A novel tool wear state recognition framework based on graph neural networks

The tool wear state significantly influences the surface quality of alloy workpieces during milling, thereby affecting working performance and service life. Given the continuous characteristics of tool wear, collected data is often imbalanced, and fuzzy regions exist between different wear stages, which hinder accurate identification of wear states. Existing studies seldom exploit the non-Euclidean structural relationships among multi-sensor signals to enhance the accuracy of tool wear state recognition. To address these limitations, a novel tool wear state recognition framework based on graph neural networks is proposed. A method for defining nodes is introduced to characterize the tool wear state by leveraging the latent relationships between nodes. A multi-receptive fields fusion graph attention network is employed to capture more global information by utilizing important nodes to generate weight coefficients for multiple graphs. This approach effectively extracts meaningful features and improves classification accuracy for data in fuzzy regions between wear stages. The final output is further strengthened through the linear combination of outputs from two parallel fully-connected layers. The proposed framework’s effectiveness is validated using the PHM2010 dataset, which achieved 99.68 %, 98.41 %, and 98.41 % accuracy on three D cross-datasets, respectively. This framework enables precise recognition of three tool wear states based on force and vibration sensor signals and facilitates flexible tool replacement strategies in intelligent manufacturing.