Physics-Guided Dual-Branch Domain Generalization with Confidence-Adaptive Decision Fusion for Cross-Speed Fault Diagnosis of Planetary Gearboxes

Data distribution shift across different rotational speeds limits domain generalized fault diagnosis for planetary gearboxes. Existing single-representation domain generalization (DG) methods degrade severely in cross-speed extrapolation, with different representations showing speed-dependent advantages. The core contribution of this paper is a sample-wise confidence-adaptive decision fusion mechanism under the pure DG constraint. For each test sample, the fusion weights are computed online from the relative softmax prediction confidence of the two branches. The mechanism requires neither target-domain data nor calibration sets and introduces no fusion hyperparameters. A physics-guided dual-branch representation is constructed to support this mechanism. The physics-guided order-domain feature (PGF) branch extracts speed-independent features via triple-band demodulation and computed order tracking resampling. Physical enhancement is further applied to improve cross-domain robustness. A homologous time-domain branch shares the IEDGNet architecture and training strategy. Experiments on the WT planetary gearbox dataset show that the proposed method achieves 92.23% average accuracy in forward speed extrapolation, which is 5.62 percentage points higher than the best single branch and only 0.72 percentage points lower than the global Oracle upper bound, and reaches 96.48% accuracy in backward generalization. Complementarity analysis and feature visualization reveal non-overlapping misclassification regions between branches, confirming that the confidence arbitration mechanism effectively exploits dual-representation complementarity in cross-speed extrapolation scenarios.