Modified independent component analysis and Bayesian network-based two-stage fault diagnosis of process operations

Statistical fault detection techniques are able to detect fault and diagnose root-cause(s) from the monitored process variables. For complex process operations, it is not feasible to screen all the process variables due to monitoring cost and flooding of alarms. Thus, if a fault is originated from a process variable that is not monitored, conventional statistical techniques are incapable of locating the true root-cause. To relax this limitation, a two-stage fault diagnosis technique is proposed for process operations. In the first-stage, the modified independent component analysis is used for fault detection and to identify the faulty monitored variable. In the second-stage, a Bayesian Network model is constructed considering the process variables and their dependence obtained from the process flow diagram. Evidence is then generated at the network node corresponding to the faulty variable identified in the first-stage. Subsequently, the network is updated and analyzed using deductive and abductive reasoning to identify the true root-cause. To verify the applicability of the proposed technique it is tested on two process models. The results of both case studies have demonstrated the effectiveness of the proposed technique to diagnose the true root-cause that originated from process variables that are not monitored. Once integrated with process loss functions, the proposed technique will serve as an important element of dynamic operational risk management framework.