Thesis title: Satellite On-Board Solutions for Precise Orbit Determination on Earth and Moon Orbit
Precise Orbit Determination, which is the problem of finding the satellite ephemeris by estimating the satellite position and velocity based on Earth ob- servations data, has always been one of the most important aspects of satellite navigation. Satellite positioning is used daily by smartphones to provide several features and also by military personnel. Both of these users require different levels of accuracy. In the last decades the increasing interest in space exploration has brought forward the necessity to provide the satellites with on–board estimation algorithms. The ability to self–estimate their position and velocity is of utmost importance in scenarios in which data from Earth are not available, like in Moon or Mars orbiting. Artificial Intelligence has proven to be one of the best solutions, able to provide the satellite with non–standard measurements. The abstraction and generalization capability of neural networks allow to perform complex tasks while satisfying real–time constraints. In this context Crater Matching is one of the most promising solutions for orbit determination.
In this thesis two different approaches for on–board Precise Orbit Determination will be proposed: one making use of standard GNSS measurements coming from Earth and the other one making use of non–standard ones provided by neural networks. In the former case an end–to–end analysis, going from satellite propagation to satellite visibility has been performed. In the latter case a first step toward the development of a full Terrain Relative Navigation system has been carried out: a benchmarking of different neural networks architectures has been performed, by using a space–qualified processor, in order to identify the best on–board solution to deal with the crater detection problem.
Two additional research projects tackling important aspects of the space domain, satellite communication and Earth observation respectively, will also be detailed. A mixed Artificial Intelligence and Reinforcement Learning solution has been proposed for the first topic with extensive simulations and comparisons to validate the approach. A full Artificial Intelligence approach has instead been taken to tackle the second project, in which a Convolutional Neural Network has been trained to detect wildfires in real–time and has been tested over a space–qualified processor to verify its feasibility to be deployed on–board.