John Ebeling, Data Fusion & Neural Networks, LLC; Duane DeSieno, Data Fusion & Neural Networks, LLC; Jacob Hansen, Data Fusion & Neural Networks, LLC; Christopher Tschan, Data Fusion & Neural Networks, LLC; Carolyn Sheaff, Air Force Research Laboratory/RIED
Keywords: Electro-Optical (E-O), space surveillance observations, space surveillance imagery, space domain awareness, machine learning
Abstract:
As space grows ever more congested, achieving and maintaining Space Domain Awareness has been the focus of many research efforts. Electro-Optical (E-O) space surveillance observations are growing in importance and quantity. This paper presents a prototype implementation of a modular pipeline that characterizes space situations by leveraging both electro-optical data and imagery. The pipeline is designed to be deployed at the space telescope sensor location in order to immediately process and autonomously alert operators to data and images that contain interesting events, such as multiple satellites in proximity. Ultimately, this algorithm would also provide automated near-real time imagery annotation significantly reducing time to respond to emerging space events. This allows individual E-O sensors to immediately contribute to Space Domain Awareness and opens new opportunities for higher level fusion to be performed on centralized data processing platforms.
This work culminates two Air Force Research Laboratory (AFRL) projects, the Rapid Discovery of Evasive Satellite Behavior (RDESB) and Electro-optical Pre-custody Threat Warning (EPTW). The current iteration of the prototype processes a single E-O image in under 2.5 seconds. The pipeline prototype was first trained and tested using simulated E-O data from the SJ-21 (satellite catalog object 49330) and Compass G2 (satellite catalog object 34779) close approach and grappling scenario in January of 2022. As the objects’ visual separation decreased the pipeline correctly identified and annotated a visual proximity event. A Focus of Attention algorithm was used to identify real data of interest from the Unified Data Library for additional testing. Using this data, the pipeline prototype produced similar promising results.
The prototype pipeline utilizes multiple sets of Neural Networks (NNs) with Machine Learning and Data Fusion algorithms to extract, detect, and categorize scenarios. Convolutional, traditional feed forward, and autoencoder NNs are all employed for object extraction, light curve analysis, and anomaly detection respectively. We investigate Gaussian Mixture Model fitting as a fast solution to source extraction. The Hungarian algorithm is used to solve the assignment problem. A heuristic-based clustering approach is used to classify time series of error scores. This paper also discusses plans to replace the heuristic-based clustering with NN-based clustering as more annotated event timelines become available. Imagery and derived data analysis subprocesses run in parallel pipelines with their outputs being fused to produce an event timeline that can be resampled to an operator’s desired scale. Recommendations for future optimization and research are also presented.
Date of Conference: September 17-20, 2024
Track: Machine Learning for SDA Applications