Alexander Cabello, EO Solutions Corp; Jeff Houchard, EO Solutions Corp; Cameron Harris, EO Solutions Corp; Zach Gazak, USSF SSC/SZG; Jonathan Kadan, USSF SSC/SZG; Justin Fletcher, USSF SSC/SZG
Keywords: Machine Learning, Initial Orbit Determination, Simulation
Abstract:
Accurate Initial Orbit Determination (IOD) is a cornerstone of Space Domain Awareness (SDA), serving as the foundation for detecting, tracking, and predicting the movements of space objects. In the context of SDA, one of the critical missions, especially for autonomous sensors, is detecting and tracking uncorrelated targets (UCTs)—objects that appear unexpectedly and are not immediately identifiable. When a UCT is detected, it is essential to establish a reliable orbit determination to facilitate follow-up observations. A well-executed IOD improves the chances of reacquiring these targets and allows for greater flexibility in sensor tasking, as the prediction accuracy provides a larger time window for future observations.
This research introduces a machine learning-based IOD method to enhance the accuracy of orbit predictions derived from short-arc angle-only measurements obtained via electro-optical (EO) sensors. The primary objective of this work is to develop and rigorously evaluate a machine learning framework tailored for IOD, using a dataset of simulated EO observations. These simulations encompass a wide range of orbital regimes, including Low-Earth Orbit (LEO), Medium-Earth Orbit (MEO), Geostationary Orbit (GEO), and Highly Elliptical Orbit (HEO). By comparing the performance of this machine learning approach with traditional IOD methods such as Laplace, Gauss, and Gooding, this research seeks to demonstrate the efficacy and applicability of the proposed method across diverse orbital environments.
The innovative aspect of this research is the integration of machine learning with the principles of orbital mechanics, forming an end-to-end differentiable framework. This approach allows the model to learn directly from simulated orbital trajectories and their corresponding initial observations, leveraging data-driven insights and physics-based constraints. The machine learning model is trained on a diverse dataset of 15 million observations, which includes various orbital regimes, to ensure its generalizability and robustness. Specifically, the model is structured around a neural network architecture that includes two Long Short-Term Memory (LSTM) layers followed by a dense layer. This configuration was selected after wide experimentation with different architectures and output modes.
One of the key challenges in IOD is dealing with the inherent imperfections in observational data. While effective under ideal conditions, traditional methods are often constrained by the quality of the input data. They typically assume that the observation angles are precise, and thus, they lack mechanisms to account for noise and other inaccuracies in the measurements. This limitation can lead to significant errors in orbit determination, particularly when dealing with short-arc observations. In contrast, the learned methods developed in this research are shown to be more resilient to such noisy data, predicting position and velocity vectors more reliably than traditional methods.
In conclusion, this research demonstrates the feasibility and advantages of applying neural network-based approaches to Initial Orbit Determination (IOD). By overcoming the limitations of traditional methods, particularly in handling noisy and imperfect observational data, the proposed machine learning framework offers a promising path forward for more accurate and reliable orbit determinations. As space traffic increases and the need for effective SDA grows, such advancements will be critical in ensuring the safety and sustainability of space operations.
Date of Conference: September 17-20, 2024
Track: Astrodynamics