Tool Wear Monitoring in Robotic Milling Based on a Physics-Informed Data Fusion Model

To address tool wear monitoring in industrial robotic milling, a fusion-based monitoring method was proposed. The method was termed the Physics-Informed Transformer–Kolmogorov–Arnold Network (PI-Transformer-KAN). First, a physical mapping model between tool wear and marginal spectral amplitude was established through feature correlation analysis. A coefficient of determination greater than 0.9772 was achieved. Second, the linear layers of the Transformer were replaced with a Kolmogorov–Arnold Network (KAN), allowing the output weights to be dynamically adjusted. Monotonicity constraints were further introduced as a physics-based loss function. This constraint enabled the predicted wear trend to better conform to the actual evolution of tool wear. In addition, physical information and data-driven features were fused using a nonlinear regression network to guide and improve the prediction process. Robotic milling experiments were conducted to evaluate the proposed method. Compared with the data-driven model, the proposed method reduced the root mean square error (RMSE), mean absolute error (MAE), and width of the 95% confidence interval by 30.70%, 28.82%, and 30.08%, respectively.