Scholarlink

<p>To address the degradation in grinding accuracy and the difficulty of achieving precise workpiece dimensional control caused by grinding wheel wear, this study proposes a multi-attention-enhanced dual-branch temporal convolutional network (MAE-DBTCN) model. The model enables real-time prediction of grinding wheel wear and workpiece material removal using multi-sensor signals, thereby providing critical data for machine tool error compensation. By constructing a dual-branch heterogeneous TCN architecture for high-frequency vibration and low-frequency pressure signals, the proposed method reduces cross-sensor interference and improves prediction accuracy. The incorporation of cross-attention mechanisms and spatio-temporal joint modeling further enhances performance. Experimental results show a root mean square error of 0.0017, mean absolute percentage error of 3.86 %, mean absolute error of 0.001, and coefficient of determination of 0.9978. The single inference time ranges from 9 to 9.55 ms, confirming the effectiveness of the proposed architecture. This study also demonstrates realtime grinding wheel position compensation based on dual predictive outputs.</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 prediction method of grinding wheel wear and workpiece removal based on multi-attention-enhanced dual-branch TCN model

  • 期刊论文

  • 发表日期：

    2026年5月7日

查看全文
  • Jianhua Tang,
  • Gaotian Hong,
  • Li Jiang,
  • Yuanxuan Huang,
  • Yinjun Li,
  • Shaodai Huang,
  • Junhao Fu,
  • Jia Pan

摘要

To address the degradation in grinding accuracy and the difficulty of achieving precise workpiece dimensional control caused by grinding wheel wear, this study proposes a multi-attention-enhanced dual-branch temporal convolutional network (MAE-DBTCN) model. The model enables real-time prediction of grinding wheel wear and workpiece material removal using multi-sensor signals, thereby providing critical data for machine tool error compensation. By constructing a dual-branch heterogeneous TCN architecture for high-frequency vibration and low-frequency pressure signals, the proposed method reduces cross-sensor interference and improves prediction accuracy. The incorporation of cross-attention mechanisms and spatio-temporal joint modeling further enhances performance. Experimental results show a root mean square error of 0.0017, mean absolute percentage error of 3.86 %, mean absolute error of 0.001, and coefficient of determination of 0.9978. The single inference time ranges from 9 to 9.55 ms, confirming the effectiveness of the proposed architecture. This study also demonstrates realtime grinding wheel position compensation based on dual predictive outputs.