A Novel Consistent Regularized Domain Smoothing Generalized Networks for Cross-Domain Fault Diagnosis
IEEE Sensors Journal · 01.01.2026 · 26 (1) · 181-193
DOI: 10.1109/jsen.2025.3633066 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Sensors Journal, erschienen am 01.01.2026. Verfasst von Haoyang Jian und 3 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 51 Literaturangaben. Gefördert von: Natural Science Foundation of Shanxi Province, Doctoral Research Start-Up Foundation of Taiyuan University of Science and Technology.
Abstract (Originalsprache)
In recent years, domain generalization methods have gained significant attention in fault diagnosis to address performance degradation caused by data distribution shifts under varying operating conditions. Practical industrial applications—such as wind turbine gearbox monitoring under fluctuating wind speeds and chemical pump bearing diagnosis with variable load conditions—face substantial challenges where diagnostic models often demonstrate significantly degraded performance when encountering unseen operational scenarios. To overcome these limitations, a novel consistent regularized domain smoothing generalized network (CRDSGN) is proposed for cross-domain fault diagnosis. The network first trains pure feature extractors for each domain and then performs hybrid adversarial training. During domain alignment, the consistency regularization term is employed to measure the distance between features extracted by the mixed feature extractor and those extracted by the pure feature extractors. By maximizing the consistency regularization error, the network facilitates more effective learning of domain-invariant features. In addition, a weighted environment label smoothing (WELS) strategy is proposed to improve training smoothness while providing more informative soft labels for the domain discriminator. Comprehensive experiments on CWRU and JNU bearing datasets demonstrate that CRDSGN achieves a notable improvement in diagnostic accuracy, with an average increase of at least 4.54% on unseen target domains compared to state-of-the-art methods. This work provides an effective solution for reliable intelligent fault diagnosis of industrial rotating machinery operating under diverse and variable conditions, demonstrating significant practical value for real-world industrial applications.
Förderung
- Natural Science Foundation of Shanxi Province
- Doctoral Research Start-Up Foundation of Taiyuan University of Science and Technology
Zitieren
APA
Jian, H., Chen, G., Zhang, C., & Du, X. (2026). A Novel Consistent Regularized Domain Smoothing Generalized Networks for Cross-Domain Fault Diagnosis. IEEE Sensors Journal, 26(1), 181-193. https://doi.org/10.1109/jsen.2025.3633066
Vancouver
Jian H, Chen G, Zhang C, Du X. A Novel Consistent Regularized Domain Smoothing Generalized Networks for Cross-Domain Fault Diagnosis. IEEE Sensors Journal. 2026;26(1):181-193. doi:10.1109/jsen.2025.3633066
BibTeX
@article{Jian2026,
author = {Jian, Haoyang and Chen, Gaohua and Zhang, Chunmei and Du, Xudi},
title = {A Novel Consistent Regularized Domain Smoothing Generalized Networks for Cross-Domain Fault Diagnosis},
journal = {IEEE Sensors Journal},
year = {2026},
volume = {26},
number = {1},
pages = {181--193},
doi = {10.1109/jsen.2025.3633066}
}
EUVIMED