Scholarlink

<p>Conventional CNC machining systems suffer from spatial rigidity, limiting workpiece size and process integration. Mobile machining systems offer flexibility but are prone to chatter vibrations due to lower dynamic stiffness and variable cutting conditions. Chatter, being nonlinear and nonstationary, poses significant challenges for real-time detection. This study introduces a lightweight CNN-based chatter detection system tailored for mobile machining environments with limited computational resources. Milling data from a wheel-based mobile machine tool (MMT) were processed using variational mode decomposition (VMD) with Bayesian optimization to extract chatter modes. These were transformed into time–frequency images via wavelet synchro-squeezing transform (WSST) and classified using a transfer-learned SqueezeNet v1.1 model. The resulting system achieves 92% accuracy, distinguishes stable and chatter states with zero misclassification, and maintains a compact footprint of 2.8&#xa0;MB—making it ideal for embedded deployment in mobile machining platforms.</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

Real-time chatter recognition in the mobile milling systems using lightweight CNN

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月19日

查看全文
  • Donggeon Lee,
  • Chang Hyeon Mun,
  • Hyung Wook Park

摘要

Conventional CNC machining systems suffer from spatial rigidity, limiting workpiece size and process integration. Mobile machining systems offer flexibility but are prone to chatter vibrations due to lower dynamic stiffness and variable cutting conditions. Chatter, being nonlinear and nonstationary, poses significant challenges for real-time detection. This study introduces a lightweight CNN-based chatter detection system tailored for mobile machining environments with limited computational resources. Milling data from a wheel-based mobile machine tool (MMT) were processed using variational mode decomposition (VMD) with Bayesian optimization to extract chatter modes. These were transformed into time–frequency images via wavelet synchro-squeezing transform (WSST) and classified using a transfer-learned SqueezeNet v1.1 model. The resulting system achieves 92% accuracy, distinguishes stable and chatter states with zero misclassification, and maintains a compact footprint of 2.8 MB—making it ideal for embedded deployment in mobile machining platforms.