<p>Accurate tool wear state monitoring is essential for enhancing machining efficiency, prolonging tool life, and ensuring precision in smart machining systems. However, existing methods face significant challenges: physics-based methods often struggle to adapt to nonlinear and dynamic machining environments, while data-driven methods lack physical consistency and require extensive labeled data. To address these limitations, this study proposes a deep learning method with embedded physical knowledge for tool wear state monitoring. By integrating physical domain knowledge into a data-driven framework, the proposed method balances prediction accuracy with physical consistency. The model incorporates several innovations: a physics-guided loss function embeds the physical mechanisms of tool wear into the learning process; periodic and segmented attention mechanisms improve temporal feature extraction from cutting force signals; a data augmentation strategy leverages physics-based simulations to generate diverse and physically consistent datasets, enriching the model’s training data; and architectural improvements, such as monotonicity constraints, further enhance the model’s robustness and alignment with real-world tool wear behavior. Experimental results demonstrate that the proposed method significantly outperforms baseline approaches. It achieves a 74.1% improvement in prediction accuracy and maintains strong robustness, as further validated through experiments conducted under mildly varying machining conditions. This work presents a scalable and efficient solution for tool wear state monitoring in smart machining systems, bridging the gap between physics-based and data-driven paradigms.</p>