Scholarlink

<p>Despite the widespread adoption of convolutional neural networks (CNNs) in bearing fault diagnosis due to their superior nonlinear feature extraction capabilities, the inherent opacity of their end-to-end architecture raises interpretability concerns. To address the mismatch between conventional wavelets (e.g., Morlet, Laplace) and actual bearing fault pulse responses, this study proposes a novel softmax-activated bi-damped wavelet and establishes a physics-informed wavelet convolution layer. We further develop BWKNet – an interpretable model architecture integrating the proposed bi-damped wavelet convolution layer with a ResNet18 backbone enhanced by a plug-and-play lightweight local attention module. Comprehensive evaluations demonstrate the model’s effectiveness, achieving recognition accuracies of 99.86 % on the Case Western Reserve University dataset and 98.15 % on the Nanchang Railway Bureau dataset. The proposed model combines physical interpretability through customized wavelet convolution layer design with data-driven learning capabilities, supported by systematic interpretation via priori empowerment and attributional explanations. This work provides a solution for fault diagnosis interpretability.</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

BWKNet: Toward an interpretable neural network with bi-damped wavelet for fault diagnosis of rolling bearing

  • 期刊论文

  • 发表日期：

    2025年10月8日

查看全文
  • Hongxue Bi,
  • Mingkang Zhang,
  • Yiwen Xiao

摘要

Despite the widespread adoption of convolutional neural networks (CNNs) in bearing fault diagnosis due to their superior nonlinear feature extraction capabilities, the inherent opacity of their end-to-end architecture raises interpretability concerns. To address the mismatch between conventional wavelets (e.g., Morlet, Laplace) and actual bearing fault pulse responses, this study proposes a novel softmax-activated bi-damped wavelet and establishes a physics-informed wavelet convolution layer. We further develop BWKNet – an interpretable model architecture integrating the proposed bi-damped wavelet convolution layer with a ResNet18 backbone enhanced by a plug-and-play lightweight local attention module. Comprehensive evaluations demonstrate the model’s effectiveness, achieving recognition accuracies of 99.86 % on the Case Western Reserve University dataset and 98.15 % on the Nanchang Railway Bureau dataset. The proposed model combines physical interpretability through customized wavelet convolution layer design with data-driven learning capabilities, supported by systematic interpretation via priori empowerment and attributional explanations. This work provides a solution for fault diagnosis interpretability.