Kang Li · Minrui Fei

Dajun Du · Zhile Yang

Dongsheng Yang (Eds.)

Communications in Computer and Information Science 924

Intelligent Computing
and Internet of Things
First International Conference on Intelligent Manufacturing
and Internet of Things and 5th International Conference on Computing
for Sustainable Energy and Environment, IMIOT and ICSEE 2018
Chongqing, China, September 21–23, 2018
Proceedings, Part II

123
Communications
in Computer and Information Science 924
Commenced Publication in 2007
Founding and Former Series Editors:
Phoebe Chen, Alfredo Cuzzocrea, Xiaoyong Du, Orhun Kara, Ting Liu,
Dominik Ślęzak, and Xiaokang Yang

Editorial Board
Simone Diniz Junqueira Barbosa
Pontifical Catholic University of Rio de Janeiro (PUC-Rio),
Rio de Janeiro, Brazil
Joaquim Filipe
Polytechnic Institute of Setúbal, Setúbal, Portugal
Igor Kotenko
St. Petersburg Institute for Informatics and Automation of the Russian
Academy of Sciences, St. Petersburg, Russia
Krishna M. Sivalingam
Indian Institute of Technology Madras, Chennai, India
Takashi Washio
Osaka University, Osaka, Japan
Junsong Yuan
University at Buffalo, The State University of New York, Buffalo, USA
Lizhu Zhou
Tsinghua University, Beijing, China
More information about this series at http://www.springer.com/series/7899
Kang Li Minrui Fei
•

Dajun Du Zhile Yang

•

Dongsheng Yang (Eds.)

Intelligent Computing
and Internet of Things
First International Conference on Intelligent Manufacturing
and Internet of Things and 5th International Conference on Computing
for Sustainable Energy and Environment, IMIOT and ICSEE 2018
Chongqing, China, September 21–23, 2018
Proceedings, Part II

123
Editors
Kang Li Zhile Yang
The University of Leeds Shenzhen Institute of Advanced Technology
Leeds Chinese Academy of Sciences
UK Shenzhen
China
Minrui Fei
Shanghai University Dongsheng Yang
Shanghai Northeastern University
China Shenyang
China
Dajun Du
Shanghai University
Shanghai
China

ISSN 1865-0929 ISSN 1865-0937 (electronic)

Communications in Computer and Information Science
ISBN 978-981-13-2383-6 ISBN 978-981-13-2384-3 (eBook)
https://doi.org/10.1007/978-981-13-2384-3

Library of Congress Control Number: 2018953021

© Springer Nature Singapore Pte Ltd. 2018

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, 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.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, express or implied, with respect to the material contained herein or for any errors or
omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
 Preface

This book constitutes the proceedings of the 2018 International Conference on Intel-
ligent Manufacturing and Internet of Things (IMIOT 2018) and International Confer-
ence on Intelligent Computing for Sustainable Energy and Environment (ICSEE 2018),
which were held during September 21–23, in Chongqing, China. These two interna-
tional conference series aim to bring together international researchers and practitioners
in the fields of advanced methods for intelligent manufacturing and Internet of Things
as well as advanced theory and methodologies of intelligent computing and their
engineering applications in sustainable energy and environment. The new conference
series IMIOT is jointly organized with the well-established ICSEE conference series,
under the auspices of the newly formed UK-China University Consortium in Engi-
neering Education and Research, with an initial focus on intelligent manufacturing and
sustainable energy.
At IMIOT 2018 and ICSEE 2018, technical exchanges within the research com-
munity took the form of keynote speeches, panel discussions, as well as oral and poster
presentations. In particular, two workshops series, namely the Workshop on Smart
Energy Systems and Electric Vehicles and the Workshop on Communication and
Control for Distributed Networked Systems, were held again in parallel with
IMIOT 2018 and ICSEE 2018, focusing on the two recent hot topics of the integration
of electric vehicles with the smart grid, and distributed networked systems for the
Internet of Things.
The IMIOT 2018 and ICSEE 2018 conferences received 386 submissions from over
50 different universities, research institutions, and companies from both China and UK.
All papers went through a rigorous peer review procedure and each paper received at
least three review reports. Based on the review reports, the Program Committee finally
selected 135 high-quality papers for presentation at the IMIOT 2018 and ICSEE 2018.
These papers cover 22 topics and are included in three volumes of the CCIS series,
published by Springer. This volume of CCIS includes 52 papers covering 9 relevant
topics.
Located at the upstream Yangtze basin, Chongqing constitutes the most important
metropolitan area in the southwest of China. It has a glorious history and culture and
serves as a major manufacturing center and transportation hub. Chongqing is also
well-known for its spicy food and hotpot, attracting tourists and gourmets from around
the world. In addition to academic exchanges, participants were treated to a series of
social events, including receptions and networking sessions, which served to build new
connections, foster friendships, and forge collaborations. The organizers of
IMIOT 2018 and ICSEE 2018 would like to acknowledge the enormous contribution
of the Advisory Committee, who provided guidance and advice, the Program Com-
mittee and the numerous referees for their efforts in reviewing and soliciting the papers,
and the Publication Committee for their editorial work. We would also like to thank the
editorial team from Springer for their support and guidance. Particular thanks are of
VI Preface

course due to all the authors, as without their high-quality submissions and presenta-
tions the conferences would not have been successful.
Finally, we would like to express our gratitude to our sponsors and organizers, listed
on the following pages.

September 2018 Fusheng Pan

Shilong Wang
Mark Price
Ming Kim Lim
Kang Li
Yuanxin Luo
Yan Jin
 Organization

Honorary Chairs
Fusheng Pan Chongqing Science and Technology Society/Chongqing
University, China
Shilong Wang Chongqing University, China
Mark Price Queen’s University Belfast, UK

General Chairs
Ming Kim Lim Chongqing University, China
Kang Li Queen’s University Belfast, UK

Advisory Committee Members

Erwei Bai University of Iowa Informatics Initiative, USA
Zhiqian Bo China Xuji Group Corporation, China
Tianyou Chai Northeastern University, China
Phil Coates Bradford University, UK
Jaafar Elmirghani University of Leeds, UK
Qinglong Han Swinburne University of Technology, Australia
Deshuang Huang Tongji University, China
Biao Huang University of Alberta, Canada
Guangbin Huang Nanyang University of Technology, Singapore
Minrui Fei Shanghai University, China
Sam Ge National University of Singapore, Singapore
Shaoyuan Li Shanghai Jiaotong University, China
Andy Long University of Nottingham, China
Dong Yue Nanjing University of Posts and Communication, China
Peter Taylor University of Leeds, UK
Chengshan Wang Tianjin University, China
Jihong Wang University of Warwick, UK
Xiaohua Xia Petoria University, South Africa
Yulong Ding University of Birmingham, UK
Yugeng Xi Shanghai Jiaotong University, China
Sarah Supergeon University College London, UK
Derong Liu University of Illinois, USA
Joe Qin The Chinese University of Hong Kong, Hong Kong, China
Savvas Tassou Brunel University London, UK
Qinghua Wu South China University of Technology, China
Yusheng Xue China State Grid Electric Power Research Institute, China
Jiansheng Dai King’s College London, UK
VIII Organization

I-Ming Chen Nangyang Technological University, Singapore

Guilin Yang Institute of Advanced Manufacturing Technology, Ningbo,
China
Zhuming Bi Indiana University Purdue University Fort Wayne, USA
Zhenyuan Jia Dalian University of Technology, China
Tian Huang Tianjin University, China
James Gao University of Greenwich, UK
Weidong Li Coventry University, UK
Stan Scott Queen’s University Belfast, UK
Dan Sun Queen’s University Belfast, UK

International Program Committee

Chairs
Yuanxin Luo Chongqing University, China
Yan Jin Queen’s University Belfast, UK

Local Chairs
Xuda Qin Tianjin University, China
Fuji Wang Dalian University of Technology, China
Yingguang Li Nanjing University of Aeronautics and Astronautics, China
Adam Clare University of Nottingham, UK
Weidong Chen Shanghai Jiaotong University, China
Rui Xiao Southeast University, China
Furong Li Bath University, UK
Min-Sen Chiu National University of Singapore, Singapore
Petros Aristidou University of Leeds, UK
Jinliang Ding Northeastern University, China
Bing Liu University of Birmingham, UK
Shan Gao Southeast University, China
Mingcong Deng Tokyo University of Agriculture and Technology, Japan
Zhengtao Ding The University of Manchester, UK
Shiji Song Tsinghua University, China
Donglian Qi Zhejiang University, China
Wanquan Liu Curtin University, Australia
Patrick Luk Cranfield University, UK
Guido Maione Technical University of Bari, Italy
Chen Peng Shanghai University, China
Tong Sun City University London, UK
Yuchu Tian Queensland University of Technology, Australia
Xiaojun Zeng The University of Manchester, UK
Huaguang Zhang Northeastern University, China
Shumei Cui Harbin Institute of Technology, China
Hongjie Jia Tianjin University, China
Youmin Zhang Concordia University, USA
 Organization IX

Xiaoping Zhang University of Birmingham, UK

Peng Shi University of Adelaide, Australia
Kay Chen Tan National University of Singapore, Singapore
Yaochu Jin University of Surrey, UK
Yuchun Xu Aston University, UK
Yanling Tian University of Warwick, UK

Organization Committee
Chairs
Congbo Li Chongqing University, China
Minyou Chen Chongqing University, China
Adrian Murphy Queen’s University Belfast, UK
Sean McLoone Queen’s University Belfast, UK

Special Session Chairs

Qian Tang Chongqing University, China
Xin Dai Chongqing University, China
Johannes Schiffer University of Leeds, UK
Wenlong Ming Cardiff University, UK

Publication Chairs
Zhile Yang Chinese Academy of Sciences, China
Jianhua Zhang North China Electric Power University, China
Hongjian Sun Durham University, UK
Trevor Robinson Queen’s University Belfast, UK

Publicity Chairs
Qingxuan Gao Chongqing University, China
Junjie Chen Southeast University, China
Brian Falzon Queen’s University Belfast, UK
Ben Chong University of Leeds, UK

Secretary-General
Yan Ran Chongqing University, China
Dajun Du Shanghai University, China
Rao Fu Queen’s University Belfast, UK
Yanxia Wang Queen’s University Belfast, UK

Registration Chairs
Guijian Xiao Chongqing University, China
Shaojun Gan Queen’s University Belfast, UK
X Organization

Program Committee Members

Stefan Andreasson Queen’s University Belfast, UK
Andy Adamatzky University of the West of England, UK
Petros Aristidou University of Leeds, UK
Vijay S. Asirvadam Universiti Teknologi Petronas, Malaysia
Hasan Baig University of Exeter, UK
Lucy Baker University of Sussex, UK
John Barry Queen’s University Belfast, UK
Xiongzhu Bu Nanjing University of Science and Technology, China
Jun Cao University of Cambridge, UK
Yi Cao Cranfield University, UK
Xiaoming Chang Taiyuan University of Technology, China
Jing Chen Anhui University of Science and Technology, China
Ling Chen Shanghai University, China
Qigong Chen Anhui Polytechnic University, China
Rongbao Chen HeFei University of Technology, China
Weidong Chen Shanghai Jiaotong University, China
Wenhua Chen Loughborough University, UK
Long Cheng Chinese Academy of Science, China
Min-Sen Chiu National University of Singapore, Singapore
Adam Clare University of Nottingham, UK
Matthew Cotton University of York, UK
Xin Dai Chongqing University, China
Xuewu Dai Northeastern University, China
Li Deng Shanghai University, China
Mingcong Deng Tokyo University of Agriculture and Technology, Japan
Shuai Deng Tianjin University, China
Song Deng Nanjing University of Posts and Telecommunications, China
Weihua Deng Shanghai University of Electric Power, China
Jinliang Ding Northeastern University, China
Yate Ding University of Nottingham, UK
Yulong Ding University of Birmingham, UK
Zhengtao Ding University of Manchester, UK
Zhigang Ding Shanghai Academy of Science and Technology, China
Dajun Du Shanghai University, China
Xiangyang Du Shanghai University of Engineering Science, China
Geraint Ellis Queen’s University Belfast, UK
Fang Fang North China Electric Power University, China
Minrui Fei Shanghai University, China
Dongqing Feng Zhengzhou University, China
Zhiguo Feng Guizhou University, China
Aoife Foley Queen’s University Belfast, UK
Jingqi Fu Shanghai University, China
Shaojun Gan Queen’s University Belfast, China
Shan Gao Southeast University, China
 Organization XI

Xiaozhi Gao Lappeenranta University of Technology, Finland

Dongbin Gu University of Essex, UK
Juping Gu Nantong University, China
Zhou Gu Nanjing Forestry University, China
Lingzhong Guo Sheffield University, UK
Yuanjun Guo Chinese Academy of Sciences, China
Bo Han Xi’an Jiaotong University, China
Xuezheng Han Zaozhuang University, China
Xia Hong University of Reading, UK
Guolian Huo North China Electric Power University, China
Weiyan Hou Zhengzhou University, China
Liangjian Hu Donghua University, China
Qingxi Hu Shanghai University, China
Sideng Hu Zhejiang University, China
Xiaosong Hu Chongqing University, China
Chongzhi Huang North China Electric Power University, China
Sunan Huang National University of Singapore, Singapore
Wenjun Huang Zhejiang University, China
Tan Teng Hwang University College Sedaya International University, Malaysia
Tianyao Ji South China University of Technology, China
Yan Jin Queen’s University Belfast, UK
Dongyao Jia University of Leeds, UK
Jongjie Jia Tianjin University, China
Lin Jiang University of Liverpool, UK
Ming Jiang Anhui Polytechnic University, China
Youngwook Kuo Queen’s University Belfast, UK
Chuanfeng Li Luoyang Institute of Science and Technology, China
Chuanjiang Li Harbin Institute of Technology, China
Chuanjiang Li Shanghai Normal University, China
Dewei Li Shanghai Jiao Tong University, China
Donghai Li Tsinghua University, China
Guofeng Li Dalian University of Technology, China
Guozheng Li China Academy of Chinese Medical Science, China
Jingzhao Li Anhui University of Science and Technology, China
Ning Li Shanghai Jiao Tong University, China
Tongtao Li Henan University of Technology, China
Weixing Li Harbin Institute of Technology, China
Xiaoli Li Beijing University of Technology, China
Xin Li Shanghai University, China
Xinghua Li Tianjin University, China
Yunze Li Beihang University, China
Zhengping Li Anhui University, China
Jun Liang Cardiff University, UK
Zhihao Lin East China University of Science and Technology, China
Paolo Lino University of Bari, Italy
Bin Liu University of Birmingham, UK
XII Organization

Chao Liu Centre national de la recherche scientifique, France

Fei Liu Jiangnan University, China
Guoqiang Liu Chinese Academy of Sciences, China
Mandan Liu East China University of Science and Technology, China
Shirong Liu Hangzhou Dianzi University, China
Shujun Liu Sichuan University, China
Tingzhang Liu Shanghai University, China
Wanquan Liu Curtin University, Australia
Xianzhong Liu East China Normal University, China
Yang Liu Harbin Institute of Technology, China
Yunhuai Liu The Third Research Institute of Ministry of Public Security,
China
Patrick Luk Cranfield University, UK
Jianfei Luo Chinese Academy of Sciences, China
Yuanxin Luo Chongqing University, China
Guangfu Ma Harbin Institute of Technology, China
Hongjun Ma Northeastern University, China
Guido Maione Technical University of Bari, Italy
Marion McAfee Institute of Technology Sligo, Ireland
Sean McLoone Queen’s University Belfast, UK
Gary Menary Queen’s University Belfast, UK
Gillian Menzies Heriot-Watt University, UK
Wenlong Ming Cardiff University, UK
Wasif Naeem Queen’s University Belfast, UK
Qun Niu Shanghai University, China
Yuguang Niu North China Electric Power University, China
Bao Kha Nyugen Queen’s University Belfast, UK
Ying Pan Shanghai University of Engineering Science, China
Chen Peng Shanghai University, China
Anh Phan Newcastle University, UK
Meysam Qadrdan Imperial College London, UK
Donglian Qi Zhejiang University, China
Hua Qian Shanghai University of Engineering Science, China
Feng Qiao Shenyang Jianzhu University, China
Xuda Qin Tianjin University, China
Yanbin Qu Harbin Institute of Technology at Weihai, China
Slawomir Raszewski King’s College London, UK
Wei Ren Shaanxi Normal University, China
Pedro Rivotti Imperial College London, UK
Johannes Schiffer University of Leeds, UK
Chenxi Shao University of Science and Technology of China, China
Yuntao Shi North China University of Technology, China
Beatrice Smyth Queen’s University Belfast, UK
Shiji Song Tsinghua University, China
Yang Song Shanghai University, China
Hongye Su Zhejiang University, China
 Organization XIII

Guangming Sun Beijing University of Technology, China

Tong Sun City University of London, UK
Xin Sun Shanghai University, China
Zhiqiang Sun East China University of Science and Technology, China
Wenhu Tang South China University of Technology, China
Xiaoqing Tang The University of Manchester, UK
Fei Teng Imperial College London, UK
Yuchu Tian Queensland University of Technology, Australia
Xiaowei Tu Shanghai University, China
Gang Wang Northeastern University, China
Jianzhong Wang Hangzhou Dianzi University, China
Jingcheng Wang Shanghai Jiaotong University, China
Jihong Wang University of Warwick, UK
Ling Wang Shanghai University, China
Liangyong Wang Northeastern University, China
Mingshun Wang Northeastern University, China
Shuangxin Wang Beijing Jiaotong University, China
Songyan Wang Harbin Institute of Technology, China
Yaonan Wang Hunan University, China
Kaixia Wei NanJing XiaoZhuang University, China
Lisheng Wei Anhui Polytechnic University, China
Mingshan Wei Beijing Institute of Technology, China
Guihua Wen South China University of Technology, China
Yiwu Weng Shanghai Jiaotong University, China
Jianzhong Wu Cardiff University, UK
Lingyun Wu Chinese Academy of Sciences, China
Zhongcheng Wu Chinese Academy of Sciences, China
Hui Xie Tianjin University, China
Wei Xu Zaozhuang University, China
Xiandong Xu Cardiff University, UK
Juan Yan University of Manchester, UK
Huaicheng Yan East China University of Science and Technology, China
Aolei Yang Shanghai University, China
Dongsheng Yang Northeastern University, China
Shuanghua Yang Loughborough University, UK
Wankou Yang Southeast University, China
Wenqiang Yang Henan Normal University, China
Zhile Yang Chinese Academy of Sciences, China
Zhixin Yang University of Macau, Macau, China
Dan Ye Northeastern University, China
Keyou You Tsinghua University, China
Dingli Yu Liverpool John Moores University, UK
Hongnian Yu Bournemouth University, UK
Kunjie Yu Zhengzhou University, China
Xin Yu Ningbo Institute of Technology, Zhejiang University, China
Jin Yuan Shandong Agricultural University, China
XIV Organization

Jingqi Yuan Shanghai Jiao Tong University, China

Hong Yue University of Strathclyde, UK
Dong Yue Nanjing University of Posts and Communications, China
Xiaojun Zeng The University of Manchester, UK
Dengfeng Zhang University of Shanghai for Science and Technology, China
Huifeng Zhang Nanjing University of Posts and Communications, China
Hongguang Zhang Beijing University of Technology, China
Jian Zhang State Nuclear Power Automation System Engineering
Company, China
Jingjing Zhang Cardiff University, UK
Lidong Zhang Northeast Electric Power University, China
Long Zhang The University of Manchester, UK
Qianfan Zhang Harbin Institute of Technology, China
Xiaolei Zhang Queen’s University Belfast, UK
Xiaoping Zhang University of Birmingham, UK
Youmin Zhang Concordia University, USA
Yunong Zhang Sun Yat-sen University, China
Dongya Zhao China University of Petroleum, China
Guangbo Zhao Harbin Institute of Technology, China
Jun Zhao Tianjin University, China
Wanqing Zhao Cardiff University, UK
Xingang Zhao Shenyang Institute of Automation Chinese Academy
of Sciences, China
Min Zheng Shanghai University, China
Bowen Zhou Northeastern University, China
Huiyu Zhou Queen’s University Belfast, UK
Wenju Zhou Ludong University, China
Yimin Zhou Chinese Academy of Sciences, China
Yu Zhou Shanghai Tang Electronics Co., Ltd., China
Yunpu Zhu Nanjing University of Science and Technology, China
Yi Zong Technical University of Denmark, Demark
Kaizhong Zuo Anhui Normal University, China

Sponsors

Chongqing Association for Science and Technology, China

Shanghai University, China

Organizers

Chongqing University, China

Queen’s University Belfast, UK
 Organization XV

Co-organizers

Southeast University, Beijing Institute of Technology, Dalian University of Technol-

ogy, Harbin Institute of Technology, Northwestern Polytechnical University, South
China University of Technology, Tianjin University, Tongji University, Shanghai
University, University of Birmingham, Cardiff University, University College London,
University of Nottingham, University of Warwick, University of Leeds.
 Contents – Part II

Advanced Evolutionary Computing Theory and Algorithms

Improved Shuffled Frog Leaping Algorithm for Multi-objection

Flexible Job-Shop Scheduling Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Mingli Gou, Qingxuan Gao, and Su Yang

Hysteretic Model of a Rotary Magnetorheological Damper in Helical

Flow Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Jianqiang Yu, Xiaomin Dong, Shuaishuai Sun, and Weihua Li

Dynamic Production Scheduling Modeling and Multi-objective

Optimization for Automobile Mixed-Model Production . . . . . . . . . . . . . . . . 25
Zhenyu Shen, Qian Tang, Tao Huang, Tianyu Xiong,
Henry Y. K. Hu, and Yi Li

Tandem Workshop Scheduling Based on Sectional Coding

and Varying Length Crossover Genetic Algorithm. . . . . . . . . . . . . . . . . . . . 34
Hao Sun and Xiaojun Zheng

Robust Bi-level Routing Problem for the Last Mile Delivery

Under Demand and Travel Time Uncertainty . . . . . . . . . . . . . . . . . . . . . . . 44
Xingjun Huang, Yun Lin, Yulin Zhu, Lu Li, Hao Qu, and Jie Li

A Comprehensive Fault Diagnosis System and Quality Evaluation Model

for Electromechanical Products by Using Rough Set Theory. . . . . . . . . . . . . 55
Jihong Pang, Ruiting Wang, and Yan Ran

The Research of Improved Wolf Pack Algorithm Based

on Differential Evolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
Yingxiang Wang, Minyou Chen, Tingli Cheng,
and Muhammad Arshad Shehzad Hassan

Multi-objective Optimization Genetic Algorithm

for Multimodal Transportation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Xiong Guiwu and Xiaomin Dong

Big Data Analytics

The Research of 3D Power Grid Based on Big Data . . . . . . . . . . . . . . . . . . 89

Bing He, Jin-xing Hu, and Ge Yang
XVIII Contents – Part II

Identifying Service Gaps from Public Patient Opinions

Through Text Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
Min Tang, Yiping Liu, Zhiguo Li, and Ying Liu

How to Verify Users via Web Behavior Features: Based

on the Human Behavioral Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Jiajia Li, Qian Yi, Shuping Yi, Shuping Xiong, and Su Yang

Authentication Using Users’ Mouse Behavior in Uncontrolled

Surroundings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Fan Mo, Shiquan Xiong, Shuping Yi, Qian Yi, and Anchuan Zhang

Construction of Network User Behavior Spectrum

in Big Data Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Mengyao Xu, Fangfei Yan, Biao Wang, Shuping Yi, Qian Yi,
and Shiquan Xiong

Relation Analysis of Heating Surface’s Steam Temperature Difference

and Fouling Degree Based on the Combination of Thermodynamic
Mechanism and Production Data Mining . . . . . . . . . . . . . . . . . . . . . . . . . . 144
Jiahui Wang, Hong Qian, and Cheng Jiang

Variable Selection Methods in Dredger Production Model . . . . . . . . . . . . . . 155

Yinfeng Zhang, Zhen Su, and Jingqi Fu

Fault Diagnosis and Maintenance

A New Data Analytics Framework Emphasising Pre-processing

in Learning AI Models for Complex Manufacturing Systems . . . . . . . . . . . . 169
Caoimhe M. Carbery, Roger Woods, and Adele H. Marshall

Performance Assessment of Multivariate Control System Based

on Data-Driven Covariance Historical Benchmark . . . . . . . . . . . . . . . . . . . . 180
Hong Qian, Gaofeng Jiang, and Yuan Yuan

Quantitative Safety Assessment Method of Industrial Control System

Based on Reduction Factor. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
Haoxiang Zhu, Jingqi Fu, Weihua Bao, and Zhengming Gao

Intelligent Computing in Robotics

The Electric Field Analysis and Test Experiments of Split Type Insulator
Detection Robot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
Pengxiang Yin, Xiao Hong, Lei Zheng, Biwu Yan, and Hao Luo
 Contents – Part II XIX

Study on Early Fire Behavior Detection Method for Cable Tunnel

Detection Robot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Biwu Yan, Guangzhen Ren, Xiaowei Huang, Junfeng Chi, Lei Zheng,
Hao Luo, and Pengxiang Yin

Speech Endpoint Detection Based on Improvement Feature

and S-Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
Lu Xunbo, Zhu Chunli, and Li Xin

A Localization Evaluation System for Autonomous Vehicle . . . . . . . . . . . . . 236

Yuan Yin, Wanmi Chen, Yang Wang, and Hongzhou Jin

Cooperative Slip Detection Using a Dual-Arm Baxter Robot . . . . . . . . . . . . 246

Shane Trimble, Wasif Naeem, and Seán McLoone

Intelligent Control and Automation

Study on Swimming Curve Fitting of Biomimetic Carangiform

Robotic Fish. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
Baodong Lou, Yu Cong, Minghe Mao, Ping Wang, and Jiangtao Liu

H1 Filter Designing for Wireless Networked Control Systems

with Energy Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
Lisheng Wei, Yunqiang Ma, and Sheng Xu

Distance Overestimation Error Correction Method (DOEC) of Time

of Flight Camera Based on Pinhole Model . . . . . . . . . . . . . . . . . . . . . . . . . 281
Le Wang, Minrui Fei, Hakuan Wang, Zexue Ji, and Aolei Yang

Discrete Event-Triggered H1 State-Feedback Control

for Networked Control Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
Weili Shen, Jingqi Fu, Weihua Bao, and Zhengming Gao

Brief Technical Analysis of Facial Expression Recognition. . . . . . . . . . . . . . 302

Lei Xu, Aolei Yang, Minrui Fei, and Wenju Zhou

Application of Intelligent Virtual Reference Feedback Tuning

to Temperature Control in a Heat Exchanger . . . . . . . . . . . . . . . . . . . . . . . 311
Yalan Wen, Ling Wang, Weiqing Peng, Muhammad Ilyas Menhas,
and Lin Qian

Design and Realization of 108MN Multi-function Testing System. . . . . . . . . 321

Shutao Zheng, Yu Yang, Zhiyong Qu, and Junwei Han

Research on Vibration Suppression for Boom Luffing of Telescopic

Boom Aerial Work Platform. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
Ru-Min Teng, Xin Wang, Ji-Zhao Wang, and Ji-Fei Liang
XX Contents – Part II

Intrusion Detection in SCADA System: A Survey . . . . . . . . . . . . . . . . . . . . 342

Pu Zeng and Peng Zhou

Intelligent Servo Feedback Control for Hydrostatic Journal Bearing. . . . . . . . 352

Waheed Ur Rehman, Jiang Guiyun, Nadeem Iqbal, Luo Yuanxin,
Wang Yongqin, Shafiq Ur Rehman, Shamsa Bibi, Farrukh Saleem,
Irfan Azhar, and Muhammad Shoaib

Iterative Feedback Tuning for Two-Degree-of-Freedom System . . . . . . . . . . 365

Hui Pan, Yanjin Zhang, and Ling Wang

IoT Systems

Data Monitoring for Interconnecting Microgrids Based on IOT . . . . . . . . . . . 383

Weihua Deng and Shufen Wang

A Hybrid Routing Control Mechanism for Dual-Mode Communication

of Streetlight Information Acquisition System . . . . . . . . . . . . . . . . . . . . . . . 390
Min Xiang, Xudong Zhao, and Yongmin Sun

An RF Energy Harvesting Approach for Secure Wireless

Communication in IoT. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400
Cheng Yin, Emiliano Garcia-Palacios, and Hien M. Nguyen

Low-Cost, Extensible and Open Source Home Automation Framework . . . . . 408

Che Cameron and Kang Li

Spectrum Utilization of Cognitive Radio in Industrial Wireless Sensor

Networks - A Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
Mingjia Yin, Kang Li, and Min Zheng

A Network Coding Against Wiretapping Attacks of the Physical

Layer Security Based on LDPC Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
Yujie Zheng and Jingqi Fu

Neural Networks and Deep Learning

A Weighted KNN Algorithm Based on Entropy Method . . . . . . . . . . . . . . . 443

Hui Zhang, Kaihu Hou, and Zhou Zhou

Control Strategy and Simulation for a Class of Nonlinear Discrete

Systems with Neural Network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
Peng Liu

Research on Joint Nondestructive Testing Based on Neural Network . . . . . . . 458

Junyang Tan, Dan Xia, Shiyun Dong, Binshi Xu, Yuanyuan Liang,
Honghao Zhu, and Engzhong Li
 Contents – Part II XXI

A New Real-Time FPGA-Based Implementation of K-Means Clustering

for Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
Tiantai Deng, Danny Crookes, Fahad Siddiqui, and Roger Woods

Neural Network Identification of an Axial Zero-Bias Magnetic Bearing . . . . . 478

Qing Liu, Li Wang, and Yulong Ding

Precision Measurement and Instrumentation

A Novel Bench of Quarter Vehicle Semi-active Suspension . . . . . . . . . . . . . 491

Xiaomin Dong, Wenfeng Li, Chengwang Pan, and Jun Xi

Research on Nondestructive Testing for the Depth of Hardening Zone. . . . . . 502

Ping Chen, Tao Xiang, Song Guo, Jie Fang, and Xin Xu

Distributed Fusion Estimation Based on Robust Kalman Filtering

for 3D Spatial Positioning System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 512
Li Liu, Wenju Zhou, Aolei Yang, Jun Yue, and Xiaofeng Zhang

Image Processing

A Novel 3D Head Multi-feature Constraint Method for Human

Localization Based on Multiple Depth Cameras . . . . . . . . . . . . . . . . . . . . . 525
Feixiang Zhou, Haikuan Wang, Zhile Yang, and Dong Xie

Head Ternary Pattern-Head Shoulder Density Features Pedestrian Detection

Algorithm Based on Depth Image . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536
Haikuan Wang, Haoxiang Sun, Kangli Liu, and Minrui Fei

Maximally Stable Extremal Regions Improved Tracking Algorithm

Based on Depth Image. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 546
Haikuan Wang, Dong Xie, Haoxiang Sun, and Wenju Zhou

Dynamic Hand Gesture Recognition Based on the Three-Dimensional

Spatial Trajectory Feature and Hidden Markov Model . . . . . . . . . . . . . . . . . 555
Kangli Liu, Feixiang Zhou, Haikuan Wang, Minrui Fei, and Dajun Du

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 565

Advanced Evolutionary Computing
Theory and Algorithms
 Improved Shuffled Frog Leaping Algorithm
for Multi-objection Flexible Job-Shop Scheduling
Problem

Mingli Gou, Qingxuan Gao ✉ , and Su Yang

( )

College of Mechanical Engineering, Chongqing University, Chongqing 400044, China

[email protected]

Abstract. This paper proposes an improved shuffled frog leaping algorithm

(SFLA-PSO) to solve multi-objective flexible job-shop problem. Pareto optimi‐
zation strategy is used to balance three objectives including minimum the
maximum completion time, maximum workload of all machines and the total
processing time of all job. In the new algorithm, the particle swarm flight strategy
is innovatively embedded into the local update operator of the SFLA-PSO. A
dynamic crowding density sorting method is used to update external elite archive
which holds the non-dominated population. To further improve the quality of the
solution and the diversity of the population, a new local optimization strategy is
developed associated with a neighborhood search strategy for achieving a high
development ability of the new algorithm. The test results of benchmark instances
illustrate the effectiveness of the new algorithm.

Keywords: Flexible job-shop scheduling · Multi-objective optimization

Pareto optimum solution · Shuffled frog leaping algorithm
Particle swarm algorithm

1 Introduction

The multi-objective flexible shop scheduling problems (MOFJSP) evolved from flexible
shop scheduling problems (FJSP) which has been proven to be NP-hard problem [1]. It
can comprehensively consider multiple targets and better meet modern production
requirements, but it is more complicated combinatorial optimization problem than FJSP
because of its increased machine selection, extended search scope as well as multiple
choices of optimization objectives. Therefore, this issue has attracted more and more
attention from academia and industry.
Two optimization approaches are ususlly developed for multi-objective flexible job
shop scheduling problem. The first method is weighted sum approach, namely, the multi-
objective optimization problems is transformed into a single objective optimization
problem by assigning different weights to each objective, and then solved by some meta-
heuristic approaches. Neverthless, this method only obtains a single solution, thus it is
difficult to find Pareto-optimal solutions for decision maker’s choice. The second
approach is called Pareto-based multi-objective evolutionary algorithms which is

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 3–14, 2018.
https://doi.org/10.1007/978-981-13-2384-3_1
4 M. Gou et al.

developed to reduce the influence of decision-maker’s preference. Instead of finding a

single optimal solution, this method generates Pareto non-inferior solutions with suffi‐
cient number and uniform distribution that satisfy multiple objectives. The most typical
algorithms is called NSGA- II proposed by Deb et al. [2], which is the improved version
of NSGA that is highly dependent on the parameters of the fitness sharing. Another
modified version of the NSGA-II is called non-dominated ranking genetic algorithm
(NRGA) [3] using Rolette-wheel strategy instead of tournament selection strategy. In
solving MOFJSP, Shao et al. [4] proposed a hybrid discrete particle swarm optimization
(DPSO) and simulated annealing (SA) to identify an approximation of the Pareto front
for FJSP. Zhong et al. [5] proposed a niching and particle swarm optimization algorithms
for multi-objective flexible job-shop scheduling problem. a new meta-heuristic is
designed based on ICA and VNS to minimize total tardiness and makespan under the
constraint by Lei et al. [6]
Since Shuffled frog leaping algorithm (SFLA) [7] put forward, SFLA has made great
progress in its applications [8, 9] because of its advantages of fewer parameters, fast
computation and global optimization ability. In solving job shop scheduling. Xu et al. [10]
proposed the shuffled frog leaping algorithm (SFLA) to solve the hybrid flow-shop problem
with multiprocessor tasks. A novel SFLA, where the search process within each meme‐
plex is done on its non-dominated member, is proposed based on a three-string coding
method to solve two conflicting goals by Xiuping Guo et al. [11] To our knowledge, the
research of the shuffled frog leaping algorithm hybriding other algorithms to find Pareto
solutions for multi-objective job shop scheduling problem seem relatively less.
In this paper, we proposed an improved shuffled frog leaping algorithm to solve
multi-objective job scheduling problem. Most importantly, the particle swarm flight
strategy was embedded into the local location update operator of the shuffled frog leaping
algorithm to increase global exploration capability, a new local optimization strategy
for external elite archive is developed associated with a neighborhood search strategy
for achieving a high development ability of the new algorithm. The objectives in this
paper are to minimize makespan, the workload of the critical machine and the total
workload of machines simultaneously.
This paper is organized as followed. We define the multi-objective flexible job shop
scheduling problem in Sect. 2, Sect. 3 presents the framework of the SFLA-PSO algo‐
rithm, SFLA-PSO algorithm is explained in Sect. 4 including the meme evolution, the
external elite document’s maintenance and so on. In Sect. 5, SFLA-PSO is compared
with other heuristic algorithm for classic instance. Section 6 gives the conclusion of the
paper.

2 Multi-objective Flexible Job Shop Scheduling Problem

The MOFJSP can be described as a set of n jobs Ji where i = 1, 2, … n which have to be

processed on a set of m machines Mk where k = 1, 2, … m. Each job contains a prede‐
termined sequence of fixed operations, each operation can be processed on one or more
different machines and the processing time varies with the machine. The goal of the
MOFJSP is to select the right machine for each operation and determine the optimum
 Improved Shuffled Frog Leaping Algorithm 5

processing order of each operation on the machine, so as to optimize the manufacturing

system as much as possible. The following assumptions and constraints are considered.
(1) All jobs and machines are available at time zero and each machine can only execute
one operation at a given time.
(2) There are no sequential requirements for the operations between the different jobs,
the operation of the same job must be finished in the previous operation before
starting the next process 3.
(3) Each operation cannot be interrupted during processing
(4) Once a job is completed on a machine, it is transported to the next machine imme‐
diately and the transportation time is negligible.
(5) Breakdowns are not considered for the machines.
The notation used in this paper are as following,
{ }
(1) Let J = Ji , 1 ≤ i ≤ n, indexed I, be a set of n jobs to be scheduled.
{ }
(2) Let M = Mj , 1 ≤ j ≤ m, indexed j, be a set of m machines.
(3) Let Oi,j be the jth operation of Ji, ni donates the total number of operation of job Ji.
(4) let Pi,j,k be the processing time of Oi,j on the machine Mk, Ti,j be the completion time
of the operation Oi,j.
(5) Decision variables Xi,j,k = 1 if machine Mk is selected for operation Oi,j, otherwise
Xi,j,k = 0.
In this paper, three objectives are simultaneously considered, which are given as
follows.

Min Tmax = max{Ti,j } (1)

1≤i≤n

n m
∑ ∑
Min Wmax = max {Wk } = max Pi,j,k Xi,j,k (2)
1≤k≤m 1≤k≤m
i=1 j=1

n ni m
∑ ∑ ∑
Min Lmax = Pi,j,k Xi,j,k (3)
i=1 j=1 k=1

Where Tmax is the maximum completion time of all jobs; Wmax is the maximum
workload of all machines; Lmax is the total processing time of all jobs.

3 Framework of SFLA-PSO Algorithm

In this paper, we propose an improved SFLA (SFLA-PSO) as showed in Fig. 1, which

is divided into three phases including initializing phase, evolution phase and update
external elite document phase.
6 M. Gou et al.

Parameter initialization
Initialize phase

Randomly generate the velocity and

displacement of the frog population, Update external elite
Evolution phase
produce scheduling operation document phase

Decoding and calculate muti-

Copy the nondominated
objective(Tmax,Wmax,C), determine the
population to External
domination rank by fast nondominated
elite document
sorting approach in NSGA- and
compute the crowding distance
Is it up to the Y
Executive
specified size? deletion strategy
Execute grouping operator
N
Execute meme evolution operator: the selection strategy
particle swarm flight strategy
embedded into the local location
update operator to update frog local optimization
population

Execute mixed operator

Perform neighborhood search strategy

N
Is the stop criteria satisfied?

Y
Output Pareto non-inferior solutions

Fig. 1. The framework of SFLA-PSO

4 The Hybrid Algorithm SFLA-PSO

4.1 The Meme Evolution Operator

The SFLA is a novel meta-heuristic algorithm, which makes use of the cultural envi‐
ronment information named as “meme” among the population to perform its evolution.
The searching process is performed using the instruction of memes generated by the
cooperation and competition among the population. In addition, the redistribution of the
sub-population makes searching process less possible to be trapped in local minima. The
traditional meme evolution operator are shown below.

ΔX = r and () ∗ (Xb − Xw ) (4)

( )
newXw = Xw + ΔX ΔXmax ≥ ΔX ≥ −ΔXmin (5)
 Improved Shuffled Frog Leaping Algorithm 7

( )
ΔX = r and () ∗ Xg − Xw (6)
( )
newXw = Xw + ΔX ΔXmax ≥ ΔX ≥ ΔXmin (7)

Where, in Eq. (4) ΔX denotes the moving step of frog individual, rand() represents
the random number uniformly distributed between 0 and 1, Xb and Xw represents the
best and worst frog in each meme group respectively, Eq. (5) denotes the update of the
location of the worst frog individuals in the meme group, if the frog could produce a
better position after the above update, then replace the original position with a new
position, otherwise replace Xg with Xb, and update with (6) (7) formula above. A new
solution is randomly generated instead of the worst frog if the above method still fails
to produce better frog individuals.
In this paper, the fly strategy of particles in PSO is embedded into the local update
operator of SFLA to guide the evolution of frog individuals in the meme group, the local
position update operator is adjusted as follows,
( k ) ( )
Vk+1
id
= wV k
id
+ c 1 r1 Pid
− Xk
id
+ c2 r2 Pk
gd
− Xk
gd (8)

k+1
Xid = Xkid + Vk+1
id (9)
( )
Vk+1
id
= wVkid + c3 r3 Pkgd − Xkid (10)

k+1
Xid = Xkid + Vk+1
id (11)

Where, in Eqs. (8), (9), (10) and (11) the parameters such as c1 , c2 , c3 , r1 , r2 , r3 and w
are the same meaning as those in the standard PSO algorithm. But the difference is that
Xkid represents the worse frog individual in the meme group, Pkid and Pkgd denotes the
optimal position of the meme group and all frog individuals respectively. According to
Eqs. (8) and (9),the frog individual’ speed and displacement is updated, The objective
function (Objv’) of the updated individual is calculated and compared with objective
function (Objv) of frog individual before update in Pareto dominance relation,
(a) if Objv’≻ Objv, then the speed and displacement of the individual is replaced with
the updated frog individual’s.
(b) if Objv ≻ Objv’, then update the speed and displacement according to Eqs. (10) and
(11), a new individual is randomly generated instead of the original individual if
the frog is still dominated by the original individual.
(c) if Objv’∼ Objv, then accept the speed and displacement of the updated individual
at a probability of 50%.
The Pareto relation can be described as followed, ( ) for multi-objective optimization
problem, min[fi (x), i = 1, 2, ⋯ , n], if fi (Xa ) ≤ fi Xb and a strict inequality exists at
least, then Xb is dominated by Xa, symbolized as Xa ≻ Xb, if the two vectors do not
dominate each other, it can be symbolized as Xa ∼ Xb.
8 M. Gou et al.

4.2 External Elite Document’s Maintenance

In order to solve the multi-objective optimization problem, the key is to find enough
uniformly distributed Pareto sets which are close to the real Pareto front. This paper uses
elitist strategy to store the non-dominated individuals generated during the evolution of
each generation and delete the dominant individuals that may exist in the elite document,
The algorithm steps are as follows.
(1) Copy the non-dominated individual generated by the individual flight to the elite
document, and delete the duplicated individuals and dominating ones in the docu‐
ment.
(2) If the number of individuals in the document exceeds the specified scale N, the
individuals of the same rank are descending order by the crowding-distance, and
then the first N individuals are retained. The method of computing crowding-
distance is the fast non-dominated sorting approach proposed by Deb et al. [8] in
NSGA II, however, the crowding-distance of the Pareto solution is calculated only
once in each iteration, thereby resulting in poor diversity and evenness of Pareto
solution. Therefore, in this paper each individual is removed and the crowding-
distance of the remaining individuals is recalculated until the specified size is satis‐
fied in the process of maintaining elite documents. Otherwise, the elite document
is not reduced. The proposed method avoids the increase of the number of non-
dominated individuals in the process of external document iteration, and removes
the redundant crowding individuals to ensure the uniformity of Pareto solution
distribution.

4.3 A Local Optimization Strategy for External Elite Documents

The outside elite documents stored non-dominated individuals in the iteration which
overcome the defects of NSGA II algorithm without using external elite population to
preserve the historical non-dominated individuals which caused some better solutions
to be lost, falling into premature convergence. However, the external elite document is
only used to store the non- dominated individuals searched in the iteration, and these
individuals do not react to the evolutionary population, resulting in reducing the conver‐
gence speed of the algorithm and decreasing the diversity of Pareto solutions. Therefore,
this paper proposes the local optimization strategy including selection strategy and local
optimization to optimize the external Pareto solution set. The selection strategy decides
how to choose the important search area to avoid blind search and the local optimization
is to optimize the important search regions. The selection strategy is as follows.
(1) Strategy 1, In order to avoid the repeated search for a region, a non-dominated
individuals on the current stored in the external elite document is designated the
largest number of local search, once more than the search times, the search area is
discarded.
(2) Strategy 2, A sphere radius (R) is specified in advance, aiming at guiding the search
toward promising regions. For non-dominated individuals that do not reach the
search upper limit (L), the number of non-dominated individuals within the radius
 Improved Shuffled Frog Leaping Algorithm 9

of the hyper-sphere is counted as the center, the smaller amount means the less non-
dominated individuals around it, indicating the new non-dominated individuals are
discovered with the high probability. If the minimum number exceeds one, the
average euclidean distance is calculated within the radius of the hyper-sphere, The
larger distance means the more sparse distribution of solutions, resulting in more
likely to find new non-dominated solutions. The local optimization algorithm for
external elite documents is given in Fig. 2.

Fig. 2. Local optimization algorithm for external elite documents

4.4 Selection of Global Optimal Solution

In the SFLA-PSO algorithm, the worst individual in the meme group uses the fly strategy
of the particle in PSO to track its historical optimum position and global optimum posi‐
tion to update its speed and displacement. For a single objective optimization problem,
the global optimal solution can be obtained by directly comparing the fitness values of
individual populations, but for multi-objective optimization, the Pareto solution sets
obtained in each evolution process are mutually exclusive, and there is no global optimal
solution in the traditional angle of view. In order to ensure the diversity of the Pareto
solutions and the uniformity, we can select non-dominated individuals randomly from
10 M. Gou et al.

the descending order based on crowding-distance after the first 20% as a global optimal
solution to guide the meme within the group of the worst individual updates.

4.5 The Encoding and Decoding of Frog Populations

In order to deal with the chromosome encoding in MOFJSP, we should not only choose
the machine to process, but also arrange the order of the operations on the machine.
Therefore, the double coding based on operation and machine is adopted in this paper.
The position vectors of individual frogs can be represented as vectors of a dimension
∑n
(total operation number L = i=1 nj), each L dimension vector is represented by
Xp [L], Xm [L] respectively. Take 3 × 4 FJSP as an example, O1 = {O11 , O12 , O13 },
O2 = {O21 , O22 , O23 }, O3 = {O31 , O32 , O33 }, The chromosome encoding and decoding
corresponding to a feasible solution is shown in Table 1.

Table 1. Examples of chromosome coding and decoding

Xm[ L] 3 4 2 4 3 3 2 2 1
encoding
Xp[ L] 1 3 2 2 1 3 3 1 2

Xp[ L] O11 O31 O21 O22 O12 O32 O33 O13 O23
decoding
Xm[ L] M3 M4 M2 M4 M3 M3 M2 M2 M1

Where Oij represents the j-th operation of the i-th job, when decoding, the procedure
firstly reads the operation coding successively and then randomly select a machine in a
given optional machine table of operation. The operations processed on each machine
after decoding are {M1 : O23 }, {M2 : O21 , O33 , O13 }, {M3 : O11 , O12 , O32 }, {M4 : O31 , O22 }.

4.6 The Adjustment of Frog Population’s Displacement

The velocity vector of the worst individual in the meme group is updated according to
the Eqs. (8) and (10), and the position vector is updated according to the Eqs. (9) and
(11). Since FJSP belongs to integer encoding, there may be a decimal fraction after
updating, so it is necessary to adjust the infeasible solution. The feasible operation is
obtained according to the reordered new displacement in ascending order. The adjust‐
ment results are shown in the following Table 2. At the same time, the process also has
machine constraint, hence there is a need to reselect the machine after the adjustment

Table 2. Example of the adjusted displacement of the frog individual

The displacement Xp [L] 1 3 2 2 1 3 3 1 2
The updated displacement Xp′ [L] 1.3 2.4 0.8 1.5 1.8 2.1 5 4.8 2.9
The ordered displacement Xp′ [L] 0.8 1.3 1.5 1.8 2.1 2.4 2.9 4.8 5
The adjusted displacement Xp [L] 2 1 2 1 3 3 2 1 3
 Improved Shuffled Frog Leaping Algorithm 11

operation, according to the procedure to find the process of optional machine set in the
optional machine table, and then randomly select a processing machine.

4.7 Neighborhood Search Strategy

SFLA-PSO has the advantages of few parameters, global searching ability and fast
convergence speed, but it is easy to fall into local optimum. In order to improve the local
search ability and increase the diversity and dispersion of Pareto front, the neighborhood
search strategy adopted in this paper is as follows.
(1) Exchange operations based on operation genes
The two positions which is selected randomly from the operation genes XP [L] are
interchanged, the rest remains the same and X′P [L] is obtained. As shown in Table 3(a),
the operation of gene position 3 and 6 interchange and the machine genes are adjusted
accordingly. The objection function of the individual after the interchange is calculated.
If the individual after the neighborhood search dominates the original individual, the
original individual is replaced.

Table 3. The operation base on operation

(2) Reverse operation based on operation genes

Reverse the operation between two positions which is selected randomly from the
operation genes XP [L], the rest remains the same, X′P [L] is obtained. As shown in
Table 3(b), the operation gene between position 3 and 6 reverse and the machine genes
are adjusted accordingly. The objection function of the individual after reversing is
calculated.If the individual after the neighborhood search dominates the original indi‐
vidual, the original individual is replaced.
(3) Mutation based on machine genes
Select a position randomly at Xm [L] and find the operation corresponding to the
position, choose a machine randomly that is different from the original machine at
12 M. Gou et al.

optional machine set. The objective function is calculated. If the mutated individual
dominates the original one, the original individual is replaced.

5 Computational Results

To evaluate the performance of the proposed SFLA-PSO algorithm, we conducted a

number of computational experiments. The AL + CGA algorithm [12], PSO + SA algo‐
rithm [13], PSO + TS algorithm [14] and SCAO algorithm [15] are used to compare
with the proposed SFLA-PSO algorithm.
In this paper, the SFLA-PSO algorithm was implemented in MATLAB environment
on an Intel Core i5-4200M 2.5 GHz PC. The Five Kacem instances are selected to test,
each instance can be characterized by the followed parameters, number of jobs (n),
number of machines (m), and each operation Oij of job i.

5.1 Problem 8 × 8

This is an instance of 8 × 8 with 27 operations. The comparison of the SFLA-PSO

algorithm with others is showed in Table 4, in which Si denote i th solution. The
comparison result shows that the proposed hybrid algorithm in this paper obtains more
Pareto solutions than other algorithm. Moreover, the optimal solution for the maximum
load of all machines is not under the premise of sacrificing the other two goals.

Table 4. Comparison of results on problem 8 × 8

AL + CGA PSO + SA PSO + TS SCAO SFLA-PSO
S1 S2 S1 S2 S1 S2 S1 S2 S1 S2 S3
Tmax 15 16 16 16 15 14 15 14 16 16 16
Lmax 79 79 75 73 75 77 76 77 73 75 77
Wmax – – 12 13 12 12 12 12 13 12 11

5.2 Problem 10 × 10
This is an instance of 10 × 10 with 30 operations taken from Kacem et al. [12]. The
comparison of the SFLA-PSO algorithm with others is showed in Table 5, in which Si
denote i th solution. The comparison result shows that the proposed SFLA-PSO algo‐
rithm obtains more and better Pareto solutions than other algorithms.

Table 5. Comparison of results on problem 10 × 10

AL + CGA PSO + SA PSO + TS SCAO SFLA-PSO
S1 S2 S1 S2 S1 S2 S1 S2 S1 S2 S3
Tmax 7 – 7 – 7 8 7 8 8 7 8
Lmax 45 – 44 – 42 42 42 42 41 42 42
Wmax 5 – 6 – 6 5 6 5 7 6 5
 Improved Shuffled Frog Leaping Algorithm 13

5.3 Problem 15 × 10
This is an instance of 15 × 10 with 56 operations taken from Du et al. (2008). The
comparison of the SFLA-PSO algorithm with others is showed in Table 6, in which Si
denote i th solution. The comparison result shows that the proposed SFLA-PSO algo‐
rithm is not bad.

Table 6. Comparison of results on problem 15 × 10

AL + CGA PSO + SA PSO + TS SCAO SFLA-PSO
S1 S2 S1 S2 S1 S2 S1 S2 S1 S2 S3
Tmax 23 24 12 – 11 11 11 11 11 11 –
Lmax 95 94 93 – 91 93 91 93 91 93 –
Wmax 11 11 11 – 11 10 11 10 11 10 –

6 Conclusions

This paper studies a multi-objective flexible job shop scheduling problem with the goal
of minimum the maximum completion time, maximum workload of all machines and
the total processing time of all job, and establishes a multi-objective optimization model.
Since it is NP-hard, an improved shuffled frog leaping algorithm were proposed to
heuristically solve them. The key contribution of this paper is that The particle swarm
flight strategy is embedded in the SFLA, and the dynamic crowding density sorting
method is used to update the non-dominant population external elite files, to improve
the performance of algorithm. Moreover,A new local optimization strategy is adopted
to achieve the high development ability of the new algorithm. The comparison results
with other typical algorithms demonstrated the effectiveness of the proposed algorithm.
The improved shuffled frog leaping algorithm proposed in this paper improves the
search performance of the shuffled frog leaping algorithm and extends the applicability
of the algorithm. How to improve the SFLA algorithm in solving more complex prob‐
lems and improve the convergence precision of the algorithm, simulated annealing
algorithm, the chaotic theory and immune algorithm with strong local search ability is
an important research direction in the future of the authors.

References

1. Garey, M.R., Johnson, D.S., Sethi, R.: The complexity of flowshop and jobshop scheduling.
Math. Oper. Res. 1(2), 117–129 (1976)
2. Deb, K.: A fast and elitist multiobjective genetic algorithm. NSGA-II. IEEE Trans. Evol.
Comput. 6(2), 182–197 (2002)
3. Jadaan, O.A., Rajamani, L., Rao, C.R.: Non-dominated ranked genetic algorithm for solving
multi-objective optimization problems, NRGA. J. Theor. Appl. Inf. Technol. 1, 60–67 (2008)
4. Shao, X.Y., Liu, W.Q., Liu, Q.: Hybrid discrete particle swarm optimization for multi-
objective flexible job-shop scheduling problem. Int. J. Adv. Manuf. Technol. 67(9–12), 2885–
2901 (2013)
14 M. Gou et al.

5. Zhong, Y., Yang, H., Mo, R., Sun, H.: Optimization method of flexible job-shop scheduling
problem based on niching and particle swarm optimization algorithms. Comput. Integr.
Manuf. Syst. 12(21), 3231–3238 (2015)
6. Lei, D., Li, M., Wang, L.: A two-phase meta-heuristic for multiobjective flexible job shop
scheduling problem with total energy consumption threshold. IEEE Trans. Cybern. PP(99),
1–13 (2018)
7. Eusuff, M., Lansey, K.: Optimization of water distribution network design using the shuffled
frog leaping algorithm. J. Water Resour. Plann. Manage. 129(2003), 210–225 (2003)
8. Luo, X.H., Yang, Y., Li, X.: Modified shuffled frog-leaping algorithm to solve traveling
salesman problem. J. Commun. 30(7), 130–135 (2009)
9. Ren, W.L., Zhao, C.W.: A localization algorithm based on SFLA and PSO for wireless sensor
network. Inf. Technol. J. 12(3), 502–505 (2012)
10. Xu, Y., Wang, L., Liu, M., Wang, S.Y.: An effective shuffled frog-leaping algorithm for
hybrid flow-shop scheduling with multiprocessor tasks. Int. J. Adv. Manuf. Technol. 68(5–
8), 1529–1537 (2013)
11. Lei, D., Zheng, Y., Guo, X.: A shuffled frog-leaping algorithm for flexible job shop scheduling
with the consideration of energy consumption. Int. J. Prod. Res. 55(11), 3126–3140 (2017)
12. Kacem, I., Hammadi, S., Borne, P.: Approach by localization and multiobjective evolutionary
optimization for flexible job-shop scheduling problems. IEEE Trans. Syst. Man Cybern. Part
C Appl. Rev. 32(1), 1–13 (2002)
13. Xia, W.J., Wu, Z.M.: An effective hybrid optimization approach for multi-objective flexible
job-shop scheduling problems. Comput. Ind. Eng. 48(2005), 409–425 (2005)
14. Zhang, G.H., Shao, X.Y., Li, P.G., Gao, L.: An effective hybrid particle swarm optimization
algorithm for multi-objective flexible job-shop scheduling problem. Comput. Ind. Eng.
56(2009), 1309–1318 (2009)
15. Xing, L.N., Chen, Y.W., Yang, K.W.: Multi-objective flexible job shop schedule, design and
evaluation by simulation modeling. Appl. Soft Comput. 9(2009), 362–376 (2009)
 Hysteretic Model of a Rotary
Magnetorheological Damper in Helical Flow
Mode

Jianqiang Yu1,2, Xiaomin Dong1(&), Shuaishuai Sun2,

and Weihua Li2
1
State Key Laboratory of Mechanical Transmission, Chongqing University,
Chongqing 400044, China
[email protected]
2
School of Mechanical, Materials and Mechatronic Engineering,
University of Wollongong, Wollongong, NSW 2522, Australia

Abstract. To capture the accurate hysteretic characteristics of a rotary mag-

netorheological (MR) damper in reciprocating motion, a new model with
reversibility is proposed and analyzed. The rotary MR damper in helical flow
mode is designed and tested on MTS under different current to obtain the
hysteretic characteristics. To portray hysteresis effectively and accurately, the
proposed model composed of a shape function and hysteresis factor is intro-
duced. To obtain the reversibility, the model is separated to the hysteretic part
and current-dependent part based on normalization method. The two parts fol-
low the multiplication rule. To improve computational efficiency, Constriction
Factor Particle Swarm Optimization (CFPSO) algorithm is used to identify the
model’s parameters. Feasibility and effectiveness of the identified model are
validated through comparison with two typical dynamic models.

Keywords: Rotary magnetorheological damper
Helical flow mode

Model
Hysteresis factor

1 Introduction

To suppress the vibration in transmission devices and suspensions, semi-active mag-

netorheological (MR) systems have been investigated in recent years. MR dampers in
the semi-active systems can provide tunable damping with rapid response and lower
energy consumption. To ensure the MR systems to be feasible and efficient, the
accurate dynamic model should be constructed to capture the MR dampers’ nonlinear
hysteresis. So far, different modeling methods and dynamic models are proposed and
applied in semi-active control procedures [1]. MR dampers in different working modes
exhibit various hysteretic characteristics. The damping in valve mode is magnetic field-
and rate-dependent while the shear mode is mainly influenced by magnetic field. The
hysteresis of MR damper in squeeze mode is rate-independent and gap width-
dependent [2]. Therefore, the dynamic model should be established based on the
unique characteristics of MR damper.

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 15–24, 2018.
https://doi.org/10.1007/978-981-13-2384-3_2
16 J. Yu et al.

The hysteretic characteristics of dampers working at valve mode can be predicted

by bi-viscous models [3], viscoelastic models [4], Bouc-wen models [5], Dahl models
[6], hyperbolic tangent models [7], sigmoid function models [8], polynomial model [9],
hysteretic-division model [10], and so on [11]. The parametric models contain the
Bouc-wen model [12, 13], Dahl model [14] and LuGre friction model [15] have been
used to describe damping characteristics of dampers in shear mode. Different from the
models of valve mode damper, the effects of stiffness are not considered in most of
shear mode models [16]. Improved Bouc-wen model can also be used to describe the
characteristics of dampers in squeeze mode [2].
As a new working mode, helical flow mode has been proposed and analyzed in our
prior work [17]. It is different from the conventional valve mode, shear mode and their
mixed mode. The dampers working at helical flow mode have high damping density
and can avoid sedimentation. Nowadays, vehicle or seat suspensions utilized rotary
magnetorheological dampers (RMRDs) have been studied because of the rotary
damper’s unique advantages [18, 19]. The rotary dampers don’t need accumulator to
provide volume and temperature compensation. Besides, the rotary dampers are ben-
eficial to integrate self-power and self-sensing systems. RMRD in helical flow mode is
also suitable for application in vehicle or seat suspensions. Considering high torque
density, the damping system with helical flow mode damper can save the force
amplifier such as gear box and rack and pinion. For its application and control, the
dynamic model should be established based on its hysteretic characteristics.
In this study, a novel dynamic model with inverse ability is proposed to capture the
characteristics of rotary MR damper in helical flow mode. Construction of this study is
as follows. The hysteretic characteristics are tested and analyzed in Sect. 2. Section 3
shows the details of the dynamic model. Parameters identification is conducted in
Sect. 3.2. Conclusions are made in Sect. 4.

2 Hysteretic Characteristics of RMRD in Helical Flow Mode

The configuration of RMRD in helical flow mode is shown in Fig. 1. It mainly consists
of outer cylinder, inner cylinder, piston, coil and rod. The inner cylinder rotates with
the rod while piston and outer cylinder keep still. Spiral groove in the rod is filled with
fluid. MR fluid flows in duct 1 and duct 2. The details of helical flow mode can be seen
in Ref. [17].
Total torque derived from three ducts can be calculated by Eqs. (1)–(4).

Tmax ¼ T1 þ T2 þ T3 ð1Þ
Z l
DP
T1 ¼ xhdx cos hðrd  0:5rh Þ ð2Þ
0 l
w
T2 ¼ 2pr23 Lg þ 2pr22 la s2 ð3Þ
d
 Hysteretic Model of a Rotary Magnetorheological Damper 17

Duct 3 Duct 2 Duct 1

Inner Outer
Rod Coil Piston
cylinder cylinder

Fig. 1. Configuration of RMRD in helical flow mode

(a) 0.2 A (b) 2.0 A

Fig. 2. Magnetic field intensity distribution in the duct 2 and duct 3

w
T3 ¼ 2pr33 Lg þ 2pr32 la s3 ð4Þ
d

here, T1 , T2 and T3 are torque in duct 1, duct 2 and duct 3, respectively. T1 is derived
from the effects of helical flow mode. T2 and T3 are both calculated based on working
principle of shear mode. In our prior work [17], a rotary damper is designed and tested.
In this study, the thickness of duct 2 and duct 3 is changed from 1.5 mm and 2 mm to
1 mm and 1 mm, respectively. Figure 2 shows the comparison of the magnetic field
intensity between the initial damper and novel one. The novel one shows higher
magnetic field intensity than the initial one both under 0.2 A and 2.0 A. The higher
magnetic field intensity is beneficial to improve coulomb damping force.
The components and assembly of the novel damper can be seen in Fig. 3. The
volume of MR fluid in RMRD is 385  106 m3 which is also smaller than the initial
damper. Test apparatus of the RMRD is shown in Fig. 4. Rack-and-pinion is applied to
convert the linear motion to rotational motion. Characteristics of torque-angle and
torque-velocity loops are displayed in Figs. 5, 6 and 7. The torque can be improved by
18 J. Yu et al.

(a) Components (b) Assembled

Fig. 3. Picture of manufactured RMRD

Rotary MR
damper

Rack and
Pinion

Control desk MTS Power supply

Fig. 4. Test apparatus of the RMRD

(a) Torque vs. angle (b) Torque vs. velocity

Fig. 5. Torque characteristics at amplitude of 6 mm and frequency of 0.6 Hz

 Hysteretic Model of a Rotary Magnetorheological Damper 19

(a) Torque vs. angle (b) Torque vs. velocity

Fig. 6. Torque characteristics at frequency of 0.8 Hz and current of 0.4 A

(a) Torque vs. angle (b) Torque vs. velocity

Fig. 7. Torque characteristics at amplitude of 10 mm and current of 0.6 A

increasing current obviously. The curve mutation appears in torque-angle loops

especially in Fig. 7 due to the cooperation problem of the rack-and-pinion at larger
damping force.

3 Modeling and Parameters Identification

3.1 Modeling Method

To capture the characteristics of torque-angle and torque-velocity curves, a model
based on a shape function shown in Eq. (5) and hysteresis factor is proposed. The
parameter sensitivity analysis of shape function is conducted and displayed in Fig. 8.

1  ebw
z¼ ð5Þ
1 þ ebw

Even though these type curves are enough to capture torque-angle or force-
displacement, the nonlinear hysteretic characteristics cannot be captured accurately. To
describe the hysteresis, hysteresis factor in Eq. (6) is introduced to replace the velocity
term in Eq. (5). To realize the reversibility of model, normalization method is used to
20 J. Yu et al.

separate hysteretic part and current-dependent part. The two parts follow the multi-
plication rule which is shown in Eq. (7).
 
w ! n1 w þ n2 sgn / € or w ! n1 w þ n2 sgnð/Þ þ n3 /
€ þ n3 / ð6Þ

Td ¼ TI G ð7Þ

€ and / are angular acceleration and angle, respectively. TI and G are current-
here, /
dependent part and hysteretic part, respectively. The maximum torque vs. velocity
under different current is shown in Fig. 9. The maximum torque is mainly controlled by
current while the velocity has little effect on maximum torque. The mathematical
expression of current-dependent part is shown in Eq. (8).
2
þ c3 I
T1 ¼ c 1 e c 2 I þ c4 ð8Þ

here, I is current. c1 , c2 , c3 and c4 are model’s parameters.

  1  en1 w þ n2 sgnð/Þ þ n3 /
T1 ¼ c1 ec2 I þ c3 I þ c4
2
ð9Þ
1 þ en1 w þ n2 sgnð/Þ þ n3 /

Fig. 8. Parameters analysis of the shape function

Fig. 9. Maximum force vs. velocity under different current

 Hysteretic Model of a Rotary Magnetorheological Damper 21

The inverse model is

pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
c3  c23 þ 4c2 j
I¼ ð10Þ
2c2
a w þ a sgnð/Þ þ a /
Here, j ¼ ln½ðTd =G  c4 Þ=c1  and G ¼ 11e 1 2 3
þ ea1 w þ a2 sgnð/Þ þ a3 /
.

3.2 Parameter Identification

To identify the dynamic model parameters efficiently, Constriction Factor Particle
Swarm Optimization (CFPSO) algorithm is introduced. Constriction factor in CFPSO
is used to ensure convergence of conventional PSO algorithm. The update mode is
dependent on Eq. (11).
    
vkidþ 1 ¼ v vkid þ c1 rand1k pbestid
k
 xkid þ c2 rand2k pbestdk  xkid ð11Þ
 
 0:5 
in which, v ¼ 2=2  r  ðr2  4rÞ  is constriction factor. r ¼ C1 þ C2 and r [ 4.
The values of parameters are selected based on prior knowledge: C1 ¼ C2 ¼ 2:05 and
convergence tolerance is 1:0e20 . Objective function is as follows.

X
N  2
J¼ T^t  Ts ð12Þ
k¼1

here, N is the number of data. T^t and Ts are the model torque and test data, respectively.
The optimization results can be seen in Table 1.

Table 1. Values of parameters in the novel dynamic model

Parameters c1 c2 c3 c4 n1 n2 n3
Values −1.5938 −8.1292 9.4075 −2.0948 0.0995 0.8785 0.1447

Experimental and theoretical results are compared in Fig. 10. The hysteresis
characteristics can be captured by the model in the visual. To further evaluate the
accuracy of the novel model, the arctangent model and hyperbolic tangent model are
identified and compared. Their equations are shown in Eqs. (13) and (14). The mean
deviation, root-mean-square error (RMSE) and coefficient of determinations R2 [20] are
set as evaluation coefficients. Values of the arctangent model, the hyperbolic tangent
model and the novel model about mean deviation are 1.6638, 1.6744 and 1.3808,
respectively. The values about RMSE are 96.7687, 96.5282 and 86.1562, respectively.
The lower values of mean absolute error and RMSE show that the novel model has
more accuracy than the other two models. Besides, the values about R2 are 0.9658,
0.9660 and 0.9729, respectively. The higher determination coefficient verifies the
trustable ability of the novel model. The differences between experimental results and
22 J. Yu et al.

theoretical results are also calculated and shown in Fig. 11. It is noted that the novel
model owing better accuracy.

a1 2 a4 a8 sgnð/Þ
Td ¼ /_ þ a tan½a7 ð/_ þ Þ ð13Þ
1 þ a2 e a 3 I p 1 þ ea5 ða6 þ I Þ 1 þ a9 ea10 I
 
Td ¼ ða1 I þ a2 Þ/_ þ ða3 I þ a4 Þ/ þ a5 I 2 þ a6 I þ a7 tanh½a8 /_ þ a9 sgnð/Þ ð14Þ

(a) Torque vs. angle /0.6 Hz (b) Torque vs. angle /0.8 Hz (c) Torque vs. angle /1.0
Hz

(d) Torque vs. velocity /0.6 Hz (e) Torque vs. velocity /0.8 Hz (f) Torque vs. velocity
/1.0 Hz

Fig. 10. Comparison of theoretical and experimental results

Fig. 11. Comparison of the accuracy of models

 Hysteretic Model of a Rotary Magnetorheological Damper 23

4 Conclusion

To capture the hysteretic characteristics of the RMRD in helical flow mode, a novel
model is proposed and analyzed based on the experimental results. The hysteretic
characteristics are captured by a shape function and hysteresis factor. The current-
dependent part and the hysteretic part are separated to ensure the reversibility. The
model are identified by CFPSO algorithm and compared with two typical models. The
mean absolute error, RMSE and coefficient of determinations are introduced to evaluate
accuracy of the models. Results show that the novel model can describe the nonlinear
hysteretic characteristics with feasibility.

Acknowledgments. We would like to thank the authors of the references for their enlighten-
ment. This research is also supported financially by the National Natural Science Foundation of
People’s Republic of China (Project No. 51675063), the Program for New Century Excellent
Talents in University (No. NCET-13-0630) and the State Scholarship Fund of China Scholarship
Council (No. 201706050094). These supports are gratefully acknowledged.

References
1. Wang, D.H., Liao, W.H.: Magnetorheological fluid dampers: a review of parametric
modelling. Smart Mater. Struct. 20, 1–34 (2011)
2. Chen, P., Bai, X.-X., Qian, L.-J., Choi, S.-B.: A new hysteresis model based on force–
displacement characteristics of magnetorheological fluid actuators subjected to squeeze
mode operation. Smart Mater. Struct. 26, 1–10 (2017)
3. Wereley, N.M., Panng, L., Kamath, G.M.: Idealized hysteresis modeling of electrorheolog-
ical and magnetorheological dampers. J. Intell. Mater. Syst. Struct. 9, 642–649 (1998)
4. Li, W.H., Yao, G.Z., Chen, G., Yeo, S.H., Yap, F.F.: Testing and steady state modeling of a
linear MR damper under sinusoidal loading. Smart Mater. Struct. 9, 95–102 (2000)
5. Ismail, M., Ikhouane, F., Rodellar, J.: The hysteresis Bouc-Wen model, a survey. Arch.
Comput. Methods Eng. 16, 161–188 (2009)
6. Dahl, P.R.: Solid friction damping of mechanical vibrations. AIAA J. 14, 1675–1682 (1976)
7. Şahin, İ., Engin, T., Çeşmeci, Ş.: Comparison of some existing parametric models for
magnetorheological fluid dampers. Smart Mater. Struct. 19, 1–11 (2010)
8. Ma, X.Q., Wang, E.R., Rakheja, S., Su, C.Y.: Modeling hysteretic characteristics of MR-
fluid damper and model validation. In: Proceedings of the IEEE Conference on Decision and
Control, vol. 2, pp. 1675–1680 (2002)
9. Choi, S.-B., Lee, S.-K., Park, Y.-P.: A hysteresis model for the field-dependent damping
force of a magnetorheological damper. J. Sound Vib. 2, 375–383 (2001)
10. Yu, J., Dong, X., Zhang, Z.: A novel model of magnetorheological damper with hysteresis
division. Smart Mater. Struct. 26, 1–15 (2017)
11. Chen, P., Bai, X.-X., Qian, L.-J., Choi, S.B.: An approach for hysteresis modeling based on
shape function and memory mechanism. IEEE/ASME Trans. Mech. 23, 1270–1278 (2018)
12. Pawlus, W., Karimi, H.R.: A comparative study of phenomenological models of MR brake
based on neural networks approach. Int. J. Wavelets Multiresolut. Inf. Process. 11, 1–30
(2013)
13. Miah, M.S., Chatzi, E.N., Dertimanis, V.K., Weber, F.: Nonlinear modeling of a rotational
MR damper via an enhanced Bouc-Wen model. Smart Mater. Struct. 24, 1–14 (2015)
24 J. Yu et al.

14. Tse, T., Chang, C.: Shear-mode rotary magnetorheological damper for small-scale structural
control experiments. J. struct. Eng. ASCE 130, 904–911 (2004)
15. Boston, C., Weber, F., Guzzella, L.: Modeling of a disc-type magnetorheological damper.
Smart Mater. Struct. 19, 1–12 (2010)
16. Imaduddin, F., Mazlan, S.A., Zamzuri, H.: A design and modelling review of rotary
magnetorheological damper. Mater. Des. 51, 575–591 (2013)
17. Yu, J., Dong, X., Wang, W.: Prototype and test of a novel rotary magnetorheological damper
based on helical flow. Smart Mater. Struct. 25, 1–15 (2016)
18. Sun, S.S., Ning, D.H., Yang, J., Du, H., Zhang, S.W., Li, W.H.: A seat suspension with a
rotary magnetorheological damper for heavy duty vehicles. Smart Mater. Struct. 25, 1–10
(2016)
19. Yu, J., Dong, X., Zhang, Z., Chen, P.: A novel scissor-type magnetorheological seat
suspension system with self-sustainability. J. Intell. Mater. Syst. Struct. 29, 1–12 (2018)
20. Dong, X., Yu, J., Wang, W., Zhang, Z.: Robust design of magneto-rheological (MR) shock
absorber considering temperature effects. Int. J. Adv. Manuf. Tech. 90, 1735–1747 (2017)
 Dynamic Production Scheduling Modeling
and Multi-objective Optimization
for Automobile Mixed-Model Production

Zhenyu Shen, Qian Tang, Tao Huang(&), Tianyu Xiong,

Henry Y. K. Hu, and Yi Li

State Key Laboratory of Mechanical Transmissions,

Chongqing University, Chongqing 400044, China
[email protected]

Abstract. Due to inventory redundancy problem caused by automakers mixed-

model production mode, a practical scheduling modeling and multi-objective
optimization strategy is presented to increase production and inventory efficiency
in this paper. Numerous factors including the general assembly shop, the painting
shop, and linear buffer between two workshops have been considered, and a novel
dynamic production scheduling model is proposed to achieve three optimization
goals: (i) equalize parts consumption rate in the general assembly shop so changes
in parts inventory can be predicted; (ii) reduce color switching frequency in the
painting shop’s production queue; (iii) reduce waiting time in car body’s buffer
zone. Based on this model, an embedded heuristic algorithm with NSGA-2 (No-
domination Sorting Genetic Algorithms-II) is employed to solve multi-objective
optimization problem. Simulations are finally conducted, when compared with a
traditional algorithm, the results are obviously better than traditional algorithm,
which validate effectiveness of the proposed model and optimization algorithm.

Keywords: Shop scheduling
Inventory control
Multi-objective optimization

Genetic algorithm

1 Introduction

Due to end customers’ diversified demands, more companies have adopted production
by order strategy, resulting in multiple types of products are collinearly produced on
mixed-model production lines. In consequence, wide variety of parts need to be
assembled which results in higher inventory cost. When formulating a production plan,
the amount of parts inventory needs to be known, so production sequence can be
reasonably arranged. Since production flow sequence determines parts usage sequence,

Foundation items: This work was supported in part by the Key Technology Research and System
Integration of Discrete Intelligent Manufacturing Workshop, China (No. cstc2016zdcy-
ztzx60001), the Fundamental Research Funds for the Central Universities of China under Grant
2018CDXYJX0019, and the National Nature Science Foundation of China under Grant
51805053 as well as Grant 51575069.

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 25–33, 2018.
https://doi.org/10.1007/978-981-13-2384-3_3
26 Z. Shen et al.

optimizing production queue can help to balanced parts consumption ratio. The pur-
pose of equalization is to reduce fluctuations in consumption rate. This not only
increases the robustness of production line, it is also possible to predict the con-
sumption rate of parts, which is the key to keeping inventory at a low level. Through
the strategy of milk run [1], stocks will eventually reach a dynamic balance, and it will
ultimately achieve the goal to reduce inventory costs.
Prior to the use of intelligent algorithms, most of production scheduling was done
manually. Shop scheduling problem is a typical NP-hard problem with exponential
explosion characteristics [2], and it is difficult to find the optimal solution relies on
manual scheduling. However, the emergence of intelligent algorithms solves this
problem. The results filtered by tens of millions of iterations are often much better than
the results of manual scheduling [3].
Therefore, this article introduces a dynamic scheduling model that covers the
general assembly shop, the painting shop, and painted body storage (PBS) which is a
buffer zone connects the two shops. Parts inventory in the general assembly shop will
dynamically change as time goes. The optimization goal of the general assembly shop
is to equalize the consumption rate of various parts as much as possible while ensuring
each car can be assembled on time; the painting shop is to reduce the number of color
switching in production queue and lower spraying cost; and PBS is to rearrange the
production queues coming out of the painting shop, and to reduce waiting time of
vehicles in the buffer zone, under the premise of meeting the general assembly shop’s
requirements. The NSGA-2 has been selected to achieve these multiple optimization
goals. This algorithm has a reliable performance in solving multi-objective optimiza-
tion problems and is widely used [4].

2 Mathematical Model

Many car companies are currently running on order production mode. An order infor-
mation represents a specific model of car. In the workshop, an order i corresponds to a
list of parts Mi , so the optimization of production queue is essentially the optimization of
fMi g. Pre-definition is mandatory for each model of car. Considering extensibility of the
mathematical model, 42 different models are predefined, shown in Table 1.

Table 1. Configuration table.

Model Number of configurations Number of colors
A 5 4
B 4 3
C 3 2
D 2 2

Meanwhile, a collection  of parts lists required for 42 models

M : Mi;j 2 M; i 2 N1 ; j 2 N2 is defined, and all parts are assigned to 10 stations.
Dissimilar parts have been divided into 4 categories, totaling 52, as shown in Table 2.
 Dynamic Production Scheduling Modeling and Multi-Objective Optimization 27

Table 2. Parts list.

Station Number of parts Type Station Number of parts Type
1 10 T1 6 4 T3
2 3 T2 7 3 T2
3 1 T4 8 4 T3
4 4 T3 9 3 T2
5 10 T1 10 10 T1

Among them, T1 are customized parts that can be used only for the specified model
of the specified configuration, such as seats and engines. T2 are common parts in
different car models that can be used only for the specified configuration, like in various
electronic auxiliary systems. T3 are common parts in different configurations that can
be used only for specified models, like interior trimming panels and wheels. T4 are
common parts for all vehicles, like different fasteners.
For the general assembly shop, the parts consumption speed needs to be more
balanced. The theoretical consumption rate of each part is

J ¼ P N1
EVj ¼ V ð1Þ
N1
i¼1 Mi;j

Deviation between actual consumption rate and theoretical consumption rate is the
optimization goal of the general assembly shop.
P N2
 J Þ2
ðVj  V
j¼1
min f1 ¼ ð2Þ
N1

For the painting shop, reducing the frequency of color switching can reduce coating
costs. The number of color switching can be expressed as
8
< 0; i¼1
Si ¼ 0; colorðiÞ ¼ colorði  1Þ ð3Þ
:
1; colorðiÞ ¼
6 colorði  1Þ

Therefore, the optimization goal of the painting shop is

XN1
min f2 ¼ i¼1
Si ð4Þ

There are many kinds of buffers for connecting the general assembly shop and the
painting shop, linear, circular, backward and so on [5]. In general, waiting buffer time
of car body is always proportional to the optimization effect of buffer. Therefore, linear
buffer has been selected for our model.
For PBS, the less time car bodies are waiting in PBS, the more efficient PBS will
be. The waiting time can be expressed as
28 Z. Shen et al.

8
< iði0 Þ, i0 ¼ 1
Wi ¼ iði0 Þ, iði  1Þ\iði0 Þ
0
ð5Þ
: 0
iði  1Þ, iði0  1Þ  iði0 Þ

So, the optimization goal of PBS is

X N1
min f3 ¼ Tt2 i0 ¼1
ðWi  iÞ ð6Þ

3 NSGA-2 and Heuristic Buffer Algorithm

Three optimization objectives have strong correlations and involve many variables.
Changes in any workshops’ production queue will affect another workshop queue and
buffer. Therefore, a multi-objective optimization algorithm is necessary. The typical
multi-objective evolutionary algorithms are NSGA-2, PESA-2 and SPEA-2. Each of
these three algorithms has its advantages and disadvantages. The advantage of NSGA-2
is that it has high operational efficiency and good distribution of solution sets in low-
dimensional problems; its disadvantage is that the diversity of solution sets is not ideal
in high-dimensional problems. The advantage of PESA-2 is that the convergence of its
solution is very good; but the disadvantage is that the selection operation can only
select one individual at a time, the time consumption is very large, and the class
diversity is not good. The advantage of SPEA-2 is that it can obtain a well-distributed
solution set, but its clustering process takes a long time to maintain diversity, and the
operating efficiency is not high.
In order to balance the running time with the quality of reconciliation, NSGA-2 was
eventually selected. Besides, we have designed a heuristic algorithm that can directly
calculate the downstream shop queue within constraints of the linear buffer, based on
the queue of upstream shop and the optimization goal of downstream shop which are
embed it in the NSGA-2.
The heuristic algorithms are divided into inbound rules and outbound rules. The
inbound rules are as follows:
1. If the last car in a lane has the same color of the waiting car, enter the lane.
2. Enter a lane with least car.
Its outbound rules are as follows:
1. In the waiting outbound car, select the same color as the last outbound car.
2. Unless the number of waiting cars in PBS is greater than M, otherwise there is no
car outbound.
3. When the number of waiting cars in PBS is greater than M, select the lane with the
highest number of cars, and outbound.
It should be pointed out that the general assembly shop is set to be the upstream
workshop and the painting shop as the downstream workshop. The reason is that the
optimization of the painting shop is simpler than the general assembly shop. A simple
 Dynamic Production Scheduling Modeling and Multi-Objective Optimization 29

heuristic algorithm can be selected if it achieves a good optimization effect. By

adjusting the parameter M, we can balance with waiting time and optimization effect.
When M is larger, there will be reduce color switching in the painting shop production
queue, but longer waiting time in PBS.
For the actual buffer, its inbound and outbound order are opposite to our algorithm.
When the algorithm is used in the real buffer, actual outbound rule will be different to
the outbound rule that we have designed, but inbound rule stays the same. The actual
outbound has only one rule: outbound is order specified from the general assembly
shop.
This heuristic algorithm needs to be embedded in NSGA-2, and through iterative
filtering, the scheduling plan which meets our optimization goals can be identified.
Algorithm flow chart is shown in Fig. 1.

Fig. 1. Algorithm flow chart.

4 Simulation Examples

The simulated data is based on the production plan of 1,000 cars in a car factory within
one day. In addition, a traditional algorithm is designed to compare with the above
algorithm. This traditional algorithm is actually applied to the factory, rely on PLC
control and will determine the inbound and outbound sequence of body in PBS
according to a fixed priority. Both that embedded heuristic algorithm based on NSGA-2
and traditional algorithms are programmed by Matlab.
30 Z. Shen et al.

Due to computational complexity, the production plan has been divided into 10
equal proportions smaller plans, so only one portion (100 cars) production queue needs
to be calculated. Then connect them end to end in a cyclic manner to get the total
production queue, so this total production queue is “equalized”. Inventory changes of
parts should conform to the rule in Fig. 2.

Fig. 2. Inventory Changes.

Two example was run 50 times. The operating environment of the program is Intel
Xeon CPU E3-1240 3.50 GHz, and the RAM is 32 GB. The average time of operation
of embedded heuristic algorithm was 278.8 s, which is sufficient for a one-day pro-
duction plan of a car factory. And traditional algorithm uses the same production
sequence of painting shop with embedded heuristic algorithm.
The results of the two algorithms are shown in Tables 3 and 4 respectively.

Table 3. Non-dominated solution set of embedded heuristic algorithm.

Pareto solution Optimization goal
f1 f2 f3
P1 2203.3 11 700
P2 2209.8 11 500
P3 2297.4 11 450
P4 2309.6 10 900
Average 2582.7 12.8 810

Taking P3 as an example, cars are numbered from 1 to 42 according to their

models, configurations and colors. Starting from the time when first car arrives at
station 10 in the general assembly shop, record the time t ¼ T0 . Then parts inventory
status and production line status at t ¼ T1 ¼ T0 þ 10  Tt1 , t ¼ T2 ¼ T1 þ 10  Tt1 are
respectively as shown in Figs. 3 and 4.
The solution obtained by the traditional algorithm have lower waiting time, but the
consumption rate of parts is very uneven. The consumption rate of parts will directly
 Dynamic Production Scheduling Modeling and Multi-Objective Optimization 31

Table 4. Corresponding solution set of traditional algorithm.

Solution Optimization goal
f1 f2 f3
0
P1 7354.8 11 450
P02 8785.4 11 450
P03 7927.5 11 450
P04 7651.7 10 450
Average 7885.1 12.8 450

Fig. 3. The status of inventory at different times.

affect the inventory of the assembly shop, but the waiting time will not be. In order to
study what negative impact the waiting time would have on the production line, some
inspection work was done.
When observing the results of the calculation, an unrealistic situation was dis-
covered. The results show that when waiting time is bigger, a kind of defect is more
likely to appear. This defect will result vacancies at the production queue of the general
assembly shop. The reason for this defect is that there are fewer bodies available in the
PBS when the queue is first started. If there is no body that meets the outbound rules,
there will be a vacancy in the assembly shop queue.
To avoid this situation, the production speed of the painting shop needs to be
greater than the general assembly shop, and the M value increases over time for a short
period of time when the queue is just starting to run. When the available car body in the
PBS is sufficient, the M value no longer increases.
Actually, as long as the car body can meet the constraints of continuous outbound,
they are all equivalent regardless of the waiting time. If the embedded heuristic
algorithm can ensure continuous outbound car body in PBS. The traditional algorithms
do not have obvious advantages in Optimization goal 3. Instead, there is a clear
disadvantage in optimizing goal 1. It can be considered that the embedded heuristic
algorithm has a comparative advantage over the traditional algorithms.
32 Z. Shen et al.

Fig. 4. The status of the production line at different times.

5 Conclusion

In this paper, a dynamic scheduling model is proposed for automobile mixed-model

production. By scheduling the sequence reasonably, this model can balance the con-
sumption rate of parts in the general assembly shop, reduce the number of color
switching in the painting shop, and improve reordering efficiency of the linear buffer.
When compared with a traditional algorithm, the result is obviously better than the
traditional algorithm. In addition, this is beneficial to achieve a more accurate inventory
control and sequence tracking by monitoring status of parts inventory and production
queues. It can be known for any order, it is where and how to complete, and even
which batch of parts are being used at the time. The value calculated by this model can
be serve as an important reference for the automakers who want to reduce their pro-
duction cost.
 Dynamic Production Scheduling Modeling and Multi-Objective Optimization 33

References
1. Sadjadi, S.J., Jafari, M., Amini, T.: A new mathematical modeling and a genetic algorithm
search for milk run problem. Int. J. Adv. Manuf. Technol. 44(1–2), 194–200 (2009)
2. Jalilvand-Nejad, A., Fattahi, P.: A mathematical model and genetic algorithm to cyclic
flexible job shop scheduling problem. J. Intell. Manuf. 26(6), 1085–1098 (2015)
3. Soleimani, H., Kannan, G.: A hybrid particle swarm optimization and genetic algorithm for
closed-loop supply chain network design in large-scale networks. Appl. Math. Model. 39(14),
3990–4012 (2015)
4. Deb, K.: A fast elitist non-dominated sorting genetic algorithm for multi-objective
optimization: NSGA-2. Lect. Notes Comput. Sci. 1917, 849–858 (2000)
5. Chen, GY: Research on buffer design and resequence in automobile production line.
Huazhong University of Science and Technology (2007)
 Tandem Workshop Scheduling Based
on Sectional Coding and Varying Length
Crossover Genetic Algorithm

Hao Sun and Xiaojun Zheng(&)

School of Mechanical Engineering, Dalian Jiaotong University, Dalian, China

[email protected]

Abstract. For the tandem workshop scheduling problem, the objective of

optimization is to obtain minimum total distribution time. To achieve that goal,
we propose an optimization model, considering the rated load of automated
guided vehicles (AGV) and the different regional transportation speeds. This
model has three features. First, the sectional coding rules are adopted because
materials need to be transported in batches between machines. Second, the
crossover operation with varying length is used because the superior charac-
teristics of the previous generation population could be better passed down to
the offspring, thus accelerating the convergence rate of the population. Finally,
the mutation operation combining insertion and reverse can maintains the
diversity of the population and improve the local search ability of the algorithm.
The tandem workshop scheduling problem can apply our algorithm, and the
effectiveness of the improvement is demonstrated.

Keywords: Tandem workshop scheduling
Genetic algorithm

Sectional coding

1 Introduction

In the Flexible Workshop Scheduling Problem (FJSP), the material transportation

system is one of the important design aspects. Research shows that the cost of material
transportation links takes up 30–95% of the total production cost [1]. As a tool for
material transportation and distribution, AGV can improve the automation, flexibility,
efficiency and safety of the modern flexible workshops. It is mainly to design the route
of transportation by considering the constraints (Such as speed, the number of AGV,
deadweight and so on) in order to achieve the goal of the shortest path or time.
With the increasing amount of material transportation in the flexible workshops, the
demand for AGVs from the system has also increased, which greatly increases the
complexity of the scheduling tasks, and it is easy to cause traffic jam, collision and
other problems. Bozer and Srinivasan [2] proposed the concept of a tandem workshop
scheduling system, effectively avoiding the conflicts and collisions. Zhang [3] proposed
a tandem workshop hybrid scheduling model based on cloud computing environment,
the model could efficiently realized the load balance. Zhou et al. [4] putted the tabu
search into the genetic algorithm to solve single-field tandem workshop scheduling

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 34–43, 2018.
https://doi.org/10.1007/978-981-13-2384-3_4
 Tandem Workshop Scheduling Based on Sectional Coding 35

optimization problem. Bai et al. [5] proposed a tandem workshop control model, which
reduced the complexity of AGVs scheduling and improved the efficiency of the system.
Tang et al. [6] solved the problem of task scheduling and coordination control in
tandem AGV systems by neuro-endocrine coordination mechanisms. Hou et al. [7]
developed a collaborative optimization method to solve the problem of machine
allocation and loop layout for tandem AGV systems. Reza et al. [8] putted forward a
method based on tabu search and genetic algorithm to avoid infeasible solutions from
crossover operation.
The paper will be organized as follows: in Sect. 2, a tandem workshop scheduling
system model was defined. And we established the mathematical model in Sect. 3.
Next, we explained the proposed sectional coding and the varying length crossover
genetic algorithm (VGA) as detailed in Sect. 4. In the Sect. 5, we compared the result
of the proposed algorithm with the fixed length crossover genetic algorithm (FGA) to
verity the effectiveness of the algorithm. At last, we made a conclusion.

2 Problem Description

In a tandem workshop scheduling system, the distribution of the machines are given,
the AGVs move in a bi-directional path, and only one AGV is worked in different areas
to avoid collisions. In different areas, the location of a machine is designated as a
transfer station, which is used to handle transportation tasks between different areas,
and all transfer stations constitute a transfer center. AGV has different operating speeds
in different areas, which can save transportation time, increase production efficiency,
and save costs. The tandem workshop scheduling model is shown in Fig. 1.

Fig. 1. The tandem workshop scheduling model

The optimization for tandem AGV workshop aims to provide the excellent scheme
of path planning, so that the material distribution time of tandem workshop scheduling
system is reduced effectively.
36 H. Sun and X. Zheng

3 Mathematical Model

In a tandem workshop scheduling system, the machine layout is given, There is a

material transfer station (A machine can be regarded as a transfer station) in different
areas, and only one AGV is available for scheduling, without considering the loading
and unloading time of the material. The speed of AGVs for material transport between
machines is v1 , the speed of AGVs for material transport between different transfer
stations is v2 . A kind of material can be transported between different machines at a
time.
The maximum load capacity of AGV is Q, The number of AGV is
i 2 f1; 2; . . .; Rg, The number

of different
 areas (Which include machine areas and
transfer centers) is rj ; sj 2 1; 2; . . .; Mj , Mj is the total number of machines in area j.
A preferable AGV path scheduling optimization scheme is obtained, and the
shortest time for the entire tandem workshop scheduling system is calculated. We
establish the following mathematical model:
Objective function:

R X
X Mj X
N X Mj
Min dijrj sj Tijrj sj : ð1Þ
i¼1 j¼1 rj ¼1 sj ¼1

Constraints:

X Mj X
N X Mj
dijrj sj qjrj sj  Q 8i 2 f1; 2; . . .; Rg: ð2Þ
j¼1 rj ¼1 sj ¼1

dijrj sj
tijrj sj ¼ 8i 2 f1; 2; . . .; Rg 8j 2 f1; 2; . . .; N g: ð3Þ
v1
0
0
dijr j sj
tijr j sj
¼ 8j 2 f1; 2; . . .; N g 8i 2 f1; 2; . . .; Rg: ð4Þ
v2
Mj X
X N Mj X
X N
0
Tijrj sj ¼ dijrj sj tijrj sj þ dijrj sj tijr j sj
8i 2 f1; 2; . . .; Rg: ð5Þ
rj ;sj ¼1 j¼1 rj ;sj ¼1 j¼1

Mj
X
dij1sj ¼ 1 8i 2 f1; 2; . . .; Rg 8j 2 f1; 2; . . .; N g: ð6Þ
sj ¼1

Mj
X
dijrj 1 ¼ 1 8i 2 f1; 2; . . .; Rg 8j 2 f1; 2; . . .; N g: ð7Þ
rj ¼1
 Tandem Workshop Scheduling Based on Sectional Coding 37
Mj
X Mj
X
dijrj h  dijhsj ¼ 0 8h 2 f1; 2; . . .; M g 8i 2 f1; 2; . . .; Rg
rj ¼1 sj ¼1
ð8Þ
8j 2 f1; 2; . . .; N g:

Where:
Formula (1) shows the total time for AGV to perform all transportation tasks in
tandem workshop.
Formula (2) indicates that the load of a AGV cannot exceed the load-carrying quota.
Formula (3) indicates that in the machine area, the operating time of AGV transport
materials between two machines.
Formula (4) indicates that in the transfer center, the operating time of AGV
transport materials between two machines.
Formula (5) represents the total time of different areas to accomplish tasks.
Formula (6)–(8) means that different AGV will return to the starting position after
completing all the transportation tasks from the transfer station.
Notes:
Tijrj sj : In the area j, the time for the i-th AGV to complete the transportation tasks
from the machine rj to the machine sj .
qjrj sj : In the machine area j, transport the weight of the material from machine rj to
machine sj .
0
dijrj sj : In the transfer center j, the distance of the i-th AGV from the machine rj to the
machine sj .
dijrj sj : In the transfer center j, the distance of the i-th AGV from the machine rj to the
machine sssssj .
8
< ¼ 1 If the i-th AGV is arranged for material transportation from
dijrj sj machine rj to machine sj :
:
¼ 0 Other

4 Sectional Coding and Varying Length Crossover Genetic

Algorithm

In this paper, a tandem workshop AGV scheduling model is established based on the
minimum material transportation time as the objective function. Compared with the
traditional genetic algorithm, the coding method of chromosomes is redesigned, which
could generate chromosomes to meet the needs of the optimization problem. The
objective function is used as the fitness function of the algorithm. With improved
crossover and mutation operation, the certain cross rate and mutation rate are set, so
that the superior individuals of the previous generation are not lost, and this can ensure
the variety.
38 H. Sun and X. Zheng

4.1 Coding Design

There are many encoding strategies for genetic algorithms. A sectional coding based on
the order of transportation tasks is used in this paper. One AGV is only responsible for
transport tasks in the current area, so the chromosome should be segmented according
to the number of AGV in the system. Each segment represents the sequence of an AGV
transportation path. Each gene represents the transportation tasks required between two
machines. Since the load-carrying quota of different AGVs cannot exceed the load
capacity, there will be the same serial number of the transportation path. Figure 2
shows an example of chromosome coding design, which indicates that the entire
scheduling system consists of two areas, in which the numbers represent the trans-
portation path from start to finish. The same number represents the same path, and the
times indicate the number of transportation along the path.

Fig. 2. An example of chromosome coding design

4.2 Selection Operation

For the selection operation, tournament selection method is adopted. Selecting a certain
size of individuals from the population (Which is known as the scale of the tournament,
the value is two), comparing different individuals, and retaining individuals with higher
fitness values to the next generation, and this operation is repeated until the scale of
next generation reaches the preset value.

4.3 Crossover Operation

For genetic algorithms, chromosomes are cross-processed to generate new individuals.
Due to real-valued sectional coding is adopted, it is unsatisfactory to employ the cross-
rules of traditional genetic algorithm for addressing this complex optimization problem,
and it is easy to produce infeasible solutions. In literature [9], the use of the crossover
operation fixed cross-block size may make the population converge slowly and prone
to local optimum. However, the crossover operation of the varying length crossover
can not only make the excellent characteristics of the two parent retained, but also
accelerate the convergence rate of the population. The crossover operation improved is
as follows:
Step 1: In different gene coding segments of the chromosome, two random numbers
(Between 0 and L, where L represents the length of the coding segment) are
generated, respectively, as the cross-block size of the two individuals of the
parent.
Step 2: In the parental chromosome Parent1, a crossing place A1 is randomly
generated, then the length of the cross-block is added to generate the second
crossing place A2. Therefore, the genes between the two crossing places are
the cross-block A12.
 Tandem Workshop Scheduling Based on Sectional Coding 39

Step 3: In the chromosome Parent2, the operation of step 2 is repeated to form a

cross-block B12.
Step 4: The value of each gene bit of the cross-block A12 is placed into the
offspring2 in turn, and the value of each gene bit of the cross-block B12 is
placed into offspring1 in turn.
Step 5: Finding the extra genes in Parent1 and Parent2 relative to B12 and A12
respectively, and putting them into the offspring1 and offspring2.

The following Figs. 3, 4 and 5 show the specific operation of the crossover
operation:

Fig. 3. Selecting the crossing place and the cross-block

4.4 Mutation Operation

The mutation operation applying the insertion (The gene of one place is inserted into
another place) and reverse sequence (Reversing sequence of genes between different
places) were considered for the VGA. This method can improve the search ability of
the whole algorithm, avoid premature phenomena, and maintain the diversity of the
whole population. The specific mutation operation for different gene encoding segment
is as follows:
Step 1: Selecting a gene at a certain position (Where P1 is located) in the Parent,
randomly.
Step 2: A place P2 (P1 and P2 are not equal) is randomly selected in the Parent, and
the gene selected in the previous step is inserted into the place of P2, and the
rest of the genes are moved forward.
Step 3: Then, two places randomly selected in the Parent are P3 and P4. If P3 is
equal to P4, then P4 is the value of the L-2 gene place (L is the length of this
chromosome segment).
Step 4: Reversing the sequence of genes between P3 and P4, and replacing the
original gene block. Thus, a new chromosome offspring Offspring is formed.

As shown in Figs. 6 and 7, the specific operation for the mutation operation.
40 H. Sun and X. Zheng

Fig. 4. Finding the extra genes

Fig. 5. Generating new offspring

Fig. 6. Insertion mutation operation

Fig. 7. Reversing sequence mutation operation

5 Experimental Verification

In order to analyze the advantages of VGA, a tandem workshop material transportation

is taken as an example. The workshop consists of five areas (Which include four
machine areas and one transfer center) and 24 machines, and different areas contain the
machine serial number as shown in Table 1, the space position of different machines is
given (As shown in Table 2 below). There are 5 AGVs in the workshop for trans-
portation work. The AGV speed is 40 m/min in the four machine areas and the AGV
speed in the transfer center area is 60 m/min. The material loading and unloading time
in the workshop is negligible. The maximum load of different AGVs is 100 kg. The
 Tandem Workshop Scheduling Based on Sectional Coding 41

materials required to be transported between the machines in the workshop are shown
in Table 3 below.

Table 1. Machine serial number in different areas

Area Machine Area Machine
Machine area 1 1, 2, 3, 4, 5, 6 Machine area 4 19, 20, 21, 22, 23, 24
Machine area 2 7, 8, 9, 10, 11, 12 Transfer center 5 4, 11, 13, 20
Machine area 3 13, 14, 15, 16, 17, 18

The problem is solved by applying our algorithm. The detailed parameters are as
follows: The population size is N = 100, the crossover rate is 0.6, and the maximum
iteration number of the algorithm is 10000 times. The algorithm of DE, the fixed length
crossover genetic algorithm and the genetic algorithm of this paper were performed 50
times respectively. The data results are shown in Table 4 below:
The VGA proposed in this paper adopts the sectional coding rules, which can avoid
illegal solutions in the crossover and mutation operation. The crossover operation with
varying length crossover enables the population to have more potential, retain the fine
quality of the individual and improve the computing capability of the algorithm. At the
same time, the mutation operation combining insertion mutation and reverse sequence
mutation can maintains the diversity of the population and increases the local search
ability of the algorithm to avoid premature phenomena.
Figure 8 shows the average fitness value curve of different algorithms when
reaching the termination condition. Since the VGA that we proposed employs the
mutation operation combining two mutation methods, the DE algorithm with different
mutation rates is used for comparison. For the DE, the optimal value decreases as the
mutation rate increases. In the case of high mutation rate, the optimal value of DE is
obviously larger than that of FGA and VGA. According to the experimental results,
sectional coding and varying length crossover genetic algorithm have faster conver-
gence rate and better value of the objective function than the traditional algorithm.

Table 2. The space position of the machine

Machine Coordinate Machine Coordinate Machine Coordinate
1 (2500, 21000) 9 (19100, 17431) 17 (13650, 3000)
2 (5000, 20000) 10 (20375, 11654) 18 (12000, 4000)
3 (5600, 14500) 11 (9958, 10580) 19 (3600, 5300)
4 (5300, 9800) 12 (10079, 16650) 20 (5600, 5300)
5 (4780, 10000) 13 (10150, 7250) 21 (5600, 4000)
6 (3100, 15000) 14 (15600, 7250) 22 (5600, 3000)
7 (11340, 23450) 15 (19800, 5600) 23 (4300, 3000)
8 (16434, 22341) 16 (19800, 3100) 24 (3600, 3500)
42 H. Sun and X. Zheng

Table 3. Material transport table

Route Weight Route Weight Route Weight Route Weight
1–2 300 11–7 550 16–13 224 24–21 300
1–3 450 9–11 280 17–14 245 19–20 180
3–5 50 12–9 70 18–15 77 19–22 95
4–6 980 7–10 360 18–16 114 4–20 460
2–5 510 9–7 240 20–24 350 4–11 160
1–4 320 13–15 810 20–23 650 11–13 190
4–3 780 13–16 970 21–19 80 13–20 310
11–8 760 13–18 260 22–24 110 20–4 490
8–10 200 14–15 315 23–20 450 13–11 100

Table 4. Experimental results

Algorithm The optimal value Average value Mutation rate
Insertion mutation Reverse mutation
FGA 37875 38132 0.9 0.7
VGA 37579 37796 0.9 0.7
DE 38631 39505 0.9

Fig. 8. The comparison of average fitness value

6 Conclusion

An optimization model is proposed to optimize the tandem AGV workshop with the
goal of minimizing the total distribution time. Only one AGV is responsible for
material transportation in different areas, the coding rules, crossover and mutation
operation of traditional genetic algorithms are improved, and the proposed algorithm is
 Tandem Workshop Scheduling Based on Sectional Coding 43

validated by the example of tandem workshop scheduling problem. Experimental

results show that the average fitness curve of the proposed algorithm falls faster, its
convergence speed is faster, and a more excellent objective value can be obtained,
compared with the fixed length crossover genetic algorithm and DE.
To sum up, the genetic algorithm proposed in this paper is suitable for the selection
of work schemes for the tandem workshop scheduling problem, which can effectively
improves the production efficiency and improves the economic efficiency of the
enterprises.

Acknowledgements. This work was supported by the Guidance Program for Natural Science
Foundation of Liaoning (No. 20170540138).

References
1. Pen C.T., Du Z.J.: Design of AGV based on AT89S52 MCU. Wireless Internet Technology
13 (2017)
2. Bozer, Y.A., Srinivasan, M.M.: Tandem configuration for automated guided vehicle systems
and the analysis of single vehicle loops. IIE Trans. 23, 72–82 (1991)
3. Zhang, H.: Design and implementation of a hybrid scheduling model for tandem workshop
resources in cloud computing environment. CIT 5, 8–11 (2017)
4. Zhou, Q., Liu, J., Wei, F.L.: Single-field tandem workshop scheduling optimization based on
genetic taboo search algorithm. I J. Chang. Univ. Sci. Technol. (Nat. Sci.) 4, 32–38 (2014)
5. Bai, S.F., Tang, D.B., Gu, W.B., Zheng, K.: Research of multiple AGV systems based on
tandem workshop control module. CNEU 3, 8–12 (2012)
6. Tang, D.B., Lu, X.C., Zheng, K.: Research on tandem AGV scheduling based on neuro-
endocrine coordination mechanism. Mach. Build. Autom. 4, 112–115 (2015)
7. Hou, L.Y., Liu, Z.C., Shi, Y.J., Zheng, X.J.: Optimizing machine allocation and loop layout in
tandem AGV workshop by the collaborative optimization method. Neural Comput. Appl. 4,
959–974 (2016)
8. Rezapour, S., Zanjirani-Farahani, R., Miandoabchi, E.: A Machine-to-loop assignment and
layout design methodology for tandem AGV systems with single-load vehicles. Int. J. Prod.
Res. 49, 3605–3633 (2011)
9. Chen, Z.T.: Research and application on job shop scheduling problem based on improved
genetic algorithm, p. 32. Dalian University of Technology, Dalian
 Robust Bi-level Routing Problem
for the Last Mile Delivery Under Demand
and Travel Time Uncertainty

Xingjun Huang1 , Yun Lin1(B) , Yulin Zhu1 , Lu Li1 , Hao Qu1 ,

and Jie Li2
1 College of Mechanical Engineering, Chongqing University, Chongqing, China
[email protected]
2 School of Automotive Engineering, Chongqing University, Chongqing, China

Abstract. Designing the last mile delivery system in a lean way has
become an important part of serving customers efficiently and econom-
ically. However, in practice, the uncertainty in customer demand and
travel times often means vehicles capacity may be exceeded along the
planed route and vehicles miss theses time windows, increasing the cost,
reducing efficiency and decreasing the customer satisfaction. Previous
studies have lacked an uncertainty-based view, and few studies have
discussed how to develop an uncertain model. To address this issue,
the bi-level routing problem for the last mile delivery is formulated
as a robust vehicle routing problem with uncertain customer demand
and travel times. In addition, a modified simulated annealing algo-
rithm is proposed and tested in computational experiments. The results
show that the proposed model has good performance for uncertainty
processing.

Keywords: Last mile delivery · Uncertainty

Modified simulated annealing algorithm · Robust optimization

1 Introduction
The bi-level multisized terminal location-routing problem (BL-MSTLRP ) for
the last mile delivery, as a classical combinatorial optimization problem, first is
introduced by Zhou [8] under the e-commence environment, that aims to find the
optimal set of terminals and the optimal set of routes for a fleet of vehicle deliv-
ering products or service to a given set of customers groups. It is worthy that the
Supported by the National Social Science Fund of China (Grant No.
18BJY066), Fundamental Research Funds for the Central Universities (Grant No.
106112016CDJXZ338825), Chongqing key industrial generic key technological inno-
vation projects (Grant No. cstc2015zdcy-ztzx60009), Chongqing Science and Tech-
nology Research Program (Grant No. cstc2015yykfC60002).
c Springer Nature Singapore Pte Ltd. 2018
K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 44–54, 2018.
https://doi.org/10.1007/978-981-13-2384-3_5
 Robust Bi-level Routing Problem for the Last Mile Delivery 45

BL-MSTRP, as the classical optimization for last mile delivery, assumes that all
the input data, such as customer demands and travels time, are both determin-
istic and are known in advance. However, the solutions derived by deterministic
model are often infeasible when applied to the real-world situation, particularly
in the era of on-time delivery [7]. To tackle problem, this paper proposes robust
bi-level routing problem (RBLRP ) for the last mile delivery under demand and
travel time uncertainty, an extended version of the BL-MSTRP for the last mile
delivery. Note that robust optimization is advantageous due to its computational
tractability and its ability to deal with the practical applications, in which only
partial information or a small amount of historical data about the uncertain
parameters is available [1–4]. In this study, the uncertainty sets of customer
demand and uncertainty sets of travel times of bi-level were derived with partial
information about the distributions of the uncertain parameters and a modified
SA algorithm was proposed to tackle this issue.
The reminder of this paper is organized as follows. Section 2 describes the
considered RBLRP and the process of uncertainty transformation. In addition,
the corresponding robust mathematical model is also presented. In Sect. 3, a
modified SA algorithm is designed for the RBLRP. Section 4 presents a compu-
tational experiment using the modified versions of Zhou ’s instances [8]. Finally,
conclusions are drawn and presented in Sect. 5.

2 Problem Description and Model Formulation

2.1 Problem Description

The RBLRP is defined on a complete undirected digraph G = (N, A), where

the set of vertex is represented by N = {N0 ∪ NC ∪ ND } and the set of the arcs
is represented by A = {AC ∪ AD }. Vertex N0 represents the DC where several
identical vehicles k with capacity Q are located. The set of customer groups is
denoted as NC = {1, 2, ..., nc }, where the distance between the two customer
groups is defined as the AC . Similarly, the set of terminals is denotes as ND ,
where the distance between the two terminals is defined as the AD . For the first
level, vehicles start at the DC and deliver parcels to a set of terminals and the ver-
tex set of the level 1 is denoted by N1 = {N0 ∪ ND }; for the second level, vehicles
deliver parcels from the DC but directly to customer groups, serving individual
customers and the vertex set of the level 2 is denoted by N2 = {N0 ∪ NC }. In
addition, a subset of N is denoted by Nl = {NC ∪ ND }. Customers in each group
can freely choose between HD and CP services. Vehicles serve customers with
HD demand directly, and those with CP demand pick up parcels themselves at
the closest terminals. An example of the RBLRP is illustrated to describe such
situation in Fig. 1, which consists of one DC, four terminals and four customer
groups. Note that customer demand and travel times will fluctuate within a cer-
tain range, and the uncertain customer demand and travel times is modeled as
symmetric and bounded random variables.
46 X. Huang et al.

qiH - qiH , qiH + qiH

d 23

q5 - q5 , q5 + q5
q5

d12

Fig. 1. The problem description of RBLRP with uncertainty for last mile delivery with
HD and CP service programming.

2.2 Uncertain Transformation of Customer Demand

The customer demand uncertainty set Uqki and the travel distance uncertainty
set Udk is defined as follows based on the robust optimization approach [4].

uqi = ×k∈K Uqki , qi = {qiC ∪ qiH } (1)

with ⎧
⎨ 
= q̃i ∈ R|Nl | |q̃i = q̄i + αi q̂i ,
k
Uqki |αi | ≤ Γkqi ,
⎩ (2)
i∈Nlk
  k 
|αi | ≤ 1, Γqi = θqi NC  , ∀i ∈ NC
k

Equation (1) reflects the overall demand uncertainty set uqi of customer groups
i ∈ NC is the Cartesian product of the demand uncertainty set for each vehi-
cle, which consists of CP demand and HD demand. Equation (2) denotes the
uncertain demand of customer groups with HD demand; q̄i represents the nom-
inal value of uncertain demand; and q̂i denotes the maximum deviation from
the nominal value for each vertex i ∈ NC . αi is the auxiliary variable, and
Γkqi is the uncertainty budget that controls the level of uncertain demand of
customer groups. Since the uncertainty experienced by each vehicle is related
k
to the number  customer group visited, Γqi change by at
of customers in each
  k  
most θqi NC  , where  k
  θqi NC represents the smallest integer that is larger
than or equal to θqi NCk . And θqi is the demand uncertainty budget coefficient
of customer groups, and takes value in the interval [0, 1]. If θqi = 0, Γkqi = 0
and q̃i = q̄i – the influence of the demand uncertainty
 of customer groups can
be ignored completely. If θqi = 1, Γkqi = NCk  – each customer demand q̃i of
customer groups can take any value in the interval [q̄i − q̂i , q̄i + q̂i ].
 Robust Bi-level Routing Problem for the Last Mile Delivery 47

In addition, we defined the actual HD demand in customer group i ∈ NC and

the actual CP demand in vertex i ∈ Nl , namely qiH and qiC . Each is defined as
nominal customer demand polytope, as discussed by Zhou et al. [8].

pijCH = f (dij ), i ∈ NC , j ∈ ND (3)

q̃iC = q̃i − q̃iH , ∀i ∈ NC (4)

wk1 + wk2 ≤ 1, ∀k ∈ K (5)

Equation (3) reflects that probability of customers in groups i ∈ NC with initial

CP demand tending to choose HD service when terminal j is selected to serve
it. In Eq. (4), q̃iH denotes the overall uncertain HD demand in customer group
i ∈ NC ; piH and piC are the initial percentages of the two types of services,
respectively, where piH + piC = 1 and zij is a binary variable that 1 if customer
group i choose the terminal j to serve itself; 0, otherwise. Equation (5) show the
overall CP demand in customer group i ∈ NC . Equation (6) guarantee that each
vehicle can only be allocated neither level 1 or level 2.

2.3 Uncertain Transformation of Travel Distance

For the distance within the customer groups, because of the spatial aggregation
feature of customers locations for the last mile delivery, the overall delivery
demand of a customer group keeps constant while individual demand constantly
changes, which is difficult to calculate the tour travel distance for all customers.
Here, the approximate continuous model for customer group (modified from
Zhou et al. [8]) was introduced to model the nominal travel distance L̄ (n) of the
tour in customer group with n customers, which is given by

L̄ (n) = k1 S n/η (6)

qiH
n = nC pI pS · · pE (7)
qiC
where S is the customer group distribution area. η denotes the agglomeration
degree which represents the customer group’s customer density or degree of
aggregation; k1 is a constant, generally k1 = 0.57. n is the number of the cus-
tomers with HD demand in certain customer group; nC is the number of cus-
tomer in a customer group; pI is the probability of the internet users; pS is the
probability of online shopping of customer group; qiH/qiC is the probability that
the customer group selects HD service; pE is the probability that the customer
group buys online every day. Based on the construction process of uncertainty
set of robust optimization, the travel time uncertainty set uL(n) of the tour in
customer group is denoted.

uL(n) = ×k∈K UL(n)

k
(8)
48 X. Huang et al.

with
⎧
⎨ k 
 
k
UL(n) = L̃ (n) ∈ R|NC | L̃ (n) = L̄ (n) + λi L̂ (n) , |λi | ≤ ΓkL(n) ,
⎩ k (9)
i∈NC
   
|λi | ≤ 1, ΓkL(n) = θL(n) NCk  , ∀i ∈ NCk

Equation (9) shows the overall distance uncertainty set uL(n) of the tour in cus-
tomer groups is the Cartesian product of the travel time uncertainty set UdkA of
the tour in a customer group for each vehicle. NCk denotes the set of customer
groups on the route of the vehicle. Similarly, the parameters in the Eq. (10) has
the resemble meanings to those in Eq. (2). NCk represents the nominal value of
uncertain demand L̃ (n) and L̂ (n) denotes the maximum deviation from the
nominal value for each customer group ∀i ∈ NCk . λi is the auxiliary variable,
and ΓkL(n) is the uncertainty budget that controls the level of uncertain travel
distance of the tour in the customer groups and ΓkL(n) is defined as equaling
     k 
θL(n) NCk  , where θL(n)  
 k NC represents the smallest integer that is larger
than or equal to θL(n) NC . And θL(n) is the travel time uncertainty budget

coefficient of the tour in a customer group, and takes value of between 0 and 1.
For the distance between terminals in level 1 and between customer groups in
level 2, without loss of generality, the travel time uncertainty set udA between
terminals and customer groups is given by

udA = ×k∈K UdkA , A = {AC ∪ AD } (10)

with
⎧
⎨ k 
 
UdkA = d˜ ∈ R|A | d˜ij = d¯ij + βij dˆij , |βij | ≤ ΓkdA ,
⎩ (11)
(i,j)∈Ak
  
|βij | ≤ 1, ΓkdA = θdA Ak  , ∀ (i, j) ∈ Ak

Equation (11) shows the overall distance uncertainty set udA is the Cartesian
product of the travel time uncertainty set UdkA for each vehicle. Similarly, in
Eq. (12), Ak denotes the set of arcs on the route of the vehicle k, which consists
of the set between customer groups and the set between terminals. d¯ij repre-
sents the nominal value of uncertain distance d˜ij and d˜ij denotes the maximum
deviation from the nominal value for each arc ∀ (i, j) ∈ Ak . βij is the auxiliary
variable, and ΓkdA is the uncertainty budget that controls the level of uncertain
k
travel time
 of customer  and terminals and ΓdA is defined as equaling
groups
 
θdA A  – where
k   represents the smallest integer that is larger than
 k  θdA A
k

 
or equal to θdA A . And θdA is the distance uncertainty budget coefficient, and
takes value of between 0 and 1.
 Robust Bi-level Routing Problem for the Last Mile Delivery 49

2.4 Mathematical Formulation

The considered RBLRP given the route-dependent uncertainty sets is presented,

starting with the objective function:
     
(RBLRP ) min Z = t̃ij xijk wk1 + t̃ij xijk wk2
i∈N1 j∈N1 k∈K i∈N2 j∈N2 k∈K
  
+ t̃i (n)xijk wk2
i∈Nc j∈N2 k∈K
(12)
Objective (13) minimizes the total travel time, including vehicle travel time
between vertexes in level 1, vehicle travel time between customer groups, travel
time and service time in customer groups of level 2.

xijk ≤ 1, ∀i ∈ ND , k ∈ K (13)
j∈N1

x0jk = 1, ∀k ∈ K (14)
j∈N1

xi0k = 1, ∀k ∈ K (15)
i∈N1

 
xjik − xijk = 0, ∀i ∈ ND , k ∈ K (16)
j∈N1 j∈N1
 
U0uk = q̃jC xijk , ∀u ∈ ND , k ∈ K (17)
i∈N1 j∈ND
 
Ujik − Uijk = q̃iC , ∀i ∈ ND , k ∈ K (18)
j∈N1 j∈N1

Uijk ≤ Q, ∀i, j ∈ N1 , k ∈ K (19)

   
d˜ij xijk /s1 + tjC xijk ≤ Tk , ∀k ∈ K (20)
i∈N1 j∈N1 i∈N1 j∈ND

Ui0k =0, ∀k ∈ K, k ∈ K (21)
i∈N1

Constraints (13)–(22) are used for limiting the vehicle routing of level 1. Con-
straints (14) guarantee that each terminal is served at most once. Constraints
(15)–(17) are the flow conservation constraints, which ensure that vehicle starts
from DC and ends at DC. Constraints (18) ensure calculate the load of vehicles.
Constraints (19) are the flow constraints for demand. Constraints (20) make sure
that remaining demand of a vehicle should not exceed its capacity. Constraints
(21) stipulate that vehicle’s maximum working hours should not be exceeded.
Constraints (22) ensure that the remaining demand of a vehicle is zero after
50 X. Huang et al.

serving the final vertex. Note that constraints (20)–(22) are both related to both
level 1 and level 2.

xijk = 1, ∀i ∈ N2 , k ∈ K (22)
j∈NC

zim = 1, ∀i ∈ NC (23)
m∈ND

x0jk = 1, ∀k ∈ K (24)
j∈N2

xi0k = 1, ∀k ∈ K (25)
i∈N2
 
xjik − xijk = 0, ∀i ∈ NC , k ∈ K (26)
j∈N2 j∈N2
 
U0uk = q̃jH xijk , ∀u ∈ NC , k ∈ K (27)
i∈N2 j∈NC
 
Ujik − Uijk = q̃iH , ∀i ∈ NC , k ∈ K (28)
j∈N2 j∈N2

Uijk ≤ Q, ∀i, j ∈ N2 , k ∈ K (29)

 
  
 
d˜ij xijk s1+ xijk k1 S ni /ηi  s2 + q̃iH tH ≤ Tk , ∀k ∈ K
i∈N2 j∈N2 i∈NC i∈NC

 (30)
q̃iC zim ≤ bm ym , ∀k ∈ K, m ∈ ND (31)
i∈NC

Constraints (23)–(32) are used for limiting the level 2. Constraints (23) guaran-
tee that each customer group is served exactly once by a level 2. Constraints (24)
ensure that each customer group is assigned to a terminal exactly once. Con-
straints (25)–(31) have a similar meaning with constraints (15)–(21). Constraints
(32) make sure that the terminal’s capacity cannot be exceeded by customers
assigned.

3 The Modified SA Algorithm for the RBLRP

3.1 Solution Representation and Encoding

In this work, permutation representation is used and the splitting procedure [5,6]
is integrated into the encoding strategy. As far as the chromosome of level 1 and
level 2, the DCs, denoted as 0 between the head and tail of the chromosome,
are generated at random, aiming to split the giant tour into some small routes
without violating the vehicle capability and working time constraints.
 Robust Bi-level Routing Problem for the Last Mile Delivery 51

3.2 The Neighborhood Search Structure

For the SA heuristic, we employ a random neighborhood structure that features

three types of moves [4], including intra-route swap, intra-route relocate and
intra-route 2-opt. The local optimum solution of Rsearch is generated from the
solution of Rinitial based on the neighborhood search operators and a graphic
illustration of the three neighborhood operators appears in Fig. 2.
Intra-route Swap. Intra-route swap is executed by randomly selecting the i th
and j th positions located in the same chromosome and therefore exchanging the
genes in the two positions.
Intra-route Relocate. We implement the move by randomly selecting the j th
position, remove it from the chromosome and reinsert it into the middle of the
j th and j-1 th positions of the same chromosome.
Intra-route 2-opt. This neighborhood operator is executed by randomly select-
ing the i th and j th positions located in the same chromosome and reverse the
order of (i, j).

i1 i i+1 i1 i+1 i1 i i+1

j1 j j+1 j j1 j1 j j+1

(a) Intra-route swap. (b) Intra-route relocate. (c) Intra-route 2-opt.

Fig. 2. Three neighborhood operators.

3.3 The Modified SA Method

As is shown in the Algorithm 1, the modified SA procedure is depicted. For

the parameters setting, initialize the terminate temperature Tmin and initial
temperature T that is calculated, T = 100 ∗ n, where n is the number of the
vertexes. For the acceptance decision, the Metropolis criterion is used in the
local search step. Generate a neighborhood solution Rsearch from Rcurrent and
calculate its fitness function. The solution generated form the local search Rsearch
is compared to the current solution Rcurrent . If the fitness function difference
of Rsearch and Rcurrent , namely , is greater than 0, the Rsearch replaces the
current solution Rcurrent . Otherwise, Rsearch is accepted with the probability of
acceptance exp(−Δ/T) when exp(−Δ/T) is greater than 0–1 random number.
Implement the iterations until i = iter. If the temperature is greater than Tmin ,
the current temperature is updated by T = T ∗ r, where r is the temperature
update factor.
52 X. Huang et al.

Algorithm 1. The procedure of modified SA

1: Define neighborhood research structure Nsearch J , J = 1, · · ·, Jmax ;
2: Set the solution of giant tour as the initial solution Rinitial and improve Rinitial by local search;

3: Set current solution Rcurrent ← Rinitial ;

4: repeat
5: repeat
6: {Local search}
7: Generate a local optimal solution Rsearch ∈ Nsearch
J (Rcurrent );
8: {Acceptance decision} STATE Δ = F (Rsearch ) − F (Rcurrent );
9: if Δ < 0 then
current
10: R ← Rsearch ;
11: else
12: if exp(−Δ/T) > rand then
current
13: R ← Rsearch ;
14: end if
15: end if
16: i←i+1
17: until i ≤ iter
18: T = T ∗ r (r < 1);
19: until T < Tmin

4 Computational Experiments

4.1 Experiment Description and Parameter Setting

The computational experiment instance was derived from the latest research
instance of Zhou [8] designed for the BL-MSTLRP with deterministic customer
demands and travel times. To simplify the calculation, the datasets of terminal
size information and vehicle type information were ignored. The speed of vehicles
is 30 km/h between nodes and 20 km/h in customer groups with working time
about 8 h. In addition, the customer demand transfer coefficient is subject to
distribution as follows.
⎧
⎪ 0, 0 ≤ dij ≤ 500
⎪
⎨
pijCH = f (dij ) = 0.3, 500 < dij ≤ 800 (32)
⎪
⎪ 0.7, 800 < dij ≤ 1000
⎩
1, dij > 1000
For the robust purpose, the nominal value q̄i (i ∈ NC ) and t̄ij ((i, j) ∈ A) was
assumed to equal to the corresponding customer demand and travel time in
case set data from the Zhou, respectively. Moreover, we assumed that the max-
imal demand deviation q̂i was 0.2q̄i , and the maximal travel time deviation t̂ij
was 0.2t̄ij and θqi = θL(n) = θdA was 0.6. For the SA, The terminal tempera-
ture Tmin = 0.001, maximal evolutional generation Iiter = 100 and temperature
update factor r = 0.99. Finally, the computational experiments were imple-
mented using MATLAB R2014a, and the experiments on a computer with an
AMD A8–7650K Radeon R7 CPU at 3.30 GHz and 8 GB RAM under Windows
X64 system.
 Robust Bi-level Routing Problem for the Last Mile Delivery 53

4.2 Computational Results

The result of comparison with two strategies is shown in the Table 1. Compared
with the deterministic solution, the robust strategy will result in a lower vehicle
travel time, which is reduced by 1.63%. This is mainly because the disturbance
of the uncertainty factor makes the distance from the customer to the terminal
increase, and customers also prefer CP services. This can be seen from the aver-
age load rate of vehicles transferred from (54.78%, 89.66%) to (55.41%, 84.61%).
To examine the impact of customer demand and travel time uncertainty, two sets
of experiments were added, Rob-T and Rob-D. One interesting finding is that
travel time uncertainty has a stronger influence on deterministic solution strat-
egy than the demand uncertainty, for example, 1.60% vs 0.15%. This is mainly
because the uncertainty of demand does not affect the choice of customer service,
while the uncertainty of the path directly leads to the shift of customer service
selection. Finally, The result of robust optimization is shown in the Fig. 3.

Table 1. Comparison of two solutions strategies on the considered instances

Type NO. N.V. OBJ Reduce Avg.loadrate Level-1 Level-2

Sobj11 Sobj12 Sobj21 Sobj22 Sobj23 Sobj24
Det 2\4 6 28.1019 - 54.78%\89.66% 4.7693 0.5270 4.5405 5.9162 6.1567 6.1921
Rob 2\4 6 27.6440 1.63% 55.41%\84.61% 1.2462 4.0485 7.2201 5.6179 4.3403 5.1710
Rob-T 2\4 6 27.6512 1.60% 57.72%\88.14% 0.8779 4.4156 5.1710 6.5562 5.6711 4.9595
Rob-D 2\4 6 28.0595 0.15% 52.59%\86.07% 4.7624 0.5355 5.7462 5.8996 6.5549 4.5609

1600

1400

1200

1000
Y-coordinate

800

1
600

400
Terminal

200 Distribution center

Customer groups

0
0 500 1000 1500 2000 2500 3000
X-coordinate

Fig. 3. The best solution of robust optimization of the instance.

5 Conclusion
The last mile delivery problem has gained widespread attention in recent years,
but the previous studies lack enough analysis of the uncertain phenomenon. This
paper focus on the research on the uncertainty optimization of the last mile. We
found that robust optimization performed outstandingly on uncertainties and
also had good performance in dealing with large-scale network uncertainties.
54 X. Huang et al.

Future research should consider more uncertainty and establish a last-mile

optimization model that is closer to reality. In addition, future research should
also pay more attention to empirical research and enrich relevant research results.

References
1. Ben-Tal, A., Ghaoui, L., Nemirovski, A.: Robust Optimization (2009)
2. Bertsimas, D., Sim, M.: The price of robustness. Oper. Res. 52(1), 35–53 (2004).
https://doi.org/10.1287/opre.1030.0065
3. Han, J., Lee, C., Park, S.: A robust scenario approach for the vehicle routing problem
with uncertain travel times. Transp. Sci. 48(3), 373–390 (2014). https://doi.org/10.
1287/trsc.2013.0476
4. Hu, C., Lu, J., Liu, X., Zhang, G.: Robust vehicle routing problem with hard time
windows under demand and travel time uncertainty. Comput. Oper. Res. 94, 139–
153 (2018). https://doi.org/10.1016/j.cor.2018.02.006
5. Kloimüllner, C., Papazek, P., Hu, B., Raidl, G.R.: A cluster-first route-second
approach for balancing bicycle sharing systems. In: Moreno-Dı́az, R., Pichler, F.,
Quesada-Arencibia, A. (eds.) EUROCAST 2015. LNCS, vol. 9520, pp. 439–446.
Springer, Cham (2015). https://doi.org/10.1007/978-3-319-27340-2 55
6. López-Santana, E., Rodrı́guez-Vásquez, W., Méndez-Giraldo, G.: A hybrid expert
system, clustering and ant colony optimization approach for scheduling and routing
problem in courier services. Int. J. Ind. Eng. Comput. 9(3), 369–396 (2018). https://
doi.org/10.5267/j.ijiec.2017.8.001
7. Taş, D., Dellaert, N., Van Woensel, T., De Kok, T.: Vehicle routing problem with
stochastic travel times including soft time windows and service costs. Comput. Oper.
Res. 40(1), 214–224 (2013). https://doi.org/10.1016/j.cor.2012.06.008
8. Zhou, L., Lin, Y., Wang, X., Zhou, F.: Model and algorithm for bilevel multisized
terminal location-routing problem for the last mile delivery. Int. Trans. Oper. Res.
(2017). https://doi.org/10.1111/itor.12399
 A Comprehensive Fault Diagnosis System
and Quality Evaluation Model
for Electromechanical Products by Using
Rough Set Theory

Jihong Pang1,3(&), Ruiting Wang1, and Yan Ran2

1
College of Mechanical and Electronic Engineering, Wenzhou University,
Wenzhou 325035, China
[email protected]
2
College of Mechanical Engineering, Chongqing University, Chongqing
400030, China
3
College of Mechanical Engineering, Zhejiang University, Hangzhou 310027,
China

Abstract. Electromechanical product is an important part of mechanical and

electrical control system, and its quality plays a key role in the normal operation
process. In this paper, a comprehensive fault diagnosis system and quality
evaluation model for electromechanical products is analyzed. Firstly, the feature
extraction of different faults is carried out, and the fault features of electrome-
chanical products are simplified by using the approximation set information
system properties of rough set theory. Secondly, the subjective weight index
model is determined based on the rough information system properties of rough
set theory. Then, the evaluation weight of each index of quality evaluation
model for electromechanical products is obtained by the importance measure-
ment of information system properties. Finally, this paper illustrates that the
results of fault diagnosis and quality evaluation of ball valves as well as the
availability of scientific.

Keywords: Fault diagnosis system
Quality evaluation model

Electromechanical products
Rough set

1 Introduction

Reliability and quality performance for electromechanical products will directly affect
the stability of the operational stability. Thus, every company should make the most of
its sources and regulating the cost from time to time to product quality [1]. It is
introduced that during the manufacturing process of electromechanical products, the
fault diagnosis system and quality evaluation of their key procedures must be

J. Pang—This work was supported by the National Natural Science Foundation of China
(No. 71671130, No. 71301120, No. 51705048).

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 55–64, 2018.
https://doi.org/10.1007/978-981-13-2384-3_6
56 J. Pang et al.

strengthened to ensure the qualification and stability to materials and workmanship [2].
The fault diagnosis system is a key part of product quality evaluation, and product
working performance is affected by the processing quality in through their manufac-
turing process [3]. Moreover, the quality evaluation goals can be translated into a set of
preference or utility functions that represent the value of alternative sets of conse-
quences by decision theorists [4]. Through introducing the reliability specifications and
quality test methods specified for electromechanical products, the basic fault diagnosis
and quality control requirements can be understood by the application [5]. This paper
presents a fault diagnosis system for alarm processing and fault diagnosis in product
manufacturing process.
On the other hand, a comprehensive fault diagnosis system and quality evaluation
model for electromechanical products is the basis of scientific decision-making. Fault
diagnosis system is an important part for quality evaluation model, a comprehensive
evaluation with rough set theory can improve the electromechanical products quality
[6]. Furthermore, combination with rough sets and decision-making theory has enor-
mous development potentials in the field of fault diagnosis system and quality evalu-
ation for electromechanical products [7]. To ensure the accuracy and stability of quality
evaluation, this comprehensive fault diagnosis system sets up fault recognition
mechanism before acting [8]. Domestic and foreign scholars have done a lot of research
work on comprehensive quality evaluation methods. In the next, under the framework
of rough set theory, this study presents a multi-stage innovation quality evaluation
model with the theory of rough set analysis. A quality evaluation model for identifi-
cation, diagnosis, estimation and forecast of fault diagnosis system using the data
fusion technique and the rough set theory was established [9]. This paper describes how
to apply rough set theory in quality evaluation model validity of fault diagnosis sys-
tems. Objective weighting evaluation method based on actual data source to make
quantitative evaluation of indicators, the weight is determined according to the cor-
relation or variation coefficient of each index.
Considering the above aspects, a new method for quality evaluation and its
application in ball valve fault diagnosis with rough set theory is proposed in this study.
This study will be arranged as follows. The implement process comprehensive fault
diagnosis system and quality evaluation model for electromechanical products by using
rough set theory in the following section. In the case study, a testing attribute evalu-
ation is set up, the attributes of t quality evaluation model are calculated, and the
experimental data are analyzed to verify or identify the model’s key attributes for a
high fault diagnosis system. In the last section, some conclusions on the approved
diagnosis system and evaluation model are given.

2 Implement Processes by Using Rough Set Theory

Rough set theory is an effective tool to deal with vagueness and uncertainty. It provides
effective means for people to understand things correctly and make scientific decisions.
The rough set theory was proposed by Poland scientist in 1982, and after the contin-
uous development of scientific researchers on rough set theory [10]. Rough set theory is
a new mathematical tool with strong knowledge obtaining ability after probability
 A Comprehensive Fault Diagnosis System and Quality Evaluation Model 57

theory, fuzzy sets, mathematical theory of evidence [11]. At present, rough set theory
has been applied to various fields, such as machine learning, fault diagnosis, control
algorithm acquisition, process control and knowledge acquisition in relational database,
and has achieved great success [12].
In brief, the rough set theory and method is quite an efficient means to deal with
complex system, which is applied to a lot of fields. An algorithm of attribute reduction
based on attribute importance of rough set was used to fault diagnosis system [13]. An
extraction method of decision rules for fault diagnosis by using rough set theory is
presented to extract simple and effective diagnostic rales for quality evaluation in this
paper. The implement process of fault diagnosis system and quality evaluation model
for electromechanical products by using rough set theory is shown in Fig. 1.

Comprehensive fault diagnosis system and quality

evaluation model for electromechanical products

Collecting fault information of electromechanical products

Determining information system and knowledge reduction

Confirming approximation set and weight determination

Decisioning dependency of attributes

Calculating importance of attributes

Making simplification and classification

Importance measurement of information system properties

Determination of index weight by using rough set theory

Making quality evaluation with the result of computation

Fig. 1. Implement process of quality evaluation by using rough set theory

58 J. Pang et al.

2.1 Information System and Knowledge Reduction

Knowledge reduction is the core of rough set theory. In this paper, the association and
their relation between information system and knowledge reduction were described.
A new object of value reduction is proposed to get the rules with maximal support
measurement. Rough set theory has two important objectives of classification and
reduction [14].
Let S ¼ ðU; A; V; f Þ is an information system, where U represents a nonempty
finite set of objects, U ¼ fx1 ; x2 ; . . .; xn g is called a domain. A represents a nonempty
finite set of attributes, A ¼ fa1 ; a2 ; . . .; am g, aj ðj  mÞ is called an attribute. Va is the
range of the attribute a, f: U  A ! V is an information function for each attribute of
object given a value of information, namely a 2 A; x 2 U; f ðx; aÞ 2 Va.
Step one: Determining information system and knowledge reduction
A rough set model to mine default rules was used to solve the decision question with
incomplete information. Rough set theory is a new mathematical tool to deal with
fuzzy and uncertain knowledge [15]. In the process of classification, individuals with
little difference are attributed to the same class, and their relationship is an indistin-
guishable relationship, which is shown as:

INDðPÞ ¼ fðx; yÞ 2 U  U : f ðx; aÞ ¼ f ðy; aÞ; a 2 Pg ð1Þ

Step two: Confirming approximation set and weight determination

The next approximation of set X about I is the largest set of objects, which must be
based on the existing knowledge belongs to X, and remember to make POS(X), the
specific calculation formula is as follows (2):

I ðXÞ ¼ fx 2 u : IðxÞ  Xg ð2Þ

The upper approximation of set X with respect to I is x-intersecting non-null

equivalent I(X), the following formula can be used as:

I ðXÞ ¼ fx 2 u : IðxÞ \ X 6¼ /g ð3Þ

Step three: Decisioning dependency of attributes

The decision attribute set D depends on the conditional attribute set C.

jPOSc ðDÞj
cc ðDÞ ¼ ð4Þ
jUj

Step four: Calculating importance of attributes

Different attributes play a different role in determining the degree of dependency
between the decision attributes. Add attribute a to the conditional attribute set R, and
define the importance of classification U/IND(P):
 A Comprehensive Fault Diagnosis System and Quality Evaluation Model 59

sgf ða; R; PÞ ¼ cR ðPÞ  cRfag ðPÞ ð5Þ

Step five: Making simplification and classification

In a decision-making system, there are often some dependencies and classifications
between the conditional attributes. Simplification can be understood as the easiest way
to represent the decision attribute of the decision system to the set of condition
attributes without losing the information Dependencies and classification.
A given decision system S ¼ fU; A ¼ C [ D; V; f g. The simplification of the
conditional attribute C is a nonempty subset C″ of C,
(1) INDðC0 ; fdgÞ ¼ INDðC; fdgÞ.
(2) This explains that INDðC 00 ; fdgÞ ¼ INDðC; fdgÞ.
It satisfies that all the reduced sets of C is denoted RED (C), and the intersection of
all reduced sets of C is called a kernel, denoted CORE (C), CORE (C) = \ RED (C).
Therefore, any simplification B of C can be used instead of C without losing any
information in the information table, resulting in a simplified message.

2.2 Importance Measurement of Information System Properties

Based on the information system of rough set theory, we can suppose that
U=INDðPÞ ¼ fX1 ; X2 ; . . .; Xn g. And P is defined as:

X
n
jXi j jXi j 1 X n
IðPÞ ¼ j1  ¼1 jXi2 j ð6Þ
i¼1
jUj jUj 2
jUj i¼1

Where X represents the cardinal number of the set X, Xi/U represents the relative
cardinality of the equivalent class Xi in U.
And the importance of attribute Aa is defined as:

SGFAfag ðaÞ ¼ IðAÞ  IðA  fagÞ ð7Þ

The above definition shows the importance of attribute a in A is a change in the

amount of information caused by A removed after {a} size measurement.

2.3 Determination of Index Weight by Using Rough Set Theory

Due to the shortcomings of the single evaluation method, the subjective weight is
organically combined by the rough set theory in this paper.
And the attribute of ai 2 A is representative of SGFA  fai g  ðai Þ, ai 2 A are
defined as weight:

X
m
xðai Þ ¼ SGFAfai g ðai Þn SGFAfai g ðai Þ ð8Þ
i¼1
60 J. Pang et al.

3 A Case Study

A method of fault diagnosis of ball valve by using rough set theory is proposed in the
case study. A ball valve is widely used in the pneumatic and hydraulic control system,
which controls the circulation of the medium in the pipe through opening and closing
of the ball valve. The existing problems of manufacturing process of ball valve were
analyzed, including unstable machining quality, imperfect technics, incongruous pro-
cess design and low assembly quality. The comprehensive evaluation method consists
of objective weighting method and subjective weighting method. Rough set theory was
introduced to solve this problem, and further the quality evaluation model for fault
diagnosis ability was proposed.
In this paper, the comprehensive evaluation method is used to evaluate the index of
the quality of the ball valve. With the use of rough set theory, the operational quality of
ball valve was improved and optimized. In the process of work, if the valve leakage or
failure cannot start normally, it will cause the system fail to achieve the prescribed
workflow and affect the environmental pollution. Typical application of the ball valve
of quality improving and enhancement of processing capacity were introduced in this
case study. Rough set theory is applied to evaluate the quality of ball valve. The index
that affects the quality of ball valve is determined through comprehensive index weight
values, and the key part is controlled during the production process of ball valve
quality. To realize the complementary advantages of uncertainty, the final compre-
hensive evaluation result is more reasonable.
To determine the reliability correction factor, it is necessary to comprehensively
consider the composition structure, manufacture and assembly, and use of the product.
However, the above indicators are too general and difficult to quantify in practice.
Therefore, it is necessary to divide the above several large segments into easily
quantified indicators, establish a more comprehensive index system to evaluate the
degree of difference.
In view of the quality characteristics and consulting related manuals and large
number of failure data, the ball valve can be obtained the factors which affecting the
quality of life cycle. Then, the comprehensive evaluation hierarchy model of numerical
control machine tool subsystem quality correction factor is established, which is shown
in Fig. 2.
The layered model consists of three layers, the top layer is the target layer, the
middle layer is the criterion layer and the bottom layer of the index layer. Rule layer
contains leakage of packing, leakage of main seal, side snapping or bite motion,
abnormal pneumatic actuator. And each factor contains number of factors for the
specific comparative indicators, which is used to build the hierarchical model of index
layer. Then, the simplification of rough set theory to the index system is described
below. Firstly, the indexes of two level in Fig. 2 are set as the attributes of the
information system, that is, A ¼ fX1 ; X2 ; . . .; X11 g. Ball valve can be evaluated as a
collection of objects in the system, for the convenience of processing, there are eight
ball valves, namely U = {I, II, …, VIII}.
In the evaluation process, the corresponding evaluation results of each secondary
index are measured in four grades, which is shown in Table 1.
 A Comprehensive Fault Diagnosis System and Quality Evaluation Model 61

Fig. 2. Quality evaluation index system of ball valve

Table 1. Information system of ball valve

X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11
I 3 2 2 1 2 3 1 2 2 1 1
II 1 4 2 3 2 1 3 2 2 2 2
III 2 2 4 3 1 2 2 4 4 1 2
IV 4 2 3 2 1 4 2 3 3 3 3
V 1 4 2 3 3 1 3 2 2 2 2
VI 2 2 3 2 1 2 2 3 3 3 3
VII 4 2 4 2 1 4 1 4 3 3 3
VIII 3 4 2 1 2 3 1 2 2 1 1

In the next, the information system S = (U, A, V, f) of rough set is simplified, and a
simple set is R1 = {X1, X2, X5, X7, X9, X11}. And the Table 1 corresponding to the
information system S can be further reduced to Table 2.
From Table 2, comprehensive assessment standards are namely “excellent”,
“good”, “general” and “poor”, and can be respective by 4, 3, 2, 1, which are used to
build an information system S of rough set. And the corresponding level one and two
levels index sets are {F1, F2, F3, F4} = {{X1, X2}, {X5}, {X7}, {X9, X11}}.
Then, the computing process of secondary classes is simplified with the framework
among U = {I, II, III, IV, V, VI, VII, VIII} and A = {X1, X2, X5, X7, X9, X11}.
62 J. Pang et al.

Table 2. The brief information system

X1 X2 X5 X7 X9 X11
I 3 2 2 1 2 1
II 1 4 2 3 2 2
III 2 2 1 2 4 2
IV 4 2 1 2 3 3
V 1 4 3 3 2 2
VI 2 2 1 2 3 3
VII 4 2 1 1 3 3
VIII 3 4 2 1 2 1

Let F1 = {X1, X2}, then we can get U/IND(F1) = {{I}, {II, V}, {III, VI}, {IV, VII},
{VIII}} through the Formula (6):

12 þ 22 þ 22 þ 22 þ 12 50
IðF1Þ ¼ 1  ¼ :
82 64

So we have a hypothesis that U/IND(F1 − {X1}) = {{I, III, IV, VI, VII}, {II, V,
VIII}} is calculated with Formula (6):

52 þ 32 30
IðF1  fX1gÞ ¼ 1  ¼ :
82 64

Next, the calculating results can be get through the Formula (7):

50  30 20
SGFF1fX1g ðX1) = I(F1)  I(F1  fX1gÞ¼ ¼ :
64 64

So similarly, the results can be obtained:

50  48 2
SGFF1fX2g ðX2) = I(F1)  I(F1  fX2gÞ¼ ¼ :
64 64

Then, the weight can be counted through the Formula (8):

X
2
20
xðX1Þ ¼ SGFF1fX1g ðX1Þn SGFF1fXig ðXiÞ ¼  0:9091:
i¼1
20 þ 2

In much the same way, the other weights can be obtained as:

xðX5Þ ¼ 1:0000; xðX5Þ ¼ 1:0000:

xðX9Þ ¼ 0:3333; xðX11Þ ¼ 0:6667:

 A Comprehensive Fault Diagnosis System and Quality Evaluation Model 63

xðF1Þ ¼ 0:4000; xðF2Þ ¼ 0:2000:

xðF3Þ ¼ 0:2000; xðF4Þ ¼ 0:2000:

After that, depending on the results, we can meet with the comprehensive fault
diagnosis system and quality evaluation model of ball valve to discuss further action.
Data collection module is the base of the condition monitoring and quality evaluation
model in the manufacturing process. Next, we point out the merits and main problems
of the ball valve quality evaluation model. Development of the multiparameter con-
dition monitoring and quality evaluation model for ball valve. The experimental results
demonstrate that the proposed fault diagnosis system and quality evaluation model is
correct and effective.

4 Conclusions

This paper studies the decision rules acquisition methodology by using rough set theory
for fault diagnosis and quality evaluation. The application of comprehensive quality
evaluation model of ball valve in a manufacturing firm was discussed. This study offers
one new approach for automatic diagnosis rules knowledge in intelligent quality
evaluation. Rough set theory not only can enrich intelligent decision theory, but also
effectively provides the reference to the research of relative areas. Moreover, a fault
diagnosis system and quality evaluation model for electromechanical products is
developed based on the information fusion technology and rough set theory. Examples
show that the evaluation is feasible effective and provides a new way for the evaluation
of the quality of the ball valve. The method of fault diagnosis and quality evaluation
model can be also prospected for other electromechanical products by using rough set
theory.

Acknowledgments. This work was supported by the National Natural Science Foundation,
China (No. 71671130, No. 71301120, No. 51705048), the postdoctoral program of Zhejiang
University & Zhejiang Linuo Fluid Control Technology Company of China (No. 174102), the
key project of teaching reform of Wenzhou University (No. 15jg07), engineering practice edu-
cation center of Wenzhou University & Zhejiang Linuo Fluid Control Technology Company of
China.

References
1. Bassi, F., Clerici, R., Aquario, D.: Evaluating quality of the didactics at university: the
opportunities offered by latent class modeling. TQM J. 30(2), 168–180 (2018)
2. Huang, H.B., Li, R.X., Yang, M.L., Lim, T.C., Ding, W.P.: Evaluation of vehicle interior
sound quality using a continuous restricted Boltzmann machine-based DBN. Mech. Syst.
Sig. Process. 84, 245–267 (2017)
3. Oliveira, J.C.M., Pontes, K.V., Sartori, I., Embirucu, M.: Fault detection and diagnosis in
dynamic systems using weightless neural networks. Expert Syst. Appl. 84, 200–219 (2017)
64 J. Pang et al.

4. Vidal Legaz, B., Maia De Souza, D., Teixeira, R.F.M., Anton, A., Putman, B., Sala, S.: Soil
quality, properties, and functions in life cycle assessment: an evaluation of models. J. Clean.
Prod. 140, 502–515 (2017)
5. Ma, C., et al.: Sound quality evaluation of the interior noise of pure electric vehicle based on
neural network model. IEEE Trans. Ind. Electron. 64(12), 9442–9450 (2017)
6. Gao, Z., Chin, C.S., Chiew, J.H.K., Jia, J., Zhang, C.: Design and implementation of a smart
lithium-ion battery system with real-Time fault diagnosis capability for electric vehicles.
Energies 10(10), 1503 (2017)
7. Zhao, R., Li, C., Tian, X.: A novel industrial multimedia: rough set based fault diagnosis
system used in CNC grinding machine. Multimed. Tools Appl. 76(19), 19913–19926 (2017)
8. Moayyedian, M., Abhary, K., Marian, R.: Optimization of injection molding process based
on fuzzy quality evaluation and Taguchi experimental design. CIRP J. Manuf. Sci. Technol.
21, 150–160 (2018)
9. Nazari, R., Seron, M.M., De Dona, J.A.: Actuator fault tolerant control of systems with
polytopic uncertainties using set-based diagnosis and virtual-actuator-based reconfiguration.
Automatica 75, 182–190 (2017)
10. Chiaselotti, G., Ciucci, D., Gentile, T., Infusino, F.: Rough set theory and digraphs. Fundam.
Inform. 153(4), 291–325 (2017)
11. Gardiner, E.J., Gillet, V.J.: Perspectives on knowledge discovery algorithms recently
introduced in chemoinformatics: rough set theory, association rule mining, emerging
patterns, and formal concept analysis. J. Chem. Inf. Model. 55(9), 1781–1803 (2015)
12. Hu, D., Yu, X., Wang, J.: Statistical inference in rough set theory based on Kolmogorov-
Smirnov goodness-of-fit test. IEEE Trans. Fuzzy Syst. 25(4), 799–812 (2017)
13. Huang, C.-C., Tseng, T.-L., Tang, C.-Y.: Feature extraction using rough set theory in service
sector application from incremental perspective. Comput. Ind. Eng. 91, 30–41 (2016)
14. Khan, M.A.: A probabilistic approach to rough set theory with modal logic perspective. Inf.
Sci. 406–407, 170–184 (2017)
15. Yao, Y., Zhang, X.: Class-specific attribute reducts in rough set theory. Inf. Sci. 418–419,
601–618 (2017)
 The Research of Improved Wolf Pack
Algorithm Based on Differential Evolution

Yingxiang Wang(&), Minyou Chen, Tingli Cheng,

and Muhammad Arshad Shehzad Hassan

State Key Laboratory of Power Transmission Equipment and System Security

and New Technology, School of Electrical Engineering, Chongqing University,
Chongqing 400044, China
[email protected]

Abstract. Aiming at the problems of traditional wolf pack algorithm (WPA):

easy to fall into local optimal, large computational resource cost and low
robustness, an improved wolf pack algorithm based on differential evolution
(DIWPA) is proposed. By introducing the search factor for search wolves,
maximum number of raid wolves, adaptive siege step size and differential
evolution strategy, the proposed algorithm can not only reduce the computa-
tional cost but also improve the global search ability. The DIWPA is used to
conduct optimization test on 12 benchmark functions and compare to 3 typical
optimization algorithms. The test results show that DIWPA has great robustness
and global search ability, especially has excellent performance in multi-peak,
high-dimension, indivisible functions.

Keywords: Wolf pack algorithm
Local optimal
Differential evolution

Robustness
Global search ability

1 Introduction
1.1 A Subsection Sample
Swarm intelligence algorithms such as Particle Swarm Optimization (PSO) [1–3], Fish
Swarm Optimization (FSA) [4, 5], Ant Colony Optimization (ACO) [6], Differential
Evolution Algorithm (DE) [7], are widely used to solve optimization problems and
have their own merits in solving different problems. The wolf pack search (WPS) was
proposed by Yang et al. [8] in 2007. The Wu et al. [9] proposed a new Wolf Pack
algorithm (WPA) on the basis of analyzing the behavior of hunting and prey among
wolves in 2013. The WPA simulates the characteristics of wolves and groups to
collaboratively hunt. In addition, the abstract behavior of walking, summoning, raids,
siege, “winner is king” head wolf generation mechanism and the “strong survival”
group update their mechanism to solve the optimization problems. Although the
algorithm has a short time of introduction due to its good performance in multi-peak,
high-dimension functions. In addition, it has been widely used in three-dimensional
underwater track planning [10], hydropower reservoir optimization scheduling [11] and
other human production activities. However, there are inevitably some shortcomings:

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 65–76, 2018.
https://doi.org/10.1007/978-981-13-2384-3_7
66 Y. Wang et al.

easy to fall into local optimal solution, large computational resource cost and low
robustness. Authors in [12] proposed an improved wolf pack algorithm, which pro-
posed the wolf updating rules and introduced the phase factor according to the basic
idea of the traditional wolf pack algorithm. At the same time, it optimized the types of
step lengths of the traditional wolves and designed a new wolves. The location update
formula is validated by the test function simulation. Literature [13] proposed a cultural
wolf pack algorithm combined with cultural algorithm. This algorithm can effectively
solve the blindness problem of artificial wolf search. Through testing and analyzing
three complex functions, the validity of the culture wolf pack algorithm is verified. In
Literature [14], the idea of solving the current local optimum in the PSO algorithm was
introduced into the walking and summoning behavior of the wolf pack algorithm, and
the suboptimal solution obtained by the chaos method is optimized. The improved
algorithm greatly improves the accuracy of the search and avoids falling into local
optimal. The effectiveness of the algorithm is verified by simulation.
Aiming at the deficiencies of WPA, this paper proposes an improved wolf pack
algorithm based on differential evolution (DIWPA). Based on the intelligent behavior
improvement of WPA, DIWPA introduces crossover, mutation and selection in dif-
ferential evolution, which greatly improves the optimization performance and robust-
ness of the algorithm.
This paper introduces the basic principle of DIWPA firstly. Then, DIWPA is used
to conduct optimization test on 12 benchmark functions and compare to 3 typical
optimization algorithms. The test results show that DIWPA has good performance for
all kinds of test function optimization, and can avoid premature algorithm.

2 Basic Principle of the DIWPA

The DIWPA inherits all the behavior of the WPA and improves it. The main
improvements are as follows:
① Due to the lack of guidance in wandering in WPA, the algorithm is unable to fully
search the solution space and easily fall into a local optimum. The DIWPA
introduces the search factor to the search wolves to make the algorithm traverse the
entire space.
② In the face of high-dimensional function optimization, it may inevitably consume a lot
of computing resources and may be difficult to jump out of the optimization process
and reduce the robustness in WPA. The DIWPA introduces the maximum number of
raid wolves to improve computational efficiency and algorithm robustness.
③ All the artificial wolves will siege at the same determined step size in siege
behavior in WPA, so the adaptive siege step length is introduced in the siege
behavior and the differential evolution strategies is introduced to improve the
global search capabilities in DIWPA.
 The Research of Improved Wolf Pack Algorithm Based on Differential Evolution 67

In summary, the DIWPA consists of seven kinds of behaviours such as walking,

summoning, raid, siege, mutation, crossover, and selection, as well as two kinds of
mechanisms, such as the generation of the wolf with the winner as the king and the
group renewal with the survival of the strong.
The following minimum optimization problem as an example to introduce the main
process of the DIWPA algorithm in detail:
(1) Initialization
Using the formula (1) to initialize the wolves, producing n artificial wolfs:


Xit ¼ XiL þ rand  XiU  XiL ; i ¼ 1; 2;


; n ð1Þ

where Xit represents the ith artificial wolf in the tth generation; X L and X U are the lower
and upper bounds of the variable X ¼ ðx1 ; x2 ;


; xD Þ respectively; D is the variable
dimension.
(2) The head wolf produces rules
In the initialized population, the fitness value of each artificial wolf is calculated by
Eq. (2). The artificial wolf with minimal fitness value will be elected to be the head
wolf in the iterative process and the fitness value of head wolf is Qlead. During the
process, the artificial wolf with the optimal fitness value after each evolution need to
compare with the head wolf in the previous generation. If the current fitness value of
the artificial wolf is better, the artificial wolf will replaces the head wolf and to be a
new head wolf. For the minimum problem, Qi is considered better than Qlead when
Qi < Qlead.

Q ¼ f ðX Þ ð2Þ

where f ð X Þ is the objective function of the minimum optimization problem.

(3) Walk behaviour
In addition to the head wolf, the Snum wolves with the best fitness value are selected as
the search wolves in the current prey group to search for prey in space. Snum takes an
integer between ½n=ða þ 1Þ; n=a, where a is the scale factor of search wolf. The search
wolves heads in the h directions respectively and records the fitness value of
each position. And then they return to the original position, and proceeds to the pth
(p = 1, 2, …, h) direction. The location of wolf i in the dth (d = 1, 2, …, D) dimen-
sional space is shown in Eq. (3):

xPid ¼ xid þ c  stepda ð3Þ

68 Y. Wang et al.

where stepa is the walking distance of the search wolf, c is the search factor of search
wolf, and is a random number within [−1, 1].
At that time, the fitness value of the position where the search wolf is located is Qip.
The direction in which the fitness value is the smallest and smaller than the current
position fitness value Qio is selected to be further advanced, and the search wolf
information Xi is updated. Comparing the artificial wolf with the minimal fitness value
Qmin with the head wolf Qlead after the end of one-time walking behaviour. If
Qmin < Qlead, the wolf with the minimal fitness value becomes the new head wolf and
initiate the summoning; otherwise, repeat the walking behaviour until a certain wolf i
with the fitness value Qi < Qlead, or the number of walks T1 reaches the maximum
number of walks T1max. It should be pointed out that each search wolf has different
fitness values so the search directions h of each search wolf is different, it will be taken
within [hmin, hmax].
(4) Raid behaviour
The head wolf initiates the summoning action by howling, calling the raid wolves to
approach the head wolf quickly. The position of the raid wolf i in the dth (d=1, 2, …,
D) dimensional space at the (k + 1)th evolution is shown in Eq. (4):

  
xkidþ 1 ¼ xkid þ stepdb  gkd  xkid =gkd  xkid  ð4Þ

where stepb is the step size of the raid wolf, and gkd is the position of the head wolf of
the kth generation in the dth (d = 1, 2, …, D) dimensional space.
In the process of wolves rushing, the wolf with the minimal fitness value Qmin is
compared with the head wolf Qlead after each raid. If Qmin < Qlead, the wolf with the
minimal fitness value becomes a new head wolf, the raid behaviour ends; otherwise, the
raid continues until the fitness value of a certain wolf i is smaller than the head wolf
fitness value or the raid reaches the maximum raid T2max.
(5) Siege behaviour
The raid wolves will unite all the search wolves and move to the position where the
head wolf is after the raid behaviour. For all the wolves of kth generation, assuming
that the position of the head wolf in the dth (d = 1, 2, …, D) dimensional space is Gkd ,
the siege behaviour of wolves is shown in Eq. (5):
 
xid ¼ xkid þ k  stepdi  ðGkd  xkid Þ=Gkd  xkid  ð5Þ

where k is a random number distributed uniformly between [−1, 1]; stepi is the
adaptive siege step size of wolf i. Taking the difference between the fitness difference in
the search space upper and lower bound and the search space and the optimal value as
 The Research of Improved Wolf Pack Algorithm Based on Differential Evolution 69

the reference value, a positive correlation relationship between the adaptive siege step
length and the difference between the fitness value of wolf i and the fitness value of
head wolf is established. Adaptive siege step size stepi is shown in Eq. (6):

XU  XL
stepi ¼ ðQi  Qlead Þ  ð6Þ
maxQ0  minQ0

where maxQ0 and minQ0 are the worst fitness value and optimal fitness value of the
artificial wolf produced by the algorithm initialization.
If the fitness value of the position of the artificial wolf after the siege is smaller than
the fitness value of the original position, the position information of the artificial wolf
will be updated; otherwise, the position of the artificial wolf is not changed.
In particular: the stepa and stepb have the relationship is shown in Eq. (7):
 
stepa ¼ stepb =2 ¼ X U  X L =S ð7Þ

where S is the step size factor, it is the physical quantity that describes the artificial wolf
to search for fine degree in the space to be searched for.
(6) Variation behaviour
After the siege behaviour, two different artificial wolf (kth generation) X1k , X2k are
randomly selected in the group, and the vector difference is assigned to the weight and
then the third randomly selected artificial wolf X3k is added. Artificial wolf variation
vector vki þ 1 for any artificial wolf target vector Xik has expressed in Eq. (8):


vki þ 1 ¼ X3k þ F  X1k  X2k ð8Þ

where F is a scaling factor, it’s a value between [0, 2], indicating the degree of scaling
of the vector difference. The larger of F, the greater influence of the difference vector on
vki þ 1 , which is beneficial to keeping the diversity of wolves; on the contrary, F can
promote the local refined search.
(7) Crossover behaviour
After the variation behaviour, the artificial wolf variation vector vki þ 1 and the artificial
wolf target vector Xik are mixed according to the Eq. (9) to generate the artificial wolf
test vector uki þ 1 :
(
vkij þ 1 randðjÞ  CRorj ¼ randnðiÞ
ukij þ 1 ¼ ð9Þ
xkij randðjÞ [ CRorj 6¼ randnðiÞ

where j represents the jth variable of the artificial wolf, randnðiÞ 2 ½1; 2;


; D is the
identification of the randomly selected dimension variable; CR is the crossover prob-
ability factor, generally in the value between [0, 1]. The larger the CR is, the more
70 Y. Wang et al.

favourable of the algorithm performs local search and also to speed up the convergence
rate; the smaller the CR is, the more conducive to maintaining the diversity of wolves
and global search.
(8) Choice behaviour
After the crossover behaviour, the artificial wolf test vector uki þ 1 and the artificial wolf
target vector Xik are obtained. If the fitness value of the test vector is smaller than the
target vector, uki þ 1 is replaced by Xik to become the (k + 1)th generation and the
artificial wolf position is updated. Otherwise, use Xik directly as the (k + 1)th gener-
ation and update. It is shown in Eq. (10):

uki þ 1 Qðuki þ 1 Þ \ QðXik Þ
Xik þ 1 ¼ ð10Þ
Xik Qðuki þ 1 Þ  QðXik Þ

For the artificial wolf vector Xik þ 1 , the artificial wolf with the minimal fitness value
Qmin should be found and compared with the head wolf fitness Qlead. If Qmin < Qlead,
the artificial wolf with the minimal fitness value is selected as the head wolf.
(9) Wolves update mechanism
The prey hunted by the wolves is not evenly distributed to every wolf and will be
distributed according to the principle of “weak meat and strong food”. This may lead to
the weak wolves to be starved to death. This principle will be simulated by removing R
wolves with worst fitness and produced R artificial wolves according to formula (1).
R is a random integer between ½n=ð2  bÞ; n=b and b is the updates scale factor of
wolves.
Through the above actions, the wolves constantly update and evolve until the
algorithm reaches the maximum number of iterations Gmax or when the optimal value
obtained by the algorithm reaches the preset precision.

3 Simulation Test and Analysis of DIWPA

3.1 Introduction of Benchmark Functions
In order to test the performance and validity of DIWPA, this paper selects 12 benchmark
functions shown as Table 1 to test it. In Table 1, “U” represents unimodal function; “M”
represents multimodal function, it is usually used to detect the abilities of global opti-
mization and avoid precocity [15] of the algorithm; “S” represents a separable function,
“N” represents an inseparable function. The inseparable function will be more complex
than the separable function, and the performance requirement of the algorithm is higher
in the optimization process. In the definition field column of Table 1, the set part is the
 The Research of Improved Wolf Pack Algorithm Based on Differential Evolution 71

range of the variable value, and the index is the dimension D of the function. The higher
the test function dimension D, the greater the computational complexity of the algorithm
and the higher the performance requirement of the algorithm.

Table 1. 12 benchmark functions

Function Function expressions Feature Domain Global
name optimum
Sphere P
D US [−10, 10]2 0
f1 ¼ x2i
i¼1


Matyas f2 ¼ 0:26 x21 þ x22  0:48x1 x2 UN [−10, 10]2 0
Booth 2
f3 ¼ ðx1 þ 2x2  7Þ þ ð2x1 þ x2  5Þ 2 MS [−10, 10]2 0
Eggcrate f4 ¼ x21 þ x22 þ 25ðsin2 x1 þ sin2 x2 Þ MN 2
½p; p 0
Trid6 P
D P
D UN [−36, 36]6 -50
f5 ¼ ðxi  1Þ2  xi xi1
i¼1 i¼2
Sumsquares PD US [−10, 10]10 0
f6 ¼ ix2i
i¼1
Rastrigin PD   MS [−10, 10]60 0
f7 ¼ x2i  10cosð2pxi Þ þ 10
i¼1
Griewank P
D Q
D MN [−600, 0
f8 ¼ 4000
1
x2i  cosðpxiffiiÞ þ 1 600]100
i¼1 i¼1
Quadric P
D Pi MS [−30, 30]100 0
f9 ¼ ð x k Þ2
i¼1 k¼1
Rosenbrock Ph
D1
2 i MN [−30, 30]100 0
f10 ¼ 100 xi þ 1  x2i þ ðxi  1Þ2
i¼1
0 vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1
Ackley-1 u D MN [−32, 32]100 0
u1 X
f11 ¼ 20exp@0:2t x2i A
D i¼1
!
1X D
exp cosð2pxi Þ þ 20 þ e
D i¼1
0 vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1
Ackley-2 u D MN [−32, 32]200 0
u1 X
f11 ¼ 20exp 0:2t
@ x2i A
D i¼1
!
1X D
exp cosð2pxi Þ þ 20 þ e
D i¼1
72 Y. Wang et al.

3.2 Algorithm Verification and Analysis

In order to test the performance of DIWPA, this paper uses DIWPA, WPA, PSO, and
DE algorithms to perform 100 optimization tests on the 12 functions listed and records
the optimal value, the worst value, the average value, the standard deviation,the cal-
culation success rate and average consuming time to compared and analysed. Exper-
imental environment are as follows: Intel(R) Core(TM) i5-4590 CPU @3.30 GHz,
16 GB of memory, Win7 Professional 64-bit operating system, Matlab R2013b.
The calculation success rate is defined as the percentage of successful calculations
in the total number of calculations [9]. For each result Q obtained by the optimization
calculation, when Q satisfies formula (11), it means the search is successful:
(  
QQbest 
 Qbest  \ 1E  03 Qbest ¼
6 0
ð11Þ
jQ  Qbest j \ 1E  03 Qbest ¼ 0

where Qbest is the theoretical global optimal value of the function.

The maximum number of iterations for the optimization calculation are set to 2000,
and the population size of the DIWPA, WPA, PSO, and DE algorithms are 50, where:
the parameters of the WPA algorithm are selected according to the literature [9], the
parameters of the PSO algorithm are selected according to the literature [1], and the
parameters of the DE algorithm are selected according to the literature [7], as shown in
Table 2.

Table 2. Parameters of each algorithm

Algorithm Parameters
PSO Inertia weight x ¼ 0:7298, learning factor c1 ¼ c2 ¼ 2, individual velocity
range [−0.5,0.5]
DE Scale factor F = 0.6, crossover probability factor CR = 0.9
WPA Scale factor of search wolf a ¼ 4, maximum number of walks Tmax = 20,
distance determination factor x ¼ 500, step size factor S = 1000, update scale
factor b = 6
DIWPA Scale factor of search wolf a ¼ 4, maximum number of walks T1max = 10,
maximum number of raids T2max = 10, step size factor S = 1500, update scale
factor b = 3, scale factor F = 0.6, crossover probability factor CR = 0.9

The standard deviation, average value, optimal value, calculation success rate,
worst value, and average consuming time of the four algorithms on the 12 functions are
shown in Table 3. The average convergence curves of the functions are shown in
Fig. 1.
 The Research of Improved Wolf Pack Algorithm Based on Differential Evolution 73

Table 3. Optimization results of each algorithm

Function Algorithm Optimal Worst Average Standard Calculation Average
value value value deviation success rate consuming time(s)
Sphere PSO 3.92E−87 4.64E−81 1.69E−81 2.09E−81 100% 1.74
DE 0 0 0 0 100% 4.98
WPA 0 2.44E−187 1.22E−188 0 100% 16.8
DIWPA 0 0 0 0 100% 8.4
Matyas PSO 1.35E−71 2.85E−65 7.63E−66 1.08E−65 100% 4.89
DE 0 0 0 0 100% 9.09
WPA 0 0 0 0 100% 34.95
DIWPA 0 0 0 0 100% 16.64
Booth PSO 0 0 0 0 100% 4.03
DE 0 0 0 0 100% 6.93
WPA 3.77E−07 1.26E−04 4.22E−05 5.23E−05 100% 27.40
DIWPA 0 0 0 0 100% 13.04
Eggcrate PSO 8.03E−88 2.17E−75 1.09E−76 4.73E−76 100% 2.62
DE 0 0 0 0 100% 6.01
WPA 1.27E−171 1.25E−164 2.65E−165 0 100% 21.58
DIWPA 0 0 0 0 100% 11.35
Trid6 PSO −4.49E+01 −4.93E+01 −4.75E+01 1.31E+00 0 2.47
DE −5.00E+01 −5.00E+01 −5.00E+01 0 100% 5.88
WPA −5.00E+01 −5.00E+01 −5.00E+01 0 100% 20.88
DIWPA −5.00E+01 −5.00E+01 −5.00E+01 0 100% 9.08
Sumsquares PSO 3.22E−17 1.33E−15 5.48E−16 5.91E−16 100% 2.40
DE 1.08E−70 5.29E−67 4.64E−68 1.17E−67 100% 7.08
WPA 1.54E−56 8.99E−55 4.26E−55 3.81E−55 100% 23.7
DIWPA 0 0 0 0 100% 11.86
Rastrigin PSO 7.63E+01 1.56E+02 1.13E+02 2.16E+01 0 3.07
DE 2.28E+02 4.98E+02 3.83E+02 7.03E+01 0 16.9
WPA 4.41E−09 3.37E−06 3.75E−07 7.48E−07 100% 49.93
DIWPA 0 2.41E−12 1.58E−13 5.20E−13 100% 22.69
Griewank PSO 4.72E+00 2.03E+00 3.55E+00 8.68E−01 0 6.16
DE 1.99E+01 1.97E+01 1.98E+01 3.28E−04 0 18.90
WPA 3.11E−12 3.88E−01 1.49E−01 9.73E−02 94% 62.65
DIWPA 2.44E−15 6.68E−12 1.59E−12 1.95E−12 100% 32.98
Quadric PSO 3.16E−28 4.65E−23 4.45E−24 1.10E−23 100% 15.67
DE 2.08E−10 4.29E−02 4.01E−03 9.25E−03 50% 32.79
WPA 5.93E−10 8.64E−07 2.33E−07 3.06E−07 100% 121.15
DIWPA 4.15E−12 8.48E−08 1.11E−08 1.89E−08 100% 57.70
Rosenbrock PSO 1.46E+02 4.43E+02 2.78E+02 7.86E+01 0 3.45
DE 3.83E+02 2.81E+03 8.06E+02 5.62E+02 0 26.90
WPA 2.59E−11 5.26E−02 3.83E−02 8.48E−03 90% 75.88
DIWPA 3.23E−13 2.82E−10 2.39E−11 6.20E−11 100% 39.93
Ackley-1 PSO 5.92E−01 4.30E+00 2.43E+00 9.00E−01 0 4.28
DE 1.99E+01 2.01E+01 2.00E+01 5.95E−03 0 25.29
WPA 7.30E−13 6.36E−02 5.73E−02 1.76E−02 87% 73.93
DIWPA 8.88E−16 8.88E−16 8.88E−16 0 100% 36.97
Ackley-2 PSO 2.26E+00 4.91E+00 3.04E+00 6.29E−01 0 6.19
DE 1.96E+01 1.99E+01 1.98E+01 8.26E−04 0 46.14
WPA 1.52E−11 7.17E−01 2.39E−01 4.68E−01 75% 136.06
DIWPA 8.88E−16 8.88E−16 8.88E−16 0 100% 68.11
74 Y. Wang et al.

50 0 10
10 10 10
WPA DIWPA PSO DE WPA DIWPA PSO DE WPA DIWPA PSO DE

0 -50 0
10 10 10

The optimal value/log

The optimal value/log

-50
The optimal value/log

10 -100 -10
10 10
-100
10
-150 -20
10 10
-150
10
-200 -30
10 10
-200
10

-250 -40
10 -250 10 10
0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000
Number of iterations Number of iterations Number of iterations

(a) The average iteration (b) The average iteration (c) The average iteration
curve of Sphere curve of Matyas curve of Booth
20
50 500 10
10 WPA DIWPA PSO DE WPA DIWPA PSO DE
WPA DIWPA PSO DE

400 0
0 10
10
The optimal value/log

The optimal value/log

300
The optimal value/log

-20
-50 10
10
200
-100 -40
10 10
100

-150 -60
10 10
0

-200 -80
10 -100 10
0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000
Number of iterations Number of iterations Number of iterations

(d) The average iteration (e) The average iteration (f) The average iteration
curve of Eggcrate curve of Trid6 curve of Sumsquares

5 10
10 5 10
10
WPA DIWPA PSO DE WPA DIWPA PSO DE
WPA DIWPA PSO DE
The optimal value/log

0 0 5
10 10 10
The optimal value/log

The optimal value/log

-5 -5
10 10 0
10

-10 -10
10 10
-5
10

-15 -15
10 10
0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000
Number of iterations Number of iterations Number of iterations

(g) The average iteration (h) The average iteration (i) The average iteration
curve of Rastrigin curve of Griewank curve of Quadric

10 5
10 5 10
WPA DIWPA PSO DE 10
WPA DIWPA PSO DE
WPA DIWPA PSO DE

5 0
10 0 10
10
The optimal value/log
The optimal value/log

The optimal value/log

0 -5
10 -5 10
10

-5 -10
10 -10 10
10

-10 -15
-15
10 10 10

-15 -20 -20

10 10 10
0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000
Number of iterations Number of iterations Number of iterations

(j) The average iteration (k) The average iteration (l) The average iteration
curve of Rosenbrock curve of Ackley-1 curve of Ackley-2

Fig. 1. The average iteration curve of each function

 The Research of Improved Wolf Pack Algorithm Based on Differential Evolution 75

From Table 3 and Fig. 1, we can get the results as follows:

① For the single-peak, low-division and separable function of Sphere, the four
algorithms are all successfully searched. Among them, DIWPA and DE have the
best search capability, and the search capability of WPA is worse than DIWPA.
② For the single-peak, low-dimension and indefinable function of Matyas, the four
algorithms are all successfully searched. Among them, DIWPA, WPA and DE
have the same search capability and excellent performance.
③ For the multi-peak, low-dimension and separable function of Booth, the four
algorithms are all successfully searched, among which DIWPA, PSO, and DE have
the same search capability, and WPA has worse ability to search than the other
three algorithms.
④ For the multi-peak, low-dimension and indefinable function of Eggcrate, the four
algorithms are all successfully searched, among which DIWPA and DE have better
search capability, and WPA has worse ability than DIWPA.
⑤ For the single-peak and indefinable function of Trid6, DIWPA, WPA, and DE are
successfully searched and have good performance. The PSO is trapped in a local
optimal and lead to fail.
⑥ For the single-peak and separable function of Sumsquares, the four algorithms are
all successfully searched, among which the DIWPA has the best search capability
and the WPA is worse than it.
⑦ For the multi-peak and separable function of Rastrigin, The PSO and DE fall into a
local optimal lead to fail. DIWPA and WPA are successful and the DIWPA has
better performance.
⑧ For the high-dimensional, multi-modal and indefinable functions of Griewank and
Rosenbrock, the PSO and DE were failed. The DIWPA optimization was all
successful and has good performance. The WPA may fallen into a local optimal
and made the optimization ability unstable at some time.
⑨ For the high-dimensional, multi-peak and separable function of Quadric, the PSO,
WPA and DIWPA are all successfully searched, among which the PSO has the
best search capability. The DIWPA and WPA have the same optimization capa-
bility, but the DE suffers from local optimal values lead to the optimization ability
unstable.
⑩ For high-dimensional, multi-modal and indefinable function of Ackley, the PSO
and DE completely fail. The DIWPA has been successfully optimized. WPA has
fallen into a local optimal situation, which results in a decrease in the search
capability at some time. With the increase of the dimensionality of Ackley
function, The probability of falling into a local optimum is further increased, and
its ability to find optimization will be further reduced.
In summary, the DIWPA has a good ability for computing various types of
uni-/multi-peak, low-dimensional/high-dimensional, and separable/inseparable func-
tions. Compared with WPA, the DIWPA has better performance in multi-peak, high-
dimension, indivisible functions. The robustness and accuracy of the solution are
strengthened, and takes less time of the DIWPA.
76 Y. Wang et al.

4 Conclusion

In order to improve the optimization performance of traditional wolf pack algorithm, this
paper proposes an improved wolf pack algorithm based on differential evolution
(DIWPA). The main improvements are as follows: (1) The DIWPA introduces the search
factor of search wolf in the walking behaviour, which makes the algorithm optimization
can traverse the entire space; (2) The DIWPA introduces the maximum number of raid
wolves to improves computational efficiency and algorithm robustness; (3) the adaptive
siege step size and the differential evolution strategies is introduced to improve the global
search capabilities in DIWPA. The DIWPA is used to conduct optimization test on 12
benchmark functions and compare to 3 typical optimization algorithms. The test results
show that DIWPA has great robustness and global search ability, especially has excellent
performance in multi-peak, high-dimension, indivisible functions.

References
1. Kennedy, J., Eberhart, R.: Particle swarm optimization. In: 1995 Proceedings of IEEE
International Conference on Neural Networks, vol. 4, pp. 1942–1948. IEEE (2002)
2. Tehsin, S., Rehman, S., Saeed, M.O.B., et al.: Self-organizing hierarchical particle swarm
optimization of correlation filters for object recognition. IEEE Access PP(99), 1 (2017)
3. Wei, L.X., Li, X., Fan, R., et al.: A hybrid multi-objective particle swarm optimization
algorithm based on R2 indicator. IEEE Access PP(99), 1 (2018)
4. Singhal, P.K., Naresh, R., Sharma, V.: Binary fish swarm algorithm for profit-based unit
commitment problem in competitive electricity market with ramp rate constraints. Gener.
Transm. Distrib. IET 9(13), 1697–1707 (2015)
5. Li, X.L., Shao, Z.J., Qian, J.X.: An optimizing method based on autonomous animate: fish
swarm algorithm. Syst. Eng.-Theory Pract. 22, 32–38 (2002)
6. Li, Y., Zhang, C., Yang, Q., et al.: Improved ant colony algorithm for adaptive frequency-
tracking control in WPT system. IET Microwaves Antennas Propag. 12, 23–28 (2017)
7. Wu, L.H.: Differential Evolution Algorithm and Its Application. Hunan University (2007)
8. Yang, C., Tu, X., Chen, J.: Algorithm of marriage in honey bees optimization based on the
wolf pack search. In: International Conference on Intelligent Pervasive Computing, pp. 462–
467. IEEE Computer Society (2007)
9. Wu, H.S., Zhang, F.G., Wu, L.H.: A new swarm intelligence algorithm—wolf pack
algorithm. Syst. Eng. Electron. 35(11), 2430–2438 (2013)
10. Zhang, L., Zhang, L., Liu, S., et al.: Three-dimensional underwater path planning based on
modified wolf pack algorithm. IEEE Access 5(99), 22783–22795 (2017)
11. Wang, J.Q., Jia, Y.Y., Xiao, Q.Y.: Wolf pack algorithm in the optimal operation of
hydropower station reservoirs. Adv. Sci. Technol. Water Resour. 35(3), 1–4 (2015)
12. Hui, X.B., Guo, Q., Wu, P.P., et al.: An improved wolf pack algorithm. Control Decis. 32(7),
1163–1172 (2017)
13. Qian, R.X.: A wolf algorithm based on cultural mechanism. Inf. Technol. 2015(12), 98–102
(2015)
14. Cao, S., Jiancheng, A.: Otsu image segmentation method for improved wolf group
optimization algorithm. Microelectron. Comput. 34(10), 16–21 (2017)
15. Solano-Aragón, C., Castillo, O.: Optimization of benchmark mathematical functions using
the firefly algorithm with dynamic parameters. In: Castillo, O., Melin, P. (eds.) Fuzzy Logic
Augmentation of Nature-Inspired Optimization Metaheuristics. SCI, vol. 574, pp. 81–89.
Springer, Cham (2015). https://doi.org/10.1007/978-3-319-10960-2_5
 Multi-objective Optimization Genetic Algorithm
for Multimodal Transportation

Xiong Guiwu1,2 ✉ and Xiaomin Dong2

( )

1
School of International Business, Sichuan International Studies University,
Chongqing 400031, China
[email protected]
2
College of Mechanical Engineering, Chongqing University, Chongqing 400044, China

Abstract. The multimodal transportation is an effective manner in reducing the

transportation time and cost. However, the programming of multimodal trans‐
portation is very complex and belongs to NP difficulty problems. Therefore, an
optimal model based on the graph structure was firstly formulated for the multi‐
modal transportation with two optimal objectives including the transportation
time and the transportation cost. An optimized algorithm with two layers was then
proposed after characterizing the formulated model. The upper level was applied
to find the global optimal Pareto fronts and the transportation path, whereas the
lower level was to find the optimal path and the transportation manner. At last, a
numerical simulation was performed to validate the model and the proposed
algorithm. The results show that the proposed algorithm can find a series of Pareto
front solutions, which indicates that the formulated model and proposed algorithm
are effective and feasible.

Keywords: Fourth party logistics · Time window · Multi-agent

Hybrid Taguchi genetic algorithm

1 Introduction

As the economic globalization continues to intensify, resources, products, and markets

are scattered all over the world. Market demands are changing rapidly and competition
is becoming increasingly fierce. How to make all members in the logistics business
closer to each other in the supply chain? The information in the process of business
processing is highly integrated to achieve collaborative operation of the supply chain,
shorten the relative length of the supply chain, make the logistics business on the supply
chain smoother, yield higher, and respond faster. Meeting the needs of customers has
become an urgent issue in management science and engineering. The multimodal oper‐
ation is an important task of logistics and transportation organization and has become
an effective measure for the fourth party logistics company to improve its competitive‐
ness and thus reduce logistics costs. Compared to a single mode of transport, multimodal
transport can increase efficiency, enhance safety and flexibility, and at the same time
meet the requirements of the cargo carrier. However, compared to a single mode of
transport, due to multi-participation, multiple objectives need to be taken into account.
This makes multimodal transportation planning much more complicated and difficult.

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 77–86, 2018.
https://doi.org/10.1007/978-981-13-2384-3_8
78 X. Guiwu and X. Dong

Therefore, it is very important to develop a multi-objective optimization algorithm for

multimodal transportation planning.
At present, scholars have studied the modeling and algorithm problems of multimodal
transport planning from different perspectives. However, most of works belong to single-
objective optimization problems. There are few studies on multi-objective optimization
problems that consider cost and transit time. For example, Angelica Lozano et al. [1] studied
the shortest feasible path problem under multimodal transportation and solved it by using the
sequential algorithm. Boussedjra and Moudni [2] adopted a bi-directional research strategy
to establish an intermodal transport network model and presented the shortest path algo‐
rithm between source and destination. Some scholars consider multi-objective optimization
when passengers use multiple transport modes. Brands et al. [3] studied the multi-objective
optimization problem of multimodal passenger transportation in urban passengers consid‐
ering uncertain demand. The results show that the demand is uncertain. The Pareto frontier
solution of multi-objective optimization has a great influence on them. They [4] also applied
this algorithm to solve multimodal network design problems. Zhang et al. [5] applied the
uncertain optimization method to study the shortest path of passenger multimodal trans‐
port. They considered the difference in transit time and transportation cost caused by the
uncertainties of different transport modes. The goal of the double-objective optimization
problem was verified by numerical methods. Osuji et al. [6] studied the multi-objective
transportation problem based on fuzzy optimization. Zhang et al. [7] proposed a general‐
ized shortest path algorithm for solving optimal transportation routes based on the multi‐
modal transport model established by Reddy. Wei et al. [8] considered the time-varying
conditions of cost, transit time, and risk in the transportation process, and established the
short-circuit model of time-varying multimodal transportation through the deformation of
the transportation network and solved it. Zhang et al. [9] established an optimal distribution
model for multimodal transport networks, and quantitatively analyzed the rational organi‐
zation mode of multimodal transport to minimize total costs. Wang et al. [10] proposed an
optimization model for transportation modes based on the analysis of the transport charac‐
teristics of various transportation modes and presented a solution algorithm. Wei et al. [11]
established the shortest time path-transportation cost model for multimodal transportation,
and proposed the iterative algorithm to solve the shortest path. Yang et al. [12] studied the
multimodal transport model with time windows, proposed a bi-level optimization model,
and used ant colony algorithm to achieve path optimization. These studies have established
and solved the multimodal transport model for a single logistics transport task under certain
conditions, which belongs to the single-objective optimization problem.
From the above analysis, it can be seen that there are still few researches on the multi-
objective optimization of logistics multimodal transport planning problems. The
research on modeling methods and algorithms still needs to be in-depth. Therefore, based
on the previous research [13], this paper intends to investigate multi-objective modeling
and solving algorithms for multimodal transport planning. In the previous research,
considering the integration of multimodal transport and multi-agent logistics operations
with time windows, a multimodal transportation optimization model based on graph
structure was established and a two-layer search algorithm was designed to solve it. For
further in-depth research, this paper proposes to build a multi-objective optimization
model based on a multimodal transport network that considers transportation time and
 Multi-objective Optimization Genetic Algorithm 79

transportation cost in Sect. 2. In Sect. 3, a two-layer multi-objective hybrid Taguchi

algorithm is proposed. The numerical examples verify the established models and algo‐
rithms in Sect. 4. Section 5 will conclude it.

2 Modeling of Multimodal Transportation

In order to facilitate modeling and analysis, the virtual network is defined as follows.
Let G = (V, E, M) be the network graph under consideration. It is the abstraction of
logistics nodes and traffic connections in the geographical map under consideration. V
is a set of the nodes in the actual logistics transportation network correspond. E is a set
of edges, corresponding to all possible transportation paths. M is a collection of trans‐
portation modes. To facilitate problem solving, the following assumptions are made.
Hypothesis 1 the same shipment cannot be transported separately;
Hypothesis 2 there are multiple modes of transport between two adjacent transfer
stations to choose from, but only one mode of transport can be selected;
Symbol description:

xijm commodity is shipped at (i, j) with mode m;

cijm unit shipment price of commodity at (i, j) with mode m;
dijk shipment length at (i, j) with mode m;
vm unit shipment velocity for mode m;
ttikl the unit transfer time from mode k to l at node i;
tcikl the unit transfer cost from mode k to l at node i;
𝜔kli changing sign from mode k to l at node i;
qijm Commodity volume at (i, j);
Di Schedule time at node i, Di = 0 for no schedule time;

G = (V, E, M) is a multimodal transport network with time windows, which includes

reducing total transportation time and transportation costs. The decision variables are
xijm and 𝜔kli. The multi-objective optimization model with the goal of transportation cost
and transportation time is as follows:
{ }
∑∑ ∑∑∑
Z1 = min cijm qijm dijm xijm + qijm 𝜔kli tcikl
ij m ij i kl

{ }
∑∑ ∑∑∑
Z2 = min dijm ∕vm + max(Di , qijm ttikl ) (1)
ij m ij i kl

S.T.

⎧Q if i = O
∑ ∑ ⎪
qijm − qjim = ⎨ −Q if i = D ∀ i ∈ V, m ∈ M (2)
j i ⎪0 otherwise
⎩
80 X. Guiwu and X. Dong

max(Di , qijm ttikl ) + dijm ∕vm − Aj = 0 ∀ i, j ∈ V, (i, j) ∈ E, m ∈ M (3)

li ≤ Ai ≤ ui , ∀ i ∈ V (4)
∑
xijm ≤ 1 ∀ (i, j) ∈ E, m ∈ M, (5)
m

{
1 k≠l
𝜔kli = , (6)
0 k=l

xijm ∈ {0, 1} ∀ (i, j) ∈ E, m ∈ M. (7)

The first optimization objective function in Eq. (1) is to reduce the total transportation
cost, and the second optimization goal is to reduce the total transportation time. Equa‐
tion (2) is the traffic balance condition in the network flow. Formula (3) ensures the
tolerance of time requirements. Equation (4) represents that the node should meet the
predetermined time window constraints. Formula (5) indicates that there is at most one
transport mode for each link. The values of variables (6) and (7) are limited.

3 Multi-objective Genetic Algorithm Based on Time Window

In order to make the multimodal transportation model closer to the real scene, we studied
the customized time window, the transportation distances of different transportation
modes, different transportation speeds, the conversion costs and conversion time of
different nodes, which is different from the previous issue of multimodal transportation
routes. There are multiple edges between some nodes. Moreover, once the mode of
transportation has changed at certain nodes, additional conversion costs and time are
still included. The problem has become very complicated, and the key is to find it difficult
to find a best multimodal transport network. Therefore, a two-layer multi-objective
hybrid Taguchi algorithm is proposed to find the optimal transport path.
The dual-layer multi-target hybrid Taguchi algorithm consists of two layers. The
upper layer aims to find the overall global optimal Pareto frontier and transportation
mode, while the lower layer searches for the local optimal Pareto frontier of the fixed
path and transport mode. The lower level optimization method determines the upper
feasible solution space. The multimodal transport path problem model can be reduced
to the following two-level multi-objective optimization problem.

min Z = {Z1 , Z2 )
(xijm ,𝜔kli )

S.T.

𝜔kli ∈ arg min {Z1 , Z2 }, (8)

kl
𝜔i
 Multi-objective Optimization Genetic Algorithm 81

⎧Q if i = O
∑ ∑ ⎪
qijm − qjim = ⎨ −Q if i = D ∀ i ∈ V, m ∈ M, (9)
j i ⎪0 otherwise
⎩

max(Di , qijm ttikl ) + dijm ∕vijm − Aj = 0, ∀ i, j ∈ V, (i, j) ∈ E, m ∈ M, (10)

li ≤ Ai ≤ ui , ∀ i ∈ V, (11)
{
1 k≠l
𝜔kli = , (12)
0 k=l
∑
xijm ≤ 1 ∀ (i, j) ∈ E, m ∈ M, (13)
m

xijm ∈ {0, 1} ∀ (i, j) ∈ E, m ∈ M. (14)

3.1 Solution Algorithm for the Upper Level

The task of this layer is to find the global Pareto frontier and provide effective and
feasible transport paths for the lower levels. As far as multimodal transport path opti‐
mization is concerned, if network agents have only one mode of transport to choose
from, the shortest path algorithm should be the optimal solution. Taking into account
the existence of various modes of transport, it is very difficult to directly apply the k
shortest path algorithm. To this end, multimodal transport networks are transformed,
standardized, and weighted.
Algorithm 1. Initial feasible path population generation
Step 1: Read the data of the multimodal network.
Step 2: Calculate the normalized weighting matrix according to the weights
Step 3: Calculate k shortest paths using the K-shortest path algorithm
Algorithm 2. Check and replace operation
Step 1: If the path is feasible, stop it;
Step 2: Generate a random weight factor. Calculate normalization matrix and apply
K shortest path algorithm to calculate k shortest paths;
Step 3: Select a feasible path from the K feasible paths to replace the infeasible path.

3.2 Optimization Algorithm for the Lower Transportation Mode

Since the position of the Pareto front before the optimization is not known, it is hoped
that the chromosomes of the initial population can be evenly dispersed throughout the
feasible solution space, and there is a potential advantage of using the Taguchi experi‐
mental method to generate these chromosomes.
Algorithm 3. Generation of the initial population of transport modes
Step 1: Read the path provided by the upper layer
82 X. Guiwu and X. Dong

Step 2: All the edges in the path have the same transport method available? If appli‐
cable, apply to select chromosomes and go to step 4. If not, go to step 3
Step 3: Divide the path into several sections based on the same available transport
method. Sides with the same available transport methods are connected together.
Each section chooses its own orthogonal table.
Step 4: Calculate the shipping cost and time for each chromosome.
Step 5: Sort the chromosomes according to the fast non-dominated sorting method.
Step 6: Select the chromosome based on chromosome rankings.
The Taguchi experiment method also incorporates crossovers in order to inherit good
genes from the father. This article uses a two-level orthogonal array with l factors, where
l is the length of the feasible path. Two levels correspond to the parents of two chro‐
mosomes.
Algorithm 4. Crossing based on taguchi experimental method
Step 1: Randomly select the parents of the two paired pools;
Step 2: Select the orthogonal table according to the length of the chromosome;
Step 3: Use orthogonal experiments to generate progeny;
Step 4: Sort children according to the fast non-dominated sorting method;
Step 5: Select two chromosomes based on the chromosomes.

3.3 Two-Level Multi-target Taguchi Genetic Algorithm

Based on the above analysis, the two-level multi-object Taguchi genetic algorithm can
be summarized as follows.
Algorithm 5. Bi-level multi-target Taguchi genetic algorithm
Step 1: Read the array and graphics structure. Read the initial value of all parameters.
Step 2: parameter settings: For the upper level, including population size, crossover
rate and mutation rate. For the lower layer, population size, crossover rate, mutation
rate
Step 3: Apply algorithm 1 to generate the initial population of feasible paths
Step 4: Upper Population Evolution
If (stopping conditions are not met) proceed
Start
Step 4.1: Apply Algorithm 3 to generate the initial population of transport methods
for each feasible path
Step 4.2: Lower population evolution
If (stopping conditions are not met) proceed
Start
Step 4.3: Select: Select chromosomes to cross, crossover probability. Continue to
choose until the number of selected chromosomes is half of it.
Step 4.4: Crossover: Crossing selected chromosomes using algorithm 4
Step 4.5: Variation: The probability of variation in each chromosome
Step 4.6: Sort all chromosomes using fast non-dominated sorting
Step 4.7: Keep the next generation of the previous chromosome
 Multi-objective Optimization Genetic Algorithm 83

End
Step 5: Select: Intersect each chromosome with crossover probability. Continue to
choose until the selected number of chromosomes reach to half of it
Step 6: Cross: Crosses the selected chromosome. If the crossed path is feasible, then
go to Step 8
Step 7: Execute Algorithm 2
Step 8: Variation: The probability that each chromosome will mutate. If the crossover
path is feasible, then go to Step 10.
Step 9: Execute Algorithm 2
Step 10: Use the fast non-dominated sorting method to sort all chromosomes
Step 11: Preserve the next chromosome of the next generation
End

4 Numerical Simulation

In order to verify the models and algorithms constructed, based on the logistics network
in Eastern China, it is assumed that there are a batch of containers that need to be trans‐
ported. The transport can be carried out in 35 city nodes. The transportation methods
include roads, railways and waterways, but not all cities have three links. An example
was constructed with 35 nodes and 136 edges. Figure 1 shows the map of logistics and
transportation in eastern China. Many major cities along the Yangtze River rely on
waterways, railroads and trucks for transportation. There is a simple and effective model
in the multimodal network environment. The model can provide some good trade-off
solutions. These nodes next to the Yangtze River can be connected to each other by
barges, such as to nodes 1, 3, 9, …, 93. The following parameters
G1 = 220, pc1 = 0.1, pm1 = 0.1, G2 = 50, pc2 = 0.2 and pm2 = 0.2 are used in the
simulation. The stop condition is simply to reach the highest iteration of the predeter‐
mined upper and lower levels. The two predetermined maximum iterations are 20
generations.

Fig. 1. Logistics transportation network in China eastern region

There are three common modes of transportation, namely roads, railways and water‐
ways. The total transportation costs include transportation costs and transportation costs.
84 X. Guiwu and X. Dong

The mode of transportation determines the cost of transportation. Transit costs include
labor costs and equipment costs during transportation. The transfer container is a
standard container, 8 feet high, 8 feet wide and 20 to 55 feet long. Containers can be
transported by trucks, railroads and barges. In general, containers can hold 15 tons.
Because there is no universal standard to describe the shipping rate, transshipment rate,
transportation speed and transit time. This article uses the relevant data provided by [14]
as the basis for calculation, as shown in Tables 1, 2 and 3. The transport rates and speeds
given in Table 1 are shown in Tables 2 and 3, respectively.

Table 1. Transportation rates and speed

Transportation rates (USD dollars/container) High way Rail way Water
Cost (USD dollars) 0.90 0.3510 0.1234
Speed (miles/hour) 42 37 6.7

Table 2. Transfer rate (cost for each container)

High way Rail way Water
High way 0 58 100
Rail way 58 0 100
Water 100 100 0

Table 3. Transfer time (time for each container)

High way Rail way Water
High way 0 0.1 0.17
Rail way 0.1 0 0.17
Water 0.17 0.17 0

Consider one task is 1000 containers from node 1 (Shenzhen) to node 93 (Dandong).
The goods are more focused on the quality of transportation. Assuming that the trans‐
portation continues until it reaches the destination, simulations are performed. The
results are shown in Fig. 2. The Pareto solution and other non-dominated optimal solu‐
tions are given in the figure. These solutions become alternatives to the carrier’s decision.
The carrier needs to select according to the shipper’s requirements (Table 4).
 Multi-objective Optimization Genetic Algorithm 85

Pareto front solution

Second order
Third order
Others

Fig. 2. No-dominated solution plan

Table 4. Transportation path and manner from Shenzhen to Dandong

Plan Transportation path and manner Cost (USD Time (hours)
Dollars)
1 1-9-10-15-31-32-33-47-45-44-43-51-56-62-69 7.9405 e + 005 388.35
-74-75-80-89-85-84-93
Water-Water-Water-Water-Water-Rail-Rail-
Rail-Rail-Rail-Rail-Rail-Rail-Rail-Rail-Rail-
Rail-Rail-Rail-Rail-Rail
2 1-10-15-31-46-47-48-50-57-56-62-69-74-75-8 7.156 e + 005 412.97
0-89-85-84-93
Water-Water-Water-Water-Water-Rail-Rail-
Rail-Rail-Rail-Rail-Rail-Rail-Rail-Rail-Rail-
Rail-Rail
3 1-10-15-31-32-46-45-44-50-57-56-62-69-74-7 7.4644 e + 005 403.81
5-80-89-85-84-93
Water-Water-Water-Water-Water-Rail-Rail-
Rail-Rail-Rail-Rail-Rail-Rail-Rail-Rail-Rail-
Rail-Rail-Rail
4 1-9-10-15-31-46-58-93 3.4699 e + 005 419.68
Water-Water-Water-Water-Water-Water-
Water
5 1-9-10-15-31-46-45-44-50-51-56-62-69-74-75 7.5566 e + 005 404.52
-80-89-85-84-93
Water-Water-Water-Water-Water-Rail-Rail-
Rail-Rail-Rail-Rail-Rail-Rail-Rail-Rail-Rail-
Rail-Rail-Rail
86 X. Guiwu and X. Dong

5 Conclusion

A multi-objective modelling and optimization algorithm for multi-modal transport opti‐

mization of fourth-party logistics was proposed. Based on the description of model
problems, a multi-objective optimization model for multimodal transport planning with
time window constraints was established using graph-like structures. The optimization
of different task requirements was solved. The calculation results show that the algorithm
can solve a series of Pareto frontier solutions that meet the requirements based on the
user’s demand for transportation time and transportation costs. This shows that the
proposed algorithm is feasible and effective. This article does not consider the issue of
the transfer of jobs between agents. This will be investigated further in the future.

References

1. Lozano, A., Storchi, G.: Shortest viable path algorithm in multimodal networks. Transp. Res.
Part A 35(3), 225–241 (2001)
2. Boussedjra, M., Bloch, C., EI Moudni, A.: An exact method to find the intermodal shortest
path. In: Proceedings of the 2004 IEEE International Conference on Networking, Sensing &
Control, Taiwan, China (2004)
3. Brands, T., Wismans, L.J.J., Berkum, E.C.V.: Multi-objective optimization of multimodal
passenger transportation networks: Coping with demand uncertainty. In: Papadrakakis, M.,
Karlaftis, M.G., Lagaros, N.D. (eds.) An International Conference on Engineering and
Applied Sciences Optimization, Kos Island, Greece, 4–6 June 2014, pp. 547–561 (2014)
4. Brands, T., Berkum, E.C.V.: Performance of a genetic algorithm for solving the multi-objective,
multimodal transportation network design problem. Int. J. Transp. 2(1), 1–20 (2014)
5. Zhang, Y., et al.: A bi-objective model for uncertain multi-modal shortest path problems. J.
Uncertainty Anal. Appl. 3(8), 1–17 (2015)
6. Osuji, G.A., Okoli Cecilia, N., Opara, J.: Solution of multi-objective transportation problem
via fuzzy programming algorithm. Sci. J. Appl. Math. Stat. 2(4), 71–77 (2014)
7. Yunhe, Z., Dong, L.B.L., Hongyan, G.: Research on a generalized shortest path method of
optimizing intermodal transportation problems. J. China Railway Soc. 28(4), 22–26 (2006)
8. Wei, H., Li, J., Liu, N.Z.: An algorithm for shortest path with multimodal in time - varying
network. Chin. J. Manag. Sci. 14(4), 56 (2006)
9. Jianyong, Z., Yaohuang, G.: A multimodal transportation network assignment model. J. China
Railway Soc. 24(4), 114–116 (2002)
10. Tao, W., Gang, W.: A combined optimization model for transportation modes of multimodal
transport. Eng. Sci. 7(10), 46–50 (2005)
11. Zhong, W., Jinsheng, S., Ailing, H., et al.: Research on model of the shortest time path and
transport cost in multimodal transportation. Eng. Sci. 8(8), 61–64 (2006)
12. Wendong, Y., Wenfang, W.: Analyzing and modeling of multimodal transportation with time
window. J. Nanjing Unv. Aeronaut. Astronaut. 41(1), 111–115 (2009)
13. Guiwu, X., Yong, W.: Optimization algorithm of multimodal transportation with time
window and job integration of multiagent. J. Syst. Eng. 26(3), 379–387 (2011)
14. Manuel, D., Rossetti, H.N.: WebShipCost - Intermodal Transportation Linkage Cost
Assessment Via the WWW. The Mack-Blackwell Transportation Center (2005)
Big Data Analytics
 The Research of 3D Power Grid Based on Big
Data

Bing He(&), Jin-xing Hu, and Ge Yang

Shenzhen Institutes of Advanced Technology, CAS, Shenzhen, China

[email protected]

Abstract. The visual graphics can reflect the hiding relationships between
different data by color, location, and the topological structure. This paper pro-
poses a 3D power grid framework by integrating big data and 3D visualization
technologies to achieve flexible, multidimensional, dynamic, and panoramic
display of power data. The proposed framework consists of data acquisition,
data analysis and mining and 3D visualization module. We detailed the 3D
method from modeling, management and visualization techniques, and analyzed
the modeling elements from geospatial data, electric graphic data, electric
attribute data and data mining result dataset. Finally, we give two scenarios of
3D power system. Most of tools used in this paper are open source software.
This can ensure the system stability, flexible and easy to use for the management
of large distributed data resources and avoid duplication system development.

Keywords: Geospatial database
Time series data
CityGML

Power GIS

1 Introduction

As the power system becomes more and more complex and complicated, the control of
the safe and stable operation of power grid becomes very complicated. In order to
improve the operation level of power grid, the power system imposes higher require-
ments on high-performance computing and visualization technologies. But, due to the
heterogeneous nature of power systems, it’s difficult to achieve interoperation between
various power databases. This results in the phenomenon of information islands [1].
However, a new trend in IT service is managing, analyzing and visualizing the multi-
source power big data by high-performance computer clusters to construct smart power
grid [2–5]. Big data has 4 V characteristics, focusing on the fusion, analysis and mining
of multi-source and multi-type data. It has high data throughput speed, distributed
computing, and multi-source data association mining analysis capabilities. Therefore,
the characteristics of big data technology fit the needs of massive power data analysis
and calculation.
Currently, 2D visualization technology has been widely used in the power industry.
The power data is displayed by two-dimensional lines, charts, scatter plots and images
[6, 7], or by geographic information system (GIS) [8–10]. However, big data tech-
nology and GIS technology have been decentralized in the power sector. Also, 2D GIS
is based on two-dimensional abstract graphical symbols, which have limited ability to

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 89–98, 2018.
https://doi.org/10.1007/978-981-13-2384-3_9
90 B. He et al.

present information to users. And it’s difficult to accurately deal with many 3D
problems in 2D GIS. Therefore, if the high-performance computing capability of big
data technology can be effectively combined with geographical 3D visualization
technology, the real-time power grid operation and control will become clear, intuitive
and effective. Therefore, this paper represents a power grid framework based on big
data and 3D technologies to achieve flexible, dynamic visualization of power data.

2 3D Power Data Visualization Framework

This section introduces a framework based on big data technologies and 3D techniques
to manage and visualize large amounts of power data. This framework consists of data
acquisition and conversion modules, big data platform, power data analysis and mining
modules, and 3D visualization modules.
(1) Data acquisition and conversion modules: in the proposed framework, power data
is mainly extracted from the internal system [11], for example, wide area mea-
surement system, energy management system, device detection and monitoring
system, etc. External data is collected from socio-economic, urban population,
geographic information system. The two kinds of data are imported into big data
platform by three ways. Traditional relational data is converted by Apache Sqoop
[12]. Streaming data or timing data is processed by Apache Storm [13].
Unstructured data could be imported by Apache Flume [14]. Specifically, the data
which will be managed and processed in big data platform includes power
topology elements, 3D models, related attribute data (such as electric equipment
information) and geospatial data (e.g. streets, buildings, roads, water, etc.)
(2) Big data platform: We use Apache Hadoop [15] to build big data platform. The
underlying storage form is HDFS which will be analyzed directly. In order to
facilitate the scheduling and management of cluster resources, we adopt YARN
[16] as the manager. Above data management module is data processing services,
including off-line processing, machine learning, memory computing. Offline pro-
cessed data can be stored by HBASE [17]. HIVE [18] generates SQL query
statement and submits to other analysis and processing modules. Offline processing
is mainly used to meet non-real-time requirements. But for time-critical computing,
we use memory computing, such as Impala [19] or Spark [20].
(3) Data analysis and mining modules: Multi-source heterogeneous data often implies
valuable relationship. For this purpose, machine learning library (such as Spark R
and Spark Mllib) is employed. Except for providing simple data query and general
analysis capabilities, this framework includes the 3D modeling method for domain
application, For example, the data mining results of equipment state evaluation,
risk assessment, load capacity assessment and fault prediction. The modeling
results will be submit to visualization client (such as web browser, desktop 3D
viewer) (Fig. 1).
(4) 3D visualization of analysis modules: 3D visualization module is the core of this
framework. It consists of two parts. One part is power data spatialization module,
the other is the power data visualization module. Spatialization module access to
 The Research of 3D Power Grid Based on Big Data 91

Power Web3D GIS 3D Symbol Annotation

Visualization

Spatialization of Analysis Results

State Evaluation & Fault

Data Analysis Load Capacity Assessment
Prediction
and Mining
Equipment Risk Assessment

JDBC ODBC

Memory Computing Machine Learning Offline Processing

Impala Spark Spark R MLlib HIVE HBASE

Big Data
Platform Cluster Resource Scheduling Management (YARN)

HDFS HDFS HDFS HDFS

Distributed File System

Migration Migration Migration

Storm Sqoop Flume

Data Collection Time Traditional

Unstructured
Series Relational
Data
Data Database

Fig. 1. 3D power data visualization framework

power big data platform through JDBC, ODBC or Hive. The spatialized objects
include data mining and analysis results and static attribute data. Then, the spa-
tialized results are transferred into three-dimensional symbols (such as points, lines,
and surfaces) and annotations. Finally, three-dimensional client utilizes digital
terrain models, remote sensing images, power equipment, 3D symbols and anno-
tations to construct a panoramic virtual scene.

3 3D Techniques

Comparing with the 2D power GIS, 3D modelling process is complex and requires a
combination of different software to construct power and geospatial objects. This
section will detail the visualization techniques for 3D power system (Fig. 2).

3.1 Modeling Techniques

• SketchUp [21] is an open source software for 3D modelling. It has been used in a
wide range of applications such as architectural and civil design. Power tower and
electric equipment can be easily modelled in SketchUp. However, due to lack of
geographical projection information, 3D models created by SketchUp can’t be
directly import into 3D power system. In order to ensure that the model can be
92 B. He et al.

Internet
HTML JavaScript Java Applet
Explorer
Mobile Client Desktop Client

Visualization Techniques

WebGL Earth Cesium Ossimplanet WorldWind

WebGL OpenGL

OpenLayers

Map,GML CityGML

Google SketchUp
3D Geodatabase Skyline TerraExplorer
GeoServer 3D Models
(3DCityDB)
ESRI CityEngine
Autodesk Revit
PostGIS
Modeling Techniques

Fig. 2. 3D techniques for power GIS

displayed properly, the usual practice is firstly defining the local coordinates of the
area to be modelled, and then convert it to system’s geographic coordinate.
• Skyline TerraExplorer Pro [22], ESRI CityEngine [23] are two commercial 3D GIS
desktop viewer and creator. They provide a wide range of powerful terrain analysis
tools to improve decision making. They can manage and visualize geospatial fea-
tures, imageries, elevation raster, urban models and dynamic objects. These data
could be published and utilized by web or desktop client. Secondary development
power 3D system could have rich functionality based on this two software, and the
development process is easier.
• 3D GIS has been used to the acquisition, management, analysis in macroscopic
geographical environment, and show buildings generalization appearance and
location. It’s unable to obtain further constructional details (such as indoor pattern,
various pipelines and facilities configuration and materials, etc.). The inner com-
ponents of 3D objects haven’t explicit spatial relationship. This results in today’s
power 3D GIS staying in viewing three-dimensional appearance. For the city power
3D map based on big data technologies, the detailed power information is very
helpful for power management staff to know the panorama operation status of
power network. Autodesk Revit [24] software is a such tool to build Building
Information Modelling (BIM). In order to achieve three-dimensional multi-scale
power map visualization, it needs BIM models integrated into 3D GIS.

3.2 3D Geospatial Database

With the growth of 3D power models, it’s need effective storage and management
tools. Meanwhile, in order to support data access into big data platform, the 3D model
management method should be migrated to web environment, and ultimately realizing
 The Research of 3D Power Grid Based on Big Data 93

3D models sharing. Currently model data management often adopts relational database
such as SQL Server, Oracle, and associates model files with database to achieve read
and display. For the model retrieval, each model will be classified, labelled and added
property, and then establish text index database to achieve quick search. This method is
inefficient. It requires an efficient way for large data management in big data platform.
The 3D City Database (3DCityDB) is a free Open Source package consisting of a
database schema and a set of software tools to import, manage, analyse, and visualize
virtual 3D city models according to the CityGML standard. 3D power models can be
transformed to CityGML. The geometry, topology, semantics between the different
classes of the power objects are kept in the CityGML file. Then, 3DCityDB import tool
could be used to save the CityGML file into relational database, and tile the large,
complex 3D power objects to easily visualize in virtual globes like Google Earth or
Cesium WebGlobe.

3.3 Web3D Visualization

Common 3D power system is tightly coupled to the 3D commercial software. It’s
difficult for sharing modules between different power systems and departments. For the
cross-platform demand, there are two techniques to achieve 3D visualization in web
browser. One is based on WebGL, the other is based on OpenGL. Each produces
corresponding visualization software or library.
Cesium [26] and WebGL Earth [27] are two types of JavaScript library for creating
3D globes and 2D maps in a web browser without a plugin. They provide a mechanism
to inserting interactive maps into webpages. They use WebGL for hardware-
accelerated graphics. The power 3D objects in the format of CityGML are viewable.
The geospatial data sources of 3D power GIS can be other web map services as well as
image tiles services. Power management staff can rotate and zoom in or out on the
globe, or even tilt the viewing angle.
World Wind [28] and ossimPlanet [29] are based on object-oriented programming
language. They are open source. World Wind is a Java library and uses satellite
imagery and elevation data to allow users to experience earth terrain in visually rich
3D. ossimPlanet is a C++ library. It supports native file access to a wide variety of
geospatial data sets - including raster, vector, and view elevation data sets and OGC
Web Mapping Services over the web (Fig. 3).

We3D Visualization

WebGL OpenGL

Ceisum WebGL Earth WorldWind OSSIMPlanet

Fig. 3. 3D visualization in web browser

94 B. He et al.

4 Power 3D GIS Modeling

4.1 Modeling Elements
In the general power system, geographic and power data are visualized by 2D
geometries, colors and symbols, and the power assessment results are pie charts, bar
charts, graphs. Power management staff can’t grasp the spatial distribution of analysis
results. In this paper, power 3D data will be divided into four categories: geospatial
data, electric graphic data, electric attribute data and power analysis and data mining
result dataset.
(1) Geospatial data: geospatial data is mainly used to build a three-dimensional virtual
scene, including DEM, remote sensing images and other geographic features, such
as buildings, roads, bridges, and related information. There data will be modelled
by 3D software and saved into 3D database. Based on LOD technology and spatial
indexing technology, the display efficiency can be improved (Fig. 4).

Fig. 4. 3D visualization of geographic data

(2) Electric graphic data: this type of data includes geographic wiring diagram for
power facilities such as (overhead lines, cable lines), equipment (substations,
switching stations, towers, substations, transformers, switches, etc.). The other
three types of data could be associated with electric graphic data to build a 3D
panoramic visualization environment (Fig. 5).
(3) Electric attribute data: include monitoring data, power sales data, social statistics,
and etc.
(4) Data Analysis and Mining Results: Power system management process involves a
lot of data analysis. This type of data sets comes from load capacity assessment,
equipment state evaluation and fault prediction, equipment risk assessment, etc.

4.2 Modeling Method

According to the 3DCityDB characteristics, the modelling method includes static
symbols and dynamical spatialization. When static symbolization, different types of
visualization objects will be separately established the appropriate static symbol library,
and stored into 3D model database. The visualization process of these types of objects
 The Research of 3D Power Grid Based on Big Data 95

Fig. 5. 3D modeling of power devices

3D Modeling Method

Static Symbolization Dynamic Generation

3D Symbols 3D Spatialization

Spatial Association Unique Value

Categories Unique Value
with Many Fields
Graduated Graduated Proportional
Quantities
Colors Symbols Symbols
Composite Object
Charts Pie Bar Stacked

Point Line Polygon Multiple Attributes

3D Entity
Object Object Object

Fig. 6. 3D modeling method for power system

will be displayed by setting colors, geometries and symbol size by analysed data value.
Dynamic spatialization is used to visualize the data analysis and mining results. The
obtained graphics data sets will be directly graded and classified for visualization. The
spatialization method includes inverse distance weighted method, ordinary kriging,
cubic splines, and etc. The visualization method includes categories (unique value,
unique value with many Fields), quantities (graduated colors, graduated symbols,
proportional symbols), charts (pie, bar, stacked), and multiple attributes (Fig. 6).

5 Scenarios

This section will describe the using of 3D power visualization framework in the fol-
lowing two scenarios:

5.1 Operation State Monitoring of Power Grid

The sensors deployed in the power network can real-time access the status of related
equipment. The acquired data is based on the form of flow and time series. This paper
96 B. He et al.

uses Storm to process it. According to the purpose of the application, the visualization
method is divided into three types: abnormal state detecting, real-time operation state
visualization, post hoc analysis. During the abnormal status detecting process, the
abnormal value combine with the devices coordinate will be transmit to visualization
client and symbolized by abnormal types to achieve equipment malfunction status
alerts. Real-time status could be visualized by time interval and associated with the
observed value of the related equipment and power lines. The status value need to the
space-time transformation to convert the indicator data or time data into three-
dimensional space, and the format index axis is the Z-axis three-dimensional graphics,
and can also be used hierarchical, categorical displayed. The post hoc analysis is
focused on the need for large amounts of data analysis requirements. According to the
specified interval, the analysis results will be symbolized, colored and loaded into 3D
scene.

5.2 Load Rate Analysis

Line load rate analysis can provide reference for transmission scheduling. Power
network topology is the basis of the load rate analysis. Therefore, in the 3D modeling
process, network features, topology, and semantics will be stored directly in the
CityGML, and then converted into relational database. The analysis process is divided
into the following steps: (1) obtaining network topology and line attribute information;
(2) calculating the real-time load rate of each line according to line attribute and
measurements; (3) using the frequent sub-graph mining algorithm to analyze temporal
and spatial distribution pattern; (4) coloring and symbolizing the analysis results
directly in three-dimensional network connection model or 3D virtual scene as
geospatial map layers. Thus, according to the load rate of overhead lines and cable lines
and taking the value as the Z-axis coordinate, the results can be dynamical visualized
by line types, width, colors, or the combinations form. As well, the continuous-time
mining results can constitute a three-dimensional time-series map animation, and
shown the change of load rate, as well as spatial distribution at different time periods
(Fig. 7).

Fig. 7. Dynamic visualization of load status

 The Research of 3D Power Grid Based on Big Data 97

6 Conclusions

Considering the drawback of common 2D power system, this paper proposes a 3D

power visualization framework based on big data technologies and 3D modeling
techniques to achieve flexible display of large multidimensional power data in dynamic
panoramic scene. The proposed framework consists of data acquisition and conversion,
big data platform, power data analysis and mining, and 3D spatialization of analysis
results. Then, we detailed the visualization techniques and modeling elements in 3D
power system from geospatial data, electric graphic data, electric attribute data and
power analysis and data mining result dataset. Finally, we give two use case of power
system. Most of the mentioned tools and software are open source and integrated
together to ensure high performance and presentation capabilities of the 3D framework
while avoiding unnecessary duplication of development work. This research can guide
researchers and power company for the development of 3D big data power system.

Acknowledgment. The work described in this paper is supported by the National High Tech-
nology Research and Development Program of China (863 Program) (2015AA050201), and the
Chinese National Natural Science Foundation (41701167) and (6433012). We gratefully
acknowledge the valuable cooperation and help of Guangzhou Power Supply Co. Ltd.

References
1. Liu, Q.: Data fusion and application of power equipment based on big data on Panorama.
Informatization Constr. (9), 63–65 (2015)
2. Zhou, J.Z., Hu, R.Q.Y., Qian, Y.: Scalable distributed communication architectures to
support advanced metering infrastructure in smart grid. IEEE Trans. Parallel Distrib. Syst. 23
(9), 1632–1642 (2012)
3. Fang, X., Yang, D.J., Xue, G.L.: Evolving smart grid information management cloudward: a
cloud optimization perspective. IEEE Trans. Smart Grid 4(1), 111–119 (2013)
4. Vu, H., Liu, J.K., Huang, X.: A secure cloud computing based framework for big data
information management of smart grid. IEEE Trans. Cloud Comput. 3(2), 233–244 (2014)
5. Maheshwari, K., Birman, K., Wozniak, J., Zandt, D.V.: Evaluating cloud computing
techniques for smart power grid design using parallel scripting. In: Proceedings of 2013 13th
IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGrid),
Persistent Link, pp. 319–326 (2013)
6. Ravikumar, G., Pradeep, Y., Khaparde, S.A.: Graphics model for power systems using
layouts and relative coordinates in CIM framework. IEEE Trans. Power Syst. 28(4), 3906–
3915 (2013)
7. W.Zhangjin,H.Xingwen,D.Ying,Z.Rui,Z.Qingguo, “A CGI + AJAX + SVG based moni-
toring method for distributed and embedded system,” in Proc. 2008 IEEE 1st Int. Conf. UBI-
Media Computing,2008,pp. 144–148
8. Tang, K., Zhong, Y.J., Wu, H.F., He, P.Y., Xu, L.J., Qi, W.Q.: Design and implementation
of GIS basic data quality management tools for power network. Electr. Power Inf. Commun.
Technol. 14(2), 98–101 (2016)
9. Bai, Y.Q., Nie, J., Chen, Y., Li, Z., Chen, T.: Construction of GIS platform for Chongqing
grid. J. Geomatics 37(5), 73–75 (2012)
98 B. He et al.

10. Wei, W., Liu, B.H., Dai, J.Z., Xia, H., Li, Z.: Design and implementation of GIS-based
Wuhan Grid management systems. China Electr. Power 37(1), 161–163 (2009)
11. Zhang, D.X., Miao, X., Liu, L.P., Zhang, X., Liu, K.Y.: Research on development strategy
for smart grid big data. Proc. CSEE 35(1), 2–12 (2015)
12. Apache Sqoop. http://sqoop.apache.org/
13. Apache Storm. http://storm.apache.org/
14. Apache Flume. http://flume.apache.org/
15. Apache Hadoop. http://hadoop.apache.org/
16. YARN. http://hadoop.apache.org/docs/current/hadoop-yarn/hadoop-yarn-site/YARN.html
17. HBASE. http://hbase.apache.org/
18. HIVE. http://hive.apache.org/
19. Impala. http://impala.io/
20. Spark. http://spark.apache.org/
21. SketchUp. http://www.sketchup.com/
22. Skyline TerraExplorer Pro. http://www.skylinefrance.com/en/home
23. ESRI CityEngine. http://www.esri.com/software/cityengine/
24. Autodesk Revit. http://www.autodesk.com/products/revit-family/ overview
25. 3D City Database. http://www.3dcitydb.org/3dcitydb/3dcitydbhomepage/
26. Cesium. https://cesiumjs.org/
27. WebGL Earth. http://www.webglearth.com/
28. World Wind. https://goworldwind.org/
29. ossimPlanet. https://trac.osgeo.org/ossim/
 Identifying Service Gaps from Public Patient
Opinions Through Text Mining

Min Tang1, Yiping Liu1 ✉ , Zhiguo Li1, and Ying Liu2

( )

1
School of Management, Chongqing Technology and Business University,
Chongqing 400067, China
[email protected], [email protected]
2
School of Engineering, Cardiff University, Cardiff, CF24 3AA, UK

Abstract. Nowadays, healthcare systems have become increasingly patient-

centered and the unstructured, open-ended and patient-driven feedback has drawn
a significant attention from medical and healthcare organizations. Based on this,
we are motivated to harness various machine learning algorithms to process such
a large amount of unstructured comments posted on public patient opinion sites.
We first used sentiment analysis to automatically predict the concerns of patients
from the training set which was already labelled. Then, with the help of the clus‐
tering, we extracted the hot topics related to a specific domain to reflect the service
issues that patients concern most. Through experimental studies, the performance
of different algorithms and the influence of different parameter were compared.
Finally, refering to the survey and previous studies, the results were analyzed to
obtain the conclusions.

Keywords: Text mining · Sentiment analysis · Clustering analysis

Public health service

1 Introduction

At the present stage, it is widely accepted that collecting, responding and processing the
comments from customers is extremely important to the organizations. The feedback
from customers has been a crucial source to evaluate the quality of service and products.
With the development of the technology, reporting their experience and public opinions
on the Internet has become a popular way for patients and customers [1]. In the area of
the service industry, large organizations have performed a huge number of studies to
discover the reasonable approaches to process the comments from consumers. Sentiment
analysis and text mining based on the natural language processing and machine learning
algorithms play a critical role in these studies [2]. Obviously, it could be meaningful if
sentiment analysis can be conducted in the field of healthcare. Through this, the infor‐
mation of sentiment and emotion can be captured from the textual data of patients’
experience. Academic research in this area has largely focused on establishing the
models with higher accuracy for ‘automatically predicting the sentiment embedded
within the text’. Additionally, based on the research of the National Health Service
(NHS) (2013) in England, the sentiment analysis witnessed it could become the novel

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 99–108, 2018.
https://doi.org/10.1007/978-981-13-2384-3_10
100 M. Tang et al.

resource for patients when they were making choice of hospitals. Based on this, the test
for patients’ satisficing to the service of hospitals will be conducted by analysis a vast
amount of textual data on the Patient Opinion website in this project. Three questions
will be focused on, which are whether this hospital could be proposed by the specific
patient, whether the attitude of the staff is appropriate and whether the waiting time is
reasonable.

1.1 Feedback Collection

In recent years, the study of recipient experience has drawn a significant interest. And
the researchers continue in the field of extraction and analysis. As introduced by the
study performed by Sun and Li (2011), it is important for organizations to collect and
analyze the recommendations of customers, since it allows companies to adopt the
service to customers’ expectations in a developing manner. Based on the research of
McAfee and Brynjolfsson (2012), there is five percent growth of productivity and six
percent growth of profitability in the companies can process the comments. The tech‐
nological advances expand the source for hospitals to collect the customer comments.
Then the sentiment analysis, which is helpful to assess people’s feelings for the services
and products, is convenient to be carried out. In addition to these, comparing with the
traditional methods, which often have the low response rate [3], the quantity of the
materials from the Internet and social media is extremely wide. Above all, the usage of
the Internet can obtain the achievement of the wide source, timeliness and convenience.

1.2 Text Analysis

Textual comments have attracted considerable attention for the past few years. Literature
highlights that marketing department had used manual and machine learning methods
to process the information. The deeper understanding of the comments can be gained if
the analysis is performed with the manual manners. However, the review for the huge
quantity of text might be inconsistent without ‘procedural models’ [4].
In terms of the medical service, online health communities and the patient-authored text
they contained, accompanied by the potential value, have attracted more and more atten‐
tion. Graebner et al. (2012) suggested that the major direction of consumer comments anal‐
ysis with text mining has basically focused on establishing more accurate models for fore‐
casting the sentiment embedded within the textual recommendations. Previous studies
showed that the linguistic text mining approaches had better performance than that of
manual methods in the categorization of the comments because they performed better in the
precision of prediction. In the field of medicine, enterprises like Google Flu [5] and
HealthMap [6] have already demonstrated that the PAT (patient-authored text) is a depend‐
able source for illness tracking and medical research. Additionally, websites such as Cure‐
Together and PatientLikeMe have put great effort into the research for effect exerted by the
patient opinions on the healthcare service.
A large number of works focused on the sentiment analysis, which was also regarded
as opinion mining. It involved the study of the analysis of person’s opinion, sentiments
to the product, services etc., and the factors influence them. Machine learning means the
 Identifying Service Gaps from Public Patient Opinions 101

model is trained on the structure of the data. In terms of sentiment analysis, the literature
displays several key aspects of it. The bag-of-words approach can be used to process
the text with unigrams, bigrams and trigrams in the analysis. With consideration of
potential emotional comments, the sentiment could be determined in several ways by
the algorithm. As can be seen in the research of the feedback of the healthcare experience
on Twitter in the US, monitoring social media can obtain the meaningful, unsolicited
and real-time data, while the traditional survey system cannot capture it. One of the most
important reasons is social media can provide the opinions from patients that cannot be
obtained from other routes with the cautious observation [7].
Besides those manners of classification, automatic topic detection based on docu‐
ment clustering is an alternative approach for extracting health-related hot topics in
online communities [8]. Patterson (2011) illustrated that document clustering is a valid
manner to clearly differentiate the groups of clinical narratives. Recently, many studies
applied clustering manners to this user-generated medical text to discover the topics that
interest patients who post their experience online. The study of Denecke focused on
blogs related to healthcare and separated the topics of these blogs into two groups:
informative and affective [9]. Brody (2010) processed the texts based on Latent Dirichlet
Allocation (LDA) topic models to explore the salient directions of online evaluations of
healthcare systems. Therefore, it is valuable to perform the clustering approaches in text
mining.
In summary, the main mission of the service industry, such as the healthcare systems,
is offering the satisfactory service to meet the requirements of the recipients. Addition‐
ally, since people rely increasingly upon the internet and social networks, the amount
of the text present online continues to rise sharply, which provides organizations with
a source to exploit business intelligence. Moreover, collecting and analyzing consumer
comments witness their importance have a continuous increase. Hence improving the
service quality based on the advice from the Internet and social media has drawn signif‐
icant attention because of its vast potential value, outstanding convenience and high
efficiency. Previous literature highlights several research limitations, such as the manual
preprocess, size of the dataset and the variety of specialities, which also point out the
directions of further research.

2 Methodology

The data analysis techniques were conducted to the textual comments from Patient
Opinion about the service experience of cancer patients. The objective was to test
whether the prediction for patients’ views on several topics could be obtained automat‐
ically based on their feedback. The machine learning classification approach was chosen
and trained from the labelled data, which was regarded as the training group. And the
comments can be classified into various categories based on the given examples, with
the help of the open-source data mining software—Weka, which had been widely
adopted in the previous researches of text mining in many domains. To scale the relia‐
bility of the experiments, the results of this research were compared with the quantitative
rating provided by the Patient Opinion. There were two domains of the service of
102 M. Tang et al.

healthcare predicted, which were waiting for time and attitude of the staff. In terms of
clustering, after converting the textual data into the valid format of Weka, the expectation
maximization (EM) clustering was conducted.

2.1 Source of Data and Pre-processing

In this study, Patient Opinion website was used as the data source. Next, the represen‐
tative comment fields should be determined. Based on the statistical widgets of Patient
Opinion, cancer was one of the most common diseases of the comments. There were
the tools can analyse the 5 most concerned words mentioned in the feedback for ‘What
could be improved’ in the last week. ‘Waiting time’ and ‘communication’, which had
the most occurrence, were chosen as the measurements - ‘whether the waiting time is
reasonable’ and ‘whether the communication between patients and staffs is unblocked’.
Thus, this paper chose cancer patients as research subjects and collects the comments
related to waiting time and communication.
In the area of sentiment analysis, we conducted data processing techniques on all the
online free-text comments posted by cancer patients about the treatment attitude and
waiting time on a Patient Opinion website. The Patient Opinion was asked for helping
to collect the textual data as the training set for the research, which included the works
like removing the meaningless labels, lowercasing the text and tuning each sentence into
a separate line. The texts were split into two parts, which establishes two contents - ‘neg’
and ‘pos’. The comments with positive sentiment labels were stored in ‘pos’ content. In
contrast, the ‘neg’ content was composed of the negative evaluation. Each comment was
an individual file. For clustering analysis, Offline Explorer, one of the web crawler tools,
was used to receive all the web pages from the comment platform. Additionally, to obtain
the better topic clustering results, some sentences unrelated to the topics were dislodged.

2.2 Sentiment Analysis and Text Classification

Sentiment analysis, which is also named opinion mining, is to process people’s opinions
and extract the sentiments towards the products and services [10]. Machine learning is
about learning from the structure of predefined data and discover the principles and
functions can process the unlabelled information [11]. There are two main types of
machine learning algorithms, which is supervised and unsupervised respectively. Senti‐
ment analysis or text classification should be involved by supervised fields, which means
that the training data should have already had the class labels by the start of the operation.
The fundamental process of supervised machine learning method for sentiment analysis
was shown in Fig. 1, the labelled data should be provided as the training set and the
algorithm can learn and output the trained model. The reliability of the model should be
tested with the unlabelled data.
 Identifying Service Gaps from Public Patient Opinions 103

Labelled Data
(Training set) Unlabelled Data

Machine Learning
Algorithm Model Trained Model

Prediction

Fig. 1. The fundamental process of supervised machine learning method for sentiment analysis

A large number of different technical methods could be conducted in classification

in machine learning. In this paper, 3 popular supervised machine learning algorithms
were selected to perform the sentiment analysis, which was ‘Naïve Bayes Multinomial’
(NBM), Decision Trees and Support Vector Machines.

2.3 Clustering Analysis

Recently, text clustering has been extensively adopted to discover the most concerned
topics related to medical service. Since there were a significant amount of medical-
related words occur in those comments, UMLS Metathesaurus, which had often been
used in previous studies, was adopted to extract medical concepts. As listed in Table 1,
the most frequent health-related and service-related concepts were extracted. And all
the features’ domains used in this research are also shown in Table 1. However, to input
the received features to the topic clustering, the results should be quantified as feature
vectors, which are characterized by ‘high dimensionality, redundancy and high corre‐
lation among individual attributes, which is unfavourable for clustering’ [12]. Principle
component analysis (PCA), which had been proved by many experiments as the effective
method to reduce features, was adopted to solve this problem.

Table 1. All the features’ domains used in this research

Category Types of key phrases Category Types of key phrases
Keywords n-gram Semantic terms
1-gram Service domain Service terms
2-gram Semantic terms
Medical domain Medical terms

In this paper, one of the typical probability-based clustering-Expectation Maximi‐

zation (EM) clustering was selected. To enhance the performance of distinguishing the
different clusters and identifying the topics from the clusters, many keywords were
selected to differentiate the topics. These key phrases were n-gram words and included
the medical concepts and terms in service evaluation from UMLS lexicon. Several
approaches to extract these keywords had been used, in this paper, they were selected
from the function below, where f is the frequency of w in the cluster ci, and the bottom
of the formula is ‘the total number of clusters with a term w frequency greater than or
equal to the term w frequency of the cluster being evaluated’. Once the key phrases have
104 M. Tang et al.

high scores were ‘ranked and combined with expert opinions’, the topics of the clusters
were identified.

N
f (w, ci) ∗ log (1)
|{cj|f (w, cj) ≥ f (w, ci), j = 1, 2, … , N|}|

2.4 Performance Measurement

A large number of different technical methods could be conducted in classification in

machine learning. In this paper, 3 popular supervised machine learning algorithms were
selected to perform the sentiment analysis, which was ‘Naïve Bayes Multinomial’
(NBM), Decision Trees and Support Vector Machines. The principle of Decision Trees
is classifying the data into different classes by ‘recursively separating the feature space
into two parts and assigning different classes based upon which region in the divided
space the sentence is’ [13]. The Support Vector Machines method obtains the classes
with the help of the support vectors, which represents the boundary of different groups.
These vectors are continuously maximized until it is large enough to differentiate the
difference [14]. In addition to this, the comparison was also applied to detect which gave
the quickest and most accurate outcomes. In feature extraction, the feature matrix was
established by separating the sentences and texts into words. Within the matrix, the
sentences and texts were the indexes in a row and the column consisted of all of the
words occurred in the documents. The measurement of the value was calculated based
on the frequency count of each word in each comment, the size of the comment and the
occurrence of every word in the whole document. After the matrix had been built up,
the feature matrix can be operated by each model to output the results and the perform‐
ance could also be evaluated. The measurements are accuracy (the percentage of correct
outputs from the whole database), F-measure (‘the harmonic mean of precision and
recall’ [11]), Recall index and cost time. Finally, the J48 Decision Trees was selected
to process the test set.

3 Results and Discussion

3.1 Sentiment Analysis

To reduce the processing time of classification, the number of words in the learning
process was limited to no more than 10,000 based on the frequency of each word. After
the experiment, the prediction performance of every method was compared with that of
others. The results outputted by Naïve Bayes Multinomial, J48 Decision Trees and
Support Vector Machine of 3 domains of healthcare service (overall evaluation to the
service, whether the communication is good and whether the waiting time is reasonable)
were shown as follow (Tables 2 and 3).
 Identifying Service Gaps from Public Patient Opinions 105

Table 2. The results of the overall evaluation, waiting time, communication and the overall
evaluation without ‘Stop Word List’
Overall rating Waiting time Communication Overall rating without ‘Stop
Word List’
NBM J48 SVM NBM J48 SVM NBM J48 SVM NBM J48 SVM
Accuracy 86.60% 76.30% 78.60% 80.20% 80.30% 81.60% 83.60% 81.60% 80.70% 85.90% 72.60% 72.90%
F-measure 0.87 0.81 0.82 0.83 0.85 0.82 0.85 0.8 0.81 0.85 0.81 0.8
Recall 0.87 0.81 0.82 0.83 0.85 0.82 0.85 0.8 0.81 0.85 0.81 0.8
Time (s) 0.08 226 316 0.03 36 61 0.03 92 106 0.09 252 332

Table 3. The performance of three algorithms in overall evaluation without ‘Tokenizer’

NBM Accuracy F-measure Recall Time No. of extraction
2-gram 73.60% 0.74 0.74 0.06 2676
3-gram 71.20% 0.71 0.71 0.05 1360
N(1-3)-gram 83.20% 0.84 0.84 0.04 7960
WordTokenizer 83.60% 0.85 0.85 0.03 863
Alphabetic 80.90% 0.85 0.85 0.03 792
J48 Accuracy F-measure Recall Time No. of extraction
2-gram 72.90% 0.73 0.73 151 2676
3-gram 72.10% 0.71 0.71 140 1360
N(1-3)-gram 82.80% 0.8 0.8 106 7960
WordTokenizer 83.50% 0.8 0.8 92 863
Alphabetic 80.30% 0.8 0.8 87 792

Based on the results, the NBM algorithm observed that it had the best performance
in the overall assessment. Then, obviously, the accuracy of 3 approaches changed in the
experiment for ‘waiting time’ and ‘communication’. On the one hand, the J48 Decision
Trees and Support Vector Machine performed better, which increased roughly 3%. The
comparison also suggests that the NBM could perform better with the large database.
Conversely, the J48 DT and SVM algorithms could outperform the NBM if the size of
the dataset is not large enough. In terms of the influence of ‘StopWordList’, it displayed
that the list had a helpful effect on the work of the algorithms. All 3 methods became
more accurate with using the Weka’s list of stop words. Moreover, the F-measure and
Recall index also increased slightly with the time taken to complete the process
becoming less. For the final prediction of the test set (Table 4), there was 77% comments
showed the service satisfied the patients, which was fundamentally in a similar range
with the quantitative survey and the statistical analysis of Patient Opinion.

Table 4. The results of the trained model

Overall rating Positive Negative
Overall 770 230
Waiting time 926 74
Communication 756 244
106 M. Tang et al.

The results indicate that the output of the sentiment analysis of the comments can
be efficient and have reasonable accuracy [15]. And it can accomplish the effective
extraction and identification of the critical issues of the review. It can also prove that
there are dramatic meaningful information in the unstructured feedback of patients and
it could significantly benefit the organizations if processed appropriately. With the help
of the sentiment analysis, in both overall view and certain domains, an aggregate eval‐
uation of objectives can be provided.

3.2 Clustering Analysis

In the area of clustering analysis, after pre-processing and features extraction, the EM
clustering was conducted by Weka. After operating the EM cluster, the results were
organized based on their ranks and semantic relation to the medical care and service
evaluation. The final results were displayed as follow (Table 5).

Table 5. Clustering analysis results

Label Key phrases
Staff Good, bad, brilliant, excellent, professional, amazing, impressed,
shortage
Communication Respect, lack of efficient, nice, ignore, wait, staff manner
Care Good, patient, professional, empathy, dignity, compassion,
brilliant
Nurse Nice, excellent, lack, patient, amazing, ignored, professional
End of life Well treated, respect, appreciate
waiting time (negative) Hour, still, emergency, appointment
Communication (negative) Lack of, ignore, rude, waiting

In the ‘staff’ group, basically, the patients’ views to doctors and staffs were satis‐
factory. There was one word deserved to be mentioned - ‘professional’, which was
mentioned more times in writers’ praises than other words. It suggested that the patients
may pay more attention to attitudes and skills in the professional field instead of char‐
acters and personalities. On the negative side, patients and reporters often criticized the
topic of ‘shortage’, which showed that the shortage of staffs was still the critical problem
in the healthcare system. Comparing the topics around ‘staff’, the second group labelled
‘communication’ was comprised of features much more specific. ‘Respect’ and ‘effi‐
cient’ are two domains with which the criticisms were most concerned. There were also
4 words displayed the drawbacks of the communication, which was ‘poor’, ‘ignored’,
‘lack of’ and ‘no’, and showed the ignorance, might partly as a result of overworked
status, of the staffs and nurses is the most serious problem. In terms of ‘care’ label, it
involved the key features of the care service itself and the results observed the similar
situation. The praises from patients were almost full of the texts about care and
‘patience’, ‘empathy’, ‘dignity’ and ‘compassion’ are the aspects with the most concern
of recipients. The number of comments about nurses was considerably more than that
about other workers in the healthcare systems, which made the nurse become the
 Identifying Service Gaps from Public Patient Opinions 107

essential part of the hospitals. Among all of these words, ‘patience’ played the most
important role in service provided by nurses. There was the other special label is ‘end
of life’, which represented the status of cancer patients in their last time. These comments
were often written by their relatives and families. The praise also occupied most of the
texts and the most important word in this field was ‘respect’, which suggested, to the
patients in their final time, respect was crucial to them.
To discover the drawbacks of the healthcare service, the comments with negative
sentiment in the fields of ‘waiting time’ and ‘communication’ were selected as our
objectives. Obviously, it can be observed that patients with emergency and appointments
were more likely to complain about the waiting time. Since the words involve ‘hour’
were much more the phrases have ‘minute’. Therefore, the further study of the limit
needs to be performed here. In the area of ‘communication’, which the ignorance and
rude manners of staffs mentioned most times in the comments.
Above all, professional skills of staff, nurses’ patience and the manners of commu‐
nication have drawn the most attention from patients. The shortage of staff, overworked
schedules and ignorance for an effective communication are the problems that need to
be improved immediately.

4 Conclusion

This work demonstrates that it is possible for organizations to predict the evaluation
embedded with the huge amount of comments through sentiment analysis and explore
the hot medical related topics through clustering analysis. By conduct these approaches
based on text mining, several valuable conclusions can be drawn from the obtained
results. Most of the reporters from Patient Opinion have a basically satisfactory expe‐
rience but the comments mention the ‘waiting time’ observe that the negative feedback
occupies the larger part than the positive group, which is mostly because of the appoint‐
ment and emergency. Then, in the field of communication, more than 70% of the
comments tagged with positive sentiment and efficiency and manners of staff play the
crucial roles in communication between patients and hospitals. The criticism for the
communication mainly focuses on the ignorance of the help workers. In addition, for
nurses, patients and their families, they pay more attention to their patience while doctors
and staffs are praised mainly because of their professional skills. Finally, respect plays
the most important role in the healthcare for cancer patients in their last days.

References

1. Greaves, F., Millett, C.: Consistently increasing numbers of online ratings of healthcare in
England. J. Med. Internet Res. 14(3), e94 (2012)
2. Tumasjan, A.: Predicting elections with Twitter: what 140 characters reveal about political
sentiment. In: Fourth International AAAI Conference on Weblogs and Social Media,
Washington DC, pp. 178–185 (2010)
3. Zimlichman, E., Levin-Scherz, J.: The coming golden age of disruptive innovation in health
care. J. Gen. Intern. Med. 28, 865–867 (2013)
108 M. Tang et al.

4. Ziegler, C., Skubacz, M., Viermetz, M.: Mining and exploring unstructured customer
feedback data using language models and treemap visualizations. In: IEEE/WIC/ACM
International Conference on Web Intelligence and Intelligent Agent Technology, pp. 932–
937. IEEE, Sydney (2008)
5. Ginsberg, J.: Detecting influenza epidemics using search engine query data. Nature 457,
1012–1014 (2008)
6. Freifeld, C.C.: HealthMap: global infectious disease monitoring through automated
classification and visualization of internet media reports. J. Med. Res. 15, 150–157 (2008)
7. Greaves, F., et al.: Use of sentiment analysis for capture patient experience from free-text
comments posted online. J. Med. Internet Res. 15(11), e239 (2014)
8. Lin, Y., et al.: A document clustering and ranking system for exploring MEDLINE citations.
J. Am. Med. Inform. Assoc. 14, 651–661 (2007)
9. Denecke, K., Nejdl, W.: How valuable is medical social media data? Content analysis of the
medical web. Inf. Sci. 179, 1870–1880 (2009)
10. Pang, B., Lee, L.: Opinion mining and sentiment analysis found. Trends Inf. Retr. 2(1–2), 1–
138 (2008)
11. Ivanciue, O.: Weka machine learning for predicting the phospholipidosis including potential.
Curr. Top. Med. Chem. 8(18), 1691–1709 (2008)
12. Witten, I.H., Frank, E.: Data Mining: Practical Machine Learning Tools and Techniques.
Morgan Kaufmann Publishers, San Francisco (2005)
13. Frank, E., et al.: Data mining in bioinformatics using Weka. Bioinformatics 20(15), 2479–
2481 (2004)
14. Li, J., et al.: Discovery of significant rules for classifying cancer diagnosis data.
Bioinformatics 19(Suppl. 2), 1193–2103 (2003)
15. Alemi, F., et al.: Feasibility of real-time satisfaction surveys through automated analysis of
patients’ unstructured comments and sentiments. Qual. Manag. Health Care 21(1), 9–19
(2012)
16. Abegaz, T., Dillon, E., Gilbert, J.E.: Exploring affective reaction during user interaction with
colors and shapes. Proc. Manuf. 3(Suppl. C), 5253–5260 (2015)
17. Dong, A., Lovallo, D., Mounarath, R.: The effect of abductive reasoning on concept selection
decisions. Des. Stud. 37(Suppl. C), 37–58 (2015)
18. Evans, P.: From deconstruction to big data: how technology is reshaping the corporation.
MIT Technol. Rev. (2015). Stanford, California
19. Hsu, F.-C., Lin, Y.-H., Chen, C.-N.: Applying cluster analysis for consumer’s affective
responses toward product forms. J. Interdiscip. Math. 18(6), 657–666 (2015)
20. Chen, R., Xu, W.: The determinants of online customer ratings: a combined domain ontology
and topic text analytics approach. Electron. Commer. Res. 17(1), 31–50 (2017)
21. Holy, V., Sokol, O., Cerny, M.: Clustering retail products based on customer behaviour. Appl.
Soft Comput. 60(Suppl. C), 752–762 (2017)
 How to Verify Users via Web Behavior
Features: Based on the Human Behavioral
Theory

Jiajia Li, Qian Yi(&), Shuping Yi, Shuping Xiong, and Su Yang

College of Mechanical Engineering, Chongqing University, Chongqing 400044,

China
[email protected]

Abstract. Nowadays, the Internet has penetrated into people’s daily life in
many ways, and people usually use the form of web account to carry out web
activities. However, the problem of user account being illegally occupied by
others has become increasingly prominent, and the security of user account is
not guaranteed. On the other side, the network’s advantage of recording data
easily provides a valuable opportunity to comprehensively record a tremendous
amount of human behaviors. Based upon the status and problems, we present an
approach to distinguish genuine account owners from intruders. This study
constructs a series of web behavior features based on web usage logs, and uses
machine learning to identify users based on Bayesian theorem and structural risk
minimization criterion. Experiments show that web behavior analysis is a fea-
sible way for user verification.

Keywords: Human behavior
Human web behavior
User verification

Machine learning

1 Introduction

With the rapid development of Internet technology, people are becoming more and
more inseparable from the Internet. According to the 39th report of CNNIC (China
Internet Information Center) “By the end of 2016, the number of netizens in China has
reached 731 million, and the total number of new netizens has increased 42 million 990
thousand over the whole year” [1]. The development of mobile Internet has brought
great convenience to people’s life. It also provides support platform for information
sharing, which further reduces transaction costs and facilitates communication. In
general, it improves the efficiency of the overall resource utilization. But at the same
time, there are also various security problems of network. So far, the common risk we
faced is that the user’s account number is embezzled. Under the condition of account
password leakage, there are some shortcomings in traditional protection methods,
which makes user verification become one of the key issues in account security. In
order to solve these problems, we decide to use the real time updated web log and adopt
the web behavior features which based on the behavior analysis to carry on the con-
tinuous security monitoring to the user.

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 109–120, 2018.
https://doi.org/10.1007/978-981-13-2384-3_11
110 J. Li et al.

The field of human web behavior, elucidation of human behavior in the web
environment is quickly gaining eminent position [2]. Guo et al. establish a dynamics
model of human interest attenuation based on the data of blog releasing. The results
show that blog arrival time interval distribution is a mixed distribution with exponential
and power-law feature. It demonstrates the variety of human web behavior dynamics
[3]. Neelima et al. dig the frequency of users accessing each page, and find users’
sessions to analyze users’ time spent on a specific page [4]. Zhong et al. calculate the
features for every page based on analysis of association rules, then build up a per-
sonalized user browsing behavior model [5]. Via click stream analysis, Leung et al.
contrast the different behavioral patterns of consumers which during weekdays and
weekends [6]. Cui et al. conducts a pioneering research in how to use model or method
to represent human behaviors. They mention that forecasting and predicting are major
goals in modeling human behaviors, but it also faces various challenges [7]. In most of
existing research, some of which discuss web behavior from the time dimension, and
others focus on the analysis of the connection between users and browsing pages. But
unlike those studies, this paper maps the behavior theory of the real world into the
virtual world, so as to explore the user web behavior from different dimensions.
On Classification and Prediction, Lin et al. use three integrated feature sets
(EEG + Modeling, EEG + Task, and EEG + Task + Condition) training a linear dis-
criminant analysis classifier, then integrating human behavior modeling and data
mining to predict human errors [8]. By leveraging observable information of the user’s
Twitter feedback, Pennacchiotti and Popescu use Gradient Boosted Decision Trees-
GBDT to build user classification model. Finally, automatically infer the values of user
attributes, such as political orientation or ethnicity [9]. According to the characteristics
of user interaction with the mouse, Feher et al. introduce a method to verifies users
[10]. Pao et al. propose a user verification approach based on user trajectories. Such as
on-line game traces, mouse traces, handwritten characters, and traces of the movements
of animals in their natural environments [11]. These studies not considered user
behavior related to physiological and psychological needs comprehensively, therefore,
the current methods do not fully reflect the individual differences of users, nor can they
guarantee enough effectiveness. Based on the web behavior which is refined by web
log analysis, we propose a general user verification approach, which does not require
any extra sensing equipment.
Based on the characteristics of individual differences in human web behavior, this
paper is no longer just a study of the relationship between users and web pages but
make use of Web logs to construct a series of behavior features which are related to
user habits, and then, it draws support from two kinds of machine learning methods to
verify users identity.

2 Feature Construction

2.1 Human Web Behavior

Human behavior is a meaningful reaction, which is in order to adapt the external
environment. Generally speaking, most of these daily activities are some customary
 How to Verify Users via Web Behavior Features 111

patterns of action, besides, people’s attitude will change their behavior accordingly
when they are in a state of discomfort [12]. In this study, web behavior refers to the
action that takes place on human-computer interface when people surf the Internet.
Compared with the traditional entity scene, the existence of the web log to make it
possible for accurately and comprehensively collection of human behaviors in the
absence of other additional equipment. The internal drive of web behavior includes
psychological and physiological demands together with the way of thinking, which
same as the offline behavior. Moreover, it combined with the constraints on actual
conditions can decide the possibility and trend of behavior. In the same way, web
behavior also shows five major characteristics as bellows: Adaptability; combined with
the work and rest time, user always choose the most convenient time to surf on Internet,
besides, the size of browser window is also depend on their vision and operating
environment. Relevance; in fact, most of the time user care more about what they are
interested in, and the occurrence of web behavior is related to it, meanwhile, it also
affects the persistence and frequency of behavior. Diversity; user can perform web
browsing, mouse clicking, input text and other multiple forms of action on the human-
computer interaction interface. Dynamic and Controllability; If the external environ-
ment or user’s motivation changes, user can decide whether to change the current
action and how to change it.

2.2 Definitions
Based on the above theory, we analyze the web logs of AML (Advanced Manufac-
turing Technology) to get 13 features of user behavior. The definitions and descriptions
of these web behavior features used in this study are given below.
(1) Week: It indicates that the behavior is happened in which day of the week. It
transforms by the ‘timestamp’ which is recorded on web log. First, read the data
of ‘timestamp’ from web log, such as 2017-05-25T19:47:19.123Z. Next, the
information of year, month and date was selected. Finally, convert it to ‘week’.
For example: ‘2017-05-25T19:47:19.123Z’ is transformed into ‘Thur’.
(2) Time: It indicates that the behavior is happened in which time of the day. It
transforms by the ‘timestamp’ which is recorded on web logs. First of all, we
divide the 24 h of a day into 6 parts, and the partition rule is shown in Table 1.
After that, we extract the information in the ‘timestamp’ which represents hours,
and transform it into the Time’s value in the end.

Table 1. The partition rule of a day

No. Time Hours
1 Before dawn 00:00–06:00
2 Morning 06:00–11:00
3 Noon 11:00–13:00
4 Afternoon 13:00–16:00
5 Evening 16:00–18:00
6 Night 18:00–24:00
112 J. Li et al.

For example: ‘2017-05-25T19:47:19.123Z’ is transformed into ‘Night’.

(3) W/H: The ratio of width of height of user browser window, and it can reflect the
interface shape of human-computer interaction. The size of the browser window is
variable, and which is determined by the user’s usage habits. We set 5 ranges of
this value according to the shape that has appeared, as showed in Fig. 1. The
width and height of the browser window are directly recorded by web logs. After
calculating their ratios, the results are classified into 5 ranges.

Fig. 1. The shape of the browser window

(4) P-Web: The ratio of web browsing height to the total height of the web page. It
should be noted that the height of user’s mouse sliding on a Web page is called
web browsing height. Generally, users will browse more content when the page is
more consistent with his interests. The height of web browsing and the total height
of web page are directly given by web logs.
(5) Event: The type of operation event that user does on web pages. There are 8 types
involved in this study, as illustrated in Table 2.

Table 2. The event types of web page

Event Explain
View Users browse web pages
Click Click behavior of the mouse
Behave Stay for a time on web pages
Fomel The act of filling out a form
Formsub The act of submitting a form
Download_paper Downloading operations
Trail Mouse movement
Error Do the wrong operation

(6) Language: The language used by the user for typing. In our study, this feature’s
value is mostly English or Chinese, which reflects users’ habitual language and is
related to their living environment.
 How to Verify Users via Web Behavior Features 113

The features involved in this paper also include:

(7) NumS: It’s the number of views of the current session.
(8) UP: The web page where the user is located.
(9) URL: The destination page of a jump in a web site.
(10) Tp: The staying time of a web page.
(11) LTp: The loading time of a web page.
(12) Click-X: The X axis position of the mouse click.
(13) Click-Y: The Y axis position of the mouse click.
The above web behavior features directly or indirectly reflect the characteristics of
web behavior. Figure 2 describes the correspondence between the 13 features and the 5
characteristics.

Fig. 2. The correspondence between features and characteristics

3 Methodologies

In this section, we describe the process of user verification and explain the two machine
learning methods that we use: Naïve Bayes and Support Vector Machine (SVM). The
flow chart of the process is shown in Fig. 3. The process is implemented in five steps.
The first step is to obtain the data from the data source, then select the high frequency
operation users as the research objects. The second step is to deal with the data
according to the principle of feature construction in the Sect. 2. The third step is to
further process the data so that it meets the input requirements of machine learning. The
fourth step is to use the classifier called Naïve Bayes and SVM to determine whether
114 J. Li et al.

Fig. 3. The flow chart of the process to user verification

the web behavior belongs to the account owner. The fifth step is to discuss the per-
formance of the classifiers.

3.1 Data Collection

AML is an academic exchange website of the industrial engineering of Chongqing
University: http://www.cquieaml.com. The web log of AML is the Data Source for our
study. Users can do various operations in different sections of the website, such as
browse web pages, upload or download the datum, publish a topical subject, comment
interaction and so on. It is worth mentioning that the levels of the user are divided into
administrators, team members or others, and different levels are set up different
functional privileges. In addition, a person who has not landed will be unable to browse
the website.

3.2 Classification Model

The essence of user verification is to determine whether users are themselves through
web behavior features. It is a problem of binary classification, and the mathematical
definitions of the classification problem is as follows:

C : C ¼ fy1 ; y2 ; . . .; yn ; . . .g

The set C represents the set of all categories, yn represents the category n.
 How to Verify Users via Web Behavior Features 115

I : I ¼ fx1 ; x2 ; . . .; xm ; . . .g

The elements in set I are waiting to be classified, xm represents the element m.

There is a mapping rule: y ¼ f ðxi Þ

It makes any xðxi 2 IÞ has one and only one y ðyj 2 f ðxi ÞÞ corresponded to it.
Among them, f is called classifier, and the task of classification algorithm is to con-
struct classifier f.
A. Naïve Bayes
Classification ideology of Naïve Bayes is to solve the probability of occurrence of
category yj ðj ¼ 1; 2. . .; n; . . .Þ when element xi appears. After that, the probable
maximum category y will be selected as the category of element xi . The classification
process of Naïve Bayes is divided into three phases, and the specific content
arrangements are as follows:
(1) Preparation; above all, the feature attributes are determined and partitioned
according to the practical matter, next a part of the elements is selected and
classified though known data, in the last step, we take it as a training set. In this
phase, the input is all the data to be classified, and the output includes both the
feature attributes and the training set. The quality of the features and training
samples plays a decisive role in the design of classifier, therefore, this is a crucial
phase.
(2) Classifier Design; in this phase, a classifier will be generated after input feature
attributes and training dataset. The main work of this phase is not only to calculate
the probability of each category in the training set, but also to solve the conditional
probability, that is to calculate the probability of occurrence of all categories when
the feature iði ¼ 1; 2; . . .; nÞ occurs. Ultimately, the results are recorded to design a
classifier.
(3) Application; the test data is used to verify the classification models which build in
phase (2).
Naïve Bayes classification is a typical generative learning method which has the
advantages of simple logic and easy implementation. Bayes faster than other complex
methods in learning and classification, indeed, it has excellent performance in text
classification and spam filtering.
B. Support Vector Machine
Support vector machine (SVM) is a binary model by Computing support vector
which has strong generalization capability. By the way, it is widely applied into fields
such as face recognition and speech recognition. This algorithm maximizes the interval
between different types of samples in the feature space by learning a linearity classi-
fication facet, and then transforms it into a convex two order programming problem.
That means maximizing the interval of two classes of sampling points under constraint
conditions. Similar to Naïve Bayes, the classification process of SVM has been three
116 J. Li et al.

phases: Preparation, Classifier Design and Application. In the second phases, the main
task is to find a hyperplane of optimal decision for classification.

3.3 Evaluation Metrics

In this study, we use accuracy and confusion matrix to compare the performance of two
classifiers.
A. Accuracy
In general, we utilize accuracy to judge the performance of the classifier. The
definition of accuracy is that the ratio of the correct classification numbers to the total
samples in a given dataset.
B. Confusion Matrix
Confusion matrixes are usually used to compare the true class and predicted class.
In a general way, the class we will focus on is defined as the positive case (+1), and the
other classes are negative cases (−1). There are 4 situations about classifiers prediction
on the test set, as shown in Table 3.

Table 3. Prediction performance of classifier

−1(true class) +1(true class)
−1(predicted class) True negative(TN) False negative(FN)
+1(predicted class) False positive(FP) True positive(TP)

The interpretations of each symbol are as follows:

TN – The numbers that the model predicts negative sample to negative sample.
FN – The numbers that the model predicts positive sample to negative sample.
FP – The numbers that the model predicts negative sample to positive sample.
TP – The numbers that the model predicts positive sample to positive sample.
The precision, recall and F1-score are common metrics for binary classification.
(1) Precision (P):

P ¼ TP=ðTP þ FPÞ

Precision can measure the correct rate of the model in predicting positive samples.
(2) Recall (R):

R ¼ TP=ðTP þ FN Þ

Recall can measure the reliability of the model in predicting positive samples
 How to Verify Users via Web Behavior Features 117

(3) F1 -score ðF1 Þ:

2=F1 ¼ 1=P þ 1=R

F1 ¼ 2TP=ð2TP þ FP þ FN Þ

F1 is a harmonic mean which comprehensive reflects the precision and recall.

4 Experimental Result

Firstly, we collected the web usage log of AML which in the period of May 2017 to
March 2018. Secondly, according to the user’s operation frequency, the top ten user’s
data set has been retained. And then, the data are preprocessed by the principle of
feature construction in part 2 of this paper. Thirdly, since most of the machine learning
methods can only accept the input of quantitative features, we use the encoding method
which called One-Hot to convert the non-numeric features into numerical. After that, in
order to gain as much effective information as possible from the data set, we use cross

Fig. 4. Prediction results of SVM

118 J. Li et al.

validation to partition the training set and test set, and the proportion of the two is 7:3.
At the same time, this can also help us evaluate the predictive performance of the
model. After the data preparation is completed, we use Naïve Bayes and SVM algo-
rithms to implement user identity recognition based on web behavior. This process is
implemented in Python Programming.
Figure 4 below shows the prediction results of SVM for the test set, and the
prediction accuracy is 65%. According to the predicted results by SVM, there are 812
and 1212 web behaviors that belong to user 2 and user 7, of which has 801 and 1196
are correctly classified to their owns, and only a few are misjudged as other users. But
for other users, the prediction is not satisfactory. Among them, 203 web behaviors
belong to user 4 are wrongly awarded to users 7. Moreover, only 18 of the 130 web
behaviors that belong to user 8 are correctly classified. On the whole, many behaviors
that are not user 7 are wrongly classified to him.
The accuracy of Naïve Bayes is 64%, as shown in Fig. 5. The prediction result for
user 2 is the most ideal, and 800 of web behavior records which belongs to him are
correctly identified. But from the point of view of the user 1, the prediction results are
not good. There are 333 web behavior records in the test set which belongs to the user
1. Among it has 146 records are classified to user 8, and there is also having 59 and 53

Fig. 5. Prediction results of Naïve Bayes

 How to Verify Users via Web Behavior Features 119

Table 4. Model performance

SVM Naïve Bayes
Precision Recall F1-score Precision Recall F1-score
0 1 0.38 0.55 0.8 0.78 0.79
1 0.95 0.32 0.48 0.46 0.14 0.21
2 1 0.99 0.99 0.99 0.99 0.99
3 0.95 0.34 0.5 0.66 0.63 0.65
4 0.98 0.23 0.37 0.37 0.37 0.37
5 0.97 0.19 0.32 0.81 0.81 0.81
6 1 0.3 0.46 0.76 0.78 0.77
7 0.42 0.99 0.59 0.65 0.58 0.61
8 0.95 0.14 0.24 0.14 0.82 0.24
9 0.98 0.75 0.85 0.73 0.48 0.58
Avg 0.84 0.65 0.64 0.7 0.64 0.65

records are misjudged to user 4 and user 7. It is disappointing that only has 46 records
are correctly identified.
Table 4 shows the overall performance of SVM and Naïve Bayes. The value of
prediction, recall and F1 -score of each user is presented in this report, and the last row
express the model performance with regard to the whole test set. From this one report
we can see that SVM is better than Naïve Bayes in precision and recall, but for F1 -
score, Bayes is slightly higher than SVM.
As the result of analysis, SVM has higher accuracy but less reliable for positive
samples. But for Bayes, it is more balanced for the distribution of accuracy and
reliability.

5 Conclusions

We present 13 web behavior features which associated with the physiological, psy-
chological and living environment of users to verify their identity. It is proved that this
is a feasible task under the support of machine learning method, although the results
vary with the choice of the model. It is a valuable study to verify user identity with the
help of web behavior features, which is because the user’s equipment information and
network environment are very easy to fake, but the behavior of someone is hard to copy
owing to his uniqueness, so that we could guarantee the security of accounts by the user
web behaviors, without using other additional devices.
The future direction of work includes the following points: (1) We will try diverse
classification and prediction methods to improve the accuracy and reliability of user
identification. (2) On the basis of the existing research, we will further explore more
features of human web behavior. (3) The importance and contribution of every web
behavior feature under different approaches will be discussed in the next step, and the
web user is presented in a comprehensive way from the perspective of human
behavioral science.
120 J. Li et al.

Acknowledgment. This work was supported by Fundamental Research Funds for the Central
Universities NO. 106112016CDJXY110003, 2016.1-2017.12 and the National Natural Science
Foundation of China under Grant No. 71671020.

References
1. CNNIC: The thirty-ninth statistical report on the development of Internet in China (2017).
http://www.cnnic.net.cn/hlwfzyj/hlwxzbg/hlwtjbg/201701/t20170122_66437.htm
2. Géczy, P., Izumi, N., Akaho, S., et al.: Human behavior on web: what is known. In:
International Proceedings of Economics Development & Research (2011)
3. Guo, J.L., Fan, C., Guo, Z.H.: Weblog patterns and human dynamics with decreasing
interest. Eur. Phys. J. B 81(3), 341–344 (2011)
4. Neelima, G., Rodda, S.: Predicting user behavior through sessions using the web log mining.
In: International Conference on Advances in Human Machine Interaction, pp. 1–5. IEEE
(2016)
5. Zhong, J., Yan, C., Yu, W., et al.: A kind of identity authentication method based on
browsing behaviors. In: Seventh International Symposium on Computational Intelligence
and Design, pp. 279–284. IEEE Computer Society (2014)
6. Leung, R., Law, R., Masiero, L., et al.: Behavior of online visitors to hotel ICON: a
weekday-weekend analysis. e-Rev. Tour. Res. 7 (2016)
7. Cui, P., Liu, H., Aggarwal, C., et al.: Uncovering and predicting human behaviors. IEEE
Intell. Syst. 31(2), 77–88 (2016)
8. Lin, C.J., Wu, C., Chaovalitwongse, W.A.: Integrating human behavior modeling and data
mining techniques to predict human errors in numerical typing. IEEE Trans. Hum.-Mach.
Syst. 45(1), 39–50 (2017)
9. Pennacchiotti, M., Popescu, A.M.: A machine learning approach to Twitter user
classification. In: International Conference on Weblogs and Social Media, Barcelona,
Catalonia, Spain. DBLP, July 2011
10. Feher, C.: User identity verification via mouse dynamics. Inf. Sci. 201(19), 19–36 (2012)
11. Pao, H.K., Fadlil, J., Lin, H.Y., et al.: Trajectory analysis for user verification and
recognition. Knowl.-Based Syst. 34(5), 81–90 (2012)
12. Von Mises, L., Greaves, B.B.: Human Action: The Treatise on Economics. Liberty Fund,
Carmel (2007)
 Authentication Using Users’ Mouse Behavior
in Uncontrolled Surroundings

Fan Mo, Shiquan Xiong ✉ , Shuping Yi, Qian Yi, and Anchuan Zhang
( )

Department of Industrial Engineering, Chongqing University,

Chongqing, People’s Republic of China
[email protected]

Abstract. Trusted interaction mechanism is very crucial to ensure the online

security, but almost all studies of authentication using behavioral biometrics are
based on controlled experiment. In order to generalize the authentication, data
were continuously gathered from websites. The experimental system, which was
designed as websites, can be regarded as an uncontrolled experiment. Eight users
used the websites for more than three months. No specific tasks were asked to be
finished, so the users could use the websites as their will. The system gathered
users’ mouse data automatically, and based on that, mouse behavior models were
built. Only left click and movement sequence of mouse events are considered,
but error rates are lower than 3.36% in terms of left click and 4.21% in terms of
the movement sequence. The results of a case study show that the authentication
accuracy using users’ mouse behavior in uncontrolled surroundings is quite high.
This research has verified a rapid and general approach to authentic user behavior
on the network environment.

Keywords: Trusted interaction mechanism · Authentication

Uncontrolled surroundings · Network environment · Mouse behavior

1 Introduction

Human’s behavior is continuous in both real world or online. Before using the Internet,
they already develop behavioral patterns on many things from using a pen to pressing
a button. After they get online, their behaviors continue with offline behavioral patterns,
develop into online patterns, and become stable gradually. Based on that, analysis of
their online behavior is meaningful and feasible, since each people’s behavior online
probably could only match with himself/herself. Studies on trusted interaction mecha‐
nism are extremely important nowadays, because of the rapid development of the
Internet and security needs of persons and institutions online.
Authentication using behavioral biometrics is a very important issue waiting to be
tackled nowadays. People use access control mechanisms, like username, password,
token, fingerprints, or signature, to protect against another person’s unauthorized access.
This means that a user needs to give his/her identity’s proof when logging in a web.
Another drawback of user verification methods is that they require dedicated hardware
devices such as fingerprint sensors are expensive and are not always available. Although

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 121–132, 2018.
https://doi.org/10.1007/978-981-13-2384-3_12
122 F. Mo et al.

fingerprint verification is becoming widespread in laptops, it is still not popular enough

and it cannot be used in web applications.
Authentication in the web environment is also a critical issue because of the devel‐
opment of the Internet. Existing authentication usually focuses on a specific device, such
as using the identical computer by one participant, which is not suitable for network
environment. Today’s people tend to use diverse devices, and the authentication of their
behavioral biometrics should be more tolerant and should be established based on
complex surroundings, may be based on front-end web development.
There are numerous studies focus on behavioral authentication in experimental
surroundings, but a real situation of invading usually is in complex conditions and
uncontrolled environments. These previous research explored controlled authentication
from different aspects, including keyboard, mouse, etc. Using experiments to analysis
users’ behaviors is able to get a higher performance in terms of prediction results, since
the sample size is similar even identical, and the equipment that participants used is the
same. In uncontrolled surroundings, users are allowed to use different devices (tablets,
smartphones, personal computers, etc.). Also, the data sizes of different users are not
similar. It means that the authentication need to be adopted in real conditions. This can
develop prior studies on behavioral authentication, and make them more practical.
Therefore, in this research, authentication using behavioral biometrics based on
front-end web development which collects data from the uncontrolled experimental
environment is explored.

2 Literature Review

Behavioral biometrics is the process of measuring behavioral tendencies of a user that

results from both physiological and psychological differences from person to person. It
does not require special designated hardware since they use common devices such as
the keyboard and mouse [1]. The common behavioral biometric authentication methods
include keystroke dynamics [2, 3], mouse dynamics [4–6], and Graphical User Interface
(GUI) usage analysis [7]. A major benefit to usage-based biometrics is that authentica‐
tion can occur actively throughout the session of a user as opposed to once during initial
login [8, 9]. After the initial login has occurred, this can prevent session of a user from
being hijacked. This is extraordinarily important for online authentication such as shop‐
ping on the internet.
Gaines, Lisowski, Press, and Shapiro [10] introduced the idea of using behavioral
biometrics as a supplement to traditional authentication. Keystroke timing data was
initially used to supplement password entry [11, 12] this evolved into being able to
analyze long structured text as a basis for authentication [13–15], and finally long free
text samples [3, 16]. Research on mouse dynamics that are based biometric authentica‐
tion [17] that is started in 2003 when Everitt and McOwan [18] showed a promising
result of static authentication, but most recently Shen, Cai, Guan, Du, and Maxion [19]
showed mouse dynamics’ potential for static authentication as an alternate access control
solution. Ahmed and Traore [4, 20] integrated mouse dynamics and keyboard into a
single architecture that could act as an intrusion detection system. Twenty-two subjects
 Authentication Using Users’ Mouse Behavior 123

were asked to install a monitoring system and ran the software for nine weeks. These
fusion methods were able to achieve high accuracy [21]. However, the time for collecting
data was too long, and their methods cannot be used on websites.
Overall, authentication for websites generally needs more varied than that on a
traditional computer. Studies that used behavioral biometrics designed experiment to
collect data from users who used the same computer, and their variable control is strict.
This means that the generalization of these authentications models is not so easily into
the daily usage of web development. First, one person may have several devices to access
the websites. The mouse or computer used by a person may be different. Second, on the
condition of the Internet, invaders are highly possible to use different computers and
measures to remote control users’ data. To deal with this problem, this research considers
the complex environment of using computers and tries to build models to verify mouse
behavior of individuals.

3 Uncontrolled Experimental System Design

3.1 Participants

There were more than twenty participants in this experiment. The participants were
recruited from the Department of Industry Engineering, Chongqing University. They
were asked to use online websites to search information, to read texts and pictures, and
to download files if they want. There were no specific tasks they are asked to finish, and
there is no target amount of tasks, so the participants can choose to use the web a lot,
only use it a little, or even not use it. In that conditions, only data from eight participants
(two are females and six are males) were chosen to be analyzed in next analysis because
of their data were from the period more than 3 months. The average age of participants
was 31.25 (SD = 12.848). Six of them were postgraduate students, and two were profes‐
sors. All the participants had more than 10 years’ experience of using internet and mouse.
Participants used mouse normally because there is no target in this experiment. The data
were collected when they used websites to browse webs, so it was a normal mouse
surfing.

3.2 Procedure

A behavioral biometric system fusing user data from mouse, which is based on front-
end web development, is presented in this research. There are mainly four steps (the last
three steps are as shown Fig. 1.
124 F. Mo et al.

Fig. 1. The process of features acquisition, modeling, and verification.

First, users log into the websites, and use it of his/her own free will; meanwhile, the
system records users mouse data. An example JavaScript code can be seen as follows.
The second phase is features acquisition. The gathered data are processed by R, and the
specific actions (e.g. click, or move) are extracted. Then, features of the action are
calculated. These features are saved as feature data sets. The third phase is modeling
and the fourth phase is verification. More details about the last three parts are described
in Sect. 4.

3.3 Equipment and Programming

A data collection system was developed. Javascript and Java EE in front-end web are
used to develop the system which is official websites of the research group (http://
www.cquieaml.com/) as shown in Fig. 2. The websites include several parts, such as
academic research, business exchanges and cooperation, blog article, about us, and the
management system. Javascript codes recorded users mouse data, like click and move,
and then transformed the data into back-end server and saved them. The data that are
used in following analyses can be downloaded in the management system.

Fig. 2. The system framework of the uncontrolled experimental system.

 Authentication Using Users’ Mouse Behavior 125

Two personal computers (Dell Inspiron 3268-R1308) are used for controlling the
management system. Participants access the online websites using their own computers,
including MacBook Air (2015 early), Lenovo (ThinkPad X1 CAR), Acer (4750G), and
Acer (TMP238-M-P6J9). The sampling rate of the mouse data is not consistent, from
10 Hz to 200 Hz. By resampling, the rate is fixed at 100 Hz, so the interval time between
two data records (e.g. two movement segments) is 0.01 s.
Eight users’ behavior of this system is recorded for more than 3 months, from April
2017 to July 2017. All of them are members of the research group and have their own
accounts on the websites. Although the users know the websites can record their
behavior, they do not pay attention to it since they generally do some basic work, such
as downloading papers or posting a message on the forum, and the recording of their
behavior does influence their feelings, so it can be regarded as uncontrolled surround‐
ings.

4 Data Processing

4.1 Features Acquisition

Data on mouse behavior are cleaned, classified, and featured through R programming.
User characteristics that relate to the way that a particular user interacts with the websites
are collected by the system. This is done by monitoring mouse movements. Features are
calculated on these actions. After data collection, the raw data is processed to create
features for active authentication testing and identification. The calculated features are
selected from prior works. These features are used in many related studies associated
with authentication, since they have relatively high reliable.
The original data only involved five values, including type of the mouse event, x-
coordinate of the mouse pointer (x), y-coordinate of the mouse pointer (y), time of the
mouse event (t), and users ID. R codes are developed to classify the type of mouse event
and extract the features. The main mouse events in webs are left-click (LC) and the
movement sequence (MS). Unlike the mouse behaviors when interacting with files
systems or games, right-click (LC) is not commonly used on websites. The right click
motion on pages is to open the webpage ‘tab’, which helps users to have advanced
functions with the websites. These functions are usually not used when surfing the web,
so the right click is not going to be analyzing in follow analysis.
Therefore, the research focused on the two behavior and extracted features of them.
Features used in this research are from existing studies. The data format changes as
shown in Fig. 3.
126 F. Mo et al.

Fig. 3. Changed data format.

The basic events of mouse behavior include pressing the mouse button (including
left button and right button), lifting the mouse button (also including left button and right
button), and the movement of the mouse which is recognized by a photoelectric detector.
( )
⎧ EventPR = PRi , xi , yi , ti , IDuser
⎪ ( )
⎪ EventLL = LLi , xi , yi , ti , IDuser
⎪ ( )
⎨ EventLR = LRi , xi , yi , ti , IDuser (1)
⎪ ( )
⎪ EventM = (Mi , xi , yi , ti , IDuser
⎪ Event = MS , x , y , t , ID )
⎩ MS i 1:n 1:n 1:n user

Left click usually is composed of three sequential basic events: pressing the mouse
button (PL), small movement (M), and lifting the mouse button (LL). Users may not
keep still when pressing the button, which is the reason why there is a small movement.
Features of left click involved interval time between pressing and lifting the mouse
button and the slide distance between them.
{( ) ( ) ( ) }
EventLC = LCi , [(PLi , xi , yi , ti , MSi , x1:n , y1:n , t1:n , LLi , xi , yi , ti ], IDuser (2)

Interval time in LC equals that ti of LLi minus ti of PLi, and that is, interval time in
LC = tLLi − tPLi. Similarly, slide distance in LC is:
√
( )2 ( )2
xLL − xPL + yLL − yPL
i i i i
(3)

Features of movement sequence involved 49 aspects in this research. Nine features

are shown in Table 1.
 Authentication Using Users’ Mouse Behavior 127

Table 1. Nine features used to evaluate a motion curve.

Variables Definition Details
( )
Angle of movement The angle of the path tangent with θi = arctan δy1 ∕δx1 + ∑ij=1 δθj
the x-axis ( )
δθ = δartan δy ∕δx is the
j 1 1
minimal angle change given in
[−π, π] between the j and j + 1
points
Curvature The relative angle change to the c = δθ∕δs
traveled distance
Curvature change rate The rate of the curvature change Δc = δc∕δs
Horizontal velocity Velocity with respect to the x-axis vx = δx∕δt
Vertical velocity Velocity with respect to the y-axis vy = δy∕δt
√
Velocity The first displacement moment v = v2x + v2y
Acceleration The second displacement moment v̇ = δv∕δt
Jerk The third displacement moment v̈ = δv∕δt
̇
Angular velocity To the angle change rate w = δθt ∕δt

Each movement has values of all above features, so the minimum, maximum, range,
mean, and standard deviation of them are used to reflect the features of the movement
sequence. In this part, 45 features are used to measure the movement sequences.
There are also four features that are able to use to reflect the movement, including
duration
√( of movement tn, traveled distance Sn, jitter S′ ∕Sn, straight distance
) 2 ( )2
S′ = x1 − xn + y1 − yn . Therefore, there are 49 features in total.

4.2 Learning

R codes are used to build the machine learning methods. Classifiers are constructed for
both left-click (LC) and movement sequence (MS). Specifically, the two actions instance
forms a row whose columns contain the features that are associated with the action, and
its label is given by the ID of the user who performed the action, so it is the supervised
learning. Support Vector Machine, Random Forest, and K-Mean Clustering are tested,
and then the method of random forest is used to establish models, because of its relatively
high accuracy and lesser time consumption. This research not aims to get the best results
of machine learning methods; it just shows a good learning results can be gained in
uncontrolled surroundings. The mouse data is organized using the techniques of classi‐
fication and regression trees. An algorithm in the random forest runs, and then randomly
selects the predictor and repeats the process. Hundreds of regression trees are generated.
The random selection of the predictive factor is averaged as a result of the classification.
The data from feature data sets are split into two parts: training sets (60%) and testing
sets (40%), using random sampling methods. This method picks sample randomly and
put it into a set. Each training sample is used to construct a decision tree. ‘randomForest’
library of R is used. The number of trees to grow is 500. The importance of predictors
is set as needed to be assessed, and the proximity measure among the rows is set as
128 F. Mo et al.

needed to be calculated. Other parameter settings in the randomForest function is default.

The models are firstly built on the data of training sets by classifier; features of users’
mouse behavior are categorized by the random forest. The results of classification are
the mouse behavior models of the user, so the results can reflect this specific user. After
training, testing sets are used to test the accuracy of prediction. Prediction results are
visualized though confusion matrix. The model costs about 20 s to get training models
and costs around 1 min to get prediction results. ‘corrplot’ library of R is used to visualize
the prediction result.

4.3 Verification
In the end, authentication is conducted based on the two behavior sub-models of LC and
MS. The authors comprise an inducer used to build verification model for the user by
training on past behavior, often given by samples. During verification, the induced model
is used to classify new samples that are acquired from the user. There are two main kinds
of models which are models of click and models of movement. Invaders’ mouse behavior
is recorded by using the account of members in the group, the reliability of his/her click
behavior and movement behavior are calculated respectively.
To give the final reliability of users’ mouse behavior, the two reliabilities are mixing.
To reasonably and simply mix the two authentication results from the models, below
equation is used.

α ≥ A&𝛽 ≥ B (4)

Where α is the reliability of models of left click, and β is the reliability of models of
the movement sequence. A is the toleration of reliability of left click, and B is the toler‐
ation of reliability of movement sequence. When both reliabilities from the two models
reach the threshold, the individual’s behavior can be considered as credible.

5 Results and Discussions

5.1 Models Without Sample Balance

Data of mouse behavior are gathered from March 29th, 2017 to August 25th, 2017, and
there are more than 1.6 million records of mouse events from eight people. Users’ click
behavior and movement behavior are modeling using Random Forest based on the
training data sets, respectively. There are 500 trees for both of the two models. After
that, testing data sets are used to verify the accuracy of the models. As shown in Fig. 4,
the accuracy is not very high. In terms of the event of left click, the only accuracy of
three participants’ behavior is up to 50%, including MF (72%), YS (55%), and ZHJ
(79%). The results are not good even worse than random guessing. Several predictions
are very bad. For instance, YJ’s left clicks are all classified into XHQ (55%) and ZAC
(45%). None of LJJ left clicks are grouped to other. With regards to the event of move‐
ment sequence, there are five participants who have accuracy up to 50%, including LJJ
(66%), MF (64%), XSQ (64%), ZAC (52%), and ZSJ (51%). These results are better than
 Authentication Using Users’ Mouse Behavior 129

the results of left click but still very poor. The reasons for it is the unbalanced sample
size of different participants, because the data are collected from uncontrolled experi‐
mental surroundings.

Fig. 4. Left figure: Pie figure of classifying events of LC. Pie figure of classifying events of MS.

5.2 Models with Sample Balance

In order to improve the prediction accuracy, the sample size is balanced by multiplying
the size of a small sample group, and random sampling the large sample group. Then,
the models are rebuilt using these new data sets. As shown in Tables 2 and 3, the accuracy
significantly increases. The last column is the error rates of grouping. As to the left click,
the error rates of MF, SHY, and XSQ are 0%, which means that all the predictions are
absolutely right. The error rates of LJJ are the highest, up to 15.58%. The overall error
rate is 3.36%.

Table 2. The accuracy of left click using testing sets.

LJJ MF SHY XSQ YS YSP ZAC ZSJ Error rates
LJJ 195 0 0 0 34 2 0 0 15.58%
MF 0 216 0 0 0 0 0 0 0.00%
SHY 0 0 550 0 0 0 0 0 0.00%
XSQ 0 0 0 498 0 0 0 0 0.00%
YS 1 0 0 0 638 0 1 0 0.31%
YSP 0 0 0 0 12 168 0 0 6.67%
ZAC 2 0 0 0 26 2 320 0 8.57%
ZSJ 1 3 0 0 2 0 9 149 9.15%

As to the accuracy of the movement sequence, the overall error rate is 4.21%, which
is also low. All the predictions of participants’ movement sequence behavior are rela‐
tively high, except for the predictions of ZAC. The error rate of predicting ZAC’s
movement sequence is 62.96%. It should be noticed that ZAC is the most active partic‐
ipants in the data collection system, so the data of ZAC is maximum; nevertheless, in
the results of the accuracy of movement sequence, the sample size of ZAC is the
minimum. That is why the result is not reasonable.
130 F. Mo et al.

Table 3. The accuracy of movement sequence using testing sets.

LJJ MF SHY XSQ YS YSP ZAC ZSJ Error rates
LJJ 569 0 0 0 9 0 1 1 1.90%
MF 0 204 4 2 0 0 0 0 2.86%
SHY 0 2 584 2 0 0 0 1 0.85%
XSQ 0 1 1 602 0 0 0 0 0.33%
YS 8 1 0 0 540 0 5 1 2.70%
YSP 0 0 0 0 5 167 1 0 3.47%
ZAC 15 0 2 0 63 1 60 21 62.96%
ZSJ 0 0 0 0 0 0 0 619 0.00%

Data on mouse behavior are cleaned, classified, and featured through R program‐
ming. User characteristics that relate to the way that a particular user interacts with the
websites are collected by the system. This is done by monitoring mouse movements.
Features are calculated on these actions. After data collection, the raw data is processed
to create features for active authentication testing and identification. The original data
only involved five values, including the type of the mouse event, x-coordinate of the
mouse pointer (x), y-coordinate of the mouse pointer (y), time of the mouse event (t),
and users ID. R codes are developed to classify the type of mouse event and extract the
features. The main mouse events in webs are left click (LC) and the movement sequence
(MS), so the researchers focused on the two behavior and extracted features of them.
Features used in this research are from existing studies. The data format changes as
shown in Fig. 3. Interviews of one participant showed that he used both mouse and the
MacBook touchpad. As we can see, behavior of the two interaction ways must be very
different, but results indicated that all mouse behavior from one person can be learned,
and the classifiers can be strong. That is, although participants used several different
devices to log in the system, it seems that the models could handle this problem. There‐
fore, in some extent, this finding may be extended to other studies, so researchers do not
need to control the experiment strictly, which can reduce the cost and also make models
fit with real conditions better.

5.3 A Case of Unauthorized Behavior from an Invader

However, the results are high enough, so a test of unauthorized behavior from an invader
is conducted. There is an invader whose code is ZSJ. ZSJ used MF’s account to login
to the websites and tried to simulate MF’s mouse behavior. After about ten minutes’
use, 35 left click events and 42 movement events are recorded. No matter left click events
and movement events, none of ZSJ’s mouse behavior is recognized as MF’s, and the
accuracy of authentication is 100%.
Prediction results in Sect. 5.2 are hard to reach a really high accuracy (i.e.
100%), but when the system knows the behavior from the identical account, the
accuracy of authentication could be much higher. This situation conforms to the real
environment that an invader knows others’ account by stealing or password breaker.
The uncontrolled mouse behavior models provide the second protection to prevent
 Authentication Using Users’ Mouse Behavior 131

invader’s unauthorized behavior. However, the system needs to collect enough data
in seconds in order to extract them into mouse action, so it may cost several minutes
to ensure the accuracy of recognition.

6 Conclusion

This research uses data from websites which can be considered as an uncontrolled
experiment. Data are continuously gathered from websites, which means the data are
unbalanced and quite conform to the real situation. The data for the uncontrolled experi‐
ment is further used to build mouse behavior models through the machine learning
method (i.e., random forest). Only left click and movement sequence of mouse events
are considered, but error rates are still very high. In terms of left click, error rates lower
than 3.36%, and in terms of movement sequence, the error rates lower than 4.21%. In
the end, a case study of unauthorized behavior from an invader is conducted. The results
institute that the authentication using users’ mouse behavior is quite high. This research
provides a rapid and general approach to authentic user behavior on the network, which
is meaningful for the development of authentication using behavioral biometrics in the
uncontrolled environment.
Further work should focus on conducting this experiment on larger sample size. Real
Environment usually involves a large group of users on websites, and the identification
of potential invaders need to base on all these users’ behavior. Besides, mass data might
lead to more processing time, so how to improve the classifiers is vital. In addition to
that, the influence of users’ psychological characteristics, as well as their age and expe‐
rience of using smart devices, on their mouse behavior may deserve further research.
This could explore more about how online behavior related to the specific users who
take the behavior.

Acknowledgment. This work was supported by Fundamental Research Funds for the Central
Universities NO. 106112016CDJXY110003, 2016.1-2017.12 and the National Natural Science
Foundation of China under Grant No. 71671020.

References

1. Multi factor user authentication on multiple devices (2017). https://patents.google.com/

patent/US20180097806A1/en. Accessed 26 Apr 2018
2. Joyce, R., Gupta, G.: Identity authentication based on keystroke latencies. Commun. ACM
33(2), 168–176 (1990)
3. Monrose, F., Rubin, A.D.: Keystroke dynamics as a biometric for authentication. Future
Gener. Comput Syst. 16(4), 351–359 (2000)
4. Ahmed, A.A.E., Traore, I.: A new biometric technology based on mouse dynamics. IEEE
Trans. Dependable Secur. Comput. 4(3), 165–179 (2007)
5. Muthumari, G., Shenbagaraj, R., Pepsi, M.B.B.: Mouse gesture based authentication using
machine learning algorithm. In: 2014 IEEE International Conference on Advanced
Communications, Control and Computing Technologies, pp. 492–496 (2014)
132 F. Mo et al.

6. Zheng, N., Paloski, A., Wang, H.: An efficient user verification system via mouse movements.
In: Proceedings of the 18th ACM Conference on Computer and Communications Security,
CCS 2011, pp. 139–150. ACM, New York (2011). http://doi.acm.org/
10.1145/2046707.2046725. Accessed 25 Aug 2017
7. Gamboa, H., Fred, A.: A behavioral biometric system based on human-computer interaction,
pp. 381–392 (2004). http://adsabs.harvard.edu/abs/2004SPIE.5404..381G. Accessed 25 Aug
2017
8. Mondal, S., Bours, P.: Combining keystroke and mouse dynamics for continuous user
authentication and identification. In: 2016 IEEE International Conference on Identity,
Security and Behavior Analysis, ISBA, pp. 1–8 (2016)
9. Park, J., Han, S.H., Kim, H.K., Moon, H., Park, J.: Developing and verifying a questionnaire
for evaluating user value of a mobile device. Hum. Factors Ergon. Manuf. Serv. Ind. 25(6),
724–739 (2015)
10. Gaines, R.S., Lisowski, W., Press, S.J., Shapiro, N.: Authentication by keystroke timing: some
preliminary results. Report No. RAND-R-2526-NSF. Rand Corp Santa Monica CA, May
1980. http://www.dtic.mil/docs/citations/ADA484022. Accessed 25 Aug 2017
11. Brown, M., Rogers, S.J.: User identification via keystroke characteristics of typed names
using neural networks. Int. J. Man-Mach. Stud. 39(6), 999–1014 (1993)
12. Haider, S., Abbas, A., Zaidi, A.K.: A multi-technique approach for user identification through
keystroke dynamics. In: 2000 IEEE International Conference on Systems, Man, and
Cybernetics, vol. 2, pp. 1336–1341 (2000)
13. Han, S.J., Kim, S.-U.: Placement of a touchpad and click-buttons to relieve arm fatigue and
discomfort in a laptop PC. Hum. Factors Ergon. Manuf. Serv. Ind. 27(3), 131–137 (2017)
14. Monrose, F., Rubin, A.: Authentication via keystroke dynamics. In: Proceedings of the 4th
ACM Conference on Computer and Communications Security, CCS 1997, pp. 48–56. ACM,
New York (1997). http://doi.acm.org/10.1145/266420.266434. Accessed 25 Aug 2017
15. Pusara, M.: An examination of user behavior for user re-authentication (2007). https://
search.proquest.com/openview/1bd296a5f59df64c30768646f674eeab/1?pq-origsite=gsch
olar&cbl=18750&diss=y. Accessed 25 Aug 2017
16. Bergadano, F., Gunetti, D., Picardi, C.: User authentication through keystroke dynamics.
ACM Trans. Inf. Syst. Secur. 5(4), 367–397 (2002)
17. Pusara, M., Brodley, C.E.: User re-authentication via mouse movements. In: Proceedings of
the 2004 ACM Workshop on Visualization and Data Mining for Computer Security, VizSEC/
DMSEC 2004, pp. 1–8. ACM, New York (2004). http://doi.acm.org/
10.1145/1029208.1029210. Accessed 26 Apr 2018
18. Everitt, R.A.J., McOwan, P.W.: Java-based internet biometric authentication system. IEEE
Trans. Pattern Anal. Mach. Intell. 25(9), 1166–1172 (2003)
19. Shen, C., Cai, Z., Guan, X., Du, Y., Maxion, R.A.: User authentication through mouse
dynamics. IEEE Trans. Inf. Foren. Secur. 8(1), 16–30 (2013)
20. Ahmed, A.A.E., Traore, I.: Anomaly intrusion detection based on biometrics. In: Proceedings
from the Sixth Annual IEEE SMC Information Assurance Workshop, pp. 452–453 (2005)
21. Bailey, K.O., Okolica, J.S., Peterson, G.L.: User identification and authentication using multi-
modal behavioral biometrics. Comput. Secur. 43, 77–89 (2014)
 Construction of Network User Behavior
Spectrum in Big Data Environment

Mengyao Xu, Fangfei Yan, Biao Wang, Shuping Yi(&), Qian Yi,
and Shiquan Xiong

College of Mechanical Engineering, Chongqing University, Chongqing 400044,

China
[email protected]

Abstract. Studying the behavior patterns of network users is important for

understanding the individual needs and identifying the identity of users. In this
paper, the behavior patterns of network users are built by constructing the
behavior spectrum of network users. Network users’ behavior spectrum is
constructed by dividing the behaviors of users into perceptual state and physi-
ological state. The perceptual state is divided into other features according to the
actual situation. The data of the enterprise is used to establish a method of user
behavior spectrum based on perceptual state. The physiological state is repre-
sented by the features of mouse behavior. The data of self-built website is used
to explore a method of user behavior spectrum based on the physiological state.
Finally, an example is used to establish a user’s behavior spectrum based on two
methods.

Keywords: Network user behavior spectrum
Perceptual state

Physiological state
Big data

1 Introduction

With the development of Internet technology, people’s lives have become more con-
venient. At the same time, information security issues have also received more and
more attention. Identification of user identities can be achieved by studying the
behavior patterns of network users, thus a new method can be offered for the estab-
lishment of information security protection mechanism. At present, many scholars have
explored how to reflect the behavior patterns of network users. Frhan [1] developed a
visualization model named as Social Pattern Clustering WebClickviz
(SPCWebClickviz) to visualize the social networking data based on user activities and
then clusters them into specified groups. Qiang and Ji-min [2] analyzed Web logs to
build user behavior patterns. Zhu et al. [3] proposed a method for defining user roles
based on behavior patterns and clusters the behavior of telecommunication network
users during socialization. Katerina and Nicolaos [4] investigated the potential corre-
lation between mouse behavioral patterns or keystroke dynamics and a set of End-User
Development (EUD) behavioral attributes. These studies are currently only from a
single point of view to establish the user behavior patterns, either only use the front-end
data (mouse or keyboard related data), or only use the back-end data (log data),

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 133–143, 2018.
https://doi.org/10.1007/978-981-13-2384-3_13
134 M. Xu et al.

research is too one-sided, therefore, the user’s behavior patterns should be described in
a more comprehensive and three-dimensional manner by establishing a network user
behavior spectrum.
The concept of behavior spectrum was first used by Makkink’s in 1936 in the study
of European pied Avocet. The behavior spectrum is widely seen in animal behavior.
Nishida et al. [5] used behavior spectrum to list and describe all behavior patterns
recorded by chimpanzees in the Mahale Mountains National Park, Hall and Heleski [6]
demonstrated that the horse’s behavior spectrum plays an important role in assessing
training methods; Ivanov and Krupina [7] studied Contagion behavior cause changes in
the behavior spectrum in Rats. There are few literatures that human behavior is studied
by using behavioral spectrum, only one paper can be found so far. The behavior
spectrum was defined by Qin et al. [8] as an effective user network access behavior
monitoring and measurement method. However, the data that reflect the physiological
behavior is not collected, that is, the data collected in the article does not fully represent
the behavior of the user.
All the behaviors of the network user are recorded by using the behavior spectrum
of the network user in this paper, and then the behavior pattern of the network user are
constructed. Perceived behavioral features and physiological behavioral features can be
used to construct a network user’s behavioral spectrum. Perception [9] is the organi-
zation, identification, and interpretation of sensory information in order to represent and
understand the presented information, or the environment. In this paper, perception
refers to the direct reaction of an objective thing through the sensory brain in the human
brain. After the brain perceives the object, user will be guided to make an action. For
example, when the operating system’s own language is unfamiliar to the user, the
language will immediately be transformed into his familiar language. Therefore, the
user’s perceived behavior is defined that the influence of the external environment,
personality traits, and the habits formed during the growth process on the user’s net-
work behavior. Physiology [10] is the scientific study of normal mechanisms, and their
interactions, which works within a living system. In this paper, physiology refers to the
life activities of living organisms and the functions of various organs. Therefore, the
user’s physiology is different and the user’s behavior is also different. For example:
The elderly move the mouse generally slower than young people. Therefore, the user’s
actions using the input tools (mouse, keyboard), including movement direction and
movement speed, movement trajectory, movement time, tapping rate, tapping time are
defined as the user’s physiological behavior.
The remainder of this paper is organized as follows. The data collection is
described in Sect. 2. The method for constructing a network user behavior spectrum is
described in Sect. 3 and an example is given in Sect. 4. Finally, the concluding remarks
are addressed in Sect. 5.

2 Data Collection

Data collection is an essential step before data analysis. In this paper, the data is
analyzed from the third person point of view to get the user’s behavior features and a
one-to-one mapping with the actual behavior is achieved, so different types of data are
 Construction of Network User Behavior Spectrum 135

needed to collect. At present, data is obtained through the self-built website and the
enterprise management system, and data from the system’s background log and using
embedding code technology is collected.
The enterprise management system data is obtained from the background log of the
cooperative enterprise. The operations of all employees in the system are recorded,
these employees belong to different departments and have different responsibilities, but
each user has a clear operation task. As shown in Fig. 1, the collected features are
divided into four basic categories (time attributes, operational attributes, individual
basic attributes and company related attributes) and used to analyze user behaviors. As
shown in Fig. 2, the self-built website data is obtained from the background log and
embedding code technology of the academic exchange website established by the
laboratory.

Fig. 1. Enterprise data

3 Method of Constructing Network User Behavior Spectrum

From the Sect. 2, the user behavior features are divided into perceptual state and
physiological state. The network user behavior spectrum is constructed by analyzing
the network behavior of the perceptual state and physiological state. Next, the enter-
prise data is used to establish a network behavior analysis method based on the per-
ception state and the mouse data of a self-built website is used to establish a network
behavior analysis method based on physiological state.

3.1 Network Behavior Analysis Method Based on the Perception State

As mentioned above, a company’s ten-year log records are collected by us. In this part,
one-month log records are used, including 6738 log records.
136 M. Xu et al.

Fig. 2. The composition diagram of the network user behavior spectrum

3.1.1 Statistical Analysis of Features

Since the purpose of this paper is to study the individual’s network behavior spectrum,
five more relevant features are selected: operatorname, logintime, operatorid, enter-
fuction and enterbutton. First, this month’s record is filtered into two parts according to
the operatorname, and these two parts represent the log records of two users (user A
and user B). Some information can be obtained for statistical analysis from two users.
From Fig. 3a and b and time, it can be seen that user A always uses the system from
9:30 am to 6:00 pm, and occasionally works overtime to around 9:00 pm. In addition,
the frequently used buttons are “save” and “modify”, and the frequently used functions
are “making lists” and “sales invoices”. Therefore, it can be roughly judged that user
A’s position is related to finances.

(a) (b)
enterbutton enterfuction
ennull
Make a list
save
cancellation
delete
modify Sales Invoice
increase
invalid

Fig. 3. (a) “enterbutton” for user A. (b) “enterfuction” for user A

 Construction of Network User Behavior Spectrum 137

3.1.2 Users Behavior Spectrum Based on Perceptual State

The correlation between the various features is tried to be found by us, and a user’s
behavior spectrum based on perceptual state is expressed as follows:

operatename ¼ A  logintime ¼ pm  operateid ¼ 192:168:11:43  enterfuction ¼

order management  enterbutton ¼ save

In this expression, the processing of the logintime is expressed as 8:00–12:00 as

am, 12:01–18:00 as pm, 18:01–00:00 as night, and 00:01–7:59 as representing for early
morning. The correlation strength is expressed by calculating the degree of confidence,
the confidence calculation formula: confidence(R) = NR/Noperatename = A, NR
indicates the number of records that satisfy the behavior spectrum, and Noperatename
= A indicates the number of records for user A. The behavior spectrum result for user A
is as follows (Table 1).

Table 1. Behavior spectrum result for user A

Confidence NR Examples Instructions
64.18% 1557 A*pm*192.168.2.55 User A often logs in to the system
using ID = 192.168.2.55 in the
afternoon
48.27% 1171 A*pm*192.168.2.55*make User A often logs in to the system
a list with ID = 192.168.2.55 in the
afternoon. The often used
enterfuction is “make a list”
22.67% 550 A*pm*192.168.2.55*make User A often logs in to the system
a list*modify with ID = 192.168.2.55 in the
afternoon. The often used
enterfuction is “make a list”, the
often used enterbutton is
“modify”

3.2 Network Behavior Analysis Method Based on the Physiological State

Physiological behavior refers to the moving, clicking and tapping of input tools
(mouse, keyboard) used by the user in the human-computer interaction environment,
including the input tool’s moving direction, moving distance, moving speed, click rate,
tap rate, etc. The mouse movement is used as an example to explain the analysis
method of physiological behavior in this section.

3.2.1 The Features Value of Mouse Moving

Five eigenvalues are calculated to represent the physiological and temporal behavior
characteristics in time and space. The eigenvalues are described as follows: ① Sx
(k) = x(k + 1)−x(k), k = 1, 2,.., n, The moving distance on the X axis during the Kth
mouse moving; ② Sy(k) = y(k + 1) − y(k), k = 1, 2, .., n, The moving distance on the
Y axis during the Kth mouse moving; ③ T(k) = t(k + 1) − t(k), k = 1, 2.., n, The time
required for the Kth mouse moving; ④ Vx(k) = Sx(k)/T(k), k = 1, 2.., n, The moving
138 M. Xu et al.

rate per unit time on the X-axis during the Kth mouse moving; ⑤ Vx(k) = Sy(k)/T(k),
k = 1, 2.., n, The moving rate per unit time on the X-axis during the Kth mouse
moving.
From the eigenvalues, not only the user’s range of moving distance and moving
rate can be gotten, but also the standard deviation of the sample eigenvalues can be
calculated, from which you can see the difference in user behavior.

3.2.2 Mouse Moving Analysis

200 data are randomly selected from 2 users (33rd and 40th) during the period of
2017.7.17–2018.1.3 from both spatial and temporal dimensions as samples, and the
first two eigenvalues are selected on the spatial dimension, the second eigenvalues are
selected on the temporal dimension.
(1) Analyze the user’s mouse moving from the spatial dimension
The moving distance refers to the displacement distance of the mouse in a specific
direction each time the mouse is moved. Everyone has different habits of operating the
mouse, and the length of the mouse’s moving distance is also different.
From Fig. 4a and b, it can be seen that there is a significant difference in the moving
distance of the mouse in different directions between the NO. 33 and the NO. 40. The
horizontal moving distance range [−1040, 1040] of NO. 40 is wider than [−18, 16] of
NO. 33, the moving distance standard deviation 202.5287 is greater than 5.4268 of NO.
33, and the vertical moving distance range [−245, 236] is wider than [12,100] of
No. 33, the standard deviation of the moving distance 55.2566 is greater than 8.1125 of
No. 33.
(2) Analyze user mouse moving from the temporal dimension
The mouse’s moving rate refers to the speed of the mouse’s moving in a specific
direction within a unit time. Each person’s physical function is different, so each
person’s moving rate is also different.
From Fig. 5a and b, it can be seen that there is a significant difference in the moving
rate of the mouse in different directions between the NO. 33 and the NO. 40. The
horizontal moving rate range [−0.0622, 7.5] of the NO. 40 is narrower than [−3, 10] of
the NO. 33, the moving distance standard deviation 1.4975 is greater than the 0.9027 of
the NO. 33, and the vertical moving rate range [−4.64, 2.5] of the NO. 40 is wider than
[−1.125, 1.5] of the NO. 33, the moving distance standard deviation 1.1518 is greater
than 0.4136 of NO. 33.

4 Case Study

Based on the analysis method of the Sect. 3 for network user’s perceptual state
behavior and physiological state behavior, an example is used to illustrate the user’s
network behavior spectrum building process. 150 data from No. 46 are used for
experimental analysis.
In terms of perceptual behavior, according to the four attributes, five features are
selected: user name, date, IP, head and browser; In terms of physiological behavior,
 Construction of Network User Behavior Spectrum 139

(a) 40 33
Horizontal moving distance
1050
700
350
0

145
103
109
115
121
127
133
139

151
157
163
169
175
181
187
193
199
1
7

31
13
19
25

37
43
49
55
61
67
73
79
85
91
97
-350
-700
-1050

(b) 40 33
Vertical moving distance

150

-50 103
1
7

109
115
121
127
133
139
145
151
157
163
169
175
181
187
193
199
49
13
19
25
31
37
43

55
61
67
73
79
85
91
97

-250

Fig. 4. (a) The moving distance of the mouse in the horizontal direction of the number 33 and
40. (b) The moving distance of the mouse in the vertical direction of the number 33 and 40

(a) 40 33
Horizontal moving rate
10

0
113
106

120
127
134
141
148
155
162
169
176
183
190
197
15
22
29
36
43
50
57
64
71
78
85
92
99
1
8

-5

(b) 40 33
3
Vertical moving rate
2
1
0
127
106
113
120

134
141
148
155
162
169
176
183
190
197
15
22
29
36
43
50
57
64
71
78
85
92
99
1
8

-1
-2
-3
-4
-5

Fig. 5. (a) The moving rate of the mouse in the horizontal direction of the number 33 and 40.
(b) The moving rate of the mouse in the vertical direction of the number 33 and 40
140 M. Xu et al.

according to the user’s mouse usage, four features are selected: coordinates of the
mouse on the X axis, coordinates of the mouse on the Y axis, time, user name.
Statistical analysis and correlation analysis are performed on the features. The results
are shown in Figs. 6, 7, 8 and 9 and Table 2.

head Control
center

Homepage

Account
management

Fig. 6. The analysis of “head” for user 46

IP
106.91.50.77
14.106.175.192
14.110.128.84
222.178.10.230
222.178.10.232
222.178.10.238
222.178.10.239

Fig. 7. The analysis of “IP” for user 46

Moving distance Horizontal direction Vertical direction

150
100
50
0
11
16
21
26
31
36
41
46
51
56
61
66
71
76
81
86
91
96
1
6

101
106
111
116
121
126
131
136
141
146

-50
-100
-150
-200
-250

Fig. 8. Mouse moving distance in horizontal and vertical directions

 Construction of Network User Behavior Spectrum 141

Moving rate Horizontal direcƟon VerƟcal direcƟon

10

0
11
16
21
26
31
36
41
46
51
56
61
66
71
76
81
86
91
96
1
6

101
106
111
116
121
126
131
136
141
146
-5

-10

-15

Fig. 9. Mouse moving rate in the horizontal and vertical directions

Table 2. Behavior spectrum result for user 46

Confidence NR Examples Instructions
82.00% 123 C*pm*Chrome User C often uses
chrome to log in to
the website in the
afternoon
31.30% 47 C*pm*Chrome*222.178.10.232 User C often uses
chrome to log in to
the website in the
afternoon. The often
used IP is
“222.178.10.232”
16.67% 25 C*pm*Chrome*222.178.10.232*control User C often uses
center chrome to log in to
the website in the
afternoon. The often
used IP is
“222.178.10.232”.
The often used head
is “control center”

The user behavior spectrum of No. 46 is shown in the Fig. 10.

142 M. Xu et al.

Fig. 10. The user behavior spectrum of No. 46

5 Conclusions

In this article, a method of representing user behavior patterns based on the user
behavior spectrum is proposed by us. Firstly, the related definitions of perceptual
behavior and physiological behavior are given. Then, the feature values of the user’s
perceptual behavior and physiological behavior are used for statistical analysis and
correlation analysis. Thus, analytical methods based on perceptual behavior and
physiological behavior are established. Finally, these analysis methods are used to
establish one user’s behavior spectrum, and the results showed that the combination of
two behaviors to construct the user’s behavior pattern is more comprehensive and
three-dimensional than analyzing only a single behavior. The purpose of describing the
behavior can be achieved by establishing the behavioral spectrum. Only after the
description of the behavior, problems such as the behavior’s development, mechanism,
function and evolution can be proposed. Moreover, the establishment of user behavior
spectrum is significant for the identification of user identities.
In the future, there is still a lot of work that needs to be done by us. The analysis of
the combination of perception and physiological behavior needs to be more thorough,
the mining of behavioral data needs to be deeper, and the construction of the user’s
behavioral spectrum needs to be more complete. In addition, only the data from self-
built websites is used by us to establish a behavioral spectrum. However, self-
constructed websites are short-lived and have less data. In the future, more data from
cooperative companies will be collected and the establishment of user behavior
spectrum for employees in the company will be completed. Also, corporate data will be
used to study the evolution of user behavior.
 Construction of Network User Behavior Spectrum 143

Acknowledgment. This work was supported by Fundamental Research Funds for the Central
Universities NO. 106112016CDJXY110003, 2016.1-2017.12 and the National Natural Science
Foundation of China under Grant No. 71671020.

References
1. Frhan, A.J.: Visualization and analysis of user behaviour patterns for multimedia content
view in social networks. In: 5th ISEEE International Symposium on Electrical and
Electronics Engineering, pp. 1–7. IEEE Press, Galati, Romania (2017)
2. Qiang, W., Ji-min, L., Wan-hai, Y., et al.: User identification in the preprocessing of web log
mining. Comput. Sci. 29(4), 64–66 (2002)
3. Zhu, T., Wang, B., Wu, B., et al.: Role defining using behavior-based clustering in
telecommunication network. Expert Syst. Appl. 38(4), 3902–3908 (2011)
4. Katerina, T., Nicolaos, P.: Mouse behavioral patterns and keystroke dynamics in end-user
development: what can they tell us about users’ behavioral attributes? Comput. Hum. Behav.
83, 288–305 (2018)
5. Nishida, T., Kano, T., Goodall, J., et al.: Ethogram and ethnography of Mahale chimpanzees.
Anthropol. Sci. 107(2), 141–168 (1999)
6. Hall, C., Heleski, C.: The Role of the ethogram in equitation science. Appl. Anim. Behav.
Sci. (2017)
7. Ivanov, D.G., Krupina, N.A.: Changes in the ethogram in rats due to contagion behavior.
Neurosci. Behav. Physiol. 47(8), 987–993 (2017)
8. Qin, T., et al.: Behavior spectrum: an effective method for user’s web access behavior
monitoring and measurement. In: IEEE Global Communications Conference, pp. 961–966
(2013)
9. Schacter, D.L.: Psychology. Worth Publishers, New York (2011)
10. Prosser, C.L.: Comparative Animal Physiology, Environmental and Metabolic Animal
Physiology, 4th edn, pp. 1–12. Wiley-Liss, Hoboken (1991). ISBN 0-471-85767-X
 Relation Analysis of Heating Surface’s Steam
Temperature Difference and Fouling Degree
Based on the Combination of Thermodynamic
Mechanism and Production Data Mining

Jiahui Wang1(&), Hong Qian1,2, and Cheng Jiang1

1
School of Automation Engineering,
Shanghai University of Electric Power, Shanghai, China
[email protected], [email protected],
[email protected]
2
Shanghai Power Plant Automation Technology Key Laboratory,
Shanghai, China

Abstract. This paper aims to monitor heating surface fouling degree of boiler
without fume temperature data recorded. Based on the thermodynamics, this
paper analyzes the heat balance mechanism of the heat transfer characteristics of
the convection heating surface of the boiler and the endothermic characteristics
of the steam and obtains the relationship between the heat transfer coefficient
and the temperature difference and it is proved that the temperature difference
between the heating surface and heat transfer coefficient has a strong correlation
through the gray correlation analysis method. This paper analyzes the actual
production data by correlation analysis and regression analysis: the amount of
heat released from the combustion of the boiler is certain when the load is stable,
and based on the model for the temperature difference of the working fluid to
represent the degree of dust accumulation is established. This paper validates the
result that the model is able to reflect the degree of dust deposition on the
convective heating surface well by simulation and field data verification
methods.

Keywords: Steam temperature difference
Fouling degree
Data mining

Characteristics of heating surface
Ebsilon simulation

1 Introduction

It is a common phenomenon that ash particles produced during solid fuel combustion
adhere to surfaces of heat transfer tubes inside boilers. How to estimate fouling degree
is one of the most urgent puzzles for electric engineers in industry. As for now, the soot
blowing of large coal-fired power station boilers generally employs a fixed quantitative
method to purge each heated surface every certain time [1]. This blind and centralized
soot-blowing method is likely to result in insufficient soot-blowing or excessive soot-
blowing, which will not only damage boiler tube but also affect the stability of the
operation, and reduce the generator unit’s operating efficiency [2]. Therefore, it is very

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 144–154, 2018.
https://doi.org/10.1007/978-981-13-2384-3_14
 Relation Analysis of Heating Surface’s Steam Temperature Difference 145

necessary to transform the “timed and quantitative” soot-blowing method into “on-
demand distribution” soot-blowing methods so as to improve the safety and economy
of boiler operation.
Researchers at home and abroad have optimized the boiler soot-blowing system
through various methods [3–5]: Prof. Wei of North China Electric Power University
proposed the concept of clean factor [3], the ratio of real heat absorption and maximum
transferable heat of ideal heat exchanger [4], that is, the ratio of heat transfer effective-
ness to characterize the degree of dust accumulation; the method of exporting the
temperature of flue gas as a whole reflects the total degree of ash pollution in the
furnace and its effect. And as for specific heated surface, there is no way to know its
fouling degree.
This paper takes a 300 MW coal-fired power plant boiler (DC furnace) as the
research object, and analyzes the gray correlation between the temperature difference
between the inlet and outlet working surface of the convective heating surface and the
cleaning factor calculated based on the thermodynamic mechanism based on the
excavation of a large amount of actual production data. The characteristics prove that
the temperature difference on the side of the heating surface can reflect the degree of
dust accumulation. This method can characterize the degree of dust accumulation based
on limited measurement points of heated temperature without calculating the
endothermic enthalpy value. This method is proved to be reliable through field test.

2 Modeling of the Relationship Between Ash Fouling

and Temperature Difference of the Working Fluid

The characteristics of convective heat transfer and endothermic absorption of the

working fluid are analyzed. The model of the cleaning factor that characterizes the ash
content of the heated area was determined and the idea of using the temperature
difference at the inlet and outlet of the heated surface to characterize the degree of dust
accumulation was proposed.

2.1 Analysis of Heat Transfer Characteristics of Convective Heating

Surface
For the type P boiler studied in this paper, heating surfaces mainly absorb the heat in
the flue gas by convection. The basic formula for convection heat transfer is [6]

Q1 ¼ KHDt=Bj : ð1Þ

In the formula, H is heat surface area, m2; Dt is heat transfer temperature and
pressure,  C; Bj Calculated fuel consumption, kg/s; K is heat transfer coefficient,
kW=ðm2
 CÞ.
For the heated surface of the fouling external surface of the pipe, the heat transfer
coefficient K, which does not take into account the thermal resistance and scaling of the
metal pipe, is calculated as
146 J. Wang et al.

1
K¼ ð2Þ
1=a1 þ e þ 1=a2

In the above formula, a1 and a2 are the exothermic coefficient of the flue gas
against the wall and the wall of the working medium, kW=ðm2
 CÞ; e is the coefficient
of ash pollution, ðm2
 CÞ=kW. As the degree of dust accumulation on the heated
surface increases, the fouling coefficient e will gradually increase, and the value of heat
transfer coefficient K will gradually decrease.

2.2 Analysis of Endothermic Characteristics of Convective Heating

Surface
For superheaters, reheaters, and economizers, the convective heat absorption of the
working fluid Q2 is generally calculated as follow:

Q2 ¼ Dðh2  h1 Þ=Bj : ð3Þ

In the formula, Bj is calculated fuel volume, kg/s; D is steam flow, kg/s; h1 ; h2 is

enthalpy of heated inlet and outlet surface respectively, kj/kg.
Calculation of the enthalpy of the working fluid can refer to the “Water and Steam
Thermodynamic Properties Chart” [7] written by Yan Jialu. Through the look-up table,
the steam enthalpy corresponding to temperature and pressure can be obtained.
To obtain the relationship between difference in enthalpy and that in temperature,
this paper sets the enthalpy value at 265 °C as calculated base, and calculates the
difference between enthalpy at different temperature every 25 °C and that at 265 °C in
3 different pressure conditions, as the diagram below shows in the left. In the same
way, the diagram in the right is drawn, and its base is the enthalpy value at 350 °C, and
temperature difference is every 15 °C in 7 different pressure conditions Fig. 1.

700 800

1MPa 11MPa
700
600 12MPa
2MPa
3MPa 13MPa
600
500 14MPa
15MPa
500
16MPa
400
enthalpy difference
enthalpy difference

400
300
300

200
200

100 100

0 0
0 50 100 150 200 250 300 0 20 40 60 80 100 120 140 160 180 200
temperature difference temperature difference

Fig. 1. Characteristic of enthalpy difference Fig. 2. Characteristic of enthalpy difference

caused by temperature difference. caused by temperature difference.
 Relation Analysis of Heating Surface’s Steam Temperature Difference 147

From the above two figures, it can be found that the enthalpy difference is
approximately linear with a certain pressure and temperature difference Fig. 2.

Dh ¼ aDT: ð4Þ

Thus, the convective heat absorption can also be expressed by the following
formula

Q2 ¼ DaDT=Bj ð5Þ

2.3 The Relationship Between Dust Accumulation and Temperature

Difference

Qualitative Analysis of Relevance. For a specific convective heating surface, the heat
Q1 transferred by the convection and the heat Q2 absorbed by the working fluid should
be in balance [7]. The conclusion can be made from (1) and (5),

KHDt DaDT
¼ : ð6Þ
Bj Bj

So, the formula of heat transfer coefficient K can be inferred,

DaDT
K¼ : ð7Þ
HDt

From the formula (7), the heat transfer temperature and pressure can be calculated
through the inlet and outlet temperature parameters on the steam side and the flue gas
side, however, it will inevitably be limited in practical applications. Based on the above
formula, it can be implied that there is a certain correlation between heat transfer
coefficient K and the temperature difference ΔT.
Correlation is defined as the measure of the size of systems’ association with time or
different objects [8]. In the process of system development, if the trend of heat transfer
coefficient K is consistent with that of temperature difference ΔT of the working fluid, it
means the degree of synchronization changes is relatively high, vice versa.
Quantitative Analysis of Relevance. The degree of correlation is essentially the
degree of difference between the geometric shapes of the curves. The correlation
coefficient for the reference sequence K and the comparison sequence ΔT at each time
can be calculated by the following formula:

Dmin þ qDmax
n¼ : ð8Þ
DðtÞ þ qDmax

Among them, the resolution coefficient q is generally between 0 and 1; DðtÞ is the
absolute difference between each point on the K curve and the ΔT curve, Dmin is
minimum of the absolute difference at all time, and Dmax is the maximum one.
148 J. Wang et al.

The correlation r is a number that represents the correlation degree which is cal-
culated by merging the correlation coefficients at each time (i.e., each point in the
curve) into one value. And take this value as the index to evaluate the relationship
between ΔT and K. The formula is as follows:

1X N
r¼ nðtÞ: ð9Þ
N t¼1

In this paper, the heat transfer coefficient K model is established using the low-
temperature reheater with a wide range of working surface heat and temperature
measurement points in the research object, and the correlation between heat transfer
coefficient and the temperature difference at the working side is analyzed by the gray
correlation analysis method to verify that the temperature difference in the working side
reflects the degree of dust accumulation.

3 Analysis of the Correlation Between the Temperature

Difference and the Degree of Fouling Degree on the Side
of the Convective Heating Surface

Take the above 300 MW coal-fired power plant boiler (DC furnace) as an example. In
the load range of [160 MW, 300 MW], the load stability section was selected to
analyze the soot blowing system.

3.1 Change in Fouling Degree in Heated Area

The form of fouling in heated area is a dynamic and balanced process. Heat transfer
coefficient is determined based on the flow rate of the low-temperature reheater, the
temperature difference between the inlet and outlet sides of the cryogenic heat exchanger
and the flue gas side when a stable operating condition is selected. According to the soot-
blowing records, the load curve of the day and the Heat transfer coefficient (Pu) curve of
the low-temperature reheater were plotted, as shown in Figs. 3 and 4:

300 Start blowing

1

280 0.9

0.8
eat transfer coefficient

260
0.7
load/MW

0.6
240
0.5

220 Start blowing 0.4

0.3
200
5:47 8:47 11:47 14:47 0.2
time/h 5:47 8:47 11:47 14:47
time/h

Fig. 3. Real-time load data of coal-fired

Fig. 4. Heat transfer coefficient (Pu) of Low-
boiler.
temperature reheater curve.
 Relation Analysis of Heating Surface’s Steam Temperature Difference 149

From Fig. 4, it can be seen that the per-unit value of heat transfer coefficient (K) of
the heating surface in the low-temperature reheater increases from 0.3 to 0.9 after the
start of soot blowing. As the time go by, the degree of dust accumulation in the low
reheating area gradually increases, and the K value gradually decreases.

3.2 Correlation Between Heat Transfer Coefficient and Temperature

Difference on Heating Surface Side
Similarly the curve of the temperature difference between the inlet and outlet of the
cryogenic reheater on day A is drawn, as shown in Fig. 5. It can be seen that the curves of
Figs. 4 and 5 are very similar. Using the gray correlation analysis method, the relation-
ship coefficients of 541 moments in this time period are calculated, as shown in Fig. 6.
The average value of the 541 relationship coefficients, known as correlation degree
r, is obtained. The correlation between the temperature difference at the inlet and outlet
side of the low-temperature reheater and the heat transfer coefficient is 0.61, which is a
strong correlation.
110 1

108
0.9
106
0.8
temperature difference/

104

102 0.7

100
0.6
98
0.5
96

94 0.4

92
5:47 8:47 11:47 14:47
time/h 0 100 200 300 400 500 600
time/min

Fig. 5. Temperature difference curve diagram Fig. 6. Correlation coefficient n between

for the inlet and outlet of the low-temperature heat transfer coefficient and temperature
reheater. difference on heating surface side at each
time.

Through the analysis of the field data, it is proved that under certain load condi-
tions, the temperature difference ΔT on the working side is positively correlated with
the heat transfer coefficient K, so the temperature difference on the working side can
reflect the degree of dust accumulation on the heated surface.

4 Fouling Degree Model in Convection Heating Surface Side

Through Temperature Difference

4.1 Fundamentals of Fouling Degree Model

This study does not consider influence of the types of coal, and the heat released from
fuel combustion is mainly determined by the amount of coal. According to the actual
data, the data of 1440 moments in a certain day were taken, and the load P was used as
150 J. Wang et al.

the independent variable, and the coal supply B and the wind volume W were used as
the dependent variables to perform the linear regression analysis.
The relationship between the amount of actual coal B on one day and the amount of
^ is shown in Fig. 7. And the relationship between the amount of
coal fitted by the load B
actual wind W and the amount of wind fitted by the load is shown in Fig. 8.

Fig. 7. Linear fitting of coal quantity to load. Fig. 8. Linear fitting of wind W to load

Analysis result is shown in Table 1.

Table 1. Correlation analysis and linear fitting results

X Y r R2 a b
P W 0.987 0.975 0.295 29.618
P B 0.962 0.914 0.205 9.320

From the results above, it can be concluded that there is a good linear relationship
between the coal feed amount B, the total air flow rate W, and the load P. That is, when
the load is constant, the amount of heat released from the combustion of the boiler is also
constant, and the magnitude of the load is corresponding. Therefore, the difference in
heat absorption per unit working medium on the convective heating surface under the
same load can reflect the degree of fouling degree on the convective heating surface.

4.2 Fouling Degree Model Establishment

The difference in heat absorption per unit working medium at the convective heating
surface under the same load can reflect the degree of dust accumulation on the con-
vective heating surface. The coefficient that defines the degree of fouling of the con-
vective heating surface is b, and the calculation formula for b is:
8
< 1 DT  TL :
DTTL
b¼ TL  DT  TH : ð10Þ
: TH TL
0 TH  DT:

In the above formula, DT is the actual value of the temperature difference between
the inlet and the inlet of the convective heating surface whose unit is °C. From formula
 Relation Analysis of Heating Surface’s Steam Temperature Difference 151

(10), it can be inferred that the fouling degree coefficient b is between 0 and 1, and b
approaching 0 means that the heated area is severely covered by ash; while b
approaching 1 means that the heated surface tends to be clean.
According to the boiler’s 696,960 samples in 484 days of operation, 287 samples
with stable load in the range of [160 MW, 300 MW] are analyzed. Since for different
heat receiving surfaces under different loads, the working medium heat absorption
interval values differs as well, and the temperature difference between the import and
export of the lower limit value of the TL, the upper limit value TH and the load
relationship is shown in Table 2:

Table 2. Upper and lower limit of inlet and outlet heater surface temperature.
Heating surface Lower limit TL Upper limit TH
Screen superheater −0.0136  P+86.1 −0.155  P+149.7
High temperature superheater −0.017  P+37.66 −0.066  P+58.09
Low temperature superheater −0.08  P+55.26 −0.23  P+115.34
High temperature reheater 0.028  P+74 −0.13  P+139.5
Low temperature reheater −0.09  P+169.73 −0.3  P+207.33

That is, the temperature difference of the working surface of the heating surface
indicates the degree of dust accumulation, which needs to be analysed and determined
in conjunction with the load at that time. The larger the load, the greater the temper-
ature difference of the working fluid will be correspondingly larger, and the difference
in the dust level represented by the temperature difference of the working fluid under
different loads is different.

5 Verification of Fouling Degree Model

5.1 Software Simulation

EBSILON Professional is used to calculate the physical quantities such as heat, work,
thermal cycle efficiency, and thermal state parameters in the thermal process. Of course,
the actual process should be revised on the basis of the ideal process.
There are many component models in the EBSILON component library, which can
be used to build the power station based on the complexity of the power station system
or the accuracy of the research problem. Therefore, two models are established, which
are low-temperature reheater heat transfer model and high-temperature superheater heat
exchange model respectively shown in Figs. 9 and 10.
In order to verify that the temperature difference of the working surface of the
heating surface represents the degree of dust accumulation, the real-time data of five
groups of low retest points under the same conditions of the load are input into the two
models to determine the heat transfer coefficient. Though lack of data on the high-
temperature superheater flue gas side in the field data, it is still possible to obtain heat
transfer coefficient by simulation based on the boundary condition of the data on the
152 J. Wang et al.

Fig. 9. Low-temperature reheater heat excha- Fig. 10. High-temperature superheater heat
nge model. exchange model.

flue gas side, the working side of the low temperature superheater, and the flue gas side
of the rear screen superheater. The results are shown respectively in Table 3 and
Table 4.
According to the output results in Tables 3 and 4, it can be concluded that whether
on the surface of the low temperature reheater or on that of high temperature

Table 3. Simulation results of low-temperature reheater model.

Part of the input Part of the output
Steam inlet/outlet Temperature Coefficient of heat
temperature(°C) difference(°C) transfer(kW/K)
305.74 467.674 161.934 485.873
306.48 465.27 158.79 461.458
321.54 477.55 156.01 460.721
307.15 460.260 153.11 434.006
317.14 469.915 152.775 399.982

Table 4. Simulation results of high temperature superheater model.

Part of the input Part of the output
Steam inlet/outlet Temperature Coefficient of heat
temperature(°C) difference (°C) transfer(kW/K)
489.537 524.536 34.999 34.978
485.648 521.820 36.172 50.830
486.938 524.873 37.935 80.673
485.889 524.463 38.574 95.746
485.833 525.893 40.06 119.800
 Relation Analysis of Heating Surface’s Steam Temperature Difference 153

superheater, when the temperature difference between the inlet and outlet working
fluids decreases, the heat transfer coefficient will also decrease.

5.2 Verification of Field Data

According to the soot-blowing records, the load is relatively stable from 11:05 to 21:05
on a certain day, and the soot-blowing time is 17:05. The soot blowing area is the
heated surface of the high-temperature superheater (taking the heated surface as an
example). The temperature difference curve of the outlet steam and the high-
temperature superheater, b curve of the high-temperature superheater is shown in
Figs. 11 and 12, respectively.
41 0.7

40 0.6

39 0.5
temperature difference/

38 0.4

37 0.3

36 0.2

35 0.1

34 0
11:05 13:05 15:05 17:05 19:05 21:05 11:05 13:05 15:05 17:05 19:05 21:05
time/h time/h

Fig. 11. Temperature difference curve of work- Fig. 12. Fouling degree coefficient b curve
ing fluid in high temperature superheater of high-temperature superheater.

From the above figure, it can be seen that the temperature difference of the working
surface of the high-temperature superheater heating surface after soot blowing rises
from 35 °C to 40 °C, and the fouling degree coefficient b increases from 0.1 to 0.5. As
time went by, the high reheating area gradually accumulated ash, and the temperature
difference and coefficient of dust accumulation gradually decreased.

6 Conclusion

The temperature difference varies between the inlet and outlet side on different heating
surfaces under different load. However, it is not difficult to find the relationship
between the temperature difference limit of each heating surface and the load through
the production data. And to some extent, the real-time temperature difference data can
effectively reflect the degree of fouling degree. The results show that:
(1) The degree of correlation between the temperature difference on the heating
surface side of the working surface and the heat transfer coefficient is strong.
(2) The temperature difference on the heating surface of the heating surface can
effectively reflect the degree of dust accumulation.
154 J. Wang et al.

(3) In the case of a smokeless temperature measurement point of the boiler, it can
provide guidance to the working personnel and help achieve on-demand blowing.
This study only focused on stable conditions, the unit parameters themselves have a
certain degree of volatility, and further processing was done using data filtering
methods. Under varying conditions, the parameters of the unit have a wide range of
changes and needs to be further explored in the future.

References
1. Shi, Y.H., Yun, F., Wei, T.Z.: Research on ash fouling monitoring of heating surface in coal
fired utility boiler. J. Power Syst. Eng. 26(06), 49–50 + 53 (2010)
2. Qian, H., Song, L., Chen, Q.Q.: Research on the believe rule base of soot blowing demand
based on production data mining. J. Therm. Power Gener. 46(06), 113–118 (2017)
3. Yan, W.P., Liang, X.J., Zhou, J., Ye, X.M., Zheng, Z.: Ash monitoring and sootblowing
optimization of a 300 MW coal-fired utility boiler. Proc. CSEE 20(01), 58–61 + 129 (2007)
4. Wu, S.Y.: Effect of fouling on heat transfer performance of heat exchanger. J. Petro-Chem.
Equip. 29(1), 10–12 (2000)
5. Xu, X.H., Zhou, K.Y., Wei, H.Q., Xu, J.Q.: Increment of furnace exit gas temperature for
monitoring ash fouling in the furnace of coal-fired boilers. J. Proc. CSEE 29, 21–26 (2011)
6. Zhou, Q.T.: Principle of Boilers, 2nd edn, pp. 183–185. China Electric Power Press, Beijing
(2009)
7. Yan, J., Yu, X., Wang, Y.: Water and Steam Thermodynamic Properties Chart, 3rd edn,
pp. 13–53. Higher Education Press, Beijing (2015)
8. Sun, C.Z., Song, Y.T.: Theory discussion about grey interconnect degree model. J. Global
Geol. 19(03), 248–252 + 270 (2000)
 Variable Selection Methods in Dredger
Production Model

Yinfeng Zhang, Zhen Su, and Jingqi Fu(&)

Department of Automation, College of Mechatronics Engineering and

Automation, Shanghai University, No.149, Yanchang Rd., Shanghai 200072,
China
[email protected]

Abstract. The production of earthwork is an important index to evaluate the

performance of dredgers. Because the parameters affecting production are
numerous and not independent of each other, it is easy to overfit the production
model and have low accuracy. In view of this problem, based on the measured
data of a Trailing Suction Hopper Dredger (TSHD), three variable selection
methods are applied to select the parameters that can affect the yield most and
the inputs of the final production model are determined. The results show that
the deleted and retained parameters conform to the actual working conditions.
Finally, the advantages and disadvantages of these three methods and their
applicability under different working conditions of dredgers are analyzed.

Keywords: Hopper dredging
Yield model
Variable selection

1 Introduction

Dredgers are designed and manufactured for underwater earth and rock works. They
can be used to excavate canals, widen waterways, remove siltation from reservoirs, dig
trenches, collect ore and so on. The TSHD is more popular than other kinds of dredgers
because of its ability to complete the work of digging, loading, transporting, dis-
charging (blowing) independently and change the site itself. Nowadays, the dredging
equipment and auxiliary equipment of TSHDs are becoming gradually perfect. How-
ever, the operation and management level of TSHDs is generally not high, which
affects the dredging efficiency of TSHDs directly.
The optimization of the dredging parameters of TSHD is to obtain the relevant
construction parameters, so that the dredger can be in good operation condition and
achieve the purpose of high efficiency, high quality and low consumption when
dredging under the guidance of these parameters. However, giving proposed con-
struction parameters are not simple, the main difficulties are as follows: 1. It is unre-
alistic to establish an accurate input-output mechanism model for a complex nonlinear
multivariable time-varying system such as dredging operation system from a basic
principle. 2. The characteristics of the actual dredging equipment are greatly influenced
by the construction environment and operation conditions. 3. The dredging process
must ensure the safety of the slurry transport system and unreasonable parameter

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 155–165, 2018.
https://doi.org/10.1007/978-981-13-2384-3_15
156 Y. Zhang et al.

adjustment will cause abnormal situations such as blockage, rupture and forced
shutdown.
Holland IHC company is one of the most important dredger manufacturing and
research units in the world. They also studied the optimization of dredging operations
and put forward new concepts such as “Efficient Dredging” and “Accurate Dredging”
[1]. But the results of their dredging optimization are not detailed, testing results and
key technologies not transparent, usually only a concept. In China, the research reports
on the optimization of dredging work are rare and floating on the surface, so it is
difficult to study their essence and specific performance.
Modeling and optimization of dredging operations often use model predictive
control [2], and the relationship between production and influencing factors must be
understood before modeling. In this paper, we focus on maximizing the production of
dredging. According to the experience of actual operator and some simulation exper-
iments, the factors that affect the yield can be summarized as follows: speed to ground,
vacuum degree of dredge pump, depth and pressure of drag head, soil characteristics [3,
4], pump speed, high pressure flushing system, overflow using [5] and so on.
Accordingly, there are many parameters related to production. If taking all the
parameters as the inputs of TDS model, it is easy to have problems such as over fitting,
low precision and long modeling time. The purpose of variable selection is to select a
certain feature set from production-related parameters as inputs to establish an effective
TDS model, which can improve the accuracy while removing redundant information,
reducing the complexity, improving the generalization performance and simplifying the
model. Three methods are applied to realize the screening of production-related
parameters, namely, least absolute shrinkage and selection operator (LASSO), mean
impact value (MIV) and genetic algorithm (GA).

2 Method Description

Taking the magnitude and dimensional difference of measured data of dredgers into
account, the data is first normalized in the early stage of variable screening. This paper
uses deviation normalization to process data and the formula is as follows:
xk  xmin
xk ¼ ð1Þ
xmax  xmin

where xmin is the minimum number of data series and xmax is the largest number.

2.1 Variable Screening by LASSO

Define the dredging output with tones of dry soil (TDS):

mt ðtÞ
vt ðtÞ  qw mt ðtÞ  qw vt ðtÞ
TDSðtÞ ¼ vt ðtÞqq ¼ qq ð2Þ
qq  qw q q  qw
 Variable Selection Methods in Dredger Production Model 157

where qw (1.024 ton/m3) and qq (2.65 ton/m3) denote the density of water and quartz;
mt and vt are loading quality and capacity.
When establishing the yield model, it is assumed to be a linear model. The
production-related parameters are recorded as x1 ; x2 ; . . .. . .xp as independent variables,
and the TDS is calculated according to the real data and Eq. (2) and recorded as y as the
dependent variable. Supposing that n groups measurement data of TSHD have been
standardized as follows: y1 ; x11 ; x12 ; . . .x1p ; y2 ; x21 ; x22 ; . . .x2p …… yn ; xn1 ; xn2 ; . . .xnp .
Create a linear model of production and production-related parameters:

y ¼ b1 x1 þ b2 x2 þ . . .. . .bp xp þ e ð3Þ
01 0 1 0
1 0 1
y1 x11 ... x1p b1 e1
B .. C B .. .. .. C ¼ ð x B .. C B .. C
where y ¼ @ . A x ¼ @ . . . A 1


xp Þ b ¼ @ . A e ¼ @ . A
yn xx1


xnp bp en
ð4Þ

The above production model can be written as:

y ¼ xb þ e ð5Þ

It is easy to obtain coefficients b according to the thought of ordinary least square

(OLS). The basic idea of LASSO [6] is to impose ‘1 punishment on basis of OLS, that
is, to limit the length of parameters and the expression is as follows:

  X
P
^ arg min ðy  xbÞT ðy  xbÞ s:t:
b¼ jbi j  t ð6Þ
b2RP i¼1

Where t  0 and it is a regulating parameter that controls the degree of compres-

sion. When t is smaller, it means stronger compression and more coefficients will
become 0. The above expression can be written as the following equivalent form:

  X
P
^ arg min ðy  xbÞT ðy  xbÞ þ k
b¼ jbi j ð7Þ
b2RP i¼1

Where k is a penalty parameter and k  0.

Solving LASSO is essentially a solution of a quadratic programming problem with
linear inequality constraints. In addition to modern optimization theory and methods
based on constraints, least angle regression (LAR) is also commonly used in LASSO.
The solution flow of LAR [7] is presented in Algorithm. 1.
158 Y. Zhang et al.

Algorithm 1. Least angle regression

1. Standardize the independent variables to have mean zero and unit norm. Start with the
residual r = y − y , β , β , ......β = 0 .
1 2 p

2. Find the independent variable x j most correlated with r .

3. Move β j from 0 towards its least-squares coefficient x j , r ,until some other competi-
tor xk has as much correlation with the current residual as does x j .
4. Move β j and β k in the direction defined by their joint least squares coefficient of the cur-
rent residual on x j , xk ,until some other competitor xl has as much correlation with the
current residual.
5. Continue in this way until all p predictors have been entered. After min ( n − 1, p )steps,
algorithm stops.

2.2 Variable Screening by MIV

The idea of MIV is to fluctuate the measured value of the production-related parameters
to affect the TDS output, and then sort the parameters according to the magnitude of the
output change. It is based on the TDS model and improved BP network [8] with GA is
used for TDS modeling.
Using GA to improve BP neural network is divided into 3 stages: determining the
structure of BP network, GA optimizing the weight and threshold of BP network and
predicting BP network. The structure of BP network is determined firstly according to
the number of input and output parameters, and then determine the length of GA
individual. Individual coding is implemented by real number, so each individual is a
real string, which is composed of 4 parts: the connection weight of the input layer and
the hidden layer, the threshold of the hidden layer, the connection weight of the hidden
layer and the output layer, and the threshold of the output layer.
GA optimization refers to optimizing the weights and thresholds of BP network.
Each individual in population contains all the weights and thresholds of a network. GA
evaluates the pros and cons of the individual through fitness function and then finds the
individual who has the optimal fitness value by selection, crossover and mutation
operation. The sum of absolute errors between real TDS and predicted TDS is used as
individual fitness function.
The prediction of BP neural network is the assignment of the initial weights and
thresholds of BP network by the optimal individuals obtained by GA, and then the BP
network is trained to predict TDS.
MIV is considered as one of the best indexes in neural network to evaluate the
degree of input affecting output. The symbol of MIV value represents the relative
direction (positive and negative correlation) and absolute value represents the relative
degree. The specific calculation process [9] is shown in Algorithm. 2.
 Variable Selection Methods in Dredger Production Model 159

Algorithm 2. Selecting variables with MIV

1. Use training sample P to train improved BP network.
2. Training sample P is added 10% to get new training sample P1 and reduced 10% to get P2.
P1 = P × 1+10% P 2 = P × 1-10%
3. Test the network with P1 and P2 to get new outputs A1 and A2.
A1 = sim ( P1) A2 = sim ( P 2)
4. The impact value(IV) due to the change of training sample P is the difference between A1
and A2. IV = A1 − A2
IV
5. Calculate the mean impact value. MIV =
N
6. Sort the independent variables according to the absolute value of the MIV.

2.3 Variable Screening by GA

In the field of variable selection, GA could be utilized to search for latent variables
spontaneously [10]. When using GA to solve the problem of variable selection in
TSHD, two designs are mainly considered: one is how to use production-related
parameters for chromosome coding, and the other is how to design the fitness function.
In this paper, the encoding length is designed to be 18 (corresponding to the
number of parameters in Table 1). A population of binary strings (i.e. chromosomes) is
created randomly from all the variables. Each position (i.e. gene) of the binary string
corresponds to one specific variable, which is coded as “1” if the variable is selected
and “0” if not [11].
The fitness function in this paper is used to evaluate the effectiveness of a certain set
of production-related parameters in the establishment of TDS model. Therefore, the
reciprocal of the squared sum of the errors of the actual TDS and the network predicted
TDS can be chosen as fitness function and the expression is as follows:

1 1 1
f ðXÞ ¼ ¼ ¼P ð8Þ
^
SE sseðT  TÞ n
ð^t  t Þ2 i i
i¼1

^ f^t1 ; ^t2 ; . . .^tn g is the TDS of test set; T ¼ ft1 ; t2 ; . . .tn g is the real
In the formula, T¼
TDS of test set; n is the number of samples for the test set.
To avoid the influence of randomness of the initial weights and thresholds of BP
network on the fitness function calculation, the weights and thresholds of BP neural
network are optimized by GA for each individual in calculating the fitness function
value, that is, the improved BP network using GA mentioned above.
160 Y. Zhang et al.

Table 1. Primary selection of production-related parameters.

Variable number Variable name Variable type
1 Depth of portside drag Control parameter
2 Angle of portside drag Control parameter
3 Differential pressure of portside drag Measurement parameter
4 Pressure of portside wave compensator Control parameter
5 Motor speed of portside jet water pump Control parameter
6 Vacuum degree of portside dredge pump Measurement parameter
7 Speed of portside dredge pump Control parameter
8 Position of portside diversion window Control parameter
9 Depth of starboard drag Control parameter
10 Angle of starboard drag Control parameter
11 Differential pressure of starboard drag Measurement parameter
12 Pressure of starboard wave compensator Control parameter
13 Motor speed of starboard jet water pump Control parameter
14 Vacuum degree of starboard dredge pump Measurement parameter
15 Speed of starboard dredge pump Control parameter
16 Position of starboard diversion window Control parameter
17 Speed of dredger Control parameter
18 Current depth of overflow weir Control parameter

The total design flow of variable selection with GA is shown in Algorithm. 3.

Algorithm 3. Selecting variables with GA

1. Encode individuals and generate initial populations.
Encoding rules: If a bit value of the chromosome is "1", it means that the corresponding
input variable of the bit participates in the final TDS modeling; conversely, "0" means not
participating in modeling.
2. Use population to train improved BP network.
3. Calculate fitness function value.
4. If the termination condition is satisfied, the optimization result is output; otherwise, the
selection, crossover, and mutation operations are performed to obtain a new population and
jump to step 2.
5. Decode the optimal individual and obtain the selected parameters.

3 Experimental Result
3.1 Sources of Data
Modern TSHDs are equipped with advanced equipment, such as DGPS high-precision
positioning systems, draft gauge indicators, drag arm position indicators, concentration
meters, flowrate indicators and output calculating devices. So various parameters in the
process of dredging can be measured.
 Variable Selection Methods in Dredger Production Model 161

The experimental data in this paper are taken from the measured dredging data of a
TSHD in Xiamen. According to the actual operator’s experience and some simulation
results, the primary selection of production-related parameters is listed in Table 1.

3.2 The Result of Variable Selection

LAR and modern optimization theory are introduced to solve LASSO, and the results
are shown in Figs. 1 and 2. The horizontal axis of the graph represents the penalty
terms and the vertical axis represents the coefficients of production-related parameters.
The paths of the different color lines represent coefficients solution and we can see that
the coefficients solution paths of LAR and LASSO are highly similar. From right to
left, with the increase of penalty, the coefficients of parameters gradually decrease to 0.
Selecting appropriate penalty item and deleting the production-related parameters with
coefficients of 0 can achieve the purpose of variable selection.

Fig. 1. LAR coefficient solution path.

More generally, it is possible to sort the importance of production-related param-

eters according to the LASSO coefficients by selecting the appropriate penalty and the
result is shown in Table 2.
The MIV algorithm is trained for 5 times to obtain average value (AVG) and the
specific training results are shown in Table 3. The average MIV value of motor speed
of jet water pump and diversion window position are 0, indicating that their changes do
not affect the TDS.
According to the absolute value of MIV, the importance ranking of production-
related parameters is arranged, and the result is shown in Table 4.
Figure 3 is the evolution curve of population fitness function, which is optimal in
the 75th generation.
The results of variable selection with GA cannot sort the variables according to the
degree of importance. Compared with the two methods above, the deleted variables are
angles of portside and starboard drags, pressures of portside and starboard wave
compensators positions of portside and starboard diversion windows and motor speed
of starboard jet water pump.
162 Y. Zhang et al.

Fig. 2. LASSO coefficient solution path.

Table 2. Variable screening results of LASSO.

Importance number 1 2 3 4 5 6 7 8 9
Variable number 6 14 18 15 3 9 11 17 2
Importance number 10 11 12 13 14 15 16 17 18
Variable number 1 10 12 4 7 16 5 13 8

Table 3. Statistics list of MIV.

1 2 3 4 5 AVG
1 0.066 0.082 0.042 0.069 0.075 0.067
2 −0.016 −0.047 −0.044 −0.022 −0.034 −0.033
3 −0.056 −0.095 −0.103 −0.068 −0.075 −0.080
4 0.024 0.006 0.007 0.003 0.014 0.011
5 0 0 0 0 0 0
6 −0.104 −0.059 −0.063 −0.159 −0.105 −0.098
7 0.022 0.198 0.061 0.099 0.069 0.090
8 0 0 0 0 0 0
9 0.095 0.094 0.016 0.066 0.080 0.070
10 −0.060 −0.035 −0.035 −0.052 −0.159 −0.068
11 −0.029 −0.038 −0.074 −0.063 −0.063 −0.053
12 0.004 0.022 0.032 0.002 0.026 0.017
13 0 0 0 0 0 0
14 −0.107 −0.092 −0.104 −0.077 −0.092 −0.094
15 0.132 0.364 0.138 0.046 0.342 0.204
16 0 0 0 0 0 0
17 −0.077 −0.039 −0.015 −0.047 −0.067 −0.049
18 0.231 0.238 0.198 0.217 0.261 0.229
 Variable Selection Methods in Dredger Production Model 163

Table 4. Variable screening results of MIV.

Importance number 1 2 3 4 5 6 7 8 9
Variable number 18 15 6 14 7 3 9 10 1
Importance number 10 11 12 13 14 15 16 17 18
Variable number 11 17 2 12 4 5 8 13 16

Fig. 3. Evolution curve of fitness function.

According to the 3 methods above and the order of importance, pressures of

portside and starboard wave compensators positions of portside and starboard diversion
windows and motor speed of starboard jet water pump deleted in GA are ranked in the
end of the other two methods and the motor speed of portside jet water pump in
LASSO and MIV also ranks behind, so delete these 6 variables. In addition, angles of
portside and starboard drags deleted in GA are ranked in the middle of the other two
methods, so these two are preserved temporarily.
Stated thus, the final selected variables can be shown in Table 5.
In combination with the actual situation, the variables removed are all related to soil
properties. The wave compensator is used to compensate the ship’s pressure on the
river bed due to the rise and fall of waves. Usually, when the dredging soil is hard, the
pressure of the wave compensator should be reduced and if soil is loose and easy to dig,
the pressure should be increased. The high pressure flushing system can change the soil
moisture content, making the dredging soil loose and easy to dig. For water diversion
window, when the dense soil is excavated, the opening of the water diversion window
should be enlarged to increase the inflow and should be reduced if soil is loose.
Therefore, these three items should not be deleted.
However, according to the measured data, it is found that the pressure of the wave
compensator at the stage of dredging is kept near a certain value and the speed of the
high pressure flushing motor is 0. Besides, the measured value of the position of
diversion window remains unchanged. This indicates that the high pressure flushing
system and the opening of the water diversion windows has not been activated or
changed. It is preliminarily concluded that the sediment characteristics and the wind
and wave currents have not changed significantly. Therefore, their changes have no
effect on yield, so that the importance of ranking is behind and especially in the MIV
164 Y. Zhang et al.

Table 5. The optimal production-related parameters.

Variable number Variable name Variable type
1 Depth of portside drag Control parameter
2 Angle of portside drag Control parameter
3 Differential pressure of portside drag Measurement parameter
6 Vacuum degree of portside dredge pump Measurement parameter
7 Speed of portside dredge pump Control parameter
9 Depth of starboard drag Control parameter
10 Angle of starboard drag Control parameter
11 Differential pressure of starboard drag Measurement parameter
14 Vacuum degree of starboard dredge pump Measurement parameter
15 Speed of starboard dredge pump Control parameter
17 Speed of dredger Control parameter
18 Current depth of overflow weir Control parameter

screening method, their mean value being 0. But these three items must not be con-
sidered as having no effect on TDS, because their functions are obvious in the sea area
where sediment characteristics and wind waves change significantly. Especially in the
maritime space where soil is dense and influence of wave flow is great, it is necessary
to open the high-pressure flushing system to break the soil and increase the pressure of
the wave compensator and the opening of the water diversion window.
Accordingly, it can be concluded that the results of these three variable selection
methods above conform to actual measurement data. However, the results cannot be
used as input for all TDS models, because the soil characteristics and construction
environments encountered in each working condition are different.

4 Conclusion

The screening results of three methods above are not entirely consistent, but the
variables that should be kept are roughly the same. The main difference is reflected in
the ranking of importance, which is caused by the algorithm principle of LASSO and
MIV. LASSO pays attention to the linear fitting degree of production-related param-
eters to TDS, while MIV method takes the sensitivity of parameters as the importance.
LASSO has a greater impact on the penalty item. When the penalty parameter k is
larger, more coefficients tend to be 0 and less production-parameters are persisted. In
addition, it assumes that the TDS model is a linear model. Although it is not in line
with the actual situation, it has obvious advantages in solving the collinearity problem.
MIV takes the sensitivity of production-related parameters as importance, leading
to some important but constant parameters missing. But for the specific working
environment, if the soil or environmental factors in dredging process have not changed
significantly, these important control variables can remain unchanged.
 Variable Selection Methods in Dredger Production Model 165

GA is the best way to establish effective TDS model, but it only works for specific
conditions, and has no universality like MIV. What’s more, MIV and GA are based on
the BP network model, and there are inherent defects in the BP network, such as the
randomness of initial weight and threshold value, local minimization, slow conver-
gence speed, different choice of network structure and sample dependence. Although
the BP network is optimized by GA, it cannot make up for all the defects of the BP
network, at the same time, the application of GA algorithm increasing the running time
of the program.
Concluding, these three methods above have their own advantages and disadvan-
tages, and GA is the best in terms of modeling effect. They also have one thing in
common: the final input variables are not universal, but they are in line with the actual
situation of the specific dredging operation. Therefore, for the different sea areas, these
three methods above should be reused to determine the best input parameters.

Acknowledgments. This work was financially supported by the Science and Technology
Commission of Shanghai Municipality of China under Grant (No. 17511107002).

References
1. Increased accuracy in data for Dutch dredging operations. World Pumps, 2017(26), 28–29
(2017)
2. Braaksma, J., Babuska, R., Klaassens, J.B., et al.: Model predictive control for optimization
the overall dredging performance of a trailing suction hopper dredger. In: Proceedings of the
18th World Dredging Congress (2007)
3. Stano, P.M., Tilton, A.K., Babuska, R.: Estimation of the soil-dependent time-varying
parameters of the hopper sedimentation model: The FPF versus the BPF. Control Eng. Pract.
24, 67–78 (2014)
4. Chen, X., Miedema, S.A., van Rhee, C.: Numerical modeling of excavation process in
dredging engineering. Procedia Eng. 102, 804–814 (2015)
5. Lendek, Z., Babuska, R., Braaksma, J., de Keizer, C.: Decentralized estimation of overflow
losses in a hopper dredger. Control Eng. Pract. 16, 382–406 (2008)
6. Tibshirani, R.: Regression shrinkage and selection via the LASSO: a retrospective. R. Statist.
Soc. B 73(3), 273–282 (2011)
7. Zhang, L., Li, K.: Forward and backward least angle regression for nonlinear system.
Automatic 53, 94–102 (2015)
8. Moradi, M.H., Abedini, M.: A combination of genetic algorithm and particle swarm
optimization for optimal DG location and sizing in distribution systems. Int. J. Electr. Power
Energy Syst. 1, 66–74 (2011)
9. Zhang, Z.: An estimation of coal density distributions by weight based on image analysis and
MIV-SVM. In: 2015 IEEE Advanced Information Technology Electronic and Automation
Control Conference (2015)
10. Chen, D., Hu, B., Shao, X., Su, Q.: Variable selection by modified IPW (iterative predictor
weighting)-PLS (partial least squares) in continuous wavelet regression models. Analyst 129
(7), 664–669 (2004)
11. Duan, F., Xiao, F., Jiang, J.J., Ma, L.: Automatic variable selection method and a
comparison for quantitative analysis in laser-induced breakdown spectroscopy. Spectrochim.
Acta, Part B 143, 12–17 (2018)
Fault Diagnosis and Maintenance
 A New Data Analytics Framework
Emphasising Pre-processing in Learning
AI Models for Complex Manufacturing
Systems

Caoimhe M. Carbery1,2(B) , Roger Woods1 , and Adele H. Marshall2

1
Electronic Computer Engineering, Queen’s University Belfast, Belfast, UK
{ccarbery02,r.woods}@qub.ac.uk
2
Mathematical Sciences Research Centre, Queen’s University Belfast, Belfast, UK
[email protected]

Abstract. Recent emphasis has been placed on improving the processes

in manufacturing by employing early detection or fault prediction within
production lines. Whilst companies are increasingly including sensors to
record observations and measurements, this brings challenges in inter-
pretation as standard approaches for artificial intelligence (AI) do not
highlight the presence of unknown relationships. To address this, we pro-
pose a new data analytics framework for predicting faults in a large-scale
manufacturing system and validate it using a publicly available Bosch
manufacturing dataset with a focus on pre-processing of the data.

1 Introduction

Manufacturing is highly competitive and companies have made considerable

investments to improve their production analysis capabilities by adding sen-
sors to record information as products undergo manufacture [1]. The abundance
of data from multiple sources and in different formats that is monitored con-
tinuously, creates challenge for analysis. Moreover, the data may not have been
recorded properly, resulting in missing data which has the potential of severely
impacting subsequent modeling systems and biasing results. Missingness can be
due to faults in a machine or sensor, occurrence of noise during processing, power
shortages, or some other issues [2]. In addition, imbalanced classes can occur as
a result of under-represented classes, such as in binary classification where there
can be a majority and minority class. Research has highlighted the problems in
assessing classifiers as errors result in inaccuracy as the system is biased towards
the majority class. The work here focuses on binary classification where we want
to determine whether a product will be grouped into the minority (failure) class
dependent on the input parameters.

c Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 169–179, 2018.
https://doi.org/10.1007/978-981-13-2384-3_16
170 C. M. Carbery et al.

Review of Relevant Literature: The data in modern manufacturing chal-

lenges can suffer from high dimensionality, complexity, non-linearity and incon-
sistencies [1–4]. To address these challenges, machine learning and data analytics
methods have been employed [2–5], which concentrate on predictive maintenance
and rare event prediction [6].
Lee et al. [1] presented a cyber-physical system with a case study as an
approach to monitor the behaviour of machines using sensor data for Industry
4.0; they also highlighted the need for further work to improve generalisability
of the system. Susto et al. [3] presented a new multiple classifier model for
predictive maintenance along with a simulation study and benchmark dataset;
the data needed to be pre-processed in order to allow for a suitable classifier such
as k-NN and support vector machines (SVM), to be trained. In [5], different
ML methods were compared for a semiconductor manufacturing dataset and
highlighted the benefits of reducing the data through feature selection. Work in
[7] emphasised the benefits of AI and ML for manufacturing, but there has been
little investigation into the statistical basis of data preprocessing to improve
model performance and learning procedures. Kotsiantis et al. [8] has shown the
major impact that inefficient data can have on machine learning models.

Motivation and Overview: with increasing complexity in manufacturing pro-

cesses, machine learning algorithms are being used to ensure earlier detection
of defects, improve production performance and prediction of future perfor-
mance [5]. A framework is presented that collates, pre-processes and generates
training data for manufacturing and allows behaviour to be identified that can
influence the production outcome. We present an approach that allows AI sys-
tems to be built for behavioural analysis and information extraction to be per-
formed which will help engineers to improve machine performance and aid future
decision making
This paper is organised as follows; Sect. 2 outlines the framework for
analysing manufacturing systems which can be applied to large, inconsistent,
imbalanced datasets. Section 3 presents the Bosch case study used to test the
framework and create results. Section 4 presents the results of our algorithm and
covers conclusions.

2 Manufacturing Data Framework

In this research, we focus on the crucial stages of pre-processing, selecting suit-

able algorithms and interpretation of the results to generate a suitable method
for providing feedback to manufacturing engineers. Our framework shown in
Fig. 1, involves four key stages to produce an appropriate learning model [9].
Raw, real-world data which is unstructured and inconsistent in nature, often
involves large dimensions, class imbalance issues and missing instances. There-
fore its collation is often challenging and involves combining data from different
sources e.g. from sensors of varying machines etc. and processing it appropriately.
 A New Data Analytics Framework Emphasising Pre-processing 171

Fig. 1. Visualisation of proposed data analysis design flow for manufacturing data

2.1 Exploratory Data Analysis and Pre-processing

Exploratory data analysis (EDA) and pre-processing are crucial stages in prepar-
ing data for AI algorithms [8]. Manufacturing data can contain a large amount
of redundant information which if blindly fed into a learning model, can result
in a biased or unreliable outcome. Pre-processing can have a critical impact on
model performance, therefore effort has been spent on standard approaches e.g.
filtering and normalising to ensure that the training dataset is of an appropriate
format whilst ensuring that no bias has been introduced.

Fig. 2. Detail of the EDA and pre-processing steps required in the framework

Data cleaning removes redundant and unsuitable variables by investigating

feature variance and removing those which have near to zero variance as these
would not provide useful information to the model build and would only com-
plicate the learning process. Incomplete data is unavoidable but we must try to
have reasoning behind our choice for handling missingness whilst trying to not
influence our model. We can choose from a number of methods which aim to
handle missing data [10]. The most common are:
1. Remove any instance with at least one unknown variable value;
2. Mean substitution;
3. Treat missing values as a unique value;
4. K-nearest neighbour imputation method.
For our analysis, there is too much missingness to merely discard these instances,
so the chosen approach, as demonstrated by other researchers [6], is to rescale
the data and assign the missing instance an unique value, or if the data is of a
discrete/categorical format, assign the missing instances an independent group.
172 C. M. Carbery et al.

To improve model performance and reliability, two feature selection

approaches are performed to reduce the number of variables, namely wrapper
and embedded methods both of which produce results to indicate the features
that are most influential and important. The features retained must encode as
much information about the system as possible in order for the final classifier
model to perform well. As it is clear that a reduced feature count will improve
both performance and accuracy, we want to ensure that we do not remove any
features that could be influential to the model outcome [8].
Wrapper methods use a classifier model and conduct an extensive search in
the space of subsets of features to determine optimal performance to produce
a ranking of features. Often they are superior to filtering approaches, yet they
require a larger amount of computation as they involve investigating a large
search space. Embedded methods, however, can be seen as a balance between
the two approaches as they use internal information of the model. Thus, we
have implemented an embedded method to identify the key features for our
learning model namely an extreme gradient boosting tree (XGBoost) and thus
can determine the most influential and important features for building a suitable
classifier [11]. XGBoost generates importance measures based on the number of
occasions that a feature is selected for splitting trees in the algorithm.
Sampling methods are used if the chosen AI model cannot handle imbalanced
data and re-sample the data to either increase the minority class or else reduce
majority class [10]. Imbalanced data is prevalent in cases of anomaly detection
or rare events, where some ML algorithms could provide biased and inaccurate
results. This is a result of ML algorithms aiming to improve accuracy and not
considering the distribution of the class variable. The most common approaches
are the synthetic minority over-sampling technique (SMOTE) or ensemble meth-
ods which combine weak learners to create stronger learning models.

2.2 Learning Model

Extreme gradient boosting classification trees have the ability to not only
uncover important data features, but to construct a robust classification model.
It is a popular choice among classification models due to its simple implementa-
tion [11]. XGBoost involves the construction of an ensemble of multiple weaker
trees i.e. small trees. In order to utilise XGBoost models, the data must be in
a numeric format. In the first instance of feature selection (see Sect. 2.1), now
we can run the learning algorithm on these important variables with multiple
iterations to generate a powerful classification model.

2.3 Model Validation

The suitability of an intelligent classification model can be evaluated using stan-
dard statistical metrics such as accuracy, sensitivity, specificity, precision and
F-measure [12]. We utilised our model and testing dataset to produce values
to asses the suitability of our chosen classifier and assess its performance. We
calculated the number of correctly classified positive samples (true positives),
 A New Data Analytics Framework Emphasising Pre-processing 173

number of correctly recognised as not in the class (true negative), count of sam-
ples that were incorrectly assigned a class (false positive) and those who were
not recognised as being in the correct class (false negatives), each denoted by
tp, tn, f p, f n respectively [12]. These are used to construct confusion matrices
which provide values that are used for calculating the common performance
measures to evaluate classification models, for this paper, binary classification
(Table 1).

Table 1. General format of confusion matrix.

True Value
Positive Negative Total
Predictive Positive tp fp tp + f n
Negative f n tn f n + tn
Total tp + f n f p + tn N

The measures are highlighted as follows:

– Accuracy: indicating overall effectiveness of a classifier, calculated using the
formulae tp+tn
N but is biased when class imbalance is not addressed;
– Sensitivity and specificity analysis provide values to evaluate the effectiveness
of the classifier to identify positive and negative labels respectively, and are
given by Sensitivity = tp+f tp
n and Specif icity = f p+tn ;
tn

– Precision is a measure of a class agreement of the data labels with the clas-
tp
sifiers output labels, calculated by tp+f p;
precision.sensitivity
– F-measure is calculated by 2 precision+sensitivity and is more robust to imbal-
anced data [12].

3 Bosch Manufacturing Case Study

Bosch provided a large anonymised dataset representing one of their production
lines with an aim of utilising methods to try to predict the outcome of products
and is available on Kaggle [13]. This dataset is one of the largest publicly avail-
able manufacturing datasets (14.3 Gb), containing approximately 1.2 million
observations and over 4,000 features. The only information provided is the man-
ufacturing line and station associated with each feature which is contained within
the variable names e.g. L1 S24 F1695 indicates that Feature 1695 was observed
at Station 24 on Line 1.
The datasets were split into three categories; date, categorical and numeric.
Within each of these groups, Bosch have provided the data separated for training
and testing thus avoiding in this case, the third stage in Fig. 1. The training sets
contain the variable Response where a value of 1 indicates a product has failed
quality control, and 0 otherwise. No response variable is included in the test
174 C. M. Carbery et al.

dataset as this is the value that our model aims to predict. The quality of the
products is extremely high as only 0.58% of products fail at the final testing
stage, thus introducing a major class imbalance issue with the data. Figure 3
depicts an example of the flow of a product across the factory floor, highlighting
the numerous stations associated with different lines in the build.

Fig. 3. Example flow of Bosch factory floor depicting stations as circular nodes.

Table 2. Overview of data used for the analysis of Bosch manufacturing.

Data characteristics Total

Variables 986
Rows 1183747
Lines 4
Stations 51
Percentage missing 78.5
Percentage fail 0.58

3.1 Exploratory Analysis

In the first instance, we perform EDA to identify key properties of the Bosch
dataset to identify correlation, redundant variables, underlying structure and
issues within the data.

Data Properties: The Bosch manufacturing dataset consists of over 2.4M jobs,
each of which have an associated ID and 4364 variables. These variables/features
represent either numeric, categorical or date measurements. We performed anal-
ysis to determine the proportion of missing observations per feature and also a
count of missing observations per ID. Initial investigation into the categorical
 A New Data Analytics Framework Emphasising Pre-processing 175

features indicate an issue of extreme sparsity (around 99% missing) and thus
is not included in this paper as done in [6]. Our analysis has focused on the
numeric data as preliminaries found it to be most influential, therefore categor-
ical and date variables were not within the scope of this study. Table 2 provides
a summary of the dataset used for the research in this paper.

Fig. 4. Bar plot showing the number of features associated with the individual stations
of the Bosch production line.

Alongside the properties of the dataset content, a number of other character-

istics should be noted prior to any analysis. The chosen processing stages and
algorithms must be able to account for each of these challenges if we are to
appropriately model the data without introducing bias.

– As the data is anonymised, no expert knowledge can be employed to indicate

the higher importance features and learning is fully data-driven.
– Missing observations represents up a large proportion of the data and could
be where a product may not pass through a particular station.
– No information is related to each ID, so we could postulate that the manu-
facturing process involves a number of different products where they may not
undergo the same processing steps.
– As the data set is large, any learning procedure must have the capabilities of
processing the data of this scale.
– High class imbalance is present within the response variable as only 0.58% of
products fail at the final testing stage.
176 C. M. Carbery et al.

Fig. 5. Visualisation of missing versus observed instances within the Bosch data

Figure 4 shows the count of numeric features associated with each station. Sta-
tions 24, 25, 30 and 31 contain the largest number of features, so we assume that
these stations process more products and could be more influential.

3.2 Pre-processing
Initial analysis was performed to check for outliers in the features through visual-
isations of the distributions. Correlations between features as well as the response
were calculated. This demonstrated that features from the later stages of the
build were more highly correlated than those from earlier in the process. The
class imbalance is high, therefore if this is not handled appropriately, any model
built with this data will result in a biased approach predicting that the product
to be in the majority class i.e. pass. Before implementing sampling methods to
handle class imbalance, a number of stages of preprocessing are necessary.

Data Cleaning: Duplicated rows were removed as they provided no further

information. Variances for each feature were calculated allowing removal of
redundant features with zero variance. Our feature count reduced to 158. Whilst
this reduces the dimensionality of a dataset, the relation of the features with the
response variable can also be investigated. Figure 5 shows missing data observa-
tions in the dataset where the lighter shaded portion represents missing data. It
is clear that the later stages are where more information is recorded and would
appear in the final model. This needs to be accounted for and our approach was
to create an independent category for when an instance was unobserved, by per-
forming discretization on the 158 features to include another factor representing
unobserved instances.

Feature Selection: Feature selection allows selection of key influential vari-

ables which influence the outcome whilst improving the predictive accuracy and
 A New Data Analytics Framework Emphasising Pre-processing 177

improving interpretability. Here we used the top 50 features indicated from the
algorithm and their associated observations to train a new XGBoost classifier
model. Table 3 shows an example of the ‘Gain’ values produced by XGBoost.

Sampling: To account for the extreme class imbalance, one must consider sam-
pling methods to rebalance the class variable, but investigation into the XGBoost
algorithm demonstrated its robustness to imbalanced data and was not per-
formed for this initial analysis. However, sampling methods must be considered
for our general data analysis framework when implementing alternative learning
algorithms (Fig. 6).

Table 3. Example of six variables from XGBoost which show the accuracy of model
gained by retaining these features.

Feature Gain
L1 S24 F1723 0.5328070
L1 S24 F1846 0.2248599
L1 S24 F1632 0.1162531
L1 S24 F1695 0.0611954
L3 S34 F3876 0.0403588
L2 S26 F3036 0.0132253

Fig. 6. Output from XGBoost showing the top 15 features of importance

178 C. M. Carbery et al.

4 Conclusion
In this paper, a new framework has been presented which combines useful ana-
lytics tools into a different format from those previously implemented. Using a
widely available dataset from Bosch, an appropriate training dataset contain-
ing 50 features was produced, allowing an extreme gradient boosting tree to
be used as a classification prediction model. Using the framework (Fig. 2), data
preprocessing and exploratory analysis was used to create a reduced data size
highlighting the most influential features. This allowed us to perform an R imple-
mentation of an extreme gradient boosting (XGBoost) model [11] and employ
R’s inbuilt performance metrics to demonstrate a high accuracy and F-measure.
The research highlights the necessity to pre-process and we are currently work-
ing with a commercial partner to apply the research to their system to produce
an automated system for performing the analysis on manufacturing data.

References
1. Lee, J., Kao, H., Yang, S.: Service innovation and smart analytics for industry 4.0
and big data environment. Proc. CIRP 16, 3–8 (2014)
2. He, Q.P., Wang, J.: Statistical process monitoring as a big data analytics tool for
smart manufacturing. J. Proc. Control 67, 35–43 (2017). https://doi.org/10.1016/
j.jprocont.2017.06.012
3. Susto, G., Schirru, A., Pampuri, S., McLoone, S., Beghi, A.: Machine learning for
predictive maintenance: a multiple classifier approach. IEEE Trans. Industr. Inf.
11(3), 812–820 (2015)
4. Wuest, T., Weimer, D., Irgens, C., Thoben, K.D.: Machine learning in manufactur-
ing: advantages, challenges, and applications. J. Prod. Manufact. Res. 4(1), 23–45
(2016). https://doi.org/10.1080/21693277.2016.1192517
5. Moldovan, D., Cioara, T., Anghel, I., Salomie, I.: Machine learning for sensor-based
manufacturing processes. In: Intelligent Computer Communication and Processing
(ICPP) 2017, pp. 147–154. IEEE (2017)
6. Zhang, D., Xu, B., Wood, J.: Predict failures in production lines. In: IEEE Inter-
national Conference on Big Data, pp. 2070–2074. Washington, USA (2016)
7. Lee, K., Cheon, S., Kim, C.: A convolutional neural network for fault classification
and diagnosis in semiconductor manufacturing processes. IEEE Trans. Semicond.
Manufact. 25(5), 1167–1180 (2014)
8. Kotsiantis, S.B., Kanellopoulos, D., Pintelas, P.E.: Data preprocessing for super-
vised learning. Int. J. Comput. Sci. 1(2), 111–117 (2006)
9. Carbery, C.M., Woods, R. Marshall, A.H.: A Bayesian network based learning
system for modelling faults in large-scale manufacturing. In: IEEE International
Conference on Industrial Technology 2018, pp. 1357–1362. France (2018)
10. Chawla, N.V.: Data mining for imbalanced datasets: an overview. In: Maimon,
O., Rokach, L. (eds.) Data Mining and Knowledge Discovery Handbook. Springer,
Boston (2009). https://doi.org/10.1007/978-0-387-09823-4 45
 A New Data Analytics Framework Emphasising Pre-processing 179

11. Chen, T., Guestrin, C.: Xgboost: a scalable tree boosting system. In: 22nd Inter-
national Proceedings on knowledge discovery and data mining, pp. 785–794. ACM
(2016)
12. Sokolova, M., Lapalme, G.: A systematic analysis of performance measures for
classification tasks. Inf. Proc. Manag. 45(4), 427–437 (2009)
13. Kaggle.com: Bosch production line performance (2016). https://www.kaggle.com/
c/bosch-production-line-performance. Accessed Nov 2017
 Performance Assessment of Multivariate
Control System Based on Data-Driven
Covariance Historical Benchmark

Hong Qian1,2, Gaofeng Jiang1(&), and Yuan Yuan1

1
Shanghai University of Electric Power, Changyang Road 2588, 200090
Shanghai, China
{qianhong.sh,godfery_Jiang,yuanyuan_shdl}@163.com
2
Shanghai Power Station Automation Technology Key Laboratory, Shangda
Road 99, 200072 Shanghai, China

Abstract. In order to better meet the on-site monitoring requirements and

ensure the control system is running well, the performance assessment for
multivariate control system (MIMO) is an increasing concern. According to the
generalized eigenvalue analysis, multiple performance indexes based on the
eigenvalues was proposed, and big data of process output was used to obtain a
history data set and establish a historical benchmark, which can not only
monitor the adjustment ability of single closed-loop, but also get the change of
the overall control performance. A user-definable indicator is put forward to
select the historical data set, then, based on the statistical analysis theory, the
benchmark is improved and multiple performance levels are divided. Finally, the
performance evaluation method is presented for MIMO. An example for mon-
itoring the performance of the coordination control system (CCS) is illustrated to
show the use and effectiveness and objectivity of the proposed method.

Keywords: Performance assessment
Data-driven

Multivariate control system
Historical benchmark
Statistical analysis theory
Performance levels

1 Introduction

Modern industrial process controllers can meet design requirements in the early run-
ning, but without regular maintenance, the controller performance will decline with
time and eventually lead to a huge gap [1]. At present, for MIMO systems on site, such
as CCS, the control performance can only be judged by manual analysis, which is
difficult to adapt to the requirements of power production automation. It is also nec-
essary to realize the timely detection of performance degradation, provide guidance for
the optimization. Therefore, it is an urgent need to implement automatic and effective
online performance assessment and monitoring of a MIMO.
Research on control performance was first proposed by Harris [2], which shows the
construction of performance benchmark is the key to evaluation index. Depending on
the benchmark, performance evaluation methods can be divided into two types [3]. One
is a method based on mathematical modeling. The performance benchmark of this

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 180–190, 2018.
https://doi.org/10.1007/978-981-13-2384-3_17
 Performance Assessment of Multivariate Control System Based 181

method is mostly from some of the best controllers designed for theoretical goals. Such
methods as IMC [4], MVC [5, 6] and LQG [7] are relatively mature, but the monitoring
effect depends on the accuracy of the process model. The other is data-driven evalu-
ation methods, such as based on multivariate covariance and principal component
analysis statistics [8]. These methods have strong versatility and do not rely on process
models. Yu and Qin propose a statistical method for the MIMO process based on
historical data [9, 10], in industrial applications, the availability of benchmarks has
been validated [11]. However, a practical problem is that the selection of benchmark
data for this method requires a certain amount of expert knowledge [12]. Based on
data-driven technology, this paper presents a performance evaluation method based on
historical benchmarks for MIMO. According to the actual operation of the system, only
I/O operation data is needed and it is performed under the closed-loop conditions. It
does not require in-depth understanding of the system and does not hinder the normal
operation of the industrial process. In order to overcome the limitations of historical
benchmarks relying on expert experience, a user-definable indicator was proposed to
select historical data sets and perform distribution fitting to obtain a range of index
values divided by performance levels.
This paper is organized as follows. In Sect. 2, the performance method based on
data-driven is proposed. In Sect. 3, an experiment is illustrated to show the rationality
of this method. Finally, the conclusions are given in Sect. 4.

2 Proposed Method
2.1 The Framework of the Proposed Method
In this section, a flowchart of the proposed method is given in Fig. 1. Data mining
technology is used to analyze a large number of process data off-line, obtain the
benchmark value corresponding to the historical data sets, calculate the current value
Jact of the system operation online, and compare with the benchmark Jhist to obtain the
performance indexes, and further multiple performance levels are divided. Finally, a
comprehensive analysis of the performance of MIMO is performed.

2.2 Performance Indexes Based on Historical Benchmark for MIMO

The most common approach to loop monitoring is to compare the performance of the
monitored loop with that of a benchmark [13], based on historical benchmark, the
performance index can be expressed as

Jhist
cðkÞ ¼ : ð1Þ
Jact

For a MIMO process, let the benchmark period be I and the monitored period II, to
assess the overall control performance during period II, a covariance index similar to
that proposed by Mcnabb and Qin [14] as follows:
182 H. Qian et al.

Testing data The actual performance J act

The historical Performance

Benchmark J hist indexes

Data mining Performance

Training data
algorithm grades

Evaluation result

Fig. 1. Flowchart of the proposed method

jcovðyI Þj
Iv ¼ : ð2Þ
jcovðyII Þj

The covariance matrix used to characterize performance can be transformed into

generalized eigenvalues [9, 15], which is a quantitative indicator. Since MIMO system
with q controlled variates has

covðyI ÞP ¼ covðyII ÞPK: ð3Þ

  
Where K ¼ diag k1 ; k2 ;


; kq and P¼ p1 ; p2 ;


; pq .
By Eqs. (2) and (3), then we have

jcovðyI ÞjjPj ¼ jcovðyII ÞjjPjjKj: ð4Þ

Noticing that P is non-singular [9], we further deduced the overall performance

index of the control system:

jcovðyI Þj Y q
Iv ¼ ¼ jKj ¼ ki : ð5Þ
jcovðyII Þj i¼1

Besides, the generalized eigenvalue analysis can be also simplified as follow:

 
cov1 ðyII Þ
covðyI Þ p ¼ kp: ð6Þ

According to the mathematical definition of the eigenvalue, it has:

covðyI ÞPI ¼ KI PI : ð7Þ

covðyII ÞPII ¼ KII PII : ð8Þ

 Performance Assessment of Multivariate Control System Based 183



Where KI ¼ diag kI;1 ; kI;2 ;


; kI;q and KII ¼ diag kII;1 ; kII;2 ;


; kII;q ,kI;i are
eigenvalues during period I corresponding to a controlled variate, and kII;i are eigen-
values during period II.
Comparing Eqs. (4), (7) and (8) can yield:
2k 3
I;1
kII;1
6 7
6 kI;2 7
6 kII;2 7
K¼ KI K1
II ¼6 .. 7: ð9Þ
6 7
4 . 5
kI;q
kII;q

So we get the single closed-loop performance index:

kI;i
Ii ¼ki ¼ : ð10Þ
kII;i

2.3 Historical Data Set Selection Rules and Performance Classification

Combining with practical applications, the historical data set to be selected is a series of
continuous process data, and at the same time, the indicator does not show a large
jump. Therefore, we have proposed an indicator that can be customized according to
different needs. It is used to select a historical data set and defines it as follow:

Yq X
q
aðCÞ¼x0 k þ
i¼1 C;i
xi kC;i : ð11Þ
i¼1

Where C is a certain operating phase, kC;i are the eigenvalues correspondingly. The
weight coefficient x0 represents the importance of overall control performance,
x0 ; x1 ;


xq represents the importance of each closed-loop control performance.
In this paper, a large amount of historical process data is used to traverse all
historical stages according to Eq. (11) for off-line calculations to obtain index values
aðCÞ for any stage and regard it as a random variate, denoted as X. Through mathe-
matical statistics analysis, we can the probability distribution of X and further fit the
distribution to f ð xÞ, take multiple data sets corresponding to multiple stages corre-
sponding to the peak value as historical data sets, in that way, the average value of the
eigenvalue values in the multiple stages corresponding to the historical data sets serves
as the evaluation benchmark, i.e.k;1 ; k;2 ;


; k;q .
After fitting, in most cases, X is possible to obey the lognormal distribution, thus
Y ¼ lnðXÞ  Nðl; r2 Þ, the probability density function expression of X is
( h i
pffiffiffiffi
1
exp  2r1 2 ðln x  lÞ2 x[0
f ðx; l; rÞ¼ 2prx : ð12Þ
0 x0
184 H. Qian et al.

When y ¼ lnðxÞ ¼ l, which is x ¼ el , fmax ðx; l; rÞ exists. As there are M stages in

the history data set, the benchmark is

1X M
k;i ¼ kk;i : ð13Þ
M k¼1

When y ¼ lnðxÞ ¼ l þ pr, which is x ¼ el þ pr , corresponding to N data sets, there

are

1X N
kl þ pr;i ¼ kpk;i : ð14Þ
N k¼1

Then the improved single closed-loop control performance indexes are

k;i
I i ¼ ki ¼ : ð15Þ
kII;i

The improved overall control performance index can be expressed by

Yq Yq k;i
Iv ¼ i¼1
ki ¼ i¼1 k
: ð16Þ
II;i

In practical work, Y obeys the normal distribution and has 3r criteria, so the
lognormal distribution X has

Pðelr  x  el þ r Þ¼68:27%
Pðel2r  x  el þ 2r Þ¼ 95:45% ð17Þ
Pðel3r  x  el þ 3r Þ¼ 99:73%:

Therefore, the range of indexes can be divided into 5 intervals in the direction of
ð0; 1Þ, corresponding to 5 grades, respectively poor, fair, good, nice and great, as
shown in Table 1. In the end the rating criteria is summarized as follows

Table 1. Performance grades

Item Great Nice Good Fair Poor
Ii ð1; 1Þ ð k; i ; 1 k; i
ðkl þ 2r; i ;
k; i k; i k; i k; i
kl þ r; i kl þ r; i  ðkl þ 3r; i ; kl þ 2r; i  ð0; kl þ 3r; i 
Iv ð1; 1Þ Qq Qq Q
q Qq Q q Q
q
k; i k; i k; i k; i k; i k; i
ðQ
q
i¼1
; 1 ðQ
q
i¼1
;Q
q
i¼1
 ðQ
q
i¼1
;Q
q
i¼1
 ð0; Q
q
i¼1

kl þ r; i kl þ 2r;i kl þ r; i kl þ 3r; i kl þ 2r; i kl þ 3r; i
i¼1 i¼1 i¼1 i¼1 i¼1 i¼1

The performance of MIMO is a comprehensive consideration of single closed-loop

indexes and the overall index. Therefore, If Ii and Iv are all great, the controller of
MIMO system works well, on the contrary, when they all are poor, the adjustment
effect is difficult to meet the control requirements.
 Performance Assessment of Multivariate Control System Based 185

3 Experiment and Discussion

The process output data from a 600 MW coal-fired heating unit of a thermal power
plant under normal operation is used as the data source of the experiment. The dataset
contains the set-point and output values of the two controlled variates (output power
and main steam pressure). Then, an example for monitoring the performance of CCS is
illustrated in this section.

3.1 Performance Indexes for Coordinated Control System

Generally, coordinated control system of the thermal power unit is a tow input tow
output control object. From Eqs. (15) and (16), the control performance indexes are

k;N
I N ¼ kN ¼ : ð18Þ
kII;N

k;P
IP ¼ kP ¼ : ð19Þ
kII;P

k;N k;P
I v ¼ kN kP ¼ : ð20Þ
kII;N kII;P

Where IN and IP is the single closed-loop performance index of power and main
steam pressure, respectively, Iv represents the overall performance of the system.

3.2 Experimental Procedure

Due to the large data, only 6 h were shown below. The output power (MW) was in
Fig. 2. The main steam pressure (MPa) was in Fig. 3.

Fig. 2. The set-point and output of electric power

186 H. Qian et al.

Fig. 3. The set-point and output of main steam pressure

Fig. 4. Statistical summary report of a

By using Eq. (11) ðx0 ¼ 1:5; x1 ¼ 2 and x2 ¼ 2Þ to perform offline calculations

on multiple historical phases (1 h), the corresponding index values for 2 857 phases are
obtained. Through data analysis software MINITAB, the frequency of a was counted
and fitted to a lognormal distribution a  f ðx; 1:431; 0:620Þ, showed in Fig. 4.
Hypothesis test by using Anderson-Darling shows P ¼ 0:107 [ 0:05 in Fig. 5, it meets
the null hypothesis and the distribution is established. Finally, we can get the bench-
mark k;1 ¼ 0:0511 and k;2 ¼ 2:1403. The performance levels are shown in Table 2.

3.3 Results and Analysis

The control performance change of 12 h on December 22 was monitored using the field
operation data. Finally, the dynamic changes of the performance indexes for a single
controlled quantity (output power, main steam pressure) are shown in Figs. 6 and 7, the
overall system performance changes as shown in Fig. 8. Pick four times t1, t2, t3, and
t4 during performance monitoring, as shown in Fig. 9, we can get the corresponding
index value, showed in Table 3.
 Performance Assessment of Multivariate Control System Based 187

Fig. 5. Anderson-darling hypothesis test

Table 2. Classification of the coordinated control performance

Item Great Nice Good Fair Poor
IN ð1; 1Þ (0.592, 1] (0.406, 0.592] (0.229, 0.406] (0, 0.229]
IP ð1; 1Þ (0.568, 1] (0.307, 0.568] (0.164, 0.307] (0, 0.164]
Iv ð1; 1Þ (0.339, 1] (0.157,0.339] (0.054,0.157] (0,0.054]

Fig. 6. Single closed-loop performance index of electric power control

The simulation results show that the electric power control performance is great for
more than 360 min, and is still nice for a long time, and the performance is good and
fair in a short time within 200 min to 400 min. The main steam pressure control
performance is great or nice most of the time, nice state time is longer, and performance
is good and general in a short time. The overall control performance is great or nice
most of time and it is good or fair in a very short time.
At time t1, IN ,IP and Iv are all greater than 1, control quality of the system better
than the benchmark, and is in the great state; At t2, the control performance is general,
which needs to be improved. There is a severe attenuation trend in t1 to t2, which
188 H. Qian et al.

Fig. 7. Single closed-loop performance index of main steam pressure control

Fig. 8. Overall performance index

Fig. 9. The dynamic change curve of performance indexes

should be taken seriously. Adjustment quality of the system is nice at t3; it is good at
t4. The situation does not occur, that is, the power, main steam pressure, and overall
control performance are all in poor, which means that the system is out of control.
 Performance Assessment of Multivariate Control System Based 189

Table 3. Performance index values at t1–t4

Item IN IP Iv
t1 1.904 2.593 4.937
T2 0.325 0.265 0.086
T3 0.790 0.927 0.732
T4 0.616 0.560 0.345

4 Conclusion

This paper proposes evaluation indexes for MIMO based on data-driven technology
combined with the actual operation of the control system. It can not only obtain the
control performance of a single closed-loop, but also get the overall performance of
MIMO. And a rule for selecting historical sets is also proposed, by which an improved
historical benchmark is obtained, and uses distribution fitting to divide the control
performance into five levels. The performance evaluation of a 600 MW unit coordi-
nated control system is simulated, and the objectivity and rationality of the method are
verified. For the characteristics and control quality requirements, this method can be
applied to a large number of multivariate control systems in the field. It can realize
online automatic evaluation to monitor the change of control quality, quickly respond
to severe performance degradation, and provide guidance for the optimization of
control system.

References
1. Desborough, L., Miller, R.: Increasing customer value of industrial control performance
monitoring-honeywell’s experience.In: AIChE Symposium Series, pp. 169–189 (2002)
2. Harris, T.J.: Assessment of control loop performance. Can. J. Chem. Eng. 67(5), 856–861
(1989)
3. Xin, Q., Yang, C., et al.: A review of control loop monitoring and diagnosis: prospects of
controller maintenance in big data era. Chin. J. Chem. Eng. 24(8), 952–962 (2016)
4. Zhang, G.M., Li, N., Li, S.Y.: A data driven performance monitoring method for predictive
controller. J. Shanghai Jiao Tong Univ. 45(8), 1113–1118 (2011)
5. Zhang, Q., Li, S.Y.: Performance monitoring and diagnosis of multivariate model predictive
control using statistical analysis. Chin. J. Chem. Eng. 14(2), 207–215 (2006)
6. Wang, X., Huang, B., Chen, T.: Multirate minimum variance control design and control
performance assessment: a data-driven subspace approach. IEEE Trans. Control Syst.
Technol. 15(1), 65–74 (2006)
7. Bernstein, D.S., Haddad, W.M.: LQG control with an H∞ performance bound: a Riccati
equation approach. In: American Control Conference, pp. 796–802. IEEE (2009)
8. Alghazzawi, A., Lennox, B.: Model predictive control monitoring using multivariate
statistics. J. Process Control 19(2), 314–327 (2009)
9. Yu, J., Qin, S.J.: Statistical MIMO controller performance monitoring, Part I: data-driven
covariance benchmark. J. Process Control 18(3–4), 277–296 (2008)
10. Yu, J., Qin, S.J.: Statistical MIMO controller performance monitoring, part II: Performance
diagnosis. J. Process Control 18(3–4), 297–316 (2008)
190 H. Qian et al.

11. Li, Q., Whiteley, J.R., Rhinehart, R.R.: A relative performance monitor for process
controllers. Int. J. Adapt. Control Sig. Process. 17(7–9), 685–708 (2010)
12. Wang, L., Li, N.: Performance monitoring of the data-driven subspace predictive control
systems based on historical objective function benchmark. Acta Autom. Sin. 39(5), 542–547
(2013)
13. Desborough, L., Harris, T.: Performance assessment measures for univariate feedback
control. Can. J. Chem. Eng. 70(6), 1186–1197 (1992)
14. Mcnabb, C.A., Qin, S.J.: Projection based MIMO control performance monitoring: I—
covariance monitoring in state space. J. Process Control 13(8), 739–757 (2003)
15. Yuan, Q., Lennox, B., Mcewan, M.: Analysis of multivariate control performance
assessment techniques. J. Process Control 19(5), 751–760 (2009)
 Quantitative Safety Assessment Method
of Industrial Control System Based
on Reduction Factor

Haoxiang Zhu1(&), Jingqi Fu1, Weihua Bao2, and Zhengming Gao2

1
Department of Automation, College of Mechatronics Engineering and
Automation, Shanghai University, No.149,Yanchang Rd, 200072 Shanghai,
China
[email protected], [email protected]
2
Shanghai Automation Instrumentation Co Ltd, Shanghai, China

Abstract. Information security assessment is important to the smooth and

stable operation of industrial control system (ICS), and provides valuable
advices for security policy and measurement. On the basis of analyzing the
safety risk assessment standards of industrial control system, this paper puts
forward a method of risk assessment for industrial control system based on
fuzzy analytic hierarchy process, and creatively introduces fuzzy consistent
matrix and entropy method to overcome the lack of fuzziness in the evaluation
result of traditional analytic hierarchy process. This method can effectively
assess the importance of system assets, the severity of vulnerability and the
threats it faces. Through the above steps, the risk of the system is comprehen-
sively evaluated.

Keywords: Industrial control system
Risk assessment
Entropy method

Vulnerability

1 Introduction

With the arrival of intelligent manufacturing and the deep integration of industrial-
ization and information technology [1], in view of the increasing security threat and
attack of industrial control system, the network security risk assessment is becoming
more and more concerned as a method to recognize the risk of industrial control system
[2] in advance. It has become a new hot spot in the research of industrial control
network security [3].
Most of the industrial enterprises in our country have not established the infor-
mation security management system of industrial control system, the safety manage-
ment system is not perfect, and the establishment of the safety management system [4]
is lack of basis and guidance. With the continuous development of the Internet, the
proportion of information security risk is becoming more and more serious. Therefore,
it is necessary to develop the relevant standards for safety management, put forward the
specific process of safety management [5], and put forward the safety requirements of
the industrial control system from many aspects such as information security and

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 191–201, 2018.
https://doi.org/10.1007/978-981-13-2384-3_18
192 H. Zhu et al.

functional safety, and guide the industrial enterprises to establish the suitable condi-
tions according to their own actual conditions [6].
In the field of industrial control security, most of them are the combination of
analytic hierarchy process and normalization. In this paper, the fuzzy matrix method
and entropy method are adopted to reduce the consistency test [7]. The reduction factor
can also improve the accuracy of the final evaluation. It is a feasible method.

2 Factors Analysis of Industrial Control Security Assessment

2.1 Asset Identification

First, an object asset library is needed to create, which can not be limited to the
hardware and software of the system. The assets of a complete control system should
include information, data, hardware, software, communication facilities, personnel, and
manufactured products. Generally speaking, the assets of industrial control network can
be divided into physical assets, logical assets and personnel assets.
Destroying the availability, integrity and confidentiality of information assets is the
main way to affect the system, and most of the system losses also arise. In the process
of assigning assets, the value of assets can be expressed qualitatively or quantitatively,
and some systems can describe the value of assets in the process of risk analysis by
quantitative analysis.
The final assignment is assigned to a level, and asset classification shows a com-
prehensive description of the importance of different levels. The higher the level is, the
more important the assets are, the evaluators can determine the scope of the important
assets [8] according to the results of the assets, then the next risk assessment is carried
out through the important assets.

2.2 Vulnerability Identification

The vulnerability of industrial control system can be divided into technical vulnera-
bility and management vulnerability according to different attributes. For a long time,
industrial control systems have been considered as isolated and isolated systems, and
information security is not the first issue to be considered. This makes the industrial
control system have many vulnerabilities in the technology of [9] platform and net-
work. In the industrial field, the awareness of information security and protection is
weak, and a large part of the industrial control system lacks effective safety manage-
ment. It is mainly manifested in the imperfect safety management system, the unclear
operation process of the safety equipment, the lack of effective supervision, inspection
of the security risks faced by the system, the lack of training for the personnel, the
backward emergency plan and so on.The degree of vulnerability of the identification is
classified to the degree of membership, and then classify the severity of vulnerability
according to the difficulty of vulnerability being exploited.
 Quantitative Safety Assessment Method of Industrial Control 193

2.3 Threat Recognition

Industrial control system threats mainly come from external environment and human
factors. The external environment threats are mainly manifested in the physical envi-
ronment effects such as hardware and software failures, network attacks and serious
natural disasters, such as earthquakes and fires, and human factors are mainly mani-
fested in poor management, operation error, unauthorized access and operation.
The two indicators for assigning threats are the frequency of threats and the impact
on assets. The frequency of the earthquake threat is very low, but it is very destructive
to the system, and the frequency of network sniffing is very high, but most of the effects
on the system can be ignored.
After the identification of assets, vulnerabilities, threats and existing security
measures, according to certain rules, an appropriate method is adopted to assign the
identified risk factors. The risk values of the organization are related to the importance
of the assets, the severity of the vulnerability and the intensity of the threat.

3 Safety Quantitative Analysis

3.1 Level Analysis Method

A hierarchical structure model is established. On the basis of in-depth analysis of
practical problems, various factors are decomposed from top to bottom into several
levels according to different attributes. The factors of the same layer are influenced by
the factors belonging to the upper layer or on the upper factors, and at the same time,
the factors of the next layer or the role of the lower factors are dominated. The top level
is the target layer, usually has only 1 factors, the lowest level is usually the plan or the
object layer, and there is one or several levels in the middle, it is usually the criterion or
the index layer. When there are too many criteria (for example, more than 9), the sub
criterion level should be further decomposed.
The weight vector is calculated and the consistency check is carried out. For each
pair of paired comparison matrices, consistency index is used, random consistency
index and consistency ratio to test the consistency of the maximum characteristic root
and corresponding eigenvector. If the test is succeed, the feature vector (normalized) is
the weight vector; if it is not succeed, the paired comparison matrix needs to be
reconstructed.
Calculate the combined weight vector and check the consistency. The lowest
combination weight vector is calculated to tend to the target., and do the combination
consistency check according to the formula. If the test is succeed, the decision can be
made according to the result expressed by the combination weight vector. Otherwise,
the model or the pair comparison matrix with large consistency ratio need to be
reconsidered.
The analytic hierarchy process is used to decompose the industrial control system
into several levels, construct the information security risk hierarchy structure model,
and establish multiple evaluation items. The factors of safety evaluation are divided
into target level, criterion level and index level. The work of pre-risk assessment
includes asset identification, vulnerability identification and threat identification. It’s
194 H. Zhu et al.

necessary to transform the complex multi factor problem into the contrast problem. By
comparing the factors, the fuzzy complementary [10] judgement matrix constructed by
the expert is converted into a fuzzy consistent judgment matrix [11].

3.2 Level Analysis Method

The judgement matrix of the analytic hierarchy process needs many times when the
consistency is poor. The fuzzy complementary judgement matrix constructed is con-
verted by the expert to the fuzzy consistent judgment matrix. It overcomes the defect
that the subjectivity of the judgement matrix is too strong in the analytic hierarchy
process, and the difference between consistency [12] and matrix consistency, the dif-
ficulty and lack of consistency check.
Judgment matrix:
2 x2 xn 3
1 x1


x1
6 .. 7
6 x1


. 7
6 1 7
Rðrij Þnn ¼ 6 x2 7
6 .. .. 7
4 . . 5
x1
xn


1

The comparison of the two factors is divided into 6 grades: the same, important,
relatively important, obviously important, and absolutely important. It’s natural to use
digital 0.1-0.9 to quantify, as shown in Table 1.

Table 1. Numerical meaning table

rij Meaning
0.5 The two factors are equally important
0.6 Compared with the two elements, the former is slightly more important
0.7 Compared with the two elements, the former is obviously important
0.8 Compared with the two elements, the former is strongly important
0.9 Compared with the two elements, the former is extremely important
0.1–0.4 Contrary to the above

In order to determine the relative membership degree of a single evaluation index,

the fuzzy evaluation matrix is used to eliminate the dimensional effect of each eval-
uation index to make modeling universal.

R ¼ ðrij Þnn ð1Þ

The fuzzy complementary judgement matrix is calculated by line and recorded as:
 Quantitative Safety Assessment Method of Industrial Control 195

X
n
ri ¼ rik ð2Þ
k

The mathematical transformation is applied to the fact:

ri  rj
fij ¼ þ 0:5 ð3Þ
2n

The records are as follows:

RM ¼ ðfij Þnn ð4Þ

The consistency condition of the judgement matrix is not completely satisfied, and
it is objective or even impossible to eliminate in practice.
Use the test index:

kmax  n
CI ¼ ð5Þ
n1

When less than 0.1, the judgement matrix conforms to the consistency standard
[13], and the result of hierarchical single ranking is acceptable. Otherwise,CIwill be
revised until it meets the consistency standard.kmax is the maximum eigenvalue of a
fuzzy judgment matrix.

3.3 Element Normalization

When the value of a certain index of each evaluation object [14] is different, the
entropy value is smaller, which indicates that the effective information provided by the
index is larger and its weight should be larger. On the contrary, if the difference of a
certain index value is small and the entropy value is larger, the amount of information
provided by the index is smaller and the weight of the index should be smaller. When
the values of each target are the same, the entropy is maximum, which means that the
index has no useful information, and can be removed from the evaluation index system.
Since the measurement units of each index are not unified, it is necessary to
standardize them before they are used to calculate the comprehensive indexes [16],
which means that the absolute value of the index is converted to the relative value.
 
fij ¼ fij  ð6Þ

In order to determine the fuzzy membership matrix of the relative membership

degree of a single evaluation index, the dimensional effect of each evaluation index is
eliminated, so that the modeling is universal.

RM ¼ ðfij Þnn ð7Þ

196 H. Zhu et al.

The matrix RM is used as the original matrix, and linear normalization of it can be
got, so as to solve the homogenization problem of different quality indicators. More-
over, because the meaning of the positive and negative index values is different (The
higher the positive index is, the lower the negative index value is.). Therefore, different
algorithms is used to standardize the data for high and low index. The specific methods
are as follows:
For positive indicators:

fij  minðfij Þ
j
lij ¼ ð8Þ
maxðfij Þ  minðfij Þ
j j

For negative indicators:

 
max fij  fij
lij ¼ ð9Þ
maxðfij Þ  minðfij Þ
j j

The entropy of the i index is:

X
n
hi ¼ k hij ln hij ð10Þ
j¼1

lij
hij ¼ P
n ð11Þ
hij
j¼1

k is defined as a super parameter and can be adjusted according to the actual

situation. After calculating the entropy of the i, the entropy weight of the i can be
obtained through the following formula:

1  hi X m
xi ¼ P
m ð xi ¼ 1Þ ð12Þ
m  hi 1
i¼1

Wi ¼ ðx1 ; x2 . . .. . .xn Þ ð13Þ

The result of the evaluation:

Bi ¼ Wi  RAi ð14Þ
 Quantitative Safety Assessment Method of Industrial Control 197

3.4 Anti Model Gelatinization Method

According to the negative rule idea, the weight of each part in the industrial control
system is obtained by using the reduction factor inverse model gelatinization method,
and the risk of the whole system is quantified through it.
The result of fuzzy comprehensive evaluation is a fuzzy vector which belongs to the
risk comment set {high, high, general, low, low}. Then, based on the barycenter
method, the reduction factor is used to improve the fuzzification method. The factor is
derived from the negative rules and li is called the reduction factor. li is the reduction of
the normal weight.
Reduction factors make up a reduction vector l .
Using the reduction factor li to change the normal weight in the anti-gelatinization
process:

P
n
ti xi
j¼1
B1 ¼ Pn ð15Þ
xi
j¼1

P
n
ti ðxi Þa
j¼1
B2 ¼ n ð16Þ
P
ðxi Þa
j¼1

a is a compromise parameter that is distributed in the interval½1; þ 1.

ti takes the intermediate values of the grade of membership degree vector of the
evaluation grade.

si ¼ aTi
qi ð17Þ

ai is the unit amount of each element, qi is the reduction, If it’s not zero, ti was cut.
To ensure that the weight after reduction is not negative,si  ð0; 1Þ.
The evaluation score of the upper layer is obtained by the formula 17. Parameter Bi
is the weight of the various factors of the I, Si is the evaluation score for the I factor. In
this way, the total score of the system can be finally obtained.

4 Case Analysis

The network risk assessment is carried on to a specific power system. First, the
evaluation object is analyzed, and the evaluation target is established by the analysis.
Aiming at the total risk target of the system, it is refined into multiple sub targets, and
establish a hierarchical structure model of industrial control system with multi-
objective, multi-level and multi criteria layer by improved analytic hierarchy process.
The level structure model is divided into four layers: the target layer is the security risk
198 H. Zhu et al.

of the enterprise control system; the first level index divides the factors affecting the
security risk of the control system into three aspects: assets, vulnerabilities and threats;
the second index is the identification of the three risk factors and subdivides according
to the attributes of each category of risk factors. In order to refine the mapping rela-
tionship between target level and target level, the second level index is further refined
into families, forming the underlying factor layer. Finally, the hierarchical structure
model is determined.
Specifically, the relative importance of each factor index is divided according to the
quantitative scaling method of fuzzy matrix priority relation, and the fuzzy judgment
matrix is constructed.
Referring to the assignment structure, the numerical results is given in the form of
matrix, and get the corresponding fuzzy judgement matrixRP1 ,RP2 ,RP3 ,RP4 andRP5 .
2 3 2 3
0:5 0:7 0:8 0:9   0:5 0:5 0:2 0:3
6 0:3 0:5 0:6 0:8 7 6 0:5 0:4 7
RP1 ¼6 7; RP2 ¼ 0:5 0:3
; RP3 ¼6
0:5 0:3 7
4 0:2 0:4 0:5 0:6 5 0:7 0:5 4 0:8 0:7 0:5 0:6 5
0:1 0:2 0:4 0:5 0:7 0:6 0:4 0:5

2 3
0:5 0:2 0:3 0:4 0:6
6 0:8 0:5 0:4 0:4 0:7 7
6 7
RP4 ¼6
6 0:7 0:6 0:5 0:5 0:8 7
7
4 0:6 0:6 0:5 0:5 0:6 5
0:4 0:3 0:2 0:4 0:5
2 3
0:5 0:4 0:7 0:6 0:6 0:5 0:8 0:4 0:2 0:8 0:9
6 0:6 0:5 0:7 0:8 0:7 0:6 0:9 0:3 0:2 0:5 0:6 7
6 7
6 0:3 0:3 0:5 0:4 0:3 0:4 0:6 0:2 0:3 0:7 0:8 7
6 7
6 0:4 0:2 0:3 0:5 0:4 0:4 0:7 0:5 0:5 0:6 0:7 7
6 7
6 0:4 0:3 0:7 0:6 0:5 0:6 0:8 0:5 0:5 0:7 0:8 7
6 7
RP5 ¼6
6 0:5 0:4 0:6 0:6 0:4 0:5 0:7 0:5 0:5 0:6 0:8 7
7
6 0:2 0:1 0:4 0:3 0:2 0:3 0:5 0:6 0:7 0:5 0:5 7
6 7
6 0:6 0:7 0:8 0:5 0:5 0:5 0:5 0:5 0:7 0:8 0:8 7
6 7
6 0:8 0:8 0:7 0:5 0:5 0:5 0:3 0:3 0:5 0:2 0:1 7
6 7
4 0:1 0:5 0:3 0:4 0:3 0:4 0:5 0:2 0:8 0:5 0:5 5
0:2 0:4 0:2 0:3 0:2 0:2 0:5 0:2 0:9 0:5 0:5

According to formula 8 and 9, each matrix is normalized and transform the fuzzy
complementary judgement matrix RPi into fuzzy consistent matrix RMPi .i ¼ 1; 2. . .5.
According to formula 10, the entropy weight xi of the i index is calculated. Finally, the
weight of each factor bi is calculated based on formula 11. The overall framework is
shown in Fig. 1.
 Quantitative Safety Assessment Method of Industrial Control 199

Fig. 1. Distribution of asset importance

It can be known that the weights of each index are as follows:

xP1 ¼ ð 0:405 0:313 0:275 0:007 Þ

xP2 ¼ ð 0:375 0:625 Þ

xP3 ¼ ð 0:224 0:315 0:337 0:124 Þ

xP4 ¼ ð 0:012 0:121 0:221 0:485 0:161 Þ

xP5 ¼ ð 0:135 0:151 0:081 0:094 0:117 0:108 0:071 0:111 0:103 0:013 0:017 Þ

Finally, according to the comprehensive evaluation method: B ¼ x  R, the final

distribution of the importance of assets can be got. As shown in Fig. 2. Effective safety
measures adopted by industrial control systems can reduce the possibility of vulnera-
bility to threats. Effective security measures can reduce the possibility of risk occur-
rence and reduce the damage caused by security events, while invalid security
measures may further create vulnerabilities and provide an intrusion approach to
threats. It is found that although the deployed security performance is good, there are
still some fragile points without any security measures. This is the main approach to the
external threat of intrusion. For example, some commonly used industrial protocols use
plaintext transmission, and the common control equipment lacks the encryption
authentication mechanism and network when communicating with the host computer.
The opening of the port is not standard and so on.
200 H. Zhu et al.

Fig. 2. Risk distribution diagram

5 Conclusion

In this paper, an improved fuzzy analytic hierarchy process is applied to assess the risk
of information security in specific power systems. The results of the inverse model
gelatinization are improved by entropy method and improved reduction factor, effec-
tively assessing the importance of the system assets, the severity of the vulnerability
and the threats faced, and evaluating the system risk synthetically, thus verifying the
effectiveness and practicability of the proposed method. At the same time, it provides a
good theoretical basis for the risk management of the industrial control system, and
further enriches the theory of risk assessment of industrial control system. It has high
theoretical value and practical value for the establishment of the strengthening scheme
of industrial control system and the improvement of the risk management system.
In the comprehensive fuzzy evaluation of the relative importance of each element
and the factors of the evaluation layer, it is one of the key points of the future research
to design a better evaluation standard to deal with the evaluation data better and reduce
the subjective uncertainty.

Acknowledgments. This work was financially supported by the Science and Technology
Commission of Shanghai Municipality of China under Grant (No. 17511107002).

References
1. Peng, y, Jiang, C., Xie, F.: Progress in information security of industrial control systems.
J. Tsinghua Univ. (Nat. Sci. Ed.) 52(10), 1396–1408 (2012)
2. Peng, J., Liu, L.: Information security analysis of industrial control systems. Autom. Instrum.
33(12), 36–39 (2012)
3. Chin, Z.: Safety and management of industrial network control system. Meas. Control
Technol. 32(2), 87–92 (2013)
 Quantitative Safety Assessment Method of Industrial Control 201

4. Xiong, Q., Peng, Y., Dai, Z.: Preliminary study on information security risk assessment of
industrial control system. Chin. Inf. Secur. 27(3), 57–59 (2012)
5. Zhu, X.: Information security risk assessment risk analysis method. Inf. Secur. Technol. 28
(8), 87–89 (2010)
6. Jia, C., Feng, D.: Safety assessment of industrial control system based on fuzzy analytic
hierarchy process. J. Zhejiang Univ. 50(4), 759–765 (2016)
7. Qin, C., Chen, X., Yang, Y.: Research and application of three level assessment on the safety
of tailing bank based on FAHP IE algorithm. Control Proj. 21(6) 995–1000 (2014)
8. Lu, H., Chen, D., Peng, Y.: Quantitative research on information security risk assessment of
industrial control system. Autom. Instrum. 35(10), 21–25 (2014)
9. Yuan, Y.: Research on risk management of construction projects based on fuzzy analytic
hierarchy process (FAHP). Chongqing University, Chongqing (2013)
10. Ma, L., Zhang, L., Yang, Y., et al.: Research on the security risk assessment of information
systems based on fuzzy neural network. J. Chin. Secur. Sci. 22(005), 164–169 (2012)
11. Li, J., Yin, X., Hutu, et al.: Research on network security assessment based on extension
theory. Comput. Eng. Appl. 48(21), 79–82 (2012)
12. Krutz, R.L.: Securing SCADA Systems. Wiley, Hoboken (2005)
13. Fan Hong, Feng Dengguo, Wu Yafei. Methods and applications of information security risk
assessment. Tsinghua University press, Beijing 2006 49–50
14. Kefeng, F., Ruikang, Z., Lin, L.: Research on information security standard system of
industrial control system. Top. Netw. Secur. Stand. 6 (4), 17–21 (2016)
15. Xiaofeng, Z., Xiuzhen, C.: Grey level information security evaluation model. Ind. Control
Syst. 1(6), 15–20 (2014)
16. Nicholson, A., Webber, S., Dyer, S., et al.: SCADA security in the light of cyber-warfare.
Comput. Secur. 31(4), 418–436 (2012)
Intelligent Computing in Robotics
 The Electric Field Analysis and Test
Experiments of Split Type Insulator Detection
Robot

Pengxiang Yin1,2,3(&), Xiao Hong1,2,3, Lei Zheng1,2,3, Biwu Yan1,2,3,

and Hao Luo1,2,3
1
NARI Group Corporation Ltd., Nanjing 211106, China
[email protected]
2
Wuhan NARI Limited Liability Company, State Grid Electric Power Research
Institute, Wuhan 430074, China
3
Hubei Key Laboratory of Power Grid Lightning Risk Prevention, Wuhan
430074, China

Abstract. The insulator strings of Extra High Voltage (EHV) and Ultra High
Voltage (UHV) transmission lines are longer. It is more convenient and effective
to detect insulators with robot. Currently, insulator detection robots mainly work
by climbing along the insulator string, the structure is bulky and complex, and
needs further improvement. Therefore, this paper proposes a split insulator
detection robot which is suitable for the detection of suspended insulator string.
The electric field distribution around the insulator robot is simulated and ana-
lyzed. The test experiment of 220 kV insulator string is carried out. The results
indicate that the maximum electric field strength around the split type robot is
1753 kV/m. Partial discharge will not be generated. The measured potential
value of the insulator detection robot is less than the original voltage distribution
value of the insulator. The actual voltage distribution of insulators can be
obtained by compensation and correction, then the detection of low and zero
insulators is carried out.

Keywords: Split type robot
Insulator detection
Electric field analysis

Test experiments

1 Introduction

Transmission line insulators work in outdoor environment. Due to severe weather or

industrial pollution, they may suffer from erosion of dust, natural hydrochloric acid,
haze, acid rain, industrial waste gas and other pollutants [1, 2]. As time goes by, the
insulation performance decreases. To ensure the safe operation of transmission line
insulators, preventive inspection should be carried out regularly. In addition, insulators
should be detected before live working such as replacing insulators and electric power
fittings on transmission towers [3, 4]. The disc porcelain insulator string need to be
tested for low/zero state, the detection methods can be classified as contact type and
non-contact type [5, 6]. At present, the non-contact type detection is mainly carried out
by using infrared or ultraviolet detection. The detected temperature difference of the

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 205–214, 2018.
https://doi.org/10.1007/978-981-13-2384-3_19
206 P. Yin et al.

insulator is small through infrared detection method, while the ultraviolet detection
performance is easily affected by the environment of the insulators [7, 8]. Both methods
have their own limitations, the data analysis methods need further studied in order to
improve the detection accuracy [9]. The data of contact detection is more reliable.
Traditional detection methods mainly rely on spark fork method or voltage distribution
method by manual work to distinguish the condition of insulators, which are very
inconvenient for long string insulators detection because the operating lever is usually
too long [10, 11]. With the rapid development of EHV and UHV transmission lines, the
insulator strings are too long to use manual detection methods [12, 13]. Insulator
detection by robot is convenient and effective for long insulator strings, and also
reduces the manpower investment in insulator detection work, reducing the operation
and maintenance costs of transmission lines [14, 15].
At present, insulator detection robot moves up and down along the suspension
insulator string mainly by climbing. Electric field distribution of each part of the
insulator detection robot is simulated and analyzed by researchers. Simulation results
show that the robot has a certain influence on electric field distribution of the insulator
string, but does not affect its insulation performance [16, 17]. The electric field strength
of the robot can be reduced through structural improvement and optimization, it can
avoid interference to the detection system and the malfunction of robot by corona
discharge and suspended potential discharge [3, 18]. However, due to the combination
of power and detection module on robot body, resulting in large volume, heavy weight
and complex structure of the robot body.
This paper designs a kind of split insulator detection robot for the suspension string
of the disc porcelain insulator in the straight-line tower, which has simple structure,
light weight and small volume. By simulating the electric field of the robot body and
the insulator string when the robot moves along insulator string, researchers study the
influence of the robot on insulator string and the largest field strength of each structure
of robot body. The test of 220 kV transmission line insulator string voltage detection
by designed insulator robot is carried out, the results of voltage distribution mea-
surement are analyzed.

2 Operating Principle of Split Type Insulator Detection

Robot

The structure of split insulators low/zero live detection robot is showed in Fig. 1. It
contains two parts: power module and detection module. The power module and the
detection module are connected by a thin insulation rope, which can be dissembled and
installed flexibly. The detection module transmits instructions to the power module
through WIFI communication network, which controls the detection module to move
up and down and stop, then insulator detection according to the predetermined pro-
cedures is accomplished.
The power module includes cable drive motor, cable rack, guide pulley and fixed
plate. The driving motor is consisted of control chip, WIFI communication module,
motor driver, rotate speed sensor, grinding motor, power management unit and the
battery. The structure of the cable rack and guide pulley are simple and small in size, it
 The Electric Field Analysis and Test Experiments 207

Fig. 1. Structure of the split type insulator detection robot.

is easy to carry onto the tower. The connecting plate can be fixed on the cross arm
stably due to the structure of the cross arm, the connecting part is metal, so as to ensure
that the power module and the tower body are equipotential.
The detection module includes distance sensor, frame, stepper motor, probe, guide
stripes and WIFI communication module. All modules are installed on the frame and
are dismountable. The guide stripes can effectively prevent the insulators detection
robot from being stuck by the insulator’s shed while moving along the insulator string.
The detection module identifies the steel cap, shed and steel foot of the insulator
through the pre-supposed algorithm by using the range sensor. When the range sensor
recognizes the steel cap, the communication module sends the stop instruction to the
cable drive motor. When the detection module stops at the detection position, stepper
motor controls the two probes to touch the adjacent two insulators’ steel caps and
measure the voltage of the insulators. Then the voltage data can be transmitted to the
hand-held terminal through the communication module, and draw the distribution
voltage curve on the terminal which is used to judge low/zero value insulators.

3 Electric Field Analysis of Insulator Detection Robot

Because the power module of the insulator detection robot is installed on the cross arm
of the tower and is equipotential with respect to the arm, the influence of the electric
field on power module is negligible. But the detection module may change the dis-
tribution of the surrounding electric field while moving, and the module body may also
produce partial discharge due to the high local field strength, which needs simulation
on the electric field for further analysis. Take the disc type porcelain insulator string on
the 220 kV single-loop cathead tower as simulation object, the electric field distribution
of the robot’s body, insulators and gold fittings when the insulator detection robot
208 P. Yin et al.

moves at different positions along the 220 kV disc type porcelain insulator string are
analyzed.

3.1 Establishment, Subdivision and Loading of 3D Model

The 3D model includes the tower-line system model and the insulator robot model, as
shown in Fig. 2. The tower model adopts 220 kV straight line cathead tower, it height
42 m, and tower size according to drawing. Insulator string type is XP-70. The height
of single insulator is 146 mm, the disc diameter is 255 mm, the creepage distance is
280 mm, and one insulator string is composed of 15 insulators. Bundled conductor is
considered (vertically arranged). The length of the conductor model is 20 meters, the
split spacing is 400 mm, and the diameter of the sub conductor is 21.66 mm. In order
to subdivision the model, it is necessary to build a wrapped air bag outside the
conductor.

Fig. 2. The 3D model of the tower-line system and the insulator robot.

For the insulator robot, detection part including probe, stepper motor, connecting
plate, rack, guide stripes, communication module are built in the model. The probe is
divided into front metal and insulated connecting rod, the length of the front section is
120 mm, the length of the connecting rod is 130 mm, the inner radius of the rack is
140 mm, the thickness is 5 mm, the height is 20 mm, the guide stripes is 200 mm long
and 20 mm wide, and the radius of the circular arc is 20 mm; The size of the
communication module is 30  40  100 (mm), and the step motor size is
30  40  45 (mm). For the convenience of model subdivision, the model of the
detection part is simplified and the components with less influence on the electric field
simulation results are removed.
Then the whole model is subdivided, air bags are added respectively above the
insulator string and the tower in order to control their density. The insulator string and
the robot get encryption subdivision. The subdivision diagram of the insulator robot is
shown in Fig. 3 for example. At last, potential is loaded and the maximum field
pffiffi
intensity is studied. The maximum voltage Um ¼ 220 pffiffi 2 ¼ 179:629 kV is loaded in the
3
phase of the insulator robot, the rest phases are loaded with 0:5Um according to phase
 The Electric Field Analysis and Test Experiments 209

sequence. The insulator is located at the high voltage end, the high potential Um is
loaded on the steel foot and wire connected to it, and 0 potentials are loaded on the low
voltage end which is connected with metal fittings, the iron cap, the pole tower, the
ground and the outsourced air boundary. The steel foot and iron cap with unknown
potential in the calculation model are suspended potential conductors. Coupled free
degree method is applied, which means the connected steel feet and steel cap are
coupled at equal potential. The probe of the insulator robot also needs to be coupled
when measuring the equipotential of the steel cap connected with measuring device.

Fig. 3. Mesh of insulator detection robot.

3.2 Simulation Results and Analysis

Finite element method is used to simulate the static field and to number the insulators
starts from the high voltage end. Taking the third insulator for example, the influence of
insulator on the surface strength of insulator strings and the distribution characteristics
of the electric field of the insulators robot are analyzed. The field strength comparison
of the surface of insulator strings of the third insulation with and without robot is
shown in Fig. 4. It indicates that although the insulator robot has a certain influence on
the local field strength of the third and the fourth insulators, the overall change is little,
and the maximum field intensity change of the surface of the insulator is only about
1%, which means that the effect of insulator robot on surface field strength of insulators
is small.
When measuring the third insulator, the electric field strength on the rack and guide
stripes of the insulator, as shown in Fig. 5, is about 240 kV/m, which is much less than
the halo field strength of the air. The field strength at the probe and connecting rod is
shown in Fig. 6, with a maximum value of about 1372 kV/m, which is less than that of
the air. By comparing and analyzing the maximum field strength of the probe, the field
strength of the insulator robot is larger at both ends of the insulator string, and the probe
field intensity is the largest near the high end, while the insulator robot’s probe field
strength is smaller in the middle of the insulator. This is the same with the voltage
distribution characteristics of insulators, because the insulators at both ends bear a
higher voltage and the insulators in the middle bear a lower voltage, so the voltage
210 P. Yin et al.

a) Without robot b）Robot measuring No. 3 insulator

Fig. 4. The influence of detection robot on the electric field intensity of insulator surface.

Fig. 5. The electric field distribution of the rack.

difference between two probes is different. The larger the voltage difference is, the
greater the field strength is. And this also verifies the effectiveness of the calculation.
The electric field simulation shows that the insulator robot has a certain influence
on the distribution of the electric field of the insulators, but it has little influence on the
maximum field strength of them, so this influence can be ignored. The maximum field
strength of the insulator is about 1753 kV/m. At the probe and the insulated connecting
rod, the radius of the probe wire is 0.3 cm. The halo field strength is about 37.6 kV/cm
calculated with the Peek formula, and the field strength of the probe surface is
17.53 kV/cm, which is a large margin. So the corona discharge will not occur.
 The Electric Field Analysis and Test Experiments 211

Fig. 6. The electric field distribution of probe and connecting rod.

Fig. 7. Field layout of the test.

4 Test and Analysis of 220 KV Insulator String

The 220 kV insulator string was used for testing in the laboratory. The insulator string
has 13 insulators, the field layout is shown in Fig. 7. By using the iron tower to
simulate the tower cross arm, the iron frame is grounded through copper wire, the
lowest fittings are connected to a high-voltage generator, and the AC voltage effective
value is 127 kV. The insulator detection robot is set to measure twice of each insulator
when working.
212 P. Yin et al.

As there are some differences between the environment in the laboratory and for the
actual tower, the difference of the distribution voltage in the industry standard between
the measured value and the standard value is quite different. The simulation data with
the measured data are compared in Table 1. Firstly, the sum of voltage of 13 insulators
is 110.03 kV, which is about 17 kV less than the applied voltage. This result is caused
by the capacitance of the probe and the impact of the measurement circuit on the
distribution voltage. When the insulator robot is measuring the insulator, it is equiv-
alent to a parallel branch, resulting in the reduction of the equivalent impedance, and
the distribution voltage is reduced. When the robot moves along the insulator string, the
measured values are smaller than that of the original insulators, so the recorded total
voltage is about 17 kV less than the sum of the applied voltages.

Table 1. Comparison of measurement and simulation data.

No. The simulation The simulation Measured The difference of measured
data without data with value/kV value and simulation data with
robot/kV robot/kV robot/kV
1 29.71 26.26 24.75 1.51
2 20.27 17.16 15.68 1.48
3 15.03 12.85 11.70 1.15
4 11.47 10.01 9.24 0.77
5 9.02 7.87 7.78 0.09
6 7.32 6.39 6.73 −0.34
7 6.09 5.52 5.96 −0.44
8 5.28 4.87 5.42 −0.55
9 4.73 4.26 4.72 −0.46
10 4.45 4.02 4.46 −0.44
11 4.38 3.94 4.33 −0.39
12 4.53 4.11 4.49 −0.38
13 4.71 4.31 4.77 −0.46
SUM 126.99 111.47 110.03 1.44

The robot and the insulator string forms a equivalent parallel circuit structurally.
The configuration of this equivalent parallel circuit is constantly varying during
measurement, and the measured value from the detection robot has a complex non-
linear relationship with the original voltage distribution.
The whole model of insulator robot detection is established and verified by sim-
ulation. Firstly, the three-dimensional model according to the test field layout is set up,
distribution parameters are simulated according to the finite element method in Sect. 2,
and the equivalent branch of the insulator robot is added to the road model. The
insulators at high potential are added to the equivalent branch of the robot to calculate
the voltage values at both ends respectively, and the insulators without the robot branch
are calculated. Results of voltage distribution is shown in Table 1. When the robot is
close to one certain insulator, the voltage distribution of the simulation is different from
the original insulator voltage distribution, the maximum single voltage difference
 The Electric Field Analysis and Test Experiments 213

appears on the first insulator, the difference is 3.45 kV, which is about 13.14%. The
overall difference of the whole series voltage is about 15.5 kV. When the robot is close
to insulators, the maximum difference between the simulated data and the measured
data is about 1.51 kV, and the difference is about 6.1%.
Because the insulator model is idealized, the material parameters are different from
data of the actual insulators, and the measurement error is considered at the same time,
the above error values can meet engineering requirements. It shows that the insulator
detection robot can effectively measure the insulator distribution voltage. However, due
to the difference between the total value of measurement and the actual voltage value, it
is not convenient for the staff to analyze. The voltage compensation method can be
studied so that the measurement results reflect the original voltage distribution of the
insulators as accurate as possible.

5 Conclusion

This paper designed a split type insulator detection robot which is applicable for
suspension insulator string detection, and carried out electric field analysis and test
study of insulator robot detection. Conclusions are as follows:
(1) The insulator robot has some influence on the local electric field distribution of
insulators, but it has little influence on the maximum field strength of insulators
which can be neglected. The maximum field strength of the insulator robot itself is
less than halo field strength of the air. The same research method can be used to
detect the low and zero insulators of the suspension porcelain insulator strings of
500 kV and 1000 kV transmission lines by further improving the insulation of the
insulators.
(2) The insulator robot can accurately and effectively measure the distribution voltage
of insulators. However, because of the influence of the probe and measuring cir-
cuit, the measurement value is different from the actual voltage distribution of the
insulator. The measured results should be compensated.
(3) The insulator robot can detect the distribution voltage effectively under the sim-
ulated environment with low and zero voltage value insulators. The distinguishing
and analysis of low and zero insulators can be carried out with the help of the
measured voltage distribution curve.

Acknowledgement. Fund Project: Technology project of State Power Grid Co., Ltd.
(No. 524625170046)

References
1. Zhibin, Q., Jiangjun, R., Daochun, H., et al.: Study on aging modes and test of transmission
line porcelain suspension insulators. High Volt. Eng. 42(4), 1259–1267 (2016)
2. Xian, Z., Richeng, L., Ming, Z., et al.: Online detection device of low or zero value
insulators in UHVDC Transmission line. Electr. Power 49(6), 90–94 (2016)
214 P. Yin et al.

3. Juan, J., Guanbin, W., Shiyou, M., et al.: The electric field analysis and optimal design on a
robot for insulator detection. Insul. Surge Arresters 2, 180–185 (2017)
4. Yuntu, J., Jun, H., Jian, D., et al.: The identification and diagnosis of self-blast defects of
glass insulators based on multi-feature fusion. Elect. Power 50(5), 52–58 (2017)
5. Feng, W., Hongcai, L., Xiang, P., et al.: Optimization research on condition-based
maintenance of overhead transmission line with online detection without power cut. Elect.
Power 49(10), 84–89 (2016)
6. Haipeng, S., Weiguo, L., Zhanzhan, Q.: Study on the on-line detection device of hand-held
insulator. Insul. Surge Arresters 1, 22–26 (2012)
7. Jiangang, Y., Ye, Z., Tangbing, L.I., et al.: Analysis of high-voltage ceramic insulators
infrared detection blind areas. High Volt. Eng. 43(9), 2903–2910 (2017)
8. Jing, C., Nan, Z., Zhong, M., et al.: Ultraviolet detection technology of 750 kV resistant to
porcelain insulator based on artificial pollution test. Elect. Power 49(9), 23–29 (2016)
9. Peng, L.: Study on Heating and Discharge Characteristics of Zero Value Insulators Based on
Infrared and Ultraviolet Imaging. North China Electric Power University, Beijing (2016)
10. Yong, F., Zongren, P., Peng, L., et al.: Voltage-sharing characteristics of porcelain insulators
for UHV AC transmission lines. High Volt. Eng. 36(1), 270–274 (2010)
11. Xiaowei, L., Linong, W., Junhua, W., et al.: Research on automatic detection technology for
faulty porcelain insulators on AC transmission lines. Electr. Meas. Instrum. 53(11), 110–115
(2016)
12. Liang, Z., Yong, L., Zhigang, R., et al.: Analysis on the influence factor of insulators
detection by robots and electric field method. Insul. Surge Arresters 2, 148–155 (2017)
13. Zhengzi, C., Weiguo, L., Wenbin, C.: Design and implementation of insulator resistance
measurement robot based on WiFi wireless remote control. Elect. Power 48(12), 23–26
(2015)
14. Rui, G., Bing, T., Lei, Z., et al.: Study on cap and pin porcelain insulator detecting
technology suitable for robots on transmission line. Insul. Surge Arresters 2, 141–147 (2017)
15. Yao, C., Dehua, Z., Jie, N.: The impact of zero and low resistance insulator on potential and
electric field distribution of long insulator strings in the 500 kV transmission line. Insul.
Surge Arresters 3, 29–34 (2015)
16. Ming, Z., Richeng, L., Jie, X., et al.: Feasibility research and develop of working detecting
robot for low/zero insulator. High Volt. Appar. 52(6), 160–166 (2016)
17. Linhua, Z., Yandong, C.: The research and design of overhead transmission line insulator
detecting robot. Electron. Des. Eng. 23(16), 164–166 (2015)
18. Tao, C., Daqing, S., Deli, Z., et al.: Research on the application of disc type porcelain
insulators detection robot in overhead transmission lines. Insul. Surge Arresters 2, 11–16
(2013)
 Study on Early Fire Behavior Detection
Method for Cable Tunnel Detection Robot

Biwu Yan1,2,3(&), Guangzhen Ren1,2,3, Xiaowei Huang1,2,3,

Junfeng Chi1,2,3, Lei Zheng1,2,3, Hao Luo1,2,3, and Pengxiang Yin1,2,3
1
NARI Group Corporation Ltd., Nanjing 211106, China
[email protected]
2
Wuhan NARI Limited Liability Company, State Grid Electric Power Research
Institute, Wuhan 430074, China
3
Hubei Key Laboratory of Power Grid Lightning Risk Prevention, Wuhan
430074, China

Abstract. With the increase of cable tunnels and improvement of inspection

requirements, different types of inspection robots have been designed and
applied to the inspection and fire protection of cable tunnels. Due to the complex
environment of the cable tunnel, it is difficult for existing robots to conduct fire
detection and diagnosis effectively. Therefore, a fire detection method of cable
tunnel robot based on support vector machine (SVM) is proposed. According to
the actual tunnel size, a simulation model is established to calculate the smoke
concentration and temperature variation of tunnels under different fire power.
Considering the dynamic detection of robot, the characteristics of the measured
data are analyzed. The SVM model is trained by using the sample characteristic
data, and then the test data is used for testing. The accuracy of the test is up to
95%.

Keywords: Split type robot
Insulator detection
Electric field analysis

Test experiments

1 Introduction

With the rapid development of cities, more and more cable tunnels have replaced
overhead lines. The power cables work under high voltage and high temperature
environment for a long time. Breakdown and short circuit may occur with insulation
aging. Further, the cable tunnel fire may be produced. In order to ensure the safe and
stable operation of power cables, monitoring and maintenance should be carried out.
Due to the narrow space and complex environment, it is very inconvenient to operate
and maintain the cable tunnel manually. Especially when a tunnel fire accident occurs,
it is not easy to find the hidden danger of the fire. Once lighting a fire, the fire will
develop rapidly along the cable line. Combustion produces a large amount of smoke,
toxic and harmful substances and high temperatures. These pose a threat to the personal
safety of the staff in the tunnel. The cable tunnel has a long distance, narrow pas-
sageways and few outlets. Rescuers are hard to get close to the source of fire.
Therefore, the cable tunnel inspection robot emerges as the times require [1–3].

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 215–224, 2018.
https://doi.org/10.1007/978-981-13-2384-3_20
216 B. Yan et al.

Intelligent robots are used for inspection work, which can effectively reduce staff
burden. Furthermore, there are fixed-point detection systems in general tunnels, but
they have monitoring blind area. As a mobile monitoring system, the tunnel robot can
effectively compensate for the shortcomings of the fixed-point detection system. The
cable tunnel robot can be divided into walking and track type [4, 5]. The walking robot
is mainly a caterpillar type. The cable tunnels may have accumulated water, debris,
large slopes and complex pipelines, which are not conducive to walking. The orbital
robot walking on the track at the top of the tunnel is not affected by the above factors.
In 2009, the Shanghai Jiao Tong University developed a lightweight, small crawler
patrol robot [6]. The upper part arranged infrared cameras, gas sensors and signal
transceivers to collect and transmit the physical environment and gas environment in
the tunnel. In 2012, the Hangzhou power company produced a wheeled robot [7]. It
carries the gas sensor and the smoke sensor, can effectively perceive the potential
danger in the tunnel, and provide the data base for the fire prevention. The fire
extinguishing device can take effective measures to extinguish the fire according to the
feedback signal of the actual sensing device. However, because the robot carries too
many equipment, its weight and size exceed the standard. In 2014, the Xiamen Electric
Power Exploration Institute studied and produced a fixed track cable tunnel inspection
system [8]. The mobile mechanism marched along the predetermined track. Two
infrared camera devices were arranged to collect the images in the tunnel and an
ultrasonic sensing device to measure the obstacles. The whole mobile mechanism also
carries several gas sensors to detect the gas environment. However, the above research
focuses on the robot’s motion and data acquisition function. There are few studies on
the tunnel fire behavior judging by the collected data. If the fire source is detected in
time, the fire can be extinguishing as soon as possible.
In this article, the simulation model is established according to the actual size of the
cable tunnel. The tunnel fire is simulated under different fire power sources, and the
data of flue gas concentration and temperature in the tunnel are obtained. Consider the
dynamic sampling data of robots, the characteristic parameters are analyzed and
extracted. A fire detection method of cable tunnel robot based on SVM is proposed.
The effectiveness of the above method is verified by simulation data.

2 Modeling and Simulation of Fire in Cable Tunnel

In order to study the early fire behavior detection method, the development law of the
tunnel fire needs to be obtained. The cable tunnel robot can measure temperature,
characteristic gas, particle concentration, etc. Unlike the fixed-point monitoring device,
the tunneling robot detects relevant data in the process of moving. It is necessary to
obtain data on temperature and smoke concentration measured by robots. However, the
fire tests are destructive and non-recoverable, and full-scale tests are expensive. A more
efficient method is as follows: first, the fire development rules are obtained by
numerical simulation, and then the simulation data are verified by typical tests. This
section will introduce the model building and simulation results.
 Study on Early Fire Behavior Detection Method 217

2.1 Modeling and Parameter Setting

According to the structure and size of an electric single cabin of integrated corridor in
Hangzhou, the 1:1 numerical model of cable tunnel is established. The length, width
and height of the cable tunnel are 200  2.3  2.3 m. There is one column on each
side of the tunnel, with three rows in each column. The height of the lowest cable to the
ground is 0.5 m, and the two rows are separated by 0.4 m. The robot is located on the
top of the middle of the tunnel. Its measuring device is 0.8 m from the top, so the probe
is set at the same height. 220 kV cables are arranged according to the actual layout. The
cable diameter is about 0.11 m. Because of the setting requirement of the calculation
software, the rectangle with equal cross section area is used for simulation. Electric
power cable is generally copper core XLPE cable. The material is allocated according
to the mass ratio of copper and plastic to 6:4. The ignition source is selected in the test
simulation, and the fire source size is installed on the cable with a size of 0.1  0.1 m.
The overall model is shown in Fig. 1. In order to calculate accurately, the mesh is more
detailed near the fire source.

Fig. 1. Fire simulation model of cable tunnel

According to the actual situation, there is ventilation equipment in the cable tunnel
under normal working conditions. In the early stage of the fire, the ventilation
equipment may be still working. Wind speed has a certain effect on the spread of the
fire. When the fire develops to a certain stage, trigger the linkage and close the fire
door. It is considered that the wind speed of the port is zero. It can delay the spread of
fire and reduce oxygen content. The two cases above are simulated respectively.
According to results of previous studies, the heat release rate of the cable material is
obtained. Combined with the shape and size of the cable in this article, the maximum
heat release rate of different numbers of cables is calculated. The fire power
218 B. Yan et al.

corresponding to different fire sizes is set to 1467 kW/m2, 2200 kW/m2 and
3385 kW/m2, respectively. The fire source is set as the T2 fire growth model. The
source of the fire rises to maximum power after 200 s, and then the power remains
unchanged. In order to simulate the situation of fire in different layers of cable, The fire
source height is set to 0.5 m, 0.9 m, 1.3 m, respectively. Considering the wind speed,
the power and the height of the fire source, different working conditions are set up as
shown in Table 1.

Table 1. Setting of simulation conditions

No. Height of the fire source (m) Heat release power (kW/m2) Wind speed (m/s)
1 0.5 1467 1
2 0.5 2200 1
3 0.5 3385 1
4 0.9 1467 1
5 0.9 2200 1
6 0.9 3385 1
7 1.3 1467 1
8 1.3 2200 1
9 1.3 3385 1
10 0.5 1467 0
11 0.5 2200 0
12 0.5 3385 0
13 0.9 1467 0
14 0.9 2200 0
15 0.9 3385 0
16 1.3 1467 0
17 1.3 2200 0
18 1.3 3385 0

2.2 Simulation Result

(1) Analysis of flame spread

In the simulation of different conditions, comparing the conditions 2, 5 and 8 with
conditions 10, 11 and 12, it can be found that the spread rate of fire will grow with the
increase in the size of the fire source power in the same of fire source setting height and
wind speed. For example, in conditions 2, 5 and 8, the fire spread quickly in the
vertical direction while it is 77th seconds, 59th seconds, and 46th seconds respectively.
Comparing conditions 7, 8, and 9, the spread speed of fire is also related to the space
height of the fire. When the cables set at the bottom cable holder is on fire, the flame
will rise with the flowing of hot air, and quickly ignites the overhead cable. If the tunnel
is well ventilated, it will also accelerate the lateral spread of fire. When the cables in the
top is on fire firstly, during the initial stage of the fire, the flame rises up against the
ceiling, quickly forming a ceiling jet and moving horizontally along the ceiling. After
 Study on Early Fire Behavior Detection Method 219

200 s, the cables in the top are all ignited. Typical fire spread situation is shown in
Fig. 2.

Fig. 2. Typical flame spread schematic

Fig. 3. Smoke diffusion in the early stage of fire

(2) Analysis of fire smoke diffusion

It is very important for early warning of fire detection to study the diffusion of
smoke in cable tunnels. As shown in Fig. 3, during the smoldering stage of the cable,
although the cable only ignited small flames, the smoke had spread rapidly in the
tunnel. In the fire simulation of cable tunnels, it can be observed that under other
conditions unchanged, only changing the size of the fire source power, such as con-
ditions 2, 5 and 8, the smoke used 50.8 s, 40.2 s and 31.4 s respectively through the
cable tunnel with 20 m, that is, the smoke spread rate grows with the increase of the fire
source power.
Consider the effect of ventilation in the cable tunnel on the spread of smoke. When
the tunnel is ventilated, lateral wind will make the smoke downstream of fire source
pass through the tunnel model first. With growing of the fire, the smoke in the upper
reaches of the fire source will gradually increase. As in conditions 8 and 17, it takes
58.6 s and 40.6 s respectively for the smoke to pass through tunnel model with 20 m.
220 B. Yan et al.

When the tunnel is not ventilated, the smoke spreads more evenly in the tunnel, and the
ventilation of the tunnel will intensify the burning of the cable and make the smoke
accumulate faster in the tunnel.
(3) Analysis of temperature variation
At the beginning of the combustion, the temperature distribution throughout the
tunnel is relatively uniform. During the development of the fire, the temperature grows
with the rise of smoke, which causes the high temperature to appear at the top of the
tunnel and spreads rapidly over the tunnel. The temperature declines with the distance.
This makes the suspended orbit inspection robot better able to detect changes in
temperature and smoke concentration on fire than the robot walking on the ground. The
typical tunnel temperature distribution is shown in in Fig. 4.

Fig. 4. Typical tunnel temperature distribution schematic

3 Simulation Data Analysis and Fire Behavior Detection

Method

Take the following working conditions as an example: height of the fire source 0.9 m,
heat release power 3385 kW/m2 and wind speed 1 m/s. The variation of smoke con-
centration and temperature in cable tunnel fires is analyzed. The probe is arranged at
the height of the robot sensor. The smoke concentration and temperature measured by
the probe arranged in proper order from the fire source point are shown in Figs. 5 and
6. It can be seen from Fig. 5 that the temperature of the detection point above the fire
source point rises slowly due to the effect of ventilation. The temperature at the second
detection points in the downwind direction increased rapidly and the temperature value
is higher. For the turbulence produced by combustion has a great influence on the
temperature change. The temperature variation of several detection points near the fire
source point is obviously different from that of other detection points. The trend of
temperature of the detection point beyond the fire source 5 m is similar changing with
time. The difference in the rate of rise is small. But there are some differences in the
starting time of temperature rise at different locations. Therefore, when sampling
 Study on Early Fire Behavior Detection Method 221

sequence is selected for analysis, the point with a certain distance from the fire source is
more regular. It can be seen from Fig. 6 that the smog concentration curve is rather
chaotic compared with the temperature. This is mainly due to the fact that the gener-
ation and diffusion of particles are very irregular under the action of flame turbulence.
Especially when the flame is burning fiercely, it is more difficult to make use of smoke
concentration. The smoke concentration changes more smoothly at a certain distance
from the fire source and when the concentration is low. Therefore, when selecting
smoke concentration data, a lower trigger value can be set. When the smoke concen-
tration exceeds the trigger value, a certain length of data is recorded and analyzed.

Fig. 5. Temperature variation at different locations from the fire point

When the fire behavior is detected, the tunnel robot moves towards the fire source.
The data collected during the movement are different from those obtained directly by
simulation. It is necessary to analyze and process data with the speed of mobile robot.
In order to conduct fire source detection, it is necessary to determine the starting
position of the analysis data. Due to the random time and location of the fire, the robot
may begin to judge in different locations. For the temperature varies greatly with the
outside world, it is more reasonable to use smoke concentration as trigger criterion. It is
assumed that the analysis data begin to record when the smoke concentration exceeds
10%. Take the following working conditions as an example: height of the fire source
0.9 m and wind speed 1 m/s. The heat release power is 1467 kW/m2, 2200 kW/m2 and
3385 kW/m2, respectively. The smoke concentration and temperature detection value
under different heat release power and initial measurement position are shown in
Figs. 7 and 8. It shows that the sampling data sequence of smoke concentration and
temperature under different fire source power has certain rules. But it is difficult to
distinguish only through the curve.
222 B. Yan et al.

Fig. 6. Smoke concentration variation at different locations from the fire point

Fig. 7. The simulation of temperature sequence of robot measurement

From the above analysis, it can be seen that fire detection is more accurate based on
robot detection data at the early stage of fire. In this paper, a fire detection method based
on SVM is proposed. When the smoke concentration exceeds the threshold, 10 tem-
perature values of robot sampling and 10 smoke concentrations are selected to form a
characteristic vector. It is input to SVM for classification and judgement. SVM
 Study on Early Fire Behavior Detection Method 223

Fig. 8. The simulation of Smoke concentration sequence of robot measurement

classifier constructs a hyper-plane as the edge of classification decision by using

support vectors in training set. The training set is given as Formula 1.


D ¼ ðxi ; yi Þ
xi 2 RP ; yi 2 f1; 1g; i ¼ 1;


; n ð1Þ

Where: xi is a P dimension characteristic vector, yi is various classifications, n is the

total number of characteristic vector.
Support vector machines can maximize the difference between parallel hyper-
planes of two categories. Two hyper-planes can be expressed by Formula 2.

H1 =xi ; w þ b ¼ 1
ð2Þ
H2 =xi ; w þ b ¼ 1

Where: w is a normal vector of hyper-plane, b is a bias vector.

The optimization problem can be expressed by formula 3. The constraint condition
is formula 4.

1 Xn
Min kW k2 þ P e
i¼1 i
ð3Þ
2
yi ðw  /ðxi Þ þ bÞ  1  ei ; i ¼ 1; 2;


; n ð4Þ

The classification of support vector machines can be expressed by formula 5.

224 B. Yan et al.

Xl 
f ð xÞ ¼ sgn a y KðX; Xi Þ þ b
i¼1 i i
ð5Þ

Where: K a kernel function, a* and b* is the optimal coefficient.

Support vector machine classifier is constructed based on the above formula. In this
paper, there are 180 sample data, of which 120 are used as training samples and 60 as
test samples. First, training samples are used to train the SVM model, and then the test
samples are used for testing. The accuracy of the test is 91.7%. The validity of the fire
detection method is verified.

4 Conclusion

The simulation model is established according to the actual size of the cable tunnel in
this article. The tunnel fire is simulated under different fire power sources, and the data
of flue gas concentration and temperature in the tunnel are obtained. The variation law
of smoke concentration and temperature under different combustion conditions is
analyzed. Data sequence collected by the robot during a mobile process is analyzed.
The sampling data sequence of smoke concentration and temperature under different
fire source power has certain rules. The characteristic values are selected through
simulation data, and the fire detection model based on support vector machine is
trained. The accuracy of the test is 91.7%. The validity of the fire detection method is
verified.

Acknowledgments. This study was funded by the State Grid Corporation Headquarters Science
and Technology Project (524625160014).

References
1. Hui, Z., Ruiming, Q.: Mechanism design of a new kind of inspection robot for cable tunnel
and track optimization simulation. Mach. Des. Manuf. Eng. 47(3), 61–65 (2018)
2. Ingason, H., Li, Y.Z.: Model scale tunnel fire tests with longitudinal ventilation. Fire Saf. J. 45
(8), 371–384 (2010)
3. Rie, D.H., Hwang, M.W., Kim, S.J., et al.: A study of optimal vent mode for the smoke
control of subway station fire. Tunn. Undergr. Space Technol. 21(3), 300–301 (2006)
4. Victores, J.G., Martinez, S., Jardon, A., et al.: Robot-aided tunnel inspection and maintenance
system by vision and proximity sensor integration. Autom. Constr. 20(5), 629–636 (2011)
5. Li, Y.Z., Lei, B., Ingason, H.: The maximum temperature of buoyancy-driven smoke flow
beneath the ceiling in tunnel fires. Fire Saf. J. 46(4), 204–210 (2011)
6. Fu, Z., Chen, Z., Zheng, C., et al.: A cable-tunnel inspecting robot for dangerous
environment. Int. J. Adv. Robot. Syst. 5(3), 243–248 (2008)
7. Wei, H., Zhen, R.G., Ge, J., et al.: Research of patrol robot for electrical cable corridor.
Autom. Instrum. 12, 13–16 (2013)
8. Ling, F., Huaxia, Y., Dong, F.W.: Research on the moving structure of split cable tunnel
inspection robot. Process Autom. Instrum. 38(6), 46–50 (2017)
 Speech Endpoint Detection Based
on Improvement Feature and S-Transform

Lu Xunbo(&), Zhu Chunli(&), and Li Xin(&)

School of Mechatronic Engineering and Automation, Shanghai University,

Shanghai 200072, China
{xunbolu,zhuspringwinner}@163.com,
[email protected]

Abstract. In the low SNR and non-stationary noise environment, traditional

feature detection methods will lead to a sharp drop in detection performance.
This paper proposes an improved speech endpoint detection algorithm based on
S-transform (ST). ST has the advantages of both Short Fast Fourier Transform
(SFFT) and Wavelet Transform (WT). It can extract more robust MFCC fea-
tures. In this paper, the ST is combined with spectral subtraction to transform the
speech into the time-frequency joint domain in order to obtain a purer speech.
Then the dynamic threshold updating mechanism is used to detect the noisy
speech with two-parameter double threshold method. Through Matlab simula-
tion, the improved algorithm presented in this paper is compared with two other
algorithms. The experimental results reveal that this algorithm has a higher
accuracy in endpoint detection. Moreover, it has a great advantage both in
detection rate and error rate.

Keywords: Endpoint detection
S-transform
Spectral subtraction

MFCC
Uniform sub-band variance

1 Introduction

Speech enhancement technology plays an important role in speech processing such as

speech coding, recognition and endpoint detection. The speech signal is a typical non-
stationary signal, but the existing methods usually divide the signal into multiple
frames and process each frame. This kind of method causes the loss of information in
the signal, especially the key information between two frames. Therefore, it is nec-
essary to make some improvements to the traditional feature extraction methods. The
conventional Fourier transform cannot effectively analyze unsteady signals because it is
only a global transformation of a signal, and for speech signal processing, the fre-
quency domain characteristics at each moment are also very important. Therefore, it is
necessary to combine the time domain and frequency domain when processing speech
signals. The s-transform proposed by Stockwell [1] has the advantages of SFFT and
WT. The s-transform can not only preserve the phase information of the signal but also
provide the changeable time-frequency accuracy. It is a linear transformation and
suitable for processing non-stationary signals. Compared with SFFT, WT and other

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 225–235, 2018.
https://doi.org/10.1007/978-981-13-2384-3_21
226 L. Xunbo et al.

methods, ST is more intuitive in time-frequency performance, and decomposes more

meticulously in the high-frequency part, which means it can extract more features.
Meanwhile, for the problem of unsatisfactory single-feature speech endpoint
detection and waveform distortion under low SNR environment, feature parameter
improvement techniques and multi-feature parameter fusion techniques become the
focus of endpoint detection. Huang and Yang [2] used MFCC as a parameter for
judging speech and noise, which creates a new idea [3]. References [4, 5] introduce
wavelet transform (DWT) into the extraction of MFCC parameters. Literature [6]
combines MFCC cepstrum distance with logarithmic energy for endpoint detection;
literature [7] points out that the dual threshold setting of MFCC cepstrum parameters is
not suitable and proposes a feature parameter of the cosine of the MFCC; WANG
Haibin [8] proposes a novel method called MFCC based on s-transform (SMFCC) for
speech feature extraction. These studies have achieved certain improvements in the
endpoint detection under low SNR conditions, but still need to be improved in the
accuracy and stability of the detection. Therefore, this paper uses ST combined with
spectral subtraction to transform the speech into the time-frequency joint domain in
order to obtain a purer speech, then the dynamic threshold updating mechanism is used
to detect the noisy speech with two-parameter double threshold method.

2 Concept of S-Transform (ST)

S-transform is an effective time-frequency representation method, which combines the

advantages of SFFT and WT, and is suitable for the analysis of nonlinear and non-
stationary signals (such as speech signals).
For a given signal xðtÞ, its S-transform is defined as:

þZ1
Sðt; f Þ ¼ xðsÞgðt  s; f Þej2pf s ds: ð1Þ
1

s and t are time variables, f is a frequency variable and gðt  s; f Þ is a Gaussian

function as follows:

j f j ðtsÞ2 f 2
gðt  s; f Þ ¼ pffiffiffiffiffiffi e 2 : ð2Þ
2p

The S-transform is reversible and its inverse transform is:

Z þ1 Z þ1
xð t Þ ¼ fSðt; f Þdtgej2pft df : ð3Þ
1 1

From Eq. (1), we can see that ST can be regarded as a phase correction of a
continuous wavelet transform(CWT), it changes the corresponding scale parameters,
translation parameters of the wavelet transform, thus making up for the disadvantage of
phase information lacking in CWT. From Eq. (2), we can see that the window function
is characterized by both time and frequency characteristics, and the variance of the
 Speech Endpoint Detection Based on Improvement Feature and S-Transform 227

window function is equal to the inverse of the frequency. Based on these character-
istics, satisfactory results obtained if a signal is transformed into the time-frequency
domain by ST and then processed by the method of spectral subtraction.

3 SMFCC Cepstral Distance Feature Extraction

3.1 MFCC Principle

The Mel Frequency Cepstrum Coefficient (MFCC) analysis is based on the auditory
mechanism of the human ear, describes the distribution of speech signals in energy, and
can well simulate the human auditory system’s perception ability [8], which has good
recognition and robustness. It is one of the most important and effective feature
parameters in the current speech signal processing. MFCC is a cepstrum parameter
extracted in the frequency range of the Mel scale, where the Mel scale describes the
nonlinear characteristics of the human ear frequency, and the relationship between the
Mel frequency and the linear frequency can be approximated by Eq. (4), where the
frequency is f, the unit is Hz.
 
f
fmel ¼ 2595: log 1 þ : ð4Þ
700

3.2 SMFCC Feature Extraction Steps

SMFCC feature extraction steps are as follows:
(1) Preprocessing: Including pre-emphasis, frame division, and windowing functions.
The purpose of pre-emphasis is to compensate for the loss of high-frequency
components, enhance the high-frequency components. The usual pre-emphasis
filter is:

H ðzÞ ¼ 1  az1 : ð5Þ

(2) S-transform
Transform the speech into the time-frequency joint domain.
(3) Through the Mel filter bank
Transform the signal into the Mel spectrum.
(4) Take logarithms for energy
The log energy output for each filter bank is:
!
X
N 1
2
E ðmÞ ¼ ln jXa ðkÞj Hm ðkÞ ; 0  m  M: ð6Þ
k¼0

where M means the number of trigonometric filters.

228 L. Xunbo et al.

(5) Calculate discrete cosine transform (DCT) to get SMFCC cepstrum parameters:

X
N 1  
pnðm  0:5Þ
smfccðnÞ ¼ E ðmÞcos : ð7Þ
m¼0
M

(6) Extract dynamic difference parameters.

The standard cepstrum parameter MFCC only reflects the static characteristics of
the speech parameters, and the dynamic characteristics of the speech can be described
by the static characteristics of the differential spectrum. Combining dynamic and static
features can effectively improve the system’s recognition ability.

3.3 SMFCC Cepstrum Distance

Assuming the speech signal is xs ðnÞ, the SMFCC of each frame of the speech signal is
smfccspeech ði; nÞ, the leading speechless segment is used as the background noise frame,
the average value of several previous frames SMFCC is taken as the estimated value of
the background noise of SMFCC. Since it is irrelevant to i after the value is averaged,
we denoted it as smfccnoise ðnÞ, so the SMFCC cepstrum distance is:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X p
 2
dsmfcc ðiÞ ¼ smfccspeech ði; nÞ  smfccnoise ðnÞ : ð8Þ
n¼1

In Eq. (8), p refers to the order used in the SMFCC cepstrum analysis.

4 SMFCC Cepstrum Distance Feature Extraction Based

on Spectral Subtraction

4.1 Discrete S-Transform Based on Spectral Subtraction

Spectral subtraction is one of the most commonly used methods for speech noise
reduction. Assumed that the speech signal xðnÞ can be expressed as:

yðnÞ ¼ xðnÞ þ d ðnÞ: ð9Þ

Which xðnÞ is a pure speech signal, d ðnÞ is a noise signal, and the s-transform is
performed on both sides of Eq. (9):

Ys ðm; nÞ ¼ Xs ðm; nÞ þ Ds ðm; nÞ: ð10Þ

where m denotes the m-th sample point in the ST, and n denotes the n-th frame.
Ys ðm; nÞ can be obtained by direct discrete s-transform. However, the noise energy
spectrum Ds ðm; nÞ cannot be obtained accurately and has to be obtained approximately
by the way of estimation.
h The inormal estimation method in [9, 10] is to take the
statistical average E jDs ðm; nÞj2 in the absence of speech as jDs ðm; nÞj2 .
 Speech Endpoint Detection Based on Improvement Feature and S-Transform 229

It is known that the length of the noise segment is IS, and the corresponding frame
number is NIS. The average energy value of the noise segment can be obtained as:

h i 1 XNIS
E jDs ðm; nÞj2 ¼ jXs ðm; nÞj2 : ð11Þ
NIS i¼1

By obtaining the power spectrum of the silent speech segment and subtracting the
power spectrum of the original speech, then the power spectrum of the relatively pure
speech signal is processed by square root, and after the inverse S-transform (IST), a
relatively pure speech signal xei ðkÞ can be obtained:
!
X 1XN1
xei ðkÞ ¼ f i2pnk
Xs ðm; nÞ e N ; 0  k  N  1: ð12Þ
N m¼0

4.2 Spectral Subtractive Feature Extraction of SMFCC

Before extracting the SMFCC cepstrum feature, spectral subtraction is performed on
the noisy-added speech signal to remove most of the noise and suppress the distortion
of the speech waveform. The principle is illustrated in Fig. 1.

Noise
estimation

Quadratic
Framing ST
amplitude

Keep phase Spectrum

position reduction

square root
IST
of amplitude

Fig. 1. Principle diagram for spectral subtractive feature extraction of SMFCC.

5 Bandwidth Variance Feature Extraction of Uniform

Subbands

The characteristics of noise and speech in the frequency domain are quite different. In
general, the voice energy of a voiced segment has a large change with the frequency
band, and there is a larger peak at the resonance peak, but the energy value in the noise
segment is relatively low, and more evenly distributed within the frequency band.
According to this characteristic of the speech signal, the band variance becomes the
230 L. Xunbo et al.

most commonly used characteristic parameter. The uniform subband separation fre-
quency band variance is proposed as the speech characteristic parameters of the signal
based on this characteristic [11].

5.1 Uniform Subband Separation Band Variance

In order to eliminate the effect of noise on the amplitude of each spectral line, a method
based on uniform subband separated band variance algorithm is proposed. Assuming
the length of each frame is N, there are ðN=2 þ 1Þ spectral lines in positive frequency
after FFT, we can obtain amplitude spectrum Xi ¼ fXi ð1Þ; Xi ð2Þ; . . .; Xi ðN=2 þ 1Þg
after DFT, then split it into q subbands, the number of lines contained in each subband
is:

p ¼ fix½ðN=2 þ 1Þ=qðfix½ means to get the integer partÞ: ð13Þ

p is so called uniform subband, which means each subband is equal bandwidth. The
amplitude of the m-th subband in the i-th frame is:

X
1 þ ðm1Þp þ ðp1Þ
XXi ðmÞ ¼ jXi ðkÞj: ð14Þ
k¼1 þ ðm1Þp

Assuming XXi ¼ fXXi ð1Þ; XXi ð2Þ; . . .; XXi ðqÞg, Then we can get the sub-band
mean value of the i-th frame speech signal:

1X
q
Ei;1 ¼ XXi ðk Þ: ð15Þ
q k¼1

The sub-band variance of the i-th frame speech signal can expressed as:

1 X
q
Di;1 ¼ ½jXXi ðkÞj  Ei;1 2 : ð16Þ
q  1 k¼1

6 Dynamic Threshold Update and Endpoint Detection

The most commonly method used in endpoint detection is a dual threshold method. By
extracting typical characteristic parameters of the speech, an proper threshold is
selected to determine the start and end points of speech and to detect speech frames and
non-speech frames [12]. This paper selects the improved spectral subtractive SMFCC
cepstral distance and uniform subband variance as features, adopts a dynamic threshold
estimation mechanism and uses the two-parameter double threshold method to perform
endpoint detection. The procedure of the spectral subtraction algorithm based on ST is
given as follows:
 Speech Endpoint Detection Based on Improvement Feature and S-Transform 231

The detection algorithm flow is shown in Fig. 2.

7 Experiment Results

The speech database used in the experiments here is 126 pure voices in the TIMIT
voice database. The noise signals used in the simulation includes 5 kinds of noise
(white, pink, volvo, babble and factory noises) of NOISEX-92 Datebase. The pure
speech signals and various noise signals are mixed at 3 different SNR(10 dB, 5 dB, and
0 dB). In order to verify the correctness and robustness of this method, we compared
the proposed algorithm with the algorithms mentioned in literature [4] and literature
[13]. All three algorithms perform well when SNR is 10 dB, the detection results with
SNR of 5 dB and 0 dB are shown in Figs. 3, 4 and 5, it can be seen that when the SNR
is 5 dB, the detection method based on the DWT-MFCC cepstrum distance missed
speech signals between 3 s and 3.5 s; when the SNR is 0 dB, the two comparison
algorithms were not able to detect the behind speech segments while the improved
232 L. Xunbo et al.

Noise-added Speech Signal

Pretreatment

Spectral Subtraction Based on S-transform to

Reduce Noise

Calculate the SMFCC DFT

Calculate the SMFCC Each frame is divided

cepstral distance into q uniform subbands

Thresholds T1, T2 update Calculate the

dynamically subband variance

No First-level
judgment Thresholds T3 T4
update dynamically
Exceeded the
threshold?
No
Yes
Second-level
judgment Exceeded the
threshold?
Yes
Speech frame

Fig. 2. Improved algorithm endpoint detection flowchart.

Pure speech signal Pure speech signal

Amplitude

1
Amplitude

1
0 0
-1 -1
0 0.5 1 1.5 2 2.5 3 3.5 0 0.5 1 1.5 2 2.5 3 3.5
Ɵme/s Ɵme/s
Noise-added speech signal (SNR=5dB) Noise-added speech signal (SNR=0dB)
Amplitude

Amplitude

1 1
0 0
-1 -1
0 0.5 1 1.5 2 2.5 3 3.5 0 0.5 1 1.5 2 2.5 3 3.5
Ɵme/s Ɵme/s

Short-term spectrum reducƟon SMFCC cepstrum distance Short-term spectrum reducƟon SMFCC cepstrum distance
Amplitude

Amplitude

20 20
0 0
0 0.5 1 1.5 2 2.5 3 3.5 0 0.5 1 1.5 2 2.5 3 3.5
Ɵme/s Ɵme/s
Bandwidth variance for short-Ɵme uniform subband separaƟon
Amplitude

Bandwidth variance for short-Ɵme uniform subband separaƟon

Amplitude

200 200
100 100
0 0
0 0.5 1 1.5 2 2.5 3 3.5 4
0 0.5 1 1.5 2 2.5 3 3.5 4 Ɵme/s
Ɵme/s

Fig. 3. The endpoint detection results of improved algorithm based on ST in SNR = 5 dB and
SNR = 0 dB.

algorithm in this paper can still detect the position of each segment of speech. The
experiment results can clearly show that the proposed algorithm has higher robustness
for speech endpoint detection.
In order to further verify the stability and effectiveness of the proposed algorithm,
the correct rate results are used to evaluate the detection results under different noise
 Speech Endpoint Detection Based on Improvement Feature and S-Transform 233

Pure speech signal Pure speech signal

1 1

Amplitude
Amplitude

0 0

-1 -1
0 0.5 1 1.5 2 2.5 3 3.5 0 0.5 1 1.5 2 2.5 3 3.5
Noise-added speech signal (SNR=5dB) Noise-added speech Signal (SNR=0dB)
1 1

Amplitude
Amplitude

0 0

-1 -1
0 0.5 1 1.5 2 2.5 3 3.5 0 0.5 1 1.5 2 2.5 3 3.5
DWT-MFCC Cepstrum distance DWT-MFCC Cepstrum distance
Amplitude

Amplitude
10
10

0
0 0.5 1 1.5 2 2.5 3 3.5 0
Ɵme/s 0 0.5 1 1.5 2 2.5 3 3.5
Ɵme/s

Fig. 4. The endpoint detection results based on DWT-MFCC cepstrum distance in SNR = 5 dB
and SNR = 0 dB.

Pure speech signal

Pure speech signal 1
1
Amplitude
Amplitude

0 0

-1 -1
0 0.5 1 1.5 2 2.5 3 3.5 0 0.5 1 1.5 2 2.5 3 3.5
Noise-added speech Signal(SNR=5dB) Noise-added speech signal(SNR=0dB)
1 1
Amplitude

Amplitude

0 0

-1 -1
0 0.5 1 1.5 2 2.5 3 3.5 0 0.5 1 1.5 2 2.5 3 3.5
Spectral subtraction and uniform sub-band spectrum variance Spectral subtraction and uniform sub-band spectrum variance
200 200
Amplitude

Amplitude

100 100

0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
Ɵme/s Ɵme/s

Fig. 5. The endpoint detection results based on spectral subtraction and uniform sub-band
spectrum variance in SNR = 5 dB and SNR = 0 dB.

Table 1. The correct rate of different algorithms in different noises (SNR = 5 dB).
Noise Proposed DWT- Spectral Proposed algorithm
algorithm MFCC subtraction + uniform based on ST in pure
based on ST cepstrum sub-band spectrum speech
distance variance
white 92.7% 84.4% 90.5% 96.2%
pink 91.6% 82.8% 90.6% 99.1%
volvo 91.3% 80.4% 90.3% 97.6%
babble 90.4% 81.3% 89.9% 96.4%
factory 90.2% 79.7% 89.5% 96.0%
234 L. Xunbo et al.

Improved algorithm based on ST

DWT-MFCC cepstrum distance

Spectral subtracƟon + uniform
sub-band spectrum variance

Error rate/%
DetecƟon rate/%

Improved algorithm based on ST

DWT-MFCC cepstrum distance
Spectral subtracƟon + uniform
sub-band spectrum variance

Fig. 6. The detection rate and error rate of different algorithms under different SNR

environments when the SNR is 5 dB. As shown in Table 1, it can be clearly seen that
the improved algorithm in this paper has better detection accuracy.
The detection rate and error rate are used to evaluate the superiority of this algo-
rithm. It is easy to see that the improved algorithm has a higher detection rate and lower
error rate relative to the compared algorithms from Fig. 6. The error rate of the pro-
posed algorithm is relatively stable under different SNR, which illustrates that this
algorithm has more stable and reliable performance, stronger anti-noise robustness in
speech endpoint detection.

8 Conclusion

In the low SNR and non-stationary noise environment, traditional feature detection
methods will lead to a sharp drop in detection performance. This paper proposed a
feature-improved speech endpoint detection algorithm based on s-transform. This
method combines s-transform with MFCC and spectral subtraction to reduct noise. The
experimental results show that this algorithm has a high accuracy and has advantages in
the detection rate and error rate compared to the other two algorithms in Literature [4,
13]. It can effectively suppress noise and improve the quality of speech.

References
1. Stockwel, R.G., Mansinha, L., Lowe, R.P.: Localization of the complex spectrum: the S-
transform. J. IEEE Trans. Signal Procesing. 44(4), 998–1001 (1996)
2. Huang, L., Yang, C.: A novel approach to robust speech endpoint detection in car
environments. In: IEEE International Conference on, vol. 3, pp. 1751–1754. IEEE (2000)
3. Nakagawa, S., Wang, L.: Speaker identification and verification by combining MFCC and
phase information. In: IEEE International Conference on Acoustics, vol. 20, no. 4, pp. 4529–
4532 (2009)
4. Yin, R., Cheng, J.: Improved feature extraction algorithm based on DWT-MFCC. J. Modern
Electron. Technol. 40(9), 18–21 (2017)
 Speech Endpoint Detection Based on Improvement Feature and S-Transform 235

5. Zhang, Z., Yao, E., Shi, Y.: Audio endpoints detection algorithm based on wavelet analysis
and MFCC. J. Electron. Meas. Technol. 39(7), 62–66 (2016)
6. Zeng, S., Jingxiang, L.: Speech endpoint detection method based on fusion of MFCC
distance and logarithmic energy parameter. J. Audio Eng. 40(9), 51–55 (2016)
7. Cao, D., Gao, X., Gao, L.: An improved endpoint detection algorithm based on MFCC
Cosine Value. J. Wirel. Pers. Commun. 95, 2073–2090 (2017)
8. Wang, H., Yu, Z.: SMFCC: a novel feature extraction method for speech signal. J. Comput.
Appl. 36(6), 1735–1740 (2016)
9. Goh, Z., Tan, T.-C., Tan, B.T.G.: Postprocesing method for suppressing musical noise
generated by spectral subtraction. IEEE Trans. Speech Audio Process. 6(3), 28–292 (1998)
10. Paliwa, K., Wojcicki, K., Schwerin, B.: Single-channel speech enhancement using spectral
subtraction in the short-time modulationdomain. J. Speech Commun. 52(5), 450–475 (2010)
11. Wang, Z.-F.: Speech endpoint detection method research based on double threshold-
frequency band variance. Electron. Des. Eng. 24(19), 86–88 (2016)
12. Sun, Y., Wu, Y., Li, P.: Research on speech endpoint detection based on the improved dual-
threshold. J. Chang. Univ. Sci. Technol. (Nat. Sci. Ed.) 39(1), 92–95 (2016)
13. Wang, W., Hu, G., Yang, L., et al.: Research of endpoint detection based on spectral
subtraction and uniform sub-band spectrum variance. Audio Eng. 40(5), 40–43 (2016)
 A Localization Evaluation System
for Autonomous Vehicle

Yuan Yin, Wanmi Chen(&), Yang Wang, and Hongzhou Jin

School of Mechatronic Engineering and Automation, Shanghai University,

Shanghai 200072, China
{Alexx_Yin,wanmic,wy7ang}@163.com

Abstract. The autonomous vehicle is a kind of intelligent robot. At present,

there are many algorithms for autonomous vehicle localization, but few methods
for localization evaluation. To solve this problem, a grid hypothesis model using
the existing prior information of the surrounding environment and posterior
information of current laser real-time collection of autonomous vehicle is pro-
posed, Kullback–Leibler divergence and Fourier transform are methods to
evaluate the current location results. The above two methods can give relatively
accurate evaluation results and corresponding evaluation method can be selected
according to the actual speed and accuracy requirements.

Keywords: Autonomous vehicle
Localization evaluation

Grid hypothesis model
Kullback–Leibler divergence
Fourier transform

1 Introduction

The autonomous vehicle is an intelligent robot system that is driven by wheels or

crawlers and has the ability to autonomously move on the ground under road or non-
road conditions. An autonomous vehicle has been extensively studied in plenty of field
in recent years. Localization, as one of the core of autonomous vehicle research, is the
focus of many scientific research institutions and scholars. The accuracy and robustness
of the localization is an important core indicator to measure the intelligence of the
autonomous vehicle, and it is also one of the bottlenecks for the practical application of
autonomous vehicle technology. Many researchers have developed different algorithms
and models to solve this problem.
In [1], a self-localization algorithm consisting of a 2D laser range finder and the
particle filter is proposed. The particle filter estimates the state of the vehicle by
approximating the probability distribution of the system using a set of particles [2].
Each particle represents a possible state of the vehicle. While the literature [3] proposes
a modified Markov localization algorithm for terrain-based navigation control of a light
autonomous vehicle. Markov localization is a probabilistic algorithm. Instead of
maintaining a single hypothesis as to which cell the vehicles might be, Markov
localization maintains a probability distribution over the space of all such hypotheses
[3]. Similarly, Kalman filter and its improved algorithm are also widely used in the
positioning of autonomous vehicles [4–7].

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 236–245, 2018.
https://doi.org/10.1007/978-981-13-2384-3_22
 A Localization Evaluation System for Autonomous Vehicle 237

However, due to the uncertainty of the mobile robot itself, it will inevitably affect
the accuracy of localization. Scholars have also conducted research on this issue. As the
researches did in [8], using entropy and mutual information metrics to estimate
quantitative uncertainty estimates and the effectiveness of information theory concepts
within a probabilistic framework. In [9], it suggests a position estimation algorithm
based upon an adaptive fading Kalman filter to overcome the inaccurate results due to
limited linear approximation, which the adaptive fading factor enables the estimator to
change the error covariance according to the real situation. And the literature [10]
associates the new observations to existing estimates in terms of the Kullback-Leibler
divergence to evaluate the proposed method to augment standard dynamic object
trackers which solve the problem that other traffic participants are often occluded from
sensor measurements by buildings or large vehicles in autonomous driving scenarios.
But at present, there are few methods to evaluate the results of the localization of
autonomous vehicles.
To solve this problem, we construct a localization evaluation system based on
Gaussian distribution model that is independent of its own localization system and is
related to the actual environmental information. The Kullback–Leibler divergence and
Fourier transform methods are used to process the localization data to complete the
evaluation of the localization effect.

2 Localization and Evaluation Model

2.1 Localization Model

Localization model is a three-dimensional Gaussian distribution model(X-Y-Z). The
probability density function of a multidimensional Gaussian distribution is as follows:

1
pðxÞ ¼ detð2pRÞ2 expf ðx  lÞT R1 ðx  lÞg:
1
ð1Þ
2

In the formula, l is the mean vector and R is a semi-positive definite symmetric

matrix, which is the covariance matrix (in this model, the covariance matrix is a 3  3
semi-positive definite symmetric matrix). Gaussian model has a wide range of appli-
cations in robot localization, like in [11]. Gaussian techniques all share the basic idea
that beliefs are represented by multivariate Gaussian distributions. The commitment to
represent the posterior by a Gaussian has important ramifications. Most importantly,
Gaussians are unimodal; they possess a single maximum. Such a posterior is charac-
teristic of many tracking problems in robotics, in which the posterior is focused around
the true state with a small margin of uncertainty [2].
In order to describe the pose of the unmanned vehicle more clearly, we introduce
Euler angles (roll, pitch, yaw) to construct a model of pose6D (x, y, z, roll, pitch, yaw).
Yaw represents the angle of rotation about the Y axis, Pitch the angle of rotation about
the X axis, and Roll represents the angle of rotation about the Z axis, as shown in
Fig. 1. Any rotation angle can be obtained by rotating these three times in order.
238 Y. Yin et al.

Fig. 1. Euler angle

2.2 Evaluation Model

Grid Hypothesis Model. Agrid hypothesis model using the existing prior information
of the surrounding environment and posteriori information of current laser real-time
collection of autonomous vehicle is constructed and it is the basis and focus of this
localization evaluation system. The prior information is the set of existing high-
definition maps and the array of data collected by the laser sensor in the area with a
certain accuracy. The posterior information is the array of data of the current auton-
omous vehicle collected by the laser sensor in real time. Comparing the posteriori
information with the priori information on the collected data of the corresponding
position of the autonomous vehicle, the matching degree of the two can be obtained.
We use the matching degree to evaluate the localization results of the autonomous
vehicle. Figure 2 illustrates how to get the priori data array and posterior data array.

Fig. 2. Data collection schematic diagram

As shown in Fig. 2, if the data are collected in the A and B directions, the resulting data
array is A = [0.3, 0.5, 0, 0.2, 0.7, 0.4, 0, 0], B = [0, 0.2, 0.7, 0.4, 0, 0, 0.3, 0.5]. The
priori data array and an array of posterior data can be obtained by collecting data by
laser. Then we construct a hypothesis grid to describe the model.
The output from the localization algorithm as a center to construct a localization
frame composed of m  n grids. Each grid acquires data in l directions by laser and a
total of m  n  l sets of data are obtained. Figure 3 is a top view of the grid local-
ization frame.
 A Localization Evaluation System for Autonomous Vehicle 239

Fig. 3. Grid positioning frame top view

Fig. 4. A part of grid hypothesis (Color figure online)

Comparing the above m  n  l sets of data with the corresponding current real-
time laser acquisition data to obtain the matching degree, in order to better describe the
matching degree in each grid, we simulated a visualized image of the matching degree.
For the convenience to display, one of the l directions in each grid is taken, the red
arrow shown in Figs. 4 and 5, and the height(ordinate) of the red arrow is the matching
value of prior data and posterior data.
Finally, on the basis of constructing a real-time grid hypothesis evaluation model,
Kullback–Leibler divergence and Fourier Transform methods are used to evaluate the
current localization results.
240 Y. Yin et al.

Fig. 5. Grid hypothesis after scaling (Color figure online)

Kullback-Leibler Divergence. The Kullback–Leibler divergence (also called relative

entropy) is a measure of how one probability distribution diverges from a second,
expected probability distribution [12]. The definition formula is:
X PðiÞ
DKL ðPjjQÞ ¼ PðiÞ ln : ð2Þ
i
QðiÞ

the value of DKL ðPjjQÞ indicates the information loss when using Q to approximate P,
if it is close to 0, the two distributions have similar behavior, while a value close to 1
means two very different distributions. In the evaluation model, P denotes the data
distribution collected in the localization hypothesis grid, Q denotes the theoretical data
distribution of the localization hypothesis grid in the collection data set of the region
and DKL ðPjjQÞ denotes the relative entropy of the two distributions. By calculating the
distance between the data of the localization hypothesis grid and the priori data of the
localization hypothesis grid in the coordinate system, the confidence is obtained by
combining the DKL ðPjjQÞ parameter, thereby completing the localization evaluation.
Fourier Transform. Fourier Transform is an algorithm that samples a signal over a
period of time (or space) and divides it into its frequency components. Fourier analysis
converts a signal from its original time domain to a representation in the frequency
domain and vice versa. The following are the formulas for the discrete Fourier forward
and inverse transforms:

X
N1
i2pkn
XðkÞ ¼ xðnÞe N

n¼0
ð3Þ
X
N 1
2pkn 2pkn
¼ xðnÞ½cos  i sin ; k ¼ 0; 1; . . .; N  1:
n¼0
N N
 A Localization Evaluation System for Autonomous Vehicle 241

1XN 1
i2pkn
xðnÞ ¼ XðkÞe N
N k¼0
ð4Þ
1XN 1
2pkn 2pkn
¼ XðkÞ½cos þ i sin ; n ¼ 0; 1; . . .; N  1:
N k¼0 N N

where the latter expression follows from the former one by Euler’s formula:

eix ¼ cos x þ i sin x: ð5Þ

In this model, evaluation with Fourier Transform method in three steps:

Step 1: Get error. First, we find the maximum data in grid hypothesis, then record
the serial number and subtract the central serial number.
Step 2: Fourier Transform. Through the previous step we get the error sequence in
time domain and then use the Fourier Transform to get the error result in frequency
domain.
Step 3: Analyze the results. Low-frequency signal indicates false result to complete
the evaluation of the localization result.

3 Experiments
3.1 Evaluation Steps

Step 1: Using the pose6D localization model introduced to Euler angles to locate the
autonomous vehicle. A high-definition map and a collection of arrays of data col-
lected by the laser sensor in that area are ready.
Step 2: After several experiments on the localization of the autonomous vehicle, the
accuracy of the optimal localization frame is:
X dimension: 63; Y dimension: 63; theta dimension: 32; theta resolution: 0.00436;
0.16 meters per pixel.
The data in its localization frame are collected by laser sensor.
Step 3: Comparing the data obtained in the previous step with the priori data set
corresponding to the first step to obtain the heat map.
In this heat map, red is darkest than other colors, from red to background blue is from
dark to light. The darker the color, the closer the two sets of data are. The smaller
offset in the first heat map in Fig. 6 indicates that the localization is more accurate, the
offset in the second heat map is centered, and the third is larger, indicating poor
localization results.
Step 4: Using Kullback–Leibler divergence and Fourier transform respectively for
localization evaluation. We define data from grid hypothesis½i as dfgh½i and
data from grid pose½i as dfp½i. According to (2), we can obtain the relative
entropy:
242 Y. Yin et al.

Fig. 6. Heat map (Color figure online)

 
 dfgh½i
KLparam 
¼ 100  dfgh½i  ln : ð6Þ
dfp½i 

Then obtain confidence for localization evaluation with the KL parameter.

According to the definition (3) and (4), one-dimension complex discrete Fourier
Transform is performed on the difference discrete value over time, and the real part of
the output value is judged. If it is greater than zero, the localization result is True,
otherwise it is False. And calculate confidence for evaluation.

3.2 The Experimental Results

The Kullback–Leibler divergence and Fourier transform methods are used to evaluate
the localization results of the autonomous vehicle, and calculating the confidence
obtained by the two methods to judge whether the localization is correct or not, 1 for
correct localization and 0 for localization failure. Figures 7 and 8 shown the evaluation
confidence and localization results obtained by using the Fourier transform method in
two time periods, Figs. 9 and 10 shown the evaluation confidence and localization
results obtained by using the Kullback–Leibler divergence method.

Fig. 7. Evaluation confidence and localization results by using the Fourier transform
 A Localization Evaluation System for Autonomous Vehicle 243

Fig. 8. Evaluation confidence and localization results by using the Fourier transform

Fig. 9. Evaluation confidence and localization results by using the Kullback–Leibler divergence

Fig. 10. Evaluation confidence and localization results by using the Kullback–Leibler
divergence

In Figs. 7, 8, 9 and 10, the y axis represents the value of evaluation confidence and
localization result that change with time. In Fig. 7, when the localization fails, the
confidence changes obviously and the localization failure can be immediately judged
based on the result, while in Fig. 9, the confidence changes slowly and the evaluation
speed is slower. In Fig. 8, when some small localization errors occur (such as the
impact of pedestrians on the road, changes in some buildings, etc.), result is judged as
244 Y. Yin et al.

localization failure, as shown in Fig. 10 that the evaluation analysis using the Kull-
back–Leibler divergence method is more accurate.

4 Conclusion

In this note, we have considered the evaluation of the localization effect of autonomous
vehicles under real environmental information using the Kullback–Leibler divergence
and Fourier transform evaluation models.
The experimental results show that the two evaluation models can evaluate the
localization effect. The Fourier transform evaluation model evaluates the localization of
the autonomous vehicle faster than the Kullback–Leibler divergence evaluation model.
However, the Kullback–Leibler divergence evaluation model is more effective than the
Fourier transform evaluation model in dealing with some unexpected conditions
encountered by autonomous vehicles under the real environmental information, and it
is more accurate in localization evaluation.

Acknowledgments. This work is supported by Shanghai University, and we would like to

appreciate the Prof. Wanmi Chen and M.E Yang Wang for the support of our paper.

References
1. Kurashiki, K., Fukao, T., Ishiyama, K., Kamiya, T., Murakami, N.: Orchard traveling UGV
using particle filter based localization and inverse optimal control. In: The 3rd Symposium
on System Integration, pp. 31–36. IEEE Computer Society, Japan (2010)
2. Thrun, S., Burgard, W., Fox, D.: Probabilistic Robotics. The MIT Press, England (2006)
3. Shoukry, Y., Abdelfatah, W.F., Hammad, S.A.: Real-time Markov localization for
autonomous UGV. In: 4th International Design and Test Workshop. IEEE Computer
Society, Saudi Arabia (2009)
4. Tripathi, S.K., Sapre, R.M.: Robust target localization and tracking using Kalman filtering
for UGV-UAV coordinated operation. In: 2016 IEEE International Conference on Recent
Advances and Innovations in Engineering, India (2016)
5. Tsalatsanis, A., Valavanis, K.P., Kandel, A., Yalcin, A.: Multiple sensor based UGV
localization using fuzzy extended Kalman filtering. In: 2007 Mediterranean Conference on
Control and Automation. IEEE Computer Society, Greece (2007)
6. Kwon, J.-W., Park, M.-S., Kim, T.-U., Chwa, D.-K., Hong, S.-K.: Localization of outdoor
wheeled mobile robots using indirect Kalman filter based sensor fusion. J. Inst. Control. Rob.
Syst. 14, 800–808 (2008)
7. Nguyen, N., Tyagi, D., Shin, V.I.: An observation model based on polyline map for
autonomous vehicle localization. In: Proceedings of the IEEE International Conference on
Industrial Technology, India, pp. 2427–2431 (2006)
8. Messina, E., Madhavan, R.: Intelligent unmanned ground vehicle navigation via information
evaluation. Adv. Rob. 20, 1375–1400 (2006)
9. Sung, W.J., Choi, S.O., You, K.H.: TDoA based UGV localization using adaptive Kalman
filter algorithm. In: 2008 2nd International Conference on Future Generation Communica-
tion and Networking Symposia, China, vol. 4, pp. 99–103. IEEE Computer Society (2008)
 A Localization Evaluation System for Autonomous Vehicle 245

10. Galceran, E., Olson, E., Eustice, R.M.: Augmented vehicle tracking under occlusions for
decision-making in autonomous driving. In: IEEE/RSJ International Conference on
Intelligent Robots and Systems, Germany, pp. 3559–3565 (2015)
11. Pfaff, P., Plagemann, C., Burgard, W.: Gaussian mixture models for probabilistic
localization. In: IEEE International Conference on Robotics and Automation, USA,
pp. 467–472 (2008)
12. Wikipedia. https://en.wikipedia.org/wiki/Kullback%E2%80%93Leibler_divergence#cite_
note-1
 Cooperative Slip Detection Using
a Dual-Arm Baxter Robot

Shane Trimble(B) , Wasif Naeem, and Seán McLoone

School of Electronics, Electrical Engineering and Computer Science,

Queen’s University Belfast, Belfast, UK
[email protected]

Abstract. When dealing with robotic manipulation tasks, slip detection

and control is vital to overcome payload uncertainties and to compensate
for external disturbances. Many modern smart manipulators are shipped
with integrated joint torque sensing capabilities providing a potential
means of detecting slip and hence generating a feedback signal for slip
control, without the need for additional external sensors. This paper
investigates preliminary results on slip detection with an aim to extend
the existing work on single manipulator to a dual arm cooperative manip-
ulator system. The slip signal is obtained through filtering of onboard
joint torque measurements when different materials are held in a friction
grip between the two cooperating robotic end-effectors of a Baxter robot.
Frequency domain analysis of data from preliminary experiments
shows a substantial frequency component in the vicinity of 12 Hz when
slip is taking place. Consequently, a 9–15 Hz band-pass Chebyshev fil-
tered torque signal is proposed as a real-time slip detection signal and
its performance evaluated in experiments using a selection of materials
with differing mass, coefficient of friction and thickness.
Analysis of the resulting filter output and threshold signal for each
material revealed that the proposed approach was effective at detect-
ing slip for some materials (laminated card, MDF, PVC and Perspex)
but not for others (sheet steel, cardboard and polypropylene). This is
attributed to the variation in consistency in slip behaviour of different
material/weight combinations.

1 Introduction
The manipulation of objects by robots is typically accompanied by events such
as slippage between a robot’s fingers or, in the case of two non-dexterous end-
effectors, robotic arms. Humans are able to identify such events using a combi-
nation of superficial and deep mechanoreceptors [1], but for robots, with more
elementary mass, stiffness and friction sensing, such events can be hard to dis-
tinguish [2].
In industry many items manipulated by robots are too wide or heavy to
be practically held in the grasp of a single enclosed end-effector due to size
or payload limitations. In certain scenarios manipulators may work together in
c Springer Nature Singapore Pte Ltd. 2018
K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 246–258, 2018.
https://doi.org/10.1007/978-981-13-2384-3_23
 Cooperative Slip Detection Using a Dual-Arm Baxter Robot 247

pairs, where an object may be held in a friction only grasp between the end-
effectors, without additional support employed [3].
In the field of robotic slip detection and control, much work has been done
on single dexterous end-effectors. While several methods exist to detect slip,
such as the ones using accelerometers [4,5] and optical sensors [6,7] the main
method of detecting slip is via pressure signal analysis of piezoresistive force
sensing resistors (FSR) mounted on contact points [8]. These sensors encounter
a change in resistivity which is directly proportional to the change in pressure.
A typical time domain FSR sensor signal with slip occurring can be seen in
Fig. 1. Such a sensor is incorporated into a control loop to detect slippage based
on empirical mode decomposition in [2]. This approach relies on the analysis of
vibrations generated during slippage, however challenges are faced in discrim-
inating between noise and the small signals generated by slippage because of
low signal-to-noise ratio, necessitating the use of time domain filtering on any
feedback signals. In [9], a novel flexible, thin and lightweight FSR was developed
in order to detect slip. The authors encountered issues differentiating between
slip and changes in normal force, and so in [10], they investigated a method
for identifying object slip by analysing the frequency components of the output
measurements from the sensor. End-effector mounted strain gauge signals were
analysed after Fast Fourier Transformation in [11]. Seven band-pass filters with
different frequency bands were designed based on these signals and a threshold
for each defined. Once slip is detected, the parameters of a sliding mode slip
prevention controller are updated so that the grip force is increased in discrete,
predetermined amounts to prevent the object slipping further.

Fig. 1. A typical FSR time domain signal response with obvious high frequency com-
ponents while slip is occurring, from 5 to 7 s [2].

Similarly, in [12], FSR sensor slip signals were transformed using FFT. This
data was then used to train a standard artificial neural network for incipient
slip detection, which successfully detected slippage of five different objects, of
varying materials.
The optimal sensing layout for a single-end effector with two load cells mea-
suring normal and tangential force used in conjunction with an optical sensor is
investigated in [13]. Slip is induced through the modulation of grasp force applied
to an object. The paper concludes that it is preferable to monitor normal force,
248 S. Trimble et al.

tangential force and vertical slippage (optically) simultaneously. With feedback

coming from this sensor combination no uncontrolled slip was observed.
Cooperating manipulators is currently an active area of study. In [6], interac-
tion effect characterization during manipulation of a common object on such a
system is explored. Kinematic constraints for the system are taken into account
when modelling the overall system dynamics. The fundamental properties of the
system during a manipulation task are derived and an experiment is conducted
with two cooperating anthropomorphic robotic arms, which support the findings.
Coupled dynamics are presented in [14] for two cooperating mobile robotic
arms manipulating a moving object with additional factors of external distur-
bances and other uncertainties accounted for. Centralized robust adaptive con-
trols are introduced to guarantee set motion and force trajectories of an object.
Closed-loop stability and the boundedness of tracking errors of the system are
validated using Lyapunov stability synthesis.
A model for a cooperative continuum robot (CCR) concept is presented in
[15]. These robots working in tandem allow for a large variety of new applications
due to their flexibility, such as maintenance, search and rescue operations and
medical interventions.
When selecting an appropriate controller feedback signal, it is desirable to
utilize existing built-in sensors for both simplicity and economical reasons, as
opposed to adding additional external circuitry.
Modern smart collaborative robots have spring loaded joints with torque
sensing capabilities [16]. While these robots have many benefits which make
them more suitable to a human workspace, such a system presents a number of
challenges. Firstly, actuators can store unwanted energy in the elastic element
which can potentially be dangerous and, secondly, system uncertainties and non-
linearities are present. The authors of [17] make use of a Variable Boundary
Layer Sliding Mode Controller (VBSMC) to account for this on a single Baxter
manipulator, showing an increase in compliance when the arm is moved by a
human.
For object manipulation, a multi-purpose blank end-effector allows for a mul-
titude of tasks to be done without the need for retooling for different jobs. While
every task would not be suitable for this type of end-effector, most pushing, lift-
ing and grasping jobs could be performed by robots with such devices attached.
Two of these acting together allows for a wide range of objects to be manipu-
lated, through the application of a normal force acting orthogonally to an object
and the use of friction to overcome gravitational and motion acceleration forces.
Collaborating and general purpose robotics, two of the major focuses of
Industry 4.0, are set to become more prevalent in the future [18]. It is therefore
important to optimally design these systems for general purpose use on the shop
floor.
The addition of slip control functionality increases the universal nature of a
cooperating robot system, and so a natural extension to the field is slip detection
and control using a pair of collaborating arms. In this paper, a preliminary study
has been conducted to investigate the feasibility of using the built-in torque
 Cooperative Slip Detection Using a Dual-Arm Baxter Robot 249

sensing capability of a Baxter dual arm robot, to be used ultimately as a feedback

signal for control of slip between two collaborating manipulators. This has been
analysed with a range of materials with differing weights for comparison.
The remainder of the paper is organised as follows. Section 2 presents a
description of the robot and the related software and hardware including the
end-effector, filter design and overall system architecture. Whilst the methodol-
ogy of slip detection is discussed in Sect. 3 with Sect. 4 providing details of the
experiments. Results are presented in Sect. 5 whereas Sect. 6 highlights ensuing
discussions. Finally, Sect. 7 draws the concluding remarks and future work.

Table 1. ROS topics used in experiment

Rostopic Built-in Experiment usage Pub freq [Hz]

.. /gravity compensation torques/actual effort Yes S0 & W 0 torque values 95
.. /suppress collision avoidance Yes Deactivate 10
.. /filter output No Filter W 0 orque 200

2 System Description
This section discusses the robotic platform employed in this study including the
custom end-effectors and the Robot Operating System (ROS). The filter design
and test objects used in the experiments are also described.

Fig. 2. Baxter robot with dual 7 degree-of-freedom manipulators [19].

2.1 The Baxter Robot

Baxter is a collaborative robot with two 7 degree-of-freedom manipulators
designed to work on production lines side by side with human workers. The
robot is safe to use around humans as each of it’s 14 spring-loaded manipulator
joints, as seen in Fig. 2, have built-in torque sensing capability providing feed-
back in case of a collision. Baxter also has a ring of ultrasonic sensors encircling
the head module, allowing proximity detection of people or moving objects.
250 S. Trimble et al.

Fig. 3. 3D printed end-effector fitted to Baxter. Note the three rubber contact points
for increased gripping.

Experimentation has revealed that the end-effector pose position accuracy

deviates by a mean value of +/−15 mm [20]. The robot thus continues to cor-
rect for this error by constantly modifying the joint torque values. This makes
it challenging to extract the desired frequency signal thus necessitating time
domain filtering, as discussed in Sect. 2.4.

2.2 End-Effectors

For this experiment two customised 3D-printed end-effectors are attached to

both manipulators, as shown in Fig. 3. Three rubber bumpers are added to each
end-effector to increase contact quality and grip. The joint torque and gravity
compensation values of the robot are recalibrated after attachment, to compen-
sate for the slight increase in mass (circa 90 g).

2.3 System Architecture

In order to promote the platform, a research version of the robot was released
in early 2013 along with an open-source Software Developement Kit (SDK),
integrated with the Robot Operating System (ROS) [21]. ROS is an open-source,
meta-operating system of libraries and tools built onto Ubuntu, which provides
a networked framework for controlling and communicating with robots through
publishing/subscribing to/from Rostopics. The Rostopics employed in this study
are summarized in Table 1.
The ROS network and robot block diagram is shown in Fig. 4. This peer-
to-peer network consists of one master and one client machine. The Rosmaster
PC is an Ubuntu machine with ROS installed. On this machine three custom
nodes are running. The torque control node is a custom script used to change the
torque values of the Rostopic .../gravity compensation torques/actual ef f ort.
It is commanded with an external hand-held controller, connected over USB. The
collision avoidance suppression node publishes a simple message to the collision
avoidance node to deactivate the subsystem. This normally-on safety feature is
constantly reset, so the custom node needs to publish at a minimum of 10 Hz.
 Cooperative Slip Detection Using a Dual-Arm Baxter Robot 251

Fig. 4. System block diagram.

The Baxter robot communicates with the ROSmaster PC using TCP-IP pro-
tocol over ethernet, with a low latency of <1 ms. The robots S0 torque joint top-
ics are passed torque values using ROSmessages sent from the torque control
node in the ROSmaster PC. This applies normal force, F , through the end-
effectors onto the test object. F can then be decremented by the user to entice
slip.

2.4 Chebyshev Filter

Preliminary testing revealed a vast amount of undesirable signals and noise when
attempting to detect slip, due to end-effector motion and constant error correc-
tion. The torque value of the joint W 0 is measured in Nm at a sampling rate
of 95 Hz, similar to the 100 Hz sampling rate used in [22], but below the 400 Hz
used in [5,23].

Fig. 5. Frequency response of 9 to 15 Hz passband filter applied to the torque value of

joint W 0.

Due to the real-time operation of the system, a low order infinite impulse
response (IIR) filter was preferable to minimise excessive computation and ensure
that the rosnode containing the filter would operate at the correct frequency of
200 Hz. This value was chosen as it is high enough to prevent aliasing (>2F s),
yet sufficiently low that the node scan-cycle does not have any latency issues.
252 S. Trimble et al.

A steep roll-off is desirable to ensure that intermodulation distortion is min-

imised. A 6th order Chebyshev 9 to 15 Hz IIR bandpass filter, with the fre-
quency response shown in Fig. 5, was found to be satisfactory and it was applied
to the torque output signal of joint W 0, published as Rostopic/f ilter output.
The effects of passband ripple associated with such a filter are insignificant in
this case (≈−1 dB), due to the fact that a large attenuation was achieved for the
rejected frequencies (≈−21 dB & ≈−28 dB at passband ±5 Hz respectively).

2.5 Experimental Objects

In order to conduct experiments with varying coefficients of friction, several

everyday materials with different masses were selected. Among these, four mate-
rials which yielded consistent slip detection results were further tested with addi-
tional weights attached to examine the effect of an increased tangential force with
a constant coefficient of friction on the magnitude of the frequency response.

3 Slip Detection Methodology

As mentioned earlier, in order to design a slip control system for a dual arm
system, it is necessary to find an appropriate feedback signal for control purposes.
Determination of normal force, F , the force applied to the object by the end-
effectors during manipulation, is necessary to enable optimisation of the force in
order to prevent slip.

Fig. 6. Tangential and normal forces acting on an object while being manipulated by
two non-dexterous collaborating end-effectors, in blue (Color figure online)

It can be established from Fig. 6 that if one or both manipulators are equipped
with force feedback, F can be adjusted to a safe minimum in order to hold
the object securely without any additional support. By applying minimal F
the object can be held safely without the risk of crushing [13]. This can be
calculated when the tangential force acting on the object, Ft , and the interface
static coefficient of friction, µs , are known, as in Eq. 1:
Ft
F = (1)
µs
 Cooperative Slip Detection Using a Dual-Arm Baxter Robot 253

where µs is dependant on the frictional properties of the two materials, their

interaction and environmental factors [24,25]. In a non-laboratory environment,
it is usually impractical to have the mass, and rare to have the coefficient of
friction information available for different objects being manipulated.
In the case of this experimental platform, discussed in Sect. 2, the shoulder
joint S0 (see Fig. 2) acts as a fulcrum applying torque, T in Newton-meters.
F can be calculated from the torque at the joint, TS0 , and the perpendicular
distance from the axis of rotation to the point of contact, d, in meters, as:
TS0
F = (2)
d
The tangential force, Ft , pulling down on the object is given by:

Ft = m(g + a) (3)

Z-plane acceleration of grippers, a, must be taken into account when the robot
is in motion, along with g, acceleration due to gravity.
In order to apply the minimum grasping force to hold an object where µs
and m are unknown, F cannot be calculated. When Ft is greater than F , the
object will begin to move, therefore, the point at which minimum F is applied
can be assumed as the point of impending slippage [13]. TS0 , torque at shoulder
joint S0, and m is recorded for each test. Combining Eqs. 1, 2 and 3 gives the
relationship:

TS0 d(g + a)
= (4)
m µs
Therefore for an object of the same µs and d acting on the same system but
different m, it can be said that the ratio of threshold torque to mass remains
the same, that is:
TS01 TS02
= (5)
m1 m2

4 Experimental Methodology
The two end-effectors were commanded to identical Cartesian coordinates in the
X and Z planes in order to keep them at the same height and distance from
the robot. Both were also set to the same roll and yaw, with opposing pitch in
order to get the flat surface of the end-effectors to face each other to allow for a
grasping action as left and right F were increased.
Each object was held manually between the end-effectors and TS0 incre-
mented, decreasing the difference in Y until a firm grasp was achieved, in the
manner of Fig. 6. Torque was monitored to ensure a constant steady-state value
before slipping is induced.
The torque level on the right S0 joint was decremented manually using an
external hand-held controller until slipping was observed. The real-time torque
254 S. Trimble et al.

value of left joint W 0 was input into the filter discussed in Sect. 2.4 and the
output recorded, along with all left arm joint torque values. Slip quality was
monitored by the observer, watching for any sudden drops or unwanted parox-
ysmal movement, signifying a material that normal force is difficult to modulate
for. The observer description of the contact properties of the intersection between
the end-effectors and the object is recorded.
The experiment was repeated five times for each of the eight material/mass
combinations where slip could be successfully detected. The average slip thresh-
old torque values of left joint S0 and the filter output were recorded for each
repetition, and the ratio of these values to mass calculated to validate Eq. 5.

5 Results
Quantitative analysis of four of the materials where slip was successfully recorded
is provided in Table 2, which lists the type and description of materials used,
their masses, average threshold forces at which slip occurs, average peak values
of vibrations in joint W 0 during slip and ratios of threshold and vibration forces
to mass for each material. Figure 7 depicts torque (Nm) against time (samples)
for time windows, in which slip occurs, of the eight experiments where the top
rows (i) and (ii) correspond to the standard material whilst rows (iii) and (iv)
correspond to the same material but with additional mass. The plots reveal
the difference in slip threshold force at joint S0 and magnitude of vibrations
observed at joint W 0 as mass is increased and are typical of the materials where
a consistent good quality slip signal was observed. It can be seen that when the
slip threshold measured at S0 is crossed as in rows (i) and (iii) of Fig. 7, a spike

Fig. 7. Rows (i) and (ii) show the S0 force and absolute filter output respectively for
time windows, in which slip occurs, of the lighter mass experiments; rows (iii) and (iv)
for the heavier. The dashed lines mark the point of impending slip.
 Cooperative Slip Detection Using a Dual-Arm Baxter Robot 255

in the absolute filtered signal value W 0 is observed (see corresponding rows (ii)
and (iv)). The magnitude of the spike varies depending on a combination of mass
and surface friction properties.

Table 2. Experimental torque/mass analysis

Material Object Mass T threshold T threshold/mass Peak Peak/mass

description [kg] [Nm] [Nm/kg] [Nm] [Nm/kg]
Laminated card glossy, sticky 0.8270 8.3024 10.0392 3.2904 3.9787
1.3500 12.4563 9.2269 5.6785 4.2063
MDF high friction 0.9900 5.4730 5.5283 0.9243 0.9336
1.5150 9.7596 6.4420 2.6854 1.7725
PVC glossy, sticky 1.1330 5.5640 4.9109 4.4427 3.9212
1.6650 8.4640 5.0835 5.5422 3.3286
perspex glossy, sticky 0.7440 6.8950 9.2675 0.8750 1.1761
1.2730 10.9540 8.6049 1.0732 0.8430

Fig. 8. Average frequency magnitude compared to the mass for four different objects.

The average magnitude of vibration against object’s mass for the four materi-
als is depicted in Fig. 8, showing that as mass increases, the frequency component
of the force in the 9–15 Hz range during slip at W 0 increases.
Additional materials including sheets of steel, cardboard, polypropylene and
a hard card box were also tested but proved to be challenging with regard to slip.
These are excluded from Table 2 as due to their slip behaviour it was not possible
to acquire a satisfactory slip signal with such measurements. These issues are
further discussed in the following section.

6 Discussion
As can be expected, for objects with the same coefficient of friction where slip
could be modulated consistently, the magnitude of vibration is directly propor-
tional to the mass due to additional tangential force. There is also a relationship
256 S. Trimble et al.

between the threshold holding torque at joint S0 and this tangential force, as
seen in Table 2. This was true in general for the experiments either with a glossy,
but sticky material such as plastics or laminated card, and materials with a high
friction surface, such as MDF. The ratios of both normal force to mass and
vibration magnitude to mass are approximately adhered to, as per Eq. 5.
As mentioned earlier, it was not possible to detect slip for all the materials
tested due to the intermittent or sudden nature of the slip observed. ‘Slip-stick’
is a complex known problem due to dynamic instabilities [4]. Materials with this
characteristic are not best suited to a noisy moving robotic platform, and in the
case of this study they made joint S0 force control difficult to modulate.
The strip of cardboard exhibited behaviour in this manner due to it being an
easily compressible material with a non-uniform thickness. The polypropylene
material exhibited similar behaviour, because, although smooth, it was very
sticky and required a large mass in order to cause slip. This created undesirable
torsion forces acting on the end-effectors which disrupted the contact shape
between the material and the robot. The steel sheet, with it’s relatively high
mass and very slippery surface slid out of the grasp suddenly on all five occasions,
preventing a distinct slip signal from being observed. The three-dimensional hard
card box experienced changes in orientation as slipping occurred, thus shifting
the angle of the contact surfaces. This caused sporadic vibrations forces acting
on W 0, making a slip signal problematic to capture.

7 Conclusions and Future Work

In this paper a dual-arm Baxter collaborative robot with custom friction grip
end-effectors is introduced. Experiments were conducted with different materials
with varying surface friction properties and weight and real-time time-domain fil-
tering implemented on a joint torque signal to detect slip. Time domain analysis
of the filtered and threshold signals revealed that the approach was effective at
detecting slip for the laminated card, MDF, PVC and Perspex, as these materials
slipped in a smooth manner. The steel, cardboard and polypropylene exhibited
‘slip-stick’ behaviour, making it difficult to capture a slip signal. It would be
valuable to further investigate data in Fig. 7, by increasing data points with
additional materials and weights. In future work, improved signal conditioning
will be performed to enhance the slip detection signal. It would also be useful
to investigate the filter time delay, as the filtered torque signal is to be used as
a feedback signal for a fully automated closed-loop slip controller on the dual
manipulator system.
 Cooperative Slip Detection Using a Dual-Arm Baxter Robot 257

References
1. Gerling, G.J., Rivest, I.I., Lesniak, D.R., Scanlon, J.R., Wan, L.: Validating a
population model of tactile mechanotransduction of slowly adapting type I afferents
at levels of skin mechanics, single-unit response and psychophysics. IEEE Trans.
Haptics 7, 216–228 (2014)
2. Deng, H., Zhong, G., Li, X., Nie, W.: Slippage and deformation preventive control
of bionic prosthetic hands. IEEE/ASME Trans. Mechatron. 22, 888–897 (2017)
3. Xian, Z., Lertkultanon, P., Pham, Q.C.: Closed-chain manipulation of large objects
by multi-arm robotic systems. IEEE Rob. Autom. Lett. 2, 1832–1839 (2017)
4. Won, H.-I., Chung, J.: Numerical analysis for the stick-slip vibration of a trans-
versely moving beam in contact with a frictional wall. J. Sound Vibr. 419, 42–62
(2018)
5. Khurshid, R.P., Fitter, N.T., Fedalei, E.A., Kuchenbecker, K.J.: Effects of grip-
force, contact, and acceleration feedback on a teleoperated pick-and-place task.
IEEE Trans. Haptics 10, 40–53 (2017)
6. Erhart, S., Hirche, S.: Model and analysis of the interaction dynamics in coopera-
tive manipulation tasks. IEEE Trans. Rob. 32, 672–683 (2016)
7. Saen, M., Ito, K., Osada, K.: Action-intention-based grasp control with fine finger-
force adjustment using combined optical-mechanical tactile sensor. IEEE Sens. J.
14, 4026–4033 (2014)
8. Aqilah, A., Jaffar, A., Bahari, S., Low, C.Y., Koch, T.: Resistivity characteristics
of single miniature tactile sensing element based on pressure sensitive conductive
rubber sheet. In: 2012 IEEE 8th International Colloquium on Signal Processing
and its Applications, pp. 223–227, March 2012
9. Teshigawara, S., Ishikawa, M., Shimojo, M.: Development of high speed and high
sensitivity slip sensor. In: 2008 IEEE/RSJ International Conference on Intelligent
Robots and Systems, pp. 47–52, September 2008
10. Teshigawara, S., Tadakuma, K., Ming, A., Ishikawa, M., Shimojo, M.: High sensi-
tivity initial slip sensor for dexterous grasp. In: 2010 IEEE International Conference
on Robotics and Automation, pp. 4867–4872, May 2010
11. Engeberg, E.D., Meek, S.G.: Adaptive sliding mode control for prosthetic hands
to simultaneously prevent slip and minimize deformation of grasped objects.
IEEE/ASME Trans. Mechatron. 18, 376–385 (2013)
12. Schoepfer, M., Schuermann, C., Pardowitz, M., Ritter, H.: Using a Piezo-resistive
tactile sensor for detection of incipient slippage. In: ISR 2010 (41st Interna-
tional Symposium on Robotics) and ROBOTIK 2010 (6th German Conference
on Robotics), pp. 1–7, June 2010
13. Dzitac, P., Mazid, A.M., Ibrahim, M.Y., Appuhamillage, G.K., Choudhury, T.A.:
Optimal sensing requirement for slippage prevention in robotic grasping. In: 2015
IEEE International Conference on Industrial Technology (ICIT), pp. 373–378,
March 2015
14. Li, Z., Tao, P.Y., Ge, S.S., Adams, M., Wijesoma, W.S.: Robust adaptive control
of cooperating mobile manipulators with relative motion. IEEE Trans. Syst. Man.
Cybern. Part B (Cybern.) 39, 103–116 (2009)
15. Lotfavar, A., Hasanzadeh, S., Janabi-Sharifi, F.: Cooperative continuum robots:
concept, modeling, and workspace analysis. IEEE Rob. Autom. Lett. 3, 426–433
(2018)
16. Arshad, A., Badshah, S., Soori, P.K.: Design and fabrication of smart robots. In:
2016 5th International Conference on Electronic Devices, Systems and Applications
(ICEDSA), pp. 1–4, December 2016
258 S. Trimble et al.

17. Makrini, I.E., Rodriguez-Guerrero, C., Lefeber, D., Vanderborght, B.: The variable
boundary layer sliding mode control: a safe and performant control for compliant
joint manipulators. IEEE Rob. Autom. Lett. 2, 187–192 (2017)
18. Chen, B., Wan, J., Shu, L., Li, P., Mukherjee, M., Yin, B.: Smart factory of industry
4.0: key technologies, application case, and challenges. IEEE Access 6, 6505–6519
(2018)
19. ReThink Robotics: Baxter hardware specifications, May 2014. Accessed 14 Jan
2018
20. Cremer, S., Mastromoro, L., Popa, D.O.: On the performance of the baxter research
robot. In: 2016 IEEE International Symposium on Assembly and Manufacturing
(ISAM), pp. 106–111, August 2016
21. Rethink robotics opens up baxter robot for researchers (2018). Accessed 5 Feb 2018
22. Stachowsky, M., Hummel, T., Moussa, M., Abdullah, H.A.: A slip detection and
correction strategy for precision robot grasping. IEEE/ASME Trans. Mechatron.
21, 2214–2226 (2016)
23. Damian, D.D., Newton, T.H., Pfeifer, R., Okamura, A.M.: Artificial tactile sensing
of position and slip speed by exploiting geometrical features. IEEE/ASME Trans.
Mechatron. 20, 263–274 (2015)
24. Tuononen, A.J.: Onset of frictional sliding of rubber–glass contact under dry and
lubricated conditions. Sci. Rep. 6(1) (2016)
25. Xu, F., Yoshimura, K.-I., Mizuta, H.: Experimental study on friction properties of
rubber material: influence of surface roughness on sliding friction. Procedia Eng.
68, 19–23 (2013)
Intelligent Control and Automation
 Study on Swimming Curve Fitting
of Biomimetic Carangiform Robotic Fish

Baodong Lou1, Yu Cong2, Minghe Mao3(&), Ping Wang2,

and Jiangtao Liu1
1
School of Mechanical and Electrical Engineering, Hohai University,
Nanjing 210098, China
2
School of Energy and Electrical, Hohai University, Nanjing 210098, China
3
School of Computer and Information, Hohai University,
Nanjing 210098, China
[email protected]

Abstract. At present, the biomimetic carangiform robotic fish with infinite

flexible structure has not been presented. In order to facilitate the robotic fish
swimming mode control, based on a finite flexible structure a number of rigid
joints are used instead. In order to overcome the existing curve fitting method
which endpoint of each joint falls on the swimming curve, causing the deviation
between the fitting curve and the real swimming curve is larger. In this paper, a
minimum error criterion is proposed to fit the swimming curve for a three-joint
biomimetic carangiform robotic fish. Experimental results show that the pro-
posed method which causes the endpoints of robotic fish moving joints not
falling on the fitted swimming curve can obtain more accurate swimming
motion according to the real fish than that uses the conventional endpoints-on-
fitted-curve method. Moreover and thus, using the proposed method the
swimming velocity of the robotic fish is improved rapidly.

Keywords: Biomimetic carangiform robotic fish
Curve fitting

Motion equation
Minimum error

1 Introduction

After a long period of natural continuous choice, fabulous evolution of the fishes makes
their swimming ability meet the needs of survival. According to the different swimming
duration time the fish movement can be divided into two patterns: cruise and
maneuvering. The cruise refers to the long-term steady-state swimming over a long
period of time, while the maneuvering refers to generating a great acceleration or a
large steering torque in a short period of time. Also, the propulsion patterns can be
divided into two categories depending on the different parts of fish body that can
generate thrust: Body and/or Caudal Fin propulsion (BCF) and Media and/or Paired Fin
propulsion (MPF). The carangidae fish studied in this paper uses the BCF propulsion
pattern, which relies on the swing of the 1/3 parts of the body and the caudal fins, to
generate thrust. Based on the prototype of carangidea fish the study is implemented on
both structural appearance and athletic stance of swimming in this paper. The

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 261–271, 2018.
https://doi.org/10.1007/978-981-13-2384-3_24
262 B. Lou et al.

propulsive efficiency of fish is more than 80%. Therefore, in order to generate high
speed with high propulsive efficiency the motion control for the biomimetic robotic fish
should be well studied.
The kinematics theory of robotic fish has been studied since 1920s. In 1960,
Lighthill introduced the “slender body theory” in aerodynamics into the hydrodynamic
analysis of fish swim. Using carangidae fish as the research object, he initially estab-
lished the carangidae fish kinematic model. Experimental studies on fish dynamics
mainly focused on kinematic parameters and various forms of fish movement. In 1933,
Gary recorded the movement of fish by camera, and the theory of undulatory propulsion
from the motion of the cross section. Domenici and Blake studied the C-shape start of
the angel fish, suggesting that the size of the fish is independent of the duration of the
escape process but linearly related to the turning radius. In 1999, Spierts gave the C-
sharp and S-sharp turning curves of crucian carp, and studied the kinematics difference
between the two starting methods respectively [11]. In 2001, Rosenberger observed the
morphological parameters and swimming patterns of the myliobatis-shaped fish, and
compared the effects of pectoral fin motion parameters on the swimming speed and
propulsion performance. At present, there are two main control algorithms for robotic
fish motion control: Central Pattern Generator (CPG) and curve fitting. CPG is a control
algorithm using the neural network, which is based on the motion mechanism of the
animal, and it is used as the bottom controller to generate the swimming posture of the
robotic fish. Since CPG has a high degree of non-linear characteristic, which leads to big
obstacles for its theoretical research and practical application. Moreover, the basic
movement of the CPG is composed of a series of rhythmic movements, which can
respond quickly for the robotic fish cruise movement, but it is difficult to achieve its
maneuvering. The curve fitting is to fit the trajectory of the fish through each joint of the
robotic fish, which is suitable for all kinds of motion forms, meanwhile the design
process is simple and convenient. In the traditional curve fitting, the end points fall on
the curve in the fitting process, which lead to larger fitting error. In this paper, the
minimum error criterion is proposed for curve fitting as a metric.
In this paper, we use the curve fitting method to control the three-joint biomimetic
carangiform robotic fish Based on the metric of minimum error criterion. Therefore, the
end point of the joint does not fall on the fitted curve as the traditional methods.
Experimental results show that the minimum error curve fitting method can make the
robotic fish movement more smoothly and improve the velocity of biomimetic robotic
fish movement compared with the traditional curve fitting method.

2 The Kinematic Equations of Biomimetic Robotic Fish

According to Lighthill’s theory of slender body [8], the fish wave is a traveling wave
group with a gradually increasing wave amplitude. It can be approximated by a
polynomial and a sinusoidal curve:

FB ðx; tÞ ¼ ðc1 x þ c2 x2 Þ sinðxt þ kxÞ ð1Þ

 Study on Swimming Curve Fitting of Biomimetic Carangiform Robotic Fish 263

Where FB ðx; tÞ is the lateral position of the fish;

c1 is the primary coefficient of the fish wave envelope;
c2 is the quadratic term coefficient of the fish wave envelope;
k is the number of fish waves (k ¼ 2p=k, k is the wavelength of the fish body);
x is the frequency of fish wave (x ¼ 2pf ¼ 2p=T).

Through the biological studies we can see that the swing amplitude of the tail
motion joint is 0.075–0.1 times of the body length, and fish wave wavelength is usually
k  1LB , so the fish wave number is k ¼ 2p=k  2p=LB [16]. Therefore, in the absence
of special instructions, the parameters used in this paper are assumed as c1 ¼ 0:05,
c2 ¼ 0:03, k ¼ 8, x ¼ 2p=9. The values here are negative because the origin point of
the model is set the connection point of the head and tail movement joints, and then the
fish head direction is negative. The body wave motion equations of the biomimetic
robotic fish using FB ðx; tÞ can be drawn by MATLAB shown in Fig. 1. Each curve in
Fig. 1 corresponds to the movement posture of the robotic fish at a given moment, and
the fluctuation starting point of the tail movement joint is 0 point.

Fig. 1. Body wave curve cluster of biomimetic robotic fish

3 Kinematic Curve Fitting of Three-Joint Robotic Fish

Due to technical limitations, robotic fish can not be infinitely flexible at present, but
instead of flexible curves it can be approximated by finite rigid joints. Because of the
limitation of the biomimetic robotic fish joint length, it is difficult to use the analytical
method to fit the equation of the multi-joint biomimetic robotic fish’s body wave
motion. Therefore, the numerical approximation method is used. The method firstly
divides the body wave motion equation of the biomimetic robotic fish into N move-
ments with time varying as FB ðx; iÞði ¼ 0. . .. . .NÞ, and the gestures of each moment are
fitted using three rigid joints as ðl1 ; l2 ; l3 Þ. In this way, the fish postures at all times can
be replaced by approximated rigid joints of robotic fish. Figure 2 shows the robotic fish
motion at a certain moment during cruise, while Fig. 3 shows the body wave motion
curve corresponding to this posture.
264 B. Lou et al.

Fig. 2. Schematic diagram of biomimetic robotic fish at a certain time

In Fig. 3, hi;j ði ¼ 0. . .M  1; j ¼ 1. . .3Þ is the slip angle of each joint relative to the
horizontal line, /i;j ði ¼ 0. . .M  1; j ¼ 1. . .3Þ is rotation angle which the individual
joint of the biomimetic robotic fish need to be rotated, that is, the deflection angle of the
back joint relative to the front joint. For example, “/i;2 ” means the angle at which the
second joint is required to be rotated relative to the first joint at the time of the
movement posture FB ðx; iÞ. If this formula is expanded to a K-joint biomimetic robotic
fish, the result of numerical approximation for the rotation angle using the body wave
motion equation FB ðx; tÞ can be expressed as /i;j ði ¼ 0. . .M  1; j ¼ 1. . .KÞ.

Fig. 3. The motion curve of biomimetic robotic fish

In references [18, 19], a method of curve fitting is proposed. In the case of known
joint length, the starting point of the first joint coincides with the starting point of the
wave, and then the circle is rounded with the length of the joint. The intersection with the
curve is the coordinate of the other end point of the joint. The starting point of the second
joint is the end point of the first joint, and the end point coordinate of the second joint is
obtained by using the above mentioned method. Similarly, the coordinate of the third
joint is obtained, then the fitted curves of the three-joint robotic fish motion are obtained.
 Study on Swimming Curve Fitting of Biomimetic Carangiform Robotic Fish 265

In the process of fitting, the end of each joint falls on the theoretical fish wave curve,
which causes large fitting errors, and the fluctuation curve of biomimetic robotic fish is
not smooth enough to generate accurate biomimetic robotic fish swimming motion.
In this paper, the minimum error fitting method is proposed to fit the body wave
motion curve of biomimetic robotic fish. Assuming that the biomimetic robotic fish has
k joints with the length of each joint is lj ðj ¼ 1. . .kÞ; the starting coordinate of each
joint is ðxb ; yb Þ, the end point coordinate is ðxe ; ye Þ, which means that the starting
coordinate of the j th joint is the end coordinate of the j  1 th joint; finally, each joint
can be represented by a linear equation y ¼ gðxÞ. Then, the fitting error equation can be
written as:
 Z xe 
 
eðxÞ ¼  ½gðxÞ  FB ðxÞdx: ð2Þ
xb

Although the fitting error includes amplitude error and direction error, only the
amplitude error is considered in this paper, since the direction error is minimized to a
certain small value at the same time when the amplitude error is minimized. Formula
(6) represents only one joint fitting error, and it can be expanded to k th joint error
equation as:
Z 
 xei;j 
 
ei;j ðxÞ ¼  ½gi;j ðxÞ  FB ðx; iÞdx; ð3Þ
 xb 
i;j

where i ¼ 0. . .M  1; j ¼ 1. . .k,

gi;j ðxÞ ¼ ki;j x þ bi;j ð4Þ

ye yb
where ki;j ¼ xei;j xbi;j bi;j ¼ ybi;j  ki;j xbi;j :
i;j i;j

From the above analysis, we can see that the coordinates of each joint are related,
so the starting coordinates of each biomimetic robotic fish joint can be described by the
following equation:
8
>
> xb ¼ 0 j¼0
< y i;j ¼ 0 j¼0
bi;j
ð5Þ
>
> x b ¼ xei;j1 j ¼ 2. . .k:
: i;j
ybi;j ¼ yei;j1 j ¼ 2. . .k

The initial coordinate of the first joint is (0, 0), which is the connection point
between the biomimetic robotic fish head and the tail movement joints. The connection
point is the starting point of the fish wave motion. And since the length of each joint of
the biomimetic robotic fish is given, the following equation can be obtained:

ðxei;j  xbi;j Þ2 þ ðyei;j  ybi;j Þ ¼ l2j ð6Þ

266 B. Lou et al.

Therefore, the problem of curve fitting is transformed into the problem of finding
the joint end point coordinates in order to minimize the error equation ei;j ðxÞ when the
motion attitude equation FB ðx; iÞ with the starting point coordinate and length of the
joint are known at a certain moment. So that the following formula can be established:

dei;j ðxÞ
¼0 ð7Þ
dx

It is obviously impossible to solve Eq. (11) by simultaneous Eqs. (7), (8), (9) and
(10), because the constraint condition (9) and (10) themselves are interdependent and
not completely independent of each other. Therefore, we assume that a length pro-
portional correlation coefficient R 2 ½0; 1 the ratio between the joint length between the
starting point and the intersection of the rigid joint and the body wave curve and the
total length of the joint. Then, for an arbitrary value of R, the intersection of the joint j
and the motion curve ðxci;j ; yci;j Þ can be expressed by the following equation:

ðx  xbi;j Þ2 þ ðy  ybi;j Þ2 ¼ ðRlj Þ2
ð8Þ
y ¼ FB ðx; iÞ;

where: i ¼ 0. . .M  1; j ¼ 1. . .k
ðxci;j ; yci;j Þ can be obtained according to the above formula. Therefore, the linear
equation gi;j ðxÞ of the joint j can be easily obtained when knowing the coordinate of
starting point and the intersection coordinate. Rp ¼ pn ; p ¼ 0. . .:n is introduced to
calculate the minimum fitting error:

ei;jmin ¼ min ðei;j ðxÞjRp Þ ð9Þ

p¼0...n

The fitting curve is obtained using the above equations, and then the rotation angle
equation of each joint of the biomimetic robotic fish can be obtained by fitting the
curve. The rotation angle equation is given as:

/i;j ¼ hi;j j¼1
ð10Þ
/i;j ¼ hi;j  hi;j1 j ¼ 2. . .k

In order to control the motion of a robotic fish, the rotation angle /i;j at each
moment is integrated into a time varying equation /j ðtÞ. Since the joint rotation of the
robotic fish is a periodic motion that changes over time, the shape of the motion curve
of each joint is similar to a sine function. It can be expressed by Fourier series as:

X
1
/j ðtÞ ¼ an;j sinðnxt þ un;j Þ; j ¼ 1. . .k: ð11Þ
n¼1
 Study on Swimming Curve Fitting of Biomimetic Carangiform Robotic Fish 267

In order to limit the number of parameters to facilitate online optimization during

engineering debugging, the results only retain the first term, and the joint angle control
equations of the robotic fish can be obtained by discarding the higher order compo-
nents, one has:

/j ðtÞ ¼ aj sinðxt þ uj Þ; j ¼ 1. . .k ð12Þ

The fitting curves of the traditional fish joint are plotted by using the angle control
equation, as shown in Fig. 4. While the minimum error method for the robotic fish joint
curves are shown in Fig. 5. One can see that there are no significant differences on the
shape of the joint fitting curve using traditional fitting method and the proposed
minimum error method. However, the initial phase angle and amplitude of each joint
movement are different. And the difference between these two parameters will directly
affect both the swimming posture and swimming speed of the robotic fish.

Fig. 4. Traditional fitting curve

Fig. 5. Minimum error fitting curve

4 Experiment and Analysis

In order to verify the effect of the initial phase angle and amplitude of each joint on the
swimming posture and swimming speed of the biomimetic robotic fish, a three-joint
biomimetic carangiform robotic fish prototype is developed, as shown in Fig. 6, and
the following experiment was carried out in the experimental system shown in Fig. 7.
268 B. Lou et al.

Fig. 6. Three-joint biomimetic carangiform robotic fish

The whole system consists of five parts: a three-joint biomimetic carangiform

robotic fish, a PC client, an IDS industrial camera, a communication module and a
60 cm  12 cm  20 cm water tank. The three-joint biomimetic carangiform robotic
fish is composed of five parts: the head, the internal control warehouse, the movement
joint 1, the movement joint 2, the movement joint 3. The robotic fish communicates
with the PC client using a 433 MHz RF module.

Fig. 7. Experimental environment and equipment

The fitting curves shown in Figs. 4 and 5 are realized on the three-joint biomimetic
carangiform robotic fish prototype. Figure 8 is a set of continuous experimental pictures
using the traditional fitting curve method for experimentation. While Fig. 9 is a set of
 Study on Swimming Curve Fitting of Biomimetic Carangiform Robotic Fish 269

continuous experimental pictures using the proposed minimum error fitting curve method
for experimentation. Compared with the two sets of pictures, one can see that using the
minimum error fitting curve to control the robotic fish swimming trajectory has a higher
similarity with the swimming trajectory of real fish, and the curve is more smoothly in the
process of movement. Comparing the two sets of pictures with the corresponding theo-
retical curve, one can find that the swimming curve of the robotic fish with the minimum
error fitting curve is closer to the theoretical curve and has better linearity. Moreover,
comparing the duration time taken by robotic fish after swimming the same distance, it
can obviously be found that the duration time of the robotic fish using the minimum error
fitting curve method is less than that of using the traditional method. The average velocity
is improved by 7%, and the swimming posture is getting more stable.

Fig. 8. Traditional curve fitting experiment

Fig. 9. Minimum error curve fitting experiment

270 B. Lou et al.

5 Conclusion

In this paper, the curve fitting method of biomimetic carangiform robotic fish has been
studied, and the experimental results have been validated. The following conclusions
can be drawn as: A minimum error curve fitting method has been proposed for the
three-joint biomimetic carangiform robotic fish. The method does not require the end
point of the joint falling on the swimming curve. Experimental results have proved that
the proposed method can make the fish swimming curve closer to that of the real fish.
The fitted motion curve has been applied to a robotic fish prototype, which has proved
that the proposed method can make the robotic fish swim more smoothly and improve
the swimming speed rapidly.

References
1. Yu, J., Ding, R., Yang, Q., Tan, M., Wang, W., Zhang, J.: On a bioinspired amphibious
robot capable of multimodal motion. IEEE/ASME Trans. Mechatron. 17(5), 847–856 (2012)
2. Lu, X.Y., Yin, X.Z.: Propulsive performance of a fish-like travelling wavy wall. Acta Mech.
175(1), 197–215 (2005)
3. Hong, C.: Kinematic mechanism research on the swimming and maneuvering of robot fish.
Ph.D. dissertation, University of Science and Technology of China, China (2006)
4. Sfakiotakis, M., Lane, D.M., Davies, J.B.C.: Review of fish swimming modes for aquatic
locomotion. IEEE J. Oceanic Eng. 24(2), 237–252 (1999)
5. Hu, T., Lin, L., Zhang, D., Wang, D., Shen, L.: Effective motion control of the biomimetic
undulating fin via iterative learning. In: 2009 IEEE International Conference on Robotics
and Biomimetics (ROBIO), Guilin, pp. 627–632 (2009)
6. Low, K.H., Willy, A.: Biomimetic motion planning of an undulating robotic fish fin. J. Vib.
Control 12(12), 1337–1359 (2006)
7. Liu, J., Dukes, I., Knight, R., et al.: Development of fish-like swimming behaviours for an
autonomous robotic fish. Univ. Bath 15(1), 6–9 (2004)
8. Lighthill, M.J.: Note on the swimming of slender fish. J. Fluid Mech. 9(2), 305–317 (1960)
9. Taylor, G.: Analysis of the Swimming of Microscopic Organisms. Proc. Roy. Soc. Math.
Phys. Eng. Sci. 209(209), 447–461 (1951)
10. Domenici, P., Blake, R.W.: The kinematics and performance of the escape response in the
anglefish (Pterophyllum eimekei). J. Exp. Biol. 156, 187–205 (1991)
11. Spierts, I.L., Leeuwen, J.L.: Kinematics and muscle dynamics of C- and S-starts of carp
(Cyprinus carpio L.). J. Exp. Biol. 202(Pt 4), 393 (1999)
12. Rosenberger, L.J.: Pectoral fin locomotion in batoid fishes: undulation versus oscillation.
J. Exp. Biol. 204(2), 379–394 (2001)
13. Ijspeert, A.J.: Central pattern generators for locomotion control in animals and robots: a
review. Neural Netw. 21(4), 642–653 (2008)
14. Chen, W.H., Ren, G.J., Zhang, J.B., Wang, J.H.: Smooth transition between difierent gaits of
a hexapod robot via a central pattern generators algorithm. J. Intell. Rob. Syst. 67(3–4), 255–
270 (2012)
15. Chuan, L.K., Kit, T.W.: Task learning utilizing curve fitting method for kinect based
humanoid robot. In: 2014 IEEE International Conference on Robotics and Biomimetics
(ROBIO 2014), Bali, pp. 1505–1511 (2014)
 Study on Swimming Curve Fitting of Biomimetic Carangiform Robotic Fish 271

16. Liu, Y.-X.: The entity design and dynamic research on the two-joint robot fish. Master
dissertation, Harbin Institute of Technology, China (2007)
17. Liu, J., Hu, H.: Biological inspiration: from carangiform fish to multi-joint robotic fish.
J. Bionic Eng. 7(1), 35–48 (2010)
18. Liu, J., Hu, H.: Mimicry of sharp turning behaviours in a robotic fish. In: 2005 IEEE
International Conference on Robotics and Automation, pp. 3318–3323. IEEE (2005)
 H1 Filter Designing for Wireless Networked
Control Systems with Energy Constraints

Lisheng Wei1(&), Yunqiang Ma2, and Sheng Xu3

1
Anhui Polytechnic University, Wuhu 241000, An-Hui, China
[email protected]
2
Anhui Technical College of Mechanical and Electrical Engineering,
Wuhu 241002, An-Hui, China
3
Nantong Vocational University, Nantong 226007, Jiang-Su, China

Abstract. The issue of H1 filter designing for a class of Wireless Networked

Control System (WNCS) with communication constraints was discussed in this
paper. Firstly, the energy consumption of wireless sensor nodes in the data
acquisition stage was reduced by using time-varying sampling strategy. And the
wireless sensor nodes’ energy consumption was reduced based on reducing-
frequency in the data sending stage. Then the mathematical model of WNCS
was described as the Asynchronous Dynamic System (ADS). By using Linear
Matrix Inequality (LMI) technique, the sufficient condition of the WNCS to be
stability was obtained. Finally, the numerical example was given to show the
effectiveness of the proposed filter.

Keywords: Energy constraints  WNCS  Filtering  Time-varying sampling

1 Introduction

Over the past twenty years, WNCS has been widely expended in various kinds of
applications including remote operations, aircraft, marine monitoring, intelligent
transportation, and manufacturing plants [1]. These processes are often linked to pass
the data collected through the sending of a real-time wireless network medium. WNCS
has numerous advantages to energy saving optimization and actual industrial control
system [2]. Nevertheless, a great many new problems and difficulties are created for the
wireless network when it is applied to industrial control system. As the wireless net-
work is an unstable shared channel, it will make stability analysis difficult and decrease
the system control performance [3–5]. And the limited energy of wireless sensor nodes
will reduce the lifetime of the WNCS. Consequently, it is of enormous scientific
significance value to do research on the problem of filtering technique and controller
design for WNCS with the energy constraints [3].
Two of the central issues are delay and packet-dropout, which has been extensively
investigated in the last decade. A lot of researchers have tried to solve the control
performance with delay and packet-dropout [4, 5]. Several good results have been
available in the existing literatures. For instance, Zhen et al. [6] researched the problem
of modeling and stabilization for WNCS with both time-varying delay and packet-
dropout, which are modeled as an asynchronous dynamic system with three subsystems.

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 272–280, 2018.
https://doi.org/10.1007/978-981-13-2384-3_25
 H∞ Filter Designing for Wireless Networked Control Systems 273

And the sufficient condition for the system was obtained. Peng et al. [7] investigated the
problem that the energy efficiency could be much improved while preserving the desired
H1 performance. And the self-triggered sampling scheme was proposed. Dong et al. [8]
designed the robust filter for a class of discrete-time WNCS with communication
constraints and packet dropouts by using switched system theory. In conclusion, the
majority of the above researches center on the issues of the delay and packet losses. The
filtering technology is adopted in the external noise disturbance. On the foundation of
the research achievements above, we are aimed at investigating the problem of filter
designing and stabilization analysis for a class of WNCS with energy constraints.
The rest of paper is organized as follows. In Sect. 2, we propose a method to build a
new modeling of WNCS with energy constraints. In Sect. 3, the sufficient conditions of
the H1 filtering are given. In addition, and the gain used in the H1 filtering can be
derived by LMI technology. Then the effectiveness of the proposed method is proved
by the simulation results in Sect. 4. At last, the conclusions are given in Sect. 5.

2 Problem Formulations

The block diagram of a class of WNCS with filter is shown in Fig. 1. We can know that
wireless sensor nodes contains a fixed topology previously, and transmits data from
plant to filter nodes by reducing-frequency transmission.

Fig. 1. The Block diagram of WNCS

Because the energy in WNCS is limited, the energy saving strategy is shown in
Fig. 1 to prolong the service life of wireless sensor. The time-varying sampling is used
to reduce the consumption of data acquisition of wireless sensor nodes, and then the
reducing-frequency transmission is used to save the energy consumption of data
transmission phase. The plant R1 can be described as follows:
8

< x_ ðtÞ ¼ AxðtÞ  1 wðtÞ
þB

yðtÞ ¼ CxðtÞ þ B 2 wðtÞ ð1Þ
: 
zðtÞ ¼ C0 xðtÞ
274 L. Wei et al.

where A, B1 , B2 , C and C 0 are constant matrices with appropriate dimensions.

xðkÞ¼½x1 ; x2 ;    ; xm T 2 Rm is the system state vector. yðkÞ ¼ ½y1 ; y2 ;    ; yn T 2 Rn is
 T
the output vector. zðkÞ ¼ z1 ; z2 ;    ; zp 2 Rp is the estimated matrix of system and
wðkÞ 2 l2 ½0; 1Þ is the disturbance noise with limited energy.

2.1 Energy Saving Strategy Based on Time-Varying Sampling

Without loss of generality and for convenience of discussion, the following assump-
tions are given for system (1) [8].
(1) Time driven is used in sensor nodes;
(2) The network induced delay is less than one sampling period;
(3) tk represents the system sampling time, and k ¼ 0; 1; 2      . hk represents the
system sampling period at the kth time. Then we have hk ¼ tk þ 1  tk 2 ½d1 ; d2 .
d1 is the allowable minimum sampling period of WNCS when the energy is
sufficient. d2 is the maximum allowable sampling period of WNCS when the
energy is limited.
Using active variable sampling period strategy is applied to analyze time-varying
sampling process of the system. The sampling period ½d1 ; d2  is divided into c equal
interval, where c is positive integer. If the current sampling time is tk , then the next
sampling time tk þ 1 is satisfied the following Equation.
8
< b 1 ; k ¼ t k þ b1
t k þ 1 ¼ b 2 ; k 2 ð t k þ b1 ; t k þ b2  ð2Þ
:
d2 ; k [ tk þ d2

where b1 ¼ d1 þ aðd2 d d2 d1

c Þ; b2 ¼ d1 þ ða þ 1Þð c Þ; a ¼ 0; 1; 2;    ; c  1. Then we
1

have the Eq. (3).

d2  d1
hk ¼ d1 þ eð Þ e ¼ 0; 1; 2;    ; c: ð3Þ
c

So the range of the system sampling period hk is in the finite set X ¼ fd1 ;    ;
d1 þ eðd2 d
c Þ;    ; d2 g.
1

The plant R1 can be described as follows:

8
< xðk þ 1Þ ¼ AxðkÞ þ B1 wðkÞ
yðkÞ ¼ CxðkÞ þ B2 wðkÞ ð4Þ
:
zðkÞ ¼ C0 xðkÞ

 R hk 
where A ¼ eAhk ; B1 ¼ 1.
eAS dsB
0
R Dh  
Setting Dh ¼ hk  d1 and Fk ¼  1 ðeADh
eAS ds ¼ A  IÞ, then we have
0
  k þ I.
eADh ¼ AF
 H∞ Filter Designing for Wireless Networked Control Systems 275

     1
 k þ I Þ ¼ eAd   ~
A ¼ eAðd1 þ DhÞ ¼ eAd1  eADh ¼ eAd1 ðAF þ eAd1 AF k ¼ A þ M1 Fk N1 ð5Þ
 
~ ¼ eAd1 ; M1 ¼ eAd1 A;
 N1 ¼ I.
where A
Similarly, we have
Z d1 Z hk
 
B1 ¼ 1 þ
eAS dsB 1
eAS dsB
0 d1
Z d1 Z hk d1
 
¼ AS 1 þ
e dsB 1
eAðS þ d1 Þ dsB
0 0 ð6Þ
Z d1 Z Dh
  1 
¼  1 þ eAd
eAS dsB 1
eAS dsB
0 0
~ þ M2 F k N 2
¼B
R  R
~ ¼ d1 eAS
where B dsB
 1
 1 ; M2 ¼ eAd  1 ; Fk ¼ Dh eAS
; N2 ¼ B

ds.
0 0
According to the above analysis, the discrete model of system (4) can be described
as:
8    
< xðk þ 1Þ ¼ A ~ þ M1 Fk N1 xðkÞ þ B~ þ M2 Fk N2 wðkÞ
yðkÞ ¼ CxðkÞ þ B2 wðkÞ ð7Þ
:
zðkÞ ¼ C0 xðkÞ
R 
where A
 1
~ ¼ eAd  
; M1 ¼ eAd1 A; ~ ¼ d1 eAS
N1 ¼ I; B dsB
 1
 1 ; M2 ¼ eAd  1 ; Fk ¼
; N2 ¼ B
R Dh AS

0

0 e ds
For the system (7), rmax ðAÞ is represent the largest singular value of system
T
parameters. Then Fk Fk  k I, k [ 0 and satisfied the following Equation.
2

 
ermax ðAÞðd2 d1 Þ  ermax ðAÞð0Þ
k¼  ð8Þ
rmax ðAÞ

2.2 Energy Saving Strategy Based on Reducing-Frequency Transmission

Considering the energy limitation of WNCS, the reducing-Frequency Transmission
strategy can be used in this paper. The energy consumption of sensor nodes can be
saved by reducing the amount of transmission data, that is, only updated data is allowed
to be sent to filter nodes at the sampling time, and the remaining unupdated data are
keeping the last sampling data. To simplify the description system, the function #ðkÞ 2
N; N , f1; 2;    ; ng is introduced to represent the state of sending data from sensor
nodes to the filter nodes. If #ðkÞ ¼ d, that is the dth sensor node can send the updated
data to filter node, then the data of filter node is yd ðkÞ ¼ yd ðkÞ, the data of other filter
node keep the last sampling data yd ðkÞ ¼ yd ðk  1Þ. Setting the matrix form
N#ðkÞ 2 fN1 ; N2 ;    ; Nn g representing the variable function #ðkÞ, then we have
276 L. Wei et al.

N#ðkÞ ¼ diag½dðd  1Þ; dðd  2Þ    dðd  nÞ, d is the Kronecker delta function. The
data input of the filter for WNCS with energy constraints can be described as:

yd ðkÞ ¼ #ðkÞyd ðkÞ þ ð1  #ðkÞÞyd ðk  1Þ

  ð9Þ
yðkÞ ¼ N#ðkÞ yðkÞ þ I  N#ðkÞ yðk  1Þ

Considering the following filter,


xf ðk þ 1Þ ¼ Af xf ðkÞ þ Bf yðkÞ
ð10Þ
zf ðkÞ ¼ Cf xf ðkÞ

where Af , Bf and Cf are constant matrices of filter with appropriate dimensions.

 T 
xf ðkÞ¼ x1f ; x2f ;    ; xmf 2 Rm , yðkÞ ¼ ½y1 ; y2 ;    ; yn T 2 Rn and zf ðkÞ ¼ z1f ; z2f ;
   ; zp f T 2 Rp are state vector, input vector and output vector respectively.
Combining the Eqs. (7), (9) and (10), the closed loop model of the filter error can
be expressed by:

Xðk þ 1Þ ¼ UXðkÞ þ BwðkÞ
ð11Þ
eðkÞ ¼ CXðkÞ
h iT
where XðkÞ ¼ xT ðkÞ xTf ðkÞ yT ðk  1Þ is the state augmented vector. eðkÞ ¼
zðkÞ  zf ðkÞ is the error vector. And
2 3 2 3
~ þ M1 Fk N1
A 0 0 ~ þ M2 Fk N2
B
 
U¼ 4 Bf N#ðkÞ C Af Bf I  N#ðkÞ 5; B ¼ 4 Bf N#ðkÞ B2 5
N#ðkÞ C 0 I  N#ðkÞ N#ðkÞ B2
 1
~ ¼ eAd  
C¼½ C0 Cf 0 ; A ; M1 ¼ eAd1 A; N1 ¼ I
Z d1 Z Dh
  1 
~¼
B AS  1 ; M2 ¼ eAd
e dsB  1 ; Fk ¼
; N2 ¼ B eAS ds
0 0

Considering the scheduling matrix N#ðkÞ is a random matrix with time varying, the
closed loop filtering error system can be described as a discrete system with n mode. In
order to design the H1 filter for the discrete control system (7), which is shown in
Eq. (10), we should get the following two conditions [7].
(1) When wðkÞ ¼ 0, the closed loop model of the filter error (11) is stochastic sta-
bility, if and only if

X
1
li m xðkÞT xðkÞ\1 ð12Þ
0
 H∞ Filter Designing for Wireless Networked Control Systems 277

(2) At the condition of zero initial state, the filtering error system (11) is used to suppress
the disturbance level c [ 0, then the H1 performance c can be described as:

keðkÞk22  c2 kwðkÞk22 ð13Þ

P
1 P
1
where keðkÞk22 ¼ eðkÞT eðkÞ , kwðkÞk22 ¼ wðkÞT wðkÞ.
k¼0 k¼0

3 Main Results

Our work is aiming at designing the filter for WNCS so as to solve the problems of
energy constraints. And prior to giving a detailed introduction of the main results, the
sufficient convergence condition for WNCS is given.
Theorem 1: Considering the system (11), given positive constant c [ 0. Then the
augmented system is asymptotical stable if there exist symmetric constant matrix
Pj [ 0, Pi [ 0 and
2 3
Pi   
6 0 c2 I   7
6 7\0 ð14Þ
4 U B Pj 1
 5
C 0 0 I

where  denotes the symmetric terms and

2 3 2 3
~ þ M1 Fk N1
A 0 ~ þ M2 Fk N2
B
 0 
U ¼ 4 Bf Nu C Af Bf I  Nu 5; B ¼ 4 Bf Nu B2 5 ; C¼½ C0 Cf 0
Nu C 0 I  Nu Nu B2

Theorem 2: Considering the system (7), given a positive scalar c [ 0 and e [ 0, if

there exist Pi [ 0 and Pj [ 0 ði; j 2 ½1;    ; N Þ, J11 ; J12 ; J13 ; J22 ; J31 ; J32 ; J33 ; V
and W with proper dimensions such that the LMI (15) holds.
2 3
P~  
4M~ k1 e1 I  5\0 ð15Þ
~T
N 0 k1 eI

Hence the system (11) is asymptotically stable as well as having the H1 perfor-
mance index.
where
2 3
P1   
6  7
~ 1 ¼ 6 0 c I 
2
P 7
4 P2 P3 P4  5 ; P1 ¼ Pi
P5 0 0 I
278 L. Wei et al.

2   3
T ~ T T
J11 A þ V N#ðkÞ C þ J31 N#ðkÞ C U V I  N#ðkÞ þ J31 ðI  N#ðkÞ Þ
6 T~ T
  T 7
P2 ¼ 4 J12 A þ V N#ðkÞ C þ J32 N#ðkÞ C U V I  N#ðkÞ þ J32 ðI  N#ðkÞ Þ 5
T ~ T T
J13 A þ J33 N#ðkÞ C 0 J33 ðI  N#ðkÞ Þ
2 3
T ~ T
J11 B þ VN#ðkÞ B2 þ J31 N#ðkÞ B2
6 T~ T 7
P3 ¼ 4 J12 B þ VN#ðkÞ B2 þ J32 N#ðkÞ B2 5
T ~ T
J13 B þ J33 N#ðkÞ B2

P4 ¼ Pj  J  J T ; P5 ¼½ C0 W 0
T
2 3 2
T
3
J11 M1 0 0 J11 M2 0 0
P6 ¼ 4 0 0 0 5 ; P7 ¼ 4 0 0 05
0 0 0 0 0 0
2 3 2 3
0 0 0 0 ^1 0
N 0 0
6 0 0 0 07 6 0 N ^2 0 07
~ ¼6
M ~ ¼6
7; N 7
4 P6 P7 0 05 4 0 0 0 05
0 0 0 0 0 0 0 0
2 3 2 3 2 3
N1 0 0 N2 J11 J12 J13
^1 ¼ 4 0
N 0 ^ 2 ¼ 4 0 5; J ¼ 4 J22
0 5; N J22 0 5
0 0 0 0 J31 J32 J33

T T
And the parameters matrix of filter are Af ¼ J22 U; Bf ¼ J22 V and Cf ¼ W.
Remark: The optimal performance c can be obtained by solving the convex opti-
mization problem as follows:

min l subject to ð15Þ with l ¼ c2

J11  J33 ; Pi ; Pj ; U; V; W
pffiffiffiffiffiffiffiffiffi
So the optimal performance is c ¼ lmin .

4 Numerical Simulations

In this segment, an illustrative instance, taking the following discrete-time control

systems into account, is showed to testify the validity of the raised theorems.
 
12 7 13
x_ ðtÞ ¼ xðtÞ þ wðtÞ
9 10 8
 
0:6 0:2 0:8
yðtÞ ¼ xðtÞ þ wðtÞ
0:3 0:4 0:3
zðtÞ ¼ ½ 0:5 0:5 xðtÞ
 H∞ Filter Designing for Wireless Networked Control Systems 279

Assuming the minimum allowed sampling period of the system d1 ¼ 0:1s when
WNCS energy is sufficient, and the maximum allowed sampling period of the system
d2 ¼ 0:2s when WNCS energy is limited. The sampling period h1 ¼ 0:1s, h2 ¼ 0:15s
and h3 ¼ 0:2s. So the sampling period is randomly switched over in a finite set
X ¼ f0:1; 0:15; 0:2g. Setting #d ðkÞ ¼ f1; 2; 1; 2;   g, N1 ¼ diag½ 1 0  and
N2 ¼ diag½ 0 1 , then by using Eq. (8), we have k ¼ 0:3014. When choosing the
noise wðkÞ ¼ 2e0:3k sinð0:5pkÞ and using theorem 2 and Matlab LMI Toolbox, we
obtain the optimal suppression level c ¼ 0:6427. At the same time, the feasible
solution of a set of H1 filter parameters under the optimal suppression level state is
obtained as follows.
 
0:1586 0:0983 1:6748 0:5342
Af ¼ ; Bf ¼
0:2366 0:1466 1:1469 0:3615

Cf ¼ ½0:6331  0:6664

Assuming that the initial values xf ð0Þ ¼ ½ 0 0 T and xð0Þ ¼ ½0:5  0:5T , the state
tracking curves of system and filter can be shown in Figs. 2 and 3 respectively.

1 0.5
x 1(k) x 2(k)

x 1f (k) x 2f (k)

0.5
x (k)and x (k)
x (k)and x (k)

2f
1f

0
2
1

-0.5
-0.5

0 10 20 30 40 0 10 20 30 40
k k

Fig. 2. State trajectory of x1 ðkÞ and x1f ðkÞ Fig. 3. State trajectory of x2 ðkÞ and x2f ðkÞ

In Figs. 2 and 3, x1 ðkÞ and x2 ðkÞ are the state of plant, and x1 f ðkÞ and x2 f ðkÞ are
the state of filter. It can be observed that xif ðkÞ can track xi ðkÞ effectively in 20 time
steps.

5 Conclusions

In this paper, the filtering problem for WNCS with energy constraints was considered.
By using time-varying sampling strategy, the energy consumption of wireless sensor
nodes in the data acquisition stage was reduced. And the mathematical model of
280 L. Wei et al.

WNCS could be described as an asynchronous dynamic system. The sufficient con-

ditions of H1 filter were derived. Then by the simulation, the effectiveness of the
proposed method was proved.

Acknowledgments. This work was supported by natural science research programme of col-
leges and universities of an-hui province under grant KJ2016A062, natural science foundation of
an-hui province under grant 1608085MF146, foundation for talented young people of an-hui
polytechnic university under grant 2016BJRC008, visiting study Foundation for outstanding
young talent of an-hui educational committee under grant gxfxZD2016108, and the applied basic
research plan of nan-tong under grant GY12017015.

References
1. Ge, X., Yang, F., Han, Q.-L.: Distributed networked control systems: A brief overview. Inf.
Sci. 380, 117–131 (2017)
2. Zhang, X., Han, Q.: Event-triggered H∞ control for a class of nonlinear networked control
systems using novel integral. Int. J. Robust Nonlinear Control 27(4), 679–700 (2017)
3. Sadi, Y., SC, E.: Joint optimization of wireless network energy consumption and control
system performance in wireless networked control systems. IEEE Trans. Wirel. Commun. 16
(4), 2235–2248 (2017)
4. Bai, J., Renquan, L., Hongye, S., Xue, A.: Modeling and control of wireless networked
control system with both delay and packet loss. J. Franklin Inst. 352(10), 3915–3928 (2015)
5. Makled, E.A., Halawa, H.H., Daoud, R.M., Amer, H.H., Refaat, T.K.: On the perform ability
of hierarchical wireless networked control systems. Intell. Control Autom. 6(2), 126–133
(2015)
6. Hong, Z., Gao, J., Wang, N.: Output-feedback controller design of a wireless networked
control system with packet loss and time delay. Math. Probl. Eng. 1, 1–7 (2014)
7. Peng, C., Han, Q.: On designing a novel self-triggered sampling scheme for networked
control systems with data losses and communication delays. IEEE Trans. Ind. Electron. 63(2),
1239–1248 (2016)
8. Dong, H., Wang, Z., Gao, H.: Robust H1 filtering for a class of nonlinear networked systems
with multiple stochastic communication delays and packet dropouts. IEEE Trans. Signal
Process. 58(4), 1957–1966 (2010)
 Distance Overestimation Error Correction
Method (DOEC) of Time of Flight Camera
Based on Pinhole Model

Le Wang, Minrui Fei, Hakuan Wang(&), Zexue Ji, and Aolei Yang

Shanghai Key Laboratory of Power Station Automation Technology,

School of Mechatronic Engineering and Automation, Shanghai University,
Shanghai 200072, China
{wangle,mrfei,HKWang,Zexueji,aolei}@shu.edu.cn

Abstract. Depth cameras with Time of Flight (ToF) technology are widely
used in machine vision and various measurement tasks. However, due to
hardware conditions and imaging characteristics, multiple errors limit the further
application of the ToF camera. This paper classified errors into errors caused by
non-imaging principle and error caused by imaging principle. In order to sim-
plify the experimental procedure and improve the efficiency of errors correction,
a simple and feasible method is used to correct errors caused by non-imaging
principle, and an evaluation function is proposed to determine the optimal ref-
erence distance, so as to select appropriate integration time and global offsets.
To tackle the radial distance error, Distance Overestimation Error Correction
method (DOEC) based on the principle of pinhole imaging is proposed to
further improve the accuracy of depth data. Finally, error correction methods
proposed in this paper are verified by experiments, and the segmentation of
different depth ranges is successfully achieved by using the modified data, prove
the effectiveness of the proposed methods.

Keywords: Time of flight
Error correction
Pinhole imaging

Plan segmentation
Data accuracy

1 Introduction

With the rapid development of machine vision system, different types of cameras play
significant roles in industrial intelligence more and more. Because of compact structure
and being less sensitive to illumination, time-of-flight (ToF) camera is widely applied
into many fields, such as mobile robot avoidance and navigation [1], object detection
and recognition [2, 3], 3D reconstruction and gesture recognition [4, 5].
Time-of-flight camera is a kind of active sensors, it emits near infrared (NIR) and
receive reflected light [6], and calculate the distance from the optical center of the
camera to an target object surface. There are some noise and errors exist in TOF camera
[7, 8], which cause negative impact on the accuracy of distance measurement [9]. Thus
it is essential to correct the data error of ToF camera at the initial procedure. There are
several source of errors in ToF camera, generally the error can be classified as sys-
tematic errors and non-systematic errors [10]. Systematic errors can be predicted and

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 281–290, 2018.
https://doi.org/10.1007/978-981-13-2384-3_26
282 L. Wang et al.

corrected by different calibration methods, but the non-systematic errors are complexed
and unpredicted and generally removed by filtering.
Several phenomena from the idealistic model of ToF camera are analyzed, but the
temperature, multi-path reflections and ambient light effects are still hard to solve [11].
An error model is utilized to estimate ground and wall planes, acquisition geometry and
noise characteristic are analyzed in this error model [12]. A 2.5D pattern board with
holes is used to capture both color and ToF depth images for feature detection. The
transformation of the ToF camera is reset. In order to get accurate bias after ray
correction, k-means clustering and B-spline functions are utilized [1].
In ToF camera, the radial distance is acquired between the optical center to an target
object surface, which is longer than the actual distance. To remove the radial distance
error, this paper divides the various errors in the TOF camera into errors caused by the
non-imaging principle and error caused by the imaging principle. When compensating
errors caused by the non-imaging principle, data error caused by temperature changes
and integration time are analyzed and corrected, multipath errors and other noises are
compensated by taking offsets, and an evaluation function is proposed to determine the
optimal reference distance. To tackle the radial error caused by imaging principle,
Distance Overestimation Error Correction method (DOEC) based on the principle of
pinhole imaging is proposed. By analyzing the positional relationship between the
center pixel and other pixels in depth image, the angle relationship between the two is
obtained, and the radial distance is convert to vertical distance. Thus the right distance
of ToF camera is obtained.
The remainder of this paper is organized as follows. Section 2 is errors caused by
non-imaging principle in ToF camnera. Section 3 corrects the radial error caused by the
imaging principle. Section 4 states the experimental results.

2 Errors Caused by Non-imaging Principle in ToF Camnera

TOF cameras are able to obtain distance information by utilizing the time of flight
measurement of photons. Photons are emitted by modulated infrared light. The mod-
ulated infrared light determines the phase shift u between reference signal and reflected
light. where c is the speed of light, f is the modulation frequency. The distance value d
is formulated as:

1 u
d¼ c ð1Þ
2 2pf

The temperature of the chip changes with the operating time, which will cause drift
of depth data, especially when the temperature of the chip changes significantly when
the ToF camera is just started. In order to reduce the deviation caused by temperature
changes, warm-up time of the chip is required before the measurement of ToF camera.
Observed from the experimental results of EPC660 chip, it takes about 15–20 min to
warm up before the depth image is acquired, and the obtained depth data is relatively
more stable and accurate.
 DOEC Method of Time of Flight Camera Based on Pinhole Model 283

Integration Time (IT) is the time span when ToF camera computes multiple phase
of photons and obtain the distance. Thus the Integration Time is important to the
accuracy of ToF camera data. In this paper, when the integration time is 1500 ls, the
effect of integration time on the depth data is studied by taking pictures of flat walls at
different distances. As shown in Fig. 1(a), (b), (c), (d) is depth images acquired at
different distance. Different colors represent different distances and black represents
invalid points. It can be seen that when the integration time is fixed, the larger the
measurement distance is, the more invalid points are, and the more pixels that appear
jitter at the edge of the image, the accuracy of the measurement decreases. Therefore, it
is necessary to select an appropriate integration time.

(a)Depth image of 3meter (b)Depth image of 3.3meter

(c)Depth image of 3.6meter (d)Depth image of 3.9meter

Fig. 1. Depth image of different distance with integration time is 1500 ls (Color figure online)

Multipath errors and other noises can be eliminated through precise modeling, but
the implementation process is more complex. Some nonlinear errors will change with
distance, which are difficult to solve. In order to simplify the errors correction proce-
dure, this paper sets offsets to correct such errors.
Through the above analysis of the source of error, this paper corrects the error of
ToF camera under the range of 0–4 m. Since the nonlinear errors are related to the
distance, 1 m, 2 m, 3 m, and 4 m are utilized as the distance baseline for the experi-
ment. The experiment steps of errors correction caused by non-principle are shown as
follows, where 1 m is sets as the distance baseline firstly.
Step1: Warm up before starting. To reduce errors due to temperature changes,
image acquisition should start after 20 min so that the ToF camera has enough time
to warm up, meanwhile the data fluctuation range is small.
284 L. Wang et al.

Step2: Adjust the integration time. The integration time was adjusted enable the
amplitude at 1 m distance baseline reach about 1000LSB regarded as suitable
amplitude. Integration time did not change during the subsequent experiment.
Step3: Set the global offset. Since the data accuracy of the pixel at the optical center
is less affected by various errors than other pixels in the depth image, the trueness of
the data is higher. Therefore, the measured depth data at the optical center denoted
as basemeai is used to compute the global offset at the distance baseline of 1 m. i is
the representation of different distance baseline from 1 m to 4 m. The value of
global offseti can be formula as follow:

offseti ¼ distancebaselinei  basemeai i ¼ 1; 2; 3; 4 ð2Þ

Where the distancebaseline1 is 1 m. After the value of global offset1 is acquired, it

can be used to compensate errors caused by multiple path and other noises.
Step4: Measure the depth data at other distances. To analyze the data accuracy at
different distance after the integration time and global offset were set at the distance
baseline of 1 m, the camera position was adjust at the actual distance of 2 m, 3 m,
and 4 m respectively, denoted as acti . Measured distance data of optical center at
different actual distance were marked and denoted as meai .
Step5: Calculate the evaluation function v. To evaluate the effect on data accuracy
by changing integration time and global offset set at different distance baseline,
evaluation function v is used and formula as follow:

X
4
v ¼ ðacti  meai Þ2 i ¼ 1; 2; 3; 4 ð3Þ
i¼1

If the difference between the actual distance and measured distance is smaller,
indicates that the accuracy of the measured data is higher, so the value of v is
smaller.
Step6: Repeat the above process by changing the distance baseline to 2 m, 3 m, and
4 m respectively.
Experiment results are shown in Table 1. When the distance baseline is 2 m, the
value of evaluation function v is the smallest than that of other distance baseline. It
indicates that the distance data obtained are more reliable, and errors caused by the
non- imaging principle have been improved to some extent. Thus the integration time
and global offset set at the distance baseline of 2 m are chosen as the better experiment
parameters in the 0–4 m operating range of ToF camera. Depth data acquired at
distance baseline of 2 m are also saved as data sample, which are used to correct the
errors caused by imaging principle in the next section.
 DOEC Method of Time of Flight Camera Based on Pinhole Model 285

Table 1. Experiment results and value of v

Actual distance meai at 100 cm meai at 200 cm meai at 300 cm meai at 400 cm v
100 cm 100 98.7 97.1 94.7 62.82
200 cm 203.5 200 198.1 196.3 16.61
300 cm 305.4 301.4 300 298.2 19.31
400 cm 407.1 403.6 402.7 400 45.02

3 Error Caused by Imaging Principle

After correcting errors caused by the non-imaging principle, a more accurate depth
image is obtained. In this section, we analyze the error caused by the imaging principle
and propose a Distance Overestimation Error Correction method (DOEC) based on the
pinhole imaging principle.

3.1 Distance Overestimation Error

The distance overestimation error means that the distance measured by ToF camera is
longer than the actual distance. In actual measure process, as shown in Fig. 2, there are
three test points A; B; C on the same object plane need to be measured, just imaging
that there are three test points A; B; C on a smooth wall, test point B is the intersection
of the line started from optic center OC and the object plane. The straight line distance
of test point A, test point B, test point C between optic center OC denoted as rA , rB , rC ,
and the vertical distance between object plane and camera plane is denoted as d. it is
clear to see that d is equal to rB , rA and rC are longer than the true distance rB , this
phenomena is called distance overestimation error. The overestimated distance rA and
rC can be corrected by the following formula:

d ¼ rB ¼ rA  sin hA ¼ rC  sin hC ð4Þ

Where hA is the angle between the straight line rA to camera plane, hC is the angle
between the straight line rC to camera plane. This kind of distance overestimation will
have a great impact on reconstruction of the 3D image.

A d
rA
A

B rB

C rC

Fig. 2. Distance overestimation error Fig. 3. Distance data at 1 m of wall

286 L. Wang et al.

Experiment has been done in front of smooth wall, the vertical distance between the
wall to the camera optical center Oc is 1 m. Figure 3 is the scatter diagram of the wall
after using the depth data collected from the experiment. It can be seen that the farther
the test point is away from the optic center Oc , the more obvious the overestimation of
the depth in the 3D image is, that means the depth value of points near the optical center
are closer to 1 m and the depth value near edge points are greater, even though they are
in same plane. So it is necessary to compensate such error in accurate measurement.

3.2 Distance Overestimation Error Correction Method (DOEC)

Aim to remove the distance overestimation error, this paper proposed Distance
Overestimation Error Correction method (DOEC) by exploiting the principle of pinhole
imaging, a model is built to correct such error, which is shown in Fig. 4. Where
imaging plane, camera plane and object plane are shown in Fig. 4. Ob is the center
point of imaging plane, Oc is the optic center of camera plane, and Oa is the center
point of object plane. p is an arbitrary point on object plane, p0 is the corresponding
point of p on imaging plane. f is the focal length, d is the vertical distance from the
object plane to the depth camera plane, r is the actual distance measured by the depth
camera. h is the angle between the laser light and the line started from optic center Oc .

p
ob oc
r oa
θ
θ
p
f d
Fig. 4. Imaging model of ToF

According to the principle of linear propagation of light, it is easy to know that

h ¼ h0 . The actual distance of camera plane and object plane is d, however, under the
impact of distance overestimated, distance of point p to optic center Oc is r, which is
longer than d. The geometric relationship between the vertical distance d and the TOF
camera measurement r can be expressed as (5):

d ¼ r  cos h ð5Þ

The distance of each point can be corrected if h is knowable. h0 can be measured as:

O b p0
h0 ¼ arctanð Þ ð6Þ
f
 DOEC Method of Time of Flight Camera Based on Pinhole Model 287

where Ob p0 is the image distance between p and image plane center Ob . The depth map
exists as a two-dimensional pixel matrix in the depth camera. Since there is a conversion
relationship between successive imaging plane and discrete image pixel coordinate
system, Ob p0 actually represents the true distance of pixel point in the discrete image
pixel coordinate system, and the position of p0 in the continuous imaging plane coor-
dinate system is known as (x, y). The coordinate transformation between the continuous
imaging plane and the discrete image pixel coordinate is shown in Fig. 5.
where u  v represents the discrete pixel coordinate system, x  y represents the
continuous image coordinate system, and from a continuous imaging plane to a discrete
image pixel coordinate system, p0 ðx; yÞ will be transformed into ðup0 ; vp0 Þ, which is
shown in (7):
x y
up 0 ¼ þ u0 vp 0 ¼ þ v0 ð7Þ
dx dy

where dx represents the true physical width of one pixel in the x-axis direction, and dy
represents the same in the y-axis direction. These two widths are intrinsic parameters of
the imaging chip and can be found by consulting the chip manual. If the chip is stable,
dx and dy will not change. ðu0 ; v0 Þ is the center point of the image plane, which is the
projection point of the camera aperture center on the image plane and can be obtained
by calibration. So Ob p0 can be calculated by (8).
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ob p0 ¼ ðup0  u0 Þ2 þ ðvp0  v0 Þ2 ð8Þ

Then the final depth vertical distance can be computed as

pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2
ffi
2
p ðu 0  u0 Þ þ ðv 0  v0 Þ
p
d ¼ r  cosðarctanð f ÞÞ
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð9Þ
ðx=dxÞ þ ðy=dyÞ2
2
z ¼ r  cosðarctanð f ÞÞ

ouv u

ob (u0 , v0 ) x
dx
p dy

v y
Fig. 5. Coordinate transformation Fig. 6. Experiment result of DOEC
288 L. Wang et al.

Through the above steps, the distance overestimated has been corrected, final result
of the correction are shown in Fig. 6. It can be seen that depth data of points far from
the optical center point are corrected, and the measured data fluctuate around 1 m about
2 cm of 1 m. Thus the effectiveness of the DOEC has been proved.

4 Experiment Results

By correcting errors caused by the non-imaging principle and the error caused by
imaging principle, the accuracy of the depth data is greatly improved, and the depth
data in the same plane can be kept within a small fluctuation range. To prove the
effectiveness of the above method, objects on the desktop and the desktop are taken as
the experimental objects. By analyzing the scope of the depth data, the plane where the
desktop is located and the plane where the objects are located are divided.
Figure 7(b) is a grayscale image, as shown in Fig. 7(a) is a depth map after non-
imaging principle error correction, it can be seen that the overall image data after the
error correction caused by the non-imaging principle is relatively stable, but due to the
impact of errors caused by imaging principle, radiating radially from the center of the
depth image, the color gradually darkens from yellow to blue. This phenomenon
indicates that the depth data is different at different locations although they are in the
same plane. The accuracy of the depth data is not high, the fluctuation range is large,
and it is difficult to distinguish between the target object and the background. There-
fore, in order to further improve the accuracy of the depth data, it is necessary to correct
the depth overestimation error caused by the imaging principle.

(a)Depth image (b)Grayscale image (c)Correct image

(d)Background image (e)Target objects

Fig. 7. Experiment process and results of errors correction (Color figure online)

After correcting errors caused by non-imaging principle and error caused by

imaging principle, the corrected depth image is finally obtained as shown in Fig. 7(c)–
(e). In order to facilitate the analysis and processing of the image, this paper maps the
 DOEC Method of Time of Flight Camera Based on Pinhole Model 289

(a)Depth image (b)Grayscale image (c)Correct image

(d)Background image (e)Target objects

Fig. 8. Experiment process and results of errors correction

data after the overestimation error correction to the range of 0–255, so that the cor-
rected depth image is represented by a grayscale image, as shown in Fig. 7(c). Different
depth data corresponds to different gray values in the range of 0–255. The greater the
depth distance is, the greater the gray value is. It can be seen from the Fig. 7(c) that the
gray value on the same plane after the error correction is relatively uniform, indicating
that the fluctuation range of the depth data on the same plane is small, so the back-
ground and the target object can be separated according to the depth distance range of
different planes. Figure 7(d) is a divided background, and Fig. 7(e) is a divided target
object. Because the distance from the center of the camera to the desktop is greater than
the distance from the surface of the article, the gray value corresponding to the desktop
after correction is greater than the corresponding gray value of the surface of the target
object.
After change the number and shape of target objects, final results are shown in
Fig. 8. The background planes and the target objects are well separated, proved that the
two error correction methods are valid.

5 Conlusion

Due to the imaging characteristics and various factors in the external environment,
there are many kinds of errors exist in the depth data of ToF camera. Starting from the
imaging principle, this paper classifed errors into two categories: errors caused by the
non-imaging principle and errors caused by imaging principle. To improve errors
correction efficiency of errors caused by non-imaging principle, temperature changes,
amplitude intensities and integration time are analyzed and corrected in this paper,
multipath errors and other noises are compensated by taking offsets, an evaluation
function is proposed to determine the optimal reference distance, so as to select
appropriate integration time and global offsets. In the correction of errors caused by
290 L. Wang et al.

imaging principle, Distance Overestimation Error Correction method (DOEC) based on

the pinhole imaging principle is proposed to remove the radial distance error of the ToF
depth camera, which contributes to limit the data on the same plane within a certain
range. Finally, objects of different heights in different planes were taken as experiment
objects, and plans are segment by depth data range. Even the shape of the item are
complicated, the background plan can be split, which demonstrates that the correction
methods proposed in this paper are effective.

Acknowledgement. This work is supported by National Science Foundation of China

(61473182, 61773253), Science and Technology Commission of Shanghai Municipality
(15JC1401900), Natural Science Foundation of Shanghai (No. 18ZR1415100).

References
1. Pertile, M., Chiodini, S., Giubilato, R., Debei, S.: Calibration of extrinsic parameters of a
hybrid vision system for navigation comprising a very low resolution time-of-flight camera.
In: IEEE International Workshop on Metrology for Aerospace, pp. 391–396 (2017)
2. Lin, J., Liu, Y., Suo, J., Dai, Q.: Frequency-domain transient imaging. IEEE Trans. Pattern
Anal. Mach. Intell. 39, 937–950 (2016)
3. Adam, A., Dann, C., Yair, O., Mazor, S., Nowozin, S.: Bayesian time-of-flight for realtime
shape, illumination and albedo. IEEE Trans. Pattern Anal. Mach. Intell. 39, 851–864 (2017)
4. Anwer, A., Ali, S.S.A., Khan, A., Mériaudeau, F.: Underwater 3D scene reconstruction
using Kinect v2 based on physical models for refraction and time of flight correction. IEEE
Access 5, 1–11 (2017)
5. Shim, H., Lee, S.: Recovering translucent objects using a single time-of-flight depth camera.
IEEE Trans. Circuits Syst. Video Technol. 26, 841–854 (2016)
6. Francis, S.L.X., Anavatti, S.G., Garratt, M., Shim, H.: A ToF-camera as a 3D vision sensor
for autonomous mobile robotics. Int. J. Adv. Robot. Syst. 12, 1 (2015)
7. Illadequinteiro, J., Brea, V.M., López, P., Cabello, D., Doménechasensi, G.: Distance
measurement error in time-of-flight sensors due to shot noise. Sensors 15, 24–42 (2015)
8. Fürsattel, P., et al.: A comparative error analysis of current time-of-flight sensors. IEEE
Trans. Comput. Imaging 2, 27–41 (2016)
9. Belhedi, A., Bartoli, A., Bourgeois, S., Gay-Bellile, V.: Noise modelling in time-of-flight
sensors with application to depth noise removal and uncertainty estimation in three-
dimensional measurement. Comput. Vis. IET 9, 967–977 (2015)
10. Ghorpade, V.K., Checchin, P., Trassoudaine, L.: Line-of-sight-based ToF camera’s range
image filtering for precise 3D scene reconstruction. In: European Conference on Mobile
Robots, pp. 1–6 (2015)
11. Hertzberg, C., Frese, U.: Detailed modeling and calibration of a time-of-flight camera. In:
International Conference on Informatics in Control, Automation and Robotics, pp. 568–579
(2015)
12. Konno, Y., et al.: Accurate plane estimation based on the error model of time-of-flight
camera. In: Second IEEE International Conference on Robotic Computing, pp. 304–307
(2018)
 Discrete Event-Triggered H1 State-Feedback
Control for Networked Control Systems

Weili Shen1, Jingqi Fu1(&), Weihua Bao2, and Zhengming Gao2

1
Department of Automation, College of Mechatronics Engineering
and Automation, Shanghai University, No. 149, Yanchang Road,
200072 Shanghai, China
[email protected]
2
Shanghai Automation Instrumentation Co., Ltd., Shanghai, China

Abstract. This paper is concerned with H1 controller design problem for

event-triggered networked control systems with bounded time-varying trans-
mission delay and packet losses. The state-feedback controller proposed in this
paper deals with event-triggered model and guarantees a required disturbance
attenuation level c subject to external bounded disturbances. Firstly, an
improved event-triggered communication scheme is proposed. Secondly, an
event-based system is transformed into a delay system. Thirdly, the asymptotical
stability criterion in the sense of mean-square for the delay-dependent H1
control, and the collaborative approach of the state-feedback control gain and
the event-triggered scheme parameters are derived by using LMIs technique and
Lyapunov stability theory which lead to less conservatism. Finally, two exam-
ples are provided to illustrate the effectiveness of the proposed event-triggering
transmission scheme and the co-design method.

Keywords: H1 controller  Event-triggered communication scheme

Lyapunov-Krasovskii functional  Time-delay  NCSs

1 Introduction

The cross-infiltration and development of Sensors, microelectronics and communica-

tion technologies contribute to the research and accelerate the application of the net-
work control systems (NCSs). The components of NCSs are modularized and share the
communication networks, which brings a series of advantages, such as low installation
and maintenance costs, high information integration, flexible scalability, and the
capacity of meeting the requirements of some special occasions [1–3]. Up to now, the
NCSs have a wide range of applications in various fields [2]. However, its limited
energy resources are the bottleneck restricting its forward development.
In order to solve the above problem, the event-triggered mechanism(ETM) comes
into being. Event-triggered techniques for control design guarantee the signals which
are really needed to be sent or the control actions which are really necessary to be
executed by using a judgement function [3, 4]. Typically, the function contains the state
or output information of the systems, when the information satisfies a certain condition,
the task is executed. Thus, the amount of the sent state signals is relatively less

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 291–301, 2018.
https://doi.org/10.1007/978-981-13-2384-3_27
292 W. Shen et al.

compared with time-triggered method. With the reduction of transmission data, the
problems of limited bandwidth, packet loss and network delay in NCSs will be alle-
viated and the lifespan of the battery of the nodes will also increase.
However, compared with the theory of time sampling, the theoretical research on
ETM is relatively lacking, and the main reason caused that is the theoretical analysis on
the latter is much more complex. Even for a simple system, the complexity and
challenge for analysis and design is apparently. Time delay system method is one of the
main modeling and analysis methods of event triggered control system [1, 5–8]. The
linear system is modeled as a system model with internal delay, which facilitates the
consideration of network constraints such as induced delays and designing controllers.
In [1], theorems are proposed to preserve the desired control performance of the event-
triggered linear system and evaluate the upper bound of the event-triggered condition;
and to avoid the sensor nodes keeping radio on to wait for reception of the sampled
data, a novel self-triggered sampling scheme is developed based on event-triggered
theorems. In [5], a sufficient condition is derived to co-design both the desired dynamic
output feedback controls and the event-triggering parameters considering not all state
of the system is measurable. In [9], a more general forms of the event-triggered
paradigm than [5] refers is used for both state-feedback and static output-feedback
control; the maximum verifying period is obtained by solving a generalized eigenvalue
problem. For the linear system with constant delay, a state feedback controller based on
event triggered mechanism is designed, which can offer the optimal control perfor-
mance while reducing the control task executions and the data transmissions [6]. In [7],
an improved event-triggered mechanism is proposed which can obtain low transmis-
sion rate during the whole operation time by using both the state and state-independent
information. In [8], for the case that the statistical information on the external noises is
not precisely known, an event-triggered method to ensure the exponential stability as
well as prescribed H1 performance for the filtering error system are derived. However,
there are a few studies on event-triggered control considering time delay and packet
loss simultaneously, which is the motivator of the study in this paper.
Referring to the above literature, the time-delay system modeling method is applied
in this paper to the stability analysis and controller design of event-triggered systems.
Considering that most research is a continuous trigger mechanism depending on
hardware, a discrete event-triggered communication mechanism with variable weight,
which can be implemented by software and guarantee the minimum lower limit of the
release period in principle, is proposed in this paper. Take the effect of both time delay
and packet losses into account, the NCS is remodeled. The results show that the joint
design method of the event-triggered method and the state feedback controller pro-
posed in this paper can achieve longer average control period, i.e. reduce more
resources in NCSs.
Notation: Rn and N denote the n-dimensional Euclidean space and natural number,
respectively. The notation X [ 0ðX  0Þ stands for a real symmetric positive definite
(positive semidefinite). X T denotes the transpose of matrix X. Diag{}denotes the block-
diagonal matrix. “*” is used as the ellipsis for terms induced by symmetry. Sym { X }
represents X þ X T .
 Discrete Event-Triggered H∞ State-Feedback Control 293

2 Problem Statement
2.1 System Description
Consider the following continuous-time linear system governed by


x_ ðtÞ ¼ AxðtÞ þ B^uðtÞ þ Bx xðtÞ
ð1Þ
zðtÞ ¼ CxðtÞ

where u^ðtÞ 2 Rm is the control input vector, xðtÞ 2 L2 ½0; 1Þ is the disturbance, and
zðtÞ 2 Rp is the control output; A, B, C, D and Bx are constant matrices with appro-
priate dimensions; the initial condition of the system (1) is given by xðt0 Þ ¼ x0 .
The structure of NCSs with network latency and ETM is constructed in Fig. 1.

uˆ(t )

x(t )

x(kh)

x(tk h)

τ ca (tk ) τsc (tk )

Fig. 1. The structure of the NCSs based on event-triggered scheme

The sensor collects the information of a plant; In this paper, it is used to monitor
changes in the state xðtÞ. The sampler detects the continuous signal acquired by the
sensor in a fixed period h, and obtains discrete signals xðkhÞ. The sampling time is set
as Ss ¼ f0; h; 2h; . . .gðk 2 NÞ. Whether the sampled data is transmitted to the con-
troller depends on the event-triggered condition predefined by event generator.
The control signal is transmitted in a single packet. First do not consider the packet
loss, the successful transmitted instant sequence of the event generator can be described
as Sg= {0, b1h, b2h, …} ðbk 2 NÞ. With the rapid development of electronic technology
and chips, and the wide use of high-speed digital processors, the computational delay is
negligible compared with transmission delays in the NCSs [2]. The network induced
delay can be expressed and sk ¼ ssc ðtk Þ þ sca ðtk Þ and sk 2 ½sm ; sM Þ, where ssc ðtk Þ,
sca ðtk Þ are the transmission delay between sampler to controller and controller to
actuator respectively; sm = min{sk }, sM = max{sk }. The input of the actuator is gen-
erated by the ZOH, and the holding time is t 2 ½tk ; tk þ 1 Þ(tk ¼ bk h þ sk ).

2.2 An Event-Triggered Mechanism

In order to reduce the data transmission in the NCSs, the following event-triggered
algorithm is designed for the event generator in Fig. 1:
294 W. Shen et al.

½xðbk h þ jhÞ  xðbk hÞT W ½xðbk h þ jhÞ  xðbk hÞ [ dNT WN ð2Þ

where N ¼ ð1  rÞxðbk hÞ þ rxðbk h þ jhÞ;xðbk h þ jhÞ represents the value of the

plant state at the current sampling instant, and xðbk hÞ represents the value of the latest
trigger instant of the plant state; W is a symmetric positive definite matrix, known as
the event-triggered matrix; d  0 is a given scalar parameter named threshold;
0  r  1 is a given scalar parameter as well. Once xðbk h þ jhÞ satisfies (2), it will be
sent immediately and the control law will be updated accordingly; otherwise, it will be
completely discarded. In terms of the above ETM, the controlled system can achieve
low average transmission rate while ensuring a satisfactory system performance by
adjusting the parameters d, r, W. By changing the magnitude of the parameter r, the
weight of the current sampled data and the latest release data can be adjusted. A larger
threshold can bring a smaller transmission data rate under the premise of a stable
control system.
The trigger condition (2) is state-dependent discrete condition. As described, it only
needs to determine whether the sampled signals meet the trigger condition at sampling
instants, thus ensuring the trigger interval greater than a sampling period, i.e. strictly
positive, by which avoiding zone phenomenon in essence, and there is no need to the
extra hardware devices monitoring the changes of system status.

2.3 Closed Loop System Modeling

Assume that the all state of the system can be measured, a static linear state feedback
controller uðtÞ ¼ Kxðbk hÞ (t 2 ½tk ; tk þ 1 Þ) is designed for the system. And the ZOH
retention time X ¼ ½tk ; tk þ 1 Þ can be divided into subintervals according to the time
delay analysis method in reference [7]:

d
X ¼ [ Xik ð3Þ
ik ¼ 0

where d ¼ bk þ 1  bk  1,


½bk h þ ik h þ sk ; bk h þ ðik þ 1Þh þ sk Þ; ik ¼ 1; 2; . . .; d  1
Xik ¼
½bk h þ dh þ sk ; bk þ 1 h þ sk þ 1 Þ; ik ¼ d

Defining a time delay function:

8
>
> t  bk h; t 2 ½bk h þ sk ; bk h þ h þ sk Þ
>
< t  ðbk h þ hÞ; t 2 ½bk h þ h þ sk ; bk h þ 2h þ sk Þ
gðtÞ ¼ . ð4Þ
>
> ..
>
:
t  ðbk h þ dhÞ; t 2 ½bk h þ dh þ sk ; bk þ 1 h þ sk þ 1 Þ

Obviously gðtÞ is a piecewise linear function and bounded, that is g1  gðtÞ  g2 ,

where g1 ¼ sm , g2 ¼ ssM þ h; and the derivative in the subinterval is 1.
 Discrete Event-Triggered H∞ State-Feedback Control 295

Define an error function eðtÞ:

8
>
> xðbk hÞ  xðbk hÞ; t 2 X0
>
< xðbk h þ hÞ  xðbk hÞ; t 2 X1
eð t Þ ¼ .. ð5Þ
>
> .
>
:
xðbk h þ dhÞ  xðbk hÞ; t 2 Xd

Thereby control input can be represented as uðtÞ ¼ K ½xðt  gðtÞÞ  eðtÞ

ðt 2 ½tk ; tk þ 1 ÞÞ. Take the packet losses into account next. Assume that the random
packet losses under the network condition of this paper is described with a stochastic
variable bðtÞ which follows the Bernoulli distribution:



PðbðtÞ ¼ 1Þ ¼ EfbðtÞg ¼ b
 ð6Þ
PðbðtÞ ¼ 0Þ ¼ 1  b

where 0\b\1. data not be lost when bðtÞ ¼ 1. Then the final input signal translates
into: ^uðtÞ ¼ bðtÞKxðbk hÞ, and the spatial model of the system is as follows:
8
< x_ ðtÞ ¼ AxðtÞ þ B^uðtÞ þ Bx xðtÞ
zðtÞ ¼ CxðtÞ ð7Þ
:
^uðtÞ ¼ bðtÞKxðbk hÞ; t 2 ½tk ; tk þ 1 Þ

In combination with the above analysis, a closed loop system model based on event
triggered mechanism is obtained:


 ðtÞ
x_ ðtÞ ¼ w1 ðtÞ þ ðbðtÞ  bÞw 2 ð8Þ
zðtÞ ¼ CxðtÞ

In which, the initial function of the system is xðtÞ ¼ uðtÞ, t 2 ½g2 ; 0. Where
 ½xðt  gðtÞÞ  eðtÞ þ Bx xðtÞ, w ðtÞ ¼ BK ½xðt  gðtÞÞ  eðtÞ.
w1 ðtÞ ¼ AxðtÞ þ bBK 2

3 H1 Control Stability Analysis

In this section, for a given disturbance attenuation level c, we will obtain a sufficient
condition on the design of the state feedback controller such that the system described
by (8) with the event generator (2) meets the following two requirements.
(1) The closed-loop system described by (8) with xðtÞ ¼ 0 is asymptotically stable;
(2) Under zero initial condition, the controlled output zðtÞ satisfies kzðtÞk2  c
kxðtÞk2 for any nonzero xðtÞ 2 L2 ½0; 1Þ.

 and the feedback gain

Theorem 1: For the given positive parameters g1 , g2 , c, r, d, b
K, under the event-trigged communication scheme (2), the closed-loop system (8) is
asymptotically stable with an H1 performance constraint index c for the disturbance
296 W. Shen et al.

attention, if there exist matrices P ¼ PT [ 0, W [ 0, Zi ¼ ZiT [ 0, Qi ¼ QTi [ 0,

Ri ¼ RTi [ 0ði ¼ 1; 2Þ, and matrix S with appropriate dimensions such that
 
Z2 
[0 ð9Þ
S Z2
 
U11 
\0 ð10Þ
U21 U22

where
 T    T    T    T    T  
e1 e1 e2 e2 e1 e1 e3 e3 e1 e1
U11 ¼ symfCT1 Pe1 g þ Q1  Q1 þ Q2  Q1 þ R1
e2 e2 e3 e3 e2 e2 e4 e4 e6 e6
     T    T  
e6 T e6 e1 e1 e7 e7  T   
 R1 þ R2  R2  e1  eT2 Z1 ðe1  e2 Þ  c2 eT10 e10  eT4  eT5 Z2 ðe4  e5 Þ
e7 e7 e6 e6 e8 e8
     
 eT5  eT8 Z2 ðe5  e8 Þ þ eT5  eT8 Sðe4  e5 Þ þ eT4  eT5 Sðe5  e8 Þ  eT9 We9
 T T
  
þ d ðr  1Þe9 þ e5 W ðr  1Þe9 þ e5

g1 g
UT21 ¼ colf C1 ; g21 C1 ; C2 ; 1 C3 ; g21 C3 g; U22
n
3 3 o
¼ diag Z11 ; Z21 ; I; b ^1 Z 1 ; b^1 Z 1
1 2

Proof: Construct a Lyapunov-Karsovskii functional as following

VðtÞ ¼ V1 ðtÞ þ V2 ðtÞ þ V3 ðtÞ ð11Þ

where V1 ðtÞ ¼ xT ðtÞPxðtÞ

Z t Z t
V2 ðtÞ ¼ xT1 ðsÞQ1 x1 ðsÞds þ xT1 ðsÞQ2 x1 ðsÞds
g 2g
t 31 t 31
Z t Z t
þ xT2 ðsÞR1 x2 ðsÞds þ xT2 ðsÞR2 x2 ðsÞds
g 2g2
t 32 t 3

Z Z t Z g1 Z t
g1 0 T
V3 ðtÞ ¼ x_ ðsÞZ1 x_ ðsÞdsdh þ g21 x_ T ðsÞZ2 x_ ðsÞdsdh; g21
3 g31 t þ h g2 tþh
¼ g2  g 1

Taking the derivation of VðtÞ for t 2 ½bk h þ sk ; bk þ 1 h þ sk þ 1 Þ along the trajectory

of the system, obtain:

EfV_ 1 ðtÞg ¼ Ef2xT ðtÞP_xðtÞg ¼ 2xT ðtÞPw1 ðtÞ ð12Þ

 Discrete Event-Triggered H∞ State-Feedback Control 297

g1 g
EfV_ 2 ðtÞg ¼ xT1 ðtÞQ1 x1 ðtÞ  xT1 ðt  ÞQ1 x1 ðt  1 Þ þ xT1 ðtÞQ2 x1 ðtÞ
3 3
T 2g1 2g1 T T g g
 x1 ðt  ÞQ1 x1 ðt  Þ þ x2 ðtÞR1 x2 ðtÞ  x2 ðt  2 ÞR1 x2 ðt  2 Þ ð13Þ
3 3 3 3
T T 2g2 2g2
þ x2 ðtÞR2 x2 ðtÞ  x2 ðt  ÞR2 x2 ðt  Þ
3 3
Z t
g2 ^ T ðtÞZ1 w ðtÞ  g1
EfV_ 3 ðtÞg ¼ 1 ½wT1 ðtÞZ1 w1 ðtÞ þ bw 2 2 x_ T ðsÞZ1 x_ ðsÞds
9 3 th31
Z tg1 ð14Þ
T ^ T
þ g21 ½w1 ðtÞZ2 w1 ðtÞ þ bw2 ðtÞZ2 w2 ðtÞ  g21
2
x_ T ðsÞZ2 x_ ðsÞds
tg2

 T  T
^ ¼ bð1
where x1 ðtÞ ¼ xT ðtÞ xT ðt  g31 Þ ; x2 ðtÞ ¼ xT ðtÞ xT ðt  g32 Þ ; b   bÞ.

The formula (14) contains two integral terms. In order to reduce conservatism, we
use Jensen’s inequality [7] and reciprocally convex method to scale the two integral
terms respectively:
Z h
g t
g iT h g i
 1 _ T ðsÞZ1 x_ ðsÞds   xðtÞ  xðt  1 Þ Z1 xðtÞ  xðt  1 Þ
x ð15Þ
3 g
t 31 3 3
R tg1 T
g21 tg2 x_ T ðsÞZ2 x_ ðsÞds   ½xðt  g1 Þ  xðt  gðtÞÞ Z2 ½xðt  g1 Þ  xðt  gðtÞÞ
T

½xðt  gðtÞÞ  xðt  g2 ÞT Z2 ½xðt  gðtÞÞ  xðt  g2 Þ

þ sym ½xðt  gðtÞÞ  xðt  g2 Þ S½xðt  g1 Þ  xðt  gðtÞÞ
ð16Þ

_
For VðtÞ, plus and minus the corresponding items, obtains:

_
E VðtÞ _
¼ E VðtÞ  eT ðtÞWeðtÞ þ eT ðtÞWeðtÞ  zT ðtÞzðtÞ
ð17Þ
þ zT ðtÞzðtÞ  c2 xT ðtÞxðtÞ þ c2 xT ðtÞxðtÞ

Within the trigger interval, the following formula is satisfied:

eT ðtÞWeðtÞ  d#T ðtÞW#T ðtÞ ð18Þ

where #ðtÞ ¼ ð1  rÞ½xðt  gðtÞÞ  eðtÞ þ rxðt  gðtÞÞ.

Combine (15), (16), (17) and (18), obtain:

_
E VðtÞ  vT ðtÞHvðtÞ  zT ðtÞzðtÞ þ c2 xT ðtÞxðtÞ ð19Þ

where vðtÞ ¼ fxðtÞ; xðt  g31 Þ; xðt  g1 Þ; xðt  gðtÞÞxðt  g32 Þ; xðt  2g32 Þ; xðt  g2 Þ;
g2 g2 ^ T 2 ^ T
eðtÞ; xðtÞg; H ¼ U11 þ 91 CT1 Z1 C1 þ g221 CT1 Z2 C1 þ CT2 C2 þ 91 bC 3 Z1 C3 þ g21 b C3 Z2

C3 ; C1 ¼ Ae1 þ bBKe 
5  bBKe9 þ Bx e10 ; C2 ¼ Ce1 ; C3 ¼ BKe5  BKe9 ; ei ði ¼ 1; 2;
. . .; 10Þ is a block coordinate matrix, for example e2 ¼ ½0 I 0 0 0 0 0 0 0 0.
298 W. Shen et al.

By applied the Schur complements, obtains:

 
U11 
\ 0 ) H\0 ð20Þ
UT12 U22

Stated thus, the system (18) satisfying Theorem 1 with x  0 is asymptotically

stable and meets the H1 performance index c. This completes the proof.
On the basis of Theorem 1, the design method of controller K is given by
Theorem 2.
 under the event-
Theorem 2: For the given positive parameters g1 , g2 , c, r, d and b,
trigged communication scheme (2), the closed-loop system (8) is asymptotically stable
with an H1 performance index c for the disturbance attention, and the feedback gain
K ¼ YX 1 , if there exist matrices X ¼ X T [ 0, W ~ ¼ W ~ T [ 0, Z~i ¼ Z~iT [ 0,
~ ~ T ~ ~ T
Qi ¼ Qi [ 0, Ri ¼ Ri [ 0ði ¼ 1; 2Þ, and matrix S with appropriate dimensions such
that
 
Z~2 
~S ~2 [ 0
Z
ð21Þ

 
~ 11
U 
~
U21 ~ 22 \0
U
ð22Þ

Where
 T    T    T    T  
e1 e1 e2 e2 e1 e1 e3 e3
~ 11 ¼symfeT C
U ~ ~1 ~1 ~2 ~1
1 1g þ Q  Q þ Q  Q
e2
e2 e3 e3 e2 e2 e4 e4
 T    T
   T    T  
e1 e1 e6 e6 e1 e1 e7 e7  T 
þ ~1
R  ~1
R þ ~2
R  ~2
R  e1  eT2 Z~1 ðe1  e2 Þ
e6 e6 e7 e7 e6 e6 e8 e8
     
~ 9  c2 eT10 e10  eT4  eT5 Z~2 ðe4  e5 Þ  eT5  eT8 Z~2 ðe5  e8 Þ þ eT5  eT8 ~Sðe4  e5 Þ
eT9 We
     
þ eT4  eT5 ~Sðe5  e8 Þ þ d ðr  1ÞeT9 þ eT5 W ~ ðr  1Þe9 þ e5

~ 21 ¼ colfg1 C
U ~ 1 ; g21 C ~ 2 ; g1 C
~ 1; C ~ 3 ; g21 C
~ 3g
3 3
n o
~ 22
U ¼ diag X Z~11 X; X Z~21 X; I; b ^1 X Z ~11 X; b ^1 X Z~ 1 X
2

C 
~ 1 ¼ AXe1 þ bBYe  ~
5  bBYe9 þ Bx e10 ; C2 ¼ CXe1 ; C3 ¼ BYe5  BYe9 :

Proof: Define matrices n = diag{X, X, X, X, X, X, X, X, X, I, I, I, I, I}, 1 = diag{X, X},

X = P−1, W ~ = XWX, Z~i = XZiX, Q ~ i = XQiX, R~ i = XRiX(i = 1, 2). The congruent
transformation is then performed for the condition (9) and (10) in Theorem 1.
 Discrete Event-Triggered H∞ State-Feedback Control 299

   
Z~2  T Z2 
~S ¼1 1[0 ð23Þ
Z~2 S Z2
   
~ 11
U  U11 
~ ~ 22 ¼ nT n\0 ð24Þ
U21 U U21 U22

This completes the proof.

CCL algorithm is adopted to solve the non-linear terms X Z~11 X, X Z~21 X in
inequality (22), details of which can be found in reference [10].

4 Illustrative Examples

Assume that the systems described in Fig. 1 is the satellite control system [5]. The state
equation of the satellite system is described as follows:
8 2 3 2 3 2 3
>
> 0 1 0 0 0 1
>
> 6J J J J 7
k d k d 607 617
<
x_ ðtÞ ¼ 6
4 0
2 2 2 7
xðtÞ þ 6 7 6 7
4 0 5uðtÞ þ 4 1 5xðtÞ
0 15
2
0 ð25Þ
>
> k d k d
>
>  J1  J1 J1 J1
1
J1 1
:
zðtÞ ¼ ½ 1 0 0 0 xðtÞ

where J1 ¼ J2 ¼ 1; k ¼ 0:09 and d ¼ 0:0219. Our task is to design an H1 state

feedback controller for the satellite control system so that the system is stable with H1
performance norm bound c when existing the disturbance of external noise The
parameters are set as sm ¼ 20ms, sM ¼ 60ms; h ¼ 100ms, d ¼ 0:1; r ¼ 0:5; c ¼ 10.
For t 2 [10 s, 20 s], imposing a perturbation wðtÞ ¼ e2t on the system. Using
Theorem 2 to solve the state feedback parameter K and the event-triggered matrix W.
Under this condition, setting b ¼ 0:05 yields:

K ¼ ½0:0273  0:8072  0:3140  0:7013

2 3
1:3671 0:6341 0:6846 0:0483
6 0:6341 1:3326 0:6341 0:0643 7
W ¼64 0:6846
7
0:6341 1:3668 0:0481 5
0:0483 0:0643 0:0481 0:1607

In Fig. 2, the abscissa and the ordinate show the release instants for the signal and
the time interval between two transmission instants. From the diagram, we can see that
under the event-triggered communication mechanism, the control cycle of the system is
not fixed, and the maximum execution cycle is 1.6 s, which is larger than the sensor
sampling period h. For t 2 (0, 60 s), the number of data obtained from the sampler is
600. If the traditional periodic trigger control strategy is adopted, 600 packets have to
be transmitted, although some of which is unnecessary. However, the simulation results
show that the amount of transmitted data is 78 with proposed ETM, that is, the average
data transmission rate is 13.00%. Compared with the time-triggered mechanism, it
300 W. Shen et al.

saves 87.00% of the bandwidth resources. The frequency of task execution is reduced
in the NCS with ETM, but the control effect is not affected when some control data lost
(Fig. 3).

Fig. 2. The release instants and release interval

Fig. 3. The event-triggered state responses

5 Conclusion

To save the NCSs with limited resources and increase its life-span, in this paper, we
have studied a kind of discrete event-triggered scheme for continuous-time system
when existing time-varying delay and packet loss constrains. The data is transmitted in
the form of a single packet in the network, and the loss of packets is described by a
stochastic variable that satisfies the Bernoulli distribution. For the resulting delay
system, a Lyapunov-Karsovskii functional with the delay information has been con-
structed. A new stabilization criterion in the sense of mean-square expressed in the
form of matrix inequalities with H1 norm bounds and H1 control design has been
developed. As shown by simulation results, the proposed method can significantly
reduce data transmission while normal operation of the system, despite delay and
packet loss.
 Discrete Event-Triggered H∞ State-Feedback Control 301

Acknowledgments. This work was financially supported by the Science and Technology
Commission of Shanghai Municipality of China under Grant (No. 17511107002).

References
1. Peng, C., Han, Q.L.: On designing a novel self-triggered sampling scheme for networked
control systems with data losses and communication delays. IEEE Trans. Ind. Electron. 63
(2), 1239–1248 (2016)
2. Zhang, L., Gao, H., Kaynak, O.: Network-induced constraints in networked control systems
—a survey. IEEE Trans. Ind. Inform. 9(1), 403–416 (2013)
3. Liu, Q., Wang, Z., He, X., Zhou, D.H.: A survey of event-based strategies on control and
estimation. Syst. Sci. Control. Eng. 2(1), 90–97 (2014)
4. Trimpe, S., Campi, M.C.: On the choice of the event trigger in event-based estimation. In:
2015 International Conference on Event-based Control, Communication, and Signal
Processing (EBCCSP), pp. 1–8, 17–19 June 2015
5. Zhang, X.M., Han, Q.L.: Event-triggered dynamic output feedback control for networked
control systems. IET Control Theory Appl. 8(4), 226–234 (2014)
6. Wu, W., Reimann, S., Görges, D., Liu, S.: Suboptimal event-triggered control for time-
delayed linear systems. IEEE Trans. Autom. Control 60(5), 1386–1391 (2015). https://doi.
org/10.1109/TAC.2014.2347214
7. Li, F., Fu, J., Du, D.: An improved event-triggered communication mechanism andL∞con-
trol co-design for network control systems. Inf. Sci. 370–371, 743–762 (2016)
8. Hu, S., Yue, D.: Event-based H∞ filtering for networked system with communication delay.
Signal Process. 92(9), 2029–2039 (2012)
9. Chen, X., Hao, F.: Periodic event-triggered state-feedback and output-feedback control for
linear systems. Int. J. Control Autom. Syst. 13(4), 779–787 (2015)
10. Peng, C., Han, Q.L.: A novel event-triggered transmission scheme and L2 control co-design
for sampled-data control systems. IEEE Trans. Autom. Control 58(10), 2620–2626 (2013)
 Brief Technical Analysis of Facial Expression
Recognition

Lei Xu, Aolei Yang(&), Minrui Fei, and Wenju Zhou

Shanghai Key Laboratory of Power Station Automation Technology,

School of Mechatronic Engineering and Automation, Shanghai University,
Shanghai 200072, China
{XuLeiaa,aolei}@shu.edu.cn, [email protected],
[email protected]

Abstract. Facial expression recognition (FER) is the current hot research topic,
and it is widely used in the fields of pattern recognition, computer vision and
artificial intelligence. As it is an important part of intelligent human-computer
interaction technology, the FER has received widespread attention in recent
years, and researchers in different fields have proposed many approaches for it.
This paper reviews recent developments on FER approaches and the key
technologies involved in the FER system: face detection and preprocessing,
facial expression feature extraction and facial expression classification, which
are analyzed and summarized in detail. Finally, the state-of-the-art of the FER is
summarized, and its future development direction is pointed out.

Keywords: Computer vision
Artificial intelligence
Facial expression

Feature extraction
Classification

1 Introduction

As an aspect of emotion recognition, the FER plays a very important role in inter-
personal communication. Facial expression is an important way to express emotion and
information exchange. According to research conducted by the social psychologist
Albert Mehrabian, in the daily human communication, emotional information trans-
mitted through facial expressions is as high as 55% of the total amount of information
[1], while information transmitted through voice and language accounts for 38% and
7% of the total amount of emotional information. The FER is the foundation of
emotional understanding, and it has injected new blood for artificial intelligence.
In 1971, Ekman and Friesen made groundbreaking work [2]. They defined six basic
types of facial expression for human expression recognition: happiness, sadness, sur-
prise, fear, anger, disgust. In 1978, Ekman and Friesen et al. developed the famous
Facial Action Coding System (FACS) to detect subtle changes in facial expressions [3].
In 1991, MASE and Kenji proposed a new theory that uses the optical flow method to
recognize facial expressions. Since then, researchers have used computer technology to
extract and classify facial expression features that have attracted widespread attention.
FER system includes three major technical steps: face detection and preprocessing,
facial expression feature extraction and facial expression classification. As shown in

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 302–310, 2018.
https://doi.org/10.1007/978-981-13-2384-3_28
 Brief Technical Analysis of Facial Expression Recognition 303

Fig. 1, FER system is established to first detect and preprocess the face, which includes
image rotation correction, face location, and scale normalization of the expression
image, etc. The feature information that can represent the essence of the input
expression is extracted from the static images or the dynamic image sequences. In order
to avoid excessive dimensions, features such as dimension reduction and feature
decomposition are required in the process of extracting feature data. Finally, the
relationship between features is analyzed and the input facial expressions are classified
into corresponding categories. At present, the FER plays an important role in inter-
active game platforms, safe driving, intelligent recommendation, distance education,
criminal investigation and security, auxiliary medical and other fields, and it has
tremendous space for development and application.

Facial
Face Detection Facial
Expression
and Expression
Feature
Preprocessing Classification
Extraction

Original feature Feature Feature

extraction Reduction Decomposition

Fig. 1. Facial expression recognition system framework

According to the relevant literatures on the FER studies in recent years, this paper
analyzes face detection and preprocessing, facial expression feature extraction,
expression classification technology. It reviews the research results of facial expression
recognition, and then looks forward to its development trend.

2 Face Detection and Preprocessing

The most importance for FER is to be able to completely extract feature information of
the entire face on an original image. Before feature extraction, it is necessary to
preprocess the original image. Image preprocessing is mainly divided into two aspects:
removing the unwanted background of the facial expression image and image
normalization.
The development of face detection technology provides an effective way to remove
the complex background of the original image in the FER system. Viola and Jones used
cascaded classifiers on Haar-like features to detect face [4]. However, sometimes it is
failed to detect contours or partially occluded faces by this method. Due to the different
image acquisition methods, the size and format of the generated images are different.
The original images need to be normalized to make the full use of the images. The
image normalization includes three parts: face normalization, plane face rotation cor-
rection and depth face rotation correction. The problem of large difference in gradation
due to uneven illumination can also be solved by gray-scale normalization after the
geometric normalization is achieved. After a series of steps, the normalization of the
original picture is finally completed. For example, as shown in Fig. 2, facial images in
304 L. Xu et al.

the Japanese Female Facial Expression (JAFFE) database are detected and prepro-
cessed, and they are classified into seven basic expressions.

happy angry fear sadness disgust suprise neural

Fig. 2. Sample images of one subject with seven expressions on JAFFE database

3 Facial Feature Extraction

Feature extraction plays an important role in the entire FER system. Under the premise
of ensuring the integrity of the original image information of the images, the approach
for extracting useful information of facial expressions can greatly improve the rate of
FER. The extraction of expression features is divided into static image feature
extraction and dynamic sequence image feature extraction according to the different
nature of the image. However, most feature extraction processes have excessive
dimensions or the high volume of data, and it is necessary to perform dimension
reduction, feature decomposition and etc. At present, various feature extraction algo-
rithms have been continuously proposed and improved.

3.1 Extraction of Facial Expression Features in Static Images

Static images are unchanged still pictures, and they are prepared for follow-up pro-
cessing after a series of preprocessing. Usually, it is divided into two major categories
including geometric feature extraction and appearance feature extraction.
The method of extracting geometric features is a macroscopic method that generally
represents the structural changes of human faces. According to geometric relationship
between the feature points, the main task of this method is to extract the feature points
of human face and facial expression features. The three representative geometric fea-
ture extraction methods include Active Shape Model (ASM), Active Appearance
Models (AAM) and Scale Invariant Feature Transform (SIFT). ASM uses a local
texture model to find the best position of the feature point in the facial expression
image, and adjusts the parameters of the statistical model. The model gradually
approaches the outline of the facial expression and achieves the accurate position of the
target. It can extract facial features rapidly and have high accuracy [5]. Antonakos et al.
built AAMs and experimental results prove that it has a good recognition effect for
unseen faces such as poses and occlusion. SIFT is a kind of local descriptor, and it has
scale invariance and can detect key points in the image. Li et al. [6] proposed the new
method based on SIFT, and it improved the performance of feature extraction in multi-
spectral graphics.
The appearance feature extraction method is a microscopic method. The appearance
feature is based on the feature of using the pixel properties of the entire face image,
 Brief Technical Analysis of Facial Expression Recognition 305

which reflects the information of the bottom face of the face image, and particularly
shows the information of the local minute change. The appearance feature extraction
methods mainly include Local Binary Pattern (LBP) and Gabor Filter. LBP is used to
compare the local features of an image. It can extract and measure the texture infor-
mation of the local neighborhood in a grayscale image, and can effectively suppress the
influence of illumination. Chao et al. [7] proposed the expression-specific LBP with
symmetric extension features, and it can better extract important face points of the
human face. Gabor Filter is a classic method for facial expression feature extraction.
A set of filters are used to filter the image. The result can reflect the relationship
between local pixels. The Gabor filter can detect texture changes of the multi-scale and
multi-directional, and it is less affected by light changes. A novel approach for FER is
proposed, and the facial feature is extracted from image by applying Gabor wavelet
filter on an image and computing local transitional pattern (LTP) code from the
transformed image. Extensive experiments illustrate that it is effective and efficient for
the FER.

3.2 Extraction of Facial Expression Features in Dynamic Image

Sequences
The process of continuous facial expression movement is reflected in the dynamic
sequence images. The expression features of the dynamic image sequences are mani-
fested by the deformation of the human face and the muscle movement in each area of
the face. Currently, feature extraction methods based on dynamic image sequences
mainly include optical flow, feature point tracking method, model tracking method and
the elastic map matching.
The optical flow is often used to extract the expression features of the dynamic
image sequence. It has the advantages of highlighting the face deformation and
reflecting the movement tendency of the face image sequences. In facial expression
recognition, the dynamic descriptor itself is more suitable than the static descriptor
because different parts of the facial expression change with time. The proposed method
based on optical flow by Fan et al. has a better recognition rate than using separate
descriptors in the FER.
The feature points are selected in regions where the gray level changes are rela-
tively large, such as corners of the eyes, corners of the mouth, etc. Displacement or
deformation information of facial features is easily obtained by tracking these points.
Ghimire et al. [8] proposed facial system can achieve good result of the FER. Unlike
the tracking of feature points, the goal based on model tracking is the entire face. The
model here can be 2D or 3D, but more complex calculations are required in most
models. Dewan et al. [9] proposed a novel method based on model tracking, and it can
gradually learn the trajectory model of human face and improve the performance of the
system.
The basic principle of the elastic map matching method is that select feature points
on the input facial expression image to construct an elastic map with a certain topo-
logical structure. Then the feature vectors on each node and structure are extracted and
a similarity function is constructed to compare the degree of similarity between the
feature vector and the topology. If the similarity is the greatest, the match could be
306 L. Xu et al.

completed. Zafeiriou and Pitas [10] propose a new technique based on the elastic map
matching improves the performance of the selection of the most discriminant facial
landmarks for the FER.

3.3 Feature Reduction Method

After the facial expression image is extracted from the original features, some features
often have very large spatial dimensions, such as appearance features. These features
need to be represented by converting from high-dimensional features to low-
dimensional subspaces through some mapping and transformation. The dimension of
the feature is significantly reduced, and the effectiveness of these features is improved.
Feature reduction methods can be further divided into two categories including linear
and nonlinear methods.
Linear dimensionality reduction methods commonly used in the FER include
principal component analysis (PCA), linear discriminant analysis (LDA), etc. The PCA
can effectively extract the local features that make a great contribution to the FER,
eliminate redundancy and reduce the dimension, but the distinction between different
types of data in this method is not considered. The LDA based on the Fisher criterion is
a more successful feature extraction algorithm. It selects the proper projection direction
by maximizing the dispersion of the data between classes and minimizing the intra-
class dispersion of data, mainly to find the direction with the greatest resolution. In
[11], stepwise linear discriminant analysis which is based on the LDA is proposed, and
it selects the localized features with the partial F-test value from the expression frames.
It is a significant improvement for the FER.
There are two representative manifold learning methods for nonlinear dimension-
ality reduction including local linear embedding and isometric mapping (Isomap).
Local linear embedding (LLE) which focus on preserving the local structures can also
reduce the dimension of data, and it can also make the data better retain the original
geometric relationship in the local sense after dimension reduction. A new kernel-based
supervised manifold learning algorithm based on LLE is proposed for the FER [12],
and it has the effectiveness and promising performance. Isomap is based on the Mul-
tidimensional Scale Transform (MDS). The goal is to maintain the intrinsic geometric
properties of the data points by keeping the geodesic distance between the two points
constant. Isomap based on expression weighted distance is proposed [13]. In order to
take the advantage of knowledge of expression classes, Isomap assigns weighed values
according to different sample distance, and it performs very well on the JAFFE
database.

3.4 Feature Decomposition Method

Face images contain a wealth of information, and the information used varies for
different identification requirements. For example, what is needed in face detection is
common information in face images, and what is needed in face recognition is face
difference information. In the FER, it is necessary to use the information of the dif-
ferences between various expression features, and perhaps information that is favorable
for one identification requirement interferes with another. Therefore, a solution is to
 Brief Technical Analysis of Facial Expression Recognition 307

separate the different face factors, so that the various types of recognition can be
performed in different subspaces to avoid interference from other factors. A dictionary-
based approach for the FER by decomposing expressions is proposed by Taheri et al.
[14]. They achieve expression decomposition and recognition by computing the facial
expression sparse code matrix.

4 Facial Expression Classification

The expression classification method and the expression feature extraction method are
closely related. It refers to defining a set of categories and designing a corresponding
classification mechanism to recognize expressions and classify them into correspond-
ing categories. At present, the main expression classification methods include Hidden
Markov Model (HMM), K-Nearest Neighbor (KNN), Support Vector Machines
(SVM), AdaBoost, Artificial Neural Network (ANN), and so on.

4.1 Hidden Markov Modeling (HMM)

The HMM contains hidden unknown parameters, and it is a Markov process based on
the Markov chain. One of them is the Markov chain, and it is a random process that
describes the transfer of states; another one describes the statistical correspondence
between observations and states. The whole process of the HMM algorithm is that the
observed values are generated with the transition of the state, and we are concerned
with judging the hidden state through the existing observations. It uses a long series of
observation sequences to extrapolate possible state sequences that lead to this result.
The random signal information can be effectively described in the statistical model by
the HMM. Wu et al. [15] proposed a novel method based on the HMM for the FER. It
can locate apex frames of sequences with multiple peaks and label the sequences
effectively.

4.2 K-Nearest Neighbor (KNN)

The KNN algorithm is a mature classification method and it is also one of the machine
learning algorithms. The principle of this method is that if the majority of the k most
similar samples (the nearest neighbors in the feature space) in a sample belong to the
same class. The main idea of KNN is to determine its own category based on the
similar neighbor categories. The premise of the algorithm is that it needs to have a set
of training data that has been marked. First, it calculates the distance between the test
object and all objects in the training set. Then, it finds the nearest K objects in the
distance calculated in the previous step as the neighbors of the test object. Finally, it
finds out the most frequently occurring objects among the K objects, and its category is
the category to which the test object belongs. Bahari et al. [16] apply the KNN to
classify the features into the emotional states, and it has better classification accuracy.
308 L. Xu et al.

4.3 Support Vector Machine (SVM)

The SVM is a classifier with strong generalization ability. It has many advantages in
solving identification problems such as nonlinearity, small sample sets and high-
dimensional patterns. The basic model of SVM is to find the best separating hyperplane
in the feature space so that the interval between positive and negative samples is the
largest in the training set. SVM is a supervised learning algorithm used to solve the
two-category problem. SVM can also be used to solve nonlinear problems after the
introduction of the kernel method. The SVM not only increases the computational
complexity, but also avoids the curse of dimensionality compared with the linear model
[17]. It proves that the polynomial kernel function has good real-time performance in
FER, while radial basis kernel function has better generalization and learning ability.

4.4 AdaBoost
AdaBoost is an iterative algorithm, and its basic principle of Adaboost algorithm is to
combine several weak classifiers (weak classifiers generally use a single-level decision
tree) to make it a strong classifier. It adopts an iterative idea that only one weak
classifier is trained per iteration and the weak classifier trained will participate in the
use of the next iteration. The AdaBoost algorithm which is proposed by Freund and
Schapire, is first practical boosting algorithm, and it is one of the most widely used and
studied [18]. Owusu et al. proposed a new method which is based on the AdaBoost,
and it can enhance accuracy and speed with the expression dataset reduced [19].

4.5 Artificial Neural Networks (ANN)

The ANN massively parallel computing system composed of a large number of con-
nected neurons, and learns complex nonlinear input-output relations through a training
process. Its principle can be summarized as follows: for nonlinear separable samples,
the input space is first converted to a high-dimensional space by nonlinear transfor-
mation, and then the optimal linear interface is obtained in the high-dimensional space,
and this nonlinear transformation is achieve by defining a proper kernel function.
Different architectures and models of ANN are used for the FER in the recent years,
such as Convolutional Neural Network (CNN), Polynomial Neural Network (PNN),
Recurrent Neural Networks (RNN), Retinal Connected Neural Network (RCNN), etc.
For example, Lopes et al. [20] combined the CNN and the specific image pre-
processing steps to recognize the facial expressions, and the CNNs achieved better
accuracy with big data.

5 Conclusion

The FER is a cross-disciplinary research topic and has become a very popular research
topic in the fields of psychology, medicine, computer vision, pattern recognition, etc. It
has important application value in the field of intelligent human-computer interaction.
This paper describes the key technologies involved in facial expression recognition:
 Brief Technical Analysis of Facial Expression Recognition 309

face detection and preprocessing, expression feature extraction and expression classi-
fication. The paper categorizes the research results on facial expression recognition
technology at home and abroad, and introduces the problems existing in some tech-
nologies and the latest technologies related to facial expression recognition.
Although many achievements have been made in the study, many problems still
need to be solved. For example, expression classification is too restrictive and human
expressions are not only six basic expressions, but there are also many complex and
superimposed expressions. Face expressions in dynamic images and facial expressions
in 3D real life are easily affected by illumination, occlusion, etc. The robustness of the
FER is poor, and the effectiveness of the recognition algorithm still needs improve-
ment. Therefore, many expression recognition algorithms have not yet been applied to
specific projects or products, and it is still a challenging topic. With the arrival of the
era of big data, the demand for real-time facial expressions has increased dramatically.
The application of the FER technology has a broader scope of development and uti-
lization value. In the future, facial expressions can be assisted by integrating sound,
heat, and environmental factors. Expression recognition system adds depth information
based on 2D to support 3D face analysis. Taking time changes into consideration helps
to improve expression recognition rates by analyzing the correlation between con-
secutive adjacent frames.

Acknowledgements. This work is supported by Natural Science Foundation of Shanghai

(No. 18ZR1415100), National Science Foundation of China (61473182, 61773253), Science and
Technology Commission of Shanghai Municipality (15JC1401900) and Key research and
development project of Yantai (2017ZH061).

References
1. Mehrabian, A.: Communication without words. Psychol. Today 2, 53–55 (1968)
2. Ekman, P., Friesen, W.V.: Constants across cultures in the face and emotion. J. Pers. Soc.
Psychol. 17, 124–129 (1971)
3. Ekman, P.: Facial Action Coding System. A Technique for the Measurement of Facial
Action (1978)
4. Hu, P.: Application research on face detection technology based on OpenCV in mobile
augmented reality. Int. J. Sign. Process. Image Process. Pattern Recogn. 8, 249–256 (2015)
5. Shbib, R., Zhou, S.: Facial expression analysis using active shape model. Int. J. Sign.
Process. Image Process. Pattern Recogn. 8, 9–22 (2016)
6. Li, Y., Liu, W., Li, X., Huang, Q., Li, X.: GA-SIFT: a new scale invariant feature transform
for multispectral image using geometric algebra. Inf. Sci. 281, 559–572 (2014)
7. Chao, W.L., Ding, J.J., Liu, J.Z.: Facial expression recognition based on improved local
binary pattern and class-regularized locality preserving projection. Signal Process. 117, 1–10
(2015)
8. Ghimire, D.: Recognition of facial expressions based on tracking and selection of
discriminative geometric features. Int. J. Multimed. Ubiquitous Eng. 10, 35–44 (2015)
9. Dewan, M.A.A., Granger, E., Marcialis, G.L., Sabourin, R., Roli, F.: Adaptive appearance
model tracking for still-to-video face recognition. Pattern Recogn. 49, 129–151 (2016)
310 L. Xu et al.

10. Zafeiriou, S., Pitas, I.: Discriminant graph structures for facial expression recognition. IEEE
Trans. Multimed. 10, 1528–1540 (2008)
11. Siddiqi, M.H., Ali, R., Khan, A.M., Park, Y.T., Lee, S.: Human facial expression recognition
using stepwise linear discriminant analysis and hidden conditional random fields. IEEE
Trans. Image Process. 24, 1386–1398 (2015)
12. Zhao, X.: Facial expression recognition using local binary patterns and discriminant kernel
locally linear embedding. EURASIP J. Adv. Signal Process. 2012, 20 (2012)
13. Wang, S., Yang, H., Li, H.: Facial expression recognition based on incremental isomap with
expression weighted distance. J. Comput. 8, 2051–2058 (2013)
14. Taheri, S., Qiu, Q., Chellappa, R.: Structure-preserving sparse decomposition for facial
expression analysis. IEEE Trans. Image Process. 23, 3590–3603 (2014)
15. Wu, C., Wang, S., Ji, Q.: Multi-instance hidden Markov model for facial expression
recognition. In: IEEE International Conference and Workshops on Automatic Face and
Gesture Recognition, pp. 1–6 (2015)
16. Bahari, F., Janghorbani, A.: EEG-based emotion recognition using recurrence plot analysis
and K nearest neighbor classifier. In: Biomedical Engineering, pp. 228–233 (2013)
17. Wang, F., He, K., Liu, Y., Li, L., Hu, X.: Research on the selection of kernel function in
SVM based facial expression recognition. In: Industrial Electronics and Applications,
pp. 1404–1408 (2013)
18. Schapire, R.E.: Explaining AdaBoost. Springer, Heidelberg (2013)
19. Owusu, E., Zhan, Y., Mao, Q.R.: A neural-AdaBoost based facial expression recognition
system. Expert Syst. Appl. 41, 3383–3390 (2014)
20. Lopes, A.T., Aguiar, E.D., Souza, A.F.D., Oliveira-Santos, T.: Facial expression recognition
with convolutional neural networks: coping with few data and the training sample order.
Pattern Recogn. 61, 610–628 (2017)
 Application of Intelligent Virtual Reference
Feedback Tuning to Temperature Control
in a Heat Exchanger

Yalan Wen1, Ling Wang1(&), Weiqing Peng1,

Muhammad Ilyas Menhas1,2, and Lin Qian1,3
1
Shanghai Key Laboratory of Power Station Automation Technology,
School of Mechatronics Engineering and Automation, Shanghai University,
Shanghai 200072, China
{phoenixwen,wangling,candidates}@shu.edu.cn,
[email protected], [email protected]
2
Department of Electrical Engineering,
Mirpur University of Science and Technology, Mirpur A.K., Pakistan
3
Shanghai Power Construction Testing Institute, Shanghai 200031, China

Abstract. Heat exchangers are frequently incorporated in industrial processes.

Temperature control in heat exchangers is very important for the safety and
economic benefits of the industrial system. However, it is still challenging to
control the temperature of heat exchangers because of the complicated thermal
dynamic phenomena. In this paper, an improved data-driven control method, i.e.
intelligent virtual reference feedback tuning based on multi-population coop-
erative human learning optimization (IVFRT-MCHLO), is developed to design
the optimal controller for a water heat exchanger. The controller is designed
based on IVRFT method and a novel multi-population cooperative human
learning optimization (MCHLO) algorithm is proposed to find out the optimal
controller. The experimental results demonstrate that the proposed IVFRT-
MCHLO has better control performance as the multi-population cooperation
strategy of MCHLO improves the global search ability greatly.

Keywords: Heat exchanger
Temperature control

Human learning optimization
Intelligent virtual reference feedback tuning
Data-driven

1 Introduction

Heat exchangers are extensively utilized in many industrial fields, including power
generation systems [1], chemical processes [2] and district heating systems [3]. Effi-
cient temperature control can improve the dynamic performance and enhance stability
of the system involving the heat exchanger significantly [4]. However, it is a major
challenge to control the fluid temperature being heated at a specific and stable reference
value because heat exchangers are extremely complex devices involving a large
number of thermal dynamic phenomena, for which the prediction of their operation
from first principles seems to be impossible [5]. Therefore, researchers proposed

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 311–320, 2018.
https://doi.org/10.1007/978-981-13-2384-3_29
312 Y. Wen et al.

various advanced approaches to design and improve temperature controllers for heat
exchangers over past few decades. However, most of those methods [4, 6, 7] are
designed based on mathematical models, which means that plenty of effort has to be
spent on system modelling. Furthermore, it is difficult to construct a perfect model and
some complex nonlinear systems are even unlikely to be modelled.
Virtual reference feedback tuning (VRFT) is a data-driven control (DDC) approach.
As a hotspot, VRFT approaches have been improved substantially and applied to a
wide range of control systems [8–10]. However, the optimal controller cannot be
designed successfully because the performance indices are not included in its objective
function in the standard VRFT method. Therefore, Wang et al. [11] developed an
intelligent VRFT (IVRFT), in which a potential performance metric, i.e., the reference
model parameter, was included in the controller design to achieve the optimal control
performance. Compared with VRFT approaches, the controller design problem in the
IVRFT is transformed from a parameter identification problem into an optimization
problem. Traditional derivative-based methods including the least square method,
which is adopted in VRFT approaches, cannot be used to solve such an optimization
problem.
Obviously, the optimization method has a significant impact on the controller
design based on IVRFT method [11]. Human learning optimization (HLO) is an simple
yet efficient optimization tool, which mimics the human learning process [12]. In
previous works, HLO has obtained the best-so-far results on two well-studied sets of
multi-dimensional knapsack problems, i.e. 5.100 and 10.100 [13], which shows its
advantages. In this paper, a multi-population cooperative human learning optimization
(MCHLO) is proposed as an optimization tool to design the optimal controller for a
water heat exchanger based on IVRFT.
The reminder of this paper is organized as follows. Section 2 introduces the IVRFT
briefly. MCHLO is described in Sect. 3. Section 4 presents the optimal controller
design based on IVRFT and MCHLO (IVRFT-MCHLO) in details. Then IVRFT-
MCHLO is utilized to control the temperature in a water heat exchanger and compared
with the other control approaches in Sect. 5. Finally, we draw the conclusions in
Sect. 6.

2 Intelligent Virtual Reference Feedback Tuning

The VRFT approach [14] aims to solve a model-reference problem in discrete time.
The objective function adopted in the VRFT method does not include any performance
index, and thus it is almost impossible to design the optimal controller through the
original VRFT method. However, for practical engineering applications, the dynamic
performance is of great importance because it significantly influences the efficiency of
systems and the benefit of enterprises. Considering that the reference model determines
the closed-loop control performance of a controller tuned by the VRFT method, Wang
et al. [11] presented a novel intelligent VRFT approach in which the reference model
and the controller are both optimized. Consequently, the controller with the best per-
formance can be found in the framework of IVRFT.
 Application of Intelligent Virtual Reference Feedback Tuning 313

Consider a first-order reference model as Eq. (1), where A is a parameter related to

the speed of response and d refers to the delay time.

ð1  AÞz1 d
Mðz1 Þ ¼ z ð1Þ
1  Az1

Apparently, the reference model with a relatively small A will have a better
dynamic performance. The controller design based on such a reference model is bound
to result in better control performance. Encouraged by that fact, a novel objective
function is proposed in IVRFT as Eq. (2)

JIVRFT ¼ JVRFT þ k  A ð2Þ

where JVRFT denotes the objective function of the standard VRFT approach, and
k denotes the weight of A which should be small enough to make sure that JVRFT
dominates this objective function.

3 Multi-population Cooperative Human Learning

Optimization

Social entities such as organizational units are often drawn by boundaries of knowledge
[15]. Individuals shape their own organizations and are also impacted by them. Levels of
cooperation have great impacts on knowledge sharing among organizational units [16].
It is natural that proper knowledge sharing can speed up the knowledge conversion
process. Inspired by this fact, a multi-population cooperative strategy is introduced into
HLO, and a novel multi-population cooperative human learning optimization is pro-
posed, where the whole population is divided into a master subpopulation and S slave
subpopulations as Fig. 1. The master population involves S individuals and the
remaining individuals are shared by the slave subpopulations equally.

Slave
subpop

Gib
SKD

Gib Gib
Slave SKD Master SKD Slave
subpop subpop subpop

Fig. 1. Schematic of MCHLO strategy

In MCHLO, the cooperation between the master subpopulation and each slave
subpopulation works together. The master subpopulation and slave subpopulation are
evolved in parallel, which can take full advantage of the emerging solutions saved in
314 Y. Wen et al.

the master subpopulation. Besides, the proposed multi-population cooperative strategy

can increase the population diversity. Therefore, the global search ability is further
improved by the enhanced exploration and exploitation.
In MCHLO, an individual is encoded as a binary string like Eq. (3), in which each
bit stands for a basic component of knowledge.

xi ¼ ½ xi1 xi2


xij


xiM ; xij 2 f0; 1g; 1  i  N; 1  j  M ð3Þ

where xi is ith individual, N refers to the number of individuals in a population, and

M denotes the length of solutions. Each bit is stochastically initialized with 0 or 1.

3.1 Learning Operators

Random Learning Operator (RLO). Human learning processes are always accom-
panied by randomness. When learning new things, people usually attempt to solve
problems by using random methods for lack of prior knowledge [17]. The individual
has random learning behaviors to explore new knowledge during the search process,
which are defined in MCHLO as Eq. (4).

0; 0  r1  0:5
xij ¼ Rð0; 1Þ ¼ ð4Þ
1; else

where r1 is a stochastic number between 0 and 1.

Individual Learning Operator (ILO). External stimuli can motivate people to build
skills and knowledge through individual reflection [18]. For each subpopulation,
individual best solutions are saved in the individual knowledge database (IKD) like
Eqs. (5)–(6)

2 3
ikd 1
6 ikd 2 7
6. 7
6. 7
6. 7
IKD ¼ 6 7; 1  i  N ð5Þ
6 ikd i 7
6. 7
4. 5
.
ikd N
2 3 2 3
ikd i1 ik i11 ik i12


ik i1j


ik i1M
6 ikd i2 7 6 ik i21 ik i22


ik i2j


ik i2M 7
6 7 6 7
6 .. 7 6 .. .. .. .. 7
6. 7 6. . . . 7
6
ikdi ¼ 6 7¼6 7; 1  p  L ð6Þ
ikd 7 6 ik ip1 ik ip2


ik ipj


ik ipM 7
6 ip 7 6 7
6 .. 7 6. .. .. .. 7
4. 5 4 .. . . . 5
ikd iL ik iL1 ik iL2


ik iLj


ik iLM
 Application of Intelligent Virtual Reference Feedback Tuning 315

where ikdi is the IKD of individual i, L indicates the number of solutions stored in the
IKD, and ikdip is known as pth best solution for individual i. For the slave subpopu-
lation, when MCHLO performs the individual learning operator (ILO), it generates new
solutions according to the knowledge in the IKD as Eq. (7)

xij ¼ ikipj ð7Þ

For the master subpopulation, the IKDmaster directly inherits Gib, i.e., the iteration-
best solution generated by the slave subpopulation.

ikdmaster;i ¼ ½ ik i1 ik i2


ik ij


ik iM ; 1  i  S; 1  j  M ð8Þ

where ikdmaster, i is the IKD of the ith individual in the master subpopulation. When an
individual in the master subpopulation yields the new candidate, it generates new
solutions from the knowledge stored in the IKDmaster as Eq. (9).

xij ¼ ikij ð9Þ

Social Learning Operator (SLO). As human beings are social animals, human
learning cannot be separated from interaction among each other. To accelerate the
learning processes, it is natural for an individual to learn from his/her own organization.
For the slave population in MCHLO, the social knowledge database (SKD) is utilized
to reserve the knowledge of the subpopulation as Eq. (10)

2 3 2 3
skd1 sk11 sk12


sk1j


sk1M
6 skd2 7 6 sk21 sk22


sk2j


sk2M 7
6 7 6 7
6 .. 7 6 .. .. .. .. 7
6 . 7 6 . . . . 7
SKD ¼ 6 7 6
6 skdq 7¼6 skq1
7; 1  q  H
7 ð10Þ
6 7 6 skq2


skqj


skqM 7
6 .. 7 6 . .. .. .. 7
4 . 5 4 .. . . . 5
skdH skH1 skH2


skHj


skHM

where skdq is the qth solution, and H denotes the size of the SKD. For an individual in
the slave subpopulation, the social learning operator (SLO) is performed as Eq. (11)

xij ¼ sk qj ð11Þ

For the master subpopulation, only the best-so-far solution is stored in the
SKDmaster as Eq. (12). Each individual in the master subpopulation executes SLO as
Eq. (13)

SKDmaster; j ¼ ½ sk 1 sk 2


sk j


sk M ; 1  j  M ð12Þ

xj ¼ sk j ð13Þ
316 Y. Wen et al.

In summary, an individual in MCHLO employs the RLO, ILO and SLO in the
search process to generate new candidates, which can be formulated as Eq. (14)
8
< RLO; if 0  r  pr
xij ¼ ILO; if pr\r  pi ð14Þ
:
SLO; if pi\r  1

where pr, (pi − pr) and (1 − pi) represent the probability of executing RLO, ILO and
SLO, respectively. Note that ILO and SLO for the master subpopulation are different
from those for each slave subpopulation, which has been described before.

3.2 Update of IKDs and SKD

IKDs and SKDs are updated so as to simulate people evaluating their performance and
updating the experience in the learning process. The sizes of IKDs and SKDs were both
equal to 1 as recommended in [17].
For the slave subpopulation, the solution in IKDs is substituted by a new candidate
if the new one has a better fitness value. SKD is updated in the same way. For an
individual in the slave subpopulation, its IKD will be re-initialized if its optimal
solution is not updated in the consecutive rl iterations to avoid premature. For the
master subpopulation, as mentioned above, IKDmaster is updated through copying the
iteration-best solution Gib generated by each slave subpopulation. SKDmaster is updated
when the new best solution is found by any subpopulation.

4 Controller Design Based on IVRFT-MCHLO

Consider a typical PID controller as Eq. (15)

uðkÞ ¼ uðk  1Þ þ Kp ½eðkÞ  eðk  1Þ þ Ki eðkÞ

ð15Þ
þ Kd ½eðkÞ  2eðk  1Þ þ eðk  2Þ

where u(k) represents the manipulated variable at the kth sampling instant, e(k) refers to
the error between the process output and its set-point at the kth sampling instant, and
Kp, Ki and Kd are parameters of a PID controller. The corresponding controller transfer
function can be formulized as Eq. (16), and its parametric form is as Eq. (17)

  Ki  
C z1 ; h ¼ Kp þ 1
þ Kd 1  z1 ð16Þ
1z
 1 
C z ; h ¼ bT ðzÞh ð17Þ
 Application of Intelligent Virtual Reference Feedback Tuning 317

where
 
T 1
b ðzÞ ¼ 1 1  z1 ð18Þ
1  z1

h ¼ ½ Kp Ki K d T ð19Þ

From Eqs. (1) and (16), uvir ðtÞ can be deduced as Eqs. (20)–(21)
   
vir
 1
 Ki  1
 1  z1 d
u z ¼ Kp þ þ Kd 1  z  z y ð20Þ
1  z1 ð1  AÞz1
2 3
  Kp
uvir ðz1 Þ ¼ uT ðtÞh ¼ up ðtÞ ui ðtÞ ud ðtÞ 4 Ki 5 ð21Þ
Kd

where up ðtÞ, ui ðtÞ, ud ðtÞ denote the ratio of Kp, Ki and Kd, respectively.
In IVRFT-MCHLO, MCHLO is used to search for the best Kp, Ki, Kd and A by
following Eq. (2), in order to tune the controller with superior performance. The
implementation of IVRFT-MCHLO can be summarized as follows:
Step 1: initialize the algorithm parameters included in MCHLO;
Step 2: set the length of each decision variable; in this paper, each decision variable
is coded by 30 bits;
Step 3: initialize the master subpopulation and each slave subpopulation randomly;
Step 4: initialize the IKDs and SKD of each slave subpopulation, as well as the
IKDsmaster and SKDmaster;
Step 5: generate new solutions by performing learning operators as Eq. (14);
Step 6: evaluate the fitness of each individual according to Eq. (2);
Step 7: update the IKDs and SKD of each slave subpopulation, as well as the
IKDsmaster and SKDmaster according to the updating rules of MCHLO;
Step 8: output the optimal solution in SKDmaster if the termination criterion is met;
otherwise go back to step 5.

5 Temperature Control in a Water Heat Exchanger

As Fig. 2 shows, the valve V that manipulates the input water is utilized to control the
temperature T of output water in a water heat exchanger. The input water is from a
storage tank, whose temperature can be kept at a certain point by an independent
controller. The behavior of this process can be modeled as Eq. (22) [19]. Note that
IVRFT-MCHLO and VRFT are model-free approaches, and the model is only utilized
to collect the I/O data and validate the controller designed by IVRFT-MCHLO.

T ðsÞ 0:12 3s

¼ PðsÞ ¼ e ð22Þ
V ðsÞ 1 þ 6s
318 Y. Wen et al.

FC

Cold water
Hot water

FH

TT

Fig. 2. Schematic diagram of the water heat exchanger

IVRFT-MCHLO was implemented to control the water temperature in the heat

exchanger, as well as IVRFT [11] and the other model-based control approaches, i.e.
the Z-N method [20] and model-based optimal PID controller design with harmony
search (MPID) [21]. The parameter settings for MCHLO are obtained by trial and error,
that is, pr = 5/M, S = 2, rl = 100, pi = 0.85 + 2/M. For the MPID approach, its
objective function is formulated as Eq. (23)

F ¼ w1  IAE þ w2  tup þ w3  OS ð23Þ

where IAE, tup, and OS refer to the integral of absolute error, rise time and overshoot,
respectively. w1, w2 and w3 are the corresponding weight coefficients which are the
recommended values in [21]. Note that the ideal model as well as the model with 0.5%
error is taken as a competitor in this paper. The control performance of the above-
mentioned approaches is depicted in Fig. 3, and the corresponding results are listed in
Table 1. Non-overshoot control is achieved by all approaches except the Z-N method.
It is undoubted that the results obtained by the IVRFT approaches are inferior to the
MPID method with the perfect model. However, it should be emphasized that the
results obtained by IVRFT-MCHLO are very close to that of MPID and significantly
better than that of MPID with 0.5% model error. IVRFT-MCHLO outperforms IVRFT
on each performance index except OS, which reflects that the optimization algorithm is
important to the control performance in the framework of IVRFT.

Fig. 3. Step response curves of all the control methods

 Application of Intelligent Virtual Reference Feedback Tuning 319

Table 1. Results of MCHLO-IVRFT and other control approaches on the temperature control
Methods Tup(s) OS(%) IAE Fitness
Z-N 12 21.88 17.5284 1123.5000
MPID with 0.5% model error 9 0 8.6685 17.6685
MPID 6 0 6.4707 12.4707
IVRFT-MCHLO 7 0 6.8822 13.8822
IVRFT 8 0 7.6210 15.6210

6 Conclusions

Heat exchangers widely exist in industrial processes. However, its temperature control
is still challenging in engineering applications. Therefore, we proposed an improved
intelligent virtual reference feedback tuning based on a novel multi-population coop-
erative human learning optimization algorithm (IVFRT-MCHLO), which can design
the optimal temperature controller of heat exchangers only based on the collected data
of the system efficiently. With the enhanced exploration and exploitation abilities,
MCHLO can be combined with the IVRFT to find out the optimal controller. Exper-
imental results display that IVRFT-MCHLO can be utilized to design the temperature
controller for a water controller successfully and achieve the desirable control
performance.

Acknowledgments. This work is supported by National Natural Science Foundation of China

(Grant No. 61633016 & 61703262), Key Project of Science and Technology Commission of
Shanghai Municipality under Grant No. 16010500300 and 15220710400, Shanghai Sailing
Program under Grant No. 16YF1403700, and Natural Science Foundation of Shanghai
(No. 18ZR1415100).

References
1. May, P., Ehrlich, H.-C., Steinke, T.: ZIB structure prediction pipeline: composing a complex
biological workflow through web services. In: Nagel, W.E., Walter, W.V., Lehner, W. (eds.)
Euro-Par 2006. LNCS, vol. 4128, pp. 1148–1158. Springer, Heidelberg (2006). https://doi.
org/10.1007/11823285_121
2. Anxionnaz, Z., Cabassud, M., Gourdon, C., Tochon, P.: Heat exchanger/reactors (HEX
reactors): concepts, technologies: state-of-the-art. Chem. Eng. Process.: Process. Intensif. 47,
2029–2050 (2008)
3. Tol, Hİ., Svendsen, S.: A comparative study on substation types and network layouts in
connection with low-energy district heating systems. Energy Convers. Manag. 64, 551–561
(2012)
4. Wang, Y., You, S., Zheng, W., Zhang, H., Zheng, X., Miao, Q.: State space model and
robust control of plate heat exchanger for dynamic performance improvement. Appl. Therm.
Eng. 128, 1588–1604 (2018)
́
5. Dıaz, G., Sen, M., Yang, K., McClain, R.L.: Dynamic prediction and control of heat
exchangers using artificial neural networks. Int. J. Heat Mass Transf. 44, 1671–1679 (2001)
320 Y. Wen et al.

6. Vasičkaninová, A., Bakošová, M., Čirka, Ľ., Kalúz, M., Oravec, J.: Robust controller design
for a laboratory heat exchanger. Appl. Therm. Eng. 128, 1297–1309 (2018)
7. Wang, S., Yu, X., Liang, C., Zhang, Y.: Enhanced condensation heat transfer in air-
conditioner heat exchanger using superhydrophobic foils. Appl. Therm. Eng. 137, 758–766
(2018)
8. Formentin, S., De Filippi, P., Corno, M., Tanelli, M., Savaresi, S.M.: Data-driven design of
braking control systems. IEEE Trans. Control Syst. Technol. 21, 186–193 (2013)
9. Formentin, S., Karimi, A.: Enhancing statistical performance of data-driven controller tuning
via L2-regularization. Automatica 50, 1514–1520 (2014)
10. Yan, P., Liu, D., Wang, D., Ma, H.: Data-driven controller design for general MIMO
nonlinear systems via virtual reference feedback tuning and neural networks. Neurocom-
puting 171, 815–825 (2016)
11. Wang, L., Ni, H., Yang, R., Pardalos, P.M., Jia, L., Fei, M.: Intelligent virtual reference
feedback tuning and its application to heat treatment electric furnace control. Eng. Appl.
Artif. Intell. 46, 1–9 (2015)
12. Sörensen, K., Sevaux, M., Glover, F.: A history of metaheuristics. In: Martí, R., Panos, P.,
Resende, M. (eds.) Handbook of Heuristics, pp. 1–18. Springer, Cham (2018). https://doi.
org/10.1007/978-3-319-07153-4_4-1
13. Wang, L., Yang, R., Ni, H., Ye, W., Fei, M., Pardalos, P.M.: A human learning optimization
algorithm and its application to multi-dimensional knapsack problems. Appl. Soft Comput.
34, 736–743 (2015)
14. Campi, M.C., Lecchini, A., Savaresi, S.M.: Virtual reference feedback tuning: a direct
method for the design of feedback controllers. Automatica 38, 1337–1346 (2002)
15. von Krogh, G.: Individualist and collectivist perspectives on knowledge in organizations:
implications for information systems research. J. Strat. Inf. Syst. 18, 119–129 (2009)
16. Hansen, M.T.: The search-transfer problem: the role of weak ties in sharing knowledge
across organization subunits. Adm. Sci. Q. 44, 82–111 (1999)
17. Wang, L., Ni, H., Yang, R., Fei, M., Ye, W.: A simple human learning optimization
algorithm. In: Fei, M., Peng, C., Su, Z., Song, Y., Han, Q. (eds.) LSMS/ICSEE 2014. CCIS,
vol. 462, pp. 56–65. Springer, Heidelberg (2014). https://doi.org/10.1007/978-3-662-45261-
5_7
18. Wang, L., Ni, H., Yang, R., Pardalos, P.M., Du, X., Fei, M.: An adaptive simplified human
learning optimization algorithm. Inf. Sci. 320, 126–139 (2015)
19. Normey-Rico, J.E., Camacho, E.F.: Unified approach for robust dead-time compensator
design. J. Process Control 19, 38–47 (2009)
20. Åström, K.J., Hägglund, T.: Revisiting the Ziegler-Nichols step response method for PID
control. J. Process Control 14, 635–650 (2004)
21. Wang, L., Yang, R., Pardalos, P.M., Qian, L., Fei, M.: An adaptive fuzzy controller based on
harmony search and its application to power plant control. Int. J. Electr. Power Energy Syst.
53, 272–278 (2013)
 Design and Realization of 108MN
Multi-function Testing System

Shutao Zheng(&), Yu Yang, Zhiyong Qu, and Junwei Han

Harbin Institute of Technilogy, Harbin Heilong Provience, China

{zhengst77,hjw}@hit.edu.cn, [email protected],
[email protected]

Abstract. With the rapid development of China economics, large buildings are
beginning to appear. In order to ensure safety of the large buildings, 108MN
Multi-Function Testing System (MFTS) built by HIT with 6-DOF and large
loading force was proposed by China State Construction Engineering Corpo-
ration (CSCEC), which can research and evaluate the seismic performance of
structures such as long columns building, cross shape joints, shear walls and 8
floors full scale building. This paper shows you the base function, system main
structure parameter and performance characteristics of MFTS, detailed design
process was given, inverse kinematics model based Hydraulic servosystem and
compound control model of force and displacement were built to analysis the
performance, simulation analysis results show that the MFTS can meet design
requirements and the proposed controller has good abilities. The related
experiments will be tested after MFTS is built.

Keywords: Seismic
Inverse kinematics model
Hydraulic servosystem

Mixed controller

1 Introduction

As everyone knows, earthquake has extensive damage for building structure etc. which
lead to loss of many people’s lives and impact on the local economy (e.g.
1994 Northridge earthquake which occurred on January 17 with 60 deaths and property
damage between $13 billion and $50 billion; 1995 Hyogo-ken Nambu earthquake
which occurred on January 17 with over 5000 deaths and property damage $150 billion
[1]; 2008 wenchuan earthquake which occurred on May 12 with 69225 deaths and
property damage RMB 8451 hundreds of millions [2]). In order to reduce the destroy
associated with large earthquakes, the seismic performance of building structures
should be evaluated. Because of the simulation models are often inconformity to
capture complex response, and the seismic characteristics is difficult to be precise
evaluated when multi-dimensional and continuously varying are acting on large and
complex building structure during earthquakes [3, 4], many large scale structural
seismic tables have been built and used for performance evaluation. The University of
Minnesota’s Multi-Axial Subassemblage Testing (MAST) [5] System provides a
powerful tool for investigating effects of earthquakes with 6-DOF control system and
high speed data acquisition system, the Multi-Axial Testing System (MATS) in the

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 321–329, 2018.
https://doi.org/10.1007/978-981-13-2384-3_30
322 S. Zheng et al.

NCREE [6], E-Defense [7] which is the largest 3-D shaking table in the world, the
Multi-Axial Full-Scale Sub-Structured Testing and Simulation Facility (MUST-SIM) is
a state-of-the-art physical and analytical simulation facility [8], SRMD [9] consists of
prestressed concrete reaction frame and a moving platen with four hydrostatic low
friction hydraulic has the ability to test dampers and isolator bearings. Above those
loading facilities have the capacity of multiaxis, multifunctional and full-scale tests
which greatly fulfil the experiment demand for buildings.
China is the country with frequent earthquakes, large high-rise buildings and
complex structural bridges are becoming more and more with the rapid economic
development, 108MN Multi-Function Testing System (MFTS) with high speed and
large-scale loading force was proposed and developed jointed by CSCEC (China State
Construction Engineering Corporation) and HIT (Harbin Institute of Technology),
which can provide real-time, full-scale, 6-DOF dynamic and static performance testing
for bearings and all kinds of structural components including shear walls, long col-
umns, and complex cross shape joints. MFTS is full scale and multi scope powerful
experiment tool in earthquake engineering research field when it is finished.
First the paper gives the main construct and characteristics of MFTS,then shows the
detailed design and analysis process,the inverse kinematics model based hydraulic
servosystem of MFTS is derived for understanding the characteristics of the system, a
mixed controller with loading force in out loop and displacement in inner loop is
designed to resolve the control problem of MFTS, simulation and experiment shows
the MFTS has good control performance.

2 MFTS Description

Figure 1 shows the main components of the MFTS, 6DOF loading system is the core of
MFTS. The MFTS is placed in a 25 m  25 m  9 m (length * width * depth) base
pit. In order to ensure the system performance requirement, Hydraulic Power Unit
(HPU) is provided with 1750 L/min at 21 MPa for system supply and 350 L/min at
14 MPa for pre-stage supply for servo valve and 700 L/min at 21 MPa for system
supply or at 31 MPa for accumulator charging. The accumulator is composed of
15,480 L accumulator volume and 19,440 L gas volume pressurized up to 31 MPa.
The oil tank is 68 m3 for oil storage. The 6DOF loading system is made up of 18
actuators, and a loading platform which can simultaneously impose different vertical
loading force and double direction horizontal displacement. Isolation system with 52
vibration isolators is placed between reinforced concrete base and concrete base to
prevent the impact on the surrounding facilities when the MFTS is working. Four lift
cylinders can drive the movable platform move to adjust the distance between 6-DOF
loading system and movable platform for adapting to different parts.
The performance parameter of the MFTS are summarized in Table 1. MFTS can
satisfy four kinds of specimens including long column, shear wall, complex cross shape
joints, and rubber bearings. Testing objectives range from slow-speed unidirectional
testing for basic performance characterization to high-speed, 3-D dynamic testing for
energy-based analysis. The clearance for specimen installation is up to 9.1 m  6.6 m
in plane and up to 10 m in height between movable platform and 6-DOF loading
 Design and Realization of 108MN Multi-Function Testing System 323

LiŌ cylinder 4 sets

X Cylinder(2 set)
2*3MN
1500mm Stroke
Loading Frame 630L/min Servovalve

Movable Plaƞorm Loading Plaƞorm

6DOF Loading
System Z Slide Bearing Cylinder(6 set)
6*18MN
250mm Stroke
800L/min Servovalve

Reinforced concrete base

Hold-down Cylinder(4 set)
Pipeline and 4*3MN
maintenance channel 365mm Stroke
Y Cylinder(6 set)
IsolaƟon System
6*1.5MN
Concrete base 500mm Stroke
7600L/min+630L/min Servovalve

Fig. 1. MFTS overall structure and 6-DOF loading system

system. In order to solve the problem of 108MN large loading force in Z direction,
Static pressure support actuators with very low friction are adopted which are fixed
together with the loading platform.

Table 1. Perfomance parameter of MFTS

Degree of freedom x y z
Loading force 6MN(tension) ±6MN (dynamic) 108MN
10MN (compression) ±9MN (Compression)
Velocity (mm/s) ±53 ±1571 ±22
Displacement (mm) ±1500 ±500 250
Relative platform rot. ±2° (u) ±2° (h) ±10° (w)
Capacity of specimen 9.1 m (length)  6.6 m (width)  10 m (height)

3 System Modeling

According to the above description, MFTS is a complex loading system. System

modeling should be derived to analyze the system characteristics. This section shows
the system model and mixed controller adopt displacement and force.

3.1 Inverse Kinematics

6DOF loading system of the MFTS is composed of Y-direction hydraulics cylinder
with 6, X-direction hydraulic cylinder with 2, Z-direction hydraulic cylinder with 6, 4
324 S. Zheng et al.

hold-down actuators, and a loading platform. Figure 2 shows the structure diagram of
6-DOF loading system. Two coordinate systems were built to analyses the performance
of 6DOF loading system with body coordinate system Op-xpypzp fixed on the loading
platform and the inertial coordinate system Ob-xbybzb fixed on the ground. In the initial
condition, two coordinate systems are overlapped, Ob and Op at the top surface center
of the loading platform.

x2 x1

7205mm
y4 y1
z4 z1

2250mm
xp(xb)
6000mm

y5 h3 h1 2250mm y2
z5 z2
yp(yb)Op (Ob)
y6 h4 h2 y3
z6 z3

2500mm
4000mm 4550mm 5550mm

zp(zb)
yp(yb) Op(Ob)
1090mm

850mm

8600mm

Fig. 2. Schematic diagram of 6DOF loading system

Figure 3 shows you the three kinds of articulated joints in the system: x direction
(a) is SPS (P-prismatic joint, S-spherical joint, E-epitaxy), y direction (b) is PSS, and
the z direction ((c) is EPS which including z1 * z6 and h1 * h4).
The inverse kinematics model of x direction is as following:

li ¼ Rðqo Þai  bi þ qp ¼ kgi  bi k ði ¼ 1; 2Þ ð1Þ

where ai, and gi are the (3  1) vectors of upper joint point Ai in two coordinate system
respectively, bi is the (3  1) vector of the lower joint point Bi in Ob-xbybzb. Ri is a
rotation matrix of transformation from Op-xpypzp to Ob-xbybzb. qp ¼ ½ x y z T is the
 Design and Realization of 108MN Multi-Function Testing System 325

vai lni vai cni Ob

Ob
Ai gi
Ai gi
ci

bi lni
pi bi
Ci

vai
lni gi Ob

Ai
bi

Bi Bi
Bi

(a) x1 x2 Cylinder in x (b) y1-y6 Cyinder in y (c) z1-z6 Cylinder in z

SPS style PSS style EPS style

Fig. 3. Structure diagram of three types kinematic chains

position vector, qo ¼ ½ u h w T the orientation vector, k k is the Euclidean norm

operator, li is the ith actuator length.
For Y direction, the jth actuator length vector lj is as following:

pj ¼ pj  cj ¼ gj  bj  cj ðj ¼ 1; 2; 3; 4; 5; 6Þ ð2Þ

For Z direction, the kth actuator length vector lk is as following:

lk ¼ zak  zbk =znk ðk ¼ 1; 2; 3; 4; 5; 6Þ ð3Þ

According to the Eqs. (1), (2), and (3), 6DOF loading system’s inverse kinematic
can be obtained.

3.2 6DOF Loading System Jacobian Matrix

6DOF loading system Jacobian matrix shows the relationship between the loading
platform and the actuator lengths.
For x DOF, the velocity vai of upper joint point Ai and the velocity of x cylinder are
as following:

vai ¼ q_ p þ w  Rai ð4Þ

 
_l ¼ Jx q_ p x T ð5Þ
326 S. Zheng et al.

For y DOF, the velocity vcj of the pushrod and the velocity of y cylinder are as
following:


vci ¼ cTnj vaj ¼ cTnj q_ p þ cTnj w  Raj ð6Þ
 
_l ¼ Jy q_ p x T ð7Þ

Where cni is the unit vector.

For z DOF, the velocity vnk of the point Ak and the velocity of z cylinder are as
following:

vnk ¼ q_ p þ w  Rak ð8Þ

 
_l ¼ Jz q_ p x T ð9Þ

According to Eqs. (5), (7), (9), 6DOF loading system velocity Jacobian matrix J is
as following:

J ¼ ½ Jx Jy Jz  ð10Þ

3.3 Controller Design

Controller of MFTS is mainly used to control the displacement and pose of the loading
platform accurately, according to the system requirement Z direction is force control in
outer loop and position control in inner loop and other five direction is position control.
Figure 4 shows you the structure of controller, in order to speed up the simulation and

• •

Fig. 4. Structure diagram of MFTS controller

 Design and Realization of 108MN Multi-Function Testing System 327

check the control effect, Matlab/Simulink was used to develop the control system
model and XPC-target real time kernel are adopted to run control model, the sample
time is 1 ms.

4 Simulation Experiments

According to the requirement, seven cycle of Y direction is designed to performed to

validate the performance of the proposed controller, Fig. 5 shows the neoprene pad test
of MFTS. Four hold-down cylinders can provide 12MN downward force to prevent the
loading platform tip-over. Only z DOF is force mode and other five DOFs are position
mode. Because of the max velocity need the large flow and large power, only 7 cycles
(A = 0.2 m, f = 0.5 Hz) in y direction were designed to guarantee test and the start and
the end are half cycle and six whole cycle. Figure 6 shows that the curve has a good
position dynamic tracking ability as (a), z direction force keeps changing(c) between
57.5MN and 62.5MN because of the displacement fluctuation.

Fig. 5. Neoprene pad test of MFTS

328 S. Zheng et al.

a) Y displacement

b) Z displacement

c) Z force (Direct Force Control, DFC)

Fig. 6. Test curve of Neoprene pad based MFTS

5 Conclusion

In order to satisfy the requirement of China’s economic development, MFTS was

proposed to build. The system structure and performance characteristics of MFTS were
given which is a 6DOF static and dynamic real-time testing facility. In order to better
analysis the system characteristics, the 18 cylinders inverse kinematics model and
Jacobian matrix are built. Mix controller structure was given and simulation model was
built. Finally, an experiment of neoprene pad to check whether the mixed controller can
be used. The results demonstrate that MFTS can use the proposed controller to gain
good effect.
 Design and Realization of 108MN Multi-Function Testing System 329

References
1. Elnashai, A.S., Di Sarno, L.: Fundamentals of Earthquake Engineering, pp. 18–19. Wiley
Press, Chicheste (2008)
2. http://www.baike.com/wiki/wenchuanearthquake
3. Gao, C., Zheng, S., Cong, D., Han, J., Yang, Z., Sun, J.: Modeling and control of the CSCEC
multi-function testing system. J. Earthq. Eng. 22, 1–14 (2016)
4. Nakata, N., Spencer, B.F.: Multi-dimensional hybrid simulation using a six-actuator self-
reaction loading system. In: The 14th World Conference on Earthquake Engineering, China
(2008)
5. http://nees.umn.edu
6. Chen, J.C., Lin, T.H., Chen, P.C., Lin, K.C., Tsai, K.C.: Advanced seismic testing using the
multi-axial testing system (MATS) in NCREE. In: Proceedings of the 3rd International
Conference on Advances in Experimental Structural Engineering, San Francisco (2009)
7. Xianchang, L.: E-defense and research tests. J. Earthq. Eng. Eng. Vib. 28(4), 111–116 (2008)
8. http://www.structures.ucsd.edu/
9. Shortreed, J.S., Benzoni, G.: Characterization and testing of the caltrans seismic response
modification device test system. Philos. Trans. Roy. Soc. Math. Phys. Eng. Sci. 359(1786),
1829–1850 (2001)
 Research on Vibration Suppression for Boom
Luffing of Telescopic Boom Aerial Work
Platform

Ru-Min Teng(&), Xin Wang(&), Ji-Zhao Wang(&),

and Ji-Fei Liang(&)

Mechanical and Engineering Department, Dalian University of Technology,

Dalian 116024, Liaoning, People’s Republic of China
[email protected], [email protected],
[email protected], [email protected]

Abstract. With the increase in the length of the boom for aerial work vehicles,
the boom vibrates in rapid motion. Therefore, how to suppress the boom
vibration and achieve good stability and safety is significant. Based on the
research of telescopic boom aerial vehicles, the mathematical models of the link
hydraulic system of telescopic boom aerial vehicles are established. According
to the state equation of the boom system, the feedback gain vector is designed
based on the linear quadratic optimal control theory, the feedback value is fed
back to the signal input end of the solenoid valve to construct the closed loop
control system which restrains the boom vibration, and the simulation of the
closed loop control system is made to analyze the vibration suppression effect of
the boom. Through the control of luffing motion, a vibration control method of
telescopic boom aerial vehicle is put forward, which has a good reference
function for similar products.

Keywords: Aerial work platform
Hydraulic system
Vibration suppression

Feedback

1 Introduction

Aerial work platforms are widely used in building, manufacturing and maintenance.
The vibration will be generated on the platform when the boom luffing motion occurs.
Especially while the boom works with full boom extension, the vibration is obvious. So
it is necessary to study the vibration suppression for improving the quality of the aerial
work platform.
Now the researches mainly aimed at the study of the state equation, dynamic
analysis and vibration suppression. For example, the mathematical model of the folding
aerial platforms was established by Jing through the Lagrange equations [1]; Tan
proposed a rapid method of acquiring modal parameters for the high platform fire
engine [2]; Sawodny of the University of Stuttgart, Germany designed a kind of motion
planner for the boom elevation angle of the fire ladder car [3], and got the state space
equation of the boom system [4–6]. Based on study of vibration suppression, this paper
adds the electro-hydraulic proportional control model and a new control system is built.
© Springer Nature Singapore Pte Ltd. 2018
K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 330–341, 2018.
https://doi.org/10.1007/978-981-13-2384-3_31
 Research on Vibration Suppression for Boom Luffing 331

2 Modeling Electro-Hydraulic Proportional Control System

An electro-hydraulic proportional control system is applied to determinate the dis-

placement of the hydraulic cylinders, and the specific working theory of the system is
shown in Fig. 1.

Fig. 1. The schematic diagram of electro-hydraulic proportional position control system of

hydraulic system

1. Transfer function of the proportional amplifier

I0 ðsÞ
G1 ðsÞ ¼ ¼ Ka ð1Þ
Ii ðsÞ

where, Ka - Proportional amplification factor.

2. Transfer function of electro-hydraulic proportional valve
The transfer function of the proportional valve output flow Q to the input current I
is [8]:

QðsÞ Ksv
Gsv ¼ ¼ ð2Þ
IðsÞ ws22 þ 2f
wsv s þ 1
sv
sv

where, Ksv - Proportional valve gain, m3/s
A; m3/s
A

wsv - The natural frequency of the proportional valve, rad/s;
fsv - Proportional valve damping ratio.
3. The transfer function of the hydraulic cylinder
Modeling the transfer function based on symmetrical valve control theory [8–11].
Figure 2 is a schematic diagram of the asymmetric hydraulic cylinder based on the
four-way valve. Table 1 shows the parameters of the cylinder.
In the symmetrical valve asymmetric cylinder system, the effective area acting on
the hydraulic cylinder is inconsistent when the hydraulic cylinder is extended and
retracted. As a result, the open loop gain of the transfer function is not equal when the
hydraulic cylinder stays extended and retracted. Therefore, different situations need to
be considered separately [8].
332 R.-M. Teng et al.

xp

K
A1 A2

mt FL

V1 V2
Bp
p1 p2
xv
qL qL

ps p0

Fig. 2. The schematic diagram of four-way valve-controlled asymmetrical cylinder

Table 1. The parameters of luffing cylinder

Name Symbol Value Unit
Variable luff hydraulic Hydraulic cylinder inner diameter D 200 mm
cylinder Piston rod diameter d 140 mm
Cylinder stroke L 1902 mm
Rodless cavity area A1 31415.93 mm2
Rod cavity area A2 16024.65 mm2
Effective area ratio of two cavities n 0.51
Fluid equivalent elastic modulus be 1.4  109 Pa
Drive equivalent load m 11551.02 kg
Hydraulic damping ratio fn 0.3
Natural frequency wh 37.09 rad/s

4. The transfer function of the displacement sensor

Imax
Ks ¼ ð3Þ
Lmax

among, Imax - Maximum output current value of the sensor output, A;

Lmax - Maximum length of the sensor measurement, m.
5. Luffing hydraulic control system mathematical model
Through the solution of the transfer function, the final system block diagram is
shown in Fig. 3.
 Research on Vibration Suppression for Boom Luffing 333

Fig. 3. The block diagram of valve-controlled cylinder electro-hydraulic proportional position

control system

3 The Geometric Relationship Between the Expansion

of the Hydraulic Cylinder and the Luffing Angle

The luffing angle is achieved by controlling the displacement of the luffing hydraulic
cylinder, so the geometric relationship of the expansion and luffing angle of the
hydraulic cylinder can be derived. The structure dimension diagram of the telescopic
boom aerial platform is shown in Fig. 4.

Fig. 4. The dimension of telescopic boom aerial platform

d1 - the distance between A and B. The A indicates the hinges between the boom
and the turntable, B indicates the hinge point between the boom and the variable
luff hydraulic cylinder;
d2 - the distance between A and C. C indicates the hinge point between the
turntable and the variable luff hydraulic cylinder;
l - luffing hydraulic cylinder length;
h0 - When the boom is in horizontal position, the angle between AB and AC;
hðtÞ - luffing angle.

According to the triangular relationship:

d12 þ d22  l2
cosðh þ h0 Þ ¼ ð4Þ
2d1 d2
334 R.-M. Teng et al.

4 Vibration Suppression Based on Linear Quadratic Optimal

Control

According to the state equation of the boom system [12], the feedback gain vector is
designed by linear quadratic optimal control theory, and the feedback signal is input
into the electro-hydraulic proportional valve and the closed loop control system is
constructed to suppress the vibration of the boom. The control system is simulated in
the Simulink and the effect of vibration suppression is analyzed.

4.1 Design Based on Linear Quadratic Controller

In Wang master’s thesis [13], the state equation of the boom is:

x_ ¼ Ax þ Bu
y ¼ Cx þ Du

among,
2 3 2 3
0 1 0 0 0
6 17:99 0:424 0 0 7 6 7
6
A ¼4 7; B ¼ 6 10355 7
0 0 0 1 5 4 0 5
0 0 538:9 0:232 65822

C ¼ ½ 0:00191 0 0:00095 0 ; D ¼ 0

Output deviation is

eðtÞ ¼ yðtÞ ¼ CxðtÞ þ DuðtÞ ¼ ½ 0:00191 0 0:00095 0 xðtÞ  uðtÞ ð5Þ

Because output y(t) and input u(t) are one-dimensional functions, the weight
function Q and R can be represented by the constant Q and R. The optimal objective
function (6) is

Z1
1  
J¼ xT ðtÞðC T qCÞxðtÞ þ uT ðtÞðDT qD þ rÞuðtÞ þ 2xT ðtÞðCT qDÞuðtÞ dt ð6Þ
2
0

further,

Z1
1  T 
J¼ x ðtÞQ1 xðtÞ þ uT ðtÞR1 uðtÞ þ 2xT ðtÞN1 uðtÞ dt ð7Þ
2
0

among, Q1 ¼ CT qC; R1 ¼ DT qD þ r; N ¼ C T qD
 Research on Vibration Suppression for Boom Luffing 335

For q = 0.05, r = 1 and C, D is known, the feedback gain vector is obtained

through MATLAB optimal control toolbox:

½ 0:5318 0:6443 0:5337 0:0578   104

4.2 Vibration Response Analysis Based on Linear Quadratic Optimal

Control
Figure 5 is a simulation model based on the linear quadratic optimal control. After the
state equation coefficient matrix C and D are adjusted, the state variable of the boom is
obtained on the basis of the boom output angular displacement and the speed of the
boom tip. According to the previous feedback gain vector obtained, the feedback value
of the LQR optimal control is calculated and fed back to the signal input end of the
solenoid valve, and the flow of the solenoid valve is adjusted to control the motion of
the hydraulic cylinder, then, the vibration suppression of the boom is achieved.
Meanwhile, the linear acceleration response of boom tip can be obtained by deriving.

Fig. 5. The simulation model based on linear quadratic suboptimal control

4.2.1 Design of Luffing Angle

The boom luffing motion is designed that the boom luffing angle rises from 0° to 32° at
the first stage and stays with a interval. At the second stage, the angle changes to 47°,
then the boom luffing angle declines to 32°, then to 16°. The acceleration changed
curve of the boom luffing angle is shown in Fig. 6, and the specific data is shown in
Table 2.

4.2.2 Response Analysis of the Boom Angular Displacement

Figure 7 is the angular displacement response curve of the boom. At the beginning, it is
not easy to observe because of the superposition of the vibration and the luffing.
Therefore, in order to facilitate the expression, this paper uses the method of analyzing
336 R.-M. Teng et al.

Fig. 6. The curve of acceleration variation of boom luffing angel

Table 2. Acceleration variation of the boom angle

Operation Time Change time of Maximum Boom working time at
stage (s) boom acceleration acceleration uniform speed (s)
(s) (°/s2)
First rise 0–10 2 ±4 6
Second 24–30 1 ±6 4
rise
First 50–56 1 ±6 4
decline
Second 75–81 2 ±4 22
decline

Fig. 7. The response of angel displacement of boom tip

the residual vibration, and extracts the boom tip angular displacement curve in the four
stages. Figures 8 and 9 are respectively the boom angular displacement response curve
in rising stage and decline stage.
 Research on Vibration Suppression for Boom Luffing 337

Figure 8, The specific change of expected angular displacement, angular dis-

placement under uncontrolled condition and the angular displacement under LQR
control are shown in Tables 3 and 4.

Fig. 8. The angel displacement response of boom tip at the rising stage

Table 3. Comparison of data for the first ascending phase

Control method Peak Phases The difference between peaks Steady state
(°) (°) and troughs (°) value (°)
Desired angular – – – 32
displacement
uncontrolled 31.59 31.25 0.34 31.41
LQR optimal 32.14 31.97 0.17 31.97
control

Table 4. Data comparison for the second ascending phase

Control method Peak Phases The difference between peaks Steady state
(°) (°) and troughs (°) value (°)
Desired angular – – – 47
displacement
uncontrolled 46.78 45.58 1.2 46.21
LQR optimal 47.5 46.65 0.85 47
control

Figure 9, The specific change of expected angular displacement, angular dis-

placement under uncontrolled condition and the angular displacement under LQR
control are shown in Tables 5 and 6.
338 R.-M. Teng et al.

Fig. 9. The angel displacement response of boom tip at the decline stage

Table 5. Comparison of data for the first decline phase

Control method Peak Phases The difference between peaks Steady state
(°) (°) and troughs (°) value (°)
Desired angular – – – 32
displacement
uncontrolled 31.96 30.76 1.2 31.41
LQR optimal 32.32 31.47 0.85 31.97
control

Table 6. Comparison of data for the second decline phase

Control method Peak Phases The difference between peaks Steady state
(°) (°) and troughs (°) value (°)
Desired angular – – – 16
displacement
uncontrolled 15.87 15.53 0.34 15.7
LQR optimal 16.12 15.81 0.31 15.99
control

Through the response analysis of the angular displacement of the boom tip, it can
be found that the vibration of the boom is suppressed by the LQR optimal control. As
time increases, the suppression effect is more and more obvious. In addition, although
the angle displacement of the boom tip has some overshoot, the steady-state value is
basically same with the expected value, which eliminates the steady-state deviation.
And through data analysis, it can be found that the angular displacement response of
 Research on Vibration Suppression for Boom Luffing 339

the boom tip is basically consistent when the LQR optimal control system stays at the
rising and decline stages.

4.2.3 The Response Analysis of the Boom Tip Line Speed

The boom line speed analysis process is consistent with the analysis process of the
boom angle displacement, and no specific data analysis is performed here. Figure 10
shows the boom tip line speed response curve.
Figure 11 is the response curve of the boom line speed in the rising phase. Through
observation, it is known that the system based on LQR optimal control has more and
more obvious advantage, and the speed of the tip line speed tends to be stable after fifth
cycles.

Fig. 10. The velocity response of boom tip

Fig. 11. The velocity response of boom tip at rising stage

340 R.-M. Teng et al.

Figure 12 is the response curve of the tip line speed at the decline stage. Under the
same variable luffing angle, the effect of LQR optimal control is almost consistent with
the rising stage, namely, the vibration luffing and vibration attenuation time are basi-
cally consistent.

Fig. 12. The velocity response of boom tip at decline stage

Through the analysis of the line speed of the boom tip, the effect of the vibration
suppression of the boom line speed is not obvious in the first phase. As time increases,
the vibration suppression effect of the boom line speed based on the LQR optimal
control system is more obvious, and the line speed of the boom tip tends to be a stable
value of 0 after about four to five cycles. Compared with the uncontrolled system, the
system has a good effect on suppressing the vibration of the boom.

5 Conclusion

On the basis of the hydraulic system of the telescopic boom aerial vehicles, the
mathematical modeling about the electro-hydraulic proportional valve, the valve con-
trolled asymmetric hydraulic cylinder and the displacement sensor is carried out in this
paper for the luffing vibration; According to the structure dimension of the telescopic
boom aerial vehicle, the geometric relationship between the expansion of the hydraulic
cylinder and the elevation angle of the boom is established. And on the basis of the
state equation of the boom system, the feedback gain vector is designed by using the
linear quadratic optimal control theory. A complete system simulation model is built in
the Simulink and through the simulation analysis, the system based on the linear
quadratic optimal control system can eliminate the steady-state deviation between the
angular displacement of the boom tip and the expected angle displacement, and then
can achieve a good vibration suppression effect and provide ideas for faster and more
stable motion of the aerial platform vehicle.
 Research on Vibration Suppression for Boom Luffing 341

References
1. Jing, Y.L., Xiong, J.Q., Wang, F.L.: Application of input shaper in anti-swing control of
folding aerial work vehicles. Noise Vibr. Syst. (4), 38–42 (2011). (in Chinese)
2. Tan, J.H.: Dynamic analysis and vibration control of fire truck boom based on platform.
Xiangtan University, Xiangtan (2014). (in Chinese)
3. Sawodny, O., Aschemann, H., Bulach, A., Hofer, E.P.: Online generation of trajectories
considering kinematic constraints. In: Proceedings of 7th IEEE International Conference on
Methods & Models in Automation & Robotics, Miedzydroje, Poland, pp. 7133–7188 (2001)
4. Sawodny, O., Lambeck, S., Hildebrandt, A.: Trajectory generation for the trajectory tracking
control of a fire rescue turntable ladder. In: Proceedings of the Third International Workshop
on Robot Motion and Control, RoMoCo 2002, pp. 411–416 (2002)
5. Zuyev, A., Sawodny, O.: Observer design for a flexible manipulator model with a payload.
In: Proceedings of 45th IEEE International Conference on Decision and Control, pp. 4490–
4495 (2006)
6. Kharitonov, A., Zimmert, N., Sawodny, O.: Active oscillation damping of the fire-rescue
turntable ladder. In: Proceedings of the IEEE Conference on Control Applications, pp. 391–
396 (2007)
7. Yang, Z.R., Hua, K.Q., Xu, Y.: Electro-Hydraulic Ratio and Servo Control. Metallurgical
Industry Press, Beijing (2009). (in Chinese)
8. Wu, Z.S.: Hydraulic Control System. Higher Education Press, Beijing (2008). (in Chinese)
9. Zhao, C.L.: Adaptive controller and its application in valve-controlled asymmetric cylinder
speed system. Harbin Institute of Technology, Heilongjiang (2006). (in Chinese)
10. Zhang, Y.: Discrete adaptive controller in valve-controlled asymmetric cylinder electro-
hydraulic servo system. Harbin Institute of Technology, Heilongjiang (2008). (in Chinese)
11. Zhang, L.C.: Adaptive controller and its application in valve-controlled asymmetric cylinder
force system. Harbin Institute of Technology, Heilongjiang (2006). (in Chinese)
12. Gao, L.C.: Research on vibration suppression of vertical arm system of aerial work vehicle.
Dalian University of Technology, Liaoning (2015). (in Chinese)
13. Wang, J.Z.: Study on amplitude suppression of arm boom for telescopic boom aerial work
truck. Dalian University of Technology, Dalian (2016). (in Chinese)
 Intrusion Detection in SCADA System:
A Survey

Pu Zeng and Peng Zhou(&)

School of Mechatronic Engineering and Automation, Shanghai University,

Shanghai 200072, China
{plovehui,pzhou}@shu.edu.cn

Abstract. Nowadays, the industrial systems are more and more interconnected
with the outside world. However, the interconnection of Supervisory Control
and Data Acquisition (SCADA) systems with the outside world using Internet-
based standards introduce numerous vulnerabilities to these systems. Although
awareness is constantly rising, the SCADA systems are still exposed to serious
threats. In this paper, a review of Intrusion Detection and report results is
conducted in the surveyed works. In the end, we also discuss the potential
research directions on this topic.

Keywords: Intrusion Detection
SCADA system
Survey
Machine learning

1 Introduction

The importance of Supervisory Control and Data Acquisition (SCADA) systems has
increased in the past years [1]. And there are more and more cyber incidents in recent
years, which have caused serious damage to the critical infrastructures with very
important economic losses, especially after the Stuxnet incident in 2010 [2]. The
malware which was targeting the PLCs connected to a nuclear centrifuge, and its
ultimate goal was to sabotage the nuclear plant. Besides, a city water utility was hacked
(2011) [3], fully demonstrated the great threat of attacks to industrial control systems.
Intrusion Detection technology is becoming a major research direction for solving
the problem of attacks to SCADA systems which detect attacks by analyzing network
characteristics (such as network traffic, packet loss rate, etc.) after occurring of attacks.
During the last years, many papers about the Intrusion Detection in SCADA system
were reported [9–29]. Therefore, we will make a survey about them in this paper.
This paper is organized as follows: background on SCADA systems is reviewed in
Sect. 2. Taxonomy of detection approach is overviewed in Sect. 3. Intrusion Detection
methodologies are discussed in Sect. 4. This paper is ended with the final conclusion in
Sect. 6.

2 Background

In this section, we propose some necessary introduction of SCADA systems to support

the argumentation in next sections.

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 342–351, 2018.
https://doi.org/10.1007/978-981-13-2384-3_32
 Intrusion Detection in SCADA System: A Survey 343

2.1 Supervisory Control and Data Acquisition

Supervisory control: SCADA systems are highly distuributed, so that they can be used
to control critical infrastructure, which often distributes over thousands of square
kilometers. Though SCADA system, managers can monitor and control field sites over
long-distance communication networks.
Data acquisition: SCADA systems collect data from one or more distant facilities.
Then they display the data to the operators graphically or textually and transfer it to a
central computer facility so that to record it and log it in the system databases.
Data acquisition allows the operators not only monitor but also control the system
in real time, and to send control instructions to those facilities [1]. Therefore, the
system makes supervisory control possible. Common SCADA systems include
SCADA electric power grids, SCADA gas pipelines, SCADA water networks and
SCADA storage tanks. In this paper, we mainly survey papers which make research in
electric power systems.

2.2 Components of SCADA System

There are four main components of SCADA systems [39]:
• The Human Machine Interface (HMI) is not only a hardware but also a software
which can display process status information, reports, and other information to
systems.
• The Master Terminal Unit (MTU), which can also be called SCADA Server, give all
commands to other different components of SCADA system. The MTU gathers data,
stores some information about system and pass other to associated components.
• The Remote Terminal Units (RTUs) or remote telemetry units, can converts sensor
signals to digital data. So it can support SCADA remote stations as a data acqui-
sition and control unit designed. The RTUs gather information from different
devices and keep it in the memory. When MTU needs it, RTUs code and transmit
the information back to the MTU.
• The Programmable Logic Controllers (PLCs) are small industrial computers which
are connected to the physical process sensors and convert sensor signals to digital
data. Because PLCs are more economical, flexible and versatile than RTUs, they are
sometimes implemented as field devices instead of RTUs.

2.3 Security of SCADA System

Threats by insiders include employees who introduce malware into control systems
which are extremely dangerous. Knowledgeable insiders could control systems without
connected networks [5]. And other kinds of security threats would also cause serious
damage to critical infrastructure, especially threats by nation-states which could get
many resources for attacking the SCADA system by highly capable organizations [6]
(Fig. 1).
Before we make research of intrusion detection, we should make investigation of
SCADA vulnerabilities, which can help us understand about how they are exploited for
attacks. The SCADA systems can be divided into four major types [7]: security policy,
344 P. Zeng and P. Zhou

Fig. 1. The SCADA system [1]

communication protocol, architectural and software and hardware vulnerabilities.

Nowadays, SCADA system have become more and more complicated not only in
software but also in hardware. The model of attacks is shown in Fig. 2.

Fig. 2. Attacks in SCADA system [4]

 Intrusion Detection in SCADA System: A Survey 345

3 Taxonomy of Detection Method

In this section, we introduce taxonomy of detection approach, which can be roughly

divided into three categories: data-based method, model-based method and information
security method [3].

3.1 Data-Based Method

The data-based method uses pattern recognition techniques and machine learning to
analyze data which we acquires from SCADA system. It do not require models of
systems and attacks for the detection. The training dataset is observed, such as com-
mand signals, sensor measurements and control signals. The benefit of data-based is we
can study classification without using the machine learning model directly. There are
two main categories of machine learning which is divided into supervised and unsu-
pervised. Classification and regression represent the supervised learning, while clus-
tering represents the unsupervised learning.

3.2 Model-Based Method

The model-based method which develops the SCADA systems models under both
proper functioning and different attack scenarios. The comparison between system
outputs and observations model makes a decision about whether system is under attack.
The system could be considered to be suffer from attacking, when data are not con-
sistent with the normal. This method often requires professional prior knowledge which
is based to build a model for attacks detection. And then, we can check the consistency
of the observed data with the output of this attack pattern, if the model we build are
available.

3.3 Information Security Method

Information security method focuses broadly on three main categories [8]. Firstly, data
confidentiality, which is generally performed by access control methods or authenti-
cation methods. Secondly, the integrity of data is related to the trustworthiness of data.
And the integrity of data is implemented by prevention mechanisms and detection
mechanisms. Thirdly, the availability of data is when you need the information or
resources you can get them from system. The Dos attacks is a classic example of the
data which is unavailability. By using the characteristic of information security, we can
improve the security of SCADA systems against attacks

4 Intrusion Detection Methodologies

In this section, we survey existing Intrusion Detection methodologies recent years.

Proposals are divided according to the taxonomy described in Sect. 3.
Lahza et al. [9] propose a method which use knowledge of domain to construct new
features which include three SCADA-protocol-specific steps. The result of this method
346 P. Zeng and P. Zhou

have a better performance than those described in previous research used with several
other classifiers.
Lin et al. [10] propose a semantic analysis framework. The framework combines
Intrusion Detection System (IDS) with a power flow analysis that can estimate the
execution results of control commands.
Wan et al. [11] propose a method which use One Class Support Vector Machine
(OC-SVM) as the first layer to detect misbehaviors. For the process data behavior, they
use Reconstruction Error based on Kernel Principal Component Analysis (RE-KPCA)
method to detect data exceptions. This approach has a better performance in classifi-
cation accuracy than those methods mentioned in the paper.
Shitharth et al. [12] propose an Intrusion Weighted Particle-based Cuckoo Search
Optimization (IWP-CSO) method which combines Hierarchical Neuron Architecture
based Neural Network (HNA-NN) techniques is proposed in this paper. The IWP-CSO
method is used to select the best feature when the input data is clustered. And then, the
HNA-NN is used to classify the intrusions in SCADA system.
Sadhasivan et al. [13] propose an approach which uses the Linear Weighted
Cuckoo Search Optimization (LWCSO) algorithm to get the best features. A Proba-
bilistic Kernel Model (PKM) is used to update the weight function to form a cluster that
represents the best features.
Kleinmann et al. [14] use algorithms which first build a Discrete-Time Markov
Chain (DTMC). Then, the symbols are split into sets. And they use the Deterministic
Finite Automata (DFA) in the work to build a graph model.
Giuseppe et al. [15] introduce models that are developed with the Control Theory
perspective. This research highlights the role of determining the reaction that the
monitoring system may experience when it undergoes different types of cyberattacks.
Yang et al. [16] propose detection method based on expert knowledge. The
advantage of this method is that it does not take into account the actual operating
environment because it has clear advantages over existing more general proposals.
Cruz et al. [17] use an IT-OCSVM method which perform intrusion detection with
high accuracy and low overhead in a time step. The algorithm introduced at most a
96.3% detection accuracy and a 2.5% false alarm rate in the experiments.
Bermudez et al. [18] propose a system named FieldHunter. The system which
extracts fields and infers their types automatically. And it can detect realistic zero-day
attacks on SCADA networks, as result of the valuable information about network
protocol specification it provides.
Almalawi et al. [19] propose a method which uses a data-driven clustering method
and the DBSCAN is selected as the base clustering algorithms. And the method can
identify whether the given system is normal or critical.
Zhou et al. [20] propose a method which use the hidden Markov model. The
method is designed to distinguish between attacks and failures and overcome the
shortcomings of anomaly detection. The research limits the false alarm rate to less than
1.61%. Besides, the designed system has little influence on the performance of control
system.
Erez et al. [21] propose a new domain-aware intrusion detection system. The
system can detect unusual changes in SCADA control register values within
Modbus/TCP protocols, which use Incremental classification and Single-window
 Intrusion Detection in SCADA System: A Survey 347

classification. The experimental result shows that the system have a 93% classification
accuracy. Besides, it has a 0.86% false alarm rate.
Nader et al. [22] propose a work which can detect anomaly. The anomaly will
happen after malicious have already passed the IDS. This paper uses machine learning
method for anomaly detection by using one-class classification method including
SVDD and KPCA.
Yang et al. [23] propose a system which provide a considerate method to solve
intrusion detection, by analyzing multiple attributes. And the system uses whitelists and
behavior-based concept, so as to make SCADA systems more safety.
Goldenberg et al. [24] propose a method which use the DFA algorithm by an HMI-
PLC channel. The DFA-based intrusion detection system delve into Modbus/TCP
packets and generate a very detailed model.
Carcano et al. [25] present a method which is based on the critical state analysis and
state proximity concepts. Because the detection is based on the system evolution
analysis, and not on the analysis of the attack evolution, the IDS can detect also “zero-
day attacks”.
Yoo et al. [26] propose a method which extracts the main fields from the
GOOSE/MMS packets using 3-phase preprocessing. And then they removed outliers
by an EM algorithm. Besides, it uses the OCSVM algorithm to learn and detection
normal behavior.
Hink et al. [27] propose a power system classification framwork by using machine
learning method. With the JRipper + AdaBoost approach, they could dependably
classify SCADA system with low false positive rates.
Pan et al. [28] propose a method which uses a Bayesian network graph and gets
data logs from the perspective of relay in the SCADA system. The resulting IDS was
used to classify data get from the perspective of each relay station.
Pan et al. [29] propose another approach which uses the common paths mining-
based method. The common paths is complex specifications which could describe
patterns of system behavior connected with SCADA systems. The IDS which is
mentioned in this paper correctly classifies 90.4% of tested instances. And the average
detection accuracy of this IDS for zero-day attack is 73.43%.

5 Discussion

In this section, we make a discussion about intrusion detection methodologies provided

in Sect. 4. We compare recent papers and make Tables 1 and 2 as follows:
Table 1 shows the trend of writing papers about Intrusion Detection in SCADA
systems.
As Table 1 shows, the Data-based approach is more used than the other two
approaches because of the rapid development of machine learning techniques. And we
can see that there are more and more researchers which focus on the security of
SCADA systems in recent years.
Table 2 shows the surveyed works of Intrusion Detection in SCADA systems by
listing the following metrics: protocol, datasets or testbed, and main detection tech-
nique. As Table 2 lists, the systems which use the protocol of Modbus are studied most
348 P. Zeng and P. Zhou

Table 1. Papers about Intrusion Detection in SCADA systems

Years Data-based method Model-based method Information security method
2015–2018 [9, 11–13, 17, 19–21, 29] [10, 14–16, 18]
2011–2014 [22, 26, 27] [24, 28] [23, 25]

Table 2. Comparison of the surveyed works

Author Ref. Protocol Datasets or testbed Main detect. technique
Lahza et al. [9] MMS/GOOSE Power system SVM/DT/NN
Lin et al. [10] DNP3 Power system Semantic analysis framework
Wan et al. [11] Modbus/TCP TEP system OCSVM and RE-KPCA
Shitharth [12] No mentioned ADFA IDS datasets IWP-CSO and HNA-NN
et al.
Sadhasivan [13] Modbus MSU SCADA LWCSO and PKM
et al. datasets
Kleinmann [14] S7-0x72 S7-0x72 datasets DTMC and DFA
et al.
Giuseppe [15] Modbus/TCP Water distribution Control Theory
et al. system
Yang et al. [16] GOOSE/SMV Power system multi-layered IDS
Cruz et al. [17] TCP/IP Power system IT-OCSVM
Bermudez [18] Binary Power system FieldHunter
et al. protocols
Almalawi [19] Modbus/TCP Water treatment DBSCAN
et al. system
Zhou et al. [20] Powerlink/CAN TEP system HMM
Erez et al. [21] Modbus/TCP Power system Incremental classification and
Single-window classification
Nader et al. [22] No mentioned Gas Pipeline Testbed SVDD and KPCA
and water treatment
Yang et al. [23] IEC 60870-5 Photovoltaic system Access-Control Whitelists
Goldenberg [24] Modbus/TCP Power system DFA
et al.
Carcano [25] Modbus Boiling Water Reactor The critical state validation
et al.
Yoo H et al. [26] MMS/GOOSE Power system EM and OCSVM
Hink et al. [27] Modbus MSU SCADA JRipper + AdaBoost
datasets
Pan et al. [28] Modbus MSU SCADA Bayesian network
datasets
Pan et al. [29] Modbus MSU SCADA Common path algorithm
datasets
 Intrusion Detection in SCADA System: A Survey 349

frequently in the papers we survey. And then, the power systems are the main field that
investigators study in, which causes frequent using of Modbus protocol. Mississippi
State University SCADA Security Laboratory integrates control systems within
numerous critical infrastructure industries to create a testbed and datasets (MSU
SCADA datasets) [30]. These datasets are very widely used in various related studies.

6 Conclusion

We present a survey paper comprising three main contributions: (i) a review of current
proposals of Intrusion Detection in SCADA system, (ii) a taxonomy to classify existing
intrusion detection methodologies and (iii) a comparison of the surveyed works by
which we can see the research trends.
Intrusion Detection in SCADA system is still a developing field. With the devel-
opment of machine learning algorithms [31], researchers achieve a higher detection
accuracy. In the future research, we can combine the machine learning methodologies
with other approach (such as trust management systems [32, 33], adversarial learning
[34, 35] and game theory [36–38]) to get better practical results.

Acknowledgments. This work was partially supported by the National Natural Science Foun-
dation of China (Nos. 61502293, 61775058 and 61633016), the Shanghai Young Eastern Scholar
Program (No. QD2016030), the Young Teachers’ Training Program for Shanghai College &
University, the Science and Technology Commission of Shanghai Municipality (Nos. 18ZR141
5000 and 17511107002) and the Shanghai Key Laboratory of Power Station Automation
Technology.

References
1. Nader, P.: One-class classification for cyber intrusion detection in industrial systems. IEEE
Trans. Ind. Inf. 10(4), 2308–2317 (2015)
2. Chen, T.M., Abu-Nimeh, S.: Lessons from Stuxnet. Computer 44(4), 91–93 (2011)
3. Do, V.L., Fillatre, L., Nikiforov, I., et al.: Feature article: security of SCADA systems
against cyber-physical attacks. IEEE Aerosp. Electron. Syst. Mag. 32(5), 28–45 (2017)
4. Dong, Y., Zhou, P.: Jamming attacks against control systems: a survey. In: Yue, D., Peng,
C., Du, D., Zhang, T., Zheng, M., Han, Q. (eds.) LSMS/ICSEE -2017. CCIS, vol. 762,
pp. 566–574. Springer, Singapore (2017). https://doi.org/10.1007/978-981-10-6373-2_57
5. Cardenas, A., Amin, S., Sinopoli, B., et al.: Challenges for securing cyber physical systems.
In: First Workshop on Cyber-physical Systems Security, vol. 2010, pp. 363–369 (2010)
6. Kushner, D.: The Real Story of Stuxnet. IEEE Spectr. 50(3), 48–53 (2013)
7. Fovino, I.N., Coletta, A., Masera, M.: Taxonomy of security solutions for the SCADA
sector. Joint Research Centre of the European Commission (2010)
8. Bishop M.: Introduction to computer security. In: Euromicro International Conference on
Parallel, Distributed, and Network-Based Processing, pp. 170–174. IEEE (2004)
9. Lahza, H., Radke, K., Foo, E.: Applying domain-specific knowledge to construct features for
detecting distributed denial-of-service attacks on the GOOSE and MMS protocols. Int.
J. Crit. Infrastruct. Prot. 20, 48–67 (2017)
350 P. Zeng and P. Zhou

10. Lin, H., Slagell, A., Kalbarczyk, Z.T., et al.: Runtime semantic security analysis to detect
and mitigate control-related attacks in power grids. IEEE Trans. Smart Grid 9(1), 163–178
(2017)
11. Wan, M., Shang, W., Zeng, P.: Double behavior characteristics for one-class classification
anomaly detection in networked control systems. IEEE Trans. Inf. Forensics Secur. 12(12),
3011–3023 (2017)
12. Shitharth, S., Winston, D.P.: An enhanced optimization based algorithm for intrusion
detection in SCADA network. J. Comput. Secur. 70, 16–26 (2017)
13. Sadhasivan, D.K., Balasubramanian, K.: A novel LWCSO-PKM-based feature optimization
and classification of attack types in SCADA network. Arab. J. Sci. Eng. 42(8), 3435–3449
(2017)
14. Kleinmann, A., Wool, A.: Automatic construction of statechart-based anomaly detection
models for multi-threaded industrial control systems. ACM Trans. Intell. Syst. Technol. 8(4),
55 (2016)
15. Giuseppe, B., Miciolino, E.E., Pascucci, F.: Monitoring system reaction in cyber-physical
testbed under cyber-attacks. Comput. Electr. Eng. 59, 86–98 (2017)
16. Yang, Y., Xu, H.Q., Gao, L., et al.: Multidimensional intrusion detection system for IEC
61850-based SCADA networks. IEEE Trans. Power Deliv. 32(2), 1068–1078 (2017)
17. Cruz, T., Rosa, L., Proença, J., et al.: A cybersecurity detection framework for supervisory
control and data acquisition systems. IEEE Trans. Ind. Inf. 12(6), 2236–2246 (2017)
18. Bermudez, I., Iliofotou, M., et al.: Towards automatic protocol field inference. Comput.
Commun. 84(C), 40–51 (2016)
19. Almalawi, A., Fahad, A., Tari, Z., et al.: An efficient data-driven clustering technique to
detect attacks in SCADA systems. IEEE Trans. Inf. Forensics Secur. 11(5), 893–906 (2016)
20. Zhou, C., Huang, S., Xiong, N., et al.: Design and analysis of multimodel-based anomaly
intrusion detection systems in industrial process automation. IEEE Trans. Syst. Man Cybern.
Syst. 45(10), 1345–1360 (2015)
21. Erez, N., Wool, A.: Control variable classification, modeling and anomaly detection in
Modbus/TCP SCADA systems. Int. J. Crit. Infrastruct. Prot. 10, 59–70 (2015)
22. Nader, P., Honeine, P., Beauseroy, P.: Ip-norms in one-class classification for intrusion
detection in SCADA systems. IEEE Trans. Ind. Inf. 10(4), 2308–2317 (2014)
23. Yang, Y., Mclaughlin, K., Sezer, S., et al.: Multiattribute SCADA-specific intrusion
detection system for power networks. IEEE Trans. Power Deliv. 29(3), 1092–1102 (2014)
24. Goldenberg, N., Wool, A.: Accurate modeling of Modbus/TCP for intrusion detection in
SCADA systems. Int. J. Crit. Infrastruct. Prot. 6(2), 63–75 (2013)
25. Carcano, A., Coletta, A., Guglielmi, M., et al.: A multidimensional critical state analysis for
detecting intrusions in SCADA systems. IEEE Trans. Ind. Inf. 7(2), 179–186 (2011)
26. Yoo, H., Shon, T.: Novel approach for detecting network anomalies for substation
automation based on IEC 61850. Multimed. Tools Appl. 74(1), 303–318 (2015)
27. Hink, R.C.B., Beaver, J.M., Buckner, M.A., et al.: Machine learning for power system
disturbance and cyber-attack discrimination. In: International Symposium on Resilient
Control Systems, pp. 1–8. IEEE (2014)
28. Pan, S., Morris, T., Adhikari, U.: A specification-based intrusion detection framework for
cyber-physical environment in electric power system. Int. J. Netw. Secur. 17(2), 174–188
(2015)
29. Pan, S., Morris, T., Adhikari, U.: Developing a hybrid intrusion detection system using data
mining for power systems. IEEE Trans. Smart Grid 6(6), 3104–3113 (2015)
30. Morris, T., Srivastava, A., Reaves, B., et al.: A control system testbed to validate critical
infrastructure protection concepts. Int. J. Crit. Infrastruct. Prot. 4(2), 88–103 (2011)
 Intrusion Detection in SCADA System: A Survey 351

31. Mitchell, T.M., Carbonell, J.G., Michalski, R.S.: Machine Learning. McGraw-Hill, New
York (2003)
32. Zhou, P., Chang, R., Gu, X., et al.: Magic train: design of measurement methods against
bandwidth inflation attacks. IEEE Trans. Dependable Secure Comput. PP(99), 1 (2018)
33. Zhou, P., Jiang, S., Irissappane, A., et al.: Toward energy-efficient trust system through
watchdog optimization for WSNs. IEEE Trans. Inf. Forensics Secur. 10(3), 613–625 (2015)
34. Lowd, D., Meek, C.: Adversarial learning. In: Eleventh ACM SIGKDD International
Conference on Knowledge Discovery in Data Mining, pp. 641–647. ACM (2005)
35. Dalvi, N., Domingos, P., Sanghai, S., et al.: Adversarial classification. In: Tenth ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 99–108.
ACM (2004)
36. Li, Y., Shi, L., Cheng, P., et al.: Jamming attacks on remote state estimation in cyber-
physical systems: a game-theoretic approach. IEEE Trans. Autom. Control 60, 2831–2836
(2015)
37. Yuan, Y., Yuan, H., Guo, L., et al.: Resilient control of networked control system under DoS
attacks: a unified game approach. IEEE Trans. Ind. Inf. 12(5), 1786–1794 (2016)
38. Huang, J.Y., Liao, I.E., Chung, Y.F., et al.: Shielding wireless sensor network using
markovian intrusion detection system with attack pattern mining. Inf. Sci. 231, 32–44 (2013)
39. Stouffer, K., Falco, J., Kent, K.: Guide to supervisory control and data acquisition (SCADA)
and industrial control systems security, Spin (2006)
 Intelligent Servo Feedback Control
for Hydrostatic Journal Bearing

Waheed Ur Rehman1 , Jiang Guiyun1(&), Nadeem Iqbal1,

Luo Yuanxin1, Wang Yongqin1, Shafiq Ur Rehman2, Shamsa Bibi2,
Farrukh Saleem1, Irfan Azhar3, and Muhammad Shoaib1
1
The State Key Laboratory of Mechanical Transmission, Chongqing University,
Chongqing 400030, China
{wrehman87,L1600316,yxluo,Wyq,L1700258,
L1700263}@cqu.edu.cn, [email protected]
2
Chemistry Department, University of Agriculture Faisalabad,
Faisalabad, Pakistan
{Shafiq.urrehman,Shamsa.shafiq}@uaf.edu.pk
3
Koc University, Istanbul, Turkey
[email protected]

Abstract. The purpose of current research work is to improve dynamics

characteristics of hydrostatic journal bearing which is integral part of high speed
and heavy load machinery nowadays. Current work presents hydrostatic journal
bearing with servo control. Mathematical model is derived for hydrostatic
journal bearing and two control strategies were presented. Result shows that
hydrostatic journal bearing with self-tuning PID control has better results than
PID control. The performance of active hydrostatic journal bearing under two
different control strategies were checked with respect to different conditions of
speed, viscosity, load, pressure, bearing clearance. The numerical result shows
that proposed hydrostatic journal because of active lubrication has better per-
formance including stability and controllability, high stiffness, faster response,
strong resistance under Self tuning PID control. Furthermore, proposed hydro-
static journal bearing with servo control has no eccentricity under new equi-
librium position after being applied load which is big advantage on conventional
hydrostatic journal that always face some amount of eccentricity under external
load due to absent of servo feedback. All simulations were performed in
Matlab/Simulink. The numerical result shows that proposed active hydrostatic
journal due to active lubrication has good performance including stability and
controllability, high stiffness, faster response, strong resistance in
Matlab/Simulink which shows effectiveness of proposed system.

Keywords: Self tuning PID
PID
Tribology
Fluid mechanics

Hydrostatic journal bearing
Fluid film lubrication theory
Servo valve etc.

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 352–364, 2018.
https://doi.org/10.1007/978-981-13-2384-3_33
 Intelligent Servo Feedback Control for Hydrostatic Journal Bearing 353

1 Introduction

Rotary equipment’s for precision applications have been long utilized hydrostatic
journal bearing due to their many useful features. These journal bearing offer greater
position accuracy, enhanced stability, and low friction. For the reasons mentioned
earlier, numerous researchers have conducted studies on hydrostatic journal bearings.
Matthew et al. [1] studied the dynamic performance using linear and non-linear
numerical methods for complaint cylindrical journal bearings. Kim et al. [2] derived
equations of perturbation and general Reynold equation. He accomplished this by
analyzing the behavior of journal bearings and studying effects of different recess
shapes on it. Thomsen and Klit [3] further focused their studies on analyzing the effects
of compliant layers over the behavior of hydrodynamic journal bearings. De Pellegrin
et al. [4] derived isothermal and iso-viscous models for hydrostatic journal bearing by
keeping in view the behavioral characteristics of tilt pad type hydrostatic journal
bearing. Phalle et al. [5] studied the effect of wear on performance of a hybrid journal
bearing system which employ membrane type of compensation. Above mentioned
research findings demonstrated to play a vibrant role in successful improvement of the
performance under normal operational environmental conditions in traditionally used
journal bearings.
Considering the operational conditions, these conditions severely restrict behavior
of journal bearings and also play an important part too. So numerous studies have been
conducted in this field by great number of researchers to enhance the dynamic char-
acteristics of these bearing, the rigidness of fluid film and rigid rotor’s stability. Bouaziz
et al. [6] studied the dynamic behavior of the misaligned rotor. Sharma et al. [7]
examined the wear-tear phenomena on a hybrid conical journal bearing with four
pockets. The research findings of this study proposed that with augmented radial
position accuracy and stationarity of traditional journal bearings, we can directly
influence the behavior of mechanical equipment under harsh operational conditions. It
is also tried to control oil film thickness in [8, 9] and some control strategies were
presented.
After this above conclusion, it is probable to increase the performance of a journal
bearing using an active control technology. Using several types of actuators like
piezoelectric, magnetic & hydraulic to control the vibrations in rotor (Nicoletti and
Santos [10]. The Morosi and Santos [11] have effectively validated that it is likely to
apply active lubrication to the gas journal. The Santos and Watanabe [12] further
explored the applications of the active lubrication in journal bearings, in order to get
enhanced damping coefficients and to have a reduced cross-coupling stiffness. With
that it considerably augmented the threshold for the stability of the journal bearing.
Santos suggested an active control for the control of vibration in a flexible rotor, using
an output feedback controller and computed gains of the controllers. Kim and Lee [13]
focused on improvement of dynamic pressure by examining the properties of a sealed
squeeze film type damper. And the results of their study showed that active control
technology is effective in dampers. Ho et al. [14] worked and analysed the effects of
thrust magnetic bearings on the stability and bifurcation of the flexible rotor with active
magnetic bearing system. Nicoletti and Santos [15] focused on the analysis of the
354 W. U. Rehman et al.

frequency response tilt type active hydrostatic bearing. Estupinan and Santos [16]
effectively studied multi-body systems which were inter-connected through thin films
of fluid and derived the mathematical modeling and showed that it can be easily
controlled by using several types of actuators. Haugaard and Santos [17] and Nicoletti
and Santos [15] studied the dynamic performance in theory and focused on active
lubrication of tilting-pad journal bearings. These research studies were concentrated on
customary analysis of stiffness in bearings and the damping in the frequency domain.
Most of the precision rotary equipment have been extensively and more commonly
employing these hydrostatic journal bearings, which led to the requirement of having a
bearing with improved performance and reliability. Santos and Watanabe [12] worked
on the extension of the stability-threshold of the hydrostatic journal bearing and studied
the viability of reduced vibrations in theory. With most of the researchers discovering
these areas, none of them have previously worked or studied on controlling the journal
position accuracy and precision of a hydrostatic journal bearing under external load
conditions and not much existing literature addresses this topic.
In our pursuit to realize our objectives, first we made mathematical model for
hydrostatic journal bearing using servo control and then two different control strategies
are presented. Self-tuning PID control and PID control. A numerical comparison has
been done in Result and discussion part which show effectives of proposed system.

2 Problem Formulation for Coupled System

The Structure of hydrostatic journal bearing is shown in Fig. 1, it consists of servo

valve, filter, regulator and some electronics components. Hydrostatic journal bearing
has advantage over hydrodynamic bearing that is surface of journal remain separated
from bearing under different speeds and loads. So that is why, hydrostatic journal
bearing has less wear and less friction as compared to hydrodynamic bearing under low
speed or under starting condition of rotary machinery. The performance of hydrostatic
journal bearing depends upon gap between journal and surface. Throttlers have big
influence on efficiency of hydrostatic journal bearing. Fixed throttlers have problem as
they are limited in term of stiffness and load carrying capacity. This paper present
active throttling device called servo valve to control clearance gap between inner and
outer periphery of hydrostatic journal bearing.

3 Modelling for Hydrostatic Journal Bearing System

Servo valve is key component of proposed system. It controls the flow towards bearing.
Flow through servo valve is describe by second order differential equation in [18].
The coefficients such as gain, frequency and damping are obtained from manufacturers
[19, 20].
Let suppose non linearities are ignored for derivation of a mathematical model of
Electro hydraulic servo valve (EHSV). The EHSV current is is related to spool dis-
placement xv by first order transfer function [21, 22].
 Intelligent Servo Feedback Control for Hydrostatic Journal Bearing 355

Arduino UNO Hardware

Ux

Servo
valve
Buffer op amp

Comparator

Control Output

Download Simulink Program into Arduino UNO Hardware

PRV Oil recess 1 Sensor
Buffer op amp
Oil recess 2
PRV
Uy
sor
Sen

PRV
Sha

o
e
Oil recess 4 J

Senso
Servo

r
valve

PRV

Comparator Sensor Oil recess 3

Oil Tank

Ps Computer

Filter

Oil Tank

Fig. 1. Hydrostatic journal bearing attached with feedback servo control

sv x_ vi ¼ Ki ii  xvi i 2 x; y ð1Þ

Flow for hydrostatic bearing along recess aligned in vertical and horizontal
direction is given by [23];
8 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
>
> 2
>
>
< Q2 ¼ Cd wxvx q ðPs  P0  Px Þ
>
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð2Þ
>
>
>
> 2
> Q4 ¼ Cd wxvx
: ðPs  P0 þ Px Þ
q

Similarly for Y direction, is given by;

8 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
>
> 2
>
>
< Q1 ¼ Cd wxvy q ðPs  P0  Py Þ
>
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð3Þ
>
>
>
> 2
>
: Q3 ¼ Cd wxvy q ðPs  P0 þ Py Þ
356 W. U. Rehman et al.

Hydrostatic journal bearing which has thin land normally we obtain its transfer
function by analyzing effect of changes of pressure and oil film thickness into and out
of bearing recess. Whenever there is offset distance then squeezing effect come into
play which is given by [23],

dhi
Qsqueeze ¼ Ae ð4Þ
dt

Flow is due to effect of pressure at one end. Fluid between two circular surfaces
faces restriction due to shear stress. The motion of journal produces surface velocity
which is zero at boundary and is maximum at place where oil film thickness center is
present. Let suppose pressure across flow is constant then equilibrium force on element
of width b is
   
dp ds
pbdy  p þ dx bdy  sbdx þ s þ dy bdx ¼ 0
dx dy

This leads to

ds dp
¼ ð5Þ
dy dx

Substituting s ¼ gðdu=dyÞ into Eq. (5).

d 2 u dp
g ¼ ð6Þ
dy2 dx

Applying boundary conditions and integrating such as u ¼ 0 & u = U when y ¼ 0

& y ¼ h; respectively, to get velocity distribution that is;

Uy 1 dp  
u¼  yh  y2 ð7Þ
h 2g dx

Flow rate is often integral of velocity that is given by;

1 bh3 dp
QPU ¼ Uhb  ð8Þ
2 12g dx

Let suppose there is a recess or pocket of width b and length l, then pressure
gradient in an Eq. (8) will take the form of PL =l

1 bh3 PLi
QPU ¼ Uhb  ð9Þ
2 12g l

Equation (9) is very useful; it is employed to flow that is due to surface velocity and
pressure. Pressure is obtained from external source of energy while surface velocity is
 Intelligent Servo Feedback Control for Hydrostatic Journal Bearing 357

produced because of relative motion between two surfaces such as journal and bearing
surface.
Compressibility has good influence on bearing stiffness and damping. It depends
upon bulk modulus which is relationship between change in volume and change in
pressure, is described by relationship;

Ve
dV ¼ dPLi ð10Þ
be

Rearranging Eq. (10) while taking changes with respect to time

dV Ve dPLi
¼ ð11Þ
dt be dt

Time rate of change of volume is called is flow for hydrostatic bearing termed as
compressible flow, so Eq. (11) takes form that is given by;

Ve dPLi
QCompressibility ¼ ð12Þ
be dt

One can get overall flow in hydrostatic bearing by employing equation of conti-
nuity [24, 25]. Where flow from servo valve is employed in term of squeezing flow,
pressure velocity flow and compressibility flow. Motion dynamics of hydrostatic
bearing can be described by considering factors such as; arrangement of masses, fluid
film stiffness and structure linkages. The equation of motion for hydrostatic bearing in
term of masses, stiffness and coefficients, is given by;

Ffilm ¼ m€hi þ Bd h_ i þ FLi ð13Þ

4 Design Strategy

In this paper two design strategies are proposed for HJBSC, one is PID control and
second is Intelligent Control.

4.1 PID Design Strategy

PID control is best controller which is used in a lot of engineering aplications due to
their ease of operations and good performance characteristics. PID consists of Pro-
portional (reduces rise time, make system fatser), Derivative (smooth signal over time)
and Integral (reduce staedy state error part of signal). PID controller takes error signal
and produces corresonding controller signal for plant. Error signal is genereted with
differential amplifier which takes output feedback signal and reference bearing clear-
ance for comparison. The PID controller drives servo valve by generating appropriate
signal. A simulink program is written in the form of algorithim to find efficient and
358 W. U. Rehman et al.

optimal parameters. A fittness function is used to find precise and accurate values of
controller parameters. This fitness function finds sum of errors between refrence and
current value and it is given by;

F ðei Þ ¼ sumðabsðei  e0 ÞÞ ð14Þ

4.2 Proposed Design Staretgy

In order to improve further effeciency of active hydrostatic journal bearing, a self
tuning PID control method is introduced in this part.
Self tuning PID controller is best option to compensate error caused by external
laoding and inaccuracies. Conventional PID controller are used in a lot of industrial
process and automatic control system due to their ability of simpilcity and ease of
operation. Fuzzy control is a good method for systems such as complex systems, time-
delay linear systems and higher-order nonlinear systems that have no accurate and
precise mathematical models. Keeping in view the problems to get an accurate
mathematical model of the proposed system and to achieve good load rejection per-
formance and good stiffness. A self tuning PID controller is designed in this paper. The
structure of self tuning PID controller or Intelligent controller is shown in Fig. 2. It is
just like a parallel structure where fuzzy logic is used for tuning the parameters of a PID
controller. In Fig. 2, ep is the error between reference bearing clearance and output
bearing clearance while Dp1 and Dp2 are gain that covert signal into such a form which
is suitable for fuzzy logic controller, while controller parameters are D1 , D2 , C1 , C2 ,
G1 , G2 and tuning parameters are Kp0 and Ki0 , Kd0 . The fuzzy logic controller is a parallel
structure where fuzzy logic is used to tune the parameter of PID controller. The “rule
base” keeps the knowledge in the form of rules and each rule present a certain state
called member function. The triangular member function is used which has general
formula as:

Fig. 2. Structure of self-tuning PID controller

 Intelligent Servo Feedback Control for Hydrostatic Journal Bearing 359

8
>
> 0 x  ar
<
ðx  ar Þ=ðbr  ar Þ ar \x\br
lr ¼ ð15Þ
>
> ðc  xÞ=ðcr  br Þ br \x\cr
: r
0 cr  x

Where r 2 fNB, NS, ZE, PS, PB, S, MS, M, MB, Bg.

Let suppose the range of parameters Kp , Kd and Ki of the PID controller are
bounded. A suitable range of each parameter is found by the simulation of the con-
vention PID controller. The range of each parameter for the fuzzy logic controller is
Kp 2 ½Kpmin ; Kpmax , Kd 2 ½Kdmin ; Kdmax  and Ki 2 ½Kimin ; Kimax . They can be calibrated
by using an adjustment mechanism;

Kz ¼ ðKzmax  Kzmin ÞKz0 þ Kzmin ð16Þ

Where subscript z presents the type of the tuning parameter. It may be derivative,
proportional or integral.

5 Simulation Results and Discussion

In order to check effectiveness of proposed strategy experiments have been done in

matlab/Simulink by using parameters which are given in Table 1. Simulations are
performed under different dynamics conditions which are explained in next part.

5.1 Influence of Initial Oil Pressure

Whenever we use rotational machinery then high speed and high load is always
demand in operationg conditions. Such applications where high speed and high load is
common operationg conditions, demand hydrostatic jornal bearing. The pressure of oil
is one of the most important component that has influence on the working of active
hydrostatic journal bearing. So its effects must be analyzed. A number of experiments
have been done in matlab/simulink and it is observed that the performance of self-
tuning PId controller is better than coventional PID controller under same initial oil
pressure as shown in Fig. 3.

5.2 Influence of External Load

The shaft cannot get its equilibrium position after being applied load. There are some
reasons behind this phenomena such as; throttling effect, bearing component’s mass
and stiffness as well as inertia of hydrostatic bearing.
There are two important factors such eccentricity and response time to get new
equilibrium position. These factors are very important for performance of active
hydrostatic journal bearing. Active hydrostatic journal has servo valve which adjust
flow in real time so that hydrostatic journal bearing can get zero eccentricity under
equilibrium conditions. In order to improve performance of active hydrostatic journal
bearing, two strategies are proposed one is conventional PID control and second is self-
360 W. U. Rehman et al.

Table 1. Simulation parameters

Parameter Symbol Values
Recess h0 ðmÞ 2:5  105
l ðm Þ 0:08
l 1 ðm Þ 0:01
bðmÞ 0:08
b1 ðmÞ 0:01
mðkgÞ 23
Shaft (Spindle) d ðmÞ 0:06
N ðrpmÞ 1500
Oil Be ðPaÞ 7  108
qðKg=m3 900
Bd ðNs=mÞ 5  105
gðPasÞ 0:025
Bearing LðmÞ 0:06
L1 ðmÞ 0:012
DðmÞ 0:025
B1 ðmÞ pD=16
C ðm Þ 2:5  105
Ae ðm2 Þ 1:88  102
Ve ðm3 Þ 2:8  104
F ðN Þ 1200
DðmÞ 0:06006
Servo valve Ki ðm=AÞ 3:04  105
Kq ðm2 =sÞ 2:7
Kc ðm3 s1 PaÞ 1:75  1011
sv ðsÞ 0:001

Performance of PID & self Tuning PID with respect to Initial pressure
25.05

25

24.95
PID
Bearing Clearance (um)

24.90 Self-Tuning PID

24.85 P0=3MPa
P0=3MPa
24.80

24.75

24.70

24.65

24.60

24.55
1.05 1.1 1.15 1.2 1.25 1.3 1.35
Time (seconds)

Fig. 3. Influence of initial oil pressure on performance of bearing

 Intelligent Servo Feedback Control for Hydrostatic Journal Bearing 361

tuning PID control. The external load of 800 N and 1600 N is applied at a time 1.1 s as
shown in Fig. 4. Result shows that under same external load, the stiffness and load
rejection performance of self-tuning PID control based active hydrostatic journal is
better than PID control based active hydrostatic journal bearing. This thing provides the
system good rotation and stationary characteristics as well provides rotor or shaft more
stability.

Performance of PID & self Tuning PID with respect to external loading
25.1

25
Bearing Clearance (um)

24.9

24.8

24.7

PID
24.6 Self-Tuning PID
1600N
800N
24.5
1600N
800N
24.4
1.09 1.1 1.11 1.12 1.13 1.14 1.15 1.16 1.17 1.18 1.19 1.20
Time (seconds)

Fig. 4. Influence of external load on performance of active hydrostatic journal bearing

5.3 Influence of Oil Viscosity

During working, there is some separation between journal and surface and this sepa-
ration kept constant by a pressurized oil film. This separation help hydrostatic journal
bearing to exhibit good stiffness and better rejection performance as well as low friction
and good damping characteristics. Investigation reveals that temperature has important
effect on viscosity of oil. Operations of high speed and high load produce a lot heat
which has effect to change viscosity of oil. So oil under temperature condition will
affect the efficiency of hydrostatic journal bearing. This problem has been removed in
active hydrostatic journal bearing; temperature does not affect much more due to servo
feedback. Furthermore, Self-tuning PID control is less sensitive to temperature changes
as compared to conventional PID. It has also better stiffness under same viscosity
conditions as compared to conventional PID as shown in Fig. 5.

5.4 Influence of Spindle Speed

Whenever there is eccentricity then offset distance come to play which produce
squeezing effect and squeezing effect produce film oscillation. Conventional hydro-
static journal bearing produces eccentricity even once system settled down and gets
new equilibrium position while hydrostatic journal bearing with proposed strategy has
no eccentricity under equilibrium position and also Self tuning PID controller has better
362 W. U. Rehman et al.

Performance of PID & self Tuning PID with respect to viscosity

25.50

25

24.95
Bearing Clearance (um)

24.90
Self-Tuning PID
24.85 PID

24.80 η=0.015

2.475

24.70

24.65

24.60

24.55
1.05 1.1 1.15 1.2 1.25 1.3
Time (seconds)

Fig. 5. Influence of viscosity on performance of active hydrostatic journal bearing

results under different speed as compared to PID controller as shown in Fig. 6. Servo
valve is key component to remove squeezing effect in active hydrostatic journal
bearing.

Performance of PID & self Tuning PID with respect to speed

25.50

25

24.95
Bearing Clearance (um)

24.90
Self-Tuning PID
24.85 PID

24.80 RPM=1500

2.475

24.70

24.65

24.60

24.55
1.05 1.1 1.15 1.2 1.25 1.3
Time (seconds)

Fig. 6. Influence of external load on performance of active hydrostatic journal bearing

6 Conclusions

The current work presents an accurate and precise mathematical model that consists of
active hydrostatic journal bearing and servo feedback control. The proposed mathe-
matical model is checked under two different strategies. Experiments are carried out in
Matlab/Simulink. The problem of traditional hydrostatic journal bearing is eccentricity
under different loads which is removed by proposed active hydrostatic journal bearing.
Furthermore, Proposed active hydrostatic journal bearing is checked under two
 Intelligent Servo Feedback Control for Hydrostatic Journal Bearing 363

different strategies such PID controller and intelligent Controller. Results show that
active hydrostatic journal bearing with intelligent Controller has better stiffness, load
rejection, rotor stability, less vibration and no wear under different dynamic conditions
of rotating velocity, temperature, initial oil pressure and varying load. Proposed active
hydrostatic journal bearing monitors bearing clearance in real time and adjust oil film
thickness in such a way so that eccentricity is zero under new equilibrium after being
applied load. Current work involves simulation work while future work will be to
validate proposed system on test bench.

Acknowledgement. This work is supported by the National Natural Science Foundation of

China (Grant No. 51075409).

References
1. Cha, M., Kuznetsov, E., Glavatskih, S.: A comparative linear and nonlinear dynamic
analysis of compliant cylindrical journal bearings. Mech. Mach. Theory 64, 80–92 (2013)
2. Kim, H., Jang, G., Ha, H.: A generalized Reynolds equation and its perturbation equations
for fluid dynamic bearings with curved surfaces. Tribol. Int. 50, 6–15 (2012)
3. Thomsen, K., Klit, P.: A study on compliant layers and its influence on dynamic response of
a hydrodynamic journal bearing. Tribol. Int. 44, 1872–1877 (2011)
4. De Pellegrin, D.V., Hargreaves, D.J.: An isoviscous, isothermal model investigating the
influence of hydrostatic recesses on a spring-supported tilting pad thrust bearing. Tribol. Int.
51, 25–35 (2012)
5. Phalle, V.M., Sharma, S.C., Jain, S.: Influence of wear on the performance of a 2-lobe
multirecess hybrid journal bearing system compensated with membrane restrictor. Tribol.
Int. 44, 380–395 (2011)
6. Bouaziz, S., Hili, M.A., Mataar, M., Fakhfakh, T., Haddar, M.: Dynamic behaviour of
hydrodynamic journal bearings in presence of rotor spatial angular misalignment. Mech.
Mach. Theory 44, 1548–1559 (2009)
7. Sharma, S.C., Phalle, V.M., Jain, S.: Influence of wear on the performance of a multirecess
conical hybrid journal bearing compensated with orifice restrictor. Tribol. Int. 44, 1754–
1764 (2011)
8. Rehman, W.U., Yuanxin, L., Guiyun, J., Yongqin, W., Rehman, S.U., Bibi, S., et al.:
Control of oil film thickness for hydrostatic journal bearing using PID disturbance rejection
controller. In: 2017 IEEE 3rd Information Technology and Mechatronics Engineering
Conference (ITOEC), pp. 543–547 (2017)
9. Rehman, W.U., Yuanxin, L., Guiyun, J., Yongqin, W., Yun, X., Iqbal, M.N., et al.: Control
of an oil film thickness in a hydrostatic journal bearing under different dynamic conditions.
In: 2017 29th Chinese Control And Decision Conference (CCDC), pp. 5072–5076 (2017)
10. Nicoletti, R., Santos, I.F.: Control system design for flexible rotors supported by actively
lubricated bearings. J. Vib. Control 14, 347–374 (2008)
11. Morosi, S., Santos, I.F.: Active lubrication applied to radial gas journal bearings. Part 1:
modeling. Tribol. Int. 44, 1949–1958 (2011)
12. Santos, I.F., Watanabe, F.Y.: Compensation of cross-coupling stiffness and increase of direct
damping in multirecess journal bearings using active hybrid lubrication: part I—theory.
J. Tribol. 126, 146–155 (2004)
364 W. U. Rehman et al.

13. Kim, K.-J., Lee, C.-W.: Dynamic characteristics of sealed squeeze film damper with a central
feeding groove. J. Tribol. 127, 103–111 (2005)
14. Ho, Y., Liu, H., Yu, L.: Effect of thrust magnetic bearing on stability and bifurcation of a
flexible rotor active magnetic bearing system. Trans. Am. Soc. Mech. Eng. J. Vib. Acoust.
125, 307–316 (2003)
15. Nicoletti, R., Santos, I.F.: Frequency response analysis of an actively lubricated rotor/tilting-
pad bearing system. In: ASME Turbo Expo 2004: Power for Land, Sea, and Air, pp. 735–
744 (2004)
16. Estupinan, E.A., Santos, I.F.: Linking rigid multibody systems via controllable thin fluid
films. Tribol. Int. 42, 1478–1486 (2009)
17. Haugaard, A.M., Santos, I.F.: Multi-orifice active tilting-pad journal bearings—harnessing
of synergetic coupling effects. Tribol. Int. 43, 1374–1391 (2010)
18. Merritt, H.E.: Hydraulic Control Systems. Wiley, Hoboken (1967)
19. Edelmann, H.: Schnelle Proportionalventile und ihre Anwendug, Sonderdruck aus Ölhy-
draulic und Pneumatik. Schnelle Proportionalventile und ihre Anwendug, Sonderdruck aus
Ölhydraulic und Pneumatik 30, 1 (1986)
20. Neal, T.: Performance estimation for electrohydraulic control systems. Moog Tech. Bull. 126
(1974)
21. Rehman, W.U., Wang, S., Wang, X., Shi, C., Zhang, C., Tomovic, M.: Adaptive control for
motion synchronization of HA/EHA system by using modified MIT rule. In: 2016 IEEE 11th
Conference on Industrial Electronics and Applications (ICIEA), pp. 2196–2201 (2016)
22. Rehman, W.U., Wang, S., Wang, X., Fan, L., Shah, K.A.: Motion synchronization in a dual
redundant HA/EHA system by using a hybrid integrated intelligent control design. Chin.
J. Aeronaut. 29, 789–798 (2016)
23. Rehman, W.U., Jiang, G., Luo, Y., Wang, Y., Khan, W., Rehman, S.U., et al.: Control of
active lubrication for hydrostatic journal bearing by monitoring bearing clearance. Adv.
Mech. Eng. 10, 1687814018768142 (2018)
24. Rehman, W.U., Wang, X., Wang, S., Azhar, I.: Motion synchronization of HA/EHA system
for a large civil aircraft by using adaptive control. In: 2016 IEEE Chinese Guidance,
Navigation and Control Conference (CGNCC), pp. 1486–1491 (2016)
25. Rehman, W.U., Nawaz, H., Wang, S., Wang, X., Luo, Y., Yun, X., et al.: Trajectory based
motion synchronization in a dissimilar redundant actuation system for a large civil aircraft.
In: 2017 29th Chinese Control And Decision Conference (CCDC), pp. 5010–5015 (2017)
 Iterative Feedback Tuning
for Two-Degree-of-Freedom System

Hui Pan1(&), Yanjin Zhang1 , and Ling Wang2

1
College of Automation Engineering, Shanghai University of Electric Power,
Shanghai 200090, China
[email protected]
2
Shanghai Key Laboratory of Power Station Automation Technology,
Shanghai University, Shanghai 200072, China

Abstract. This paper is concerned with Iterative Feedback Tuning (IFT) for
Two-degree-of- freedom (2-DOF) system. The IFT is a data-driven method for
tuning controller parameters, which uses the closed-loop system input-output
data directly, and without establishing a mathematical model for the controlled
system. The design of control system is a multi-objective problem, so a 2-DOF
control system naturally has advantages over a one-degree-of- freedom (1-DOF)
control system. When tuning 2-DOF system controllers, two-step method is
firstly concerned. While in this paper, the application of IFT method in a 2-DOF
control system is studied, which can make controllers’ parameters tuned at the
same time, and the IFT method is more accurate and efficient in tracking per-
formance and robustness. The feasibility and effectiveness of the method are
verified by numerical simulation and comparison.

Keywords: IFT
2-DOF
Data-driven
PID
Parameter tuning

1 Introduction

In this paper, the problem of tuning Two-degree-of-freedom (2-DOF) system con-

trollers is concerned, when no mathematical description of the plant dynamical
behavior is available. The degree of freedom of the control system can be defined as the
number of independently adjustable closed-loop transfer functions [1, 2]. When
designing a control system for a multi-objective optimization problem, 2-DOF control
has more advantages than one-degree-of-freedom (1-DOF), for example, system
tracking performance and robustness. This fact was already stated by the article [3]. In
[4, 5], various 2-DOF PID controllers were proposed for industrial use and detailed
analyses were made, also the further study were made about optimal tuning [6–9]. The
main methods are Two-step Tuning Method (TSTM) [3], Frequency-domain Method
(FDM) [9], Virtual Reference Feedback Tuning (VRFT) [10], and so on.
It has been more than twenty years since the Iterative Feedback Tuning (IFT) was
put forward by Hjalmarsson [11]. As one of the methods that belongs to data-driven
control (DDC) [12], IFT has attracted many scholars’ attention since its emergence.
After years of development, IFT has gradually become a well-established design
methodology [13, 14], and which has been applied into many control systems that

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 365–379, 2018.
https://doi.org/10.1007/978-981-13-2384-3_34
366 H. Pan et al.

input-output data is available. Similar to other DDC methods [15–19], IFT does not
need to establish the mathematical model of the controlled system, and only uses the
input-output data of the closed-loop experiments of the system to tuning the controller
parameters directly. IFT theory was originally proposed for single-loop control sys-
tems, and with the development, it has been studied and extended to many advanced
control systems, such as multi-input-multi-output systems [20], internal model control
[21], cascade control [22] and so on. IFT method has been extended to various
industrial applications [23–26].
In this paper, the IFT method is studied to tuning the parameters of the PID
controllers in the 2-DOF control system. In IFT method, at each iteration, the exper-
imental input-output data of the system are utilized to tuning the parameters of the 2-
DOF PID controllers’ parameters directly. The feasibility and effectiveness of the
method are proved by the numerical simulation and comparison of the typical 2-DOF
control system.
The paper is organized as follows. Section 2 illustrates the basic theory of 2-DOF
system and IFT method. In Sect. 3, the IFT theory is extend to the 2-DOF system,
which contains the derivation of correlation variables. Section 4 gives two numerical
examples of the given method, and Sect. 5 is the summary of this paper and some
future work.

2 Preliminaries

2.1 2-DOF Control System

A general form of the 2-DOF control system is shown in Fig. 1, where the controller
consists of two conventional controllers: CðsÞ and Cf ðsÞ. C ðsÞ is called the serial (or
main) controller and Cf ðsÞ is called the feed-forward controller. The transfer function
Pd ðsÞ from the disturbance d to controlled variable y is assumed different from the
transfer function PðsÞ from the manipulated variable u to y. Variable r is the set value
of the system, and H ðsÞ is a detector. The closed-loop transfer functions from r to y and
d to y are respectively given by:


PðsÞ CðsÞ þ Cf ðsÞ
Gr ðsÞ ¼ ð1Þ
1 þ PðsÞCðsÞHðsÞ

Pd ðsÞ
Gd ðsÞ ¼ ð2Þ
1 þ PðsÞCðsÞHðsÞ

To simplify the problem, the following two assumptions that are appropriate for
many practical design problems with some exceptions, are introduced for the control
system shown in Fig. 1.
 Iterative Feedback Tuning for Two-Degree-of-Freedom System 367

contoller plant
disturbance
Cf (s) d Pd (s)
controlled
set point manipulated
variable + variable + variable
r + + u + y
C(s) P(s)
-

H(s)
detector

Fig. 1. Two-degree-of-freedom (2-DOF) control system.

Assumption 1: The detector has sufficient accuracy and speed for the given control
purpose, i.e.,

HðsÞ ¼ 1 ð3Þ

Assumption 2: The main disturbance enters at manipulating point, i.e.,

Pd ðsÞ ¼ PðsÞ ð4Þ

As a result, the control system shown in Fig. 1 can be simplified to the feed-
forward type of 2-DOF control system shown in Fig. 2 under assumptions 1 and 2 [3].

Cf (s) d
+ +
r + + u+ y
C(s) P(s)
-

Fig. 2. Feed-forward type of 2-DOF PID control system.

The controller of the Fig. 2 system consists of C ðsÞ and Cf ðsÞ, where the C ðsÞ is a
conventional PID controller and Cf ðsÞ is a feed-forward compensation PID controller.
 
1
CðsÞ ¼ Kp 1 þ þ Td DðsÞ ð5Þ
Ti s

Cf ðsÞ ¼ Kp ða þ bTd DðsÞÞ ð6Þ

Where the Kp is the proportional gain, Ti is the integral time and Td is the differ-
ential time. a and b are 2-DOF parameters, and the above five parameters can be
adjusted. DðsÞ can be expressed by formula (7), where s ¼ Td =d, and d is a differential
gain which is a fixed value and does not affect the optimization of other parameters, but
the characteristics of the actual controlled object need to be considered.
s
DðsÞ ¼ ð7Þ
1 þ ss
368 H. Pan et al.

In the following discussion, the feed-forward 2-DOF control system shown in

Fig. 2 will be considered. According to the Fig. 2, the closed-loop transfer function
y2 ðtÞ (superscript 2 represents 2-DOF) can be expressed by formula (8).


CðsÞ þ Cf ðsÞ rðtÞ þ dðtÞ
y ðtÞ ¼
2
PðsÞ ð8Þ
1 þ PðsÞCðsÞ

According to formula (8), the output of the closed-loop system can be divided into
two parts: one is the output caused by the set value r ðtÞ, which is recorded as y2r ðtÞ; and
the other is the output caused by the disturbance d ðtÞ, which is recorded as y2d ðtÞ.

CðsÞ þ Cf ðsÞ
y2r ðtÞ ¼ PðsÞrðtÞ ð9Þ
1 þ PðsÞCðsÞ

1
y2d ðtÞ ¼ PðsÞdðtÞ ð10Þ
1 þ PðsÞCðsÞ

2.2 Iterative Feedback Tuning Theory

IFT is one of the DDC methods, which puts forward a performance criterion function
(for example, LQG type) for the control system. Based on the input-output data of the
closed-loop control system, to find the optimal controller parameters that is based on
gradient iteration. IFT does not need to model the plant and directly tunes the controller
parameters, which makes the controller parameters tuning more direct and accurate
[14].
As is shown in Fig. 3, the conventional 1-DOF closed-loop control system, the
output of the Closed-loop system y1 ðtÞ (superscript 1 represents 1-DOF) can be
expressed by formula (11) and (12).

d
+
r+ u+ y
C(s,θ ) P(s)
-

Fig. 3. Conventional closed-loop PID control system.

Where r ðtÞ is the system set value input; d ðtÞ is the disturbance; PðsÞ is the
unknown plant; uðtÞ is the control signal. The controller structure of C ðs; hÞ has been
selected beforehand, the controller parameter vector h is adjustable, and the parameter
vector h is differentiable. h ¼ ½h1


hn , n is the number of the parameters. In practice,
the controller C ðs; hÞ is usually chosen as the conventional PID controller, then n ¼ 3.
 Iterative Feedback Tuning for Two-Degree-of-Freedom System 369

y1 ðtÞ ¼ ðuðtÞ þ dðtÞÞPðsÞ ð11Þ


uðtÞ ¼ Cðs; hÞ rðtÞ  y1 ðtÞ ð12Þ

Where r ðtÞ is the system set value input; d ðtÞ is the disturbance; PðsÞ is the
unknown plant; uðtÞ is the control signal. The controller structure of C ðs; hÞ has been
selected beforehand, the controller parameter vector h is adjustable, and the parameter
vector h is differentiable. h ¼ ½h1


hn , n is the number of the parameters. In practice,
the controller C ðs; hÞ is usually chosen as the conventional PID controller, then n ¼ 3.
The expected closed-loop response of the system from set value input r ðtÞ to output
y1d ðtÞ, which is expressed by formula (13). Td1 ðsÞ is the expected closed-loop system
transfer function.

y1d ðtÞ ¼ Td1 ðsÞrðtÞ ð13Þ

The tracking error between the actual output y1 ðtÞ and the expected output y1d ðtÞ is
~y ðtÞ.
1

~y1 ðtÞ ¼ y1 ðtÞ  y1d ðtÞ

PðsÞCðs; hÞ PðsÞ ð14Þ
¼ rðtÞ  Td1 ðsÞrðtÞ þ dðtÞ
1 þ PðsÞCðs; hÞ 1 þ PðsÞCðs; hÞ

The design goal of the controller C ðs; hÞ is to minimize the tracking error ~y1 ðtÞ
between the y1 ðtÞ and the y1d ðtÞ, so that the system performance criterion function J ðhÞ
can be chosen as quadratic criterion.

N h
2 i
1 X
JðhÞ ¼ y1 ðk; hÞ  y1d ðkÞ ð15Þ
2N k¼1

h ¼ arg min JðhÞ ð16Þ

h

When the sampling data length N is fixed, the value of J ðhÞ is only related to the
controller parameters vector h. The goal of the IFT method is to find the h when the
performance function J ðhÞ gets the minimum value, and define it as the optimal
controller parameter. To find the optimal h can be regarded as the problem of finding
the minimum value of the quadratic function J ðhÞ.
To obtain the minimum of J ðhÞ, a solution for h to the Eq. (17) is found:

N 

@JðhÞ 1 X  @y1 ðk; hÞ
¼ y ðk; hÞ  yd ðkÞ
1 1
¼0 ð17Þ
@h N k¼1 @h

If the gradient @J ðhÞ=@h could be computed, then the solution of (17) would be
obtained iteratively by the following Algorithms:
370 H. Pan et al.

@Jðhi Þ
hi þ 1 ¼ hi  ci R1
i ð18Þ
@h
Here, Ri is some appropriate positive definite matrix at iteration i, typically a Gauss-
Newton approximation of the Hessian of J ðhÞ, while ci is a positive real scalar that
determines the step size. The sequence ci , must obey some constraints for the algorithm
to converge to a local minimum of the performance criterion function J ðhÞ.
As stated, this problem is intractable since it involves expectations that are
unknown. However, such problem can be solved by using a stochastic approximation
algorithm of the form (18). The gradient @J ðhÞ=@h evaluated at the current controller
can be replaced by an unbiased estimate. In order to solve this problem, one thus needs
to generate the following quantities:
(i) the signals ~y1 ðtÞ and r ðtÞ;
(ii) the gradients @~y1 ðt; hÞ=@h and @y1 ðt; hÞ=@h.
Because the PðsÞ is unknown, the input-output data of the system are required
which are obtained by closed-loop system experiments.

3 Iterative Feedback Tuning for 2-DOF Controllers

The IFT for 1-DOF system and the typical structure of 2-DOF are given in previous
section. In this section, the IFT is extended for the 2-DOF case shown in Fig. 2.

3.1 Derivation of Correlation Variables

For the 2-DOF control system shown in Fig. 2, the system output is expressed by
formula (8). To facilitate the following derivation, rewrite the conventional

 controller
C ðsÞ to Cðhc Þ, and rewrite the feed-forward controller Cf ðsÞ to Cf hf . Notice that,


C ðhc Þ and Cf hf respectively relate to the adjustable PID controller of parameters hc
and hf . Now that:
2 3 2 3
hc1 Kp    
hf 1 ahc1
hc ¼ 4 hc2 5 ¼ 4 KTp 5; hf ¼ ¼ ð19Þ
i hf 2 bhc3
hc3 Kp Td

T
Where h ¼ ½ hc hf  . The error ~y2 ðtÞ between actual output and expected output
of the Fig. 3 system can be expressed by formula (20).

~y2 ðtÞ ¼ y2 ðtÞ  y2d ðtÞ


CðhÞ þ Cf ðhf Þ rðtÞ þ dðtÞ ð20Þ
¼ PðsÞ  Td2 ðsÞrðtÞ
1 þ PðsÞCðhÞ

Similar to 1-DOF system, a performance criterion function of quadratic form like

formula (15) is selected.
 Iterative Feedback Tuning for Two-Degree-of-Freedom System 371

N h
2 i
1 X
JðhÞ ¼ y2 ðk; hÞ  y2d ðkÞ ð21Þ
2N k¼1

The gradient @J ðhÞ=@h can be obtained from the derivation of h on both sides of
the Eq. (21).

N  
@JðhÞ 1 X @~y2 ðk; hÞ
¼ ~y2 ðk; hÞ ð22Þ
@h N k¼1 @h

In order to achieve iterative optimization of controller parameters vector h, cal-

culation can be carried out according to formula (17). However, when using formula
(17), it is necessary to figure out the gradient @J ðhÞ=@h by using formula (22), but in
fact, the value of @~y2 ðt; hÞ=@h must be obtained, which greatly increases the difficulty
of the solution process. For this purpose, it is necessary to use the input-output data of
the system to solve @~y2 ðt; hÞ=@h by stochastic approximation, and it means that to
replace @~y2 ðt; hÞ=@h after the calculation and processing of the sampling data currently.
As a result, the error signals ~y2 ðt; hi Þ of the system must be obtained, which is used for
the gradient @~y2 ðt; hÞ=@h. When the controller parameters of the system is hi , the
gradient of the error signal @~y2 ðt; hÞ=@h can be obtained by the following formula.


@~y2 ðt; hi Þ @ y2 ðt; hi Þ  y2d ðtÞ @y2 ðt; hi Þ
¼ ¼ ð23Þ
@h @h @h

Obviously, by calculating the gradient @y2 ðt; hi Þ=@h to estimate @~y2 ðt; hÞ=@h. For
the following calculation, the controller parameters hc and hf are differentiated
respectively.

@y2 ðt; hi Þ @y2r ðt; hi Þ @y2d ðt; hi Þ

¼ þ
@hc @hc @hc
@Cðhc Þ PðsÞ

¼ rðtÞ  y2r ðt; hi Þ ð24Þ
@hc 1 þ Cðhc ÞPðsÞ
@Cðhc Þ PðsÞ
 y2d ðt; hi Þ
@hc 1 þ PðsÞCðsÞ

@y2 ðt; hi Þ @y2r ðt; hi Þ @y2d ðt; hi Þ

¼ þ
@hf @hf @hf
ð25Þ
@Cðhf Þ PðsÞ
¼ rðtÞ
@hf 1 þ Cðhc ÞPðsÞ

Notice that the model of the PðsÞ is unknown and there is no need to model it.
@y2 ðt; hi Þ=@hc and @y2 ðt; hi Þ=@hf are obtained by using the input-output data from three
experiments in IFT method.
372 H. Pan et al.

3.2 Closed - Loop System Experiment

Through the derivation of Sect. 3.1, the key to the iterative optimization of controller
parameter vector h is to get the values of @y2 ðt,hi Þ=@hc and @y2 ðt; hi Þ=@hf , which can
be used for the estimation of @~y2 ðt; hi Þ=@h and @J ðhi Þ=@h. To this end, three closed-
loop system tests are required to obtain relevant data.
(i) the first is called basic experiment. With r1 ¼ r ðtÞ as the input of the set value of
the system, the system output y21 ðt; hi Þ (subscript 1 represents the first experiment) with
a data length of N is obtained.

CðsÞ þ Cf ðsÞ 1
y21 ðt; hi Þ ¼ rðtÞPðsÞ þ dðtÞPðsÞ ð26Þ
1 þ PðsÞCðsÞ 1 þ PðsÞCðsÞ

Cf (s) r − y12
d
+ +
r=0 + + u+ y22 ( ŷ )′
2

C(s) C′(s)
2
P(s)
-

Fig. 4. Schematic diagram of the second experiment.

(ii) the second experiment is described in Fig. 4, set r2 ¼ 0 and r ðtÞ  y21 ðt; hi Þ as
the input of the set value of the system. The output y22 ðt; hi Þ with data length N should
be collected. The gradient @~y2 ðt; hi Þ=@hf can be calculated by formula (28).

PðsÞ 

y22 ðt; hi Þ ¼ rðtÞ  y21 ðt; hi Þ þ dðtÞ ð27Þ
1 þ PðsÞCðhc Þ

@~y22 ðt; hi Þ @Cðhc Þ 2

¼ y ðt; hi Þ
@hc @hc 2
@Cðhc Þ PðsÞ 

¼ rðtÞ  y21 ðt; hi Þ þ dðtÞ ð28Þ
@hc 1 þ PðsÞCðhc Þ
@y2 ðt; hi Þ
¼ þ wðt; hi Þ
@hc

Where

@Cðhc Þ PðsÞ

wðt; hi Þ ¼ dðtÞ þ y2d
1 ðt; hi Þ ð29Þ
@hc 1 þ PðsÞCðhc Þ

(iii) the third experiment is described in Fig. 5, to set r3 ¼ 0 and r ðtÞ as the input of
the set value of the system. Similarly, to get the output y23 ðt; hi Þ with data length N. The
gradient @~y2 ðt; hi Þ=@hf can be calculated by formula (30).
 Iterative Feedback Tuning for Two-Degree-of-Freedom System 373

Cf (s) r
d
+ + y32 ( yˆ32 )′
r=0 + C(s) + u+
P(s) C′f (s)
-

Fig. 5. Schematic diagram of the third experiment.

PðsÞ
y23 ðt; hi Þ ¼ ðrðtÞ þ dðtÞÞ ð30Þ
1 þ PðsÞCðhc Þ

@~y23 ðt; hi Þ @Cðhf Þ 2

¼ y ðt; hi Þ
@hf @hf 3
@Cðhf Þ PðsÞ
¼ ðrðtÞ þ dðtÞÞ ð31Þ
@hf 1 þ PðsÞCðhc Þ
@y2 ðt; hi Þ
¼ þ vðt; hi Þ
@hf

Where

@Cðhf Þ PðsÞ
vðt; hi Þ ¼ dðtÞ ð32Þ
@hf 1 þ PðsÞCðhc Þ

Through three experiment, the related input-output data are obtained, and can be
calculate for @~y2 ðt; hi Þ=@hc and @~y2 ðt; hi Þ=@hf .
When there is no noise or mean value of noise in the system is zero, or in another
word, the mean values of Eqs. (30) and (33) are zero, @y2 ðt; hi Þ=@hc and @y2 ðt; hi Þ=@hf
are unbiased estimates of @~y2 ðt; hi Þ=@hc and @~y2 ðt; hi Þ=@hf [14].
Based on the above analysis, the gradient @J ðhi Þ=@h can be rewritten as follows:
8 N 

>
> @Jðhi Þ 1 X  @~y2 ðk; hi Þ
>
> ¼ y ðk; hi Þ  yd ðkÞ
2 2
< @hc N k¼1 @hc
  ð33Þ
>
> @Jðhi Þ 1 X
2
N  @~y ðk; hi Þ
2
>
> ¼ y ðk; hi Þ  yd ðkÞ
2
:
@hf N k¼1 @hf

The positive definite matrix Ri can be obtained by using the Hessian estimation of
J ðhi Þ.
8 " T  2 #
>
> 1 X N
@~y 2
ðt; h Þ @~y ðt; h Þ
>
> R ¼
i i
>
< ci N @hc @hc
k¼1
"   # ð34Þ
>
> 1X N
@~y2 ðt; hi Þ
T
@~y2 ðt; hi Þ
>
>
: Rfi ¼ N
>
@hf @hf
k¼1
374 H. Pan et al.

On the basis of formula (18), the formula is obtained according to the derivation
and transformation of the above relevant variables. Finally, the iteration of controller
parameter vector h is realized according to formula (35), and the optimal value is finally
achieved.
8
>
> 1 @Jðhi Þ
< hci þ 1 ¼ hci  cci Rci @h
c
ð35Þ
>
> @Jðhi Þ
: hfi þ 1 ¼ hci  cfi R1
fi
@hf

4 Numerical Example

Based on the above theoretical derivation, this section extends the application of IFT in
2-DOF system by numerical simulation, to compare the feasibility and effectiveness of
this method. In the following discussion, two numerical examples are chosen for the
simulation, which come from reference [3].

4.1 Numerical Example 1

For a typical plant model with first-order inertia plus hysteresis, when L=T ¼ 0:2 and
T ¼ 1. Selection of the expected Closed-Loop transfer function of the system is
Td2 ðsÞ ¼ 1, and the structure of PID controllers are shown in formula (37).
The sampling time of the system is Ts ¼ 0:01 s, s ¼ 0:001 s, runtime is 10 s.
According to the reference data in the literature, the initial controller parameters are as
follows: hc1 ¼ 5, hc2 ¼ 10, hc3 ¼ 0:25, hf 1 ¼ 3:5, hf 1 ¼ 0:45. Iterative data is shown
in Fig. 6.

1
P1 ðsÞ ¼ eLs ð36Þ
1 þ Ts
2 3
hc1  
  hf 1
hc ¼ 1 1 s 4 hc2 5; hf ¼  1 s
ð37Þ
s 1 þ ss 1 þ ss hf 2
hc3

From the data in Fig. 6, it can be seen that the value of the performance criterion
function J ðhÞ obtained from the previous iterations decreases at a relatively large rate,
and the range of parameter changes is relatively large. But with the iteration i increases,
the J ðhÞ decreases slowly and tends to be stable.
The above initial parameters and optimization parameters are brought into the
simulation experiment respectively, and the optimized controller parameters in the
literature are taken as reference. To set up r ðtÞ ¼ 1, with the starting time is 0 s and the
running time is 3.5 s. The simulation figure is shown in a diagram in Fig. 7. In the
following figures, Ini represents the initial parameters of the controller, IFT represents
the optimal parameters, and Ref represents the parameters in the reference [3].
 Iterative Feedback Tuning for Two-Degree-of-Freedom System 375

0.25

J(θ)
0.2

0 2 4 6 8 10 12
5.5

c1
θ
5
0 2 4 6 8 10 12
12

c2
11
θ
10
0 2 4 6 8 10 12
0.4
c3

0.3
θ

0.2
0 2 4 6 8 10 12
3.6

3.4
f1
θ

3.2
0 2 4 6 8 10 12

0.45
0.4
f2
θ

0.35

0 2 4 6 8 10 12
Iterative times (i)

Fig. 6. Parameters iterative graph of numerical example 1.

1.2 0.3
Ini
IFT
1 0.25
Ref

0.8 0.2

0.6 0.15
y

y
Output

Output

0.4 0.1

0.2 0.05

0 Ini 0
IFT
Ref
-0.2 -0.05
0 0.5 1 1.5 2 2.5 3 3.5 0 0.5 1 1.5 2 2.5 3 3.5
Time(s) (t/s) Time(s) (t/s)

(a) set point response (b) disturbance response

Fig. 7. Comparison of set point response and disturbance response in numerical examples 1.

Setting up r ðtÞ ¼ 0, and adding the disturbance d ðtÞ ¼ 1 to the system. The start
time is 0 s, the simulation figure is shown in b diagram in Fig. 7. It can be seen from
Fig. 7 that the comprehensive control effect of the IFT parameters is better than that of
the initial parameters and the reference parameters.

4.2 Numerical Example 2

For a typical plant model with second-order inertia plus hysteresis, when L=T ¼ 0:2
and T ¼ 1. Selection of the expected Closed-Loop transfer function of the system is
376 H. Pan et al.

Td2 ðsÞ ¼ 1, and the structure of PID controllers are shown in formula (38) in numerical
example 1.

1
P2 ðsÞ ¼ 2
eLs ð38Þ
ð1 þ TsÞ

According to the reference data in the literature, the initial controller parameters are
as follows: hc1 ¼ 10, hc2 ¼ 10:5, hc3 ¼ 2:2, hf 1 ¼ 8, hf 1 ¼ 1:7. The running condition
is the same as numerical example 1. Iterative data is shown in Fig. 8.

0.4
J(θ)

0.3

0.2
0 5 10 15 20 25 30
12

11
c1
θ

10
0 5 10 15 20 25 30
20

15
c2
θ

10
0 5 10 15 20 25 30
4

3
c3
θ

2
0 5 10 15 20 25 30
10

9
f1
θ

8
0 5 10 15 20 25 30
2.5

2
f2
θ

1.5
0 5 10 15 20 25 30
Iterative times (i)

Fig. 8. Parameters iterative graph of numerical example 2.

From the data in Fig. 8, it can be seen that the value of the performance criterion
function J ðhÞ obtained from the previous iterations decreases relatively small, and the
range of controller parameters is relatively stable.
The initial parameters and optimization parameters are brought into the simulation
experiment respectively, and the optimized controller parameters in the literature are
taken as reference. Setting up r ðtÞ ¼ 1 with the starting time is 0 s, and the running
time is 7 s. The simulation result is shown in a graph in Fig. 9.
Setting up r ðtÞ ¼ 0, and adding the disturbance d ðtÞ ¼ 1 to the system. The sim-
ulation result is shown in b graph in Fig. 9.
According to Fig. 9, to conclude that the comprehensive control effect of the IFT
parameters is better than that of the initial parameters and the reference parameters,
though the anti-disturbance is not as good as reference parameters.
 Iterative Feedback Tuning for Two-Degree-of-Freedom System 377

1.4 0.14
Ini Ini
IFT 0.12 IFT
1.2 Ref
Ref
0.1
1
0.08

0.8 0.06

Output (y)
Output (y)

0.04
0.6

0.02
0.4
0

0.2
-0.02

0 -0.04
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
Time(s) (t/s) Time(s) (t/s)

(a) set point response (b) disturbance response

Fig. 9. Comparison of set point response and disturbance response in numerical examples 2.

5 Summary

In this paper, we extend the IFT method to tuning the parameters of the PID controllers
in the 2-DOF structure control system, and the typical feed-forward 2-DOF control is
considered as the research object. Compared with the conventional TSTM, the 2-DOF
PID controllers can be tuned by using the input-output data of the closed-loop control
system without establishing a mathematical model, also two PID controllers’ param-
eters can be tuned at the same time, which makes the controller parameter tuning more
direct and accurate. The simulation results show that the tracking performance of the
system is better than that of the TSTM, and the anti-interference performance of the
system is slightly worse than that of the TSTM. However, the IFT method has less
oscillation and faster stability. It is feasible and effective to tune the parameters of 2-
DOF PID controllers by integrating the control effect of the whole system.
The following two points should be noted when applying the IFT method in tuning
the 2-DOF parameters:
(i) It is necessary to select an appropriate reference model, otherwise, the local
optimum is easy to appear in the IFT iteration, and the control effect of the
obtained parameters is not satisfactory.
(ii) It is necessary to select the appropriate initial value of the controller parameters,
so that the whole control system can be in a stable state, and all the controller
parameters can reach the optimum or near the optimum synchronously, which
requires some experience in the controlled system.
Currently, the IFT is still under study.

Acknowledgments. This work is supported by Shanghai Key Laboratory of Power Station

Automation Technology (No.13DZ2273800).
378 H. Pan et al.

References
1. Taguchi, H., Araki, M.: Two-degree-of-freedom PID controllers — their functions and
optimal tuning. J. IFAC Proc. Volumes 33(4), 91–96 (2000)
2. Horowitz, I.M.: Synthesis of Feedback Systems. Academic Press, Cambridge (1963)
3. Araki, M., Taguchi, H.: Two-degree-of-freedom PID controller. Int. J. Control Autom. Syst.
1(4), 18–25 (2003)
4. Yukitomo, M., Shigemasa, T., Baba, Y.: A two degrees of freedom PID control system, its
features and applications. In: Control Conference, IEEE, 2004, pp. 456–459 (2004)
5. Araki, M.: Two-degree-of-freedom control system – I. J. Syst. Control 29, 649–656 (1985)
6. Ye, S.U., Han, P., Wang, D.F.: Chaos optimization of two degree-of-freedom PID control.
J. Comput. Simul. 21(11), 164–166 (2004)
7. Taguchi, H.: Two-degree-of-freedom PID controllers - their functions and optimal tuning.
In: IFAC Digital Control: Past, Present and Future of PID Control, 5–7 April 2000, Terrassa,
Spain (2000)
8. Taguchi, H., Kokawa, M., Araki, M.: Optimal tuning of two-degree-of-freedom PD
controllers. In: Proceedings of the 4th Asian Control Conference, 25–27 September 2002,
Singapore (2002)
9. Taguchi, H., Doi, M., Araki, M.: Optimal parameters of two-degrees-of-freedom PID control
systems. J. Trans. Soc. Instrum. Control Eng. 23(9), 889–895 (2009)
10. Lecchini, A., Campi, M.C., Savaresi, S.M.: Virtual reference feedback tuning for two degree
of freedom controllers. Int. J. Adapt. Control Signal Process. 16(5), 355–371 (2002)
11. Hjalmarsson, H., Gunnarsson, S., Gevers, M.: A convergent iterative restricted complexity
control design scheme. In: 1994 Proceedings of the IEEE Conference on Decision and
Control, pp. 1735–1740 (1994)
12. Xu, J.X., Hou, Z.S.: Notes on data-driven system approaches. J. Acta Autom. Sinica 35(6),
668–675 (2009)
13. Hjalmarsson, H., Gevers, M., Gunnarsson, S.: Iterative feedback tuning: theory and
applications. J. IEEE Control Syst. 18(4), 26–41 (1998)
14. Hjalmarsson, H.: Iterative feedback tuning—an overview. J Int. J. Adapt. Control Signal
Process. 16(5), 373–395 (2002)
15. Arimoto, S., Kawamura, S., Miyazaki, F.: Bettering operation of dynamic systems by
learning: a new control theory for servomechanism or mechatronics systems. In: 2007 IEEE
Conference on Decision and Control, pp. 1064–1069 (1984)
16. Han, Z.G., Hou, Z.S.: Robust model-free learning adaptive control for nonlinear systems.
Control Decis. (2), 137–142 (1995)
17. Spall, J.C.: Multivariate stochastic approximation using a simultaneous perturbation gradient
approximation. IEEE Trans. Autom. Control 37(3), 332–341 (1992)
18. Campi, M.C., Lecchini, A., Savaresi, S.M.: Brief virtual reference feedback tuning: a direct
method for the design of feedback controllers. Automatica 38(8), 1337–1346 (2002)
19. Safonov, M.G., Tsao, T.C.: The unfalsified control concept: a direct path from experiment to
controller. In: Francis, B.A., Tannenbaum, A.R. (eds.) Feedback Control, Nonlinear
Systems, and Complexity. LNCIS, vol. 202, pp. 196–214. Springer, Heidelberg (1995).
https://doi.org/10.1007/BFb0027678
20. Nakamoto, M.: Parameter tuning of multiple PID controllers by using iterative feedback
tuning. In: 2004 SICE 2003 Conference, IEEE, pp. 183–186. IEEE (2003)
21. Bruyne, F.D.: Iterative feedback tuning for internal model controllers. Control Eng. Pract. 11
(9), 1043–1048 (2003)
 Iterative Feedback Tuning for Two-Degree-of-Freedom System 379

22. Tesch, D., Eckhard, D., Bazanella, A. S.: Iterative feedback tuning for cascade systems. In:
2017 Control Conference, pp. 495–500. IEEE (2017)
23. Ren, Q., Xu, J., Li, X.: A motion control approach for a robotic fish with iterative feedback
tuning. In: 2015 IEEE International Conference on Industrial Technology, pp. 40–45. IEEE
(2015)
24. Navalkar, S.T., Wingerden, J.W.V.: Iterative feedback tuning of an LPV Feedforward
controller for wind turbine load alleviation. IFAC PapersOnLine 48(26), 207–212 (2015)
25. Rădac, M.B., Precup, R.E., Petriu, E.M.: Stable iterative feedback tuning method for servo
systems. In: 2011 IEEE International Symposium on Industrial Electronics, pp. 1943–1948.
IEEE (2011)
26. Precup, R., Mosincat, I., Radac, M.: Experiments in iterative feedback tuning for level
control of three-tank system. In: 2010 IEEE Mediterranean Electrotechnical Conference,
Melecon 2010, pp. 564–569. IEEE (2010)
IoT Systems
 Data Monitoring for Interconnecting
Microgrids Based on IOT

Weihua Deng(&) and Shufen Wang

Shanghai University of Electric Power, Shanghai 200090, China

[email protected]

Abstract. The real time monitoring is very important for the interconnecting
microgrids. The power production status is not only grasped by watching the
real time data, but also some optimization approach can be employed to improve
operation mode thus dispatching power reasonably. But now a full range of
monitoring of interconnecting microgrids is not implemented yet. In the present
age of big data, the computing capability and storage space have been devel-
oping fast. These advanced techniques enable the large-scale data monitoring
system. The big data from monitoring can be analyzed and applied into the
optimization of production process. Motivated by this, a real-time data moni-
toring system simulation platform is developed. Specially, ThingSpeak platform
is validated in our work.

Keywords: Microgrid
IOT
Thingspeak
Monitoring

1 Introduction

The Microgrid (MG) is considered to be the most promising intelligent power man-
agement system that can be operated in parallel or on an island. The power generated
by a single microgrid is limited and can only provide a maximum load capacity of
approximately 10 MVA. However, several MGs can be interconnected to provide more
power to meet greater load requirements. It also has more redundancy and ensures
better power supply reliability [1]. Multiple interconnected microgrids are often
referred to as Multi Microgrid (MMG) [2]. The MMG system connects multiple
individual microgrids and other distributed generation to a medium voltage distribution
network. Currently, each MG is usually based on the local control strategy and does not
coordinate with other microgrids according to local requirements and goals [3, 4].
However, it is expected that future subtransmission and distribution systems will
consist of several interconnected microgrids and form a complex grid [5].
In an interconnected microgrid, each microgrid will be connected to the next
microgrid by High Voltage DC (HVDC) transmission based on Voltage-Sourced
Converter (VSC). With the rapid development of fully-controlled power electronic
devices and power electronics, high-voltage direct current (HVDC) systems based on
voltage source converters (VSC-HVDC) and pulse width modulation technologies have
been extensively studied and applied [6]. The VSC-HVDC system provides an eco-
nomical and reliable solution for large-capacity long-distance power transmission.

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 383–389, 2018.
https://doi.org/10.1007/978-981-13-2384-3_35
384 W. Deng and S. Wang

Recently, the development process of the VSC-HVDC system shows that they are a
better alternative to the traditional thyristor-based HVDC system [7].
A large range of modeling and control of the VSC-based HVDC system have been
published in the last few years [8]. From these studies, it becomes clear that the control
and the system impedance can have the impact on the stability of the system. Usually,
we need to pre-assess their impact on the system stability before connecting to the main
grid [9, 10]. Interconnecting together multiple microgrids can lead to undesirable
dynamic behaviors, that is why we must monitor some of the important parameters.
It is, therefore, necessary to data monitoring based on IOT. In particular, this paper
investigates interconnecting together multiple microgrids, thus provide a method for
monitoring data. The remote monitoring center can read the data stored in ThingSpeak
from any microgrid at any time. The main objective in this paper is to use thingspeak
on the microgrid data monitoring, and the remote monitoring center can access
microgrid data at any time for data analysis.

2 System Model

A single microgrid system operated in islanded mode and data monitoring center can
monitor data by the communications network. In other words, each microgrid can
transmit data in real time, and for the data monitoring center can always get any time
required data from any microgrid. Figure 1 shows the comprehensive architecture of
IOT framework with ThingSpeak cloud service for microgrid data monitoring. Where
P1, P2, P3 respectively are active power of each microgrid, Q1, Q2, Q3 respectively are
reactive power of each microgrid, and Udc is DC voltage of VSC-HVDC link.

Data monitoring Data monitoring

center center

Thingspeak
P1 Udc1 P2 Udc2 P3
Q1 Q2 Q3

Microgri Microgri Microgri

d1 d2 d3

VSC- VSC-
HVDC Link HVDC Link

VSC-
HVDC Link

Udc3

Fig. 1. architecture of IOT framework with ThingSpeak cloud service

 Data Monitoring for Interconnecting Microgrids Based on IOT 385

2.1 VSC-Based HVDC System Model

The two-terminal VSC-based HVDC system under this study is depicted in Fig. 2.
The HVDC system consists of the converter transformers, ac filters, two VSC HVDC
converters named VSC1 and VSC2, and the dc cable. Both the VSC1 and VSC2 are
assumed to be identical in structure. The electrical circuit of a VSC-HVDC converter
for analytical modeling is shown in Fig. 3.

Fig. 2. VSC-based HVDC system

Overview of the control system of a VSC converter and interface to the main circuit
shows an overview diagram of the VSC control system and its interface with the main
circuit.

Fig. 3. VSC-HVDC converter

386 W. Deng and S. Wang

The converter 1 and converter 2 controller designs are identical. The two controllers
are independent with no communication between them. Each converter has two degrees
of freedom. In our case, these are used to control: P and Q in station 1 (rectifier); Udc
and Q in station 2 (inverter).

2.2 IOT with ThingSpeak

ThingSpeak is an IOT analytics platform service that allows you to aggregate, visu-
alize, and analyze live data streams in the cloud [11]. You can send data to ThingSpeak
from your devices, create instant visualizations of live data.
With MATLAB analytics inside ThingSpeak, you can write and execute MATLAB
code to perform preprocessing, visualizations, and analyses. You can collect and
analyze data from internet connected sensors using ThingSpeak. Matlab Simulink is
connected with ThingSpeak by the channel as shown Fig. 4.

Fig. 4. the diagram of connection of ThingSpeak and Matlab

3 Results

3.1 Write the Data from MATLAB Simulink

Since ThingSpeak supports up to 8 data fields, we can use ThingSpeak. write to send
more than one value to the special channel of ThingSpeak. As the simulation pro-
gresses, for example, we transfer data from microgrid 1 in real-time. The data will be
passed in ThingSpeak cloud as shown in Fig. 5.
There are matlab code to write data to ThingSpeak as follows:
function [x,y,z] = fcn(u,m,n)
x=u;
y = m;
z=n;
channelID = 242552 ;
writeKey = 'BZ57MIMSENY2F992'
ThingSpeakWrite(channelID,[x,y,z],'field',[1,2,3],'W
riteKey', writeKey);
pause(15);
 Data Monitoring for Interconnecting Microgrids Based on IOT 387

Fig. 5. the data from microgrid 1 is displayed in ThingSpeak platform with simulation.
(a) microgrid 1 Udc, (b) microgrid 1 P, (c) microgrid 1 Q

3.2 Read the Data Stored in ThingSpeak

We will use another computer as a remote monitoring center. The monitoring center
can be build in any place where it is needed. Our remote monitoring center can read the
data stored in ThingSpeak from any microgrid at any time. We can always read the data
stored in ThingSpeak, next we will verify its feasibility.
Before we have transfer the VSC-based HVDC simulation data to ThingSpeak,
now we read the data from 2017/5/17 08:00:00 to 2017/5/17 10:00:00. Firstly, we
begin by specifying a start date and an end date using a datetime object. And we use
ThingSpeak. read to read the stored in ThingSpeak. We then append the data and time
from each field into two vectors, called data and time. Next we plot the data and time as
shown in Fig. 6. (a) microgrid 1 Udc, (b) microgrid 1 P, (c) microgrid 1 Q.
388 W. Deng and S. Wang

There are matlab code to read data stored in ThingSpeak channel as follows:
endDate=datetime('5/14/2017
08:00:00','InputFormat','MM/dd/yyyy
hh:mm:ss');
startDate=datetime('5/12/2017
06:00:00','InputFormat','MM/dd/yyyy hh:mm:ss');
dateVector=[startDate, endDate];
channelID =270840;
readKey ='8NQ2MRCG53WICFTS';
[data,time]=
ThingSpeakRead(channelID,'field',[1,2,3],'DateRang
e',dateVector,'ReadKey',readKey);
plot(time, data)

(a) (b)

(c)

Fig. 6. read the data from ThingSpeak platform (a) microgrid 1 Udc, (b) microgrid 1 P,
(c) microgrid 1 Q
 Data Monitoring for Interconnecting Microgrids Based on IOT 389

4 Conclusions

Data monitoring based on internet of thing for interconnecting together multiple

microgrids in this paper. It is shown that even systems that are far apart can share
information and monitor data through ThingSpeak. This paper can be applied to
multiple microgrids, and we can remotely monitor the dynamic interaction between
microgrids thereby controlling the flow of energy. And then an intelligent power
management system is proposed to facilitate power trading among microgrids and
improved reliability.

References
1. Chowdhury, S., Chowdhury, S.P., Crossley, P.: Microgrids and Active Distribution
Networks, Renewable Energy Series, vol. 6. Institution of Engineering and Technology,
Stevenage (2009)
2. Gil, N.J., Lopes, J.A.P.: Hierarchical frequency control scheme for islanded multi-microgrids
operation. In: IEEE Lausanne Power Tech Lausanne, pp. 473–478 (2007)
3. Lopes, J.A.P., Hatziargyriou, N., Mutale, J., Djapic, P., Jenkins, N.: Integrating distributed
generation into electric power systems: A review of drivers, challenges and opportunities.
Lecct. Rower Syst. Rest. 77(9), 1189–1203 (2007)
4. Buayai, K., Ongsakul, W., Mithulananthan, N.: Multi-objective micro-grid planning by
NSGA-II in primary distribution system. Eur. Trans. Elect. Power 22(2), 170–187 (2012)
5. Koyanagi, K., et al.: Electricity cluster-oriented network: a gridindependent and autonomous
aggregation of micro-grids. In: Proceedings of the International Symposium: Modern
Electric Power System, Wroclaw, Poland, pp. 1–6, September 2010
6. Aditya, S.K., Das, D.: Load-frequency control of an interconnected hydro-thermal power
system with new area control error considering battery energy storage facility. Int. J. Energy
Res. 24, 525–538 (2000)
7. Usama, M.U., Kelle, D., Baldwin, T.: Utilizing spinning reserves as energy storage for
renewable energy integration. In: Power Systems Conference, pp. 1–5 (2014)
8. Ktiraei, F., Iravani, R., Hatziargyriou, N., Dimeas, A.: Microgrids Management: Controls
and Operation Aspects of Microgrids. IEEE Power and Energy Magazine (2008)
9. Beerten, J., Cole, S., Belmans, R.: Modeling of multi-terminal VSC HVDC systems with
distributed DC voltage control. IEEE Trans. Power Syst. 29(1), 34–42 (2014)
10. Lu, W., Ooi, B.-T.: Optimal acquisition and aggregation of offshore wind power by
multiterminal voltage-source HVDC. IEEE Trans Power Deliv. 18, 201–206 (2003)
11. Maureira, G.A.M., Oldenhof, D., Teernstra, L.: ThingSpeak—an API and Web Service for
the Internet of Things. World Wide Web, 7 November 2015
 A Hybrid Routing Control Mechanism
for Dual-Mode Communication of Streetlight
Information Acquisition System

Min Xiang(&), Xudong Zhao(&), and Yongmin Sun(&)

Key Laboratory of Industrial Internet of Things & Networked Control,

Chongqing University of Posts and Telecommunications,
Chongqing 400065, China
[email protected], [email protected],
[email protected]

Abstract. In order to improve network stability and data acquisition success rate
of streetlight information acquisition network, a hybrid routing control mecha-
nism is proposed. Combining the distribution and access mode of streetlight node,
and signal attenuation differences between wireless and power line carrier, a dual-
tree network is constructed. Firstly, wireless and carrier next hop nodes are
respectively calculated in dual-tree network by post-order traversal algorithm.
Secondly, next hop node of forwarding packet is selected from wireless and carrier
next hop node by the index of transformation rate of signal. The test results show
that hybrid routing control mechanism can improve network stability and data
acquisition success rate of streetlight information acquisition network.

Keywords: Dual-mode communication
Hybrid routing control

Dual-tree network
Streetlight information acquisition system

1 Introduction

With the development of Internet of Things (IOT) technology, it is a trend that

streetlight control system gradually turns to be intelligent [1]. Now, Power Line
Communication (PLC) and wireless technology are widely applied in streetlight
information acquisition network [2, 3], but it is hard to improve network stability and
data acquisition success rate only using PLC or wireless technology in streetlight
information acquisition network. Now, dual-mode communication technology is used
in smart streetlight, and how to build a reasonable dual-mode network and be efficient
to select routing has become a hot issue.

Min Xiang, born in 1974, Ph.D., Professor. His research interests include Smart Grid, wireless
sensor network and industrial internet of things.
Xudong Zhao, born in 1993, M.S., His research interests include Smart Grid and industrial
internet of things.
Yongmin Sun, born in 1988, M.S., His research interests include Smart Grid and industrial
internet of things.

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 390–399, 2018.
https://doi.org/10.1007/978-981-13-2384-3_36
 A Hybrid Routing Control Mechanism for Dual-Mode 391

In recent years, there has been relevant research on the application of dual-mode
communication technology in streetlight information acquisition network. In [4], a
comparison between PLC and wireless used in streetlight control system is proposed,
and the advantages and disadvantages about them are introduced. In [5], a design of
remote control system based on PLC in the internet of things is proposed. The node can
get the RSSI decision channel by obtaining the power carrier and the radio receiving
signal intensity. In [6], a method of channel selection and routing strategy is proposed
based on the ladder algorithm. The node makes the decision channel through the field
strength and the pre divided channel priority.
In order to improve network stability and data acquisition success rate of streetlight
information acquisition network, a hybrid routing control mechanism is proposed.
Firstly, a dual-tree network is constructed. Secondly, routing of streetlight node for-
warding packets is calculated by post-order traversal algorithm and index of trans-
formation rate of signal (TROS).

2 Dual-Mode Communication Architecture

In the streetlight information acquisition network, communication network includes

long-distance communication network and local dual-mode communication network.
The long-distance communication network is constructed for communicating between
the master station and the concentrator node. Local communication network is con-
structed for communicating between the concentrator node and the streetlight node [7].
The dual-mode communication modules are used in the device of concentrator and the
streetlight. The dual-mode communication architecture of streetlight information
acquisition system is shown in Fig. 1.
Concentrator node consists of concentrator and dual-mode communication module,
which builds and manages streetlight information acquisition network. Streetlight node
which consists of master streetlight and dual-mode communication module is
responsible for measuring streetlight data.

Master station

Remote
communications Ethernet GPRS CDMA Others

Concentrator

PLC PLC
Local dual-mode
communications

Wireless Wireless
Streetlight

Fig. 1. Dual-mode communication architecture of streetlight information acquisition system

392 M. Xiang et al.

3 Hybrid Routing Control Mechanism

3.1 Overall Scheme Design
Dual-tree network is proposed by combining the distribution of streetlight equipment,
access mode of streetlight equipment and signal attenuation differences between
wireless and PLC. In the streetlight information acquisition system, the distribution of
streetlight nodes have the following characteristics: streetlight nodes are in a parallel
strip distribution. Streetlight nodes are mostly tree-like access to streetlight information
acquisition network, and the network topology of streetlight nodes is designed to be
dual-tree network. There is signal attenuation differences between wireless and PLC.
Wireless signals are mainly affected by buildings, distances and weather. After
streetlight nodes are deployed, buildings and distances of two nodes are stable, but
weather would affect larger areas and lots of nodes. Carrier signals are mainly affected
by distances and power line load. After streetlight node is deployed, the distances of
two streetlight nodes is stable, but power line load only affects small area and a little
nodes. Relay is an effective method to improve the reliability of carrier network. Nodes
build wireless and carrier routing metric by hop and RSSI. Wireless routing metric
tends to RSSI, and carrier routing metric tends to hop count.
When streetlight node forwards packet, the post-order traversal algorithm is
adopted to calculate the wireless and carrier next hop node in dual-tree network.
Wireless and carrier next hop node of streetlight node forwarding packet is calculated
by post-traversing wireless and carrier routing forwarding tree. TROSw is transfor-
mation rate of signal between current node and wireless next hop node. TROSp is
transformation rate of signal between current node and carrier next hop node. The
overall scheme of hybrid routing control mechanism is shown in Fig. 2.

Start

N
Received packet?
Calculate TROSw and TROSp
Y

Receive and Y Destination address of

process packet is the current
node? N
TROSw>TROSp?
N

N Y
Wait for a moment Dual-tree network has
been built? Next hop node is Next hop node is
wireless next hop carrier next hop
Y node node

Calculate wireless and carrier

next hop node

End

Fig. 2. Overall scheme of hybrid routing control mechanism

 A Hybrid Routing Control Mechanism for Dual-Mode 393

3.2 Dual-Tree Network Construction

Streetlight node combines the distribution of streetlight equipment, access mode of
streetlight equipment and signal attenuation differences between wireless and PLC, and
dual-tree network is constructed. Four network control message frames is defined, and
the dual-tree network construction process is divided into three stages.
(1) Network control message frame. In order to build a dual-tree network, four net-
work control message frames are defined. Four network control message frames
consists of network beacon frame, route request frame, route reply frame and
routing update frame. Streetlight node broadcasts route request frame to get
network beacon frame, unicasts route reply frame to parent node to update current
node route, and unicasts routing update frame to update child node route to
concentrator node. The format of network beacon frame is shown in Fig. 3. The
Source Address is 48-bit MAC address. The MAC payload includes one-byte
HOP count, one-byte control domain C and one-byte child node number N. High
four bit of control domain C identifies the network type, and network beacon
frame consists of wireless and carrier network beacon frame.

0 7 55 63 71 79 80 87
32H Source Address C HOP N 16H
MHR MAC payload MFR

Fig. 3. Format of Network beacon frame

(2) Network initialization stage. When streetlight node and concentrator node are
powered on, firstly the network will initialized. The dual-mode communication
modules of concentrator and streetlight node are initialized on software and
hardware. For example, IPv6 protocol stack, software state machine and hardware
I/O. The IPv6 address of concentrator node is set by information from the master
station. After its IPv6 address is configured, the wireless routing forwarding tree
and carrier routing forwarding tree is built by concentrate node, and the wireless
and carrier network beacon frame is broadcasted.
(3) Network discovery stage. Streetlight node needs to wait for a while and judge
whether network beacon frame is received. If streetlight node hasn’t received
network beacon frame since node began to wait, it will broadcasts a route request
frame to its neighboring neighbor nodes and requests a network beacon frame. If
the network beacon frame is received, streetlight node will analyze network
beacon frame to obtain hop count, the content of control domain C, and source
address. Streetlight node calculates RSSI in data link layer. At the same time, the
streetlight receive IPv6 network prefix by neighbor discovery protocol (NDP).
(4) Access control stage. The streetlight node firstly evaluates communication per-
formance of the network to be accessed. The streetlight node sets wireless and
carrier routing metric by hop count and RSSI. When streetlight node evaluates the
communication performance of wireless network, node focuses on RSSI rather
394 M. Xiang et al.

hop count. When streetlight node evaluates the communication performance of

carrier network, streetlight node focuses on hop count rather RSSI. Streetlight
node evaluates communication performance of wireless and carrier network by
Eq. 1. The b is weight coefficient, 0.5 < b < 1, normally b = 0.6, the Cm is the
permitted biggest number of child node. The Lm is the permitted biggest number
of child node. Streetlight node joins the network which has the highest commu-
nication performance, and itself hop count is set as a value which is equal to
parent hop count plus one. The hop parent is the hop count of parent streetlight
node. Then, streetlight node unicasts routing reply frame to parent node, and
broadcasts wireless and carrier network beacon frame.
(
b 10rssi
lg w þ ð1  bÞð1  Cm Þ;
hop
wireless
r¼ : ð1Þ
bð1  Cm Þ þ ð1
hop
 bÞ 10rssi
lg w ; plc

When streetlight node firstly joins the network, its parent node assigns a two-byte
short address Ac to current streetlight node by Eq. 2 and Eq. 3 [8, 9]. The Ap is two-
byte short address of parent node, and the n is the number of child node of parent node.
In Eq. 3. The d is the hop count of current node. After having set two-byte short
address, the streetlight node can obtain the PANID and obtain 64-bit IPv6 network
prefix of parent streetlight node.

AC ¼ AP þ 1 þ CskipðdÞðn  1Þ; ð1  n  Cm Þ : ð2Þ

(
1 þ Cm ðLm  d  1Þ; Cm ¼ 1
CskipðdÞ ¼ 1CmLm d : ð3Þ
1Cm ; Cm [ 1

A dual-tree network is built by one concentrator and seven streetlight nodes, which
is shown in Fig. 4. The concentrator node is the root node of this dual-tree network,
and each streetlight node has one wireless and carrier parent node. When streetlight
node C joins dual-tree network, streetlight node C doesn’t choose concentrator node as
wireless parent node because there is a barrier between concentrator node and street-
light node C, but streetlight node C still choose concentrator node as carrier parent node
because barrier can’t affect carrier signal. In dual-tree network, some streetlight nodes
have the same wireless and carrier parent nodes while some streetlight nodes have
different wireless and carrier parent nodes. When concentrator node and seven street-
light nodes all use wireless channel to communicate, dual-tree network can be regarded
as wireless routing forwarding tree, and the wireless next hop count of the streetlight
node C is two.

3.3 Hybrid Routing Control

When packet is received by streetlight node, if destination address of packet is the
current node, node will receive and handle this packet. If not, node will forward this
packet. Firstly wireless and carrier next hop node is calculated. Secondly, TROS is
 A Hybrid Routing Control Mechanism for Dual-Mode 395

A B C D
Concentrate
Rode
PLC
Street Wireless
Barrier
Rode
I G FH E

Fig. 4. Dual-tree network

calculated between current node and wireless and carrier next hop node. Finally, next
hop node is calculated by TROS and wireless and carrier next hop.
(1) Calculate wireless and carrier next hop node. Wireless next hop node of for-
warding packet is calculated by post-traversing wireless routing forwarding tree,
and carrier next hop node of forwarding packet is calculated by post-traversing
carrier routing forwarding tree. The process of streetlight node post-traversing
routing forwarding tree is shown in Fig. 5.

Start

Post-traverse a node

N This node is the

destination node?

Receive and Y
This node is the current
process
node?

Post-traverse a node

Y
N

This node is the child N

This node is the current
node of current node?
node?

Y Y

Next hop node

End

Fig. 5. Process of streetlight node post-traversing routing forwarding tree

In Fig. 5, when current node post-traverses a node in the routing forwarding tree, it
will judge whether the node is destination node of packet. If this node isn’t the
396 M. Xiang et al.

destination node, current node will continue post-traversing a node and judging. If this
node is the destination node, current node will post-traverse a node and judge whether
the node is the child node of current node. If the node is child node of current node, the
node is next hop node. If not, current node won’t stop post-traversing a node and judge
whether the node is child node of current node until the node is current node or child
node of current node.
(2) Calculate TROS. In the streetlight information acquisition network, there are
different transmit powers of wireless and carrier communication modules, so the
wireless and carrier RSSI can’t be compared to judge the communication quality
of wireless and carrier channels, and the index of TROS is defined. The current
streetlight node calculates TROS by Eq. 4. The rssi is the received signal strength
of the channel by which current streetlight node communicates with node N, and
the w is the transmit power value in milliwatts. The current streetlight node should
calculate wireless and carrier transformation rate of signal.

10 lg w  rssi
TROSðNÞ ¼  100%: ð4Þ
10 lg w

(3) Hybrid routing control. When current streetlight node receives packet and the
destination address of received packet is the current node, the node receive and
process this packet. When forwarding received packet, the current streetlight node
calculates the next hop node Nnext by Eq. 5. The node N1 is the wireless next hop
node, and the node N2 is the carrier next hop node.
8
< N1 ; TROSðN1 Þ [ TROSðN2 Þ
Nnext ¼ N2 ; TROSðN1 Þ\TROSðN2 Þ : ð5Þ
:
random; TROSðN1 Þ ¼ TROSðN2 Þ

4 Test and Analysis

In order to test the performance of hybrid routing control mechanism, a test platform is
set up. This test platform consists of ten streetlight nodes, one concentrator node and
master station. Each streetlight node and concentrator node uses dual-mode commu-
nication module which has wireless and carrier communication mode. In the test, there
are walls between two streetlight nodes, and test platform is shown in Fig. 6.
Related network parameters is shown in Table 1. The Nw is the total barriers in test
area, the Rw is the communication radius of wireless modules, and the Lm is the biggest
permissible number of child nodes.
(1) Parent node update times. In order to verify the ability of improving network
stability by hybrid routing control mechanism (HRCM), HRCM is compared with
carrier communication scheme proposed in [10] and the wireless communication
 A Hybrid Routing Control Mechanism for Dual-Mode 397

Master station

GPRS Concentrator node

Streetlight node
Street
15m

20m

Fig. 6. Test platform

Table 1. Network parameters

Parameter Value Definition
Nw 10 Total barrier number
Rw 80 m Communication radius of wireless
Rp 310 m Communication radius of PLC
Cm 5 The biggest hop
Lm 6 The biggest number of child node
b 0.6 Scale factor

scheme proposed in [11]. The devices shown in test platform are adopted to build
network of streetlight information acquisition by three communication schemes,
and concentrator node calculates the parent node update times. The parent node
update times under different time is shown in Fig. 7. The parent node update times
under different numbers of wall is shown in Fig. 8.
In Fig. 7, the parent node update times in the daytime is more than it at night, but
the parent node update times of HRCM is less than the parent node update times of

Fig. 7. The parent node update times under Fig. 8. The parent node update times under
different time different numbers of wall
398 M. Xiang et al.

PLC in the same time. In Fig. 8, with the increase of the numbers of wall, the parent
node update times gradually increase, but the parent node update times of HRCM is
less than the parent node update times of wireless in the same number of streetlight
nodes. Combining the distribution of streetlight equipment, access method of streetlight
equipment and signal attenuation differences between micro power wireless and power
line communication, a dual-tree network is proposed. The dual-tree has the less parent
node update times.
(2) Packet delivery rate. In the network of streetlight information acquisition, the
packet delivery rate can directly reflect the communication performance of net-
work. In order to verify the ability of improving acquisition success rate streetlight
information, HRCM is compared with PLC and wireless by the index of packet
delivery rate. The devices shown in test platform are adopted to build network of
streetlight information acquisition by three communication schemes, and con-
centrator node calculates packet delivery rate. The packet delivery rate under
different time and wall number is shown in Table 2.

Table 2. Packet delivery rate under different time and nodes

T N Nw Size HRCM PLC WIRELESS
08:00 10 0 1000 100% 99.3% 100%
12:00 10 2 1000 100% 98.2% 100%
16:00 10 4 1000 99.8% 98.3% 99.7%
20:00 10 6 1000 99.6% 98.5% 99.5%
24:00 10 8 1000 99.9% 99.3% 98.9%
04:00 10 10 1000 100% 99.8% 98.1%

In Table 2, with the increasing number of streetlight nodes, the packet delivery rate
of wireless is gradually reduced, and the packet delivery rate of PLC in the day time is
less than at night. The packet delivery rate of HRCM is the highest in three commu-
nication schemes. In the network built by HRCM, streetlight node evaluates wireless
and carrier channel communication performance by TROS index, and chooses the
better channel to transfer packet.

5 Conclusions

In order to improve network stability and data acquisition success rate of streetlight
information acquisition network, a hybrid routing control mechanism is proposed.
Firstly, a dual-tree network is constructed by combining the distribution of streetlight
equipment, access mode of streetlight equipment, and signal attenuation differences
between wireless and power line carrier. Secondly, when node forwards packet, it
firstly post-traverses wireless and carrier routing forward tree to calculate wireless and
carrier next hop node; node secondly calculates next hop node by index of TROS and
 A Hybrid Routing Control Mechanism for Dual-Mode 399

wireless and carrier next hop node. The test results show that hybrid routing control
mechanism can improve network stability and data acquisition success rate of street-
light information acquisition network.

Acknowledgments. This work is supported by Key R&D program of common key technology
innovation for key industries in Chongqing (cstc2017zdcy-zdyfX0032).

References
1. Bai, C.-L., Ma, J.: Intelligent street lamp monitoring system based on internet of things.
J. Appl. Electron. Tech. 40(3), 82–85 (2014)
2. Yan, Z.-L., Gao, S.-J., Huang, X.-Y.: Application of wireless sensor network based on IPv6-
based low-power wireless personal area network in smart street lights. J. China Electric
Power 23, 78–79 (2016)
3. Gao, Y.-H., Liang, X.-Y.: Design of intelligent streetlight control system based on Zigbee.
J. Modern Electron. Technol. 19, 29–32 (2013)
4. Zhu, M.: Comparison of Zigbee and power line carrier in street lamp single lamp control
system. J. China New Commun. 17(19), 39–40 (2015)
5. Li, Z.-Z., Du, Y., Fan, X.-Z.: Research and design of dual mode heterogeneous
communication based on ladder algorithm. J. Electro Tech. Appl. 17, 118–121 (2015)
6. Kong, Y.-H., Li, J.-C., Chen, Z.-X.: Design and Realization of an Adaptive Dual-mode
Communication Module Based on RSSI. J. Sci. Technol. Eng. 16(23), 203–207 (2016)
7. Zhu, Y.-N., Liu, J., Xu, Q.: Application of dual-mode communication in bidirectional
interaction of smart meters. J. Appl. Electron. Tech. (z1), 210–213 (2015)
8. Qi, Z.-Y., Li, Z.-F.: Microgrid monitoring network based on Zigbee communication.
J. Comput. Eng. 43(4), 79–83 (2017)
9. Zhou, S.-Y., Gao, J.-Z., Zou, S.-H.: Energy-efficient data storage with non-uniform nodes
distributed wireless sensor networks. J. Chongqing Univ.: Nat. Sci. Edn. 40(9), 57–66
(2017)
10. Xiang, M., He, J.-X., Du, Y.-H.: An IPv6 based-tree routing control mechanism for low-
voltage power line communication. J. Power Syst. Technol. 40(06), 1874–1880 (2016)
11. Huang, Z.-C., Yuan, F., Li, Y.: Implementation of IPv6 over low power wireless personal
area network based on wireless sensor network in smart lighting. J. Comput. Appl. 34(10),
3029–3033 (2014)
 An RF Energy Harvesting Approach for
Secure Wireless Communication in IoT

Cheng Yin1(B) , Emiliano Garcia-Palacios1 , and Hien M. Nguyen2

1
Queen’s University Belfast, Belfast, UK
[email protected], [email protected]
2
Duy Tan University, Da Nang, Vietnam
[email protected]

Abstract. Mobile terminals and base stations are being deployed glob-
ally resulting in an increase in carbon emissions. Energy harvesting is
attracting considerable attention and an environment- friendly commu-
nication system is desirable. In this paper, we demonstrate such a sys-
tem. A power beacon provides energy for a transmitting source and for a
relay using time switching radio frequency energy harvesting technique.
We demonstrate that when an eavesdropper tries to wiretap the signals
transmitted from the source and the relay the system remains secure.
The closed-form expression for secrecy outage probability is derived. Our
results show how the performance varies with the SNR of the power bea-
con and the distance from the relay and the source to the power beacon.
In addition, we show that the duration of energy harvesting process in
relation to the time used for transmitting information has a significant
impact on the system security.

1 Introduction
In early Internet of Things (IoT) applications batteries are being used in plenty
of sensor devices and embedded equipment but their lifespan is limited and they
need replacement and maintenance. Most batteries end up in landfills and this
has a bad impact on the environment. To reduce the use of batteries, energy
harvesting (EH) techniques are gaining increasing interest. Devices and equip-
ment can derive energy by themselves using energy harvesting technology from
environmental resources and prolong their lifetime. Ideally sensors could derive
enough energy to become battery-less! When performing a harvesting process,
energy is gathered from the external environment, i.e., solar, wind, vibration,
electromagnetic, thermoelectric phenomena. In this way, sufficient energy could
be harvested to drive a wireless sensor network (WSN) for an IoT application.
To compare the amount of harvested energy from these resources, we show some
figures from [4,5] in Table 1. Although solar, vibration, thermal have more power
density, they are not always available, moreover, the cost of energy such as solar
is still high.
Compared with other sources in Table 1, due to the increase in mobile ter-
minals and base stations, electromagnetic energy is available anywhere (both
c Springer Nature Singapore Pte Ltd. 2018
K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 400–407, 2018.
https://doi.org/10.1007/978-981-13-2384-3_37
 An RF Energy Harvesting Approach for Secure Wireless Communication 401

Table 1. an overview of different energy sources and harvested energy [4, 5]

Source Harvested power

Vibration
Human 4 µw/cm2
Industrial 100 µw/cm2
Thermal
Human 30 µw/cm2
Industrial 1−10 mw/cm2
Solar
Indoor 10 µw/cm2
Outdoor 10 mw/cm2
Radio Frequency (RF)
Cell phone 0.1 µw/cm2
Wi-Fi 1 mw/cm2
GSM 0.1 µw/cm2

indoor and outdoor) and anytime. In addition, it can be retrieved boundlessly

from multiple power sources in various scenarios, e.g., TV, radio stations, satel-
lite stations, cell phones masts, Wi-Fi, etc. In order to capture electromagnetic
energy, some authors have already designed a system to realize it [4]. The key
modules are the antenna and rectifier circuit which can transfer RF signals or
alternating current (AC) to direct current (DC) [4]. The process of harvesting
energy from RF signals is clean and does not produce any pollution to the envi-
ronment. Several schemes have been proposed for RF energy harvesting because
simultaneously harvesting power and transmitting information is not practical
due to the limitation of hardware [2]. There are two main protocols to tackle
this problem, i.e. time switching (TS) protocol and power splitting (PS) proto-
col. In TS protocol, during a given period, time is dived into two parts for energy
harvesting and information transmitting. If PS protocol is used instead, some
power is used for energy harvesting and the rest is used for signal transmitting.
In [8] a TS protocol was considered in a cognitive radio context and the impact
of the position of the primary user (far or near) upon secondary transmitters
was studied.
In IoT and WSN where communication technologies are limited in transmis-
sion range from source to destination the use of relays is desirable. Relays help
to extend the range of sensor nodes thus helping to interconnect nodes in the
sensor network or helping a given sensor to reach a distant destination. Existing
research has shown that relaying networks can enhance the system performance
[7]. At the radio layer there are two relay alternatives, amplify-and-forward (AF)
and decode-and-forward (DF). Compared with AF, a relay network operating
with DF relays has better performance because the interference is lower [6].
402 C. Yin et al.

However, DF relays require more power to operate due to its complexity. Relay
nodes can only be equipped with finite energy to work for a functioning time
[1]. Replacing and charging batteries is not desirable. This motivate us to con-
sider exploiting RF energy harvesting for relays too. We could envisage a sensor
network where sensors and relays are now battery-less.
In future sensor applications, security is also a crucial problem that we can-
not ignore. Because of the broadcast nature of wireless channel, information is
vulnerable to eavesdropping [6]. Traditional way to secure wireless information
is to deploy cryptographic techniques at higher layers despite consuming huge
power for encrypting, decrypting data and burdening the protocol stack which
is to be avoided in energy constrained sensor networks. At the lower layer, phys-
ical layer security (PLS) has gained popularity in recent years to secure wireless
communication.
In this research, we propose a mathematical approach and system model
to obtain the system performance, i.e. secrecy outage probability (SOP). In
the system, an information source (e.g. a sensor) transmits desired signals to a
destination (e.g. a base station or sink) with the help of a relay node. In the
future we envisage IoT scenarios where the sensors and relay nodes are battery-
less. To enable such a system, a power beacon is assumed to provide energy for
the information source and the relay node. Then the information source and
the relay forward the signals with the harvested energy. To assess security, an
eavesdropper is also considered in the system, and it can wiretap signals from
the source and the relay. In our model, an outage occurs when either the system
is not reliable or not secure, hence we assess the secrecy outage probability as a
performance parameter. Our contributions are summarized as:
– We propose a system model including an information source, a relay, a des-
tination, an eavesdropper and a power beacon. The information source and
relay node harvest energy from the power beacon using TS protocol.
– The impacts of the location and the signal-to-noise ratio (SNR) of power
beacon on the system performance are investigated.
– The fraction of time dedicated for energy harvesting in a TS frame α versus
the time dedicated for information transmission to the relay and destination
has a direct impact on performance. This is investigated to find the optimum
value of alpha.
– Closed-form expression for SOP is derived. In addition, Monte Carlo method
is used to evaluate simulation results.

2 System Model
We consider a network with a power beacon B, an information source S, a DF
relay R, a destination D and an eavesdropper E, as shown in Fig. 1. All the
nodes are equipped with one antenna and all the channels are independent and
identically distributed. The additive white Gaussian noise (AWGN) at Rk and D
has zero mean and variance N0 . Assuming all the channels are Rayleigh fading
and the channel power gains follow exponential distribution with parameter λAB ,
where A ∈ {B, S, R} and B ∈ {S, R, D, E}.
 An RF Energy Harvesting Approach for Secure Wireless Communication 403

Power Beacon

Cloud

Source Relay DesƟnaƟon

Energy HarvesƟng Link

LegiƟmate Link

IllegiƟmate Link
Eavesdropper

Fig. 1. An RF energy harvesting system model for secure IoT

2.1 Energy Harvesting Scheme

In the considered system, B provides energy for S and R using TS protocol based
energy harvesting technique, as shown in Fig. 2, because of its high throughput. α
is the EH time fraction (0 < α < 1) that has been chosen as a design parameter.
In a transmission block time T , it takes S and R a period of αT to harvest energy
from B and then both S and R use (1−α)T 2 to transmit information to R and D.

Fig. 2. TS Protocol

Hence, the energy harvested at S and R are

ES = ηPB αT |hBS |2 , (1)
ER = ηPB αT |hBR | ,
2
(2)
where η is the efficiency coefficient of the energy conversion ratio between the
power delivered to the load and the harvested power at the antenna (0 < η < 1)
[4]. PB is the transmit power of power beacon. |hBS |2 is the channel power gains
of B → S and |hBR |2 is the channel power gains of B → R. The transmit power
at S and R can be derived as below,
2ηPB |hBS |2 α
PS = , (3)
(1 − α)
2ηPB |hBR |2 α
PR = . (4)
(1 − α)
404 C. Yin et al.

2.2 Security Scenarios

In the considered system, E can wiretap the information from S and R. Assume
that E is equipped with energy and B does not provide energy for E. S and R
use different code books to enhance the security performance. Hence, the secrecy
capacity can be obtained as

Cs = min(C1s , C2s ), (5)

where C1s and C2s are the achievable secrecy rate of S → R and R → D, expressed
as follows:
  +   +
1−α 1 + γ1M 1 + γ1M
C1s = log2 =  log2 , (6)
2 1 + γ1E 1 + γ1E
  +   +
1−α 1 + γ2M 1 + γ2M
C2s = log2 =  log2 , (7)
2 1 + γ2E 1 + γ2E

2 clarify the fact that in a block time T , both S → R and R → D

where  = 1−α
use (1 − α)T /2 to transmit information, γ1M is the SNR at R, γ2M is the SNR
at D, γ1E and γ2E are the SNRs of the links S → E and R → E,

γ1M = ξγM |hBS |2 |hSR |2 , (8)

γ2M = γM ξ|hBR | |hRD | ,
2 2
(9)
γ1E = γE ξ|hBS | |hSE | ,
2 2
(10)
γ2E = γE ξ|hBRk | |hRE | ,
2 2
(11)
PB PB
where γM = N 0
, γE = N E
2ηα
, ξ = (1−α) . |hSR |2 ,|hRD |2 , |hSE |2 and |hRE |2 are the
channel power gains of S → Rk∗ , Rk∗ → D, S → E, and R → E. NE is the variance
of the AWGN at E.
The secrecy capacity of the proposed system is given as,
 
1 + γM ξ|hBS |2 |hSR |2
Cs =  log2 min ,
1 + γE ξ|hBS |2 |hSE |2
+
1 + γM ξ|hBR |2 |hRD |2
. (12)
1 + γE ξ|hBR |2 |hRE |2

3 Secrecy Outage Probability

In this section, SOP is derived to obtain the secrecy performance of the consid-
ered system. The concept of SOP is that the probability of secrecy capacity is
below a certain rate [3],

Pout = P {CS < Rth } , (13)

 An RF Energy Harvesting Approach for Secure Wireless Communication 405

where Cs is the secrecy capacity of the system, and Rth is the secrecy target
rate (Rth > 0). An outage occurs when either the system is not reliable or not
secure.
From (12), we have

P {Cs < Rth } = P {γPRS < β}

= FγPRS (β) , (14)

where

1 + γM ξ|hBS |2 |hSR |2
γPRS = min ,
1 + γE ξ|hBS |2 |hSE |2

1 + γM ξ|hBR |2 |hRD |2
. (15)
1 + γE ξ|hBR |2 |hRE |2
Rth
FγPRS (β) is the CDF of γPRS , and β = 2  .
From (14) we have the following lemma.

Lemma 1. The SOP of the considered system is formulated as follows:

√
4γM (β − 1)λSE λRE λBS λBR λSR λRD
FγPRS (β) = 1 −
ξ(γE λSR β + γM λSE )(γE λRD β + γM λRE )



λRD λBR (β − 1) λSR λBS (β − 1)

× K1 2 K1 2 , (16)
γM ξ γM ξ

where K1 (·) is the modified Bessel function of the second kind.

4 Numerical Results
In this section, the simulation results using Monte Carlo approach are evaluated
to prove the accuracy of above performance analysis. The ’Sim’ curves are the
simulation results and ‘Ana’ curves are analytical results. In the figures, we can
observe that both the simulation curves and analytical curves match very well.
Moreover, the parameters are fixed as: γE = 20dB, Rth = 0.2 bits/s/Hz and
η = 0.5. It is assumed that the position of the nodes are located in Cartesian
coordinate system as S = (0, 0), R = (2, 0), D = (3, 0) and E = (1, −4) respec-
tively. The location of B is not fixed because the impact of the location of B is
investigated
 later. Hence, the distance between two nodes can be described as
dAB = (xA − xB )2 + (yA − yB )2 where A, B = {S, B, D, R, E}. A and B have
the co-ordinates (xA , yA ) and (xB , yB ). It is assumed that average SNR of each
link is dependent on the path loss as 1/λX = 1/dpl X , where pl is the path loss
exponent. In this section, pl = 4 is assumed.
In Fig. 3, the impacts of γM and the position of B are investigated. We can
observe that with the increase of γM , the system has a better performance due
to the decrease of SOP. The position of B varies from (1, 0), (1, 1) to (1, 2)
406 C. Yin et al.

which means it moves vertically, further from S and R. We can observe from
the figure that the distance between B to R and S increases, the system has a
worse performance because the SOP increases. This demonstrates the fact that
as the distance between B to R and S increases, R and S harvest less energy for
transmission, thus decreasing the system performance.
In Fig. 4, SOP is plotted as function of α. We assume that γE = 20 dB
and B = (1, 0). It can be observed that SOP is extremely high when α is too
small or too large. This is because when α is small, S and R can not harvest
enough energy. In addition, when α is large, the transmit period will be short, so
the secrecy capacity is low. Therefore, we should design α carefully in order to
achieve a secure wireless system. In our examples, α = 0.6 is nearly an optimum
point.

Fig. 3. SOP with different position of B

Fig. 4. SOP is plotted as a function of α with different γM

 An RF Energy Harvesting Approach for Secure Wireless Communication 407

5 Conclusion
In this paper, an EH system with a DF relay and a power beacon is proposed.
The information source and the relay can harvest energy from the power beacon
and then forward the signal to the destination. The eavesdropper can wiretap the
signals transmitted from the source and the relay. Time switching protocol based
EH is used. The closed-form expression for SOP is derived. The results have
shown that γM and the distance between the power beacon and the source and
the relay can affect the performance of the considered EH system significantly.
When the power beacon locates closer to the source and the relay as well as γM
is high, the system performs significantly better. Moreover, the time dedicated
to harvesting energy in a TS frame α can affect the secrecy performance of the
considered system significantly. Hence, α should be designed carefully. We show
the optimum value in our examples. In summary, we have demonstrated that it
is possible to harvest enough energy from RF sources to transmit information, to
extend the range via a relay and more significantly for the system to be resilient
to an eavesdropper attack.

Acknowledgement. This work was supported by a Research Environment Links

grant, ID 339568416, under the Newton Programme Vietnam partnership. The grant is
funded by the UK Department of Business, Energy and Industrial Strategy (BEIS) and
delivered by the British Council. For further information, please visit www.newtonfund.
ac.uk.

References
1. Akhtar, F., Rehmani, M.H.: Energy replenishment using renewable and traditional
energy resources for sustainable wireless sensor networks: a review. Renew. Sustain.
Energy Rev. 45, 769–784 (2015)
2. Benedict, F.P., Maji, P., Roy, S.D., Kundu, S.: Secrecy analysis of a cognitive radio
network with an energy harvesting AF relay. In: 2017 International Conference
on Wireless Communications, Signal Processing and Networking (WiSPNET), pp.
1358–1363. IEEE (2017)
3. Bloch, M., Barros, J., Rodrigues, M.R., McLaughlin, S.W.: Wireless information-
theoretic security. IEEE Trans. Inf. Theory 54(6), 2515–2534 (2008)
4. Tran, L.G., Cha, H.K., Park, W.T.: RF power harvesting: a review on designing
methodologies and applications. Micro Nano Syst. Lett. 5(1), 14 (2017)
5. Vullers, R., van Schaijk, R., Doms, I., Van Hoof, C., Mertens, R.: Micropower energy
harvesting. Solid-State Electron. 53(7), 684–693 (2009)
6. Yin, C., Nguyen, H.T., Kundu, C., Kaleem, Z., Garcia-Palacios, E., Duong, T.Q.:
Secure energy harvesting relay networks with unreliable backhaul connections. IEEE
Access 6, 12074–12084 (2018)
7. Yin, C., Nguyen, N.P., Garcia-Palacios, E., Tran, X.N., Le-Tien, T.: Secure energy
harvesting communications with relay selection over nakagami-m fading channels.
Mob. Netw. Appl. 1–8 (2017)
8. Zhang, J., Nguyen, N.P., Zhang, J., Garcia-Palacios, E., Le, N.P.: Impact of primary
networks on the performance of energy harvesting cognitive radio networks. IET
Commun. 10(18), 2559–2566 (2016)
 Low-Cost, Extensible and Open Source
Home Automation Framework

Che Cameron1(B) and Kang Li2

1
Queens University Belfast, University Road, Belfast BT7 1NN, UK
[email protected]
2
University of Leeds, Leeds LS2 9JT, UK

Abstract. The implementation of an IoT application usually depends

on a framework, platform or ecosystem - of which there are many choices
available, both closed- and open-source, each presenting their own vision.
By reviewing a range of different options with respect to features, price
and customisability it is shown that each provider tends to limit the
available opportunities in their platform, either intentionally or uninten-
tionally. Therefore an open-source framework with NodeRED at its core
is developed and presented which imposes minimal constraints upon the
user/developer, encourages novices to learn and safeguards the privacy
of its users.
This approach is verified by building a demonstration system which is
retrofitted to an old property and successfully provides heating system
control and finger print entry. The use of heterogeneous nodes and the
integration of 2 radio technologies serve to underscore the technological
agnosticism and flexibility of this framework.

Keywords: Home automation · IoT · Open-Source · Low cost

NodeRED · MQTT

1 Introduction
The field of IoT has been the topic of much discussion and development over
the past few years, and has arguably suffered from its own hype. The stunning
array of low-cost sensors, actuators and controllers presents a panoply of possible
applications - and in so doing, runs the risk of an over-abundance of choice.
This is reflected in the lack of harmonisation or standardisation throughout the
field [1].
However, some standard tools and techniques have started to emerge which
are helping to increase interoperability between ecosystems and devices [10] -
an important aspect of the interconnected world which is striven for. In this
fluid and constantly developing landscape, a strong strategy for platform devel-
opment focuses upon wide compatibility, technological agnosticism and quick
reconfigurability [6].
Useful technology can be defined by the impact it has upon our lives, and the
home environment presents a prime opportunity for this impact. Some schools of
c Springer Nature Singapore Pte Ltd. 2018
K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 408–418, 2018.
https://doi.org/10.1007/978-981-13-2384-3_38
 Low-Cost, Extensible and Open Source Home Automation Framework 409

architecture consider the home as a machine for living and in this vein, the most
important functions of that machine could be considered as suitable targets for
IoT technologies.
The temperature of a dwelling is one of the most important aspects of com-
fort (and even safety, in extreme conditions) and usually consumes the largest
proportion of the total energy expended in the home - both in empirical terms
and in cost. Central heating control is a natural starting point for many home
IoT systems, which often build out their functionality from this foundation [4].
Proactive energy management of a home plays a large role in setting household
running costs and influences environmental impact - and is greatly assisted and
optimised through the use of smart interconnected systems.
A selection of available IoT platforms will be overviewed (with a focus on
Open-Source) in Sect. 2, then this paper will introduce the framework in Sect. 3,
including the details of the software packages, then go on to describe the demon-
stration system in Sect. 4, with information about the implementation in hard-
ware and software. Section 5 offers some observations and thoughts on the results
of using this framework in a practical setting, and the work is briefly concluded
in Sect. 6.

2 IoT Platforms
The framework that will be presented here is more akin to an architecture than
any individual products and so it is most fruitful to consider the major Home
IoT ecosystems in use today rather than attempt to review the wide range of
individual products available.
Amazon Alexa, Google Home and Apple Homekit, nest products and Sam-
sung Smartthings make up a large segment of the market and give a fair repre-
sentation of the shape of closed-source Home IoT technologies. They are charac-
terised by expensive components, centralised servers for management and beau-
tiful, highly developed interfaces that require minimal effort from the end-user
to set up.
Initially, each company tried their best to create a closed ecosystem with a
monopoly on compatible products, however consumer pressure is forcing them to
gradually open their platforms to more devices [5]. The support that each offers
is still far from universal, ultimately restrictive in what can be accomplished and
largely pre-defines how the user interacts with their system.
Open source solutions are available for IoT and there are many, and ever-
changing choices (as is custom in this arena). These solutions vary from pre-
senting a complete management suite down to aggregation of control panels for
disparate devices/systems. Some closed-source projects also aim to integrate dis-
parate IoT ecosystems and generally monetise this effort by offering enhanced
features for a price.
The general advantages of an open-source approach are transparency of oper-
ation (helping to safeguard security and privacy), the ability to customise the
elements as desired and a reduction in cost. In a less tangible fashion, the com-
munity surrounding an open-source project is often passionate about what they
410 C. Cameron and K. Li

are building and using, which is commonly reflected in the level of help that is
available from developers and peers.
Many of these platforms provide built-in support for a wide range of end
devices and hubs, along with a community that continues to develop plug-ins.
Without a specific application in mind, it is counter-productive to attempt to
definitively rank the different projects - Table 1 shows a comparison of the impor-
tant characteristics and features of a variety of systems.

Table 1. Comparison of IoT platforms

The implementation of rules engines, scripting and programming languages is

often proprietary (or at least non-standard). This means that basic, pre-defined
or anticipated tasks are relatively easy, but a steep learning curve often presents
itself upon any attempt to customise functionality outside of the designated
development path. Therefore any large project is likely to generate significant
amounts of code and effort that cannot be reused elsewhere [9].
Each platform tends to have a particular philosophy or paradigm to cham-
pion, however therein lies a problem - the presupposition of applications, tasks
and functionality. The framework described in Sect. 3 addresses this issue by
connecting generalised tools with an open and flexible approach.

3 Home IoT Framework

3.1 Overview

This framework is composed of a networked hub, sensor/actor nodes and a web

interface suitable for computer or mobile use.
The hub is responsible for node management, data handling (aggregation,
storage and processing) and it integrates a web interface which may be optionally
forwarded to the WAN for remote access. Programming and implementation is
achieved primarily through the flow-based Node-RED software on the hub, and
 Low-Cost, Extensible and Open Source Home Automation Framework 411

inter-component communication is via the MQTT protocol making use of JSON

encoding for data structures.
Physical nodes may be composed of any technologies that the user desires, the
only requirement being that some form of MQTT is supported for integration
with NodeRED. Most modern homes incorporate a wireless network already,
which this system can connect to - alternatively a separate air-gapped network
or different radio technology can be used if desired. The intent is to demonstrate
a low-cost, hardware agnostic approach to Home IoT which trades some ease
of installation for large cost savings, fast development of applications and huge
customisability.
An important note should be made with respect to privacy - all data is
handled and stored locally, owned by the user and never needs to traverse the
internet unless explicitly instructed to do so. With growing awareness of the
dangers posed by careless distribution of personal data [3] (reflected by the new
GDPR regulations), a system which defaults to safeguarding the people who
use it - empowering them with ownership and responsibility - is a step towards
improving our collective attitude towards privacy versus convenience.

Interface Hub

Web GUI

Database

Computer Tablet Phone

Mosquitto Broker

RF24SN nRF24L01+

WiFi LAN

Nodes
WiFi
nRF
2.4GHz

Lolin32 Lite : ESP32 Arduino Nano : Atmel ATmega328P

WiFi, UART, i2c, SPI, GPIO,

ADC, PWM, Capacitive Sense
... UART, i2c, SPI,
GPIO, ADC, PWM
RF24SN nRF24L01+

Battery Sensors Actuators Battery Sensors Actuators

eg. 1200mAh LiPo eg. Temperature eg. Relay eg. 1200mAh LiPo eg. Temperature eg. Relay

Fig. 1. Block diagram of the implemented framework

3.2 Software

The presented cocktail of open-source software forms the essential components of

this framework, providing for simple and fast development along with the ability
to create complex and rich applications if so desired. The suite of software and
412 C. Cameron and K. Li

its interrelations will be explained, along with some of the notable qualities of
each component.
Communication - The nodes and hub communicate using the MQTT protocol
which implements a topic based publish/subscribe model. The hub acts as a
broker (which routes and buffers all network messages) and each client wishing
to communicate connects to this hub. A single message is composed of a topic
and a payload, each of which is a plaintext string of variable length.
The hub makes use of the free and open-source broker mosquitto, and the
only requirement placed upon nodes is that they must be able to communi-
cate via MQTT - however this does not necessarily need to be accomplished
on the node itself as the hub can implement a translation layer (note RF24SN
in Fig. 1). Complex data structures may be transferred using JSON encoding,
which enables rich communication or lightweight messaging without the need for
multiple communication protocols.
This choice of communication protocols and data structures reflects the flex-
ible nature of the framework - using the most established lightweight stan-
dards maximises interoperability and minimises the burden upon hardware which
enables the widest range of devices to be integrated into a synthesised system [7].
Logic - Node management, business logic, data processing and display are all
accomplished via Node-RED - a programming tool originally developed at IBM
that was open-sourced in 2013 and continues to grow in popularity and func-
tionality. This flow-based programming language runs on top of Node.js and
enables unparalleled development speed and reconfigurability. Instead of tradi-
tional code based programming, the user drags and drops nodes on to a canvas
and wires them together create functionality. In the most basic conception, nodes
are classified into Input, Function and Output - with a rich set of features already
included in the base installation.
Program operation is defined by messages which propagate through a given
set of nodes - transformations of that message or actions triggered by it are
defined by the logic that the user has wired into the program. Each set of nodes
may include branches, conditional blocks or message routing to enhance the
complexity of the program. By defining behaviour in this way, sections of a
program naturally decouple and the relationship between different elements can
be clearly visually seen. The created ‘codebase’ is organised into one or more
labelled tabs which are scoped individually.
There is a thriving open-source community constantly contributing to Node-
RED, both with custom nodes and also full fledged programs (known as Flows)
- the integration of which is often single click effort, again underscoring the
ease-of-use for novice users.
This flow based paradigm is becoming an increasing focus in IoT research
due to its quick reconfigurability and ease of use [2] - applications are not static
and IoT implementations are made considerably more responsive and useful by
enabling a larger proportion of the user base to modify their operation without
needing to learn a complex programming language.
 Low-Cost, Extensible and Open Source Home Automation Framework 413

More advanced users can write custom functions in Javascript, create custom
widgets using html/Angular code, group functionality into subflow nodes for easy
re-use, and make completely new nodes with customised functionality. Scoped
variables and objects are also available for buffering and communication of data
within the program.
Node-RED is an excellent match for the requirements of a reconfigurable logic
and data handling framework, and is being increasingly recognised as a powerful
tool [8]. The fact that it is accessible to new users with excellent community
support further enhances its strong position as the heart of a home automation
framework (Fig. 2).

NODES DEPLOY
PALLET (Commit
and Run)

Custom
Subflows

Input DEBUG TAB

Nodes (Show error
messages
and output
of debug
nodes)
Output
Nodes

FLOW
FLOW SWITCHING
CANVAS TABS
(Program
Area)

Fig. 2. Flow based programming in Node-RED

Data Storage - Small sets of data can be stored as JSON encoded text files
by Node-RED, however this strategy quickly becomes unsustainable for a highly
connected home - especially when data analytics are desirable to enhance the
intelligence and utility of an IoT system. Therefore the free and open-source
MySQL relational database is used for data storage.

4 Demonstration System
4.1 Hub
A Raspberry Pi 3 Model B serves as the hub for the system, integrating WiFi
with an additional NRFL4201+ radio for wireless communication. The hub runs
the RF24SN service that implements an MQTT translation layer between the
MQTT Broker and NRFL2401+ radios. A preconfigured Raspberry Pi disk
image minimises set up time.
414 C. Cameron and K. Li

4.2 Nodes

The node types can be divided into three groups - sensors (regularly report
some feature of the environment), actuators (listen for commands from the
hub and then take some action) and actors (perform some action in response
to an event but are not available for command unless externally triggered).
A mix of Arduino Nano and Lolin32 Lite microcontrollers are used to imple-
ment nodes. Battery powered nodes use their onboard ADC to read the battery
voltage and report back to the hub each time they communicate. If the battery
voltage falls below a programmed threshold they will automatically power down
and issue an alert (Fig. 3).

(a) Arduino Nano, NRFL2401+ (b) Lolin32 Lite (ESP32) with battery and
radio and DS18B20 tempera- door latch driver circuitry
ture sensor

Fig. 3. (a) Temperature Sensor node and (b) Finger Print Scanner controller

Definition of Operations:

MEASURE - Measure some variable, eg. temperature, battery voltage.

ACT - Take some action eg. operate a relay.
REPORT - Report some information to the hub eg. sensor reading, node
state, details of an event.
LISTEN - Poll the hub for instructions and act upon any valid reply received,
otherwise take no action eg. OTA Update, Shutdown, Restart
SLEEP - Go to an idle state (wired power) or deep sleep mode (battery
power) for a programmable period of time or until externally interrupted.

Sensors
Order of Operations: MEASURE - REPORT - LISTEN - SLEEP
Sensors are configured to take readings of some quantity of interest, and are
expected to be available to the hub at regular intervals.
 Low-Cost, Extensible and Open Source Home Automation Framework 415

Actuators
Order of Operations: LISTEN - ACT - REPORT - SLEEP
Actuators are expected to be always online, with minimised latency - and there-
fore are generally mains powered. The hub will issue commands to these nodes
and receive a report of the outcome of the node action. In the case of important
systems, the nodes should have suitable fallback/offline behaviour to cater for
network failures.
Actors
Order of Operations: ACT - REPORT - LISTEN - SLEEP
Actors are distinct from actuators because the hub does not command them
directly - there is no requirement for a minimum contact interval. They are semi-
autonomous devices whose functionality is enhanced by a server connection but
not dependent upon it.

4.3 System Functionality

Data Acquisition and Storage - Data from the nodes is reported at regu-
lar intervals via MQTT messages. Upon receipt in Node-RED, the message is
timestamped and its topic set to an appropriate label. If the data is a sensor
reading then a MySQL query is generated and the information is inserted into
the appropriate database.
A NODESTATES object tracks the status of all nodes in the system (includ-
ing battery voltage, last transmission and online/offline status), and is updated
upon receipt of information from each component. Instantaneous or otherwise
non-persistent data can be delivered directly to output nodes (charts, gauges,
text) and will not bloat the database with undesired information.
Data Processing - The system calculates a windowed derivative to estimate
the rate of heating or cooling for each reported location, which can be displayed
as a graph or as an instantaneous estimate.
Hysteresis reporting is also used to remove noise and smooth the graph of
historical temperature data (which is queried from the MySQL database at a
regular interval).
These are simple examples and it should be noted that NodeRED and
MySQL can integrate with external tools for more advanced analytics.
System Operations
Node Error Identification:
The NODESTATES object is interrogated on a regular basis to set a flag for
radio errors, low voltage and offline state. A human readable string is produced
to alert the user that some corrective action should be taken.
Heating Controller:
A multi-room thermostat regulates the heat (using customisable width hystere-
sis) based upon the lowest reported temperature in a user defined set of rooms
and thus ensures a comfortable temperature in all the living areas.
416 C. Cameron and K. Li

The integration to the boiler is via replacing a traditional ‘heating clock’

with a set of relays - this approach would also extend to more complex ported
valve heating systems.
Access Control via Finger Print Scanner and Electric Door Latch:
A finger print scanner is affixed to the outside of the house, whilst the node
is mounted in the hallway. It holds up to 20 IDs and is capable of operating
autonomously after users have been enrolled via the webapp.
Interface - The lowest bar for IoT in the home is that it is not more difficult than
the traditional technology that it replaces. Where it is advantageous, individual
nodes should provide feedback without the need to access anything on a tablet,
smartphone or computer - for example, in this system the heating controller
activates LEDs when the relays are on and the access control node incorporates
a piezo buzzer to inform the user of the lock state.
With that stipulation being met, an IoT system should then go on to offer
enhanced information, insightful knowledge and new functionality to the user.
Node-RED provides an excellent set of user interface and data display widgets
that can be used to quickly build an interface that is suitable for all major mobile
and desktop platforms - without the need to write any code.
The demonstration system is organised into tabs, two of which are shown;
Home (Fig. 4), the default display which details overall status, 2 day historical
temperature and live readings, and Management (Fig. 5) which offers a manage-
ment interface for the heating system and communications. There are additional
tabs showing 7 day historical temperature, access logs and detailed node status.

Error
Indicator
Gauges
Showing
Rate of
Change
Temperature
(ºC per Hour)
History
for Each
(2 Days)
Node

Current
Temperature, Gauges
Setpoint and Showing
Estimated Current
Time to Temperature
Reach Target of Each
Node

Fig. 4. Web interface - home tab

 Low-Cost, Extensible and Open Source Home Automation Framework 417

ON/OFF
Central
Table
Heating
showing
System
time since
last
Setpoint transmission
Adjustment (in minutes)
from each
System node
Mode

Hysteresis:
Heating or
Cooldown

Control
Reference
Room Reset
NRF Radio
manually
Demand
and State of
Boiler
Relays

Fig. 5. Web interface - management tab

5 Discussion
The low cost NRFL2401+ 2.4 GHz radio system from Nordic Semiconductor does
not support WiFi or encryption and its use here should not be considered an
endorsement of this outdated technology which has largely been supplanted by
low cost devices integrating 802.11 WiFi. Espressif has made a large impact with
their low-cost WiFi microcontrollers, firstly with the ESP8266, followed by the
upgraded ESP32. The SoCs can be procured for less than £2 and development
boards for less than £5 - this price point makes the aforementioned NRF24L01+
practically obsolete for IoT applications and due to the fact that it must be paired
with a microcontroller, the total cost is usually higher than a single ESP SoC.
Examining the collected time series of temperatures (along with knowledge
of the house activities) has made it clear that rich knowledge is available from
this data - it is possible to easily see when the shower has been used, to deter-
mine if doors between rooms are open or closed, the thermal characteristics and
performance of individual rooms are readily apparent and faults in the heat-
ing system leave unique features. The potential for intelligent behaviour when
more advanced analytics are applied is quite significant - additional sensors and
measurands will only increase this. Minimising the fuel used for heating through
smarter scheduling and recommendations for targeted insulation efforts (iden-
tified by modelling the performance of the various house areas) could make a
significant impact on energy bills.

6 Conclusion
A framework based upon flexibility, wide compatibility and quick reconfiguration
was proposed for Home IoT applications. Using open-source software through-
out reduces cost and increases transparency of each component, safeguarding
personal data.
418 C. Cameron and K. Li

Due to the changeable and unpredictable nature of IoT, the framework does
not seek to impose limitations or anticipate end uses - instead it offers a gener-
alised set of tools to easily accomplish tasks and respond to changing application
requirements.
A demonstration system which has improved the performance of the home in
which it was installed has been described, and some basic real-time analysis of the
data has been incorporated without the need for any third party tools. Informal
observation of the data suggests that there is significant scope for knowledge
extraction from low-grade sensor information and with more advanced analytics
it would be possible to make suggestions for physical improvements that target
heating efficiency, as well as automatically minimising the fuel required.

Acknowledgement. Kind acknowledgements to the UK-China University Consor-

tium on Engineering Education and Research for their support.

References
1. Bandyopadhyay, D., Sen, J.: Internet of things: applications and challenges in tech-
nology and standardization. Wirel. Pers. Commun. 58(1), 49–69 (2011)
2. Belsa, A., Sarabia-Jácome, D., Esteve, M.: Flow-based programming interoperabil-
ity solution for IoT Platform Applications (2018)
3. Connolly, A.: Freedom of Encryption. Real-World Crypto, pp. 102–103 (2018)
4. Dharur, S., Hota, C., Swaminathan, K.: Energy efficient IoT framework for Smart
Buildings, pp. 793–800 (2017)
5. Leminen, S., Westerlund, M., Rajahonka, M., Siuruainen, R.: Towards IOT ecosys-
tems and business models. In: Andreev, S., Balandin, S., Koucheryavy, Y. (eds.)
NEW2AN/ruSMART -2012. LNCS, vol. 7469, pp. 15–26. Springer, Heidelberg
(2012). https://doi.org/10.1007/978-3-642-32686-8 2
6. Lu, C.H.: Improving system extensibility via an IoT-interoperable platform for
dynamic smart homes. In: Proceedings of the 2017 IEEE International Conference
on Applied System Innovation: Applied System Innovation for Modern Technology,
ICASI 2017, pp. 1300–1303 (2017)
7. Naik, N.: Choice of effective messaging protocols for IoT systems: MQTT, CoAP,
AMQP and HTTP. In: 2017 IEEE International Symposium on Systems Engineer-
ing, ISSE 2017 - Proceedings (2017)
8. Rajalakshmi, A., Shahnasser, H.: Internet of Things using Node-Red and alexa.
In: 2017 17th International Symposium on Communications and Information Tech-
nologies (ISCIT), pp. 1–4 (2017). http://ieeexplore.ieee.org/document/8261194/
9. Ruiz, M.C., Mcarmenruizuclmes, E., Olivares, T., Teresaolivaresuclmes, E., Lopez,
J.: Evaluation of Cloud Platforms for Managing IoT Devices (2017)
10. Zitnik, S., Jankovic, M., Petrovcic, K., Bajec, M.: Architecture of standard-based,
interoperable and extensible IoT platform. In: 24th Telecommunications Forum,
TELFOR 2016, pp. 1–4 (2017)
 Spectrum Utilization of Cognitive Radio
in Industrial Wireless Sensor
Networks - A Review

Mingjia Yin1(B) , Kang Li1 , and Min Zheng2

1
School of Electronic and Electrical Engineering,
University of Leeds, Leeds LS2 9 JT, UK
{elmyi,k.li1}@leeds.ac.uk
2
School of Mechatronic Engineering and Automation,
Shanghai University, Shanghai, China
[email protected]

Abstract. The increasing demand for intelligent control and automa-

tion in industry requires better use of the radio spectrum due to the use
of industrial wireless sensor networks (IWSNs). Cognitive Radio (CR)
is a promising technology to improve the spectrum utilization by sens-
ing spectrum holes. Research in this area is still in its infancy, but it
is progressing rapidly. In this paper, industrial environment with dif-
ferent wireless technology, such as WirelessHART and ISA 100.11a is
investigated. Various sensing schemes and the challenges associated for
the cognitive radio are reviewed. In addition, the paper discussed the
methods relevant to industrial applications, covering architecture, spec-
trum access, interference management, spectrum sensing and spectrum
sharing.

Keywords: IWSN · Cognitive radio · Spectrum sensing

Spectrum utilization

1 Introduction
In the current information-intensive society, the development of distributed mon-
itoring and control system and industrial and factory automation is featured by
the characteristics of flexibility, integration, robotization and intelligentization.
Industrial applications have no longer been confined to closed plant environ-
ment, the networks tend to coexist with other industrial wireless systems and
existing commercial wireless systems [1]. Given the growing number of inter-
connected industrial systems and devices for agile industrial manufacturing,
industrial wireless sensor networks (IWSNs) are playing an increasingly more
important role. In an industrial environment, wireless sensor nodes are installed
on on-site equipment and used to monitor diverse parameters such as tempera-
ture, pressure, humidity, location and vibration [2]. The advantage of IWSNs are
appealing over traditional wired communication systems such as cost-efficiency,
c Springer Nature Singapore Pte Ltd. 2018
K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 419–428, 2018.
https://doi.org/10.1007/978-981-13-2384-3_39
420 M. Yin et al.

self-organization, scalability and mobility [3]. It has been seen as a vital com-
ponent in the Industry 4.0 framework, and can be used for smart factories,
networked manufacturing and industrial internet of things.
Current spectrum allocation mechanism assigns wireless spectrum to licensed
operators based on the static spectrum allocation policy. The cost of the license
is very high, however, only parts of the spectrum is utilized effectively. Most
of the distributive spectrum is utilized intermittently. According to report of
Federal Communications Commission (FCC), the current utilization of a licensed
spectrum varies from 15% to 85% [4]. The inefficient use and spectrum scarcity
definitely hinder the development of wireless communication systems. The rest
of industrial and home applications have to compete for the even more crowded
free Industrial Scientific Medical (ISM) bands not only with each other but also
with other wireless communication devices in the same area.
Cognitive Radio (CR) technology is recognized as an emerging technology
to solve the dilemma spectrum utilization is facing: the contradiction between
rapid-growing number of wireless devices and the scarcity of available spectrum
resources. The definition of CR by FCC is “Cognitive radio: A radio or system
that senses its operational electromagnetic environment and can dynamically and
autonomously adjust its radio operating parameters to modify system operation,
such as maximize throughput, mitigate interference, facilitate interoperability,
access secondary markets [5].” In this paradigm, sensor nodes can sense the idle
frequency bands (namely, the spectrum holes) and transmit data by re-using
the frequency bands occupied by the authorized users. Thus, the utilization of
existing spectrum resources is enhanced. The authorized users are recognized as
primary users (PUs) while sensors nodes are secondary users (SUs). SUs carry
radio devices that have the ability of scan and sense of the surrounding spec-
trum utilization. Once spectrum holes found, SUs can access the idle spectrum
opportunistically by altering their spectrum bands subsequently [6].
In this paper, we start with a discussion on the challenges of industrial envi-
ronments, focusing on different interference sources, coexistence, and heteroge-
neous environment. Next, an overview of spectrum sensing techniques is given.
This is then followed by a presentation on a range of methods relative to spec-
trum utilization. Finally, we discuss some open problems and future trends of
IWSNs.

2 Challenges of Wireless Industrial Network

2.1 Interferences in Wireless Industrial Networks

Industrial environment is often more complicated than public and private envi-
ronments. It has higher quality of service (QoS) requirements than that applica-
tions at homes and offices. Reliability, latency, and availability are some major
aspects and they can be quite specific for different applications. For example,
monitoring system is usually time-sensitive. Data with long latency may lead to
wrong decisions.
 Spectrum Utilization of Cognitive Radio in IWSN - A Review 421

The signal strength in the industrial environment is heavily affected by mul-

tipath fading (e.g., the reflections from floors and walls), the interferences from
other devices (e.g., motors), and noise. Specifically, interference signals can be
divided into two classes: broadband or narrowband [7]. Broadband interferences
usually come from unintentional radiating sources. They have high energy and
constant energy spectrum across a large range of frequencies whereas narrow-
band interference signals have lower energy at specific frequencies.
In a typical industrial site, there are sufficient wireless communication sys-
tems and networks that might generate interference against the radio signals [8].
The co-existence issues between various wireless networks also need to be consid-
ered. The deployed networks include wireless local area networks (WLANs) and
wireless personal area networks (WPANs) which are short range networks. The
common protocols used in industrial environment like Zigbee, bluetooth, Wi-Fi
(Wi-Fi also works in the 5 GHz band), WirelessHART, ISA 100.11a and wireless
interface for sensors and actuators (WISA) are operating at 2.4 GHz band. The
emerging low power wide area network (LPWAN) allows small amounts of data
to be transmitted within a long distance at lower frequencies. Table 1 presents
the key characteristics and make a comparison among different platforms. Zig-
bee, WirelessHART and ISA 100.11a use the same physical layer and media
access control protocol (IEEE 802.15.4). The frequency spectrum is divided into
16 channels at 2.4 GHz with each channel having a bandwidth of 2 MHz. IEEE
802.11 is adopted by Wi-Fi which also works at 2.4 GHz and normally has 13
channels. The bandwidth of each channel is 20 MHz and most routers choose
channel 1,6 and 11 because they hardly overlap.

Table 1. Wireless industrial communication protocols

Zigbee Wi-Fi WirelessHART ISA 100.11a LoRa

Bandwidth 2 MHz 40 MHz, 20 MHz 2 MHz 2 MHz 125 KHz
Frequency 2.4 GHz 2.4 GHz 2.4 GHz 2.4 GHz 433,915 MHz
Channels 16 13 16 16 32
Radio 802.15.4 802.11a, b, g 802.15.4 802.15.4 Proprietary
Data rate 250 kbps 11–54 Mbps 250 kbps 250 kbps 10 kbps
Range 10–100 m 1–100 m 50–100 m 50–100 m 50 m–10 km

In order to reduce the effect of interference, spectrum allocation without CR

technology chooses non-overlapping channels. Figure 1 illustrates the channel
distribution and highlight the non-overlapping bands between IEEE 802.15.4 and
IEEE 802.11. In [9], a multi-dimensional resource allocation strategy (MRAS) is
used to meet the different industrial requirements. It can improve the utilization
of available wireless network resources, but the improvement is not significant
enough compared to its complexity.
422 M. Yin et al.

Fig. 1. The channel distribution of IEEE 802.15.4 and IEEE 802.11

2.2 Complex Heterogeneous Environment

Traditionally, the applications of the IWSN systems can be classified into three
categories: safety systems, control systems, and monitoring systems [10]. In these
systems, a wide range of data are collected or processed over a given region for
a long duration and the data are exploited thoroughly to make certain conclu-
sions. In recent years, the data sources are more abundant which may include
on-demand or live video streaming, audio, and still images. These multimedia
applications set high demands for transmission rate and bandwidth. Moreover,
in a dense network environment such as indoor plants, the media control access
mechanism should be carefully designed because different sensor nodes in the
area are likely to access a channel simultaneously. In addition, the topology of
the network and connection of the nodes change over time due to the link fail-
ures and battery depletion. The idea of heterogeneous sensor networks thus come
naturally with two of more nodes working with different energy and hardware
complexity.

3 Spectrum Sensing Technologies

In Sect. 2, the difficulty of spectrum scarcity is discussed and the deployment
problem of several objects connected to infrastructure through radio links is
introduced. Recently, integration of CR technology in wireless sensor nodes has
gained much attention as it enables sensors transmit data packets over the
licensed spectrum bands as well as the free ISM bands. Spectrum sensing is
a critical component for CR. Through spectrum sensing, CR can obtain neces-
sary observations about its surrounding radio environment, such as the presence
of PUs and appearance of spectrum holes. Basically, there are two types of sens-
ing techniques: signal processing techniques and cooperative sensing techniques.
Figure 2 shows the classification of spectrum-sensing techniques.

3.1 Signal Processing Techniques

(1) Matched Filter Detection: The matched filter (also known as coherent detec-
tor), is considered as the optimum method because of its accuracy. As the
 Spectrum Utilization of Cognitive Radio in IWSN - A Review 423

Fig. 2. Classification of spectrum-sensing techniques

parameters of the transmitted signals are known, the matched filter accord-
ingly maximizes the received signal-to-noise ratio (SNR). The presence or
absence of PU is determined by the comparison between the result of corre-
lation and a predetermined threshold. However, matched-filtering also has
limitations. It requests the features of PU signal in advance, which include
bandwidth, operating frequency, modulation type and order, pulse shap-
ing, and frame format [11]. Also, the computational complexity and power
consumption are very high.
(2) Energy Detection: Energy detection is performed in both time and frequency
domain. It compares the energy received by PU signal with a threshold based
on the noise estimation. The method is simple and has low computational
cost because it does not need a priori information of PUs signal. But, the
drawback of the method is obvious as well. The noise needs to be accurate
which is difficult to estimate [12]. It is also not suitable to use the method
in low SNR environment and for detecting spread spectrum signals.
(3) Cyclostationary Feature Detection: The cyclostationary feature detector dif-
ferentiates the PU signal from the additive noise since the method deals with
the inherent cyclostationary properties or features of a signal. A cyclosta-
tionary signal is said to be periodic when considering the mean and auto-
correlation [13]. The technique is simple and effective. Only one or two
periodic features of the PU signal is needed. It can be used to detect at very
low SNR condition. The robustness to noise of the method performs better
than energy detector, nevertheless, the computational complexity is higher
than that of energy detector.

3.2 Cooperative Sensing

Cooperative sensing literally means that the sensor nodes with CR can cooper-
ate and exchange their sensed information in different manners. It can improve
sensing accuracy by exploiting spatial diversity.
(1) Centralized Cooperation: In centralized cooperation, a central cooperator,
which can be a collector, or server, is used to manage the information from
424 M. Yin et al.

distributed CR sensor nodes and determine the absence or presence of the

PU. Firstly, each sensor node with CR senses the surrounding spectrum by
themselves and distinguish the presence of the PU from the absence at a
specific frequency. Secondly, the nodes forward the results to the central
cooperator. At last, the central cooperator aggregates and organize data
together. Once a sensor node requests for channel information, the cooper-
ator can transmit their final decision after data fusion.
(2) Decentralized Cooperation: The decentralized cooperation works in a decen-
tralized manner with no central cooperator. Instead, it uses a spectrum
information table to record the sensing results. Each node performs local
spectrum sensing and make a decision as to the presence or absence of the
PU. Then the nodes spread the information to neighbouring ones and cre-
ate spectrum information table in a predicted period. The scheme is more
flexible, but the storage and computation cost are high due to the updated
table [14].
(3) Hybrid Cooperation: The hybrid cooperation combines the advantages of
centralized cooperation and decentralized cooperation. Each sensor node
sense and transmit information in the same way as decentralized manner and
it may share channel information with centralized cooperator as requested.

4 Spectrum Utilization

The cognitive capability is a promising solution to enhance the spectrum sens-

ing of dynamic environment by adjusting the parameters used to determine the
transmitted signal. Meanwhile, it provides an elastic feature, known as reconfig-
urability, that is suited to the changing parameters in a dynamic radio environ-
ment without modifying the hardware structure. In this section, many techniques
that can be used to improve the spectrum efficiency are introduced. The network
architecture, spectrum access, interference management, spectrum sensing and
spectrum sharing should always be considered together.

4.1 Clustering

The complexity of the architecture can be reduced by forming clusters with

neighbouring nodes. The physically nearby sensor nodes are grouped into single-
hop clusters with a star topology. Each cluster elects a cluster head (CH) to
perform the intra- and inter-cluster management and data processing. CHs then
forward their data to sinks or base stations (BS) which are supposed to have
higher energy and stronger processing capabilities.
The existing clustering algorithms [15–18] consider homogeneous sensor
nodes, and in [19] a cluster-based approached is formed for a meshed network.
Within the cluster, the CH is responsible for the spectrum management. In
[20], a distributed spectrum-aware clustering (DSAC) protocol is proposed. The
simulation results show the preferable scalability and stability under dynamic
spectrum access.
 Spectrum Utilization of Cognitive Radio in IWSN - A Review 425

4.2 Hierarchical Topology

A hierarchical network allows nodes at different tiers to perform different tasks.
Among them, a typical example is the two-tier structure. The lower tier is respon-
sible for sophiscated task such as monitoring nodes without CR. In the upper
tier, it comprises a set of SUs opportunistically access the spectrum to send
the detection results to the sink node. In [21], the author incorporated two con-
strained Markov Decision Processes (MDP) with the two-tier network for event
detection and channel allocation. One MDP is used to adjust the detection cri-
terion for the delay constraint. Meanwhile, the other MDP is used for optimum
spectrum allocation of SUs.

4.3 Fuzzy Cognitive Map

When multiple objectives, environment variables, and processes are considered,
fuzzy cognitive map (FCM) is a simple way to present the system in a param-
eterized and directed form. The goals can be expressed as causal relationships
which are in low complexity under conflicting constraints [22]. FCM is a unsu-
pervised model which can be viewed as a intelligent tool. The authors of [23]
present a cognitive sensor network framework for highway safety applications.
FCM is used to address the challenges of highway safety, network lifetime and
throughput. The Q-learning algorithm is also incorporated in the system so as
to enhance the learning ability of the FCM after evaluation and improve the
system performance by a well-designed reward system.

4.4 Optimization
The challenges of power and hardware limitations of sensor nodes should be over-
looked as its inherent characteristics. Combining CR functionalities with sensor
nodes is a self-evident choice when energy and network lifetime is co-considered.
The existing networks and protocols are seldom aware of cognitive capability,
many open optimization issues are left for research. For example, authors in [24]
investigated the optimal packet size for a cognitive wireless sensor network. It
defines the system and gives the simulation result of the optimal packet size for
the proposed system with maximum energy efficiency while maintaining reason-
able interference level for the licensed PU. In [25], a multi-objective optimization
framework is proposed for reducing CO2 emissions. The cooperative cognitive
system jointly optimizes the relay assignment and power allocation with two
conflicting objectives which are to maximize the sum-capacity and to minimize
the CO2 emissions. Although a hybrid evolutionary scheme of EDA is proposed,
the real-time performance is poor due to the priori knowledge of the distribution
of relay nodes and the constraints of greedy algorithm.

4.5 Random Processes

A random process is often referred to a whole set of random variables that depend
on the parameters. The parameter is usually time. Random variables are the
426 M. Yin et al.

mathematical models of random phenomena, and their values will change with
the affecting factors. [26] proposes a hybrid sensing method using a continuous
time Markov chain model to adjust the parameters in avoid of contradictory
factors. According to the characteristics of PU and SUs, the best sensing period
is found. Considering that sensing results are affected by uncertain noise and
contradictory factors, [27] introduces the partially observable Markov decision
processes to build the channel model. It enables the cognitive nodes with the
ability of channel sensing, channel switching, and data transmission.

4.6 Game Theory

Game theory mainly studies the interaction between the formulated incentive
structure, which is a mathematical theory and method for studying the phe-
nomena of competition. It considers the predicted behavior and actual behavior
of individuals and studies their optimization strategies [28]. Cognitive radio net-
work is a good instance of game theory as the SUs are competing for spectrum
usage based on the sensed spectrum and actions of other users. Non-cooperative
game theory can be used for spectrum sharing schemes since the SUs are assumed
to be selfish who only care about their own spectrum. [29] analyzes how to max-
imize the benefits of cognitive users (i.e., maximize spectrum utilization). In the
premise of maintaining the benefits of the authorized users, the choice of cog-
nitive nodes are evaluated by the revenue function. On this basis, the decision-
making interactions among cognitive nodes are formulated as a non-cooperative
game process and Nash equilibrium (NE) corresponding to a stable decision is
obtained. The proposed mechanism is shown to have a relatively low complexity.
Auction mechanisms are also widely used for spectrum sharing. In [30], the spec-
trum usage is defined as a function of SNR. [31] proposes a real-time spectrum
auction framework. In the mechanism, spectrum channels are assigned to proper
SUs under interference constraints.

5 Conclusion
In this paper, the potential benefits and current studies of using CR technology
to improve the spectrum utilization in IWSN have been discussed. We give a brief
introduction of the CR technology and its different sensing techniques, namely
signal processing and cooperative sensing techniques. Further, the potential tech-
niques from the aspects of network architecture, spectrum access, interference
management, spectrum sensing and spectrum sharing are presented. However,
there are still some open problems in this area. One thing should be noted is
that there is no universal test platform and test environment to implement the
approaches, evaluate the cooperation mechanism and verify the accuracy. Hence,
standardized protocols and practical methods need to be proposed.

Acknowledgments. This work was financially supported by UK EPSRC under the

Optimising Energy Management in Industry - ‘OPTEMIN’ project EP/P004636/1. M.
YIN would like to thank the EPSRC for sponsoring her research (project reference
1951147).
 Spectrum Utilization of Cognitive Radio in IWSN - A Review 427

References
1. Chiwewe, T.M., Mbuya, C.F., Hancke, G.P.: Using cognitive radio for interference-
resistant industrial wireless sensor networks: an overview. IEEE Trans. Ind. Inform.
11(6), 1466–1481 (2015)
2. Gungor, V.C., Hancke, G.P.: Industrial wireless sensor networks: challenges, design
principles, and technical approaches. IEEE Trans. Ind. Electron. 56(10), 4258–4265
(2009)
3. Cena, G., Seno, L., Valenzano, A., Zunino, C.: On the performance of ieee 802.11E
wireless infrastructures for soft-real-time industrial applications. IEEE Trans. Ind.
Inform. 6(3), 425–437 (2010)
4. Akyildiz, I.F., Lee, W.Y., Vuran, M.C., Mohanty, S.: A survey on spectrum man-
agement in cognitive radio networks. IEEE Commun. Mag. 46(4), 40–48 (2008)
5. Federal Communications Commission: Notice of proposed rule making and order:
facilitating opportunities for exible, efcient, and reliable spectrum use employing
cognitive radio technologies. ET Docket, February 2005
6. Joshi, G.P., Nam, S.Y., Kim, S.W.: Cognitive radio wireless sensor networks: appli-
cations, challenges and research trends. Sensors 13(9), 11196–11228 (2013). http://
www.mdpi.com/1424-8220/13/9/11196
7. Low, K.S., Win, W.N.N., Er, M.J.: Wireless sensor networks for industrial environ-
ments. In: International Conference on Computational Intelligence for Modelling,
Control and Automation and International Conference on Intelligent Agents, Web
Technologies and Internet Commerce (CIMCA-IAWTIC 2006), vol. 2, pp. 271–276,
November 2005
8. Werb, J., Newman, M., Berry, V., Lamb, S., Sexton, D., Lapinski, M.: Improved
quality of service in IEEE 802.15.4 mesh networks. In: Proceedings of International
Workshop on Wireless and Industrial Automation, pp. 1–4 (2005)
9. Qu, F., Zhang, J., Shao, Z., Qi, S.: Research on resource allocation strategy of
industrial wireless heterogeneous network based on IEEE 802.11 and IEEE 802.15.4
protocol. In: 2017 3rd IEEE International Conference on Computer and Commu-
nications (ICCC), pp. 589–594, December 2017
10. Somappa, A.A.K., Ovsthus, K., Kristensen, L.M.: An industrial perspective on
wireless sensor networks x2014; a survey of requirements, protocols, and challenges.
IEEE Commun. Surv. Tutor. 16(3), 1391–1412 (2014)
11. Zhang, X., Chai, R., Gao, F.: Matched filter based spectrum sensing and power
level detection for cognitive radio network. In: 2014 IEEE Global Conference on
Signal and Information Processing (GlobalSIP), pp. 1267–1270, December 2014
12. Xuping, Z., Jianguo, P.: Energy-detection based spectrum sensing for cogni-
tive radio. In: 2007 IET Conference on Wireless, Mobile and Sensor Networks
(CCWMSN07), pp. 944–947, December 2007
13. Aparna, P.S., Jayasheela, M.: Cyclostationary feature detection in cognitive radio
using different modulation schemes (2012)
14. Bera, D., Maheshwari, S., Chakrabarti, I., Pathak, S.S.: Decentralized coopera-
tive spectrum sensing in cognitive radio without fusion centre. In: 2014 Twentieth
National Conference on Communications (NCC), pp. 1–5, February 2014
15. Pritom, M.M.A., Sarker, S., Razzaque, M.A., Hassan, M.M., Hossain, M.A., Ale-
laiwi, A.: A multiconstrained QoS aware MAC protocol for cluster-based cognitive
radio sensor networks. Int. J. Distrib. Sens. Netw. 11(5), 262871 (2015). https://
doi.org/10.1155/2015/262871
428 M. Yin et al.

16. Rauniyar, A., Shin, S.Y.: A novel energy-efficient clustering based cooperative spec-
trum sensing for cognitive radio sensor networks. Int. J. Distrib. Sen. Netw. 11(6),
198456 (2015). https://doi.org/10.1155/2015/198456
17. Singhal, D., Barjatiya, S., Ramamurthy, G.: A novel network architecture for cogni-
tive wireless sensor network. In: 2011 International Conference on Signal Process-
ing, Communication, Computing and Networking Technologies, pp. 76–80, July
2011
18. Li, X., Wang, D., McNair, J., Chen, J.: Residual energy aware channel assignment
in cognitive radio sensor networks. In: 2011 IEEE Wireless Communications and
Networking Conference, pp. 398–403, March 2011
19. Chen, T., Zhang, H., Maggio, G.M., Chlamtac, I.: Topology management in
CogMesh: a cluster-based cognitive radio mesh network. In: 2007 IEEE Interna-
tional Conference on Communications, pp. 6516–6521, June 2007
20. Zhang, H., Zhang, Z., Dai, H., Yin, R., Chen, X.: Distributed spectrum-aware
clustering in cognitive radio sensor networks. In: 2011 IEEE Global Telecommuni-
cations Conference - GLOBECOM 2011, pp. 1–6, December 2011
21. Jamal, A., Tham, C.K., Wong, W.C.: Event detection and channel allocation in
cognitive radio sensor networks. In: 2012 IEEE International Conference on Com-
munication Systems (ICCS), pp. 157–161, November 2012
22. Mougy, A.E., Ibnkahla, M.: Achieving end-to-end goals of WSN using weighted
cognitive maps. In: 37th Annual IEEE Conference on Local Computer Networks,
pp. 328–331, October 2012
23. El Mougy, A., Ibnkahla, M.: A cognitive WSN framework for highway safety based
on weighted cognitive maps and q-learning. In: Proceedings of the Second ACM
International Symposium on Design and Analysis of Intelligent Vehicular Networks
and Applications, DIVANet 2012, pp. 55–62. ACM, New York (2012). https://doi.
org/10.1145/2386958.2386967
24. Oto, M.C., Akan, O.B.: Energy-efficient packet size optimization for cognitive radio
sensor networks. IEEE Trans. Wirel. Commun. 11(4), 1544–1553 (2012)
25. Naeem, M., Ashrafinia, S., Lee, D.: Estimation of distribution algorithm for green
resource allocation in cognitive radio systems. In: 2012 6th International Confer-
ence on Signal Processing and Communication Systems, pp. 1–7, December 2012
26. Khattab, A., Perkins, D., Bayoumi, M.A.: Opportunistic spectrum access: from
theory to practice. IEEE Veh. Technol. Mag. 7(2), 62–68 (2012)
27. Han, J.A., Jeon, W.S., Jeong, D.G.: Energy-efficient channel management scheme
for cognitive radio sensor networks. IEEE Trans. Veh. Technol. 60(4), 1905–1910
(2011)
28. Wang, B., Wu, Y., Liu, K.R.: Game theory for cognitive radio networks: an
overview. Comput. Netw. 54(14), 2537–2561 (2010). http://www.sciencedirect.
com/science/article/pii/S1389128610001064
29. Yuan, W., Leung, H., Chen, S., Cheng, W.: A distributed sensor selection mech-
anism for cooperative spectrum sensing. IEEE Trans. Sig. Process. 59(12), 6033–
6044 (2011)
30. Huang, J., Berry, R.A., Honig, M.L.: Auction-based spectrum sharing. Mob. Netw.
Appl. 11(3), 405–418 (2006). https://doi.org/10.1007/s11036-006-5192-y
31. Gandhi, S., Buragohain, C., Cao, L., Zheng, H., Suri, S.: A general framework for
wireless spectrum auctions. In: 2007 2nd IEEE International Symposium on New
Frontiers in Dynamic Spectrum Access Networks, pp. 22–33, April 2007
 A Network Coding Against Wiretapping
Attacks of the Physical Layer Security Based
on LDPC Code

Yujie Zheng(&) and Jingqi Fu

Department of Automation, College of Mechatronics Engineering and

Automation, Shanghai University, No.149, Yanchang Rd., Shanghai 200072,
China
[email protected]

Abstract. The secure transmission of wireless network information has

received wide attention. To solve Hacking threat problems that exist in the
wireless network, this paper presented a construction method which is based on
the classic Low-Density Parity-Check (LDPC) encoding method, the transmis-
sion model of the physical security was studied, and the safe transmission
scheme was developed under the gaussian eavesdropping channel, the code with
a nested structure was designed, and the random selection of code words was
determined, message processing is carried out through an improved large-
column weight and low complexity check matrix, not only effectively reduce the
requirement for storage space and the complexity of the information processing,
but also effectively decrease the error rate of the system. Simulation experiments
and comparative analysis show that the bit error rate (BER) of the encoding
method in this paper is significantly lower than that of Progressive Edge Growth
(PEG) algorithm, which achieves the anti-eavesdropping effect.

Keywords: LDPC
Physical layer security
Network coding

Eavesdropping

1 Introduction

With the rapid development of wireless communication, more and more attention was
paid to how to realize the safe transmission of information. As the lowest layer of
communication network, the physical layer has attracted attention for its security
problems due to the eavesdropping hazards. In 2010, Wyner [1] proposed the eaves-
dropping channel model.
The traditional network security problem relies on the method of cryptography, but
there are some disadvantages such as large amount of computation. Therefore, the
security of information theory has attracted people’s attention. Security in the infor-
mation theory sense is security in the absolute sense. In the past, it was generally
believed that security in the sense of information theory could not be realized, so,
security in the sense of information theory was not considered in the actual commu-
nication system. Shannon has pointed out that when the quality of the eavesdropping
channel is worse than that of the legitimate user channel, the physical layer security can
© Springer Nature Singapore Pte Ltd. 2018
K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 429–440, 2018.
https://doi.org/10.1007/978-981-13-2384-3_40
430 Y. Zheng and J. Fu

be realized by certain coding methods [2, 3]. The LDPC code can approach Shannon’s
limit and realize the security of information theory. Network Coding (NC) came into
being in 2000, which was proposed by Dr. Ahlswede [5] from Hong Kong based on
Shannon’s theory of “maximum flow and minimum cut”, it was proved theoretically
that nodes in the network can complete the storage and forwarding function in the
traditional relay network, it can also carry out some coding processing on the trans-
mitted signals to achieve the theoretical maximum of network multicast, such as
exclusive OR (XOR) in the finite domain. In 2002, Cai and Yeung [6] first studied
security network coding, it was assumed that in a single-source network, an eaves-
dropper can eavesdrop on a subset of a channel that the sender and receiver do not
know about. This network code is considered secure when the eavesdropper does not
receive any information transmitted over the network through the eavesdropper. In the
literature, the wire-tap model (CSWN) is proposed, this model is aimed at eaves-
droppers with certain eavesdropper capability, when the number of eavesdropper
channels is less than a certain value, no useful information of source information can be
obtained, this model has been proved to be safe in information theory. In 2006, the
application of network coding ideas was extended to the physical layer, Shengli Zhang
from Chinese university of Hong Kong proposed the concept of Physical layer network
coding (PNC) in literature [4]. The essence is to make full use of the superposition
characteristic of radio waves to improve the throughput of the network. The specific
method is to make use of the superposition characteristic of radio waves, and then carry
out XOR coding processing on the signal that needs to be relayed, finally, send it to the
destination node.
The Progressive Edge Growth (PEG) algorithm is a greedy algorithm. Currently,
the most commonly used LDPC construction method is PEG algorithm proposed by
Hu et al. [11] it’s a classic random construction method, it is considered to be the
optimal algorithm in the design of short and medium code. The method presented in
this paper is based on the improved PEG algorithm and a structured construction
approach is used, it’s made up of a circulant replacement unit matrix, the matrix is
sparser, the constructed check matrix is not limited by column weight and can effec-
tively reduce the demand for memory space, the matrix has no short loop and the check
part matrix has the structure of quasi dual-diagonal lines, it has the characteristics of
fast coding.

2 The Basic Principles of Physical Layer Network Coding

The core idea of physical layer network coding is under the premise of strict carrier
synchronization and code element synchronization, introducing the appropriate map-
ping mechanism, after the physical layer of relay node superimposes and receives the
electromagnetic waves arriving at the same time, by modulation mapping, this
superimposed electromagnetic wave is mapped to bit information flow on the network
layer, it is equivalent to directly in the network layer carried out XOR coding.
 A Network Coding Against Wiretapping Attacks 431

2.1 The Comparison of Three Transmission Models

In wireless channels, because of its own attributes, multiple messages may be trans-
mitted simultaneously in the channel, radio waves with similar frequencies tend to
interfere with each other and affect the receiver’s reception and judgment of signals.
Therefore, a useful signal transmitted simultaneously may be interference noise for
another signal. So, in the bidirectional relay model, the traditional information trans-
mission mode staggers the transmission time of information. That is, in the same time
slot, only a unique user can send information. Due to the limitation of power coverage,
the information exchanged between two nodes must be forwarded through an inter-
mediate node. Considering the three-node linear network model. In this model, nodes
S1 and S3 are nodes that need to interact with each other, and node R is a relay. X1, X2
and X3 respectively represent information sent by nodes S1, R and S3, and s1 ðtÞ, r2 ðtÞ,
s3 ðtÞ respectively represent physical electromagnetic signals sent by nodes S1, R and
S3, S3 sends a “frame”, a time slot defined as the time it takes to send a frame. In the
traditional scheme, it takes four time slots to complete the mutual transmission (see
Fig. 1.).

Fig. 1. Schematic diagram of traditional transmission scheme

Figure 2 is a schematic diagram of a simple network coding scheme. In this

scheme, only three time slots are required to complete one signal transmission. The
third time slot is R sending X2 ðX2 ¼ X1  X3Þ to both S1 and S3, then S1 and S3
solve each other’s messages based on their own information. Due to the introduction of
network coding, the throughput is 1/3 higher than the traditional scheme.

Fig. 2. A simple network coding scheme

The physical layer network coding is shown in Fig. 3. In this scheme, only two
times gaps are needed to realize the mutual transmission of information. The first time
slot is S1 and S3 sending messages to R at the same time. X2 ðX2 ¼ X1  X3Þ is
encoded by the physical layer network. The second time slot is R broadcasting X2 to
432 Y. Zheng and J. Fu

S1 and S3, and S1 and S3 solve each other’s information according to their own
information. Therefore, compared with the traditional scheme, the throughput is
improved by 100%.

Fig. 3. Physical layer network coding scheme

2.2 Physical Layer Network Coding Mapping

If all nodes are QPSK modulated, the symbol level and carrier phase are synchronized,
and power control is applied to make the signal from S1 and S2 to R have the same
amplitude and phase. The band communication number received by R within a symbol
period can be expressed as:

r2 ðtÞ ¼ s1 ðtÞ þ s3 ðtÞ

¼ ½a1 cosðxtÞ þ b1 sinðxtÞ þ ½a3 cosðxtÞ þ b3 sinðxtÞ ð1Þ
¼ ða1 þ a3 ÞcosðxtÞ þ ðb1 þ b3 ÞsinðxtÞ

In the formula, s1 ðtÞ and s3 ðtÞ are the band communication Numbers sent by N1
and N3 within a symbol period, ai and bi (i = 1or3) are the information bits modulated
by QPSK. R received the communication number with the direction of R is a pair of
signals that are orthogonal to each other, namely ða1 þ a3 ÞcosðxtÞ and
ðb1 þ b3 ÞsinðxtÞ. QPSK signal is regarded as two orthogonal BPSK signals, the same
analysis is done for the homologous component and the orthogonal component.
Let the sending bit data be mi , The range of bit data of source node S1 and S2 in the
network layer corresponding to the modulated electromagnetic wave signal is
ei 2 f1; 1g, its modulation mapping mode is:ei ¼ 2mi  1, electromagnetic waves
are sent by the physical layers of the two nodes. That is, in the case of complete
synchronization, the physical layer of relay node receives the superposition of the
electromagnetic wave amplitude of the signal sent by two signal source nodes, whose
range is ðe1 þ e2 Þ 2 f2; 0; 2g, equivalent to the specific or value directly carried out at
the network layer, the network coding function at the physical layer is realized.
The relay node modulates the mapped signal and sends it, the two source nodes are
demodulated after receiving, then the encoded bit data m ¼ m1  m2 is geted,
according to the signals sent respectively m1 and m2 , S1 and S2 get the messages sent
by the other party through XOR operation.
 A Network Coding Against Wiretapping Attacks 433

2.3 Physical Layer Security System Transport Model

In a wireless two-way relay system, Nodes S1 and S2 need to complete mutual
information exchange, information exchange with the help of relay node R. The
physical layer network and LDPC co-coding system are shown in Fig. 4.

Node S1
QPSK N Node R
LDPC encode
ModulaƟon
SoŌ decision
decoding
PNC map decode
QPSK
LDPC encode
ModulaƟon
Node S2

Fig. 4. Block diagram of physical layer network and LDPC co-coding system

In Fig. 4, the transmission of the entire system consists of two stages: Multiple
access phases (MAC) and broadcast phases (BC). At the multi-access stage, the source
nodes S1 and S2 simultaneously send coded modulated signals to the relay R, but the
relay R receives signals from S1 and S2 and mixed signals with noise; During the
broadcast phase, relay R simultaneously sends network-coded messages to nodes S1
and S2.

3 LDPC Security Coding Design for Gaussian

Eavesdropping Channel

3.1 Gaussian Eavesdropping Channel Model

The wiretapping channel model proposed by Wyner is shown in Fig. 5. Gaussian
eavesdropping channel refers to the gaussian distribution of the main channel and the
noise in the eavesdropping channel model. The channel model consists of sender, legal
receiver and eavesdropper.

Legimate
Sender X main channel Y user
Encoder Decoder
(Gaussian noise)

Hacking channel
(Gaussian noise)

Eavesdropper Z

Fig. 5. Eavesdropping channel model

434 Y. Zheng and J. Fu

As shown in Fig. 5, the legitimate users received messages yn and eavesdroppers

received zn , the gaussian channel model could be represented as:

yn ¼ xn þ vb ; z n ¼ xn þ ve ð2Þ

Where vb and ve respectively represent channel noise of legitimate users and

eavesdroppers, they are all gaussian noises, whose means are 0, the variances are r2b
and r2e . If the message to be sent is expressed as u and the resulting matrix is G, then
the encoded code word is xn ¼ G
u, namely

yn ¼ G
u þ vb ; z n ¼ G
u þ ve ð3Þ

3.2 Security Coding Condition

According to Shannon’s perfect secrecy theory, the plaintext data and ciphertext data
are statistically independent, the ciphertext cannot give any information in clear text,
that is, the eavesdropper cannot get any useful information about the confidential
information from the code word. The mathematical description of physical layer
security is as follows:

HðWÞ ¼ HðWjZ n Þ ð4Þ

Where W is the confidential message being sent; H(W) is the information contained
in the source; Signal through the channel transmission with noise to eavesdroppers,
source of information loss is HðWjZ n Þ.
Wyner used conditional entropy ðHðWjZ n ÞÞ to represent the eavesdropper’s
uncertainty about confidential information. In practical communications, it is difficult to
calculate the uncertainty of the eavesdropper. Therefore, the error rate of eavesdropper
is approximately defined as equivalent instead of uncertainty. It’s not hard to prove
from the definition of information entropy that the code error rate of the eavesdropper is
0.5 when the uncertainty is maximum. That is, the decoding error rate of the eaves-
dropper is equal to 0.5. Wyner showed that as long as the quality of the main channel is
better than that of the eavesdropping channel, then a coding scheme that meets this
requirement can always be found. but the encoding method that meets the requirement
has a transmission upper limit, namely secret capacity. I. Csiszar and S.K.L eung’s
study shows that the secret channel capacity of gaussian channel is the difference
between the capacity of legal channel and the capacity of eavesdropping channel. that
is, the upper bound of the channel secret transmission rate. Therefore, the security
capacity is the maximum transmission efficiency when the eavesdropper has the
greatest uncertainty. The reduction of the minimum bit error rate (the increase of
reliability) is achieved by adding redundancy, this must reduce the transmission rate,
confidentiality conditions try to limit this redundancy. However, reliability and con-
fidentiality can be well compromised, and the measurement index is channel confi-
dentiality capacity.
 A Network Coding Against Wiretapping Attacks 435

According to the random binning coding technology proposed by Wyner, messages

to be sent correspond to boxes of large Numbers of code words, when given a message
to be transmitted, randomly select a code word from the corresponding code word box
and send it. In order not to let the eavesdropper know the specific message sent, if the
eavesdropper can find the random code word sent correctly from the corresponding
code word box of the specific message, then the eavesdropper’s decoding ability is
consumed. When a message is sent in k bits and the length of the code word is n + k
bits, the design of the security coding scheme requires the following three conditions:
① The code can be divided into 2k sub codes, each of which corresponds to send a k-
bit message length.
② The actual transmission efficiency of this code is less than the channel capacity of
the main channel ðCðSNR1 ÞÞ, and the actual transmission efficiency of the sub-
code is equal to the channel capacity of the eavesdropping channel ðCðSNR2 ÞÞ.
③ Given the sent message bit k, a code word is randomly selected from its corre-
sponding sub code and sent.

3.3 Security Coding Based on LDPC

The core of physical layer security technology is to design a secure coding scheme for
it. According to the above requirements, the following LDPC security codes can be
designed to meet the requirements. First, parameter description: In additive gaussian
white noise (AWGN) channel, N1 is the noise variance of the main channel; N2 is the
noise variance of the eavesdropping channel. P is the transmission power of coded
characters. The signal-noise ratio of the main channel is SNR1 ¼ P=N1; The signal-
noise ratio of the eavesdropping channel is SNR2 ¼ P=ðN1 þ N2Þ; The capacity of the
main channel is CðSNR1 Þ ¼ 0:5log2 ð1 þ SNR1 Þ; The capacity of the eavesdropping
channel is CðSNR2 Þ ¼ 0:5log2 ð1 þ SNR2 Þ.
Define the code word format (see Fig. 6.), Suppose the source message is S, and the
length of S is k bits. The encoded code word is C, and the length of C is n + k bits. The
random number generator randomly generated d, whose length is l − k bits. Therefore,
the length of the message bit is l bit, and the length of the parity bit is n + k − l, and
k < l < n + k. Design a LDPC coding matrix called H, who has n + k − l rows and
n + k columns.

l-k k n+k-l
d s p
Fig. 6. The code words

According to random packing technology, the verification matrix H and LDPC

codes are divided into 2k subcodes according to k bit real message, the length of each
subcode is n. This kind of subcode is a linear packet code, the message of which is l-k
bit random message. The code word box can be regarded as a code book. If a nested
436 Y. Zheng and J. Fu

code book is designed, each message corresponds to the sub code book one by one, and
the random number generator is used to generate the random code word that grows into
l − k.
The LDPC security coding scheme based on AWGN channel is as follows:
In the verification matrix H, element “0” accounts for the majority, while element
“1” accounts for the rest, which is a sparse construction matrix. The check matrix is
designed as system code, which is divided into the information part matrix H1 and the
parity part matrix H2 . The information part constructs the cyclic replacement matrix
based on PEG algorithm. the calibration part adopts the structure of quasi dual-
diagonal, thus, the constructed LDPC code has the characteristics of low complexity
and fast coding. The structure of H is as follows:

H ¼ ½H1 H2  ð5Þ

Where H1 is the matrix array of ðn þ k  lÞ  l, corresponding to the information

bit part of the verification matrix H, H2 is the quasi dual-diagonal matrix of
ðn þ k  lÞ  ðn þ k  lÞ, which corresponds to the parity bit of the calibration matrix
H.
The H1 construct can be expressed as:
2 3
Ix


Ix
6 .. .. .. 7
H1 ¼ 4 . . .5 ð6Þ
Ix


Ix

Ix represents the unit matrix of P  P that moves x bits to the right, where
x 2 ð0; 1; . . .; p  1Þ. X equals 0 is the identity matrix I.
H2 is constructed as follows:
2 3
I I


0 0 0
6 .. 7
6 0 I


. 0 07
6 .. .. 7
6 Ix 0


I . . 7
6 7
H2 ¼ 6 .. 7 ð7Þ
6 .


I I 07
6 0 7
6 .. 0 I I 7
4 . 0


5
0 0 I
I 0

Where, I and 0 respectively represent the identity matrix and zero matrix of P  P.
Because H2 is the structure of bidiagonal, it satisfies the nonsingularity. And it can be
used directly for simple and fast coding.
When the information bits are multiplied by the generated matrix, the code word C
encoded by LDPC can be obtained, which is denoted as C = (d, s, p). By the definition
of check matrix H
C T ¼ 0:
 A Network Coding Against Wiretapping Attacks 437

2 3
dT
H
C T ¼ ½ H1 H2 4 sT 5 ¼ 0 ð8Þ
dpT

Check bits of accessible code words:

 T
p ¼ ðd; sÞ
H21
H1T ð9Þ

Therefore, the code word after LDPC code encoding is:

  T 
C ¼ ðd; s; pÞ ¼ d; s; ðd; sÞ
H21
H1T ð10Þ

To meet the conditions of the security encoding scheme ②, make that:


ðl  kÞ=n ¼ CðSNR2 Þ
ð11Þ
ðl  k Þ=n\CðSNR1 Þ

The actual transmission efficiency of sub codes is (l-k)/n. For legitimate users, the
actual transmission efficiency of LDPC code is l/(n + k), it is less than the channel
capacity, therefore, the legitimate user can simultaneously translate the real message s
and the randomly generated message d with the decoding error probability approaching
0. For the eavesdropper, it is hoped that all the decoding ability can be used for the
correct sub code translation (d, p). For the sub code (d, p), where the message is d, the
transmission efficiency of the sub code ðl  kÞ=n ¼ CðSNR2 Þ is required. So

l=ðn þ k Þ [ CðSNR2 Þ ð12Þ

Therefore, for the eavesdropper, the actual transmission efficiency of the code word
is greater than the capacity of the eavesdropper channel. According to Shannon’s
theorem, the probability of eavesdropper decoding error cannot approach 0.
Finally, the classical Belief Propagation (BP) decoding algorithm is used to com-
plete the LDPC coding algorithm.

4 Simulation Experiment and Comparative Analysis

To verify the effect of this method, MATLAB was used to simulate LDPC code in the
gaussian channel. The security network coding scheme designed by the eavesdropping
channel model should meet the requirements of any small bit error rate of the legitimate
receiver (approximately 0) and any large bit error rate of the eavesdropper (approxi-
mately 0.5). Setting the number of iterations to 5, constructed the matrix H, so that
there are 8 cyclic permutations in each row, and 4 cyclic permutations in each column,
the order p of the cyclic replacement matrix is 127, so the matrix has n þ k  l ¼ 508
rows and n þ k = 1016 columns. During the simulation, a 127-bit real message and a
random message with l  k ¼ 381 bits were generated at a time. The code word of 508
438 Y. Zheng and J. Fu

bits message bits and 508 bits check bits was then encoded with LDPC, and the code
length is 1016 bits.
After generating a 127-bit binary number from the binary random sequence gen-
erator, this binary number was received and encoded by the LDPC coding function
module, after coding, the data obtained was modulated with QPSK modulation module,
and then the modulated data entered the AWGN channel module to generate noise
signal, the simulation of the signal generation and transmission process was completed.
Then, QPSK demodulation module was used to demodulate the noise signal at the
receiving end, the receiver transcoded the demodulated signal data through the BP
decoding module, the signal from the transmission to the reception of the entire process
was completed. Figure 7 is the simulation curve of the signal to noise ratio (SNR) and
bit error rate of the main channel (The LDPC code constructed in this paper is com-
pared with PEG algorithm under the same code length and the same parameters).

Fig. 7. The relationship between the main channel error rate and SNR

As can be seen from Fig. 7, with the increase of the main channel SNR, the main
channel error rate decreases. When the SNR is less than 1.5 dB, the encoding scheme
presented in this paper is like the error code performance of PEG algorithm; with the
increase of SNR, the code word performance constructed in this paper is better than
that of PEG code, at the BER of 105 , the codes get an improvement of 0.2 dB
compared with the PEG algorithm.
To study the relationship between the SNR and BER of the eavesdropping channel,
if the SNR of the main channel is fixed, and the SNR of the eavesdropping channel is
variable, the SNR1 of the main channel is taken as 14, SNR2 takes 15 different values
(0.5, 0.4, 0.3, 0.2, 0.1, 0.075, 0.05, 0.02, 0.01, 0.0085, 0.005, 0.0035, 0.002, 0.001).
The simulation curves of SNR and BER of gaussian channel are obtained:
Bit rate l=ðn þ kÞ ¼ 1=2 \CðSNR1 Þ ¼ 0:5log2 ð1 þ SNR1 Þ  2, at this point, the
LDPC code’s actual transmission efficiency is far less than the channel capacity of
main channel. As shown in Fig. 8, the lower the SNR of the eavesdropping channel,
the higher the bit error rate of the eavesdropper is. That is, the closer the error prob-
ability of the eavesdropper is to 0.5, the more uncertain the eavesdropper is, and the
more secure the security encoding scheme is. Moreover, when the SNR of the
 A Network Coding Against Wiretapping Attacks 439

eavesdropping channel increases, the code rate should be reduced, because the main
channel SNR is fixed, the security capacity decreases with the increase of the SNR2 .
Therefore, when the eavesdropper error rate reaches the maximum, the code rate can be
increased to make the transmission rate close to the secret capacity of the channel.
Based on channel security, the maximum speed transmission can be realized.

Fig. 8. The relationship between the BER and SNR of the eavesdropping channel

5 Conclusion

This paper mainly introduces the construction of LDPC code verification matrix and
the realization process of security coding, quasi dual-diagonal matrix reduces the
coding complexity, the structured unit circulant replacement matrix effectively reduces
the memory space, unrestricted column weight improves LDPC code decoding per-
formance. Finally, a simple verification simulation experiment was conducted for
LDPC code through MATLAB, under the AWGN and BP decoding algorithm, the
simulation results show that the proposed LDPC codes get an improvement of 0.2 dB
compared with the PEG algorithm at the BER of 106 , under the same BER condition,
this method is closer to zero. In the eavesdropping channel, as the SNR of the
eavesdropping channel decreases, the bit error rate approaches 0.5. This method can
effectively prevent eavesdropping.

Acknowledgments. This work was financially supported by the Science and Technology
Commission of Shanghai Municipality of China under Grant (No. 17511107002).

References
1. Shannon, C.E.: Communication theory of secrecy systems.In: International Conference on
Autonomous Agents and Multiagent Systems. International Foundation for Autonomous
Agents and Multiagent Systems, pp.1951–1952 (2015)
2. Wyner, A.D.: The wire-tap channel. Bell Labs Tech. J. 54(8), 1355–1387 (2014)
3. Csiszar, I., Korner, J.: Broadcast channels with confidential messages. IEEE Trans. Inf.
Theory 24(3), 339–348 (2003)
440 Y. Zheng and J. Fu

4. Zhang, S., Liew, S.C., Lam, P.P.: Hot topic: physical-layer network coding. In: International
Conference on Mobile Computing and Networking, MOBICOM 2006, Los Angeles, Ca,
USA, pp. 358–365, September 2006
5. Ahlswede, R., Cai, N., Li, S.Y.R., et al.: Network information flow. IEEE Trans. Inf. Theory
46(4), 1204–1216 (2002)
6. Cai, N., Yeung, R.W.: Secure network coding. In: 2002 Proceedings of the IEEE
International Symposium on Information Theory, p. 323. IEEE (2008)
7. Wang, J., Wang, W.: Probability Theory and Mathematical Statistics. Advanced Education
Press, Beijing (2015)
8. Mahdavifar, H., Vardy, A.: Achieving the secrecy capacity of wiretap channels using polar
codes. IEEE Trans. Inf. Theory 57(10), 6428–6443 (2011)
9. Dorf, R., Simon, M., Milstein, L., et al.: Digital communication. IEEE Trans. Acoust.
Speech Signal Process. 32(1), 190 (2017)
10. Liu, L., Yan, Y., Ling, C.: Achieving secrecy capacity of the gaussian wiretap channel with
polar lattices. IEEE Trans. Inf. Theory 64(3), 1647–1665 (2018)
11. Hu, X.Y., Eleftheriou, E., Arnold, D.M.: Regular and irregular progressive edge-growth
tanner graphs. IEEE Trans. Inf. Theory 51(1), 386–398 (2005)
Neural Networks and Deep Learning
 A Weighted KNN Algorithm Based
on Entropy Method

Hui Zhang(&), Kaihu Hou, and Zhou Zhou

Kunming University of Science and Technology, Kunming 650000, China

[email protected]

Abstract. Aiming at the problem that the classification accuracy of K-nearest

neighbor algorithm is not high, this paper proposes a K-nearest neighbor
algorithm that uses the weighted entropy method of Extreme value (EEM-KNN
algorithm). The entropy method assigns weight to the sample’s feature index,
and then introduces the weight of the feature index when calculating the distance
between the query sample vector and the training sample vector. The four
groups of classification data sets are used as test samples to test the effectiveness
of the improved KNN algorithm, it also compares the difference between the
improved algorithm and the traditional algorithm under different K values.
Algorithms are implemented and tested on the Jupyter Notebook interactive
platform. The improved KNN algorithm is verified by experiments, and the
classification accuracy is improved.

Keywords: KNN algorithm
Distance metric
Entropy method

Weighting

1 Introduction

K-Nearest Neighbor (KNN) algorithm, as the supervised learning method, has been a
relatively mature classification method in theory and is commonly used in the classi-
fication of machine learning and data mining classification. The classification of the
KNN algorithm is realized by the training sample k, an arbitrary number greater than 3,
that is closest to the classification object. The spatial distance between the samples
generally adopts the Euclidean distance measurement method. All the samples are
classified into the same category when least three samples belong to the category [1].
KNN algorithm relies on training samples, one of the commonly used classification
algorithms, becoming widely used in classification problems owning to its simplifi-
cation, high efficiency, avoiding training and establishing the predictive model [2–6].
There is an inductive bias in KNN [7], and the classification results may be inaccurate
due to the uneven distribution of samples. The optimizing of KNN algorithm have
developed three angles, such as feature attribute pruning, rough set, improved distance
formula. The clustering, principal component analysis, etc., as the common ways of the
feature attribute pruning, have the disadvantage of possibly cutting too much in the
process of pruning. The disadvantage of rough set classification is that it is sensitive to
noise data. The improved distance formula is a common direction of improvement. The
traditional KNN algorithm adopts Euclidean distance. Some scholars have proposed

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 443–451, 2018.
https://doi.org/10.1007/978-981-13-2384-3_41
444 H. Zhang et al.

using chi-square distance and Markov distance instead of Euclidean distance calcula-
tion. These improved ideas are based on the traditional KNN algorithm to improve the
classification accuracy.
The high dimension of the data is one of the reasons for the low classification
accuracy, some attributes in the attribute dimension are of little reference value to the
classification results, this part of the noise data reduces the accuracy of the classifi-
cation. When the traditional KNN algorithm calculates the similarity between samples
by various distance formulas, it will appear that the feature index with large data
difference has a great influence on the classification result, and the weight of the
classification feature index has not been considered. Therefore, considering the weight
of feature index in the classification of the algorithm is an important direction of the
improvement of the KNN algorithm. Some scholars have improved the algorithm in
this direction. For example, a KNN algorithm based on particle swarm optimization
(PSO) weighting is proposed, by using PSO-KNN algorithm and traditional KNN
algorithm, GA weighted KNN algorithm are used to classify the sample data, which
verifies that PSO-KNN algorithm is superior to traditional KNN algorithm and GA
weighted KNN algorithm [8]. Using attribute correlation to calculate weight improves
the nearest neighbor algorithm of K [9]. In reference [10], the traditional KNN algo-
rithm only takes into account the distance between the same attributes, neglecting the
correlation between the attributes, which leads to the problem of low classification
accuracy, thus, the KNN algorithm of feature entropy correlation difference is pro-
posed. A KNN algorithm for information entropy of attribute values is introduced in
reference [11], which based on the sample with the same maximum attribute value
between the samples to be classified and training samples to classify the samples to be
classified, suitable for the classification of non-numerical attribute values. Literature
[12] introduces the sparse learning theory and proposes a K-valued adaptive SA-KNN
algorithm. In reference [13], an improved algorithm for KNN algorithm is proposed,
which uses the combined feature entropy to calculate the information entropy, form a
new feature attribute, and then classify the samples. In order to solve the problem of
low classification accuracy brought about by multidimensional information, using
information entropy to reduce the dimension of the feature attribute and using the
attribute with large contribution degree to classify the KNN algorithm in reference [14].
A KNN algorithm based on attribute weighted entropy is proposed in reference [15],
the method of normalized entropy value is used to calculate the weight of characteristic
attributes, and the factor of attribute weight is added to calculate the distance between
samples.
To sum up, most of the KNN optimization algorithms which join weights are based
on entropy method to give weights to attributes. Entropy method is the most primitive
entropy nondimensionalization standard, which is the normalization method, and the
interval stability and translation independence of this method are weak. In the nondi-
mensionalization method of the entropy method, the extreme value processing method
is the best [16]. Therefore, based on the existing research, the KNN algorithm (EEM-
KNN algorithm) based on the extreme value entropy method is proposed, and the
EEM-KNN classification is used to classify 4 data sets in the Jupyter Notebook
 A Weighted KNN Algorithm Based on Entropy Method 445

interactive WEB end based on Python language, to test the rationality of the algorithm.
The improved algorithm is superior to the traditional KNN algorithm in improving the
classification accuracy.

2 KNN Algorithm

The KNN algorithm is a case-based learning algorithm. Given a set of local query
vector sets and a set of feature vector sets, the algorithm searches K nearest local
feature vectors. The working mechanism of the KNN algorithm is to first calculate all
the distances between the query vector and all database vectors, then sort the calculated
distances, and select the smallest K reference amounts from the ranking to determine
the classification of the query vectors [17].

2.1 Distance Metrics

The core of the KNN algorithm is to calculate the distance between the sample to be
classified and the training sample. The common methods of distance measurement
include Euclidean distance, Manhattan distance, Minkowski distance, chi-square dis-
tance. The traditional KNN algorithm uses the Euclidean distance. Let there be two n-
dimensional vectors X, Y, let X = (X1, X2, …, Xn), Y = (Y1, Y2, …, Yn). The
Euclidean distance of X and Y is defined as:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X n
D¼ ðXi  Yi Þ2 ð1Þ
i¼1

3 EEM-KNN Algorithm

3.1 Entropy Method

Entropy method first appeared in thermodynamics, then was introduced into infor-
mation theory by Shannon. At present, entropy method is widely used in engineering,
economy and other fields. Entropy measures the amount of information provided by
data, and small entropy value indicates that the amount of information provided is
large. Then the weight ratio of the corresponding index should be larger, and if the
entropy value is large, the corresponding index weight should be set a little smaller.
Entropy value method is an objective and effective weighting method [18].
The calculation steps of the entropy method are:
There are mdata samples and n sample characteristic indexes, and there is a data
matrix M ¼ aij mn .
446 H. Zhang et al.

(1) tij indicates the proportion of the i-th sample data under the j-th index to the total
sample under the j-index
aij  mj
tij ¼ Pn ð2Þ
j¼1 ðaij  mj Þ

Among them, mj is the minimum value under this feature attribute

(2) Calculate the information entropy hj of j-th index

X
n  
hj ¼  tij ln tij ð3Þ
j¼1

(3) Calculate the entropy value ej of the j-th index

ej ¼ khj ð4Þ
 
1
k¼
ln m

(4) Calculate the difference coefficient gj :

gj ¼ 1  e j ð5Þ

(5) Define the weight of the feature index:

gj
bj ¼ Pm ð6Þ
j¼1 gj

The entropy method can effectively avoid the influence of subjective randomness
on the importance of the index. It depends on the sample data and has a strong
theoretical basis. The performance of extreme value method in the nondimensional-
ization method of entropy value method is the best. Therefore, this paper uses the
extreme entropy method to give weight to the sample characteristic index [17].

3.2 Improved KNN Algorithm

The steps of the improved KNN classification algorithm are as follows:
(1) Create a sample data set S, let the Xi is the training sample set, i 2 ½1; . . .. . .; m,
the dataset contain n different classification categories., Yi is the test sample,
i 2 ½1; . . .. . .; t, The data set contains m characteristic indexes in each sample. The
method of extreme entropy value is used to calculate the weight of the charac-
teristic attribute of the sample data, the process is carried out in the form of
formulas (2)–(6).
 A Weighted KNN Algorithm Based on Entropy Method 447

(2) Set an initial value for k in the KNN algorithm.

(3) The distance between each test sample and the training sample is calculated using
an improved distance metric formula, such as:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X n
D¼ bj ðXi  Yi Þ2 ð7Þ
i¼1

(4) Sort the calculated distance and select k training samples nearest to the sample to
be classified.
(5) According to the category of k selected training samples, the largest sample of k
samples is used as the category of samples to be classified.
(6) According to the classification results of the samples to be classified and com-
pared with the actual category, the accuracy of the classifier is calculated. The
formula for calculating the accuracy of the classifier is as follows:

B
p¼  100% ð8Þ
C

In the formula: B is the correct number of test samples and C is the total number of
test samples.

4 Example Verification

In this paper, the improved algorithm is verified by data, and the KNN classifier is
written with python on the Jupyter Notebook interactive WEB. The data set is imported
into the classifier, and different K values are selected to verify the data. The whole
verification process is divided into two parts. In the first part, the original KNN
algorithm is used to verify the data. In the second part, the improved KNN algorithm is
used to compare the accuracy of different K values with the results of the first
part. Finally, the experimental results are analyzed.

4.1 Experimental Data

In this paper, four groups of data are used to verify the validity of the algorithm. Data
set 1 is derived from the chemical composition analysis of tobacco leaves in a tobacco
re-drying plant, data set 2, data set 3 and data set 4 are derived from UCI data sets.
Table 1 shows the composition of the test data. One third of the total data of each data
set is selected as the test sample and the remaining 2/3 as the training sample.
Table 1. Data composition of algorithm validation
Data Set Samples Features Classes
1 tobacco leaves 749 6 5
2 Iris 150 4 3
3 wine 178 13 3
4 Glass-identification 214 9 7
448 H. Zhang et al.

4.2 Comparative Analysis

In order to verify the validity of the improved KNN classification algorithm and to
change the K value during the verification process, the algorithm was verified with
different data sets. In the verification process, each dataset was tested 10 times, in order
to verify the effectiveness of the improved algorithm, the average of 10 times of
verification is taken, as shown in Table 2:

Table 2. Test accuracy of different data sets for different K values (%)
Data sets Algorithm K=3 K=5 K=8
Tobacco leaves KNN 92.1594 93.3741 93.2211
EEW-KNN 95.0244 96.6005 94.246
Iris KNN 95.6733 95.456 95.2647
EEW-KNN 96.6321 96.01 95.923
Wine KNN 73.1842 73.7775 72.8113
EEW-KNN 90.9656 93.6511 90.425
Glass-identification KNN 70.9336 68.4688 69.3213
EEW-KNN 82.3282 88.6889 80.0948

It can be seen from Table 2 that in the data set test of the original KNN algorithm,
the classification accuracy of the data sets numbered 3 and 4 is low, mostly around
70%. while for data set 1 and data set 2, the classification accuracy is above 90%. From
the classification results of the two classification algorithms, it can be seen that the
classification accuracy of the four data sets is improved by using the improved KNN
classification algorithm, and the changes in the wine data set are most obviously. The
largest increase of accuracy is the test in the Glass-Identification data set when K = 5.
There is a low accuracy of classification owning to more sample features contained
in the data sets 3 and 4, leading to strong interference feature data. In order to improve
the accuracy of classification, low weight is given to the feature data with poor overall
performance to reduce their impact on the classification results when improved KNN
algorithm is adopted.
In order to evaluate the classification effect of the improved KNN algorithm more
intuitively, a broken line graph of the classification accuracy with different K values
was plotted. When K = 3, the accuracy rate result is shown in Fig. 1. When K = 5, the
result is shown in Fig. 2, and when K = 8, the result is shown in Fig. 3.
It can be seen from Figs. 1, 2 and 3 that the classification accuracy of the improved
algorithm is greatly improved no matter how much K value is taken by adding the
eigenvector weight of entropy method. It can be seen from the Fig. 1 that the KNN
algorithm with Euclidean distance is more effective in the classification of data set Iris.
When K = 3, the classification accuracy is the highest. For data set 1, the classification
accuracy is the highest when K = 5. Figures 1, 2 and 3 show that the stability of the
improved algorithm in the data training process is not very good, and the accuracy of
the algorithm varies greatly.
 A Weighted KNN Algorithm Based on Entropy Method 449

100 tobacco leaves KNN

95 tobacco leaves EEW-KNN
90
Iris KNN
85
Iris EEW-KNN
80
wine KNN
75
70 wine EEW-KNN

65 Glass-IdenƟficaƟon KNN

60 Glass-IdenƟficaƟon EEW-KNN
1 2 3 4 5 6 7 8 9 10

Fig. 1. Curve of the accuracy of the result when K = 3

100 tobacco leaves KNN

95
tobacco leaves EEW-KNN
90
Iris KNN
85
80 Iris EEW-KNN
75 wine KNN
70
wine EEW-KNN
65
60 Glass-IdenƟficaƟon KNN

55 Glass-IdenƟficaƟon EEW-KNN
1 2 3 4 5 6 7 8 9 10

Fig. 2. Curve of the accuracy of the result when K = 5

100 tobacco leaves KNN

95 tobacco leaves EEW-KNN
90
Iris KNN
85
Iris EEW-KNN
80
wine KNN
75
70 wine EEW-KNN

65 Glass-IdenƟficaƟon KNN

60 Glass-IdenƟficaƟon EEW-KNN
1 2 3 4 5 6 7 8 9 10

Fig. 3. Curve of the accuracy of the result when K = 8

450 H. Zhang et al.

4.3 Experimental Results

According to the classification of the improved KNN algorithm and the traditional
KNN algorithm in four data sets, the improved KNN algorithm has greatly improved
the classification accuracy. The results of classification are influenced directly by the
selection of K value in the KNN algorithm. A small K value will result in too little
information for the sample to be classified and affect the classification accuracy while a
great K also cause a unreasonable classification owning to the decision-making power
being held in the hands of the “majority” of the selected training samples according to
the “majority” principle when classify the sample to be tested by KNN algorithm. So
the choice of K value is very important when classifying. From Figs. 1, 2 and 3, It can
be seen that when using the improved KNN algorithm based on Euclidean distance, the
classification of the data set Iris should select K = 3. When K = 5, Data set tobacco
Leves, data set wine and data set Glass-Identification have the highest classification
accuracy.

5 Conclusion

In this paper, we propose an improved K-nearest neighbor classification algorithm

(EEM-KNN) to solve the feature index problem, which has an effect on classification
accuracy. Before classifying data samples, the method of extreme value entropy, which
has good nondimensionalization effect, is used to give weight to the characteristic
index of the data, and then the weight is introduced into the distance formula to reduce
the influence of noisy data on the classification result, it also avoids the excessive
clipping of feature indicators, and the accuracy is used to measure the effectiveness of
the improved algorithm. The experimental results of four groups of classification
datasets show that the improved algorithm keeps the integrity of the data set and
improves the accuracy of classification. But the algorithm is not perfect enough, the
algorithm is more suitable for the classification of data sets with numerical index, it is
not widely used, the robustness of the algorithm is not good enough, and the accuracy
of classification fluctuates greatly. In the future study, we should improve on these two
aspects.

References
1. Li, L., Weinberg, C.R., Darden, T.A., et al.: Gene selection for sample classification based
on gene expression data: study of sensitivity to choice of parameters of the GA/KNN
method. Bioinformatics 17(12), 1131 (2001)
2. Beaudoin, A., Bernierp, Y., Guindon, L., et al.: Mapping attributes of Canada’s forests at
moderate resolution through. Can. J. Forest Res. 44(5), 521–532 (2014)
3. Pandya, D.H., Upadhyay, S.H., Harsha, S.P.: Fault diagnosis of rolling element bearing with
intrinsic mode function of acoustic emission data using APF-KNN. Expert Syst. Appl. 40
(10), 4137–4145 (2013)
4. Aslam, M.W., Zhu, Z., Nandi, A.K.: Automatic modulation classification using combination
of genetic programming and KNN. IEEE Trans. Wirel. Commun. 11(8), 2742–2750 (2012)
 A Weighted KNN Algorithm Based on Entropy Method 451

5. Steenwijk, M.D., Pouwels, P.J., Daams, M., et al.: Accurate white matter lesion
segmentation by k nearest neighbor classification with tissue type priors (kNN-TTPs).
Neuroimage Clin. 3(3), 462–469 (2013)
6. Homaeinezhad, M.R., Atyabi, S.A., Tavakkoli, E., et al.: ECG arrhythmia recognition via a
neuro-SVM-KNN hybrid classifier with virtual QRS image-based geometrical features.
Expert Syst. Appl. 39(2), 2047–2058 (2012)
7. Tan, S.: An effective refinement strategy for KNN text classifier. Expert Syst. Appl. 30(2),
290–298 (2006)
8. Ren, J.T., Zhuo, X.L., Xu, S.C.: PSO based feature weighting algorithm for KNN. Comput.
Sci. 34(5), 187–189 (2007)
9. Wang, Z.; Wang, K.: Improve KNN algorithm based on entropy method. Comput. Eng.
Appl., 45, 129–131,160 (2009)
10. Zhou, J., Liu, J.: KNN algorithm based on feature entropy correlation difference. Comput.
Eng. 37(17), 146–148 (2011)
11. Tong, X., Zhou, Z.: Enhancement of K-nearest neighbor algorithm based on information
entropy of attribute value. Comput. Eng. Appl. 46(3), 115–117 (2010)
12. Sun, K., Gong, Y., Deng, Z.: An efficient K-value adaptive SA-KNN algorithm. Comput.
Eng. Sci. 37(10), 1965–1970 (2015)
13. Zhou, J., Liu, J.: Improved K-nearest neighbor algorithm for feature union entropy. Comput.
Appl., 31(7), 1785–1788,1792 (2011)
14. Wei, L., Zhao, X., Zhou, X.: An enhanced entropy-K-nearest neighbor algorithm based on
attribute reduction. In: 2014 4th International Conference on Computer Engineering and
Networks, pp 381–388 (2014)
15. Xiao, X., Ding, H.: Enhancement of K-nearest neighbor algorithm based on weighted
entropy of attribute value. In: 2012 5th International Conference on BioMedical Engineering
and Informatics, Chongqing, pp. 1261–1264 (2012)
16. Zhu, X., Wei, G.: Discussion on the non-dimensional method’s good standard in entropy
method. Stat. Decis. 02, 12–15 (2015)
17. Mejdoub, M., Amar, C.B.: Classification improvement of local feature vectors over the KNN
algorithm. Multimed. Tools Appl. 64, 197–218 (2013)
18. Zou, Z.H., Yun, Y., Sun, J.N.: Entropy method for determination of weight of evaluating
indicators in fuzzy synthetic evaluation for water quality assessment. J. Environ. Sci. 18(5),
1020–1023 (2006)
 Control Strategy and Simulation for a Class
of Nonlinear Discrete Systems with Neural
Network

Peng Liu(&)

Department of Communication Engineering, Chongqing College of Electronic

Engineering, Chongqing, China
[email protected]

Abstract. A PID algorithm based on multi-layer neural network training is

presented in this paper. The indirect automatic tuning controller for nonlinear
discrete systems adopts a learning algorithm. The problem is to select bounded
control so that the system output is as close as possible to the required value.
Finally, an example is given to show that the proposed controller is effective.

Keywords: Neural network
Automatic tuning
Nonlinear discrete systems

PID control

1 Introduction

Due to its robustness and simplicity, the PID controller has been widely used in
industry. For ordinary control processes, they show satisfactory control results. How-
ever, due to the unpredictability of the dynamic process, the controller parameters are
not adjusted in time, which will lead to poor control performance. Many of the changes
in parameters are caused by nonlinearity. Since 2010s, using artificial intelligence to
study nonlinear adaptive PID control has become a research hotspot. The use of neural
networks has been shown to be successful in identifying and controlling complex
nonlinear dynamics systems [4, 5]. In particular, a new PID adaptive control scheme is
proposed for a nonlinear system [6]. The model discussed is a class of nonlinear and
time-varying systems. Therefore, the PID control is applied to the dominant model to
achieve satisfactory control effect. The neural network uses radial basis function net-
work. Documentation [7–10] Several different BP learning algorithms are proposed,
such as an online neural network learning algorithm for time varying processing, fast
learning algorithm based on neuronal spatial gradient decline and LM algorithm.
The overall arrangement of this article is as follows: In Sect. 2, the control system
is expounded. In Sect. 3, we proposed the learning algorithm. The design and the rules
for the PID tuning are explained. Simulation result and conclusions are given in
Sects. 4 and 5, respectively.

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 452–457, 2018.
https://doi.org/10.1007/978-981-13-2384-3_42
 Control Strategy and Simulation for a Class 453

2 Problem Statement

A SISO discrete time nonlinear system

yðj þ 1Þ ¼ g½yðjÞ;


; yðj  n þ 1Þ;

ð1Þ
uðjÞ;


; uðj  m þ 1Þ

which u, y are the control and the output, respectively, n is non-negative integer
(n  m), m is non-negative integer, g is an unknown nonlinear function. Assume that
the output rðjÞ is single boundary.
The purpose of the control problem is to determine a bounded control u so that the
system output y can tracks the required value r.
Denotes the tracking error between the required value r and the system output y as

eðkÞ ¼ rðjÞ  yðjÞ ð2Þ

The incremental expression of PID controller is described by

uðjÞ ¼ uðj  1Þ þ kp ðeðjÞ  eðj  1ÞÞ

ð3Þ
þ ki eðjÞ þ kd ðeðjÞ  2eðj  1Þ þ eðj  2ÞÞ

where kp , ki and kd are the parameters which are tuned by the BPNN algorithm with
optimal learning rate.

3 Math the Learning Algorithm

In the proposed methodology, the adjustment criterion for the weights w of the neural
network is the negative gradient of the variance performance index function. The
variance performance index of systems (1) be defined as:

1
EðjÞ ¼ ½rðj þ 1Þ  yðj þ 1Þ2
2 ð4Þ
1 2
¼ e ðj þ 1Þ
2

The numbers of the wki and vlk are arbitrary. The vlk has the input regression vector,
composed of the desired input, the past states of the plant output and the control signal.
Three neurons in the output layer are jp ; ji ; and jd . The structure of NN is shown in
Fig. 1.
The NN weights are computed by the following equation

@E
Dwlk ðjÞ ¼ r ð5Þ
@wlk
454 P. Liu

Fig. 1. The architecture of the NN.

In the learning process, small learning rate parameters r ð0  r\1Þ should be

selected. In this way, the smaller the weight change in the network, the smoother the
trajectory of the weight change. However, this improvement will lead to an increase in
the amount of calculations while slowing down the learning speed. To solve the above
problem, we can modify the delta rule of (5) by adding a momentum item [11].

@E
Dwlk ðjÞ ¼ r þ aDwlk ðj  1Þ ð6Þ
@wlk

where 0  a\1 is the momentum constant to increase the learning speed. It controls
the feedback loop acting around Dwlk ðj  1Þ.

@E @yðj þ 1Þ @uðjÞ @outl @netl

¼ eðj þ 1Þ ð7Þ
@wlk @uðjÞ @outl @netl @wlk

NN

PID Pl

Fig. 2. T the feedback closed-loop system

The structure of the overall feedback closed-loop system is shown in Fig. 2.

Remark 1. We give the following algorithm for system (1) with the variance per-
formance index (4).
Step1: Assuming that no prior information is available, pick the synaptic weights
and thresholds from (0, 1), randomly.
 Control Strategy and Simulation for a Class 455

Step2: Calculate eðjÞ from (2).

Step3: The actual weight changes from the out node j to the hidden node k are
defined as

Dwk ðjÞ ¼ rdl outk þ aDwlk ðj  1Þ ð8Þ


@yðj þ 1Þ @uðjÞ 0
dl ¼ eðjÞSign f ðnetl Þ ð9Þ
@uðjÞ @outl

where netl is the total input of the neuron k; wlk ðjÞ is the neuron weights between the l
and k.
The actual Dwki ðjÞ changes from the hidden node k to the input node i are defined as

Dwki ðjÞ ¼ rdk outi þ aDwki ðj  1Þ ð10Þ

X
dk ¼ f 0 ðnetk Þ dl wlk ð11Þ
l

where outk is the output of the k; wki ðjÞ is the neuron weights between the hidden
neuron k and the input neuron i. f is the sigmoid function can be defined by

1
f ðnetk Þ ¼ ð12Þ
1 þ expðnetk Þ
P
Step4: Give some positive real constant e. If eðjÞ2 \e, then stop and output uðjÞ
j
in (3), else go to step 3.
Step5: Letting j ¼ j þ 1, go to step 2.

4 Simulation

Next, we use the MATLAB to simulate the above algorithm. In this example, the
nonlinear system is described by

5yðjÞ
yðj þ 1Þ ¼ þ u3 ðjÞ ð13Þ
2:5 þ yðjÞ

The desired values are sinð0:25pjÞ for 0  j  40 and sinð0:5pjÞ for 40\j  80.
Each simulation, using 0 to 80 steps as a simulation cycle. NN consists of 1 middle
layer with 10 nodes, and 3 output nodes. Which r ¼ 0:1, a ¼ 0:4.
The desired value figure, the error between the system outputs and expected values
are shown in Fig. 3, respectively.
From Fig. 3 it can be seen that the plant outputs is perfectly coincident with the
desired values.
456 P. Liu

Fig. 3. The system outputs (solid lines) and the expected values (dashed lines). (a) The desired
value, (b) The plant outputs, (c) The error between the system outputs and expected values
 Control Strategy and Simulation for a Class 457

5 Conclusion

Neural Network PID control of nonlinear systems using a fast learning algorithm is
present. By comparing the error of the system output and the expected value, the
weight of NN is reversed. In the simulation, the output of the system quickly finds the
expected value, so the effectiveness of the learning algorithm is proved.

References
1. Al-Assadi, S.A.K., Al-Chalabi, L.A.M.: Optimal gain for proportional integral derivation
feedback. IEEE Control Syst. Mag. 7(2), 16–19 (2007)
2. Hwang, C., Hsiao, C.-Y.: Solution of a non-convex optimization arising in PI/PID control
design. Automatica 38(6), 1895–1904 (2012)
3. Daley, S., Li, G.P.: Optimal PID tuning using direct search algorithms. Comput. Control
Eng. J. 10(3), 51–56 (2009)
4. Juditsky, A., Hjalmarsson, H., Benveniste, A., Delyon, B., Ljung, L.: Nonlinear black-box
models in system identification. Mathematical foundations. Automatica 31(3), 1725–1750
(2015)
5. Narendra, K.S.: Neural networks for control: theory and practice: mathematical foundations.
Proc. IEEE 84(1), 1385–1406 (2016)
6. Huang, S.N., Tan, K.K., Lee, T.H.: A combined PID/adaptive controller for a class of
nonlinear systems. Automatica 37(6), 611–618 (2011)
7. Zhao, Y.: On-line neural network learning algorithm with exponential convergence rate.
Electron. Lett. 32(1), 1381–1382 (2016)
8. Zhou, G., Si, J.: Advanced neural network training algorithm with reduced complexity based
on Jacobian deficiency. IEEE Trans. Neural Netw. 9(3), 448–453 (2017)
9. Parisi, R., Di Claudio, E.D., Orlandi, G., Rao, B.D.: A generalized learning paradigm
exploiting the structure of feedforward neural networks. IEEE Trans. Neural Netw. 7(2),
1450–1459 (2016)
10. Hagan, M.T., Menhaj, M.B.: Training feedforward neural networks with the Marquardt
algorithm. IEEE Trans. Neural Netw. 5(1), 95–99 (2014)
11. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning representations of back-propagation
errors. Nature 323(2), 533–536 (2016)
 Research on Joint Nondestructive Testing
Based on Neural Network

Junyang Tan1, Dan Xia1(&), Shiyun Dong1, Binshi Xu1,

Yuanyuan Liang2, Honghao Zhu1, and Engzhong Li1
1
National Key Laboratory for Remanufacturing, Fengtai District, Dujiakan no.
21, Beijing 100072, China
[email protected]
2
Beijing Special Vehicle Research Institute, Fengtai District, Huaishuling no. 3,
Beijing 100072, China

Abstract. Based on the ultrasonic and Barkhausen non-destructive testing

methods, the feature extraction and recognition capabilities of deep learning
methods were obtained to quantitatively characterize the surface hardness
characteristics of typical materials. Firstly, the hardness value of 45 steel after
different heat treatment was measured by Brinell hardness tester, and then the
speed characteristics of ultrasonic propagation in 45 steel after different heat
treatment and the Barkhausen noise signal in different steel specimens were
measured through experiments. A polynomial fitting method was used to cali-
brate the mapping curve and verify its detection accuracy. The deep learning
method was used to realize the multi-dimensional comprehensive characteri-
zation of ultrasonic steel speed and Barkhausen noise signal to 45 steel hardness
value, and the feasibility and accuracy advantages of the joint detection method
were verified.

Keywords: Ultrasound
Barkhausen
Hardness
Deep learning

Nonlinear mapping

1 Introduction

Material hardness is one of the key factors influencing the use of materials. The
traditional method for measuring the surface hardness of materials is to use the
indentation test method to calculate the surface hardness value by measuring the size of
the indentation. However, the indentation test method is a destructive test method that
will cause irreversible damage to parts and test blocks and affect subsequent use. With
the continuous emergence of related advanced technologies and instruments, non-
destructive testing technology has gradually been applied to the characterization and
evaluation of the mechanical properties of materials [1], avoiding the irreversible
damage to materials. Conventional nondestructive testing methods including: ultra-
sonic testing, magnetic particle testing, eddy current testing, penetration testing, and ray
testing. In addition, other nondestructive testing methods such as incremental perme-
ability method, metal magnetic memory method, Barkhausen noise method, etc. are
also used.

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 458–467, 2018.
https://doi.org/10.1007/978-981-13-2384-3_43
 Research on Joint Nondestructive Testing Based on Neural Network 459

Ultrasonic non-destructive testing is the most widely used, most frequently used,
and most studied non-destructive testing technology at the present stage [2]. It is widely
used in product quality evaluation, defect evaluation, and performance monitoring [3].
Some scholars also use this method to qualitatively or quantitatively evaluate the
surface hardness of materials. Magnetic Barkhausen Noise (MBN) is also a widely
used non-destructive testing method that can be applied to the detection of mechanical
properties of ferromagnetic materials and used as an internal organizational structure
criterion [4].
There are many influencing factors in the non-destructive testing process, and the
detection results and the required parameters may have a nonlinear mapping rela-
tionship [5]. The use of traditional methods is more complicated and not effective [6].
Due to its excellent feature extraction capability, self-learning ability, and high fault-
tolerance capability, deep neural networks can achieve tasks that cannot be quickly
completed by many conventional methods [5–10]. Through deep neural networks, data
fusion, performance prediction, and intelligent identification of parameters can be
performed [11–17].
In this paper, ultrasonic non-destructive testing method and Barkhausen noise
method were used to detect 45 steel specimens under different heat treatment condi-
tions, and a deep neural network was used to establish a complex mapping relationship
between the two signals and surface hardness, and the accuracy of joint detection was
verified.

2 Experimental Equipment, Materials and Basic Principles

2.1 Experimental Equipment and Materials
Changes in the microstructure of the material under different heat treatment conditions
will lead to different surface hardness. This article selected 45 steel blocks for a total of
six groups, respectively, through six different heat treatment methods.
Ultrasonic detection process using RITEC’s RAM-5000 nonlinear high-energy
ultrasound testing system, SIUI ultrasound instrument Institute of SIUI probe by
sinusoidal signal excitation 2.25 MHz ultrasound, respectively, the longitudinal wave,
plane wave, pulse wave (Fig. 1).

Fig. 1. RITEC’s RAM-5000 nonlinear high-energy ultrasound testing system

460 J. Tan et al.

The Barkhausen detection process was carried out using a portable Barkhausen
detector developed by Nanjing University of Aeronautics and Astronautics and a self-
made probe.

2.2 Indentation Testing

The Brinell hardness of the test piece was measured with the Brinell hardness tester
using the indentation test method. According to the diameter of the hard alloy ball, the
indentation diameter and the experimental force, the Brinell hardness can be calculated:

0:204F
HBW ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; ð1Þ
pDðD  D2  d 2 Þ

Among them, D is the alloy ball diameter, d is the indentation diameter, and F is the
experimental force (take 1000 kgf).

2.3 Ultrasonic Nondestructive Testing

According to the measured thickness of the sample and the ultrasonic waves propagate
the sound of the adjacent pulse bottom wave in the sample, the speed of sound
propagated by the ultrasonic wave in the test block is calculated. Use sound speed to
establish the following mapping relationship:

2h
v¼ ; ð2Þ
s

Where v is the speed of ultrasonic sound, h is the thickness of the sample, and t is
the sound time.

2.4 Barkhausen Testing

The signal parameters of the Barkhausen noise non-destructive testing method include
root mean square, peak-to-peak, half-width, and peak position. In this paper, the root
mean square (RMS) value is selected as the analysis parameter, and the instrument is
used to measure the data and obtain its root mean square to establish the subsequent
mapping relationship. The excitation voltage frequency is 20 Hz, the excitation voltage
is 10 V, the pre-power amplifier is 20 times, and the sampling frequency is 200 K.

2.5 Deep Learning Modeling

In this paper, the feature extraction and learning ability of deep neural network are used
to multidimensionally model the sound velocity data obtained by ultrasonic detection
and the root mean square data of noise signals detected by Barkhausen. Two signal
features are extracted and the hardness is quantitatively characterized. Since the
detection signal has a global correlation characteristic, a full-connection network is
used for modeling. And the fully-connected networks are usually trained using the bp
 Research on Joint Nondestructive Testing Based on Neural Network 461

back-propagation algorithm. Use the logistics function as an activation function (as

shown in Fig. 2).

Fig. 2. The curve of a sigmoid function.

The network structure is 2-100-200-200-1000-200-200-1. The two inputs are the

wave speed and the Barkhausen noise (RMS), and the output is the hardness. There are
6 full-connection hidden layers in the middle, and the network structure is shown in the
Fig. 3.

Fig. 3. The topology of the network. The nodes are fully connected with other notes in the
adjacent layer.

Before the network training, the two input and output data are Z-score normalized.
Randomly increase the Gaussian noise perturbation in the six groups of measured data,
increase the number of samples, and make the network have better robustness and
generalization ability.
462 J. Tan et al.

3 Experimental Results and Analysis

Due to the unevenness of the heat-treated test blocks, the Brinell hardness, ultrasonic
velocity, and Barkhausen noise signals measured at different test points on the surface
will also be different. For the entire multidimensional mapping calibration process, the
network training is more robust because it provides more noise-containing data in
different heat treatment states, making the trained characterization network have better
generalization capabilities. In the experiment, it is assumed that the thickness of each
test piece (line cut) measured by the vernier caliper is the same. The following table
shows the thickness measurement results (Table 1):

Table 1. The hardness of test blocks under six different heat treatment conditions.
Groups A N 600 °C 400 °C 200 °C WQ
Thickness/mm 18.95 18.97 18.98 19.00 18.96 18.97

For the measurement of ultrasonic speed, Barkhausen noise signal, and Brinell
hardness, 200 points were selected for each test block, a total of 1200 sets of data.

3.1 Analysis of Ultrasonic Test Results

According to the actually measured 1200 sets of data, the ultrasonic detection signal
and Brinell hardness were calibrated and analyzed. The scatter points in Fig. 4 show
the measured data distribution. It can be seen that 200 inspection points of each steel
block have fluctuations in a small range, and 6 groups are formed in the
map. According to the distribution in the figure, it can be generally seen that the
calibration curve is monotonously decreasing. The curve function
(y = −1297600 + 447.29698 * x − 0.03853 * x2) can be obtained from the com-
monly used polynomial fitting, and the fitting degree reaches 0.975 (see Fig. 4).
In order to verify the reliability of the calibration curve, in addition to the six test
blocks, a verification block was set up to take a 45 steel test piece, which was heated to
a certain temperature and then water-cooled. The surface Brinell hardness was 312
HBW, and the ultrasonic sound speed was 5875 m/s. When the speed of sound is taken
into the above polynomial, the calculated hardness is 382.7 HBW, and the relative error
is about 22.66%.

3.2 Barkhausen Test Results Analysis

The Barkhausen noise signal (root mean square) and hardness were also calibrated by
1200 sets of measured data. The scatter points in Fig. 5 show the measured data
distribution. It can be seen that similar to ultrasonic testing, the distribution of data
points is divided into several clusters by six different heat treatment test blocks, and the
calibration curve will also be monotonously decreasing, but the unevenness is different
from that of ultrasonic testing. The polynomial fitting gives a fitting function of
y = 598.6186 − 14.13972 * x + 0.11715 * x2, as shown below.
 Research on Joint Nondestructive Testing Based on Neural Network 463

600

500

400
Hardness/HBW

300

200

100

5760 5780 5800 5820 5840 5860 5880 5900 5920

Ultrasonic velocity/(m/s)

Fig. 4. Sound velocity-hardness mapping curve.

600

500

400
Hardness/HBW

300

200

100

0 20 40 60 80 100
barkhausen noise/RMS

Fig. 5. Barkhausen noise-hardness mapping curve.

The root mean square of the Barkhausen noise signal measured by the verification
block is 19.8, and the hardness calculated by the function is 364.6 HBW with an error
of 15.45%.
464 J. Tan et al.

3.3 Deep Learning Results Analysis by Using Ultrasound

and Barkhausen Joint Detection
This paper proposes the joint detection of ultrasonic and Barkhausen noise, using two
signals to calibrate the hardness at the same time, the resulting nonlinear relationship
contains three parameters - ultrasonic speed, Barkhausen noise signal (root mean
square) and hardness, the three. The group data is plotted in the XYZ three-dimensional
coordinate system. The scattered results of the measured results are shown in the
Fig. 6.

500

400
Hardness/HBW

300

200

100 100
5780 80 S
Ult 58005820 60 M
ras /R
oni 58405860 40 ise
cv
elo 20 no
/(m city 58805900 n
/s) se
5920
0
hau
r k
Ba

Fig. 6. Discrete distribution of experimental results by joint detection.

The above figure shows the three-dimensional scatter distribution of three param-
eters. The left and right planar projections have the same distribution as a single non-
destructive testing method.
Through the deep neural network, the input is the ultrasonic detection signal and
the Barkhausen detection signal, the output is the surface hardness, and the amount of
data is expanded by ten times to 12,000 training data using Gaussian noise. After 5000
iterations of training, the network converged. The multi-dimensional mapping curve
established should be on the surface shown in the figure below (Fig. 7).
In fact, the size of the ultrasonic wave and the magnetic domain distribution that
causes the magnetic Barkhausen signal are related to the microstructure of the material.
There is a correlation between the two detection signals. When considering the cor-
relation according to the actual detection data, the multidimensional mapping The
relationship curve is as shown in the Fig. 8.
The root mean square values of ultrasonic velocity and Barkhausen noise signal
measured by the verification block were normalized and brought into the calibration
 Research on Joint Nondestructive Testing Based on Neural Network 465

Fig. 7. Multi-dimensional mapping relationship surface map.

Fig. 8. Multi-dimensional mapping curve.

depth network. The surface hardness was 333.7 HBW, the relative error was 6.96%,
and the accuracy was improved.
Multiple methods combined with non-destructive testing extract material properties
from multiple signals and increase the tolerance to test noise. This can not only
improve the detection accuracy, but also can avoid to some extent that some test pieces
are not sensitive to the change of the parameters of a certain method, which leads to the
466 J. Tan et al.

detection failure. In the future, more parameters will be jointly modeled through more
methods, and feature detection and deep-learning capabilities will be used to implement
joint detection in a broader sense.

4 Conclusion

The hardness of 45 steel specimens under six different heat treatment conditions was
tested using the lossy method. The ultrasonic and Barkhausen noise methods were used
for the quantitative non-destructive evaluation of hardness. Through a large number of
experiments to obtain data, the ultrasonic-hardness, Barkhausen noise-hardness map-
ping relationship was calibrated, and the test block was used to evaluate the calibration
curve error. Afterwards, an ultrasonic-Barkhausen joint detection method based on
deep learning network was proposed and a joint detection network model was estab-
lished. After training, the network can accurately represent the hardness of 45 steel. The
experimental results show that the joint detection method to characterize the hardness
of 45 steel is feasible and has higher accuracy, and this joint detection method provides
a new direction for the comprehensive and intelligent evaluation of the mechanical
properties of materials in the future.

Acknowledgments. This work was supported by the National Key Research and Development
Program of China, no. 2016YFB1100205.

References
1. Xia, X.L., Zhang, Y.Y., Yang, H.X.: Development of mechanical fault diagnosis technology
under the condition of modern science and technology. J. Xuzhou Inst. Technol. 21(9), 67–
69 (2006). (in Chinese)
2. Cheng, S.B., Yao, J., Cai, P.: Application of ultrasonic measuring technology in the rigidity
measurement. Chin. Meas. Technol. 30(1), 12–13 (2004). (in Chinese)
3. Vasudevan, M., Palanichamy, P., Venkadesan, S.: A novel technique for characterizing
annealing behavior. Scr. Metall. Mater. 30(11), 1479–1483 (1994)
4. Puppin, Ezio: Statistical properties of Barkhausen noise in thin Fe films. Phys. Rev. Lett. 84
(23), 5415 (2000)
5. Chen, Y.H., Li, M.X.: Application of artificial neural network to ultrasonic nondestructive
testing. J. Appl. Acoust. 3, 40–44 (1996). (in Chinese)
6. Lorenz, M., Wielinga, T.S.: Ultrasonic characterization of defects in steel using multi-SAFT
imaging and neural networks. NDT and E Int. 26(3), 127–133 (1993)
7. Liu, W.J., Wang, X.M.: Study on the defect determination method of ultrasonic
nondestructive testing based on neural network. J. Dalian Univ. Technol. 5, 548–552
(1998). (in Chinese)
8. Yang, L.J., Ma, F.M., Gao, S.W.: Quantitative recognition of pipeline defects based on
neural network and data fusion. Nondestr. Test. 28(6), 281–284 (2006). (in Chinese)
9. Dai, G., Qiu, F., Chen, R.G., et al.: Artificial neural network intelligent evaluation method of
tank bottom corrosion status. Nondestr. Test. 34(6), 5–7 (2012). (in Chinese)
10. Wang, P., Zhu, L., Zhu, Q., et al.: An application of back propagation neural network for the
steel stress detection based on Barkhausen noise theory. NDT and E Int. 55(3), 9–14 (2013)
 Research on Joint Nondestructive Testing Based on Neural Network 467

11. Haykin, S.: Neural Network and Learning Machines. Pearson, London (2008)
12. Arel, L., Rose, D.C., Karnowski, T.P.: Deep machine learning a new frontier in artificial
intelligence research. Comput. Intell. Mag. 5(4), 13–18 (2010)
13. Deng, L., Yu, D.: Deep Learning: Method and Application. China Machine Press, Beijing
(2015). (in Chinese)
14. Yu, D., Deng, L., Seide, F.T.B., et al.: Discriminative pretraining of deep neural networks.
US, US9235799 (2016)
15. Zheng, W.H., Wang, Z.H.: Ant Colony algorithm and application in inspection of concrete
structure defects. Nondestr. Test. 35(5), 4–7 (2013). (in Chinese)
16. Yan, W.X.: Research on deep learning and its application on the casting defects automatic
detection. South China University of Technology (2016). (in Chinese)
17. Lu, H.: Study on fault diagnosis of flexible manufacturing system based on neural network.
Chang’an University, Xi’an, China (2016). (in Chinese)
 A New Real-Time FPGA-Based
Implementation of K-Means Clustering
for Images

Tiantai Deng(&), Danny Crookes(&), Fahad Siddiqui,

and Roger Woods

Queen’s University Belfast, University Road, Belfast, UK

{tdeng01,d.crookes,f.siddiqui,r.woods}@qub.ac.uk

Abstract. As an unsupervised machine-learning algorithm, K-means clustering

for images has been widely used in image segmentation. The standard Lloyd’s
algorithm iteratively allocates all image pixels to clusters until convergence. The
processing requirement can be a problem for high-resolution images and/or real-
time systems. In this paper, we present a new histogram-based algorithm for K-
means clustering, and its FPGA implementation. Once the histogram has been
constructed, the algorithm is O(GL) for each iteration, where GL is the number
of grey levels.
On a Xilinx ZedBoard, our algorithm achieves 140 FPS (640
480 images,
running at 150 MHz, 4 clusters, 25 iterations), including final image recon-
struction. At 100 MHz, it achieves 95 FPS. It is 7.6 times faster than the
standard Lloyd’s algorithm, but uses only approximately half of the resources,
while giving the same results. The more iterations, the bigger the speed-up. For
50 iterations, our algorithm is 10.2 times faster than the Lloyd’s approach. Thus
for all cases our algorithm achieves real time performance whereas Lloyd’s
struggles to do so. The number of clusters (up to a user-defined limit) and the
initialization method (one of three) can be selected at runtime.

Keywords: Unsupervised machine learning  Data processing

K-means clustering  FPGA acceleration

1 Introduction

Clustering problems arise in many applications [1]. K-means clustering is an important

pre-processing step in applications such as medical image segmentation [2], pattern
recognition, machine learning, and bio-informatics [3, 4]. K-means clustering is the
process of grouping elements of a data set into several subsets, based on minimizing the
total Euclidean distance between cluster elements and the cluster means. It is an
unsupervised computational clustering algorithm to distinguish different objects or
regions in the image. The standard algorithm for K-means clustering is Lloyd’s
algorithm, which iteratively adjusts the clusters until convergence (see Sect. 3). The
common implementation platform is a CPU. Lloyd’s algorithm can be efficiently
coded.

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 468–477, 2018.
https://doi.org/10.1007/978-981-13-2384-3_44
 A New Real-Time FPGA-Based Implementation of K-Means Clustering 469

However, when real-time processing is required, for videos or for images with very
high resolution, achieving real time performance of the standard Lloyd’s algorithm
becomes a problem, because of the iterative processing of the complete image [5].
Thus, an alternative way of implementing K-Means clustering is needed for real-time
processing. The standard Lloyd’s algorithm is naturally parallel [6]. We can improve
the performance either by using different hardware (e.g. multiple CPUs, or a GPU, or
an FPGA), or by using a faster algorithm. In this paper, we propose a new imple-
mentation for implementing K-means clustering which achieves significant speed-up
both by algorithm innovation and by hardware implementation on an FPGA.
The FPGA architecture can achieve real time performance for video processing or for
large image resolutions. The algorithm is considerably faster than the standard Lloyd’s
algorithm. Indeed, the algorithm gives significant speed-up even on a CPU. The main
contributions of the paper are as follows:
(1) A new histogram-based implementation of K-means clustering. Our method gives
the same result as the standard Lloyd’s algorithm for image data, but is much
more efficient.
(2) An FPGA implementation on the Xilinx ZedBoard. Our method is 7.6 times faster
but uses 50% of the resources compared to the Lloyd’s FPGA implementation.
(3) A comparison of two different FPGA implementations: our new Histogram-based
K-means method, and the standard Lloyd’s K-means method, plus a performance
comparison with three FPGA implementations from the literature.

2 Related Work
A. Standard Lloyd’s Algorithm for K-means Clustering
The K-means problem is to find K clusters such that the means lc of each cluster
Clc satisfy Eq. (1).

K X
X
arg min kpx  lc k2 ð0  c \ K Þ ð1Þ
Clc c¼1 px 2Clc

The standard Lloyd’s algorithm is an iterative algorithm to solve the K-means

problem. The data set needs to be processed several times until the minimum distance
over all clusters is achieved. There are two steps in Lloyd’s algorithm: The first is to
assign the elements of the data set to the cluster with the closest mean. The second is to
update the cluster means. These steps are defined by Eqs. (2) and (3), for iteration
t. The typical pseudocode is shown in Fig. 1.
The Assignment Step:
 
 2 
2

Cltc ¼ px : Px  ltc   Px  ltj  8j; 0  j \ K ð2Þ
470 T. Deng et al.

The Update Step:

1 X
ltcþ 1 ¼  
Clt  pClc ð3Þ
c pi 2Clt
c

Pseudocode for Standard Lloyd’s Pseudocode for Histogram-based

Fig. 1. Pseudocode for the standard Lloyd’s algorithm

Since the algorithm needs to access the data set to calculate the means and distances
at each iteration, it is necessary to keep a copy of the original data set in memory
throughout the FPGA implementation.
B. FPGA Implementation of K-means Clustering
A number of researchers have already studied the FPGA implementation of K-
means clustering. In [7], Lavenier feeds the image into the distance calculation units,
and indexes every pixel in another array. Due to limited resources on the FPGA, the
image is stored in off-chip memory, and must then be streamed multiple times. This
entails extra communication overheads between the FPGA and off-chip memory.
Estlick et al. implemented K-Means in hardware using software/hardware co-design.
The distance calculation is computed in hardware and the rest of the algorithm is
completed on a CPU to avoid using a huge amount of hardware resources [8]. And in
[4], Yuk-Ming and Choi etc. use three FPGAs which work together to deal with the
large amount of data because of the limited resources on each FPGA. In [9], Benkrid
implemented a highly parameterized parallel system, and his implementation is not
dedicated to image clustering but shows the potential of using FPGAs to accelerate the
clustering process. Takashi, Tsutomu et al. implemented K-means on a Xilinx FPGA,
using KD-Tree, but they are limited by the resources on the FPGA, and they have to
store the image in an off-chip SRAM. They report a performance of 30FPS on an RGB
24-bit image [10, 11]. In [11], the approach replaces the Euclidean distance into the
multiply-less Manhattan distance and reduces the word length to save resources. The
approaches above focus on achieving greater efficiency by adjusting the word-length,
simplifying the distance calculation or using a heterogeneous platform.
Others try to reduce the amount of data. In [1], a filter algorithm is applied first
before K-means to reduce the amount of data. Tsutomu reduces the scanning data size
by doing partial K-means, whereby the algorithm is applied on 1/8, 1/4, and 1/2
sampled data. This simplifies the K-means algorithm to get better performance but also
introduces inaccuracy into the result [12].
 A New Real-Time FPGA-Based Implementation of K-Means Clustering 471

Another way to speed up Lloyd’s algorithm is to reduce the number of iterations.

The performance of Lloyd’s is very dependent on the starting point. In [13], Khan and
Ahmad used a small amount of data to execute K-means to get a better estimate of the
initial mean for each cluster. Pena, Lozano, and Larranaga compared four classical
initialization methods for the K-means algorithm: Random, Forgy, MacQueen, and
Kaufman. The author found that the Random and Kaufman initialization make the
algorithm more effective and independent of the initial value which means it is more
robust and effective [14].
The Soft-processor approach is another way of using FPGAs. This retains the high
level programming approach, and hides much of the complexity of hardware from
developers. In [15], a multi-processor architecture is proposed for K-means based on a
network on chip, and using a shared memory to support communication between
processors.
To summarize the above approaches, there are three existing ways to speed up the
process of K-means clustering using an FPGA. One is to focus on the system archi-
tecture, parallelizing the system and finding more effective ways to allocate memory,
such as in [7, 9, 10]. The second way is to reduce the amount of data, such as in [1, 12].
The third way is to reduce the number of iterations by finding better initial means for
each cluster, as in [13, 14].
We now present our novel approach to implementing K-means clustering, and then
present its implementation on a single chip of FPGA including camera and display
control.

3 Histogram-Based K-Means Clustering

We first note that the K-means problem for image processing is more constrained than
K-means for other kinds of data. The range of the input data is discrete, e.g. for an 8-bit
grayscale image, the grey levels are from 0 to 255.
Secondly, the clustering of pixels does not need to take the location of pixels into
account. Therefore, we can represent the input image by the image histogram (for the
purposes of clustering). The image histogram (which does not change throughout the
iterative clustering process) is effectively a compressed version of the image. For any
grey level value g, all pixels with that value can be clustered at the same time, with a
single calculation.
Thirdly, we can represent each cluster c merely as a sub-range of the histogram,
from column Bc to column B(c+1)−1 inclusive, where Bi is the lower bound of cluster i.
Fourthly, given the means of each cluster lc, the new bounds, which define the new
clusters, will be simply half way between the respective means. Figure 2 shows a
typical histogram and the representation of four clusters.
For an Nx
Ny image, and doing L iterations, the complexity of the standard
Lloyd’s algorithm is Nx
Ny
L. The complexity of our new histogram-based
algorithm will be Nx
Ny + 256
L (assuming 256 grey levels), where the Nx
Ny
component is for calculating the histogram. For example, for a 512
512 image with
L = 20, the standard method will require around 5.2M steps, whereas our histogram-
based method will require around 0.267M steps – an improvement, in theory, of the
472 T. Deng et al.

Fig. 2. An image histogram and representation of four clusters

order of 20x. In practice, there will be the additional overhead of image input and
perhaps reconstruction of the output image and the number of iterations will be image-
dependent. These factors will typically have the effect of reducing the apparent speed-
up.
For an 8-bit image, the histogram is a one-dimensional vector with 256 elements,
where each element in the histogram represents the number of pixels having that grey
level. After generating the histogram, bounds are needed to divide the histogram into
different clusters. Each cluster is defined by a pair of bounds, which define the part of
the histogram corresponding to the cluster. If we have K clusters, there will be K + 1
bounds, where B0 = 0, and BK = 256. Cluster c (0  c < K) is the part of the his-
togram from column Bc to B(c+1)-1 inclusive. The inner bounds (0 < c < K) can be
initialized in different ways (see later). Then, the means for the current clusters are
calculated as defined by Eq. (4).

BCP
þ 1 1
ð Hi
iÞ
i¼BC
lc ¼ BCP
ð0\C\KÞ ð4Þ
þ 1 1
ð Hi Þ
i¼BC

Once the updated means are calculated, the next iteration first assigns the clusters
by merely updating the bounds. To assign a pixel to the correct new cluster, we have to
see which mean it is closest to. Consider a pixel with value g (say, 150) in an image
with a histogram such as in Fig. 2, and the means are [0.0, 27.5, 90.5, 165.0, 220.0,
256.0]. Column g will be part of the cluster whose mean is closest to g. We can achieve
this simply by moving the bounds so that each bound is half way between the two
respective means. The updating of the clusters is defined by Eq. (5).
 A New Real-Time FPGA-Based Implementation of K-Means Clustering 473

ðlC1 þ lC Þ
BC ¼ ð0  C\K Þ ð5Þ
2
The iteration can then be completed by updating the means using Eq. (4).
We use integers to hold the bounds, so that they correspond to actual pixel grey
levels. In our FPGA implementation, we use simple fixed point to represent the means.
Termination of the iterative algorithm could be brought about under several con-
ditions. The total distance can be readily calculated as in the standard Lloyd’s algo-
rithm. However, in our preferred approach, the termination condition is that none of the
bounds are changed by the updating (Eq. (5)). This has the advantage that the total
(squared) distance does not need to be calculated. A third option is to set a fixed
number of iterations (say, 50), which can be sufficient for many applications in practice.
This number may be application–specific, and can be set at runtime.
There are several ways to initialize the K-means algorithm such as Random, or
MacQueen and Kaufman, as mentioned in [14]. Actually, in our histogram method, we
initialize the cluster bounds rather than the means. We provide two particular ways to
do this. The first is simply to divide the grey level range into K equal parts. The second
way is to divide all the pixels into K equally sized clusters. The latter can be achieved
by using the cumulative histogram and setting the bounds to the columns which have
the appropriate height.
Dividing the histogram range into K equal parts takes the least resources in a
hardware implementation. However, in an image with an unevenly distributed his-
togram, some clusters may be empty (for instance, if the image is overexposed or too
dark). The danger condition of calculating the mean of a zero-element cluster can be
overcome with a suitable condition; however, if we have two adjacent zero-element
clusters, the bounds would not move and those clusters would not grow. Thus, the
safest way to initialize the clusters is to set each cluster initially to have the same
number of pixels (approximately), using the cumulative histogram. We can build the
cumulative histogram from the histogram during initialization. A refinement which is
easy to implement is for the user to supply an initial, application-specific estimate of the
relative sizes of each cluster as a parameter.

4 System Implementation

We implemented the histogram-based K-means Clustering algorithm on a Xil-

inx FPGA, using Vivado HLS on a ZedBoard and running at either 150 MHz or
100 MHz. And the architecture is shown in the Fig. 3 below.
The architecture of the FPGA implementation is simple. The system is made up of
two parts, the I/O part and the processing part. The whole system is controlled by the
AXI-Lite bus. The camera repeatedly captures a frame and stores it in the frame buffer.
Then the data is streamed to the input FIFO and fed into the K-means hardware. After
the image is processed, the bounds and means are sent to the image reconstruction
hardware and the output frame buffer. We use a physical VGA interface for display.
The architecture is shown in Fig. 3. The K-means hardware and the image recon-
struction hardware are programmed in HLS. The input and output images are
474 T. Deng et al.

Fig. 3. System architecture

transferred using the AXI-Stream interface. The means and bounds of each cluster are
held in a BRAM.
Our architecture is parameterized to provide some flexibility for the user. Our
K-means hardware can support any number of clusters up to a defined maximum (in
our case, 10 clusters). Within this limit, the actual number of clusters can be selected,
and changed during the run time by setting the control registers in the control bus.
Extending the maximum number requires re-synthesis the hardware and pay more
hardware resources.
The method for setting the initial value of the bounds, as mentioned above, can be
selected at runtime via a control register as well.
For comparison purposes, we run both the Lloyd’s and our K-means at both
100 MHz and 150 MHz. The ARM output 100/150 MHz clock is divided into
50 MHz and 25 MHz. 50 MHz is used in the camera and display, and 25 MHz is used
in the data transfer through the FIFOs and frame buffers. The FIFOs are used to help
synchronise data across the clock domain between the frame buffer and our K-means
hardware.
Both Lloyd’s and our histogram-based algorithm can run at 150 MHz. However,
the hardware resources for both these higher speed implementations are noticeably
higher than for the 100 MHz version. In the next section, we discuss the performance
and resources in detail.

5 Results and Comparison

We compare the performance of two designs: the standard Lloyd’s algorithm and our
new histogram-based algorithm. For the hardware versions, we consider several fac-
tors: hardware usage, the clock cycles required to process an image (4 clusters), and the
 A New Real-Time FPGA-Based Implementation of K-Means Clustering 475

actual number of frames per second (FPS). The utilization of the different designs is
shown in Table 1. It can be seen that the histogram-based algorithm takes about half
the LUT (Look up table) and FF (Flip-Flop) resources.

Table 1. Comparison of performance (including reconstruction of output image)

100 MHz LUTs FFs DSPs BRAMs
Histogram 1194 749 1 40.5
Lloyd’s 2641 1535 3 38.5
150 MHz LUTs FFs DSPs BRAMs
Histogram 1807 2770 4 40.5
Lloyd’s 3793 4350 3 38.5

Since even at 100 MHz our histogram-based architecture is able to achieve real
time performance, the clustering itself does not demand using the fastest clock speed;
and since the 100 MHz version uses rather less hardware resources, its use would
therefore normally be preferred (Table 2).

Table 2. Comparison with other literatures

100 MHz, L = 25 FPS Iterations
Histogram (HW) 95 25
Lloyd’s (HW) 12.5 25
[13] 30 –
[14] 38.6 –
[12] 45.8 –

We also compare the performance of our algorithm to the implementations in [13–

15]. The work in [14] takes a 24-bit RGB image of size 640
480 and has a per-
formance of 30 FPS; the pixels in RGB channel are clustered in parallel. And the work
in [13, 15] is using 640
480 grayscale image. [15] uses the Manhattan distance to
replace Euclidean distance, and [13] reduces the data set by sampling the image to get
better performance. Those two approaches are based on the standard Lloyd’s algorithm
but cannot guarantee full accuracy, and have less performance than our new imple-
mentation. The superiority of our new method can be seen.
Hence, the new histogram-based algorithm is the best choice when multi-clusters is
needed. It has more advantages over the existing algorithms, it takes fewer resources,
better performance than full K-means, and it is easy to adapt the stream
implementation.
476 T. Deng et al.

6 Conclusion and Future Work

This paper has presented a new K-means clustering algorithm, based on iteratively
processing the image histogram rather than the image. The main contributions are as
follows:
1. A new histogram-based K-means clustering algorithm, which transforms the 2-D
image into a 1-D histogram. The histogram retains all the details which are needed
to perform K-means clustering. The method is equivalent to Lloyd’s, and gives the
same results.
2. An FPGA implementation of the new histogram-based K-means clustering algo-
rithm. On an FPGA, the histogram-based takes approximately half the resources
compared to Lloyd’s, and a speed-up of approximately 7.6. Even at 100 MHz and
with 50 iterations, we achieve 82 FPS for 640
480 images.
Future work will look at providing greater flexibility in the FPGA architecture,
including a wider range of cluster initialization strategies, and some possible further
optimizations. Using the histogram-based algorithm to cluster RGB color image will be
challengeable, we are going to find a way to optimize the algorithm and make it
suitable for RGB color image.

Acknowledgement. This work was supported by the Chinese Scholarship Council.

References
1. Kanungo, T., Mount, D.M., Netanyahu, N.S., et al.: An efficient K-means clustering
algorithm: Analysis and implementation. IEEE Trans. Pattern Anal. Mach. Intell. 24(7),
881–892 (2002)
2. Ng, H.P., Ong, S.H., Foong, K.W.C., et al.: Medical image segmentation using K-means
clustering and improved watershed algorithm. In: Proceedings of 2006 IEEE Southwest
Symposium on Image Analysis and Interpretation, pp. 61–65. IEEE (2006)
3. Duda, R.O., Hart, P.E., Stork, D.G.: Pattern Classification and Scene Analysis, 2nd edn.
Wiley Interscience, Hoboken (1995)
4. Choi, Y.M., So, H.K.H.: Map-reduce processing of K-means algorithm with FPGA-
accelerated computer cluster. In: Proceedings of the 2014 IEEE 25th International
Conference on Application-Specific Systems, Architectures and Processors, ASAP, pp.
9–16. IEEE (2014)
5. Neshatpour, K., Koohi, A., Farahmand, F., et al.: Big biomedical image processing hardware
acceleration: a case study for K-means and image filtering. In: 2016 IEEE International
Symposium on Circuits and Systems, ISCAS, pp. 1134–1137. IEEE (2016)
6. Farivar, R., Rebolledo, D., Chan, E., et al.: A parallel implementation of K-means clustering
on GPUs. In: PDPTA 2008, vol. 13, no. 2, pp. 212–312 (2008)
7. Lavenier, D.: FPGA implementation of the K-means clustering algorithm for hyperspectral
images. In: Proceedings of Los Alamos National Laboratory LAUR (2000)
8. Estlick, M., Leeser, M., Theiler, J., et al.: Algorithmic transformations in the implementation
of K-means clustering on reconfigurable hardware. In: Proceedings of the 2001 ACM/SIGDA
Ninth International Symposium on Field Programmable Gate Arrays, pp. 103–110. ACM
(2001)
 A New Real-Time FPGA-Based Implementation of K-Means Clustering 477

9. Hussain, H.M., Benkrid, K., Erdogan, A.T., et al.: Highly parameterized K-means clustering
on FPGAs: comparative results with GPPs and GPUs. In: Proceedings of 2011 International
Conference on Reconfigurable Computing and FPGAs, ReConFig, pp. 475–480. IEEE
(2011)
10. Saegusa, T., Maruyama, T., Yamaguchi, Y.: How fast is an FPGA in image processing? In:
Proceedings of International Conference on Field Programmable Logic and Applications,
FPL 2008, pp. 77–82. IEEE (2008)
11. Saegusa, T., Maruyama, T.: An FPGA implementation of K-means clustering for color
images based on Kd-tree. In: Proceedings of International Conference on Field Pro-
grammable Logic and Applications, FPL 2006, pp. 1–6. IEEE (2006)
12. Maruyama, T.: Real-time K-means clustering for color images on reconfigurable hardware.
In: Proceedings of 18th International Conference on Pattern Recognition, ICPR 2006, vol. 2,
pp. 816–819. IEEE (2006)
13. Khan, S.S., Ahmad, A.: Cluster center initialization algorithm for K-means clustering.
Pattern Recogn. Lett. 25(11), 1293–1302 (2004)
14. Pena, J.M., Lozano, J.A., Larranaga, P.: An empirical comparison of four initialization
methods for the K-means algorithm. Pattern Recogn. Lett. 20(10), 1027–1040 (1999)
15. Khawaja, S.G., Akram, M.U., Khan, S.A., et al.: A novel multiprocessor architecture for
K-means clustering algorithm based on network-on-chip. In: Proceedings of 2016 19th
International Multi-topic Conference, INMIC, pp. 1–5. IEEE (2016)
 Neural Network Identification of an Axial
Zero-Bias Magnetic Bearing

Qing Liu1, Li Wang1(&), and Yulong Ding1,2(&)

1
School of Energy and Environmental Engineering, University of Science and
Technology Beijing, Beijing 100083, China
[email protected], [email protected]
2
Birmingham Center for Energy Storage, School of Chemical Engineering,
University of Birmingham, Birmingham B15 2TT, UK

Abstract. System identification for magnetic bearings can be used for various
purposes, such as aggressive control design, on-line monitoring and early
diagnosis of upcoming faults in rotating machinery, etc. Especially for the zero-
bias active magnetic bearing (AMB) system, which omits the bias current/flux to
reduce power loss and rotor heating, whereas brings out the problem of highly
nonlinearity. Here, we applied the NARX neural network to the identification of
an axial zero-bias magnetic bearing system. The I/O data are acquired through a
stabilized magnetic bearing test rig, which are more accurate than data that
generated by simulation. Two kinds of training methods are compared in dif-
ferent working conditions. The results show a good ability of generalization,
which prove the feasibility of the NN identification.

Keywords: Zero-bias
Magnetic bearing
Neutral networks

System identification

1 Introduction

Active magnetic bearings (AMBs) generate controlled electromagnetic forces to levi-

tate the rotor to balancing positions (Fig. 1), thus there is no mechanical contact or
abrasion to the magnetic bearings. Therefore, AMB provides many unique and prac-
tical merits compared with traditional bearings, such as free of lubrication, no con-
tamination, higher rotating speed and higher precision. Consequently, AMBs have been
widely utilized in various fields, including high-speed electrical spindle, flywheel
energy storage systems, turbine engines and vacuum pumps. In addition, the apparent
noncontact and frictionless characteristics make the AMB irreplaceable at harsh
working conditions such as in extreme temperature or vacuum environment. Further,
both the bearing stiffness and damping are controllable which make the AMB adjus-
table in case of sudden imbalance or vibration.
Due to the intrinsic nonlinear characteristics of magnetic forces, a constant current
or flux is supplied to each electromagnet (EM), which is referred as the bias
current/flux, to improve the linearity of the force-current relationships around the
operating point and allow for a higher slew rate of force. Meanwhile, a variable
current/flux which is calculated by the controller is superimposed on the bias

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 478–488, 2018.
https://doi.org/10.1007/978-981-13-2384-3_45
 Neural Network Identification of an Axial Zero-Bias 479

Fig. 1. Schematic of one-DOF active magnetic bearings

current/flux, referring as the control current/flux. Although a bias current/flux linearizes

the relationship between the force and the control current, it also results in rotor heating
and power loss [1] (the copper loss in the windings, eddy-current loss and hysteresis
loss). In certain situations, large bias current/flux is designed to improve bearing
stiffness and load capacity, yet causes huge copper losses and cuts down the running
time of the AMB system [2].
Based on this, there are several viable drive modes to reduce power loss of the
AMBs. For one, the bias currents/fluxes are lowered and the control current is merely
superimposed on one side of the pair of coils. Consequently, the energy consumption
can be reduced efficiently. Nevertheless, this operating mode is merely appropriate for
low bearing stiffness and low vibration applications due to poor slew rate and con-
trollability of the force. For the other, there is no bias current/flux and only the control
current/flux is exerted on one side of the pair of coils, which is also called the zero-bias
AMB. In spite of the least rotor heating of all kinds, high nonlinearity of the elec-
tromagnetic force degrades the control characteristic greatly.
According to the highly nonlinearity of the zero-bias magnetic bearings, precise
modelling is of essence for several reasons. First is that it enables prediction of the
system dynamics under different working situations. Then, identification for AMB
systems could be used for on-line monitoring and early diagnosis of upcoming faults in
rotating machinery [3]. Next, a concise model of the zero-bias AMB plant is required
for the controller design [4]. Otherwise the controllers have to be robust with respect to
the plant uncertainty [5].
Generally, the dynamic model of the AMB is established on the basis of theoretical
knowledge, which mainly relies on the identification of the model parameters. For
example, the linearized magnetic force around the equilibrium position is f ¼ ki i þ kx x,
where ki and kx are the current-force coefficient and the displacement-force coefficient,
which are related to coil turns, geometry, nominal air gap and the bias current.
Moreover, the dynamics of the flexible rotor is calculated by FE model. Nonetheless,
this kind of methods are not accurate enough for controller design. Besides, many parts
are not covered, including sensors, power amplifiers, eddy current and hysteresis
effects.
Apart from that, the linearized model based on Taylor expansion around the
equilibrium position is no longer suitable for the zero-bias AMBs. Accordingly,
480 Q. Liu et al.

effective ways to identify the nonlinear system of zero-bias AMB is of significance to

controller design. Here, we apply neural networks (NN) to the modelling of zero-bias
magnetic bearings.

2 The One-DOF Zero-Bias Magnetic Bearing

As shown in Fig. 2, the zero-bias AMBs are comprised of the rotor, displacement
sensors, the controller, power amplifiers and electromagnets. The displacement sensors
detect the rotor position and send the position signals to the controller. After AD
conversion and data filtering, the controller calculates the control value and outputs it
as PWM wave. The power amplifiers connect the controller with the electromagnets,
which act as an electronic switch and meanwhile drive the EM.

Fig. 2. Block diagram of zero-bias AMBs

Compared with conventional AMBs, there is no constant bias current and only one
electromagnet being electrified at any moment. In an attractive type bearing configu-
ration of electromagnets demonstrated in Fig. 1, when the rotor is deviating from the
balanced position due to a disturbance, the coil current of the EM that the rotor is
approaching reduces to zero, while the other EM on the opposite is switched on to
produce attractive magnetic force to drag the rotor back to the equilibrium position and
stabilized. The motion equation for the one-DOF zero-bias AMB system is derived as

m€x ¼ f1  f2 : ð1Þ

where m is the rotor mass and f1 , f2 are the nonlinear control force of each EM, which
can be denoted as

ki21 ki22
f1 ¼ 2
f2 ¼ : ð2Þ
ðx0 þ xÞ ðx0  xÞ2

where k is a constant ð¼ l0 N 2 S=4Þ, i1 ; i2 are the EM’s currents, x0 is the nominal air
gap and x is the rotor displacement. The total force to the rotor f is

f1 ðx [ 0; i2 ¼ 0Þ
f ¼ : ð3Þ
f2 ðx\0; i1 ¼ 0Þ
 Neural Network Identification of an Axial Zero-Bias 481

As demonstrated, the relationship between the magnetic force and the control
current is nonlinear. Meanwhile, only one control input is effective at any moment
which is depended on the rotor position [6, 7].

3 Architecture and Algorithm of the NARX Network

System identification is a process of estimating parameters of a system model, which

also allows fast and accurate modelling of magnetic bearings [5]. Identification for
AMB systems could be used for on-line monitoring and early diagnosis of upcoming
faults in rotating machinery [3]. Artificial intelligence theory such as neural networks
and fuzzy logic, have also been used in identification and control of AMBs [8].
Compared with classical system identification methods, such as prediction error
(PE) method, the output error (OE) method, the instrumental variable (IV) method [9],
NN has an ability of approaching any continuous nonlinear mapping, therefore the
NN is generally applied to system identification and control.
NN comprises of many simple processing elements operating in parallel whose
function is determined by network structure, connection strength, and the processing
preformed at computing elements or nodes [10]. Common approaches such as BP, RBP
or SVM are static neural networks, which add TDL (Tapped Delay Line) in front of the
input to realize dynamic characteristics. Nonetheless, the TDL is not an internal
component of the neural networks, therefore increases the training complexity. Besides,
the order of the TDL is generally 2 order, which reduces the model accuracy. Hence,
dynamic neural networks are preferred. Here NARX network is presented since the
TDL unit is the interior structure of the NN and can be adjusted dynamically, therefore
reduce the training complexity. Besides, compared with another dynamic neural net-
works, i.e. NAR, which only uses past values of network output, NARX is more
accurate. The NARX model can be expressed as [11]

yðtÞ ¼ f ðuðt  1Þ; . . .; uðt  nu Þ; yðt  1Þ; . . .; yðt  ny ÞÞ: ð4Þ

where uðtÞ and yðtÞ denote the I/O of the network at time t, nu and ny represent the input
and output order, and function f is a nonlinear function. When function f can be
approximated by a multilayer perception, the resulting system is called a NARX net-
work. As shown in Fig. 3, the NARX network mainly consists of the input layer, the
hidden layer, the output layer, and the TDL. The output of the hidden layer i at time
k can be described as

X
nu ny
X
pi ðkÞ ¼ fsig ð wir ðkÞuðk  rÞ þ wil ðkÞyðk  lÞ þ bi Þ: ð5Þ
r¼1 l¼1

where fsig is the Sigmoid function of the hidden layer, nu is the length of time delay of
the external input, and ny is the length of time delay of the output feedback. wir ðkÞ is
the weight value between the input u(t − r) and the hidden unit i at time k; wil ðkÞ is the
weight value between the output feedback y(t − l) and the hidden unit i at time k; bi is
482 Q. Liu et al.

Fig. 3. The structure of NARX neural network model

the threshold value of the unit i. Hence, the output yj ðkÞ of the output layer j can be
expressed as

X
N
yi ðkÞ ¼ wji ðkÞdi þ Tj : ð6Þ
i¼1

where wji ðkÞ is the weight value between the hidden unit i and the output layer j; Tj is
the threshold value of the output layer j; N is the number of the hidden units. Therefore,
the expression of NARX network can be expressed by combining Eqs. (5) and (6).
The NN configurations are demonstrated in Fig. 4. There are two kinds of con-
figurations. One is called parallel architecture and the other is series-parallel archi-
tecture. For the former structure, the output of the NARX network is fed back to the
input of the feedforward neural network as part of the standard NARX architecture. For
the latter one, the true output is used instead of feeding back the estimated output,
which possesses two advantages. Firstly, the input to the feedforward network is more
accurate. Secondly, the resulting network has a purely feedforward architecture, and
static backpropagation can be used for training.

Fig. 4. Two kinds of architectures of NARX model

 Neural Network Identification of an Axial Zero-Bias 483

As shown in Fig. 4, the aim of the NN training algorithm is to minimize the error
between the sample output and the corresponding NN model output estimation. The
training algorithm applies some parameter-searching strategy, such as the Levenberg-
Marquardt (LM) and back-propagation (BP) training function with momentum
weight/bias learning function, to adjust the parameters of all neurons in the NN. These
parameters mainly contain two groups, i.e. bias and weight.
The performance criteria for the network model can be classified into two types:
mean square error (MSE) and error range, which can be described as follows

1X N
^ 2
MSE ¼ ðyðiÞ  yðiÞÞ : ð7Þ
N i¼1

^ ^
Dd = [min(yðiÞ  yðiÞÞ; maxðyðiÞ  yðiÞÞ: ð8Þ

^
where yðiÞ is the output of the zero-bias AMB system, and yðiÞ is the prediction of the
model output. N is the number of the samples. The smaller the MSE and Dd are, the
more accurate the network model is.
As for the training algorithms for the NN, there are many choices. Here we compare
two kinds of algorithms, i.e. the LM and Scaled Conjugate Gradient (SCG). The LM
interpolates between the Gauss–Newton algorithm (GN) and the method of gradient
descent. For well-behaved functions and reasonable starting parameters, the LM tends
to be a bit slower than the GN. Nevertheless, the LM is more robust than the GN, which
means that in many cases it finds a solution even if it starts very far off the final
minimum. In addition, a supervised learning algorithm (SCG) with super linear con-
vergence rate is applied, which is based upon Conjugate Gradient Methods. SCG uses
second order information from the neural network yet demands only O(N) memory
usage (N is the number of weights in the network) [12].

4 Simulation Results

In this section, system identification was applied to a one-DOF zero-bias magnetic

bearing test rig based on different working conditions. Moreover, different training
algorithms are compared.
The AMB plant is open loop unstable, thus to fulfil system identification, it must be
stabilized [13]. An embedded control system was implemented with a micro-controller
(Fig. 5). The test rig is comprised of two EMs, a rotor, a displacement sensor, a
controller and two power amplifiers. The EMs are two small attracted-disk type with EI
shaped cores. The rotor is an iron disc with wall thickness 2.5 mm and diameter
60 mm. The displacement sensor is eddy current type with an output range of 0–10v,
1%FS. The controller is based on K60 microprocessors (Freescale 32-bit), and the
control algorithm is designed in IAR development environment. In addition, two
switching power amplifiers which mainly constitute of MOSFET (IRF540N) are
applied to actuate the EMs.
484 Q. Liu et al.

Fig. 5. The test rig of the one-DOF zero-bias AMB

Two group of data were acquired, the first was under the steady state with 2000
timesteps and a sampling time of 1 ms (Fig. 6). The second group covered an external
disturbance and the relaxation process, which contained 6000 timesteps with a sam-
pling time of 1 ms (Fig. 10). The number of hidden neurons is ten, and both the input
order and output order of the NARX model are two.

Fig. 6. Snapshot of: (a) input data sequences; (b) output data sequences (steady state)

The simulation results of the steady state were displayed in Figs. 7 and 8. The total
time of training, validation, and test for LM is 1 s, whereas that of the SCG is more
than 2 s. Specifically, Fig. 7. is about the MSE, which represents the average squared
difference between actual output and the targets [14]. Lower values are better, and zero
means no error. Figure 8 is about the error histogram. Obviously, the majority of the
error concentrated on zero, however, the error distribution was more dispersed using
LM. Moreover, the overall regression (R) values for LM is 0.97065, while that of SCG
is 0.9552. The regression measures the correlation between the output and target. An R
value of 1 means a close relationship, 0 a random relationship. Overall, LM has a far
 Neural Network Identification of an Axial Zero-Bias 485

faster convergence speed and smaller regression value, whereas the epoch iteration is
more. Besides, both algorithms are sensitive to initial parameters.

Fig. 7. Mean squared error and epochs number of the steady state: (a) LM; (b) SCG

Fig. 8. Error histogram with 20 bins (steady state): (a) LM; (b) SCG

As for the second group, the simulation results of the steady state are demonstrated
in Figs. 9, 10, 11 and 12. The total time of training, validation, and test for LM is
around 10 s, whereas that of the SCG is more than 20 s. The LM algorithm has smaller
MSE and regression value compared with SCG. Overall, both algorithm can approach
the nonlinear system very well. Although the LM requires large calculation, it prevails
in convergence speed and regression.
486 Q. Liu et al.

Fig. 9. A snapshot of: (a) input data sequences; (b) output data sequences (under disturbance)

Fig. 10. Mean squared error and epochs number of: (a) LM; (b) SCG (under disturbance)

Fig. 11. Error histogram with 20 bins (under disturbance): (a) LM; (b) SCG
 Neural Network Identification of an Axial Zero-Bias 487

Fig. 12. The regression values of the training, validation, test process (under disturbance):
(a) LM; (b) SCG

5 Conclusion

Zero-bias magnetic bearing is a highly-nonlinear system and the design of the con-
troller needs to include the nonlinear characteristics. Based on the fact that the NN can
approach any nonlinear mapping, here we present the system identification of the zero-
bias AMB with a NARXNET. Two kinds of training algorithms are compared in
different operation conditions. The results show the feasibility and effectiveness of NN
identification.

References
1. Kato, Y., Yoshida, T., Ohniwa, K.: Self-sensing active magnetic bearings with zero-bias
current control. Electr. Eeg. Jpn. 165, 69–76 (2008)
2. Rong, H., Zhou, K.: Nonlinear zero-bias current control for active magnetic bearing in power
magnetically levitated spindle based on adaptive backstepping sliding mode approach. Proc.
Inst. Mech. Eng. Part C: J. Mech. Eng. Sci. 231, 3753–3765 (2017)
3. Gahler, C., Mohler, M., Herzog, R.: Multivariable identification of active magnetic bearing
systems. JSME Int. J. Ser. C Mech. Syst. Mach. Elem. Manuf. 40, 584–592 (1997)
4. Cho, Y.M., Srinavasan, S., Oh, J.-H., Kim, H.S.: Modelling and system identification of
active magnetic bearing systems. Math. Comput. Model Dyn. 13, 125–142 (2007)
5. Gähler, C., Herzog, R.: Identification of magnetic bearing systems. Math. Model. Syst. 1,
29–45 (1995)
6. Sivrioglu, S., Saigo, M., Nonami, K.: Adaptive backstepping control design for a flywheel
zero-bias AMB system. In: Proceedings of 2003 IEEE Conference on Control Applications,
CCA 2003, pp. 1106–1111. IEEE (2003). B 7
488 Q. Liu et al.

7. Kato, Y., Yoshida, T., Ohniwa, K., Miyashita, O.: A self-sensing active magnetic bearing
with zero-bias-current control. In: European Conference on Power Electronics and
Applications, p. 9 (2005)
8. Mohd-Mokhtar, R., Liuping, W.: System identification of MIMO magnetic bearing via
continuous time and frequency response data. In: IEEE International Conference on
Mechatronics, ICM 2005, 10–12 July 2005, p. 191–196 (2005)
9. Srinivasan, S., Cho, Y.M.: Modeling and system identification of active magnetic bearing
systems. In: Proceedings of International Conference on Control Applications, pp. 252–260
(1995)
10. Jayaswal, P., Verma, S.N., et al.: Development of EBP-artificial neural network expert
system for rolling element bearing fault diagnosis. J. Vib. Cont. 17, 1131–1148 (2010)
11. Siegelmann, H.T., Horne, B.G., Giles, C.L.: Computational capabilities of recurrent NARX
neural networks. IEEE Trans. Syst. Man Cybern. Part B Cybern. 27, 208–215 (1997).
A Publication of the IEEE Systems Man & Cybernetics Society
12. Møller, M.F.: A scaled conjugate gradient algorithm for fast supervised learning. Neural
Netw. 6, 525–533 (1993)
13. Gibson, N.S., Heeju, C., Buckner, G.D.: H∞ control of active magnetic bearings using
artificial neural network identification of uncertainty. In: 2003 IEEE International
Conference on Systems, Man and Cybernetics, vol.1442, pp. 1449–1456 (2003)
14. Hamamreh, J., Dama, Y.: Self-organizing schedulers in LTE system for optimized pixel
throughput using neural network (2013)
Precision Measurement and
Instrumentation
 A Novel Bench of Quarter Vehicle Semi-active
Suspension

Xiaomin Dong1,2(&), Wenfeng Li1,2, Chengwang Pan1,2,

and Jun Xi1,2
1
State Key Laboratory of Mechanical Transmission, Chongqing University,
Chongqing, China
[email protected]
2
School of Mechanical Engineering, Chongqing University, Chongqing,
People’s Republic of China

Abstract. According to the requirements of automotive performance evalua-

tion standards, a typical quarter vehicle suspension model was established, and
the correlation between the suspension model transfer function and the vehicle
performance evaluation criteria was analyzed. Because the current lack of a
standard bench of car suspension, this paper has established a universal and
efficient standard bench of quarter car semi-active suspension. The bench is
composed of the hardware system part, the sensing scheme and software system
part. The bench can not only be used as a passive suspension development
platform, but also can be used as a semi-active suspension control test platform.
With the technology of Rapid Control Prototype (RCP), a variety of control
schemes for verifying the semi-active suspension can be compared to find a
suitable control algorithm, which is very significant for improving ride comfort
and safety of vehicle suspension. In addition, the bench is equipped with a force
sensor on the piston rod of the damper to monitor the real-time output force
value of the damper, and to achieve real-time tracking and correction of the
control output in the semi-active suspension control test.

Keywords: Quarter vehicle suspension model

Bench of quarter car suspension
RCP
Real-time tracking

1 Introduction

Suspension system is one of the most important assembly of modern automobiles. It is

the mechanism of elastic connection between the frame and the wheels. The main task
of suspension system is to transfer all forces and moments between the wheel and the
body. At the same time, it will buffer the impact load transmitted from the road to the
body, which causes the passenger’s discomfort. Suspension system has a significant
impact on the performance of a car such as ride stability, driving stability, ride comfort,
etc. [1]. Therefore, the suspension system has always been one of the most important
issues for car designers and researchers [2]. For the traditional design of automotive
passive suspension, the task of the matching between the stiffness coefficient and
damping coefficient requires much time and effort. An effective vehicle suspension

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 491–501, 2018.
https://doi.org/10.1007/978-981-13-2384-3_46
492 X. Dong et al.

bench is promising but difficult to achieve. Some researchers investigated the devel-
opment of vehicle suspension test rig, e.g., Liu et al. [3] constructed a performance
testing system that can measure the absorption efficiency, vibration frequency, phase
difference and corresponding vibration wave form of suspension. It provides a basis for
comprehensive evaluation and fault diagnosis of suspension performance. Based on the
existing automobile suspension test platform, Chen et al. [4] designed a high precision
force sensor and the adjustment module of the output signal of the speed sensor, which
improved the accuracy and real time of the test system. By simplifying the actual
vehicle, Geng et al. [5] established the physical model of the two degree of freedom
Mcpherson suspension system. On this basis, the research and construction of the
performance test bench for the Mcpherson suspension system were designed and
constructed. Omar et al. [6] designed a universal suspension test rig for electrohy-
draulic active and passive automotive suspension system, and demonstrated the
effectiveness of the platform through simulation and experimentation.
However, these suspension test rig has less adaptive ability. There is a difficulty in
tuning damping coefficient. They are not suitable to study the semi-active suspension
and control algorithms. Consequently, the contribution of this study is to develop a
bench for study vehicle suspension. According to the vehicle suspension performance
evaluation criteria, a two-degree-of-freedom vehicle suspension physical model was
developed. This bench can not only be used for passive suspension training to match
the test of passive vehicle suspension stiffness and damping, but also for semi-active
suspension testing. By adopting rapid control prototype (RCP) technology, the com-
parison of different control algorithms on the impact of suspension performance can be
realized, the cost and time of the control algorithm test can be saved. The effectiveness
of the quarter car suspension bench system was verified through experiments.
The organization of this study is as following. Test standard for vehicle suspension
performance is addressed in the Sect. 2. In Sect. 3, the theoretic analysis for the bench
test system of a quarter car suspension is analyzed. The design of the bench test system
of a quarter car suspension is proposed in the Sect. 4. The prototype test and discussion
are investigated in the Sect. 5. Section 6 concludes this study.

2 Study on Evaluation Standard of Automobile Suspension

Vehicle suspension evaluation methods can be roughly divided into subjective sensory
evaluation and objective physical quantity evaluation methods. Since subjective
evaluation methods vary from person to person, the discrete type is very large, and it is
unfavorable to explain the performance of the automotive suspension. With the
development of measurement technology and signal processing technology, the
objective physical quantity evaluation method has been widely used.
 A Novel Bench of Quarter Vehicle Semi-active Suspension 493

2.1 International Standard Evaluation Method of Vehicle Drive Stability

and Riding Comfort
The drive stability of the car is closely related to the ride comfort. The total weighted
acceleration root mean square (RMS) value is generally used to evaluate the car ride
stability in the international standard.
For the evaluation method of ride comfort, vibration dose values (EVDV) are
generally used in the international standard to evaluate human response to vibration.
When the crest factor is less than 12, the frequency-weighted acceleration RMS
value is used to evaluate the person’s response to vibration and to calculate the esti-
mated vibration dose value (EVDV).
 Z T 12
1
m=s1:75
1
EVDV ¼ 1:4
aw ðtÞdt
T 4
2
ð1Þ
T 0

where aw is the equivalent vibration value for the entire exposure time (m=s2 ); T is the
vibration analysis time (s).
When the crest factor is greater than 12, the method of average quadripartite value
can better estimate the severity of the high crest factor vibration, and vibration dose
value (VDV) can be obtained.
Z T 12
VDV ¼ a4w ðtÞdt m=s1:75 ð2Þ
0

If the vibration exposure consists of two or more vibrations of different intensity,

the total exposure EVDV (or VDV) is.
hX i14
EVDV ¼ EVDVi4 ð3Þ

hX i14
VDV ¼ VDVi4 ð4Þ

For the same vibration with a crest factor less than 12, EVDV is basically the same
as VDV. If the difference between the two values is less than 25%, the RMS method
can be used [7].

2.2 Other Commonly Used Vehicle Suspension Evaluation Methods

Many scholars have concluded a lot of suspension evaluation methods through
experiments and experiences. At present, the most common method to objectively
evaluate the suspension performance is based on the RMS value of the sprung mass
acceleration, the root mean square value of the suspension dynamic deflection, and the
RMS value of the dynamic tire load [8].
The RMS value of sprung mass acceleration indicates the ride stability and ride
comfort of the car. The RMS of suspension dynamic deflection reflects the deformation
494 X. Dong et al.

of the suspension. The RMS value of the dynamic load of the tire directly affects the
grounding of the vehicle. The excessive dynamic load of the tire will seriously affect
the safety of the grounding of the vehicle, which will degrade the steering stability.
Based on the above evaluation indicators, it can be seen that the ride stability and
comfort of the vehicle are affected by the sprung mass acceleration and suspension
dynamic deflection. Stability of vehicle operation is affected by tire dynamic load.
According to different vehicle suspension requirements, different vehicle suspension
parameters are selected to obtain the required suspension performance.

3 Theoretic Analysis for Bench Test System

The working principle of the vehicle suspension system. The unevenness of the road
surface is transmitted to four dampers of suspension systems through the wheels, and
then transmitted to the vehicle body through the suspension system. The vehicle system
in motion is a multi-degree-of-freedom vibration system. Research on the car sus-
pension is based on the dynamic model of the suspension. The vehicle model can
accurately predict the impact of the suspension parameters on its performance, thus it
can reflect the performance of real suspension system. However, the establishment of
an accurate model is a complex process. The establishment of a complete vehicle
model requires more relevant parameters, and the accuracy of the parameters will
inevitably affect the accuracy of the model.
The quarter-vehicle suspension model is the most basic model for the research of
the suspension system. Although it cannot be used for the attitude control of the
vehicle, it is sufficient to reflect the relevant indicators that reflect the performance of
the suspension: the sprung mass acceleration and the sprung mass displacement, sus-
pension dynamic deflection and tire dynamic load. In addition, the quarter vehicle
suspension model involves fewer parameters, which facilitates in-depth study on the
suspension system. Therefore, the test object of the vehicle suspension characteristic
test system in this paper is the quarter vehicle suspension model [9, 10].

3.1 Establishment of Physical Model of a Quarter Vehicle Suspension

The simplified model of an automotive passive suspension model and magneto-
rheology (MR) suspension model are shown in Figs. 1 and 2.

ms ms xs
xs
ks
cs fd

mu z
mu z
kt kt
d d

Fig. 1. 1⁄4 car passive suspension model Fig. 2. 1⁄4 car semi-active suspension model
 A Novel Bench of Quarter Vehicle Semi-active Suspension 495

In the figure: ms , xs are body masses and displacements, respectively. mu , z are tire
mass and displacement, respectively. ks is the stiffness of the suspension. kt is the tire
stiffness. cs is the viscous damping coefficient of the suspension. d is random road
input. fd is the damping force of a MR damper.

3.2 Establishment of Two-Degree-of-Freedom Mathematical Model

for the Quarter-Vehicle Suspension
The quarter vehicle suspension mathematic model with two-degree-of-freedom is
usually represented by the Newtonian mechanics method. The dynamic equation of the
passive suspension is.

ms€xs þ cs ð_xs  z_ Þ þ ks ðxs  zÞ ¼ 0 ð5Þ

ms€z  cs ð_xs  z_ Þ  ks ðxs  zÞ þ kt ðz  dÞ ¼ 0 ð6Þ

Where €xs , x_ s are the sprung mass accelerations and speeds, respectively. €z, z_ are un-
sprung mass accelerations and speeds, respectively.
For a MR semi-active suspension, the corresponding passive damper can be
replaced with a magneto-rheological damper, and its kinetic equation is.

ms€xs þ ks ðxs  zÞ þ fd ¼ 0 ð7Þ

mu€z  ks ðxs  zÞ þ kt ðz  dÞ  fd ¼ 0 ð8Þ

Here, only the passive suspension is taken as an example for illustration.

Performing Laplace transformation on Eqs. 1 and 2. The transfer function of the
sprung mass acceleration with respect to the road surface input.

€s ðsÞ kt ðCs s þ ks Þs2

X
HA ¼ ¼ ð9Þ
DðsÞ Ds

in which,

Ds ¼ mu ms s4 þ ðms Cs þ mu Cs Þs3 þ ðms ks þ ms kt þ mu ks Þs2 þ kt Cs s þ ks kt

The transfer function of the sprung mass displacement with respect to the road
surface input:
Xs ðsÞ kt ðCs s þ ks Þ
HDI ¼ ¼ ð10Þ
DðsÞ Ds

The transfer function of the suspension deflection relative to the road surface input,

Xs ðsÞ  ZðsÞ kt ms s2
HDE ¼ ¼ ð11Þ
DðsÞ Ds
496 X. Dong et al.

The transfer function of the tire dynamic load with respect to the road surface input,

kt ðZðsÞ  DðsÞÞ kt ZðsÞ k2 ðms s2 þ Cs s þ ks Þ

HF ¼ ¼  kt ¼ t  kt ð12Þ
DðsÞ DðsÞ Ds

As described in Sect. 2, the sprung mass acceleration is often used as an important

index for evaluating the ride comfort in the evaluation of the automobile performance.
The suspension deflection seriously affects the ride stability and riding comfort. The
tire dynamic load affects the stability of the car’s operation. In the process of vehicle
suspension test, due to the small rolling of the tire, it is difficult to test the dynamic load
of the tire. Therefore, in the process of vehicle suspension test, the dynamic load of
vehicle suspension is indirectly obtained by testing the tire dynamic deflection and un-
sprung mass acceleration.

4 Quarter Vehicle Semi-active Suspension Bench System

Design

The quarter car semi-active suspension bench system is mainly designed to simulate the
vibration characteristics of the quarter vehicle suspension system. Its overall structure is
shown in the Fig. 3.
The quarter vehicle semi-active suspension bench system consists of two parts, the
hardware system part and the sensing and software system part. The hardware system
part completes all physical movements of the vehicle suspension system. The sensing
scheme and software system part complete signal acquisition, feedback and control of
hardware.
The quarter vehicle semi-active suspension bench system working principle is as
follows, the electric servo vibration excitation system sends excitation signals for
simulating the road surface excitation to the vibration table, and the vibration table
produces corresponding vertical vibration, the vibration is transmitted to the inner
frame by the tire assembly and the damper assembly. The inner frame is vibrated by the
roller to move relative to the guide rail of the outer frame, thus the vertical vibration of

Data processing and suspension control system

Suspension main frame and sensing scheme system

Inner frame of vehicle suspension Under line test

MR damper Damper test Bench

sprung mass Test of force value
Acceleration sensor Signal input/out port characteristics
Outer frame sensor signal
Spring force sensor modeling
pulley
xs xs
Electric servo excitation fd
system fd Control current
Acceleration sensor xs − z z
displacement sensor xs − z
z−d MR damper
MR Damper
Vibration table z z Current driver
digital model
Analog control signal
control center
Signal processor Acceleration sensor Link arm z−d
Laser displacement sensor
sensor signal
A/D or D/A
conversion Control algorithm

dSPACE
d Wheel assembly
Control algorithm
interactive platform
Vibration table control force
Driving source
Signal Input/output port

Inverse MR damper digital model

Fig. 3. The quarter vehicle semi-active suspension bench system block diagram
 A Novel Bench of Quarter Vehicle Semi-active Suspension 497

the inner frame is achieved. Through the sensors at the specified position of the inner
frame, the status information of different parts of the suspension is transmitted to
dSPACE. The signal is converted by dSPACE and transmitted to the control algorithm
interaction platform. The control algorithm interactive platform outputs the control
signal to the damper after the signal is calculated and processed. The control signal is
applied through the current driver to the coil of the semi-active MR damper, thus the
suspension damping is changed and the suspension control is achieved.

4.1 Hardware System Part

The hardware system part of the quarter car semi-active suspension bench system is
mainly composed of two parts, the vibration control cabinet and the vibration platform
part, the suspension main frame part. The vibration control cabinet and the vibration
table part are used to provide excitation input for the semi-active suspension platform
of the quarter vehicle. The suspension main frame part is used to simulate 1/4 semi-
active suspension and related motion.

4.1.1 The Vibration Control Cabinet and the Vibration Table Part
The vibration control cabinet includes servo amplification system and vibration table
system, etc. The servo amplifier system is designed to drive a vibration table, including
a power amplifier, a power amplifier cabinet, a system control box, etc. By modulating
the signal, conducting electrical control, measuring and controlling the output voltage
and output current of the power amplifier, the excitation function of the vibration table
can be completed.
The vibration table system is mainly composed of an excitation part, a moving coil
and its supporting and guiding, a vibration isolation mechanism and a bearing. In the
moving coil and its support and guide, an air spring is used for support.

4.1.2 Suspension Main Frame Part

The suspension main frame is mainly composed of an inner frame of a simulated
vehicle suspension and a fixed vertical guide rail outer frame for restraining the inner
frame of the vehicle suspension. The inner frame includes the wheel assembly, damper
assembly, connecting arms, the frame that simulates the sprung mass of the suspension,
the fork arms that connect the wheels to the frame and the pulley of the inner frame for
sliding freely in the vertical direction. The outer frame part consists of a fixed pillar and
a guide rail mounted on the pillar surface and matched with the inner frame roller. The
specific structure is shown in Fig. 3.
The main points of the technical design are as follows.
To ensure that the un-sprung mass is closer to the real vehicle, the system adopts a
real-tire assembly.
In order to pledge the normal operation of the damper assembly part, a real car
damper assembly is adopted in the passive suspension; when the MR semi-active
damper assembly is configured, the travel of the MR damper is designed to satisfy
suspension requirements, outer contours and passive dampers are matched, and spring
of damper assemblies are used to ensure suspension stiffness matching.
498 X. Dong et al.

The pulley guide is used between the internal and external frames. The pulley is
evenly arranged along the vertical direction of the guide rail, a certain pre-pressure is
exerted between the pulley and the guide rail to eliminate the shake due to the
movement of the inner frame.
With the aim of guaranteeing that the sprung mass is close to the real vehicle, the
counterweights are evenly and symmetrically arranged on the top platform of the inner
frame according to requirements. The counterweight and the top platform of the inner
frame are fixed by set screws to prevent vibration.

4.2 Sensing Scheme and Software System Part

4.2.1 Sensing Scheme System
According to the requirements of vehicle suspension performance evaluation in the
Sect. 3.2 and the requirements of ensuring the completeness of vehicle suspension
performance parameters, the sensing scheme of the platform is worked out.
The installation of the sensor and the testing parameters of the sensor are detailed in
Fig. 3.

4.2.2 Software System

The software part is composed of two parts. One is the excitation control center for
controlling the excitation platform, and the other is the data processing and suspension
control system.
The vibration platform control center consists of human-computer interaction
interface and signal receiver. The input of different excitation signals can be achieved,
and the final excitation value can be corrected by adjusting the excitation input value in
real time according to the feedback of the acceleration signal.
Data processing and suspension control system is mainly established in this paper
for the MR semi-active suspension, including dSPACE system, offline MR damper
modeling and interactive platform for control algorithm. As shown in Fig. 3.
The dSPACE system is used for analog-to-digital conversion of the sensor signal,
digital-to-analog conversion of the control signal, and also for establishment of inter-
face with control algorithm program software, interaction between dSPACE and
control software.
With the dSPACE system, RCP technology can be achieved, the comparison of
different control algorithm tests for semi-active suspension can be accomplished. By
replacing the control algorithm module with the required control algorithm, and the
control experiment of the relevant control algorithm can be carried out based on the
rapid control prototype technology.
The offline MR damper modeling is a preparatory work. For details, please refer to
[11] Sect. 3.1.3. Through the treatment of the damper mathematical model, the inverse
model of the damper is obtained, and the control algorithm is used to control the output
characteristics of the damper.
The control algorithm interaction platform is used to operation the control algo-
rithm program. Download the required control algorithm into the software and establish
the interaction between the control program and dSPACE through the interface pro-
gram between the software and dSPACE. Through the force sensor installed between
 A Novel Bench of Quarter Vehicle Semi-active Suspension 499

the damper piston rod and the frame bottom plate of the vehicle suspension, the output
force value of the damper can be monitored in real time, the real-time correction of the
control signal can be realized.

5 Prototype Test and Discussion

According to the function and structure of the semi-active suspension characteristic test
platform system of the vehicle as description in the third section, the prototype of the
test platform system is designed and constructed as shown in Fig. 4.

1 6
7

9
3
10
4

1-MR suspension system; 2-Vibration table; 3-dSPACE; 4-DC power supply;

5-Control algorithm interactive platform; 6, 9-Acceleration sensor; 7-Displacement sensor;
8-MR damper assembly; 10-Laser displacement sensor.

Fig. 4. Auto semi-active suspension characteristic test platform system prototype schematic

By the vehicle semi-active suspension characteristics test platform system, various

suspension vibration tests can be carried out. In this paper, the simulation of the
sinusoidal spectrum excitation input is taken as an example to test the system, the
suspension characteristics of the MR semi-active suspension at different currents are
experimentally investigated.
The input excitation which is a displacement excitation of the vibrating table is as
A ¼ 5
sinð2
p
8
tÞ. The results of the test are shown in Fig. 5. The test results
show that the test platform system can simultaneously test multiple parameters of
suspension system, such as suspension spring load acceleration, un-sprung load
acceleration, suspension dynamic deflection, tire displacement and damper force value.
From the test results, we can clearly know the characteristics of the MR semi-active
suspension with different currents. We can combine and calculate the above test results
to evaluate the suspension performance according to the evaluation index of the second
section. This test proves the integrity and effectiveness of the test platform system.
500 X. Dong et al.

(a) Sprung acceleration (b) Un-sprung acceleration

(c) Suspension deflection (d) Damper force

Fig. 5. Characteristics of semi-active suspension under different current

6 Conclusion

In this paper, a universal and efficient standard bench of quarter car semi-active sus-
pension is proposed, designed and configured, and the effectiveness of the bench is
verified through experiments. Four main conclusions can be drawn as followings.
(1) During the design process of the quarter car semi-active suspension bench, it is
ensured that the bench is close to the real vehicle model, and that the testing result
of the test platform is relatively high relative to the spoilage of the real vehicle
model.
(2) The bench can be equipped with a passive suspension training platform to match
the stiffness coefficient and the damping coefficient of passive suspension.
(3) The bench based on dSPACE system can realize RCP technology and multiple
real-time tests of control algorithms, thus avoiding the time and cost loss caused
by designing and manufacturing the control algorithm hardware.
(4) By arranging the force sensor on the MR damper, real-time tracking and cor-
rection of the control force value can be realized, and the real-time performance of
the control effect can be detected.

Acknowledgment. The authors disclosed receipt of the following financial support for the
research, authorship, and/or publication of this article: This work was financially supported by
the National Natural Science Foundation of People’s Republic of China (Project NO. 51675063),
this research is also supported by graduate research and innovation foundation of Chongqing,
China (Grant No. CYB17023). These supports are gratefully acknowledged.
 A Novel Bench of Quarter Vehicle Semi-active Suspension 501

References
1. Dong, X., Zhang, Z., Yu, J., Tuo, F.: Partial fault sliding mode fault-tolerant control for
magneto-rheological semi-active suspension actuators considering temperature effect. In:
13th National Rheology Conference Proceedings, pp. 36–38, Xi’an (2016)
2. Dong, X., Yu, M., Liao, C., Chen, W., Huang, S.: Research on adaptive fuzzy logic control
for automobile magneto-rheological semi-active suspension. J. China J. Highw. Transp. 19
(2), 111–115 (2006)
3. Liu, Y., et al.: Study on testing method and testing system of vehicle suspension
performance. J. Highw. Transp. Res. Dev. 23(9), 126–130 (2006)
4. Chen, G., et al.: Research and design of the suspension performance test system for
automobiles. J. Chang. Univ. Sci. Technol. (Nat. Sci. Ed.) 32(3), 388–391 (2009)
5. Geng, X., Xu, W., Li, B., Wang, Z., Chen, Q.: The design of McPherson suspension system
performance test platform. J. Mach. Des. Manuf. (11), 163–165 (2012)
6. Omar, M., et al.: A universal suspension test rig for electrohydraulic active and passive
automotive suspension system. J. Alex. Eng. J. 56(4), 1–11 (2012)
7. ISO 2631-1: Mechanical vibration and shock-Evaluation of human exposure to whole-body
vibration-part 1: General requirements. S, pp. 3–10 (2003)
8. Miao, Y.: Study on control system of automobile magneto-rheological semi-active
suspension, pp. 46–51. Chongqing University, Chongqing (2003)
9. Tuo, F.: Research on fault-tolerant control of MR semi-active suspension with partial fault in
actuator considering temperature effect, pp. 32–34. Chongqing University, Chongqing
(2015)
10. Zhao, M., Liu, X., Pan, G.: Study on magneto-rheological damper model and semi-active
suspension control. J. Vib. Eng. 17, 1051–1053 (2004)
11. Dong, X.: Human simulated intelligent control of automobile magneto-rheological semi-
active suspension, pp. 44–49. Chongqing University, Chongqing (2006)
 Research on Nondestructive Testing
for the Depth of Hardening Zone

Ping Chen, Tao Xiang(&), Song Guo, Jie Fang, and Xin Xu

School of Mechanical Engineering, Chongqing University, Chongqing 400030,

People’s Republic of China
[email protected]

Abstract. In view of the problem of depth of hardening zone measurement for

quenching work-piece, this paper analyzes the advantages of new nondestructive
testing methods compared with traditional destructive testing methods. The
nondestructive testing method, principle and achievable precision index of depth
of hardening zone based on eddy current, ultrasonic and magnetic effects are
reviewed. Combining the deficiencies in the research status of nondestructive
testing technique, the development trend of nondestructive testing technology
for depth of hardening zone is prospected.

Keywords: Quenching work-piece
Depth of hardening zone

Nondestructive testing
Development trend

1 Introduction

Quenching is an important heat treatment process for the performance enhancement of

metal parts, and then the mechanical properties and physical properties of the parts can
be greatly improved with proper tempering process. Therefore, it is widely used in the
heat treatment process, such as the rotary support track, the gear rack surface, the
machine tool guide, the important shaft parts and so on [1]. The combination of suitable
hardening depth and suitable inner ductility can make these parts perform the best
comprehensive property. Therefore, accurate measurement of hardening depth is of
great significance in industrial production.
The traditional measurement of quenching depth usually uses the metallographic
method (from the quenching surface to the martensitic zone’s half of the depth), the
microhardness method (the hardness of semi-martensitic zone as the standard) or the
macro fracture method [2]. But these methods all need to damage the parts to varying
degrees. This not only affects the property of parts but also greatly increases the cost of
inspection, which this do not meet the development needs of Industry 4.0, either. In
particular, as for the large or high-precision quenched parts, the above method can’t be
effectively detected, which urgently requires an efficient nondestructive testing tech-
nique to measure quenching depth of the work-piece. Therefore, the researchers tried to
detect the quenching depth through different detection principles and methods. These
testing methods include nondestructive testing methods of based on eddy current,
ultrasonic and magnetic effects [3]. Finally, this paper will discuss the principle,
research status and development trend of the quenching depth’s nondestructive testing
method.

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 502–511, 2018.
https://doi.org/10.1007/978-981-13-2384-3_47
 Research on Nondestructive Testing for the Depth of Hardening Zone 503

2 Eddy Current Testing Technology

The basic principle of eddy current nondestructive testing: the exciting coil will gen-
erate an alternating magnetic field around the coil after energizing the exciting coil. The
eddy current shown in Fig. 1(a) will be generated, when the work-piece is placed in the
alternating magnetic field. However, the eddy currents will in turn generate a magnetic
field in the opposite direction compared with the excitation magnetic field so that the
original magnetic field in the coil is reduced, thus causing the change of the coil
impedance. Since the distribution of the eddy current depends on excitation conditions.
Therefore, when some exciting parameters are kept constant, the change of exciting
coil impedance or the voltage in another dedicated detecting coil can be measured so
that the researcher calculate the parameter that the test can need.
The eddy current in the measured work-piece is unevenly distributed because the
frequency of the exciting current, the conductivity and the magnetic permeability of the
tested work-piece can influence the eddy current generated in the work-piece. The eddy

Exciting coil Exciting coil

Eddy Current Depth Depth

Skin Effect Eddy Current

Workpiece

(a) (b)

Fig. 1. Principle of eddy current generation

current density is the largest and the densest on the work-piece surface, and the eddy
current density gradually decreases with the increase of depth. The attenuation
amplitude shows an exponential decay function with the increase of depth and also
shows the skin effect of eddy current, as shown in Fig. 1(b) [4].
The distance that the eddy current penetrates into the work-piece is called the depth
of penetration. Usually the depth that the eddy current density is attenuated to 1/e
(37%) of work-piece surface’s eddy current density is defined as the standard pene-
tration depth, also called the skin depth. Its calculation formula is as follows [4]:

1
d ¼ pffiffiffiffiffiffiffiffiffiffi : ð1Þ
pf lr

In the formula, d is the standard penetration depth (skin depth), r is the conduc-
tivity, l is the magnetic permeability, and f is the frequency of the exciting current.
According to the calculation formula, it can be known that the skin depth is
inversely proportional to the 1/2 order power of the current frequency for a given
quenched work-piece. Therefore, the reasonable detecting frequency can be selected by
504 P. Chen et al.

the calculation formula so that the skin depth includes the entire quenched layer and the
detection made can be more accurate and faster.
The measurement of hardening depth by using eddy current is essentially based on
establishing mathematical model between the hardening depth and the induced voltage
of detecting coil, so as to achieve the goal of measuring the hardening depth. For
example, the calculation formula of induction voltage for cylindrical quenching parts is
as follows [5]:

pd 2
Vm ¼ ð2pf ÞNs l l H0  108 : ð2Þ
4 r eff

Where Vm is the induced voltage of detecting coil, Ns is the number of coil turn, f is
the frequency of the exciting current, lr is the relative magnetic permeability, leff is the
effective magnetic permeability, and H0 is the magnetic field strength.
It can be known from the above formula that the induced voltage increases with the
increase of excitation frequency, but the excitation frequency can also affects the
effective permeability. So the induced voltage will show a non-linear increasing trend
when the excitation frequency increases. Therefore, the mathematical model between
the hardening depth and the detecting coil’s induced voltage is established under the
condition that the frequency of excitation current remains unchanged by this formula,
and the depth of hardening zone is finally measured by solving the mathematical
model. For example, in the literature [6], the hardening depth of 40Cr steel was
measured by using this method. Through the analysis of the measured eddy current, a
mathematical model between the hardening depth and the induced voltage was
established. The maximum relative deviation of the measured results of the hardening
depth was less than 6% compared with the metallographic method on average.

3 Ultrasonic Testing Technology

Ultrasonic nondestructive testing is a new nondestructive testing method with many

advantages, which is widely used in detection of defects, cracks and thicknesses. The
attenuation measurement and backscatter measurement are the two main technical
means of ultrasonic nondestructive testing [7–12], and ultrasonic backscatter mea-
surement is just used for nondestructive testing of hardening depth based on ultrasonic
[1]. Beecham and Fay were the earliest scholars who used ultrasonic scattering tech-
nique to evaluate grain size of metal material [13–15], then Du and Lonsdale used
ultrasonic backscattering technique to measure the quenching depth of railroad wheels
on the basis of Beecham’s and Fay’s studies and good results have been achieved [16].
Specifically, the grain size of quenching work-piece is from the small grain size of the
hardening layer in transition to the large grain size of substrate material. So the
microstructure size of the work-piece surface’s quenched layer is small, the energy loss
of ultrasonic wave is less and the ultrasonic reflection is very little when the ultrasonic
wave enters the work-piece surface through different methods. But it will encounter
with relatively bulky pearlite and ferrite when the ultrasonic wave penetrates into the
boundary between the hardened layer and the substrate material layer, so there will
 Research on Nondestructive Testing for the Depth of Hardening Zone 505

form backscattered echo. Therefore, the depth of hardening zone can be calculated by
measuring the time when the ultrasonic wave is emitted to the receiving echo.
At present, the most widely used method in the field of ultrasonic nondestructive
testing is the piezoelectric ultrasonic nondestructive testing technology. Figure 2 shows
the detection principle. After the ultrasonic wave enters into work-piece by the cou-
plant, the backscattering phenomenon will occur when ultrasonic wave passes through
the interface. Therefore, the time that the ultrasonic wave transmits from the work-
piece surface to the interface can be measured by the detection of backscatter signals,
and then the depth of hardening zone is calculated by the path formula.

Fig. 2. Principle of the piezoelectric ultrasonic non-destructive testing

According to the theory of acoustic scattering, the size of microstructure and the
frequency of ultrasonic wave both have a great influence on the detection. Therefore, it
is necessary to determine the frequency range that can generate backscattered echo
according to the Eq. (3) before testing [1]. It can not only improve the efficiency of
experimental detection, but also make the detection more accurate.
v
dk¼ : ð3Þ
f

Where d is the microstructure sizes, k is the wavelength, v is the wave velocity, f is

the ultrasonic frequency.
For example, in the literature [17], the hardening depth of specimen was detected
by the piezoelectric ultrasonic nondestructive testing method, and the final measure-
ment error was about 1 mm.

4 Magnetic Effect Testing Technology

4.1 Magnetic Coercivity Testing Technology
The magnetic hysteresis loop is a curve describing the magnetization phenomenon and
reverse magnetization phenomenon of the ferromagnetic material under the external
magnetic field. That is to say, it describes that the magnetic induction intensity B
changes with the external magnetic field intensity H, which reflects the basic magnetic
properties of the material and is one of the fundamental bases of the material for the
use. The magnetic hysteresis loop of magnetic material is shown in Fig. 3, showing
506 P. Chen et al.

three important parameters: Br (remanent magnetic induction intensity), Bm (saturated

magnetic intensity), Hc (magnetic coercive force). On the magnetic hysteresis loop, Br ,
Bm and Hc are all the characteristic parameters which can reflect the properties of the
material [20–23]. What’s more, the magnetic coercive force is related to the grain size
and the carbon content in the material. Therefore, the hardening depth of the quenched
work-piece can be characterized by the magnetic coercive [18, 19].

B
Bm
1
Br
2

-Hm -Hc Hc H
3 O 3 Hm

2
1 2 The reversible
-Br Barkhausen jump
1 3 The irreversible
-Bm Barkhausen jump

Fig. 3. The magnetic hysteresis loop

In order to obtain the value of the magnetic coercive force to characterize the depth
of hardening zone, the magnetic hysteresis loop is measured by the RC indirect
measurement method shown in Fig. 4. According to the following formula [18]:

CR2
B¼ UB : ð4Þ
SN2

The magnetic induction intensity in the tested piece can be calculated by measuring
the output voltage UB at the two ends of the integration capacitor C. Where B is the
magnetic induction intensity, C is the integration capacitor, R2 is the Integral resistance,
S is the effective magnetic flux area of detecting coil and N2 is the turn number of
detecting coil.

Exciting coil Detecting coil(N2)

I1 Integration capacitor C UB
Power source
I2
Sampling resistance R1 Workpiece IIntegration resistance R2

Fig. 4. The principle of nondestructive testing based on magnetic coercive force

 Research on Nondestructive Testing for the Depth of Hardening Zone 507

After measuring the magnetic induction intensity B under different the external
magnetic field intensity, the magnetic hysteresis loop can be obtained and the value of
magnetic coercive force can be similarly obtained. Since the magnetic coercive force is
closely related to the carbon content and the grain size (the grain size increases
gradually from the work-piece surface to the interface) of material, the depth of
hardening zone can be obtained by establishing the mathematical model between the
magnetic coercive force and the hardening depth and then solving the model. For
example, the carburizing and quenching hardening layer depth is measured by mag-
netic coercive force in the literature [18], and the relationship between the magnetic
coercive force and the quenching depth was established.

4.2 Barkhausen Effect Testing Technology

The magnetic domain is many tiny regions formed by the spontaneous magnetization
of ferromagnetic materials. In these regions, the orientation of the atomic magnetic
moment are the same, but the magnetic moments of the magnetic domain between
different magnetic domains are inconsistent. Thus, there is a transition region where the
atomic magnetic moments change from one direction to another direction among the
different magnetic domains is called the domain wall. In the magnetization process of
ferromagnetic materials, the Barkhausen jump is induced by the discontinuous
movement of the domain wall [24]. It is found that the magnetization is a stepwise
irreversible jump process in the steepest region of the magnetic hysteresis loop when
the ferromagnetic material is magnetized, as shown in Fig. 3. This discontinuous
jumping phenomenon is called the Barkhausen effect, and the discontinuous jumping
signal can be received in the detecting coil on the material surface, called Barkhausen
Noise [28–30]. In the movement process of the domain wall, the magnetization of an
external magnetic field is the motive force of the domain wall moving, besides the
change of microstructure inside the material and the residual stress are the resistance of
the domain wall moving. Since the Barkhausen Noise is induced by the discontinuous
movement of the domain wall, the Barkhausen Noise signal carries information on the
microstructure and residual stress inside the material. Therefore, the Barkhausen Noise
signal can be used to detect the depth of hardened layer [25–27].

Power amplifier Signal generator

Personal
computer
Signal processing circuit Data acquisition card
Magnetization ciol

The U-Shaped magnetic yoke

The detecting coil

Workpiece

Fig. 5. Principle of non-destructive testing based on Barkhausen effect

508 P. Chen et al.

The method shown in Fig. 5 is used to measure the Barkhausen Noise. After
establishing the mathematical model between the strength of noise signal and the depth
of hardening zone, and solving the mathematical model can finally achieve the goal of
measuring the depth of hardening zone. For example, the Barkhausen effect was used
to detect the depth of hardening zone in the literature [25]. The fitting straight line
equation of the strength of noise signal and the depth of hardening zone was
MP ¼ 14:688H 0:6828 , and the relative error of the fitting straight line equation was
measured to be within 4% at last.

5 Research Status and Deficiency

The detecting methods of hardening depth that are referred to above are all non-
destructive testing method. These methods can directly detect the work-piece and the
operation is simple and quick. What’s more, it can achieve the goal of large-scale
detection for the quenched work-piece and eliminate the disadvantage that the tradi-
tional method needs to damage the specimen and needs to do sampling detection.
So these methods have a wide application prospect.
Although nondestructive testing of the hardening depth has many advantages, there
are still many limitations. For example, the eddy current nondestructive testing tech-
nology can only detect work-piece with the shallow depth of hardening zone because of
the skin effect. If the tested work-piece has the large depth of hardening zone, the
measurement result will be inaccurate due to the skin effect. What’s more, the eddy
current nondestructive testing technology can only be used for conductive materials.
With regard to ultrasonic nondestructive testing technology, the excitation frequency of
ultrasonic wave has great influence on the detection. If the inappropriate excitation
frequency is selected, the scattered echo signal may be extremely weak and difficult to
recognize. So the reasonable excitation frequency must be selected in the experiment.
In addition, the couplant must be also used between the piezoelectric ultrasonic probe
and the work-piece to measure. Then, the type of couplant, its thickness, and other
factors must also be taken into account. This makes the detection more complicated.
Since the nondestructive testing technology based on magnetic coercive force is limited
to the detection principle, it can only detect the work-piece made of ferromagnetic
material and be limited by the shape of the detected work-piece. According to the
literature [25], the nondestructive testing technology based on Barkhausen effect also
limits the detecting depth because of the detection principle. When the hardening depth
is more than 4 mm, the signal strength of the Barkhausen Noise no longer changes.
This makes the Barkhausen effect be limited when detecting the large depth of hard-
ening zone. In addition, these nondestructive testing methods all require that the
detection probe is separately designed for the work-piece of complex shape, which also
brings about many difficulties in detection.
 Research on Nondestructive Testing for the Depth of Hardening Zone 509

6 The Development Tendency and Prospects

In view of the above deficiency, the development of nondestructive test methods for the
depth of hardening zone can be considered according to the following points. For the
quenched work-piece with shallow depth of hardening zone, the eddy current nonde-
structive testing and the Barkhausen Noise nondestructive testing can be used, and the
optimization research is carried out for two kinds of nondestructive testing technology
on this basis. For the optimization research, it can be considered from the following
two aspects:
(1) For traditional eddy current nondestructive testing technology, the detection of
depth is limited due to the skin effect. So pulsed eddy current testing technology
can be considered to take place of traditional eddy current testing technology. The
pulse eddy current nondestructive testing technology is a new detection method
developed on the basis of traditional eddy current nondestructive testing tech-
nology. It uses pulsed current signal as excitation signal. Compared with single-
frequency excitation signal, the frequency spectrum of pulsed excitation signal
has a wider range. The secondary induced electric field is easier to identify and
detect. The final induced voltage signal carries more abundant and diverse
information, which can increase its detection depth.
(2) For the nondestructive testing of hardening depth based on Barkhausen effect, a
mathematical model between the hardening depth and grain size can be consid-
ered to establish. Because the Barkhausen Noise signal carries grain size infor-
mation, this can make the depth of hardening zone that is measured further
increase by the grain size information carried in the noise signal.
For the quenched work-piece with deep hardened layer, ultrasonic nondestructive
testing technology and nondestructive testing technology based on magnetic coercivity
can be considered to use, and the optimization research was carried out for two kinds of
nondestructive testing technology on this basis. The optimization research can be
considered from the following two aspects:
(1) For the piezoelectric ultrasonic testing technology, it is difficult to control the
thickness of couplant. Therefore, the electromagnetic ultrasonic transducer can be
used to replace the piezoelectric ultrasonic transducer. Electromagnetic ultrasonic
transducer is a new type ultrasonic receiving and transmitting device. It relies on
the electromagnetic effect to transmit and receive ultrasonic. What’s more, it
doesn’t need the couplant in the detection process and contact with the work-piece.
So it can avoid the problems and disadvantages because of using the couplant.
(2) For the nondestructive testing technology based on magnetic coercive force, the
detection process is limited to the detection probe. When the shape of work-piece
is complicated, different structure size’s probes need to be designed for different
work-pieces each time the detection is performed. Therefore, it can consider the
miniaturized and flexible design of the probe, which can make the probe adapt to
most of the different shapes of work-piece. If the problem of probe can be solved,
the nondestructive testing technology based on magnetic coercive force can be
applied more widely.
510 P. Chen et al.

The above four detection methods all measure the depth of hardening zone by
establishing a mathematical model of the special signal parameters and the depth of
hardening zone. But they all ignore the relationship between grain size and the depth of
hardening zone. Therefore, the next research can focus on the relationship between the
grain size and the depth of the hardened layer. That is, the mathematical model between
the hardening depth and grain size is established by the grain size information carried in
the detected signal. Finally, solving the mathematical model achieves the goal of
measuring the depth of hardening zone.
When the nondestructive testing technology is used to measure the depth of
hardening zone of quenched work-piece, it can greatly improve the production effi-
ciency of actual quenched work-piece and greatly reduce the cost of metallographic
detection because of its nondestructive, convenient, rapid, efficient and low-cost
characteristics. Therefore, there is extremely important economic value and industrial
value to carry out further research, exploration and optimization on the nondestructive
testing method for the depth of hardening zone.

References
1. Shi, J.C.: Nondestructive testing technology of hardened layer of 45 steel based on ultrasonic
wave. Anhui University of Technology (2016)
2. Guan, Y.Q.: The ultrasonic measurement of the depth of hardening zone. J. Constr. Mach. 6,
106–107 (2015)
3. Xu, G.Q., Zhong, X.L., Zhan, S.F.: Ultrasonic testing technology for quenching depth of
cummins crankshaft. 28(4), 196–197 (2016)
4. Cheng, X.M.: Research on carburized layer characteristics and eddy current testing system of
carburized materials. Wuhan University of Technology (2003)
5. Li, J.W.: Nondestructive Testing Manual. China Machine Press, Beijing (2012)
6. Wang, C.Y., Chen, T.Q., Song, L., et al.: Eddy current nondestructive testing of the
hardening depth of 40Cr steel. J. Nondestruct Test. 31(6), 8–13 (2007)
7. He, X.P., Tian, Y.P., Zhang, H.P.: Ultrasonic nondestructive evaluation on the grain size of
metal materials. J. Tech. Acoust. 6, 445–451 (2013)
8. Lobkis, O.I., Rokhlin, S.I.: Characterization of polycrystals with elongated duplex
microstructure by inversion of ultrasonic backscattering data. J. Appl. Phys. Lett. 96(16),
161905 (2010)
9. Weaver, R.L.: Diffusivity of ultrasound in polycrystals. J. Mech. Phys. Solids 38, 55–86
(1990)
10. Papadakis, E.P.: Ultrasonic attenuation caused by scattering in polycrystalline media. In:
Mason, W. (ed.) Physical Acoustics, vol. IV, no. Part B, pp. 269–328. Academic Press, New
York (1968)
11. Merkulov, L.G.: Investigation of ultrasonic scattering in metals. Sov. J. Tech. Phys. 26, 59–
69 (1956)
12. Hirsekorn, S.: The scattering of ultrasonic waves by polycrystals. J. Acoust. Soc. Am. 72,
1021–1031 (1982)
13. Beecham, D.: Ultrasonic scatter in metals its properties and its application to grain size
determination. J. Ultrason. 4(2), 67–76 (1966)
14. Fay, B.: Theoretical considerations of ultrasound backscatter. J. Acta Acust. United Acust.
28(19), 354–357 (1973)
 Research on Nondestructive Testing for the Depth of Hardening Zone 511

15. Fay, B., Brendel, K., Ludwig, G.: Studies of inhomogeneous substances by ultrasonic back-
scattering. J. Ultrasound Med. Biol. 2(3), 195–198 (1976)
16. Du, H., Lonsdale, C., Oliver, J., et al.: Evaluation of railroad wheel steel with lamellar
duplex microstructures using diffuse ultrasonic backscatter. J. Nondestruct. Eval. 32(4), 331–
340 (2013)
17. Zhao, Z.K., Kong, C.H., Bao, W.B.: The principle and application of ultrasonic backscatter
hardening layer depth detector. In: 8th National Electromechanical Technology Annual
Meeting and Machinery Industry Energy Saving and Emission Reduction Process
Technology Seminar, pp. 95–101, Hangzhou (2014)
18. Li, D.P.: Research on detection technology of carburized and hardened depth of the magnetic
coercivity. Nanchang Hangkong University (2015)
19. Luo, X., Wu, W., Li, D.P., et al.: Detection of 20CrMnTi steel carburizing hardening depth
by magnetic coercive force. J. Nondestruct. Test. 38(7), 47–50 (2016)
20. Sun, J.M.: Study on measurement technology of stress measurement system on coercive
force measuring for ferromagnetic materials. Shenyang University of Technology (2016)
21. Lv, Z.: Ferromagnetic material stress testing technology research based on the coercivity.
Shenyang University of Technology (2016)
22. Zhou, Y.: Based on coercivity of ferromagnetic materials stress detecting technology.
Shenyang University of Technology (2015)
23. Zhang, L.: Coercive force of ferromagnetic materials testing technology research. Shenyang
University of Technology (2014)
24. Wang, W.T.: Research on the key technology of nondestructive testing based on Barkhausen
effect. Zhengzhou University (2017)
25. Liu, Y.B., Shi, H.J., Meng, F.Q., et al.: The research for Barkhausen noise to test the
hardened depth. J. Coal Mine Mach. 10, 24–25 (2001)
26. Liu, Y.B., Yang, Z.Y., Shi, X.J., et al.: Nondestructive determination of hardening depth in
an induction hardened bearings by Barkhausen noise. In: 1th International Mechanical
Engineering Conference, p. 547. Machinery Industry Press, Shanghai (2000)
27. Du, F.M., Xu, X.C., Xu, Y.H., et. al.: Nondestructive testing of hardening depth with
Barkhausen effect. In: 8th National Acoustic Emissions Academic Symposium, pp. 62–65,
Shanghai (1999)
28. Yuan, H.J., Fu, J., Li, H.M., et al.: Application status of nondestructive testing technology
based on Barkhausen effect. J. Acta Metrol. Sin. 37(01), 75–78 (2016)
29. Pan, L., Zhang, S.J., Wan, G., et al.: The detection of ferromagnetic material hardness based
on Barkhausen effect. J. Nondestruct. Test. 38(12), 28–31 (2016)
30. Zhang, Y.: Study on nondestructive testing of high strength steel plate based on Barkhausen
effect. Huazhong University of Science and Technology (2015)
 Distributed Fusion Estimation Based
on Robust Kalman Filtering for 3D
Spatial Positioning System

Li Liu1(B) , Wenju Zhou1(B) , Aolei Yang2 , Jun Yue1 , and Xiaofeng Zhang1
1
School of Information Science and Electrical Engineering, Ludong University,
Yantai 264025, China
[email protected], [email protected]
2
Shanghai Key Laboratory of Power Station Automation Technology,
School of Mechatronic Engineering and Automation,
Shanghai University, Shanghai 200072, China

Abstract. Because of the requirements of growing measurement scale,

the issue of the networked high-precision positioning has been developed
rapidly, and this paper designs a 3D spatial positioning system. With the
aid of the spatial positioning principle, the 3D spatial positioning system
is used to enlarge communication constraint and increase signal coordi-
nation processing. A information fusion estimation method is presented
for the distributed networked systems with data transmission delays. The
proposed distributed fusion estimation scheme employs the transforma-
tion of measurement and the weighted fusion of innovation sequence. To
reduce the communication burden and computational cost with transmis-
sion delays, a re-optimal weighted fusion estimator is designed. Moreover,
the proposed method reduces the information redundancy and maintains
the higher measurement accuracy. An illustrative example obtained from
the 3D spatial positioning system is given to validate the effectiveness of
the proposed method.

1 Introduction
In recent decades, to solve the problem of high-precision measurement and posi-
tioning, the space measurement and positioning systems with relevant measure-
ment methods have been developed rapidly. With the increasing of the network
scale, the scope of the measurement and positioning systems has been growing
much larger; meanwhile, the coverage mode has been growing much more flexible.
Therefore, the application of distributed network structure can able to increase
L. Liu—This work was supported by Natural Science Foundation of China
(61772253, 61633016, 61473182, 61472172), Key Research and Development Project
of Yantai (2017ZH061), Key Project of Science and Technology Commission of
Shanghai Municipality (15411953502), Key Technical of Key Industries of Shan-
dong (2016CYJS03A02-1), and Key Research and Development Project of Shandong
(2016ZDJS06A05).
c Springer Nature Singapore Pte Ltd. 2018
K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 512–521, 2018.
https://doi.org/10.1007/978-981-13-2384-3_48
 Distributed Fusion Estimation Based on Robust Kalman Filtering 513

the reliability of information, and improve the precision of collaborative calcu-

lation. Moreover, the distributed network structure reduces the burden of com-
munication [1]. Large-scale networked systems have been successfully applied in
some important research fields, such as network physical systems (cyber-physical
systems, CPSs), Smart grid, and intelligent transportation.
The distributed estimation strategy [2–4] is applied to target tracking and
positioning, fault diagnosis and so on. The objective is that state estimation
accuracy after fusion is higher than that of each local estimation. Research on
filtering is the basis of the information perception and fusion, according to differ-
ent performance indicators, the relevant distributed filtering has aroused more
and more attention, including distributed particle filtering [5], distributed H∞
filtering [6], distributed Kalman Filtering [2,3,7,8] and so on. Furthermore, dis-
tributed Kalman filtering is used to realize the coordination and fusion of error
estimation, which is an important scheme to solve the state estimation for large-
scale systems [2].
Due to the restriction of communication facilities and scale for networked sys-
tems, the measurement information is inevitably affected by the unpredictable
interference. Note that the fundamental Kalman filtering method can not sat-
isfy the performance requirement, such that the robust Kalman filtering method
is designed considering various noise sequences. For example, the uncertainty
of system parameters can be described by state-dependent noise [9,10]. Costa
et. al. [11] investigated robust pattern independent filtering for discrete-time
Markov jump systems with state-correlated noises. Considering the correlation
for noise, the distributed Kalman filtering fusion method [7,12] was presented
with cross correlation between measurement noise and process noise. In addi-
tion, for a class of noisy sequences with uncertain variance, the method of the
least upper bound in [13–15] for all admissible uncertainties was proposed. In
order to reduce the influence of transmission delays for the measured value, the
measurement transformation [15] was proposed, which used the Kalman filter
of recombination the innovation sequence to deal with multiple delays system.
Therefore, it is transformed into the correlative system with no time delay, and
the computational complexity is reduced.
According to the above analysis, based on the principle of spatial positioning
of linear CCD, the information measured by 3D photoelectric sensor is consid-
ered to be influenced by the uncertain factors, such as transmission delays and
cross correlation noise. In order to improve the robustness of the estimator, the
innovation sequence of weighted fusion is used to optimize the interference of
cross correlation noise during the information exchange. Developing the reorga-
nized measurement sequence and the innovation sequence, the Kalman filtering
is transformed into Kalman prediction to reduce the computational complex-
ity. The proposed weighted fusion method of filtering error covariance is used to
coordinate and exchange the measurement information between two subsystems.
Therefore, the distributed fusion estimation is based on the optimal information
fusion criterion. It is worth noting that the re-optimization of the local estimate
is able to get higher estimation precision than the local one.
514 L. Liu et al.

The remainder of the paper is organized as follows. The space positioning

system model is established in Sect. 2. Section 3 presents the distributed fusion
method using weighted innovation sequence method. Simulation results and anal-
ysis are presented in Sect. 4 and the concluding remarks are given in Sect. 5.

2 Space Positioning System Model

Based on the principle of spatial positioning, the photoelectric sensor is designed
in this section. Furthermore, considering the uncertainty of the distributed sys-
tems, the system model is established from the system state and measure infor-
mation, which is obtained by the sensor with noise correlation and transmission
delays.

2.1 Principle of Spatial Positioning

Firstly, the Metris Krypton Measurement approach based on the three-
coordinate positioning measurement is introduced, which is obtained by using
the optical tracker and the special light emission device (LED).
The 3D photoelectric sensor, as shown in Fig. 1(a), is composed of three linear
CCD, which is embedded three fixed-focus cylindrical optical mirrors. The three
linear CCD has a “Y” shape distribution on the same plane.

(a) 3D photoelectric sensor. (b) Positioning principle.

Fig. 1. 3D photoelectric sensor and its principle.

The system principle of 3D photoelectric sensor positioning is shown in

Fig. 1(b). The measured target is composed of the special LED. And then, each
cylindrical optical mirror through the light only project to its corresponding
plane under the CCD. A point light source of space is projected into each CCD
plane by a cylindrical optical mirror, which is a straight line light intersecting
with the CCD, and can perceive its position information on the CCD. If the
target is in the measurement range of the 3D photoelectric sensor, the target
will form three planes with three CCD lenses. Therefore, processor computes the
intersection point of three planes, that is, the coordinates of the light source for
the space points (X, Y, Z).
 Distributed Fusion Estimation Based on Robust Kalman Filtering 515

2.2 System Model Establishment

The information measured by 3D photoelectric sensor is disturbed by the uncer-
tain factors in the process of network communication. Based on the experimental
analysis, the distributed uncertainty system for sensor networks can be described
as following:
xk+1 = (Ak + ΔAk ) xk + Bk wk , k = 1, 2, · · · (1)
zki = Hki xk−τk + ki , i = 1, · · · , n, (2)
r
in which, xk ∈ R is the estimated state, zki
∈ R mi
represents the measured
value of the i-th sensor, and τk > 0 denotes the transmission delay. Ak ∈ Rr×r
expresses the state transition matrix, Bk ∈ Rr is a time-varying vector with
known dimensions, and Hki ∈ Rmi ×r is the measurement matrix. Meanwhile,
the state-dependent matrix ΔAk = Ck ξk is used for expressing the uncertainty
within the system [7], where Ck ∈ Rr×r is the state transition matrix, and the
state-dependent noise ξk ∈ R is independent of zero mean white noise. wk ∈
R represents the process noise with variance Qk , and ki ∈ Rmi represents
the measurement noise of the i-th sensor with variance Rki . Furthermore, the
initial state x0 is not correlated with other noise signals with the mean μ0 , and
covariance are set as P0 .
Considering the effect of transmission delay, the noise from the previous sam-
pling time is more important than the current measurement value. Therefore, it
is assumed that the process noise and measurement noise for the subsystems and
different sensors are one-step related. Accordingly, the relation has the following
statistical properties:
 i
k ) = 0, E k = 0,
E (w
wk  i T
E i wlT l
 k  (3)
Qk δk−l + Qk,l δk−l+1 Ski δk−l + Sk,li
δk−l+1
=
0 Rki δk−l + Rk,l
i
δk−l+1
 T
where δk−l is the Kronecker function, satisfying Qk = Qk T and Rki = Rki .

3 Distributed Fusion Estimation

Based on the Kalman filtering, a distributed weighted fusion estimation method
is proposed for weighted fusion of the innovation sequence and error covariance
matrix.

3.1 Filter Parameters

At the time instant of k, the i-th subsystem employs the minimized one-step
prediction error covariance, to probe the appropriate filter parameters. For the
local one-step prediction x̂ik+1|k , the covariance Pk+1|k
i
from the prediction error
is computed by the following equation:
516 L. Liu et al.
  T 
i
Pk+1|k =E x̃ik+1|k x̃ik+1|k
 i T  T
= Φik Pk|k−1
i
Φk + Bk Qk BkT + Kki Rki Kki + Ck Xk CkT
 T  i T  i T (4)
+ Φik Wki BkT + Bk Wki i
Φk − Bk Sk,k+1 Kk
 i T  i T  i T  T
i
−Kk Vk Φk − Φk Vk Kk − Kk Sk,k+1 BkT
i i i i

i
in which, Pk|k−1 denotes the one-step prediction error covariance at the time
instant of k − 1.
Since the local estimate x̂ik+1|k satisfies as following:

x̂ik+1|k = Φik x̂ik|k−1 + Kki yki (5)

the expectation Xk is solved according to the state xk . It is defined as:

Xk+1 = Ak Xk ATk + Ck Xk CkT + Bk Qk BkT

(6)
+Ak Bk−1 Qk−1,k BkT + Bk Qk,k−1 Bk−1
T
ATk

Δ
and the initial value is set to X0 = μ0 μT0 + P0 [7].
Then, Wki and Vki are obtained by the following means, respectively.
   i T
Wki = E x̃ik|k−1 wkT = Bk−1 Qk−1,k − Kk−1
i
Sk (7)

  T 
Vki = E x̃ik|k−1 vki i
= Bk−1 Sk−1,k+1 i
− Kk−1 i
Rk−1,k (8)

Thus, the optimal filter gain matrix Kki is evolved, i.e.

  i T −1
Kki = Ξki Hki Pk|k−1
i
Hk + Rki + Ξ̃ki (9)

 i T  T  i T
where Ξki = Ak Pk|k−1i
Hk + Bk Wki Hk + Ak Vki + Bk Sk,k+1
i
and Ξ̃ki =
  T   T
Hki Vki + Vki Hki .
i
Substituting Eqs. (9) into (4), error covariance matrix Pk+1|k is rewritten as

i i
 T
Pk+1|k = Ak Pk|k−1 ATk + Γki − Kki Ξki (10)
 T
Note that Γki = Bk Qk BkT + Ck Xk CkT + Ak Wki BkT + Bk Wki ATk . Based on
the above analysis, the matrix Σk can be calculated as following:
 T  T
Σki,j = Hki Pk|k−1
i,j
Hkj + Hki Vki,j + Hkj Vkj,i + Rki,j (11)

here Σki,j is an nm × nm-dimensional matrix.

 Distributed Fusion Estimation Based on Robust Kalman Filtering 517

3.2 Estimation Error Cross-Covariance

The weighted fusion strategy based on the Kalman one-step prediction x̂ik+1|k
and filter gain Kki , the predicted error covariance between the i-th and j-th
subsystems, at the time instant k can be obtained as following:

x̂ik+1|k = Ak x̂ik|k−1 + Kki Ỹk (12)

  i T −1
Kki = Ξki Hki Pk|k−1
i
Hk + Rki + Ξ̃ki (13)

 T
i,j i,j
Pk+1|k = Ak Pk|k−1 ATk + Bk Qk BkT + Ck Xk CkT + Kki Pk Kkj
 T  T
+Ak Wki BkT + Bk Wkj ATk − Ak Mki Kkj (14)
 T  T
− Bk LTk Kkj − Kki Mkj ATk − Kki Lk BkT

where,  
Pk = E Ỹk ỸkT
 i T
Wki = Bk−1 Qk−1,k − Kk−1 i
Sk
Xk+1 = Ak Xk ATk + Ck Xk CkT + Bk Qk BkT
+ Ak Bk−1 Qk−1,k BkT + Bk Qk,k−1 Bk−1 T
ATk (15)
n  
i,q T
Mki = Pk|k−1 (Hkq ) + Vki,q θkq
q=1   T 

n
T
Lk = (θkq ) Hkq Wkq + Sk−1,k
q
q=1

Next, the weighted fusion strategy is used for fusing the cross-correlated
covariance, which is defined as
 T −1  
J = Ixk − x̂k|k−1 Σk|k−1 Ixk − x̂k|k−1 (16)

Since the optimal Kalman filter uses the weighted fusion method of linear
minimum variance [12], the corresponding optimal information fusion estimates
i
covariance Pk|k−1 satisfies Pk|k−1 ≤ Pk|k−1 , i = 1, 2, · · · , n.

4 Experimental Results and Analysis

In order to verify the robustness of the proposed weighted fusion estimation,
a 3D photoelectric sensor positioning measurement platform is designed. The
experimental results of this paper is supported by the relation of the algorithm
and platform, the research of the distributed fusion estimation algorithm is the
518 L. Liu et al.

(a) State estimation from RKFDFE, OSF(b) State estimation of Sensor 1 from
and WRKF. RKFDFE.

(c) State estimation of Sensor 2 from(d) State estimation of Sensor 3 from

RKFDFE. RKFDFE.

Fig. 2. Comparison of state estimation for acceleration.

(a) Estimation error covariances from (b) Estimation error covariances of

OSF, WRKF and RKFDFE. RKFDFE from each sensor and fusion
center.

Fig. 3. Comparison of filtering error covariance.

 Distributed Fusion Estimation Based on Robust Kalman Filtering 519

support of the 3D photoelectric sensor positioning system. The mathematical

description of the target tracking system used in the experiment is:
⎛⎛  ⎞ ⎛ ⎞ ⎞ ⎛ 2 ⎞
0.96 T T 2 2 0.01 0 0 T 2
xk+1 = ⎝⎝ 0 0.96 T ⎠ + ⎝ 0 0.01 0 ⎠ ξk ⎠ xk + ⎝ T ⎠ wk
0 0 0.96 0 0 0.01 1
k = 1, 2, · · ·
(17)
zki = Hki xk−τk + ki , i = 1, 2, 3, 4 (18)
wk = ηk + ηk−1 (19)
ki = βi wk−1 + βi ηk , i = 1, 2, 3, 4 (20)
T
The sample period is T = 0.1s. The state xk = (sk ṡk s̈k ) represents the
position, velocity, and acceleration of the target at the instant of kT . ξk ∈ R
represents a state-dependent noise with a zero mean value of and variance σξ2 = 1.
ηk ∈ R is independent on the variance for white noise ση2 = 0.09. zki represents
the time delay value measured by the i-th sensor. In which, H1 = [1 0.6 0.8],
H2 = [0.8 1 0.5], H3 = [0.8 0.3 1] and H4 = [0.6 0.5 1]. Variable βi is
used to determine the intensity of the noise correlation, and set β1 = 1, β2 = 0.8,
β3 = 2 and β4 = 5.
T
The initial value is x̂z0|0
0
= μ0 = E (x0 ) = (1 1 1) and P0|0 z0
= 0.01I3 .
Compare the proposed distributed fusion estimation (RKFDFE) with the
weighted robust Kalman filtering (WRKF) [7] and the optimal sequential fusion
(OSF) method [12]. In the mathematical analysis of signal and system, the noise
sequence adopts the white noise processing method.
In order to demonstrate the estimation results clearly, the state of the acceler-
ation and its estimation are shown in Fig. 2. In contrast to these three methods,
Fig. 2(a) shows the best performance of RKFDFE in state estimation with cross-
correlation noise and transmission delays. Figures 2(b)–(d) show the state esti-
mate of the acceleration computed by each sensor. Because the cross-correlation
for noise, the filtering estimation for target tracking is more dependent on the
RKFDFE method, when the cross-correlated strength is small.
Figure 3 compares and analyzes the corresponding estimation error covari-
ance results. Figure 3(a) illustrates the trace of estimated error covariance by the
above three methods, and verifies that the trace of the estimated error covariance
matrix obtained by RKFDFE is less than the other two methods. Figure 3(b)
using the RKFDFE method, the trace (i.e. T r (P i) , i = 1, · · · , 4) of the error
covariance matrix between each filter and the trace (i.e. T r (P d)) of the fused
matrix is compared. Furthermore, the weighted fusion method is able to improve
the robustness for the state estimation and get the optimal steady value, and
then obtain a higher measurement accuracy.

5 Conclusion
Based on the principle of spatial positioning of linear CCD, the 3D photoelectric
sensor has been designed. The spatial target information has been measured,
520 L. Liu et al.

and the trajectory has been tracked and estimated. This paper has studied the
state estimation problem of distributed uncertain systems with cross correla-
tion noise and transmission delays. To deal with the stochastic transmissions
and reduce the network communication burden, the reorganized weighted fusion
estimation and innovation sequence method has been proposed. It has realized
the information interaction between any two neighboring subsystems due to the
communication constraints. Finally, numerical experiments have shown that the
proposed distributed fusion estimation algorithm has a better dynamic tracking
effect for state estimation, to improve the spatial positioning accuracy.

References
1. Ge, X., Han, Q.-L.: Distributed event-triggered H∞ filtering over sensor networks
with communication delays. Inf. Sci. 291, 128–142 (2015)
2. Li, D., Kar, S., Moura, J.M., Poor, H.V., Cui, S.: Distributed Kalman filtering over
massive data sets: analysis through large deviations of random Riccati equations.
IEEE Trans. Inf. Theory 61(3), 1351–1372 (2015)
3. Song, E., Xu, J., Zhu, Y.: Optimal distributed Kalman filtering fusion with sin-
gular covariances of filtering errors and measurement noises. IEEE Trans. Autom.
Control 59(5), 1271–1282 (2014)
4. Hu, J., Wang, Z., Chen, D., Alsaadi, F.E.: Estimation, filtering and fusion for
networked systems with network-induced phenomena: New progress and prospects.
Inf. Fus. 31, 65–75 (2016)
5. Zhong, X., Mohammadi, A., Premkumar, A., Asif, A.: A distributed particle filter-
ing approach for multiple acoustic source tracking using an acoustic vector sensor
network. Sig. Process. 108, 589–603 (2015)
6. Song, B., Wu, Z.-G., Park, J.H., Shi, G., Zhang, Y.: H∞ filtering for stochastic
systems driven by Poisson processes. Int. J. Control 88(1), 2–10 (2015)
7. Feng, J., Wang, Z., Zeng, M.: Distributed weighted robust Kalman filter fusion for
uncertain systems with autocorrelated and cross-correlated noises. Inf. Fus. 14(1),
78–86 (2013)
8. Song, H., Yu, L., Zhang, W.-A.: Distributed consensus-based Kalman filtering in
sensor networks with quantised communications and random sensor failures. IET
Sig. Process. 8(2), 107–118 (2014)
9. Zhang, W., Zhao, Y., Sheng, L.: Some remarks on stability of stochastic singular
systems with state-dependent noise. Automatica 51, 273–277 (2015)
10. Tian, T., Sun, S., Li, N.: Multi-sensor information fusion estimators for stochastic
uncertain systems with correlated noises. Inf. Fus. 27, 126–137 (2016)
11. Costa, O., Benites, G.: Robust mode-independent filtering for discrete-time Markov
jump linear systems with multiplicative noises. Int. J. Control 86(5), 779–793
(2013)
12. Yan, L., Li, X.R., Xia, Y., Fu, M.: Optimal sequential and distributed fusion for
state estimation in cross-correlated noise. Automatica 49(12), 3607–3612 (2013)
13. Qi, W., Zhang, P., Deng, Z.: Robust weighted fusion Kalman filters for multisensor
time-varying systems with uncertain noise variances. Sig. Process. 99, 185–200
(2014)
14. Zhang, P., Qi, W., Deng, Z.: Hierarchical fusion robust Kalman filter for clustering
sensor network time-varying systems with uncertain noise variances. Int. J. Adapt.
Control Sig. Process. 29(1), 99–122 (2015)
 Distributed Fusion Estimation Based on Robust Kalman Filtering 521

15. Qi, W.-J., Zhang, P., Deng, Z.-L.: Robust sequential covariance intersection fusion
Kalman filtering over multi-agent sensor networks with measurement delays and
uncertain noise variances. Acta Automatica Sinica 40(11), 2632–2642 (2014)
16. Rezaei, H., Esfanjani, R.M., Sedaaghi, M.H.: Improved robust finite-horizon
Kalman filtering for uncertain networked time-varying systems. Inf. Sci. 293, 263–
274 (2015)
Image Processing
 A Novel 3D Head Multi-feature Constraint
Method for Human Localization Based
on Multiple Depth Cameras

Feixiang Zhou1, Haikuan Wang1(&), Zhile Yang2, and Dong Xie1

1
School of Mechatronics Engineering and Automation, Shanghai University,
Shanghai 200444, China
[email protected]
2
Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences,
Shenzhen 518055, Guangdong, China

Abstract. Up to date, the majority of existing spatial localization methods is

based on visual positioning methods and non-visual positioning methods. In the
vision-based positioning method, the traditional 2D human detection method are
vulnerable to tackle with environmental changes including illumination, com-
plex background, object occlusion, shadow interference and other factors, due to
which the algorithm is less robust and difficult to achieve accurate target posi-
tioning. In respects to 3D human positioning, binocular vision or multi-vision
approaches have been widely used to acquire depth information. The complexity
of the algorithm is high, and the detection range is limited. To deal with this, a
spatial location method based on multiple depth cameras is proposed in this
paper. Multiple 3D-TOF depth cameras are jointly used to directly obtain depth
information. Histograms of Oriented Depth (HOD) features are then extracted
and trained to find the human head and shoulder region. Moreover, Spatial
Density of Head (SDH) and the Convexity and Square Similarity of Head
(CSSH) features are combined to determine the human target. Finally, the
positioning data of multiple cameras are determined by using Nearest Center
Point (NCP) to obtain the final human body positioning information. Experi-
mental results show that the proposed method can not only obtain higher
recognition rate and positioning accuracy, but also enable a larger detection
range, meeting the needs of large-scale spatial positioning.

Keywords: Multiple depth cameras
Human localization
SDH

CSSH
NCP

1 Introduction

With the unprecedented development of information technology and artificial intelli-

gence, human detection and positioning has been widely used in video surveillance,
virtual reality, human-computer interaction, planetary detection, behaviour under-
standing, and other areas, achieving public safety monitoring and management, acci-
dents Precaution, detection and treatment, nursing care for young and old patients, and
autonomous navigation.

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 525–535, 2018.
https://doi.org/10.1007/978-981-13-2384-3_49
526 F. Zhou et al.

At present, various methods [1–3] have been proposed in addressing this research
area, the majority of which are based on 2D images. Ma et al. [4] first obtained the
motion vector by the optical flow method, and then applied it into a shape model for
human detection and tracking. However, such methods are easily affected by the
external environment. The depth image has many advantages over the 2D gray image
and the color image. First, the depth image is robust to the changes in color and
illumination. In addition, the depth image directly represents the target’s 3D infor-
mation. Therefore, researchers began to use depth information to study the detection
and positioning of moving targets. The RDSF algorithm (Relational Depth Similarity
Features) proposed by Ikemura et al. [5] makes full use of the difference in the fre-
quency distribution of the depth values of different objects, and divides the depth map
into blocks that do not overlap each other. Yu et al. [6] proposed SLTP (Simplified
Local Ternary Patterns) to perform histogram statistics on the edge information of
depth images. Wang et al. [7] proposed a depth feature extraction algorithm PDSS
(Pyramid depth self-similarities) due to strong local self-similarity of depth information
of the human body. Spinello et al. [8] proposed HOD depth feature extraction algo-
rithm that mimics the HOG algorithm, and performs histogram statistics on the mag-
nitude and direction of depth difference to extracts depth features. Shen et al. [9]
improved the HOG features and proposed an HDD feature extraction algorithm. The
algorithm extends the gradient direction from the original 180-degree statistical space
to 360-degree, and makes statistics on the gradients in different directions more precise.
When shooting vertically, the head area is closest to the camera, and the human body
target can be detected and positioned by recognizing the head area. Xia et al. [10]
proposed a model-based human detection system. The algorithm uses 2D head contour
model and 3D head surface geometry model for human detection. The literature [11]
relies on the relationship between the pixels at the edge of the head region to create a
head template to realize the recognition of the head. Meanwhile, the author also pro-
posed a human template matching algorithm [12] to further improve the accuracy of the
head recognition. The algorithm has a strong recognition effect on multi-person
environment and strong ability to eliminate human-like targets.
Aiming at the practical needs of large space, this paper proposes a 3D head multi-
feature constraint method based on multiple depth cameras to detect and locate the
human body. According to the multi-camera spatial structure, NCP is used to determine
the positioning data of multiple cameras to obtain the final body positioning
information.
The rest of the paper is organized as follows: Sect. 2 presents the human posi-
tioning method based on multiple depth cameras; Sect. 3 proposes the nearest center
point method. The Sect. 4 analyses the experimental results, followed by that Sect. 5
concludes the paper.

2 Multi-feature Constraint Method for Human Localization

Figure 1 shows the flow chart of human positioning method based on multiple depth
cameras. Using multiple cameras avoids the disadvantages of a single camera with a
small coverage and improves the detection accuracy.
 A Novel 3D Head Multi-feature Constraint Method 527

Multiple depth images

Segmentation of deep image

HOD feature extraction and training

Conforming to CSSH feature

constraints

Conforming to SDH feature

constraints

Nearest center point method

Outputing location information Discarding

Fig. 1. Flow chart of human location method

2.1 Segmentation of Depth Image

In image processing, we normally have interest only in some particular areas of an
image. It is assumed that the actual scene in this paper is fixed, and depth background
difference is used to segment the human target. In this paper, the five depth images
acquired by the five cameras in their respective positions are reserved as fixed back-
ground images. The spatial structure and background image are shown in Fig. 2.
Due to that the camera shoots downwards, the legs of the human body can be
considered

as invalid areas. This paper controls the quantified depth range at
Dnear ; Dfar . Firstly, Gaussian filtering is performed on the original depth data to filter
out noise and suppress the drift of depth data. Then the original depth image is sub-
tracted from the background image, and the foreground target is extracted according to
the threshold T, as shown in Formula (1).
Di ðx; yÞ ¼ jIi ðx; yÞ  Bðx; yÞj

1; Di ðx; yÞ  T ð1Þ
Ti ðx; yÞ ¼
0; Di ðx; yÞ\T

where Bðx; yÞ is the background image, Ii ðx; yÞ is the i-th frame image, Ti ðx; yÞ is the
binary image, and it is the effective target point.
Then the depth image of the corresponding area is extracted, as shown in
Formula (2).
528 F. Zhou et al.

(a) (b) (c)

(d) (e) (f)

Fig. 2. Spatial structure and background images of 5 cameras. (a) spatial structure; (b–f)
background images.

Ii ðx; yÞ; Di ðx; yÞ  T
Si ðx; yÞ ¼ ð2Þ
0; Di ðx; yÞ\T


where Si ðx; yÞ is the effective human region, and Si ðx; yÞ Dnear ; Dfar .

2.2 Extraction and Training of HOD Feature

The traditional HOG feature is generally used to describe the local area of the image. It
composes the target feature by calculating the gradient direction histogram on the local
area. This article utilises the HOD feature to detect depth images, which is very similar
to HOG. The only difference is that HOD calculates the depth difference in the depth
image, and HOG calculates the gray value or RGB value of the pixel in the 2D image.
After obtaining the HOD feature vector, SVM classifier is used in this paper to train
the features. In this paper, SVM is used to classify and recognize the HOD features of
human head and shoulders in real space. The positive and negative samples are trained
to obtain the decision function, and different feature vectors are distinguished.

2.3 3D Head Feature Constraints

According to Sect. 2.2, head and shoulder detection of human targets can be accom-
plished, but over-fitting or under-fitting is often caused by feature training, resulting in
missing or false detection. Therefore, two human character constraints (SDH and
CSSH) are proposed in this paper to further improve detection accuracy.
In the depth image, the entire human head area can be considered as a standard
sphere, and the upper half of the human head is like a circular arc convexity, as shown
 A Novel 3D Head Multi-feature Constraint Method 529

in Fig. 3(a). The depth of the edge is greater than the depth of the middle. The convex
surface can be considered to be composed of an infinite number of circular arc curves,
and for a single circular arc curve, the slope of positive and negative changes only once
during the process from one side to the other. According to this feature, a CSHA-based
head feature discriminate method is proposed. The n points on the edge of the con-
vexity d1 ; d2 ; d3 . . .dn are selected as shown in Fig. 3(b). Then the n points are not
repeatedly linked to obtain n=2 circular arc curves, as shown in the colored curves in
Fig. 3(b), and the number of points that conform to the law of slope change is counted.
Finally the head convexity similarity Scon is calculated.

N
Scon ¼ ð3Þ
n=2

(a) (b)

Fig. 3. Circular arc convexity of head area: (a) circular arc convexity; (b) circular arc curves.

In the 2D image, an approximate ellipse target head region will also be obtained, as
shown in the red circle in Fig. 6. To simplify the calculation, the rectangular region of
M  N is taken as the head Harea, as shown in the blue rectangle in Fig. 4. The pixels
in the rectangular area Hi ðx; yÞ can be represented as

Fig. 4. Top area of head. (Color figure online)

530 F. Zhou et al.


1; Di ðx; yÞ 2 Harea
Hi ðx; yÞ ¼ ð4Þ
0; Di ðx; yÞ 62 Harea

The square similarity of head is

P M1
N1 P
16  Hi ðx; yÞ
x¼0 y¼0
Ssqu ¼ ð5Þ
ðChead Þ2

where Chead is the perimeter of a rectangular region. When the head is more like a
standard circle, the rectangle resembles a square, and the value of Ssqu is closer to 1.
Finally, CSSH is defined by combining these two features, as shown in Formula (6).

Scs ¼ k1  Scon þ k2  Ssqu ð6Þ

where k1; k2 are the weight coefficients, and Tk 2 ½0; 1 is the threshold. When Scs  Tk ,
the current area satisfies the CSSH feature. Otherwise, it does not satisfy the feature.
The spatial density of head (SDH) proposed in this paper is based on a fixed ratio
relationship between the head region of the human body and the space area of the
shoulder and head, as shown in Fig. 5. This feature does not change with height and
weight. When a person walks in the scene, there will be no huge deformation on the
shoulders and above. The spatial density of the head Dhead can be calculated with
Formula (7).

Fig. 5. Spatial distribution of head area.

VHead
Dhead ¼ ð7Þ
VRec

where VRec is the volume of space where the human head is located; VHead is the
volume of the head, which can be approximated as the sphere volume, and it can be
calculated with Formula (8).
 A Novel 3D Head Multi-feature Constraint Method 531

0v ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi13
u P
u Ti ðx; yÞ
4 B tðx;yÞ2hz C
VHead ¼ pB
@
C
A ð8Þ
3 p

When VHead  Vk and Vk 2 ½0; 1, the current region satisfies the SDH feature, and
otherwise, it do not satisfy the feature.
The effects of over-fitting and under-fitting caused by training can be prevented by
combining the above two features, and the detection accuracy of the human body can
be improved.

3 Nearest Center Point Method

After
 determining the head area of the human body, the central point of the head region
lx ; ly is selected as the positioning point of the human body, as shown in formula (9).

P M1
N1 P
x
Hi ðx; yÞ
x¼0 y¼0
lx ¼
P M1
N1 P
Hi ðx; yÞ
x¼0 y¼0
ð9Þ
P M1
N1 P
y
Hi ðx; yÞ
x¼0 y¼0
ly ¼
P M1
N1 P
Hi ðx; yÞ
x¼0 y¼0

Since multiple cameras may detect the head region at the same time, it is necessary
to determine which camera data is valid. In this paper, the nearest center point method
is proposed to select an effective location point. The pixel distance between the center
point of the head and the center point of the image ðu0 ; v0 Þ is calculated, and the closest
point is selected as the final human body target location point, as shown in formula
(10). This can prevent the problem of incomplete depth data when the target deviates
from the center of the image, thus ensuring the accuracy of positioning information.
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2ffi
D¼ ðlx  u0 Þ2 þ ly  v0 ð10Þ

Finally, according to the camera coordinate transformation, the actual spatial

position of the center point of the head can be obtained, that is, the position coordinate
of the human body.
532 F. Zhou et al.

4 Experimental Results and Analysis

In order to verify the stability of the system and the effectiveness of proposed algo-
rithm, the positioning test of the human target was performed in the actual space. In this
experiment, a total of 10 individuals were tested in sequence. Firstly, 5000 frames of
depth images were collected for training, including head-up, head-down, head-turn,
squat, and other movement states. Some of the positive and negative samples used in
this paper are shown in Figs. 6 and 7 respectively, and all training samples are nor-
malized to the same size ð32  32Þ. The algorithm is written in C++ and compiled and
run in Visual Studio 2013. The test computer uses an Intel core i5-4570 processor
clocked at 3.20 GHz. The five TOF depth cameras used in this paper have a resolution
of 320  240, and they are fixed at a height of 3 m above the ground. The lens of the
middle camera is parallel to the ground, and the remaining four cameras are oriented in
the four directions towards the east, west, south and north respectively, and have a
certain angle with the horizontal plane.

Fig. 6. Positive samples. (Color figure online)

SVM classifier combining the SDH and CSSH features is used to determine the
human head region. The results are shown in Fig. 8, and the point on the head is the
positioning point of this target.
In addition, the detection effect of human detection algorithm combining HOD with
SVM is compared with the algorithm combining SDH and CSSH feature constrains, as
shown in Table 1.
It can be seen from Table 1 that the 3D head feature constraint method proposed in
this paper can effectively improve the detection precision of the human target, and the
total running time of the algorithm is controlled within 100 ms, e.g. 10 frames of depth
images can be processed per second.
In order to further verify the accuracy of location, the relationship between the
outputs of the 3D coordinates of the center of the head and the location in the actual
space is compared in this paper. Figure 9(a) shows the change in the height of a person
with 1.70 m height, and he is positioned in real time when walking in the space. The
results show that the height error is 3 cm, while Fig. 9(b) shows the 2D track of
human target is compared with the track of walking in real space. It can be found that
the positioning is more accurate near the central point (0, 0), while in the place close to
 A Novel 3D Head Multi-feature Constraint Method 533

Fig. 7. Negative samples.

(a ) (b )

(c ) (d )

Fig. 8. Human location: (a) normal walking; (b) head-up; (c) head-down; (d) squat.

Table 1. Comparison of accuracy and real time of two algorithms for human detection
Algorithm Training Test False False Detection Running
samples samples positive negative rate time
rate rate
HOD + SVM 3000 270 4.67% 4.83% 95.33% 82 ms
HOD + SVM + SDH + CSSH 3000 350 2.87% 3.96% 97.13% 94 ms
534 F. Zhou et al.

(a) (b)

Fig. 9. Comparison between experimental measurements and actual values: (a) height distri-
bution of human body; (b) track of human head.

the corner, there is an obvious deviation. The main reason is that when people approach
the wall, the information of the target is not complete.

5 Conclusion

In order to address the drawbacks of algorithms for large scene space location, a 3D
head multi-feature constraint method based on multiple depth cameras is proposed in
this paper to detect and locate the human target. Multiple depth cameras are jointly
applied to obtain depth images of the human body and a large number of depth images
of human head are divided to extract the HOD features of the area. Then the head and
shoulder classifier is trained by SVM, and SDH and CSSH features are used to further
improve the accuracy of human detection and positioning. Finally, the nearest center
point method is used to determine the final human body positioning information.
Experimental results show that the system has high stability and competitive real-time
performance, and the proposed algorithm also has good robustness and accuracy,
which is of great potential to future specific applications.

References
1. Dalal, N., Triggs, B.: Histograms of oriented gradients for human detection. In: IEEE
Computer Society Conference on Computer Vision & Pattern Recognition, pp. 886–893
(2005)
2. Dalal, N., Triggs, B., Schmid, C.: Human detection using oriented histograms of flow and
appearance. In: Leonardis, A., Bischof, H., Pinz, A. (eds.) ECCV 2006. LNCS, vol. 3952,
pp. 428–441. Springer, Heidelberg (2006). https://doi.org/10.1007/11744047_33
3. Schwartz, W.R., Kembhavi, A., Harwood, D., Davis, L.S.: Human detection using partial
least squares analysis. In: IEEE International Conference on Computer Vision, pp. 24–31
(2010)
 A Novel 3D Head Multi-feature Constraint Method 535

4. Ma, J., Ren, F.: Detect and track the dynamic deformation human body with the active shape
model modified by motion mectors. In: IEEE International Conference on Cloud Computing
and Intelligence Systems, pp. 587–591 (2011)
5. Ikemura, S., Fujiyoshi, H.: Real-time human detection using relational depth similarity
features. In: Kimmel, R., Klette, R., Sugimoto, A. (eds.) ACCV 2010. LNCS, vol. 6495,
pp. 25–38. Springer, Heidelberg (2011). https://doi.org/10.1007/978-3-642-19282-1_3
6. Yu, S., Wu, S., Wang, L.: SLTP: a fast descriptor for people detection in depth images. In:
IEEE Ninth International Conference on Advanced Video and Signal-Based Surveillance,
pp. 43–47 (2012)
7. Wang, N., Gong, X., Liu, J.: A new depth descriptor for pedestrian detection in RGB-D
images, pp. 3688–3691 (2012)
8. Spinello, L., Arras, K.O.: People detection in RGB-D data. In: International Conference on
Intelligent Robots & Systems, vol. 38, no. 2, pp. 3838–3843 (2011)
9. Shen, Y., Hao, Z., Wang, P., Ma, S., Liu, W.: A novel human detection approach based on
depth map via kinect. In: IEEE Conference on Computer Vision and Pattern Recognition
Workshops, pp. 535–541 (2013)
10. Xia, L., Chen, C. C., Aggarwal, J.K.: Human detection using depth information by kinect.
In: Computer Vision and Pattern Recognition Workshops, pp. 15–22 (2011)
11. van Oosterhout, T., Bakkes, S., Kröse, B.: Head detection in stereo data for people counting
and segmentation (2011)
12. van Oosterhout, T., Englebienne, G., Kröse, B.: RARE: people detection in crowded
passages by range image reconstruction. Mach. Vis. Appl. 26(5), 561–573 (2015)
 Head Ternary Pattern-Head Shoulder Density
Features Pedestrian Detection Algorithm
Based on Depth Image

Haikuan Wang, Haoxiang Sun(&), Kangli Liu, and Minrui Fei

School of Mechatronics Engineering and Automation, Shanghai University,

99 Shangda Road, BaoShan District, Shanghai 200444, China
[email protected]

Abstract. The existing pedestrian detection algorithms are mostly based on 2D

image processing, which are susceptible to complex background, the changes of
light intensity, and many other factors. For these reasons, the robustness and
accuracy of traditional algorithms are not ideal. Moreover, for other few algo-
rithms that use 3D vision, the core process is to extract features by converting
3D images to grayscale images, which cannot make full use of depth infor-
mation. To solve these problems, a pedestrian detection algorithm based on
depth image is proposed in this paper. For depth image preprocessing, a 3D
dilation and erosion algorithm based on probability density is proposed. Com-
bined with the Head Ternary Pattern-Head Shoulder Density (HTP-HSD) fea-
tures proposed in this paper, the implement of pedestrian detection can be
ensured using SVM classifier to classify the features. Experimental results show
that the HTP-HSD features meet the actual needs of pedestrian detection with
higher recognition rate and recognition speed.

Keywords: Depth image
Pedestrian detection
3D dilation and erosion

HTP-HSD

1 Introduction

With the enhancement of hardware and the richness of AI application scenarios,

pedestrian detection plays an increasingly important role in image processing. Besides,
pedestrian detection has made great contributions in many fields, such as traffic acci-
dent warning, intelligent service robots, etc. According to the cameras, pedestrian
detection algorithms can be divided into (a) pedestrian detection algorithm based on
color images and (b) pedestrian detection algorithm based on depth images.
In the pedestrian detection based on color images, the Histogram of Oriented Gra-
dient (HOG) proposed by Dalal and Triggs [1] is currently the most widely used
pedestrian feature descriptor. However, the HOG algorithm is based on gradient features,
which leads to high dimensions and slow calculations. In order to solve this problem,
Zhu et al. [1] allow the block size in the HOG to be variable to increase the speed of
operation. Malysiak and Markard [2] used parallel technology to implement HOG on the
GPU. The Local Binary Pattern (LBP) is a feature extraction method for texture clas-
sification, which is first proposed by Liu et al. [3]. Based on the characteristics of
© Springer Nature Singapore Pte Ltd. 2018
K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 536–545, 2018.
https://doi.org/10.1007/978-981-13-2384-3_50
 Head Ternary Pattern-Head Shoulder Density Features 537

pedestrians, Mu et al. [4, 5] proposed two variants of LBP: Semantic-LBP (S-LBP) and
Fourier LBP (F-LBP). Considering the basic features detected by the color images are
susceptible to light, Cosmo et al. [6] used the relative stability of human body structure to
propose a color self-similarity (CSS) feature, which can be combined with the HOG
feature to greatly improve the detection performance.
In recent years, with the development of hardware technology, depth cameras can
capture three-dimensional information of the shooting scene with a greater range and
higher accuracy. In terms of the fact, depth cameras can meet the requirements of
system stability on the industrial site, and therefore have more advantages in obstacle
positioning and detection. Compared with the color image, the depth information only
depends on the three-dimensional features of the object, which has nothing to do with
the light intensity and contrast. All the above advantages let depth cameras bypass the
problems faced by existing pedestrian detection algorithms. Xia et al. [7] proposed a
template-matching method based on depth image for human detection, using Canny
operator for edge extraction, and then calculating template matching of depth image to
locate the human body. Spinello and Arras [8] attempted to use depth information
instead of gray information for calculating traditional HOG feature, and proposed a
Histogram of Oriented Depth (HOD). Zhang et al. [9] proposed the Water-fill algorithm
which is based on the flow characteristics of water in reality. This algorithm obtains the
local extremum of the head by imitating the flow direction of the water droplets and the
stopping point of the water droplets and has a high recognition rate. The literature [10]
relies on the relationship between the pixel points at the edge of the head region to
create a head template and to achieve head recognition.
Based on the above research, this paper proposes an HTP-HSD features pedestrian
detection algorithm based on depth vision collected by ToF cameras. After the pre-
process of original images, the HTP-HSD combined features is used for pedestrian
detection and localization.
The structure of this paper is organized as follows: this section is an introduction;
specific scheme of pedestrian detection is introduced in Sect. 2; experiments for the
algorithm and analyzes the results are conducted in Sect. 3; Sect. 4 is a summary.

2 Scheme of Pedestrian Detection

The depth-based pedestrian detection process is shown in Fig. 1.

2.1 Depth Image Preprocessing

Depth image preprocessing, which provide support for subsequent detection algo-
rithms, is an important part of the entire machine vision inspection system, including
spatial filtering, morphological processing and preliminary extraction of some primary
features. However, when using the depth camera to capture the image of the application
scene, the complex background environment will bring difficulties to the classification
and recognition. First, the uncertainty of the ambient light source causes depth infor-
mation anomalies. Second, the materials, finish, and flatness of different surfaces are
different from each other, leading to the fluctuations or lack of depth information.
538 H. Wang et al.

3D Dilation and Erosion

Scene Image
Based on Probability
segmentation Preprocessing
Density

HTP HSD Pedestrian

Feature Feature Feature
Extraction

SVM Object
Recognition

Fig. 1. The flow chart of pedestrian detection algorithm.

Based on these problems, spatial domain and morphological filtering in image pre-
processing can weaken the fluctuation of depth information, fill in the missing depth
information, and thus improve the quality of the image.
3D Dilation and Erosion Based on Probability Density
Traditional morphological processing mostly uses fixed templates to process images in
binary images. Therefore, when dealing with depth images with complex scenes, the
traditional morphological filtering has more limitations [11]. In this paper, a 3D dilation
and erosion method based on probability density is proposed, which is used to fill
missing depth information in depth images and remove isolated points and noises.
First, the depth histogram of the original image is counted, and then the depth
interval between the obstacle and the pedestrian is located. Second, the original image
is projected onto the corresponding depth interval DR . The standard deviation of the
normal distribution is ra and expected Dði; jÞ defined as follows:
P
i;j ½Pði; jÞ
Dði; jÞ
Dðx; yÞ ¼ P ð1Þ
i;j Pði; jÞ

And the probability density function Hði; j; ra Þ of the normal distribution is

jDm  Dði; jÞj2

Hði; j; ra Þ ¼ expð Þ
k ð2Þ
2r2a

1
k ¼ pffiffiffiffiffiffi ð3Þ
2pra  1

Let the dilation factor a and the erosion factor b are defined as:
P
i;j Pði; jÞ
Hði; j; ra Þ
a¼ ð4Þ
m
n1
b¼1a ð5Þ
 Head Ternary Pattern-Head Shoulder Density Features 539

Thus, the symbolic function Pðx; yÞ be defined as


1; Dðx; yÞ 2 DR
Pðx; yÞ ¼ ð6Þ
0; Dðx; yÞ 62 DR

Let TP and TF be thresholds of dilation and erosion respectively. When the dilation
factor of the invalid point a [ TP , the depth of the point will be filled according to the
Eq. (1). And when the erosion factor of valid point b [ TF , this point will be removed.
An adaptive threshold is obtained by learning the probability distribution of all pixels
with the dilation factor and the erosion factor in the neighborhood of the center point.
The application of this method can effectively suppress noise of the depth image after
depth projection and can fill in the black hole to improve the effect of preprocessing.
Increasing the number of iterations of 3D dilation and erosion can better fill in the black
holes, remove noise, and enhance the effect of image preprocessing in the depth image.
The experiments show that, in Fig. 2, the human profile is more obvious traditional
method. It is more conducive to the final identification and classification.

Fig. 2. Comparison of effects of different dilation and erosion methods. (a) the original image.
(b) the image after 3D Dilation and Erosion Based on Probability Density processing. (c) the
image after traditional template-based processing.

Scene Segmentation
The objects themselves have a certain spatial thickness, therefore the use of depth
information for segmentation can effectively separate the complex foreground and
background in the depth image. Because the interference around the object within a
single scene after segmentation is relatively small, it is beneficial to improve the
accuracy of obstacle location. The positioning of obstacles uses the following
equations:

P
X P
Y
i
Dði; jÞ
i¼1 j¼1
Cx ¼ ð7Þ
P
X P
Y
Dði; jÞ
i¼1 j¼1
540 H. Wang et al.
P
X P
Y
j
Dði; jÞ
i¼1 j¼1
Cy ¼ ð8Þ
P
X P
Y
Dði; jÞ
i¼1 j¼1

where Dði; jÞ is the depth value of the point ði; jÞ, X and Y respectively denotes the
total number of pixels in the horizontal and vertical directions. The region of the
obstacle can be completely located by using the region growing method.
The result of image segmentation using the depth information is shown in Fig. 3.

Fig. 3. The result of image segmentation. (a) the original depth image. (b) three images after
segmentation.

2.2 HTP-HSD Pedestrian Features Extraction

As the preprocessing is described in Sect. 2.1, the scenario can be described more
accurately and completely in depth image which prepares for the next step of pedestrian
features extraction.
The traditional detection methods are mostly based on the pedestrian’s outline
features and shape features. [12] proposed a Head-Shoulder-Body Density (HSBD)
based pedestrian detection method. The application of these features shows good
robustness in dealing with images with complex environments. These features can
discriminate pedestrians and other obstacles effectively and accurately.
In the application scenario, the HSBD feature may be falsely detected or missed
because of the blockage of the body part. Analysis of the actual application sample
shows that there is a relatively fixed ratio between the width of the head and the
shoulder. In addition, the situation that the human head is shielded is very rare,
therefore Head-Shoulder-Density (HSD) features can improve the accuracy of pedes-
trian detection. So, a model based on depth image is introduced and a Head Ternary
Pattern-Head Shoulder Density (HTP-HSD) features in proposed. According to the
depth interval to which the object belongs, the HSD can be calculated as:
P
Pði; jÞ
ði;jÞ2H
HSD ¼ ð9Þ
H
where H is the selected region of head and Pði; jÞ is the symbolic function:
 Head Ternary Pattern-Head Shoulder Density Features 541


1; Dði; jÞ 2 DR
Pði; jÞ ¼ ð10Þ
0; Dði; jÞ 62 DR

DR denotes the depth interval where the object is located. The extraction result of
HSD feature is shown in Fig. 4.

The Area
of Head

The Area of
Pedestrian

Fig. 4. The result of HSD feature extraction.

In the positioned region of head, HTP feature are extracted from its outline.
Because the detected region and the number of pixels is small, the stability of the
gradient features can be improved by using the Sobel operator to calculate the gradient
of the depth around the center point. The extraction HTP feature is as shown in Fig. 5.

Fig. 5. HTP feature extraction diagram.

Let Dx ¼ dðx; y þ 1Þ  dðx; y  1Þ be the gradient of X axis and Dy ¼

dðx þ 1; yÞ  dðx  1; yÞ be the gradient of Y axis. HTP feature of center point is
computed as:

HTPðxc ; yc Þ ¼ hðtx ; ty Þ ð11Þ

where
8
< 1; Dx [ Td
tx ¼ 0; jDxj  Td ð12Þ
:
1; Dx  Td
542 H. Wang et al.

8
< 1; Dy [ Td
ty ¼ 0; jDyj  Td ð13Þ
:
1; Dy  Td

in which Dxc ¼ ð2Dx þ Dðx  1Þ þ Dðx þ 1ÞÞ=4 is denoted as the horizontal gradient
and Dyc ¼ ð2Dy þ Dðy  1Þ þ Dðy þ 1ÞÞ=4 is denoted as the vertical gradient of the
center pixel calculated by Sobel operator and Td denotes the threshold for adjusting the
HTP feature.
Using the HTP-HSD feature can detect pedestrians accurately and effectively
because this feature is less affected by the environment. It is suitable for implementing
pedestrian detection in application scenarios with complex backgrounds. The ToF
camera directly obtains the depth information of each pixel in the scene. The adoption
of scale information avoids the disadvantage of the long time-consuming disadvantage
of the traditional image detection that requires the construction of an image pyramid to
obtain scale-invariant features.

2.3 Support Vector Machine

In order to distinguish between pedestrians and other obstacles, features of objects
extracted from image should be classified and identified. Most classification algorithms
are based on the theory of statistics. By training the positive and negative samples in
advance, the criteria for classification are obtained. Then the new samples are identified
according to the criteria.
Support Vector Machine (SVM) was first proposed by Cortes and Vapnik in the
1990s [13]. It is a supervised machine learning method that trains a large amount of
sample data with tagged information. The support vector, which minimizes the error
and maximizes the separation interval, is usually used for classification and regression
analysis. This support vector is also known as the maximum interval classifier.
Assuming the number of data samples is n: x1  xn , the classified hyperplane can
be represented as:

wT x þ b ¼ 0 ð14Þ

where x denotes the point on the classification hyperplane, and w denotes the vector
perpendicular to the classification hyperplane, and b denotes the displacement which is
used to improve the flexibility of classified hyperplane. In order to make all the training
samples outside two parallel hyperplane spacing regions, it is necessary to ensure that
all samples x1  xn satisfies:

wT xi þ b  1; or
ð15Þ
wT xi þ b   1

According to the content of Sects. 2.1 and 2.2, the HTP-HSD features are extracted
from massive depth images collected by ToF cameras, which are preprocessed by 3D
 Head Ternary Pattern-Head Shoulder Density Features 543

dilation and erosion based on probability density. HTP-HSD features are trained by the
SVM classifier to achieve pedestrian detection and classification.

3 Experiments

In order to verify the effectiveness of the algorithm, 2400 depth images collected in the
actual application scene are divided into three parts in this experiment, in which 1440
training samples are for SVM training, and 480 cross verified samples are for validating
model validity, and the rest 480 test samples are for testing the generalization capability
of the model. The results of some different positions about obstacles and humans are
shown below (Fig. 6):

Fig. 6. The original image, segmented image, HTP-HSD features image and result of pedestrian
detection in different situation. (a) pedestrian in front of an obstacle. (b) pedestrian between two
obstacles. (c) pedestrian image is incomplete. (d) pedestrian standing sideway.

After training 1440 training samples, the parameters of the SVM were obtained.
Then cross verified samples are used to verify that whether the model parameters are
reasonable. The precision is 98.81% and there is no false alarm which indicates that the
trained SVM model is effective. Test samples are used to test the generalization ability
of the pedestrian detection method proposed in this paper. The false detection rate is
0.74% which shows that the HTP-HSD pedestrian detection algorithm has better
generality and universal applicability. The result and analysis of experiments are shown
in Tables. 1 and 2.
544 H. Wang et al.

Table 1. The result of experiments.

Positive Negative True False False True
samples samples positive positive negative negative
Cross 252 228 249 0 3 228
validation
samples
Test samples 208 272 208 2 0 270

Table 2. The analysis of experiments

Precision Recall F Score
Cross validation samples 100.00% 98.81% 99.40%
Test samples 99.04% 100.00% 99.52%

Precision, Recall, and F Score are usually used in machine learning to evaluate
classification models and algorithms. The definitions are as follows:

TruePositive
Precision ¼ ð16Þ
TruePositive þ FalsePositive
TruePositive
Recall ¼ ð17Þ
TruePositive þ FalseNegative

P
R
FScore ¼ 2
ð18Þ
PþR

Experimental results show that the HTP-HSD features can be used for pedestrian
detection based on depth images. By analyzing the false detection, images that are false
alarmed or missed were taken during outdoor operations. It is inspected that the HTP
feature cannot be matched due to that the intense sunlight caused abnormal fluctua-
tions. In normal situations, the HTP-HSD feature can accurately classify pedestrian.
This method meets the requirements of high detection rate and low false alarm rate in
application scenarios.
The test platform is a personal computer with an i5-4570 CPU and the development
environment is Windows. The software development platform is Microsoft Visual
Studio 2013. The average time of each image is 35 ms, and 28 frames of depth images
can be processed every second.

4 Conclusion

An HTP-HSD features pedestrian detection algorithm based on depth images is pro-

posed in this paper. It avoids the disadvantages of traditional 2D pedestrian positioning.
By filtering and preprocessing the original depth image using the 3D dilation and
erosion based on probability density, the HTP-HSD features proposed in this paper are
 Head Ternary Pattern-Head Shoulder Density Features 545

extracted. The pedestrian detection is implemented using criteria trained by SVM.

Experiments show that this method can detect and identify pedestrian more accurately,
and it meets the requirements of practical applications for real-time pedestrian detection.

References
1. Dalal, N., Triggs, B.: Histograms of oriented gradients for human detection. In: 2005 IEEE
Computer Society Conference on Computer Vision and Pattern Recognition, pp. 886–893.
IEEE, San Diego (2005)
2. Malysiak, D., Markard, M.: Increasing the efficiency of GPU-based HOG algorithms
through tile-images. In: Nguyen, N.T., Trawiński, B., Fujita, H., Hong, T.P. (eds.) ACIIDS
2016. LNCS (LNAI), vol. 9621, pp. 708–720. Springer, Heidelberg (2016). https://doi.org/
10.1007/978-3-662-49381-6_68
3. Liu, L., Fieguth, P., Wang, X., Pietikäinen, M., Hu, D.: Evaluation of LBP and deep texture
descriptors with a new robustness benchmark. In: Leibe, B., Matas, J., Sebe, N., Welling, M.
(eds.) ECCV 2016. LNCS, vol. 9907, pp. 69–86. Springer, Cham (2016). https://doi.org/10.
1007/978-3-319-46487-9_5
4. Xu, F., Zhang, J., Wang, J.Z.: Microexpression identification and categorization using a
facial dynamics map. IEEE Trans. Affect. Comput. 8, 254–267 (2017)
5. Mu, Y., Yan, S., Liu, Y., Huang, T., Zhou, B.: Discriminative local binary patterns for
human detection in personal album. In: 2008 IEEE Conference on Computer Vision and
Pattern Recognition, pp. 1–8. IEEE, Anchorage (2008)
6. Cosmo, D.L., Salles, E.O.T., Ciarelli, P.M.: Pedestrian detection system based on HOG and
a modified version of CSS. In: Seventh International Conference on Machine Vision (ICMV
2014), pp. 5–15. SPIE, Milan (2015)
7. Xia, L., Chen, C.C., Aggarwal, J.K.: Human detection using depth information by kinect. In:
2011 Conference on Computer Vision and Pattern Recognition, pp. 15–22. IEEE, Colorado
Springs (2011)
8. Spinello, L., Arras, K.O.: People detection in RGB-D data. In: 2011 International
Conference on Intelligent Robots and Systems, pp. 3838–3843. IEEE, San Francisco (2011)
9. Zhang, X., Yan, J., Feng, S., Lei, Z., Yi, D., Li, S.Z.: Water filling: unsupervised people
counting via vertical kinect sensor. In: 2012 IEEE Ninth International Conference on
Advanced Video and Signal-Based Surveillance, pp. 215–220. IEEE, Beijing (2012)
10. Hou, Y.-L., Pang, G.K.H.: Multi-cue-based crowd segmentation in stereo vision. In: Real,
P., Diaz-Pernil, D., Molina-Abril, H., Berciano, A., Kropatsch, W. (eds.) CAIP 2011. LNCS,
vol. 6854, pp. 93–101. Springer, Heidelberg (2011). https://doi.org/10.1007/978-3-642-
23672-3_12
11. Hasan, S.M.A., Ko, K.: Depth edge detection by image-based smoothing and morphological
operations. J. Comput. Des. Eng. 3, 191–197 (2016)
12. Qi, X., Fei, M., Hu, H., Wang, H.: A novel 3D expansion and corrosion method for human
detection based on depth information. In: Fei, M., Ma, S., Li, X., Sun, X., Jia, L., Su, Z.
(eds.) LSMS/ICSEE -2017. CCIS, vol. 761, pp. 556–565. Springer, Singapore (2017).
https://doi.org/10.1007/978-981-10-6370-1_55
13. Wang, X., Han, T.X., Yan, S.: An HOG-LBP human detector with partial occlusion
handling. In: 2009 IEEE 12th International Conference on Computer Vision, pp. 32–39.
IEEE, Kyoto (2009)
 Maximally Stable Extremal Regions Improved
Tracking Algorithm Based on Depth Image

Haikuan Wang, Dong Xie(&), Haoxiang Sun, and Wenju Zhou

School of Mechatronics Engineering and Automation,

Shanghai University, 99 Shangda Road, BaoShan District,
Shanghai 200444, China
[email protected]

Abstract. In order to solve the problem that traditional Camshift algorithm can
easily fail to track overlapping targets and multiple similar depth targets, a new
improved maximally stable extremal regions (MSER) algorithm is presented in
this paper. Firstly, the suspected target contour is extracted and similarity
analysis is performed. Secondly, the improved MSER algorithm is used to
confirm the target contour and update the similarity library. Finally, combined
with the physical properties unique to the depth image and based on the Kalman
filter, it is possible to predict the tracking target’s moving position. The
experimental results show that the real-time performance and recognition rate
are improved, and robustness to the situation of target overlap and occlusion is
better with the improved MSER algorithm.

Keywords: Depth image
MSER algorithm
Target tracking

Camshift algorithm

1 Introduction

Traditional digital image processing technology [1] has gradually matured, but it is
greatly affected by light, and meanwhile stability is difficult to control at times, which
greatly restricts its further development. The depth information of the target can be
reflected directly and identified with the structural information due to that the depth
image [2] is not affected by the light intensity nor the surface characteristics of the
object, and binocular imaging can be replaced by depth images, because of which more
and more applications are developed.
Nowadays, there are many classic methods, e.g. eigen pictures and eigen shapes
[3], using the depth information for target tracking, among which the continuous
adaptive mean shift (Camshift) [4] algorithm attracts much attention. The algorithm
effectively solves human target deformation and achieves better results in the simple
background environment. However, when there are multiple targets of similar depth in
the image, simple layering alone cannot be multi-targeted.
The MSER algorithm [5] marks a series of discernible detection areas as suspicious
target areas for the current frame image by processing the images, and determines the
exact boundary of the detected area. However, when the pixel change is not obvious,
the algorithm will not be able to extract valid information, resulting in tracking failure.

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 546–554, 2018.
https://doi.org/10.1007/978-981-13-2384-3_51
 Maximally Stable Extremal Regions Improved Tracking Algorithm 547

Kalman filter [6], which is low computational complex and better real time, is an
algorithm calculating the linear minimum variance estimate of the state sequence of the
dynamic system.
Based on the content above, a new MSER improved tracking algorithm based on
depth image is proposed in this paper. Firstly, the depth image is extracted and the
contour curve and local surface features [7] are extracted, and the similarity is analyzed.
The contour will be considered invalid and ignored when the similarity is low and
considered valid when the similarity is high. If the similarity stays in the middle, the
improved MSER algorithm will be used to determine the target contour, and to extract
the contour features of the confirmed target, and to update the similarity database, so as
to improve the accuracy of the similarity analysis. The Kalman filter at last is used to
predict the position of the target in the three-dimensional scene which keeps the target’s
motion consistent and coherent.

2 Features and Acquisition of Depth Image

The depth image is usually converted into a grayscale image [8] for intuitive display,
but the meaning of the depth image and the ordinary grayscale image is obviously
different. Every point of the depth image represents the distance between the point and
the depth camera, which is a solid physical meaning that will be used to determine the
position of the target in a three-dimensional scene.
There are many popular methods to acquire depth image such as stereo matching
[9] method of a binocular camera, ToF method [10, 11] based on the principle of time-
of-flight, depth measurement method based on the structure light [12] and the moiré
fringe characteristics. These methods have their own strengths in different fields. In this
paper, depth information is collected with time-of-flight method, which has high real-
time performance more advantages compared to the other measurement methods, and is
less influenced by the environment and less complex.

3 Specific Algorithm Flow

The overall framework of this algorithm is shown in Fig. 1.

3.1 Similarity Analysis

The outer contour of objects is acquired and then extracted into contour curves, on
which the largest curvature point is set as a feature and projected on the original
contour as marked point [13]. The contour, presented by the sequence of the marked
points, is analyzed by similarity with the contours in the similarity library. The pre-
cision of the object is realized by ruling out the contours of low similarity.
A discrete set p ¼ fp1 ; p2 ; p3 ; . . .pn g is collected from the contour. The relative
positions between pi and the rest points are reduced to the distribution counts of every
sector on the model. Denote k ¼ f1; 2; . . .Kg as a number set in which K ¼ M  N.
The statistical distribution histograms is denoted by hi ðkÞ which satisfies
548 H. Wang et al.

Target depth
image

Extract target
outline

L M
Mser
Similarity
improved
analysis
algorithm

H
Throw away

Update
similarity
library

Kalman filter

Forecast
target
location

Target
tracking

Fig. 1. Algorithm flow chart

hi ðkÞ ¼ #fq 6¼ pi : ðq  pi Þ 2 binðkÞg ð1Þ

Then calculate the cost matrix which satisfies

X
K
½hi ðkÞ  hj ðkÞ2
Ci;j ¼ Cðpi ; qj Þ ¼ 0:5  ð2Þ
k¼1
hi ðkÞ þ hj ðkÞ

where hj ðkÞ denotes the point qi shape histogram of the target Q what satisfies
 Maximally Stable Extremal Regions Improved Tracking Algorithm 549

X
HðpÞ ¼ Cðpi ; qp ðiÞÞ ð3Þ

The larger the value of H is, the lower the similarity will be, and vice versa. Target
Depth map and outline map are simply shown in Fig. 2.

Fig. 2. Three target depth maps and contour maps

3.2 Improved MSER Algorithm

The MSER detection algorithm is based on the orderly arrangement of the pixel values
in the image, and then sets the sorted pixel values into a component tree. The maximum
stable extremum area of the merged area is determined by the determination criterion to
clean up the invalid area.
Bucket sorting [14], also called Binsort sorting, can greatly reduce the complexity
of conventional sorting methods. When the pixels are bucket sorted, they will be
distributed at its position in the image where the extremum area can be found by
merging pixels with similar properties via union-find method. The two sets must be
examined if they are intersected before merging. As is shown in Fig. 3.

0 1

2 5 8 4 6

2 1
5 8

4 6

Fig. 3. Union-find sets

550 H. Wang et al.

When using the recursive method to find the root node, the complexity of each
algorithm is oðnÞ, which greatly increases the overall algorithm’s consumption time.
Therefore, a path compression method is used to reduce the complexity (see Fig. 4).
When all the algorithms are completed, only the final connected component remains.
The neighborhood quadtree method is used to extract the MSER in order to save the
originally connected regions at different times in the image and improve the efficiency
of the algorithm.

3 3
1 1 2 3 1 2 3
1 1 1 2 1 2 1 2 1 2
3 3

i=1 i=2 i=3

Fig. 4. Extremum area search

Before nodes are merged, function Make Set is used to complete the creation of the
initial node, and the presence of neighboring nodes within the four-neighborhood of the
pixel should be checked. The invalid region is deleted via the stability condition, and
the maximum stable extremum region that meets the conditions is obtained.
When the areas are merged, all the extremum regions in the image are obtained, but
not all of the regions are the maximum stable extremum region. For invalid areas, the
specific removal rules are as follows:
(1) Removal of oversized areas: the threshold number of pixels occupied by the
largest area is set to 3000 through experimental analysis of the tested images pixel
numbers in this paper.
(2) Removal of small areas: Similar to the method of large areas, the analysis of
experimental results shows that the threshold number of pixels occupied by the
minimum area is set to 400.
(3) Removal of larger change rate areas and smaller area: The area change rate is set
according to (3.4), and the threshold change rate is set to (0.5, 1). Extremum areas
that exceed this range will be cleared. Denote Ra and Rb as the areas of adjacent
areas respectively that change with depth values and that satisfy

jjRa j  jRb jj
0:5\ \1 ð4Þ
jRa j

When the maximum stable extremum area is found, the remaining area in the depth
image is considered as the moving object detection area. By performing the AND
 Maximally Stable Extremal Regions Improved Tracking Algorithm 551

operation with the suspected target contour previously analyzed by similarity, the target
contour to be tracked can be confirmed.

3.3 Predicting the Target Location

In the tracking process, the target may move in the vertical direction, so the depth value
may also change. In order to solve this problem, according to the physical meaning of
depth, the actual position of the target in the scene is displayed by establishing a spatial
rectangular coordinate system. The noise matrices of the observation matrix and the
state transition matrix are mutually independent Gaussian white noise matrices.
Denote ðxk ; yk Þ as the centroid of the horizontal plane, and ðvxk ; vyk Þ as its speed.
Denote G as the top point of the target in the Z direction G, which moves at the speed
of vzk , is at zk . The interval between the two continuous frames is short and it can be
considered as uniform motion. Denote Dt as the interval between two adjacent images
as.
The target’s motion state vector Xk satisfies

Xk ¼ ½xk yk zk vxk vyk vzk T ð5Þ

The observation state vector Zk satisfies:

Zk ¼ ½xk yk zk T ð6Þ

zk ¼ zk1 þ ðDtÞvzk1 ð7Þ

vzk ¼ vzk1 ð8Þ

Therefore, the system state transition matrix A satisfies:

0 1
1 0 0 Dt 0 0
B0 1 0 0 Dt 0C
B C
B0 0 1 0 0 Dt C
A=B
B0
C ð9Þ
B 0 0 1 0 0C C
@0 0 0 0 1 0A
0 0 0 0 0 1

The observation matrix C is:

0 1
1 0 0 0 0 0
C ¼ @0 1 0 0 0 0A ð10Þ
0 0 1 0 0 0

So the Kalman filter can be used to predict the motion of the target in the Z axis.
The highest point as a reference point is shown in Fig. 5.
552 H. Wang et al.

Camera

The highest
point G

Tracking goals

Fig. 5. Tracking target

4 Experimental Results and Analysis

Overlapping and occlusion experiments were carried out using the ToF camera in order
to validate the accuracy of the improved MSER tracking algorithm. The ToF camera
was placed on the top of the testing area to collect the images of the moving targets.
Tracking effect images with two and three targets can be seen in Fig. 6.

Fig. 6. Track multiple targets

 Maximally Stable Extremal Regions Improved Tracking Algorithm 553

The overlapped targets will be considered as one with traditional Camshift for that
the targets to be tracked cannot be distinguished from the disturbance of similar
characteristics and failed to be tracked when the target is sheltered (see Fig. 7). The
traditional Camshift algorithm works well when the targets do not overlap. However,
when the targets overlap, the target cannot be distinguished correctly, and the two
targets are mistaken as one. These problems are solved with an improved MSER
tracking algorithm in this paper where the targets can be precisely tracked when there
are many targets or the targets are sheltered (Tables 1 and 2).

Fig. 7. Traditional camshift algorithm tracking error map

Table 1. Track statistics of experimental results with overlapping targets

Tracking the number of Tracking Test Misuse rate Detection rate
targets algorithm samples (%) (%)
2 Camshift 500 2.6% 97.4%
Improved 1.5% 98.5%
MSER
3 Camshift 500 7.2% 92.8%
Improved 3.6% 96.4%
MSER

Table 2. Track statistics of experimental results of target occlusion

Tracking the number of Tracking Test Misuse rate Detection rate
targets algorithm samples (%) (%)
2 Camshift 500 5.7% 94.3%
Improved 2.5% 97.5%
MSER
3 Camshit 500 11.2% 88.8%
Improved 5.6% 94.4%
MSER
554 H. Wang et al.

5 Conclusion

In this paper, using the unique physical properties of depth images, an improved
tracking algorithm for MSER is proposed, and Kalman filter is used to achieve accurate
prediction and tracking of target locations. The target’s contour is extracted and con-
firmed iteratively. The accuracy of the similarity analysis is improved with the simi-
larity database update. However, this algorithm based on the contour curve and local
area feature matching in the depth image has a large amount of computation and
increases the complexity. At the same time, the effect is good in the stable area while
many problems remain in the complex ones which will be improved in the future
research.

References
1. Baqersad, M., Hamedi, A., Mohammadafzali, M., Ali, H.: Asphalt mixture segregation
detection: digital image processing approach. Adv. Mater. Sci. Eng. (2017)
2. Shotton, J., et al.: Efficient human pose estimation from single depth images. IEEE Trans.
Pattern Anal. Mach. Intell. 35, 2821–2840 (2013)
3. Biskup, M., Procaccia, E.B.: Eigenvalue vs perimeter in a shape theorem for self-interacting
random walks (2018)
4. Hou, H.N., Wu, T.: Pedestrian tracking algorithm based on camshift multi-feature adaptive
fusion. Electron. Qual. 3, 007 (2017)
5. Donoser, M., Bischof, H.: Efficient maximally stable extremal region (MSER) tracking. In:
2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition,
pp. 553–560 (2006)
6. Nguyen, T., Mann, G.K.I., Vardy, A., Gosine, R.G.: Developing a cubature multi-state
constraint Kalman filter for visual-inertial navigation system. In: Conference on Computer
and Robot Vision, pp. 321–328 (2017)
7. Lo, T.W.R., Siebert, J.P.: Local feature extraction and matching on range images: 2.5D
SIFT. Comput. Vis. Image Underst. 113, 1235–1250 (2009)
8. Qiao, B., Jin, L., Yang, Y.: An adaptive algorithm for grey image edge detection based on
grey correlation analysis. In: International Conference on Computational Intelligence and
Security, pp. 470–474 (2017)
9. Sun, C.: A Fast Stereo Matching Method. Digital Image Computing Techniques and
Applications, pp. 95–100 (1997)
10. Eichhardt, I., Jankó, Z., Chetverikov, D.: Novel methods for image-guided ToF depth
upsampling. In: IEEE International Conference on Systems, Man, and Cybernetics (2017)
11. Bevilacqua, A., Stefano, L.D., Azzari, P.: People tracking using a time-of-flight depth
sensor. In: IEEE International Conference on Video and Signal Based Surveillance, p. 89
(2006)
12. Sitnik, R.: New method of structure light measurement system calibration based on adaptive
and effective evaluation of 3D-phase distribution. In: Proceedings of SPIE - The
International Society for Optical Engineering, vol. 5856, pp. 109–117 (2005)
13. Salve, S.G., Jondhale, K.C.: Shape matching and object recognition using shape contexts. In:
IEEE International Conference on Computer Science and Information Technology, pp. 483–
507 (2010)
14. Astrachan, O.: Bubble sort: an archaeological algorithmic analysis. ACM (2003)
 Dynamic Hand Gesture Recognition Based
on the Three-Dimensional Spatial Trajectory
Feature and Hidden Markov Model

Kangli Liu1, Feixiang Zhou1, Haikuan Wang1(&), Minrui Fei1,2,

and Dajun Du1
1
School of Mechatronics Engineering and Automation, Shanghai University,
Shanghai, China
[email protected]
2
Shanghai Key Laboratory of Power Station Automation Technology,
Shanghai University, Shanghai 200444, China

Abstract. In the process of the traditional hand gesture recognition, complex

backgrounds, illumination and hand gesture changes in space can seriously
affect the accuracy of the dynamic hand gesture recognition. In this paper, an
improved dynamic hand gesture recognition method based on Three-
dimensional Time of Flight (3D-TOF) camera is proposed. Firstly, an adap-
tive segmentation algorithm is used for the hand gesture segmentation com-
bining the frame difference method and the depth threshold. Then a dynamic
hand gesture recognition algorithm based on the three-dimensional (3D) spatial
trajectory feature and Hidden Markov Model (HMM) is proposed which takes
full advantage of the depth data. And in order to improve the recognition rate of
the dynamic hand gesture, the misidentified samples are trained again. Experi-
mental results show that the algorithms proposed in this paper have a high
recognition rate of the dynamic hand gestures and a good robustness to the
different backgrounds and illumination.

Keywords: 3D-TOF
Dynamic hand gesture recognition

3D spatial trajectory feature
HMM

1 Introduction

With the progress of science and technology and the development of computer
application technology, the traditional mouse, keyboard and other human-computer
interaction means have been unable to meet the requirements of current users.
Therefore, how to develop a natural, friendly and harmonious human-computer
interaction which conforms to the human’s behaviour becomes a hot issue in the
present research. In recent years, hand gesture recognition has attracted much attention
as one of the most natural and friendly human-computer interaction [1, 2]. And it has
been widely applied to various fields such as virtual reality, unmanned vehicles and
smart home, etc.

© Springer Nature Singapore Pte Ltd. 2018

K. Li et al. (Eds.): ICSEE 2018/IMIOT 2018, CCIS 924, pp. 555–564, 2018.
https://doi.org/10.1007/978-981-13-2384-3_52
556 K. Liu et al.

In the traditional hand gesture recognition based on vision, the accuracy of

recognition can be easily affected by the complex background and illumination, etc.
With the constant progress of the depth sensing technology [3, 4], a hand gesture
recognition method based on the depth vision is proposed using the depth information
acquired from a depth camera, which successfully overcomes the problems existing in
the traditional hand gesture recognition to some extent. At present, more and more
scholars use the depth camera to accomplish the dynamic hand gesture recognition.
Hsieh et al. [5] adopted Haar features and support vector machine (SVM) to recognize
the dynamic hand gestures. Rahmani et al. [6] proposed a real-time dynamic hand
gesture recognition method based on histogram of depth gradients (HDG) and RF.
Yang et al. [7] proposed a method based on depth motion map (DMM). First, calculate
the three views of the depth map of each frame, then normalize the hand regions in each
view, after that divide each frame and accumulate to obtain the DMM, then extract the
corresponding HOG features, and finally use the SVM to classify. Premaratne et al. [8]
employed the changes of the direction angles of the hand gesture trajectory as the
features, and then achieved the recognition based on HMM. In the above literatures,
depth information is mostly applied to segment hand gesture regions, but not directly
used to extract the three-dimensional (3D) features of the hand gesture. Scholars
gradually introduced depth information into gesture description features, such as 3D
HOG (Histogram of Oriented Gradient) features and 3D HOF (Histograms of Oriented
Optical Flow) features based on 2D image features. Oreifej et al. [9] proposed a four-
dimensional (4D) HOG feature by adding a one-dimensional time factor on the basis of
the 3D HOG features. Geng et al. [10] proposed the hand shape features extracted from
the depth images and the spherical coordinate (SPC) feature extracted from the 3D
hand motion trajectories for sign language recognition. Yang et al. [11] presented a
flexible 3D trajectory indexing method for complex 3D motion recognition. Trajec-
tories are represented in the subprimitive level, the level between the point level and
primitive level.
The dynamic hand gesture recognition method based on 3D-TOF camera [12, 13]
studied in this paper takes full advantage of the depth information. In order to over-
come the changes of the complex background and illumination, an adaptive hand
gesture segmentation method is presented which combines the frame difference method
with the depth threshold. Then based on the HMM, this paper proposed a dynamic
hand gesture recognition algorithm based on the direction angle features of the 3D
spatial trajectory. Experimental results show that the average recognition rate of
dynamic hand gestures using the above algorithms reaches 94.5%, and the algorithms
have a good robustness to the different backgrounds and illumination.
The structure of this paper is organized as follows: This section is an introduction.
Section 2 describes the hand gesture segmentation under complex background and
different illumination. Section 3 describes a dynamic gesture recognition method based
on the 3D spatial trajectory feature and HMM. In Sect. 4, the experimental results are
analysed. Finally, some important conclusions and future work are discussed.
 Dynamic Hand Gesture Recognition 557

2 Hand Gesture Segmentation Under Complex Background

and Different Illumination

Hand gesture segmentation is the first and crucial step of the dynamic hand gesture
recognition. It means detecting and segmenting the motion hand region of interest
accurately from the complex background. The result of the hand gesture segmentation
will directly affect the recognition rate of dynamic hand gestures.
In this paper, a TOF camera with a resolution of 320  240 is used to obtain depth
images. Considering the actual situation, the hand gestures are closer to the camera than
the background and the human body. Therefore, the method based on the depth
threshold is usually applied for the hand gesture segmentation. In this paper, an
adaptive hand gesture segmentation method with the combination of the frame dif-
ference method and the depth threshold is put forward. This method not only over-
comes the influence of complex fore-background and different illumination, but also
realizes the adaptive hand gesture segmentation. The specific steps are as follows:
Step 1: Detect the depth image of the motion region of the current frame.


It ðx; yÞ jIt ðx; yÞ  It1 ðx; yÞj [ T1
Dðx; yÞ ¼ ð1Þ
0 jIt ðx; yÞ  It1 ðx; yÞj  T1

Where It ðx; yÞ is the depth image of the current frame, It1 ðx; yÞ is the depth image
of the previous frame, Dðx; yÞ is the depth image of the motion region.
Step 2: Calculate the average depth d of the depth image of the motion region, and
take the average d as the depth threshold T for the hand gesture segmen-
tation, as shown in the formula (2):

1 X X
T¼d¼ Dðxi ; yj Þ ð2Þ
240  320 i¼240 j¼320

Step 3: According to the depth threshold T, the hand gesture segmentation is

performed on the current frame, as shown in formula (3):


255 It ðx; yÞ\T  d
T ðx; yÞ ¼ ð3Þ
0 otherwise

Where Tðx; yÞ is the binary image of the segmented hand gesture region, d is the
given error of the depth threshold.
The result of the hand gesture segmentation is shown in Fig. 1. Figure 1(a) shows
the depth image from the TOF camera. Figure 1(b) shows the segmented hand region
by the above method. It can be seen that the hand gesture region is completely
segmented.
558 K. Liu et al.

(a) (b)

Fig. 1. (a) The original depth image (b) the binary image of the segmented hand gesture region

3 Dynamic Gesture Recognition Based on 3D Spatial

Trajectory Feature and HMM

3.1 Hand Gesture Tracking

Unlike the static hand gesture recognition in which the changes of the shapes and
geometric features of the hand are focused mostly, the dynamic hand gesture recog-
nition is more concerned about the motion trajectory of the hand gesture. In this paper,
the trajectory of the gravity center of the dynamic hand gesture represents the trajectory
of the hand gesture. According to the features of the moment of the contours, the
gravity center of the contour can be extracted, as shown in formula (4):
m10 m01
p ¼ ðx0 ; y0 Þ ¼ ð ; Þ ð4Þ
m00 m00

Where m10 and m01 are the first-order geometric moments of the contour image and
m00 is the zero-order geometric moment of the contour image.
Then the trajectory points are extracted every three frames denoted as Pðx; y; zÞ,
where z represents the depth value of the point.

3.2 3D Spatial Trajectory Feature Extraction

The extraction of hand gesture features is a very critical procedure in the hand gesture
recognition and can directly decide the accuracy of the recognition. In general, the
extraction of the dynamic hand gesture features is transformed into the extraction of the
dynamic hand gesture trajectory features. The common hand gesture trajectory features
include position, direction angles, and movement rate. In this paper, the direction angle
which contributes most to the recognition rate is selected as the hand gesture trajectory
feature. And in order to take full advantage of the depth information provided by the
TOF camera, the original the two-dimensional (2D) trajectory is replaced by a 3D
trajectory to increase the motion features in the depth direction. Thus, the 3D trajectory
features can not only reflect the plane hand gesture trajectory features but also the
spatial trajectory features, which is very effective for improving the accuracy of the
dynamic hand gesture recognition. Figure 2 shows the spatial trajectory and projections
of a dynamic hand gesture.
 Dynamic Hand Gesture Recognition 559

100 50
50 100

y
0
150
0
100 200
200
200
y 0 x 0 100 200 300
x

100 100

50 50

z
0 0
0 100 200 300 0 50 100 150 200
x y

Fig. 2. 3D trajectory and projection of the gesture

Assuming that the hand gesture trajectory points obtained at t  1 and t are
Pt1 ðxt1 ; yt1
 ; zt1 Þ and Pt ðxt ; yt ; zt Þ respectively. The direction angle of Pt1 Pt is
defined as hxyt ; hyzt ; hxzt , where hxyt , hyzt , and hxzt respectively are the orientation
angles of the trajectory points Pt1 and Pt in the XOY, YOZ, and XOZ plane. The
calculation method is shown in formula (5), (6) and (7):
8
>
> arctanðyxtt y xt1 Þ þ p
t1
x2  x1 \0
<
yt yt1
hxyt ¼ arctanðxt xt1 Þ þ 2p y2  y1 \0 : ð5Þ
>
>
: arctanð t t1 Þ
y y
otherwise
xt xt1

Where hxyt is the direction angle in the XOY plane.

8
>
> arctanðyztt z
yt1 Þ þ p
t1
y2  y1 \0
<
hyzt ¼ arctanðyztt z
yt1 Þ þ 2p
t1
z2  z1 \0 ð6Þ
>
>
: arctanð zt zt1 Þ otherwise
yt yt1

Where hyzt is the direction angle in the YOZ plane.

8
>
> arctanðxztt z
xt1 Þ þ p
t1
x2  x1 \0
<
hxzt ¼ arctanðxztt z
xt1 Þ þ 2p
t1
z2  z1 \0 ð7Þ
>
>
: arctanð zt zt1 Þ otherwise
xt xt1

Where hxzt is the direction angle in the XOZ plane.

Then the extracted direction angle features are quantized into 12 levels using 12-
direction chain codes. As shown in Fig. 3, the direction angle of each level is expressed
as a corresponding number between 0 and 11. Finally, the discrete eigenvectors are
obtained as the input to the HMM.
560 K. Liu et al.

9
8 10

7 11

5 1
2
4
3

Fig. 3. Direction angle quantization diagram

3.3 Dynamic Hand Gesture Recognition Based on the 3D Trajectory

Feature and HMM
The HMM describes a double stochastic process that randomly generates observation
sequences from a hidden Markov model. Because HMM has a strong ability to model
the atypical hand gestures and can describe the complex gestures and it has the
advantage of predefining the gesture models [14]. Therefore, the HMM algorithm [15–
17] has always been dominant in the dynamic hand gesture recognition. This paper
selects the HMM algorithm based on the 3D trajectory features to implement the
dynamic hand gesture recognition.
The complete set of parameters for an HMM can be represented as
k ¼ fM; N; A; B; pg, where N is the number of hidden states of the HMM, M is the
number of the observed values of the HMM, A is the state transition probability matrix,
B is the observation probability matrix, and p is the initial probability vector. This
paper adopts the simpler parameter representation of the HMM, that is, k ¼ fp; A; Bg.
Once the model parameters of the HMM are determined, the HMM is developed.
The dynamic hand gesture recognition based on the HMM is mainly divided into
two stages, namely the training stage of the HMM and the recognition stage of the
dynamic hand gestures. The basic steps are as follows:
Step 1: Use the left-right strip HMM topology and initialize the model parameter to
k ¼ fp; A; Bg.
Step 2: Using the Baum-Welch learning algorithm, estimate the new model
 
parameters k ¼ p; A; B through continuous iterative operations based on
the 3D gesture spatial trajectory feature vectors which is input as the
observation sequence O.

Step 3: Use the fore-backward algorithm to calculate PðOjkÞ and PðOkÞ, that is, the
probability of the observation sequence O under the model parameters k and
    
k, and then calculate log PðOkÞ  log PðOjkÞ. If log PðOkÞ  log P
ðOjkÞj\e (e is a given convergence threshold value), then converges.
The HMM trained at this time is closest to the observed value. So far, the
training process is completed. Otherwise making k ¼ k, continue the step 2

until PðOkÞ converges.
 Dynamic Hand Gesture Recognition 561

Collecting hand gestures

Segmentation and tracking of hand gestures

Extraction of the 3D trajectory features

HMM1 P(O | λ 1) HMM2 P(O | λ 2) HMMV P(O | λ V )

Maximum probability

Recognition result

Fig. 4. Flow chart of dynamic hand gesture recognition

An HMM is trained for each given dynamic hand gestures, then the motion tra-
jectory direction angle feature vector of a sample is taken as the input observation
sequence of the HMM. Finally calculate PðOjki Þ (0  i  V, where V is the number of
the defined dynamic hand gestures). The hand gesture of the corresponding HMM with
the highest probability is considered as the recognition result. The dynamic hand
gesture recognition flow chart is as shown in Fig. 4.
Since all the samples are treated uniformly by using the Baum-Welch classic
training algorithm, PðOjki Þ may eventually converge to the local optimal value
resulting in the poor recognition of some samples. On the basis of this circumstance,
some improvements have been made in this paper. The training process of the hand
gesture samples is linked with the test results of the model, and the HMM is corrected
by using the test results of the model. If some hand gesture samples are misrecognized
as other gestures, then those samples will be input into the corresponding HMM again
for a second training to correct the model parameters. Finally, after performing this
operation several times, the optimal parameters of the HMM is determined.

4 Experimental Results and Analysis

Based on the principles of the practicality, simplicity and identifiability of the hand
gestures and conforming to people’s habits, this paper defines four basic dynamic hand
gestures, as shown in Fig. 5, which are the trajectories of the four specific dynamic
hand gestures respectively. The small solid black circle represents the starting position
of the hand gesture, and the direction of the black solid arrow is the direction of the
motion trajectory of the dynamic hand gesture, wherein the gesture 4 represents
moving hand from far to close.

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

Fig. 5. The trajectories of the four specific dynamic hand gestures: (a) gesture1, (b) gesture2,
(c) gesture3, (d) gesture4
562 K. Liu et al.

In order to verify the accuracy of the algorithms in this paper, a real-time dynamic
hand gesture recognition experiment was performed. In the experiment, a total of ten
volunteers were arranged to stand in front of the 3D-TOF camera and repeated four
kinds of dynamic gestures shown above 20 times each. Therefore, each volunteer need
to complete 80 gestures and the total test number for each gesture is 200.

4.1 Experimental Results Analysis and Comparison

The algorithms proposed in this paper are adopted to detect and recognize the dynamic
hand gestures in the real time. The experimental results show that the accuracy of these
four specific dynamic hand gesture recognition is higher than 90%, and the average
recognition accuracy is 94.5%.
When processing a relatively simple gesture, such as gesture 1, the recognition rate
is highest up to 100%. When processing a gesture with complicated changes, such as
the gesture 3, the accuracy of the recognition is reduced to some extent. But generally,
the recognition effect satisfies the requirements. The specific recognition results are
shown in Table 1.

Table 1. The accuracy of dynamic hand gesture recognition

Dynamic hand gesture Test samples Correct Accuracy Average accuracy
gesture1 200 200 100% 94.5%
gesture2 200 190 95%
gesture3 200 180 90%
gesture4 200 186 93%

At the same time, the recognition accuracy of dynamic hand gesture based on the
3D trajectory features is evaluated by comparing with the 2D trajectory features used in
the past. The comparison result is shown in Fig. 6.

Fig. 6. Comparison of hand gesture recognition accuracy based on 3D trajectory features and
2D trajectory features.
 Dynamic Hand Gesture Recognition 563

The experimental results show that when recognizing gesture 2, gesture 3 and
gesture 4, the recognition accuracy is certainly improved by extracting the 3D trajec-
tory features proposed in this paper instead of the previous 2D trajectory features. This
shows that the depth information does increase the recognition accuracy.

4.2 Robust Verification

In order to verify the robustness of the algorithms under different backgrounds and
illumination, the defined dynamic hand gesture 2 was recognized in the four cases of
natural light & simple background, natural light & complex background, dim light &
simple background and dim light &complex background. And 50 experiments were
performed in each case. The experimental results are shown in the following Table 2:

Table 2. Recognition accuracy under different backgrounds and lighting conditions

Simple background Complex background
Natural light 95.1% 94.3%
Dim light 94.2% 93.5%

From the table above, it can be seen that there is no dramatic change of the
recognition rate of the dynamic hand gestures when the background and illumination
change. Therefore, the proposed algorithms for the dynamic hand gesture recognition
have a good robustness to the different backgrounds and illumination.

5 Conclusion

In this paper, the dynamic hand gesture recognition method based on 3D-TOF camera
is principally investigated. In the preprocessing stage, an adaptive hand gesture seg-
mentation method combining the frame difference method and depth threshold is
adopted to successfully overcome the influence of complex background and different
illumination. In the process of the dynamic hand gesture recognition, the 3D spatial
trajectory feature vectors of the dynamic gestures are input into the HMM for training
and recognition. In addition, the misidentified samples are trained for the second time
to improve the training method of the HMM. The experimental results show that the
dynamic gesture recognition algorithms based on 3D trajectory features and HMM are
very effective for improving the accuracy of the hand gesture recognition and
increasing the range of the hand gesture recognition, and have a good robustness to
different backgrounds and illumination.
In the actual human-computer interaction, more abundant hand gesture changes are
needed. At the same time, more 3D space trajectory features should be explored to
improve the accuracy and generalizability of the algorithm presented in this paper.
564 K. Liu et al.

References
1. Rautaray, S.S., Agrawal, A.: Vision based hand gesture recognition for human computer
interaction: a survey. Artif. Intell. Rev. 43, 1–54 (2015)
2. Song, Y., Demirdjian, D., Davis, R.: Continuous body and hand gesture recognition for
natural human-computer interaction. ACM Trans. Interact. Intell. Syst. (TiiS)-Spec. Issue
Affect. Interact. Nat. Environ. 2, 1–28 (2012)
3. Liu, K., Chen, C., Jafari, R., Kehtarnavaz, N.: Fusion of inertial and depth sensor data for
robust hand gesture recognition. IEEE Sens. J. 14, 1898–1903 (2014)
4. Munaro, M., Basso, A., Fossati, A., Gool, L.V., Menegatti, E.: 3D reconstruction of freely
moving persons for re-identification with a depth sensor. In: 2014 IEEE International
Conference on Robotics and Automation (ICRA), pp. 4512–4519. IEEE, Hong Kong (2014)
5. Hsieh, C.-C., Liou, D.-H.: Novel Haar features for real-time hand gesture recognition using
SVM. J. Real-Time Image Process. 10, 357–370 (2015)
6. Rahmani, H., Mahmood, A., Huynh, D.Q., Mian, A.: Real time action recognition using
histograms of depth gradients and random decision forests. In: IEEE Winter Conference on
Applications of Computer Vision, pp. 626–633. IEEE, Steamboat Springs (2014)
7. Yang, X., Zhang, C., Tian, Y.: Recognizing actions using depth motion maps-based
histograms of oriented gradients. In: Proceedings of the 20th ACM international conference
on Multimedia, pp. 1057–1060. ACM, Nara (2012)
8. Premaratne, P., Yang, S., Vial, P., Ifthikar, Z.: Dynamic hand gesture recognition using
centroid tracking. In: Huang, D.-S., Bevilacqua, V., Prashan, P. (eds.) ICIC 2015. LNCS,
vol. 9225, pp. 623–629. Springer, Cham (2015). https://doi.org/10.1007/978-3-319-22180-
9_62
9. Oreifej, O., Liu, Z.: HON4D: histogram of oriented 4D normals for activity recognition from
depth sequences. In: 2013 IEEE Conference on Computer Vision and Pattern Recognition,
pp. 716–723. IEEE, Portland (2013)
10. Lubo, G., Xin, M., Haibo, W., Jason, G., Li, Y.: Chinese sign language recognition with 3D
hand motion trajectories and depth images. In: Proceeding of the 11th World Congress on
Intelligent Control and Automation, pp. 1457–1461. IEEE Shenyang (2014)
11. Yang, J., Yuan, J., Li, Y.F.: Flexible trajectory indexing for 3D motion recognition. In: 2015
IEEE Winter Conference on Applications of Computer Vision, pp. 326–332. IEEE,
Waikoloa (2015)
12. Corti, A., Giancola, S., Mainetti, G., Sala, R.: A metrological characterization of the kinect
V2 time-of-flight camera. Robot. Auton. Syst. 75, 584–594 (2016)
13. Li, L., Xiang, S., Yang, Y., Yu, L.: Multi-camera interference cancellation of time-of-flight
(TOF) cameras. In: 2015 IEEE International Conference on Image Processing (ICIP),
pp. 556–560. IEEE, Quebec City (2015)
14. Kumar, P., Gauba, H., Roy, P.P., Dogra, D.P.: Coupled HMM-based multi-sensor data
fusion for sign language recognition. Pattern Recogn. Lett. 86, 1–8 (2017)
15. Raheja, J.L., Minhas, M., Prashanth, D., Shah, T., Chaudhary, A.: Robust gesture recognition
using Kinect: a comparison between DTW and HMM. Opt.-Int. J. Light. Electron Opt. 126,
1098–1104 (2015)
16. Liu, K., Chen, C., Jafari, R., Kehtarnavaz, N.: Multi-HMM classification for hand gesture
recognition using two differing modality sensors. In: 2014 IEEE Dallas Circuits and Systems
Conference (DCAS), pp. 1–4. IEEE, Richardson (2014)
17. Zhang, X.H., Wang, J.J., Wang, X., Ma, X.L.: Improvement of dynamic hand gesture
recognition based on HMM algorithm. In: 2016 International Conference on Information
System and Artificial Intelligence (ISAI), pp. 401–406. IEEE, Hong Kong (2016)
 Author Index

Adams, Jonathan I-239 Cheng, Tingli II-65

Afonso, Daniel I-40 Chi, Junfeng II-215
Agarwal, Dheeraj I-73 Chunli, Zhu II-225
Allen, David I-108 Collins, R. I-3
An, Ran III-3 Cong, Yu II-261
Armstrong, Cecil G. I-73 Crookes, Danny II-468
Azhar, Irfan II-352
Dai, Wentao I-363
Bai, Long I-571
de Sousa, Ricardo Alves I-40
Bao, Weihua II-191, II-291
Deng, Lei I-97
Barden, John I-108
Deng, Tiantai II-468
Barnes, David I-239
Deng, Weihua II-383
Behera, Amar Kumar I-40
Ding, Yulong II-478, III-214
Bi, Guangjian I-418, I-428
Dong, Jing III-253
Bi, Ruyu III-247
Dong, Shiyun II-458
Bi, Yining III-173
Dong, Xiaomin I-257, II-15, II-77, II-491
Bibi, Shamsa II-352
Drummond, Stephen I-108
Bo, Zhengjun I-325
Du, Dajun II-555, III-36, III-94, III-253,
Buchanan, Fraser I-228
III-314
Butterfield, Joe I-3, I-239
Du, Qianqian I-332
Butterfield, Joseph I-108

Cai, Rong III-3 Fan, Hong-Wei I-290

Cai, Wenyue III-264 Fang, Jie II-502
Cai, Yunlong I-484 Fei, Minrui II-281, II-302, II-536, II-555,
Cameron, Che II-408 III-14
Cao, Huajun III-129 Feng, Sheng I-529
Cao, Le I-318, I-325 Fu, Hao I-325
Cao, Zhanyong III-61 Fu, Jingqi II-155, II-191, II-291, II-429
Carbery, Caoimhe M. II-169 Fu, Rao I-249
Chen, Chen I-418, I-440 Fu, Yanxiang I-308
Chen, Chin-Yin I-548, I-561 Fu, Zhiqiang III-314
Chen, Chin-yin I-84
Chen, Erheng III-129 Gao, Qingxuan II-3
Chen, Guodong I-548 Gao, Wenzhong III-151
Chen, Jinquan I-571 Gao, Yang I-290
Chen, Minyou II-65, III-284 Gao, Yuan I-13
Chen, Ping II-502 Gao, Zhengming II-191, II-291
Chen, Rui I-571 Garcia-Palacios, Emiliano II-400
Chen, Si-Lu I-84 Ge, Hong-Yu I-290
Chen, Wanmi II-236 Geng, Haipeng I-529
Chen, Yi I-394 Gou, Honggang I-221
Chen, Youling I-279 Gou, Mingli II-3
Cheng, Lin III-151 Gu, Yaqi III-46
566 Author Index

Guan, Tianming I-121 Jiang, Feiyang I-465, I-474, I-518

Guiwu, Xiong II-77 Jiang, Gaofeng II-180
Guiyun, Jiang II-352 Jiang, Zhu III-214
Guo, Song II-502 Jin, Hongzhou II-236
Guo, Yanchun I-300 Jin, Yan I-40, I-239, I-249
Guo, Yue III-192
Guo, Zhen I-146, III-70 Leng, Guanghui III-214
Li, Chen III-105
Han, Junwei II-321 Li, Chunyan III-105, III-264
Han, Zhi III-14 Li, Deshun I-394
Hao, Liangliang III-225 Li, Engzhong II-458
Hassan, Jawad Ul I-267 Li, Huilin III-61
Hassan, Muhammad Arshad Shehzad I-267, Li, Jiajia II-109
II-65, III-284 Li, Jie I-146, II-44, III-70
He, Bing II-89 Li, Jing I-221
He, Feng III-61 Li, Junjie I-548
He, Qinyi III-129 Li, Kang I-343, II-408, II-419, III-183,
He, Shui I-363 III-201
He, Siyuan I-332 Li, Lieqiang I-325
He, Yan III-139 Li, Longxiao I-131
He, Zeyin I-506 Li, LongXiao I-191
Higgins, Colm I-3, I-249 Li, Lu II-44
Higgins, Peter I-3, I-108 Li, Mengyun I-300
Hong, Jie III-36 Li, Qiang III-284
Hong, Xiao II-205 Li, Quan I-363
Hou, Kai III-192 Li, Shawn III-183
Hou, Kaihu II-443, III-163 Li, Tao I-300
Hu, Bo III-24 Li, Wei I-540
Hu, Henry Y. K. I-169, II-25 Li, Weihua II-15
Hu, Jin-xing II-89 Li, Wenfeng II-491
Hu, Linming III-139 Li, Xiang I-22, III-201
Huang, GuoQin III-24 Li, Xue III-36, III-94, III-253, III-295,
Huang, Tao I-169, II-25 III-314
Huang, Xiaowei II-215 Li, Yan I-179, III-105
Huang, Xingjun I-146, II-44, III-70 Li, Yanxue III-183
Huang, Yun I-363 Li, Yi I-169, II-25
Li, Yihuan III-183
Li, Yulong I-63, I-371
Iqbal, Nadeem II-352 Li, Zexin I-97
Li, Zhiguo II-99
Jafar, Salman III-129 Li, Zhihao I-408
Jan, Y. I-3 Liang, Ji-Fei II-330
Ji, Zexue II-281 Liang, Yuanyuan II-458
Jia, Bin III-306 Liao, Jianquan III-274
Jia, Hongjie III-192 Liao, Wenzhu I-308
Jia, Siqi I-332 Lim, Ming K. I-157, I-179, I-198, I-209,
Jia, Zhenyuan I-382 III-139
Jiahua, Suolang I-363 Lin, Tao III-247
Jiang, Cheng II-144 Lin, Tengjiao I-465, I-474, I-484, I-494,
Jiang, Dianke III-236 I-506, I-518
 Author Index 567

Lin, Yun I-131, I-146, II-44, III-70 Mu, Zongyi I-63

Ling, Lei I-279 Murphy, Adrian I-3, I-40, I-108, I-249
Liu, Boyu III-236
Liu, Chunyao III-306 Naeem, Wasif II-246
Liu, Hao I-382 Nguyen, Hien M. II-400
Liu, Huazhen III-116 Ni, Hangpeng III-247
Liu, Jian I-279 Ni, Zhonghua I-22
Liu, Jianbo I-474 Nisar, M. Usman I-50
Liu, Jiangtao II-261 Nisar, Muhammad Usman I-267
Liu, Jun I-494 Niu, Shilin I-318, I-325
Liu, Kaipeng I-131 Niu, Yufei I-279
Liu, Kangli II-536, II-555
Liu, Li II-512 Pan, Chengwang II-491
Liu, Lin III-225 Pan, Hui II-365
Liu, Peng II-452 Pang, Jihong II-55
Liu, Qiang I-548 Pang, Zezhao I-394
Liu, Qing II-478 Peng, Quancheng I-506
Liu, Weiqi I-257 Peng, Shitong I-300
Liu, Wen I-484, I-494 Peng, Weiqing II-311
Liu, Xiaohui III-173
Liu, Xiaojun I-22 Qi, Hu III-14
Liu, Xuan III-201 Qian, Feng I-13
Liu, Ying I-363, II-99 Qian, Hong II-144, II-180, III-46
Liu, Yiping II-99 Qian, Lin II-311
Liu, Yuqing I-157 Qian, Wen III-247
Lou, Baodong II-261 Qiu, Xiaoli I-22
Lu, Xingxing I-465, I-518 Qiu, Yilin I-571
Luo, Geshan I-363 Qu, Hao II-44
Luo, Hao II-205, II-215 Qu, Zhiyong II-321
Luo, Hongfei III-264
Luo, ShaQi III-24 Rakhio, Allah I-257
Luo, Simin III-81 Ran, Yan I-63, I-371, I-451, II-55
Luo, Yuanxin I-267 Rehman, Shafiq Ur II-352
Lv, Hesheng I-494, I-506 Rehman, Waheed Ur I-267, II-352
Ren, Guangzhen II-215
Ma, Guocai I-408 Ren, Xianlin I-394
Ma, Haiping III-173 Robb, Stephen I-108
Ma, Jingjie III-295 Robinson, Trevor T. I-73
Ma, Ling I-267
Ma, Shiqian III-192 Saleem, Farrukh I-50, I-267, II-352
Ma, Xiangjin III-236 Shan, Weilu III-94
Ma, Yunqiang II-272 Shao, Meiyang III-225
Mao, Minghe II-261 Shen, Weili II-291
Marshall, Adele H. II-169 Shen, Zhenyu I-169, II-25
McLoone, Seán II-246 Sheng, Qi III-284
Menary, Gary I-228 Shi, Chao III-151
Menhas, Muhammad Ilyas II-311, III-14 Shi, Chen III-3
Mo, Fan II-121 Shi, Suilin I-221
Mu, Yunfei III-192 Shoaib, Muhammad I-267, II-352
568 Author Index

Siddiqui, Fahad II-468 Wang, Ruiting II-55

Song, Chengze III-225 Wang, Shufen II-383
song, Sunhao I-548 Wang, Weijun I-548
Su, Zhen II-155 Wang, Xiaonan I-418
Sun, Dan I-249 Wang, Xin II-330
Sun, Hao II-34 Wang, Xu I-97, I-131, I-157, I-191
Sun, Haoxiang II-536, II-546 Wang, Xudong III-192
Sun, Li I-121 Wang, Yacan I-332
Sun, Peiliang I-343 Wang, Yan III-139
Sun, Shuaishuai II-15 Wang, Yang II-236
Sun, Yi III-70 Wang, Yingxiang II-65, III-284
Sun, Yongmin II-390 Wang, Yongqin I-371
Wang, Yue III-192
Tan, Junyang II-458 Wang, Yulin III-139
Tang, Min II-99 Wang, Zegang I-418, I-428
Tang, Qian I-169, I-571, II-25 Wei, Huidong I-228
Tang, Yingjia I-131, I-191 Wei, Jing I-474, I-506
Taylor, C. I-3 Wei, Lisheng II-272
Teng, Ru-Min II-330 Wen, Peihan I-50
Tian, Liting III-151 Wen, Yalan II-311
Tian, Xiao III-306 Wen, Yuhao III-274
Tian, Yan III-81 Weng, Daiyun I-221
Tong, Xiaoying I-121 Woods, Roger II-169, II-468
Trimble, Shane II-246 Wu, Junyong III-3, III-225
Wu, Li I-571
Ul Hassan, Jawad I-50 Wu, Pengcheng III-139
Wu, Shanjin III-46

Wan, Piao I-325 Xi, Jun II-491

Wang, Biao II-133 Xi, Yawen III-3
Wang, Chongchong I-561 Xia, Dan II-458
Wang, Dong I-418, I-428, I-440 Xiang, Min II-390
Wang, Fuji I-418, I-428, I-440 Xiang, Tao II-502
Wang, Haikuan II-525, II-536, II-546, II-555 Xiang, Tianchun III-192
Wang, Hakuan II-281 Xiao, Guijian I-221, I-363
Wang, Hao I-209 Xiao, Wenlei I-408
Wang, Hongqing III-173 Xiao, Youqiang III-247
Wang, Hongwei I-63 Xie, Dong II-525, II-546
Wang, Jiahui II-144 Xie, Linjun I-318
Wang, Ji-Zhao II-330 Xie, Wenrui III-116
Wang, Kun III-129 Xie, Yunhan I-332
Wang, Le II-281 Xin, Li II-225
Wang, Lei I-382 Xin, Qiang I-561
Wang, Li II-478 Xiong, Shiquan II-121, II-133
Wang, Ling II-311, II-365, III-14 Xiong, Shuping II-109
Wang, Long I-279 Xiong, Tianyu I-169, II-25
Wang, Pan I-50 Xiong, Weiqing I-179, I-198
Wang, Ping II-261 Xu, Binshi II-458
Wang, Qi I-428, I-440 Xu, Junfeng I-267
Wang, Qianggang III-274 Xu, Lei II-302
 Author Index 569

Xu, Mengyao II-133 Zhang, Gen-bao I-451

Xu, Sheng II-272, III-314 Zhang, Hui II-443, III-163
Xu, Xin II-502 Zhang, Ji I-318
Xu, Zhu III-61 Zhang, Jianhua III-306
Xunbo, Lu II-225 Zhang, Jie I-84, I-548
Zhang, Jingxia III-81
Yan, Biwu II-205, II-215 Zhang, Jinhong I-484
Yan, Fangfei II-133 Zhang, Li I-382
Yan, Shiyong I-228 Zhang, Qian I-50, III-105, III-116
Yang, Aolei II-281, II-302, II-512 Zhang, Runzhi III-70
Yang, Baisong I-529 Zhang, Shaohua III-295
Yang, Ge II-89 Zhang, Sheng-yong I-451
Yang, Guilin I-548, I-561 Zhang, Xiaofeng II-512
Yang, Lujia I-249 Zhang, Xiaogang I-371
Yang, Min I-408 Zhang, Xiaoheng III-81
Yang, Su II-3, II-109 Zhang, Yanjin II-365
Yang, Wenwen I-146 Zhang, Yanjun I-484, I-494
Yang, Xiao III-236 Zhang, Yinfeng II-155
Yang, Yu II-321 Zhang, Yunhuai I-97
Yang, Zhile II-525 Zhang, Zhehan I-394
Yi, Qian II-109, II-121, II-133 Zhao, Gang I-408
Yi, Shijia I-465 Zhao, Xudong II-390
Yi, Shuping II-109, II-121, II-133 Zhao, Yuanfang I-13
Yin, Cheng II-400 Zhao, Zirui I-465, I-474, I-518
Yin, Mingjia II-419 Zheng, Lei II-205, II-215
Yin, Pengxiang II-205, II-215 Zheng, Min II-419
Yin, Xiaoyu I-332 Zheng, Shutao II-321
Yin, Yuan II-236 Zheng, Tianjiang I-548
Yongqin, Wang II-352 Zheng, Wei III-81
Yu, Gaofu III-46 Zheng, Xiaojun II-34
Yu, Guobing I-484, I-494 Zheng, Yujie II-429
Yu, Jiangyou I-318, I-325 Zhou, Bowen III-81, III-236
Yu, Jianqiang II-15 Zhou, Feixiang II-525, II-555
Yu, Jin III-24 Zhou, Fuli I-157
Yu, Lie I-529 Zhou, Hao I-169
Yu, Ziheng III-236 Zhou, Jie I-84
Yuan, Fang I-97 Zhou, Nianchen III-274
Yuan, Li-Ming I-84 Zhou, Peng II-342
Yuan, Yuan II-180 Zhou, Wenju II-302, II-512, II-546
Yuanxin, Luo II-352 Zhou, Zhou II-443, III-163
Yue, Jun II-512 Zhu, Haoxiang II-191
Zhu, Honghao II-458
Zeng, Pu II-342 Zhu, Minjie III-173
Zhan, Qiang I-540 Zhu, Xiaowen III-3
Zhang, Anchuan II-121 Zhu, Yi III-116
Zhang, Bangjie I-318 Zhu, Yulin II-44
Zhang, Chi I-84, I-548 Zou, Ben III-306
Zhang, Chong I-440 Zou, Lai I-363
Zhang, Genbao I-63, I-371 Zuo, Lidan I-279