Unsupervised Cross-Condition Fault Diagnosis Using Dual-Branch Time–Frequency Adaptive Fusion and MMD Alignment

Abstract

Cross-condition bearing diagnosis is challenged by simultaneous changes in transient impacts, spectral-energy patterns, and source–target feature distributions as speed and load vary. Handcrafted indicators and single-representation networks often capture only part of this coupled variation, whereas global alignment alone may reduce marginal discrepancy without preserving diagnostic structure. This study proposes DBAF-Net, which couples a raw vibration sequence with a five-channel, multi-resolution, source-domain-calibrated Kaiser short-time Fourier transform tensor. The time-domain branch models impact timing and long-range dependence, while the time–frequency branch describes local texture and band-energy redistribution. A feature selection and fusion module maps both representations into a shared space, produces sample-conditioned channel gates, and retains pre-refinement information through a residual path. Maximum mean discrepancy is then imposed on the fused feature used by the classifier under a transductive unsupervised domain-adaptation protocol. Across six Paderborn transfers, DBAF-MMD achieves 98.96% mean target accuracy. On T3 and T4, the full model obtains Macro-F1 scores of 98.40% and 99.70% and MCC values of 0.9832 and 0.9900, respectively. Structured occlusion produces larger decision changes for time–frequency perturbations than for local raw-signal masking, while the selected T4 checkpoint yields an expected calibration error of 6.44%. In a bounded coal-mill case, the method obtains 98.00% accuracy, 97.98% Macro-F1, and 0.9706 MCC for three-speed cross-machine transfer under fixed misalignment. DBAF-MMD requires 4.9976 ms/sample under the common profiling protocol.