Concurrent Multilayer Fault Monitoring with Nonlinear Latent Variable Regression

Abstract

Latent variable regression (LVR) constructs the latent structure by maximizing the projection of quality variables on the latent spaces of the process variables, which has demonstrated its effectiveness over other multivariate statistical methods in terms of quality-relevant monitoring. However, it cannot be applied to model most real-world industrial processes due to its implicit linear assumption on the data. To overcome this issue, the kernel trick is designed to develop its nonlinear counterpart. Further, a novel concurrent kernel LVR (CKLVR) method is proposed for multilayer modeling with subsequent decomposition, which separates the process and data spaces comprehensively into five subspaces. The corresponding monitoring statistics are also developed for each subspace extracted by CKLVR, with each serving a particular monitoring purpose. Case studies including a numerical simulation, Dow’s Refining Process, and the Tennessee Eastman Process are employed to illustrate the performance of the proposed algorithm.