Milling Force Prediction Based on Spindle Current Signal

Spindle-current-based force estimation provides a nonintrusive alternative to dynamometer-based milling-force measurement, but its accuracy is limited by the nonlinear and time-dependent relationship between spindle current and cutting force. This study proposes a CNN–ResNet–RF model for instantaneous milling-force prediction using only spindle current signals as input. In the proposed architecture, CNN layers extract local temporal features from windowed current sequences, residual blocks refine multiscale force-related representations, and a random-forest regressor performs nonlinear force regression. Milling experiments were conducted on 7075 aluminum alloy and steel 45 using a five-axis machining center. To prevent temporal data leakage, the synchronized and preprocessed current–force data were divided at the continuous cutting trial level into training, validation, and independent test subsets. On the independent test subset, the proposed model achieved an R2 value of 0.952, an MAE of 2.793 N, and an RMSE of 4.301 N, outperforming the CNN, CNN–ResNet, and RF baseline models in terms of prediction accuracy and error reduction. These results demonstrate that the CNN–ResNet–RF framework improves test-set milling-force prediction within the tested machining range.