A Physics-Informed Spatio-Temporal Graph Neural Network with Self-Supervised Representation Learning for Industrial IoT Anomaly Detection

Abstract

Industrial Internet of Things (IIoT) environments consist of interconnected and evolving data, and the complexities of dependence, small labeled anomalies and changes in operating conditions make accurate anomaly detection difficult. The Physics-Informed Spatio-Temporal Graph Neural Network with Self-Supervised Representation Learning (PI-STGNN-SSRL) is proposed in this paper to tackle these challenges. The framework combines physics-informed graph construction, self-supervised representation learning, spatio-temporal graph reasoning and adaptive anomaly scoring, to capture and localize abnormal industrial behaviors. Experimental results show that the proposed method, PI-STGNN-SSRL, outperforms the compared anomaly detection techniques in terms of accuracy, precision, recall, F1-score, and AUROC with values of 98.2%, 97.8%, 97.5%, 97.6% and 99.1% respectively. Root cause localization is complemented by component level anomaly scoring. The results show that the proposed framework is accurate, robust and interpretable anomaly detection in complex IIoT environments.