Machine Learning for Agile Aircraft Control
Lead Research Organisation:
University of Bristol
Department Name: Aerospace Engineering
Abstract
Control of aircraft post-stall and at high angles of attack presents a significant challenge for many control techniques. Machine learning has the potential to effectively shift the problem of designing control algorithms to a neural network or other machine learning architecture. However, this still presents significant challenges in defining both the machine learning architecture and the associated objective functions.
While aircraft control is well established for conventional control surfaces, it is hoped that a machine learning approach will be readily extendable to novel control surface arrangements. If the machine learning process can be made sufficiently robust and is able to learn a suitable control scheme for a new configuration quickly enough, it should
allow for rapid testing of novel control surface arrangements.
Moving the learning onboard would provide scope for allowing the aircraft to adapt to changes in flight conditions. This has potential applications for damage tolerant control systems or control systems able to adapt to changing centre of gravity of the aircraft.
Initial work will likely focus on extending previous research undertaken at the University of Bristol into using machine learning to perform perching manoeuvres. The work looked at a relatively narrow range of starting conditions. Extending these starting conditions to a a wider range is a likely first step. The work can further be extended by changing the manoeuvre to be learned and attempting to achieve similar results for a wider range of manoeuvres. This work is likely to begin incorporating flight data into the training data, initially performing the processing offline and later developing onboard techniques.
Later work may look at novel control surface configurations. If a machine learning approach that begins with no knowledge of the underlying system can be developed for a conventional configuration, applying the same approach to novel configurations should allow development of a control scheme with minimal intervention.
While aircraft control is well established for conventional control surfaces, it is hoped that a machine learning approach will be readily extendable to novel control surface arrangements. If the machine learning process can be made sufficiently robust and is able to learn a suitable control scheme for a new configuration quickly enough, it should
allow for rapid testing of novel control surface arrangements.
Moving the learning onboard would provide scope for allowing the aircraft to adapt to changes in flight conditions. This has potential applications for damage tolerant control systems or control systems able to adapt to changing centre of gravity of the aircraft.
Initial work will likely focus on extending previous research undertaken at the University of Bristol into using machine learning to perform perching manoeuvres. The work looked at a relatively narrow range of starting conditions. Extending these starting conditions to a a wider range is a likely first step. The work can further be extended by changing the manoeuvre to be learned and attempting to achieve similar results for a wider range of manoeuvres. This work is likely to begin incorporating flight data into the training data, initially performing the processing offline and later developing onboard techniques.
Later work may look at novel control surface configurations. If a machine learning approach that begins with no knowledge of the underlying system can be developed for a conventional configuration, applying the same approach to novel configurations should allow development of a control scheme with minimal intervention.
Organisations
People |
ORCID iD |
Thomas Richardson (Primary Supervisor) | |
Robert Clarke (Student) |
Studentship Projects
Project Reference | Relationship | Related To | Start | End | Student Name |
---|---|---|---|---|---|
EP/N509619/1 | 30/09/2016 | 29/09/2021 | |||
1939762 | Studentship | EP/N509619/1 | 30/09/2017 | 16/03/2024 | Robert Clarke |
EP/R513179/1 | 30/09/2018 | 29/09/2023 | |||
1939762 | Studentship | EP/R513179/1 | 30/09/2017 | 16/03/2024 | Robert Clarke |