EUVIMEDEUVIMED entdecken
Crossref

A novel semi-supervised domain generalization framework based on domain knowledge-guided and adversarial learning

Mengzhen Geng, Shida Yu, Jiabing Zhou, Yixuan Sun

Measurement Science and Technology · 31.07.2026 · 37 (30) · 306109

DOI: 10.1088/1361-6501/ae8bc9 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 31.07.2026. Verfasst von Mengzhen Geng und 3 weiteren Autor:innen. Die Arbeit nennt 40 Literaturangaben. Gefördert von: Research Foundation for Youth Scholars of Beijing Technology and Business University, Beijing Municipal Commission of Education.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Extending intelligent fault diagnosis models to unseen operating conditions is an important research topic in predictive maintenance for mechanical systems. Existing domain generalization fault diagnosis (DGFD) methods usually require sufficient labeled samples to be available in all source domains, whereas obtaining high-quality labels is prohibitively expensive. To alleviate the scarcity of labeled samples in practical industrial scenarios, this paper focuses on semi-supervised domain generalization (SSDG) for fault diagnosis and proposes a semi-supervised DGFD framework based on domain-guided pseudo-labeling and multi-objective adversarial learning, termed DMSSDG. First, to address noise accumulation caused by conventional pseudo-labeling mechanisms, this paper constructs a domain-guided pseudo-labeling (DGPL) mechanism, which integrates model predictions and intra-domain class similarity to assign pseudo-labels to unlabeled data and further selects reliable pseudo-labels through confidence-based filtering. Second, to make better use of low-confidence samples, a low-confidence knowledge mining (LCKM) strategy is proposed. By introducing entropy minimization regularization and a dynamic weighting mechanism, LCKM fully mines fault-discriminative information contained in low-confidence samples. Finally, to tackle the challenge of cross-domain alignment, an adversarial domain alignment (ADA) architecture is developed. By introducing a gradient reversal layer, ADA enables adversarial optimization between the feature extractor and the domain discriminator, thereby suppressing domain-related information while preserving fault-discriminative information. In this way, the model learns feature representations that are both fault-discriminative and domain-invariant. Extensive experiments on the CWRU and PU bearing datasets show that DMSSDG achieves superior performance on 24 cross-condition diagnosis tasks, with average accuracies of 96.92% and 90.08%, respectively. Its performance approaches that of fully supervised domain generalization methods, demonstrating the effectiveness and superiority of the proposed framework for fault diagnosis under unseen operating conditions with limited labeled data.

Förderung

  • Research Foundation for Youth Scholars of Beijing Technology and Business University
  • Beijing Municipal Commission of Education

Zitieren

APA

Geng, M., Yu, S., Zhou, J., & Sun, Y. (2026). A novel semi-supervised domain generalization framework based on domain knowledge-guided and adversarial learning. Measurement Science and Technology, 37(30), 306109. https://doi.org/10.1088/1361-6501/ae8bc9

Vancouver

Geng M, Yu S, Zhou J, Sun Y. A novel semi-supervised domain generalization framework based on domain knowledge-guided and adversarial learning. Measurement Science and Technology. 2026;37(30):306109. doi:10.1088/1361-6501/ae8bc9

BibTeX
@article{Geng2026,
  author = {Geng, Mengzhen and Yu, Shida and Zhou, Jiabing and Sun, Yixuan},
  title = {A novel semi-supervised domain generalization framework based on domain knowledge-guided and adversarial learning},
  journal = {Measurement Science and Technology},
  year = {2026},
  volume = {37},
  number = {30},
  pages = {306109},
  doi = {10.1088/1361-6501/ae8bc9}
}
Kostenlos starten →