<p>With the advancement of intelligent manufacturing, tool wear monitoring grows increasingly important. However, in milling operations, noise interference degrades signal quality, while traditional deep learning models demonstrate limited capability in capturing the inherent long-term temporal dependencies in tool wear. These issues collectively constrain monitoring accuracy. To address these issues, this paper proposes a novel monitoring model that integrates variational mode decomposition (VMD) with hierarchical adaptive wavelet thresholding to improve signal quality, and fuses an improved whale optimization algorithm (IWOA) with the gated recurrent unit (GRU). Through the synergistic action of multiple novel mechanisms, the IWOA achieves an excellent balance between global exploration and local exploitation, thereby enhancing its ability to optimize the GRU, while an attention mechanism is embedded to adaptively prioritize decisive temporal segments for wear progression modeling. Firstly, raw signals are processed through VMD-based denoising with hierarchical adaptive wavelet thresholding method and windowing. Subsequently, extracting multi-domain features from both time and frequency domains, then filtering highly correlated tool wear datasets via Pearson correlation analysis. Finally, inputting the filtered data into the IWOA-GRU-Attention model and validating reliability through three-fold cross-validation. Experiments on the PHM2010 dataset demonstrate that the proposed IWOA-GRU-Attention model achieves an average monitoring accuracy of 99.22%, with MAE, RMSE, and MSE values of merely 0.99, 1.38, and 1.94 respectively. These metrics represent average reductions of 86.71%, 85.10%, and 98.17% compared to other benchmark models. Additionally, the classification accuracy for tool wear states reached 98.7%. This study provides a robust support for high-precision tool wear monitoring technology.</p>