<p>Tool wear state monitoring is of significant importance for ensuring machining accuracy, enhancing production efficiency, and reducing manufacturing costs. In practical machining processes, due to the complex and variable machining conditions coupled with multifactor interactions, existing physics-driven and data-driven methods encounter limitations in accuracy and generalization in monitoring tool wear states under diverse machining conditions. To address the aforementioned issues, this paper proposes a novel hybrid knowledge-based deep learning method for tool wear state monitoring. Initially, knowledge from the physical model assists wear label expansion. Subsequently, a modified Fisher score method, which guarantees the minimized intra-class distances and maximized inter-class distances, is used for feature extraction. Then, dimension reduced features are integrated with machining parameters to form fusion features, which are imported into a deep neural network to obtain the observation probability matrix of the hidden Markov model. Finally, the proposed method establishes a classifier for tool wear state monitoring. The proposed method significantly improves accuracy, particularly under varying machining conditions, by leveraging both physics knowledge and machine learning. Unlike traditional methods, which struggle with generalization across different environments, this approach adapts effectively to diverse conditions, offering reliable performance in real-world industrial settings. Experimental results demonstrate an outstanding average accuracy of 97.15% and robustness across multiple operating scenarios, making it suitable for practical applications in modern machining processes.</p>