A spatio-temporal feature fusion network based on deep learning for collaborative prediction of tool wear and machining precision

In precision metal cutting processes, tool wear is a critical factor constraining machining accuracy and surface quality. Conventional monitoring approaches predominantly rely on offline measurements of macroscopic wear parameters, which often fail to capture the dynamically evolving micro-topographical characteristics and spatio-temporal correlations inherent in the wear process. This study introduces a deep learning framework called the Spatial–Temporal Feature Fusion Network (STF-Net). Unlike generic module stacking, STF-Net adopts a problem-driven, collaboratively embedded architecture tailored to tool wear monitoring. Specifically, STF-Net implements a task-specific feature decomposition strategy: the CNN front-end is explicitly designed to extract wear-sensitive spatial signatures (e.g., cutting-edge micro-geometry, crater morphology), while the LSTM back-end captures the temporal evolution of process parameters (e.g., cutting force, temperature)—a division grounded in wear physics. These streams are then dynamically fused through an attention mechanism that adaptively weights features based on current machining conditions (e.g., prioritizing thermal responses during high-speed cutting). This integrated framework enables synergistic optimization of wear state identification and accuracy prediction within a unified multi-task learning objective. Experimental results demonstrate that STF-Net significantly outperforms standalone CNN or LSTM models in two key tasks: tool wear state identification and machining accuracy prediction. For tool wear state classification (differentiating between mild, moderate, and severe wear), STF-Net achieves an average identification rate of 83.5%. For machining accuracy prediction—evaluated as the proportion of predictions within ± 2 μm error—the proposed model improves performance metrics (accuracy, precision, ROC-AUC) by an average of over 8% compared to baseline models. These findings confirm that the proposed spatio-temporal feature fusion strategy effectively enhances holistic perception and predictive capability in complex machining processes. This work provides a data-driven paradigm for synchronized online tool wear monitoring and machining precision control, offering potential applications in predictive maintenance and quality assurance for intelligent manufacturing systems.