Vibra-FusionNet: A fusion framework based on vibration-adaptive decoupling and feature extraction for wear-state monitoring of micro-textured tools

Existing tool-wear monitoring methods often depend on experience-based preprocessing, show unstable prediction in wear-state transition regions, and have limited generalization under single-sensor conditions. To address these issues, this study proposes a hierarchical tool-wear monitoring framework, Vibra-FusionNet, for micro-textured tool milling of titanium alloys based on a single vibration source. The framework consists of front-end signal enhancement, middle-end feature screening, and back-end robust prediction. First, Northern Goshawk Optimization (NGO) is used to optimize the key parameters of Variational Mode Decomposition (VMD), and Wavelet Thresholding (WT) is introduced to adaptively denoise non-stationary vibration signals. Second, frequency-domain and wavelet-packet feature sets are constructed, and wear-sensitive features are selected using RReliefF weight ranking and cumulative contribution analysis. Finally, Transformer is employed to represent vibration-feature sequences, while Bayes-optimized Support Vector Regression (Bayes-SVR) is used for wear-value prediction and wear-state recognition. Furthermore, a fuzzy boundary tolerance criterion for adjacent wear stages is proposed, wherein the tolerance coefficient is defined as 5% of the wear value range. This criterion is employed to assess the stability of state recognition within the continuous wear transition region. The proposed method was validated using both the PHM2010 public dataset and a self-constructed milling vibration dataset. The latter was generated through dry side milling experiments on Ti6Al4V alloy, employing an AlSiTiN-coated, micro-textured cemented carbide ball-end milling cutter mounted on a three-axis CNC machine tool. Tool wear was quantified by measuring the flank wear width using an industrial camera. On the PHM2010 dataset, the MAE and RMSE are 2.5715 μm and 2.8156 μm, respectively, and the coefficient of determination reaches 0.995. On the self-built dataset, the MAE and RMSE are 2.8786 μm and 3.4225 μm, respectively, and the coefficient of determination reaches 0.993. The F1 scores on the two datasets reached 98.20% and 96.98%, respectively. These results demonstrate that Vibra-FusionNet achieves accurate wear prediction, stable wear-state recognition, and good cross-dataset adaptability under a single vibration source.