Scholarlink

<p>As a core component of modern industrial equipment, the health status of rolling bearings directly impacts the reliability and safety of mechanical systems. Although deep learning has achieved notable progress in bearing fault diagnosis, existing models still struggle to concurrently achieve high classification accuracy, maintain a lightweight architecture with low computational complexity, and exhibit strong robustness under noisy environments. In this paper, a bearing fault diagnosis method based on Mixed Knowledge Distillation (MKD) is proposed to enable effective knowledge transfer from the teacher model to the student model. First, to address the limitations of traditional DenseNet in multi-scale modeling and connection redundancy, a teacher model named DualDenseFormer is designed with multi-scale dense blocks and a Vision Transformer (ViT). The teacher model achieves collaborative global-local feature modeling through dynamic noise suppression and enhanced feature representation, which reduces model redundancy and improves robustness in noisy environments. Second, a lightweight student model architecture, TinyDenseFormer, is designed with reduced depth and improved residual depthwise separable convolution (RES-DWConv), informed by the structural characteristics of the teacher model. The model achieves significant compression in both parameters and computational cost, with feature adapters in the distillation process enabling effective feature space alignment between teacher and student representations. Finally, to further enhance distillation effectiveness, the MKD approach incorporates hard target distillation, soft target distillation, intermediate layer feature distillation, and attention distillation, where the weighting of these strategies is optimized with bayesian optimization. It facilitates multi-strategy collaborative learning and comprehensively guides the student model to learn the deep semantic representations of the teacher model. Experimental validation on the CWRU, PU, and MFPT bearing datasets demonstrates average diagnostic accuracies of 99.81%, 91.50%, and 92.81%, respectively.</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

Bearing fault diagnosis with knowledge distillation between DualDenseFormer and TinyDenseFormer

  • 期刊论文

  • 发表日期：

    2026年2月11日

查看全文
  • Wei Lin,
  • Xinlong Zhao

摘要

As a core component of modern industrial equipment, the health status of rolling bearings directly impacts the reliability and safety of mechanical systems. Although deep learning has achieved notable progress in bearing fault diagnosis, existing models still struggle to concurrently achieve high classification accuracy, maintain a lightweight architecture with low computational complexity, and exhibit strong robustness under noisy environments. In this paper, a bearing fault diagnosis method based on Mixed Knowledge Distillation (MKD) is proposed to enable effective knowledge transfer from the teacher model to the student model. First, to address the limitations of traditional DenseNet in multi-scale modeling and connection redundancy, a teacher model named DualDenseFormer is designed with multi-scale dense blocks and a Vision Transformer (ViT). The teacher model achieves collaborative global-local feature modeling through dynamic noise suppression and enhanced feature representation, which reduces model redundancy and improves robustness in noisy environments. Second, a lightweight student model architecture, TinyDenseFormer, is designed with reduced depth and improved residual depthwise separable convolution (RES-DWConv), informed by the structural characteristics of the teacher model. The model achieves significant compression in both parameters and computational cost, with feature adapters in the distillation process enabling effective feature space alignment between teacher and student representations. Finally, to further enhance distillation effectiveness, the MKD approach incorporates hard target distillation, soft target distillation, intermediate layer feature distillation, and attention distillation, where the weighting of these strategies is optimized with bayesian optimization. It facilitates multi-strategy collaborative learning and comprehensively guides the student model to learn the deep semantic representations of the teacher model. Experimental validation on the CWRU, PU, and MFPT bearing datasets demonstrates average diagnostic accuracies of 99.81%, 91.50%, and 92.81%, respectively.