Robust modeling method for thermal error of CNC machine tools based on random forest algorithm

• Published in: Journal of intelligent manufacturing Vol. 34; no. 4; pp. 2013 - 2026
• Main Authors: Zhu, Mengrui, Yang, Yun, Feng, Xiaobing, Du, Zhengchun, Yang, Jianguo
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.04.2023; Springer Nature B.V
• Subjects: Accuracy; Advanced manufacturing technologies; Algorithms; Artificial intelligence; Business and Management; Control; Deformation effects; Errors; Machine learning; Machine tools; Machines; Manufacturing; Mathematical models; Mechatronics; Modelling; Parameters; Prediction models; Processes; Production; Robotics; Robustness; Time lag; Workpieces; Thermal error; Random forest; Machine tool; Artificial intelligence; Machine learning
• ISSN: 0956-5515; 1572-8145
• DOI: 10.1007/s10845-021-01894-w

Thermal error of machine tools has a huge influence on the accuracy of the workpiece. However, the nonlinearity of the thermal error limits the accuracy and robustness of the prediction model. With the rapid advancement in artificial intelligence, this paper presents a novel thermal error modeling method based on random forest. The model’s hyper-parameters are easy to be optimized by grid searching method integrating with cross validation. The temperature features are measured as the model input. Based on the out-of-bag data generated during modeling process, the proposed model itself can simultaneously evaluate the temperature feature importance through comparing the decrease in model’s the prediction accuracy after randomly shuffling the value of the target feature. Moreover, to enhance the model performance and reduce the measurement and computational cost, the method of selecting key temperature points are presented to exclude the redundant features through iteratively eliminating the least important feature and comparing the prediction accuracy under different feature combinations. Furthermore, the hysteresis effect between temperature and deformation is also considered. The method of determining the time lag is proposed through permuting the original time series of the target feature while keeping the remainder constant and comparing the resultant relative importance. A thermal error experiment validates the accuracy and robustness of the proposed model which can continuously maintain the prediction accuracy of over 90% in spite of varying operation conditions. Compared to conventional machine learning methods, the proposed model requires less training data, enables faster and more intuitive parameter tuning, achieves higher prediction accuracy, and has stronger robustness.