Triet Tran, Cornerstone Consulting & Services, LLC; Anthony N. Dills, L3Harris Technologies, Inc.; James Crowley, L3Harris Technologies, Inc.
Keywords: Machine Learning, Data Association, Maneuver Classification
Abstract:
The ability to perform near real-time data association and automatic detection and classification of Resident Space Object (RSO) maneuvers is highly desirable. The problem of mis-tagging and Uncorrelated Tracks (UCTs) is still a challenge in processing observations today. This problem is in part due to unknown maneuvers that occur between successive revisits of the tracking source on the RSOs. Advanced techniques in statistical filtering have shown various degrees of success. Artificial Intelligence/Machine Learning (AI/ML) techniques have seen significant growth in recent years and pose viable approaches within the space domains solution space to address this challenge.
Feasibility of using AI/ML to perform Data Association (DA) and Maneuver Detection/Classification (MD/MC) of the Galaxy-15 in the Geosynchronous (GEO) orbit regime has been demonstrated and reported previously using publicly available Wide Area Augmentation System (WAAS) data. The various ML techniques such as Tree-Based Pipeline Optimization Tool (TPOT) and Autoencoder (AE) were shown to achieve better than 90% accuracy in both data association and maneuver classification [1].
The current work demonstrates improved performance over our previous reported techniques by using the Ensemble ML technique. This technique combines the various ML models into a single model in a way that improves the overall performance over individual models. We show that this technique is applicable to both DA and MD/MC using publicly available data on the Galaxy-15 satellite. Characterizing the Pattern of Life (POL) of the maneuvering RSO is the key to the success of the AI/ML model to recognize whether a set of optical astrometric measurements can be associated with an RSO candidate, and whether a maneuver type can be classified for that RSO based on the same optical measurement set. Using open-source template ML models from Keras and Scikit-Learn libraries, we compare the performance of individual models to that of the ensemble model using feature input observational data derived from WAAS data from Galaxy-15 as well as data from a fictitious satellite artificially placed close to the Galaxy-15 location. Our work also addressed the training issue of data staleness that points to the practice of model retraining using the streaming data technique. Preliminary results using ensemble ML technique indicate that for both data association and maneuver classification, the accuracy improved from 88% and 85%, respectively in previous work to at least 94% overall.
In summary, with the encouraging results from the current work, we surmise that there is real potential for improved information gain by combining statistical filtering techniques and ML ensemble techniques that will be of benefit to the space domain community. In our next step, we will consider the enabling capability of building training data from historical TLEs for any catalogued RSOs. This activity is crucial for the success of applying ML techniques to address current challenges, since historical TLEs are abundantly available for most RSOs but precision ephemerides are rare for most RSOs of interest.
Date of Conference: September 27-20, 2022
Track: Machine Learning for SSA Applications