Abstract: To address the challenges of poor adaptability, incomplete data acquisition, and weak generalization in existing tool wear monitoring methods under complex working conditions, an online monitoring approach that integrates a mechanistic model with a data-driven model is proposed. Firstly, a digital twin platform for data acquisition and virtual modeling is established to collect vibration signals and wear data in real time during machining. Secondly, a digital twin mechanistic model is developed from four dimensions, namely geometry, physics, rules, and behavior. Concurrently, a data-driven model is constructed by embedding a self-attention mechanism into a temporal convolutional network (TCN) to extract deep features from sensor signals. Finally, a particle filter algorithm fuses the mechanistic and data-driven models, which enhances the model's robustness and prediction accuracy in complex machining environments. Experimental results demonstrate that the wear predictions obtained by the proposed method are in close agreement with actual measurements, confirming its effectiveness and accuracy for wear monitoring.