Scholarlink

<p>To address the issues of multi-fault feature coupling, single-domain feature incompleteness, and poor fixed-threshold adaptability in rolling bearing compound fault diagnosis, this paper proposes an intelligent method integrating multi-source feature fusion, three-stage feature screening, heterogeneous ensemble learning, and dynamic threshold optimization. First, time-domain, frequency-domain, and time–frequency domain features are fused to overcome the limitations of single-domain features. A three-stage screening strategy combining variance thresholding, mutual information, and recursive feature elimination is adopted to construct a low-redundancy, high-discriminability feature set. A heterogeneous ensemble learning framework integrating Convolutional Neural Network with Attention mechanism (CNN-Attention), Densely Connected Networks (DenseNet), and CNN-Long Short-Term Memory (CNN-LSTM) is established, with an improved attention-CNN as the final classifier. A dynamic threshold mechanism is introduced to adaptively adjust decision boundaries based on validation set statistics. On the Case Western Reserve University (CWRU) dataset, the proposed method achieves an average accuracy of 99.74%. Under no-load operation and 10&#xa0;dB SNR noise conditions on the laboratory platform, the accuracy exceeds 98%. Confusion matrix and t-SNE visualizations confirm good inter-class separability and intra-class clustering. The proposed method demonstrates favorable accuracy, adaptability, and anti-interference capability for rolling bearing compound fault diagnosis.</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

A Rolling Bearing Compound Fault Diagnosis Method Based on Ensemble Learning and Dynamic Threshold

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • Peng Liu,
  • Ziqiang Sun,
  • Lingyu Jiang,
  • Haoqi Zhang

摘要

To address the issues of multi-fault feature coupling, single-domain feature incompleteness, and poor fixed-threshold adaptability in rolling bearing compound fault diagnosis, this paper proposes an intelligent method integrating multi-source feature fusion, three-stage feature screening, heterogeneous ensemble learning, and dynamic threshold optimization. First, time-domain, frequency-domain, and time–frequency domain features are fused to overcome the limitations of single-domain features. A three-stage screening strategy combining variance thresholding, mutual information, and recursive feature elimination is adopted to construct a low-redundancy, high-discriminability feature set. A heterogeneous ensemble learning framework integrating Convolutional Neural Network with Attention mechanism (CNN-Attention), Densely Connected Networks (DenseNet), and CNN-Long Short-Term Memory (CNN-LSTM) is established, with an improved attention-CNN as the final classifier. A dynamic threshold mechanism is introduced to adaptively adjust decision boundaries based on validation set statistics. On the Case Western Reserve University (CWRU) dataset, the proposed method achieves an average accuracy of 99.74%. Under no-load operation and 10 dB SNR noise conditions on the laboratory platform, the accuracy exceeds 98%. Confusion matrix and t-SNE visualizations confirm good inter-class separability and intra-class clustering. The proposed method demonstrates favorable accuracy, adaptability, and anti-interference capability for rolling bearing compound fault diagnosis.