Motor-Current-Based Bearing Fault Detection Under Unseen Operating Conditions

Reliable motor-current-based bearing diagnosis requires evaluation on unseen physical bearings and operating conditions. This study uses the Paderborn University benchmark, acquired from a 425 W permanent-magnet synchronous motor (PMSM) test rig, to evaluate time–frequency deep transfer learning under strict bearing-level grouping. Four representations—continuous wavelet transform (CWT), short-time Fourier transform (STFT), wavelet synchrosqueezed transform (WSST), and Fourier synchrosqueezed transform (FSST)—are combined with pretrained CNN backbones across four binary targets: aged-only A/B and artificial-plus-aged C/D, with mixed-fault bearings excluded/included within each pair. The workflow includes pooled-condition candidate discovery, exploratory Main-split leave-one-operating-condition-out (LOCO) screening, and a retrospective multi-split LOCO audit. The audit contains 288 crossed condition–split–seed evaluations. Because pooled test summaries and Main-split LOCO results informed later stages, these evaluations provide descriptive robustness evidence rather than an independent post-selection test. Target A achieved the highest all-split mean balanced accuracy (0.736 for CWT–EfficientNetB0). The pairs for Targets B and C were near-ties, and the Target D ordering reversed when Main was excluded. Across the eight audited candidates, mean sensitivity ranged from 0.618 to 0.948, whereas specificity ranged from 0.092 to 0.564. Target D combined approximately 0.89 sensitivity with an approximately 0.90 false-alarm rate. Thus, operating condition, fault-class composition, bearing split, and error-cost priorities all affect model interpretation. A matched current-domain baseline audit added 336 evaluations using handcrafted-feature RBF–SVM and Random Forest models and a compact raw-current 1D-CNN. The results show that instability is broader than the TF–CNN pipeline but is not uniform across model families: TF candidates were clearly stronger for Targets A and C, feature-based models were stronger for Target B, and Target D remained mixed and protocol-sensitive. A complementary bearing-level source-group analysis quantified six healthy/fault-source categories across the 37 current features; among the eight features with the largest mean between-group variance fraction, only spectral entropy and dominant power fraction preserved the same mixed-versus-non-mixed contrast direction across all four operating conditions. An additional matched Target D reference held the binary target, physical-bearing split, held-out condition, seed, fourth-order FSST representation, ResNet-50 backbone, and training settings fixed while changing the sensing channel. Across 24 matched runs, vibration showed descriptively higher mean balanced accuracy (0.642 vs. 0.503) and specificity (0.476 vs. 0.146), while sensitivity was slightly lower (0.807 vs. 0.859); the comparison does not establish universal modality superiority. Pooled-condition performance is useful for candidate discovery, but credible condition-generalization claims require explicit separation of exploratory selection and confirmatory testing. The numerical findings are specific to the evaluated PMSM benchmark and do not establish universal performance across electric-machine types.