Predicting Satellite Collision Risk Using Artificial Intelligence

Abstract

This thesis investigates the use of machine learning for satellite conjunction risk assessment using the ESA Collision Avoidance Challenge dataset. It proposes a hybrid model that combines an XGBoost regressor for miss-distance prediction with a Neural Network classifier for critical-event detection, and integrates them through a soft-vote ensemble. The study addresses practical issues such as feature engineering, preprocessing consistency, leakage control, and threshold selection for safety-oriented triage. The system is evaluated on both a development split and an independent holdout set to assess generalization under data shift. The results show that the proposed approach can support first-level conjunction screening as a decision-support tool in space situational awareness.