Neuro-fuzzy-based reduced-order modeling for aerodynamic loads computation at high-speed buffet buffeting

在飞机飞行包线极限处精确确定空气动力载荷对于减少设计过程中的不确定性以及根据质量和刚度分布产生优化结构至关重要。因此，在研究小组FOR 2895中，这个子项目的总体目标是开发一类基于神经模糊的降阶模型（ROMs），可用于跨音速冲击/抖振的有效负载计算和分析。这些模型的准确性、鲁棒性和对马赫数、雷诺数和攻角变化的敏感性将被深入研究。因此，将降低阶模型的结果与实验和数值参考数据进行比较。为了能够以足够的精度再现非定常气动载荷的频率和幅值相关特征，rom的训练过程必须适应占主导地位的流动物理效应，这给rom的训练带来了特殊的挑战。因此，可以记录实际运输机构型（XRF-1）在高速失速条件下的瞬态气动载荷特性。在rom方面，目前的子项目侧重于神经网络（“多层感知器网络”、“长短期记忆网络”、“卷积神经网络”），这些神经网络用于映射随机激波运动结合局部流动分离导致的非定常气动载荷，因此，代表了强非线性相互作用。在第一个项目阶段，有可能获得相对于参考数据非常精确的结果。该模型将在第二个项目阶段继续进行，并通过扩展来考虑振动对流动物理机制的影响。这些研究的动机之一是，在ETW中为风洞模型测量的压力谱在特征机翼结构特征模态上显示出峰值。在ETW中进行的有目标的结构振动激励的附加实验应该提供关于这一点的进一步数据。通过实验和数值模拟，以及受其影响的ROM方法，分析了叠加振动对冲击情景的影响。此外，为了进一步详细分析，为了表示动力气动弹性系统，进行了结构响应的单向和双向耦合形式的计算，从而开发了有关结构动力学分析的耦合系统的ROM方法。

The accurate aerodynamic loads determination at the limits of the aircraft’s flight envelope is essential for reducing uncertainties within the design process and to yield optimized structures with respect to mass and stiffness distributions. The overall goal of this subproject within the research group FOR 2895 is, therefore, to develop a class of neuro-fuzzy-based reduced-order models (ROMs) that can be used for efficient load computation and analysis at transonic buffet/buffeting. These models will be thoroughly investigated concerning their accuracy, robustness and sensitivity against variations of Mach number, Reynolds number and angle of attack. Therefore, the results of the reduced-order models will be evaluated in comparison to experimental and numerical reference data. Particular challenges arise for the training process of the ROMs, which must be adapted to the dominating flow physics effects in order to be able to reproduce the frequency and amplitude related characteristics of the unsteady aerodynamic loads with sufficient accuracy. Hence, the transient aerodynamic load characteristics of a realistic transport aircraft configuration (XRF-1) under high-speed stall conditions can be recorded. With regard to the ROMs, the present subproject focuses on neural networks (“Multi-Layer Perceptron Network”, “Long-Short-Term Memory Network”, “Convolutional Neural Network”), which are used to map unsteady aerodynamic loads as a result of stochastic shock motions combined with local flow separation, thus, representing strongly nonlinear interactions. In the first project phase, it was possible to achieve results of very good accuracy relative to the reference data. This modeling will be continued in the second project phase and supplemented by an extension to consider the influence of vibrations on the flow-physical mechanisms. These investigations are motivated, among others, by the fact that the pressure spectra measured for the wind tunnel model in the ETW showed peaks at characteristic wing structural eigenmodes. Additional experiments in the ETW with targeted excitation of structural vibrations should provide further data on this. The influence of superimposed vibrations on the buffet scenario is thus analyzed by experimental and numerical simulations and ROM methods conditioned by them. In addition, for a further detailed analyses calculations in the form of one-way and two-way couplings for the structural response are carried out in order to represent the dynamic aeroelastic system and, thereby, to develop ROM methods for the coupled system with regard to structural dynamics analyses.