A hybrid expert-data causal Bayesian network approach for fault diagnosis in complex industrial processes
Year: 2026
Type: article
Abstract: Fault diagnosis and root cause analysis are critical challenges in modern manufacturing, where increasing process complexity makes it difficult to identify variables responsible for observed anomalies. Data-driven approaches achieve strong predictive performance but rely on statistical associations, making them unable to distinguish causal relationships from spurious correlations induced by unobserved confounders, a critical limitation when the goal is causal attribution rather than prediction. This paper proposes a Causal Bayesian Network (CBN)-based framework addressing the full pipeline from structure learning to causal inference. Structure learning is performed using the proposed DrBO-LC algorithm (DAG Recovery via Bayesian Optimisation with Latent Confounders), which relaxes the causal sufficiency assumption by accounting for hidden confounders while using Bayesian optimisation to handle large search spaces. Hyperparameter tuning is carried out through a two-stage design of experiment combining a Plackett–Burman screening design with a Box–Behnken response surface design. Parameter learning employs a hybrid model accommodating both continuous and discrete variables, reflecting the heterogeneous nature of industrial process data. Root cause analysis is performed via do-calculus-based interventional simulation, quantifying the causal influence of each variable through interventional lift, specificity, interaction, and sensitivity analyses. The framework is evaluated on a benchmark dataset for structure learning under latent confounding and on the Tennessee Eastman Process for the full fault diagnosis pipeline. Results demonstrate superior diagnostic performance and interpretability compared to standard baselines, confirming the suitability of the proposed approach for industrial fault diagnosis.
Source: Expert Systems with Applications
Institution: Institut Supérieur de Mécanique de Paris
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,500
OpenAlex ID: W7222786154