Recent Developments in Intelligent Computing, Communication and Devices: Proceedings of ICCD 2019 [1st ed.] 9789811558863, 9789811558870

Citation preview

Advances in Intelligent Systems and Computing 1185

C. H. WU Srikanta PATNAIK Florin POPENTIU VLÃDICESCU Kazumi NAKAMATSU   Editors

Recent Developments in Intelligent Computing, Communication and Devices Proceedings of ICCD 2019

Advances in Intelligent Systems and Computing Volume 1185

Series Editor Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland Advisory Editors Nikhil R. Pal, Indian Statistical Institute, Kolkata, India Rafael Bello Perez, Faculty of Mathematics, Physics and Computing, Universidad Central de Las Villas, Santa Clara, Cuba Emilio S. Corchado, University of Salamanca, Salamanca, Spain Hani Hagras, School of Computer Science and Electronic Engineering, University of Essex, Colchester, UK László T. Kóczy, Department of Automation, Széchenyi István University, Gyor, Hungary Vladik Kreinovich, Department of Computer Science, University of Texas at El Paso, El Paso, TX, USA Chin-Teng Lin, Department of Electrical Engineering, National Chiao Tung University, Hsinchu, Taiwan Jie Lu, Faculty of Engineering and Information Technology, University of Technology Sydney, Sydney, NSW, Australia Patricia Melin, Graduate Program of Computer Science, Tijuana Institute of Technology, Tijuana, Mexico Nadia Nedjah, Department of Electronics Engineering, University of Rio de Janeiro, Rio de Janeiro, Brazil Ngoc Thanh Nguyen , Faculty of Computer Science and Management, Wrocław University of Technology, Wrocław, Poland Jun Wang, Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Shatin, Hong Kong

The series “Advances in Intelligent Systems and Computing” contains publications on theory, applications, and design methods of Intelligent Systems and Intelligent Computing. Virtually all disciplines such as engineering, natural sciences, computer and information science, ICT, economics, business, e-commerce, environment, healthcare, life science are covered. The list of topics spans all the areas of modern intelligent systems and computing such as: computational intelligence, soft computing including neural networks, fuzzy systems, evolutionary computing and the fusion of these paradigms, social intelligence, ambient intelligence, computational neuroscience, artificial life, virtual worlds and society, cognitive science and systems, Perception and Vision, DNA and immune based systems, self-organizing and adaptive systems, e-Learning and teaching, human-centered and human-centric computing, recommender systems, intelligent control, robotics and mechatronics including human-machine teaming, knowledge-based paradigms, learning paradigms, machine ethics, intelligent data analysis, knowledge management, intelligent agents, intelligent decision making and support, intelligent network security, trust management, interactive entertainment, Web intelligence and multimedia. The publications within “Advances in Intelligent Systems and Computing” are primarily proceedings of important conferences, symposia and congresses. They cover significant recent developments in the field, both of a foundational and applicable character. An important characteristic feature of the series is the short publication time and world-wide distribution. This permits a rapid and broad dissemination of research results. Indexed by SCOPUS, DBLP, EI Compendex, INSPEC, WTI Frankfurt eG, zbMATH, Japanese Science and Technology Agency (JST), SCImago. All books published in the series are submitted for consideration in Web of Science.

More information about this series at http://www.springer.com/series/11156

C. H. WU Srikanta PATNAIK Florin POPENTIU VLÃDICESCU Kazumi NAKAMATSU •

•

•

Editors

Recent Developments in Intelligent Computing, Communication and Devices Proceedings of ICCD 2019

123

Editors C. H. WU Department of Supply Chain and Information Management (SCM) The Hang Seng University of Hong Kong (HSUHK) Shatin, New Territories, Hong Kong Florin POPENTIU VLÃDICESCU Information and Communication Engineering Faculty Automatic Control and Computer Science University of Oradea Oradea, Romania

Srikanta PATNAIK Department of Computer Science and Engineering Faculty of Engineering and Technology Siksha ‘O’ Anusandhan University Bhubaneswar, Odisha, India Kazumi NAKAMATSU School of Human Science and Environment University of Hyogo Himeji, Japan

ISSN 2194-5357 ISSN 2194-5365 (electronic) Advances in Intelligent Systems and Computing ISBN 978-981-15-5886-3 ISBN 978-981-15-5887-0 (eBook) https://doi.org/10.1007/978-981-15-5887-0 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 This work is subject to copyright. All rights are solely and exclusively licensed 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, expressed 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

The integration of Machine Learning, Big Data and cloud computing has lead to evolution of new paradigms of Intelligent Computing, Communication and Devices, as well as their extensive applications. Although, technically these technologies are distinct and very vast, but when integrated with each other, together they lead to disruptive and next generation devices and technologies with applications in various sectors. This edition of proceedings cover a collection of recent research works presented at the 5th International Conference on Intelligent Computing, Communication and Devices (ICCD 2018) held at Xi’an, China during November 22–24, 2019. The research papers provide a wide coverage to all the aspects of next generation computing, communication and intelligent devices. The papers are broadly categorized into three major parts (i) Intelligent Computing (ii) Next Generation Communication and Networking and (iii) Intelligent Devices. The first part “Intelligent Computing” addresses research works on intelligent computing in distributed networks and cloud-based networks. It further includes papers on soft computing, data mining, data exchange systems, data crawling, cluster analysis, context-based adaptive filters, automatic decision support systems, collaborative systems etc. The second part namely, “Next generation Communication and Networking” covers recent research and trends being developed in various areas such as M2M communications and networks, signal transmissions, signal processing of mechanical vibrations, artificial intelligence based network security, fault detection and diagnosis. It further exhibits work on ad-hoc networks, radar signals, wideband differential phase shifters, Dual-band antenna array, miniaturized antenna units and their transmission lines etc.

v

vi

Preface

Last but not the least, part on “Intelligent Devices” covers various hybrid devices developed for various industrial sectors to increase productivity of the sector while coping with various challenges and circumstances.

Shatin, Hong Kong Bhubaneswar, India Oradea, Romania Himeji, Japan

Editors C. H. WU Srikanta PATNAIK Florin POPENTIU VLÃDICESCU Kazumi NAKAMATSU

Acknowledgements

The contributions covered in this proceeding are the outcome of the contributions from more than one hundred researchers. We are thankful to the authors, paper contributors of this volume and the departments which support the event a lot. We are thankful to the Editor-in-Chief of the Springer Book series on “Advances in Intelligent Systems and Computing” Prof. Janusz Kacprzyk for his support to bring out the fourth volume of the conference i.e. ICCD 2019. It is noteworthy to mention here that constant support from the Editor-in-Chief and the members of the publishing house makes the conference fruitful for its fifth edition. We would like to extend our heartfelt thanks to Dr. Aninda Bose, Senior Editor, Interdisciplinary and Applied Sciences, Computational Intelligence, Springer and his publishing team. for his encouragement and support. We are thankful to Hamido Fujita, Director of Intelligent Software Systems Iwate Prefectural University, Japan, for his well-researched keynote address “New Challenges in Machine Learning: Multiclass-Classification for Risk Predictions in Health Care Applications” and equally thankful to Prof. Gang Sun, University of Electronic Science and Technology of China (UESTC), for his path breaking talk. Lastly but not the least, we are thankful to our friend Prof. Dr. Sc. Kazumi Nakamatsu, School of Human Science and Environment, University of Hyogo, Japan, Prof. Dr. Andrew W. H. Ip, University of Saskatchewan, Department of Mechanical Engineering, Canada, for their support and guidance, We are also thankful to the experts and reviewers who have worked for this volume despite of the veil of their anonymity. We are also thankful our academic partners: Interscience Research Network (IRNet), Hunan University of Finance and Economics, Changsha, China; School of Electrical and Information Engineering, Hunan University of Technology, Zhuzhou, China; IRnet International Academic Communication Center, China; Financial Big Data Science and Technology Key Laboratory of Hunan Province, Changsha, China; Hunan provincial 2011 Collaborative Innovation Center for Development and Utilization of Finance and Economics Big Data Property, Changsha, China; Hunan Province Higher Educational Institutions Key Laboratory

vii

viii

Acknowledgements

“Information Technology and Information Security”, Changsha, China; Key Laboratory for Electric Drive Control and Intelligent Equipment of Hunan Province, China; Guangdong Graphic Image Association, China; Guangdong Massive Biometric Information Processing Engineering Technology Research Center; Jiangmen Computer Federation, China and lastly, but not the least Interscience Institute of Management and Technology, Bhubaneswar, India. We look forward to your valued contribution and support to next editions of the 6th International Conference of Intelligent Computing, Communication and Devices (ICCD 2020), whose venue will be announced shortly. We are sure that the readers shall get immense benefit and knowledge from the fourth volume of the area Intelligent Computing, Communication and Devices. Editors C. H. WU Srikanta PATNAIK Florin POPENTIU VLÃDICESCU Kazumi NAKAMATSU

Contents

Intelligent Computing Multi-SoftMax Convolutional Neural Network and Its Application in the Diagnosis of Planetary Gearbox Complicated Faults . . . . . . . . . . Shixin Zhang, Qingquan Lv, Shenlin Zhang, and Jianhua Shan

3

Signal Complexity Measure Based on Construction Creep Rate in 3s Feature Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jinwei Cai, Xueqin Chen, Peng Ming, and Wenshi Li

12

Research on Partition Technology of Real-Time Database in Big Data Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yuxuan Li and Suling Li

18

RCS Sequence Period Estimation Method Based on Constant False Alarm Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xinbing Fang, Sifang Liu, and Xiquan Wang

23

Target Recognition of RCS Sequence Based on Neural Network . . . . . . Gaofeng Pan, Sheng Liang, Sifang Liu, and Xie Yong

29

Encouraging Cooperation in Participatory Sensing with Reputation and Game Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tianqi Zhang, Rong Zhang, and Jin Wang

36

Multi-granularity Attribute Sentiment Analysis Based on Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Runxue Chen, Peng Yan, and Lian Xu

43

Study on Fuzzy Crop Planning Problems Based on Credibility Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hai Tao Zhong, Ming Fa Zheng, Fang Chi Liang, and Ya Nan Wang

55

ix

x

Contents

Empirical Research on the Antecedents of Enterprise IS Users’ Voice Behavior Based on the Work Behavior Theory . . . . . . . . . . . . . . . . . . . Manhui Huang and Changxian Li

64

Image Fusion Based on Masked Online Convolutional Dictionary Learning with Surrogate Function Approach . . . . . . . . . . . . . . . . . . . . . Chengfang Zhang and Xingchun Yang

70

The Liquid Cooling Technique Based on Foam Metal Copper . . . . . . . . Xue Wang, Gang Huang, NaiKuo Chen, and Changsheng Jing Study on the Forecast of Driving Reaction Based on PNN and FCM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jinjian Xiao, Yubo Weng, and Yingna Xie Android Malware Detection via Behavior-Based Features . . . . . . . . . . . Yanxia Li and Yihong Liu Analysis of the Power Converter Used in Wind Power Generator System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jinping Zhang, Xushan Han, Long Zhao, Dingmei Wang, Jin Li, and Chengzhi Ma

75

84 92

99

Evaluation of New Energy Theory Based on Data Recovery . . . . . . . . . 105 Qingquan Lv, Qiang Zhou, Long Zhao, Mingsong Wang, Pengfei Gao, and Jin Li Summary, Reflection, and Prospect of Wind Power Development in China . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 Qiang Zhou, Xushan Han, Qingquan Lv, Chenyun Shen, Mingsong Wang, and Zhenzhen Zhang Time-Varying Probability Distribution Model of High Permeability Wind Power Considering Non-stationary Characteristics . . . . . . . . . . . . 117 Zhenzhen Zhang, Qiang Zhou, Pengfei Gao, Jin Li, Dingmei Wang, and Jianmei Zhang An Adaptive T-Distribution Variation Based HS Algorithm for Power System ED . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 Zhili Ma, Xun Zhang, Hui Yuan, Xiaoqin Zhu, Fan Yang, Enzhan Zhang, Qiwen Zhang, and Yukai Yao Application of Improved Particle Swarm Optimization in Economic Dispatch of Power System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 Jianglong Nie, Zhouqiang He, Hui Yuan, Zhiru Li, Lina Gao, Qiwen Zhang, Enzhan Zhang, and Yukai Yao Design of an Ultra-Short-Term Wind Power Forecasting Model Combined with CNN and LSTM Networks . . . . . . . . . . . . . . . . . . . . . . 141 Yinghui Zhang, Shiyuan Zhou, Ziqiang Zhang, Liang Yan, and Li Liu

Contents

xi

An Algorithm for Intelligent Attack Decision-Making . . . . . . . . . . . . . . 146 Shouwei Hu, Wenjun Huang, Ying Cai, Qi Tang, and Xiaogang Qi Wavelet Packet Energy Entropy Based Feature Analysis of Seismic Signals from Vehicle Targets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 Kai Ding, Xinghua Li, Hao Li, Huayuan Ma, Lei Fan, and Xiaogang Qi Group Decision-Making Approach Under Interval-Valued Dual Hesitant Fuzzy Unbalanced Uncertain Linguistic Environments and Its Application to Community Tourism Resilience Evaluation . . . . 159 Junling Zhang, Ying Hong, Jun Ma, and Xiaowen Qi Analysis on the Operational Effectiveness of Certain Landmine Based on the AHP-ADC Combination Model of Computing . . . . . . . . . 166 Zhang Nan, Wang Xiao, Wang Xin, and Wu Zhiqi Research on a Monitoring and Alarm System for Safe Distance of Aerial Work Based on Single-Chip Microcomputer . . . . . . . . . . . . . . 174 Lang Pei and Jinhua Cai Design and Implementation of Database Query System for Civil–Military Integration Regulations Based on B/S Mode . . . . . . . 180 Huang Tianming Decision Problems for Three Subclasses of Regular Languages . . . . . . . 186 Hui Xu, Jing Tian, and Jia Liu Research on Clustering Identification Method Based on Path Sampling in Support Vector Clustering . . . . . . . . . . . . . . . . . . . . . . . . . 193 Wang Shiqiang, Gao Caiyun, Zeng Huiyong, Bai Juan, Zong Binfeng, and Cai Jiliang Just-in-Time Purchasing and Its Application in Manufacturing Enterprises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198 Shengchun Liu and Hong Zhao Exploring New Ways of Calligraphy Education Communication in the Age of Digital Media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203 Xiaoqing Zhou Dynamic Modeling and Analysis of Ammunition Loading Robot Based on Screw Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209 Yang Li, Da Xu, and Cheng Zhou Vision-Based Irregular Car Parking Behaviors Detection in the Underground Garage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219 Pengfei Wang, Zhenyuan Xu, Libin Cen, Jianxiang Xiang, Wenqing Wang, and Fei Gao

xii

Contents

Early Warning of Regional Energy Security Exogenous Source Based on CBR with Relative Ranking Entropy . . . . . . . . . . . . . . . . . . . 225 Hu Jian, Hu Jiali, and Xu linlin Using Hybrid Method Based on Machine Learning for Energy Consumption Prediction of Oil and Gas Production . . . . . . . . . . . . . . . 234 Jun Li, Yidong Guo, Yiran Wang, and Ying Qi Next Generation Communication and Networking Path Tracking Control for Ship Collision Avoidance When Overtaking in Narrow Channel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241 Renqiang Wang, Keyin Miao, Jianming Sun, Yue Zhao, Hua Deng, and Jiabao Du Research on the Position Controlling the Silhouette of a Style Women’s Coat Based on Eye Tracking Technology . . . . . . . . . . . . . . . . 249 Yang Wang, Yue Li, Xiaogang Wang, and Lingfeng Song Estimating Model Parameters of Concrete Materials Based on Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256 He Yu and Shouju Li Hierarchical Extraction Algorithm of Video Summary Based on Multi-feature Similarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263 Zhou Ju and Luo Bing Signal Complexity Measures Based on Ising Model . . . . . . . . . . . . . . . . 271 Meng Zhang, Hao Wu, Jinwei Cai, and Wenshi Li A High-Precision Method of Time Delay Estimation in Complex Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277 Xiaogang Qi, Lieping Yuan, Yu Chen, and Lifang Liu Sleeping-Based Opportunistic Routing Algorithm for Relay Node Selection in Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . 283 Xiaogang Qi, Haijun Dong, Hao Li, and Lifang Liu Multiple-Model Based EM Algorithm for Multirate ARX Model with Time-Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290 Yingjiao Rong, Yuelin Xu, Kai Ding, and Xiaogang Qi Application of Orientation Determination Technology Using Navigation Satellite System in Intelligent Minefield . . . . . . . . . . . . . . . . 298 Ning Yu, Shoujun Zheng, Wenjun Huang, Kai Ding, and Xiaogang Qi Research on ECG Signal Compression Algorithm Based on Compressed Sensing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305 Xiao Jian, Hu Fang, and Zhang Kai

Contents

xiii

Passive Image Copy–Move Forgery Detection Based on ORB Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312 Zhao Xue, Lihua Tian, and Chen Li A Novel Star Image Preprocessing Method Based on Image Subdivision and Multi-threshold Segmentation . . . . . . . . . . . . . . . . . . . 318 Yiyang He, Hongli Wang, Lei Feng, Sihai You, and Yongqiang Xiao Comparison of Semiparametric Method and BP Neural Network in Missing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327 Fanghui Zhai, Sanzhi Shi, and Simin Fan Construction and Application of Bilingual Terminology Database of Political Field Based on Chinese–English Parallel Corpus . . . . . . . . . 332 Jiaxin Lin and Wei Li Real-Time Diagnosis Algorithm Design of Household Wage Survey Data Accuracy Based on Fisher Discriminant Method . . . . . . . . . . . . . . 339 Chengfang Shen, Dalong Mo, and Wentao Huang A Review of Application of Computer Vision in Fruit Picking Robot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346 Huiping Si, Jianghong Lv, Kaiyan Lin, Junhui Wu, and Jie Chen Construction Method of Knowledge Base for Power Grid-Aided Decision Based on Knowledge Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . 356 Ying Chen, Zhifang Liao, Bin Chen, Hui Shi, and Huanwen Chen An Automatic Removal Method of Ocular Artifacts in EEG . . . . . . . . . 362 Mingai Li, Fan Liu, Yanjun Sun, and Lina Wei Research and Application of Clustering Algorithm in Battlefield Scheduling Genetic Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372 Jialun Li, Peng Wang, Xiaoyan Li, Zhigang lv, and Bowen Fu DeepPlaner: Query Optimization Using Recurrent Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379 Shize Xu, Yongming Zhang, Rong Bian, Fei Cheng, and Xiaofei Li A Novel Voiceprint Verification Technology Through Deep Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389 Wen Jun, Zhan Yu, and Song Wenhao Study on the Network Information Security Problems Under the Environment of Big Data Cloud Computing . . . . . . . . . . . . . . . . . . 396 Feng Zhao and Manling Chen

xiv

Contents

Intelligent Devices Improved H∞ Filtering Method for Pulsar Position Error Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405 Yongqiang Xiao, Hongli Wang, Lei Feng, Sihai You, and Yiyang He Stochastic Resonance Driven by Fractional Gaussian Noise in Neuron Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413 Gao Fengyin, Li Lin, and Ren Jinshen Research of Radar Emitter Signal Model and Clustering Sorting Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 418 Wang Shiqiang, Gao Caiyun, Zeng Huiyong, Bai Juan, Zong Binfeng, and Cai Jiliang The Action Evaluation of Model Catwalk Based on Gait Segmentation and Template Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424 Mingdong Wang, Haitao Fan, Jijun Tong, and Lurong Jiang Hesitant Fuzzy Linguistic MAGDM with Prioritization Relation Among Attributes and Application to Sustainability Evaluation of Ecotourism Development Practices . . . . . . . . . . . . . . . . . . . . . . . . . . . 431 Junling Zhang, Ying Hong, Jun Ma, and Xiaowen Qi Research on Nonlinear Characteristics of Ship Flow in Y-Type Waterway . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438 Wan Jianxia and Zhang Shukui Jatropha Seed Selection Technology Based on Machine Vision . . . . . . . 444 Baolin Zhu, Yanbo Zhang, Xinyu Zhang, Ruonan Xing, and Mingyue Zhao The Influence of Hypocrite on Cluster Motion Under Vicsek Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450 Xinyu Zhang, Yanbo Zhang, Baolin Zhu, Ruonan Xing, and Zidong Chai Lyapunov Inequality for a Fractional Differential Equation Modeling Damped Vibrations of Thin Film MEMS . . . . . . . . . . . . . . . . . . . . . . . . 454 Liana Eneeva, Arsen Pskhu, Alexander Potapov, Tianhua Feng, and Sergo Rekhviashvili Modeling Damped Vibrations of Thin Film MEMS: Fractional Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 458 Sergo Rekhviashvili, Arsen Pskhu, Alexander Potapov, Tianhua Feng, and Liana Eneeva Fast Measurement Method of Non-contact Measurement for Oil Pipe with Screw Thread . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464 Gengpei Zhang, Fenfen Liu, Xianggui Ming, Ren Lu, Hanyang Wang, Liangliang Li, Tao Luo, and Kui Tian

Contents

xv

Error-Correction Coding of QR Code Based on Guruswami–Sudan Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470 Rongjun Chen, Yongxing Yu, Yan Liu, Songjin Liu, Hong-Zhou Tan, and Huimin Zhao Analysis Method of Sensing Information in Electric Power Distribution Based on Edge Computing . . . . . . . . . . . . . . . . . . . . . . . . . 477 Wang Chengliang, Shui Weilian, Yang Qingsheng, and Ning Yan Simulation and Optimization of Automobile Back Door Lock . . . . . . . . 482 Wang Song, Ma Xiuying, and Ma Qiang Thermal Elastohydrodynamic Lubrication of Line Contact Considering the Thermal Elastic Deformation . . . . . . . . . . . . . . . . . . . . 488 Zunyou Lu, Yanjun Lu, and Yongfang Zhang Research on the Robot Uncalibrated Visual Servo Method Based on the Kalman Filter with Optimized Parameters . . . . . . . . . . . . 494 Qian Gao and Wei Xiao Design of Visual Adaptive Control System for Mobile Robot Based on Sliding Mode Speed Tracking . . . . . . . . . . . . . . . . . . . . . . . . . 504 Qian Gao and Wei Xiao Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517

About the Editors

Dr. C. H. WU is an assistant professor at the Department of SCM, Hang Seng University of Hong Kong. He has presented his projects and findings on IoT, engineering optimisation and digital health at 15+ international conferences and published on them in 60+ international refereed journals. In 2019, Dr WU was the recipient of the Outstanding Paper Award (Emerald Literati Network Awards) and the Top Peer Reviewer Award (Publons). He is the managing editor of the journal Enterprise Information Systems and associate editor of both the Journal of Organizational and End User Computing, and the International Journal of Engineering Business Management. Dr. Srikanta PATNAIK is a full professor at the Department of Computer Science and Engineering, Faculty of Engineering and Technology, Siksha ‘O’ Anusandhan (SOA) University, Bhubaneswar, India. Dr PATNAIK has published 60+ research papers in international journals and conference proceedings. He is the author of 2 textbooks and has edited 12 books and numerous book chapters published by leading international publishers. He is the editor-in-chief of the International Journal of Information and Communication Technology, of the International Journal of Computational Vision and Robotics, and of the book series Modeling and Optimisation in Science and Technology, published by Springer. Dr. Florin POPENTIU VLÃDICESCU is an associate professor at the Department of Control Engineering and Industrial Informatics, University “Politehnica” of Bucharest and Co-Chairholder of the UNESCO Chair in Information Technologies, University of Oradea. He has been a Visiting Professor at various universities, e.g. the Université Pierre et Marie Curie Paris, Delft University of Technology, etc. He has worked for many years on problems associated with software reliability. He has published 100+ papers in international journals and conference proceedings and is the author of 1 book and co-author of 4 more. He is currently an associate editor of the International Journal of Information and Communication Technology.

xvii

xviii

About the Editors

Dr. Kazumi NAKAMATSU received his Doctor of Sciences (D.Sc.) from Kyushu University, Japan. He has developed a number of paraconsistent annotated logic programs and applied them to various intelligent systems. He has contributed over 150 journal/conference papers and book chapters, edited/authored 12 books, and is the Founding Editor of the International Journal of Reasoning-based Intelligent Systems.

Intelligent Computing

Multi-SoftMax Convolutional Neural Network and Its Application in the Diagnosis of Planetary Gearbox Complicated Faults Shixin Zhang, Qingquan Lv, Shenlin Zhang, and Jianhua Shan(B) School of Mechanical Engineering, Anhui University of Technology, Maanshan 243002, China [email protected]

1 Introduction Planetary gearbox is widely used in wind power generation, metallurgy, shipping and lifting transportation. To keep safety, fault diagnosis of planetary gearbox is particularly important [1, 2]. The planetary gearbox is in typical compound motion, whose vibration response is more complex than that of the other gearboxes with bearings and fixed shaft drive [3]. So Scholars have done a lot of research on these fault diagnosis problems. Yu [4] from the university of Toronto, Canada, proposed a fault diagnosis method of planetary gearbox based on wavelet transform and time-domain average. In 2017, Cheng [5] from Hunan university proposed a fault diagnosis method of planetary gearbox based on ASTFA and SDEO regulation. Above methods are only applicable to the steady-state condition and complicated mathematics method. In order to overcome the shortage of these methods, an innovative method is required to propose to solve the problems of fault diagnosis of planetary gearbox. In recent years, development of deep learning has been fast, where image recognition, natural language processing, and other fields have had a great success. Convolutional neural network [5] was an important member of the deep learning, which was used widely. The scholars began to try to apply it in fault diagnosis of rotating machinery, mainly in bearing fault diagnosis. In 2016, Wang [6] from China university of petroleum adopted convolutional neural network in rotating mechanical fault diagnosis, whose input directly adopted the truncated one-dimensional vibration data. In 2017, Chen Lu of Beijing university of aeronautics and astronautics [7], Zhang Wei of Harbin Institute of Technology [8], and Liu Ruonan of Xi’an Jiaotong University [9] assembled one-dimensional vibration signals into two-dimensional signals and then inputted into convolutional neural network for fault diagnosis. In 2017, Xia [10] from the University of British Columbia, Canada, proposed the rotating mechanical fault diagnosis based on convolutional neural network with multi-sensor fusion. There are some deficiencies in the above papers: (1) The length of input sample is arbitrarily set by humans in a fixed value, which is difficult to generalize to the conditions with different sampling frequencies and rotating speeds. (2) Comparing with bearing diagnosis, due to small difference of vibration waveform, the fault diagnosis of planetary gearbox is more difficult. (3) The working condition is relatively simple; the rotation speed is fixed or the range of

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 C. H. WU et al. (eds.), Recent Developments in Intelligent Computing, Communication and Devices, Advances in Intelligent Systems and Computing 1185, https://doi.org/10.1007/978-981-15-5887-0_1

4

S. Zhang et al.

variation is small; the diagnosis problem of the working condition is not studied. (4) For complex faults, the effect of single fault diagnosis cannot be shown. In view of the above shortcomings, this paper propose an improvement scheme to solve the problem of fault diagnosis of planetary gearbox. (1) The calculation formula of sample data length is given, which can adapt to the working conditions of different rotating speed and sampling frequency. And random interception method is used to create samples to improve the network generalization ability. (2) The fault tree structure is proposed, which can deal with single and compound faults uniformly and check the diagnosis effect of each node. (3) In order to complete diagnose by the fault tree structure, the Multi-SoftMax Convolutional Neural Network is proposed. (4) In this paper, large convolutional window, size of pooling step, and the DenseNet are adopted to simplify the network structure and solve the complex fault diagnosis problem of planetary gearbox in variable working conditions.

2 Fault Tree Structure and Muti-SoftMax Convolutional Neural Network 2.1 Fault Tree Structure In fact, fault types may be particularly complex, including single fault, compound fault, and various fault degrees. In order to deal with various fault situations uniformly, a fault tree structure is designed, as shown in Fig. 1. SoftMax1

Normal

Fault

SoftMax2

Single fault

Compound fault SoftMax4

SoftMax3

Sun wheel+Bearing Planet wheel+Bearing

Sun wheel Bearing SoftMax5

SoftMax6

SoftMax8

SoftMax7

Rolling elements

Slight

Medium

Outer ring

Inner ring

Heavy

Serious

Slight

Medium

Wear

Cracks

Missing teeth

Broken teeth

Outer ring

Rolling elements

Inner ring

Missing teeth

Wear

Cracks

Broken teeth

SoftMax9

Heavy

Serious

Heavy

Serious

Slight

Medium

Heavy

Serious

Slight

Medium

Fig. 1 The structure of planetary gearbox fault tree

In the fault tree structure, normal (no fault) and fault are the root nodes of the fault tree, then to branch down to the leaf nodes. Each node corresponds a SoftMax, showing a fault type.

Multi-SoftMax Convolutional Neural Network and Its Application …

5

2.2 Working Parallel Structure In addition to different fault types, the planetary gearbox also has variable working conditions, including the rotational speed and load. As shown in Fig. 2, according to the actual working conditions, several typical rotational speed and load can be added to the working condition parallel structure.

Revolving speed SoftMax10

Revolving speed SoftMax11

Load 2

Load 3

Load 1

Revolving speed 3

Revolving speed 1

Revolving speed 2

Fig. 2 Working parallel structure

2.3 Structure of Multi-SoftMax Convolutional Neural Network Multi-SoftMax convolutional neural network structure includes input layer, convolutional layer, maximum pooling layer, average pooling layer, and Multi-SoftMax output layer. The multi-SoftMax output layer is the output layer with multiple SoftMax score vectors, and the score vector set is S = {SoftMaxi |i = 1, . . . , m}. As shown in Fig. 3, each SoftMaxi is fully connected with the last feature layer. Input layer

Convolutional Layer Pooling layer

The last feature layer

Mutil-SoftMax output SoftMax 1

SoftMax 2

Fig. 3 The structure of Multi-SoftMax convolutional neural network

The empirical risk loss of Multi-SoftMax convolutional neural network is hugely different from the traditional convolutional neural network. For example, in Fig. 1, the

6

S. Zhang et al.

fault type of planetary gearbox is the sun wheel failure, but it is not known exactly where the sun wheel is failed. And, the subscript set of the score vector SoftMax is {1, 2, 3}. The fault type is the crack fault of sun wheel, and the degree is medium. The subscript set of SoftMax score vector is {1, 2, 3, 5, 9}. In addition, SoftMax score vector in parallel structure of working conditions is also included. It is obvious that the sample corresponds to multiple SoftMax score vectors, and the number of SoftMax score vectors is indefinite.

3 Process of Planetary Gearbox Fault Diagnosis SoftMax convolutional neural network can be divided into four steps as follows: (1) Create training database. The length of the vibration data intercepted at random locations is larger than a maximum data period, that is number of data points of the planetary gearbox rotating once at the lowest speed. Repeat the above process to create enough sample data as training database. The label of the sample is created according to the fault tree structure and the working condition parallel structure. (2) Create Muti-SoftMax convolutional neural networks. The network structure except SoftMax output layer is based on DenseNet to design, whose size of input layer is H * 1 * K, where H is sample data length, width 1 is data width, and K is the data dimensions (number of sensor). Using multiple sensors to collect signals and using these signals for fault diagnosis can improve the accuracy of diagnosis, which is equivalent to multi-sensor fusion in the data layer. (3) Train SoftMax convolutional neural network. In each mini-batch, the contained samples were different. Until the end of the training, each sample is only trained once. All these mini-batch samples constitute the training set. According to debug the superparameters, better parameters are obtained and network generalization performance is improved. (4) Diagnose planetary gear box fault. Input the sample data into the trained MutilSoftMax convolutional neural network and get the score vector of SoftMax. Start with the SoftMax score vector of the root node and judge from the maximum position down.

4 Example of Planetary Gearbox Fault Diagnosis This example adopts the self-test fault data of the planetary gearbox in the laboratory. The device is shown in Fig. 4. Vibration data are collected by two sensors. There are four kinds of sun wheel speed in the experiment: 10, 15, 20, and 25 Hz. Due to experimental limitations, compared with the fault tree structure in Fig. 1, the fault types of the sun wheel are relatively simple. The single fault of the sun wheel is broken tooth, cracks, wear, and missing tooth. The rotation frequency is 10, 20, and 25 Hz. The compound faults include the sun wheel fault and the fault of bearing outer ring. The sun wheel fault is the broken tooth, cracks, wear, and missing tooth. The rotating frequency is 15 and 20 Hz. Loads are divided into two types: load and no-load.

Multi-SoftMax Convolutional Neural Network and Its Application …

Motor

Sensor 1

Planetary Gearbox

7

Sensor 2

Fig. 4 Experimental device

The vibration waveforms in different working conditions are shown in Fig. 5, which can be found that the vibration waveform of broken teeth and cracks are very close, in the same working condition, indicating that the fault diagnosis of planetary gearbox is difficult. The vibration mode of missing teeth is obvious and easy to recognize.

Fig. 5 Vibration waveform in different working conditions. a 10 Hz, no-load, single fault. b 15 Hz, load, compound fault. c 20 Hz, no-load, compound fault. d 25 Hz, load, single fault

4.1 Structure of Convolutional Neural Network Vibration data of the sampling frequency is 8192 Hz, the minimum frequency for 10 Hz. At this time, the sun wheel rotate once, the number of data points is 820, which is a data cycle. 1280 is beyond a data cycle, which are selected as number of continuous sample data points at random locations, as shown in Fig. 6. 4.2 Structure of Mutil-SoftMax Convolutional Neural Network Convolutional layer using large convolutional kernels and step length, pooling layer also uses large pooling window and step length, which can reduce the number of network

8

S. Zhang et al.

Fig. 6 Create sample data

layers. Activation function is Max (−1, X) in activation layer. To solve the problem that too deep network layer leads to difficulty in training and too shallow network leads to insufficient feature extraction, the double path is appended in the network referring to DenseNet, whose structure is as shown in Fig. 7. Input layer Convolutional Layer 1 Maximum pooling layer 1

Parameters of each layer

Convolutional Layer 2 Maximum pooling layer 2 Convolutional Layer 3 Maximum pooling layer 3

Average pooling layer 2

Route 1 Convolutional Layer 4

Route 2

Average pooling layer 1

Fully Connected Layer Mutil-SoftMax output layer Output layer

Fig. 7 Mutil-SoftMax convolutional neural network used in the experiment

5 Experiment When the number of iterations i < 7000, the learning rate was 0.001. The regularization coefficient was 0.00001. The mini-batch number was 32, and the momentum coefficient was 0.9. When i ≥ 7000, the learning rate was 0.0005. The regularization coefficient was 0.00001. The mini-batch number was 32, and the momentum coefficient was 0.9.

Multi-SoftMax Convolutional Neural Network and Its Application …

9

Due to that weight was initialized as a random number of gaussian distribution, the mean value was zero and the variance was 0.1. And the bias was initialized as zero. The data multiplied by 10, as the input data. The early termination method was adopted to stop the training in 21,400 times, and the accuracy of the test set was tested after the training. 100 samples were collected for each working condition as the test data set. And the diagnosis effect was determined by two indicators. The first indicator was the accuracy of all samples, which means result was completely consistent with the given label. And the accuracy was 97.55%. Another was the recall rate of each node in the tree structure diagram and the working condition parallel structure, as shown in Fig. 8. Based on above two indicators, it was obvious that the method can accurately diagnose the complex fault types of planetary gearbox. Normal

Fault

97.26

99.89

Single fault

Compound fault

97.26

99.89

Broken Missing Cracks Wear teeth teeth 99.71

99.61

99.86

99.86

Revolving speed

Broken Missing Cracks Wear teeth teeth 97.60

98.44

97.36

96.56

10

15

20

25

100

99.28

99.07

100

Fig. 8 The recall rate of each node in experiment 1

In order to test the generalization ability of Muti-SoftMax convolutional neural network, vibration data of planetary gearbox in no-load condition were taken as training samples and vibration data in load condition as test samples. The sample labels and network was unchanged. The number of iterations was increased to 4000. The learning rate was changed, and training was stopped after 6000 times. The number of iterations was significantly lower, which indicated that the training difficulty was reduced. In Fig. 9, the sample accuracy was 88.83%.

Broken Cracks teeth 89.22

73.56

Normal

Fault

99.41

98.67

Single fault

Compound fault

96.92

99.83

Wear

Missing teeth

99.78

100

Broken Cracks teeth 85.39

94.56

Revolving speed

Wear

Missing teeth

10

15

20

25

89.78

90.1

100

99.28

99.07

100

Fig. 9 The recall rate of each node in experiment 2

10

S. Zhang et al.

Due to load, the training sample and test sample were different. The accuracy was significantly less than that of the experiment 2. But multiple samples can be captured in each condition. The fault type of the most occurrence samples represented the final fault type of planetary gearbox. Even though the accuracy of a single sample was 88.83%, the probability of misjudgment of fault types was very low. In Fig. 9, it was found that the recall rate of compound faults was much higher than that of its lower nodes, which indicates that the samples of compound faults were basically correctly diagnosed, but the specific fault types were wrongly diagnosed sometimes. If the fault type was not considered and only considered the presence or absence of fault and single or compound fault, the recall rate was very high, whose lowest value was 96.92%.

6 Conclusion The fault tree and working condition parallel structure are proposed for the first time, which can deal with various complex fault types and variable working conditions uniformly. At the same time, diagnosis effect of each node can be checked. Mutil-SoftMax convolutional neural network is proposed for the first time, which can deal with fault tree and working condition parallel structure. Referring to the DenseNet network structure, it makes training easier. Meanwhile, the activation function is Max(−1, X). Each kind of fault types of planetary gearbox in the laboratory are successfully diagnosed by the above method, and the accuracy of experiment 1 reaches 97.55%.

References 1. Lei, Y.G., He, Z.J., et al.: Research advances of fault diagnosis technique for planetary gearboxes. J. Mech. Eng. 47(19), 59–67 (2011) 2. Zhou, Z.K.: Summary of research on fault diagnosis technology of wind turbine gearbox. Technol. Mark. 23(4), 25–28 (2016) 3. Zou, J.C.H., Shen, Y.D.: Review of gearbox fault diagnosis techniques under variable conditions. Mech. Drive 36(8), 124–127(132) (2012) 4. Yu, J., Yip, L., Makis, V.: Wavelet analysis with time-synchronous averaging of planetary gearbox vibration data for fault detection, diagnostics, and condition based maintenance. In: 2nd International Conference on Mechanical and Electronics Engineering, Kyoto, Japan, 1–3 Aug 2010, pp. 132–136. IEEE (2010) 5. Cheng, J.S., Yang, X.K., et al.: A method of planetary gearboxes fault diagnosis based on ASTFA and SDEO demodulation. Noise Vib. Control 37(2), 137–142 (2017) 6. Li, Y.D., Hao, Z.B., Lei, H.: Summary of the convolutional neural network. J. Comput. Appl. 37(9), 1946–1953 (2016) 7. Wang, J.J., Zhuang, J.H., Duan, L.X., et al.: A multi-scale convolutional neural network for featureless fault diagnosis. In: 2016 International Symposium on Flexible Automation, Cleveland, Ohio, USA, 1–3 Aug 2016 8. Wang, J.J., Zhuang, J.H., Duan, L.X., et al.: A multi-scale convolutional neural network for featureless fault diagnosis.In: 2016 International Symposium on Flexible Automation, Cleveland, Ohio, USA, 1–3 Aug 2016

Multi-SoftMax Convolutional Neural Network and Its Application …

11

9. Zhang, W., Peng, G.L., Li, C.H.: Bearings fault diagnosis based on convolutional neural networks with 2-D representation of vibration signals as input. In: MATEC Web of Conferences vol. 95, p. 13001 (2017) 10. Liu, R.N., Meng, G.T., et al.: Dislocated time series convolutional neural architecture: an intelligent fault diagnosis approach for electric machine. IEEE Trans. Industr. Inform. 13(3), 1310–1320 (2017)

Signal Complexity Measure Based on Construction Creep Rate in 3s Feature Space Jinwei Cai, Xueqin Chen, Peng Ming, and Wenshi Li(B) School of Electronics and Information Engineering, Soochow University, Shizi Road No. 1, Suzhou City, China [email protected]

1 Introduction In the year of Periodic Table of The Elements, we have to think deeply about the similar evolution of chaotic features. We need to bravely propose and apply new chaotic criteria in the perspective of signal complexity research, comparing known chaotic criteria with each other, which hang on the chaos criterion tree embedded in various dimensionentropy dynamics and in the simplest 0-1 test for chaos [1–5]. Compared to 0-1 test for chaos and Kaplan–Yorke dimension, the goals of this work include that (1) to illustrate our new features of chaos measure, 3s plot and construction creep (CC) rate [6–8], (2) to search for new maximal complexity in simple attractor reported by Ref. [9].

2 Methodology The growing rule of chaos criteria tree refines the collection of chaotic features. The main characterization paths of nonlinear dynamics involve dimensions, entropies, and stochastic processes, as well as noise-inserting strategies [6]. As a primer chaotic feature, 0-1 test for chaos in essence is based on group extension for picking up order or chaotic state. Here its 3D bi-compression type embedded in hyperbolic tangent function is our idea of chaotic signal complexity measures. 2.1 0-1 Test for Chaos Review To see Formulae (1), (2), and (3), we know that 0-1 test for chaos uses sine and cosine functions to form 2D artificial random process (pq-plot), wherein the time factor c = 0–2π is difficult to select for different raw data of ϕ(n) in data length of N [5]. p(n) =

n 

ϕ(j) cos(θ (j)), n = 1, 2, 3, . . .

(1)

ϕ(j) sin(θ (j)), n = 1, 2, 3, . . .

(2)

j=1

q(n) =

n  j=1

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 C. H. WU et al. (eds.), Recent Developments in Intelligent Computing, Communication and Devices, Advances in Intelligent Systems and Computing 1185, https://doi.org/10.1007/978-981-15-5887-0_2

Signal Complexity Measure Based on Construction Creep …

θ (j) = jc +

j 

ϕ(i), j = 1, 2, 3, . . . , n

13

(3)

i=1

If we open the third axis for pq-plot, we will find new information in 3D feature space depicted as followed. 2.1.1 3D Representation of 0-1 Test for Chaos To build up Formula (4) with hyperbolic tangent function, we gained the third axis for expanding into a three-dimensional feature space. Herein the most efficient classification function in deep learning was utilized [7]. s(j) =

n 

ϕ(j)tansig(θ (j)), n = 1, 2, 3, . . .

(4)

j=1

2.2 New Algorithm (Spring Test) Block of 3s Plot and CC Rate 3D bi-compression type with activated hyperbolic tangent function and its hard multithreshold copy was used as 3s-plot feature space (see Fig. 1). The numerical metric of the “spring shape feature” was obtained by the self-similarity measure.

Fig. 1 New algorithm (spring test) block of 3s plot and CC rate

The key techniques of self-similarity measure are: (1) The Euclidean distances between the spring points be made into a two-dimensional array based on the delay parameter m; (2) Based on new threshold γ in range of (0.01–0.99), the differences between the adjacent two rows were calculated as average CC rate; and (3) After traversing m from 1 to int(N/2), the minimum CC rate had been touched.

14

J. Cai et al.

2.3 Two Equations as Test Cases To check the recognition effects of our spring test (3s plot and CC rate) for complexity measures, we chose two following nonlinear equations [9]. 2D equation: Henon map is as formula in (5). Xn+1 = 1 − aXn2 + bYn , Yn+1 = Xn

(5)

3D equation: Maximum KY dimension Lorenz in parameter R is as (6). x˙ = y − x, y˙ = −xz, z˙ = xy − R

(6)

3 Results and Discussions 3.1 Results from pq-Plot of 0-1 Test to 3s Plot and CC Rate in Spring Test Figure 2 shows the power of data visualization and characterization. Key conditions: Probing factor c = 1, γ = 0.02; Initial values [0.1, 0.1, 0.1]; Raw data head cut with 1000 pints and followed 2000 points be used. From 2D old feature to 3D new features, ring versus cloud are characterized into cup versus dragon, or good spring versus bad spring, for distinguishing periodicity or chaotic state. The most prominent advantage in spring test for complexity measures exists in the solid key parameters of golden ratio threshold mapping in 3s plot and self-similarity threshold γ = 0.02 in CC rate. Figure 3 illustrates that the key self-similarity threshold γ scanning for Henon map in dynamic cloud maps. In Fig. 3 top sub-figure shows that higher complexity region with CC rate near 80% is more than in Fig. 3 middle and bottom sub-figures. Novel thinking is that if we select γ = 0.618 then can we reveal more complicated singular attractor performance? 3.2 Key Self-similarity Measure Threshold γ Locking in Max Complexity Reference [9] once reported that “the maximum Kaplan-Yorke dimension occurs for R = 3.4693 where the LE are λ = (0.30791, 0, −1.30791)” and DKY = 2.23542. Here we show while we utilize γ = (0.02, 0.618, 0.80), then what can we grasp for complexity enhancement in CC-R plots. During scanning threshold γ , we can capture more areas in CC-R plot maintaining the reported old max complexity except R = 0.43 (red dot in Fig. 4 upper sub-plot @ γ = 0.02), R = 0.57 (red dot in Fig. 4 middle sub-plot @ γ = 0.618), and R = 0.49 (red dot in Fig. 4 bottom sub-plot @ γ = 0.80). So far we can find the narrower R-value ranges into (0.43, 0.57) for the new higher complexity in Formula (6). The further evidence is (R, Information Entropy) = (0.49, 7.9975) versus (3.4693, 7.9969) while running bit full scrambling and pixel diffusion-based encryption algorithm with Lena image 256 × 256.

Signal Complexity Measure Based on Construction Creep …

15

Fig. 2 Henon map: Phase plots, pq plots, pqs plots, and 3s plots (from top to bottom). Left column for period state @ (a = 1.2, b = 0.4), right column for chaos state @ (a = 1.4, b = 0.3)

4 Conclusions One novel chaos feature had been gained by opening the third axis of pq-plot in 0-1 test for chaos and by inserting bi-activity functions of hyperbolic tangent and its hard approximation. Our spring test for complexity measure provided 3 s plot (good spring signs periodicity versus bad spring marks chaos) and CC rate (less than 7% means periodicity, between 8 and 84% points chaos).

16

J. Cai et al.

Fig. 3 Dynamic Cloud Map of Henon map: CC rate, from top to bottom with γ @ (0.02, 0.30, 0.618, 0.80)

The key self-similarity threshold in CC rate used 0.02 for preferred solid value which can mind out the new max complexity in Maximum KY dimension Lorenz equation. Based on the normalized test data, two-layer compression entropy and Euclidean distance self-similarity measure ensured the foundation of successful simulation research. This case study on signal complexity measure will accelerate the growth of chaos criteria tree.

Signal Complexity Measure Based on Construction Creep …

17

Fig. 4 Maximum KY dimension Lorenz in parameter R: CC rate, from top to bottom with γ @ (0.02, 0.30, 0.618, 0.80) Acknowledgements. This work is supported by Technological Innovation of Key Industries in Suzhou City Prospective Application Study [No. SYG201701], Graduate Research & Practice Innovation Program of Jiangsu Province [No. KYCX18_2509], and the Open Projects of Laboratory of Modern Acoustics of MOE [No. 2017-001]. Thank Mr. Peng Ming for running our new encryption algorithm.

References 1. Li, T.Y., Yorke, J.A.: Period three implies chaos. Am. Math. Mon. 82(10), 985–992 (1975) 2. Sprott, J.C.: A proposed standard for the publication of new chaotic systems. Int. J. Bifurcat. Chaos 21(9), 2391–2394 (2011) 3. May, R.: Simple mathematical models with very complicated dynamics. Nature 261, 459–467 (1976) 4. Tang, X.L., Lv, H.L., Yang, F.M., Yu, L.A.: Complexity testing techniques for time series data: a comprehensive literature review. Chaos Solitons Fractals 81, 117–135 (2015) 5. Bernardini, D., Litak, G.: An overview of 0-1 test for chaos. J. Braz. Soc. Mech. Sci. Eng. 38, 1433–1450 (2016) 6. Li, W.: Micro-nano-electronics Modeling Case Study (in Chinese), pp. 286–296. Soochow University Press (2019) 7. Wu, S., Li, Y., Li, W., Li, L.: Chaos criteria design based on modified sign functions with one or three-threshold. Chin. J. Electron. 28(2), 364–369 (2019) 8. Cai, J., Li, Y., Li, W., Li, L.: Two entropy-based criteria design for complexity measures. Chin. J. Electron. 28, 1113–1117 (2019) 9. Sprott, J.C.: Maximally complex simple attractors. Chaos 17, 033124-1-6 (2007)

Research on Partition Technology of Real-Time Database in Big Data Environment Yuxuan Li(B) and Suling Li Nanchang Institute of Technology, Nanchang, China [email protected]

1 The Introduction The application of big data technology has led internet data to an order of magnitude increase. As for ensuring big data, the background management and disposal of data is the foundation. The formal commercial use of cloud database especially increases the difficulty of this process. From this aspect, the load capacity of database for data is likely to become one of the bottlenecks in the application of big data technology. In view of this situation, it is a feasible way to use M/R architecture to solve data calculation in a batch solution. However, it is found from the practical application process that the transactional ACID feature in this architecture can not meet the present trend of data expansion. As a result, it makes a local impact on database server. Meanwhile, there still exists a certain force that cannot be incorporated into the database computing system. The above problems can effectively allocate the system under the flow computing framework, so that the database load can be more utilized.

2 Design of Tual-Time Partition Database System for Flow Computing From the perspective of the basic characteristics of the streaming computing framework, when the data is connected to the database cluster, effective cluster management is formed and secondary allocation is performed. According to the real-time partition system of different database, the cluster can form an effective horizontal expansion. Under different levels of extended architecture units, the database units with system attributes and weights can be implemented independently. Forming an effective data processing process inside the database unit, specifically, according to the optimal load of a single database partition several effective processing units are formed under the internal flow of data processing. The processing unit includes two necessary processes: load preprocessing and partition algorithm engine. The system and the partitioning strategy are formed inside different data processing units, and fixed time nodes are updated to complete effective data storage. From the overall design framework, there are five specific functional modules in this system as follows. First, the cluster management module. Cluster management is the core of the functional requirements of the system. The rational allocation of data is effectively achieved

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 C. H. WU et al. (eds.), Recent Developments in Intelligent Computing, Communication and Devices, Advances in Intelligent Systems and Computing 1185, https://doi.org/10.1007/978-981-15-5887-0_3

Research on Partition Technology of Real-Time Database …

19

through cluster management. Similar to the previous general design, there are two partition modes in the real-time database partition process; one is a partition algorithm based on the inside database group, and the other is based on horizontal expansion. The cluster management module is the key to the implementation of horizontal expansion. Firstly, we define different horizontal expansion units in the unified weight system during the specific design process. When the data stream is connected to the cluster management module, the module selects the corresponding data group randomly and sets the data to the data group for disposal. Meanwhile the database load is fed back and the weight value is raised. If the load of the database is relatively high, and the cluster management module randomly selects the data again in the same low weight horizontal extended database. Second, the data access module. This module is relatively simple and has no difference from the ordinary database management system. The architecture is usually used to obtain information data from console, etc., and pack it for transmission to subsequent cluster management module. Third, the data processing module. The module is divided into two parts, one is to dispose of the data. This module is implemented by general database processing module, and there has no difference between function and load. The other part is the load preprocessing module, which is responsible for calculating the reality of the current real load level of disposal unit, and communicate and feedback the database load with the cluster management module. The packet size/real-time characteristics of feedback information. Fourth, partition policy and dynamic management module. Partition strategy is the basis of dynamic management, which can form an effective partition algorithm under the overall framework of the data stream. In the framework of data flow, single classification of data format is used as the basis for different. As for the data processing process, it is carried out by means of database group paralled and data type queuing serial wave disposal. This way of working enables each piece of data to form an effective database correspondence under the component in the farthest dispose way so that the ability of the database to dispose of the data at the same time is effectively increased and achieve the basic situation of dynamic management. Fifth, horizontal expansion module. Horizontal expansion is an effective way to form real-time zoning. There are two kinds of logic in this process. One is to allocate the current database load in the value, and the other is to allocate directly and unconditionally under the current load conditions. The former helps to adjust overall effectiveness, and the latter is more conducive to the balance under different modules of the server. Under the first logical framework. Generally setting is that more than 80% of the database to form a database group. When the group load reaches the que value once again and then allocate subsequently.

3 Implementation of Real-Time Partition System Based on the analysis of the above five functional modules, specific modules are designed and implemented according to the principle of Storm system. The design step is to invoke the general architecture of WSPS under Storm in the overall implementation of

20

Y. Li and S. Li

the database. After completing the control of Flume and Kalfa framework after the data interface under Zook protocol, the purpose of MySQL data storage protocol was finally realized. The realization of different modules needs to establish their own module attributes. There is no essential difference between a data access module and a data processing module. So no more retelling. In addition, in the cluster management module implementation, Storm is used to effectively match Zookeeper to realize the recalculation of data packets and reasonably divide task nodes [1]. In a word, in this part of the design, the implementation of the whole data package processing. The large overall data flow is subcontracted to separate data grouping systems for efficient allocation in subsequent parallel extended database partitioning architectures. The main function of the liability preprocessing module is to perform real-time calculation and feedback on the load status of the database node to which it belongs. In the process of implementation [2], ReadBolt protocol is used to map the workload of work packets calculating, thus feedback real-time workload. The realization of this computing model has two advantages. One is that the load representation by load mapping can rarely occupy the load range of data processing, and then enables the database to implement more load resources to the data processing level. The positive test load map is positively correlated with the actual data processing capacity of the database to provide highfrequency refresh function for the module to ensure the real-time characteristics of the data. The implementation of specific aspects of the realization of TPCC table using a database query to achieve vertical mapping information extraction. The principle of weight randomness is realized through parallel extension module. That is, random selection between database groups under the same weight. While the weight involved follows the way of initialization and adjustment. The way of the adjustment is adjusted is positively correlated with the load value by the database partition [3]. The initialization system sets the weight of all database units to 1 in the specific implementation process. After obtaining the external packet, the random selection is made in the database unit with the weight of 1 (at this time all the database units). After that, the selected database processes all packets and feedback the real-time load value of the database node. The horizontal expansion module adjusts the weight value of the database to “1 + (real load 100)/100”. When the real-time load exceeds 90%, the second random selection is made in the database group whose weight is still “1”. In order to achieve the balance to database load, another parallel database node is introduced to allocate packet twice. According to this, it can be concluded that the real-time load is used to form the real-time partitioning function of the database. With the progress of data operation, the selected database will gradually reduce, and the server load will return to the initial state with the completion of data processing.

4 Audit and Effect of System Effectiveness Through the realization of the above design, the real-time database partitioning system based on streaming computing framework is theoretically completed. The system can effectively improve the load distribution of database group, improve the data processing capacity, and further play the load capacity of database group. In order to verify the

Research on Partition Technology of Real-Time Database …

21

effectiveness of the system [4], this paper adopts the traditional database architecture and the database partition system designed in this paper to process the same data and takes the total load ratio and data processing time as the evaluation basis to verify the effectiveness of the system and the actual effect. The experimental data unit is divided into two parts. A set is data with 10 tables and 5 sets of transaction-type data. The other set is a set of 100 table fields and 50 sets of transaction-type data. The latter set of data is more complex both quantitatively and structurally. In the experiment, the traditional real-time database partition system based on PC and the real-time database partition system based on the design of this paper are established. External data import is adopted, respectively to form a stable data flow. There is little difference in system configuration, data import hardware, and specific mode. The test comparison showed that when processing the first set of data, the average load of the traditional database was 35%, the peak load was 89%, and the total processing time was 14 s. The average load of the database system in this paper was 33%, the peak load was 82%, and the total processing time was 12 s. It can be seen that real-time partitioning of the database system through the streaming computing framework can effectively reduce the peak load of the database and improve the efficiency and objectivity of data processing. When processing the second set of data, the average load of the traditional database was 61%, the peak load was 98%, and the total processing time was 123 s. The average load of the database system in this paper was 58%, the peak load was 81%, and the total processing time was 77 s. Experimental data show that the real-time partitioned database system of the streaming computing framework is superior to the traditional database architecture in terms of load peak and processing time [5]. On the other hand, it can be found that the more complex the data type and the larger the amount of data, the more obvious the advantages of the database system in this paper are.

5 Summary The purpose of this paper is to design a real-time database partitioning system based on the streaming computing framework that can solve the data load and processing in the current big data environment. By comparison with experimental data shows that both to reduce the peak load of the database and the data processing time, the real-time database partition system based on streaming computing framework, compared with traditional database architecture, has obvious advantages, the future under the environment of big data, the data processing work can play a greater role.

References 1. Zhang, Y., Yu, J., Lu, L., et al.: Big data stream computing framework, Heron based classification task scheduling strategy. Comput. Appl. 39(4), 178–188 (2019) 2. Guo, M., Kang, H., Yuan, X.: Real-time database partitioning system based on streaming computing framework. Comput. Eng. 11, 14–21 (2017) 3. Poikane, S., Phillips, G., Birk, S., Free, G., Kelly, M., Willby, N.: Deriving nutrient criteria to support ‘good’ ecological status in European lakes: an empirically based approach to linking ecology and management. Sci. Total Environ. (2018)

22

Y. Li and S. Li

4. Lu, L., Yu, J., Chen, B., et al.: Task migration strategy of big data streaming computing framework storm. J. Comput. Res. Dev. 55(1), 71–92 (2018) 5. Yi, L., Hou, Y., Chen, C., et al.: Overview of task management technology for big data streaming computing. Comput. Eng. Sci. 39(2), 215–226 (2017) 6. This paper belongs to the scientific and technological research project of Jiangxi Education Department: application and Practice Research Based on database system. Topic number: GJJ171050, one of the results of the conclusion

RCS Sequence Period Estimation Method Based on Constant False Alarm Detection Xinbing Fang, Sifang Liu(B) , and Xiquan Wang Marine Department of Satellite Tracing and Metering, Jiangyin 214431, China [email protected]

1 Introduction In the Radar Cross Section (RCS) sequence of periodic targets [1], there is a unique periodic phenomenon as shown in Fig. 1, which is due to the periodic phenomenon caused by the target roll or the periodic motion of the target component. Although this type of RCS sequence has obvious periodic phenomena, it cannot be extracted by spectrum analysis tools. For this kind of data, this paper proposes a target period discrimination method based on Constant False Alarm Rate (CFAR) detection.

Fig. 1 Special RCS sequence when the target rolls over

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 C. H. WU et al. (eds.), Recent Developments in Intelligent Computing, Communication and Devices, Advances in Intelligent Systems and Computing 1185, https://doi.org/10.1007/978-981-15-5887-0_4

24

X. Fang et al.

2 Periodic Movement Target Characteristics (1) Rolling of fast flying targets or debris Targets such as engines or debris that lose power during the target launch are most likely to roll over. Judging the rollover phenomenon of the target, the rollover period of the extraction target, and the length of the warp direction can provide important information for target recognition. (2) Spin phenomenon of fast flying target When the target is in the middle inertia flight, it needs to adopt the spin motion for attitude control to ensure that it can reenter the small angle of attack and improve the hit accuracy. At the same time, due to the action of the heavy moment, the separation of the projectile, the release of the bait, and other external disturbances will produce a cone rotation and a swing on the basis of the spin, thereby forming micro motions such as precession and nutation, which will cause radar echo. Doppler modulation and power modulation provide important features for distinguishing between warhead targets and non-warhead targets [2]. (3) Spin phenomena of spin-stable satellites Spin stabilization is a way for the satellite to maintain its attitude. The satellite uses the gyro’s fixed axis around the spin cycle to orient the spin axis in the inertial space. In general, a spin-stable satellite can reach about 100 revolutions per minute. Only in this way can it be stable in inertial space and relatively stable on Earth observation. By extracting the spin period of the target and combining the orbital height of the target, it can be judged whether the target is a spin-stable state, thereby providing a basis for the target attribute discrimination. (4) Rolling phenomenon of failed satellites “Tumble” is a special form of motion of a space object that is characterized by a target rolling around a certain spatial axis and the roll plane remains unchanged. The failed satellite performs complex rotational motion under the action of gravity and perturbation of the Earth, and eventually tends to roll, such as the rolling of the Tiangong-1 before the crash and the rolling of the failed satellite. The rolling cycle is an important feature of the space rolling target. By discriminating the rolling cycle of the target, it can provide a basis for the attribute determination of the target. (5) Periodic movement of satellite components Based on the purpose of detecting targets in specific areas, some satellites have freely movable components. By identifying the presence of these components on the satellite, it can provide a basis for the satellite’s attributes and threat level.

RCS Sequence Period Estimation Method Based on Constant …

25

This paper mainly identifies the periodicity of (4) and (5) targets.

3 Periodic Discriminant Method for Constant False Alarm Detection 3.1 Periodic Discriminant Method for Constant False Alarm Detection The so-called periodic discriminating method for constant false alarm detection means that the occurrence of a maximum value is equivalent to detecting a target. The maximum value judgment block diagram is shown in Fig. 2. Use x n , yn , n = 1, 2, …, n, respectively, to represent the reference RCS sampling points on both sides of the RCS sampling point D (before and after the edge) to determine the maximum value, and the M nearest to the leftmost D The sampling point is similar to the protection unit in the CFAR detection, and the RCS value is also large in order to deal with the extreme values. The number of protection units can be an empirical value. The RCS maximum value judgment thresholds Thx and Thy are, respectively, calculated according to the RCS sampling points of the leading edge and the trailing edge, and the final threshold is taken as Th = max(Thx , Thy ). If D > Th, it is judged that D is a maximum value, otherwise D is considered not to be a maximum value. The front and back thresholds are calculated as follows:

Fig. 2 Block diagram of maximum value judgment

Thx = mx + Aσx Thy = my + Aσy

(1)

where mx and σ x are the mean and standard deviation of the leading edge sampling point x n , respectively, and my and σ y are the mean and standard deviation of the trailing edge sampling point yn , respectively, and A takes the empirical value. The sliding window maximum value is judged for the RCS sequence in a period of time, and the position and number of the maximum value are counted. If the interval between the maximum value positions is similar, and the number is three or more, the target can be considered as Do tumbling exercise.

26

X. Fang et al.

3.2 Effectiveness Analysis Radar parameter settings include: center frequency 10 GHz, pulse repetition frequency 120 Hz, bandwidth 1 GHz. The target of the moving warhead is a round cone with a length of 1 m and a diameter of 1 m. The warhead motion parameters are spin frequency 3 Hz, cone spin frequency 1 Hz, and precession angle 6° [3]. The non-warhead tumbling target selects a cylindrical three-stage rocket and irregular shape fragments. The threestage rocket is 2.35 m long and 1.3 m in diameter. The fragments are irregular shapes in the rectangular body. The two sides of the rectangular parallelepiped are 0.7 m and the other side is 0.06 m long. The rollover period for both types of non-warhead targets is set to 2.5 s. The attitude angle curve and the RCS sequence of the moving warhead target within 10 s are shown in Fig. 3. According to the method introduced in the previous section, the position and number of maximum values of the RCS sequence are judged. The number of protection units M is 3, the number of reference units is 20, and the threshold calculation parameter A is 3. The final maximum number is 0. It can be judged that the target does not roll.

Fig. 3 Target angle variation curve and RCS sequence of the cone head

The RCS sequences of the two types of targets are shown in Fig. 4, respectively. According to the above parameters, the roll characteristics are judged. Non-vertebral targets have 8 maxima, corresponding to the 8 maxima of the sequence. The axisymmetric nature of the target causes the number of maxima to be twice the rollover period, which does not prevent the rollover judgment of the target. The fragmentation judgment result has 16 maxima, because each of the maximum values of the RCS sequence has two amplitudes corresponding to each other and is taken as a protection unit when calculating the threshold, so that it can be judged as a maximum value [4].

RCS Sequence Period Estimation Method Based on Constant …

(a) RCS sequence of a cylinder three-stage rocket rolling

27

(b) RCS sequence when irregular fragments are rolled

Fig. 4 RCS sequences of the two types of targets

4 Experimental Verification The CFRS method was used to process the measured RCS sequences of Fengyun 3B and Fengyun 3C satellites at a data rate of 20 Hz. The sliding window method was used for periodic discrimination [5]. The results are as follows in Figs. 5 and 6. 30 25 1.8195

20 1.819

10

(s)

RCS(dB)

15

5

1.8185

0 1.818

-5 -10

1.8175

-15 -20

1.817

0

20

40

60

80

100

120

Time(s)

140

160

180

200

50

100

150

200

(s)

Fig. 5 RCS sequence and estimation of Fengyun 3B satellite

It can be seen that the method can accurately discriminate the periodicity of such special RCS sequences.

28

X. Fang et al. 1.86

25 1.84

20 15

(s)

RCS(dB)

1.82

10 5

1.8

0 1.78

-5 -10

1.76

-15 180

200

220

240

260

Time(s)

280

300

320

200

220

240

260

280

300

320

340

(s)

Fig. 6 RCS sequence and estimation of Fengyun 3C satellite

Acknowledgements. An automatic periodic discriminant method based on target periodic motion is studied for a special periodic RCS sequence. Electromagnetic simulation was carried out on a class of round head cone precession warhead targets and cylindrical three-stage rockets and irregular debris two types of non-warhead rolling targets. The RCS sequence was used to determine the rollover characteristics by the method of maximal position and number judgment. Simulation results verify the effectiveness of the proposed method. In addition, the effectiveness of the algorithm is verified by the measured RCS sequences of Fengyun 3B and Fengyun 3C satellites.

References 1. Skolnik, M.I.: Introduction to Radar Systems. Publishing House of Electronics Industry, Beijing (2006) 2. Chen, B.: Radar Target Reflection Characteristics. National Defence Industry Press, Beijing (1993) 3. Wang, S., et al.: Calculation method of radar cross section for micro-motion complex target. Modern Radar (2008) 4. Xing, W.: Nonparametric Statistics. Tsinghua University Press, Beijing (2014) 5. Blandford, D., Parr, J.: Digital Signal Processing and MATLAB Simulation. Mechanical Industry Press, Beijing (2015)

Target Recognition of RCS Sequence Based on Neural Network Gaofeng Pan, Sheng Liang, Sifang Liu(B) , and Xie Yong Marine Department of Satellite Tracing and Metering, Jiangyin 214431, China [email protected]

1 Introduction When the space object moves along the orbit, its attitude changes continuously with respect to the radar, so that its radar cross-sectional (RCS) data can be obtained, and the variation law reflects the physical characteristics of the target structure [1]. Through the estimation of the spatial target RCS statistical parameters, the stable state of the target can be roughly inferred, the typical spatial target can be built and matched, and the target type can be identified. Classification identification of spatial targets refers to the use of an RCS sequence of unknown targets to determine the type of target. The measured data analysis shows that the RCS variation range is large in the whole observation arc of the radar to the space target, and the RCS sequence of the same target has a large difference in the form of the front, middle, and back sections [2]. Aiming at this problem, this paper proposes a scheme for constructing a template according to the attitude angle of the radar relative to the target. The angle between the radar line of sight and the target radial axis and the tangential axis is used to roughly describe the attitude angle of the radar relative to the target, thus constructing the target. Templates, using neural networks for training, improve target recognition.

2 Build the Attitude Angle Template Since the target is a 3-dimensional structure, it needs two angles of azimuth and pitch to describe it strictly. However, we cannot determine the specific attitude of the satellite in space, because it is impossible to obtain accurate radar line of sight azimuth and pitch information under the target coordinates. In this paper, the Angle between radar line of sight and target radial axis and tangential axis, respectively, is proposed to roughly describe the attitude Angle of radar relative to target. Radial attitude theta is defined as the Angle between the radar-satellite line and the satellite and geocentric line, as shown in Fig. 1. The radial attitude Angle corresponding to the RCS sequence in the figure is shown in Fig. 2. The attitude Angle varies from 62° to 23°, and then to 55°. The template construction combined with attitude Angle and target recognition can reduce the search range of template library and improve the accuracy of target recognition. The definition of tangential attitude Angle is the included Angle of the connection between radar and satellite and the flight direction of satellite as shown in Fig. 3. © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 C. H. WU et al. (eds.), Recent Developments in Intelligent Computing, Communication and Devices, Advances in Intelligent Systems and Computing 1185, https://doi.org/10.1007/978-981-15-5887-0_5

30

G. Pan et al.

Fig. 1 Schematic diagram of radial attitude Angle of space target 65 60 55

Theta

50 45 40 35 30 25 20

0

100

200

300

400

500

600

Time(s)

Fig. 2 Changes of radial attitude Angle of space target

The tangential attitude Angle corresponding to the RCS sequence in the figure is shown in Fig. 4. The attitude Angle varies from 30° to 140°, so the sum of the two attitude angles is obviously not equal to 90°. The template construction combined with attitude Angle and target recognition can reduce the search range of template library and improve the accuracy of target recognition.

3 Feature Extraction and Classifier Design Given a segment of RCS sequence and its attitude Angle, the classification scheme adopted in this paper is feature extraction and classifier design [3]. In the feature extraction process, the main 16 features are adopted:

Target Recognition of RCS Sequence Based on Neural Network

31

Fig. 3 Schematic diagram of tangential attitude Angle of space target 150

Phi

100

50

0

0

100

200

300

400

500

600

Time(s)

Fig. 4 Tangential attitude Angle changes of spatial target

(1)

(2)

(3) (4)

Mean feature difference of RCS sequences: the absolute value of the difference between the mean of target RCS sequences and the mean of template library target RCS sequences; Standard deviation characteristics of RCS sequences: absolute value of the difference between standard deviation of target RCS sequences and standard deviation of template library target RCS sequences; KL divergence of RCS sequences: KL divergence of target RCS sequences and target RCS sequences of template library was tested; Spectral standard deviation characteristics of RCS sequences: the absolute value of the difference between the spectral standard deviation of the target RCS sequence and the spectral standard deviation of the template library target RCS sequence;

32

(5)

(6)

(7)

(8)

(9)

(10)

(11)

(12)

(13)

(14)

(15)

(16)

G. Pan et al.

Spectral energy ratio characteristics of RCS sequences: the absolute value of the difference between the spectral energy ratio of target RCS sequences and the spectral energy ratio of target RCS sequences in template library; Spectral distance of the RCS sequence: the spectrum of the RCS sequence is defined as the fast Fourier transform of the RCS sequence, and the spectral distance is defined as the root mean square of the spectrum difference of the two sequences; Mean characteristics of first-order difference sequences: the absolute value of the difference between the mean of first-order difference of the target and the mean of first-order difference of the template library is tested; Characteristics of standard deviation of first-order difference sequences: the absolute value of the difference between standard deviation of the target first-order difference sequence and standard deviation of the template library target first-order difference sequence; KL divergence of the first-order difference sequence: test the KL divergence of the target first-order difference sequence and the template library target first-order difference sequence; Spectral standard deviation characteristics of the first-order difference sequence: the absolute value of the difference between the spectral standard deviation of the first-order difference sequence of the target and the spectral standard deviation of the template library target; Spectral energy ratio characteristics of first-order difference sequences: the absolute value of the difference between the spectral energy ratio of the target firstorder difference sequences and the spectral energy ratio of the target first-order difference sequences in the template library is tested; Spectral distance of the first-order difference sequence: the spectrum of the firstorder difference sequence is defined as the fast Fourier transform of the first-order difference sequence, and the spectral distance is defined as the root mean square of the spectrum difference of the two sequences; Mean characteristics of second-order difference sequences: the absolute value of the difference between the mean of the target second-order difference and the target second-order difference of the template library is tested; Characteristics of standard deviation of second-order difference sequence: the absolute value of the difference between standard deviation of the target secondorder difference sequence and standard deviation of the template library target second-order difference sequence is tested; KL divergence of the second-order difference sequence: test the KL divergence of the target second-order difference sequence and the template library target second-order difference sequence; Spectral standard deviation characteristics of second-order difference sequence: the absolute value of the difference between the spectral standard deviation of the target second-order difference sequence and the spectral standard deviation of the template library target second-order difference sequence is tested.

The above 16 features constitute an 16-dimensional feature vector, which is trained and classified by using neural network classifier.

Target Recognition of RCS Sequence Based on Neural Network

33

4 Neural Network Design Neural network consists of input layer, output layer, and hidden layer. During the learning process, the input vector is transmitted to the input layer and transmitted to the output layer through the hidden layer [4]. By learning to adjust the weight value of each neuron node, the error between the actual output and the target vector is minimized. In this paper, a 3-layer feedforward network is constructed. The number of input layer nodes is equal to the number of features, which is 18. The number of nodes in the output layer is determined by the number of categories to be judged. In this work, the identification of warheads is judged, so the number of nodes in the output layer is 1. If the number of hidden layer nodes is too small, the relationship between features may not be accurately extracted, leading to poor fault tolerance. If there are too many hidden nodes, the network will be complex, the training time will increase, and even the feature space may be divided too finely, and the network lacks generalization ability (Fig. 5).

Fig. 5 Schematic diagram of network structure

The sample generation method is as follows: currently there are 4 RCS sequences, which belong to two targets FY03 and X37B. For each section of the RCS sequence, using the sliding window method to produce a sample, sliding window length 1200 points (60 s data), sliding window interval (0.5 s) 10, for each piece of data, a sliding window to find the corresponding attitude Angle from 4 period of RCS sequence data segment (may not exist), and calculate the two paragraphs above 18 characteristics of RCS sequence, 18 to form a feature vector, to get a sample. The sample labels are numbered 1 (same type) or 0 (different type) according to the types of the two samples. The sample situation is as in Table 1. During the test, samples of 3 RCS sequences were used for training, and the remaining 1 RCS sequence was tested. The results are shown in Table 2. The average accuracy reached 97.80%. The training process of the training set is shown in Fig. 6.

G. Pan et al. Table 1 Sample conditions of plus and minus RCSS NSL1 NSL0 Sum SA

1516

1536

3052

SB

1339

1173

2512

SC

764

654

1418

SD

530

253

783

Note NSL1 number of samples labeled 1; NSL0 number of samples labeled 0

Table 2 Classification accuracy of cross-verified RCS sequences RCSS NSL1 CDQ ACC SA

1516

1479

1339 100

97.56 1536

SB

1339

SC

764

731

SD

530

530 100

Stat

4149

4079

NSL0 CDQ ACC TACC 1438 93.62 95.58

1173

1172 99.91 99.96

654

653 99.85 97.6

253

252 99.6

95.68

96.34 3616

99.87

3515 97.26 97.8

Note RCSS RCS Sequence; CDQ correctly discriminate quantity; ACC accuracy; TACC total accuracy; SA-SD sequence A-sequence D Best Validation Performance is 9.3408e-06 at epoch 11 Train Validation Test Best

10 0

Mean Squared Error (mse)

34

10 -2

10 -4

10 -6

0

2

4

6

8

10

12

14

16

17 Epochs

Fig. 6 Convergence of mean square heel error of training set

Target Recognition of RCS Sequence Based on Neural Network

35

Acknowledgements. This paper gives a method to build the target template, extract the 18dimensional characteristics of the RCS sequence, construction of neural network in target recognition based on RCS sequence, the recognition accuracy of the test of 96.77%, the accuracy is a sliding window, through the continuous sliding window weighted average [5] for many times, can further improve the recognition of the time. Based on the current feature extraction and recognition framework, more data can be collected for verification in the future, and the features can be increased or decreased according to the statistical results, so as to achieve the ideal recognition results for the measured data.

References 1. Chen, B.: Analysis and Design of Modern Radar System. Xi’an University of Electronic Science and Technology Press, Xi’an (2016) 2. Deng, J.: Research on Measurement of Target Reflection Characteristics of Millimeter Wave Short Range Radar. Nanjing University of Science and Technology, Nanjing (2007) 3. Chen, L., et al.: Modeling and Detection of Typical Complex Micro-motion Targets. Telecommun. Technol. 56(11) (2016) 4. Shen, T.: Research and Discussion of Artificial Neural Network Based on Nonparametric Statistical Method. East China Normal University, Shanghai (2010)

Encouraging Cooperation in Participatory Sensing with Reputation and Game Analysis Tianqi Zhang, Rong Zhang, and Jin Wang(B) School of Information Science and Technology, Nantong University, Nantong, China [email protected]

1 Introduction Participatory sensing refers to an interactive and participatory sensing network formed by mobile devices. It is able to capture, analyze, and share local sensing data. Participatory sensing is the extension and expansion of wireless sensing network. At present, the participatory sensing has been used in certain applications, such as sharing of personal diet [1], perception and real-time display of urban noise pollution [2], collection and sharing of bicycling experience [3], etc. In order to protect privacy, users can exchange data with each other before uploading data [4]. However, how to encourage cooperation in this process is still a problem. Trust management [5] is introduced into participatory sensing for building trust or reputation. Most researches on participatory sensing assume that the task center is trustworthy. The Incognisense [6] pointed out that the task center may be untrustworthy. Enlightened by the hybrid network [7], this paper adopts the scenario that participants collaboratively exchange their sensing data when they encounter with each other. The collaborative data exchange scheme requires the cooperation among the participants. However, the malicious participants may not cooperate with others, such as abandoning the exchanged data. This may result in the failure of participatory sensing applications. To cope with the uncooperative behaviors between the participants, this paper proposes a reputation management mechanism to stimulate the cooperation among the participants. On the other hand, this paper proposes the “tit for tat” game strategy to punish those uncooperative behaviors. We use the game theory to analyze the motivation of malevolent participants for their uncooperative behaviors. This mechanism can prompt the rational participants to choose the “cooperative” strategy.

2 System Model The model of the participatory sensing system application mainly comprises three characters, namely, (1) task provider; (2) task center; and (3) participant. The collaborative exchange strategies assume that the participants are trustworthy and will honestly and cooperatively complete the data exchange agreement. Actually, this assumption is a strong constraint. In fact, some participants may be malicious or selfish. The solution proposed in this paper is to establish the trust relations among the participants through

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 C. H. WU et al. (eds.), Recent Developments in Intelligent Computing, Communication and Devices, Advances in Intelligent Systems and Computing 1185, https://doi.org/10.1007/978-981-15-5887-0_6

Encouraging Cooperation in Participatory Sensing …

37

the trust management mechanism to assist the participants in the collaborative data exchange. The participants will firstly judge whether the opposite party is trustworthy or not when they conduct data exchange, and then they will decide whether to conduct the data exchange according to their own decision strategy.

3 Trust Model and Game Analysis 3.1 Trust Model Assume that the assessment information uploaded by the participants is true. Main steps of trust management are described as follows. (1) Participants U1 and U2 register pseudonyms in the pseudonym server to generate a pseudonym verification certificate. (2) Reputation Initialization and Data Exchange. When participants U1 and U2 are encountered, U1 and U2 will send the request of reputation initialization to the reputation server with the pseudonyms. The reputation servers will firstly verify the pseudonym certificate and then make response to the request for reputation initialization. After getting the initial reputation, the participants will judge whether to conduct data exchange according to the reputation value. If they intend to exchange the data, the exchanged data will be uploaded to the task provider. U1 and U2 will, respectively, upload their evaluations into the reputation server which can update the reputations of U1 and U2 according to the following trust model. (3) Calculation and Update of Reputation The reputation is calculated mainly base on beta trust model [8]. Firstly, the parameters of Beta density function are selected. It is assumed that there are n participants, the number of cooperation and non-cooperation times of participant i at the data communication stage of the kth round are, respectively, ri and si . For any i ∈ {1, 2, . . . , n}, the coefficient of reputation value corresponding to the participant is αik = ri + 1

(1)

βik = si + 1

(2)

The reputation value of any participant i at the kth round is calculated according to two parameters of beta density function. Definition 1 The increment of the reputation value of any participant at the kth round is defined as         αik  βik xαi k (1 − x)βi k Trust(i, k) =  αik + βik (3) wherein x is a probability variable which satisfies E(x) = αik

 k αi + βik .

38

T. Zhang et al.

It is assumed that the reputation value of participant i after the k − 1th round is expressed by Trust(i, k − 1). Definition 2 after k rounds, the reputation value of participants is defined as Trust(i, k) = ω1∗ × Trust(i, k − 1) + ω2∗ × Trust(i, k)

(4)

wherein: ω1∗ and ω2∗ are the balance factors of the reputation value, and ω1∗ + ω2∗ = 1. It is assumed that the initial Trust(i, 0) = 0.5. After acquiring the latest reputation value, the reputation server will update, save, and broadcast this value to the participants. Participants will evaluate whether to conduct data exchange according to the evaluation on reciprocal reputation value. 3.2 Game Model The participants work in the participatory sensing system can be regarded as a game. It is assumed that participants of the game are rational, and then the game players are participants in the participatory sensing system in this game. Behaviors of the participants in the communication process can be divided into two behavioral strategies, namely, cooperation behavior and non-cooperation behavior. These two behaviors are also the strategic space of the participants in the game, written as S = {cooperation, non-cooperation}. In the participatory sensing system, it is assumed that most participants in the network are honest and trustworthy. The reputation value of participants will be dynamically updated according to their strategies. The game model is composed of three basic elements, which are, respectively, participant, strategy space, and payoff function. The game model is built in the following for the scene of untrustworthy participant. (1) One-stage Game Analysis Assumptions: (a) The participants in the participatory sensing system are all deemed as rationalist. (b) Strategies of the participants can be divided into two kinds, cooperation and non-cooperation, which are written as S = {cooperation, non-cooperation}. When a participant fails to receive the rewards from the task provider, the other participant involved in the exchange shall be deemed as non-cooperation act. If the participant receives the rewards, the other participant involved in the exchange shall be deemed as cooperation act. Participants of the one-stage game modeling are the participants in the participatory sensing system. The strategic space is S = {cooperation, non-cooperation}. The payoff parameters of payoff function are defined as follows: R refers to the payoff of participants after they successfully upload the data packages. CR refers to the payoff of the betraying party, and CR > R. The payoff matrix of the one-stage game is as shown in Table 1. It can be known from the definition of stage game and Table 1 that in the stage game, on the premise that participant U1 and U2 are both rationalists, the strategy {cooperation, non-cooperation} is the Nash equilibrium of this game. It can be known through the analysis on the payoff that this strategy pair is not the optimal in the network.

Encouraging Cooperation in Participatory Sensing …

39

Table 1 Payoff matrix U1

U2 Cooperation Non-cooperation

Cooperation

(R, R)

Non-cooperation (CR , 0)

(0, CR ) (0, 0)

(2) Payoff Function of One-stage Game The prospective payoff if the participant chooses the cooperation strategy is U (i) = 1/2 · R + 1/2 · 0 = 1/2 · R

(5)

The payoff of the participants who choose the non-cooperation strategy is U (j) = 1/2 · CR + 1/2 · ×0 = 1/2 · CR

(6)

Since CR > R, the payoff under the non-cooperation strategy is higher than that under the cooperation strategy. In the long run, the rational participants in the network may gradually become malicious participants to pursue higher payoff. As a result, the network will be in an unbalanced status, which will be further paralyzed and unable to provide normal services. Therefore, a reasonable mechanism needs to be designed to deal with the non-cooperative behaviors. 3.3 Repeated Game Analysis (1) Repeated Game Description The tit for tat strategy is adopted in this paper. That is, if one participant adopts the non-cooperation strategy at the time slot t, all the neighbor participants will adopt the non-cooperation strategy against this participant as a punishment in the following l time slots. To solve the Nash equilibrium problem in the repeated game, Folk Theorem [9] is introduced. This theorem proves that in a repeated game, the game players can probably reach cooperation equilibrium under the conditions of satisfying the personally rational constraints of the game players as long as they have sufficient patience (shown as the discount factor σ is large enough). In the repeated game, the payoff of the game player in a single stage at the moment t  t t is μi Si , Sj , then the prospective payoff of the participants in the entire repeated game can be expressed as Ui =

L

  σ t μi Sit , Sjt

(7)

t

In the game theory, the infinite game is used to simulate a limited game, wherein the participants are unaware of the game process. Obviously, this often occurs in strategic

40

T. Zhang et al.

interactions, especially in network operation. To simulate the inscrutable result of the game, the payoff value in the future periods is reduced. This technique is referred to as discount. The parameter σ is the discount factor, and 0 ≤ σ ≤ 1. Participants will pay more attention to the long-term interests if they are active for a long time. Correspondingly, the σ will be larger; otherwise, σ will be smaller if the participants focus on immediate interests. The parameter L refers to the number of time slots which the participants intend to be active in the network. (2) Repeated Game Payoff If the participant chooses the cooperation strategy, the payoff can be expressed by Ui =

L

σ t μi (C, C) =

L

t=0

σ tR

(8)

t=0

If the participant adopts the cooperation strategy at the moment of 0 → t1 − 1, adopts the non-cooperation strategy at the moment of t1 , and is in the punishment period at the moment of t1 + 1 → t2 , the participant will adopt the cooperation strategy at the moment of t2 + 1 → L after the punishment due to non-cooperation behaviors. (It is assumed that this participant adopts the non-cooperation strategy only at the moment of t1 ). The payoff of the participant is defined as Uj , which is specifically as follows Uj =

t

1 −1

σ t R + σ t1 CR +

t2

0+

t=t1 +1

t=0

= σ t1 CR +

t

1 −1

σ tR +

L

L

σ tR

t=t2 +1

σ tR

(9)

t=t2 +1

t=0

The punishment period l = t2 − t1 , and the non-cooperation payoff CR = nR. The payoff under the cooperation strategy and non-cooperation strategy are compared as Ui − Uj =

L

σ R − σ CR − t

t1

t=0

=

t2

t

1 −1

σ R− t

L

σ tR

t=t2 +1

t=0

σ t R − σ t1 CR

(10)

t=t1 t2

  σ t R − σ t1 CR = R σ t1 + σ t1 +1 + · · · σ t2 − σ t1 CR

t=t1



1 − σ l+1 −n =σ R 1−σ t1

(11)

Encouraging Cooperation in Participatory Sensing …

41

(3) Analysis of Cooperation Conditions To make Ui − Uj > 0, it only needs to ensure (1 − σ l+1 )/(1 − σ ) − n > 0. Since the discount factor σ is close to 1, closer to 1 of σ will indicate that the participants pay more attention to the long-term interests. In this paper, σ = 0.9, which indicates that the participants all focus on the long-term interests. The punishment period l ≥ 2, which shall not be too short, otherwise, it cannot awe the malicious participants. If the punishment to the malicious participants is not strict enough, the adopted punishment strategy will be out of action, and the malicious participants will continue to adopt the non-cooperation strategy. If it goes on like this, the malicious participants in the network will be gradually increased, the network may be paralyzed, and therefore this paper assumes the punishment period l = 4. The payoff under the non-cooperation strategy is n times of that under the cooperation strategy, n = 2 in this paper. Then,   1 − σ l+1 /(1 − σ ) − n = 1 − 0.95 /(1 − 0.9) − 2 = 2.0951 > 0. It is known from the above analysis that Ui > Uj , namely, the payoff of the participants adopting the cooperation strategy is higher than adopting the non-cooperation strategy after the repeated game. After the repeated game, if one participant adopts the non-cooperation strategy, the other participants will execute punishment measures against this participant. This participant will enter into a punishment period, during which the malicious participants cannot gain any payoff through the tasks or cooperation with other participants. After the punishment period ends, the malicious participants will consider their long-term interests and adopt the cooperation strategy. For the participants, the cooperation strategy is their best choice, and therefore both parties of the game will adopt the (cooperation, cooperation) strategy at last. Therefore, the “tit for tat” strategy can force the malicious participants to adopt the cooperation strategy in the tasks in the future.

4 Conclusion This paper proposes a trust management schema and “tit for tat” strategy to resist the noncooperation behaviors of the participants. According to the analysis, the (cooperation, cooperation) strategy is the most beneficial choice for the participants in the long run, and therefore it can effectively resist the non-cooperation participants.

References 1. Reddy, S., Parker, A., Hyman, J., Burke, J., Estrin, D., Hansen, M.: Image browsing, processing, and clustering for participatory sensing: lessons from a DietSense prototype. In: Proceedings of the 4th Workshop on Embedded Networked Sensors, pp. 13–17 (2007) 2. Rana, R.K., Chou, C.T., Kanhere, S.S., Bulusu, N., Hu, W.: Ear-phone: an end-to-end participatory urban noise mapping system. In: Proceedings of the 9th ACM/IEEE International Conference on Information Processing in Sensor Networks, pp. 105–116 (2010) 3. Eisenman, S.B., Miluzzo, E., Lane, N.D., Peterson, R.A., Ahn, G.-S., Campbell, A.T.: BikeNet: a mobile sensing system for cyclist experience mapping. ACM Trans. Sensor Netw. (TOSN) 6(1), 6 (2009)

42

T. Zhang et al.

4. Christin, D., Guillemet, J., Reinhardt, A., Hollick, M., Kanhere, S.S.: Privacy-preserving collaborative path hiding for participatory sensing applications. In: 2011 IEEE Eighth International Conference on Mobile Ad-Hoc and Sensor Systems, pp. 341–350 (2011) 5. Guo, J., Chen, I.-R., Wang, D.-C., Tsai, J.J.P., Al-Hamadi, H.: Trust-based IoT cloud participatory sensing of air quality. Wirel. Pers. Commun. 105(4), 1461–1474 (2019) 6. Christin, D., Roßkopf, C., Hollick, M., Martucci, L.A., Kanhere, S.S.: IncogniSense: an anonymity-preserving reputation framework for participatory sensing applications. Pervasive Mob. Comput. 9(3), 353–371 (2013) 7. Chaum, D.L.: Untraceable electronic mail, return addresses, and digital pseudonyms. Commun. ACM 4(2), 84–90 (2003) 8. Josang, A., Ismail, R.: The beta reputation system. In: Proceedings of the 15th Bled Electronic Commerce Conference, vol. 5, pp. 2502–2511 (2002) 9. Kandori, M.: Introduction to repeated games with private monitoring. J. Econ. Theory 102(1), 1–15 (2002)

Multi-granularity Attribute Sentiment Analysis Based on Neural Network Runxue Chen, Peng Yan(B) , and Lian Xu School of Computer, Sichuan University of Science & Engineering, Zigong 643000, China [email protected]

1 Introduction With the advancement of technology, people communicate their opinions and opinions through the Internet. For example, comment on the goods purchased by the e-commerce platform, and comment on topics of interest on micro-blog. These remarks can reflect the public’s attitudes and feelings towards products and hot events. It is difficult to achieve massively processed text data by hand. Excavating emotional information from large amounts of text through machine learning can create enormous commercial and social value. Sentiment analysis is to analyze the text in order to mine the emotional tendency of the text. Attribute sentiment analysis is an important aspect of emotion classification. It mainly analyzes the emotional tendency of an attribute in the context, such as positive and negative emotions. In the sentence “This wine tastes good, but the cup doesn’t look good”, The “wine” attribute is a positive emotion, while the “cup” is a negative emotion. By extracting feedback from users, the feedback information helps companies identify specific defects in the product and prompts the company to improve the product. Traditional emotional classification algorithms have great challenges when faced with multiple target attributes. When a sentence contains multiple objectives, it is critical to accurately model and extract such structural information.

2 Related Work Attribute sentiment classification [1] needs to consider both the context of the sentence and the information of the target attribute. The traditional machine learning [2] method is to use sentiment analysis as a text classification task and design an effective feature extraction method to train classification. Manek et al. [3] designed a support vector machine based on Geordie coefficient to achieve good results in sentiment analysis. Vo et al. [4] Design specific emotional words embedding and emotional words to extract features to predict emotional tendency. These methods require a large number of features to be collected manually and are easy to achieve performance bottlenecks. In recent years, original features can be encoded as continuous and flexible sized vectors without the need for complex feature engineering. Tang et al. [5] modeled the left and right by designing two neural networks, and finally stacked the left and right outputs together as the final predicted output. This method does not consider the attributes of © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 C. H. WU et al. (eds.), Recent Developments in Intelligent Computing, Communication and Devices, Advances in Intelligent Systems and Computing 1185, https://doi.org/10.1007/978-981-15-5887-0_7

44

R. Chen et al.

the target entity. When there are multiple target entities in the sentence, it is difficult to classify them accurately. Cheng et al. [6] enhanced the role of each target attribute in classification by adding attention to the model. The attention mechanism generally uses the mean of the vector to learn the attention weight of the context. Based on the attention mechanism, Ma et al. [7] designed a model in which text and target attributes can pay attention to each other. This mechanism learns the relationship between extra attention weights and context words. These methods do not consider multi-target attributes, which will lead to the loss of some target attributes and context-related information. Aiming at the above problems, a multi-granularity attribute sentiment analysis model based on neural network is proposed. Using bidirectional attention to extract the correlation between context and target attributes, a strong attention mechanism is proposed to capture important information inside the target. Combine bidirectional attention and strong attention to better emotional classification of goals.

3 Multi-granularity Sentiment Analysis Model The model proposed in this paper combines bidirectional attention and strong attention to learn the emotional characteristics of different attributes more fully. The whole frame is shown in Fig. 1. The model mainly consists of four parts: embedded layer, hidden layer, multi-granularity attention layer, and output layer.

Fig. 1 Multi-granularity sentiment analysis model

Multi-granularity Attribute Sentiment Analysis …

45

3.1 Task Definition A sentence S = {s1 , s2 , . . . , sn } consisting of n words and a list of attribute aspects A = {a1 , a2 , . . . , ak } and a sentence subsequence ai = {si1 , si2 , . . . , sim }, m ∈ [1, n] corresponding to each aspect are given. The multi-attribute sentiment analysis is an emotional tendency analysis of multiple target attributes in a text. 3.2 Model Embedded Layer The input layer maps words in the dataset to low-dimensional, continuous, and realvalued word vector spaces. All word vectors are stacked into an embedded matrix Lw ∈ Rd ∗|V | , where d is the dimension of the word vector and |V | is the vocabulary. Use the pre-trained word vector “Glove” [8] to get a fixed embedding of each word. Hidden layer Vectorize the words and enter them into the hidden layer. The context and target associations in the hidden layer are extracted by bidirectional long short-term memory (BiLSTM) [9]. Standard RNNs experiences problems with gradients disappearing or exploding, where gradients may exponentially increase or decay over long sequences. So using BiLSTM unit as transfer function can better simulate the long-distance semantic correlation in the sequence. Compared with the standard RNN, the BiLSTM unit contains three additional gates: input gate, forgetting gate, and output gate. The update process of BiLSTM is as follows:    i · [ht−1 ; xt ] + bi it = σ W

(1)

     f · ht−1 ; xt + bf ft = σ W

(2)

     o · ht−1 ; xt + bo ot = σ W

(3)

     r · ht−1 ; xt + br gt = tanh W

(4)

ct = it  gt + ft  ct−1

(5)

ht = ot  tanh(ct )

(6)

where  is the multiplication of elements, σ is sigmod function, it , ft and ot are input f , W  o, W  r is the weight matrix,  i, W gate, forgetting gate and output gate, respectively. W bi , bf , bo , br is the deviation term. Similar transformation process for reverse sequence model. An embedded word for a given context statement s and corresponding aspect a, BiLSTM can be used to get the output of context output C = [h1c , h2c , . . . , hnc ] ∈ R2d ∗N 2d ∗M , respectively. and target attribute square A = [h1a , h2a , . . . , hm a]∈R

46

R. Chen et al.

Multi-granularity Attention Layer Attention mechanism [10] plays an important role in the interaction between mining context and target attributes. The coarse-grained attention mechanism has a wide range of applications when extracting the attention weights of context and target attributes. A simple averaging will result in some loss of information, especially for multiple target sentences. In this paper, a multi-grained attention mechanism is proposed, which combines bidirectional attention and strong attention to extract the context and target association. The implementation steps are as follows: (1) Feature extraction in context and target For input sequence S = {s1 , s2 , . . . , sn }, BiLSTM is used to extract target context C ∈ R2d ∗N and target aspect A ∈ R2d ∗M , as follows: −−−−−→ C = BiLSTM(s1 , s2 , . . . , sn )

(7)

←−−−−− A = BiLSTM(s1 , s2 , . . . , sn )

(8)

(2) Adding Bidirectional Attention First, the obtained feature representation is averaged to obtain Cavg and Aavg , as follows: 1 i hc n

(9)

1 i = ha n

(10)

n

Cavg =

i=1 n

Aavg

i=1

Second, get the attention weight of the context αica and the attention weight of the target βiac , as follows:   exp f (hic , Aavg ) ca   αi = (11) j n j=1 exp f (hc , Aavg )   exp f (hia , Cavg ) ac   βi = (12) j m j=1 exp f (ha , Cavg ) where the f function is a function that calculates the importance of hic and hia , defined as follows:     (13) f hic , Aavg = tanh hic · Wa · ATavg + ba where Wa and ba are weighted matrices and biased matrices, respectively, f (hia , Cavg ) is calculated by a similar method.

Multi-granularity Attribute Sentiment Analysis …

47

Finally, the context Cr and target aspect Ar are obtained from the attention weight matrix, as follows: Cr = Ar =

n  i=1 m 

αica hic

(14)

βiac hia

(15)

i=1

(3) Adding strong attention Firstly, the context and target information extracted by BiLSTM are multiplied by matrix to get the interaction matrix I. I = C · AT

(16) strong

Second, add attention to the rows and columns of the interaction matrix to get αi strong and βj , respectively. The calculation formula is as follows: strong

αi

strong

βj

exp(Iij ) = i exp(Iij )

(17)

exp(Iij ) = j exp(Iij )

(18)

Finally, the context and target representations Cstrong and Astrong are obtained according to the obtained strong attention weight. As shown below Cstrong = Astrong =

n  i=1 n 

strong i hc

(19)

strong i ha

(20)

αi βj

i=1

Output layer The context and target representations are concatenated into m ∈ R8d , then embed them into the softmax layer for classification, and use the classification of the last layer as the judgment of emotional tendency. The calculation is as follows: 

(21) m = Cr ; Ar ; Cstrong ; Astrong   p = softmax Wp ∗ m + bp

(22)

where p ∈ RC is the probability of the target sentiment classification. Wp ∈ RC∗8d and bp ∈ RC are weight matrix and bias matrix, respectively. Setting C = 3 represents positive, neutral, and negative.

48

R. Chen et al.

3.3 Multi-granularity Attention Mechanism Bidirectional attention mechanism In this paper, bidirectional attention mechanism is added to simulate observation. Obtain αica and βiac , respectively, by mutual attention between context and target attributes. To obtain the formulas such as Eqs. (11) and (12), the bidirectional attention mechanism can get more extra information. Strong attention mechanism The attention mechanism is mainly to enable the mining of important feature attributes by letting the model automatically learn the relationship between context and current attributes. In order to extract more context and target attribute information, a strong attention mechanism is designed. Strong attention mechanism is to add attention mechanism to one’s own rows and columns, as shown in Formulas (17) and (18), which can dig deep into the important parts. 3.4 Model Training The model of this paper uses a cross entropy loss function, using L2 regularization to avoid overfitting problems. By minimizing the loss function, the model is optimized to perform Attribute-level sentiment analysis tasks on text, as follows:  P g (yi = c) · log(P(yi = c)) + λθ 2 (23) loss = − i

c∈C

Among them, λ is the L2 regularization parameter and θ is the weight matrix of the linear layer and LSTM network. Use dropout to avoid overfitting, minimize the weight matrix in the model by adding a small batch random gradient descent algorithm to the algorithm.

4 Experiment The proposed model is tested on SimEval 2014 and Twitter datasets. Word embedding is introduced by introducing pre-trained Glove word vectors, where the dictionary size and training dimensions are 1.9 MB and [25–300] dimensions, respectively. The graphics card model is Tesla T4, the driver version SIM is 410.79, the CUDA version is 10.0, the video memory is 15G, the Linux system server is used, and the deep neural network framework is Pytorch 1.1.0. 4.1 Experimental Data and Parameter Configuration The Laptop, restaurant, and twitter datasets are classified into positive, neutral, and negative categories. The distribution by emotional polarity is shown in Table 1. Parameter configuration: randomly initialize the weight matrix from the uniform distribution, set all the bias items to 0, set the regularization coefficient to 0.0001, set dropout to 0.2, and train every English vocabulary into 25–300 dimensional vector.

Multi-granularity Attribute Sentiment Analysis …

49

Table 1 Experimental datasets Dataset

Positive

Neutral

Negative

Train Test Train Test Train Test Laptop

994 341

870 128

464 169

Restaurant 2164 728

807 196

637 196

Twitter

1561 173 3127 346 1560 173

4.2 Method Comparison The algorithm of this paper and the following four models were tested in the SemEval 2014 and Twitter datasets, mainly using accuracy to measure performance. The LSTM model primarily models the sentences in the document and embeds them in the last layer for classification. A single LSTM model can achieve some effect on some special emotional classifications. The TD-LSTM [4] model is modeled mining the aspects and context of the text, Then connect the hidden states of the two LSTM networks to predict emotional polarity. The AT-LSTM [11] model mainly uses the attention mechanism to obtain context information, and adds attention mechanism to the LSTM model to semantically classify emotions. The ATAE-LSTM [11] model is an extension of the AT-LSTM model, which introduces specific target information when calculating attention weights. This paper proposes a multi-granularity neural network model, which combines the bidirectional attention and strong attention mechanism to better mine the association between context and target attributes, so that the model can classify multiple targets. 4.3 Analysis of Experimental Results Experimental comparison of the proposed model with four basic models (LSTM, TDLSTM, AT-LSTM, and ATAE-LSTM), Train and cross-validate models on SemEval 2014 and twitter datasets. In the experiment, the text is mapped into 25–300 dimension word vectors. The experimental results of the SemEval 2014 dataset (laptop and restaurant) and the twitter dataset are, respectively, plotted as a histogram as shown in Figs. 2, 3, and 4. The experimental results are recorded as shown in Tables 2, 3, and 4, respectively. From Figs. 2, 3, and 4, we can see that these models have achieved good results on SimEval 2014 and Twitter datasets. With the increase of text dimension, the classification accuracy of different algorithms increases linearly. The standard LSTM algorithm only classifies the emotional tendency of sentences. As the text dimension increases, the classification accuracy of LSTM does not increase significantly. Compared with the LSTM algorithm, both AT-LSTM and ATAE-LSTM algorithms add attention mechanisms to achieve higher classification accuracy. For the TD-LSTM algorithm to extract the association between context and target, more information of the target can be mined, so as to obtain better classification accuracy.

50

R. Chen et al.

Fig. 2 Comparison of different text dimensions in laptop

Fig. 3 Comparison of different text dimensions in restaurant

By adding bidirectional attention and strong attention mechanism, the model in the paper can better capture context and target attributes and has better classification accuracy than TD-LSTM using unidirectional attention mechanism. Compared with ATAE-LSTM and AT-LSTM with unidirectional LSTM model, this paper uses a bidirectional long-term memory model, which can better extract context features for better accuracy. Experiments were conducted on datasets in different areas of the SemEval 2014 dataset (laptop and restaurant) and the twitter dataset, with classification accuracy rates of 72.88%, 80.53%, and 73.49%, respectively. The accuracy of different algorithms in Laptop, Restaurant, and Twitter is shown in Fig. 5. In different dimensions, AT-LSTM with attention mechanism is added. ATAELSTM algorithm has higher classification accuracy than LSTM algorithm. Attention mechanism can improve the mining of target information. Compared with AT-LSTM, ATAE-LSTM algorithm adds specific target information, gets more context and target

Multi-granularity Attribute Sentiment Analysis …

51

Fig. 4 Comparison of different text dimensions in twitter

Table 2 Accuracy of different text dimensions in laptop Model

25d

50d

100d

200d

300d

LSTM

0.6424 0.6556 0.6649 0.6654 0.6668

AT-LSTM

0.6488 0.6701 0.6775 0.6854 0.6876

ATAE-LSTM 0.6504 0.6739 0.6853 0.6896 0.6934 TD-LSTM

0.6543 0.6771 0.6969 0.7084 0.7145

Ours

0.6787 0.7101 0.7130 0.7210 0.7288

Table 3 Accuracy of different text dimensions in restaurant Model

25d

50d

100d

200d

300d

LSTM

0.7323 0.7434 0.7541 0.7556 0.7576

AT-LSTM

0.7532 0.7665 0.7667 0.7677 0.7689

ATAE-LSTM 0.7544 0.7685 0.7741 0.7756 0.7776 TD-LSTM

0.7565 0.7688 0.7765 0.7857 0.7911

Ours

0.7694 0.7866 0.7920 0.7937 0.8053

association, increases the accuracy of classification, but has poor accuracy for multiobjective classification. Figure 5 shows that the TD-LSTM algorithm of bidirectional LSTM has higher classification accuracy than the AT-LSTM and ATAE-LSTM algorithm of unidirectional LSTM. The bidirectional LSTM can extract the association between context and target better and make the accuracy of emotion classification higher. This algorithm uses bidirectional LSTM, adds bidirectional attention, and strong attention

52

R. Chen et al. Table 4 Accuracy of different text dimensions in twitter Model

25d

50d

100d

200d

300d

LSTM

0.6604 0.6763 0.6779 0.6822 0.6935

AT-LSTM

0.6788 0.6805 0.6835 0.6875 0.6943

ATAE-LSTM 0.6807 0.6836 0.6894 0.6936 0.7049 TD-LSTM

0.6936 0.7124 0.7137 0.7225 0.7283

Ours

0.6965 0.7125 0.7225 0.7245 0.7349

mechanism to mine the association between context and target. A better emotional classification of multiple target attributes in sentences.

0.84

0.72

0.82

Accuracy

0.74

LSTM AT-LSTM ATAE-LSTM TD-LSTM ours

0.7

0.8

0.76 0.74

Accuracy

LSTM AT-LSTM ATAE-LSTM TD-LSTM ours

0.76

Accuracy

Accuracy of twitter

Accuracy of restaurant

Accuracy of laptop 0.78

0.78

0.72

LSTM AT-LSTM ATAE-LSTM TD-LSTM ours

0.7

0.68 0.76

0.68

0.66 0.74

0.64 0

25

50

100

200

300

0.66

0

25

50

100

200

0

300

25

50

100

200

300

Text dimension

Text dimension

Text dimension

Fig. 5 Trends in different text dimensions in laptop, restaurant, and Twitter

Figure 6 shows the variation of the loss function of the algorithm in the training process of datasets laptop, restaurant, and twitter. It can be seen from the figure that the model iterates after 20,000 iterations. restaurant

twitter 0.6

0.4

0.4

0.4

loss

0.6

loss

loss

laptop 0.6

0.2

0.2

0.2

0

0 0

4000

8000

step

12000

0 0

0.4

0.8

1.2

step

1.6

2

2.4

x 10

0 4

0.4

0.8

1.2

1.6

2

step

Fig. 6 Loss function trained by our algorithm on laptop, restaurant, and twitter

2.4

x 10

4

Multi-granularity Attribute Sentiment Analysis …

53

5 Conclusion On the issue of processing sequences, the circulating neural network has a better effect. The proposed model incorporates bidirectional long-term and short-term memory to mine the association between context and target, which is more effective than unidirectional long-term and short-term memory. By combining bidirectional attention mechanism with strong attention mechanism, we can better mine the relationship between context and target attributes and have better classification accuracy than general models. In this paper, bidirectional attention and strong attention mechanism are added to the attention model, and depth model (such as BiLSTM) is used to extract context-content correlation. Construct a model that can analyze the emotion of multiple target objects. Experiments show that the proposed model works well in the SemEval 2014 dataset and the twitter dataset. Contributions of this paper: (1) This paper proposes a strong attention mechanism, which can pay more attention to itself and dig out more important information. (2) Combining bidirectional attention and strong attention mechanism, a multigranularity attribute sentiment analysis model based on neural network is proposed to solve the sentiment analysis of multiple target attributes. (3) Experiments in SemEval 2014 (laptop and restaurant) and Twitter datasets gave better classification results.

Acknowledgements. This project is supported by the Science and Technology Support Program of Sichuan Science and Technology Department (19ZDYF1078), Zigong City Key Technology Plan (2018GYCX33). Please contact Chen Runxue for any questions.

References 1. Wang, B., Lu, W.: Learning latent opinions for aspect-level sentiment classification. In: ThirtySecond AAAI Conference on Artificial Intelligence (2018) 2. Tang, D., Qin, B., Liu, T.: Document modeling with gated recurrent neural network for sentiment classification. In: Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (2015) 3. Manek, A.S., Shenoy, P.D., Mohan, M.C., et al.: Aspect term extraction for sentiment analysis in large movie reviews using Gini Index feature selection method and SVM classifier. World Wide Web 20(2), 135–154 (2017) 4. Vo, D.T., Zhang, Y.: Target-dependent twitter sentiment classification with rich automatic features. In: Twenty-Fourth International Joint Conference on Artificial Intelligence, pp. 1347–1353 (2015) 5. Tang, D., Qin, B., Feng, X, et al.: Effective LSTMs for target-dependent sentiment classification. arXiv preprint arXiv:1512.01100 (2015) 6. Cheng, J., Zhao, S., Zhang, J., et al.: Aspect-level sentiment classification with heat (hierarchical attention) network. In: Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, pp. 97–106. ACM (2017)

54

R. Chen et al.

7. Ma, Y., Peng, H., Khan, T., et al.: Sentic LSTM: a hybrid network for targeted aspect-based sentiment analysis. Cogn. Comput. 10(4), 639–650 (2018) 8. Bojanowski, P., Grave, E., Joulin, A., et al.: Enriching word vectors with subword information. Trans. Assoc. Comput. Linguist. 5, 135–146 (2017) 9. Zhou, P., Shi, W., Tian, J., et al.: Attention-based bidirectional long short-term memory networks for relation classification. In: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, vol. 2: Short Papers, pp. 207–212 (2016) 10. Han, H., Li, X., Zhi, S., et al.: Multi-attention network for aspect sentiment analysis. In: Proceedings of the 2019 8th International Conference on Software and Computer Applications, pp. 22–26. ACM (2019) 11. Wang, Y., Huang, M., Zhao, L.: Attention-based LSTM for aspect-level sentiment classification. In: Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pp. 606–615 (2016)

Study on Fuzzy Crop Planning Problems Based on Credibility Theory Hai Tao Zhong, Ming Fa Zheng(B) , Fang Chi Liang, and Ya Nan Wang Department of Basic Sciences, Air Force Engineering University, Xi’an 710051, China [email protected]

1 Introduction Crop planning problems are widely used in agricultural production. To provide the optimal planting plan for the maximum income is the urgent need of farmers and limited social resources. However, in the real life, these problems often involve some fuzzy factors which motivate scholars to research fuzzy crop planning problems. “Fuzzy Set” [1] published by Zadeh in 1965 marks the birth of fuzzy mathematics. The proposed fuzzy variables study the non-random phenomenon, and possibility measure is used to describe a fuzzy event. Based on the possibility theory, many researchers have given many fuzzy programming models and studied the properties and solution methods [2–8]. But, as we known, possibility measure has no self duality. In practical optimal problems, there maybe happen the possibility measure of a fuzzy uncertainty event is 1, and the possibility measure of its complementary event is also 1 [9] that lead to unreasonable result. Thus, in 2002, a self-dual measure proposed by Liu and Liu [5, 9], called the credibility measure [5, 9]. Moreover, Liu [5, 10] gives the axiomatic system of credibility theory. Since then, credibility theory has become a novel tool for studying the fuzzy programming problem [11, 12]. Therefore, based on credibility theory, the paper has an emphasis on the solution method of fuzzy crop planning problem. Firstly, maximizing credibility measure model is presented, which converts the crop planning problem with income coefficients into a certainty programming problem. By using the convexity and the method of bisection, we can obtain a good solution procedure to solve. Secondly, without loss of generality, for crop planning with bi-fuzzy income coefficients, we extended the proposed maximizing credibility measure model by setting credibility measure of constraint function not less than a certain number to convert the fuzzy model into certainty programming problem and also give a solving method. Finally, an example is given to obtain the optimal fuzzy decision-making. This paper is formed as follows. Section 2, we review some basic concepts and properties of credibility theory. We present the crop planting problem with fuzzy income coefficients and its maximizing credibility measure model in Sect. 3. In addition, based on the bisection method, we give an effective solution procedure. Finally, a brief summary is given.

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 C. H. WU et al. (eds.), Recent Developments in Intelligent Computing, Communication and Devices, Advances in Intelligent Systems and Computing 1185, https://doi.org/10.1007/978-981-15-5887-0_8

56

H. T. Zhong et al.

2 Basic Principles Definition 1 [5, 10] Let  be a nonempty set, and p() the power set of , and Pos a set function defined on p() is called a possibility measure if it satisfies the following conditions: (1) Pos{∅}  = 0,Pos{} = 1; (2) Pos i∈I Ai = sup Pos{Ai }, for any subclass {Ai |i ∈ I } of p(). i∈I

The (, p(), Pos) is called a possibility space. Definition 2 [9, 10] Let (, p(), Pos) be a possibility space. The credibility measure denoted by Cr is defined as Cr{ϑ} =

  1 1 + Pos{ϑ} − Pos ϑ c , for any ϑ ∈ p(), 2

where ϑ c is the complement of ϑ. The (, p(), Cr) is called a credibility space. Definition 3 [10, 11] Let (, p(), Pos) be a possibility space, and ξ fuzzy variable defined by ξ : (, p(), Pos) → R. (1) Membership function μ(x) = Pos{ϑ ∈ |ξ(ϑ) = x}, (2) Credibility distribution (x) = Cr{ϑ ∈ |ξ(ϑ) ≤ x}.

3 Fuzzy Crop Planning Problem Farmers should decide what crops they plant before seeding in order to obtain maximum income. Usually, they set aside a certain plot of land for each crop to cultivate. Of course, the total farmland is limited, the labor and profit for each crop are different. Now, we suggest that there are n kinds of crops, the cultivation area of the i th crop denoted by xi (i = 1, 2, . . . , n), respectively. Let γ t = (γ1 , γ2 , . . . , γn ) be the profit per unit of crops (profit coefficients), l t = (l1 , l2 , . . . , ln ) the labor force pre unit for the crops. And the total farmland is limited by M, so we have x1 + x2 + · · · + xn ≤ M (Farmland Constraint) The total labor force is limited is less than or equal to L, we have l1 x1 + l2 x2 + · · · + ln xn ≤ L(Labor Force Constraint). Therefore, a crop planning problem is formed as the following linear programming: ⎧ max γ1 x1 + γ2 x2 + · · · + γn xn = γ t x ⎪ ⎪ ⎨ s.t. x1 + x2 + · · · + xn ≤ M (P1) ⎪ l1 x1 + l2 x2 + · · · + ln xn ≤ L ⎪ ⎩ x ≥ 0. where x = (x1 , x2 , . . . , xn )t .

Study on Fuzzy Crop Planning Problems Based on Credibility Theory

57

3.1 Problem with Fuzzy Income Coefficients In real life, the crop planning problem often involve some fuzzy factors. We assume that the income coefficient γi for ith crop, respectively, is a variable defined γi : (, p(), Pos) → R(i = 1, 2, . . . , n). So the vector γ t = (γ1 , γ2 , . . . , γn ) is properly defined a fuzzy income coefficient vector. The membership function of γ = (γ1 , γ2 , . . . , γn )t can be the following unimodal function [8]: μγ (x) = Pos{ϑ ∈ |γ (ϑ) = x} = ρ((x − q)t D(x − q)).

(1)

Here qt = (q1 , q2 , . . . , qn ) is a constant value vector whose ith component denotes mean value of corresponding γi , respectively, and D is an n × n diagonal positive definite matrix denoting the spreads of γ t = (γ1 , γ2 , . . . , γn ). Let ρ be a continuous decreasing function satisfying: ρ : [0, +∞) → [0, 1], ρ(0) = 1, existing α such that α = inf {α|ρ(α) = 0}. Based on credibility theory, the membership function of γ t x = Y is presented as follows: 



Theorem 1 If μγ (s) = ρ((s − q)t D(s − q)), then the membership function of Y is

(y − qt x)2 . (2) μY (y) = ρ xt D−1 x Proof We all know, γ is a fuzzy vector. Let γ (ϑ) = s, for any ϑ ∈ . By using the extension principle [10], we have μY (y) =

sup γ t (ϑ)x=y

ρ((s − q)t D(s − q)).

(3)

Here the function ρ is a decreasing function, we can construct the lagrangian function,   (4) ρ(s, k) = (s − q)t D(s − q) + k y − st x . Partial differentiating ρ(s, k) by s and k, we have 2D(s − q) − kx = 0,

(5)

y − st x = 0.

(6)

2xt DD−1 (s − q) − kxt D−1 x = 0 ⇔ 2xt (s − q) − kxt D−1 x = 0, .

(7)

Multiply (5) by xt D−1 , then

From (6), we have 2y − 2xt q − kxt D−1 x = 0.

(8)

Then,   2 y − qt x y − qt x x k= , s = + q. xt D−1 x xt D−1 x D

(9)

58

H. T. Zhong et al.

Substituting (3) for (9), we prove this theorem. As γ is a fuzzy vector, so we can’t get an optimal value through classical optimization theory. Based on credibility theorem, we can define the fuzzy optimal value by trying to obtain the maximum credibility measure of Y. There the credibility distribution and the goal credibility distribution of γ t x = Y are presented as follows: ⎧  ⎨ 0, y < a, μY (y) t x, , y < q y−a 2 Y (y) = (y) =  , a ≤ y < b, G ⎩ b−a 1 − μY2(y) , y ≥ qt x. 1, y ≥ b. Then we convert the problem (P1) into maximizing credibility measure model (P2), ⎧ max Y (y) = G (y) ⎪ ⎪ ⎨ s.t. x1 + x2 + · · · + xn ≤ M (P2) ⎪ l1 x1 + l2 x2 + · · · + ln xn ≤ L ⎪ ⎩ x ≥ 0.

3.2 Method Let



∗

ρ (w) = and ∗G (w)

 =

min{r|ρ(r) ≤ w}, 0 ≤ w < 1, 0, w = 1.

(10)

0, w = 0, min{r|G (r) ≥ w}, 0 < w ≤ 1.

(11)

Theorem 2 If η = 0, 1, then Y (y) = G (y) ≥ η ⎧   ∗ ⎪ ⎪ ρ ∗ 2G (η)−2a xt D−1 x − qt x + 1 (b + a) ≥ 0, 0 < η < 1 , ⎨ b−a 2 2   ⇔  ∗ ⎪ 2b−2G (η) t −1 1 ⎪ ⎩ − ρ∗ x D x − qt x + b ≥ 0, b−a 2 ≤ η < 1. Proof

Y (y) = G (y) ≥ η ⇔ G (y) ≥ η, Y (y) = G (y).

(12)

(13)

From (11), we have ⇔ ∃y : ∗G (η) ≤ y, Y (y) = G (y). When 0 < η < 1/2, y ≤ (b + a)/2, from (13), we have   (y−qt x)2 ρ y−a xt D−1 x = . ∃y : b−a 2

(14)

Study on Fuzzy Crop Planning Problems Based on Credibility Theory

From (14), ∗G (η)

≤y=

ρ



(y−qt x)2 xt D−1 x

2

 (b − a) + a,

59



2∗G (η) − 2a (y − qt x)2 ≤ρ . b−a xt D−1 x

As ρ is a decreasing function, from (10), we have ∗

(y − qt x)2 ∗ 2G (η) − 2a ≥ t −1 . ρ b−a xD x

(15)

Simplifying (15), then  

∗ (η) − 2a 2 2∗H (η) − 2a t −1 t G ∗ t −1 ∗ x D x+q x ≤y ≤ ρ x D x + qt x. − ρ b−a b−a Above all, when 0 < η < 1/2, 

2∗G (η) − 2a t −1 1 ∗ x D x − qt x + (b + a) ≥ 0. (13) ⇔ ρ b−a 2 When 1/2 ≤ η < 1, y ≤ b, from (13), ρ y−a =1− ∃y : b−a



(y−qt x)2 xt D−1 x

2

 .

By (14), we have ∗G (η) ≤ y =

2−ρ



(y−qt x)2 xt D−1 x

2

 (b − a) + a,



2b − 2∗G (η) (y − qt x)2 ≥ρ . b−a xt D−1 x

As ρ is a decreasing function, from (10), we have

2b − 2∗G (η) (y − qt x)2 ρ∗ ≤ t −1 . b−a xD x

(16)

Simplifying (16), then  

∗ (η) 2b − 2 2b − 2∗G (η) t −1 t G y ≤ − ρ∗ xt D−1 x + q x or x D x + qt x ≤ y. ρ∗ b−a b−a So, we have



ρ∗

2b − 2∗G (η) t −1 x D x + qt x ≤ y ≤ b. b−a

Then, when 1/2 ≤ η < 1, 

2b − 2∗G (η) t −1 ∗ x D x − qt x + b ≥ 0. (13) ⇔ − ρ b−a

60

H. T. Zhong et al.

In the end, we prove the theorem. t ⎧   ⎪ 2ΦG∗ (η)−2a ⎪ ⎨ ρ∗ xt D−1 x − qt x + 21 (b + a), 0 < η < 21 , b−a   Q(x, η) =  ⎪ 2b−2ΦG∗ (η) t −1 1 ⎪ ⎩ − ρ∗ x D x − qt x + b, b−a 2 ≤ η < 1. where Y (y) and G (y) are continuous functions, we can convert (P2) into the following (P3): ⎧ ⎪ max η ⎪ ⎪ ⎪ ⎪ s.t. x1 + x2 + · · · + xn ≤ M ⎪ ⎪ ⎨ l1 x1 + l2 x2 + · · · + ln xn ≤ L (P3) ⎪ Q(x, η) ≥ 0 ⎪ ⎪ ⎪ ⎪ x≥0 ⎪ ⎪ ⎩ 0 < η < 1, There ∗G (η) is an increasing function of η, and ρ ∗ (η) is a decreasing function of η, 

2∗G (η) − 2a t −1 1 x D x − qt x + (b + a), ρ∗ b−a 2 

2b − 2∗G (η) t −1 x D x − qt x + b, − ρ∗ b−a are decreasing functions of η. In other words, Q(x, η) is a decreasing function of η. Where X1 satisfying x1 + x2 + · · · + xn ≤ M , l1 x1 + l2 x2 + · · · + ln xn ≤ L, x ≥ 0. Theorem 3   If max Q x, η > 0, then η∗ > η, x∈X1   If max Q x, η < 0, η∗ < η, x∈X1   If max Q x, η = 0, then η∗ = η. 









(17)



x∈X1

Here η∗ is the optimal value of the problem (P3).     Proof Let maxx∈X1 Q x, η = Q xˆ , η . The corresponding maxx∈X1 Q(x, η∗ ) = Q(x∗ , η∗ ) = 0, x∗ is the optimal solution. If Q(x∗ , η∗ ) > 0, as Q(x, η) is a decreasing η) ≥ 0 satisfying the confunction for η, so there exists other  η > η∗ such that Q(x∗ ,  straint. This is conflicts with optimal value η∗ . Therefore, we have maxx∈X1 Q(x, η∗ ) = Q(x∗ , η∗ ) = 0. 



The following cases are demonstrated:  case (1): max Q x, η > 0, if η∗ ≤ η 



x∈X1

          max Q x, η = Q xˆ , η ≤ Q xˆ , η∗ ≤ Q x∗ , η∗ = 0 ⇒ max Q x, η ≤ 0. 

x∈X1





x∈X1

Study on Fuzzy Crop Planning Problems Based on Credibility Theory

61

  This is conflicts with max Q x, η > 0, so we have η∗ > η.   x∈X1 case (2): max Q x, η < 0, if η∗ ≥ η, 







x∈X1

          max Q x, η = Q xˆ , η ≥ Q x∗ , η ≥ Q x∗ , η∗ = 0, ⇒ max Q x, η ≥ 0. 







x∈X1

x∈X1

  This is conflicts with max Q x, η < 0, so we have η∗ < η. 



x∈X1

Algorithm 1 √ √ Step 1: Calculate − α xt D−1 x + qt x − b, α xt D−1 x + qt x − a. 



√ √ If maxX1 − α xt D−1 x + qt x − b ≥ 0 or minX1 α xt D−1 x + qt x − a ≤ 0, then terminate with the optimal value 1 and corresponding optimal solution x. 



Step 2: Set Z = 0, R = 1. Step 3: Calculate η = (Z + R)/2.   Step 4: When η = 1/2, calculate max Q x, 21 = max b − qt x. 



x∈X1

x∈X1

If max b − qt x > 0: x∈X1

case 1: max b − qt x − x∈X1

√ α xt D−1 x ≥ 0 and max 21 (b + a) − qt x > 0, 

x∈X1

then terminate with the optimal value 1and corresponding xˆ ; √ case 2: max b − qt x − α xt D−1 x < 0, then set Z = η and turn to Step 3; 

x∈X1

case 3: max b − qt x − x∈X1



√ α xt D−1 x ≥ 0 and max 21 (b + a) − qt x < 0, 

x∈X1



then set R = η and turn to Step 3; √ case 4: max b − qt x − α xt D−1 x ≥ 0 and max 21 (b + a) − qt x = 0, 

x∈X1

x∈X1

then terminate with the optimal solution xˆ and optimal value 1/2. 

If max b − qt x < 0, then set R = η and turn to Step 3. x∈X1

Otherwise max b − qt x = 0, terminate with the optimal value 1/2 and the corresponding x∈X1

optimal solution xˆ .   Step 5: When η = 1/2, calculate max Q x, η . 



x∈X1

62

H. T. Zhong et al.

  If max Q x, η > 0, then set Z = η and turn to Step 3. 



x∈X1

  If max Q x, η < 0, then set R = η and turn to Step 3. 

x∈X1





Otherwise the value is 0, terminate with η as the optimal value and the corresponding xˆ as the optimal solution. Furthermore, we can know that some extreme point of X1 is the optimal solution. So we don’t calculate all x ∈ X1 to simplify Step 5 to obtain the solution effectively.

4 Conclusion For the crop planning problem, we consider the real condition that the future profit is fuzzy. In the paper, we suppose the income coefficients of planning problem are fuzzy variables whose membership is a unimodal function. And the credibility distribution is presented to construct certain model. Based on credibility theory, we give some theoretical analysis for the planning problem with fuzzy variable and propose an useful algorithm. However, in real life, for the crop planning problem there are many factors affecting crop planting, such as light, rain, and soil. This article only considers the future profit but ignores these factors that influence each other and even contradict with. So it is necessary to consider more factors to construct fuzzy programming model for formulating better planting plan. More, such as multi-objective programming or fuzzy random programming. In the long term, we can study the crop planning problems with some other fuzzy factors, we also study on other optimization problems with fuzzy factors based on credibility theory. Acknowledgments. Supported by the Research Fund of Department of Basic Sciences at Air Force Engineering University (2019101). Supported by the Research Fund of University at Air Force Engineering University (2019028). Supported by the Natural Science Foundation of Shaanxi Province of China (2019JM-271).

References 1. Zadeh, L.A.: Fuzzy sets. Inf. Control 8, 338–353 (1965) 2. Dantzig, G.B.: Linear programming under uncertainty. Manage. Sci. 1(3–4), 197–206 (1955) 3. Chen, W.T., Li, Y.P., Huang, G.H.: A two-stage inexact-stochastic programming model for planning carbon dioxide emission trading under uncertainty. Appl. Energy 87(3), 1033–1047 (2010) 4. Wang, S., Huang, G.H.: Interactive two-stage stochastic fuzzy programming for water resources management. J. Environ. Manage. 92(8), 1986–1995 (2011) 5. Kumar, T., Bajaj, R.K.: Expected Value Based Ranking of Intuitionistic Fuzzy Variables. AIP Publishing (2017) 6. Zadeh, L.A.: Fuzzy sets as a basic for a theory of possibility. Fuzzy Sets Syst. 1, 3–28 (1978) 7. Wen, B., Hongguang, L.I., Chen, X.C.: A modeling approach for fuzzy programming with echelon form membership functions. J. Beijing Univ. Chem. Technol. (Nat. Sci. Ed. ) 45(1), 78–83 (2018)

Study on Fuzzy Crop Planning Problems Based on Credibility Theory

63

8. Toyonaga, T.: A Crop Planning Problem with Fuzzy Random Profit Coefficients, vol. 4, pp. 51–69. Springer Science+Business (2005) 9. Liu, B., Liu, Y.K.: Expected value of fuzzy variable and fuzzy expected value models. IEEE Trans. Fuzzy Syst. 10, 445–450 (2002) 10. Liu, B.: Uncertainty Theory. Springer Nature (2004) 11. Liu, B.: A Course in Uncertainty Theory. Tsinghua University Press (2005) 12. Bai, H., Zhang, P., Su, J.: Model study for determining the wind farm penetration limit based on corrected energy function and credibility theory. Trans. China Electrotech. Soc. 32(8), 264–273 (2017)

Empirical Research on the Antecedents of Enterprise IS Users’ Voice Behavior Based on the Work Behavior Theory Manhui Huang(B) and Changxian Li Guangdong University of Finance and Economics, Guangzhou 510320, China [email protected]

1 Introduction IS Users’ behavior is the key effect factor of IS success. However, extant research has focused on the in-role behaviors of IS users such as IS adoption and usage, while paid little attention to the voice behavior which belongs to the extra-role behaviors [1, 2]. People are one of the key resources for achieving organizational performance, and it is important to play the role of employees’ creativity and encourage employees’ voice behavior for organizational innovation [3]. Similarly, in the IS context, IS users’ voice behavior is important for achieving IS performance [4]. Voice is a contextual factor [5]. Past research has analyzed the antecedents of voice behavior from different perspectives such as personality, emotion, leadership, organizational culture in different contexts [6– 9]. However, research on the antecedents of voice behavior in the IS context is still limited. To fill the research gap, this paper tries to analyze the antecedents of IS users’ voice behavior based on the work behavior theory. Work behavior theory has been widely applied to explain employees’ work behavior [10], as voice behavior belongs to work behavior, it can be used to analyze IS users’ voice behavior. Voice behavior refers to expressing rather than withholding relevant ideas, information, and opinions about possible work-related improvements [5]. In this paper, IS users’ voice behavior refers to express relevant ideas, information, and opinions about IS-related improvements.

2 Theoretical Background and Hypotheses Work behavior theory is widely used to explain employee’s work behavior [11], which states that opportunity, willingness, and capacity will predict the work behavior of an employee [10]. Capacity refers to the capabilities that enable an individual to complete a task effectively, which may include specific abilities and personality [12]. Willingness refers to the psychological and emotional situations that influence an individual to perform a task. Opportunity refers to the environmental factors beyond an individual’s control that influence the employee’s work behavior [10, 13, 14]. First, according to the work behavior theory, in the IS context, IS user needs the opportunity to conduct IS voice behavior. If the enterprise doesn’t support innovation and © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 C. H. WU et al. (eds.), Recent Developments in Intelligent Computing, Communication and Devices, Advances in Intelligent Systems and Computing 1185, https://doi.org/10.1007/978-981-15-5887-0_9

Empirical Research on the Antecedents of Enterprise …

65

doesn’t encourage her employees to express different opinions about possible improvements of IS, IS users will tend not to conduct voice behavior. Therefore, an innovative atmosphere is an important opportunity for conducting IS voice behavior. Past research has shown that an innovative atmosphere has a positive effect on voice behavior [15]. Therefore, in the IS context, the following hypothesis is proposed: H1: Innovative atmosphere has a positive effect on IS users’ voice behavior. Second, according to the work behavior theory, an individual’s willingness to do the work will predict her work behavior. Emotion may reflect willingness. If one is eager to do the work, she will be happy to do it. Past research has shown that emotion will predict attitude and then influence work behavior [16]. Therefore, in the IS context, we suggest that if one is happy to use the IS, she will tend to conduct voice behavior toward the application of IS. Therefore, the following hypothesis is proposed: H2: IS users’ positive emotion toward IS usage has a positive effect on their voice behavior toward the application of IS. Third, according to the work behavior, the ability will predict work behavior. Ability refers to the skills to complete a specific task [12]. In the IS context, if one has the skill to operate the information systems, she will have more knowledge about the application of IS, and then it is more possible to conduct IS voice behavior. In the extant IS research, the construct perceived ease of use is used to reflect one’s ability to operate the information systems, which means one’s perception that to operate the information systems is free of effort [17]. Therefore, the following hypothesis is proposed: H3: IS users’ perceived ease of use will predict IS voice behavior. Fourth, work behavior theory shows that personality will predict work behavior. To finish a specific task, some kinds of personalities are necessary. Past research has shown that extroversion is important to conduct voice behavior [18–20]. If one has an extrovert personality, she will tend to share opinions with some other people, and then tend to conduct voice behavior during her work. Similarly, in the IS context, if the employee has an extrovert personality, she will tend to express her opinions about the application of IS. Therefore, the following hypothesis is proposed: H4: IS users’ extrovert personality will predict IS voice behavior.

3 Research Method To test the hypotheses, a survey on the IS users in enterprises in an anonymous manner was conducted to collect the data. Totally 187 questionnaires were finally collected. Among the 187 IS users, 48.7% are female, 70.1% are young people with the age below 25, 87.7% have the education with undergraduate degree, 25.6% have 2 years work experience, and 17.1% have IS usage experience with over 5 years. The constructs in the current research are all measured in 5-point Likert scale. All the measurements used in this research are adapted from the extant research [17, 21–24] to assure the reliabilities and validities of the measurements.

66

M. Huang and C. Li

4 Research Results To test the reliabilities of the measurements, Cronbach’s α test is conducted. Usually, if Cronbach’s α coefficient is over 0.7, the reliability of the measurements will be acceptable [25]. With SPSS19.0, the results show that Cronbach’s α of all the measurements of the constructs in the current research including IS voice behavior, innovative atmosphere, perceived ease of use, positive emotion, and extrovert personality are over 0.7, which indicates the acceptable reliabilities. To test the validity of the questionnaire, EFA (Exploratory Factor Analysis) is conducted. Totally 5 factors emerged with 84.904% variance explained. The loadings of items on their respective factors are all over 0.4, and the cross-loadings are small, which indicate that the validities of the measurements are acceptable. To test the hypotheses, regression is conducted. To exclude the effect of demographic factors on the dependent variable, the demographic factors including gender, age, education level, years of work experience, and years of IS usage experience are added in the model as the control variables. We conducted a two-step regression. In the first step, the control variables are added in the model as independent variables. In the second step, the predictor variables including innovative atmosphere, positive emotion, perceived ease of use, extrovert personality are added to the model. The results are shown in Table 1, which shows that innovative atmosphere has a positive effect on IS user voice behavior (β = 0.283, p < 0.001), positive emotion has a positive effect on IS user voice behavior (β = 0.163, p < 0.05), perceived ease of use has a positive effect on IS user voice behavior (β = 0.167, p < 0.01), and extrovert personality has a positive effect on IS user voice behavior (β = 0.204, p < 0.001). Therefore, hypotheses 1–4 are all supported.

5 Discussion and Conclusion We conducted this research based on the research gap that most of the extant research has focused on in-roles behaviors such as IS adoption and usage, while paid little attention to IS voice behavior which belongs to extra-role behaviors in IS context [3]. As one kind of creative behavior, IS voice behavior is important for improving IS performance [4]. Based on the work behavior theory in the organizational behavior field, we analyzed the effect factors of IS voice behavior. The results show that the organizational factor innovative atmosphere, the individual factors including positive emotion, perceived ease of use, and extrovert personality have an important effect on IS voice behavior. First, innovative atmosphere has a positive effect on IS voice behavior. This implied that if organizations encourage their employees to express their different creative opinions about how to apply the IS to better support the business, the employees will have more opportunity to do so, and they will feel more relaxed to voice if they find the opportunities to improve the application of IS. Therefore, to improve the IS performance, enterprises need to create innovative atmosphere within the organizations and encourage their employees to voice. Second, positive emotion has a positive effect on IS voice behavior. This implied that enterprises need to interfere with employees’ emotion during the application of IS, e.g., design incentive mechanism about IS voice so that IS users may have positive emotion

Empirical Research on the Antecedents of Enterprise …

67

Table 1 Results of regression Variables

Model 1 Model 2

Gender

−0.073

−0.065

Age

0.245*

0.220*

Education level

0.014

0.007

YW

−0.215

−0.088

YU

0.189

0.021

IA

0.283***

PE

0.163*

PEU

0.167**

EP

0.204***

R2

0.059

0.448

Adjusted R2

0.033

0.420

R2

0.059*

0.389***

Notes YW years of work experience, YU years of IS use, IA innovative atmosphere, PE positive emotion, PEU perceived ease of use, EP extrovert personality. Dependent Variable is IS users’ voice behavior. Entries are standardized regression coefficient *