(305f) A Novel Dynamic Bayesian Network Approach for Networked Process Monitoring and Fault Propagation Diagnosis

Bayesian networks are graphical models for performing probabilistic inference that can be used for a large number of problems under systematic uncertainty such as reasoning, learning causality analysis, etc. Bayesian networks provide graphical structures that effectively characterize the complex casual relationships among different variables for evaluating system uncertainty using probabilistic inference. They are directed acyclic graphs consisting of a set of nodes that represent variables with a finite set of states. The arcs or edges connecting the various nodes represent the probabilistic causal dependence among the variables.

A novel networked process monitoring and fault propagation diagnosis method is developed in this study. Dynamic Bayesian networks (DBN) are used to monitor and diagnose abnormal events in chemical processes as well as to model the propagation of faults throughout the processes. Despite plenty of literature research on process monitoring, inadequate attention has been paid on diagnosing the root causes of abnormal events when multiple potential sources of faults exist. The proposed process monitoring scheme integrates preliminary process knowledge with operational data to identify the root causes and diagnose the propagation of process faults. The preliminary process knowledge is used to map process flow diagrams into the Bayesian network structure and the DBN is then trained using plant historical data. Due to the dynamic and stochastic natures of chemical processes, a probabilistic analysis technique based on Bayesian inference is developed to compute the likelihood of abnormality of process variables for fault identification. The trained network model is also used to determine the fault propagation throughout the processes due to intricate variable interactions and thus diagnose the root causes of process faults. The presence of uncertainty is further quantified by the node relationships and the adaptive model parameters. The identified fault propagation sequence provides a better understanding as to the abnormal process operation and the interactive effects of faults throughout the processes.

The proposed approach is demonstrated using the Tennessee Eastman Chemical process and the fault identification and diagnosis performance is evaluated against conventional principal component analysis (PCA) and independent component analysis (ICA) based variable contribution plots. The fault identification and fault propagation diagnosis capability of the proposed networked process monitoring approach is proven through accurate fault detection and root-cause diagnose of the abnormal events. Moreover, the propagation pathways of process faults provide a deep understanding regarding the effects of the abnormal event throughout the processes. It is shown that the proposed networked process monitoring and fault propagation diagnosis method accurately identifies the faulty process variables and effectively determines the fault propagation pathways to diagnose the root causes of process faults.