A tool wear prediction model integrating mechanism-data fusion with transfer learning for machine tool digital twins

Digital Twin enables high-fidelity data integration, dynamic state prediction, and adaptive decision-making in machining process. A robust tool wear monitoring (TWM) model is pivotal for ensuring data fidelity in digital twin systems and providing accurate evaluation of tool wear states. However, existing methods generally rely on idealized assumptions or high-quality datasets, while neglecting the physical wear mechanisms, exhibiting weak physical consistency and limited generalization under variable machining conditions. To overcome these challenges, this study proposes a TWM model that integrates the physics-informed Gaussian process regression with transfer learning within digital twin intelligent monitoring of milling process. Concretely, multi-domain features are extracted from monitoring signals and a high-correlation feature subspace is constructed by comprehensive correlation analysis. Then, a physics-based model is developed and embedded into the Gaussian process regression framework to provide prior knowledge of tool degradation. Furthermore, a two-stage transfer learning strategy is incorporated to facilitate effective cross-condition knowledge adaptation, significantly enhances the model's generalization capability. Finally, the proposed method is validated on the PHM2010 dataset and milling experiments. Experimental results indicate that the method achieves RMSE between 0.20 μm and 0.63 μm and MAE between 0.14 μm and 0.39 μm across varying machining conditions, confirming its prediction accuracy and robust cross-condition adaptability. The proposed model provides reliable tool wear data for digital twin monitoring systems, facilitating intelligent decision-making and maintenance.