Structural Health Monitoring, Год журнала: 2025, Номер unknown
Опубликована: Май 24, 2025
In rotating machinery fault diagnosis, the distribution difference of vibration signals across working conditions and insufficient feature extraction significantly reduce diagnostic accuracy are often accompanied by negative transfer, making transfer learning from source domain to target complicated unstable. this paper, an innovative cross-working condition network, Joint weighted multi-scale network (JWMS-NET), is designed. The JWMS-NET model innovatively combines a dynamic convolutional residual as extractor, pseudo-label classification loss, Jensen–Shannon divergence (JSD) loss. First, JSD-based loss metric proposed enhance confusion alignment cross-domain features, thereby reducing impact inconsistent distributions in two domains. Second, response problem that may occur during extraction, paper designs improved mechanism. This mechanism effectively adjusts weights samples different domains through adaptive weight allocation, suppresses effect, prompts extractor learn more robust consistent features. Finally, can further improve optimizing representation flexibly controlling decision boundary. average reached 99.66%, 99.84%, 99.41% 26 cross-speed migration tasks on three datasets. Through subsequent experimental comparisons, outperforms existing adaptation model, verifying its superiority robustness bearing diagnosis.
Язык: Английский