Wear and life predictions for bearings considering simulation-to-reality variability

Abstract

Bearing wear in mechanical systems often remains unmeasurable, establishing physical simulations as the primary method for investigating wear mechanisms and remaining useful life (RUL). However, discrepancies often emerge between simulated wear and real-world observations, even under identical lubrication and operational conditions. This phenomenon, termed simulation-to-reality variability (StRV), undermines the accuracy of simulation-based wear and RUL predictions. To address this challenge, we propose a hybrid framework that characterizes StRV as an uncertainty source and incorporates stochastic processes to enhance predictive robustness. The framework utilizes multiple nonlinear autoregressive exogenous models (NARXs) to distribute uncertainty in simulated wear volumes. Additionally, a state-dependent Wiener process, induced by a neural network, is formulated to model the dynamic evolution of bearing wear. By introducing a stochastic parameter and neural network modeling, the method accounts for inherent uncertainties while leveraging data-driven insights to infer wear patterns. This approach captures the two-stage wear evolution, comprising an initial rapid running-in stage followed by a steady wear stage. Final RUL predictions are derived through Monte Carlo simulations, enabling the propagation of stochastic uncertainties embedded in the Wiener process. The bearing wear model, constructed via coupled elasto-hydrodynamic simulations and experimentally validated, demonstrates the efficacy of the proposed methodology in comparative analyses.