Giovanni Lavezzi, Massachusetts Institute of Technology; Peng Mun Siew, Massachusetts Institute of Technology; Di Wu, Massachusetts Institute of Technology; Zachary Folcik, MIT Lincoln Laboratory; Victor Rodriguez-Fernandez, Universidad Politécnica de Madrid; Jeffrey Price, United States Air Force; Richard Linares, Massachusetts Institute of Technology
Keywords: Machine Learning, Early Time Series Classification, SSA
Abstract:
Space Situational Awareness (SSA) has gained prominence owing to its criticality in national defense and growing number of Resident Space Objects (RSOs) due to commercialization of space. The identification and classification of RSOs is a desirable objective for SSA. AI techniques, specifically Machine Learning (ML) and Deep Learning (DL) algorithms, now facilitate classification of space objects based on sensor observations. However, these AI algorithms are often constrained by the lack of a sufficiently large training dataset.
This work aims to use supervised ML algorithms for early time series classification using Vector Covariance Message (VCM) data covering 22,303 RSOs over six months. ML techniques such as Random Forest, Logistic Regression, K-Nearest Neighbor, Support Vector Machine, Naive Bayes, along with a Multi-Layer Perceptron DL technique are initially tested. Subsequently, advanced time series classification algorithms like InceptionTime and TSiT, for regularly and irregularly sampled time series, respectively, are examined. Given the variability in RSO’s VCM estimates and the problem of data imbalance between classes, time regularization and synthetic data are introduced. However, the use of synthetic data creates problems in the classification accuracy of the ML algorithms. The irregular time series classification technique can overcome this issue.
Considering the novelty of using the VCMs as datasets for the ML algorithms, the study explores input combinations and variations of the datasets. Early classification of space objects is of key interest to space system operators, therefore the acceptable trade-off between accuracy and timeliness of the prediction is investigated.
Date of Conference: September 17-20, 2024
Track: Astrodynamics