Multi-condition milling cutter wear prediction based on split-channel information re-fusion and domain adaptation

Abstract

Accurate tool wear prediction is crucial for ensuring superior quality and operational efficiency in cutting processes, avoiding part defects, and minimizing economic losses. However, due to variations in cutting parameters such as cutting speed, feed rate, and depth of cut under different working conditions, the collected sensor signals (e.g., current, acceleration, acoustic emission) exhibit significant differences in amplitude, spectral density, and temporal feature distribution. This results in a distribution shift of the signal data, making it difficult for traditional models to generalize across multiple working conditions and leading to a notable decline in prediction performance. As a result, traditional models struggle to adapt to these changes and fail to capture wear patterns under diverse conditions, leading to low prediction accuracy across varying working scenarios. Therefore, a multi-condition wear prediction method for the milling cutter is proposed in this paper. This method is based on split-channel information re-fusion and domain adaptation, utilizing cutting signal and machining process data. A split-channel information re-fusion module is proposed to align the spatial wear feature distribution across different working conditions. The module first separates the multidimensional cutting signal and processes data along channel directions. Then, different convolutions for heterogeneous feature extraction are applied, and the multidimensional correlated features along the channels are fused for re-extraction. This process ensures that the diverse working condition features are fully captured and aligned. Moreover, the Mamba is improved due to the inconsistent distribution of temporal wear features under different working conditions. Furthermore, the GLMamba is constructed to fuse global–local attention in parallel configuration with selective state space models (SSMs). The loss of local information during the global compression of the selective SSMs is avoided by updating global information with local information features, ensuring accurate temporal feature extraction and alignment. The Maximum Mean Discrepancy (MMD) algorithm is utilized to quantify the distributional differences between domains in the fully connected layer, aligning feature distributions between the source and target domains. Experimental results on the NASA and Nanjing University of Aeronautics and Astronautics (NUAA) datasets demonstrate that the proposed method achieves significant performance advantages under various working conditions. Specifically, the average RMSE and MAE on the NASA dataset are reduced to 0.0311  mm and 0.0253  mm, respectively, while on the NUAA dataset they are reduced to 0.0058  mm and 0.00425  mm, outperforming several mainstream benchmark models. This study provides a novel solution for predicting milling cutter wear under various working conditions.