High Precision Prediction Of Workpiece Surface Roughness In Precision Machining

With the development of industry and manufacturing industry,the requirements for operation in high speed,high temperature,high pressure and heavy load environment are increasing,and the requirements for workpiece processing quality are also increasing.Surface roughness is regarded as an important factor to measure the quality of workpiece processing.Therefore,it is of great significance to realize realtime monitoring of workpiece surface roughness to improve processing quality and production efficiency.Therefore,this paper systematically studies the high-precision prediction method of workpiece surface roughness based on traditional machine learning and deep learning.Combined with the actual processing conditions,the centerless lathe is used to cut the titanium alloy wire,and the vibration signal is collected in multiple channels.Based on machine learning and deep learning,the surface roughness prediction model of the workpiece is constructed.The main research contents include:(1)Considering various cutting conditions,the test scheme is designed,and the test platform for cutting titanium alloy wire by centerless lathe is built.The test data samples are collected during the cutting process,including the vibration signal samples of different sampling channels of the centerless lathe test platform and the workpiece surface roughness samples of the corresponding measuring points.After collecting samples,the sensitive factors between the centerless lathe and the surface roughness of the workpiece are deeply analyzed and studied.Firstly,the correlation between the vibration signals of different sampling channels in the centerless lathe cutting system and the surface roughness of the workpiece is analyzed to determine the sensitive sampling channel.Then,the vibration signal of the sensitive channel is decomposed into wavelet packet coefficients of different frequency bands by wavelet packet decomposition,and the sensitive frequency band is determined by correlation analysis.(2)The prediction models of workpiece surface roughness are established based on multi-layer feedforward network(BP)and support vector machine(SVM)respectively.The time domain analysis and wavelet packet decomposition technology are used to extract data features from the vibration signals collected from the sensitive sampling channels,and the features are screened by correlation sorting,so as to construct the input parameters of BP neural network and SVM.In terms of model optimization,genetic algorithm(GA)is used to optimize the weight,threshold and number of hidden layer neurons of BP neural network prediction model,and grid search method is used to optimize the penalty factor C and kernel function parameter g in SVM.The results show that the relative percentage error of BP neural network model prediction results is not more than 11.4 %,the root mean square error is 0.0235,the average absolute error is 0.0175,and the determination coefficient is 0.8401.The relative percentage error of the SVM model prediction results is not more than 13.0 %,the root mean square error is 0.0313,the average absolute error is 0.0292,and the coefficient of determination is 0.7744.(3)A high-precision prediction model of workpiece surface roughness is established based on traditional convolutional neural network(CNN)and residual network(Res Net)respectively.Combined with sensitive factors,the method of removing redundant information is selected,and the bottom coefficients of wavelet packet corresponding to the sensitive frequency band are fused.According to the order of different sampling channels,it is arranged into one-dimensional sequence,and then the one-dimensional sequence is reconstructed into two-dimensional coefficient matrix as input parameters to realize prediction.The results show that the relative percentage error of CNN model prediction results is less than 10.3 %,the root mean square error is0.0235,the average absolute error is 0.0188,and the determination coefficient is0.8100.The relative percentage error of the prediction results of the Res Net model is not more than 5.8 %,the root mean square error is 0.0159,the average absolute error is0.0133,and the determination coefficient is 0.9148.