Fault Detection And Diagnosis Based On Kernel Principal Component Analysis

With the development of science and technology, and the improvement of system complexity and automation, faults tend to occur more frequently and more severely. Fault detection and diagnosis technique in dynamical systems has become one of the important means to improve system safety and reliability. Data driven fault detection and diagnosis approaches which need neither to establish the system mathematical model nor to require accurate prior knowledge, are researched with real process data which are obtained by information technique. The study on data driven fault detection and diagnosis methodologies has both theoretical and application significance.In this dissertation, based on the novel research achievement, the fault detection and diagnosis design methods based on kernel principal component analysis (KPCA) were deeply considered. Aimed at some existing shortages of KPCA, KPCA was applied with particle swarm optimization (PSO) algorithm, wavelet packet transform and support vector machine (SVM). Respectively, the fault detection and diagnosis approaches of kernel principal component analysis (KPCA) based on particle swarm optimization (PSO), KPCA based on wavelet packet transform, KPCA and SVM multi-classification were proposed. And the simulation researches and the result analysis have been done and discussed by Tennessee-Eastman (TE) model. The main achievement of this dissertation can be summarized as follows:1. The kernel function parameter optimization method based on particle swarm optimization algorithm was proposed for the performance of KPCA influenced by its kernel function parameters. The particle swarm optimization algorithm was applied to optimize by constructing the kernel function parameter optimization model, and the numerical simulation results were presented to validate the proposed approaches. The fault condition recognition of KPCA based on PSO was researched in simulation by TE model.2. The basic principle of KPCA and monitoring statistics SPE and T2 were studied, and the process algorithm of fault detection based on KPCA was presented. Some numerical simulation results were presented to validate the approach by TE model. On the basis of feature extraction based on KPCA optimized by PSO algorithm, combined with fault identification approach based on contribution diagram, the fault detection and diagnosis method of KPCA based on PSO was proposed in order to reduce the blindness of kernel function parameter setting and improve the efficiency of fault detection and diagnosis. By the simulating researches by TE model, it has noted that the changes of statistics SPE and T2 have been used to determine whether fault happened and the contributions to statistics of each variable were used to identify the fault source for fault detection and diagnosis.3. Aimed at process data containing noises, the wavelet packet denoising algorithm was presented by studying the basic theory of wavelets and wavelet packet. Integrated with the fault detection approach of KPCA based on PSO, the fault detection method of KPCA based on the wavelet packet denoising was investigated. The simulation results by TE model have showed that the proposed method can effectively reduce the numbers of the nonlinear principal component, smooth fault monitoring curves and improve fault monitoring.4. By researching the basic principle of SVM and the SVM multi-classification and applied on the KPCA fault diagnosis, the new fault diagnosis approach based on KPCA and SVM multi-classification was presented. It was proven that the fault monitoring was realized by fault detection based on KPCA and the fault classification was obtained by training SVM multi-classification to identify fault. The numerical simulation results have been presented to validate the proposed approaches.