A Preliminary Study of Optimization Strategies for Ultra-precision Robot Polishing Process based on Machine Learning
Robot polishing technology has been adopted in modem ultra-precision manufacturing industries.Given a set of polishing parameters which define a polishing procedure with robot manipulator, the material removal function can be estimated based on theoretical models such as Prestons empirical equations.Meanwhile, the polishing parameters can also be estimated if the required removal fimction is given.There will be differences between the estimated values of removal function or polishing parameter and the real ones due to the existing errors including systematic errors and random errors.Resulting from certain conditions of robot manipulator and circumstances and other factors, systematic error greatly contributes to this differences.Unlike random error, expectation of systematic error is nonzero.Although it is difficult to describe or predict systematic error, it can be learntfrom past experiences.Therefore, this paper aims to study the optimization strategies by using algorithms which can learn from past experiences to predict and figure out the optimum polishing parameters and the material removal function.Machine learning algorithms are employed to predict removal function based on polishing parameters.Three classical algorithms of machine learning are used, which include the algorithms of k-nearest neighbor (kNN),regression decision tree (RDT), and support vector machine (SVM).The training data set are learnt by using the above three algorithms, then three different learning models are got and their performance is tested respectively.Furthermore, one of ensemble learning algorithms called Adaboost algorithm can be used as to take the advantages of the three algorithms so as to improve the overall performance.Besides, Convolution Neural Network (CNN) is also to be taken to scanthe material removal function matrixes and establish certain neural transfer weight from which polishing parameters is then estimated by minimizing loss function.Simulation and experimental studies have been undertaken and the results show that, although each of the selected machining learning algorithms, including kNN, RDT and SVM, can predict the material remove function corresponding to specific polishing parameters considering the systematic errors well, the three algorithms have their advantages and disadvantages.Ensemble learning algorithm is supposed to be able to combine all the advantages of the three types of algorithms to improve the prediction accuracy.Furthermore,algorithm of CNN also exhibits a promising performance in predicting polishing parameters in ultra-precision robot polishing.The research work will provide a promising method to find out optimization strategies in ultra-precision robot polishing process.
Ultra-precision robot polishing Machine learning Material removal function Deep learning Ensemble learning Optimization
Yi Yu Lingbao Kong Haitao Zhang Min Xu Liping Wang
Shanghai Engineering Research Center of Ultra-precision Optical Manufacturing, Department of Optical State Key Laboratory of Applied Optics, Changchun Institute of Optics, Fine Mechanics and Physics, C
国际会议
上海
英文
87-91
2017-11-19（万方平台首次上网日期，不代表论文的发表时间）