中国机械工程 ›› 2021, Vol. 32 ›› Issue (23): 2850-2860,2889.DOI: 10.3969/j.issn.1004-132X.2021.23.009

• 高端装备智能制造技术 • 上一篇    下一篇

航空发动机外形点云的保特征去噪方法

闫杰琼;周来水;胡少乾;文思扬   

  1. 南京航空航天大学机电学院,南京,210016
  • 出版日期:2021-12-10 发布日期:2021-12-23
  • 通讯作者: 周来水(通信作者),男,1962 年生,教授、博士研究生导师。研究方向为数字化制造技术。E-mail:zlsme@nuaa.edu.cn。
  • 作者简介:闫杰琼,女,1994 年生,博士研究生。研究方向为数字化设计与制造、点云数据处理。E-mail:yanjieqiong@nuaa.edu.cn。
  • 基金资助:
    国家科技支撑计划(2020YFB2010702)

Feature-preserving Denoising Method for Aero-engine Profile Point Cloud

YAN Jieqiong;ZHOU Laishui; HU Shaoqian;WEN Siyang   

  1. College of Mechanical and Electrical Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing,210016
  • Online:2021-12-10 Published:2021-12-23

摘要: 三维激光扫描设备可以提供航空发动机外形实测点云,但其中包含的噪声会直接影响后期外形几何模型的重建精度。为保证在去除噪声的同时不模糊或破坏掉发动机复杂的外形几何特征,提出了一种基于深度学习的点云保特征去噪方法。将航空发动机外形噪声点云分割成特征数据和非特征数据之后,分别设计了特征去噪网络和非特征去噪网络,用于预测特征噪声点和非特征噪声点的位置修正向量,噪声点沿预测向量移动后被投影回模型真实的底层表面上,实现去噪。构建了用于特征去噪学习和非特征去噪学习的数据集。验证结果表明,在将该方法应用于各种噪声尺度的发动机外形点云时,相比现有的学习基方法,去噪效果得到提高,且有更好的几何特征保护能力,可以为后续重建提供高质量点云。

关键词: 航空发动机, 保特征去噪方法, 深度学习, 去噪损失函数, 迭代去噪

Abstract: The 3D laser scanning devices might provide the raw data of aero-engine profiles,but the noise would directly affect the reconstruction accuracy of the 3D geometric model of aero-engine profiles.In order to ensure that the noise was removed without blurring or destroying the complex geometry features of the aero-engine profiles,a feature-preserving point cloud denoising method was proposed based on deep learning.The noisy point cloud of aero-engine profiles was divided into feature data and non-feature data.The feature denoising network and the non-feature denoising network were designed to predict the position correction vectors of the feature noisy points and the non-feature noisy points respectively.The noisy points moved along above vectors and were projected back onto the original clean surfaces of the model.The data sets were constructed separately for feature denoising learning and non-feature denoising learning.The verification results show that when the proposed method is applied to the point cloud of aero-engine profiles with various noise scales,the denoising effectiveness is improved respectively compared with the existing learn-based methods.The proposed method has better geometric feature protection ability and may provide high quality point cloud for subsequent reconstruction.

Key words: aero-engine, feature-preserving denoising method, deep learning, denoising loss function, iterative denoising

中图分类号: