Emma Stevenson, Universidad Politécnica de Madrid; Victor Rodriguez-Fernandez, Universidad Politécnica de Madrid; Hodei Urrutxua, Universidad Rey Juan Carlos
Keywords: Space Debris, Conjunction Assessment, Machine Learning, Graph Neural Networks
Abstract:
The prevention of in-orbit collisions is crucial both in the near-term, to protect current-day space assets, and for long-term space sustainability. To this end, identifying close approaches between all catalogued objects, whether active or debris, is of vital importance for early detection of potentially catastrophic events. However, this problem of all-vs-all conjunction assessment is computationally challenging, with hundreds of millions of possible conjunction pairs already present today, and threatens to become ever-more so in the face of increasing space traffic and observational capabilities in the New Space era. With a pressing need to look for new solutions to this problem, one emerging approach is the adoption of recent advancements in the field of machine learning.
In the all-vs-all case, interactions (links, or edges) between pairs of individual objects (nodes) over a whole catalogue can be naturally described using a graph. Graph-structured data such as this is prevalent in a wide variety of different domains, from social networks, to transport networks, to molecular graphs. Following the successes of Graph Neural Networks (GNNs) in these areas, the application of machine learning to graph-structured data has become one of the fastest growing research areas in machine learning. Unlike for grid-like data, such as sequences or images, which have a standard data representation, one of the main challenges in this domain is to build a representation of the graph that can be successfully exploited for tasks such as link prediction (e.g., recommending new social network connections, or predicting conjunctions between two space objects). In this work, we present a graph-based, global representation of the all-vs-all scenario that is able to profit from recent advancements in Graph Neural Networks, and make a step towards efficient, machine learning based conjunction assessment.
Date of Conference: September 27-20, 2022
Track: Machine Learning for SSA Applications