Hybrid Global and Local Neural Network With Lightweight Multiscale Fusion for Tool Wear Prediction

Tool wear status is crucial for maintaining high workpiece quality and enhancing productivity. Numerous deep learning-based approaches have been developed to monitor tool wear. A single deep-learning model, however, often faces limitations in feature extraction. In order to address this issue, we propose a hybrid global and local neural network with lightweight multiscale information fusion for tool wear prediction tasks. First, Stockwell transform (ST) is applied to comprehensively extract time-frequency features from raw signals. Next, the lightweight multiscale feature extraction module arranges spatial separable convolution (SSConv) blocks with different receptive fields in parallel. This design allows the model to fully exploit spatial features at multiple scales to enhance dynamic signal sensitivity. Subsequently, a hybrid global and local feature extraction module comprises two parallel branches, one focused on extracting fine-grained local detail and the other on modeling global long-range dependencies. Finally, a novel orthogonal attention feature fusion (OAFF) strategy is proposed to enable effective interactions among features and capture a comprehensive high-level representation of wear-related characteristics. In order to validate the superior predictive capability of the proposed model, a milling experiment was designed and conducted. The experiment results demonstrate that the proposed model achieves an average root mean square error of 9.369 and a coefficient of determination of 0.907. Ablation experiments, moreover, demonstrated the effectiveness of the proposed multiscale feature extraction module and hybrid global and local architecture.