Neural-Network Based Thermal Modeling of Small Satellites
A First-Principles Approach
U.U. Bhat (TU Delft - Aerospace Engineering)
S. Speretta – Mentor (TU Delft - Space Systems Egineering)
Alessandra Menicucci – Graduation committee member (TU Delft - Space Systems Egineering)
Ines Uriol Balbin – Graduation committee member (TU Delft - Aerospace Structures & Computational Mechanics)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
The use of small satellites, enabled by the standardization of the CubeSat specifications and miniaturization in electronics, has seen a rapid increase in the past decades. The low-cost and short development time of these satellites has made them an attractive option for both commercial and academic applications, making space exploration more accessible. However, these small satellites are prone to failures, leading to lost scientific potential. Mitigation of these failures forms the motivation for this thesis. Recent advances in neural networks have shown promise in the field of anomaly detection. The black-box nature of such models, however, makes it challenging to understand the reasoning behind their predictions.
Constraining the data-driven models with known physics can not only help us understand the reasoning behind their predictions, but also ensuring the model is consistent with the real-world behavior of the system. The work presented in this Master's thesis aims to demonstrate the advantages of such first-principles neural networks over purely data-driven models in thermal behavior modeling of small satellites. Baseline performance of data-driven Long Short-Term Memory (LSTM) networks is established using FUNCube-1 telemetry data, quantifying the temperature prediction accuracy of the models under ideal conditions. The limitations of these models, especially with sparse data, are then investigated, to highlight the need for more robust models.
First-principles models, based on a physics-informed curve-fit and simplified thermal network models, are then developed to constrain the data-driven model predictions. The first-principles models are shown to be more robust to sparse data, with the predictions on data not seen during training being more consistent with the real-world thermal behavior of the satellite. Methods to relate the first-principles model parameters to the physical properties of the satellite are also proposed and explored, to help extract the evolution of the thermal behavior of the satellite over time.