A novel Causality-Constrained Synchronous Spatio-Temporal Graph Convolutional Networks for fault diagnosis in large-scale industrial processes

Complex industrial plants exhibit intricate spatio-temporal dynamics across interconnected units, posing significant challenges for fault diagnosis and interpretability. Existing methods often struggle to simultaneously capture spatio-temporal features while ensuring causal consistency. To address this, this paper proposes a method based on a Causality-Constrained Synchronous Spatio-Temporal Graph Convolutional Network(CC-Sync-STGCN), which integrates prior causal knowledge with synchronous spatio-temporal feature learning to deliver accurate detection and transparent root-cause analysis. First, process knowledge is used to partition the plant into sub-units, and transfer entropy quantifies directional relationships among variables within each sub-unit. These relations inform a synchronous, sparsified spatio-temporal adjacency structure that captures long-range, time-varying dependencies while keeping the model lightweight. A gated learning module jointly extracts spatial and temporal representations and is trained with a penalty that rewards causal consistency, so that learnt features remain physically meaningful. For online monitoring, prediction residuals feed a fused statistic that combines complementary indicators of deviation, enabling reliable alarms. A dual-level analysis, spanning plant-wide localisation and variable-level inspection, then traces the fault origin and propagation pathway for interpretable diagnosis. Case studies on a simulated chemical process and an operating hot-strip mill demonstrate improved detection reliability and clear identification of the initiating variable and its downstream effects, while preserving computational efficiency. The results indicate that embedding causal structure into synchronous graph learning is an effective route to robust, interpretable fault diagnosis in large-scale industrial environments.