Control d’actitud de satèl·lits artificials mitjançant aprenentatge automàtic

Abstract

[en] When a satellite moves around any orbit, it will encounter constant environmental perturbations that will produce small changes in the attitude (i.e. the orientation and angular velocity) that, in the long term, will cause unstable flying and rolling. Thus, the satellite will have to perform constant maneuvers in order to correct its attitude. An attitude controller is an algorithm that derives the appropiate maneuvers needed to correct the attitude error (i.e. the difference between the current and desired attitude), usually by calculating how much torque must be applied by the satellite actuators. The purpose of this project is the study of the attitude control procedures of artificial satellites (as rigid bodies) by fusing classical control algorithms with modern machine-learning techniques. This project consists of two clearly differentiated parts: the study of the attitude (Ch. 2-3) and the study of the attitude control (Ch. 4-5-6). In Chapters 2 and 3 we will explore the mathematical tools that will allow us to express a rotation in three dimensions: Euler rotation theorem and the unit quaternions. We will also derive the differential equations that dictate the attitude changes of the satellite when subject to external torques. In Chapter 4 we will study the different types of attitude errors and we will introduce some basic concepts and definitions about control algorithms. In Chapter 5 we will discuss the implementation of a virtual simulation environment that will allow us to test the attitude changes of a virtual satellite when subject to various torques. We will also implement a control law that will allow us to stabilize the satellite. In Chapter 6, we will explore the attitude control problem as a decision problem and how the Deep-Q Learning algorithm can be used to train a controller.