FTPC-FL: A forgetting-suppressed two-stage personalized contrastive federated learning method for industrial fault diagnosis
Journal of Computational Design and Engineering · 12.11.2026
DOI: 10.1093/jcde/qwag100 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Computational Design and Engineering, online erschienen am 12.11.2026. Verfasst von Kai Zheng und 4 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
As industrial equipment becomes increasingly complex, fault diagnosis is essential for system safety, downtime reduction, and operational reliability. Conventional centralized methods are limited by data aggregation challenges and privacy concerns, making federated learning an effective alternative. However, existing federated fault diagnosis methods still struggle to jointly handle label distribution skew and partial client participation, degrading diagnostic performance under coupled data and system heterogeneity. To address this issue, this paper proposes FTPC-FL, a forgetting-suppressed two-stage personalized contrastive federated learning method for industrial fault diagnosis. FTPC-FL first uses local historical class prototypes to adapt the shared feature extractor, thereby reducing its mismatch with the local classifier when long-inactive clients rejoin training. It then introduces global prototype feedback contrastive learning to improve fault representations under label skew and employs forgetting-suppressed conflict-aware aggregation to preserve historical client knowledge under partial participation. We conducted experiments on the HUST and PU bearing datasets with 40 and 100 clients under different degrees of label distribution skew and client participation rates. The experimental results show that, under the representative setting of 100 clients, Dir(0.5), and a participation rate of 0.05, based on three runs with independently generated data partitions and reported as mean ± standard deviation, FTPC-FL achieved accuracies of 80.09 ± 0.48% and 70.44 ± 0.57% on HUST and PU, respectively, outperforming the strongest baseline, FedPKDA, by 5.07 and 4.33 percentage points. The overhead analysis indicates that, under the investigated experimental settings, FTPC-FL introduces only a small additional communication overhead but increases the computational overhead of server-side aggregation. In addition, controlled client-absence experiments demonstrate that the proposed method effectively mitigates the forgetting of fault knowledge associated with long-inactive clients. Overall, FTPC-FL outperforms FedPKDA and other advanced federated learning methods across different client scales, degrees of label distribution skew, and participation rates, while maintaining stable fault diagnosis performance.
Zitieren
APA
Zheng, K., Li, S., Yang, L., Zhang, X., & Bai, Q. (2026). FTPC-FL: A forgetting-suppressed two-stage personalized contrastive federated learning method for industrial fault diagnosis. Journal of Computational Design and Engineering. https://doi.org/10.1093/jcde/qwag100
Vancouver
Zheng K, Li S, Yang L, Zhang X, Bai Q. FTPC-FL: A forgetting-suppressed two-stage personalized contrastive federated learning method for industrial fault diagnosis. Journal of Computational Design and Engineering. 2026. doi:10.1093/jcde/qwag100
BibTeX
@article{Zheng2026,
author = {Zheng, Kai and Li, Shaobo and Yang, Lei and Zhang, Xiangjie and Bai, Qiang},
title = {FTPC-FL: A forgetting-suppressed two-stage personalized contrastive federated learning method for industrial fault diagnosis},
journal = {Journal of Computational Design and Engineering},
year = {2026},
doi = {10.1093/jcde/qwag100}
}
EUVIMED