In-Flight Learning Based Flight Control of an Unmanned Aircraft System
Abstract
Unmanned Aerial Vehicles (UAVs) popularity has increased substantially in the last few
years. UAVs capabilities continue to improve as a result of advances in battery technology, communication, navigation systems and electronics. Increased popularity has driven
researchers to improve UAVs reliability and safety which is reflected by the number of
publications and accelerating educational programs interest. UAVs are suited for a wide
range of civilian and military applications; however, UAVs currently can not integrate
with civilian airspace because of stringent safety requirements. Hence, it is necessary to
push the envelope for UAVs design and control so that they can learn from nature and have
more self-aware capabilities to improve safety and reliability. This dissertation addresses
some challenges involved with flight controller learning based on real-time modeling of
UAV.
Plenty of UAV applications require different operational capabilities within a composite mission. These capabilities include landing and taking off using short runways,
while being able to perform missions that require a high cruise speed i.e. tracking applications. A composite mission also requires the aircraft to be able to hover or operate with
low cruise speeds for applications involving stationary moments. All of these different
operational modes require a hybrid aircraft design that combines fixed wing aircraft capabilities and Vertical Take-Off and Landing (VTOL) aircraft capabilities. However, extensive resources required for hybrid aircraft design prohibited the discovery of different
revolutionary designs. The work presented in this dissertation describes the development
of a rapid modeling, prototyping and controller design platform of an unmanned quadrotor aircraft. Three main objectives are investigated: intelligent excitation input design,
real-time parameter estimation, and learning control.
Real-time estimation of dynamic model parameters is important for control adaptation. However, the aircraft model estimation performance can be severely degraded
by an active control system and highly collinear model terms such as those found on a
quadrotor unmanned aircraft. Recursive Fourier Transform Regression was applied to estimate parameters of different model forms/structures and using different excitation levels.
The generated models are utilized to reconfigure a Nonlinear Dynamic Inversion (NDI)
controller considering different testing conditions: normal, failure, and learning flights.
Finally,an intelligent input design technique is proposed which enables autonomous identification of the vehicle’s response modal frequencies and emphasizes excitation power
accordingly.
Table of Contents
Introduction -- Literature review -- Real-time closed loop system identification of a Quad-copter -- Flight controller learning based on real-time model estimation of a quadrotor aircraft -- Unmanned aircraft system intelligent system identification experiment design -- Conclusion and future work -- Appendix A. Power spectrum of a multisine signal -- Appendix B. Power spectrum of a multisine signal
Degree
Ph.D. (Doctor of Philosophy)