Non-adversarial cross-condition bearing fault diagnosis from vibration measurements using Gabor wavelet convolution and sliced Wasserstein distance

Rolling bearings in rail-transit rotating machinery frequently operate under variable speeds, fluctuating loads, and strong background noise, resulting in substantial distribution shifts in vibration measurements and degraded cross-condition diagnostic performance. Existing domain-adaptation methods commonly rely on adversarial learning or Maximum Mean Discrepancy (MMD); however, adversarial optimization may be unstable, while MMD has limited capability to characterize complex distribution geometry. To address these issues, this paper proposes a non-adversarial cross-condition bearing fault diagnosis method, termed Gabor Wavelet Feature learning with Sliced Wasserstein Distance (GWF-SWD). Parallel learnable Gabor wavelet convolutions are first employed to extract complementary multiscale time-frequency features and enhance transient fault impulses. A Frequency-aware scale-channel Attention Mechanism (FAM) then adaptively recalibrates Gabor-derived feature channels, emphasizing fault-relevant responses while suppressing operating-condition and noise interference. Finally, global and class-conditional Sliced Wasserstein Distance (SWD) alignment is introduced to reduce source–target distribution discrepancies through efficient stochastic projections. Experiments were conducted on the 6205 and NU205EM bearing vibration datasets under four cross-condition scenarios, including variable-speed, variable-load, weak-fault, and noisy conditions. The average diagnostic accuracy was calculated over all transfer tasks in these scenarios, achieving 98.85%, which demonstrates the robustness of the proposed GWF-SWD framework for cross-condition vibration measurement analysis.