Abstract: In the milling process of critical aero-engine components such as blades and casings, the rapid tool wear caused by difficult-to-cut materials like titanium alloys and nickel-based superalloys significantly affects the geometric accuracy and surface quality of workpieces. To address the limitations of traditional end-to-end black-box models, including poor interpretability and weak cross-working condition generalization capability, a mechanism-data fusion tool wear prediction model is proposed. Firstly, a physically meaningful tool wear prediction mechanism model is established by developing a cutting force model considering tool wear and a power-based cutting force prediction model. Secondly, for the multi-dimensional information in machine tool processing, time-frequency domain processing and mutual information screening methods are employed to extract features highly correlated with tool wear and construct data feature vectors. Then, a mechanism-data fusion gated recurrent unit neural network model (MDF-GRU) is designed based on GRU model, and through the construction of a physically constrained loss function, the model optimization is achieved within the physical consistency space. Machining experiments demonstrate that the proposed model reduces the root mean square error and mean absolute error by 71.5% and 68.7% respectively compared to traditional methods, significantly improving prediction accuracy. This research provides a novel technical solution for tool wear prediction in difficult-to-cut material machining.

Key words: tool wear, mechanism-data fusion, sensing signal characteristics, loss function, neural network