Scholarlink

<p>The dynamic evolution of abrasive belt wear directly affects grinding efficiency and workpiece surface quality. Accurate and real-time monitoring of abrasive belt wear is therefore essential for optimising the grinding process and enhancing machining accuracy. This study proposes a 12-stage wear classification method based on the abrasive wear area ratio, enabling precise characterisation of the wear evolution process. Sound signals are employed as monitoring input and transformed into time–frequency spectrograms using the short-time Fourier transform, thereby avoiding the subjectivity and uncertainty of manual feature extraction. Additionally, a depthwise squeezed residual transformer module is designed, which integrates the local feature extraction capability of convolutional neural networks with the global information modelling capability of vision transformers, significantly boosting recognition accuracy while reducing computational costs. The effectiveness of the proposed method is verified through grinding experiments on GCr15 workpieces using a GXK51-B 60# abrasive belt. The results indicate that the method achieves an overall recognition accuracy of 97.7%. Even for fine-grained 177-stage wear state classification, the recognition accuracy remains at 89.8%. Under varying grinding parameters, the accuracy consistently exceeds 94%. These findings fully demonstrate the practicality, robustness, and broad application potential of the method for real-time abrasive belt wear monitoring.</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

Abrasive belt wear condition monitoring using sound signals: a fusion of convolutional neural networks and vision transformers

  • 期刊论文

  • 发表日期：

    2025年9月30日

查看全文
  • Wanjing Pang,
  • Guangpeng Zhang,
  • Lijuan Ren,
  • Nina Wang

摘要

The dynamic evolution of abrasive belt wear directly affects grinding efficiency and workpiece surface quality. Accurate and real-time monitoring of abrasive belt wear is therefore essential for optimising the grinding process and enhancing machining accuracy. This study proposes a 12-stage wear classification method based on the abrasive wear area ratio, enabling precise characterisation of the wear evolution process. Sound signals are employed as monitoring input and transformed into time–frequency spectrograms using the short-time Fourier transform, thereby avoiding the subjectivity and uncertainty of manual feature extraction. Additionally, a depthwise squeezed residual transformer module is designed, which integrates the local feature extraction capability of convolutional neural networks with the global information modelling capability of vision transformers, significantly boosting recognition accuracy while reducing computational costs. The effectiveness of the proposed method is verified through grinding experiments on GCr15 workpieces using a GXK51-B 60# abrasive belt. The results indicate that the method achieves an overall recognition accuracy of 97.7%. Even for fine-grained 177-stage wear state classification, the recognition accuracy remains at 89.8%. Under varying grinding parameters, the accuracy consistently exceeds 94%. These findings fully demonstrate the practicality, robustness, and broad application potential of the method for real-time abrasive belt wear monitoring.