面向飞行试验认知不确定性的气动数据融合方法

Aerodynamic data fusion method under epistemic uncertainty in flight tests

在飞机设计领域中,不同的气动数据获取手段各有利弊,仅靠单一手段难以精确预测飞机的气动特性。因此,在实际工程应用中通常需要融合多种来源的数据,以获得更为准确和全面的气动特性描述。针对这一需求,以典型喷气式飞机为例,采用真实飞行数据、模拟飞行数据以及计算流体力学(CFD)仿真数据,结合深度神经网络,提出了一种认知不确定性的气动数据双层深度证据融合算法。该算法通过引入两种标准的置信分配方法,并将深度神经网络的输出与变分狄利克雷分布参数相结合,来表达和量化模型融合过程中的认知不确定性,并借助Dempster-Shafer理论有效地融合不同来源的数据及其不确定性。研究结果表明,该算法有效地融合了多源气动数据,所得结果不仅更加符合物理规律,而且提供了更高精度和更全面的气动数据,相比于单一数据源具有明显优势。

In the field of aircraft design,various methods for obtaining aerodynamic data have their own advantages and disadvantages,making it a challenge to accurately predict an aircraft's aerodynamic characteristics using a single approach.Therefore,in practical engineering applications,it is often necessary to fuse data from multiple sources to achieve a more accurate and comprehensive description of aerodynamic characteristics.In response to this need,this study takes the typical jet aircraft as an example and employs real flight data,simulated flight data,and Computational Fluid Dynamics(CFD)data.By combining these with deep neural networks,we propose a dual-level deep evidential fusion algorithm for aerodynamic data under epistemic uncertainty.In this algorithm,two standard confidence distribution methods are introduced,and the output of deep neural networks is combined with variational Dirichlet distribution parameters to express and quantify epistemic uncertainty during the model fusion process.Utilizing the Dempster-Shafer theory,this algorithm effectively fuses data from different sources and their associated uncertainties.The results indicate that this algorithm successfully fuses multi-source aerodynamic data,producing outcomes that not only conform to physical laws but also provide more accurate and comprehensive aerodynamic data,which demonstrate significant advantages over single-source data methods.