EUVIMEDEUVIMED entdecken
Crossref

An integrated deep learning model with reinforcement learning for rolling bearing fault diagnosis under small-sample conditions

Hui Shao, Gang Han, Huaitao Shi

Journal of Vibration and Control · 05.03.2026

DOI: 10.1177/10775463261429084 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Vibration and Control, online erschienen am 05.03.2026. Verfasst von Hui Shao und 2 weiteren Autor:innen. Die Arbeit nennt 17 Literaturangaben. Gefördert von: The Key Research and Development Program of China, National Natural Science Foundation of China, “Unveiling and Leading” Special Science and Technology Research Project of Liaoning Province.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

This paper tackles the challenge of accurately diagnosing rolling bearing faults under limited data conditions by introducing an integrated deep learning framework that synergistically combines Temporal Convolutional Networks (TCN), Bidirectional Gated Recurrent Unit (BiGRU), Attention mechanism, and Deep Q-Network (DQN)-based reinforcement learning. Conventional approaches frequently exhibit overfitting and poor generalization when labeled fault data is scarce. To address this limitation, our model leverages TCN’s capacity for parallelized long-term dependency modeling, BiGRU’s capability to capture bidirectional contextual features, and the Attention mechanism’s adaptive feature weighting capability. Unlike traditional methods that rely on static architectures or manual tuning, this study innovatively introduces a Deep Q-Network (DQN) agent as an external controller. This design theoretically enables the dynamic self-optimization of hyperparameters and network structures during training, ensuring robust convergence even under severe data scarcity. Additionally, a reinforcement learning strategy dynamically optimizes hyperparameters and network architecture during training, thereby enhancing convergence and robustness under data scarcity. The framework integrates advanced signal processing techniques, including Variational Mode Decomposition (VMD) for noise-resistant signal decomposition and maximum singular value energy entropy for effective feature fusion. Experimental results on benchmark bearing fault datasets demonstrate that the proposed model achieves a detection rate (DR) exceeding 80% with an error threshold of 0.195 and 90% with an error threshold of 0.2, significantly outperforming comparative models including LVQ, CNN, LSTM, and standalone TCN-BiGRU-Attention. Notably, at a 30% training sample size, the model improves DR by 3.21% and reduces the false alarm rate (FAR) by 4.13% compared to baseline methods. These results validate the method’s efficacy in achieving high diagnostic accuracy and stability under data-limited scenarios, providing a practical solution for intelligent fault diagnosis in industrial applications.

Förderung

  • The Key Research and Development Program of China
  • National Natural Science Foundation of China
  • “Unveiling and Leading” Special Science and Technology Research Project of Liaoning Province

Zitieren

APA

Shao, H., Han, G., & Shi, H. (2026). An integrated deep learning model with reinforcement learning for rolling bearing fault diagnosis under small-sample conditions. Journal of Vibration and Control. https://doi.org/10.1177/10775463261429084

Vancouver

Shao H, Han G, Shi H. An integrated deep learning model with reinforcement learning for rolling bearing fault diagnosis under small-sample conditions. Journal of Vibration and Control. 2026. doi:10.1177/10775463261429084

BibTeX
@article{Shao2026,
  author = {Shao, Hui and Han, Gang and Shi, Huaitao},
  title = {An integrated deep learning model with reinforcement learning for rolling bearing fault diagnosis under small-sample conditions},
  journal = {Journal of Vibration and Control},
  year = {2026},
  doi = {10.1177/10775463261429084}
}
Kostenlos starten →