Application of Neural Networks in the Altitude Propagation of CubeSats

  • Shaima Z. N. Bahumaish

Student thesis: Master's Thesis

Abstract

Due to the many benefits of CubeSats, nowadays many universities and other educational institutes have their own space related programs and activities. As a result, altitude propagators started to draw more attention. Such propagators describe a satellite's dynamic motion as a result of the forces acting on it. Existing propagators such as SGP4 and HPOP are reliable but the prediction error could be decreased by applying new methods of machine learning. To minimize such error, in this thesis, MYSAT-1 TLE data were utilized to create a propagator function using a relatively new approach which is the 'Neural Network Fitting Tool'. MYSAT-1 is a 1U CubeSat currently in Low Earth Orbit (LEO). The semi-major axis of MYSAT-1 was the main orbital parameter considered in this study. Data collection started on the 23rd of April 2019 at a semi-major axis of 6840.1 km. The inputs and the targets are the semi-major axis at a certain day and that at the following day. The training algorithm used in this study is the Levenberg-Marquardt showing a mean square error with an order of magnitude of 10-5. A function was generated from the Neural Network activation functions then tested and compared with the available propagators used by researchers all around the world. The well-known SGP4 and HPOP propagators showed lower accuracy than the Neural Network generated function created using the same input for testing and validation. A second function was generated using both the semi-major and semi-minor axes of a certain day as inputs and semi-major and semi-minor axes on the following day as targets. The same range of data of the first function was used. Training was carried out using Levenberg-Marquardt and the results showed a mean square error of an order of magnitude 10-4 and of 10-3 in the first testing and second testing procedures, respectively. For the conditions considered, the first function showed more compatibility with the deep-learning-based propagators.
Date of AwardJun 2020
Original languageAmerican English

Keywords

  • Two-Line-Element (TLE); 1U CubeSat; Disturbances; Low Earth Orbit (LEO); Propagator; Neural Network (NN);

Cite this

'