Astrometric and Photometric Data Fusion in Machine Learning-Based Characterization of Resident Space Objects

Abstract

In this dissertation, a topic fundamental to Space Situational Awareness (SSA) is considered: resident space object (RSO) physical and dynamical property estimation (characterization). This topic is explored by leveraging both simulated and observational data of RSOs in conjunction with physics- and machine learning-based (ML) models to illustrate novel characterization methods for SSA. This is accomplished in three main parts: simulated and real-world case studies of object-specific dynamical and physical property inversion through Bayes’ theorem; initial results from a novel sensor system designed to autonomously collect data en masse on near-geosynchronous RSOs; and a deep machine learning approach to orbital regime identification of cislunar RSOs from telescope observational data. These topics represent significant advancement in the field of SSA in three different areas associated with the RSO characterization problem, where each provide many opportunities for future research and further advancement of the field. The first work put forward is a new method of RSO property estimation using Bayes’ theorem through the process of Markov Chain Monte Carlo (MCMC) sampling. This process will yield the most likely distributions for a desired set of space-object properties conditioned by observational data of the target. The feasibility of this method, as well as the performance for inversion of various physical and dynamical properties of an RSO is explored through both real-world case studies and simulation. The results show this to be a powerful method of property estimation reliant on sufficient measurements and a robust forward model. The first case study applies the technique to the near-Earth object (NEO) 2020 SO, detected by planetary defense survey telescopes. This object, which turned out to be the Centaur upper stage of the Surveyor 2 mission launched in 1966, was temporarily recaptured by the Earth out of heliocentric orbit. After two close fly-bys of the Earth, it was kicked back into heliocentric orbit by an encounter with the Moon. With a closest approach just beyond the geosynchronous distance, and a close fly-by of the Moon, this object became of particular interest for characterization. Through telescope observations taken from around the world of 2020 SO, coupled with a simulation validation, this first case study establishes this Bayesian light curve inversion as an important characterization technique. With the lessons learned from 2020 SO in-hand, the second case study was of another object discovered by planetary defense surveys, WE0913A, which impacted the Moon on March 4, 2022. This object, discovered in cislunar space, is shown conclusively through the characterization efforts in this work to be the upper stage of the Long March 3C rocket that launched the Chinese Chang’e 5-T1 lunar fly-by mission. The second work discussed is the design and implementation of the Stingray sensor system, which is comprised of 15 CMOS detectors fixed-staring at the Geosynchronous Earth Orbit (GEO) belt visible above Tucson, Arizona. The design and function of the system is described in detail, along with verification of initial survey results. Every clear night, the Stingray system is capable of producing near-real time astrometry and photometry of all RSOs it can see in near-geosynchronous orbit simultaneously. Creating a huge wealth of data, this system enables training of ML models for SSA with real measurements. A feat which historically, has been very difficult to accomplish due to the volume of data required for traditional ML techniques and the comparatively low throughput of most telescope systems. Stingray also provides an avenue for validation of physics-based models simulating the translational, rotational, or reflective dynamics of spacecraft and debris. The access to near-real time data on so many RSOs at once allows for analysis and alerting of behavior in an operational cadence that was not previously accessible to researchers. The data produced by the Stingray system also provides the potential for a myriad of other future research topics in SSA that are not explored here, and the system will continue to drive research for years to come. The third portion of this work uses deep recurrent neural networks (RNNs), where the feasibility of recovering orbital regime from astrometric data of RSOs is explored for 32 different orbital families that arise naturally in the circular restricted three body problem (CR3BP) framework of the Earth-Moon system. Additionally, stable and unstable invariant manifolds of the three colinear Lagrange point planar Lyapunov orbit families and the asso- ciated northern and southern halo orbit families are considered to gauge performance of this method in differentiating between manifolds and periodic orbits. In doing so, the potential of this method to differentiate objects drifting/repositioning from those on periodic orbits is shown, a capability critical to the future of cislunar space situational awareness.