Research On Fault Detection And Classification Of Complex Systems Based On Kernel Independent Component Analysis

In order to ensure the product quality and safety of the process,complex industrial process monitoring and fault diagnosis have become hot research topic in the field of process control.The method which is based on data driven research needs no accurate analytical model of the process and less priori knowledge.And this method can build models on normal condition and fault conditions respectively according to the process data.Compared with the method which is based on analytical model and knowledge,this method is more suitable for fault diagnosis in the modern complex industrial process.At present,there are a lot of fault diagnosis researches on single characteristic data,such as nonlinear,non-Gaussian process,dynamic process and so on.However,there are less study on the variety of data characteristics coexist,especially the non-Gaussian and nonlinear process.Most of the complex industrial processes have nonlinear and non-Gaussian characteristics,so it has very important theoretical and engineering significance in doing fault detection and diagnosis research on the complicated process which full of nonlinear and non-Gaussian characteristics.Aiming at the problems of the actual industrial process research on fault diagnosis,this paper combine with the latest research results make some research and some put forward a number of methods.The aim of the study is to research the problems of the actual industrial process research on fault diagnosis,and this paper combines with the latest research results and puts forward a number of methods.The traditional fault diagnosis which is based on principal component analysis(PCA)needs to assume the process variables satisfied the Gaussian distribution and it is just fitted for the linear processing.However,the real complex industrial process data usually have strong nonlinear and non-Gaussian characteristics.The independent component analysis(ICA)has poorly effect in the process with complex nonlinear even though it could diagnoses fault on non-Gaussian distribution data.Therefore,fault diagnosis which is based on the kernel independent component analysis(KICA)has become an effective method to deal with this problem.This method was applied to fault diagnosis experiment,which verified the KICA method has better detecting effect than KPCA and ICA method.And this method is more suitable for real complex industrial process data which have nonlinear and non-Gaussian characteristics.Considering the industrial process,the data usually have strong non-Gaussian and nonlinear characteristic.In order to further improve the detection effect,combining with support vector data description(SVDD)which is not limit for data distribution,a complex system fault detection method based on KICA and SVDD was proposed.Firstly,KICA was carried out to extract independent component on process data.Then we use the SVDD to model the extracted leading independent component and to calculate the statistics and the control limits.So that the fault on the nonlinear and non-Gaussian system could be detected.Finally,compared with the single KICA and SVDD method,experimental results on the Tennessee-Eastman(TE)process' s simulation study show that the proposed method reduces the fault misclassification ratio and miss detection ratio,which verify the proposed method's feasibility and validity.The multivariate statistics based on Kernel Independent component analysis(KICA)is mainly used for industrial process of fault detection.While,it is not effective in fault classification.For this reason,a fault classification method of complex system with Stacked Sparse Auto encoder(SSAE)and the softmax classifier was proposed.First of all,the KICA is utilized to extract leading independent component among the data set,and then model was set up on the training data according the complete SSAE algorithm.Meanwhile,Greedy Layer-Wise Unsupervised Learning algorithm was used to initialize the weights of the network and L-BFGS algorithm optimized the parameters of the network.Finally,the test data was fed to the trained model to realize fault classification.According to the experimental result on Tennessee Eastman(TE)simulation platform,the average classification rate of KICA-SSAE shows higher performance,and this method proved a better classification effect.