基于迁移学习的飞行器高低阶精度数据融合方法  

High and low order precision data fusion method for aircraft based on transfer learning

作　　者：崔榕峰 李鸿岩 王祥云 刘哲 郭承鹏 CUI Rongfeng;LI Hongyan;WANG Xiangyun;LIU Zhe;GUO Chengpeng(AVIC Aerodynamics Research Institute,Shenyang 110034,China)

机构地区：[1]航空工业空气动力研究院,辽宁沈阳110034

出　　处：《飞行力学》2024年第4期7-12,20,共7页Flight Dynamics

基　　金：航空科学基金资助(2022Z006026004,2023M071027001)。

摘　　要：为提升飞行器气动性能分析效率、降低所需成本,采用了两种基于迁移学习的数据融合智能预测方法:一种是根据样本特征进行权重配比的多保真度融合方法,利用自适应提升算法进行样本错误率评估,并依据结果进行加权融合;另一种是基于模型的参数冻结迁移方法,将高低阶精度数据进行神经网络分层训练,实现模型意义上的数据融合。两种方法均考虑将高阶精度的风洞试验数据与低阶精度的CFD数据进行深度融合训练,从而实现精准预测。以YF-16飞机标模为例进行预测分析,结果表明基于迁移学习的数据融合方法能够实现气动力的准确预测,并在精度上超过了传统CoKriging融合方法。In order to enhance the efficacy of aircraft aerodynamic performance analysis and reduce expenditure,two data fusion methods based on transfer learning are implemented.The first one is a multi-fidelity fusion method with weight distribution based on sample features.In this case,an adap-tive boosting algorithm is employed to assess the sample eror rate,and the weighted fusion is con-ducted by using the error rate.Another method is a parameter freezing transfer method based on deep neural network models.This method trains high-order and low-order precision data into network layers with the objective of achieving data fusion in the sense of the model.Both methods consider deep fusion training of high-order precision wind tunnel test data and low-order precision CFD data,fully leveraging the correlation between high-order and low-order data,thereby improving prediction accuracy.The results of the analysis of the YF-16 airplane model demonstrate that the data fusion method based on transfer learning can achieve accurate prediction of aerodynamic forces and returns more precise predictions compared with the traditional CoKriging method.

关 键 词：数据融合 迁移学习 风洞试验 CFD计算 深度神经网络 

分 类 号：V211.4[航空宇航科学与技术—航空宇航推进理论与工程]