Representation, normalization and architecture interact across operating- and hardware-domain shifts in source-only fault diagnosis

Fault-diagnosis studies often change signal representation, normalization, architecture and training objective together, obscuring attribution. We crossed raw waveforms, log-magnitude spectra and affine-LayerNorm spectra with four objectives on a matched ResNet18 under source-only, target-isolated domain holdouts, and compared compact spectral references. On four pump-frequency holdouts, empirical-risk-minimization (ERM) Macro-F1 was 0.801 for Raw, 0.608 for LogFFT and 0.706 after affine LayerNorm; FFT-Linear and the normalized log-spectral classifier (NLSC) reached 0.892 and 0.950. A post-review non-affine LayerNorm control reached 0.708, indicating that learned per-bin affine weights were not required for the observed pump recovery. Record-level aggregation over 64 channel records from 32 paired acquisitions preserved the ordering. In four HUST leave-one-bearing-type-out folds (84 records), ResNet18 marginal means across objectives were 0.251, 0.332 and 0.400, respectively, whereas NLSC reached 0.244; non-affine and affine ERM arms were nearly equal (0.386 and 0.387). Fixed frequency-axis scaling caused large losses, and source-dispersion-conditioned margins failed to outperform matched controls consistently across pump, JNU and HUST. Representation, normalization and architecture therefore interact with shift type; compact-model and margin benefits are not universal.