A Novel Graph Convolutional Network Approach for Anomaly Detection in Industrial IoT Time Series Data with Spatio-Temporal Dependencies

The proliferation of the Industrial Internet of Things (IIoT) has catalyzed the deployment of dense sensor networks monitoring mission-critical physical infrastructure. Detecting anomalies in these complex cyber-physical environments requires identifying subtle deviations within high-dimensional multivariate time series while modeling intricate spatial inter-sensor correlations and temporal dynamics. Existing deep learning approaches often treat sensor dimensions independently or rely on static, pre-defined topology graphs that fail to capture time-varying physical dependencies. In this paper, we propose an Adaptive Spatio-Temporal Graph Convolutional Network (AST-GCN) tailored for unsupervised anomaly detection in IIoT systems. Our architecture integrates a dynamic graph learning mechanism that adaptively infers latent inter-sensor topological structures directly from continuous measurement streams without requiring prior domain graphs. This spatial representation is coupled with stacked dilated causal convolutions to model multi-scale temporal dependencies. By combining a joint forecasting and reconstruction optimization framework, AST-GCN generates robust anomaly scores that discern localized sensor failures from plant-wide systemic disruptions. Extensive empirical evaluations conducted on real-world industrial benchmarks, including the Secure Water Treatment (SWaT) and Water Distribution (WADI) datasets, demonstrate that AST-GCN consistently outperforms state-of-the-art baselines, achieving F1-scores of 0.884 and 0.826, respectively. The results highlight the framework's superior sensitivity, precision, and robustness against severe sensor noise.