Modal Parameters Prediction Of The End Of The Milling Robot Based On Deep Transfer Learning

When the robot is performing milling processing,due to the low rigidity,it is prone to vibration,which will have a negative influence on the accuracy of processing and the quality of the workpiece surface.In order to lower the working vibration,the stability region of the processing parameter is often obtained through vibration stability analysis,and the vibration interval is avoided by optimizing the processing parameter.The important premise of vibration stability analysis is modal parameters,which are generally obtained by means of experimental modal analysis.However,the modal parameters of the robot in different poses are very different.Sometimes even a change in angle will cause the modal parameters to change,so performing modal experiments in different postures will take a lot of time and labor costs.Thus,how to quickly obtain the modal parameters of the specified pose under the corresponding working conditions is of great significance to industrial practice.Therefore,this paper proposed a deep transfer learning modal parameters prediction model based on Generative Adversarial Nets,which can quickly predict the modal parameters of a specified pose.The main research contents of this article are as follows:(1)Deep transfer learning modeling based on generative adversarial networks.This paper focused on the transfer learning method and the generation of adversarial networks.The loss function of GAN was derived in detail.This paper clarified how to add GAN ideas to transfer learning,defined the transfer learning task of this article,and clarified "what to transfer" and "how to transfer";The data format of the model was defined: the robot's end pose and joint variables are features,and the robot's modal parameters are used as labels.The loss function of each part of the model was defined,and finally the deep transfer learning modeling was completed based on the adversarial generation network.(2)D-H kinematics modeling of robots.The reference coordinate system of the COMAU robot was established through the standard D-H method,and the D-H kinematics parameters table of the robot was obtained through coordinate transformation;kinematics modeling and forward and inverse kinematics solution were obtained with MATLAB,and the feature data of the source domain and target domain were finally obtained.(3)Modal experiment design and frequency response function curve fitting.The robot impact hammer experimental platform was built;the frequency response function curves of the sample points in the source and target fields were obtained.Finally,the modal parameters of each order were identified by the rational fractional polynomial method,which completed the labeling of the source and target domain samples.(4)Validation of the transfer learning prediction model.This article brought the source domain data and target domain data with complete annotations into the built deep transfer learning model for training and testing,and took corresponding measures to optimize and adjust the model based on the preliminary prediction results,and acquired the prediction consequences of the test data from the target domain and comparison results before and after optimization.The contradistinction consequences confirmed the feasibility of the model optimization method.By comparing with the prediction results of the test data without adversarial transfer learning,the effectiveness of the proposed method for modal parameter prediction was evidently proved.