Physics-informed orthogonal network with hierarchical time-frequency feature refining strategy for tool wear recognition

Abstract

Tool wear recognition is critical to improve the safety and reliability of machining operations with real-time tool status assessment. Conventional deep learning-based (DL-based) recognition approaches map time-frequency representations (TFRs) of the monitoring signals to tool wear with neural networks. However, data-driven mapping suffers from hardships in excavating wear-sensitive information due to the lack of explicit constraints on wear mechanisms, resulting in inferior recognition performance and recognition results against physical laws. To address this issue, this paper develops a time-frequency refining physics-informed orthogonal network (TFRPION). Firstly, a hierarchical time-frequency refining strategy consisting of energy concentration and adaptive amplitude modulation is conducted to emphasize machining dynamics-related characteristic frequency components in TFRs, highlighting wear-sensitive signal features. Secondly, an orthogonal network module maps features from the refined TFRs by capturing temporal amplitude variations within characteristic frequency components, improving the physical representational capability for the TFR mapping process. Thirdly, a physical network module and a modified loss function that take wear mechanisms into account are integrated to regularize the calculation path and optimization of the proposed network, enhancing the physical consistency between its mapping process and wear mechanisms. The feasibility and effectiveness of the proposed network are verified with collected spindle current signals in high-speed milling tool wear recognition experiments.