航空发动机外形点云的特征分割方法

Feature Segmentation Method of Aero-Engine Profile Point Cloud

目前我国存在较多外购发动机的情况,外购发动机存在只有实物及安装尺寸等信息,而没有三维数字化模型的问题,这给飞机与发动机的装配协调设计带来较大困难,因此飞机设计部门对快速重构航空发动机的外形几何模型提出了迫切需求。为了使重建出的发动机外形几何模型尽可能地保留准确的结构特征,提出了一种基于深度学习的航空发动机外形点云特征分割方法,该方法将整体点云分割成特征数据与非特征数据,这有利于后续采用不同的方法重建出各种复杂的结构特征。设计了一种迭代密度均衡算法用于构建特征分割数据集,该算法为特征分割网络的训练、测试和性能评估提供基础;设计了一种特征分割网络,从多尺度局部表面片中收集形状结构和局部邻域信息,用于判断其中心是否是特征点。将训练好的特征分割网络模型应用于发动机外形点云,验证结果表明,特征分割精度达到95.16%,所提算法实现了高精度语义分割。

At present,many domestic aero-engines are bought from abroad.Only physical objects and installation dimensions are provided for such aero-engines,and the lack of three-dimensional digital models brings great difficulties to the assembly coordination design of aircraft and aero-engines.Therefore,aircraft design departments urgently need to quickly reconstruct geometric models of aero-engine profiles.To enable a reconstructed geometric model of the aero-engine profile to retain exact structural features,this paper proposes a feature segmentation method of the aero-engine profile point clouds based on deep learning.It divides the whole point clouds into feature data and non-feature data,which is conducive to the subsequent reconstruction of various complex structural features by different methods.An iterative density equalization algorithm designed to create a feature segmentation dataset provides a basis for the training,testing,and performance evaluation of the feature segmentation network.A feature segmentation network is designed to collect the shape structure and local neighborhood information from multi-scale patches and thereby determine whether the center is a feature point.The trained feature segmentation network model is then applied to the profile point cloud of an aero-engine.The verification results show that the accuracy of feature segmentation reaches 95.16%,which means the proposed algorithm achieves high-precision semantic segmentation.