Research On Fault Detection Based On Wavelet Packet De-noising And Principle Component Improvement Analysis

Because of the continuing improvement of scientific technology anddevelopment of modern manufacturing industry, the structures of the heavyequipments are designed to be more and more complex, and the functions ofthem are more and more comprehensive as well. However, at them same time, thepossibility of occurring technical faults is also increased and their consequencesare getting more serious. Therefore, the research of fault detection and diagnosistechnology has become one of the most important elements to the system safetyand reliability. An accurate mathematical model of complex industrial process isvery difficult to built, therefore the data-driven fault detection and diagnosismethod is based on analysis of the processing data, and it is accomplished bypicking up the hidden information among the massive data base. This report istargeting on the high complexity, nonlinearity, dynamics, and noise disturbanceissues within the TE process, and to research on the data-driven fault detectionand diagnosis method and its improvements. In addition, the feasibility analysisof it is also conducted from the simulated experiments.First of all, the principal component analysis (PCA) and kernel principalcomponent analysis (KPCA) are comprehensively researched, and the actualalgorithms, as well as the conditions of applying them, are carried out. Thewavelet packet theory is introduced, and by comparing the decomposition andreconfiguration patterns of the wavelet analyses, the advantage of it duringdealing with the frequency conversion detailed information is clearly observed.Secondly, the best basis wavelet packet de noising method is pointed out inorder to solve the noising issues within the actual industrial processing data. Thecollected data is multi-layered decomposed with high and low frequencies by wavelet packets. As a result, the original information is retained maximum andthe noising signals are taken out. Thus the mistaking and failure rates of faultdetection are significantly reduced. In addition, to solve the nonlinear issues ofthe actual system, the kernel principle component analysis (KPCA) method,combined with the best basis wavelet packet de noising concepts, is introduced,and the actual operation procedures are given out as well. This method is thenapplied on the TE process. The Q statistics of PCA, KPCA, and KPCA withwavelet packet de noising are analyzed, and their variance contribution graphsare compared. As a result, it can be concluded that the effectiveness andadvantages of the introduced method are proved.At last, due to the disadvantage of original Principle Component Analysiswhen it is used to deal with nonlinear or dynamic processes, an exponentiallyweighted dynamic autoregressive element was added into the kernel principlecomponent model, and the data with unique interval is updated in real time. Theupdated data is used to build up a new Kernel Principle Element model, with theapplication of weighting factor, a diagnosis tool with dynamic self-adaptivecharacter is founded by both the new and the original models, and the actualprocedures of building the model is listed in this report as well. The effectivenessof this method is again verified by picking up different weighting factors andcomparing with the original KPCA method.