Scholarlink

<p>Targeting the issues of low bearing fault identification accuracy and challenging vibration signal feature extraction, this study proposes an intelligent diagnostic method that combines improved multi-scale residual networks with Optuna-optimized LightGBM (OptGBM) for rolling bearing fault diagnosis. Firstly, the Gramian Angular Difference Field (GADF) is employed to convert one-dimensional vibration signals into two-dimensional representations, thereby preserving temporal information. Secondly, we combine the wavelet convolution with the depthwise separable convolution to develop Depthwise Separable Wavelet Convolution (DSWTC), which expands the network receptive field and reduces the computational overhead. In addition, DSWTC is innovatively introduced into ResNet to construct an improved residual network DSWTC-Res, which combines the advantages of ResNet in deep feature propagation and learning, and significantly improves the accuracy of the model in diagnosing bearing faults. A feature extractor integrating multi-scale DSWTC-Res, attention mechanism, and adaptive pooling is designed to effectively extract the discriminant features of vibration signals at different scales. Finally, the extracted deep features are input into OptGBM to achieve the high-precision classification of bearing faults. Experiments are conducted on three datasets from Case Western Reserve University (CWRU), Southeast University (SEU), and Xi'an Jiaotong University (XJTU). The fault diagnosis accuracy of the proposed method on these datasets reaches 99.31%, 99.64%, and 99.82%, respectively. The proposed method has excellent anti-noise performance and still has high detection accuracy even with small sample sizes.</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

Rolling bearing fault diagnosis method based on improved multi-scale residual networks and light gradient boosting machine

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • Linjun Wang,
  • Zijia Wang,
  • Zhenxiong Wu,
  • Xifa Yang,
  • Youxiang Xie

摘要

Targeting the issues of low bearing fault identification accuracy and challenging vibration signal feature extraction, this study proposes an intelligent diagnostic method that combines improved multi-scale residual networks with Optuna-optimized LightGBM (OptGBM) for rolling bearing fault diagnosis. Firstly, the Gramian Angular Difference Field (GADF) is employed to convert one-dimensional vibration signals into two-dimensional representations, thereby preserving temporal information. Secondly, we combine the wavelet convolution with the depthwise separable convolution to develop Depthwise Separable Wavelet Convolution (DSWTC), which expands the network receptive field and reduces the computational overhead. In addition, DSWTC is innovatively introduced into ResNet to construct an improved residual network DSWTC-Res, which combines the advantages of ResNet in deep feature propagation and learning, and significantly improves the accuracy of the model in diagnosing bearing faults. A feature extractor integrating multi-scale DSWTC-Res, attention mechanism, and adaptive pooling is designed to effectively extract the discriminant features of vibration signals at different scales. Finally, the extracted deep features are input into OptGBM to achieve the high-precision classification of bearing faults. Experiments are conducted on three datasets from Case Western Reserve University (CWRU), Southeast University (SEU), and Xi'an Jiaotong University (XJTU). The fault diagnosis accuracy of the proposed method on these datasets reaches 99.31%, 99.64%, and 99.82%, respectively. The proposed method has excellent anti-noise performance and still has high detection accuracy even with small sample sizes.