Louis Penafiel, Aptima, Inc.; William Dupree, Aptima, Inc.; Thomas Gemmer, Aptima, Inc.
Keywords: Machine Learning, clustering algorithms, Unified Data Library, Uncorrelated Tracks, Unsupervised Machine Learning
Abstract:
Every year the boundaries of space exploration are pushed to new heights. From 2019 to 2028, about 10,000 satellites are set to be launched into space. These along with other rocket launches will vastly increase the number of objects (including active satellites and debris) in space to track. This makes the demand for more capable and more automatic tools and analytics to analyze these objects a primary concern. A specific area of focus is Space Situational Awareness (SSA). It is necessary to keep track of what satellites are doing to know if a satellite is in danger of collision or being threatened in some other way. A major task in that area is associating uncorrelated tracks (UCTs) to known resident space objects (RSOs). UCTs mainly arise when an object is poorly tracked (making it difficult for sensors to pick up) or an object has maneuvered(making its trajectory significantly different from the most recent orbit solution). This task has mainly been performed manually by analysts, making it time and resource intensive.
In this study, we focus on automating the UCT correlation procedure through analyzing historical data to characterize the relationships between UCTs and cataloged objects, such as origin (e.g. debris or loss of custody) or threat (UCT will threaten a cataloged object). We apply this study in the context of increasing the effectiveness of space operators by creating a microservice called WhoAmI. WhoAmI performs the UCT association, then produces human-understandable alerts containing a rank-ordered list of the most probable RSOs that the UCT might be associated to. Furthermore, this tool can help discover new objects that have not been previously tracked and entered in the public catalog. Lastly, we use the data from the Unified Data Library (UDL), where we query their OpenAPI for observational data and receive data of various types, ranging from raw observations, to two-line element sets. These capabilities of WhoAmI enable operators to perform UCT resolution tasks more effectively by automating the process of filtering the number of RSOs to consider, more confidently by prioritizing UCTs with the highest threat assessment, and more adaptively by reducing reliance on a single data type.
WhoAmI performs associations of UCTs to cataloged RSOs using a joint approach of machine learning combined with astrodynamic analysis. We use a two-step, unsupervised machine learning model to filter the Space Catalog in order to find objects with similar orbital elements to the UCT. The first step uses density based clustering of the catalog in orbital element space to find cataloged objects in the astrodynamic vicinity of a UCT. The second step uses propagation models to perform a more traditional pass of filtering by directly analyzing the orbital paths of the filtered candidates. For the latter, we performed various orbit determination techniques based on the type of data received from UDL. In addition to directly associating UCTs to a catalog, we use a similar approach to matching the UCTs to their most likely launches.. This allows WhoAmI to automatically find origin information for debris objects so that they can be entered into the catalog as well as making it potentially useful in breakup analysis. The results of this process are then passed via alerts for space operators with visualizations of probability of association to act on. We validate our approach on metrics of accuracy (using historical tracking data and maneuvers as baselines) and processing time.
Date of Conference: September 14-17, 2021
Track: Machine Learning for SSA Applications