Ian McQuaid, Air Force Institute of Technology; Laurence D. Merkle, Air Force Institute of Technology; Brett Borghetti, Air Force Institute of Technology; Richard Cobb, Air Force Institute of Technology; Justin Fletcher, Air Force Research Laboratory/Directed Energy Directorate
Keywords: Space situational awareness, artificial neural networks, deep neural networks, machine learning, GEO track association, non-resolved optical observations
Abstract:
Ground-based non-resolved optical observations of resident space objects (RSOs) in geosynchronous orbit (GEO) represent the majority of the space surveillance networks (SSNs) deep-space tracking. Reliable and accurate tracking necessitates angular separation of the observations, and thus separation in time as well. This requires that subsequent observations be associated with prior observations of a given RSO before they can be used to create or refine an ephemeris. When using astrometric data (e.g. angles) alone this association task is complicated by RSO maneuvers between observations, or when RSOs are found in close proximity to one another. Accurately associating an observation with an RSO thus motivates the use of photometric space object identification (SOI) data along with astrometric estimates. Contemporary machine learning, specifically deep neural networks (DNNs), offers mechanisms that may be used to perform this association autonomously by first learning patterns between observations and their parent objects. This research assesses the extent to which a trained DNN can autonomously associate previously unseen observations with the RSO they represent. The DNN architecture used in this work is comprised of a recurrent neural network (RNN) astrometric component and a convolution neural network (CNN) photometric component. The RNN component infers kinematic information, such as angle rates and potential orbit geometries from sequences of angle observations. The CNN component extracts features that distinguish RSOs, such as average brightness and periodic artifacts, from photometric light curves. The interred kinematic information and photometric feature maps are then used as inputs to a classification layer which jointly maps those inputs to a probability mass function over RSO labels. The performance of the network is evaluated via simulation of two scenarios. In the first, a geostationary catalog maintains station throughout one year. In the second, a fraction of the catalog conducts periodic rendezvous and proximity operations (RPO) throughout the GEO belt over the course of the year. In both cases the astrometric and photometric observations are generated as if observed by a GEODSS telescope at one of the three GEODSS sites (Maui, Diego Garcia, and Socorro NM). For each scenario, the network is trained using the first several months of pre-labeled data and then evaluated based on its effectiveness in labeling the remainder of the observations. Specifically, the percentage of observations associated to the correct RSO serves as the primary measure of network effectiveness. This research contributes to the development of autonomous SSN telescopes, and systems to autonomously and rapidly update an RSOs ephemeris after it maneuvers.
Date of Conference: September 11-14, 2018
Track: Astrodynamics