Abstract
Deep feature learning and domain adaptation have improved bearing fault diagnosis by enhancing signal representations and reducing cross-domain distribution discrepancies. However, preserving fault information and class structure during alignment remains challenging under varying operating conditions. This study proposes a stationary wavelet-based conditional subdomain adaptation network (SWCSAN). Its statistical–deep feature fusion module combines deep features extracted from undecimated multiscale wavelet representations with seven log-transformed time-domain descriptors, integrating waveform patterns and signal-level statistics. A confidence-gated joint alignment strategy couples conditional adversarial adaptation (CDAN) with local maximum mean discrepancy (LMMD) in the fused feature space. The adversarial branch aligns feature prediction representations, while the subdomain branch matches corresponding classes using source labels and high-confidence target pseudo-labels. Source classification and both adaptation objectives are optimized jointly. Experiments on six transfer tasks per dataset yield average accuracies of 98.67% on PU and 99.53% on HUST, exceeding those of the evaluated adaptation baselines.