Scholarlink

<p>Addressing limitations in existing tool wear condition recognition methods concerning feature extraction, vanishing gradients in deep networks, and multi-scale signal adaptability, this study proposes a method combining multi-domain feature extraction with a Multi-scale Residual Network. In this approach, multi-channel signals such as cutting force and vibration are collected to extract 238-dimensional time-domain, frequency-domain, and time-frequency-domain features, forming the input feature set. The Multi-scale Residual Network employs a 3 × 1/5 × 1/7 × 1 parallel convolutional branch with a multi-level residual connection structure. It captures wear features at different scales through multi-scale convolutional kernels and mitigates the vanishing gradient problem in deep networks via the residual structure, enabling feature extraction and propagation. Experimental results show that the proposed method achieves an average recognition accuracy of 98.41% on the PHM2010 dataset under within-tool evaluation, and an average accuracy of 93.68% under leave-one-tool-out cross-tool validation. Ablation studies further verify the effectiveness of the multi-scale design and residual learning components. External validation on the MWMM dataset demonstrates an accuracy of 92.48% on an unseen tool.</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

Tool wear condition recognition using multi-scale residual networks based on feature fusion

  • 期刊论文

  • 发表日期：

    2026年4月10日

查看全文
  • Yaqin Tian,
  • Xiaowei Liu,
  • Jiepeng Wang,
  • Zengshun Wang,
  • Lidong Ma

摘要

Addressing limitations in existing tool wear condition recognition methods concerning feature extraction, vanishing gradients in deep networks, and multi-scale signal adaptability, this study proposes a method combining multi-domain feature extraction with a Multi-scale Residual Network. In this approach, multi-channel signals such as cutting force and vibration are collected to extract 238-dimensional time-domain, frequency-domain, and time-frequency-domain features, forming the input feature set. The Multi-scale Residual Network employs a 3 × 1/5 × 1/7 × 1 parallel convolutional branch with a multi-level residual connection structure. It captures wear features at different scales through multi-scale convolutional kernels and mitigates the vanishing gradient problem in deep networks via the residual structure, enabling feature extraction and propagation. Experimental results show that the proposed method achieves an average recognition accuracy of 98.41% on the PHM2010 dataset under within-tool evaluation, and an average accuracy of 93.68% under leave-one-tool-out cross-tool validation. Ablation studies further verify the effectiveness of the multi-scale design and residual learning components. External validation on the MWMM dataset demonstrates an accuracy of 92.48% on an unseen tool.