Counterfactual Robust Order-Domain Generalization for Cross-Dataset Bearing Fault Diagnosis

Single-source cross-dataset bearing diagnosis is challenged by changes in machines, sensors, and operating conditions. We propose CRODG, which creates virtual source environments from angular-resampled Hilbert-envelope signals and combines shared representation learning with environment-level objectives. CRODG and four baselines are compared on 12 directed transfers using five seeds and 200 updates. CRODG reaches 64.46% mean accuracy, 65.69% balanced accuracy, and 57.68% macro-F1. The method achieves the highest mean accuracy, balanced accuracy, and macro-F1 among the four compared baselines, leading the strongest baseline in mean accuracy by 7.59 percentage points and ranking first on nine of the 12 transfers. Exploratory three-seed component comparisons show numerical differences across settings, but none of the 16 paired comparisons remains significant after Holm correction. These analyses do not establish the independent contribution of individual components. Exploratory sensitivity analysis finds the highest tested means with six environments and a moderate distortion limit.