Thomas G. Roberts, Massachusetts Institute of Technology; Victor Rodriguez-Fernandez, Universidad Politécnica de Madrid; Peng Mun Siew, Massachusetts Institute of Technology; Haley Solera, Massachusetts Institute of Technology; Richard Linares, Massachusetts Institute of Technology
Keywords: geosynchronous, station-keeping, pattern of life, maneuver detection, satellite behavior, unsupervised machine learning
Abstract:
Satellites in the geosynchronous orbital regime (GEO) typically spend the majority of their operational lifetimes performing station-keeping maneuvers to maintain a near-fixed position in the Earth-centered, Earth-fixed inertial reference frame. Although almost all satellite operators command their GEO satellites to perform some kind of station-keeping maneuvers, there is a remarkable diversity of station-keeping protocols across the regime’s population. Some satellites perform infrequent, but high-magnitude maneuvers using chemical thrusters, while others exercise more frequent, long-duration burns using electric propulsion to accomplish the same task. Despite decades of heritage operations in the GEO environment, there is no comprehensive rulebook for GEO satellite behavior. Due to the availability of GEO satellite data and the periodic nature of both perturbative and artificial forces in the geosynchronous space environment, machine learning techniques are particularly well suited to identify and describe patterns of behavior amongst the GEO satellite population.
This work uses the GEO satellite patterns of life (PoL) formulation [1]—where satellite behavior can be described as periods of behavioral modes, separated by nodes during which behavior switches from one mode to another—to identify and characterize GEO satellite behavioral modes in a single, end-to-end algorithm via unsupervised machine learning. Using historical orbital elements (OEs) and geographic longitude, latitude, and altitude measurements (LLAs) derived from publicly available two-line elements, satellite behavior between PoL nodes can be sorted into groups by a k-means clustering algorithm adapted for time-series data using dynamic time warping (DTW) [2]. In this work, unlabeled historical OEs and LLAs can be used to cluster GEO satellites into particular behavioral modes—often, but not necessarily associated with well-known operations, such as east-west station-keeping, libration orbits, or retirement drift in the graveyard orbit—at each time step during a study period [3]. When more clusters are identified by the algorithm, more behavioral modes beyond those associated with well-known operations can be identified. For this paper, a study period of January 1, 2010, to December 31, 2022, was selected, with a two-hour time step: meaning OEs and LLAs are computed from TLEs every two hours, when available.
The results of the algorithm described and evaluated in this work can be represented in two ways: a ledger and a map. The ledger lists the Satellite ID numbers included in each of the identified clusters at each time step. The map offers a visualization of the cluster populations in two-dimensional space over time, where objects that appear closer in the visualization are also closer together in the DTW-informed distance metric embedded in the clustering algorithm.
Both sets of results can be used to describe a GEO satellite’s PoL as a history of cluster assignments. A satellite that is first inserted into the GEO belt, then maintains a single longitudinal position with high-frequency east-west and north-south station-keeping, and finally retires to the graveyard orbit, all within the study period, would likely be grouped into a only a small number of clusters: those associated with its three behavioral modes and the two nodes during which it was transitioning.
In addition to contributing to the burgeoning literature on satellite pattern-of-life characterization, this work can also be used to describe historical GEO satellite station-keeping behavior at low resolution. That is, a satellite’s historical behavior could be described by behavioral mode classes and short lists of relevant parameters, as opposed to the high-resolution OE and LLA data on which the algorithm was trained.
Date of Conference: September 19-22, 2023
Track: Machine Learning for SDA Applications