Physics-constrained domain-decoupled transfer method: Application to cross-condition tooth-wise monitoring in multi-tooth tools
Bowen Zhang, Xianli Liu, Caixu Yue, Steven Y. Liang, Lihui Wang, Tianxiang Zhou
Abstract
In the machining process, tool wear monitoring (TWM) technology is crucial for ensuring machining safety and optimizing production costs. However, due to tool manufacturing precision, workpiece material characteristics, and tool runout, multi-tooth tools often exhibit uneven wear among individual cutting edges. Existing monitoring methods primarily focus on overall tool wear, lacking precise evaluation of individual cutting-edge wear, which hinders a detailed understanding of tool wear behaviour. Moreover, current monitoring approaches fail to incorporate the physical mechanism of edge-wise cutting processes into transfer learning, limiting the adaptability of models in multi-condition scenarios. To this end, this paper proposes a physics-constrained transfer learning network for multi-condition tooth-wise monitoring (PCTLNet). By embedding a physics-informed model, the proposed network implicitly learns and captures tool wear characteristics, ensuring that the learning process of tooth-wise wear features adheres to physical laws in a self-consistent manner. Specifically, the network is based on the DenseNet architecture and integrates a multi-modal one-dimensional convolutional neural network (1D-CNN) with a multi-head self-attention mechanism (MHSA), effectively extracting static features from cutting parameters and dynamic features from time-series signals. This paper proposes an adversarial high-order moment alignment mechanism to reduce domain discrepancies and improve the model’s adaptability under disturbances across different machining conditions. Furthermore, a multi-axis machining force model considering tooth-wise wear is established to reinforce physical consistency in the transfer learning process and ensure conformity with cutting physics. Based on this model, a novel physics-constrained domain adaptation strategy is proposed to enhance the network’s ability to learn tooth-wise wear features in the target domain. The effectiveness of the proposed method is validated through multi-condition milling wear experiments, demonstrating its advantages over existing state-of-the-art models in tooth-wise monitoring under varying machining conditions.