(345g) A Modified Bayesian Network to Handle Cyclic Causal Network in Root Cause Diagnosis of Rare Events

In the chemical process industry, rare events are low-frequency high-consequence events caused by process disturbances. To reduce the severity of consequences, identification of the root causes of rare events is of great importance to provide an approach for effective troubleshooting. Due to the difficulty in developing accurate first-principle models, data-based techniques are widely used for root cause diagnosis in the chemical process industry. Specifically, Bayesian model is popularly used for root cause diagnosis of rare events [1]. In this method, firstly, a Bayesian network (BN) is constructed offline to represent the causal relations among process variables; then, it is updated online using discrete alarm data to diagnose the root cause [2]. However, due to acyclic nature of BN, the existing BN-based methods do not account for cyclic causal relationship of process variables, which are prevalent in chemical processes due to material and heat integration, recycle streams, feedback control, and coupling among process variables [3, 4]. Unaccountability for these cyclic loops in the existing BN-based methods leads to inaccurate root cause diagnosis. To address this limitation, this work proposes a novel modified BN framework to handle cyclic causal relationship of process variables in root cause diagnosis of rare events.
To handle a cyclic causal network, the proposed method first identifies the weakest causal relation in the cyclic loop using a score based on transfer entropy (TE). Here, the TE score quantifies the strength of a causal relation using the amount of information flow from parent variables to child variables that are connected by a causal relation. Then, the weakest causal relation is converted into a temporal relation where the parent variable at the present time step directly affects the child variable at the subsequent time instant. Specifically, for a slowly evolving rare event, this conversion of the weakest causal relation into a temporal relation is valid as it takes some time for an effect to be realized when the causal relation is weak. Since one of the causal relations of cyclic loop is converted into a temporal relation, the cyclic causal network is decomposed into an acyclic network throughout the duration of a rare event. As a result, the obtained acyclic network can effectively account for the cyclic causal relationships among process variables, which improves the accuracy of root cause diagnosis using BN. The performance of the proposed modified BN framework was demonstrated through a case study of the industrial benchmark Tennessee Eastman process.

Keywords: cyclic causal network, Bayesian network, root cause diagnosis, transfer entropy

References:
1. Hu, J., Yi, Y., 2016. A two-level intelligent alarm management framework for process safety. Safety Science 82, 432-444
2. Hu, J., Zhang, L., Cai, Z. and Wang, Y., 2015. An intelligent fault diagnosis system for process plant using a functional HAZOP and DBN integrated methodology. Engineering Applications of Artificial Intelligence, 45, pp.119-135.
3. Suresh, R., Sivaram, A., Venkatasubramanian, V., 2019. A hierarchical approach for causal modeling of process systems. Computers and Chemical Engineering 123, 170-183.
4. Zhu, Q.X., Ding, W.J., He, Y.L., 2020. Novel multimodule Bayesian network with cyclic structures for root cause analysis: Application to complex chemical processes. Industrial & Engineering Chemistry Research 59, 12812-12821.