Lecture Notes in Electrical Engineering Volume 100
Ming Ma (Ed.)
Communication Systems and Information Technology Selected Papers from the 2011 International Conference on Electric and Electronics (EEIC 2011) in Nanchang, China on June 20-22, 2011, Volume 4
ABC
Ming Ma NUS ACM Chapter 81 Victoria Street, Singapore 188065, Singapore E-mail:
[email protected] ISBN 978-3-642-21761-6
e-ISBN 978-3-642-21762-3
DOI 10.1007/978-3-642-21762-3 Lecture Notes in Electrical Engineering
ISSN 1876-1100
Library of Congress Control Number: 2011929654 c 2011 Springer-Verlag Berlin Heidelberg 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, reuse of illustrations, recitation, broadcasting, reproduction on microfilm 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. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Typeset & Cover Design: Scientific Publishing Services Pvt. Ltd., Chennai, India. Printed on acid-free paper 987654321 springer.com
EEIC 2011 Preface
The present book includes extended and revised versions of a set of selected papers from the International Conference on Electric and Electronics (EEIC 2011), held on June 20–22, 2011, which is jointly organized by Nanchang University, Springer, and IEEE IAS Nanchang Chapter. The goal of EEIC 2011 is to bring together the researchers from academia and industry as well as practitioners to share ideas, problems and solutions relating to the multifaceted aspects of Electric and Electronics. Being crucial for the development of Electric and Electronics, our conference encompasses a large number of research topics and applications: from Circuits and Systems to Computers and Information Technology; from Communication Systems to Signal Processing and other related topics are included in the scope of this conference. In order to ensure high-quality of our international conference, we have high-quality reviewing course, our reviewing experts are from home and abroad and low-quality papers have been refused. All accepted papers will be published by Lecture Notes in Electrical Engineering (Springer). EEIC 2011 is sponsored by Nanchang University, China. Nanchang University is a comprehensive university which characterized by "Penetration of Arts, Science, Engineering and Medicine subjects, Combination of studying, research and production". It is one of the national "211" Project key universities that jointly constructed by the People's Government of Jiangxi Province and the Ministry of Education. It is also an important base of talents cultivation, scientific researching and transferring of the researching accomplishment into practical use for both Jiangxi Province and the country. Welcome to Nanchang, China. Nanchang is a beautiful city with the Gan River, the mother river of local people, traversing through the whole city. Water is her soul or in other words water carries all her beauty. Lakes and rivers in or around Nanchang bring a special kind of charm to the city. Nanchang is honored as 'a green pearl in the southern part of China' thanks to its clear water, fresh air and great inner city virescence. Long and splendid history endows Nanchang with many cultural relics, among which the Tengwang Pavilion is the most famous. It is no exaggeration to say that Tengwang Pavilion is the pride of all the locals in Nanchang. Many men of letters left their handwritings here which tremendously enhance its classical charm. Noting can be done without the help of the program chairs, organization staff, and the members of the program committees. Thank you. EEIC 2011 will be the most comprehensive Conference focused on the various aspects of advances in Electric and Electronics. Our Conference provides a chance for academic and industry professionals to discuss recent progress in the area of Electric and Electronics. We are confident that the conference program will give you detailed insight into the new trends, and we are looking forward to meeting you at this worldclass event in Nanchang.
EEIC 2011 Organization
Honor Chairs Prof. Chin-Chen Chang Prof. Jun Wang
Feng Chia University, Taiwan Chinese University of Hong Kong, HongKong
Scholarship Committee Chairs Chin-Chen Chang Jun Wang
Feng Chia University, Taiwan Chinese University of Hong Kong, HongKong
Scholarship Committee Co- chairs Zhi Hu Min Zhu
IEEE IAS Nanchang Chapter, China IEEE IAS Nanchang Chapter, China
Organizing Co-chairs Jian Lee Wensong Hu
Hubei Normal University, China Nanchang University, China
Program Committee Chairs Honghua Tan
Wuhan Institute of Technology, China
Publication Chairs Wensong Hu Zhu Min Xiaofeng Wan Ming Ma
Nanchang University, China Nanchang University, China Nanchang University, China NUS ACM Chapter, Singapore
Contents
The Design and Implementation of DDR PHY Static Low-Power Optimization Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Wei Ge, Mengnan Zhao, Cheng Wu, Jun He Design on Triple Bi-directional DC/DC Converter Used for Power Flow Control of Energy Storage in Wind Power System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yanlei Zhao, Naiyong Xia, Housheng Zhang ESL Based SoC System Bandwidth Estimation Method . . . . . . . . Zhen Xie, Xinning Liu, Weiwei Shan, Wei Ge Cellular Automaton for Super-Paramagnetic Clustering of Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Zhang Botao, Zhang Shuqiang, Yu Zhongqiu Embedded Memory Wrapper Based on IEEE 1500 Standard . . . . Maryam Songhorzadeh, Rahebeh Niaraki Asli A SVPWM Control Strategy for Neutral Point Potential Compensation in Three-Level Inverter . . . . . . . . . . . . . . . . . . . . . . . . . . Jinsong Kang, Yichuan Niu Process Mining: A Block-Structured Mining Approach . . . . . . . . . Yan-liang Qu, Tie-shi Zhao China RoHS: How the Changing Regulatory Landscape Is Affecting Process Equipment Reliability . . . . . . . . . . . . . . . . . . . . . . . . Chris Muller, Henry Yu Building Process Models Based on Interval Logs . . . . . . . . . . . . . . . . Yan-liang Qu, Tie-shi Zhao
1
7 15
23 31
41 49
57 71
X
Contents
Fostering a Management Model of Librarians at Vocational College e-Libraries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ling-Feng Hsieh, Mu-Chen Wu, Jiung-Bin Chin
79
Extension 2D to 3D of FAN Transform in Image . . . . . . . . . . . . . . . . Fan Jing, Xuan Ying, Li Honglian
87
A Design of Low Cost Infrared Multi-touch System . . . . . . . . . . . . . Juan Wu, Yao-hui Hu, Guo-qiang Lv, Jing Yin
95
A Review of Research on Network-on-Chip Simulator . . . . . . . . . . 103 Haiyun Gu Design Methodology of Dynamically Reconfigurable Network-on-Chip . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 Haiyun Gu Numerical Simulation Study on a Flat-Plate Solar Air Collector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 Wenfang Li, Shuhui Xu, Haiguang Dong, Jun You Artificial Immune for Harmful Information Filtering . . . . . . . . . . . . 125 Yan Sun, Xue guang Zhou Identification and Pre-distortion for GaN-PA . . . . . . . . . . . . . . . . . . . 133 Yuanming Ding, Yan Wang, Akira Sano On-Line Monitoring System of Methane Reaction Generator . . . 141 Ma Baoji Security Flaws in Two RFID Lightweight Authentication Protocols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 Wang Shaohui Study on Mechanical Parameters in Finite Element Analysis of Children’s Orbital-Bone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157 Weiyuan Lu, Zhixiang Liu, Xiuqing Qian, Tingting Ning, Huagang Yan 3D Model Retrieval Based on Multi-View SIFT Feature . . . . . . . . 163 Shungang Hua, Qiuxin Jiang, Qing Zhong Current Issues and Future Trends in Analysis of Automotive Functional Safety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 Chunyang Mu, Xing Ma, Hongxing Ma, Xianlian Huang, Ling Zhang, Rong Fan Development of an Automatic Steering System for Electric Power Steering (EPS) System Using Fuzzy Control Theory . . . . . 179 Tao Hong, Xijun Zhao, Yong Zhai
Contents
XI
Design and Implementation of Interactive Digital Shadow Simulation System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 Qian Li, Qing-yi Hua, Jun Feng, Wei Niu, Hao Wang, Jie Zhong Weighing Machine Modifications for Purblind and Sightless People . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 Vladimir Kasik, Martin Stankus Active RFID Based Infant Security System . . . . . . . . . . . . . . . . . . . . . 203 Lin Lin, Nan Yu, Tao Wang, Changan Zhan Modified Segmentation Prony Algorithm and Its Application in Analysis of Subsynchronous Oscillation . . . . . . . . . . . . . . . . . . . . . . 211 Yujiong Gu, Dongchao Chen, Tiezheng Jin, Zhizheng Ren Group Control Strategy of Welding Machine Based on Improved Active Set Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219 Chen Shujin, Zhang Junjun Simulation and Experiment Research on the Effects of DC-Bias Current on the 500kV Power Transformer . . . . . . . . . . . . . 227 Feng-hua Wang, Jun Zhang, Cheng-yu Gu, Zhi-jian Jin A HID Lamp Model in Simulink Based on the Principle of Electric Arc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235 Xiaohan Guan, Zhongpeng Li Coordinated Control for Complex Dynamic Interconnected Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243 Xin-yu Ouyang, Xue-bo Chen Study on Calibration of Transfer Character of Ultrasonic Transducer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 Qiufeng Li, Quanhong Zhang, Min Zhao, Lihua Shi A Novel Non-contact Pulse Information Detection Method Based on the Infrared Sequence Images . . . . . . . . . . . . . . . . . . . . . . . . . 259 Weibin Zhou, Bin Jing, Dian Qu, Guihong Yuan, Chunyan Wang, Haiyun Li A Method of Neurons Classification and Identification . . . . . . . . . . 267 Xingfu Li, Donghuan Lv A Broadband Image-Rejection Sub-harmonically Pumped Mixer MMIC for Ka-band Applications . . . . . . . . . . . . . . . . . . . . . . . . 275 Fang-Yue Ma, Yang-Yang Peng, Xiao-Ying Wang, Wen-Quan Sui
XII
Contents
Study of Phase-Shift Laser Ranging on Travelling Crane Anti-collision System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283 Bisheng Cao, Zhou Wan, Zhongguo Jing, Xin Xiong Bearing Condition Monitoring and Fault Diagnosis of a Wind Turbine Using Parameter Free Detection . . . . . . . . . . . . . . . . . . . . . . . 289 Shenggang Yang, Xiaoli Li, Ming Liang An Approach of K-Barrier Coverage of WSN for Mine . . . . . . . . . 295 Qianping Wang, Liangying Wang, Rui Zhou, Dong Jiang Design and Implementation of Intrusion Detection System . . . . . 303 Tai-ping Mo, Jian-hua Wang Self Tuning PID Controller for Main Steam Temperature in the Power Plant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309 Amin Aeenmehr, Alireza Sina A Novel Transformerless Single-Phase Three-Level Photovoltaic Inverter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317 Xiaoguang Yang, Youhua Wang Image Reconstruction Algorithm Based on Fixed-Point Iteration for Electrical Capacitance Tomography . . . . . . . . . . . . . . . . 325 Cuihuan Li, Xiaoguang Yang, Youhua Wang Research on Data Preprocessing in Exam Analysis System . . . . . 333 Ming-hua Zhu The Development and Application of Environmental Art Project Which Based on Semiotics in Information Age . . . . . . . . . 339 Ke Zuo Rotor Time Constant Estimation for the Vector Controlled Induction Motor Drive Based on MARS Scheme . . . . . . . . . . . . . . . 345 Hua Li, Shunyuan Zhou Small-World Request Routing System in CDNs . . . . . . . . . . . . . . . . . 353 Lan Li Experimental Study on Simulated Cerebral Edema Detection with PSSMI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361 Gui Jin, Mingxin Qin, Chao Wang, Wanyou Guo, Lin Xu, Xu Ning, Jia Xu, Dandan Gao Medium Choice of Chinese Consumers in Obtaining Advertising Information about Minitype Automobile . . . . . . . . . . . 369 Dao-ping Chen
Contents
XIII
The Finite Element Modeling for Mechanical Feature Analysis of Human Lumbar L4-L5 Segment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379 Juying Huang, Haiyun Li, Hao Wu Hybrid Control Using Sampling PI and Fuzzy Control Methods for Large Inertia and Time Delay System . . . . . . . . . . . . . 387 Jia Xie, Shengdun Zhao, Zhenghui Sha, Jintao Liang An Operator Controllable Authentication and Charging Solution for Application Store . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395 Jianbing Xing, Zhaoxia Li, Xiongwei Jia, Zizhi Qiao A Modified Kalman Filter for Non-gaussian Measurement Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401 Muhmmad J. Mirza A Novel Framework for Active Detection of HTTP Based Attacks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411 Liang Jie, Sun Jianwei, Hu Changzhen Model Predictive Control for Single Phase Inverters . . . . . . . . . . . . 419 Yecheng Lv, Ningxiang Xie, Kai Wang Evaluation Method and Standard on Maintainability of Cockpit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427 Junwei Zhao, Lin Zhou, Haitao Zhao, Xuming Mao, Youchao Sun Smart Mobile User Adaptive System with Autonomous Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433 Ondrej Krejcar, Robert Frischer Reduction of Player’s Weight by Active Playing Using Motion Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439 Ondrej Krejcar, Dalibor Janckulik Use of Neural Networks Library for Material Defect Detection Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447 Ondrej Krejcar Multi-sensor Measurement Fusion via Adaptive State Estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455 Li-Wei Fong A Simple Automatic Outlier Regions Detection . . . . . . . . . . . . . . . . . 463 Kitti Koonsanit The Feature Parameters Algorithm of Digital Signal in Circuit Based on Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471 Xiaodong Ma, Guangyan Zhao, Yufeng Sun
XIV
Contents
Research of Stereo Matching Based on Improved Median Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479 Huiyan Jiang, Rui Gao, Xiaojie Liu Generating Method of Three-Phase Voltage Sag in EV Charging System Performance Testing . . . . . . . . . . . . . . . . . . . . . . . . . . 487 Xiaoming Yue, Hui Fan, Jin Pan, Xiaoguang Hao, Zhimeng Zhang Improved Predictive Control of Grid-Connected PV Inverter with LCL Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 497 Huijie Xue, Wei Feng, Zilong Yang, Chunsheng Wu, Honghua Xu Study on Application of Demand Prognosticating Model Based on Grey Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503 Han Qingtian, Li Lian, Zhang Yi Research on Bayes Reliability Assessment for Test Data . . . . . . . . 509 Han Qingtian, Li Lian, Cui Jia Improved Particle Swarm Optimization by Updating Constraints of PID Control for Real Time Linear Motor Positioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515 Ying-Hao Li, Yi-Cheng Huang, Jen-Ai Chao Small Form-Factor Driver for Power LEDs Powered with One Element Battery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525 Garate Jose Ignacio, de Diego Jose Miguel, Araujo Jose Angel The Empirical Analysis of the Impact of Industrial Structure on Carbon Emissions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533 Mei Liao, Tianhao Wu An Ant-Routing Algorithm for Wireless Sensor Networks . . . . . . 541 Pengliu Tan, Xiaojun Deng Transient Sensitivity Computations for Large-Scale MOSFET Circuits Using Waveform Relaxation and Adaptive Direct Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 549 Chen Chun-Jung Incremental Circuit Simulation for Large-Scale MOSFET Circuits with Interconnects Using Iterated Timing Analysis . . . . 557 Chen Chun-Jung Inversion of Array Lateral-Logging Based on LSSVM . . . . . . . . . . . 563 Yu Kong, Li Zhang, Linwei Feng, Yueqin Dun
Contents
XV
Voice Recognition Based on the Theory of Transmission Wave and LSSVM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569 Yu Kong, Yue Min, Jing Xiao Design of a Two-Phase Adiabatic Content-Addressable Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 577 Meng-Chou Chang, Yen-Ting Kuo Branch Importance Assessment under Cut-Off Power Flow Based on EM Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 585 Feng Yuan, Wang Li-ming, Xia Li, Bu Le-ping, Shao Ying Knowledge Discovery of Energy Management System Based on Prism, FURIA and J48 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 593 Feng Yuan, Xia Li, Wang Li-ming, Pu Le-ping, Shao Ying Study on the Medium-Term and Long-Term Forecast Technology of Wind Farm Power Generation . . . . . . . . . . . . . . . . . . . 601 Yang Gao, Li Liu, Guoyan Liang, Shihai Ma, Chenwei Tian Optimal Space Vector Modulation Control for Three-Phase Inverter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 609 Su Xiaofang, Rao Binbin, Li Shijie, Zheng Ruilan Optimal Dead-Time Elimination for Voltage Source Inverters . . . 617 Su Xiaofang, Rao Binbin, Zeng Yongsheng Research of the Influence Factors in Single-Phase Inverter Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 625 Xiaofang Su, Binbin Rao, Chong Chen, Junbo Liu The Inspiration of the Economic Restructuring in Ruhr of Germany to the Sustainable Development of Mining Cities in Henan Province of China . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633 Wang Liping Design of Metal Pipe Defect Detection Device Based on Electromagnetic Guided Wave . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 639 Shuwang Chen, Deliang Li, Zhangsui Xu Implement of Explosion-Proof and Intrinsic Safe Model Mult-protocol Converter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 645 Ming-san Ouyang, Cheng-jie Zhu, Zhe Liang The Application in Mobile Computing of Spatial Association Rules Mining Algorithm Based on Separating Support Items . . . 651 Gang Fang
XVI
Contents
Automatic Test Case Generation for Web Applications Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 657 Lulu Sun, Junyi Li, Shenglan Liu Design of Holmium Laser Treatment Instrument Control System Based on STM32 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 667 Youjie Zhou, Chunhua Xiong, Changbo Lu Research on Attribute Granular Computing and Its Fuzzification Method Based on Qualitative Mapping . . . . . . . . . . . . 673 Ru Qi Zhou, Yi Lin Wu, Yi Qun Chen A CT Image Denoise Method Using Curvelet Transform . . . . . . . 681 Junmin Deng, Haiyun Li, Hao Wu Research System Optimization of the Wireless Mesh Networks Based on WLAN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 689 Jin Wen Stabilization Control of Chaotic System Based on LaSalle Invariant Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 697 Chunchao Shi, Yanqiu Che, Jiang Wang, Xile Wei, Bin Deng, Chunxiao Han Parameter Estimation in a Class of Chaotic System via Adaptive Steady State Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 705 Chunchao Shi, Yanqiu Che, Chunxiao Han, Jiang Wang, Bin Deng, Xile Wei Algorithms of 3D Segmentation and Reconstruction Based on Teeth CBCT Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 713 Wenjun Zhang A Calculation Method for Series Lagging Correction Based on Root Locus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 721 Li Wang The Computer Simulation and Real-Time Control for the Inverted Pendulum System Based on PID . . . . . . . . . . . . . . . . . . . . . . 729 Yong Xin, Bo Xu, Hui Xin, Jian Xu, Lingyan Hu Application of Fractal Dimension in Analysis of Soil Micro Pores . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 737 Shu-hua Zhang Assessment of Loess Collapsibility with GRNN . . . . . . . . . . . . . . . . . 745 Shu-hua Zhang
Contents
XVII
Grey Comprehensive Relational Analysis of Fighting Efficiency Influencing Factors of AEW Radar . . . . . . . . . . . . . . . . . . . 753 Guangdong Liang, Guangshan Lu, An Zhang, Yanbin Shi Diesel Misfire Fault Diagnosis Using Vibration Signal over Cylinder Head . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 761 Liu Li-tian, Liao Hong-yun, Chen Xiang-long, Feng Yong-min, Xiao Yun-kui Proof of a Conjecture about k-Graceful Graph . . . . . . . . . . . . . . . . . . 769 Li Wuzhuang, Yan Qiantai High Dynamic GPS Signal Analysis and Acquisition Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 773 Xiaoming Han, Guangyu Zheng, Shusheng Peng On the Mutual Information and Capacity of Coded MIMO with Interference Alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 781 Haitao Li, Haiying Yuan Interdisciplinary Education Model of Fashion Marketing and E-Commerce . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 789 Zhaoyan, Renli Research on the Image Registration Algorithm Based on Regional Feature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 795 Xiaoyan Cao, Xiangxin Shao, Xinying Li, Chunying Wang, Yan Zhou Accurate Curvature Approximation of 3-Dimension Discrete Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 803 Shilin Zhou, Jianping Yin, Xiaolin Yang, Junyun Wu The Research of a Symmetrical Component Method and Dynamic Reactive Power Compensation of Electric Arc Furnace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 811 Xuezhe Che, Fenglong Shen, Jianhui Wang Numerical Simulation of Electrokinetic Flow in a Nanotube with Variable Physical Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 819 Davood D. Ganji, Mofid Gorji-Bandpy, Mehdi Mostofi Physical Properties of Camellia Oleifera Heated by Micro-wave . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 827 Jian Zhou, Lijun Li, Yihua Hu, Ke Cheng, Zhiming Yang, Ye Xue Research on Trust-Based Dynamic Role Access Control Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 833 Dongping Hu, Guohua Cui, Aihua Yin, Liang Chen
XVIII
Contents
Design of Special Welding Machine Based on Open CNC System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 839 Haibo Lin Optimization Algorithm on the Intelligence Schedule of Pubic Traffic Vehicles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 845 Cui Yonghong, Wang Qingrong The Criterion of Intrinsic Safe Circuit Characteristic Based on Electric Arc Power . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 853 Ligong Wang Multiple View Locality Preserving Projections with Pairwise Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 859 Xuesong Yin, Qi Huang, Xiaodong Chen Analysis of E-SCM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 867 Huang Hua, Peng Cong Information Sharing of Partnership in Supply Chain and Enhancement of Core Competitiveness of Enterprises . . . . . . . . . . . 875 Huang Hua, Peng Cong Novel Spatial Diversity Equalizer Suitable for 16-QAM Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 885 Rao Wei Fuzzy Logic Controller for a Pneumatic Rotary Actuator . . . . . . . 893 Logah Perumal Evolutionary Neural Network Based on Immune Continuous Ant Colony Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 903 Gao Wei Reliable and Delay-Efficient Routing Algorithm in Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 911 Mohammad Ali Jabraeil Jamali, Adel Fathi, Javad Pashaei Interfacial Impedance Sensor Employing Bio-activated Microbeads and NiHCF-Coated Interdigitated Microelectrodes: A Model Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 921 Nuno M.M. Pires, Tao Dong, Zhaochu Yang, Lei Zhang An Improved Fp-Tree Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 929 Haijun Zhang, Changchang Zhang, Bo Zhang
Contents
XIX
A Decoupling Method for Evaluating Lightning-Induced Overvoltages on One Certain Line of an Overhead Transmission Lines System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 937 Dongming Li, Cheng Wang, Xiaohu Liu Power Flow Calculation in the Voltage Output Stage . . . . . . . . . . . 945 Sun Qiuye, Li Zhongxu, Ma Dazhong, Zhou Jianguo Research on Trajectory Control of Polishing Robot Based on Secondary Deceleration and Dynamic Model . . . . . . . . . . . . . . . . . . . 953 Tongying Guo, Languang Zhao, Haichen Wang Security of Industrial Control System . . . . . . . . . . . . . . . . . . . . . . . . . . . 959 Peng Jie, Liu Li The Application of Automatic Control Theory in Software Engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 965 Wensong Hu, Xingui Yang, Min Zhu, Yan Zhang Portable MIT-BIH Physiological Signal Playback System Based on Android . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 969 Wensong Hu, Li Ming, Min Zhu, Linglin Xia The Application of Network Model on Power Station . . . . . . . . . . . 975 Min Zhu, Wensong Hu, Yan Zhang The Application of Algorithm in Power Station . . . . . . . . . . . . . . . . . 981 Min Zhu, Wensong Hu, Yan Zhang The Application of IP/UDP Protocols in 51 Microcontrollers . . . 987 Wensong Hu, Yuyuan Zhu, Min Zhu, Ke Zuo A Collaborative Filtering Recommendation Algorithm Based on User Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 993 Daner Chen The Personalized Recommendation Algorithm Based on Item Semantic Similarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 999 Yulong Ying Application Research of Four-Branch Antenna Design by Improved Orthogonal Multi-objective Evolutionary Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1005 Jincui Guo, Xiaojuan Zhao, Jinxin Zou
XX
Contents
Research on Aviation Forewarning Systems Based on Business Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1013 Liu Yongjun, Hu Qing, Ruan Wenjuan The Need for Teleradiology System in Medical Remote-Diagnosis Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1021 Mohammad H. Al-Taei, Ziyad T. Abdul-Mehdi, Subhi H. Hamdoon The Clustering Methods Based on the Most Similar Relation Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1027 Yan Yu, Wei Hong Xu, Yu Shan Bai, Min Zhu Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1035
The Design and Implementation of DDR PHY Static * Low-Power Optimization Strategies Wei Ge1, Mengnan Zhao1, Cheng Wu1, and Jun He2 1 National ASIC System Engineering Research Center, Southeast University, Nanjing 210096 People's Republic of China {duiker,zmn,wcccc}@seu.edu.cn 2 Huawei Technol. Co., Ltd., Nanjing, China
[email protected] ,
Abstract. The static power of DDR PHY has increasingly become the limit of the low-power application of system-on-a-chip (SoC). An optimization of static power based on "behavior" and "state" of DDR PHY static power is proposed, considering the design principle and physical properties. Experimental results show that the proposed optimization strategy can achieve the highest 59.12% reduction in work mode and only 0.723uW power consumption in sleep mode. Keywords: DDR PHY, DDR PAD, Static Power, UPF.
1 Introduction The power consumption of DDR controller has been drawn more and more attention while providing faster transfer rate and higher data bandwidth. At present, low power consumption of DDR PHY research has the following two aspects. 1) Improve the control logic to minimize the number of row opens and closes, and lower the energy consumption during read/write operations [1]. Reshape the memory traffic to coalesce short idle periods into longer ones, thus enabling existing techniques to effectively exploit idleness in the memory [2]. 2) Improve the design of DDR PHY by using the direct clock pulses to latch the data, without a need for additional pulse generator circuitry for the clock signal, which lowers the clock dynamic power consumption by factor of 2x [3]. Employ a clock branch-sharing scheme to reduce the number of clocked transistors in the design. The newly proposed design also employs conditional discharge and split-patch techniques to further reduce switching activity and short-circuit currents, respectively [4]. The research above optimized the access method and the structure design, but did not consider the DDR PHY static power of activities and non-activities. We propose innovatively low-power strategies based on "behavior" and "state" which effectively reduce the DDR PHY static power consumption of activities and nonactivities. *
This work was sponsored by the National Scientific Foundation of China (Grant No. 61006029) and Jiangsu Scientific Foundation (Grant No. BK2010165).
M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 1–6. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
2
W. Ge et al.
2 Design Principle In order to reduce the DDR PHY static power consumption in work mode and support the ultra-low-power mode, both the DDR PAD and the DDR PHY need to be optimized according to different "behavior" and "state". 2.1 Principle of DDR PHY Low-Power Based on Behavior DDR PHY does the jobs of exchanging data with external chip, multilevel conversion, DQS phase shift, etc., and its static Power consumption is mainly comprised of Leakage Power, which does not change with frequency. Static power consumption formula is expressed as:
Psta = Vdd * I SUB
,which I
SUB
W =C V e L 2 ox th
VGS −VT nVth
(1)
From the formula (1), the static power would be close to zero by cutting off the Vdd, thus greatly reducing the static power consumption of DDR PHY.
Fig. 1. Optimization design of DDR PAD
DDR PAD shows high static power consumption under no flipping circumstances. Therefore, the control signals OutputShutDown (OSD) and InputShutDown (ISD) are added to cut off the DDR PAD output and input leakage paths respectively, as shown in figure (1). So that, the leakage current will decrease to nearly zero. 2.2 Principle of DDR PHY Low-Power Based on State The UPF language provides a new way to specify the power requirements of a design and specifies how to create a power supply network to each design element, the behavior of supply nets with respect to each other, and how the logic functionality is extended to support dynamic power switching to design elements.
The Design and Implementation of DDR PHY Static Low-Power Optimization Strategies
3
In order to switch chip from sleep mode to work mode rapidly, the DDR controller should be special designed. "Sub Power Domain" need to be controlled respectively, as shown in figure (2), DDR PHY Domain (DDRPD) and DDR Retention Domain (DDRRD) separated from the power network on chip are provided different power domain optimization methods by the controller in the Always on Domain (AOD).
Fig. 2. DDR PHY power domain design base on UPF
It makes ultra-low-power available by designing different working states for each power domain: 1. NORMAL: each power domain works normally. DDR PHY reduces power consumption through frequencies adjusting. 2. STOP: the kernel of SoC is powered down, and DDRPD power domain works normally, meanwhile the DLL of DDR PHY (Delay Locked Loop) works in Bypass Mode. 3. SLEEP: only AOD and DDRRD power domain are active, DDR PHY offers the least control signals, to ensure the data integrity of the external memory chip.
3 Implementation of Low-Power Design 3.1 DDR PAD Static Optimization Strategy Based on Behavior Based on the design of the DDR interface timing specified in JEDEC specifications, we propose various optimization strategies for different behavior (command, data) operations, which control the switch of DDR PAD static leakage path. For the unidirectional command signals, CMD_OSD always shut down the input channel, and the output path is switched by the CMD_ISD signal when the command is valid. For the bidirectional data signals, DATA_OSD and DATA_ISD make effect only in the valid data operating phase. In order to realize these functions, low-power control logic is added to generate the control signals (OSD and ISD), which make relative responses according to each DDR operation such as read, write, refresh, activate, etc. At the same time, considering the recovery time of PADs after enabling OSD and ISD, there must be enough margin to satisfy the various frequencies requirement in different working states. Figure (3) shows the HSIM simulation result of OSD and ISD under the condition of tsmc 65nm standard library (1.2v Core Vdd, 1.8v PAD Vdd, 25°C). After the validation of OSD and ISD, the PAD current between the power (vmvddq) and ground (vmvssq) changes correspondingly. When the OSD and ISD are valid, the data transfers between
4
W. Ge et al.
the chip external and the pad internal. On the contrary, the data path is been cut off. There is a slight current disturbance, which will not impact the function.
Fig. 3. HSIM simulation of DDR PAD
3.2 DDR PHY Power Management Based on State In NORMAL mode, the DDR PHY works normally in SoC. In STOP and SLEEP mode, the external memory chip enters the Self Refresh or Power Down mode, and the DDR PHY need guarantee the control signals (CKE, RESETB) valid to avoid data loss. The kernel of SoC is powered down in STOP mode, so the registers in the DDR PHY are used for latching the command and data signals. In SLEEP mode, except the AOD and DDRRD power domains, the logic of the DDR PHY controller and most of the IO PADs are powered down. For this reason, the signal control and isolation should be taken into consideration in the DDRRD power domain. The detail design of DDRRD is shown in Figure (4). The independent power supply network is separated by ISOLATION PAD, so the DDRRD can work without being influenced when other parts of the chip are powered down. The RETLEC PAD provides internal control signal in SLEEP mode, which automatically lapses when the chip wakes up. The CKE PAD and RESETB PAD provide the necessary control signals for the external memory chip.
Fig. 4. PAD placement of DDR Retention Domain
The Design and Implementation of DDR PHY Static Low-Power Optimization Strategies
5
4 Design Result The DDR PHY static power consumption optimization strategies, based on “behavior” and “state”, can significantly reduce power consumption in work and standby modes. By using tsmc 65nm standard library and Prime Time PX power analysis tools, we can obtain DDR PHY power consumption simulation results in Table (1). The available work modes are defined as follows: SLEEP: the power domains except the AOD and DDRRD are powered down and the chip is in “sleep” state. STOP: the kernel is powered down, only the necessary interrupt applications are reserved to wake up the system. IDLE: the kernel clock is gated, and the chip enters the standby state. NOMAL: chip works normally, typical applications of computing and display work properly. FASTHOT: chip works in high speed, most of the scientific computing, image processing and display applications occupy the memory bandwidth as much as possible. Table 1. DDR PHY power consumption comparison in defferent mode
STATE SLEEP STOP IDLE NORMAL FASTHOT
DDR_PHY_OLD(mW) NULL NULL 227.73 314.84 385.92
DDR_PHY(mW) 0.000723 3.6 92.95 218.4 312
Optimization(%) NULL NULL 59.18 30.63 19.15
Through the above comparison of power consumption simulation in different modes, we can see that the power optimization of the DDR PHY changes antilinear with increasing of the memory access load. The power optimization ratio in FASTHOT mode and IDLE mode are up to 19.15%and 59.18%. Through the low-power domain design based on the “state”, power consumption on DDR PHY just reaches 3.6mW in STOP mode, and ultra-low-power of 0.723uW in SLEEP mode.
5 Conclusions This paper proposed static low-power optimization strategies based on “behavior” and “state”, which effectively reduce the power consumption of DDR PHY in both working state and sleep state. The simulation results show that the strategies not only improve the DDR PHY power characteristic in different operation modes (19.15%~59.18%), but also provide ultra-low-power application methods for SoC (only 0.723uW in SLEEP mode).
6
W. Ge et al.
References 1. Dongwook, L., Sungjoo, Y., Kiyoung, C.: Entry control in network-on-chip for memory power reduction. In: International Symposium on Low Power Electronics and Design (ISLPED), pp. 171–176. ACM/IEEE (2008) 2. Hai, H., Shin, K.G., Lefurgy, C., Keller, T.: Improving energy efficiency by making DRAM less randomly accessed. In: International Symposium on Low Power Electronics and Design, pp. 393–398 (2005) 3. Devarapalli, S.V., Zarkesh-Ha, P., Suddarth, S.C.: A robust and low power dual data rate (DDR) flip-flop using c-elements. In: 11th International Symposium on Quality Electronic Design, pp. 147–150 (2010) 4. Peiyi, Z., McNeely, J., Golconda, P., Bayoumi, M.A., Barcenas, R.A., Weidong, K.: LowPower Clock Branch Sharing Double-Edge Triggered Flip-Flop. IEEE Transaction on Very Large Scale Integration (VLSI) Systems 15, 338–345 (2007)
Design on Triple Bi-directional DC/DC Converter Used for Power Flow Control of Energy Storage in Wind Power System* Yanlei Zhao, Naiyong Xia, and Housheng Zhang School of Electrical and Electronic Engineering, Shandong University of Technology Zhangdian District, Zibo, Shandong Province, China
[email protected] Abstract. A triple bi-directional DC/DC converter used for power flow control of energy storage in wind power system is designed. Firstly, the paper analyzes energy flow characteristic of wind power system containing hybrid energy storage (the battery and the supercapacitor). Secondly, working principle of the triple bi-directional DC/DC converter is explained and mathematical model is constructed. Then, based on the model, control unit consisting of voltage outer loop and current inner loop is designed. To ensure the consistency of working state of batteries in parallel, a novel current-sharing strategy is proposed based on battery SOC (state of charge). To realize energy distribution between the battery and the supercapacitor, a one-order low pass filter is used to detach the low frequency component in feedback voltage. The simulation results show that the converter can effectively regulate the load voltage and reasonably distribute the power between both energy storage elements. Keywords: wind power system; energy storage; power flow optimization; bidirectional DC/DC converter.
1 Introduction The ransom fluctuation of output power of wind power system changes grid power flow frequently in terms of direction and size, which will lead to voltage instability, frequency fluctuation and some other power quality problems, threaten the security and stability of power system. Consequently, wind power penetration level in power grid is confined. Equipping with certain amount of energy storage in wind power system can resolve the problems mentioned above. Battery energy storage system has been widely used in renewable energy sources power system, which can smooth power fluctuation of the generation system, enhance dispatching ability of the generation system, improve static and dynamic characteristic of the grid[1]. However, the battery, if solely used in wind power system, has its inherent shortcomings: Firstly, the battery charges and discharges continually, which reduces *
Project Supported by National Natural Science Foundation of China (50807034).
M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 7–14. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
8
Y. Zhao, N. Xia, and H. Zhang
cycle life of the battery; secondly, the battery has small power density and low dynamic response speed, so the battery has no method to compensate many dynamic power quality problems occurring in short time span. The supercapacitor, as a new fast energy storage element, has high power density and long cycle life[2]. The supercapacitor, combinated with the traditional battery and used for power conditioning or power flow optimization in wind power system, can take full advantage of the battery’s superiority of high energy density and the supercapacitor’s superiority of high power density. As a result, energy storage unit in wind power system can be optimized in respects of its performance and cost. The combination of the supercapacitor and the battery generally has three patterns, including direct parallel connection, parallel connection through the inductor and parallel connection through the power converter (DC-DC converter)[3]. With regard to the first two hybrid energy storage types of direct parallel connection and of parallel connection through the inductive, their port voltages vary in wide range in the process of charging and discharging. More adverse is, the energy distribution between two energy storage elements can not be controlled flexibly, so that it’s difficult to fully play the both ones’ respective advantages. While the energy storage unit of parallel connection through power converter can easily control energy flow between both energy storage elements, so the capacity and the performance of energy storage unit can be flexibly optimized. Power converters used in the hybrid energy storage system can be divided into two types: the one of unilateral energy flow and the one of bidirectional energy flow. For power flow optimization and control of wind power system, the energy flow between energy storage unit (DC link) and the public power grid (AC Link) is two-way, correspondingly, the current of the DC-DC converter used in hybrid energy storage unit is also bidirectional. On the basis of the energy flow characteristic in wind power system containing power flow optimization by virtue of hybrid energy storage, the paper designs a triple bi-directional DC/DC converter used for energy flow control of the hybrid energy storage unit.
2 Energy Flow Characteristic in Wind Power System Containing Power Flow Optimization The ultimate goal of power flow optimization of wind power system is to restrain the power fluctuation, which is based on the balances of the active power as well as the reactive power. Therefore, the fundamental functions of power flow optimization system have two: one is to smooth the active power flowing into the grid in real time; another is to compensate reactive power flowing into the grid in real time. According to these basic functions, the overall layout of the wind power system containing power flow optimization is shown in Figure 1. Suppose, in the whole wind power system, output power of wind generation is denoted as PWG ; charging( or discharging) power of energy storage unit is denoted as PS , of which, charging( or discharging) power of the supercapacitor is denoted as
PC , the battery PB ; the power flowing into the grid is denoted as PG . Because of
Design on Triple Bi-directional DC/DC Converter
9
randomness and uncertainty of wind energy, PWG is obviously randomly fluctuant. If, without any compensation,
PWG is directly transported into the grid, the voltage, the
frequency and some other parameters of the local grid are will be influenced negatively. Fluctuations components in wind energy of above 1Hz can be absorbed by the inertia of generation system, while the components of 0.01Hz ~ 1Hz have the maximum impact on the grid in terms of the voltage, the frequency and other performances[4]. And the ones of below 0.01Hz directly affect the capacity reliability of generation system. To restrain the negative impact of wind power, its power flow optimization system should smooth the active power injected into the power grid. That is to say, of the output power of the wind generation system, only the components of relatively low frequency(lower than 0.01Hz) is transported into the public grid, while the higher frequency components should be absorbed by the energy storage. Apparently that the power absorbed by energy storage is PS = PWG − PG , which may be flexibly controlled by a four-quadrant DC-AC converter. Power Grid Grid Voltage
Wind Power System output current
Wind Power System
Wind Power System gridconnected Converter
Power Flow Optimization Control Strategy
Controller)
Control Strategy for Power Flow of Energy Storage
Battery State of Charge(SOC)
Storage Battery
Four-quadrant converter (Power Flow Optimization
Bi-direction DC converter Power Flow of Energy Storage
DC Voltage
Super-Capacitor
Fig. 1. The overall layout of the wind power system containing power flow optimization
Of two energy storage elements, the supercapacitor has high power density and lower energy density; while the battery has low power density, high energy density and short cycle life. Hence, in the process of power flow optimization, the charging (discharging) times of the battery should be reduced to less as far as possible. That is,
10
Y. Zhao, N. Xia, and H. Zhang
the relatively lower frequency components of the wind power are taken by the battery, while the remaining higher frequency components are borne by the supercapacitor. The energy between the two energy storage devices is controlled by the bi-directional DC-DC converter designed by the paper.
3 Design on Triple Bi-directional DC/DC Converter 3.1 Working Principle and Small-Signal Model In the wind power system of comparatively large power, the batteries generally need to be paralleled to meet the power demand. Due to individual differences in feature, the charging (discharging) current of various sub-unit may be different, which easily lead to the problems of over-charge or under charge. Obviously, the number of directly parallel individuals is larger, the imbalance problem is more serious. To reduce the number of batteries directly in parallel and to ease the charging (discharging) imbalance problem, in this paper, the triple two-way DC-DC converter is adopted, which can decrease the number of batteries directly in parallel and flexibly control charging (discharging) currents of the three groups of batteries. At the same time, multiple combinations of the conversion circuit can increase the transmission power and reduce current ripple as well as the filter components (inductors, capacitors) parameters. The topology of the converter is shown by Fig. 2.
vo v1i
v 2i
v 3i
Fig. 2. The topology of the triple two-way DC-DC converter
The control of the converter is based on PWM. The switches S1~S6 and D1~D6 are composed of power electronics device such as IGBT and diode. If the switches S1, S3, S5 operate in PWM mode, and the switches S2, S4, S6 are always off, the converter runs in Buck chopper mode. While the switches S2, S4, S6 operate in PWM mode, and the switches S1, S3, S5 are always off, the converter runs in Boost chopper mode. When the switching state of S1, S3, S5 and the ones of S2, S4, S6 are opposite each other, and all operate in PWM mode, the converter is in the unified working mode, i.e. Buck-Boost chopper. In a switching cycle of Buck-Boost chopper, its inductor current may be two-way, which can increase response speed[5], so the third mode is chosen in the paper.
Design on Triple Bi-directional DC/DC Converter
11
Because the topology and the control method of three sub-units are identical in the triple converter, as for modeling and analysis, only an independent sub-unit is taken into account. An independent bi-directional DC-DC converter sub-unit is shown in Fig. 3.
L
vo
S1
i
C
S2
vi
R
Fig. 3. An independent bi-directional DC-DC converter ∧
L
Vo d (t ) −
+
D′ :1
+
vˆi (t )
+
∧
i (t )
Idˆ (t )
C
R
vˆo (t ) −
Fig. 4. Small signal AC equivalent circuit model of Buck-Boost
In Fig. 3, vi , i , vo and R respectively stand for the battery voltage, the battery current, the supercapacitor voltage and the equivalent load resistance. By virtue of average switch modeling method, the small signal AC linear equivalent circuit model of Buck-Boost can be obtained and shown as Fig. 4. In Fig. 4, d (t ) is the duty ratio of the bottom switch (S2), D is the value of ∧
d (t ) in quiescent operation point, and d (t ) is the perturbation of d (t ) . ∧
Based on the linear equivalent circuit, the transfer function of duty cycle
d ( s ) to
∧
output voltage
vo ( s) Gvd ( S ) =
vˆo ( S ) dˆ ( S ) vˆi ( S )=0
=
D′Vo (1 − LCS 2 +
LS ) RD′2
L S + D′2 R
(1)
3.2 Control Strategy According to the overall structure of wind power system with energy storage, the energy exchange between power flow optimization system and the public grid can be reflected in the fluctuation of the DC bus voltage. If the relatively low frequency components of PS is expected to be absorbed by the battery, the lower frequency
12
Y. Zhao, N. Xia, and H. Zhang
components of the DC bus instantaneous voltage should be detached by a low pass filter and be stabilized to a given value. Considering response speed, an one-order low pass filter is chosen in the paper. To ensure that the battery absorbs the relatively lower frequency energy fluctuation, firstly, the voltage regulation control loop for the low frequency components of transient DC bus voltage is necessary, which ensure super-capacitor voltage is constant at the low frequency. To guarantee the charge state of three paralleled battery groups are identical, this paper proposed a current-sharing strategy based on the battery SOC (state of charge). For the battery, SOC
=
QC
QI
, where: QC is the residual capacity,
QI is the rated
capacity under a constant discharge current. In the process of current-sharing, for an independent sub-unit of the triple converter, the SOC of the corresponding battery group is higher, its average inductor current is larger, Conversely, the SOC lower, the current smaller. As a result, this can make the SOC of three battery groups tend to no difference. On the basis mentioned above, the paper proposes the control strategy of triple bidirectional DC-DC converter consisting of voltage outer loop and current inner loop, whose structure is shown as Fig. 5.
I ref
1 SOC1 I ref ∑ SOCi
LS ) RD′2 L LCS 2 + S + D′2 R D′Vo (1 −
kip s + kii s
i =1, 2,3
ki * U dc
kup s + kui s
I ref
3 I ref
SOC3
∑ SOC
i =1, 2,3
LS ) RD′2 L LCS 2 + S + D′2 R D′Vo (1 −
kip s + kii s
i
ki 1 1+τ 2s
ku
Fig. 5. The control structure of triple bi-directional DC-DC converter
According to [6], the parameter of the regulator
kip , kii , kup and kui can be
determined.
4 Simulation Results Based on Matalab / Smulink, the triple bi-directional DC converter is simulated. In the converter, the nominal voltage of the battery is 72V, the expected voltage of the supercapacitor (load voltage) is 200V, kip = 10 , kii = 60 , kup = 6
Design on Triple Bi-directional DC/DC Converter
and
13
kui = 10 , SOC1 = SOC2 = SOC3 , τ 2 = 0.01 . When the equivalent load
current (discharging current) varied in the case shown by
Fig. 6. (A) Charging current of the supercapacitor (B) Real voltage of the supercapacitor
Fig. 7. The currents of three sub-units of triple bi-directional DC-DC converter
Fig. 6 (A), under the control strategy proposed by the paper, the real-time voltage of the supercapacitor is illustrated as Fig. 6 (B), and the currents of three sub-units are shown as Fig. 7. Seen from the two figures, the relatively lower frequency components of load current (power) are undertaken by the battery, while the remaining higher frequency
14
Y. Zhao, N. Xia, and H. Zhang
components are undertaken by the supercapacitor. The simulation results indicate that the converter can not only reasonably distribute the power between two energy storage elements but effectively regulate the load voltage.
5 Conclusions The design on a triple bi-directional DC/DC converter used for power flow control of energy storage in wind power system are discussed. The paper designs control unit of the converter with voltage outer loop and current inner loop, which includes a novel current-sharing strategy based on battery SOC and uses a one-order low pass filter to detach the low frequency component in feedback voltage. The simulation results show that the converter can effectively regulate the load voltage and reasonably distribute the power between both energy storage elements with comparatively good static and dynamic characteristic.
Acknowledgments This paper and its related research are supported by National Natural Science Foundation of China (50807034) and Support Program for Young Teachers Development of Shandong University of Technology.
References 1. Yang, Z., Shen, C., Zhang, L., et al.: Integration of a StaCom and Battery Energy System Storage. IEEE Trans on Power System 16(2), 254–260 (2001) 2. Andrew, B.: Ultracapacitors: Why, How, and Where is the Technology. Journal of Power Sources 91, 37–50 (2000) 3. Tang, X., Qi, Z.: Study on the Ultracapacitor/Battery Hhybrid System. Chinese Journal of Power Sources 30(11), 933–936 (2006) 4. Luo, C., Ooi, B.-T.: Frequency Deviation of Thermal Power Plants Due to Wind Farms. IEEE Trans. on Energy Conversion 21(3), 708–716 (2006) 5. Wang, Z., Liu, J.: Power Electronics Technology. China Machine Press (May 2009) 6. Xu, D.: PowervElectronics System Modeling and Control. China Machine Press (January 2006)
ESL Based SoC System Bandwidth Estimation Method* Zhen Xie, Xinning Liu, Weiwei Shan, and Wei Ge National ASIC System Engineering Research Center, Southeast University, Nanjing 210096 People's Republic of China {xz,xinning.liu,wwshan,duiker}@seu.edu.cn
,
Abstract. The increasing complexity of system-on-a-chip (SoC) design is challenging the design engineers to estimate the system bandwidth. A method of SoC system bandwidth estimation based on electronic-system-level (ESL) is proposed, which estimates the system bandwidth by transaction-level-modeling (TLM) and analysis depended on simulation. Compared with the traditional RTL simulation, the estimation is effective and the simulation speed is more than two orders of magnitude faster. Keywords: ESL, SoC, TLM, Bandwidth estimation.
1 Introduction Bandwidth is always one of the bottlenecks in system-on-a-chip (SoC) systems. The increasing complexity of SoC design is challenging the design engineers to estimate the bandwidth. If the bandwidth is unable to well fit the requirement, the designer can modify the architecture in the early stage of the product development cycle, thus reducing the potential risk of system re-design [1]. In [2], the authors used statistical models to calculate execution cycles for bus traffic analysis and system performance estimation. However, the main drawback of static analysis is the lack of dynamic analysis information, such as bus contention, arbitration, dynamic scheduling, etc [3]. So some designers explore architecture by dynamic simulation at register-transfer-level (RTL). While this was possible for designs that were relatively simple, exploring complex SoC designs at RTL is an intimidating prospect. The speed of RTL simulation is too slow to allow adequate coverage of the large design space. Besides, making small changes in the design will probably require considerable re-engineering effort due to the complex nature of these systems [4]. To overcome these limitations, system designers should raise the level of abstraction, which gives an early estimation of the system characteristics before committing to RTL development. In this paper, a method of SoC system bandwidth estimation based on electronicsystem-level (ESL) is proposed, which estimates the system bandwidth by transaction-level-modeling (TLM) and analysis depended on simulation. In Section 2, basic principles of ESL design methodology are introduced. In Section 3, the process that *
This work was sponsored by the National Scientific Foundation of China (Grant No. 61006029) and Jiangsu Scientific Foundation (Grant No. BK2010165)
M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100,. pp. 15–21. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
16
Z. Xie et al.
using the method based on ESL to model the target system is described. In Section 4, the analysis depended on simulation and the bandwidth estimation of the target system are given. Section 5 is the conclusion.
2 Principles of ESL Design Methodology In current SoC design, the model established at the level of algorithm-and-function (ALF) is lack of timing information, and has little relationship with the system structure and implementation, which cannot be used to estimate the system performance. While the RTL model needs to focus on details of signal processing and design implementation, the speed of modeling and simulation is slow. ESL design methodology introduces a new level, transaction level, between ALF and RTL to abstract systems. Transaction-level-modeling (TLM) provides software and hardware engineers a virtual platform for the exploration of architecture and embedded software development.
Fig. 1. Comparison of the three modeling methods
ESL design methodology enables SoC designers to describe the system abstractly, with no need to prematurely be involved in the specific of implementation. Through ESL design methodology, designers can model the system in the early stage, explore the system architecture and guarantee the advancement. So that the risk of re-design due to performance cannot meet the requirement is reduced. In this paper, the architects view TLM is proposed, which has enough timing information and is mainly used for architecture exploration. Modules of the system will be abstracted as bus, memory or device to simulate system load. This method considerably reduces the workload required in the modeling, and greatly speeds up the system simulation.
ESL Based SoC System Bandwidth Estimation Method1
17
Fig. 2. Architects view TLM
3 Modeling of the Target System In this paper, ESL design methodology is used to analyze the performance, estimate the bandwidth by TLM on the ESL design platform, and optimize the architecture during the research to meet the requirement. 3.1 General Situation of the Target System The block diagram of the target system is shown at Fig. 3. 1. The system uses DDR2-800 as memory, data bus width is 32-bits, main timing parameters are CL=4, tRCD=4; 2. The interface frequency of AXI based LCDC, VPU and DDR2 is 200 MHz; 3. LCDC supports resolution up to 1920x1080 at refresh rate of 60Hz, and has three overlays, considering the worst case, bandwidth demand is about 1200 MB/s; 4. VPU accesses data every 4085ns, bandwidth demand is up to 900 MB/s, of which 700 MB/s is read access and 200 MB/s is write access. A rough estimate, LCDC and VPU consume 65% of the peak bandwidth. However, because of the complexity of access to DDR2, evaluation for the bandwidth is difficult. DDR2 storage can be divided into bank, row and column. According to the current and previous access addresses of the bank and row, access to DDR2 can be classified as continuous access and with-precharge access. Time spans of the two types of access are different. LCDC and VPU simultaneously request transaction whose required address is a random variable. On the other hand, bus transfer and arbitration will consume some bandwidth. So in fact, the bandwidth consumed by the two modules will significantly exceed 65%. To evaluate the bandwidth effectively, a correction factor must be involved.
18
Z. Xie et al.
Fig. 3. The block diagram of the target system
However, due to the randomness of data access, evaluation of the factor is a great challenge. Using ESL design methodology, modeling the target system abstractly, and depending on the ESL design platform, the factor and the system bandwidth can be obtained accurately. 3.2 Modeling of the Target System To estimate the system bandwidth, the amount of data transfer is concerned, while data content and processing are not cared. Architects view TLM can greatly speed up the modeling, and evaluate the system bandwidth effectively. 1. Data access to DDR2 takes 38 bus cycles for data transfer in the bus. The depth of task buffer in the DDR2 controller is 16. LCDC owns higher priority than VPU. Considering the two types of access, continuous access needs 2 bus cycles to complete burst-4 operation, while with-precharge access takes 4 bus cycles. 2. LCDC will request access to DDR2 when any of the three overlay FIFOs is not full. If data of continuous access belong to the same overlay, access is continuous access. Otherwise access is with-precharge access. 3. VPU will always request access to DDR2, which is continuous access. Because LCDC and VPU request access simultaneously, access may be interrupted by each other, continuous access may be interrupted to become with-precharge access often. Thus, the access to DDR2 is similar to a random process. Taking the advantage of the ESL design platform, the actual situation of access to DDR2 can be simulated and analyzed. Thus the system bandwidth can be evaluated effectively. The system model built in the ESL design platform is shown at Fig.4.
ESL Based SoC System Bandwidth Estimation Method1
19
Fig. 4. The system model built in the ESL design platform
4 Simulation and Analysis Result The simulation result of the target system model in the ESL design platform is shown in Table 1. Because of the existence of the with-precharge access, the utilization of the system bandwidth rises significantly. When LCDC is not optimized, the utilization of bandwidth is about 83%, the system bandwidth margin is about 17%. Table 1. The contrast of the Utilization of system bandwidth between ideal estimation and actual simulation (LCDC is not optimized)
The utilization of system bandwidth
Ideal estimation
Actual simulation
62%
83%
LCDC can be optimized to effectively improve the utilization of system bandwidth by arranging the continuous access and reducing the number of with-precharge accesses. On the other hand, the more continuous accesses, the requirement of FIFO depth is higher. The depth of FIFO will affect the implementation area, so there should be a trade-off between the bandwidth and the area. The simulation results of LCDC with different transmission modes are shown at Fig.5. Through optimizing the LCDC design, the utilization of system bandwidth drops from 83% to 64%, in the worst case the margin of system bandwidth is still about 36%, which can meet the requirement.
20
Z. Xie et al.
Fig. 5. The simulation results of LCDC with different transmission modes
Fig.5 also shows the utilization of system bandwidth and the requirement of FIFO depth drops and rises in an exponential manner. Comparing Case 4 with Case 3, the utilization of system bandwidth drops by only 2%, but the requirement of FIFO depth rises by 24. Thus it is reasonable to access data belong to the same overlay 3 times continuously. In the late stage of the development, RTL simulation results show that, when LCDC accesses data belong to the same overlay 3 times continuously, the requirement of FIFO depth is 27, and the utilization of system bandwidth is 76%. These are in agreement with ESL simulation results, indicating the efficiency of the ESL based SoC system bandwidth estimation method. The time spent on ESL simulation and RTL simulation of the target system is shown in Table 2. Compared with the traditional RTL simulation, the time is reduced from 33 hours to 9 minutes, more than two orders of magnitude shorter. Table 2. The time spent on ESL simulation and RTL simulation of the target system
The time spent on data output of 24 frames of 800x640 image
ESL simulation About 9 minutes
RTL simulation About 33 hours
5 Conclusions In this paper, a method of SoC system bandwidth estimation based on ESL is proposed, which estimates the system bandwidth by TLM and analysis depended on simulation. Compared with the traditional RTL simulation, the estimation is effective
ESL Based SoC System Bandwidth Estimation Method1
21
and the simulation speed is more than two orders of magnitude faster. Practice shows that the effective use of ESL design methodology can improve system performance and success rate of the implementation. With the improvement of complexity in SoC designs, the introduction of ESL design methodology has become an irresistible trend, and strongly promotes the development of SoC designs.
References 1.
2.
3.
4.
Ruei-Xi, C., Wei, Z., Qinyi, L., Fan, J.: Efficient H.264 Architecture Using Modular Bandwidth Estimation. In: International Conference on Embedded Software and Systems, pp. 277–282 (2008) Cho, Y.S., Choi, E.J., Cho, K.R.: Modeling and analysis of the system bus latency on the SoC platform. In: International Workshop on System-Level Interconnect Prediction, pp. 67–74. ACM, New York (2006) Zhe-Mao, H., Jen-Chieh, Y., Chuang, I.Y.: An accurate system architecture refinement methodology with mixed abstraction-level virtual platform. In: Design, Automation & Test in Europe Conference & Exhibition, pp. 568–573 (2010) Shin, C., Grun, P., Romdhane, N., Lennard, C., Madl, G., Pasricha, S., Dutt, N., Noll, M.: Enabling heterogeneous cycle-based and event-driven simulation in a design flow integrated using the SPIRIT consortium specifications. Des. Autom. Embed. Syst. 11, 119–140 (2007)
Cellular Automaton for Super-Paramagnetic Clustering of Data Zhang Botao, Zhang Shuqiang, and Yu Zhongqiu Institute of Science, Information Engineering University, 450001 Zhengzhou, China
[email protected],
[email protected] Abstract. Using the basic idea of Super-paramagnetic Clustering (SPC), we propose a cellular automaton approach for data clustering: A data set is regarded as a Potts magnetic system and a short-range interaction is introduced between the neighboring spins. Let the system evolve automatically under the function of the spin-spin interaction and the thermal motion. Finally, at a proper temperature the system will reach the super-paramagnetic phase, in which the spins (data points) will form into a number of ‘magnetic domains’ (data clusters). We apply this method to some data sets with different structures and get satisfactory results. Keywords: Data clustering, Potts magnetic system, Cellular Automaton.
1 Introduction Data clustering is a basic technique in data analyzing and in a variety of scientific and engineering fields. It is widely used in the fields such as pattern recognition, artificial intelligence, computer image processing, biology, astrophysics, etc. [1]. The general definition of data clustering is as follows: partition N points given in the feature space of the data into M different groups so that two points that belong to the same group are, in some sense, more similar than two points that belong to different ones [1]. The procedure of the traditional and typical clustering methods is as follows: firstly some knowledge about the clusters’ structure is assumed (e.g. each cluster is represented by a center and some point around it), and then the similarity between data points is calculated. Finally, partition data points into different clusters according to their similarity. It should be noted that, in the above procedure a successful clustering is based on a priori knowledge and a correct assumption about the structure of data set. That is, in the traditional procedure, some external factors is imposed to the set structure no matter it is right or not. In 1996, Eytan Domany presented a new approach for clustering: Super-paramagnetic Clustering (SPC)[2-4]. The basic idea of SPC is to build clusters naturally by utilizing the physical properties of an inhomogeneous ferromagnetic model. Here a Potts spin variable is assigned to each data point and a short range-interaction is introduced. Thus, the Potts spins will evolve automatically under the function of the spin-spin interaction and the thermal motion. At a proper temperature, the system will reach the super-paramagnetic M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 23–29. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
24
Z. Botao, Z. Shuqiang, and Y. Zhongqiu
phase. Spins with strong interaction to each other will form a magnetic domain in which all spins have a same state. This magnetic domain is thus corresponding to a data cluster. Domany explored a way to identify the magnetic domain by using Monte Carlo procedure [5]. Domany’s idea inspires us to consider a “Cellular Automaton (CA) for data clustering”. Contrast to the Monte Carlo procedure, the CA model is a dynamic model and is based on the spatial sequences [6], which is very similar to the procedure of magnetizing of a Potts system. Furthermore, CA model may be run on the special CA machine. Thus, we expect to build a CA model to realize SPC in a more useful way. In the following sections we first introduce the idea of the Super-paramagnetic Clustering (Sec2), and then describe the CA model for data clustering and corresponding algorithm (Sec3). In Sec4 we apply the method to two data sets with different typical structures. At last in Sec5 we give some conclusions and discussions.
2 The Idea of SPC A data point is a Potts spin when the data set is regarded as inhomogeneous Potts system. The spin-spin correlations (presented by the coupling function J) let the uniformity of the system but thermal motion destroys it. The magnetic domain will be formed and separated from others in the super-paramagnetic phase. This is the data clusters. Here, the coupling function J represents the similarity between data points. The closer two points are to each other, the more they ‘like’ to belong to the same class. Obviously, J is some positive decreasing function of the distance, which is chosen as follows [2]: 2
d 1 J i = exp(− i 2 ) K 2a
(1)
Where a stands for the average nearest-neighbor distance. K is the average number of neighbors per set. The thermal motion will destroy the “bonds” among spins. At the low temperature enough, all spins of the system are in the same state, Obviously, the temperature plays the role of clustering resolution here.
3 The Simulation of Cellular Automaton Clustering In this section we propose a Cellular Automaton [8] model for the clustering of data. 3.1 Model Firstly, we consider the effect of the interaction Ji and the temperature T by setting a link between the center site and its neighbor i [5]. The probability that a spin can’t link up with its neighborhood i is as follows:
Cellular Automaton for Super-Paramagnetic Clustering of Data
Pni = exp(−
25
Ji ) T
(2)
and the probability of linking-up is:
Pbi = 1 − Pni = 1 − exp(−
Ji ) T
(3)
The next state of a site is determined by the configuration of the links between the site and its neighbors. The model is described as follows: The possible states of each lattice site (spin, data point) is s (s=1, 2,... , q). In order to apply the method to different dimension data sets especially in high dimension data sets, we use the mutual K-nearest-Neighbor[9] which is defined as follows: two points, A and B, are defined as neighbors if they have a mutual neighborhood value K; that is ,if A is one of the K nearest neighbors of B and vice-versa. This definition ensures that the local interaction is symmetric; the number of bonds of any site is less than K. In general the value of K determines the range of interactions between a site and its neighbors. The rules are defined as: (i) If a site doesn’t link up with all of its neighbors, it updates its state randomly with probability Pn. otherwise it will be ‘magnetized’ by its linked neighbors with probability 1 –Pn (converts its state to the state of one of its linked neighbors), Here Pn is the probability of the configuration of a site with no links (see Fig 1 (a)): K
K
Pn = ∏ Pni = exp(−
∑J
i =1
ķ
ķ
Ĵ
Ĵ
Ĺ
Ĺ (a)
Ĺ
T
i
ķ Ĺ
(4)
)
ķ
Ĵ Ĺ
(b)
i =1
ķ
Ĵ Ĺ
(c)
()
Ĺ
Ĵ Ĺ
(d)
Fig. 1. An example of realized configurations. a a site with no link, of s=1, c d e a site with links of s=3
( )、( )、( )
Ĺ
Ĺ (e)
(b) a site with a link
(ii) If a site has links with its neighbors with state s, it will be ‘magnetized’ to the state s with probability Ps. Suppose it has m neighbors with state s, Ps is the combined probability of the configurations the site and its neighbors with state s:
26
Z. Botao, Z. Shuqiang, and Y. Zhongqiu
K
m
Ps = [1 − exp( −
∑ Ji i =1
T
)] ⋅ exp( −
∑J j =m
T
j
)
(5)
Here, the configurations that a site links up with different states are forbidden, because it is impossible that a site be ‘magnetized’ by different states simultaneously. Fig 1 shows all the realized configurations of a site with three neighbors. 3.2 Algorithms At a proper range of T, the steps of the procedures of data clustering are: (i) Find out the neighbors of each site (Sec 3.1.2), and calculate K , di and Ji; (ii) Generate an initial configuration randomly; (iii) Calculate the converted probability of each site to different possible states; (iv) Update the states of all sites according to the above probabilities; (v) Repeat steps (iii) and (iv), until the system goes to stable; (vi) Class the sites (data points) with the same state into a cluster.
4 Tests 4.1 Random Numbers of the Two-Dimensional Normal Distribution This is a group of artificial data. Two groups of random numbers, which are evenly distributed in angle and normally distributed in radial, are generated by random number generator. This is a dataset of two natural kinds and two class central coincide annular, including 100 data points of average radius of 0.5 and 200 data points of average radius of 2. Clustering parameters are chosen to be: q = 15, K = 8. Figure 2 to figure 5 give four clustering results from low to high temperature (resolution). When the temperature is very low, the system is in ferromagnetic phase, all the datum being grouped into one cluster (figure 2). When the temperature increases to about 0.007, the system will experience a phase change, and the data points will be split into two clusters (figure 3). Then, when the temperature increases to about 0.1, the outer data class is further divided into two mass, and all the data is divided into three clusters (figure 4) , because there is an area of low density both above and below the outer datum. When the temperature continues to increase, the data mass continuously split, until the system turns into the paramagnetic phase (figure 5). As far as the natural class of the datum is concerned, proper clustering happens at an interval of T = 0.005 to 0.009.
Cellular Automaton for Super-Paramagnetic Clustering of Data
Fig. 2. T=0.001 into one cluster
,the data points are grouped
Fig. 4. T=0.01, all the data is divided into three clusters
Fig. 3. T=0.007 two clusters
27
,the data points are split into
,
Fig. 5. T=0.5 paramagnetic phase
4.2 Iris Data Iris data is a complete data set, including 150 samples of each Iris species (50 samples each), got by Fisher in 1936 after he collected the four features (the length and width of calyx leaves and the length and width of petals)of three different Iris genus(Iris Setosa, Iris Versicolor & Iris Virginica) . In the feature space, Iris Setosa and the samples of the other two species separate linearly, and the samples of Iris Versicolor and Iris Virginica have a little overlap. Clustering parameters are chosen to be: q = 15, K = 35. After a clustering test on this data set, we found that a roughly correct classification can be obtained at the interval of T = 0.01 to 0.1, and the clustering effect is the most stable when T≈0.018.At this time, the data were divided into three clusters, the samples of Iris Setosa being correctly divided but the samples of Iris Versicolor and Iris Virginica is more or less wrongly divided. Repeated testing results show that the number of wrong division is 5 to 25. Figure 6 shows the projection of the four-dimensional Iris number upon the first dimension and the fourth dimension. This is the best example of clustering, only 5 data
28
Z. Botao, Z. Shuqiang, and Y. Zhongqiu
Fig. 6. The projection of the Iris number upon the first dimension and the fourth dimension. T =0.018, five wrongly divided points are marked by rectangular square full sample.
being wrongly divided (see the five rectangular squares in Figure 6). In most tests, the number of wrong division is 13 or so. It shows in the chart that there are probably less than 50 samples in each cluster, which is the result of projection overlap of the data points.
5 Conclusions and Remarks We’ve constructed the “Cellular Automaton for data clustering” with the help of the similarities between super-paramagnetic phase formation and data clustering in the inhomogeneous Potts magnetic system. This method is based on the density of data points, which is suitable for different structure data set, and there is no need to make assumption about the data structure. In the clustering process, the rules of Cellular Automaton has played a decisive role, it provides a dynamic mechanism, letting the data points joint with each other according to their similarity. And the introduction of thermal motion functions as a kind of adjustment, letting the data point make multiple and repeated choices. The system will eventually evolve into a relatively stable configurations, and it has little to do with the initial configuration. This work is just a preliminary realization of clustering, and some questions need further research or optimization, including clustering stability, similarity and the best functional relation of the distance between data points, the best Cellular Automaton rules and the objective method to determine a candidate neighbor number K, etc.
Cellular Automaton for Super-Paramagnetic Clustering of Data
29
References 1. 2. 3. 4. 5. 6. 7. 8. 9.
Theodoridis, S., Koutroumbas, K.: Pattern Recognition. Publishing House of Electronics Industry, BeiJing (2010) Blatt, M., Wiseman, S., Domany, E.: Super-paramagnetic Clustering of Data. Physical Review Letters 76, 3251–3254 (1996) Blatt, M., Wiseman, S., Domany, E.: Data Clustering Using a Model Granular Magnet. Neural Computation 9, 1805–1842 (1997) Domany, E.: Superparamagnetic Clustering of Data-The Definitive Solution of an Ill-Posed Problem. Physica A 263, 158–169 (1999) Wang, S., Swendsen, R.H.: Cluster Monte Carlo alg. Physica A 167, 565 (1990) Dan Mueller, D.C., Chen, K., et al.: Solving the advection-diffusion equations in biological contexts using the cellular Potts model. Physical Review E 72(4), 1–10 (2005) Wu, F.Y.: The Potts model. Reviews of Modern Physics 54(1), 235–265 (1982) Zhang, B.T., Liu, C.H.: Cluster-approximation mean field theory of a class of cellular automation models. Physical Review E 59(5), 4939–4944 (1999) Blatt, M., Wiseman, S., Domany, E.: Clustering data through an analogy to the Potts model. In: Touretzky, D.S., Mozer, M.C., Hasselmo, M.E. (eds.) Advances in Neural Information Processing Systems, vol. 8, MIT Press, Cambridge (1996)
Embedded Memory Wrapper Based on IEEE 1500 Standard Maryam Songhorzadeh and Rahebeh Niaraki Asli Guilan University Guilan, Rasht, Iran
[email protected] Abstract. IEEE Std 1500 defines a modular and scalable test interface for embedded cores of a system-on-chip (SoC) which simplify test challenges. In this paper, we present a specialized wrapper compatible with IEEE Std 1500 to implement at-speed testing for embedded memory cores. The proposed embedded memory wrapper (EMW) supports test diagnosis with reasonable area overhead which makes it suitable for memory BIST applications. All required test control signals of EMW is generated on-chip by a single centralized memory Built-In-Self Test (BIST) controller. The BIST controller can be used in a hierarchical test design and implement parallel test to handle multiple test wrappers concurrently. Simulation and synthesis results on a group of embedded memory cores confirm that the proposed wrapper has been effectively reduces the test time and area overhead. Keywords: Embedded memory testing; System-on-a-Chip (SoC); Built-in-SelfTest (BIST); IEEE Std 1500; at -speed testing.
1 Introduction Today, the fast innovation in VLSI technology enables the design of complex system on a single chip. Although system chips offer advantages such as higher performance, lower power consumption and smaller volume and weight, the recent methods of design like synchronous production of different cores and using reusable and heterogeneous IP cores, poses new test constraints to the test community [1]. On the other hand, basic features of testability like controllability and observability are hard to achieve because direct access to core ports is virtually impossible [1]-[3]. To address these problems, IEEE Std 1500 was proposed which defines a scalable structure for independent, modular test development and test application for embedded design blocks. This standard also enables testing the external logic surrounding cores [4]-[6]. One of the core types which occupy a significant amount of SOCs is memory cores. The high density of embedded memory cores make them more prone to manufacturing defects than other types of on-chip circuitries. Since direct access to memory ports is virtually impossible and at-speed testing is difficult to achieve, BIST is a more practical method to test and diagnose embedded memories [7], [8]. But a M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 31–39. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
32
M. Songhorzadeh and R.N. Asli
serial BIST like BIST based on IEEE std 1500, has a long and unacceptable test and diagnosis time, because this standard can’t implement a parallel test and manage multiple cores concurrently. Considering all these issues, several BIST schemes using IEEE Std 1500 serial links has been proposed [8]-[12]. In [8], a modular test wrapper for small wide memories has been introduced. The interfaces between the memories and BIST circuit are based on IEEE std 1500. This scheme allows for at-speed test at low area overhead. In [9]-[11], a serial test interface based on IEEE 1500 for small memories has been proposed. But the area overhead of the IEEE std 1500 wrappers is high. On the other hand, testing multiple memories using a serial BIST based on IEEE std 1500 requires a long and intolerable test time. In [12], another wrapper based on IEEE std 1500 is introduced which is in complete accordance with the standard but has no modification and optimization for memory testing, so it can’t support at-speed and parallel test. In this paper, the structure of a wrapper based on IEEE Std 1500 for Built-In SelfTest of embedded memory cores is proposed. The proposed structure is capable of implementing at-speed test and also supports diagnosis which is an important requirement in repairing procedure. By using a centralized BIST controller to generate control signals of the wrapper, the proposed scheme can be imported in a hierarchical test methodology and handle multiple test wrappers concurrently. On the other hand, the area overhead of the design is improved comparing the existing schemes. The organization of this paper is as follows: An overview of the general specification of the scheme is first given in section 2. In section 3 we propose different operational modes of the wrapper and Section 4 is allocated to the results of simulation and synthesis on a group of embedded memory cores. The paper is concluded in section 5.
2 Embedde Memory Wrapper (EMW) Figure 1 shows a memory core surrounded by the proposed IEEE-1500-compliant wrapper called EMW. The structure is composed of a Wrapper Instruction Register (WIR) circuitry, a Wrapper Boundary Register (WBR), a Wrapper Bypass register
Fig. 1. The overall structure of the IEEE-1500-compliant wrapper (EMW)
Embedded Memory Wrapper Based on IEEE 1500 Standard
33
(WBY), a Wrapper Data Register (WDR) and two multiplexers. All components of EMW are specialized for memory testing while adopt with IEEE Std 1500. In the reminder of this section, we introduce each part in detail. 2.1 WIR Circuit Figure 2 shows the structure of the proposed WIR circuit. It is composed of three main parts: WIR controller, WIR register and selector. According to IEEE Std 1500, WIR circuit must produce the signals required for controlling the other components of the wrapper and this can be done through initializing the wrapper serial port and WIR register.
Fig. 2. Structure of the proposed WIR circuit
WIR Controller. WIR controller generates control signals of the wrapper. From hierarchical point of view, this controller is the last level of test hierarchy. The BIST controller constitutes the upper level controller and generates all of the wrapper serial port signals and an additional signal for parallel shift. The BIST controller initializes input signals to the WIR controller and shifts the necessary instructions to WIR register through WSI signal. When there is no instruction to shift, WSI signal is used to transfer the related data of the current mode. Generating the input signals of WIR controller through the BIST controller can help producing a hierarchical test structure in which a memory BIST controller manages all WIR controllers of the wrappers. This relation can reduce test complexity while causes higher compatibility between the two controllers. On the other hand, the test methodology which also affects the wrapper will be more flexible because the BIST controller is completely definable by designer. WIR Register. In IEEE Std 1500, WIR register receives required instructions and test data. The 3 bit register proposed in figure 3 is in complete accordance with the standard and composes of two stages: shift and update. WIR_WSI is the serial input which shifts data and instructions to WIR register. After shifting operation, the contents of shift register are loaded into the update register. WIR_controller produces WIR_shift and WIR_update control signals.
34
M. Songhorzadeh and R.N. Asli
Fig. 3. Structure of WIR Register
WIR Selector. WIR selector is another part of WIR circuit. It is used to connect WSI signal to any of the wrapper registers. Based on the IEEE Std 1500, only one register can be set between the WSI and WSO signals. The selector circuit separates the respective register and connects WSI signal to it. 2.2 Wrapper Boundary Register (WBR) WBR is an important element of IEEE Std 1500 which has the responsibility of applying test and functional stimulus to the core and receiving the core responses. Figure 4 shows the proposed structure of WBR which is specialized for embedded
Fig. 4. The proposed structure for Wrapper Boundary Register (WBR)
Embedded Memory Wrapper Based on IEEE 1500 Standard
35
Fig. 5. The proposed structure for WBR input cell
Fig. 6. The proposed structure for WBR output cell
memory core BIST applications and composed of several input and output cells, multiplexers and some glue logics. The operation of input and output cells is controlled by control signals from memory BIST controller. Figures 5 and 6 show input and output cell architectures, respectively. Each cell has two data and test input and output ports and can be set in three different operational modes. CFI and CFO ports are used to receive and shift out the functional data, while the CTI and CTO ports are allocated to receive and shift out the test data. There are three control signals for controlling the operation of the cells: WBR_s_shift, WBR_p_shift and WBR_capture. WBR_s_shift and WBR_capture signals are derived from the shiftWR and captureWR signals of WSP respectively and WBR_p_shift is resulted from an optional signal called WPP_shift which is added to WSP port for enabling parallel shift of WBR cells. Table 1 shows the value of each control signal in different operational modes. Table 1. Control signals of WBR Mode Functional Internal Test External Test
WBR_S_shift
WBR_P_shift
WBR_capture
0 0 1
0 1 0
0 1 1
36
M. Songhorzadeh and R.N. Asli
One of the most important issues in memory testing is at-speed test to detect delay faults. BIST architectures based on IEEE Std 1500 need many clock cycles to shift in test commands, update operation to the memory and capture and shift out the memory responses. In the proposed architecture, due to parallel operation of WBR cells, the test data is sent to the memory in each clock cycle, and the memory outputs is compared in real time. So by such an approach, the new test data is applied to the memory in each clock cycle. This time scheduling not only reduces test time, but also applies valid data to memory in each cycle. To realize the approach, a design like figure 4 is used. As can be seen, a 2-to-1 multiplexer is set between two input cells. One input of this multiplexer is used only for parallel test and the other comes from CTO of the previous cell. WBR_s_shift and WBR_p_shift signals control multiplexers and set WBR in appropriate configuration. 2.3 Wrapper Bypass Register (WBY) WBY is a one bit register sets between WSI and WSO signals and bypasses the wrapper in functional mode. WBY block diagram with input and output signals is shown in figure 7. WIR controller produces the only control signal of WBY which is called WBY_shift.
Fig. 7. WBY structure
2.4 Wrapper Data Register (WDR)
WDR is an optional register of IEEE std 1500 which shifts test outputs. When the memory is under test, WDR is set between WSI and WSO signals and shifts the diagnosis information out. The volume and content of this data is dependent on the structure of the designed BIST. WIR controller generates WDR_shift control signal.
Fig. 8. WDR structure
3 Operational Modes of EMW Functional mode: In this mode, the wrapper is bypassed and the memory implements the functional operation. The instruction code of this mode is 111. According to figure 3, when the reset signal is enabled, this value is loaded in the update register of WIR and
Embedded Memory Wrapper Based on IEEE 1500 Standard
37
WBY is set between WSI and WSO signals. According to figure 4, the three control signals WBR_p_shift, WBR_s_shift and WBR_capture are all set to 0 and the functional data can be shifted through the paths WFI_n into input_n and output_n into WFO_n for input and output cells respectively. Internal test mode: This mode is assigned to testing the memory and has an instruction code of 001. When the wrapper sets in this mode, the test data is shifted to the input cells through WPI parallel port and the memory responses are also shifted out from the memory through WPO parallel port. WDR is set between WSI and WSO signals and shifts the test results out. The values of the control signals are shown in table 1. External test mode: In external test mode, the logic surrounded the memory core is tested by shifting 010 to WIR. The values of the control signals change as shown in table 1. By setting WBR_s_shift and WBR_capture to 1 and WBR_p_shift to 0, WBR configures as a scan chain and sets between WSI and WSO signals. So applying the test data through WFI or receiving the response through WFO from the logics surrounding the memory, can be done by setting the WBR_capture to 0.
4 Simulation and Synthesis Results To evaluate the design, EMW has been simulated and synthesized with Xilinx 11.1 for memories with different configurations. It is synthesized using Xilinx 11.1 with 0.18 CMOS standard cell library. At first, EMW is simulated for a 16K×16 SRAM to measure test time. Table 2 shows the results of the synthesis for some march algorithms with different length. The proposed structure guarantees an at-speed test with the 147.1MHz clock frequency and it has an access time of 6.8 ns which are results from the synthesis. To evaluate the area overhead of the proposed approach, EMW is synthesized with Xilinx 11.1 and the results are compared with [12]. The wrapper structure of [12] is completely designed based on IEEE std 1500 without any modification. In this structure, WSP of the wrapper comes from a TAP controller and not from any the memory BIST controller, so the test procedure has less flexibility compared to the proposed structure. Table 3 summarizes the results on area overhead for different blocks in EMW and [12]. These wrappers are both synthesized for a 16K×16 SRAM and using 0.18 CMOS standard cell library. The results show the area overhead of the EMW is 120.79% improved compared to [12]. Also to evaluate the area overhead of the proposed wrapper for different memories, it is simulated and synthesized using Xilinx 11.1 for memories with various size and configurations. The memories are high density embedded SRAMs and EMW structure are synthesized using 0.18um CMOS standard cell library. Table 4 summarizes the synthesis results of area overhead of the proposed wrapper. In this table a N×W memory configuration denotes a memory with N words and each word has W bits. As can be seen, by increasing the size of the memory, the area overhead is significantly decreasing. Because by changing the configuration of the memory, only the number of the input and output cells of the WBR needs to be changed and the rest of the structure is the same for memories with different configuration and due to the results of table 3, WBR has the most area overhead.
38
M. Songhorzadeh and R.N. Asli Table 2. Test time for three different algorithm using EMW on a 16×16K SRAM March Algorithm March RABL [13] March BDN [14] March LR [15]
Test Time 7.616 ms 4.787 ms 3.046 ms
Table 3. Wrapper area overhead compared to [12] Wrapper Components WBY WDR WBR WIR circuit Total(Wrapper 16K×16) optimization
# of gates in Proposed Wrapper 16 50 864 426 1356
# of gates in Previous Wrapper NR* NR NR NR 2994 120.79%
Table 4. Wrapper area overhead for different memory sizes Memory size
EMW Number of gates
4 K×16 16 K×16 4 K×32 16 K×32 4 K×64 16 K×64
1320 1356 1896 1932 3048 3084
EMW area (mm2) 0.0136 0.0146 0.0196 0.0200 0.0316 0.0319
Area Overhead 3.59% 0.96% 2.59% 0.66% 2.08% 0.52%
5 Conclusion The structure of an at-speed Embedded Memory Wrapper (EMW) has been proposed. The proposed EMW structure is based on IEEE Std 1500 and suitable for Built-In Self-Test applications. One of the main issues in memory testing is at-speed test to detect delay faults and in the proposed wrapper this feature is realized by using parallel ports for input and output cells of the introduced embedded memory BIST wrapper boundary (MBWBR). Also diagnosis can be supported by shifting the diagnostic information out through WDR. The results shows EMW structure also has reasonable area overhead. Also any number of wrappers in EMW structure can receives their control signals from one memory BIST controller so the presented structure has the capability of diagnosis and parallel testing of multiple wrappers concurrently. In the future, we hope to develop a MBWBR which can share between multiple identical memories to reduce the area overhead of EMW.
Embedded Memory Wrapper Based on IEEE 1500 Standard
39
References 1. Hong Fang, B.: Embedded Memory BIST for Systems-on-a-Chip. MS Thesis, McMaster University, Hamilton, Ontario,Canada (2003) 2. Larsson, A.: Test Optimization for Core-based System-on-Chip. PHD Thesis, Linköpings university (2008) 3. Naddeau-Dosite, B., Adham, M.I.S., Abbott, R.: Improved Core Isolation and Access for Hierarchical Embedded Test, vol. 26(1), pp. 18–25. IEEE Computer Society Press, Los Alamitos (2009) 4. Wang, L., et al.: Turbo1500: Toward Core-Based Design for Test and Diagnosis Using the IEEE 1500 Standard. In: International Test Conference 2008 (ITC 2008), pp. 1–9 (2008) 5. Li, J., et al.: A Hierarchical Test Methodology for Systems on Chip. IEEE Micro 22(5), 69–81 (2002) 6. IEEE Computer Society, IEEE Standard Testability Method for Embedded Core-based Integrated Circuits- IEEE Std 1500TM-2005. IEEE, New York (2005) 7. Dean Adams, R.: High Performance Memory Testing:Design Principles, Fault Modeling and Self –Test. Kluwer Academic Publishers, New York (2005) 8. Aitken, R.C.: A Modular Wrapper Enabling High Speed BIST and Repair for Small Wide Memories. In: International Test Conference (ITC 2004), pp. 997–1005 (2004) 9. Vadeau-Dostie, B., Silburt, A., Agarwal, V.K.: A serial interfacing technique for external and built-in self-testing of embedded memories. IEEE Design & Test of Computers 7(2), 5–64 (1990) 10. Jone, W.B., Huang, D.C., Wu, S.C., Lee, K.J.: An efficient bist method for small buffers. In: Proc. IEEE VLSI Test Symp (VTS), pp. 246–251 (1999) 11. Huang, D.C., Jone, W.B.: A Parallel Built-in Self-Diagnostic Method for Embedded Memory Arrays. IEEE Trans. Computer-Aided Designed of Integrated Circuits and Systems 21(4), 44–465 (2002) 12. Squillero, G., Rebaudengo, M.: Test Techniques for Systems-on-a-Chip, Politecnico di Torino (2005) 13. Benso, A., Bosio, A., Di Carlo, S., Di Natale, G., Prinetto, P.: Automatic March Tests Generations for Static Linked Faults in SRAMs. In: Proceedings Design, Automation and Test in Europe DATE 2006, pp. 1–6 (2006) 14. Bosio, A., Natale, G.D.: March Test BDN: A new March Test for Dynamic Faults. In: IEEE International Conference on Automation, Quality and Testing, Robotics, AQTR 2008, pp. 85–89 (2008) 15. Van de Goor, A.J.: March LR: A test for realistic linked faults. In: Proc. IEEE VLSI Test Symp., pp. 272–280 (1996)
A SVPWM Control Strategy for Neutral Point Potential Compensation in Three-Level Inverter Jinsong Kang and Yichuan Niu Tongji University, College of Electronic and Information Engineering, 200092 Shanghai, China
Abstract. This paper introduces the topology of three-level inverter and analyses its operational process. According to the analysis, traditional SVPWM control strategy will cause neutral point potential shifting, thereby an improved control strategy is designed to solve the potential shifting by resynthesizing vectors in particular sectors. This control strategy reduces the neutral current to a value near zero to compensate the neutral point potential shifting. Simulation and experiment have verify the effect of this SVPWM control strategy. Keywords: three-level neutral point potential SVPWM.
1 Introduction Three-level inverter is widely applied in frequency control for high power scale, utility systems and rail transit etc. In comparison with traditional two-level inverter, three-level inverter possesses the advantages such as lower output voltage harmonics and lower switching loss, besides the electrical stress of power electronic devices is also minimized. Two dominant kinds of topology for three-level inverter are diode clamped and capacitor clamped between which the diode clamped is the most common topology in pragmatic use. Control strategies for three-level inverter could also be divided into three parts: single-pulse control method, PWM control based on carrier wave and space vector pulse width modulation (SVPWM). Because of easier implement in program and a higher DC voltage utilization, SVPWM is applied more extensively. Besides, neutral point potential balancing in three-level inverter needs to be taken into account and realized by modulating operating period of voltage vectors, thereby the mid-point voltage of capacitor can be stabilized.[1] Based on the operational principle of diode clamped three-level inverter (hereinafter referred to as the three-level inverter), this paper focuses on SVPWM control strategy and neutral point potential compensation for three-level inverter. It, starting from the voltage space vector distribution, analyzes the influence for neutral point voltage induced by voltage vector, thus a SVPWM control strategy based on vector synthesis is promoted and its effectiveness is verified by simulations.
2 Analysis of Three-Level Inverter Topology A three-phase three-level inverter topology is shown in Fig. 1. Each phase has four main switching devices, four freewheeling diodes and two clamping diodes. DC rail M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 41–48. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
42
J. Kang and Y. Niu
voltage is divided into two levels by two series capacitors C1 and C2. When power electronic devices VTa1 and VTa2 are conducting, the output voltage of terminal A is UDC/2, while VTa2 and VTa3 are conducting, the output voltage is 0. The conduction of VTa3 and VTa4 means the output voltage has its magnitude of -UDC/2.[2]
Fig. 1. Main circuit of diode clamped three-level inverter
In Fig. 1, the capacitor mid-point O is considered to be the reference neutral point, while point O’ is the mid-point of three-phase load. The relations between phase voltage of three-phase load and output voltage of inverter are
:
⎧U a = U A − U O ' ⎪ ⎨U b = U B − U O ' ⎪U = U − U ⎩ c C O'
(1) D
On the basis of space vector theory, introduce complex operator g( g = e j120 )into these equations, and combine the three-phase voltage equations. Then the vector representation of load phase voltage could be derived as Uref : U ref =
2 (U a + gU b + g 2U c ) 3
2 = [(U A − U O '' ) + g (U B − U O '' ) + g 2 (U C − U O '' )] 3 2 = (U A + gU B + g 2U C ) = u sα + ju sβ 3
In α and β stationary coordinates system,
(2)
usα and usβ are expressed as:
2 1 1 (U A − U B − U C ) 3 2 2 1 = (U B − U C ) 3
u sα = u sβ
(3)
Therefore, the 27 switching states known as output voltage vectors in stationary coordinates for three-phase three-level inverter are shown in Fig. 2(a).
A SVPWM Control Strategy for Neutral Point Potential Compensation
43
β
β
α
α
a. Space vector diagram
b. Vector synthesis in sector I
Fig. 2. Voltage space vector of three-level inverter
There are 19 voltage space vectors valid in the total 27, the other 8 vectors are redundant vectors. According to their magnitude, all space vectors can be divided into four categories: large vectors, medium vectors, small vectors and zero vectors. The 6 large vectors with a magnitude of 2U DC / 3 are: PNN, PPN, NPN, NPP, NNP, PNP. The 6 medium vectors are: PON, PNO, ONP, NOP, NPO, OPN which possess a magnitude of 3U DC / 3 . There are 6 groups of small vectors, each group has 2 vectors with a magnitude of U DC / 3 : (PPO, OON), (POO, ONN), (POP, ONO), (OPP, NNO), (OPP, NOO), (OPO, NON). The three zero vectors are: OOO, PPP and NNN.[5]
3 Neutral Point Potential Balancing of Three-Level Inverter In the operational process of three-level inverter, the inaccuracy of output voltage is caused by the non-conformity in charging and discharging of capacitors, which due to the discrepancy between the current flows into and out of the capacitors. [3] The heterogeneous influences upon neutral point potential may vary with the different vectors. Large vectors connect the phase terminals either to the positive or the negative dc rail. They do not cause the potential shifting in neutral point. The potential in each phase of three-phase load is equal in the 3 switching states of zero vectors. Even when all the loads are connecting to neutral point simultaneously, it will not cause the potential ripple. But, in medium vectors and small vectors, there is at least one phase connected to neutral point while others connected to the dc rail. Thus, the current flows through neutral point causing the unbalancing of neutral point potential in a form of potential ripple that fluctuates around the zero potential over time in a cycle. Fig. 3. shows the analysis of medium vector PON and small vectors POO, ONN. Equivalent circuits are given in Fig. 3, in order to explicate the cause of neutral point potential. iO represents neutral point current, while it is assumed the direction that the current flows out of neutral point is positive. As it's shown in Fig. 3(a), when the medium vector PON is on, neutral point current equals to the phase B load current, that is iO=ib. When ib>0, capacitor C1 is charging, hence its voltage UC1 increases; while capacitor C2 is discharging, its voltage UC2 decreases.
44
J. Kang and Y. Niu
a. PON
b. POO
c. ONN
Fig. 3. Equivalent circuit of voltage vectors β NPN
NON NPO
Ⅲ
OPN
Ⅱ
ξ
OPO
OON
PPO
NNN
NPP
OPP
NOO NOP
Ⅳ NNO
NNP
OOP
PPN
PON
Ⅰ
POO
ONN
PPP POP
Ⅵ
α
PNO
ONO
ONP
PNN
Ⅴ
OOO
PNP
Fig. 4. Voltage space vectors with UDC/4 under neutral point potential shifting
Thereby, neutral point potential UO=(UC2-UC1)/2 will decrease. Similarly, when ib>W b task a followed by task b directly or indirectly – a >W b if and only if there is a trace σ = t1t2t3 . . . tn−1 and i ∈ {1, . . . , n−2} such that σ∈ W and ti = a and ti+1 = b – a →W b if and only if a >W b and b ≯W a – a W b if and only if a >W b and b >W a
‖
Process Mining: A Block-Structured Mining Approach
51
3 The Block-Structured Mining Method Because the flow of tasks is to be portrayed, control-flow mining techniques need to support the correct mining of the common blocks (e.g., control-flow constructs) that appear in process models. These blocks are: sequences, parallels, alternatives, loops, and non-free-choice, invisible tasks and duplicate tasks. In structured WF-nets, there are only four blocks: sequences, parallels, alternatives and loops. The block-structured mining method is focus on the four blocks above. Of cause, the approach is also helpful to other algorithms that face other blocks. 3.1 Route of the Mining Approach For four blocks: sequences, parallels, alternatives and loops, four mining algorithms are designed. The idea is: aim directly at logs, four mining algorithms are applied circulative in order and found blocks to make logs shrink through replace. The approach show in the below: Do while { do while{ algorithmsⅠ /* Sequence blocks mining } until can’t find new mining blocks do while{ algorithmsⅡ /* Alternative blocks mining } until can’t find new mining blocks do while{ algorithmsⅢ /* Loop blocks mining } until can’t find new mining blocks do while{ algorithmsⅣ /* Parallel blocks mining do mining sequence blocks in parallel blocks mining alternative blocks in parallel blocks } until can’t find new mining blocks } until can’t find new mining blocks Each out loop, one or a few blocks can be found. Through substitution, logs can get shrink and simplifying logs is as the basic data sets for the next loop. 3.2 Four Algorithms Algorithm I: Sequence blocks mining There are many sequence blocks in logs. Sequence blocks can be divided two kinds: explicit and implicit. Explicit sequence constructs can be shown in the logs conspicuously. Some sequence constructs is difficult to be found intuitionally due to the disturbing of other tasks (i.e., tasks in parallel). Next is the explicit sequence blocks mining algorithm. Implicit sequence blocks will be given in parallel blocks mining algorithm.
52
Y.-l. Qu and T.-s. Zhao
Input: logs W, m cases, W={σ1, σ2,…,σm,}, n tasks, T={ t1,t2,t3, . . . tn } Output: a set of new tasks T’, a set of new cases W’ For each pair ∈T×T and a!=b { If a→wb //exist sequence relation between a and b { Found sequence block B=ab Replace ab(series) with B in each case Modify task set T: add B, drop tasks a,b Record B=a+b in document table } } Algorithm II: Alternative blocks mining Alternative construct has two characteristics. First, each alternative block has one start task and one end task in process model. Second, there is only one task between start task and end task. Also, alternative blocks can be divided two kinds: explicit and implicit. Explicit alternatives constructs can be shown in the logs conspicuously. Some alternatives constructs is difficult to be found intuitionally due to the disturbing of other tasks (i.e., tasks in parallel). Next is the explicit alternatives blocks mining algorithm. Implicit sequence blocks will be given in parallel blocks mining algorithm. Input: logs W, m cases, W={σ1, σ2,…,σm}, n tasks, T={ t1,t2,t3, . . . tn } Output: a set of new tasks T’, a set of new cases W’ For each pair ∈T×T and a!=b { Sign=.T. A=Φ Forσ=σ1 to σm { If exist si=a in σ =s1s2s3… { If si+2=b A=A {si+1} Else {sign=.F. Break For } If exist si=b in σ =s1s2s3… { If si-2=a A=A∪{si-1} Else {sign=.F. Break For} }
∪
If sign and |A|>1 //a is start task and b is end task in alternative block. // the elements in set P are alternative items. { B=a+A+b Replace aAb(series) with B in each case Modify T, add B, drop tasks a,b
Process Mining: A Block-Structured Mining Approach
53
Record B=a+A+b in document table } } Algorithm III: Loop blocks mining After running of algorithmⅠand algorithmⅡ, loop constructs have only two kinds in Figure 2 at this time. The length of loop is 0 or 1. So, algorithmsⅢ is composed of part1 and part2.
Fig. 2. Two Kinds of Loop Constructs
Part1: mining loop construct C1 Input: logs W, m cases, W={σ1, σ2,…,σm}, n tasks, T={ t1,t2,t3, . . . tn } Output: a set of new tasks T’, a set of new cases W’ For each t∈T { Sign=.F. For σ = σ1 to σm If exist si=t in σ=s1s2s3… { If si+1=t { Sign=.T. Break For } } If sign{ B=ttt… Replace ttt… with B in each case Modify T, add B, drop task t Record B=t in document table } } Part2: mining loop construct C2 Input: logs W, m cases, W={σ1, σ2,…,σm}, n tasks, T={ t1,t2,t3, . . . tn } Output: a set of new tasks T’, a set of new cases W’ For each t∈T { Sign=.F. L=Φ Forσ=σ1 to σm If exist si=t in σ =s1s2s3… { If si+1t and si+2=t {
54
Y.-l. Qu and T.-s. Zhao
Sign=.T. L=L si+1 }
∪
} If sign{ B=t+L Replace t+L with B in each case Modify T, add B, drop tasks t,L Record B=t+L in document table } } Algorithm IV: Parallel blocks mining Firstly, it is necessary to part the parallel tasks. The key of partition is to create the equivalence classes of parallel tasks. It assumes that A={T1,T2,…} is a equivalence classes of parallel tasks. ∀ x1,x2∈Ti, x1 wx2. The algorithm of equivalence classes of parallel tasks:
‖
(1) Let T’ be the set of parallel tasks and R be the set of parallel tasks relations on logs W. T’=Φ; R=Φ.
‖
(2) ∀ a,b∈T, if a Wb then R=R∪(a,b),T’=T’∪{a}∪{b} (3) Extended relation R until make R be a equivalence relation. It is easy to make R satisfy reflexive, symmetric and transitive. (4) Partition equivalence classes T’/R.
∀ a∈T’, set Ti=[a]R={x|x∈T,a
‖
w
x}
So, A=T’/R={T1,T2,…}. Ti is a equivalence class that is the set of all elements that are related to a task a. In other words, the tasks in the set of parallel tasks may be not parallel tasks. Maybe, there exist sequence or alternative relations. For example, it is in Figure 3.
Fig. 3. An example of parallel constructs.
Logs are {BCFDG, BCDFG, BDCFG}. Tasks CFD are parallel tasks. But task C never appears in the behind of task F. So there has the sequence relation between C and F. Next is the rule to find out sequence relation among parallel tasks equivalence classes of A:
∀ Ti∈A, a, b∈Ti if a>>b then the relation of a and b is sequence. Also, there is the rule to find out alternative relation among parallel tasks equivalence classes of A: ∀ Ti∈A, a, b∈Ti if a>>b & b>>a then the relation of a and b is alternative.
Process Mining: A Block-Structured Mining Approach
55
Also, parallel construct have one characteristic. Each parallel block has one start task and one end task in process model. Input: logs W, m cases, W={σ1, σ2,…,σm}, n tasks, T={ t1,t2,t3, . . . tn }, A={T1,T2,…} is the equivalence classes of parallel tasks on W. Output: a set of new tasks T’, a set of new cases W’ For ∀ m, n∈T { Sign=.T. TS=Φ Forσ=σ1 to σm { If tasks(σ,m, n)Φ and ∃ Ti∈A,tasks(σ,m, n) ∈Ti { do case TS=Φ: TS=tasks(σ,m, n); TStasks(σ,m, n): Sign=.F.; break for; Endcase } Else { Sign=.F.; Break for;} } If sign { Found parallel block B=mSn Replace mSn with B in each case Modify task set T: add B, drop tasks m, S, n Record B=m+S+n in document table } } Function Tasks(σ,a, b): caseσ= t1t2t3…tn, σ∈W. Exists a=ti,b=tj,i>j then Tasks(σ,a, b)={ti+1,ti+2,…,tj-1}
4 Conclusion In this paper we have presented an approach on discovering process models from process logs. This approach has been validated using logs of transactional information systems. It shows that the approach has an obvious advantage in getting the reasonable, security and understandable model Process mining is not restricted to creating new formal process models. Any organization's process will evolve over time, and thus their process models will need to evolve as well. Methods for discovery process may give a process engineer clues as to when and in what direction the process model should evolve, based on data from the currently executing process. While we have focused here on the use of these methods for generating process models, we also believe that they are useful in visualizing the data collected on a process. An engineer may simply be interested in a way to better understand the current process, as captured by the event data. Discovering patterns of behavior may be of help.
56
Y.-l. Qu and T.-s. Zhao
References 1. van der Aalst, W.M.P., van Hee, K.M.: Workow Management: Models, Methods, and Systems. MIT press, Cambridge (2002) 2. Fischer, L. (ed.): Workow Handbook 2001, Workow Management Coalition. Future Strategies, Lighthouse Point, Florida (2001) 3. Herbst, J.: A machine learning approach to workflow management. In: Lopez de Mantaras, R., Plaza, E. (eds.) ECML 2000. LNCS (LNAI), vol. 1810, pp. 183–194. Springer, Heidelberg (2000) 4. Agrawal, R., Gunopulos, D., Leymann, F.: Mining Process Models from Workflow Logs. In: Sixth International Conference on Extending Database Technology, pp. 469–483 (1998) 5. van der Aalst, W.M.P., van Dongen, B.F., Herbst, J., Maruster, L., Schimm, G., Weijters, A.J.M.M.: Workflow Mining: A Survey of Issues and Approaches. Data and Knowledge Engineering 47(2), 237–267 (2003) 6. van der Aalst, W.M.P., Weijters, A.J.M.M., Maruster, L.: Workow Mining: Which Processes can be Rediscovered, WP 74. BETA Working Paper Series. Eindhoven University of Technology, Eindhoven (2002) 7. Hammori, M., Herbst, J., Kleiner, N.: Interactive workflow mining. In: Desel, J., Pernici, B., Weske, M. (eds.) BPM 2004. LNCS, vol. 3080, pp. 211–226. Springer, Heidelberg (2004) 8. Cook, J.E.: Process Discoverying and Validation through Event-Data Analysis. Technical Report CU-CS-817-96, University of Colorado, Boulder, Colorado, Novermber (1996) 9. Agrawal, R., Gunopulos, D., Leymann, F.: Mining process models from workflow logs. In: Sixth International Conference on Extending Database Technology, pp. 469–483 (1998) 10. Schimm, G.: Generic linear business process modeling. In: Mayr, H.C., Liddle, S.W., Thalheim, B. (eds.) ER Workshops 2000. LNCS, vol. 1921, pp. 31–39. Springer, Heidelberg (2000) 11. Schimm, G.: Process miner - A tool for mining process schemes from event-based data. In: Flesca, S., Greco, S., Leone, N., Ianni, G. (eds.) JELIA 2002. LNCS (LNAI), vol. 2424, p. 525. Springer, Heidelberg (2002) 12. Herbst, J., Karagiannis, D.: Integrating machine learning and workflow management to support acquisition and adaptation of workflow models. International Journal of Intelligent Systems in Accounting, Finance and Management 9, 67–92 (2000) 13. Herbst, J.: Dealing with concurrency in workflow induction. In: Baake, U., Zobel, R., AlAkaidi, M. (eds.) European Concurrent Engineering Conference. Society of Computer Simulation (SCS), Europe (2000) 14. van der Aalst, W.M.P., Weijters, A.J.M.M., Maruster, L.: Workflow Mining: Discovering Process Models from Event Logs. IEEE Transactions on Knowledge and Data Engineering (12), 369–378 (2004) 15. de Medeiros, A.K.A., van der Aalst, W.M.P., Weijters, A.J.M.M.T.: Workflow mining: Current status and future directions. In: Chung, S., Schmidt, D.C. (eds.) CoopIS 2003, DOA 2003, and ODBASE 2003. LNCS, vol. 2888, pp. 389–406. Springer, Heidelberg (2003)
China RoHS: How the Changing Regulatory Landscape Is Affecting Process Equipment Reliability Chris Muller1,* and Henry Yu2 2
1 Purafil, Inc., 2654 Weaver Way, Doraville, GA 30340 USA Purafil Asia, Room 602B, Tengda Plaza, Haidian District, Beijing 100044 China
[email protected] Abstract. In 2006, China promulgated a law entitled “Administration on the Control of Pollution Caused by Electronic Information Products.” The purpose of this law is similar to that of the European Union’s RoHS Directive (2002/95/EC, “restriction of the use of certain hazardous substances in electrical and electronic equipment”). These regulations require the elimination of lead in electronic products and manufacturers now have to comply with RoHS if they want to continue to do business in the EU and China. Corrosion-induced failures were frequent in industrial process control systems even before RoHS regulations with a typical failure mechanism being the reaction of atmospheric sulfur with exposed metals. Corrosion can occur quite rapidly in humid environments especially in the presence of small amounts of atmospheric sulfur and chlorides resulting in e.g., intermittent equipment malfunctions, unplanned shutdowns, or failure of critical systems. This paper will discuss issues related to RoHS compliance, and China RoHS in particular, and the resulting potential for corrosion-related problems. Air quality and failure analysis data will be presented from several sites in China illustrating the fact that in addition to industrial environments, corrosive environments exist in locations that would otherwise be considered benign if not for the changes in electronic equipment mandated by RoHS legislation. Keywords: China RoHS, copper corrosion, corrosion control, electronic equipment, ISA Standard 71.04-1985, process controls, reactivity monitoring, reliability, RoHS, silver corrosion.
1 Introduction In 1998, the European Union (EU) discovered that alarmingly large amounts of hazardous waste were being dumped into landfill sites. Trends also indicated that the volumes were likely to grow 3-5 times faster than average municipal waste. This highlighted a massive, and growing, source of environmental contamination. In order to address these issues, the member states of the EU decided to create the Waste Electrical and Electronics Equipment (WEEE, 2002/96/EC) directive, whose purpose was to: *
Corresponding author.
M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 57–69. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
58
1. 2. 3. 4.
C. Muller and H. Yu
Improve manufacturers’ designs to reduce the creation of waste, Make manufacturers responsible for certain phases of waste management, Separate collections of electronic waste (from other types of waste), and Create systems to improve treatment, refuse, and recycling of WEEE.
The WEEE directive laid the groundwork for additional legislation and a proposal called EEE (Environment of Electrical & Electronics Equipment) was also introduced along the same lines. However, now this policy is generally referred to as the RoHS Directive and is often referred to as “Lead-Free” legislation. This is not a very accurate nickname, because it extends to other pollutants as well. The European Union (EU) directive 2002/95/EC “on the Restriction of the use of certain Hazardous Substances in electrical and electronic equipment” or RoHS was implemented in July 2006. This directive applies to electrical and electronic equipment designed for use with a voltage rating not exceeding 1,000 volts for alternating current and 1,500 volts for direct current. The requirements of this directive are applicable to the member states of the European Union. The purpose of the directive is to restrict the use of hazardous substances in electrical and electronic equipment and to contribute to the protection of human health and the environmentally sound recovery and disposal of waste electrical and electronic equipment. The EU's RoHS Directive restricts the use of six substances in electrical and electronic equipment: mercury (Hg), lead (Pb), hexavalent chromium (Cr(VI)), cadmium (Cd), polybrominated biphenyls (PBB) and polybrominated diphenyl ethers (PBDE). In order to comply with the EU ROHS legislation, all of these substances must either be removed, or must be reduced to within maximum permitted concentrations, in any products containing electrical or electronic components that will be sold within the European Union. Manufacturers have made significant investments in new processes that will eliminate these substances – especially lead. All applicable products in the EU market must now pass RoHS compliance. In short, RoHS impacts the entire electronics industry and compliance violations are costly – product quarantine, transport, rework, scrap, lost sales and man-hours, legal action, etc. Non-compliance also reflects poorly on brand and image and undercuts ongoing environmental and “due diligence” activities.
2 RoHS – The EU and Beyond Companies selling a broad range of electrical goods in the EU must now conform to WEEE and those same companies must also conform to RoHS. WEEE and RoHS rules, while laid down at the European level, are put into law at the national level. When exporting to Europe, it is essential to comply with national law in each relevant country.1 The EU law simply serves as a template for national laws, which may differ considerably. At the end of February 2006, China promulgated a law entitled “Administration on the Control of Pollution Caused by Electronic Information Products.” The purpose of 1
Croatia, Norway, and Switzerland are not part of the EU. They may nevertheless have legislation implementing EU WEEE and RoHS rules, or similar legislation.
China RoHS
59
this law is similar to that of the European Union’s so-called RoHS Directive (2002/95/EC, “restriction of the use of certain hazardous substances in electrical and electronic equipment”). In fact, the Chinese law is simply called “China RoHS” in the industry. While there is some commonality between the RoHS requirements in the EU and those in China, there are also significant differences that must be recognized and dealt with. However, in both instances these RoHS regulations require the elimination of lead in electronic products and manufacturers have to comply with RoHS if they want to continue in to do business in the EU and China. Many consider China RoHS regulations to be considerably more restrictive than those passed in the EU. As described by a potentially-impacted customer: “Without exemptions, it is impossible to build a compliant board” [1]. Although, the regulations are different and are based on different processes, the aims are similar. Convergence of the regulations is not foreseen at present, as it would require high-level negotiations between the EU and China and changes of approach [2]. Some of the key differences between China RoHS and EU RoHS [3]:
The scope is different. The requirements are different. There are no exemptions ... yet. Labels, marks, and disclosure are required. The concept of "Put on the market" is different. The penalties are different. The responsibilities dictated by the law are different. Material testing down to the homogeneous materials in every single part you use to build your product may be required. The regulation has been in force since March 1, 2007. You will have to design labels and issue change orders in order to comply. The standards that you have to comply with just became available in finalized versions.
RoHS regulations are also either in effect or pending in many countries – including Argentina, Australia, Brazil, Japan, Korea, Taiwan, and the United States.
3 Unintended Consequences An aim shared by almost all RoHS legislation is the elimination of lead in electronic products. Thus the main issue for the electronics industry became the use of lead in the manufacture of components and circuit board assemblies. A printed circuit board, or PCB, is used to mechanically support and electrically connect electronic components using conductive pathways, or traces, laminated onto a non-conductive substrate. Alternative names are printed wiring board (PWB), and etched wiring board. A PCB populated with electronic components is a printed circuit assembly (PCA), also known as a printed circuit board assembly). All PCBs have conducting layers on their surface typically made of thin copper foil. If the copper is left unprotected, it will oxidize and deteriorate. Traditionally, any exposed copper was plated with lead (-based) solder by the hot air solder leveling (HASL) process.
60
C. Muller and H. Yu
HASL has been working well for many years, is the predominant surface finish used in the industry, and is also the cheapest PCB available. Now RoHS essentially makes PCBs using the HASL process obsolete. Failure modes on other common leadfree PCB finishes such as Organic Solder Preservative (OSP) and electroless-nickel immersion gold (ENIG) make these technologies undesirable. As a result, alternatives such as immersion silver (ImmAg) and organically coated copper (OCC) are currently used as board finishes. Due to inherent processing difficulties with OCC boards, ImmAg boards are becoming the standard PCB finish in the electronics industry [4]. Immersion silver would seem to have a bright future under RoHS [5]. It is easy to apply to the boards, relatively inexpensive, and usually performs well. While ENIG presently has a larger market share, over the past 12 months more immersion silver process lines have been installed in PCB facilities than any other finish. However, some manufacturers have complained about issues with corrosion. If severe enough, this could lead to shorts and ultimate failure of the board. The Internationsl Society for Automation (ISA) Standard 71.04 [6] classifies several levels of environmental severity for electrical and electronic systems: G1, G2, G3 and GX, providing a measure of the corrosion potential of an environment (Table 1). G1 is benign and GX is open-ended and the most severe. Table 1. Classification of Reactive Environments Class G1
Severity Copper Level Reactivity2 Mild
G2 Moderate
G3
Harsh
GX
Severe
Comments
An environment sufficiently well-controlled such that corrosion is not a factor in determining equipment reliability. An environment in which the effects of corrosion are measurable 0 k =1. Step 2. Solving the problem (12), if p k < ε , turn to step 3, else turn to step 4.
,
Step 3. Calculate λ* which satisfied formula (13), choose λ*i = min{λi i ∈ wκ }. If
λ*i ≥ 0 , ∀i ∈ wk ∩ U (u ) , then u k is approximate optimized solution. If λ*i < 0 , delete the relevant constraint from wk , turn to step 6. Step 4. Calculate β k , then u k +1 = β k p k + u k . Step 5. If α k < 1 , add the relevant constraint into wk , then turn to step 6. Step 6. k = k + 1 , turn to Step 2.
4 Experimental Results Ship welding workshop is equipped with SZ9-1600/10 transformer, welding equipment such as single-phase, three-phase CO2 welding machine, are connected to power line by the triangle method. Then the group control experiments on 120 sets of welding equipment were carried out. In order to achieve reactive power compensation, and avoid fluctuations of power factor, voltage and current, welding system on the two-phase power were compensated by the use of three-phase delta connection imbalance compensation system. After the operation application of each welding machine, the number of online welding machine is calculated by the master controller according to the optimization algorithm, power factor compensation is also realized based on the upload parameters of each welding machine. 440
1 0.95 3
400
1
PF
voltage(V )
420
0.9
380 0.85
360 2 340
1
5
9
13 17 time(h)
21 24
Fig. 2. Voltage curve of 24 hours power supply
0.8 1
4 5
9
13 17 time(h)
21 24
Fig. 3. Power factor curve of 24 hours
As shown in Figure 2, curve 1 and 2 is the voltage curve of different times in a one day, respectively, curve 1 represents the voltage curve using the optimized algorithm,
Group Control Strategy of Welding Machine Based on Improved Active Set Algorithm
225
curve 2 represents the voltage curve when optimization algorithm is not used. Figure 2 also shows that, the average supply voltage workshop was increased 1.1% and reached 394V after group optimization controlling.There are two troughs on the curve: 10-11 points in the morning and afternoon between 3-4 points, the maximum voltage drop of two stages appeared at the same time in these two stage. But the voltage return to the maximum in at about midnight, this indicating that welding work had stopped, we can call this stage idle time. The power factor curve of different times in a day is shown in Figure 3. Curves 3 and 4, respectively, reprensent using the optimized algorithm and optimization algorithm is not used.through controlling the number of on-line capacitance and welding machine, dynamic power factor compensation is achieved, the average power factor increased to 90%, 95% transient, about 4% higher than curve 4.
5 Conclusion In this paper, we have successfully controlled the number of online welding machine under the condition of constrained input and output of network group control system. The optimization model considered n and AWT, experiment show it is simple and effective. The results of the experiment also indicate that CAN-Fieldbus technology is fit for welding group controlling, the perfect performance of this group control system was realized by the active set algorithm, the entire control process followed the principle of continuous and smooth. The active set algorithm is fit for solving the inequality constraints problem in this text; it has a strong ability to search along with constraint border. During each iteration process, the original question is equivalent to the corresponding equation constraint problems. Because of the linear search process, those non-functional inequality constraints can ensure the feasibility of iterative sequence. Finally, the results of the experiment indicate that this method is simple and feasible. If this method is combined with the other control technology, it will reduce the dependence of electricity demand of the production, reduce energy consumption and increase energy efficiency.
References 1. 2.
3. 4.
Cui, y., Zhao, j.: Computer group control system for resistance welding. Welding Technology 35(5), 50–51 (2006) (in Chinese) Cheng, H., Yu, X., Guo, J., Yin, G.: Development of Networking System and GroupControl System for Resistance Welding Machines. Modern Electronics Technique 7, 148–150 (2008) (in Chinese) Wang, R.: The resistance welding machine control system based on microcontroller. Welding Technology 39(12), 41–43 (2010) (in Chinese) Lin, T., Chen, H.B., Li, W.H., et al.: Intelligent methodology for sensing, modeling, and control of weld penetration in robotic welding system. Industrial Robot 36(6), 585–593 (2009)
226 5.
6. 7.
C. Shujin and Z. Junjun Xin, L., Xu, Z., Zhao, M., et al.: Analysis and research on mismatch in tailored blank laser welding for multi-group. In: Proceedings of the 2008 IEEE International Conference on Information and Automation, ICIA 2008, pp. 1034–1039 (2008) (in Chinese) Sagues, P.: Adaptive control techniques advance automatic welding. Welding Journal 89(8), 26–28 (2010) Wylie, N., Wylie, S.R., Cullen, J.D., et al.: NDE system for the quality control of spot welding in the automotive industry. In: Proceedings IEEE Sensors Applications Symposium, SAS 2010, pp. 73–78 (2010)
Simulation and Experiment Research on the Effects of DC-Bias Current on the 500kV Power Transformer Feng-hua Wang1, Jun Zhang1, Cheng-yu Gu2, and Zhi-jian Jin1 1
Key Laboratory of Control of Power Transmission and Conversion, Ministry of Education, Department of Electrical Engineering, Shanghai Jiaotong University, Shanghai 200240, China {fhwang7723,junzhang,zjjin}@sjtu.edu.cn 2 East China Electric Power Test & Research Institue, Shanghai 200437, China
[email protected] Abstract. In the paper, the effects of DC-bias current on the 500kV power transformer with three-phase five-limb core structure are investigated by calculation and experiment. First, the coupling field-circuit model is applied to simulate the DC-bias characteristics of power transformer, where the traditional magnetic circuit models are replaced by Maxwell's equations in the magnetic circuit. Then the no-load DC-bias experiment is made to investigate the endure ability of 500kV power transformer. It is seen that the calculated results are agreed well with the experimental results. The no-load currents of the 500kV power transformer are distorted and the THD increases under DC-bias. The even harmonic is increased first and then decreased with the increasing of DC injected current. The obtained results are significant for the manufacturer of power transformer and operation staff of power system. Keywords: power transformer, DC bias, magnetizing current, harmonic.
1 Introduction When the monopolar operation with ground return is adopted or the bipolar operation with unbalanced currents is occurred in the High Voltage Direct Current (HVDC) systems or the phenomenon of magnetic storm is occurred, DC currents or the geomagnetic induced currents (GICs) which are essentially direct currents due to its extremely low frequency are found to enter and leave the directly earthed neutrals of high-voltage star-connected windings, causing a DC-bias in the magnetizing current of the transformer. Those DC currents cause the transformer core to saturate during the half cycle in which the bias current is in the same direction as the magnetizing current. Consequently, some undesirable effects such as increased noises, excessive core vibrations and overheating etc are brought out, which poses a potential threat for the integrity and longevity of the transformers [1]. In the past several years, many researchers have paid more attentions to analysis the performance of power transformer influenced by DC-bias through the simulation studies such as the coupled field-circuit method [2] where the electric circuit equations and magnetic circuit equations are solved together and the finite element M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 227–234. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
228
F.-h. Wang et al.
method [3] or through the experimental studies [4] and some important conclusions have been obtained. However, three-phase five-limb power transformer is always adopted in the high voltage and large capacity power transmission. Due to the large volume and high cost, it is difficult to make the DC-bias experiment in the 500kV power transformer. Moreover, the foundation of magnetic circuit model is a hard work because of the complex magnetic structure and few work has been done. In the paper, the coupled field-circuit method is applied to simulate the DC-bias characteristics of a 500kV power transformer, where the traditional magnetic circuit models are replaced by Maxwell's equations in the magnetic circuit. The no-load DCbias experiment is made to investigate the endure ability of 500kV power transformer. The results are significant for the manufacturer of power transformer and operation staff of power system.
2 Description of the Calculation Model of 500kV Power Transformer under DC-Bias The DC current is flowed into the AC system through the neural point of the two AC power transformer with star-grounded connection. Fig.1 is the schematic diagram of the coupling model of electric and magnetic circuits of a three-phase five-limb power transformer under dc current inrush. In the figure, us1, us2 and us3 are the three-phase power supply, B1, B2 and B3 are the magnetic induction intensity of three-phase core column, B4 and B5 are the magnetic induction intensity of upper return yoke, B6 and B7 are the magnetic induction intensity of low return yoke, i1, i2 and i3 are the primary current of three-phase winding, ZL1, ZL2 and ZL3 are the three-phase load reactance, e1(t), e2(t) and e3(t) are the electromotive force of three-phase winding to the neutral line, U0 is the DC voltage. 2.1 Electric Circuit Equations According to the electrical circuit model shown in Fig.1, the following equations are satisfied with the loop constituted by each of the three phases and the neutral line:
usj (t ) + U 0 − R j i j (t ) − L j
di j (t ) dt
= e j (t ) = − N
d φ j (t ) dt
j = 1, 2,3
(1)
where, φ j ( j = 1, 2,3) are the magnetic flux of three-phase core column separately, R j ( j = 1, 2,3) are the linear resistances of the transmission line and the three-phase windings and R1=R2=R3, L j ( j = 1, 2,3) are the three-phase winding inductance and L1=L2=L3, N is turn of the primary winding. Equation (1) could also be represented in matrix form with the rearrangement. [Q ] = R[ I ] + L[ I ][di / dt ] + N [ I ][dφ / di ]
(2)
where, [Q] is the summer of the AC power supply and DC voltage, [I] is the identity matrix.
Simulation and Experiment Research on the Effects of DC-Bias Current
i1
i3
i2 us 2
us1
229
us 3
e2 (t )
e1 (t )
e3 (t ) U0
B1
B3
B2
e1 (t ) B5
B4
B6
B7
ZL2
Z L1 Z L3
Fig. 1. Coupling model of the electric circuit and magnetic circuit for a three-phase five-limb power transformer under DC-bias
2.2 Magnetic Circuit Equations
Maxwell's integral equations, which are suitable for any situation in electromagnetic fields, are represented as follows: G G G G G ∂D G (3) v∫ l H ⋅ dl = ∫ J ⋅ dS + ∫ ∂t ⋅ dS G
G
v∫ B ⋅ dS = 0
(4)
The schematic diagram of magnetic circuit of three-phase five-limb power transformer is shown in Fig.2. Here, H1, H2 and H3 are the magnetic field intensity of three-phase core column, H4 and H5 are the magnetic field intensity of upper return yoke, H6 and H7 are the magnetic field intensity of low return yoke. For the low frequencies in the core, the displacement current is ignored, so the second term on the right side of Eq.(3) is ignored. By applied the Eq.(3) to the magnetic circuit loop , , and marked in Fig.2, the following equation can be obtained.
①
②③ ④
g j ( H1 , H 2 , H 3 , H 4 , H 5 , H 6 , H 7 , i1 , i2 , i3 ) = 0
j = 1, 2,3, 4
(5)
where, g is the function of magnetic field intensity, winding current and the length of magnetic circuit.
230
F.-h. Wang et al.
H6
H7
H4
H5 H1
H2
H3
Fig. 2. Magnetic circuit of three-phase five-limb power transformer
Similarly, by applied the Eq.(4) to the three independent magnetic circuit node in Fig.2, then we have h j ( B1 , B2 , B3 , B4 , B5 , B6 , B7 ) = 0
j = 1, 2,3
(6)
where, h is the function of magnetic induction intensity and section area of the core. The single-value curve used for expressing the nonlinearity of each power transformer core [5] is represented as follows. B = f (H )
(7)
Substituting Eq.(7) into variable B of Eq.(6), the seven nonlinear equations are formed with Eq.(5). Finally, the three nonlinear equation including three variable H1, H2 and H3 are obtained. g j ( H1 , H 2 , H 3 , i1 , i2 , i3 ) = 0
j = 1, 2,3
(8)
Through programming with the iterative algorithm, the waveforms of no-load current and winding current of three phases are obtained. Then, the magnetic field intensity and flux density under DC-bias are found using Eq.(5) and Eq.(6).
3 DC-Bias Experiment of 500kV Power Transformer The DC-bias experiment are made in two 500kV power transformer connected in parallel form and the connection diagram is figured in Fig.3. The type of test transformer is YSFP-31500/220. In the figure, Tr1 and Tr2 is the two test power transformer with same structure, G is the generator. The DC source is composed of DC source and M&C system, where the M&C system is the measurement and control system to measurement the magnetic current of power transformer through Hall current sensor and to control the output of direct current of DC source [4]. The direct current output from the M&C system is applied to simulate the DC component in the earth flowed into the operated transformer. By adjusted the values of direct current in the M&C system, the distortion degree of magnetizing current and the noise of test transformer can be observed and consequently to estimate the ability of test
Simulation and Experiment Research on the Effects of DC-Bias Current
231
transformer to endure the DC-bias. Meanwhile, the sound intensity of test transformer is measured through the sound level meter which is placed near the test transformer to investigate the noise variations of test transformer.
Fig. 3. Connection diagram of DC-bias experiment of 500kV power transformer.
4 Results and Discussions 4.1 Calculation Results
Fig.4 and Fig.5 is the calculated results of the no-load current and magnetic flux under different DC voltage for 500kV power transformer based on the coupling fieldcircuit method. It is seen that the three-phase no-load current are distorted and their peak values become larger with increasing DC voltage. The magnetic flux move forward with increasing voltage which is implied the DC component in the power transformer. Using the FFT, the harmonic distributions of three-phase exciting currents with different DC bias are shown in Tab.1, Tab.2 and Tab.3. It is seen that the even harmonic and odd harmonic and THD all increase with DC voltage.
(a) U0=0
(b) U0=100
Fig. 4. Waveforms of excitation current of 500kV transformer under different DC voltage.
232
F.-h. Wang et al.
(b) U0=100
(a) U0=0
Fig. 5. Waveforms of magnetic flux of 500kV transformer under different DC voltage. Table 1. Harmonic characteristic of excitation current of A phase under different DC voltage. U0(V) 0 50 100 200
First 0.4551 0.4655 0.4919 0.6023
Second 0.0048 0.0416 0.0838 0.2065
Third 0.0476 0.0522 0.0670 0.1494
Fourth 0.0028 0.0263 0.0532 0.1372
Fifth 0.0461 0.0509 0.0636 0.1233
Sixth 0.0005 0.0055 0.0162 0.0679
Seventh THD(%) 0.0122 14.8581 0.0145 19.2292 0.0214 28.2089 0.0575 54.7439
Table 2. Harmonic characteristic of excitation current of B phase under different DC voltage. U0(V) 0 50 100 200
First 0.4781 0.4908 0.5162 0.5690
Second 0.0098 0.0572 0.0986 0.1565
Third 0.0281 0.0293 0.0294 0.0172
Fourth 0.0048 0.0299 0.0574 0.1037
Fifth 0.0485 0.0522 0.0601 0.0749
Sixth 0.0013 0.0044 0.0025 0.0146
Seventh THD(%) 0.0129 12.2590 0.0165 18.2980 0.0219 26.0107 0.0324 36.2536
Table 3. Harmonic characteristic of excitation current of C phase under different DC voltage. U0(V) 0 50 100 200
First 0.4564 0.4675 0.4980 0.2027
Second 0.0098 0.0440 0.0901 0.6263
Third 0.0490 0.0545 0.0731 0.2292
Fourth 0.0063 0.0281 0.0580 0.1708
Fifth 0.0466 0.0518 0.0666 0.1551
Sixth 0.0020 0.0066 0.0193 0.1377
Seventh THD(%) 0.0125 15.2951 0.0151 19.9317 0.0235 30.0605 0.0800 59.3634
4.2 Experimental Results
Fig.6 is the experimental results of the no-load current under different DC current for 500kV power transformer. It is seen that the no-load current are distorted clearly and their peak values become larger with increasing DC current. The harmonic distributions of exciting currents are shown in Tab.4. It is shown that the THD increases with DC current. The odd harmonic increases with DC current, while the
Simulation and Experiment Research on the Effects of DC-Bias Current
233
even harmonic increases firstly and then decreases, which may imply the deep saturation of the core. At the same time, the sound intensity is 59dB measured by the sound level meter with no DC current. When the injected DC current is 4A, the sound intensity is 82dB.
(a) Idc=3A
(b) Idc=4A
Fig. 6. Waveforms of excitation current of 500kV transformer under different DC currents. Table 4. Harmonic characteristic of excitation current of A phase under different DC current. Idc(A)
First
Second
Third
Fourth
Fifth
1 2 3 4
1 1 1 1
0.0050 0.0128 0.0137 0.0011
0.046 0.059 0.093 0.761
0.001 0.013 0.026 0.003
0.0006 0.029 0.028 0.325
Sixth 0.0001 0.0044 0.016 0.003
Seventh THD(% ) 0.0007 4.58 0.0084 6.883 0.020 10.946 0.0242 83.947
5 Conclusion The coupling field-circuit model is applied to simulate the DC-bias characteristics of a 500kV power transformer, where the traditional magnetic circuit models are replaced by Maxwell's equations in the magnetic circuit. The DC-bias experiments are made in the 500kV power transformer with the developed DC controlled current source to simulate the DC component in the earth and to the power transformer experimentally. It is seen that the calculated results are agreed well with the experimental results. The no-load current of the three-phase five-limb transformer is distorted clearly and the THD increases under DC-bias. The variation of even harmonic is increasing firstly and then decreasing, which implies the variation of B-H curve. Therefore, it is necessary to model the B-H curve carefully including the linear segment and saturation segment to investigate the DC endure ability of 500kV power transformer by simulation. This is our next works.
234
F.-h. Wang et al.
References 1.
2.
3. 4. 5.
Cao, L.: Study on the Influence of DC biasing Current on Power Transformer Magnetization Characteristics. Ph. D. thesis, Department of Electrical Engineering. Tsinghua University, Beijing (2007) Cao, L., Zhao, J., He, J.L.: Research on the Withstand Performance of Three-phase Threelimb Power Transformer under DC Current Biasing. High Voltage Engineering 33, 71–75 (2007) Tong, L.: Research of DC Magnetic Bias on 500kV Auto-transformer. Master Thesis, Department of Electrical Engineering. Shanghai Jiaotong University, Shanghai (2007) Wang, F.H., Zhang, J., Jin, Z.J.: Experimental Study of the Effects of DC-bias Current on the Power Transformer. In: APPEEC 2010, pp. 1–4. IEEE Press, New York (2010) Pedra, J., Sainz, L., Corcoles, F.: PSPICE Computer Model of a Nonlinear Three-phase Three-legged Transformer. IEEE Trans. Power Del. 19, 200–207 (2004)
A HID Lamp Model in Simulink Based on the Principle of Electric Arc Xiaohan Guan* and Zhongpeng Li College of Information Engineering, North China University of Technology 5#, Jinyuanzhuang Road, Shijingshan District, Beijing, China, 100144
[email protected] Abstract. A dynamic conductance model of HID lamp, based on the classical principle of electric arc, will be established in this paper. Followed by the model and corresponding results, this type of HID lamp model will be simulated at low and high frequency in Simulink. The V-I characteristic, the relationship of current and voltage output, also the conductance of HID lamp model will be measured in this paper. It shows that the proposed model faithfully emulates external electrical properties of HID lamp at low and high frequency. Keywords: HID lamp; model; Simulink; electric arc
1 Introduction With the development of Computer-Aided Design, simulation tools are more significant in the power electronic systems currently. In the design process of electronic ballasts, the electronic ballast circuit would be seriously considered, and then simulated in order to test the feasibility and optimization of the circuit. Therefore, an equivalent model of HID lamp must be provided to facilitate the simulation. Matlab/Simulink is the high-performance numerical software of Mathworks Company in United States in the mid-80s in 20th century. After developing within three decades, Matlab has become a basic mathematical tool of mathematical statistics, automatic control theory, dynamic system simulation and many other courses. In fact, there are many articles on HID lamp model in Pspice. However, the model of HID lamp in Simulink is still less. Therefore, in the second section of this paper, a new dynamic conductance model of HID lamp based on the principle of electric arc will be established. The model combined the mathematical model proposed by Cassie and Mayr. This model will be simulated at low and high frequency in Simulink. Simulation results show that the model in this paper can simulate the external electrical characteristics of HID lamp and verify the *
This work is supported by Beijing Education Committee Technology Development Plan Project (KM200810009011) and Funding Project for Academic Human Resources Development in Institutions of Higher Learning under the Jurisdiction of Beijing Municipality PHR201008185.
M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 235–242. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
236
X. Guan and Z. Li
applicability of the model. The model of this paper can provide numerous guidance and reference value for electronic ballasts.
2 Establishment of the New Dynamic Conductance Model Before a model of HID lamp is established, it is extremely significant to understand some basic knowledge of the HID lamp’s electrical properties, such as different electronic properties [1] of HID lamp at low and high frequency. In the other word, when the low frequency sinusoidal current is treated as input, such as tens of Hertz, lamp voltage will cause the phenomenon of re-ignition. From another view, when the sinusoidal current is changing around zero, the voltage will increase suddenly. This phenomenon is similar as beginning ignition. And with the current increase, the voltage reaches maximum value dramatically and followed by decreasing slowly to the normal level. This phenomenon occurs periodically as the sinusoidal current. So its V-I characteristics show a classic hysteresis phenomenon. When the input current is a high-frequency sine wave, for instance, tens of kHz, current and voltage of lamp are sinusoidal and at the same phase. The V-I characteristics at this time is a linear relationship, however, the slope is not fixed. As it can be seen, HID lamp model should be designed to simulate these two characteristics simultaneously. Since discharge process of HID lamp is an electric arc, researchers had carried out much study of electric arc followed by lots of results. The study of HID lamp arc can be referenced from the research of switch arc mathematical model. The classical Cassie mathematical model established in 1939 and Mayr mathematical model proposed in 1943, based on simplifications of principal power-loss mechanisms and energy storage in the arc column, have been recognized for many years. In the Cassie model, Cassie assumed that the current density in an electric arc model is a constant, so the cross-section of the arc varied directly with the arc current. The resistivity and stored energy per unit volume are constants. Based on these basic assumptions, the famous Cassie arc model can be obtained, showed in equation (1). This model has a drawback that the modeled arc cannot be ceased. It describes the behavior of the arc when the current is strong; however it is not suitable for the description of arc characteristics when the arc-current is near to zero. θ
d ln G v 2 = 2 −1 dt EO
(1)
In Mayr’s model, Mayr assumed that the heat loss occurs from the outer arc only. Also the conductance of the arc changed with the energy stored in it. The Mayr mathematical equation is
θ
d ln G i2 = −1 dt PO G
(2)
This equation does allow the arc to cease. With the decrease of conductance G, i2/POG can be still more than unity. Hence, dlnG/dt is positive and the conductance
A HID Lamp Model in Simulink Based on the Principle of Electric Arc
237
continues to decrease until the arc is extinguished. The fitting of model results to measured data is achieved by means of a proper selection of arc parameters like the time constant and the current-dependent cooling power, which are normally taken as a function of arc current and voltage. The modified Mayr model is used in this article, shown as equation (3).
1 dG d ln G 1 vi = = ( − 1) θ P( PO + Ci i ) G dt dt
(3)
G is the instantaneous arc conductance; θ is the arc time constant; EO is the constant steady-state arc voltage; v is the arc voltage; i is the arc current; PO is the constant power loss of temporarily stable; P is the fill pressure of the circuit breaker; Ci is the current constant. Obviously, the Cassie model and modified Mayr model are not applicable in all cases. But two models are complementary with each other. As the modified Mayr model is more feasible for zero and low arc current region, Cassie model is more suitable for high arc current region. If these two models can be combined reasonably, a more applicable mathematical arc model can be obtained. Therefore, the hypothesis is presented as follow: 1. The time constant in these two models is the same. 2. A complete arc discharge process can be described by the combination of the Cassie model and modified Mayr model. But the conversion between two models lacks a transition point. It can be assumed that the transition current is I. If the arc current is greater than I, the arc discharge process is described by the Cassie model; otherwise, it is described by the modified Mayr model. 3. This paper supposes that the transition current is continuous. The transition is smooth. It is possible to define a transition factor f, which is an exponential function of the arc current. The transition factor f can be taken as: f = exp( −
i2 ) I2
(4)
A new mathematical arc model, equation (5), was got using equations (1), (3) and (4):
θ
1 dG v2 vi = ( 2 − 1)(1 − f ) + ( − 1) f G dt EO P ( PO + C i i )
(5)
In addition, the arc inherent conductance Gm between the electrodes should be considered in equation (5). Considered equations (4) and (5), the complete model is thus given by
θ
dG i2 i2 i2 i2 i2 G= G exp( − 2 ) − G 2 + Gm − exp( − 2 ) + dt EO EO I P ( PO + Ci i ) I
(6)
238
X. Guan and Z. Li
The equation (6) is the new dynamic conductance mathematics model of HID lamp. A 250W of HID lamp was tested at 50Hz in order to obtain the parameters. The parameters were taken as: θ =2*10-4S, EO=120V, I=0.45A, Gm =1.5*10-8S, PO = 250W. P and Ci were obtained by genetic algorithms and other mathematical calculations. P=4.99bar, Ci =499.8V/bar. According to equation (6), the dynamic conductance model of HID lamp in Simulink was established as Figure 1.
Fig. 1. Dynamic Conductance model of HID lamp in Simulink
3 Simulation of the New Dynamic Conductance Model The simulation circuit was shown as Figure. 2. According to the HID lamp current in practical work, the input current of simulation model is set to a sinusoidal current that the peak is 1.5A. In order to observe the voltage and current waveforms better, the input of simout1 in Figure. 2 was amplified to certain multiple.
Fig. 2. Simulation Test Circuit of model
A HID Lamp Model in Simulink Based on the Principle of Electric Arc
239
3.1 Simulation of the Model at Low Frequency
First, the conductance of the model at low frequency is observed, which was shown as Figure.3. It shows that the conductance changes periodically with the sinusoidal current. The peak is not fixed, which changes with the peak of the current. Figure 4 shows the voltage and current waveforms of the model at low frequency. It can be seen that the output voltage of the model is similar to a square wave. Current and voltage are in the same phase, but the voltage has a peak at each zero-crossing point. That is because the phenomenon of re-ignition of HID lamp appears at low frequency. However, as the frequency increases, the peaks gradually become much smaller and then disappear. At this time, the performance of voltage waveform will be improved.
Fig. 3. Conductance at 50Hz
Fig. 4. Voltage and Current at 50Hz
Figure 5 is the V-I characteristics of model at low frequency. It shows that the V-I characteristic performed as the classic hysteresis phenomenon, that is, the current reach the maximum, and then voltage reach the maximum as well. After that, both current and voltage are close to zero. Basically, the parameters of HID lamp are changing with this principle regularly at low frequency. All mentioned above are due to the gas thermal inertia in HID lamp. HID lamp at low frequency performed the nonlinear V-I characteristics. In order to analyze the frequency characteristics of the model deeply, the model in this paper was simulated at 5Hz. Simulation results of voltage and current were shown in Figure 6. It can be seen that the dynamic conductance model at 5Hz can reflect the negative incremental impedance characteristics of HID lamp more clearly. In other words, the lamp voltage decreases as the current increases.
240
X. Guan and Z. Li
Fig. 5. V-I characteristics at 50Hz
Fig. 6. Voltage and Current at 5Hz
3.2 Simulation of the Model at High Frequency (50 kHz)
The conductance of the model at high frequency was shown in Figure.7 and Figure.8. It can be seen in Figure 7 that conductance of the model gradually increased from zero to a constant value. When the model reached at high frequency in stable state, the equivalent conductance was shown in Figure.8. The conductance fluctuated in small range, and small-scale periodic fluctuations could be ignored, that is, the conductance was regarded as a constant. This fact is coincided with the actual situation of HID lamp. When HID lamp start at high frequency, the resistance is infinite. Otherwise, when it achieves stable state, the equivalent resistance is a constant value.
Fig. 7. Conductance at 50 kHz
Fig. 8. Conductance at the stable state
Figure 9 shows voltage and current of the model. It shows that the voltage becomes a sine wave and the peak disappears. The voltage and current are in same phase. This indicates that the model appears a pure resistance at high frequency. Figure 10 shows the V-I characteristics which was a linear relationship, while the slope was not fixed, but changed in certain range. It was verified that resistance with HID lamp could be considered as a variable resistor at high frequency.
A HID Lamp Model in Simulink Based on the Principle of Electric Arc
Fig. 9. Voltage and Current at 50 kHz
241
Fig. 10. V-I characteristics at 50 kHz
4 Conclusion The arc discharge process of HID lamp was described by the classical Cassie model and the modified Mayr model. The new mathematical model was established in this paper followed by the new dynamic conductance model based on the principle of an electric arc. Simulation results showed that the new dynamic conductance model could simulate the main characteristics of HID lamp. The model’s characteristics at low and high frequency are in great agreement with the external characteristics presented by HID lamp. Moreover, the model at a more low-frequency (5Hz) better reflects the negative incremental impedance of HID lamp. Some specific guidance and reference value in future will be provided in the process of electronic ballasts designed. The Cassie model and modified Mayr model describe the arc characteristics only from the physical concept. Therefore, a description of some complex features about HID lamp, such as the phenomenon of acoustic resonance and stroboscopic, has not yet contained in the new model. Such model establishment will become the following direction of HID lamp model.
References 1. 2.
3.
4.
Wang, Y., Wu, W., Wang, J., Yang, L.: Pspice Models of High-Intensity Discharge Lamps. Journal of Shanghai University (Natural Science), 11(6) (2005) Parizad, A., Baghaee, H.R., Tavakoli, A., Jamali, S.: Optimization off Arc Models Parameters Using Genetic Algorithm. In: International Conference on Electric Power and Energy Conversion Systems EPECS 2009, pp. 1–7 (2009) Shvartsas, M., Sam, B.-Y.: A SPICE compatible model of high intensity discharge lamps. In: 30th Annual IEEE Power Electronics Specialists Conference (PESC 1999), pp. 1037–1042 (1999) Wei, Y., Hui, S.Y.R.: A Universal Pspice Model for HID Lamps. In: 37th IAS Annual Meeting,Conference Record of the Industry Applications Conference 2002, pp. 1475–1482 (2002)
242 5. 6. 7.
8. 9.
X. Guan and Z. Li Herrick, P.R.: Mathematical models for high-intensity discharge lamps. IEEE Transactions on Industry Applications, 1A 16(5), 648–654 (1980) Wei, Y., Hui, S.Y.R., Chung, H., Cao, X.H.: Genetic algorithm optimized high-intensitydischarge lamp model. Electron. Lett. 38(3), 110–112 (2002) Zissis, G., Damelincourt, J.J., Bezanahary, T.: Modelling discharge lamps for electronic circuit designers: A review of the existing methods. In: Proceedings of the 2001 Industrial Application Society (2001) Anton, J.C.: An equivalent conductance model for high-intensity discharge lamps. In: IEEE Industry Applications Society-Annual Meeting, vol. 2, pp. 1494–1498 (2002) Tseng, K.J.: Dynamic Model of Fluorescent Lamp Implemented in PSpice. In: Proceedings of Power Conversion Conference PCC 1997, vol. 2, pp. 859–864 (1997)
Coordinated Control for Complex Dynamic Interconnected Systems Xin-yu Ouyang1,2 and Xue-bo Chen2 1
School of Control Science and Engineering, Dalian University of Technology, Dalian Liaoning 116024, China
[email protected] 2 School of Electronics and Information Engineering, Liaoning University of Science and Technology, Anshan Liaoning 114051, China
[email protected] Abstract. It provides a coordinated control method for complex systems with dynamic interconnections, that is, by using the dynamic inclusion principle and a desired reverse order permutation transformation, complex system with information constrains of dynamic topology structure is decomposed as a group of pair-wise decoupled subsystems in the expanded space. Then decentralized controllers and coordinators are designed, according to the conditions of dynamic inclusion principle, the overall coordinated compensator also are given. The obtain coordinated controller of the overall system can be contracted and implemented in the original space. Keywords: Complex Systems; Coordinated Control; System Decomposition.
1 Introduction The research on complex systems has been a focus in modern science currently[1-12], there are many literatures discussed this issue. However, there hasn’t uniform definition for complex systems. According to the motive of this paper, we call the system as complex system, which consists of multiple homogeneous subsystems and has many dynamic interconnections between subsystems, such as multi-agent system, electric power systems, multi-vehicle system, etc.. Since this kind of complex systems have the characteristics of high dimensions and variable topology structure constraints, coordinated control for them is a difficult problem. According to these facts, we will introduce dynamic inclusion principle and permuted transformations, and provide a coordinated control method for dynamic interconnected complex systems.
2 Descriptions of Complex Dynamic Interconnected Systems Consider an n -order complex dynamic interconnected system S = {Si } with N subsystems described by M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 243–249. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
244
X.-y. Ouyang and X.-b. Chen
N
Si : xi = Aii xi + ∑ eij (t , x) Aij x j + Bii ui ; yi = Cii xi , i = 1, 2," , N ( N ≥ 3 )
(1)
j =1
where xi ∈ R ni , ui ∈ R mi and yi ∈ R li are the state, input and output vectors of the i th subsystem respectively, Aii ∈ R ni ×ni , Aij ∈ R ni ×ni , Bii ∈ R ni ×mi and Cii ∈ R li ×ni are constant matrices; eij (t , x) is dynamic interconnected coefficient between subsystem i and j , which is a function with respect to time t and/or state x ; eij = 0 denotes it hasn’t self-connection in the i th subsystem at i = j . Assume all subsystems of the system are homogeneous, then the matrix form of the system (1) can be described by S : x = Ax + Ea x + Bu ; y = Cx
(2)
The variables satisfying N
N
N
i =1
i =1
i =1
n = ∑ ni , m = ∑ mi , l = ∑ li , x = [ x1T ," , x NT ]T , u = [u1T ," , u NT ]T , y = [ y1T ," , y NT ]T
(3)
Here x ∈ R n , u ∈ R m and y ∈ R l are the state, input and output vectors of the system respectively. Coefficient matrices are
A = blockdiag ( A11 , A22 ," , ANN ), B = blockdiag ( B11 , B22 ," , BNN ), C = blockdiag (C11 , C22 ," , CNN ), Ea = (eij Aij ),
(i, j = 1, 2," , N )
(4)
Let AE = A + Ea , the system (2) also can be rewritten as S : x = AE x + Bu , y = Cx
(5)
In fact, Ea represents dynamic topology structure of the system (2), i.e. dynamic interconnections between subsystems. In order to understand the characteristics of the complex system, let’s give the following notion of multi-overlapping[3,4,13]. Definition 1: The system S is said to possess N ( N − 1) / 2 multi-overlapping pair-wise dynamic interconnected subsystems
Sij : xi = Aii xi + eij Aij x j + Bii ui ;
yi = Cii xi ⎫ ⎪ x j = e ji A ji xi + A jj x j + B jj u j ; yi = C jj x j ⎬ ⎪ i = 1, 2,..., N − 1 j = i + 1,..., N ⎭
(6)
Coordinated Control for Complex Dynamic Interconnected Systems
245
if eij ≠ 0 and/or e ji ≠ 0 , where the subsystem Si in (1) is a multi-overlapping part of the pair-wise subsystem Sij in (6). If the subscripts of these pair-wise subsystems are arranged as Sij : S12 , S 23 , S13 , S34 , S 24 , S14 ," , Sij , " , S 2 N , S1N i = j − k , j = 2,3," , N , k = 1, 2," , j − 1.
. ()
(7)
we call the special sequence as recurrent reverse order subscripts. According to the definition 1, we know the dynamic interconnected system is a system with structure information overlapping, thus the system can be decomposed.
3 Decomposition of Systems With the system S in (5), we associate a corresponding expanded system described by
, y = Cx S = {Si } : x = A E x + Bu
(8)
where x ∈ R n , u ∈ R m and y ∈ R l are the state, input and output vectors of the expanded system S ; A E , B and C are matrices with appropriate dimensions. It is crucial to assume that n < n , m < m , l < l . In order to analyze the complex system, we give the following definition [4,7]. Definition 2: The system S dynamic includes the system S , or S ⊃ S , if there exists a group of full rank matrices {Vn × n ,U n× n , Rm × m , Qm× m , Tl×l , Sl ×l } satisfying UV = I n ,
QR = I m , ST = I l , such that for any x0 ∈ R n and any fixed u (t ) , the conditions x0 = Vx0 and u = Ru imply x(t ; t0 , x0 , u ) = Ux (t ; t0 , x0 , u ) and y[ x(t )] = Sy[ x (t )] for all t ≥ t0 . Here x(t ; t0 , x0 , u ) , x (t ; t0 , x0 , u ) are the unique solution of the first equation
in (5) and (8) for the initial time t0 respectively. The coefficient matrices of the system S can be obtained by A E = VAEU + M 1 , B = VBQ + M 2 , C = TCU + M 3 .
(9)
246
X.-y. Ouyang and X.-b. Chen
where T
N -1 N -1 ⎛
⎞ V = blockdiag ⎜ I n1 I n1 " I n1 , " , I nN I nN " I nN ⎟ ,
⎟ ⎜ N ⎝ ⎠ N N 1 1 ⎛
⎞ 1 U= blockdiag ⎜ I n1 I n1 " I n1 , " , I nN I nN " I nN ⎟ ,
⎟ ⎜ N −1 N ⎝ ⎠ T
N -1 N -1 ⎛
⎞ R = blockdiag ⎜ I m1 I m1 " I m1 , " , I mN I mN " I mN ⎟ , ⎜ ⎟ N ⎝ ⎠ 1 1 N N ⎛
⎞ 1 Q= blockdiag ⎜ I m1 I m1 " I m1 , " , I mN I mN " I mN ⎟ , ⎜ ⎟ N −1 N ⎝ ⎠
(10)
T
N -1 N -1 ⎛
⎞ T = blockdiag ⎜ I l1 I l1 " I l1 , " , I lN I lN " I lN ⎟ , ⎜ ⎟ N ⎝ ⎠ N -1 N -1 ⎛
⎞ 1 S= blockdiag ⎜ I l1 I l1 " I l1 , " , I lN I lN " I lN ⎟ ⎜ ⎟ N −1 N ⎝ ⎠
and M 1 , M 2 and M 3 are complementary matrices with appropriate dimensions, satisfying M 1V = 0 , M 2 R = 0 and M 3V = 0 . Let’s introduce permuted inclusion and permutation matrices[3] in the following. Definition 3: Consider
SP : x p = A Ep x p + B P u p , y p = C p x p
(11)
Assume there exists a group of full rank matrices {Vn × n , U n× n , Rm × m , Qm× m , Tl×l , Sl×l } such that S ⊃ S , then there exists a group of full rank matrices {U p , V p , R p , Tp } such that the system S p permuted includes S , or S p ⊃ S . Where x p ∈ R n , u p ∈ R m ,
y p ∈ R l ; x p = PA−1 x , u p = PB−1u , y p = PC−1 y ; U p = UPA , V p = PA−1V , R p = PB−1 R and Tp = PC−1T ; PA , PB and PC are non-singular permutation transformation matrices with appropriate dimensions. Definition 4: By partitioning an identity matrix I n× n into M sub-identity matrices,
I1 ," , I k ," , I M , with proper dimensions, we call
Coordinated Control for Complex Dynamic Interconnected Systems
⎡ 0 pk ( k +1) = blockdiag ( I1 ," , I k −1 , ⎢ ⎣ I k +1 ⎡0 p −1k ( k +1) = blockdiag ( I1 ," , I k −1 , ⎢ ⎣Ik
Ik ⎤ , I ," , I M ), 0 ⎥⎦ k + 2 I k +1 ⎤ , I ," , I M ) 0 ⎥⎦ k + 2
247
(12)
as basic column exchange matrix and basic row exchange matrix respectively, and H P = pi ( i +1) p(i +1)( i + 2) " p( j −1) j = Π kj −=1i pk ( k +1) , G (13) P −1 = p −1( j −1) j " p −1(i +1)( i + 2) p −1i (i +1) = Π kj −=1i p −1k ( k +1) , (i ≥ 1, j ≤ M )
are column group permutation matrix and row group permutation matrix respectively. In order to obtain the special sequence of Sij in Definition 1, the following transforms can be used: H H H P = Π iN=1− 2 Π Nj =−1i −1Π kN=(1N+−i ( ji −) −1)i ( j +1) pk ( k +1) G G G (14) P −1 = Π iN=1− 2 Π Nj =−1i −1Π kN=(1N+−i ( ji −) −1)i ( j +1) pkT( k +1) Here permutation matrices P and P −1 represent nonsigular column transformation and nonsigular row transformation respecitively. After the decomposition and permutation above, we can obtain pair-wise subsystems with recurrent reverse order subscripts as in (6) and (7).
4 Coordinated Control of Systems Consider the pair-wise subsystems in (6), we can design decentralized controllers and coordinators for them. The decentralized controllers to control each subsystem can be described by ⎡ kii K ij : uij = ⎢ ⎣0
0 ⎤ ⎡ xi ⎤ k jj ⎥⎦ ⎢⎣ x j ⎥⎦
(15)
and the coordinators can be described by ⎡0 Lij : uij = ⎢ ⎣l ji
lij ⎤ ⎡ xi ⎤ 0 ⎥⎦ ⎢⎣ x j ⎥⎦
(16)
which are used to coordinate the interconnected relations between subsystems i and j . Then the pair-wise closed loop subsystems can be described by ⎡ xi ⎤ ⎡ Aii + Bii kii Scij : ⎢ ⎥ = ⎢ ⎣ x j ⎦ ⎣e ji Aji + l ji B ji
eij Aij + lij Bij ⎤ ⎡ xi ⎤ Ajj + B jj k jj ⎥⎦ ⎢⎣ x j ⎥⎦
(17)
However, these pair-wise coordinators are not contractible to and implemented in the original space S with respect to dynamic inclusion principle. So we need to design a coordinated compensator ΔL for contraction in the expanded space.
248
X.-y. Ouyang and X.-b. Chen
In the expanded space, the controller is K = blockdiag ( K ii ) , i = j − k , j = 2,3," , N , k = 1, 2," , j − 1.
(18)
and the coordinator is L = blockdiag ( Lij ) , i = j − k , j = 2,3," , N , k = 1, 2," , j − 1.
(19)
where ⎡ kii K ij = ⎢ ⎣0
0⎤ ⎡0 , Lij = ⎢ ⎥ k jj ⎦ ⎣l ji
lij ⎤ 0 ⎥⎦
(20)
According to the dynamic inclusion principle and decomposition of the complex system stated above, it is obvious that the structure of ( K + L ) should be the same as the one of A Ep . Therefore, once the position of Aij is fixed in the expanded space, locations of lij will be determined. In this way, the coordinated compensator
ΔL can be established. Let LΔ = L + ΔL , it have ⎡ LΔ11 ⎢L LΔ = ( LΔ ij ) = ⎢ Δ 21 ⎢ # ⎢ ⎣ LΔ N 1
LΔ12 LΔ 22 # LΔ N 2
" LΔ1N ⎤ " LΔ 2 N ⎥⎥ , i, j=1,2,…,N. % # ⎥ ⎥ " LΔ NN ⎦
(21)
here LΔ ij are block matrices with a dimension of (N–1) I ni × (N–1) I n j and have
LΔij
⎧0, ⎪ ⎪ ⎪⎪ ⎡ 0 = ⎨⎢ ⎪⎢ 0 ⎪⎢ # ⎪⎢ ⎩⎪ ⎣⎢ 0
i = j, " lij " 0 ⎤ " lij " 0 ⎥⎥ , # #⎥ ⎥ " lij " 0 ⎦⎥
i < j , non-zero element is in column j − i,
(22)
i > j, non-zero element is in column N − ( j − i ).
Thus we can obtain the coordinator and coordinated compensator by L = P −1 LΔ P
(23)
Here P is recurrent reverse order permuted transformation matrix. At last, we obtain the coordinated controller ( K + L of overall complex system,
)
it can be contractible to and implemented in the original space S by Lk = Qp ( K + L )V p
(24)
Coordinated Control for Complex Dynamic Interconnected Systems
249
5 Conclusions The paper discussed coordinated control method of dynamic interconnected complex systems by using dynamic inclusion principle and permutation transform. The obtain coordinated controller of the overall system can be contracted and implemented in the original space.
Acknowledgment This research reported herein was supported by the NSF of China under grant No. 60874017.
References [1] Zhang, Z.D., Jia, L.M., Chai, Y.Y.: On General Control Methodology for Complex Systems. In: Proceedings of the 27th Chinese Control Conference, Kunming,Yunnan, China, pp. 504–508 (2008) [2] Ouyang, X.Y., Chen, X.B., Wang, W.: Modeling and decomposition of complex dynamic interconnected systems. The 13th IFAC Symposium on Information Control Problems in Manufacturing, Moscow, Russia, p.1006–1011 (2009) [3] Chen, X.B., Stankovic, S.S.: Decomposition and decentralized control of systems with multi-overlapping structure. Automatica 41, 1765–1772 (2005) [4] Chen, X.B., Stankovic, S.S.: Dual inclusion principle for overlapping interconnected systems. Int. J. Control 77(13), 1212–1222 (2004) [5] Ikeda, M., Šiljak, D.D., White, D.E.: Decentralized control with overlapping information sets. Journal of Optimization Theory and Applications 34(2), 279–310 (1981) [6] Chen, X.-B., Stankovic, S.S.: Overlapping decentralized approach to automation generation control of multi-area power systems. International Journal of Control 80(3), 386–402 (2007) [7] Chen, X., Stankovic, S.S.: Inclusion principle of stochastic discrete-time systems. Acta Automatica Sinica 23(1), 94–98 (1997) [8] Šiljak, D.D.: Large scale dynamic systems: stability and structure. North Holland, New York (1978) [9] Ikeda, M., Šiljak, D.D.: Lotka-Volterra Equations: Decomposition, Stability, and Structrue. Journal of Mathematical Biology 9(1), 65–83 (1980) [10] Tan, X.L., Ikeda, M.: Decentralized stabilization for expanding construction of large-scale systems. IEEE Transactions on Automatic Control 35(6), 644–650 (1990) [11] Šiljak, D.D.: Dynamic graphs. Nonlinear Analysis: Hybrid Systems 2, 544–567 (2008) [12] Wang, Q., Chen, X.B.: Connective Stability Analysis for a Class of Pseudo-Linear Interconnected Swarm Systems. In: Proceedings of the 8th World Congress on Intelligent Control and Automation, Jinan, China, pp. 1195–1199 (2010) [13] Chen, X.-B., Xu, W.-B., Huang, T.-Y., Ouyang, X.-Y., Stankovic, S.S.: Pair-wise decomposition for coordinated control of complex systems. Submitted to Information Sciences (2010)
Study on Calibration of Transfer Character of Ultrasonic Transducer Qiufeng Li1, Quanhong Zhang1, Min Zhao1, and Lihua Shi2 1
Key Laboratory of NDT of Ministry of Education, Nanchang Hangkong University Nanchang, China 2 Engineering Institute of Corps of Engineers, PLA Univ. of Science and Technology Nanjing, China
[email protected] Abstract. Transfer character of ultrasonic transducer often influences on the test signal partly, and then test errors arise. To the problem, a compact method is proposed to calibrate the transfer character in this paper. The experiment data was obtained in water- immerging test of the transducers, and a discrete transfer function is established based on system identification algorithms and then used for transducer calibration. The method is validated effective by experiment. Not only can the characteristic of transducers be indicated, but also a referenced method is presented for calibrating the transfer character of LTI system. Keywords: Ultrasonic transducer, transfer character, LTI system, system identification.
1 Introduction During the course of ultrasonic test, the transfer character of ultrasonic transducer is that test signals could be filtered by a band-pass filter, which would pass the resonant frequency and restrain the deviation frequency. And thus excitation signal selected is always close to the resonant frequency at which greater response signal and SNR would be obtained [1]. Then the transfer character of transducer can be reflected by frequency response of transducer concretely. Nevertheless, the higher resolution needs to be obtained with wideband excitation signal. So the frequency band of excitation signal from transducer is requested more wide, which can obtain more frequency signals [2,3]. It is important to measure the frequency response of transducer. But the measuring device is so expensive and complicated that domestic producers cannot afford to it and provide the graph. For this reason, a compact method of calibrating frequency response of transducer, which could be achieved with common instruments and material, is presented in the contribution. The theory of the method is introduced in next section. M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 251–258. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
252
Q. Li et al.
2 Theory of Calibrating Transducer To continuous linear time-invariant (LTI) system, the relationship between input x(t) and output y(t) can be described by a linear differential equation with constant coefficients as below: N
∑ a (i ) ⋅ i =0
d i y (t ) M d j x (t ) = b ( j ) ⋅ ∑ dt i dt j j =0
(1)
It can express dynamic characteristics of the system. Any linear time-invariant system can be described by a difference equation similarly [4,5]: N
y (n) + ∑ a (i ) ⋅ y (n − i ) i =1
(2)
M
= ∑ b( j ) ⋅ x ( n − j ) j =0
Here a(i) (i=1,2,3…N) and b(j) (j= 1,2,…M) are constant coefficients of the equation. Taking (2) with Z-transform can transform as below: N
M
i =1
j =1
Y ( z )[1 + ∑ a (i) z −i ] = X ( z )[b(0) + ∑ b( j ) z − j ]
(3)
,
From the trait of Z-transform, Y(z)=X(z)H(z) and the discrete transfer function (DTF) of system H(z) can be given by comparing with (3). M
H ( z) =
Y ( z) = X ( z)
∑ b( j ) z
−j
j =0
N
1 + ∑ a (i ) z
(4) −i
i =1
Actually, the calibration of transducer system is the process of calculating H(z), which can be given after evaluating constant coefficients a(i) and b(j) through the system identification algorithms [6-8]. For major actual system, the order is unknown, so important problem is to ensure the order. In theory, higher order can describe the system more accurate. But the oversize order would bring more complicated model, calculation and error in practice. Thus higher order is not advantaged at all time [9]. For lower frequency transducer system, lesser order of the DTF model is required, and the feasible value is M N 5 which has been verified through experiments.
==
Study on Calibration of Transfer Character of Ultrasonic Transducer
253
3 Verification of LTI System 3.1 LTI System Character Before calibration, a precondition, the detecting system is LTI system, must be achieved. The results are effective and significant in the condition of assuring the precondition. Therefore, the detecting system established must be assured as LTI system. The detecting devices usually have working range. And these devices are working as LTI system when its work under the range. That is the reason why the model is established with LTI system in this paper. To assuring the precondition, the experiment has been achieved to verifying the LTI system. And the theory of verification can be described as follow. On the assumption that the unit impulse response of detecting system is h(n) the response to input signal x1(n) is y1(n) and the response to x2(n) is y2(n).
,
,
y1 (n) = x1 (n) ∗ h(n)
(5)
y2 ( n) = x2 (n) ∗ h(n)
(6)
If the response to ax1(n)+bx2(n) is ay1(n)+by2(n), as follow equation:
y (n) = [ax1 ( n) + bx 2 ( n)] ∗ h(n) = ax1 (n) ∗ h(n) + bx2 (n) ∗ h( n)
(6)
= ay1 ( n) + by 2 (n) Then the system is linear system. If the input signal is delayed k sampling cycles, the output signal is delayed k sampling cycles correspondingly. And then the system is also a time-invariant system.
y1 ( n − k ) = x1 (n − k ) ∗ h(n)
(7)
The upper content could be described directly as follow. For the given input, the system output has nothing to do with moment of input signal. The system is called LTI system if it is linear and time-invariant [5]. 3.2 Verification Experiment A test system shown in Fig. 1 includes an arbitrary function generator, a pair of 50kHz ultrasonic transducers produced by Kcrt company and signal collection and display unit. signal collection and display unit collects signals by 9812 card and inputs the signals into computer, and then displays the image after the signals are processed with imaging algorithm.
254
Q. Li et al.
ֵSignalো
Signal ֵো䞛䲚
Excitation
collection and
থ⫳఼
Ϣᰒ⼎ display
water ∈
Fig. 1. Sketch of test system.
The output signal from transducer need be collected in the experiment. However, in the course of propagation from receiver and collector, the signal would be affected not only by transducer system and but also by the media, and then the established model could not reflect the frequency response of transducer system because it involves the disturbance of the media. Therefore, water-immersion method is introduced for water is a homogeneous and isotropic media. When ultrasonic wave propagates in water, it can be thought that the wave has just attenuation in amplitude, and the effect of attenuation could be removed through adjustment of signal amplitude [10-12]. To bring out the reflected wavepackets facilely, a water tank with 300mm depth is selected which avoids the overlap of reflected waves.
1 0.8
Scaledx(n) amplitude
0.6 0.4 0.2 0
-0.2 -0.4 -0.6 -0.8 -1 -1.5
-1
-0.5
0
0.5
1
1.5
2
2.5 -4
t t/ s
x 10
18
|X(jamplitude )| Scaled
16
14
12
10
8
6
4
2
0
2
4
6
8
10
12 4
f / Hz
x 10
Fig. 2. The waveform and spectrum of excitation signal.
Study on Calibration of Transfer Character of Ultrasonic Transducer
255
Two excitation signals, impulse signalδ(n) and 50kHz modulated signal x(n), are selected as input signal respectively. δ(n) is given by difference operation with step signal in fact. And x(n) is a cosine function modulated by Gaussian pulse, which equation is shown as below:
x(n) = cos(2πft ) ⋅ e ( − ((t −ti )⋅ω )
2
)
(8)
Here f is the main frequency, w is a pulse-width coefficient of Gaussian pulse and ti is positional parameter of wavepacket in the transmitting cycle. Actual waveform and its spectrum are shown in Fig. 2. According to the input signals, system responses should be as below respectively.
y1 (n) = δ (n) ∗ h(n)
(9)
y 2 ( n ) = x ( n) ∗ h( n)
(20)
In the light of LTI system trait, it should be y2(n)=x(n)y1(n). And then the system is verified as LTI system if measured output of x(n) is y3(n)=y2(n)=x(n)h(n). Comparison of measured data is shown in Fig. 3.
0.8
Scaled amplitude
0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 1000
2000
3000 4000
5000
Sampling point
Fig. 3. Comparison of measured output data.
In the figure, solidline represents the collected waveform y2(n) after generated by x(n), and dashline describes the convolution result with x(n) and the collected signal after generated byδ(n). In view of existing some errors in the experiment, wavepacket and amplitude of two waveforms could be thought as uniform and anastomotic approximately, which could be verified that the test system is a LTI system.
4 Experiment of Calibrating Transducer 4.1 Signal Collection In the experiment, x(n) is selected as excitation signal. A pair of input and output signals is collected as the data for system identification. From the collected output
256
Q. Li et al.
signal, the first wavepacket is the reflected wavepacket from bottom. Contrasting with input signal as Fig. 4, the reflected wavepacket has been changed obviously, and ringing has comparative length.
3
2
1.5
Output signal y(n)
1
H(j2πf) |H(j )|
Scaled amplitude
2.5 2.5
2 1.5 1
0.5
0.5
0
Input signal x(n)
-0.5
0 4
-1
150
200
250
300
Sampling point Fig. 4. Comparison of input and output signals.
4.5
5
f /Hz
350
5.5 4
x 10
Fig. 5. The frequency response of system.
4.2 Calibrating Rresult After amplitude of the output signal is adjusted according to the input signal, the model of the system is established in the light of system identification, and then DTF H(z)=B(z)/A(z) could be obtained. The graph of H(jω) is shown in Fig. 5. The model parameters of measured transducer system are listed in Table I. And then the calculated DTF is calibration of the system. Table 1. Parameter of DTF of 50kHz transducer sytem. a(1) a(2) a(3) a(4) a(5)
0.5739 -2.129 3.064 -2.023 0.5186
b(1) b(2) b(3) b(4) b(5)
-3.338 3.952 -1.520 -0.418 0.3364
4.3 Verification for Calibrating Result To verifying the upper calibrating results, another modulated signal, which waveform is shown in Fig. 6, is selected as input signal to validate the DTF. After the input signal passes through the test system, the output signal compares with the output waveform which is obtained after the input signal passes through the calibrated DTF. The contrasting graph is shown in Fig. 7. In the Fig. 7, solidline represents the output waveform after input signal passes through the test system, and dashline describes the output signal after input signal passes through the calibrated DTF. After comparison with two output signals, it could be discerned that those signals are anastomotic approximately, which shows that the calibrating result is effective.
Study on Calibration of Transfer Character of Ultrasonic Transducer
257
1
1 0.8
Scaled amplitude
Scaled amplitude
0.6 0.4 0.2 0 -0.2 -0.4 -0.6
0.5
0
-0.5
-0.8 -1 0
0.2
0.4
0.6
0.8
1 -3
x 10
t /s
-1 0
1000
2000
3000
Sampling point
Fig. 6. Waveform of verified excitation signal. Fig. 7. Comparison of output signal and filtered signal.
5 Conclusion A compact method to calibrate transfer character of transducer has been presented in this paper. At first, the system for detection is verified as a LTI system. And then DTF of the transducer system is established based on system identification algorithms. At last, the transfer function is verified with other test signal, and the result shows that the calibration effect is valid. This kind of method can be used for calibrating others LTI systems. Acknowledgment. This work is supported by Natural Science Foundation of China (10872217), by Open Foundation of Key Laboratory of NDT of Ministry of Education (ZD200929003), and by the Graduate Innovation Base of Jiangxi Province.
References 1. Yuan, Y.Q.: Ultrasonic transducer. Nanjing University Press, Nanjing (1992) 2. Chandrana, C., Kharin, N.A., Nair, A.: High resolution fundamental and harmonic imaging using a MEMS fabricated ultrasonic transducer. In: 2007 IEEE Ultrasonics Symposium, New York, October 28-31, pp. 1183–1187 (2007) 3. Benenson, Z.M., Elizaroy, A.B., Yakovleva, T.V., et al.: Approach to 3-D ultrasound high resolution imaging for mechanically moving large-aperture transducer based upon Fourier transform. IEEE Transactions on Ultrasonics, Ferroelectrics and Frequency Control 49(12), 1665–1685 (2002) 4. Parks, T., Burrus, C.S.: Digital filter design. John wiley and sons, Chichester (1987) 5. Hu, G.S.: Digital Signal Processing Theory, Algorithm and implementation. Tsinghua university press, Beijing (2003) 6. Ljung, L.: System Identification: Theory for the User, 2nd edn. Prentice-Hall, NJ (1999) 7. Woo, S.-H., Doo, H.-L.: ‘System identification of structural acoustic system using the scale correction. Mechanical Systems and Signal Processing 20(1), 389–402 (2006)
258
Q. Li et al.
8. Dong, X.-J., Meng, G., Peng, J.-C.: Vibration control of piezoelectric smart structures based on system identification technique: Numerical simulation and experimental study. Journal of Sound and Vibration 297(6), 680–693 (2006) 9. Zhang, C.X., Ren, J.S.: The Order Discernment of Transfer Function for Linear System. Journal of Nanjing University of Science and Technology 21(2), 106–109 (1997) 10. Li, Q.F., Shi, L.H., Liang, D.K.: Method of Compensating Transducers Based on Digital Filtering in Concrete Test. Journal of Nanjing University of Aeronautics and Astronautics 40(1), 55–59 (2008) 11. Ohara, Y., Kawashima, K.: Detection of Internal Micro Defects by Nonlinear Resonant Ultrasonic Method Using Water Immersion. Japanese Journal of Applied Physics 43(5), 3119–3120 (2004) 12. Hak-joon, K.I.M., Sung-Jin, S.O.N.G., Lester, W.S.: Modeling Ultrasonic Pulse-Echo Signals from a Flat-Bottom Hole in Immersion Testing Using a Multi-Gaussian Beam. Journal of Nondestructive Evaluation 23(1), 11–19 (2004)
A Novel Non-contact Pulse Information Detection Method Based on the Infrared Sequence Images Weibin Zhou1,2, Bin Jing1,*, Dian Qu1, Guihong Yuan1,3 Chunyan Wang1, and Haiyun Li1 1
School of Biomedical Engineering, Capital Medical University, No.10, Xitoutiao,Youanmen, Fengtai District, Beijing, China 2 School of Pharmacy & Bioengineering, Chong Qing University of Technology, No.69, Hongguang Rd, Banan District, Chongqing, China 3 School of Foundational Education, Peking University Health Science Center, Xueyuan Rd.Haidian District, Beijing, China
[email protected],
[email protected],
[email protected],
[email protected],
[email protected],
[email protected] Abstract. In this paper, we propose a novel non-contact method to detect the pulse information from the radial artery using the infrared sequence images. First, a ROI (region of interest) can be located on the radial artery by the infrared imaging system because of its higher temperature than surrounding tissue. Then, a short time of AVI video of the radial artery is recorded and the area of the ROI is calculated in every frame, and a time-lapse signal is constructed using the calculated results, which can reflect the pulse information. Compared to the pulse wave from the pressure sensor, our result is acceptable. Our method reveals a novel non-contact way to obtain the pulse information and shows a significant value in the Chinese medicine. Keywords: radial artery, pulse information, infrared sequence images.
1 Introduction Pulse condition is very important in Chinese medicine, and it could give a lot of diagnostic messages. Traditionally, it depends on doctors’ experiences to feel the pulse exactly, but the experiences are relative and not objective. Many kinds of contact sensors have also been used to measure the pulse, but they may bring inconvenience to the subjects. Recently, some new pulse wave measurement methods have been proposed in [1][2][3].Infrared imaging system already has many applications in biomedical measurement. We have used it to detect physiological signals on Lumbar Vertebra and temporal artery in [4][5]. In this paper, we propose a novel objective pulse information detection method from the infrared sequence images. The paper is organized as follows: Section2 introduces the experiment setup. Section 3 introduces the processing method. Section 4 shows the discussion and conclusion. *
Co-first author.
M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 259–265. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
260
W. Zhou et al.
2 Experiment Setup First, an infrared dynamic image acquisition system is constructed. It has an infrared sensor (Fluke, NETD 0 , β > 0 (9) is well- posed. So the problem turn to select α and β which makes the solution of (8) is also the approximate solution of (3).
4 Fixed-Point Iteration The nonlinearity of (8) poses number of computation challenge, so the fixed-point iterative algorithm is introduced to overcome the difficult. Its basic idea is to successive linearization of nonlinear equation. Equation (8) can be rewritten into the following first-order nonlinear system
G
Where v =
G S T Sg − α∇ ⋅ v = S T λ
(10)
G G 2 − ∇g + ∇ g + β 2 v = 0
(11)
∇g ∇g + β 2
2
.When g of square root is fixed, i.e. set g=g(m) (m is the
G
number of iterations), (8) is linearized. When eliminate v in (10), the format of the fixed point iteration expressed as
( S T S + αL( g ( m ) )) g (m +1) = S T λ
(12)
Image Reconstruction Algorithm Based on Fixed-Point Iteration
329
Note that at each iteration, one must solve a linear diffusion equation, whose diffusivity depends on the previous iterate g(m). Thus
g (m +1) = ( S T S + αL( g ( m ) )) −1 S T λ
(13)
Equation (13) can be rewritten as
~ g ( m +1) = g ( m ) − ( H ( g ( m ) )) −1 u ( g ( m ) )
(14)
u ( g ) = ( S T S + αL( g )) g − S T λ
(15)
Where
~ H ( g ) is the approximation of Hessian matrix. ~ H ( g ) = S T S + αL ( g )
(16)
Fixed-point iterative algorithm is convergent and with increase in iteration the method converges linearly. But the numerical results show that the convergence is fast, only need one or two steps to get a good approximation of exact solutions. And fixed-point iteration is globally convergent, so it has nothing to do with the choice of initial value.
5 Simulation and Analysis of Results To verify the efficiency of the new algorithm, numerical experiments have been carried out by simulating the typical gas/oil two-phase patterns. In the patterns, dielectric constant of oil is ε oil = 3 , dielectric constant of gas is ε gas = 1 .The number of measurement electrodes is 12, the inner and outer radius of the pipe is 62mm and 75mm respectively, the radius of the grounded shielding is 80mm, and the measurement angle is 26°. The image area (the cross-section area of the pipeline of two-phase flow) is divided into 804 elements. While in experiment iteration error usually satisfied
Sg − λ ≤ δ
(17)
Where δ is observation error of original data and δ = 0.001 , β = 0.1 in this paper. While analysis reconstructing image quality, selecting spatial image error as evaluation index of image quality, its definition as follows:
e= Where
gδ − g g
(18)
g is the grey value of the simulation model, and gδ is the grey value of the
reconstruction image. The experiment results are shown as Table.1.
330
C. Li, X. Yang, and Y. Wang
Table 1. Comparison of reconstructed images.
Phantom(a)
Phantom (b)
Phantom(c)
LBP Algorithm
Landweber Iteration Algorithm
Tikhonov Regularization Algorithm
TV Regularization Algorithm
Phantom(d)
Image Reconstruction Algorithm Based on Fixed-Point Iteration
331
Table 1 shows the comparison of the reconstructed image. From Table 1, it can be seen that LBP has poor precision, and can not clearly distinguish the target area for the complex flow pattern; its position has obvious deviation. It can only be used for qualitative analysis; Landweber iteration method generally closes to the original flow for simple flow pattern, but can not distinguish multiple target areas clearly. Tikhonov regularization can distinguish multiple targets basically, but artifacts are more and precision is low. TV regularization can clearly distinguish between single or multiple target area, and the contrast and sharpness of reconstructed images increased significantly. Table 2 is the comparison of the reconstructed image of the error. It can be seen from the table, LBP least accurate, TV regularization algorithm has the highest accuracy. Table 2. Image error (%). Phantom
(a)
LBP
87.63 46.37
33.72
61.46
50.70
Landweber
31.49
56.38
13.77
Tikhonov
44.16
23.54
26.64
13.11
TV
41.05
17.10
22.49
12.19
(b)
(c)
(d)
6 Conclusion In this paper TV regularization based on bounded variation function is introduced to overcome the ill-posed of ECT, and the range of approximate solution can be extended to bounded variation function space. In order to overcome the difficult of solving the nonlinear Euler equations, a fixed-point iteration algorithm is proposed to get TV regularization solution. Numerical experiments show that: the contrast and sharpness of reconstructed images increased significantly, and the proposed algorithm has the advantages both in imaging speed and quality. Selection of regular parameters of regularization method directly affect the accuracy and speed of the reconstructed image, at present parameters selected mainly from experience, which has a certain randomness and uncertainty, so the selection of regular parameter, in particular the more reasonable selection principle and more efficient numerical implementation method is still to be further important issue to be addressed.
Acknowledgment This work was supported by the Hebei Province Natural Science Foundation of China under Grant No. E2007000048.
332
C. Li, X. Yang, and Y. Wang
References 1. Loser, T., Wajman, R., Mewes, D.: Electrical capacitance tomography: image reconstruction along electrical field lines. J. Measurement Science and Technology 12(8), 1083–1091 (2001) 2. Jaworskia, J., Dyakow Ski, T.: Application of electrical capacitance tomography for measurement of gas-solid flow characteristics in a pneumatic conveying system. J. Measurement Science and Technology 12(8), 89–98 (2001) 3. Zhu, K., Rao, M., Wang, C. H.: Electrical capacitance tomography measurements on vertical and inclined pneumatic conveying of granular solids. J. Chemical Engineering Science 58(18), 4225–4245 (2003) 4. Liu, S., Li, T.J., Chen, Q.: Visualization of flow pattern in the rmosyphon by ECT. J. Flow Measurement and Instrumentation 18, 216–222 (2007) 5. Wang, Z.Y., Jin, N.D., Wang, C., et al.: Temporal and spatial evolution characteristics of two-phase flow pattern based on image texture analysis. J. Journal of Chemical Industry and Engineering (China) 59(5), 1122–1130 (2008) 6. Yang, W.Q., Peng, L.H.: Image reconstruction algorithms for electrical capacitance tomography. J. Meas. Sci. Technol. 14, R1–R13 (2003) 7. Wang, H., Zhu, X., Zhang, L.: Conjugate Gradient Algorithm for Electrical Capacitance Tomography. Journal of Tianjin University 38(1), 1–4 (2005) 8. Chen, D.Y., Chen, Y., Wang, L.: A Novel Gauss-Newton Image Reconstruction Algorithm for Electrical Capacitance Tomography System. J. Acta Electronica Sinica. 37(4), 739–743 (2009) 9. Chen, Y., Chen, D.Y., Wang, L., et al.: Image reconstruction algorithm accelerated by polynomial for electrical capacitance tomography system. J. Chinese Journal of Scientific Instrument. 29(12), 2538–2542 (2008) 10. Zhao, J.C., Fu, W.L., Li, S.T., et al.: Image reconstruction new algorithm for electrical capacitance tomography. J. Computer Engineering 30(8), 54–82 (2004) 11. Wang, L., Chen, Y., Chen, D.Y., et al.: Improved trust region based image reconstruction algorithm for electrical capacitance tomography system. J. Chinese Journal of Scientific Instrument 31(5), 1077–1081 (2010) 12. Sun, N., Peng, L.H., Zhang, B.F.: Tikhonov Regularization Based on Near-Optimal Regularization Parameter with Application to Capacitance Tomography Image Reconstruction. J. Journal of Data Acquisition & Processing 19(4), 429–432 (2004) 13. Wang, H.X., He, Y.B., Zhu, X.M.: Regularization Parameter Optimum of electrical capacitance tomography Based on L-curve Method. Journal of Tianjin University 39(3), 306–309 (2006) 14. Jing, L., Shi, L., Zhihong, L.: mage reconstruction iteration algorithm based on 1-norm for electrical capacitance tomography. Chinese Journal of Scientific Instrument 29(7), 1355–1358 (2008) 15. Wang, H., Tang, L., Yan, Y.: Total variation regularization for electrical capacitance tomography. Chinese Journal of Scientific Instrument 28(11), 2015–2018 (2007) 16. Xiao, T.Y., Yu Sh, G., Wang, Y.: Numerical solution of inverse problems. Science Press, Beijing (2003)
Research on Data Preprocessing in Exam Analysis System Ming-hua Zhu College of Computer and Information Engineering, Jiangxi Normal University, Nanchang330022, china
[email protected] Abstract. Data preprocessing is the key to provide high quality data for Exam Analysis System. In order to get the better useful information from the complex and uncertain exam data, it’s necessary to preprocess the source data. In this paper, the source data of Exam Analysis System was analyzed in detail, and found that the source data is inconsistent, redundancy and so on. Therefore, a general method of data preprocessing is introduced in this paper. Keywords: data preprocessing, data mining, exam analysis system.
1 Introduction Data Mining is a kind of database technology developing with database and artificial intelligence. Currently researches on data mining are mainly concentrated on the discussion of algorithms, as a result that data preprocessing has been neglected. However, data of an actual system can seldom meet the requirement of data mining, which has seriously affected the efficiency of data mining algorithms, even lead to a results deviation. According to statistics, the time and cost on data preprocessing has accounted for 60% to 80% of a whole data mining process. As a result, an effective way of analysis and preprocessing on data source has become a key issue to the achievement of a data mining system. Exam Analysis System has accumulated large amounts of data during years of examinations. Exam data can provide a lot of important information after being processed, which makes contribution to guiding teaching, accurate assessment and making the education standardized, modernized and specific. Decisions of high quality must depend on data of that, so it is necessary to preprocess these data before sent into database. The structure of the Exam Analysis System which depends on data mining is shown as Fig. 1.
2 Analysis of Data Source in Exam Analysis System Data preprocessing is an important section in data mining, which provides clean, accurate and concise data for data mining. But actually the original accumulated data are so-called ‘dirty’ with characters of messy, repetitive and incomplete. M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 333–338. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
334
M.-h. Zhu
Top: front-end tool
DSO decision support object
OLAP service
Middle layer: Analysis Services
data warehouse Botton layer: SQL-SERVER Servers Data preprocessing
Old warehouse
Registration information
Return data Data Source
Fig. 1. Architecture of the Exam Analysis System
Data of the Exam Analysis System are widely accumulated, mainly for three sections: the first are data form paper system, including information of questions and exams, which are saved in SQL-Servers; the second is information gathered from each exam center, which are saved as Excel files or SQL files; the third are data of exams gathered from each exam center, which are saved as text files. After analyzing these original data, those characters can be concluded below: (1) The original database is designed depending on rules of relational database, so that those data are entire and consistent with small redundancy except for null numeric in some of fields. As a result, some transform and integrated work should be done in order to fulfill those blank fields. (2) Those data of students’ information from each exam center are basically the same in structure, but totally different in the entity and consistency, for example some major codes are presented in registration form while not in majors’ information form. (3) The same kind of data from different exam centers is different in date representation. For example, some use ‘M’ or ‘F’ in the sex field in information table, while some use ‘Male’ or ‘Female’. (4) The same kind of data from different exam centers is different in date type. For example, some use type of date while some use type of integer to represent date field. (5) The same kind of data from different exam centers is presented in some of exam centers, while others don’t have or uncompleted, such as student’s ID. (6) Examinee registration data from each exam center may more or less have noise data so they should be cleaned before sent into database. (7) Some new information will be added in data delivered from each exam center after examination so these kinds of information should be refreshed automatically.
Research on Data Preprocessing in Exam Analysis System
335
The following form 1 and form 2 are listed the data uploaded from two different exam centers. Differences in data, blank numeric fields and data redundancy have appeared. Table 1. The examinee registration data of examination site A Registeration number 260605100325
Name
Identity card
Sex
Exam course
Major
Li mei
3601011980
F
260605100029
Fu Min
3601031984
M
260610100287
3601021978
F
professional diploma of CS Undergraduates of CS Network
260605100212
Zhang ping Xu Li
260610100116
Hu Wen
3601011979
MC applications comprehensive subject one comprehensive subject one comprehensive subject two comprehensive subject two …
…
…
F
…
M …
Network Undergraduates of CS …
Table 2. The examinee registration data of examination site B Registeration number 260610100102 260605100208 260610100478 260610100038 260605100291 260610100074 …
Name
Identity card
Sex
Age
Huang an Chen feng Cui En-zan Liu Kai Wan Yan Yang Tan …
3601011983 3601021976 3601011980 3601031985 3601011982 3601021979 …
Male Male Male Male Female Male …
27 34 25 28 31 …
Exam course MC applications subject one subject one subject two subject one subject two …
Major U of CS PD of CS Network U of CS Network Network …
3 Data Preprocessing in Exam Analysis System It is convinced after analyzing the data source from exams that it is necessary to do data preprocessing. Data preprocessing in the Exam Analysis System is implemented by the following four ways that is data extraction, data cleaning, data transformation and data integration: (1)Data extraction is the entrance of database for all data. Because a database is an individual data environment, it extracts data from inner database and external database by the way of extracting procedure. Data extraction is technically involved with interconnection, copying and increment. It is unnecessary for data in database to keep the same with the former database real-timely, so data can be extracted regularly. However, when data extractions are executed at the same time, the intervals, sequences and success or not play an important role to the effectiveness of data in the database. (2)Data cleaning is the standardization before data come into data warehouse which is the identification of data’s integrity and consistency. Data cleaning is technically involved with filling in missing values, smoothing noisy data and resolving inconsistencies. There are many way on filling in missing values, such as
336
M.-h. Zhu
ignoring the tuple, filling in missing values manually, using a global constant or attribute mean to fill in the missing value and so on. All these methods above should be selected depending on the detail analysis on the source data. The mainly way to deal with noisy data is by smoothing processing. While dealing with data differences, there are data processing in the check on the effectiveness and uniform representation of data. The check on the data effectiveness depends on checking the length of numeric values, while the check on the data’s uniform is usually in the use of value replacing. (3)Data transformation is a way including the transformation of data type, data aggregation and the generalization and normalization of data in order to make the data more appropriate for data mining. The transformation of data type to numeric values is an easy way, for example to transform type of character to type of numeric for the speed of searching numeric is one order of magnitude faster than searching characters. Data aggregation is to summarize and aggregate data in order to reduce the quantity of issues and accelerate the speed of query and analysis in the future. At the same time, aggregating data out of data in the database will help reduce the size of historical data without loss. The conceptualization of data is to stratify data by concepts that the original in the concept of low-level will be replaced with that in the concept of high-level such as replacing the title session property with chapter property. The normalization is to scaling its values so that they fall within a small specified range, such as 0.0 to 1.0. The commonly used methods for normalization are min-max normalization, Z-Score normalization (or zero-average normalization), normalization by decimal scaling, attribute construction and so on. (4)Data integration is a way to combine data from multiple sources into a coherent data store, as in data warehousing. These sources may include multiple databases, data cubes, or flat files. There are a number of issues to consider during data integration:First is data selection problem. How can the data analyst or the computer be sure that studentid in one database and stuid in another refer to the same attribute? Commonly, it can be solved by convince the source data from database or data warehouse or communicating with those business people directly.Second is the detection and resolution of data value conflicts. When there is attribute values from different sources may differ, which one should be chosen as a standard? For example, in data sources of this system, data gathered from different exam centers, it may happen that some student information is added in one exam center while the student id is the same as another student in another exam center. This can result in a data confliction.Third is data loss. In the data source, there are some missing values which make no effect on the original system. However, in the data warehouse, it is much better that it is filled with values than be blank for misconception may come into the result.Fourth is derived data definition. It is involved with functions for calculating sum and average values and analysis of complex business. The derived data are usually redundant for data using for calculating are stored in data warehouse. However, derived data can much simplify the query, so when data coming into data warehouse, those derived data should be checked for consistency and correctness. It can be known from table 1 and table 2 that in the fields of sex, major and exam course, there is a same content presenting inconsistent data and non-standard. Such as ‘M’and ‘Male’, ‘U of CS’ and ‘Undergraduates of CS’ etc. The field of sex should be represented uniformly by using ‘M’ or ‘F’ and coded referring to major and
Research on Data Preprocessing in Exam Analysis System
337
examination subjects in order to normalize the data in fields of major and exam course. The contents of ID code uploaded from different exam centers are different so that it can be transformed into numeric to accelerate the speed of query. When the field of age values is missing, it can be filled with the values dropped from the ID code. The field of age is regarded redundantly, but it can simplified query and statistics. According to the data preprocessing by data extraction, data cleaning, data transformation and data integration on examinee registration data, the information is shown as Table 3 that achieving the purpose of the same data format and data types, clear data, in order to content to the requirement for data mining and provide supportive to efficient data mining. Table 3. The examinee registration data after data preprocessing Registeration number 260605100029 260605100208 260605100212 260605100291 260605100325 260610100038 260610100074 260610100102 260610100116 260610100287 260610100478 …
Name Fu Min Chen feng Xu Li Wan Yan Li mei Liu Kai Yang Tan Huang an Hu Wen Zhang ping Cui En-zan …
Identity card
Sex
Age
3601031984 3601021976 3601041981 3601011982 3601011980 3601031985 3601021979 3601011983 3601011979 3601021978 3601011980 …
M M F F F M M M M F M …
26 34 29 28 30 26 31 27 31 32 30 …
Exam course 80702 80701 80709 80709 80701 80902 80709 80702 80702 80709 80709
Major
…
…
2331 2210 2332 2331 2210 2332 2332 2331 2332 2331 2331
Microsoft SQL Server 2005 Data Transformation Services (DTS) is a set of graphical tools and powerful programming objects, that can be taken from completely different sources data extraction, transformation, merging into a single or multiple purposes. Data preprocessing tools of this system is programming with DTS package, in the process, we can design a number of steps to complete the data extraction and conversion, which can be a parallel between steps can also be a serial, you can also According to the results of the previous step to determine the next step of the process. To ensure the data extraction and conversion of integrity and consistency, the design of the system will be a few steps in a transaction through the various steps of the returned results to determine whether to roll back the transaction. In the data preprocessing tools use DTS powerful data extraction and transformation capabilities of the system three different data sources to extract data. DTS data collected will be stored in a temporary table first, then through data cleaning and data conversion functions to the data in the temporary table to clean up and converted into the data warehouse, data preprocessing time work to complete.
338
M.-h. Zhu
4 Conclusion During the construction of the Exam Analysis System, data preprocessing is an important and necessary work after design the data warehouse, which is the effective guarantee for data mining. This paper has made a detail analysis on the data source and concluded a detail report on the ways of data preprocessing according to the complexity and inconsistency of data. Among these methods, there are still manual parts. During the further work, it can be explored effective detection and automatic resolving methods for data preprocessing to deal with dirty data in order to provide complete entire solutions.
References 1. 2. 3. 4. 5. 6. 7. 8.
Han, J., Kamber, M.: Data Mining: Concepts and Techniques. Higher China Machine Press, Beijing (2007) Liu, M.-j., Wang, X.-f., Huang, Y.-l.: Data Preprocessing in Data Mining. Computer Science 4, 54–57 (2000) Inmon, W.H., Rudin, K.: Building the Data Warehouse. In: Wang, Z.-h. (ed.), 4th edn., Higher China Machine Press, Beijing (2006) Jian, Z.-g., Jin, X.: Research on Data Preprocess in Data Mining and Its Application. Application Research of Computers (7), 117–118 (2004) Yang, F.x., Liu, Y.c., Duan, Z.h.: An Overview of Data Cleaning. Application Research of Computers 3, 3–5 (2002) You, X., Lou, N.-l., Wang, Y.-x.: Research of data preprocessing in education decision support system. Computer Engineering and Design 28(16), 3985–3988 (2007) Yin, J., Chen, Y., Zhang, G.: An OLAM System Based on Data Warehouse. Computer Engineering (19), 49–51 (2004) SQL Server 2005 Online Help
The Development and Application of Environmental Art Project Which Based on Semiotics in Information Age Ke Zuo School of software, Nanchang University, China
[email protected] Abstract. Environment art engineering development faster and faster in the information age, the paper analyzes the semiotic characteristics of the information age, to promoting the rational use of environment art design engineering,. and to promote the cause of China's environment art engineering for further development. Keywords. Information age; environmental art engineering; software; semiotics.
1 Introduction This thesis mainly from the thinking of the digital age way, From semiotics to study the environmental art design. Due to the environmental functions to modeling forms of transformation is used symbols process, in this process modeling form to become transfer information carrier, i.e., the form of the design symbol, it conveys architectural environment to people with various meanings. While in symbols in the process of innovation and composition, because its diversity, repeatability and the reconstruction in digital era under the influence of the produced a new way of thinking. Thus, based on the analysis of semiotics, analyzes the digital age of design thinking mode, the further deepening digital era under the environmental art design method.
2 Concept and Characteristics of EAE Thinking Digital art design and the traditional close relationship between art and design. Creativity is considered in the design of scientific work of art after the session, is to make the visual design goals can be achieved visualization of ideas can be heard. 2.1 The General Concept of Design Engineering The human brain is thinking, the nature of objective things, and regular reflection of the neurons in the physical, chemical, physical exercises in the form of synthesis, and design thinking is the practice of each designers to achieve design goals Provide important subjective design performance conditions. In the design process to achieve the desired results and the effectiveness of play is directly related to thinking, therefore, demands and designers from the creative aspects of practice, designers have to learn to M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 339–344. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
340
K. Zuo
create a psychology based on the results, summarized some with Regular phenomenon, subject to grasp the essence of good design, and better develop their own creative thinking, the ability to improve the innovative design. 2.2 The Features of Design Project 1) Abnormalities: The abnormal form of thinking and often reflect the nature of the mutation for the development of thinking, across or interrupt logic, this is because creative thinking is not primarily on existing concepts, knowledge of the result of a gradual and logical reasoning process, but on inspiration, Such as intuition or insight implemented as non-logical thinking. 2) Differentiation: Composition between the two are opposites, not only different from each other, negative, antagonistic, and complementary, interdependent, united, the resulting contradictions of creative thinking, and promote the development of creative thinking. This dialectic of thought process and often reflected in the form of integrated creative thinking, creative thinking that is actually a synthesis of various forms of thinking. 3) Open: Mainly refers to the need for creative thinking from many angles and sides, in all directions to examine the issue, but no longer confined to logic, single, linear thinking, thus forming a divergent thinking, reverse thinking, lateral thinking, seeking Different thinking, non-linear thinking and open-minded and other forms of creative thinking. 4) Originality: Is a direct manifestation of creative thinking or signs, and often the outcome of specific performance for the creation of novelty and uniqueness. 5) Initiative: Shows that the subject of creative thinking is to create a purposeful activity, rather than the objective world in the human brain is simple, passive direct reflection, it shows the dynamic nature of human activity and initiative. 6) Comprehensive: Knowledge is the foundation of creative thinking, thinking of the main wealth of knowledge to stand above, easy to create new associations and insights. Often create their own "intellectual cross "results, it is both kinds of knowledge of the mutual penetration, combined with, but also a variety of forms and methods of synthesis of thinking. Art and Design in the digital environment, the richer and more intuitive performance fast. As the information age may be, we can reach more of the design works of art, have a more wonderful fantasies and desire for the creation of a stronger, more creative glow of inspiration and insight, and thus a better image of the rich and quick thinking of our intuition , artistic creation in the environment more like a duck.
3 Information Age Environment Art Engineering Characteristics New social form, the design content of the art form has undergone great changes. Design from the static, rational, single, material to create the dynamic, emotional, complex, non-material to create change. Embodies the essence of art created for the free, non-material design features of the past makes the development of a strong artistic nature of the design elements in art, more and more content becoming more and more artistic designs. The development of non-material design, is both an expression of the
The Development and Application of Environmental Art Project
341
art digital technology on the traditional way of impact, is the perfect combination of technology and art of expression. Material from the traditional design of the transition to non-material design, not only reflects the technology development, but also reflect and meet the people desire for the diversification of lifestyles.Design-led the design direction, different times have different design. When humans gradually from the post-industrial era into the digital age, the interior design concept has changed naturally, mainly in the following two aspects: First, the contradiction from the environment aptitude coordination environment changes,As earth non-renewable resources, environment of increasingly poor escalation, the environmental protection consciousness of people got greatly enhances, hence worldwide open a wave green consumption tide, more and more people in adornment environment, will look to whether it used materials for the whole human environment harmful, and as far as possible need not nonrenewable resources of the earth. To sum up, in the digital environment, environmental art design concept to harmony with environment direction (that is, ecological design concept) change will become a necessity. Second, from the "form follows function" to "form follows emotion " change, environmental design is to meet human "functional requirements" as the core of the movement, and "Spirituality" is only interior design accessories. Therefore, in the design of modernism that "form follows function" design philosophy, and popular. But in the digital environment, network and virtual community does not make the relationship between the people become more closely, but strengthened the solitary and personal survival way, so indoor design also carry the consolation of human spirit and heart of responsibility! Therefore, the famous frog design is put forward "form following emotion" design idea!This equipment will not only be able to design real-time drawing sketches presented in the computer, it can be very realistic to simulate the traditional brush strokes, strength and color. Interior design sketches into a computer brings many benefits, mainly in the following points: (1) computer sketch easily modified so that the designer of the program easier deformation and expansion, thus indirectly inspired the designer's creative inspiration; (2) digital sketch, sketch on the network to get timely and efficient delivery, so that the remote interior design becomes more perfect; (3) sketches the image can be inserted into AutoCAD, as a base map with a fuzzy reference to engineering drawings for the AutoCAD drawing convenience. Design performance is to express the final design, design thinking for designers to visualize the process and specific, and through color renderings, graphic layout, flat layout, flat pattern in elevation, section , The node graph and other means to reflect, its purpose is to allow owners to further understand the designer's design intent, as well as construction workers, construction basis. In the digital environment, environmental art design produced a huge change in performance: 1) The development of computer technology in the past painted flat layout, flat layout, flat pattern in elevation, section, node graph, completely used AutoCAD to draw.Therefore, the design of environmental art show presents a simple, fast, easy to modify, intuitive, and complete benefits. 2) With the three-dimensional network of technology (Network Virtual Reality) of the mature, designers can use Cult3D, Pulse3D, Ser, 3DML other network three-dimensional software, environmental art and design three-dimensional modeling,
342
K. Zuo
so that the owners and stakeholders through the network Environmental art program designed multi-angle, full of observation and evaluation. In this way, to the environmental art design provides remote technical protection. 3)The target of four-dimensional design,As the digital era, computer art design on the environment has shifted from the secondary to, semiotics and other disciplines to support art and design environment to gradually move toward the front, and to awareness of environmental art and design ways of thinking and theory led to innovative Wave.
4 The Using of Environment Art Engineering Building Semiotics In the semiotic sense, the appearance of buildings, materials, uses, etc., are the use of functions from their abstracts, access to the cultural significance of non-architecture, creating a similar system means the system of linguistic signs. 4.1 Architectural Symbols and Cultural Meaning Determinants of the art of architecture culture.Different cultures have a different architecture, thus forming a different architectural symbol of that culture is a symbol of the determinants that affect the building. Such as traditional Chinese architecture in general are neighbor to create the image of the scale, space and environment, rather than the Gothic architecture as the West to exaggerate the extraordinary scale to symbolize God's space and atmosphere, giving the impression that their own Small and bowed to the feet of God, reflecting on the minds of religious shock. Building a symbol of the human spirit is not only the culture, it is also representative of human emotions. Alone building, it is the provision of human and social activities in the residential function of the carrier, all cultural phenomena have occurred in them. Meanwhile, in order to adapt to a wide range of social needs, building also must reflect the times, geographical, national, public and social life of cultural identity and social order to keep pace.Building a more entrusted with people's thoughts and feelings. 4.2 Environmental Art Project under the Information Age in the Use of Architectural Semiotics The digital age, the form and content of environmental art design, great changes have occurred. Design is no longer the focus of the art of some kind of tangible material products, but from the material level gradually move closer to the spiritual level.Embodies the essence of art created for the free, non-material design features of the past makes the development of a strong artistic nature of the design elements in art, more and more content becomes increasingly designed around art. New technology means not only to bring a new way of thinking space and visual space, it also brings a new sensory needs and psychological needs.People-oriented, demand for services will undoubtedly keep the Art of Design and create one personalized to meet the diversified needs, which will lead to the design would be to diversify the face of art, personal. Gehry design works in small doping social and ideological things. He usually polygonal surface, tilt the structure, form and inverted form of a variety of substances
The Development and Application of Environmental Art Project
343
applied to the design and visual effects to go. Gehry uses fault geometry to break tradition, to him, breaking a means to explore the social order is not clear. Gehry out in the form of features, created not a whole structure, but a successful abstract ideas and the city agencies. In many ways, he treated the same as building work as sculpture, this structure through three focus on a variety of forms. Art is often inspired by Gehry's birthplace, his interest in the arts can be learned of his architectural works. At the same time, art made him the first time the use of open architecture structure, and people think it is an invisible change, rather than deliberately. Gehry's building is often surreal, abstract and occasionally people are deeply confused by the message so it is often misleading. Even so, Gehry's building is showing its unique, elegant and mysterious atmosphere. Gehry materials using a variety of substances, using a variety of architectural forms, and humor, mystery, and his dreams of building systems integration. He said: "I like to see this in the beauty of the building process, which the U.S. has often lost in the manufacturing process technology. " Gehry in the early work on the bold use of open space, a variety of raw materials And not to carry out the construction of formality. Gehry's architecture also includes a common process, a continuation of life, the evolution of life and the lives of such growth. Presentation of information in this chapter by examples of semiotics in the construction era of environmental art project in the new development. As Roland Barthes in the "semantic object" wrote: "The meaning has always been a cultural phenomenon, is the product of culture; However, in our society, this cultural phenomenon are constantly being Naturalization. words that we believe in a pure object in the transitive situation, and again into the natural meaning of the phenomenon. We believe we find ourselves at a purpose, function of objects, the complete control of the formation of the practice of the world, in fact, through the objects, we find ourselves in a sense, reasons, excuses posed by the world: features derived from Symbols, and this symbol has been transformed into functional re-display. I believe it is this will be a natural process of cultural transformation was established ideology of our society.” Therefore, we conduct multi-disciplinary environment, art and design process, the need to support multi-discipline, but also the so-called Open design, and semiotics, as construction on the environment, the use of art and design, art semiotics also played Environmental Art Design A great role, and has unlimited potential.
5 Conclusion and Outlook As awareness of human expression, the means and methods of conveying information is one of the environmental design is similarly dependent on the support of various disciplines. As designers we have to learn to use other disciplines to design more effective function. Environmental Art and Design, for its part, to include architecture, art and other symbols, by symbols of these elements at different levels, the success of the designer selection, combination, conversion, regeneration of these elements, together become the referent of his thoughts Symbols, as their common recognition of the symbols and the audience, this is the real form of communication, information accurate and complete communication, design of the thinking process is perfect. The purpose of environmental art design is the exchange of people, including the
344
K. Zuo
application of semiotics in various subject areas, including the principles and methodology in environmental art and design in research and practice is quite common and in-depth, but also undoubtedly become inevitable in our design Tools, and environmental art design to truly become a no limit design.
References 1. 2.
3.
Xige : Art multimedia technology applications in the environment (January 2004) Feng, Y.: Architectural design concept of the software supporting the three key technologies, based on Creator & Vega, 3ds max and AotoCAD, Sichuan Construction (October 2006) Feng, G., Zhang, J.: Architectural decoration based on virtual reality system design and managent. Hebei Architectural Science and technology (June 2006)
Rotor Time Constant Estimation for the Vector Controlled Induction Motor Drive Based on MARS Scheme Hua Li and Shunyuan Zhou Institute of Electric Power System and Motor Drives, College of Information Science and Engineering Northeastern University Shenyang, Liaoning Province, China
[email protected],
[email protected] Abstract. In this paper, Model Reference Adaptive System (MRAS) is presented for the rotor time constant estimation for induction motor based on the regulation of instantaneous reactive power. The estimated rotor time constant is used as feedback for calculating the slip speed in an indirect vector control system. This method avoids using pure integration and stator resistance in the reference model. Moreover, the steady state of the reactive power eliminates the derivative terms in the adjustable model. The structure of estimator is very simple and the technique is robust to variations of motor parameters. Simulation results are presented the robustness and accuracy of the proposed schemes and show that good tracking capability and fast responses have been achieved. Keywords: Rotor Time Constant, MRAS, Induction Motor, Vector Control, Reactive Power.
1 Introduction In recent years, the induction motor (IM) has been widely used in industrial application due to its simple structure, great reliability and low costs. Control techniques of these drives are well treated in the literatures. The vector control is a sophisticated control method. It is based on rotor field oriented control according to rotating frame transformation, and has a decoupling control between torque and flux of the IM drive and consequently dynamic performances similar to those of a DC machine. However, rotor time constant (τr) is a very important parameter, which is required in indirect field-oriented control (IFOC) system of IM drive. The rotor time constant may have variation with working condition change, especially temperature change, which may lead to improper flux orientation, improper stator current decoupling, and hence deterioration of dynamic performance of the IM drive. In order to achieve good performance of vector controlled IM drive, many different rotor time constant estimation schemes have been proposed, for example, signal injection-based method [1], Model Reference Adaptive System (MRAS), Extended M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 345–352. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
346
H. Li and S. Zhou
Kalman filter [2]. In these methods, MRAS is popular due to its simplicity, less computation time and good stability [3-7]. In this paper, the instantaneous reactive power has been used in the proposed technique for estimating the rotor time constant of IM drive. Most of the MRAS schemes for estimation of rotor speed or rotor time constant require flux estimation. However, this method is not like that. So, it overcomes the reflections from pure integration and stator resistance [5]. The reference model has no influence from the parameter variations. The steady state of the reactive power eliminates the derivative terms in the adjustable model. The structure of the model is simplified. Therefore, the accuracy of parameter estimation is verified and is not suffered from integratorrelated problems at low speed. The validity and robustness are proved by simulation.
2 Inverse Rotor Time Constant Estimation 2.1 Mathematical Model of Induction Motor The electromagnetic behavior of IM in the d-q synchronously rotating reference frame can be expressed by equation (1). ⎛ ⎜ Rr + ( p + jω ) σ Ls ⎛ vs ⎞ ⎜ ⎜ ⎟=⎜ RL ⎝0⎠ ⎜ − r m ⎜ Lm ⎝
( p + jω ) Lm
⎞ ⎟ Lr ⎟ ⎛ is ⎞ ⎟ ⎜ψ r ⎟ Rr + ⎡⎣ p + j ( ω − ωr ) ⎤⎦ ⎟⎟ ⎝ ⎠ ⎠
(1)
where ω = ωr + ωsl , σ = 1 − L2m / ( Ls Lr ) and p is the differential operator. The voltage, current and flux space vectors are given as x = xd + jxq . The instantaneous reactive power can be expressed as Q1 = vqs ids − vds iqs
(2)
According to the equation (1), the (2) can be rewrote, a new expression of Q1 is
(
)
(
)
Q2 = σ Ls piqs ids − pids iqs + σ Lsω ids2 + iqs2 −
(
)
(
Lm L pψ dr iqs − pψ qr ids + ω m ψ qr iqs + ψ dr ids Lr Lr
)
(3)
In the IFOC system, ψqr=0 and ψdr=ψr. In the running process of motor, ids=ψr/Lm. When the IM drive is working in the steady state, the derivative terms are zero. The equation (3) can be simplified, the expression of reactive power resolves to
(
)
Q3 = σ Lsω ids2 + iqs2 + ω
L2m 2 ids Lr
(4)
In the equation (1) and (4), it is found that the stator resistance does not affect the reactive power. Also, pure integration terms are eliminated, and the question of integral drift can be resolved.
Rotor Time Constant Estimation for the Vector Controlled Induction Motor Drive
347
2.2 Model Reference Estimation Structure The structure of MRAS consists of a reference model, an adjustable model and an adaptation mechanism which is found out to minimize the error between two models, as illustrated in Fig 1. In this paper, Q1 is considered as the reference model, which represents the components of the reactive power in terms of accessible stator variables, that is, stator currents and voltages. Q3 is chosen as the adjustable model which is free from rotor flux and derivative terms, and it is dependent on slip speed (ωsl). The error signal is fed to the adaptation mechanism. According to the conditions of rotor fieldorientation, the following expression can be gained.
ωsl = Lm iqs / (τ rψ r )
(5)
According to equation (5), the rotor time constant can be calculated. The structure of MRAS using the reactive power is illustrated in Fig 1.
Fig. 1. Structure of the new MRAS using reactive power
2.3 Model Reference Estimation Structure
The proposed MRAS based on rotor time constant have been shown in the above part. The reactive power variation is defined by
ε = Q1 − Q3
(6)
Substituting Equation (2) and (4),
ε = vqs ids − vds iqs − σ Lsω ( ids2 + iqs2 ) − ω
L2m 2 ids Lr
(7)
This variation is used by the adaptation mechanism to generate the estimated slip speed and make it converge towards its actual value. The adaptation mechanism must be designed in order to obtain a fast and stable time response. The adaptation mechanism is based on the Popve’s hyperstability concept, which mainly concerns the stability properties of a class of feedback systems as illustrated in
348
H. Li and S. Zhou
Fig.2. Usually, r=0 , gain u=-w. Two conditions are required when the system is said asymptotic hyperstable. (a) The transfer function of the feed forward linear time invariant block must be strictly positive real. (b) The nonlinear time varying block must satisfy the Popov’s integral inequality η ( t0 , t1 ) = ∫ yT ( t ) w ( t )dt ≥ −γ 02 t1
t0
∀t1 ≥ t0 , γ 02 > 0
(8)
where w is the feedback block. The term of w is input and ε is the output of the linear forward block.
Fig. 2. Standard block diagram of nonlinear feedback system
According to the hyperstability concept, a state model is represented as pε = Aε − w
(9)
Assume y=Cε, where C=1/[σLs(idr2+iqr2)+(Lm2/Lr)idr2], and constitute the linear part of the standard nonlinear feedback systems, whereas rotor time constant estimation constitutes nonlinear part. The equivalent MRAS is shown in the Fig. 3.
Fig. 3. Equivalent MRAS for the proposed scheme
Substituting y and w into equation (8), the following expression can be gained ⎡
η ( 0, t1 ) = ∫ yT (ωsl − ωsl ) ⎢( ids2 + iqs2 ) + ω t1
0
∧
⎣
L2m 2 ⎤ ids ⎥dτ ≥ −γ 02 Lr ⎦
(10)
Rotor Time Constant Estimation for the Vector Controlled Induction Motor Drive
349
The adaptation mechanism can be chosen by the expression ∧
∧
ωsl = ∫ F1 ( y ,τ , t )dt + F∧2 ( y, t ) − ωsl (0) 0 ωsl (0) t1
According to the actual situation, choose the
, and substitute (11) into (10), then
∧
η ( 0, t1 ) = ∫ ε T ( ∫ F1 ( y ,τ , t )dt − ωsl (0) − ωsl )dτ + ∫ ε T F2 ( y , t )dτ ≥ −γ 02 t1
t1
t0
0
(11)
t1
(12)
0
To make Equation (12) be satisfied, the following inequality can be used
∫
t1
0
pf ( t )kf ( t ) dt =
k 2 k ⎡ f ( t1 ) − f 2 ( 0 ) ⎤⎦ ≥ − f 2 ( 0 ) 2⎣ 2
(13)
So, have pf ( t ) = ε T ∧
kf ( t ) = ∫ F1 ( y ,τ , t )dt − ωsl (0) − ωsl t1
(14)
0
According to the equation (14), the following result can be obtained F1 ( y,τ , t ) = kτ I ε T
kτ I > 0
(15)
F2 ( y, t ) = kτ P ε T
kτ P > 0
(16)
and
The adaptation mechanism can be gained ∧
t1
∧
ω sl = ∫ kτ I ε T dt + kτ P ε T + ω sl (0)
(17)
0
where kτ I , kτ P are the PI controller gains of the adaptation mechanism. Therefore, a PI controller is sufficient to satisfy Popov’ s integral inequality. These satisfy both the Popov’s criterion and confirm the stability of the system.
3 Simulation Results The above presented procedure has been simulated using MATLAB/Simulink to verify the effectiveness of MRAS for estimating the rotator time constant. Fig.4 shows the block diagram of IFOC based on MRAS using the reactive power. The proposed estimation scheme has been simulated on a 1.1-kW four-pole squirrel cage induction motor, and the parameters are summarized in the Table 1. The motor model is represented by equation (1) as well as the following mechanical equation pωr = −
np B ωr + (Te − Tl ) J J
(18)
The performance of the estimator has been studied in terms of its ability to converge to the actual rotor time constant and the response of the estimator towards parameters change. Fig.5 shows the simulation result of rotor time constant for the vector controller IM. The rotor time constant estimator was simulated for a load of 2 Nm and a
350
H. Li and S. Zhou
Table 1. Induction motor parameters Parameters Nominal power (kW) Stator line voltage (V) Frequency (Hz) Stator resistance (Ω) Rotor resistance (Ω) Mutual inductance (H) Stator leakage inductance(H) Rotor leakage inductance(H) Pole pairs
Value 1.1 380 50 6.6 5.5 0.454 0.021 0.021 2
Fig. 4. Block diagram of IFOC based on MRAS
reference speed of 800 rpm. At time t=2s, the rotor resistance (Rr) of the induction motor has a step response which is increased by 50% of Rr , and at time t=4s, the rotor time constant is reduced to the normal value, Rr However, the rotor time constant change slowly with time change in the real drive. In this simulation, the step variation is to verify the robustness of the proposed estimator. From the Fig.5., it is found that the value estimated of reactive power ( Qest) can track the reference reactive power (Qref) better compared by the method using flux calculator in [10]. Therefore, the method has good tracking capability and fast response. Fig.6 shows the rotor time constant estimation at low speed. At time t=2s, the rotor resistance is increased by 50%. Compared the method using the rotor flux information, the proposed technique can estimate the rotor resistance at low speed accurately.
Rotor Time Constant Estimation for the Vector Controlled Induction Motor Drive
351
0.4 Qref 0.3
0.2
0.1
0 Qest -0.1 0
1
2
3
4
5
6
5
6
time (sec)
(a) 1.2 1 0.8
d-axis stator current
0.6 0.4 0.2 0
q-axis stator current
-0.2 0
1
2
3
4
time (sec)
(b) 20
15
10
5
0 1
1.5
2
2.5
3
3.5
4
4.5
5
time (sec)
(c) Fig. 5. Simulation results of the proposed system: (a) the reactive powers, (b) d and q-axis stator current, (c) inverse rotor time constant Tr-1
50 40 30 20 10 0 -10 0
0.5
1
1.5
2
2.5
3
3.5
time (sec)
Fig. 6. Inverse rotor time constant estimation at a low speed
4
352
H. Li and S. Zhou
4 Conclusion In this paper, the rotor time constant estimator based on MRAS is proposed using the function obtained from instantaneous reactive power. The technique has no integrators in the implementation and it is robust to the variation of stator resistance. The rotor time constant can be estimated correctly at steady state of the reactive power. From the simulation results, the rotor resistance estimation has proved the effectiveness of MRAS. The adaptation mechanism has good tracking capability to parameter variations.
References 1. Wade, S., Dunnigan, M.W., Willianms, B.W.: A new method of rotor resistance estimation for vector controlled induction machines. IEEE Trans. Ind. Electron. 44(2), 247–257 (1997) 2. Garcia Soto, G., Mendes, E., Razek, A.: Reduced-order observers for rotor flux, rotor resistance and speed estimation for vector controlled induction motor drives using the extended kalman filter technique. IEE Proc. Electr. Power Appl. 146(3), 282–288 (1999) 3. Peng, F.Z., Fukao, T.: Robust speed identification for speed-sensorless vector control of induction motor. IEEE Trans. Ind. Applicat. 30(5), 1234–1240 (1994) 4. Zorgain, Y.A., Koubaa, Y., Boussak, M.: Rotor resistance estimation for indirect stator flux oriented induction motor drive based on MRAS Scheme.In: STA 2009, REM-627, pp. 1347–1362 (2009) 5. Bin, H., Wenlong, Q., Haifeng, L.: A novel on-line rotor resistance estimation method for vector controlled induction motor drive. In: Proc. Conf. Rec. IEEE IPEMC Conf., vol. 2, pp. 655–660 (2004) 6. Marcetic, D.P., Vukosavic, S.N.: Speed-sensorless AC drive with the rotor time constant parameter update. IEEE Trans. Ind. Applicat. 54(5) (2007) 7. Haron, A.R., Idris, N.: Simulation of MRAS-based speed sensorless estimation of induction motor drive using MATLAB/SIMULINK. In: PECON 2006, pp. 411–415 (2006) 8. Maiti, S., Chakraborty, C., Hori, Y.: Model reference adaptive controller-based rotor resistance and speed estimation techniques for vector controlled induction motor drive utilizing reactive power. IEEE Trans. Ind. Electronics 55(2), 594–601 (2008) 9. Peng, F.Z., Fukao, T., Lai, J.S.: Low-speed performance of robust speed identification using instantaneous reactive power for tacholess vector control of induction motors. In: Conference Record of the 1994 IEEE-IAS Annal Meeting, vol. 1, pp. 509–514 (1994) 10. Ta, C.-M., Uchida, T., Hori, Y.: MRAS-based speed sensorless control for induction motor drive using instantaneous reactive power. In: Conf. Rec. of the IEEE-IECON, pp. 1417– 1422 (2001)
Small-World Request Routing System in CDNs Lan Li* 235 Nanjing Road East, Qingshanhu Distric School of software,Nanchang University 330029 Nanchang, PR. China
[email protected] Abstract. In this paper, we present a novel small-world method for request routing in CDNs. CDNs have multiple servers carry the same content. Request routing system is used to choose a proper replica server that has the requested content and route the incoming request to that sever. To present the proper strategies on choosing the replica server, we use small-world distributed hash tables to create the new replica server and relax the hot spots in replica servers. According to the experiment results, the trade-off between the scalability and robustness gained from small-world distributed hash tables reduce the average query delay. Keywords: We would like to encourage you to list your keywords in this section.
1 Introduction Content delivery(or distribution)networks(CDNs)[1, 2] is an effective approach to improve Internet service quality. It has recently been proposed to improve the performance of the response time, bandwith and the accessibility by using index, cache, stream splitting, multicast and other technologies[2-4]. Request routing system is used to choose a proper replica server that have the requested content and route the incoming request to that sever. There are many techniques have been proposed to guide users to use a suitable server among a set of replica servers, such as client multiplexing, HTTP redirection, DNS indirection, anycasting and peer-to-peer routing, in which peer-to-peer systems build the information retrieval network one the members of replica servers themselves instead of relying on a dedicated infrastructure like the traditional CDNs do. Recent work also has shown that the request interests is a kind of social network and exhibits small-world behavior. Characterizing such behavior is important for the proper evaluation of large-scale content delivery techniques. Inspired by this, this paper present a novel request routing scheme based on small-world theory.
2 Related Work If one object is popular, then the probability of finding more than one replica server to store it becomes higher. Content Distribution Networks (CDN) have multiple servers *
This work is supported by Youth Science Foundation(GJJ11038), Department of Education, Jiangxi Province.
M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 353–359. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
354
L. Li
carry the same content whether it is web content or any other application. Such a problem of multiple serving nodes is similar to peer-to-peer network, however, in contrast to peer-to-peer network, CDNs servers are well maintained by some professional personnel and lost self-organizing features and dynamics. In order to present a suitable content distribution mechanism for the individuals content producers, peer-to-peer network can be introduced into CDNs. When providing servers from individuals, CDNs’ servers are often very dynamic, and might leave a network even if they were in the middle of serving an object to a client node. Thus, it is necessary to provide certain number of replica servers for a client node. In the event that one or more of the serving nodes disappear, a client does not have to restart the download of the entire object from another serving node. According to this, Distributed hash tables(DHT) can be introduced to store index information in CDNs[5]. Distributed hash tables(DHT) associate with an address based on the numerical distance between their hash keys. DHT is widely used to either store index information about the location of data or to store the data itself. However, because a message traverses a typical DHT on links that are completely location unaware, a message originating in two very near nodes may take lots of hops. This problem can be solved by integrating the location information into the DHT. However, the integration of location information often brings the problem of trade-off between the cost of robustness and scalability by reducing the randomness that was intentionally integrated into the design of DHT. In order to achieve the balance of structure and randomness, small-world graphs may give out the solution. Following Watts and Strogatz’ findings[6], the structural properties of small-world graphs typically exhibit a short path length between any two vertices and strong clustering behavior. Small-world graphs exhibit connectivity properties that are between random and regular graphs. Like regular graphs, they are highly clustered; yet like random graphs, they have typically short distances between arbitrary pairs of vertices. It has been shown that many networks have similar small-world property. After Watts and Strogatz’ finding, Kleinberg[7] proposed two engaging questions: “why should there exist short chains of acquaintances linking together arbitrary pairs of strangers?” and “why should arbitrary pairs of strangers be able to find short chains of acquaintances that link them together?” Kleinberg’s studies on the networks with small world characteristics show that searches can be efficiently conducted when the network exhibits the following properties: 1) each node in the network knows its local neighbors; 2) each node knows a small number of randomly chosen distant nodes, with probability proportional to 1/d where d is the distance. A search can be performed in O(log2N) steps on such networks, where N is the number of nodes in a network. Inspired by Kleinberg's work, DHT Symphony is designed by Manku et al.'s. In a Symphony network with n nodes, each node randomly chooses a node identifier from the unit interval [0,1) and positions itself on a ring. The key idea of symphony[8] is to arrange all participants along a ring and equip them with long distance contacts drawn from a family of harmonic distributions. In symphony, each node establishes links with its two immediate neighbors on the ring as well as q long distance links chosen with the help of a probability density function pn(x)=1/xln(x), in which x [1/n,1]. In order to establish a long link, a node draws a distance d from pn(x) and then searches for the manager of the point at distance d away from itself on the ring. Symphony's pn(x)
∈
Small-World Request Routing System in CDNs
355
implies that the probability density of a long distance link being established at a distance d is inversely proportional to d. According to the recent work, the proposed content distribution mechanism contains two parts: firstly, construct a content mechanism network based on the collections of popular routes queries. Secondly, give the solutions of request routing system in CDNs.
3 Small-World Content Distribution Mechanism 3.1 Construction and Initialization According to the discussion above, obviously, it is important to make data stored close to the requestors in content distribution networks. The proposed small-world DHT determines the distance between nodes by underlining DHT proximity routing, which is different from the measurement of distance between two ip addresses. In order to record the incoming queries, each node maintains a routing table to record the next-hop nodes these queries traveling through before they arrive in the key node. This routing table is a Hash table, in which stores the key and the location of the data. In order to generate a content delivery network based on DHT, each node is given a node identifier based on the content weights and nodes in CDNs is linked to the direct nearest neighbors nodes they know. The identifier of a node is generated by a hash function based on the contents it provides. If there are other links to that node in CDNs, these links connected to the other neighbors. The content weight is decided by the number of queries in CDNs. Algorithm1 describe the initial process of constructing a content distribution network. Algorithm 1: construction of content distribution network
program Construction( int max_query) begin If a key’s query count equals to max_query Choose a new node act as server; New server send joining message; give the node a identifier Ni according to its content weights; end if search the identifier Ni in the routing table with maximum query rate; If no such identifier Ni Add Ni to the routing table with maximum query rate; Inform all nodes in system by sending messages to them; End if end. 3.2 Reinforce the Structure with Small-World Feature According to Algorithm 1 described above, a initial content delivery structure is built. In order to utilize the small-world phenomenon of the interests from users, some additional links are added based on a probability density function over the weights of contents between nodes in the content distribution network. Inspired by symphony, the
356
L. Li
proposed method makes use of the weights of queries’ interests and contents. The smaller the differences between two nodes’ content weights, the higher probability that they are joined by a link. According to this, a node add a new link by computing the difference from the probability density function pn(x)=1/xln(x) and search a node with the difference in the whole network. Then the node sends messages to a number of selected neighbors with its coordinate and the content weight difference. These selected neighbors check their routing tables for any content nodes with the computed content weight difference by probability density function. If find nothing, then a new value is drawn by probability density function and begin the process again. If there are some hits return, then an addition link between the node and the found node is established.
4 Query and Routing Mechanism This section details the query and routing operation in CDNs. Because of the proposed content distribution network construction mechanism adds the content weight according to the query interests, queries can search the replicas in the relevant interests servers. Therefore, when routing a query, a client node looks for the key-k forwards the lookup to its neighbor nodes with identifier-n and these neighbor nodes look up the minimized difference between k and n in the identifier space. If a matched node is found, it returns the query results to the handler, otherwise forwards the query to its next neighbor. Because of using the small-world feature to make the CDN structure revised, the content node make the decision based on the normal links and the added links when selecting the next hop. Algorithm2 present the details of query and routing mechanism. Algorithm 2: query and routing in content distribution network
program Nexthop (key k) begin if Islowest(k-n)//look for the neighbor node n with the lowest difference return nexthop as n ; increase the query weights of node n; end if end. Theorem 1. Given a CDN of N nodes, with the number of maximum query size M, the average search path length for search across both normal links and added lins is O(log(2N/M)). Proof: Note that the basic structure is formed by the normal links and the added links, query is built on all links in the network. During the search process, the algorithm maintains a search-hit list in each server. For every queries, a server node replaces the lowest hit rate in the search-hit list with the highest hit rate with probability of N/(N + M).
Small-World Request Routing System in CDNs
357
5 Experiments Results and Analysis The proposed strategy inspired by the small-world DHT protocol in peer-to-peer network present solutions for routing a query to the nearest server node in CDNs. In this section, the experiments will show that the proposed algorithm is very efficient at minimizing the query delay. We measure the content distribution system by using a simulator built on the top of Chord. The nodes are generated by the function based on uniform distribution. The parameters of the experiment are given in Table 1. Table 1. The parameters used in the experiment. Descriptions The rate of key production The rate of query arrival The latency of next hop The average throughput of replica data The content weight per node
Value
α=4 qa 5ms 100/s 90% of each capacity
In figure1, we compare the average query delay among normal network without CDN, basic CDN and CDN with small-world links. It shows that the average query delay in CDN with small-world feature has the advantage in he average query delay.
350
) s m 300 ( y a l 250 e d g 200 n i y r 150 e u q 100 e g a r 50 e v A
0
Normal CDN CDN with small-world link
2
4
6
8
10
12
14
Times(s)
16
18
20
22
Fig. 1. Average query delay (qa = 100/s)
Figure2 shows the average query hops comparison among normal network without CDN, basic CDN and CDN with small-world links.
358
L. Li
4 3.5 3
s p o h 2.5 y r e u 2 q e g a 1.5 r e v A
Normal
1
CDN
0.5
CDN with small-world link
0 2
4
6
8
10
12
14
16
18
20
22
Times(s)
Fig. 2. Average query hops (qa = 100/s)
It shows that the CDN with small-world links performs better than the normal network without using content distribution method.
6 Conclusions In this paper, we present a content distribution mechanism by using the overlay DHT with small-world features. The initial construction of content distribution network is based on a structured overlay model, afterwards, reinforce the structure by adding additional links according to the query interests. The mechanism present in the paper shows an efficient way of request routing. The experiments show that the delay and hops in CDN with small-world features have advantages compared to the normal network without using CDN and normal CDN without small-world features.
References 1. 2. 3. 4. 5.
Peng, G.: CDN: Content distribution network. Arxiv preprint cs/0411069 (2004) Pathan, M., Buyya, R., Vakali, A.: Content Delivery Networks: State of the Art, Insights, and Imperatives. Content Delivery Networks, 3–32 (2008) Pathan, M., Buyya, R.: A Taxonomy of CDNs. Content Delivery Networks, 33–77 (2008) Chen, Y.: Dynamic, Scalable, and Efficient Content Replication Techniques. Content Delivery Networks, 79–104 (2008) El Dick, M., Pacitti, E., Kemme, B.: Flower-CDN: a hybrid P2P overlay for efficient query processing in CDN, pp. 427–438. ACM, New York (2009)
Small-World Request Routing System in CDNs 6. 7. 8.
359
Watts, D.J., Strogatz, S.H.: Collective dynamics of small-world networks. Nature 393, 440–442 (1998) Kleinberg, J.: Small-world phenomena and the dynamics of information, p. 431. MIT Press, Cambridge (2002) Manku, G.S., Bawa, M., Raghavan, P.: Symphony: Distributed hashing in a small world, pp. 10–10. USENIX Association (2003)
Experimental Study on Simulated Cerebral Edema Detection with PSSMI Gui Jin1, Mingxin Qin1,*, Chao Wang2, Wanyou Guo2, Lin Xu1, Xu Ning1, Jia Xu1, and Dandan Gao1 1 College of Biomedical Engineering and Medical Imaging, Third Military Medical University, Gaotanyanzheng street 30, Shapingba district, 400030 ChongQing, China 2 College of Electronic Engineering, Xidian University, Taibai South Road 2, 710126 Xi’an, China
[email protected],
[email protected],
[email protected] Abstract. Based on PSSMI method, one new detection system and the physical model of cerebral edema, the experimental study of simulated detection of cerebral edema was carried out. We applied three kinds of excitation signals with different frequencies. Three kinds of NaCl solutions were used to simulate the brain tissue, cerebral edema and calibration solution, whose conductivities were 0.133s/m, 0.194s/m and 3.6s/m respectively. To simulate the volume change of cerebral edema, the solutions were increased from 5 to 100 ml with an infusion pump. The phase resolution was up to 0.005° and the range of the gain was -10 35dB in the detection system. The experimental results show that the phase shift is directly proportional to volume, conductivity and frequency. The experimental study suggests that the PSSMI method has the potential of being a simple method for cerebral edema detection.
~
Keywords: magnetic induction; phase shift; cerebral edema.
1 Introduction Phase shift spectroscopy of magnetic induction (PSSMI) method applies a certain frequency range of magnetic fields to induce eddy current in biological tissues and the phase shift spectroscopy between the excitation magnetic field and the inductive magnetic field is then detected. The phase shift spectroscopy will be used to measure the occurrence and progress of cerebral edema and have an important significance for the detection of cerebral edema [1-3]. PSSMI needs the high precision of phase shift. In the previous work, we have established a detection system based on phase locking amplifier(SR844), whose precision was only 0.02° and structure was complicated. So we redesigned a new detection system of cerebral edema. In order to test the new system performance a physical model was used to simulate cerebral edema and detection experiments were made. *
Corresponding author.
M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 361–367. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
362
G. Jin et al.
2 Configuration of Our Detection System The detection system includes an excitation source, excitation and detection coils, a physical model of cerebral edema and a phase detector. The excitation current from the excitation source flows into the excitation coils to generate the magnetic field, the magnetic field generates the inductive magnetic field in the physical model. The excitation field and the inductive field are measured by the detection coil. The phase shift between the signal from the detection coil and the reference signal from the excitation source is measured by the phase detector and displayed on a screen.
Fig. 1. Detection System includes (1) phase detector, (2) excitation source, (3) brain edema physical model, (4) excitation coil, (5) detection coil, (6) skin dilator, (7) infusion pump and (8) beaker.
Fig. 2. Schematic Diagram of Detection System.
2.1 Excitation Source The excitation source can output two sinusoidal signals. One signal is used as the excitation signal and the other is the reference signal. The excitation source provides three frequencies and the range of output power from the excitation source is
Experimental Study on Simulated Cerebral Edema Detection with PSSMI
~
363
1.5dB 33db [2-3]. The frequency stability is the order of magnitude of 10-4, the distortion reaches 10-2 10-4 and the SNR is 30 60dB in the excitation source.
~
~
2.2 Physical Model The physical model includes an organic glass ball and a skin dilator[6], which can simulate the irregular geometry, different volumes and different conductivities of cerebral edema. The excitation coil and detection coil are fixed to organic glass ball’s neck and middle separately. The diameter of excitation coils and detection coils are 68mm and 220mm, the number of turns of both coils are 10 and the distance of two coils is 100mm [4-5]. The solution in the beaker is evenly input into the skin dilator by the infusion pump (ZNB-XY1). 2.3 The Phase Detector The phase detector for measuring the phase shift includes the band-pass filter, amplifier, AD, FDGA, DSP, flash memory and LCD. The phase detector can be calibrated and the temperature drift and the noise can be eliminated by a calibration software. You can adjust frequency, set gain, view waveform, measure phase shift and transmite data to computer on the phase detector. The parameters of our phase detector are as follows: the range of phase measurement: 0 180°, the phase precision: 0.005°, the range of gain: -10 35dB, once measurement time: 3 7s. The 12h data measured can be saved in the phase detector.
~
~
~
3 Experiments on Simulated Cerebral Edema Detection 3.1 Design of Experiments The excitation source was connected to phase detector as Fig 2. The solution of 2800ml NaCl was injected into the physical model to simulate the brain tissue and the conductivity of the solution was 0.133s/m. Two beakers of NaCl solution was prepared to simulate the cerebral edema and physiological saline and the volume of each beaker was 100ml. The conductivities of two solutions were 0.194 s/m, 3.6 s/m respectively. The skin dilator made of plastic film was connected to the infusion pump. The infusion pump evenly transported the solution from the beaker to the skin dilator, the speed of pump was set to 2000ml/h and each time volume was set to 5ml. The phase shifts were measured by the phase detector under the conditions of two different solution volumes and three different frequencies. Process: First, one simulated solution and operating frequency were selected. Whenever each 5ml solution was injected, the phase detector measured the phase shift in time. The solution in the skin dilator was increased from 0 to 100ml, the phase detector got 21 datas in each measurement. After this measurement was over, the infusion pump drew out 100ml. Again, 20 measurements were repeated in the above method. Second, The operating frequency was changed and the above step was repeated. Third, The solution was changed. Step one and two were repeated.
364
G. Jin et al.
The data was processed as follows: 20
∑ ( θ ji − θ j 0 ) X obi =
j =1
(1)
.
20
Here, j is the measurment number from 1 to 20, i is the data number in one measurement. θji stands for the phase shift and θj0 represents the phase shift when the skin dilator is empty. Xobi represents the average phase shift over 20 measurments[3]. 3.2 Experimental Results Fig 3 shows the phase shift curves generated by the simulated cerebral edema solution in three kinds of frequencies. From the curves, the phase shift increases as solution volume increases and the change ratios are different. The change of phase shifts are 0.007°, 0.479° and 0.533° using three operating frequencies respectively when the volume increases from 0 to 100ml. 0.6 f1
phase shift ( °)
0.5
f2
f3
0.4 0.3 0.2 0.1 0 0
20
( )
40 60 volume of the solution mL
80
100
phase shift ( °)
Fig. 3. The phase shift of cerebral edema.
0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
f1
0
f2
20
f3
( )
40 60 volume of the solution mL
Fig. 4. The phase shift of physiological saline.
80
100
Experimental Study on Simulated Cerebral Edema Detection with PSSMI
365
Fig 4 shows the curves generated by the physiological saline. From the curves the phase shift increases also as the volume increases, and the change ratios are also different. The change of phase shifts are 0.011°, 0.584° and 0.853° using three operating frequencies respectively when the volume increases from 0ml to 100ml. Fig 5, 6 ,7 are the curves of two solutions using three operating frequencies, in which the experiment datas are the same as Fig 3, 4. The changes of phase shift generated by two solutions are very little, the maximum is only 0.011° when the frequency is f1. The change is larger and the maximum is 0.584° as to frequency f2. The change is the largest and the maximum is up to 0.853° as to frequency f3. From the curves, we know the phase shift of the physiological saline is larger than the cerebral edema’s in the same volume and frequency. 0.012
phase shift(°)
0.01 0.008 0.006 0.004 0.002 0 0
10
20
30
40 50 60 the volume of the solution(ml)
cerebral edema
70
80
90
100
80
90
100
physiological saline
Fig. 5. The phase shift caused by frequency f1.
0.7 0.6
phase shift(°)
0.5 0.4 0.3 0.2 0.1 0 0
10
20
30
40
50
60
70
the volume of the solution(ml) cerebral edema physiological saline
Fig. 6. The phase shift caused by frequency f2.
366
G. Jin et al. 0.9 0.8
phase shift ( °)
0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0
10
20
30
40
50
the volume of the solution cerebral edema
60
(mL)
70
80
90
100
physiological saline
Fig. 7. The phase shift caused by frequency f3.
4 Discussion According to our experimental results, it can be seen: (1) From the curves in Fig3, 4 our new detection system can detect the phase shift less than 0.01° for frequency f1. The phase precision of our system is higher than the phase locking amplifier(SR844). (2) From Fig 3, 4 the greater the volume, the greater the phase shift generated is, under the condition of the same conductivity solution and the same frequency. The higher the signal frequency, the greater the phase shift generated is, under the condition of the same conductivity solution and the same volume. For frequency f1 the phase shift generated is less than 0.1° and for frequency f3 the phase shift generated is more than 0.5°. (3) From Fig 5, 6, 7 the phase shift generated by physiological saline is larger than cerebral edema with regard to the each frequecncy. (4) According to the formula (2) the phase shift is directly proportional to the conductivity, the volume and the frequency [6] and our experimental results are consistent with the formula (2).
θ ∝ kωσ
.
(2)
θ is the phase shift, k is the geometry coefficient, σ is the conductivity, ω is the angular frequency. From above discussions our new experimental results are consistent with the results of our previous tests, so the new system is proved to be feasible. There are two problems in our experimental study. One is that the excitation and detection coils are exposed to the surrounding environment and suffered from various EMI, such as human body influence and power interference and so on. The other is that the self-calibration needs too long time in our phase detector. In order to solve the above problems the physical model should be shield to eliminate EMI and the selfcalibration time of the phase detector should be reduced by optimizing our selfcalibration software.
Experimental Study on Simulated Cerebral Edema Detection with PSSMI
367
Acknowledgments. This work was supported by National Natural Science Foundation of China (61072254).
References 1. Hu, X., Qin, M., Liang, W., et al.: Study on techniques of phase detection in brain magnetic induction tomography. In: Proceedings of ICBBE 2007, pp. 782–785 (2007) 2. Zhou, W., Qin, M., Li, K., et al.: Study on detection phase deviation in brain MIT system. In: IFMBE Proceedings, vol. 25, pp. 328–330 (2009) 3. Liang, W., Qin, M., Jiao, M., et al.: Phase Detection Based on the Lock-in Amplifier SR844 and Experiments of Brain Neuron Cells in MIT System. In: Proceedings of ICBMEI 2008, pp. 638–642 (2008) 4. Li, K., Qin, M., et al.: The calculation and Measurement on phase shift in Single-channel BMIT system. Chinese Journal of Medical Physics 26(3), 1097–1101 (2009) 5. Jiao, M., Qin, M., Liang, W., et al.: Design and implementation of a new type excitation source and the optimal excitation coil for MIT. In: Proceedings of ICBBE, pp. 538–541 (2008) 6. Gfiffiths, H., Stewart, W.R., Cough, W.: Magnetic induction tomography: A measuring system for biological tissues. Ann. N Y Acad. Sci. 873, 335–345 (1999)
Medium Choice of Chinese Consumers in Obtaining Advertising Information about Minitype Automobile* Dao-ping Chen** School of Economics and Management, Chongqing Normal University, Chongqing 400047, P.R. China
[email protected] Abstract. This paper explores the media choice of Chinese consumers in obtaining advertising information about minitype automobile. A worldwide survey which involves a majority of Chinese areas is conducted. The survey focuses on consumers’ demographic characteristics, media choice and consumption things (first or repeated consumption). The result of the survey shows that consumers for obtaining information about minitype automobile chiefly choose friend or relative, newspaper, TV or automobile dealer as the media. The study finds that there isn’t a significant difference between first and repeated consumers in the media choice and that the demographic characteristics significantly affect the media choice of consumers. Consumers’ living city, occupation and education are the most important characteristics which affect the media choice of consumers. Keywords: medium choice, advertising information, minitype automobile, consumer.
1 Introduction A steady and rapid growth of China economy in longer time provides basic condition for development of Chinese automobile industry. China takes near 40 years, from 1953 to 1992, to make its automobile production volume reach 1 million units; However, China’s automobile production volume reaching 2 million units only spent 8 years, from 1992 to 2000, and then it only takes 2 years that China's automobile production volume reaches 3 million units in the end of 2002 (Zhang and Sun 2004). Weng (2004), vice-minister of Chinese Ministry of Communications, predicts that the total amount of civil automobile will reach 140 million or so in 2020. Facing such huge chance, multinational automotive companies, such as GM, Ford and Kreisler, invest one after another in China. In order to lure consumers to their brands, automakers place lots of advertisements in media. Statistics shows that advertising expenditure by automotive firms reach RMB 4.6 billion in 2003, doubling in size since 2002, with RMB 2 billion on TV and RMB 2.6 billion spent on print (Savage 2004). In this situation, automakers are confronted with an important *
Funded by Doctor Foundation of Chongqing Normal University (No. 11XWB004). ** Dao-ping Chen: PhD; Asso. Prof.; Research interests applied statistics. M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 369–377. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
370
D.-p. Chen
decision: media decisions, namely, what medium should be chosen to advertise about automobile? If a manufacturer advertises in a certain medium, but consumers don’t like to choose the medium to obtain the advertising information. The advertising information in the media can’t reach or can’t effectively reach these target consumers even though the manufacturer has taken a lot of expenditure for it. Accordingly, it is important to study the medium choice of consumers’ obtaining advertising information. This paper studies medium choice of Chinese minitype automobile consumers. In the remainder, the previous relevant literatures are briefly reviewed. Next, a questionnaire is designed and accounted for responses. Then, the results are present and interpreted. Finally, conclusion and implication are given.
2 Literature Review Advertisers and media planners generally think that media differentially impact the effectiveness of advertising and media context has an important influence on the value of advertising (Ducoffe 1995). Among conventional media, Magazine advertising is especially efficient because it targets consumers by demographics and lifestyle (LaReau 2005). Therefore, automobile companies are increasingly working with food and travel magazines to sponsor consumer promotions and special events (Bernstein 2004). A research (Marketing (UK) 1999) indicates that potential automobile buyers find television and newspaper the most valuable media when it comes to making a new purchase, while radio and posters are least helpful. Since Internet appeared, there have been a lot of researches about the web medium. It is reported that more automobile buyers turn to Web for information, especially in the initial stage of purchasing automobile (Milsom 2003). Yoon and Kim (2001) think that one of the most significant differences may be the interactivity of Internet advertisements. In fact, Internet, as one of all kinds of media, is only a tool that closes relationships with automotive customers even if it possesses powerful function (Washington1998). Thus, mass media advertising is still very important in strengthening the bond between automakers and consumers (Serafin 1994). Therefore, the consumers’ needs and preferences on the advertising medium are vital for automakers to be faced with medium decisions. A study (Pollay, Tse, and Wang 1990) suggests that Chinese consumers are more positive about advertising than consumers in the West. Another study (Zhou, Zhang, and Vertinsky 2002) has also gotten similar results that urban Chinese has similar or more positive attitudes toward advertising than their America counterparts. Chinese households now possess more TV, newspapers and computers than before, which offers a condition for advertisement reaching consumers. Moreover, television viewing is considered the most popular leisure time activity in China (Wei and Pan 1999). Chinese local firms, therefore, tend to place a special emphasis on television advertising (Li 2004). Meanwhile, newspapers in China have also grown in ways similar to television stations, although they tend to take on a stronger regional than national feature (Yao 2004). Like television and newspapers, magazines are changing rapidly in China (Zenith Media 2000/2001). After compared with television, newspapers, and magazines’ effectiveness to reach particular consumer segments in
Medium Choice of Chinese Consumers in Obtaining Advertising Information
371
China, Hung, Gu and Tse (2005) find that magazines have the highest targetability among the three most popular mass media. Another survey (Zenith Media 2000/2001) on the proportion of reaching target audience reported that there are the different reaching rates for different media in China: television reaches 89%, newspapers reaches 72%, and magazines reaches 13% of the population. It is not doubt that these research results are very helpful for automakers to make media decision on advertisements. But authors don’t more consider consumers’ preference to advertising media.
3 Questionnaire and Response Nine media are selected in this questionnaire. They are newspaper, magazine, TV, street poster or billboard, Internet, automobile dealer, automobile manufacturer, automobile exhibition and friend or relative. That friend or relative is included in the questionnaire is because compared with the West, China is considered a high context culture (Hall Hall 1990). A research (Prendergast et al. 2001) shows that personal relationships with friends and family are more important, and these personal relationships can be extended to affect business relationships in a high context culture. In questionnaire, gender, living city, occupation, education, age and monthly household income are used to describe consumer characteristics. Gender has two levels (1=male, 2=female), living city four levels (1=big city, 2=city of middling size, 3=county seat, 4=villages and towns), occupation seven levels (1=government servant, 2=employee of national enterprise, 3=employee of private enterprise, 4=employer of individual enterprise or partnership enterprise, 5=farmer, 6=professional (lawyer, accountant, teacher, doctor, athlete, reporter etc.), 7=other), education six levels (1=junior high school or below, 2=senior high school, 3=technical secondary school, 4=junior college, 5=college or university, 6=graduate student), age seven levels (1=18-21 years, 2=22-25, 3=26-29, 4=30-34, 5=35-39, 6=40-59, 7=60 or above 60 years), monthly household income four levels (1=less than RMB 2000, 2=RMB 2000 to less than RMB 5000, 3=RMB 5000 to less than RMB 8000, 4=more than RMB 8000). This survey is a large what is supported by an automotive group company. The area involved in the survey amounts to 29 provinces, municipalities directly under the central government or autonomous regions. The survey is conducted between December 2003 and March 2004 by means of 263 automobile dealers that sell minitype automobile. Each dealer is with responsibility for providing and regaining 10 questionnaires, thus the questionnaires add up to 2630. The object of the survey is the consumer selected randomly from consumers entering the shop of dealer and having intention to purchase minitype automobile. A total of 2623 questionnaires are returned. Out of 2623, 280 are discarded due to their incompleteness and the remaining 2343 questionnaires are used for the final analysis. The data are analyzed using SPSS 15.0.
&
372
D.-p. Chen
4 Results 4.1 Media Choice of Chinese Minitype Automobile Consumers The result of the media choice of consumers is shown in able 1. Of total 2343 cases, 1305 are from consumers of first purchasing automobile and 1038 are from consumers of repeatedly purchasing automobile. According to Table 1, the percentage of the media choice to consumers of first purchasing automobile is as follows: friend or relative (64.7%), newspaper (45.1%), automobile dealer (43.2%), TV (39.8%), automobile manufacturer (27.0%), automobile exhibition (26.6%), Internet (23.1%), magazine (13.6%), and street poster or billboard (5.2%). For the repeated consumers, the percentage of media choice breakdown is as follows: friend or relative (69.2%), automobile dealer (47.2%), newspaper (43.4%), TV (37.4%), automobile manufacturer (26.3%), automobile exhibition (23.2%), Internet (18.5%), magazine (11.3%), and street poster or billboard (6.6%). The percentage of media choice between first consumers and repeated consumers only has a small difference. Is the difference significant A MANOVA is conducted with two levels (1=repeatedly and 2=first) as the factor on the full set of 9 dependent variables, friend or relative, newspaper, magazine, TV, street poster or billboard, Internet, automobile dealer, automobile manufacturer and automobile exhibition (Wilks' Lambda=0.988 (Exact F=2.950, Sig.=.002 RN (output) be a continuous mapping, then hidden layers and weights (thresholds) values can be found in such way that φ is realised by this network with arbitrary precision [3]. If some network (its weights) exists to such arbitrary continuous mapping φ, it is possible to find x to approximate value y ≅ φ(x) by this algorithm: •
• • •
to the inputs of neurons in input layer values of all all elements of vector x are submitted. Outputs of neurons in input layer obtain values yi = xi, i = 1,…, m, and are transferred along the lines with weights, which are marked wji to the inputs of all k-neurons of hidden layer (j = 1, …, k) these values are computed by neurons from the hidden layer to its response and proceed them to inputs of the neurons in higher layer this activity is continuously spread until from the output layer neuron the response vector x of the entire network y is created.
3 New Solution This chapter is devoted to detailed description of the C# neural network library, which, as aforesaid, is possible to use for design of application that classifies outputs of spectral analysis of material diagnostics. Library AForge.Neuro.dll is based on Microsoft Visual Studio Solution called AForge, which contains elementary classes, methods, components and other elements that can be used not only for design of neural networks, but even for design genetic algorithms, fuzzy systems, image processing, robotics and many other practical applications. AForge.Neuro.dll namespace contains interface and classes for computing purposes of neural networks. Following description of all classes that are contained in the library helps for better insight of neural network architectures design. 3.1 Classes in AForge.Neuro Library Neuron class is elementary class of the whole library. Method Neuron has only one input parameter in its argument which is count of neuron’s inputs. This class also contains vector of all neuron inputs weights and one output variable. After neuron is created, all its input weight values are set by Randomize method. The range of these randomized values is from 0 to 1. Weight value are during following learning process changed, so that the results and its errors are minimalised. Layer class is basically similar to Neuron Class and summarizes common functions of all neural layers. In this class particular parameters of the whole layer are determined, that means count of neurons in layer, overall count of layer inputs. Resulting vector of layer outputs is generated by using method Compute.
452
O. Krejcar
In standard practice is Network class the most popular class. If some specific kind of neural network architecture ought to be implemented, it is necessary to include this class and extend it with required parameters of given architecture, e.g. Kohonen selforganizing maps that are capable to recognize colors. Network class contains declaration of overall number of inputs to neural network, number of layer in network and number of neurons in each layer. From this generally designed class inherits classes ActivationNetwork and DistanceNetwork, which can be used for particular design of applications.
Fig. 3. Sigmoid progression [7].
3.2 SigmoidFunction Class SigmoidFunction class which implements non-linear sigmoid with 0 to 1 range is much more common. Presicion of neural network learning depends on its steepness that is defined as λ = tg α. In this class default setting of λ value is 2. It is possible to change this value if the neural network is not trained in a good quality.
4 Conclusions In this project we provided support for the design and realisation of application developed in object orientated programming language C#, which will classify data obtained on basis of the material defects diagnostics. The application implements three layer perceptron architecture that was so far used pro classification purposes in commercial programme Statistica. After realization of this application it will be possible to substitute Statistica programme. Support for programming of the application is C# neural network library called AForge.Neuro.dll, which was introduced on codeproject.com server by programmer Andrew Kirillov in 2006. Author of this library reserves rights to its usage and it is forbidden to use it for the commercial purposes without his agreement. When compiling the theoretical part of this work, we used particularly expert materials focused on neural networks and we understood the basic principles of their
Use of Neural Networks Library for Material Defect Detection Diagnosis
453
design. As far as my opinion is concerned, we think that the future of neural network is very perspective, because very complicated and complex tasks can be solving by these algorithms. Acknowledgement. This work was supported in part by (1) „Centre for Applied Cybernetics“, Ministry of Education of the Czech Republic under project 1M0567, (2) „SMEW – Smart Environments at Workplaces“, Grant Agency of the Czech Republic, GACR P403/10/1310, (3) „SCADA system for control and monitoring of processes in Real Time“, Technology Agency of the Czech Republic, TACR, TA01010632 and (4) "User Adaptive Systems", VSB - Technical University of Ostrava under project SP/2011/22. We also acknowledge support from student Jakub Hlavica.
References 1. Jirsik, V., Horacek, P.: Neural networks, expert systems and speech recognition. Study support FEKT VUT, 7–46 2. Krejcar, O., Frischer, R.: Detection of the Internal Defects of Material on the Basis of the Performance Spectral Density Analysis. Journal of Vibro Engineering 12(4), 541–551 (2010) 3. Krejcar, O.: Problem Solving of Low Data Throughput on Mobile Devices by Artefacts Prebuffering. EURASIP Journal on Wireless Communications and Networking, Article ID 802523, 8 pages (2009) DOI 10.1155/2009/802523 4. Neural Networks in Statistica program (2010), http://www.statsoft.com/textbook/neural-networks/ (quoted 3/11/2010) 5. Neural Networks on C# - The Code Project 1(November 9,2006), http://www.codeproject.com/KB/recipes/aforge_neuro.aspx (quoted 25/10/2010) 6. Krejcar, O., Janckulik, D., Motalova, L.: Complex Biomedical System with Biotelemetric Monitoring of Life Functions. In: Proceedings of the IEEE Eurocon 2009, St. Petersburg, Russia, May 18-23, pp. 138–141 (2009) DOI 10.1109/EURCON.2009.5167618 7. Mikulecky, P.: Remarks on Ubiquitous Intelligent Supportive Spaces. In: 15th American Conference on Applied Mathematics/International Conference on Computational and Information Science, pp. 523–528. Univ. Houston, Houston (2009) 8. Krejcar, O., Frischer, R.: Non Destructive Defects Detection by Performance Spectral Density Analysis. Journal Sensors, MDPI Basel 11(3), 2334–2346 (2011) 9. Brida, P., Machaj, J., Duha, J.: A Novel Optimizing Algorithm for DV based Positioning Methods in ad hoc Networks. Elektronika Ir Elektrotechnika (Electronics and Electrical Engineering) 1(97), 33–38 (2010) 10. Augustynek, M., Penhaker, M., Korpas, D.: Controlling Peacemakers by Accelerometers. In: 2010 The 2nd International Conference on Telecom Technology and Applications, ICTTA 2010, Bali Island, Indonesia, March 19-21, vol. 2, pp. 161–163. IEEE Conference Publishing Services, NJ (2010) ISBN 978-0-7695-3982-9, DOI: 10.1109/ICCEA.2010.288
454
O. Krejcar
11. Pindor, J., Penhaker, M., Augustynek, M., Korpas, D.: Detection of ECG Significant Waves for Biventricular Pacing Treatment. In: 2010 The 2nd International Conference on Telecom Technology and Applications, ICTTA 2010, Bali Island, Indonesia, March 19-21, vol. 2, pp. 164–167. IEEE Conference Publishing Services, NJ (2010) ISBN 978-0-76953982-9, DOI: 10.1109/ICCEA.2010.186 12. Tucnik, P.: Optimization of Automated Trading System’s Interaction with Market Environment. In: 9th International Conference on Business Informatics Research. LNBIP, vol. 64, pp. 55–61. Univ. Rostock, Rostock (2010) 13. Labza, Z., Penhaker, M., Augustynek, M., Korpas, D.: Verification of Set Up DualChamber Pacemaker Electrical Parameters. In: 2010 The 2nd International Conference on Telecom Technology and Applications, ICTTA 2010, Bali Island, Indonesia, March 19-21, vol. 2, pp. 168–172. IEEE Conference Publishing Services, NJ (2010) ISBN 978-0-76953982-9., DOI: 10.1109/ICCEA.2010.187
Multi-sensor Measurement Fusion via Adaptive State Estimator Li-Wei Fong No. 168, Shiuefu Rd., Tanwen Village, Chaochiao Township,36143 Miaoli, Taiwan
[email protected] Abstract. An adaptive state estimator is developed to fuse the measurements extracted from multiple sensors for tracking the same maneuvering target. The proposed approach consists of Multi-Band Standard Kalman Filter (MBSKF) and a learning processor. Based on Bayesian estimation scheme, the likelihood function of learning processor can be approximated by the Gaussian basis function whose smoothing factor is related to the estimated bandwidth by taking an average of innovation error covariance matrices of MBSKF. Based upon learning processor and MBSKF, adaptive state estimation is extended to handle the switching plant in the multi-sensor environment. The simulation results are presented which demonstrate the effectiveness of the proposed approach. Keywords: Measurement fusion, adaptive state estimator, learning processor.
1 Introduction In the modern surveillance systems, multi-sensor data fusion algorithms have been applied to many fields such as air traffic control, tactical weapon defense, and C3I (Command, Control, Communication and Intelligence) systems where measurements extracted from multiple sensors are used to estimate the states (position, velocity, and acceleration etc.) of the moving objects [1]. Tracking targets by fusing measurements from Standard Kalman Filter (SKF) in a central level fusion process (a global estimation process) is acknowledged to be one of the most powerful tracking techniques [2]. It is known that the target dynamics may vary rapidly. Using only one single-level of process noise, the SKF can be resulted in performance degradation. In order to accommodate the possibly varying accuracy requirement, different levels of system process noise are required for the adaptation of state estimator. In the literature, a few adaptive approaches [3], [4], [5] introduced to overcome the filtering dilemma were to use a group of two parallel filters set up for two different models, namely, the system process noises of narrow bandwidth and wide bandwidth. In this paper, we extend the works of [6] and [4] by using Multi-Band SKF (MBSKF) and the learning processor to develop an Adaptive State Estimator (ASE) for central measurement fusion, as shown in Fig. 1.
M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 455–462. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
456
L.-W. Fong
Fig. 1. The adaptive state estimator in the measurement fusion center.
In particular, a semi-Markov process is incorporated into the learning processor this ensuing adaptive bandwidth capability, that is, the abilities to switch between low-level-band SKF in the absence of maneuvers and middle- or high-level-band SKF in the presence of maneuvers or very highly skilled maneuvers. Hence, the learning processor becomes an unsupervised Bayesian learning algorithm [7]. The resulting estimator makes natural for target tracking that provides significantly better tracking performance than each individual SKF.
2 Adaptive State Estimation In this section, the target model, measurement model, and the computational structure of the measurement fusion center are described. In the fusion center, the estimation architecture of ASE is divided into two parts, MBSKF and learning processor. The candidate central fused estimates are produced by the MBSKF. The learning processor is used to generate the model probabilities to classify which one of the candidate estimates for output. Meanwhile, Bayesian estimation scheme of ASE is developed and briefly outlined. 2.1 Target and Measurement Models Consider a dynamical target which is tracked by a multi-sensor system. Assume that the target dynamics can be modeled by one of M hypothesis models. The model set is
Multi-sensor Measurement Fusion via Adaptive State Estimator
457
denoted as ΨM := {1,2, …, M } . The event that model j is in effect during the sampling
period (k,k+1] is denoted by α kj+1 . For the j-th hypothesized model α kj+1 , the target dynamics and the N sensor measurements are modeled by the following discrete-time state-space model: xk +1 = Φ kj xk + Γkj wkj ,
j = 1,2, … , M
(1)
z ki +1 = H ki +1 xk +1 + vki +1 ,
i = 1,2,… , N
(2)
where xk is nx -dimensional state vector, Φ kj is the state transition matrix, Γkj is input matrix, z ki is the n z -dimensional measurement vector of the i-th sensor, and H ki is the observation mapping matrix. The wkj and vki vectors are assumed to be zeromean, white sequences with known covariance matrices. The covariance matrices for the wkj and vki vectors are given by cov[wkj , wlj ] = Qkjδ kl and cov[vki , vli ] = Rki δ kl
(3)
where the Kronecker delta δ kl = 0 for k ≠ l and δ kl = 1 for k = l . The two noises are also assumed to be uncorrelated as cov[wkj , vli ] = 0,
∀k , l .
(4)
By measurement fusion, the N sensor models can be integrated into following single model: z k +1 = H k +1 xk +1 + v k +1
(5)
where z k +1 = [( z1k +1 ) T , ( z k2+1 ) T , H k +1 = [( H k1+1 ) T , ( H k2+1 )T , v k +1 = [(vk1 +1 ) T , (vk2+1 ) T ,
, ( z kN+1 ) T ]T ,
(6)
, ( H kN+1 ) T ]T , , (vk3+1 ) T ]T ,
R k +1 = cov[ v k +1 , v k +1 ] = diag[ Rk1+1 , Rk2+1 ,
, RkN+1 ].
(7)
(8)
(9)
458
L.-W. Fong
2.2 Multi-Band Standard Kalman Filter
Given the system model described by (1) and (5), the MBSKF is based on a bank of M parallel standard Kalman filters set up for M different models. The equations of the j-th standard Kalman filter are given by [8] Pk +j 1|k = Φ kj Pk |jk (Φ kj ) T + Γkj Qkj (Γkj ) T
(10)
S kj+1 = H k +1 Pk +j 1|k (H k +1 ) T + R k +1
(11)
K kj+1 = Pk j+1|k (H k +1 ) T ( S kj+1 ) -1
(12)
Pk j+1|k +1 = Pk j+1|k − K kj+1 S kj+1 ( K kj+1 ) T
(13)
xˆkj+1|k = Φ kj xˆ kj|k
(14)
d kj+1 = z k +1 − H k +1 xˆ kj+1|k
(15)
xˆkj+1|k +1 = xˆ kj+1|k + K kj+1d kj+1
(16)
where xˆ kj+1|k and xˆkj+1|k +1 are predicted and filtered state vectors of the α kj+1 , Pk +j 1|k ( Pk +j 1|k +1 ) is the covariance matrix of the estimation errors before (after) processing the measurement, d kj+1 is the innovation , S kj+1 is the covariance of the innovation, K kj+1 is the filter gain matrix, Rkj+1 is the variance of the measurement noise, and Qkj+1 is the variance of the process noise. 2.3 Learning Processor Assume that the unknown target maneuver model randomly switches at random times between a finite set of M possible models. The rate of switching is assumed to be considerably slower than that of the observation sampling rate, and the random switching process among the models will be determined by a semi-Markov process. The model probabilities are calculated by the learning processor. The likelihood function of learning processor can be approximated by the Gaussian basis function whose smoothing factor is related to the estimated bandwidth by taking an average of innovation error covariance matrices of MBSKF. Based on the Bay's estimation procedure, a weighted state estimate can be written as follows: xˆk +1|k +1 =
M
∑μ j =1
j k +1
xˆ kj+1|k +1 .
(17)
Multi-sensor Measurement Fusion via Adaptive State Estimator
459
The probability μ kj+1 of the α kj+1 is given by Bayes’ rule as below M
μ kj+1 =
f (α kj+1 | d kj+1 , S k +1 )∑ θ jm μ km m =1 M
M
∑ f (α
j k +1
j =1
(18)
| d , S k +1 )∑ θ jm μ km j k +1
m =1
where θ jm is the transition probability. The likelihood function of the α kj+1 is given by the innovation d kj+1 and the averaged covariance of the innovation S k +1 as f (α kj+1 | d kj+1 , S k +1 ) =
1 (2π )
nz / 2
S k +1
1/ 2
⎞ ⎛ −1 exp⎜ ( d kj+1 ) T ( S k +1 ) −1 d kj+1 ⎟ 2 ⎠ ⎝
(19)
where S k +1 =
1 M
M
∑S
j k +1
.
(20)
j =1
3 Simulation Results The results of computer simulation are presented for the performance comparison of ASE, low-level-band SKF (denoted as SKF1; maneuver variance given by 133m2/s4), middle-level-band SKF (denoted as SKF2; maneuver variance given by 833m2/s4) and high-level-band SKF (denoted as SKF3; maneuver variance given by 2133m2/s4). The results of 100 runs of Monte Carlo simulation are processed and presented in the Figures of following case study. The descriptions of scenario and system parameters are similar to but a little different in [3], [4], [5]; the interesting readers may consult these papers for realizing the method of simulation processing. The performance is evaluated by using the Root Mean Square Error (RMSE). Averaged Root Mean Square Error (ARMSE) is defined as the performance index. There are two maneuvering turns of the target trajectory called circular-turn and U-turn shown in Fig. 2. The errors of the position estimate are plotted in Fig. 3. Similar RMS-type quantities are plotted in the other figures. Figs. 3-5 show the estimation errors of ASE, SKF1, SKF2 and SKF3, respectively. For a quantitative performance comparison, the timeaverages of estimation errors in Figs. 3-5 are listed in Table 1. As shown, ASE demonstrates adaptive capability in the tracking process. ASE performs better tracking performance than SKF1, SKF2 and SKF3. Table 1. Time-Averages of Estimation Errors. Method SKF1 SKF2 SKF3 ASE
Position (m) 21.4431 19.4546 19.9763 18.5923
Velocity (m/s) 31.0124 29.2598 32.4038 25.5496
Acceleration (m/s2) 22.0567 22.7735 26.8376 19.1526
L.-W. Fong
12000 Target 10000
z-axis (m )
8000
6000
4000
2000
0
2000
4000
6000
8000 10000 x-axis(m)
12000
14000
16000
Fig. 2. The target trajectory with both circular- and U-maneuvering turns.
POSITION 40 SKF1 SKF2 SKF3 ASE
35
RMS ERROR (m)
460
30
25
20
15
10
0
10
20
30
40
50 60 Time(sec.)
70
80
Fig. 3. RMS errors in position estimate.
90
100
Multi-sensor Measurement Fusion via Adaptive State Estimator
VELOCITY 90 SKF1 SKF2 SKF3 ASE
80
RMS ERROR (m/s)
70 60 50 40 30 20 10
0
10
20
30
40
50 60 Time(sec.)
70
80
90
100
Fig. 4. RMS errors in velocity estimate.
ACCELERATION 110 SKF1 SKF2 SKF3 ASE
100 90
RMS ERROR (m/s 2)
80 70 60 50 40 30 20 10 0
0
10
20
30
40
50 60 Time(sec.)
70
80
Fig. 5. RMS errors in acceleration estimate.
90
100
461
462
L.-W. Fong
4 Conclusion In this paper, an adaptive state estimator for use of the centralized measurement fusion is introduced to improve tracking accuracy of the multi-senor surveillance systems. With a semi-Markov process, learning processor is developed as an on-line learning algorithm to choose an appropriate target model for the standard Kalman filter performing sub-optimal state estimation. Simulation results are provided for comparison of tracking accuracy between proposed estimator and each individual standard Kalman filter. For target trajectory with both circular- and U-maneuvering turns, the proposed estimator demonstrates better tracking accuracy than the averaged estimates of multi-band standard Kalman filter with about 8.37%, 17.29%, and 19.83% improved in position, velocity, and acceleration, respectively. The results indicate that proposed estimator provides significant performance improvements.
References 1. Hall, D.L., Llinas, J.: An Introduction to Multisensor Data Fusion. Proc. of the IEEE 85(1), 6–23 (1997) 2. Gan, Q., Harris, C.J.: Comparison of Two Measurement Fusion Methods for Kalman-filterbased Multisensor Data Fusion. IEEE Trans. Aerosp. Electron. Syst. 37(1), 273–280 (2001) 3. Fong, L.W.: An Adaptive Filter for Multi-sensor Track Fusion. In: Proceedings of International Conference on Signal Processing, pp. 231–235 (2008) 4. Fong, L.W.: Multi-sensor Track Fusion via Multiple-model Adaptive Filter. In: Proceeding of the 48th IEEE Conference on Decision and Control and 28th Chinese Control Conference, pp. 2327–2332 (2009) 5. Fong, L.W.: Adaptive Information Matrix Filtering Fusion with Nonlinear Classifier. In: Proceedings of SICE Annual Conference, pp. 2214–2219 (2010) 6. Moose, R.L., Wang, P.P.: An Adaptive Estimator with Learning for a Plant Containing Semi-Markov Switching Parameters. IEEE Trans. Syst. Man Cybern. 3(3), 277–281 (1973) 7. Duda, R.O., Hart, P.E., Stork, D.G.: Patter Classification, 2nd edn. Wiley, New York (2001) 8. Bar-Shalom, Y., Li, X.R.: Estimation and Tracking, Principles, Techniques, and Software. Artech House, Norwood (1993)
A Simple Automatic Outlier Regions Detection Kitti Koonsanit Department of Computer Science, Faculty of Science, Kasetsart University, Bangkok, Thailand
[email protected] Abstract. Automatic determination of the outlier regions is often needed to eliminate that outlier region. In this paper, a method has been developed to determine the outlier regions in satellite image using a data mining algorithm based on the co-occurrence matrix technique in order to determinate that outlier. Our method consists of four stages, the first stage estimate a number of region by co-occurrence matrix, the second stage cluster dataset by automatic clustering algorithm, the third stage detect outlier regions by automatic threshold value and the final stage defines outlier regions, which are lower than threshold value, to be outlier regions. The proposed method was tested using data from unknown number of regions with multispectral satellite image in Thailand. The results from the tests confirm the effectiveness of the proposed method in finding the outlier regions. Keywords: Outlier regions, Anomaly Detection, determination outlier regions, co-occurrence statistics, Outlier regions detection.
1 Introduction Clustering is a popular tool for data mining and exploratory data analysis. One of the major problems in cluster analysis is the determination of the outlier clusters for unlabeled data, which should be eliminated. Outlier is the data which has obviously difference with clustering. In 1980, Hawkins made the definition of it: an outlier is an observation that deviates so much from other observations as to arouse suspicion that it was generated by a different mechanism [1]. Usually, this kind of data has special behavior or model. In effective data set, outlier is a small part and recognized as the byproduct of clustering [2]. So, outlier is always canceled or neglected simply. However, certain outlier probably is the real reflection of normal data. These data are worthy to be study more. In this paper, we propose a new easy method for automatically estimating the outlier regions in unlabeled data set. Pixel clustering technique in a color image is a process of unsupervised classification of hundreds thousands or millions pixels on the basis of their colors. In this paper, a method has been developed to determine the outlier regions in satellite image clustering application using a data mining algorithm based on the co-occurrence matrix technique. Therefore, automatic determination of the outlier regions can greatly help with the unsupervised classification of satellite Image. M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 463–470. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
464
K. Koonsanit
2 Related Work 2.1 Background Mutispectral imaging is characterized by its ability to record detailed information about the spectral distribution of the received light. Mutispectral imaging sensors typically measure the energy of the received light in tens or hundreds of narrow spectral bands in each spatial position in the image, so that each pixel in a mutispectral image can be represented as a high-dimensional vector containing the sampled spectrum. Since different substances exhibit different spectral signatures, mutispectral imaging is a wellsuited technology for numerous remote sensing applications including target detection. When no information about the spectral signature of the desired targets is available, a popular approach for target detection is to look for objects that deviate from the typical spectral characteristics in the image. This approach is commonly referred to as anomaly detection [3], and is related to what is often called outlier detection in statistics. If targets are small compared to the image size, the spectral characteristics in the image are dominated by the background. An important step in outlier detection is often to compute a metric for correspondence with the background, which then can be thresholded to detect objects that are unlikely to be background objects. Two approaches are of particular interest. One was developed by Reed and Yu [4][5][6] and is referred to as the RX detector (RXD), which has shown success in outlier detection for multispectral and hyperspectral images [7][8]. Another was proposed in [9][10] and is referred to as low probability detection (LPD), which was designed to detect targets with low probabilities in an image. The benchmark of anomaly detection is RX algorithm which is derived from the Generalized Likelihood Ratio Test (GLRT) with the assumption of Gaussian background [11]. However, background may be consisted of different ground cover types in real remote sensing images, such as water body, grass land, trees. This will lead to miss detection in complex background. Many researches attempt to use the Gaussian mixture model [12][13]. In reference [12], Ashton employs K-means cluster clustering the image into a number of statistical clusters and models each cluster with the Gaussian distribution. Subsequently, Carlotto proposes a similar approach using vector quantization to reduce the computational time [13]. However, the K-means based algorithm only considers the spectrum information of background. This will lead to miss clustering background pixels during the cluster process. Besides, all the methods, these methods are unsuitable for our application that needs to implement software fast and to ease the difficulty of implement software for beginners. In this paper, propose an outlier region detection algorithm. The new method is compatible with the k-means algorithm and it overcomes the limitation of having to indicate the outlier regions by co-occurrence matrix and automatic threshold define which is a apply technique in this proposed paper.
3 The Proposed Algorithm 3.1 Our Algorithm In this paper, we propose a new method for determination of the outlier regions, which is based on co-occurrence matrix scheme. While a traditional co-occurrence
A Simple Automatic Outlier Regions Detection
465
matrix specifies only the transition within an image on horizontal and vertical directions. The proposed method can be used to automatically select a k range in multispectral satellite image as shown in Table. 1. Table 1. Algorithm finding outlier region Algorithm Simple Outlier region detection Input: A spectral image Output: Outlier regions Processing 1. Estimation: Estimate a number of region by co-occurrence matrix [14][15][16][17] 2. Clustering: Cluster image by a clustering algorithm and input a number of region from step 1. 3. Threshold: Detect outlier regions by automatic threshold or input from user 4. Detection : Define regions which are lower than thresholding value as outlier regions
The proposed technique consists of four main steps: estimation, clustering, threshold and detection. 3.2 Estimate Number of Region The proposed technique first, the co-occurrence matrix scheme is employed to automatically segment out the object region in an image. Then, the local maximum technique is used to count a number of regions, which a number of region. Our definition of a co-occurrence matrix [15][16][17][14] is based on the idea that the neighboring pixels should affect region of clusters. Hence, we define a definition for a co-occurrence matrix by including the transition of the gray-scale value between the current pixel and adjacent pixel into our co-occurrence matrix illustrated in figure 1 and figure 2.
Fig. 1. right and bottom of pixel in a cooccurrence matrix
Fig. 2. Creating a Co-Occurrence Matrix
Let F be the set of images. Each image has dimension of P×Q. Let tij be an element in a co-occurrence matrix depending upon the ways in which the gray level i follows gray level j
466
K. Koonsanit
P
tij = ∑ x =1
Q
∑δ y =1
( Fk ( x, y ) = i ) and ( Fk ( x, y + 1) = j )
or
( Fk ( x, y ) = i ) and ( Fk ( x + 1, y ) = j )
Where δ = 1 ,δ = 0
otherwise .
(1) th
where Fk denotes the k band in the image set, F If s,0 ≤ s ≤ L − 1 is a threshold. Then s can partition the co-occurrence matrix into 4 quadrants, namely A, B, C, and D shown in Figure.3. Clusters
Fig. 3. An example of blocking of co-occurrence matrix
Since two of the quadrants shown in Figure.3, B and D, contain information about edges and noise alone, they are ignored in the calculation. Because the quadrants, which contain the object and the background, A and C, are considered to be independent distributions. The idea of proposed method is to select the results of co-occurrence matrix into a diagonal matrix. After threshold processing, the result of diagonal matrix was shown in figure 5. Diagonal matrix is used to show some clustered pixels. The gray level corresponding to local maximum which give the optimal number for object- classification in image as shown in figure 6. 3.3 Clustering After K estimate processing, we got a number of region, we cluster data by clustering algorithm such as k-means which is the most popular clustering techniques. 3.4 Automatic Thresholding After cluster processing, thresholding techniques was assigned an outlier score to each instance in the test data depending on the degree to which that instance is considered an outlier. In this paper, we apply a method for automatic thresholding, which is based on standard deviation. Standard deviation is a statistical evaluate of spread or variability. First, we have to find the mean for standard deviation. Mean is represented by the division of sum of all values and the total number of values. The standard deviation is the root mean squares deviation of the values from their arithmetic mean. It is calculated by take the square root of the variances and is symbolized by s.d, or s. as shown in (2).
A Simple Automatic Outlier Regions Detection
∑ (x − Mean )
467
2
σ= where
(2)
n −1
Ʃ = sum of Mean = Mean of all point
x = individual point n = sample size (Number of point)
The proposed method can be used to select Mean − σ is an appropriate automatically threshold values in order to estimate a cut-off threshold value to select the outlier region as shown in Figure.6. Thus the output of our techniques use a cut-off threshold to select the outlier regions which less than Mean − σ value. 3.5 Automatic Outlier Region Detection Finally, after threshold processing, we get outlier region which is an outlier region as shown in fig 4- fig 7.
Fig. 4. an example of original image
Fig. 5. an example of co-occurrence matrix of image from Figure 4.
Figure 5. illustrates outlier in a simple 2-dimensional data set. The data has 4 normal regions, A, B, C and D, since most observations lie in this one region. Points that are different from the regions, e.g., “A” points in red region, are outlier region.
Fig. 6. Example histogram from diagonal of cooccurrence matrix (with threshold value = 30)
Fig.7. X Region = a result of outlier detection
468
K. Koonsanit
4 Experiment and Results 4.1 Dataset We used the eight sets of raw data from different CCD Multi-spectrum images [18]. The dataset are obtained from small multi mission satellite project (SMMS), a department of Electrical Engineering, Kasetsart University. We would like to analyze data, which was registered in Thailand and thus try to determinate of the outlier region in interesting areas of Thailand. 4.2 Unsupervised Classification Method The experiments performed in this paper use the simple K-mean from the Weka software package. The simple K-Mean is the common unsupervised classification method used with remote sensing data. 4.3 Experimental Result Our experiment was tested with CCD Multi-spectrum images and shown in Table.2. The experiments demonstrate the robustness and effectiveness of the proposed algorithm. Table 2. Algorithm finding outlier region Original image
Outlier Region Detection
Outlier
Outlier
From the experimental result, it was found that outlier region solved by cooccurrence statistics techniques gives the nearest outlier regions with ground truth. It can be noticed that the oulier region in clustering between original images and solving by co-occurrence statistics techniques are very closed.
A Simple Automatic Outlier Regions Detection
469
5 Conclusions In this paper, a method has been developed to determine the outlier regions in satellite image using a data mining algorithm based on the co-occurrence matrix technique in order to determinate that outlier. Our method consists of four stages, the first stage estimate a number of region by co-occurrence matrix, the second stage cluster dataset by automatic clustering algorithm, the third stage detect outlier regions by automatic thresholding and the final stage defines regions, which are lower than threshold value, to be outlier regions. The proposed method was tested using data from unknown number of regions with multispectral satellite image in Thailand. The results from the tests confirm the effectiveness of the proposed method in finding the outlier regions.
Acknowledgment This work was supported by budget for overseas academic conference from the graduate school, Kasetsart University. The authors would like to thank TGIST.
References 1. Hawkins, D.: Identification of Outliers. Chapman and Hall, London (1980) 2. Liu, J.: Study and implementation of clustering and outlier detection algorithm : (Master thesis).LIAO NING: Shenyang Institute of Computing Technology Chinese Academy of Sciences (2006) 3. Stein, D.W., Beaven, S.G., Hoff, L.E., Winter, E.M., Schaum, A.P., Stocker, A.D.: Anomaly detection from hyperspectral imagery. IEEE Signal Process. Mag. 19, 58–69 (2002) 4. Reed, I.S., Yu, X.: Adaptive multiple-band CFAR detection of an optical pattern with unknown spectral distribution. IEEE Trans. Acoust., Speech, Signal Processing 38, 1760–1770 (1990) 5. Yu, X., Reed, I.S., Stocker, A.D.: Comparative performance analysis of adaptive multispectral detectors. IEEE Trans. Signal Processing 41, 2639–2656 (1993) 6. Yu, X., Hoff, L.E., Reed, I.S., Chen, A.M., Stotts, L.B.: Automatic target detection and recognition in multispectral imagery: A unified ML detection and estimation approach. IEEE Trans. Image Processing 6, 143–156 (1997) 7. Ashton, E.A., Schaum, A.: Algorithms for the detection of sub-pixel targets in multispectral imagery. Photogram. Eng. Remote Sens., 723–731 (July 1998) 8. Stellman, C.M., Hazel, G.G., Bucholtz, F., Michalowicz, J.V., Stocker, A., Scaaf, W.: Real-time hyperspectral detection and cuing. Opt. Eng. 39, 1928–1935 (2000) 9. Harsanyi, J.C.: Detection and classification of subpixel spectral signatures in hyperspectral image sequences. Ph.D. dissertation, Dept. Elect. Eng., Univ. Maryland-Baltimore County, Baltimore, MD (1993) 10. Harsanyi, J.C., Farrand, W., Chang, C.-I.: Detection of subpixel spectral signatures in hyperspectral image sequences. Proc. Amer. Soc. Photogram. Remote Sens., 236–247 (1994) 11. Reed, R., Yu, X.: Adaptive multi-band CFAR detection of an optical pattern with unknown spectral distribution. IEEE Trans. Acoust., Speech, Signal Process. 38, 293–305 (1990) 12. Ashton, E.A.: Detection of Subpixel Anomalies in Multispectral Infrared Imagery Using an Adaptive Bayesian Classifier. IEEE Trans. Geosci. Remote Sensing 36, 506–517 (1998)
470
K. Koonsanit
13. Carlotto, M.: A cluster-based approach for detecting man-made objects and changes in imagery. IEEE Trans. Geosci. Remote Sensing 43, 374–387 (2005) 14. Koonsanit, K., Jaruskulchai, C.: Automatic Determination of the Initialization Number of Clusters in K-means Clustering Application by Using Co-occurrence Statistics Techniques for Multispectral Satellite Image. In: The 2010 International Conference on Information Security and Artificial Intelligence (ISAI 2010), Chengdu, China, December 17-20 (2010) 15. Pal, N.R., Pal, S.K.: Entropic thresholding. Signal processing 16, 97–108 (1989) 16. Chanwimaluang, T., Fan, G.: An efficient algorithm for extraction of anatomical structures in retinal images. In: ICIP 2003 Proceedings, September 4-17 (2003) 17. Koonsanit, K., et al.: Metal Artifact Removal on dental CT Scanned Image by Using Multi-layer Entropic Thresholding and Label Filtering Technique for 3-D Visualization of CT images. In: Proc. of International Conference on Biomedical Engineering: ICBME 2008.IFMBE Proceedings, 13th International Conference on Biomedical Engineering, Singapore (December 2008) 18. Small Multi-Mission Satellite (SMMS) from the World Wide Web: http://smms.ee.ku.ac.th/index.php (data Retrieved: May 26, 2010)
The Feature Parameters Algorithm of Digital Signal in Circuit Based on Simulation Xiaodong Ma, Guangyan Zhao, and Yufeng Sun School of Reliability and System Engineering, Beihang University, 100191 Beijing, China
[email protected],
[email protected],
[email protected] Abstract. Aiming at fault judgment issue of digital signal in circuit fault simulation, a fault judgment method of time domain feature parameters based on Cadence simulation was made deep research in this paper. Firstly, the commonly used feature parameters of digital signal were analyzed. Secondly, as per data structure feature of Cadence software, the digital signal was divided into two kinds, one was expressed by analog form, and the other was expressed by numeric form. Then the specific algorithm of relative parameters of these two kinds of digital signal in Cadence was provided respectively. By the end, a representative case was made simulation analysis to validate the correctness of the algorithm. The result also indicates that the method is practical in reliability design of circuit. Keywords: feature parameter, digital signal, algorithm, fault simulation.
1 Introduction Circuit fault simulation technology based on EDA is an effective way to achieve integrated circuit design and analysis. By using of this technology, we can realize batch injection and simulation for component failure, then got large numbers of fault response. On basis of judgment made on such response signal, we can achieve circuit reliability analysis then provide foundation for improving the circuit design. It is very simple to make manual judgment on output signals of fault simulation; however, automatic judgment is necessary facing hundreds of thousand of simulation results. Signal analyzing and processing, as well as fault criterion technology are mainly applied in fault judge of circuit signal. At present, there are 3 main methods for signal analysis and processing, they are times domain method, frequency domain and time-frequency analysis method [1]. Frequency domain analysis takes frequency of input signal for variables, and studies the relationship between system structural parameters and performance in frequency domain. It reveals the inherent frequency features of signal and the relationship between time features of signal and its frequency ones. Frequency domain analysis is widely used in communications, automatic control and other fields [2,3]. However, for the purposes of judging the fault of circuit signal, just knowing its frequency feature is not enough. Thus, frequency domain method is not the focus of this paper. Time-frequency analysis triggerred from 1940’s is a new method which can clearly M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 471–478. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
472
X. Ma, G. Zhao, and Y. Sun
indicate the evolution relations of spectrum over time, and is mainly used in nonsteady signals or time-varying signal analysis and processing, it is used extensively in speech signal analysis and processing[4,5]. Although it is already applied in system fault judge and diagnosis[6,7], it is not easy to achieve due to complicated principle and large amount of calculation. In contrast, time domain analysis is a direct and more accurate analysis method. Time-domain feature parameters can reveal system’s instantaneous state and steady state performance, this method is widely used in automatic control, system fault diagnosis and other fields [8,9]. In view of the contents of this paper is fault judge of circuit signal in Cadence-based environment, time domain signal feature parameters can effectively reflect variation of amplitude over time. We can extract a series of feature parameters from the given signal and analyze whether these parameters meet requirements then realize signal fault judgment. Thus, this paper mainly studies the algorithm of time-domain feature parameters based on Cadence.
2 Feature Parameters of Circuit Signal Circuit signal typically contains analog signal and digital signal. Different types of signals have different feature parameters. Analog signal’s feature parameters and its algorithms have been given in literature [10] and this paper does not mention it again. Digital signal can be furtherly divided into single-pulse signal and multi-pulse signal, and its commonly used feature parameters are, rise/fall time, slew rate, period, duty cycle, pulse width, delay time, over/under overshoot, time interval of two positive /negative pulses, time interval between two positive/negative jump, and so on. Some feature parameters’ definition and algorithm are given as follows. (1) Delay Time It refers to time deflection of the measured signal against the reference one under the selected high and low level. Here it does not require that the measured signal keep same shape as the reference one, but they must be in the same graph region. Specifically: under selected high and low level, it’s the difference of jump points between measured signal and reference one in the time coordinates. (2) Slew Rate Also known as conversion rate, it refers to variable range of output voltage within unit time, expressed as:
SR=
V(H)-V(L) t (R)
(1)
Where, SR refers to slew rate, V (H) means high level potential, V (L) means low level potential, and t (R) is the rise time of signal. In Fig.1, V (H) = 3.7566v, V (L) =0.011494v, the got t (R) = 540.17us (see Fig.1) as per rise time calculation method, slew rate can be obtained by formula (1), SR = 5571.8 (v / s). (3) Period Period is the time interval between two adjacent rising edge (or falling edge). We take the median point of rising (or falling) edge in calculation. The determination of the rising / falling edge should be noted. As shown in Fig.2. The period is the time interval between the left point and the right one.
The Feature Parameters Algorithm of Digital Signal in Circuit Based on Simulation
Fig. 1. Definition of Slew Rate
473
Fig. 2. Definition of Period
(4) Duty Cycle It is the ratio of duration of high level to total period.
DC=
t (1) T
(2)
Where, DC refers to duty cycle, t (1) is the duration of high level, T is the period of the signal. (5) Over / Under Overshoot Over overshoot refers to the difference between the maximum of a signal and the high level within a signal cycle. Under overshoot refers to the difference between the minimum of a signal and the low level within a signal cycle.
V(Over)=V(max)-V(H)
(3)
Where, V (Over) refers to the over overshoot, V (max) refers to the maximum voltage of a signal, V (H) is the voltage of high level.
V(Under)=V(L)-V(min)
(4)
Where, V (Under) refers to the under overshoot, V (min) refers to the minimum voltage of a signal, V (L) is the voltage of low level.
3 Data Structure of Circuit Signal in Cadence Cadence is a kind of common used EDA simulation software, which can simulate digital/analog mixed circuit. Its output signal types include analog signal and digital signal. All signal information during the whole simulation period is recorded in simulation output file .dat. Data structure of the output file of time domain in Cadence is shown in Table 1. As shown in Table 1, analog signal is described as the amplitude of the signal at different time, more data points are recorded in region with higher variation rate of amplitude, and vice versa, that is, analog signal is recorded in the way of non-equal time spacing in Cadence. Digital signal is described as 0,1, R, F, X and Z, where 0 represents the low level, 1 represents the high level, R on behalf of the rising edge, F on behalf of the falling edge, X on behalf of uncertain state, Z represents highimpedance state, and record signal state after jump when digital signal is jumping.
474
X. Ma, G. Zhao, and Y. Sun Table 1. Data structure of circuit signal in Cadence
Signal Type
Analog signal
Digital signal
Data Structure #H <settings*> #N [‘] or #C [‘] or #C … [circle descriptions for analysis once again] #H <settings*> #N ‘ #C ‘ #C … [circle descriptions for analysis once again]
Notice
“**”—more than one, “”—Required, “[]”— Options, “[…] or […] or […]”— — at least one.
“**”—— more than one, “”—— Required, “[]”—— Options.
4 Cadence-Based Algorithm for Digital Signal Parameters In view of the feature of digital/analog mixed simulation in Cadence, an invisible DtoA interface exists between digital device and analog one where they are connected, it switches original digital signal to analog one then connect it with the analog device. Therefore, when these signals are displayed or judged, user sees the analog signals while they concern about its digital properties. For such signals, we treat it differently from digital signal or analog signal, and call it as digital signal expressed by analog form. For the digital signal expressed by 0, 1 etc discrete variables, we call it as digital signal expressed by digital form. Algorithm for feature parameters of these two types of digital signals is introduced as follows. 4.1 Algorithm of Digital Signal Expressed by Digital Form For convenience, the meanings of symbols are agreed as follows:
—— - —— + —— —— —— - ——
S(i) The i-th state of a specified state sequence. S i (y ) The state before the i-th state “y” in a specified state sequence. S i (y ) The state behind the i-th state “y” in a specified state sequence. n(y) The number of state "y" in a specified state sequence. ti(y) The time of the i-th state “y” in a specified state sequence. ti(y ) The time of the state before the i-th state “y” in a state sequence.
The Feature Parameters Algorithm of Digital Signal in Circuit Based on Simulation
475
+)—— The time of the state behind the i-th state “y” in a state sequence. —— The total number of state in a specified state sequence.
ti(y N
From Part 3 of this paper, we know that state of digital signal after jump is recorded with 0, 1 and other discrete variables, it is lack of description of the features of signal amplitude, and therefore we can only get part of the feature parameters. The specific algorithms are given in Table 2. Table 2. Algorithm of feature parameters of digital signal expressed by digital form Feature parameters
Algorithm
Note
n
Pulse Width
Delay Time
∑ [t (1+) − t (1)] i
PW =
i
1
n
DT = M ax[ t b ( y ) − ti ( y )] n [ S (1)]− 1
Period
∑ {[t
T =
n −1
Duty Cycle
∑
DC =
[ S (1)] − ti [ S (1)]}
1
n[ S (1)] − 1 [ t i (1+ ) − t i (1)] ti + 1 [ S (1)] − t i [ S (1)] n −1
n −1
Time Interval between t= Two Positive Jumps
i +1
1
∑ [t
Pulse numbers after n= one pulse is triggered
i +1
(0 + ) − ti (0 + )]
1
{
tb(y) is the jump time of reference signal at each state; ti(y) is the jump time of measured signal at each state. For periodic signal, it is the average of periods. It has no sense for non-periodic signal. where , n
n(0), the first edge after t is rise edge n(1), the first edge after t is fall edge
= n[ S (1)]
where , n
n −1
{
n (1)−1, S ( N ) =1 others
where , n = n (1),
−1, S ( N ) =0 = { nn(0) (0), others
t is action end time of excitation signal——t2[S(1)]. n(0)/ n(1) is the number of state 0/state 1.
4.2 Algorithm of Digital Signal Expressed by Analog Form Part of feature parameters of digital signal expressed by analog form can be converted to digital signal expressed by digital form via certain algorithm, and then use the algorithm listed in Table 2 to calculate feature parameters. However, some parameters cannot be converted; it should be calculated directly by means of the output data of simulation. Digital signal expressed by digital form is used to record the state of signal after jump, thus we just need consider the two data points near high / low level threshold. Because of data points with amplitude equal to high or low level threshold may not exist, this paper takes time points as state transition ones, time points are got by means of linear interpolation algorithm. Low level threshold is set as UL, high threshold is set as UH. Two adjacent state points of digital signals expressed by analog quantity are S1 (t1, u1) and S2 (t2, u2). Transformation rules are given in Table 3.
476
X. Ma, G. Zhao, and Y. Sun Table 3. Transformation Rules
conditions
< ≦ u <U ≦ <U ≦u < <U ≦ u
u 1 UL u 1 UL UL u1
2
H
H
2
H
2
State
Time
conditions
R R, 1 1
t(R) t(R), t(1) t(1)
UL u2 u 2 UL u 2 UL
< ≦ U <u ≦ <U ≦ u ≦ <u <U H
1
H
1
1
H
State
Time
F F, 0 0
t(F) t(F), t(0) t(0)
Where, t(R) and t (0) can be obtained from:
、 t (0) = ut -−tu 1
t (R )
2
1
× (U L -u1 )+ t1
(5)
× (U H -u1 )+ t1
(6)
2
t (F) and t(1) can be obtained from:
、 t (1) = ut -−tu
t (F)
1
1
2
2
Over / under overshoot of digital signal expressed by analog form can be calculated by formula (7) and formula (8). V (O ver)=M ax ( u i )-V (H )
(7)
V (U nder)= V (L)-M in ( u i )
(8)
Where, UL refers to low level threshold, UH refers to high level threshold, ui is the amplitude of digital signal expressed by analog form.
5 Case Study Figure 3 is a typical digital/analog mixed circuit. We apply the arithmetic of feature parameters described in this paper to determine circuit fault. In which, OUT is digital signal expressed by digital form, and U3B:Y is digital signal expressed by analog form. Normal simulation and fault one is made for circuit above by Cadence respectively, we observe the waveform of OUT and U3B:Y. The results of normal simulation and fault simulation are shown in Fig.4 and Fig.5 respectively.
Fig. 3. Circuit Schematic Diagram
The Feature Parameters Algorithm of Digital Signal in Circuit Based on Simulation
477
Fig. 4. Waveform of normal simulation
(a) C1- Parameter drifts
(b) C1-Short-circuited
Fig. 5. Waveform of fault simulation
U3B:Y is a clock signal, and period and duty cycle are two important feature parameters for clock signal. Therefore, we choose period and duty cycle as judgment parameters. OUT is a square signal, we also choose period and duty cycle as judgment parameters. Feature parameters for two observation points in fault simulation output calculated as per algorithm of digital signal time domain feature parameters mentioned in this paper, together with fault judgment result is shown in Table 4. Table 4. Value of feature parameters and the results Signal Name
OUT
U3B˖Y
Feature parameters
period
DC
period
DC
Minimum permitted
2.3us
0.4
1.2 us
0.4
5us
0.6
2.5 us
0.6
Maximum permitted C1- Parameter drifts
C1-Short-circuited
Value Parameter Signal Value Parameter Signal
1.2588us 0.5122 Fault Normal Fault -1 -1 Fault Fault Fault
0.6293 us 0.5232 Fault Normal Fault -1 -1 Fault Fault Fault
478
X. Ma, G. Zhao, and Y. Sun
The values in the table are calculated by arithmetic given by this paper. By comparison, the results keep same as simulation graph, which indicates that the algorithm is correct. Since there is no amplitude variation after waveform achieves stable with C1 shortcut, it cannot calculate relevant feature parameter, which is expressed as -1, which also proves circuit fault judgment could be achieved by use of this algorithm.
6 Conclusion Aiming at the data structure of Cadence, this paper divides digital signal into signal expressed by analog quantity and signal expressed by digital quantity, and gives specific algorithms respectively for relevant feature parameters of these two types of digital signals in Cadence-based environment, By using a typical case, it validates that the arithmetic is correct and then provide necessary technical support for the circuit fault simulation based on Cadence. On the base of realization of fault simulation to obtain large numbers of fault simulation judgment, it then can make further reliability analysis, including fault mode effect analysis (FMEA), reliability parameter prediction, test-purpose parameter prediction and etc.
References 1. Zhao, G.: Signal Analysis and Processing. China Machine Press, Beijing (2010) 2. Lu, K., Zhuo, Y.: Extraction of Multipath Timedelay in Frequency Domain to UWB Chirp Signal. J. Communications Technology. 43(12), 15–17 (2010) 3. Li, L.M.: Analysis of Nonlinear Oscillators Using Volterra Series in the Frequency Domain. Journal of Sound and Vibration 330(2), 335–337 (2011) 4. Xu, Y., Wang, G., Guo, Y.: Wavelet Package Based Speech Enhancement Algorithm Using Time-Frequency Threshold. Journal of Electronics & Information Technology 30(6), 1363–1366 (2008) 5. Ayat, S., Manzuri, M.T.: Wavelet based speech enhancement using a new thresholding algorithm. In: International Symposium On intelligent Multimedia, Video and Speech Processing, pp. 238–241. IEEE Press, Hong Kong (2004) 6. Yang, C.: Study on Feature Extraction of Fault Circuits Based on Wavelet. Jilin University, Jilin (2004) 7. Yu, G.: A cluster-based wavelet feature extraction method for machine fault diagnosis. J. Applied Mechanics and Materials 10, 522–548 (2008) 8. Sreejith, B., Verma, A.K., Srividya, A.: Fault diagnosis of rolling element bearing using time-domain features and neural networks. IEEE Press, New York (2008) 9. Lee, H., Nguyen, N., Kwon, J.: Bearing diagnosis using time-domain features and decision tree. In: Region 10 Colloquium and 3rd International Conference on Industrial and Information Systems, pp. 952–960. IEEE Press, New York (2008) 10. Zhang, M., Zhao, G., Sun, Y.: Research on Arithmetic of Fault Judge in typical circuit. Journal of Naval Aeronautical and Astronautical University 24(1), 115–117 (2009)
Research of Stereo Matching Based on Improved Median Filter Huiyan Jiang, Rui Gao, and Xiaojie Liu Software College, Northeastern University, 110819 Shenyang, China
[email protected] Abstract. In the refinement process of traditional median filter for the disparity map, the estimation of disparity for some pixels is not always precise. In order to solve this problem, the thesis presents an improved median filter for disparity refinement method. The proposed method first uses the left/right consistency test and color segmentation to detect the mismatch regions and discontinuities sections, and then filter out these regions in the filter window so as to achieve a better estimate of disparity for current pixel. Experimental results clearly indicate that this method can obtain a more precise disparity map compared with conventional method. Keywords: Stereo matching; Binocular vision; Filter; Median filtering.
1 Introduction Stereo vision is the research focus of computer vision [1]. With the development of computer technology, stereoscopic vision in robot vision, industrial measurement, object recognition and military field have been widely used. Stereo matching plays an important role on the three-dimensional visual field, but as a result of noise and other factors during the imaging process, we have not found a universally applicable method for stereo matching currently [2]. Thus, the research for the stereo matching method has important theoretical significance. At present, the research for the stereo matching algorithm mainly concentrates in two directions. One is a local optimal algorithm based guidelines for local support neighborhood, and the other is an algorithm based global strategy to implement global optimization. Both of algorithms have advantages and disadvantages. For the local optimal algorithm, the main advantage is easy to implement and the algorithm is fast, while the disadvantage is that the accuracy of obtained disparity map is low. The disparity map accuracy of the algorithm based global strategy to implement global optimization is high, but it is difficult to obtain a global optimal solution [3]. Therefore, accurate and fast stereo matching algorithm has been an important research direction. After the analysis of the problem that traditional median filter applied to the disparity map refinement process, this thesis presents an improved median filter for M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 479–486. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
480
H. Jiang, R. Gao, and X. Liu
disparity refinement method, and then make a comparison between traditional filter and improved median filter. Moreover, introduce the evaluation standard of the stereo matching. Experimental results show that this method can obtain a more precise disparity map compared to conventional method, and it is more effective than the traditional median filter based on stereo matching.
2 Filters on Image Processing Gauss filter is a linear smoothing filter based on the shape of Gaussian function to select the right value [4]. It’s very effective on wiping out normally distributed noises. Filter is a mathematical model, and it can convert the image data to the energy. When the energy is low, exclude low energy which stands for the noise. Generally, by choosing proper weights and template, the value of a specific pixel is the result of weighted mean on its own and nearby pixels [5]. Mean filter algorithm is one of the classic linear filtering algorithms by replacing the value of the central pixel with the average value of its most closely surrounded eight other neighbor pixels and itself. The normal processing course is choosing a kind of template consisted with a certain amount of neighbor pixels for current pixel ( x, y ) , then calculating the mean value of this template as the final gray scale value g ( x, y ) of ( x, y ) which is
g ( x , y ) = 1 ∑ f ( x, y ) m
(1)
m here is the amount of pixels in the template [6]. Median filter algorithm based on statistical theory of sorting, is one of the nonlinear signals processing technology to inhibit noises effectively. Weak textures in some areas, due to relatively low signal to noise ratio images will cause the failure of matching, so there will be some mismatch point, so the median filter can be filled on the point of these errors The basis course of Median filter algorithm is replacing a specific value of a numeric sequence or digital image with the middle value of its neighborhood, so the pixels values of surrounding area will be close to the actual one so that eliminate isolated noise points [7]. Median filter in image processing commonly is a classic method of smoothing the noise, and it is used to protect the edges. Firstly, sort the value of pixels in a specific shaped 2-D floating template monotonously in ascending order or reverse order, then replace the current pixel value, as in
g ( x, y ) = med { f ( x − k , y − 1), (k , l ∈ W )}
(2)
Where f ( x, y ) is the original image while g ( x, y ) is the processed one. W stands for 2-D floating template which generally is an area with size of 2*2 or 3*3[8], but it also can be shapes like bars, circles, crosses and rings, etc.
Research of Stereo Matching Based on Improved Median Filter
481
3 Stereo Matching Based on Improved Median Filter By analyzing and comparing different filter algorithms, it is not hard to find Gauss filtering is suitable for the image containing gauss noise. Mean filter may produce some influence for the image edges, and the disparity contains some isolated point, while the median filter is more suitable for the process of disparity map refinement [9]. There may be two problems in the process of disparity map refinement. The one problem is that filter window may contain some errors matching area, and the other second problem is that filter window may contain some disparity discontinuity region. Therefore, estimating the parallax of current pixel, if taking the parallax of the mismatching and discontinuity pixel into account, it may lead to a further error of the current pixel evaluation. For the purpose of solving the above problem, the paper proposes an improved median filter method. Although there are many kinds of stereo matching algorithms, their solving steps are basically the same. Scharstein divides it into four steps: matching cost calculation, cost value accumulation, disparity calculation and disparity refinement [1, 10]. The paper uses two images each time during the experiments. One is the reference image as left view and the other is the target image as right view. Firstly process the two images, and cost the value accumulation. Then achieve the disparity calculation based on Winner-Take-All (WTA). The most important step is to detect the mismatch regions and discontinuities sections through using the left/right consistency test and color segmentation, and then filter out these regions in the filter window so as to achieve a better estimate of disparity for current pixel. The specific steps are described as follows. Matching cost calculation. For each pixel ( x, y ) of the reference image L* ( x, y ) ,
calculating the cost ( x + d , y ) of maxim to the corresponding pixel in the target image R* ( x, y ) , d is (0, d max ) , d max is a different value for different stereo images, we can get the matching cost for each pixel in the reference image, the calculation formula is as follow [11]
C0 (x, y, d) =
∑
c∈{r, g,b}
Lc (x + d, y) − Rc (x, y)
(3)
Cost value accumulation. The disparity of one pixel has certain correlation with the pixels around it. Local algorithm usually sets the matched pixel as the center of the window whose size is (2 L + 1) ∗ (2 L + 1) , then choose the mean value of the pixels in the window as the value of the center pixel.
C ( x, y , d ) =
1 x+ L y+ L ∑ ∑ C0 ( x , y , d ) 9 i= x− L j= y− L
(4)
482
H. Jiang, R. Gao, and X. Liu
Disparity calculation and refinement. From the last step, we can get C ( x, y, d ) . For * each pixel (u , v) in the disparity, choose the minimum value of C ( x, y, d ) among
(0, d max ) , and then d * is the value of DL (u, v) . Before the process of median filter, we can use the left/right consistency test and color segmentation to detect the mismatch regions and discontinuities regions. After eliminating these regions which are detected in the filter window, use the remaining pixels as the elements of the median filter to estimate the parallax of current pixels. There are three steps as follows. The first step is left/right consistency test. Exchange the reference image and target image, making the right image as the reference image and making the left image as the target image, and then get the disparity image DR of the right image R according to step C and step D. For each pixel of disparity image DL of the left image, if
DL (x, y) ≠ DR (x + dL (x, y), y) , then DL ( x, y ) = 0 ,and if DL (x, y) = DR ( x + d L ( x, y ), y ) , else DL (x, y) = DL (x, y) . The second step is detection of discontinuous region based on Mean Shift algorithm. Use Mean Shift to segment the reference image L* ( x, y ) in order to get the tagged image Label . Those regions in Label whose colors are similar will be assigned the same label. The third step is improved median filter. According to the disparity image DL which is obtained under the first and second step and the tagged image Label , use the median filter to process DL . Detailed steps are as follows. For each pixel ( x, y ) in the disparity image DL , choose a window which size is (2 L + 1) ∗ (2 L + 1) and identify ( x, y ) which belong to the U = {( u, v) | u ∈ ( x − L, x + L), v ∈ ( y − L, y + L)} as the
centre. Then select (u , v) within the set of coordinate point and make that DL (u , v) is not zero and Label (u, v) = Label(x, y) .That means the set is U* = {( u, v) | u ∈ (x − L, x + L), v ∈ ( y − L, y + L), DL (u, v) ≠ 0, Label (u, v) = Lable( x, y )}
Finally, choose the middle value of coordinate set DL (U* ) as the final disparity of the DL (x, y) , and DL is the final disparity image.
4 Experiment Results This paper uses the four pairs of stereo image pairs: map, tsukuba, sawtooth, venus, they are chosen from the Middlebury stereo image library. The results shown in Fig. 1, where (a), (e), (j), (o) for the corresponding threedimensional images on the left view , (b), (f), (k), (p) for the corresponding threedimensional view of the right image, (c), (g), (l), (q) for the three-dimensional view of the true image of the left disparity map, (d), (h), (m), (r) for the traditional algorithm the experimental results obtained, (e), (i), (n), (s) for the experimental results obtained by the paper .
Research of Stereo Matching Based on Improved Median Filter
(a)
(b)
(c)
(e)
(f)
(g)
(h)
(i)
(j)
(k)
(l)
(m)
Fig. 1. Compared experiment results.
483
484
H. Jiang, R. Gao, and X. Liu
(n)
(o)
(p)
(q)
(r)
(s)
Fig. 1. (continued)
Compared with the traditional median filter in Fig.1 and the paper eliminates most of the obvious mistakes region and relatively gets smooth disparity in the absence of texture regions. The effect is quite satisfactory, as it is smooth on the background at the top right corner of the image tsukuba, venus the left of the image, etc. Table1 is the table of false match rate .The data appears from the results that this method resulting in false match rate of non-matching block area, no texture regions and disparity discontinuities is superior to the traditional methods. In particular, the average error matching rate increases 9.36% in non- block area and achieves satisfactory results. But in the right half of the lamp in Figure tsukuba can see that the color segmentation result is not accurate because of the shadow, and then disparity estimation error occurs. On the other hand, the paper uses color segmentation and detection of about consistency during the process of disparity map refinement, so there is an increase in running time of the algorithm. From the running time of view in Table2, despite the increase in the running time, the time complexity is still within the acceptable range. In the last line the Mean Shift color segmentation algorithm is used in each process of image segmentation, and it takes almost half the whole time. Looking for faster and more accurate color segmentation approach is still to this next step to conduct research.
Research of Stereo Matching Based on Improved Median Filter
485
Table 1. Comparison of algorithm mistakes rate.
Parameter
Tsukuba (%)
Map (%)
Sawtooth (%)
Venus (%)
Traditional Method
Paper Method
BO
8.31
4.37
BT
11.2
4.92
BD
15.7
10.7
BO
3.40
2.21
BT
0.71
0.57
BD
13.9
6.77
BO
3.91
5.96
BT
1.84
8.50
BD
3.91
5.96
BO
1.84
8.50
BT
3.91
5.96
BD
1.84
8.50
Table 2. The runtime(s) comparison of different methods
Parameter
Mean Shift
Traditional Method
Paper Method
Tsukuba
2.44
1.43
5.27
Map
1.56
1.20
3.92
Sawtooth
4.03
2.42
8.31
Venus
3.92
2.40
8.41
In Table1 and Table2, this method is superior to the traditional methods in false match rate of non-matching block area, no texture regions and disparity discontinuities. There is an increase in the running time, but the time complexity is still within the acceptable range.
5 Conclusion Based on the Improved Median Filter this paper is compared with the traditional median filter and implements the Stereo Matching, and introduce evaluation of stereo matching algorithm proposed by Scharstein. The experiment results can eliminate most obvious mistake area and relatively get smooth parallax in the absence of texture areas. Moreover, it can obtain a more accurate matching result. In the process of
486
H. Jiang, R. Gao, and X. Liu
disparity refinement, mean-shift algorithm is time consuming, so researching a faster and more accurate color segmentation method is the future work.
Acknowledgement This research is supported by the National Nature Science Foundation of China (No: 60973071, No: 50834009) and the Liaoning province Natural Science Foundation (No: 20092004).
References 1. Scharstein, D., Szeliski, R.: A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. International Journal of Computer Vision, 7–42 (2002) 2. Van Meerbergen, G., Vergauwen, M., Pollefeys, M., Van Gool, L.: A hierarchical stereo algorithm using dynamic programming. In: Proceedings of IEEE Workshop on Stereo and Multi-Baseline Vision, pp. 166–174. IEEE Computer Society Press, Washington, USA (2001) 3. Birchfield, S., Tomasi, C.: A pixel dissimilarity measure that is insensitive to image sampling. In: IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 401–406. Springer, Heidelberg (1998) 4. Hartley, R.I.: Theory and practice of projective rectification. In: International Journal of Computer Vision, pp. 115–127. IEEE Computer Society Press, Los Alamitos (1998) 5. Chan, R.H., Ho, C.W., Nikolova, M.: Salt and pepper noiseremoval by median type noise detectors and detail preserving regularization. IEEE Trans.Image Process, 1479–1485 (2005) 6. Bobick, A.F., Intille, S.S.: Large occlusion stereo. International Journal of Computer Vision, 181–200 (1999) 7. Intille, S.S., Bobick, A.F.: Disparity-space images and large occlusion stereo. In: Eklundh, J.-O. (ed.) ECCV 1994. LNCS, vol. 801, Springer, Heidelberg (1994) 8. Dornaika, F., Chung, R.: Cooperative Stereo Motion: Matching and Reconstruction. In: Computer Vision and Image Understanding, pp. 408–427. Springer, Heidelberg (1998) 9. Fukunaga, K., Hostetler, L.D.: The estimation of the gradient of a density function with application in Pattern recognition. IEEE Trans. Information Theory, 32–40 10. Gong, M.L., Yang, Y.H.: Fast stereo matching using reliability-based dynamic programming and consistency constraints. In: Proceedings of the Ninth IEEE International Conference on Computer Vision, pp. 610–617. IEEE Computer Society Press (2003) 11. Felzenszwalb, P.F., Huttenlocher, D.P.: Efficient belief propagation for early vision. International Journal of Computer Vision, 261–268 (2006)
Generating Method of Three-Phase Voltage Sag in EV Charging System Performance Testing Xiaoming Yue, Hui Fan, Jin Pan, Xiaoguang Hao, and Zhimeng Zhang Hebei Electric Power Research Institute, Shijiazhuang, Hebei 050021, China
[email protected] Abstract. In order to analyze the impact between Electric Vehicle (EV) charging system and grid influence. In this paper, a method of three-phase voltage sag generating based on DQ transformation is proposed. With the ‘Back to Back’ linked two PWM voltage-source converters, normal and fault state of infinite grid is simulated. Aiming at seven typical voltage sag faults (including all the voltage sag types in symmetric, asymmetric excluding zero-sequence and asymmetric contain zero-sequence circumstances), compound operation is completed in synchronous rotational axes (dq0) under symmetrical component method applied. Especially, abc-dq0 transform is amended for asymmetric contain zero-sequence voltage sag situation. So the space vector pulse width modulation (SVPWM) control signal is generated. Three-phase voltage sag generated with this method could satisfy the phase relationship and amplitude relationship of the inter-phase automatically. As a result, common connection point (PCC) fault voltage could be accurately generated. And the full scope of the depth of voltage sag adjustable and fault duration of any settings are realized. Simulation and experiment results show that this method is the rationality and feasibility. Keywords: Electric Vehicle, Voltage Sag, Voltage-source Converter, DQ Transformation.
1 Introduction Along with the progress of the society and the development of economy, the energy issue gradually becomes the focus. Electric power is essential in modern society. Meanwhile, with the increasing of cars, pollution from cars becomes more and more serious. So Electric Vehicle becomes a good solution for that. But wide-scale EV charging will bring a lot of trouble to grid, and the charging performance of the electric vehicle charger will be influenced by grid, too. Thus, the mutual influence between the wide-scale EV charging and the grid has important significance. In the analysis of the electric car battery and grid interaction, the grid voltage sag is one of the common problems. But the grid voltage sag fault has its randomness and uncertainty, so how to correctly generated grid fault situations becomes a problem must be solved. M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 487–496. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
488
X. Yue et al.
Generally, a lot of different grid fault types can occur, such as single-phase-toneutral, phase-to-phase, 2-phase-to-neutral, and 3-phase faults. The form of grid voltage sag can be divided into seven types (A, B, C, D, E, F and G) [1-5]. There are some methods of generating voltage sag, but they all have their own shortcomings and couldn’t accurately reflect the actual situation of power grid faults. This paper proposes a voltage sag generating method based on DQ transformation. The main topology of this method is two three-phase PWM voltage-source converter (VSC) connected by the back-to-back structure, generating a symmetrical, asymmetric excluding zero sequence and asymmetric contain zero sequence three circumstances of all types of voltage sag, realized the full scope of the voltage sag depth and fault duration adjustable of any settings.
2 Main Topology and Principle The principle diagram of the generating method in this paper is shown in Fig. 1. Main circuit adopts two three-phase PWM voltage-source converters (VSC), which is connected in "back-to-back" form. This could simulate the power supply characteristics of infinite power grid in normal and various fault conditions. The energy between grid and converter can flow bilaterally. Thereby fault generating side PWM voltage-source converter can connect passive load, also can connect active load. The grid side PWM voltage-source converter, used SVPWM double closed loop control method [6], produces a constant DC voltage in the DC side, realizes the AC unit power factor controlled, and reduces grid harmonics pollution.
Fig. 1. The principle diagram
The vector control principle diagram of the fault generating side converter is shown in Fig. 2. The grid fault voltage is decomposed into positive sequence voltage, negative sequence voltage and zero sequence voltage by the symmetric component method [7-8]. Seven kinds of typical of voltage sag signals (A, B, C, D, E, F and G) come through the stator side 3-phase stationary reference frame to synchronous revolution frame by abc-dq0 transformation. With the compound operation in dq0 frame and the dq0-abc transformation, space vector pulse width modulation (SVPWM) control signals are generated. Accordingly various-type three-phase voltage sag waveform of arbitrary voltage sag depth and arbitrary duration are gained.
Generating Method of Three-Phase Voltage Sag in EV Charging System
489
L
Grid Side Converter
Charger C
abc
SVPWM
Fault Type
dq0
Symmetrical Phasor Synthesis
Eigenvalues Phasor
Fault Time Sag Depth
Phasor Decompositio n
Fig. 2. SVPWM control schematic diagram of fault side converter
3 Generating Method of Three-Phase Voltage Sag Grid short-circuit fault can be divided into two kinds such as symmetrical fault and asymmetric fault. The most common asymmetric short-circuit fault are single-phaseto-neutral, phase-to-phase, 2-phase-to-neutral. And asymmetric short-circuit faults are divided into two kinds. One kind contains zero sequence components; another kind is excluding zero sequence components. All kinds of voltage sag fault has his different features. Voltage sag fault is divided into seven typical types, such as Type A, Type B, Type C, Type D, Type E, Type F and Type G. Type A is the three-phase symmetric voltage sag caused by three-phase short-circuit fault; Type B is the asymmetric voltage sag caused by single-phase-to-neutral fault; Type C, D are the asymmetric voltage sag caused by phase-to-phase fault; Type G, E, F are the asymmetric voltage sag caused by 2-phase-to-neutral fault. Each phase voltage of the entire voltage sag fault at the PCC
. . .
、、
point [9-12] is shown in Table 1. U a U b U c are the phase voltage phases at the point of common connection (PCC). U is the characteristic value of the voltage sags. Based on the symmetrical component method, voltage sag in PCC can be resolved into positive-sequence component, negative-sequence component and zero-sequence component. The transform equation as follows:
⎡ . ⎤ ⎢U1 ⎥ ⎡1 a ⎢ . ⎥ 1⎢ ⎢U 2 ⎥ = ⎢1 a 2 ⎢ . ⎥ 3⎢ 1 1 ⎢U ⎥ ⎣⎢ ⎢⎣ 0 ⎥⎦ D
Where a is defined as a = e j120 .
a
2⎤
⎥ a⎥ ⎥ 1⎥ ⎦
⎡ . ⎤ ⎢U a ⎥ ⎢ . ⎥ ⎢U b ⎥ ⎢ . ⎥ ⎢U ⎥ ⎢⎣ c ⎥⎦
(1)
490
X. Yue et al. Table 1. Voltage-sag types and the corresponding phase voltages of PCC Type
U a
U b
U c
A
U
1 3 − U + j U 2 2
1 3 − U + j U 2 2
B
U
1 3 − −j 2 2
C
1
−
D
U
1 3 − U − j 2 2
1 3 − U + j 2 2
E
1
1 3 − U − j U 2 2
1 3 − U + j U 2 2
F
U
G
2 1 + U 3 3
1 3 − j U 2 2
−
−
1 3 + j 2 2
1 3 + j U 2 2
1 3 3 1 3 3 − U − j ( + U ) − U + j ( + U) 2 3 6 2 3 6
1 1 3 − − U − j U 3 6 2
1 1 3 − − U + j U 3 6 2
Vector Chart
Generating Method of Three-Phase Voltage Sag in EV Charging System
491
The positive sequence components, negative sequence components and zero sequence components of all the typical voltage sag fault corresponding PCC point can be seen in Table 2. Obviously voltage sag in Type B and E contain zero sequence components, Type A, Voltage sag in Type C, Type D, Type F and Type G exclude zero sequence components. Only voltage sag in Type A is symmetrical, the other types are asymmetric situation. Table 2. Positive, negative and zero sequence voltage for each sag type
Type
.
.
.
Np
Nn
N0
A
U
0
0
B
(2 + U ) / 3
(U − 1) / 3
(U − 1) / 3
C
(1 + U ) / 3
(1 − U ) / 2
0
D
(1 + U ) / 2
(U − 1) / 2
0
E
(1 + 2U ) / 3
(1 − U ) / 3
(1 − U ) / 3
F
(1 + 2U ) / 3
(U − 1) / 3
0
G
(1 + 2U ) / 3
(1 − U ) / 3
0
.
As. shows in Table 2, N P is the characteristic value of positive-sequence voltage, . the N n is characteristic value of negative-sequence voltage, the N 0 is characteristic value of zero- sequence voltage. Therefore, the phase voltage of PCC voltage sag in three-phase static axes (a, b, c) can be expressed as: .
.
.
.
U PCC = N p e jωet + N n e− jωet + N 0 e jωet
(2)
With the abc-dq0 transformation, the phase voltage of PCC voltage sag in the synchronous rotational axes (d, q, 0) shown in (3): .
.
.
.
U PCC = N p + N n e− j 2ωet + N 0
(3)
Thus, phase voltage sag at PCC is composed of the symmetrical components from seven typical voltage sag decomposition. Then d-axis, q-axis, o-axis components can be received by abc-dq0 transformation. After the compound operation, three-phase fault voltage modulation signal is got by the dq0-abc transformation. For excluding the zero sequence components of voltage sag, the traditional dq0-abc transformation matrix [13] as follows:
492
X. Yue et al.
⎡ ⎤ ⎢ cos ω t − sin ω t 1⎥ ⎡ ua ⎤ ⎢ ⎥ ⎡ud ⎤ ⎢u ⎥ = ⎢ cos(ω t − 2π ) − sin(ω t − 2π ) 1⎥ ⎢ u ⎥ ⎢ b⎥ ⎢ ⎥⎢ q⎥ 3 3 ⎢⎣uc ⎥⎦ ⎢ ⎥ ⎢⎣ u0 ⎥⎦ ⎢cos(ωt + 2π ) − sin(ωt + 2π ) 1⎥ ⎥⎦ 3 3 ⎣⎢
(4)
And for the voltage sag containing zero sequence components, dq0-abc transformation matrix (4) should be revised to ⎡ ⎤ ⎢ cos ω t − sin ωt cos ωt ⎥ ⎡ ua ⎤ ⎢ ⎥ ⎡ud ⎤ ⎢ u ⎥ = ⎢ cos(ω t − 2π ) − sin(ω t − 2π ) cos ω t ⎥ ⎢ u ⎥ ⎢ b⎥ ⎢ ⎥⎢ q⎥ 3 3 ⎢⎣ uc ⎥⎦ ⎢ ⎥ ⎢⎣ u0 ⎥⎦ ⎢cos(ω t + 2π ) − sin(ω t + 2π ) cos ω t ⎥ ⎢⎣ ⎥⎦ 3 3
(5)
In the generating method of three-phase voltage sag in this paper, the dq0-abc transformation uses transform matrix (5). So any type of voltage sag is generated, including symmetric, asymmetric excluding zero sequence and asymmetric containing zero sequence.
4 Simulation and Experimental Results Grid side converter side: Power frequency f Z is 50Hz; Source voltage Us is 220V; The dc given voltage Vdc is 400V; The carrier frequency f C is 3kHz. Fault generating converter side: Filtering inductances L is 0.4mH; Filter capacitance C is 140μF ; Symmetric three-phase resistor load R is 50Ω . The duration of voltage sag is 100ms. The simulation results is shown in Fig. 3-Fig. 9. And part of experimental results is shown in Fig. 10-Fig. 11.
(a) Voltage sag 70% UN
(b) Voltage sag 30% UN
Fig. 3. Waveforms of three-phase voltage sag in type A
Generating Method of Three-Phase Voltage Sag in EV Charging System
(a) Voltage sag 70% UN
(b) Voltage sag 30% UN
Fig. 4. Waveforms of three-phase voltage sag in type B
(a) Voltage sag 70% UN
(b) Voltage sag 30% UN
Fig. 5. Waveforms of three-phase voltage sag in type C
(a) Voltage sag 70% UN
(b) Voltage sag 30% UN
Fig. 6. Waveforms of three-phase voltage sag in type D
(a) Voltage sag 70% UN
(b) Voltage sag 30% UN
Fig. 7. Waveforms of three-phase voltage sag in type E
493
494
X. Yue et al.
(a)Voltage sag 70% UN
(b)Voltage sag 30% UN
Fig. 8. Waveforms of three-phase voltage sag in type F
(a) Voltage sag 70% UN
(b) Voltage sag 30% UN
Fig. 9. Waveforms of three-phase voltage sag in type G
Part of the experimental results is depicted in Fig. 10 and Fig. 11. Fig. 10 shows the waveforms of three-phase voltage sag in type A with voltage sag 70%UN. Fig. 11 shows the waveforms of three-phase voltage sag in type G with voltage sag 80%UN.
Fig. 10. Waveforms of three-phase voltage sag sag with voltage sag 70% UN in type A
Fig. 11. Waveforms of three-phase voltage with voltage sag 80%UN in type G
The simulation and experimental results show the three-phase voltage sag waveforms of seven typical voltage sag with different voltage sag depth. The simulation and experimental results show that the generating method of the three-phase voltage
Generating Method of Three-Phase Voltage Sag in EV Charging System
495
sag in this paper can generate all the kinds of voltage sag fault, with voltage sag depth and fault duration set arbitrary. Meanwhile each phase voltage automatic meet phase relationship and amplitude relationship between any two phases. So the power system fault condition is reflected factually.
5 Conclusion This paper adopts "back-to-back" PWM converter as the main circuit of three-phase voltage sag generator with which energy can flow two-way. The method of threephase voltage sag proposed in this paper is based on DQ transformation. The voltage sag signal is decomposed into the positive and negative, zero sequence voltage using symmetric component method. The compound operation is completed with abc-dq0 transformation. And the traditional dq0-abc transform type is revised, and then the voltage sag of SVPWM modulation signal is generated. The three-phase voltage sag generated by this method automatic meet the phase relationship and the amplitude relationship between any two phases. The PCC voltage is reproduced more accurately, which includes symmetric, asymmetric excluding zero sequence and asymmetric containing zero sequence. And the full scope of the depth of voltage sag adjustable and fault duration of any settings are realized. The simulation and experimental results verify the correctness of the generating method described in this paper.
References 1. Liu Wanshun, C.: Electric Power System Fault Analysis. China Electric Power Press, Beijing (1998) 2. Xiao, X.: C: Power Quality Analysis and Control. China Electric Power Press, Beijing (2004) 3. Bollen, M.H.J.: C.: Understanding Power Quality Problems: Voltage Sags and Interruptions. IEEE Press, Piscataway (2000) 4. Xiao, X., Tao Shun, M.S.: Voltage sags types under different grounding modes of neutral and their propagation Part I. J. Transactions of China Electrotechnical Society 22(9), 143–147 (2007) 5. Xiao, X., Tao, S., M.S.: Voltage sags types under different grounding modes of neutral and their propagation Part I. J. Transactions of China Electrotechnical Society 22(10), 156–159 (2007) 6. Zhang, C., Zhang, X.: C: PWM Rectifier and Control. China Mechanic industry Press, Beiing (2003) 7. Yan, X., Zhang, B., Gu, X., et al.: M.S.: Closed-loop control of three-phase five-level PWM current source inverter. J. Transactions of China Electrotechnical Society 22(sup1), 60–64 (2007) 8. Collins, E.R., Morgan, R.L.: M.S.: A three-phase sag generator for testing industrial equipment. J. IEEE Trans- actions on Power Delivery 11(1), 526–532 (2003) 9. Yan, X., Venkataramanan, G., Yang, W.: M.S.: Fault Tolerance of DFIG Wind Turbine with a Series Grid Side Passive Impedance. Industry Applications Society Annual Meeting, 1–8 (2009)
496
X. Yue et al.
10. Boonmee, C., Kumsuwa, Y., Premrudeepchacharm, S.: M.S.: Implementation of real time three phase balanced voltage sag generator 1 kVA: using microcontroller and PC control. In: ICROS- SICE International Joint Conference, pp. 903–907 (2009) 11. Oranpiroj, K., Premrudeepreechacham, S., et al.: M.S.: The 3-phase 4-wire voltage sag generator based on three dimensions space vector modulation in abc coordinates. In: IEEE International Symposium on Industrial Electronics (ISIE 2009), pp. 275–280 (2009) 12. Collins, E.R., Morgan, R.L.: M.S.: A three-phase sag generator fortesting industrial equipment. IEEE Trans. Power Delverry II, 526–532 (1996) 13. Nanjing Institute. C.: Power System. China Electric Power Press, Bei Jing (1982)
Improved Predictive Control of Grid-Connected PV Inverter with LCL Filter Huijie Xue, Wei Feng, Zilong Yang, Chunsheng Wu, and Honghua Xu Department of Renewable Energy Power Generation Institute of Electrical Engineering, CAS No.6 Beiertiao, Zhongguanchun Haidian District, Beijing, China {xuehuijie,fengwei,zlyang,wcsxg hxu}@mail.iee.ac.cn
Abstract. LCL filter has high insertion loss and is expected to replace LC filter in the grid-connected PV inverter. However, the inverter with LCL filter is hard to be control and instability is liable to be incurred. So it is necessary to research reliable control strategy for the inverter with LCL filter. This paper researched the problem in detail. Discrete state space model of the inverter with LCL filter was firstly obtained with Clarke transformation and ZOH method. Then, improved predictive control strategy was proposed and explained in detail. Simulation was carried on and the results verified control effect of the proposed control strategy. Compared with prior control strategies, the proposed control strategy has the merits of simplicity, fast response and robustness. Keywords: PV Inverter LCL Filter Predictive Control.
1 Introduction As a kind of clean renewable energy resource, solar PV is undergoing rapid growth in the past several years [1, 2]. Grid-connected operation is the main utilization method of the PV system. The inverter is an important equipment of the grid-connected PV system. As a switching mode power electronics converter, the inverter will produce plenty of high frequency ripple current during normal operation. If not filtered out adequately, the ripple current will be injected into the grid and decreases the power quality even produces EMI to other electric and electronic equipment. So the filter with high insertion loss is desirable in the PV inverter. At present, most PV inverters adopt the LC inverter. In the view of the filter performance, the LCL filter is prior to the LC filter because of higher insertion loss, lower profile and lower cost. However, the LCL filter is hard to be controlled and liable to oscillate if not controlled appropriately [3]. The scholars have proposed many different control strategies for the inverter with the LCL filter [4, 5]. However, the strategies have respective shortcoming such as extra signal sensing [4] or limitation on the filter parameter [5]. This paper proposed an improved predictive control strategy of the PV inverter with the LCL filter. The proposed control strategy has explicit physical sense and the merit of simplicity, fast response and robustness. M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 497–502. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
498
H. Xue et al.
The rest of the paper is organized as follows: Section 2 models the PV inverter with the LCL filter. Section 3 proposes the improved predictive control of the PV inverter with the LCL filter in detail. Simulation results are given in Section 4.
2 Modeling the Inverter with LCL Filter The 3-phase PV inverter with the LCL filter is shown in Fig. 1. The line frequency transformer may be included into the model if necessary. The leakage inductance of the transformer can be included in L2 in that case.
Fig. 1. Schematic of the Inverter with LCL Filter
Assume that the parameters of the LCL filter of every phase of the inverter are identical. Take the voltage across the filter capacitor C , the current through the filter inductor L1 and the L2 as the state variables, the output voltage of the inverter bridge U inv as the input variable and the grid voltage U grid as the external disturbance, the state equation of Fig.1 can be expressed as: i
X o = Ao X o + BoU o + N oWo , Yo = C o X o where
X o = ⎡⎣uCa
iL1a
iL 2a
uCb
iL1b
iL 2b
uCc
iL 2c ⎤⎦
iL1c
(1) T
U o = ⎡⎣uinva
T
uinvb
uinvc ⎤⎦
T
Ao = diag{ Ao a, Aob, Ao c} Yo = ⎡⎣iL 2a iL 2b iL 2c ⎤⎦ Wo = ⎡⎣u grida ugridb u gridc ⎤⎦ Bo = diag{Bo a, Bo b, Bo c} No = diag{No a, Nob, Noc} Co = diag{Co a, Cob, Co c} T
⎡ ⎢ 0 ⎢ ⎢ 1 Ao a = Ao b = Ao c = ⎢ − ⎢ L1 ⎢ 1 ⎢ ⎣ L2
⎡ Bo a = Bo b = Bo c = ⎢ 0 ⎣ C o a = C o b = C o c = [0
1 ⎤ ⎥ C ⎥ ⎥ 0 ⎥ ⎥ R ⎥ − 2⎥ L2 ⎦
1 C R − 1 L1
−
0
1 L1 0 1] .
⎤ 0⎥ ⎦
T
T
⎡ 1⎤ , N o a = N ob = N o c = ⎢0 0 − ⎥ L2 ⎦ ⎣
Improved Predictive Control of Grid-Connected PV Inverter with LCL Filter
499
Principally, the current through L2 should be controlled to inject sinuous current into the grid. However, the current through L1 is almost the same with that through the inductor L2 because of the low capacitance of C. Moreover, the current through it is easy to be compensated if necessary. So, satisfactory results can be achieved by controlling the current through L1 in the grid-connected inverter. With the control objective of the current through L1 , the model of (1) can be simplified as (2) by looking the current through L2 as external disturbance and eliminating the redundancy with Clarke transformation. i
X t = At X t + BtU t + N tWt , Yt = Ct X t
(2)
where X t = ⎡uCα ⎣
iL1α
uC β
iL1β ⎤ ⎦
T
U
t
= ⎡ u in v α ⎣
u in v β ⎤ ⎦
T
T
Yt = ⎡iL1α iL1β ⎤ W t = ⎡ iL 2α iL 2 β ⎤ ⎣ ⎦ ⎣ ⎦ At = diag{ Atα , At β } Bt = diag{Btα , Bt β } N t = diag{N tα , N t β } Ct = diag{Ctα , Ct β } T
⎡ ⎢ 0 Atα = At β = ⎢ ⎢− 1 ⎢⎣ L1
1 ⎤ C ⎥ ⎥ R − 1⎥ L1 ⎥⎦
⎡ B tα = B t β = ⎢ 0 ⎣
1 ⎤ ⎥ L1 ⎦
T
⎡ 1 N tα = Nt β = ⎢ − ⎣ C
⎤ 0⎥ ⎦
T
C tα = C t β = [ 0 1] .
In order to design digital controller directly, discrete model should be obtained. Assuming the sampling period is Ts , the discrete model can be obtained by applying zero order holder method [6] and can be expressed as:
X d ( n + 1) = Ad X d ( n) + BdU d ( n) + N dWd ( n) Yd ( n) = Cd X d ( n)
,
(3)
where Ad = e At ⋅Ts , Bd = e AtTs BtTs , N d = e A T N tTs and Cd = Ct . t s
3 Proposed Control Strategy of Inverter with LCL Filter The proposed predictive control of the inverter with the LCL filter is easy to understand according to the physical sense of the controlling behavior and is explained as follows. In order to control the current through L1 at the instant (n + 1)Ts reach I ref ( n ) , the reference at the instant of nTs , the control input, i.e. the output voltage of the inverter U d (n) , can be obtained according to (4).
500
H. Xue et al.
U d (k ) = (Cd Bd )−1 ( I ref (k ) − Cd Ad X d (k ) − Cd N dWd (k )) ,
(4)
Generally, the sampling instant of the control system occurs at the middle of the control period to avoid alias effect. However, the updating of the modulation signal is finished at the end of the control period due to the computation delay. Besides, updating at the end of the control period is desirable to avoid asymmetric pulse width modulation of the inverter. So the delay of half switching period is produced by the computation and updating behavior. Moreover, one switching period is necessary for the inverter to finish the current control. So total one and half switching periods delay is unavoidable in the predictive control. In the high power PV inverter, the time delay can’t be omitted because the switching frequency is limited to only several kilohertz to avoid high switching loss. So it is necessary to overcome the delay. The delay can be compensated with the following measure. Denote the components of control input of the α axle and β axle with U dα and U d β respectively, the improved strategy can be expressed with U dαβ (k ) = e jθ (U dα (k ) + jU d β (k )) ,
(5)
where the real part and imaginary part is the control input of the α axle and β axle respectively and θ = 2π f ref t delay .The delay compensation is achieved by introducing a pure imaginary number e jθ , which is a phase shift factor corresponding to the time delay in αβ coordinate. It can be implemented easily with simple complex number operation and has no much operation time consuming.
4 Simulation Results In order to verify the proposed control strategy, simulation was carried on PSIM. The schematic is shown in Fig. 2. The parameters of the inverter used in the simulation are as follows: L1 = 1.5mH R1 = 0.1Ω L2 = 900 μ H R2 = 0.1Ω C = 66 μ F f sw = 5.4kHz .
Fig. 2. Schematic of the Simulation
Improved Predictive Control of Grid-Connected PV Inverter with LCL Filter
501
Fig. 3. Simulation Result with Nominal Parameters
Simulation result of the nominal parameter and ideal U grid is shown in Fig.3. The upper waveform is U grid and the lower one is the iL1 respectively. It can be seen that the current is in phase with the voltage and has no distortion except for the current ripple. The control effect in the case of distortion and unbalance of the grid voltage are shown in Fig. 4 and Fig. 5 respectively. In Fig. 4, the grid voltage has 5th harmonic distortion. In Fig. 5, it has negative sequence component. The results show that both distortion and unbalance has no influence on iL1 .
Fig. 4. Simulation Result with Grid Voltage Distortion
Fig. 5. Simulation Result with Grid Voltage Unbalance
502
H. Xue et al.
To verify the robustness of the proposed control strategy, L1 is changed to 2.0mH from the nominal 1.5mH , which means a variation of over 33%.
Fig. 6. Simulation Result with Grid Voltage
The simulation result with the revised parameter in Fig. 6 manifests the good robustness of the proposed control strategy. Variation in other parameters has similar results.
5 Conclusion The LCL filter has higher insertion loss than the LC filter. However, control of the inverter with the LCL filter is much more difficult than that with the LC filter. An improved predictive control of the grid-connected PV inverter with LCL filter was proposed in this paper. The control strategy has explicit physical sense and is easy to understand and implement. Simulation results show that the proposed strategy has good reduction effect to grid unbalance and distortion. Moreover, the robustness to the parameter variation is satisfactory.
References 1. International Energy Agency, Solar Photovoltaic Energy Technology Roadmaps, pp. 6–10 (2010) 2. European PV Industry Association, Market Outlook for Photovoltaic until 2014 (March 2011), http://www.epia.org/ 3. Wang, T.C., Zhihong, Y., Gautam, S., Xiaoming, Y.: Output Filter Design for a Gridinterconnected Three-Phase Inverter. In: Proceeding of 34th Annual IEEE Power Electronics Specialist Conference, vol. 2, pp. 779–784 (2003) 4. Liu, F., Yan, Z., Duan, S., Ji, Y., Liu, B., Liu, F.: Parameter Design of a Two-Current-Loop Controller Used in a Grid-Connected Inverter System With LCL Filter. IEE Transactions on Industrial Electronics 56(11) (November 2009) 5. Gabe, I.J., Montagner, V.F., Pinheiro, H.: Design and Implementation of a Robusr Current Controller for VSI Connected to the Grid Through an LCL Filter. IEEE Transactions on Power Electronics 24(6) (2009) 6. Dorf, R.C., Bishop, R.H.: Modern control systems, 10th edn. Addison Wesley Longman, Menlo Park (2005)
Study on Application of Demand Prognosticating Model Based on Grey Theory Han Qingtian, Li Lian, and Zhang Yi Naval Aeronautical and Astronautical University, Yantai, China
[email protected] Abstract. Grey theory has been widely used in the predicting, estimation and assessment. In the paper, GM(1,1) model was introduced for demand prognosticating. Firstly grey theory model was studied on, including data transform, data modeling. Secondly, the prognosticating steps were given out. Finally, a spare prognosticating application example was carried out, and the prognosticating result and test are presented. The result shows that prognosticating model is more adaptive and comprehensive than other models. Keywords: grey theory; exponential smoothing; prognosticating model.
1 Introduction The gray theory treats the whole random variables all make gray a number, the transaction of the gray number is not to seek probability distribution or statistics law, it is the way that makes use of data handling to look for law between data. Carry on a transaction through the data in the logarithm row, creation new array, with this to excavation and seek number of law method, be called number of born. Number of the born mode has a variety and mainly introduce here tired apply born and tired reduce born and all be worth born. A novel grey-based modeling strategy for a dynamically turned gyroscope random drift model is studied in [1]. Fu YU-sun et. gives grey system theory, data preparation and their application [2]. And the grey model is also used for prognosticating in some combination methods [3]. In the paper, GM(1,1) model is introduced for demand prognosticating.
2 Grey Theory Model The gray model is to make use of discrete random values. Through transforming the data can follow some low. Then the pattern of the differential equation form of establishment, so it becomes easy for the change process progress search and present. 2.1 Data Transform For assure the mass of the model and the accurate result of the systematic analysis. Data transform and transaction to the raw data should be carried on for accumulating to make its cancellation measure a key link and have comparability. M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 503–508. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
504
H. Qingtian, L. Lian, and Z. Yi
Suppose there is a sequence and the data transform
x = ( x(1), x(2)," , x(n)) f :x→ y f ( x( k )) = y ( k ), k = 1, 2," , n
(1)
2.2 GM(1,1) Model
Let
array
x (0) = ( x (0) (1), x (0) (2)," , x (0) (n))
,AGO k
x (1) = ( x (1) (1), x (1) (2)," , x (1) ( n)) . Here x (1) (k ) = ∑ x (0) (i ) i =1
array
of
x (0)
is
( k = 1, 2,", n ).
Then d (k ) = x (0) (k ) = x (1) (k ) − x (1) (k − 1)
(2)
Let z (1) be the average array of x (1) , z (1) (k ) = 0.5 x (1) (k ) + 0.5 x (1) (k − 1) . So z (1) = ( z (1) (2), z (1) (3)," , z (1) ( n)) . Define the differential equation of GM(1,1) d (k ) + az (1) (k ) = b , that is x (0) ( k ) + az (1) ( k ) = b
(3)
When k = 2,3," , n , it becomes ⎧ x (0) (2) + az (1) (2) = b ⎪ (0) (1) ⎪ x (3) + az (3) = b ⎨ ⎪""" ⎪ x (0) (n) + az (1) (n) = b ⎩ ⎡ − z (1) (2) ⎢ (1) Let Y = ( x (0) (2), x (0) (3)," , x (0) (n))T , u = (a, b)T , B = ⎢ − z (3) ⎢ # ⎢ (1) ⎣⎢ − z ( n) model can be denoted as matrix equation
Y = Bu
(4)
1⎤ ⎥ 1⎥ , then GM(1,1) #⎥ ⎥ 1⎦⎥
(5)
Using LSE it can be calculate
uˆ = (aˆ , bˆ)T = ( BT B) −1 BT Y
(6)
Study on Application of Demand Prognosticating Model Based on Grey Theory
505
2.3 Prognosticating Steps 2.3.1 Data Test For assuring the feasibility of model method, it is needed to make a necessary inspection transaction to the known data sequence. Suppose original data x (0) = ( x (0) (1), x (0) (2)," , x (0) (n)) , then
λ (k ) = 2
x (0) (k − 1) , k = 2,3," , n x (0) (k )
(7)
2
If all λ (k ) are in (e− n +1 , e n +1 ) , then x (0) can be used for the prognosticating. Oth-
erwise, x (0) should be transformed. The usual data transform is parallel moving transform. For a constant c , y (0) (k ) = x (0) (k ) + c , k = 1, 2," , n . So for y (0) = ( y (0) (1), y (0) (2)," , y (0) (n)) .
λ y (k ) =
y (0) (k − 1) ∈ X , k = 2,3," , n . y (0) (k )
(8)
2.3.1 Prognosticating Model b⎞ b ⎛ xˆ (1) (k + 1) = ⎜ x (0) (1) − ⎟ e − ak + , k = 1, 2," , n − 1 a a ⎝ ⎠
(9)
And xˆ (0) ( k + 1) = xˆ (1) ( k + 1) − xˆ (1) ( k ) , k = 1, 2," , n − 1
(10)
2.3.2 Prognosticating Value Test
Let ε (k ) be the error, calculate
ε (k ) =
x (0) (k ) − xˆ (0) (k ) , k = 1, 2," , n x (0) (k )
(11)
Then the test is taken. Firstly, using x (0) (k − 1) and x (0) (k ) , λ0 (k ) can be calculated. Then the index a is used for calculate the ρ (k ) . ⎛ 1 − 0.5a ⎞ ⎟ λ0 ( k ) ⎝ 1 + 0.5a ⎠
ρ (k ) = 1 − ⎜
(12)
Lastly the prognosticating precise level can be estimated, and Table 1 can be referenced.
506
H. Qingtian, L. Lian, and Z. Yi Table 1. Model test precise level
Precise level
ε (k )
ρ (k )
Bad Ordinary Better
ε (k ) > 0.2 ε (k ) < 0.2 ε (k ) < 0.1
ρ (k ) > 0.2 ρ (k ) < 0.2 ρ (k ) < 0.1
3 Application of Demand Prognosticating Model For an enterprise, the spare consume quantity from 1998 to 2009 is shown in Table 2. The demand prognosticating is needed for the quantity of 2010. Table 2. Spare consume quantity from 1998 to 2009
Year 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009
1 season 35 34 33 38 36 35 40 37 39 42 43 40
Spare consume quantity 2 season 3 season 30 29 30 30 31 30 32 33 30 31 34 33 32 37 32 33 34 32 39 37 37 39 32 35
4 season 18 20 23 21 24 26 26 30 29 25 27 29
3.1 Ratio Test
Create the sequence of the spare consume data of spring season as follows. x (0) = ( x (0) (1), x (0) (2), x (0) (3)," , x (0) (11), x (0) (12))
= (35, 34, 33, " , 43, 40)
(13)
Firstly, λ (0) (k ) is calculated.
λ (0) = (λ (0) (2), λ (0) (3)," , λ (0) (12)) (14) = (1.029,1.030,0.868,1.056,1.029, 0.875,1.081, 0.948,0.928, 0.976,1.075) (0) Then, the value is estimated. since λ (k ) ∈ [ 0.8574, 1.1663] , k = 2,3," ,12 , so
x (0) is satisfactory model.
Study on Application of Demand Prognosticating Model Based on Grey Theory
507
3.2 GM(1,1) Model
Step 1: Accumulate the data x (0) . x (1) = (35, 69,102,140,176, 211, 251, 288,327,369, 412, 452) Step 2: Construct the array B and vector Y . ⎡ 1 (1) (1) ⎢ − 2 ( x (1) + x (2)) ⎢ ⎢ − 1 ( x (1) (2) + x (1) (3)) B=⎢ 2 ⎢ # ⎢ ⎢ 1 (1) (1) ⎢ − ( x (11) + x (12)) ⎣ 2
⎤ 1⎥ ⎡ x (0) (2) ⎤ ⎥ ⎢ (0) ⎥ 1⎥ x (3) ⎥ ⎥ ,Y = ⎢ ⎢ # ⎥ ⎥ #⎥ ⎢ (0) ⎥ ⎣⎢ x (12) ⎦⎥ ⎥ 1⎥ ⎦
(15)
Step 3: Calculate uˆ . ⎛ −0.0215 ⎞ u = (a, b)T = ( BT B) −1 BT Y = ⎜ ⎟ ⎝ 32.8503 ⎠ Then a = −0.0215 , b = 32.8503 . Step 4: Prognosticating modeling
dx (1) − 0.0215 x (1) = 32.8503 dt
(16)
(17)
And
b b x (1) (k + 1) = ( x (0) (1) − )e − ak + = 1562.9e0.0215 k − 1527.9 a a
(18)
3.3 Prognosticate Result and Test
We can get xˆ (1) (k + 1) and xˆ (0) (k + 1) . xˆ (1) (1) = xˆ (0) (1) = x (0) (1) = 35 . Let k = 1, 2,...,14 , then we have xˆ (1) (1) = 35.0000, xˆ (1) (2) = 68.9662, xˆ (1) (3) = 103.6705, xˆ (1) (4) = 139.1291 xˆ (1) (5) = 175.3583, xˆ (1) (6) = 212.3749, xˆ (1) (7) = 250.1959, xˆ (1) (8) = 288.8389 xˆ (1) (9) = 328.3217, xˆ (1) (10) = 368.6626, xˆ (1) (11) = 409.8802,xˆ (1) (12) = 451.9935 xˆ (1) (13) = 495.0221,xˆ (1) (14) = 538.9859 So, the prognosticating model is as follows xˆ (0) (k + 1) = xˆ (1) (k + 1) − xˆ (1) (k )
(19)
The prognosticating results are shown in Table 3. Model test can also be carried out based on tab. 1. as validated, the relabive error is smaller than 10%, so the precise is satisfactory.
508
H. Qingtian, L. Lian, and Z. Yi Table 3. Demand forecasting and test results Year 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
Original value 35 34 33 38 36 35 40 37 39 42 43 40 — —
Model value 35 33.9662 34.7043 35.4586 36.2292 37.0166 37.8210 38.6430 39.4828 40.3409 41.2176 42.1133 43.0286 43.9638
error 0 0.0338 -1.7043 2.5414 -0.2292 -2.0166 2.1790 -1.6430 -0.4828 1.6591 1.7824 -2.1133 — —
Relative error 0 0.09 -5.16 6.69 -0.64 -5.76 5.45 -4.44 -1.24 3.95 4.15 -5.28 — —
Precise 100 99.91 94.84 93.31 99.36 94.24 94.55 95.56 98.76 96.05 95.85 94.72 — —
References 1. Fan, C.-l., Tian, W.-f., Jin, Z.-h.: A novel grey-based modeling strategy for a dynamically turned gyroscope random drift model. Journal of Shanghai Jiaotong University 38(10), 1741–1743 (2004) 2. Fu, Y.-s., Tian, Z.-h., Shi, S.-j., et al.: Grey system theory, data preparation and their application. Journal of Shanghai Jiaotong University 35(2), 267–271 (2001) 3. Coulson, N.E., Robins, R.P.: Forecast Combination in a Dynamic Setting. Journal of Forecasting 12(1), 63–67 (1993)
Research on Bayes Reliability Assessment for Test Data Han Qingtian, Li Lian, and Cui Jia Naval Aeronautical and Astronautical University, Yantai, China
[email protected] Abstract. Inverse Weibull distribution model has been recently proposed as a model in the analysis of life testing data. According to the distribution function of the Inverse Weibull distribution, concerning characters of failure data, the uniform distribution was taken as an prior distribution of the failure probability, in order to estimate the failure probability using the Bayes method. The least squares estimates of the distribution parameters were given, and Bayes estimation of failure probability was presented, so that the reliability assessment was obtained. Keywords: reliability assessment; Bayes inference; inverse Weibull distribution.
1 Introduction In the time-censored life tests, especially in high reliability and small sampling tests. there maybe small number items failure, sometimes, there only one failure occurs. [1,2] It is also very important of the statistic analysis for this condition. In the existing literatures many familiar distributions such as Exponential, Weibull, Normal and Log Normal distributions have been studied as the life distributions of the test items. In this paper, a new life distribution, Inverse Weibull distribution, is introduced into the reliability assessment for data with only one failure.[3,4,5]. Inverse Weibull distribution model has been recently proposed as a model in the analysis of life testing data. Keller et al. derived this model on the basis of physical considerations on some failures of mechanical components subject to degradation phenomena. Erto has provided other physical failure processes leading to such distribution. Furthermore, he has shown that the Inverse Weibull distribution gave a good fit to life testing data reported in Nelson. Estimation procedures, in classical and Bayesian approaches for such distribution, are given in the literature. Calabria and Pulcini have been investigated the statistical properties of the maximum likelihood estimators (MLE’s) of the parameters and reliability for a complete sample. Erto (1989) used the Least-Square (LS) method for obtaining the estimators of the parameters and reliability. Calabria and Pulcini derived the MLE’s and the LS of the parameters. Calabria and Pulcini derived the Bayes estimator of the parameters and reliability. Finally, Calabria and Pulcini derived the prediction for some future variables. The probabilistic design of complex systems requires estimating the reliability function of any part (or component) of the whole system. The system component fails when the stress induced by the operating conditions exceeds the stress resisting capacity (strength) of the M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 509–513. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
510
H. Qingtian, L. Lian, and C. Jia
component. Both stress and strength are, generally, random variables since they depend on several operating and manufacturing variables such as temperature, size, surface finish, etc. Once distributions of these r.v.s are determined over the complete range of stress r.v. or only in the inference range, then the reliability of each component can be evaluated. In this framework, Keller and Kamath and Calabria and Pulcini proposed a statistical model, Inverse Weibull, to evaluate the reliability function of a component subjected to a given degradation phenomenon[6,7]. Distribution function of Inverse Weibull distribution
F (t ) = e
η
− ( )m t
, m > 0,η > 0
(1)
2 Test Data Analysis Supposed that the life of the product follows Inverse Weibull distribution. n products are chosen and separated into k groups, and the numbers in every group are n1 , n2 ," , nk . The start time is the same and the end time is different. Test time is recorded as t1 .t2 ," , tk , and t1 < t2 < " < tk . For the products in i group, test is ended at the predefined time ti . In the whole test, there is only one failure occurs. It is supposed that the failure occurs during (tm −1 , tm ) , ant the group has and only has one failure. Let k
sl = ∑ ni , so the situation can be denoted as tl , sl , rl , l = 1, 2," , k ,. Here sl , rl denote i =l
the number of test products and the number of failure products. Obviously s1 ≥ s2 ≥ " ≥ sk , when l ≤ m − 1 , we have rl = 0 , and when l ≤ m − 1 , we have rl = 1 . From the information above, we can know that: When t = 0 , the failure probability p0 = 0 (or approximately to 0). Because of 0 < t1 < t2 < " < tk , let pi = P (T < ti ) , then p1 < p2 < " < pk , when sk is large, pi (i = 1, 2," , k ) are all small. Under the suppose of the Inverse Weibull life distribution and the test information, we can estimate the parameters m and η and assess the reliability of the product.
3 Failure Probability Parameters Estimation 3.1 Prior Distribution of Failure Probability
Because there is only one failure occur until time tk , so we can judge that the reliability of the product is very high during time (0, tk ) . That is failure probability pk is small, especially when sk is large pk approximately to 0. So the failure probability pk ≤ 0.5 before time tk . In generally, through collect the design experience from experts, we can give the more precise upper limit λk for pk . So pk is in the area of (0, λk ) , and λk ≤ 0.5 .
Research on Bayes Reliability Assessment for Test Data
511
As the life follows Inverse Weibull distribution, then
f (t ) =
mη m − (ηt )m e t m +1
(2)
so ∂ 2 f (t ) mη m (m 2η 2 m − 3mη m (m + 1)t + (m + 1)(m + 2)t m ) − (ηt )m e = t 3( m +1) ∂t 2
(3)
∂ 2 f (t ) > 0 , distribution function F (t ) is concave function of t . ∂t 2 Because pk < λk ≤ 0.5 , so 0 < t1 < t2 < " < tk < μ , based on the quality of concave function, we know that
In some situation,
0
VOUT discontinuous VIN>VOUT continuous VIN + b] − 1 + ξ i i 2 i=1 i i=1 i
(6)
Where α i (i = 1,2,..., n ) is Lagrange multiplier. In equation (6), the partial derivatives of w, ξ i , b, α i are as follows.
Voice Recognition Based on the Theory of Transmission Wave and LSSVM n ⎧ ∂L = 0 → w = ∑ α i y i φ(x i ) ⎪ ∂w i =1 ⎪ ∂L n α i yi = 0 ⎪ =0→∑ i =1 ⎪ ∂b ⎨ ∂L ⎪ = 0 → α i = γξ i i = 1,2,..., n . ∂ ξ i ⎪ ⎪ ∂L = 0 → y [w T φ(x ) + b] + ξ − 1 = 0 i i i ⎪ ∂α ⎩ i
n
∑α y x
Due to w =
i =1
i
i
i
, ξi =
573
(7)
αi , equation (7) can be translated to matrix form γ
as follows.
y ⎢0 ⎥ ⎡ b ⎤ ⎡ 0T ⎤ (8) ⎢⎣ y Ω + γ −1I ⎥⎦ ⎢⎣α T ⎥⎦ = ⎢⎣1 ⎥⎦ . Where x = [x 1 ,..., x n ] , y = [y1 ,..., y n ] , 1 = [1,...,1] , α = [α1 ,..., α n ] , n Ω = {y y φ(x ) φ(x ) }i, j=1 , K(x , x ) = φ(x ) φ(x ) is kernel function, the RBF T
T
i
j
T
j
T
i
i
j
i
j
kernel function (seeing equation (7)) is selected in this paper.
⎛ xi − x j K(x i , x j ) = exp⎜ − 2 ⎜ 2σ ⎝
2
⎞ ⎟ ⎟. ⎠
(9)
4 Experimental Results In this experiment, glottis excitation signal of sonant is simulated by periodic triangle pulse sequence. The sound samples are extracted from a Chinese male. Since the length of sound channel is about 17cm and the acoustic velocity is 360m/s, then the propagation time of voice in channel is about 0.5ms. The sampling rate of voice is 11.025 kHz and quantification precision is 12 bit. So there are 5 sampling points in 0.5ms. The experiment samples are “1,2,3,4,5,6,7,8,9,10” said in Chinese. 40 coefficients extracted from each sound are taken as classified data. If the number of coefficient is less than 40, the coefficient should be extended symmetrically. Fig.4 shows the transmission coefficient of sonant “1”. Fig.5 shows the transmission coefficient of sonant “7”. Fig. 6 gives the transmission coefficient of sonant “10”. It is clear that
574
Y. Kong, Y. Min, and J. Xiao
there have obvious differences between the transmission coefficients of sonant “10” and “1” or “7”. Though the transmission coefficients of “1” and “7” are similar, they still have tiny differences.
Fig. 4. Transmission coefficient of sonant “1”.
Fig. 5. Transmission coefficient of sonant “7”.
Fig. 6. Transmission coefficient of sonant “10”.
After extracting transmission coefficients, LSSVM is used to classify the coefficients. γ and σ are two kernel parameters which are very important for LSSVM. The kernel parameters are selected by 5-fold cross-validation method in the paper. In 5fold cross-validation the data is first partitioned into 5 equally (or nearly equally) sized segments or folds. Subsequently 5 iterations of training and validation are performed such that within each iteration a different fold of the data is held-out for validation while the remaining 4 folds are used for learning. The generalization error is estimated according to the average value of mean squared error (MSE) of 5 iterations. Finally, the optimal parameters are selected. If surd and sonant are both considered, the recognition rate of BP neural network is 84.5%, while that of LSSVM can reach to 87.8%, as shown in table 1. If only sonant is considered, the recognition rate of BP neural network is 73%, and that of LSSVM reaches to 77%, as shown in table 2.
Voice Recognition Based on the Theory of Transmission Wave and LSSVM
575
Table 1. Recognition rate of “1” to “10” considering both surd and sonant /6690 %3
Table 2. Recognition rate of “1 ,7 ,10” and “ 6 ,9” only considering sonant /6690 %3
5 Conclusion This paper mainly discussed the classification of voice parameters extracted from one-dimensional transmission wave based on the method of LSSVM, compared the identification results with that of BP neural network algorithm. LSSVM doesn’t need to determine the number of hidden nodes which is a hard problem for neural network. LSSVM avoids the local minimum trap and can achieve global optimization. LSSVM has better generalization effect. The experiment demonstrated that the recognition rate of LSSVM is better than BP neural network. LSSVM has better research value for the classification of voice. Making full use of the merits of LSSVM and neural network and combining the two methods to further improve recognition rate is the research direction in future.
References 1. 2. 3. 4. 5.
Mallat, S.: A Wavelet Tour of Signal Processing. China Machine Press, Beijing (2002) Rabiner, L.R.: Fundamentals of Speech Recognition. Prentice Hall, Inc., Englewood Cliffs (1999) Huang, G.: Seismic Reflection Wave Imaging New Method and Its Theoretical Basis. J. Scientia Sinica(D) 30(6), 650–655 (2000) Kong, Y., Huang, G., Ning, F.: Pilot Study of Extracting Voice Feature Parameters by Transmitted Wave. Journal of Shandong University(Natural Science) 39(2), 56–61 (2004) Rabiner, L.R., Schafer, R.W.: Digital Processing of Speech Signals. Prentice Hall, Inc., Englewood Cliffs (1978)
Design of a Two-Phase Adiabatic Content-Addressable Memory Meng-Chou Chang and Yen-Ting Kuo National Changhua University of Education, Department of Electronic Engineering, No.2, Shida Rd., Changhua 50074, Taiwan, R.O.C.
[email protected] Abstract. This paper presents the design of a two-phase adiabatic CAM, which achieves low-power dissipation by employing adiabatic operation and two-step data matching. The proposed adiabatic CAM can recycle the charge on match lines and keep the voltage drop between the power clocks and the output nodes close to zero during the charging/discharging process, leading to lower power dissipation. Also, the match-line in each CAM word is partitioned into two segments, and the second segment is selectively charged/discharged according to the match result of the first segment. If the match result of the first word segment is mismatch, the charging/discharging of the second segment of the match-line will be eliminated, further reducing the power dissipation. Simulation results show that the proposed two-phase adiabatic CAM with 64 words×144 bits can achieve a power reduction of 16.9% compared to the traditional single-phase adiabatic CAM at an operating frequency of 500 MHz. Keywords: Content-addressable memory (CAM), adiabatic logic, adiabatic CAM.
1 Introduction Content-addressable memory (CAM) compares input search data in parallel against a table of stored data, and returns the address of the matching data. CAMs can be used in a wide variety of applications requiring high-speed parallel search. These applications include pattern recognition, data compression, and network address translation [1]. However, the parallel search operations in CAMs cause high power consumption due to frequent switching of highly capacitive match lines (MLs). Traditional CAM power reduction techniques include low swing schemes, selective precharge schemes, and precharge low schemes. Recently, adiabatic switching principle has also been applied to the design of lowpower CAMs [2-6]. An adiabatic logic uses AC power sources instead of traditional DC sources to charge and discharge its output nodes. When an adiabatic logic charges/discharges its output nodes, the switch elements on the charging/discharging path consume almost zero power by keeping the voltage drop between the AC power M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 577–583. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
578
M.-C. Chang and Y.-T. Kuo
sources and the output nodes close to zero. In [6], Q. Xu et al. had implemented an adiabatic CAM using CPAL (Complementary Pass-transistor Adiabatic Logic) [7] and showed that the adiabatic CAM can achieve 86% energy saving compared to the conventional static CMOS implementation at 100 MHz. In this paper, we propose a two-phase adiabatic CAM that employs both the adiabatic switching techniques and match-line segmentation to reduce the energy consumption. In the proposed architecture, the match line in each CAM word is partitioned into two segments, and the second match-line segment is activated only if the corresponding first match-line is matched.
2 The Proposed Two-Phase Adiabatic CAM Figure 1 shows the block diagram of the proposed two-phase adiabatic CAM with a storage array of 64 words×144 bits. Each word in the CAM array is partitioned into two segments, the first word segment with 6 bits and the second word segment with 138 bits. That is, the CAM array is organized as two sub-arrays, the first sub-array with 64 words×6 bits and the second sub-array with 64 words×138 bits. As used in the adiabatic CAM proposed by Q. Xu et al. [6], CPAL (Complementary Pass-transistor Adiabatic Logic) [7] is also used to implement the circuits in our two-phase adiabatic CAM. Figure 2(a) shows the structure of the CPAL buffer/inverter. A CPAL gate is supplied by a single phase power clock, denoted by PC in Fig. 2(a). Cascaded CPAL gates, as shown in Fig. 2(b), can be driven by fourphase power-clocks, PC1, PC2, PC3, and PC4. The power clocks can be sinusoidal or trapezoidal, and there is a 90° phase lag between adjacent power clocks, as depicted in Fig. 2(c). D144
D7 D1
PC2
D6
bit-line drivers
PC1
PC2 PC3
bit-line drivers
PC1
PC2 PC3
PC1 BL1
BLb1
BL6
BL7
BLb6
BLb7
BL144
PC4
BLb144
SWL1
FWL1 Cell
Cell
FMDL1
Cell
FML1 ML recovery
Cell
SML1
Match
SMDL1
Sensor
ML recovery
Match Result Merge
Out1
Match Result Merge
Out64
SWL64
FWL64
Cell
Cell
Cell
FML64 FMDL64
ML recovery
1st storage sub-array
Cell
Match Sensor
SMDL64
2nd storage sub-array
Fig. 1. The proposed two-phase adiabatic CAM.
SML64 ML recovery
Design of a Two-Phase Adiabatic Content-Addressable Memory
PC1
PC D
N1
P2
P1
Db
PC4
OUT
IN
N2
PC3
PC2
579
INb
OUTb
(b) OUTb
OUT
t1
t2
t3
t4
t5
t6 t7
t8
N4
N3
PC1 PC2 IN
N5 INb
INb N8
N6 N7
IN
INb
IN
IN
PC3 PC4
INb
(c)
(a)
Fig. 2. (a) A CPAL inverter/buffer, (b) Cascaded CPAL buffers, (c) Waveforms of power clocks.
The operation of a CPAL gate is composed of four phases: wait, evaluate, hold, and recovery. During the wait phase, the power clock keeps low, and both the outputs of the CPAL gate remain low. During the evaluate phase, the power clock rises from 0V to the peak voltage, and one of the gate outputs is charged to high according to the input values. During the hold phase, the power clock remains high, and the gate outputs hold their values. During the recovery phase, the power clock falls from the peak voltage to 0V, and the charge on the charged output node is recycled to the power clock. Figure 3 (a) shows the structure of the CAM storage cell implemented with CPAL. If the stored bit in the CAM cell matches the search bit, the transistor N5 is turned off and the CAM cell does not provide a conducting path between the match-driving line (MDL) and the match line (ML); if the stored bit in the CAM cell mismatches the search bit, the transistor N5 is turned on and the CAM cell provides a conducting path between MDL and ML. That is, if the stored data in a CAM word mismatches the search word, there is at least a conducting path between the corresponding MDL and ML. The MDL is connected to a power clock, and thus the corresponding ML will be
BL
BLb
WL
PC VDD
N2
N1
MR
ML N3
X
N4 N5
ML
MDL
MDL (a)
(b)
Fig. 3. (a) CAM storage cell, (b) the match-line recovery circuit.
580
M.-C. Chang and Y.-T. Kuo
PC3
PC2 N3 P1
N5
N7 P3
N4
N6 P2
SMDL N9
FML N1
N2
N8
Fig. 4. The circuit of the match sensor block.
PC4 SMDL P1
PC3
Out
SML N1
N2
N3
PC2
N4
PC1
Fig. 5. The circuit of the match result merge block.
driven to high by the power clock during the hold phase if the corresponding stored data mismatches the search data. On the contrary, the corresponding ML will be kept low during the hold phase if the corresponding stored data matches the search data. Figure 3(b) shows the match-line recovery circuit, which is used to provide a conducting path from ML to MDL during the recovery phase to recycle the charge on the match line. The CAM storage cell and the match-line recovery circuit used in our two-phase adiabatic CAM are the same as those used in [6]. Data matching in our two-phase adiabatic CAM is a two-step process. In the first step, the first 6-bit of the search data is compared with each CAM word of the first storage sub-array. Only those matching CAM words in the first step will activate the second-step data matching of the corresponding CAM words in the second storage sub-array. As shown in Fig. 1, in the proposed two-phase adiabatic CAM, each match line is composed of FML (first-stage match line) and SML (second-stage match line). Also, each match driving line is composed of FMDL (first-stage match driving line), driven by PC2, and SMDL (second-stage match driving line), driven by PC3. The match sensor block in Fig. 1 detects the match result on FML and determines whether power clock PC3 should drive SMDL. The circuit of the match sensor block is shown in Fig. 4. If the match result on FML is match, the voltage of FML will be low during the hold phase of PC2, resulting in a conducting path between PC3 and the corresponding SMDL, and thus the corresponding CAM word in the second storage subarray can participate the second-step data matching. On the contrary, if the match result on FML is mismatch, the voltage of FML will be high during the hold phase of PC2, blocking the conducting path between PC3 and the corresponding SMDL, and
Design of a Two-Phase Adiabatic Content-Addressable Memory
581
thus the corresponding CAM word in the second storage sub-array cannot participate the second-step data matching. The match result merge block in Fig. 1 detects the voltages on SMDL and SML to determine the final match result for the corresponding CAM word. Table 1 shows the three possible cases of data matching in the proposed two-phase adiabatic CAM. In case 1, the result of first-step matching is mismatch, and the second-step matching for the corresponding CAM word is cancelled. In this case, FML is high during the hold phase of PC2, and both SMDL and SML keep low during the hold phase of PC3. In case 2, the result of first-step matching is match, and the result of the second-step matching for the CAM word is mismatch. In this case, FML is low during the hold phase of PC2, and both SMDL and SML keep high during the hold phase of PC3. In case 3, the result of first-step matching is match, and the result of the second-step matching for the CAM word is also match. In this case, FML is low during the hold phase of PC2, and SMDL is high and SML is low during the hold phase of PC3. Figure 5 shows the circuit of match result merge block, which samples the voltage on SMDL and SML during the hold phase of PC3 and generates a valid match result Out during the hold phase of PC4. The match result Out goes high during the hold phase of PC4 only if SMDL is high and SML is low during the hold phase of PC3.
3 Simulation Results The Hspice simulations for the proposed two-phase adiabatic CAM with 64 words×144 bits were performed using BPTM (Berkeley Predictive Transistor Model) 45nm technology with a Vdd of 1.0V. Figure 6 shows the waveforms of the proposed two-phase adiabatic CAM for the three cases of data matching described in Table 1. In Fig. 6, trapezoidal power clocks with a frequency of 500 MHz were used for the adiabatic CAM. For comparison, the Hspice simulations for the traditional single-phase adiabatic CAM proposed by Q. Xu et al. [6] were also performed using the same CAM size and spice parameters. Table 2 lists the power consumption comparisons for the traditional single-phase adiabatic CAM and the proposed two-phase adiabatic CAM. Both CAMs use CPAL gates and the same structure of CAM storage cells. The difference between them is that data matching in the proposed CAM is a two-step process and most of the second-step comparison are never activated due to mismatch in the first step. From Table 2, it can be seen that the proposed two-phase adiabatic CAM can achieve a power reduction of 16.9% compared to the traditional single-phase adiabatic CAM when the frequency of power clocks is 500 MHz. Table 1. Three cases of data matching in the proposed two-phase adiabatic CAM.
Case 1 Case 2 Case 3
First-step comparison mismatch match match
Second-step comparison inactive mismatch match
FML
SMDL
SML
high low low
low high high
low high low
Match result mismatch mismatch match
582
M.-C. Chang and Y.-T. Kuo Table 2. Power consumption comparisons for two adiabatic CAMs.
Frequency of power clocks 100 MHz 200 MHz 300 MHz 400 MHz 500 MHz
Traditional singlephase adiabatic CAM [6] (64 words×144 bits) 125.7 uW 153.9 uW 189.4 uW 222.7 uW 262.0 uW
Case 3
The proposed twophase adiabatic CAM (64 words×144 bits) 114.7 uW 133.1 uW 159.0 uW 184.2 uW 217.5 uW
Case 2
Power saving 8.7% 14.2% 16.1% 18.8% 16.9%
Case 1
Fig. 6. Waveforms for the proposed two-phase adiabatic CAM.
4 Conclusions We have presented the design of the proposed two-phase adiabatic CAM. Adiabatic CAMs can recycle the charge on match lines and keep the voltage drop between the power clocks and the output nodes close to zero, leading to lower power dissipation. In our two-phase adiabatic CAM, the match line is further segmented into two segments, FML (first-stage match line) and SML (second-stage match line), and two-step data matching are employed. If the result of the first-step comparison is mismatch, the second-step comparison will be cancelled to further reduce energy dissipation. The circuit of the match sensor has been designed to detect the result on FML and to control the driving of SML. Simulation results have shown that the proposed two-phase adiabatic CAM can achieve a power saving of 16.9% compared to the traditional single-phase adiabatic CAM at an operating frequency of 500 MHz.
Design of a Two-Phase Adiabatic Content-Addressable Memory
583
References 1. Pagiamtzis, K., Sheikholeslami, A.: Content-Addressable Memory (CAM) Circuits and Architectures: A Tutorial and Survey. IEEE J. Solid-State Circuits 41, 712–727 (2006) 2. Natarajan, A., Jasinski, D., Burleson, W., Tessier, R.: A Hybrid Adiabatic Content Addressable Memory for Ultra Low-Power Applications. In: 2003 Great Lakes Symposium on VLSI, pp. 72–75 (2003) 3. Bala, G.J., Perinbam, J.R.P.: A Novel Low Power 16x16 Content Addressable Memory Using PAL. In: 18th IEEE International Conference on VLSI Design, pp. 791–794 (2005) 4. Wu, Y., Hu, J.: Low-Power Content Addressable Memory Using 2N-2N2P Circuits. In: IEEE 2006 International Conference on Communications, Circuits, and Systems, pp. 2657–2661 (2006) 5. Zhang, S., Hu, J., Zhou, D.: A low-Power Adiabatic Content-Addressable Memory. In: IEEE 2007 International Midwest Symposium on Circuits and Systems, pp. 1285–1288 (2007) 6. Xu, Q., Ye, L., Hu, J., Huang, L.: The Implementation of Low-Power CAM with Fully Adiabatic Driving for Large Node Capacitances. In: 2009 World Congress on Computer Science and Information Engineering, pp. 413–417 (2009) 7. Hu, J., Xu, T., Li, H.: A Low-Power Register File Based on Complementary Pass-Transistor Adiabatic Logic. IEICE Trans. Inf. & Syst., E88-D (7), 1479–1485 (2005)
Branch Importance Assessment under Cut-Off Power Flow Based on EM Clustering Feng Yuan, Wang Li-ming, Xia Li, Bu Le-ping, and Shao Ying College of Electric and Information Engineering, Naval University of Engineering, Wuhan, Hubei, 430033
[email protected] Abstract. Assessment on branch importance under cut-off power flow helps to analyze the size of system disturbances caused by breaking of different branches. This article has proposed four factors which affect the branch importance based on cut-off power flow. After a branch importance assessment form is formed, according to the actual characteristics of the data sheet, this article has put forward an attribute weight assessment algorithm based on EM clustering, and has verified it with an example of a 38-node power system. It can be found from the weight vector of an attribute that branch importance is mainly determined by DifV, DifGenPower and DifBrPower. Through the comparison of the original score sheet and the assessment form, we can find that the ultimate branch importance sequencing is consistent with the above conclusion, which thus has proved the correctness of the algorithm. Keywords: power system, cut-off power flow, branch importance assessment, data mining, EM algorithm.
1 Introduction In power system operation, we may often encounter a variety of disturbances, and component failures may also occur, so as to lead to out of operation for some components. Studying the distribution of power flow in the system under a branch breaking, and inspecting whether a branch breaking may cause overload of other branches in the system, or out of range for the voltages of some load buses, will help to guide operators to take timely measures to eliminate overload or out of range[1-4]. Besides, research on changes in system parameters under a branch breaking will also help to analyze the size of system disturbances caused by breaking of different branches. Therefore, it is of great significance to study branch importance assessment under cutoff power flow in the power system.
2 The Basic Idea of the Algorithm Suppose
the
A = {a1 , a2
attribute
set
of
the
multi-attribute
am } ; the instance set is L = {l1 , l2 ,
assessment
form
is
ln } ; sik is the assessed
M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 585–592. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
586
F. Yuan et al.
value under the attribute of
ak which corresponds to the instance of li . Then, each
element in L can generate m assessed values under the corresponding A, which constitute the assessed value matrix S as shown in Table 1. Suppose the weight vector corresponding to each attribute is vector of the instance set is
W = [ w1 , w2
D = [ d1 , d 2
D = [ d1 , d 2
wm ]T , and the final decision
d n ] . Then
dn ]
= S iW
(1)
⎛ a11 … a1n ⎞ ⎛ w1 ⎞ ⎟⎜ ⎟ = ⎜⎜ ⎟⎜ ⎟ ⎜a amn ⎟⎠ ⎜⎝ wn ⎟⎠ ⎝ m1 It is very easy to understand the above assessment criteria. That is, for a multiobjective (or multi-attribute, multi-factor) decision problem, we just need to assess a single objective of each instance, and give different weights to different objectives. At last, through the use of Formula (1), the final decision (assessment) results can be obtained. A variety of multi-objective optimization problems are all using this method or a variant of this method to establish assessment criteria. The criteria are very simple, but the difficulty lies in how to give weight to each attribute. With the method of analytic hierarchy process, decision and scoring can be made according to expert experiences, so as to obtain the weight vector of the attribute set. However, this method can only rely on expert experiences, so it is inevitably subject to personal preferences, and the conclusion may be somewhat biased. This article proposes a method that using the clustering algorithm to cluster such data like in Table 1, and obtain the weight vector of the attribute set. Assume that we have obtained the decision vector D of Table 1, and add it into Table 1 as the decision attribute. Then, we can use the EM clustering algorithm. The clustering results have nothing to do with the weight determining the assessment, but the performance evaluation indicators of the clustering results, such as log-likelihood, have objectively shown the clustering effects. The greater the log-likelihood, the better is the clustering effect. If we remove one attribute, the performance of EM clustering will definitely be affected, which is reflected in changes in log-likelihood. Thus, the importance of the attribute is beyond no doubt related to the increment of loglikelihood after this attribute is eliminated. The larger the increment, the more important the attribute, and the greater the corresponding weight value; vice versa. xm } through the Assume that the class obtained from an instance set {x1 , x2 , EM algorithm is
{C1 , C2 ,
Cn } , and the definition of its log-likelihood is as
follows:
⎛ m n ⎞ log-likelihood = log 2 ⎜ ∏ ∑ P(C j ) P( xi | C j ) ⎟ ⎝ i =1 j =1 ⎠
(2)
Branch Importance Assessment under Cut-Off Power Flow Based on EM Clustering
Among them,
587
P (C j ) represents the probability of the class C j , and P ( xi | C j )
represents the probability of the instance
xi that belongs to the class C j .
Therefore, attribute weights can be determined by EM clustering, so as to assess branch importance of Table 1. The flow chart of the algorithm is shown in Figure 1.
Fig. 1. The flow chart of the clustering-based attribute weight assessment algorithm
Among the weight vectors obtained based on the above method, each weight value determines the importance of each attribute. Therefore, in the attribute set A, the length of each attribute interval must be consistent. It is not feasible to directly put the assessed value matrix in Table 1 into Formula (1), so we must normalize each column vector in Table 1 before calling Formula (1).
3 Generation and Pre-treatment of the Branch Importance Assessment Form In the process of power system running, when a branch breaking occurs, the system changes caused can be measured by the following four indicators: (1) DifV which represents the degree of changes in voltage magnitude of each node (bus). Assume that the node set of a power system is {n1 , n2 , nm } , and the voltage magnitude vector of the node at
t0 is
V (t 0 ) = ( v1 (t 0 ), v2 (t 0 ),
vm (t 0 ) ) . Suppose
the voltage magnitude vector of the node after a branch breaking occurs at
t1 is
588
F. Yuan et al.
V (t1 ) = ( v1 (t1 ), v2 (t1 ),
vm (t1 ) ) , and then DifV that is the degree of changes
in voltage magnitude of each node can be defined by the following formula:
DifV =
sum
(( V
(t1 ) − V (t 0 ) )
2
)
=
m
∑ ( v (t ) − v (t ) ) i
i
1
2
0
(3)
i =1
Similarly, the other three indicators can be defined as: (2) DifAngle which represents the change in voltage phase difference at both ends of each branch:
(
DifAngle = sum ( d Θ(t1 ) − d Θ(t0 ) )
2
) ∑( dθ (t ) − dθ (t )) =
s
i
1
2
0
i
(4)
i =1
(3) DifBrPower which represents the change in power flow of each branch:
(
DifBrPower = sum ( dP(t1 ) − dP(t0 ) ) + ( dQ(t1 ) − dQ(t0 ) ) 2
s
∑( dp (t ) − dp (t )) + ( dq (t ) − dq (t ))
=
2
i
1
0
i
1
i
i
2
) (5)
2
0
i =1
(4) DifGenPower which represents the change in throughput of each electric generator:
(
DifGenPower = sum ( dPG(t1 ) − dPG(t0 )) + ( dQG(t1) − dQG(t0 )) =
2
r
∑( dpg (t ) − dpg (t )) + ( dqg (t ) − dqg (t )) 2
i
1
i
0
i
1
i =1
Fig. 2. Network topology of a power system
i
0
2
) (6)
2
Branch Importance Assessment under Cut-Off Power Flow Based on EM Clustering
589
The above four indicators can be obtained by changes in cut-off power flow before and after a branch breaking occurs. Taking a 38-node power system in Figure 2 as an example, we verify the algorithm in Section 1. In order to assess the importance of each branch in the power system as shown in Figure 2, we cut off each branch once, and then calculate the values of the above four indicators of this branch. In this way, the assessment form formed has 37 instances, each standing for a branch. The attribute set of instances is the set of the above four indicators, namely DifV DifAngle DifBrPower DifGenPower , thus, the branch importance assessment form formed at last is shown in Table 1. Discretization schemes of DifV DifAngle DifBrPower and DifGenPower are shown in Table 2-5.
{
,
,
,
,
,
}
Table 1. Branch importance assessment form of the 38-node example No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37
DifV 0.00340432600125688 8.35695074179583E-5 5.79922563950256E-6 4.52266483535953E-7 4.70369428579647E-6 1.63826242767072E-6 8.4937409851007E-4 0.00127406114776511 0.00397347345954063 0.00524554884118672 0.00581685269274162 5.25050784172493E-4 2.14962039715691E-6 7.54057324361026E-6 2.29661214432493E-5 0.00544167874110104 0.00581215174258109 0.00608113567356488 0.00528995353171892 0.00340040174878768 8.34222458096973E-5 5.80113727376007E-6 4.54207889779117E-7 4.66150168089895E-6 1.57755124248942E-6 4.39409655938969E-4 0.00425203376119709 0.00535237865824362 0.00500876843243892 5.2388962991251E-4 2.23104061824839E-6 7.17878215082896E-6 2.31425659015219E-5 0.0051846872567289 0.00511754748673742 0.00574470018280519 0.00608058093239142
DifAngle 0.225061204314816 0.00175500756847545 1.21788255305471E-4 9.48868580944875E-6 1.05221282594555E-4 3.918663462142E-5 0.00724489758680611 0.0160705565193942 0.2849713667938 0.255524998164968 0.247852380591799 0.00597767204820271 3.32611539859744E-4 0.00132706353555697 0.00422876127676194 0.274829966591299 0.241066856607496 0.279314886385203 0.290905053314283 0.230132233688245 0.244469622137357 0.0169930520004376 0.00134492582399308 0.0134701961779425 0.00527617707997256 0.0133921304328285 0.290105456338376 0.254233157641671 0.246636267768716 0.0514066542511433 0.00145445802603794 0.00560294186581023 0.0182117928423102 0.235708773084351 0.279479943427911 0.245544129957989 0.265668195528536
DifBrPower 0.182368646034108 0.00176621523158616 2.09065388086655E-4 1.67574692252627E-5 1.81867998524467E-4 6.88491458626745E-5 8.8298148461953E-4 0.00141277037539125 0.0306387974169379 0.0087547183967317 0.00892276050498212 0.00125494564166862 2.86774184654777E-5 1.29354798877109E-4 5.53700607705125E-4 0.00935134841663105 0.00907613031447038 0.0373839870242216 0.036358974806495 0.181524120640651 0.196050928331789 0.00124950765565013 3.33934117103194E-5 8.41713237232106E-4 1.93844728179513E-4 3.85213613052834E-4 0.0308657343411399 0.00844027082940741 0.00819027052682039 0.00751048775465395 4.33707853158693E-5 2.32243852117149E-4 0.0014421898896895 0.00826738179897265 0.00882394266644837 0.0345394717320888 0.0368089637468814
DifGenPower 81.9581968284197 44.0992484601377 3.06015258092036 0.238419053323693 2.64663900361579 0.984654611149533 2.33169762488723 3.81472368639318 0.530793499162651 0.487844307193314 0.5011026542472 11.4781453362922 0.34985442882718 1.42652531449544 4.46068375514486 0.739382602508517 0.69198848196906 0.761174785552933 0.751143181512689 82.5066180433843 69.1421012136608 4.80700796346065 0.378141330327426 4.00965511443702 1.50034814197802 2.50087120345261 0.861611087799057 0.768432301551797 0.736458398036631 15.1739420068954 0.379329253476743 1.45037847231635 4.72610557633655 0.685668552534714 0.769166073278181 0.716357544597963 0.769467493329999
590
F. Yuan et al. Table 2. DifV attribute discretization scheme
before discretization after discretization
DifV≤1
1 Decision=U1 (CF = 0.83) (PL2 = HIGH) and (V1 = HIGH) => Decision=U1 (CF = 0.66) It can be included that the result of FURIA is a little different from that of Prism.
6 Results and Performance of J48 The results and evaluation of J48 are shown in Tab.8-10.
Knowledge Discovery of Energy Management System
599
Table 8. Overview of Evaluation of J48 Correctly Classified Instances Incorrectly Classified Instances Kappa Statistic Mean Absolute Error Root Mean Squared Error Relative Absolute Error Root Relative Squared Error Coverage of Cases(0.95 level) Mean Rel. Region Size(0.95 level)
84% 16% 0.7585 0.1029 0.3036 23.2134% 64.2903% 92% 37.3333%
Table 9. Detailed Accuracy by Class of J48 Class S U2 U1 Weighted Average
TP Rate 1 0.857 0.625 0.84
FP Rate 0 0.167 0.059 0.065
Precision 1 0.667 0.833 0.853
Recall 1 0.857 0.625 0.84
F-Measure 1 0.75 0.714 0.839
ROC Area 1 0.877 0.886 0.929
Table 10. Confusion Matrix of J48 Class S U2 U1
S 10 0 0
U1 0 6 3
U2 0 1 5
J48 can generate decision tree when the algorithm is finished. Thus the corresponding decision tree is shown as Fig.2.
Fig. 2. Decision Tree Generated by J48
In Fig.2, the non-leaf nodes of decision tree present conditional attribute, the leaf nodes present class value, numbers in the bracket after the class value present the number of supported instances. The branches present conditional attribute value.
600
F. Yuan et al.
7 Conclusion No matter whether table 1 is real power system state parameter set or simulation data set, the results of classifier in section 3-5 make quite guiding significance on EMS. If the if-then rules generated by Prism and FURIA, or the decision trees generated by J48 can form a intelligent module of EMS by way of an expertise database, it can not only provide decision support and reference for EMS to improve response performance of EMS, but also apply in 3D power system simulation training. When certain characteristic of power system is hard to be generated by real-time simulation, the expertise rules generated by offline history data or assumptive contingency data can be helpful to generate the responding characteristic. Therefore, it can provide another feasible solution for power system simulation training.
References 1. Madan, S., Son, W.K., Bollinger, K.E.: Applications of Data Mining for Power Systems. In: Proceedings of 1997 IEEE Canadian Conference on Electrical and Computer Engineering (CCECE 1997), Canada, pp. 403–406 (1997) 2. Ordieres Mere, J., Ortega, F., Bello, A., et al.: Operational Information System in a Power Plant. In: Proeeedings of the IEEE International Conference on Systems, Man and Cybernetics, Computational Cybernetics and Simulation, Orlando, USA, pp. 3285–3288 (1997) 3. Steele, J.A., McDonald, J.R., D’Arcy, C.: Knowledge Discovery in Databases: Applications in the Electrical Power Engineering Domain. IEE Colloquium(Digest.) 340, 33–38 (1997) 4. Lambert-Torres, G.: Application of Rough Sets in Power System Control Center Data Mining. IEEE Transaction on Power Delivery 17(3), 1368–1373 (2002) 5. Witten, I.H., Frank, E.: Data Mining: Practical Machine Learning Tools and Techniques, 2nd edn., pp. 1–324. Morgan Kaufmann Publishers, Amsterdam (2005) 6. Cendrowska, J.: PRISM An algorithm for inducing modular rules. Man-Machine Studies 27(2), 349–370 (1987) 7. Huehn, J., Huellermeier, E.: FURIA: An Algorithm for Unordered Fuzzy Rule Induction. Data Mining and Knowledge Discovery 47(19), 293–319 (2009) 8. Quinlan, R.: C4.5: Programs for Machine Learning, pp. 1–100. Morgan Kaufmann, San Francisco (1993)
Study on the Medium-Term and Long-Term Forecast Technology of Wind Farm Power Generation Yang Gao1, Li Liu1, Guoyan Liang1, Shihai Ma1, and Chenwei Tian2 1 Shenyang Institute of Engineering, Shenyang 110136, Liaoning Province, China 2 Beijing Institute of Technology, Haidian District, Beijing 100081, China
Abstract. In view of analyzing forecast method of long-term forecast of generating capacity of wind power, this paper puts forward the long-term forecast model of wind power generation, which is based on the grey systems theory. Considering that wind power generation depends on wind speed, first we may establish wind - power function through the existing wind speed and wind power of the pre-installed wind turbine, resulting in years of wind power generation data. And then, a grey information renewal GM (1 ,1) model for predicting calculated capacity of annual wind power generation is established by using the calculated annual wind power in this paper. Moreover, the model is applied to predict the calculated capacity of annual output of a wind turbine generator system in FuJin wind farm. Keywords: Grey Forecast; Grey Model; Forecast of Wind Power Generation; Mean Absolute Error.
1 Introduction In recent years, wind power generation has made an immensely rapid headway. Developing wind power, a king of energy without carbon, is of great significance to carrying out a strategic plan of energy in our country, ensuring Energy Safety, developing low-carbon economy and dealing with climate change. In order to make the invest benefit evaluation on the large-scale wind farm, the feasibility study on the design of a wind power plant, the choice of the address, the arrangement of overhaul and the management examination, the long-term forecast of the generating capacity of wind power will be needed. In this way (one month to one year), we can provide the evidence for the management of wind farm. By predicting the generating capacity of wind power in a year, we may work out the assessing method on the production of the wind farm and meanwhile make an effective evaluation about the effect after the establishing of the farm. By forecasting the generating capacity of wind power (one month to a quarter), we can also supply the evidence for the arrangement of the overhaul and maintenance the equipment in the wind farm. In a word, the medium-term and long-term forecast of wind power generation is one of the problems which needs to be solved immediately in the operation management nowadays[1,2]. M. Ma (Ed.): Communication Systems and Information Technology, LNEE 100, pp. 601–607. springerlink.com © Springer-Verlag Berlin Heidelberg 2011
602
Y. Gao et al.
2 The Medium-Term and Long-Term Forecast Method of Wind Power Generation Capacity The cycle of long-term forecast of wind power generation needs a period of time which is as long as more than one year, but the numerical weather forecast can not make the forecast in such long period of time[3-4]. To solve the problem of the long-term forecast of wind power generation, this paper brings forward the long-term forecast model of wind power generation which based on the Rough Set Theory. The wind-driven generating unit mostly adopts the maximum power point tracking (MPPT) control method so as to obtain the clean energy resources farthest. Therefore, there are direct and certain relation to the generating capacity the wind-driven generating unit and speed. To get the long-term capacity forecast of wind power generation, the best way is to make the long-term forecast on average wind speed and then we can get the long term generating capacity forecast of wind power by making use of the fundamental characteristics curve of wind-mill generator. The common medium-term and long-term predicting methods are Support Vector Machine, Neural Networks, Time Series Analysis, Grey Forecast etc, of which Grey Forecast is the basic thinking and method based on the grey systems theory[5-6]. The subject investigated of Grey Systems Theory is the grey system of which part of the information is known, while the other is unknown. The factors influencing the change of generating capacity of wind power are uncertain, and it seems that they are irregular, so the system of wind power generation can be regarded as a grey system. Because the grey forecast model has the advantages of having less historical data needed, simple model, high precision of forecast, easily calculating, no need to take the distributing order etc., thus it has got the extensive application. The routine greater precise data of forecast model GM 1,1 are only the latest several data, and in order to cover the shortage, this paper introduces a Metabolic Grey Model –GM (1, 1) which makes forecast and verification of the annual generating capacity of some wind-driven generating units in Fujin wind farm.
( )
3 Metabolic Grey Model –GM (1,1) 3.1 The Principles Of Metabolic Grey Model –GM (1,1) The Grey Systems Theory accumulates the seemingly irregular historical data serial 1-AGO , this make it have the exponential growth feature. Because the solutions of differential equation has the feature of exponential growth Consequently, by using the differential equation model, it is quite natural to get the regular historical data serial with the growth of exponential growth after accumulating the data serial (1-IAGO). Making use of the solutions of differential equation, we can easily make forecast on the regular historical data serial of the generating capacity of wind power and then make the reverse forecast value, that is, we can get the actual capacity forecast value by descending the regular historical data serial (1-IAGO). The most common model of grey forecast is GM 1,1 model, but as a medium-term and long-term forecast method, it has less precision. For the sake of high precision, this paper introduces Metabolic Grey Model –GM (1,1). On the one hand the model keeps the advantages of the routine
(
)
,
( )
Study on the Medium-Term and Long-Term Forecast Technology
603
GM (1,1) model, on the other hand it can also take the disturbance factors coming into the system into account promptly. To use the Metabolic Grey Model to make forecast is not to use keep forecast with one model, it is to predict a value according to the model GM(1,1) established on the basis of the known serials, and then add the value to the known serials, meanwhile delete the oldest datum so as to keep the serials invariable. In this way, we can predict one by one and fill vacancies in the proper order until we finish the forecast target. This method overcomes the shortage of the regular GM(1,1) model[7]. 3.2 Grey Model GM
(1,1)
( )
GM represents Grey Model, In the bracket of GM 1,1 , the first 1 stands for First Order Equations, and the second 1 for a variable. Suppose that the known historical annual generating capacity is X, the data serial is as the following:
X (0) = [ x (0) (1), x (0) (2), …, x (0) (n)]
(1)
Making the accumulation of the serial, we get 1-AGO, and the new serial is:
X (1) = [ x (1) (1), x (1) (2),…,x (1) (n)]
(2)
among which k
x (1) (k ) = ∑ x (1) (i ) i =1
and from this new serial we got the close average value serial
Z(1)
Z (1) = [ z (1) (1), z (1) (2),…, z (1) (n)] among which
z (1) (k ) = 0.5 x (1) (k )+0.5x (1) (k -1) , k = 2,3, …, n
(3)
(4)
(5)
To set up the grey model GM (1, 1), the first degree Whitening differential equation is
dx (1) + ax (1) = b dt
(6) in which a and b stand for parameter. The least square in the parameter serial A = [a
b]T is A = [a b]T = ( BT B) −1 BT Y
among which
⎡ − z (1) (2) ⎢ (1) − z (3) B=⎢ ⎢ # ⎢ (1) ⎣ − z ( n)
⎡ x (0) (2) ⎤ 1⎤ ⎢ (0) ⎥ ⎥ x (3) ⎥ 1⎥ Y =⎢ ⎢ # ⎥ #⎥ ⎢ (0) ⎥ ⎥ 1⎦ ⎣ x ( n) ⎦
604
Y. Gao et al.
Put the calculated parameter a, b into equation (6), assume
x (1) (0) = x (0) (0) , we
solve the differential equation, and then get the equation for model GM(1,1) Λ (1)
b b x (k + 1) = [ x (0) (0) − ]e − ak + a a
(7)
After that we make the first degree decreasing restoration to the equation and we mark it as 1-IAGO and then we can get the grey forecast model GM(1,1)for the original serial X
,
Λ (0)
Λ (1)
Λ (1)
x (k ) = x (k ) − x (k − 1) ,
k = 1, 2,…, n
(8)
3.3 Metabolic Grey Model –GM (1, 1) [8]
In the original data serial, we insert new information information x
(0)
x (0) (n + 1) and delete the oldest
( )
(1) , and then we can set up the grey model GM 1,1 according to the (0) steps mentioned in section 2.2 by taking X = [ x (0) (2),…, x (0) (n), x (0) (n + 1)] as the original serial. We can do it again and again and then fill vacancies in the proper order until we finish the forecast target, that is, the Metabolic Grey Model –GM (1,1). On the forecast side, the metabolic model GM (1, 1) is an ideal model. With the development of the systems, the old information is becoming less important, so it is natural to delete the old information while the new information is constantly supplied. The establishing of the serial model can reflect the characteristic of the present system, especially when the system changes beyond the past with the accumulation of the quantitative change comparing with the old-timely system, so it is clearly reasonable to delete the old information which can not reflect the current characteristic of the system. Besides, by constant metabolism we can avoid the expanding continuously of computer memory because of the increase of information, meanwhile we can also avoid the difficulty of the increasing forecast budget in establishing model[9].
4 Analysis of the Case of Long-Term Generating Capacity Forecast of Wind Power Let’s take some wind farm in Fujin area as an example. Here we have the time serial of daily mean wind speed from 1954 to 2009 in this area. See Fig.1. The long-term generating capacity forecast of wind power is the forecast of the energy capture process, it has direct connection with the forecast of mean wind speed. When we get the changing forecast data of wind speed, we can calculate the actual generating capacity by the curve of design feature of wind-driven generating units. We can also find that the output power of wind power generation has certain relation with wind speed. This mainly depends on the MPPT control strategy of the wind power generating.
605
daily mean wind speed(m/s)
Study on the Medium-Term and Long-Term Forecast Technology
Time(day)
Fig. 1. daily mean wind speed from 1954 to 2009 in Fujin Area
By five-time polynomial combining the measured data of the relation between the output power of wind power generation and wind speed, we obtain the feature of the output power of wind-driven generating unit, see formula 9. Putting the data of the daily mean wind speed in Fujin Area from National Meteorological Information Centre to formula 9 and then integrating in one year cycle, we can obtain the calculated values of the generating capacity of some wind-driven unit from 1954 to 2009 in Jinjiang Area. Therefore, in view of the establishing of wind-driven farm in Jinjiang in a short time, we may take the calculated values as the sample for the study on the forecast methods of, instead of the actual generating capacity.
P = 6.61925 −3.54685v − 3.16953v2 + 2.40084v3 −0.19352v4 + 0.00436v5
(9) In the formula, V stands for speed, the unit is m/s, P for power and the unit is kw. First of all, we take the calculated values of the annual generating capacity from 1954 to 1961 as the sample data to set up the grey model GM(1,1), making the forecast on the annual generating capacity in 1962 of the unit. Then we establish the grey model GM (1,1) by using the calculated values of the annual generating capacity from 1955 to 1962 as the sample data, making the forecast on the annual generating capacity in 1963 of the unit. Every time when insert the new data of the generating capacity and delete the old data, we update the data in turn until we get the forecast generating capacity in 2009 of the unit. The comparing figure of the predicting curve and the curve of the original sample is shown as Fig.2.
power generation (MW h)
3400
calculate values 计算值 预测值 forecast values
3200 3000 2800 2600 2400 2200 2000 1800 1600 1400
1955
1965
1975
1985
1995
2005
2015
sample point(year)
Fig. 2. medium and long-term wind power generation forecast
606
Y. Gao et al.
The evaluation criterion of model is the precision of forecast. This paper chiefly applies the generalized mean absolute error NMAE to evaluate the precision of forecast of the model. It is defined as installed capacity
(
(
MAE =
1 N
NMAE =
) )
N
∑x −x i =1
i
(10)
ip
MAE CAP
(11)
The generalized mean absolute error of the generating capacity of some wind-driven generating units in Fujin Area from 1962 to 2009 is 7.8806%. It is obvious that metabolic grey model –GM (1, 1) have the better combining precision, we can come to the conclusion that the model can be applied to the medium-term and long-term forecast of the generating capacity of wind power generation. By using the model to calculate, We can obtain the stimulant value of the generating capacity of the wind-driven generating units from 2001 to 2008 in Fujin Area and comparing with its original serial X, the relative mean error is 2.99976% a =0.03001