PHYSICS‑INFORMED RUL PREDICTION WITH ALEATORIC/EPISTEMIC UNCERTAINTY AND MAINTENANCE SYSTEM INTEGRATION

Abstract

Accurate prediction of Remaining Useful Life (RUL) in safety-critical machinery is essential for transitioning from reactive to predictive maintenance strategies. While deep learning approaches have demonstrated empirical progress, they suffer from two fundamental limitations: the inability to extrapolate beyond training distributions and the conflation of fundamentally distinct uncertainty sources. This paper presents a Physics-Informed Neural Network (PINN) framework that integrates thermodynamic degradation models and fatigue crack propagation mechanics directly into the neural network loss function, producing RUL predictions that are both physically consistent and probabilistically calibrated. We decompose predictive uncertainty into its aleatoric component — irreducible noise arising from sensor imprecision and environmental stochasticity — and its epistemic component, quantifying reducible model uncertainty via Monte Carlo Dropout. The resulting uncertainty bounds are propagated into a cost-optimized maintenance scheduling module that selects intervention timing by minimizing the expected total maintenance cost under the full predictive distribution. Experiments on the NASA CMAPSS turbofan degradation benchmark and a real-world wind turbine gearbox dataset demonstrate that our Full PINN-UQ model achieves an RMSE of 11.6 cycles and MAPE of 7.3%, representing improvements of 59.2% and 62.0% respectively over LSTM baselines, with a calibration score of 0.93 — the highest among all seven models benchmarked. The integrated maintenance planner reduces unnecessary preventive maintenance actions by 34% while maintaining a false-negative rate below 2.1%, demonstrating clear operational value beyond point-prediction accuracy.