just by dividing the latter by the member cross sectional areas (Wang 1986). Figure 20.3 shows how the method of superposition is implemented in the case of two uncertain parameters P1 and P2 . In sum, to account for uncertainties in the structural design, it is sufficient to substitute the convex responses for the nominal cases in the structural analysis routine. One important note is that, although the uniform bound convex model is implemented here jointly with genetic algorithms, it can also be introduced in the conventional (non-optimal) design of structures. As mentioned, the superposition method is easy to implement, but it presents a few drawbacks. The main one is that it can be applied only when the conditions to use superposition are met. For example, it cannot handle nonlinear structures. Another drawback is that the superposition method requires the solution of multiple structural analyses, adding computational burden.

554

Structural design optimization considering uncertainties

P1

P1

P2

P2

xi, con = xi (P10 , P20)



{b1 |xi(P10)|



b2 |xi(P20)|}

Fj, con = Fj (P10 , P20)



{b1 |Fj(P10)|



b2 |Fj(P20)|}

Figure 20.3 The convex displacements, x i,con , and forces F j,con , are obtained using superposition. β1 and β2 are the percents of uncertainty for P 1 and P 2 respectively convex domain.

3 Optimal structural design using genetic algorithms Genetic algorithms have been a well-established optimization technique for a decade (Haupt and Haupt 1998). They offer at least two advantages over calculus-based optimization algorithms. The function to be optimized, known as the objective function, must be continuous. Since genetic algorithms are based loosely on the process of natural selection found in nature, they do not require a continuous objective function. Though aspects of natural processes can be modeled using continuous functions, of course, the GA works at a lower level, modeling not just behavior, but the evolutionary processes that produces behavior. For example, GA loosely mimics differential reproduction in nature, one result of which is adaptation. Further, in structural design, the design variables are often the member cross-sectional areas. Since these are manufactured only in specific, discrete values, considering values other than these imposes an unnecessary computational burden on the optimization process. It is extremely difficult to determine if the solution to an optimization problem is truly optimal or nearly optimal. In the case of volume minimization for a truss structure, the problem space grows exponentially with the size of the truss. The true optimal solution could be found performing an exhaustive search of the problem space. But in practical terms, this is impossible. A problem whose optimal solution is theoretically possible, but practically impossible, is called intractable. Another famous problem that is classified as intractable is the traveling salesman problem. It can be stated like this: given a set of cities with known distances between them, construct a tour of minimum distance visiting each city once and returning to the origin. The traveling salesman problem belongs to the class of NP-complete problems, the most difficult one to tackle from a computational point of view. Overbay et al. (2006) demonstrated that the truss problem belongs to the NP-complete set. Heuristic techniques, such as GA, are the only practical solution to intractable problems. Here is another reason to use GA in optimal structural design. Although no solution found with GA is guaranteed to be optimal, using GA helps find a nearly optimal solution. For all practical purposes, this represents an improvement with respect to a solution that is not optimized.

Genetic algorithms in structural optimum design using convex models

555

In the following section, examples of structural optimization are presented. The goal is to minimize the volume of a truss with fixed geometry and load conditions. The design process sets constraints on the maximum stresses and displacements, so that the structure’s safety and serviceability do not fall below a specified minimum. The optimal design problem for a truss can be stated like this: min f (A, P) such that gk (A, P) ≤ gk , allowable

(k = 1, . . . , m)

(5)

where f = objective function, the truss volume, expressed as a function of the cross-sectional areas (A) and the external loads (P), A and P = design parameters, gk (A, P) = constraints, k = number of constraints to be satisfied by the optimal design, gk,allowable = allowable value for constraint gk . If gk (A, P) exceeds gk,allowable for any constraint, the particular configuration under consideration is unfit. As already stated, GA is an optimization method that mimics the natural selection process. An initial population of individuals (trusses) is randomly generated and ranked based on desired characteristics. Genes are the truss’ cross-sectional areas. The population can be ranked in a fitness hierarchy, where the fittest truss is the one whose members have the least volume while still meeting the structural constraints. Those trusses that do not meet the constraints are assigned a cost penalty, pushing them lower in the fitness hierarchy. After the initial population has been generated, individuals have to be paired so that they can “mate’’ and produce the next generation. Since the strength of differential reproduction consists in a parent passing characteristics to its offspring, GA next needs to establish how these characteristics, these “genes,’’ will be inherited. This is referred to as crossover. An important element is random mutation. A fixed percentage of the genes present in the population are mutated. Mutation ensures that some alleles (gene values) will be introduced that were not randomly generated at the beginning. This reduces the chance that the algorithm will converge on a local minimum. The cycle is repeated until an acceptable optimal solution is reached. The genetic algorithm that produced the results presented in this chapter was customdesigned, using object-oriented programming techniques. However, standard libraries are available for the implementation of GA. Worth mentioning is GAlib, a collection of GA routines that are available at no cost (Wall 1995). In the next section examples are solved. A benchmark ten-bar truss is included along with a 64-bar truss where the 64 cross-sectional areas are treated as independent variables.

4 Applications including the use of convex models of uncertainty together with genetic algorithms to optimize structural design This section includes examples of truss optimal design considering both the nominal and convex model solutions. A 10-bar truss is presented first as a benchmark. To demonstrate the efficacy of GA in handling large structures, a 64-bar truss is shown.

556

Structural design optimization considering uncertainties

X8 F

X6

E

D

1 9

8

2

X7

5

X5 6

9144 mm (360 in)

10 X2

7

X4

3

4 X1

B

A

P1 9144 mm (360 in) A node

3 bar

X3

C P2 9144 mm (360 in)

X2 Degrees of freedom X1

Load P2

Figure 20.4 10-bar truss: geometry; loading conditions; degrees of freedom.

In this example, all of the cross sectional areas are independent variables. The examples feature a nominal and a convex model solution, where uncertainties are accounted for in the static loads. The solutions are obtained using both discrete optimization, i.e. integer values, and continuous optimization, i.e. floating points. For comparison with the literature, examples were solved using US units. However, SI units or their conversion factor, are given as well. 4.1 1 0-b ar trus s The 10-bar truss of Figure 20.4 is a commonly used benchmark. The truss is composed of aluminum, with a Young’s modulus of 68.9 GPa (104 ksi). The static loads P1 and P2 have a nominal value of 444.8 kN (100 kip). The design variables, represented by genes, are the cross-sectional areas of the ten members. The problem is to minimize the total volume of the truss without violating any constraint. Constraints are imposed on member stresses that may not exceed 172.4 MPa (25 ksi), except member 9, whose stress may not exceed 517.1 MPa (75 ksi). The allowable limits for stresses are set for both tension and compression. The search range for the design variables is limited to between 64.5 mm2 (0.1 in2 ) and 6451.6 mm2 (10 in2 ). The results obtained are presented in Table 20.1. The 10-bar truss is solved providing three optimal designs. The first two consider nominal loads and no uncertainties. Design 1, presented in Column 2, is obtained using integers. Design 2, shown in Column 3, is achieved using floating points. The opportunity given by GA to perform discrete optimization, as well as continuous, is significant in structural design. The integer solution converges faster than the floating point one and can be used for a preliminary estimate of member sizes. Design 3, presented in Column 4, differs from the others in that uncertainties are accounted for.

Genetic algorithms in structural optimum design using convex models

557

Table 20.1 10-bar Truss. Resultsa . Member

Design 1 Nominal integer Squared inches

Design 2 Nominal floating Squared inches

Design 3 Convex floating Squared inches

1 2 3 4 5 6 7 8 9 10 Volume (in3 ) Volume (m3 )

8 1 4 9 1 1 7 5 4 1 17 300 0.283

7.437 0.574 3.437 8.576 0.101 0.576 6.469 4.853 3.230 0.813 15 270 0.250

8.286 0.554 3.888 9.315 0.100 0.554 6.950 5.495 4.054 0.784 16 970 0.278

a Note: 1 in2 = 645.2 mm2 .

Each load, P1 and P2 , has an uncertainty of 10%, which means that βn in Eqs. 1 through 4 is equal to 0.1. It is clear that the convex model solution is more conservative than the nominal one obtained with floating points. However, this sacrifice in cost results in robustness against uncertainties. Convex models are useful in identifying the worst case scenario due to uncertainty in the static load condition. It is reasonable to hypotize that the worst case scenario could be obtained with a truss whose loads are increased by the percentage of uncertainty of their nominal values. However, (Elishakoff et al. 1994) and (Ganzerli and Pantelides 2000) have demonstrated that the case scenario obtained using convex models is the most robust against variation of the loads with respect to their nominal values. Optimal designs obtained for both the convex model and the “assumed’’ worst load condition, obtained by increasing or decreasing each load magnitude according to its percentage of uncertainty, were compared. Whereas the convex model design does not violate the stress constraints in any member for any load combination, the “assumed’’ worst case scenario presents a large number of violations, for the majority of load combinations. This is true with respect to both stress and displacement constraints. 4.2 64-bar trus s A 64-bar truss is used to demonstrate the algorithm efficiency for large problems. See Figure 20.5. The truss is composed of aluminum, with a Young’s modulus of 68.9 GPa (104 ksi). Adjacent nodes are 5080 mm (200 in) apart in the horizontal and vertical directions. Nodes 21, 22, 27, and 28 have no degrees of freedom, but all other nodes have two. Two designs are proposed. In one case, the design variables represent the member cross-sectional areas and constitute 64 independent variables. In the other case, the 64 variables are linked so that they are reduced to only 19 independent variables. The linking is shown in Figure 20.6. A set of members is represented by one variable in this case, and each member in the set will have the same value. Set 1, for

558

Structural design optimization considering uncertainties

2

21

22

19

20

17

18

4

6

8

16

24

26

28

3

5

7

15

23

25

27

13

14

11

12

311.4 kN (70 kip) (LC1) 1 88.9 kN (20 kip) (LC3)

88.9 kN (20 kip) (LC3)

9

10 444.8 kN (100 kip) (LC2)

3 node

88.9 kN (20 kip) (LC3)

Load magnitude, 88.9 kN (20 kip), and load condition, (LC3)

Figure 20.5 64-bar truss. Node numbering and loading conditions.

example, includes members 1–3, 3–5, 2–4, and 4–6. The members are named using the two end joint numbers. For example, the notation 1–3 refers to the member included between nodes 1 and 3. In addition, the truss is solved using both integer and floating point values. Each design variable, A, has the following range: • •

For integers: 645.2 mm2 (1 in2 ) ≤ A ≤ 20645.1 mm2 (32 in2 ) For floating points: 645.2 mm2 (1 in2 ) ≤ A ≤ 12903.2 mm2 (20.0 in2 ).

The constraints on the truss are: • •

Displacement: nodes 1 (vertical) and 9 (horizontal) are limited to 254 mm (10 in) Stresses: no member’s stress may exceed 172.4 MPa (25 ksi) in either tension or compression.

Genetic algorithms in structural optimum design using convex models

16

559

16

14

14 16 15

15 14

14 11 11

11 8

8 1 2

1 2

2

3 3

7

7 3

9

9

9

7

17 12 18 18 12 19

7

17

13 12 15

13 8

17 19 19

7 1

1

7

10

10

17

8

8 10 6

6

4

4 6 5

5 4

4 5 1 design variables linked in group 1

Figure 20.6 64-bar truss. Design variables linking.

The loads are imposed simultaneously for both the nominal and convex cases. The loads for both cases are: • • •

Loading condition 1 (LC1): 311.4 kN (70 kips) horizontally to the right on nodes 1 and 2 Loading condition 2 (LC2): 444.8 kN (100 kips) vertically downwards on nodes 9 and 10 Loading condition 3 (LC3): 88.9 (20 kips) vertically downwards on node 1 and 88.9 (20 kips) horizontally to the right on node 9.

For the convex case, the uncertainties are: • • •

LC1: ±20% (β1 = 0.2) LC2: ±10% (β2 = 0.1) LC3: ±10% (β3 = 0.1)

560

Structural design optimization considering uncertainties

In summary, three variations are used on the trial runs, with two options for each variation. Variations and options are: • • •

Value type (integer or floating point) Model (nominal or convex) Number of variables (64 when not linked or 19 when linked).

The results for the eight designs are shown in Table 20.2. Column 1 contains the member names and the set of linked variables. Columns 2 to 5 show the integer solutions. Columns 2 and 3 are the nominal designs solved using 64 independent variables and 19 independent variables. Similarly, columns 4 and 5 display the results for the convex model design. Column 4 shows the values for 64 independent variables and Column 5 shows the design with 19 independent variables. The last four columns are analogous to columns 2–5 but solved with floating points. A ranking of the volumes obtained shows expected results. The integer results lead to a higher volume than the floating points. The nominal case displays a lower volume than the convex model case. The extra-volume for convex models is a structural cost added to safeguard against uncertainties. The 64 independent variables give a more detailed description of the design and show lower volumes with respect to the designs solved with the linked variables. However, the computational cost is higher for the 64 independent variables, because the genetic algorithm converges at a slower rate. The 64-bar truss was previously solved for the nominal case with linked variables in (Ghasemi et al. 1999). Some variations from (Ghasemi et al. 1999) in the application of the load condition were made in this chapter. The authors have also solved the 64-bar truss with the same criteria as (Ghasemi et al. 1999) and have obtained results that are in agreement. It is worth mentioning that increasing the independent variables from 19 to 64 increases the complexity of the problem substantially. The examples demonstrate the power and flexibility of the method for large structures under severe uncertainties.

5 Suggestions for further studies The implementation of convex models offers many research opportunities. Especially interesting is the study of nonlinear structures, an area that has not yet received much attention. An innovative approach to the convex model would be to study its application for plasticity under uncertain loads. It would also be valuable to proceed with the study of robustness, allowing the uncertainties to vary. Convex models are an attractive alternative to the study uncertainties, an area that is growing increasingly important in engineering. 5.1 C o nv ex mo d e ls fo r plas t ic it y It is relevant to underline that the superposition method is just one of the options available to implement convex models. At the same time, the uniform bound convex model is one convex set among others. Using superposition would not permit the solution of nonlinear structures. To overcome this problem, it is possible to use an energy bound method suitable for the purpose of handling plasticity (Ben-Haim and Elishakoff 1990; Pantelides and Tzan 1996). Nevertheless, the method has not been

Genetic algorithms in structural optimum design using convex models

561

Table 20.2 64-bar Truss. Resultsa,b . Member/ linked set

Integer results for areas

Floating points results for areas

Squared inches

Squared inches

Nominal

2-4/1 1-3/1 4-6/1 3-5/1 1-2/2 2-3/2 1-4/2 3-4/3 4-5/3 3-6/3 10-12/4 9-11/4 12-14/4 11-13/4 9-10/5 9-12/5 11-10/5 11-12/6 11-14/6 13-12/6 6-8/7 5-7/7 13-7/8 14-15/8 8-17/8 16-18/8 16-24/7 15-23/7 5-6/9 6-7/9 5-8/9 13-14/10 7-14/13 13-15/13 17-18/11 17-16/11 8-18/11 23-24/12 16-23/12 15-24/12 7-8/8 15-16/8 8-16/7

Convex

Nominal

Convex

Un-linked U-N-I

Linked L-N-I

Un-linked U-C-I

Linked L-C-I

Un-linked U-N-F

Linked L-N-F

Un-linked U-C-F

Linked L-C-F

3 4 2 4 1 1 1 1 1 1 4 5 3 5 1 1 1 1 1 1 1 5 6 2 8 2 1 6 1 1 1 1 1 1 1 1 1 1 1 1 7 2 1

4 4 4 4 1 1 1 1 1 1 5 5 5 5 1 1 1 2 2 2 6 6 7 7 7 7 6 6 1 1 1 1 1 1 1 1 1 1 1 1 7 7 6

3 4 3 5 1 1 1 1 1 1 4 5 3 6 1 1 1 1 1 2 2 6 7 3 9 2 1 7 1 1 1 1 1 1 1 1 1 1 1 1 8 3 2

4 4 4 4 1 1 1 1 1 1 5 5 5 5 1 1 1 2 2 2 6 6 8 8 8 8 6 6 1 1 1 1 1 1 1 1 1 1 1 1 8 8 6

2.071 2.917 1.786 4.143 1.736 1.053 0.125 0.208 0.382 1.232 3.396 4.316 2.473 5.036 1.904 0.360 0.986 0.889 0.083 1.087 1.000 4.595 6.263 2.541 7.203 1.847 0.265 6.102 0.271 0.907 0.263 0.143 0.430 1.236 0.043 0.191 0.118 0.625 0.052 0.109 7.035 2.886 0.858

3.862 3.862 3.862 3.862 1.114 1.114 1.114 0.953 0.953 0.953 5.071 5.071 5.071 5.071 0.667 0.667 0.667 0.781 0.781 0.781 5.794 5.794 7.212 7.212 7.212 7.212 5.794 5.794 0.916 0.916 0.916 0.590 0.590 0.590 0.204 0.204 0.204 0.206 0.206 0.206 7.212 7.212 5.794

2.859 3.692 3.342 4.431 1.753 0.917 0.465 0.548 1.128 0.238 3.881 4.679 3.084 6.586 1.731 0.449 0.940 0.637 0.083 1.196 1.935 6.033 7.343 3.154 8.737 1.101 2.203 6.692 0.687 0.015 2.885 0.428 0.091 2.260 0.825 0.464 0.616 1.443 1.338 0.223 7.092 1.645 2.541

3.851 3.851 3.851 3.851 0.642 0.642 0.642 0.809 0.809 0.809 5.062 5.062 5.062 5.062 0.665 0.665 0.665 0.830 0.830 0.830 5.997 5.997 7.533 7.533 7.533 7.533 5.997 5.997 1.771 1.771 1.771 0.904 0.904 0.904 0.165 0.165 0.165 1.239 1.239 1.239 7.533 7.533 5.997 (Continued)

562

Structural design optimization considering uncertainties

Table 20.2 (Continued) Member/ linked set

Integer results for areas

Floating points results for areas

Squared inches

Squared inches

Nominal

7-15/7 8-15/13 7-16/13 17-19/14 18-20/14 19-21/14 20-22/14 19-18/15 17-20/15 19-20/16 21-20/16 19-22/16 24-26/17 23-25/17 26-28/17 25-27/17 24-25/18 23-26/18 25-26/19 26-27/19 25-28/19 Variable Volume (in3 ) Volume (m3 )

Convex

Nominal

Convex

Un-linked U-N-I

Linked L-N-I

Un-linked U-C-I

Linked L-C-I

Un-linked U-N-F

Linked L-N-F

Un-linked U-C-F

Linked L-C-F

6 1 1 7 2 7 3 1 2 2 1 1 1 6 1 6 1 1 1 1 1 U-N-I 31840 0.522

6 1 1 7 7 7 7 1 1 1 1 1 5 5 5 5 2 2 1 1 1 L-N-I 43090 0.706

7 1 1 8 2 8 3 1 2 1 1 1 1 7 1 7 1 1 1 1 1 U-C-I 35320 0.579

6 1 1 7 7 7 7 1 1 1 1 1 6 6 6 6 1 1 1 1 1 L-C-I 44520 0.730

5.393 0.553 0.010 7.312 2.022 7.326 1.433 0.116 0.128 0.174 0.267 0.152 0.121 6.315 0.060 6.088 0.391 0.194 0.918 0.405 0.534 U-N-F 25160 0.412

5.794 0.423 0.423 7.076 7.076 7.076 7.076 0.241 0.241 0.223 0.223 0.223 5.670 5.670 5.670 5.670 0.259 0.259 0.222 0.222 0.222 L-N-F 37970 0.622

6.071 2.705 0.026 8.815 1.814 8.850 1.035 0.048 0.160 0.434 0.160 0.137 1.297 6.943 0.418 7.094 0.612 0.587 0.322 1.676 0.090 U-C-F 31950 0.524

5.997 0.051 0.051 7.436 7.436 7.436 7.436 0.280 0.280 0.096 0.096 0.096 5.557 5.557 5.557 5.557 0.748 0.748 0.832 0.832 0.832 L-C-F 40460 0.663

a Note: 1 in2 = 645.2 mm2 . b Note: In this Table the following symbols are used. U = unlinked (64 independent variables); L = linked

(19 independent variables); N = nominal case; C = convex model case; F = floating points; I = integer values.

fully explored. If convex models could be shown to be viable for nonlinear structures, their reputation in the uncertainty arena would grow. As with other techniques for handling uncertainty, probability methods, for example, nonlinear structures present a higher degree of difficulty. 5.2 Ro b ustn ess o f s t r uc t ur es The convex model method requires that the uncertainties are bound within a convex set. Therefore, the percentage of uncertainty βn is fixed. It is desirable to study different degrees of uncertainty. In order to do so, several values of βn should be considered. Allowing the percent of uncertainty to vary, a nested series of convex sets is obtained. The larger the uncertainty, the larger is the set. Robustness expresses the greatest level

Genetic algorithms in structural optimum design using convex models

563

of uncertainty at which failure cannot occur. Therefore, it is advantageous to allow the uncertainties to have a large variation from the nominal value without the collapse of the structure. In other words, a large robustness is sought (Ben-Haim 2001). However, to tolerate a large uncertainty, it is necessary to sacrifice the performance of the design. One can think of the performance as the structural volume expressed as a function of the cross-sectional areas and the load magnitudes. In optimal design, we would say that the performance is the value of the objective function dependent upon the design parameters and the uncertainties. The critical performance is the minimum level of acceptable performance. The robustness can be plotted versus the performance to obtain a design curve. It has been demonstrated that the design curve is monotonic (Ben-Haim 1996). This implies that there is a trade-off in deciding which point on the design curve is the “working point,’’ i.e., the optimal design. The decision of where to choose the working point on the design curve takes place during the design process, and it is dependent on the desired robustness for the structure. Robust design of trusses has gained recent attention in (Au et al. 2003) and (Kanno and Takewaki 2006a and 2006b) and is an area open to further investigation.

Acknowledgments Funding for this project has been provided by the McDonald Work Award and the Gonzaga Research Council. The results for the optimal structural designs were obtained by Matthew Burkhart, Andrew Burton, and Jared Smith, Gonzaga University alumni of the Department of Computer Science. The authors would like to acknowledge the other members of GUCEA (Gonzaga University Center for Evolutionary Algorithms) who are currently collaborating with them on research. A special thank you goes to Dr. Shannon Overbay for her involvement in the research on genetic algorithms.

References Attoh-Okine, N.O. 2002. Uncertainty analysis in structural number determination in flexible pavement design – A convex model approach. Construction and Building Materials 16(2): 67–71. Au, F.T.K., Cheng, Y.S., Tham, L.G. & Zeng, G.W. 2003. Robust design of structures using convex models. Computers and Structures 81(28–29):2611–2619. Ben-Haim, Y. 1996. Robust reliability in the mechanical sciences. Berlin: Springer-Verlag. Ben-Haim, Y. 2001. Information-gap decision theory: Decisions under severe uncertainty. New York: Academic Press, Inc. Ben-Haim, Y. & Elishakoff, I. 1996. Convex models of uncertainty in applied mechanics. New York: Elsevier. Ben-Haim, Y., Chen, G. & Soong, T.T. 1996. Maximum structural response using convex models. ASCE J. Engineering Mechanics 122:325–333. Cho, M. & Rhee, S.Y. 2003. Layup optimization considering free-edge strength and bounded uncertainty of material properties. AIAA Journal 41(11):2274–2282. Cho, M. & Rhee, S.Y. 2004. Optimization of laminates with free edges under bounded uncertainty subject to extension, bending and twisting. International Journal of Solids and Structures 41(1):227–245. Elishakoff, I. 1995. An Idea on the uncertainty triangle. The Shock and Vibration Digest 22(10): 1–1.

564

Structural design optimization considering uncertainties

Elishakoff, I., Haftka, R.T. & Fang, J. 1994. Structural design under bounded uncertainty optimization with anti-optimization. Computers and Structures 53(6):1401–1405. Elseifi, M.A, Gurdal, Z. & Nikolaidis, E. 1998. Convex and probabilistic models of uncertainties in geometric imperfections of stiffened composite panels. In Anon (ed.), AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference and Exhibit and AIAA/ASME/AHS Adaptive Structures Forum, Long Beach,CS, USA, April 20–23 1998 Part 2 (of 4):1131–1140. Ganzerli, S. & Pantelides, C.P. 2000. Optimum structural design via convex model superposition. Computers and Structures 74(6):639–647. Ganzerli, S. & Burkhart, M.F. 2002. Genetic algorithms for optimal structural design using convex models of uncertainties. In Spanos & Deodatis (eds), Fourth International Conference on Computational Stochastic Mechanics (CSM4); Proc. Intern. Conf., Kerkyra (Corfu), Greece, 9–12 June 2002. Rotterdam: Millpress. Ganzerli S., De Palma, P., Stackle, P. & Brown, A. 2005. Info-gap uncertainty in structural optimization via genetic algorithms. In G. Augusti, G.I. Schuëller & M. Ciampoli (eds), ICOSSAR’05, 9th International Conference on Structural Safety and Reliabilit; Proc. Intern. Conf., Rome, Italy, June 19–22, 2005. Rotterdam: Millpress: 2325–2330. Ganzerli, S., DePalma, P., Smith, J.D. & Burkhart, M.F. 2003. Efficiency of genetic algorithms for optimal structural design considering convex models of uncertainty. In Armen Der Kiureghian, Samer Madanat & Juan M. Pestana (eds), Ninth International Conference on Statistic and Probability on Civil Engineering (ICASP9); Proc. Intern. Conf., Berkeley, CA, July 6–9, 2003. Rotterdam: Millpress: 1003–1010. Ghasemi, M.R., Hinton, E. & Wood, R.D. 1999. Optimization of trusses using genetic algorithms for discrete and continuous variables. Engineering Computations 16(3):272–301. Goldberg, D.E. 1989. Genetic algorithms in search optimization and machine learning. New York: Addison-Wesley. Gonzaga University Center for Evolutionary Algorithms (GUCEA), http://www.cs.gonzaga.edu/ gucea/ Haupt, L.H. & Haupt, S.E. 1998. Practical genetic algorithms. New York: John Wiley & Sons, Inc. Holland, J.H. 1975. Adaptation in natural and artificial systems. Ann Arbor: The University of Michigan Press. Kanno, I. & Takewaki, Y. 2006a. Robustness analysis of trusses with separable load and structural uncertainties. International Journal of Solids and Structures 43(9):2646–2669. Kanno, I. & Takewaki, Y. 2006b. Sequential semidefinite program for maximum robustness design of structures under load uncertainty. Journal of Optimization Theory and Applications 130(2):265–287. Kim, T.-U. & Sin, H.-C. 2001. Optimal design of composite laminated plates with the discreteness in ply angles and uncertainty in material properties considered. Computers and Structures 79(29–30):2501–2509. Kirsch, U. 1981. Optimum structural design. New York: McGraw-Hill, Inc. Overbay, S., Ganzerli, S., De Palma, P., Stackle, P. & Brown, A. 2006. Trusses, NPCompleteness, and Genetic Algorithms. In F.A. Charney, D.E. Grierson, M. Hoit & J.M. Pestana (eds), 17th Analysis and Computation Specialty Conference; Proc. Conf., Saint Louis, MO, May 18–20, 2006. Reston: ASCE Publications. Pantelides, C.P. & Tzan, S.-R. 1996. Convex models for seismic design of structures – I. Earthquake Eng. Struct. Dyn. 25(9):927–944. Rajeev, S. & Krishnamoorthy, C.S. 1997. Genetic algorithms-based methodologies for design optimization of trusses. ASCE J. of Structural Engineering 123(3):350–358.

Genetic algorithms in structural optimum design using convex models

565

Schemit, L.A. & Lai, Y.C. 1994. Structural optimization based on preconditioned conjugate gradient analysis methods. International Journal for Numerical Methods in Engineering 37(6): 943–964. Spletzer, J.R. & Taylor, J.C. 2003. A bounded uncertainty approach to multi-robot localization. In Anon (ed.), IEEE International Conference on Intelligent Robots and Systems; Proc. Intern. Conf., Las Vegas, October 27-312003. Institute of Electrical and Electronics Engineers Inc. Tonon, F., Bernardini, A. & Elishakoff, I. 2001. Hybrid analysis of uncertainty: Probability, fuzziness and anti-optimization. Chaos, Solutions and Fractals 12(8):1403–1414. Qiu, Z. 2003. Comparison of static response of structures using convex models and interval analysis method. Numerical Methods in Engineering 56(12):1735–1753. Qiu, Z. & Wang, X. 2006. Interval analysis method and convex models for impulsive response of structures with uncertain-but-bounded external loads. Acta Mecanica Sinica 22(3):265–276. Wall, M. 1995. GAlib. A C++ library of genetic algorithm components. Available at http://lancet.mit.edu/ga Wang, C.K. 1986. Structural analysis on microcomputers. New York, NY: Macmillan.

Chapter 21

Metamodel-based computational techniques for solving structural optimization problems considering uncertainties Nikos D. Lagaros National Technical University of Athens, Athens, Greece

Yiannis Tsompanakis Technical University of Crete, Chania, Greece

Michalis Fragiadakis University of Thessaly, Volos, Greece

Vagelis Plevris National Technical University of Athens, Athens, Greece

Manolis Papadrakakis National Technical University of Athens, Athens, Greece

ABSTRACT: Uncertainties are inherent in engineering problems due to various numerical modeling “imperfections’’ and due to the inevitable scattering of the design parameters from their nominal values. Under this perspective, there are two main optimal design formulations that account for the probabilistic response of structural systems: Reliability-based Design Optimization (RBDO) and Robust Design Optimization (RDO). In this work both type of problems are briefly addressed and realistic engineering applications are presented. The optimization part of the proposed probabilistic formulations is solved utilizing efficient evolutionary methods. In both types of problems the probabilistic analysis is carried out with the Monte Carlo Simulation (MCS) method incorporating the Latin Hypercube Sampling (LHS) technique for the reduction of the sample size. In order to achieve further improvement of the computational efficiency a Neural Network (NN) is used to replace the time-consuming FE analyses required by the MCS. Moreover, various sources of randomness that arise in structural systems are taken into account in a “holistic’’ probabilistic perception by implementing a Reliability-based Robust Design Optimization (RRDO) formulation, where additional probabilistic constraints are incorporated into the standard RDO formulation. The proposed RRDO problem is formulated as a multi-criteria optimization problem using the non-dominant Cascade Evolutionary Algorithm (CEA) combined with the weighted Tchebycheff metric.

1 Introduction The basic engineering task during the development of any structural system is, among others, to improve its performance in terms of constructional or life-cycle cost and structural behaviour. Improvements can be achieved either by using design rules based

568

Structural design optimization considering uncertainties

on the experience of the engineer, or via an automated manner by using optimization methods that lead to optimum structural designs. Strictly speaking, optimal means that for the formulation considered, no better solution exists. Taking into account the complexity of a structural optimization problem it is obvious that finding the global optimum solution is not an easy task. In real world applications, if uncertainties have not been taken into account, the significance of the optimum solutions would be limited. This is because, although nearly perfect structural models can be simulated in a computing environment, real world structures always have imperfections or deviations from their nominal state defined by the design codes. The optimum that is obtained through the numerical simulation is never materialized in an absolute way and as a result a near optimal solution is always applied in practice. A formulation of a structural optimization problem that ignores the scattering of the various design parameters is defined as a deterministic one. A numerically feasible optimum design, according to the deterministic formulation, once applied in a real physical system, may lose its feasibility due to the unavoidable dispersion on the values of structural parameters (material properties, dimensions, loads, etc). This happens because the performance of the applied design may be far worse than expected. In order to account for the randomness of the most important parameters that affect the simulation and the response of a structure, a different formulation of the optimization problem based on stochastic analysis methodologies has to be used. The recent developments on the stochastic analysis methods (Schuëller 2005), has stimulated the interest for the probabilistic optimum design of structures. Over the last decade efficient probabilistic-based optimization formulations have been developed in order to account for the various uncertainties that are involved in structural design. There are two distinguished design formulations that account for the probabilistic systems response: Robust Design Optimization (RDO) (see Messac and Ismail-Yahaya (2002), Jung and Lee (2002), Doltsinis and Kang (2004), Lagaros and Papadrakakis (2006), among others), while detailed literature overview on RDO problems can be found in the work of (Park et al. 2006), and Reliability-based Design Optimization (RBDO) (see Frangopol and Soares (2001), Agarwal and Renaud (2004), Tsompanakis and Papadrakakis (2004), Youn et al. (2005), Agarwal and Renaud (2006), Ba-abbad et al. (2006) among others). RDO methods primarily seek to minimize the influence of stochastic variations on the nominal values of the design parameters. On the other hand, the main goal of RBDO methods is to design for minimum weight/cost, which satisfies the allowable probability of failure for certain limit state(s). In this study three characteristic probabilistic optimization problems of realistic steel structures are presented, in which efficient metamodels based on Neural Networks (NN) are incorporated in order to improve the computational efficiency of the proposed methodologies. In all test examples considered, the randomness of loads, material properties, and member dimensions is taken into consideration using the Monte Carlo Simulation (MCS) method combined with Latin Hypercube Sampling (LHS). In order to deal with the increased computational cost required, despite the use of the LHS technique, by the MCS for lower limits of the probability of violation of the constraints, a NNbased methodology is adopted for obtaining computationally inexpensive estimates of the response required during the stochastic analysis. The use of NN is motivated by the approximate concepts inherent in stochastic analysis and the time consuming repeated analyses required for MCS. In each case a specially tailored NN is trained, utilizing

M e t a m o d e l-b a s e d c o m p u t a t i o n a l t e c h n i q u e s i n p r o b a b i l i s t i c o p t i m i z a t i o n

569

available information generated from selected conventional analyses. Subsequently, the trained NN is used to fast and accurately predict the output data for the next sets of random variables. It appears that the use of a properly selected and trained NN can eliminate any limitation on the sample size used for MCS and on the dimensionality of the problem, due to the drastic reduction of the computing time required for the repeated finite element analyses. Firstly, the reliability-based sizing optimization of large-scale multistorey 3D frames is investigated. The objective function is the weight of the structure while the constraints are both deterministic (stress and displacement limitations) and probabilistic (the overall probability of failure of the structure). Randomness of loads, material properties, and member geometry are taken into consideration in the reliability analysis using Monte Carlo simulation. The probability of failure of the frame structures is determined via a limit elasto-plastic analysis. The optimization part is solved using Evolution Strategies (ES), while the limit elasto-plastic analyses required during the MCS are replaced by fast and accurate NN predictions. Secondly, an efficient methodology is presented for performing RBDO of steel structures under seismic loading. Optimum earthquake-resistant design of structures using probabilistic analysis and performance-based design criteria is an emerging field of structural engineering. The modern conceptual approach of seismic structural design constitutes the so-called Performance-based Earthquake Engineering or PBEE (for details see the excellent book by (Bozorgnia and Bertero 2004)). An important ingredient of PBEE is structural reliability (Wen 2000): a straightforward consideration of all uncertainties and variabilities that arise in structural design, construction and serviceability in order to be able to calculate the level of confidence about the structure’s ability to meet the desired performance goals. Due to the uncertain nature of the earthquake loading, structural design is often based on design response spectra of the region of interest and on some simplified assumptions on the structural behaviour under earthquake. In this test example the reliability-based sizing optimization of multistorey steel frames under seismic loading is investigated, in which the optimization part of RBDO is solved utilizing Evolution Strategies (ES) algorithm. The objective function is the weight of the structure, while the constraints are both deterministic (stress and displacement restrictions imposed by the design codes) and probabilistic (limitation on the overall probability failure of the structure which is defined in terms of maximum interstorey drift). Finally, a hybrid Reliability-based Robust Design Optimization (RRDO) formulation is presented, where probabilistic constraints are incorporated into the standard RDO formulation. A similar RRDO formulation has been used in the work of Youn and (Choi 2004), where a performance moment integration method is proposed that employs a numerical integration scheme for output response to estimate the product quality loss. The proposed RRDO is formulated as a multi-criteria optimization problem using the non-dominant Cascade Evolutionary Algorithm (CEA) combined with the weighted Tchebycheff metric. The main goal of this approach is to account for the influence of probabilistic constraints in the framework of structural RDO problems, by comparing the RRDO formulation with the standard one. For this purpose, a characteristic test example of a 3D steel truss is investigated, where the objective functions considered in the RRDO formulation are the weight and the variance of the response of the structure, represented by a characteristic node displacement. During

570

Structural design optimization considering uncertainties

the optimization process each structural design is checked whether it satisfies the provisions of the European design codes for steel structures (EC3 2003) with a prescribed probability of violation.

2 Formulations of probabilistic structural optimization problems Generally, in structural optimization problems the aim is to minimize the weight of the structure under certain deterministic behavioural constraints usually imposed on stresses and displacements. The significant developments of stochastic analysis methods have stimulated the interest for their application in structural design resulting to two main categories of probabilistic optimum design formulations: Reliability-based Design Optimization (RBDO) and Robust Design Optimization (RDO). The main goal of RBDO methods is to achieve increased safety levels of the structure with respect to variations of the random design parameters, while RDO methods primarily seek to minimize the influence of stochastic variations on the mean design of a structural system. Since the aforementioned method can be complementary to each other, hybrid Reliability-based Robust Design Optimization (RRDO) formulations have also been presented, where probabilistic constraints are incorporated into the standard RDO formulation. There are also several other probabilistic optimization formulations, for example those based on convex set models, evidence theory, possibility theory, etc, which are described in other chapters of the present volume. In the sequence, the three aforementioned major types of stochastic optimization formulations are briefly described. 2.1 Rel i a b i l i ty-b as e d d es ig n o pt imizat i o n In reliability-based optimal design additional probabilistic constraints are imposed in the standard deterministic formulation, in order to take into account various random parameters and to ensure that the probability of failure for the whole structure or some of its critical members is within acceptable limits. The probabilistic constraints enforce the condition that the probability of exceeding a certain limit state’s threshold value is smaller than a certain value (usually from 10−3 to 10−5 ). Under this perspective, a discrete RBDO problem can be formulated in the following form: min CIN (s, x) subject to gj (s, x) ≤ 0 j = 1, . . . , m pf (s, x) ≤ pall

(1)

where CIN (s, x) is the objective function (i.e. the structural weight or the initial construction cost) to be minimized, s (which can take values only from the given discrete set Rd ) and x are the vectors of the design and random variables, respectively. Regarding the constraints, gj (s, x) are the deterministic constraint functions and pf (s, x) is the probability of failure of the design that it is bound by an upper allowable probability equal to pall . Most frequently, the deterministic constraints of the structure are the member stresses and nodal displacements or interstorey drifts. Furthermore, due to

M e t a m o d e l-b a s e d c o m p u t a t i o n a l t e c h n i q u e s i n p r o b a b i l i s t i c o p t i m i z a t i o n

571

engineering practice demands, the members are divided into groups having the same design variables. This linking of elements results in a trade-off between the use of more material and the need of symmetry and uniformity of structures due to practical considerations. Furthermore, it has to be taken into account that due to manufacturing limitations the design variables are not continuous but discrete since cross-sections belong to a certain set. 2.2

Robust des ign optimization

In practical applications, optimizing a single objective function, most often the material weight or cost, cannot capture every aspect related to the performance of the structure. Actually, in real world optimization problems, there are several conflicting and usually incommensurable criteria that have to be dealt with simultaneously. Such problems are called multi-objective or multi-criteria optimization problems. In addition, in the majority of cases the objective functions are conflicting and as a result there exists no unique point which represents the optimum for all of them. Consequently, the common optimality condition used in single-objective optimization must be replaced by a “multi-collective’’ concept, the so-called Pareto optimum. Thus, in the multi-criteria formulation of a robust design structural sizing optimization problem, implemented in this work, an additional objective function is considered which is related to the influence of the random nature of the structural parameters on the performance of the structure. The aim is to minimize both the weight and the variance of the response of the structure. The mathematical formulation of the RDO problem is as follows min [CIN (s, x), StDevu (s, x)]T subject to gj (s, x) ≤ 0 j = 1, ..., k

(2)

where CIN (s, x) is the initial construction cost and StDevu (s, x) is the standard deviation of the response that correspond to the two objectives to be minimized, s and x are the vectors of the design and random variables respectively and gj (s, x) are the deterministic constraint functions. 2.3

Reliability-bas ed robus t des ign opt i mi zati o n

In a combined RRDO formulation the constraint functions can also vary, due to the random nature of the structural parameters. In the proposed RRDO formulation the probability of violation of the constraints is taken into account as an additional constraint function. The mathematical formulation of the RRDO problem implemented in this work is as follows min subject to

[CIN (s, x), StDevu (s, x)]T gj (s, x) ≤ 0 j = 1, ..., k pv,max (s, x) ≤ pall

(3)

where CIN (s, x) is the initial construction cost and StDevu (s, x) is the standard deviation of the response that correspond to the two objectives to be minimized, s and x are

572

Structural design optimization considering uncertainties

the vectors of the design and random variables respectively, gj (s, x) are the deterministic constraint functions, while pv,max (s, x) is the maximum probability of violation, among the k behavioural constraint functions, that it is bound by an upper allowable probability equal to pall . In this study three types of deterministic behavioural constraints are imposed to the sizing optimization problem of the truss structure examined: (i) stress, (ii) compression force (for buckling) and (iii) displacement constraints. On the other hand, the employed probabilistic constraint enforces the condition that the probabilities of violation of certain limit state functions are smaller than a certain value.

3 Solving the optimization problem As mentioned in the previous section, two types of optimization problems are encountered in the framework of this study: a single and a multi-objective one. Evolutionary based algorithms are employed for tackling both of them. The two most widely used optimization algorithms belonging to the class of evolutionary computation that imitate nature by using biological methodologies are the Genetic Algorithms (GA) and Evolution Strategies (ES). Initially the ES method was introduced in the seventies for mathematical type of optimization problems (see Schwefel 1981). In this work ES are used as the optimization tool for addressing demanding probabilistic optimization problems. Both GA and ES imitate biological evolution in nature and have three characteristics that make them differ from mathematical optimization algorithms: (i) instead of the usual deterministic operators, they use randomised operators, (ii) instead of a single design point, they work simultaneously with a population of design points, (iii) they can handle continuous, discrete and mixed optimization problems. The second characteristic allows for a natural implementation of ES on parallel computing environments (Papadrakakis et al. 1999). Structural optimization problems have been mainly treated with mathematical programming algorithms, such as the sequential quadratic programming (SQP) method, which need gradient information. In structural optimization problems, and especially when uncertainties are considered, the objective function and the constraints are particularly highly non-linear functions of the design variables, thus the computational effort spent in gradient calculations is usually excessive. In studies by (Papadrakakis et al. 1999) and (Lagaros et al. 2002), it was found that probabilistic search methods are computationally more efficient than mathematical programming methods, even though more optimization steps are required in order to reach the optimum. In the former case the optimization steps are computationally less expensive than in the latter case since there is no need for gradient information. 3.1 So l v i n g th e s ing le o b je c t ive o pt imi z a t i o n p r o b l e m The absence of sensitivity analysis in evolutionary methods has even greater importance in the case of probabilistic problems, since the calculation of the derivatives of the reliability constraints is very time-consuming. Furthermore, these methodologies can be considered, due to their random search, as global optimization methods because they are capable of finding the global optimum, whereas mathematical programming algorithms may be trapped in local optima. As it can be seen in Flowchart 21.1, the ES optimization procedure initiates with a set of parent vectors. If any of these

M e t a m o d e l-b a s e d c o m p u t a t i o n a l t e c h n i q u e s i n p r o b a b i l i s t i c o p t i m i z a t i o n

1. 2. 3. 4. 5. 6. 7. 8.

573

Selection step: selection of si (i=1, 2, . . . , µ) parent design vectors Analysis step: perform structural analysis (i=1, 2, . . . , µ) Constraints check: all parent become feasible Offspring generation: generate sj , ( j=1, 2, . . . , λ) offspring design vectors Analysis step: perform structural analysis ( j=1, 2, . . . , λ) Constraints check: if satisfied continue, else go to step 4 Selection step: selection of the next generation parent design vectors Convergence check: If satisfied stop, else go to step 4

Flowchart 21.1 The ES algorithm for single-objective optimization problems.

parent vectors gives an infeasible design, then it is modified until it becomes feasible. Subsequently, the offspring design vectors are generated and checked if they are in the feasible region. According to the (µ+λ) selection scheme, in every generation the values of the objective function of the parent and the offspring vectors are compared and the worst vectors are rejected, while the remaining ones are considered to be the parent vectors of the new generation. This procedure is repeated until the chosen termination criterion is satisfied. 3.2 Solving the multi-objective optimizati o n p ro bl e m A number of techniques have been developed in the past, that adequately deal with the multi-objective optimization problem (Coello-Coello 2000, Mattson et al. 2004, Marler and Arora 2004). The multi-objective algorithm employed in this work belongs to the hybrid methods, where an evolutionary algorithm is combined with a scalarizing function. In general, when using scalarizing functions, locally Pareto optimal solutions are obtained. Global Pareto optimality can be guaranteed only when the objective functions and the feasible region are both convex or quasi-convex and convex, respectively. For non-convex cases, such as the majority of structural multi-objective optimization problems, a global single objective optimizer is required. In this work the non-dominant Cascade Evolutionary Algorithm using the augmented Tchebycheff metric (CEATm) is employed for the solution of the Pareto optimization problem at hand. This implementation was proposed by the authors in a previous work by (Lagaros et al. 2005), where more details of the present implementation can be found. The basic steps of the CEATm algorithm are outlined below in Flowchart 21.2, where it is obvious that the CEATm optimization scheme can easily be applied in two parallel computing levels, an external and an internal one. In addition, the multi-objective optimization problem is converted into a series of single objective optimization problems, where the solution of each subproblem can be performed concurrently.

4 Probabilistic analysis using Monte Carlo simulation The reliability of a structure or its probability of failure is an important factor in the design procedure since it quantifies the probability under which a structure will fulfill

574

Structural design optimization considering uncertainties

Independent run i, i=1, . . . , nrun Generate/update the weight coefficients w i,j j=1, . . . , m of the Tchebycheff metric. CEATm LOOP 1. Initial generation: 1a. Generate sk (k=1, . . . , µ) vectors 1b. Structural analysis step 1c. Evaluation of the Tchebycheff metric 1d. Constraint check: if satisfied k=k+1 else k = k. Go to step 1a 2. Global non-dominant search: Check if global generation is accomplished. If yes, then non-dominant search is performed, else wait until global generation is accomplished. 3. New generation: 3a. Generate s (=1, . . . , λ) vectors 3b. Structural analysis step 3c. Evaluation of the Tchebycheff metric 3d. Constraint check: if satisfied =+1 else =. Go to step 3a 4. Selection step: selection of the next generation parents according to (µ + λ) or (µ, λ) scheme 5. Global non-dominant search: Check if global generation is accomplished. If yes, then non-dominant search is performed, else wait until global generation is accomplished. 6. Convergence check: If satisfied stop, else go to step 5 END OF CEATm LOOP End of Independent run i Flowchart 21.2 The CEATm algorithm for multi-objective optimization problems.

its design requirements. Structural reliability analysis, or probabilistic analysis is a tool that assists the design engineer to take into account all possible uncertainties during the design, construction phases and lifetime of a structure in order to calculate its probability of failure pf , or probability of a limit state violation pviol . In structural reliability analysis problems, the probability of violation of a limit state function, expressed as G(x) < 0, can be written as
pviol =

fx (x) dx

(4)

G(x)≥0

where x = [x1 , x2 , . . . , xM ]T is a vector of the random structural parameters and fx (x) denotes the joint probability of violation for all random structural parameters. In probabilistic analysis of structures the Monte Carlo Simulation (MCS) method is very popular and particularly applicable when an analytical solution is not attainable. This is mainly the case in problems of complex nature with a large number of random variables where all other probabilistic analysis methods are not applicable. Despite its simplicity, MCS method has the capability of handling practically every possible case regardless of its complexity; it requires, though, excessive computational effort. In order to improve the computational efficiency of MCS, various techniques have been proposed.

M e t a m o d e l-b a s e d c o m p u t a t i o n a l t e c h n i q u e s i n p r o b a b i l i s t i c o p t i m i z a t i o n

575

Since MCS is based on the theory of large numbers (N∞ ) an unbiased estimator of the probability of violation is given by ∞ 1  I(xj ) N∞

N

pviol =

(5)

j=1

in which xj is the j-th vector of the random structural parameters, and I(xj ) is an indicator for successful and unsuccessful simulations defined as  I(xj ) =

1 if G(xj ) ≥ 0 0 if G(xj ) < 0

(6)

In order to estimate pviol an adequate number of N independent random samples are produced. The value of the violation function is computed for each random sample xj and the Monte Carlo estimation of pviol is given in terms of sample mean by NH pviol ∼ = N

(7)

where NH is the number of successful simulations and N the total number of simulations. In general, a vast number of simulations have to be performed in order to achieve great accuracy, especially for low values of probability of failure. In an effort to reduce the excessive computation cost of crude MCS using purely random sampling methodologies, which is considered as the drawback of the method, various sampling reduction techniques have been proposed. Among them are the importance sampling, adaptive sampling technique, stratified sampling, antithetic variate technique, conditional expectation technique, and Latin Hypercube Sampling (LHS), which was introduced by (MacKay et al. 1979). Although LHS is generally recognized as one of the most efficient size reduction techniques it has been proven to be efficient only in the case that relatively large probability of violation is to be calculated and in the case of the calculation of statistical quantities like the mean value and the standard deviation. In most other cases MCS-LHS performs like the crude MCS (Owen 1997). In the LHS method, the range of probable values for each random variable is divided into M non-overlapping segments of equal probability of occurrence. Thus, the whole parameter space, consisting of N parameters, is partitioned into MN cells. Then the random sample generation is performed, by choosing M cells from the MN space with respect to the density of each interval, and the cell number of each random sample is calculated. The cell number indicates the segment number that the sample belongs to with respect to each of the parameters. Using LHS technique, sampling is realized independently, whereas, matching of random samples is performed either randomly, or in a restricted manner. All necessary random samples are produced and they are accepted only if they do not agree with any previous combination of the segment numbers. The advantage of the LHS approach is that the random samples are generated from all the ranges of possible values, thus giving a more thorough insight into the tails of the probability distributions.

576

Structural design optimization considering uncertainties

5 NN-based MCS for stochastic analysis Over the last ten years artificial intelligence techniques like neural networks (NNs) have emerged as a powerful tool that could be used to replace time consuming procedures in many engineering applications (Lagaros and Tsompanakis 2006), (Tsompanakis et al. 2007). Some of the fields where NNs have been successfully applied are: pattern recognition, regression (function approximation/fitting), optimization, nonlinear system modelling, identification, damage assessment, etc. Function approximation involves approximating the underlying relationship from a given finite input-output data set. Feed-forward NNs, such as multi-layer perceptrons (MLP) and radial basis function networks, have been widely used as an alternative approach to function approximation since they provide a generic functional representation and have been shown to be capable of approximating any continuous function with acceptable accuracy. A trained neural network presents some distinct advantages over the numerical computing paradigm. It provides a rapid mapping of a given input into the desired output quantities, thereby enhancing the efficiency of the structural analysis process. This major of a trained NN over the conventional procedure, under the provision that the predicted results fall within acceptable tolerances, leads to results that can be produced in a few clock cycles, representing orders of magnitude less computational effort than the conventional computational process. In this work the application of NNs is focused on the simulation (i.e. probabilistic analysis of structures) of demanding computational problems of probabilistic mechanics. Many sources of uncertainty (material, geometry, loads, etc) are inherent in structural design and functioning. Probabilistic analysis of structures leads to safety measures that a design engineer has to take into account due to the aforementioned uncertainties. Probabilistic analysis problems, especially when earthquake loadings are considered, are highly computationally intensive tasks since in order to calculate the structural behaviour under seismic loads a large number of numerical analyses are required. In general, soft computing techniques are used in order to reduce the aforementioned computational cost. The aim of the present study is to train a neural network to provide computationally inexpensive estimates of analysis outputs required during the MCS process. In the present work the ability of neural networks to predict characteristic measures that quantify the response of a structure considering uncertainties is presented. This objective comprises the following tasks: (i) select the proper training set, (ii) find suitable network architecture, and (iii) perform the training/testing of the neural network. The learning algorithm, which was employed for the training, is the well-known Back-Propagation (BP) algorithm (Rummelhart and (McClelland 1986). An important factor governing the success of the learning procedure of NN architecture is the selection of the training set. A sufficient number of input data properly distributed in the design space together with the output data resulting from complete structural analyses are needed for the BP algorithm in order to provide satisfactory results. Overloading the network with unnecessary similar information results to over training without increasing the accuracy of the predictions. The required training patterns are generated randomly using the LHS technique, where a parametric study is performed for defining the size of the training set for the efficient training of NN. The basic NN configuration

M e t a m o d e l-b a s e d c o m p u t a t i o n a l t e c h n i q u e s i n p r o b a b i l i s t i c o p t i m i z a t i o n

Input layer

Hidden layer

577

Output layer

Figure 21.1 Typical neural network configuration.

employed for all the test cases examined in this study is selected to have one hidden layer, as shown in Figure 21.1.

6 Numerical results 6.1

RBDO of s teel 3D frames under s tati c l o adi ng us i ng elasto-plas tic analys is

Firstly, the reliability-based sizing optimization of multistorey 3D frame structures under static loading is investigated. The objective function is the weight of the structure while the constraints are both deterministic (stress and displacement limitations) and probabilistic (the overall probability of failure of the structure). Randomness of loads, material properties, and member geometry are taken into consideration in reliability analysis using the MCS method. The probability of failure of the frame structures is determined via a limit elasto-plastic analysis. The optimization part is solved using ES and two methodologies combining evolution strategies and neural networks (ES-NN) are examined. In the first one, a trained NN utilizing information generated from a number of properly selected design vectors, computed by conventional finite element and reliability analyses, is used to perform both deterministic and probabilistic constraints checks during the optimization process. The data obtained from these analyses are processed in order to obtain the necessary input and output pairs which are subsequently used for training the NN. The trained NN is then applied to predict the response of the structure in terms of deterministic and probabilistic constraints checks due to different sets of design variables. The NN training is considered successful when the predicted values resemble closely the corresponding values of the conventional analyses which are considered exact. In the second methodology the limit elasto-plastic

578

Structural design optimization considering uncertainties

analyses required during the MCS are replaced by NN predictions of the structural behaviour up to collapse. For every MCS that is required in order to perform the probabilistic constraints check, a NN is trained utilizing available information generated from selected conventional elasto-plastic analyses. The limit state analysis data are processed to obtain input and output pairs, which are used for training the NN. The trained NN is then used to predict the critical load factor due to different sets of basic random variables. A fully connected network, as shown in Figure 21.1, is used.

6.1.1 R eli a b il ity-b a s e d s t r u ct u r a l o p t im iza t io n us i ng M C S, E S and N N In reliability analysis of elasto-plastic structures using MCS the computed critical load factors are compared to the corresponding external loading leading to the computation of the probability of structural failure. The probabilistic constraints enforce the condition that the probability of a local failure of the system or the global system failure is smaller than a certain value (i.e. 10−5 to 10−3 ). In this work the overall probability of failure of the structure, as a result of limit elasto-plastic analyses, is taken as the global reliability constraint. The probabilistic design variables are chosen to be the cross-sectional dimensions of the structural members and the material properties (E, σy ). MCS requires a number of limit elasto-plastic analyses that can be dealt independently and concurrently. This allows the natural implementation of the MCS method in parallel computing environment as well. The most straightforward parallel implementation of the MCS method is to assign one limit elasto-plastic analysis to every processor without any need of inter-processor communication during the analysis phase. In the present study the parallel computations were performed on a Silicon Graphics Power Challenge shared memory computer where the number (p) of activated processors is equal to the number of the parent or offspring design vectors since µ = λ.

6.1.2 N N u sed f o r d e t e r m in is t ic a n d p r o b a bi l i s t i c c o ns t r ai nt s c he c k In this methodology, a trained NN utilizing information generated from a number of properly selected design vectors is used to perform both the deterministic and probabilistic constraints checks during the optimization process. After the selection of the suitable NN architecture the training procedure is performed using a number (M) of data sets, in order to obtain the I/O pairs needed for the NN training. The trained NN is then applied to predict the response of the structure in terms of deterministic and probabilistic constraint checks due to different sets of design variables. The combined ES-NN optimization procedure is performed in two phases. The first phase includes the training set selection, the corresponding structural analysis and MCS for each training set required to obtain the necessary I/O data for the NN training, and finally the training and testing of a suitable NN configuration. The second phase is the ES optimization stage where the trained NN is used to predict the response of the structure in terms of the deterministic and probabilistic constraint checks due to different sets of design variables. This algorithm is summarized in Flowchart 21.3.

M e t a m o d e l-b a s e d c o m p u t a t i o n a l t e c h n i q u e s i n p r o b a b i l i s t i c o p t i m i z a t i o n

579

• NN training phase: 1. Training set selection step: select M input patterns 2. Deterministic constraints check: perform the check for each input pattern vector 3. Monte Carlo Simulation step: perform MCS for each input pattern vector 4. Probabilistic constraints check: perform the check for each input pattern vector 5. Training step: training of the NN 6. Testing step: test the trained NN • ES-NN optimization phase: 1. Parents Initialization 2. NN (deterministic-probabilistic) constraints check: all parents become feasible 3. Offspring generation 4. NN (deterministic-probabilistic) constraints check: if satisfied continue, else go to step 3 5. Parents’ selection step 6. Convergence check Flowchart 21.3 The ES-NN1 methodology.

6.1.3

NN predicti on of the cri ti cal l oad i n st ru ct u ra l fa i l u re

In the second methodology the limit elasto-plastic analyses required during the MCS are now replaced by NN predictions of the structural behaviour up to collapse. For every MCS an NN is trained utilizing available information generated from selected conventional elasto-plastic analyses. The limit state analysis data is processed to obtain input and output pairs, which are used for training the NN. The trained NN is then used to predict the critical load factor due to different sets of basic random variables. At each ES cycle (generation) a number of MCS are carried out. In order to replace the time consuming limit elasto-plastic analyses by predicted results obtained with a trained NN, a training procedure is performed based on the data collected from a number of conventional limit elasto-plastic analyses. After the training phase is concluded the trained NN predictions replace the conventional limit elasto-plastic analyses, for the current design. This algorithm is summarized in Flowchart 21.4.

6.1.4 Twent y-st orey space frame RBDO exam p l e A characteristic 3D building frame shown in Figure 21.2, has been tested in order to illustrate the efficiency of the proposed methodologies for reliability-based sizing optimization problems. The cross section of each member of the space frame considered is assumed to be a W-shape and for each structural member one design variable is allocated corresponding to a member of the W-shape data base. The objective function is the weight of the structure. The deterministic constraints are imposed on the interstorey drifts and for each group of structural members, on the maximum stresses due to axial forces and bending moments. The values of allowable axial and bending stresses are Fa = 150 MPa and Fb = 165 MPa, respectively, whereas the allowable interstorey drift is restricted to 1.5% of the height of each storey.

580

Structural design optimization considering uncertainties

1. Parents Initialization 2. Deterministic constraints check: all parents become feasible 3. Monte Carlo Simulation step: 3a. Selection of the NN training set 3b. NN training for the limit load 3c. NN testing 3d. Perform MCS using NN 4. Probabilistic constraints check: all parents become feasible 5. Offspring generation 6. Deterministic constraints check: if satisfied continue, else go to step 5 7. Monte Carlo Simulation step: 7a. Selection of the NN training set 7b. NN training for the limit load 7c. NN testing 7d. Perform MCS using NN 8. Probabilistic constraints check: if satisfied continue, else go to step 5 9. Parents’ selection step 10. Convergence check Flowchart 21.4 The ES-NN2 methodology.

The probabilistic constraint is imposed on the probability of structural collapse due to successive formation of plastic hinges and is set to pall = 0.001. The probability of failure caused by uncertainties related to material properties, geometry and loads of the structures is estimated using MCS with the LHS technique. External loads, yield stresses, elastic moduli and the dimensions of the cross-sections of the structural members are considered to be random variables. The loads follow a log-normal probability density function, while random variables associated with material properties and cross-section dimensions follow a normal probability density function. The twenty-storey space frame shown in Figure 21.2 consists of 1,020 members with 2,400 degrees of freedom. This example is selected in order to show the efficiency of the proposed methodologies in relatively large-scale RBDO problems. The basic load of the structure is a uniform vertical load of 4.78 kPa at each storey and a horizontal pressure of 0.956 kPa acting on the x-z face of the frame. The members of the frame are divided into eleven groups, as shown in Figure 21.4, and the total number of design variables is eleven. The deterministic constraints are twenty-three, two for the stresses of each element group and one for the interstorey drift. The type of probability density functions, mean values, and variances of the random parameters are shown in Table 21.1. A typical load-displacement curve of a node in the top-floor is depicted in Figure 21.3, corresponding to the following design variables: 14WF176, 14WF158, 14WF142, 14WF127, 12WF106, 12WF85, 10WF60, 8WF31, 12WF27, 16WF36, 16WF36. For this test case the (µ + λ)-ES approach is used with µ = λ = 10, while a sample size of 500 and 1,000 simulations is taken for the MCS-LHS. Table 21.2 depicts the performance of the optimization procedure for this test case. As can be seen, the probability of failure corresponding to the optimum computed by the deterministic optimization procedure is much larger than the specified value of 10−3 . For this example the

M e t a m o d e l-b a s e d c o m p u t a t i o n a l t e c h n i q u e s i n p r o b a b i l i s t i c o p t i m i z a t i o n

group 7

group 9

group 8

group 11 1

5

9 13 group 10

17

2

6

10 14 group 10

18

3

7

11

15

19

group 11 8

12 Plan view

16

20

12

group 1

group 2 group 3 group 4

group 5

group 6

4

24

24

24

Front elevation

Figure 21.2 Description of the twenty-storey frame.

Table 21.1 Characteristics of the random variables. Random variable

Probability density function (pdf )

Mean value

Standard deviation (σ)

E σy Design variables Loads

N N N Log-N

200 25.0 si 5.2

0.10E 0.10σy 0.1si 0.2

581

582

Structural design optimization considering uncertainties

Load factor 2.5

2.0

1.5

1.0

0.5

10

20

30

40

(inches)

x-displacement – node 1 top storey

Figure 21.3 Load-displacement curve. Table 21.2 Performance of the methods. Optimization procedure

ES Gens.

pf **

DBO RBDO (500 siml.) RBDO-NN1 (500 siml.) RBDO-NN2 (500 siml.) RBDO (1,000 siml.) RBDO-NN1 (1,000 siml.) RBDO-NN2*

83 126 129 126 120 127 122

0.197 10−0 0.103 10−2 0.102 10−2 0.103 10−2 0.103 10−2 0.101 10−2 0.97 10−3

Optimum weight (kN)

Sequential time (h)

Parallel time (h) p=5

p = 10

p = 20

6,771 9,114 9,121 9,114 9,156 9,172 9,255

2.0 141.0 34.5 15.8 250.3 68.5 17.0

0.7 28.4 7.2 3.3 50.1 13.8 4.1

0.3 14.1 3.5 1.7 25.1 6.9 2.2

0.3 7.1 1.8 0.9 12.6 3.5 1.2

* For 100,000 simulations. ** For each final design and with 100,000 simulations using the NN2 scheme.

increase on optimum weight achieved, when probabilistic constraints are considered, is approximately 26% of the deterministic one, as can be observed in Table 21.2. In Table 21.2 showing the results of the test example, DBO stands for the conventional Deterministic-based Optimization approach, RBDO stands for the conventional Reliability-based Optimization approach, while RBDO-NNi corresponds to the proposed Reliability-based Optimization with NN incorporating algorithm i (i = 1, 2). For the application of the RBDO-NN1 methodology the number of NN input units is equal to the number of design variables. Consequently, the NN configuration used in this case has one hidden layer with 15 nodes resulting in an 11-15-1 NN architecture which is used for all runs. The training set consists of 200 training patterns capturing

M e t a m o d e l-b a s e d c o m p u t a t i o n a l t e c h n i q u e s i n p r o b a b i l i s t i c o p t i m i z a t i o n

583

the full range of possible designs. For the application of the RBDO-NN2 methodology the number of NN input units is equal to the number of random variables, whereas one output unit is needed corresponding to the critical load factor. Consequently the NN configuration with one hidden layer results in a 3-7-1 NN architecture which is used for all runs. The number of conventional step-by-step limit analysis calculations performed for the training of NN is 60 corresponding to different groups of random variables properly selected from the random field. As can be seen in Table 21.2 the proposed RBDO-NN2 optimization scheme manages to achieve the optimum weight in one tenth of the CPU time required by the conventional RBDO procedure in sequential computing implementation. Table 21.2 also depicts the performance of the proposed methodologies in a straightforward parallel mode, with 5, 10 or 20 processors in which 5, 10 or 20 Monte Carlo simulations are performed independently and concurrently. It can be seen that the parallel versions of RBDO, RBDO-NN1 and RBDO-NN2 reached the perfect speedup irrespectively of the number of processors used. The aim of the proposed RBDO procedure was threefold. To reach an optimized design with controlled safety margins with regard to various model uncertainties, while at the same time minimizing the weight of the structure and reducing substantially the required computational effort. The solution of realistic RBDO problems in structural mechanics is an extremely computationally intensive task. In the test example considered in this study the conventional RBDO procedure was found over seventy times more expensive than the corresponding deterministic optimization procedure. The goal of decreasing the computational cost by one order of magnitude in sequential mode was achieved using: (i) NN predictions to perform both deterministic and probabilistic constraints check, or (ii) NN predictions to perform the structural analyses involved in MCS. Furthermore, the achieved reduction in computational time was almost two orders of magnitude in parallel mode with the proposed NN methodologies.

6.2

RBDO of s tructures under s eis mic l o adi ng

In this section the reliability-based sizing optimization of multistorey framed structures under earthquake loading is investigated. The discrete RBDO problem is formulated in the form of Eq. (1), where CIN (s, x) is the initial construction cost to be minimized, s and x are the vectors of the design and random variables respectively, gj (s) are the deterministic stress and displacement constraints. The overall probability of failure of the structure, as a result of multi-modal response spectrum analysis, is taken as the global reliability constraint. Failure is detected when the maximum interstorey drift exceeds a threshold value, here considered as 4% of the storey height, defined as p(θ10/50 > θall ) the probability that the drift θ10/50 for the 10/50 hazard level exceeds the allowable drift θall , that is bound by an upper allowable probability equal to pall . For rigid frames with W-shape cross sections as in this study, the design constraints were taken from the design requirements specified by Eurocode 3 (2003) and Eurocode 8 (2004). During the solution of the optimization problems a number of MCS runs are carried out for each different set of design variables. In order to replace the time consuming FE analyses by predicted results obtained with a trained NN, a training procedure is performed based on the data collected from a number of previously performed FE

584

Structural design optimization considering uncertainties

analyses. After the training phase is concluded the NN predictions replace all conventional FE analyses, for the current design. For the selection of the suitable training pairs, the sample space for each random variable is divided into equally spaced distances. The central points within the intervals are used as inputs for the FE analyses. The random variables considered are the cross-sectional dimensions of the structural members, the material properties (E, σy ) and the loading conditions. Under the proposed approach the FE analyses required during the MCS are replaced by NN predictions of the structural response. For every design a NN is trained utilizing available information generated from selected conventional FE analyses. The trained NN is then used to predict the structural response for different sets of random variables depending on the type of problem examined. 6.2.1 Ea rth q u a ke lo a d in g o f s t e e l f r a m e s In Eurocodes earthquake loading is taken as a random action, therefore it must be considered for the structural design with the following loading combination: Sd =



G “+’’ Ed “+’’



kj

ψ2i Qki

(8)

where “+’’ implies “to be combined with’’,  implies “the combined effect of’’, Gkj denotes the characteristic value of the permanent action j, Ed is the design value of the seismic action, and Qki refers to the characteristic value of the variable action i, while ψ2i is the combination coefficient for quasi permanent value of the variable action i, here taken as 0.30. Design code checks are implemented in the optimization algorithm as constraints. Each structural member should be checked for actions that correspond to the most severe load combination obtained from Eq. (8) and the permanent load combination: Sd = 1.35



Gkj “+’’ 1.50



Qki

(9)

It should be pointed out that the seismic action is obtained from the elastic spectrum reduced by the behaviour factor q. This is done because the structure is expected to absorb energy by deforming inelastically. Maximum values for the q-factor are suggested by design codes and vary according to the material and the type of the structural system. For the framed steel structures considered in this study q = 4.0. The most common approach for the definition of the seismic input is the use of design code response spectra, a general approach that is easy to implement. However, if higher precision is required, the use of spectra derived form natural earthquake records is more appropriate. Since significant dispersion on the structural response due to the use of different natural records has been observed, these spectra must be scaled to the same desired earthquake intensity. The most commonly applied scaling procedure is based on the peak ground acceleration (PGA). Dynamic analysis of simple frames is most frequently performed using the multi-modal response spectrum analysis, which is based on the mode superposition approach and is briefly described in the next section.

M e t a m o d e l-b a s e d c o m p u t a t i o n a l t e c h n i q u e s i n p r o b a b i l i s t i c o p t i m i z a t i o n

6.2.2

585

D yn a mic anal y si s usi ng Mul ti-modal Re s p o n s e Sp e ct ru m

In general, the equations of equilibrium for a finite element system in motion can be written in the usual form Mü(t) + C u(t) ˙ + Ku(t) = R(t)

(10)

where M, C, and K are the mass, damping and stiffness matrices; R(t) is the external load vector, while u(t), u(t) ˙ and u(t) ¨ are the displacement, velocity, and acceleration vectors of the finite element assemblage, respectively. A design approach based on the Multi-modal Response Spectrum (MmRS) analysis, which, in turn, is based on the mode superposition approach, has been used in the present study. The MmRS method is based on a simplification of the mode superposition approach with the aim to avoid time history analyses which are required by both the direct integration and mode superposition approaches. In the case of the multi-modal response spectrum analysis Eq. (10) is modified according to the modal superposition approach to a system of independent equations Mi ÿi (t) + C i y˙ i (t) + Ki yi (t) = Ri (t)

(11)

where Mi =
Ti M
i ,

C i =
Ti C
i ,

Ki =
Ti K
i

and R(t) =
Ti R(t)

(12)

are the generalized values of the corresponding matrices and the loading vector, while
i is the i-th eigenmode shape matrix. According to the modal superposition approach the system of N differential equations, which are coupled with the off-diagonal terms in the mass, damping and stiffness matrices, is transformed to a set of N independent normal-coordinate equations. The dynamic response can therefore be obtained by solving separately for the response of each normal (modal) coordinate and by superposing the response in the original coordinates. In the MmRS analysis a number of different formulas have been proposed to obtain reasonable estimates of the maximum response based on the spectral values without performing time history analyses for a considerable number of transformed dynamic equations. The simplest and most popular one is the Square Root of Sum of Squares (SRSS) of the modal responses. According to this estimate the maximum total displacement is approximated by

umax =

N  i=1

1/2 u2i,max

(13)

ui,max =
i yi,max where ui,max corresponds to the maximum displacement vector corresponding to the i-th eigenmode.

Victoria Mexico (06.09.80) Kobe (16.01.95) Imperial Valley (19.05.40) Duzce (12.11.99) San Fernando (09.02.1971) Gazli (17.05.1976) Friuli (06.05.1976) Aigion (17.05.90) Central California (25.04.54) Alkyonides (24.02.81) Northridge (17.01.94) Athens (07. 09.99) Cape Mendocino (25.04.92) Erzihan,Turkey (13.03.92) Kalamata (13.09.86) Iran (16.09.78) Loma Prieta1 (18.10.89) Loma Prieta2 (18.10.89) Mammoth Lakes (27.05.80) Irpinia, Italy (23.11.80)

*Ms : Surface moment magnitude.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Earthquake name (Date)

Table 21.3 List of natural accelerograms.

Cerro Prieto\Alluvium Kobe\Rock El Centro Array\CWB: D, USGS: C Bolu\CWB: D, USGS: C Pacoima dam\Rock Karakyr, CWB:A Bercis\CWB: B OTE building\Stiff soil Hollister City Hall\CWB: D, USGS: C Korinthos OTE building\Soft soil Jensen filter Plant\CWB: D, USGS: C Sepolia (Metro Station)\Unknown Petrolia\CWB: D, USGS: C Erzikan East-East Comp\CWB: D, USGS: S Kalamata, Prefecture\Stiff soil Tabas\CWB: S Hollister Diff Array\CWB: D Coyote Lke dam\CWB: D McGee Creek\CWB: D Sturno\Unknown

Site\Soil Conditions 45 0 180 90 164 90 0 90 271 90 292 0 90 270 0 0 255 285 0 270

Orientation 6.4 6.95 7.2 7.3 6.61 7.3 6.5 4.64 – 6.69 6.7 5.6 7.1 6.9 5.75 7.4 7.1 7.1 5 6.5

MS 0.62 0.82 0.31 0.82 1.22 0.72 0.03 0.20 0.05 0.31 0.59 0.24 0.66 0.49 0.21 0.85 0.28 0.48 0.33 0.36

PGA (g)

31.57 81.30 29.80 62.10 112.49 71.56 1.33 9.76 3.90 22.70 99.10 17.89 89.72 64.28 32.90 121.40 35.60 39.70 8.55 52.70

PGV (cm/sec)

19.30 9.91 10.32 12.99 10.69 9.83 21.17 20.00 12.77 13.34 5.86 13.32 7.24 7.56 6.41 6.89 7.69 11.95 37.29 6.66

a/v (sec)

M e t a m o d e l-b a s e d c o m p u t a t i o n a l t e c h n i q u e s i n p r o b a b i l i s t i c o p t i m i z a t i o n

5

5

5 2

1

4

4

2

1

2 3

3 2

1

5 2

2 

4

1 4

2

1 3

3 2

587

2

1

Figure 21.4 RBDO Test example – Geometry, member grouping.

6.2.3 Three-storey plane frame under earthquake loads RBDO example One test example has been considered in the present study in order to illustrate the efficiency of the proposed methodology for reliability-based sizing optimization problems under earthquake loading. This test example is a four-bay, three-storey moment resisting plane frame shown in Figure 21.4. The frame has been previously studied by Gupta and Krawinkler (2000), where a detailed description of the structure is given. The frame consists of rigid moment connections and fixed supports. Each bay has a span of 9.15 m (30 ft), while each storey is 3.96 m (13ft) high. The permanent action considered is equal to 5 kN/m2 while the variable action is equal to 2 kN/m2 , both distributed along the beams. The frame is considered to be part of a 3D structure where each frame is 4.5 m (15ft) apart. The median spectrum used for the determination of the base shear corresponds to a peak ground acceleration of 0.32 g. Structural members are divided into five groups, as shown in Figure 21.4, corresponding to the five design variables of a discrete structural optimization problem. The cross-sections are W-shape beam and column sections available from manuals of the American Institute of Steel Construction (AISC). The objective function is the weight of the structure, to be minimized. In this study a suite of twenty natural accelerograms, shown in Table 21.3, is used. It can be seen that each record corresponds to different earthquake magnitudes and soil conditions. The records of this suite comprise a wide range of PGA and peak acceleration over peak displacement ratio (a/v) values. The latter parameter is considered to describe the damage potential of the earthquake more reliably than PGA. The records are scaled to the same PGA and their response spectra that are subsequently derived are shown in Figure 21.5. It has been observed that the response spectra follow the lognormal distribution. Therefore the median spectrumx, ˆ also shown in Figure 21.5, and the standard deviation δ are calculated from the above suite of spectra using the following expressions:  n xˆ = exp

i=1

ln (Rdi (T)) n

(14)

588

Structural design optimization considering uncertainties

1.6

Spectral acceleration (g)

1.4 1.2 1 Median spectra 0.8 0.6 0.4 0.2 0

0.5

0

1

1.5

2 Period T (sec)

2.5

3

3.5

4

Figure 21.5 Natural record response spectra and their median.

Table 21.4 Characteristics of the random variables. Random variable

Probability density function

Mean value

Standard deviation

E σy Seismic load

N N Log-N

2.1 106 MPa 235 MPa Median Spectrum (Eq. 14)

0.10E 0.10σy δ (Eq. 15)

 n ! δ=

i=1

"2 1/ 2 ln (Rdi (T)) − ln (x) ˆ n−1

(15)

where Rdi (T) is the response spectrum value for period equal to T of the i-th record (i = 1, . . . , n, where n = 20 in this study). For a given period value, the acceleration Rd is obtained as a random variable following the log-normal distribution with its mean value equal to xˆ and the standard deviation equal to δ. The deterministic constraints are related to stress and displacement constraints for steel frames according to Eurocodes. The probabilistic constraint is imposed on the probability of structural collapse which is set equal to pall = 0.001. The probability of failure caused by uncertainties related to seismic loads and material properties of the structure is estimated using MCS with the LHS technique. The earthquake ground motion parameter, as described in Eq. (14), the yield stress and the elastic modulus are considered to be random variables. The type of probability density functions, mean values, and variances of the random parameters are shown in Table 21.4. The seismic action follows a log-normal probability density function, while the rest of

M e t a m o d e l-b a s e d c o m p u t a t i o n a l t e c h n i q u e s i n p r o b a b i l i s t i c o p t i m i z a t i o n

589

Table 21.5 Performance of the methods. Optimization procedure

ES cycles

pf

Time sequential (h)

Time parallel (p = 5) (h)

DBO RBDO-MCS (5,000 siml.) RBDO-LHS (1,000 siml.) RBDO-NN (100,000 siml.)

157 65 72 68

0.0932 0.0008 0.001 0.0009

0.3 557.3 149.1 42.1

0.08 140.1 40.6 16.2

the random variables follow a normal probability density function. For more details on probabilistic formulations of uncertainties the reader is referred to JCSS (2001) guidelines. For this test case the (µ + λ)-ES approach is used with µ = λ = 5 (equal to the number of design variables), while a sample size of 5000 simulations is taken. Table 21.5 depicts the performance of the optimization procedure for this test case. As it can be seen, the probability of failure corresponding to the optimum computed by the deterministic optimization procedure is much larger than the specified value of 10−3 , thus unacceptable. On the other hand, the increase in safety results also in a significant increase on optimum weight. When probabilistic constraints are considered the weight increase is approximately 26% compared to the deterministic one, from 125.3 to 167.4 tn. Furthermore, the computation times are also enlarged in the case of RBDO, however, the use of NN as well as parallel computation reduces drastically the excessive computational cost of the process. As far as the NN implementation is concerned, it was performed in a similar manner as the ES-NN1 algorithm that was described in the previous section. The NN configuration used has the typical architecture shown in Figure 21.1. It consists of three layers: one input, one hidden, and one output layer with varying number of nodes per layer. After an initial investigation on the optimum number of hidden layers and their nodes, one hidden layer was used having 10 nodes. The input data of the NN are the eleven random variables (two for each of the five element groups plus the seismic coefficient), while the output is one, i.e. the maximum interstorey drift value, which defines the limit-state violation. Thus, the NN configuration that was used was the following: 16-20-1. The training-testing set of the NN consisted of two hundred input/output pairs, twenty of which were used for testing the generalization capabilities of the trained NN. The application of NN reduces the computing time in a fraction of the time required for the conventional FE analyses. In addition, it does not affect the accuracy of the MCS method, in fact it can increase it since the fast NN approximations allow the use of much greater sampling size.

6.3

Hybrid RRDO 3D trus s tes t exampl e

For the purposes of this study a 3D steel truss structure has also been considered. For this test example, two objective functions have been taken into account, the initial construction cost and the standard deviation of a characteristic node displacement

590

Structural design optimization considering uncertainties

representing the response of the structure. Two sets of constraints are enforced, deterministic constraints on stresses, element buckling and displacements imposed by the European design codes and probabilistic ones. Furthermore, due to manufacturing limitations the design variables are not continuous but discrete since cross-sections belong to a certain predefined set provided by the manufacturers. The discrete design variables are treated in the same way as in single optimum design problems using the discrete evolution strategies. The design variables considered are the dimensions of the members of the structure taken from the Circular Hollow Section (CHS) table of the Eurocode. The random variables related to the cross-sectional dimensions, for both test examples, are two per design variable: the external diameter D and the thickness t of the circular hollow section. Apart from the cross-sectional dimensions of the structural members, the material properties (modulus of elasticity E and yield stress σy ) and the lateral loads have also been considered as random variables. The robustness of the constraints is also considered using the overall probability of maximum violation of the behavioural constraints, as a result of the variation of the uncertain structural parameters. The test example considered is the 3D truss tower shown in Figures 6(a) to 6(c). The height of the truss tower is 128 m, while its basis is a rectangle of side 17.07 m. The FE model consists of 324 nodes and 1254 elements which are divided into 12 groups that play the role of the design variables. The applied loading consists of: (i) self weight (dead load), (ii) live loads and (iii) wind actions according to the (Eurocode 1 2003). The type of probability density function, the mean value, and the variance of the random variables are shown in Table 21.6. In the present implementation an investigation is performed on the ability of the NN to predict the required data for the evolution of the RRDO process. The inputs of the NN correspond to the random variables, while the outputs are the characteristic node displacement and the maximum displacement, stress and compression force required for the calculation of the probability of violation. The appropriate selection of I/O training data constitutes the most important parts in the NN training. The number of training patterns may not be the only concern, as the distribution of samples is of great importance also. Having chosen the NN architecture and trained the neural network, the probability of violation and the standard deviation of the response can be obtained in orders of magnitude less computing time. The modulus of elasticity, yield stress, diameter D and thickness t of the circular hollow cross-section as well as the loading have been considered as random variables for the structures examined. The inputs of the NN correspond to the random variables, while the outputs are the characteristic node displacement and the maximum displacement, stress and compression force required for the calculation of the probability of violation. The previously described multi-criteria optimization (CEATm) algorithm employed is denoted as CEATm(µ + λ)nruns,csteps where µ, λ are the number of the parent and offspring vectors used in the ES optimization strategy, nruns is the number of independent CEA runs and csteps is the number of cascade stages employed. The basic steps inside an independent run of the multi-objective algorithm when the NN is embedded in the optimization process, as adopted in this test case, are described in Flowchart 21.5. For the solution of the multi-objective optimization problem in question the non-dominant CEATm(µ + λ)nrun,csteps optimization scheme was employed, where µ = λ = 5, nrun = 10 and csteps = 3. The resultant Pareto front curves for the RDO

M e t a m o d e l-b a s e d c o m p u t a t i o n a l t e c h n i q u e s i n p r o b a b i l i s t i c o p t i m i z a t i o n

591

Independent run do i = 1, nrun CEATm LOOP 1. Initial generation: do while sk not feasible k = 1, µ Generate sk (k=1,… , µ) vectors Analysis step Evaluation of the Tchebycheff metric Deterministic constraints check: if satisfied continue else regenerate sk design Monte Carlo Simulation step: Selection of the NN training set NN training for the limit load NN Monte Carlo Simulations Probabilistic constraint check end do 2. Global non-dominant search: Check if global generation is accomplished. If yes, then non-dominant search is performed, else wait until global generation is accomplished. 3. New generation: do while s not feasible  = 1, λ Generate s (k = 1, . . . , µ) vectors Analysis step Evaluation of the Tchebycheff metric Deterministic constraints check: if satisfied continue else regenerate sk design Monte Carlo Simulation step: Selection of the NN training set NN training for the limit load NN Monte Carlo Simulations Probabilistic constraint check end do 4. Selection step: selection of the next generation parents according to (µ + λ) or (µ, λ) scheme 5. Global non-dominant search: Check if global generation is accomplished. If yes, then non-dominant search is performed, else wait until global generation is accomplished. 6. Convergence check: If satisfied stop, else go to step 3 END OF CEATm LOOP End do of Independent run Flowchart 21.5 The CEATm algorithm combined with NN. Table 21.6 3D truss tower example: Characteristics of the random variables. Random variable

Description

Probability Mean density value (µ) function

E (kN/m2 ) σy (kN/m2 ) F (kN) D t

Young’s Modulus Allowable stress Horizontal loading CHS Diameter CHS Thickness

Normal Normal Normal Normal Normal

2.10E + 08 355000 Fµ d ∗i t ∗i

Standard σ/µ deviation (σ)

95% of Values interval

1.50E + 07 35500 0.4 Fµ 0.02 d i 0.02 t i

(1.81E + 08, 2.39E + 08) (2.85E + 05, 4.25E + 05) (2.16 Fµ , 17.84 Fµ ) (0.9618 d i , 1.039 d i ) (0.9618 t i , 1.039 t i )

7.14% 10.00% 40.00% 2% 2%

* Taken from the Circular Hollow Section (CHS) table of the Eurocode, for every design.

592

Structural design optimization considering uncertainties

29.26

4.27

40.23

x y

58.52

128.01

Characteristic node

17.07 19.21 (a)

(b)

(c)

Figure 21.6 3D truss tower example: (a) 3D view, (b) Side view, (c) Top view.

formulations are depicted in Figure 21.7, with the structural weight on the horizontal axis and the standard deviation of the characteristic node displacement on the vertical axis. The displacement in the x-direction of the top node is selected as the characteristic one (Figure 21.6c). As can be seen in Figure 21.7 the trend on the influence of the probabilistic constraint is similar to that of the first example, where the Pareto front curves coincide in different parts. Four different formulations of the RDO problem have been considered in this study: (i) the standard RDO formulation, (ii) RRDO with allowable probability equal to 2% denoted as RRDO_2%, (iii) RRDO with allowable probability equal to 0.1% denoted as RRDO_0.1% and (iv) RRDO with allowable probability equal to 0.01% denoted as RRDO_0.01%. As can be seen in Figure 21.7 the presence of the probabilistic constraint influences the Pareto curves near the DBO area (designs Ai, i = 1, . . . , 4) of the Pareto front, where the weight of the structure is the dominant criterion. On the contrary, the four Pareto front curves almost coincide at the areas

M e t a m o d e l-b a s e d c o m p u t a t i o n a l t e c h n i q u e s i n p r o b a b i l i s t i c o p t i m i z a t i o n

593

5.5E-02

Standard deviation (m)

5.0E-02

A4 (4497.3, 5.1E-02) A2 (4231.2, 5.0E-02) A1 (3874.7, 4.9E-02) A3 (4314.6, 4.9E-02)

RDO RRDO 2% RRDO 0.1% RRDO 0.01%

4.5E-02

4.0E-02

3.5E-02 B1 (4997.3, 3.0E-02) B3 (4982.8, 2.9E-02) B2 (4964.7, 3.0E-02) B4 (4983.7, 2.9E-02)

3.0E-02

2.5E-02 3750.0

4000.0

4250.0

4500.0 Weight (kN)

4750.0

5000.0

5250.0

Figure 21.7 3D truss tower example: Comparison of the Pareto front curves.

where the importance of the second criterion (standard deviation of the response) increases. The performance of the NN prediction is depicted in Figures 21.8a to 21.8d, where the prediction of the characteristic displacement, the maximum displacement, the maximum compressive force and the maximum tensile force are shown, respectively. Three different training sets, of size 100, 200 and 500, respectively have been examined, randomly generated using LHS, while 50 patterns have been used for testing. As can be seen in Figure 21.8, 100 samples are enough for efficiently training the NN. The MCS sample sizes used in this test example are 10,000, 100,000 and 500,000. In the RDO and RRDO_2% formulations a sample size of 10,000 simulations has been used, while in RRDO_0.1% and RRDO_0.01% a sample size of 100,000 and 500,000 simulations has been employed. The different formulations and consequently the different sample sizes lead to a significantly different computing cost. In order to reduce the increased computing cost, especially of the last two formulations, a neural network formulation has been applied. The NN configuration implemented in this example has one hidden layer with 50 nodes resulting in a 27-50-4 NN architecture (see Figure 21.1), which is used for all runs. The computing cost is depicted in Table 21.7, where the conventional and the corresponding NN computing times are reported. It has to be mentioned that the denoted basic computing costs for the RRDO_0.1% and RRDO_0.01% formulations are estimations due to the excessive computing cost required for these two cases. It can be seen that the NN-based methodology requires up to four orders of magnitude less computing time compared to the conventional one.

594

Structural design optimization considering uncertainties Table 21.7 3D truss tower example: Computing times. Formulation

No of simulations

RDO RRDO 2% RRDO 0.01% RRDO 0.001%

Time (hours)

10,000 10,000 100,000 500,000

Basic

NN

5.33E+01 5.42E+01 5.59E+02* 2.79E+03*

5.87E−01 5.96E−01 6.15E−01 6.14E−01

* Estimated.

6.0E-01

6.0E-01

100 200 500

4.0E-01

5.0E-01

100 200 500

Real

Real

5.0E-01

4.0E-01 3.0E-01 2.0E-01 2.0E-01

3.0E-01

4.0E-01 Predicted (a)

5.0E-01

3.0E-01 3.0E-01

6.0E-01

100 200 500

50.0 40.0

Real

180.0

60.0 Real

6.0E-01

200.0

70.0

30.0 30.0

4.0E-01 5.0E-01 Predicted (b)

100 200 500

160.0 140.0 120.0

40.0

50.0 Predicted (c)

60.0

70.0

100.0 100.0

120.0

140.0 160.0 Predicted (d)

180.0

200.0

Figure 21.8 3D truss tower: Performance of NN with respect to the number of the training patterns (a) characteristic displacement, (b) maximum displacement, (c) maximum compressive force, and (d) maximum tensile force.

7 Conclusions In most cases the optimum design of structures is based on nominal values of the design parameters and is focused on the satisfaction of the deterministically defined design code provisions. The deterministic optimum is not always a “safe’’ design, since there are many random factors that affect the design, i.e. manufacturing and performance of a structure during its lifetime. In order to find a “real’’ optimum the designer has to take into account all necessary random variables. In order to alleviate this deficiency, two types of formulations have been proposed in the past: RBDO and RDO. In the present work, apart from presenting successful RBDO applications, the combined RRDO is

M e t a m o d e l-b a s e d c o m p u t a t i o n a l t e c h n i q u e s i n p r o b a b i l i s t i c o p t i m i z a t i o n

595

also proposed, where probabilistic constraints are incorporated into the robust design optimization formulation. In the examined RBDO formulations, under static or dynamic loads, the reliability analysis of the structure has to be performed in order to determine its optimum design taking into account a desired level of probability of structural failure. Only after forming and solving this RBDO problem, even with additional cost in weight and computing time, can a “global’’ and realistic optimum structural design be found. The aim of the proposed RBDO procedure is to increase the safety margins of the optimized structures under various uncertainties, while at the same time minimizing its weight, and reducing substantially the required computational effort. The solution of realistic RBDO problems in structural mechanics is an extremely computationally intensive task. As it can be observed from the numerical results, the computational cost for the solution of realistic RBDO problems is orders of magnitude larger than the corresponding cost for a deterministic optimization problem. Due to the size and complexity of RBDO problems, a non-conventional, stochastic evolutionary optimization method – such as ES – appears to be a suitable choice. In a similar manner, in order to implement the hybrid RRDO formulation, structural reliability analyses for every candidate design have to be performed for the evaluation of the probability of violation. Depending on the value of the allowable probability of violation, different sample sizes are employed in order to calculate with sufficient accuracy the statistical quantities under consideration i.e. the standard deviation of the response and the probability of violation of the constraints. The Pareto front curves obtained for the presented RRDO formulation and the RDO formulation appear to be different when the weight objective function is predominant, while they approach each other in the areas of the Pareto fronts where the significance of the standard deviation of the response criterion increases. In other words, for the same standard deviation value, the optimum weight achieved by the RRDO formulation are larger than the corresponding weight achieved by the conventional RDO approach. Furthermore, it was observed that the presence of the standard deviation as an objective function forces the RDO formulation to produce results very close to those obtained by the RRDO formulation close to the right end of the Pareto front curve. Concluding, the aim of this work was twofold: to examine the influence of the probabilistic parameters and constraints in structural optimization, and to deal with computationally demanding tasks in probabilistic mechanics. The computational effort involved in the conventional MCS becomes excessive in large-scale problems, especially when earthquake loading is considered, due to the enormous sample size and the computing time required for each Monte Carlo run. Although the LHS technique has been implemented for improving the computational efficiency of the MCS method, the computational cost remains excessive, making the solution of large-scale probabilistic optimization problems computationally unsolvable. Thus, a neural network assisted methodology has been proposed in order to obtain the structural response results required during the Monte Carlo simulations inexpensively. The achieved reduction in computational time was several orders of magnitude compared to the conventional procedure making tractable the optimization of real world structures under probabilistic constraints. The use of NN can practically eliminate any limitation on the scale of the problem and the sample size used for MCS without deteriorating the accuracy of the results.

596

Structural design optimization considering uncertainties

References Agarwal, H. & Renaud, J.E. 2006. A new decoupled framework for reliability based design optimization. AIAA J. 44(7):1524–1531. Agarwal, H. & Renaud, J.E. 2004. Reliability based design optimization using response surfaces in application to multidisciplinary systems. Eng. Optim. J. 36(3):291–311. Ba-abbad, M.A., Nikolaidis, E. & Kapania, R.K. 2006. New approach for system reliabilitybased design optimization. AIAA J. 44(5):1087–1096. Bozorgnia, Y. & Bertero, V.V. (eds) 2004. Earthquake engineering: from engineering seismology to performance-based engineering. CRC Press Publications. Coello-Coello, C.A. 2000. An updated survey of GA-based multi-objective optimization techniques. ACM Computing Surveys 32(2):109–143. Doltsinis, I. & Kang, Z. 2004. Robust design of structures using optimization methods. Comput. Methods Appl. Mech. Engrg. 193:2221–2237. EC1. 2003. Eurocode 1 : Basis of design and actions on structures – Part 2–4: Actions on Structures – Wind actions. CEN, ENV 1991-2-4, European Committee for Standardization, Brussels. EC3. 2003. Eurocode 3: Design of steel structures, Part 1.1: General rules for buildings. CEN, ENV 1993-1-1/1992, European Committee for Standardization, Brussels. EC8. 2004. Eurocode 8: Design of structures for earthquake resistance – Part 1. European standard. CEN-ENV-1998-1, European Committee for Standardization, Brussels. Frangopol, D.M. & Soares, C.G. 2001. Reliability-oriented optimal structural design. J. Reliab. Eng. Syst. Saf. 73(3):195–3069. Gupta, A. & Krawinkler, H. 2000. Behaviour of ductile SMRFs at various seismic hazard levels. ASCE Journal of Structural Engineering 126(1):98–107. JCSS. 2001. Probabilistic Model Code. Joint Committee on Structural Safety, http://www.jcss. ethz.ch [accessed 15 September 2006]. Jung, D.H. & Lee, B.C. 2002. Development of a simple and efficient method for robust optimization. Int. J. Numer. Meth. Engng. 53:2201–2215. Lagaros, N.D. & Papadrakakis, M. 2006. Robust seismic design optimization of steel structures, J. Struct. Multidisc. Optim. Available on-line, Doi: 10.1007/s00158-006-0047-5. Lagaros, N.D. & Tsompanakis, Y. (eds) 2006. Intelligent computational paradigms in earthquake engineering. Idea Publishers. Lagaros, N.D., Papadrakakis, M. & Kokossalakis, G. 2002. Structural optimization using evolutionary algorithms. Computer & Structures 80(7–8):571–587. Lagaros, N.D., Plevris, V. & Papadrakakis, M. 2005. Multi-objective design optimization using cascade evolutionary computations. Comput. Methods Appl. Mech. Engrg. 194(30–33):3496–3515. Marler, R.T. & Arora, J.S. 2004. Survey of Multi-objective Optimization Methods for Engineering. J. Struct. Multidisc. Optim. 26(6):369–395. Mattson, C.A., Mullur, A.A. & Messac, A. 2004. Smart pareto filter: Obtaining a minimal representation of multiobjective design space. Eng. Optim. J. 36(6):721–740. McKay, M.D., Beckman, R.J. & Conover, W.J. 1979. A comparison of three methods for selecting values of input variables in the analysis of output from a computer code. Technometrics 21(2):239–245. Messac, A. & Ismail-Yahaya, A. 2002. Multiobjective robust design using physical programming. J. Struct. Multidisc. Optim. 23:357–371. Owen, A.B. 1997. Monte Carlo variance of scrambled net quadrature. SIAM J. Num. Anal. 34(5):1884–1910. Papadrakakis, M., Tsompanakis, Y. & Lagaros, N.D. 1999. Structural shape optimization using evolution strategies. Eng. Optim. J. 31:515–540.

M e t a m o d e l-b a s e d c o m p u t a t i o n a l t e c h n i q u e s i n p r o b a b i l i s t i c o p t i m i z a t i o n

597

Park, G.-J., Lee, T.-H., Lee, K.H. & Hwang, K.-H. 2006. Robust design: An overview. AIAA J. 44(1):181–191. Rummelhart, D.E. & McClelland, J.L. 1986. Parallel Distributed Processing, Volume 1: Foundations. The MIT Press, Cambridge. Schuëller, G.I. 2005. Special Issue on Computational Methods in Stochastic Mechanics and Reliability Analysis. Comput. Methods Appl. Mech. Engrg. 194(12–16):1251–1795. Schwefel, H.P. 1981. Numerical optimization for computer models. Wiley & Sons, Chichester, UK. Tsompanakis, Y. & Papadrakakis, M. 2004. Large-scale reliability-based structural optimization. J. Struct. Multidisc. Optim. 26:1–12. Tsompanakis, Y., Lagaros, N.D. & Stavroulakis, G. 2007. Soft computing techniques in parameter identification and probabilistic seismic analyses of structures, J. Advances in Eng. Software (in press). Wen, Y.K. 2000. Reliability and performance-based design. Proceedings of the 8th ASCE Specialty Conference on Probabilistic Mechanics and Structural Reliability, University of Notre Dame, Indiana, USA, 24–26 July. Youn, B.D. & Choi, K.K. 2004. The performance moment integration method for reliabilitybased robust design optimization. Proceedings of the ASME Design Engineering Technical Conference, Salt Lake City, Utah, USA, September 28–October 2. Youn, B.D., Choi, K.K. & Du, L. 2005. Enriched performance measure approach for reliabilitybased design optimization. AIAA J. 43(4):874–884.

References

AASHTO, American Association of State Highway and Transportation Officials 1992. Standard specifications for highway bridges. Washington, D.C.: American Association of State Highway and Transportation Officials. 15th edition. Abramowitz, M. & Stegun, I.A. 1972. Handbook of mathematical functions. 10th ed. New York: Dover. Adams, B.M., Eldred, M.S. & Wittwer, J.W. 2006. Reliability-based design optimization for shape design of compliant micro-electro-mechanical systems. In Proceedings of the 11th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, Number AIAA-2006-7000, September 6–8, Portsmouth, VA. Adams, B.M., Bichon, B.J., Eldred, M.S., Carnes, B., Copps, K.D., Neckels, D.C., Hopkins, M.M., Notz, P.K., Subia, S.R. & Wittwer, J.W. 2006. Solution-verified reliability analysis and design of bistable mems using error estimation and adaptivity. Technical Report SAND20066286, Sandia National Laboratories, 2006, October, Albuquerque, NM. Adelman, H. & Haftka, R. 1986. Sensitivity analysis of discrete structural systems. AIAA Journal 24:823–832. Agarwal, H. & Renaud, J.E. 2004. Reliability based design optimization using response surfaces in application to multidisciplinary systems. Eng. Optim. J. 36(3):291–311. Agarwal, H. & Renaud, J. 2006. A new decoupled framework for reliability based design optimization. AIAA Journal 44(7):1524–1531. Agarwal, H., Renaud, J.E., Lee, J.C. & Watson, L.T. 2004. A unilevel method for reliabilitybased design optimization. In Proceedings of the 45th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, Number AIAA-2004-2029, April 19–22, Palm Springs, CA. Agarwal, H., Renaud, J.E., Preston, E.L. & Padmanabhan, D. 2004. Uncertainty Quantification Using Evidence Theory in Multidisciplinary Design Optimization. Reliability Engineering and System Safety 85:281–294. Agmon, N., Alhassid, Y. & Levine, R.D. 1979. An algorithm for finding the distribution of maximal entropy. Journal of Computational Physics 30:250–258. Aitchison, J. & Dunsmore, I.R. 1975. Statistical Prediction Analysis. Cambridge University Press, Cambridge. Akgul, F. & Frangopol, D.M. 2003. Probabilistic analysis of bridge networks based on system reliability and Monte Carlo simulation. In Applications of Statistics and Probability in Civil Engineering. A. Der Kiureghian, S. Madanat, & J.M. Pestana (eds), Rotterdam, The Netherlands, pp. 1633–1637. Millpress. Akiyama, H. 1985. Earthquake Resistant Limit-State Design for Buildings. University of Tokyo Press, Tokyo, Japan. Akpan, U.O., Rushton, P.A. & Koko, T.S. 2002. Fuzzy Probabilistic Assessment of the Impact of Corrosion on Fatigue of Aircraft Structures, Paper AIAA-2002-1640. Allen, J.J. 2005. Micro Electro Mechanical System Design. Boca Raton: Taylor and Francis.

600

References

Allen, M. & Maute, K. 2004. Reliability-based design optimization of aeroelastic structures. Structural and Multidisciplinary Optimization 27(4):228–242. Allen, M., Raulli, M., Maute, K. & Frangopol, D. 2004. Reliability-based analysis and design optimization of electrostatically actuated MEMS. Computers and Structures 82(13–14): 1007–1020. Ambartzumian, R., Der Kiureghian, A., Ohaniana, V. & Sukiasiana, H. 1998. Multinormal probability by sequential conditioned importance sampling: Theory and application. Probabilistic Engineering Mechanics 13(4):299–308. Ananthasuresh, G.K., Kota, S. & Gianchandani, Y. 1994. A methodical approach to the design of compliant micromechanisms. In Proc. IEEE Solid-State Sensor and Actuator Workshop, Hilton Head Island, SC, pp. 189–192. Ang, A.H.-S. & De Leon, D. 1997. Determination of optimal target reliabilities for design and upgrading of structures. Structural Safety 19:91–103. Ang, H.-S.A. & Lee, J.-C. 2001. Cost optimal design of R/C buildings. Reliability Engineering and System Safety 73:233–238. Ang, H.-S. & Tang, W.H. 1975. Probabilistic concepts in engineering planning and design, Vol. I and II, Wiley. Ang, A.H.-S. & Tang, W.H. 1984. Probability Concepts in Engineering Planning and Design, Volume II; Decision, Risk, and Reliability, New York, John Wiley & Sons. Aoues, Y. & Chateauneuf, A. 2007. Reliability-based optimization of structural systems by adaptive target safety application to RC frames. Structural Safety. Article in Press. Arias, A. 1970. A measure of earthquake intensity. In Seismic Design for Nuclear Power Plants, R.J. Hansen (ed.), The MIT Press, Cambridge, MA, 438–469. Atkinson, G.M. & Silva, W. 2000. Stochastic modeling of California ground motions. Bulletin of the Seismological Society of America 90(2):255–274. Attoh-Okine, N.O. 2002. Uncertainty analysis in structural number determination in flexible pavement design – A convex model approach. Construction and Building Materials 16(2):67–71. Au, S.K. 2005. Reliability-based design sensitivity by efficient simulation. Computers and Structures 83:1048–1061. Au, F.T.K., Cheng, Y.S., Tham, L.G. & Zeng, G.W. 2003. Robust design of structures using convex models. Computers and Structures 81(28–29):2611–2619. Au, S.K. & Beck, J.L. 1999. A new adaptive importance sampling scheme. Structural Safety 21:135–158. Au, S.-K. & Beck, J.L. 2001. First excursion probabilities for linear systems by very efficient importance sampling. Probabilistic Engineering Mechanics 16(3):193–207. Au, S.K. & Beck, J.L. 2001. Estimation of small failure probabilities in high dimensions by subset simulation. Probabilistic Engineering Mechanics 16(4):263–277. Au, S.K. & Beck, J.L. 2003. Importance sampling in high dimensions. Structural Safety 25(2):139–163. Au, S.K. & Beck, J.L. 2003. Subset simulation and its applications to seismic risk based on dynamic analysis. Journal of Engineering Mechanics 129(8):901–917. Augusti, G., Ciampoli, M. & Frangopol, D.M. 1998. Optimal planning of retrofitting interventions on bridges in a highway network. Engineering Structures 20(11):933–939. Elsevier. Austrell, P.-E., Dahblom, O., Lindemann, J., Olsson, A., Olsson, K.-G., Persson, K., Petersson, H., Ristinmaa, M., Sandberg, G. & Wernbergk, P.-A. 1999. Structural Mechanics, LTH, Sweden. CALFEM: A finite element toolbox to MATLAB, Version 3.3. http://www. byggmek.lth.se/Calfem/index.htm.

References

601

Ayhan, H., Lim’on-Robles, J. & Wortman, M.A. 1999. An approach for computing tight numerical bounds on renewal functions. IEEE Transactions on Reliability 48(2):182–188. Ba-abbad, M.A., Nikolaidis, E. & Kapania, R.K. 2006. New approach for system reliabilitybased design optimization. AIAA J. 44(5):1087–1096. Bae, H.-R., Grandhi, R.V. & Canfield, R.A. 2004. An Approximation Approach for Uncertainty Quantification Using Evidence Theory. Reliability Engineering and System Safety 86:215–225. Bae, H.-R., Grandhi, R.V. & Canfield, R.A. 2004. Epistemic Uncertainty Quantification Techniques Including Evidence Theory for Large-Scale Structures. Computers and Structures 82:1101–1112. Barlow, R.E. & Proschan, F. 1965. Mathematical Theory of Reliability. New York: John Wiley & Sons. Barlow, R.E. & Proschan, F. 1975. Statistical Theory of Reliability and Life Testing: Probabilistic Models. New York: Holt, Rinehart and Winston. Barndorff-Nielsen, O. & Cox, D.R. 1979. Edgeworth and saddle-point approximations with statistical applications. Journal of the Royal Statistical Society 41:279–312. Beck, J.L., Chan, E., Irfanoglu, A. & Papadimitriou, C. 1999. Multi-criteria optimal structural design under uncertainty. Earthquake Engineering and Structural Dynamics 28(7):741–761. Beck, J.L. & Katafygiotis, L.S. 1998. Updating models and their uncertainties. I: Bayesian statistical framework. Journal of Engineering Mechanics 124(4):455–461. Bendsoe, M.P. 1989. Optimal shape design as a material distribution problem. Comp. Mth. Appl. Mech. Engrg. 1:193–202. Bendsoe, M.P. & Kikuchi, N. 1988. Generating optimal topologies in optimal design using a homogenization method. Comp. Mth. Appl. Mech. Engrg. 71:197–224. Bendsøe, M.P. & Sigmund, O. 1989. Topology Optimization: Theory, Methods and a applications. Springer-Verlag, Berlin. Ben-Haim, Y. 1985. The Assay of Spatially Random Material. Dordrecht: D. Reidel Publishing Company. Ben-Haim, Y. 1996. Robust Reliability in the Mechanical Sciences. Berlin: Springer Verlag. Ben-Haim, Y. 2001. Information-gap decision theory: Decisions under severe uncertainty. New York: Academic Press, Inc. Ben-Haim, Y. 2004. Uncertainty, probability and information-gaps. Reliability Engineering and System Safety 85:249–266. Ben-Haim, Y. 2005. Value at risk with Info-gap uncertainty. Journal of Risk Finance 6:388–403. Ben-Haim, Y. 2005. Info-gap decision theory for engineering design. In Engineering Design Reliability Handbook, E. Nikolaidis, D. Ghiocel & S. Singhal (eds), CRC Press. Ben-Haim, Y. 2006. Information-gap Decision Theory: Decisions Under Severe Uncertainty. 2nd edition. London: Academic Press. Ben-Haim, Y. & Elishakoff, I. 1990. Convex Models of Uncertainty in Applied Mechanics. New York: Elsevier. Ben-Haim, Y. & Elishakoff, I. 1996. Convex models of uncertainty in applied mechanics. New York: Elsevier. Ben-Haim, Y., Chen, G. & Soong, T.T. 1996. Maximum structural response using convex models. ASCE J. Engineering Mechanics 122:325–333. Benjamin, J.R. & Cornell, C.A. 1970. Probability, Statistics and Decision for Civil Engineers. Mc-Graw-Hill. Bennet, J. & Botkin, M. 1985. Structural shape optimization with geometric description and adaptive mesh refinement. AIAA Journal 23:458–464. Ben-Tal, A. & Nemirovski, A. 2001. Lectures on Modern Convex Optimization: Analysis, Algorithms, and Engineering Applications. Philadelphia: SIAM.

602

References

Ben-Tal, A. & Nemirovski, A. 2002. Robust optimization – methodology and applications. Mathematical Programming B92:453–480. Berens, A.P., Hovey, P.W. & Skinn, D.A. 1991. Risk Analysis for Aging Aircraft Fleets, Air Force Wright Lab Report, WL-TR-91-3066, Vol. 1. Bichon, B.J., Eldred, M.S., Swiler, L.P., Mahadevan, S. & McFarland, J.M. 2007. Multimodal reliability assessment for complex engineering applications using efficient global optimization. In Proceedings of the 9th AIAA Non-Deterministic Approaches Conference, Number AIAA-2007-1946, April 23–26, Honolulu, HI. Bjerager, P. 1988. Probability integration by directional simulation. Journal of Engineering Mechanics 114(8):1288–1302. Bjerager, P. 1990. On computational methods for structural reliability analysis. Structural Safety 9(2):79–96. Bjerager, P. 1991. Methods for structural reliability computation. In Reliability problems: general principles and applications in mechanics of solid and structures. F. Casciati (ed.), New York: Springer-Verlag. Boore, D.M. & Joyner, W.B. 1997. Site amplifications for generic rock sites. Bulletin of the Seismological Society of America 87:327–341. Boore, D.M. 2003. Simulation of ground motion using the stochastic method. Pure applied Geophysics 160:635–676. Box, G.E.P. & Cox, D.R. 1964. An analysis of transformations. J. Royal Stat. Soc. 26:211–252. Boyd, S. & Vandenberghe, L. 2004. Convex Optimization. Cambridge: Cambridge University Press. Bozorgnia, Y. & Bertero, V.V. (eds) 2004. Earthquake engineering: from engineering seismology to performance-based engineering. CRC Press Publications. Breipohl, A.M. 1970. Probabilistic Systems Analysis, Wiley, New York. Breitung, K. 1984. Asymptotic approximation for multi-normal integrals. ASCE Journal of Engineering Mechanics 10(3):357–366. Brill, E.D. 1979. The use of optimization models in public-sector planning. Mangement Science 25(5):413–422. Brown, M. & Proschan, F. 1983. Imperfect repair. Journal of Applied Probability 20:851–859. Bucher, C.G. 1988. Adaptive Sampling – An Iterative Fast Monte Carlo Procedure, Structural Safety 5:119–126. Bucher, C.G. & Bourgund, U. 1990. A fast and efficient response surface approach for structural reliability problems. Structural Safety 7:57–66. Bucher, C. & Frangopol, D.M. 2006. Optimization of lifetime maintenance strategies for deteriorating structures considering probabilities of violating safety, condition, and cost thresholds. Probabilistic Engineering Mechanics 21(1):1–8. Elsevier. Burke, J.V. 1991. Calmness and exact penalization. SIAM J. Control and Optimization 29(2):493–497. Burks, A. 1970. Essays on Cellular Automata, Chapter Von Neumann’s self-reproducing automata, pp. 3–64. University of Illinois Press. Calafiore, G. & El Ghaoui, L. 2004. Ellipsoidal bounds for uncertain linear equations and dynamical systems. Automatica 40:773–787. Chan, K.Y., Kokkolaras, M., Papalambros, P.Y., Skerlos, S.J. & Mourelatos, Z. 2004. Propagation of uncertainty in optimal design of multilevel systems: Piston-ring/cylinder-liner case study. In Proceedings of the SAE World Congress, Detroit, Michigan, March 8–11. Paper No. 2004-01-1559. Chandu, S. & Grandhi, R. 1995. General purpose procedure for reliability based structural optimization under parametric uncertainties. Advances in Engineering Software 23:7–14. Chateauneuf, A. & Noret, E. 2005. System reliability-based optimization of redundant timber trusses. In Reliability and optimization of structural systems, J.D. Sørensen, (ed.), Proceedings

References

603

of the IFIP WG7.5 Working conference on reliability and optimization of structural systems, Aalborg, Denmark. Cheah, P.K., Fraser, D.A.S. & Reid, N. 1993. Some alternatives to edgeworth. Canadian Journal of Statistics 21:131–138. Chen, L. & Rao, S.S. 1997. Fuzzy finite element approach for the vibration analysis of imprecisely defined systems. Finite Elements in Analysis and Design 27:69–83. Chen, S., Lian, H. & Yang, X. 2002. Interval static displacement analysis for structures with interval parameters. International Journal for Numerical Methods in Engineering 53:393–407. Chen, X. & Lind, N.C. 1983. Fast probability integration by three-parameter normal tail approximation. Struct. Saf. 1:269–276. Chen, X., Hasselman, T.K. & Neill, D.J. 1997. Reliability based structural design optimization for practical applications. In Proceedings of the 38th AIAA/ASME/ASCE/AHS/ASC structures, structural dynamics and material conference, Kissimmee, Florida, AIAA-97-1403. Cheng, G., Xu, L. & Jiang, L. 2006. A sequential approximate programming strategy for reliability-based structural optimization. Computers and Structures. Article in Press. Cherng, R.H. & Wen, Y.K. 1994. Reliability of uncertain nonlinear trusses under random excitation, I., II. J. of Engineering Mechanics ASCE 120(4):733–757. Chernousko, F.L. 1999. What is ellipsoidal modelling and how to use it for control and state estimation? In Whys and Hows in Uncertainty Modelling, I. Elishakoff (ed.), pp. 127–188. Wien: Springer-Verlag. Ching, J. & Hsieh, Y.-H. 2007. Local estimation of failure probability function and its confidence interval with maximum entropy principle. Probabilistic Engineering Mechanics 22: 39–49. Ching, J. & Hsu, W.-C. 2006. Transforming reliability limit state constraints into deterministic limit state constraints. Structural Safety. Article in Press. Cho, M. & Rhee, S.Y. 2003. Layup optimization considering free-edge strength and bounded uncertainty of material properties. AIAA Journal 41(11):2274–2282. Cho, M., & Rhee, S.Y. 2004. Optimization of laminates with free edges under bounded uncertainty subject to extension, bending and twisting. International Journal of Solids and Structures 41(1):227–245. Choi, K.K. & Youn, B.D. 2002. On probabilistic approaches for reliability-based design optimization. In 9th AIAA/NASA/USA/ISSMO symposium on Multidisciplinary Analysis and Optimization, September 4–6, Atlanta, GA, USA. Choi, K.K., Youn, B.D. & Yang, R. 2001. Moving least square method for reliabilitybased design optimization. In Proceedings of the Fourth World Congress of Structural and Multidisciplinary Optimization (WCSMO-4), June 4–8, Dalian, China. Choi, K.K., Du, L. & Youn, B.D. 2004. A New Fuzzy Analysis Method for Possibility-Based Design Optimization. In 10th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, AIAA 2004-4585, Albany, NY. Christian, J.T. & Baecher, G.B. 1998. Point-estimate method and numerical quadrature. Journal of Geotechnical and Geoenvironmental Engineering 125:779–786. Clarke, F. 1983. Optimization and nonsmooth analysis. New York: John Wiley & Sons. Coello-Coello, C.A. 2000. An updated survey of GA-based multi-objective optimization techniques. ACM Computing Surveys 32(2):109–143. Cox, D.R. 1962. Renewal Theory. Monographs on Applied Probability and Statistics. London: Chapman & Hall. Cox, D.R. & Isham, V. 1980. Point Processes. Monographs on Applied Probability and Statistics. London: Chapman & Hall. Cramer, H. & Leadbetter, M.R. 1967. Stationary and Related Stochastic Processes. New York: John Wiley & Sons.

604

References

Creveling, C.M. 1997. Tolerance design: A handbook for developing optimal specification. Cambridge, MA: Addison-Wesley. Cunha, S.B., De Souza, A.P.F., Nicolleti, E.S.M. & Aguiar, L.D. 2006. A Risk-Based Inspection Methodology to Optimize Pipeline In-Line Inspection Programs. Journal of Pipeline Integrity, 5: Q2. Dailey, R. 1989. Eigenvector derivatives with repeated eigenvalues. AIAA Journal 27: 486–491. Daniels, H.E. 1954. Saddlepoint approximations in statistics. Annals of Mathematical Statistics 25:631–650. Deak, I. 1980. Three digit accurate multiple normal probabilities. Numerische Mathematik 35:369–380. Deqing, Y., Yunkang, S., Zhengxing, L. & Huanchun, S. 2000. Topology optimization design of continuum structures under stress and displacement constriaints. Applied Mathematic and Mechanics 21:1–26. Der Kiureghian, A. 1996. Structural reliability methods for seismic safety assessment: a review. Engineering Structures 95:412–24. Der Kiureghian, A. 2005. First- and Second-Order Reliability Methods, Chapter 14 In Engineering Design Reliability Handbook, E. Nikolaidis, D.M. Ghiocel & S. Singhal (eds), CRC Press, Boca Raton, FL. Der Kiureghian, A. & Dakessian, T. 1998. Multiple Design Points in First and Second-order Reliability. Structural Safety 20:37–50. Der Kiureghian, A. & Liu, P.L. 1986. Structural reliability under incomplete probability information. J. Eng. Mech. ASCE 112(1):85–104. Der Kiureghian, A., Lin, H.-Z. & Hwang, S.-J. 1987. Second-order reliability approximations. ASCE Journal of Engineering Mechanics 113(8):1208–1225. Der Kiureghian, A. & Polak, E. 1988. Reliability-based optimal design: a decoupled approach. In Reliability and optimization of structural systems. A.S. Nowak (ed.), Proceedings of the 8th IFIP WG7.5 Working Conference on reliability and optimization of structural systems, Chelsea, MI, USA: Book Crafters. pp. 197–205. Der Kiureghian, A. & Zhang, Y. 1999. Space-variant finite element reliability analysis. Comput. Methods Appl. Mech. Engrg. 168:173–183. D’Errico, J.R. & Zaino, N.A. 1988. Statistical tolerancing using a modification of Taguchi’s method. Technometrics 30(4):397–405. Diaz, A. & Kikuchi, N. 1992. Solution to shape and topology eigenvalue optimization problems using a homogenization method. International Journal for Numerical Methods in Engineering 35:1487–1502. Ditlevesen, O. 1979. Narrow reliability bounds for structural systems. Journal of Structural Mechanics 7:453–472. Ditlevsen, O. & Madsen, H.O. 1996. Structural reliability methods. New York, New York: Wiley. Ditlevsen, O., Bjerager, Olesen, R. & Hasofer, A.M. 1989. Directional Simulation in Gaussian Space. Probabilistic Engineering Mechanics 3(4):207–217. Ditlevsen, O., Oleson, R. & Mohr, G. 1987. Solution of a class of load combination problems by directional simulation. Structural Safety 4:95–109. Doltsinis, I. 1999. Elements of plasticity – Theory and computation. WIT Press, Southampton. Doltsinis, I. (ed.), 1999. Stochastic Analysis of Multivariate Systems in Computational Mechanics and Engineering. CIMNE, Barcelona. Doltsinis, I. 2003. Inelastic deformation processes with random parameters – methods of analysis and design. Comput. Methods Appl. Mech. Engrg. 192:2405–2423. Doltsinis, I. 2003. Large deformation processes of solids – From fundamentals to computer simulation and engineering applications. WIT Press, Southampton.

References

605

Doltsinis, I. & Kang, Z. 2004. Robust design of structures using optimization methods. Comput. Methods Appl. Mech. Engrg. 193:2221–2237. Doltsinis, I. & Rodic T. 1999. Process design and sensitivity analysis in metal forming. Int. J. Numer. Meth. Engrg. 45:661–692. Doltsinis, I., Kang, Z. & Cheng, G. 2005. Robust design of non-linear structures using optimization methods. Comput. Methods Appl. Mech. Engrg. 194:1779–1795. Drenick, R.F. 1970. Model-free design of aseismic structures. J. Engrg. Mech. Div. ASCE 96(EM4):483–493. Drezner, Z. & Erkut, E. 1995. Solving the continuous p-dispersion problem using nonlinear programming. The Journal of the Operational Research Society 46(4):516–520. Du, X. & Chen, W. 2000. An Integrated Methodology for Uncertainty Propagation and Management in Simulation-Based Systems Design. AIAA Journal 38(8):1471–1478. Du, X. & Chen, W. 2002. Sequential optimization and reliability assessment method for efficient probabilistic design. In ASME Design engineering technical conferences and computers and information in engineering conference, DETC2002/DAC-34127, Montreal, Canada. Du, X. & Chen, W. 2004. Sequential optimization and reliability assessment for efficient probabilistic design. ASME Journal of Mechanical Design 126(2):225–233. Du, X., Sudjianto, A. & Huang, B. 2005. Reliability-Based Design with a Mixture of Random and Interval Variables. ASME Journal of Mechanical Design 127:1068–1076. Dubois, D. & Prade, H. 1988. Possibility Theory. New York: Plenum Press. Dunnett, C.W. & Sobel, M. 1955. Approximations to the probability integral and certain percentage points of multivariate analogue of Student’s t-distribution. Biometrika 42: 258–260. Duysinx, P. 1997. Layout optimization: A mathematical programming approach. Technical Report DCAMM report No. 540, University of Liege. EC1. 2003. Eurocode 1: Basis of design and actions on structures – Part 2–4: Actions on Structures – Wind actions. CEN, ENV 1991-2-4, European Committee for Standardization, Brussels. EC3. 2003. Eurocode 3: Design of steel structures, Part 1–1: General rules for buildings. CEN, ENV 1993-1-1/1992, European Committee for Standardization, Brussels. EC5. 2005. Eurocode 5: Design of timber structures; Part 1–1: general rules and rules for buildings. EN 1995-1-1, European Committee for Standardization, Brussels. EC8. 2004. Eurocode 8: Design of structures for earthquake resistance – Part 1. European standard. CEN-ENV-1998-1, European Committee for Standardization, Brussels. Eldred, M., Giunta, A. & Collis, S. 2004. Second-order corrections for surrogate-based optimization with model hierarchies. In AIAA 2004-44570, 10th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, 30 August–1 September, Albany, NY. Eldred, M.S. & Bichon, B.J. 2006. Second-order reliability formulations in DAKOTA/UQ. In Proceedings of the 47th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference, Number AIAA-2006-1828, May 1–4, Newport, RI. Eldred, M.S. & Dunlavy, D.M. 2006. Formulations for surrogate-based optimization with data fit, multifidelity, and reduced-order models. In Proceedings of the 11th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, Number AIAA-2006-7117, September 6–8, Portsmouth, VA. Eldred, M.S., Adams, B.M., Copps, K.D., Carnes, B., Notz, P.K., Hopkins, M.M. & Wittwer, J.W. 2007. Solution-verified reliability analysis and design of compliant microelectro-mechanical systems. In Proceedings of the 9th AIAA Non-Deterministic Approaches Conference, Number AIAA-2007-1934, April 23–26, Honolulu, HI. Eldred, M.S., Agarwal, H., Perez, V.M., Wojtkiewicz, S.F. Jr. & Renaud, J.E. 2007. Investigation of reliability method formulations in DAKOTA/UQ. Structure & Infrastructure Engineering: Maintenance, Management, Life-Cycle Design & Performance. To appear.

606

References

Eldred, M.S., Brown, S.L., Adams, B.M., Dunlavy, D.M., Gay, D.M., Swiler, L.P., Giunta, A.A., Hart, W.E., Watson, J.-P., Eddy, J.P., Griffin, J.D., Hough, P.D., Kolda, T.G., MartinezCanales, M.L. & Williams, P.J. 2006. DAKOTA, a multilevel parallel object-oriented framework for design optimization, parameter estimation, uncertainty quantification, and sensitivity analysis: Version 4.0 users manual. Technical Report SAND2006-6337, Sandia National Laboratories, Albuquerque, NM. Eldred, M.S., Giunta, A.A., Wojtkiewicz, S.F. Jr. & Trucano, T.G. 2002. Formulations for surrogate-based optimization under uncertainty. In Proceedings of the 9th AIAA/ISSMO Symposium on Multidisciplinary Analysis and Optimization, Number AIAA-2002-5585, September 4–6, Atlanta, GA. Elishakoff, I. 1991. Essay on reliability index, probabilistic interpetation of safety factor and convex models of uncertainty. In Reliability Problems: General principles and Applications in Mechanics of Solids and Structures. F. Casciati & J.B. Roberts (eds), pp. 237–271. Wien: Springer-Verlag. Elishakoff, I. 1995. An idea on the uncertainty triangle. The Shock and Vibration Digest 22(10):1–1. Elishakoff, I. 1999. Are probabilistic and anti-optimization approaches compatible? In Whys and Hows in Uncertainty Modelling, I. Elishakoff (ed.), pp. 263–355. Wien: SpringerVerlag. Elishakoff, I. 2005. Safety Factors and Reliability: Friends or Foes? New York: Kluwer. Elishakoff, I., Haftka, R.T. & Fang, J. 1994. Structural design under bounded uncertainty optimization with anti-optimization. Computers and Structures 53(6):1401–1405. Elishakoff, I. & Ren, Y. 2003. Finite Element Methods for Structures with Large Stochastic Variations. Oxford: Oxford University Press. Elseifi, M.A., Gurdal, Z. & Nikolaidis, E. 1998. Convex and probabilistic models of uncertainties in geometric imperfections of stiffened composite panels. In AIAA/ASME/ ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference and Exhibit and AIAA/ASME/AHS Adaptive Structures Forum, Anon (ed.), Long Beach,CS, USA, April 20–23, Part 2 (of 4):1131–1140. Enevoldsen, I. 1994. Reliability-based optimization as an information tool. Mech. Struct. & Mach 22:117–135. Enevoldsen, I. 1994. Sensitivity Analysis of a Reliability-Based Optimal Solution. ASCE Journal of Engineering Mechanics 120(1):198–205. Enevoldsen, I. & Sørensen, J. 1992. Optimization algorithms for calculation of the joint design point in parallel systems. Structural and Multidisciplinary Optimization 4(2): 121–127. Enevoldsen, I. & Sørensen, J.D. 1993. Reliability-based optimization of series systems of parallel systems. ASCE Journal of Structural Engineering 119(4):1069–1084. Enevoldsen, I. & Sørensen, J. 1994. Reliability-based optimization as an information tool. Mechanics of Structures & Machines 22(1):117–135. Enevoldsen, I. & Sørensen, J. 1994. Reliability-based optimization in structural engineering. Structural Safety 15:169–196. Enevoldsen, S. & Sørensen, J.D. 1998. A Probabilistic Model for Chloride-Ingress and Initiation of Corrosion in Reinforced Concrete Structures. Structural Safety 20:69–89. Engels, H. 1980. Numerical quadrature and qubature. New York: Academic Press. Engelund, S. 1997. Probabilistic models and computational methods for chloride ingress in concrete. Ph.D. Thesis, Dept. of Building Technology and Structural Engineering, Aalborg University. Engelund, S. & Sørensen, J.D. 1998. A Probabilistic Model for Chloride-ingress and Initiation of Corrosion in Reinforced Concrete Structures. Structural Safety 20:69–89. Enright, M.P. & Wu, Y.-T. 1999. Probabilistic Fatigue Life Sensitivity Analysis of Titanium Rotors. In Proceedings of the AIAA 41st SDM Conference, Atlanta, GA.

References

607

Erstafr, G.K. 1998. A method for multi-parameter PDF estimation of random variables. Structural Safety 20:25–36. Eschenaueuer, H.A. & Olhoff, N. 2001. Topology optimization of continuum structures: A review. Applied Mechanics Reviews 54:331–390. Estes, A.C. & Frangopol, D.M. 1999. Repair optimization of highway bridges using system reliability approach. Journal of Structural Engineering ASCE 125(7):766–775. Estes, A.C. & Frangopol, D.M. 2001. Minimum expected cost-oriented optimal maintenance planning for deteriorating structures: Application to concrete bridge decks, Reliability Engineering & System Safety 73(3):281–291. Estes, A.C. & Frangopol, D.M. 2005. Life-cycle evaluation and condition assessment of structures, Chapter 36 In Structural Engineering Handbook, 2nd Edition, W.-F. Chen & E.M. Lui (eds), CRC Press, 36-1 to 36-51. Evans, D.H. 1972. An application of numerical integration techniques to statistical tolerancing, III – general distributions. Technometrics 14:23–35. Everett, R.A. Jr. 2002. Crack-Growth Characteristics of Fixed and Rotary Wing Aircraft. In 6th Joint FAA/DoD/NASA Aging Aircraft Conference. Faber, M.H., Engelund, S., Sørensen, J.D. & Bloch, A. 1989. Simplified and generic risk based inspection planning. Proc. OMAE2000, New Orleans. Fadel, G.M., Riley, M.F. & Barthelemy, J.-F.M. 1990. Two point exponential approximation method for structural optimization. Structural Optimization 2(2):117–124. Faravelli, L. 1989. Response surface approach for reliability analysis. ASCE Journal of Engineering Mechanics 115(12):2763–2781 Feng, Y.S. & Moses, F. 1986. A method of structural optimization based on structural system reliability. J. Struct. Mech. 14:437–453. Field, R. 2002. Numerical methods to estimate the coefficients of the polynomial chaos expansion. In 15th Engineering Mechanics Conference, Columbia University, NY. ASCE. Field, R., Red-Horse, J. & Paez, T. 2000. A nondeterministic shock and vibration application using polynomial chaos expansions. In 8th ASCE Joint Specialty Conference on Probabilistic Mechanics and Structural Reliability, South Bend, IN. Fiessler, B., Neumann, H. & Rackwitz, R. 1979. Quadratic limit states in structural reliability. ASCE Journal of Engineering Mechanics 105(4):661–676. Forth, S.C., Everett, Jr. R.A. & Newman, J.A. 2002. A Novel Approach to Rotorcraft Damage Tolerance. 6th Joint FAA/DoD/NASA Aging Aircraft Conference. Fox, B. 1966. Age replacement with discounting. Operations Research 14(3):533–537. Frangopol, D.M. 1984. A reliability-based optimization technique for automatic plastic design, Computer Methods in Applied Mechanics and Engineering 44:105–117. Frangopol, D.M. 1984. Interactive reliability-based structural optimization. Computers and Structures 19(4):559–563. Frangopol, D.M. 1985. Multicriteria reliability-based structural optimization. Structural Safety 3(1):23–28. Frangopol, D.M. 1985. Sensitivity of reliability-based optimum design, Journal of Structural Engineering ASCE 111(8):1703–1721. Frangopol, D.M. 1985. Structural optimization using reliability concepts. Journal of Structural Engineering ASCE 111(11):2288–2301. Frangopol, D.M. 1990. System reliability in structural analysis, design, and optimization, Structural Safety (Special Issue), D.M. Frangopol (guest ed.), 7(2–4):83–309. Frangopol, D.M. 1995. Reliability-based structural design. Chapter 16 In Probabilistic Structural Mechanics Handbook, C.R. Sundararajan (ed.), pp. 352–387. New York: Chapman & Hall. Frangopol, D.M. 1997. How to incorporate reliability in structural optimization, Chapter 11 In ASCE Manual on Engineering Practice No. 90: Guide to Structural Optimization, J.S. Arora (ed.), ASCE, New York, 211–235.

608

References

Frangopol, D.M. 1998. Probabilistic structural optimization. Progress in Structural Engineering and Materials 1(2):223–230. Frangopol, D.M. (ed.) 1998. Optimal Performance of Civil Infrastructure Systems ASCE (ISBN 0-7844-0315-5), Reston, Virginia, 222 pages. Frangopol, D.M. (ed.) 1999. Case Studies in Optimal Design and Maintenance Planning of Civil Infrastructure Systems ASCE (ISBN 0-7844-0420-8), Reston, Virginia, 272 pages. Frangopol, D.M. 1999. Life-cycle cost analysis for bridges. In Bridge safety and reliability. ASCE, Reston, Virginia, 210–236. Frangopol, D.M. 2000. Advances in life-cycle reliability-based technology for design an maintenance of structural systems. In Computational mechanics for the twenty-first century. Edinburgh: Saxe-Coburg Publishers, pp. 257–270. Frangopol, D.M. & Cheng, F.Y. (eds). 1997. Advances in Structural Optimization, ASCE (ISBN 0-7844-0221-3), New York, 225 pages. Frangopol, D.M. & Corotis, R.B. 1994. Reliability-based structural system assessment, design and optimization. Structural Safety (Special Issue), D.M. Frangopol & R.B. Corotis. (guest eds), Elsevier, 16(1+2):1–174. Frangopol, D.M. & Corotis, R. 1996. Reliability-based structural system optimization: State-ofthe-art verse state-of-the-practice. In Analysis and Computation: Proceedings of the Twelfth Conference held in Conjunction with Structures Congress XIV, Cheng (ed.), pp. 67–78. Frangopol, D.M. & Estes, A.C. 1999. Optimum lifetime planning of bridge inspection and repair programs, Structural Engineering International, Journal of IABSE 9(3):219–223. Frangopol, D.M. & Furuta, H. (eds) 2001. Life-Cycle Cost Analysis and Design of Civil Infrastructure Systems, ASCE (ISBN 0-7844-0571-9), Reston, Virginia, 336 pages. Frangopol, D.M. & Guedes Soares, C. 2001. Reliability oriented optimal structural design, Reliability Engineering & Systems Safety (Special Issue), D.M. Frangopol & C. Guedes Soares, (guest eds), 73(3):195–306. Frangopol, D.M. & Hendawi, S. 1994. Incorporation of corrosion effects in reliability-based optimization of composite hybrid plate girders. Structural Safety 16(1+2):145–169. Frangopol, D.M. & Liu, M. 2006. Multiobjective optimization of risk-based maintenance and life-cycle cost of civil infrastructure. System Modeling and Optimization (text of Plenary Lecture) E. Ceragioli, A. Dontchev, H. Furuta, K. Marti, & L. Pandolfi (eds), Boston, Springer, 123–136. Frangopol, D.M. & Liu, M. 2007. Maintenance and management of civil infrastructure based on condition, safety, optimization, and life-cycle cost. Structure and Infrastructure Engineering 3(1):29–41. Frangopol, D.M. & Maute, K. 2003. Life-cycle reliability-based optimization of civil and aerospace structures. Computers & Structures 81(7):397–410 (invited review article). Frangopol, D.M. & Maute, K. 2004. Reliability-Based Optimization of Civil and Aerospace Structural Systems. Engineering Design Reliability Handbook. Chapter 24, CRC Press, Boca Raton, Florida. Frangopol, D.M. & Maute, K. 2005. Reliability-based optimization of civil and aerospace structural systems, Chapter 24 In Engineering Design Reliability Handbook, E. Nikolaidis, D.M. Ghiocel & S. Singhal (eds), CRC Press, Boca Raton, 24-1 to 24-32. Frangopol, D.M. & Moses, F. 1994. Reliability-based structural optimization, Chapter 13 In Advances in Design Optimization, H. Adeli (ed), Chapman & Hall, London, pp. 492–570. Frangopol, D.M. & Neves, L.C. 2004. Probabilistic maintenance and optimization strategies for deteriorating civil infrastructures, Chapter 14 In Progress in Computational Structures Technology, B.H.V. Topping & C.A. Mota Soares (eds), Saxe-Coburg Publ., Stirling, Scotland, pp. 353–377. Frangopol, D.M. & Rondal, J. 1977. Reliability-based structural optimization. Het Ingenieursblad, Acta Technica Belgica, Anvers, Belgium, 46(7):189–195.

References

609

Frangopol, D.M. & Rondal, J. 1978. Optimum probabilistic design of structures, (in French), Annales Institut Technique du Batiment et des Travaux Publics, No. 363, Serie: Theories et Methodes de Calcul, France, 218:23–30. Frangopol, D.M., Brühwiler, E., Faber, M.H. & Adey, B. (eds) 2004. Life-Cycle Performance of Deteriorating Structures: Assessment, Design and Management, ASCE (ISBN 0-7844-0707-X), Reston, Virginia, 456 pages. Frangopol, D.M., Corotis, R.B. & Rackwitz, R. (eds) 1997. Reliability and Optimization of Structural Systems, Pergamon (ISBN 0-08-042826-6), Elsevier, 363 pages. Frangopol, D.M., Estes, A.C., Augusti, G. & Ciampoli, M. 1997. Optimal bridge management based on lifetime reliability and life – cycle cost, Chapter 8 In Optimal Performance of Civil Infrastructure Systems, D.M. Frangopol (ed), ASCE, New York, 98–115. Frangopol, D.M., Gharaibeh, E.S., Kong, J.S. & Miyake, M. 2000. Optimal network-level bridge maintenance planning based on minimum expected cost, Journal of the Transportation Research Board, Transportation Research Record, 1696(2), National Academy Press, 26–33. Frangopol, D.M., Klisinski, M. & Iizuka, M. 1991. Optimization of damage – tolerant structural systems. Computers and Structures 40(5):1085–1095. Frangopol, D.M., Lin, K-Y. & Estes, A.C. 1997. Life-cycle cost design of deteriorating structures. Journal of Structural Engineering ASCE 123(10):1390–1401. Frangopol, D.M., Miyake, M., Kong, J.S., Gharaibeh, E.S. & Estes, A.C. 2002. Reliabilityand cost-oriented optimal bridge maintenance planning. Chapter 10 In Recent Advances in Optimal Structural Design, S. Burns (ed.), ASCE, Reston, Virginia, pp. 257–270. Fu, G. & Frangopol, D.M. 1990. Balancing weight, system reliability and redundancy in a multiobjective optimization framework. Structural Safety 7(2–4):165–175. Fu, G. & Frangopol, D.M. 1990. Reliability-based vector optimization of structural systems. Journal of Structural Engineering ASCE 116(8):2143–2161. Fujimoto, Y., Itagaki, H., Itoh, S., Asada, H. & Shinozuka, M. 1989. Bayesian Reliability Analysis of Structures with Multiple Components. Proceedings ICOSSAR 89, pp. 2143–2146. Fujita, M., Schall, G. & Rackwitz, R. 1989. Adaptive Reliability Based Inspection Strategies for Structures Subject to Fatigue. Proceedings ICOSSAR 89, pp. 1619–1626. Furuta, H., Kameda, T., Nakahara, K., Takahashi, Y. & Frangopol, D.M. 2006. Optimal bridge maintenance planning using improved multi-objective genetic algorithm, Structure and Infrastructure Engineering 2(1):33–41. Gamerman, D. 1997. Markov Chain Monte Carlo. Chapman & Hall. Ganzerli, S., De Palma, P., Stackle, P. & Brown, A. 2005. Info-gap uncertainty in structural optimization via genetic algorithms. In Proceedings of the 9th International Conference on Structural Safety and Reliabilit; Proc. Intern. Conf., Reliability (ICOSSAR05), G. Augusti, G.I. Schuëller & M. Ciampoli (eds), Rome, Italy, June 19–22. Rotterdam: Millpress: 2325–2330. Ganzerli, S., DePalma, P., Smith, J.D. & Burkhart, M.F. 2003. Efficiency of genetic algorithms for optimal structural design considering convex models of uncertainty. In Proceedings of the 9th International Conference on Statistic and Probability on Civil Engineering (ICASP9), A. Der Kiureghian, S. Madanat & J.M. Pestana (eds), Berkeley, CA, July 6–9. Rotterdam: Millpress, pp. 1003–1010. Ganzerli, S. & Burkhart, M.F. 2002. Genetic algorithms for optimal structural design using convex models of uncertainties. In Proceedings of the 4th International Conference on Computational Stochastic Mechanics (CSM4); P. Spanos & G. Deodatis (eds), Kerkyra (Corfu), Greece, 9–12 June. Rotterdam: Millpress. Ganzerli, S. & Pantelides, C.P. 2000. Optimum structural design via convex model superposition. Computers and Structures 74(6):639–647. Gasser, M. & Schuëller, G.I. 1997. Reliability-based optimization of structural systems. Mathematical Methods of Operations Research 46:287–307.

610

References

Gasser, M. & Schuëller, G.I. 1998. Some basic principles in reliability-based optimization (RBO) of structures and mechanical components. In Stochastic programming methods and technical applications, K. Marti & P. Kall (eds), Lecture Notes in Economics and Mathematical Systems p. 458, Springer-Verlag, Berlin, Germany. Gayton, N., Mohamed-Chateauneuf, A., Sørensen, J.D., Pendola, M. & Lemaire, M. 2004. Calibration methods for reliability-based design codes. Structural Safety 26(1):91–121. Gentle, J.E. 1998. Random Number Generation and Monte Carlo Methods. Springer-Verlag, New York. Genz, A. 1992. Numerical computation of multivariate normal probabilities. Journal of Computational and Graphical Statistics 1:141–149. Ghanem, R.G. & Spanos, P.D. 1991. Stochastic Finite Elements: A Spectral Approach. Springer, Berlin. Ghasemi, M.R., Hinton, E. & Wood, R.D. 1999. Optimization of trusses using genetic algorithms for discrete and continuous variables. Engineering Computations 16(3):272–301. Gill, P., Murray, E.W., Saunders, M.A. & Wright, M.H. 1998. User’s guide for npsol 5.0: A fortran package for nonlinear programming. Technical Report SOL 86-1, System Optimization Laboratory, Stanford University, Stanford, CA. Gill, P.E., Murray, W. & Wright, M.H. 1981. Practical Optimization. Academic Press. Giunta, A.A. & Eldred, M.S. 2000. Implementation of a trust region model management strategy in the DAKOTA optimization toolkit. In Proceedings of the 8th AIAA/USAF/NASA/ISSMO Symposium on Multidisciplinary Analysis and Optimization, Number AIAA-2000-4935, September 6–8, Long Beach, CA. Glasserman, P. & Yao, D.D. 1992. Some guidelines and guarantees for common random numbers. Management Science 38:884–908. Goldberg, D.E. 1989. Genetic algorithms in search optimization and machine learning. New York: Addison-Wesley. Gollwitzer, S. & Rackwitz, R. 1983. Equivalent components in first-order system reliability. Reliability Engineering 5:99–115. Gollwitzer, S. & Rackwitz, R. 1988. An efficient numerical solution to the multinormal integral. Probabilistic Engineering Mechanics 3(2):98–101. Goulet, C.A., Haselton, C.B., Mitrani-Reiser, J., Beck, J.L., Deierlein, G., Porter, K.A. & Stewart, J.P. 2007. Evaluation of the seismic performance of code-conforming reinforcedconcrete frame building-From seismic hazard to collapse safety and economic losses. Earthquake Engineering and Structural Dynamics. Article in Press. Grandhi, R. & Wang, L. 1998. Reliability-based structural optimization using improved two-point adaptive nonlinear approximations. Finite Elements in Analysis and Design 35–48. Greenwood, W.H. & Chase, K.W. 1990. Root sum squares tolerance analysis with nonlinear problems. ASME Journal of Engineering for Industry 112:382–4. Gu, X., Renaud, J.E. & Batill, S.M. 1998. An Investigation of Multidisciplinary Design Subject to Uncertainties. In 7th AIAA/USAF/NASA/ISSMO Multidisciplinary Analysis & Optimization Symposium, St. Louis, Missouri. Guan, X.L. & Melchers, R. 2001. Effect of response surface parameter variation on structural reliability estimates. Structural Safety 23:429–444. GUCEA, 2007. Gonzaga University Center for Evolutionary Algorithms (GUCEA), http://www. cs.gonzaga.edu/gucea/. Gupta, A. & Krawinkler, H. 2000. Behavior of ductile SMRFs at various seismic hazard levels. ASCE Journal of Structural Engineering 126(1):98–107. Gurdal, Z. & Tatting, B. 2000. Cellular automata for design of truss structures with linear and nonlinear response. In Proceedings of the 41st AIAA Structures, Strucutural Dynamics, and Materials Conference, Number 2000-1580, April 3–6, Atlanta, Georgia. Haftka, R.T. & Kamat, M.P. 1985. Elements of Structural Optimization. Martinus Nijhoff, The Hague.

References

611

Haftka, R.T., Gurdal, Z. & Kamat, M.P. 1990. Elements of Structural Optimization. Dordrecht: Kluwer Academic Publishers. Hahn, G.J. & Shapiro, S.S. 1967. Statistical Models in Engineering. John Wiley & Sons: New York. Haimes, Y.Y., Tarvainen, K., Shima, T. & Thadathil, J. 1990. Hierachical Multiobjective Analysis of Large-Scale Systems. Hempisphere Publishing Corporation, pp. 41–42. Hajela, P. & Kim B. 2001. On the use if energy minimization for ca based analysis in elasticity. Struct. Multidisc. Optim. 23:24–33. Haldar, A. & Mahadevan, S. 2001. Probability, Reliability and Statistical Methods in Engineering Design. Wiley. Hamming, R.W. 1973. Numerical Methods for Scientists and Engineers. New York: Dover Publications. Hansen, E. & Walster, G.W. 2004. Global Optimization using Interval Analysis. New York: Marcel Dekker, Inc. Harbitz, A. 1986. An Efficient Sampling Method for Probability of Failure Calculation. Structural Safety 3:109–115. Harkness, H.H., Fleming, M., Moran, B. & Belytschko, T. 1994. Fatigue Reliability With InService Inspections. FAA/NASA International Symposium on Advanced Structural Integrity Methods for Airframe Durability and Damage Tolerance. Harr, M. 1989. Probabilistic estimates for multivariate analysis. Applied Mathemathical Modelling 13:313–318. Hasofer, A. 1974. Design for infrequent overloads. Earthquake Engineering and Structural Dynamics 2(4). Hasofer, A.M. & Lind, N.C. 1974. Exact and invariant second order code format. ASCE Journal of Engineering Mechanics 100(1):111–121. Hasofer, A.M. & Rackwitz, R. 2000. Time-dependent models for code optimization. In Proceedings of the 8th International Conference on Applications of Statistics and Probability (ICASP8), R.E. Melchers & M.G. Stewart (eds), Sydney, Australia, December 1999, Volume 1, Rotterdam/Brookfield, pp. 151–158. CERRA: A.A. Balkema. Haug, E.J. & Choi, K.K. 1982. Systematic occurrence of repeated eigenvalues in structural optimization. Journal of Optimization Theory and Applications 38:251–274. Haupt, L.H. & Haupt, S.E. 1998. Practical Genetic Algorithms. New York: John Wiley & Sons. He, L. & Polak, E. 1990. Effective diagonalization strategies for the solution of a class of optimal design problems. IEEE Transactions on Automatic Control 35(3):258–267. Helmberg, C. 2002. Semidefinite programming. European Journal of Operational Research 137:461–482. Hernandez, S. 1990. Métodos de diseño óptimo de estructuras. Madrid: Colegio de Ingenieros de Caminos, Canales y Puertos. Hisada, T. & Nakagiri, S. 1981. Stochastic finite element method developed for structural safety and reliability. In Proceedings of the Third International Conference on Structural Safety and Reliability, pp. 395–408. Rotterdam: Elsevier. Hohenbichler, M. & Rackwitz, R. 1981. Non-normal dependent vectors in structural safety. ASCE Journal of the Engineering Mechanics Division 107(6):1227–1249. Hohenbichler, M. & Rackwitz, R. 1983. First-order concepts in system reliability. Structural Safety 1(3):177–188. Hohenbichler, M. & Rackwitz, R. 1986. Sensitivity and importance measures in structural reliability. Civil Eng. Syst. 3:203–209. Hohenbichler, R. & Rackwitz, R. 1988. Improvement of Second-order Reliability Estimates by Importance Sampling. J. Eng. Mech. ASCE 114(12):2195–2199. Holicky, M. & Markova J. 2003. Reliability analysis of impacts due to road vehicles. In Applications of Statistics and Probability in Civil Engineering, A. Der Kiureghian, S. Madanat & J.M. Pestana (eds), Rotterdam, Netherlands, pp. 1645–1650. Millpress.

612

References

Holland, J.H. 1975. Adaptation in natural and artificial systems. Ann Arbor: The University of Michigan Press. Hong, H.P. 1998. An efficient point estimate method for probabilistic analysis. Reliability Engineering and System Safety 59:261–267. Hong, H.P. 1999. Simple approximations for improving second-order reliability estimates. J. Eng. Mech. ASCE 125(5):592–595. Hong, H.P. 1996. Point-estimate moment-based reliability analysis. Civil Engineering Systems 13(4):281–294. Hong, H.P., Escobar, J.A. & Gomez, R. 1998. Probabilistic assessment of the in seismic response of structural asymmetric models. In Proceedings of the Tenth European Conference on Earthquake Engineering, Paris, 1998, Rotterdam. Balkema. Housner, G.W. 1959. Behavior of structures during earthquakes. Journal of the Engineering Mechanics Division ASCE 85(4):109–129. Housner, G.W. & Jennings, P.C. 1975. The capacity of extreme earthquake motions to damage structures. Structural and geotechnical mechanics: A volume honoring N.M. Newmark edited by W.J. Hall, 102–116, Prentice-Hall Englewood Cliff, NJ. Huh, J.S., Kim, K.H., Kang, D.W., Gweon, D.G. & Kwak, B.M. 2006. Performance evaluation of precision nanopositioning devices caused by uncertainties due to tolerances using function approximation moment method. Review of Scientific Instrument 77:015103. Hurtado, J.E. 2001. Neural networks in stochastic mechanics. Archives of Computational Methods in Engineering 8:303–342. Hurtado, J.E. 2004. An examination of methods for approximating implicit limit state functions from the viewpoint of statistical learning theory. Structural Safety 26:271–293. Hurtado, J.E. 2004. Structural Reliability. Statistical Learning Perspectives. Heidelberg: Springer. Hurtado, J.E. 2006. Optimal reliability-based design using support vector machines and artificial life algorithms. In Intelligent Computational Paradigms in Earthquake Engineering, N.D. Lagaros & Y. Tsompanakis (eds), Hershey: Idea Group Inc. Hurtado, J.E. 2007. Filtered importance sampling with support vector margin: a powerful method for structural reliability analysis. Structural Safety 29:2–15. Hurtado, J.E. & Alvarez, D.A. 2001. Neural network-based reliability analysis: A comparative study. Computer Methods in Applied Mechanics and Engineering 191:113–132. Hurtado, J.E. & Barbat, A. 1998. Fourier-based maximum entropy method in stochastic dynamics. Structural Safety 20:221–235. Igusa, T. & Wan, Z. 2003. Response surface methods for optimization under uncertainty. In Proceedings of the 9th International Conference on Application of Statistics and Probability, A. Der Kiureghian, S. Madanat & J. Pestana (eds), San Francisco, California. Igusa, T. & Der Kiureghian, A. 1988. Response of uncertain systems to stochastic excitations. J. Eng. Mech. ASCE 114(5):812–832. Inou, N., Shimotai, N. & Uesugi, T. 1994. Cellular automaton generating topological structures. In 2nd European Conference on Smart Structures and Materials, Number 2361-08, 1994, October Glasgow, United Kingdom, pp. 47–50. Inou, N., Uesugi, T., Iwasaki, A. & Ujihashi, S. 1998. Self-organization of mechanical structure by cellular automata. Fracture and Strength of Solids 145(9):1115–1120. ISO 19902. 2001. Petroleum and natural gas industries – Fixed steel offshore structures. Itoh, Y. & Liu, C. 1999. Multiobjective optimization of bridge deck maintenance. In Case Studies in Optimal Design and Maintenance Planning if Civil Infrastructure Systems, D.M. Frangopol (ed.), ASCE, Reston, Virginia. Iwan, W.D. & Cifuentes, A.O. 1986. A model for system identification of degrading structures. International Journal of Earthquake Engineering and Structural Dynamics 14:877–890.

References

613

Jaynes, E.T. 1957. Information Theory and Statistical Mechanics. The Physical Review 106: 620–630. Jaynes, E.T. 2003. Probability theory: the logic of science. Cambridge, UK: Cambridge University Press. JCSS. 2001. Probabilistic Model Code. Joint Committee on Structural Safety, http://www.jcss. ethz.ch [accessed 15 September 2006]. Jensen, B.D., Parkinson, M.B., Kurabayashi, K., Howell, L.L. & Baker, M.S. 2001. Design optimization of a fully-compliant bistable micro-mechanism. In Proc. 2001 ASME Intl. Mech. Eng. Congress and Exposition, New York, NY. Jensen, H. 2000. On the structural synthesis of uncertain systems subjected to environmental loads. Structural and Multidisciplinary Optimization 20:37–48. Jensen, H.A. 2005. Structural optimization of linear dynamical systems under stochastic excitation: a moving reliability database approach. Computer Methods in Applied Mechanics and Engineering 194(12–16):1757–1778. Jensen, H. & Iwan, W.D. 1992. Response of systems with uncertain parameters to stochastic excitations. J. Eng. Mech. ASCE 114:1012–1025. Joanni, A.E. & Rackwitz, R. 2006. Stochastic dependencies in inspection, repair and failure models. In Proceedings of the European Safety and Reliability Conference, C. Guedes Soares & E. Zio (eds), Estoril, Portugal, London, pp. 531–537. Taylor & Francis. Johnson, N.L., Kotz, S. & Balakrishnan, N. 1994. Continuous Univariate Distributions, Vol. 1. New York: John Wiley and Sons. Jones, D., Shonlau M. & Welch W. 1998. Efficient global optimization of expensive black-box functions. INFORMS J. Comp. 12:272–283. Jones, D.R., Perttunen, C.D. & Stuckman, B.E. 1993. Lipschitzian Optimization Without the Lipschitz Constant. Journal of Optimization Theory and Applications 73(1):157–181. Jung, D.H. & Lee, B.C. 2002. Development of a simple and efficient method for robust optimization. Int. J. Numer. Meth. Engng. 53:2201–2215. Kale, A., Haftka, R.T. & Sankar, B.V. 2007. Efficient Reliability Based Design and Inspection of Stiffened Panels Against Fatigue. Journal of Aircraft. Kang, Z. 2005. Robust Design Optimization of Structures under Uncertainties. Doctoral Thesis, University of Stuttgart. Kanno, Y. & Takewaki, I. 2006. Robustness analysis of trusses with separable load and structural uncertainties. International Journal of Solids and Structures 43:2646–2669. Kanno, Y. & Takewaki, I. 2006. Sequential semidefinite program for maximum robustness design of structures under load uncertainties. Journal of Optimization Theory and Applications 130:265–287. Kanno, Y. & Takewaki, I. 2007. Worst-case plastic limit analysis of trusses under uncertain loads via mixed 0-1 programming. Journal of Mechanics of Materials and Structures 2(2): 247–273. Kanzow, C., Nagel, C., Kato, H. & Fukushima, M. 2005. Successive linearization methods for nonlinear semidefinite programs. Computational Optimization and Applications 31: 251–273. Kapur, J.N. 1989. Maximum Entropy Models in Science and Engineering. New York: John Wiley and Sons. Karamchandani, A. 1990. New Methods in Systems Reliability, Ph.D dissertation, Stanford University. Karamchandani, A. & Cornell, CA. 1991. Adaptive Hybrid Conditional Expectation Approaches for Reliability Estimation. Structural Safety 11:59–74. Karamchandani, A. & Cornell, C.A. 1992. Sensitivity estimation within first and second order reliability methods. Struct. Saf. 11:95–107.

614

References

Kaymaz, I. & Marti, K. 2006. Reliability-based design optimization for elastoplastic mechanical structures. Computers and Structures. Article In Press. Kemeny, D.C., Howell, L.L. & Magleby, S.P. 2002. Using compliant mechanisms to improve manufacturability in MEMS. In Proc. 2002 ASME DETC, Number DETC2002/DFM-34178. Kennedy, C.A. & Lennox, W.C. 2000. Solution to the practical problem of moments using nonclassical orthogonal polynomials with applications for probabilistic analysis. Probabilistic Engineering Mechanics 15:371–379. Kennedy, C.A. & Lennox, W.C. 2001. Moment operations on random variables, with applications for probabilistic analysis. Probabilistic Engineering Mechanics 16: 253–259. Kharitonov, V. 1997. Interval uncertainty structure: Conservative but simple. In Uncertainty: Models and Measures, H. Gunther-Naske & Y. Ben-Haim (eds), pp. 231–243. Berlin: Akademie Verlag. Kharmanda, G. 2004. Two points of view for developing reliability-based design optimization. NT2F4 (New Trends in Fatigue and Fracture IV), Aleppo, Syria, 10–12 May. Kharmanda, G., Altonji, A. & El-Hami, A. 2006. Safest point method for reliability-based design optimization of freely vibrating structures, 1st International Francophone Congress for Advanced Mechanics, IFCAM01, Aleppo, Syria, 02–04 May. Kharmanda, G., El-Hami, A. & Olhoff, N. 2004. Global Reliability–Based Design Optimization, In: Frontiers on Global Optimization, C.A. Floudas (ed.), 255(20), Kluwer Academic Publishers, January. Kharmanda, G., Makhloufi, A. & El-Hami, A. 2007. Efficient computing time reduction for reliability-based design optimization, Qualita, 20–22 March, Tanger, Maroc. Kharmanda, G., Mohamed, A. & Lemaire, M. 2001. New hybrid formulation for reliabilitybased optimization of structures, The 4th World Congress of Structural and Multidisciplinary Optimization, WCSMO-4, Dalian, China, June 4–8. Kharmanda, G., Mohamed, A. & Lemaire, M. 2002. Efficient reliability-based design optimization using hybrid space with application to finite element analysis. Structural and Multidisciplinary Optimization 24:233–245. Kharmanda, G., Mohamed-Chateauneuf, A. & Lemaire, M. 2002. CAROD: Computer-Aided Reliable and Optimal Design as a concurrent system for real structures. Journal of Computer Aided Design and Computer Aided Manufacturing CAD/CAM 1(1):1–12. Kharmanda, G., Mohamed, A. & Lemaire, M. 2003. Integration of reliability-based design optimization within CAD and FE models. In Recent Advances in Integrated Design and Manufacturing in Mechanical Engineering, Kluwer Academic Publishers. Kharmanda, G., Olhoff, N., Mohamed, A. & Lemaire, M. 2004. Reliability-based topology optimization. Structural and Multidisciplinary Optimization 26:295–307. Kharmanda, G., Olhoff, N. & El-Hami, A. 2004. Recent Developments in Reliability-Based Design Optimization (Keynote Lecture). In Computational Mechanics, Proc. Sixth World Congress of Computational Mechanics (WCCM VI in conjunction with APCOM’04), Sept. 5–10. Beijing, China. Tsinghua University Press & Springer-Verlag. Kharmanda, G., Olhoff, N. & El-Hami, A. 2004. Optimum values of structural safety factors for a predefined reliability level with extension to multiple limit states. Structural and Multidisciplinary Optimization 27:421–434. Kharmanda, G. & Olhoff, N. 2007. Extension of optimum safety factor method to nonlinear reliability-based design optimization. Journal of Structural and Multidisciplinary Optimization (to appear). Kim, H.M. 2001. Target Cascading in Optimal System Design. Ph.D. Thesis, University of Michigan. Kim, H.M., Kokkolaras, M., Louca, L.S., Delagrammatikas, G.J., Michelena, N.F., Filipi, Z.S., Papalambros, P.Y., Stein, J.L. & Assanis, D.N. 2002. Target cascading in vehicle redesign: A class VI truck study. International Journal of Vehicle Design 29(3):1–27.

References

615

Kim, H.M., Michelena, N.F., Papalambros, P.Y. & Jiang T. 2003. Target cascading in optimal system design. SME Journal of Mechanical Design 125(3):474–480. Kim, H.M., Rideout, D.G., Papalambros, P.Y. & Stein, J.L. 2003. Analytical target cascading in automotive vehicle design. ASME Journal of Mechanical Design 125(3):481–489. Kim, T.-U. & Sin, H.-C. 2001. Optimal design of composite laminated plates with the discreteness in ply angles and uncertainty in material properties considered. Computers and Structures 79(29–30):2501–2509. Kirjner-Neto, C., Polak, E. & Der Kiureghian, A. 1998. An outer approximations approach to reliability-based optimal design of structures. Journal Optim. Theory Appl. 98(1):1–16. Kirsch, U. 1981. Optimum Structural Design. New York: McGraw-Hill, Inc. Kirsch, U. 1993. Structural Optimization. Fundamentals and Applications. Heidelberg: Springer Verlag. Kirsch, U. 1999. Efficient, accurate reanalysis for structural optimization. AIAA Journal 37(12):1663–1669. Kirsch, U. 2000. Combined approximations – a general reanalysis approach for structural optimization. Structural and Multidisciplinary Optimization 20(2):97–106. Kirsch, U. 2001. Exact and accurate solutions in the approximate reanalysis of structures. AIAA Journal 39(11):2198–2205. Kirsch, U. 2002. A unified reanalysis approach for structural analysis, design, and optimization. Structural and Multidisciplinary Optimization 25(1):67–85. Kirsch, U. 2003. Approximate vibration reanalysis of structures. AIAA Journal 41(3): 504–511. Kita, E. & Toyoda, T. 2000. Structural design using cellular automata. Struct. Multidisc. Optim. 19:64–73. Kleiber, M. & Hien, T.D. 1992. The Stochastic Finite Element Method. Chichester: John Wiley and Sons. Kleinmann, N.L., Spall, J.C. & Naiman, D.C. 1999. Simulation-based optimization with stochastic approximation using common random numbers. Management Science 45(11):1570–1578. Klir, G. 1997. Uncertainty theories, models and principles: An overview of personal views and contributions. In Uncertainty: Models and Measures, H. Gunther-Natke & Y. Ben-Haim (eds), pp. 27–43. Berlin: Akademie Verlag. Klir, G.J. & Filger, T.A. 1988. Fuzzy Sets, Uncertainty, and Information. Prentice Hall. Klir, G.J. & Yuan, B. 1995. Fuzzy Sets and Fuzzy Logic: Theory and Applications. Prentice Hall. Koch, P.N., Yang, R.J. & Gu, L. 2004. Design for six sigma through robust optimization. Struct. and Multidisc. Optim. 26:235–248. Koch, P.N., Yang, R.J. & Gu, L. 2007. Design for six sigma through robust optimization. Structural and Multidisciplinary Optimization. Article in Press. Kocvara, M. 1997. Topology optimization with displacement constraints: a bilevel programming approach. Struct. Optim. 4:256–263. Kojima, M. & Tunçel, L. 2000. Cones of matrices and successive convex relaxations of nonconvex sets. SIAM Journal on Optimization 10:750–778. Kokkolaras, M., Fellini, R., Kim, H.M., Michelena, N.F. & Papalambros, P.Y. 2002. Extension of the target cascading formulation to the design of product families. Structural and Multidisciplinary Optimization 24(4):293–301. Kokkolaras, M., Louca, L.S., Delagrammatikas, G.J., Michelena, N.F., Filipi, Z.S., Papalambros, P.Y., Stein, J.L. & Assanis, D.N. 2004. Simulation-based optimal design of heavy trucks by model-based decomposition: An extensive analytical target cascading case study. International Journal of Heavy Vehicle Systems 11(3–4):402–432. Kokkolaras, M., Mourelatos, Z.P. & Papalambros, P.Y. 2004. Design optimization of hierarchically decomposed multilevel systems under uncertainty. In Proceedings of the

616

References

ASME Design Engineering Technical Conferences, Salt Lake City, Utah, Paper No. DETC2004/DAC-57357. Kokkolaras, M., Mourelatos, Z.P. & Papalambros, P.Y. 2006. Design optimization of hierarchically decomposed multilevel systems under uncertainty. ASME Journal of Mechanical Design 128(2):503–508. Kokkolaras, M., Mourelatos, Z.P. & Papalambros, P.Y. 2006. Impact of uncertainty quantification on design decisions for a hydraulic-hybrid powertrain engine. In Proceedings of the 47th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, Newport, Rhode Island, Paper No. AIAA-2006-2001. Kolassa, J.E. 1997. Series Approximation Methods in Statistics. New York: Springer-Verlag. Kong, J.S. & Frangopol, D.M. 2003. Life-cycle reliability-based maintenance cost optimization of deteriorating structures with emphasis on bridges. Journal of Structural Engineering ASCE 129(6):818–828. Kong, J.S. & Frangopol, D.M. 2004. Cost-reliability interaction in life-cycle cost optimization of deteriorating structures. Journal of Structural Engineering ASCE 130(11): 1704–1712. Kong, J.S. & Frangopol, D.M. 2005. Probabilistic optimization of aging structures considering maintenance and failure costs. Journal of Structural Engineering ASCE 131(4):600–616. Kosko, B. 1992. Neural Networks and Fuzzy Systems. Englewood Cliffs: Prentice Hall. Koyluoglu, H.U. & Nielsen, S.R.K. 1994. New approximations for SORM integrals. Structural Safety 13:235–246. Koyluoglu, H.U., Cakmak, A.S. & Nielsen, S.R.K. 1995. Interval algebra to deal with pattern loading and structural uncertainties. J. Eng. Mech. ASCE 121(11):1149–1157. Kramer, S.L. 2003. Geotechnical Earthquake Engineering. Prentince Hall. Krishnamurthy, T. 2003. Response Surface Approximation with Augmented and Compactly Supported Radial Basis Functions. Proceedings of the AIAA 44th SDM Conference. Kroon, I.B. 1994. Decision Theory Applied to Structural Engineering Problems. Ph.D. Thesis, Dept. of Building Technology and Structural Engineering, Aalborg University. Kuschel, N. & Rackwitz, R. 1997. Two basic problems in reliability-based structural optimization. Mathematical Methods of Operations Research 46:309–333. Kuschel, N. & Rackwitz, R. 1998. Structural optimization under time-variant reliability constraints. In Proceeding of the 8th IFIP WG 7.5 Working conference on Reliability and Optimization of Structural Systems, edited by Nowak, University of Michigan, Ann Arbor, Michigan, USA, pp. 27–38. Kuschel, N. & Rackwitz, R. 2000. Optimal design under time-variant reliability constraints. Structural Safety 22(2):113–127. Kuschel, N. & Rackwitz, R. 2000. Time-variant reliability-based structural optimization using SORM. Optimization 47(3/4):349–368. Kuschel, N. & Rackwitz, R. 2000. A new approach for structural optimization of series system. In: Proceedings of the 8th International conference on applications of statistics and probability (ICASP) in Civil engineering reliability and risk analysis, R.E. Melchers & M.G. Stewart (eds), Sydney, Australia, December 1999, (2):987–994. Kushner, H.J. & Yin, G.G. 2003. Stochastic approximation and recursive algorithms and applications. New York: Springer Lagaros, N. & Papadrakakis, M. 2003. Soft computing methodologies for structural optimization. Applied Soft Computing 3:283–300. Lagaros, N.D. & Papadrakakis, M. 2006. Robust seismic design optimization of steel structures, J. Struct. Multidisc. Optim. Available on-line, Doi:10.1007/s00158-006-0047-5. Lagaros, N.D., Papadrakakis, M. & Kokossalakis, G. 2002. Structural optimization using evolutionary algorithms. Computers and Structures 80(7–8):571–589. Lagaros, N.D., Plevris, V. & Papadrakakis, M. 2005. Multi-objective design optimization using cascade evolutionary computations. Comput. Methods Appl. Mech. Engrg. 194(30–33):3496–3515.

References

617

Lagaros, N.D. & Tsompanakis, Y. (eds) 2006. Intelligent computational paradigms in earthquake engineering. Idea Publishers. Lange, K. 1999. Numerical Analysis for Statisticians. New York: Springer-Verlag. Lawrence, C., Zhou, J.L. & Tits, A.L. 2007. User’s guide for CFSQP version 2.5. Available from http://www.aemdesign.com. Lee, K.H. & Park, G.J. 2001. Robust optimization considering tolerances of design variables. Computers and Structures 7:77–86. Lee, J.O., Yang, Y.O. & Ruy, W.S. 2002. A Comparative Study on Reliability Index and Target Performance Based Probabilistic Structural Design Optimization. Computers and Structures 80:257–269. Lee, T.W. & Kwak, B.M. 1987–88. A reliability-based optimal design using advanced first order second moment method. Mechanics of Structures and Machines 15(4):523–542. Lee, S.H. & Kwak, B.M. 2005. Reliability based design optimization using response surface augmented moment method. Proceedings of 6th World Congress on Structural and Multidisciplinary Optimization, Rio de Janeiro, Brazil. Lee, S.H. & Kwak, B.M. 2006. Response surface augmented moment method for efficient reliability analysis. Structural Safety 28:261–272. Lee, W.J. & Woo, T.C. 1990. Tolerances: their analysis and synthesis. ASME Journal of Engineering for Industry 112:113–121. Legresley, P. & Alonso, J. 2001. Investigation of nonlinear projection for POD based reduced order models for aerodynamics. In AIAA 2001-16737, 39th Aerospace Sciences Meeting & Exhibit, January 8–11, Reno, NV. Lemaire, M., in collaboration with Chateauneuf, A. & Mitteau, J.C. 2006. Structural Reliability. ISTE, UK. Leverant, G.R., Littlefield, D.L., McClung, R.C., Millwater, H.R. & Wu, Y.-T. 1997. A Probabilistic Approach to Aircraft Turbine Rotor Material Design. The International Gas Turbine & Aeroengine Congress & Exhibition, Paper No. 97-GT-22, Orlando, FL. Lewis, K. & Mistree, F. 1997. Collaborative, Sequential and Isolated Decisions in Design. Proceedings of ASME Design Engineering Technical Conferences, Paper# DETC1997/ DTM-3883. Li, K.S. & Lumb, P. 1985. Reliability analysis by numerical integration and curve fitting. Structural Safety 3:29–36. Liang, J., Mourelatos, Z.P. & Tu, J. 2007. A Single-Loop Method for Reliability-Based Design Optimization. In press International Journal of Product Development. Also, Proceedings of ASME Design Engineering Technical Conferences, 2004, Paper # DETC2004/DAC57255. Liang, J., Mourelatos, Z.P. & Nikolaidis, E. 2007. A Single-Loop Approach for System Reliability-Based Design Optimization. ASME Journal of Mechanical Design (accepted). Liang, J., Mourelatos, Z.P. & Tu, J. 2004. A Single-Loop Method for Reliability-Based Design Optimization. International Journal of Product Development. Interscience Enterprises Limited, Great Britain (in press). Lin, K.-Y. & Frangopol, D.M. 1996. Reliability – based optimum design of reinforced concrete girders. Structural Safety 18(2/3):239–258. Lind, N.C. 1977. Reliability based structural codes, practical calibration. Safety of structures under dynamic loading, Trondheim, Norway, pp. 149–160. Lindley, D.V. 1976. Introduction to Probability and Statistics from a Bayesian Viewpoint, Vol. 1+2. Cambridge University Press, Cambridge. Liu, H., Chen, W., Kokkolaras, M., Papalambros, P.Y. & Kim, H.M. 2006. Probabilistic analytical target cascading – a moment matching formulation for multilevel optimization under uncertainty. ASME Journal of Mechanical Design 128(4):991–1000. Liu, P.-L. & Der Kiureghian, A. 1991. Optimization algorithms for structural reliability. Structural Stafety 9:161–177.

618

References

Liu, M. & Frangopol, D.M. 2004. Optimal bridge maintenance planning based on probabilistic performance prediction. Engineering Structures 26(7):991–1002. Liu, M. & Frangopol, D.M. 2005. Maintenance planning of deteriorating bridges by using multiobjective optimization. Transportation Research Record, Journal of the Transportation Research Board, CD 11-S, Transportation Research Board of the National Academies, Washington, D.C., pp. 491–500. Liu, M. & Frangopol, D.M. 2005. Balancing connectivity of deteriorating bridge networks and long-term maintenance cost through optimization. Journal of Bridge Engineering ASCE 10(4):468–481. Liu, M. & Frangopol, D.M. 2005. Bridge annual maintenance prioritization under uncertainty by multiobjective combinatorial optimization. Computer Aided Civil and Infrastructure Engineering 20(5): 343–353. Liu, M. & Frangopol, D.M. 2005. Multiobjective maintenance planning optimization for deteriorating bridges considering condition, safety and life-cycle cost. Journal of Structural Engineering ASCE 131(5):833–842. Liu, M. & Frangopol, D.M. 2006. Optimizing bridge network maintenance management under uncertainty with conflicting criteria: Life-cycle maintenance, failure, and user costs. Journal of Structural Engineering ASCE 132(11):1835–1845. Liu, P.-L. & Kuo, C.-Y. 2003. Safety evaluation of the upper structure of bridge based on concrete non-destructive tests. In Applications of Statistics and Probability in Civil Engineering, A. Der Kiureghian, S. Madanat & J.M. Pestana (eds), Rotterdam, The Netherlands, pp. 1683–1688. Millpress. Liu, P.-L. & Der Kiureghian, A. 1986. Multivariate Distribution Models with Prescribed Marginals and Covariances. Probabilistic Engineering Mechanics 1(2): 105–112. Liu, W.K., Belytschko, T. & Lua, Y.J. 1995. Probabilistic finite element method. In Probabilistic Structural Mechanics Handbook, C.R. Sundararajan (ed.), pp. 70–105. New York: Chapman & Hall. Lombardi, M. & Haftka, R.T. 1998. Anti-Optimization Technique for Structural Design under Load Uncertainties. Computer Methods in Applied Mechanics and Engineering 157:19–31. Louca, S., Kokkolaras, M., Delagrammatikas, G.J., Michelena, N.F., Filipi, Z.S., Papalambros, P.Y. & Assanis, D.N. 2002. Analytical target cascading for the design of an advanced technology heavy truck. In Proceedings of the 2002 ASME International Mechanical Engineering Congress and Exposition, New Orleans, LA. Paper No. IMECE-2002-32860. Lugannani, R. & Rice, S. 1980. Saddle point approximation for the distribution of sums of random variables. Advances in Applied Probability 12:475–490. Luo, X. & Grandhi, R.V. 1997. ASTROS for reliability-based multidisciplinary structural analysis and optimization. Computers and Structures 62:737–745. Lyon, R.H. 1975. Statistical energy analysis of dynamical systems, The MIT Press, Cambridge, MA. Madsen, H.O., Krenk, S. & Lind, N.C. 1986. Methods of Structural Safety, Englewood Cliffs, New Jersey: Prentice Hall. Madsen, H.O. & Friis Hansen, P. 1991. Comparison of some algorithms for reliability-based structural optimization and sensitivity analysis. In Reliability and Optimization of Structural Systems, C.A. Brebbia & S.A. Orszag (eds), Springer-Verlag, Germany, pp. 443–451. Madsen, H.O. & Friis-Hansen, P. 1992. A comparison of some algorithms for reliability-based structural optimization and sensitivity analysis. In Proceedings of the 4th IFIP WG 7.5 Working Conference on Reliability and Optimization of Structural Systems, R. Rackwitz & P. ThoftCristensen (eds), Munich, Germany, September 1991, Berlin, pp. 443–451. IFIP: Springer. Madsen, H.O., Sørensen, J.D. & Olesen, R. 1989. Optimal Inspection Planning for Fatigue Damage of Offshore Structures. Proceedings ICOSSAR 89, pp. 2099–2106.

References

619

Madsen, H.O. & Sørensen, J.D. 1990. Probability-Based Optimization of Fatigue Design Inspection and Maintenance. Presented at Int. Symp. on Offshore Structures, University of Glasgow. Madsen, H.O., Skjong, R.K., Talin, A.G. & Kirkemo, F. 1987. Probabilistic Fatigue Crack Growth Analysis of Offshore Structures, with Reliability Updating Through Inspection. SNAME, Arlington, VA. Marler, R.T. & Arora, J.S. 2004. Survey of Multi-objective Optimization Methods for Engineering. J. Struct. Multidisc. Optim. 26(6):369–395. Marsh, P.S. & Frangopol, D.M. 2007. Lifetime multi-objective optimization of cost and spacing of corrosion rate sensors embedded in a deteriorating reinforced concrete bridge deck. Journal of Structural Engineering ASCE 133(6):777–787. Marti, K. 1996. Differentiation formulas for probability functions: the transformation method. Mathematical Programming 75:201–220. Marti, K. 2005. Stochastic Optimization Methods. Berlin: Springer. Martin, J.D. & Simpson, T.W. 2005. Use of Kriging Models to Approximate Deterministic Computer Models. AIAA Journal 3(4). Masur, E. 1984. Optimal structural design under multiple eigenvalue constraints. International Journal of Solids and Structures 20:117–120. Mathworks, Inc. 2004. Matlab reference manual, Version 7.0. Natick, Massachusetts: Mathworks, Inc. Mathworks, Inc. 2007. www.mathworks.com. Mattson, C.A., Mullur, A.A. & Messac, A. 2004. Smart pareto filter: Obtaining a minimal representation of multiobjective design space. Eng. Optim. J. 36(6):721–740. Maute, K. & Frangopol, D.M. 1998. Reliability-based design of mems mechanisms by topology optimization. Computers and Structures 81:813–824. Maute, K. & Frangopol, D.M. 2003. Reliability-based design of MEMS mechanisms by topology optimization. Computers & Structures, 81(8–11), K.J. Bathe 60th Anniversary Issue pp. 813–824. McAllister, C.D. & Simpson, T.W. 2003. Multidisciplinary Robust Design Optimization of an Internal Combustion Engine. ASME Journal of Mechanical Design 125(1):124–130. McKay, M.D., Beckman, R.J. & Conover, W.J. 1979. A comparison of three methods for selecting values of input variables in the analysis of output from a computer code. Technometrics 21(2):239–245. Mead, L.R. & Papanicolau, N. 1984. Maximum entropy in the problem of moments. Journal of Mathematical Physics 25:2404–2417. Melchers, R.E. 1987. Structural Reliability: Analysis and Prediction, Wiley. Melchers, R.E. 1989. Importance sampling in structural systems. Structural Safety 6(1):3–10. Melchers, R.E. 2001. Structural Reliability Analysis and Prediction. John Wiley & Sons. Merriam-Webster on-line. 2007. www.m-w.com, accessed April 2007. Messac, A. & Ismail-Yahaya, A. 2002. Multiobjective robust design using physical programming. J. Struct. Multidisc. Optim. 23:357–371. Meza, J.C. 1994. OPT++: An object-oriented class library for nonlinear optimization. Technical Report SAND94-8225, Sandia National Laboratories, March, Albuquerque, NM. Michelena, N.F., Kim, H.M. & Papalambros, P.Y. 1999. A system partitioning and optimization approach to target cascading. In Proceedings of the 12th International Conference on Engineering Design, Munich, Germany. Michelena, N.F., Louca, L., Kokkolaras, M., Lin, C.-C., Jung, D., Filipi, Z., Assanis, D., Papalambros, P.Y., Peng, H., Stein, J. & Feury, M. 2001. Design of an advanced heavy tactical truck: A target cascading case study. SAE 2001 Transactions – Journal of Commercial Vehicles. Also appeared in the Proceedings of the 2001 SAE International Truck and Bus Meeting and Exhibition, Chicago, IL, Paper No. 2001-01-2793.

620

References

Michelena, N.F., Park, H. & Papalambros, P.Y. 2003. Convergence properties of analytical target cascading. AIAA Journal 41(5):897–905. Miller, A.C. & Rice, T.R. 1983. Discrete approximations of probability distributions. Management Science 29:352–362. Mills-Curran, W. 1988. Calculation of eigenvector derivatives for structures with repeated eigenvalues. AIAA Journal 26(7):867–871. Millwater, H.R., Fitch, S., Riha, D.S., Enright, M.P., Leverant, G.R., McClung, R.C., Kuhlman, C.J., Chell, G.G. & Lee, Y.-D. 2000. A Probabilistically-Based Damage Tolerance Analysis Computer Program for Hard Alpha Anomalies In Titanium Rotors. Proceedings of the 45th ASME International Gas Turbine & Aeroengine Technical Congress, Munich, Germany. Millwater, H.R., Wu, Y.-T., Cardinal, J.W. & Chell G.G. 1996. Application of Advanced Probabilistic Fracture Mechanics to Life Evaluation of Turbine Rotor Blade Attachments. Journal of Engineering for Gas Turbines and Power 18:394–398. Mogami, K., Nishiwaki, S., Izui, K., Yoshimura, M. & Kogiso, N. 2006. Reliability-based structural optimization of frame structures for multiple failure criteria using topology optimization techniques. Struct. Multidisp. Optim. 32(4):299–311. Mohsine, A. 2006. Contribution à l’optimisation fiabiliste en dynamique des structures mécaniques. Thèse de doctorat, INSA de Rouen, France (French version). Mohsine, A., Kharmanda, G. & El-Hami, A. 2006. Improved hybrid method as a robust tool for reliability-based design optimization, Structural and Multidisciplinary Optimization 32:203–213. Moilanen, A. & Wintle, B.A. 2006. Uncertainty analysis favors selection of spatially aggregated reserve structures. Biological Conservation 129:427–434. Moore, R.E. 1966. Interval Analysis. Prentice-Hall. More, J.J., Garbow, B.S. & Hillstrom, K.E. 1980. User guide for MINPACK-1. Argonne National Labs Report ANL-80-74. Argonne. Illinois. Moré, J.J. & Sørensen, D.C. 1983. Computing a Trust Region Step. SIAM Journal on Scientific and Statistical Computing 3:553–572. Mori, Y. & Ellingwood, B.R. 1993. Time-dependent system reliability analysis by adaptive importance sampling. Structural Safety 12(1):59–73. Moses, F. 1973. Desing for Reliability: Concepts and Applications. John Wiley & Sons. Moses, F. 1977. Structural System Reliability and Optimization. Comput. Struct. 7:283–290. Moses, F. 1995. Probabilistic Analysis of Structural Systems. Probabilistic Structural Mechanics Handbook: Theory and Industrial Applications, edited by C. Raj Sundararajan, Chapman & Hall, 166–187. Moses, F. 1997. Problems and prospects of reliability based optimization. Engineering Structures 19(4):293–301. Mourelatos, Z.P. & Zhou, J. 2005. Reliability Estimation with Insufficient Data Based on Possibility theory. AIAA Journal 43(8):1696–1705. Mourelatos, Z.P. & Zhou, J. 2006. A Design Optimization Method using Evidence Theory. ASME Journal of Mechanical Design 128(4):901–908. Muhanna, R.L. & Mullen, R.L. 2001. Uncertainty in Mechanics Problems – Interval-Based Approach. Journal of Engineering Mechanics 127(6):557–566. Mullen, R.L. & Muhanna, R.L. 1999. Bounds of Structural Response for all Possible Loadings. ASCE Journal of Structural Engineering 125(1):98–106. Murotsu, Y. & Shao, S. 1989. Optimum shape design of truss structures based on reliability. Struct. Multidisc. Optim. 2(2):65–76. Murotsu, Y., Kishi, M., Okada, H., Yonezawa, M. & Taguchi, K. 1984. Probabilistically optimum design of frame structures. Proc. 11th IFIP Conf. on System Modeling and Optimization, Springer-Verlag, pp. 545–554. Murotsu, Y., Shao, S. & Watanabe, A. 1994. An approach to reliability-based optimization of redundant structures. Structural Safety 16:133–143.

References

621

Muscolino, G. 1993. Response of linear and non-linear structural systems under gaussian or non-gaussian filtered input. In Dynamic Motion: Chaotic and Stochastic Behaviour, F. Casciati (ed.), pp. 203–299. Wien: Springer-Verlag. Myers, R.H. & Montgomery, D.C. 1995. Response surface methodology: process and product optimization using designed experiments. New York: John Wiley & Sons. Nakamura, H., Miyamoto, A. & Kawamura, K. 2000. Optimization of bridge maintenance strategies using GA and IA techniques. In Reliability and Optimization of Structural Systems, Proceedings IFIP WG 7.5, A.S. Nowak & M.M. Szerszen (eds), Ann Arbor, Michigan. Nakib, R. & Frangopol, D.M. 1990. RBSA and RBSA-OPT: Two computer programs for structural system reliability analysis and optimization. Computers and Structures 36(1): 13–27. Nakib, R. & Frangopol, D.M. 1990. Reliability-based structural optimization using interactive graphics. Computers and Structures 37(1):27–34. Neves, L.A.C., Frangopol, D.M. & Cruz, P.J.S. 2006. Probabilistic lifetime-oriented multiobjective optimization of bridge maintenance: single maintenance type. Journal of Structural Engineering ASCE 132(6):991–1005. Neves, L.A.C., Frangopol, D.M. & Petcherdchoo, A. 2006. Probabilistic lifetime-oriented multiobjective optimization of bridge maintenance: combination of maintenance types. Journal of Structural Engineering ASCE 132(11):1821–1834. Nie, J. & Ellingwood, B.R. 2005. Finite element-based structural reliability assessment using efficient directional simulation. ASCE Journal of Engineering Mechanics 131(3):259–267. Nieuwenhof, B. & Coyotte, J. 2002. A perturbation stochastic finite element method for the time-harmonic analysis of structures with random mechanical properties. In 5th World Congress on Computational Mechanics, Vienna, Austria. Nikolaidis, E. & Burdisso, R. 1988. Reliability-based optimization: a safety index approach. Computers and Structures 28(6):781–788. Nikolaidis, E., Chen, S., Cudney, H., Haftka, R.T. & Rosca, R. 2004. Comparison of Probability and Possibility for Design Against Catastrophic Failure Under Uncertainty. ASME Journal of Mechanical Design 126:386–394. Nikolaidis, E., Ghiocel, D.M. & Singhal, S. (eds) 2005. Engineering Design Reliability Handbook, CRC Press, Boca Raton, FL. Nurdin, H. 2002. Mathematical modeling of bias and uncertainty in accident risk assessment. Mathematical Sciences, University of Twente, The Netherlands. Oberkampf, W., Helton, J. & Sentz, K. 2001. Mathematical Representations of Uncertainty. AIAA Non-Deterministic Approaches Forum, AIAA 2001-1645, Seattle, WA, April 16–19. Oberkampf, W.L. & Helton, J. 2002. Investigation of Evidence Theory for Engineering Applications. AIAA Non-Deterministic Approaches Forum, AIAA 2002-1569, Denver, CO. Ohsaki, M., Fujisawa, K., Katoh, N. & Kanno, Y. 1999. Semi-definite programming for topology optimization of truss under multiple eigenvalue constraints. Computer Methods in Applied Mechanics and Engineering 180:203–217. Ordaz, M. 1988. On the use of probability concentrations. Structural Safety 5:317–318. Ordaz, M., Huerta, B. & Reinoso, E. 2003. Exact computation of input-energy spectra from Fourier amplitude spectra. Earthquake Engineering and Structural Dynamics 32: 597–605. Overbay, S., Ganzerli, S., De Palma, P., Stackle, P. & Brown, A. 2006. Trusses, NPCompleteness, and Genetic Algorithms. In Proceedings of the 17th Analysis and Computation Specialty Conference; Proc. Conf., F.A. Charney, D.E. Grierson, M. Hoit & J.M. Pestana (eds), Saint Louis, MO, May 18–20. Reston: ASCE Publications. Owen, A.B. 1997. Monte Carlo variance of scrambled net quadrature. SIAM J. Num. Anal. 34(5):1884–1910. Page, C.H. 1952. Instantaneous power spectra. Journal of Applied Physics 23(1):103–106.

622

References

Palmberg, B., Blom, A.F. & Eggwertz. 1987. Probabilistic Damage Tolerance Analysis of Aircraft Structures. In Probabilistic Fracture Mechanics and Reliability, J.W. Provan (ed.), Martinus Nijhoff Publishers. Pandey, M.D. 1998. An effective approximation to evaluate multinormal integrals. Structural Safety 20(1):51–67. Pandey, M.D. & Ariaratman, S.T. 1996. Crossing rate analysis of non gaussian response of linear systems. Journal of Engineering Mechanics 122:507–511. Pantelides, C.P. & Tzan, S-R. 1996. Convex models for seismic design of structures – I. Earthquake Eng. Struct. Dyn. 25(9):927–944. Pantelides, C.P. & Ganzerli, S. 1989. Design of trusses under uncertain loads using convex models. Journal of Structural Engineering (ASCE) 124:318–329. Papadimitriou, C., Beck, J.L. & Katafygiotis, L.S. 2001. Updating robust reliability using structural test data. Probabilistic Engineering Mechanics 16:103–113. Papadrakakis, M., Lagaros, N. & Tsompanakis, Y. 1998. Structural optimization using evolution strategies and neural networks. Computer Methods in Applied Mechanics and Engineering 156:309–333. Papadrakakis, M., Papadopoulos, V. & Lagaros, N. 1996. Structural reliability analysis of elastic-plastic structures using neural networks and Monte Carlo simulation. Computer Methods in Applied Mechanics and Engineering 136:145–163. Papadrakakis, M., Tsompanakis, Y. & Lagaros, N.D. 1999. Structural shape optimization using evolution strategies. Eng. Optim. J. 31:515–540. Papalambros, P.Y. 2001. Analytical target cascading in product development. In Proceedings of the 3rd ASMO UK/ISSMO Conference on Engineering Design Optimization, Harrogate, North Yorkshire, England. Papalambros, P.Y. & Wilde, D.J. 2000. Principles of Optimal Design; Modeling and Computation. 2nd Edition, Cambridge University Press. Papoulis, A. 1991. Probability, Random Variables and Stochastic Processes. New York: McGraw-Hill. Park, G.-J., Lee, T.-H., Lee, K.H. & Hwang, K.-H. 2006. Robust design: An overview. AIAA J. 44(1):181–191. Patel, N.M., Agarwal, H., Tovar, A. & Renaud, J. 2005. Reliability based topology optimization using the hybrid cellular automaton method. In 1st AIAA Multidisciplinary Design Optimization Specialist Conference, April 18–21 Austin, Texas. Penmetsa, R.C. & Grandhi, R.V. 2002. Estimating Membership Response Function using Surrogate Models. Paper AIAA 2002-1234. Penmetsa, R.C. & Grandhi, R.V. 2002. Efficient Estimation of Structural Reliability for Problems with Uncertain Intervals. Computers and Structures 80:1103–1112. Pierce, S.G., Worden, K. & Manson, G. 2006. A novel information-gap technique to assess reliability of neural network-based damage detection. Journal of Sound and Vibration 293:96–111. Polak, E. 1997. Optimization. Algorithms and consistent approximations. New York: Springer-Verlag. Polak, E. & Royset, J.O. 2007. Efficient sample sizes in stochastic nonlinear programming. J. Computational and Applied Mathematics. To appear. Porter, K.A., Beck, J.L., Shaikhutdinov, R.V., Au, S.K., Mizukoshi, K., Miyamura, M., Ishida, H., Moroi, T., Tsukada, Y. & Masuda, M. 2004. Effect of seismic risk on lifetime property values. Earthquake Spectra 20:1211–1237. Powell, M.J.D. 1982. VMCWD: A FORTRAN Subroutine for Constrained Optimization. Report DAMTP 1982/NA4, Cambridge University, England. Powell, M.J.D. 1994. A direct search optimization method that models the objective and constraint functions by linear interpolation. In Proceedings of the 6th Workshop on Optimization

References

623

and Numerical Analysis, S. Gomez & J.-P. Hennart (eds), Oaxaca, Mexico, January 1992, Dordrecht, pp. 51–67. Kluwer Academic Publishers. Pradlwater, H.J., Schuëller, G.I., Koutsourelakis, P.S. & Champris, D.C. 2007. Application of line sampling simulation method to reliability benchmark problems. Structural Safety (Article in Press). Pshenichnyj, B.N. 1994. The Linearization Method for Constrained Optimization, Volume 22 of Computational Mathematics. Berlin: Springer. Qiu, Z. 2003. Comparison of static response of structures using convex models and interval analysis method. Numerical Methods in Engineering 56(12):1735–1753. Qiu, J. & Slocum, A.H. 2004. A curved-beam bistable mechanism. J. Microelectromech. Syst. 13(2):137–146. Qiu, Z. & Wang, X. 2003. Comparison of dynamic response of structures with uncertainbut-bounded parameters using non-probabilistic interval analysis method and probabilistic approach. Int. J. Solids and Structures 40:5423–5439. Qiu, Z. & Wang, X. 2006. Interval analysis method and convex models for impulsive response of structures with uncertain-but-bounded external loads. Acta Mecanica Sinica 22(3):265–276. Qu, X. & Haftka, R.T. 2003. Design under uncertainty using Monte Carlo simulation and probabilistic sufficiency factor. In Proceedings of DET’03 Conference, Chicago, IL, USA. Rackwitz, R. 2000. Optimization – the basis of code making and reliability verification. Structural Safety 22(1):27–60. Rackwitz, R. 2001. Risk control and optimization for structural facilities. Proc. 20th IFIP TC7 Conf. On System modeling and optimization. Trier, Germany. Rackwitz, R. 2001. Reliability Analysis – A Review and Some Perspectives. Structural Safety 23:365–395. Rackwitz, R. 2002. Optimization and risk acceptability based on the Life Quality Index. Structural Safety 24:297–331. Rackwitz, R. & Fiessler, B. 1978. Structural reliability under combined random load sequences. Comput. Struct. 9:489–494. Rackwitz, R., Lentz, A. & Faber, M.H. 2005. Socio-economically sustainable civil engineering infrastructures by optimization. Structural Safety 27(3):187–229. Raiffa, H. & Schlaifer, R. 1961. Applied Statistical Decision Theory. Harward University Press, Cambridge, Mass. Rajashekhar, M.R. & Ellingwood, B.R. 1993. A new look at the response surface approach for reliability analysis. Structural Safety 12:205–220. Rajeev, S. & Krishnamoorthy, C.S. 1997. Genetic algorithms-based methodologies for design optimization of trusses. ASCE J. of Structural Engineering 123(3):350–358. Rao, S.S. & Cao, L. 2002. Optimum Design of Mechanical Systems Involving Interval Parameters. ASME Journal of Mechanical Design 124:465–472. Rao, S.S. & Sawyer, J.P. 1995. A Fuzzy Finite Element Approach for the Analysis of Imprecisely Defined Systems. AIAA Journal 33:2264–2370. Ravindran, S. 1999. Proper orthogonal decomposition in optimal control of fluids. Technical report, NASA TM-1999-209113. Reid, N. 1988. Saddlepoint methods and statistical inference. Statistical Science 3:213–238. Rencher A.C. 1995. Methods of Multivariate Analysis. New York: John Wiley & Sons. Robert, C.P. & Casella, G. 2004. Monte Carlo Statistical Methods (2nd edition). New York: Springer. Robinson, J. 1982. Saddlepoint approximations for permutation tests and confidence intervals. Journal of the Royal Statistical Society Series B 44:91–101. Rosen Group. 2004. Metal Loss Inspection Performance Specifications, www.Roseninspection. net,Standard_CDP_POFspec_56_rev3.62.doc.

624

References

Rosenblatt, M. 1952. Remarks on a Multivariate Transformation. The Annals of Mathematical Statistics 23(3):470–472. Rosenblueth, E. 1975. Point estimates for probability moments. Proceedings of the National Academy of Sciences of the USA 72:3812–3814. Rosenblueth, E. 1976. Optimum Design for Infrequent Disturbances. ASCE Journal of the Structural Division 102(9):1807–1825. Rosenblueth, E. 1981. Two-point estimate in probabilities. Applied Mathematical Modeling 5:329–335. Rosenblueth, E. & Mendoza, E. 1971. Reliability optimization in isostatic structures. ASCE Journal of the Engineering Mechanics Division 97(6):1625–1642. Ross, T.J. 1995. Fuzzy Logic with Engineering Applications. McGraw Hill. Royset, J.O. & Polak, E. 2004. Implementable algorithm for stochastic programs using sample average approximations. J. Optimization Theory and Application 122(1):157–184. Royset, J.O. & Polak, E. 2004. Reliability-based optimal design using sample average approximations. J. Probabilistic Engineering Mechanics 19(4):331–343. Royset, J.O. & Polak, E. 2007. Extensions of stochastic optimization results from problems with simple to problems with complex failure probability functions. J. Optimization Theory and Application 133(1):1–18. Royset, J.O. & Polak, E. 2007. Efficient sample size in stochastic nonlinear programming. Journal of Computational and Applied Mathematics (Article in Press). Royset, J.O., Der Kiureghian, A. & Polak, E. 2001. Reliability-based optimal design of series structural systems. Journal of Engineering Mechanics 607–614. Royset, J.O., Der Kiureghian, A. & Polak, E. 2001. Reliability-based optimal structural design by the decoupling approach. Reliability Engineering and System Safety 73: 213–221. Royset, J.O., Der Kiureghian, A. & Polak, E. 2006. Optimal design with probabilistic objective and constraints. J. Engineering Mechanics 132(1):107–118. Rozvany, G.I.N. 1997. Topology Optimization in Structural Mechanics. Springer. Rozvany, G.I.N., Bendsoe, M.P. & Kirsh, U. 1995. Optimality criteria: a basis for multidisciplinary optimization. Appl. Mech. Rev. 48:41–119. Rubinstein, R. & Shapiro, A. 1993. Discrete Event Systems: Sensitivity Analysis and Stochastic Optimization by the Score Function Method. New York: John Wiley & Sons. Rummelhart, D.E. & McClelland, J.L. 1986. Parallel Distributed Processing, Volume 1: Foundations. The MIT Press, Cambridge. Ruszczynski, A. & Shapiro, A. 2003. Stochastic Programming. New York: Elsevier. Sacks, J., Schiller, S.B. & Welch, W.J. 1989. Design for Computer Experiments. Technometrics, 31(1). Schemit, L.A. & Lai, Y.C. 1994. Structural optimization based on preconditioned conjugate gradient analysis methods. International Journal for Numerical Methods in Engineering 37(6):943–964. Schittkowski, K. 1986. NLPQL: A FORTRAN Subroutine Solving Non-Linear Programming Problems. Annals of Operations Research. Schuëller, G. 1997. A state-of-the-art report on computational stochastic mechanics. Probabilistic Engineering Mechanics 5:485–500. Schuëller, G.I. 1998. Structural Reliability – Recent Advances. Proc. 7th ICOSSAR’97, pp. 3–33. Schuëller, G. 2001. Computational stochastic mechanics – recent advances. Computers & Structures 79:2225–2234. Schuëller, G.I. 2005. Special Issue on Computational Methods in Stochastic Mechanics and Reliability Analysis. Comput. Methods Appl. Mech. Engrg. 194(12–16): 1251–1795.

References

625

Schuëller, G.I. & Stix, R. 1987. A critical appraisal of methods to determine failure probabilities. Structural Safety 4:293–309. Schwefel, H.P. 1981. Numerical optimization for computer models. Wiley & Sons, Chichester, UK. Sentz, K. & Ferson, S. 2002. Combination of Evidence in Dempster – Shafer Theory. Sandia National Laboratories Report SAND2002-0835. Seo, H.S. & Kwak, B.M. 2002. Efficient statistical tolerance analysis for general distributions using three-point information. International Journal of Production Research 40(4): 931–944. Sexsmith, R.G. 1999. Probability-based safety analysis – value and drawbacks. Structural Safety 21:303–310. Shannon, C.E. 1948. A Mathematical Theory of Communication. The Bell System Technical Journal 27:379–423. Shiao, M.C. 2006. Risk-Based Maintenance Optimization. Proceedings of the International Conference on Structural Safety and Reliability. Shiao, M.C. & Wu, Y.-T. 2004. An Efficient Simulation-Based Method for Probabilistic Damage Tolerance Analysis With Maintenance Planning. Proceedings of the ASCE Specialty Conference on Probabilistic Mechanics and Reliability. Shinozuka, M. 1970. Maximum structural response to seismic excitations. J. Engrg. Mech. Div. ASCE 96(EM5):729–738. Shore, J.E. & Johnson, R.W. 1980. Axiomatic derivation of the principle of maximum entropy and the principle of minimum cross-entropy. IEEE Transactions on Information Theory 26(1):26–37. Skjong, R. 1985. Reliability-Based Optimization of Inspection Strategies. Proc. ICOSSAR’85 II:614–618. Slotta, D., Tatting, B., Watson, L., Gurdal, Z. & Missoum, S. 2002. Convergence analysis for cellular automata applied to truss design. Engineering Computations 19(8): 953–969. Sobczyk, K. & Trebicki, J. 1990. Maximum entropy principle in stochastic dynamics. Probabilistic Engineering Mechanics 5:102–110. Sørensen, D.C. 1994. Minimization of a Large Scale Quadratic Function Subject to an Ellipsoidal Constraint. Department of Computational and Applied Mathematics, Rice University. Technical Report TR94-27. Sørensen, J.D. & Tarp-Johansen, N.J. 2005. Reliability-based optimization and optimal reliability level of offshore wind turbines. International Journal of Offshore and Polar Engineering (IJOPE) 15(2):1–6. Sørensen, J.D. & Tarp-Johansen, N.J. 2005. Optimal Structural Reliability of Offshore Wind Turbines. CD-rom Proc. ICOSSAR’2005, Rome. Sørensen, J.D. & Thoft-Christensen, P. 1988. Inspection Strategies for Concrete Bridges. Proc. IFIP WG 7.5, Springer-Verlag 48:325–335. Sørensen, J.D. & Thoft-Christensen, P. 1985. Structural optimization with reliability constraints. Proc. 12th IFIP Conf. On ‘System modeling and optimization. Springer-Verlag, pp. 876–885. Sørensen, J.D. & Frangopol, D.M. (eds) 2006. Advances in Reliability and Optimization of Structural Systems (ISBN 0-415-39901-7), Taylor & Francis Group plc, London, 308 pages. Sørensen, J.D., Kroon, I.B. & Faber, M.H. 1994. Optimal reliability-based code calibration. Structural Safety 15:197–208. Sørensen, J.D., Thoft-Christensen, P., Siemaszko, A., Cardoso, J.M.B. & Santos, J.L.T. 1995. Interactive reliability-based optimal design. Proc. 6th IFIP WG 7.5 Conf. On Reliability and optimization of structural systems. Chapman & Hall, pp. 249–256. Spall, J.C. 2003. Introduction to stochastic search and optimization. New York: WileyInterscience.

626

References

Spletzer, J.R. & Taylor, J.C. 2003. A bounded uncertainty approach to multi-robot localization. In IEEE International Conference on Intelligent Robots and Systems; Proc. Intern. Conf., Anon (ed.), Las Vegas, October, 27–31. Institute of Electrical and Electronics Engineers Inc. Stadler, W. 1984. Multicriteria optimization in mechanics (A Survey). Appl. Mech. Rev. 37: 277–286. Steihaug, T. 1983. The Conjugate Gradient Method and Trust Regions in Large Scale Optimization. SIAM Journal on Numerical Analysis 20:626–637. Streicher, H. & Rackwitz, R. 2002. Structural optimization – a one level approach. Proc. Workshop on Reliability-based design and optimization – rbo02. IPPT, Warsaw. Streicher, H. 2004. Zeitvariante zuverlässigkeitsorientierte Kosten-Nutzen-Optimierung für Strukturen unter Berücksichtigung neuer Ansätze für Erneuerungs- und Instandhaltungsmodelle. PhD dissertation, Technische Universität München, Munich, Germany. In German. Streicher, H. & Rackwitz, R. 2004. Time-variant reliability-oriented structural optimization and a renewal model for life-cycle costing. Probabilistic Engineering Mechanics 19(1–2):171–183. Streicher, H., Joanni, A. & Rackwitz, R. 2006. Cost-benefit optimization and target reliability levels for existing, aging and maintained structures. Structural Safety. Accepted for publication. Sturm, J.F. 1999. Using SeDuMi 1.02, a MATLAB toolbox for optimization over symmetric cones. Optimization Methods and Software 11/12:625–653. Svanberg, K. 1987. The method of moving asymptotes a new method for structural optimization. Int. J. Numer. Meth. Engrg. 24:359–373. Szidarovszky, F. & Bahill, A.T. 1992. Linear Systems Theory. Boca Raton: CRC Press. Taflanidis, A.A. & Beck, J.L. 2007. Stochastic subset optimization for optimal reliability problems. Journal of Probabilistic Engineering Mechanics (Article in Press). Taflanidis, A.A. & Beck, J.L. 2007. Stochastic subset optimization for stochastic design. ECCOMAS Thematic Conference on Computational Methods in Structural Dynamics and Earthquake Engineering, Rethymno, Greece, 13–16 June. Taguchi, G. 1978. Performance analysis design. International Journal of Production Research 16:521–530. Takewaki, I. 2001. A new method for nonstationary random critical excitation. Earthquake Engineering and Structural Dynamics 30(4):519–535. Takewaki, I. 2001. Probabilistic critical excitation for MDOF elastic-plastic structures on compliant ground. Earthquake Engineering and Structural Dynamics 30(9):1345–1360. Takewaki, I. 2002. Critical excitation method for robust design: A review. Journal of Structural Engineering ASCE 128(5):665–672. Takewaki, I. 2002. Robust building stiffness design for variable critical excitations. Journal of Structural Engineering ASCE 128(12):1565–1574. Takewaki, I. 2004. Bound of Earthquake Input Energy. Journal of Structural Engineering ASCE 130(9):1289–1297. Takewaki, I. 2006. Critical Excitation Methods in Earthquake Engineering. Elsevier Science Publishers, Amsterdam. Takewaki, I. & Ben-Haim, Y. 2005. Info-gap Robust Design with Load and Model Uncertainties. Journal of Sound and Vibration 288(3):551–570. Tamma, K.K., Zhou, X. & Sha, D. 2000. The time dimension: a theory towards the evolution, classification, characterization and design of computational algorithms for transient/dynamic applications. Archives of Computational Methods in Engineering 7:67–290. Thanedar, P.B. & Kodiyalam, S. 1992. Structural Optimization Using Probabilistic Constraints. Structural Optimization 4:236–240. Thoft-Christensen, P. & Murotsu, Y. 1986. Application of Structural Systems Theory. Springer.

References

627

Thoft-Christensen, P. & Sørensen, J.D. 1987. Optimal Strategies for Inspection and Repair of Structural Systems. Civil Engineering Systems 4:94–100. Thomas, J., Dowell, E. & Hall, K. 2001. Three-dimensional transonic aeroelasticity using proper orthogonal decomposition based reduced order models. In 42nd AIAA/ASME/ASCE/ AHS/ASC Structures, Structural Dynamics and Materials (SDM) Conference, April, Seattle, WA, AIAA Paper 2001-1526. Tonon, F., Bernardini, A. & Elishakoff, I. 2001. Hybrid analysis of uncertainty: Probability, fuzziness and anti-optimization. Chaos, Solutions and Fractals 12(8): 1403–1414. Torczon, V. & Trosset, M.W. 1998. Using approximations to accelerate engineering design optimization. In Proceedings of the 7th AIAA/USAF/NASA/ISSMO Symp. on Multidisciplinary Analysis and Optimization, AIAA Paper 98-4800, St. Louis, Missouri. Torng, T.Y. & Yan, R.J. 1993. Robust structural system design using a system reliability-based design optimization method. In Probabilistic Mechanics: Advances in structural reliability methods, P.D. Spanos & Y.-T. Wu (eds), Springer-Verlag, NY:534–549. Tovar, A. 2004. Bone Remodeling as a Hybrid Cellular Automaton Optimization Process. Ph.D. Thesis, University of Notre Dame. Tovar, A., Patel, N.M., Kaushik, A.K. & Renaud, J.E. 2007. Optimality conditions of the hybrid cellular automata for structural optimization. AIAA Journal 43(3):673–683. Tovar, A., Quevedo, W., Patel, N. & Renaud, J. 2006. Topology optimization with stress and displacement constraints usng the hybrid cellular automaton method. In Proceedings of the 3rd European Conference on Computational Mechanics, June 5–8, Lisbon, Portugal. Trebicki, J. & Sobczyk, K. 1996. Maximum entropy principle and non-stationary distributions of stochastic systems. Probabilistic Engineering Mechanics 11:169–178. Tretiakov, G. 2002. Stochastic quasi-gradient algorithms for maximization of the probability function. A new formula for the gradient of the probability function. In Stochastic Optimization Techniques pp. 117–142. Springer, New York. Tsompanakis, Y. & Papadrakakis, M. 2004. Large-scale reliability-based structural optimization. J. Struct. Multidisc. Optim. 26:1–12. Tsompanakis, Y., Lagaros, N.D. & Stavroulakis, G. 2007. Soft computing techniques in parameter identification and probabilistic seismic analyses of structures. J. Advances in Eng. Software (Article in Press). Tu, J. & Jones, D.R. 2003. Variable Screening in Metamodel Design by Cross-Validated Moving Least Squares Method. Proceedings 44th AIAA/ASME/ASCE/ AHS/ASC Structures, Structural Dynamics and Materials Conference, AIAA-2003-1669, Norfolk, VA. Tu, J., Choi, K.K. & Park, Y.H. 1999. A New Study on Reliability-Based Design Optimization. ASME Journal of Mechanical Design 121:557–564. Tu, J., Choi, K.K. & Park, Y.H. 2000. Design potential method for robust system parameter design. AIAA Journal 39(4):667–677. Uang, C.M. & Bertero, V.V. 1990. Evaluation of seismic energy in structures. Earthquake Engineering and Structural Dynamics 19:77–90. Uryasev, S. 1995. Derivatives of probability functions and some applications. Annals of Operations Research 56:287–311. Van Noortwijk, J.M. 2001. Cost-based criteria for obtaining optimal design decisions. In Proceedings of the 8th International Conference on Structural Safety and Reliability, Newport Beach, CA, U.S.A., June. Vandenberghe, L. & Boyd, S. 1996. Semidefinite programming. SIAM Review 38:49–95. Volker, A.W.F., Dijkstra, F.H., Heerings, J.H.A.M. & Terpstra, S. 2004. Modeling of NDE Reliability; Development of A POD-Generator. 16th WCNDT 2004 – World Conference on NDT.

628

References

von Neumann, J. & Morgenstern, O. 1943. Theory of Games and Economical Behavior. Princeton University Press. Wall, M. 1995. GAlib:. A C++ library of genetic algorithm components. Available at http://lancet.mit.edu/ga. Wang, C.K. 1986. Structural analysis on microcomputers. New York, NY: Macmillan. Wang, G. 2003. Adaptive Response Surface Method Using Inherited Latin Hypercube Design Points. ASME Journal of Mechanical Design 125:1–11. Wang, L. & Grandhi R.V. 1994. Efficient safety index calculation for structural reliability analysis. Comput. Struct. 52(1):103–111. Wen, Y.K. 2000. Reliability and performance-based design. Proceedings of the 8th ASCE Specialty Conference on Probabilistic Mechanics and Structural Reliability, University of Notre Dame, Indiana, USA, 24–26 July. Westermo, B.D. 1985. The critical excitation and response of simple dynamic systems. Journal of Sound and Vibration 100(2):233–242. White, D.J. 1996. A heuristic approach to a weighted maxmin dispersion problem. IMA Journal of Mathematics Applied in Business and Industry 7:219–231. White, P., Barter, S. & Molent, L. 2002. Probabilistic Fracture Prediction Based On Aircraft Specific Fatigue Test Data. 6th Joint FAA/DoD/NASA Aging Aircraft Conference. Willcox, K. & Peraire, J. 2001. Balanced model reduction via the proper orthogonal decomposition. In 15th AIAA Computational Fluid Dynamics Conference, June 11–14, Anaheim, CA, AIAA 2001–2611. Wittwer, J.W., Baker, M.S. & Howell, L.L. 2006. Robust design and model validation of nonlinear compliant micromechanisms. J. Microelectromechanical Sys. 15(1). To appear. Wojtkiewicz, S.F. Jr., Eldred, M.S., Field, R.V. Jr., Urbina, A. & Red-Horse, J.R. 2001. A toolkit for uncertainty quanti_cation in large computational engineering models. In Proceedings of the 42rd AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, Number AIAA-2001-1455, April 16–19, Seattle, WA. Wolfram, S. 2002. A New Kind of Science. Wolfram Media. Wolkowicz, H., Saigal, R. & Vandenberghe, L. (eds) 2000. Handbook of Semidefinite Programming – Theory, Algorithms, and Applications. Dordrecht: Kluwer. Wu, Y.-T. 1994. Computational methods for efficient structural reliability and reliability sensitivity analysis. AIAA Journal 32(8):1717–1723. Wu, Y.-T. & Wirsching, P.H. 1987. A new algorithm for structural reliability estimation. J. Eng. Mech. ASCE 113:1319–1336. Wu, Y.-T., Shin, Y., Sues, R. & Cesare, M. 2006. Probabilistic Function Evaluation System (ProFES) for Reliability-Based Design. Journal of Structural Safety 28(1–2):164–195. Wu, Y.-T., Millwater, H.R. & Cruse, T.A. 1990. An Advanced Probabilistic Structural Analysis Method for Implicit Performance Functions. AIAA Journal 28(9):1663–1669. Wu, Y.-T. & Mohanty, S. 2006. Variable Screening and Ranking Using Several Sampling Based Sensitivity Measures. Journal of Reliability Engineering & System Safety 91(6):634–647. Wu, Y.-T., Shiao, M., Shin, Y. & Stroud, W.J. 2005. Reliability-Based Damage Tolerance Methodology for Rotorcraft Structures. Transactions Journal of Materials and Manufacturing. Wu, Y.-T. & Shin, Y. 2004. Probabilistic Damage Tolerance Methodology For Reliability Design And Inspection Optimization. Proceedings of the AIAA 45th SDM Conference, Paper 2004-01-0681. Wu, Y.-T. & Shin, Y. 2005. Probabilistic Function Evaluation System for Maintenance Optimization. Proceedings of the AIAA 46th SDM Conference. Wu, Y.T., Shin, Y., Sues, R. & Cesare, M. 2001. Safety factor based approach for probabilitybased design optimization. In Proceedings of the 42nd AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, Seattle, WA, USA, Paper no AIAA 2001-1522.

References

629

Wu, Y-T., Enright, M.P. & Millwater, H.R. 2002. Probabilistic Methods for Design Assessment of Reliability With Inspection. AIAA Journal 40(5):937–946. Xiao, Q., Sues, R. & Rhodes, G. 1999. Computational strategies for reliability based Multidisciplinary optimization. Proceedings of the 13th ASCE EMD Conference. Xie, Y.M. & Stevens, G. 1997. Evolutionary Structural Optimization. Springer-Verlag, London. Xiu, D., Lucor, D., Su, C.-H. & Karniadakis, G. 2002. Stochastic modelling of flow-structure interactions using generalized polynomial chaos. Journal of Fluids Engineering 124:51–59. Xu, S. & Grandhi, R.V. 1998. Effective two-point function approximation for design optimization. AIAA J. 36(12):2269–2275. Yager, R.R., Fedrizzi, M. & Kacprzyk, J. (eds) 1994. Advances in the Dempster – Shafer Theory of Evidence. John Wiley & Sons, Inc. Yang, R.J., Chuang, C., Gu, L. & Li, G. 2005. Experience with approximate reliabilitybased optimization methods II: an exhaust system problem. Structural and Multidisciplinary Optimization 29:488–497. Yang, S.-I., Frangopol, D.M., Kawakami, Y. & Neves, L.C. 2006. The use of lifetime functions in the optimization of interventions on existing bridges considering maintenance and failure costs. Reliability Engineering & System Safety 96(6):698–705. Yang, S.-I., Frangopol, D.M. & Neves, L.C. 2006. Optimum maintenance strategy for deteriorating structures using lifetime functions. Engineering Structures 28(2):196–206. Ye, K.Q., Li, W. & Sudjianto, A. 2000. Algorithmic Construction of Optimal Symmetric Latin Hypercube Designs. Journal of Statistical Planning and Inference 90:145–159. Yi, P., Cheng, G. & Jiang, L. 2006. A sequential approximate programming strategy for performance-measure based probabilistic structural design optimization. Structural Safety. Article in Press. Youn, B.D., Kokkolaras, M., Mourelatos, Z.P., Papalambros, P.Y., Choi, K.K. & Gorsich, D. 2004. Uncertainty propagation techniques for probabilistic design of multilevel systems. In Proceedings of the 10th AIAA/ISSMO Symposium on Multidisciplinary Analysis and Optimization, Albany, New York, Paper No. AIAA-2004-4470. Youn, B.D. & Choi, K.K. 2003. Hybrid Analysis Method for Reliability-Based Design Optimization. ASME Journal of Mechanical Design 125(2):221–232. Youn, B.D. & Choi, K.K. 2004. A new response surface methodology for reliability-based design optimization. Computers and Structures 82:241–256. Youn, B.D. & Choi, K.K. 2004. Selecting probabilistic approaches for reliability-based design optimization. AIAA Journal 42(1):124–131. Youn, B.D. & Choi, K.K. 2004. The performance moment integration method for reliabilitybased robust design optimization. Proceedings of the ASME Design Engineering Technical Conference, Salt Lake City, Utah, USA, September 28–October 2. Youn, B.D., Choi, K.K. & Du, L. 2005. Enriched performance measure approach for reliability-based design optimization. AIAA J. 43(4):874–884. Youn, B.D., Choi, K.K. & Park, Y.H. 2001. Hybrid Analysis Method for Reliability-Based Design Optimization. ASME Journal of Mechanical Design 125(2):221–232. Yu, X., Chang, K. & Choi, K. 1998. Probabilistic structural durability prediction. AIAA Journal 36(4):628–637. Yu, X., Choi, K. & Chang, K. 1997. A mixed design approach for probabilistic structural durability. Journal of Structural Optimization 14(2–3):81–90. Yu, X., Choi, K. & Chang, K. 1997. Reliability and durability based design sensitivity analysis and optimization. Technical Report R97-01, Center for Computer Aided Design University of Iowa. Yuge, K. & Kikuchi, N. 1995. Optimization of a frame structure subjected to a plastic deformation. Struct. Optim. 10:197–208. Yuge, K., Iwai, N. & Kikuchi, N. 1999. Optimization of 2-d structures subjected to nonlinear deformations using the homogenization method. Struct. Optim. 17(4):286–299.

630

References

Zadeh, L.A. 1965. Fuzzy Sets. Information and Control 8:338–353. Zadeh, L.A. 1978. Fuzzy Sets as a Basis for a Theory of Possibility. Fuzzy Sets and Systems 1:3–28. Zhang, W.-H., Beckers, P. & Fleury, C. 1995. A unified parametric design approach to structural shape optimization. International Journal for Numerical Methods in Engineering 38:2283–2292. Zhao, Y.G. & Ono, T. 2001. Moment methods for structural reliability. Structural Safety 23:47–75. Zhou, J. & Mourelatos, Z.P. 2007. A Sequential Algorithm for Possibility-Based Design Optimization. In press ASME Journal of Mechanical Design. Also, Proceedings of ASME Design Engineering Technical Conferences, 2006, Paper# DETC2006-99232. Zhou, M. & Rozvany, G.I.N. 1991. The COC algorithm, Part II: Topological, geometrical and generalized shape optimization. Comp. Meth. Appl. Mech. Engrg. 89:309–336. Zou, T., Mahadevan, S. & Sopory, A. 2004. A reliability-based design method using simulation techniques and efficient optimization approach. ASME Design Engineering Technical Conferences, Salt Lake City, Utah, DETC2004/DAC-57457. Zou, T., Mahadevan, S. & Rebba, R. 2004. Computational efficiency in reliability-based optimization. In Proceedings of the 9th ASCE Specialty Conference on Probabilistic Mechanics and Structural Reliability, July 26–28, Albuquerque, NM.

Author index

Adams, B.M. 401 Agarwal, H. 281 Allen, M. 135 Aoues, Y. 217

Huh, J.S. 57 Hurtado, J.E. 435

Beck, J.L. 155 Ben-Haim, Y. 531 Bichon, B.J. 401

Kanno, Y. 471 Kharmanda, G. 189 Kokkolaras, M. 115 Kwak, B.M. 57

Chateauneuf, A. 3, 217 De Palma, P. 549 Doltsinis, I. 499 Eldred, M.S. 401 Fragiadakis, M. 567 Frangopol, D.M. 135 Ganzerli, S. 549

Patel, N.M. 281 Plevris, V. 567 Polak, E. 307

Joanni, A.E. 335

Lagaros, N.D. 567 Lee, S.H. 57 Liang, J. 87 Mahadevan, S. 401 Maute, K. 135 Mourelatos, Z.P. 87, 247 Nikolaidis, E. 87 Papadrakakis, M. 567 Papalambros, P.Y. 115

Renaud, J.E. 281 Royset, J.O. 307 Rackwitz, R. 335 Sørensen, J.D. 31 Taflanidis, A.A. 155 Takewaki, I. 471, 531 Tillotson, D. 281 Tovar, A. 281 Tsompanakis, Y. 567 Weickum, G. 135 Wu, Y.-T. 369 Zhou, J. 247

Subject index

Advanced Mean Value (AMV) 402, 406 Age-dependent Repairs 350 AMV+ 402, 406–407 AMV2 + 402, 406–407 Analytical Target Cascading 117, 121 Bi-level RBDO 413 Block repairs 353 Cascade Evolutionary Algorithm (CEA) 573, 590 Cellular automata 290 Chloride attack 362 Combined approximation 142 Common Random Numbers 168, 171, 180 Convergence 221, 223 Convex model 472, 549, 551, 554, 560 Cost function 5, 22 Cost-benefit optimization 34, 44, 343 Critical excitation 532, 538 DAKOTA 403 Decision theory 32 Decomposition-based Design 115 Design of Experiments 59 Desirability function 500, 509 Deterioration failure models 336–337 Deterministic Design Optimization 189, 199, 233 Earthquake input energy 534, 536, 546 Efficient global reliability analysis (EGRA) 411, 419 Elasticity 512 Entropy 444, 459 Evidence theory 248, 261 Evidence-based Design Optimization 261 Evolution Strategies (ES) 572

Failure-Region Sampling 378, 381 First-order reliability method (FORM) 146, 402, 408 Genetic Algorithms 554 Global reliability methods 411, 419 Hybrid Cellular Automaton method 282, 290–291 Importance Sampling 174, 180 Info-gap theory 471, 475, 532, 539, 546 Inspection and repair 349, 354 Inspection optimization 371, 386 Interval Analysis 127 Large displacements 514, 526 Latin Hypercube Sampling (LHS) 575 Life-cycle Cost 156, 175 Limit state 574 Maintenance Optimization 369 Maintenance strategies 348 Markov Chain Monte Carlo 381 Mean-Value First-Order Second-Moment (MVFOSM) 403 Microelectromechanical systems (MEMS) 422 Model Prediction 157, 159, 170 Moment estimation 448, 450 Monte Carlo Simulation (MCS) 307, 314, 574 Most Probable Failure Point 8, 15, 219, 222, 226 Multi-objective optimization 573 Nataf transformation 405 Neural networks (NN) 576 Nonlinear dynamics 516

634

Subject index

Optimal inspection 36 Optimization 31, 36 Optimization methods 359 Optimization under uncertainty 116 Optimum Safety Factor 191, 199, 209 Pareto optima 509, 526, 573 Passively controlled structure 532, 542, 547 Pearson System 64 Performance measure approach (PMA) 402, 405 Performance-based Earthquake Engineering (PBEE) 569 Plasticity 517 Poissonian disturbances 347 Polynomial Chaos expansion 147 Possibility theory 248, 253 Possibility-based Design Optimization 249, 258 Probabilistic Damage Tolerance 370 Probabilistic Design 124 Probabilistic Ground Motion Model 176 Probabilistic transformation 221, 223, 226 Probability estimation 441 Probability of detection 371, 387 Quasiconvex optimization 473, 487 Reduced order model 137–138 Reliability 89–90 Reliability analysis 191–192 Reliability index 6, 18, 22 Reliability index approach (RIA) 402, 405 Reliability methods 403 Reliability versus robustness 509 Reliability-Based Design Optimization (RBDO) 34, 76, 149, 189, 199, 233, 283, 309, 412, 570 Reliability-Based Robust Design Optimization (RRDO) 571, 589 Reliability-Based Topology Optimization 282, 296 Renewal model 340, 344 Response Surface Method 68 Risk Sensitivity Analysis 386 Robust analysis 455, 459

Robust Design Optimization (RDO) 439, 507, 509, 571 Robustness function 471, 477, 485, 490, 533, 541 Safety factor 3, 10, 26, 226, 228 Sample average approximations 314 Sample-adjustment rules 308, 315–316 Second-order reliability method (SORM) 402, 408 Semidefinite programming 471, 473, 490 Sensitivity analysis 42, 139, 413 Sequential Optimization and Reliability Assessment (SORA) 92 Sequential PBDO 265 Sequential RBDO 415 Series systems 339, 357 Simultaneous-perturbation Stochastic Approximation 171, 180 Single-Loop RBDO approach 95 Statistical Moments 58–59, 76 Stochastic constraints 504, 509 Stochastic Design 157, 171 Stochastic Optimization 499 Stochastic Simulation 157, 173 Stochastic Subset Optimization 159, 178 Structural Dynamics 138 System Reliability 9, 21, 156, 159, 230–231 System Reliability-based Design Optimization 89 Target reliability 221, 229 Taylor series approximation 499, 503 Topology Optimization 286 Two-point adaptive nonlinearity approximation (TANA) 402, 406–407 Two-Stage Importance Sampling (TIS) 377 Uncertainty analysis 144 Uncertainty Propagation 122 Unified RDO-RRDO 456 Updating 342 Wind turbines 48