Modelado de la dinámica de satélites como cuerpos rígidos mediante redes neuronales artificiales para aplicaciones de control.
Loading...
Date
2026
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
En la actualidad, el ambiente espacial es uno de los recursos más explotados, aumentando cada año las misiones que se llevan a cabo en el. Estas misiones demandan un alto nivel de exactitud y precisión a la hora de controlar las naves espaciales. Sin embargo, en las naves donde se tienen grandes apéndices flexibles como antenas, paneles solares, velas solares, entre otros, se ha evidenciado que los algoritmos de control actuales no son capaces de mantener un correcto funcionamiento debido a las vibraciones inducidas por los apéndices. Esto dado que los modelos generalmente utilizados corresponden a modelos rígidos o flexibles simplificados. Es aquí, donde las redes neuronales artificiales (ANN) se presentan como una buena opción para aproximar su dinámica de una manera eficiente. En este trabajo se aborda este desafío al desarrollar un modelo en base a redes neuronales artificiales de la dinámica de cuerpo rígido de un cubesat-6U como primer acercamiento y estudio de su capacidad de trabajo con estos sistemas para su posterior implementación como modelo de dinámica flexible. La metodología contempló la implementación de una estructura de simulación de alta fidelidad en el software Basilisk. Estas simulaciones se llevaron a cabo para una ´orbita baja (LEO) y contemplaron las cuatro perturbaciones ambientales más significativas: arrastre atmosférico, presión por radiación solar, gradiente de gravedad y campo electromagnético terrestre. Además, se incorporaron tres ruedas de reacción (RW) como controladores de actitud, las cuales mediante señales pseudoaleatorias (PRBS) excitaron el sistema, asegurando la captura de su dinámica. Se realizaron 2744 simulaciones con condiciones iniciales de velocidad angular y acciones de control diferentes. Con los datos extraídos de las simulaciones se procedió al entrenamiento de diferentes arquitecturas de redes Perceptrón Multicapa (MLP), las cuales se sometieron a tres procesos de discriminación que determinaron el impacto de cada variable en el desempeño de las redes, entregando cuatro modelos funcionales. Los resultados indican que la dinámica interna de las RW constituye la mayor fuente de complejidad no lineal del sistema. Además, se demostró que incrementar el número de variables de entrada no se traduce necesariamente en mejores predicciones; de hecho, la exclusión de variables como los cuaterniones o inercias constantes benefició el ajuste . El modelo con mejor rendimiento global (ANN C) requirió únicamente tres variables de entrada: la velocidad angular del satélite (ωsat), las acciones de control (uctrl) y la velocidad angular de las RW (ωrw). Este modelo logró un Error Relativo Medio (MRE) de 2.14% tanto en la predicción del instante siguiente como en ventanas recursivas, requiriendo tan solo 62.691 parámetros entrenables y 123.904 operaciones de punto flotante (Flop).
The space environment is currently one of the most utilized resources, with the number of missions there increasing annually. These missions demand a high level of accuracy and precision in spacecraft control. However, for spacecraft with large flexible appendages such as antennas, solar panels, and solar sails, existing control algorithms have proven incapable of maintaining proper operation due to the vibrations induced by these components. This limitation arises because the models typically employed are either rigid-body models or simplified flexible models. Artificial Neural Networks (ANNs) offer a promising solution for efficiently approximating this dynamic behavior. This study addresses this challenge by developing an ANN-based model of the rigid-body dynamics of a cubesat-6U, acting as an initial approach to evaluating the technology’s suitability for such systems. The methodology involved implementing a high-fidelity simulation framework using Basilisk software. Simulations were made for a Low Earth Orbit (LEO) scenario and were added the four most significant environmental perturbations: atmospheric drag, solar radiation pressure, gravity gradient, and the Earth’s electromagnetic field. Additionally, three reaction wheels (RW) were incorporated as attitude actuators; these excited the system using pseudorandom binary sequences (PRBS) to ensure the capture of its dynamic behavior. A total of 2,744 simulations were performed, varying initial angular velocity conditions and control inputs. Data extracted from these simulations were used to train various Multilayer Perceptron (MLP) network architectures; after three stage discrimination process, these converged into four functional models. The results indicate that the internal dynamics of the reaction wheels are the primary source of nonlinear complexity within the system. Furthermore, the study demonstrated that increasing the number of input variables does not necessarily lead to improved predictions; actually, excluding variables such as quaternions or constant inertia values improved the model fit. The model with the best overall performance (ANN C) required only three input variables: the satellite’s angular velocity (ωsat), the control actions (uctrl), and the angular velocity of the reaction wheels (ωrw). This model achieved a Mean Relative Error (MRE) of 2.14% for both single-step-ahead prediction and recursive windows, requiring only 62,691 trainable parameters and 123,904 floating-point operations (FLOPs).
The space environment is currently one of the most utilized resources, with the number of missions there increasing annually. These missions demand a high level of accuracy and precision in spacecraft control. However, for spacecraft with large flexible appendages such as antennas, solar panels, and solar sails, existing control algorithms have proven incapable of maintaining proper operation due to the vibrations induced by these components. This limitation arises because the models typically employed are either rigid-body models or simplified flexible models. Artificial Neural Networks (ANNs) offer a promising solution for efficiently approximating this dynamic behavior. This study addresses this challenge by developing an ANN-based model of the rigid-body dynamics of a cubesat-6U, acting as an initial approach to evaluating the technology’s suitability for such systems. The methodology involved implementing a high-fidelity simulation framework using Basilisk software. Simulations were made for a Low Earth Orbit (LEO) scenario and were added the four most significant environmental perturbations: atmospheric drag, solar radiation pressure, gravity gradient, and the Earth’s electromagnetic field. Additionally, three reaction wheels (RW) were incorporated as attitude actuators; these excited the system using pseudorandom binary sequences (PRBS) to ensure the capture of its dynamic behavior. A total of 2,744 simulations were performed, varying initial angular velocity conditions and control inputs. Data extracted from these simulations were used to train various Multilayer Perceptron (MLP) network architectures; after three stage discrimination process, these converged into four functional models. The results indicate that the internal dynamics of the reaction wheels are the primary source of nonlinear complexity within the system. Furthermore, the study demonstrated that increasing the number of input variables does not necessarily lead to improved predictions; actually, excluding variables such as quaternions or constant inertia values improved the model fit. The model with the best overall performance (ANN C) required only three input variables: the satellite’s angular velocity (ωsat), the control actions (uctrl), and the angular velocity of the reaction wheels (ωrw). This model achieved a Mean Relative Error (MRE) of 2.14% for both single-step-ahead prediction and recursive windows, requiring only 62,691 trainable parameters and 123,904 floating-point operations (FLOPs).
Description
Tesis presentada para optar al título de Ingeniero/a Civil Aeroespacial.
Keywords
Satélites, Artificial neural networks, Naves espaciales, Controladores programables