Edge device deployment for intelligent machine tools: A lightweight and interpretable tool wear monitoring method considering wear behavior

Tool wear condition monitoring is essential for reducing production costs and improving machining precision, serving as a key strategy for achieving machine tool intelligence. However, existing methods often depend on empirically designed complex networks to achieve high recognition accuracy, which results in high computational costs, poor performance during later wear stages, and limited interpretability. To address these challenges, a lightweight and interpretable tool wear recognition method is proposed. The feature self-adaptive reconstruction strategy based on wear behavior improves feature quality, while a nonlinear cumulative wear model provides physics guidance, ensuring the model remains lightweight and interpretable. To improve recognition accuracy and robustness during later wear stages, an adaptive loss adjustment mechanism driven by error uncertainty is proposed. Additionally, the influence of reconstructed features on model output is analyzed using shapley additive explanations (SHAP) values, while dependency graphs explore interactions between features across signals and domains, reinforcing physical interpretability. Results show reconstructed features in the y-direction have the greatest influence on model output during side milling. Time-domain and time-frequency domain features dominate, with frequency-domain features providing complementary information. Experiments show the proposed method reduces RMSE by 4.77 and 6.63, MAE by 2.89 and 5.17, and improves R² by 0.06 and 0.12 compared to models with different loss weights and feature processing methods. Recognition accuracy was further improved during later wear stages, achieving RMSE, MAE, and R² values of 3.58, 2.73, and 0.92, respectively. Moreover, the model uses only four fully connected layers, reducing parameters by over 9.32 times, demonstrating the feasibility of edge deployment.