Anne Adriano, University of Waterloo; K. Andrea Scott, University of Waterloo; Nasser Lashgarian Azad, University of Waterloo
Keywords: Classification, space debris, light curve, Blender, Orekit, machine learning, deep learning, Wavelet Scattering Transform, attitude, shape
Abstract:
The field of Space Domain Awareness (SDA) is becoming increasingly important, particularly amidst the growing population of space debris in low Earth orbit (LEO). Previous research has explored the classification of various types of attitudes exhibited by both defunct and operational satellites and rocket bodies, yet there is a noticeable gap in the literature concerning the analysis of attitude complexities exhibited by smaller orbital debris, such as fragments and detached satellite components, specifically. This lack of focus on debris sub-classification could be attributed to the scarcity of labeled debris data.
To address this challenge, a light curve simulation framework was created, which combined the capabilities of both Orekit for numerical orbit propagation, and Blender for physical environment modelling and image rendering. Using this simulator, light curves of different debris shapes were generated based on initial conditions defined by LEO Two-Line Elements (TLEs), and a series of physical object parameters. While there is an infinite number of shapes that orbital debris can assume following a breakup event, this study focused on photometric signatures produced by five shape classes in particular: rods, panels, cuboids, dishes, and cones.
The first task aimed to classify the light curve attitude complexity as either tumbling or stable, where tumbling objects are characterized by a rotation about two or more body axes, and stable objects are either nonrotational or rotating about a single axis. Data pre-processing for this step involved the extraction of frequency information using the Wavelet Scattering Transform (WST). Running an Extreme Gradient Boosting (XGBoost) algorithm on the extracted features for this binary classification task produced an accuracy of 90%. The second task involved the shape classification of tumbling debris using a set of deep learning models, namely a Convolutional Neural Network (CNN), a Long Short-Term Memory network (LSTM), and a fully convolutional LSTM (LSTM-FCN). The highest performer among these networks was the LSTM-FCN, achieving a test accuracy of 95%.
This study employed novel approaches to the attitude and shape classification of debris light curves, producing promising results when using a unique dataset of LEO debris. Extensions of this work will involve the implementation of attention mechanisms to further improve the performance of the deep learning models, as well as the application of more sophisticated light curve inversion techniques to extract precise attitude rates from the light curves.
Date of Conference: September 17-20, 2024
Track: Space Debris