L1 adaptive control within Learn-to-Fly

Adaptive Control;L1;Learn-to-fly

Snyder, Steven

关键词: Adaptive Control;    L1;    Learn-to-fly;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/106252/SNYDER-DISSERTATION-2019.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

This research presents the development and flight testing of a learning control framework that was combined with real-time modeling and real-time guidance algorithms to test the feasibility of the NASA Learn-to-Fly concept.The objective of Learn-to-Fly is to reduce or eliminate the need for ground-based aerodynamic modeling in favor of in-flight modeling and control law determination. Due to the initial (high) uncertainty levels, an L1 adaptive control law is used to provide robustness.In order to apply the L1 controller within this Learn-to-Fly framework, extensions to the L1 adaptive control theory were required to enable the linear reference systems to be updated (which leads to switched reference systems) as the modeling algorithm adjusts the vehicle model and to be parameterized (which can be captured by linear parameter-varying (LPV) models) by different flight conditions.This document extends L1 control theory for switched linear reference systems and LPV reference systems and proves that the actual switched or LPV system can be made to behave arbitrarily close to its corresponding switched or LPV reference system by increasing the adaptation gain.Development of these extensions also resulted in new stability criteria, which are more suited for these types of reference systems than the traditional L1 norm stability criteria prevalent in the L1 adaptive control literature.In the simple case of a linear time-invariant reference system, the new conditions are shown to be less conservative than the existing L1 norm conditions, allowing more freedom in the control design.The developed learning control framework was tested on more than 30 flights.The Learn-to-Fly architecture successfully navigated for both stable and unstable vehicles without a representative apriori model.In the case of the unstable vehicle, the developed framework was able to autonomously fly it, despite a poor initial model that inaccurately indicated that the aircraft was stable.The flight tests show promising results and successfully establish the feasibility of the Learn-to-Fly concept.