Research On Thermal Error Modeling Of Machine Tool Spindle Based On Deep Learning

Today,with the transformation of Chinese manufacturing into Chinese intelligent manufacturing,high-precision CNC machining technology is still the goal pursued in the domestic manufacturing field.Among the many factors that affect the machining accuracy of machine tools,thermal error of the machine tool spindle accounts for a large proportion,which is one of the main reasons for the machining error of machine tools.The traditional thermal error modeling method is based on the selection of temperature-sensitive points,and thermal error modeling research is carried out by establishing a machine learning algorithm or a simple neural network.It is mainly used for training and prediction of some simple working condition data,and the generalization and robustness of the model need to be further improved.Therefore,the thermal error model with high accuracy,strong robustness and excellent generalization under multiple working conditions still needs to be further explored.This paper focuses on the key technologies of thermal error modeling for CNC machine tools under multiple working conditions.The specific contents are as follows:(1)The VMC850 CNC machining center is taken as the experimental object to carry out the thermal error experiment.Firstly,according to the principle of temperature sensor arrangement in the ISO230-3 national standard,the temperature sensors are reasonably installed,and the thermal deformation acquisition instrument is arranged according to the "five-point method".Secondly,the experimental conditions are designed,which are divided into three parts: "fixed speed","variable speed" and "complex variable speed".Then,the experimental data under multiple conditions are obtained according to the experimental conditions.Finally,the thermal error data under different working conditions are displayed and analyzed.(2)Aiming at the problem of poor generalization of traditional thermal error modeling methods in thermal error prediction under multiple operating conditions,a thermal error modeling method based on convolutional neural network(CNN)and mayfly algorithm(MA)is proposed.First of all,the advantages of one-dimensional convolutional neural network in feature extraction of time series data are utilized to build a CNN framework based on onedimensional convolutional neural network,batch normalization layer,maximum pooling layer and fully connection layer.The “Early Stopping” method and “Dropout” method are combined to prevent the model from overfitting.Meanwhile,MA is used to optimize the combination of the kernel and the kernel-size of convolutional layer of CNN.During the optimization process,the mean square error is taken as the fitness function,and the constructed sample set is input into the CNN model for updating and iterating fitness values to further improve the robustness of the model.Secondly,several evaluation indexes are established to evaluate the prediction performance from the validity and generalization of the model.Finally,the z-direction thermal error under different working conditions are predicted by the trained MA-CNN model,and the results are compared with the traditional and classical thermal error models,such as support vector regression(SVR),feedback neural network(BP)and multiple linear regression(MLR).By analyzing the fitting curve,residual curve,validity and generalization evaluation index values,the thermal error prediction results of various models under different working conditions are compared to verify the prediction performance of the models.The results show that the proposed MA-CNN model has high prediction accuracy and good generalization,but the generalization in variable speed needs to be further improved.(3)Aiming at the problem that the generalization of MA-CNN model needs to be improved in the thermal error prediction of variable speed and the hysteresis of thermal error,a thermal error modeling method based on deep gated recurrent unit and mayfly algorithm(MA-DGRU)is proposed.Firstly,a deep network(DGRU)with multi-layer GRU network and full-connected layer network is constructed by combining with the thermal hysteresis effect and the excellent historical memory function of GRU.In addition,MA is used to optimize the number of hidden layer units,batch size and dropout rate in the DGRU network to further improve the prediction accuracy and robustness of the network.Secondly,MADGRU is trained and predicted using the same dataset as the MA-CNN model,and compared with LSTM and MA-CNN model.The prediction performance of the three models in different working conditions is verified by analyzing the fitting curve,residual curve,effectiveness and generalization index values.The results show that the proposed MA-DGRU model has better prediction accuracy at both fixed and variable speeds,and has better generalization than MACNN and LSTM models.(4)A CAGA hybrid network spindle thermal error modeling method is proposed for the data of complex variable speed.Firstly,the CAGA network framework is built by combining the advantages of CNN,GRU and self-attention.Through the first self-attention,the underlying features extracted by CNN are given different weights.Then,the calculated new feature matrix is inputted to GRU to improve the training efficiency of the model.Through the second self-attention,the high-level spatiotemporal features extracted by GRU are further simplified and optimized to further improve the fitting effect of the model.At the same time,the model is trained by the data of complex working conditions.Secondly,the t-SNE dimension reduction method is used to visualize the hidden layer features before and after the double self-attention in the CAGA network.Therefore,the feature extraction ability of CAGA model is verified in reverse.Meanwhile,the high-level hidden layer features of the CAGA,MA-DGRU and MA-CNN models are mapped and visualized in low-dimensional to show the good feature extraction capabilities of the proposed three models.Then,the prediction performance of the CAGA model is verified by comparing the Z-direction thermal error prediction results of the MA-DGRU and MA-CNN models under different complex variable speeds.The results show that the proposed CAGA model has good generalization and high training efficiency in thermal error prediction of complex operating conditions.Finally,by comparing with the traditional thermal error model,the effectiveness of the three thermal error modeling methods based on deep learning proposed in this paper in other directions of thermal error is verified.