Scholarlink

<p>To solve the problem of low fault diagnosis accuracy in bearing systems under strong noise, this paper proposes a deep learning framework based on weighted differential attention and Gabor-BiLSTM, called WGLformer. First, a Gabor-BiLSTM module is used at the input. The input signal is processed in parallel by learnable Gabor filters and Bidirectional Long Short-Term Memory (BiLSTM) network, and their outputs are summed and passed to the next layer. This module combines local frequency with global time features. Then, in deeper layers of the network, a Multi-Headed Weighted Differential Attention (W-MHDA) mechanism and an Refined Separable Multi-Scale Convolution Block (ReSMCB) are introduced. W-MHDA uses the concept of differential amplifier circuits to suppress noise and enhance features. ReSMCB combines Transformer structure with multi-scale techniques to extract global and local features. Finally, experiments on two datasets show this method achieves higher fault identification accuracy and robustness in high-noise environments.</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

WGLformer: a deep learning framework based on weighted differential attention and Gabor-BiLSTM for bearing fault diagnosis in noisy environments

  • 期刊论文

  • 发表日期：

    2025年10月7日

查看全文
  • Tingting Fang,
  • Dechen Yao,
  • Jianwei Yang,
  • Tao Zhou,
  • Yuteng Ren,
  • Bin Zhu

摘要

To solve the problem of low fault diagnosis accuracy in bearing systems under strong noise, this paper proposes a deep learning framework based on weighted differential attention and Gabor-BiLSTM, called WGLformer. First, a Gabor-BiLSTM module is used at the input. The input signal is processed in parallel by learnable Gabor filters and Bidirectional Long Short-Term Memory (BiLSTM) network, and their outputs are summed and passed to the next layer. This module combines local frequency with global time features. Then, in deeper layers of the network, a Multi-Headed Weighted Differential Attention (W-MHDA) mechanism and an Refined Separable Multi-Scale Convolution Block (ReSMCB) are introduced. W-MHDA uses the concept of differential amplifier circuits to suppress noise and enhance features. ReSMCB combines Transformer structure with multi-scale techniques to extract global and local features. Finally, experiments on two datasets show this method achieves higher fault identification accuracy and robustness in high-noise environments.