Fault information-aided principal component subspace construction for efficient small fault detection for chemical processes

Abstract

Principal component analysis (PCA) is widely used for fault detection for chemical processes; however, the efficient principal component (PC) selection remains an challenge. The effect of PC selection on PCA-based fault detection performance is analyzed within the statistical framework of hypothesis testing. A performance-driven fault-relevant PC (FRPC) subspace construction integrated with Bayesian fusion method for efficient chemical process monitoring is then proposed. First, the FRPC subspace is constructed through genetic algorithm-based performance-driven FRPC selection, which achieves the best possible fault detection results for each fault from the PC selection aspect. Second, process measurements are examined in each FRPC subspace as well as the residual subspace. Then the Bayesian fusion is employed to integrate fault detection results from all subspaces. The proposed fault detection approach is tested on a simulated numerical example and the Tennessee Eastman process. The efficiency of the proposed approach is demonstrated.