Scholarlink

In view of the insufficient generalization performance and overfitting problems of the existing aerodynamic prediction model based on the PointNet architecture in aerodynamic prediction tasks, this study proposes an aerodynamic parameter prediction model based on the PointNet++ architecture and fused with feature-preserving downsampling - Feature-Preserving Aerodynamic Network (FPANet). Based on the strong correlation between the aerodynamic characteristics and geometric features of the aircraft, the feature-preserving downsampling algorithm is first used to pre-process the original point cloud to construct a low-redundancy input with significant geometric features; then, a feature-preserving downsampling module based on the PointNet++ architecture is innovatively constructed to replace the original farthest point sampling module, and through the multi-scale geometric feature enhancement mechanism, the network is explicitly guided to focus on the key local features related to aerodynamics. Experimental results show that FPANet shows a significant improvement in generalization performance in unknown data testing. This study provides a deep learning solution with strong generalization ability for the intelligent prediction of aircraft aerodynamic characteristics.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Research on Missile Aerodynamic Prediction Based on Feature Preserving Downsampling

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Qing Li,
  • Sumei Xiao,
  • Zhiyong Wang,
  • Qisong Xiao,
  • Shijiexi Gao,
  • Kaiting Li

摘要

In view of the insufficient generalization performance and overfitting problems of the existing aerodynamic prediction model based on the PointNet architecture in aerodynamic prediction tasks, this study proposes an aerodynamic parameter prediction model based on the PointNet++ architecture and fused with feature-preserving downsampling - Feature-Preserving Aerodynamic Network (FPANet). Based on the strong correlation between the aerodynamic characteristics and geometric features of the aircraft, the feature-preserving downsampling algorithm is first used to pre-process the original point cloud to construct a low-redundancy input with significant geometric features; then, a feature-preserving downsampling module based on the PointNet++ architecture is innovatively constructed to replace the original farthest point sampling module, and through the multi-scale geometric feature enhancement mechanism, the network is explicitly guided to focus on the key local features related to aerodynamics. Experimental results show that FPANet shows a significant improvement in generalization performance in unknown data testing. This study provides a deep learning solution with strong generalization ability for the intelligent prediction of aircraft aerodynamic characteristics.