Modeling Of Milling Machine’s Spindle Power Based On Transfer Learning

Intelligent NC machining is an important development direction of intelligent manufacturing.Predicting machine response through large data modeling is one of the important aspects to improve the accuracy and efficiency of NC machining.The previous data-driven machine response(such as spindle power)modeling method is usually based on a given working condition with a specific tool,material and cooling condition,so it’s difficult to apply the model to different working conditions.However,it spends too much on the data acquisition time and experimental cost of the re-modeling process for some new conditions.To reduce the data demand for spindle power modeling under new working conditions,and control the data acquisition cost of various working conditions in actual production,this thesis applied transfer learning technology to realize spindle power modeling by using the processing data under different working conditions.To reduce the data demand of new working conditions and improve the utilization rate of historical data,this thesis used the instance-based transfer learning method.By transferring historical spindle power data instances,this thesis has built the spindle power prediction model with good performance in new working conditions,only a small amount of data was needed.This thesis proposed a domain adaptation method for regression problems to reduce the distribution difference between the historical data(source domain data)and the new data(target domain data)for better modeling results.By building a domain adaptation neural network(Da NN)and fine tuning the network,the knowledge of spindle power data in the source domain could be learned and the distribution difference between the source domain and target domain could be reduced.Based on the transfer source data and the target domain data,a transfer method based on Tr Ada Boost was designed,whose optimal parameters are determined through an iterative test to realize the target domain spindle power prediction.The method proposed in this thesis was tested in the experiments of different working conditions,such as tool size,tool type,and blank material.When sufficient spindle power modeling data has been obtained in one working condition,the demand of spindle power training instances in other new working conditions was successfully reduced to about 10%of the original with the average spindle power prediction error less than 5%.The method proposed in this thesis showed obvious advantages in the comparison experiment and verified the effectiveness and advantage of the work.