Explainable Causal Graph‐based Method for Diagnosis and Root Cause Analysis in Process Systems

Abstract

The growing adoption of automation and intelligent maintenance has intensified the complexity of industrial process systems, characterized by high-dimensional nonlinear dynamics and tightly coupled multivariate interactions. This evolution exacerbates cross-level fault propagation patterns, necessitating precise anomaly pathway identification and root cause diagnosis for improving operational safety and reliability. This study proposes an interpretable root cause tracing framework for process systems based on causal graphs. A single-variable equivalent intervention evaluation model is developed to effectively screen critical fault indicators and implement extended convergent cross mapping for causal relationship extraction and abnormal causal graph modelling. In addition, redundancy detection and pruning strategies are designed to comprehensively reconstruct anomaly propagation pathways. Finally, the discriminant criteria of root cause variables are constructed through anomaly triggering and propagation analysis. A root cause variable traceability algorithm is proposed to improve diagnostic accuracy and mechanistic interpretability. Validated through the Tennessee Eastman process, the proposed method demonstrates superior performance in root cause identification and propagation path recognition within tightly coupled systems compared to conventional traceability methodologies.