Physics-Informed Machine Learning for Predictive Reliability Assessment of Smart Engineering Systems

Abstract

Smart engineering systems instrumented with dense sensor networks of rotating machinery, aerospace propulsion, energy storage, and power-grid assets, among others, generate high-volume condition-monitoring data, but purely data-driven machine learning models can lead to physically inconsistent predictions, limited extrapolations from the training distribution, and brittle performance in the face of distribution shifts. On the other hand, purely physics-based degradation models are interpretable and safe for extrapolation but rely on several simplifying assumptions, leading to inaccurate modelling of complex, multi-physics degradation mechanisms. This paper proposes and assesses a physics-informed machine learning (PIML) framework to learn a prognostic function for predictive reliability assessment, while systematically reviewing 83 papers at the intersection of physics-informed learning and prognostics and health management (PHM). In the proposed hybrid PINN–LSTM architecture, we use two loss terms to learn the physics of degradation: a data-fidelity loss and a physics-residual loss, with an empirical degradation ordinary differential equation whose RUL output is propagated into the reliability function using the Weibull distribution. The proposed model outperformed traditional LSTM and ARIMA baselines, yielding a root-mean-square error of 11.6 cycles compared with 21.7 and 38.4 cycles, respectively, on a C-MAPSS-type turbofan degradation benchmark; it also produced an asymmetric prognostic score on the PHM08 benchmark that was about 2.6 times lower than the conventional LSTM. Importantly, outside the observation horizon, the physics-informed model's predictive uncertainty increased at a significantly slower rate than the purely data-driven model, and its implied reliability function was much closer to the empirical fleet run-to-failure function. The results show that even a simplified physical degradation law, used as a soft constraint, can significantly improve point-prediction performance and long-horizon extrapolation confidence, with applications to condition-based and predictive maintenance of safety-critical smart systems. Limitations and directions for reliability/uncertainty aware extensions and computational trade-offs are discussed.