Please use this identifier to cite or link to this item:
Title: Stable flight control of a tri-rotor drone using artificial neural networks
Authors: Sciberras, Mark
Keywords: Drone aircraft -- Control systems
Neural networks (Computer science)
Computer algorithms
Issue Date: 2018
Citation: Sciberras, M. (2018). Stable flight control of a tri-rotor drone using artificial neural networks (Bachelor's dissertation).
Abstract: Unmanned air vehicles, more specifically quad-rotor drones, are currently being extensively researched. This is due to the wide range of uses they present in multiple fields, such as search and rescue operations, research, and leisure. The quad-rotor drone is the most common type of drone, but it is also heavy and expensive to produce, hindering the performance and affordability of the product. By reducing the number of motors to three, and improving the control system of the drone, one can achieve improved agility at a reduced price point. The aim of this study is to show that a control system consisting of artificial neural networks can match and outperform the widely used proportional integral derivative control system, especially when considering systems consisting of a large number of inputs and outputs. Multiple machine learning methods and optimisation algorithms, more specifically, back propagation, genetic algorithms and proximal policy optimisation, were implemented and tested to determine which one of the algorithms produced the best step input response. The proximal policy optimisation algorithm was then deemed to be the best of the three and was used to train a tri-rotor drone model using real-world specifications. This was done by making use of a widely used game development engine, namely the Unity3D game engine, which contains a full-fledged graphics engine and a real-world physics engine. The artificial neural networks were trained, and the results were observed visually in real time. Learning curves of the training processes were also produced in order to be able to observe the results and the artificial neural networks’ training progress. Using these methods, it was observed that using reinforcement learning; more specifically the proximal policy optimisation algorithm, produced an adequate response in 6 degrees of freedom, when taking into consideration the limitations in experimentation and the complexity of the model. These results indicate that tri-rotor drones controlled by artificial neural networks can have comparable flight stability and reliability to those of quad-rotor and PID controlled drones. Further research can be conducted as the fields of machine learning and artificial intelligence grow, so as to reduce learning times and increase flight stability.
Appears in Collections:Dissertations - FacICT - 2018
Dissertations - FacICTCCE - 2018

Files in This Item:
File Description SizeFormat 
  Restricted Access
2.75 MBAdobe PDFView/Open Request a copy

Items in OAR@UM are protected by copyright, with all rights reserved, unless otherwise indicated.