Fault Diagnosis and Prediction of Continuous Industrial Processes Based on Hidden Markov Model-Bayesian Network Hybrid Model

Abstract

Hidden Markov models (HMMs) have been recently used for fault detection and prediction in continuous industrial processes; however, the expected maximum (EM) algorithm in the HMM has local optimality problems and cannot accurately find the fault root cause variables in complex industrial processes with high-dimensional data and strong variable coupling. To alleviate this problem, a hidden Markov model-Bayesian network (HMM-BN) hybrid model is proposed to alleviate the local optimum problem in the EM algorithm and diagnose the fault root cause variable. Firstly, the model introduces expert empirical knowledge for constructing BN to accurately diagnose the fault root cause variable. Then, the EM algorithm is improved by sequential and parallel learning to alleviate the initial sensitivity and local optimum problems. Finally, the log-likelihood estimates (LL) calculated by the improved hidden Markov model provide empirical evidence for the BN and give fault detection, prediction, and root cause variable detection results based on information about the similar increasing and decreasing patterns of LL for the training data and the online data. Combining the Tennessee Eastman (TE) process and the continuously stirred tank reactor (CSTR) process, the feasibility and effectiveness of the model are verified. The results show that the model can not only find the fault in time but also find the cause of the fault accurately.