关键词: 点云, 注意力机制, 迁移学习, 误差补偿, 三维飞行器, 气动性能, 智能预测

Abstract: This paper proposes a point cloud dual-branch Transformer aerodynamic prediction model based on transfer learning and error compensation for intelligent end-to-end prediction of aerodynamic forces of three-dimensional rocket and missile type axisymmetric vehicles. To address the challenges of difficulty in acquiring high-fidelity simulation data and insufficient sample size for three-dimensional vehicles, a point cloud self-attention backbone network is first trained on large-scale low-fidelity data as pre-trained model to capture the fundamental characteristics and trends of aerodynamic variations. Subsequently, a point cloud branch network combined with self-attention and cross-attention is employed to learn the residuals between small-scale low-fidelity and high-fidelity data as a fine-tuning model, enabling error correction of low-fidelity data and fusion of multi-fidelity data. Low-fidelity and high-fidelity aerodynamic datasets are constructed using engineering calculation methods and computational fluid dynamics (CFD) simulations, respectively. Experimental results on these datasets demonstrate that the proposed method reduces the mean relative error by 74.8% as compared to models trained directly with high-fidelity data, and achieves a 16.5% reduction compared to baseline models with direct fine-tuning. These results thoroughly validate the effectiveness of combining transfer learning with error correction. The proposed method provides an effective approach for fast and reliable prediction of aerodynamic performance of complex aircraft configurations during the aircraft design phase, and is expected to be extended to few-shot learning problems in other engineering domains.

Key words: point cloud, attention mechanism, transfer learning, error correction, 3D aerospace vehicles, aerodynamic performance, intelligent prediction