Abstract: Tool wear prediction is of great significance for improving machining accuracy and production efficiency. Tool wear prediction models include physics-based models and data-driven models. The physics-based models often apply empirical or simplified formula to model the tool wear process which often loss prediction accuracy under changing cutting parameters. On the other side, the data-driven models estimate the tool wear by monitor data without considering the mechanisms of tool wear, resulting in low model generalization and results interpretation. To address these issues, a novel Gaussian process latent force model for tool wear prediction is proposed. The proposed model applies Gaussian process to model the unknown parameters of the tool wear physical model, and establishes a physics-informed machine learning model. The Gaussian process latent force model not only avoids identifying parameters in the physical model, but also explores hidden information from physics and data domains. Moreover, by integrating the physical model with the Gaussian process covariance function, a physics-informed covariance function is constructed to constrain the outputs of the model and improve the prediction accuracy. The multi-condition experiments results show that the mean absolute error and root mean square error of the proposed method are 2.5 945 and 3.7 408, respectively, which are smaller than the prediction error of the traditional data-driven models and further improve the prediction accuracy.

Key words: tool wear prediction, Gaussian process, latent force model