• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2019, Vol. 55 ›› Issue (15): 48-55.doi: 10.3901/JME.2019.15.048

• 特邀专栏:增材制造技术 • 上一篇    下一篇

基于机器学习的电子束选区熔化成形件密度预测

亓欣波1, 李长鹏2, 李阳3, 林峰3, 李勇1, 程宣2, 陈国锋2   

  1. 1. 清华大学摩擦学国家重点实验室 北京 100084;
    2. 西门子中国研究院 北京 100102;
    3. 清华大学生物制造与快速成形技术北京市重点实验室 北京 100084
  • 收稿日期:2018-12-16 修回日期:2019-01-14 出版日期:2019-08-05 发布日期:2019-08-05
  • 通讯作者: 亓欣波(通信作者),男,1988年出生,博士后。主要研究方向为智能制造。E-mail:qixinbo@gmail.com
  • 作者简介:李长鹏,男,1977年出生,高级工程师。主要研究方向为智能制造。E-mail:changpeng.li@siemens.com;林峰,男,1966年出生,教授。主要研究方向为增材制造和生物制造。E-mail:linfeng@tsinghua.edu.cn;李勇,男,1962年出生,教授。主要研究方向为微细电加工工艺与装备、微流控器件、超精密切削加工及其应用。E-mail:liyong@mail.tsinghua.edu.cn

Machine Learning Algorithms on Density Prediction of Electron Beam Selective Melted Parts

QI Xinbo1, LI Changpeng2, LI Yang3, LIN Feng3, LI Yong1, CHENG Xuan2, CHEN Guofeng2   

  1. 1. State Key Laboratory of Tribology, Tsinghua University, Beijing 100084;
    2. Corporate Technology, Siemens Ltd., Beijing 100102;
    3. Bio-manufacturing and Rapid Forming Technology Key Laboratory of Beijing, Tsinghua University, Beijing 100084
  • Received:2018-12-16 Revised:2019-01-14 Online:2019-08-05 Published:2019-08-05

摘要: 电子束选区熔化技术,以其快速加工复杂几何形状、真空熔炼、成形件残余应力小、粉末可回收等优点,是近年来快速发展的一种增材制造技术。但该技术加工参数众多,目前很难建立起直接且准确的加工参数与成形件性能的关系。通过调节电子束选区熔化技术的扫描速度、束流、基板温度、粉层厚度等四种参数,制备了一系列Inconel718立方体样品,并测试获得其密度性能。利用线性回归、支持矢量回归和神经网络等三种机器学习算法建立了四种加工参数与密度的关系。算法结果表明:线性回归因其模型容量小,预测精度最差;神经网络模型容量大,但易出现过拟合,预测精度较好;支持矢量回归的模型容量适当,且解释性良好,预测精度最高。

关键词: 电子束选区熔化, 机器学习, 密度, 预测, 增材制造

Abstract: Electron Beam Selective Melting (EBSM) is a novel additive manufacturing technology, which is developing very fast nowadays. It has many advantages:building up parts with complex morphology; processing under vacuum to get rid of the impurity; its manufactured part with small residual stress and powder recycling. However, it is difficult to establish the relationship between EBSM's processing parameters and parts' properties. Here a series of Inconel 718 cubic specimens are manufactured through adjusting EBSM's scanning speed, beam current, plate temperature and layer thickness. Then the densities of these specimens are measured. Three kinds of machine learning algorithms, including linear regression, support vector regression and neural network, have been utilized to build the relationship between these four processing parameters and density. The results show that:Linear regression has the worst prediction skill as a result of its small model capacity; neural network has a better prediction accuracy, but it is easily overfitting; support vector regression has appropriate model capacity and good physical interpretation, and it behaves best in the density prediction.

Key words: additive manufacturing, density, electron beam selective melting, machine learning, prediction

中图分类号: