Intelligent Computing and Information Science, Part II

Communications in Computer and Information Science

135

Ran Chen (Ed.)

Intelligent Computing and Information Science International Conference, ICICIS 2011 Chongqing, China, January 8-9, 2011 Proceedings, Part II

13

Volume Editor Ran Chen The Key Laboratory of Manufacture and Test Chongqing University of Technology Chongqing, 400054, P. R. China E-mail: [email protected]

Library of Congress Control Number: Applied for CR Subject Classification (1998): I.2, H.4, H.3, C.2, H.5, D.2 ISSN ISBN-10 ISBN-13

1865-0929 3-642-18133-3 Springer Berlin Heidelberg New York 978-3-642-18133-7 Springer Berlin Heidelberg New York

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting, reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. springer.com © Springer-Verlag Berlin Heidelberg 2011 Printed in Germany Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India Printed on acid-free paper 06/3180

Preface

The 2011 International Conference on Intelligent Computing and Information Science (ICICIS 2011) was held in Chongqing, China, during January 8–9, 2011. The aim of the conference series is to provide a platform for researchers, engineers, academicians as well as industrial professionals from all over the world to present their research results and development activities in intelligent computing and information science. This two-volume set of CCIS 134 and CCIS 135 communicates the latest progress and research results in the theory, methods, and technology in the field of intelligent computing and information science if it also presents an update on the international trends, which will drive the global communication and cooperation in production, education and research in this field. We received more than 600 submissions which were reviewed by international experts, and 230 papers were selected for presentation. We believe the proceedings provide the reader with a broad overview of the latest advances in the field of intelligent computing and information science. On behalf of the guest editors for this special issue, I would like to thank the National 863 Program of China and the National Science Fund of China. I also thank the conference organization staff and the members of the International Technical Committees for their hard work. We look forward to seeing all of you next year at ICICIS 2012.

October 2010 Ran Chen

ICICIS 2011 Committee

Conference Chairman Ran Chen

Chongqing University of Technology, China

Publication Chair Wenli Yao

Control Engineering and Information Science Research Association, Hong Kong International Frontiers of Science and Technology Research Association

International Technical Committees Peifu Chen Lin Mi Viranjay M. Srivastava Liu Yunan Mir Mahdi Zalloi Wang Liying Zhou Liang Chenggui Zhao Rahim Jamian Li-Xin GUO Wen-Sheng Ou Hsiang-Chuan Liu Mostafa Shokshok Ramezan ali Mahdavinejad Wei Fu Anita Kova Kralj Tjamme Wiegers Gang Shi Zengtao Chen Bhagavathi Tarigoppula

Chongqing Aerospace Polytechnic College, China Chongqing University of Technology, China Jaypee University of Information Technology, Solan, India University of Michigan, USA Iran Institute of Water Conservancy and Hydroelectric Power, China Donghua University, China Yunnan University of Finance and Economics, China Universiti Kuala Lumpur Malaysian Spanish Institute, Malaysia Northeastern University, China National Chin-Yi University of Technology, Taiwan R.O.C. Asia University, Japan National University of Malaysia, Malaysia University of Tehran, Iran Chongqing University, China University of Maribor, Slovenia Delft University of Technology, The Netherlands Inha University, South Korea University of New Brunswick, Canada Bradley University, USA

VIII

Organization

Co-Sponsored by Control Engineering and Information Science Research Association International Frontiers of science and technology Research Association Chongqing Xueya Conferences Catering Co., Ltd Chongqing University of Technology

Table of Contents – Part II

Unified Electronic Currency Based on the Fourth Party Platform Integrated Payment Service . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xu Yong and Hu Qiqi

1

Modeling and Power Flow Analysis for Herringbone Gears Power Dual-Branching Transmission System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xiaofang Yang, Yanxiang Zhu, Zongde Fang, and Jiangong Gu

7

Research on SaaS and Web Service Based Order Tracking . . . . . . . . . . . . . Jianhua Jiang, Buyun Sheng, Lixiong Gong, and Mingzhong Yang

16

Research and Analyses on Unsteady Heat Transfer of Inner Thermal Insulation Wall During Multi-temperature Refrigerated Transportation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Guanghai Liu, Ruhe Xie, and Yongcai Sun Evaluation of the Industrial Economic Benefits Based on TOPSIS . . . . . . Baolu Wei, Feng Dai, and Jingxu Liu Optimization and Distributing Research of Measuring Points for Thermal Error of CNC Machine Based on Weighted Grey Relative Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Qin Wu, Jianjun Yang, Zhiyuan Rui, and Fuqiang Wang A DS-UWB Cognitive Radio System Based on Bridge Function Smart Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yafei Xu, Sheng Hong, Guodong Zhao, Fengyuan Zhang, Jinshan Di, and Qishan Zhang Problems and Countermeasures of Zhejiang High-Tech Enterprises Industry-University-Institute Cooperation in China . . . . . . . . . . . . . . . . . . . Qing Zhou, Chong-Feng Mao, and Lin Hou Research on Puncture Area Calibration of Image Navigation for Radio Frequency Ablation Robot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Peng Wang, Wei Cao, Wenhao Jiang, Jinglei Xin, Shaochen Kang, and Xin Li Low Power and Robust Domino Circuit with Process Variations Tolerance for High Speed Digital Signal Processing . . . . . . . . . . . . . . . . . . . Jinhui Wang, Xiaohong Peng, Xinxin Li, Ligang Hou, and Wuchen Wu

23 30

35

41

47

53

59

X

Table of Contents – Part II

Detection of Attention-to-Rest Transition from EEG Signals with the Help of Empirical Mode Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Cheng Man Ng and Mang I. Vai

66

A Traffic Information Estimation Model Using Periodic Location Update Events from Cellular Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bon-Yeh Lin, Chi-Hua Chen, and Chi-Chun Lo

72

The Application of Virtual Reality on Distance Education . . . . . . . . . . . . . Zehui Zhan

78

Framework Design of Unified Cross-Authentication Based on the Fourth Platform Integrated Payment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xu Yong and He Yujin

84

Analysis towards VMEM File of a Suspended Virtual Machine . . . . . . . . . Zheng Song, Bo Jin, and Yongqing Sun

89

Optimization and Reconfiguration of Advanced Manufacturing Mode Based on Object-Based Knowledge Mesh and Improved Immune Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chaogai Xue and Haiwang Cao Modeling and Simulation of Water Allocation System Based on Simulated Annealing Hybrid Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . Jiulong Zhu and Shijun Wang Study on Feed-Forward MAP-Based Vector Control Method of Vehicle Drive Motor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yafu Zhou, Xiaoyong Shen, Jing Lian, Xinhan Sun, Jun Li, Minghui Liu, and Ziliang Zhao

98

104

110

Condition Monitoring and Fault Diagnosis of Wet-Shift Clutch Transmission Based on Multi-technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . Man Chen, Liyong Wang, and Biao Ma

116

Circulant Graph Modeling Deterministic Small-World Networks . . . . . . . . Chenggui Zhao

124

Research on Risk Manage of Power Construction Project Based on Bayesian Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Zhengyuan Jia, Zhou Fan, and Yong Li

128

The Design of Logistics Information Matching Platform for Highway Transportation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Daqiang Chen, Xiaoxiao Zhu, Bing Tong, Xiahong Shen, and Tao Feng

135

Table of Contents – Part II

XI

An Intelligent Prediction Method Based on Information Entropy Weighted Elman Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tao Chen, Xiao-li Xu, and Shao-hong Wang

142

A Multi-layer Dynamic Model for Coordination Based Group Decision Making in Water Resource Allocation and Scheduling . . . . . . . . . . . . . . . . . Wei Huang, Xingnan Zhang, Chenming Li, and Jianying Wang

148

Analysis of Mode Choice Performance among Heterogeneous Tourists to Expo Shanghai 2010 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Shengchuan Jiang, Yuchuan Du, and Lijun Sun

154

Hybrid Schema Matching for Deep Web . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Kerui Chen, Wanli Zuo, Fengling He, and Yongheng Chen

165

Study on DS/FH Technology Used in TT&C System . . . . . . . . . . . . . . . . . Chang Xiaoming, Zhang Xiaolin, and Huang Zhengjing

171

RFID-Based Critical Path Expert System for Agility Manufacture Process Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Haifang Cheng and Yuli Xiang

177

Research on the Evolutionary Strategy Based on AIS and Its Application on Numerical Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Li Bei

183

Prediction of the NOx Emissions from Thermal Power Plant Based on Support Vector Machine Optimized by Chaos Optimization Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jingmin Wang, Junjie Kang, and Huaitao Liang

189

Optimizing of Bioreactor Heat Supply and Material Feeding by Numerical Calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Zhiwei Zhou, Boyan Song, Likuan Zhu, Zuntao Li, and Yang Wang

195

Research on Technique of the Cyberspace Public Opinion Detection and Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yongping Du, Jiangli Liu, and Ming He

203

A High Performance 50% Clock Duty Cycle Regulator . . . . . . . . . . . . . . . . Peng Huang, Hong-Hui Deng, and Yong-Sheng Yin

208

Study on Risk of Enterprise’ Technology Innovation Based on ISM . . . . . Hongyan Li

215

Study on Reservoir Group Dispatch of Eastern Route of South-to-North Water Transfer Project Based on Network . . . . . . . . . . . . . . . . . . . . . . . . . . . Hongyan Li

221

XII

Table of Contents – Part II

Immune Clone Algorithm to Solve the Multi-object Problems . . . . . . . . . . Liang Zhou and Jianguo Zheng

227

Foreign Language Teachers’ Professional Development in Information Age . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xiying Fan and Gang Wu

233

Active Learning Framework Combining Semi-supervised Approach for Data Stream Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mahnoosh Kholghi and MohammadReza Keyvanpour

238

Sensorless Vector Control of the Charging Process for Flywheel Battery with Artificial Neural Network Observer . . . . . . . . . . . . . . . . . . . . . . . . . . . . Honglin Qin, Meng Huang, Zhixiong Li, and Shuangqing Tang

244

Framework for Classifying Website Content Based on Folksonomy in Social Bookmarking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Shih-Ming Pi, Hsiu-Li Liao, Su-Houn Liu, and Chen-Wen Lin

250

Research on Internal Damping Algorithm of Marine Inertial Navigation System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Li Kui, Liu Fang, and Xu Yefeng

256

Design and Implementation of Process Migrating among Multiple Virtual Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Si Shen, Zexian Zhang, Shuangxi Yang, Ruilin Guo, and Murong Jiang

262

The Adaptability Evaluation of Enterprise Information Systems . . . . . . . . Junjuan Liu, Chaogai Xue, and Lili Dong

268

Structural Damage Alarm Utilizing Modified Back-Propagation Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xiaoma Dong

273

Computation of Virtual Regions for Constrained Hybrid Systems . . . . . . Jianqiang Li, Zhen Ji, and Hai-long Pei

279

Fault Diagnosis of Diesel Engine Using Vibration Signals . . . . . . . . . . . . . . Fengli Wang and Shulin Duan

285

Influences on the Directivity of Acoustic Vector Sensor by Soft Finite Cylinder Baffle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ji Jianfei, Liang Guolong, Zhang Guangpu, and Li Yang

291

The Method of Intervenient Optimum Decision Based on Uncertainty Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Lihua Duan

296

Table of Contents – Part II

The Deflection Identify of the Oil Storage Tank . . . . . . . . . . . . . . . . . . . . . . Jingben Yin, Hongwei Jiao, Jiemin Zhang, Kaina Wang, and Jiahui Ma

XIII

302

PID Control of Miniature Unmanned Helicopter Yaw System Based on RBF Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yue Pan, Ping Song, and Kejie Li

308

Identity-Based Inter-domain Authentication Scheme in Pervasive Computing Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Shi-Wei Huo, Chang-Yuan Luo, and Hong-Zhi Xin

314

Computer Simulation of Blast Wall Protection under Methane-Air Explosion on an Offshore Platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Changjian Wang, Weigang Yan, Jin Guo, and Changming Guo

321

Throughput Analysis of Discrete-Time Non-persistent CSMA with Monitoring in Internet of Things . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hongwei Ding, Dongfeng Zhao, and Yifan Zhao

327

The Effect of Product Placement Marketing on Effectiveness of Internet Advertising . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hsiu-Li Liao, Su-Houn Liu, Shih-Ming Pi, and Hui-Ju Chen

332

A Modular Approach to Arithmetic and Logic Unit Design on a Reconfigurable Hardware Platform for Educational Purpose . . . . . . . . . . . Halit Oztekin, Feyzullah Temurtas, and Ali Gulbag

338

A Variance Based Active Learning Approach for Named Entity Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hamed Hassanzadeh and MohammadReza Keyvanpour

347

Research on Routing Selection Algorithm Based on Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Guohong Gao, Baojian Zhang, Xueyong Li, and Jinna Lv

353

Optimization of the Performance Face Recognition Using AdaBoost-Based . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mohsen Faghani, Md Jan Nordin, and Shahed Shojaeipour

359

Design and Implementation Issues of Parallel Vector Quantization in FPGA for Real Time Image Compression . . . . . . . . . . . . . . . . . . . . . . . . . . . Krupa R. Rasane and Srinivasa Rao R. Kunte

366

Discriminative Novel Information Detection of Query-Focused Update Summarization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jinguang Chen and Tingting He

372

XIV

Table of Contents – Part II

Visualization of Field Distribution of the Circular Area Based on the Green Function Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Gang Wu and Xiying Fan Efficient Genetic Algorithm for Flexible Job-Shop Scheduling Problem Using Minimise Makespan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hamid Ghaani Farashahi, B.T.H.T. Baharudin, Shahed Shojaeipour, and Mohammad Jaberi

378

385

Core Image Coding Based on WP-EBCOT . . . . . . . . . . . . . . . . . . . . . . . . . . Yongdan Nie, Yan Zhang, and Jinghui Li

393

A New Method of Facial Expression Recognition Based on SPE Plus SVM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Zilu Ying, Mingwei Huang, Zhen Wang, and Zhewei Wang

399

Multiple Unmanned Air Vehicles Control Using Neurobiologically Inspired Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yong Zhang and Li Wang

405

The Use of BS7799 Information Security Standard to Construct Mechanisms for the Management of Medical Organization Information Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Shu-Fan Liu, Hao-En Chueh, and Kuo-Hsiung Liao

411

An Improved Frame Layer Rate Control Algorithm for H.264 . . . . . . . . . . Xiao Chen and Feifei Lu

417

Some Properties in Hexagonal Torus as Cayley Graph . . . . . . . . . . . . . . . . Zhen Zhang

422

Modeling Software Component Based on Extended Colored Petri Net . . . Yong Yu, Tong Li, Qing Liu, and Fei Dai

429

Measurement of Software Coupling Based on Structure Entropy . . . . . . . . Yong Yu, Tong Li, Qing Liu, and Qian Yu

435

A 3D Grid Model for the Corridor Alignment . . . . . . . . . . . . . . . . . . . . . . . . Kun Miao and Liang Li

440

The Consumers’ Decisions with Different Delay Cost in Online Dual Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Shengli Chen

445

Fuzzy Control for the Swing-Up of the Inverted Pendulum System . . . . . Yu Wu and Peiyi Zhu

454

A Novel OD Estimation Method Based on Automatic Vehicle Identification Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jian Sun and Yu Feng

461

Table of Contents – Part II

Stress Field in the Rail-End during the Quenching Process . . . . . . . . . . . . Siqiang Xu, Jianyi Kong, Gongfa Li, Jintang Yang, Hegen Xiong, and Guozhang Jiang

XV

471

A Novel Adaptive Target Tracking Algorithm in Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xuewen Wu, Guan Huang, Dunye Tang, and Xinhong Qian

477

Computer Management of Golden Section Effect of Physical Pendulum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jia Jing, Hui Lin, Caixia Liu, and Ying Zhu

487

Simulation Study of Gyrotron Traveling Wave Amplifier with Distributed-Loss in Facilities for Aquaculture . . . . . . . . . . . . . . . . . . . . . . . . Xufeng Hua, Chengxun Chen, Dawei Xu, and Kezhi Xing

492

The Worm Propagation Model with Dual Dynamic Quarantine Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yu Yao, Xiao-wu Xie, Hao Guo, Fu-xiang Gao, and Ge Yu

497

Keyword Extraction Algorithm Based on Principal Component Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chang-Jin Li and Hui-Jian Han

503

A Web Information Retrieval System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tae-Hyun Kim, Dong-Chul Park, Woong Huh, Hyen-Ug Kim, Chung-Hwa Yoon, Chong-Dae Park, Dong-Min Woo, Taikyeong Jeong, Il-Hwan Cho, and Yunsik Lee

509

An Effective Intrusion Detection Algorithm Based on Improved Semi-supervised Fuzzy Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xueyong Li, Baojian Zhang, Jiaxia Sun, and Shitao Yan

515

What Does Industry Really Want in a Knowledge Management System? A Longitudinal Study of Taiwanese Case . . . . . . . . . . . . . . . . . . . . Liang-Chih Yang and Hsi-Peng Lu

521

A Process Positioning System for Sheet-Fed Offset Press . . . . . . . . . . . . . . Li E. Ma, Hai Yan Zhang, and Wei Li

532

Research of Electronic Image Stabilization Algorithm Based on Orbital Character . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xiaodong Xian, Peipei Hou, Shan Liang, and Ping Gan

538

Study on USB Based CAN Bus for Data Measurement System . . . . . . . . . Weibin Wu, Tiansheng Hong, Yuqing Zhu, Guangbin He, Cheng Ye, Haobiao Li, and Chuwen Chen

544

XVI

Table of Contents – Part II

Research of WLAN CAN Bus Data Test System . . . . . . . . . . . . . . . . . . . . . Weibin Wu, Yuqing Zhu, Tiansheng Hong, Cheng Ye, Zhijie Ye, Guangbin He, and Haobiao Li

550

A Framework of Simple Event Detection in Surveillance Video . . . . . . . . . Weiguang Xu, Yafei Zhang, Jianjiang Lu, Yulong Tian, and Jiabao Wang

556

ESDDM: A Software Evolution Process Model Based on Evolution Behavior Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Na Zhao, Jian Wang, Tong Li, Yong Yu, Fei Dai, and Zhongwen Xie

562

A Loading Device for Intervertebral Disc Tissue Engineering . . . . . . . . . . Lihui Fu, Chunqiu Zhang, Baoshan Xu, Dong Xin, and Jiang Li

568

The Optimization of Spectrum Sensing Frame Time . . . . . . . . . . . . . . . . . . Dongmei Shu, Jinkuan Wang, Xin Song, and Bin Wang

574

A Novel Model for the Mass Transfer of Articular Cartilage: Rolling Depression Load Device . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Zhenmin Fan, Chunqiu Zhang, Haiying Liu, Baoshan Xu, Jiang Li, and Lilan Gao

580

Sensor Scheduling Target Tracking-Oriented Cluster-Based . . . . . . . . . . . . Dongmei Yan, JinKuan Wang, Li Liu, and Bin Wang

586

Improved Stack Algorithm for MIMO Wireless Communication Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Li Liu, Jinkuan Wang, Dongmei Yan, Ruiyan Du, and Bin Wang

592

The Cooperative Hunting Research of Mobile Wireless Sensor Network Based on Improved Dynamic Alliance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Daming Pei, Ping Song, Xiaobing Han, and Kejie Li

599

Identification of Matra Region and Overlapping Characters for OCR of Printed Bengali Scripts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Subhra Sundar Goswami

606

New Energy Listed Companies Competitiveness Evaluation Based on Modified Data Envelopment Analysis Model . . . . . . . . . . . . . . . . . . . . . . . . . Chong Gao, Zhou Fan, and Jian-ze Zhang

613

A SMA Actuated Earthworm-Like Robot . . . . . . . . . . . . . . . . . . . . . . . . . . . Y.K. Wang, C.N. Song, Z.L. Wang, C. Guo, and Q.Y. Tan

619

Design of Instruction Execution Stage for an Embedded Real-Time Java Processor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Guang Hu, Zhilei Chai, and Wenke Zhao

625

Table of Contents – Part II

XVII

A Survey of Enterprise Architecture Analysis Using Multi Criteria Decision Making Models (MCDM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mehmooda Jabeen Zia, Farooque Azam, and Maria Allauddin

631

Research of Manufacture Time Management System Based on PLM . . . . Ni Jing, Zhu Juan, and Zhong Liangwei

638

A Non-rigid Registration Method for Dynamic Contrast Enhancement Breast MRI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yangping Wang, Jianwu Dang, Xiaogang Du, and Sha Li

644

Iterative Algorithms for Nonexpansive Semigroups with Generalized Contraction Mappings in Banach Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Qiang Lin

650

Water Distribution System Optimization Using Genetic Simulated Annealing Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Shihu. Shu

656

Distributed Workflow Service Composition Based on CTR Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Zhilin Feng and Yanming Ye

662

Offline Optimization of Plug-In Hybrid Electric Vehicle Energy Management Strategy Based on the Dynamic Programming . . . . . . . . . . . Shichun Yang, Ming Li, Haigang Cui, Yaoguang Cao, Gang Wang, and Qiang Lei

668

Development of Field Information Monitoring System Based on the Internet of Things . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ken Cai, Xiaoying Liang, and Keqiang Wang

675

System Design of Real Time Vehicle Type Recognition Based on Video for Windows (AVI) Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Wei Zhan and Zhiqing Luo

681

Subjective Assessment of Women’s Pants’ Fit Analysis Using Live and 3D Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Linli Zhang, Weiyuan Zhang, and Hong Xiao

687

A Geographic Location-Based Security Mechanism for Intelligent Vehicular Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Gongjun Yan, Jingli Lin, Danda B. Rawat, and Weiming Yang

693

XVIII

Table of Contents – Part II

Intrusion-Tolerant Location Information Services in Intelligent Vehicular Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Gongjun Yan, Weiming Yang, Earl F. Shaner, and Danda B. Rawat

699

The Role of Network and Mobility Simulators in Evaluating Vehicular Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Gongjun Yan, Jingli Lin, Danda Rawat, and Justin C. Enyart

706

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

713

Table of Contents – Part I Some Analysis and Research of the AdaBoost Algorithm . . . . . . . . . . . . . . Peng Wu and Hui Zhao

1

Simulation the Heat Flow Field in Cooler of Rotary Calcining Kiln . . . . . Yonggang Liu, Jishun Li, Xianzhao Jia, Weimin Pan, and Zhiming Liao

6

Teaching Platform Scheme of Engineering Testing Curriculums Based on LabVIEW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tingrui Liu, Xiuhua Sui, and Huibin Liang

12

Software Development Cost and Time Forecasting Using a High Performance Artificial Neural Network Model . . . . . . . . . . . . . . . . . . . . . . . . Iman Attarzadeh and Siew Hock Ow

18

A Weighted Combination Method of Target Identity Identification . . . . . Jin Hongbin, Lan Jiangqiao, and Li Hongfei

27

Fault Diagnosis of Power Transformer Based on BP Combined with Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Weiguo Zhao, Yanning Kang, Gangzhu Pan, and Xinfeng Huang

33

The Complexities of Implementing Cluster Supply Chain – Case Study of JCH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xiao Xue, Jibiao Zhang, and Yang Wang

39

Fault Diagnosis of Rolling Bearing Based on Lyapunov Exponents . . . . . . Liying Wang, Huang Meng, and Yanning Kang

45

New Encryption Scheme of One-Time Pad Based on KDC . . . . . . . . . . . . Xin Xie, Honglei Chen, Ying Wu, Heng Zhang, and Peng Wu

51

Three Dimensional Canonical Correlation Analysis and Its Application to Facial Expression Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Lei Gang, Zhang Yong, Liu Yan-Lei, and Deng Jing

56

Technology Barriers Analysis on Bearing Industry Based on Relational Matrix Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xianzhao Jia, Feng Lv, Yonggang Liu, and Juan Wang

62

Optimizational Study on Computing Method of Channel Earthwork Based on MATLAB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yuqiang Wang, Jianqiang Zhang, and Yinghua Wang

69

A New Co-training Approach Based on SVM for Image Retrieval . . . . . . Hui Liu, Hua Han, and Zhenhua Li

77

XX

Table of Contents – Part I

Recognizing Human Activities Using Non-linear SVM Decision Tree . . . . Haiyong Zhao, Zhijing Liu, and Hao Zhang

82

Research on Logistics Service Providers Selection Based on AHP and VIKOR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Lu Shan

93

Research on Digitized Scenario for Tactical Internet Simulative Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jian-jun Shen, Hua Tian, and Zhi-chun Gan

99

Numerical Simulation for the Timoshenko Beam Equations with Boundary Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dian-kun Wang and Fu-le Li

106

A Study on Technology Architecture and Serving Approaches of Electronic Government System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ChunNian Liu, YiYun Huang, and Qin Pan

112

Construction and Development of CRM Technology and Industry Chain in China . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ChunNian Liu, YongLong Wang, and Qin Pan

118

Multi Scale Adaptive Median Filter for Impulsive Noise Removal . . . . . . . Xiangzhi Bai, Fugen Zhou, Zhaoying Liu, Ting Jin, and Bindang Xue

124

GA-Based Optimization Approach of Fractional Linear Neural Network and Its Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Guowei Yang and Lei Guo

130

Study of Unicyclic Graph with Maximal General Randi´c Index Rα for α < 0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Deqiang Chen

136

Study on Molecular Dynamics Simulation of Calcium Silicate Hydrate (C-S-H) Gels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Peng Hu and Wei Dai

142

Apply Ensemble of Lazy Learners to Biomedical Data Mining . . . . . . . . . Liu Pengfei and Tang Wulei

148

Molecular Dynamics Simulation on Calcium Silicate Hydrate Doped Organic Molecules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Wei Dai, Zhonghe Shui, and Ping Duan

155

Construction of Energy Management and Control Information System in Iron and Steel Enterprise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Wen Qiang Sun, Jiu Ju Cai, and Yong Liu

161

Table of Contents – Part I

XXI

Review of Dynamic Modeling and Simulation of Large Scale Belt Conveyor System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Qing He and Hong Li

167

Identification of Network Traffic Based on Radial Basis Function Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yabin Xu and Jingang Zheng

173

Thermal Hotspots in CPU Die and It’s Future Architecture . . . . . . . . . . . Jian Wang and Fu-yuan Hu

180

A Modified Nearest Neighbor Classification Approach Based on Class-Wise Local Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Deqiang Han, Chongzhao Han, and Yi Yang

186

The kinematics and Dynamics Simulation of the Collision between the Rigid Rod and the Ground . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Lulu Gao and Wenli Yao

193

Research on Innovative Practice Teaching System Based on the High-End Practice Teaching Environment for Software Engineering Speciality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jianli Dong, Cunhua Li, Zhaohui Ji, and Junming Wu

199

Chaotic Artificial Bee Colony Used for Cluster Analysis . . . . . . . . . . . . . . . Yudong Zhang, Lenan Wu, Shuihua Wang, and Yuankai Huo

205

A DVE Time Management Simulation and Verification Platform Based on Causality Consistency Middleware . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hangjun Zhou, Wei Zhang, Yuxing Peng, and Sikun Li

212

Ways of and Gains from Foreign Communication through Computer in China’s College Bilingual Teaching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ailing Gong, Tingrui Liu, and Yan Liu

219

Correction Method for Photoelectric Theodolite Measure Error Based on BP Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hai-yan Li and Yun-an Hu

225

The Application of LOGO! in Control System of a Transmission and Sorting Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jian Liu and Yuan-jun Lv

231

Design of Firmware Update Strategy in Tower Mounted Amplifier . . . . . . Yi Lv and Shuqin Han

237

Extraction of Urban Built-Up Land in Remote Sensing Images Based on Multi-sensor Data Fusion Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chengfan Li, Jingyuan Yin, Junjuan Zhao, and Lan Liu

243

XXII

Table of Contents – Part I

Coordinated Control and Localizing Target System for Multi-UAVs Based on Adaptive UKF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hengyu Li, Jun Luo, Lei Li, Jin Tong, and Shaorong Xie

249

Applying Fuzzy Data Mining to Telecom Churn Management . . . . . . . . . . Kuo-Hsiung Liao and Hao-En Chueh

259

Towards a Pre-computed Relation Matrix for Semantic Web Service . . . . Luokai Hu, Shi Ying, and Kai Zhao

265

Precise Localization of Facial Features Based on Cascade Fusion . . . . . . . Ying Chen, Chunlu Ai, and Chunjian Hua

271

Multi Scale Toggle Contrast Operator Based Image Analysis . . . . . . . . . . Xiangzhi Bai, Fugen Zhou, Zhaoying Liu, Bindang Xue, and Ting Jin

278

Prototype Design and Motion Analysis of a Spherical Robot . . . . . . . . . . . Shengju Sang, Ding Shen, Jichao Zhao, Wei Xia, and Qi An

284

An Algorithm of Fast Mining Frequent Neighboring Class Set . . . . . . . . . Gang Fang, Hong Ying, and Jiang Xiong

290

Research and Development Trends of Car Networking . . . . . . . . . . . . . . . . Wei He, Zhixiong Li, and Guotao Xie

296

A Research on MIMO Radar Based on Simulation . . . . . . . . . . . . . . . . . . . Zeng Jiankui and Dong Zhiming

302

Curriculum Reform Research of Computer Network Technology Based on School-Enterprise Cooperation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Peng Liu

308

Industry Cluster’s Adaptive Co-competition Behavior Modeling Inspired by Swarm Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Wei Xiang and Feifan Ye

312

Empirical Study on Relationship Capital in Supply Chain-Based on Analysis of Enterprises in Hunan Province . . . . . . . . . . . . . . . . . . . . . . . . . . Lu Shan and Ou-yang Qiang-bin

320

Hard-Failure Diagnosis Using Self-tuning Kalman Filter . . . . . . . . . . . . . . . Xiuling Xu and Xiaodong Wang

326

Research on the Structure Transformation of Landing Craft . . . . . . . . . . . Linfang Su, Xiao Liang, and Zhibin Li

335

Dynamic Reasoning under Probabilistic Uncertainty in the Semantic Web . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Limin Chen and Zhongzhi Shi

341

Table of Contents – Part I

PLS Regression on Wavelet Transformed Infrared Spectra for Prediction of Coal Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yanming Wang, Deming Wang, Haihui Xin, Xiaoxing Zhong, and Gouqing Shi

XXIII

348

The Case of Web-Based Course on Taxation: Current Status, Problems and Future Improvement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Zhigang Qin

354

A Novel Block Encryption Based on Chaotic Map . . . . . . . . . . . . . . . . . . . . Pengcheng Wei, Huaqian Yang, Qunjian Hang, and Xi Shi

360

Fuzzy Evaluating Customer Satisfaction of Jet Fuel Companies . . . . . . . . Haiying Cheng and Guoyi Fang

368

The Research of the BDAR for ATS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chengjin Gao, Sheng Sheng, and Lin Wang

374

Secure Biometric E-Voting Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Taha Kh. Ahmed and Mohamed Aborizka

380

Glowworm Swarm Optimization Algorithm for Solving Numerical Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yan Yang and Yongquan Zhou

389

An Energy Aware Ant Colony Algorithm for the Routing of Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Deqiang Cheng, Yangyang Xun, Ting Zhou, and Wenjie Li

395

A Kind of Decay-Curve Inertia Weight Particle Swarm Optimization Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yan Sun, Shishun Zhu, Qiang Li, Daowei Zhu, and Shujun Luo

402

Performance Analysis of Cooperative Virtual Laboratory Based on Web . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Zheng Gengzhong and Liu Qiumei

410

Numerical Simulation of Gas Leaking Diffusion from Storage Tank . . . . . Hongjun Zhu and Jiaqiang Jing

416

An Orthogonal Wavelet Transform Blind Equalization Algorithm Based on the Optimization of Immune Clone Particle Swarm . . . . . . . . . . . . . . . . Guo Yecai and Hu Lingling

422

A Support Vector Machine Blind Equalization Algorithm Based on Immune Clone Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Guo Yecai and Ding Rui

428

A Service Access Security Control Model in Cyberspace . . . . . . . . . . . . . . . Li Qianmu, Yin Jie, Hou Jun, Xu Jian, Zhang Hong, and Qi Yong

434

XXIV

Table of Contents – Part I

A Motor Speed Measurement System Based on Hall Sensor . . . . . . . . . . . . Wen-cheng Wang

440

Fault Reporting Based on Geographic Information of Distribution Network Fault Locating System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sun Quande, Li Kaiyang, and Wang Chunsheng

446

Key Curriculum Reform Research on Numerical Analysis . . . . . . . . . . . . . Zhong Li and Chensong Peng

453

The Topic Analysis of Hospice Care Research Using Co-word Analysis and GHSOM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yu-Hsiang Yang, Huimin Bhikshu, and Rua-Huan Tsaih

459

An Approach to Feature Selection Based on Ant Colony Optimization and Rough Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Junyun Wu, Taorong Qiu, Lu Wang, and Haiquan Huang

466

A New Adaptive Deformable Model Using Gradient Vector Flow . . . . . . . Bin Zhao, Siyuan Cheng, and Xiangwei Zhang

472

A Comparative Study of Different Distances for Similarity Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Zhong Li, Qiaolin Ding, and Weihua Zhang

483

Reliability Analysis for CNC Machine Tool Based on Failure Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yingzhi Zhang, Rui Zheng, Guixiang Shen, and Bingkun Chen

489

Distribution Network Information System of Jiaozuo Based on ArcObject . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Wang Chunsheng, Sun Quande, and Li Kaiyang

497

Design and Implementation for Low Network Loaded and Automatically Photographic Surveillance System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Linying Jiang and Heming Pang

502

The Research on Incentive Mechanism of Knowledge Creation in IT Enterprise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Si-hua Chen

509

Computational Simulation of Submarine Oil Spill with Current . . . . . . . . Wei Li, Yongjie Pang, and Hongwei Li

515

Study on Engineering Consciousness and Ability Fostering of Applied Talents in Engineering Education . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Wang dong ping

521

Application Research of Improved Grey Forecasting Model in Load Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yuansheng Huang, Wei Fang, and Zhou Fan

529

Table of Contents – Part I

XXV

An Adaptive Multi-Agent System for Project Schedule Management . . . . Yongyi Shou and Changtao Lai

535

Using F-Metric to Construct a New Niederreiter Public Key Cryptosystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mu Han, Hong Zhang, and Chun-gen Xu

541

Flax Fibers as Reinforcement in Poly (Lactic Acid) Biodegradable Composites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yuan Yuan, Minghui Guo, and Yong Wang

547

A Novel Framework to Maximum Lifetime for Wireless Sensor Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Feng Sheng, Qi Xiao-gang, and Xue Ji-long

554

Enhancing Security by System-Level Virtualization in Cloud Computing Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dawei Sun, Guiran Chang, Chunguang Tan, and Xingwei Wang

565

Enabling Two Levels of Adaptation: A Framework for Traffic-Aware and Context Management Deployment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chin-Ling Chen and Jia-Ching Wang

571

Bacteria Foraging Based Agent Feature Selection Algorithm . . . . . . . . . . . Dongying Liang, Weikun Zheng, and Yueping Li

581

Research on the Interrupt Functions Based on the CAN Controller of LPC2300 Series ARM Chips . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Li Tu, Juanjuan Song, and Jun’an Liu

589

Data Recognition and Filtering Based on Efficient RFID Data Processing Control Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hsu-Yang Kung, Chiung-Wen Kuo, and Ching-Ping Tsai

596

Research on Operating Performance Evaluation of Electric Power Enterprises Based on F-AHP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yuansheng Huang, Wei Fang, and Mingxi Shi

602

Efficient Reference Picture Management Schemes for H.264/AVC Decoding System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chunshu Li, Kai Huang, Min Yu, and Xiaolang Yan

609

Research on the Algorithm for 3L-CVRP with Considering the Utilization Rate of Vehicles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Han-wu Ma, Wei Zhu, and Sen Xu

621

Simulation of a Signalized Intersection Delay Model . . . . . . . . . . . . . . . . . . Minghui Wu, Lian Xue, Hui Yan, and Chunyan Yu

630

XXVI

Table of Contents – Part I

Matlab for Forecasting of Electric Power Load Based on BP Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xi-ping Wang and Ming-xi Shi

636

Application of Bayesian Networks in Situation Assessment . . . . . . . . . . . . Xi Su, Peng Bai, Feifei Du, and Yanping Feng

643

Preliminary Analysis on the Relative Solution Space Sizes for MTSP with Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Junling Hao

649

Hybrid Artificial Fish Swarm Algorithm for Solving Ill-Conditioned Linear Systems of Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yongquan Zhou, Huajuan Huang, and Junli Zhang

656

A Survey on Social Image Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Zheng Liu

662

Binocular Vision-Based Position and Pose of Hand Detection and Tracking in Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chen Jun, Hou Wenjun, and Sheng Qing

668

Linear Feedback Anti-control of Chaos in Permanent Magnet Synchronous Motor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jingyue Wang, Haotian Wang, and Lixin Guo

676

An Infrastructure for Personalized Service System Based on Web2.0 and Data Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 686 Yu Liu, Weijia Li, Yuan Yao, Jing Fang, Ruixin Ma, and Zhaofa Yan Research of Optimized Agricultural Information Collaborative Filtering Recommendation Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Fang Kui, Wang Juan, and Bu Weiqiong

692

An IPv6 Multihomed Host for Outbound Traffic . . . . . . . . . . . . . . . . . . . . . Chin-Ling Chen and Sheng-Lung Cao

698

A New Practical Electric Vehicle Battery Management System . . . . . . . . . Yanpeng Shi and Guoxin Wu

704

Dynamic Methods of Component Composite Service Selection Based on Trust-Aware . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yuan Bo, Wang BinQiang, Zhao Bo, and Song Shasha

711

An Optimization Multi-path Inter-Session Network Coding in Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Zhuo-qun Xia, Chao Liu, Xue-han Zhu, Pin-chao Liu, and Li-tong Xie

717

Table of Contents – Part I

Exploring the Intrinsic Motivation of Hedonic Information Systems Acceptance: Integrating Hedonic Theory and Flow with TAM . . . . . . . . . Zhihuan Wang Vision-Guided Robot Tracking Algorithm Based on Characteristic Description of Maneuvering Target . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yuan Zhang, Peng Wang, Xin Li, Shaochen Kang, Jinglei Xin, and Wenhao Jiang

XXVII

722

731

Fuzzy Comprehensive Evaluation of Peak Load Unit Based on Entropy Weight . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Zhengyuan Jia, Zhou Fan, and Shaohui Chen

737

Molecular Variation of Potato Virus Y in China . . . . . . . . . . . . . . . . . . . . . Hongyi Yang, Nana Zhang, Debin Li, and Lili Li

744

The Overview of Entity Relation Extraction Methods . . . . . . . . . . . . . . . . . Xian-Yi Cheng, Xiao-hong Chen, and Jin Hua

749

Detailed Design and Analysis on Error Handing and Controlling Protocol in Mobile Payment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xu Yong and Yan Tingting

755

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

761

Unified Electronic Currency Based on the Fourth Party Platform Integrated Payment Service* Xu Yong1 and Hu Qiqi2 1

School of Economics and Commerce, South China University of Technology, Guangzhou University City, Guangzhou, China [email protected] 2 School of Economics and Commerce, South China University of Technology, Guangzhou University City, Guangzhou, China [email protected] Abstract. This paper presents a solution of unified e-currency based on the fourth party platform integrated payment service. The purpose of the paper is to solve the problem of distribution and resource-wasting caused by the lack of unified electronic currency, and to solve regulatory difficulties due to regulation size caused by a wide variety of e-currency. Methods: This article first analyzes the problems in the development of electronic money, and then proposes the concept of a unified electronic currency based on the fourth party platform integrated payment service. Besides, it proposes a unified mechanism and transaction procedures for unified e-currency, and analyzes the liquidation process, security and regulatory requirements, which are involved in using unified electronic currency. Keywords: The fourth party platform integrated payment service, Unified electronic currency, Unified billing.

1 Introduction Electronic currency has a strong vitality. Compared with the traditional currency, it has unique advantages, such as improving the operational efficiency of capital, reducing the transaction costs of clearing, beyond the constraints of time space and so on. However, we cannot ignore the problems which appeared in the development of e-currency. For example, the problems of unity, security issues and regulatory issues, etc. To solve the above problems of electronic money, we need to propose a complete solution from the entire payment chain perspective. Therefore, this paper will develop unified e-currency based on the fourth party platform integrated payment service.

2 Introduction of the Fourth Party Platform Integrated Payment Service The fourth party platform integrated payment service is a comprehensive payment service platform for the payment chain which is based on the theory of The Fourth Party *

The National Soft Science Research Program, 2010B070300016.

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 1–6, 2011. © Springer-Verlag Berlin Heidelberg 2011

2

X. Yong and H. Qiqi

Payment. The Fourth Party Payment is an integrated service provider of electronic payment and value-added service, it can provide a set of solution to customers by integrating and managing various resources as well as the capabilities of complementary providers, standardizing the E-commerce process and providing the supervision interfaces through its own payment and value-added service. The service provided by the fourth party payment enterprises are not solely from a perspective angle of payment, but for uniting various protocols, providing a third-party cross-authorization and electronic payment supervision modes, and more value-added services. [1]

3 Unified Electronic Currency Based on the Fourth Party Platform Integrated Payment Service 3.1 The Concept of Unified Electronic Currency Unified e-currency is a kind of electronic cash in pure electronic form. It is issued by the fourth party platform integrated payment service under the supervision of the relevant regulatory authorities and can be generic used in the whole network (Internet and mobile networks) as well as common industry. The fourth party platform integrated payment service reaches agreement with various electronic currency issued subject, which can make the unified e-currency realize the direct exchange and indirectly achieve these mutual exchange between the electronic cash. Besides itself has very broad range of flow. Unified e-currency itself has strong versatility and convenience. It have the features of wide coverage and the bank independence. 3.2 Unified Electronic Currency Mechanisms and Transaction Process Unified e-currency includes two functions: one is using the traditional currency recharge for consumption; the other is the use of other electronic cash recharge. (A) Obtaining process of unified electronic currency The users store a certain amount of cash after they apply to the fourth party platform integrated payment service for account. And then they will receive the appropriate client software, and that is the electronic wallet. After that they can use electronic cash to make purchases. As shown in figure1:

Fig. 1. Obtain process of unified electronic currency

Unified Electronic Currency Based on the Fourth Party

A. B. C. D. E. F.

3

Users apply to the fourth party integrated payment service for account. Users login account and buy unified e-currency and download electronic purse through the net. And then unified e-currency will be stored in electronic purse. Users choose commodities, then transfer product information and unified electronic currency to the businessman. Businessmen confirm legitimacy of unified electronic currency to the fourth party integrated payment service. Confirm the legality of unified electronic currency, businessmen delivery. The fourth party integrated payment service makes unified settlement of funds according to financial information.

(B) Exchanging with other e-currency (a). Dead e-currency exchange At present, there is Dead e-currency. The user makes an e-currency exchange, which can only buy the goods or services provided by the e-currency issuer, and cannot counter against the traditional currency. When the user no longer uses goods and services provided, the user account balance will be idle, resulting in waste. In order to reduce this waste, the fourth party platform integrated payment service will reach agreement with all e-currency issuers, so that unified electronic currency can be as a bridge to realize the exchange between different dead e-currency. Unified e-currency exchanged from the dead e-currency can only be stored in the Internet account for the purchase of other e-currency, but it cannot be stored in the electronic wallet to purchase other commodities. The process shown in Figure 2:

Fig. 2. Dead e-currency exchange process

A. B.

C.

D.

User logins the account of unified e-currency and makes dead e-currency exchange function available. Then selects dead e-currency A for exchange. User recharges unified e-currency with dead e-currency A; Then the fourth party platform integrated payment service adds the funds in the internet unified ecurrency account according to e-currency exchange rate, and send encrypted recharge messages to dead e-currency issuer A; Dead e-currency issuer A deducts the funds in dead e-currency A account according to the encrypted recharge messages. User recharges the account B with the unified e-currency; then the dead ecurrency issuer B adds funds in the dead e-currency B account depending on the encryption information. The fourth party platform integrated payment service makes unified settlement and do hedge liquidation according to the recharge message.

4

X. Yong and H. Qiqi

(b). Undead e-currency exchange Undead e-currency is a kind of e-currency which can exchange with traditional currency. When the user no longer uses this kind of e-currency, e-currency account balance can be converted into bank account balances, and will not cause waste. Unified e-currency issuer reaches agreement with different e-currency issuers, through which other undead e-currency can be directly used to recharge and funds in the user electronic wallet will add. The funds can be used for shopping. The process shown in Figure 3:

Fig. 3. Undead e-currency exchange processes

In the process of exchange, unified e-currency purse must link to the Internet. Exchange funds are directly deposited into the electronic wallet, which can help separate unified e-currency and funds in Internet account out. A. B. C. D.

E.

User logins unified e-currency wallet, select the recharge way: electronic cash recharge. User recharges unified e-currency purse with e-currency A. Issuers of e-currency receive and sent recharge data through a special interface, then reduce the amount of e-currency A. The fourth party platform integrated payment service receives data through the special interface, send data to unified e-currency purse to increase the amount of e-purse. The fourth party platform integrated payment service makes unified settlement, then the funds of e-currency A in the bank A go to the temporary platform bank account.

4 Unified e-Currency Settlement Mechanisms and Processes 4.1 Unified e-Currency Shopping Capital Clearing and Settlement User uses the bank account A to buy unified e-currency, the funds remains in the fourth party platform integrated payment service temporary clearing account. When using the unified e-currency shopping, the liquidation process is as follows:

Unified Electronic Currency Based on the Fourth Party

5

Fig. 4. Unified e-currency shopping capital settlements

A. B.

C.

D. E.

F.

G.

H.

I.

User buys unified e-currency and the purchase information is passed to the fourth party platform integrated payment service. The fourth party platform integrated payment service receives purchase information which includes user account information, funds information and so on. The fourth party platform integrated payment service sends finished information to the central bank, and then the central bank transfers the funds from User Banks to the fourth party platform integrated payment service temporary clearing account. User uses unified e-currency purse to pay, and the secure payment information is passed to the fourth party platform integrated payment service The fourth party platform integrated payment service receives users’ secure payment information. This information includes user payment amount information, the user bank information, and business information. The fourth party platform integrated payment service based on business information; determine the business bank account information, and business information. The fourth party platform integrated payment service clears up the financial information and bank account information received, analyzes the flow of funds between various banks, and makes the transmission of finished information to the central bank. According to information received, the central bank makes the capital settlement, and then sends the funds from the fourth party platform integrated payment service temporary clearing account to merchant banks clearing account. The central bank transfers liquidation Information to the fourth party platform integrated payment service, then the information is cleared up and sent to the merchants head bank by the platform. The head band sends the specific account information and financial information to the merchant bank branches, making the primary account and fund settlement.

4.2 Non-unified e-Currency Shopping Capital Clearing and Settlement When making purchases, the user can directly use the non-unified e-currency. Platform can judge issuer and currency exchange rates of RMB, according to the type of e-currency, and clears up the funds directly with the issuer bank account. It needn’t to

6

X. Yong and H. Qiqi

be converted into unified e-currency. The liquidation process is the same as the direct use of unified e-currency. 4.3 Dead e-Currency and Undead e-Currency Clearing and Settlement Dead e-currency. The amount of Dead e-currency exchange is usually little, and the exchange involves interests of various businessmen. Therefore, dead e-currency uses the way of hedging accounts directly to clear up. The way of hedging accounts directly to clear up refers that it directly hedges user’s account among dead e-currency issuers, is not to involve in the actual transfer of funds. Undead e-currency clearing and settlement. The users choose the function of undead e-currency exchange. The fourth party platform integrated payment service must reach agreement with undead e-currency issuers, provide the interface for them, clearing up thought the platform. Its liquidation mainly relates to inter-bank settlement; the process is the same as using Internet Bank to buy unified e-currency.

5 Summary Unified e-currency is the ultimate trend in the development of e-currency, the full realization of the ultimate form is industry-wide and entire network coverage function, and replaces all other types of e-currency. The purpose of e-currency exchange is to replace the other e-currency smoothly in the early development. Security and supervision of unified e-currency must be able to keep up with development of unified e-currency, which requires the State and the community to pay adequate attention.

References 1. 2. 3. 4. 5. 6. 7.

Xu, Y., Fang, C.: A theoretical framework of fourth party payment. ICEE, EI Retrieval Li, Y.: Promote the healthy development of the cause of payment and settlement. Financial Computer of China (03) (2010) Ye, Y., Xiong, Z., Chen, Y.: Proposal concerning reconstruction of Chinese foreign currency clearing system based on analysis of its drawbacks. Shanghai Finance (03) (2008) Zhou, Z.: Payment and settlement system for cross-national perspective. China Foreign Exchange (08) (2009) Zhou, J.: The Development of Modern Payment System and the Adjustment of Monetary Regime. Journal of Financial Research (01) (2007) Lacker, J.M.: Clearing, settlement and monetary policy. Journal of Monetary Economics 40, 347–381 (1997) Committee on Payment and Settlement Systems: The role of central bank money in payment systems. In: BIS (2003)

Modeling and Power Flow Analysis for Herringbone Gears Power Dual-Branching Transmission System Xiaofang Yang1, Yanxiang Zhu2, Zongde Fang1, and Jiangong Gu1 1

School of Mechatronics, Northwestern Polytechnical University, Xi’an 710072, China 2 Patent Bureau of State Intellectual Property Office, Beijing 100088, China [email protected]

Abstract. Based on power dual-branching transmission system of herringbone gears, the mechanical structural model was established. This study represented the simplified algorithm to obtain its power flow situations through formulating the deformation compatibility condition for the linear relationship between the torque and transverse deformation of tooth surface and the torque equilibrium condition. Besides, the effects on the power flow of system were calculated under all kinds of the installation error and processing error of gear pairs. Finally, the power flow situations of dual branches were solved via Programming. A numerical example that illustrated the developed theory was provided. The research results can be applied to analyze the actual application of herringbone gears power split-path transmission system. Keywords: herringbone gears; power-split; torque equilibrium; deformation compatibility; error.

1 Introduction The herringbone gears power flow transmission system has been widely used in the aeroengine and reducer to achieve the high-speed and reverse spilt-flow and confluence high-power [1-3]. The power dual-branching gear transmission system which is composed by herringbone gears of dual-branching structure to implement dual-path power of cross-flow is simple and compact [4]. Besides, the herringbone gears as transmission component has unique characteristics of large contact ratio, smooth transmission and low noise, which make the system on the one hand, to meet the requirements of heavy duty working conditions. On the other hand, it has broad application prospects on the fields such as aviation and marine transmissions due to its size reduced and weight lightened. In this study, the mechanical structural model of power flow transmission system is presented. Based on the torque analysis of the herringbone gears, the mutually meshing influence between related gear pairs of system is researched during the analysis of deformation compatibility. The power flow situations are solved by programming. Actually, the clearance is unavoidable during meshing of gear pairs due to the processing error and installation error of gear system [5-6], and analyzing the influence on power split. R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 7–15, 2011. © Springer-Verlag Berlin Heidelberg 2011

8

X. Yang et al.

2 Modeling and Analysis Power dual-branching gear transmission system can be defined as the gear system to divide the torque or power into dual-branching and generally with two stage transmission considering the requirements of structural characteristics. An analytical method is developed and used to study the load-sharing properties of such designs. 2.1 Outline Drawing of Power Dual-Branching System The outline drawing model is adopted by herringbone gears power dual-branching transmission system [4]. The whole system is comprised of first-level transmission and second-level transmission system. In this system, input power is given to firstlevel gear though pinion split-flow which transfers the torque to second-level pinion, and then converges into second-level gear for output though second-level. In order to achieve better load-sharing result and minimize interaction between two levels, the shafts connecting two levels are made into slender load-sharing torsion shaft parts. 2.2 Mechanical Structural Model In the process of modeling, according to herringbone gears transmission characteristic, the system is simplified as lumped-parameter model and the wheel is treated as rigid body. The mechanical structural model established is as shown in Figure 1. The entire system consists of six herringbone gears, T1 as input torque, T6 as output torque, and other torque subjected to every gear as T2 , T3 , T4 , T5 , respectively. As to gear pair 12, gear pair 13, gear pair 46 and gear pair 56, the torque in the two gears of every gear pair are Tij , T ji ( i = 1, 2,3, 4,5, 6; j = 1, 2,3, 4,5, 6 ), respectively,

thereinto, Tij defined as function torque of gear j versus gear i, active torque defined as positive, load torque defined as negative. T24

T42 T46

T21

T64

T1 2

T13

T1

T6 T65 T56

T31 T35

T53

Fig. 1. Mechanical structural model

2.3 Conditions of Torque Equilibrium [7]

The conditions of torque equilibrium which can be obtained based on gears and shafts illustrated in figure 2 are represented with the following equations.

Modeling and Power Flow Analysis for Herringbone Gears Power

⎧T1 + T12 + T13 = 0, T24 + T21 = 0 ⎪ ⎨T35 + T31 = 0, T42 + T46 = 0 ⎪T + T = 0, T + T + T = 0 ⎩ 53 56 6 64 65

9

(1)

The equation of gear 6 is non-independent, due to six herringbone gears formed closed system. The base circle radius of herringbone gears is ri ( i = 1, 2,3, 4,5, 6 ). Owing to gear 2 and pinion 4, gear 3 and pinion 5 both coaxial, according to action and reacting force equilibrium condition, the equations are as follows, (2)

T24 = −T42 , T35 = −T53

Other herringbone gears pairs meshing, the conditions of torque equilibrium are as follows, (3) T ji = −Tij rj / ri Equations (1), (2), (3) yield ⎧T1 + T12 + T13 = 0, −T12 r2 / r1 + T46 = 0 ⎨ ⎩−T13 r3 / r1 + T56 = 0, T6 − T46 r6 / r4 − T56 r6 / r5 = 0

(4)

3 Conditions of Deformation Compatibility The relationship model of various gear pairs meshing rotation angle and gear shafts twist angle of system established is shown in figure 2.

24

(T21 )

2

4

46 12

6

1

13

(T46 )

(T12 )

(T13 ) 56

3

(T56 )

5

35

(T31 )

Fig. 2. Relationship model of various gear pairs meshing rotation angle and gear shafts twist angle

3.1 Conditions of Deformation Compatibility [8]

The relationship of various gear pairs meshing rotation angle and gear shafts twist angle is represented as follows, ⎧Δφ12 (T12 ) = φ1 − φ2i12 , Δφ24 (T21 ) = φ2 − φ4 ⎪ ⎨Δφ46 (T46 ) = φ4 − φ6i46 , Δφ13 (T13 ) = φ1 − φ3i13 ⎪Δφ (T ) = φ − φ , Δφ (T ) = φ − φ i 3 5 56 56 5 6 56 ⎩ 35 31

(5)

10

X. Yang et al.

Here, φi is the rotation of the gear i . Δφij ( Tij ) is the angular distortion of gear i with respect to j under the action of torque Tij , that is the loading transmission error,

and Δφij ( Tij ) is the function of Tij .

Gear 2 and 3, pinion 4 and 5 are as same level gear, respectively. Therefore, the numbers of teeth of gear 2 and 3, pinion 4 and 5 equal, respectively, yield the following equations,

i12 = i13 , i46 = i56

(6)

where i12 = z2 / z1 = r2 / r1 , i13 = z3 / z1 = r3 / r1 , i46 = z6 / z4 = r6 / r4 , i56 = z6 / z5 = r6 / r5 Equations (5) yield T46 = T12 r2 / r1 , T56 = T13 r3 / r2

(7)

Equations (5), (6) and (7) yield the deformation compatibility equations of the system as follows, Δφ12 (T12 ) r1 / r2 − Δφ24 (T12 ) r2 / r1 + Δφ46 (T12 ) r2 / r1 = Δφ13 (T13 ) r1 / r3 − Δφ35 (T13 ) r3 / r1 + Δφ56 (T13 ) r3 / r1

(8)

3.2 Torsion Meshing Stiffness

The relationship between the torsion meshing stiffness K ij and load Tij is nonlinear, however, in the vicinity of a certain load, so it can be approximately considered as linear [9-10]. Accordingly, the meshing angular distortion of gear pairs can be simplified represented as follows, (9) Δφij = Tij / Kij ri 2 In the computational model of gear transmission system, the torsional meshing stiffness is analyzed based on the variation of line of action of gear pair, N / mm as unit, besides, the meshing stiffness in this study all as linear. The torsional meshing stiffness between gear i and gear j is

K ij = 2 × 103 Cγ ij bi

(10)

(

)

Here: b is total width of herringbone gears, mm as unit; Cγ ij = 0.75ε αij + 0.25 C 'ij is defined as the total rigidity mean value in the end section of gear i and j , abbreviation “meshing stiffness”, mm ⋅ μ m / N as unit; ε α is transverse contact ratio; Cij ' = 1 q is the maximum stiffness in the normal section of one pair gears, q = 0.04723 + 0.15551cos3 β / z1 + 0.25791cos3 β / z2 − 0.00635 x1 − 0.11654 x1 cos3 β / z2 − 0.00193 x2 − 0.24188 x2 cos3 β / z2 + 0.00529 x12 + 0.00182 x22

Modeling and Power Flow Analysis for Herringbone Gears Power

11

3.3 Twist Angle Deformation of Load-Sharing Torsion Shaft

Gear 2 and pinion 4, gear 3 and pinion 5 are connected though the load-sharing torsion shafts and their angular deformation is represented by twist deformation of the shaft [11]. Δφ24 = −(32r2 l24T12 ) / (Gπ r1 (d 2 4 − d 4 4 )), Δφ35 = −(32r3l35T13 ) / (Gπ r1 ( d34 − d5 4 ))

(11)

where, G is the material shear elastic modulus of shaft, MPa as unit, and for steel, G = 8.1× 10 4 MPa; l AB ( A = 2,3; B = 4,5) is the length of the shaft under torque, mm as unit; d i (i = 2,3, 4,5) is the shaft diameter connecting first-level and second-level. 3.4 Equations of Deformation Compatibility

As mentioned above, Δφij ( Tij ) is the angular distortion of gear i with respect to

gear j under the action of torque Tij . The variable Δφij (Tij ) includes initial gap between

gears correspond to the relative twist angle Δφij0 before tooth surface of gear i and j come into contact and the relative twist angle Δφij1 ( Tij ) at the mesh point after tooth surface of gear i and j contact deformation. Δφij ( Tij ) =Δφij1 ( Tij ) + Δφij0

(12)

Equations (8), (9), (10), (11) and (12) yield the deformation compatibility equations under the action of torque Tij of meshing gear i and gear j as follows, T12 [1 / ( K12 r1r2 ) − (32r2 l24 ) / (Gπ r1 (d 2 4 − d 4 4 )) + r2 / ( K 46 r4 2 r1 )] − T13 [1 / ( K13 r1r3 ) − (32r3l35 ) / (Gπ r1 (d34 − d5 4 )) + r3 / ( K56 r5 2 r1 )] = 0

(13) The linear equations (4) and (13) conveniently yield load bearing torque of each gear pair Tij , which is the condition of power split of system.

4 Error on the Influence of Power Split Envision that the output gear of the gear train is rigidly fixed from rotating and a torque is applied to the input pinion. As the torque is applied, the input shaft will rotate some amount because of deformations [4]. This rotation of the input pinion relative to the output gear is called the load windup. In the condition of gear parameter and meshing parameter invariable, due to the parallel error of the axes of gear pair made by manufacture and installation [11-12] (thereinafter abbreviation axis based error), the deviation of center distance among gear pairs is produced directly, leading to the gap inevitably in the process of meshing of gear pairs, corresponding to the initial twist angle Δφij0 non-zero.

12

X. Yang et al.

4.1 Initial Twist Angle

The gap of center distance error separately projects along and vertically the direction of line of action, in which the projection along the direction of line of action can transform as the angle of clearance, while the projection vertically the direction of line of action possible ignore[13]. According to the geometric al relationship of tooth face of herringbone gears, when the axial movement of gear i (the gap of center distance error) is ΔEi , the relative twist angle corresponding to initial clearance between gears of gear pair is represented as follows: Δφij 0 = ΔEi sin α t cos β / Rij

（14)

where Rij is the mean radius of gyration for the meshing point with respect to the shaft of gear. 4.2 Error Occurring in Gear Pair

If the gap of center distance axial error ΔE2 occurs between pinion 1 and gear 2, the deformation compatibility equations as follows, T12 [1/ ( K12 r1r2 ) − (32r2 l24 ) / (Gπ r1 ( d 2 4 − d 4 4 )) + r2 / ( K 46 r4 2 r1 )] −T13 [1/ ( K13 r1 r3 ) − (32r3l35 ) / (Gπ r1 ( d 34 − d 5 4 )) + r3 / ( K 56 r5 2 r1 )] = ΔE2 sin α t cos β1r1 / r2 2 (15) If the gap of center distance axial error ΔE3 occurs between pinion 1 and gear 3, the deformation compatibility equations as follows, T12 [1/ ( K12 r1r2 ) − (32r2 l24 ) / (Gπ r1 (d 2 4 − d 4 4 )) + r2 / ( K 46 r4 2 r1 )] −T13 [1/ ( K13 r1r3 ) − (32r3l35 ) / (Gπ r1 ( d34 − d 54 )) + r3 / ( K 56 r52 r1 )] = ΔE3 sin α t cos β r1 / r32

(16)

If the gap of center distance axial error ΔE4 occurs between pinion 4 and gear 6, the deformation compatibility equations as follows, T12 [1/ ( K12 r1r2 ) − (32r2 l24 ) / (Gπ r1 (d 2 4 − d 4 4 )) + r2 / ( K 46 r4 2 r1 )] −T13 [1/ ( K13 r1r3 ) − (32r3l35 ) / (Gπ r1 (d 3 4 − d 5 4 )) + r3 / ( K 56 r52 r1 )] = ΔE4 sin α t cos β / r4 (17) If the gap of center distance axial error ΔE5 occurs between pinion 5 and gear 6, the deformation compatibility equations as follows, T12 [1 / ( K12 r1r2 ) − (32r2 l24 ) / (Gπ r1 ( d 2 4 − d 4 4 )) + r2 / ( K 46 r4 2 r1 )] −T13 [1 / ( K13 r1r3 ) − (32r3l35 ) / (Gπ r1 ( d34 − d 54 )) + r3 / ( K 56 r5 2 r1 )] = ΔE5 sin α t cos β / r5

(18)

If the gap of center distance axial error occurs in every gear pair simultaneously, the deformation compatibility equations as follows, T12 [1 / ( K12 r1r2 ) − (32r2 l24 ) / (Gπ r1 ( d 24 − d 4 4 )) + r2 / ( K 46 r4 2 r1 )] − T13 [1 / ( K13 r1r3 ) − (32r3l35 ) / (Gπ r1 ( d 34 − d 5 4 )) + r3 / ( K 56 r52 r1 )] = ΔE2 sin α t cos β r1 / r2 2 − ΔE4 sin α t cos β / r4 + ΔE3 sin α t cos β r1 / r3 2 − ΔE5 sin α t cos β / r5

(19)

Modeling and Power Flow Analysis for Herringbone Gears Power

13

5 Calculation Results The input power of pinion 1 is 1556 kw , rotational speed 6000 r / min. The parameters of gears of transmission system are listed in table 1, and the parameters of middle torsion shaft listed in table 2. Table 1. Parameters of gears Items

Pinion 1

Gear 2 and 3

Pinion 4 and 5

Gear 6

Number of teeth z Normal module mn / mm

24 8

80 8

65 8

176 8

Face width b / mm Helical angle

/( )

Modification coefficient

120

340

260

770

29.5

29.5

29.5

29.5

x1

x2

0.53

x1

x2

0.53

x1

0.53

x2

x1

x2

0.53

Table 2. Parameters of shafts Items Load effective shaft part length/mm

Shaft 24

l24

Shaft diameter/mm

d2

Shaft 35

1500

260, d 4

l35 180

d3

1500

260, d 5

180

5.1 The Effect of Power Flow Caused By the Gap of Center Distance Axial Error

The gap of center distance axial error of gear pairs, leading to the initial twist angle corresponding to gear pairs, impacts on power split of gear pairs of system, which is according to the torque transferred change. In the conditions of standard installation and existing different errors, the situation of power split in this study is listed in table 3 by programming. Table 3. Situation of power spilt in different conditions

items

Standard installation

Error in pair gear 12

Error in pair gear 13

Error in pair gear 46

Error in pair gear 56

Error in every gear pair

Ei / mm

0

T12 / N m

-1246

-1231

-1216

-1061

-938

-1242

T13 / N m

-1246

-1261

-1276

-1431

-1554

-1250

T46 / N m

-4154

-4104

-4054

-3537

-3126

-4142

T56 / N m

-4154

-4204

-4254

-4771

-5182

-4166

0.01

0.02

0.01

0.02

5.2 The Effect of Power Flow Caused by Installation Error

The load of every gear is diverse in the different engaging position of the system, which explained particularly in author’s other papers. The power flow charts of gear

14

X. Yang et al.

pairs in various situations are fitted through loaded tooth contact analysis [14]. The power flow of system in the various situation of two mesh period (correspondence ten engaging position) is as shown in figure 3.

Fig. 3. Power split diagram of the system under different situation

6 Conclusions This study was done to better understand split-path transmission and support their application in the future. From the analytical study, the following conclusions can be drawn: (1) The method developed in this study greatly simplifies the process for solving the situation of power spilt of the complex gear transmission system, increasing the computational efficiency, with the guidance to the three and four branch transmission structure. (2) The situation of power split affect though adjusting the gap of tooth flank can be conveniently calculated using the method developed. In normal conditions, the amount of power split of dual-branch is equivalent. The insignificant gap results in the tremendous transformation. Therefore, the branched structure is sensitive to the gap of center distance axial error. (3) The variation of power flow caused by the gap of center distance axial error is consistent with the graphs power split caused by installation error. The loaded imbalance of two sides of gear is caused by errors, which is easy to cause one-sided to overload. And consequently, the author proposes the method to achieve load-sharing of power spilt as far as possible though changing the gap of tooth flank of adjusting the axial gear position. (4) The precision required for gap and installation error is within the capabilities of available and proven manufacturing processes. The model provides important theoretical instruction for achieving valid error compensation of load-sharing.

Modeling and Power Flow Analysis for Herringbone Gears Power

15

References 1. Litvin, F.L., Fuentes, A.: Gear Geometry and Applied Theory. Cambridge University Press, Cambridge (2004) 2. Litvin, F.L., Zhang, J., Handschuh, R.F., Coy, J.J.: Topology of Modified Helical Gears. Surface Topography 3, 41–58 (1989) 3. White, G.: Design Study of A 375-kw Helicopter Transmission with Split-torque Epicyclic and Bevel Drive Stage. J. Mech. Eng. Sci. 197C, 213–224 (1983) 4. Timothy, L., Krantz, A.: Method to Analyze and Optimize the Load Sharing of Split Path Transmissions. In: Seventh International Power Transmission and Gearing Conference, San Diego (1996) 5. Zhang, Y., Fang, Z.: Analysis of Transmission Errors Under Load of Helical Gears with Modified Tooth Gears. ASME Journal of Mechanical Design 119, 120–126 (1997) 6. Umeyama, M., Kato, M., Inoue, K.: Effects of Gear Dimensions and Tooth Surface Modifications on The Loaded Transmission Error of A Helical Gear Pair. ASME Journal of Mechanical Design 120, 119–125 (1998) 7. Yao, Y., Yan, H.S.: A New Method for Torque Balancing of Planar Linkages Using Noncircular Gears. Proceedings of the Institution of Mechanical Engineers Part C-Journal of Mechanical Engineering Science 217, 495–503 (2003) 8. Gu, J.G., Fang, Z.D., Zhou, J.X.: Modeling and Power Flow Analyzing for Power Split System of Spiral Bevel Gears. In: Feng, G., Xijun, Z. (eds.) Proc. of 2009 International Workshop on Information Security and Application, pp. 401–404. Acadmy Publisher, Finland (2009) 9. Peng, H., Liu, G.L.: Tooth Contact Analysis of Face-gear Meshing. Mechanical Science and Technology for Aerospace Engineering 27, 92–95 (2008) 10. Vecchiato, D.: Tooth Contact Analysis of a Misaligned Isostatic Planetary Gear Train. Mechanism and Machine Theory 41, 617–631 (2006) 11. Shao, X.R.: Influence of Load Sharing Coefficient on the Manufacture and Assemble Error of the Planetary Gear NGW. Journal of Northeast Heavy Machinery Institute 18, 306–309 (1994) 12. Li, T., Pan, C.Y., Li, Q., Zhang, L.J.: Analysis of assembly error affecting on directing precision of spherical gear attitude adjustment mechanism. Acta Armamentarii. 30, 962– 966 (2009) 13. Chang, S.L., Tsay, C.B., Tseng, C.H.: Kinematic Optimization of a Modified Helical Gear Train. ASME Journal of Mechanical Design 119, 307–314 (1997) 14. Gu, J.G., Fang, Z.D., Pang, H., Wang, C.: Modeling and Load Analysis of Spiral Bevel Gears Power Split System. Journal of Aerospace Power 24, 2625–2630 (2009)

Research on SaaS and Web Service Based Order Tracking Jianhua Jiang1,*, Buyun Sheng1, Lixiong Gong2, and Mingzhong Yang1 1

School of Mechanical and Electronic Engineering, Wuhan University of Technology, Wuhan 430070, China 2 Chong Qing Automobile College, Chongqing University of Technology, Chongqing 400054, China

Abstract. To solve the order tracking of across enterprises in Dynamic Virtual Enterprise (DVE), a SaaS and web service based order tracking solution was designed by analyzing the order management process in DVE. To achieve the system, the SaaS based architecture of data management on order tasks manufacturing states was constructed, and the encapsulation method of transforming application system into web service was researched. Then the process of order tracking in the system was given out. Finally, the feasibility of this study was verified by the development of a prototype system. Keywords: dynamic virtual enterprise, software-as-a-service, web service, order tracking.

1 Introduction DVE takes outsourcing as the main business mode and it allocates customer orders to appropriate supplier by the way of bidding or negotiation. The real-time query of the status of order is named order tracking. Provide order tracking service in DVE helps for customers to participate in the manufacturing process of their orders so as to improve their loyalty to DVE. The service is benefit for the core enterprise of DVE to reallocate orders as to ensure orders completion. The research on order track is rare, such as for order tracking in discrete manufacturing oriented enterprise, an order execution process analysis and model method based on key node tracking of information in MTO environment was proposed [1]. To develop an order tracking system for auto assemblage, the hardware structure and software design based on barcode technology were discussed [2]. Mobile agent was used to develop online order tracking across global logistic alliances [3]. As to order tracking for alliance members within a supply chain, mobile agent was employed to develop a prototype of order tracking on the web [4]. Previous research is focus on single enterprise or alliance enterprise with stable membership. However, the existing solutions and systems is powerless for DVE for member enterprises join or leave dynamically and with distributed geographical location, heterogeneous hardware and software, different information level. So, this paper is to seek the method for order tracking in DVE. *

Corresponding author.

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 16–22, 2011. © Springer-Verlag Berlin Heidelberg 2011

Research on SaaS and Web Service Based Order Tracking

17

2 System Design 2.1 Problem Description After received orders, core enterprise of DVE will classify and divided them into subtasks and allocated to suppliers. Suppliers arrange production for orders and store the related data. In processing of order, customer can query its status and feedback the suggestion for the manufacturing method. In addition, core enterprise of DVE may reallocate orders according orders real-time status. The process is shown in figure 1. Because of suppliers’ different information system and distributed globally, how to collect manufacturing data from suppliers and provide order tracking on web is the main problem to be solved.

Fig. 1. Order management process

2.2 System Design SaaS provides online software and customers purchase service by need while the copyright belongs to software provider. The provider is responsible for maintenance, daily operations and technical support services to software [5]. Web serve is used to integrate and share data or application in heterogeneous platform. So, according application features of SaaS and technical characteristics of web service, a SaaS and web service based order tracking solution was designed as shown in figure 2. The system includes SaaS based manufacturing data management system, UDDI and order tracking & querying. Its idea is suppliers without information system uses SaaS service to store production date into manufacturing data DB. While suppliers have information system encapsulates it into web service and publishes it to UDDI. The order tracking & querying module is responsible for finding out the supplier ID from order allocated DB related to the order firstly. Then it acquires the data from manufacturing data DB ID directly or by invocating web service deployed in remote supplier enterprise.

18

J. Jiang et al.

Fig. 2. SaaS and web service based order tracking solution

3 Key Problem 3.1 SaaS Based Manufacturing Data Management Technically, developer of SaaS cares about the characteristics of multi-tenant, configurable and scalable of software. Multi-tenant refers to customers share a set of code while the data of different tenants is isolated from each other for safe and to meet customers’ personalized needs. Configurable means the tenants’ personalized requirements is realized by configurable operations without changing source code. Scalable makes it can support any scale of applications just by add hardware. According to the technical features of SaaS, the architecture of SaaS based manufacturing data management system is shown in figure 3. Sys. Management

SaaS based manufacturing data management system Presentation layer Collaboration Layer

IE Business rule

Portable terminal Workfolw control

Business component

Order management Order quote Order bid Order prewarning

Common component

Message service DB Access Message managemeng Security service

Mobile terminal

Instant messaging device

Customization and configuration

Role and permission

Quality management workshop management Material management Material purchase Production inspection Task assignmeng Material inspection Package inspection Process scheduling Input workload Material in & out Product in & out inspection Report management Product statistic report Order statistic report Material statistic report Warehouse report Product schedule report Supplier info. report Product quality report Workload report

Infrastructure

Data query Report service

Monitor service Interface service

Network/OS/DB/Server/Cluster

XML parse Log service

Tenant managent Cost management Application customization Account management Security management Role management Permission management System monitor Common management Sys. Maintenance Copyright management Backup management Log management Statistic analysis

Fig. 3. Architecture of SaaS based manufacturing data management system

Research on SaaS and Web Service Based Order Tracking

19

In the architecture, presentation layer is responsible for submitting user needs to business component and it’s the interface for user login. Collaboration layer provides customization and configuration service for different tenants to meet their personalized needs. Business component layer includes all the components needed for manufacturing data management while common component layer provides basic services for all tenants. Infrastructure includes the software and hardware for running the system. To ensure normal running of the system, the system management provides multi-tenant management, system security and maintenance functions. In SaaS mode, the isolation of system data, scalable and configurable of data structure must be taken in account. Data storage method including independent database, shard database and independent structure, shared database and shared structure. In this paper, the shared database and shared structure is chose. For the consistent data model is not suit for the storage method, a meta-data table and extended data table is added to extend the data model. The field and data-type customized by tenant is stored in meta-data table while the value of customized filed is stored in extended data table. The extended information is expressed by the relationship between customer information table, meta-data table and extended data table as shown in figure 4.

Fig. 4. Extension of data model by name-value pairs

In addition, it is important to keep the security of tenants’ data. The security includes the authorization of customer account and accessing database. To the security of customer account, centralized authentication is used. Software provider manages all customer accounts and manager of customer account is authorized to manage user account as shown in figure 5. As to the security of accessing database, MD5 encrypt algorithm and the three-tier architecture was used as shown in figure 6.

System manager

Create customer role

Allocate module for customer role

Account audit

Account & role binding

Apply account Create user role

Verify account N

Y

Inherit role module Apply account failed

Allocate module for user role

customer Create user account

Account & role binding

Inherit role module

Fig. 5. Account management process

customer Presentation layer

Encryption & decryption query module

request Encrypted DB return Web service of data query

Business layer

Data layer

Fig. 6. Data accessing method

20

J. Jiang et al.

The process of accessing encrypted database is customer inputs query to encryption & decryption query module in bushiness layer. It is transferred to data query web service and will be executed on encrypted database if the query is legal and the customer account is authorized. 3.2 Encapsulate Application System into Web Service Application system can be encapsulated into web service directly if it developed by component technology while the component must be extracted from application system if it is developed by object oriented of procedure oriented technology. Application system can be encapsulated into web service from three layer of data layer, business layer and application program. The encapsulation of data layer into web service is to operate database in application system directly by open data connection technology such as ODBC, OLEDB, JDBC, ADO.NET and ADO. In .NET platform, the ADO.NET can be used to access heterogeneous database. The data acquired by ADO.NET is stored in dataset as XML format in local machine. Data in dataset can be operated and updated to real data source or transferred as a parameter. Encapsulate business layer into web service depends on the development language of application. .NET platform is suitable for encapsulating system developed by language such as VB, C or C++ while Java Native Interface (JNI) is suitable for system developed in Linux platform. In .NET platform, the method to encapsulate managed DLL and native DLL is different. For managed DLL, just copy it to bin subdirectory in web service application, instantiating an object and invocating its method. Wile for native DLL such as COM, ActiveX or Win32 API, the p-invoke in .NET platform is used to encapsulate it into web service as shown in figure 7. The process is to load it into memory firstly, acquiring its address and pushing needed parameters into stack, then sends parameters to it and transfers the control to it. metadata Managed code

compiler

P-Invoke

MSIL code assembly

Native DLL DLL function

CRL

Fig. 7. Invocation process of native DLL

The method of encapsulating application program is added an interface for active the process of .exe file. The web service is only transfer the invocation to application without any return.

4 Prototype System To meet the needs of virtual business model, a platform named product customization of mechanical parts in network environment was developed for an enterprise. Its

Research on SaaS and Web Service Based Order Tracking

21

modules including 2D/3D production display on web, customer needs submit, need quoting, classify and allocate orders, supplier management and production management. For the disadvantages of complex tracking process, high cost of communications, constrained by working hours and slow response in current order tracking method by telephone, fax or email, the SaaS and web service based order tracking system was added to the platform and the prototype was developed. Process of order tracking in the prototype is in figure 8.

Fig. 8. Process of order tracking

Step 1 Customer submits an orderId to order tracking system by client IE. Step 2 The system acquired all the supplierId from allocated order information database by orderId and read out the data acquiring method of each supplier from supplier information database. Step 3 Acquire order state from SaaS based manufacturing data management system for suppliers without production management system. While for suppliers has production management system and has been encapsulated as web service, searching for the WSDL by supplierId from UDDI and invocating the web service deployed in remote supplier enterprise by URI in WSDL to acquire order state. Set 4 Merge the acquired order state if it comes from manufacturing data management system and web service respectively. Format it by XSLT and return the result to client IE of customers.

5 Conclusions As to order tracking across the boundary of enterprises, different in information level and heterogeneous in software & hardware, a SaaS and web served based solution was put forward. In realize the solution, the storage of customer data in SaaS mode was solved by adding metadata table and extend data table while MD 5 encryption

22

J. Jiang et al.

algorithm and three-tier architecture was used to ensure the security of customer data. To encapsulate application system into web service, the solutions from three layers of data layer, business logic layer and application program were researched. On basis of the research in this paper, a web based order tracking prototype was developed and it realized the order tracking across enterprises. Next, we will modify and improve it until to put it into practical application. The research in this paper has reference to the development of similar application system.

Acknowledgment It is a project supported by National Natural Science Foundation for the Youth Project of China (contract no. 50905133).

References 1. 2. 3.

4. 5.

Chen, X., Tang, R., Wang, Z.: Research on Order Track Manage System Oriented to Discrete Manufactruing. J. Light Industry Machinery 28, 111–115 (2010) (in Chinese) Chen, S., Liu, T.: Barcode Order Tracking System for Auto Assemblage. J. Journal of Wuhan University of Technology 27, 75–78 (2005) (in Chinese) Trappey, A.J.C., Trappey, C.V., Hou, J.L., et al.: Mobile agent technology and application for online global logistic service. J. Industrial Management & Data System 104, 168–183 (2004) Cheng, C.-B., Wang, C.: Outsourcer selection and order tracking in a supply chain by mobile agents. J. Coupter & Industrial Engineering 55, 406–422 (2008) Choudhary, V.: Software as a Service: Implications for Investment in Software Development. In: Proceedings of the 40th Hawaii International Conference on System Sciences, pp. 209–218. IEEE Press, Hawaii (2007)

Research and Analyses on Unsteady Heat Transfer of Inner Thermal Insulation Wall during Multi-temperature Refrigerated Transportation Guanghai Liu, Ruhe Xie, and Yongcai Sun Research Center of Modern Service Industry, Guangzhou University, Guangzhou, Guangdong, 510006, China

Abstract. There are lots of differences between multi-temperature refrigerated vehicles and ordinary ones in many aspects such as structure and running circumstances, hence the hugely different heat-transfer characters and response coefficients. This paper researched the unsteady heat transfer process of the multi-temperature refrigerated vehicle by response coefficients method, numerically calculated the unsteady heat transfer of inner thermal insulation materials in the multi-temperature refrigerated vehicle by computers, and studied the influences of different term numbers of equation root, term numbers of response coefficients on load calculation accuracy. It showed that the accuracy requirement could be met when the root of equation was -25 and the term number of response coefficient was 30. Keywords: Multi-temperature refrigerated transportation; Unsteady heat transfer; Response coefficients method.

1 Introduction Multi-temperature transportation is a kind of special refrigerated transportation methods that can transport goods of different kinds and storage temperatures simultaneously by dividing the refrigerated transportation into several areas with different temperatures. Since the concept multi-temperature transportation appeared in the end of last century, it has been noticed by experts and has been preliminarily applied rapidly in developed countries because such transportation can meet the small-lot, multi-variety requirement of market, repair the deficiency of traditional refrigerated transportation that food with different temperatures cannot be transportation simultaneously, lower the requirements about transportation burden and kinds of goods, reduce transportation, improve efficiency and cut down energy consumption while food quality is guaranteed during transportation. Considering heat transfer conditions between multi-temperature units can affect food quality and running energy consumption during multi-temperature transportation, the dynamic load model should be adapted in initial design to solve the problems above fundamentally. Moreover, unsteady heat transfer algorithm should be used to calculate the heat of thermal insulation materials in order to achieve the aim of design optimization and energy-saving. R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 23–29, 2011. © Springer-Verlag Berlin Heidelberg 2011

24

G. Liu, R. Xie, and Y. Sun

2 Determine the Model of Unstable Heat Transfer By the basic theory of heat transfer, when study unstable heat transfer problems of single homogeneous material (shown in fig.1), we can obtain the answer to solve the partial differential equations: ⎧ ∂t ( x,τ ) ∂t 2 ( x,τ ) = a⋅ ,0 < x < l ,τ > 0 ⎪ ∂x 2 ⎪ ∂τ ∂t ( x,τ ) ⎪ ,0 < x < l ,τ > 0 ⎨ q( x,τ ) = −λ ⋅ ∂x ⎪ t ( x , 0 ) = 0 ⎪ ⎪⎩

℃

(1)

And t was temperature in , q was heat flux in W/m2, α was thermal diffusivity of the wall materials in m2/h, λ namely thermal conductivity of the wall materials in W/m. , x was the wall calculation coordinate in m, τ namely time variable in h.

Fig. 1. Single homogeneous wooden partition

℃

Fig. 2. Multi-layer wooden partition

Laplace transform the variables x and τ in equation (1) that transform the original differential equations to algebraic equations, for x=l, re-use inverse Laplace transform to obtain the mathematical expression about the wall temperature distribution and heat flux distribution, shown in (2). ⎡T (l , s) ⎤ ⎡ch( s / al ),− sh( s / al ) /(λ s / a )⎤ ⎡T (0, s ) ⎤ ⎥⎢ ⎢Q(l , s)⎥ = ⎢ ⎥ ⎣ ⎦ ⎢⎣− λ s / a sh( s / al ), ch( s / al ) ⎥⎦ ⎣Q(0, s)⎦

(2)

And T(l.s) was the Laplace transform with time for temperature on l calculation section, Q(l,s) was the Laplace transform with time for heat flux on l calculation section, s was the Laplace transform with time. If given the boundary conditions, we can obtain the Laplace transform with temperature and heat flux in any position of wooden partition by (2), re-use inverse Laplace transform with T(l.s) and Q(l,s),then obtain the final solution. Because formula (2) only for single homogeneous material of the wall, the wall of multi-temperature refrigerated trucks be combination of multi-layer insulation materials(shown in fig.2), then the transfer matrix of thermal system wooden partition should include the air layer of inside and outside and multi-layer wooden partition etc.

Research and Analyses on Unsteady Heat Transfer of Inner Thermal Insulation Wall

25

⎡T (l , s) ⎤ ⎡1 − 1 / α n ⎤ ⎡ An ( s) − Bn ( s )⎤ ⎡ An −1 ( s ) − Bn −1 ( s) ⎤ LL ⎢Q(l , s) ⎥ = ⎢0 1 ⎥⎦ ⎢⎣− C n ( s) Dn ( s ) ⎥⎦ ⎢⎣− C n −1 ( s ) Dn −1 ( s ) ⎥⎦ ⎣ ⎦ ⎣

(3)

⎡ A1 ( s) − B1 ( s) ⎤ ⎡1 − 1 / α w ⎤ ⎡T (0, s) ⎤ ⎢− C ( s ) D ( s ) ⎥ ⎢0 1 ⎥⎦ ⎢⎣Q(0, s) ⎥⎦ ⎣ 1 1 ⎦⎣

And αn was the convection heat transfer coefficient of air inside of refrigerated trucks in W/m2. , αw was the convection heat transfer coefficient of air outside of refrigerated trucks in W/m2. . Ai ( s ) = ch( s / al i ) ; Bi ( s) = sh( s / ali ) / λ s / a ;

℃

℃

C i ( s ) = λ s / a sh( s / ali ) ; Di ( s ) = ch( s / al i ) .

，

For B(s)=0 when solution the heat transfer coefficient of wall obtain the solution ai of the transcendental equation. Used expand law Heaviside and superposition method to discrete time, we can receive the Y(j) of heat transfer coefficient in wall under the disturbance of the units triangular wave, shown in (4). ∞ Bi ⎧ − ai Δτ ) ⎪Y (0) = K + ∑ Δτ (1 − e ⎪ i =1 ⎨ ∞ ⎪Y ( j ) = − Bi (1 − e − ai Δτ ) 2 e − ( j −1) ai Δτ ∑ ⎪⎩ i =1 Δτ

j=0

(4)

j ≥1

℃

And Y(j) was the j reaction coefficient in W/m2. , K was the heat transfer coefficient of wooden partition in W/m2. , i was the amount of equation root, ai was the i root with B(s)=0 that transfer matrix with wooden partition thermal system. τ was the discrete time interval in h, Bi was the coefficient.

℃

Bi = −1 /[ ai B′(−ai )] 2

△

(5)

Then, the heat through the wooden partition can be determined by (6). ∞

HG ( n) = ∑ Y ( j )t z ( n − j ) − Kt n j =0

(6)

HG(n) was the heat transfer gains with n time in W/m2, tz(n) was the integrated temperature outside the car with n time in , tn was the air temperature inside the car in . In short, the essence of analysis method above was that the modern control theory be introduced and Laplace transform be adopted, which regards the wall as a thermodynamic system, the temperature boundary conditions changing with any time are decomposed into a series of triangular pulse, solved for the heat flux change caused by the external disturbance of units of temperature wave triangle in advance, which regard as the reaction coefficient of thermal system, and each pulse is linearly dependence, but the thermodynamic system itself is linear system. Based on the application of superposition principle, the heat flux change caused by each pulse is superimposed according to the convolution principle, namely, acquired the thermal condition of overall system. Due to a series of final, so this method is called reaction coefficient method. In the heat transfer of construction field, this method has been applied in a certain, and recognized as the optimization calculation method and the development direction of unstable heat transfer, but due to the great differences between construction and refrigerator in the

℃

℃

26

G. Liu, R. Xie, and Y. Sun

conditions of use, structure, material etc, thus the related knowledge of reaction coefficient method acquired in construction field can not directly apply in the refrigerator, which must concrete analysis according to specific problems.

3 Research on Heat Transfer of Heat-Barrier Material in the Multi-temperature Refrigerated Vehicles

（－ ）

From (4) to (6), the heat transfer reactivity coefficient Y(j) of the batter wall is the sum of an infinite index which is related to the root ai of the transition matrix element B(s). The number of Y(j) selected in the calculation directly affect the accuracy and the complexity level. At the same time, when the number of the heat transfer reaction coefficient get large enough, Y(j) and the index related to the root above mentioned will tend to be zero, so the problem of how to compute the capacity of heat transfer in unsteady state accurately can be translated into problems that how to select the range of the root value of B(s) element and what number of the heat transfer reaction coefficient should be calculated. In architecture, there are more researches on how to define the range of the root value about B(s) element, but the opinions are various for the types of the architecture and the accuracy standard are various: the ASHRAE says the accuracy can be guaranteed generally while calculating to the root is more than (-50) in air conditioning buildings[1]; references [2, 3] suggest that (-40) is a proper value; besides, Japanese HASP program says (-25) is appropriate while the Canadian program set the standard that (-100) is necessary[4,5]. In the aspect of the number about heat transfer coefficients, there are fewer studies since the relevant difference can be eliminated by adopting periodic reaction coefficient which depends on the nature that there are no area changes about the common buildings[6]. But this method is not suitable for the calculation of the heat transfer of the multi-temperature refrigerated vehicles’ wall because of its nature of moving all the time[7]. Therefore, to define the number of reaction coefficients is necessary in the calculation of the unstable heat transfer. Besides, how the special velocity nature impact the heat transfer nature of the vehicle while it is moving is also the aspect which should be considered in the analysis as well as the impaction level. To sum up, since the large difference between the vehicle and the common building, their heat transfer nature is different, and even larger between their reaction coefficients. All of these make the adopting the method of reaction coefficient should be given more comprehensive analysis on different velocities, different numbers of the equation root, different numbers of the reaction coefficient and any other variety elements according to its nature in the progress of the research on unstable heat transfer in multi-temperature refrigerated vehicles.

4 Reaction Coefficient Analysis and Determination 4.1 Research Parameters and Calculation The internal insulation structure of multi-temperature refrigerated truck is various. In order to keep universality in this research, three common insulation materials in our

Research and Analyses on Unsteady Heat Transfer of Inner Thermal Insulation Wall

27

country were chosen to analyze. (1)1.5mm Weathering Steel, 200mm Insulation and 1.5mm Alloy Steel (2)4mm Rubber Floor Cloth, 19mm Plywood, 4mm Glass Floor, 115mm Insulation, 2mm Alloy Steel (3)1.5mm Weathering Steel, 135mm Insulation, 1.5mm Alloy Steel. 4.2 The Precision Influence from the Number of Equation Root For comprehensive analyzing solution precision influence of B(s) element of the root value set the range of ai from (0) to (-10000) in the beginning. Through the study about multi-temperature refrigerated truck, we found just only the top 5 of the number of the B(s) element equation roots are able to greater impact the reaction coefficient. When the range is from ai to 10000, the value of Y(j) is standard value and obtain relative results after putting into the 4th equation. Compare the results with the Y(j) which result from the top 5 above. The Y(j)analysis curve of material 1 as picture 5. As the picture show, the less the number of Y(j) will be, the more sensitive the result to the number of the B(s) element equation roots will be and the more the number of roots increase, the less the influence will be. In the other hand, the better the thermal insulation of exterior-protected construction will be, the greater the influence to the B(s) element which depend on the number of equation root. For material 1, if only take a root, there are 32 reaction coefficients which exist more than 1.0 × 10-15% calculation errors. For material 2, there are 15 and for material 3, there are 11. The trend will still exist when the number of equation root become more. The top 5 of B(s) elements and relevant roots of the multi-temperature refrigerated truck exterior-protected construction are shown in Table 1. The ai of the B(s) element root in roof, wall and underneath of the truck should be set (-10), (-22), (-24) respectively and it will satisfy the requirement of calculation accuracy. For simplicity, the range of ai should be set to (-25) in calculation.

（－ ）

－ －

（－ ）

（－ ）

Table 1. the range with -ai of different materials equation root

i

Material 1

（-a ） i

（s）

B

Material 2

（-a ） i

（s）

B

Material 3

（-a ） i

（s）

B

1

-0.39638148 -0.68370221 -1.17013029 -0.49063269 -0.87937917 -0.53068790

2

-1.58547510 0.17344120

3

-3.56712075 -0.07902275 -10.34793579 -0.15963721 -7.90937811 -0.06893153

4

-6.34102397 0.04607084 -17.62221088 0.14606765 -14.05040327 0.04546968

5

-9.90670575 -0.03093426 -23.94561618 -0.11724395 -21.92145553 -0.03732472

-4.65790452

0.15978138

-3.51675725

0.14028678

28

G. Liu, R. Xie, and Y. Sun

Fig. 3. Error with different number roots

Fig. 4. The K with different materials

4.3 The Influence of Reaction Coefficient Number of Terms on Solution Precision

，

Using the basic properties of reaction coefficient Make the sum of former j-ary Series respectively is K(j), make(6)defined the former 100 series of the reaction coefficient is standard heat transfer coefficient, so by choosing the different number of terms reaction coefficient, the existence relative error δj is: δj =

K − K ( j) × 100% K

(7)

Where, δj is relative error by choosing the different number of terms reaction coefficient in energy consumption computation, %. K(j) is considered as heat transfer coefficient value by select the former n-ary series reaction coefficient in energy consumption computation, W/m2. Based on the foregoing analysis, take the root value of elements B(s) from ai to ( 25). By (7), we can calculate the relative error when various walls surface selects the different reaction coefficient, expressed as the image, As shown in Figure 4. The figure shows in respect to reaction coefficient number of terms when the number of items taken from less than 10, each additional one, errors were significantly reduced, but its trend is more and more slowly. As far as car roof, when the number of items are more than 16, error is less than one percent; When reaction coefficient number of terms take to 29, error is less than one ten-thousandth; For wall and underneath of the car, when reaction coefficient number of terms take to 15 and 13, error is less than one ten-thousandth. So, reaction coefficient take to 30th can sufficient to meet the requirements for multi-temperature refrigeration truck.

－

℃。

，

（－ ）

，

5 Summary Calculating the unsteady heat transfer of the exterior-protected construction by response coefficient method is an objective method to reflect the dynamic characters of

Research and Analyses on Unsteady Heat Transfer of Inner Thermal Insulation Wall

29

refrigerated vehicles. However, it has been only applied in construction area but refrigerated vehicles. Accordingly, this paper numerically calculated unsteady heat transfer of the exterior-protected construction by principles of response coefficient method and computer-aided analysis, studied the influences of various factors such as different velocities, different term numbers of equation root and different term numbers of response coefficients on load accuracy calculation of refrigerated vehicles. Here are the conclusions as follows.

（－ ）

（－ ）

(1) The range of ai should be set to 25 in the selection of v element's root. (2) When the term number of response coefficient was determined, it is reasonable to take values until the 30th term. (3) This paper improved the heat transfer response coefficient method, created situations for the application of this method in multi-temperature refrigerated transportation, and provided supports for the design optimization of multi-temperature refrigerated vehicles and the energy-saving of food refrigerated chain.

Acknowledgements The project was supported by the Natural Science Foundation of Guangdong Province (No.9451009101003189) and Natural Science Foundation of China (No. 51008087).

References 1. 2. 3. 4. 5. 6. 7.

ASHRAE. AHSRAE Handbook-Fundamentals (1997) Yan, Q., Zhao, Q.: Building thermal process. Construction Industry Press, Beijing (1986) Long, W.: Energy conservation of building and construction equipment. Construction Industry Press, Beijing (1990) Long, W., Zhao, J.: Computer simulation of annual dynamic thermal load in hig-rise building. Heating Ventilating & Air Conditioning 20(3), 25–28 (1990) Lian, Z.: Effect of number of root on loads calculated by response factor method. Journal of xi an University of Architecture & Technology 28(3), 245–247 (1996) Boyer, H.: Thermal building simulation and computer generation of nodal models. Building and Environment 31(3), 207–214 (1996) Liesen, R.J., Pedersen, C.O.: An evaluation of inside surface heat balance models for cooling load calculations. ASHRAE Trans. 103(1), 485–502 (1997)

Evaluation of the Industrial Economic Benefits Based on TOPSIS Baolu Wei*, Feng Dai, and Jingxu Liu Institute of Information Engineering, Information Engineering University, Zhengzhou China [email protected]

Abstract. At first, this article gives an introduction to the principle and evaluation procedures of TOPSIS. Secondly, Using TOPSIS method to assess 8 prefecture-level cities of a province of China based on the basis of statistical data. The evaluation results demonstrate that the application of TOPSIS method to Industrial economic benefits is Reasonable. TOPSIS has some practical value in the economic field, can provide scientific basis for the scientific management. Keywords: TOPSIS; Industrial economic benefits; Economic results.

1 Introduction Industrial enterprises are the pillar industry of national economy, the evaluation of economic benefit is not only an important part of industrial policy, but also an important responsibility of managers. How to find a practical, quantitative and accurate multi-objective comprehensive evaluation method in recent years is an research topics of common concern. This paper introduces systems engineering TOPSIS method, the economic efficiency of industrial enterprises a comprehensive evaluation, and achieved satisfactory results.

2 Principle and Steps of TOPSIS TOPSIS is Short for Technique for Order Preference by Similarity to Ideal Solution, The method is close to the ideal solution to the principle of selecting the best technical solution, the final evaluation results should be the ideal solution is the closest, and farthest away from the worst solution. Using this method, firstly should define the ideal solution for the positive and negative assessment, although both programs could not exist in the assessment (otherwise no assessment), but we both can be idealized as the best, the worst point, in order to other possible options for balancing the distance between the two. Secondly, general the Assessment factor as so to sort, evaluate and analysis. Overall, the algorithm process of the method is as follows: *

This work is supported by the National Social Science Foundation of China (No. 09&ZD014).

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 30–34, 2011. © Springer-Verlag Berlin Heidelberg 2011

Evaluation of the Industrial Economic Benefits Based on TOPSIS

31

Assume the property values from Ai to C j is yij ( i = 1, 2, L , n; j = 1, 2,L , m ) , and Y

⎡ y11 y12 L y1m ⎤ ⎢y y22 L y2 m ⎥⎥ is the initial evaluation matrix. Y = ⎢ 21 . ⎢L L O L ⎥ ⎢ ⎥ ⎣ yn1 yn 2 L ynm ⎦ 1. Standardized the evaluation matrix and comparing the index value, eliminate the impact of the dimensionless between different indicators, the attribute value normalization is as follows: If c j is the efficiency index, the greater the

index value, the better the results of the evaluation indicators. decrees zij =

yij

max { yij | 1 ≤ i ≤ n}

(1)

If c j is the cost-based indicator, the smaller the index value, the better the results of the evaluation indicators, decrees zij =

yij

max { yij | 1 ≤ i ≤ n}

(2)

If c j is the moderate type of indicators, to assess the indicators were most satisfied with is that for the more favorable assessment of indicators, decrees max = max { yij − a #j | 1 ≤ i ≤ n} ; min = min { yij − a #j | 1 ≤ i ≤ n} ;

zij =

2.

max − | yij − a #j | max − min

Units of the evaluation matrix, make the index value into the dimensionless quantity. Decrees rij =

zij

(4)

m

∑z j =1

3.

(3)

2 ij

Unit the weighted Matrix by Expert survey or AHP we can obtain the nomalization of the index weight vector. The weighted matrix is: ⎡ x11 ⎢x X = ⎢ 21 ⎢L ⎢ ⎣ xn1

L x1m ⎤ x22 L x2 m ⎥⎥ ; x = yij ⋅ w j ; L O L ⎥ ij ⎥ xn 2 L xnm ⎦ x12

32

B. Wei, F. Dai, and J. Liu

4.

Determine the positive and negative ideal solution, decree xi+ = max xij , state

； decree = { x , x ,L x } is the negative ideal solution；

x + = { x1+ , x2+ , L , xm+ } is the ideal solution x−

5.

− 1

− 2

i

xi− = min xij , state i

− m

Calculated the distance between positive and negative ideal solution, m

Li = ∑ ( xij − x +j )2

(5)

j =1 m

Di = ∑ ( xij − x −j )2

(6)

j =1

Li and Di are the closeness of the ideal solution to the positive and negative points of Ai . 6. Determine the assessment factor, comparative and analysis the program evaluation factor, define Ci =

Di Di + Li

(7)

Clearly, the closer the assessment factor Ci closer to 1, the closer the program Ai is the ideal solution; the closer the assessment factor Ci closer to 0, the closer the negative ideal solution.

3 Analysis of the Assessment Examples 3.1 The Choice of Evaluation Index System

According to the current of the actual situation of industrial development, comprehensive assessment of its economic benefits, to establish evaluation index system should include: X1—Labor productivity (yuan/person) X2—Net output rate (%) X3— Output rate of fixed assets (%) X4—Interest rate of fixed assets (%) X5—Capital tax rate (%) X6—Sales tax rate (%).

、

、

、

、

、

3.2 The Choice of Evaluation Index System

According to an annual 8 of province-level cities in all of independent accounting of the relevant statistical data calculated by the author of the evaluation order (see Table 1), the application of TOPSIS method for comprehensive evaluation of the economic benefits of industrial cities.

Evaluation of the Industrial Economic Benefits Based on TOPSIS

33

Table 1. 8 cities in a province on the table of economic benefit evaluation Indicators Labor productivity Net output rate Output rate of fixed assets Interest rate of fixed assets

A

B

C

D

E

F

G

H

14250

27362

12521

8560

26895

16850

17520

7950

49.97

81.52

64.35

52.64

85.89

82.61

79.20

74.76

52.42

75.64

58.64

68.50

78.10

65.72

49.67

64.51

11.24

28.56

31.00

9.68

20.78

19.85

24.69

10.83

Capital tax rate

15.64

19.88

20.69

19.67

20.13

11.35

9.99

9.97

Sales tax rate

12.35

20.10

17.35

16.85

19.99

12.34

15.66

11.31

The evaluation index system this title established has been meet the same trend, Directly normalized it, ⎡ 0.2817 ⎢ 0.5409 ⎢ ⎢ 0.2475 ⎢ 0.1692 Z =⎢ ⎢ 0.5316 ⎢ ⎢ 0.3331 ⎢ 0.3463 ⎢ ⎣⎢ 0.1571

0.2436 0.3975 0.3137 0.2566 0.4188 0.4013 0.3862 0.3645

0.2858 0.4124 0.3197 0.3734 0.4258 0.3583 0.2708 0.3517

0.1886 0.4794 0.5203 0.1625 0.3488 0.3332 0.4144 0.1818

0.3437 0.4369 0.4547 0.3730 0.4424 0.2494 0.2196 0.1971

0.2716 ⎤ 0.4421⎥⎥ 0.3816 ⎥ ⎥ 0.3706 ⎥ 0.4397 ⎥ ⎥ 0.2714 ⎥ 0.3445 ⎥ ⎥ 0.2488 ⎦⎥

Obtain the optimal value vector and the vector of worst values: Z + = ( 0.5409, 0.4188, 0.4258, 0.5203, 0.4547, 0.4221) ; Z − = ( 0.1571, 0.2436, 0.2708, 0.1625, 0.1971, 0.2488 ) ;

Calculated the distance between positive and negative ideal solution, take city A for example: 2 2 2 Da+ = ( 0.2817 − 0.5409 ) + ( 0.2436 − 0.4188 ) + L + ( 0.2716 − 0.4421) = 0.5186 ; Da− =

( 0.2817 − 0.1571)2 + ( 0.2436 − 0.2436 )2 + L + ( 0.2716 − 0.2488 )2

C = 0.1961* ( 0.5186 + 0.1961) = 0.2744 .

Calculate the remaining city as to A, detailed results see Table 2: Table 2. Industrial economic benefits of 8 results city A B C D E F G H

Di+

Di−

Ci

0.5186 0.0512 0.3347 0.5541 0.1722 0.3928 0.3728 0.6116

0.1961 0.6216 0.4770 0.2379 0.5712 0.3096 0.3595 0.1467

0.2744 0.9239 0.5877 0.3004 0.7684 0.4408 0.4909 0.1934

results 7 1 3 6 2 5 4 8

= 0.1961 ;

34

B. Wei, F. Dai, and J. Liu

From table2 we can see B, E and C city is the finest city, G, D, F, A and H city is the Medium city in the industrial economic benefits. There is no inferior city. It explains that the industrial economic benefits of the 8 cities’ development is Balanced.

4 Summary TOPSIS is a systems engineering approach for the comprehensive evaluation. In recent years, it has begun to be used for economic and health fields. The law of the original data with the trends and normalized, not only to eliminate the influence of different indicators of dimension, but also make full use of the original data can be quantitatively evaluate the pros and cons of different units level, the results of objective and accurate. In this paper, the development of the current status quo of industrial enterprises, the establishment of the industrial economic benefit evaluation index system, the method is introduced into the field of industrial economy, the economic benefits of comprehensive evaluation of multiple units, results in good agreement with the actual situation.

References 1. 2. 3.

Yue, C.-y.: Decision-making theory and approach. Science Press, Beijing (2002) Xu, Z.-s.: A new method of multiple objective decision under uncompleted information. Operations and Management 10(2), 25–28 (2001) Olson, D.L.: Comparison of weights in TOPSIS models. Mathematical and Computer Modeling (40), 721–727 (2004)

Optimization and Distributing Research of Measuring Points for Thermal Error of CNC Machine Based on Weighted Grey Relative Analysis Qin Wu 1,2, Jianjun Yang 1,2, Zhiyuan Rui 1, and Fuqiang Wang1 1

School of Mechanical and Electronical Engineering, Lanzhou University of Technology, Lanzhou, 730050, China 2 Key Laboratory of Digital Manufacturing Technology and Application, The Ministry of Education, Lanzhou University of Technology, Lanzhou, 730050, China [email protected], [email protected], [email protected], [email protected]

Abstract. The optimization of modeling method and distributing research of choosing the measuring points for thermal error of CNC machine were presented based on the theory of weighted grey relative analysis in the article. The model of weighted grey relation was set up firstly, then, by calculating the matrix of grey relation between the temperature data serial of thermal points and thermal error data serial, after classify the thermal points which placed on the CNC machine, 5 critical points which should be in different classes were picked out from whole16 temperature points according to the degree of grey relation. At last, the thermal forecasting model was founded using the 5 critical points. The model was validated to be more accurate and robust. Keywords: thermal error; weighted grey correlation; machine center; measuring points; modeling.

1 Introduction As the components of CNC machine tools being uneven heated in operation, so the distribution of temperature field of machine tools being complex, the resulting thermal deformation will cause the change of relative position between cutting tools and workpiece , it can be said that the thermal error is the largest error source of CNC machine tool. According to the statistics, the thermal error in the proportion of the total error of machine tool can reach 70% [1]. So how to improve the prediction accuracy and robustness of the error model, become the issue that so many scholars from different countries developing the large number of fruitful research. Chen J. S. of University of Michigan integrated the geometric error and thermal error, defined 32 new machine errors, and conducted the error model of machine tool by using multiple regression analysis[2]. Yang Jian-Guo of Shanghai Jiao Tong University proposed the means of thermal error modeling robustly based on the thermal modal analysis, and R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 35–40, 2011. © Springer-Verlag Berlin Heidelberg 2011

36

Q. Wu et al.

established the thermal error model applying multiple linear regression method and the gray correlation method [3]; Du Zheng-Chun, etc. advanced the thermal error compensating and modeling method based on Radial Basis Function (RBF) neural network for CNC lathe. Although these analysis methods improved the identification speed and accuracy of thermal error, but the prediction accuracy and generalization ability of the model is not high, and it is difficult to achieve the requirements of precision machining or ultra-precision machining. As the temperature field with non-stationarity and time variability, its influencing factors are complex. In order to avoid too little temperature measuring points leading to the lack of useful information, it is required to arranging a large number of temperature sensors on the machine tools. But these measuring points not only increases the workload of measuring and calculating data, but also coupling could occur between the nearby measuring points. So it requires to selecting the critical temperature measuring points for thermal error modeling. As the feature of thermal error is obviously gray, while the measurement of limited temperature measuring points have the character of small sample and poor information relative to the overall temperature distribution of the machine tool, so the article will establish the weighted gray correlation analysis model based on the measured temperature data, and select the critical temperature measuring points for the establishment of thermal error prediction model.

2 The Model of Weighted Grey Relative Analysis Grey relative analysis does not require a large sample of the typical distribution of the data column. In the experimental data processing of each temperature measuring point, the analysis method can use very little experimental data. The key of Grey relative analysis is the establishment of Grey correlation, and the Grey relative grade is quantum that reflects the correlative degree among each factor [4]. However, the formula which commonly calculating the correlative coefficient and the relative grade only focusing on the area between the two curves to identify semblable extent, ignoring the change trend of the curves, and not considering the weighted difference of all factors. In this paper, the weighted coefficient will be added based on the optimization of the sample data sequence, this method makes the results better reflect the proximity degree between the two curves. Make X = {xθ θ = 0,1,2,L, m} as the relative factor aggregate of the data sequence, x0 as the reference function, xi as the comparison function, i = 1,2,3, L , m , xθ (k ) for the value of xθ in No. k points, where k = 1, 2,3, L , n . In order to obtained Increment information and expanded density of the information in the modeling, the data sequence had Inverse Accumulating Generation Operator (IAGO). Set up y = ( y(1), y(2), y(3),L y(n)) as the inverse accumulating generated sequence of the sequence x = ( x (1), x (2), x (3), L x( n)), where y (1) = x (1), y (k ) = x (k ) − x (k − 1),, (k = 2, 3, L n) .

Optimization and Distributing Research of Measuring Points

37

对于 x , x ，令： 0

i

ξ ⋅ max max x0 (k ) − xi (k ) i∈m, k∈n

ζ i (k ) =

λ1 x0 (k )− xi (k ) + λ2 y0 (k )− yi (k ) + ξ ⋅ max max x0 (k ) − xi (k ) i∈m

k∈n

i∈m

i∈m

k∈n

(1)

k∈n

the gray relational coefficient x i of x 0 at the time of k, ξ for the differentiate ratio, 0 < ξ < 1 , λ1 for the weighted displacement coefficient, λ 2 for the In the formula (1),

ζ i (k ) for

且

weight coefficient of rate of change, λ1 , λ 2 ≥ 0, λ1 + λ 2 = 1 . In practice, all coefficients can be appropriately adjusted according to the emphasize particularly on the specific issues. According to the formula (1), it met the definition of the gray relational space. The calculating formula of the gray relational grade given as follows [5]:

γ i = γ ( x 0 , xi ) =

1 n ⋅ ∑ ζ i (k ) n k =1

(2)

In the formula (2), γ i for the gray relative grade that xi of x 0 . 2.1 Experimental Analysis of Examining the Thermal Error of CNC Machine In this experiment, a vertical machining center was regarded as the subject investigated, each measuring point and the spindle thermal error data were collected. As temperature field of the tool is subject to many factors, in order to fully detect the changes in temperature field, 16 temperature sensors were disposed on the machining center in the experiment, their distribution was as follows: No.1, 2 sensor were placed at the spindle box near the spindle bearings location; No.3, 4 sensors were placed at the spindle nose; No.5, 6, 7 sensors were placed at the rolling guides of X, Y, Z axes; No.8, 9, 10 sensors were placed separately at the ball screws of X, Y, Z axes close with screw nut; the 11th sensor was placed at the front uprights; the 12th sensor to monitor temperature changes of surroundings;No.13, 14 sensors were placed on both sides of the bed; the 15th sensor was placed on the table; the 16th sensor was placed at the cooling fluid box under the bed. Three non-contact eddy flow displacement sensors were used to measuring thermal error of the vertical machining center. They were installed in specially designed experimental device, arranged around the spindle, were respectively used to measure the thermal drift error of the spindle at the direction of the X, Y and Z axes [6]. The experiment adopted the style of dry run to simulate the cutting cycle. The rotating speed of spindle was 3000r/min; the feed speed was 25m/min; and cooling fluid keeping circulation. Sampling time was 5 hours of continuous operation; the sampling interval was 15 min. The curve diagram changes along with time of the temperature of each measuring point and the thermal error at the direction of z axis separately shown in Fig. 1 and Fig. 2.

38

Q. Wu et al. 36

0

34

-5 9 7 16

1

3 2

11

Error/um

4

32

T/

5 10

8

30

6

13 15

-10

-15

28 14 12

-20 26

24

0

50

100

150

200

250

300

-25

t/min

0

50

100

150

200

250

300

t/min

Fig. 1. The curve diagram changes along with Fig. 2. The curve diagram changes along with time of the temperature of each measuring time of the thermal error at the direction of z axis point

2.2 Establishment and Analysis of the Weighted Grey Relative Model Analysis of the Experimental Data. Make the data sequence of the temperature measuring points as a comparison function of the gray system (son-factors), and makes the data sequence of the thermal error at z direction of the spindle as a reference function (mother-factors), each data series takes 11 time nodes. The data were processed to dimensionless first, and substituted into equation (1), then calculated the relative coefficient between the comparative series xi (k ) i=1,2,3,…,16, k=1,2,3,…,10 and the reference series x0 (k ) .The gray relative coefficient matrix B formed as follows:

（

⎡0.8721 ⎢0.8721 ⎢ ⎢0.8721 ⎢ ⎢0.8721 ⎢0.8721 ⎢ ⎢0.8721 ⎢0.8721 ⎢ B = ⎢0.8721 ⎢0.8721 ⎢ ⎢0.8721 ⎢ ⎢0.8721 ⎢0.8721 ⎢ ⎢0.8721 ⎢0.8721 ⎢ ⎣⎢0.8721

）

0.9041 0.9773 0.8651 0.8016 0.7686 0.7760 0.7515 0.6715 0.6036 0.5553 ⎤ 0.9043 0.9769 0.8654 0.8011 0.7680 0.7759 0.7518 0.6713 0.6037 0.5556 ⎥⎥ 0.9043 0.9769 0.8654 0.8012 0.7678 0.7756 0.7508 0.6706 0.6030 0.5549 ⎥ ⎥ 0.9042 0.9769 0.8659 0.8022 0.7691 0.7766 0.7525 0.6717 0.6039 0.5557 ⎥ 0.9039 0.9774 0.8666 0.8028 0.7705 0.7784 0.7540 0.6732 0.6051 0.5567 ⎥ ⎥ 0.9039 0.9772 0.8669 0.8030 0.7707 0.7787 0.7540 0.6731 0.6052 0.5569 ⎥ 0.9039 0.9773 0.8668 0.8031 0.7706 0.7780 0.7538 0.6734 0.6053 0.5572 ⎥ ⎥ 0.9036 0.9779 0.8671 0.8037 0.7720 0.7799 0.7555 0.6743 0.6060 0.5574⎥ 0.9036 0.9780 0.8671 0.8036 0.7701 0.7773 0.7530 0.6725 0.6046 0.5562⎥ ⎥ 0.9042 0.9766 0.8662 0.8026 0.7697 0.7773 0.7530 0.6723 0.6044 0.5561⎥ ⎥ 0.9034 0.9079 0.8674 0.8033 0.7710 0.7795 0.7557 0.6753 0.6071 0.5585⎥ 0.9034 0.9066 0.8673 0.8034 0.7707 0.7791 0.7556 0.6752 0.6071 0.5583 ⎥ ⎥ 0.9037 0.9052 0.8669 0.8031 0.7709 0.7798 0.7557 0.6756 0.6075 0.5590⎥ 0.9035 0.9040 0.8677 0.8038 0.7715 0.7804 0.7562 0.6761 0.6079 0.5591⎥ ⎥ 0.9027 0.9011 0.8697 0.8065 0.7752 0.7848 0.7605 0.6793 0.6106 0.5615⎦⎥

The relative coefficient substituted into equation (2), gives the relative grade that were normalized between each temperature measuring point and the thermal errors:

Optimization and Distributing Research of Measuring Points

39

， ， ， ， ， ， = 0.8906，γ = 0.9688，γ = 1，γ = 0.8438，γ = 0.7969，γ = 0.0313， = 0，γ = 0.0156，γ = 0.0469，γ = 0.3594

γ 1' = 0.6875 γ 2' = 0.6719 γ 3' = 0.6250 γ 4' = 0.75 γ 5' = 0.8750 γ 6' = 0.8438 γ

' 7

γ 13'

' 8

' 14

' 9

' 10

' 15

' 11

' 12

' 16

After the normalization, values were all dimensionless quantities [7]. Application and Validation of the Model. Generally, the measuring points which were arranged in the same parts should be united a class, and the number of selected critical temperature measuring points should be as much as possible with the number of key components of machine tools. In this case, measuring points can be divided into six categories, namely: (1,2,3,4); (5,6,7); (8,9,10); (11); (12,13, 14,15); (16). In each category, choose the critical temperature measuring point which with the maximum relative grade. Finally, No.4, 7, 9, 11 and 16, the five sensors were chose as the key temperature measuring points for modeling the thermal error, they corresponded to the five locations such as the spindle end, z-axis rolling guide, y axis screw nut, column and cooling tank. Finally, based on 16 temperature measuring points and the selected 5 critical temperature measuring points, the modified GM modeling approach was used for the thermal error modeling, the results shown in Fig.3 and Fig.4. Obviously, modeling with the 16 temperature measuring points, the coupling with the data of each temperature measuring point, affected the accuracy of the model; and modeling with the finally identified five key temperature measuring points that optimized by the weighted gray relative analysis, the accuracy of the model has been greatly improved, the fitting precision of the curve was relatively high, residual error was small.

Fig. 3. The thermal error model based on 16 temperature variables

Fig. 4. The thermal error model based on 5 temperature variables

40

Q. Wu et al.

3 Conclusion In this study, the data that collected by data acquisition system had gray nature relative to the thermal distribution of the overall machine tool, and the data columns with small samples and uncertain typical distribution. It was very suitable to analyzing and modeling the thermal error with the gray system theory [8]. Using the model of weighted gray relative analysis before-mentioned, the distribution of temperature field that affected the thermal error of machine tool was optimized. Validated by the modeling, the forecast precision of the thermal error model was quite high.

Acknowledgements It is a Special project supported by the national science and technology Program of China (2009ZX04001-015).

References 1. 2.

3.

4. 5.

6.

7.

8.

Jun, N.: A Perspective Review of CNC Machine Accuracy Enhancement through Real-time Error Compensation. China Mechanical Engineering 8, 29–33 (1997) Hong, Y., Jun, N.: Dynamic neural network modeling for nonlinear, non-stationary machine tool thermally induced error. International Journal of Machine Tool Manufacture 45, 455–465 (2005) Jianguo, Y., Weiguo, D.: Grouping Optimization Modeling by Selection of Temperature Variables for the Thermal Error Compensation on Machine Tools. China Mechanical Engineering 15, 10–13 (2004) Youxin, L., Longting, Z., Min, L.: The Grey System Theory and The Application in The Mechanical Engineering. Press of National University of Defense Technology (2001) Yongxiang, L., Hengchao, T., Jianguo, Y.: Application of Grey System Theory in Optimizing the Measuring Points of Thermal Erron on Machine Tools. Machine Design & Research 2, 78–81 (2006) Jiayu, Y., Hongtao, Z., Guoliang, L.: Optimization of Measuring Points for Machine Tool Thermal Error Modeling Based on Grouping of Synthetic Grey Correlation Method. Journal of Hunan University (Natural Sciences) 35, 37–41 (2008) Jiayu, Y., Hongtao, Z., Guoliang, L.: Application of a New Optimizing Method for the Measuring Points of CNC Machine Thermal Error Based on Grey Synthetic Degree of Association. Journal of Sichuan University (Engineering Science Edition) 40, 160–164 (2008) Bryan, J.B.: International Statures of Thermal Error Research. Annals of CIRP 39, 645–656 (1990)

A DS-UWB Cognitive Radio System Based on Bridge Function Smart Codes Yafei Xu1,2, Sheng Hong1, Guodong Zhao1, Fengyuan Zhang2, Jinshan Di1, and Qishan Zhang1 1

Beihang University, XueYuan Road No.37, HaiDian District, BeiJing, China [email protected], [email protected] 2 Beijing University of Chemical Technology 15 BeiSanhuan East Road, ChaoYang District, Beijing, China

Abstract. This paper proposes a direct-sequence UWB Gaussian pulse of cognitive radio systems based on bridge function smart sequence matrix and the Gaussian pulse. As the system uses the spreading sequence code, that is the bridge function smart code sequence, the zero correlation zones (ZCZs) which the bridge function sequences' auto-correlation functions had, could reduce multipath fading of the pulse interference. The Modulated channel signal was sent into the IEEE 802.15.3a UWB channel. We analysis the ZCZs's inhibition to the interference multipath interference (MPI), as one of the main system sources interferences. The simulation in SIMULINK/MATLAB is described in detail. The result shows the system has better performance by comparison with that employing Walsh sequence square matrix, and it was verified by the formula in principle. Keywords: UWB; Bridge function; Smart code; BER; Cognitive radio; MPI.

1 Introduction Recent years, international academia and IEEE standardization organizations are more and more interested in cognitive radio technology, calling it as the next big thing in the future [1] [2] [5], ultra-wideband (UWB) wireless communication systems is the first way in the level of technical and practical steps to achieve the cognitive radio [3][4][6]. UWB is Mainly used in intensive multipath indoor channels. Multipath interference (MPI) is the main source of interference in systems [3]. In the third generation mobile communication technology, mainly in code division multiple access technology (wave division technology) [7]. In traditional CDMA systems, multiple encoding are widely used with orthogonal codes and PN code, limited by the Walsh matrix, the traditional CDMA system is a strong self-interference system, people began to study all have a zero correlation zones (ZCZs) and low correlation zones multiple access codes [7], that is, smart code sequence, which can reduce the multipath fading of the pulse interference. R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 41–46, 2011. © Springer-Verlag Berlin Heidelberg 2011

42

Y. Xu et al.

This paper has the following main aspects. Section II describes the bridge function smart code sequence matrix and the nature of autocorrelation. Section III elaborates the SIMULINK / MATLAB system model. Section IV gives the simulation results and makes a qualitative description. Section V is the conclusion.

2 Bridge Function Smart Code Sequence Matrix Sequence with the following characteristics could be called smart sequences [7]: autocorrelation function ⎧1 τ = 0 Raa (τ ) = ⎨ ⎩0 τ = 1

in a small intervalτ 0

(1)

cross-correlation function, to some zone τ 0 ,all of the τ

Rab (τ ) = δ

(2)

Where δ far less than 1. Paper [4] discusses the first sort of the bridge function correlation which is first copying and after shifting, obtains a class of smart code sequences, and received the following two theorems, quoting as follows: Theorem 1 (1) Each sequence in the group B r i q , k ( m ) , have the same zero-correlation zone

length, when 0 < τ < 2k , RBri ( m ) (τ ) = 0 (2) Any two different sequences in the group have the same length of zero correlation zone, when 0 ≤ τ ≤ 2 k , RBri ( m ), Bri ( n ) (τ ) = 0 By the definition of smart code, we know that the bridge function sequences in the group is a special kind of smart code sequence. Theorem 2

(1)Each sequence in the group B r i q , k ( m ) , have the same zero-correlation zone length, when 0 < τ < 2k , RBri ( m ) (τ ) = 0 (2) Any two different sequences in the group, it’s cross-correlation function identically zero RBri ( m ), Bri ( n ) (τ ) = 0

A DS-UWB Cognitive Radio System Based on Bridge Function Smart Codes

43

3 System Structure of the DS-UWB The diagram of DS-UWB system model for one user is shown

Fig. 1. The diagram of cognitive radio system model for one user

In the UWB pulse generator, Gaussian pulse waveform was generated. The output single pulse signal is then modulated into modulation. Signal model The transmission signal can be expressed as follows

s (t ) =

∞

∑d

n =−∞

⎡n⎤ ⎣⎢ G ⎦⎥

c n p (t − nTc )

(3)

p (t ) is the normalized pulse, and the duration is Tc . cn is the spreading code, and the period is G , symbol period is Ts = GTc , so G is the spreading gain. Sending the modulation signal into the standard of UWB Channel(SV/IEEE 802.15.3a) By the SV / IEEE 802.15.3a channel model given [3] L −1 K −1

h(t ) = ∑∑ α k ,l (t − Tl − τ k ,l )

(4)

l =0 k =0

where L is the number of clusters, K is the number of cluster-rays,

α k ,l is the multi-

path gain, T l is delay time in cluster l . τ k ,l is the relative delay from the section

k rays in cluster l Received signal,

r (t ) = s (t ) * h(t ) + n(t )

(5)

n (t) is additive white Gaussian noise which the mean is 0, and bilateral power spectral density is N 0 / 2 .

44

Y. Xu et al.

Literature [4] also described as multi-path interference

Eu [nIPI ,u (i ' ) + nISI ,u (i ' )] Where

nIPI ,u (i ' ) is inter-symbol interference, nISI ,u (i ' )

(6) is Inter- Pulse interference.

We can see that multipath interference is made up by every multipath interference, each of the multipath interference is divided into two. Based on the definition of IEEE 802.15.3a channel model, the energy of the multipath components is in accordance with the exponential rate of decay, By literature [5], we know that the variance of multipath interference is

σ 2 MPI =

1 G2

N −1

M −1

∑ ∑

l =0 k =0 M + K ≠m

l 2 (m − M − k )) le − M Δτ / Γ e − k Δτ /γ ( Rc 2 (m − M − k ) + R c

(7)

l (τ ) = R (T − τ ) Rc (τ ) is the Spreading code autocorrelation, R c c s Finally, from the BPSK error rate formula, we can obtain the m’th synchronization of multi-path matched filter’s receiver error probability ⎡ ⎤ 2 Em Pe = Q ⎢ ⎥ 2 ⎣ σ M PI + N 0 / 2 ⎦

(8)

E m is the energy of each reception symbol in the multi-path and

Q(a) = ∫

∞

a

1 − x2 / 2 e dx 2π

(9)

By judging and error rate calculation block, we could get the bit error rate.

Because of Rc (τ ) is the spreading code autocorrelation, with the smart code sequence value k value increases, the zero correlation zones begin to get longer. By the formula (9), we can see that the bit error rate performance getting better and better.

4 Simulation Results The system simulation structure is modeled in Simulink/Matlab software environment. UWB channel impulse response has four types (CM1, CM2, CM3 or CM4). In our simulation, we select the CM2.The transmitted binary data is set to be 10000 bits with symbol period of 4ns. Because the simulation results depend on the channel impulse response of a specific implementation, the channel influence should be correctly amended by averaging a number of the channel realizations. However, this will bring more time costs. Therefore, we make an appropriate balance between impact of the channel and time costs. The simulation results were showed in figure 2, 3. Figure 2, using bridge function smart code sequence matrix which belongs to theorem 1, and parameters is from k=1 to k= 4. Figure 3, using special bridge function smart code sequence matrix which

A DS-UWB Cognitive Radio System Based on Bridge Function Smart Codes

45

belongs to theorem 2, and parameters is from k=1 to k=4. It can be seen from the figure that using the bridge function smart code sequence matrix system has the better bit error rate performance than using Walsh function sequence. And with the values of k in the bridge function sequence increases, the zero correlation zone sequences of smart sequences get large too, and the systems have better system performance. This is because the bridge function smart code autocorrelation function, the existence of zero correlation zones, can effectively reduce the multipath interference. From the paper [7] we can see that the IEEE802.15.3a channel’s multipath interference is determined by the spreading code cycle, multipath gain and code spreading sequence code autocorrelation function .And when the channel parameters and the spreading sequence code determines, the spreading code’s cycle, multipath gain is determined too. Therefore, multipath interference is mainly determined by the correlation function of the spreading code sequences. With the increase of zero correlation zone of the correlation function, by the formula (9), we know that system has better BER performance. 0

10

-1

Bit Error Rate

10

-2

10

-3

10

Walsh With zeors padding Walsh(8) Bridge(3,1) Bridge(3,2)

-4

10

Fig. 2

0

5

10

15

Eb/No(dB)

Fig. 2. Using bridge function smart code sequence matrix which belongs to theorem 1 and Fig. is2 from k=1 to k=4 q = 4, and parameters 0

10

-1

Bit Error Rate

10

-2

10

Walsh(8) With one zero padding Walsh(16) Bridge(4,1) Bridge(4,2) Bridge(4,3)

-3

10

0

5

10

15

Eb/No(dB)

Fig. 3. Using special bridge function smart code sequence matrix which belongs to theorem 2 and q = 4, and parameters is from k=1 to k=4

Fig. 3 Fig. 3

46

Y. Xu et al.

5 Conclusion This paper proposes an ultra wideband (DS-UWB) cognitive radio system model witch make the bridge function smart code as the spread spectrum sequences. And compared bit error rate (BER) which use the Walsh sequence as the spreading code. The non-periodic autocorrelation function of each sequence that in the bridge function smart code sequence matrix has the zero correlation zones (ZCZS). Using the bridge function smart code sequence matrix, and chose the appropriate value of k, we can get a better anti-multipath fading and BER performance. The simulation results also verify this conclusion. This paper presents a single-user system, and verifies the zero correlation zones (ZCZS) of bridge function smart code sequence autocorrelation function have a good inhibition to the multipath interference. We know that, in the Code Division Multiple Access(CDMA) system, wish to use small cross-correlation function of sequences, ideally, the correlation function is to 0, that is, two code sequences is orthogonal, so that we can distinguish between different users, so smart code sequence in the communication system has broad application prospects.

Acknowledgments This work is supported by the Fundamental Research Funds for the Central Universities under grant No. YWF-10-02-023. China.

References 1.

2.

3.

4.

5.

6.

7.

Hong, S., Liu, K., Qi, Y.: A new direct-sequence UWB transceiver based on Bridge function sequence. In: 2010 Second International Conference on Computational Intelligence and Natural Computing, September 13-14, pp. 209–212 (2010) Di, J., Hong, S., Zhang, Q.: An UWB Cognitive Radio System Based on Bridge Function Sequence Matrix and PSWF Pulse Waveform. In: 2010 Second International Conference on Computational Intelligence and Natural Computing, September 13-14, pp. 162–165 (2010) Fisher, R., Kohno, R., Ogawa, H., Zhang, H., Takizawa, K., Mc Laughlin, M., Welborn, M.: DS-UWB physicallayer submission to 802.15 task group 3a, pp. 15–40. IEEE P, Los Alamitos (2005) Shaterian, Z., Ardebilipour, M.: Direct Sequence and Time Hopping ultra wideband over IEEE.802.15.3a channel model. In: 16th International Conference on Telecommunications and Computer Networks, SoftCOM 2008, September 25-27, pp. 90–94 (2008) Yang, Z., Qian, Y., Li, Y., Bi, X.: IEEE 802.15.3a channel DS-UWB multipath interference analysis. Journal of Natural Science of Heilongjiang University 24(1) (February 2007) Sablatash, M.: Adaptive and Cognitive UWB Radio and a Vision for Flexible Future Spectrum Management. In: 10th Canadian Workshop on Information Theory, CWIT 2007, June 6-8, pp. 41–44 (2007) Zhang, F., Zhang, Q.: A new type of smart code sequence and the correlation function. Journal of Telemetry;Tracking and Command (September 2005)

Problems and Countermeasures of Zhejiang High-Tech Enterprises Industry-University-Institute Cooperation in China Qing Zhou1, Chong-Feng Mao2, and Lin Hou1 1

Institute of Management Decision and Innovation, Hangzhou Dianzi University, Xiasha Higher Education Zone, Hangzhou, Zhejiang P.R. China [email protected] 2 School of Business, Central South University, 932, Yuelu South Road, Changsha, Hunan P.R. China [email protected]

Abstract. Industry-university-institute cooperation is an important means to accelerate technical development and achievements for high-tech enterprises. Considering that Zhejiang high-tech enterprises existed some problems which included low cooperative level, single distribution, weak secondary R&D ability, obvious risk and so on, government should play an guiding role on improving information service system, enhancing cooperative level, promoting scientific intermediary service organization system construction, and building better environment for Industry-university-institute cooperation. Keywords: High-tech enterprises; Industry-University-Institute Cooperation; Problems; Countermeasures.

1 Introduction Industry-university-institute cooperation is valuable experience getting from the economic and technical development. At the moment, Industry-University- Institute cooperation has become the strategic measure which speeds up technical development and achievement, increases overall national strength, and strengthens economic competitiveness. In 2008, Chinese Industry-University-Institute cooperation technology development contract amounted to 900 billion yuan, accounting for 37% of technology market transactions in total contract amount. The whole nation has established 8 Industry-University-Institute strategic alliances, 129 productivity promotion centers, 67 university science parks, with 1115 colleges and universities participated in the research cooperation. In recent years, industry-university-institute cooperation has enjoyed the rapid development, but also some new problems; many researchers on these issues conducted analysis and discussion from all sides. Cohen, Nelson and Walsw (2002) found that not only the formal cooperation was very important, but informal cooperation was also very important, even more important [1]. Monjon and Waelbroeck (2003) considered that commissioned R&D in a variety of R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 47–52, 2011. © Springer-Verlag Berlin Heidelberg 2011

48

Q. Zhou, C.-F. Mao, and L. Hou

industry-university-institute cooperation technical innovation model be in a more important position[2]. Zhong Wei-jun, Mei Shu-e, Xie Yuan-yuan(2009) obtained conclusions that among all the technical innovation modes in the industry-university-institute cooperation, joint and cooperation R&D is in an absolutely dominant position, commissioned R&D and consultation also play an important role, while transforming of the science research and technology development achievements and new ventures are in a secondary position[3]. Mao Jia-qiang, Liu Lu (2009)concluded that in the combination of industry study and research there are such problems as the improper enterprise understanding the low turning ratio of technology the shortage of the research and development funds the imperfect planning and coordination mechanism and the imperfect laws regulations and policies. In the development of industry study and research, governments should make strenuous efforts to support the constructing of regional innovation group of combination of industry study and research should perfect the related policies laws and regulations[4]. This paper used the benchmarking test audit model designed by Chiesa et al 1996 to analyze measured problem[5]. Empirical data originated from “Survey of Zhejiang Enterprises Industry-University-Institute Cooperation in China” organized by Zhejiang Province Science and Technology Department, which obtained 817 sample enterprises. By controlling the variables in the total sample as analysis of high-tech enterprises to extract the object, we obtained 111 researched high-tech enterprises. During empirical analysis, this paper analyzed problems of industry-university-institute cooperation in Zhejiang high-tech enterprises from motivation, the main influencing factors and so on, and gave some countermeasures to solve these problems.

，

，

，

，

，

（ ）

，

，

， ， ，

2 Problem Analysis Between 2006 and 2008, the average operating income of Zhejiang high-tech enterprises was 513.5 million yuan. The average R & D investment was 1327 million yuan, accounting for 2.5% of the average operating income. The average cooperative R & D investment was 5.15 million yuan, accounting for 0.97% of the average operating income. About 86 enterprises hammered at industry–university-institute cooperation, and the total item was 351 projects, nearly 3 projects per enterprise. These data suggested that industry-university-institute cooperation of Zhejiang high-tech enterprises had a boom trend, industry-university-institute cooperation become an important way of Zhejiang high-tech enterprises technical innovation. But through questionnaires and interviews, we also found there were many problems. Firstly, industry-university-institute cooperation still remains at a low level. The main motivation of industry-university-institute cooperation of Zhejiang high-tech enterprises was to develop new products, improve product quality, improve technology equipments and develop new markets, the proportion separately reaching more than 30%. But the object of patent development and technical standards , which was main way to enhance core competitive advantages, separately held only 20% and 6% proportion. Secondly, the mode of cooperation and the way of interest distribution were relatively single. The main mode of industry-university-institute cooperation of Zhejiang high-tech enterprises was cooperative development patterns, accounting for 86.0%. Other patterns, such as commissioned development, building R&D institutions and

Problems and Countermeasures of Zhejiang High-Tech Enterprises

49

technology transfer etc, were accounting to less than 10%. At the same time, 62.0% Zhejiang high-tech enterprises completed cooperation through providing R&D fund for universities and research institutes. Other patterns, such as assigned profit according to sales, retained profits, invest proportion, respectively held 18% proportion. Thirdly, the secondary R&D ability of ZheJiang high-tech enterprises is not strong. Cooperative approach was always the important innovation method selected by ZheJiang high-tech enterprises, but the method always resulted in short-term goal on industry-university-institute cooperation, which mass cooperation remained at imitative phase. Meanwhile, because many decision makers couldn’t understand industry-university-institute cooperation effectively, many enterprises didn’t want to invest a lot of R&D funds for industry-university-institute cooperation. This was a direct result for high-tech enterprises lacking "secondary innovation" capacity, which resulted in weak R&D achievements. Finally, the risk of industry-university-institute cooperation was obvious and serious. At present, the situation of industry-university-institute cooperation of Zhejiang high-tech enterprises was relatively satisfied. More than 40% enterprises considered that their industry-university-institute cooperation was fine; more than 10% enterprises intended to cooperation again after current cooperation. But More than 40% enterprises considered that there were many problems with industry-university-institute cooperation, and among them 14% enterprises often encountered disputes on the distribution of interest. 10% enterprises often disputed with partners because of intellectual property. At the same time enterprises would face risk in the field of market, technology etc during industry-university-institute cooperation. 50% enterprises considered the main risk of industry-university-institute cooperation was technical risk, although 60% enterprises considered marketing risk. These risks directly lead to uncertainty during industry-university-institute cooperation: 70% industry-university-institute cooperation projects couldn’t meet the expected target, 13% projects couldn’t distribute interest reasonably, 9% projects faced with conflict of inconsistent relationships and so on.

3 Countermeasures Enterprise is the body of innovation, and enterprise innovation need intellectual support of universities, research institutes, and it is more important for government supports to provide necessary innovation public goods and services. Through government support, enterprises, research institutes and universities can requires more innovation to be innovative. This section analyzed how different factors affected Zhejiang high-tech enterprises to absorb technology and knowledge from universities and research institutions. Then this section analyzed the role of government on industryuniversity-institute cooperation and we gave some suggestion for government to help Zhejiang high-tech enterprises cooperative R&D. As we can see from Table 1, Zhejiang high-tech enterprises industry-universityinstitute cooperation were influenced by many factors. The main influence factors including lack of information on technology, the high fee of technology transfer, lack of intermediate test and engineering capability and so on. These results indicated that Zhejiang high-tech enterprises industry-university-institute cooperation required

Q. Zhou, C.-F. Mao, and L. Hou

50

Table 1. The Influent Factors of Enterprises Achieving Technology from Universities and Research institutes % Lack of information on technology The

high

fee

of

technology transfer Not mature of technology Unclear

of

technology

ownership

Very small (1)

Smaller (2)

General (3)

More Very large mean large (4) (5)

5.95

15.86

40.17

26.49

11.53

3.22

5.41

13.33

34.41

33.15

13.70

3.36

3.78

9.55

27.57

36.58

22.52

3.65

12.79

18.56

32.25

22.16

14.24

3.06

5.59

13.87

29.55

31.35

19.64

3.46

6.30

13.15

28.65

33.16

18.74

3.45

19.46

25.59

37.66

13.33

3.96

3.58

7.75

16.76

30.63

30.63

14.23

3.29

No advantage compared to

similar

foreign

technology Unclear

of

technology

application prospect Not

strong

of

intermediary services Lack of intermediate test and

engineering

capability

government policy efficient utilization, which can guide cooperation among enterprises, universities and research institutes to improve R&D performance. As can be seen from Table 2, high-tech enterprises of ZheJiang government had very high expectations of countermeasures from government policy to promote industry-university-institute cooperation. Many enterprises hoped that government should establish an information service platform to strengthen exchanges and cooperation, establish fixed department responsible for technology transfer, encouragement to establish R&D consortium, and so on. The enterprises had strong expectation of more than 4 points value, indicated that ZheJiang high-tech enterprises hoped government can play an effective role in information communication, and could become information bridge of industry-university-institute cooperation. Therefore, government should develop effective guide and support mechanism for Zhejiang high-tech enterprises industry-university-institute cooperation. Firstly, government should promote the information service system vigorously. Because information was important resource for technical innovation, especially technology and information market information, government needed to provide perfect kinds of information, such as cooperative information of potential members, patent information, and build information database, etc. Government should ensure all information was authoritative, authenticities and efficient sources.

Problems and Countermeasures of Zhejiang High-Tech Enterprises

51

Secondly, government should urge enterprises to enhance level of industry-university-institute cooperation. Government should highlight enterprises as dominant part in industry-university-institute cooperation, promote innovative elements accumulate for enterprises innovation. Let enterprises play important role as main body to R&D invest, benefit from R&D investment and take a risk. Government should promote enterprises to become not only main body of innovation, but also the accumulation of innovation resources, fund, technology and talents. Table 2. The expectation of countermeasures from government policy

% Establish information service platform Establish fixed department responsible for technology transfer Encouragement to establish R&D consortium Establish database of intellectual property Incubating technology enterprises through intermediary Improving regulations and laws Government’s direct participation

Very small (1)

Smaller General (3)

More large (4)

Very large (5)

mean

(2)

1.08

5.42

14.80

44.04

34.66

4.06

1.80

10.99

31.35

38.56

17.30

3.59

0.54

3.24

18.02

42.34

35.86

4.10

1.62

5.41

26.49

43.78

22.70

3.81

4.32

16.40

41.80

26.31

11.17

3.24

1.44

7.75

28.11

37.84

24.86

3.77

6.49

13.33

33.15

33.51

13.52

3.34

Thirdly, government should play important role in constructing and improving technical intermediary service organization system. In the condition of market economy, science and technology intermediary service institutions was a bridge of knowledge and technology’s flow, diffusion and transfer, a service link of technology and application, production and consumption, and a clear sign of high-tech achievements transfer to practical productive forces, reflecting the level of technical innovation. In course of practice, government needed to urge intermediary service organization to develop long-term plan, establish modern science and technology intermediary network services platform, create actively market demand for technology intermediary, and energetically talent science and technology agency personnel, improve overall quality of team. Finally, government should build better external environment of industryuniversity-institute cooperation. Government needed to further change the functions, and carry out some measures to promote the combination of science and technology transformation and innovation. Through government investment and tax incentives, government induced and encouraged transformation of science and technology,

52

Q. Zhou, C.-F. Mao, and L. Hou

advocated for political combination of industry-University-Institute, coordinate relative government departments to play role. At the same time, government also needed to improve policies, regulations to guide participants to establish risk investment mechanism, which should ensure that all participants could get fair investment income.

4 Conclusion Zhejiang high-tech enterprises industry-university-institute cooperation has achieved some achievements, but still remained some questions, such as low cooperative level, single distribution, weak secondary R&D ability, obvious risk and so on, etc. This paper summarized and analyzed these issues through empirical test, and offered some policy suggestions for government to improve Zhejiang high-tech enterprises industry-university-institute cooperation, which included promoting the information service system vigorously, urging enterprises to enhance level of cooperation, playing important role to construct and improve technical intermediary service organization system, building better external environment of industry-university-institute cooperation, and so on.

Acknowledgement The authors would like to thank NSFC, China, for funding the project coded 70903020 and 71010107007. And we also would like to thank science research project supported by Science and technology Department of Zhejiang Province, for funding the project coded GK090903005.

References 1. 2.

3. 4. 5.

Cohen, Nelson, Walsw: Links and Impacts: the Influence of Public Research on Industrial R&D. Management Science 48, 1–23 (2002) Monjon, Waelbroeck: Assessing Spillovers from Universities to Firms: Evidence from French Firm-level Data. International Journal of Industrial Organization 21, 1255–1270 (2003) Zhong W.-j., Mei S.-e., Xie Y.-y.: Analysis of Technical Innovation Modes for the Industry-university-institute cooperation. China Soft Science 8, 174–181 (2009) Mao J.-q., Liu L.: Problems& Development Measures in Combination of Industry, Study& Research in China. Theory and Practice of Education 29, 23–25 (2009) Chiesa, Coughlan, Voss: Development of a Technical Innovation Audit. Journal of Production Innovation Management 13, 105–136 (1996)

Research on Puncture Area Calibration of Image Navigation for Radio Frequency Ablation Robot Peng Wang1, Wei Cao2, Wenhao Jiang1, Jinglei Xin1, Shaochen Kang1, and Xin Li1 1

School of Mechanical & Power Engineering, Harbin University of Science and Technology, Xuefu road. 52, 150080 Harbin, People’s Republic of China 2 Department of Cardiovascular Disease, The Second Affiliated Hospital of Harbin Medical University, Xuefu road. 246, 150081 Harbin, People’s Republic of China [email protected], [email protected]

Abstract. Aiming at puncture image characteristic of robot auxiliary radio frequency ablation, a modified tradition camera calibration method is proposed in this paper. Based on the comprehensive analysis for puncture image distortion characteristic, there are three geometric distortions in the puncture image system: radial distortion, centrifugal distortion and thin lens distortion, we research the distortions and establish the correction matrixes one by one, and then establish the camera mathematical model under the puncture condition, complete the coordinate systems’ conversion calculation to obtain the camera internal and external parameters, finally to get the puncture system calibration polynomial. Through the comparative experiment that takes the area as parameter, we get that the puncture system calibration method after image correction can meet the accuracy requirement in actually. Keywords: Puncture image, Vision system calibration, Image distortion, Vision-guided robot.

1 Introduction The radio frequency ablation is an interventional operation which releases radio frequency current at some part of intracardiac with electrode catheter to make local endocardium and subendocardial myocardium coagulation necrosis to destroy some original points of rapid arrhythmia. The eradication rate of this operation for paroxysmal supraventricular premature tachycardia is above 90 , so it has been the important method and effective means to treat atrial fibrillation in department of cardiology. The puncture and implantation of ablation electrode catheter are the long time operations under the radiation of X-ray, and the ray harms doctors seriously, so it’s necessary to research how to use image navigation robot to do the operations instead. The primary problem for reaching the surgery accuracy requirement is how to calibrate the surgery system. During the calibration process, the mathematical model is often used to describe the camera it describes the process from scene projection and image. The pin-hole model is a common ideal state model, from the physics, the

％

，

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 53–58, 2011. © Springer-Verlag Berlin Heidelberg 2011

54

P. Wang et al.

image point is the intersection point of the image plane and the line connecting between the optical center and the object point, its best advantage is the concisely, practically and accurately linear imaging relationship. Simulating the camera imaging process linearly with the pin-hole imaging is a kind of approximation for the camera applied in the actual, while the obtained puncture image has the seriously geometric distortion, all the geometric distortions will change the image point position in the scene, so they must be compensated. The lens distortion coefficient, which reflects the distortion effect, is brought in the ideal pin-hole perspective model in this paper.

2 Puncture Image Distortion Corrections The radial distortion is caused by the lens shape which can’t meet the theory requirement, when the radial distortion is existing, the image point shifts towards the radial compared with ideal position, while the straight line, which crosses the optical axis in the image plane, only has variation on length. The common mathematical model of radial distortion is shown as:

δ r = k1r 3 + k2 r 5 + k3r 7 + k4 r 9 + ⋅ ⋅ ⋅ Where

δ r is

(1)

the nonlinearity distortion at the image point whose polar coordinate

is (r , ϕ ) , r is the radial distance between the image center and the pixel point and k1 , k2 , k3 , k4 ,⋅ ⋅ ⋅ is distortion coefficient.

During the radial distortion correction, we find that the complicated distortion model not only fails to enhance the measurement accuracy, it appears to make the numerical calculation instability. Its accuracy contribution rate is less than the disturbance for the correction beginning from the third term of the polynomial, so only the first two terms are adopted in this paper, and the finally radial distortion correction model is shown as:

( (

⎧ xr = xd 1 + k1r 2 + k 2 r 4 ⎨ 2 4 ⎩ yr = yd 1 + k1r + k 2 r

) )

(2)

Actually, the varying degrees eccentricity necessarily exists, namely all the optical centers of the lens can’t be strictly collinear. This defect causes the so-called centrifugal distortion, which makes the image characteristic parameter instable. The simplified distortion components on x axis and y axis are:

⎡δ x ⎤ ⎡sin ϕ cos ϕ ⎤ ⎡δ r ⎤ ⎢δ ⎥ = ⎢ ⎥⎢ ⎥ ⎣ y ⎦ ⎣cos ϕ − sin ϕ ⎦ ⎣δ t ⎦ Where

δ r and δ t

(3)

are the distortional components at the radial and the tangential

respectively, m1 and m 2 are the constant coefficients, r is the radial distance between the image center and the pixel point, ϕ is the angle between the radial line, where the

Research on Puncture Area Calibration of Image Navigation

55

y axis positive direction, while ϕ 0 is the angle between the radial line at the maximum tangential distortion and y axis positive direction. Only image point locates, and

the first two order distortions are taken, so from the formula (3) we can get:

( (

) )

⎧ xt = m1 3 x 2 + y 2 + 2m2 xy ⎨ 2 2 ⎩ yt = m2 x + 3 y + 2m1 xy

(4)

With the consideration of thin lens distortion, the nonlinearity distortions mentioned above exist in the puncture images taken by the optical lens; the nonlinearity distortion is the superposition of the three distortions, so the total distortion is:

( (

) )

( (

) )

⎧ Δx = xd 1 + k1r 2 + k 2r 4 + m1 3x 2 + y 2 + 2m2 xy + n1r 2 ⎨ 2 4 2 2 2 ⎩Δy = yd 1 + k1r + k 2 r + m2 x + 3 y + 2m1 xy + n2 r

(5)

3 The Puncture Camera Parameters Calibration The work model of radio frequency ablation robot is shown as Fig.1. Its model schematic diagram is shown as Fig.2. The operation accuracy is directly influenced by the system coordinate calibration and conversion, there are 6 coordinates can be built in the system: word coordinate, camera coordinate, image coordinate, pixel coordinate, robot coordinate and patient coordinate. The conversion from the puncture point to the image coordinate needs 4 matrixes. According to the basic system working principle, the movement locus of surgical robot is determined by the target point correctly calibrated and mapped in each coordinate.

Fig. 1. The work station of radio frequency ablation robot

56

P. Wang et al.

Fig. 2. Puncture camera based on correction

OC ⋅ X CYC Z C is camera coordinate , the origin OC is the optic center of the camera, Z C axis coincides with optical axis, o ⋅ xy image physics coordinate, the origin o is the intersection point of optical axis and image plane, the x and y parallel with X C and YC ; O ⋅ uv is pixel coordinate, the position of point P in world coordinate is ( X W , YW , ZW ) , (u , v ) is actual image coordinate position of imaging point p, whose unit is pixel. The focal distance f is the distance from image plane to the optical center. Through perspective projection, the geometrical relationship of the imaging position p is defined as below:

xu = f × ( X c / Z c ) , Where

yu = f × (Yc / Z c )

(6)

( X C ,YC , Z C ) is the position of P in camera coordinate, (xu , yu ) is the ideal

pin-hole camera model; the unit of p in physics image coordinate is

mm .because

(u , v ) shows the line number and the column number of the pixel locates in digital

image array without any physical units, o ⋅ xy whose unit is built. The relationship between two coordinates depends on the size and the shape of the pixel, in addition to the pixel position in the technical grade camera. In the two coordinates, any pixel in the image has the relationship as below:

⎡u ⎤ ⎡ f u ⎢v ⎥ = ⎢ 0 ⎢ ⎥ ⎢ ⎢⎣1 ⎥⎦ ⎢⎣ 0

s fv 0

u0 ⎤ ⎡ X c / Z c ⎤ v0 ⎥⎥ ⎢⎢ Yc / Z c ⎥⎥ 1 ⎥⎦ ⎢⎣ 1 ⎥⎦

(7)

f u = f × s x , f v = f × s y , f u is the scale factor at u axis, f v is the scale factor at v axis. s is defined as the comprehensive distortion factor, when the pixel is a rectangle, a = 0, s = 0 or else s ≠ 0 . The conversion between the patient coordinate and the Where

world coordinate can be done by translation and rotation, and the camera can be placed at any position. Where R and T are rotation transformation and translation transformation, it shows the relationship between the camera coordinate and the patient coordinate. A group of error equations can be listed with a certain amount of control points in the image coordinate and the reference coordinate and the coefficients in the

Research on Puncture Area Calibration of Image Navigation

57

polynomial can be obtained with the least square method, so get the polynomial shown as below:

⎧ x = cos ϕX + k1r 2 X + sin ϕY + 3 sin ϕm1 X 2 + sin ϕn1r 2 + sin ϕ cos ϕ ⎨ 2 2 2 ⎩ y = sin ϕY + k1r Y + cos ϕX + 3 cos ϕm2Y + cos ϕn1r + sin ϕ cos ϕ

(8)

4 Calibration Experiment and Result Assessment Though the projection model based on comprehensive distortion correction, the calculative image eigenvalue after calibration is more accurate. The mainly thought is that fix the calibration block on the worktable, the calibration data of noncoplanar calibration point can be obtained with the accurate moving at the horizontal and the vertical. Analyzing whether the distortion correction has influence on measurement accuracy after calibration, the specific compare steps are: 1. Set the standard reference. The puncture region sign is a circle, so many circular coins are adopted in the experiment 2. Measure the area of the standard reference.The measuring result of non-corrected image is shown as Fig.3; the measuring result of our calibration is shown as Fig.4.

Fig. 3. The area measurement result of distortional image

Fig. 4. The area measurement result of correction image

58

P. Wang et al.

From the two figures above, we get that, the originally crooked side boundary is eliminated after correcting the distortion image; with the area as parameter, the measured area data has variation during the correction and the measured area after correcting reaches the real value better. So imaging system correction should be done to ensure the puncture accuracy.

5 Conclusions In this paper an image navigation puncture region calibration method is proposed, it is realized with a comprehensive distortion correction projection model, and we test the method with the experiments. The result shows that, this method is simple and effective, what’s more, it can enhance the location accuracy of the puncture point, so it lays a foundation for realizing auxiliary radio frequency ablation surgical robot accurately control the patient pose and enhances the safety of the participation, and meanwhile this method solves the radiation problem when the doctors directly operate in the X-ray. The robot coordinate system will be further researched to effectively realize the puncture point fast positioning. Acknowledgments. This work is supported by post-doctoral Foundation of Heilongjiang Province(LBH-Z09212).

References 1. 2. 3. 4.

Fei, B.: The safety issues of medical robotics. Reliability Engineering and System Safety 73, 183–192 (2001) Zhang, Z.: Camera calibration with dimensional objects. IEEE Transactions on Pattern Analysis and Machine Intelligence 26(7), 892–899 (2004) Machacek, M.: Two-step calibration of a stereo camera system for Measurements in large volumes. Measurement Science and Teehnology 14, 1631–1639 (2003) Devernay, F.: Automatic calibration and removal of distortion from scenes of structured environments. Machine Vision and Application 13, 14–24 (2001)

Low Power and Robust Domino Circuit with Process Variations Tolerance for High Speed Digital Signal Processing Jinhui Wang, Xiaohong Peng, Xinxin Li, Ligang Hou, and Wuchen Wu VLSI & System Lab, Beijing University of Technology, Beijing 100124, China [email protected] Abstract. Utilizing the sleep switch transistor technique and dual threshold voltage technique, a source following evaluation gate (SEFG) based domino circuit is presented in this paper for simultaneously suppressing the leakage current and enhancing noise immunity. Simulation results show that the leakage current of the proposed design can be reduced by 43%, 62%, and 67% while improving 19.7%, 3.4 %, and 12.5% noise margin as compared to standard low threshold voltage circuit, standard dual threshold voltage circuit, and SEFG structure, respectively. Also, the inputs and clock signals combination static state dependent leakage current characteristic is analyzed and the minimum leakage states of different domino AND gates are obtained. At last, the leakage power characteristic under process variations is discussed. Keywords: Domino circuits, SEFG, power, noise margin.

1 Introduction Domino circuits are commonly employed in register, cache, and high performance microprocessors [1]. As technology scales down, the supply voltage must be reduced to keep the dynamic power within acceptable levels [2]. However, to meet the performance requirements, the threshold voltage (Vt) and gate oxide thickness (tox) of the transistors must be reduced to accompany the supply voltage scaling down, which leads to exponential growth of the sub-threshold leakage (Isub) and the gate leakage (Igate) [3]. Especially, for domino circuits, the excess leakage current also degrades the noise immunity. This further highlights the importance of leakage current reduction. There has been lots of relative work on leakage reduction in domino circuits, such as dual Vt technique [4], sleep transistor technique [5], P-type domino circuit [6], source following evaluation gate (SEFG) [7], and so on. However, each technique could not separately solve the leakage and robustness problem completely. Therefore, in this paper, a novel domino design using several techniques is proposed for simultaneously suppressing the leakage current and enhancing noise immunity.

2 Proposed Circuits As described in Section 1, the excess leakage current has become an important issue to threaten the performance of domino circuit [2]. Dual Vt technique is an effective R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 59–65, 2011. © Springer-Verlag Berlin Heidelberg 2011

60

J. Wang et al.

technique to reduce Isub by using low Vt and high Vt transistors on the timing critical paths and non-critical paths, respectively. Utilizing this technique, gating all the initial inputs of the domino gates into a low Isub state is required. But dual Vt domino circuits do not take the effect of the Igate into account and therefore it could not minimize the total leakage current. To reduce Igate, P-type domino circuit adopting low leakage PMOS transistors instead of high leakage NMOS transistors in input network and sleep transistor technique based on adding current switch in sleep state are proposed [5], [6]. What’s more, to improve the robustness of the circuits, Kim proposed SEFG (Source Following Evaluation Gate) structure [7]. In SEFG, the output of source follower is limited by the gate input voltage and does not depend on the width of discharging path even if noise input lasts for a long time, as shown in Fig. 1 (b) and (c). In this paper, the SEFG structure, the dual Vt technique, and the sleep transistor are employed in P-type domino gate, as shown in Fig. 1 (d). When the gate works, sleep transistor is turned on. In the pre-discharge phase, clock is set high. Dynamic node is discharged to ground. Evaluation phase begins when the clock signal is set low. Provided that the necessary input combination to charge the evaluation node is applied, the circuit evaluates and the dynamic node is charged to Vdd, otherwise, preserving until the following pre-discharge phase. When the gate sleeps, sleep transistor is cut-off to lower the leakage current. Thus, compared to other existing domino circuits, the proposed circuit could realize low power and high noise immunity design with a little active power penalty. Vdd Vdd

Vout + Vgs Gnd

Ileak Gnd

+ Vgs Gnd

Vdd Low-Vt device

Vdd

High-Vt device sleep

Ileak Vout

int1

int2 Dynamic node

clk Gnd

P-keep

int1 Out

N-keep

int2 Dynamic node

clk

P-keep Out N-keep

Gnd

Fig. 1. (a) Common Source (b) Source Following (c) SEFG structure (d) Proposed design

3 Simulation Results The analysis in this paper is based on H-spice tool and 45 nm BSIM4 models [8]. The device parameters are listed in Table 1. To evaluate the performance of the proposed design, the following four two-input AND domino gates are simulated: the standard low Vt domino gate (s_Low), the standard dual Vt domino gate (s_dual), the SEFG gate, and the proposed design. Each domino gate drives a capacitive load of 8 fF. All AND gates are turned to operate at 1 GHz clock frequency. To achieve a reasonable comparison, all of the circuits are sized to have the similar delay. The delay is calculated from 50% of signal swing applied at the inputs to 50% of signal swing observed at the output. The leakage current, active power, noise immunity of these gates are simulated and compared. To analyze leakage current, two typical sleep temperatures are considered:

Low Power and Robust Domino Circuit with Process Variations Tolerance

61

(1) 110oC which is assumed that the sleep mode is short and the temperature keeps 110oC during the short sleep period; (2) 25oC which is assumed that the sleep period is long and the sleep temperature has fallen to the room temperature. While considering the noise immunity, the same noise signal is coupled to all of the AND gates, so this situation represents the worst case noise condition. In order to quantify the noise margins, the noise signal is assumed to be a 2.5 GHz square wave with 87.5% duty cycle. The maximum tolerable noise amplitude is defined as the signal amplitude at the inputs which induced a 10%-Vdd rising/falling in the voltage at the output of the AND gate.

(a)

(b)

Fig. 2. (a) Comparison of the active power and the noise immunity of four gates (b) Comparison of leakage current of four gates

Table 1. Parameters of devices The Vt value of four different devices High Vt NMOS High Vt NMOS Low Vt NMOS Low Vt NMOS 0.35V -0.35V 0.22V -0.22V Process: 45nm Working Temperature: 110oC.

Table 2. Normalized Leakage Current of the Devices NMOS Transistor PMOS Transistor Low-Vt High-Vt Low-Vt High-Vt 34.9 2.7 22.8 1.09 Ileak(Isub,Igate) [ 110 oC] (33.3,1.6) (1.2,1.5) (22.7,0.09) (1,0.09) Igate [110 oC] 4.7 3.5 0.1 0.1 126.2 60.4 56.3 4.4 Ileak(Isub,Igate) [25 oC] (66.5,59.6) (0.8,59.6) (52.8,3.4) (1,3.4) Igate [25 oC] 159.1 124.0 5.3 5.3 Currents are normalized to the gate leakage current produced by PMOS Transistor in Ileak state. Transistor width=1μm, length=45 nm. |Low-Vt|=0.22V. |High-Vt|=0.35V. Vdd=0.8V. Igate: |Vgs|=|Vgd|=|Vgb|=Vdd. Ileak: Vgs=0 and |Vds|=Vdd.

62

J. Wang et al.

Table 3. Leakage current of four gates in the different input vector and clock states at 25 oC and 110 oC (A) Input vector and clock state (25 oC) Input vector and clock state (110oC) CLIH CHIH CLIL CHIL CLIH CHIH CLIL CHIL Proposed 6.02e-8 6.23e-8 1.76e-7 1.39e-7 2.94e-7 9.37e-7 3.17e-7 1.55e-6 SEFG 1.86e-7 2.14e-7 1.97e-7 2.49e-7 5.32e-5 1.05e-6 1.45e-6 1.61e-6 s_Low 1.84e-7 1.66e-7 1.59e-7 2.01e-7 2.07e-6 6.62e-7 1.21e-6 1.27e-6 s_Dual 1.81e-7 1.44e-7 1.07e-7 1.79e-7 1.57e-7 6.59e-7 1.67e-7 1.27e-6 CLIH (clock=low, inputs=high), CHIH (clock=high, inputs=high). CLIL (clock=low, inputs=low), CHIL (clock=high, inputs=low).

Fig. 2(A) shows the normalized active power and noise immunity of four different domino gates. It shows that the noise margin of the proposed circuit is 19.7%, 3.4%, and 12.5% higher than that of the low Vt domino circuit, the dual Vt domino circuit, and the SEFG structure, respectively. This results from to the additional P-keeper transistor in the proposed circuit. Though this P-keeper is only half size of the transistor in the pull-up network, it improves the noise immunity effectively. As also shown in Fig. 2(A), the active power of the proposed design is increased by 45.8% as compared to standard dual Vt domino circuit. As discussed in Section 2, the proposed circuit has more transistors, thereby consuming more active power. Also, in order to achieve the same delay time, the size of the sleep transistor and PMOS transistor in the inverter must be increased, which leads to more active power in proposed design. However, the proposed circuit shows better leakage characteristics and noise immunity than other circuits including SEFG structure. Table 3 lists the leakage current of four gates in different input vectors and clock states at typical sleep temperatures. It can be observed that the leakage current characteristic depends strongly upon both of input vector and clock states. Therefore, to obtain the minimum leakage state which can be used to optimize the leakage current, a detailed analysis is required. On the one hand, when the sleep temperature is 25 oC, the minimum leakage states of all AND gates share one common character: the clock signal is low. This is can be explained as followed. When clock signal is low, low-Vt clock PMOS transistor is turned on and produces the total leakage current is 5.3, as shown in Table 2; lowVt/high-Vt clock NMOS transistor is turned off and produces the total leakage current is 126.2/60.4. When clock signal is high, low-Vt clock PMOS transistor is turned off and produce the total leakage current is 56.3; low-Vt/high-Vt clock NMOS transistor is turned on and produce the total leakage current is 159.1/124. Therefore, the low clock signal decreases total leakage current, as can be seen in Table 3. The leakage current of standard domino circuit at 25oC is minimized when inputs are low, as shown in Table 3. When the inputs are high, the PMOS transistors connected to inputs (input-PMOS) are turned off and produce both Isub and Igate (totally 56.3). But with low inputs, these PMOS transistors only produce Igate (only 5.3). Thus, low inputs minimize the total leakage current. However, in the SEFG structure and the proposed design, there are several additional transistors including the P-keeper and sleep transistor. Low inputs would turn on input-PMOS, which would result in the conduction of these additional transistors. Then the current loop is formed and stack effect [9] is vanished, thereby increasing the leakage current significantly. Thus, high input signals to input-PMOS could achieve minimum leakage current in both the SEFG structure and the proposed design.

Low Power and Robust Domino Circuit with Process Variations Tolerance

63

On the other hand, when the temperature is 110 oC, high inputs would minimize leakage current. This is because Isub increases exponentially as temperature increases, due to the reduction of Vt and increasing of the thermal voltage. But Igate, unlike Isub, has a very weak dependence on temperature. Therefore, at 110 oC, Isub makes a larger contribution to the total leakage current than Igate. When the clock input signal is high or low, one of the transistors connected to clock input is non-conduced produce large Isub. To suppress this large Isub, the stack effect is required. The input-PMOS with high signal realizes this important effect. Isub is only produced in turning-off state. In standard dual Vt domino circuit and the proposed design, Isub of high-Vt off-NMOS connected to clock (the values is 1.2) is less than that of low-Vt off-PMOS connected to clock (the values is 22.7). Low clock signal is efficient to suppress Isub. On the contrary, in the standard low-Vt domino circuit and the SEFG structure, Isub of the low-Vt off-NMOS connected to clock (the values is 33.3) is larger than that of low-Vt off-PMOS connected to clock (the values is 22.7). High clock signal is helpful to suppress leakage current, as can be seen in Table 3. From the above analysis, the minimum leakage state at 25 oC of the proposed design and the SEFG structure is CLIH and the minimum leakage state of standard domino circuits is CLIL. Alternatively, the minimum leakage state at 110oC of the proposed design and standard dual Vt domino circuit is CLIH. But in SEFG structure and standard low Vt domino circuits, the leakage is minimum in CHIH state. Fig. 2(B) compares the leakage current of four different gates at their minimum leakage states. It is observed that the leakage current of the proposed design is smallest in four gates due to dual Vt technology and sleep transistor technology. At 25 o C, the leakage current of the proposed design decrease by 43%, 62%, and 67% as compared to standard dual Vt domino circuit, standard dual Vt domino circuit and the SEFG structure. However, Isub increases exponentially with the increasing of temperature. Although the sleep transistor suppresses leakage current efficiently, the leakage current of the proposed design oversteps standard dual Vt domino circuit. But simulation results in Fig. 2(B) indicates that at 110 oC the proposed design decrease leakage current by 55%, and 72% as compared to standard low Vt domino circuit and the SEFG structure. In conclusion, the proposed design has great advantage to decrease leakage power at typical sleep temperatures.

4 Leakage Power Characteristic under Process Variations As the CMOS process advances continually, scaling has resulted in significant increase in the variations of the process parameters, including gate length (Lgate), channel doping concentration (Nch), and gate oxide thickness (tox). All of these process variations have a significant effect on the threshold voltage of the devices, which result in variation of the leakage power. Therefore, in order to evaluate the impact of process variations on leakage power characteristics of different domino circuits, Monte Carlo analysis is applied [10]. In the simulation, Lgate, Nch and tox are all assumed to have normal Gaussian statistical distributions with a three sigma (3σ) fluctuation of 10%. And 1000 Monte Carlo simulations are run.

64

J. Wang et al.

150

55 W

Samples

85 %

Proposed (CLIH) Dual Vt (CLIL)

76 %

200

Proposed (CLIH) Low Vt (CHIH)

180

T=25

100

0.4 W

160

99 %

140 120

89 %

T=11

100 80

50

60 40 20 0

0 0.2

0.4

0.6

0.8

1

1.2

0

1.4

0.2

0.4

1.6

0.6

1.8

0.8

1

1.2

1.4

2

Fig. 3. Leakage power distribution curves of the standard low threshold voltage circuit, the dual threshold voltage circuit, and the proposed circuit under process variations

Fig. 3 shows the leakage power distribution curves of the standard low threshold voltage circuit, the dual threshold voltage circuit, and the proposed circuit, respectively, in two typical temperatures under process variations. And three kinds of circuits are set in the minimum leakage power state (standard low threshold voltage circuit - CHIH, dual threshold voltage circuit - CLIL, proposed circuit - CLIH). It can be seen that the leakage current distribution curves of the dual threshold voltage circuit vs. the proposed circuit cross at 55 nW in 25 oC. 85% of the samples of the proposed circuit are lower than 55 nW and 76% of the samples of the dual threshold voltage circuit are higher than 55 nW. The leakage current distribution curves of the standard low threshold voltage circuit vs. the proposed circuit cross at 0.4 uW in 110 o C. 99% of the samples of the proposed circuit are lower than 0.4 uA and 89% of the samples of the standard low threshold voltage circuit are higher than 0.4 uW. These results indicate that the proposed design is preferable to reduce the leakage current in majority of the samples even under process variations, which is similar result to the analysis of the normal case.

5 Summary In this paper, a novel domino circuit structure is proposed to suppress the leakage current and enhance the noise immunity. Based on the simulation results, input vector and clock state of the gate is discussed to obtain the minimum leakage state. At last, the leakage characteristic under process variations is analyzed.

References 1. Stackhouse, B., et al.: A 65 nm 2-Billion Transistor Quad-Core Itanium Processor. IEEE Journal of Solid-State Circuits 44, 18–31 (2009) 2. Gong, N., et al.: Analysis and Optimization of Leakage Current Characteristics in Sub65nm Dual Vt Footed Domino Circuits. Microelectronics Journal 39, 1149–1155 (2008)

Low Power and Robust Domino Circuit with Process Variations Tolerance

65

3. Wang, J., et al.: Low Power and High Performance Zipper CMOS Domino Full-adder Design in 45nm Technology. Chinese Journal of Electronics 37, 266–271 (2009) 4. Kao, J.T., Chandrakasan, A.P.: Dual-threshold voltage techniques for low-power digital circuits. IEEE Journal of Solid-State Circuits 35, 1009–1018 (2000) 5. Liu, Z., Kursun, V.: IEEE transactions on Circuits and Systems. Leakage Biased PMOS SleepSwitch Dynamic Circuits 53, 1093–1097 (2006) 6. Hamzaoglu, F., Stan, M.R.: Circuit-level techniques to control gate leakage for sub-100nm CMOS Proc. In: Int. Symp. on Low Power Electronics and Design, pp. 60–63. IEEE Press, New York (2002) 7. Kim, J., Roy, K.: A leakage tolerant high fan-in dynamic circuit design technique. In: Solid-State Circuits Conference, ESSCIRC 2001, pp. 309–313. IEEE Press, New York (2001) 8. Predictive Technology Model (PTM), http://www.eas.asu.edu/~ptm 9. Lee, D., et al.: Analysis and Minimization Techniques for Total Leakage Considering Gate Oxide Leakage. In: ACM/IEEE Design Automation Conference, pp. 175–180. IEEE Press, New York (2003) 10. Wang, J., et al.: Monte Carlo Analysis of Low Power Domino Gate under Parameter Fluctuation. Journal of Semiconductors 30, 125010-1–125010-5 (2009)

Detection of Attention-to-Rest Transition from EEG Signals with the Help of Empirical Mode Decomposition Cheng Man Ng and Mang I. Vai Department of Electrical and Electronics Engineering, Faculty of Science and Technology, University of Macau, Macau SAR, China {ma36578,fstmiv}@umac.mo

Abstract. In this paper, an empirical mode decomposition (EMD) scheme is applied to analyze the steady-state visually evoked potentials (SSVEP) in electroencephalogram (EEG). Based on EMD method, the oscillatory activities of the decomposed SSVEP signal are analyzed. It is observed that the 6th IMF showed the features of the attention-to-rest transition response. In other words, high powers are observed instantly after the volunteer turns from an attentively focusing stage into an unfocused attention stage. Having made the point that the 6th IMF of the SSVEP signals corresponds to very low frequency (0.5 – 2 Hz), this drives us to look into that frequency range of the SSVEP signal. All of this reflects that a very low frequency seems to occur during the attention-to-rest transitions. Experiments are performed with different people. The result shows that the attention-to-rest transition can be detected with an accuracy of 82.6%. Keywords: Electroencephalogram (EEG); Empirical Mode Decomposition (EMD); Intrinsic mode functions (IMF); Steady-state visually evoked potentials (SSVEP).

1 Introduction Electroencephalography is the neurophysiologic measurement of the electrical activity of the brain by records from electrodes put on the scalp or the cortex. The resulting traces, namely electroencephalogram (EEG), reflect the electrical activity of a multitude of neural populations in the brain. EEG signal is a complicated nonlinear and non-stationary signal, containing different types of rhythmic waves: alpha wave (8–12 Hz), beta wave (12–30 Hz), delta wave (up to 4 Hz), theta wave (4–8 Hz) [1]. This paper is intended to analyze the SSVEP based BCI system in EEG. SSVEP are signals that are natural responses to visual stimulation at specific frequencies. They are stable responses that can be measured when a person is attentively focusing on a flashing light source with flickering frequency. The database of this paper is obtained from different volunteers who are asked to continuously focus on the flashing lights for every 6 seconds, with 4 seconds of resting time between each 6 seconds of attention. Our concern is to consider the transition point when the volunteer turns from the 6s attention stage into the 4s resting stage. R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 66–71, 2011. © Springer-Verlag Berlin Heidelberg 2011

Detection of Attention-to-Rest Transition from EEG Signals with the Help of EMD

67

The obtained SSVEP signals are decomposed into IMF based on the EMD method. EMD has been demonstrated as a method for the data processing of nonlinear and non-stationary signals. From the oscillatory activities of the decomposed SSVEP, it is noticed that the attention-to-rest transitions are obviously shown in the 6th IMF. Due to this observation, we will focus on examining the 6th IMF so as to lead to the detection of the attention-to-rest transitions. As a consequence, this may be used for detecting the idle period of a SSVEP based BCI system.

2 Methodology 2.1 Data Acquisition The signal of the SSVEP based BCI system is obtained from 7 volunteers at the ages of 18–30. Each volunteer was stimulated with a series of flashing lights, corresponding to 20Hz, 15Hz, 10Hz, 8Hz and 7.5Hz, by the use of electrodes PO3, PO4, POz, Oz, O1 and O2. The signals were recorded with a sampling rate of 600Hz. Firstly, the volunteer looked at the flashing light attentively, with the first stimulus frequency of 20Hz for 6s, and then rested for 4s. Secondly, the volunteer looked at the flashing lights at stimulus frequency of 15Hz for the next 6s, and rested for 4s, and so on. The above 5 stimulus frequencies were repeated for 1 time, resulting in a total experiment time of 100s. Signal processing technology is used to obtain the EEG signals. Fig.1 offers a complete set of the original SSVEP-EEG signals. The red areas refer to the time when the volunteer is gazing at the flickering lights of the specified stimulus frequency. Every volunteer repeats the experiment for 6 times. Therefore, the EEG database is composed of a total of 42 full-set signals. 2.2 EMD Application to the SSVEP Signals EMD is applied to decompose the SSVEP signals. In Fig. 2, it denotes 11 IMFs and one residue decomposed from the second channel data (PO4) of the original EEG signal. From the oscillatory activities of the 5th IMF and the 6th IMF, each time after the volunteer gazed at the flickering light and turned into a rest condition, the amplitudes are greatly raised in each transition period. All the experiments show that this transition phenomenon is most obviously shown in the 6th IMF. Fig. 3 depicts the enlargement of the 6th IMF from the corresponding EEG signal. The red dotted lines indicate the locations of 6s, 16s, 26s, 36s, 46s, 56s, 66s, 76s, 86s and 96s. We can see clearly high amplitudes are found at the start of each transition. Therefore, it may be useful to look more closely at some of the more important features of the 6th IMF. The 6th IMF is further analyzed by Fast Fourier Transform (FFT). Fig. 4 denoted the corresponding Fourier Spectrum of the 6th IMF, showing that there is a peak at 1Hz. The frequency contents of the 6th IMF from all the signals in our EEG database are found to be at a very low frequency, between 0.5Hz – 2Hz. In this way, we try to extend this observation into the idea that during the period of attention-to-rest transition, there is a very low frequency occurs in EEG.

68

C.M. Ng and M.I. Vai

Fig. 1. Six channels of the original SSVEP signals

Fig. 2. EMD decomposed signals of channel PO4 of EEG signal with 11 IMFs and residue

Fig. 3. Enlargement of the 6th IMF

Fig. 4. Fourier Spectrum of the 6th IMF

Detection of Attention-to-Rest Transition from EEG Signals with the Help of EMD

69

2.3 Detection of the Attention-to-Rest Transition A finite impulse response (FIR) equiripple band-pass filter (with low frequency at 0.5Hz and high frequency at 2 Hz) is therefore designed and applied to the original EEG signal in order to preserve the frequency content of the 6th IMF. The power spectrum of the filtered EEG signal is performed by an application of window with length 500 ms moving along the signal. The filtered EEG signal is then divided into 10 segments of 10s duration (including 6s of time focusing on the flickering lights and 4s of time resting). The highest power of each 10s duration are found to be located at 6.832s, 17.03s, 26.77s, 36.99, 46.74s, 56.82s, 66.89s, 77.78s, 86.58s and 96.88s. All of them belong to the resting durations, the time is within 0.58s – 1.78s right after the volunteer stop gazing at the flickering lights. Fig. 5 illustrates the power spectrum of the EEG signal after the band-pass filter is applied, in which the highest power locations of each 10s duration are also marked.

Fig. 5. Power Spectrum of the filtered EEG signal

3 Experimental Results 3.1 Accuracy Fig. 6 indicates a better demonstration on the locations of the highest powers for the filtered EEG signal. In this figure, the original SSVEP-EEG signal is shown in blue color, and the green areas are the 6s durations of gazing at the flickering lights, finally the red dotted lines are the occurrences of the highest power found in the filtered EEG signal. As expected, all the highest powers are located in the resting duration and they occur right after the volunteer stop watching the flickering lights. It can be concluded that a very low frequency occurs during the attention-to-rest transition. The same analysis procedures are applied to all the SSVEP signals in the EEG database. Table 1 summarizes the results of the accuracy for detecting the attention-to-rest transition point. The mean accuracy is 82.6%. Accordingly, our method is able to detect the idle period of the SSVEP based BCI system.

70

C.M. Ng and M.I. Vai

Fig. 6. Original EEG signal with (i) green areas are the time of watching flickering lights. (ii) the red lines are the occurrences of the highest power from the filtered EEG signal. Table 1. Accuracy of detecting the attention-to-rest transition on all volunteers Volunteers Trial 1 Trial 2 Trial 3 Trial 4 Trial 5 Trial 6 Accuracy

1

2

3

4

5

6

7

90% 80% 100% 90% 90% 100% 91.67%

90% 70% 70% 70% 90% 80% 78.33%

100% 70% 90% 80% 70% 80% 81.67%

80% 70% 80% 80% 70% 70% 75.00%

90% 80% 100% 100% 100% 100% 95.00%

70% 70% 70% 100% 60% 70% 73.33%

90% 80% 80% 90% 80% 80% 83.33%

3.2 Evoked Delta Waves The experimental results lead to the conclusion that a very low frequency occurs at the transition point of the idle period. Since this frequency band (less than 2Hz) belongs to the delta waves, it is reasonable to suppose that the attention-to-rest transition might be related to an increase in delta EEG activity. Delta activity has been found during some continuous attention task [1]. An increase in delta power has been reported in different types of mental tasks [2], it is neither due to the ocular movements nor to any other artifact. Research has been devoted that an increase in delta EEG activity may be related to attention to internal processing during the performance of a mental task [2]. Delta power will be increased in conditions such as attention, activation of working memory, letter identification etc [3][4]. It is pointed out in some Go/No-Go analysis, which are studied with event-related potentials (ERP), that there is a power increase at 1 Hz in both Go and No-Go conditions. The increase in delta activities during the No-Go conditions is related to the inhibition of non-relevant stimuli (Harmony et at., 1996), signal matching and decision making (Basar-Eroglu et at., 1992) [3]. On the other hand, delta power will become high during target relative to non-target processing, namely, in relation to the rest condition [5].

Detection of Attention-to-Rest Transition from EEG Signals with the Help of EMD

71

4 Discussion In this paper, we begin with the analysis of the SSVEP based BCI system by the EMD method. The transition response is found in the 6th IMF of the SSVEP signals. Therefore, a band-pass filter (0.5 – 2Hz) is used to preserve only the very low frequency of the original SSVEP signals. Consequently, the phenomenon of the occurrence of very low frequency during the attention-to-rest transition is demonstrated. This phenomenon is examined with different SSVEP signals obtained from different person. As a result, the attention-to-rest transitions are being successfully detected and the mean accuracy is found to be 82.6%. This result leads to the conclusion that during the attention-to-rest transition, a very low frequency occurs which means that delta waves are being evoked. To put it the other way round, when the volunteer turns from an attentively focusing stage to an unfocused attention stage, there is an increase in delta EEG activity. This phenomenon may be related to the inhibition of non-relevant stimuli, signal matching and in a state of non-target processing [3]. As a consequence, our method is able to detect the idle period of the SSVEP based BCI system.

5 Conclusion EMD method offers a powerful tool for analyzing nonlinear and non-stationary signal such as EEG. It offers the key to understand the components of the SSVEP signals, in which the attention-to-rest transition is able to be detected by means of the features of the chosen IMF. The most likely explanation is that the attention-to-rest transition is accompanied by the occurrence of delta activities. Since there is room for further improvement on the detection accuracy, we will further analyze the behavior of the EMD method, as well as the IMFs.

References 1. 2.

3.

4. 5.

Wikipedia, http://en.wikipedia.org/wiki/Electroencephalography Harmony, T., Fernández, T., Silva, J., Bernal, J., Díaz-Comas, L., Reyes, A., Marosi, E., Rodríguez, M., Rodríguez, M.: EEG Delta Activity: An Indicator of Attention to Internal Processing during Performance of Mental Tasks. International Journal of Psychophysiology 24(1-2), 161–171 (1996) Harmony, T., Alba, A., Marroquín, J.L., González-Frankenberger, B.: Time-FrequencyTopographic Analysis of Induced Power and Synchrony of EEG Signals during a Go/NoGo Task. International Journal of Psychophysiology 71(1), 9–16 (2009) Schroeder, C.E., Laktos, P.: Low-Frequency Neuronal Oscillations as Instruments of Sensory Selection. Trends in Neurosciences 32(1), 9–18 (2009) Doege, K., Bates, A.T., White, T.P., Das, D., Boks, M.P., Liddle, P.F.: Reduced EventRelated Low Frequency EEG Activity in Schizophrenia during an Auditory Oddball Task. Psychophysiology 46(3), 566–577 (2009)

A Traffic Information Estimation Model Using Periodic Location Update Events from Cellular Network Bon-Yeh Lin1,2, Chi-Hua Chen1, and Chi-Chun Lo1 1

Institute of Information Management, National Chiao-Tung University, 1001, University Road, Hsinchu, Taiwan 300, ROC 2 Telecommunication Laboratories, Chunghwa Telecom Co., Ltd, 12, Lane 551, Min-Tsu Road Sec. 5, Yang-Mei, Taoyuan, Taiwan 326, ROC [email protected], [email protected], [email protected]

Abstract. In recent years considerable concerns have arisen over building Intelligent Transportation System (ITS) which focuses on efficiently managing the road network. One of the important purposes of ITS is to improve the usability of transportation resources so as extend the durability of vehicle, reduce the fuel consumption and transportation times. Before this goal can be achieved, it is vital to obtain correct and real-time traffic information, so that traffic information services can be provided in a timely and effective manner. Using Mobile Stations (MS) as probe to tracking the vehicle movement is a low cost and immediately solution to obtain the real-time traffic information. In this paper, we propose a model to analyze the relation between the amount of Periodic Location Update (PLU) events and traffic density. Finally, the numerical analysis shows that this model is feasible to estimate the traffic density. Keywords: Periodic Location Update, Intelligent Transportation System, Cellular Network, Traffic Density Estimation.

1 Introduction In recent years considerable concerns have arisen over building Intelligent Transportation System (ITS) which focuses on efficiently managing the road network. One of the important purposes of ITS is to improve the performance of transportation so as extend the durability of vehicle which can be reduced the fuel consumption and travel time. Before this goal can be achieved, it is to obtain the correct and real-time traffic information which includes traffic density, traffic flow, speed, travel time, traffic conditions, and traffic accidents so that traffic information services can be provided in a timely and effective manner. At present, the methods of collecting real-time traffic information can be classified into three categories as follows. (1). Stationary traffic information detectors [1] (2). Global Position System (GPS)-based probe car reporting [2] (3). Tracking the location of mobile users through the cellular network [3-9] R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 72–77, 2011. © Springer-Verlag Berlin Heidelberg 2011

A Traffic Information Estimation Model Using PLU Events from Cellular Network

73

By far the most intuitive way is to set up the stationary traffic information detectors on important road segments to gather the information, but this way need excessive setup and maintenance costs [1]. In addition, several researches using GPS-based cars as probes. But the sample size of GPS-based probe cars needs to be high enough to infer the traffic information accurately [2]. Additional costs are incurred when these GPS-based probe cars periodically report the traffic information through the air. To reduce the aforementioned building and maintenance costs, cost-effective and immediately alternatives need to be found. As we can see nearly everyone has his own Mobile Stations (MS), it would seem advisable to collect traffic information by using MS as a probe [3-9]. Cellular network has rigorous management processes to keep track of the movement of MSs. There are events triggered by mobility management for cellular network such as Location Update (LU). By grouping neighboring cells, Location Areas (LA) can be defined to describe high level locations of MSs. When a MS moves from a LA to another LA, a Normal LU (NLU) event is triggered to inform the cellular network the MS’s latest LA information. Besides, a Periodic LU (PLU) event is triggered to update the MS’s LA information periodically if the NLU event hasn’t been triggered in a period of time. Through the PLU and NLU process, the network always knows the current cell and LA of a MS. Therefore, the amount of PLU events is related with the density of MSs in specific cell. For this reasoning, analyzing the occurred time and related information of these mobility management events can infer the real-time traffic information. In this paper, we propose a model to analyze the relation between the amount of PLU events and traffic density information. We also derive a specific formula to describe this relation and provide a numeric analysis of it. The remainder of the paper is as follows. In Section 2, we propose the analytical model and the derived formula. The numeric analysis is provided in section 3. Finally, conclusions are given in Section 4.

2 A Traffic Information Estimation Model When considering the probability of a PLU event in a specific cell, we have to take the following two scenarios: (1). There is no call between two consecutive PLU events. (2). There are several calls between two consecutive PLU events. Finally, we consider and sum up the two scenarios about the probability of PLU event in a specific cell for traffic density estimation. 2.1 Scenario (1): No Call between Two Consecutive PLU Events Figure. 1 shows the space diagram for scenario (1). There is a car within a MS moves along the road. The first PLU event is triggered at time t0, and then the car enters the cell at time t1. The second PLU event is triggered at time t2, and then the car leaves the cell at time t3. The timing diagram is illustrated in Figure. 2. The following assumptions and parameters in the model are defined below. • There is only one road residing in the cell. • c (hr): The cycle time of the PLU.

74

B.-Y. Lin, C.-H. Chen, and C.-C. Lo

• x (hr): The time difference between the first PLU event and entering the cell. We assume the time function f(x) is an uniform distribution function as f(x) = 1/c. • d (km): The length of road segment covered by a cell. • v (km/hr): The average speed of a car crossing a cell. In this scenario, the probability of PLU triggered in the cell is as formula (1). d ⎛ ⎞ Pr( Scenario 1) = Pr ⎜ x + > c ⎟ v ⎝ ⎠ =

∫

c

c−

d v

f ( x ) dx = ∫

c

c−

d v

(1)

1 d dx = vc c

Cell

Road

1st Location Update enter Cell at t1 at t0

2nd Location Update at t2

leave Cell at t3

Fig. 1. The scenario diagram for vehicle movement and PLU events on the road when there is no call between two consecutive PLU events 1st Location Update

2nd Location Update c

enter Cell x t0

leave Cell d/v

t1

t2

t3

Fig. 2. The time diagram for vehicle movement and PLU events on the road when there is no call between two consecutive PLU events

2.2 Scenario (2): Several Calls between Two Consecutive PLU Events The first PLU event is triggered at time t0. There is a call arrived at time t1, and then the car enters the cell at time t2. The second PLU event is triggered at time t3, and then the car leaves the cell at time t4. After that, a second call arrives at time t5. The space diagram is illustrated in Figure. 3 and the timing diagram is illustrated in Figure. 4.

A Traffic Information Estimation Model Using PLU Events from Cellular Network

75

The following assumptions and parameters in the model are defined below. • The call arrivals to/from one MS per one car along the road can be evaluated. The call arrival rate to a cell is λ (call/hr). • The call inter-arrival time function g(t) is an exponential distribution function as g(t) = λ e−λ t. • The call inter-arrival time tCIA is exponentially distributed [10] with the mean 1/ λ. • y (hr): The time difference from first call arrival to entering cell. • The time function h(y) is an uniform distribution function so h(y) = 1/(2c). In this scenario, the probability of PLU triggered in the cell is as formula (2). d⎞ d ⎛ ⎞ ⎛ Pr(Scenario2) = Pr⎜ tCIA > c ∩ y + > c ⎟ = Pr(tCIA > c ) × Pr⎜ y > c − ⎟ v⎠ v ⎝ ⎠ ⎝ ∞ ∞ c d d = ∫ g (t )dt ×∫ d h( y)dy = ∫ λe − λt dt × = e − λc c c− c 2vc 2vc v

(2)

Cell

Road

1st Location Update 1stcall arrival at t1 at t0

2nd Location Update at t3

Enter Cell at t2

Leave Cell at t4

2nd call arrival at t5

Fig. 3. The scenario diagram for vehicle movement and PLU events on the road when there are several calls between two consecutive PLU events

1st Location Update 1st call arrival

2nd Location Update

enter Cell y t0

t1

c

2nd call arrival

tCIA leave Cell d/v

t2

t3

t4

t5

Fig. 4. The time diagram for vehicle movement and PLU events on the road when there are several calls between two consecutive PLU events

76

B.-Y. Lin, C.-H. Chen, and C.-C. Lo

2.3 Summary: Traffic Density Estimation We use the formula (3) to consider and sum up the two scenarios about the probability of PLU event in a specific cell for traffic density estimation. Pr( PLU ) = Pr (tCIA > c ) × Pr( Senario 1) + Pr (tCIA < c ) × Pr( Senario 2) = e − λc ×

(

)

(

)

d d ⎞ d ⎞ ⎛ ⎛ − λc + 1 − e − λc × ⎜ e − λc × ⎜ e − λc ⎟ = 3−e ⎟ vc 2 vc 2 vc ⎠ ⎝ ⎠ ⎝

(3)

Formula (3) is the probability of PLU event triggered in a specific cell of one car. To find the probabilities of PLU events triggered in a specific cell of all cars, we have to multiply formula (3) with the traffic flow f (car/hr). The amount of PLU events r (event/hr) on the road segment can be expressed as formula (4) to estimate the traffic density D (car/hr).

(

)

d ⎞ ⎛ r = f × Pr(PLU) = f × 3 − e −λc × ⎜ e − λc ⎟ 2vc ⎠ ⎝ d ⎞ ⎛ = D × 3 − e − λc × ⎜ e − λc ⎟ 2c ⎠ ⎝

(

(4)

)

3 Numeric Analysis

500

500

Density PLU

400

400

f = 5000 (car/hr) λ = 1 (call/hr) d = 1 (km) c = 1 (hr)

300 200

300 200

100

100

0

0 10

20

30

40

50

60

70

80

90

Amount of PLU (event/hr)

Traffic Density (car/km)

In this section, we analyze the relation between the amount of PLU events and traffic density to evaluate the feasibility of our traffic information estimation model. For the purpose of demonstration, we adopt some parameters as followings to estimate the traffic density and the amount of PLU events: f = 5000 car/hr, λ = 1 call/hr, d = 1 km, and c = 1 hr. Fig. 5 shows that a positive relationship with the amount of PLU events and traffic density. Therefore, this model can be used to estimate traffic information for ITS to analyze the traffic information.

100

Speed (km/hr)

Fig. 5. The relation between the amount of PLU events and traffic density with different vehicle speeds

A Traffic Information Estimation Model Using PLU Events from Cellular Network

77

4 Conclusions This paper studied an analytic model to analyze the speed report rate with considering communication behavior and traffic information for the feasibility evaluation of traffic information estimation from cellular data. In experiments, the results shows that a positive relationship with the amount of PLU events and traffic density. This model can be used to estimate traffic information for ITS to analyze the traffic congestion, accidents, transportation delays, and etc.

References 1. Middleton, D., Parker, R.: Vehicle Detector Evaluation. Report 2119-1. Project Number 02119. Texas Transportation Institute (2002) 2. Cheu, R.L., Xie, C., Lee, D.H.: Probe Vehicle Population and Sample Size for Arterial Speed Estimation. Computer-Aided Civil and Infrastructure Engineering (17), 53–60 (2002) 3. Ygnace, J., Drane, C., Yim, Y.B., de Lacvivier, R.: Travel time estimation on the SanFrancisco bay area network using cellular phones as probes. University of California, Berkeley, PATH Working Paper UCB-ITS-PWP-2000-18 (2000) 4. Fontaine, M.D., Smith, B.L.: Probe-based traffic monitoring systems with wireless location technology: an investigation of the relationship between system design and effectiveness. Transportation Research Record: Journal of the Transportation Research Board (1925), 3–11 (2005) 5. Bar-Gera, H.: Evaluation of a cellular phone-based system for measurements of traffic speeds and travel times: A case study from Israel. Transportation Research Part C (15), 380–391 (2007) 6. Caceres, N., Wideberg, J.P., Benitez, F.G.: Deriving origin-destination data from a mobile phone network. IET Intelligent Transport Systems 1(1), 15–26 (2007) 7. Logghe, S., Maerivoet, S.: Validation of travel times based on cellular floating vehicle data. In: Proceedings of the 6th European Congress and Exhibition on Intelligent Transportation Systems, Aalborg, Denmark (2007) 8. Caceres, N., Wideberg, J.P., Benitez, F.G.: Review of traffic data estimations extracted from cellular networks. IET Intelligent Transport Systems 2(3), 179–192 (2008) 9. Gundlegard, D., Karlsson, J.M.: Handover location accuracy for travel time estimation in GSM and UMTS. IET Intelligent Transport Systems 3(1), 87–94 (2009) 10. Bolotin, V.A.: Modeling call holding time distributions for CCS network design and performance analysis. IEEE Journal on Selected Areas in Communications 12(3), 433–438 (1994)

The Application of Virtual Reality on Distance Education Zehui Zhan Center of Educational Information Technology, South China Normal University, Guangzhou, 510631, China [email protected]

Abstract. The features and classifications of Virtual Reality Techniques have been summarized and recommendation of applying Virtual Reality on distance education has been made. Future research is needed on the design and implementation of virtual classroom and courseware. Keywords: Virtual Reality; distance education; virtual classroom; application.

1 Introduction With the development of network technology such as broadband transport rate, Virtual Reality Techniques have been gradually taken consideration by distance education organizations. Modern educational technologies are making Virtual Reality a promising instructional means, where modeling and simulations can be used to display the structure and trends of natural, physics or social system, so as to provide an experiential and observable environment for students. The purpose of this paper is to analyze the applicability of Virtual Reality in Distance Education.

2 Virtual Reality As described on Wikipedia, Virtual Reality was first proposed by Jaron Lanier in 1989 [1]. It is also known as "Artificial Reality", "Cyberspace", "Artificial Environments", "Synthetic Environments" and "Virtual Environments". Virtual Reality is a kind of perceived environment usually simulated by computer. Most virtual reality environments are primarily visual experiences, displayed either on a computer screen or through special stereoscopic displays, but some simulations include additional sensory information, such as sound through speakers or headphones. Some advanced and experimental systems have included limited tactile information, known as force feedback. Virtual Reality is based on computer technology and hardware equipments, to implement a kind of illusional spaces that can be seen, listened, touched and smelled by users. It can be a set of computer simulation of 3D environment, which might be either entities exist in the world or those fictive characters never exist. In virtual reality environment, by dint of the computer hardware, network techniques, broadband and computer 3D calculating capacity, users can enter the virtual world through HumanComputer Interface. R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 78–83, 2011. © Springer-Verlag Berlin Heidelberg 2011

The Application of Virtual Reality on Distance Education

79

Besides, Virtual Reality is a subject that integrates human factors and information technology. Its main purpose is to denote information through artificial work and experience. Complex and abstract objects can be divided into sub-objects and expressed as specific symbols in virtual environment. VIRTUAL REALITY will introject many human factors, and intends to magnify the affects that it carries out to personal feelings. Therefore, it is built up on the integration of psychology, controlism, computer graphics, database design, real-time distributed system, electronics, robotization, multi-media techniques, etc. Moreover, distinguished from the static CAD model, Virtual Reality is a dynamic open Environment, which will response to the users' input such as gestures, sounds and keyboard commands. Through these interactions, users will feel themselves being immersing in this virtual space. Also, administer can supervise and control the environment by another interface as well.

3 Features of Virtual Reality The basic features of Virtual Reality have been sum up by Grigore Burdea in a model called "Virtual Reality Triangle", into three perspectives—Immersion, Interaction and Imagination[2]. 3.1 Immersion Immersion emphasizes the "real" experience. Ideally, users should totally fling themselves into the virtual system without feeling they are inside of it. This perfect illusional effect is a sense of presence, which should be contributed by three factors-Imagery, Interaction and Behavior. Imagery guarantees the High Fidelity of 3D effects. For example, we should control the parallax, vision scope, depth and angle to generate the feeling of "presence". Interaction is various in Virtual Reality, users should be able to feel natural when they execute an action in virtual environment. Behavior means that objects in virtual environment should be able to obey the order of nature or the regulation set by designers when they are moving or interacting. It is also called Autonomy of Virtual Reality system. For example, when an object receives a force, it will move to the opposite direction and give a force back at the same time, according to physics regulation. According to human cognitive theory, these three factors connect to as well as act on one another, and finally lead to the Immersion effect. 3.2 Interaction Interaction is a goal to provide convenient and vivid interactive means, which enable users to act inside virtual space and get real-time response; at the same time, enable the system to collect feedback from users as well. 3.3 Imagination Imagination is a factor to make the Virtual Reality system stand out. Most of the Virtual Reality system is not one kind of high-end interface, but the application focus on specific fields and question domain. Therefore, it requires not only the understanding and digesting of techniques, but also the brave imagination.

80

Z. Zhan

4 Classification of Virtual Reality 4.1 Desktop Virtual Reality Desktop Virtual Reality implements the simulation via personal computer and lowend work station. Usually, computer screen is used as the window for users to observe virtual space; the peripheral equipments are used to control the stand-ins and other interaction. In desktop Virtual Reality, 3D vision effects (such as using stereo-glasses) can enhance users' immersion. Besides, the CRT screen and stereo-image techniques lead to a high screen solution power and a low price. Therefore, this kind of virtual reality is the easiest to spread widely and has a strongest life force. 4.2 Immersive Virtual Reality Immersive Virtual Reality uses interaction equipments such as headset screen and data gloves to temporary block the users' vision and acoustical sensory perception into a close-up environment, so as to immerse deeply into the virtual system. There are three kinds of typical immersion Virtual Reality --cave immersion, cabin-seat immersion and screen projection immersion. Cave immersion is a 360 degree immersion system. Cabinet seat immersion makes the user sit in a cabinet, where there is a window-like screen entitle them to "look outside". So the user can turn around and browse in this space without wearing any Virtual Reality equipments. Projection immersion uses the projector to reflect the user him/herself to the screen, so that they can see themselves and interact with the virtual environment around them in the screen. 4.3 Distributed Virtual Reality Distributed Virtual Reality is to connect different users that stay in different location, let them share the same environment and work together in virtual space. Different from other Virtual Reality that just enable people to interact with virtual environment, it realizes the communication between different users as well. To sum up, Virtual Reality can display a rich and colorful space, enable users studying, working, living and entertaining inside. It brings a brand-new taste and view to the society and education. Now, some kinds of virtual digital world, such as virtual community, virtual family, virtual tourism and virtual language guide are already a part in our life that cannot be neglected. At the same time, modern education is trying to make good use of this technology and improve the education quality and learning effects.

5 Virtual Reality Applied in Online Learning Environment 5.1 Drawbacks of Today’s Online Learning Environment The drawbacks of the existing online learning environment can be sum up as follows: First, today's online education is short of the real experience. According to the constructivism theory mentioned earlier in this paper, knowledge should be organized within an environment, so that learners can link them actively and fulfill the process

The Application of Virtual Reality on Distance Education

81

of reconstruction. The learning outcome is combined with the specific situation, which can be met in real life or in learning environments. However, most of today's online learning are based on website and follow the HTTP protocol. They are lack of reality and fail to show the situation as authentic as it should be. Second, today's online education is not perceivable and natural enough. The collaborative learning environment base on internet should have at least three basic functions--sharable information resource, interactive tools and cooperative learning space. A multi-channel interaction with rich information and favorable effect is greatly needed. However, nowadays, the user interface and interactive modes are not friendly enough in online learning environment. It fails to satisfy the cooperators' emotional expression and collective apperceive, affects the quality of cooperation and communication. A good collaborative learning environment should be able to achieve individual improvement through the collective force. Anyone who acts excellent or falls behind will be noticed by his/her teacher and classmates, so as to build up a good psychological atmosphere to help everyone make progress together. Third, today's online education doesn't attach enough importance on the learners' nonintellective factors. Different learner will have different characteristic, different social and cultural background, and their original knowledge structure. So there are usually quite large gap on thinking patterns and learning styles differences. However, it is hard for us to identify the learners' differences on most of the online education website, because the only difference between learners is the ID that assigned to them automatically by the system. This shortage hinders the teaching in accordance of students' aptitude. Moreover, learners might not be able to learn more about each other as well. 5.2 Solution and Trend—Virtual Reality Applied in Online Learning Environment One of the best possible solutions for the drawbacks mentioned before is to combine cooperative learning techniques and virtual reality techniques, set up a distributed virtual learning environment, shows information by a three-dimension dynamic format, so as to achieve better reality, natural interaction and immersive effects. There are several advantages for applying Virtual Reality in online virtual classroom: First, virtual reality enriches the multimedia representation format in virtual classroom. 2D and 3D format take the place of the dull interaction. They can make the interface real and vivid, convenient for users to get information. An enjoyable feeling will be transferred to users and increase their cooperation efficiency. Second, virtual reality improves the cooperative users interface in virtual classroom. 3D distributed cooperative learning environment is much friendlier than the simple 2D software interface. Users can experience a deeper immersion, which make the interface more acceptable and easy to use. User’s status will be more transparent for others as well, so teachers and classmates will be able to know each other and collaborate better. Third, Virtual reality is good for the cultivation of learners' non-intellective factor. In virtual classroom, stand-ins will be used to stand for learners. Users can select the stand-in that similar to their own characteristic or the characteristic they would like to

82

Z. Zhan

cultivate. This step enables the system to identify user's personality, culture background and knowledge structure in the first stage. Everyone will use "first person" to communicate with classmates or teachers, and then the system will be able to record and analyze their inclination and interest, so as to choose the best way for them to learn. Fourth, virtual reality increases the whole system's efficiency. The reality threedimensional senses give students a quick entry to the study status and keep them concentrate inside. In this way, the learning efficiency is increased. In addition, the harmonious human computer interface transfers the real world interaction almost directly into virtual classroom, therefore, users don't have to passively adapt to the computer interface. In this way, Virtual Reality save the time and effort by avoid making cognitive burden for users. Actually, some countries have already pay attention on the Virtual learning environment. For example, United States is the first country that applies Virtual Reality onto education. In 1992, East Carolina University set up the Virtual Reality Educational Lab, aims to confirm the feasibility of Virtual Reality -Education, evaluate the hardware and software of Virtual Reality, study the effects on education generated by Virtual Reality and the application in reality, also, it compare the effects between Virtual Reality and other education medias. Also, the United Kingdom develops Virtual Reality-Education with great passion as well. The first Educational Virtual Reality project is set up in Newcastle-Upon-Type middle school, based on the Dimension International technology. Virtual Reality applied on language training and Industrial training has been explored here. Meanwhile, the VITART (Virtual Reality Application Research Team) project in Nottingham University is also doing the research on virtual learning system, which focuses on Virtual Reality input equipments and disable people training. 3D technique is much more complex than 2D. That’s the reason why it is not as prevail now. However, with the development of technology, when the data communication is more advanced and it is not going to take us as much effort to set up a 3D space, 3D virtual environment will certainly become popular for online education.

Summary This paper analyzed the possibility and necessity of applying Virtual Reality on distance education. The definition, features and classification of Virtual Reality have been sum up and merits of building up virtual classroom for distance education has been pointed out. Future research is needed on the design and implementation of virtual classroom and courseware, especially on the design of virtual classroom interface.

Acknowledgement This work is supported by the Natural Science Foundation of Guangdong province in China No. 8451063101000690, and the Foundation of Scientific and Technological Planning Project of Guangdong province in China No. 2009B070300109.

The Application of Virtual Reality on Distance Education

83

References [1] Information on, http://en.wikipedia.org/wiki/Virtual_reality [2] Burdea, G., Coiffet, P.: Virtual Reality Technology. John Wiley & Sons, New York (1994) [3] Robertson, G., Czeminski, M., van Dantzich, M.: Immersion in Desktop Virtual Reality, attrieved from http://www.research.microsoft.com/en-us/um/people/ marycz/uist97.pdf

Framework Design of Unified Cross-Authentication Based on the Fourth Platform Integrated Payment* Xu Yong and He Yujin School of Economics and Commerce, South China University of Technology, Guangzhou University City, Guangzhou, China [email protected], [email protected]

Abstract. The essay advances a unified authentication based on the fourth integrated payment platform. The research aims at improving the compatibility of the authentication in electronic business and providing a reference for the establishment of credit system by seeking a way to carry out a standard unified authentication on a integrated payment platform. The essay introduces the concept of the forth integrated payment platform and finally put forward the whole structure and different components. The main issue of the essay is about the design of the credit system of the fourth integrated payment platform and the PKI/CA structure design. Keyword: The fourth integrated payment platform, Unified authentication, PKI/CA.

1 Background The data released by iResearch, a professional consulting company, shows that the deal size in first quarter reached 2 trillion 12 billion and 100 million yuan, which increased 17.8% compared with the deal size in the last quarter and 93.5% compared with the same period of last year. It is obvious that the network payment industry remains in rapid development and it has become one of the fastest growing industries in the world of Internet [1].While the network payment industry is growing at a high speed, two main problems are still unsolved. One lies in the incompatibility of its rapid development and its related infrastructure. The other is contradiction between the massive need for online payment security and the poor compatibility of existing online payment tools [2]. According to the Computer Security Net, the number of government CA and licensed brand CA is 36 [3]. One of the key elements in protecting and securitizing the ID authentication is a stable and secured network; however, the over-sophisticated and duplicated net security instrument also negatively affects the efficiency and compatibility of network payment. *

The National Soft Science Research Program, 2010B070300016.

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 84–88, 2011. © Springer-Verlag Berlin Heidelberg 2011

Framework Design of Unified Cross-Authentication

85

2 Introduction of Unified Cross-Authentication in TFPP 2.1 What Is the Fourth Integrated Payment Platform? The fourth integrated payment platform is also called the fourth party payment (TFPP for short). The fourth party payment is integrated service provider offering electronic payment and its value added service. As Figure 1 shows below. Integrated Accessing Interface For platform Of E-Business Merchant

Control Of Security And Risk

The Middleware Of Products Display Interface Of Payment

Integrated Accessing Interface For Added-value SP

Unified Payment Protocol Interface Of Banks

Credit Evaluation

Unified Cross-Authorization Protocol

Interface Of E-Payment Supervision

Unified Settlement

Interface Of Taxation

compreh ensive service informati on Manage ment

Taxation

Fig. 1. The Composition Of TFFP

Through its payment service, value added service, and by integrating different resources and technologies owned by different service providers, standardizing the process of electronic commerce and by offering electronic payment supervision interfaces. The fourth party payment consists of interface service module and support system service module. The interface service includes unified access interface for electronic business vendors, unified access interface for value added service provider, products show medium, payment interface, bank interface, electronic payment supervision interface, unified settlement, credit assessment, tax control inferface. Value added service includes security and risk management module and general information management module. 2.2 A PKI Framework for TFFP The new PKI/CA based on TFFP is very similar with the PKI/CA in structure. PKI is responsible for the part of user identity that fits the public key managed by PKI. CA is responsible for the management of certificate [5]. With its payment function, the authentication center of the fourth party payment based on PKI/CA has a structure as shown in the fig.2

86

X. Yong and H. Yujin

Fig. 2. Unified cross-authentication center framework of TFPP

As it is shown in the fig.2 that the fourth party payment unified cross-authentication is linked with six major substances Domain. Foreground requests for authentication are usually sent by authentication center of buyers and sellers, who initiate the authentication process. Between the buyers/sellers and the backgrounders are CAs of the intermediary service provider and third-party payment platform provider, which do not actually trade commodities but benefit by offering intermediary service. CAs of finance and government administration operates in the background and are responsible for supervision and control. Government administration CA links to the crossauthentication mainly because the current government online payment supervision facilities are not able to keep pace with the evolution of the online payment and because online trade tax issues. Government agencies are able to levy taxes, and work in cooperation with financial institutions to supervise and manage the deposit fund. The two main information flows can be categorized into two kinds. The first kind is the request information and authentication information that transports from the substance domain to the fourth party payment. The first kind of information stream is indicated by dotted lines in the above fig.2. The second, as shown by the real line, is the response from the fourth party payment.

3 Detail Design of TFPP Authentication Center 3.1 Detail PKI/CA Framework of TFPP In the fourth party payment platform, unified authentication need to get the point that a single user can achieve more identity. Using the way of “domain” and declaring the

Framework Design of Unified Cross-Authentication

87

attribute status can access the permissions after verification. The most important function of the fourth party payment center is carrying on the credit management to the participation entities and guaranteeing the participation gain the permission and information assigned by the role. Next is providing authentication service which is cross domain, cross system and cross application to the participation. 3.2 Functions of TFPP Authentication Center The fourth party payment PKI is similar to the traditional PKI in case of the system, but has more flexibility in the credit administration and the authentication. The framework of PKI in TFPP has shown below.

Fig. 3. Framework of PKI in TFPP

1. The fourth party payment CA The fourth party payment CA is the trust center of the whole authentication system and also the highest credit center. Credit administration's object include domain CA as well as the participation entity in the domain, the way of management is evaluating the credit line by initial registration to the status of entity and previous activity log and determining to permit, forbid or restore the cross domain application. 2. Domain CA Domain CA is similar to the secondary CA in the traditional PKI system, but for a single domain it is root CA. Domain CA is responsible for issuing certificate to the participation in every single domain, distributing, updating and restoring the key and managing the certificate.

88

X. Yong and H. Yujin

3. Attribute confirmation Just as the preceding text said, every domain has the sole attribute value. The attribute value is saved in the expand part of the certificates which make the X.509 V3 as the standard platform. So the fourth party payment need to verify the attribute values, one purpose is confirming the authenticity of the status, determining which domain it belong and assigning jurisdiction, another is transmitting information to the fourth party payment CA, which providing the data for the authentication. 4. Certificate warehouse Resembling the PKI, the fourth party payment PKI also use certificate to save and retrieval certificate, certificate cancelled list and so on. The ostensible certificate information in the certificate warehouse can be inquired by the participant and public, but regarding the information in the secret part, its authenticity and security can be safeguarded by effective reliable gathering to the certificate holder made by the organization issuing the certificate. We can see that the fourth party payment system includes two major functions: identity authentication and cross credit transmission.

4 Conclusion This article put forward the unified cross authentication of the fourth party payment based on PKI/CA by studying the problems in the network payment, especially the contradiction in the safe authentication aspect. Its characteristic lies in the ductibility of the trust relationship in the PKI system. Different domain and different participation establish mutual trust relationship by bridge CA the fourth party payment, reducing cross authentication certificate between each other, raising the efficiency.

References 1.

2. 3. 4. 5.

iResearch China, The market scale of China network payment has reached 212.1 billion Yuan in the first quarter of 2010 (2010), iResearch China in http://irs.irDEearch.com.cn Financial Times, A report on China integrated liquidation platform. Cio360 in, http://www.cio360.net/ The list of brand CA in China, http://www.infosec.org.cn/zydh/ca.html Xu, Y., Fang, C.: A theoretical framework of fourth party payment. In: The International Conference on E-Business and E-Government (ICEE 2010), May 7-9 (2010) Ming, Q.: Electronic commerce Security, 2nd edn. Higher Education Press (2006)

Analysis towards VMEM File of a Suspended Virtual Machine Zheng Song, Bo Jin, and Yongqing Sun* Key Laboratory of Information Network Security, Ministry of Public Security, People’s Republic of China (The Third Research Institute of Ministry of Public Security), Shanghai 201204, China [email protected], [email protected], [email protected]

Abstract. With the popularity of virtual machines, forensic investigators are challenged with more complicated situations, among which discovering the evidences in virtualized environment is of significant importance. This paper mainly analyzes the file suffixed with .vmem in VMware Workstation, which stores all pseudo-physical memory into an image. The internal file structure of .vmem file is studied and disclosed. Key information about processes and threads of a suspended virtual machine is revealed. Further investigation into the Windows XP SP3 heap contents is conducted and a proof-of-concept tool is provided. Different methods to obtain forensic memory images are introduced, with both advantages and limits analyzed. We conclude with an outlook. Keywords: VMEM, Virtual Machine, Forensic.

1 Introduction Memory is an important basic component in modern computer architecture. Programs are generally loaded into memory and then executed by CPU, inputs and outputs are transferred via memory. Many sensitive information of a running system exists in memory. In the past, analyzing a memory image means using commands such as strings and grep to search the whole image for ASCII or UNICODE strings like passwords, IP addresses, or some other plain text contents [1]. As the result of the first forensics challenge issued before fifth annual conference of Digital Forensic Research Workshop (DFRWS 2005) [2], there are more tools and methods to Windows memory analysis. However, with the advent of virtualization era, traditional tools and methods may not work correctly any longer and memory image analysis faces more challenges. This paper mainly analyzes the file suffixed with .vmem in VMware Workstation, which stores all pseudo-physical memory into an image. The functionality of the .vmem file is explained as follows. By default, the .vmem file is enabled by the VMware Workstation product. Every time a user starts a virtual *

Corresponding author.

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 89–97, 2011. © Springer-Verlag Berlin Heidelberg 2011

90

Z. Song, B. Jin, and Y. Sun

machine, a temporary file suffixed with .vmem is created. When the suspend button is pressed, the Virtual Machine Monitor (VMM) saves all the memory contents of the virtual machine’s pseudo-physical memory into this file, which has a size that is same to memory size in virtual machine’s configuration file (.vmx). So when the virtual machine is resumed, VMM fetches data from this file and can create an environment that is exactly the moment the virtual machine is suspended without any differences or modifications. So our investigation starts from this .vmem file. The next section discusses the VMEM structure. Section 3 studies the useful information that can be obtained from VMEM. The demos show the steps how we are able to find useful information from Windows XP SP3 heap structure.

2 Overview of VMEM Most contemporary general-purpose operating systems commonly adopts paging mechanisms to support multi-process environments hence allows the physical address space of a process to be noncontiguous. Each process is executed in its virtual address space, fetching or saving data according to its virtual addresses. However, there is only one physical address space to which multiple virtual address spaces are mapped. So bridging the gap between virtual and physical address is a prerequisite before deeper investigation. VMEM is a kind of memory image file that is slightly different from memory dump file (*.dmp) under Windows OS. It is identical to a dd-style memory image without any header in Windows dump files as we discovered in our experiments. The structure of .vmem file can be regarded as the same to pure physical memory, which depends on guest operating system.

3 Useful Information from VMEM An enormous number of useful information resides in memory image, such as kernel structures from the OS, processes and threads, heaps. Other sensitive information include user’s password input, records of instant messaging (IM), and lots of browser-related information. The whole world seems chaos before what every byte means in memory image is figured out, they will soon be viewed in order. The first thing to do is to identify the processes, which is represented by an executive process (EPROCESS) block in Windows OS. As long as all the processes are recognized in the memory image, it is easier to obtain information such as threads, heaps and various other attributes relating to a process. Because all of them have some relation with EPROCESS which can be illustrated as follows.

Analysis towards VMEM File of a Suspended Virtual Machine

91

_KPROCESS +0x000 Header : _DISPATCHER_HEADER

_EPROCESS +0x000 Pcb

: _KPROCESS

...

...

+0x018 DirectoryTableBase : [2] Uint4B

+0x084 UniqueProcessId : Ptr32 Void

...

+0x088 ActiveProcessLinks : _LIST_ENTRY ...

...

+0x174 ImageFileName UChar

: [16]

... : Uint4B

...

+0x018 ProcessHeap Void

: Ptr32

+0x000 Type UChar

:

+0x001 Absolute UChar

:

+0x002 Size UChar

:

+0x003 Inserted UChar

:

+0x004 SignalState Int4B

:

+0x008 WaitListHead : _LIST_ENTRY

... : Ptr32 _PEB

...

+0x008 ImageBaseAddress : Ptr32 Void ...

+0x1a0 ActiveThreads

+0x1b0 Peb

_PEB

_DISPATCHER_HEADER

+0x088 NumberOfHeaps

: Uint4B

+0x08c MaximumNumberOfHeaps : Uint4B +0x090 ProcessHeaps Ptr32 Void

: Ptr32

...

Fig. 1. EPROCESS and its internal substructures as KPROCESS, PEB and DISPATCH_ HEADER

Processes are synchronized objects so their control structure EPROCESS begins with a substructure known as DISPATCHER_HEADER (similar situation happens to threads). The header contains a Type field which is used for identification. It functions similarly to magic numbers in some file formats in forensic field (e.g., PE magic number, ZIP magic number). In our experiments under Windows XP SP3 guest, the starting 4 bytes are always 0x03,0x00,0x1b and 0x00 (0x001b0003 in little-endian). Getting to know the specific header characteristics is not enough, because there are false positives possible. A further check on the EPROCESS candidates is a must. A lot of useful information about a process is stored inside EPROCESS such as PID (process id), create time and exit time, its threads and handle table. Andreas Schuster [3] contributed experiences to searching for processes and threads in Windows memory dumps, which is useful here. Besides, there are lsproc, lspi toolkit series which is suitable only for Windows 2000 [1]. As soon as each EPROCESS is located by its offset within the .vmem file, it is time to present more information. The 4 bytes beginning from offset 0x1b0 inside EPROCESS is a point to the PEB (Process Environment Block) structure. However, as this pointer is in virtual address space semantic, it does not work in the memory image in physical address space.

92

Z. Song, B. Jin, and Y. Sun

So the next step is to convert the virtual addresses to physical ones so we can know where it is located within the memory image. According to Intel’s publications [4], virtual address can be translated to physical address via segmentation and paging mechanisms. But we will only cover the paging because we can have two inputs and want only an output: Physical Address = TranslationFunc ( Virtual Address, CR3 ); Virtual address can be obtained obviously, while the value of CR3 register of each process is stored in the 4 bytes beginning from offset 0x018 inside EPROCESS (the first member of DirectoryTableBase array). But the conversion progress is different: this conversion is done by hardware (MMU in CPU) under real machine, but achieved by software-MMU (functions from VMM) in virtual machine. So traditional paging translation mechanisms under x86 failed to output the correct physical address in our experiments and we struggled to figure out that it PAE mode through various tests. It seemed incredible to us at first because PAE is not turned on in BIOS in host machine, besides, there were no clues about PAE in Windows XP SP3 guest OS. But we soon got to know that either kind of paging mechanism is possible if it’s achieved by software. The following figure shows the paging mechanism adopted by VMware Workstation 7.0.0 build 203739.

Fig. 2. Linear Address Translation with PAE enabled (4-KByte Pages) [4]

With the gap from virtual to physical bridged, more information can be revealed by PEB. Executive files under Windows (e.g., *.exe) are organized as PE format, and some sections are loaded into memory from disk, while others are not statically stored and are dynamically created by Windows sub-system. Heap is an obvious example. The Windows Heap Manager is a sub-system used throughout Windows to provision dynamically allocated memory. It resides on top of the virtual memory interfaces

Analysis towards VMEM File of a Suspended Virtual Machine

93

provided by the kernel, which are typically accessed via VirtualAlloc() and VirtualFree(). Basically, the Heap Manager is responsible for providing a high-performance software layer such that software can request and release memory using the familiar malloc()/free()/new/delete idioms[5]. Each process typically has multiple heaps, and software can create new heaps as required. There is a default heap for the process, known as the process heap, and a pointer to this heap is stored in the PEB. All of the heaps in a process are kept in a linked list hanging off of the PEB. The 4 bytes beginning from offset 0x090 is a pointer to a pointer array of ProcessHeaps whose size is represented by NumberOfHeaps in PEB. The first member in this array is default heap for process which equals to ProcessHeap in PEB. Each member of the array points to a HEAP structure. The HEAP structure contains a pointer array with 64 elements, each of which points to a HEAP_SEGMENT structures or filled with NULL. The FirstEntry in HEAP_SEGMENT is a pointer to HEAP_ENTRY structure, which reveals size of current block, size of previous block and unused bytes in current block. The value of Size/PreviousSize is calculated as actual bytes divided by 8. Flags represents the block is in use, the last one, or a virtual one. _PEB

_HEAP_SEGMENT

_HEAP ...

+0x008 ImageBaseAddress : Ptr32 Void

+0x008 Signature Uint4B

... +0x018 ProcessHeap Ptr32 Void

:

ProcessHeaps Pointer Arrays

...

Pointer to Heap No.0 (+0x018 ProcessHeap)

+0x088 NumberOfHeaps : Uint4B

Pointer to Heap No.1

+0x08c MaximumNumberOfHeaps : Uint4B +0x090 ProcessHeaps Ptr32 Ptr32 Void ...

:

+0x000 Entry : _HEAP_ENTRY

+0x000 Entry : _HEAP_ENTRY

+0x00c Flags Uint4B ...

+0x058 Segments[0]: Ptr32 _HEAP_SEGMENT +0x05c Segments[1]: Ptr32 _HEAP_SEGMENT ...

Pointer to Heap No.(NumberOfHeaps - 1)

+0x154 Segments[63]: Ptr32 _HEAP_SEGMENT ...

: :

_HEAP_ENTRY

: Ptr32

+0x000 Size : Uint2B

+0x00c Flags Uint4B

:

Pointer to Heap No.2 ...

+0x008 Signature Uint4B

:

+0x010 Heap _HEAP

+0x002 PreviousSize : Uint2B

... +0x018 BaseAddress Ptr32 Void

:

+0x01c NumberOfPages Uint4B

:

+0x020 FirstEntry : Ptr32 _HEAP_ENTRY +0x024 LastValidEntry : Ptr32 _HEAP_ENTRY ... +0x038 LastEntryInSegment : Ptr32 _HEAP_ENTRY

+0x004 SmallTagIndex : UChar +0x005 Flags : UChar +0x006 UnusedBytes : UChar +0x007 SegmentIndex : UChar

Fig. 3. Heap Organization

Although there are some insightful researches into Windows heap exploitation [5], they are of few help here because their usages are all in live state but we see data in a static way. Bridging the gap between live and static is much more difficult than addresses translation.

4 Demos With the detailed description in last section, we use Perl scripts to analyze the .vmem file. The procedures can be summarized as follows.

94

Z. Song, B. Jin, and Y. Sun

1. 2. 3.

Indentify all the processes from the .vmem file by specific characteristics in DISPATCHER_HEADER. Find out physical location of a process’s PEB within .vmem file by address translation. Get detailed information directly or indirectly from PEB. For example, searching for heap contents by ProcessHeaps form PEB.

The following demo is done under VMWare Workstation 7.0.0 build 203739 with a guest OS of Window XP SP3. The guest virtual machine is configured with a memory size of 256MB. Scenario I Start the virtual machine, log in to Windows XP SP3 guest, and open the default notepad program in XP. Type in one line “This is a demo”, and DO NOT save it to any .txt files on disk. Then press the suspend button from the VMware Workstation control panel immediately. Now follow the steps below: 1.

Identify the processes from VMEM by using a lsproc-series Perl script that we modified from lsproc tools. From the picture captured below, we can see the offset of EPROCESS structure of process notepad.exe is 0x0186e1b8.

Fig. 4. Identify the processes from VMEM

2.

Use a Perl script named lsheap_xpsp3.pl to display the structures of heap organization and their physical offsets within the VMEM file. There are altogether 10 heaps in our demo, and we just showed 3 of them in the picture captured below (Fig. 5) due to the length of this paper.

Analysis towards VMEM File of a Suspended Virtual Machine

3.

95

By a deeper walk of the heap contents beginning from the first entry in the default heap (starting from 0x000a0000), we found that the UNICODE string of “This is a demo” locating at virtual address 0x000b1b08, which is offset 0x07e02b08 in VMEM file as showed in Fig. 6. Note that the 8 bytes from 0x0x07e02b00 is the HEAP_ENTRY part.

Fig. 5. Display the heap organizations of the process notepad.exe

Fig. 6. The contents of VMEM showed by WinHex

96

Z. Song, B. Jin, and Y. Sun

Scenario II This is a similar scenario, but the outcome of our findings is attractive. Start the virtual machine, log in to Windows XP SP3 guest, and open the latest Windows Live Messenger (version 14.0.8117.416) in XP. Type in the password of an account that is stored beforehand, and log in normally. Then press the suspend button from the VMware Workstation control panel immediately. This time when we are walking through the heap contents of the process msnmsgr.exe, we found the plain-text password in the context of the MSN account information accidentally. We assume that the password was used in the login process and released without cleaning the contents. For privacy reasons, the results will not be published here. This may become a potential vulnerability and users of MSN will risk privacy leakage.

5 Conclusion and Future Work The method of analysis to .vmem file proposed in this paper has several significances. 1. 2. 3.

It is a general method to search for useful information step by step, thus can be ported to different paging mechanism or different operating system. Other virtualization products with pause function by other vendors may have similar architectures thus some attention should be paid. As .vmem file is a binary file representing physical memory on the bottom of the guest operation system, tricks taken by rootkits to hide in OS does not work any longer. It is another way to detect whether there are hidden processed in your system that you do not know yet.

Our method of searching sensitive data in .vmem file still has some limits: 1.

2.

3.

Diversity. The actual structure of a .vmem file varies depending on the paging mechanism, different operating system version or service pack, and it is difficult to develop an all-in-one program to solve the diversity problem. Script language like Perl is more powerful here, and is agile and flexible to changes. Besides, the analysis to .vmem file depends on the ability of analyst, and more scripts can be prepared to deal with different situations (e.g., different versions and service packs of Windows). Efficiency. Identifying the specific bytes with characteristics from such a large file and then checking those candidates is a time-consuming process. It usually takes several minutes. However, the next steps can be done via scripts in seconds because no comparison is needed. So optimizing the scripts to improve efficiency is a future direction. Granularity. In our demo, the plain-text contents of the heap can be found by the process it belongs to, i.e., it is on process granularity. We hope future investigation can reveal more information so it can be used in forensic field without any question.

Analysis towards VMEM File of a Suspended Virtual Machine

97

Acknowledgement This paper is supported by the Special Basic Research, Ministry of Science and Technology of the People's Republic of China (No. 2008FY240200), the Key Project Funding, Ministry of Public Security of the People's Republic of China (No. 2008ZDXMSS003).

References 1. 2. 3. 4. 5.

Carvey, H.: Windows Forensic Analysis. Elsevier, Amsterdam (2007) DFRWS 2005 (2005), http://www.dfrws.org/2005/index.shtml Schuster, A.: Searching for process and threads in Microsoft Windows memory dumps. Digital Investigation (2006) Intel. Intel 64 and IA-32 Architectures Software Developer’s Manual, vol. 3A (2007) McDonald, J., Valasek, C.: Practical Windows XP/2003 Heap Exploitation. Blackhat (2009)

Optimization and Reconfiguration of Advanced Manufacturing Mode Based on Object-Based Knowledge Mesh and Improved Immune Genetic Algorithm Chaogai Xue1 and Haiwang Cao2 1

Department of Management Engineering, Zhengzhou University, Zhengzhou, China [email protected] 2 Department of Electronic and Communication Engineering, Zhengzhou Institute of Aeronautical Industry Management, Zhengzhou, China [email protected]

Abstract. This paper deals with an approach to the optimization and reconfiguration of advanced manufacturing mode based on the object-based knowledge mesh (OKM) and improved immune genetic algorithm (IGA). To explore the optimization and reconfiguration of the new OKM by the user’s function requirements, an optimization procedure of an OKM aiming at the user’s maximum function-satisfaction is proposed. Firstly, based on the definitions of the fuzzy function-satisfaction degree relationships of the users’ requirements for the OKM functions and the multiple fuzzy function-satisfaction degrees of the relationships, the optimization model of the OKM multiple set operation expression is constructed. And the OKM multiple set operation expression is optimized by the immune genetic algorithm, with the steps of the OKM optimization presented in detail as well. Based upon the above, the optimization and reconfiguration of an advanced manufacturing mode are illustrated by an actual OKM example. The proposed approach proves to be very effective. Keywords: Advanced manufacturing mode, optimization, user satisfaction degree, object-based knowledge mesh, improved immune genetic algorithm.

1 Introduction At present, various advanced manufacturing modes and new concepts emerge constantly. Though these advanced manufacturing modes are of different advantages, there has been no advanced manufacturing system that can contain all of the advanced manufacturing modes suitable for all kinds of manufacturing enterprises. If all kinds of complementary advanced manufacturing modes are transformed into their corresponding advanced manufacturing knowledge in the advanced manufacturing system, the enterprise can be allowed to select the most appropriate combination of advanced manufacturing modes or the best mode for operation. Therefore, a knowledge mesh (KM) was brought forward to formally represent complex knowledge such as advanced manufacturing modes, information systems. And to solve the information exploration in KM representation, OKM was brought forward [1~3]. R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 98–103, 2011. © Springer-Verlag Berlin Heidelberg 2011

Optimization and Reconfiguration of Advanced Manufacturing Mode

99

The reconfiguration technique provides an effective approach to the adaptation of the enterprises to outward conditions. The self-reconfiguration methods based on KM and AM are studied in Yan and Xue [2,3]. Only if an optimal OKM multiple set operation expression is obtained can the new optimal OKM be inferred from the expression to realize the optimal self-reconfiguration of the KMS. Thus, optimization of an OKM multiple set operation expression is the key to those of the reconfiguration. The optimization problem of an OKM multiple set operation expression is no longer the general linear, nonlinear or dynamic programming one. But it can be solved by connotative enumerative searching methods like genetic algorithm. Therefore, an improved IGA based on niche algorithm is adopted for optimization of OKM multiple set operation expressions, and the feasible OKM multiple set operation expression with user’s maximal fuzzy function-satisfaction degree is obtained. And then, reconfiguration is conducted to realize the new optimized advanced manufacturing modes.

2 OKM Multiple Set Operation Expression Optimization Since users usually focus attention on functions of a manufacturing mode, it is necessary to build the function-satisfaction degree model of an OKM and define its operations according to the relevant knowledge of fuzzy math [4], which can be referred in reference [5]. 2.1 Mapping Problem of the Solution Space To solve the optimization problem, relative evaluation criteria need to be established, and the optimization of function-satisfaction degree expressions for OKMs is realized. Then acquired function-satisfaction degree expressions are mapped into their OKM multiple set operation expressions. The optimization process is shown in Fig. 1. Evaluated OKM multiple sets

Function-satisfaction degrees for OKMs Combined with satisfaction degree operators

Combined with OKM multiple set operators

Function-satisfaction degree expression Perform immune genetic algorithm

OKM multiple set operation expression with optimal function-satisfaction degree

According to mapping rules

Function-satisfaction degree expression with optimal function-satisfaction degree

Fig. 1. Optimization process of OKM multiple set operation expression

Theorem 1: There is a one-to-one mapping between the OKM multiple set operation expression with brackets and N non-bracket operators and the function-satisfaction degree expression with N operators and the non-repetitive permutation of operator priorities.

100

C. Xue and H. Cao

2.2 Optimal Model of OKM Multiple Set Operation Expression The objective of optimization is to obtain the optimized chromosome with the best fitness vector. Thus, the objective function is

J = max ε ( f m'n' ( x ) , x 0 ) , f

Where,

m'n'

f m'n' is a fitness vector determined by the n ' th chromosome in the m' th

generation, which varies with the chromosomes. The notation some in the

x is the n ' th chromo-

m' th generation. x0 is the ideal fitness vector and a row vector made up

of 1 for the ideal satisfaction degree of each OKM-function is 1. 2.3 Immune Genetic Algorithm for OKM Multiple Set Operation Expression Optimization Compared with genetic algorithm based on immune principle and standard GA, IGA has the following advantages: (1) immune memory function; (2) maintaining antibody diversity function; (3) self-adjustment function. When fitness in IIGA is calculated, Niche algorithm can be used to select the best chromosome, and maintain the diversity of population. Improvements proposed in this paper are as follows: Each fitness is adjusted according to (1).

d ⎧ ⎪1 − sh ( d ) = ⎨ σ share ⎪0 ⎩ d=

d ≤ σ share

(1)

d > σ share

d ( opi, opj )

(2)

m × popsizel

Where, σ share is the radius of Niche, and generally σ share = 0.1 ; d ( opi , opj ) is Ham-

ming distance, m is the number of genes, popsize1 is the number of chromosomes in each sub-group. Then the new method of calculating fitness is shown in (3).

ε ( xopi ) =

ε ' ( x opi )

∑ sh ( x N

opi

opj =1

, xopj )

(3)

To overcome Niche algorithm’s deficiency of fast convergence, which leads to premature convergence, convergence function is introduced as follows.

⎛

f =

⎞

log ⎜ ∑ sh ( xopi , xopj ) ⎟ N

⎝ opj =1 ⎠ ⎛ N ⎞ log ⎜ ∑ sh ( xopi , xopj ) − 0.02 ⎟ ⎝ opj =1 ⎠

(4)

Optimization and Reconfiguration of Advanced Manufacturing Mode

101

ALGORITHM 1: OKM multiple set operation expression optimization algorithm: STEP 1 Initialize the parameters of IGA, and population size is

popsize .

STEP 2 Dividing population into qs sub-spaces uniformly, which includes: (1) Calculating and adjusting fitness according to above (1) and (2). (2) Taking the first five optimal chromosomes as initial vaccine bank. (3) Crossover operation. The Pc of any two chromosomes in popi ( i = 1, 2, L , s ) is calculated according to (3), then the pair with maximum Pm will be selected to perform crossover operation, and the first two optimal chromosomes will be reserved. (4) Mutation operation. Judging evolution generation, if the maximum generation is divisible by present generation, then one chromosome is selected from popi , and will be vaccinated. Or the chromosomes are selected via a simulated roulette wheel. (5) Adjusting vaccine bank. When the chromosomes’ fitness values are lower than a given threshold φ , then the corresponding chromosomes will be preserved, or will be excluded from the bank. (6) Updating sub-population and vaccine bank. Any chromosome’s fitness is higher than a given value δ in popi , the corresponding chromosome is submitted to vaccine bank, and when the size of vaccine bank has been reached, the chromosome being submitted will replace the one with the lowest fitness in the bank. STEP 3 After each sub-population is calculated in sequence, judging the present generation, if it’s greater than the maximum generation, then end the process, else n = n + 1 , and go to step 2. STEP 4 According to Theorem 1, the best chromosome obtained above is transformed into the bracket-including OKM multiple set operation expression with the highest function-satisfaction degree.

3 Optimization of Advanced Manufacturing Modes Given below is the procedure of advanced manufacturing mode optimization. PROCEDURE 1: The steps of optimization of advanced manufacturing modes STEP 1 The user’s function requirements are layered by modules and sub-modules to facilitate the evaluation and management. STEP 2 The existing advanced manufacturing modes are represented as OKM and ITRMs according to the representation approach. And according to the user’s requirement, other OKMs with complementary functions are selected from the base. STEP 3 Intensive selection is to select the OKMs satisfying the user’s function requirements from the OKMs selected preliminarily in STEP 2. The layer structure is adopted to evaluate the OKM functions, and the first two OKMs with the top evaluation of function i are selected for i = 1,L , n . STEP 4 The OKMs selected through STEP 3 are evaluated and the functionsatisfaction degree vectors for the OKMs are obtained by fuzzy evaluation methods.

102

C. Xue and H. Cao

STEP 5 The OKM multiple set operation expression is optimized by ALGORITHM 1. The chromosome with the best fitness vector and the OKM multiple set operation expression with the highest function-satisfaction degree is then obtained. STEP 6 A new OKM can be inferred from the step 5-obtained OKM multiple set operation expression with the highest function-satisfaction degree by the developed OKM-based inference engine so as to realize the reconfiguration of the OKM. STEP 7 The new OKM and its ITRMs are mapped into the new advanced manufacturing mode, and the optimized manufacturing mode is obtained.

4 Examples Following the steps in optimization of an advanced manufacturing mode, a simple example is given to show the optimization process. The user requirements are first transformed into layer structure according to step 1 of Procedure 1, as shown in Fig. 2. Production management Equipment management Material management

Function requirement of an enterprise

Quality management technic management Financial management Tool management

Production plan & schedule Production control Production record Equipment files Repair and maintenance Earlier stage management Equipment pre-check Bill management Part management Plate management Labor protection management Metering management Quality inspection Quality information Product route Product process Process change In-tally report Credence management Purchase audit Detailed tooling cost Detailed account Tooling historical card Process equipment list

Fig. 2. Function requirement layering

The OKMs corresponding to advanced manufacturing mode are represented according to step 2 of Procedure 1. Suppose that the OKMs are : 1 , : 2 , and they are conducted the union operation. Then they will be taken as operands of an OKM multiple set operation expression during the evaluation process. Following step 3 and step 4 of Procedure 1, the OKMs are evaluated and the function-satisfaction degree vectors for the OKMs are obtained. The final functionsatisfaction degree and the multiple function-satisfaction degree for OKM : 1 are T

⎛ ⎞ h1L = w L ⎜ H1 ⎟ = ( 0,0,0.6594,0, 0.6857, 0, 0.6535 ) ⎝ ~ ⎠

， (h )

L * w1

= ( 0, 0,1, 0,1, 0,1) .

Optimization and Reconfiguration of Advanced Manufacturing Mode

In the same way, other OKMs : 2 , : T

3

and :

4

103

are evaluated as follows.

h2L = w L ⎛⎜ H 2 ⎞⎟ = ( 0.832,0, 0, 0.642, 0,0.71, 0.538 ) ⎝ ~ ⎠

， (h )

L * w2

= (1, 0, 0,1,0,1,1) .

T

， (h )

L * w3

= ( 0,1,0,1,1, 0,1) ..

T

， (h )

= (1, 0,1,1,0,1, 0 ) .

⎛ ⎞ h3L = w L ⎜ H 3 ⎟ = ( 0,0.751, 0, 0.542,0.496,0, 0.816 ) ⎝ ~ ⎠ ⎛ ⎞ h4L = w L ⎜ H 4 ⎟ = ( 0.685, 0,0.579, 0.785, 0, 0.573,0 ) ⎝ ~ ⎠

L * w4

Following step 5 of Procedure 1, the best OKM multiple set operation expression is (( M 2 - ( M1 + M 3 )) + M 0 ) - M 3 , with best fitness vector {0.832, 0.751, 0.6594, 0.785, 0.6857, 0.71, 0.816}. We can see that after the reconfiguration of OKM, user satisfaction degree is improved compared with the original OKMs, because functions are more richened.

5 Conclusions The optimization of advanced manufacturing modes is studied based on OKM and IGA. Based on the user function requirements of OKM and the optimization of the OKM multiple set expressions, optimization problem of advanced manufacturing modes aiming at maximum user satisfaction is solved. As is verified, the method proposed can help the enterprise select the optimal combination of advanced manufacturing modes or the best mode for operation. Acknowledgments. This work is supported by National Natural Science Foundation of China Grant #70971119, #70901066 and Aviation Science Foundation of China Grant #2008ZG55020.

References 1. 2.

3. 4. 5.

Yan, H.S.: A new complicated-knowledge representation approach based on knowledge meshes. IEEE Transactions on Knowledge and Data Engineering 18, 47–62 (2006) Xue, C.G., Cao, H.W.: Formal representation approach to enterprise information system based on object knowledge mesh. In: 2008 Chinese Control and Decision Conference (2008) Cao, H.W., Xue, C.G.: A Study on reconfiguration of enterprise information system based on OKM. In: The 8th World Congress on Intelligent Control and Automation (2010) Yager, R.R.: Fuzzy modeling for intelligent decision making under uncertainty. IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics 30, 60–70 (2000) Xue, C.G., Cao, H.W.: Evaluation and Decision Making on Advanced Manufacturing Modes Based on Object-based Knowledge Mesh and User Satisfaction (in press)

Modeling and Simulation of Water Allocation System Based on Simulated Annealing Hybrid Genetic Algorithm Jiulong Zhu1 and Shijun Wang2 1

Zhongyuan University of Technology, 450007 Zhengzhou, China 2 HoHai University, 210098 Nanjing, China [email protected]

Abstract. Presently water resource in most watersheds in China is distributed in terms of administrative instructions. This kind of allocation method has many disadvantages and hampers the instructional effect of market mechanism on water allocation. The paper studies South-to-North Water Transfer Project and discusses water allocation of the node lakes along the Project. Firstly, it advanced four assumptions. Secondly, it analyzed constraint conditions of water allocation in terms of present state of water allocation in China. Thirdly, it established a goal model of water allocation and set up a systematic model from the angle of comprehensive profits of water utilization and profits of the node lakes. Fourthly, it discussed calculation method of the model by means of Simulated Annealing Hybrid Genetic Algorithm (SHGA). Finally, it validated the rationality and validity of the model by a simulation testing. Keywords: water; allocation system; simulated hybrid genetic algorithm (SHGA); modeling.

1 Introduction By transferring water from upstream, midstream and downstream areas of the Changjiang River to North China, three transferring lines of South-to-North Water Transfer Project come into being [1,2]. At the same time, by connecting the Changjiang River, the Huaihe, the Yellow River and the Haihe River, total structure of water in China is formed. The construction of the first stage of eastern route has begun and how to allocate water reasonably and effectively causes public concern. To simplify research, we choose a simple node lake as object and discuss water allocation by means of product distribution theory [3,4]. In order to develop research, the following assumptions have been made: (1) Suppose water consumed to maintain ecological environments in areas around the node lake is provided by the Project and priority-of-guarantee principle is followed. (2) Suppose the distribution of water for residential living use follows population distribution pattern and water buying price is the same when the node lake allocates water. (3) When the node lake allocates water for other uses, idea and competitive mechanism of supply chain management (SCM) is used and water is distributed according to operational mode of quasi-market economy. (4) The conflict between the lake and node districts is solved by consultation, which is illustrated in [4]. R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 104–109, 2011. © Springer-Verlag Berlin Heidelberg 2011

Modeling and Simulation of Water Allocation System Based on SHGA

105

2 Water Allocation Model 2.1 Constraint Conditions The first condition is survival constraint condition which mainly refers that the living of the residents within the areas along the Project must be solved firstly during the process of water allocation. Survival constraint condition is defined as: Qis≥Kibi .

(1)

Where Qis denotes water quantity for living use allocated to the ith district, Ki denotes total population of the ith district, bi denotes per capita water shortage quantity for living use, n is the number of node districts. The second constraint condition is total water resource which is defined as:

∑ (Q n

i =1

is

+ Q ig + Q in + Q it ) ≤ Q .

(2)

Where Q is allocable water amount in the node lake, Qig, Qin, Qit are water quantities allocated to the ith district for industrial, agricultural and ecological uses. The third constraint condition is water demand ability which is defined as: Qis+Qig+Qin+Qit≤Qi max .

(3)

Qis+Qig+Qin+Qit≥Qi min .

(4)

Where Qi max and Qi min are maximum and minimum water shortage quantity in the ith district. The fourth constraint condition is development coordination degree. The paper introduces market operation mechanism into water allocation of the Project. According to [5], constraint of development coordination degree is defined as:

Ξ B1 (ℑ1 )Ξ B2 (ℑ 2 ) ≥ Ξ ∗ . n

Where

1

∑z

i 1

i

,

i 1

(5)

2 Qis Qig Qin Qit Qimin Qimax ,

i 1

B1

1

⎧ 1 ⎨ ⎩exp 4 1

n

2

∑z

i

i 2

,

2

.

i 1

1

,

1 2 1

1

i 2

!i !i0 "i2 "i20 ,

B2

2

exp 4

2

1

B1 is a fuzzy subset of coordination degree between water utilization and economic development, B2 is a fuzzy subset of coordination degree between economic development and environmental improvement, Ξ B1 (ℑ1 ) and Ξ B2 (ℑ 2 ) are membership

ℑ1∗ and ℑ∗2 are optimal ratios, Ξ* is optimal coori dination degree, zi is the ratio of the ith district and zi=1/n, and h is economic growth i i indexes during baseline period and planning period respectively, l 20 and l 2 are functions of coordination degree,

values of environmental improvement of the ith district during baseline period and planning period respectively.

106

J. Zhu and S. Wang

The fifth condition is constraint of water quantity for ecological use. According to the principle of guaranteeing water quantity, this constraint condition is: Qis≥Qismin .

(6)

Where Qismin denotes minimum water shortage quantity for ecological use in the ith district. The sixth condition is non-negative constraint of parameters, so there is: Qis≥0, Qig≥0, Qin≥0, Qit≥0 .

(7)

2.2 Objective Function The goals include comprehensive water utilization and income of the lake. The first goal is comprehensive water utilization. Comprehensive benefit of water utilization includes economic, environmental and social benefits. The paper only discusses economic benefits of water utilization which includes industrial and agricultural production benefits, living benefit and ecological benefit. This goal is defined as: n n n ⎧ n ⎫ f1 (Q ) = Max ⎨∑ eis Qis + ∑ eig Qig + ∑ ein Qin + ∑ eit Qit ⎬ . i =1 i =1 i =1 ⎩ i =1 ⎭

(8)

Where eis, eig, ein, eit is net benefit coefficient of water for living use, for industrial use, for agricultural use and for ecological use in the ith district respectively, n is the number of node districts. The second goal is income of the lake. According to [4,5], each node lake is an independent legal entity and plays the role of distributing water resource. Therefore, the goal of the node lake includes not only maximizing comprehensive benefit of water utilization but maximizing incomes of the node lake. Income goal of the lake is: n ⎫ ⎧ n p Q + ⎪ ∑ s is ∑ p i (Q ig + Q in + Q it )⎪ ⎪ . ⎪ i =1 i =1 f 2 (Q ) = max ⎨ ⎬ n ⎪ ⎪− C ( ) Q Q Q Q + + + c∑ is ig in it ⎪⎭ ⎪⎩ i =1

(9)

Where ps is price of water for living use. According to [4], water price for all districts is the same. In (8), ps is price of water resource for other uses expect living use and price information to all districts is asymmetrical, Cc is unit water cost.

3 Solving Method of the Model In this paper, the solving of the model includes the solving of net benefit coefficient of water utilization and the holistic model. Solving method of the former refers to [5]. This paper only discusses solving method of the latter. Based on [3], this paper adopts Simulated Annealing Hybrid Genetic Algorithm (SHGA) method to solve the model. SHGA mainly includes the following steps:

Modeling and Simulation of Water Allocation System Based on SHGA

107

The first step is encoding and decoding method. In the above model, there are many continuous variables and value ranges of the variables are broad. To improve operational efficiency and accuracy of solutions of SHGA method, we use decimal floating point number to encode tenderly [6,7]. Value ranges of continuous variables are divided into n aliquots and each genic value of chromosomes is denoted by the integers in [1,n+1]. Each variable is encoded according to floating point number. Transformation equation from genic value JYi to true value JCij of genic value of chromosome is as follows:

Where

[JC

[

(

)]

JYij = Qij min + (JYi −1) JCij max − JCij min n . k ij max

]

(10)

, JCijk min is value range of JCij.

The second step is construction of fitness degree function. This paper uses exactly non-differentiable penalty function to deal with constraint conditions: m1 + m2 ⎧ m1 ⎫ + ( ) S JM min (0, S i ( JM )) ⎪ ∑ ⎪∑ i i = m1 +1 ⎪ i =1 ⎪ . G ( JM ) = g ( JM ) ± ς 1 ⎨ m + m + m ⎬ 1 2 3 ⎪+ ⎪ max (0, S i ( JM )) ⎪ i = m∑ ⎪ ⎩ ⎭ 1 + m 2 +1

(11)

Where JM is decision variable after encoding, G(JM) is fitness degree, g(JM) is goal function. If the goal is a minimum one, g(JM) is negative; if the goal is a maximum one, g(JM) is positive. ζ1 is penalty factor, m1, m2 and m3 are the number of constraint conditions of “=”, “≥,>” and ““≤,0.8, which shows that the coordination extent between economic development and environmental improvement in areas around the lake. Thirdly, not all water resources in the node lake are distributed, because the lake must hold some buffer water. This is the reason why allocable water quantity is less than 2×108m3 in the case study. Acknowledgments. This work is supported by Ministry of Education of China (Human Social Science Fund, No.09YJC790069), Decision Bidding Project of Henan

Modeling and Simulation of Water Allocation System Based on SHGA

109

Government (No.B794)(No.B795), Technology Research and Development Project of Zhengzhou Science and Technology Bureau (No.0910SGYG25230-5).

References 1. 2.

3. 4.

5. 6.

7.

Hu, J.L., Ge, Y.X.: Research on the Water Distribute Mode and Harmonized System of Huanghe River. Management World 20, 43–53 (2004) (in Chinese) Zhou, L., Huang, Z.H.: Hybrid Genetic Algorithm for the Multi-objective Nonlinear Water Resources Optimization Model. Water Resources and Power 23, 22–26 (2005) (in Chinese) Zhao, J.S., Wang, J.Z.: Theory and Model of Water Resources Complex Adaptive Allocation System. Acta Geographica Sinica 69, 39–48 (2002) (in Chinese) Huang, F.: Optimal Operation Model for Lagerge-Scale Water Resources System Having Multiple Resources and Multiple Users. Journal of Hydraulic 47, 91–96 (2002) (in Chinese) Li, X.P.: Research Review on Water Configuration. Haihe Water Resource 21, 13–15 (2002) (in Chinese) Ren, C.X., Zhang, H., Fan, Y.Z.: Optimizing Dispatching of Public Transit Vehicles Using Genetic Simulated Annealing Algorithm. Journal of System Simulation 17, 40–44 (2005) (in Chinese) Wen, P.C., Xu, X.D., He, X.G.: Parallel Genetic Algorithm/Simulated Annealing Hybrid Algorithm and Its Applications. Computer Science 30, 21–25 (2003) (in Chinese)

Study on Feed-Forward MAP-Based Vector Control Method of Vehicle Drive Motor Yafu Zhou1, Xiaoyong Shen1, Jing Lian1, Xinhan Sun1, Jun Li2, Minghui Liu2, and Ziliang Zhao2 1

School of Automotive Engineering, Faculty of Vehicle Engineering and Mechanics, State Key Laboratory of Structural Analysis for Industrial Equipment, Dalian University of Technology, Dalian, 116024, P.R. China 2 China FAW Group Corporation R&D Center, Changchun, 130033, P.R. China

Abstract. Contraposing to the shortage of narrow efficient area and over current when vector control method is applied to vehicle drive motors, this paper proposes a feed-forward MAP-based vector control method of vehicle drive motor. According to required motor torque and speed, directly control the magnitude and direction of voltage space vector to realize the aim of motor torque control. So, the calculation of torque current component and field current component is no need, which not only avoids over current that the PID closed-loop control leads to, improving the reliability of controller, but also avoids the dependence on current measurement, improving control precision and motor efficiency. Finally, simulation results and motor bench test prove this method can significantly enlarge efficient area, and it is suitable for vehicle running conditions. Keywords: MAP-based; vector control; drive motor.

1 Introduction With the increase of automobile yield and quantities existed, increasingly pressure on oil demand and environment protection forces the global vehicle industry to seek new energy-saving power system, which accelerates the research progress of electric vehicle technology [1]. Nowadays, many technical problems still need to be solved, one of which is to seek for an efficient and reliable motor control method that is suitable for drive motor operating characteristics [2]. Vector control and direct torque control are two kinds of vehicle drive motor control methods widely used [3]. Vector control is a feed-back loop control. Obtain torque current component and magnetic current component by coordinating transform based on measuring the three-phase current, then carry out PID control according to the deviation between actual current component and demand current component to adjust voltage output vector to control motors [4,5]. However, because the integral link in PID control is hard to meet the rapid response, over current phenomena is inevitable, which will shorten the life of IGBT in motor controller [6]. In addition, current sensor’s measuring error is relatively bigger under small current operating R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 110–115, 2011. © Springer-Verlag Berlin Heidelberg 2011

Study on Feed-Forward MAP-Based Vector Control Method of Vehicle Drive Motor

111

condition, so the motor control performance is poor [7].Direct torque control uses the hysteresis control and switch choice list to control the power inverter to output reasonable voltage vector through observation on motor torque and stator flux to realize the purpose of stator flux linkage and torque control [8]. But the effect of coupling is needed to eliminate during magnetic flux control, so the decoupling model is necessary [9]. And because of obvious current and torque ripple, its dynamic response characteristic is poor [10]. Thus, this paper proposes an improved feed-forward MAP-based vector control method of vehicle drive motor. Directly control the magnitude and direction of voltage space vector to realize the magnetic directional control according to the demand torque and speed. Not only avoid current over caused by closed-loop control, but also avoid the dependence on current measurements, which improves the control precision and efficiency.

2 Control Principle and Process Motor vector control coordinate is shown as in figure1, in which α - β is static coordinate, d-q is synchronous rotation coordinate, U is voltage space vector amplitude, θ is voltage vector phase earlier than q axis, ϕ is rotor machine angle position. During the operating process, d-q coordinate synchronously rotates with the rotor. The magnitude and direction of magnetic field can be controlled by changing voltage amplitude U and phase θ .

q

E

U

d

T

M

D

Fig. 1. Motor vector control coordinate

Drive motor control system includes speed and angle position sensor, motor control three-dimensional MAP inquiring module, space voltage vector PWM module, inverter module, battery and motor, shown as in figure 2, in which T* is command torque, n* is motor speed, U* is optimal voltage amplitude, θ * is optimal voltage phase, ϕ * is rotor machine angle position. First, through calibration experiment form the optimal motor control threedimensional MAP which is stored in motor control program in format of tables. Second, MAP inquiring module builds curved surface differential model, Third, optimal voltage amplitude U*and optimal voltage phase θ * can be obtained by curved surface differential calculation according to T*, n* and parameters in each vertex of the model.

112

Y. Zhou et al.

Last, space voltage vector PWM module gives voltage vector phase θ *+ ϕ *+90° under α - β coordinate, according to θ *and ϕ *. Then produces six PWM control signals to control inverter module to work. battery

T*

motor control threedimensional MAP inquiring module

n*

U*

T*

space voltage vector PWM module

PWM control signals inverter module

motor

M* speed and angle position sensor

Fig. 2. Principle diagram of drive motor control system

3 Simulation Analysis Feed-forward MAP-based vector control simulation model for permanent magnet synchronous motor is built under Simulink, shown in figure 3. The model mainly includes permanent magnet synchronous motor module, inverter module, MAP inquiring module, space voltage vector PWM module, speed and position detection module, etc. MAP inquiring module is used for giving the optimal voltage and phase; space voltage vector PWM module is used for producing six PWM control signals.

Fig. 3. Simulation model

Simulation experiment is carried out under the model. Experimental conditions are as follows: initial speed is 700rpm, initial load is 6Nm, and load suddenly drops to 2Nm at 0.04s. Speed curve and torque curve are respectively shown in figure 4 and 5. Speed increases rapidly at first, and becomes maximum value 730rpm at 0.005. Then fluctuates around 700rpm, and quickly steady at 700rpm at 0.23s. In the case of load reduction suddenly at 0.04s, speed fluctuates, but soon stabilizes at 700rpm. Torque fluctuates greatly at first, and peak value is 33Nm. Soon it stabilizes at 6Nm at 0.23s. In the case of load reduction suddenly at 0.04s, torque fluctuates, but soon stabilizes at 2Nm. From simulation results we can see that the system has good speed and torque control performance.

Study on Feed-Forward MAP-Based Vector Control Method of Vehicle Drive Motor

Fig. 4. Speed curve

113

Fig. 5. Torque curve

4 Experiment Analysis 4.1 Drive Motor Experiment Platform Building In order to further verify the feasibility of vehicle drive motor control method and indepth study motor control method and calibration method, we build drive motor experiment platform. Figure 6 is experimental platform system diagram, in which 1 is host-computer, 2 is battery, 3 is power analyzer, 4 is on-line battery charger, 5 is eddy current dynamometer, 6 is bench, 7 is the tested motor, 8 is the tested motor controller, 9 is cooling system, 10 is CAN communication transceiver, 11 is dynamometer controller. The actual experimental platform is shown in figure 7.

Fig. 6. Experiment platform system

Fig. 7. Picture of experiment platform

4.2 Experimental Results and Analysis Motor calibration experiment, motor efficiency experiment and generator efficiency experiment are carried out under the built experimental platform. Motor Calibration Experiment. Motor parameters are very important to control performance. But the calculated parameters during motor design process often drift from the actual value as a result of material parameter errors, ignoring secondary

114

Y. Zhou et al.

factor, environmental temperature’s changing, machining precision influence, etc. Thus, motor calibration experiment has to be taken under different cycle conditions to form the optimal motor control three-dimensional MAP before high-performance motor control. Motor calibration experiment data are shown in Table 1. From the data we can find that motor efficiency at the same motor operating point is different with different voltage amplitude and angle, which is not considered by traditional motor control methods. If select optimal voltage amplitude and phase which obtain the highest efficiency at each motor operating point to form motor control MAP, then carry out motor control based on the MAP, advantages can be fully exerted and the highest efficiency can be obtained. Table 1. Motor MAP calibration experiment data Number Speed/r Torque/ Angle/° pm Nm

1 2 3 4 5 6 7

2048 2048 2048 2048 2048 2048 2048

48 48 48 48 48 48 48

26 27 29 30 31 32 33

Amplitude /V

Efficiency

158 156 154 152 149 147 144

0.89 0.90 0.89 0.89 0.88 0.87 0.87

Motor Efficiency and Generator Efficiency Experiment. Motor efficiency can be calculated by measuring motor’s input voltage, input current, output speed, output torque at each motor operating point on n-T map. The motor efficiency threedimensional surface is shown in figure 8. When the motor runs as a generator, generator efficiency can be calculated by measuring generator’s input speed, input torque, output voltage, output current at each motor operating point. The generator efficiency three-dimensional surface is shown in figure 9.

Fig. 8. Motor efficiency graphic

Fig. 9. Generator efficiency graphic

Study on Feed-Forward MAP-Based Vector Control Method of Vehicle Drive Motor

115

What can be seen from figures above is motor efficient area is enlarged, up to 80%, which can meet the demand of keeping high efficiency during large speed and torque range for vehicle drive motors.

5 Conclusions This paper studies MAP-based feed-forward vector control method deeply which is suitable for vehicle drive motor. Motor efficiency at the same motor operating point is different with different voltage amplitude and angle, so it is possible to improve motor efficiency based on MAP formed by the optimal voltage amplitude and angle. Simulation and bench experiments show that the motor control method significantly expands motors’ efficiency area. It is an efficient and reliable control method for vehicle drive motor. Acknowledgments. This project is supported by China Postdoctoral Science Foundation (20100471431) and 863 National High Technology Research and Development Program of China (2008AA11A140).

References 1. Haddoun, A., Benbouzid, M.H., Diallo, D.: A loss-minimization DTC scheme for EV induction motors. J. IEEE Transactions on Vehicular Technology 56(1), 81–88 (2007) 2. Timko, J., Zilková, J., Girovský, P.: Shaft sensorless vector control of an induction motor. J. Acta Technica CSAV (Ceskoslovensk Akademie Ved) 52(1), 81–91 (2007) 3. Kumar, R., Gupta, R.A., Bhangale, S.V.: Vector control techniques for induction motor drive: A review. J. International Journal of Automation and Control 3(4), 284–306 (2009) 4. Nait Seghir, A., Boucherit, M.S.: A new neural network based approach for speed control of PM synchronous motor. J. WSEAS Transactions on Circuits and Systems 6(1), 87–93 (2007) 5. Badsi, B., Masmoudi, A.: DTC of an FSTPI-fed induction motor drive with extended speed range. J. COMPEL - The International Journal for Computation and Mathematics in Electrical and Electronic Engineering 27(5), 1110–1127 (2008) 6. Vaez-Zadeh, S., Jalali, E.: Combined vector control and direct torque control method for high performance induction motor drives. J. Energy Conversion and Management 48(12), 3095–3101 (2007) 7. Trentin, A., Zanchetta, P., Gerada, C.: Optimized commissioning method for enhanced vector control of high-power induction motor drives. J. IEEE Transactions on Industrial Electronics 56(5), 1708–1717 (2009) 8. Kadjoudj, M., Taibi, S., Benbouzid, M.E.H.: Permanent-magnet-synchronous-motor speed control using fuzzy adaptive controller. J. Advances in Modeling and Analysis C 62(3-4), 43–55 (2007) 9. Vavrus, V., Vittek, J., Malek, M.: Velocity vector control of a linear permanent magnet synchronous motor. J. Komunikacie 9(4), 14–19 (2007) 10. Singh, B., Jain, P., Mittal, A.P.: Sensorless DTC IM drive for an EV propulsion system using a neural network. J. International Journal of Electric and Hybrid Vehicles 1(4), 403– 423 (2008)

Condition Monitoring and Fault Diagnosis of Wet-Shift Clutch Transmission Based on Multi-technology* Man Chen1, Liyong Wang2, and Biao Ma1 1

School of mechanical and vehicle engineering, Beijing Institute of Technology, Beijing 100081, China 2 School of Mechanical & Electrical engineering, Beijing Information Science & Technology University, Beijing 100192, China

Abstract. Based on the construction feature and operating principle of the wet-shift clutch transmission, the condition monitoring and fault diagnosis for the transmission of the tracklayer with wet-shift clutch were implemented with using the oil analysis technology, function parameter test method and vibration analysis technology. The new fault diagnosis methods were proposed, which are to build the gray modeling with the oil analysis data, and to test the function parameter of the clutch press, the rotate speed of each gear, the oil press of the steer system and lubrication system and the hydraulic torque converter. It’s validated that the representative function signals were chosen to execute the condition monitoring analysis, when the fault symptoms were found, and the oil analysis data were used to apply the gray modeling to forecast the fault occurs time can satisfy the demand of the condition monitoring and fault diagnosis for the transmission regular work. Keywords: Wet-shift Clutch, Transmission, Oil analysis, Condition Monitoring, Fault Diagnosis.

1 Introduction At present, wet shift clutch transmission is widely used in tracked armored vehicle and engineering machinery industry. This transmission device is also called wet-shift clutch transmission which is with a compact structure, a high transmission efficiency, and is easy to operate, However, to test and control the running state of the transmission quickly has been the most important issue that is imperative to be solve, which is also the key for ensuring the riding quality of the transmission and insuring to launch the maintenance in time[1].

，

*

This work were supported by national natural science foundation under Grant No. 50975020 and Key Laboratory foundation of BeiJing under Grant No. KF20091123205.

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 116–123, 2011. © Springer-Verlag Berlin Heidelberg 2011

Condition Monitoring and Fault Diagnosis of Wet-Shift Clutch Transmission

117

1.1 Structural Principle of Wet-Shift Clutch Transmission Wet-shift clutch transmission mainly consists of non-stop speed-change mechanism and hydraulic steering mechanism. Non-stop mechanism includes six wet shift clutches and several gear pairs. Clutches are respectively placed on I axis and III axis, power is input from I axis, transmitted through II axis, and output from III axis, and the combination of clutches can shape several forward gears and reverse gears[2]. The construction features of the straight drive shift and steering equipment of the transmission of the tracklayer with wet-shift clutch are introduced firstly, as shown in Fig.1.

Fig. 1. Structural Principle Diagram of A Transmission Device

1.2 Hydraulic Lubrication System of Wet-Shift Steering Transmission The hydraulic lubrication system is an import part of wet-shift steering transmission, it mainly completes the following functions: (1) provide compensatory oil for hydraulic torque converter to transmit the power. (2) provide control oil for the transmission manipulation system of the transmission device. (3) provide cutting oil for cooling and friction elements and lubrication transmission clutch’s friction plates. (4) provide pressure oils for steering pump motors. See Figure 2 for the sketch map of oil circulation of hydraulic lubrication system.

118

M. Chen, L. Wang, and B. Ma

Fig. 2. Sketch Map of Hydraulic Lubrication System of Power-shift steering transmission

A shown in Figure 2, oil is mainly divided into three circulating subsystems: (1) hydraulic torque converter’s hydraulic transmission oil, radiator oil, and transmission lubrication oil; (2) transmission clutch control oil and transmission lubrication oil; (3) hydraulic pump motor system’s steering power flow transmission force oil and bearing lubrication and cooling oil. Three oil circulating systems above have different flows and pressure changes. The working process of Hydraulic lubrication system was analyzed in this paper, and pointed out that the oil analysis technology plays an important role in wet-shift clutch transmission wear site and process monitoring, wear failure type and wear mechanism studying, and oil evaluation; besides, it is an important means to perform condition monitoring and fault diagnosis on mechanical equipment without stopping and disassembling[3].

2 Fault Diagnose Method of Wet-Shift Clutch Transmission The method and steps of condition monitoring and fault diagnosis for the transmission of the tracklayer with wet-shift clutch with using the oil analysis technology, function parameter test method and vibration analysis technology were introduced, as shown in Fig.3. As show in fig.3, function parameter test , oil analysis and vibration test were used to monitoring wet-shift clutch trans- mission, and function parameter test is intuitive and high efficiency, but there are many kinds of function parameters of wet-shift clutch transmission, and it is impossible to test all the parameter. So it is very important to

Condition Monitoring and Fault Diagnosis of Wet-Shift Clutch Transmission

119

Fig. 3. Schematic drawing for state monitoring and fault diagnosis of wet-shift clutch transmission

chose representative signals to test. Because faults symptom can be catch by using oil analysis, so it is another efficiency method to monitor- ring and diagnose transmission fault. Vibration monitoring is the third method in transmission state monitoring, but because the influence of running, it is difficult to test and analysis the vibration of wet-shift clutch transmission, so vibration monitoring is only the supplementary means.

120

M. Chen, L. Wang, and B. Ma

3 Gray Modeling Fault Diagnose Based on Oil Analysis In this study, considering the finiteness of the data quantity and the randomness of the variety of the sample[4], the gray theory was selected to forecast the wear extent and the select method and modeling steps of the GM(1,1) gray modeling with using the oil analysis data was proposed, as shown in Fig.4.

Fig. 4. the gray modeling steps for fault diagnose by oil analysis data

As show in fig.4, as to the sampled data with unequal distance, judge whether the oil data and sampled data collected from each sensor meet the scale requirements. If not, judge again after calculating the logarithm of concentration of elements in oil and other data. If it meets the scale condition, continue to judge whether it meets the sampling

(

interval for modeling PD[ ΔW N scale condition is

：

)

(

N

 . Δt k is the sampling interval. The

 ⎛ −  ⎞ Q + Q + ⎟ ⎜ and δ is scale function, n is the δ (N ) ∈ H H ⎜ ⎟ ⎝ ⎠

，

number of sampled data. When it meets the

(

) PLQ(ΔW ) <

)

sampling

interval

requirements

for

modeling:

PD[ ΔW N PLQ ΔW N <  , build the gray model through the difference quotient method. If it doesn’t meet the modeling condition, sampling intervals shall be optimized.

Condition Monitoring and Fault Diagnosis of Wet-Shift Clutch Transmission

121

After the optimization of sampling intervals, it should to judge whether the sampling rule is smooth or not. If it is, the model shall be built as per grey modeling in time-sequence fitting method. If there is data fluctuation, the grey model shall be built through the improved time-sequence fitting method based on the wear curve. After the grey model is built through the difference quotient method, time-sequence fitting method and time-sequence fitting method based on the wear curve, the fitting precision shall be checked. If it passes the check, grey prediction could proceed. If it doesn’t pass the check, it shall be remodeled by changing the modeling dimension until it meets the requirements for fitting precision.

4 Fault Diagnose Based on Function Parameter The test parameters for transmission in this study mainly include: control oil press, shift clutch press, rotate speed of each axis, oil system of the steering system and the hydraulic torque converter. The sequence and method to test each parameter were demonstrated as shown in Fig.5. As show in fig.5, check firstly the gear transmission pressure based on signals of control pressure provided by the front pump of gearbox, the pressure of gear shift clutch and lockup clutch. If the pressure of clutches at gears being tested is normal, determine the gear of existing gearbox. In case of any anomaly, record the amplitude of abnormal signals to indicate the anomaly of the existing gear. When the existing gear is determined, convert and display the current input of revolution speed and car speed based on the input revolution speed and three-axis revolution speed collected; calculate the transmission ratio of current gearbox based on the input and output revolution speed. If the calculated transmission ratio equals to that for this gear, it indicates “normal state” for the current gear. If the calculated transmission ratio doesn’t equal to that for this gear, it will diagnose and find the specific position for the faulty position and display the diagnostic result. After the current gear is displayed, check the pressure of the hydraulic circuit at the high-pressure outlet and low-pressure inlet of the steering system. If the pressure of the high/low-pressure hydraulic circuit is within the normal ranges, it will indicate “normal pressure” for the steering system. If the pressure of the high/low-pressure hydraulic circuit exceeds the range of technical indexes, it will record and indicate the faulty position and specific value. After the steering system is detected as normal, check whether the lubrication pressure in the gearbox’s lubrication system is normal, measure mainly the lubrication pressure at first, second and third transmission shafts. If the lubrication pressure of each shaft is within the range of normal technical indexes, it will indicate “normal lubrication pressure” for the corresponding shaft. If the lubrication pressure at each shaft exceeds the range of technical indexes, it will record and indicate the faulty position and specific value. The gearbox’s temperature sensor is installed inside the lubrication system’s hydraulic circuit. Before detecting the lubrication system’s pressure, the diagnosis system will display the gearbox’s temperature.

122

M. Chen, L. Wang, and B. Ma

Fig. 5. Test method for each function parameter

After the lubrication system is detected as normal, the hydraulic pressure at inlet and outlet of the hydraulic torque converter is detected. If the pressure difference between inlet and outlet of the torque converter is within the range of normal technical indexes,

Condition Monitoring and Fault Diagnosis of Wet-Shift Clutch Transmission

123

it will indicate “normal pressure” for the hydraulic torque converter. If the pressure exceeds the range of technical indexes, it will record and indicate the faulty position and specific value. At last, store signals and faulty messages collected, generate inspection report and print.

5 Conclusions In this study, it’s indicated that the oil analysis and function parameter are the mainly methods for wet-shift clutch transmission condition monitoring, and the vibration analysis is the assistant method. It’s validated that the representative function signals were chosen to execute the condition monitoring analysis, when the fault symptoms were found, and the oil analysis data were used to apply the gray modeling to forecast the fault occurs time can satisfy the demand of the condition monitoring and fault diagnosis for the transmission regular work.

References [1]

[2]

[3] [4]

Wang, L., Ma, B., Zheng, C., et al.: A Study on Running-in Quality Evaluation Method of Power-shift Steering Transmission based on Oil Monitoring. Lubrication Engineering 7(33), 35–38 (2008) Li, H.-y., Wang. L.-y.: A Study on no-load Running-in Wear of Power-shift Steering Transmission based on Spectrum Analysis Ma Biao. Spectroscopy and Spectral Analysis 29(4), 1013–1016 (2009) Wang, L., Ma, B., Li, H., et al.: A Study on Running-in QualityEvaluation Method of Power-shift. Steering Transmission based on Performance Parameter 3(33), 86–88 (2008) Deng, J.: Grey control system, Wu Han. Huazhong Institute of Technology Publisher, Chinia (1985)

Circulant Graph Modeling Deterministic Small-World Networks Chenggui Zhao* School of Information, Yunnan University of Finance and Economics, Kunming, 650221, China [email protected]

Abstract. In recent years, many research works have revealed some technological networks including internet to be small-world networks, which is attracting attention from computer scientists. One can decide if or not a real network is Small-world by whether it has high local clustering and small average path distance which are the two distinguishing characteristics of small-world networks. So far, researchers have presented many small-world models by dynamically evolving a deterministic network into a small world one by stochastic adding vertices and edges to original networks. Rather few works focused on deterministic models. In this paper, as a important kind of Cayley graph, the circulant graph is proposed as models of deterministic small-world networks, thinking if its simple structures and significant adaptability. It shows circulant graph constructed in this document takes on the two expected characteristics of small word. This work should be useful because circulant graph has serviced as some models of communication and computer networks. The small world characteristic will be helpful to design and analysis of structure and performance. Keywords: Circulant graph, small-world networks, clustering coefficient average path length.

1 Introduction The research on small world networks has received a number of literatures in recent years. In 1998, Watts and Strogatz [3] called those real networks with phenomenon of six degrees of separation as small world networks. They described some stochastic models of small-world networks. These models have clustering coefficients much larger than those of random networks, but a small average path length. Some regular networks have large clustering coefficient and large average path length, while the random networks with the same size and average node degree have much smaller clustering coefficient and average path length. D.J. Watts et al. randomly rewire the edges with probability p on regular network such as a loop network to construct small world network such that average path length dramatically decrease as p increases but the clustering coefficient decreases slowly. *

Corresponding author.

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 124–127, 2011. © Springer-Verlag Berlin Heidelberg 2011

Circulant Graph Modeling Deterministic Small-World Networks

125

Comellas and Sampels [2] discovered that constant-degree deterministic smallworld networks are also feasible. Deterministic models for small-world networks can facilitate the understanding of their behavior. This approach also permits a direct calculation of relevant network parameters and thus allows the design of specific small-world networks for applications such as new communication systems and computer architectures. In this paper we also begin with loop network to construct a small world network called circulant graph; but we rewire edges on loop networks according to a way of construction of Cayley graph other than stochastic way used in [3]. The results show circulant graph holds characteristics of small world networks and can be adapted to model them. For simple structures as well as small-world characteristics of our construction, this model can facilitates some members of Cayley graph family to be selected as technological network models.

2 A Cayley Graph Adapted to Model Small World Networks 2.1 Basic Definitions Most of Cayley-graph have been used to design interconnection networks. The circulant graph is a important class of Caylay graph family.A digraph X=(V,E) is defined by a set V of vertices and a set E={(u,v)| u, v∈V} of arcs. The subset E is symmetric, so we can identify two opposite arcs (u, v) and (v, u) by the undirected edge (u,v). Let G be a finite group and S a subset of G. If very element of G can be expressed as a finite product of the powers of the elements in S, the subset S is said to be a generating set for G and elements of S are called generators of G, In this case, we also say that G is generated by S. Let Cay (G, S) denotes a graph with vertices that are elements of G and arcs that are ordered pairs (g, gs) for g∈G, s∈S. Cay (G, S) is called Cayley digraph of group G and the subset S, If S is a generating set of G then Cay (G, S) is called the Cayley digraph of G generated by S. If the identity 1 element of G is not include in S and －1 S=S , then Cay (G, S) is a simple and undirected graph. Unless noted otherwise, all graphs in this paper are undirected graphs. 2.2 A Special Construction of Circulant Graph Firstly, let us have a special class of Cayley graph by following definition. It has been applied to model computer networks for many years. Definition. A circulant graph X (n, S) is a Cayley graph Cay (Zn, S) on Zn where n is a power of 2. That is, it is a graph whose vertices are labelled {0, 1,…,n − 1}, with two vertices labelled i and j adjacent if and only if i − j (mod n) ∈S, where S as a subset of Zn has S = −S and 0 ∉ S. Secondly, there will be constructed a particular circulant graph by a deliberate way to select generating set S of Zn. if |n|=dt, where d is a factor of n and you can select d

126

C. Zhao

to obtain different node degree of Cay(Zn, S). Let S={1,t, 2t,…, dt}.Then Cay (Zn, S) is a cayley graph because S is clearly a generating set of Zn by noting it includes the generator 1 of Zn . Together with S =− S, It is sure Cay (Zn, S) is undirected and connected. Let symbol X substitute for Cay (Zn, S) for simplicity. Next section shows X has the Small-World characteristics: large coefficient and small average path length.

3 The Small-World characteristics of Cay (Zn, S) 3.1 Clustering Coefficient Distinctly, X is a Cayley graph of constant degree | S|=d+1. Between d+1 neighbors of node v in X, the ratio Cv of actually existing edges to possible edges is known as the clustering coefficient of node v. The average of Cv over all node v of G is the clustering coefficient C of the graph X. For the symmetry of X, It holds Cv=C such that only the clustering coefficient of identity node e need to be considered. The set of neighbors of node e is S. There exists an edge between element s1 and s2 of S if and only if s1s2− 1 is also an element of S. If H is a subset of S and H {1} becomes a group then as long as s1, s2∈H and sequentially s1s2− 1∈H node s1, and s2 will be joining. Consequentially, there are at least |H|(|H|− 1) edges existing among the neighbors of 1, which leads to X has the clustering coefficient

∪

C ≥

| H | (| H | −1) | S | (| S | −1)

(1)

Now, for circulant graph constructed in before section, let H=S −{1} then |H|=d. By (1) we have C ≥(d− 1)/(d+1). It means one can get a large C by select a large d such that resulting circulant graph becomes small-world network. 3.2 Average Path Length Obviously, the undirected Cayley Graph X(G,S) of a group G with respect to the generating set S o is a regular graph of degree d. Babai et al. proved that every group of order n has log2 n+ O(log log n) elements x1,...,xt forms a generating set of G. It follows that G has a set of log2 n+ O(log log n) generators such that the resulting Cayley graph has a logarithmic diameter. So X has average path length no lager than log2n. For a general survey on Cayley graphs with small diameters see [4].

4 Conclusion We have constructed a good model for small-world network by a simple way. It can be believed that this model will improve some designs based on circulant graph in technological network according such that they become small-world networks. On the other way, this model may also service as model of deterministic small-world network.

Circulant Graph Modeling Deterministic Small-World Networks

127

References 1. 2. 3. 4. 5.

6.

Wenjun, X., Parhami, B.: Cayley graph as models of deterministic small-world networks. Information Processing Letters 97, 115–117 (2005) Comellas, F., Sampels, M.: Deterministric small-world networks. Physica A 309, 231–235 (2002) Watts, D.J., Strogatz, S.H.: Collective dynamic of small-world networks. Nature 393, 440– 442 (1998) Heydemann, M.C., Ducourthial, B.: Cayley graphs and interconnection networks. Graph Symmetry, Algebraic Methods and Applications. NATO ASI C 497, 167–226 (1997) Xiao, W.J., Chen, W.D., Parhami, B.: On Necessary Conditions for Scale-Freedom in Complex Networks, with Applications to Computer Communication Systems. Int’l J. Systems Science (to appear) (e-publication in March 2010) Xiao, W.J., Peng, L., Parhami, B.: On General Laws of Complex Networks. In: Zhou, J. (ed.) Proc. 1st Int’l Conf. Complex Sciences, Part 1, pp. 118–124. Springer, Heidelberg (2009)

Research on Risk Manage of Power Construction Project Based on Bayesian Network Zhengyuan Jia, Zhou Fan, and Yong Li School of Business Administration, North China Electric Power University, Baoding 071003, China [email protected], [email protected]

Abstract. With China's changing economic structure and increasingly fierce competition in the market, the uncertainty and risk factors in the projects of electric power construction are increasingly complex, the projects will face huge risks or even fail if we don't consider or ignore these risk factors. Therefore, risk management in the projects of electric power construction plays an important role. The paper emphatically elaborated the influence of cost risk in electric power projects through study overall risk management and the behavior of individual in risk management, and introduced the Bayesian network to the project risk management. The paper obtained the order of key factors according to both scene analysis and causal analysis for effective risk management. Keywords: power construction projects; risk management; Bayesian networks.

1 Introduction Risk management of power construction project is to identify, analyze, evaluate, predict and treat effectively the risk in construction project. It must deal with the uncertain factors and reduce the cost of project, so as to finish the project security and scientific. Because of the high input and high risk, the risk management is an important part of project management, and advanced technology is must be used in project management. This article discusses the risk management of power construction projects based on Bayesian network, and carry out the quantitative management for risk management.

2 The Bayesian Network As an important probability model, Bayesian networks is proposed by professor earl J.P, the university of California in 1986. Bayesian network usually consists of two parts, the first part is the Bayesian network structure, called a directed acyclic graph (DAG directed acyclic graph). Each node of structure represents a risk event, the condition of node is corresponding to the reliability of risk events, the arc represents a causal relationship between risk events. The second part is the probability set of Bayesian networks, and each node has a conditional probability that represents the relationship with the parent node. Bayesian network is shown in Figure 1. R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 128–134, 2011. © Springer-Verlag Berlin Heidelberg 2011

Research on Risk Manage of Power Construction Project Based on Bayesian Network

129

Fig. 1. Bayesian network

Figure 1, Node E1 , E 2 ,

E 3 represent the different risk events, P ( E 3 / E1 ) and P ( E 3 / E 2 ) denote the event in the E1 and E 2 with the probability of E3 , it indicate that exist certain causal relationship among events E1 , E 2 , E 3 . Use Bayesian networks for risk management, is to take the risk events for the center, detect the probability of events based on existing data and determine the probability according to the formula.

P(e1 , e2 ,LL , e n ) = P (ei P( E k )) P (e1 , e 2 ,LL , e n ) is the probability of corresponding of (e1 , e2 ,LL , en ) , P (ek ) is the status of parent node for E k . Where,

(1) state

3 Construction of Bayesian Network for Risk Management 3.1 The Relationship between Bayesian Network and the Power Risk Management The uncertain factors is the main problem of risk management for modern electrical engineering, and because of the advantage of the Bayesian network, it is more agile and scientific. Then, as building power projects important basis for risk management models, Bayesian networks is mainly carried out based on the following points: (1)Bayesian network has a solid theoretical foundation, and its reasoning result is more acceptable. At the same time, the independencies of Bayesian network can express the relationship between power engineering project and risk. (2)There is mature software and reasoning algorithm for Bayesian network, the key is to get the value of conditional probability. Application of the Bayesian networks, the probability of reasoning algorithm of information, can be required at low under the condition of incomplete information on reasoning. (3)The expression of Bayesian network is more suitable for risk prediction. We can improve the network structure and parameters through reasoning process, and update the information of probability.

130

Z. Jia, Z. Fan, and Y. Li

3.2 Construction of Bayesian Network for Cost Management of Power Project Bayesian network construction can study and artificial construct two ways. In the training sample fully circumstance, can learn from these data structure of the network. Construction of Bayesian network can be constructed by self learning and artificial. In the case of the full training sample, you can learn network structure from the data. For risk management system, the risk is uncertain, and there have no sufficient training data, therefore, it must be personally establishment and application for Bayesian network. Constructing Bayesian Networks for risk management is divided into the following three steps: (1)Determine the contents of the node Bayesian network composed of nodes, different nodes corresponds to different risk events. Therefore, we must determine the risk event exist in project implementation process. Project manage contain nine knowledge areas: scope management, time management, cost management, human resources management, risk management, quality management, purchasing management, communication management and integrated management. For the knowledge of the importance of the project area, depth analysis should be made to analyze the impact of factors at all levels and the corresponding factors of the properties, but little impact on project implementation to run the knowledge, can be used for shallow analysis. The hierarchical structure refers to a field of knowledge in risk events. (2)Determine the relationship between nodes Determine the node content need to follow a certain method to determine the relationship between each node and the Bayesian network inference. In Bayesian network inference, the common are 3 types: causal reasoning, diagnostic reasoning and support reasoning. Here, the hierarchical structure refers to a field of knowledge in this field. (A)Causal inference. The conclusion is obtained by reason from top to down, the objective is to get the result. Known for a reason (evidence), through using the Bayesian network inference, we can obtain the probability. (B)Diagnosis reasoning. By structure of reasoning from down to up, the purpose is to find the factors which may happen and find out the causes of risk as well as risk of probability. The purpose of using the reasoning is control the risk in time, find its source and prevent the recurrence when the risk occurs. (3)The probability distribution Probability distribution consists of two parts: determine the probability of the toplevel parent node P ( Ei ) , and determine the conditional probability P ( E i / Pb ( E i )) . Probability distribution between the events needs to have extensive knowledge of risk management, and the probability is determined by expert of risk management according to their experience. Therefore, in building Bayesian networks, the system already contains an authoritative expert knowledge. Because the feature of project is complex and reversible, the above step 3 can be alternately and repeatedly modified until the construction of Bayesian networks is finished.

Research on Risk Manage of Power Construction Project Based on Bayesian Network

131

4 Empirical Analysis There are so many risks in electrical construction project that the paper takes the risk management of cost for example. The cost risk is an important part of the management in electric power construction project, cost risk would lead to direct economic risk and the project will faces enormous economic losses, it will also cost large amounts of money if we do not take cost risk into account. 4.1 Establishing the Model of Bayesian Network for Cost Risk To establish a network, it needs to do the following steps: Because the power project risk management of historical data, and in lack of relations between nodes, according to expert knowledge of setting, causal inference and diagnosis reasoning, establishing risk two-way tracking. As the lack of historical data in determining the relationship between the nodes, most relationship is set based on expert’s knowledge, then, establish a two-way tracking map according to re-use causal reasoning and diagnostic reasoning.

Fig. 2. Bayesian network of cost control

In the Bayesian network of cost control, we can see the influence and relationship among different nodes. Determination of node status is still obtained by the experts, based on experience and knowledge, and it also can be characterized according to the node's own analysis of the data obtained. In our model, there are 3 status in each node 1, 2, 3, corresponding to low-level, intermediate, advanced.

132

Z. Jia, Z. Fan, and Y. Li

4.2 The Dynamic Assessment of Risk Based on Bayesian Network This figure is two-way tracking Bayesian due to the original acyclic graph on the basis of the development of risk, so the Bayesian reasoning is performed at the same time. According to the Figure 2, reason the result according to the expert knowledge, causal inference and diagnosis reasoning. When the technology factor risk rating is E = l, probabilistic relationship is as follows:

P( E = 1 K = 1, W = 1, C1 = 1, C 2 = 1, C 3 = 1, J = 1) P( E = 1 K = 2, W = 1, C1 = 1, C 2 = 1, C 3 = 1, J = 1) …

…

… …

…

… …

…

… …

P( E = 1 K = 3, W = 3, C1 = 3, C 2 = 3, C 3 = 3, J = 3) When the risk level is 2 or 3, the probability of relationships is as above. Thus, to C 3 = 1, E = 1 , for example, the probability of technical risk:

P = (C 3 = 1 E = 1) =

P (C 3 = 1, E = 1) P ( E = 1)

Other situations are similar. According to the assumption network diagram, the resources are composed of six overlapping part, each part can produce risk and the probability are P ( B 1 ) , P ( B 2 ) ,…, P ( B6 ) , P ( Bi ) there

P ( Bi ) > 0

are

(i = 1,L ,6)

and

>0

(i = 1,L ,6) and then,

Bi ∩ B j = φ (i ≠ j )

，

⎡ 6 ⎤ n P (Ω) = P ⎢∪ B ⎥ = ∑ P( Bi ) = 1 . ⎣ i =1 ⎦ i =1 P (K ) =0.1, P (W ) =0.2 P (C1 ) =0.1 P(C 2 ) =0.3 P(C 3 ) =0.2 and P (J ) =0.1, after further analysis , the error caused by the control and improvement measures is K , the incorrect in error analysis is W , inappropriate cost control is C1 , inappropriate cost forecasting methods is C 2 , the probability of C0 caused by misconduct of cost conRespectively, on the assumption that the probability of the above

，

，

，

，

trol C 3 and change records J are:

P(C 0 K ) = 0.2 , P(C 0 W ) = 0.1 , P(C 0 C1 ) = 0.25 , P(C 0 C 2 ) = 0.1 , P(C 0 C 3 ) = 0.15 , P(C 0 J ) = 0.2 .

Research on Risk Manage of Power Construction Project Based on Bayesian Network

133

Thus, according to the previous algorithm, the assessment result is the following: (1) Basic evaluation results If the cost control, the risk control and improvement measures "K" is caused by the possibility of risk for: In the event of risk control for cost, the probability of K is: P1 = P ( K C0 ) =

P( K ∩ C0 ) = P(C0 ) P(C0 K ) P( K )

P( K ) P(C0 K ) + P(W ) P (C0 W ) + P(C1 ) P(C0 C1 ) + P(C2 ) P(C0 C2 ) + P(C3 ) P(C0 C3 ) + P( J ) P(C0 J ) =

0.1× 0.2 = 0.14 0.1 × 0.2 + 0.2 × 0.1 + 0.1 × 0.25 + 0.3 × 0.1 + 0.2 × 0.15 + 0.1 × 0.2

Similarly, the probability of W is:

P2 = P (W C 0 ) =

P(W ∩ C 0 ) = 0.14 P (C 0 )

The probability of C1 is:

P3 = P (C1 C 0 ) =

P (C1 ∩ C 0 ) = 0.17 P (C 0 )

The probability of C2 is:

P4 = P (C 2 C 0 ) =

P (C 2 ∩ C 0 ) = 0.21 P (C 0 )

The probability of C3 is:

P5 = P (C 3 C 0 ) =

P (C 3 ∩ C 0 ) = 0.21 P (C 0 )

The probability of J is:

P6 = P ( J C 0 ) =

P( J ∩ C 0 ) = 0.14 P (C 0 )

134

Z. Jia, Z. Fan, and Y. Li

(2)Overall assessment According to the above analysis, we can draw the conclusion: the project cost control risks level:

P(C0 ) = P(K)P(C0 K) + P(W )P(C0 W ) + P(C1 )P(C0 C1 ) + P(C2 )P(C0 C2 ) + P(C3 )P(C0 C3 ) + P(J )P(C0 J ) = 0.1× 0.2 + 0.2 × 0.1 + 0.1× 0.25+ 0.3× 0.1 + 0.2 × 0.15+ 0.1× 0.2 = 0.145 Because there are many factors affect the cost control, the paper just study few factors of risk based on Bayesian networks, discussion of local risk factors can also evaluate the overall risk. Because the risk factors of overall cost are more complex, this article only gives the cost of local risk assessment model and corresponding algorithm. With reference to the example of calculation and analysis, the comprehensive assessment of actual cost can be completed.

5 Conclusion Risk management based on the application of Bayesian network model can solve the problem caused by lacking of historical data, and we also can get the order of key factors through scenario analysis and causal analysis, the aim is to control the project more effectively. However, the analytical method in the paper is hypothetical, we must use real data in practical work so as to get more valid results and make full use of the Bayesian network in risk management.

References 1. 2. 3. 4. 5. 6.

Li, D.J.: Based on Bayesian network serial decoding method. Communication Technology 115(4), 38–40 (2001) Zhang, S.: Bayesian Networks in Decision Support Systems. Computer Engineering (10), 1–3 (2004) Jia, Z., Fan, Z., Jiang, M.: Distribution Network Planning Based on Entropy Fuzzy Comprehensive Method. In: 2010 International Conference on AMM, pp. 26–28, p. 780 (2010) Evergreen, F.: Engineering factors affecting the cost of construction of the project. Shijiazhuang Railway Institute (11), 158–160 (2003) Zang, W.Y.: Bayesian Networks in stock index futures early warning of the risk. Science of Science and Technology Management (10), 122–125 (2003) Cooper, G.: Computational complexity of probabilistic inference using Bayesian belief networks. Artificial Intelligence 42(2), 393–405 (1990)

The Design of Logistics Information Matching Platform for Highway Transportation* Daqiang Chen1, Xiaoxiao Zhu1, Bing Tong1, Xiahong Shen1, and Tao Feng2 1

College of Computer Science & Information Engineering, Zhejiang Gongshang University, No.18, Xuezheng Str.,Xiasha University Town, Hangzhou, 310018, China 2 College of Economics and Business administration, China University of Geosciences, No. 388 Lumo Road, Wuhan, 430000, China [email protected]

Abstract. The development status of logistics in the financial crisis requires the shippers and carriers’ overall goal focus on cost reduction. This paper firstly analyzes the problem of information mismatch between shipper and carrier in nowadays, and describes the shippers and carriers’ demand for information platform. Then based on requirement investigation and questionnaire statistics, the specific demands for logistics information matching platform are analyzed. Finally, logistics information matching platform system for highway transportation is designed. Keywords: logistics information matching platform; highway transportation; system design.

1 Introduction With the development of modern logistics industry, the shortcoming in establishment and application of the logistics information system has turned out to be the "bottleneck" of logistics development in China, which directly influences the communication and cooperation between logistics enterprises and users, logistics enterprises and related government departments, and hampers the development of logistics service quality [1]. The shortcoming in establishment and application of the logistics information system are also acts as a serious impact on the competitiveness of China’s logistics enterprises [2]. With modern information technology and logistics information matching platform, the horizontal integration between logistics enterprises and manufacturing companies can be achieved, the regional logistics information resources can be shared, the configuration of social logistics resources can be maximum optimized, logistics cost can be reduced, and the whole process of logistics operation can be upgraded. Because the optimization of logistics supply chain needs the participation of business partners in various types (especially the supplier and demander of related logistics service, for example the shipper and carrier in highway transportation) and involves complex and various types of operations. And functions, such as information *

This paper is supported by Zhejiang Provincial University Students Scientific Innovative Project and Zhejiang Gongshang University Students Scientific Innovative Project.

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 135–141, 2011. © Springer-Verlag Berlin Heidelberg 2011

136

D. Chen et al.

release, exchange, and optimal matching between the relevant participants, which serve as a support logistics supply chain logistics service based on the Internet, can be easily accepted by the shipper and carrier. Therefore, making the best use of information of shipper and carrier, and constructing logistics information platform are of great significance in promoting the service level and efficiency of logistics industry. Based on the detailed analysis of the development status of logistics in the financial crisis with information demand analysis and questionnaire statistics, this paper analyze the problem of the mismatch of information between shipper and carrier in current domestic and their overall cost reduction goal and demand for information platform, and proposed a overall design framework for highway transportation logistics information matching platform. The organization of the paper is as follows. In section 2, the problem of the mismatch of information between shipper and carrier in highway transportation is described and analyzed, and the function demand of the logistics information matching platform for highway transportation is suggested. In section 3, the structure framework, system function and database of this information matching platform are designed. Section 4 accomplishes six main system functions according to its E-R diagram. The final section summarizes the major conclusions and suggests further application prospect of this system.

2 Analysis and Design of the Logistics Information Matching Platform's Function In order to integrate resources and realize the resource utilization, various forms of freight logistics platform appears, the main forms can be listed as: information website, information release, cargo distribution center of information. One of the outstanding performances is the "Bagualaiwang" logistics information website (http://www.8glw.com/) in Anyang city, Henan province, which is a carrier-shipper information providing website for free and can improve the vehicle real load rate. For its obvious energy saving effect, the system is classified by the Ministry of Transport as the second batch of "energy conservation and emission reduction demonstrative project throughout the national transportation industry"[3]. Another example is Huai’an logistics platform, zj56 (http://www.zj56. com.cn), and chinawutong (http://www. chinawutong. com) [4-6]. But these systems also have their shortcomings. Firstly, these systems only provide the collection function and release function of information without information processing and matching. As a result, the shipper and the carrier will be in trouble with looking for the right information. On the other hand, this system can not examine the provided information. Although there are registered memberships, but not being real name registered, anyone can register account and release relevant information, which makes it a doubt in the authenticity of information provided by "Bagalaiwang". For the purpose of information matching between shipper and carrier, we designed a questionnaire for a survey according to the market. In this survey, nearly one hundred questionnaires were distributed to logistics transport companies and 100 questionnaires were distributed to 100 manufacturing companies, and the response rate are 36% and 64% respectively (except 8 invalid questionnaires), .So as to ensure

The Design of Logistics Information Matching Platform for Highway Transportation

137

the quality and effectiveness of the questionnaire directly, 85 percents was proceeding by deep visits and sending out questionnaire face to face. The objects of the survey include professional carriers, the intermediary organizations of transportation, the transportation department of manufacturing enterprises, and large and medium-sized manufacturing enterprises. 2.1 Overall Demand These questionnaires separate the objects into the shipper and the carrier. Shipper investigation shows: 78% shippers were in trouble with too many goods to send; and 95% shippers are willing to share their information with the carriers. Carrier investigation shows: only 13% can find the freight demand information by Internet currently; and 92% of the carriers are willing to search freight demand information by Internet. The results show that the overall demand for shipper and carrier’s information by logistics information platform based on Internet is huge and urgent. 2.2 Functional Demand Functional demands of shipper and carrier have the same points: a) quickly searching the matching source, b) information release, c) member service, d) credit evaluation, e) online transactions, and f) intelligent distribution solutions. In addition, the shippers also want to gain functions of fast search and match enterprises’ true information source, shipping quote, preponderant routes and logistics chain coordinated guidance services.

3 The Design of Logistics Information Matching Platform 3.1 Structure Framework Design According to the basic demand of shipper and carrier, a B/S structure for modern logistics information matching management system (as showed in Fig. 1) is suggested, which integrates user browser, WEB server, MYSQL database, to achieves the purpose of nationwide logistics information registration, inquiry, optimized match, online transactions and feedback.

Fig. 1. B/S structure diagram

138

D. Chen et al.

The advantage of adopting this B/S structure is that all applications are in application servers, which can retain the superiorities of maintenance, management and cost reduction. Data updating in workstations can reduce the workload of system maintenance and modification, and is suitable for the Internet age's mobile applications. 3.2 System Function Design Logistics information matching platform should be excellent in compatibility, efficiency, accuracy, safety and stability, etc. It is an effective and efficient design method to adopt the modularized design to complete logistics information platform for matching and realizing the corresponding function, as Fig. 2 shows. Information release and query Information services module Basic function

Online transactions module

Online transactions Member credit management

Maintenance and management

Information matching platform

Information collection and maintenance

Maintenance and management Distribution route optimizing

Intelligent distribution

Information matching GIS GPS Vehicle tracking

Expand function

Enterprise logistics operation scheme Forecast decision supporting Logistics future prediction

Fig. 2. Information platform of matching function modules

3.3 Database Design The key of the database is the table structure design, and data modeling method which employ the E-R method. Thus the logistics information matching platform for highway transportation should mainly include the following four basic tables: z z z z

Car resource information tables, which are mainly used to record the basic information of carrier’s vehicle resource; Goods resource information tables, which are mainly used to record the basic information of shipper’s freight resource; Complain information tables, which can be divided into complain information of carrier’s vehicle and those of shipper’s freight; Feedback information tables, which are mainly used to record the feedback information of the platform.

Figure 3 is the system's E-R diagram, in which the user entity is involved in the processing of complain information of carrier’s vehicle, complain information of shipper’s freight, information of carrier’s vehicle resource, information of shipper’s freight resource and feedback information, and the information of carrier’s vehicle resource and those of shipper’s freight are correlate with each other by information matching.

The Design of Logistics Information Matching Platform for Highway Transportation

139

Fig. 3. Platform data relationship E-R chart

4 The Realization of the Function of System 4.1 Platform Login Interface To ensure the safety of platform, only inputting the correct username and password can the users be allowed to enter the system. "Register"--can be a new user to register, which is default to be an ordinary users. After registration, the user’s statue and its accessibility are in a locked position, only be used after unlocked by the super user or full administrator user of the system. 4.2 Information Input of Carrier’s Vehicle Resource and Shipper’s Freight Resource There is no difference between super users and ordinary users in this function. The owner's name, ownership, insurance, destination, start date, models, quantity, volume items of the carrier’s vehicle are required for information input, empty input would cause an error of incorrect information. The information input of shipper’s freight resource is in a similar setting. 4.3 Complain Information of Both Carrier’s Vehicle and Shipper’s Freight There is no difference between super users and ordinary users in this function as well. The carrier’s vehicle resource ID and reason are required; empty input would cause an error of incorrect information. The shipper’s freight resource complains information is in a similar setting.

140

D. Chen et al.

4.4 Feedback Information Submit There is no difference between super users and ordinary users in this function. The name and mailbox are required; the default username would be the login user, which can be changed. 4.5 Information Matching After information matching processing, the results would show up, if there is no suitable information matching between carrier’s vehicle and shipper’s freight. Otherwise, the system will pop up matching interface. The default matching method is a normal mode and scores are sorted by descend. The information of carrier’s vehicle resource or that of shipper’s freight resource has a highest score would be the optimal matching one, which can be inquired according to its ID. Fig.4 is an information matching result of carrier’s vehicle with shipper’s freight.

Fig. 4. Information matching result of carrier’s vehicle with shipper’s freight

4.6 User Management Function In this function, super user or full administrator user of the system can do operation such as add user, delete user, lock or unlock user, modify password and return. Add user operation can add super users and ordinary users. For the delete function, the super user can delete super user, ordinary users and also themselves (except full administrator user). The function of lock and unlock user depends on the current status of the object, i.e. super users can lock or unlock ordinary users and super users. All locked users can not be allowed to login the system.

5 Conclusion In the age of Internet economy, with the logistics information matching platform by Internet, the carrier and shipper can reduce their operative cost, reduce the empty rate of vehicle and improve the production capacity, to a certain extent avoid the resources waste in regional logistics operation, and can indirectly reduce traffic volume, release the traffic congestion, which is the key to improve operation efficiency.

The Design of Logistics Information Matching Platform for Highway Transportation

141

The information matching platform proposed in this paper has a strong faulttolerance, safety, convenient operation and the characteristics of stability and comprehensiveness, and is easy to expand and transplantation, which can enhance the cooperation between nodes enterprises in logistic chain and promote the development of logistics informalization, even would further promote e-business development in logistics industry. Although it is restrained from by the factors, such as fund, technology and enterprise reputation, etc., which make some insufficient information matching functions still exist in this platform, and remain to perfect further, we still believe it would have a significant application prospect.

References 1. 2.

3. 4. 5. 6.

Zhao, J.J., Yu, B.Q.: Modern logistics management. Peking University Press, Beijing (2004) Qin, W.Q., Wu, J.M.: New Access to Carries and Shipper Information: Logistics Information Real-time Intelligent Matchmaking Platform. Logistics Technology (210-211), 190– 193 (2010) Liu, C.: Analysis into the demands and Functions of Huai’An Logistics Information Platform. Logistics Engineering and Management 31(9), 30–31 (2009) Bagualaiwang logistics information, http://www.8glw.com Public Logistics Information Service Platform of Zhejiang Province, http://www.zj56.com.cn China Wutong Logistics Information Website, http://www.chinawutong.com

An Intelligent Prediction Method Based on Information Entropy Weighted Elman Neural Network* Tao Chen1, Xiao-li Xu1,2, and Shao-hong Wang1,2 1

Key Laboratory of Modern Measurement & Control Technology, Ministry of Education, Beijing Information Science & Technology University, Beijing, China 2 School of Mechanical Engineering, Beijing Institute of Technology, Beijing, China

Abstract. Neural network is an intelligent method in conditon trend prediction, while the condition trend prediction is important to guarantee the safe operation of large equipments. In order to overcome the deficiency of basically the same probability contribution of neural network input to output predicted, this paper proposes an intelligent prediction method based on information entropy weighted neural network, taking Elman neural network as the basis, combining with information entropy theory to construct the prediction model based on information entropy weighted Elman neural network. Condition trend prediction results of the flue gas turbine showed that the proposed new method has better prediction precision and real time performance. Keywords: intelligent, trend prediction, neural network, information entropy weighted.

1 Introduction Artificial neural network is an intelligent tool used to deal with complex nonlinear problems. Neural network has the characteristic of approximating the arbitrary continuous non-linear function and its all order derivatives with arbitrary precision by the appropriate selection of network layer and the cell number of hidden layers, thus being widely used in prediction during the industrial process. In fault prediction, the contribution of network input values to network output values has basically the same probability. In order to overcome the deficiency of basically the same probability contribution of neural network input to output predicted, this paper proposes a prediction method based on information entropy weighted Elman combing with information entropy theory and Elman neural network. *

National Natural Science Foundation of China (50975020) and Funding Project (PHR20090518) for Academic Human Resources Development in Institutions of Higher Learning under the Jurisdiction of Beijing Municipality.

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 142–147, 2011. © Springer-Verlag Berlin Heidelberg 2011

An Intelligent Prediction Method

2

143

Information Entropy Weighted Elman Neural Network

2.1 Calculation of Information Entropy Weight The proposed information entropy weighted neural network prediction method in this paper makes use of information entropy to indicate the contribution weight of network input to the network output predicted. Information entropy is used to measure the average amount of uncertainty of information source from the overall and objective perspective. The smaller the information entropy is, the more definite the information is[1-3]. Suppose one condition monitoring and fault prediction system has n sources of information (sensors), i.e. x1,x2,…,xn, the probabilities of the required information provided by the information sources are p1,p2,…,pn, the information structure of the system is:

⎛ X ⎞ ⎛ x1 S = ⎜⎜ ⎟⎟ = ⎜⎜ ⎝ P ⎠ ⎝ p1

x2 p2

x3 L xn ⎞ ⎟ p3 L pn ⎟⎠

(1)

Where pi is the information probability provided by information source, the calculation is as follows:

pi =

xi

(i = 1,2,L , n )

n

∑x i =1

(2)

i

In this system, the information entropy Ei which characterizes various information source, is calculated as follows:

Ei = − pi log 2 pi

(i = 1,2,L, n)

(3)

Entropy weighted neural network prediction method conducts information entropy weight of the raw information from sensors. The information entropy weighted coefficient wi is calculated as follows:

wi =

Ei n

∑ Ei

(i = 1,2, L , n)

(4)

i =1

The weight coefficient calculated based on formula (4) reflects the information amount carried by neural network input. 2.2 Entropy-Weighted Elman Neural Network Structure Elman neural network used in information entropy weighted Elman neural network is a typical feedback neural network, which has simple structure and has stronger computing power and stronger ability to adapt to time-varying characteristics than the

144

T. Chen, X.-l. Xu, and S.-h. Wang

forward neural network [4]. Different from RBF, BP and other forward neural network, besides the input layer, hidden layer and output layer, Elman network still adds a context layer in the hidden layer. The context layer acts as one-step delay operator, and is used to memorize the output value at the previous moment for hidden layer unit[5,6]. The structure of Elman neural network is shown in Figure1.

Fig. 1. Structure of Elman Neural Network

，

As shown in Figure1 k stands for moment, y, x, u ,xc respectively represents output of the network, the output of the hidden layer, external input of network, and output of context layer. w1,w2,w3 respectively stands for the connection weight matrixes from the context layer to hidden layer, from the input layer to hidden layer, from the hidden layer to the output layer respectively. b1 and b2 are the threshold values of input layer and hidden layer. Information entropy weighted Elman neural network, based on the Elman neural network, adds an information entropy weighted processing layer between the input layer and hidden layer. The structure of information entropy weighted Elman neural network is shown in Figure 2, the information entropy weighted processing layer identifies the contribution of the network input to the network output predicted, and can gain consistent description of running condition of mechanical equipments.

Fig. 2. Structure of Information Entropy Weighted Elman Neural Network

An Intelligent Prediction Method

145

Information entropy weighted Elman neural network prediction method is an intelligent dynamic prediction method. The method calculates information entropy weight of network input value, and determines the contribution amount of network input value to network output prediction value.

3 Condition Trend Prediction of Flue Gas Turbine Based on Information Entropy Weighted Elman Neural Network 3.1 Parameter Setting in Information Entropy Weighted Neural Network This paper takes the flue gas turbine in large-scale petrochemical company as the research object. The flue gas turbine is a kind of key devices in the catalytic cracking energy recovery system. In order to ensure the safe operation and scientific maintenance, conducting fault prediction will effectively avoid the contingency, save a great deal of maintenance fees and increase the equipment utilization rate. In accordance with the operating characteristics for the flue gas turbine, we collect the historical vibration data measured at YT-7702A bearing point of the turbine, and extract the vibration attribute value of 1 double frequency component to build 3-layer information entropy weighted Elman to make condition trend prediction. In the constructed network, we select the vibration attribute for every five days as the information input into the neural network, and select the vibration attribute at the sixth day as the output, that is, the number of neurons at the input layer is 5, and the number of neurons at the output layer is 1. The iterative prediction way is used which is formed from single-step prediction iteration. The transfer function from the network input layer to hidden layer adopts the hyperbolic tangent S- function. In order to effectively use the S-function, and to ensure that the nonlinear approximation ability of neural network, we conduct the normalization processing to transform the sample data to the (-0.9,0.9)interval.

xˆi =

1.8( xi − x min ) − 0.9 x max − x min

(5)

After training, the actual value on network output result can be obtained by the inverse transformation, that is, the anti-normalized processing should be made on the prediction results, then the actual prediction result will be obtained. LM algorithm is used in neural network training. As LM algorithm is not strongly dependent on the initial value, it can greatly improve the inherent flaws and shortcomings of BP network; it has the speed and precision of Newton method while not calculating Hessian matrix. As for the number and accuracy of training times, LM algorithm is obviously superior to conjugate gradient method and BP algorithm with variable learning rate [7,8]. The optimal number of nodes in the hidden layer is determined by trial-and-error method. Table 1 shows the pediction performance of different numbers of neurons in hidden layer in the constructed neural network.

146

T. Chen, X.-l. Xu, and S.-h. Wang Table 1. Prediction performance of different numbers of neurons in hidden layer

Number of neurons in the hidden layer 9 10 11 12 13

training steps 50 40 36 34 34

MAPE

MSE

6.4354 × 10 −4 6.5183 × 10 −4 5.8701 × 10 −4 6.2309 × 10 −4 6.4452 × 10 −4

2.9346 × 10 −5 2.9587 × 10 −5 2.9134 × 10 −6 2.9294 × 10 −5 2.9629 × 10 −5

As can be seen from Table 1, when the number of neurons in hidden layer is 11, the prediction errot of MAPE and MSE are minimum, and the number of training steps is intermediate; after the comprehensive assessment of training steps and errors, and prior consideration of the error performance, the optimal number of nodes in the hidden layer in the neural network is finally determined to be 11. 3.2 Prediction Results and Analysis The information entropy weighted Elman neural network is used to predict the vibration attribute value of flue-gas turbine, and the comparison with Elman neural network is made. The prediction results of informaton entropy weighted Elman and Elman are shown in Figure 3. The following table 2 shows the prediction performance of information entropy weighted Elman and Elman prediction methods.

Fig. 3. Prediction results of Information Entropy Weighted Elman and Elman Table 2. Prediction Performance of Information Entropy Weighted Elman and Elman

Types of Neural network Elman Information Entropy Weighted Elman

Iterations Times 29 26

MSE

MAPE

4.5036 × 10 −4 3.1890 × 10 −4

3.9803 × 10 −3 2.8706 × 10 −3

An Intelligent Prediction Method

147

As shown in Figure 3 and Table 2, information entropy weighted Elman is superior to Elman in terms of the iteration times and the prediction error. The information entropy weighted Elman has stronger approximation to the actual value, and improves the prediction accuracy and real-time performace, which fully illustrates the prediction performance of the proposed prediction method.

4 Conclusion Information entropy weighted neural network is an intelligent prediction mthod combining with information entropy theory. It can overcome the deficiency of basically the same probability contribution of neural network input to output predicted. The prediction result shows that the Information entropy weighted Elman neural network has higher prediction accuracy, better prediction real time performance. The analysis indicates that the proposed method is feasible in the condition trend prediction for large equipments with a broad application prospect.

Acknowledgments The authors appreciate the comments of the anonymous reviewers. Thanks to the scholars listed in the references, for their wisdom and creative achievement giving us inspiration.

References 1. 2. 3. 4. 5. 6.

7.

8.

Barnum, H., Barrett, J., Clark, L.O., et al.: Entropy and information causality in general probabilistic theories. New Journal of Physics 3, 1–32 (2010) Zhang, Q.-R.: Information conservation, entropy increase and statistical irreversibility for an isolated system. Physica A (388), 4041–4044 (2009) Zhang, J., Mi, X.: Neural Network and Its Application in Engineering. China Machine Press, Bejing (1996) Elman, J.L.: Finding Structure in Time. Cognitive Sci. (14), 179–211 (1990) Raja, S., Toqeer, N., Suha, B.: Speed Estimation of an Induction Motor using Elman Neural Network. Neurocomputing (55), 727–730 (2003) Ciarlini, P., Maniscalco, U.: Wavelets and Elman Neural Networks for Monitoring Environmental Variables. Journal of Computational and Applied Mathematics (221), 302–309 (2008) Arab, C.M., Beglari, M., Bagherian, G.: Prediction of Cytotoxicity Data (CC50) of AntiHIV 5-pheny-l-phenylamino-1H-imidazole Derivatives by Artificial Neural Network Trained with Levenberg–Marquardt Algorithm. Journal of Molecular Graphics and Modelling (26), 360–367 (2007) Bahram, G.K., Susan, S.S., Troy, N.H.: Performance of the Levenberg–Marquardt Neural Network Training Method in Electronic Nose Applications. Sensors and Actuators B (110), 13–22 (2005)

A Multi-layer Dynamic Model for Coordination Based Group Decision Making in Water Resource Allocation and Scheduling Wei Huang1,2, Xingnan Zhang1,2, Chenming Li3, and Jianying Wang3 1

National Engineering Research Center of Water Resources Efficient Utilization and Engineering Safety Hohai University, Nanjing, P.R. China 2 College of Hydrology and Water Resources Hohai University, Nanjing, P.R. China 3 College of Computer and Information Engineering Hohai University, Nanjing, P.R. China [email protected], [email protected]

Abstract. Management of group decision-making is an important issue in water source management development. In order to overcome the defects in lacking of effective communication and cooperation in the existing decision-making models, this paper proposes a multi-layer dynamic model for coordination in water resource allocation and scheduling based group decision making. By introducing the scheme-recognized cooperative satisfaction index and scheme-adjusted rationality index, the proposed model can solve the problem of poor convergence of multi-round decision-making process in water resource allocation and scheduling. Furthermore, the problem about coordination of limited resources-based group decision-making process can be solved based on the effectiveness of distance-based group of conflict resolution. The simulation results show that the proposed model has better convergence than the existing models. Keywords: Group Decision-Making, Optimization, System Simulation, Water Resource Allocation and Scheduling.

1 Introduction The water management institutional reform in China is in the direction of basin water resource unified management and region water affair integration management, and requires building up cooperative mechanism of parts participant, democratic consultation, common decision, individual responsibility and efficient execution mechanism. However, the current water allocation and management and decision support system almost stay on the individual decision-making layer, the practical requirements of water resources regulation should transform individual decision-making mode to group decision-making mode. Many researchers such as the references [1-5] expatiate the selection of water resource decision-making scheme is not a individual decision-making problem but a R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 148–153, 2011. © Springer-Verlag Berlin Heidelberg 2011

A Multi-layer Dynamic Model for Coordination Based Group Decision Making

149

group decision-making problem, which mainly pay attention to water resource field such as water resource bearing capacity, instead of group decision-making problem of how to cooperate efficiently. According to the characteristic of water resource configuration, this paper proposes a multi-layer dynamic model of coordination based group decision making for water resource allocation and scheduling. This model is a multi-objection, multi-layer, multi-time interval, multi-restriction and multi-round decision-making process on the basis of cooperation. In order to solve the confliction problem of group scheme in the model, this paper proposes a conflict resolution scheme based on group utility distance optimization. At last, this model is preliminary validated on the Swarm simulation platform.

2 The Multi-layer Dynamic Model The essence of water resource transfer group decision-making is group opinions progressively converge into consistent for limited resource conflicted decision-making. The cooperative multi-layer dynamic group decision-making model needs introducing scheme-recognized cooperative satisfaction degree index and scheme-adjusted rationality index to impel multi-round convergence. The algorithm of this model is described as follows. Step1k According to the (K-1)th layer’s guidance restriction and the Kth layer’s

macroscopical objective concept, each decision maker’s scheme advices and so on, the cooperative group organizes layer K expert group to build up the Kth layer’s objective system, feasible scheme group and the Kth layer’s constraint conditions of each scheme. Stepk2 Adjustable threshold interval θ min ,θ ave of group satisfaction degree scheme recognized is set. Cooperators organize the round(i+1) negotiation. Stepk3 With the organization of cooperative group, each decision makers group begins to the round i scheme negotiation. If group satisfactory degree exceeds the threshold θ ave , then turn to Stepk5 , otherwise, turn to Stepk4 . Stepk4 Cooperative group organizes experts to analyze conflict and adjust objectives,

schemes and constraints. If the groups’ adjusting intension for objectives, schemes and constraints is unreasonable, then the threshold of satisfactory degree is adjusted. If it is no less than θ min , then turn to Stepk5 . Otherwise, turn to Stepk6 . Stepk5 According to the threshold of group’s satisfactory degree, satisfied schemes

are sorted in the group satisfied solution set, the optimal scheme is found, the lower decision-making’s constraint set is arranged, the approbatory protocol of decision-making groups’ scheme for this layer is impelled, and penalized cost μ protocol violated is promised. Then the algorithm comes into the layer(K+1), when this layer ends. Stepk6 The cooperative group of the Kth layer organizes experts to compute cost of choice scheme. If the whole utility of the scheme is lower than μ , then the scheme

150

W. Huang et al.

and constraints opinions for the upper (K-1)th layer come into being, which is fed back to (K-1)th layer’s cooperative group, otherwise, objectives, schemes and constraints should be adjusted, and bonus-penalty factor should be added, then turn to Stepk2 .

3 Evolved Cooperative Mechanism Based on Distance Optimization of Group Utility The scheme attribute sequencing globally observed by player is assumed as A = (a1 , a2 ,..., am ) , the attribute values is vai0 ∈ [min ai0 , max ai0 ] under global constraints, and the attribute values is vai0 ∈ [min ai0 , max ai0 ] under objective constraints of decision maker j. As well, the unit effectiveness concession on attribute ai is ΔU l* (ai ) . Definition 3-1: Normalized utility distance is boundary distance U t j (ai ) , denoted by U t j (ai ) =| vai0 − vaij | U t j (ai ) =|| min ai0 − max ai0 | / 2− | min aij − max aij | / 2 | .

(1)

where U t j (ai ) is the distance between center points of two normalized utility region, where t expresses the tth round adjusting. Definition 3-2: Group adjusting inclination is described as index vector distance optimization after group adjusting, and is denoted as B( ai ) N

N

j =1

j =1

B( ai ) =| ∑U t j ( ai ) | − | ∑U t j−1 ( ai ) |, i = 1..M .

(2)

Definition 3-3: Attribute value intersection function f 0 j ⎪⎧1 vai ∩ vai ≠ Φ f ( i, j ) = ⎨ . 0 j ⎪⎩0 vai ∩ vai = Φ

(3)

① Cooperators’ cooperative strategy

Rule 1: Sort attributes by the values of B( ai ) from big to small, abstract adjustable N

attributes from the sequence, if

∑ f (i, j ) > α

( α can choose most rules, such as N/2),

j =1

N

then find out the next attribute from the sequence, until the attribute of

∑ f (i, j ) > α j =1

can not be satisfied, which is taken as global optimized adjusting attribute, and do ΔU l* (ai ) adjusting. Rule 2: If the decision maker doesn’t do any adjusting, then give the penalty equals to K ΔU l* (ai ) , where K is more than 1.

A Multi-layer Dynamic Model for Coordination Based Group Decision Making

151

Rule 3: If the decision maker does the optimized adjusting, the judgment is the effects of the sharpen degree for current conflict inclined to smooth, i.e., B( ai ) has inclined trend, then preference compensated encouragement can be given. Rule 4: If the decision maker’s adjusting is not the optimized one, then no rewards and punishment is given.

② Decision makers’ cooperative strategy

Rule 1: Observe global optimized adjusting attribute, if property with its own strength in the tolerance range of adjustment of preferences. The preference sequence is adjusting according to global optimized adjusting attribute. Rule 2: If the utility is incompatible with attribute adjusting preference, i.e., loss utility is too much, then by observed other conflicted attributes, the most adjacent personal preference structural attributes is chosen, and yielded, which can help to bring about new optimized group inclination. Because group preference opinions embody in cooperators’ adjusting, decision makers’ game is implicated in decision makers and cooperators.

4 Simulations This paper does simulated experiment on Swarm simulated platform making use of parts statistic data published in references [6-9], which simulates group decision-making composed of three decision makers and a cooperator. (1) Satisfied concentration in preference independent Taking the initial configuration set in reference [10], the assumption that decision makers prefer the independence of decision-makers concerned only with the coordinator of all the constraints, the satisfaction depends on the coordination tendency of those. Simulated results are shown as Figure 1. From the Figure 1 we can see that if the cooperators consider mandatory step, i.e., prescriptive plan, the random distribution of satisfactory degree can hardly converge. Reference [10] configuration can cause high satisfactory degree upstream, but low satisfactory degree downstream, and the average satisfactory degree is not high. From the algorithm design idea, under independent assumption and mandatory strategy, the reference standard of decision makers is unique, without considering later effect. As well, each decision maker’s strategy is irrelevant, while the conditions are independent, which can not be interacting, thus, the random characteristic of satisfactory degree shows the rationality of algorithm function designing and implementation. (2) Satisfied concentration in harmony Taking the initial configuration set in [10], the decision maker and the cooperator are assumed to adopt correlated preference, that is, other participants’ preference are considered, and the personal strategy is modulated according to group preference inclination, and the simulated results are shown in Figure 2.

152

W. Huang et al.

Fig. 1. Satisfied concentration in preference independent

Fig. 2. Satisfied concentration in harmony

If the different price of water used in peak period and valley period is introducing, in the situation of complementary configuration in the time of water used, satisfactory degree in long term is high, and increases year by year, while the velocity of convergence is fast, which are shown in Figure 2. Meanwhile, the results accord with the long term benefit configuration of preference in the algorithm design.

5 Conclusion This paper proposes a group decision-making method which is MCGD (Multi-layer Cooperative Dynamic Group Decision) corresponding to the need of water resource allocation and scheduling, which combines multi-objective group decision-making, multi-layer group decision-making, multi-period group decision-making, multi-attribute group decision-making, multi-round group decision-making and so on. The characteristic of cooperative multi-layer dynamic group decision-making is adopting group decisions by the greatest extent, the results satisfying multi-part by cooperation, instead of non-completed compromised equivalent solution. By cooperation and compromise, the decision makers are impelled to develop avoiding conflicts, thus, the integral optimized solution is got on the condition of satisfactory to every party. This decision-making mode corresponds to the dynamic configuration of limited water resource. Acknowledgment. This work is supported by the National Nature Science Foundation of China (No.50479018), and the Fundamental Research Funds for the Central Universities of China (No. 2009B20414).

References 1. Wang, H., Qin, D.Y., Wang, J.H.: Concept of system and methodology for river basin water resources programming. Shuili Xuebao 2002(8), 1–6 (2002) (in Chinese) 2. Chen, S.Y., Huang, X.C., Li, D.F.: A multi-objective and group-decision-making model of water resources and macro-economic sustainable development in Dalian. Shuili Xuebao 2003(3), 42–48 (2003) (in Chinese)

A Multi-layer Dynamic Model for Coordination Based Group Decision Making

153

3. Xu, Z.M.: A Scenario-Based Framework for Multicriteria Decision Analysis in Water Carrying Capacity. Journal of Glaciolgy and Geocryology 21(2), 99–106 (1999) 4. Cheng, G.D.: Evolution of the Concept of Carrying Capacity and the Analysis Framework of Water Resources Carrying Capacity in Northwest of China. Journal of Glaciolgy and Geocryology 24(4), 361–366 (2002) 5. Hajkowicz, S.: Cutting the cake: Supporting environmental fund allocation decisions. Journal of Environmental Management 90(8), 2737–2745 (2009) 6. China National Bureau of Statistics Of. China City Statistical Yearbook. China Statistics Press (2007) (in Chinese) 7. China National Bureau of Statistics Of. Anhui Statistical Yearbook 2007 [ISO]. China Statistics Press (2007) (in Chinese) 8. China China National Bureau of Statistics Of. Jiangsu Statistical Yearbook 2007 [ISO]. China Statistics Press (2007) (in Chinese) 9. China National Bureau of Statistics Of. Shandong Statistical Yearbook 2007 [ISO]. China Statistics Press (2007) (in Chinese) 10. Wang, S.B.: Study on Rational Water Resources Allocation Oriented to User’s Demand. PhD thesis, Xi’an University of Technology (2007) (in Chinese)

About the Authors Wei Huang is a PhD student in National Engineering Research Center of Water Resources Efficient Utilization and Engineering Safety, and College of Hydrology and Water Resources, Hohai University, Nanjing, P.R. China. He received his MSc in Water Science and Engineering Department of UNESCO-IHE in the Netherlands in 2007. His current research interests include hydroinformatics, water resources system modeling and environmental assessment. Xingnan Zhang is a professor in National Engineering Research Center of Water Resources Efficient Utilization and Engineering Safety, and College of Hydrology and Water Resources, Hohai University, Nanjing, P.R. China. His current research areas include water science and engineering, hydroinformatics. Chenming Li is an associate professor in College of Computer and Information Engineering, Hohai University, Nanjing, P.R. China. He is a senior member of China Computer Federation, and Chinese Institute of Electronic, his current research interests include information processing system and its applications, complex system modelling and simulation. Jianying Wang received his PhD from Hohai University, P.R.China in 2008. He is an associate professor in College of Computer and Information Engineering, Hohai University.

Analysis of Mode Choice Performance among Heterogeneous Tourists to Expo Shanghai 2010 Shengchuan Jiang, Yuchuan Du*, and Lijun Sun Key Laboratory of Road and Traffic Engineering of the Ministry of Education, Tongji University, Shanghai, China [email protected]

Abstract. Tourist is a conventional kind of commuters in the urban transport system. During some mega-events, such as Olympic Games or Expos, tourists would become the most important and sizable part of commuters in the host city. Expo 2010 Shanghai will attract 70 millions tourists during the whole 184 days duration. The large number of tourists expected to be carried, combined with the congested urban road network and limited parking spaces, will make it difficult for individual transport to be used during the Expo; as such, high rates of utilization of public transport will be necessary. Hence, exploring the trip mode choice behavior of Expo tourists is the keystone for traffic planning of Expo 2010 Shanghai, especially for the difference of heterogeneous tourists from various departure areas. A joint model system, in the form of associating clustering analysis and HL model, is developed in this paper to investigate the differences of trip mode choice behaviour among heterogeneous tourist groups. The clustering analysis method is used in this paper to distinguish the various types of Expo tourists for the choice sets are variable with the attributes of groups. Sorted by the attribute of departure area, tourists of Expo Shanghai could be divided into three kinds: local visitors, out-of-town one-day-trip visitor and out-of-town lodging visitors. The statistic parameters of three kind models constructed by this joint system show that this modeling method improves analytic precision effectively. Keywords: Expo Tourists, choice behaviour, clustering analysis.

1 Introduction Tourist is a conventional kind of commuters in the urban transport system. During some mega-events, such as Olympic Games or Expos, tourists would become the most important and sizable part of commuters in the host city. It is reported that Expo 2010 Shanghai will attract 70 millions tourists during the whole 184 days duration, namely 400,000 tourists daily and 800,000 tourists in peak day (Yin et al., 2007). Considering the location of the Expo Park is in the city center of Shanghai, this mega-event is regarded by many experts as one of the great, world-wide transport and logistics challenges. The large number of tourists expected to be carried, combined with the *

Corresponding author.

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 154–164, 2011. © Springer-Verlag Berlin Heidelberg 2011

Analysis of Mode Choice Performance among Heterogeneous Tourists

155

congested urban road network and limited parking spaces, will make it difficult for individual transport to be used during the Expo; as such, high rates of utilization of public transport will be necessary. Hence, exploring the trip mode choice behavior of Expo tourists is the keystone for traffic planning of Expo 2010 Shanghai, especially for the difference of heterogeneous tourists from various departure areas. Over the past few decades, research interest in the link between travel choice behaviour and the contribution of various independent variables has blossomed. More recently, discrete choice models based on SP methods have become popular among academics, governments and consulting companies to explore many aspects of transportation, including mode choice behaviour, urban forms, levels of service, prices and so on (Fowkey and Preston 1991; Bhat 1997, 2005; Raney etl. 2000; Punj and Brookes 2001; Johansson et al. 2006; McMillan 2007; Lu and Peeta 2009). Tourists differ from daily commuters in several ways that suggest different analysis method may be necessary. First, tourists usually have no acquaintance with the urban transport system in the visited city, they may prefer direct and comfortable trip mode. Second, because of the budget difference in the aspect of tourist time and expense, the variability of tourists’ choice behaviours may be greater than urban commuters. Third, feature of tourist’s origins and destinations are more complicated than daily commuters, therefore tourist-oriented management strategies must be flexible in location as well as scale, to account for spatially shifting demand. There are many strategies have been proposed for distinguishing among groups of travellers, including ones based on attribute cutoffs, clusters of travel attitudes, motivations or preferences, behavioural repertoires for different activities and hierarchical information integration (Swait, 2001; Johansson et al., 2006; Steg, 2005; Van Exel et al., 2005; Anable and Gatersleben, 2005; Bamberg and Schmidt, 2001; Tam et al., 2008; Molin and Timmermans, 2009). This paper presents a joint model system to investigate the differences of trip mode choice behaviour among tourist groups. The model system takes the form of a joint clustering analysis and hierarchical logit (HL) model. The remainder of this paper is organised as follows. The next section describes the clustering analysis method is used to distinguish the various types of tourists for the choice sets are variable with the attributes of groups. This is followed by the structure of HL model is characterised by grouping all subsets of tourist-correlated options in hierarchies. Each nest of the HL model is represented by certain tourist features which are differ from the others. Then, the joint model system is used to estimate trip shares on a survey sample of potential tourists for Expo 2010 Shanghai, which have been conducted among tourists in an airport, a train station and highway service stations in Shanghai. The last section includes the concluding comments.

2 Cluster Analysis for Potential Expo Tourists The level of transport demand in Shanghai during Expo 2010 will be very high because of the great numbers of local and out-of-town visitors, including those from the Yangtze Delta, other regions of China including Hong Kong, Macao and Taiwan and foreign countries. The registration report of the Expo 2010 Shanghai Committee

156

S. Jiang, Y. Du, and L. Sun

indicates that Shanghai visitors will account for 25% of total visitors; visitors from the Yangtze Delta, 40%; visitors from other regions of China, 30%; and visitors from overseas, 5%, as shown in Figure 1.

Yangtze Delta, 40%

Overseas, 5%

Other Regions Domestic, 30%

Shanghai Local, 25%

Fig. 1. Distribution of Expo Tourists

Because of the range of tourists to Expo 2010 Shanghai and their various attributes, it will obviously be difficult to obtain satisfactory results putting all Expo tourists into a single group for analysis and modelling. In addition, as the World Expo has not previously been held in China, there is no reference to aid in the understanding or prediction of tourist trip mode choice behaviour over the duration of this mega-event. Given such a backdrop, this paper developed a two-stage gradual stated preference survey

Research into trip mode choice and design of stated preference survey

Trip mode choice behaviour without constraints 

Stage 1

Data collection and homogeneity analysis

Cluster analysis and determine different groups

Survey of multi-scenario-comparison choice

Stage 2

Chose the structure of NL model by certain tourist features

Data collection and analysis

Parameter estimation and testing

Fig. 2. Procedure of the Two-Stage Gradual Stated Preference Survey

Analysis of Mode Choice Performance among Heterogeneous Tourists

157

method for the in-depth study of Expo tourist trip mode choice behaviour. The method of procedure is presented in Figure 2. Based on the Stage 1 survey, the cluster analysis is utilized to classify the potential Expo tourists. Cluster analysis groups data objects based only on information found in the data that describes the objects and their relationship. The aim of cluster analysis is to find out the objects within a group be similar to one another and different from the objects in other groups. The procedure of Cluster Analysis is: Step 1 is transforming the variables to standard scores. In clustering, the measure of original date often affects the comparative and operation; hence, transforming the variables to standard scores is necessary. There are n subjects, and each subject has p variables. Cluster analysis begins with a n*p original matrix as followed. ⎡ x11 ⎢x 21 X=⎢ ⎢ M ⎢ ⎣⎢ xn1

x12 L x1 p ⎤ x22 L x2 p ⎥⎥ M M M ⎥ ⎥ xn 2 L xnp ⎦⎥

(1)

where Xij (i = 1, L , n; j = 1, L , p ) is the value of ith subject’s jth attribute. Here we use Z scores to transform the variables to standard scores. Z scores are sometimes called "standard scores". The z score transformation is especially useful when seeking to compare the relative standings of items from distributions with different means or different standard deviations. It can be expressed as:

Xij' = where

Xij' is standard score;

Xij − X j

Xj =

dard deviation of jth attribute: S j =

Sj

(i = 1, L, n; j = 1, L, p)

1 n ∑ Xij is the mean of jth attribute; n i =1

(2)

S j is the stan-

1 n ∑ (Xij − X j )2 . n − 1 i=1

Step 2 is constructing a distance matrix using Euclidean distances. If ith subject and kth subject are two points in Euclidean p-space, then the distance from i to k is given by:

dik =

p

∑ (x j =1

ij

− xkj ) 2

where xij and xkj are the jth variable of ith subject and kth subject. Then the distance matrix is: ⎡ d11 d12 L d1n ⎤ ⎢d d 22 L d 2 n ⎥ ⎥ D = (dij ) = ⎢ 21 ⎢ M M M M ⎥ ⎢ ⎥ ⎣ d n1 d n 2 L d nn ⎦ where

dii =0 and d ij = d ji

，

(3)

(4)

158

S. Jiang, Y. Du, and L. Sun

Step 3 is choosing a certain clustering algorithms, whose subjects are sorted into significantly different groups where the subjects within each group are as homogeneous as possible, and the groups are as different from one another as possible. There are several different types of clusters that prove useful in practice, just like agglomerative hierarchical cluster analysis, concept cluster analysis, k-means cluster analysis, fuzzy cluster analysis. The K-means cluster analysis, which is suitable in large samples, is chosen in this paper. The k-means algorithm assigns each point to the cluster whose center is nearest. The center is the average of all the points in the cluster — that is, its coordinates are the arithmetic mean for each dimension separately over all the points in the cluster. The main advantages of this algorithm are its simplicity and speed which allows it to run on large datasets. According to data analysis of Stage 1 survey, we find that trip mode shares vary a lot in distinct groups. Therefore, this paper uses the first choice of trip without constraints, travel time and travel costs as characteristic variables to assess. There are 3 types of groups after 10 iterations. The ratios of tourists from different departure area in each type are shown in Table 1. The analysis results of variance are shown in Table 2. Table 1. Cluster Analysis Results

Local visitors Out-of-town a-day-trip visitors Out-of-town lodging visitors Type I Type II Type III

76.82%

4.71%

18.46%

14.38% 17.51%

11.09% 69.59%

74.53% 12.90%

Table 2. ANOVA

Cluster

Error F

Sig.

1185

742.731

.000

.833

1185

582.508

.000

.775

1185

1290.372

.000

Mean Square

df

Mean Square

df

Time

452.784

2

.610

Costs

484.976

2

1000.144

2

First Choice

Analysis of Mode Choice Performance among Heterogeneous Tourists

159

These results reveal that 3 variables of each type have statistical significance (p20). 2) Relative Complex Degree (RCD). This index is a value describing the probability of modules involved in the adjustment among all the modules. This index can be classified into four levels: A(0.07~1), B(0.05~0.07), C(0.03~0.05), D(0.01~0.03). 3) Personal numbers involved in the adjustment (PNI). This index means the number of people involved in the adjustment. This index can be classified into four levels: A(5~15), B(15~25), C(25~35),D(>35). (4) Risk: this index set includes the following sub-indices. 1) The meeting degree (MD) This index can be classified into four levels: A(completely meet), B(Basically meet), C(Partially meet),D(unable to meet). 2) The Influence degree (ID) This index can be classified into four levels: A(no influencing), B(Partially influencing), C(largely influencing), D(terribly influencing). 3) The confusion degree (CD) It is a qualitative index, and can be quantified according to the order degree with the pre-and post adjustment. In this paper, this index is classified into four levels: A(no confusion), B(Partially confusion), C(largely confusion), D(terribly confusion).

270

J. Liu, C. Xue, and L. Dong

3 Evaluation Model In this paper, based the similarity to ideal solution, the modeling process is shown in Fig.2, which is illustrated in detail below via the indices proposed in this paper.

Fig. 2. Modeling process of EISA

Step 1: according the Part 2, the grading standards for evaluation are as follows: 1) As for Time, the four levels A, B, C and D are 7, 5, 3, and 1 respectively. 2) As for Cost, the four levels A, B, C and D are 7, 5, 3, and 1 respectively. 3) As for absolute complex degree, the four levels A, B, C and D are 1, 3, 5, and 7 respectively. 4) As for relative complex degree, the four levels A, B, C and D are 7, 5, 3, and 1 respectively. 5) As for personal numbers involved in the adjustment, the four levels A, B, C and D are 7, 5, 3, and 1 respectively. 6) As for the meeting degree involved in the adjustment, the four levels A, B, C and D are 7, 5, 3, and 1 respectively. 7) As for the influence degree involved in the adjustment, the four levels A, B, C and D are 7, 5, 3, and 1 respectively. 8) As for the confusion degree involved in the adjustment, the four levels A, B, C and D are 7, 5, 3, and 1 respectively. To be mentioned, if the degree is between two standards, the grades are 2,4,6 respectively. Step 2: make sure the weight of the indices. As there are different weights among different indices, it’s necessary to give them different weights according to the Delphi method. As for the indices proposed in this paper, the weights of them are as follows: w = ( w1 , w2 , w3 , w4 , w5 , w6 , w7 , w8 ) = ( 0.1, 0.1, 0.05, 0.05, 0.1, 0.4, 0.1, 0.1) . where

w1 , w2 , w3 , w4 , w5 , w6 , w7 , w8 represent Time, Cost, absolute complex degree, relative complex degree, personal numbers involved in the adjustment, the meeting degree, the influence degree, the influence degree, and the confusion degree respectively. Step 3: make sure the sample matrix SM . In this step, according to the grading standards of the evaluation in step 1, each index is graded by experts, constituting SM ', then multiplying the weight of each index, and SM is obtained. Step 4: comprehensive evaluation of the sample matrix. Firstly, obtain the ideal matrix SM . Then calculate the distance d (1, 2, L , n ) between each EIS and the I

corresponding ideal matrix. And in this paper, the Frobenius matrix norm is adopted.

The Adaptability Evaluation of Enterprise Information Systems

271

According to the above steps, the final adaptability value is obtained, and the larger the final value is, the better adaptability is.

4 Examples The evaluation indices set and model are exemplified via the following case. The related data are shown in Table 1. Table 1. The index value

EIS

Indices Complexity T

C

ACD

RCD

0.1

0.1

0.05

0.05

1

5

4

3

2

4

1

3

3

3

Risk PNI

MD

ID

CD

0.1

0.4

0.1

0.1

4

1

5

1

2

2

2

5

7

3

4

5

1

2

6

1

6

According to the above process, the sample matrix is

⎡5 4 3 4 1 5 1 SM ' = ⎢ 4 1 2 2 5 7 3 ⎢ ⎣⎢3 3 5 1 2 6 1

2⎤ 4⎥

⎥

6 ⎦⎥

w = ( w1 , w2 , w3 , w4 , w5 , w6 , w7 , w8 ) = ( 0.1, 0.1, 0.05, 0.05, 0.1, 0.4, 0.1, 0.1) .

⎡ 0.5 0.4 0.15 0.20 0.1 2 0.1 0.2 ⎤ ⇒ SM = ⎢ 0.4 0.1 0.10 0.10 0.5 2.8 0.3 0.4 ⎥ ⎢ ⎥ ⎢⎣ 0.3 0.3 0.25 0.05 0.2 2.4 0.1 0.6 ⎥⎦ As the final value is the larger the better, the ideal matrix of this three EIS is

⎡0.5 0.4 0.25 0.20 0.5 2.8 0.3 0.6 ⎤ SM = ⎢0.5 0.4 0.25 0.20 0.5 2.8 0.3 0.6 ⎥ ⎢ ⎥ ⎣⎢0.5 0.4 0.25 0.20 0.5 2.8 0.3 0.6 ⎥⎦ I

Then the distance between each EIS and the ideal matrix d i ( i = 1, 2, 3 ) is calculated via Matlab 6.5, and the results are shown as follows: d1 = 1.005, d 2 = 0.4153, d 3 = 0.6021.

272

J. Liu, C. Xue, and L. Dong

As d 2 < d 3 < d1 , so the second EISA is the best, and the first EISA is the worst. What’s more, from the data in the table, the meeting degree in EIS2 is the highest. In other words, although the other indices may be better, but the meeting degree is the most important index, or the investment, such as time, money and so on will have no meaning.

5 Conclusions According to GQM, this paper proposes a set of adaptability index system, which includes five aspects: Time, Cost, Complexity, and Risk. Then the evaluation model is proposed to evaluate EISA and exemplified though a case. The research in this paper can rich the EISA theory and lay a foundation for EISA optimization. Acknowledgments. This work is supported by National Natural Science Foundation of China Grant #70971119, #70901066 and Aviation Science Foundation of China Grant #2008ZG55020.

References 1.

2. 3.

4.

5.

Meglich-Sespico, P.: Exploring the key factors for justifyingthe adoption of a human resource information system. In: Proceedings – Annual Meeting of the Decision Sciences Institute, pp. 535–540 (2003) Liu, W.-g.: An Evaluation Model and Its Implementation of Information System. Computer Applications 23, 33–35 (2003) Naing, T., Zainuddin, Y., Zailani, S.: Determinants of information system adoptions in private hostitals in Malaysia. In: 2008 3rd International Conference on Information and Communication Technologies: From Theory to Applications, pp. 1–2 (2008) Yu, C., Ma, Q., Ma, X.-X., Lv, J.: An Architecture-oriented Mechanism for Selfadaptation of Software Systems. Journal of Nanjing University (Natrual Sciences) 42, 120–130 (2006) Pan, J., Zhou, Y., Luo, B., Ma, X.-X., Lv, J.: An Ontology-based Software Self-adaption Mechnism. Computer Science 34, 264–269 (2007)

Structural Damage Alarm Utilizing Modified Back-Propagation Neural Networks Xiaoma Dong School of Civil Engineering & Architecture Zhengzhou Institute of Aeronautical Industry Management Zhengzhou 450015, China [email protected]

Abstract. Damage alarm is an important step among structure damage identification. Its objective is to evaluate the structure health. The existing damage alarm methods are mostly based on Back-Propagation Neural Networks without thinking over testing noise. Therefore, in order to avoid the disadvantages of conventional Back-Propagation Neural Networks, a modified Back-Propagation Neural Networks was proposed for structure damage alarm system in this paper. The experiment results of steel truss girder bridge show that the improved method is better than BPNN for structural damage alarm. Keywords: damage alarm, modal frequency ratio, steel truss girder bridge, modified BPNN.

1 Introduction The casualty of large engineering structure are often arose by some minuteness fatigue cracks, so it becomes an investigative hotspot to utilize some efficiency undamaged methods to inspect the structure damage beforehand[1~8]. In order to avoid compound factor identification and predigest the complexity of identification, a multilevel damage identification strategy was proposed. The strategy was to dispart the integrity process of damage identification into three steps. The first step was damage alarm, the second step was damage location, and the third step was damage degree identification. Damage alarm is an important step among structure damage identification. Its objective is to evaluate the structure health and give an alarm signal. By the vend literatures[1~5], in the structure damage identification field the investigator mostly fasten on the research of damage location and damage degree identification. And that damage alarm research is less attended due to its easy realization. By the vend literatures[6~8], the existing damage alarm methods were mostly based on conventional BPNN, and these methods didesn’t think over testing noise. Moreover there were testing noise in true testing signal. Therefore, in order to avoid the disadvantages of conventional BPNN, a modified BP neural network was proposed for structure damage alarm system in this paper. The experiment results of steel truss girder bridge show that the new method is better than BPNN for structural damage alarm. R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 273–278, 2011. © Springer-Verlag Berlin Heidelberg 2011

274

X. Dong

2 Modified BPNN Theory Artificial neural networks provide a general non-linear parameterized mapping between a set of inputs. A network with three layers of weights and sigmoidal activation functions can approximate any smooth mapping and such a type will also be used here. A BP neural network is schematically illustrated in figure 1.

Fig. 1. BPNN model

The first layer is an input layer, the second layer is a hidden layer, and the third layer is an output layer. The hidden layer node function adopts the non-linear sigmoidal function, as follows:

f ( x) =

1 . 1 + e−x

(1)

where x is the neural node input vectors. The output of the kth node in the hidden and the output layers can be described by

⎞ ⎛ ok = f (net k ) = f ⎜⎜ ∑ wkj o j ⎟⎟ . ⎠ ⎝ j where the

(2)

net k is the input of the kth node.

The interconnection weights, adjusted in such a way that the prediction errors on the training set can be minimized, are gived by

Δw ji (t ) = ηδ j oi . ⎧ f ' (netj )(y j − o j ) ⎪ δ j = ⎨ f ' (net ) δ w j ∑ k kj ⎪⎩ k

( node j is in output layer) ( node j is in hidden layer)

(3)

.

(4)

Structural Damage Alarm Utilizing Modified Back-Propagation Neural Networks

where

275

0 < η < 1 is the learning rate coefficient, Δw ji (t ) is the actual change in the

weight and

δ j is the error of the jth node, o j

output layer,

is the actual output of the jth node of

y j is the corresponding target output.

In order to control the network oscillations during the training process, a momentum coefficient 0 < α < 1 is introduced to the definition of the weight change:

Δw ji (t + 1) = αΔw ji (t ) + (1 − α )ηδ j oi .

(5)

Once the change is computed, the new weight is given by

Δw ji (t + 1) = αΔw ji (t ) + (1 − α )ηδ j oi . where the value of the momentum coefficient

(6)

α is given by

e(n) < e(n − 1) × 1.05 ⎧0 ⎪ α = ⎨0.9 e(n) > e(n − 1) . ⎪α other ⎩

(7)

where e( n) is the difference between the nth actual output and target output.

3 Steel Girder Bridge Damage Simulation Figure 2 shows a finite element model of plane steel girder bridge. There are thirteen elements and eight nodes in the finite element model. The structural material elasticity E is 7.2 × 1010 Pa , and mass density ρ is 2.8 × 10 3 kg / m 3 . The bridge length and height are respectively 2000mm and 300mm. In order to simulate structure damage, this paper adopts reducing the finite element elasticity. Table 1 shows five damage simulation cases.

Fig. 2. Finite element model of steel girder bridge

、

、

Because frequency is a simple economical easily gained modal parameter and its precision is easily guaranteed, this paper chooses MFCR (modal frequency change ratio) qua modified BPNN input character parameter[9]. Figure 3 shows former four MFCRs at undamage and five damage condition.

276

X. Dong Table 1. Steel girder bridge damage simulation cases Case

Damage element

Damage degree %

η/%

Case1 Case2 Case3 Case4 Case5

E2 E4 E5 E6 E7

40 30 20 30 40

5.1 2.7 1.3 3.6 7.2

In view of measure noise influence, a normal school random data is added on every last MFCR to simulate actual measure data. The random data mean and mean squared error are 0 and 0.005. The random data length is 300. Three hundred datum sample got at undamaged condition are used to train modified BPNN, and other datum sample got at five damage condition are used to test modified BPNN.

Fig. 3. Former four MFCRs of steel girder bridge

4 Damage Alarm Result Analysis Figure 4 show the training and testing results that got by using modified BPNN, and figure 5 show the training and testing results that got by using BPNN. In figure 4 and 5, the former half part are alarm indexes, and other part are the singular indexes. From figure 4, singular indexes of other damage cases obviously depart from alarm indexes except for damage case 3, which give a definitude damage alarm. From figure 5, singular indexes of other damage cases can’t give a definitude damage alarm except for case 1 and 5. The above analysis results show that modified BPNN classifying ability is better than BPNN, and the modified BPNN is more suitable than conventional BPNN for structure damage alarm system. Singular indexes in figure 4 can give a definitude damage alarm except for damage case 3, which relates with the error due to measure noise and modal sensitivity due to damage. All measure datum noise level is 0.005 that its corresponding most measure error is ± 1.5% . Here, the max. η in former four modal frequency sensitivity (MFCR due to damage) of every damage case is shown by table 1. From table 1, MFCRs in damage cases where alarm index can give a definitude alarm are respectively 5.1 , 2.7 , 3.6 , 7.2 . All of them are more than 1.5 , and that MFCR in

％ ％ ％ ％

％

Structural Damage Alarm Utilizing Modified Back-Propagation Neural Networks

％

277

％

damage case 3 is 1.3 that is less than 1.5 . Through frontal analysis, the conclusion are gained that RBFNN can give definitude alarm if MFCR due to damage isn’t less than measure error.

Fig. 4. Damage alarm result of modified BPNN

Fig. 5. Damage alarm result of BPNN

5 Conclusion In order to insure bridge structure safety, it is very important to detect and repair damage as soon as possible. Damage alarm is the first step among structure damage identification, and an more important step also. The traditional damage alarm methods are mostly based on conventional BPNN. BPNN is a global approach Neural Network with weak mode classifying ability and anti-noise ability. So damage alarm effect is not good based on conventional BPNN. This paper proposes a new modified BPNN for structure damage alarm system. The experiment results of steel truss girder bridge show that the proposed method is better than the old one for structural damage alarm.

278

X. Dong

In addition, through modal frequency sensitivity analysis, the conclusion are gained that RBFNN can give definitude alarm if MFCR due to damage isn’t less measure error. Acknowledgments. This research is sponsored by the Aviation Science Foundation of china (No. 2008ZA55004).

References 1. 2. 3. 4. 5. 6.

7. 8.

9.

Dutta, A., Talukdar, S.: Damage detection in bridges using accurate modal parameters. Finite Elements in Analysis and Design 40, 287–304 (2004) Zhao, J., Ivan, J.N., DeWolf, J.T.: Structural damage detection using artificial neural networks. J. Infrastruct. Syst. 4, 93–101 (1998) Shi, Z.Y., Law, S.S.: Structural Damage Location From Modal Strain Energy Change. Journal of Sound and Vibration 218, 825–844 (1998) Stubbs, N.S., Osegueda, R.A.: Global non-destructive damage detection in solids. The Int. J. of Analytical and Exp. Modal Analysis, 81–97 (1990) Dong, X., Sun, Q., Wei, B., Hou, X.: Research on Damage Detection of Frame Structures Based on Wavelet Analysis and Norm Space. In: ICIII 2009, pp. 39–41 (2009) Ko, J.M., Ni, Y.Q., Chan, T.H.T.: Feasibility of damage detection of Tsing Ma bridge using vibration measurements. In: Aktan, A.E., Gosselin, S.R. (eds.) Nondestructive Evaluation of Highways, Utilities and Pipelines IV. SPIE, pp. 370–381 (2000) Worden, K.: Structural fault detection using a novelty measure. Journal of Sound and Vibration 1, 85–101 (2001) Chan, T.H.T., Ni, Y.Q., Ko, J.M.: Neural network novelty for anomaly detection of Tsing Ma bridge cables. In: International Conference on Structural Health Monitoring 2000, Pennsylvania, pp. 430–439 (1999) Dong, X.-m., Zhang, W.-g.: Improving of Frequence Method and Its Application in Damage Identification. Journal of Aeronautical Materials 26, 17–20 (2006)

Computation of Virtual Regions for Constrained Hybrid Systems Jianqiang Li1, Zhen Ji1, and Hai-long Pei2 1

Shenzhen City Key Laboratory of Embedded System Design, College of Computer Science and Software Engineering, Shenzhen University, Shenzhen, 518060, China 2 Department of Automation, South China University of Technology, Guangzhou 510641

Abstract. An efficient method for computing invariant sets of mode transitions systems is proposed. The method is based on the convex optimization technique, the linear matrix inequality techniques and an iterative method. Computation of ellipsoidal invariant sets and polyhedral invariant sets is given. An invariant set for mode transitions systems which includes a finite-duration transition is computed by an iterative procedure. Finally, an example is given to validate this method. Keywords: mode transitions systems, convex optimization technique, linear matrix inequality (LMI), iterative method.

1 Introduction Hybrid systems are systems which include discrete and continuate dynamics. In many applications, hybrid systems have multiple operating modes, each described by a different dynamical structure. In this paper, a situation of hybrid system is introduced where transitions between the modes are caused by external events or disturbances and the mode transitions include a finite-duration transient phase, as a mode transition may correspond to a failure of the system [1].Examples of systems with this type of protective switching action include power system [12] [13]. Invariant sets of hybrid system play an important role in the many situations when dynamics system is constrained in some way. In this paper situations where transitions between the modes are caused by external events or disturbances are studied. The purpose of this paper is to study the transient behavior and establish the invariant sets of the system. The mode transitions are modeled, and the variability defines the region of the dynamic system. Efficient computation for the invariant set is proposed. This method is attractive as the invariant set is useful to design the switching strategies.

2 Mode Transition with Transient Behaviors Roughly speaking, hybrid systems are dynamic systems that involve the interaction of different types of dynamics. Consider a system

x(t ) = f ( x(t )) R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 279–284, 2011. © Springer-Verlag Berlin Heidelberg 2011

(1)

280

J. Li, Z. Ji, and H.-l. Pei

The function f: R → R is Lipschitz continuous and x(t) ∈ R is the continuous state vector are the sufficient condition for the existence and uniqueness of solutions. A mode transition due to external events such as a fault or control switching can be described by a sequence of discrete states. When a transition happen, for example, Si → Si +1 , there may exit a reset function Ri ,i+1 (i) to reset the value of the system to a new value. However, a state transition will not cause the instant reset of the continuous part of the system, there may exit a transient phase between two discrete states. The model can be shown as Figure 1 where the system has three phases which are represent by M 0 , M 1 , M 2 . A mode transition is defined as follow. n

n

n

Definition 1 (Mode Transition). A mode transition caused by an event at time t = t1 is described by three system structures [1].

M 0 (pre-transition): x(t ) = f 0 ( x(t )) , t0 ≤ t ≤ t1 ,

M 1 (transient-phase): x(t ) = f1 ( x(t ), w(t )) t1 ≤ t ≤ t2 M 2 (post-transition): x(t ) = f 2 ( x(t )) t2 ≤ t ≤ t∞

(2)

In definition 1, when the event such as a fault happen, it cause the mode transition occur. The system dynamic change from x(t ) = f 0 ( x(t )) to x(t ) = f1 ( x(t ), w(t )) , where the signal w ∈ W represents the uncertainty in the transient dynamic. The closed set W represents the range of uncertainty in the transient dynamic. The mode transition is completed and changed to x(t ) = f 2 ( x(t )) .

x = f 0 ( x)

x = f1 ( x, w) x = f 2 ( x )

Fig. 1. Mode transition dynamic

The system modes

f i : i =0,1,2 are Lipschitz continuous in x , and the invariant set of

the system is discussed in the latter sections.

3 The Invariant Sets An (positive) invariant set of a dynamic system is a subset of the state space that once the state enters this set it will remain in it for all future times in it [8]. i.e

x (0) ∈ X → x (t ) ∈ X

for all t>0.

Where x(t) is the subset of the dynamic system at time of t and X is a subset of the state space. Consider the continuous dynamic system (3) x(t ) = f ( x(t ))

Computation of Virtual Regions for Constrained Hybrid Systems

281

Where f: n → n is a continuous Lipschitz function. An sufficient and necessary for X is the invariant set is that the differential equation is directed into the set at each point an the boundary: ∂X . Theorem 1 (Nagumo,1942). Assume the system (3) admits a unique solution for

x0 ∈ n .The closed set κ is positively set for the system (3) if and only if for all x ∈κ . f ( x) ∈ κ x (4) From the theorem, a necessary and sufficient condition for system (3) is every point on the boundary ∂κ is directed into the set. This can be expressed as below:

nκ ( x)T f ( x) ≤ 0 ∀x ∈ κ Where nκ ( x) denotes a normal to ∂κ at inequality

(5)

x . The invariant set is described by an

κ = {x ∈

n

| V ( x) ≤ b}

(6)

V ( x) which defines the invariant set is a function of x. There are two important families of invariant sets. These are the classes of ellipsoidal sets and polyhedral sets. Mode transition dynamic system or continuous systems have these types of invariant sets. Ellipsoidal sets are used widely as invariant sets in continuous system. From the existence of a quadratic Lyapunov function for such system and that levelsets of Lyapunov functions are invariant sets [8]. A corollary can be deduced from it: n× n

Theorem1 A system x = A( x) , x ∈ , A∈ , if A has all non-positive real-part eigen -values, then the system has ellipsoidal invariant set. Ellipsoidal sets are popular invariant sets. An ellipsoidal invariant set can be expressed as follow: n

δ = {x ∈ Or

P

δ = {x ∈

n

n

| xT Px ≤ 1}

| ( x − xa ) P ( x − xa ) ≤ 1} T

(7) (8)

0 is symmetry matrix, and xa is the center of the ellipsoidal invariant. This is

convex optimization problem and can be solved by LMI tools [14]. As an ellipsoidal invariant problem, this set can be computed as follows:

P −1 T Subject to A P + PA ≺ 0 , P 0 viT Pvi ≤ 1 (9) Given a set of initial states δ 0 , the condition δ 0 ⊆ δ can be formulated as a linear Minimize

log det

matrix inequality using the so called S-procedure[14].

282

J. Li, Z. Ji, and H.-l. Pei

In fact, polyhedral sets are often natural expressions of physical constraints on states and control variable to the invariant sets. However, the shape of the polyhedral sets is more flexible than that of the ellipsoid, this leads to better approximation to the invariant sets and domain of dynamic systems. This flexible property makes polyhedral sets more representation in the computation. A polyhedral set can be represented in the following form [10]:

δ = {x : Fx ≤ 1}

(10)

Eigenstructure analysis/assignment method is another efficient method. Some contributions show how to determine invariant sets included in polyhedra of the form

δ (G, ρ ) { X : − ρ ≤ Gx ≤ ρ}

(11)

A stabilizing control law u = Kx is assigned [10]

4 The Invariant Sets of Mode Transitions The purpose of the study is to identify the invariant sets of mode transition which caused by external events and includes a finite-duration transition. For the system (1), let

φ f = (t , t1 , x 0 , w)

set P ost f ( x

0

be the solution of the system, the forward reachable

) and backward reachable set Pr e f ( x f ) is defined as follows: P ost f ( x 0 ) =

∪ {φ

f

= (t , t1 , x 0 , w) : x 0 ∈ χ 0 }

w∈W

Pr e f ( x f ) =

∩ {x : x

f

= φ f (t , t1 , x0 , w) : x f ∈ χ f }

(12)

w∈W

The virtual viability region of

Si +1 in state Si are defined as [1], properties of these

operators as well as their computation are discussed in [3].Considering separately at the given t0 and for the duration [t0 , t f ] , Post −1 (t0 , vi +1 ) =

∩ {x | x

w∈W

Post −1 (vi +1 ) =

∩

t∈[ t0 ,t f ]

f

= φw (t f , t0 , x) ∈ vi +1}

Post −1 (t , vi +1 )

(13) (14)

The computation of the safe region and the virtual region are important for applications, and the algorithm of the computation has been shown in [3]. The invariant sets of the mode transition can be computed by follow procedure. 1.Given the initial set of the mode transition system. 2.Compute the invariant set of the pre- transition system by the methods of ellipsoidal sets or polyhedral sets. 3.The invariant sets of the pre-transition system is use as the initial set of the transition system, compute the viability region as the invariant set.

Computation of Virtual Regions for Constrained Hybrid Systems

283

4.The post-transition system is changed to after the duration of transition. The invariant set of the post-transition system is computed from the final viable region of the transition system. The computation of the invariant sets of the mode transition is important for application, but it is difficult to compute and represent high- dimensional systems. The method in subsection 4 is efficient for low dimension system.

5 Example 2

In this sector, a continuous dynamics in R is chosen as the trajectories and sets can be easy to visualize. Convex computation and computational procedures of invariant sets of mode transitions is based on. The computation can be complete by the Matlab’s toolbox. Consider the mode transition system in Figure 1. Mode M 0 , M 2 are the pre-transition and post-transition modes, and mode M 1 is a transition mode which caused by the disturbance and will last for a certain time. Let the systems be given as follows:

⎛0 0⎞ ⎛1 0 ⎞ , ⎛ −1 0 ⎞ , A1 = ⎜ ⎟ B1 = ⎜ ⎟ , A2 = ⎜ ⎟ 1 1 0 1 ⎝ ⎠ ⎝ ⎠ ⎝ 1 −1⎠ ⎛ 0 0 ⎞ , | u |≤ 1 , ⎛ 1 0.5 ⎞ || x ||∞ ≤ 1 B3 = ⎜ A3 = ⎜ i ⎟ ⎟ ⎝ 0.5 1 ⎠ ⎝0 1⎠

，

f1 ( x ) , the invariant set can be computed by iterative procedure. The initial set of the system is given. After iteration, δ 3 = δ 2 The invariant set of the pre-transition system f1 ( x ) is δ 3 . Let the invariant set of f1 ( x ) is the initial set of the transition system f 2 ( x) . With the computation algorithm proposed, the invariant sets of f 2 ( x) has been computed. From the pre-transition system

Fig. 3. Computation of the invariant sets

284

J. Li, Z. Ji, and H.-l. Pei

After a certain time, the integration of viability region for system f 2 ( x ) evolves backward. The invariant sets of the mode transition systems have been shown in Figure 3.

6 Conclusion In this paper, a method to compute the invariant sets for mode transition dynamic is studied. Since the invariant sets can be computed efficiently, the proposed invariant sets make it possible to model-predictive control, protection, decision for mode transitions before the transient actually. A simple example is given in this paper. The applications to the realistic problem are currently being studied. The computation for complex systems is difficult, and it may have a large event sets. More efficient methods are investigated in the next step. The authors gratefully acknowledge the contribution of the National Science Foundation of China [61001185] [61003271][60903114].

References 1. Pei, H.-L., Krogh, B.H.: Stability Regions for Systems with mode Transition. In: Proc. of ACC 2001 (2001) 2. Branicky, M.S.: Multiple Lyapunov Functions and other Analysis Tools for Switched and Hybrid Systems. IEEE Transactions on Automatic Control 43(4) (April 1998) 3. Pei, H.-L., Krogh, B.H.: On the operator Post− 1 Technical Report, Dept. of Electrical and Computer Engineering, Carnegie Mellon University (2001) 4. Lygeros, J.: Lecture Notes on Hybrid Systems, Dept. of Electrical and Computer Engineering, University of Patras (February 2-6, 2004) 5. Donde, V.: Development of multiple Lyapunov functions for a hybrid power system with a tap changer, ECE Dep., University of Illinois at Urbana Champaign (2001) 6. Mayne, D.Q., Rawings, J.B.: Constrained Model Predictive: Stability and Optimality, Automatic 36 789–814 (2000) 7. Zhang, P., Cassandras, C.G.: An Improved Forward Algorithm for Optimal Control of a Class of Hybrid Systems. In: Proc. Of the 40th IEEE CDC (2001) 8. Jirstrand, M.: Invariant Sets for a Class of Hybrid Systems. In: IEEE, CDC 1998 (1998) 9. Girard, A., Le Guernic, C., Maler, O.: Efficient computation of reachable sets of linear time-invariant systems with inputs. In: Hespanha, J.P., Tiwari, A. (eds.) HSCC 2006. LNCS, vol. 3927, pp. 257–271. Springer, Heidelberg (2006) 10. Blanchini, F.: Controlled-invariant sets (2006) 11. Chutinan, A., Krogh, B.H.: Compuring Polyhedral Approximations to Flow Pipes for Dynamic Systems. In: 37th IEEE Conference on Decision & Control 12. Kunder: Power System Stability and Control. McGraw-Hill, New York (1994) 13. Pai, M.A.: Power System Stability. North-Holland Publishing Co., Amsterdam (1981) 14. Boyd, S., et al.: Linear Matrix Inequalities in System and Control Theory. SIAM, Philadelphia (1994) 15. Bertsekas, Rhodes: Min-max infinite-time reachability problem (1971a)

Fault Diagnosis of Diesel Engine Using Vibration Signals Fengli Wang* and Shulin Duan College of Marine Engineering, Dalian Maritime University, Dalian 116026, P.R. China [email protected], [email protected]

Abstract. Aiming at the characteristics of the surface vibration signals measured from the diesel engine, a novel method combining local wave decomposition (LWD) and lifting wavelet denoising is proposed, and is used for feature extraction and condition evaluation of diesel engine vibration signals. Firstly, the original data is preprocessed using the lifting wavelet transformation to suppress abnormal interference of noise, and avoid the pseudo mode functions from LWD. Obtaining intrinsic mode functions(IMFs) by using LWD, the instantaneous frequency and amplitude can be calculated by Hilbert transform. Hilbert marginal spectrum can exactly provide the energy distribution of the signal with the change of instantaneous frequency. The vibration signals of diesel engine piston-liner wear are analyzed and the results show that the method is feasible and effective in feature extraction and condition evaluation of diesel engine faults. Keywords: local wave decomposition, lifting wavelet, Hilbert transform, feature extraction, diesel engine.

1 Introduction The piston-liner wearing will lead to the change of the surface vibration response of diesel engine [1]. When the piston-liner wearing occurs, the vibration signals of the engine is non-stationary. The spectrum based on Fourier transform represents the global rather than any local properties of the signals. For measured signals in practical application with finite data length, a basic period of the data length is also implied, which determines the frequency resolution. Although non-stationary transient signals can have a spectrum by using Fourier analysis, it resulted spectrum for such signals is broad band. For example, the spectrum of a single pulse has a similar spectrum to that of white noise. Consequently, the information provided by Fourier analysis for transient signals were limited. In this paper, local wave analysis is introduced. Instead of relying on convolution methods, this method use local wave decomposition (LWD) and the Hilbert transform [2]. For a non-stationary signal, the Hilbert marginal spectrum offers clearer frequency energy decomposition than the traditional Fourier spectrum. However, the piston-liner wearing characteristics is always submerged in the background and noise signals, which will cause the mode mixture and generate undesirable intrinsic mode functions (IMFs). In order to decrease unnecessary noise *

Corresponding author.

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 285–290, 2011. © Springer-Verlag Berlin Heidelberg 2011

286

F. Wang and S. Duan

influence on LWD, it is important to denoise first before decomposing. In the denoising of traditional wavelet transform, the result of wavelet decomposing is related with wavelet basis function. Moreover, an inappropriate wavelet will overwhelm the local characteristic of vibrating signal, and lost some useful detail information of original signal. To circumvent these difficulties, we present a lifting scheme to construct adaptive wavelets by the design of prediction operator and update operator. The simulation and application analysis results show that the method is feasible and effective for analyzing non-stationary signals and the piston-liner wearing of diesel engine.

2 LWD and Hilbert Marginal Spectrum The local wave analysis is performed into two steps. First, the LWD decomposes the time-series into a set of functions designated as IMFs, and secondly applying the Hilbert transform to those IMFs for generation of the Hilbert spectrum. For any signal, to get a meaningful instantaneous frequency using Hilbert transform, the signal has to decompose a time-series into IMFs which must satisfy two conditions [2]: (1) In the entire data set, the number of extrema and the number of zero crossings must either be equal or differ at most by one; (2) At any point, the mean value of the envelope defined by the local maxima and the envelope defined by the local minima is zero. A practical procedure, known as sifting process, is employed for this purpose. Details are given in [2]. Any signal x(t) can be decomposed into IMFs c1(t), c2(t), . . . , cn(t), and a residue rn(t), n

x (t ) = ∑ ci (t ) + rn (t ) .

(1)

i =1

Applying the Hilbert transform to each IMFs, the original data can be expressed as, n

x (t ) = Re ∑ a j (t)e

iϕ j (t)

.

(2)

j =1

This frequency–time distribution of the amplitude is designated as Hilbert time– frequency spectrum, n

H (ω , t ) = Re ∑ a j (t)e ∫

i ω j (t)dt

.

(3)

j =1

We can also define Hilbert marginal spectrum, T

h(ω ) = ∫ H (ω , t )dt . 0

(4)

where T is the total data length. The Hilbert marginal spectrum offers a measure of the total amplitude distribution from instantaneous frequency.

3 Lifting Wavelet Denosing The lifting scheme can be used to construct adaptive wavelets by the design of prediction operator and update operator [3-5]. It does not rely on the Fourier transform. The principle of lifting scheme wavelet transform is described as,

Fault Diagnosis of Diesel Engine Using Vibration Signals

287

(1)Split: Split the original signal X(k) with the length of L into even sets Xe(k) ={x(2k) , k ∈ Z } and odd sets Xo(k)={x(2k+1), k ∈ Z}. (2)Update: Using a one-point update filter, the approximation signal is computed, c(k)=(Xe(k)+Xo(k))/2. (3)Select prediction operator: Design three different prediction operators N=l: d(k) =Xo(k)-c(k). N=3: d(k) =Xo(k)-[-c(k-1)/8+c(k)+ c(k+1)/8].

(5) (6)

N=5: d(k) =Xo(k)-{3[c(k-2)- c(k+2)]/128+c(k) -11[c(k-1)- c(k+1)]/64-c(k+2)}. (7) Where, N is the number of neighboring c(k) while applying the prediction operator, k=1~L/2. An optimal prediction operator is selected for a transforming sample according to minimizing the [d(k)]2. (4) Predict: Compute the detail signal d(k) by using the optimal prediction operator. Because we update first and the transform is only iterated on the low pass coefficients c(k), all c(k) depend on the data and are not affected by the nonlinear predictor. Then reuse these low-pass coefficients to predict the odd samples, which gives the high-pass coefficients d(k). We use a linear update filter and let only the choice of predictor depend on the data. The selection criterion of minimizing the squared error, an optimal prediction operator is selected for a transforming sample so that the used wavelet function can fit the transient features of the original signal. In the signal denoising, apply various thresholds to modify the wavelet coefficients at each level. The wavelet coefficients are modified via soft-thresholding with universal threshold at each level [6].

4 Simulation Let us consider a signal consisting of amplitude and frequency modulation component: x(t ) = 1.5(1 + 0.2 sin(2π × 7.5t )) cos(2π × 30t + 0.2 sin(2π × 15t )) + 3sin(2π ×10t ) + 0.025randn . The total number of data points n=1024, the sampling frequency is 640Hz, and randn is an n × 1 vector of normally distributed random numbers with a mean of zero and a standard deviation of one. The amplitude of the modulation signal is a (t ) = 1.5(1 + 0.2 sin(2π × 7.5t )) . The variation of frequencies of the modulation signal is 27≤f(t)≤33. Fig.1 a) shows Fourier spectrum of simulation signal. The spectrum based on Fourier transform is not capable of representing the characteristics of frequencies and amplitude of the modulation signal. With the comparisons given in Fig.1 b) c), we can see that the straightforward LWD can get a better results than that of LWD after denosing. The Hilbert marginal spectrum shown in Fig.1 d) represents the amplitude distribution changing with each instantaneous frequency and represents the modulation characteristic of simulation signal.

、

288

F. Wang and S. Duan

Fig. 1. Analysis of simulation signal. a) Fourier spectrum. b) results by using the straightforward LWD. c) results by using LWD after denosing. d) Hilbert marginal spectrum.

5 Application The proposed method is applied to diagnosing the diesel engine piston-liner wearing faults. According to the fundamentals of diesel engines, vibrations have a close relationship with the impact of the piston-liner. The characteristics of vibrations generated by a 6BB1 diesel engine were measured by accelerometer mounted on the cylinder body of cylinder 3 correspond to the top dead center, we collected three kind vibration signals from the same cylinder, which represent the engine no wearing, slightly wearing, and serious wearing states. All data were sampled at 25.6 kHz, and the analyzing frequency is 10 kHz. The rotating speed of the diesel engine is 1100 r /min around. Fig.2 a) ~ c) show the vibration signals of the engine no wearing, slightly wearing, and serious wearing situation. From the comparison in the time domain, we can see that the amplitude peaks of no wearing and slightly wearing signals are about the same in the time domain, no distinctness features. But the serious wearing signal’s is the highest. From the Hilbert marginal spectrum shown in Fig.2 d) ~ f). we can see that the marginal spectrum offers a measure of the amplitude distribution from each instantaneous frequency. For no wearing cylinder, the energy of the signal obvious distributes in a lower frequency area which is limited to a range of 2kHz. For slightly wearing cylinder, the lower frequency energy content is low due to leakage of combustion, and

Fault Diagnosis of Diesel Engine Using Vibration Signals

289

much energy distributes in a higher frequency area (5kHz~7kHz) generated by the occurrence of piston slap. For serious wearing cylinder, the peaks of energy of the signal are concentrated on the higher frequency area due to increasing the strength of piston slap generated by the piston-liner wearing, whereas the lower frequency energy content decrease due to increasing the leakage of combustion.

Fig. 2. Vibration signals of diesel engines and Hilbert marginal spectrum. a)~c) vibration signals of no wearing, slightly wearing, Serious wearing, d)~f) Hilbert marginal spectrum of corresponding a)~c).

6 Summary The lifting wavelet transform can overcome the denoising disadvantage of traditional wavelet transform and is adopted to remove noise. It can reduce the mode mixture in LWD, improve the quality of decomposition and obtain a much better decomposition performance. The proposed method can be applied to extract the fault characteristic information of the piston-liner wearing vibration signal effectively.

References 1. 2.

Geng, Z., Chen, J., Barry Hull, J.: Analysis of engine vibration and design of an applicable diagnosing approach. International Journal of Mechanical Sciences 45, 1391–1410 (2003) Huang, N.E., Shen, Z., Long, S.R.: The empirical mode decomposition and the Hilbert spectrum for nonlinear and nonstationary time series analysis. Proceedings of the Royal Society of London 454, 903–995 (1998)

290 3. 4. 5. 6.

F. Wang and S. Duan Sweldens, W.: The Lifting scheme: A custom-design construction of biorthogonal wavele. Appl. Comput. Harmon. Analo. 2, 186–200 (1996) Claypoole, R.L., Geoffrey, M.D., Sweldens, W.: Nonlinear wavelet transforms for image coding via lifting. IEEE Transactions on Image Processing 12, 1449–1459 (2003) Kong, G.J., Zhang, P.L., Cao, J.J.: Signal denoising based on lifting wavelet transform and its application. Computer Engineering and Applications 44, 234–237 (2008) Donoho, D.L.: De-noising by soft-thresholding. IEEE Transactions on Information Theory 41, 613–627 (1995)

Influences on the Directivity of Acoustic Vector Sensor by Soft Finite Cylinder Baffle∗ Ji Jianfei, Liang Guolong, Zhang Guangpu, and Li Yang National Laboratory of Underwater Acoustic Technology, Harbin Engineering University, Harbin, China

Abstract. In free field, the directivity of sound pressure is omni directional, and the directivity of vibration velocity is dipole directional, it can bring many advantages for acoustic measurements by combining utilization of both sound pressure and vibration velocity information. However, under the condition of boundary, the directivity of pressure and vibration velocity will be distorted by diffraction wave. In this paper, the soft boundary element model of finite cylinder baffle is established; the sound diffraction field of a plane wave from it at different frequencies and incident angles is calculated, the characteristics of directivity of pressure and vibration velocity at different frequencies and different distances are analyzed. Keywords: acoustic vector sensor, directivity, baffle, boundary element method (BEM), sound diffraction field.

1 Introduction The directivity of pressure and vibration velocity of acoustic vector sensors will be affected by sound diffraction wave under the condition of boundary [1][2]. Recently, many scholars have carried out the research about it home and abroad. The theoretical and experimental research about the influences of near-field acoustic diffraction caused by the system consisted of the elastic spherical shell filled with air and AVS on the measurement results of acoustic vector sensors were carried out [3-6]. The influences on the beam pattern of the acoustic vector sensor line array by the rigid spherical baffle showed that the beam pattern was seriously distorted by the sound diffraction wave [7]. The sound diffraction field of rigid prolate spheroid at arbitrary incident wave was calculated, and the characteristics of it are concluded. The sound diffraction field of a plane acoustic wave from an impenetrable, soft or hard, prolate or oblate spheroid was also calculated [8-13]. The influences of the infinite circular cylinder baffle on the direction of vector sensor show that the directivity of the vector sensor was seriously distorted by the sound diffraction of infinite circular cylinder [14-15]. In this paper, sound diffraction models of sphere, prolate spheroid, and cylinder are established; the influences on the directivity of acoustic vector sensor by these baffles are analyzed and compared, respectively. The conclusion of the paper will be useful for the design of the shape and the geometric of baffle. ∗

Project supported by the National Natural Science Foundation of China (Grant No.51009042).

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 291–295, 2011. © Springer-Verlag Berlin Heidelberg 2011

292

J. Jianfei et al.

2 Principle The model of sound diffraction by finite cylinder baffle is shown in Fig. 1(a). The height and radius of the cylinder are h and r , respectively. The observation point is located the in the x axis. l is the distance from the observation point to the origin (center of the cylinder). The sound source incidents at a distance and rotates around the observation point 360 ° in the xoy plane. It is difficult to calculate analytical solutions of sound diffraction field for finite cylinder baffle. The numerical solutions are solved by utilizing boundary element method (BEM). The BEM models of cylinder baffles are shown in Fig.2 (b). The height and radius of the cylinder baffles is h = 0.5m and r = 0.5m , respectively. The models are meshed fine enough to meet the calculation accuracy.

(a) schematic diagram of model

(b)BEM model of finite cylinder

Fig. 1. Sound diffraction model of finite cylinder baffle

2.1 Calculation Results The medium inside the cylinder is air and the medium outside the cylinder is water, because the wave impedance of water is much larger than that of air, the boundary can be approximated to absolute soft. As shown in Fig.1 (a), the observation point is located in x axis, and the incident plane wave rotates around the observation point 360 ° in xoy plane with an interval of 5° , the pressure, the vertical (in x axis direction) vibration velocity and the horizontal (in y axis direction) vibration velocity are calculated at every incident angles. When the sound source is located in the positive half axis of x axis, the incident angle is defined as 0 ° . The amplitude of the directivity below has all been normalized. Fig.2 shows that when the distance from the observation to the origin l is 0.5m, the directivity of pressure and vibration velocity at 100Hz, 700Hz, 1300Hz and 3000Hz. As shown in Fig.2, at 10Hz, the directivity of sound pressure is symmetrical, but the amplitude of the pressure is relatively small. As the frequency increasing, after reversed-phase superposition, the depression appears in the directivity pressure. The directivity of vertical (in x axis direction) vibration velocity basically loses the natural

Influences on the Directivity of Acoustic Vector Sensor

pressure

293

vertical vibration velocity horizontal vibration velocity (in x axis direction) (in y axis direction)

Fig. 2. l = 0 .5m , the directicvity of pressure and vibration velocity

pressure

vertical vibration velocity horizontal vibration velocity (in x axis direction) (in y axis direction)

Fig. 3. l = 3.225 m , the directicvity of pressure and vibration velocity

dipole directivity and as the frequency increasing, the grating lobes appear in the directivity pattern. In low frequency, the directivity of horizontal (in y axis direction) vibration velocity of diffraction wave is shown dipole in shape. When the sound source is in the other side of the baffle relative to the observation point, the intensity of diffraction wave at the observation point is enhanced as the frequency increasing, so the dipole directivity is right deviated. Fig.3 shows when the distance from the observation point to the origin l is 3.225m, the directivity of pressure and vibration velocity at different frequencies. Because the distance l becomes longer, the intensity of diffraction wave is decreased, and the directivities of pressure and vibration velocity are slightly distorted.

3 Conclusions The BEM model of finite cylinder is established, and the directivity of pressure and vibration velocity of the vector sensor by soft finite cylinder baffle at different frequencies and distances is analyzed. The conclusions are as follows:

294

1.

2.

J. Jianfei et al.

When the distance from vector sensor to the finite cylinder baffle l is relatively small, the intensity of diffraction wave is strong, the directivities of pressure and vibration velocity are affected by the diffraction wave. When the distance from the vector sensor to the finite cylinder baffle l becomes longer, the intensity of diffraction wave is decreased, so the directivities of pressure and vibration velocity are slightly distorted. The influences of diffraction wave on vertical (in x axis direction) vibration velocity and horizontal (in y axis direction) are different. Most of the diffraction wave energy concentrates in the vertical (in x axis direction) direction, the vertical (in x axis direction) vibration velocity is more seriously affected than the horizontal (in y axis direction) vibration velocity by diffraction wave.

References 1. Sun, G., Li, Q., Yang, X., Sun, C.: A novel fiber optic hydrophone and vector hydrophone. Physics 35(8), 645–653 (2008) 2. Jia, Z.: Novel sensor technology for comprehensive underwater acoustic informationvector hydrophones and their applications. Physics 38(3), 157–168 (2009) 3. Kosobrodov, R.A., Nekrasov, V.N.: Effect of the diffraction of sound by the carrier of hydroacoustic equipment on the results of measurements. Acoust. Phys. 47(3), 382–388 (2001) 4. Shi, S., Yang, D., Wang, S.: Influences of sound diffraction by elastic spherical shell on acoustic vector sensor measurement. Journal of Harbin Engineering University (27), 84–89 (2006) 5. Shi, S.: Research on vector hydrophone and its application for underwater platform. Doctor Dissertation of Harbin Engineering University (2006) 6. Sheng, X., Guo, L., Liang, G.: Study on the directivity of the vector sensor with spherical soft baffle plate. Technical Acoustics (9), 56–60 (2002) 7. Zhang, L., Yang, D., Zhang, W.: Influence of sound scattering by spherical rigid baffle to vector-hydrophone linear array. Technical Acoustics 28(2) (April 2009) 8. Barton, J.P., Nicholas, L.W., Zhang, H., Tarawneh, C.: Near-field calculations for a rigid spheroid with an arbitrary incident acoustic field. J. Acoust. Soc. Am. 113(3), 1216–1222 (2003) 9. Rapids, B.R., Lauchle, G.C.: Vector intensity field scattered by a rigid prolate spheroid. J. Acoust. Soc. Am. 120(1), 38–48 (2006) 10. Roumeliotis, J.A., Kotsis, A.D., Kolezas, G.: Acoustic Scattering by an Impenetrable Spheroid. Acoustical Physics 53(4), 436–447 (2007) 11. Ji, J., Liang, G., Wang, Y., Lin, W.: Influences of prolate spheroidal baffle of sound diffraction on spatial directivity of acoustic vector sensor. SCIENCE CHINA Technological Sciences 53(10), 2846–2852 (2010) 12. Ji, J., Liang, G., Liu, K., Li, Y.: Influences of soft prolate spheroid baffle on Directivity of Acoustic Vector Sensor. In: IEEE International Conference on Information and Automation, ICIA 2010, pp. 650–654 (2010)

Influences on the Directivity of Acoustic Vector Sensor

295

13. Ji, J., Liang, G., Huang, Y., Li, Y.: Influences on spatial directivity of acoustic vector sensor by soft spherical boundary. In: The 2010 International Conference on Computational and Information Sciences (2010) 14. Li, C.: Combined signal processing technology with acoustic pressure and particle velocity. Doctor Dissertation of Harbin Engineering University (2000) 15. Chen, Y., Yang, B., Ma, Y.: Analysis and experiment study on directivity of vector sensors located on complicated boundaries. Technical Acoustics (25), 381–386 (2006)

The Method of Intervenient Optimum Decision Based on Uncertainty Information Lihua Duan Department of Information, Liaoning Police College Dalian, 116036, P.R. China [email protected]

Abstract. This paper discusses the method of intervenient optimum analysis based on non-optimum theory of system, and points out main problems of exploring uncertainty decision. Then analyses the key is short of the intervenient optimum to the problems of uncertainty decision. The article establishes drowsy set and intervenient optimum principle based on uncertainty problems analysis; and puts forward that decision system is degree measured along with non-optimum traced up to systematic optimum and so on. Keywords: non-optimum, intervenient optimum analysis, drowsy set, intervenient optimum principle, uncertainty decision.

1 Introduction One of motives of traditional optimization theory is to express mankind seek perfection of things. Practice has showed that people could not have an accurate judgment for the perfection. Thus limitations of traditional optimization theory are to be reflected. Literature [1] puts forward system non-optimum theory, and point out system non-optimum decides its optimum, and it proved the motto “Failure is the mother of success” from the angle of quantitative analysis. Literature [2] discussed translation from non-optimum to optimum system, and conversion disorder to order of relations from perspective of self-organization, and obtained the conclusion, i.e. non-optimum system recognition and control equal to system achieving synergies. Literature [3] expounded non-optimum analysis is a generalized interval analysis, it is a development based on sub-optimum theory and the third optimum method. Literature [4, 5, 6] studied the method of foundation, recognition and control of non-optimum system. The author resolved the scheduling and decision-making problems of Dalian Chemical Industries Parent Company in 1992 by using non-optimum analysis method. Actual application shows that this method is more practical and reliable than traditional optimization method, and it has been achieved more economic benefits [7]. Practice proves that system non-optimum theory is a system decision-making method with realistic backgrounds and practical significance. Meanwhile, it embodies characteristics of mankind intelligent behaviors. So system non-optimum theory has important practical background and broad application prospect. In this paper, we bring forward the concept of drowsy set in the basis of the previous study and R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 296–301, 2011. © Springer-Verlag Berlin Heidelberg 2011

The Method of Intervenient Optimum Decision Based on Uncertainty Information

297

practice, and put forward the intervenient optimum decision-making method from the angle of system optimization, then discuss its practical significance and concrete solution methods.

2 Basic Concepts of Intervenient Optimum Analysis 2.1 Background and Meaning The previous system analysis committed that it is impossible to realize optimum under a limited condition of time and resources. At the same time, behind the optimum, there is definitely a series of hypotheses, middle-way decisions, and predigesting of data. Under most conditions, the hypotheses of optimum do not exist. Although people have generalized this method to many fields, the results obtained can be only temporary, and sometimes cannot achieve the final goals [1]. In real life, there are no absolute optimums, and only under certain conditions, is there differentiated relative optimum. Relative optimum can be seen as satisfactory result, because there are a great deal of uncertainty and non-linearity as explained by Simone. There are three defects in the traditional decision disciplines: to ignore the uncertainty of the economic life; to ignore the non-linear relationship of the real life; to ignore the subjective limitations of the decision maker. Simone held that in the complicated real world, only a minority of problems can be solved through the calculus of the maximum and minimum value. Sometimes there are not the most optimal solutions at all, and under most conditions people just try to find out a satisfactory approximate solution (relative optimum solution) [2]. In real problems, the heart of decision analysis is how to define the boundary of problem P between non-optimum and optimum. It is also the starting point of nonoptimum analysis studies. The meaning of non-optimum lies in definition of boundary. It is a basic point of systems transformation, and it is called intervenient optimum. States and behaviors of systems are in intervenient optimum under the most conditions. Moreover so-called optimum and non-optimum can not exist independently. Anything has optimum and non-optimum with various degrees. When optimum degree is greater than non-optimum, system has optimum attributes; when optimum degree is less than non-optimum, system has non-optimum attributes. If the problem we researched is a choice between optimum and non-optimum, then the kind of studies method is called intervenient optimum. 2.2 Concepts and Expressions of Intervenient Optimum

Ｃ Ｃ ︱ ，Ｃ ︱ ， ，Ｃ ︱ Ｐ ， ， ， → f (C ) λ ∈ (−n, n) , θ ∈{θ ， θ ， … ， θ }, Where f (C ) is

Definition 1. Suppose ={ 1 θ1 2 θ2 … n θn} be needed character set of problem under the circumstances of uncertain degree θ={θ1 θ2 … θn}, then

∀C i θ i

i

i

1

needed eigenvalue under recognized specification trusted degree.

2

n

Z (C ) λ .λ= r K x (C )

i

is real

Z r (C ) expresses recognition degree of subjectivity for characters of

298

L. Duan

problems.

K x (C ) expresses total recognition degree for characters of prob-

lems(general recognition, objective recognition, past recognition for same problems)

Ｐ

S (C i ) of the problem under recognized specification λ , and S (C i ) − f (C i ) = N O (C i ) is the value of quantity of character C i about non-optimum attributes under recognized specification λ , then we have Definition 2. If it exists real eigenvalue

S (C i ) − f (C i ) > 0 ⎧O(C i ) ⎪ P(C i ) = ⎨SO(C i ) S (C i ) − f (C i ) = 0 ⎪ NO(C ) S (C ) − f (C ) < 0 i i i ⎩

Ｐ

O(C i ) expresses that problem has optimum attributes on the needed character C i ; SO (C i ) expresses that problem has intervenient optimum attributes on the needed character C i ; NO (C i ) expresses that problem has non-optimum attributes on the needed character C i . Where

Ｐ

Ｐ

Definition 2 puts forward a method of judging optimum and non-optimum in the real analyzing problems. The setting standard of target is ditions of specification λ . If

，

N O (C i ) =0 under the con-

Ｐ is belong to non-optimum

N O (C i ) < 0 the problem

N O (C i ) . Accordingly the value of S (C i ) is also increased. If N O (C i ) > 0 the problem is belong to λ optimum state, the degree of specification λ decides the optimization standard of problem . In fact, N O (C i ) =0 is basic conditions. That is to say, optimization problems in the reality should be intervenient optimum based on specification λ . attribute, the studied angle is to decrease the value of

，

Ｐ

Ｐ

3 Basic Methods of Sub-optimum 3.1 Hesitation Set According to the above discussion, we can find that any problem P has a eigenvalue f (C i ) under recognized specification λ , it is a description of object states. The critical issue, whether or not it reflects the principles of authenticity reliability, is defined the value of λ . The determining method of the value of λ is given as the following: Uncertainty is an essential attribute of mankind cognition, it is divided into subjective uncertainty and objective uncertainty. Subjective uncertainty embodies sensibility and perceptibility of mankind. If the uncertainty of character attributes of things may be appeared alternatively and duplicated in brains. Then it is called hesitant degree of

The Method of Intervenient Optimum Decision Based on Uncertainty Information

299

things (drowsy degree, for shot). It is a recognized process from uncertainty to certainty, then from certainty to uncertainty. Thus we have:

Ｐ

= {c1 , c 2 , L, c n } be needed character field of problem . It exists a perceivable recognition set M for . For any ci ∈ C (i = 1,2,L , n) , if it exists uncertain degree θ i ∈ M → [0 1], it could decide (ci ) directly, then M is called hesitant set. Uncertain degree θ i , which is belong to needed character ci ∈ C (i = 1,2,L , n) Definition 3. Suppose C

：

，

Ｐ

Ｐ

Ｐ，can be calculated on the basis of the definition of hesitant sets.

of the problem Thus we have:

Definition 4. Suppose P ( f (λ ci )) be a probability of needed character eigenvalue

P ( S (λ ci )) be a probability of real character eigenvalue under recognized specification λ . After limited hesitant process went by, it can be obtained a uncertain degree θ i (i = 1,2, L , n) . Thus we have: under recognized specification λ , and

n

θi = lim ∑ n →∞

i =1

{P( f n (λi ci )) ≠ P( Sn (ci ))} − {P( f n (λi ci )) = P( Sn (ci ))} P{Sn (Ci ) = f n (Ci )}

，

λ = Z r (C ) K (C ) x In the analysis of uncertainty, hesitant sets should be built firstly. Then uncertain degree could be attained by the way of hesitant sets. In reality, uncertainty has some distributions, it can be obtained through statistical laws of limited hesitant processes θ 1 ,θ 2 ,L ,θ n . 3.2 Principles of Intervenient Optimum If a decision problems Pi is composed of conditions (Ci ) and targets (Oi ) in any systems S, then D{P: (C,O)}is called decision space. If a problem P could be divided into optimum category and non-optimum category under the recognized conditions. And it can be decreased non-optimum degree and increased optimum degree, then the system is called intervenient optimum system.

﹤

﹥ ﹤

Definition 5. When it has a choice interval of optimization Oab= ao ,bo for problems P in systems S, it must exist influence interval of non-optimum Nab= an ,bn , if ∀o ∈ ao ,bo ∀n ∈ an ,bn then Sab= Oab ∩Nab is called intervenient optimum interval. According to the above discussions, it is well known that so-called coexistence is decided by the relevance between optimum and non-optimum system. Therefore, intervenient problems could be studied by using correlated functions of Extenics.

﹤

﹥，

﹤

﹥，

﹥

300

L. Duan

Definition 6. Suppose s 0 be any point on R, and

S 0 = be any intervenient op-

timum interval for systems S on the field of real numbers. Under the conditions of specification λ ,

ρ (s0 , S λ ) = s0 − denotes the distance about the point

a+b b−a − 2 2

s0 and intervenient interval S 0 λ0 . Where

may be an open interval or closed interval, and also a half-open interval. Let d ( s0 , S 0 ) denote the distance of a point and interval in the classical mathematics. The relations between

①When s ②When s

ρ ( s 0 , S 0 λ0 ) and d ( s0 , S 0 ) is as the following:

0

∉ S 0 or s0 = a, b , then ρ ( s 0 , S 0 λ0 ) = d ( s 0 , S 0 ) ≥ 0 ;

0

∈ S 0 and s 0 ≠ a, b , then ρ ( s 0 , S 0 λ0 ) < 0 , d ( s0 , S 0 ) = 0

With the introduction of a concept of distance, the location relations of a point and its intervenient interval could be described accurately by the quantitative analysis method. When a point is within the interval, the distance between the point and interval is equal to zero in the classical mathematics. Whereas in the analysis of intervenient optimum, different points have different locations according to different values of distance. The concept of distance describes the located relations between points and intervals. That “It is the same within a class” is an original recognition, now it has developed that “it is different degree within a class” by the quantitative analysis method. In the analysis of system intervenient optimum, we must think not only the located relations between points and intervenient intervals (or non-optimum and optimum), but also the located relations between non-optimum intervals and optimum interval, as well as the located relations between a point and two intervals. Therefore, we have:

N 0 = 〈 a, b〉, O = 〈c, d 〉 , and N 0 ⊂ O , then the nested interval location of a point x , composed by interval N 0 and O , is defined as: Definition 7. Suppose

⎧ ρ ( x, O ) − ρ ( x, N 0 ) x ∉ N 0 ⎪ D ( x, N 0 , O ) = ⎨ ⎪− 1 x ∈ N0 ⎩

D( x, N 0 , O) describes the located relations between a point x and the nested interval of N 0 and O . Based on the analysis of values of distance, the degree of intervenient optimum is expressed as the following:

J (u ) =

ρ ( x, N 0 )

D( x, N 0 , O)

λ

The Method of Intervenient Optimum Decision Based on Uncertainty Information

Where N 0

301

⊂ O , and there is no common end vertex. The range of intervenient op-

（－ ， ）

timum degree is ∞ +∞ . The above mentioned formula expresses intervenient optimum degree of non-optimum analysis, it expands recognition of non-optimum from qualitative description to quantitative description. In the analysis of intervenient optimum degree , J(x)≥0 expresses x belongs to optimum system, k(x)≤0 expresses x belongs to non-optimum system. The value of J(x) and k(x) expresses its degree individually. k(x)=0 expresses x belongs to the boundary of system. Therefore, intervenient degree is a transferred tool, which is a quantitative description about things’ transformation from non-optimum to optimum. In this way, a new thoughtway comes into being, namely intervenient optimum principle. What is called intervenient optimum principle means that in the decision analysis, any target and behavior have optimum and non-optimum attributes, in varying degrees. The optimum attributes under non-optimum state is called intervenient optimum. It coexists with optimum and non-optimum.

4 Conclusions Intervenient optimum analysis is used to study problems of system optimizations from the angle of non-optimum analysis. It offers a fresh thoughtway for optimal decision. Research indicates that the kernel of uncertainty decisions is to seek the measurement of uncertainties. Because of the existence of uncertainties, the system’s optimization can not be optimum but intervenient optimum. It exists drowsy attribute in the process of decision for uncertain problems. Therefore drowsy set is an effective method in resolving this kind of uncertain problem with intervenient optimum attribute. If it has abilities of controlling drowsy attribute and judging drowsy number, then the reliability of decision will be improved.

References 1.

2. 3.

4. 5.

6.

Qu, Z., He, P.: Intelligence Analysis Based on Intervenient Optimum learning Guide System. In: International Conference on Computational Intelligence and Natural Computing, pp. 363–366. IEEE Computer Society, Los Alamitos (2009) Ping, H.: Theories And Methods of Non-optimum Analysis on systems. Journal of China Engineering Science 5(7), 47–54 (2003) He, J., He, P.: A New Intelligence Analysis Method Based on Sub-optimum Learning Model. In: ETP International Conference on Future Computer and Communication, pp. 116–119. IEEE Computer Society, Los Alamitos (2009) He, P.: Method of system non-optimum analysis based on Intervenient optimum. In: Proc. of ICSSMSSD, pp. 475–478 (2007) He, P.: Methods of Systems Non-optimum Analysis Based on Intervenient Optimum. In: Wang, Q. (ed.) Proc. of the Int’l conf on System Science, Management Science and System Dynamic, pp. 661–670 (2007) He, P., Qu, Z.: Theories and Methods of Sub-Optimum Based on Non-optimum Analysis. ICIC Express Letters, An International Journal of Research and Surveys 4(2), 441–446 (2010)

The Deflection Identify of the Oil Storage Tank* Jingben Yin**, Hongwei Jiao, Jiemin Zhang, Kaina Wang, and Jiahui Ma Department of Mathematics, Henan Institute of Science and Technology, Xinxiang 453003, China [email protected]

Abstract. In this paper, we consider question A in higher education community cup 2100 mathematical contest in modeling competition. We lucubrate the deflection identify of the oil storage tank and the demarcation matter of the oil storage tank’s volume, and builds models to solve the relationship between volume and depth storing up oil. We build the corresponding model with the method of linear regress, data close, interpolation, finally we obtain the required value with MATLAB software. We build an integral model with integral formula, and get the relationship between volume and depth storing up oil, basing on the integral formula with the method of closing polynomial functions to build a new function formula which makes it easier to be solved. By analysis the error and comparing the curve generated with the real data, the error is less. Keywords: Fitting model, Interpolation model, Polynomial model, Newton alternate era method.

1 Introduction Usually, the gas station has some underground oil storage tank of deposited fuel, and equips “depth storing up oil measure manages system”. People adopt flowmeter and oil digit meter to measure data of in/out oil capacity and depth storing up oil and so on. Via beforehand demarcating tank capacity table (the relationship between volume and depth storing up oil) to put up real time calculation, receiving transformation instance of volume and depth storing up oil. Owing to groundwork distortion, many oil storage tanks location will occur portrait lean and landscape orientation deflection after a short time, which lead to transformation of oil capacity table. According to related regulation, we need to demarcate the oil tank after a period of time. In order to master the impact of oil capacity table after deflection, we impose a little ellipse type oil storage tank (two ends crop ellipse cylinder).We do experiment on tank without deflection and slope bevel serves as α = 4.1 when it occurs portrait deflection. After the study we obtain the numerical tabular of the depth of oil each 1cm when the oil tank deflected [1,2]. 0

*

This paper is supported by the National Natural Science Foundation of Education Department for Henan Province (2008A110006, 2008B110004). ** Corresponding author. R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 302–307, 2011. © Springer-Verlag Berlin Heidelberg 2011

The Deflection Identify of the Oil Storage Tank

303

2 Model Supposition 1) The hypothesis adopt flowmeter and the oil level plan to come to measure data such as entering/ out oily amounts and painting altitude within the jar; 2) Assumes an oil storage tank after using a period of time, because foundation deformation waits for cause, may change such as longitudinal incline and lateral deflection happened in using the location that the jar experiences and observes (the following be called deflection); 3) Assumes the relation not considering the oil volume with the temperature change; 4) Assumes the small ellipse type oil storage tank is both ends oval-shaped closely cropped hair cylinder; 5) Oil storage tank considers as when assuming calculation being that deformation happened in the standard oval-shaped cylinder; 6) Assumes that the oil storage tank thinking because time is long not but bringing about corrodes condition; 7) The hypothesis does not consider an impact of intensity of pressure over volume.

3 Sign Explanation Y is the summation when we add certain oil every time ( L ), X is the depth of the oil when we add the certain oil into the tank ( cm ), a is half of the long axis of elliptic cylinder, b is half of the short axis of elliptic cylinder, l is half of the elliptic cylinder, m is the distance from the probe to the oil tanks side, d is Paint carburetor float altitude after incline and paint the carburetor float altitude difference when horizontal, V is the volume of the oil in the tank, H is the oil level probe demonstrated oil level altitude, L is the length of the elliptic, Δh when H = 0 , Paint the altitude facing in oil tank side, α is longitudinal incline angle, β is landscape orientation

a1 is column shape oil tank’s two sides global hat style locality spheriform radius, b1 is column shape oil tank’s two sides global hat style locality spheriform radius, c1 is column shape oil tank’s two sides global hat style altitude, other incline angle,

signs’ explanation will be showed when they are used.

4 Model Building and Solving Since the above integral model built is complex in calculation, we consider polynomial calculation is simple and convenient and calculate the relationship between volume and depth storing up oil using closing polynomial equation method Owing to

V=

[3]

.

a ⎡ ⎤ ⎛H ⎞ 1 L ⎢( H − b ) H ( 2b − H ) + b 2 arcsin ⎜ − 1⎟ + π b 2 ⎥ b ⎣ ⎝b ⎠ 2 ⎦

304

J. Yin et al.

Let

H −b (0 ≤ H ≤ b, −1 ≤ x ≤ 1) , b

x= then

⎛x ⎞ (1) V ( x ) = ⎜ + x 1 − x 2 + arcsin x ⎟ abL ⎝2 ⎠ ⎛x ⎞ 2 We let a polynomial close ⎜ + x 1 − x + arcsin x ⎟ and get the approximation ⎝2 ⎠ ⎛x ⎞ 2 of the oil in the tank. ⎜ + x 1 − x + arcsin x ⎟ is continue on [ −1,1] , ⎝2 ⎠ x ⎛ ⎞ 2 2 ⎜ + x 1 − x + arcsin x ⎟ =2 1 − x is boundary on [−1,1] , the Chebyshev ⎝2 ⎠ ∞ 1 2 series of x 1 − x + arcsin x is a0T0 ( x ) + ∑ anTn ( x ) , where 2 n =1

)

(

an =

2

π

∫

( x 1− x

1

2

−1

)

+ arcsin x Tn ( x ) 1 − x2

Tn ( x ) is Chebyshev polynomial of n th order), convergences

( x 1− x

2

)

dx ,

∞ 1 a0T0 ( x ) + ∑ anTn ( x ) 2 n =1

+ arcsin x consistent. From (1) we get

∞ ⎡π 1 ⎤ ⎛x ⎞ V ( x ) = ⎜ + x 1 − x 2 + arcsin x ⎟ abL = ⎢ + a0T0 ( x ) + ∑ anTn ( x ) ⎥ abL. 2 2 ⎝2 ⎠ n =1 ⎣ ⎦

We substitute

( x 1− x

2

)

+ arcsin x for

1 a0T0 ( x ) + a1T1 ( x ) + a2T2 ( x ) + ⋅⋅⋅ + anTn ( x ) , 2 from (2) we get

⎡π 1 ⎤ V ( x) ≈ ⎢ + a0T0 ( x ) + a1T1 ( x ) + a2T2 ( x ) + ⋅⋅⋅ + anTn ( x ) ⎥ x ∈ [ −1,1] ⎣2 2 ⎦ Case 1

n is even number an = ∫

1

−1

( x 1− x

2

)

+ arcsin x Tn ( x ) 1 − x2

dx

(2)

The Deflection Identify of the Oil Storage Tank

305

)

(

Tn ( x ) is even number too, x 1 − x 2 + arcsin x is odd number, 1 − x 2

Here

is even number,

[ −1,1] is symmetrical, then

an = a2 k = 0

(k = 0,1, 2 ⋅⋅⋅⋅⋅⋅)

(3)

Case 2 n is odd number Let x = cos t , then

a2 k +1 = =

2

π∫

π

0

sin t cos t cos ( 2 k + 1)tdt +

16 2 ⎡ ( 2k + 1) ⎣ 4 − ( 2k + 1) ⎤⎦ π 2

π

2∫

π

0

⎛π ⎞ ⎜ − t ⎟ cos ( 2k + 1) tdt ⎝2 ⎠

( k = 0,1, 2 ⋅⋅⋅⋅⋅⋅)

(4)

Form (3) and (4), we obtain

n = 2k ⎧ 0, ⎪ 16 an = ⎨ , n = 2k + 1 ( k = 0,1, 2L) 2 ⎪ ( 2k + 1) ⎡ 4 − ( 2k + 1) 2 ⎤ π ⎣ ⎦ ⎩

(5)

（ ）and（5）, we get

From 2

⎧ 1 1 ⎡1 ⎤⎫ ⎪ ⎢ 3 T1 ( x ) − 45 T3 ( x ) − 525 T5 ( x ) + ⋅⋅⋅⎥ ⎪ ⎪ π 16 ⎥ ⎪ abL ( k = 0,1, 2 ⋅⋅⋅ ) (6) V ( x) ≈ ⎨ + ⎢ ⎬ 1 ⎢+ 2 π T ( x) ⎥⎪ ⎪ ⎢ ( 2k + 1)2 ⎡ 4 − ( 2k + 1)2 ⎤ 2 k +1 ⎥ ⎪ ⎪ ⎣ ⎦ ⎣⎢ ⎦⎥ ⎭ ⎩ We substitute ( x 1 − x 2 + arcsin x +

P3 ( x ) = when

π for ) 2

x + a1T1 ( x ) + a2T2 ( x ) + a3T3 ( x ) 2

k = 1 , we get 32 ⎡π ⎤ V ( x) ≈ ⎢ + 9 x − 2 x 2 ) ⎥ abL ( ⎣ 2 45π ⎦

We substitute

π

( x 1 − x 2 + arcsin x + ) for 2 x P5 ( x ) = + a1T1 ( x ) + a3T3 ( x ) + a5T5 ( x ) 2

306

when

J. Yin et al.

k = 2 , we get 16 ⎡π ⎤ V ( x) ≈ ⎢ + 615 x − 80 x 3 − 48 x 5 )⎥ abL ( ⎣ 2 1575π ⎦

According to the above approximately function, we can easily calculate oil storage capacity under the circumstance of given oil level altitude. We can handle MATLAB to close the function between volume and depth with the given measured oil storage capacity and altitude[4-6].

y = −0.0025 x 3 + 0.5394 x 2 + 4.4910 x + 42.1041 According to the request in subject, we can make use of the useable data and closing polynomial model to obtain the numerical tabular of the depth of oil each 1cm when the oil tank deflected. By analysing the error we can get results as below (Table 1). Table 1. Error analyse between model data and real data

oil storage capacity (unit: L) 57.1841 52.138 47.0557 41.9222 36.7225 31.5416 26.4645 21.3762 16.3617 11.506 6.5941 1.811 -2.8583 -7.5288 -11.9155 -16.3334 -20.5975 -24.7228 -28.7243 -32.617 -36.4159 -40.036 -43.4923 -46.9 -50.1 -53.1 -56.2

Error value 2.133159 0.148483 0.125363 0.104721 0.086198 0.069714 0.055184 0.042131 0.030533 0.020364 0.011086 0.002897 -0.00436 -0.01095 -0.01655 -0.02169 -0.0262 -0.03015 -0.03362 -0.03667 -0.03938 -0.04168 -0.04362 -0.04537 -0.04678 -0.04789 -0.04901

oil storage capacity (unit: L) -61.6 -64.3 -66.6 -68.9 -71.1 -73.2 -75.2 -76.9 -78.6 -80.2 -81.7 -82.9 -84.2 -85.3 -86.3 -87.3 -88 -88.8 -89.4 -89.9 -90.4 -90.8 -91.1 -91.3 -91.4 -91.5 -91.5

Error value -0.05033 -0.05091 -0.05112 -0.05132 -0.05142 -0.05144 -0.05137 -0.0511 -0.05084 -0.05052 -0.05014 -0.0496 -0.04914 -0.04858 -0.04799 -0.04743 -0.04672 -0.0461 -0.04539 -0.04467 -0.04398 -0.04326 -0.04253 -0.04178 -0.04101 -0.04027 -0.03952

oil storage capacity (unit: L) -90.9 -90.5 -90.2 -89.7 -89.2 -88.7 -88.1 -87.5 -86.8 -86.1 -85.3 -84.4 -83.6 -82.7 -81.8 -80.8 -79.7 -78.5 -77.4 -76.1 -74.9 -73.5 -72 -90.9 -90.5 -90.2

Error value -0.03654 -0.035761 -0.03505 -0.034285 -0.033548 -0.032836 -0.032111 -0.03141 -0.030697 -0.030007 -0.029306 -0.028591 -0.027933 -0.027261 -0.02661 -0.025947 -0.02527 -0.024582 -0.023944 -0.023263 -0.022631 -0.021956 -0.021269 -0.03654 -0.035761 -0.03654 -

The Deflection Identify of the Oil Storage Tank

307

Work out dispel dot chart by two groups data with MATLAB software(Figure 1). One is image of model data, anther Image of real data.

Fig. 1. Contrast image between real data and model data

From figure 1 and table 1, we know that the model curve is really close to real curve. Despite individual dots, the error is less between numeration data and practice data, the model we built is Accurate. So the model we built accord with fact.

5 Model Estimate As for problem one, we build closing polynomial model, and analyze the result’s error, and compare the model chart with real data. The error is less. As for problem two, we calculate the real data with Lagrange interpolation method, subsection linear interpolation method, thrice sample bar interpolation method. According to the calculated data, we obtain the numerical tabular of the depth of oil each 1cm when the oil tank deflected, but the calculation quantity is huge.

References 1. 2. 3. 4. 5. 6.

Wu, L., Li, B.: Mathematics experiment and modeling. National Defense Industry Press, Beijing (2008) Huadong Normal University Mathematics Department: Mathematics Analysis. Higher Education Press, Beijing (2008) Sun, H., Guan, J.: Calculate oil storage volume when the cross section is ellipse with approaching method. Tube Piece and Equipment 3, 29–31 (2001) Yang, M., Xiong, X., Lin, J.: MATLAB foundation and mathematics software. Dalian Natural Science University Press, Dalian (2006) Jiang, J., Hu, L., Tang, J.: Numerical value analysis and MATLAB experiment. Science Press, Beijing (2004) Jiang, Q., Xing, W., Xie, J., Yang, D.: University mathematics experiment. Tsinghua University Press, Beijing (2005)

PID Control of Miniature Unmanned Helicopter Yaw ∗ System Based on RBF Neural Network Yue Pan, Ping Song**, and Kejie Li Intelligent Robotics Institute, School of Mechatronical Engineering Beijing Institute of Technology 100081, Beijing, China [email protected]

Abstract. The yaw dynamics of a miniature unmanned helicopter exhibits a complex, nonlinear, time-varying and coupling dynamic behavior. In this paper, simplified yaw dynamics model of MUH in hovering or low-velocity flight mode is established. The SISO model of yaw dynamics is obtained by mechanism modeling and system identification modeling method. PID control based on RBF neural network method combines the advantages of traditional PID controller and neural network controller. It has fast response, good robustness and self-adapting ability. It is suitable to control the yaw system of MUH. Simulation results show that the control system works well with quick response, good robustness and self adaptation. Keywords: Unmanned Helicopter, RBF neural network, PID, Yaw control.

1 Introduction The design of flight control system of Miniature Unmanned Helicopter (MUH) includes modeling of yaw dynamics and control algorithm. The MUH is a very complicated system, and it has highly nonlinear, time-varying, unstable, deep coupling characteristics. Two main approaches are used for MUH modeling: laws of mechanics based modeling and system identification based modeling. Laws of mechanics models are usually large and very complicated, and some parameters are very hard to measure [1].System identification modeling can produce accurate low-order models [2, 8]. In this paper, laws of mechanics method is used to analyze the yaw system characteristic, then a SISO model of MUH in hovering or low-velocity flight mode is established using system identification method. The mostly commonly used algorithm of yaw control is PID controller [3].It based on linearizing model or assumption that the model is decoupled. But the yaw dynamics is nonlinear, time-varying and coupled model. The PID control law is limited for yaw control. Nonlinear control designs include neural network control [4, 9], fuzzy control [5], these methods are too complicated, and they need accurate dynamics model. In ∗ **

This work was supported by the National “863” Project under Grant 2007AA04Z224 of China. Corresponding author.

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 308–313, 2011. © Springer-Verlag Berlin Heidelberg 2011

PID Control of Miniature Unmanned Helicopter Yaw System

309

fact, the nonlinear SISO model based control system has the advantages: simple structure, high reliability, easy to implement. So, in this paper, the control algorithm is based on nonlinear SISO model. In order to improve robustness and self-adapting ability of traditional PID controller, Radial Basis Function (RBF) neural network algorithm is introduced to adjust the parameters of PID controller. The simulation results verify RBF-PID controller’s applicability.

2 Yaw Model of MUH 2.1 Basic Mechanics Analysis of Yaw Dynamics Usually, there is an electronic gyro on the MUH, which is a damping term in the equation for yaw moment, and it can be described using K g wb 3 . The yaw moment is [6]: T\

FT LT  K gZb 3

(1)

The yaw dynamics model for MUH in hovering or low-velocity flight mode is [6]: I zz ω& b 3 = Tψ − T MR = FT LT − K g ω b 3 − T MR = a T c T ρπ R T3 Ω T2 (

B3 B2 δψ − λ T ) LT − K g ω b 3 − T MR 3 2

(2)

Transform equation (2) [6]: I zzψ&& + K gψ& =

2 K θ& sin φ − aT cT ρπRT3 ΩT2 B 2λT LT − 2T MR aT cT ρπRT3ΩT2 B 3 LT δψ + g 3 cosθ cosφ 2 cosθ cosφ

(3)

After the Laplace transformation of equation (3), the yaw transfer function of MUH is [6]:

Gψ = I zz s 2 + K g s +

aT cT ρπRT3ΩT2 B 3 LT 3 cosθ cos φ & 2 K gθ sin φ − aT cT ρπRT3ΩT2 B 2λT LT − 2T MR

(4)

2 cos θ cos φ

2.2 Simplified Yaw Dynamics of MUH When MUH is in the hovering or low-velocity flight mode, we can assume: (1) The changing magnitude of pitch and roll angle is very small, then: sin φ ≈ 0, cos θ ≈ 1, cos φ ≈ 1

(5)

(2) The angular velocity of main rotor remains unchanged. Because of the proportional relation, the angular velocity of tail rotor Ω T remains unchanged too. (3) The velocity of MUH is zero or very small, so assumes λT remains unchanged. After finish the assumption above, the simplified yaw dynamics is:

310

Y. Pan, P. Song, and K. Li

aT cT ρπRT3 ΩT2 B 3 LT 3 I zz Gψ ( s) = K − aT cT ρπRT3ΩT2 B 2λT LT − 2T MR s2 + g s + I zz 2 I zz

(6)

In equation (6), the coefficients are constants, so, the yaw movement dynamics of MUH in hovering or low-velocity flight mode can be described using a SISO model. Using output error (OE) model can get more accurate model. Using system identification method can obtain the yaw dynamics model: Gψ ( s ) =

1.33s + 31 .08 s + 3.16 s + 29.82

(7)

2

3 PID Controller Based on RBF Neural Network Design Although PID controller has advantages of simple structure, good robustness, it will not be widely used when the characteristics of model change momentarily because of its fixed coefficients. The Radial Basis Function (RBF) neural network, which is able to approach a nonlinear function arbitrarily, can be used to identify online model with high accuracy [7]. PID controller based on RBF neural network combines PID controller and RBF neural network, having the advantages of both. A RBF network is a 3-layer feed-forward neural network. RBF neural network structure is shown in Fig.1 [7]. h1

x1 x2

h2

Σ

ym



 

xn

i



hm

j

Fig. 1. Structure of RBF neural network

Fig. 2. Structure of RBF-PID control

PID controller based on RBF neural network is constructed by RBF neural network and traditional PID controller, as shown in Fig.2, using incremental as basic PID [7]. The three inputs of PID are as follows [10]: xc(1) = error ( k ) − error ( k − 1) xc( 2) = error (k ) xc(3) = error ( k ) − 2 * error ( k − 1) + error ( k − 2)

(8)

The system average square error is [10]: E (k ) =

1 error ( k ) 2 2

(9)

PID Control of Miniature Unmanned Helicopter Yaw System

311

The adjustment of k p , k i , k d parameters adopts Gradient Descent method [10]: Δk p = −η

∂E ∂E ∂y ∂Δu ∂y xc(1) = −η = ηerror (k ) ∂k p ∂y ∂Δu ∂k p ∂Δu

Δk i = −η

∂y ∂E ∂y ∂Δu ∂E xc(2) = ηerror (k ) = −η ∂Δu ∂y ∂Δu ∂k i ∂k i

Δk d = −η

∂y ∂E ∂y ∂Δu ∂E xc(3) = ηerror (k ) = −η ∂Δu ∂y ∂Δu ∂k d ∂k d

(10)

where ∂y is Jacobian information of controlled plant. It is obtained by RBF neural ∂u network identification results.

4 Simulation The differential equations of the yaw dynamics model can be described as follows: y (k ) = 1.927 y (k − 1) − 0.9385 y (k − 2) + 0.03204 x (k − 1) − 0.02 x (k − 2)

(11)

A 3-6-1 structure of RBF network is adopted. The PID parameters k p , k i , k d are adjusted by self-learning of RBF neural network. Sampling period is 0.02s, learning rateη = 0.3 , inertia coefficient α = 0.6 , the initial value of the proportional, differential and integral coefficient are 0.005, 0.5 and 0.1 separately. To compare the results, traditional PID controller is introduced. The proportional, differential and integral coefficients are 1.2, 1 and 0 separately. The simulation time is 40s. In order to verify the anti-disturbance characteristic, disturbance to the output of the model is added. y (k ) = 1.927 y (k − 1) − 0.9385 y (k − 2) + 0.03204 x (k − 1) − 0.02 x (k − 2) + ζ (k ) (12)

Fig.3 shows the simulation results with ζ (k ) showing impulse change. k ≠ 1000 ⎧ 0 ζ (k ) = ⎨ ⎩0.05 k = 1000

(13)

Fig.4 shows the simulation results with ζ (k ) showing step change. k < 1000 ⎧ 0 ζ (k ) = ⎨ ⎩0.05 k ≥ 1000

(14)

From Fig. 3, Fig. 4, the simulation results shows that, in the condition of disturbance, the output disturbance of RBF-PID controller is smaller than that of traditional PID controller’s. And the adjustment time of RBF-PID controller is shooter. In order to test the adaptive capacity of RBF-PID controller, assumes one of the model’s parameter changes with time. y (k ) = 1.927 y (k − 1) − 0.9385 y (k − 2) + (0.03204 + δ (k )) x(k − 1) − 0.02 x(k − 2) (15)

312

Y. Pan, P. Song, and K. Li

Fig. 5 shows the simulation results with δ (k ) showing impulse change. k ≠ 1000 ⎧ 0 δ (k ) = ⎨ ⎩0.01 k = 1000

(16)

Fig. 6 shows the simulation results with δ (k ) showing step change. ⎧ 0 k < 1000 δ (k ) = ⎨ ⎩0.01 k ≥ 1000

(17)

From Fig.5, Fig.6, the simulation results show that, in the condition of model parameter varying with time, the output of RBF-PID controller is more stable than that of traditional PID controller’s, this shows the robustness of RBF-PID controller.

Fig. 3. Impulsive disturbance response

Fig. 5. Model parameter impulse variation response

Fig. 4. Stepped disturbance response

Fig. 6. Model parameter step variation response

5 Conclusion The PID controller based on RBF neural network exhibits its fast response, robustness, and adaptive ability. Compared to traditional PID controller, the RBF-PID controller has higher accuracy and stronger adaptability. For the nonlinear, time-varying, coupling, complex dynamics of yaw system of MUH, the PID controller based on RBF neural network can get satisfied control result.

PID Control of Miniature Unmanned Helicopter Yaw System

313

Acknowledgment We would like to express our gratitude to all the colleagues in our laboratory for their assistance.

References 1. Padfield, G.D.: Helicopter Flight Dynamics: The Theory and Application of Flying Qualities and Simulation Modeling. AAIA Education Series (1996) 2. Shin, D.H., Kim, H.J., Sastry, S.: Control system design for rotorcraft-based unmanned aerial vehicles using time-domain system identification. In: Proceedings of the 2000 IEEE International Conference on Control Applications (2000) 3. Shim, H.: Hierarchical flight control system synthesis for rotorcraft-based unmanned aerial vehicles. University of California, Berkeley (2000) 4. Prasad, J.V.R., Calise, A.J., Corban, J.E., Pei, Y.: Adaptive nonlinear controller synthesis and flight test evaluation on an unmanned helicopter. In: IEEE Conference on Control Application (1999) 5. Frazzoli, E., Dahleh, M.A., Feron, E.: Robust hybrid control for autonomous vehicle motion planning. IEEE Transactions on Automatic Control (2000) 6. Kim, S.K.: Modeling, identification, and trajectory planning for a model-scale helicopter. The dissertation for the degree of doctor (2001) 7. Zhang, M.-g., Wang, X.-g., Liu, M.-q.: Adaptive PID Control Based on RBF Neural Network Identification. In: IEEE International Conference on Tools with Artificial Intelligence (2005) 8. Mettler, B., Tischler, M.B., Kanade, T.: System Identification modeling of a small-scale unmanned rotorcraft for flight control design. J. Journal of the American Helicopter Society, 50–63 (2002) 9. Pallett, T.J., Ahmad, S.: Adaptive neural network control of a helicopter in vertical flight. Aerospace Control Systems 2(1), 264–268 (1993) 10. Yue, W., Feng, S., Zhang, Q.: An Auto-adaptive PID Control Method Based on RBF Neural Network. In: International Conference on Advanced Computer Theory and Engineer (ICACTE) (2010)

Identity-Based Inter-domain Authentication Scheme in Pervasive Computing Environments Shi-Wei Huo, Chang-Yuan Luo, and Hong-Zhi Xin Information Engineering University, Electronic Technology Institute 450004 Zhengzhou, China [email protected]

Abstract. An Identity-based signature scheme is proposed based on additive elliptic curve group. The verification result of the signature is a constant with respect to the signer’s identifier. Then an inter-domain authentication scheme is constructed by combining the proposed signature scheme. During the authentication, a user constructs the signature of timestamp as authentication proof, which realizes secure inter-domain authentication and user anonymity. It is showed that the proposed scheme has superiority in both security and efficiency, and is more suitable for pervasive computing. Keywords: pervasive computing; inter-domain authentication; Identity-based cryptography; anonymity.

1 Introduction In a pervasive environment, mobile users often roam into foreign domains to request service. Hence, efficient and secure inter-domain authentication should be highly emphasized [1]. When users are roaming into a foreign domain, there is no trust between users and the foreign authentication server(FA), so the FA should cooperate with the users’ home authentication server(HA) to authenticate users. During interdomain authentication, user’s real identity should be concealed in order to prevent user’s sessions being tracked by malice. Besides of mutual authentication and key establishment, the inter-domain authentication scheme for pervasive computing should meet the following requirements: (1) Client anonymity: the real identity of a user should not be divulged to the FA and outsiders; (2) Non-linkability: the outsiders could not link different sessions to the same user. Lin [1] proposed an inter-domain authentication protocol based on signcryption. But it can not realize user anonymity. Peng [2] proposed an Identity-based(ID-based) inter-domain authentication scheme. It realizes anonymity and non-linkability, but it has high computation expense because users require to implement expensive bilinear pairing operations. Zhu [3] proposed a novel ID-based inter-domain authentication scheme, which reduces the count of pairing operations and has higher efficiency. However, it has drawback in anonymity. The user uses the same temporary certificate as authentication proof in the reauthentication phase, which causes user’s sessions to be tracked. R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 314–320, 2011. © Springer-Verlag Berlin Heidelberg 2011

ID-Based Inter-domain Authentication Scheme in Pervasive Computing Environments

315

This paper first presents a new ID-based signature(IBS) scheme. Then, an interdomain authentication scheme is constructed based on the new IBS scheme. It is showed that the scheme can achieve the security requirements in inter-domain authentication of pervasive computing and has higher efficiency.

2 Identity-Based Signature Scheme Bellare proposed an IBS scheme without pairing called BNN-IBS [4] with provable security. It has high efficiency. On the basis of BNN-IBS, this paper proposes a novel IBS scheme. The verification result of the signature is a constant with respect to the signer’s identity. The scheme is implemented as follows: Setup: Given the security parameter k, PKG takes the following steps: (1) Let G1 be an order-q cyclic additive group of the points on an elliptic curve E ( Fp ) with generator P. Select a system secret key s ∈ Z q* and set the system public

key Ppub = sP ∈ G1 . (2) Choose two cryptographic hash functions: H1 :{0,1}* × G1 → Z q* , H 2 :{0,1}* → Z q* . (3) Publish system parameters {G1 , q, P, Ppub , H1 , H 2} , and keep s secret. User-Key Extraction: Suppose IDA ∈ {0,1}* denotes user A’s unique identifier. PKG generates A’s private key as follows: (1) Choose at random rA ∈ Z q* and compute RA = rA P . (2) Compute s A = rA + sc , where c = H1 ( IDA , RA ) . A’s private key is the pair ( s A , RA ) , and it is sent to A via a secure channel. Signature Generation: A signs a message m ∈ {0,1}* with ( s A , RA ) as follows: (1) Choose at random y ∈ zq* , and compute Y = yP . (2) Compute z = y + s Ah , where h = H 2 ( IDA , m, RA , Y ) . Then A’s signature on m is the tuple ( RA , Y , z ) . Signature Verification: Given ( RA , Y , z ) , IDA and message m , a verifier checks the signature as follows: (1) Compute c = H1 ( IDA , RA ) and h = H 2 ( IDA , m, RA , Y ) . (2) Check whether the equality h −1 ( zP − Y ) = RA + cPpub holds. The signature is accepted if the answer is yes and rejected otherwise. In the above IBS scheme, given a user’s signature ( RA , Y , z ) , the verification result RA + cPpub is a constant with respect to the user’s identifier. This paper will realize user anonymity based on this property. The Setup, User-key Extraction and Signature Generation algorithms are the same as those of BNN-IBS. Only the Signature Verification algorithm is changed. The

316

S.-W. Huo, C.-Y. Luo, and H.-Z. Xin

equality zP = Y + h( RA + cPpub ) is changed to h −1 ( zP − Y ) = RA + cPpub . This will not affect the security of the former scheme, so the proposed scheme is secure.

3 Identity-Based Inter-domain Authentication Scheme 3.1 System Architecture The system architecture is shown as the Fig.1. There are two trust domains: Domain A and Domain B. User A are located in the domain A. HA and FA indicate the authentication server of Domain A and B respectively. A should first register in HA. When A wants to access resource in Domain B, FA must cooperate with HA to authenticate A. The security of our scheme relies on the following assumption: HA and FA are honest and they trust each other.

Fig. 1. System architecture

3.2 System Initialization HA chooses the system parameters as described in Section 2, and determines its public/private key pair ( sHA , PHA ) . HA chooses cryptographic hash functions: H1 :{0,1}* × G1 → Z q* , H 2 :{0,1}* → Z q* , H 3 : G1 → {0,1}* , and publishes system parameters {G1 , q, P, PHA , H1 , H 2 , H 3} . FA chooses the same parameters and hash functions, determines its public/private key pair ( s FA , PFA ) , and publishes system parameters {G1 , q, P, PFA , H1 , H 2 , H 3} .

3.3 Registration A sends the identifier IDA to HA. HA checks the validity of IDA . Then HA generates A’s private key ( s A , RA ) , where s A = rA + sHAc , c = H1 ( IDA , RA ) , rA ∈ Z q* . HA sends ( s A , RA ) to A. HA creates for A an account of the form < Ind A , IDA , RA > , where

Ind A = RA + H 1 ( IDA , RA ) PHA is the index of A’s account.

ID-Based Inter-domain Authentication Scheme in Pervasive Computing Environments

317

3.4 Authentication When A requests resource in Domain B for the first time, A needs to implement the authentication protocol.

，

Step1: A chooses at random x y ∈ zq* , and compute Y = yP , X = xP , Y ' = Y + xPHA . A picks up the current time TA , and compute h = H 2 ( IDA , TA , RA , Y ) , z = y + s Ah . A sends a message < IDHA , TA , X , Y ' , h, z > to the FA. Step2: After receiving the message, FA checks the validity of TA . FA rejects the request if TA is not valid. Otherwise, FA does the following: (1)

Pick

up

the

current

time

TFA

,

and

construct

a

message

mFA = {IDHA , TA , X , Y ' , h, z, IDFA , TFA} . Then, compute signature Sig FA (mFA ) , where Sig ()

denotes the elliptic curve digital signature algorithm(ECDSA). (2) Send a message < mFA , Sig FA (mFA ) > to the HA. Step3: After receiving the message, HA checks the validity of TFA and Sig FA (mFA ) . If the decision is positive, HA confirms FA is legal and does the following: (1) Compute Y = Y ' − sHA X , Ind A = h −1 ( zP − Y ) , and search client accounts with Ind A . If there is an account indexed by Ind A , obtain the corresponding identity information and check whether the equality h = H 2 ( IDA , TA , RA , Y ) holds. If the decision is positive, HA confirms A is a legal user. (2) Pick up the current time THA , and compute k = H 3 (Y ) . Then, construct a message mHA = {IDFA , IDHA , TFA , THA , EPECC (k )} , and compute signature Sig HA (mHA ) , where FA

EPECC denotes the elliptic curve encryption scheme(ECES). FA

(3) Send a message < mHA , Sig HA ( mHA ) > to the FA. Step4: After receiving the message, FA checks the validity of THA and Sig HA (mHA ) . If the decision is positive, FA confirms HA and A are legal and does the following: (1) Generate a temporary identifier IDA' and corresponding private key (s 'A , RA' ) for A. (2) Create for A an account of the form ( Ind A' , IDA' , RA' , time) , where Ind A' = RA' + H1 ( IDA' , RA' ) PFA is the index of the account and

time is the expiry date.

IDA'

and (s 'A , RA' ) can be generated in spare time. (3) Decrypt EPECC (k ) and get k . Then, pick up the current time TFA' , and send a FA

message < TFA' , Ek (TFA' , TA , s A' , RA' , IDA' ) > to A, where E is the symmetric encryption algorithm. Step5: After receiving the message, A checks the validity of TFA' . If the decision is positive, A computes k = H 3 (Y ) , and decrypts Ek (TFA' , TA , s A' , RA' , IDA' ) . Then, A checks the validity of TFA' and TA . If the decision is positive, A has known that FA is legal. A

318

S.-W. Huo, C.-Y. Luo, and H.-Z. Xin

uses k as the session key with the FA in future communications, and saves IDA' , ( s 'A , RA' ) . 3.5 Reauthentication When user A requests resource in Domain B again before the expiry date, A can implement the reauthentication protocol. In this case, the FA can fast authenticate A without the participation of HA.

，

Step1: A chooses at random x y ∈ zq* , and computes Y = yP , X = xP , Y ' = Y + xPFA . A picks up the current time TA , and computes h = H 2 ( IDA' , TA , RA' , Y ) , z = y + s 'Ah . A sends a message < TA , X , Y ' , h, z > to FA.. Step2: After receiving the message, FA checks the validity of TA . If the decision is positive, FA does the following: (1) Compute Y = Y ' − sFA X , Ind A' = h −1 ( zP − Y ) , and search client accounts with

Ind A' . If there is an account indexed by

Ind A' , obtain the corresponding identity in-

formation and check whether the equality h = H 2 ( IDA' , TA , RA' , Y ) holds. If the decision is positive, FA has known that A is a legal user. (2) Compute the session key k ' = H 3 (Y ) . Pick up the current time TFA , and send a message < TFA , Ek (TFA , TA ) > to A. '

Step3: After receiving the message, A checks the validity of TFA . If the decision is positive, A computes k ' = H 3 (Y ) , and decrypts Ek (TFA , TA ) . Then, A checks the valid'

ity of TFA and TA . If the decision is positive, A confirms FA is legal. A saves k ' as the session key with FA.

4 Security Analysis The proposed scheme can achieve the following security requirements: Mutual Authentication: In the authentication phase, entities can authenticate each other. In step 3, HA authenticates A through verifying the signature (Y ' , h, z ) . (Y ' , h, z ) is the signature over timestamp TA using the IBS scheme in section2. We

encrypt Y and get Y ' , for in this case only HA can verify the signature. Since the IBS scheme is secure and a timestamp is used for checking the freshness of signature, the authentication is secure. In step 3 and 4, HA and FA authenticate each other through verifying the other’s signature. Since the ECDSA is secure and a timestamp is used for checking the freshness of signature, the authentication is secure. In step 5, A authenticates FA through decrypting Ek (TFA' , TA , s A' , RA' , IDA' ) and checking TFA' , TA . Because HA encrypts k using FA’s public key and sends the EPECC (k ) to FA, only legal FA

FA can decrypt EPECC (k ) .In step 4, FA trusts A because FA trusts HA and HA has FA

ID-Based Inter-domain Authentication Scheme in Pervasive Computing Environments

319

authenticated A. The trust relation is established through HA. In the reauthentication phase, FA and A can also authenticate each other. The process is similar to above. Secure session key establishment: In the authentication phase, FA and A can establish session key k = H 3 (Y ) . Since only HA can compute Y from Y ' with its private key, only A and HA can compute k . HA encrypts k using FA’s public key and sends the EPECC (k ) to FA, so only the legal FA can decrypt EPECC (k ) . In step 5, A can confirm FA

FA

that FA indeed gets k through decrypting Ek (TFA' , TA , s A' , RA' , IDA' ) and checking TFA' , TA . In the reauthentication phase, FA and A can renew the session key. Client anonymity: Any outsider and FA is unable to confirm user A’s real identity. In the authentication phase, the authentication information A submits only contains TA and its signature without any identity information. Only HA can compute the index of A’s account, so any outsider and FA does not know A’s real identity. In the reauthentication phase, FA can confirm A’s temporary identifier, but does not know A’s real identity. Non-Linkability: Any outsider is unable to link two different sessions to the same user. In the authentication phase, the authentication proof is the signature over timestamp TA , so there is no linkability between different proofs and outsiders can not link different proofs to the same user. Similarly, in the reauthentication phase, nonlinkability is achieved. Our scheme can achieve the same security as the scheme in [2]. The scheme in [3] has drawback in anonymity. In the reauthentication phase, the user uses the same temporary certificate as authentication proof, which causes user’s sessions to be tracked. So our scheme has superiority in security compared with the scheme in [3].

5 Performance Analysis In this section, we compare the performance of our scheme in the authentication phase with schemes in [2,3], for they are all identity-based. In the comparison, only the time of public key operations are accounted. We suppose that the hardware platform of HA and FA is a PIV 2.1-GHZ processor, and the hardware platform of A is a 206-MHz StrongARM processor. The operation time of cryptographic primitives on the HA/FA and A are obtained by experiment [5]. The time is listed in Table 1. Table 1. Cryptographic operation time(ms)

HA/FA A

ECC Sign 3.3 33.6

ECC Verification 6.3 64.2

ECC Encryption 6.6 67.3

ECC Decryption 4.7 47.9

scalar

pairing

3.8 38.7

12.6 128.4

According to the date in Table 1, the running time of the three schemes is computed. The running time is listed in Table 2.

320

S.-W. Huo, C.-Y. Luo, and H.-Z. Xin Table 2. Performance comparison

HA(ms) FA(ms) A(ms)

Our scheme 27.6 14.3 116.1

Scheme in [2] 50.4 50.4 513.6

Scheme in [3] 29.8 13.4 193.5

The above results show that our scheme reduces the overall running time and the client’s running time. The reason is that the proposed protocol uses the new IBS scheme without pairing, and reduces the count of scalar on the client side.

6 Conclusion This paper has presented an ID-based inter-domain authentication scheme in pervasive environments. It can achieve such security requirements as mutual authentication, secure session key establishment, client anonymity and non-linkability. It has superiority in both security and efficiency, and is more suitable for pervasive computing.

References 1.

2. 3. 4.

5.

Yao, L., Wang, L., Kong, X.W.: An inter-domain authentication scheme for pervasive computing environment. J. Computers and Mathematics with Applications 59(2), 811–821 (2010) Peng, H.X.: An identity based authentication model for multi-domain. J. Chinese Journal of Computers 29(8), 1271–1281 (2006) Zhu, H., Li, H., Su, W.L.: ID-based wireless authentication scheme with anonymity. J. Journal on Communications 30(4), 130–136 (2009) Zhu, R.W., Yang, G.M., Duncan, S., Wong, D.S.: An efficient identity-based key exchange protocol with KGS forward secrecy for low-power devices. J. Theoretical Computer Science 378, 198–207 (2007) Cao, X.F., Zeng, X.W., Kou, W.D.: Identity-based Anonymous Remote Authentication for Value-added Services in Mobile Networks. J. IEEE Transactions on Vehicular Technology 58(7), 3508–3517 (2009)

Computer Simulation of Blast Wall Protection under Methane-Air Explosion on an Offshore Platform Changjian Wang1,3, Weigang Yan1, Jin Guo1, and Changming Guo2,3 1

State Key Laboratory of Fire Science, University of Science and Technology of China, Hefei, Ahhui, 230026, P.R. China 2 Department of Modern Mechanics, University of Science and Technology of China, Hefei, Ahhui, 230026, P.R. China 3 State Key Laboratory of Explosion Science and Technology, Beijing Institute of Technology, Beijing 100081, China

Abstract. An in-house explosion program is presented to evaluate blast wall protection under Methane-Air Explosion on an offshore platform, based on twodimensional, time-dependent, reactive Navier–Stokes equations including the effects of viscosity, thermal conduction and molecular diffusion. The results show that this program can successfully produce explosion process of methaneair gas cloud. Because the overpressure behind the blast wall and on the lifeboat plate is more than 1.0atm when explosion wave passes, the current blast wall is not enough to keep the person and lifeboat safe. So the blast wall needs to be redesigned. The explosion wave of methane-air gas cloud undergoes a successive process of detonation formation, detonation transmission, shock attenuation, regular reflection and Mach reflection etc. Additionally, due to high overpressure generated in gas cloud explosion, it is extremely devastating and must be avoided at all times on offshore platform. Keywords: gas cloud explosion, offshore platform, blast wall.

1 Introduction On an offshore platform, if combustible gas leaks, large gas cloud possibly forms and further leads to an accidental explosion. In some cases, the explosions will take place with the transition from deflagration to detonation. The relatively high overpressure generated in explosion will bring serious damage and potential life loss. For example, in 1988, the extraordinarily serious explosion accidence of one British oceanic offshore platform occurred. More than one hundred people lost their lives and this platform was finally discarded as uselessness [1]. So the offshore industries have spent considerable efforts in the qualification of explosion threats, the quantification of blast over-pressures and further designing against them. Since experiments can be very expensive, the computer simulation becomes very popular to evaluate the offshore explosion cases. Moreover, more scenarios can be considered for the evaluation of explosion possibility. TNO used the multi-energy model for rapid assessment of explosion overpressure [2]. UKOOA [3] employed CFD models in offshore explosion assessment and concluded that CFD models can R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 321–326, 2011. © Springer-Verlag Berlin Heidelberg 2011

322

C. Wang et al.

predict reasonably good explosion evolution and overpressure. Raman and Grillo[4] gave guidance on the application of the TNO multi-energy method and on the selection of model parameters related with the equipment layout and level of congestion. Clutter and Mathis [5] simulated vapor explosions in offshore rigs using a flamespeed based combustion model. Pula et al. [6] adopted a grid-based approach and an enhanced onsite ignition model to simulate the offshore explosion overpressure. On an offshore platform for methane exploitation in China South Sea, the blast walls were designed to separate the process areas from living quarters and life boats. In these process areas, some potential failure possibly leads to methane leakage and further methane-air cloud explosion. In this paper, we employed computer modeling to evaluate whether the blast wall is enough to protect the person and lifeboat.

2 Computation Set-Up 2.1 Governing Equations and Numerical Methods The appropriate model for offshore explosion in methane-air gas cloud is the NavierStokes equations for multiple thermally perfect species with reactive source terms, as described as follows: ∂Q ∂F ∂G 1 ⎛ ∂Fv ∂Gv ⎜ + = + + ∂y ∂x ∂y Re ⎜⎝ ∂x ∂t

⎞ ⎟+S ⎟ ⎠

(1)

where Q = {ρ1, ρ 2 ......ρ NS , ρu, ρv, ρE}T

, F = {ρ1u, ρ 2u,......ρ NS u, ρu 2 + p, ρuv, ( ρE + p)u}T ,

G = {ρ1v, ρ 2 v,......ρ NS v, ρuv, ρv 2 + p, ( ρE + p )v}T , T ∂Y ∂Y ∂Y ⎧ ⎫ Fv = ⎨ ρD 1 , ρD 2 ......ρD NS ,τ xx ,τ yx , uτ xx + vτ yx + wτ zx + q x ⎬ ∂x ∂x ∂x ⎩ ⎭

,

T

⎧ ⎫ ∂Y ∂Y ∂Y Gv = ⎨ ρD 1 , ρD 2 ......ρD NS ,τ xy ,τ yy , uτ xy + vτ yy + vτ zy + q y ⎬ ∂ y ∂ y ∂ y ⎩ ⎭ 2 ⎛ ∂u

∂v

∂w ⎞

, 2 ⎛ ∂v

∂w

∂u ⎞

S = {S1, S 2 ......S NS ,0,0,0,0}T , τ xx = μ ⎜⎜ 2 − − ⎟⎟ , τ yy = μ ⎜⎜ 2 − − ⎟⎟ , 3 ⎝ ∂x ∂y ∂z ⎠ 3 ⎝ ∂y ∂z ∂x ⎠ ∂y ∂y ⎛ ∂u ∂v ⎞ ∂θ NS ∂θ NS τ xy = τ yx = μ ⎜⎜ + ⎟⎟ , q x = k + ∑ ρDi hi i , q y = k + ∑ ρDi hi i ∂ y ∂ x ∂ x ∂ x ∂ y ∂y ⎝ ⎠ i =1 i =1

where ρ i is the i-th species density. The ratio Yi = ρ i / ρ denotes the i-th species mass fraction. NS is the species number. u and v denote the velocity components in x and y direction and p , θ and E denote pressure, temperature and total energy per

unit mass, respectively, while S i is the production rate of the i-th species. hi denotes the i-th species enthalpy per unit mass,κthe thermal conduction coefficient, Di the diffusion coefficient of the i-th species, respectively. μ = μ t + μ l , where μ t and

μ l are the turbulent and laminar viscosity coefficient, respectively. The standard k-εturbulence model [7] is used for describing the turbulent effect of methane-air gas cloud explosion in this paper. Single step chemistry is taken into

Computer Simulation of Blast Wall Protection under Methane-Air Explosion

323

account, involving reacting species of CH4, O2, CO2, H2O and N2. The second-order additive semi-implicit Runge-Kutta Method [8] was employed to discretize the time term and treat the stiffness of chemical source terms. The convective terms were integrated by 5th Weighted Essentially Non-Oscillatory (WENO) scheme [9]. The viscous, heat and diffusion terms were evaluated using second-order central finite difference. 2.2 Physical Model and Related Computation Conditions

Fig.1 presents a local diagram of blast wall. The blast wall height is 2.5m. At the left of it is gas turbine generator skid (GTGS) with 15m length, 3.2m width and 10.5m height. Here the leakage of methane gas at high pressure possibly occurs and the resulting methane-air gas cloud explosion emerges. At the right of blast wall are two lifeboats which are used to carry the workers to escape from the offshore platform in danger cases. So, current simulation aims at evaluating whether the blast wall is enough to shelter the persons and lifeboats from explosion hazards. Due to relatively larger computation scale and chemical reaction stiffness, a 2D simulation was carried out to elucidate the above problem, as shown in Fig.2. The computed domain extends 18m in X direction and 20m in Y direction. Below the height less than 10.5m in Y direction is the wall of GTGS. The lift boat plate is 7.5m away from the blast wall. Its height and width are 2.7m, 2.7m respectively. The pressure outlet bound conditions were imposed on the right and top boundaries, and the left boundary which height is more than 10.5m. A leakage with the methane amount of 0.14kg was considered and

Fig. 1. Local diagram of blast wall

(a) Explosion close to the bottom of GTGS

(b) Explosion close to the top of GTGS

Fig. 2. Schematic of computed domain

324

C. Wang et al.

located close to the bottom or top of the GTGS. So a semi-circular methane-air cloud formed with a diameter of 2m. At its center, a ignition with 4 times Chapman-Jouguet (CJ) pressure and 1 times CJ temperature was imposed, in order to generate a detonation wave in gas cloud. The diameter of this ignition region is 0.2m. Actually, such explosion was considered as the most serious case. However this explosion with current leakage is very weaker than that of large leakage.

3 Results and Discussion Fig.3 presents explosion evolution when the leakage occurs close to the bottom of GTGS. In Fig.3, an ignition takes place with high pressure and temperature, which leads to the generation of a strong shock wave. As it propagates in fresh methane-air gas cloud, due to the existence of methane-air induction time, the chemical reaction can not release the heat immediately to support the shock wave. So it degenerates. As shown in Fig.3 (a), the maximum pressure is about 8atm, much lower than CJ pressure of 17.1 atm predicted by Gordon-Mcbride code. Additionally, due to arc shock wave propagating outwards, the pressure in the central region of gas cloud steeply decreases. In order to match the low pressure in the central region with high pressure behind the shock wave, a secondary wave forms. So a pressure band can be evidently observed. As the heat release in fresh gas cloud proceeds, the chemical reaction zone catches up with the degenerated shock wave, strengthens and further couples with it. A methane-air detonation wave forms, as presented in Fig.3(b), and its maximum pressure is about 18.5atm, a little more than CJ pressure. In Fig.3(c), as the detonation wave transmits from the methane-air cloud to air, due to the loss of chemical reaction support, the transmitted shock wave decays. Since the gas cloud is very close to the ground, the precursor shock wave first reflects in normal means, and then undergoes the transition from regular reflection to Mach reflection. In Fig.3(d), the pressure behind the reflected shock or Mach stem is relatively high and more than CJ pressure. When the shock wave impinges upon the front surface of blast wall, the regular reflection occurs since the angle between the shock front and blast wall does not satisfy the Mach reflection relation. In Fig.3(e), the pressure behind the reflected shock is about 38atm. When the shock wave diffracts around the top of blast wall, due to the effects of rarefaction waves emitted from the right top corner, the shock wave decays steeply from 38atm to 18.5atm, as shown in Fig.3 (f). It travels at only several atmosphere pressure when it reaches the ground. In Fig.3(g), another transition from regular to Mach reflection occurs. However, close to the ground, the maximum pressure is only 4 to 5atm. When it climbs up to the lifeboat plate, the shock wave diffracts again. In Fig.3(h), the maximum pressure is still more than 2atm. Fig.4 presents explosion evolution when the leakage occurs close to the top of GTGS. Compared with Fig.3, Fig.4 has some special characteristics. Firstly, when the detonation wave transmits from methane-air gas cloud to air, the top part of gas cloud sphere is influenced by free boundary whereas the bottom part interacts with wall. Therefore, pressure attenuation on the top part is much more evident than on the bottom, as shown in Fig.4 (c) and (d). Secondly, in Fig.4 (e), the region at highest pressure in gas cloud interacts with the ground close to the bottom of GTGS, and shock wave front at middle pressure collides with the blast wall. Thirdly, in Fig.4 (f), shock wave diffracts around the blast wall, the pressure near the front surface is much higher

Computer Simulation of Blast Wall Protection under Methane-Air Explosion

325

than that near the back surface and about 5 to 10 times. Fourthly, on the ground behind the blast wall, the shock wave is normally reflected, and not Mach-reflected in Fig.4 (g). Lastly, in Fig.4 (h), the shock wave with the identical pressure as shown in Fig.3(h) exists on lifeboat plate. According to the general criterion, if explosion overpressure is 0.1atm, the bridges and lifeboats are impaired. If the overpressure is 1.0atm, the explosion wave leads to person death due to lung and ear damage. In current cases, close to the ground, blast wall and lifeboat plate, the overpressure values are always more than 1atm. At local place, it is more than 10 to 30atm. So the current blast wall is not enough to keep the person and lifeboat safe. Additionally, if the leakage increases more, the more

(a) t=1ms

(e) t=7.5ms

(b) t=1.6ms

(c) t=1.8ms

(f) t=11ms

(g) t=21.0ms

(d) t=6.0ms

(h) t=30.5ms

Fig. 3. Explosion evolution when the leakage occurs close to the bottom of GTGS

(a) t=1ms

(e) t=14ms

(b) t=2.5ms

(f) t=18ms

(c) t=5ms

(d) t=12ms

(g) t=26ms

(h) t=30ms

Fig. 4. Explosion evolution when the leakage occurs close to the top of GTGS

326

C. Wang et al.

dangerous cases will be faced. That is to say, the blast wall needs to be re-designed. It is suggested that an elevated height be needed or the blast wall shape be changed. This will be re-evaluated according to the further design.

4 Conclusions The computer simulations described here provide detail descriptions of the blast wall protection under the methane-air gas cloud explosion. The main conclusions can be drawn as follows: (1) The current computer program can successfully simulate gas cloud explosion. (2) Because the overpressure behind the blast wall and on the lifeboat plate is more than 1.0atm when explosion wave passes, the current blast wall is not enough to keep the person and lifeboat safe. So the blast wall needs to be re-designed. (3) The explosion wave of methane-air gas cloud undergoes a successive process of detonation formation, detonation transmission, shock attenuation, regular reflection and Mach reflection etc. (4) The maximum pressure for methane-air cloud detonation wave is about 18.5atm. Additionally, the shock wave reflects on a wall, and the local pressure is possibly more than two times the pressure behind incident shock wave. So it is extremely devastating and must be avoided at all times on offshore platform. This project was supported by grants from the Ph.D Programs Foundation of Ministry of Education of China (Grant No. 20070358072) and Open Foundation of State Key Laboratory of Explosion Science and Technology of Beijing Institute of Technology (Grant No. KFJJ06-2).

References 1.

2. 3. 4. 5.

6. 7. 8. 9.

The public Inquiry into the piper Alpha disaster. The Hon Lord Cullen, presented to parliament by the secretary of the sate for energy by command of Her Majesty, Department of the energy, London, HMSO (November 1990) The Netherlands Organization for Applied Scientific Research (TNO), Internet website (2004), http://www.tno.com.nl UKOOA (UK Offshore Operators’ Association), Fire and explosion guidance. Part 1: Avoidance and mitigation of explosions, Issue 1 (October 2003) Raman, R., Grillo, P.: Minimizing uncertainty in vapor cloud explosion modeling. Process Safety and Environmental Protection 83(B4), 298–306 (2005) Clutter, J.K., Mathis, J.: Computational Modeling of Vapor Cloud Explosions in Off-shore Rigs Using a Flame-speed Based Combustion Model. Journal of Loss Prevention in the Process Industries 15, 391–401 (2002) Pula, R., Khan, F.I., Veitch, B., Amyotte, P.R.: A Grid Based Approach for Fire and Explosion. Process Safety and Environmental Protection 84(B2), 79—91 (2006) Anderson, W.K., Thomas, J.L., Van Leer, B.: AIAA Journal 26(9), 1061–1069 (1986) Zhong, X.L.: Additive semi-implicit Runge-Kutta Methods for computing high-speed noneqilibrium reactive flows. Journal of Computational Physics 128, 19–31 (1996) Shu, C.W.: Essentially Non-Oscillatory and Weighted Essentially Non-Oscillatory Schemes for Hyperbolic Conservation Laws. ICASE Report 97-65 (1997)

Throughput Analysis of Discrete-Time Non-persistent CSMA with Monitoring in Internet of Things Hongwei Ding, Dongfeng Zhao, and Yifan Zhao Department of Communication Engineering, Yunnan University, No. 2 North Green Lake, 650091, Kunming, Yunnan, China [email protected]

Abstract. With the development of Internet of Things industry, more and more scholars start to study in the field of the Internet of things. The monitoring of the transmission state of information is one of the important fields of research in Internet of Things. This paper uses the discrete-time non-persistent CSMA with monitoring function random access mode to realize the monitoring features of the transmission state of information in the Internet of Things. And we get to the throughput of the system using the average cycle analysis method, through computer simulations to verify the correctness of the analysis. Keywords: Internet of Things, discrete-time, non-persistent CSMA, monitoring, throughput.

1 Introduction Internet of Things is defined as: The radio frequency identification (RFID), infrared sensors, global positioning systems, laser scanners and other information sensing device, according to the agreement agreed to anything connected to the Internet, the information exchange and communication, in order to achieve intelligent identify, locate, track, monitor and manage a network. Internet of things is the material objects connected to the Internet. This has two meanings: First, things are still the core and foundation of the Internet, is based on the Internet extension and expansion of the network. The second, any goods can intelligently exchange information and communications with the other goods [1]. Internet of Things broke the tradition of thinking before. Past ideas have been the physical infrastructure and IT infrastructure separately: one is the airport, roads, buildings, and the other is the data center, personal computers, broadband and so on. In the "Internet of Things" era, reinforced concrete, cable and the chip integrated into a unified broadband infrastructure, in this sense, infrastructure is more like a new earth site, the operation of the world were in on it, which including economic management, social management, production operation and even personal life[2]. In Internet of Things the persons and things, things and things as the platform for the access way into polling multiple access methods and random multiple access method. Among them, the random access method is divided into discrete time random multiple access methods and continuous-time random multiple access system access R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 327–331, 2011. © Springer-Verlag Berlin Heidelberg 2011

328

H. Ding, D. Zhao, and Y. Zhao

methods. This paper will uses the continuous-time non-persistent CSMA with monitoring function random multiple access system to achieve "automation" and "intelligent" feature, first asked that the system has the availability of a client feedback to the sender of information, monitoring functions. Enabling intelligent identification, positioning, remote monitoring, network status tracking, fault alarm, automatic measurement and control functions. In the Internet of Things, human and machine or between machines and machines must be achieved: the machine can take the initiative to report information to the people of the state information during transmission to achieve fault alarm, the system can also remotely monitor. Between the machine and the machine can automatically communicate with the data exchange, automatic measurement and control, data acquisition and transmission, status tracking, etc. Remote monitoring, status tracking, fault alarm, automatic measurement and control functions necessary requirement for the realization of the receiving end system with information feedback to the sending end, monitoring functions. This paper will uses the discrete-time non-persistent CSMA with monitoring function random multiple access mode to achieve monitoring feature. It asks that the system has the availability of a client feedback to the sender of information, monitoring functions[3].

2 Discrete-Time Non-persistent CSMA with the Monitoring The rules of non-persistent CSMA are when the channel is busy, wait a period of time, and then listen. When the channel is idle, sent immediately. In the random access system model of discrete-time non-persistent CSMA with the monitoring, we assume that the time axis is a discrete classification system, when the terminal wants to send information packets to the start time slot immediately following the channel to send messages to the packet, there are three-channel species status appears that the information packet was successfully transmitted, the information packet collision, and idle channel conditions. Our random access system with monitoring methods on the channel, the first packet of information need to monitor the situation and was successfully transferred the case of a collision [4,5]. When the information packet is successfully received, the receiver feedback from a confirmation notice to the sender information packet has been successfully received, the sender information and subject to the delay a in the channel can receive a confirmation message. When the information packet collision, the receiver feedback to the feedback that a collision inform the sender is not receiving the information packet collision, the same, the sender information in the channel after a time delay a to receive the collision feedback. Therefore, the successful transmission of information packets with the confirmation message appears at the same time; information packet collisions and collision concurrent feedback. Assuming the time length of the information packet is a unit of length of time in our random access systems, the sender and receiver of information between the delay is a ; The time length of the confirmed information of Successful transmission packet is a . The time length of the collision feedback information is a also. Time axis set by a to be divided, the system users are not limited the number, any user wants to send a message packet, it must start in the division of time slots to send.

Throughput Analysis of Discrete-Time Non-persistent CSMA with Monitoring

329

t

a

1+2a a

1+2a Idle

Busy Successful packet

a

Busy

Collision packet

Successful transmission information

Collision feedback information

Fig. 1. The discrete-time non-persistent CSMA random access system with monitoring function

3 Analysis of Discrete-Time Non-persistent CSMA with Monitoring Assuming probability of no packet arrival during transmission information time in the busy period, the probability of no packet arrival within a short time slot,

p0 = e − aG .

(1)

The probability distribution of k short slots in an idle period,

p( N = k ) = p0 k −1 (1 − p0 ) .

(2)

In a cycle period, the number of idle short slots, the average number of short slots included in an idle period

N=

1 1 = . 1 − p0 1 − e − aG

(3)

Therefore, the average length of idle period

I BU = aN =

a . 1 − e − aG

(4)

If only one packet arrival in the last short slot in idle period, then the group will be the beginning of the next short slot was successfully sent. In the literature[6], the average length of that information packet was successfully sent

U BU =

aGe− aG . 1 − e − aG

(5)

The average length of that information packet was successfully sent and information packet collision time in the channel

E[ BU ] = 1 + 2a .

(6)

330

H. Ding, D. Zhao, and Y. Zhao

So, the throughput of the system is

aGe − aG E[U BU ] aGe− aG 1 − e− aG S= = = . E[ BU ] + E[ I BU ] 1 + 2a + a 1 + 3a − (1 + 2a)e− aG 1 − e− aG

(7)

4 Simulation In the discrete-time non-persistent CSMA with monitoring function random multiple access system based on the analysis, we conducted a computer simulation[7,8], theoretical calculation and computer simulation using the same parameter values were the result of taking a = 0.1, as Fig.2. shown. 0.6 0.5 0.4

Theory

0.3

Simulation

0.2 0.1 0 0.01 0.4

0.9 1.4

1.9 2.6

3.5

7

20

G

Fig. 2. The relationship between throughput and arrival rate

5 Conclusions The simulation results have been verified to agree well with the theoretical results in Fig. 2 The discrete-time non-persistent CSMA with monitoring function random access mode be used, we can be more easily achieved the monitoring features of the transmission state of information in the Internet of Things. Throughput analysis of the discrete-time non-persistent CSMA with monitoring function random access system laid a good foundation for more in-depth understanding of the system. Throughput analysis of the system will help us to find ways to optimize.

．

References 1. 2.

Yao, W.: Basic Content of the Internet of Things. China Information Times 5, 1–3 (2010) Li, Y., Chen, H.: Pondering on Internet of Things. Value Engineering (08), 126–127 (2010)

Throughput Analysis of Discrete-Time Non-persistent CSMA with Monitoring 3.

4. 5. 6.

7. 8.

331

Kleinrock, L., Tobagi, F.A.: Packet Switching in Radio Channels: Part 1 – Carrier Sense Multiple-Access Midland Their Throughput-Delay Characteristics. IEEE Transactions on Communications 23(12), 1400–1416 (1975) Liao, B.: The Softswitch-based Personal Monitoring Communications. ZTE Communications (04), 47–50 (2006) Hu, X., Zhou, L.: Research on Signaling Monitor System of Switch Soft Networks. Telecommunications Science (01), 34–37 (2007) Zhao, D.: Study on a New Analyzing Method for Random Access Channel. In: Second International Conference and Exhibition on Information Infrastructure, Beijing, April, 1998, pp. 16–29 (1998) Zhao, D.: Study on Analyzing Method for Random Access Protocol. Journal of Electronics 16(1), 44–49 (1999) Zhou, N., Zhao, D., Ding, H.: Analysis of Multi-Channel and Random Multi-Access Ad hoc Networks Protocol with Two-dimensional Probability. In: Computational Intelligence and Industrial Applications Conference (Proceedings of ISCIIA 2006), November 22-25, pp. 26–32 (2006)

The Effect of Product Placement Marketing on Effectiveness of Internet Advertising Hsiu-Li Liao, Su-Houn Liu, Shih-Ming Pi, and Hui-Ju Chen Department of Information Management, Chung Yuan Christian University, No. 200, Chung Pei Rd., Chung Li, 320, Taiwan, ROC [email protected], [email protected], [email protected], [email protected]

Abstract. Compared to the traditional way of doing advertising, such as ad Banners, internet product placement is now emerging as a promising strategy for advertisers to do their job effectively in this Web 2.0 era. Therefore, this study focuses on the effectiveness of product placement advertising on the Internet. The results show that product prominence (Subtle or Prominent) and presentation of the advertising (Video or Images) significantly impacts the effectiveness of product placement advertising on the Internet, including brand impression, advertising attitude, and intention to click. Product prominence and presentation of the advertisement have an interactive impact. Our findings indicated that presenting the product through videos will enhance higher levels of advertising attitude, brand impression, and intention to click than presenting it through still images. Subtle placements will increase the level of advertising attitude and intention to click more so than prominent placements. But prominent placements increase the brand impression more than the subtle placements. Keywords: Internet advertising, product placement, advertising attitude, brand impression, intention to click.

1 Introduction The Internet has become the third-largest advertising medium in the US, representing 17% of the market (Snapp, 2010). Thousands of advertisers have turned to the Internet as a prospective media for promoting their brands and transacting sales. Internet advertising provides advertisers with an efficient and less expensive way of reaching the consumers most interested in their products and services, and it provides consumers with a far more useful marketing experience. Internet advertising transforms the way consumers and companies find and interact with one another (Snapp, 2010). A growing stream of product placement research has conducted surveys of consumer and practitioner views on the practice and experiments to gauge product placement’s impact on brand awareness, attitudes, and purchase intent (Wiles & Danielova, 2009). In this internet era, product placement is no longer exclusively for big companies with marketing budgets to match. From Facebook to Twitter to R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 332–337, 2011. © Springer-Verlag Berlin Heidelberg 2011

The Effect of Product Placement Marketing on Effectiveness of Internet Advertising

333

bloggers, many ways exist on the Internet to chat up and spread the word about a product (Johnson, 2009). Despite the burgeoning popularity of product placement as a marketing tool on the Internet, there is limited substantive empirical evidence regarding whether and how it is effective in impacting consumer responses. In this study, we try to answer the question: Is there any difference in the proper conducts of product placement between internet and traditional media? In an effort to enhance understanding of the impact of product placements on the Internet, our study purposefully manipulates the product prominence (Subtle or Prominent) and presentation of the advertising (Video or Images). Based on the findings of previous studies, we have proposed that these factors interact to influence the effectiveness of the product placement conducts.

2 Literature Review In the effort to enhance the effectiveness of product placement, there is an old paradox known to marketers: "If you notice it, it's bad. But if you don't notice, it's worthless" (Ephron, 2003). Findings from previous studies (Homer, 2009) indicated that customers will experience greater brand impression increases when product placements are vivid and prominent, but when placements are subtle, consumers’ attitudes toward the advertising are relatively positive. So our first question is: Does product prominence (Subtle or Prominent) have an impact on the effectiveness of product placement advertising on the Internet?” The advent of media-sharing sites, especially along with the so called Web 2.0 wave, has led to unprecedented internet delivery of multimedia contents such as images and videos, which have become the primary sources for online product placement advertising. Industry and various academic studies have acknowledged the importance of capturing a visual image of the placed product on the screen (Russell 1998, 2002). It is valuable for the product placement marketer to know the difference in effectiveness when considering integrating their advertising with images or videos. Therefore, our second research question is: Will the advertising effectiveness be impacted differently when we present the product placement through video or through still images? In this study, the following hypotheses were tested: H1: Product prominence significantly affects the effectiveness of product placement advertising. H1a: Subtle placements lead to a better advertising attitude than prominent placements. H1b: Prominent placements lead to a better brand impression than subtle placements. H1c: Subtle placements lead to higher user intention to click than prominent placements. H2: Product placement advertising that is presented through video can have a greater effectiveness than advertising that is presented through images.

334

H.-L. Liao et al.

H2a: A product placement advertising that is presented through video can lead to a better advertising attitude than presenting the advertising through images. H2b: A product placement advertising that is presented through video can lead to a better brand impression than presenting the advertising through images. H2c: A product placement advertising that is presented through video can lead to a better user intention to click than presenting the advertising through images. 2.1 The Research Model The research model empirically tested in this study is shown in Fig. 1. It represents the proposed research model drawn from the constructs of product prominence, presentation of the advertising, and effectiveness of product placement advertising. It is proposed in this model that different types of product prominence and different presentations of the advertising are potential determinants of the effectiveness of product placement advertising as the independent variable for this study.

Product Prominence - Prominent placement - Subtle placement

H1 H1c

Presentation - through Video - through Images

H1b

The effectiveness of product placement advertising Advertising attitude

H2a H2b

Brand impression

H2c

Intention to click

Fig. 1. The research model

3 Research Methodology This study used a field experiment to test the research hypotheses. This section describes the participants, procedures, instrument development, and measures. 3.1 Participants 400 volunteers were recruited from the SogiKing website (http://www.sogi.com.tw). The SogiKing website is one of the most popular community websites in Taiwan. From the 400 survey questionnaires we distributed, 242 usable responses were returned, resulting in a response rate of 60.5%. Among the 242 usable questionnaires, 136 were females (56%) and 106 were males (44%). Most of the subjects were students (61%) and the age of most subjects ranged from 21 to 25 (61%).

The Effect of Product Placement Marketing on Effectiveness of Internet Advertising

335

3.2 Procedures Random sampling was used to assign users to four groups in Table 1. Subjects in each group were asked to access several WebPages of fashion reports on the SogiKing website with different product placement settings. After browsing the WebPages, the subjects were asked to complete a survey indicating their advertising attitude, brand impression, and intention to click. Table 1. Four groups in the experiment

Video presentation Image presentation (N) : number of subjects.

Prominent-placement Group 1 (53) Group 3 (62)

Subtle-placement Group 2 (63) Group 4 (64)

3.3 Instrument Development This study developed 4 items of brand impression to ask subjects questions about product catalog, brand name, and product characteristic in the product placement advertising. For each question, subjects that had the right answer would get 1 point. Subjects that had the wrong answer got 0 points. The instrument for advertising attitude included a combination of items derived from Garretson and Niedrich (2004), Chang (2004), and Martin et al. (2003). Additionally, the subjects’ intention to click was assessed using a one item constructed questionnaire following the recommendations of Davis et al. (1989). 3.4 Measures The constructs of reliability and validity of the instruments were evaluated. The sample showed a reasonable level of reliability (α > 0.70) (Cronbach, 1970.) Factor analysis also confirmed that the construct validity of the scales could be carried out adequately. Using the principal component method with varimax rotation, construct validity was examined. The factor loadings for all items exceeded 0.8 and indicated that the individual items also had discriminant validity.

4 Analysis and Results The Pearson correlation coefficients for all research variables were found to be significant. Brand impression was negative compared to advertising attitude and intention to click. Advertising attitude was positive compared to intention to click. Data associated with brand impression, advertising attitude, and intention to click was analyzed using a two-way ANOVA test with the independent variable in Table 2. Both product prominence and presentation of the advertising caused significant differences in their effectiveness of product placement advertising, including brand impression, advertising attitude, and intention to click. Product placement prominence

336

H.-L. Liao et al.

and presentation of the advertising had an interactive impact on advertising attitude and intention to click. Compared to the image presentation, subtle placements of video presentations had substantially lower effects. Table 2. The effect of product prominence and presentation of the advertising on the effectiveness of product placement advertising Independent variable Product prominence

Dependent variable

Brand impression Advertising attitude Intention to click Presentation of the Brand impression advertising Advertising attitude Intention to click Product prominence Brand impression *Presentation of the Advertising attitude advertising Intention to click *** p < 0.01, ** p < 0.05, * p < 0.1.

F 295.132 507.331 282.915 13.852 62.650 50.542 0.737 58.607 4.953

P-value 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.391 0.000*** 0.027**

All hypotheses were examined and supported. Video presentations led to a higher level of advertising attitude, brand impression, and intention to click than image presentations. Subtle placements had a higher level of advertising attitude and intention to click than prominent placements, but prominent placements led to a higher brand impression than subtle placements.

5 Conclusion While several past experimental studies report that product placement has little impact on brand attitudes, many practitioners maintain that placement can produce "home runs," especially when certain guidelines are met (Homer, 2009). The primary goal of this study was to investigate two potential factors that may help to increase the effectiveness of the product placement conducts on the Internet, that is, the product prominence (Subtle or Prominent) and presentation of the advertising (Video or Images). Our results show that product prominence (Subtle or Prominent) and presentation of the advertising (Video or Images) both significantly affect the effectiveness of product placement advertising on the Internet. Presently, advertising through video with subtle placements can have the best result on advertising attitude and users’ intention to click. However, to get a better brand impression, you should present the advertising through video with prominent placements. Since all hypotheses were supported, and our findings indicate a consistency with previous evidence, we have concluded that the effects of product placement conducts (Product Prominence and Presentation) on the Internet are similar to the effect of product placement in other media. Our results provide further evidence that the impact of product placement is not a simple phenomenon, but rather that effects are qualified by many moderating factors. As a result, prominent placements on the Internet can have undesirable consequences. We believe

The Effect of Product Placement Marketing on Effectiveness of Internet Advertising

337

that this further highlights the importance of "integrating" your advertising with the content, which is frequently noted by both academics and industry experts but ignored by many internet marketers.

References 1. Aaker, D.A.: Managing brand equity. The Free Press, New York (1991) 2. Brackett, L.K., Carr, B.N.: Cyberspace advertising vs. other media: Consumer vs. mature student attitude. Journal of Advertising Research, 23–32 (2001) 3. Chang, C.: The interplay of product class knowledge and trial experience in attitude formation. Journal of Advertising 33(1), 83–92 (2004) 4. Cronbach, L.J.: Essentials of psychological testing. Harper and Row, New York (1970) 5. Davis, F.D., Bagozzi, R.P., Warshaw, P.R.: User acceptance of computer technology: A comparison of two theoretical models. Management Science 35(8), 982–1003 (1989) 6. Ephron, E.: The Paradox of Product Placement. Mediaweek, 20 (June 2, 2003) 7. Garretson, J.A., Niedrich, R.W.: Creating character trust and positive brand attitudes. Journal of Advertising 33(2), 25–36 (2004) 8. Gupta, P.B., Lord, K.R.: Product placement in movies: The effect of prominence and mode on recall. Journal of Current Issues and Research in Advertising 20, 47–59 (1998) 9. Homer, P.M.: Product Placements: The Impact of Placement Type and Repetition on Attitude. Journal of Advertising 38(3), 21–31 (2009) 10. Johnson, R.: Running the Show — Screen Shots: Product placements aren’t just for big companies anymore. Wall Street Journal, Eastern edition, R.9 (September 28, 2009) 11. Lutz, R.J., Mackenzie, S.B., Belch, G.E.: Attitude Toward the Ad as a Mediator of Advertising Effectiveness: Determinates and Consequences. In: Bagozzi, R.P., Tybout, M. (eds.) Advance in Consumer Research, vol. 10, pp. 532–539. Association for Consumer Research, Ann Arbor (1983) 12. Martin, B.A.S., Durme, J.V., Raulas, M., Merisavo, M.: Email Advertising: Exploratory Insights from Finland. Journal of Advertising Research 43(3), 293–300 (2003) 13. Russell, C.A.: Toward a framework of product placement: Theory propositions. Advances in Consumer Research 25, 357–362 (1998) 14. Russell, C.A.: Investigating the effectiveness of product placements in television shows: The role of modality and plot connection congruence on brand memory and attitude. Journal of Consumer Research 29(3), 306–318 (2002) 15. Snapp, M.: Principles Online Advertisers Can Thrive By. R & D Magazines (2010), http://www.rdmag.com/News/Feeds/2010/03/ information-tech-principles-online-advertisers-canthrive-by/ (last visit: May 9, 2010) 16. Vaughan, T.: Multimedia: Making it work. Journal of Marketing Research 30 (1993) 17. Wiles, M.A., Danielova, A.: The worth of product placement in Successful films: An Event Study Analysis. Journal of Marketing 73, 44–63 (2009)

A Modular Approach to Arithmetic and Logic Unit Design on a Reconfigurable Hardware Platform for Educational Purpose Halit Oztekin1,2, Feyzullah Temurtas1, and Ali Gulbag2 1

Bozok University, Electrical and Electronics Engineering Department, Yozgat, Turkey 2 Sakarya University, Institute of Science Technology, Computer Engineering Department, Sakarya, Turkey {oztekinhalit,temurtas}@gmail.com, [email protected]

Abstract. The Arithmetic and Logic Unit (ALU) design is one of the important topics in Computer Architecture and Organization course in Computer and Electrical Engineering departments. There are ALU designs that have nonmodular nature to be used as an educational tool. As the programmable logic technology has developed rapidly, it is feasible that ALU design based on Field Programmable Gate Array (FPGA) is implemented in this course. In this paper, we have adopted the modular approach to ALU design based on FPGA. All the modules in the ALU design are realized using schematic structure on Altera’s Cyclone II Development board. Under this model, the ALU content is divided into four distinct modules. These are arithmetic unit except for multiplication and division operations, logic unit, multiplication unit and division unit. User can easily design any size of ALU unit since this approach has the modular nature. Then, this approach was applied to microcomputer architecture design named BZK.SAU.FPGA10.0 instead of the current ALU unit. Keywords: Arithmetic and Logic Unit design, Educational tool, FPGA, Computer Architecture and Organization, Teaching method, Modular Approach.

1 Introduction Computer Architecture and Organization is an important basic course in computer and electrical engineering and related fields. ALU design is one of the important topics in this course. It is required to improve the students’ comprehensive capability of handling ALU design problems. Therefore, the experiment must be given sufficient attention except for good teaching in class [1]. Several ALU designs FPGA-based or object-oriented technology is available [2-6]. But these designs are provided as black box units where a user is unable to monitor the inner structure or not enough detail on how it is designed. This is unacceptable from the educational point of view. Also, these designs are not suitable as an educational tool because of non-modular nature. The students need to participate in the R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 338–346, 2011. © Springer-Verlag Berlin Heidelberg 2011

A Modular Approach to ALU Design on a Reconfigurable Hardware Platform

339

process of ALU design and to understand thoroughly the ALU’s inner structure. In order to improve the effect of experimental teaching of ALU design, we have adopted the modular approach to it. It is presented a method that allows the user to design any size of ALU unit in a few steps. The units obtained using this approach can be used both alone and on systems that need the ALU unit. This approach is then applied to microprocessor design named BZK.SAU.FPGA10.0[7] that is the FPGA implementation of BZK.SAU[8] on Altera DE2 Cyclone II development board. This structure is shown in Fig.1.

Fig. 1. Changing the Modular ALU unit obtained in this work with ALU unit in BZK.SAU.FPGA10.0 microprocessor design

2 The Proposed ALU Architecture The proposed ALU has four units: arithmetic unit except for multiplication and division operations, logic unit, multiplication unit and division unit. The top level view of ALU architecture is shown in Fig.2. Output value of ALU according to S1 and S0 selector pins is given Table 1. Table 1. Output value according to S1 and S0 selector pins S1 0 0 1 1

S0 0 1 0 1

The Output Value of ALU Arithmetic Unit Logic Unit Multiplication Unit Division Unit

340

H. Oztekin, F. Temurtas, and A. Gulbag

Fig. 2. Top level view of the proposed ALU

2.1 Arithmetic Unit Any size of arithmetic unit design consists of only two steps. The first step is to design one-bit arithmetic unit. The one-bit arithmetic unit has two 8-to-1 multiplexer and one full adder circuit. The inner structure and block diagram of this unit obtained using Quartus Web Edition Software from Altera Corporation is shown in Fig.3 and Fig.4. The final step is to decide the size of arithmetic unit. n one-bit arithmetic unit block is used to obtain n-bit arithmetic unit. Fig.3. summarizes this structure. It performs different operations according to S2, S1, S0 and Carry_In selector pins of the one-bit arithmetic unit block as shown in Table 2.

Fig. 3. Cascade structure of n-bit Arithmetic unit

A Modular Approach to ALU Design on a Reconfigurable Hardware Platform

341

Fig. 4. The inner structure of one-bit Arithmetic unit in Altera’s FPGA environment Table 2. The operation table according to S2, S1, S0 and Carry_In of selector pins S2 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1

S1 0 0 0 0 1 1 1 1 0 0 0 0 1 1 1 1

S0 0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1

Carry_In 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1

The operation OutÅInput0+Input1 OutÅInput0+Input1+1 OutÅInput0-Input1-1 OutÅInput0-Input1-1 OutÅInput0 OutÅInput0+1 OutÅInput1 OutÅInput1+1 OutÅInput0-1 OutÅInput0 OutÅInput1-1 OutÅInput1 Reserved area

S2, S1 and S0 selector pins are common pins for every unit in the n-bit arithmetic unit. When the arithmetic unit of ALU is used, “Enable” selection input is used to enable or disable the arithmetic unit. “Carry_In” selector pin status in Table 2 is only available for the uppermost one-bit arithmetic unit in Fig. 3.

342

H. Oztekin, F. Temurtas, and A. Gulbag

The occupied area to user in Table 2 is for user’s defined operations. User can define the maximum four operations. If user wants to define the more than four operations, it is sufficient to use larger multiplexer (16-to-1, 32-to-1 etc.) instead of multiplexers of one-bit arithmetic unit. 2.2 Logic Unit It is possible to design any size of logic unit in two steps same as the arithmetic unit. The first step is to build one-bit logic unit. It is necessary an 8-to-1 multiplexer, a AND gate, OR gate, XOR gate and two inverter. The proposed one-bit logic unit performs seven operations as shown Table 3 according to S2, S1 and S0 selector pins and its inner structure and block diagram are shown in Fig.5 and Fig.6.

Fig. 5. Cascade structure of n-bit Logic unit Table 3. The operation table according to S2, S1 and S0 of selector pins S2 0 0 0

S1 0 0 1

S0 0 1 0

0

1

1

1

0

0

1 1 1

0 1 1

1 0 1

The operation OutÅInput0 ∧ Input1 OutÅInput0 ∨ Input1 OutÅInput0 ⊕ Input1 OutÅ Input0 OutÅ Input1 OutÅShifted(Right) Out OutÅShifted(Left) Out Reserved Area

A Modular Approach to ALU Design on a Reconfigurable Hardware Platform

343

Fig. 6. The inner structure of one-bit Logic unit

The shifting operations are meaningful when the logic unit has more than one-bit logic unit block. Therefore “ShiftRight” and “ShiftLeft” input pins can only be used in the event that it composes of more than one-bit logic unit block. The final step to design any size of logic unit is to decide the size of logic unit. n size of logic unit has n one-bit logic unit block. This structure is shown in Fig.5. The connections to SR and SL input pins are done as follows: The uppermost SR input pin is connected to the second bit of the “Operand0”, the next SR input pin to the third bit of the “Operand0” etc and the last SR input pin to logic ‘0’. For SL input pins, the uppermost SL input pin to Logic ‘0’, the second SL input pin to the first bit of “Operand0”, the next SL input pin to the second bit of “Operand0” etc. “Enable” input pin is used to enable or disable the use of Logic unit. If user wants to define one logic function, the reserved area in Table 3 can be used. For defining more than one logic function, it is required to change the current multiplexer with the larger multiplexers (16-to-1, 32-to-1 etc.). 2.3 Multiplication Unit Multiplication unit design is a little more complicated design than other units. Since our aim is to provide uncomplicated designs to be used as an educational tool, we used parallel multiplier unit design instead of conventional multiplication algorithm. The differences between these designs are presented in section 4. Parallel multiplier circuit and its FPGA implementation in Quartus software environment are shown in Fig.7 and Fig.8. n-bit parallel multiplier circuit has n2 one-bit block and n-1 full adder circuits. The connections between these units are as shown Fig.7. There are one full adder circuit and an “and” gate in the inner structure of one-bit block unit. Its inner structure is shown in Fig.8.

344

H. Oztekin, F. Temurtas, and A. Gulbag

Fig. 7. The FPGA implementation of Parallel Mutiplier circuit in Altera’s Quartus software environment

Fig. 8. The inner structure of one-bit block in Parallel multiplier circuit on Altera’s Quartus Software environment

A Modular Approach to ALU Design on a Reconfigurable Hardware Platform

345

2.4 Division Unit Division unit design is a more complicated design than other units. In this work, the division algorithm is developed as shown Fig.9.a. For n-bit division operation, it is required n-bit register with load and shift control pins for dividend (Q), n+1-bit register with load and shift control pins for divisor (M), n+1-bit register with only load control pin for remainder (A), n-bit down-counter with load control pin (P) (initially n) and n+1-bit full adder circuit. There are four states according to this algorithm. It is sufficient four flip-flops to define these states. The outputs of these flip-flops should be designed only one to be logic “1” at the same time. 16-bit division process is implemented in this work as shown in Fig.9.b. The inner structure of this block is not given in detail. It can be examined by downloading from web-site.

(a)

(b)

Fig. 9. (a) The n-bit division algorithm (b) The block diagram of 16-bit division operation using Altera’s Quartus software

3 Results and Conclusions We presented a modular approach to ALU design in this work. Users can design any size of ALU using this method in a few steps since it has modular nature. FPGA development board is used to implement the ALU design because FPGA development boards helps computer and electrical science students learn digital logic design [9]. ALU design is divided four distinct modules. These are arithmetic, logic, and multiplication and division unit. Each of them is divided the modules in itself too. The multiplication unit design is more difficult according to other units. Multiplication operation algorithm is presented in[10]. n-bit multiplication operation lasts n*clock period (T) time when the algorithm is analyzed. If the clock period time is enough long it lasts only one clock period time with this method since it has parallel nature. In other words, it takes (n-1)*T time less. We used this algorithm to perform 16-bit

346

H. Oztekin, F. Temurtas, and A. Gulbag

multiplication operation in BZK.SAU.FPGA10.0 microprocessor architecture design. This modular method presented in this work occurred a noticeable increase in the speed of our microprocessor. In this work, the division unit design is more complicated according to multiplication unit design. We are working to make the division operation more modular same as multiplication operation. It is believed that the presented ALU design will be received since it is available at no cost[11].

Acknowledgement This work was fully supported under TUBITAK (The Scientific and Technological Research Council of Turkey) project no. 110E069.

References 1. Wang, A.: Computer Organization and Construction (3rd Version). Tsinghua University Press, Beijing (2002) 2. Shangfu, H., Baili, S., Zhihui, W., Li, L.: The Virtual Experiment Design of Arithmetic Unit Based on Object-Oriented Technology. In: 2010 Second International Conference on Multimedia and Information Technology, IEEE Conferences, pp. 159–161 (2010) 3. Paharsingh, R., Skobla, J.: A Novel Approach to Teaching Microprocessor Design Using FPGA and Hierarchical Structure. In: International Conference on Microelectronic System Education, IEEE Conferences, pp. 111–114 (2009) 4. Hatfield, B., Rieker, M., Lan, J.: Incorporating simulation and implementation into teaching computer organization and architecture. In: Proceedings 35 th Annual Conference Frontier in Education, IEEE Conferences, pp. F1G-18 (2005) 5. Xiao, T., Liu, F.: 16-bit teaching microprocessor design and application. In: IEEE International Symposium on IT in Medicine and Education, IEEE Conferences, pp. 160–163 (2008) 6. Carpinelli, J.D.: The Very Simple CPU Simulator. In: 32 nd Annual Frontiers in Education, IEEE Conferences, pp. T2F-11—T2F-14 (2002) 7. BZK.SAU.FPGA10.0, http://eem.bozok.edu.tr/database/1/BZK.SAU.FPGA.10.0.rar 8. Oztekin, H., Temurtas, F., Gulbag, A.: BZK.SAU: Implementing a Hardware and Software-based computer Architecture simulator for educational purpose. In: Proceedings of 2010 International Conference on Computer Science and Applications (ICCDA 2010), vol. 4, pp. 90–97 (2010) 9. Zhu, Y., Weng, T., Keng, C.: Enhancing Learning Effectiveness in Digital Design Courses Through the Use of Programmable Logic. IEEE Transactions on Education 52(1), 151–156 (2009) 10. Oztekin, H.: Computer Architecture Simulator Design, Msc. Thesis, Institute of Science Technology, Sakarya, Turkey (January 2009) 11. Modular ALU Design for BZK.SAU.FPGA.10.0, http://eeem.bozok.edu.tr/database/1/ALU

A Variance Based Active Learning Approach for Named Entity Recognition Hamed Hassanzadeh1 and MohammadReza Keyvanpour2 1

Islamic Azad University, Qazvin Branch, Department of Electronic, Computer and IT Qazvin, Iran and member of Young Reaserchers Club 2 Alzahra University, Department of Computer Engineering, Tehran, Iran [email protected], [email protected]

Abstract. The cost of manually annotating corpora is one of the significant issues in many text based tasks such as text mining, semantic annotation and generally information extraction. Active Learning is an approach that deals with reduction of labeling costs. In this paper we proposed an effective active learning approach based on minimal variance that reduces manual annotation cost by using a small number of manually labeled examples. In our approach we use a confidence measure based on the model’s variance that reaches a considerable accuracy for annotating entities. Conditional Random Field (CRF) is chosen as the underlying learning model due to its promising performance in many sequence labeling tasks. The experiments show that the proposed method needs considerably fewer manual labeled samples to produce a desirable result. Keywords: Active Learning, Named Entity Recognition, Conditional Random Field.

1

Introduction

Natural Language Processing (NLP) and Information Extraction (IE) with their various tasks and applications such as POS tagging, NER, NP chunking, and semantic annotation, are matters of concern in several fields of computer science for years [1]. In order to automate these tasks, different machine learning approaches such as supervised learning, semi-supervised learning and active learning are applied [2][3]. Named Entity Recognition (NER) task is one of the most important subtasks in information extraction. It is defined as the identification of Named Entities (NEs) within text and labels them with pre-defined categories such as person name, locations, organizations, etc [2]. In many works NER task is formulated as a sequence labeling problem and thus can be done with machine learning algorithms supporting sequence labeling task [4]. Moreover, in complex structured prediction tasks, such as parsing or sequence modeling, it is considerably more difficult to obtain labeled training data than for classification tasks (such as document classification), since hand-labeling individual words and word boundaries is much harder than assigning text-level class labels. One way to address this issue is to use Active Learning (AL). In AL scenario only examples of high training utility are selected for manual annotation in an iterative R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 347–352, 2011. © Springer-Verlag Berlin Heidelberg 2011

348

H. Hassanzadeh and M. Keyvanpour

manner [5]. Different approaches to AL have been successfully applied to a wide range of NLP tasks. A comparison of the effectiveness of these approaches for sequence labeling tasks of NLP is widely discussed in [3]. In this paper we present a variance based active learning approach for NER task that chooses informative entities to minimize the variance of the classifier currently built from the labeled data. By finding the entities which the current model is certain about labeling them, we use self-training to get the advantage of unlabeled instances. The rest of this paper is organized as follows: Section 2 reviews the related works. Section 3 briefly introduces the learner model variance and the method to minimize the model’s error rate based on it. Section 4 describes the proposed method in detail. Experimental results are reported in Sections 5, and we conclude in Section 6.

2

Related Work

In [6] they propose a range of active learning strategies for IE that are based on ranking individual sentences, and experimentally compare them on a standard dataset for named entity extraction. They have argued that, in active learning for IE, the sentence should be the unit of ranking. Haibin Cheng et al. propose an AL technique to select the most informative subset of unlabeled sequences for annotation by choosing sequences that have largest uncertainty in their prediction. Their active learning technique uses dynamic programming to identify the best subset of sequences to be annotated, taking into account both the uncertainty and labeling effort [7]. A method called BootMark, for bootstrapping the creation of named entity annotated corpora presents in [8]. The method requires a human annotator to manually mark-up fewer documents in order to produce a named entity recognizer with a given performance, than would be needed if the documents forming the base for the recognizer were randomly drawn from the same corpus. Kaiquan Xu et al. propose a semi-supervised semantic annotation method, which uses fewer labeled examples to improve the annotating performance. The key of their method is how to identify the reliably predicted examples for retraining [9]. Another work related to ours is that of Tomanek et al. [4]. They propose an approach to AL where human annotators are required to label only uncertain subsequences within the selected sentences, while the remaining subsequences are labeled automatically based on the model available from the previous AL iteration round. They use marginal and conditional probability as confidence measure estimator and they only apply the semi-supervised scenario as an auxiliary part beside the AL.

3

Variance Reduction for Minimizing Learner Model Error

In this section, we briefly review the model or classifier variance and then argue that minimizing the classifier variance is equal to minimizing its error rate. Let x be an input instance and c is a class label and p c |x is a classifier’s probability estimation in classifying x, and then the actual probability f x is shown as follow: |

.

(1)

A Variance Based Active Learning Approach for Named Entity Recognition

349

Where ε x , is an added error. If we consider that the added error of the classifier mainly comes from two sources, i.e., classifier bias and variance [10], then the added error ε x in (1) can be decomposed into two terms, i.e., β and η x , where β represents the bias of the current learning algorithm, and η x is a random variable that accounts for the variance of the classifier (with respect to class c ). According to [11] classifiers that trained by using the same learning algorithm but different versions of the training data suffer from the same level of bias but different variance values. Assuming that we are using the same learning algorithm in our analysis, without loss of generality, we can ignore the bias term. Consequently, the learner’s probability in classifying x into class c becomes |

η

.

(2)

According to Tumer and Ghosh’s conclusion in [11], the classifier’s expected added error can be defined as Err

(3)

.

p c |x , that p . denotes the deriWhere σ is the variance and s p c |x vate of p . . As Equation (3) indicates, the expected added error of a classifier is proportional to its variance; thus, reducing this quantity reduces the classifier’s expected error rate. Here, for a learner model which is trained on a training set L, the classifier variance for instance x is calculated by σ

|

|

∑

,

y

x

f

.

(4)

Where θ is the current learner model and Υ is a temporary set that its elements are defined by Υ L x, y: θ , while y is the estimated target label for x which is predicted by θ. In (4), |Υ | denote the number of examples in Υ . Base on this analysis, our presented confidence measure base on variance reduction is φ

4

∑

| ;

σ

.

(5)

Minimal Variance Base Active Learning Approach

Recently, several surveys have been done on application of different Active Learning methods in Named Entity Recognition and other NLP tasks [3],[12]. By reviewing these surveys and other related works, it’s revealed that variance based error reduction has been never used in active learning approaches proposed in this field. In this section we describe the way we use our variance based confidence measure in an Active Learning scenario for NER task. Here, we present a combined SSL and AL approach for annotating named entities. For comparing the proposed framework with a common AL approach for sequence labeling. Like a general AL approach, at the beginning of our method, a predicting model is trained on an initial label set. Then, by means of the current model, a label sequence is predicted for each sentence in unlabeled set, and also beside it, we use the

350

H. Hassanzadeh and M. Keyvanpour

probability of each predicted label sequence to find the least confidence sequences from unlabeled set. The rest of operations are proposed on these selected sentences. In this step, the algorithm calculates the variance and the confidence measure proposed in Section 3 for all the tokens in each selected sentences. Based on this measure, tokens with least confidence are picked up to be manually labeled. For using the advantage of other unlabeled tokens, in this step we apply a simple semi-supervised scenario to use the characteristic of those tokens which the current model is certain about them. This semi-supervised phase is described in the following. In recent years, semi-supervised learning is an efficient learning technique which applied in different domain that automation is a matter of concern. Different semisupervised methods have been developed but self-training is the most applied method among them. In our method, we apply self-training approach beside the active learning scenario to automatically label those tokens which have considerably high confident base on the proposed measure, and update the current model with these automatically labeled tokens. We use CRF as the base learner in our framework [13]. The advantage of the this combined framework in a NER task is that, even when a sequence is selected as an informative instance base on its low confident, it can still exhibit subsequences which do not add much to the overall utility and thus are fairly easy for the current model to label correctly. So, within these selected sequences only those tokens remain to be manually labeled which base on the Equation (5) have high variances.

5 5.1

Experiments and Results Experimental Settings

The data set used in our experiment is CoNLL03 English corpus which is a wellknown benchmark for Named Entity Recognition task [14]. The details of train set and test set we applied is shown in Table 1. The CoNLL03 corpus contains 9 label types which distinguish person, organization, location, and names of miscellaneous entities that do not belong to the previous three groups. In our experiment we used the “Mallet” package as a CRF implementation [15]. We employ the linear-chain CRF model in our system. All methods and classes are implemented in Java. A set of common feature functions was employed, including orthographical, lexical and morphological, and contextual ones. Unlike several previous studies, we did not employ additional information from external resources such as gazetteers. Overall experiment start from a 20 randomly selected sequences as initial label set (L). Our method picks up 50 sequences in each iteration. Table 1. Characteristic of CoNLL03 English data 6HWVRI/DEHO7\SHV6HQWHQFHV7RNHQV 7UDLQLQJVHW 7HVWVHW 'HYHORSPHQWVHW

A Variance Based Active Learning Approach for Named Entity Recognition

5.2

351

Experimental Results

The three metrics widely used in the information retrieval field, precision, recall, and F-score, were adopted in this experiment. Table 2 shows the results of our approach on CoNLL03 test set and development set. The number of manually labeled tokens used for training the model in our method is 11464 tokens which are only 5% of all available labeled tokens in training set. Beside these tokens, the model labels 44275 tokens automatically in a semi-supervised manner (overall tokens used for training is 55739, and in term of sequence is 3000 sequences). In fact, automatically labeled tokens are those tokens which the algorithm determined them as confident samples based on proposed variance based confidence measures. Table 2 shows the evaluation of a model which trained on 52147 tokens (3000 sequences). Comparing the results shows that not only our approach uses dramatically lower number of tokens to create a predicting model, but also the performance of trained model in our approach is a little higher that the model created in the same settings but through random sampling. Table 2. Results of our AL method boosted by self-training and Random Sampling (RS) (manually labeled tokens for AL is 11464 and for RS is 52147)

AL method Random sampling

6

Development Set Precision Recall F1-score 0.8831 0.8903 0.8867 PER 0.7365 0.7338 0.7352 ORG 0.8511 0.859 0.8551 LOC 0.8629 0.7375 0.7953 MISC Token Accuracy: Overall F1-score: 0.969 0.8291 0.8638 0.8572 0.8605 PER 0.7718 0.7189 0.7444 ORG 0.8543 0.8841 0.8689 LOC 0.8505 0.7592 0.8023 MISC Token Accuracy: Overall F1-score: 0.9662 0.8279

Test set Precision Recall F1-score 0.8436 0.8442 0.8439 0.6561 0.665 0.6605 0.8213 0.7992 0.8101 0.8137 0.7279 0.7684 Token Accuracy: Overall F1-score: 0.9547 0.7703 0.8071 0.8071 0.8071 0.696 0.6713 0.6834 0.8405 0.8465 0.8435 0.7014 0.6895 0.6954 Token Accuracy: Overall F1-score: 0.9487 0.7682

Conclusions

In this paper, we propose an active learning method for NER task base on minimal variance. For fully take the advantage of unlabeled data, we use self-training beside our AL algorithm. Especially, a prediction confidence measure based on minimizing the classifier’s variance is described that we studied that its minimization is equal to minimizing the classifier’s prediction error rate. CRF is chosen as the underlying model for this experiment. The new strategy we proposed makes the data with representative information have much higher selection opportunity and improve the system learning ability effectively. The experiments showed that our approach used considerably fewer numbers of manually labeled samples to produce the same result when samples are selected in a random manner. In this work, fix amount training samples are added for each iteration, we will find how to use the unlabeled data chose by self training in a more intelligent way in future work.

352

H. Hassanzadeh and M. Keyvanpour

References 1. Keyvanpour, M., Hassanzadeh, H., Mohammadizadeh Khoshroo, B.: Comparative Classification of Semantic Web Challenges and Data Mining Techniques. In: 2009 International Conference on Web Information Systems and Mining, pp. 200–203 (2009) 2. Nadeau, D., Sekine, S.: A survey of named entity recognition and classification. J. Linguisticae Investigation 30, 2–26 (2007) 3. Settles, B., Craven, M.: An Analysis of Active Learning Strategies for Sequence Labeling Tasks. In: Empirical Methods in Natural Language Processing, pp. 1069–1078 (2008) 4. Tomanek, K., Hahn, U.: Semi-Supervised Active Learning for Sequence Labeling. In: 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pp. 1039–1047 (2009) 5. Settles, B.: Active learning literature survey. Technical Report, Wisconsin-Madison (2009) 6. Esuli, A., Marcheggiani, D., Sebastiani, F.: Sentence-Based Active Learning Strategies for Information Extraction. In: 1st Italian Information Retrieval Workshop, IIR 2010 (2010) 7. Cheng, H., Zhang, R., Peng, Y., Mao, J., Tan, P.-N.: Maximum Margin Active Learning for Sequence Labeling with Different Length. In: 8th Industrial Conference on Advances in Data Mining, pp. 345–359 (2008) 8. Olsson, F.: On Privacy Preservation in Text and Document-based Active Learning for Named Entity Recognition. In: ACM First International Workshop on Privacy and Anonymity for Very Large Databases, Hong Kong, China (2009) 9. Xu, K., Liao, S.S., Lau, R.Y.K., Liao, L., Tang, H.: Self-Teaching Semantic Annotation Method for Knowledge Discovery from Text. In: 42nd Hawaii International Conference on System Sciences (2009) 10. Friedman, J.: On bias, variance, 0/1-loss, and the curse-of dimensionality. J. Data Mining Knowledge Discover. 1, 55–77 (1996) 11. Tumer, K., Ghosh, J.: Error correlation and error reduction in ensemble classifier. J. Connection Sci. 8, 385–404 (1996) 12. Tomanek, K., Olsson, F.: A web survey on the use of active learning to support annotation of text data. In: NAACL HLT Workshop on Active Learning for Natural Language Processing, pp. 45–48. Boulder, Colorado (2009) 13. Lafferty, J., McCallum, A., Pereira., F.: Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In: ICML 2001, pp. 282–289 (2001) 14. Sang, E.F.T.K., Meulder, F.d.: Introduction to the CoNLL- 2003 shared task: Languageindependent named entity recognition. In: CoNLL 2003, pp. 155–158 (2003) 15. McCallum, A.K.: MALLET: A Machine Learning for Language Toolkit (2002), http://mallet.cs.umass.edu

Research on Routing Selection Algorithm Based on Genetic Algorithm Guohong Gao, Baojian Zhang, Xueyong Li, and Jinna Lv School of Information Engineer, Henan Institute of Science and Technology Xinxiang, 453003, China [email protected]

Abstract. The hereditary algorithm is a kind of random searching and method of optimizing based on living beings natural selection and hereditary mechanism. In recent years, because of the potentiality in solving complicate problems and the successful application in the fields of industrial project, hereditary algorithm has been widely concerned by the domestic and international scholar. Routing Selection communication has been defined a standard communication model of IP version 6.This paper proposes a service model of Routing Selection communication, and designs and implements a new Routing Selection algorithm based on genetic algorithm.The experimental simulation results show that this algorithm can get more resolution at less time and more balanced network load, which enhances search ratio and the availability of network resource, and improves the quality of service. Keywords: routing select; genetic algorithm; service model.

1 Introduction Network multimedia messaging services promotes the growing demand for multimedia applications and existing network technology for further development. With the development of network applications, people on the network quality of service QoS (Quality of Service) have higher and higher requirements. How to achieve a variety of network quality of service is the world's computers, electronics and communications topics at the forefront of race. To increase the provision of network quality of service and network load balance, people often use multiple service sites to meet the requirements of network users. For example, the network often has a number of services to meet the same requirements of the service site. In this way, users on the network's service requirements, the network can be any one site to provide its services. Now the network services such as WWW's "mirror" site, SOCK server [2] belong to such a structure. To study the traffic, in recent years people put forward the "routing" of the communication model in which a given network and a set of source destination point, seeking from the source to any destination point routing path. Routing communication has been defined as a standard in the Ipv6 traffic model in [1]. There are two different modes of world's communication model for routing :one is for the network at the application level communication on the network routing [4,5], R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 353–358, 2011. © Springer-Verlag Berlin Heidelberg 2011

354

G. Gao et al.

including the routing of the communication model and selecting a target Site Strategy; the other is the network layer routing in the network of communication [3], this study has just started, mainly for routing traffic in the routing table structure and routing algorithm. This paper presents a new routing algorithm, the idea is to use the network routing algorithm for routing the request to a short delay to reach local optimum, thus improving search efficiency, so as to improve the efficiency and quality of service network.

2 Routing Service Model The basic model of routing services are as follows: G = Among them, V = V={v1

，v2，…，vn}that is a computer network.of the nodes, E

= {e1, e2,... enw }, the link is a network. Each link ej has a usable link bandwidth (or link capacity) Bj. A routing request to:

（S，D，B，N）

R=

Among them, routing address D is a group that can provide the same service to source node of the destination node, usable G(D) to a group of purpose that host G(D) = {D1, D2,... , the Dm}, including D1 D2 … Dm V. Source node to request service is a group of S the source node, usable G(S) said a source host: G (S) = {S1, S2,... , Sn}, including S1 S2 … Sn V and S1 S2 … Sn.not. G(D); B for transmission bandwidth, N for transmission delay needs, such as to simplify the delay, N demand can be expressed as the maximum transmission path allows the link (assuming all links with the same delay and transmission path delays in transmission path is delayed accumulation) all links can be simplified algorithm, so the discussion. Routing problem can be described as follows: given a network G = and a QoS routing request R = (G (S), G (D), B, N), as G(S) in each A source node Si to find a path from the source node Si to G(D) in any one node of a transmission path Pi, mak-

， ， ， ∈ ， ， ， ∈ ， ， ，

∈

n

ing each network link [k, l] bandwidth is not less than

i = 1 . and

∑

Bi

, but on the path

.[ k , l ]. in . Pi

Pi link number is not greater than Ni, and requested the total delay of each path reaches a minimum in order to achieve network load balancing. The problem of the mathematical description as follows:

∈ ，subject to：

for a path Pi for each Si G(S) for all dege[k,l] E

∈

n

min{

∑∑ delay([k , l ]) }

i =1[ k ,l ].in . Pi n

∑B < B i

i =1.and .[ k ,l ].in. Pi

kl

Research on Routing Selection Algorithm Based on Genetic Algorithm

355

Which, delay [k, l] that the path of each link on the delay, Bi is the path Pi is associated with the flow of service requests (dynamic value). Requirements of each path to satisfy the conditions Pi: Pi on the path of each link [k, l] the delay is less than Ni, the network each link [k, l] on the flow through the transmission of requests and small bandwidth in the link Bk, l.

3 Genetic Algorithm Based Routing Algorithm Design algorithm is: first of all, construct the initial routing table and routing table structure based on the initial population, then the routing calculation and genetic evolution at the same time, the evolution of the current node after the routing table is passed to the next node, next node routing table updates, an extension of corresponding chromosomes, and then continue to evolve until the get the optimal routing. The purpose is to use genetic algorithm with simpleness, universal algorithm and strong robustness, parallel and excellent characteristics, which improves group for routing algorithms searching efficiency and the network service quality. The following specific description of the algorithm: Algorithm: Routing Selection-Routing Input: network topology Re [ ] [ ], the source node list sA, dA destination list Output: the optimal routing path P Begin Step1 To take the source node s; The initial node path P s; for j = 0 to netSize (the total number of network nodes) do { if there is the link [s, j] then {Node j added to the path P; The path P into the matrix rout [m] [k + +]; } } Step2 initial population structure according to the routing table; Set generation and maxgen; Evolution maxgen generation; Characterization of the best individual path while P nodes do not reach the destination group {/ / According to network topology Re [] [], an extension of the routing table accordingly, get a new routing table; for each path P in rout [] [] do {Path P gives the last node s; for j = 0 to netSize (the total number of network nodes) do if there is the link [s, j] then {Node j added to the path P; Test the legality of the path P; if legal then the path P if (j in dA) then

356

G. Gao et al.

The path P into the matrix rout [m] [k + +]; } } Under the new routing table, the parent population has a corresponding extension of the individual; Modify the evolution of algebra maxgen; Evolution of maxgen generations; Shows the optimal path P; end. Because the algorithm is the side of route calculation, update the routing table, while genetic evolution, it does not get much network node status information to the routing, thereby ,it greatly reduces the burden on the network and improves the routing speed, meanwhile it reducesthe storage overhead.

4 Algorithm Analysis and Evaluation Intuitive view, the use of genetic algorithm is to improve search efficiency of the routing algorithm, so as to achieve the global optimal routing algorithm, as well as network load balancing using random methods is such an effective routing method, but in the use of genetic algorithm routing, we must ensure that the convergence of the algorithm, and strive to avoid the routing loop. We designed a dedicated network simulation system for analysis and evaluation of routing algorithms. In network simulation experiment, for simplicity, the source address and destination address group is a group specific random function generated. Below the experiments, each data points were randomized trials, then 100 times the average statistical all experiments. We first to network scale change circumstances efficiency of the algorithm and the convergence of the experiment, the experimental network node number by 8-48, the number of fixed for the purpose of genetic algorithm, and the initial population scale for 20, evolutionary 70 generation, Genetic algorithm is approaching the crossover probability and mutation probability from 0.5 about 0.01. Experimental results (figure 1 and figure 2).

Average evolution algebra

60 40 30 20 0 8

18

28

38

48

Network nodes Fig. 1. The number of network nodes, the average change in the evolution algebra Choose sowing expenses

Research on Routing Selection Algorithm Based on Genetic Algorithm

357

Choose sowing expenses

450 400 350 300 250 18

28

38

48

58

Network nodes Fig. 2. The number of network nodes, routing fee changes

It can be seen from Figure 1, when the network size increases, the algorithm, the average generation and growth in almost a straight line, this is due to number of nodes increases, the genetic algorithm variable length chromosome coding, the search space increases, solving the required the evolution of algebra also will increase. Figure 2 reflects the changes in the number of network nodes, the cost of routing. When the network size increases gradually, routing algorithm has increased the cost of the path gradually, which is obvious. But when the network reaches a certain size, the algorithm cost routing path of a stable trend in the value of which played its local algorithm for solving superiority.

Running time

2.0 1.5 1.0 0.5 0 3

4

5

6

7

8

9

Purpose of the number of nodes Fig. 3. Groups the number of nodes changes the purpose of running time

We also change the purpose of the case number of nodes the algorithm efficiency and convergence of the experiment, at this time the number of network nodes is fixed at 20 nodes, the purpose of the number of nodes changes from 3 to 10, the size of genetic algorithm initial population is set to 20 , Evolution 60 generations; genetic algorithm is approximately 0.5 crossover probability, mutation probability of about

358

G. Gao et al.

0.01.Figure 3 shows that the purpose of the network size fixed and changing the number of nodes, the algorithm of routing costs, but the increase is relatively stable, because the destination node increases, making it easier to meet the satisfaction of the evolution of the destination node. These network simulation experiments show :the proposed algorithm has good service performance, and can balance the network load effectively and improve search efficiency. Because of the genetic algorithm using a simple, universal, robust, parallel search, groups seeking excellent characteristics, which ensures the convergence of the algorithm and avoids the routing loop.

5 Conclusion Since a large number of network traffic, the network state is highly dynamic, so the traditional network routing methods can not avoid the temporary local network congestion, it is difficult to meet the needs of network users. This paper presents a new genetic algorithm based routing algorithm, which can achieve a shorter delay of local optimization, better balance the network load, improveing network utilization and quality of service. The algorithm is so simple and easy to implement that it can be applied to the actual network.

References 1. 2. 3. 4. 5.

Hinden, R., Deering S.: IP version 6 addressing architecture. RFC 1884, IETF (1995) Partridge, C., Mendez, T., Milliken, W.: Host Routing Selection server. RFC 1346, IETF (1993) Kwon, D.H., et al.: Design and Implementation ofan Efficient Mult icast Support Scheme for FMIPv6. In: INFOC0M 2006, pp. 1–12 (2008) Jia, W.-j., Xuan, D., Zhao, W.: Integrated routing algorithms for Routing Selection messages. IEEE Communications Magazine 38(1), 48–53 (2000) Jia, W.-j., Zhao, W.: Efficient Internet multicast routing using Routing Selection path selection. Journal of Network and Systems Management 12(4), 417–438 (2002)

Optimization of the Performance Face Recognition Using AdaBoost-Based Mohsen Faghani1, Md Jan Nordin1, and Shahed Shojaeipour2 1

Center for Artificial Intelligence Technology, Faculty of Information Science and Technology 2 Dept. of Mechanical & Materials Engineering, Faculty of Engineering &Built Environment Universiti Kebangsaan Malaysia 43600, UKM Bangi, Selangor Malaysia [email protected], [email protected], [email protected]

Abstract. In this paper, using the results of classifier composition is one of the methods of increasing efficiency of face recognition systems that many researchers paid attention to it in recent years. However AdaBoost algorithm is as one of the efficient boosting algorithm that has been used as to decrease the dimensions of characteristic space extracted from face recognition systems, it hasn’t been used as classifier in face recognition systems. In this paper paid attention to how to use this algorithm in classifying face recognition systems. At first the methods evaluated of classifier composition. Then, the result is presented of several composition methods in comparison with singular classifying methods; therefore, database has correct recognition of 96.4% and improved the results of KNN method with PCA specification. AdaBoost method is used according to weak learning, as proposed classifier system with the aim of identification validate. Keywords: Face Recognition, Classifier Composition, PCA, LDA, AdaBoost.

1 Introduction AdaBoost algorithm presented by Freund and Schapire in 1995 for the first time [1]. One of practical problems of Boosting with filtering methods needs training samples. This problem can be addressed by Adaboost. In this algorithm there is the possibility of using training information. As Boosting by filtering in Adaboost, there is weak training model available. The aim of this algorithm is to find final theory that has low error rate with respect to data D distributed to training samples. This algorithm is different from other boosting algorithms [2]. Adaboost is set as adaptable according to errors of weak theories in weak training models. Performance boundaries of Adaboost is related to performance of weak training model [3]. In this algorithm simple samples from training set from previous weak theories that was classified correctly gain less weights and hard samples from training set that was classified incorrectly gain more weights. So Adaboost algorithm concentrates to sample with hard classification. R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 359–365, 2011. © Springer-Verlag Berlin Heidelberg 2011

360

M. Faghani, M.J. Nordin, and S. Shojaeipour

Simulations done using computer with Intel® Pentium ® M-1.70GHz and with 512 Megabyte RAM and software MATLAB 7.40. 1.1 Voting Methods In these methods, declaration of each classifier due to input pattern, is calculated as one vote, and final decision making is done by total votes of different classification [4]. The input pattern will be belonged to class with the most votes. If classifiers be independent of each other and their recognition rates be more than 50%, by increasing the number of classifier, the method of voting will increase the accuracy of classification. • Unweighted Voting Method, the votes of all classifiers have the same weights. Winning criteria in these methods is the total votes extracted. Complete agreement, majority vote, absolute majority, correction and multi stage vote are voting methods without weight [5]. • Confidence Voting Method, voter shows his confidence level to each candidate. The more the confidence to candidate is the more winning probability. The candidate who gain more vote will be chosen [6]. • Ranked Voting Method, voters present their list of candidates due to their interest ranking. In this case the mean of ranks presented by voters is calculated as the final votes of that candidate, and the candidate with less rank will be chosen as the superior candidate [7]. In weighted voting methods different criteria is used to identify assigned weight to each classifier, most common of them is using performance of classifier to experimental samples. On compound system containing L independent classifier has most accuracy once every weight in classifier Di be chosen as wi=log(pi /1-pi ) That pi is the accuracy of classifier i.

2 Database For comparison different methods of face recognition has been gathered by research centers. By using set of images, we can compare different methods. Different databases have been created by different research centers, such as PIE, AR, and FERET. Different databases have unique specifications besides common specifications. Analysis and discussion should be done according to specification and results of each datasets. For better results of methods, they should be tested in different datasets. In proposed method for face recognition, Yale and ORL database are used for simulation. Each of them has specification that is useful for results analysis. We present their specification and images in the next section. 2.1 ORL Database ORL database is composed of 400 images related to 40 persons, images of them are in ten different states. Images are in 112*92 pixels and 256 grey level. Changes contain

Optimization of the Performance Face Recognition Using AdaBoost-Based

361

light, face state (cheerful, upset, open or closed eyes), details of face (with glass or without glass), face cover and circle in depth about 20 degrees. These sets of images are created in 1992 to 1994 so it contains age changes too. We made the images of this database smaller in order to decrease calculations and used in 46*46 size format. 2.2 Yale Database Yale database is composed of 165 images related to 15 persons, images of them are in 11 different state. Changes contain light, face state (cheerful, upset, sleepy, astonishing), direct sun light, light from left side, light from right side, glass state, without glass. Images haven’t any depth or level circulation. Images are in 320*243 pixels and 256 grey levels. For simulation, face side is cut and editing is done to it. We made the images of this database smaller in order to decrease calculations and used in 50*50 size format. 2.3 Training and Experimental Images For evaluating a method of face recognition in database, images need to divide to experimental and training sets. Training images are images that are edited in the training state. Experimental images are images that system should identify it correctly according to samples tested. In simulations done using ORL database, selection of images are random and run several times and mean was used from results. So all the images in each run are selected randomly for training and experiment. The results have more confidence with respect to selection mode of samples at constant state. Choosing the number of training and experimental images has direct effect to results. Here the number of training images and tests has considered equal and five. So the images of training and experimental images are 40*5. In Yale datasets the selection of images has done randomly too. The number of 6 images for training and 5 images for experiments are selected. So training set is 15*6 and experimental set is 15*5. 2.4 Extraction Method and Selection of Feature Extraction method and selection of specifications in this paper is PCA and LDA and selection of convenient coefficient. The selected specifications for ORL database and LDA method are 39 feature and for Yale and LDA is 14 features. The best result obtained for selection of 22 feature of PCA from ORL database, 22 feature from Yale and PCA. 2.5 Classifier of Database In proposed method of AdaBoost algorithm as classifier of face recognition system is used as double class. The weak learner used is binary classifier. For each class one independent AdaBoost is created. For selection of two class samples, firstly training samples quantified with lable1 (experimental set), then for later class lable2 (training set) was selected randomly from remaining class. So for ORL database with 40 class, AdaBoost40 and for Yale with 15 class, AdaBoost15 was created. For result

362

M. Faghani, M.J. Nordin, and S. Shojaeipour

comparison, face recognition system according to KNN method with both extraction method was simulated and the results was compared with proposed method. The way of selecting method of training images is one of the effective points in recognition rate. To increase the credits of results statistically, we select images randomly and calculate mean and standard deviation. Experiments were repeated 30 times for each method. Firstly the images of PCA method and then LDA method are presented. Then related charts of simulations are presented in two databases. 2.6 Comparison of Proposed Method with Othermethods First we evaluate comparison of performance of proposed method with other methods. As it is presented in Figure 1 and Figure 2 recognition results, rate of recognition with AdaBoost method and PCA specifications had significant improvements. Also in Yale database it had very effective improvements. So it can be said that proposed method is strong against light, circulation, size and state changes. By using this method according to LDA specification, significant improvement obtained from ORL and Yale database. Recognition results obtained to all data are presented in Table 1. Table 1. Recognition results Number of Number of Training Experimental Image Image Yale ORL Yale ORL 75 200 90 200 75 200 90 200 75 200 90 200 75 200 90 200

Yale database

ORL database

Method

65.0% 94.3% 87.0% 93.6%

89.3% 96.4% 88.5% 95.3%

KNN+PCA AdaBoost-PCA KNN+LDA AdaBoost-LDA

3 Evaluating Standard Deviation in Proposed Method Another important parameter that was evaluated is the rate of standard deviation in proposed method in different runs. The amounts of standard deviation are presented in Table 2. These changes can be seen in Figure 1 and 2. The amounts of standard deviation in proposed method is very less than KNN method. Kind of specification hasn’t effect on standard deviation of AdaBoost. The amount of standard deviation criteria shows algorithm stability and its independence from different conditions of tests. Table 2. Compared Averages & Criterions Yale database Criterion 3.11 0.70 3.50 0.50

Average 65.0% 94.3% 87.0% 93.6%

ORL database Criterion 2.30 0.18 2.12 0.16

Average 89.3% 96.4% 88.5% 95.3%

Method KNN+PCA AdaBoost-PCA KNN+LDA AdaBoost-LDA

Optimization of the Performance Face Recognition Using AdaBoost-Based

363

0.96 Adaboost-LDA Adaboost-KNN

0.94 0.92

Rcognition Rate

0.9 0.88 0.86 0.84 0.82 0.8 0.78

0

5

10

15 20 25 Number of classes

30

35

40

Fig. 1. Compare result for every class between LDA and KNN for ORL database 0.98 Adaboost-PCA Adaboost-KNN

0.96 0.94

Recognition Rate

0.92 0.9 0.88 0.86 0.84 0.82 0.8 0.78

0

5

10

15 20 25 Number of classes

30

35

40

Fig. 2. Compare result for every class between PCA and KNN for ORL database

4 Error Analysis By evaluating of using several different weak learning method in AdaBoost, as it is by evaluating standard deviation in different classes, our experiment in ORL database, most errors related to classes 5, 11, 15 and 40. By evaluation of images in these classes we can see that light conditions, backgrounds and face state has more changes. In Yale database, most errors related to classes 6 and 12. By evaluation of images in these classes we can see that there is uncontrolled light conditions in these images. 4.1 Evaluation of Runtime in Proposed Method One of the other important parameters in comparing performance is runtime. Runtime in proposed method of AdaBoost was less than in KNN method, it can be seen in Table 3.

364

M. Faghani, M.J. Nordin, and S. Shojaeipour Table 3. Compared the execution time Yale database

ORL database

execution time 0.037 0.010 0.034 0.004

execution time 0.130 0.035 0.160 0.017

Method KNN+PCA AdaBoost-PCA KNN+LDA AdaBoost-LDA

4.2 Evaluation of FAR and FRR Errors The amounts of FAR and FRR in all methods obtained for all images then its mean was calculated. As it can be seen in Table 4 the decrease of FAR in proposed method with respect to KNN has significant improvements. But decrease of FRR in proposed method with PCA specification in Yale database was improved and increased in the other methods. Table 4. Errors of the FAR & FRR Yale database Error FAR 32.7 28.6 27.4

Error FRR 39.9 1.4 12.7 1.5

ORL database Error FAR 0.7 15.0 29.0

Error FRR 8.3 1.12 9.6 0.4

Method KNN+PCA AdaBoost-PCA KNN+LDA AdaBoost-LDA

5 Conclusion In this paper we used extracted specifications according to PCA and LDA conversion. For training and testing this system, 400 images from ORL database and 165 images from Yale database was used. AdaBoost as compound classifier could improve results with respect to singular classifier. Final model presented is strength against different state of face, light changes and so on. The method presented in ORL database has correct recognition of 96.4% and improved the results of KNN method with PCA specification. the method presented in Yale database has correct recognition of 94.3% that has better performance with respect to KNN method with PCA specification. Finally by using proposed method with LDA specification in ORL database, the recognition percentage is 95.3% in Yale database has correct recognition of 93.6% that has better performance with respect to KNN method with LDA specification.

References 1. 2.

Freund, Y., Schapire, R.E.: A Decision-Theoretic Generalization of On-Line Learning and an Application to Boosting (1995) Zhang, T.: Convex Risk Minimization. Annals of Statistics (2004)

Optimization of the Performance Face Recognition Using AdaBoost-Based 3.

4.

5.

6.

7.

365

Yang, P., Shan, S., Gao, W., Stan, Z., Zhang, D.: Face Recognition Using Ada-Boosted Gabor Features. In: 6th IEEE International Conference on Automatic Face and Gesture Recognition, Korea, pp. 356–361 (2004) Zhang, G., Huang, X., Li, S.Z., Wang, Y., Wu, X.: Boosting Local Binary Pattern (LBP)Based Face Recognition. In: Li, S.Z., Lai, J.-H., Tan, T., Feng, G.-C., Wang, Y. (eds.) SINOBIOMETRICS 2004. LNCS, vol. 3338, pp. 179–186. Springer, Heidelberg (2004) Ivanov, Y., Heisele, B., Serre, T.: Using component features for face recognition. In: 6th IEEE International Conference on Automatic Face and Gesture Recognition, Korea, pp. 421–426 (2004) Rowley, H.A., Baluja, S., Kanade, T.: Neural Network-based face Detection. In: 1996 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, USA, pp. 203–208 (1996) Gökberk, B., Salah, A.A., Akarun, L.: Rank-based decision fusion for 3D shape-based face recognition. In: Kanade, T., Jain, A., Ratha, N.K. (eds.) AVBPA 2005. LNCS, vol. 3546, pp. 1019–1028. Springer, Heidelberg (2005)

Design and Implementation Issues of Parallel Vector Quantization in FPGA for Real Time Image Compression Krupa R. Rasane1 and Srinivasa Rao R. Kunte2 1

Department of Electronics and Communication Engineering, KLESCET, Belgaum, Karnataka, India [email protected] 2 Principal JNNCE, Navale, Shivamogga, Karnataka, India [email protected]

Abstract. In this paper a 4 codebook, Vector Quantization (VQ) core is implemented on FPGA (Field Programmable Gate Array). The proposed design has certain advantages over the earlier architecture in the form of design reuse of VQ core to build a large VQ system. The proposed core aims at increased compressing speed, modular design for design flexibility, easy reconfigurability. Modularity helps in flexile design changes for VQ with different codebook sizes and hence controls the recovered image quality. In general, the new VQ core, meets the specific and challenging needs of a single functioned, tightly constrained real time VQ encoder. The synthesis results show that a speed up of 5 is achieved. Experiments and analyses indicate that our design can satisfy the performance requirements of 30 image frames per sec for a real time image processing. The proposed VQ requires more memory and implements VQ encoder with codebook size which are in multiples of 4. Keywords: Vector Quantization, Image compression, LBG Algorithm, Image encoder, VLSI, pipelining, FPGA.

1 Introduction Demand for communication of multimedia data through the telecommunications network and accessing the multimedia data through Internet is growing widely. Image data comprises of a significant portion of the multimedia data and they occupy a large amount of the communication bandwidth for multimedia communication. As a result, development of efficient image compression techniques continues to be an important challenge to us. Vector quantization (VQ) is a lossy compression technique used to compress picture, audio or video data. Fig 1 shows the block diagram of typical vector quantizer engine. A vector quantizer maps k-dimensional vectors in the vector space into a finite set of vectors. Each of these finite vectors is called a code vector or a codeword and the set of all the codewords is called a codebook. The vector quantizer compression engine consists of the complex encoder at the transmission side and very simple decoder at the receiver side. The encoder takes the input vector and R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 366–371, 2011. © Springer-Verlag Berlin Heidelberg 2011

Design and Implementation Issues of Parallel VQ in FPGA

367

outputs the index of the codeword that offers the lowest distortion. The lowest distortion is found out by evaluating the Euclidean distance in the form of MSE (Mean Square Error) between the input vector and each of the codeword obtained from the codebook. Once the index of the winner codeword (codeword with least distortion) is sent through the communication channel the decoder replaces the index with the associated codeword from a similar codebook. In the last years we can notice a rapid development of the design methods. Advanced re-programmable units are larger and larger (in respect of gate number) so it enables the implementation of complicated digital units like SoC (System on Chip), processors or specialized controllers with enhanced speed. VQ Decoder

VQ Encoder Input Image Vector

Codebook Search

Index

Communication or Storage Channel

Index

Codebook Index Matching

Recovered Image vector

Fig. 1. VQ Engine Block Diagram

2 Codebook Training Analysis Training of a codebook usually refers to computing of a new codebook with a better quality of recovered image having a good signal-to-noise ratio. During training, Mean Square Error (MSE), between the initial codevectors from the codebook and the input image vectors is iteratively evaluated. Groups of matched codebook are indexed by the same Index. Two of the error metric used to compare the various image compression techniques are the Mean Square Error (MSE) and the Peak Signal to Noise Ratio (PSNR). The MSE is the cumulative squared error between the compressed and the original image, whereas PSNR is a measure of the peak error given as MSE

1 MN

I x, y

I x, y

PSNR 20*Log 10 255/sqrt MSE )

(1)

(2)

where I(x,y) is the original image, I'(x,y) is the recovered decompressed image and M,N are the dimensions of the images. A lower value for MSE means lesser error, and as seen from the inverse relation between the MSE and PSNR, this translates to a high value of PSNR. Logically, a HIGHER value of PSNR is good because it means that the ratio of signal to noise is higher. For fast execution of heavy matrix and vector computations, Matlab uses processor-optimized libraries, The codebooks can be first trained off line using Matlab or even C. In our implementation, the Matlab results were used for verifying the results of FPGA Implementation.

368

K.R. Rasane and S.R.R. Kunte

2.1 Training the Codebook The iterative algorithm is executed in Matlab on a Lena image Matlab results for 10 Iterations were performed on 256x256 Lena Image for a codebook size of 256 and vector dimension 16. Its MSE: 75.9049 ~ 60.1845 dBs. PSNR: 29.3281 ~ 30.3360 dBs, and were within the acceptable value of 30 db. As the code book size increases and as the vector dimension increases this complex processing consumes more time and hence not suitable for real time applications. From the design point of view, VQ encoder can be suitably designed as a Hardware - Software Co-design.

3 Hardware Realization of Proposed 4 Codeword VQ The main goal of our work is presenting a practical approach of speeding up the operation of VQ by using Hardware implementation techniques. An analysis has been made from the Matlab Profile result which shows that more than 50% time is spent on processing the Euclidean distance, which involves iterative operations of subtraction, squaring and summation for all the input and the codebook. The hardware implementation is the parallel implementation of the Euclidean distance for all codevectors in the codebook. This reduces the entire processing time to a single clock period. Further speed has been enhanced by prior computing constants with known values of codevectors, prior to the arrival of input image thus saving time. 3.1 Software Design The Software part can be implemented using the embedded MicroBalze and internal memory of a Xilinx Vertex FPGA. Implementation such as off-line training of codebooks and prior computation of constants before the arrival of a real time image data is executed without any timing constraints and can form the pre-processing or the initialization of VQ system. The complete Hardware-Software co-design is shown in Fig 2. For the sake of testing here, VQ decoder is implemented on the same module. It will be however implemented at the decoder end of the communication system. Internal Memory

VQ VQ Encoder FPGA Decoder

External Memory

Fig. 2. Proposed New VQ Engine Block Diagram

4 Modified Mathematical Analysis for Hardware Implementation The approach for both the conventional method and for a new method has been discussed in the following session.

Design and Implementation Issues of Parallel VQ in FPGA

369

4.1 Generic Algorithm for Conventional Module Let X0,X1,…etc. be the input vectors of dimension k=2 for an image X and let X0 be represented as X0= (X00, X01)

(1)

Also a 2 dimensional codeword which has been trained using Matlab by using a number of standard images and of codebook size=N, be represented as follows. Since our module is designed to find a winner codeword for every 4 codewords, let the 4 codewords be represented as 2 dimensional data. CW0 = ( CW00, CW01 ) CW1 = ( CW10, CW11 ) CW2 = ( CW20, CW21 )

(2)

CW3 = ( CW30, CW31 ) Let each codeword be associated with an Index IN for N=0,1,2,3 I0 - ( CW0 ) I1 - ( CW1 ) I2 - ( CW2 )

(3)

I3 - ( CW3 ) Algorithm: The generic algorithm used in VQ Implementation is given as follows Step 1: MSE calculation, between any input vector and the codeword N is given as, D0(N) = ( X00-CW(N)0 )2+ ( X01-CW(N)1 )2 For e.g. The MSE between the X0 and CW0 (N=0) is D00, and MSE between X0 and CW1(N=1) is given by D01 and is given as D00 = ( X00-CW00 )2 + ( X01- CW01 )2 D01 = ( X00-CW10 )2 + ( X01- CW11 )2

(4)

Step 2: Compare D00, D01……….D0 (N-1) to find the MSE. Step 3: Encode the winner codeword and hence the Index value sequentially for every input. Step 4: Repeat the above steps for all the input vectors of the image. 4.2 Algorithm for the Proposed Module The VQ encoder involves the Euclidian distance measurement. This process is modified by using a look-up-table as discussed below, which enhances the speed as compared in paper [4]. Here, the mathematical analysis and the improvement made is discussed for the hardware Implementation. Let D1 is the distortion for a 2

370

K.R. Rasane and S.R.R. Kunte

dimensional vector input and the codewords CW0 and CW1 i.e. (D1=D00-D01), for the input I0. Substituting and rewriting we get, D1 = (CW002+CW012) – (CW102 + CW112) – 2[X00(CW00 + CW10) +X01(CW01 + CW11)] …………………………………………...this can be written as D1= A01 – X00[2(CW00 - CW10)] +X01[2 (CW01 - CW11)] D1=A01-{X00[A011] +X01[A012]}.

(5)

Where A01= (CW002+CW012) – (CW102 + CW112), and A011= 2(CW00 - CW10), A012=2 (CW01 - CW11)] are the constants which is prior computed since the codevectors are known before the input arrives. Similarly, D2 is the distortion between the codeword CW2 and CW3, for I1 and is given as D2=A23-{X00[A231] +X01[A232]}

(6)

Further certain Boolean flags Ad01 and Bd01 are set as in Table 1 depending on the signs of D1 and D2. D1 and D2 can be processed in parallel on FPGA. The look-uptable 2 then decides which will be the 2 codewords out of the 4 codeword for the D3 calculation. Look –up-table 3 is used to set the Boolean flag Cd01 depending on the signs of D3. Table 1. Boolean Flag to set Ad01, Bd01, A010 Distortion Condition D1 Negative D1 Positive

Ad01 ‘1’ ‘0’

Distortion Condition D2 Negative D2 Positive

Bd01 ‘1’ ‘0’

Table 2. Distortion Table for 2 codeword selections in D3 calculation Ad01

Bd01

‘0’ ‘0’ ‘1’ ‘1’

‘0’ ‘1’ ‘0’ ‘1’

Selected Level 1 codewords Constants C0 and C3 A03 C0 and C2 A02 C1 and C3 A13 C1 and C2 A12

Table 3. Boolean Flag to set Cd0 Distortion Condition D3 Negative D3 Positive

Cd01 ‘1’ ‘0’

Level 2 Constants A031 A021 A131 A121

Level 3 Constants A032 A022 A132 A122

Distortion D03 A03-[X00.A031+X01.A032] A02-[X00.A021+X01.A022] A13-[X00.A131+X01.A132] A12-[X00.A121+X01.A122]

Table 4. Winner codeword look-up-table AD01

BD01

Cd01

‘0’ ’0’ ’0’ ’0’ ‘1’ ‘1’ ‘1’ ‘1’

’0’ ’0’ ‘1’ ‘1’ ‘0’ ’0’ ‘1’ ‘1’

’0’ ‘1’ ’0’ ‘1’ ’0’ ‘1’ ’0’ ‘1’

Winner Code Index I4 I2 I3 I2 I4 I1 I3 I1

Design and Implementation Issues of Parallel VQ in FPGA

371

The Euclidian distance measurement [2] is modified by using a look-up-table which enhances the speed as compared to our earlier design [2]. The winner Index is selected finally for the 4 Code words, from Table 4.

5 Results The VQ was tested for various Dimensions of the vectors, i.e. for k=2, 4, 6, 8, 12 and 16. The delays were analyzed for both the designs, one using the conventional approach and the other for the new architecture. Table 5 shows that this method gives comparison of delay as the ‘K’ values increases as compared to [4]. Our proposed method is best suitable for VQ of higher Dimensions and can proposes images of 30 frames in less than 1 second hence is suitable for real-time. Table 5. Delay Analysis for Codebook of size 4 core DEVICE v600ebg432-6 Dim ensio n ’k’ 2D 4D 8D 12D 16D

Conventio nal Method 40.100 44.144 55.093 60.994 67.841

LUT Method paper [4] 13.269 24.433 33.305 43.467 44.51

Proposed Method (ns) Max path delay 21.074 25.543 36.695 46.86 52.330

Delay estimate for a 512x512 Image (sec) 0.002 0.052 0.036 0.025

References [1] Al-Haj, A.M.: An FPGA-Based Parallel Distributed Arithmetic Implementation of the 1-D Discrete Wavelet Transform, Department of Computer Engineering, Princess Sumaya University. Informatics 29, 241–247 (2005) [2] Rasane, K., Kunte, S.R.: An Improved Reusable Component Architecture for ‘K’ dimensional 4 codebook VQ Encoder. Department of Electronics and Communication, KLESCET, Belgaum, Principal Institute, JNNEC Shivamoga, India, published in (ICEDST 2009), Manipal, India, pp. 271–274 (December 12, 2009) [3] Chen, P.Y., Chen, R.D.: An index coding algorithm for image vector quantization. IEEE Transactions on Consumer Electronics 49(4), 1513–1520 (2003) [4] Rasane, K., Kunte, S.R.: Speed Optimized LUT based ‘K’ dimensional Reusable VQ Encoder Core, Department of Electronics and Communication, KLESCET, Belgaum. In: Principal Institute, JNNEC Shivamogga, India, published in ICECT 2010, Kuala Lumpur, Malaysia, May 7-10, pp. 97–100 (2010)

Discriminative Novel Information Detection of Query-Focused Update Summarization Jinguang Chen1,3 and Tingting He1,2,* 1

Engineering & Research Center for Information Technology on Education, Huazhong Normal University, 430079 Wuhan, China 2 Department of Computer Science and Technology, Huazhong Normal University, 430079 Wuhan, China 3 School of Teacher Education, Huzhou Teachers College, 313000 Huzhou, China [email protected], [email protected]

Abstract. Current researches of query-focused update summarization often overlook influence of the query upon historical information. In this paper, we proposed a new information detection method which treats query-relevant and query-irrelevant old information in different way. Old information in the document set is removed in the process of sentence scoring stage instead of in additional stage after the sentences scored. Experiment results on TAC 2009 shows effectiveness of our method. Keywords: Update Summarization; Historical information removal; Novelty detecting; Query-focused Update Summarization.

1 Introduction Novel Information Detection was firstly introduced by TREC in 2002[1]. The basic task is as follows: given a topic and an ordered set of documents segmented into sentences, return sentences that are both relevant to the topic and novel given what has already been seen. This task models an application where a user is skimming a set of documents, and the system highlights new, on-topic information. The Text Analysis Conference (TAC1) is one of the most well-known series of workshops that provides the infrastructure necessary for large-scale evaluation of natural language processing methodologies and technologies. The update summarization task of TAC is introduced since 2008, which aims to generate two short and fluent summaries respectively for two chronologically ordered document sets to meet the topic-relevant information need. The summary of the second document set should be written under the assumption that the user has already read the earlier documents and should avoid repeating old information and inform the user of novel information about the specific topic. *

1

Corresponding author. http://www.nist.gov/tac

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 372–377, 2011. © Springer-Verlag Berlin Heidelberg 2011

Discriminative Novel Information Detection of Query-Focused Update Summarization

373

Fig. 1. Problem caused by ignoring query in query-focused update summarization

Previous mainstream approaches to novel information detection include cosine filtering, unseen words, WordNet synset, graphic, as well as clustering method, etc. The cosine filtering method[2] calculates the cosine similarity between a sentence and each previous sentence. The maximum similarity was used to eliminate redundant sentences. Sentences with a maximum similarity greater than a preset threshold were treated as redundant sentences. Alternatively, the unseen words method[3] considered previously unseen words as an evidence of Novelty. The WordNet synset method[4] expanded all noun phrases in a sentence using wordnet and used corresponding sysnsets for novelty comparisions. The graphic method[5] models all sentences as a graph. The topic description is deemed as the only labeled node and assigned with an initial score, and then the scores of all the sentences in the documents are learned by spreading the initial score on the graph. The clustering method[5] divides all sentences in one topic into several clusters, and selects sentence which related to all these clusters. However, the query-focused aspect of update summarization, i.e., removing old information while keeping query-focused, is rarely concerned in the current researches. Fig. 1 illustrates problems caused by ignoring query when generating queryfocused update summarization. Although both area 1 and 2 belong to old information with no regard for the query, they should be treated discriminatively when considering influence of the query. Contents which are both irrelevant to the query and belong to old information which are marked as area 2 are the right contents which should be removed. Since both document set A and B are query-related, they inevitably have some similar information revolves around the query marked as area 1. However, this part of information is very important for generating query-focused summarization and should be removed with discretion. If all sentences in area 1 are removed arbitrarily, summary of set B may turn out to be irrelevant to the query and under-performing. To solve this problem, we employ discriminative novelty detection techniques in this paper. Experiment results in participating TAC 09 show effectiveness of our method. The rest of the paper is organized as follows. Section 2 gives a brief introduction of basic summarization system, Section 3 introduces our proposed discriminative novelty detection method, section 4 presents experiments and evaluation results, and section 5 concludes with discussions and future research directions.

374

J. Chen and T. He

2 The Summarization System The basic query-focused update summarization consists of two main steps: 1. semantic sentence similarity obtaining; 2. sentence scoring. 2.1 Semantic Similarity Obtaining A log-linear part-of-speech tagger is used to extract effective words from sentences, including nouns, verbs and adjectives. After that Sentences are transferred into a vector of effective words. Considering two sentences: Si = {wi1 , wi 2 , ⋅⋅ ⋅, wik , ⋅ ⋅ ⋅, wim }, S j = {w j1 , w j 2 , ⋅⋅ ⋅, w jl , ⋅ ⋅ ⋅, w jn }

(1)

Score of words in Si and Sj can be obtained as the following: WordnetSimScore( wik ) = max{sim( wik , w jl | l = 1,2, ⋅ ⋅⋅, n)} WordnetSimScore( w jl ) = max{sim( w jl , wik | k = 1, 2, ⋅ ⋅ ⋅, m)}

(2)

Where sim( Si , S j ) is the similarity of two effective words which obtained by using WodNet synset[6,7] If synset of word wik in WordNet is U and synset of word wjl in WordNet is V, similarity of them can be obtained as the following: sim( wik , w jl ) =

U ∩V U ∪V

(3)

If a word can not be found in WordNet, its synset is itself. Similarity of sentence Si and Sj can be obtained as the following: 1 m 1 n sim( Si , S j ) = ( i∑WordnetScore( wik ) + i∑WordnetScore( w jl )) / 2 m k =1 n l =1

(4)

2.2 Sentence Scoring The basic system selects sentences by using a feature fusion method to identify the sentences with high query-relevant and high information density, i.e., more relevant a sentence is to the query and more important a sentence is in the document set, more likely the sentence is to be included in the final summary. First, we score every sentence’s representative feature by obtaining its importance in document set. The query-independent score of sentence S can be obtained as following: N

QI ( S ) = (∑ sim( S , Si )) / N

(5)

i =1

where sim( S , S j ) is semantic similarity as difined in Equ. 4. In Equ. 5, similarity to sentence S itself is also involved.

Discriminative Novel Information Detection of Query-Focused Update Summarization

375

Secondly, we re-score every sentence’s query-focused feature by obtaining its similarity with query. QF ( S ) = sim( S , Query )

(6)

The final score of sentence S is obtained by linear combination: Score(S)=σ i

QF ( S ) N

∑ QF (S ) i

i =1

+(1-σ )i

QI ( S ) N

(7)

∑ QI (S ) i

i =1

where σ is the parameter to adjust the proportion of the two parts.

3 Discriminative Novelty Detection Method In order to detect novel information from set B, we proposed a discriminative sentence scoring method. Intuition of this method is: Old information in the document set should be removed in the process of sentence scoring stage instead of in additional stage after the sentences scored. In the sentence scoring stage (as described in Equ. 7), old information should just be considered in the query-independent part, instead of both query-independent and query-focused part. Our sentence scoring method in set B can be defined as the following: Score(S)=σ i

QF ( S ) N

∑ QF (S ) i =1

i

+(1-σ )i Novelty ( S )i

QI (S ) N

∑ QI (S ) i =1

(8)

i

Where Novelty(S) is the novelty factor influenced by the degree of how sentences S related to the old information, it can be obtained as the following: Novelty ( S) = (1 − Maxsim( Si , old _ content ))

(9)

Where “old_content” is the document set A. “Maxsim” computes all the similarity values between Si and each sentence in the old_content, and returns the highest one[5]. In Ref. 5, the cosine similarity is adopted. Unlike them, we adopted the semantic similarity as described in Equ. 5. In Equ. 9, although there have been no change on the query-focused part, but its influence increases with the decrease of query-focused part when sentence S contains old information, i.e., Novelty ( S ) < 1.0 , which prevent the summarization system to remove sentences closely related to the query.

4 Experimental Results In participating TAC 2009, we submitted 2 results, ID 53, 14, respectively. To compare the effectiveness of different methods, ID 53 used the proposed discriminative

376

J. Chen and T. He

novelty detection method (DN henceforth), while ID 14 (BN henceforth) used the same summarization system as ID 53 with the only difference being that its novelty detection method is defined as the following: Score(S)=(σ i

QF ( S ) N

∑ QF ( Si )

+(1-σ )i

i =1

QI ( S ) N

∑ QI ( Si )

)i Novelty ( S )

(11)

i =1

Like most mainstream novelty detection method[1-5], BN treat the old information in an additional independent stage. Table 1 shows results of DN and BN in manual evaluation of TAC 2009 Update Summarization Task (set B). In Table 1, we can see that DN performs far better than BN when evaluated by modified pyramid score, average numSCUs, macroaverage modified score with 3 models, as well as the average overall responsiveness. Table 1. Performance of DN and BN in TAC 2009 Update Summarization Task (set B) Manual Evaluation

Average modified (pyramid) score

Average numSCUs

Macroaverage modified score with 3 models

Average overall responsiveness

BN

0.14

2.18

0.14

3.32

DN

0.24

3.50

0.23

4.07

In Figure 2, we can also see that DN performed far better than BN when evaluated by automatic evaluation metrics like ROUGE-2[8]. More importantly, since BN is outside the 95% confidence interval of DN when evaluated both by ROUGE-2 and BE, we can see that differences between BN and DN are significant, which implies effectiveness of our purposed method.

Fig. 2. ROUGE-2 Recall score with 95%confidence interval

Discriminative Novel Information Detection of Query-Focused Update Summarization

377

5 Conclusions This paper presented a new method of detecting novel information of query-focused update summarization. Unlike most current researches, we adopt differentiated treatment to the old information according to the degree of relevance between the historical information and the query. Experiment results on TAC 2009 show effectiveness of our purposed method. Although the proposed approach is simple, we hope that this novel treatment could inspire new methodologies in progressive summarization. For the future works, we plan to further validate effectiveness of our method in other benchmark large scale corpuses.

Acknowledgements This work was supported by the National Natural Science Foundation of China (No. 60773167), the Major Research Plan of National Natural Science Foundation of China (No. 90920005)t, the 973 National Basic Research Program (No. 2007CB310804), the Program of Introducing Talents of Discipline to Universities (No. B07042), and the Natural Science Foundation of Hubei Province (No. 2009CDB145), Chenguang Program of Wuhan Municipality (No. 201050231067).

References 1. 2.

3.

4.

5. 6. 7. 8.

Harman, D.: Overview of the TREC 2002 Novelty Track. In: The 11th Text Retrieval Conference (TREC 2002). NIST Special Publication 500-251, Gaithersburg (2002) Abdul-Jaleel, N., Allan, J., Croft, W.B., Diaz, F., Larkey, L., Li, X.Y.: Umass at TREC 2004, Novelty and Hard. In: The Thirteenth Text Retrieval Conference (TREC 2004). NIST Special Publication 500-261, Gaithersburg (2004) Schiffman, B., McKeown, K.R.: Columbia University in the Novelty Track at TREC 2004. In: The Thirteenth Text Retrieval Conference (TREC 2004). NIST Special Publication 500–261, Gaithersburg (2004) Eichmann, D., Zhang, Y., Bradshaw, S., Qiu, X.Y., Zhou, L., Srinivasan, P., Kumar, A., Wong, H.: Novelty, Question Answering and Genomics: The University of Iowa response. In: The Thirteenth Text Retrieval Conference (TREC 2004). NIST Special Publication 500-261, Gaithersburg (2004) Li, S.J., Wang, W., Zhang, Y.W.: TAC 2009 Update Summarization with Unsupervised Methods. In: Text Analysis Conference (TAC 2009), Gaithersburg (2009) Miller, G.A.: WordNet: A Lexical Database for English. Communications of the ACM 38(11), 39–41 (1995) Fellbaum, C.: WordNet: An Electronic Lexical Database. MIT Press, Cambridge (1998) Lin, C.Y., Hovy, E.: Automatic Evaluation of Summaries Using N-gram Co-occurrence Statistics. In: NLT-NAACL, Edmonton, Canada, pp. 71–78 (2003)

Visualization of Field Distribution of the Circular Area Based on the Green Function Method Gang Wu1 and Xiying Fan2 2

1 College of Telegraph, Pan Zhihua University College of Foreign Languages, Pan Zhihua University 617000 Pan Zhihua, China [email protected]

Abstract. The Green Function method is one of the basic methods of studying the theory of electromagnetic field. This paper, starting from the Green formula, through the establishment of the integral expression about harmonic function on the plane, draws forth Green Function on the plane, finds the Green Function formula of the circular area (mathematical model) through the “image method”, and finds potential function of any given point of the circle area, then programs the Green Function using Matlab, and finally achieves visualization of f Green unction of the electric charge of the electric field inside and outside the circular area. Keywords: Circular area, the Green Function; visualization.

1 Introduction The introduction of Green Function to solve problems of electromagnetic field is of great importance. As equations of field are linear ones, any field source distribution can be decomposed into a collection of point sources, an arbitrary source distribution in a given condition generates a field, which is equal to overlapping of the distribution of the field generated by these point sources in the same boundary conditions, and therefore, after attaining the field of point sources in the given condition (the Green Function), which can be used to seek the field of arbitrary distribution of field source in the same boundary conditions, visualization of the field can be achieved through computer simulation [1] [2] [3].

2 The Establishing of Green Function about Circular Area 2.1 The Establishment of Integral Expression about Plane Harmonic Functions In order to establish integral expressions of plane harmonic function, where the introduction of plane Green's formula is as follows[4]:

∫∫ (v∇ u + u∇ 2

D

2

v ) dσ = ∫ ( v Γ

∂u ∂v − u ) ds ∂n ∂n

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 378–384, 2011. © Springer-Verlag Berlin Heidelberg 2011

(1)

Visualization of Field Distribution of the Circular Area

Supposing M 0 ( x0 , y 0 ) is arbitrary point in region D,

v = ln

379

v is the following expression:

1 = − ln r = − ln ( x − x0 ) 2 + ( y − y 0 ) 2 r

(2)

v meets the condition ∇ 2 v = 0 everywhere except at M 0 , thus forming the neighborhood with M 0 as the center and r as the radius, whose perimeter is Γe , v is harmonic function in D − De , Thus plane Green's formula of formula (1) is as So

follows:

1 ∂ (ln ) r )ds ∂n

1 1 1 ∂u (ln ∇ 2 u − u∇ 2 ln )dσ = ∫ (ln −u ∫∫ Γ + Γe r r r ∂ n D − De In

2 D − De , ∇ ln

(3)

1 = 0 , on perimeter Γe there are the following expressions [5]: r ∂ 1 1 1 (ln ) = = ∂n r r ε

∫

Γe

ln

∫

Γe

(4)

1 ∂u 1 ∂u ds = 2πε ln ( ) r ∂n ε ∂n

u

(5)

∂ 1 1 (ln )ds = 2πε u ∂n r ε

(6)

Formula (5), (6) are obtained through middle value theorem,

n is normal outward,

∂u indicates the value in a point of De , take (5, (6) into (3), when ε → 0 , here ∂n 1 1 ∂u ε ln → 0 , So we comes 2πε ln ( ) → 0 , D − De → D , u → u ( M 0 ) , ε ε ∂n obtain the following expression through the transformation of (3)

1 u (M 0 ) = 2π When

1 ∂ ln 1 ∂u 1 r ∫Γ (ln r ∂n − u ∂n )ds − 2π

1

∫∫ ln r ∇ udσ 2

(7)

D

u is the harmonic function, then ∇ 2 u = 0 , we get 1 u(M 0 ) = 2π

1 ∂ ln 1 ∂u r ∫Γ (ln r ∂n − u ∂n )ds

(8)

380

G. Wu and X. Fan

Expression (8) is the integral expression of harmonic function established on twodimensional plane. 2.2 Green Function on the Plane In formula (1), because

u, v are both harmonic function, so ∂u

∂v

∫ (v ∂n − u ∂n )ds = 0 Γ

(9)

with (8) - (9) and obtains the following expression when satisfying the condition of

vΓ =

1 1 ln 2π r Γ . u (M 0 ) = − ∫ u Γ

∂ 1 1 ( ln − v)ds ∂n 2π r

(10)

So plane Green's function is defined as[6]:

G (M , M 0 ) =

1 1 ln − v 2π r

(11)

From the above expression it is known with the known expression (11), the potential of M 0 can be obtained through (10), whose expression is as follows:

u(M 0 ) = −∫ u(M ) Γ

∂G ds ∂n

(12)

2.3 Green Function of the Circular Area In order to get Green Function of circular domain, here assuming there is an infinitely long cylindrical of grounding conductor with radius a in space, with an infinitely

long wire of the line charge density ε 0 parallel to the conductor in the column. Because of symmetry of the column , any electric potential in cylindrical coordinates ( ρ , ϕ , z ) has nothing to do with the coordinates z in column. So the potential problem of any a point in the column can be solved by transforming it into a twodimensional circular domain problem. The potential of any point in the cross section

①the wire of line charge density ε point of the circle ; ②the image charge of charge density

of the circular field is composed of two parts:

0

M0 λ from the potential in M 1 outside the circle, in the circle, the potential of an arbitrary observation point M is the superposition of the potential of the original charge from the potential in

in the column and the potential of image charge outside the column [7], using Gauss's Law [8], Obtaining the Green's function expression[4]:

Visualization of Field Distribution of the Circular Area

381

Fig. 1. Connection diagram of circular area

G=

1 1 λ 1 ln + ln +C 2π M 0 M 2π M 1 M

(13)

，

OM 0 = ρ 0 , OM 1 = ρ1 , OP = R , OM = ρ ∠MOP = θ , When the observation point M moves to point P on the circle, from the given condition G Γ = 0 , it can be on the border [9]: Supposing

−

1 ln[ R 2 + ρ 02 − 2 Rρ 0 cos(γ − θ )] 4π

λ − ln[ R 2 + ρ12 − 2 Rρ12 cos(γ − θ )] + C = 0 4π Reorganize(14) after getting differential about

(14)

θ , and obtain：

⎧λ = −1 ⎪ R2 ⎨ ρ = ⎪ 1 ρ 0 ⎩

(15)

So Green's function of the first boundary value is got in the circular area:

G= +

1 1 ln 2π R ρ 2 − 2 ρρ cos γ + ρ 2 0 0 2

(16) 2

1 R R ln ρ 0 ρ 2 − 2 ρ ( ) cos γ + ( ) 2 2π ρ0 ρ0

382

G. Wu and X. Fan

In Formula (16), the first is the potential function of the original charge generation in the circular, the second is the potential function of image charge generation outside the circular.

3 The Potential Function of the Circular Area Through Formula (12), Dirichlet problem[10]

⎧⎪∇ 2 u = 0 ⎨ ⎪⎩u Γ = f ( R, θ ) in polar coordinates, function is:

u (M 0 ) =

(17)

ds = Rdθ in Formula (12), then the expression of the potential 1 2π

∫

2π

0

f ( R, θ )

R2 − ρ 2 R 2 + ρ 2 − 2 Rρ cos(γ − θ )

dθ

(18)

In (18), in the given boundary conditions, the potential function of the circular can be obtained. 6

4

2

0

-2

-4

-6 -2

0

2

4

6

8

10

Fig. 2. Electric Field of Charge Outside the Circular Area

4 Visualization of the Field Distribution Expression (16) is the Green Function expression of the circular area. To draw the visual graphics outside the circular area,

ρ > R , ρ1 > R

is required in expression

r (16). Here we take R = 1, ρ1 = 2 , and take OA as the direction of the polar axis, γ is included angle of OM with OM 1 , using MATLAB to carry out program design in expression (16) [12], using contour () sentences to draw contour while programming,

Visualization of Field Distribution of the Circular Area

using the formula

383

r E = −∇ϕ to get the vector of electric field intensity[10], using

sentence gridient () to draw their electric line of force, the result of which is shown in Figure 2[11]. As the expression within the Circular area is still (16), using the same method above programming, we can draw the figure of Green Function within the circular region, but ρ < R, ρ 1 < R is required in expression (16), here R = 2, ρ1 = 1 . Revision of the program design outside the Circular area can be made to get the visual graphics as in Figure 3 [11].

2 1.5 1 0.5 0 -0.5 -1 -1.5 -2 -2.5

-2

-1.5

-1

-0.5

0

0.5

1

1.5

2

2.5

Fig. 3. Electric Field of Charge inside the Circular Area

5 Conclusion As can be seen from the above discussion, the Green Function is scalar potential function of unit point source (point charge) in certain conditions and the key to potential function is to determine the Green Function based on the Green Function Method. When the Green Function is given, the potential function of the source of a random distribution may be obtained through integral. In the interactive Matlab working environment [12], the obtained mathematical expression of scalar potential function can be programmed to get accurate data of potential function and graphical visualization in the Circular region, so the abstract concept of electric field can be transformed into visual data and graphics[13]. The above research ideas can be widely applied in research on the complex electromagnetic field.

，

Acknowledgment Subsidized by Research Project of Education and Teaching Reform of Pan zhihua University (JJ0825&JJ0805).

384

G. Wu and X. Fan

References 1. Wu, G., Fan, X.: Visualization of Potential Function of the Spherical Region Based on the Green Function Method. In: The International Conference on E-Business and EGovernment (iCEE 2010), pp. 2595–2598. IEEE Press, Los Alamitos (2010) 2. Wu, G.: Discussion of the Electromagnetic Field Based on the Green Function and Dyadic Green Function. Journal of Panzhihua University 6, 38–44 (2008) 3. Feng, Z.: Dyadic Green Function of Electromagnetic Field and Application. Nanjing Institute of Electronic Technology, Nanjing (1983) 4. Wang, J., Zhu, M., Lu, H.: Electromagnetic Field and Electromagnetic Wave. Xi’an Electronic Science and Technology University Press, Xi’an (2003) 5. Wang, Y.: Mathematical Physics Equations and Special Functions. Publishing House of Higher Education, Beijing (2005) 6. Liang, K.: Methods of Mathematical Physics. Publishing House of Higher Education, Beijing (2003) 7. Yao, D.: Study Guide of Mathematical Physics Methods. Science Press, Beijing (2004) 8. Xie, X., Yuan, X., Zhang, T.: University Physics Tutorial (2000) 9. Yang, H.: Mathematical Physics Method and Computer Simulation. Electronic Industry Press, Beijing (2005) 10. Wang, J., Zhu, M., Lu, H.: Study Guide of Electromagnetic Field and Electromagnetic Waves. Xi’an Electronic Science and Technology University Press, Xi’an (2002) 11. Peng, F.: MATLAB Solution to Equations of Mathematical Physics and Visualization. Tsinghua University Press, Beijing (2004) 12. Li, N., Qing, W., Cao, H.: Simple Tutorial of MATLAB7.0. Tsinghua University Press, Beijing (2006) 13. Li, L., Wang, J.: Electromagnetic Field Teaching Using Graph Based on Matlab. Journal of Xiao Gan University 5, 120–121 (2006)

Efficient Genetic Algorithm for Flexible Job-Shop Scheduling Problem Using Minimise Makespan Hamid Ghaani Farashahi1, B.T.H.T. Baharudin1, Shahed Shojaeipour2, and Mohammad Jaberi3 1

Department of Mechanical & Manufacturing Engineering, Faculty of Engineering Univerisiti Putra Malaysia 43400, Serdang, Selangor Malaysia 2 Dept. of Mechanical & Materials Engineering, Faculty Engineering & Built Environment Universiti Kebangsaan Malaysia 43600, Bangi, Selangor Malaysia 3 School of Mechanical Engineering Iran University of Science and Technology Tehran, Iran [email protected], [email protected], [email protected], [email protected]

Abstract. The aim of this paper is to minimise the makespan. The flexible jobshop scheduling is very common in practice and parallel machine is used in jobshop environment as flexibility. These flexibilities could be used for increasing the throughput rate, avoiding the production stop, removing bottleneck problems and finally achieving competitive advantages in economical environments. In opposition to classic job-shop where there is one machine in each stage, in this problem production system consists of multistage which in each stage there are one or several parallel machines with different speeds. The flexible job-shop scheduling with parallel machines problem consists of two sub-problems for solving: assigning and sequencing sub-problem. Keywords: Flexible Job-Shop, NP-Hard, Genetic Algorithm, Makespan, Scheduling Problem, Heuristic Methods.

1 Introduction The theory of scheduling is about optimal allocation of scarce resources including machines, robots, processors, and operators, for activities over time, where the objective is the optimization of one or several performance criteria. Research on scheduling started five decades ago, initiated by [1], [2]. The main aim in scheduling theory is to decrease time based on to optimize one or several performance measures because time is a limited resource. Hence, all activities need to be scheduled consciously or unconsciously, in order to use this limited resource in optimal procedure [3]. R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 385–392, 2011. © Springer-Verlag Berlin Heidelberg 2011

386

H.G. Farashahi et al.

In a classic job-shop scheduling problem each job has a fixed and distinct routing that is not necessarily the same for all jobs. There is only one routing for each job and this refers to the lack of flexibility in this environment. The job-shop scheduling problem with parallel machine is a kind of flexible job-shop scheduling problem in which the numbers of machines are more than one in at least one stage with different speed rate. Researching the literature showed that one of the shortcomings in the classic jobshop scheduling problem research is that, there is only one machine for processing each operation and it means that there is only one feasible process routing for each job, so that there is no flexibility in this environment [4]. The problem to be investigated in this research is a critical extension of the traditional problem of job-shop scheduling, where each operation can be processed on a set of parallel machines with different speeds (uniform machine). A lot of researchers studied about flexible jobshop, but most of them are limited to situation that the speed on each machine are the same or the processing time over all machines are the same. The problem of job-shop scheduling with uniform machines accounts for an important problem that is often met in current practice of scheduling in manufacturing systems because a method or algorithm that be able to solve uniform machine also can be used to identical parallel machine. The main objective in this research is to minimise the maximum completion time (makespan). Pinedo shows that “a minimum makespan usually implies a high utilization of the machine(s)” [5]. The usage for bottleneck and near bottleneck equipment is directly associated with the system’s throughput rate. Hence, reduction of makespan should cause a higher throughput rate as well [6]. The scope of this research is to optimize the job sequences and assigning jobs to a machine in the job-shop environment based on to exist a set of uniform machine for processing each operation.

2 Definition of the Scheduling Problem The job-shop scheduling problem with uniform machines is defined as follows. There are n jobs and m stages. Each job Jj , j=1,2,…,n , includes a chain of operations (Oj,1, Oj,2,…, Oj,m) and without loss of generality, it could be assumed that the sequence of operation is the same. The number of operations to complete each job could be smaller than or equal to the number of stages. In this research, the number of operations for each job is supposed to be equal to the number of stages. At every stage v, v=1,2,…, m, there are lv ≥ 1 uniform machines and at least at one stage the number of machines must be more than one machine lv >1. Each job has a different routing. Every job must pass all stages according to its routing and is to be processed by exactly one machine at every stage. Whenever an operation Oj,i has been finished, the successor operation Oj,i+1 may start. The speed of machine Mv,r is Sv,r, Sv,r≥1, r=1,2,…, lv, v=1,2,…,m. The processing time of operation Oj,i on machine with one time unit speed (Sv,r=1) is consider Pj,i. If operation Oj,i is performed on machine with Sv,r>1 then the processing time will be decreased to Pj,i ⁄ Sv,r .The completion time of operation Oj,i is denoted Cj,i. The research objective is to minimise the maximum completion time Cmax = Max 1≤ j ≤ n , 1≤ i ≤ m Cj,i. The multistage flexible job-shop scheduling

Efficient Genetic Algorithm for Flexible Job-Shop Scheduling Problem

387

problem with v stages and lv uniform machine at each stage v, v=1,2,…,m, is denoted as follows : FJQm ║ Cmax. The job-shop scheduling problem with uniform machine included two subproblems; routing sub-problem and sequencing sub-problem. In routing sub-problem, any operation is allocated to one machine which is capable of processing and in sequencing sub-problem, the sequencing of operations is determined. Two types of approaches have been applied for solving these two sub-problems: hierarchical approaches and integrated approaches. Since the problem is NP-hard, for solving the research problem, five heuristic approaches based on priority dispatching rules and a genetic algorithm are proposed to give near optimal solution in acceptable amount of time.

3 Heuristic Methods 3.1 ECT Procedure This procedure is based on the earliest completion time (ECT) rule that is utilized for solving this scheduling problem. This heuristic is able to solve the problem integrated, it means that, the information of jobs and machines are used simultaneously. The steps for this method are presented as follow. Step 0: Create set M with the first operation of all jobs as follow: M = {Oj,1│1≤ j ≤ n }

(1)

Step 1: Calculate cj,i for each operation of set M as follow: cj,i = min{max(availv,r ,Cj,i-1 )+(Pj,i / Sv,r ) , r = 1,2,…,lv}

(2)

Step 2: Choose the first operation (O*j,i) with the minimum completion time (cj,i) from set M , and schedule it on the first machine that completed with the earliest completion time. O*j,i = {Oj,i ∈ M │ cj,i = min { cj,i ∈ M}}

(3)

Step 3: Remove scheduled operation O*j,i from set M. M = M \ {O*j,i}

(4)

Step 4: If there is operation O*j,i+1 add it to set M. M = M ∪ {O*j,i+1}

(5)

Step 5: Return to step 1 until a complete schedule is generated. 3.2 SPT-ECT Procedure This procedure is a hierarchical approach. It means assignment of operations to machines and sequencing of operations are solved separately. This heuristic is based on the shortest processing time (SPT) rule and ECT rule. At first, SPT rule is used for finding the sequence of operations and for this purpose, virtual weight is defined based on the average speed of all machines at each stage for all operations. After that,

388

H.G. Farashahi et al.

ECT rule is used for solving assigning sub-problem. The SPT-ECT procedure follows the steps: Step 0: The virtual weight is calculated by the following equation for each operation: ,

∑

(6)

,

Step 1: Compute the weighted processing time for each operation as follow: P′j,i = Pj,i ⁄ wj,i

(7)

Step 2: Create set H with the first operation of all jobs as follow: H = {Oj,1│1≤ j ≤ n}

(8)

Step 3: Create set M with operation(s) from set H which have the shortest weighted processing time. M = {Oj,i ∈ H │ P′j,i = min { P′j′,i′ │ Oj′,i′ ∈ H}}

(9)

Step 4: Run the step 1 and step 2 from ECT procedure. Step 5: Remove scheduled operation O*j,i from set H. H = H \ {O*j,i }

(10)

Step 6: If there is operation O*j,i+1 add it to set H. H = H ∪ {O*j,i+1}

(11)

Step 7: If H ≠ Ø return to step 3, otherwise terminate 3.3 LPT-ECT Procedure This heuristic is a hierarchical approach and is based on the longest processing time (LPT) rule and ECT rule. The difference between this procedure and previous one is only in step 3 that is proposed as follow: Step 3: Create set M with operation(s) from set H which have the longest weighted processing time. M = {Oj,i ∈ H │ P′j,i = max { P′j′,i′ │ Oj′,i′ ∈ H }}

(12)

3.4 MWKR-ECT Procedure This procedure is a hierarchical approach and based on the most work remaining (MWKR) rule and ECT rule. Firstly, MWKR rule is used for obtaining the sequence of operations and for this purpose virtual weight is defined based on the average speed of all machines at each stage for each operation. After that, ECT rule is used for solving assigning sub-problem. The steps for this procedure are proposed as follows: Step 0: The virtual weight is calculated by the following equation for each operation: ,

∑

,

(13)

Step 1: Compute the weighted processing time for each operation as follow: P′j,i = Pj,i ∕ wj,i

(14)

Efficient Genetic Algorithm for Flexible Job-Shop Scheduling Problem

Step 2: Create set H with the first operation of all jobs as follow: H = {Oj,1│1 ≤ j ≤ n}

389

(15)

Step 3: Calculate weighted remaining processing time for each operation of set H as: WR′j= P′j,i + P′j,i +1+…+ P′j,m (16) Step 4: Create set M with operations from set H with most weighted remaining work. (17) M = {Oj,i ∈ H │ WR′j = max { WR′j′ │ Oj′,i′ ∈ H}} Step 5: Run the step 1 and step 2 from ECT procedure. Step 6: Remove scheduled operation O*j,i from set H. H = H \ {O*j,i }

(18)

Step 7: If there is operation O*j,i+1 add it to set H. H = H ∪ {O*j,i+1} Step 8: If H ≠ Ø return to step 3, otherwise terminate.

(19)

3.5 LWKR-ECT Procedure This procedure is based on the least work remaining (LWKR) rule and ECT rule. This method is similar to MWKR-ECT procedure except for step 4 that is presented as follow. Step 4: Create set M with operation(s) from set H with the least weighted remaining work. M = {Oj,i ∈ H │ WR′j = min {WR′j′ │ Oj′,i′∈ H}}

(20)

4 Experimental Results To evaluate the performance of the proposed heuristic methods, each of the problem instances consists of the following parameters: number of jobs, number of stages, number of machines per stage, range of processing time for each operation and speed of each machine. The levels of these parameters are shown in Table 1 and follow [7]. U[x,y] denotes a discrete uniform distribution between x and y. Table 1. Parameter levels for generation random data Parameters Number of jobs Number of stages Number of machines per stage Processing time Speed of machines

Level 20-50-100-150-200-300 2-4-6-8-10 U[1,5] U[1,40] U[1,4]

By applying these parameters level, there are (6×5×1×1×1) 30 different types of test scenario. For each number of jobs and operations combination, 10 random problem instances are generated with the same parameters. Thus, 300 problem instances are generated to compare the performance of each proposed heuristics.

390

H.G. Farashahi et al.

To consider the effectiveness and efficiency of the five proposed heuristics, every heuristic ran on the same randomly generated problems (300). The experiments were carried out for minimising makespan over all jobs begins scheduled. To evaluate the performance of the proposed heuristic methods, factor “loss” is used as a merit figure. This factor is described below in which makespan denotes the maximum completion time obtained from each heuristic method separately for each problem instance. However, the lower bound is fixed for each problem instance and is independent of the using heuristic method to solve the problem instance [8]. (21) Moreover, the best and the average of loss factor and the best and the average result of makespan obtained from running in 10 replications each test scenario are used for comparison of the proposed heuristics. Table 2 shows the computational results of all 300 problem instances by running ECT, SPT-ECT, LPT-ECT, MWKR-ECT and LWKR-ECT rules. The best and the average of makespan and the best, the average and standard deviation (s.d.) of loss were reported by taking the mean over 300 problem instances in Table 2. Table 2. Summary of experimental result of all 300 problem instances Heuristic ECT SPT-ECT LPT-ECT MWKR-ECT LWKR-ECT

Average 0.245 1.186 1.619 0.198 2.474

Loss Best 0.168 0.904 1.052 0.072 1.641

s.d. 0.056 0.197 0.453 0.093 0.675

Makespan Average Best 3490.8 3329.7 6454.5 5685.2 7663.4 5951.7 3472.4 3111.2 10347.8 7652.3

Number of times minimum 106 0 1 193 0

Summary statistic results over all the randomly generated data are compared in Table 2. These values indicate that MWKR-ECT rule achieved the minimum of makespan up to 65% of all problem instances in comparison with other heuristics. Based on this data, the MWKR-ECT and ECT rules emerge to be the best for most of all the problem instances and LWKR-ECT, LPT-ECT and SPT-ECT performs worst in the most of all these cases respectively. The performance of standard deviation of loss indicated that the ECT rule is better than all in general.

5 Proposed Genetic Algorithm For designing the overall structure of the genetic algorithm for a given problem, before anything, a way must be defined for representation of each solution in form of chromosome. Then a set of chromosomes must be considered as the initial population which in fact represents a set of solutions to the problem. After this stage, it must be aimed to generate new chromosomes called offspring through the utilization of the genetic operations. After generation of some new chromosomes, best chromosomes must be selected in a way that the number of the elite chromosomes becomes equal to the initial population. The process of selection is according to the amount of fitness of

Efficient Genetic Algorithm for Flexible Job-Shop Scheduling Problem

391

each chromosome. So far, one iteration has been completed through one generation of the algorithm. The algorithm converges toward an optimal or sub-optimal solution after completion of several generations and will end upon fulfilling the algorithm’s termination criterion. The termination criterion can be set based on the maximum computation time, completion of certain number of iteration, which are set in advance, or due to no changes in several successive iterations of the algorithm, or other specific conditions [9]. The major factors in designing a GA are: chromosome representation, initializing the population, an evaluation measure, crossover, mutation, and selection strategy. Also, the genetic parameters must be specified before running the GA, such as population size (pop_size), number of generation (max_gen), probability of crossover (Pc) and probability of mutation (Pm) [10]. The performance of the proposed genetic algorithm with reinforced initial population (GA2) (with solution of the ECT and MWKR-ECT) is compared against itself with a fully randomized initial population (GA0).

6 Comparison Results between GA0 and GA2 Following [8], to generate random problem, each of the problem instance can be characterized by six parameters: number of jobs, number of stages, number of machines per stage, whether the number of machines per stages is variable or constant, range of processing time and range of machine’s speed. The levels of these parameters are shown in Table 3. Table 3. Parameter levels for generate random data Parameters Machine distribution Number of machines Number of jobs Number of stages Processing time Speed of machines

Levels Constant 2–6

Variable U[1,4] - U[1,6] 6

Constant 2 – 10

Variable U[1,4] - U[1,10] 30 – 100

2–4–8 U[50,70] - U[20,100] U[1,4]

By applying these parameters level of the problems, there are (4×3×3×2×1) 72 different types of test scenario. For each test scenario, 10 random problem instances are generated with the same parameters. Thus, 720 problem instances are generated to compare the performance of proposed GA. Table 4 shows the results obtained from algorithms GA0 and GA2, on 720 problem instances. This table includes the average amount of the loss parameter and the average amount of makespan as well as the number of times each algorithm has managed to find the optimal solution. As you can observe in all cases GA2 is superior which shows the better performance of the proposed genetic algorithm with the reinforced initial population compared to genetic algorithm with fully randomized initial population (GA0).

392

H.G. Farashahi et al. Table 4. Summary of experimental result of all 720 problem instances

ProblemSize Small Medium Large Average

Average of makespan GA0 GA2 195.2 193.1 665.3 642.1 2070.7 2022.6 977.0 952.6

Average of loss GA0 GA2 0.125 0.113 0.190 0.062 0.141 0.018 0.152 0.064

Number of times minimum GA0 GA2 164 227 115 222 125 221 404 670

7 Conclusion The three hundred problem instances have been randomly generated and taken from literature. Results over all instances indicated that the MWKR-ECT and ECT achieved the minimum of makespan up to 65% and 34% of all instances in comparison with other proposed heuristic procedures. Then, the genetic algorithm has been presented for resolving the research problem in other to find the best solution. It has been shown that proposed genetic algorithm with a reinforced initial population, via obtained solutions from MWKR-ECT and ECT has better efficiency compared to a proposed genetic algorithm with fully random initial population. In future work, it would be proposed to development of heuristic algorithms based on the solving the assignment sub-problem and then comparison of those algorithms with algorithms which solve the sequencing sub-problem at first prior to the assignment sub-problem.

References 1. Johnson, S.M.: Optimal two- and three-stage production schedules with setup times included. Naval Research Logistics Quarterly, 61–68 (1954) 2. Bellman, R.: Mathematical aspects of scheduling theory. Journal of Society of Industrial and Applied Mathematics 4, 168–205 (1956) 3. Sule, D.R.: Industrial Scheduling. PWS Publishing Company, Park-Plaza (1997) 4. Kim, Y.K., Park, K., Ko, J.: A symbiotic evolutionary algorithm for the integration of process planning and job shop scheduling. Computers & Operations Research 30, 1151– 1171 (2003) 5. Pinedo, M.L.: Scheduling: Theory, Algorithms, and Systems, 3rd edn. Springer Science + Business Media, LLC, New York (2008) 6. Cochran, J.K., Horng, S., Fowler, J.W.: A multi-population genetic algorithm to solve multi-objective scheduling problems for parallel machines. Computers & Operations Research 30, 1087–1102 (2003) 7. Nowicki, E., Smutnicki, C.: The flow shop with parallel machines: A tabu search approach. European Journal of Operational Research 106, 226–253 (1998) 8. Kurz, M.E., Askin, R.G.: Scheduling flexible flow lines with sequence-dependent setup times. European Journal of Operational Research 159, 66–82 (2004) 9. Gen, M., Cheng, R.: Network Models and Optimization: Multiobjective Genetic Algorithm Approach. Springer-Verlag London Limited, Heidelberg (2008) 10. Lee, Y.H., Jeong, C.S., Moon, C.: Advanced planning and scheduling with outsourcing in manufacturing supply chain. Computers & Indusrtial Engineering 43, 351–374 (2002)

Core Image Coding Based on WP-EBCOT Yongdan Nie, Yan Zhang, and Jinghui Li School of Computer and Information Technology, Northeast Petroleum University, Daqing, 163318, China [email protected]

Abstract. This paper proposed a new core image coding algorithm based on Wavelet Packet Embedded Block Coding with Optimized Truncation (WPEBCOT), for the characteristics of rich textures and complex edges in the core images. The entropy-based algorithm for best basis selection is used to decompose the core image, and then the wavelet packet subband structure is tested by EBCOT with various different code blocks size, as a result we find that the optimal size of the code blocks is 64*64. Results show that the proposed algorithm outperforms the base-line JPEG2000 for PSNR, and provides better visual quality for core images. Keywords: EBCOT, Core Image coding, Wavelet Packet.

1 Introduction The core samples are one of the most foundational geological data in the research of exploration and development for oil and gas fields. It is very conducive to preserve and analysis the core data that a large number of core samples have been stored by scanning approach to the digital images. However because the amount of core images is too huge, so they must be compressed before stored. Now the core images are mostly compressed by the DCT-based JPEG algorithm, which will cause the fuzzy blocking effects in the area of texture or edge region in low bit rate situation. The wavelet transform cant not only effectively overcome the limitations of Fourier transform in dealing with non-stationary signal and the shortcoming of block-effect of the DCT coding, but also provide a multi-resolution image representation, by locating the information at any resolution level to realize the embedded coding schemes of giving priority to coding and transmission of the important information in images. In 2000, Taubaman proposed EBCOT compression algorithm [1] which is finally adopted by the JPEG2000 image coding standard [2] as the core scheme. JPEG2000 operate on independent, non-overlapping blocks of quantized wavelet coefficients which are coded in several bit layers to create an embedded, scalable bit streams (Tier-1 coding). Instead of zerotrees, the JPEG2000 scheme depends on a per-block quad-tree structure since the strictly independent block coding strategy precludes structures across subbands or even code-blocks. These independent code-blocks are passed down the coding pipeline and generate separate bit streams. Transmitting each bit layer corresponds to a certain distortion level. The partitioning of the available bit budget between the code-blocks and layers (truncation points) is determined using a R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 393–398, 2011. © Springer-Verlag Berlin Heidelberg 2011

394

Y. Nie, Y. Zhang, and J. Li

sophisticated optimization strategy for optimal rate/distortion performance (Tier-2 coding)[3].In EBCOT the size of code block is an important factor affecting the coding performance, a typical choice for the nominal code block size is 64*64. For example, assume it is 2

xcb '

× 2 ycb , the xcb ' , ycb' are defined as follows: '

xcb' = min( ppx ' , xcb) ; ycb ' = min( ppy ' , ycb) ; ppx ' = ppx (r = 0) or ppx ' = ppx − 1 (r > 0) ; ppy ' = ppy (r = 0) or ppy ' = ppy − 1 (r > 0) . xcb , ycb is the parameters specified by user, r is the resolution level. The regulation of 2 < xcb < 12 , 2 < ycb < 12 , xcb + ycb < 12 is stetted in JPEG2000, and the xcb and ycb is no need to be equal. So the size of code block is related to the size of precincts ppx , ppy according to the corresponding resolution level and the parameters xcb , ycb specified by user. The JPEG2000 compression standard make a larger progress in both the architecture and the compression performance compared other efficient coding algorithm such as EZW [4], SPIHT [5], and SPECK [6]. However the dyadic DWT does not adapt to the various space-frequency properties of images, the energy compaction it achieves is generally not optimal[7].The WP generalizes the pyramidal decomposition structure by iterating the decompositions in the high-frequency subbands as well, in which the coefficients show the stronger energy accumulation than in the wavelet decomposition. It is proved that the rate-distortion performance of adaptively generated wavelet packet subband structures is superior to classical pyramidal ones, especially for images with highly textured content. Yang Yongming et al. [8] applied the block quantization technology inside subbands into the rate-distortion (RD) optimized wavelet packet decomposition framework. They used block segmentation scheme and the context modeling techniques of JPEG2000 to encode a separated block in each subband. The algorithm achieves excellent performance not only in the visual quality, but also in PSNR. So here we lay particular emphasis on the shortcoming of DWT coding schemes, thus propose the core image coding algorithm based on WP-EBCOT which is more suited for the characteristic of textures and edges in core images. This paper is organized as follows: In Section 2 wavelet packet decomposition is described. In Section 3 we explain core image compression base on WP-EBCOT algorithm. Coding results are presented in Section 4, followed by conclusion in Section 5.

2 Wavelet Packet Decomposition The WP is a wavelet transform where the signal is passed through more filters than the DWT. It generalizes the dyadic decomposition by iterating the decomposition on the high pass subbands as well. In the WP, both the detail and approximation coefficients are decomposed. For n levels of complete WP decomposition, the two

Core Image Coding Based on WP-EBCOT

(a)

395

(b)

Fig. 1. (a) Origin image. (b) 3 levels completely WP Decomposed image. 2n

dimensional image is produced 2 different subbands, for example the 3 levels completely decomposed WP of peppers is illustrated in the Figure 1. In order to achieve compression gain while keeping the computational load reasonably low, the WP decomposition needs two entities: a defined cost function for basis comparison and a fast and all-sided search scheme. The choice of cost function for best basis of WP is substantial to the coding performance and computing efficiency. Ramchandran and Vetterli proposed a practical algorithm based on the cost function of coding rate and distortion, which considered the number of bits needed to approximate an image with a given distortion in [9]. Because the selection of best basis involves embedded nonlinear optimization problems, the overall complexity of the approach is extremely high. Coifman and Wickerhauser proposed entropy-based algorithm for best basis selection [10]. According to this cost function, the choice of optimal basis is made by pruning the complete decomposition tree only once, but the cost function and the corresponding wavelet packet decomposition do not take into account the following encoding method, which to a large extent determines the performance of a coding framework. Although the entropy-based method leads to a sub-optimum basis, taking into account the computational complexity of the practical application, in this work we use Shannon entropy as the cost function for WP basis comparison instead of coding rate and distortion.

3 WP-EBCOT Unlike the dyadic subbands structure, the high-pass subbands have been transformed continuously in WP, so that the energy of high-pass subbands is accumulated to the upper left corner and the possibility of significant coefficients occurrence in the upper left corner of this subband is extremely high. Moreover the size of code block is an important factor to affect coding performance. In order to analyze the impact of code block size in wavelet packet decomposition subband on coding efficiency. A lot of experiments are taken to get the comparison of coding efficiency of EBCOT with the different size of code block partitioning. Here only we only presented the two test core images samples see in Figure 2 and the result are presented in Figure 3.

396

Y. Nie, Y. Zhang, and J. Li

(a)

(b)

Fig. 2. (a) Core image sample 1. (b) Core image sample 2.

(a)

(b)

Fig. 3. (a) Core image sample 1: comparison of four different size of block for EBCOT and WP-EBCOT coding PSNR. (b) Core image sample 2: comparison of four different size of block for EBCOT and WP-EBCOT coding PSNR.

(a) (b) Fig. 4. Best basis geometry: (a) Sample 1. (b) Sample 2.

Core Image Coding Based on WP-EBCOT

397

Our results indicate that as the size of code blocks are increased, the encoding performance of EBCOT are reduced for the same core image sample, and the PSNR declines significantly when the size of code block varying from 32*32 To 16*16 , but declines slightly when from 64*64 to 32*32. In order to take advantage of the characteristics of wavelet packet coefficients distribution while also taking into account the influence of code block size on PSNR , this work uses the size of code block is 64 * 64. The WP decomposition results of two core images as shown in figure 4. Practical tests are presented on a 5 level 9/7 biorthogonal wavelet transform and using the WP decomposition in the high pass. The rest of WP-EBCOT is same as the EBCOT. The whole algorithm of WPEBCOT is implemented in open source software jasper -1.900.1[11] of static image compression standard JPEG2000 and the compile platform is based on VC++.NET 2005.

4 Experimental Results The PSNR comparison of The WP-EBCOT and EBCOT compression are shown as Figure 5. Experimental results show that compared to the EBCOT, the propose algorithm achieve higher PSNR and obtain the better visual effects, especially at low bit rates for the core image.

(a)

(b)

Fig. 5. (a) Core image sample 1: comparison of PSNR values for EBCOT and WP-EBCOT. (b) Core image sample 2: comparison of PSNR values for EBCOT and WP-EBCOT.

5 Conclusions This paper proposed the WP-EBCOT to improve the compression performance of the core images. The results show that the proposed algorithm provides the higher PSNR, and better visual quality for core images than the EBCOT. One way to improve the performance is to design an improved cost function for WP best basis selection which would take into account the factor of coding block in EBCOT algorithm and provide optimal distortion value for a given bit rate.

398

Y. Nie, Y. Zhang, and J. Li

References [1] Taubman, D.: High performance scalable image compression with EBCOT. IEEE Trans. on Image Processing 9(7), 1158–1170 (2000) [2] Taubman, D., Marcellin, M.W.: JPEG2000-Image Compression Fundamentals, Standards and Practice. Kluwer Academic Publishers, Dordrecht (2002) [3] Reisecker, M., Uhl, A.: Wavelet-Packet Subband Structures In The Evolution of The Jpeg2000 Standard. In: Proceedings of the 6th Nordic Signal Processing Symposium NORSIG 2004, pp. 97–100 (2004) [4] Shapiro, J.M.: Embedded image coding using zerotrees of wavelet coefficients. IEEE Trans. Signal Process. 41(10), 3445–3462 (1993) [5] Said, A., Pearlman, W.A.: A new, fast and efficient image codec based on set partitioning in hierarchical trees. IEEE Trans. Circ. Syst. Video Technol. 6(6), 243–250 (1996) [6] Pearlman, W.A., Islam, A., Nagaraj, N., Said, A.: Efficient, low-complexity image coding with a set-partitioning embed-ded block coder. IEEE Trans. Circ. Syst. Video Technol. 14(11), 1219–1235 (2004) [7] Sprljana, N., Grgicb, S., Grgicb, M.: Modified SPIHT algorithm for wavelet packet image coding. Real-Time Imaging 11, 378–388 (2005) [8] Yang, Y., Xu, C.: A wavelet packet based block-partitioning image coding algorithm with rate-distortion optimization. Science in China Series F: Information Sciences 51(8), 1039–1054 (2008) [9] Ramchandran, K., Vetterli, M.: Best wavelet packet bases in a rate distortion sense. IEEE Transactions on Image Processing 2(2), 160–175 (1993) [10] Coifman, R.R., Wickerhauser, M.V.: Entropy-based algorithms for best basis selection. IEEE Trans. Inform Theory, Special Issue on Wavelet Transforms and Multires. Signal Anal. 38(3), 713–718 (1992) [11] Adams, M.D.: JasPer Software Reference Manual (Version 1.900.0) (December 2006), http://www.ece.uvic.ca/~mdadams/jasper/jasper.pdf

A New Method of Facial Expression Recognition Based on SPE Plus SVM Zilu Ying, Mingwei Huang, Zhen Wang, and Zhewei Wang School of Information Engineering, WUYI University, Jiangmen, China [email protected], [email protected], [email protected]

Abstract. A novel method of facial expression recognition (FER) is presented, which uses stochastic proximity embedding (SPE) for data dimension reduction, and support vector machine (SVM) for expression classification. The proposed algorithm is applied to Japanese Female Facial Expression (JAFFE) database for FER, better performance is obtained compared with some traditional algorithms, such as PCA and LDA etc.. The result have further proved the effectiveness of the proposed algorithm. Keywords: Facial expression recognition (FER), stochastic proximity embedding (SPE), SVM.

1 Introduction Facial expressions imply so much information about human emotions that it plays an important role in human communications. In order to facilitate a more intelligent and smart human machine interface of multimedia products, automatic facial expression recognition (FER) has become a hot issue in the computer vision and pattern recognition community. One of the many difficulties on FER is the high dimension of data. Extremely large data will increase the cost of computation and bring about the socalled “dimensionality curse" problem. Reducing data into fewer dimensions often makes analysis algorithms more efficient, and can help machine learning algorithms make more accurate predictions. To reduce the dimensionality of the hyper-dimensional data, many algorithms have been introduced. Among them, stochastic proximity embedding (SPE), which was proposed by DK Agrafiotis [1] in 2002, is an excellent self-organizing algorithm for embedding a set of related observations into a low-dimensional space that preserves the intrinsic dimensionality and metric structure of the data. This method is programmatically simple, robust, and convergent [2]. Typically, SPE can be applied to extract constraint surfaces of any desired dimension. Because it works directly with proximity data, it can be used for both dimension reduction and feature extraction [3]. In this paper, we use SPE as the dimensionality reduction method, and adopt SVM as classifier testing on JAFFE database on the study of FER. Comparing with conventional algorithms, such as PCA and LDA, our algorithm has a better performance, and its best performance was recorded 69% on FER. R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 399–404, 2011. © Springer-Verlag Berlin Heidelberg 2011

400

Z. Ying et al.

2 The Principle of SPE SPE allocates large volumes of high-dimensional data in a low-dimensional space according to their proximities or similarities. Given a set of k objects, a symmetric matrix, rij of relationships between these objects, and a set of images on a

{

}

m-dimensional display plane xi , i = 1, 2, K , k ; x ∈ ℜ m , the problem is to place xi onto the plane in such a way that their Euclidean distance d ij = xi − x j approximate as closely as possible the corresponding values rij . The quality of the projection is determined using a error function [4],

E=∑ i< j

f (dij , rij ) rij

∑r i< j

ij

(1)

where

⎧(dij − rij ) 2 if (rij ≤ rc ) ∪ ( dij < rij ) ⎪ f (dij , rij ) = ⎨ if (rij > rc ) ∪ ( dij ≥ rij ) ⎪⎩0 and

(2)

rc is the neighborhood radius. The function is numerically minimized in order to

find the optimal configuration. Stochastic proximity embedding uses a self-organizing scheme attempts to bring 2 each individual term ( d ij − rij ) rapidly to zero. The method starts with an initial configuration and iteratively refines it by repeatedly selecting tow point at random, and adjusting their coordinates so that their Euclidean distance on the map d ij matches more closely their corresponding proximity rij . The correction is proportional to the

rij − dij

, where λ is a learning rate parameter that decrease during the d ij course of refinement to avoid oscillatory behavior. The detail algorithm is as follow: disparity λ

(1). Initialize the coordinates xi . Select an initial learning rate λ . (2). Select a pair of points, i and j , at random and calculate their distance d ij = xi − x j . If d ij ≠ rij , update the coordinates xi and x j by:

xi ← xi + λ

1 rij − dij ( xi − x j ) 2 dij + ε

(3)

A New Method of Facial Expression Recognition Based on SPE Plus SVM

401

and

xj ← xj + λ

1 rij − dij ( x j − xi ) 2 dij + ε

(4)

(3). Repeat step (2) for a prescribed number of steps S . (4). Decrease the learning rate λ by prescribed decrement δλ . (5). Repeat step (2)-(4) for a prescribed number of cycles.

3 Experiment Analysis 3.1 Flows of the Experiment

The experiments of the proposed algorithm on FER are implemented in the JAFFE database. The database contains 213 images in which ten persons are expressing three or four times the seven basic expressions (anger, disgust, fear, happiness, neutral, sadness and surprise). We select 3 images of each expression of each person (210 images in all) for our experiments. To eliminating the unwanted redundant information effecting on FER, the expression images are registered using eye coordinates and cropped with a mask to exclude non-face area, as showed in Fig. 1.

Fig. 1. The samples of the JAFFE database that have been excluded the non-face area

Then, images are resized to 64×64 pixels. After this step, we process the data with histogram equalization. The step posterior to image preprocessing is dataset construction for SPE processing, we construct the dataset by reshaping each of the 64 64 data-points to 4096 1 data-points. Then all data-points constitute a dataset which contains 210 data-points with 4096 dimensions. Then we can reduce the data dimensionality of the dataset using SPE algorithm. After data dimensionality reduction, the next step is training and classifying data using SVM. We divide the database into ten equally sized sets (each set is corresponding to each specific person, the JAFFE database altogether contains ten persons’ facial expression images): nine sets are used for training and the remaining one for testing. This process is repeated so that each of the ten sets is used once as testing set and nine times ad training set. The average result over all ten rounds is considered as the final expression recognition rates of the experiment.

×

×

3.2 Result of the Experiment and Analysis

{

In this work, we uses OAA-SVM [5] model with RBF ( ψ ( x, xk ) = exp −

x − xk

2 2

σ

2

} ) kernel function as classifier in FER. At first, we optimize the parameters

402

Z. Ying et al.

on OAA-SVM model, this step insure the SVM classifier work on its best condition on FER. This step including optimize the regulation constant C and kernel argument σ 2 . The regularization constant C is a cost parameter that trades of margin size and training error in SVM. At first, the performance of FER was improved along with the increasing values of C , and recorded its peak value (69.05%) at where C =250, but afterwards, the FER recognition rates trend to fluctuate widely when C continues to increase. The recognition rate of FER rises sharply along with the increasing of σ 2 value, then it trend to be stable when the value of σ 2 is over 2. The parameter σ 2 gets its optimum parameter at 6.4, at where the recognition rates are 67.14%. An experiment was conducted to analysis the relationship between reduced dimensions and FER performance. The experiment result, which is showed in the Fig. 2, indicates that, the dimensions can be reduced to 40 and the recognition rate also keeps above 60%, and the performance of FER is stable around the dimensions between 40 and 200, and the recognition rates are almost all above 60%. The highest recognition rate is 67.14% which is attained in the dimension of 48. The best performance of FER using SPE+SVM algorithm are in a relative low dimensionality, which appears in the range of dimensions between 40 and 60. To compare with other algorithm, we used PCA to substitute SPE as dimensionality reduction toolkit and redo this experiment. The result was also showed in Fig. 2, which reveals the different abilities of SPE and PCA on feature extraction.

The expression recognition rates(%)

70 60 50 SPE+SVM PCA+SVM

40 30 20 10

0

20

40

60

80

100

120

140

160

180

200

Dimensions of the dataset (ı2=6.4 C=250)

Fig. 2. The reduced dimensions of the dataset affect the performance of FER. (the results are acquired under OAA-SVM parameters of C =250, and σ 2 = 6.4).

As we can see from this plot, apart from the FER performance is superior to PCA+SVM on the whole, SPE also exhibits its’ ability as an excellent feature extractor. For the best performance of SPE+SVM algorithm on FER is attained at the range of dimensions from 40 to 60, a relatively low dimensions compare with the most conventional dimensionality reduction algorithm, such as PCA, this suggests the SPE can produces maps that exhibits meaningful structure even when the data is embedded in fewer than its intrinsic dimensions.

A New Method of Facial Expression Recognition Based on SPE Plus SVM

403

We also made a comparison between some conventional algorithm and our proposed algorithm on FER as shown in table 1. The front four methods adopt OAA-SVM as the classifier and the last one adopt MVBoost as classifier. Table 1 shows the recognition rates of those methods. In this contrast experiment, we can see that SPE, as an effective approach to facial expression feature extraction, is much more superior to PCA, KPCA, LDA and also better than Gabor Histogram Feature+MVBoost algorithm. Table 1. The performances comparing between different FER algorithms (in percentage) Methods

SPE

PCA

KPCA [6]

LDA [6]

FER rates

69.0

60.0

53.6

55.7

Gabor Histogram Feature+MVBoost [7] 58.7

Another meaningful experiment was conducted to compare the FER performance between using SPE+SVM and PCA+SVM algorithm by contrasting their various expressions recognition rates respectively. The results are showed in Fig. 3.

The expression recognition rate(%)

90

SPE+SVM

80

PCA+SVM 70 60 50 40 30 20 10 0

anger

disgust

fear

happiness

neutral

sadness

surprise

Fig. 3. The FER performance comparison between SPE+SVM and PCA+SVM algorithms on various expressions. (the results are acquired under OAA-SVM with parameters of C =250, and σ 2 = 6.4).

As shown in Fig. 3, theses two algorithms both recorded high recognition rates on the expressions of ‘anger’, ‘disgust’, ‘happiness’ and ‘surprise’, and their performance have slightly differences. However, when it come to the rest three expressions(‘fear’, ‘neutral’, and ‘sadness’), which were recorded with low recognition rates on both algorithms, the performances of FER are quite different: the behavior of SPE+SVM algorithm is far superior to the PCA+SVM. To analysis this phenomenon, we can see the divisibility of these three expressions is low. PCA, as a conventional dimensionality reduction tool, when it reduces the dimensions of the dataset, it also reduces some useful distinction nature between these expressions, which result in poor FER performances. But SPE is not only just a dimensionality reduction tool, it also acts as a feature extractor, it can preserve the minute difference of these expressions, this

404

Z. Ying et al.

specialty enables it work well on fuzzy environments. This instance maybe illustrate the SPE algorithm is more robust than PCA etc. conventional dimensionality reduction algorithms.

4 Conclusions A new approach for facial expression recognition based on stochastic proximity embedding plus SVM was proposed. Tested on JAFFE database, the proposed algorithm obtained a satisfactory result. The FER performance of the proposed algorithm better than the traditional algorithms such as PCA, KPCA and LDA etc., and also superior to some newly introduced algorithm, such as FEA based on Gabor Histogram Feature and MVBoost. Because SPE has the ability to extract features from the dataset, compare with conventional FER algorithm PCA+SVM, SPE+SVM can attain the best performance at a very low dimension of the dataset, this advantage also enable the SPE+SVM algorithm to be more robust than PCA+SVM etc. conventional dimensionality reduction algorithms on FER.

Acknowledgment This paper was supported by NNSF (No.61072127, No. 61070167), Guangdong NSF (No. 10152902001000002, No. 07010869), and High Level Personnel Project of Guangdong Colleges (No. [2010]79).

References [1] Agrafiotis, D.K., Xu, H.: A self–organizang principle for learning nonlinear manifold. proc. Network Academy of Science, 15869–15872 (2002) [2] Rassokin, D.N., Agrafiotis, D.K.: A modified update rule for stochastic proximity embedding. J. Journal of Molecular Graphic an Modelling 22, 133–140 (2003) [3] Agrafiotis, D.K.: Stochastic Proximity Embedding. J. Comput. Chem. 24, 1251–1271 (2003) [4] Nishikawa, N., Doi, S.: Optimization of Distances for a Stochastic Embedding and Clustering of Hing-Dimensuonal Data. In: The 23rd International Technical Conference on Circuit/systems, Computers and Communications, pp. 1125–1128 (2008) [5] Abe, S.: Analysis of Multiclass Support Vector Machines. In: International Conference on Computational Intelligence for Modelling Control and Automation, pp. 385–396 (2003) [6] Ying, Z., Zhang, G.: Facial Expression Recognition Based on NMF and SVM. In: International Forum on Information Technology and Applications 2009, vol. 3, pp. 612–615 (2009) [7] Liu, X., Zhang, Y.: Facial Expression Recognition Based on Gabor Histogram Feature and MVBoost. Journal of Computer Research and Development 44(7), 1089–1096 (2002)

Multiple Unmanned Air Vehicles Control Using Neurobiologically Inspired Algorithms Yong Zhang1 and Li Wang2 1 2

Institute of Information, Tianjin University of Commerce, Tianjin 300130, China Institute of Information, Hebei University of Technology, Tianjin 300130, China [email protected], [email protected]

Abstract. In order to develop and evaluate future Unmanned Air Vehicles for the hazardous environmental monitoring, the comprehensive simulation test and analysis of new advanced concepts is imperative. This paper details an on-going proof of concept focused on development of a neurobiologically-inspired system for the high level control of a Air Vehicle team. This study, entitled Neurobiologically Enabled Autonomous Vehicle Operations, will evaluate initial System-Under-Test concept data by selecting well defined tasks, and evaluating performance based on assignment effectiveness, cooperation, and adaptability of the system. The system will be tested thoroughly in simulation, and if mature, will be implemented in hardware. Keywords: Neurobiologically-inspired Algorithms, Unmanned Air Vehicle, Reconnaissance, Surveillance, Target Acquisition.

1 Introduction The use of unmanned aerial vehicles (UAVs) for various military or civilian missions has received growing attention in the last decade. Depending on the application, there are many different ideas on how to measure the autonomy of a system. In [1] the Autonomous Control Levels (ACL) metric was introduced. The majority of the techniques that have been developed can be classified as either optimization-based methods that make use of extensive a priori information or reactive methods that use local information to define a global planning strategy. Typically these methods require offline processing due to their computational complexity and can thus show a lack of robustness in dealing with highly dynamic environments. However, in some cases, replanning phases during execution of the mission can be possible. The graph search approaches that have been used extensively typify these methods. Reactive methods take planning methods one step further by incorporating local information into the control strategy to allow for changing conditions in the environment. Rather than generating an a priori path through a given environment, reactive methods focus on using local information to define a controller for a vehicle that ultimately gives rise to the desired behavior. Potential function methods have long been exploited in the reactive paradigm. However, there is an area of research that focuses on neurobiological (or brain-based) design that has been less exploited in cooperative control. R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 405–410, 2011. © Springer-Verlag Berlin Heidelberg 2011

406

Y. Zhang and L. Wang

Brain-Based Design (BBD) may provide a unique solution to the cooperative control problem. The essence of BBD lies in its ability to adapt to changing conditions in real time. By mimicking the design and operation of the human brain, these systems can exhibit a great capacity for learning about their environment [2]. It is the ability to make reasoned decisions that separates BBD from other cooperative control strategies. The Neurobiologically Enabled Autonomous Vehicle Operations (NEAVO) Study aims to exploit the capabilities of BBD to solve problems in the area of UAV cooperative control.

2 Mission Tasks In NEAVO, Multiple UAVs will be used to perform tasks in the category of Reconnaissance, Surveillance, and Target Acquisition (RSTA). These tasks are especially important in urban settings. Given the increased need for real-time information in urban settings, perfecting RSTA UAV teams is an important research topic with applications in modern warfare and hazardous environmental monitoring. This process is shown in Figure 1. For NEAVO, a subset of important RSTA tasks is identified as test cases for the demonstration of BBD. Two basic tasks have been identified:(1)Tracking a Moving Target;(2)Cooperative Area Search. Which is shown in In the Figure 2.

Fig. 1. The Kill Chain. Each step may iterate one or more times and be performed by one or more assets.

(1)

(2)

Fig. 2. Two basic tasks (1) Tracking a Moving Target; (2) Cooperative Area Search

Multiple Unmanned Air Vehicles Control Using Neurobiologically Inspired Algorithms

407

2.1 Tracking a Moving Target Tracking a moving target requires knowledge of the target’s current position as well as the ability to predict future locations. If a UAV team is given prior knowledge of a road network, or a map of obstructions that would block the type of vehicle being tracked, The UAV team could more simply predict the likely motion of the target. However, if the team is required to “discover” the terrain while tracking, the problem becomes more complicated. This is where BBD can be exploited. 2.2 Cooperative Area Search It is common for areas to be searched using established search methods. When entering a new area, it is necessary to identify important items of interest, such as the locations of friendly or hostile troops, activity of civilian populations, the layout of road networks, or to find a specific target. Using BBD, a system can adapt to unknown terrain, learning ways to better approach areas that are to be searched, and to minimize the time to complete a mission. By employing multiple assets in a cooperative manner, the system can maximize the area searched over time. 2.3 Constraints For the above tasks, several constraints are provided. A list of constraints follows.

Fig. 3. Area search with no-fly zone

(1) The team shall stay within a specified area of operations. This area may change in shape or size during operations. (2) The team must avoid all areas marked as no-fly zones is shown in Figure 3. Nofly zones may be added, removed, or changed during operations. (3) The system shall allow for team members to be added or removed at any time. New members shall be commanded in the same way as existing team members. (4) Steady-state winds shall be considered when planning team actions.

3 System Models All aircraft in NEAVO will be small UAVs. The team may be heterogeneous, but all aircraft will be statically stable, autopilot controlled vehicles. Each vehicle will be a waypoint follower. A route will be uploaded to the aircraft by the command system.

408

Y. Zhang and L. Wang

The focus of NEAVO is cooperative task planning. BBD has shown promise in the area of video processing, and target recognition, but these technologies are not being considered in the NEAVO effort. Sensors will be modeled as polygons of fixed size, with a grid of pixels that represent a video image over that area. A notional sketch of this type of sensor is shown in Figure 4.

G

B

B

G

B

B

B B

G

G

G

G

G

G

G

G

B

B

B

B

(1) Top View

(2) Side View

Fig. 4. Sensor Description (Notional)

Data will be passed from the sensor simulation to the mission planner supplying the information seen and a confidence value. Initially, the sensor will be in a fixed orientation on board the aircraft, but sensor movement may be considered later in the NEAVO effort.When data is passed from the sensor to the planner, a confidence value is applied to the detection. The confidence can be expressed as a function C s of the number of pixels on target, the amount of time that the target spends in the sensor field-of-view, the aspect angle to the target and the level of contrast between the target and the surrounding pixels.

(

C s = f N p , t FOV ,ψ , δ

)

(1)

4 Mission Planning System A multi-UAV mission planning system can be broken into five parts, as shown in Figure 4. The user, whether human or automated, requests a task or set of tasks. A task allocator decides, based on current information regarding the UAV team, which UAV or sub-group of UAVs is assigned that task. The importance or rank of the task can be decided by the task requester. The final step is to plan a trajectory for all UAVs that have been allocated new tasks. The planning system may run once per request, or iteratively based on user analysis. One or all parts of the mission planning system can be distributed among the UAV team, or contained in one physical location. In NEAVO, the system is assumed to be centralized in one location, receiving sufficiently timely and accurate information from the UAVs to operate the entire team.

Multiple Unmanned Air Vehicles Control Using Neurobiologically Inspired Algorithms

409

NEAVO aims to apply BBD to the task allocation and path planning steps of the mission planning system. The result of the path planning step is a route instruction to one or more UAVs. Cooperation, Anticipation, Adaptability

Knowledge

User Analysis

Task Requests

Task Allocation

Path Planning

Execution

Fig. 5. The elements of mission planning

5 Neurocomputation Neurocomputation seeks to mimic the processing abilities of the brain to solve complex problems [3]. The most popular of these algorithms is the artificial neural network. ANNs consist of very simple processing units connected in a distributed fashion. Each processing unit is modeled after the neuron and is typically characterized by an activation function that may or may not produce an output based on the input presented to it. Learning is accomplished by adjusting the weights that connect the neurons to each other. In order to produce useful results, the neural network must be given some criteria for determining the goodness of its solution. This criterion varies between different learning algorithms. [4][5] Supervised learning algorithms make use of training data in the form of a set of inputs and outputs. Using information about the expected outputs given an input, the network learns how to respond to new inputs that it hasn’t seen before. The performance of the neural network when acting on its own is entirely dependent on the behaviors present in the training data. The fitness of the solutions produced is determined based on comparison to the output data in the training set. Essentially, these represent desired outputs and the network is trained to minimize the difference between the actual output and the desired output. Supervised learning mechanisms have been successfully applied to problems involving handwriting recognition, pattern recognition, and information retrieval. Unsupervised learning mechanisms make use of training data as well. However, there is no output data used in the training set. An input data set is used to fit a model to observations. By forming a probability distribution over a set of inputs, the neural network can be trained to output the conditional probability of an input given all previous inputs. As an example, consider a set of temperature data from a properly functioning power plant. A neural network could be trained to determine the likelihood of a particular reading given all previous readings, and could thus be used to monitor the operation of the power plant. Unsupervised learning techniques have also been applied to data compression problems. Reinforcement learning is unique in that it does not make use of training data. Instead, a set of possible actions is provided to the network for a given situation and a

410

Y. Zhang and L. Wang

system of rewards and penalties is used to direct behavior. At each step an estimate of the future expected reward given a particular action is formed and the neural network is trained to maximize its reward. In this way, reinforcement learning mechanisms rely on direct interaction with the environment. Reinforcement learning focuses on online performance rather than a priori training and seeks to strike a balance between the exploration of unknown areas and the exploitation of currently held knowledge. This learning scheme has been successfully applied to robot control and telecommunications problems.

6 Summary The NEAVO Program seeks to evaluate the concept of a neurobiologically inspired system controlling a RSTA UAV team. The simulation chosen to assess the system concept is called MultiUAVs. Construction of MultiUAVs satisfies the need for a simulation environment that researchers can use to develop, implement and analyze cooperative control algorithms. Since the purpose of MultiUAVs is to make cooperative control research accessible to researchers it was constructed primarily using MATLAB and Simulink. Some of the simulation functions are programmed in C++. During the simulation, vehicles fly predefined search trajectories until a target is encountered. Each vehicle has a sensor footprint that defines its field of view. Target positions are either set randomly or they can be specified by the user. When a target position is inside of a vehicle’s sensor footprint, that vehicle runs a sensor simulation and sends the results to the other vehicles. With actions assigned by the selected cooperative control algorithm, the vehicles generate their own trajectories to accomplish tasks. The simulation takes place in a three dimensional environment, but all of the trajectory planning is for a constant altitude, i.e. two dimensions. Once each vehicle has finished its assigned tasks it returns to its predefined search pattern trajectory. The simulation continues until it is stopped or the preset simulation run time has elapsed. Several issues are currently being explored in the NEAVO Program using MultiUAVs.

References 1.

2.

3. 4. 5.

Clough, B.T.: Metrics, schmetrics How the heck do you determine a UAVs autonomy anyway. In: Proceedings of the Performance Metrics for Intelligent Systems Workshop, Gaithersburg, MD (2002) Reggia, J., Tagamets, M., Contreras-Vidal, J., Weems, S., Jacobs, D., Winder, R., Chabuk, T.: Development of a Large-Scale Integrated Neurocognitive Architecture. DARPA Information Processing Technology Office (2006) Rumelhart, D., McClelland, J.: Parallel Distributed Processing: Explorations in the Microstructure of Cognition. MIT Press, Cambridge (1988) Haykin, S.: Neural Networks: A Comprehensive Foundation. Prentice-Hall, Englewood Cliffs (1999) Nunes de Castro, L.: Fundamentals of Natural Computing: Basic Concepts, Algorithms, and Applications. Chapman and Hall, Boca Raton (2006)

The Use of BS7799 Information Security Standard to Construct Mechanisms for the Management of Medical Organization Information Security Shu-Fan Liu, Hao-En Chueh, and Kuo-Hsiung Liao Department of Information Management, Yuanpei University, No. 306, Yuanpei Street, Hsinchu 30015, Taiwan [email protected], [email protected], [email protected]

Abstract. According to surveys, 80 % of security related events threatening information in medical organizations is due to improper management. Most research on information security has focused on information and security technology, such as network security and access control; rarely addressing issues at the management issues. The main purpose of this study is to construct a BS7799 based mechanism for the management of information with regard to security as it applies to medical organizations. This study analyzes and identifies the most common events related to information security in medical organizations and categorizes these events as high-risk, transferable-risk, and controlled-risk to facilitate the management of such risk. Keywords: BS7799, Medical organizations, Information security, Risk management, Access control.

1 Introduction The objective of information security is "to ensure the various interests which rely on the information system, avoid harm created by the loss of confidentiality, integrity, and availability" [5, 9, 14]. Past research on information security has constantly stressed the need for information and communication technology (ICT) such as data encryption, firewall technology, and computer viruses [2, 4, 6, 7, 8, 10, 11, 12, 16, 17, 18, 19, 20, 21, 22]. The primary responsibility of medical organizations is to provide patients with quality care, information security events quickly leave a medical organization incapable of carrying out normal operations, with negative consequences for the rights of patients. Thus, the importance of maintaining the security of medical organizations lies not only in protecting important information from theft, forgery, or damage; more importantly, information security safeguards the reputation, value, and sustainable development of the medical organization. To help enterprises achieve information security goals, the British Standards Institution (BSI) has published the BS7799 information security management standard in 1995 (which became ISO 17799 international standard in 2000) [1, 14], covering all aspects of information security. Current literature on risk management can only R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 411–416, 2011. © Springer-Verlag Berlin Heidelberg 2011

412

S.-F. Liu, H.-E. Chueh, and K.-H. Liao

provide people with applications of risk management to information security. These studies are unable to provide a clearly quantitative, indexed description or applicable methods for managing risk related to assets, personnel, or other threats and vulnerabilities created by various information security risks [18].The main purpose of this study was to implement the ten control items and one hundred twenty-seven control objectives of the BS7799 information security management standard as a foundation for the protection of information and establish a mechanism for the management of information security applicable to medical organizations.

2 Literature Review The frequency of information security events in medical organizations has been attracting increased attention to medical organization security issues. According to Smith and Eloff [15], the scope of security protection for medical organizations should include assistance in avoiding: (1) physical damage to persons or property; (2) privacy violations; (3) loss or destruction of medical information; (4) harm to operational integrity or consistency. The programming information security for medical organizations must be considered from the perspective of a comprehensive security protection , including: defining the scope of system security protection, assessing risks, establishing internal control and auditing systems, exploring other managementrelated issues, and structuring the comprehensive needs for information security of medical organizations. Harmful information security events usually include behavioral uncertainty, unpredictability, and the inability to understand what true security is. In one way, risk management improves the performance and the effectiveness of evaluation with regard to information security. After enhancing awareness of information security, organizations can use resources more efficiently, create better project management, and minimize waste. Willett (1901) [23] believed that "risk" and "chance" were different and that they should be distinguished by using "the degree of probability" for chance and "the degree of uncertainty" for risk. The general steps in managing risk are: first, identifying the risk; second, measuring the risk; third, selecting the proper tools; fourth, taking action; fifth, evaluating the performance. Because a single information security event could trigger a chain of issues, dealing with information security events arbitrarily may cause unforeseen error and losses. The United States has information security operating standards such as Control Objectives for Information and Related Technology (COBIT), Trusted Computer System Evaluation Criteria (TCSEC), while the United Kingdom has introduced the BS7799 information security management standard. Control Objectives for Information and Related Technology, the operating standard published by Information Systems Audit and Control Association (ISACA) in 1995, is a set of comprehensive considerations based on information technology control standards and information technology as security. Trusted Computer System Evaluation Criteria, proposed by the U.S. National Computer Security Commission (NCSC) in 1983, takes a systematic approach by dividing security issues into four categories, named A, B, C, and D, where the category of computer systems dictates the standard level of security required. The aim of the BS7799 information security management standard, developed

The Use of BS7799 Information Security Standard to Construct Mechanisms

413

by BSI, was to ensure the security of business information assets including software and hardware facilities as well as data and information, by avoiding information security-related damage caused by internal and external threats as well as operating mistakes of organizational staff. Simply put, the objective of BS7799 was to establish a comprehensive information security management system by ensuring information confidentiality, integrity, and availability [1, 3, 5, 9, 14]. Each standard has its own suitable application, but from the perspective of information security management for medical organizations, BS7799 is best suited to ensuring the security of medical organizations and their information-related assets. The BS7799 information security management standard includes the protection of software, hardware facilities, information, and avoiding various internal and external threats as well as operating mistakes of organizational staff. It also covers various aspects of security policy, from formulating policy related to security and delegating security related responsibilities to assessing risk and access control. As for studies applying BS7799 to the medical industry, Janczewski (2002) [15] researched the use of BS7799 to develop the Healthcare Information System (HIS), and investigated the basic security infrastructure behind the development of HIS.

3 Research Methods The main purpose of this study was to help medical organizations to use the most appropriate resources to build the most effective mechanism for information security management. This study uses the control items and the control objectives of the BS7799 information security management standard as the foundation to establishing a mechanism for the information security management suitable for medical organizations. This study first identified all information assets, information security threats, and information security vulnerabilities within the organization based on BS7799's one hundred twenty-seven control objectives, and formed a risk assessment scale through the assessment of the probability of occurrence and the degree of impact posed by security events. Then, according to the assessments of experts, we set weightings, ranked the information security events in order of priority, and structured a security management model. Carroll [1] proposed a mathematical formula for using annualized loss expectancy (ALE) to assess the necessary cost of asset protection and losses due to information threats. The formula is as follows:

ALE = TV .

(1)

Here, T is the annual value of a given threat, while V is the estimated value of assets. To evaluate the value at risk (VaR) of BS7799 control objectives, we convert the value T into the control objective's probability of occurrence (value P), and convert the value V into the control objective's degree of impact (value I).The ALE is then regarded as evaluating the VaR for each information security control objective.

P × I = VaR.

(2)

Using the 127 control objectives in BS7799 as a framework, we build an assessment table for the information security management according to each objective’s

414

S.-F. Liu, H.-E. Chueh, and K.-H. Liao Table 1. The comparative table for the value P (probability) and value I (impact) Probability (P) 0.00~0.20 0.20~0.40 0.40~0.60 0.60~0.80 0.80~1.00

Comparative Value 1 2 3 4 5

Impact (I) No Impact Slight Impact Some Impact Large Impact Significant Impact

Comparative Value 1 2 3 4 5

Table 2. The comparative table for the value P (probability) and value I (impact) Risk Quadrant Q1 Q2 Q3 Q4

Value P (Probability) 1

≦

3P 1 P 3P

≦5 ≦3 ≦5

Value I (Impact) 1I 3 3I 5 3I 5 1I 3

≦ ≦ ≦ ≦

probability of occurrence (value P) and their degree of impact. The comparative tables of probability and impact are shown in Tables 1. To more effectively evaluate information security risk and select the appropriate response, Halliday (1996) [13] again used the numerical values 1 to 5 for the probability of occurrence and the degree of impact to express respective risk quadrants, as shown in Table 2 and Fig. 1.

Fig. 1. Risk quadrants (Halliday, 1996) [13]

Information security management mechanisms constructed in this way are not always the best model. Therefore, they must be continually assessed to maintain effectiveness in risk management.

4 Experiment and Results The main purpose of this study was to establish a risk assessment table for the control items of the BS7799 information security management standard, and evaluate

The Use of BS7799 Information Security Standard to Construct Mechanisms

415

(through expert interviews), their probability of occurrence (value P) and degree of impact (value I). In this manner, we hoped to establish the VaR in the BS7799 information security management assessment table. The steps taken are described below: (1) Questionnaire Design: We produced a questionnaire based on the one hundred twenty-seven control items of the BS7799 information security management standard. (2) Expert Recommendations: We invited six information security experts to provide relevant values for the probability of occurrence and degree of impact for the questionnaire, as shown in Table 5. (3) Retrieval of Questionnaire and Data Analysis: We recovered the expert opinions and calculated the value at risk. (4) Synthesis of Risk Assessment Data: We established a mechanism for the management of information security for medical organization according to the expert opinions. This management mechanisms constructed in this study’s can be seen in Table 3. Table 3. Information security management mechanisms Risk Quadrant

Q2

Q3

Q4

Control Item

Procedural Order

1-1, 2-3, 2-4, 3-1, 4-5, 4-6, 4-7, 4-10, 5-1, 5-2, 5-3, 5-8, 5-9, 6-1, 6-2, 6-3, 6-8, 6-9, 6-10, 6-11, 6-13, 6-14, 6-15, 6-17, 6-18, 6-23, 7-2, 7-4, 7-6, 7-7, 7-8, 7-9, 7-11, 7-13, Highest priority 7-14, 7-15, 7-17, 7-19, 7-20, 7-21, 7-23, 7-25, 7-26, 727, 7-28, 8-1, 8-2, 8-10, 8-13, 8-14, 8-17, 8-18, 9-1, 9-2, 9-3, 10-2, 10-3, 10-4, 10-6. 2-2, 2-6, 2-8, 5-6, 5-12, 5-13, 6-4, 6-6, 6-12, 6-16, 6-19, 6-20, 6-21, 6-22, 7-3, 7-10, 7-12, 7-16, 7-24, 8-3, 8-4, 8- Needs to be processed 5, 8-6, 8-7, 8-8, 8-9, 8-11, 8-12. 1-2, 4-1, 4-8, 5-7, 6-5, 7-1, 7-5, 7-30, 8-15, 8-16, 10-5. Needs to be processed

5 Conclusion Given the competitive climate in the medical industry, with the expansion of hospitals, changes in government policy, and increasing demand for quality medical service, many hospitals have had to face unprecedented operational pressures, forcing hospital operators to pay closer attention to costs. The mechanisms for the information security management developed in this study could help medical organizations to assess their level of information security risk and identify appropriate improvement strategies.

References 1. Arthur, E.H., Bosworth, S., Hoyt, D.B.: Computer Security Handbook. John Wiley & Sons, New York (1995) 2. Badenhorst, K.P., Elloff, J.H.P.: Framework of a Methodology for the Life Cycle of Computer Security in an Organization. Computer & Security 8(5), 433–442 (1989)

416

S.-F. Liu, H.-E. Chueh, and K.-H. Liao

3. Christophy, A., Dorofee, A.: Introduction to the OCTAVE Method. The CERT® Coordination Center, CERT/CC (2001) 4. Ellison, R.J., Linger, R.C., Longstaff, T., Mead, N.R.: Survivable Network System Analysis: A Case Study. IEEE Software 16(4), 70–77 (1999) 5. Eloff, J.H.P., Eloff, M.M.: Information security architecture. Computer Fraud & Security 11, 10–16 (2005) 6. Eloff, M.M., Von Sloms, S.H.: Information Security Management: A Hierarchical Framework for Various Approaches. Computers & Security 19(3), 243–256 (2000) 7. Eloff, M.M., Von Sloms, S.H.: Information Security Management: An approach to Combine Process Certification and Product Evaluation. Computers & Security 19(8), 698–709 (2000) 8. Ettinger, J.E.: Key Issues in Information Security. Information Security. Chapman & Hall, London (1993) 9. Finne, T.: Information Systems Risk Management: Key Concepts and Business Processes. Computers & Security 19(3), 234–247 (2000) 10. Gehrke, M., Pfitzmann, A., Rannenberg, K.: Information Technology Security Evaluation Criteria (ITSEC)-A Contribution to Vulnerability? In: The IFIP 12th World Computer Congress Madrld on Information Processing, pp. 7–11 (1992) 11. Gollmann, D.: Computer Security. John Wiley & Sons Ltd., UK (1999) 12. Gupta, M., Chartuvedi, A.R., Metha, S., Valeri, L.: The Experimental Analysis of Information Security Management Issues for Online Financial Services. In: The 2001 International Conference on Information Systems, pp. 667–675 (2001) 13. Halliday, S., Badenhorst, K., Von Solms, R.: A business approach to effective information technology risk analysis and management. Information Management & Computer Security 4(1), 19–31 (1996) 14. ISO/IEC 17799. Information technology-code of practice for information security management. BSI, London (2000) 15. Janczewski, L.J., Shi, F.X.: Development of Information Security Baselines for Healthcare Information Systems in New Zealand. Computer & Security 21(2), 172–192 (2002) 16. Schultz, E.E., Proctor, R.W., Lien, M.C.: Usability and Security An Appraisal of Usability Issues in Information Security Methods. Computers & Security 20(7), 620–634 (2001) 17. Sherwood, J.: SALSA: A method for developing the enterprise security architecture and Strategy. Computer & Security 2(3), 8–17 (1996) 18. Smith, E., Eloff, J.H.P.: Security in health-care information systems-current trends. International Journal of Medical Informatics 54, 39–54 (1999) 19. Song, M.J.: Risk Management. Chinese Enterprise Develop Center, 33–456 (1993) 20. Trcek, D.: An Integral Framework for Information Systems Security Management. Computers & Security 22(4), 337–360 (2003) 21. Von Solms, R.: Information Security Management: The Second Generation. Computer & Security 15(4), 281–288 (1996) 22. Von Solms, R., Van Haar, H., Von Solms, S.H., Caelli, W.J.: A Framework for Information Security Evaluation. Information & Management 26, 143–153 (1994) 23. Willet, A.H.: The Economic Theory of Risk and Insurance. Ph. D. Thesis in Columbia University (1901)

An Improved Frame Layer Rate Control Algorithm for H.264 Xiao Chen and Feifei Lu School of Electronic and Information Engineering, Nanjing University of Information Science and Technology, Ningliu Road 219, 210044 Nanjing, China [email protected]

Abstract. Rate control is an important part of video coding. This paper presents an improved frame layer rate control algorithm by using the combined frame complexity and the adjusted quantization parameter (QP). The combined frame complexity can be used to more reasonable implement bit allocation for each frame, and the quantization parameter adjusted by the encoded frame information also can get more accurate rate control. Experimental results show that our proposed algorithm, in comparison to the original algorithm, reduces the act bit rate error of video sequences and gets a better average PSNR with smaller deviation. Keywords: H.264; video coding; rate control; frame complexity; QP adjustment factor.

1 Introduction The principle task of rate control in video communication is to collect the buffer status and image activity. It also allocates a certain number of bits for each image of video in the purpose of controlling the output rate and minimizing the image distortion. In the rate control algorithm for H.264/AVC, the quantization parameter is used in both the rate control and rate distortion optimization (RDO), which leads to chicken and egg dilemma [1].In order to solve the chicken and egg dilemma, many scholars have done a lot of researches. The work in [2] solves the dilemma by enhancing the ρ domain model. The relational model between rate and quantization step is advanced in [3] for the dilemma. Besides those, the scheme represents a new rate control algorithm with the comprehensive consideration of HRD consistency and the ratio of the mean absolute difference (MAD) in [4]. Typically, Li et al. in JVT-G012 [5] have proposed a linear model for MAD prediction to solve the chicken and egg dilemma, this method can obtain good coding result. Although JVT-G012 proposal well solves the dilemma, there are some problems of the proposal. Due to no explicit R-Q model for an intraframe discussed, the scheme [6] introduces an adaptive intraframe ratequantization(R-Q) model. The proposed method aims at selecting accurate quantization parameters for intra-coded frames. The work in [7] proposes separable R-D models for color video coding. Also the rate control algorithm for JVT-G012 has R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 417–421, 2011. © Springer-Verlag Berlin Heidelberg 2011

418

X. Chen and F. Lu

some lack of bit allocation in frame layer, it allocates bits based on the buffer status without considering the frame complexity, and it does not take the characteristics of the encoded frames impact on the current frame into account when calculating QP. Therefore, this paper proposes an improved frame layer rate control algorithm. By allocating bits in frame layer based on the frame complexity and adjusting the QP based on the encoded frame, our method can be more effective both in achieving rate control and improving the image quality.

2 Improved Frame Layer Rate Control Algorithm 2.1 Target Bit Allocation with the Consideration of Frame Complexity As the rate-distortion(R-D) model and the linear model for MAD prediction are not accurate, we need to reduce the differences between the actual bits and target bits by calculating a target bits for each frame. The frame layer rate control algorithm for JVT-G012 allocates the bits for non encoded frames on average based on the buffer status. So it is difficult to reach the required coding bits with the same quality due to the different frame complexity. Thus we must consider the frame complexity when allocating the bits in frame layer, which can get more accurate results. Our frame complexity FC is defines as follows [8]:

FC = μMADratio (i, j ) + (1 − μ )C j , where

(1)

MADratio (i, j ) is the ratio of the predicted MAD of current frame j to the

average

MAD

Cj = H j

of

all

H j −1 , H j

previously encoded P frames in the ith GOP, is the average difference of gray histogram between the

current frame and the previous reconstruction frame, and μ is a weighting coefficient. The target bits allocated for the jth frame in the ith group of pictures (GOP) in JVT-G012 is determined by frame rate, target buffer size, actual buffer occupancy and the available channel bandwidth ⎧ Tr (ni, j ) ⎧u(ni, j ) ⎫ +(1− β)×⎨ +γ Tbl (ni, j ) − Bc (ni, j ) ⎬ 0 ≤ FC≤1.1 ⎪0.88×FC×β × Nr ⎩ Fr ⎭ ⎪ ⎪ T (n ) ⎧u(n ) ⎫ ⎪ Ti ( j) = ⎨[0.8×(FC−1.15) +1.1]×β × r i, j +(1− β)×⎨ i, j +γ Tbl(ni, j ) − Bc (ni, j ) ⎬ 1.1< FC≤ 2.1 , N F r ⎩ r ⎭ ⎪ ⎪ Tr (ni, j ) u ( n ) ⎧ i, j ⎫ ⎪1.15×β × +(1− β)×⎨ +γ Tbl(ni, j ) − Bc (ni, j ) ⎬ FC 2.1 ⎪⎩ Nr F ⎩ r ⎭

[

]

[

[

]

(2)

]

where Ti ( j ) is the target bits allocated for the jth frame,

Fr is the predefined frame

rate, u ( ni , j ) is the available channel bandwidth for the sequences, Tbl (ni , j ) is the target buffer level, Bc (ni , j ) is the occupancy of virtual buffer,

β

its value is 0.5 when there is no B Frame and 0.9 otherwise,

is a constant and its

γ

is a constant and

An Improved Frame Layer Rate Control Algorithm for H.264

419

value is 0.75 when there is no B Frame and 0.25 otherwise, and T ' = Tr (ni , j ) , Tr ( ni , j ) r Nr

is the remaining bits of all non coded frames in the ith GOP, and N r is the number of P-frames remaining to be coded. The formula allocates smaller bits for a frame with lower frame complexity measure, whereas the more bits for a frame with higher complexity measure. 2.2 Adjustment of Quantization Parameter The quadratic R-D model has two key parameters: MAD and the header bits. The inaccuracy of two parameters should lead QP calculated by the quadratic R-D model can not produce the desired coded bits. Therefore, we must adjust the current QP by the previous frame coding information, which can achieve more accurate rate control. When we obtain the value of Ti ( j ) in the use of the improved algorithm, we calculate the quantization parameter using the R-Q model. To consider the feedback information of the coded frames, we adapt the quantization parameter adjustment factor ΔQ to adjust the value of QP. The value of ΔQ is determined by the actual bits of the previous frame and the target bits of the previous frame.

⎧2 if pretarbits ≤ preactbits 2 pretarbits . ΔQ = ⎨ 3 if preactbits ≥ 2 pretarbits ⎩

(3)

When the actual bits is larger than the target bits, it shows the previous frame gets less target bits, meanwhile, the algorithm allocates more target bits to the current frame, so the quantization parameter calculated by the quadratic R-D model is smaller, we should increase the QP appropriately. In GOP, QP of I-frame and the first P-frame are pre-defined. Because there is no bit allocation in the first P-frame, QP of the second P-frame is not adjusted, and the quantization parameter of the following P-frames is computed by the quantization parameter adjustment factor.

⎧ QPj j = 2 QPj = ⎨ . ⎩QPj + ΔQ j 2

(4)

After the adjustment, the algorithm has taken full the coded frame information into account, it achieves a good rate control, and then we perform RDO.

3 Experimental Results We have implemented our proposed rate control algorithm by enhancing the JM8.6 test model software. The JVT-G012 rate control algorithm is selected as a reference for comparison (as is implemented on reference software JM8.6). The tested sequences are in QCIF4:2:0 formats: suzie, football, mobile, foreman and coastguard. In the experiments, the frame rate is set to 15 frames per second, the target bit rate is set to 64kb/s, the total number of frames is set to 100, the initial quantization parameter is set to 28 and the length of GOP is set to 25.

420

X. Chen and F. Lu

The experimental results are shown in Tab. 1. As summarized in Tab. 1, our proposed rate control algorithm can control the bit rates more accurately. The maximum error of the actual bit rates is 0.91%, which is reduced by 2 times compared with the maximum error 2.45% of the original algorithm. Meanwhile, the average error of the actual bits is 0.48%, which is reduced by more than 50% compared with the average error 1.03% of the original algorithm. The proposed algorithm also improves the average PSNR and PSNR deviation significantly for the sequences. In Tab. 1, it shows that our method achieves an average PSNR gain of about 0.54dB with similar or lower PSNR deviation as compared to the JVT-G012 algorithm. Especially for the sequence football and the sequence mobile, which has high motion and complex texture, the proposed algorithm achieves the PSNR gain of 0.85dB and 1.31dB separately. The improvement of the sequence foreman with moderate texture is not obvious. For most sequences, the proposed algorithm obtains lower PSNR deviation compared with the JVT-G012 algorithm. This shows the proposed algorithm can smooth the PSNR fluctuation between frames to some extent. Table 1. Performance comparison for the proposed algorithm with the JVT-G012 algotithm on JM8.6 Test sequences suzie football mobile foreman coastguard

bitrate(kb/s) JM8.6 proposed 63.81 64.13 65.57 64.18 65.01 64.58 64.18 64.32 64.35 64.33

PSNR(dB) JM8.6 proposed 37.63 37.98 24.22 25.07 26.11 27.42 36.15 36.19 39.73 39.86

PSNR deviation JM8.6 proposed 1.96 1.68 4.19 3.75 3.24 2.16 0.96 0.90 0.77 0.86

4 Conclusion In this paper, we propose an improved frame layer rate control algorithm by using the combined frame complexity and the adjusted quantization parameter. The algorithm allocates bits in frame layer according to the frame complexity, and computes the quantization parameter of current frame with the consideration of the previous frame information. The experimental results show that our algorithm achieves more accurate rate control and better average PSNR.

Acknowledgments This work was supported by "Qing Lan Gong Cheng" program of Jiangsu Province of China and National Natural Science Foundation of China (No. 10904073).

References 1. Ma, S.W., Gao, W., Wu, F., Lu, Y.: Rate control for JVT video coding scheme with HRD considerations. In: Proceeding of IEEE International Conference on image and Processing, vol. 3, pp. 793–796 (2003)

An Improved Frame Layer Rate Control Algorithm for H.264

421

2. Shin, I.H., Lee, Y.L., Park, H.W.: Rate control using linear rate-ρ model for H.264. Signal Process Image Communication. 19, 341–352 (2004) 3. Ma, S.W., Gao, W., Lu, Y.: Rate-distortion analysis for H.264/AVC video coding and its application to rate control. IEEE Trans. on Circuit Syst. for Video Technol. 15, 1533–1544 (2005) 4. Li, Z.G., Gao, W., Pan, F.: Adaptive rate control for H.264. Journal of Visual Communication and Image Representation 17(2), 376–406 (2006) 5. Li, Z.G., Pan, F., Lim, K.P.: Adaptive base unit layer rate control for JVT, JVT-G012. In: Proceedings of 7th Meeting, Pattay II, pp. 7–14. IEEE Press, Thailand (2003) 6. Jing, X., Chau, L.P., Siu, W.C.: Frame complexity-based rate-quantization model for H.264/AVC intraframe rate control. IEEE Signal Processing Letters 15, 373–376 (2008) 7. Chen, Z.Z., Ngan, K.N.: Towards rate-Distortion tradeoff in real-time color video coding. IEEE Trans.Circuits Syst.Video Technol. 17, 158–167 (2007) 8. Chen, X., Lu, F.: An Improved Rate Control Scheme for H.264 Based on Frame Complexity Estimation. Journal of Convergence Information Technology (accepted)

Some Properties in Hexagonal Torus as Cayley Graph Zhen Zhang1,2 1

Department of Computer Science, Jinan University, Guangzhou 510632, P.R. China 2 Department of Computer Science, South China University of Technology Guangzhou 510641, P.R. China [email protected]

Abstract. Vertexes in the hexagonal mesh and torus network are placed at the vertices of a regular triangular tessellation, so that each node has up to six neighbors. The network is proposed as an alternative interconnection network to mesh connected computer and is used also to model cellular networks where vertexes are based stations. Hexagonal tori are known to belong to the class of Cayley graphs. In this paper, we use Cayley-formulations for the hexagonal torus to develop an elegant routing and broadcasting algorithm. Keywords: Hexagonal Broadcasting.

torus;

Cayley

graph;

Addressing;

Routing;

1 Introduction Hexagonal networks belong to the family of networks modeled by planar graphs. These networks are based on triangular plane tessellation, or the partition of a plane into equilateral triangles. The closest networks are those based on regular hexagonal, called honeycomb networks, and those based on regular square partitions, called mesh networks. Hexagonal networks and honeycomb have been studied in a variety of contexts. They have been applied in chemistry to model benzenoid hydrocarbons [5], in image processing, in computer graphics [6], and in cellular networks [2]. An addressing scheme for the processors, and the corresponding routing and broadcasting algorithms for a hexagonal interconnection network has been proposed by Chen et al.[1]. The approach proposed in [1] is a cumbersome addressing scheme which has lead to a complex routing algorithm, and similarly to a complex broadcasting scheme. Consequently, the lack of a convenient addressing scheme and the absence of elegant routing and broadcasting algorithms have discouraged further research on this type of network. Carle and Myoupo [7] recently revisited this network and attempted to simplify the addressing, routing and broadcasting schemes given in [1] with partial success. They suggested a co-ordinate system for hexagonal networks that uses two axes, at 120° between them, which are parallel to two out of the three edge directions. Using this scheme, they have described routing and broadcasting algorithms for the network. However, their scheme exhibits asymmetry R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 422–428, 2011. © Springer-Verlag Berlin Heidelberg 2011

Some Properties in Hexagonal Torus as Cayley Graph

423

which complicates the routing algorithm. Their broadcasting algorithm, on the other hand, is very elegant. Xiao, etc. propose that hexagonal mesh and torus, as well as honeycomb and certain other pruned torus networks, are belonged to the class of Cayley graphs, and they also possess other interesting mathematical properties[3, 4]. Using Xiao’s scheme, Zhang et al gave the optimal routing algorithm [8]. But their routing algorithm is very complicated. And they also gave an upper bound of the diameter. The rest of this paper is organized as follows. In section 1, we give some definitions of Cayley graphs that are used in this paper, and we also propose a co-ordinate system for hexagonal networks that uses two axes: x and y. According to this addressing scheme, a simple routing algorithm is developed in Section 3, and we will prove that this algorithm is optimal. We also give a broadcasting algorithm by using the Cayley graph and Coset graph properties in Section 4. The paper concludes with Section 5.

2 Hexagonal Mesh and Torus as Cayley Graph 2.1 Hexagonal Mesh Let G = Z × Z where Z is the infinite cyclic group of integers, and consider Г=Cay(G, S) with S={(±1, 0), (0, ±1), (1, 1), (-1, -1)}. It is evident that Г is isomorphic to the hexagonal mesh network [2]. Fig. 1 shows a small part of an infinite hexagonal mesh in which the six neighbors of the “center” node (0,0) are depicted. Using the Cayley-graph formulation of hexagonal networks, we can easily derive the distance dis((a, b), (c, d)) between the vertices (a, b) and (c, d) in such networks[3]. The routing algorithm of hexagonal mesh has been developed in [3] as the follow Lemma. Lemma 1. In the hexagonal mesh Г, dis((0, 0), (a, b)) equals max((|a|, |b|) if a and b have the same sign and |a|+|b| otherwise. Proof. See in [3].

■

By symmetry of Cayley graphs, we can easily obtain the distance between any two vertices in the graph Γ from Lemma 1, using dis((a, b), (c, d))=dis((0, 0),(c-a, d-b)). This observation and the preceding discussion lead to a simple distributed routing algorithm for Г. 2.2 Hexagonal Torus Let G = Z l × Z k , where Z l and Z k are cyclic groups of orders l and k respectively, l 0 , k 0 . Assume that S is defined as in the preceding paragraph. Then Δ=Cay(G, S) is the hexagonal torus of order l×k, which can be denoted by Hl×k, Fig. 2 shows hexagonal torus H9×5

424

Z. Zhang

Fig. 1. Connectivity pattern for hexagonal mesh network

Fig. 2. Hexagonal torus H9×5

We can adapt a coordinate system for hexagonal networks that uses two axes: x and y, as shows in Fig. 2. Using the results obtained for hexagonal meshes according to Lemma 1, we can deal with problems on Hexagonal torus which are, in general, more difficult. Then, we have the following result. Lemma 2. For the hexagonal torus Hl×k and integers a and b, l a ≥ 0 , k b ≥ 0 ,we have dis((0, 0), (a, b))=min(max(a, b), max(l-a, k-b), l-a+b, k+a-b). Proof: See in [4].

■

According to the Lemma 2 and the properties of Coset graph, we can develop routing and broadcasting algorithm of the hexagonal torus.

3 Optimal Routing Algorithm According to Lemma 2, we provide a routing algorithm Route(p=(a, b)). This algorithm is fully distributed, in the sense that it quickly determines the next vertex p’ on a shortest path from the current vertex p to the destination vertex e.

Some Properties in Hexagonal Torus as Cayley Graph

425

Route(p=(a, b)) //returns p’=(a’, b’), the first vertex on a path from p to e=(0,0) {compute d=dis((0, 0), (a, b)); if d=0 then success; if d=max(a, b) then {if b=0 then {a’=a-1; b’=b;} if a=0 then {a’=a; b’=b-1;} if a0 and b0 then {a’=a-1; b’=b-1;}} if d=max(l-a, k-b) then {if k-b=0 then {a’=a+1; b’=b;} if l-a=0 then {a’=a; b’=b+1;} if l-a0 and k-b0 then {a’=a+1; b’=b+1;}} if d=l-a+b then {if b0 then {a’=a; b’=b-1;} if b=0 then {a’=a+1; b’=b}} if d=k-b+a then {if a0 then {a’=a-1; b’=b;} if a=0 then {a’=a; b’=b+1;}} } Theorem 1. The algorithm Route is optimal. Proof. Let d’=dis((0, 0), (a’, b’)), then we only need to prove that d’=d-1. As an obvious result, we know that |d-d’|≤1. z Case 1: d=max(a, b). 1. Subcase a: b=0 and a0, that is, d=a. Let a’=a-1 and b’=b, then we have max(a’, b’)=a-1. By Lemma 2, we can get d’ ≤max(a’, b’)=a-1=d-1, that is d’=d-1. 2. Subcase b: a=0 and b0, that is, d=b. Let a’=a and b’=b-1, then we have max(a’, b’)=b-1. By Lemma 2, we can get d’ ≤max(a’, b’)=a-1=d-1, that is d’=d-1. 3. Subcase c: a0 and b0. Let a’=a-1 and b’=b-1, then we have max(a’, b’)= max(a-1, b-1)=d-1. By Lemma 2, we can get d’ ≤max(a’, b’)=d-1, that is d’=d-1. z Case 2: d=max(l-a, k-b). 1. Subcase a: k-b=0 and l-a0, that is, d=l-a. Let a’=a+1 and b’=b, then we have max(l-a’, k-b’)=l-a-1=d-1. By Lemma 2, we can get d’ ≤max(l-a’, kb’)=d-1, that is d’=d-1.

426

Z. Zhang

2.

Subcase b: l-a=0 and k-b0, that is, d=k-b. Let a’=a and b’=b+1, then we have max(l-a’, k-b’)=k-b-1=d-1. By Lemma 2, we can get d’ ≤max(l-a’, kb’)=d-1, that is d’=d-1. 3. Subcase c: l-a0 and k-b0. Let a’=a+1 and b’=b+1, then we have max(la’, k-b’)=max(l-a-1, k-b-1)=d-1. By Lemma 2, we can get d’ ≤max(l-a’, kb’)=d-1, that is d’=d-1. Case 3: d=l-a+b. 1. Subcase a: b0. Let a’=a and b’=b-1, then we have l-a’+b’=l-a+b-1=d-1. By Lemma 2, we can get d’ ≤l-a’+b’ =d-1, that is d’=d-1. 2. Subcase b: b=0. Let a’=a+1 and b’=b, then we have l-a’+b’=l-a-1+b=d-1. By Lemma 2, we can get d’ ≤l-a’+b’ =d-1, that is d’=d-1. Case 4: d=k-b+a. 1. Subcase a: a0. Let a’=a-1 and b’=b, then we have k-b’+a’=k+a-1-b=d-1. By Lemma 2, we can get d’ ≤k-b’+a’ =d-1, that is d’=d-1. 2. Subcase b: a=0. Let a’=a and b’=b+1, then we have k-b’+a’=k+a-b-1=d-1. By Lemma 2, we can get d’ ≤k-b’+a’ =d-1, that is d’=d-1. ■

z

z

4 Broadcasting Algorithm Given a connected graph G and a message originator u, the broadcast time bM(u) of the vertex u is the minimum time required to complete broadcasting from vertex u under the model M. The broadcast time bM(G) of G under M is defined as the maximum broadcast time of any vertex u in G, i.e. bM(G)= max{bM(u)|u V (G)}. In[9], Xiao proposed the upper bound of bM(G) based on the mathematical properties of Cayley graph and Coset graph.

∈

Theorem 2. Let G be a finite group and K ≤ G. Assume that Γ= Cay(G, S) and Δ= Cos(G, K, S) for some generating set S of G. For a communication model M, let bM(ΓK) be the minimum time required to complete broadcasting in the vertices of K from the identity element e. Then, we have bM(Γ) ≤ bM(Δ) + bM(ΓK). Proof. See in [15].

■

Let G= Zl×Zk, S={(±1, 0), (0, ±1), (1, 1), (-1, -1)}. According to the Theorem 2, we develop a broadcasting algorithm for the hexagonal torus Γ=Cay(G, S). Let K={(z,0)| z Zl}, and then we can get K ≤Zl×Zk. It is obviously that ΓK=Cay(K,(±1, 0)) is a cycle, and it is a subgraph of Γ. The Coset graph Δ= Cos(G, K, S) is a cycle too. Without loss of generality, we assume the identity element e=(0, 0) be the source vertex.

∈

Broadcasting Algorithm: Procedure for the source vertex e=(0, 0): {send(message, (1, 0), 0); send(message, (l-1, 0), 0); send(message, (0, 1), 1); send(message, (0, k-1), 2);} Procedure for all vertexes except the source vertex: {receive(message, (x’, y’), C);

Some Properties in Hexagonal Torus as Cayley Graph

427

if (C= ⎢ k ⎥ or ⎡ k ⎤ ) then ⎢⎣ 2 ⎥⎦ ⎢⎢ 2 ⎥⎥ stop; if (y=0) then if (x ⎢ l ⎥ ) then ⎢⎣ 2 ⎥⎦ {send(message, (x+1, y), C); send(message, (x, y+1), C+1); send(message, (x, y-1(mod k)), C+2); } else if (x ⎡ l ⎤ ) then ⎢⎢ 2 ⎥⎥ {send(message, (x-1, y), C); send(message, (x, y+1), C+1); send(message, (x, y-1(mod k)), C+2); } if (y0 and y ⎢ k ⎥ ) ⎢⎣ 2 ⎥⎦ send(message, (x, y+1), C+1); if (y ⎡ k ⎤ ) ⎢⎢ 2 ⎥⎥ send(message, (x, y-1(mod k)), C+1);} It is easy to clarify that bM(ΓK)= ⎡ l ⎤ and bM(Δ)= ⎡ k ⎤ , then the total number of com⎢⎢ 2 ⎥⎥ ⎢⎢ 2 ⎥⎥ munication steps is ⎡ l ⎤ + ⎡ k ⎤ . An illustrative example for the broadcasting algorithm ⎢⎢ 2 ⎥⎥ ⎢⎢ 2 ⎥⎥ in H9×5 is shown in Fig. 3.

Fig. 3. Broadcasting in hexagonal torus of H9×5

5 Conclusion A family of 6-regular graphs, called hexagonal mesh, is considered as multiprocessor interconnection network. Processing vertexes on the periphery of the hexagonal mesh are wrapped around to achieve regularity and homogeneity, and this type of graph is

428

Z. Zhang

also called hexagonal torus. In this paper, we use Cayley-formulations for the hexagonal torus to develop a simple routing algorithm. This routing algorithm is proved to be optimal. We also develop a broadcasting algorithm by the theory of Cayley graph and Coset graph. Then, we discuss the diameter of the hexagonal torus, and get exact value of the diameter. There are many interesting problems to be pursued for the hexagonal torus architecture, such as fault-tolerant routing, embedding properties, and the application of the hexagonal torus to solve or reduce the complexity of some difficult problems. These topics are all closely related to the associated routing and broadcasting algorithms and will be addressed in our future works.

Acknowledgments This paper is supported by the Fundamental Research Funds for the Central Universities (21610307) and Training Project of Excellent Young Talents in University of Guangdong (LYM09029).

References 1. Chen, M.S., Shin, K.G., Kandlur, D.D.: Addressing, Routing and Broadcasting in Hexagonal Mesh Multiprocessors. IEEE Trans. Computers 39(1), 10–18 (1990) 2. Nocetti, F.G., Stojmenovic, I., Zhang, J.Y.: Addressing and Routing in Hexagonal Networks with Applications for Tracking Mobile Users and Connection Rerouting in Cellular Networks. IEEE Trans. 3. Xiao, W.J., Parhami, B.: Further Mathematical Properties of Cayley Graphs Applied to Hexagonal and Honeycomb Meshes. Discrete Applied Mathematics 155, 1752–1760 (2007) 4. Xiao, W.J., Parhami, B.: Structural Properties of Cayley Graphs with Applications to Mesh and Pruned Torus Interconnection Networks. Int. J. of Computer and System Sciences, Special Issue on Network-Based Computing 73, 1232–1239 (2007) 5. Tosic, R., Masulovic, D., Stojmenovic, I., Brunvoll, J., Cyvin, B.N., Cyvin, S.J.: Enumeration of polyhex hydrocarbons up to h=17. J. Chem. Inform. Comput. Sci. 35, 181–187 (1995) 6. Laster, L.N., Sandor, J.: Computer graphics on hexagonal grid. Comput. Graph. 8, 401–409 (1984) 7. Carle, J., Myoupo, J.F.: Topological properties and optimal routing algorithms for three dimensional hexagonal networks. In: Proceedings of the High Performance Computing in the Asia-Pacific Region HPC-Asia, Beijing, China, vol. I, pp. 116–121 (2000) 8. Zhang, Z., Xiao, W., He, M.: Optimal Routing Algorithm and Diameter in Hexagonal Torus Networks. In: Xu, M., Zhan, Y.-W., Cao, J., Liu, Y. (eds.) APPT 2007. LNCS, vol. 4847, pp. 241–250. Springer, Heidelberg (2007) 9. Xiao, W.J., Parhami, B.: Some Mathematical Properties of Cayley Graphs with Applications to Interconnection Network Design. Int. J. Computer Mathematics 82, 521–528 (2005)

Modeling Software Component Based on Extended Colored Petri Net* Yong Yu1,2, Tong Li1,2, Qing Liu1,2, and Fei Dai1 2

1 School of Software, Yunnan University, Kunming, China Key Laboratory in Software Engineering of Yunnan Province, Kunming, China [email protected], [email protected], [email protected], [email protected]

Abstract. In component-based software engineering, components have become increasingly important. Software component is a unit of computing in software system. First, the definition of OR-transition Colored Petri Net is given. Second, in accordance with the related properties of software component, the formal definition of component is presented. In OR-transition Colored Petri Net, the transitions can effectively represent the operations of the software component. So, based on OR-transition Colored Petri Net, an approach is put forward to modeling the software component formally. Keywords: Petri net, Software, Component.

1 Introduction With the development of computer technology, the software requirements are growing rapidly, but the current software development is not able to fulfill the demand. In order to solve the problem of the industrialization of software, people put forward the concept of software component. Large complex software systems are composed of many software components. Building software systems from reusable software components has long been a goal of software engineers. While other engineering disciplines successfully apply the reusable component approach to build physical systems, it has proven more difficult to apply in software engineering. A primary reason for this difficulty is that distinct software components tend to be more tightly coupled with each other than most physical components [1]. A component is simply a data capsule. Thus information hiding becomes the core construction principle under*

This work has been supported by the National Science Foundation of China under Grant No. 60963007, by the Science Foundation of Yunnan Province, China under Grant No. 2007F008M, the Key Subject Foundation of School of Software of Yunnan University and the Open Foundation of Key Laboratory in Software Engineering of Yunnan Province under Grant No. 2010KS01, the promotion program for youth key teachers of Yunnan university No.21132014, by the Science Foundation of Yunnan Province Education Department No. 09J0037 and Yunnan University, China under Grant No. ynuy200920.

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 429–434, 2011. © Springer-Verlag Berlin Heidelberg 2011

430

Y. Yu et al.

lying components. A component can be implemented in (almost) any language, not only in any module-oriented and object-oriented languages but even in conventional languages [2,3]. For component, it is unavoidable to interact with the environment. The component can be obtained information from the environment, and provides relaxed services to the environment. Component interface represents the interaction with the outside world, each interface show interaction between components and connector [2,3]. Components are the special status of the software system, so the formal description of the components is very important. This component of the software system based on the characteristics and nature, according to the characteristics of component elements of the formal definition is given, and the colored Petri nets, or change components of the extended component net (C-net) in order to achieve the software components of effective description.

2 OR-Transition Colored Petri Net An OR-transition colored Petri net can be defined as follows [4]: 2.1 OR-Transition Petri Net Definition 1. An OR-transition Petri net system is a 4-tuple ∑=P where:

∪

，T，F，M , 0

1) P T≠Φ and P∩T=Φ; 2) F⊆(P×T) (T×P); 3) M0⊆P is the initial mark of the OR-transition Petri-net system; 4) A transition t∈T is enabled in a marking M iff ∃p∈˙t, M (p)=1 and ∀p′∈t˙, M(p′)=0. It is said that the transition t is enabled under the mark M and the place p.

∪

Let a transition t∈T fires under a mark M and a place p, the mark M is transformed into the mark M'; we often say that the mark M' is reachable from the mark M in a step. M′ is the successor mark of M under t and p. It is written as M(p)[tM'. where: ∀p′∈P: ⎧ M(p')-1, p'=p; ⎪ M'(p')= ⎨ M(p')+1, p' ≠ p and p' ∈ t ⋅ ; ⎪ M(p'), else. ⎩

，，，

Definition 2. In an OR-transition Petri net system ∑=P T F M0, the corresponding underlying net N=P T, F is called as OR-transition Petri net.

,

∪，

Definition 3. In an OR-transition Petri net ORPN=P, T, F, let x, y∈T P ∃b1, b2, b3, …, bk∈T P, such that x, b1, b1, b2, b2, b3, … , bk, y∈F, then we say that y is structure-reachable from x, which is denoted as xF＊y.

∪

Modeling Software Component Based on Extended Colored Petri Net

431

2.2 OR-Transition Colored Petri Net Definition 4. 1) S is a limited and non-empty type set, also known as the color set; 2) The multi-set m is a function of non-empty color set S: m∈(S→N). For the nonempty set S, m= ∑ m(s)s is the multi-set of S, m(s)≥0 is called the coefficient of s. s∈S

3) Let SMS be the set of all multi-sets of based on S, and m, m1, m2∈SMS, n∈N then: (1) m1+m2= ∑ (m1(s)+m2(s))s; s∈S

(2) n × m = ∑ (n × m (s ))s; s∈ S

(3) m1≠m2 ≡ ∃s∈S: m1(s)≠m2(s); (4) m1≤m2 ≡ ∀s∈S: m1(s)≤m2(s); (5) m1≥m2 ≡ ∀s∈S: m1(s)≥m2(s); (6) If m1≤m2, m2-m1= ∑ (m2(s)-m1(s))s. s∈S

Definition 5. An Or-transition colored Petri net (ORCPN) is a 7-tuple N=P, T, F, S, AP, AT, AF, where:

1) P, T, F is an OR-transition Petri-net, which is called as the underlying net of ORCPN; 2) S is a non-empty color set, which is called as color set of ORCPN; 3) AP P→SS, AP is a function of P, where SS is the power set of S. 4) AT T→SMS, AT is a guard function of T, where SMS is the set of all multi-sets of based on S and it meets the following condition: ∀t∈T, AT(t) ∈ ( U Ap(p)) MS .

： ：

：

，

p∈ • t

5) AF F→SMS AF is the arc expression function, where SMS is the set of all multi-sets of based on S and meet the following condition: ∀f∈F, AF(f)∈(AP(P(f)))MS, where P(f) describes the corresponding place p of arc f. Definition 6. An OR-transition colored Petri net (ORCPN) system is an 8-tuple ORCPN system ∑=P, T, F, S, AP, AT, AF, M0, where:

1) N=P, T, F, S, AP, AT, AF is an OR-transition colored Petri net, which is called as the underlying net of ORCPN system; 2) M0 is the initial marking of ORCPN system ∑=P, T, F, S, AP, AT, AF, M0, and meets the following condition: ∀p∈P: M0(p)∈(AP(p))MS.

：

Definition 7. M P→SMS is the marking of ORCPN system ∑=P, T, F, S, AP, AT, AF, M0, where ∀p∈P: M(p)∈(AP(p))MS.

3 Software Component A software component is simply a data capsule. Thus information hiding becomes the core construction principle underlying components. A component can be implemented in (almost) any language, not only in any module-oriented and object-oriented languages but even in conventional languages [5].

432

Y. Yu et al.

In software architecture, a component should include two parts: interface and implementation. Interface defines the functions and specifications provided by the component, and implementation includes a series of related operations[6]. Therefore, in this paper, the definition of component is as follows: Definition 8. A component C in software architecture is a 3-tuple C = Interface, Imp, Spec, where: 1) Interface = IP OP, is a set of component interfaces, IP represents the input interfaces of component, OP represents the output interfaces; 2) Imp is the implementation of component, and it includes a series of operations: t1, t2, ..., tn; and each operation completes specific function; 3) Spec represents the internal specification of component, and it is mainly used to describe the relationships between the implementations and the interfaces.

∪

Definition 9. Each interface in component is a 2-tuple: p = ID, DataType, where ID is the unique identifier of the interface p, DataType is the type of the information which can be accepted by the interface p. In component, each input interface represents the certain set of some operations, a component can have some input interfaces, the outside environment can request services from one or more input interfaces of the component. The output interfaces of component describe the requests of the outside environment, when the component completes a function, it may need other components to provide some help. In component, the operation is complete certain function, and it can be defined as: Definition 10. An operation t is a 5-tuple t = S, D, R, PR(X), PO(X,Y), where: S is the syntax of the operation t, and Y = t(X). X is the input vectors of the operation t, and Y is the output vectors of the operation t; X = (x1, x2, ..., xm), Y = (y1, y2, ..., yn). D = D1×D2×...×Dm is the domain of the input vectors, xi∈Di (1≤i≤m). R = R1×R2×...×Rn is the range of the output vector, yj ∈Rj (1≤j≤n). Di, Rj is a legal data type. PR (X) is called pre-assertion; PO (X, Y) is called post-assertion. Satisfy the PR(X) of the input vector X is called the legitimate input. For legal input X, to meet the PO (X, Y) as the legitimate output of the output vector Y [7]. From the definition, the implementation of the operation t needs certain conditions, when the conditions are met, the related operations are implemented.

4 Modeling the Component The transitions in ORCPN can model the operations of the component defined above, therefore, it is straight-forward to map a software component into an OR-Transition colored Petri net. In component, the operations are modeled by transitions of the ORTransition colored Petri net and the states of the software component are modeled by places of the OR-Transition colored Petri net. The arrows between places and transitions are used to specify causal relations in the software component. And based-on ORCPN, we present component-net (CN for short) to model software component. Definition 11. A component-net C-net (CN) is a 9-tuple, CN = P, T, F, S, AP, AT, AF, IP, OP, it is extended from a colored Petri net, where:

Modeling Software Component Based on Extended Colored Petri Net

433

1) ORCPN = P, T, F, S, AP, AT, AF is a or-transition colored Petri net; 2) P is a finite set of places, and it presents the states of component; 3) T is a finite set of transitions, and it presents the operations of component; 4) F⊆P×T T×P is an arc set, it describes the constraint relations between states and operations in the component; 5) S is a non-empty finite set; it describes the data types of component C; 6) AT(t) presents input vectors, which can be accepted by operation t; 7) IP, OP(⊆P) are called the input interfaces and output interfaces of the component, and ∀ip∈ IP, ˙ip=∅; ∀op∈OP, op˙=∅.

∪

In a component net CN = P, T, F, S, AP, AT, AF, IP, OP: The places P\(IP OP) present the internal states of a component; Transitions T present the various operations of component; Color-set S presents the data types of a component. The dynamic characteristics of a software component can be described by the component system defined as follows:

∪

Definition 12. A component system (CS for short) is a 10-tuple, CS=P, T, F, S, AP, AT, AF, IP, OP, M, where:

1) CN = P, T, F, S, AP, AT, AF, IP, OP is a component net, called the base net of the component system; 2) M: P→SMS is a marks set of a component system and it meet the following relationship: a) ∀p∈P: M(p)∈(AP(p))MS; b) M0 is the initial mark of a component system.

5 Conclusion In component-based software engineering, components have become increasingly important. Large complex software systems are composed of many software components. A component is simply a data capsule. Thus information hiding becomes the core construction principle underlying components. In order to descript the software component effectively, first, the definitions of OR-transition Colored Petri Net and component are given. And in OR-transition Colored Petri Net, the transitions can effectively represent the operations of the components. So an approach is presented to descript the software component formally based on OR-transition Colored Petri Net.

References 1. Weide, B.W., Hollingsworth, J.E.: Scalability of reuse technology to larege systems requires local certifiability. In: Latour, L. (ed.) Proceedings of the Fifth Annual Workshop on Software Reuse (October 1992) 2. Talor, R.N., Medvidovic, N., Anderson, K.M., et al.: A component- and message-based architectural style for GUI software. IEEE Transactions on Software Engineering 22(6), 390–406 (1996)

434

Y. Yu et al.

3. Shaw, M., Garlan, D.: Software architecture: Perspectives on an emerging discipline. Prentice Hall, Inc., Simon & Schuster Beijing Office, Tsinghua University Press (1996) 4. Yong, Y., Tong, L., Qing, L., Fei, D., Na, Z.: OR-Transition Colored Petri Net and its Application in Modeling Software System. In: Proceedings of 2009 International Workshop on Knowledge Discovery and Data Mining, Moscow, Russia, January 2009, pp. 15–18 (2009) 5. Wang, Z.j., Fei, Y.k., Lou, Y.q.: The technology and application of software component. Science Press, Beijing (2005) 6. Clements, P.C., Weiderman, N.: Report on the 2nd international workshop on development and evolution of software architectures for Product families. Technique Report, CMU/SEI98-SR-003, Carnegie Mellon University (1998) 7. Li, T.: An Approach to Modelling Software Evolution Processes. Springer, Berlin (2008)

Measurement of Software Coupling Based on Structure ∗ Entropy Yong Yu1,2, Tong Li1,2, Qing Liu1,2, and Qian Yu1,2 2

1 School of Software, Yunnan University, Kunming, China Key Laboratory in Software Engineering of Yunnan Province, Kunming, China [email protected], [email protected], [email protected], [email protected]

Abstract. In the oriented-object software, coupling is the degree of inter dependence between objects. In order to measure the coupling of a software system formally, a novel approach is presented. Firstly, the coupling of Objectoriented software system is analyzed. Second, the definition of the structure entropy is given; Entropy is the most influential concept to arise from statistical mechanics. Entropy measures the disorder in a system. Finally, based on the structure entropy, an approach to measure the coupling of the oriented-object software system is presented. Keywords: Structure Entropy, Object-oriented, Software, Coupling.

1 Introduction Software complexity measures are meant to indicate whether the software has desirable attributes such as understandability, testability, maintainability, and reliability. As such, they may be used to suggest parts of the program that are prone to errors. An important way to reduce complexity is to increase modularization [1]. As part of their structured design method, Constantine and Yourdon [2] suggested that modularity of software system be measured with two properties: cohesion and coupling. Cohesion represents the tight degree of the module in software system. High-cohesion of module is always an aim of software developer. On the other hand, coupling is the degree of inter dependence between pairs of modules; the minimum degree of coupling is obtained by making modules as independent as possible. Ideally, a well designed software system maximizes cohesion and minimizes coupling. Page-Jones gives three ∗

This work has been supported by the National Science Foundation of China under Grant No. 60963007, by the Science Foundation of Yunnan Province, China under Grant No. 2007F008M, the Key Subject Foundation of School of Software of Yunnan University and the Open Foundation of Key Laboratory in Software Engineering of Yunnan Province under Grant No. 2010KS01, the promotion program for youth key teachers of Yunnan university No.21132014, by the Science Foundation of Yunnan Province-The Research of Software Evolution Based on OOPN.

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 435–439, 2011. © Springer-Verlag Berlin Heidelberg 2011

436

Y. Yu et al.

principle reasons why low coupling between modules is desirable [3]. Myers [1] refined the concept of coupling by presenting well-defined, though informal, levels of coupling. However, how to measure the coupling of a module formally? The contribution of this paper is to propose a novel approach to measure the coupling of Object-oriented software system. The approach is based on structure entropy. Entropy is the most influential concept to arise from statistical mechanics. Entropy measures the disorder in a system. Based on structure entropy, the approach can describe the coupling better. The remainder of this paper is organized as follows. Section two analyse the coupling of Object-oriented software system. Section three gives a short introduction to the concept of structure entropy. In Section four, an approach to measure the coupling is presented.

2 The Analyses of Coupling in Object-Oriented Software In object-oriented software system, a software system (S) is composed of a series of related objects, so the coupling of software systems is mainly reflected by the coupling between objects in system. In object-oriented software system, different objects interact via message transmission and the coupling of software systems includes the coupling between the methods in an object and the attributes in an object and the coupling between the methods and the methods in an object. According to the characteristics of object-oriented software system, a software system (S) can be defined as follows: A software system S = (C1, C2, ..., Cn), Ci(i=1,2,…,n) describes an object in the software system S; For each object Ci in software system, Ci = (Ai, Mi), where Ai = {Ai1, Ai2, ..., Aip} is a set of attributes, and it describes ip attributes contained by the object Ci. Mi = {Mi1, Mi2, ..., Miq} is a set of methods, and it describes iq methods contained by the object Ci. There are two kinds of coupling relationships in object-oriented software system: (1) If the method Mik of object Ci refer to the attribute Ajl of object Cj, there is the coupling relationship between methods Mik and attribute Ajl (denoted as Mik→Ajl); (2) If the method Mik of object Ci transfers information to the method Mjl of object Cj, there is the coupling relationship between methods Mik and method Mjl (denoted as Mik→Mjl).

3 The Structure Entropy The development of the concept of entropy of random variables and processes by Claude Shannon provided the beginning of information theory. The entropy is the most influential concept to arise from statistical mechanics. Entropy measures the disorder in a system. The theory of entropy is widely used in various fields. The structure entropy is defined as fellows [4,5,6]:

Measurement of Software Coupling Based on Structure Entropy

437

Definition 1. If a system X has n sub-systems: X1, X2, …, Xn, the sub-system Xi is related to Xj, then gij=1, otherwise gij=0.

；

n

n

let N(i) = ∑ gij linking intensity ρ (i) = N(i)/ ∑ N(j). j=1

j=1

Definition 2. The structure entropy of system X:

，

n

n

H = − ∑ ρ (i)ln ρ(i) where ∑ ρ (i) = 1. i =1

i =1

4 The Coupling Based Structure Entropy For the attribute Aik of the object Ci, if there is a method Mjl of another object Cj s.t Mjl→Aik, methods Mjl is called as the coupling fan-in of the attribute Aik. For the attribute Aik, the number of the coupling fan-in is called as the fan-in degree of attribute Aik, denoted as id (Aik). For the method Mik of the object Ci, if there is a method Mjl of another object Cj s.t Mjl→Mik, methods Mjl is called as the coupling fan-in of the method Mik. For the method Mik, the number of the coupling fan-in is called as the fan-in degree of method Mik, denoted as id (Mik). Definition 3. For the object Ci (i=1,2,…,n), if the number of the attributes is ip, and the coupling fan-in degree of attribute Ak (k=i1, i2, …, ip) is id(Ak), let id ( A k )

ρ (k ) =

ip

∑ id ( A l)

l = i1

Definition 4. The structure entropy of the object based on the attribute is defined as: ip

E Ci -A = − ∑ ρ ( k ) ln ρ ( k ) k = i1 ip

where :

∑ ρ (k ) = 1.

k = i1

Definition 5. For the object Ci (i=1,2,…,n), if the number of the methods is iq, and the coupling fan-in degree of method Ak (k=i1, i2, …, iq) is id(Mk), let

ρ (k ) =

id ( M k ) iq

∑ id ( M l)

l = i1

Definition 6. The structure entropy of the object based on the method is defined as: iq

E Ci -M = − ∑ ρ ( k ) ln ρ ( k ) k = i1

iq

where : ∑ ρ ( k ) = 1 . k = i1

438

Y. Yu et al.

Definition 7. For the object Ci (i=1, 2, …, n) in the software system S, the coupling fan-in of the attribute Aik or method Mil is called as the coupling fan-in of the object Ci, denoted as id(Ci), and ip

iq

k = i1

k = i1

id ( C i ) = ∑ id ( A k ) + ∑ id ( M k ). Definition 8. If there are n objects in the software system S, and the fan-in degree of object Ci (i=1, 2, …, n) is id(Ci), let ρ (C i) =

id (C i) n

∑ id (C l)

l =1

Definition 9. The structure entropy of the software system S is defined as: n

E S = − ∑ ρ ( C i ) ln ρ ( C i ) k =1

Definition 10. If there are n objects (C1, C2, …, Cn)in the software system S, the coupling of the object Ci based on structure entropy and fan-in degree is defined as: H (C i ) = −

id (C i ) ρ (C i ) ln ρ (C i ) × n ln n

Definition 11. The coupling of the software system S based on structure entropy and fan-in degree is defined as: n id (Ci ) ρ (Ci )ln ρ (Ci) n × = ∑ H (Ci), where : ∑ ρ (Ci ) = 1. n ln n i =1 i =1 i =1 n

H ( S ) = −∑

From the definitions, we can conclude that H(S)≥0. If the value of H(S) is greater than 1, there are multiple coupling between objects in the software system S.

5 Conclusion Software complexity measures are meant to indicate whether the software has desirable attributes such as understandability, testability, maintainability, and reliability. An important way to reduce complexity is to increase modularization. Modularity of software system is measured with two properties: cohesion and coupling. Coupling is the degree of inter dependence between pairs of modules. In order to measure the coupling of a module formally, an approach is presented. In this paper, the coupling of Object-oriented software system is analyzed and the definition of the structure entropy is given. Based on the structure entropy, a novel approach to measure the coupling of the oriented-object software system is presented.

Measurement of Software Coupling Based on Structure Entropy

439

References 1. Myers, G.: Reliable Software Through Composite Design. Mason and Lipscomb Publishers, New York (1974) 2. Constantine, Yourdon: Structured Design. Prentice-Hall, Englewood Cliffs (1979) 3. Page-Jones, M.: The Practical Guide to Structured Systems Design. YOURDON Press, New York (1980) 4. Lai, Y., Ji, F.-x., Bi, C.-j.: Evaluation of the Command Control System’s Organizational Structure (CCSOS) Based on Structure Entropy Model. Systems Engineering 19(4), 27–31 (2001) (in Chinese) 5. Yong, Y., Tong, L., Na, Z., Fei, D.: An Approach to Measuring the Component Cohesion Based on Structure Entropy. In: Proceedings of 2008 International Symposium on Intelligent Information Technology Application (IITA 2008), Shanghai, China, December 2008, pp. 697–700 (2008) 6. Yu, Y., Tang, J., Li, W., Li, T.: Approach to measurement of class cohesion based on structure entropy. System Engineering and Electronics 31(3), 702–704 (2009)

A 3D Grid Model for the Corridor Alignment Kun Miao and Liang Li School of Civil Engineering and Architecture, Central South University, Changsha 410075, China [email protected]

Abstract. A corridor of highway or railway generally may be considered as a continuous trajectory in 3D space, and may refer to procedures for locating corridors in three dimensions space. However, most of methods about it used 2D grid for the problem in literature. After summarizing a few of 2D grids for the corridor problem and analyzing existing 3D spatial models, this paper proposes a 3D grid structure with axes and layers so that the corridor finding can be realized in 3D environment. The 3D grid is a natural extension of 2D square grid. Keywords: Path finding; Corridor alignment; 3D model.

1 Introduction A corridor of highway or railway generally may be considered as a continuous trajectory between two points in space through which service is provided. A corridor alignment problem may refer to procedures for locating corridors of highway or railway in three dimensions space. In the design stage, its space position is expressed by vertical and horizontal alignments. The quality of corridor alignment is frequently an important factor in determining the overall success of highway or railway engineering projects. A well designed corridor provides positive results such as reduced project costs, increased levels of service, and improved accessibility of service. A poorly designed corridor may lead to negative environmental impacts, poor level of service, and sometimes life-threatening danger. The corridor alignment problem is to find a 3D route connecting two given points. The result of the finding corridor alignment makes the total costs of the route associated with minimized, and meets some horizontal and vertical constraints. The corridor alignment is generally a piecewise linear trajectory, which is coarse and another design stage is usually employed to detail the alignment further. However, the procedure of finding a corridor has a great influence on the design stage. The research for corridor alignment problem must set up basic grid structures for search methods at first. The paper discusses different grid structures in literatures firstly. Then, 3D model in GIS and GMS (3D Geosciences Modeling System) are introduced, and from the models a new model for 3D corridor alignment is presented in part 3.

2 Search Grid Type In previous studies, the corridor alignment problem is discussed mostly by two dimensions. Models for simultaneously optimizing three-dimensional alignments are rarely R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 440–444, 2011. © Springer-Verlag Berlin Heidelberg 2011

A 3D Grid Model for the Corridor Alignment

441

found. Parker [1] developed a two-stage approach to select a route corridor, subject to gradient constraints. Yuelei [2] employed cellular automata model in design of railway route selection. In addition, a heuristic tabu search algorithm is used for solving the problem [3]. All the models employ a 2D model to realize a space alignment. A square grid with eight directions from each square(except for boundary squares which have less than eight). Turner[4] and Nicholson [5]started developing a model based on the grid where all kinds of relevant costs are evaluated for each square. The best corridor is found as the shortest path upon the grid. This grid has a blunt since the selective angular determination only integer multiplications of 45o, but it makes possible backwards bends. The grid is the most frequently used for its simplicity. Honeycomb Grid: A honeycomb grid (depicted in Fig.1) with twelve directions and the angle between the two adjacent directions is 30 degrees which is less than square grid. This leads to shifts of up to 15 degrees, and has better angular selectivity. However, it increases heavy computational effort to any search algorithm for its larger search space.

Fig. 1. Honeycomb grid

Fig. 2. Hyperbolic grid

Hyperbolic Grid: Trietsch[6] presented a hyperbolic search grid(depicted in Fig.2) where the grids are hyperbolas and the states are defined on them by orthogonal ellipses. The origin and destination are the common foci to all these. It retains the angular selectivity property of square grid, and it helps to form backwards bends. In particular, the procedures will have convergence properties when adding a bound. But it seems to be not convenient to compute its location for the structure. Orthogonal Cutting Lines: An alignment is constructed on orthogonal cutting planes at equally spaced station points. There are many rectangular search grids on each plane, where the selection for the grid point forms an alignment connecting the origin to the destination [6]. It gives us good selective angular determination in the sense that consecutive segments of our piecewise linear trajectory can be tilted relative to each other at small or large angles. But it can not make backwards bends, which is necessary sometimes in mountainous terrain. Jong[7] improved the grid and applied it to a continuous 3D model with which an alignment is constructed by connecting randomly selected intersection points on orthogonal cutting planes. As mentioned above, most grids for corridor alignment problem do not involve a 3D model. Only with a 3D model can we realized the corridor optimization about horizontal and vertical alignments simultaneously. Otherwise, all kinds of corridor optimization are not globe optimization but a local one.

442

K. Miao and L. Li

3 3D Spatial Models A corridor may locate on the earth (fill earthwork or bridge) or under the earth surface (cut earthwork or tunnel). It should be represented as a 3D spatial model. 3D geometric representation models provide geometric descriptions of objects for storage, geometric processing, display, and management of 3D spatial data by computers in 3D GIS and 3D GMS(Geosciences Modeling System).These models are classified as facial models, volumetric models and mixed models[9][10].For facial models, geometric characteristics of objects can be described by surface cells. As for volumetric models, which describe the interior of objects by using solid information, instead of surface information, the model includes voxel, needle, TEN, block, octree,3D Voronoi, tri-prism and so on. Volumetric model based on the 3D spatial partition and description of real 3D entities, emphasizes on representation both of border and interior of 3D objects [8]. Facial model emphasizes particularly on the surface description of 3D space and it is convenient for visualization and data updating, but difficult for spatial analysis. As for the corridor alignment, it is not necessary to have more knowledge about sub-surface in the present. A natural extension of the facial model will meet the need of the corridor alignment. The facial model includes grid model, TIN model, B-Rep (Boundary Representation), series sections, etc. Grid and TIN(triangular irregular network) model are two kinds of importance structures of the digital elevation model (DEM) which is a digital representation of ground surface topography or terrain. A DEM can be represented as raster (a grid of squares) or TIN. It may consist of single raster points arranged in a regular grid pattern, single points in an irregular pattern. It is a digital representation of a continuous variable over a two-dimensional surface by a regular array of z values referenced to a common datum. z = f ( x, y ), x, y ∈ D (1)

where z represents the elevation of the point (x,y) in the field D. In particular, A DEM may be defined as regular one by a gridded matrix of elevation values that represents surface form[11]. The structure of DEM with irregular triangulation is more complicated than that with regular grid in creation, calculation, data organization, while single raster points arranged in regular grid corners may be simply represented with only row and column number. The grid-based model can be extended to a 3D one on which a corridor is based. So the regular grid is chosen for the application.

4 Axis and Layer Structure The square grid is the simplest grid type. In particular, it is not necessary to store coordinates for the grids. The storage space can be saved effectively. So we set up the 3D structure based on the square grid. The study area was partitioned into grid cells such that the corner of each cell is assigned a value equal to the terrain elevation of the point. Let π1: z1=f1(i) denotes the terrain surface. Linking the origin point and destination point with a line, there is an angle θ between the line with the horizontal plane. The plan π2:

A 3D Grid Model for the Corridor Alignment

443

z2=f2(x,y) with the angle θ on the horizontal plane is a slant plane. A new curve surface is constructed from the slant plane π2 and the surface π1: π : z =f 0

0

0

(i ) = ω1f1 (i )

+ ω f (i ) 2 2

(2)

where ω1, ω2 are the weight of the two surface, and ω1+ω2=1. The curve surfaceπ0 is named as tendency surface which not only embodies ground surface for the vertical alignment of the corridor but also meets short and direct demand for the horizontal alignment of the corridor.

Axi s

Ver t ex

Layer

z

y x Fig. 3. Vertexes, axes and layers

A corner point extended up and down from the tendency surface and spaced at equal vertical intervals constructs axis. A point on axis i (i=1,…, N axis ) ( N axis is the total axis numbers) is called vertex. Let dz denote the distance between two contiguous vertexes. The elevation of vertex k on axis i ( k = 1, 2, ..., N layer ) ( N layer is the total layer numbers on one axis) is: zi ,k = f 0 (i ) − N layer ⋅ dz + k ⋅ dz

(3)

A set of points with the same k on different axis is called layer. All the points on the axes and layers are the research space of the path finding algorithm. A 3D alignment can be represented as a sequence of adjacent vertexes originating from a designated start vertex to a designated end vertex. The purpose of the algorithm based on the grid structure is to find continuous contiguous vertexes to meet some constrains in the space to form a 3D corridor alignment.

5 Application This paper presents a kind of grid structure to optimize highways or railways corridor alignment, on which some optimization algorithms may be presented based. For example, we set up a kind of ant colony optimization method based on the grid structure. The basic process to form a corridor using the method is: an ant put on the start vertex selects another vertex in its candidate set according to selection probability, repeating the action until the end vertex. The candidate set of a vertex includes all the vertexes satisfying constraints for the corridor alignment around it. A geographic map to test the

444

K. Miao and L. Li

effectiveness of the approach through an illustrative example is presented with the proposed approach for selecting a 3D highway corridor. The experiment shows the result is reasonable and practicable based on the grid. The method will be introduced in another paper.

References 1. Parker, N.A.: Rural Highway Route Corridor Selection. Transportation Planning and Technology (3), 247–256 (1977) 2. Yuelei, H., Lingkan, Y.: Cellular automata model in design of railway route selection. Journal of Lanzhou Jiaotong University 23(1), 6–9 (2004) 3. Hou, K.J.: A Heuristic Approach For Solving The Corridor Alignment Problem, p. 110. Purdue University (2005) 4. Turner, A.K., Miles, R.D.: A Computer Assisted Method of Regional Route Location. Highway Research Record 348, 1–15 (1971) 5. Nicholson, A.J., Elms, D.G., Williman, A.: Optimal highway route location. Computer-Aided Design 7(4), 255–261 (1975) 6. Trietsch, D.: A family of methods for preliminary highway alignment. Transportation Science 21(1), 17–25 (1987) 7. Jong, J.: Optimizing Highway Alignment with Genetic Algorithms. University of Maryland, College Park (1998) 8. Cheng, P.: A Uniform Framework of 3D Spatial Data Model and Data Mining from the Model. In: Li, X., Wang, S., Dong, Z.Y. (eds.) ADMA 2005. LNCS (LNAI), vol. 3584, pp. 785–791. Springer, Heidelberg (2005) 9. Lixin, W., Wenzhong, S.: GTP-based Integral Real-3D Spatial Model for Engineering Excavation GIS. Geo-Spatial Information Science 7(2), 123–128 (2004) 10. Li-xin, W., Wen-zhong, S., Gold, C.H.: Spatial Modeling Technologies for 3D GIS and 3D GMS. Geography and Geo-Information Science 19(1), 5–11 (2003) 11. Weibel, R., Heller, M.: Digital Terrain Modelling. In: Maguire, D.J., Goodchild, M.F., Rhind, D.W., Maguire, D.J., Goodchild, M.F., Rhind, D.W. (eds.) Geographical Information Systems: Principles and Applications, pp. 269–297. Longman, London (1991)

The Consumers’ Decisions with Different Delay Cost in Online Dual Channels Shengli Chen School of Management, Xian University of Finance and Economics Xian, China [email protected]

Abstract. By constructing the linear delay cost and exponential delay cost function, we formulate the consumer decision models with different delay cost based on the threshold strategies in dual-mechanism, and prove that there exists a unique symmetric Nash equilibrium in which the high-valuation consumers use a threshold policy to choose between the two selling channels. Then we extend our model under general delay cost function and find that consumers with higher valuation will choose threshold strategies to make decision once arriving at websites only if the delay cost function of them continuous strict increases with the auction remaining time. Keywords: consumers’ decisions; auction; posted price; delay cost; threshold strategy.

1 Introduction With the rise of Internet, online auctions are growing in popularity and are fundamentally changing the way many goods and services are traded. Nowadays, in the businessto-consumer market, many firms are selling the same or almost identical products online using auctions and fixed prices simultaneously. For example, airline and cruise companies sell tickets through their own websites at posted prices, but also through auctions run by Priceline.com; IBM and Sun Microsystems sell their products at their own website, but also offer some selected new and refurbished products through eBay.com auctions. In some literature, some researchers pay attention to the problem of jointly managing auction and list price channels. Wang studied the efficiency of posted-price selling and auctions [1]. Kultti studied the performance of posted price and auction [2]. Within the B2C framework, Vakrat and Seidmann compares prices paid through on-line auctions and catalogs for the same product [3]. In the infinite horizon model of van Ryzin and Vulcano, the seller operates simultaneously auctions and posted prices, and replenishes her stock in every period. However, the streams of consumers for both channels are independent, and the seller decides how many units to allocate to each of the channels [4]. Etzion et al. studied the simultaneous use of posted prices and auctions in a different model [5]. Caldentey and Vulcano considered two different variants of the problem when customers have the channels of listed price and auction [6]. Sun study Web stores selling a product at a posted price and simultaneously running auctions for R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 445–453, 2011. © Springer-Verlag Berlin Heidelberg 2011

446

S. Chen

the identical product. In this paper, they study a dual mechanism, where an online retailer combines the two conventional mechanisms (posted price and auction) for multiple units of a product [7]. In this paper, we mainly study how consumers behave and decide upon their purchasing decision when they arrive at the website in dual channels. We analyze simultaneous dual online auctions and list price channels in a B2C framework, with consumers arriving according to a Poisson process, and deciding which channel to join. We characterize the delay cost of high-valuation consumers by linear function, exponential function and general function, respectively.And we formulate the consumer decision models with different delay cost based on the threshold strategies in dualmechanism, and prove that there exists a unique symmetric Nash equilibrium in which the high-valuation consumers use a threshold policy to choose between the two selling channels. Then we find a general conclusion: consumers with higher valuation will choose threshold strategies to make decision once arriving at websites only if the delay cost function of them continuous strict increases with the auction remaining time regardless of different delay cost.

2 The Consumer’s Decisions with Different Delay Cost 2.1 The Consumer’s Decisions with Linear Delay Cost We model an online seller who offers identical items using two selling mechanisms, posted price and auctions, simultaneously. The seller’s objective is to maximize his revenue per unit time. The seller chooses the auction duration T , the quantity to auction q , and the posted price p . Without loss of generality, we assume that the marginal cost of each unit is zero. The seller’s publicly declared reserve price is r .We also assume that the seller can satisfy any demand. Consumers visit the website according to a Poisson process with rate, and each consumer is interested in purchasing one unit of the good. Consumers have independent private values for the good. We assume that each consumer’s valuation, vi , is independently drawn from a probability distribution with cumulative density function F (⋅) with support set [ vl , vh ] ,where

r ≤ vl . We assume that consumers can observe both channels on arrival, with no additional costs. Hence, consumers are fully informed. We model the auctions using the sealed-bid ( q + 1 ) -price format with risk-neutral bidders having unit demand and independent private values for the good. In a sealedbid ( q + 1 ) -price auction, the dominant strategy for each bidder is to bid his true valuation of the item. By doing so, the bidder is setting an upper bound on the price that he is willing to pay: he will accept any price below his reservation price and none above. In our setting, the posted price first splits the consumers into two parts. One part is the consumers with low valuations (i.e., valuations less than the posted price), and the other is those with high valuations (i.e., valuations greater than or equal to the posted price). All low-valuation consumers become bidders in the auctions. A fraction β of the high-valuation consumers also become bidders, while a fraction 1 − β purchase at

The Consumers’ Decisions with Different Delay Cost in Online Dual Channels

447

the posted price. The probability of participation, β , depends on the design variables ( q , T ,and p ), and much of the analysis in this section is devoted to their determination. Some bidders win the auction and some lose. The high-valuation losing bidders will purchase at the posted price at the auction’s conclusion. We characterize the consumer i by his valuation vi , and the time remaining in the auction when he arrives, t c . Low-valuation consumers, those with vi < p , cannot buy the item for its posted price because the value they get from doing so is negative, so they choose between bidding and staying out of the market. High-valuation consumers, those with vi ≥ p , who prefer to receive the item earlier rather than later, may choose not to bid, while choose to purchase at the posted price, if the remaining time of the auction is significantly long. However, low-valuation consumers must choose to bid, because they have no other option for obtaining the item. If we define U −A (vi , t c ) as the maximum expected value from participating in the auction for a consumer of type (vi , t c ) , when vi < p , the optimization problem faced by these consumers is U A (vi , t c )

Max{Pr( ZLQ b)vi

b [0, vi ]

E[auction_payment b]}

(1)

Pr( ZLQ b) is the probability that the consumer wins the item in the auction by bid-

ding b , and E[auction_payment b] is the expected auction payment by a bidder who bids b . When the consumers who participate in the auction arrive at the web site, if the time remaining in the auction is enough long, they do not obtain the item instantly; their utility from buying the item for the bidding is negative. We denote the negative utility as delay cost. And they think that the delay cost is an increasing function of the time remaining until the end of the auction. So they assume that this delay cost is linear in the time remaining until the auction ends. If Dc (t c ) denotes the delay cost, the linear delay cost is l c ⎪⎧ w t low-value consumer Dc (t c ) = ⎨ h c (2) ⎪⎩ w t high-value consumer

， ，

Where, wl denotes the delay cost per unit time of low-valuation consumers, wh denotes the delay cost per unit time of high-valuation consumers, and wh > wl .The optimization problem faced by these consumers with vi < p is Max{U −A (vi , t c ) − wl t c , 0}

(3)

In general, Low-valuation consumers prefer to receive the item earlier rather than later. So, to simplify the problem, we therefore assume that the delay cost per unit time is wl = 0 for consumers with vi < p . The decision problem faced by these consumers with vi < p is Max{U −A (vi , t c ), 0}

(4)

448

S. Chen

High-valuation consumers, those with vi ≥ p , would buy the item for its posted price if auctions were not offered. High-valuation consumers choose between buying the item for its posted price and participating in the auction. It is never optimal for these consumers to do nothing, because their utility from buying the item for the posted price is nonnegative. We assume that when high-valuation consumers purchase the item for its posted price, they obtain the item instantly. When they choose to bid, they are choosing to experience a delay in obtaining and using the item, because they must wait until the end of the auction. Hence, when choosing to bid, these consumers incur a delay cost that is an increasing function of the time remaining until the end of the auction. If we define U +A (vi , t c ) as the expected net revenue from participating in the auction for a consumer with valuation vi ≥ p , then U +A (vi , t c ) is as follows. U A (vi , t c )

Max {Pr(ZLQ b)vi

b [0, vi ]

E[auction_payment b] Pr(ORVH b˄ ) viˉp˅ wh t c }

(5)

If we define U +B (vi , t c ) as the value a consumer of type vi derives from purchasing the item for the fixed price, U +B (vi , t c ) is as follows U +B (vi , t c ) = vi − p

(6)

Because the consumer evaluates the expected payoff from bidding, using an optimal bidding strategy, and compares it with the payoff from purchasing the item for the posted price, a high-valuation consumer arriving with t c time units remaining in the auction solves the following optimization problem: Max U +j (vi , t c )

（ ）

j∈ A, B

(7)

In the sealed ( q + 1 )-price auction, it is a dominant strategy for bidders with valuations below the posted price to bid their true values, and it is dominant for bidders above the posted price to bid the posted price. In other words, if a buyer with valuation v decides to enter the auction then he will bid b(v) , where ⎧v v < p b(vi ) = ⎨ i i (8) ⎩ p vi ≥ p For consumers with V ≥ p , they may choose between buying the item for its posted price and participating in the auction. A high-valuation consumer chooses to participate in the auction if and only if his net expected revenue from participating in the auction exceeds the net expected revenue from buying the item for its posted price. If we define ΔU + (vi , t c ) as the excess of expected revenue from participating in the auction over expected revenue from buying the item for its posted price, ΔU + (vi , t c ) may be denoted as follows ΔU + (vi , t c ) = U +A (vi , t c ) − U +B (vi , t c )

(9)

The Consumers’ Decisions with Different Delay Cost in Online Dual Channels

449

According to Equation (5) and Equation (6), we may get U (vi , t c )

Pr( ZLQ p )vi

E[auction_payment p ]

Pr(ORVH p˄ ) viˉp˅ (vi

p ) wh t c

(10)

(t c ) wh t c

Where Δπ (t c ) denotes the expected discount a high-valuation consumer gets over the posted price if he does participate in an auction. So, for a high-valuation consumer i with type (vi , t c ) , when the expected discount Δπ (t c ) exceeds his delay cost Dc (t c ) , namely the following inequality (11) holds, the net expected revenue from participating in the auction is greater than he net expected revenue from buying the item for its posted price ,so, he chooses to participate in the auction. Δπ (t c ) ≥ wh t c

(11)

According to inequality (11), as long as we calculate the expected discount of a highvaluation consumer, we can make a decision his behaviors (choose to bid or choose to buy item for its posted price). Therefore, we give the following definition. Definition 1. In dual mechanism, the high-valuation consumer with type (vi , t c ) use

the following threshold policy about time remaining t ∗ to choose between buying the item for its posted price and bidding in the auction: (1)When t ∗ < T , if t c ≤ t ∗ , the high-valuation consumer chooses to bid in an auction; if t c > t ∗ , the high-valuation consumer chooses to buy the item for its posted price. (2) When t ∗ = T , the highvaluation consumer chooses to buy the item for its posted price. Although the high-valuation consumer may choose between to buy the item for its posted price and to bid in the auction according to the above threshold policy about time remaining, they must calculate the threshold time before their choices. So, the following lemma 1 is given. Theorem 1. In a dual channel with a sealed-bid q + 1 -price auction ,if bidders with linear delay cost follow the weakly dominant bidding strategy of equation (8), threshold strategies of all high-valuation consumers with threshold time t have a unique symmetric Nash equilibrium, t is given by the solution of the following fixed point equation

(1) if Δπ h (T ) ≤ whT , Δπ h ( t ) = wh t (2) if Δπ h (T ) > whT , t = T Based on Theorem 1, we conclude that under fairly general conditions high-valuation consumers use a threshold policy to choose between to buy the item for its posted price and to bid in the auction. If the remaining time of the auction observed by a high-valuation consumer upon his arrival exceeds the threshold, the consumer chooses to purchase the item for its posted price. If the remaining time of the auction is less than the threshold, t , the consumer chooses to participate in the auction.

450

S. Chen

2.2 The Consumer’s Decisions with Exponential Delay Cost

Although the above Theorem 1 make an explanation of the decision-making mechanism of all high-valuation consumers, Theorem 1 require the delay cost of all high-valuation consumers must satisfy the equation (2).However, in practice, the relationship between delay cost of the bidder and the time remaining may be not necessarily linear there may exist other forms of delay cost function, such as quadratic, exponential, etc. In this section, we will study their decisions of all high-valuation consumers with exponential delay cost. In order to analyze all high-valuation consumers’ decisions, we define exponential delay cost as follows:

，

，

⎧⎪ 0 Dc′ (t c ) = ⎨ wh t c ⎪⎩e - 1

low-value consumer

， high-value consumer

(12)

In the above equation (12), when the time remaining, t c , is equal to zero or relatively small, the delay cost of high-valuation consumer is equal to or approximately equal to the delay cost denoted by equation (2). When the time remaining, t c , is very long, the high-valuation consumer’s delay cost denoted by equation (12) is greater than the linear delay cost denoted by equation (2). In particular, when the time remaining, t c , is very long, the deviation between the linear delay cost and the exponential delay cost is very big. Therefore, the delay cost function denoted by equation (12) is used to describe the time-sensitive customers’ delay cost. When high-valuation customers’ delay cost is denoted by equation (12), will their decision behaviors be changed? The following Theorem 2 shown that all high-valuation consumers with exponential delay cost denoted by equation (12) similarly choose between bidding in auction and buying item from the posted price according to the time remaining until the end of the auction. Theorem 2. In a dual channel with a sealed-bid q + 1 -price auction ,if bidders with exponential delay cost follow the weakly dominant bidding strategy of equation (8), threshold strategies of all high-valuation consumers with threshold time tˆ have a unique symmetric Nash equilibrium, tˆ is given by the solution of the following fixed point equation

(1) if Δπ h (T ) ≤ e w T − 1 , Δπ h (tˆ) = e w t − 1 h

hˆ

(2) if Δπ h (T ) > e w T − 1 , tˆ = T h

Although all high-valuation consumers with exponential and lineal delay cost have the same decision mechanism when they arrive at the web site, they have the different threshold strategies under the different delay cost, their threshold time is different. In the following numerical experiments, we explain further the different threshold time between lineal delay cost and exponential delay cost. Fig.1 (a) depicts the two different threshold time under linear delay cost and exponential delay cost, when we assume that the arrival rate λ = 1 , the delay cost per unit time of high-valuation consumers wh = 0.025 , auction length T = 100 , the posted

The Consumers’ Decisions with Different Delay Cost in Online Dual Channels

451

Δπ h ( t ) Δπ h (t )

Dc′

D′c

Dc

Dc tˆ

t

tˆ

t

t

(b)

(a)

Fig. 1. The different threshold time under the different delay cost

price p = 60 , auctioned quantity q = 30 . In Fig. 1 (a), the threshold time of highvaluation consumer with exponential delay cost function is longer than that under lineal delay cost function. In Fig. 1 (b), when the delay cost function is lineal, the expected discount of a high-valuation consumer is greater than his delay cost. Therefore; the equation t = T is satisfied easily according to Theorem 2. Obviously, t = T > tˆ is hold easily In Figure 1 (b). From the above analysis, it is shown that the threshold time of high-valuation consumer in the threshold strategy is shorter as their delay cost is lower. This means that the changes in delay cost may induce the changes in customer decision behavior, and further have an impact on the seller’s income, when high-valuation consumers choose between bidding in auction and buy item for the posted price according to threshold strategy. 2.3 The Consumer’s Decisions with General Delay Cost

For the high-valuation consumers who arrive at website, there exist other form delay cost except for the exponential delay cost denoted by above equation (2).We define the following general form delay cost function

，

⎧⎪ 0 Dc′′ (t c ) = ⎨ c ⎪⎩ f D( t )

low-value consumer

， high-value consumer

(13)

Where f D( t c ) is continuous strictly increasing function with respect to t c . Obviously, when f D (t c ) = e w t − 1 , the equation (2) is equivalent to the equation h c

(12).When f D (t c ) = wh t c , the equation (2) ( wl = 0 ) is equivalent to the equation (13). For the general form delay cost defined by the above equation (13), it is easy draw a conclusion that is similar to the theorem 3 Theorem 3. In a dual channel with a sealed-bid q + 1 -price auction ,if bidders with exponential delay cost follow the weakly dominant bidding strategy of equation (8), ) threshold strategies of all high-valuation consumers with threshold time t have a

452

S. Chen

)

unique symmetric Nash equilibrium, t is given by the solution of the following fixed point equation ) ) (1) if Δπ h (T ) ≤ f D (T ) , then Δπ h (t ) = f D (t ) ) (2) if Δπ h (T ) > f D (T ) , then t = T From the above Theorem 3, we can see that when the delay cost of all high-valuation consumers with the type (vi , t c ) is denoted by equation (11), they similarly choose between to bid in auction and to buy item from the posted price according to the time remaining until the end of the auction.

3 Conclusions In this paper, we characterize the delay cost of high-valuation consumers by linear function, exponential function and general function, respectively. Then we formulate the consumer decision model with three different delay cost based on the threshold strategies in dual-mechanism, and prove that there exists a unique symmetric Nash equilibrium in which the high-valuation consumers use a threshold policy to choose between the two selling channels. We find that consumers with higher valuation will choose threshold strategies to make decision once arriving at websites only if the delay cost function of them continuous strict increases with the auction remaining time regardless of different delay cost. Our analysis relies on the independent private value model proposed by Riley and Samuelson (1981), many auctions have bidders that follow common value or affiliated value models. These bidders' valuations are determined, at least in part, by an unobservable but objective value for the item. In addition, it is assumed that the seller is risk neutral. However, in fact, some sellers are not risk neutral. When bidders are risk averse, the four types of auctions that follow the rules of the family of auctions in Riley and Samuelson (1981) will not generate identical expected revenues and optimal reserves will change. Therefore, these need to be considered in future research.

Acknowledgment Financial supports from a project with special fund for the construction of key disciplines in Shanxi Province, Shanxi Nature Science Foundation under Grant No 2010JQ9006 and Shanxi Department of Education Foundation under Grant No 2010JK552 .The helpful comments from anonymous reviewers are also gratefully acknowledged.

References 1. Wang, W.: Auctions versus posted-price selling. American Economic Review, 838–851 (1993) 2. Kultti, K.: Equivalence of auctions and posted prices. Games and Economic Behavior, 106– 113 (1999)

The Consumers’ Decisions with Different Delay Cost in Online Dual Channels

453

3. Pinker, E., Seidmann, A., Vakrat, Y.: Managing online auctions: Current business and research issues. Management Science, 1457–1484 (2003) 4. van Ryzin, G.J., Vulcano, G.: Optimal auctioning and ordering in an infinite horizon inventory-pricing system. Forthcoming in Operations Research 52(3), 195–197 (2004) 5. Etzion, H., Pinker, E., Seidmann, A.: Analyzing the simultaneous use of auctions and posted prices for on-line selling. Working paper CIS-03-01, William E.Simon Graduate School of Business Administration, University of Rochester (2003) 6. Caldentey, R., Vulcano, G.: Online auction and list price revenue management. Working Paper, Stern School of Business, New York University 7. Sun, D.: Dual mechanism for an online retailer. European Journal of Operational Research, 1–19 (2006)

Fuzzy Control for the Swing-Up of the Inverted Pendulum System Yu Wu and Peiyi Zhu∗ School of Electrical and Automation Engineering, Changshu Institute of Technology, Changshu 215500, China [email protected]

Abstract. The nonlinear inverted-pendulum system is an unstable and nonminimum phase system. It is often used to be the controlled target to test the qualities of the controllers like PID, optimal LQR, Neural network, adaptive, and fuzzy logic controller, etc. This paper will describe a new fuzzy controller for an inverted pendulum system. In this case, a fuzzy controller followed with a state space controller was implemented for control. It is achieved to design a control condition for the pendulum to swing up in one direction only because that the movement of throwing a bowling ball can only from one side to the unstable equilibrium point. Simulation and experimental results show that the fuzzy control can swing up the single inverted pendulum in short time with well stability and strong robustness. Keywords: inverted pendulum; swing up; fuzzy controller; LQR.

1 Introduction1 The inverted pendulum system is a standard problem in the area of control systems [1, 2]. Inverted pendulum is the organic combination of multiple technologies in different fields such as robotics, control theory and computer control. The system itself is absolutely instable, high-order, multivariable, strong-coupling and non-linear, so it can be studied as a typical control object. Nowadays fuzzy control theory is a kind of intelligent control which is the emphasis and difficulty of the control theory [3-5]. The merit is that it is independent with the accurate math model of the system and robustness; it has intelligence, self-learning ability, the feature of computer control and friendly user interface. In this page, compare the result of the fuzzy control of the linear single inverted pendulum with PID and LQR control methods, and confirm the superiority of the fuzzy control on the field of non-linear and strong coupling system like inverted pendulum.

2 Modeling A schematic of the inverted pendulum is shown in Figure 1[6]. ∗

Corresponding author.

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 454–460, 2011. © Springer-Verlag Berlin Heidelberg 2011

2L

Fuzzy Control for the Swing-Up of the Inverted Pendulum System

ϕ

455

m

M

x Fig. 1. Inverted Pendulum Setup

A vehicle equipped with a motor provides horizontal motion of the vehicle while vehicle position, x, and joint angle, φ, measurements are taken via a quartered encoder, the pendulum to the center of the length of the rotation axis, l. By applying the law of dynamics on the inverted pendulum system, the equations of motion are:

⎧ Mx&& = F − bx& − N ⎪ ⎨ d2 N = m ( x + l sin θ ) ⎪ dt 2 ⎩

(1)

It follows that

( M + m) && x + bx& + mlθ&& cos θ − mlθ& 2 sin θ = F

(2)

To the pendulum boom and vertical direction a force to have: P − mg = −mlθ&& sin θ − mlθ&&2 cos θ

(3)

Torque-equilibrium equation is − Pl sin θ − Nl cos θ = Iθ&&

(4)

( I + ml 2 )θ&& + mgl sin θ = −mlx&& cos θ

(5)

It follows that

For linearization processed by equations (2) & (5) ⎧⎪( I + ml 2 )ϕ − mgl&&ϕ = mlx&& ⎨ ⎪⎩( M + m) x + bx&& − ml&ϕ = u Let the state vector be defined as:

＝

x ( p, p& , θ , θ&)

(6)

(7)

Finally, we linearism the system about the unstable equilibrium (0, 0, 0, 0)T . Note that θ = 0 corresponds to the pendulum being in the upright position. The linearization of the inverted pendulum system around the upright position is:

456

Y. Wu and P. Zhu

⎡0 ⎢ & x ⎡ ⎤ ⎢ ⎢ && ⎥ ⎢0 ⎢x ⎥ = ⎢ ⎢ϕ& ⎥ ⎢0 ⎢ ⎥ ⎢ ⎣ϕ&&⎦ 0 ⎢ ⎢⎣

1

0

−( I + ml )b I ( M + m) + Mml 2 0 −mlb I ( M + m) + Mml 2 2

2

2

m gl I ( M + m) + Mml 2 0 mgl ( M + m) I ( M + m) + Mml 2

0⎤ 0 ⎡ ⎤ ⎥ x ⎢ ⎥ 2 ⎡ ⎤ I + ml ⎥ 0⎥ ⎢ ⎥ ⎢ ⎥ x& ⎢ I ( M + m) + Mml 2 ⎥ ⎢ ⎥ +⎢ ⎥ ⎥u 1 ⎥ ⎢ϕ ⎥ ⎢ 0 ⎥ ⎢ ⎥ ⎥ ⎣ϕ& ⎦ ⎢ ⎥ ml 0⎥ ⎢ 2 ⎥ ⎥⎦ ⎢⎣ I ( M + m) + Mml ⎥⎦

⎡x ⎤ ⎢ ⎥ ⎡ x ⎤ ⎡1 0 0 0 ⎤ ⎢ x& ⎥ ⎡ 0⎤ y=⎢ ⎥=⎢ + ⎢ ⎥u ⎥ ⎣ϕ ⎦ ⎣0 0 1 0 ⎦ ⎢ϕ ⎥ ⎣ 0⎦ ⎢ ⎥ ⎣ϕ& ⎦

(8)

3 Single Inverted Pendulum Fuzzy Controller 3.1 Design of Membership Function

In fuzzy controller design and the practical application of fuzzy control theory, the fuzzy characteristics of the object being studied to determine the membership function is a very important issue. Determination of membership function should reflect the specific characteristics of an objective fuzziness [7, 8]. 3.2 Design of Fuzzy Control Rules

Control rules are the core of fuzzy control, are generally given by experts. In this study, because the control object is a single inverted pendulum, taking into account the volume control program and operating efficiency, by many experiments try to finally choose a 3 * 3 rule. Control rules are listed in Table 1: Table 1. Fuzzy control rules E

Fuzzy controller input 1

EC

N

O

P

Fuzzy

N

Mf1

Mf3

Mf6

controller

O

Mf2

Mf5

Mf8

input 2

P

Mf4

Mf7

Mf9

3.3 Selection of Fuzzy Control Structure

The classical PID control can only be achieved on a single input control, it plans to integrate the state space and fuzzy control method or point of view, fuzzy dual-loop control of displacement are two ways to achieve a single-stage inverted pendulum. According to the actual control needs and experience, decided to fuzzy control

Fuzzy Control for the Swing-Up of the Inverted Pendulum System

457

structure for multi-input single-output (MISO). As the cascade control of their own complexity, real-time adjustments to the model have a certain degree of difficulty, achieve inverted pendulum in the case, the outer loop control of position effect is not very good, which makes the control and LQR control Xiang there is no advantage compared. Therefore, this experiment chose the combination of state space and fuzzy control method, the system block diagram shown in Figure 2.

Fig. 2. Inverted pendulum fuzzy control system

Where ke, kec and ku are quantization and scaling factor, K1T and K2T is system of feedback gain matrix K transposed matrix. Because single inverted pendulum has four state variables Angle, displacement error and error rate, and the design of fuzzy controller with only two input variables, and so on four state variables must be processed. Single inverted pendulum Angle, displacement volume K1T after two errors after an error matrix E. The same error rate is Angle, the displacement of the Ec error amount after K2T matrix derivation. 3.4 Single Inverted Pendulum Fuzzy Control Model and Real-Time Simulation

This experiment using Matlab software, therefore, the control model of this experiment is through Matlab SIMULINK. Because of this experiment to connect to high grade single inverted pendulum in real-time control, so need to join solid high realtime control module toolbox, eventually model as shown in Figure 3. One set of automatic add up the module, the more convenient, can not use hand every time, and Initi al ize GT400-SV

Vel In1

GT400-SV Initial ization

Acc

Swi ng-up Controller

Vel1

0.0

In1Out1

-K-

Pos

Acc1

Pos Ref. du/dt

Subsystem

Gain2

Sine Wave

1

Vel2

Gain

Acc2

Angle

Deri vative

Si gn 2.8

pi

-pi ~pi In1Out1

Angle Ref. du/dt

Subsystem1

-K-

Fuzzy Logic Control ler

Real Control1

Gai n4

Gain3

Deri vative1

Fig. 3. Single inverted pendulum fuzzy control model

Scope1

458

Y. Wu and P. Zhu

improve the safety of the experiment. The quantitative and scaling factor model is convenient for the design and debug model online adjustment fuzzy controller can rise, input and output of actual control effect. Swing-up module is shown in Figure 4.

1 Vel 1

-1

em

pi

Swi ng 2

In1

Acc

Gai n

Swing-up

Fig. 4. Swing-up module structures

Results were gathered from the implementation of both swing-up control methods. Data were collected from experimental runs where each control scheme swings up the pendulum from an initially downwards position to an upright position and balances the pendulum around the unstable equilibrium point. If there is excessive cause handstand down, and will again do pole into be put through the adjustment model, and once again put into stable state. A plot of the controller output during an experimental run for the single inverted pendulum fuzzy controller is shown in Figure 5 The following diagram oscilloscope waveform is the car in the above position, the following waveform is the angle of the pendulum. In the car into position to set 0.0m from 0.1m, the single inverted pendulum transition shown in Figure 6.

（

Fig. 5. Transition process of car swing-up

Fig. 6. Car position from the transition process 0.1m to 0.0m

To analyze the single inverted pendulum fuzzy control features, and classical control theory and modern control theory, PID control in the LQR control results were compared. Experimental curves obtained as follows. Inverted pendulum PID control curve shown in Figure 7 and Inverted pendulum LQR control curve shown in Figure 8:

Fuzzy Control for the Swing-Up of the Inverted Pendulum System

Fig. 7. PID Control Curve

459

Fig. 8. LQR control curve

Through comparison of these two control methods, fuzzy control is not difficult to find and PID, LQR control the difference between. PID control can only control as the single input single output system, so this inverted pendulum is used to control multi-input system will be a loss, ensure the inverted pendulum can not control the horizontal displacement, so when the pendulum moving to the limit switches Department the role of the out of control. LQR can control multiple input, multiple output system control, in the inverted pendulum control to be better than the PID, LQR control but has some limitations, in the swing phase and was disturbed, the level of displacement if the pendulum had led the General Assembly the role of the system out of control, and the system into a stable state, the rate is not quick enough.

4 Conclusion The analysis of experimental results can be seen, through the linear quadratic state feedback and fuzzy control of combined control methods to control, the LQR control with some good features, but also to participate as fuzzy control; it has more good robustness, and can more quickly into a stable state. But this test is also inadequate, that is, steady-state system, more frequent small shocks, mainly due to fuzzy control can not be achieved without static error control and subtle distortions result output control volume control, and as to swing to control the process shorter, choose the larger scale factor, therefore increasing the system's small shocks, it will be such a phenomenon. If fuzzy control rules more detailed classification, expanding the scope of the domain, you can reduce the shock, but also make the program a huge change, to control the Host Computer have higher requirements.

References 1. Butikov, E.I.: On the dynamic stabilization of an inverted pendulum. Am. J. Phys. 69(6), 1– 14 (2001) 2. Tao, C.W., et al.: Design of a Fuzzy Controller With Fuzzy Swing-Up and Parallel Distributed Pole Assignment Schemes for an Inverted Pendulum and Cart System. Control Systems Technology 16(6), 1277–1288 (2008)

460

Y. Wu and P. Zhu

3. Yorhida, K.: Swing-up control of an inverted pendulum by energy- based methods. In: Proceedings of the American Control Conference, pp. 4045–4047 (1999) 4. Anderson, C.W.: Learning to control an inverted pendulum using neural networks. Control Systems Magazine 9(3), 31–37 (1989) 5. Muskinja, N., Tovornik, B.: Swinging up and stabilization of a real inverted pendulum. Industrial Electronics 53(2), 631–639 (2006) 6. Huang, S.-J., Huang, C.-L.: Control of an inverted pendulum using grey prediction model. IEEE Trans. Ind. Appl. 36, 452–458 (2000) 7. Ma, X.-J., Sun, Z.-Q., He, Y.-Y.: Analysis and design of fuzzy controller and fuzzy observer. IEEE Transactions on Fuzzy Systems 6(1), 41–51 (1998) 8. Kovacic, Z., Bogdan, S.: Fuzzy Controller Design Theory and Applications. Taylor & Francis Group, LLC (2006)

A Novel OD Estimation Method Based on Automatic Vehicle Identification Data Jian Sun and Yu Feng Department of Traffic Engineering, Tongji University, No 4800, Cao’an road, Shanghai, China 201804 [email protected], [email protected]

Abstract. With the development and application of Automatic Vehicle Identification (AVI) technologies, a novel high resolution OD estimation method was proposed based on AVI detector information. 4 detected categories (Ox+Dy、Ox/Dy+Path(s)、Ox/Dy、Path(s)) were divided at the first step. Then the initial OD matrix was updated using the Ox+Dy sample information considering the AVI detector errors. Referenced by particle filter, the link-path relationship data were revised using the last 3 categories information based on Bayesian inference and the possible trajectory and OD were determined using Monte Carlo random process at last. Finally, according to the current application of video detector in Shanghai, the North-South expressway was selected as the testbed which including 17 OD pairs and 9 AVI detectors. The results show that the calculated average relative error is 12.09% under the constraints that the simulation error is under 15% and the detector error is about 10%. It also shows that this method is highly efficient and can fully using the partial vehicle trajectory which can be satisfied with the dynamic traffic management application in reality. Keywords: Origin-Destination Estimation; Automatic Vehicle Identification; Vehicle Trajectory; Bayesian Inference.

1 Introduction The Origin-Destination (OD) matrix is the core information for transportation system planning, design and operation. Traditionally, the OD matrix is gathered by the large-scale sample surveys of vehicle trips. In order to avoiding the problems such as labor-intensive, high-cost and time-consuming of sample survey, the theory of OD matrix estimation based on detected traffic information had became a hot research topics. Since 1978, Van Zuylen and Willumsen firstly used the theory of maximum entropy to estimate the OD matrix based on the traffic volume, the research of OD estimation had more than 30 years history. The methods include least squares state-space models information theory and so on [1][2][3]. On the other hand, the technology of traffic information collection has showed the trend of from the “fixed-point” detection to the “full-scale” Automatic Vehicle Identi-

、

、

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 461–470, 2011. © Springer-Verlag Berlin Heidelberg 2011

462

J. Sun and Y. Feng

fication (AVI) detection. The AVI technology includes the Video License Plate Recognition, Vehicle-beacon communication identification technology and so on. The fundamental of these technologies is that AVI detectors can collection the vehicle IDs time of passing the detector and vehicle location. With the application of the AVI technologies, some scholars have made some explorations on OD estimation under AVI environment. The study can be divided into two categories. One is to revise the classical model of OD estimation and add the new information from the AVI detection to improve the accuracy of OD estimation. Dixon [4] extended Kalman filter model and put the traffic volume and the travel time between the AVI detectors as observed variables to estimate the OD matrix. Xuesong Zhou[5] considered the impact of the travel time when use the nonlinear least squares estimation. Another is to use high-resolution path (OD) information to estimate the vehicle OD. Jaimyoung Kwon used the Method of Moments(MOM) to estimate the OD matrix of the highway in California Bay Area based on the toll station data[6]. Dong Jingxin analyzed the reliability of the OD matrix by using the taxi GPS data [7]. However, the above methods also have the following problems and challenges:

（）

1 OD estimation is not only affected by network topology, traffic volume, but also with the route choice (assignment matrix) and the corresponding traffic states. The previous studies were limited by the data and supporting environment so that OD was estimated by assuming some parameters on an example network. It’s practicability still need to be improved. 2 AVI detector not only can get the link volume and travel time, but also can get the partial trajectory of vehicles which is more important on OD estimation through the data matching between the multiple AVI detectors. 3 The AVI layout density and accuracy were certain limited under the field detection environment, investment budget, installation condition and other restrictions. So it’s a big challenge to improve the accuracy of the OD estimation under the limited information.

（） （）

Facing to these problems, this paper proposed a new method of OD estimation based on particle filter which can effectively integrate the traffic volume, travel time and the partial trajectory from the AVI detectors. The AVI samples of particle and the history OD were established based on probability model by using the Bayesian estimation theory. The state of random AVI samples of particle was estimated by Monte Carlo simulation. The state-space of this model was classified and updated based on the different kinds of the AVI data, so that the results of the Bayesian estimation are more realistic approximation of the actual state-space.

2 The OD Estimation Method under AVI Environment 2.1 Analysis of AVI Detection Data At the AVI locations, the vehicle IDs, time of passing the detector vehicle location can be collected. Among a numbers of AVI detectors, the travel time and the partial trajectory information can be calculated through the vehicle ID mapping. Because of

A Novel OD Estimation Method Based on Automatic Vehicle Identification Data

463

AVI facilities can’t cover all section and have detection errors, for any AVI network, the detected data can be divided into four categories. Case 1 is that the vehicle origin and destination can be detected and expressed as Ox+Dy. Case 2 is that one of vehicle origin or destination and partial trajectory can be detected and expressed as Ox/Dy+Path(S). Case 3 is that only vehicle origin or vehicle destination can be detected and expressed as Ox/Dy. Case 4 is that only partial trajectory can be detected and expressed as Path(S). 2.2 OD Estimation Process Based on different kinds of AVI data, OD estimation process can be described as figure1 and this process can be divided into six steps.

：

Step1 According to the average rate of growth to update the initial OD matrix by using the flow which was detected by AVI facilities. Meanwhile, the errors from the AVI facilities were considered. Step2 According to the updated initial OD, the initial path flow database 1 can be obtained by using the initial assignment matrix. Step3 According to case 1 to expansion the initial OD and get updated data (Ox+Dy), the path flow database 1 then can be fixed by the updated data (Ox+Dy) and the path flow database 2 was gotten as shown in figure 2. Firstly, the path flow database 1 was scanned and the data which include case 1 (Ox+Dy) was selected. Secondly, different paths in proportion which depend on the path flows were determined. In order to reduce the weight which the origin-destination had been known in the OD estimation, The new path flow database 2 could be got by using the path flow database 1 minus the partial path flow data which include case 1(Ox+Dy) samples. Step4 As for the case 2 (Ox/Dy+Path(s)) data, the first step is to search the path flow database 2 for the data (Ox/Dy+Path(s)) and these data can be used as prior information for Bayesian estimation. The second step is to get the probability of path choice through estimated and corrected based on Bayesian estimation. The third step is to make sure any vehicles actual origin-destination and trajectory by using the Monte Carlo simulation process. The last step is to accumulate the OD information which got from the previous step as the OD estimation data 1 and update the path flow database to be a new database which depends on the OD estimation data 1. The process had been showed in figure 3. The algorithm of Bayesian estimation and Monte Carlo was described in section 2.3, 2.4 respectively. Step5 As for case 3 and case 4, it has the similar method with case 2. The OD estimation data 2, OD estimation data 3, path flow database 3 can be got through above method. Step6 The final OD matrix can be gotten by accumulating all of the estimation

： ：

：

：

：

data, including updated Ox+Dy, OD estimation data 1, OD estimation data 2 and OD estimation data 3.

˖ 464˖

J. Sun and Y. Feng

Fig. 1. Flow chart of OD estimation based on AVI data

Fig. 2. Flowchart of case1 data analysis

A Novel OD Estimation Method Based on Automatic Vehicle Identification Data

Partial trajectory

Search path flow database Contains(Ox/Dy+Path(s))

（Ox/Dy+Path(s)) ……

’

……

Path i Flow =fi-fi

……

Path i Flow fi

Estimation selection probability of vehicles based on Bayesian estimation

Confirm trajectory and OD of vehicles based on Monte Carlo simulation

Update and save path flow database

Path 1 Flow =f1-f1'

Modify path flow

Path 1 Flow f1

Output relative path and path flow

Save result of OD estimation and count the amount of different path

Path 1 Flow f1' Path i Flow fi

465

’

Fig. 3. Flowchart of Case2 OD estimation

2.3 Bayesian Estimation Bayesian estimation is a probability estimation method by subjective estimating some unknown

state

under

incomplete

information.

Bayesian

formula

is

P ( H / D ) = P ( D / H ) × P ( H ) , where P(H) is called the prior probability of H and stands

for initial probability of H without any training data. P(D|H) is the prior probability of training data D when assuming H is established. This method can be much more realistic by modifying the prior probability. Supposing all prior probability of selected paths of an OD pair is P ( H 1 ), P ( H 2 ), L , P ( H i )

，and get all probability path flow by default traffic assignment

model. Every collected AVI sample can be called a particle. Based on different path flow,

assuming

a

vehicle

prior

probability of

all

probability paths

is

P ( D1 / H 1 ), P ( D2 / H 2 ),L , P ( Di / H i ) . where, P ( D1 / H 1 ) =

D1 D总

(1)

P ( D2 / H 2 ) =

D2 D总

(2)

P ( Di / H i ) =

Di D总

(3)

n

(4)

D总 = ∑ Di i =1

Where: D总 contains all path flow information through the detection section. Di

Stands for volume of path i.

（

）

All kinds of path posterior probability Case2, Case3, Case4 can be gotten based on Bayesian formulas:

466

J. Sun and Y. Feng

P ( H1 / D1 ) = P( D1 / H1 ) × P ( H1 )

(5)

P( H 2 / D2 ) = P( D2 / H 2 ) × P( H 2 )

(6)

P( H i / Di ) = P( Di / H i ) × P( H i )

(7)

P总 = P(H1 / D1 ) + P(H 2 / D2 ) + L + P(H i / Di )

(8)

Where

Supposing under a kind of detection information, the probability of path selection for a vehicle is P1 P2 L Pi . (9) P1 = P( H1 / D1 ) / P总

，，

P2 = P( H 2 / D2 ) / P总

(10)

Pi = P( H i / Di ) / P总

(11)

As the figure 4 shows, assuming the vehicle No.001 has been detected by passing the section H1 and H2, through the default traffic assignment, the trajectories and path flow of section H1 and H2 can be gathered. Naming all paths as No.1(path 1-7), No.2(path 3-9) and No.3(path 2-10). The prior probability of path selection can be got by traffic survey or other technologies, so assuming the vehicle path from the above three are, p( H 1 ) = 0.3, p( H 2 ) = 0.3

and p ( H 3 ) = 0 . 4 . The path flow of No.1, No.2 and No.3 was ex-

tracted from the path flow database. Assuming the volume of path 1 is 20, the volume of path 2 is 30 and the volume of path 3 is 50. According to the path flow, the prior probability of different paths based on the path volume would be got: P( D / H ) = 0.2 , 1

P( D2 / H 2 ) = 0.3 , P ( D3 / H 3 ) = 0.5 .

1

By using the Bayesian estimation to correct the prior prob-

ability, the posterior probability of those paths can be calculated 0.06,0.09,0.2. The last is to deal with the posterior probability in normalization to get the final probabil-

，P2=0.25，P3=0.58.

ity:P1=0.17

Fig. 4. Path estimation of sample network

A Novel OD Estimation Method Based on Automatic Vehicle Identification Data

467

2.4 Process of Monte Carlo Simulation Monte Carlo method can simulate the real physical process and approximate to the real physical results by simulation, so that the AVI sample particle choice of state-space based on the probability can be gotten by Monte Carlo simulation. For all paths which come across a AVI detector, sets:

W1 = P1

(12)

W2 = P1 + P2

(13)

Wi = P1 + P2 + L + Pi

(14)

Wi is the cumulative probability of all path posterior probability. Other symbols are the same as mentioned above. Figure 4 still was taken as an example for the process of Monte Carlo simulation by generating random number. If the random number is less than W1, the path of vehicle 001 is No.1 and corresponding OD is confirmed. If the random number is bigger than W1 and less than W2, then the path of vehicle 001 is No.2 and corresponding OD is confirmed. The other possibilities can be concluded by above method.

3 3.1

Case Study Testing Site

In 2008, Shanghai started to install video license plate recognition equipments in expressway system. The North- South expressway which has 17 entrances/exits and 9 sections with AVI equipments was chosen as the testing site. The expressway system and AVI facilities were shown in figure 5.

L u B a n In te rch a n g e Y an D o n g In te rch a n g e

T ia n M u In ter ch a n g e

A V I d e te cto r

G o n g H e X in In te rch a n g e

Fig. 5. Shanghai North-South expressway model and AVI facilities layout

468

J. Sun and Y. Feng

3.2 “True” OD Acquisitions Because of the difficulty to obtain the true OD, the evaluation of the accuracy of the OD estimation has been a difficulty in a long time. Several studies are based on assumptions of OD matrix to measurement [1]. On the other hand, because of some origins or destinations are located in the ramp of expressway interchange, the full-scale field survey can’t be organized for safe consideration and the true OD can’t be obtained. We use the simulation method to measure the precision of the OD estimation. In accordance with the field layout of the AVI detectors, the “virtual detectors” were set into the VISSIM simulation model to simulate the true AVI parameters and accuracy. The data for OD estimation which come from the virtual detectors can be compared with the “true” OD from VISSIM model. So the problem about the accuracy of OD estimation can be solved [8]. The “true” OD is gotten by using the microscopic traffic simulation “VISSIM” to establish the North-South expressway model based on the manual vehicle license survey in March 2004. The model was calibrated by using the volume data from the loop detectors and the travel time from the AVI detectors. The relative error of volume and the travel time which come from 7 sections and 9 AVI detectors were less than 15% between VISSIM model and field outputs, which can meet the practical application [9]. 3.3 Analyze Result of OD Estimation By using the data which was collected on the virtual detectors in North-South expressway VISSIM model, the OD was estimated based on the above method. The absolute error of the estimation value is shown in figure 6 and the relative error of estimation value is shown in figure 7.

Fig. 6. Absolute error of OD matrix

Fig. 7. Relative error of OD matrix

Most of the absolute errors of estimation values are from 0 to 20, and a little of absolute errors are larger than 20, such as 1-2, 19-2. The main reasons may include followings: 1 Because of some true value of OD pairs are relative large and path flow in traffic assignment is larger than some other path flow, and the probability of path

（）

A Novel OD Estimation Method Based on Automatic Vehicle Identification Data

469

selected which have larger flow may greater than those which have smaller flow. Therefore, the result of estimation will be larger; 2 Because of the random error from the simulation model, the bias would be produced by estimation process. Most of the relative errors are from 0% to 20% and a little of relative error are larger than 40%, such as 15-11, 15-12. The main reason is that true values of some OD pairs are relative small. For example the true value of OD pair 15-12 is 14 and the estimation value is 20, so the relative error is 42.86%. Finally, the relative error of true value and estimation error is 12.09% by calculating the overall average relative error of OD estimation, which means that the application of method can get an objective reality result of OD estimation.

4 Conclusion With the AVI detectors have been gradually used in urban network, this paper proposed a method of similar particle filter under AVI environment and the method was tested through the Shanghai North-South expressway model. The main conclusions of this paper were as followings:

（）

1 The data of different kinds AVI information were classified based on AVI detection features. The probability of path selection would be more approximate reality by Bayesian estimation and the accuracy of OD estimation can be improved by modifying path flow. 2 With the application of this method, a higher accuracy of OD estimation was gotten from the case study of Shanghai North-South expressway. 3 Because of lack of true OD matrix, the OD estimation error and the simulation model error may exist together. With the application of this method, the accuracy in this paper can be described as the overall average relative error is 12.09% under the model error is less than 15% and AVI detection error is about 10%. 4 The similar particle filter provided a very good tool to research OD estimation and path estimation. In the further studies, the precision of OD estimation can be improved and the impact of random parameters can be reduced by the dynamic non-linear programming method based on the link flow. The accuracy and reliability of OD estimation can also be further improved.

（） （） （）

Acknowledgement The authors would like to thank the Natural Science Foundation of China (50948056) for supporting this research.

References [1] Chang, Y.T.: Dynamic OD Matrix Estimation Model of Freeway With Consideration of Travel Time. Journal of Tongji University 37(9), 1184–1188 (2009) [2] Tsekeris, T., Stathopoulos, A.: Real-Time Dynamic Origin-Destination Matrix Adjustment with Simulated and Actual Link Flows in Urban Networks. Transportation Research Record 1857, 117–127 (2003)

470

J. Sun and Y. Feng

[3] Bierlaire, M., Crittin, F.: An efficient Algorithm for Real-Time Estimation and Prediction of Dynamic OD Tables. Operation Research 52, 116–127 (2004) [4] Zhou, X.S., Mahmassani, H.S.: Dynamic Origin-Destination demand estimation using automatic vehicle identification data. IEEE Transactions on Intelligent Transportation Systems 17(1), 105–114 (2006) [5] Dixon, M.P., Rilett, L.R.: Real-time OD estimation using automatic vehicle identification and traffic count data. Journal of Computer Aided Civil Infrastructure Engineering 17(1), 7–21 (2002) [6] Kwon, J., Varaiya, P.: Real-time estimation of origin-destination matrices with partial trajectories from electronic toll collection tag data. Transportation Research Record, 119–126 (2005) [7] Dong, J., Wu, J.: An Algorithm to Estimate OD Matrix With Probe Vehicle and Its Reliability Analysis. Journal of Beijing Jiaotong University 29(3), 73–76 (2005) [8] Sun, J., Yang, X., Liu, H.: Study on Microscopic Traffic Simulation Model Systematic Parameter Calibration. Journal of System Simulation 19(1), 48–51 (2007) [9] Sun, J., Li, K., Wei, J., Su, G.: Dynamic OD Estimation Simulation Optimization Based on Video License Plate Recognition. Journal of Highway and Transportation Research and Development 26(8), 130–135 (2009)

Stress Field in the Rail-End during the Quenching Process Siqiang Xu, Jianyi Kong, Gongfa Li, Jintang Yang, Hegen Xiong, and Guozhang Jiang College of Machinery and Automation, Wuhan University of Science and Technology, 430081 Wuhan, China [email protected]

Abstract. It is obvious that the railway develops towards the trend of high speed and overload. So the rail quality required becomes more and more strict. The quenching is highlighted as the final method to improve rail behaviors. Stress field of U71Mn rail-ends during quenching was simulated by the FEM software. And various kinds of factors that may influence the stress field’s distribution were researched. The result shows that the rail-end stress can be significantly diminished, if the heating, holding time and air-blast pressure during the wind cooling can be accurately chosen. It is significantly valuable for the choice of relevant parameters for the heat treatment of U71Mn heavy rails. Keywords: Rail-end; Quenching Process; Stress Field; FEM.

1 Introduction U71Mn heavy rails are discarded since phenomena, such as spalling and crack, occur on their ends after trains move on them for a period time. Frequently changing rails affects train schedules and diminishes the railway efficiency. Moreover, subtle crack is a potential hazard to safe movement for trains. The case that spalling and crack appear on rails results mainly from that the thermal stress and structural stress (together called the residual stress) generated during the rail-end’s quenching process were not entirely eliminated and then superimposition of rails’ residual stresses and impact stresses, which continuous hitting of hubs against rails introduces, exceeds the rail strength. In this paper, the stress distribution during the rail-end quenching and its causes were investigated. In addition, the numerical simulation was fulfilled, in order to improve the quenching process to diminish the internal stress during the quenching.

2 Causes of Thermal Stresses during Rail-End Quenching In practical manufacturing, the quenching process was that rail-ends were cooled by means of wind cooling in conjunction with air cooling after heating and holding in order to enhance production efficiency, as shown in Fig.1. R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 471–476, 2011. © Springer-Verlag Berlin Heidelberg 2011

472

S. Xu et al.

Fig. 1. The quenching process of heavy rails

The change law of the residual stress during the entire wind cooling of rail-ends was: at the early stage of cooling, surfaces were placed in tension but centers were placed in compression; at the latter stage of cooling, surfaces were compressed while centers were tensed. The final residual stress state at the end of cooling was that the compressive stress existed in the surfaces whereas the tensile stress occurred in the centers. As it showed, variety of the thermal stress introduced during the rail-end quenching was complex and instantaneous. The change law as a function of time couldn’t be illustrated precisely by formulas or experience. However, not only the thermal stress distribution at a certain time during the quenching process could be simulated, but also the stress concentration and magnitude were attained directly by the finite element method.

3 Stress Distribution during Quenching 3.1 Technology of Rail-End Quenching The chemical composition of materials used for U71Mn heavy rails was shown in Table 1. The heavy rail’s heat-treatment technology involved as follow: firstly the section 200mm apart from the rail-end was heated to 910 for 40s in the electromagnetic field, then held for 5s, and cooled for 25s by intensive convection of cold air, at last air cooled to room temperature. The wind-cooling apparatus was shown in Fig.2.

℃

Table 1. Chemical composition of materials used for U71Mn heavy rails (wt%) Brand name Chemical compositon % C Si Mn P S U71Mn 0.65~0.77 0.15~0.35 1.10~1.50 λ2 > λ3 ... > λ p ) ⎜ ⎟ M ⎪ M ⎜⎜ ⎟⎟ ⎪ ⎪ ⎪ b b ... b pp ⎠ ⎝ p1 p 2 ⎪⎩0 0 0 ... λ p ⎪⎭

(2)

We assume that the relevant matrix R is the matrix which we get after the original indicators (x1, x2 , ⋯, xp) standardized. Then λi are the eigenvalues account to p. ci1, ci2, ⋯, cip are eigenvectors. We take these eigenvalues into formula α i Then we get the contribution rate cient of keywords’ properties.

αi

p

= λi / ∑ λk . k =1

of new indicator zi. That is the weight coeffi-

Keyword Extraction Algorithm Based on Principal Component Analysis

505

3 The Properties of Keyword Before computing keywords’ weights, we must do the processes of word segmentation and POS tagging. POS tagging is a process that determines each word’s grammatical category in the given text and determines POS. It is also a process that we analyze, calculate and tag the word [2] which we have cut. We use the word components ICTCLAS [3] (Institute of Computing Technology, Chinese Lexical Analysis System) which is made by Chinese Academy of Sciences Institute of Computing Technology. After removing the stop words[4] which Ti=0, we gain the keywords. Then we will talk about keywords’ 7 properties. (1)Word Frequency Word frequency consists of two parts. First one is term frequency. The other one is inverse document frequency. Term frequency (TF) is times that the given word occurrences in knowledge base. Inverse document frequency (IDF) is a measurement of words’ general importance. The way to calculate a word’s IDF is that the number of texts which contain this word is divided by the total number of words and then we deal with the quotient in logarithm. Through statistical methods we get a weighted formula of Word Frequency by improving the way TF-IDF [5]:

Fi =

n × log( M / m) N

(3)

n is the number that the word Ci appears in the text. M is the number that the word Ci appears in other texts. M is the number of the total texts. N is the number of the total words. Through formula3, we know that TF-IDF tends to filter out common words and retain the important words. (2)Part of speech (POS) Firstly, different POS correspond to different importance of words in a text, for example cc (Coordinating Conjunction), o (Onomatopoeia), w (Punctuation), etc. These POS have little impact on the texts. We give these POS lower weighted. Other POS, for example, nr (Name), nz (Terminology), v(Verb), etc have higher impact on the texts. We give them higher weight. Secondly, Chinese sentence is composed by phrases and the relation among words in phrases is either modified or be modified. Words in different locations show different degrees of importance and then the weights among words are different. Through statistical methods, we summarize six categories phrases: Subject-predicate phrase, Modification-center phrase, Verb phrase, Verb-complement phrase, Parallel phrase, Prepositional phrase. According to their POS, we weight them separately. We take parallel phrase for example. It is a parallel relationship between two words. Then the two words have the same weight. We take Verb-complement phrase as example. Verb is a main and complement is a modification of the verb. So the weight of the verb is higher than complement. Through two areas above and statistical experiments, we divide 79 POS which we generalized into five grades. We have the right values normalized:

Xi =

posi 5

(4)

506

C.-J. Li and H.-J. Han

posi indicates that the POS weight recorded in the weight table. Through formula4, the POS like conjunction have lower weight. Whereas, the POS that is modified like noun has higher weight. (3) Word length weighted Word length also determines the importance of a word. The longer the words in Chinese are more important.

Li =

leni L max

(5)

leni is the length of word Ci and Lmax is the length of the longest word. (4) The contribution rate of the word The meaning of a sentence is composed of the meaning of the words and the relationship between these words. The co-occurrence relation is the most direct relationship among words. The meaning of the sentence is embodied in the co-occurrence relation among every word. So the co-occurrence relation in the sentence is also a property of word. We assume that the total time the word Ci appears in all sentences M within the same text is Si, that word frequency tf(Ci, M). The total time word Cj appears in the M(a set of all sentences) is Sj, namely word frequency tf(Cj, M). The co-occurrence frequency between Ci and Cj is Sij.

Pij =

Sij Sii + S jj − Sij

=

Sij Si + S j − Sij

(6)

Pij is the co-occurrence probability [6] between word Ci and Cj. We know Pij=Pji, Pii=1. Finally, we can gain a co-occurrence probability matrix in word space through the word database. The contribution rate of the word Ci is :

Di = ∑ Pij j =1

Higher contribution rate means that these words often modify other words or be modified. We believe that these words are more important and make a greater contribution to sentences. These words have higher weight and which the words associated with also should have higher weight. (5)Stop words Stop words is characteristic of high frequency in all the documents, but the attributes to the document subject was very small. However, the level of the stop words should also be different. Some words may not be stop words in this one and they may be stop words in other articles. So as to statistical, this paper will arrange the stop words into four grades, Ti respectively for 0, 0.2, 0.4 and 1. 1 means the words is not stop words; 0 means the words is stop words certainly, we can delete them directly. (6)Keywords position[7] The words in the headlines are more likely keywords, so a given position attribute value Wi = 0.2 or 0.8, 0.8 meaning the word in headline, 0.2 said not. (7) Cited stressed It represent whether the words are closed by quotes, bracketed quotation marks, etc.

Keyword Extraction Algorithm Based on Principal Component Analysis

507

4 Compute the Weight Coefficients through PCA Because of the multifarious properties, this paper hopes to use fewer variables and get more information, so we use PCA to compute the weight coefficients. 4.1 Original Data Standardization We statistic all keywords and gain a sample observations with the number n, using xij to express. i stands for sample(i=1,2,3…n). That is the keyword of the text. j stands for the indicator(j=1,2,3…p). Here p is equal to 7. We use Z-score here:

xij' =

xij − x j

(7)

sj

x j is mean value. s j is standard deviation. Then we compute Correlation coefficient matrix R.

R = (rjk ) p* p , correlation coefficient: rjk =

1 n ' ' ∑ x ij xik n − 1 i =1

，p=7.

4.2 The Eigenvalue of R, Eigenvectors and Weight Coefficients Make the main diagonal elements of the correlation matrix R obtained above instead of (1 − λ ) ,constitute the matrix determinant, and make value of the determinant for 0,then can work out p eigenvalues of λ1 , λ2 ,...λ p . Make the eigenvalues into the equation AX = λ X which can be obtained with the corresponding eigenvectors of the eigenvalues. From the front reasoning: λi means the variance of the ith new index, eigenvector

means

coefficient of the orthogonal transformation, ' .The percentage of each new factor’s corresponding zi = c x + ci 2 x 2 + ... + cip x p ' i1 1

variance λi /

'

p

∑λ k =1

k

means the relative position of this variable in all of the variables,

called contribution, for

αi . αi

is the weight coefficient of the new index for the

highway network node important degree. Keywords’ last weight scores are:

keywords (Ci ) = α F Fi + α X X i + α L Li + α D Di + αT Ti + αW Wi + α Q Qi

(8)

5 Experiment and Analysis In order to verify and test the modified keywords extraction method that the article presents preferably, we collect and organize test data from multi-datasource as far as

508

C.-J. Li and H.-J. Han

possible, include 100 conference papers, 100 News Agency text files, and 100 Internet texts, total 300 experimental texts. This article conducts artificial statistical analysis of keywords for each article and finds 20 keywords. Test 1: Contrast precision ratio Table 1. Compared with Baseline

This text Baseline

Texts of meeting 87.5% 58.5%

Comments 86.9% 52.6%

Texts of Internet 83.7% 56.5%

Test 2: The comparison of longest common subsequence methods To sort the 20 artificial finding keywords, the result of the algorithm is a size sequence, and then presenting the longest common subsequence methods detection, obtaining the matching rate for 71.2%. This experiment shows that the algorithm on the importance of keywords ranking also has a good effect.

Acknowledgement This work is supported by National 863 High-Tech program of China (2009AA01Z304), National Natural Science Foundation of China (60603077) and Shan Dong Province Natural Science Foundation of China (ZR2009GM029, ZR2009GQ004, Y2008G37, Z2006G05).

References 1. Li, Y.-S., Zeng, Z.-X., Zhang, M., Yu, S.-J.: Application of Primary Component Analysis in the Methods of Comprehensive Evaluation for many Indexes. Journal owf Hebei University of Technology 1(28), 94–197 (1999) 2. Liu, L., Zeng, Q.-T.: An Overview of Automatic Question and Answering System. Journal of Shandong University of Science and Technology (Natural Science) 26(4) (2007) 3. Information on, http://www.ictclas.org/ 4. Hua, B.-L.: Stop-word Processing Technique in Knowledge Extraction. New Technology of Library and Information Service, 48–51 (2007) 5. Jia, Z.-F., Wang, Z.-F.: Research on Chinese Sentence Similarity Computation. Science&Technology Information (11), 402–403 (2009) 6. Gong, J., Tian, X.-M.: Methods of Feature Weighted Value Computing based on Text Representation. Computer Development & Applications 21(2), 46–48 (2008) 7. Deng, Z., Bao, H.: Improved keywords extraction method research. Computer Engineering and Design 30(20), 4677–4680 (2009)

A Web Information Retrieval System Tae-Hyun Kim1,*, Dong-Chul Park1, Woong Huh1, Hyen-Ug Kim1, Chung-Hwa Yoon1, Chong-Dae Park1, Dong-Min Woo1, Taikyeong Jeong1, Il-Hwan Cho1, and Yunsik Lee2 1

Dept. of Electronics Engineering, Myong Ji University, Korea [email protected] 2 System IC R&D Division, Korea Electronics Tech. Inst., Songnam, Korea [email protected]

Abstract. An approach for the retrieval of price information from internet sites is applied to real-world application problems in this paper. The Web Information Retrieval System (WIRS) utilizes Hidden Markov Model (HMM) for its powerful capability to process temporal information. HMM is an extremely flexible tool and has been successfully applied to a wide variety of stochastic modeling tasks. In order to compare the prices and features of products from various web sites, the WIRS extracts prices and descriptions of various products within web pages. The WIRS is evaluated with real-world problems and compared with a conventional method and the result is reported in this paper. Keywords: web, information retrieval, Hidden Markov Model.

1 Introduction The internet has become a vital source for information. However, as the number of web pages continues to increase, it becomes harder for users to retrieve useful information including news, prices of goods, and research articles. Commercial search engines are widely used to locate information sources across web sites. One obstacle that has to be faced when searching the web page via a query is that commercial search engines usually return very large hit lists with a low precision. The list inevitably includes irrelevant pages and the results are ranked according to query occurrences in the documents rather than correct answers within the documents. Especially, these search engines focus on the recall ratio instead of the precision ratio. Users have to find the relevant web pages from amongst irrelevant pages by manually fetching and browsing pages to obtain specific information. For these reasons, many researchers are trying to develop solutions to perform the web information retrieval (WIR) function in a more effcient and automatic manner[1]-[4]. Early successes in WIR include the PageRank algorithm [5] and the HITS algorithm of Kleinberg [6]. Other linked-based methods for ranking web pages have been proposed including variants of both PageRank and HITS [7]. The PageRank algorithm globally analyzes the entire *

Corresponding author.

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 509–514, 2011. © Springer-Verlag Berlin Heidelberg 2011

510

T.-H. Kim et al.

web graph while HITS algorithm analyzes a local neighborhood of the web graph containing an initial set of web pages matching the user's query. The process of extracting information from the result pages yielded by a search engine is termed web information retrieval. Several automated or nearly automated WIR methods have been proposed representative methods such as Mining data records in web page (MDR) [12], Object mining and extraction system (OMINI) [13] and Information extraction based on Pattern Discovery (IEPAD) [14]. In this paper, a Web Information Retrieval System, called WIRS, is utilized to extract prices of goods on the internet. From a large number of web pages relevant to prices of goods appearing on the internet, the WIRS can help to extract the prices of goods of interest with maximal accuracy. The remainder of this paper is organized as follows: A brief review of HMM and the web information retrieval system based on HMM are presented in Section 2. Section 3 describes an actual implementation of WIRS and experiments involving a practical price extraction problem and presents results including comparisons with a conventional MDR. Finally, Section 4 concludes the paper.

2 Hidden Markov Model and Web Information Retrieval System A Hidden Markov Model (HMM) is a statistical model in which the system is assumed to be a Markov process with unknown parameters and the hidden parameters are found from the observable parameters[11]. In a HMM, the state is not directly visible, but variables in°uenced by the state are visible. Each state has a probability distribution. HMMs are especially useful for their applications in temporal pattern recognition tasks including speech recognition and data mining[11, 15, 16]. A discrete output, a first order HMM, is a finite state automaton and can be represented by a 5-tuple {S, V, ∏, A, B} where S is a set of values for the hidden states, S = {s1…sN}, N is the total number of possible states; V is a set of values for the observations V = {v1…vM}, M is the total number of possible observation values; ∏ is a set of initial probabilities for all the states, ∏ = {πi}, i = 1, 2,…, N; A is a mapping defining the probabilities of each state transition A = {P(q → q’)}; and B is a mapping defining the emission probability of each observation value on each state, B = {P(q ↑σ)}. More detailed information on HMM can be found in [11]. The WIRS used in this paper consists of the following components: 1) Page retrieval, 2) Segmentation and Parser, 3) Segment filter, 4) Observation creator, and 5) Extractor. The input to the WIRS is the URLs of a search engine’s interface and the output of the system is the list of extracted slots of each product: name, price, image, and URL. Four major segments are paragraph, table, list, and heading. The segment filter decides which segment belongs to the product segments. The adjacent segments containing product names, description, images, and prices are then grouped into a larger segment for the observation creator. The most probable state token is then found by using the Viterbi algorithm [11]. An example of the extracted data record as shown in Fig. 2 is stored in a table format. More detailed information on web information retrieval system based on HMM can be found in [17].

A Web Information Retrieval System HTML page

511

Segment tree

Document Laptop pricing

price of latop

Paragraph Table 1.6GHz 6000 PC

Table Table text Sony VAIO 1.6 GHz Pentium $1,209 price of labtop text text text text Dell Inspiron 6000 $1,440 Sony VAIO $1,029 Dell Inspiron $1,440

pentium Notebook

Fig. 1. A HTML document and the corresponding segment tree

Fig. 2. A screen of product extraction

3 Experiments and Results In this section, we demonstrate the performance of the WIRS using a real world data in comparison with a conventional system, MDR, which is a state-of-art web information extraction systems based on HTML tag structure analysis. The data set for training each product contains observations of 100 URLs returned from a general-purpose

512

T.-H. Kim et al.

search engine, Google, and the next 100 URLs are prepared for testing. Some typical web sites with sufficient information about product features and prices are listed in Table 1 while other web sites with irrelevant features are omitted. The trained HMM used in the performance evaluation includes 8 states for labeling. 3.1 Measurement for Performance Evaluation The performances are evaluated in terms of the precision and recall [18], which are widely used to evaluate information retrieval and extraction systems. These are defined as follows: Precision = CE/EO and Recall = CE/CO, where CE is the total number of correctly extracted observations, EO is the total number of extracted observations on the page, and CO is the total number of correct observations (target observations). Precision defines the correctness of the data records identified while recall is the percentage of the relevant data records identified from the web page. 3.2 Experiments The WIRS is evaluated and compared with MDR on a real-world problem. An executable program of MDR can be downloaded at [19]. MDR has a similarity threshold, which was set at 60% in our test, as suggested by the authors of MDR. The product description or data record is considered as a block containing both the product information and the noisy information in MDR. The WIRS, however, can handle the web pages deeper, because it is able to extract the specific field of the data record: image, name, price, and URL excluding noisy objects. Table 1 shows results of the performance comparison among the WIRS and MDR. A total of 10 web sites containing different formats and product information are evaluated in this experiments. In Table 1, column 3 shows the number of target products available in the corresponding URL shown in column 2. The listed web pages are Table 1. Performance Comparison of WIRS and MDR WIRS Returned Correct

No

Web sites

Products

1 2 3 4 5 6 7 8 9 10

www.flash-memory-store.com www.tigerdirect.com www.usbflashdrivestore.com www.supermediastore.com www.ecost.com www.pricespider.com www.usanotebook.com www.nextag.com www.mysimon.com shpping.aol.com

21 19 25 32 25 10 27 11 25 16

16 16 24 30 22 10 21 10 21 16

211

188 184 Rc : 87.2% Pr : 98.9%

Total Recall Performance (Rc) Precision Performance (Pr)

16 16 24 28 22 10 21 10 21 16

MDR Returned Correct 38 23 25 0 27 16 27 15 11 30

20 19 25 0 25 10 27 10 16 16

212 168 Rc : 79.6% Pr : 79.2%

A Web Information Retrieval System

513

returned from a commercial search engine, Google, and the web pages include sufficient information about product features including prices, images, and descriptions regarding the specific products such as usb flash drive, laptop, web camera, computer mouse, etc. The numbers shown in the rest of columns are the numbers of extracted products and correctly extracted products from the WIRS and MDR systems. As shown in Table 1, the average recall obtained by the WIRS is 87.2% while recall of 79.6% are obtained by MDR, respectively. With respect to the extraction precision, the WIRS proves to be a more powerful tool for the information extraction. The average precision obtained by the WIRS is 98.9% while precision of 79.2% is obtained by MDR. In our experiments, MDR provides better extracting records from HTML tables than those from non-tables while the WIRS method performs well in both cases. The WIRS method shows far better performance than MDR due to the fact that MDR was primarily designed to handle tables only. In addition, MDR does not identify the correct data sections for extracting product records and MDR sometimes extracts some advertisement records. The WIRS method is fully automated, since the extraction process is performed on URLs returned from any general-purpose search engine.

4 Conclusion In this paper, a novel and effective web information retrieval method to extract product prices from internet is evaluated. This method can correctly identify the data region containing a product record. When a data region consists of only one data record, the MDR fails to correctly identify the data region. The WIRS is evaluated and compared with MDR on a real-world problem. The WIRS method overcomes the drawbacks of the conventional MDR in processing HTML contents. Experiments show that the WIRS outperforms the MDR method significantly in terms of precision and recall.

Acknowledgment This work was supported by National Research Foundation of Korea Grant funded by the Korean Government (2010-0009655) and by MKE Configurable Device and SW R&D"(No.: KI002168).

References 1. Chorbani, A.A., Xu, X.: A fuzzy markov model approach for predicting user navigation, pp. 307–311 (2007) 2. Godoy, D., Amandi, A.: Learning browsing patterns for context-aware recommendation. In: Proc. of IFIP AI, pp. 61–70 (2006) 3. Bayir, M.A., et al.: Smart Miner: A New Framework for Mining Large Scale Web Usage Data. In: Proc. of Int. WWW Conf., pp. 161–170 (2009)

514

T.-H. Kim et al.

4. Cao, H., et al.: Towards Context-Aware Search by Learning A Very Large Variable Length Hidden Markov Model from Search Logs. In: Proc. of Int. WWW Conf., pp. 191– 200 (2009) 5. Brin, S., Page, L.: The Anatomy of a Large-Scale HypertextualWeb Search Engine. In: Proc. of Int. WWW Conf., pp. 107–117 (1998) 6. Kleinberg, J.M.: Authoritative Sources in a Hyperlinked Environment. Journal of the ACM 46(5), 604–632 (1999) 7. Tomlin, J.A.: A New Paradigm for Ranking Pages on the World Wide Web. In: Proc. of. Int. WWW Conf., pp. 350–355 (2003) 8. Rilo, E., Jones, R.: Learning Dictionaries for Information Extraction by Multi-Level Bootstrapping. In: Proc. of the 16th National Conf. on Articial Intelligence, pp. 811–816 (1999) 9. Sonderland, S.: Learning information extraction rules for semi-structured and free text. Machine Learning 34(1), 233–272 (1999) 10. Leek, T.R.: Information Extraction Using Hidden Markov Models. Master thesis, UC, San Diego (1997) 11. Rabiner, L.R.: A tutorial on hidden Markov models and selected applications in speech recognition. Proc. of IEEE 77(2), 257–286 (1989) 12. Bing, L., Robert, G., Yanhong, Z.: Mining data records in web pages. In: Proc. of ACM SIGKDD, pp. 601–606 (2003) 13. Buttler, D., Liu, L., Pu, C.: A fully automated object extraction system for the world wide web. In: Proc.of IEEE ICDCS, pp. 361–370 (2001) 14. Chang, C., Lui, S.: IEPAD: Information extraction based on Pattern Discovery. In: Proc. of WWW Conf., pp. 682–688 (2001) 15. Park, D.-C., Kwon, O., Chung, J.: Centroid neural network with a divergence measure for GPDF data clustering. IEEE Trans. Neural Networks 19(6), 948–957 (2008) 16. Jiang, J.: Modeling Syntactic Structures of Topics with a Nested HMM-LDA. In: Proc. of ICDM, pp. 824–829 (2009) 17. Park, D.-C., Huong, V.T.L., Woo, D.-M., Hieu, D., Ninh, S.: Information Extraction System Based on Hidden Markov Model. In: Yu, W., He, H., Zhang, N. (eds.) ISNN 2009. LNCS, vol. 5551, pp. 55–59. Springer, Heidelberg (2009) 18. Raghavan, V.V., Wang, G.S., Bollmann, P.: A Critical Investigation of Recall and Precision as Measures of Retrieval System Performance. ACM Trans. Info. Sys. 7(3), 205–229 (1989) 19. http://www.cs.uic.edu/~liub/WebDataExtraction/ MDR-download.html

An Effective Intrusion Detection Algorithm Based on Improved Semi-supervised Fuzzy Clustering Xueyong Li, Baojian Zhang, Jiaxia Sun, and Shitao Yan School of Information Engineer, Henan Institute of Science and Technology Xinxiang, 453003, China [email protected]

Abstract. An algorithm for intrusion detection based on improved evolutionary semi- supervised fuzzy clustering is proposed which is suited for situation that gaining labeled data is more difficulty than unlabeled data in intrusion detection systems. The algorithm requires a small number of labeled data only and a large number of unlabeled data and class labels information provided by labeled data is used to guide the evolution process of each fuzzy partition on unlabeled data, which plays the role of chromosome. This algorithm can deal with fuzzy label, uneasily plunges locally optima and is suited to implement on parallel architecture. Experiments show that the algorithm can improve classification accuracy and has high detection efficiency. Keywords: intrusion detection; semi-supervised learning; clustering; evolutionary programming.

1 Introduction With the rapid development of the Internet, network intrusion event are frequently happened, intrusion detection technology shows more and more important role. Data classification is to define the process of attack and attack recognition, more specific technology can achieve this process such as pattern matching, statistical analysis, integrity analysis and other methods, the nature of these method is to compare to get difference between normal data and detect data; According the difference to determine whether the system had been invaded [1]. Intrusion detection system is similar to most machine learning systems, to rely on existing labeled datum. But got tag data is more difficult, it needs professional worker to spend a lot of time to collect and identify large amount datum. Unlabeled data access is much easier than the label data, but only contrast classification only with the unlabeled data is less effective [2]. Clustering technology can be used to identify nature cluster of unlabeled data, but the class division is not always consistent with natural clustering [3].Fuzzy clustering algorithm has better classification accuracy and generalization ability, and evolution semi-supervised learning algorithm has better self-learning nature. The evolution semi-supervised combines with fuzzy clustering, namely ESSFC. In this paper, (Improved Evolutionary Semi-Supervised Fuzzy Clustering) [4] IESSFC is uses to intrusion detection systems, label data as the role of chromosomes, provide the class label information is used to guide each chromosome R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 515–520, 2011. © Springer-Verlag Berlin Heidelberg 2011

516

X. Li et al.

evolution. The fitness of each chromosome includes cluster variance of unlabeled data and misclassification error of labeled data. The classification structure got by IESSFC used to classify new unlabeled data. The experiments verify the efficiency of IESSFC.

2 Improved Evolutionary Semi-supervised Fuzzy Clustering We are provided with a small number of labeled data and a large number of unlabeled data. We suppose the features number is k. Set the number of labeled data is n l , the number of unlabeled data is n u .Thus, all the labeled and unlabeled documents provided can be denoted in a matrix form: ⎧ ⎫ ⎪ ⎪ X = ⎨ x1l , L x nl l x1u,L x nuu ⎬ = X 1 42 4 3 1 2 3 ⎪⎩ labeled unlabeled ⎪⎭

l

∪ X

u

(1)

Here, l indicates the designation labeled data, and u, as a superscript, indicates the designation unlabeled data. n l = X l , n u = X u , n = X = n l + n u . Assume that all data will be classified into C clusters. A matrix representation of a fuzzy c-partition of X induced by Eq. (1) has the form: U U n ⎡ 6 4 4 4 447 4 4 4 4 48 6 4 4 4 447 4 4 4 4 48 ⎤ fuzzy membership s fuzzy membership fuzzy membership s⎥ s ⎢ fuzzy membership s of labeled data n 1 of unlabeled data 1 of unlabled data n u of labeled data 1 } } } } ⎢ ⎥ ⎢ u 1l n1 L L ⎥ u 11u u 11u u 11l ⎢ ⎥ l l u u u L L u u u 2 n1 21 21 21 ⎥ U = ⎢ ⎢ ⎥ M L L M M M ⎢ ⎥ l l u u u L L u c1 u c1 u c1 ⎢ cn 1 ⎥ ⎢ ⎥ ⎢ ⎥ ⎦ ⎣⎢ l

(2)

Here, the fuzzy values in U l are determined by domain experts after a careful investigation on X l . In general, Eq. (2) should satisfy the following conditions:

∑ ∈ [0,1] , ∑

u ihl ∈ [0,1] , u ihu

c

u ihl = 1 , 1 ≤ i ≤ c,1 ≤ h ≤ nl

(3)

u ihu = 1 , 1 ≤ i ≤ c,1 ≤ j ≤ nu

(4)

i =1

c

i =1

The goal of the problem is to construct, using X, a classifier which can assign a future new pattern to one or more pre-defined classes with as least error as possible. 2.1 Misclassification Error of Labeled Data

In order to get good generalization performance, classifier to be constructed should minimize the misclassification error of labeled data. We will use the variance of the fuzzy memberships of labeled data to measure the misclassification error. Given a fuzzy c-partition on X u , the C cluster centers v1 , v 2 , L v c can be computed by Eq. (5).

An Effective Intrusion Detection Algorithm nl

nul

nl

517

nu

vi = (∑ (u ) x + ∑ (u ) x ) /(∑ (u ) + ∑ (u iku ) 2 x kl )

(5)

(

(6)

k =1

l ik

2

l k

k =1

u ik

u k

2

k =1

u ijl ' = ⎡∑ h =1 ( x lj − v i ⎢⎣ c

l ik

C

2

k =1

) /( x lj − v h

)

2 ) ⎤ ⎥⎦ C

−1

For, i = 1,2, L , c and j = 1,2, L n l , the fuzzy memberships of labeled data can be recomputed as (6), C is a covariance matrix, and x lj − v i

2 C

= ( x lj − v i ) T C ( x lj − v i ) . Accord-

ingly, the misclassification error of labeled data, denoted as E, can be measured as a weighted sum of variance between u ijl and u ijl ' .

E = ∑ j l ∑i =1 (u ijl ' − u ijl ) 2 x uj − v i n

c

2 c

(7)

Although minimizing misclassification error of labeled data is necessary for the classifier to get good generalization ability. Fuzzy within cluster variance is a well-known measurement of cluster quality in fuzzy clustering, which is defined as:

J = ∑ ju=1 ∑i =1 (u iju ) 2 x uj − vi n

c

2 c

(8)

We can see that minimizing fuzzy within cluster variance is equal to maximizing the similarity of data within the same cluster. Thus, we argue that fuzzy within cluster variance of unlabeled data can play the role of capacity control in our problem. We can define the objective function as follows:

f (U u , V ) = J + α ⋅ E

(9)

Here, α should be proportional to the ratio nu / nl .Till now, our problem has been converted to minimizing the objective function in Eq. (9). In this paper, evolutionary programming is chosen to optimize the objective function, because not only can it alleviate the local optima problem, but also it is less sensitive to initialization. More specifically, an evolutionary searching procedure is employed. 2.2 IESSFC Algorithm

The cardinal principal of ESSFC algorithm is to use labeled data information to guide evaluation of each chromosome, the fitness of each chromosome data with cluster variance of unlabeled data and misclassification error of labeled data. But this algorithm itself has two disadvantages. First: The choice of labeled data is critical to the evolution process; the original ESSFC algorithm selects the C labeled data in labeled data database by random method. If the chosen labeled data of a class in boundary zone, then these labeled data must be impact evolution process and the establishment of the classifier. In this case, to make the fitness to meet certain conditions needs more times to evolution and can cause the variance to the classifier. IESSFC we proposed overcome this disadvantage, when select the labeled data which is as the guiding role. We first calculate the center points of C classes, these points also name as mean point, and then these C center points used as the guiding labeled data to guide the evolution of unlabeled data, this

518

X. Li et al.

method can avoid selecting boundary point as the guiding data. The centre points ate calculated as follow:

vi = ∑ii=1 (∑ j i=1 xij ) /Li r

L

(10)

Here, Li indicates the data number of ith class. Second: Second, ESSFC algorithm select unlabeled data by random produce in build classification process. First, produce C real number, r1 j , r2 j , L rcj ( rij ∈ [0,1],1 ≤ i ≤ c ) as the point of one chromosome. Then calculate fuzzy degree u ijuk = rij /( r1 j + r2 j + ... + rcj ) .The data by randomly generated has uncertainty prop-

erty, the fuzzy degree generated by these real number will not in the scope of real intrusion data. These fuzzy degrees will increase the burden of evolution and to increase the variance of classification. IESSFC we proposed can also overcome this disadvantage. By calculate the distance between unlabeled and labeled data to get the r1 j , r2 j , L rcj . Specifically, calculate the distance between the unlabeled data and the centre points calculated by Eq.(10). r1 j , r2 j , L rcj are calculated as follow:

ri j = x uj − v i

2

/ x uj

(11)

This evolutionary process is more targeted, and the experimental results show that this method can effectively speed up the evolutionary process.

3 Contract Classification Based on IEFFSC Based on the above formulation, in this section, we describe the evolution press of IEFFSC to construct the classifier. Select the C centre points calculated by Eq(10) as the labeled data, and set the initial C cluster core points as v i (1 ≤ i ≤ C ) .For u ijl (i = 1,2, L c) of the labeled data can be

certain as their class label. So elements of Membership degree matrix are 0 or 1. Suppose the population size is p . The kth chromosome will be initialized through the following steps (1)-(4): (1)Generate c real numbers by Eq.(10), r1 j , r2 j , L rcj , rij ∈ [0,1],1 ≤ i ≤ c in the interval of [0,1] for the jth point of a chromosome, 1 ≤ j ≤ n u . (2)Calculate u ijuk = rij /( r1 j + r2 j + ... + rcj ) , 1 ≤ i ≤ c

, uijuk should satisfy condition

Eq(4). (3)Repeat the steps (1) and (2) n u times, j = 1,2, K , nu , and produce a chromosome. (4) Repeat the steps (1)-(3) p times to produce an initial population of size p. For k = 1,2, L p; i = 1,2, L c , determine the cluster center v ik by

vik = (∑ jl=1 (u ijlk ) 2 x lj ) / ∑ jl=1 (u ijlk ) 2 n

n

(12)

And compute the value of objective function: f

k

= j k + αE k

(13)

An Effective Intrusion Detection Algorithm

519

The following is the main step of IESSFC, set the generation counter gen=0, and the number of generations max-gen. (1)For each k = 1, K , p , generate the offspring U U ( P + K ) , here k = 1,2, L, p .

u iju ( p + k ) = (u ijlk ) 2 e

− x uj − vik

C

/ ∑ g =1 (u ijlk ) 2 e c

− x uj − vhk

C

(14)

(2)Determine the new centers and new objective function value using (15) and (16). n n n n v ip + k = (∑ l (u ijlk ) 2 x lj + ∑ u (u ijuk ) 2 x uj ) /(∑ l (u ijlk ) 2 + ∑ u (u ijuk ) 2 ) (15) j =1

j =1

f

p+k

j =1

j =1

= J p+k + α ⋅ E p+k

(16)

UK

(3)Select the p best fit form the p+p matrices of U according to the corresponding value of f , to form the next generation of population. (4) gen=gen+1. (5) Repeat the steps (1)-(4), until gen=max-gen or time allowed is exhausted. Given a new data, denoted as x, and the c cluster centers v1 , v 2 , K v c obtained with IESSFC, the fuzzy membership of x with respect to class i, u i can be computed as:

ui =

[∑ ( x − v c

g =1

2 i C

/ x − vg

2 C

)]

−1

(17)

Thus, x is assigned to c classes which has max u i .

4 Experiments Experimental data are from "KDD Cup 1999 Data"[5], including 38 kinds of attacks belong to 4 big categories. There are some redundant in 41 features, and large dimension of feature space will increase the computational complexity. In order to select some useful feature for classing, we use ID3 algorithm to construct decision tree to select important features of attributes. By ID3 we choose out 10 features, including 7 continuous features, including duration, dst_bytes, etc.; and three discrete features, including protocol type, service, flag and so on. The network data obtained must be processed to be vector form for IESSFC. For continuous features, use the standard deviation normalized to process. We define: x im

1 = n

n

∑x f =1

fm

(18) s im

1 =( n −1

n

∑ (x

fm

− x im

1 2 2 ) )

(19) xim = ( xim − xim ) / s im (20)

f =1

Where xim is the mth continuous feature of X i , s im is standard vector difference formula, xim is new continuous value after standard normalized. For discrete feature, such as{tcp, udp, icmp}convert into binary is {1,0,0},{0,1,0}{0,0,1}[6]. 4.1 Experiments Results

We selected 10000 samples from KDDCUP99, including 4000 normal data, 6000 abnormal data. Abnormal data includes seven kinds of attacks, namely: mailbomb, ipsweep, back, portsweep, smurf, snmpgetattack, mscan.

520

X. Li et al.

The parameters set as following: population size p is 10, the type of class number is 8,indicates the normal. Select the 7000 data from 10000 normalized data, generated classifier. These 7000 data includes 2500 labeled data and 4500 unlabeled data .Compute the centre points in 8 classes as the cluster centers. In calculating the objective function α =4500/2500 = 1.8.The rest 3000 data are to be test data. Iteration times were 20,40,60,80,100 respectively. In experiment, we took contrast between ESSFC and IESSFC, the experimental results shows in Table1. Table 1. Experimental results Algorithm ESSFC IESSFC

20 80.59 81.56

Iteration times and detection rate 40 60 80 100 81.06 81.75 82.19 82.46 82.97 84.32 85.63 85.96

The experimental results show that IESSFC intrusion detection algorithm is more effective than the intrusion detection algorithm based on ESSFC.

5 Conclusion This paper proposed an effective intrusion detection systems base on IESSFC, addressing the fact that labeled data is more difficult to get than the unlabeled data. IESSFC is uses to intrusion detection systems, label data as the role of chromosomes, provide the class label information is used to guide each chromosome evolution. The fitness of each chromosome data includes cluster variance of unlabeled data and misclassification error of labeled data. This algorithm can use few labeled data and large unlabeled data to construct the intrusion detection classifier. Experimental results show that the algorithm we proposed has higher detection rate than the algorithm based on ESSFC.

References 1. Jiang, J., Ma, H., Ren, D.: A Survey of Intrusion Detection Research on Network Security. Journal of Software 11(11), 1460–1466 (2000) 2. Castelli, V., Cover, T.: On the exponential value of labeled samples. Pattern Recognition Letters 16(1), 105–111 (1995) 3. Jian, Y.: An Improved Intrusion Detection Algorithm Based on DBSCAN. Microcomputer Information 25(13) (2009) 4. Liu, H., Huang, S.-t.: Evolutionary semi- supervised fuzzy clustering. Pattern Recognition Letters 24, 3105–3113 (2003) 5. KDD Cup 1999 Data. University of Califiornis, Irvine[EB/OL] (March 1999), http://kdd.ics.uci.edu/database/kddcup99/kddcup99.html 6. Wilson, D.R., Tony, R.M.: Improved heterogeneous distance functions. Journal of Artificial Intelligence Research 6(1), 1–34 (1997)

What Does Industry Really Want in a Knowledge Management System? A Longitudinal Study of Taiwanese Case Liang-Chih Yang and Hsi-Peng Lu Dept. of Information Management National Taiwan University of Science and Technology [email protected], [email protected]

Abstract. This paper depicts a longitudinal investigation of knowledge management system development from industrial perspectives. Snapshots on three surveys (2002, 2006, and 2010) of Taiwanese companies were conducted and compared, which is to explore the perceived understandings and requirements for the applications of a knowledge management system. From the surveys, it was found that the most useful applications were document management, knowledge search and retrieval, and knowledge repository and map. The emerging applications were expert management, document security, and knowledge automation such as auto-classification, auto-abstract and auto-keyword generation. The most wanted services along with KMS were consulting service, success story-sharing, and modularization while deploying knowledge management system in the enterprises. The trends and transformation of a KM system were also collected and analyzed. We suggest that a company should use different knowledge management approach according to its corporate main business function. Combing intellectual capital theories proposed by other researchers, we categorize knowledge management focus as staff-centric, system-centric, and customer-centric knowledge from industrial perspectives. Keywords: knowledge management (KM), knowledge management system (KMS), knowledge market, smiling curve, intellectual capital.

1 Introduction Knowledge management is a mechanism to help people assimilating, understanding, learning, storing, transferring, and retrieving data [1]. Knowledge management constitutes an emerging discipline aiming to support enterprise in the new business environment where the notion of economics of ideas is an important prerequisite for success and viability [2]. We may interpret knowledge management as trying to accumulate knowledge by all kinds of means, meanwhile complying with organization’s system and culture, and then to disseminate to each member of the organization. Companies with successful knowledge management can boost their competitive advantages [3]. R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 521–531, 2011. © Springer-Verlag Berlin Heidelberg 2011

522

L.-C. Yang and H.-P. Lu

Knowledge circulates within the organization all the time and creates value in use [4]. The speed might be fast or slow, and the performance might be high or low. In essence, knowledge is what employees know about customers, products, processes, and past successes and failures [5]. A key challenge in the application of knowledge is transferring it from where it was created or captured to where it is needed and should be used [6]. Thus a “knowledge market” is formed within the organization in a business context [7]. In order to manage the knowledge within the organization effectively, we need to understand market forces in the knowledge market within an organization. Most organizations think knowledge exchange is automatic and doesn’t need the driver, and there won’t be frictions as well. As a consequence, many knowledge management projects failed in the end [8]. From industrial perspectives, this paper starts with an empirical review of knowledge management development of Taiwanese companies from year 2001 to 2010. Secondly, three surveys held in 2002, 2006, and 2010 on knowledge management issues from MIS and IT professionals in Taiwan were conducted and compared, which is to explore the perceived understandings and requirements for the applications of a knowledge management system. Finally, combing intellectual capital theories of other researchers, we propose a framework based on Michael Porter’s value chain and Stan Shih’s smiling curve, and suggest that a company should use different knowledge management approach, according to its corporate main business function.

2 Review of Knowledge Management Progress in Taiwan 2.1 Year 2001-2004 The period of year 2001 to 2004 was very important to the knowledge management market in Taiwan. On the user side, the KM issues were discussed, and the concepts were gradually mature. On the KMS vendor side, various KM application systems appeared, and they tried to fit the KM issues from different perspectives. 2.1.1 KM Portal as a Major Application The concept of portal has been well developed for the past few years. Portal became the “killer apps” in Enterprise Application Integration (EAI). The main features of a Portal consist of: (1) Integrating the interface of business logics (2) Providing unified user interface In term of integrating the interface of business information, KMS plays the similar role to Portal. During this period, the “document management” is the closest to the idea of “knowledge management” among the variety of KM software features. But at the same time, another mechanism of sharing, provided by the “groupware”, can make up for the shortcoming of ‘overly static’ traditional document management. These two important mechanisms can be successfully combined by means of parallel integration of portals. Compared with other solutions, portal is an ideal one because it allows members in the organization to acquire the knowledge through the unified user interface. KM Portal combines the ideas of portal and knowledge management, and integrates the back office on a unified user interface.

What Does Industry Really Want in a Knowledge Management System?

523

2.1.2 The Needs for KM Climbed during Economic Downturns During the economic downturn, companies started to look for some other better business models. One is separating its operation units all over the world and implementing global logistics in order to reduce the cost and respond spontaneously to the market. Under this transition, the establishment of knowledge management system is endowed with the important task of integrate the team, manage the scattered, fragmentary information, and quickly set up an open communication channel. As a result, the development of software application was reaching the mature stage, and the overall environment also encouraged the trend of KM application, and that’s why the domestic need for knowledge management was climbing. 2.1.3 From Exploration to Implementation – Success Stories from All Kinds of Industries With the support from the government, industry, and academia, the development of e-Business also triggered the need for all sorts of application software integration as well as the trend for information integration and industrialization. Documents could best represent the corporate value in firms with expertise; and the productivity could be greatly increased by using the KM system to conduct sorting, organizing, and simultaneous full-text searching. A lot of time could be saved from searching, and the KM system could still be accessible outside of the office. Some other companies applied the KM system on the collecting and sharing of information, and online storage of work experience. Colleagues in charge of different functions could distribute the collected and organized information through the system, and the system would send an automatic notice to relevant people. Managers served as mediator in this process, not only staying in control of the team work through job assignments, but also keeping a track record of the practical experience in implementation through the storage of comments and modifications made by the staff. Companies in the telecom service industry in Taiwan aggressively set up the knowledge network, applying KM system on the collection of business intelligence, installation and sharing of management system of knowledge assets, even on the structural data analysis, in the purpose of assisting knowledge workers to quickly get the most useful knowledge in this fiercely competitive market. As for the manufacturing industry, companies built up the knowledge-based community in order to allow colleagues in their specialized areas to exchange their experiences. No matter it is about storing working experience or making effective use of corporate resources, the goal is to cut down on the time for work and decision-making. 2.2 Year 2005-2008 Based on the best practice of KMS in various industries, the KM market experienced a rapid growth in 2005-2008 and moved toward a more diversified and specialized direction. The further integration of portal and KM also strengthened the synergy of corporate information application integration. 2.2.1 More Integration of KM into Corporate Process Knowledge management varies according to corporate operations and corporate cultures. In 2005, the KM technology venders started to provide solutions targeting at different industry usage or operations needs. For example, in the service sector, the

524

L.-C. Yang and H.-P. Lu

knowledge management of customer service integration enabled the corporation to share its knowledge with customers, or as a feedback for internal adjustment within the corporation. In the manufacturing sector, the supply chain management could also facilitate the sharing of knowledge, and create industry value. In the financial sector, given the need of quick response to the rapidly changing market, the integration of knowledge management could shorten the response time, and even predict the market change based on data analysis of the past. 2.2.2 Web-Based KM Portal Emerged Given the features of easy to operate and with low training cost of web interface, the establishment of web-based intranet became very common in recent years. In addition, the features of cross-platform and no installment on the client side also made web-based technology an inevitable trend for corporate application software. As there was more and more corporate application software available, the integrated interface provided by corporate internal portal also became a necessary tool. As there was an increasing need from the domestic companies, a lot of international well-known brands of portal software platform invested a lot of human resources on better services in Taiwan. In addition, the Enterprise Information Portal (EIP) built in companies served better than expected. Therefore, many technology venders joined the Portal market. As a result, Portal platforms, one by one, were announced to be able to support J2EE, J2ME, Web Services, and XML, moving toward the goal of standardization and flexibility. Ovum once pointed out that after 2005, application software wouldn’t be widely accepted by users if it couldn’t be operated through integrated portal interface. For corporate users, even though market competition brought about more choices and low cost, the following concerns remained while choosing a KM software: web-based, standardized, expandable, and integrated. Only in this way can the corporate users guarantee that their investment is well-spent and the operations can be conducted on different plate forms - with the benefits of making the corporate KM system easier to development, modify, set up, and maintain, and reducing the complexity of system integration. 2.2.3 e-Taiwan Triggered the New Demands for KM Compared with common corporate organizations, the need for government agencies to implement knowledge management is even greater. From computerization of all government paper works, public service portal, to every sector of the e-Government, the government has set up a lot of administrative information system and produced a gigantic number of files during the process of digitalization. All these files (e-documents) need to be preserved and sorted properly for public use. To accelerate the flow of administrative information, it is a daunting task to establish a mechanism to manage the huge database effectively and strengthen the integration of process and information. As a result, all local governments have started to implement knowledge management since 2002, in the hope of preserving the knowledge effective through the assistance of information technology. With “e-Taiwan” as the vision, the Executive Yuan approved the “National Information and Communication Initiative” (NICI) on December 26, 2001 – the goal is to invest at least NT$40 billion in the future, and complete the “e-Government,”

What Does Industry Really Want in a Knowledge Management System?

525

“e-Industry,” “e-Society,” and infrastructure projects by 2008, in order to strengthen Taiwan’s overall information and telecommunication development, and encourage the investment and implementation of all industries. Among these are many business opportunities for the consulting and technical services related to knowledge management. 2.2.4 The Corporate Information Infrastructure and Integration Were Expanded It is no longer sufficient to divide corporation information systems into “internal” and “external” ones. For a Business Process Re-engineering (BPR) of the corporate information system, the system should be divided into “daily operational support” and “sustainable corporate value” in functions. The former includes “ERP” (Enterprise Resource Planning) or even “POS” (Point of Sales), while the latter is represented by the corporate knowledge management system. In 2007, these corporate external and internal, daily operational and sustainable corporate value systems has further integrated. As mentioned above, large-scale corporations continued to strengthen their process and information integration, in order to boost their core competency. For those small and medium sized enterprises that have implemented the basic KM infrastructure (such as document management, search engine, and personnel management, etc.) also gradually added on other KM modules to the existing platform. As a result, the introduction and application of Application Server and Portal Platform has experienced a robust growth in 2007, and gradually connected with all kinds of corporate information system applications. 2.3 KM Market in 2009-2010 2.3.1 Cloud Computing Triggered the Next Phase of Business Applications Cloud Computing refers to both the applications delivered as services over the Internet and the hardware and systems software in the datacenters that provide those services [9]. The services themselves have long been referred to as Software as a Service (SaaS), Platform as a Service (PaaS), and Infrastructure as a Service (IaaS). Business will benefit from using these services to reduce system maintenance, operation, and human labor costs; while the system and data security issues must be well taken care of in the first instance.

3 The Longitudinal Study of KM Surveys This research was incorporated with a local knowledge management system (KMS) vendor – eXpert Technologies Corporation (a pseudonym). Its knowledge management product – eKM, has been selected as No.1 Knowledge Management System since 2006 in Taiwan. During the period from year 2002 to 2010, the company held several product seminars every year. In each seminar, there were around 200 attendees, representing 100 to 150 enterprises in Taiwan. In this paper, we took a snapshot on the questionnaires collected in the year of 2002, 2006 and 2010. We passed out the questionnaires in the beginning of the seminar, and collected them at the end. The audiences were mostly from various industries, and basically they would get some ideas about KMS from the seminars. Some of them even had the experience in using two or more features of KMS. From their answers, we wanted to know how the KMS was deployed in Taiwan, and thus to figured out their further requirements.

526

L.-C. Yang and H.-P. Lu

There were 9 questions in the questionnaire, and were divided into 5 sections. The complete questionnaire is in Table 1. Table 1. The questionnaire

͘ ŽŵƉĂŶLJĂĐŬŐƌŽƵŶĚ ϭ͘ ,ŽǁŵĂŶLJĞŵƉůŽLJĞĞƐĂƌĞƚŚĞƌĞŝŶLJŽƵƌĐŽŵƉĂŶLJ͍ Ϯ͘ tŚŝĐŚŝŶĚƵƐƚƌLJĚŽĞƐLJŽƵƌĐŽŵƉĂŶLJďĞůŽŶŐƚŽ͍ ͘ ŽŵƉĂŶLJ͛ƐŝŶƚĞŶƐŝŽŶŽŶĚĞƉůŽLJŝŶŐ .Counter. assigned _ resource( resource _ id ). ACTION .release _ resource[resource _ id ].FINISH

Definition 5. A sub-process is specified by a triple {P} ∧ SW ⊗ [Q 1,…, Qn] where SW is the identifier of the sub-process, P is a set of precondition for SW , and Q1,…, Qn is a set of postcondition for SW . Definition 6. A distributed workflow specification consists of a set of CTR goals and CTR rules. A distributed workflow schema can be made up of a collection of tasks and sub-processs. Intuitively, we describe a workflow by the Definition 1~5: workflow(W ) ← task1 (W ) ⊗ [ task2 a (W ) | task2 b (W )] ⊗ subflow(W ) ⊗ task7 (W ) task1 (W ) ← : [task1a (W ) ⊗ task1b (W )]

664

Z. Feng and Y. Ye subflow(W ) ← condition(W ) ∧ [task3 (W ) ⊗ task4 (W )] subflow(W ) ← ¬condition(W ) ∧ [task5 (W ) ⊗ task6 (W )]

2.2 WS-BPEL Process Modeling Web service composition is usually expressed as a workflow process that describes the exchange of information between individual web services. The WS-BPEL is a language for describing the behavior of business processes based on web services. For the specification of a business process, WS-BPEL provides activities and distinguishes between basic and structured activities. Basic activities include activities such as sending (invoke), receiving (receive) requests and replies (reply). A structured activity defines a causal order on basic activities and can be nested in another structured activity itself. The structured activities are sequential execution (sequence), data-dependent branching (if), repeated execution (while), message-dependent branching (pick) and parallel execution (flow).

3 Mapping WS-BPEL into CRT We focus on the rich set of operators of WS-BPEL, including the basic activities and structured activities, which can be used to compose and orchestrate the workflow services. • Basic activity Definition 7. Invoke activity An activity is used to invoke the web service operations provided by partners. The following is activity.

In CTR, it will be represented by an atomic formula, marked INVOKE_ ACTIVITY. Definition 8. Reply Activity A activity is used to send a response to a request previously accepted through a activity. The following is activity.

In CTR, it will be represented by an atomic formula, marked REPLY_ACTIVITY. Definition 9. Receive activity A activity is used to receive requests in a WS-BPEL business process to provide services to its partners. The following is activity.

In CTR, it will be represented by an atomic formula, marked RECEIVE _ACTIVITY. • Structured activity

Distributed Workflow Service Composition Based on CTR Technology

665

Definition 10. Sequence Activity A activity is used to define activities that need to be performed in a sequential order. The following is activity.

name=”activityA”…/>… activity. …

It is represented by WHILE _ ACTIVITY = (cond ⊗ invoke ⊗ while) ∨ (¬cond ) in CTR. Definition 13. Pick Activity A activity is used to wait for the occurrence of one of a set of events and then perform an activity associated with the event. The following is < pick > activity. …

It is represented by PICK _ ACTIVITY = activityA ∨ activityB ∨ ... in CTR. Definition 14. Flow Activity A activity represents the implementation of sub-activities in parallel. The following is < flow > activity. …

It is represented by flow _ activity = activityA | ... | activityB in CTR.

4 Experimental Results In this section, we give an example of distributed workflow service composition and describe how it is described by WS-BPEL and translated to the CTR formulas. Fig.1 shows the corresponding application of the CTR technology to model the sample workflow. The EU$ process and US$ process handles cheque issues which enter the system via the task Begin and exit via the task End.

666

Z. Feng and Y. Ye

By the Definition 1-6, the control flow of workflow service orchestration in CTR is represented by: workflow(W ) ← ( Pay _ Req ^ Payment type choice(W )) ⊗ (( EU $ ^ Approval by Director(W )) ⊗(( Approve ^ Cheque for EU$ Bank(W )) | (Confirmation(W )))) ∨ (US $ ^ Cheque for US$ Bank(W ))) ⊗ (Get Signature from Director ) | (Confirmation(W )) .

where Payment type choice(W ) = request _ resource[ resource _ id ].START .assigned _ wait _ user[role _ id ].Payment _ Type _ Choice.release _ resource[resource _ id ].FINISH ,

(a) EU$

(b)US$

Fig. 1. Application of the CTR technology for the case of distributed workflow services composition Cheque for EU$ Bank(W ) = request _ resource[ resource _ id ].START .assigned _ resource(resource _ id ).

wait _ user[role _ id ].Check _ Cheque.release _ resource[resource _ id ].FINISH , Approval by Director(W ) = request _ resource[ resource _ id ].START .timer < begin _ time, end _ time > .

assigned _ resource( resource _ id ). Approval.release _ resource[ resource _ id ].FINISH , Confirmation(W ) = request _ resource[ resource _ id ].START .assigned _ resource( resource _ id ). Confirmation _ Cheque.release _ resource[resource _ id ].FINISH .

The following xml patch is a general definition of WS-BPEL for EU$ process by the Definition 7-15. US$ process is similar to EU$, and can be inferred from EU$.

Distributed Workflow Service Composition Based on CTR Technology

667



5 Conclusions In this paper, we are motivated by issues related to the composition of distributed workflow services by CRT technology. This paper firstly illustrates the concurrent transaction logic and WS-BPEL in summary, secondly gives the mapping rules from WS-BPEL to concurrent transaction logic. The proposed method works well for characterizing the behaviors and interactions of the distributed workflow service processes in terms of the semantics of CRT.

Acknowledgement The work has been supported by the Important National Science & Technology Specific Project, China (2009ZX01043-003-003), the National Natural Science Foundation of China under grant(No. 60703042), the Natural Science Foundation of Zhejiang Province, China (No. Y1080343), the Research and Application Plan of Commonweal Technology in Zhejiang Province (No. 2010C31027).

References 1. Scott, C.A., Ewa, D.B., Dan, G., et al.: Scaling up workflow-based applications. Journal of Computer and System Sciences 76, 428–446 (2010) 2. Jun, S., Georg, G., Yun, Y., et al.: Analysis of business process integration in Web service context. Future Generation Computer Systems 23, 283–294 (2007) 3. Chi, Y.L., Hsun, M.L.: A formal modeling platform for composing web services. Expert Systems with Applications 34, 1500–1507 (2008) 4. Wil, M.P., Kristian, B.L.: Translating unstructured workflow processes to readable BPEL: Theory and implementation. Information and Software Technology 50, 131–159 (2008) 5. Chun, O., Eric, V., Wil, M.P.: Formal semantics and analysis of control flow in WS-BPEL. Science of Computer Programming 67, 162–198 (2007) 6. Therani, M., Uttamsingh, N.: A declarative approach to composing web services in dynamic environments. Decision Support Systems 41, 325–357 (2006)

Offline Optimization of Plug-In Hybrid Electric Vehicle Energy Management Strategy Based on the Dynamic Programming Shichun Yang1, Ming Li1, Haigang Cui2, Yaoguang Cao1, Gang Wang1, and Qiang Lei1 1

School of Transportation Science & Engineering, Beihang Universtiy Beijing 100191, China [email protected] 2 College of Automotive Engineering, Jilin University Changchun 130022, Jilin, China

Abstract. By using dynamic programming (DP) which is a kind of global optimization algorithm, an energy management control strategy for a parallel PHEV on different charging depleting range (CDR) had been studied. The results show that motor-dominant control strategy should be applied to the PHEV when CDR is less than 55km, and engine-dominant control strategy should be used when CDR is more than 55km. With optimal control strategies from DP, the best economic performance can be obtained as CDR is 55km; PHEV average equivalence fuel consumption can be reduced to 2.9L/100km which is 63% lower than that of prototype vehicle. Keywords: Plug-in hybrid electric vehicle, Energy management strategy, Dynamic programming.

1 Introduction As the global oil crisis and environmental pollution become increasingly serious, improving energy use efficiency and reducing environmental pollution have become the primary task of the automotive industry development. Plug-in hybrid electric vehicle (PHEV) can not only significantly reduce human dependence on oil resources, but also be effective in reducing urban air pollution. So it has become one of most important technical means of vehicle energy saving and emission reduction[1,2]. Right now according to the customize method the hybrid vehicles energy management strategy mainly falls into three categories: Rule-based energy management strategy, based on intelligent control energy management and based on optimization algorithm energy management strategy. Different types of strategies have their advantages and disadvantages: rule-based energy management is simple and practical for real-time online control, but it main according to the designer’s experience and steady efficiency Map, so it cannot optimize the performance of the dynamic system; intelligent energy control strategy has strong robustness and good self-learning capability, very suitable for the design of the nonlinear control system, but the difficulty is how to describe the optimize problem of the energy management strategy under the framework of intelligent algorithm; instantaneous optimal control strategy can R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 668–674, 2011. © Springer-Verlag Berlin Heidelberg 2011

Offline Optimization of Plug-In Hybrid Electric Vehicle Energy Management Strategy

669

optimize the efficiency of the systems in real-time, but does not take into account of the changes in driving conditions influence on the vehicle performance; global optimization control strategy can optimize the efficiency of the system according to the whole driving conditions throughout the cycle, but the calculation is large and not suitable for the real-time control[3,4]. The energy management strategy obtain from the off-line global optimization algorithm is the best optimal control in theory, it both can be used as the Dynamic programming method (DP) is a global optimization method which can convert the complex decision problem into a series of sub-problems at different stages. It has little restrictions of the system state equation and objective cost function, the system model is a numerical model based on the experiment data and the iterative algorithm of the DP suits for solving by computer, so the DP is suitable for solving hybrid cars optimal control. DP as a multi-stage decision problem optimization method, it can divide the problem into a number of relevant stages according to the time and space. It must make a choice at each stage, this decision not only determines the effectiveness of this stage, and also the initial of the next stage, then determine the trend of the whole process (so called the dynamic programming). After the decision of each stage is determined, we get a decision sequence, known as the strategy. The so-called multi-stage decision problem is a strategy which makes the sum of every stage’s benefits maximum. Specific to the optimization problem of PHEV energy management, the selected cycle driving conditions should discretize reasonable, so that the entire cycle of the vehicle performance optimal control problem is converted into the decision problem of different time stages. Through following DP numerical solving method, we expect to obtain the PHEV optimal control under specific cycle conditions at last, and the performance which gained under optimal control is the best performance of the vehicle.

2 Results and Analysis of Off-Line Optimization Off-Line optimization must has a specified drive cycle, here we choose NEDC(New European Driving Cycle)as shown in Figure 1 and Table 1.By repeating the NEDC to achieve the different mileage of PHEV[5].

Fig. 1. Speed and corresponding gear ratio of NEDC

670

S. Yang et al. Table 1. Statistics of NEDC Parameters Speed Acceleration Decleration

Max Avg

Unit

Value

km/h

120.06 33.57

Max Avg Max Avg

1.07 m/s2

0.59 -1.43 -0.79

Time

s

1180

Distance

km

11.01

2.1 Off-Line Optimization Results As the shift schedule under the NEDC is fixed, the SOC of battery could represent the state of the vehicular electric power system of PHEV and then we can know the allocation of energy. Under the control of the different global optimal solutions for different mileages, the corresponding SOC sequences vary along with the mileage as shown in Figure 2.

Fig. 2. Results of Dynamic Programming

2.2 Analysis of the Results Unlike HEVs, to make full use of the energy charged from the electric grid, SOC of the PHEV under optimal control sequence declines with the increase of driving distance. Then according to the allocation of energy under the different global optimal solutions for different mileages, we can divide the first level of PHEV EMS into 3 cases, referred to the different driving modes[6,7]:

Offline Optimization of Plug-In Hybrid Electric Vehicle Energy Management Strategy

671

Motor-Dominant Driving Mode. When the mileage is short(less than NEDC×5, ie.55km)，SOC declines fastest，and the trajectory of SOC under the optimal control sequences for NEDC×5 completely cover those under the optimal control sequences for NEDC×1~NEDC×4. That means decline trend of SOC for each single cycle in NEDC×1~NEDC×5 is identical. The consistent decline trend of SOC indicates that driving distance has no effects on the allocation of energy in this driving mode. The reason is that the battery power from the grid at this time could meet the power demand for the driving in short distance, as thus PHEV could try to make full use of electric system to drive vehicle as much as possible. Figure 4 shows the allocation of torque under the optimal control sequence. It can be clearly seen that the motor sole provide the output power for PHEV, the engine provides load only when the request power exceeds motor power capability. Consequently, powertrain should adopt the driving mode which makes motor as the principal energy resource as shown in Figure 3 if we have known the vehicle mileage is less than 55 km beforehand.

Fig. 3. Torque allocation when mileage is less than 55 km

Engine-Dominant Driving Mode. When the vehicle mileage exceeds 55km, the decreasing rate of SOC of the optimal controlled power cell will slow down as the vehicle mileage increases, and SOC approaches the lower limit only at the end of the driving range(when the NEDC cycle ends, SOC will increase a bit by means of regenerative braking energy recovery).This is because: the energy charged from the energy grid is insufficient to meet the vehicle driving power demand for the whole trip, so both the start engine and the electrical power system are needed to drive the vehicle. As can be seen from figure 5, when the vehicle mileage exceeds NEDC× 11(121km), optimal controlled engine operating point of the best fitting curve has stabilized in the vicinity of the best efficiency performance curve. This illustrates that current PHEV energy management strategy should make the engine work in the best efficiency performance area under the circumstance that the vehicle’s power requirement is guaranteed. Whereas in the practical implementation of this strategy it can be

672

S. Yang et al.

described briefly as: as long as the power requirement exceeds the maximum power that the electrical power system can provide, or the start control logic of the engine is met, the engine will start; after the engine starts, it runs in the best efficiency condition (also on the best efficiency curve), with auxiliary drive power provided by the motor. This driving mode, which uses an engine running in best efficiency area as main power source of the vehicle, an electrical power-driven system as auxiliary power source, is called Engine-Dominant Driving Mode, as illustrated in figure 3. It is needed to be added that when the vehicle mileage exceeds 55km, the SOC only approaches the lower limit at the end of the driving range. The phenomenon that the variation trends of SOC bottom-up or bottom-maintain did not occur. This is because that the power cell’s open circuit voltage decreases and charging/discharging internal resistance increases when SOC is low, which leads to larger battery energy loss. So dynamic programming algorithm is bound to shorten the running time of the power cell under low SOC, which makes the lowest SOC only show up once at the end of the driving range.

Fig. 4. NEDC×12 engine operating points under optimal control

Fig. 5. NEDC×13 engine operating points under optimal control

Fig. 6. NEDC×14 engine operating points under optimal control

Fig. 7. NEDC×15 engine operating points under optimal control

Power-Balance Driving Mode. When the vehicle mileage is between 55km and 121km,the source of the dynamic programming optimal controlled powertrain’s total power is neither the motor nor the engine, and the two power source’s output ratio is between that of the two driving mode discussed above. This is called the PowerBalance Driving Mode of the PHEV. Under this driving mode of PHEV, it is more complicated to summarize the powertrain’s corresponding working mode switching rules and energy flow distribution strategy in various modes, and to acquire this probability statistics and multiple linear

Offline Optimization of Plug-In Hybrid Electric Vehicle Energy Management Strategy

673

regression and other methods are needed, which is the main problem to be solved in the next unit, the online energy management strategy design. From the above analysis of the dynamic programming offline global optimization computing result, we sorted out the relations between vehicle mileage and PHEV driving modes. That is to say, PHEV energy management strategy on the first level is determined. As is illustrated in figure 6, when vehicle mileage is less than s1 (55km), the vehicle will run in Motor-Dominant Driving Mode, whose main power source is electrical power driven system, and engine cannot be engaged in to drive the vehicle until vehicle demanding power exceeds the maximum output of motor. When vehicle mileage exceeds s2 (121km), the vehicle will run in Engine-Dominant Driving Mode, where engine runs along the best efficiency curve properly providing most power to drive the vehicle, whereas the electrical power driven system serves as an auxiliary drive source. When vehicle mileage is between s1 and s2, the vehicle will run in Power-Balance Driving Mode, where the output ratio of both power sources is between those of the two driving modes above, and dynamic programming optimal controlled powertrain energy flow distribution pattern is obscure which needs further research.

3 Conclusion For obtaining optimal controls, an offline global optimization method was brought. Dynamic programming was used to convert the optimal control problem of vehicle performance into step-by-step decision-making problem. By the rational discretization of reachable state set and admissible control set, we managed to get numerical solution of the global optimization algorithm. Through Matlab and C++ mixed programming code calculation, the optimal controls under different PHEV mileages were obtained. The optimization results showed that:

①. When the mileage is less than 55km (NEDC × 5), PHEV should be run under the motor-dominant driving mode; When the mileage is greater than 121km (NEDC × 11), the vehicle should be run under the engine-dominant driving mode; When the mileage is greater than 55km and less than 121km, the vehicles should be run under the power-balance driving mode of dynamic. ②. When the vehicle mileage is 55km, the energy obtained from the power grid could be used most efficiently. So, we got the best vehicle economic performance at this mileage. When the vehicle mileage is less than 165 kilometers, the vehicle average equivalent fuel consumption is 2.9 L/100km, which is improved by 63% compared with the prototype vehicle.

References 1. Chan, C.C.: The State of the Art of Electric and Hybrid Vehicles. Proceedings of the IEEE Digital Object Identifier 90(2), 247–275 (2002) 2. Nowell, G.C.: Vehicle Dynamometer for Hybrid Truck Development. SAE Paper 2002-013129 (2002)

674

S. Yang et al.

3. Simpson, A.: Cost-Benefit Analysis of Plug-In Hybrid Electric Vehicle Technology. In: 22nd International Battery, Hybrid and Fuel Cell Electric Vehicle Symposium (EVS-22), Yokohama, Japan, NREL/CP-540-40485 (October 2006) 4. Sasaki, M.: Hybrid Vehicle Technology - Current Status and Future Challenges - Effect of Hybrid System Factors on Vehicle Efficiency. Review of Automotive Engineering 26(2) (2005) 5. Gonder, J., Markel, T.: Energy Management Strategies for Plug-In Hybrid Electric Vehicles. SAE Paper 2007-01-0290 (2007) 6. California Air Resources Board: California Exhaust Emission Standards and Test Procedures for 2005 and Subsequent Model Zero-Emission Vehicles, and 2001 and Subsequent Model Hybrid Electric Vehicles, in the Passenger Car, Light-Duty Truck and Medium-Duty Vehicle Classes (March 26, 2004) (last amended December 19, 2003) 7. Markel, T., Simpson, A.: Plug-In Hybrid Electric Vehicle Energy Storage System Design. In: Advanced Automotive Battery Conference (AABC), Baltimore, MD (May 2006)

Development of Field Information Monitoring System Based on the Internet of Things* Ken Cai1,2, Xiaoying Liang3, and Keqiang Wang1 1 Information College, Zhongkai University of Agriculture and Engineering, Guangzhou 510225, China 2 School of Bioscience and Bioengineering, South China University of Technology, Guangzhou 510641, China 3 Guangdong Women's Polytechnic College, Guangzhou 511450, China [email protected]

Abstract. With the rapid development and wide application of electronics, communication and embedded system technologies, the global agriculture is changing from traditional agriculture that is to improve the production relying on the increase of labor, agricultural inputs to the new stage of modern agriculture with low yields, high efficiency, real-time and accuracy. On the other hand the research and development of the Internet of Things, which is an information network to connect objects, with the full capacity to perceive objects, and having the capabilities of reliable transmission and intelligence processing for information, allows us to obtain real-time information of anything. The application of the Internet of Things in field information online monitoring is an effective solution for present wired sensor monitoring system, which has much more disadvantages, such as high cost, the problems of laying lines and so on. In this paper, a novel field information monitoring system based on the Internet of Things is proposed. It can satisfy the requirements of multi-point measurement, mobility, convenience in the field information monitoring process. The whole structure of system is given and the key designs of system design are described in the hardware and software aspect. The studies have expanded current field information measurement methods and strengthen the application of the Internet of Things. Keywords: The Internet of Things, field information, sensor nodes, hardware, software.

1 Introduction Over the last decade Internet has been a major driver of global information and media sharing. With the advent of low cost wireless broadband connectivity, it is poised to emerge as an “Internet of Things” where the web will provide a medium for physical world objects to participate in interaction. The Internet of Things is a kind of *

This research was supported by GuangDong Provincial Science and Technology Planning Project of China under grant 2010B020315028.

R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 675–680, 2011. © Springer-Verlag Berlin Heidelberg 2011

676

K. Cai, X. Liang, and K. Wang

expansion for Ubiquitous Computing (ubiquitous comes from the Latin word and means existing in any place). It is described as a self-configuring wireless network of sensors whose purpose would be to interconnect all things [1]. A computer scientist Mark Weiser in the Xerox laboratory first proposed the concept of ubiquitous computing in 1991 [2], which describes a worldwide network of intercommunicating devices. It integrates the Pervasive Computing [3], Ambient Intelligence [4], Ubiquitous Network [5]. Although these concepts are different from the Internet of Things, their inherent ideas are quite consistent. At this point (IOT), it is easy to imagine that nearly all home appliances but also furniture, clothes, vehicles, roads and smart materials, and more, are readable, recognizable, locatable, addressable and/or controllable via the Internet. This will provide the basis for many new applications, such as Soil and plant monitoring, Monitoring of food supply chain, Monitoring of animals, Forest monitoring, Tourism support, Homeland security, Pollution monitoring. The Internet of Things represents the future development trend of computing and communications technology. Modern agriculture is the formation of the industrial revolution in agriculture, which enables farmers to utilize new innovations, research and scientific advancements to produce safe, sustainable and affordable food. It has following main features: the degree of market matures, commonly used industrial equipment, high technology is widely used, industrial system and keep improving the ecological environment seriously. In China, the agriculture is in the period of transition from traditional agriculture to modern agriculture. At the same time, the research of farmland data acquisition and processing, as a very important aspect for Modern agriculture, is at the initial stage in China. This paper mainly studied the technology of field information acquisition and process based on techniques of the Internet of Things, sensors, embedded system.

2 System Architecture The framework of field information acquisition system based on the internet of Things is shown in Figure 1. The entire system can be divided into three parts: wireless sensor nodes, sink nodes (intelligent nodes) and control centers. Field information (such as temperature, humidity, etc.) is collected by the wireless sensor nodes transmitted to the sink nodes (intelligent nodes) using the ZigBee communication network. The sink nodes (intelligent nodes) send the information through Network Bridge to the control center, and the information are displayed and stored. The sink nodes (intelligent nodes) in this system act as gateway, it is the protocol conversion used to transform a data package in ZigBee protocol to TCP/IP protocol before transmitting and a data package in TCP/IP protocol to ZigBee protocol. As the control center accesses to Internet, the operators can view and process the data of control center through the Internet-enabled computer or devices (e.g. PDA, smart phone, etc.).

Development of Field Information Monitoring System Based on the Internet of Things

677

Fig. 1. System architecture

3 System Design 3.1 Hardware Design of Wireless Intelligent Nodes The Internet of Things is a wireless sensor network where many sensor nodes and intelligent nodes are arranged in certain areas in a random way, which are the most important and basic components of the Internet of Things. For intelligent nodes, it plays a role on sinking acted as gateway. The hardware part of the intelligent nodes consists of four hardware components of sensor modules, processor modules, wireless communication modules and power supply modules. The sensor module is responsible for collecting information in sensor field. The processor module is responsible for coordinating the work of various parts of nodes. The wireless communication module can be divided into two parts. One is responsible for wireless communication with other sensor nodes. Another part is responsible for network communication with the monitor center. The power technical module is responsible for providing the power required for sensor nodes. The Figure 2 shows the structure.

Fig. 2. Block diagram of intelligent nodes

678

K. Cai, X. Liang, and K. Wang

3.2 Software Design of Wireless Intelligent Node Intelligent node is mainly responsible for running the operating system, Web Server and application procedures. It receives the user command, sends the request to the appropriate monitoring points, process data fed back by monitoring points and then give feedback to the user through the network. As the user request is sent to the wireless module by serial port, so intelligent nodes first finish the configuration settings of the serial port and make sure to open serial device. And then it sent the data in accordance with the format prescribed by the communication protocol of monitoring points and waited for data fed back by monitoring. Figure 3 shows software design flow of the intelligent node.

Fig. 3. Software design flow of the intelligent node

3.3 Design of Wireless Sensor Node The sensor nodes are responsible for collecting information in sensor field and implement the timely transmission of data. In our project, the CC2430 is choosing as for the CPU of sensor nodes. CC2430 is a true resolution to System-on-Chip (SoC), which can improve performance and meet the requirements of IEEE 802.15.4 and ZigBee applications on low costs and low power consumption. It has several key features: 2.4 GHz IEEE 802.15.4 compliant RF transceiver (industry leading CC2420 radio core), Low current consumption (RX: 27 mA, TX: 27mA, microcontroller running at 32 MHz), High performance and low power 8051 microcontroller core with 32, 64 or 128 KB in-system programmable flash, Two powerful USARTs with support for several serial protocols. After power up, CC2430 carries out the system initialization. And then it determines whether to be configured as a network coordinator or terminal nodes according to the real situation. After that, sensor nodes can collect and send data. Figure 4 shows ZigBee sensor node flow chart and circuit of CC2430.

Development of Field Information Monitoring System Based on the Internet of Things

679

E1

C1 U1

1

15pF

Y1 XTAL

2

C3

15pF

VCC

C7

10K R2

1

27pF

Y2 XTAL

RST 2

C8

C9 0.1uF

33pF

C2 0.1uF

VCC 44 43 21 19 RST

10

P0_0 P0_1 P0_2 P0_3 P0_4 P0_5 P0_6 P0_7 P1_0 P1_1 P1_2 P1_3 P1_4 P1_5 P1_6 P1_7

11 12 13 14 15 16 17 18 9 8 6 5 4 3 2 1

R4

43K

R5

56K

26 22 42

P2_3/XOSC_Q1 P2_4/XOSC_Q2 XOSC_Q1 XOSC_Q2 RESET_N P0_0 P0_1 P0_2 P0_3 P0_4 P0_5 P0_6 P0_7 P1_0 P1_1 P1_2 P1_3 P1_4 P1_5 P1_6 P1_7 RBIAS2 RBIAS1 DCOUPL

AVDD_DREG AVDD_RREG AVDD_SOC DVDD DVDD RF_N TXRX_SWITCH RF_P P2_2 P2_1 P2_0 RREG_OUT AVDD_IF1 AVDD_CHP VCO_GUARD AVDD_VCO AVDD_PRE AVDD_RF1 AVDD_SW AVDD_RF2 AVDD_IF2 AVDD_ADC DVDD_ADC AVDD_DGUARD

41 23 20 7 47

C4 0.01uF

C5 0.01uF

24 25 27 28 29 30 31 35 36 37 38 39 40

L1 1.8nH R1

34 33 32 45 46 48

C6 0.1uF

L2 Inductor 6.8nH

P2_2 P2_1

C10 0.01uF

C11 0.01uF

L3 Inductor 22nH

C12 0.1uF

R3

C13 0.1uF

C14 1uF CC2430

(a)

(b)

Fig. 4. (a) ZigBee sensor node flow chart. (b) Schematic circuit diagram of CC2430.

3.4 Design of Monitoring Center Monitoring center software mainly includes video display modules, parameter waveform display, alarm module, file service module and the network communication module. The user interface and related function modules of the system is used by QT to program. The video display module of the system is achieved by using the method of refreshing images. The video collected by the terminal to send to monitor server are the compressed JPEG images. According to the physical phenomena of human eye's vision persistence, at least playing for more than 25 frames per second, the static map can display animation. So the system set a timer with an interval of 40ms and the timer triggers a call each time to invoke image display program, refreshing 25 images per second. Parameter waveform display and alarm module also uses QT's QPainter drawing tools. The core function of QPainter is drawing and it also has very rich associated drawing functions. The system mainly uses three functions of drawLine (), drawRect (), drawText () to complete the work of drawing parameters. File service module is mainly responsible for storing related image information and names corresponding files by time for the convenience of inquiries. At the same time, it is also responsible for recording environmental parameters in the file, which are used as a historical record for later inspection and analysis. The mainly used functions are QFile, QfileInfo and QString class in QT. Network communication module is through the socket to conduct network programming. Stream socket is used to provide connection-oriented, reliable data transfer service. The service will ensure that the data can be transmitted free from errors and non-repeatedly and received according to the order. Due to the use of the transmission control protocol, namely TCP (The Transmission Control Protocol) protocol, so the Stream socket can achieve reliable data services.

680

K. Cai, X. Liang, and K. Wang

4 Conclusion The real-time monitoring system of field information can provide the scientific basis for modern agriculture management, which is important for improving crop yield and quality. In the other hand, the Internet of Things is a hot topic today, which will lead the development of the rise of new industries. In this paper, a farmland information monitoring and warning system based on the Internet of Things is proposed in order to overcome the limitations of the existing field monitoring systems by using cable for data transmission. The system can monitor real-time temperature, humidity and much other environmental information, and sensor node has low power consumption and high performance, which can satisfy the requirements of mobility, portability, multipoint-to-multipoint, multipoint-to-point, convenience in the field information monitoring process. In future, some compression algorithm will be used to reduce the data bandwidth, and the new wireless module will be used to improve the transmission speed.

References 1. Conner, M.: Sensors empower the “Internet of Things”, pp. 32—38 (2010) 2. Weiser, M.: The Computer for the 21st Century. Sci mer (September 1991) 3. Satyanarayanan, M.: Pervasive Computing: Vision and Challenges. IEEE Personal Communications, 10–17 (August 2001) 4. Mulder, E.F., Kothare, M.V.: Title of paper. In: Proc. Amer. Control Conf., Chicago, IL, vol. 5, pp. 3239–3243 (June 2000) 5. Murakami, T.: Establishing the Ubiquitous Network Environment in Japan, Nomura Research Institute Papers, No. 66 (July 2003)

System Design of Real Time Vehicle Type Recognition Based on Video for Windows (AVI) Files Wei Zhan1,2 and Zhiqing Luo2 1 College of Computer Science, Yangtze University, Jingzhou, HuBei, China 2 School of Computer Science, China University of Geosciences, Wuhan, HuBei, China [email protected]

Abstract. In this system, with technology of motion detection, the data frames include vehicle digital image can be detected automatically from a Video for Windows (AVI) File, at the same time, vehicle type will be recognized and displayed automatically. The system’s process consists of five steps: Read the AVI file and decompose it into digital image frames; Motion detection; Vehicle digital image processing; Counting number of black pixels included in vehicle body contour and project on car image; Module of vehicle type classification. In particular, algorithm of vehicle recognition through counting number of black pixels included in vehicle body contour is one innovation algorithm. Experiment on actual AVI files shows: the system design is simple and effective. Keywords: Vehicle Type Recognition; Background Image Subtraction; Anti-color Processing on image; Threshold Segmentation.

1 Introduction Automatic and real time vehicle type recognition system[1] is an important part of Intelligent Transportation System[2][3] (ITS),in china, system research on it starts behind abroad, in all weather condition, recognition accuracy of domestic vehicle recognition system is not satisfied, the primary method[4][5][6] of vehicle type recognition are: radio wave or infrared contour scanning, radar detection, vehicle weight, annular coil and laser sensor measurement. 1.1

Video Trigger

With technology of digital image processing and recognition, there are more and more research on vehicle type recognition using video. Because a wide range application and rich information of image detection, it can be use in road traffic monitoring, vehicle type classification and recognition, automatic license plate recognition, automatic highway toll, intelligent navigation, so the vehicle type recognition based on video is a hot research direction. R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 681–686, 2011. © Springer-Verlag Berlin Heidelberg 2011

682

W. Zhan and Z. Luo

2 Overall System Architecture The system consists of four main modules: Reading the AVI file and decomposing AVI file into digital image frames; Detection on moving vehicles; Processing vehicle digital image and module of vehicle type classification. First step: the system reads an AVI file generated by a certain actual HuBei Province toll station’s video capture card; second: this AVI file will be decomposed into many digital image frames; third: through module of vehicle movement detection the moving vehicle image will be find; then through module of processing digital image the length and width and the number of black pixels included in vehicle body contour will be calculated, these parameters mentioned above can be used as foundation of distinguishing vehicle type.

3 Concrete System Design 3.1 Algorithm of Decomposing AVI File into Frames Because the AVI file is with characteristic of alternative stacking frames, so the Algorithm of decomposing AVI file into frames can be designed. 3.2 Selecting a Monitoring Area A rectangular image region is selected as test area in order to detect weather a car is passing by.

Input specified image frame

Background image subtraction

Threshold segmentation

3.3 Algorithm of Judging Weather a Car Goes in or Out the Monitor Area In the case of the induction coil is not used, for judging weather a car is existing in the monitoring area, continuously detect whether the background changes is needed. One method called image subtraction is used in this system. When the system is initialized every time run, one background image and one threshold value must be set, if change of pixel number in the monitoring area exceeds the threshold value, the system thinks there is a moving car in the area, on the contrary, there is no car. If a car is found in the monitoring area, the current car image frame will be loaded into memory for further image process. In addition, the background image must be refresh constantly to ensure the accuracy of recognition.

Inverse image color

Count the number of black pixels included in vehicle body contour

Projection detection on car image

Output vehicle type Flow. 1 Flow of processing vehicle digital image

System Design of Real Time Vehicle Type Recognition

683

3.4 Flow of Processing Vehicle Digital Image As shown Flow.1. It is flow chart of processing vehicle digital image[1][6]. 3.4.1 Background Image Subtraction In Fig. 1, Fig. 2 and Fig. 3, there are respectively image need to be processed, background image, and image after subtracting.

Fig. 1. Pending image

Fig. 2. Background template

Fig. 3. Image after subtraction with background

3.4.2 Threshold Segmentation As shown in Fig.4, giving a threshold value, with algorithm of threshold segmentation, contour of the car is filled with white pixels.

Fig.4. Image after Threshold Segmentation

Fig.5. Image after Inverse color

3.4.3 Inverse Image Color Experiment shows inversing image color can get better recognition result, after inversing, all white pixels in the contour of the car will be instead of black pixels, as shown in Fig.5. 3.4.4 Counting Number of Black Pixels Included in Vehicle Body Contour The car bigger is, the more black pixels are included in vehicle body contour. 3.4.5 Projection Detection on Car Image Using projection algorithm[7], width and length of vehicle as auxiliary Parameters can be measured. As shown in Fig.6 and Fig.7.)

684

W. Zhan and Z. Luo

Width

Height

Fig. 6. Vertical projection

Fig. 7. Horizontal projection

3.5 Module of Vehicle Type Classification Extraction and classification algorithm[8] normally used are:) vehicle recognition based on neural network, models based on support vector machine identification, genetic algorithm based on wavelet decomposition and genetic algorithm (GA). Algorithm mentioned above are complex, based on video and image, a simple and effective algorithm of extracting vehicle type parameters was designed, that is by counting number of black pixels included in vehicle body contour. The bigger car is, the more number of black pixels are in the car body.

4 Result of System Test After test with a large number of AVI files shows: the system is simple and effective, the following is the test result. 4.1 Test of Non-motor Vehicle As shown in Fig.8,when one person goes into monitor area, system has not recognized this person as a car, system flag of vehicle in and out has not be changed ,which

Monitoring Area

Background Image

Pixel Change Result in Monitoring Area

Recognition result: No car Fig. 8. Test on Non-motor

System Design of Real Time Vehicle Type Recognition

685

shows that setting of the rectangular image detection region in 3.2 and the threshold value in 3.3 are reasonable and effective. 4.2 System Test on Cars Fleet As shown in Fig. 9, if there is a car fleet, system has not response timely, cars in fleet has not be separated timely, which leads to recognition error. More effective algorithm need be design to solve this difficult recognition problem, at the same time, it is a key question in the future research.

Monitoring Area

Background Iage

Pixel Change Result

Current Car Type: Six Charge: 45 yuan

Car in

Fig. 9. System test on car fleet

4.3 System Test on Cars in Different Color During testing system, often, recognition result on a dark color car is less than that of a light color car although they are the same vehicle type, suck as shown in Fig.10, the left is a white Elysee, the right is a black Passat, the right result is less than the left, but actually, result should be the same one. For this problem, solution will be proposed in the future.

(a) Recognition Result: Type 2 Charge: 10 Yuan The result is accurate.

(b) Recognition Result: Type 1,

Charge: 5 Yuan The result is small

Fig. 10. System test on cars in different color

686

W. Zhan and Z. Luo

5 Conclusion Technology of vehicle type recognition can be widely used to automatic statistics of traffic and toll. Test shows: system design is simple and effective, especially on algorithm of counting the number of black pixels included in vehicle body contour, which is an innovative approach .But on the problem of car fleet and different color cars should be lucubrated for higher recognition accuracy.

References 1. Zhan, W.: Research of Vehicle Recognition System for the Road Toll Station Based on AVI Video Flow. China University of Geosciences (2006) 2. Tian, B.: Research on Automatic Vehicle Recognition Technology in Intelligent Transportation system. pp. 23–26. XiDian University (2008) 3. Xia, W.: Video-Based Vehicle Classification Method Research. 1–2, 7, 16, 24–29. Huazhong University of Science&Technology (2007) 4. Ji, C.: Vehicle Type Recognition Based on Video Sequences. Journal of Liaoning University of Technology (Natural Science Edition) 30, 5–7 (2006) 5. Cao, Z.: Vehicle Detection and Classification Based on Video Sequence, pp. 20–46. ZheJiang University (2004) 6. Cao, Z., Tang, H.: Vehicle Type Recognition in Video. Computer Engineering and Applications 24, 226–228 (2004) 7. Xiong, S.: Research of Automobile classifying method based on Inductive Loop, pp. 2–5. ChangSha University of Science and Technology (2009) 8. Cao, Z., Tang, H.: Vehicle Detection, Tracking and Classification in Video Image Sequences. Journal of Computer Applications 3, 7–9 (2004)

Subjective Assessment of Women’s Pants’ Fit Analysis Using Live and 3D Models Linli Zhang1,2, Weiyuan Zhang1, and Hong Xiao2 2

1 Fashion institute, Donghua University 200051 Shanghai, China Xi An Polytech University, Apparel & Art Design College, 710048, Xian Shaanxi, China [email protected]

Abstract. This paper investigated the validity of three-dimensional scans as a tool for visual fit analysis. We used traditional live models and three dimensional scans to compare the fit assess on women pant. Results from different areas of pant showed different levels of accuracy. In the waistband, upper hip and hip area, the results were more accurate while the crotch and side seam were not. Keywords: Three-dimensional scan data, fit, subjective analysis.

1 Introduction Fit refers to how well the garment conforms to the three-dimensional human body. However, the principles of fit are vary from time to time, and depend on the fashion culture, industrial norm and individual perception of fit. Ashdown (2009) noted several factors impacting on decisions to understand fit within the research work [1]. LaBat and Delong suggested two external influences (social message of the ideal body and fashion figure in the industry) impacting and two personal influences (body cathexis and physical dimensional fit of clothing) which impacting on the consumer’s satisfaction with clothing fit [2]. To assess if a garment fits or not, it can be used by using some approaches, such as live models and dress forms. Live models have some advantages when used for assessing clothing fit. Because they are the real human body shape and have real movements, as well as have their comments on the clothing sensible. However, they tend to be made personal conclusion based on subjective feeling. Their judgements would be varied from this model to other model. Also, live models are always expensive. Dress forms are static and convenient to use many times, however, they tend to be impacted by personal assessment of tension. But dress forms are not convenient when the fit assessment proceeding by the photo or the video. Fit researchers have used scanned images and expert judges to evaluate the fit of pants (Ashdown et al., 2004) [3]. They concluded that there had potentials of substituting 3D scans for the live fit analysis process in research and industry for 1) recordings can be rotated and enlarged, 2) creating database of scans of a variety of body shapes and sizes wearing a single size, 3) scanning garments on fit models in multiple poses to evaluate R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 687–692, 2011. © Springer-Verlag Berlin Heidelberg 2011

688

L. Zhang, W. Zhang, and H. Xiao

garment/body relationships during natural movements and 4) holding virtual expert panels where panelists can access the fit session at any location. By using body scan data to gain insight into body/apparel relationships, Ashdown (2005) developed protocols to improve fit of existing sizing systems [4]. Ashdown (2006) developed methods to visually analyze fit using 3D scans of clothed subjects. They assessed the reliability of fit ratings at different body areas and determined that two judges are needed for reliable results overall. Fit assessment ratings of women’s pants for most body areas were reliable if fit parameters and the instrument scale were established and clearly defined for the judges [5]. The developing digital dress forms cannot solve all the clothing fit trouble yet. Today because of distributed designing centers and factories around the world, the correct fit suggestions cannot be communicated effectively. Shopping online business also caused custom confused with the fit try-on trouble. The real persons with fit and unfit questions have obsession to make definite shopping decision. By the way of live prohibit questionnaires, Zafu even doesn’t ask customer’s body dimension data to help 94% of their customer chose the perfect style and brand fit jeans [9]. Kohn and Ashdown (1998) found a positive correlation between fit analysis using a video image and traditional fit analysis using a live model [10]. Schofield, Ashdown, Hethorn, LaBat, and Salusso (2006) also used a videotape method to record and analyze pants fit for women aged 55 and older. The protocol specified wearer responses to the pants and an evaluation of the fit by an on-site expert. An additional expert fit evaluation was conducted later by a panel of geographically dispersed experts who viewed videotapes of each participant in her test pants [11]. This study investigated the possibility of subjective assessment method of women’s pants fit by using three-dimensional scans. Compare the assessment result of threedimensional scans with the assessment result with live models in order to figure out the feasibility of using three-dimensional scans. Four objectives of this study are to figure out. Compare fit problems found from three-dimension scans with live models. Assess the effectiveness of visual data from three-dimension scans in correction values. Assess the certain degree of fit from three-dimension scans. Get results to improve the women’s pants’ fit assessment.

2 Experimental Methods 2.1 Test Samples and Sizes Women pants were used as the experiment sample. The pant sloper was a trouser style with two front darts, two back darts, a center back zipper, and waistband. This style was chosen because of its ubiquity and potential for use as base pattern for other styles. The pants for test were made from medium weight cotton muslin fabric in a raw white color. Three sequential sizes pants, which according to China garment standard GB/T1335-97, were made for test. The sizes used for testing pants are shown in table 1.

Subjective Assessment of Women’s Pants’ Fit Analysis Using Live and 3D Models

689

Table 1. The sizes used for testing pants according to China garment size standard GB/T 1335-97 Number 1 2 3

Size (cm) 160/66A 165/70A 170/74A

We also prepared one kind of next-to-skin underwear for the subjects. The size and materials used for test pants are shown in table 2. Table 2. The samples of next-to-skin underwear Sample ID 1 2 3

Knitted structure Power-net Power-net Power-net

Main fabric content (%) Garment size Ny 87% U13% 160/66A Ny 87% U13% 165/70A Ny 87% U13% 170/74A

2.2 Subjects The subjects of this study were 18 healthy females chosen from 30 to 40 years old and these females height from 160cm to 165cm. their BMI (body mass index) should be within 18.5-24.9. 2.3 Experimental Methods First, we took three-dimensional scan data collection. The [TC]² three-dimension body scanner to capture the subject data. The subject was scanned two times in the standing position, once in minimal underwear clothing and second in the best-fitted test pants. The body data was captured and landmarks were acquired. Second, we fitted assessment traditionally using live models. Research by Ashdown, Loker, and Rucker (2006) showed that two fitting experts are enough to make a reliable fit analysis [5]. Two experienced judges figured out the unfitted area following fit criterion of Patty Brown’s (1998) five elements [12] and marked the correct value criterion. Third, we pretreated the three-dimensional scan analysis data with DEEP EXPLORATION, ULTRAEDIT and GEOMAGIC software tools. Three-dimensional scan files were merged, aligned, and cleaned before sending them to fit judges for threedimensional scan fit analyses. Last, we did three-dimensional scan fit analysis. The judges worked on the software Geomagic after they had been trained using software tool.

3 Results Two judges analyzed and assessed the women pants’ fit. They had 36 total results on estimating fit. Fig. 1 shows a comparison between the results from the live model fit

690

L. Zhang, W. Zhang, and H. Xiao

analyses and the three-dimensional scan analyses in each location. The area with the highest percentage of identical results between the two methods was the waistband

%

(88.9 ).There was less agreement in the back crotch (36.1%), front crotch seam (38.9%), back thigh (41.7%), left-side seam (41.7%), right-side seam (47.2%), and back crotch seam (47.2%).

Fig. 1. Percentage of agreement of live models and three-dimensional scans fit analysis

Fig. 2. Mean of confidence scores in rating misfit by body location

Subjective Assessment of Women’s Pants’ Fit Analysis Using Live and 3D Models

691

The mean of confidence scores in rating misfit was calculated to analyze the relationship between the body location and the confidence level of the judges. Fig. 2 show the level of confidence on the 5-points scale with lower confidence (1), low confidence( 2), normal (3) and very confident (5). The back crotch was the area of the lowest confidence because the crotch cloud point sometime lost when scanned. The waistband, darts, full hip were the areas of very confidence. The fit correction differences were categorized into low (0–0.5cm), medium (0.6– 1cm), high (1.1-1.5cm). When the fit assessment was in the agreement of misfit, the differences in alteration were compared in the table 3. The majority of alteration amount under 1.5cm were found in the fit correction. The alteration of front thigh and back thigh were high while the inseams, left-side seam, the right-side seam, front crotch seam, back crotch seam, darts and front crotch were low. Table 3. Differences in Alteration Amounts between Live Model Fit Analyses and ThreeDimensional Scan Location

Mean of differences(cm)

Mean confidence in estimating amount

Inseams

0.34

Left-side seam

0.12

3.4

Right-side seam

0.13

3.4

Front crotch seam

0.10

3.6

Back crotch seam

0.12

3.4

Front darts

0.05

4.2

Back darts

0.05

4.2

Waistband

0.70

4.4

Abdomen

0.64

3.9

Front crotch

0.05

3.8

Front full hip

0.90

4.2

Front thigh

1.24

4.0

Back crotch

0.06

3.4

Back upper hip

0.89

4.2

Back full hip

1.21

4.3

Back thigh

1.27

4.4

3.6

4 Conclusions Even though there was no perfect agreement in estimating the area of crotch, side seams, the high level of agreement of waistband, hip and upper hip indicates potential for using three-dimensional scans as an assist tool for these body locations. There are limitations of 3D body scanning on assess fit for missing areas, body posture and

692

L. Zhang, W. Zhang, and H. Xiao

movement, surface texture and accuracy. The future study should be improved in the fields of data process, data display and navigation interface. That would be good in utilizing three-dimensional scans data in the clothing industry.

References 1. Ashdown, S.: Introduction to Sizing and Fit Research. In: The Fit Symposium, Clemson Appearance Research, South Carolina, Clemson (2000) 2. LaBat, K.L., Delong, M.R.: Body Cathexis and Satisfaction with Fit of Apparel. Cloth. Text. Res J. 8(2), 42–48 (Winter 1990) 3. Ashdown, S., Loker, S., Schoenfelder, K., Lyman-Clarke, L.: Using 3D Scans for Fit Analysis. JTATM 4(1) (Summer 2004) 4. Ashdown, S., Loker, S.: Improved Apparel Sizing: Fit and Anthropometric 3-D Scan Data. NTC Project: S04-CR01 (June 2005) 5. Ashdown, S., Loker, S.: Improved Apparel Sizing: Fit and Anthropometric 3-D Scan Data. NTC Project: S04-CR01 (June 2006) 6. Casini, G., Pieroni N., Quattocolo S.: Development of a low cost body scanner for garment construction. In: 12th ADM Int. Conf., Rimini, Italy, September 5-7, pp. A41–A48 (2001) 7. Alvanon: Fit Conference for the apparel industry, http://www.alvanon.com 8. Song, H.K., Ashdown, S.: An Exploratory Study of the Validity of Visual Fit Assessment From Three-Dimensional Scans. Cloth. Text. Res. J. 8(4), 263–278 (2010) 9. Zafu: http://www.zafu.com/ 10. Kohn, I.L., Ashdown, S.P.: Using Video Capture and Image Analysis to Quantify Apparel Fit. Text. Res. J. 68(1), 17–26 (1998) 11. Schofield, N.A., Ashdown, S.P., Hethorn, J., LaBat, K., Salusso, C.J.: Improving Pant Fit for Women 55 and Older through an Exploration of Two Pant Shapes 55 and Older Through an Exploration of Two Pant Shapes. Cloth. and Text. Res. J. 24(2), 147–160 (2006) 12. Brown, P., Rice, J.: Ready-to-wear Apparel Analysis. Prince Hall Inc., Upper Saddle River (1998)

A Geographic Location-Based Security Mechanism for Intelligent Vehicular Networks Gongjun Yan1, Jingli Lin2, Danda B. Rawat3, and Weiming Yang4 1

Indiana University Kokomo, Kokomo, IN 46904 USA [email protected] 2 Xihua University, Chengdu China 610039 [email protected] 3 Old Dominion University, Norfolk, VA 23529 USA [email protected] 4 Chongqing Technology & Business Univ., Chongqing 400067, China [email protected]

Abstract. In Intelligent Vehicular Networks, featured as car-to-car and car-toinfrastructure wireless communication, most applications need important location information or credential information. We address a location-based encryption method that not only ensures messages confidentiality but also authenticates identity and location of communication peers. The authentication of location means that a message can only decrypted by the receiver which is “physically” present inside a decryption region specified by locations, and is moving at a specific speed, acceleration and at a specific time period. A secret key is generated by converting location, time and mobility information (such as speed and acceleration) into a unique number. The determination of the decryption region is addressed in two steps: predicting and updating. The proposed method evaluated by simulations is efficient and secure. Keywords: Geoencryption, GPS, Location based encryption, intelligent vehicular networks, location authentication, vehicular ad hoc networks.

1 Introduction In Intelligent Vehicular Networks or vehicular adhoc networks (VANET), vehicles are equipped with wireless transceivers so that they can communicate with other vehicles and roadside infrastructure. VANET promises to revolutionize the way we drive. Various car manufacturers, government agencies and standardization bodies have spawned national and international consortia devoted exclusively to VANETs. Examples include the Car-2-Car Communication Consortium, the Vehicle Safety Communications Consortium, and Honda’s Advanced Safety Vehicle Program, among others. In addition, third party providers have already started to offer Internet access, distributed gaming, as well as other features of mobile entertainment. There are ubiquitous safe and secure applications in VANETs. Most, if not all, applications in VANETs need location information to function appropriately. For safety R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 693–698, 2011. © Springer-Verlag Berlin Heidelberg 2011

694

G. Yan et al.

applications, vehicles need to know each other’s location to avoid accidents. Insurance justifications of car accidents rely on locations. For public service applications, vehicles must be aware of the location of emergency vehicles to move aside for them. For entertainment applications, both the location of vehicles and the location of resources are needed to provide a high quality service. In VANETs, most applications require protection of location information. In some environments, for example in the battlefield, the locations of vehicles need to be extremely secure. Therefore, not only the content is required to be encrypted but also the place where the encrypted message can be decrypted is required to be specified. The motivation of this paper is to address a location-based encryption method that not only ensures message confidentiality, but also authenticates the identity and location of communicating peers. In our approach, a special region (decryption region) is specified to the message receiver. The receiver must physically present in the decryption region to decrypt the received message. To achieve this idea, the receiver’s location information is converted part of a key. We design an algorithm dynamicGeoLock to convert the location information into the key. Our main contributions include: 1) Designing an algorithm to convert a location into a key. 2) Extend the decryption area from square (proposed in geolock [1]) to any shape; 3) Improving location error tolerance.

2 State of the Art Location-based encryption method is proposed by Denning et al. [2], [3] that limits the area inside which the intended recipient can decrypt messages. This geoencryption integrates geographic and mobility information (such as position, time, speed, etc) into the encryption and decryption processes. Denning proposes GeoLock that is computed with the recipient’s position, velocity, and time. Positions are signed and mapped to GeoLock which is like a value of a grid composed by xy-coordinates of positions. However, the mapping function in practice is not specified in Denning’s work. If the mapping function is pre-installed tables which are shown as example of the mapping function in [3]. It is extremely hard to ensure the synchronization of the key mapping grid in vehicular networks. Denning’s geo-encryption model did not include details of an implementation of mobility support, so Al-Fuqaha et al. [4] proposed a model to provide for mobility when using GPS-based encryption. The AlFuqaha’s method used the same mapping table pre-installed on nodes. Yan et al. [1] proposed a new geo-encryption algorithm which generates a unique location key. The key is used to encrypt and decrypt information in a decryption region where cars must physically present. But there is a restriction of the decryption region: it must be a square shape. In this paper, we design a new algorithm that can work with any shapes of decryption region. In this paper, we design the key composition/recovery in detail. No mapping tables are needed. Positions can be mapped to a lock on the fly. Since nodes in VANETs are with high dynamics, the decryption region is designed as a series of fixed-size squares in this paper. The area of the square is large enough to cover the error of the decryption region prediction. Moreover, we incorporate the prediction error by using location prediction deviation. We trade

A Geographic Location-Based Security Mechanism for Intelligent Vehicular Networks

695

freedom of the size and the shape of the decryption region in order to obtain the feasibility and the accuracy of decryption region prediction.

3 Our Location Based Geo-Encryption 3.1 An Overview Our technique involves a security key handshake stage and a message exchange stage, as shown in Figure 2(a). In the key handshake stage, the client and the server negotiate a shared symmetric key. The client generates two random numbers as keys Key_S and Key_C. Key S is used to encrypt a message composed of the aggregated location message and Key C. This encrypted message is E{Req}. The client generates a GeoLock based on the location of the server. This value is XOR-ed with Key S and then encrypted using the server’s public key Key_E to produce the ciphertext E{Key}. Both E{Req} and E{Key} are transmitted to the server through the wireless channel. When the server receives E{Key}, it is decrypted using the server’s private key Key_D to produce the XOR of the GeoLock and Key_S. The GeoLock generated from the GPS location of the server is used to recover the secret key Key_S. Then, Key_S is used to decrypt E{Req} to obtain the aggregated location message and the secret key Key_C. In the message exchange stage, the server and client use the shared Key_C to communicate. When the server wants to reply to a client, it generates a random number, Key_S’. The reply message is directly encrypted using Key S’ to generate a ciphertext, E{Rep}. Since the aggregated location message contained the client’s GPS position, the server can generate a GeoLock of the client vehicle’s decryption region. This GeoLock is XOR-ed with Key_S’ and then encrypted with Key_C to generate a ciphertext, E{Key’}. Both E{Rep} and E{Key’} are transmitted to the client through the wireless channel. E{Key’} is then decrypted using Key_C to recover the XOR of the client’s GeoLock region and Key_S’. The client generates its GeoLock based on its current location. This is used to recover the secret key Key_S’. E{Rep} is decrypted using Key_S’, and the reply message is recovered. The client repeats the algorithm in the message exchange stage to communicate with the server. 3.2 A Simple GeoLock Mapping Function The GeoLock mapping function converts geographic location, time and mobility parameters into a unique value as a lock. This unique lock value validates that the recipients satisfy certain restriction, for example the decryption region at a certain time interval. There are no preinstalled mapping tables in our proposal. The process of generating a lock value/key is shown in Figure 2(b). First of all, all the input parameters are operated respectively. The location (x0; y0) will be divided by the length of decryption region (square) L. Each of coordinate number of P0(x0; y0) will be divided by 100. The integral part after division will be obtained. Therefore, bigger L will cause less digital numbers of the output from step one. Less digital numbers will result in weaker lock key. If the value of L is small, there is a risk that a lock key may be computed by brute force attack. Second, the output of the first step is multiplexed or reshuffled. Third, the output of the second step is hashed by a hash function. The hash function in practice can be designed as mod operation or as

696

G. Yan et al.

standard hash function, like Secure Hash Algorithm (SHA) functions. An example is shown in Figure 2(c). First step, two numbers are divided by the length of the region 100, i.e. (042.00, 915.00). The integer part after division, i.e. (042, 915) is kept. Second step, the two numbers: (042,915) are multiplexed as 042915. Third step, the multiplexed number is hashed by SHA to generate the lock value. From the recipient’s view, the secret key will be recovered. The recipient’s GPS coordinates are read from the enlisted GPS receiver. The other parameters in figure 2(b) can be obtained on the recipient vehicle b. If the vehicle b is restricted by the decryption region in terms of location, time and relative speed, the exact same lock value will be generated. An example of the mapping function on the receiver’s view is shown as Figure 3(a). The receiver vehicle b is located at location (04250; 91520) (UTM 10 digital coordinates) and the decryption region L is 100 meters. Figure 3(a) shows GeoLock on recipients. First step, the xy-axis coordinates of location (04250; 91520) are divided by the length of the region 100, i.e. (042.50, 915.20). The integer part after division, i.e. (042, 915) is obtained. Second step, the two numbers (042, 915) are multiplexed as 042915. At this point, the multiplexed number is exactly same as the one in key generator on sender side. Hash function (SHA) will generate exactly same key as well. We show that the lock value generated on the receiver side is exactly same as the one computed in GeoLock from the sender’s view. It is obvious that the vehicles will pass the geographic validation.

Fig. 2. GeoEncryption

3.3 A Comprehensive GeoLock Mapping Function In the previous section, a simple mapping function deals with a square decryption region. The square region can be used in many scenarios, such as a intersection of roads, a segment of a highway, or a roadside infrastructure. But there are some cases that the shape of the decryption region is not expected to be square. Generally, the shape of decryption region can be any one, e.g. a triangle, or a irregular shape. As an example, we are going to use a triangle to represent a irregular shape, comparing with a square. The mapping function is designed to convert a location to a key. If decryption region is an irregular region, we will partition the shape into a set of square regions, as shown in figure 3(b). Although we will have small margin that is not covered by the square regions, the majority of the shape is covered. When a sender wants to encrypt a message, it will predict the reception’s decryption region which is a set of square regions representing the original triangular decryption region. Figure 3(c) shows the seven decryption regions. The sender computes the value of GeoLock by

A Geographic Location-Based Security Mechanism for Intelligent Vehicular Networks

697

randomly using one of the square regions and encrypts the message by using the GeoLock value to produce a ciphertext. The ciphertext is transmitted through wireless channel and received by a receiver. The receiver, thereafter, checks all the decryption regions of the sub-decryption-squares. If the receiver is valid, one of the square regions will produce the right key and decrypt the ciphertext.

Fig. 3. GeoEncryption

4 Simulation Results We used SUMO [5] and ns-2 for our simulation. The simulator SUMO is a microscopic road traffic simulation package and ns-2 is a network simulator for VANETs [6], [7]. The network routing protocol we were running is the Ad hoc On Demand Distance Vector (AODV). The application is Constant Bit Rate (CBR) with 16 packets every second. The total amount of vehicles is 320. The map is 3.2km x 3.2km. First, the decryption ratio over location tolerance/precision is investigated. We measured the decryption ratio as the ratio of successfully decrypted messages over those messages that were received. We varied the location tolerance because location detection has precision problem. As we expected, the increase of location tolerance will cause the decrease of the decryption rate, shown in Figure 4(a). Besides, the faster speed will cause lower decryption rate. This is because that the increase of location tolerance and the increase of speed will cause the increase of false location of vehicles. The decryption ratio is higher in our algorithm than the one in Al-Fuqaha’s algorithm which does not consider the prediction errors. Security is not free but with cost. We measured the overhead (packet size increase) and the decryption time by varying the updating pause time. Figure 4(b) shows the result that the percentage of overhead increment decreases while the updating pause increases. In our algorithm, the fixed-size square is less sensitive to the change of the pause than the Al-Fuqaha’s algorithm. We are also interested to compare the decryption ratio between a square decryption region and a triangle decryption region. We varied the location tolerance and the other simulation parameters are same as the ones in the first simulation. The result is shown in figure 4(c). We noticed that the triangle region has smaller decryption ratio for both speeds. The reason is that the summation

698

G. Yan et al.

area of sub-squares’ area is less than the triangle. In some cases, the vehicles are in the triangle region but not in any of the sub-squares, which will cause a failure of decryption. Therefore, the deviation of approximation of sub-squares will cause a degraded of performance.

Fig. 4. Simulations

5 Conclusion We describe a feasible and novel geographic location-based security mechanism for vehicular adhoc network environment on the basis of concepts proposed by [2], [3]. Comparing with the [2], [3], our algorithm is efficient on the basis of simulation. The future work will integrate the model into the existing security methods. The shape of the decryption region will be extended to any shape.

References [1] Yan, G., Olariu, S.: An efficient geographic location-based security mechanism for vehicular ad hoc networks. In: Proceedings of the 2009 IEEE International Symposium on Trust, Security and Privacy for Pervasive Applications (TSP 2009), October 12-14 (2009) [2] Denning, D., MacDoran, P.: Location-based authentication: Grounding cyberspace for better security. Computer Fraud and Security 1996(2), 12–16 (1996) [3] Scott, L., Denning, D.E.: Location based encryption technique and some of its applications. In: Proceedings of Institute of Navigation National Technical Meeting 2003, Anaheim, CA, January 22-24, pp. 734–740 (2003) [4] Al-Fuqaha, A., Al-Ibrahim, O.: Geo-encryption protocol for mobile networks. Comput. Commun. 30(11-12), 2510–2517 (2007) [5] Open source, Simulation of urban mobility, http://sumo.sourceforge.net [6] Yan, G., Ibrahim, K., Weigle, M.C.: Vehicular network simulators. In: Olariu, S., Weigle, M.C. (eds.) Vehicular Networks: From Theory to Practice. Chapman & Hall/CRC (2009) [7] Yan, G., Lin, J., Rawat, D.B., Enyart, J.C.: The role of network and mobility simulators in evaluating vehicular networks. In: Proceedings of The International Conference on Intelligent Computing and Information Science (ICICIS 2011), Chongqing, China, January 8-9 (2011)

Intrusion-Tolerant Location Information Services in Intelligent Vehicular Networks Gongjun Yan1, Weiming Yang2, Earl F. Shaner1, and Danda B. Rawat3 1

Indiana University Kokomo, Kokomo, IN 46904 USA [email protected] 2 Chongqing Technology & Business Univ., Chongqing 400067, China [email protected] 3 Old Dominion University, Norfolk, VA 23529 USA [email protected]

Abstract. Intelligent Vehicular Networks, known as Vehicle-to-Vehicle and Vehicle-to-Roadside wireless communications (also called Vehicular Ad hoc Networks), are revolutionizing our daily driving with better safety and more infortainment. Most, if not all, applications will depend on accurate location information. Thus, it is of importance to provide intrusion-tolerant location information services. In this paper, we describe an adaptive algorithm that detects and filters the false location information injected by intruders. Given a noisy environment of mobile vehicles, the algorithm estimates the high resolution location of a vehicle by refining low resolution location input. We also investigate results of simulations and evaluate the quality of the intrusion-tolerant location service. Keywords: Intrusion detection, intelligent vehicular networks, information security, vehicular ad hoc networks.

1 Introduction In the past few years, Vehicular Adhoc NETworks (VANETs), known as Vehicle-toVehicle and Vehicle-to-Roadside wireless communications, have received a huge amount of well-deserved attention in the literature. The original impetus for the interest in the intelligent vehicular networks was provided by the need to inform fellow drivers of actual or imminent road conditions, delays, congestion, hazardous driving conditions and other similar concerns. But almost all advisories and other trafficsafety related messages depend on a critical way on location information. For example, traffic status reports, collision avoidance, emergency alerts, cooperative driving, or resource availability are directly related to location information. If location information is altered by malicious attackers, these applications will not work at all and will cause some real traffic accidents under certain condition. For example, there is scenery that the line of sight to merge a highway is blocked by trees. If the drivers trust only the location information received from other vehicles, real traffic accident can happen when a roadside intruder vehicle sends false location information about other vehicles. Therefore, it is of importance to ensure intrusion-tolerant location R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 699–705, 2011. © Springer-Verlag Berlin Heidelberg 2011

700

G. Yan et al.

information. The main difficulty, however, comes from the different types of data sources which are in a noisy environment. In reality, there exists malicious attacker who can modify the location information. Malicious vehicles can report bogus position. Therefore, it is necessary to collect location information from different sources and estimate the real location among the location information. The motivation of this paper is to provide intrusion-tolerant location services in intelligent vehicular networks by specially designed mechanisms. In this paper, we describe an adaptive algorithm that detects and filters the false location information injected by intruders. Given a noisy environment of mobile vehicles, the algorithm estimates the high resolution location of a vehicle by refining low resolution location input. The collected location report as input samples are filtered by a statistic method: multi-granularity deviation factor. The contributions of this paper, therefore, include: 1) filter the false location information inserted by intruders; 2) compute highresolution location estimation; 3) improve accuracy of filtering and location estimation; 4) not require any probability distribution of the input samples. The intrusion detection algorithm, comparing with other algorithms, is more applicable to real traffic situation.

2 State of the Art Position estimation is often made by using reference devices. For example, cellular local positioning system [1] uses infrastructure (base station) or reference device (RFID) to locate a position. New devices, such as ultrasonic sensor and digital compass, camera, radar [2], [3], have been applied to locate mobile nodes. The devices can provide higher resolution observations/input. These devices can ensure the integrity of observation with line of sight. But under the non-line-of-sight (NLOS) situation, the integrity cannot be ensured. In addition, the new devices cause extra cost of the system. Our algorithm relies on devices that have already existed in intelligent vehicular networks. No extra devices are needed. Kalman filter and particle filter (Monte Carlo filter) [1] are applied in the position estimation as well. To apply these filters, the input data must follow the same distribution (normal distribution). The same distribution of all input samples is hard to satisfy in real traffic. Time-of-arrival (TOA), direction-of-arrival (DOA) [4] and received signal strength indicator (RSS or RSSI) have been originally developed to operate under line-of-sight propagation between the transmitter and the receiver. However, these algorithms are corrupted by NLOS propagation errors. In literature, there are a number of methods identifying and mitigating NLOS errors have been addressed [5].

3 Location Information Intrusion Detection 3.1 Mahalanobis Distance In another previous work [6], we applied Mahalanobis distance, a distance measure introduced by P. C. Mahalanobis [7]. Formally, it is defined as a dissimilarity measure

Intrusion-Tolerant Location Information Services in Intelligent Vehicular Networks

between two random vectors matrix S:

and

701

of the same distribution with the covariance

The intuitive explanation of Mahalanobis distance of a test point from the center of mass divided by the width of the ellipse/ellipsoid, OA/OW for test point A and OB/ OW for test point B in figure 1(b). The bigger Mahalanobis distance value is, the more likely the test point is a false location. This method needs that all the input location samples follow normal distribution. In this paper, the method we address does not require the normal distribution of location input samples.

Fig. 1. Algorithms

3.2 Box Counting Method In our previous work [8], we proposed the box counting method which filters the outliers by placing all location input into a grid area and counting the number of location in each grid. The subgrid with the biggest number of locations will be remained and refined to obtain the final location estimation. However, the box-counting method has one potential risk. If the grid granularity is not fit, there will have location estimation deviation. An example of the location estimation deviation caused by improper grid granularity is shown in figure 1(c). If we use the big grid which is composed by 4 small grids in figure 1(c), the location estimation will be covered the dotted-circle. But if we use a smaller grid granularity, say the shaded grid (one grid), the location estimation will be covered by the triangle shown in figure 1(c). The distance between the two location estimation is about 14.4m. In this paper, we proposed a statistic method that avoids the effect of the grid granularity. 3.3 Local Correlation Integral Method (LOCI) Papadimitriou et al., 2002 propose a Local Correlation Integral method (LOCI) for finding outliers in large, multidimensional data sets [9]. They introduced the multigranularity deviation factor (MDEF), which can cope with local density variations in

702

G. Yan et al.

the feature space and detect both isolated outliers as well as outlying clusters. The basic idea is that a point is selected as an outlier if its MDEF values deviate significantly (more than three standard deviations) from the local averages. Intuitively, the MDEF at radius r for a point pi is the relative deviation of its local neighborhood density from the average local neighborhood density in its r-neighborhood. Thus, an object whose neighborhood density matches the average local neighborhood density will have an MDEF of 0. In contrast, outliers will have MDEFs far from 0. The main symbols and basic definitions we are going to use are described in reference [9]. To be consistent, we use the same notation here as Papadimitriou et al.,2002. For any pi, r and α the multi-granularity deviation factor (MDEF) at radius r is defined as:

(1) The counting neighborhood (or αr-neighborhood) is the neighborhood of radius αr, over which each n(p,αr) is estimated. The sampling neighborhood (or rneighborhood) is the neighborhood of radius r, over which samples of n(p,αr) are collected in order to estimate n(p,r,α). The main outlier detection scheme of LOCI relies on the standard deviation of the αr-neighbor count over the sampling neighborhood of pi. We thus define the following quantity: (2) of which is the normalized standard deviation . Given the above definitions, the outlier detection scheme can be for outlined as follows. For each point and counting radius, the sampling neighborhood is selected to be large enough to contain enough samples. We choose α=0.5 in all exact computations. The MDEF values are examined for a wide range of sampling radius. The minimum sampling radius rmin is determined based on the number of objects in the sampling neighborhood. We always use a smallest sampling neighborhood with neighbors; in practice, this is small enough but not too small to values. A point is flagged as an introduce statistical errors in MDEF and its MDEF is sufficiently large, i.e., outlier, if for any A reasonable choice is which is suggested by Papadimitriou et al., 2002. The standard deviation-based flagging is one of the main features of the LOCI method. It replaces any arbitrary cut-offs with probabilistic reasoning based on . It takes into account distribution of pair wise distances and compares each object to those in its sampling neighborhood. Note that, even if the global distribution of distances varies significantly, the use of the local deviation successfully solves this problem. In fact, in many real data sets, the distribution of pair wise distances follows a specific distribution over all or most scales. Thus, this approach works well for many real data sets.

Intrusion-Tolerant Location Information Services in Intelligent Vehicular Networks

703

3.4 The Proposed Algorithm The algorithm of detecting location information intrusion is as follows.

Discussion: The algorithm Adaptive Intrusion Detection, in each iteration, it takes one location and put the location in a set L. After receiving a certain amount of locations (the time is up), the set of locations L is examined. If L is normally distributed, M-distance Algorithm is applied to filter the outliers. Otherwise, the LOCI algorithm [9] will be applied if the size of L is smaller than the value of multiples of traffic density, the number of lanes and the size of wireless transmission range. If none of the conditions satisfied, Box-Counting Algorithm is used to filter the outliers.

4 Simulation The proposed position filtering and estimating algorithms were simulated for straight line traffic. We used a tailored Java simulation engine. Vehicles’ mobility model is implemented as IDM [10], [11], [12]. We assumed 1% of vehicles will send out compromised position. They are malicious attackers who can create a random bogus location. The remaining 99% vehicles are honest on the location information. The input is processed by MATLAB R2009A where our algorithms are implemented. To know the accuracy of filtering outliers, we compared the filtering effect of three algorithms: LOCI, m-distance and box-counting. The size of samples is varied from 100 to 210. We recorded the number outliers filtered by the three algorithms. Figure 2(a)

Fig. 2. Comparison

704

G. Yan et al.

shows the comparison of filtering accuracy. We expected to filter a specific number of outliers (2% of the sample size). Only LOCI and m-distance filtered the expected number of outliers. The box-counting algorithm can only filter part of the outliers. Figure 2(b) shows the location estimation after the outliers are filtered by the three algorithms. We expected to obtain location estimation of (33.0, 33.0). Both m-distance and LOCI can obtain pretty precise location estimation. We investigated the complexities of the three algorithms. The comparison of the number of iterations is shown in Figure 2(c). As expected, the m-distance is more efficient, the box-counting is in the middle and the LOCI takes about O(n3) iterations to filter outliers.

5 Conclusion and Future Work We described an adaptive algorithm for filtering and estimating the location of a vehicle with the given noisy input data. Based on the features of the input location samples (distribution and the sample size), different algorithms are adopted for the intrusion location detection. Three algorithms are included in the adaptive algorithm. In the future, we will explore our algorithm by using other filters, e.g, Kalman filtering or particle filtering to filter the malicious data.

References [1] Widyawan, Klepal, M., Pesch, D.: Influence of predicted and measured fingerprint on the accuracy of rssi-based indoor location systems. In: 4th Workshop on Positioning, Navigation and Communication, WPNC 2007, March 2007, pp. 145–151 (2007) [2] Yan, G., Olariu, S., Weigle, M.C.: Providing VANET security through active position detection. Computer Communications: Special Issue on Mobility Protocols for ITS/VANET 31(12), 2883–2897 (2008) [3] Kim, H.-S., Choi, J.-S.: Advanced indoor localization using ultrasonic sensor and digital compass. In: International Conference on Control, Automation and Systems, ICCAS 2008, October 2008, pp. 223–226 (2008) [4] Bartelmaos, S., Abed-Meraim, K., Attallah, S.: Mobile localization using subspace tracking. In: Asia-Pacific Conference on Communications, October 2005, pp. 1009–1013 (2005) [5] Moraitis, N., Constantinou, P.: Millimeter wave propagation measurements and characterization in an indoor environment for wireless 4g systems. In: IEEE 16th International Symposium on Personal, Indoor and Mobile Radio Communications, PIMRC 2005, September 2005, vol. 1, pp. 594–598 (2005) [6] Yan, G., Yang, W., Olariu, S.: Data fusion for location integrity in vehicle ad hoc networks. In: Proceedings of The 12th International Conference on Information Integration and Web-based Applications and Services (iiWAS 2010), Paris, France, November 8-10 (2010) [7] Mahalanobis, P.C.: On the generalised distance in statistics. Proceedings National Institute of Science, India 2(1), 49–55 (1936) [8] Yan, G., Chen, X., Olariu, S.: Providing vanet position integrity through filtering. In: Proceedings of the 12th International IEEE Conference on Intelligent Transportation Systems (ITSC 2009), St. Louis, MO, USA (October 2009)

Intrusion-Tolerant Location Information Services in Intelligent Vehicular Networks

705

[9] Papadimitriou, S., Kitagawa, H., Gibbons, P.B., Faloutsos, C.: LOCI: fast outlier detection using the local correlation integral. In: Proceedings of 19th International Conference on Data Engineering, pp. 315–326 (2003) [10] Yan, G., Ibrahim, K., Weigle, M.C.: Vehicular network simulators. In: Olariu, S., Weigle, M.C. (eds.) Vehicular Networks: From Theory to Practice. Chapman and Hall/CRC (2009) [11] Yan, G., Lin, J., Rawat, D.B., Enyart, J.C.: The role of network and mobility simulators in evaluating vehicular networks. In: Proceedings of The International Conference on Intelligent Computing and Information Science (ICICIS 2011), Chongqing, China, January 8-9 (2011)

The Role of Network and Mobility Simulators in Evaluating Vehicular Networks Gongjun Yan1, Jingli Lin2, Danda Rawat3, and Justin C. Enyart1 1

Indiana University Kokomo, Kokomo, IN 46904 USA {goyan,jcenyart}@iuk.edu 2 Xihua University, Chengdu 610039, China [email protected] 3 Old Dominion University, Norfolk, VA 23529 USA [email protected]

Abstract. Simulators are important tools in evaluating research in vehicular networks. Typically, vehicle movements are determined by a vehicle mobility simulator, and the movement trace is then input into a network simulator to simulate communication between the vehicles. So, it is important to consider both mobility and network simulators. We present an overview of popular simulators used in vehicular networking research along with an experimental comparison of two popular vehicular mobility simulators. We show that the mobility model and topology used can greatly affect the network performance. Keywords: Vehicular networks, mobility models, simulators.

1 Introduction Vehicular networking is an emerging area of interest in the wireless networking community as well as in the transportation research community. The potential of vehicular networks to provide vital services, from real-time traffic information to advance collision warning, makes this an important area of study. Vehicular networking can comprise vehicle-to-vehicle (V2V) communication, vehicle-to-infrastructure (V2I) communication, or a combination of both. Typically, networks formed without infrastructure support are termed ad hoc networks, thus, vehicular networks with only V2V communication have been called vehicular ad hoc networks (VANETs). In order to evaluate VANETs, researchers almost always must resort to simulation as the expense of actual deployment is too high. Unfortunately, there is no standard vehicular networks simulator. Currently, most researchers generate a mobility trace using a vehicular mobility simulator and input this trace to a standard networking simulator. The choice of the mobility simulator is important as performance in vehicular networks depends highly on connectivity of the nodes. The manner in which nodes move greatly affects this connectivity. But, there has not been much work in reviewing the state of VANET simulation. The work of Harri et al. [1], [2] is an exception, with the authors focusing on classifying the available options for vehicular mobility simulators. This paper attempts to fill R. Chen (Ed.): ICICIS 2011, Part II, CCIS 135, pp. 706–712, 2011. © Springer-Verlag Berlin Heidelberg 2011

The Role of Network and Mobility Simulators in Evaluating Vehicular Networks

707

the gaps left by previous work. Inspired by Kurkowski et al. [3], [2], we review and classify the simulators used in vehicular networks studies that have been presented at the ACM VANET workshops from 2004-2007. More importantly, we present a comparison of two VANET simulation solutions and show the impact that the choice of mobility model and topology have on various network metrics.

2 Overview of VANET Simulators VANET simulations typically require a networking component and a mobility component. Before we describe the various simulators used for VANET studies, we investigate which simulators have been used for work presented at the ACM VANET workshop from 2004-2007. In the four years of the workshop, there were a total of 69 papers presented. Of these, 74% (51 papers) used some type of simulation as part of their evaluation. Out of these 51 papers, 36 papers (70%) named the specific simulator that was used, and 8 papers (16%) used a self-developed simulator. We classify each simulator according to its purpose (i.e., network simulator or mobility simulator). Table I shows the vehicular mobility and network simulators used. By far, NS-2 is the most popular choice for a network simulator. This is not surprising as NS-2 is the most widely-used simulator in the networking community. But, as we will discuss later, NS-2 may not be the best choice for VANET simulations. Table 1. Simulators Used in Papers in ACM VANET (2004-2007)

2.1 Vehicular Mobility Simulators Here we present an overview of the vehicular mobility simulators used in papers from ACM VANET as well as other publicly-available mobility simulators. SHIFT/SmartAHS [4] is an Automated Highway Systems (AHS) simulator developed as part of the California PATH project at UC-Berkeley. It was originally built to simulate automated vehicles, but a human driver model [5], based on the cognitive driver model COSMODRIVE, was later added. SHIFT/SmartAHS is still available for free download from PATH. All three papers from ACM VANET that used SHIFT/SmartAHS come from the research group at UC-Berkeley. The Microscopic Traffic Applet [6] is a Java simulator that implements the IDM car-following model. The default scenarios are ring road, on-ramp, lane closing, uphill grade, and traffic lights. As this is an applet designed to illustrate IDM, it does not

708

G. Yan et al.

include any method to import maps from other sources, set a path from source to destination, or output a trace file for input into a network simulator. VanetMobiSim [7] is an extension of the CanuMobiSim [8] project. CanuMobiSim implemented Random Waypoint movement, and its trace files can be imported into the NS-2, GloMoSim, or QualNet network simulators. VanetMobiSim adds an implementation of IDM as well as IDM-LC (lane change) and IDM-IM (intersection management). Maps can be generated by the user, randomly, or using the TIGER/Line database. In addition, the mobility model used in VanetMobiSim has been validated against the CORSIM mobility simulator. SUMO (Simulation of Urban Mobility) [9] is an open-source mobility simulator that uses Random Waypoint path movement and the Krauß car-following model. SUMO supports maps from TIGER/Line and ESRI. MOVE [10] is an extension to SUMO that allows its vehicle mobility traces to be imported into NS-2 or QualNet. 2.2 Network Simulators Since vehicular networks involve solely wireless communications, all of the networking simulators described here support performing simulations with mobile wireless nodes. We briefly describe the publicly-available simulators. NS-2 [11] is an open-source discrete event network simulator that supports both wired and wireless networks, including most MANET routing protocols and an implementation of the IEEE 802.11 MAC layer. NS-2 is the most widely-used simulator for networking research and is the most-used network simulator in the ACM VANET workshop. J-Sim [12] is an open-source simulation environment, developed entirely in Java. J-Sim provides two mobility models: trajectory-based and random waypoint. J-Sim is presented as an alternative to NS-2 because it is designed to be easier to use, but JSim has not been updated since 2006. SWANS (Scalable Wireless Ad hoc Network Simulator) [13] was developed to be a scalable alternative to NS-2 for simulating wireless networks. Based on comparisons of SWANS, GloMoSim, and NS-2, SWANS was the most scalable, the most efficient in memory usage, and consumed the least runtime [13,14]. Recently, the network model in SWANS was validated against NS-2 [14]. It was shown that along with better performance, SWANS delivered similar results as NS-2, at least for the network components that were implemented in both.

3 Comparison by Experiment In addition to analyzing what simulators have been used in previous VANET research, we present a comparison of two mobility simulators to demonstrate the differences observed at the network level caused by different mobility models. We choose to compare VanetMobiSim and SUMO not only because they are publicly available, but also because they are both well-maintained and still continue to be developed and improved. In order to keep the networking components the same, we use NS-2 as the network simulator in these experiments. We constructed a 2000m x 2000m area of city streets. Both VanetMobiSim and SUMO use a random trip generator to choose a

The Role of Network and Mobility Simulators in Evaluating Vehicular Networks

709

path between source and destination. We place 60 vehicles at five different initial positions as shown in Figure 3(a). Once moving, vehicles never leave the map, making random turning decisions at each intersection. The speed limits on the streets range from 5-20 m/s (11-45 mph), and the vehicles are constrained by the speed limit. SUMO and VanetMobiSim are run with the configuration described above. Both of these mobility simulators generate a mobility trace that can be input into NS-2. Then, NS-2 is run with the configuration as described below. For the car-following model, VanetMobiSim uses IDM, and SUMO uses the Krauß Model. Out of the 60 vehicles in the simulation, 10 are chosen (at random) as data sources and 10 are chosen as data sinks. This results in 10 constant-bit rate (CBR) data flows. Each data source sends four 512-byte packets per second, resulting in an offered load of 16 kbps (kilobits per second) from each CBR flow. Packets are routed from source to destination using the Ad-hoc On-Demand Distance Vector Routing (AODV) [15] algorithm (described below). Each vehicle will buffer packets (in a finite queue) until a route has been found to the destination. 1) Throughput: Figure 1(a) shows the throughput obtained with VanetMobiSim mobility and SUMO mobility. We also show a line for offered load, which indicates the amount of data that senders are transmitting into the network. Each dot on the offered load line represents a new CBR flow starting. In both cases, the throughput is bursty with spikes of high throughput along with periods of very low throughput. This suggests that the topology is changing and that routes are being broken. To investigate the performance further, we look at loss and delay. 2) Loss Rate: We computed the loss rate for CBR packets every 10 seconds in each experiment and show the results in Figure 1(b).The loss rates are shown on the graph in the time interval that the packets were sent. With SUMO mobility, no packets sent before time 100 seconds were ever received, and with VanetMobiSim mobility, no packets sent before time 50 seconds were received. This high rate of loss is due to vehicles being initially deployed at the borders of the map (Figure 3(a)). Since the average speed is about 20 m/s, and the area is 2000m x 2000 m, vehicles need to move about 1000 m before they are in range of one another. The time needed for this movement is about 50 seconds. Once vehicles cover the whole area of the map, more routing paths are available, and the loss rate begins to decrease.

Fig. 1. Throughput and Loss Rate

710

G. Yan et al.

3) Packet Delay: In Figures 2(a) and 3(b), we show the packet delay for each packet successfully received by the destination with VanetMobiSim mobility and SUMO mobility, respectively. In the figures, the delay for a packet is shown at the time the packet is sent by the source. We give the delays from each CBR flow a different color and symbol to distinguish them.

Fig. 2. Delay and Packet Drop

We note that the absence of packet delays up to time 50 seconds with VanetMobiSim mobility and up to time 100 seconds with SUMO mobility matches the loss rates for each, since we can only compute packet delays for packets that are actually received. In both figures there are delay spikes followed by sharp decreases in delay. This is indicative of a queue draining, especially since all of the packets in the decreases are from the same CBR flow. We looked in particular at the first spike in Figure 2(a), which is enlarged in Figure 2(c). All of the packets represented in this area are from a single flow, from source node 7 to destination node 9. Source node 7 begins sending packets at time 4.08 seconds, but according to Figure 1(b), no packets sent before time 50 seconds are received. The first packet is received 10 seconds after it was sent. Subsequent packets have a linearly decreasing delay. The issue is that AODV cannot set up a route from the source to the destination until time 61.04 seconds. Once the route is setup, all of the packets in the source’s queue can be sent and are delivered.

Fig. 3. Simulation Reasoning

To further investigate the reason for the length of time needed to setup the route between source node 7 and source node 9, we show the position of the vehicles at the time node 7 sends an AODV request (Figure 3(c)). Node 7 sends the RREQ to node 6, but the next hop, node 0, is out of communications range as the distance between the

The Role of Network and Mobility Simulators in Evaluating Vehicular Networks

711

two vehicles is about 525 m, which is larger than the transmission range of 250 m. Nodes 6 and 0 will only be able to communicate with each other once they are in range, so the vehicles need to travel about 275 m before they are within 250 m of each other and can communicate. Based on the speed that the two vehicles are traveling (about 28 m/s), it will take about 10 seconds for them to travel the required distance and be in communications range. Thus, the topology of vehicles has a dominant effect on packet delay when there is no direct route. We look at the details of packet losses to confirm this. In Figure 2(b) we show each dropped packet in the case of SUMO mobility and the reason given for the drop. We note that in our loss rate figures, some packets are not dropped, but just are never heard by the receiver. In this figure, we are looking at specific reasons why packets were dropped at a node. The reasons include queue overflow, wireless collision, packet has been retransmitted a maximum number of times, and no route to the next host. Early in the simulation, several packets are dropped because there is no route, and many are dropped because of queue overflow. These queue overflow drops are occurring at the source nodes because they are not able to deliver any of the CBR packets. Starting around 140 seconds, there are many packets dropped due to collision. This indicates that routes are available and the wireless channel has become busy.

4 Conclusion We have presented an overview of publicly-available simulators used for vehicular networking research. Typically, vehicle movements are determined by a vehicle mobility simulator, and the movement trace is then input into a network simulator to simulate communication between the vehicles. In addition to reviewing currently available simulators, we performed experiments to compare the SUMO mobility generator and the VanetMobiSim mobility generator. We show that vehicle mobility can greatly affect network performance in terms of throughput, loss, and delay.

References [1] Haerri, J., Filali, F., Bonnet, C.: Mobility models for vehicular ad hoc networks: a survey and taxonomy. IEEE Communications Surveys and Tutorials (epublication) [2] Yan, G., Ibrahim, K., Weigle, M.C.: Vehicular network simulators. In: Olariu, S., Weigle, M.C. (eds.) Vehicular Networks: From Theory to Practice. Chapman and Hall/CRC (2009) [3] Kurkowski, S., Camp, T., Colagrosso, M.: Manet simulation studies: the incredibles. ACM SIGMOBILE Mobile Computing and Communications Review 9(4), 50–61 (2005) [4] Antoniotti, M., Göllü, A.: SHIFT and SmartAHS: A language for hybrid systems engineering, modeling, and simulation. In: Proceedings of the USENIX Conference of Domain Specific Languages, Santa Barbara, CA, pp. 171–182 (1997) [5] Delorme, D., Song, B.: Human driver model for SmartAHS, Tech. rep., California PATH, University of California, Berkeley (April 2001) [6] Microsimulation of road traffic, http://www.traffic-simulation.de/ [7] Vanetmobisim project home page, http://vanet.eurecom.fr [8] Canumobisim, http://canu.informatik.uni-stuttgart.de

712

G. Yan et al.

[9] Vanet simulator sumo, http://sumo.sourceforge.net/ [10] Karnadi, F., Mo, Z.H., Lan, K.-C.: Rapid generation of realistic mobility models for VANET. In: Proceedings of the IEEE Wireless Communications and Networking Conference (WCNC), pp. 2506–2511 (2007) [11] McCanne, S., Floyd, S.: ns network simulator, http://www.isi.edu/nsnam/ns/ [12] J-sim home page, http://www.j-sim.org/ [13] Barr, R., Haas, Z., van Renesse, R.: Scalable Wireless Ad Hoc Network Simulation. In: Handbook on Theoretical and Algorithmic Aspects of Sensor, Ad hoc Wireless, and Peerto-Peer Networks, pp. 297–311. CRC Press, Boca Raton (2005) [14] Kargl, F., Schoch, E.: Simulation of MANETs: A qualitative comparison between JiST/SWANS and ns-2. In: Proceedings of the International Workshop on System Evaluation for Mobile Platforms (MobiEval), San Juan, Puerto Rico, pp. 41–46 (2007) [15] Perkins, C.E., Royer, E.M.: Ad hoc on-demand distance vector routing. In: Proceedings of the 2nd IEEE Workshop on Mobile Computing Systems and Applications, New Orleans, LA, pp. 90–100 (1999)

Author Index

Aborizka, Mohamed I-380 Ahmed, Taha Kh. I-380 Ai, Chunlu I-271 Allauddin, Maria II-631 An, Qi I-284 Attarzadeh, Iman I-18 Azam, Farooque II-631 Baharudin, B.T.H.T. II-385 Bai, Peng I-643 Bai, Xiangzhi I-124, I-278 Bei, Li II-183 Bhikshu, Huimin I-459 BinQiang, Wang I-711 Bo, Yuan I-711 Bo, Zhao I-711 Cai, Jiuju I-161 Cai, Ken II-675 Cao, Haiwang II-98 Cao, Sheng-Lung I-698 Cao, Wei II-53 Cao, Yaoguang II-668 Chai, Zhilei II-625 Chang, Guiran I-565 Chen, Bingkun I-489 Chen, Chengxun II-492 Chen, Chi-Hua II-72 Chen, Chin-Ling I-571, I-698 Chen, Chuwen II-544 Chen, Daqiang II-135 Chen, Deqiang I-136 Chen, Honglei I-51 Chen, Hui-Ju II-332 Chen, Jinguang II-372 Chen, Kerui II-165 Chen, Limin I-341 Chen, Man II-116 Chen, Shaohui I-737 Chen, Shengli II-445 Chen, Si-hua I-509 Chen, Tao II-142 Chen, Xiao II-417 Chen, Xiao-hong I-749 Chen, Ying I-271

Chen, Yongheng II-165 Cheng, Deqiang I-395 Cheng, Haifang II-177 Cheng, Haiying I-368 Cheng, Siyuan I-472 Cheng, Xian-Yi I-749 Cho, Il-Hwan II-509 Chueh, Hao-En I-259, II-411 Chunsheng, Wang I-446, I-497 Cui, Haigang II-668 Dai, Fei II-429, II-562 Dai, Feng II-30 Dai, Wei I-142, I-155 Dang, Jianwu II-644 Deng, Hong-Hui II-208 Di, Jinshan II-41 Ding, Hongwei II-327 Ding, Qiaolin I-483 Dong, Jianli I-199 Dong, Lili II-268 Dong, Xiaoma II-273 dong ping, Wang I-521 Du, Feifei I-643 Du, Ruiyan II-592 Du, Xiaogang II-644 Du, Yongping II-203 Du, Yuchuan II-154 Duan, Lihua II-296 Duan, Ping I-155 Duan, Shulin II-285 Enyart, Justin C.

II-706

Faghani, Mohsen II-359 Fan, Xiying II-233, II-378 Fan, Zhenmin II-580 Fan, Zhou I-529, I-737, II-128, II-613 Fang, Gang I-290 Fang, Guoyi I-368 Fang, Jing I-686 Fang, Liu II-256 Fang, Wei I-529, I-602 Fang, Zongde II-7

714

Author Index

Farashahi, Hamid Ghaani Feng, Tao II-135 Feng, Yanping I-643 Feng, Yu II-461 Feng, Zhilin II-662 Fu, Lihui II-568

II-385

Gan, Ping II-538 Gan, Zhi-chun I-99 Gang, Lei I-56 Gao, Chengjin I-374 Gao, Chong II-613 Gao, Fu-xiang II-497 Gao, Guohong II-353 Gao, Lilan II-580 Gao, Lulu I-193 Gengzhong, Zheng I-410 Gong, Ailing I-219 Gong, Lixiong II-16 Goswami, Subhra Sundar II-606 Gu, Jiangong II-7 Guangpu, Zhang II-291 Gulbag, Ali II-338 Guo, Changming II-321, II-619 Guo, Hao II-497 Guo, Jin II-321 Guo, Lei I-130 Guo, Lixin I-676 Guo, Minghui I-547 Guo, Ruilin II-262 Guolong, Liang II-291 Hai-yan, Li I-225 Han, Chongzhao I-186 Han, Deqiang I-186 Han, HuiJian II-503 Han, Mu I-541 Han, Shuqin I-237 Han, Xiaobing II-599 Hang, Qunjian I-360 Hao, Junling I-649 Hassanzadeh, Hamed II-347 He, Fengling II-165 He, Guangbin II-544, II-550 He, Ming II-203 He, Tingting II-372 He, Wei I-296 Hong, Li I-167 Hong, Sheng II-41 Hong, Tiansheng II-544, II-550

Hong, Zhang I-434 Hongbin, Jin I-27 Hongfei, Li I-27 Hou, Ligang II-59 Hou, Lin II-47 Hou, Peipei II-538 Hu, Fu-yuan I-180 Hu, Guang II-625 Hu, Luokai I-265 Hu, Peng I-142 Hua, Chunjian I-271 Hua, Han I-77 Hua, Jin I-749 Hua, Xufeng II-492 Huang, Guan II-477 Huang, Haiquan I-466 Huang, Huajuan I-656 Huang, Kai I-609 Huang, Meng II-244 Huang, Mingwei II-399 Huang, Peng II-208 Huang, Wei II-148 Huang, Xinfeng I-33 Huang, YiYun I-112 Huang, Yuansheng I-529, I-602 Huh, Woong II-509 Hui, Liu I-77 Huo, Shi-Wei II-314 Huo, Yuankai I-205 Jaberi, Mohammad II-385 Jeong, Taikyeong II-509 Ji, Zhaohui I-199 Ji, Zhen II-279 Jia, Xianzhao I-6, I-62 Jia, Zhengyuan I-737, II-128 Jian, Xu I-434 Jianfei, Ji II-291 Jiang, Guozhang II-471 Jiang, Jianhua II-16 Jiang, Linying I-502 Jiang, Murong II-262 Jiang, Shengchuan II-154 Jiang, Wenhao I-731, II-53 Jiangqiao, Lan I-27 Jiankui, Zeng I-302 Jiao, Hongwei II-302 Jie, Yin I-434 Ji-long, Xue I-554 Jin, Bo II-89

Author Index Jin, Ting I-124, I-278 Jing, Deng I-56 Jing, Jia II-487 Jing, Jiaqiang I-416 Jing, Ni II-638 Juan, Wang I-692 Juan, Zhu II-638 Juanjuan, Song I-589 Jun, Chen I-668 Jun, Hou I-434 Jun’an, Liu I-589 Kaiyang, Li I-446, I-497 Kang, Junjie II-189 Kang, Shaochen I-731, II-53 Kang, Yanning I-33, I-45 Keyvanpour, MohammadReza II-238, II-347 Kholghi, Mahnoosh II-238 Kim, Hyen-Ug II-509 Kim, Tae-Hyun II-509 Kong, Jianyi II-471 Kui, Fang I-692 Kui, Li II-256 Kung, Hsu-Yang I-596 Kunte, Srinivasa Rao R. II-366 Kuo, Chiung-Wen I-596 Lai, Changtao I-535 Lee, Yunsik II-509 Lei, Qiang II-668 Li, Chang-Jin II-503 Li, Chengfan I-243 Li, Chenming II-148 Li, Chunshu I-609 Li, Cunhua I-199 Li, Debin I-744 Li, Fu-le I-106 Li, Gongfa II-471 Li, Haobiao II-544, II-550 Li, Hengyu I-249 Li, Hongwei I-515 Li, Hongyan II-215, II-221 Li, Jiang II-568, II-580 Li, Jianqiang II-279 Li, Jinghui II-393 Li, Jishun I-6 Li, Jun II-110 Li, Kejie II-308, II-599 Li, Lei I-249

Li, Liang II-440 Li, Lili I-744 Li, Ming II-668 Li, Qiang I-402 Li, Sha II-644 Li, Sikun I-212 Li, Tong II-429, II-435, II-562 Li, Tu I-589 Li, Wei I-515, II-532 Li, Weijia I-686 Li, Wenjie I-395 Li, Xin I-731, II-53 Li, Xinxin II-59 Li, Xueyong II-353, II-515 Li, Yong II-128 Li, Yueping I-581 Li, Zhibin I-335 Li, Zhixiong I-296, II-244 Li, Zhong I-453, I-483 Li, Zuntao II-195 Lian, Jing II-110 Liang, Dongying I-581 Liang, Huaitao II-189 Liang, Huibin I-12 Liang, Shan II-538 Liang, Xiao I-335 Liang, Xiaoying II-675 Liangwei, Zhong II-638 Liao, Hsiu-Li II-250, II-332 Liao, Kuo-Hsiung I-259, II-411 Liao, Zhiming I-6 Lin, Bon-Yeh II-72 Lin, Chen-Wen II-250 Lin, Hui II-487 Lin, Jingli II-693, II-706 Lin, Qiang II-650 Lingling, Hu I-422 Liu, Caixia II-487 Liu, Chao I-717 Liu, ChunNian I-112, I-118 Liu, Guanghai II-23 Liu, Haiying II-580 Liu, Jian I-231 Liu, Jiangli II-203 Liu, Jingxu II-30 Liu, Junjuan II-268 Liu, Lan I-243 Liu, Li II-586, II-592 Liu, Minghui II-110 Liu, Peng I-308

715

716

Author Index

Liu, Pin-chao I-717 Liu, Qing II-429, II-435 Liu, Shu-Fan II-411 Liu, Su-Houn II-250, II-332 Liu, Tingrui I-12, I-219 Liu, Yan I-219 Liu, Yong I-161 Liu, Yonggang I-6, I-62 Liu, Yu I-686 Liu, Zhaoying I-124, I-278 Liu, Zheng I-662 Liu, Zhijing I-82 Lo, Chi-Chun II-72 Lu, Feifei II-417 Lu, Hsi-Peng II-521 Lu, Jianjiang II-556 Luo, Chang-Yuan II-314 Luo, Jun I-249 Luo, Shujun I-402 Luo, Zhiqing II-681 Lv, Feng I-62 Lv, Jinna II-353 Lv, Yi I-237 Lv, Yuan-jun I-231 Ma, Biao II-116 Ma, Han-wu I-621 Ma, Jiahui II-302 Ma, Li E. II-532 Ma, Ruixin I-686 Mao, Chong-Feng II-47 Meng, Huang I-45 Miao, Kun II-440 Ng, Cheng Man II-66 Nie, Yongdan II-393 Nordin, Md Jan II-359 Ow, Siew Hock I-18 Oztekin, Halit II-338 Pan, Gangzhu I-33 Pan, Qin I-112, I-118 Pan, Weimin I-6 Pan, Yue II-308 Pang, Heming I-502 Pang, Yongjie I-515 Park, Chong-Dae II-509 Park, Dong-Chul II-509 Pei, Daming II-599

Pei, Hai-long II-279 Peng, Chensong I-453 Peng, Xiaohong II-59 Peng, Yuxing I-212 Pengfei, Liu I-148 Pi, Shih-Ming II-250, II-332 Qian, Xinhong II-477 Qiang-bin, Ou-yang I-320 Qianmu, Li I-434 Qin, Honglin II-244 Qin, Zhigang I-354 Qing, He I-167 Qing, Sheng I-668 Qiqi, Hu II-1 Qiu, Taorong I-466 Qiumei, Liu I-410 Quande, Sun I-446, I-497 Rasane, Krupa R. II-366 Rawat, Danda B. II-693, II-699, II-706 Rui, Ding I-428 Rui, Zhiyuan II-35 Sang, Shengju I-284 Shan, Lu I-93, I-320 Shaner, Earl F. II-699 Shasha, Song I-711 Shen, Ding I-284 Shen, Guixiang I-489 Shen, Jian-jun I-99 Shen, Si II-262 Shen, Xiahong II-135 Shen, Xiaoyong II-110 Sheng, Buyun II-16 Sheng, Feng I-554 Sheng, Sheng I-374 Shi, Gouqing I-348 Shi, Ming-xi I-602, I-636 Shi, Xi I-360 Shi, Yanpeng I-704 Shi, Zhongzhi I-341 Shojaeipour, Shahed II-359, II-385 Shou, Yongyi I-535 Shu, Dongmei II-574 Shu, Shihu. II-656 Shui, Zhonghe I-155 Song, Boyan II-195 Song, C.N. II-619 Song, Ping II-308, II-599

Author Index Song, Xin II-574 Song, Zheng II-89 Su, Linfang I-335 Su, Xi I-643 Sui, Xiuhua I-12 Sun, Dawei I-565 Sun, Jian II-461 Sun, Jiaxia II-515 Sun, Lijun II-154 Sun, Wenqiang I-161 Sun, Xinhan II-110 Sun, Yan I-402 Sun, Yongcai II-23 Sun, Yongqing II-89 Tan, Chunguang I-565 Tan, Q.Y. II-619 Tang, Dunye II-477 Tang, Shuangqing II-244 Temurtas, Feyzullah II-338 Tian, Hua I-99 Tian, Yulong II-556 Tingting, Yan I-755 Tong, Bing II-135 Tong, Jin I-249 Tsai, Ching-Ping I-596 Tsaih, Rua-Huan I-459 Vai, Mang I. Wang, Wang, Wang, Wang, Wang, Wang, Wang, Wang, Wang, Wang, Wang, Wang, Wang, Wang, Wang, Wang, Wang, Wang, Wang, Wang,

II-66

Bin II-574, II-586, II-592 Changjian II-321 Deming I-348 Dian-kun I-106 Fengli II-285 Fuqiang II-35 Gang II-668 Haotian I-676 Jiabao II-556 Jia-Ching I-571 Jian I-180, II-562 Jianying II-148 Jingmin II-189 Jingyue I-676 Jinhui II-59 Jinkuan II-574, II-586, II-592 Juan I-62 Kaina II-302 Keqiang II-675 Li II-405

Wang, Lin I-374 Wang, Liying I-45 Wang, Liyong II-116 Wang, Lu I-466 Wang, Peng I-731, II-53 Wang, Shao-hong II-142 Wang, Shijun II-104 Wang, Shuihua I-205 Wang, Wen-cheng I-440 Wang, Xiaodong I-326 Wang, Xingwei I-565 Wang, Xi-ping I-636 Wang, Yang I-39, II-195 Wang, Yangping II-644 Wang, Yanming I-348 Wang, Yinghua I-69 Wang, Y.K. II-619 Wang, Yong I-547 Wang, YongLong I-118 Wang, Yuqiang I-69 Wang, Zhen II-399 Wang, Zhewei II-399 Wang, Zhihuan I-722 Wang, Z.L. II-619 Wei, Baolu II-30 Wei, Pengcheng I-360 Weiqiong, Bu I-692 Wenjun, Hou I-668 Woo, Dong-Min II-509 Wu, Gang II-233, II-378 Wu, Guoxin I-704 Wu, Junming I-199 Wu, Junyun I-466 Wu, Lenan I-205 Wu, Minghui I-630 Wu, Peng I-1, I-51 Wu, Qin II-35 Wu, Weibin II-544, II-550 Wu, Wuchen II-59 Wu, Xuewen II-477 Wu, Ying I-51 Wu, Yu II-454 Wulei, Tang I-148 Xia, Wei I-284 Xia, Zhuo-qun I-717 Xian, Xiaodong II-538 Xiang, Wei I-312 Xiang, Yuli II-177 Xiao, Hong II-687

717

718

Author Index

Xiao-gang, Qi I-554 Xiaolin, Zhang II-171 Xiaoming, Chang II-171 Xie, Guotao I-296 Xie, Li-tong I-717 Xie, Ruhe II-23 Xie, Shaorong I-249 Xie, Xiao-wu II-497 Xie, Xin I-51 Xie, Zhongwen II-562 Xin, Dong II-568 Xin, Haihui I-348 Xin, Hong-Zhi II-314 Xin, Jinglei I-731, II-53 Xing, Kezhi II-492 Xiong, Hegen II-471 Xiong, Jiang I-290 Xu, Baoshan II-568, II-580 Xu, Chun-gen I-541 Xu, Dawei II-492 Xu, Sen I-621 Xu, Siqiang II-471 Xu, Weiguang II-556 Xu, Xiao-li II-142 Xu, Xiuling I-326 Xu, Yabin I-173 Xu, Yafei II-41 Xue, Bindang I-124, I-278 Xue, Chaogai II-98, II-268 Xue, Lian I-630 Xue, Xiao I-39 Xun, Yangyang I-395 Yan, Dongmei II-586, II-592 Yan, Gongjun II-693, II-699, II-706 Yan, Hui I-630 Yan, Shitao II-515 Yan, Weigang II-321 Yan, Xiaolang I-609 Yan, Zhaofa I-686 Yang, Guowei I-130 Yang, Hongyi I-744 Yang, Huaqian I-360 Yang, Jianjun II-35 Yang, Jintang II-471 Yang, Li II-291 Yang, Liang-Chih II-521 Yang, Mingzhong II-16 Yang, Shichun II-668 Yang, Shuangxi II-262

Yang, Weiming II-693, II-699 Yang, Xiaofang II-7 Yang, Yan I-389 Yang, Yi I-186 Yang, Yu-Hsiang I-459 Yan-Lei, Liu I-56 Yao, Wenli I-193 Yao, Yu II-497 Yao, Yuan I-686 Ye, Cheng II-544, II-550 Ye, Feifan I-312 Ye, Yanming II-662 Ye, Zhijie II-550 Yecai, Guo I-422, I-428 Yefeng, Xu II-256 Yin, Jingben II-302 Yin, Jingyuan I-243 Yin, Yong-Sheng II-208 Ying, Hong I-290 Ying, Shi I-265 Ying, Zilu II-399 Yong, Qi I-434 Yong, Xu I-755, II-1, II-84 Yong, Zhang I-56 Yoon, Chung-Hwa II-509 Yu, Chunyan I-630 Yu, Ge II-497 Yu, Min I-609 Yu, Qian II-435 Yu, Yong II-429, II-435, II-562 Yuan, Yuan I-547 Yujin, He II-84 Yun-an, Hu I-225 Zhan, Wei II-681 Zhan, Zehui II-78 Zhang, Baojian II-353, II-515 Zhang, Chunqiu II-568, II-580 Zhang, Fengyuan II-41 Zhang, Hai Yan II-532 Zhang, Hao I-82 Zhang, Heng I-51 Zhang, Hong I-541 Zhang, Jianqiang I-69 Zhang, Jian-ze II-613 Zhang, Jibiao I-39 Zhang, Jiemin II-302 Zhang, Junli I-656 Zhang, Linli II-687 Zhang, Nana I-744

Author Index Zhang, Qishan II-41 Zhang, Wei I-212 Zhang, Weihua I-483 Zhang, Weiyuan II-687 Zhang, Xiangwei I-472 Zhang, Xingnan II-148 Zhang, Yafei II-556 Zhang, Yan II-393 Zhang, Yingzhi I-489 Zhang, Yong II-405 Zhang, Yuan I-731 Zhang, Yudong I-205 Zhang, Zexian II-262 Zhang, Zhen II-422 Zhao, Bin I-472 Zhao, Chenggui II-124 Zhao, Dongfeng II-327 Zhao, Guodong II-41 Zhao, Haiyong I-82 Zhao, Hui I-1 Zhao, Jichao I-284 Zhao, Junjuan I-243 Zhao, Kai I-265 Zhao, Na II-562 Zhao, Weiguo I-33 Zhao, Wenke II-625 Zhao, Yifan II-327 Zhao, Ziliang II-110 Zheng, Jianguo II-227

Zheng, Jingang I-173 Zheng, Rui I-489 Zheng, Weikun I-581 Zhengjing, Huang II-171 Zhenhua, Li I-77 Zhiming, Dong I-302 Zhong, Xiaoxing I-348 Zhou, Fugen I-124, I-278 Zhou, Hangjun I-212 Zhou, Liang II-227 Zhou, Qing II-47 Zhou, Ting I-395 Zhou, Yafu II-110 Zhou, Yongquan I-389, I-656 Zhou, Zhiwei II-195 Zhu, Daowei I-402 Zhu, Hongjun I-416 Zhu, Jiulong II-104 Zhu, Likuan II-195 Zhu, Peiyi II-454 Zhu, Shishun I-402 Zhu, Wei I-621 Zhu, Xiaoxiao II-135 Zhu, Xue-han I-717 Zhu, Yanxiang II-7 Zhu, Ying II-487 Zhu, Yuqing II-544, II-550 Zia, Mehmooda Jabeen II-631 Zuo, Wanli II-165

719