Reducing Negative Transfer in Domain Adaptation for Vibration Fault Diagnosis

Abstract

Unsupervised domain adaptation (UDA) for vibration-based fault diagnosis can improve transfer across changing operating conditions, but its reliability remains a practical concern. In particular, large domain shifts can lead to negative transfer, where an adapted model underperforms a source-only baseline. This motivates evaluation beyond average gains, with emphasis on worst-case behavior and failure modes relevant to deployment. This study proposes a lightweight safeguard for discrepancy-based UDA that does not require labeled target data. The approach augments standard adaptation with an unlabeled monitoring rule based on target prediction entropy and alignment-loss trends. When adaptation appears unstable, training is paused and the model is rolled back to a safer checkpoint. The safeguard is designed as a small reliability layer on top of existing UDA pipelines rather than as a new adaptation method. We evaluate source-only training, standard UDA, source-pretrained UDA, and safeguarded UDA on the Case Western Reserve University (CWRU) and Paderborn University (PU) bearing datasets under multiple cross-condition transfer tasks. Experiments include raw time-domain, FFT-based, and STFT-based representations with MMD- and CORAL-based adaptation. Results show that negative transfer is a repeatable phenomenon, particularly on more challenging CWRU shifts, while source-pretrained UDA substantially affects reliability. The safeguard shows partial mitigation of harmful adaptation in selected PU cases but does not consistently prevent degradation across all scenarios. Overall, the results highlight that monitoring adaptation dynamics can improve reliability in some settings, but that safe deployment of UDA for fault diagnosis still requires explicit consideration of worst-case behavior and baseline comparisons.