Artificial Intelligence-Based Risk Assessment and Maneuver Decision-Making for Satellite Conjunctions

Abstract

This thesis investigates the potential application of Machine Learning (ML) models for risk assessment, aiming to develop a prototype tool capable of supporting future Collision Avoidance (COLA) operations. The rapid growth of the space sector has resulted in a progressively more congested orbital environment, substantially increasing the workload of operators responsible for analysing Conjunction Data Messages (CDMs) and deciding on Collision Avoidance Maneuvers (CAMs). To address this challenge, key state-of-the-art ML applications for COLA were examined to assess their potential in supporting analysts and to provide a comprehensive understanding of existing approaches. Building on this knowledge, the primary objective of this work is to develop an ML model capable of analysing patterns in time series of CDMs and determining whether a CAM is necessary, specifically aiming at automating the maneuver decision-making process. This application is naturally formulated as a binary classification problem. Prior to achieving this objective, an in-depth investigation of collision avoidance processes and the fundamentals of ML was carried out, as building a solid background in these concepts is essential for the accomplishment of this work. The preliminary study confirmed the need to train the ML model on a synthetic dataset of CDMs, where the criticality of each encounter is known. COLGen, a Python-based tool, was consequently developed to generate synthetic encounters according to a well-defined underlying rationale. The encounter generation begins with a synthetic scenario in which the encounter geometry is known exactly, with no uncertainty. The scenario is then perturbed by introducing an error, referred to as the CDM error. Five neural networks based on a Multi-Layer Perceptron (MLP) architecture were trained and tested on five synthetic datasets, each characterised by a distinct CDM error, and their classification performance was subsequently compared. The promising results demonstrate the potential of this approach to support maneuver decision-making processes, thereby enhancing the safety and reliability of future COLA operations.