Experimental Research on Thermal Deformation Error of CNC Machine Tools Based on Neural Network

Abstract

During precision machining of CNC machine tools, machining accuracy is affected by multiple error factors, among which thermal deformation error accounts for 40% to 70%, becoming a key bottleneck restricting accuracy improvement. Existing thermal error prediction methods are mostly based on static mapping relationships, directly predicting thermal error at a given moment using temperature sensor data from the same moment, thereby neglecting the dynamic, time-dependent evolution characteristic of thermal error itself. In actual machining, inevitable time delays exist between temperature data acquisition, model computation, and the execution of error compensation, leading to predicted values lagging behind the actual compensation values required. This significantly reduces compensation effectiveness and machining accuracy. To address this prominent issue, this paper proposes a novel thermal error ahead-of-time prediction method based on Long Short-Term Memory (LSTM) neural network. The core of this method lies in leveraging the advantages of LSTM networks in processing time-series data. It utilizes historical and current time-series data from key temperature sensors as training samples to construct a temperature-ahead prediction model, thereby achieving early and accurate estimation of future thermal error. This study elaborates on the fundamental principles of LSTM neural networks, their gating mechanisms, and the specific application process in ahead-of-time prediction modeling. Simulation machining experiments were conducted on a self-developed aspheric surface CNC grinding machine, collecting multiple sets of temperature and thermal error time-series data for model training and validation. Experimental results show that the LSTM-based ahead-of-time prediction model can effectively predict thermal error changes within the next 10 minutes. The maximum residual, mean residual, and mean square error between the predicted and actual values are maintained at low levels, demonstrating significantly improved prediction accuracy. Compared with traditional thermal error modeling methods based on multiple linear regression, the LSTM model not only achieves timeliness in prediction but also exhibits clear advantages in prediction precision, with smaller mean residuals and variance, validating the model's effectiveness and advancement in solving the thermal error compensation lag problem and improving the machining accuracy of CNC machine tools.