EUVIMEDEUVIMED entdecken
Crossref

Accuracy-constrained automatic transfer pruning for edge-side fault diagnosis of wind-turbine gearboxes

Yanling Wu, Baoping Tang, Xiaolong Yang, Qingqing Tian

Measurement Science and Technology · 15.05.2026 · 37 (20) · 206202

DOI: 10.1088/1361-6501/ae693a ↗

Kurz gefasst

Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 18.05.2026. Verfasst von Yanling Wu und 3 weiteren Autor:innen. Die Arbeit nennt 41 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Natural Science Basic Research Program of Shaanxi Province.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

To address the challenging trade-off between accuracy and complexity in edge-side fault diagnosis models, an accuracy-constrained automatic transfer pruning (TP) method is proposed. Complex fault diagnosis baseline models were automatically pruned into a high-quality pruned model that balances high accuracy and lightweight design to realize edge-side fault diagnosis of wind-turbine gearboxes. First, based on the reinforcement learning framework, an accuracy-constrained reward function was formulated to guide a policy network and ensure high diagnostic accuracy of the pruned model. Furthermore, the layer-wise pruning sensitivity of the baseline model was automatically captured, and the pruning rate of each layer was determined. Subsequently, a TP strategy was proposed to transfer the knowledge of the pruned filters with lower scores to the retained filters with higher scores to minimize the information loss caused by filter pruning. Finally, the loss-assisted fine-tuning strategy continuously injected baseline model knowledge into the pruned model to restore its performance of the pruned model during fine-tuning. The practicability and usefulness of the method were comprehensively validated through its application to a drivetrain diagnostic simulator and real-world measurement datasets. Finally, the pruned model was deployed on practical edge hardware, Raspberry Pi 4B, to validate the advantages of edge-side fault diagnosis. Compared with competing methods, the proposed method yielded better diagnostic accuracy, with a larger percentage decrease in floating point operations (FLOPs) and faster latency. In particular, the pruned model of the proposed method on the wind-turbine gearbox dataset for the residual network-18 baseline achieved 97.9% diagnostic accuracy with a 90.53% decrease in FLOPs and 9.85 ms latency. The experimental results indicate that the proposed method is more suitable for edge-side fault diagnosis of wind-turbine gearboxes. This approach provides a potential solution for practical engineering applications.

Förderung

  • National Natural Science Foundation of China
  • Natural Science Basic Research Program of Shaanxi Province

Zitieren

APA

Wu, Y., Tang, B., Yang, X., & Tian, Q. (2026). Accuracy-constrained automatic transfer pruning for edge-side fault diagnosis of wind-turbine gearboxes. Measurement Science and Technology, 37(20), 206202. https://doi.org/10.1088/1361-6501/ae693a

Vancouver

Wu Y, Tang B, Yang X, Tian Q. Accuracy-constrained automatic transfer pruning for edge-side fault diagnosis of wind-turbine gearboxes. Measurement Science and Technology. 2026;37(20):206202. doi:10.1088/1361-6501/ae693a

BibTeX
@article{Wu2026,
  author = {Wu, Yanling and Tang, Baoping and Yang, Xiaolong and Tian, Qingqing},
  title = {Accuracy-constrained automatic transfer pruning for edge-side fault diagnosis of wind-turbine gearboxes},
  journal = {Measurement Science and Technology},
  year = {2026},
  volume = {37},
  number = {20},
  pages = {206202},
  doi = {10.1088/1361-6501/ae693a}
}
Kostenlos starten →