Tool Wear Prediction in Complex Machining Processes: A Hybrid Residual-Compensated Deep Learning Framework

Accurate tool wear prediction is essential for predictive maintenance in complex machining processes, but non-stationary sensor signals make it difficult for a single model to capture both long-term degradation trends and local transient disturbances. This study introduces a residual-compensated Hybrid CNN-Informer + LightGBM framework for tool wear prediction. The workflow first preprocesses multi-source sensor signals and selects wear-sensitive statistical descriptors to guide FiLM-based deep feature modulation. A CNN-Informer backbone then estimates the main wear trend by combining local feature extraction with long-range temporal modeling, and a LightGBM module performs secondary compensation on the remaining prediction residuals. On the PHM 2010 milling benchmark, the proposed framework achieved an RMSE of 4.0016, MAE of 2.8271, and R2 of 0.9870, reducing RMSE and MAE by 47.9% and 48.5% compared with a standard Transformer. Ablation results showed that both the CNN branch and residual compensation contributed to the final accuracy. External validation on the HMoTP dataset using an independent held-out tool further yielded an RMSE of 9.1111, MAE of 7.3522, and R2 of 0.9729. These results indicate that separating main-trend learning from residual correction provides a practical strategy for robust tool wear prediction under the tested machining conditions.