Research On Prediction Of Part Surface Roughness Based On Deep Learning

Deep learning network can be applied to the field of intelligent prediction of surface roughness of Ultra-low temperature processed parts.However,the surface roughness sample data input to the deep learning network has problems of unbalanced categories and complex data distribution,and the deep learning network itself has the disadvantages of difficulty in building and long training time,which increase the difficulty of the deep learning network model to predict the surface roughness of Ultra-low temperature processed parts.Therefore,this paper carried out the research on the intelligent prediction technology of the surface roughness of the Ultra-low temperature processed parts.The acceleration sensor was used to collect the vibration signal during the processing,a method for enhancing the surface roughness sample dataset of parts based on improved generative confrontation network(Coral GAN)is proposed,the surface roughness prediction model of the part under conventional processing conditions based on deep learning is established,a framework for predicting the surface roughness of Ultra-low temperature machined parts based on transfer learning is designed,and finally to realize the intelligent prediction of the surface roughness of Ultra-low temperature machined parts.First of all,it briefly introduces the definition of surface roughness and the selection of parameters when measuring surface roughness.The physical factors affecting the surface roughness of parts are described in detail,and the relationship between the vibration of process system and the surface roughness is clarified,so as to determine the vibration signal as the input sample of deep learning network model.Secondly,aiming at the problems of the surface roughness sample data of Ultra-low temperature machining parts and the disadvantages of the traditional generative adversarial networks,the generative adversarial network is improved,and the network parameters of the generator are optimized by Deep Coral function to improve the ability of generative adversarial network to learn complex sample data.The sample data of surface roughness of Ultra-low temperature machining parts are generated by using improved GAN(Coral GAN)network model to enhance the sample dataset.Then,aiming at the problems of deep learning network,a prediction framework for surface roughness of Ultra-low temperature machined parts based on transfer learning is proposed.The prediction model of part surface roughness based on deep autoencoder network under conventional processing conditions is established.The trained deep autoencoder network structure and parameters are transferred to the Ultra-low temperature processing conditions by using the method of model parameter migration,to extract the features of the part surface roughness sample data.Finally,the accurate prediction of the part surface roughness is realized by the method of fine-tuning with labels.Finally,the Ultra-low temperature cooling milling experiment is designed.The vibration signals in the machining process are collected,and the Coral GAN model is used to generate the surface roughness sample data of Ultra-low temperature machining parts to verify the feasibility of the proposed method,the surface roughness sample data of Ultra-low temperature machining parts are predicted based on the transfer learning method to verify the usability of the transfer learning model.The experimental results show that the Coral GAN network model and the transfer learning framework proposed in this paper can accurately predict the surface roughness of Ultra-low temperature machining parts.