Identification of probabilistic graphical network model for root-cause diagnosis in industrial processes

Identification of faults in process systems can be based purely on measurement (e.g. PCA), or can exploit knowledge of process model structure to construct a causal network. This work introduces a method to identify most likely causal network in cases when process model is not known. An incidence matrix, showing location of measurements in the plant network structure, and historical process data are used to identify the optimal causal network structure by means of maximizing Bayesian scores for alternative causal networks. Causal subnetworks, corresponding to subgraphs of the process network, are identified by finding the most probable graph based on highest posterior probability of graph features computed via Markov Chain Monte Carlo simulation. Novel Bayesian contribution indices within the probabilistic graphical network are proposed to identify the potential root-cause variables. Application to Tennessee Eastman Chemical plant demonstrates that the presented method is significantly more accurate than the current methods.