EUVIMEDEUVIMED entdecken
Crossref

G-TWDNet: A Graph Diffusion Learning Framework for Compound Fault Diagnosis in Transformer Windings

Binyu Zhu, Huan Peng, Song Wang, Yiyun Li, Siying Chen, Yilin Wang, Jun Liu, Jiawei Wang

IEEE Sensors Journal · 2026-11 · 26 (21) · 32757-32771

DOI: 10.1109/jsen.2026.3732077 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Sensors Journal, 2026-11. Verfasst von Binyu Zhu und 7 weiteren Autor:innen. Die Arbeit nennt 44 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Power transformers are core equipment in electrical grids, yet their mechanical integrity is significantly threatened by winding deformations induced by short-circuit shocks. While frequency response analysis (FRA) has become a principal nonintrusive sensing and diagnostic technique, its effectiveness in real-world scenarios is severely hindered by the complexity of compound faults. To address these bottlenecks, a novel compound fault diagnosis framework based on graph diffusion learning is proposed. To the best of our knowledge, this is the first systematic experimental study to address compound faults in transformer windings. Three primary innovations are proposed in this article. First, a graph diffusion learning framework is introduced to capture stress-related coupling among single faults in transformer windings reflected by FRA. By treating each fault as an interconnected node, it models the global topology of compound faults. Second, a supervised contrastive prototypical embedding (S-CPE) method is proposed to distill latent signatures from the raw FRA sensor signals, establishing initial node representations for each potential fault type. Finally, a probabilistic representation method and a dynamic threshold decision mechanism, combining the maximum a posteriori (MAP) principle with cost-sensitive learning, are implemented to dynamically calculate the optimal decision thresholds under limited and imbalanced compound fault samples. Experimental results on a transformer winding model demonstrate that G-TWDNet achieves a single fault diagnosis accuracy of 98.29% and a compound fault diagnosis accuracy of 95.48%. This research provides a solution for transformer winding fault diagnosis, facilitating reliable transformer maintenance and power grid safety.

Zitieren

APA

Zhu, B., Peng, H., Wang, S., Li, Y., Chen, S., Wang, Y., Liu, J., & Wang, J. (2026). G-TWDNet: A Graph Diffusion Learning Framework for Compound Fault Diagnosis in Transformer Windings. IEEE Sensors Journal, 26(21), 32757-32771. https://doi.org/10.1109/jsen.2026.3732077

Vancouver

Zhu B, Peng H, Wang S, Li Y, Chen S, Wang Y, et al. G-TWDNet: A Graph Diffusion Learning Framework for Compound Fault Diagnosis in Transformer Windings. IEEE Sensors Journal. 2026;26(21):32757-32771. doi:10.1109/jsen.2026.3732077

BibTeX
@article{Zhu2026,
  author = {Zhu, Binyu and Peng, Huan and Wang, Song and Li, Yiyun and Chen, Siying and Wang, Yilin and Liu, Jun and Wang, Jiawei},
  title = {G-TWDNet: A Graph Diffusion Learning Framework for Compound Fault Diagnosis in Transformer Windings},
  journal = {IEEE Sensors Journal},
  year = {2026},
  volume = {26},
  number = {21},
  pages = {32757--32771},
  doi = {10.1109/jsen.2026.3732077}
}
Kostenlos starten →