AMIN-TGAT-PINN: An Explainable Physics-Informed Temporal Graph Attention Framework for Remaining Useful Life Prediction in Industrial IoT Systems

Abstract

Accurate estimation of Remaining Useful Life (RUL) in Industrial Internet of Things (IIoT) environment is a crucial challenge where sensors come in different types, operating conditions are dynamic and machines interact with each other. In this paper, a physics-informed temporal graph attention framework, which is explainable, is proposed for RUL prediction in industrial systems, namely AMIN-TGAT-PINN. The proposed scheme combines Adaptive Multimodal Invariant Normalization (AMIN) to account for variations and similarity among modalities, Temporal Graph Attention Networks (TGAT) to model dynamics along the temporal dimension and between modalities, and Physics-Informed Neural Networks (PINN) to enforce degradation dynamics and operational constraints during learning. Moreover, the Explainable Artificial Intelligence module based on SHAP delivers clear maintenance advise, highlighted by the important contributions of the sensors. Experimental results on benchmark industrial prognostics datasets are shown that demonstrate superior performance in terms of prediction accuracy (98.84%), precision (97.92%), recall (98.17%), F1 score (98.04%) and AUC (0.991) and the reduction of RUL estimation error by 32.6% when compared to state-of-the-art deep learning approaches. The results validate the framework's effectiveness, robustness and interpretability of the framework for future Industry 5.0 predictive maintenance solutions.