Remaining Useful Life Prediction of Bearings Using Statistical Distributions Integrated into Neural Networks

키워드

확장요약문

Accurate remaining useful life (RUL) estimation of rolling bearings under varying operating conditions is essential for predictive maintenance. Changes in rotational speed and load alter degradation behavior and reduce the generalization of purely data-driven models. This study proposes a physics-informed neural network that integrates lifetime knowledge from the Weibull Accelerated Failure Time Regression (WAFTR) model into a Multi-Layer Perceptron (MLP) Network. It is evaluated on the PRONOSTIA dataset with multiple speed and load combinations. Three approaches are compared: WAFTR, a data-driven MLP, and the proposed physics-informed WFMLP. Models are evaluated with Leave-One-Out (LOO) and Leave-One- Condition-Out (LOCO) validation. LOCO excludes all bearings sharing the test bearing's condition, a stricter test of generalization. Performance is measured by MAE, RMSE, and MAPE. Learning-based models generally outperform WAFTR, especially under LOCO, and WFMLP further improves robustness in MAE and RMSE, showing that WAFTR-based physical knowledge supports reliable RUL estimation for unseen operating conditions.