Improved MSPCA with application to process monitoring

Abstract

This paper proposes a improved Multi-scale Principal Component Analysis (MSPCA). A key problem of fault detection in process monitoring lies in how to enhance the accuracy of fault detection to reduce detection costs. In this point, MSPCA has improved to a great degree, but it can be improved in the accuracy of selecting the coefficients of wavelet used in reconstructing. Because the accuracy of selecting the coefficients of wavelet is directly concerned with the accuracy of fault detection in using Principal Component Analysis (PCA) reconstructed. When selecting the coefficients of wavelet based on selecting the coefficients of wavelet in using MSPCA. The improved MSPCA should be proposed to detect the fault in process monitoring. It utilizes Principal-component-related Variable Residuals (PVR) statistic and Common Variable Residuals (CVR) statistic at different scales to replace the statistic Q and combine them with the statistic T2 to select the coefficients of wavelet. According to the analysis of simulation of algorithm's example, and comparing the improved MSPCA with MSPCA and conventional PCA, it shows that the improved MSPCA has enhanced the accuracy of fault detection in process monitoring.