Scholarlink

In industrial environments, vibration signals collected from rolling bearings often contain noise, and traditional Convolutional Neural Networks (CNN) exhibit poor fault feature extraction capabilities with noisy data, thereby affecting fault diagnosis accuracy. This paper proposes a Manhattan self-attention and residual neural network (MaSA-ResNet) based algorithm for bearing fault diagnosis. The algorithm incorporates Discrete Wavelet Transform (DWT) for signal denoising and feature extraction, introducing the MaSA to enhance effective feature weights and improve prediction accuracy. Additionally, Leaky rectified linear unit (Leaky-ReLU) activation functions are applied within the residual blocks to address neuron death and information loss issues. The improved residual block unit consists of 22 convolutional layers, 20 normalization layers, and 9 activation functions. This paper utilizes the bearing dataset from the Case Western Reserve University (CRWU) laboratory, with various levels of Gaussian white noise added to simulate real industrial environments. Finally, comparative experiments on test sets with CNN, SVM, and the proposed algorithm were further conducted. The results show that the proposed algorithm achieves a fault recognition rate of 99.71%, outperforming both CNN and SVM, thus demonstrating its robustness to noise and fault diagnosis accuracy.

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

Fault Diagnosis of Rolling Bearings Based on Manhattan Self-attention and Residual Neural Network

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yong Lin,
  • Yaming Liu,
  • Jiaxin Ding,
  • Pengzhe Xu,
  • Haoyu Ma,
  • Ligang Yao

摘要

In industrial environments, vibration signals collected from rolling bearings often contain noise, and traditional Convolutional Neural Networks (CNN) exhibit poor fault feature extraction capabilities with noisy data, thereby affecting fault diagnosis accuracy. This paper proposes a Manhattan self-attention and residual neural network (MaSA-ResNet) based algorithm for bearing fault diagnosis. The algorithm incorporates Discrete Wavelet Transform (DWT) for signal denoising and feature extraction, introducing the MaSA to enhance effective feature weights and improve prediction accuracy. Additionally, Leaky rectified linear unit (Leaky-ReLU) activation functions are applied within the residual blocks to address neuron death and information loss issues. The improved residual block unit consists of 22 convolutional layers, 20 normalization layers, and 9 activation functions. This paper utilizes the bearing dataset from the Case Western Reserve University (CRWU) laboratory, with various levels of Gaussian white noise added to simulate real industrial environments. Finally, comparative experiments on test sets with CNN, SVM, and the proposed algorithm were further conducted. The results show that the proposed algorithm achieves a fault recognition rate of 99.71%, outperforming both CNN and SVM, thus demonstrating its robustness to noise and fault diagnosis accuracy.