Trustworthy Predictive Maintenance: Physics-Informed XAI for Probabilistic RUL Estimation

The transition to Industry 5.0 requires predictive paradigms that transform opaque automated systems into resilient, trustworthy frameworks. In Predictive Maintenance (PdM), accurately estimating the Remaining Useful Life (RUL) of industrial assets under varying operating conditions remains a significant challenge. While deep learning models have achieved high accuracy, they typically yield deterministic point estimates and often neglect the inherent uncertainty introduced by sensor noise and the physical laws governing mechanical degradation. To address these limitations, this work introduces the Autoregressive Uncertainty and Risk-Aware architecture (AURA). The proposed architecture integrates a non-linear temporal encoder with a linear autoregressive skip-connection, culminating in a probabilistic head trained via maximum likelihood estimation. Rigorous evaluation is ensured through dynamic sequence packing, which enables comprehensive assessment of the entire dataset without data leakage. Additionally, an Explainable Artificial Intelligence (XAI) attention mechanism with strict zero-padding masks is implemented to facilitate transparent decision-making. The network is further constrained by a physics-informed loss function that penalizes RUL overestimation, thereby enforcing the cumulative nature of component wear. On the complex C-MAPSS FD004 dataset, which comprises six distinct operating conditions and dual fault modes, the model achieves a robust expected-value RMSE of 27.57 cycles. By delivering reliable $9 5 %$ confidence intervals, interpretable attention heatmaps, and physically consistent degradation curves, AURA establishes a robust framework for trustworthy prognostics in safety-critical environments.