Integrate weighted dependence and skewness based multiblock principal component analysis with Bayesian inference for large-scale process monitoring

Abstract

In a large-scale process, variables are correlated, and are Gaussian and non-Gaussian coexisted. The decentralized monitoring methods divide variables into blocks and carry out local monitoring is aimed at reducing the process complexity. Recently, the weighted copula-correlation-based multiblock principal component analysis (WCMBPCA) considering the correlation degree and correlation patterns was proposed to monitor a correlated process. While its monitoring efficiency has been shown, the WCMBPCA method has some drawbacks: 1) monitoring performance relies on the selection of copula function; 2) computation time for the weight matrix is long; 3) without considering the distribution information. This study proposes a novel weight strategy for multiblock PCA by developing a measurement, which simultaneously estimates the dependence and skewness of data, such that the distribution information is additionally considered. The proposed weight matrix is based on the use of non-parametric ranks, prompting the short computation time. The experiment results from Tennessee Eastman (TE) process monitoring shows the average fault detection rate for the proposed method is 84.19% that outperforms regular PCA, dynamic PCA, multiblock PCA, and WCMBPCA. Moreover, the weight matrix calculation takes 0.000635 s for the proposed method, while 195.040282 s for WCMBPCA.