Fault detection and diagnosis using Bayesian network model combining mechanism correlation analysis and process data: Application to unmonitored root cause variables type faults

Abstract

Risks in chemical plants can generally be divided into Black Swan incidents and Gray Rhino incidents. Black Swan events are unexpected and have a significant impact. Frequently, a large number of Black Swan events cause a huge grey rhinoceros event such as composition-related variables, which becomes a challenge for root cause diagnosis. To alleviate this problem, a strong relevant mechanism Bayesian network (SRMBN) combining mechanism correlation analysis and process state transition is proposed for fault detection and diagnosis. First, a strong relevant mechanism structure of SRMBN is constructed by combining process mechanism analysis with historical data mining for structure learning. Then SRMBN is built after conducting variable state transition and maximum likelihood estimation for parameter learning. For fault detection, a process state based Bayesian interval estimated index is developed by state transformation. Finally, Bayesian contribution index is defined to measure the contribution of each variable to the process state deviation for fault diagnosis. The variables with large index values are added as deterministic evidence to the SRMBN to update the posterior probabilities of nodes for fault propagation inference. The proposed method is applied to faults 2 and 8 (unmonitored type faults) of the Tennessee Eastman (TE) process in comparison with some other published methods. The results show its practicability and satisfactory performance in recognizing the fault propagation pathways and the root causes of faults. Meanwhile, it can provide reliability analysis for process safety and risk assessment.