Hidden Markov Model Combined with Kernel Principal Component Analysis for Nonlinear Multimode Process Fault Detection

Data-driven techniques become increasingly popular in the field of industrial fault detection. Regarding the complex nonlinear industrial process accompanied by multiple operational monitoring modes, conventional multivariate monitoring techniques such as kernel principal component analysis (KPCA) are not suitable. In this paper, a novel hidden Markov model (HMM) combined with kernel principal component analysis is proposed for nonlinear multimode process fault detection. Firstly, the HMM is built from the measurement data of different modes so as to estimate the dynamic mode sequence. Furthermore, a local KPCA model is developed to detect the fault of each mode. The effectiveness of the proposed method is shown through a numerical nonlinear multimode simulation example and Tennessee Eastman (TE) Chemical benchmark process. The comparison results demonstrate that the proposed HMM-KPCA method precedes the conventional KPCA method due to the high fault detection rate (FDR) and low false alarm rate (FAR).