机器学习数据融合方法在火箭子级栅格舵气动特性建模应用中的比较研究

机器学习数据融合方法可帮助降低飞行器气动数据库建立的成本，加快研制进度，目前已经成为飞行器设计方法领域越来越活跃的研究方向，但其在工程复杂问题方面的应用研究并不充分。将多种常见变可信度数据融合模型应用于运载火箭子级栅格舵落区控制的工程项目，在开展部分工况的风洞试验基础上，结合少量的CFD数值模拟结果，研究相关函数和不同模型预测完整工况气动特性数据的差异性。通过对比加法标度函数修正模型、Co-Kriging模型、分层Kriging模型和多可信度神经网络模型等4种不同的数据融合模型发现:高斯指数相关函数对气动建模问题的适应性更好；Co-Kriging模型对气动数据的内插表现最好；分层Kriging模型对内插的预测精度较高，外插效果不理想；多可信度神经网络模型在外插区域能获得更光滑、合理的预测结果。

Comparison of machine learning data fusion methods applied to aerodynamic modeling of rocket first stage with grid fins

Machine learning data fusion method has attracted significant attention recently in aerodynamic database construction since it makes a trade-off between high prediction accuracy and low fitting cost by fusing samples of different fidelities. But the research on methods for complex engineering project is not sufficient. In this paper, several commonly used variable-fidelity models （VFMs） of data fusion are applied to the control law design in the rocket first stage landing area control project with grid fins. Based on wind tunnel tests of partial test states, combined with CFD simulation results, VFMs successfully predict the whole aerodynamic characteristics of grid fins. Here, our objective is to compare the performances of these four VFM methods （AS-MFS, Co-Kriging, HK, MFNN） and the results show that: Gaussian exponential function is more suitable for aerodynamic modeling problems; Co-Kriging has the best performance in the interpolation of aerodynamic data; HK model has high prediction accuracy for interpolation but has poor performance for extrapolation; MFNN model can obtain smoother and more reasonable results in the extrapolation region.