Efficient Process Monitoring and Fault Isolation with the Integrated use of Markov Random Fields Learning and the Graphical Lasso

Abstract

Abstract In this study, a framework for efficiently detecting faults, and isolating fault causes with the integrated use of Markov Random Fields (MRF) and the graphical lasso, is proposed. MRFs are capable of expressing any subtle relationship that appears within a process without any a priori knowledge, and can compute the joint probabilities of the variables, taking in the nonlinearity of chemical processes. The graphical lasso, a regularization algorithm for inducing the sparse precision matrix and learning the structure of a graphical model, mitigates the high computational complexity of learning and inferencing a MRF. Also, by splitting the process variables into inter-relevant groups, it enables quick identification of the fault variable, and fault propagation path detection. The proposed algorithm was applied to the Tennessee Eastman benchmark process (TEP), to test its performance. The 28 fault cases within the TEP model were tested with the new monitoring method, and the results were compared to that of the conventional methods. The proposed method showed high accuracy and efficient fault identification performance, even with the tricky fault cases 4 and 9.