A hybrid data-driven and physics-informed model for tool wear recognition

Tool wear recognition plays a vital role in enhancing the safety and reliability of machining processes through real-time monitoring and evaluation of tool conditions. Traditional deep learning (DL)-based recognition methods typically utilize neural networks to establish mappings between time-frequency representations (TFRs) of monitoring signals and tool wear values. However, these data-driven approaches face significant challenges in effectively extracting wear-sensitive features, as they lack explicit guidance in wear mechanisms. To address this issue, this paper proposed a hybrid data-driven and physics-informed tool wear recognition model. Firstly, a time-frequency attention network is proposed to extract wear-sensitive features from TFRs. This network can dynamically perceive temporal amplitude variations within characteristic frequency components within TFRs. Secondly, a physical network is developed to regularize the mapping process of the whole model and enhance the mapping processes physical consistency. The architecture of the physical network is designed based on a representative wear mechanism. The proposed method is validated with collected spindle current signals in milling tool wear recognition experiments.