Anomaly Detection and Root Cause Identification in Multistage Industrial Processes via Causal Graph-Based Spatio-Temporal Learning

With the increasing integration of industrial production systems, multi-stage processes become more complex and exhibit strong stage heterogeneity and inter-stage error propagation, which limits conventional approaches for anomaly detection and root cause analysis. This paper proposes a causal graph neural framework for anomaly detection and root cause analysis in multi-stage industrial processes. The process is modeled as a hierarchical causal structure composed of a global graph and stagelevel subgraphs which are intended for interpretable fault propagation tracing under explicit process assumptions. Process prior knowledge is encoded as feasibility gates to constrain candidate edges, upon which learnable masks select effective connections. Global graph captures inter-stage causal dependencies, while masked subgraph learns intra-stage dynamic relationships. Then, the spatio-temporal layers with temporal encoding and graph convolution are employed for representation learning and forecasting. Residual-based anomaly detection then activates a hierarchical diagnosis pipeline, including stage-level localization via the global graph, intra-stage attribution of key parameters via dynamic subgraphs, and fault propagation path inference. Experiments on a multistage industrial dataset show that the proposed method demonstrates the feasibility of residual-based anomaly triggering and structured root-cause tracing and provides a structured and interpretable root cause analysis solution for complex manufacturing systems.