Knowledge-Assisted Multi-Graph Enhances Industrial Anomaly Detection

Industrial processes often generate highly complex and interdependent time-series data from numerous sensors across multiple operational stages. Effective monitoring and timely anomaly detection in these multivariate time series (MTAD) are crucial for preventing system failures and ensuring the reliability of automated industrial systems. While Graph Neural Networks (GNNs) have advanced MTAD by modeling complex variable dependencies through data-driven graphs, existing GNN-based methods frequently overlook or struggle to seamlessly integrate critical process knowledge, leading to suboptimal performance. To address this limitation, researchers have introduced a novel knowledge-assisted multi-graph framework. This framework explicitly incorporates domain-specific process knowledge into the graph learning process, significantly enhancing dependency modeling for MTAD in multi-stage industrial environments. The method constructs three complementary graphs: one purely data-driven and two others refined by integrating structural constraints derived directly from process knowledge. A multi-graph attention network is then employed to effectively leverage these diverse graphs, resulting in a more accurate and robust representation of complex dependencies. Comprehensive experiments on two real-world, multi-stage industrial datasets demonstrate that this integration of process knowledge substantially improves anomaly detection performance.