Scholarlink

<p>Tool wear strongly affects machining quality, cost and efficiency, so accurate wear prediction is essential for reliable production. However, purely data-driven “black-box” models may ignore underlying physical laws and yield physically inconsistent results. To address this, we propose a physics-guided deep learning method that uses the signals of each completed milling pass to predict the instantaneous flank wear at the end of that pass. First, the collected one-dimensional cutting force signals are preprocessed by removing invalid head and tail segments and applying a Hampel filter to suppress impulsive noise. Then, a hybrid model combining convolutional neural networks (CNN) and bidirectional gated recurrent units (BiGRU) is constructed to efficiently capture coupled spatial–temporal features in the time-series data. In addition, a monotonicity-aware loss function is designed based on the non-decreasing law of tool wear and used as a soft physical constraint during training to guide the predictions towards physically consistent trends. Experimental results show that the proposed CNN–BiGRU model significantly outperforms single CNN and BiGRU models in prediction accuracy, and that introducing the physical constraint further reduces the mean absolute error by up to 29.68% on the C4 test set compared with the CNN–BiGRU baseline.</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 prediction model integrating physical constraints and CNN-BiGRU

  • 期刊论文

  • 发表日期：

    2026年5月23日

查看全文
  • Long Li,
  • Nan Zhang,
  • Zhenyu Liu,
  • Jie Feng,
  • Peng Duan

摘要

Tool wear strongly affects machining quality, cost and efficiency, so accurate wear prediction is essential for reliable production. However, purely data-driven “black-box” models may ignore underlying physical laws and yield physically inconsistent results. To address this, we propose a physics-guided deep learning method that uses the signals of each completed milling pass to predict the instantaneous flank wear at the end of that pass. First, the collected one-dimensional cutting force signals are preprocessed by removing invalid head and tail segments and applying a Hampel filter to suppress impulsive noise. Then, a hybrid model combining convolutional neural networks (CNN) and bidirectional gated recurrent units (BiGRU) is constructed to efficiently capture coupled spatial–temporal features in the time-series data. In addition, a monotonicity-aware loss function is designed based on the non-decreasing law of tool wear and used as a soft physical constraint during training to guide the predictions towards physically consistent trends. Experimental results show that the proposed CNN–BiGRU model significantly outperforms single CNN and BiGRU models in prediction accuracy, and that introducing the physical constraint further reduces the mean absolute error by up to 29.68% on the C4 test set compared with the CNN–BiGRU baseline.