Process Fault Prognosis Using Hidden Markov Model–Bayesian Networks Hybrid Model

Abstract

A Hidden Markov Model–Bayesian Networks (HMM–BN) hybrid system coupled with a novel prediction technique is employed to predict and isolate 10 identified faults in the Tennessee Eastman (TE) process. HMM is trained offline with Normal Operating Condition data and then used to train the BN. Utilizing the same trained HMM, a history of Log Likelihood (LL) values of process fault data is generated. The same trained HMM determines the LL values of online data strings and compares with the LL history and predicts the most likely future state of the system. This information is then fed to BN as likelihood evidence to isolate the root cause. The system successfully predicts all the selected 10 faults of the TE process while accurately isolating 8 of them. The maximum level of noise that can be handled is presented along with the respective result. These results set the benchmark for future prediction and isolation studies of the TE process.