Milling Tool Wear Prediction Method Research Based On Transfer Learning

Tool wear has an important influence on the machining accuracy and surface quality of parts.To establish an effective tool wear prediction model can improve the qualified rate of products and reduce the production cost.The actual processing scene is complex and changeable,and the data distribution under the new scene is greatly different from the training data distribution of the existing model,so it is necessary to retrain the model with the data with wear value or wear status label in this scene.However,it is expensive and difficult to obtain data with tags in actual machining scenes.This paper takes transfer learning method as the core to study how to build tool wear prediction model in new scenes based on tool wear monitoring data and prediction model with tags in existing scenes.In this paper,different transfer learning models are constructed according to the differences between domains and the different types of prediction tasks,which can effectively solve the prediction of tool wear state and wear value in the target domain without labels.This paper mainly studies from the following three aspects:Aiming at the problem of tool wear state prediction under different tool diameters,an adaptive transfer learning model based on deep network was established in this paper.Multiple kernel-maximum Mean Disrepancy(MK-MMD)was used to measure the difference between the feature space of the source domain and the target domain.The model was verified by the prediction of the wear state of the end milling cutter with a diameter of8 mm and a diameter of 6mm.The results show the effectiveness and practicability of the proposed method.For tool wear states under different scenarios to predict cutting tool type problem,this paper constructed a dynamic adversarial transfer networks(DATN)model.The model is based on the idea of antagonism to measure the differences between domains,and can quantitatively evaluate the importance of global and local distribution adaptation.By dynamically learning domain invariant feature representation,the alignment of feature space between source domain and target domain can be realized.The experimental data of different types of tool wear monitoring were selected to verify the model,and compared with the domain adaptive migration method,the results show that the prediction effect of the proposed method is better when the source domain and the target domain differ greatly.In order to solve the problem of tool wear value prediction under different working conditions,the regression prediction models of tool wear value were constructed by using fixed feature extraction network and different domain difference measurement methods of MK-MMD and Adversarial Transfer Networks(ATN).Finally,the model was verified by migration between different working conditions,and the results show that both transfer learning methods can significantly reduce the prediction error of tool wear value under different working conditions.