Reformulating Multi-Source Domain Adaptation as Single-Source Transfer for Cross-Condition Bearing Fault Diagnosis
IEEE Access · 2026 · 14 · 82676-82691
DOI: 10.1109/access.2026.3697848 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Gaoquan Gu und 2 weiteren Autor:innen. Die Arbeit nennt 37 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Natural Science Foundation of China.
Abstract (Originalsprache)
To address the degradation in generalization performance of diagnostic models caused by data distribution shifts under variable working conditions in rolling bearings, this paper proposes a cross-condition fault diagnosis method that reformulates multi-source domain adaptation as an equivalent single-source transfer problem. Existing multi-source transfer methods usually align each source domain with the target domain separately, which not only increases the complexity of loss design and optimization, but also easily suffers from source-domain heterogeneity, leading to inconsistent alignment objectives and weak objective coordinations. To this end, a multi-source-to-single-source channelized mapping mechanism is first constructed, in which multi-source signals are stacked along the channel dimension and encoded by a shared convolutional encoder to achieve unified representation of multi-source features. On this basis, a channel-wise inter-domain discrepancy elimination module is designed to suppress the internal statistical shifts among source domains through low-order statistical normalization and discrepancy-aware weighting, thereby enhancing the consistency of the fused representations. Furthermore, a local maximum mean discrepancy constraint is introduced to perform fine-grained alignment of the conditional distributions in class-related subdomains, thus improving intra-class compactness and inter-class separability. Ablation results further indicate that LMMD plays the dominant role in reducing source–target conditional distribution discrepancies, while CIDEM improves the stability of fused multi-source representations. Experimental results on multiple cross-condition transfer tasks over a self-built dataset and the HUST dataset demonstrate that the proposed method achieves average diagnostic accuracies of 99.53% and 98.97%, respectively.
Förderung
- National Natural Science Foundation of China
Zitieren
APA
Gu, G., Yu, X., & Zhao, C. (2026). Reformulating Multi-Source Domain Adaptation as Single-Source Transfer for Cross-Condition Bearing Fault Diagnosis. IEEE Access, 14, 82676-82691. https://doi.org/10.1109/access.2026.3697848
Vancouver
Gu G, Yu X, Zhao C. Reformulating Multi-Source Domain Adaptation as Single-Source Transfer for Cross-Condition Bearing Fault Diagnosis. IEEE Access. 2026;14:82676-82691. doi:10.1109/access.2026.3697848
BibTeX
@article{Gu2026,
author = {Gu, Gaoquan and Yu, Xiang and Zhao, Cunsheng},
title = {Reformulating Multi-Source Domain Adaptation as Single-Source Transfer for Cross-Condition Bearing Fault Diagnosis},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {82676--82691},
doi = {10.1109/access.2026.3697848}
}
EUVIMED