Multi-Scale Dual-Attention Feature Network with Bidirectional Temporal Constraints for Tool Wear Monitoring

Accurate tool wear monitoring plays a decisive role in machining efficiency, product quality and reliability in modern manufacturing systems. Existing deep learning methods struggle to balance the high-frequency transient features and low-frequency evolution trends in tool wear signals, often losing key temporal evolution details when processing long-range degradation data. Therefore, this paper proposes an online prediction method of tool wear value that combines multi-scale convolution and dual-attention temporal features. This method extracts local mutation and trend features in wear signals through multi-scale convolution, captures wear evolution features through bidirectional cyclic network, and adaptively fuses local detail information and global trend through dual attention mechanism SWGC-DA to generate a multi-scale time series feature-driven prediction model. The ablation experiment based on the PHM2010 public data set verifies the effectiveness of the network structure design and demonstrates the model’s superior predictive ability. Experiments on the self-built TiAl alloy milling dataset achieved a stable prediction of R2 up to 99.1%, with MAE and RMSE of 2.29 and 2.47, respectively. The results show that this method significantly improves the accuracy and robustness of wear prediction.