Neural networks with dual-view fusion feature learning mechanism: A method to improve micro-milling tool wear monitoring performance by enhancing feature generalization capabilities

Abstract

Tool wear monitoring methods can effectively monitor cutting tool condition, workpiece machining accuracy, and machined surface quality. Current tool wear monitoring models typically focus on repeatable sensitive features in the machining signal that are consistent with the wear trend, and can only monitor from a single view. However, in actual machining processes, especially in micro-milling where signal strength is weak and signals are not sensitive to wear, such a single-view monitoring approach struggles to handle the complex nonlinear variations in signals during actual machining. Therefore, this study innovatively proposes a tool wear monitoring model based on a dual-view fusion feature learning mechanism. The feature extraction capability of the model is enhanced by introducing a Siamese neural network. First, the Kolmogorov-Arnold neural network (KAN) is constructed to extract repeatable features that align with the tool wear trend from all signals, building the original feature vector. Secondly, the Siamese neural network is embedded to capture the feature differences between the current signal and the initial machining signal, and the feature difference vector is constructed. The feature difference vector effectively avoids interference of signals caused by sudden variation of cutting state and environment during the intermediate machining process, which enhances the differentiation of the machining signals under different wear states and provides a new training perspective for tool wear monitoring. Then, the original feature vectors are fused with the feature difference vectors to form a dual-view fusion feature with excellent generalization ability. The model not only captures global features in all signals, but also adaptively learns correlations and differences between signals, providing richer feature information for tool wear monitoring. To verify the superiority of the model, ablation experiments and micro-milling experiments are performed. The results show that the Siamese network effectively enhanced the feature extraction and generalization capabilities of models. The performance of models with dual-view features for tool wear monitoring is improved compared to standalone models. Among all models, the dual-view model based on the Siamese and the KAN network has the highest accuracy, with 97.95 % prediction accuracy. In addition, it is still highly accurate under new micro-milling conditions.