Machine Learning Methods for Hypervelocity Fragment Flyout Characterization

Author

Katharine Larsen, Embry-Riddle Aeronautical UniversityFollow

Date of Award

Spring 2025

Access Type

Dissertation - Open Access

Degree Name

Doctor of Philosophy in Aerospace Engineering

Department

Aerospace Engineering

Committee Chair

Riccardo Bevilacqua

First Committee Member

Richard Prazenica

Second Committee Member

Troy Henderson

Third Committee Member

Eric Coyle

College Dean

James W. Gregory

Abstract

Resulting from breakup events, such as collisions and explosions, hypervelocity fragments create potential hazards for both terrestrial and on-orbit environments, such as terrestrial weapons explosions and satellite breakup events, respectively. To avoid unnecessary damage, an accurate understanding or characterization of hypervelocity fragmentation events is vital. Currently, publicly available two-line elements collected from on-orbit breakup events are limited, excluding pre-detonation parent body conditions, such as orientation, and information of smaller fragments. The uncertainty of these datasets varies between each collected set. Therefore, the overall goal of this work is to employ machine learning to estimate distribution characteristics of a space debris cloud resulting from on-orbit breakup events by supplementing existing space debris data with simulated fragmentation events and augmented datasets.

The first proposed machine learning technique uses Gaussian mixture models and K-nearest neighbors regression to estimate the spatial distribution and number of fragments resulting from an on-orbit collision breakup event. The technique is applied to real satellite debris information in the form of two-line element datasets.

Neural networks are also proposed to predict the location of satellite debris fragments following an explosion. Here, realistic terrestrial explosions are used to model the initial conditions of the satellite debris fragments relative to a parent body directly after an explosion occurs. Parent body detonation locations are randomly selected around the final orbit of NOAA-16, a weather satellite that experienced an explosion event in 2015. The debris is propagated under the influence of J2 through J6 zonal gravitational harmonics and atmospheric drag perturbations using analytic continuation, a highly precise and efficient semi-analytic integration method. The collected simulation data is used to train a neural network, from which one can predict the orbital elements of each of the fragments individually.

After initial tests with this simulation method, another fragmentation training dataset was propagated in the same manner with the addition of the final conditions of NOAA-16. Then, two-line element information for the debris resulting from the explosion of NOAA-16 were propagated using the SGP4 model and filtered for accuracy. A final dataset is formed by strategically up-sampling a limited filtered two-line element dataset at the initial timestep of the explosion and then propagating all fragments. Distributions are utilized again, this time in the form of generalized extreme value distributions, to model the orbital elements of the debris cloud at each timestep for all datasets. Deep neural networks can then be used to estimate the overall behavior of the debris cloud. By combining the real two-line element data with the synthetic simulation data and the augmented dataset, the training data contains realistic dynamics with consideration of new information that may be missed by the two-line element dataset, better generalizing the model.

A validation dataset is used to test the complete augmented machine learning approach, proving to successfully estimate orbital element distributions for semi-major axis, eccentricity, inclination, and right ascension of the ascending node for the NOAA-16 breakup event, which produces fragments in near-circular orbits. This approach is also validated by demonstrating an improvement in estimations when compared to a set of inaccurate two-line elements filtered from the training and testing dataset. In terms of novelties, machine learning successfully enables a fast space debris orbital element characterization tool with reduced computational costs and improved accuracy to bridge the gap between simulations and realistic data, capturing aspects excluded from the real two-line element data such as possible missed fragments. Overall, this dissertation successfully presents an approach contributing to space debris estimation methods.

Scholarly Commons Citation

Larsen, Katharine, "Machine Learning Methods for Hypervelocity Fragment Flyout Characterization" (2025). Doctoral Dissertations and Master's Theses. 893.
https://commons.erau.edu/edt/893