Research on cross-condition tool wear monitoring methods based on multi-channel fused data and transfer learning

Effective tool wear monitoring is essential for machining precision and efficiency. Traditional methods perform poorly under changing operating conditions due to limited generalization. To address this issue, a novel approach combining multi-channel fused data and transfer learning is proposed. An extended Kalman filter (EKF) integrates features from multiple sensor signals to construct a comprehensive fused dataset, and a squeeze-and-excitation residual network (SE-ResNet) is developed to predict tool wear. Under new conditions, only minimal multi-channel fused data are required to fine-tune the fully connected (FC) layer, enabling cross-condition prediction transfer. Experiments conducted on the NASA milling and PHM high-speed milling datasets demonstrate that the proposed method significantly outperforms single-channel approaches, reducing mean error by 55.2% and 46.2%, respectively. Pearson correlation analysis confirms stronger relationships between the fused features and tool wear. Ablation studies highlight the critical contribution of the SE attention mechanism, and comparative experiments show superior accuracy and generalization over mainstream fusion and transfer learning methods. The proposed method offers robust cross-condition performance, indicating strong potential as a foundational technique for tool wear monitoring.