Dynamic Graph Neural Network for Real-Time Failure Prediction in Industrial IoT Networks

Abstract

Industrial Internet of Things (IIoT) networks support mission-critical applications in manufacturing, energy, and process automation, where even a single link failure can trigger cascading failures with severe safety and operational consequences. Predicting failures before they occur requires models that can capture both the spatial dependencies between network components and the temporal evolution of link conditions. However, existing spatio-temporal graph neural networks are primarily designed for node-level prediction, often process spatial and temporal information separately, and rarely provide uncertainty estimates that operators can use to assess the reliability of predictions. This paper presents DGNN, a dynamic graph neural network for proactive link failure prediction in IIoT networks. The proposed framework combines a Gated Graph Convolutional Unit to model spatial interactions with a Temporal Attention Module to learn degradation patterns from sequential telemetry data. A multi-task learning strategy jointly performs link failure classification and link reliability regression, while MC dropout is employed to estimate predictive uncertainty, enabling risk-aware maintenance decisions based on both prediction and confidence. Experiments conducted on a graph-based synthetic IIoT dataset with realistic failure cascades and measurement noise show that DGNN achieves an F1-score of 0.911, an AUC-ROC of 0.950, and an R2 score of 0.854 for link reliability regression. The framework also satisfies real-time IIoT requirements, with an average inference latency of 5.66 ms per sample, well below the 10 ms latency budget. Furthermore, performance remains stable under severe telemetry degradation, maintaining an F1-score above 0.84 across signal-to-noise ratios ranging from 30 dB to 0 dB.