基于数据融合的气动热稀疏风洞试验数据重构方法

Data Fusion-Based Method for Reconstruction of Aerodynamic Thermal Sparse Wind Tunnel Test Data

摘要:

由于风洞试验条件的局限，气动热风洞试验数据的空间分布较为稀疏，样本量也相对较少，给构建高性能的数据驱动气动热预测模型带来了挑战，因此结合多保真度数据融合建模思路提出气动热稀疏风洞试验数据重构方法。首先，在稀疏的气动热风洞试验数据基础上引入气动热低精度仿真计算数据，构建深度神经网络的训练集；其次，设计深度神经网络的加权损失函数。加权损失函数包含风洞试验数据损失项和低精度仿真计算数据损失项两部分；最后，训练深度神经网络，得到气动热稀疏风洞试验数据的重构结果。以不同风洞的高超声速双椭球、钝锥、钝双锥和25°/55°钝双锥的气动热稀疏风洞试验数据为例开展气动热重构，结果表明：不仅气动热重构结果和风洞试验数据的归一化均方根误差均在9% 以内，而且气动热重构数据量和低精度数值仿真结果相当，可以用云图的方式对气动热的分布进行详细展示。

Abstract:

Data-driven modeling methods have become one of the main techniques for predicting aerodynamic thermal behavior in hypersonic applications. However, due to the limitations of wind tunnel experimental conditions, the spatial distribution of aerodynamic thermal wind tunnel experiment data is often sparse, and the sample size is relatively small. This poses challenges in constructing high-performance data-driven aerodynamic thermal prediction models. To address these issues, this paper proposes a reconstruction method for sparse aerodynamic thermal wind tunnel experiment data, integrating a multi-fidelity data fusion modeling approach. First, low-fidelity simulation data for aerodynamic thermal calculations are introduced based on the sparse aerodynamic thermal wind tunnel experiment data to construct a training set for the deep neural network (DNN). Then, a weighted loss function for the DNN is designed. The weighted loss function consists of two components: the loss from the wind tunnel experimental data and the loss from the low-fidelity simulation data. Finally, the DNN is trained to obtain the reconstruction results of the sparse aerodynamic thermal wind tunnel experiment data. Aerodynamic thermal reconstruction is conducted using aerodynamic thermal sparse wind tunnel experiment data from hypersonic wind tunnels for different geometries, including double-ellipsoid, blunt-cone, bicone, and 25°/55° bicone. The results indicate that not only is the normalized root mean square error of the aerodynamic thermal reconstruction results within 9% of the wind tunnel experimental data, but the volume of the reconstructed aerodynamic thermal data is also comparable to that of the low-fidelity numerical simulation results, allowing for a detailed visualization of the aerodynamic thermal distribution in cloud plot form.