Estudio e implementación de redes neuronales de bajo consumo sbre RISC-V para aplicaciones espaciales
Loading...
Official URL
Full text at PDC
Publication date
2025
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citation
Abstract
En los últimos años el número de adversidades climáticas y ambientales se ha visto en aumento. Para poder hacerles frente la teledetección nos ayuda dándonos datos para poder prevenir y mitigar sus efectos. Para llevar a cabo esta teledetección es necesario contar con un satélite. Afortunadamente, debido al avance de la tecnología, existen alternativas sin una barrera económica como sería el estándar de nanosatélites CubeSat. Estos nanosatélites cuentan con restricciones de tamaño, energía y tiempo de procesamiento muy ajustados, por lo que optimizar estos factores es vital.
A lo largo de este Trabajo Fin de Grado, estudiamos alternativas a algoritmos clásicos propuestos en un TFG previo sobre la arquitectura RISC-V, como serían el uso de redes neuronales (convolucionales en concreto), y las comparamos explorando el trade-off de tiempo de ejecución y precisión. Además, se proponen una serie de optimizaciones al modelo de red convolucional propuesto consiguiendo reducir su tamaño y tiempo de inferencia, sin ver reducida su precisión.
Al analizar los resultados obtenidos concluimos que pese a tener una menor precisión la mejora de tiempo es suficientemente significativa para contrarrestar la perdida de precisión y aumentando de esta forma la eficacia.
In recent years, the number of climatic and environmental adversities has increased. To be able to deal with them, remote sensing helps us by giving us data to be able to prevent and mitigate its effects. To carry out this remote sensing, it is necessary to have a satellite. Fortunately, due to the advancement of technology, there are alternatives without an economic barrier, such as the CubeSat nanosatellite standard. These nanosatellites have very tight size, energy, and processing time constraints, so optimizing these factors is vital. Throughout this Bachelor’s Degree Thesis, we study alternatives to classical algorithms proposed in a previous Thesis [1] on the RISC-V architecture, such as the use of neural networks (convolutional in particular), and we compare them by exploring the trade-off of execution time and precision. In addition, a series of optimizations are made to the proposed convolutional network model, managing to reduce its size and inference time, without reducing its precision. When analyzing the results obtained, we conclude that despite having a lower precision, the improvement in time is significant enough to counteract the loss of precision and thus increase the efficiency.
In recent years, the number of climatic and environmental adversities has increased. To be able to deal with them, remote sensing helps us by giving us data to be able to prevent and mitigate its effects. To carry out this remote sensing, it is necessary to have a satellite. Fortunately, due to the advancement of technology, there are alternatives without an economic barrier, such as the CubeSat nanosatellite standard. These nanosatellites have very tight size, energy, and processing time constraints, so optimizing these factors is vital. Throughout this Bachelor’s Degree Thesis, we study alternatives to classical algorithms proposed in a previous Thesis [1] on the RISC-V architecture, such as the use of neural networks (convolutional in particular), and we compare them by exploring the trade-off of execution time and precision. In addition, a series of optimizations are made to the proposed convolutional network model, managing to reduce its size and inference time, without reducing its precision. When analyzing the results obtained, we conclude that despite having a lower precision, the improvement in time is significant enough to counteract the loss of precision and thus increase the efficiency.
Description
Trabajo de fin de Grado en Ingeniería Informática, Facultad de Informática UCM, Departamento de Arquitectura de Computadores y Automática, Curso 2024/2025