现代飞行器的设计要求正逐渐向高机动性、高敏捷性、可操控性等方向发展，需要在更复杂的实际环境中具备良好的飞行性能。因此，新一代飞行器在设计过程中需要对大迎角下的非定常气动力特性进行精确描述与分析。计算流体力学（Computational Fluid Dynamics, CFD）发展至今，已经可以对复杂的非定常和非线性流动问题进行高精度数值模拟，并实现多物理场耦合分析，是飞行空气动力学领域的重要研究工具。然而CFD数值仿真技术在求解非定常流场时需要消耗大量计算资源，其计算效率也随着非定常流场数据精度和数据量的提高而迅速下降。为平衡工程应用中对海量气动数据的需求和CFD计算成本，自上世纪以来一系列非定常气动力模型被提出，利用有限的风洞试验数据或CFD仿真数据对流场特征进行提取，其流场预测结果实现了逼近CFD数值仿真精度的同时，大大降低了计算量和计算时间。

本学位论文在基于CFD技术的非定常气动力建模的研究背景下，分别探讨了数据增强和组合动态模型结构设计对模型性能提升的作用效果。主要开展了以下工作：

（1）提出了一种基于神经网络的自适应数据增强算法，对超声速下二维翼型的单自由度俯仰简谐振荡气动力数据进行数据特征提取和数据集优化，用于BP神经网络气动力模型的训练。通过提取各运动状态下非定常气动力数据的几何特征，对各组数据内部的迟滞效应和非线性特征进行自适应的特征增强，并研究了数据集相关参数对模型精度和泛化性能的影响，探讨了最适合数据增强算法的适用条件。

（2）针对跨声速下二维翼型的多级俯仰强迫振荡气动力数据，建立了一种基于Volterra泛函级数和LSTM神经网络的组合动态模型。通过低阶Volterra核函数辨识对复杂气动力响应的线性和弱非线性动态特征进行解耦分析，同时利用LSTM神经网络对Volterra级数模型的预测残差进行预测，实现对系统强非线性成分的增量建模，并在多种运动模态的非定常算例气动仿真结果上进行泛化测试。结果表明，组合动态模型针对复杂非定常流场具有较强的预测能力。

The design requirements of modern aircrafts are getting more and more focused on high mobility, high agility and maneuverability, which requires good flight performance in a complex real environment. Therefore, the unsteady aerodynamic characteristics at high angles of attack need to be accurately described and analyzed in the design process of the new generation aircrafts. The Computational Fluid Dynamics (CFD) is an important research tool in the field of flight aerodynamics, which allows to carry out high-fidelity numerical simulation of complex unsteady and nonlinear flow problems and conduct multi-physics coupling analysis. However, CFD technology consumes substantial computational resources when solving unsteady flow fields, and its computational efficiency declines rapidly with the increase of data accuracy and data size of unsteady flow fields. In order to balance the demand of massive aerodynamic data and the cost of CFD calculation in engineering applications, a series of unsteady aerodynamic models have been proposed since the last century. The flow field characteristics are extracted by using limited wind tunnel test data or CFD simulation data. The flow field prediction results achieve the approximate accuracy of CFD numerical simulation, while significantly reducing the calculation quantities and time.

In this thesis, under the background of unsteady aerodynamic modeling based on CFD technology, the effects of data augmentation and combined dynamic model structure design on model performance improvement are discussed respectively. The main work is as follows:

(1) An adaptive data augmentation algorithm based on neural network is proposed to extract data features and optimize data set of the aerodynamic data of a two-dimensional airfoil in simple harmonic pitching oscillation pattern under supersonic, which is used for the training of BP neural network aerodynamic model. By extracting the geometric features of unsteady aerodynamic data under different motion states, the hysteresis and nonlinear effects in each group of data sets are enhanced adaptively. The influences of data set parameters on model accuracy and generalization performance are studied, and the most suitable application conditions of the data enhancement algorithm are discussed.

(2) A combined dynamic model based on Volterra functional series and LSTM neural network is established for the aerodynamic data of a two-dimensional airfoil in multistage forced pitching oscillation pattern under transonic velocity. Through the identification of low order Volterra kernel functions, the linear and small nonlinear dynamic characteristics of complex aerodynamic responses are decoupled. Meanwhile, LSTM neural network is used to handle the predictive residuals of Volterra series model, and the incremental modeling of strong nonlinear part of the aerodynamic system is conducted. Generalization tests are carried out on the aerodynamic simulation results of unsteady computational examples under various motion modes. The results show that the combined dynamic model has a strong ability to predict the complex unsteady flow field.