Skip to content
Open access

Physics-Guided Representation Learning and Selective Domain Adaptation for Cross-Condition Bearing Fault Diagnosis

Oct 2026 · Machine Learning and Knowledge Extraction · 0 citations · 35 references

Abstract

Cross-condition bearing fault diagnosis remains challenging because changes in operating conditions can cause distribution shifts that degrade diagnostic performance. Moreover, domain adaptation is not always beneficial and may introduce negative transfer and unnecessary computational cost. This paper proposes a Physics-Guided Domain Representation Network (PGDRN) with discrepancy-gated deployment, combining temporal, spectral, and physics-guided representations with an unsupervised decision mechanism that determines whether target-domain adaptation should be initiated. The framework was evaluated on the Paderborn University bearing dataset across three operating-condition transfer tasks, with the final deployment and domain-adaptation comparisons repeated over five independent runs. Across five independent runs for each of the three domain-adaptation tasks, gated deployment achieved a mean Macro-F1 gain of +0.0105 relative to the frozen model, reduced the observed negative-transfer rate from 33.3% under unconditional adaptation to 13.3%, and reduced average adaptation time by 26.3%. In comparison with representative domain-adaptation methods, Gated PGDRN achieved an overall Macro-F1 of 0.7763, compared with 0.7667 for Deep CORAL, 0.7706 for DANN, 0.7495 for MDD, 0.6410 for UAC-SFUDA, and 0.7655 for PrEPoA. These results demonstrate the value of considering not only how a diagnostic model should adapt, but also when adaptation should be initiated.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.