Adaptive Flight Control to Enhance Survivability and Reduce Cost of Change

人工智能学习在控制系统设计中的一些应用已有报道。这些研究一直集中在具有相对简单动态特性的系统的控制器设计上，如汽车、机器人机械手或单摆。然而，众所周知，飞行器或无人机的动力学特性比较复杂：高度非线性、动态变化快、气动参数具有不确定性。因此，如何选择合适的人工智能学习架构，为这样一个复杂的系统找到可行和安全的解决方案，是这个博士项目的根本研究问题。另一个挑战是对解决方案的验证。基于AI学习的自动驾驶仪通常被认为是一种“积木式”的方法。这种方法的主要问题是，没有人能保证在真正的系统中实现它时是否不会产生任何问题。然而，以前的研究大多集中在展示可行性上，但缺乏对这种块盒方法进行验证的努力。也就是说，对黑箱行为和解的收敛的理解较少。在实际应用中，在实际系统中实现基于学习的自动驾驶仪时，这些对于确保基于学习的方法的性能和可靠性具有重要意义。本研究寻求一种适合于控制飞机或无人机等复杂动力学系统的人工智能学习体系结构。提出的创新将探索一种基于学习方法来理解自动驾驶行为的方法。主要贡献将是一种实用的基于学习方法的自动驾驶算法，它直接适用于实际系统，以确保可靠性的可信度。因此，科学价值和创新不仅在于开发一种新颖、实用和可重复使用的飞机或无人机自动驾驶算法，而且还在于基于理论分析、数值仿真和飞行试验对所提出的算法进行验证。本项目的主要目的是开发一种实用、安全的基于人工智能学习方法的自动驾驶算法，该算法可以根据模型变化轻松地重新配置其控制算法。总体方法是理论创新与数值模拟或飞行试验相结合。因此，具体目标将包括：a.审查最新的学习方法并调查它们对飞行控制系统的适用性；b.开发现实的飞行模型；c.基于最有前景的最先进的人工智能学习方法开发飞行控制系统；d.研究所开发的自适应飞行控制系统的适当架构；e.确定自适应飞行控制系统的相关性能指标；f.研究用于确认和验证自适应飞行控制系统的适当方法；g.使用飞行试验台和所开发的性能指标来分析和验证所设计的飞行控制系统。

A number of applications of AI learning to control system design have been reported. These studies have been focusing on the controller design for the systems that have relatively simple dynamic characteristics such as car, robot manipulator, or pendulum. However, it is well-known that the dynamics characteristics of aircraft or UAS are relatively complicated: high nonlinear, rapidly changing dynamics, and uncertainties of aerodynamic parameters. Thus, the fundamental research question of this PhD programme is how to select an appropriate architecture of AI learning in order to find feasible and safe solution for such a complicated system. Another challenge is the validation of solution. The autopilot based on AI learning is generally considered as a "block-box" approach. The main problem of such a kind of approach is that no one can guarantee whether or not it will not make any issues when implementing it in a real system. However, previous studies have mostly focused on showing the feasibility only, but there has been lack of effort to validate such a block-box approach. Namely, understandings of the behaviour of black-box and the convergence of solution have been less understood. In practice, these are important in ensuring confidence in the performance and reliability of learning-based approach when implementing the autopilot based on learning-based approach in a real system.This research seeks an appropriate AI learning architecture for controlling of complicated dynamics systems such as aircraft or UAS. The innovation proposed will investigate a way to understand the behaviour of autopilot based on learning approach. The principle contribution will be a practical autopilot algorithm based on learning approach which is directly applicable to real systems in ensuring confidence in the reliability. The scientific value and innovation thus lie in not only development of a novel and a practical and reusable autopilot algorithm for aircraft or UAS, but also validation of the proposed algorithm based on theoretical analyses, numerical simulation, and flight tests. The primary aim of this project is to develop a practical and safe autopilot algorithm based on AI learning approach that can easily reconfigure its control algorithm according to the model changes. The overall approach is a combination of theoretical innovation with numerical simulations or flight test. Therefore, the specific objectives will include: a. Review state-of-the-art learning approaches and investigate their applicability to the flight control system; b. Develop realistic flight models; c. Develop a flight control system based on the most promising state-of-art AI learning approaches; d. Investigate an appropriate architecture of the adaptive flight control system developed; e. Identify relevant performance metrics for the adaptive flight control system; f. Investigate appropriate methods for the validation and verification of adaptive flight control systems; g. Analyse and validate the designed flight control system using the flight test-bed and performance metrics developed.