Physics-Informed Domain Generalization for Bearing Prognostics Under Unseen Operating Conditions

Abstract

Modern condition-based maintenance of rotating machinery increasingly relies on data-driven prognostic models to estimate bearing health and remaining useful life (RUL). While machine-learning approaches have demonstrated strong performance under known operating conditions, their reliability often degrades under unseen loads, speeds, and degradation regimes, limiting their industrial applicability. This work addresses bearing prognostics under unseen operating conditions through a physics-informed real-to-real transfer learning framework. To improve physical consistency and long-term prognostic stability, the proposed approach incorporates constraints inspired by fatigue crack growth theory into the learning process. In particular, monotonic degradation behavior consistent with Paris-law-type dynamics is enforced by learning bias, promoting physically plausible degradation evolution and RUL estimation. Building on this physics-guided foundation, the framework further addresses domain shifts through domain generalization and feature disentanglement. The method accounts for both marginal and conditional domain shifts across multiple source domains representing different operating conditions and degradation trajectories. By disentangling domain-invariant degradation features from domain-specific operational characteristics, the model enables zero-shot generalization to previously unseen target conditions without requiring target-domain data. The proposed method is validated using experimental bearing run-to-failure datasets and demonstrates robust prognostic performance under unseen operating conditions while maintaining physically consistent degradation behavior. The results highlight the potential of combining physics-informed learning with domain generalization for reliable industrial bearing prognostics.