Scholarlink

<p>Accurate monitoring of cutting forces during milling machining is of paramount importance in the areas of tool wear detection, failure diagnosis, and process parameter optimization, especially for large and complex parts that undergo multiple machining operations. To address the challenge of acquiring labeled values for cutting forces modeling, a data-driven approach integrating convolutional neural network-long short-term memory-self-attention (CNN-LSTM-SA) is designed for predicting cutting forces during milling processes under variable cutting operations. This method leverages the inherent signals from computer numerical control monitoring and incorporates the calculated cutting forces from the mechanical force model. The acquisition of the force values using a dynamometer is not required, and the measured values only serve validation purposes. The CNN is specifically designed for localized feature extraction from the monitoring signal, while the LSTM network is employed to capture inherent long-term dependencies within the extracted features associated with the time series. In addition, this design eliminates the time lag between the machine monitoring signals and the calculated cutting forces signals. The self-attention layer focuses on the important information in the features that is relevant to the model prediction, thus improving prediction accuracy. The prediction outcomes under various cutting conditions demonstrate the efficacy of the designed methodology in adeptly extracting features from the monitoring signal, and accurately reflecting the cutting forces. The root mean square errors (RMSEs) between predicted and measured values for linear and circular trajectories are predominantly below 15% of the peak cutting forces, thus demonstrating the superior prediction accuracy achieved by the proposed model. For selected trajectories, the RMSEs are reduced by over 39.0% compared with the deep fully connected backward propagation neural network, and are reduced by more than 19.0% compared with the CNN-LSTM network, showing the superiority of the designed method.</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

Mechanism and data hybrid-driven cutting forces prediction model for end milling

  • 期刊论文

  • 发表日期：

    2025年1月7日

查看全文
  • Chang Ni,
  • Jixiang Yang,
  • Han Ding

摘要

Accurate monitoring of cutting forces during milling machining is of paramount importance in the areas of tool wear detection, failure diagnosis, and process parameter optimization, especially for large and complex parts that undergo multiple machining operations. To address the challenge of acquiring labeled values for cutting forces modeling, a data-driven approach integrating convolutional neural network-long short-term memory-self-attention (CNN-LSTM-SA) is designed for predicting cutting forces during milling processes under variable cutting operations. This method leverages the inherent signals from computer numerical control monitoring and incorporates the calculated cutting forces from the mechanical force model. The acquisition of the force values using a dynamometer is not required, and the measured values only serve validation purposes. The CNN is specifically designed for localized feature extraction from the monitoring signal, while the LSTM network is employed to capture inherent long-term dependencies within the extracted features associated with the time series. In addition, this design eliminates the time lag between the machine monitoring signals and the calculated cutting forces signals. The self-attention layer focuses on the important information in the features that is relevant to the model prediction, thus improving prediction accuracy. The prediction outcomes under various cutting conditions demonstrate the efficacy of the designed methodology in adeptly extracting features from the monitoring signal, and accurately reflecting the cutting forces. The root mean square errors (RMSEs) between predicted and measured values for linear and circular trajectories are predominantly below 15% of the peak cutting forces, thus demonstrating the superior prediction accuracy achieved by the proposed model. For selected trajectories, the RMSEs are reduced by over 39.0% compared with the deep fully connected backward propagation neural network, and are reduced by more than 19.0% compared with the CNN-LSTM network, showing the superiority of the designed method.