多可信度数据融合方法在重复使用火箭气动特性预示中的应用  

Application of Multi-fidelity Data Fusion Methods in the Prediction of Aerodynamic Characteristics for Reusable Rockets

作　　者：陈立为 秦曈 高武焕 胡彦辰 杜涛 唐伟 魏威 CHEN Liwei;QIN Tong;GAO Wuhuan;HU Yanchen;DU Tao;TANG Wei;WEI Wei(Beijing Institute of Aeronautical Systems Engineering,Beijing 100076,China)

机构地区：[1]北京宇航系统工程研究所,北京100076

出　　处：《宇航学报》2025年第2期310-319,共10页Journal of Astronautics

基　　金：国家自然科学基金(92470120)。

摘　　要：随着重复使用运载火箭技术的蓬勃发展,高效且精确地获取火箭在返回阶段的气动特性,对于加速研发进程和提升设计效率至关重要。传统神经网络建立的气动模型在遵循对称性和连续性等关键物理规律方面存在局限性。为解决这一问题,提出了一种多可信度神经网络方法。与直接将物理方程残差作为损失项的物理信息神经网络(PINN)不同,该方法以神经网络模块代替气动建模Taylor展开式系数,保持了气动系数相对攻角、侧滑角的线性和非线性关系的组合;并结合多可信度的层次化思想,有效提升了模型预示的准确度和稳定性。系统分析了数据量和网络规模对预示准确度的影响,预示结果的相对误差不超过10%,能够满足工程设计的严格要求,具有较高的工程应用价值。As reusable launch vehicle technology thrives,the accurate and efficient determination of rocket aerodynamic characteristics during re-entry is pivotal for expediting research and development,and enhancing design efficiency.Traditional neural network models,however,struggle to enforce essential physical laws like symmetry and continuity in aerodynamic models.To tackle this,a multi-fidelity neural network approach is introduced.Physical knowledge is incorporated into the network by replacing Taylor series expansion coefficients in aerodynamic modeling with neural network modules,preserving a mix of linear and nonlinear relationships between aerodynamic coefficients and angles of attack and sideslip,which is different from the physics-informed neural network(PINN)that directly uses the residual of physical equations as the loss term.Furthermore,multi-fidelity hierarchical concepts are integrated to enhance the model′s prediction accuracy and stability.A comprehensive analysis examines the influence of data quantity and network size on prediction accuracy,showing that the predictions′relative error remains below 10%,fulfilling the rigorous demands of engineering design and exhibiting substantial value for engineering applications.

关 键 词：数据融合 神经网络 知识表征 重复使用火箭 

分 类 号：V475.1[航空宇航科学与技术—飞行器设计]