Knowledge-Informed Causal Graph Deviation Network for Industrial Anomaly Detection and Root Cause Localization

Most industrial anomaly detection and localization methods rely on correlations rather than causality, which limits both the accuracy of anomaly detection and the reliability of root cause localization. Moreover, data-driven causal discovery of industrial processes presents substantial challenges. To this end, we propose a knowledge-informed causal graph deviation network (KCGDN), which uncovers a causal structure from varying correlations by invariant graph learning. The KCGDN architecture integrates two components: a spatial self-attention mechanism (SSAM) and a spatial-temporal prediction network (STPN). The SSAM learns a dynamic graph, while the STPN predicts system behavior under normal conditions, producing significant prediction errors when anomalies occur. To train the KCGDN effectively, a three-phase training algorithm is designed. In the first phase, both the SSAM and STPN are pretrained to capture the varying correlations between variables with a dynamic graph. In the second phase, the causal graph is extracted by searching for an invariant graph structure from the dynamic graph and incorporating constraints from prior knowledge. In the third phase, the STPN is fine-tuned using the extracted causal graph. Finally, we validate the effectiveness of KCGDN through extensive experiments on the DAMADICS benchmark and a real-world argon distillation process of the air separation system.