Enhancing Collision Risk Assessment with Deep Learning Models

Abstract

Recent years have seen the exponential growth of the number of artificial objects orbiting the Earth. Since space debris can cause substantial damage, measures are investigated to make space operations more sustainable. Amongst these, there is an interest in the development of methods to detect possible collisions between objects in space. Traditional methods require highly accurate propagations with large computational times, so the focus of this thesis is the modelling of faster and more accurate algorithms to enhance collision risk assessment. To avoid long propagations, neural networks have been trained for orbit prediction and uncertainty estimation, with the main goal of calculating the collision probability between two objects. Different analyses have been made to assess the accuracy, stability, applicability and generalisation of the methods developed. The results show much faster collision risk calculations than traditional methods with a similar level of accuracy. It is also demonstrated how a tool which can generalise to satellites with different geometries can be built.