Recognition of tool wear in milling using Kolmogorov-Arnold networks and multi-sensor feature fusion

Existing methods for tool wear recognition using online signal processing and deep learning techniques typically utilize support vector machines (SVM) and fully connected layers (FCL). These methods inherently struggle with capturing complex nonlinear wear patterns due to their dependence on linear transformations and fixed-weight architectures. To overcome these limitations, this study introduces a novel tool wear recognition method based on empirical wavelet transform (EWT) and Kolmogorov-Arnold Network (KAN). Through EWT, tool wear-related component signals are adaptively extracted from multi-sensor data such as cutting forces, accelerations, and acoustic emission signals. To further enhance feature extraction, this study designs a multi-scale convolutional network (MSCN) and an efficient channel attention (ECA) mechanism. The MSCN is aimed at isolating wear-sensitive features, while the ECA mechanism highlights critical wear indicators, thereby avoiding issues such as modal mixing and energy leakage. Based on these advancements, KAN has finally been adopted to construct the tool wear recognition model. Experimental results prove the feasibility of extracting tool wear component signals through EWT. Leveraging this foundation, the proposed method achieves superior recognition accuracy and feasibility compared to existing approaches. The average recognition accuracy of the proposed model exceeds 99.9%, confirming its effectiveness in tool wear recognition.