2018 年 12 巻 7 号 p. JAMDSM0131
Thermal errors account for more than 40% of the machining errors in CNC machine tools, and even reach 70% in precision and ultra-precision machine tools. One approach to reducing thermal errors is to build thermal error prediction models by monitoring the temperature field of machine tools with the data-driven modeling approaches. Usually the data-driven approaches have over-fitting and under-fitting problems. The prediction performances of the data-driven models are greatly dependent on the input of the models, namely the number of temperature measuring points (TMPs) and their locations. In this paper, a selection method of key TMPs is presented to improve the accuracy and generalization of predictive models. In this selection method, correlation analysis is used to eliminate the uncorrelated or weakly correlated TMPs to thermal errors at first; then the Minimum Redundancy Maximum Relevance (MRMR) is presented to narrow the searching scope of TMPs; finally Wrapper method is used to test the candidate set of TMPs with cross-validation accuracy to find the key TMPs. To validate this method, a test platform is built on the CNC gantry drilling machine ZK5540A. 56 Fiber Bragg Grating (FBG) temperature sensors are amounted on the body, column, spindle, and base of this machine. 7 key TMPs was selected from the 56 ones with this method. And Multiple linear regression (MLR) approach is used to build the thermal error prediction model with the 7 key TMPs, 56 TMPs and key TMPs selected by Miao's method(Miao et al., 2014) respectively. The result shows that the prediction model built with MLR using the 7 key TMPs is much more accurate and has stronger generalization.