Physics-guided virtual-real fusion and cross-level synergistic alignment for Sim2Real bearing fault diagnosis

Missing real fault observations hinder reliable bearing fault diagnosis because labelled samples for some known fault categories are often unavailable. To address this problem, a physics-guided framework integrating virtual-real fusion and cross-level synergistic alignment is proposed. At the data level, an energy-equivalent mapping strategy combines simulated fault responses with measured healthy-state background vibration under energy constraints, preserving interpretable fault impulses while reducing statistical discrepancies. At the feature level, a confidence-weighted conditional adversarial mechanism down-weights low-fidelity virtual samples and uncertain target samples during cross-domain transfer. At the semantic level, a target-anchored centre calibration strategy uses high-confidence target samples as physical anchors to correct semantic bias in virtual fault categories and guide simulated features towards real fault distributions. Experiments were conducted on a self-developed bearing test platform and corresponding finite-element simulations under cross-operating-condition and cross-fault-scale scenarios. The proposed method achieved average accuracies of 98.69% and 96.58%, respectively, outperforming representative domain adaptation methods and demonstrating strong generalisation under incomplete real fault observations.