Abstract: To achieve accurate prediction of the flank wear of milling tools for titanium alloy and other difficult-to-machine materials, a tool wear prediction model based on multi-source attention fusion was proposed. The cutting force and vibration signals were converted into time-frequency images by continuous wavelet transform. A convolutional neural network embedded with channel attention was constructed to adaptively enhance key feature channels and extract local features, and the global dependencies were modeled by a focused-attention Swin Transformer. The bidirectional deep interaction between force and vibration features was realized through cross-attention, and the optimal signal branch was dynamically selected by an adaptive signal selection module according to the wear stage, after which the flank wear value was predicted by regression. The results show that, on the PHM2010 public dataset, an average coefficient of determination (R²) of 0.994 and a root mean square error (RMSE) of 2.740 μm were achieved by the proposed model, which outperformed CNN-BiLSTM, Informer, iTransformer and other models; on the self-built TC4 titanium alloy dataset, an R² of 0.967 was reached, and the cross-attention fusion was shown by ablation experiments to contribute most significantly. High prediction accuracy and good generalization capability under varying cutting parameters are offered by the proposed model, and a reference can thereby be provided for tool wear monitoring of difficult-to-machine materials.