Spatial-temporal dual-stream fusion network with embedded knowledge: A process-generalized framework for tool wear monitoring

As a pivotal technology in intelligent computer numerical control (CNC) systems, tool wearing monitoring (TWM) significantly influences machining stability, product quality, and production efficiency. However, the complexity and randomness of the tool wear process pose significant challenges to traditional single-model methods. These methods often struggle to capture diverse data patterns, leading to low accuracy and poor processing generalization. This study proposes a Spatial-Temporal Dual Stream Fusion Network (STDSFnet) with embedded knowledge. The framework establishes strong correlations between multi-modal sensor data and tool wear states. The methodology consists of three key steps: (1) applying Spearman correlation analysis to select sensitive feature vectors, which helps reduce transient signal interference; (2) developing a dual-stream network to extract spatial and temporal wear-related information; and (3) integrating coordinated attention mechanisms and feature fusion modules to enhance the feature representation, thereby improve the TWM accuracy and processing generalization. Besides, to ensure that the trend of TWM results conforms to the actual wear pattern, domain knowledge is introduced to improve the reliability and interpretability of monitoring results. Experiments validate the effectiveness of STDSFnet on block-shaped stainless-steel workpieces and carbon fiber-reinforced polymer (CFRP) machining scenarios. Compared to baseline methods, STDSFnet reduces RMSE by 38.94 %–62.04 %, decreases MAE by 43.64 %–61.70 %, and improves R² by 4.96 %–28.59 %. These results confirm that spatial-temporal fusion with deep ensembles significantly boosts TWM accuracy and reliability.