Patch-Transformer Fault Diagnosis with Temporal-Shift Augmentation and Asymmetric Domain Alignment

Changes in operating speed and load shift vibration-signal distributions, limiting the transfer of fault diagnosis models across operating conditions. Temporal offsets between sampled windows introduce an additional source of variability. This study presents an Adaptive Cross-domain Alignment Network (ACADAN) that combines patch-based representation learning with capacity-asymmetric classifier-discrepancy adaptation. Random cyclic-shift augmentation reduces reliance on absolute waveform position. A Patch-Transformer encodes overlapping signal segments and pools their representations into a compact feature. A shallow–deep pair of task classifiers then supports alternating source-supervised learning, target-discrepancy maximisation, and feature-level discrepancy minimisation. On six Case Western Reserve University bearing transfer tasks, ACADAN achieved a reported average accuracy of 99.35%, compared with 96.40% for WDCNN. On the variable-speed gearbox Dataset 2, it achieved 99.83 ± 0.04%, compared with 96.48 ± 0.75% for MSCNN under the reported evaluation protocol. Additional studies examined single-channel inputs, temporal perturbations, patch settings, local smoothing, classifier capacity, and dropout. Synthetic time-warp tests showed that cyclic-shift augmentation alone remained sensitive to strong temporal deformation, distinguishing temporal-shift robustness from speed invariance. These findings support the combined design for closed-set cross-condition diagnosis within the evaluated tasks while leaving open-set recognition and continuous online adaptation for further evaluation.