Adaptive multi-scale heterogeneous graph learning for real-time anomaly detection in industrial cyber-physical systems

Abstract

The proliferation of IIoT devices introduces complex security vulnerabilities in cyber-physical systems. We propose AMS-HGNN, an Adaptive Multi-Scale Heterogeneous Graph Neural Network for real-time industrial anomaly detection. Our architecture features a Scale-Adaptive Graph Convolution mechanism, a DomainBridge Fusion module correlating cyber and physical entities, and a Lightweight Edge-Optimized design achieving sub-15ms inference on ARM devices. Evaluated on three industrial datasets (2.3M+ events), AMS-HGNN achieves an Fl-score of 0.938 on the SWaT benchmark - a 6.8% improvement over existing methods - while reducing computational overhead by 42%. Results confirm adaptive multi-scale learning as a viable pathway for deployable industrial security monitoring.