Abstract: Data driven methods are widely used in thermal error modeling, but the open-loop and black-box operation mode without mechanism support is difficult to ensure the robustness of the model in new operating conditions, resulting in the failure of the model. To further improve the accuracy and stability of the thermal error model of machine tools, this paper proposes a thermal error reconstruction model based on temperature similarity evaluation. By evaluating the similarity between the temperature mean vector of the prediction group and the temperature mean vector of the modeling group, a specific batch similar to the temperature of the prediction group is selected from the original modeling batch, and the thermal error model is reconstructed based on the partial least squares algorithm. To verify the effectiveness of the reconstruction model, the thermal error prediction of 31 batches was conducted. The results show that the reconstruction model can further reduce the mean root mean square error of prediction results to 2.2 μm. And on the premise of ensuring stability, the prediction accuracy can be improved by 15% and 35% respectively compared with the conventional multiple linear regression model combining fuzzy clustering and the partial least squares model, showing the significant effect. The method has reference value for evaluating the applicability of data-driven thermal error models and improving model prediction accuracy.

Key words: similarity evaluation, reconstruction model, partial least squares algorithm, stability