Comparative study of Bayesian Network-based root cause analysis methods for chemical and bioprocess systems

Root cause analysis (RCA) is critical in chemical and bioprocess industries for identifying the underlying causes of process faults. Bayesian network (BN)-based methods are widely used due to their ability to capture causal dependencies under uncertainty. Commonly, a two-stage fault detection and diagnosis-Bayesian network (FDD-BN) framework is employed: the first stage applies FDD methods to detect faults and identify symptom variables, and the second stage uses BN to perform RCA based on the identified symptoms. However, existing studies largely rely on FDD methods that primarily capture linear interactions between variables and on expert-defined BN structure, limiting their applicability to complex industrial processes. In addition, they typically consider only a few selected fault scenarios within a given benchmark and lack a systematic comparison. To address these methodological gaps, we conduct a comparative study investigating alternative data-driven methods for each stage of the FDD-BN framework, including traditional multivariate statistical methods and machine learning-based methods for FDD, as well as various causal discovery algorithms to learn BN structures without prior knowledge. The proposed methods are evaluated using two widely used industrial benchmark datasets: the Tennessee Eastman Process and the Industrial-scale Penicillin Simulation, with all fault scenarios included. The results indicate that machine learning-based FDD methods offer more robust and consistent performance across both datasets. For causal discovery, algorithm performance differs across datasets, reflecting sensitivity to differences in underlying process characteristics.