The Study Of Fault Detection And Diagnosis Based On MSPC

With the rapid development of modern industry and technology, industry structures become more and more complex. Consequently, the factors that could bring on much more potential failures increase, so it is a great challenge to ensure the reliability and security of industry system. For the sake of detecting the abnormal events and diagnosing faults, it is very important to monitor the industry process by multivariate statistical process control.According to the status of the industry applications of industrial process fault diagnosis technology at home and abroad, this dissertation emphasizes the methods of principal component analysis and partial least squares, and improves the methods based on principal component analysis(PCA) and discriminant partial least squares(DPLS). The improved methods are described as follows:1. When some faults occur in industry process, statistics of PCA model can't provide the process variational information effectively, so Q statistic is divided into principal component related variable residuals and common variable residuals, which have a better performance.2. Original signal is decomposed by wavelet in different scales, the wavelet decomposition coefficients of the real signal are held, and the wavelet decomposition coefficients of the noise are eliminated, then the signal is reconstructed by inverse wavelet transform. Kernel PCA can eliminate the relativity of variables and extract the fault information better, the feature information of the pretreatment datum is obtained by KPCA, and the performance of fault detection is improved.3. The variation information uncorrelated to dependent variables is removed from independent variables by orthogonal signal correction, the method combining discriminant partial least squares with orthogonal signal correction is introduced. This method reduces complexity and false alarm rate greatly, and improves the performance of fault diagnosis.The methods the above are proved to be feasible and effective by application in TE process, and have a better performance over PCA model and DPLS model respectively, such as false alarm rate,missing alarm rate.