Alexandra Wright, Massachusetts Institute of Technology, Lincoln Laboratory; Peter W. Boettcher, Massachusetts Institute of Technology, Lincoln Laboratory; Anye Li, Massachusetts Institute of Technology; Erin L. Main, Massachusetts Institute of Technology
Keywords: maneuvers, data fusion, orbital mechanics, machine learning, predictive analytics, space catalog, electric propulsion,
Abstract:
Modern challenges and objectives for space domain awareness look different than those decades ago when people first began launching satellites. However, we still largely rely on data models developed for reactive catalog maintenance, whose purpose is to provide a latest orbital update on each satellite. In this paper, we present modernized space data models that redefine time concepts and data representations to enable pro-active and machine-assisted decision making. A flat list of catalog updates is inadequate for this purpose, as it does not represent a history of each satellites behavior over time, instead offering a history of sensor collections. Furthermore, it does not provide the time constructs necessary for representing multiple simultaneous current or future hypotheses, which becomes critically important when assessing or anticipating satellite actions that manifest as a series of non-deterministic orbit maneuvers. Inability to represent such realistic satellite behaviors with clear mathematical constructs is a barrier to machine automation of assessment, detection, and prediction of orbital actions.?
In our model, we abandon a flat, sensor-update-focused catalog structure in favor of a state tree that constructs genealogy of a satellite’s related, consecutive orbital states and a method for alternately pruning and promoting branches of a satellites state history that incoming data either refutes or supports, respectively. Branches in the state tree connect via orbital maneuvers, which when confirmed, are promoted to the main branch of the states history. This same structure also supports owner/operator plans or hypothesized future actions, as they are also naturally expressed as a series of connected orbital states. By accommodating parallel genealogies from alternate sources, this method of describing orbits disambiguates valid start and stop times of orbital estimates, presents a representative history of a satellites past orbits and actions, allows for analysis of multiple possible future scenarios, and enables machine automation of both labeling and generating orbit behaviors and detecting when data supports specific hypotheses generated.??
We built and refined our data model by standing up multi-hypothesis tracking techniques over many years of measurement data on spacecraft and constructing tools for space operators that are fed by the results stored in our data model. This has allowed us to test, break, and improve our data model while creating functionality that demonstrates the instances where our constructs provide key scaffolding for richer data analytics. In addition to the novel predictive automation alluded to, we have found that a great strength of the flexible data model includes the incorporation of non-measured data beyond the contextual labeling of genealogical state branches. Adding other contextual metadata on maneuvers and spacecraft themselves further assists in constructing new methods for previously challenging orbital analysis tasks, such as synthesizing historically observed satellite actions into inferred intent.??
The final application in this paper is a consideration of the architectural impacts of the model if it were to scale to the broader space domain awareness community of military, commercial, and international partners. We propose an implementation of our data construct that supports decentralized catalogs of orbital state data, which addresses the fundamental challenge of supporting a diverse and distributed user base. We frame all of this with how the model we have built addresses core needs of this works sponsor, the United States Space Force, as well as the broader community of participants in the space domain.?
Date of Conference: September 27-20, 2022
Track: Machine Learning for SSA Applications