Online Fault Diagnosis of the Tennessee Eastman Process By Bayesian Network

Process disturbances that lead to changes in the physicochemical conditions of process unit, even under control, are considered as a class of fault in chemical industries. Fault diagnosis is an important issue in control and process engineering. It provides benefit of plant’s loss prevention and safe operation. One challenges of fault diagnosis in real situation is uncertainty. Probability reasoning is an approach to overcome it, and Bayesian network is a powerful tool in this regard. In present work, basic Bayesian network is developed by using sensor signals and manipulating variable (opening percentage of control valve) of Tennessee Eastman process to represent its state. Network consists of two layers which represent faults classes and symptoms. Structure and parameter learning is based on historical database and expert knowledge. Updating of basic network based on online information is a withstanding improvement. The main feature of the proposed framework is sequential model selection based on recent received signals. It would be obtain through calculating posterior distributions using conjugate prior distributions and then maximizing log_likehood. Therefore, this novel approach would provide reinforcement learning. In other word, new information about system would modify belief of conditional (in)dependency between fault and symptoms and improve faults diagnosis in inference stage.