Swin-fusion: An adaptive multi-source information fusion framework for enhanced tool wear monitoring

Abstract

Tool wear directly impacts product quality, manufacturing costs, and machining efficiency, serving as a critical factor in digital manufacturing. Existing prediction methods based on singular signals or limited features face constraints in predictive accuracy and generalizability. To address these limitations, this research proposes Swin-fusion, a multi-source information fusion framework integrating convolutional neural networks (CNNs) and Transformers. The framework innovates through an integrated CNN-Transformer architecture that enables efficient local-global feature extraction using focused attention mechanisms for individual signal processing, complemented by cross-attention based fusion for comprehensive multi-sensor information integration and adaptive feature selection for dynamic wear state monitoring. The effectiveness of the proposed approach was validated using both the public PHM2010 dataset and a self-constructed TiAl milling dataset. In tool wear life prediction, Swin-fusion achieves a mean absolute error (MAE) of 1.78 and root mean square error (RMSE) of 2.71 on PHM2010, and an MAE of 2.07 and RMSE of 3.21 on the TiAl dataset, with a coefficient of determination (R²) reaching 0.995. In tool wear state identification, the F1-score attains 98.8 % on PHM2010 and 98.0 % on TiAl. Results demonstrate that Swin-fusion markedly enhances predictive accuracy, identification precision, and generalization ability for practical tool wear monitoring applications.