Process fault diagnosis via the integrated use of graphical lasso and Markov random fields learning & inference

Abstract

In this study, a novel methodology for process fault diagnosis is proposed, where the monitored variables are modelled into pairwise Markov random fields (MRFs), and the conditional contribution values are calculated for each node, with respect to the occurring fault. First the monitored variables are modelled into a MRF framework, and the parameters of the MRF are learned using normal process data. Then when a fault occurs, the conditional marginal probability of each of the variables are obtained by using the kernel belief propagation (KBP) method, which is converted into the conditional contribution value for fault diagnosis. Compared to state-of-the art fault diagnosis methods, the proposed methodology successfully detected the root cause nodes for all of the fault types, as well as allowing detailed analysis of the characteristic of the fault. Also, the propagation paths of faults were detectable according to the conditional contribution plots.