<p>Accurately predicting tool wear is essential for maintaining high machining quality. Currently, deep learning models are extensively utilized in predicting tool wear. However, purely data-driven deep learning models are prone to local optima, and the limited tool wear sample data along with multi-sensor feature fusion also limit the models’ ability to generalize and extract valid information. To solve the above problems, a semi-supervised milling cutter wear prediction method based on the empirical formula for cutting force wear is proposed in this paper, with the model (IRM-CFAM) consisting of an Inception-ResNet module (IRM) and a channel feature adaptation module (CFAM). By introducing an empirical formula for cutting force wear, the cutting forces in the unlabeled samples are input to estimate wear values and construct wear curves. Based on the constructed wear curves, a monotonic loss function guided by an empirical curve of milling cutter wear is proposed. Meanwhile, the estimated wear values are combined with unlabeled samples to form a second data source, which is then merged with the labeled samples to expand the dataset. The constructed CFAM consists of channel attention and cross-attention mechanisms, which achieves adaptive fusion of weights. Following validation on the PHM2010 milling cutter dataset, the proposed method attained average RMSE and MAE values of 6.693 and 5.136, respectively, across three test sets, showing a significant advantage over other methods. This research facilitates the accurate prediction of milling cutter wear and introduces a novel approach for expanding sample data.</p>