飞行器多舵面气动力智能融合建模方法

An intelligent fusion modeling method for aerodynamic forces of multi-control-surface aircraft

关键词:  深度神经网络(DNN)  卷积神经网络(CNN)  数据融合建模  强非线性  数据驱动

Key words:  deep neural network (DNN)  convolutional neural network (CNN)  data-fusion modeling  strong nonlinearity  data-driven

摘要:

常规布局飞行器的数据库构建通常采用来流维度全采样,由于各个舵面距离远、干扰弱,常采用叠加舵效增量的策略,避免了舵面组合的全维度采样。但飞翼布局飞行器舵面间的气动干扰强、舵面数量多,考虑舵面间干扰的全维度采样工程代价极高。为攻克在少量舵面组合样本约束下,构建考虑舵面间非线性干扰的气动力模型这一飞机设计的主要难点。首先,针对迎角采样稠密的重要特征（如风洞吹风）,提出了融合卷积神经网络（CNN）与工程模型的气动力智能建模方法。其次,以低速飞翼后缘3组舵面组合偏转的气动力为研究对象,运用CFD方法获得单/双舵偏的气动力数据。最后,采用组合舵效与相邻舵间干扰量线性叠加方法构建低精度工程模型,再通过引入迎角序列建模机制,运用卷积神经网络进一步表征舵面间与迎角维度的非线性干扰效应。结果表明,所提方法相较于工程模型,精度提升约40%；与未嵌入工程模型的直接深度神经网络模型相比,其预测误差标准差降低50%以上。本研究显著提高了飞翼布局飞行器在大迎角、大舵偏下的强非线性气动建模的准确性与稳健性。

Abstract:

The aerodynamic database construction for conventional aircraft configurations typically employs full sampling across inflow dimensions. Given the substantial distance and weak interference between control surfaces, the strategy of superimposing control effectiveness increments is widely used, which avoids the need for a full combinatorial sampling of all control surface deflections. However, for flying-wing aircraft, the presence of strong aerodynamic interference among multiple closely-spaced control surfaces makes full combinatorial sampling that accounts for these interactions prohibitively expensive. This paper focuses on the critical challenge of constructing an aerodynamic model capable of capturing nonlinear interference between control surfaces under limited combinatorial sample. Firstly, an intelligent aerodynamic modeling method is proposed that integrates a convolutional neural network (CNN) with an engineering model, specifically for scenarios with dense angle-of-attack sampling such as wind tunnel tests. Secondly, to characterize the aerodynamic forces generated by 3 trailing-edge control surface combinations on a low-speed flying-wing configuration, high-fidelity CFD simulations are used to obtain aerodynamic data for single and dual control surface deflections. Finally, a low-fidelity engineering model is constructed using a method that linearly superimposes individual control effectiveness and interference increments between adjacent control surfaces. Then, by introducing an angle-of-attack sequence modeling mechanism, a CNN is applied to further characterize nonlinear interference effects both among the control surfaces and across the angle-of-attack dimension. Results indicate that the proposed method not only improves accuracy by approximately 40% compared to the engineering model, but also reduces the standard deviation of prediction error by over 50% relative to a direct deep neural network model without embedded engineering knowledge. This research significantly enhances both accuracy and robustness of strongly nonlinear aerodynamic modeling for flying-wing aircraft under high-angle-of-attack and large-control-surface-deflection conditions.