<p>To address the low accuracy and poor generalization ability of traditional milling-force prediction methods, this study proposes a multi-source information fusion milling force prediction model based on an improved CNN-BiLSTM-MHA framework. To overcome the limitations of conventional CNN-BiLSTM models in processing multi-source heterogeneous data, an improved CNN-BiLSTM-MHA model is proposed. The model uses CNN to extract spatial features and BiLSTM to capture temporal dependencies, and it introduces a multi-head attention mechanism (MHA) to fuse different information sources effectively. The experimental results show that, compared with the traditional CNN-BiLSTM baseline model, the proposed CNN-BiLSTM-MHA model reduces the average relative error from 5.1% to 3.8%, corresponding to an error reduction of approximately 25.5%. In addition, under cutting-speed variation conditions, the increase in prediction error of the proposed model is only 2.8%, indicating improved cross-condition adaptability. Besides, this study explored the influence of different information sources on prediction accuracy and found that tool wear and changes in workpiece temperature significantly affect the prediction results. This system can effectively improve machining efficiency, reduce tool wear, and lower production costs. This study provides a new approach for high-precision milling force prediction in intelligent manufacturing, which has theoretical and practical significance for improving machining quality and efficiency.</p>