(588c) Discovery of Cyclic Loops in Bayesian Network for Root Cause Diagnosis of Process Faults | AIChE

(588c) Discovery of Cyclic Loops in Bayesian Network for Root Cause Diagnosis of Process Faults


Shah, P. - Presenter, Texas A&M University
Wang, Q., Texas A&M University
Kumari, P., Texas A&M University
Khan, F., Memorial University of Newfoundland
Kwon, J., Texas A&M University
In the chemical process industry, root cause diagnosis of process faults is highly crucial for providing efficient troubleshooting. Failed to diagnose and act promptly, process faults can progress into high-consequence events. Bayesian-based probabilistic models are widely for fault diagnosis purpose as these models represent causality better for root cause analysis [1]. In this method, firstly, a causal network of process variables, i.e., a Bayesian network (BN), is constructed that represents causality in the process. Then, BN is updated using alarm data to diagnose the root cause. Among the data-driven approaches to construct BN, score and search algorithms are one of the most effective methods. In these algorithms, a search algorithm is utilized to create a search space of candidate networks whose fitness to data is measured by a scoring function [2]. However, these existing approaches are not equipped to effectively discover cyclic loops in BN [3]. Since cyclic loops are prevalent in chemical processes due to material and heat integration, recycle streams, feedback control, and coupling among process variables, unaccountability for these cyclic loops results in an inaccurate BN, and thereby reducing the diagnosis accuracy [4]. Therefore, it becomes important to find cyclic loops into BN for accurate diagnosis.

To this end, a direct transfer entropy (DTE)-based multiblock BN is proposed to discover cyclic loops in BN. First, the process is segmented into multiple blocks. Next, block-level BNs are learned using the DTE-based score and the Greedy search algorithm. DTE measures information transfer in the network by accounting for the effects of common source variables on parent and child variables. In doing so, DTE can distinguish if there exist direct or indirect causal relations between the parent and child variables, which was impossible to be done by the existing scores such as transfer entropy. Therefore, the DTE score provides a precise quantification of the fit of candidate networks to data, and thus aids in obtaining accurate block-level BNs. Although block-level BNs do not have cyclic loops, their fusion results in the discovery of cyclic loops. In this manner, the multiblock approach aids in the inclusion of significant cyclic loops into the BN for the process. Finally, the performance of the proposed methodology is demonstrated through a case study of the industrial benchmark Tennessee Eastman process.

Keywords: cyclic loop, Bayesian network, root cause diagnosis, direct transfer entropy


  1. Kalantarnia M., Khan, F., and Hawboldt, K., 2009. Dynamic risk assessment using failure assessment and Bayesian theory, Journal of Loss Prevention in the Process Industries 22, 5, 600-606.
  2. Scutari, M., Graafland, C. E., and Gutierrez, J. M., 2019. Who learns better Bayesian network structures: Accuracy and speed of structure learning algorithms, International Journal of Approximate Reasoning 115, 235-253.
  3. 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.
  4. Gharahbagheri, H., Imtiaz, S.A., & Khan, F., 2017. Root cause diagnosis of process fault using KPCA and Bayesian Network. Industrial & Engineering Chemistry Research, 56, 2054-2070.