Scholarlink

<p>The work proposes a hybrid approach for diagnosing faults in REB by combining Deep Learning and Reinforcement Learning approaches. The model proposed in the paper is a combination of a CNN for feature extraction, an ECA module for feature enhancement, and a PERDQN network for fault classification based on adaptation. Different operating conditions with varying speeds and loads (1000, 1500, and 2000 RPM and 0, 6, and 12&#xa0;g) were used to collect vibration data at a sampling frequency of 12,800&#xa0;Hz. This data was used to test the model for five bearing conditions: HB, ORF, IRF, BF, and CF. The model proposed in the paper overcame the drawbacks of conventional, standalone DL methods by making it possible for adaptive learning and improving class-imbalance handling through prioritized sampling. However, even though the dataset has relatively balanced fault classes, PER-DQN focuses more on the challenging training samples rather than sampling all the experiences uniformly, which leads to both better learning efficiency and classification performance. The comparison based on accuracy included various RL algorithms. Here are their scores: QR-DQN (93.96%), REINFORCE (94.32%), IQN (94.51%), Soft Q (94.54%), SARSA (94.77%), DDQN (96.89%), and PER-DQN (96.63%). PER-DQN, being th most efficient one with an interference time of 0.0005, was chosen. The reported inference time represents the average prediction time per test sample measured after model training under the same hardware configuration used for all benchmark models. The proposed PER-DQN + ECA-NET model achieved 97.24% accuracy, with 97.22% F1-score and 96.52% MCC. The hybrid model in question was able to push its macro-AUC score up to 99.86%, with each of the classes having their AUC value higher than 99% individually. The mean accuracy of the model obtained via five-fold cross-validation was 94.54%; the model was 0.001655 GFLOPs and contained 59, 529 parameters. The suggested system exhibited excellent results even after five separate runs, producing a mean accuracy of 97.21 ± 0.04% (95% CI: 97.21 ± 0.05%) and a macro-average AUC of **99.87 ± 0.01% (95% CI: 99.87 ± 0.01%). In fact, we’ve statistically validated the model through a Friedman test, confidence interval analysis, and an ablation study, which reveal improved ranking, stability, and the highest accuracy of the proposed CNN + ECA + PER-DQN model.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hybrid reinforcement learning and attention mechanism framework for intelligent fault diagnosis of rolling element bearings

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月4日

查看全文
  • Avishi Oj,
  • Suha Haroon,
  • T. Narendiranath Babu

摘要

The work proposes a hybrid approach for diagnosing faults in REB by combining Deep Learning and Reinforcement Learning approaches. The model proposed in the paper is a combination of a CNN for feature extraction, an ECA module for feature enhancement, and a PERDQN network for fault classification based on adaptation. Different operating conditions with varying speeds and loads (1000, 1500, and 2000 RPM and 0, 6, and 12 g) were used to collect vibration data at a sampling frequency of 12,800 Hz. This data was used to test the model for five bearing conditions: HB, ORF, IRF, BF, and CF. The model proposed in the paper overcame the drawbacks of conventional, standalone DL methods by making it possible for adaptive learning and improving class-imbalance handling through prioritized sampling. However, even though the dataset has relatively balanced fault classes, PER-DQN focuses more on the challenging training samples rather than sampling all the experiences uniformly, which leads to both better learning efficiency and classification performance. The comparison based on accuracy included various RL algorithms. Here are their scores: QR-DQN (93.96%), REINFORCE (94.32%), IQN (94.51%), Soft Q (94.54%), SARSA (94.77%), DDQN (96.89%), and PER-DQN (96.63%). PER-DQN, being th most efficient one with an interference time of 0.0005, was chosen. The reported inference time represents the average prediction time per test sample measured after model training under the same hardware configuration used for all benchmark models. The proposed PER-DQN + ECA-NET model achieved 97.24% accuracy, with 97.22% F1-score and 96.52% MCC. The hybrid model in question was able to push its macro-AUC score up to 99.86%, with each of the classes having their AUC value higher than 99% individually. The mean accuracy of the model obtained via five-fold cross-validation was 94.54%; the model was 0.001655 GFLOPs and contained 59, 529 parameters. The suggested system exhibited excellent results even after five separate runs, producing a mean accuracy of 97.21 ± 0.04% (95% CI: 97.21 ± 0.05%) and a macro-average AUC of **99.87 ± 0.01% (95% CI: 99.87 ± 0.01%). In fact, we’ve statistically validated the model through a Friedman test, confidence interval analysis, and an ablation study, which reveal improved ranking, stability, and the highest accuracy of the proposed CNN + ECA + PER-DQN model.