Fault-Source Transfer in Motor-Current-Based Bearing Fault Diagnosis: A Bearing-Level Evaluation

Motor-current-based bearing diagnosis is attractive because current measurements are often already available, yet a binary faulty label can combine defects with different physical origins. This study benchmarks source-defined binary evaluation under held-out physical bearings; it does not identify a causal effect of damage origin or assume that different faulty-source populations are equivalent. The proposed framework is tested on the Paderborn University bearing dataset, which includes healthy bearings and three faulty-bearing groups: artificially damaged bearings (ART), accelerated-lifetime bearings excluding mixed-type cases (AGED), and mixed-type bearings (MIX). Four healthy-plus-faulty training compositions (ART, AGED, ART+AGED, and ART+AGED+MIX) are generated and separately evaluated on ART, AGED, and MIX test sources. The primary analysis uses one operating condition, three source-aware bearing-level splits, three random seeds, and three predefined diagnostic probes combining time–frequency (TF) image representations with convolutional neural networks (CNNs); a second operating condition provides only a bounded sensitivity check. For the reference continuous wavelet transform (CWT)–EfficientNetB0 probe, ART-only training tested on AGED faults achieved a balanced accuracy of 0.6618 ± 0.1646, with healthy/faulty recalls of 0.3750/0.9486. Adding AGED faults increased the aged-test mean to 0.7490 ± 0.2201 and improved the recall balance to 0.5292/0.9688. The ART-to-AGED transfer difference was +0.0454 under the cross-source-minus-in-source convention: transfer did not collapse, but the recall imbalance does not establish source equivalence. For MIX testing, the observed inclusion contrasts were −0.0424 and +0.0312 in two settings; with only three MIX bearings and one repeated probe in the second setting, the direction of a general MIX-inclusion effect is not established. A separate size-matched control retains split-dependent aged-test balanced-accuracy gains that are not consistently accompanied by higher aged-fault recall. A further healthy-bearing role rotation shows strong unit-specific false-alarm sensitivity and sign-changing MIX effects across three physical MIX bearings. A fixed linear support vector machine (SVM) baseline on the same CWT source-image and bearing-split matrix has lower mean balanced accuracy in all 12 source cells and reverses the signs of two CNN source-inclusion contrasts, demonstrating model-dependent effect directions. Across the three original TF–CNN probes, mean AGED inclusion is positive, whereas the MIX inclusion direction is inconsistent. For the tested current channel, bearing inventory, operating conditions and predefined probes, the results favour explicit source-wise reporting; observed cross-source detection alone does not establish source interchangeability, a mechanism, or broader deployment validity.