Optimized principal component analysis and multi-state Bayesian network integrated method for chemical process monitoring and variable state prediction

Abstract

Considering the weaknesses of traditional principal component analysis (PCA) in dealing with nonlinear correlations and non-Gaussian distribution data, PCA is optimized by replacing covariance matrix with Spearman ranking correlation coefficient (SRCC) matrix and introducing Gaussian transition by Johnson transformation. Because the commonly used BN that simply identifies a node as faulty or normal states sometimes fails to diagnose critical operation information, multi-state Bayesian network (MBN) is developed to recognize a node into multiple states. To fulfill process monitoring task, the optimized PCA (OPCA) and MBN integrated method (OPCA-MBN) is proposed in this paper. OPCA is utilized to detect faults and provide evidence to MBN for diagnosing fault or normal oscillation propagation pathways. In the modeling process of MBN, the causal relationships between tangled internal variables are determined using Transfer entropy and process knowledge. The practicability and effectiveness of the proposed method are demonstrated through the application in the Tennessee Eastman (TE) process in comparison with two-state BN.