Key-performance-indicator-related state monitoring based on kernel canonical correlation analysis

Abstract

Abstract As a multivariate statistical analysis method, canonical correlation analysis (CCA) performs well for state monitoring of linear processes, but most industrial processes are nonlinear. To solve this problem, kernel canonical correlation analysis (KCCA) has been adopted; however, KCCA still has key performance indicators (KPI)-related issue. In this paper, two improved KCCA methods are proposed to deal with KPI-related issue. One is performing singular value decomposition (SVD) on the correlation coefficient matrix, then the kernel matrix can be divided into KPI-related and KPI-unrelated parts. Another one is performing general singular value decomposition (GSVD) on two coefficient matrices. In addition, this paper also performs fault detectability analysis and computational complexity analysis on these two methods. Finally, the Tennessee Eastman (TE) process is used in this study to verify the efficacy of these two proposed methods.