Peng Mun Siew, Massachusetts Institute of Technology; Haley Solera, Massachusetts Institute of Technology; Thomas G. Roberts, Massachusetts Institute of Technology; Daniel Jang, Massachusetts Institute of Technology; Victor Rodriguez-Fernandez, Universidad Politécnica de Madrid; Jonathan How, Massachusetts Institute of Technology; Richard Linares, Massachusetts Institute of Technology
Keywords: Behavioral Mode Identification, Pattern-of-Life Characterization, SSA Dataset, Astrometric Data, Machine Learning
Abstract:
Despite the attention garnered by applying innovative artificial intelligence (AI) techniques to challenging problems in aeronautics and astronautics in recent years, there is still a lack of adoption among the space situational awareness (SSA) community. Although historical data on Earth-orbiting satellites—troves of orbital state and light curve measurements derived from passive observations of space objects from the ground- and space-based sensors—may be particularly well-suited for AI-driven analysis, the disconnect between the AI and SSA research communities have prevented robust, interdisciplinary research progress. The AI research community lacks domain-specific technical knowledge and suffers from limited SSA data availability. The challenge problem introduced in this paper aims to leverage expertise from the AI research community to use multi-faceted SSA data for characterizing a satellite pattern-of-life (PoL) in new, innovative ways. To do so, the author team is providing an open-source SSA dataset and a methodology for evaluating submissions from the challenges’ entrants.
Over the course of geosynchronous (GEO) satellites’ operational lifetimes, operators issue commands to place them in various behavioral modes, ranging from station-keeping, to longitudinal shifts, to end-of-life behaviors, and perhaps many others. Satellite PoLs are descriptions of on-orbit behavior made up of sequences of both natural and non-natural behavior modes.
For the majority of GEO satellites, station-keeping is the most common class of behavioral modes. Satellites perform station-keeping maneuvers to counteract external perturbations at play in that regime, including third body effects of the sun and the moon, the variation in the Earth’s gravitational field due to the non-spherical and in-homogeneous mass distribution of Earth, and solar radiation pressure. Station-keeping maneuvers can be further classified into north-south and east-west station-keeping, referring to corrections to displacements in geographic latitude and longitude, respectively. One common way for a GEO RSO to shift from one behavioral mode to another is to perform a longitudinal-shift maneuver, which changes the satellite’s mean longitude in a controlled manner. This type of maneuver is typically driven by changes in the satellite’s mission objectives. Another time when satellites typically change their behavioral mode is at their end of life, when they may perform de-orbiting maneuvers, where the GEO satellite is maneuvered to a higher altitude, outside of the geostationary belt.
Identifying these behavioral modes—including those not associated with well-understood operational patterns—would assist the SSA research community in better understanding the behavior of satellites and thus generate a more accurate predicted trajectory and improve tracking capabilities. Furthermore, characterizing the PoLs for a diverse range of GEO satellites can help to contextualize historic on-orbit behaviors and behavior patterns, cultivate generalized maneuver prediction on a large scale, enable early identification of future satellite behaviors, enable inference of metadata from new and historic behaviors, as well as enable better mission planning.
Artificial intelligence (AI) approaches have been shown to be able to learn the relationship between time-series data and have previously been successfully applied to complex multi-dimensional problems such as spacecraft anomaly detection, orbit estimation, or weather prediction, among others. One of the main challenges in developing AI approaches for this application is the difficulties in acquiring realistic state vector data that are of sufficient quality and temporal resolution. Furthermore, there is no common dataset making it impossible to evaluate and compare the performance of different AI algorithms. Over the last decade, several space-related challenge competitions have been held with success by the Advanced Concepts Team of the European Space Agency. Through these challenges, they were able to motivate the adoption and development of AI approaches to the space community.
The Satellite Pattern-of-Life Identification Dataset (SPLID), created by the author team to support this challenge problem, consists of synthetic astrometric data generated using a high-fidelity simulator to simulate a range of operation scenarios. In addition to this public set of synthetic astrometric data, the challenge problem will also utilize a private set of real astrometric data for algorithm evaluation. The real astrometric data includes high-accuracy ephemeris data provided by satellite owners and operators as well as Vector Covariance Message (VCM) data that was provided by the United States Joint Space Operations Center (JSpOC).
A development kit coded in both Matlab and Python consisting of basic utility functions for data parsing, manipulation, and visualization has been made available alongside the SPLID. Baseline machine learning solutions will also be provided. The baseline solutions will be coded in Python using readily available packages and aim to lower the barrier of entry to AI techniques in the SSA research community.
Date of Conference: September 19-22, 2023
Track: Machine Learning for SDA Applications