24

2016-11

第17期 ②

来源:系统管理员     浏览次数:

报 告 人:Sergey Mosin 教授
                  俄罗斯喀山联邦大学
主 持 人:黄哲学
日      期:2016 年 11月 2 日
时      间:下午 2:30-4:30
地      点:计软学院623会议室

Part I: An Approach to Construction the Neuromorphic Classifier for Analog Fault Testing and Diagnosis

The alternative approach to fault dictionary (FD) construction consists in implementation the neuromorphic classifier based one artificial neural networks (ANN). Fault dictionary based on ANN allows eliminating many problems of the parametric FD. An ANN with fixed structure can be used for representing the different quantity of considered fault-free and faulty an analog circuit conditions, thus only the limited small number of coefficients describing synaptic links should be stored. The ANN provides essential time reduction on the responses comparison due to associative operation, i.e. the indicator of T&D passing is generated immediately at the output layer after application of the CUT's output response to neurons in input layer.
Main problems of neuromorphic classifier construction are the complexity of selection the structure of ANN, representation of the CUT's output responses and training of a neural network. Due to the component tolerances and nonlinearities, the output responses of fault-free and faulty analog circuits can be similar and provide wrong test and diagnosis result. Selecting the essential characteristics of output responses, which provide a distinguishability of fault-free and faulty behavior of the CUT is important in order to ensure the convergence during ANN training. The number of such characteristics for each response defines the number of neurons in the input layer of ANN. The data quantity of training set increases greatly with increasing the number of fault conditions, number of training patterns and quantity of used characteristics at each pattern. The redundancy of the essential characteristics set can lead to the overfitting of ANN and loss the accuracy at T&D.
The approach to construction the neuromorphic classifier improving the efficiency of fault T&D and solving the problems of training the appropriate ANN is presented in the paper. The approach takes into account the component tolerances and is based on the use of wavelet coefficients for representation the CUT's output responses, faults clustering and selection only such wavelet coefficients, which provide maximum distances between all pairs of considered circuit’s conditions (fault-free and all faulty).

 

Part II: Quality Improvement of Analog Circuits Fault Diagnosis Based on ANN Using Clusterization as Preprocessing

The infinite set of possible responses corresponds to a fault-free and each of faulty conditions due to component tolerances. Moreover the tolerances have casual nature and can lead to generation the subset of similar output responses for different faults. Consideration of the first circumstance at a fault dictionary construction is provided by representing each controlled parameter in the form of a range of acceptable values. The second circumstance can cause the detection error or degrading the resolving ability of a diagnostic test. The set of output responses with sufficient cardinality corresponding to each circuit condition is used in the case of ANN training. The values of output neurons are become sensitive to values of corresponding applied responses in results of ANN training. The overlapping of training patterns used for different circuit conditions can cause the problems of convergence or overfitting during an ANN training due to the second circumstance.
The technique of quality improvement of the analog circuits faults diagnosis using artificial neural network owing to: 1) preliminary clustering the output responses of considered faults before the ANN training, and 2) probabilistic prediction of presence the detected fault in the frame of each cluster during testing is proposed. The clusterization ensures localization the overlapping subsets of output responses for different faults, allowing to reduce the uncertainty of used training patterns.

 




 

相关论文发表

博士生导师

  • 黄哲学

    黄哲学

    黄哲学,瑞典皇家理工学院博士、深圳大学特聘教授、博士生导师,深圳大学大数据技术与应用研究所所长、大数据系统计算技术国家工程实验室副主任,首批广东省领军人才、深圳孔雀计划高层次人才,斯坦福大学全球“终身科学影响力排行榜”前2%顶尖科学家。符号数据快速聚类算法研究的开拓者,发表了k-modes等一系列著名聚类算法,被纳入国内外教科书和专著,进入软件产品。发表学术论文250多篇,主要论文被引用超万次。领导开发了全球首个面向算力网络的多数据中心大数据协同计算系统Octopus,最近获深圳第二十五届中国国际高新技术成果交易会“优秀产品奖”和“华为杯”第五届中国研究生人工智能创新大赛“一等奖”。
  • 黄哲学

    崔来中

    2007年6月于吉林大学获工学学士学位,同年被免试推荐直接攻读博士研究生,2012年6月于清华大学获计算机科学与技术博士学位。研究领域包括:下一代互联网体系结构、软件定义网络、边缘计算、大数据分析、机器学习和智能计算。国际电子工程师学会高级会员(IEEE Senior Member),中国计算机学会高级会员(CCF Senior Member),人工智能学会(CAAI)会员,CCF互联网专委会常委,CCF大数据专家委员会委员、CCF区块链专委会委员,CAAI知识工程与分布智能委员会副秘书长。担任SCI期刊《International Journal of Machine Learning and Cybernetics》、《International Journal of Bio-Inspired Computation 》和《Ad Hoc and Sensor Wireless Networks》的副编辑/编委。已主持国家重点研发计划课题、国家自然科学基金,广东省自然科学基金,广东省育苗工程,深圳市基础研究计划项目等项目10多项。已在国内外重要期刊以及国际会议上发表SCI/EI检索论文80余篇。《计算机网络》课程负责人,课程入选广东省一流本科课程。入选广东省青年珠江学者,深圳市优青、深圳市高层次人才和深圳大学“荔园优青”人才培养计划。
  • 黄哲学

    陈梓楠

    陈梓楠(博士,国家海外优青,IEEE会员,ACM会员)现在担任深圳大学计算机与软件学院特聘教授。在研期间一共发表了顶级会议和期刊将近30篇论文,其中CCF A类论文有19篇(第一作者有12篇),主持了国家自然科学优秀青年(海外)项目1项和国家自然科学青年基金项目1项。此外,陈老师也是各大国际会议(包括:VLDB 2022 - 2024 (demo track)、VLDB 2025 (research track)、SIGKDD 2024 、ICDE 2022和2024、EDBT 2023、IJCAI 2020、DASFAA 2021 - 2024和WISE 2019 - 2024)和国际期刊(包括:VLDBJ、TKDE、AIJ、IEEE Transactions on Computers (TC)、WWWJ、 TSAS 、TNSE、PR Journal、DKE、JCST、The Journal of Supercomputing等等)的审稿人,并担任MDM 2021 - 2024的会议论文集主席 (proceedings chair)。

科研项目

联系我们
0755-2653 0821