A causal-guided graph transformer with interpretability analysis for process fault diagnosis

Process fault detection and diagnosis (FDD) is essential for keeping chemical processes safe, but many data-driven models are hard to trust because they often rely on patterns that are not truly related to the real causes of faults. To address this, we introduce a causal-guided graph transformer with dual-layer interpretability analysis (CGTIA). It first builds a directed fault-propagation graph using a lag-aware Granger screening method and then reduces noisy or unnecessary connections with a sparsification strategy. Building on this graph, the model separates learning into two parts: a local branch that follows causal directions and a global branch that captures broader interactions and shared operating conditions. In addition, the method also includes a dual-layer interpretability system that highlights which connections matter most and provides stable, sensor-level explanations. Lastly, we validate the approach on a real multiphase flow facility dataset, which is naturally imbalanced and therefore more challenging than typical simulation datasets. Across various faults, the proposed CGTIA framework achieved a macro averaged recall of 99.20% while maintaining robust variable propagation interpretability. Overall, this combination effectively enhances confidence in the diagnostic results and gives clearer insight into how faults behave in the process. • A new method, causal-guided graph transformer with dual-layer interpretability analysis (CGTIA) is introduced • The method builds a directed fault-propagation graph using a lag-aware Granger screening • The method includes a dual-layer interpretability system • The CGTIA framework is validated using a real multiphase flow facility dataset