准确地感知飞行中的大气数据，如：马赫数、攻角、侧滑角、静压等，对于飞行器的导航、制导与控制至关重要。传统的大气数据感知基于探针式的测量技术，以空速管与角度传感器为典型代表，其原理简单，应用广泛。然而，伸出机体的探针不能承受高速飞行工况下的高热环境，同时降低了飞行器隐身性能，在某些特殊情况下还可能影响飞行器的侧向稳定性。为了解决上述问题，NASA发展了嵌入式大气数据传感（Flush Air Data Sensing，FADS）系统的概念，其基本原理是利用飞行器表面的分布式压力传感器间接解算大气数据。作为一种非侵入式的大气数据测量方式，FADS系统不会改变飞行器气动外形，不破坏飞行器隐身特性，可适用于高超声速飞行器。然而，经典的FADS解算方法大多只适用于钝头前体飞行器，导致当前FADS技术的应用场景还十分局限。近些年随着航空航天技术的发展，高超声速飞行器的气动布局由以往的大钝头升力体布局朝着尖前缘高升阻比布局方向发展，FADS技术需要有更好的通用性和普适性，以适用更多一般外形的飞行器。面向此类场景的FADS技术需要特殊设计复杂解算流程，系统构建困难，解算精度不高，发展还很不成熟，亟待展开相关研究。针对上述问题，本文面向一般外形飞行器，针对FADS系统空气动力学建模、解算方法、故障诊断及容错回归、传感器布局优化等问题展开研究，主要工作如下：

传统的FADS空气动力学模型是在理想球形假设下得到的，以此发展的FADS算法只能适用于钝头前体飞行器。为了克服这一问题，本文基于量纲分析，定义了一组用于求解大气数据的无量纲参数，提出了能够适用于一般外形飞行器的FADS广义空气动力学模型。基于符号回归方法，发展了相应的大气数据解析算法：广义三点法 。为了验证提出算法，设计了具有较尖前缘的超声速飞行器前体标准模型，利用CFD数值模拟建立了FADS系统标准仿真数据库，并在该数据库上对该算法进行了验证。结果表明，相比于传统的三点法，广义三点法可以不仅适用于钝头前体飞行器，还适用于尖楔前体飞行器。

凭借着强大的非线性拟合能力，神经网络可以拟合任意非线性复杂函数，非常适用于FADS系统。然而，传统的FADS神经网络算法需要大量的训练数据，这些数据的获取十分耗时耗力，大大限制了神经网络算法应用。为了克服上述问题，本文使用神经网络来拟合FADS广义空气动力学模型，提出了基于无量纲输入输出神经网络的嵌入式大气数据传感（Flush Air Data Sensing Based on Dimensionless Input and Output Neural Networks, DIONNFADS）算法。该算法解耦了自由来流静压，可以将神经网络的训练样本的数目降低一个数量级。通过对训练神经网络的样本数目的收敛性分析，评估了用于训练神经网络的最佳样本数目。此外，讨论了算法对不同水平噪声及偏置的容忍程度和不同压力孔组合下的精度表现。

压力传感器的故障诊断与回归算法的容错能力是FADS系统的重要组成部分，投票等常规融合决策方法仅能对简单故障进行处理，且存在误判的可能性，难以处理传感器偏置等复杂故障。为了克服上述问题，提出了基于无量纲输入输出卷积神经网络的大气数据容错回归方法，该方法在回归大气数据的同时，反向计算飞行器壁面压力分布，并度量不同的测压孔组合下与测得的压力分布之间的差异，提出了以最小误差准则筛选传感器组合进行故障诊断与容错回归。相比于传统的诊断方法，该方法通过内置物理模型极大地避免了故障误判。结果表明，该方法能够适用于各种测压孔布局形式，对传感器偏置故障具有良好的容错回归能力。

面向一般飞行器FADS系统的传感器布置在工程实践中受诸多因素约束，传感器布局形式又直接决定系统解算精度，因此传感器布局的优化研究十分重要。通过研究多种典型的机器学习回归算法，基于FADS广义空气动力学模型发展了高效的大气数据回归代理算法，在此基础上，提出了面向一般飞行器全机的压力传感器布局优化框架。对不同位置、不同传感器数目约束下的传感器布局进行了优化研究，利用DIONNFADS分析了几种优化结果的回归精度。结果表明，对于无驻点布局、机翼前缘布局等典型约束，优化后的结果均能能够满足大气数据回归需求。

嵌入物理信息神经网络（Physics-Informed Neural Networks，PINN）在偏微分方程“逆”问题上表现出巨大潜力。受此启发，提出了基于PINN的大气数据解算方法，将“基于飞行器表面压力分布求解大气数据的问题”转化为“已知流场局部观测量求解Euler方程初始或边界条件的‘逆’问题”。该方法不需要任何先验样本就可以直接求得大气数据。提出了面向PINN的自适应的迁移学习技术，通过在更高维参数空间定义最小能量路径，解决了PINN高维非凸损失函数导致的优化过程不稳定问题，实现了高效的PINN大气数据求解。

Accurate sensing of air data states, such as Mach number, angle of attack, sideslip, and static pressure, etc. is essential for guidance, navigation, and control of vehicles. Conventional air data sensing techniques involve intrusive booms, e.g., pitot probe and angular transducer, which have a simple working principle and have been widely applied in all kinds of vehicles. However, these intrusive booms cannot withstand the tremendous heat load under super or hypersonic flight conditions, increase radar cross-section area for surface vessels, and cause unwanted lateral instabilities for some vehicles with high angles of attack. To deal with the above problems, NASA developed the concept of the flush air data sensing (FADS) system that arranges pressure ports on the vehicle surface and measures pressure distribution to indirectly solve air data states. As a completely non-intrusive technique, the FADS system can avoid the hypersonic heat load caused by the flow sensing booms, does not increase the radar cross-section area and affect the instabilities of the vehicle. However, the classical FADS aerodynamic model and algorithms are only suitable for vehicles with blunt noses, resulting in the limited applications of the current FADS technology. With the development of aerospace technology in recent years, the aerodynamic configuration of hypersonic vehicles from the previous lifting-body configuration with the large blunt nose towards the direction of high lift-to-drag ratio configuration sharp with leading-edge, FADS techniques need to be applied to vehicles with more general aerodynamic configuration. Due to the lack of a clear aerodynamic model, the FADS system for vehicles with general configuration need to be specially designed for complex solution process, which has insufficient accuracy and is difficult to construct. Many problems remain to be studied and solved in the application of the FADS system for vehicles with general configuration. Aiming to solve the bottleneck problems, several key issues of the FADS system for vehicles with general configuration are researched, including aerodynamic modeling, the algorithms for air data solution, fault diagnosis and fault-tolerant regression, and the optimization of the pressure port layout. The main work is arranged as follows:

The conventional FADS aerodynamic model is obtained based on the spherical shape, and the corresponding algorithms are only suitable for vehicles with blunt noses. To overcome this problem, this paper defines a set of dimensionless parameters for solving air data states and proposes the generalized FADS aerodynamic model for vehicles with general aerodynamic configuration. Based on the symbolic regression method, the generalized triple algorithm is developed. To verify the proposed algorithm, a standard front model of a supersonic vehicle is designed, and a standard database for the FADS system is calculated by using CFD numerical simulations. Compared with the traditional triple algorithm, this generalized triple algorithm can be applied not only to vehicles with blunt nose, but also to vehicles with sharp lead-edge.

With strong fitting capabilities, neural networks can approximate complex non-linear relationships between the input and output variables of the system that is very well suited to the FADS system. However, traditional FADS neural network algorithms require a large amount of training data, which is too expensive and time-consuming to limit the application of the FADS neural network algorithm. To overcome the above problems, neural networks are used to approximate the aerodynamic model defined by dimensional analysis, and a dimensionless input and output neural networks flush air data sensing (DIONNFADS) algorithm is proposed for solving air data states. Since the model decouples the free stream static pressure, the number of training data for neural networks are effectively is reduced by an order of magnitude. The convergence analysis for the number of the training data is performed, and the optimal scale is evaluated. In addition, the tolerance to different levels of noise and bias is discussed, and the accuracy performance under different pressure port layouts is also analyzed.

The fault diagnosis of pressure sensors and the fault-tolerant capability of the regression algorithm are important components of the FADS system. The current idea is to directly fuse the possible solutions of multiple regressions by decision methods such as voting, which has the possibility of misclassification and is difficult to deal with faults such as sensor bias. To overcome the above problems, a fault-tolerant regression method based on dimensionless input-output convolutional neural networks is proposed. First, the method calculates the pressure distribution on the surface of the vehicle in reverse while regressing air data states. Second, the distance between the measured and the calculated pressure distribution under different combinations of pressure ports is measured. Finally, a minimum error principle to select a suitable combination of pressure ports is proposed. Compared with the traditional fault diagnosis and tolerance regression, this method takes the physical model into account, which greatly avoids the occurrence of misclassification and can be applied to all kinds of biases. The results show that the method can be applied to different pressure port layouts and has high diagnostic accuracy for fault sensors and good fault-tolerant capability for different biases.

The pressure port layout for vehicles with more general aerodynamic configuration directly determines the accuracy of the FADS system, so the optimization of the pressure port layout is vitally important. The classical FADS algorithm often requires a lot of modeling time, which limits the optimization of pressure port layout. To solve the problem, a fast and efficient regression agent algorithm is developed based on the FADS generalized aerodynamic model by studying several typical machine learning algorithms. Based on genetic algorithms, a pressure port layout optimization framework is proposed for vehicles with more general aerodynamic configurations. Aiming at the critical problems in engineering practice, the optimization of the layout under different position and port number constraints is studied, and the regression accuracy of several optimization results is analyzed by DIONNFADS. The results show that the optimized results can meet the regression requirements for several typical layout constraints, such as the layout without stagnation point and the layout on the leading edge.

Physics-informed neural networks (PINN) have shown great potential in solving the "inverse" problems of partial differential equations and have received much attention from researchers. Inspired by this, a method for solving air data states based on physics-informed neural networks is proposed, which transforms "the problem of solving air data states based on the pressure distribution on the aircraft surface" into "the problem of solving the initial or boundary conditions of the Euler equation with known local observations of the flow field". The method does not require any prior samples to directly solve for air data states. Because of the high-dimensional and non-convex loss function, the optimization process of PINN is unsteady and very easy to fall into local optimal solutions. To overcome this problem, this paper proposes an adaptive transfer learning method for PINN, which essentially defines a minimum energy path in the high-dimensional parameter space along which the PINN can be adaptively transferred from the source task to the target task, effectively ensuring the stability of the PINN optimization process.