Abstract: Thermal effect is the key factor restricting the machining accuracy of machine tools, in order to further improve the accuracy of thermal error prediction of machine tools, considering the relative characteristics of thermal error parameters, temperature and deformation, a thermal error prediction model for machine tools based on differential fusion long short-term memory neural network, which combines the advantages of direct prediction and differential prediction, is proposed. Taking a certain model of cylindrical grinding machine as the experimental object, four sets of experiments are designed according to different motor frequencies, the data including temperature and axial deformation of the tool (grinding wheel) are continuously collected. Then 4 temperature sensitive points are selected using random forest and recursive feature elimination with cross-validation. Training datasets and testing datasets are randomly selected, compared to conventional widely used artificial neural networks. The proposed model gets a better balance between trendiness and volatility and achieves better prediction accuracy and has stronger robustness, which demonstrates the rationality of introducing differential prediction. Thus, it has certain reference significance to the actual thermal error prediction and compensation of CNC machine tools.

Key words: thermal error prediction, relative characteristics, differential fusion, long short-term memory