Fault Detection And Identification Of Chemical Process Based On Sparse Representation And Probability Principal Component Analysis

With the rapid development of science and technology,the degree of modern industrial intelligence is deepening,the traditional fault diagnosis technology can not meet the needs of modern industrial automation.In the long-term production and operation process,it is inevitable that some damage to components and line ageing will result in failure of the system,which can result in economic gain and even explosion to danger personal and property safety.Therefore,the fault diagnosis technology has been widely concerned by experts and scholars at home and abroad.As a typical data-driven algorithm,multivariate statistical control method can successfully monitor industrial process online.This thesis takes Tennessee-Eastman Chemical Process(TE)as the research platform,through which the production simulation process is completed and the TE operation process data is collected.In characteristics of the high dimension and complex composition of the data collected in the industrial process,PCA projects the multi-index data of the high dimension space into the direction of variance which contains the dynamic characteristics of the data.However,the traditional principal element analysis has some disadvantages.The two statistics of PCA based fault detection are subject to different distribution and are insensitive to implied variables.In this paper,we introduce a fault detection model based on an improved probabilistic PCA,which can estimate the distribution of hidden variables,and then monitor the probability distribution of primary and error subspaces directly.However,due to sensor failure,network error,computer failure,data loss and data record error,the data collected in the real industrial production process has some coarse difference(outliers).The traditional multivariate statistical control method is difficult to deal with rough error data.To solve this problem,a fault detection method based on sparse subspace and probability principal element analysis is proposed.Firstly,the rough error data in Affine space is extracted by the sparse optimal model,and the probability principal element model is constructed by the rough difference data to detect the failure of TE process.Fault detection is only the first step in process monitoring and we also requires fault identification.With the success of sparse representation algorithm in face recognition,this paper introduces it into fault recognition of industrial process.Firstly,22 kinds of fault training sample data are used to construct the training sample lexicon,then the sparse matrix is obtained by the sparse representation classification algorithm,and the residual errors of each type are identified according to the test sample.Finally,the advantages of sparserepresentation algorithm are emphasized through simulation experiment of Tennessee Eastman Chemical Process(TE)research platform.