Low-Resource Spectral Feature Alignment for Cross-Condition Rolling Bearing Fault Diagnosis

Cross-condition rolling-bearing diagnosis is difficult because operating-condition shifts alter both the amplitude distribution and spectral composition of vibration signals. This study evaluates whether a fixed spectral representation changes low-resource, label-blind unsupervised domain-adaptation performance when the learning backbone and optimization budget are controlled. Each 1024-point vibration window was standardized, transformed through log(1+|rFFT|), L2-normalized, and processed by a one-channel convolutional neural network. Source-only, Deep CORAL, maximum mean discrepancy (MMD), and domain-adversarial neural-network objectives were compared under a manifest-defined protocol that permits only 10% of a target-domain alignment pool to be used without labels. Target-query labels were unavailable to training and checkpoint selection; checkpoints were selected exclusively using source-domain validation Macro-F1. Five fixed training seeds were used. On CWRU hp0-to-hp3, FFT-MMD achieved 98.92±0.34% Macro-F1, compared with 77.21±5.16% for time-domain MMD; DANN achieved 99.23±0.10%. On the reverse transfer, FFT-MMD achieved 96.84±0.68% and DANN 98.35±1.00%. On JNU-S600, FFT-MMD reached 90.48±1.58%, whereas time-domain MMD reached 67.05±3.75%. Time-domain MMD, however, outperformed FFT-MMD on the JNU-S800 and JNU-S1000 boundary tasks. The results therefore indicate a conditional representation effect, not a universal advantage of frequency-domain input.