Research On Adaptive Fault Diagnosis Method Based On Improved Dynamic Principal Component Analysis

With the rapid development of science and technology in today's society,the production equipment and systems in modern industrial processes become more and more complex.Once complex production process equipment produced a fault,finding fault sources would be a very complex process.So the research and development of new theories and methods which are able to meet the requirements of modem industrial fault diagnosis has become the most pressing issues.Fault diagnosis technology based on multivariate statistics is an important research branch in the field of fault diagnosis,which has a strong practicality.In this thesis,based on the data of the fault diagnosis method,in view of the actual industrial process,the data obtained are dynamic,nonlinear and statistical models can not be updated in a timely manner and so on,this thesis deeply explore and study the method of fault diagnosis in industrial process.The main work of this thesis is as follows:Firstly,this thesis introduces the basic knowledge of modern industrial process fault diagnosis technology and the performance evaluation index of fault diagnosis,the research situation at home and abroad,as well as several commonly used methods of fault diagnosis.Secondly,the fault diagnosis method based on principal component analysis,the main algorithm of dynamic principal component analysis(DPCA)and fault diagnosis method based on DPCA are studied respectively.The Tennessee-Eastman(TE)process simulation proves that the dynamic principal component analysis can extract the information more effectively,and it can be used to detect and diagnose the faults of the system.Thirdly,in view of the nonlinear characteristic of the data in the actual industrial process,a fault diagnosis method based on dynamic kernel principal component analysis is proposed,which is based on the kernel theory and is a combination of the kernel method and the dynamic principal component analysis.And in view of the problem of the large amount of computation of the kernel matrix,the method of feature vector extraction(FVS)is adopted to improve the DKPCA,by selecting as few features as possible in high dimensional space to describe the entire set of samples,meanwhile guaranteeing that their distribution characteristics are basically the same.After improvement,the kernel matrix computation time is greatly reduced,which greatly reduces the complexity of the computation.Finally,aiming at the problem that the traditional DPC A algorithm cannot update the monitoring model in real time,combined with the MWPCA algorithm,a fault diagnosis method based on adaptive DKPCA is proposed.The idea of the algorithm is to use a fixed length time window moving backward to extend the data of each new acquisition,at the same time,eliminate the corresponding old data,so that the length and the amount of data the entire form does not change,so as to update the statistical model.In this method,the recursive principal component analysis(RPCA)algorithm is introduced,and the recursive formula of the simplified self-correlation matrix is given,which speeds up the rate of the dynamic model on line real time update.The effectiveness of the algorithm is verified by the simulation of TE process.