ETD

Archivio digitale delle tesi discusse presso l'Università di Pisa

Tesi etd-09062021-161227


Tipo di tesi
Tesi di laurea magistrale
Autore
FERRANTE, NICOLA
URN
etd-09062021-161227
Titolo
Fault Detection Isolation and Recovery exploiting Artificial Intelligence in Satellite Systems
Dipartimento
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
COMPUTER ENGINEERING
Relatori
relatore Prof. Fanucci, Luca
relatore Prof. Bechini, Alessio
relatore Ing. Giuffrida, Gianluca
Parole chiave
  • FDIR
  • Satellite
  • Artificial Intelligence
  • Deep Learning
  • Recurrent Neural Networks
Data inizio appello
24/09/2021
Consultabilità
Non consultabile
Data di rilascio
24/09/2091
Riassunto
Fault detection, isolation and recovery (FDIR) techniques have gained in importance as the complexity and responsibilities of modern systems increase, thanks to improving electronic devices' performance while reducing power consumption. Furthermore, the application of Artificial Intelligence (AI) techniques to implement this functionalities have been evaluated, and there is evidence that these algorithms can be used to reduce the domain expertise and the time-to-market required to implement model-based FDIR systems, with the further advantage of a lower computational effort. Focusing on space systems, FDIR is essential to reach the required robustness and automation levels that allow to complete successfully a mission. The objective of this thesis is to design an on-board satellite FDIR system, exploiting Recurrent Neural Network (RNN) algorithms. The proposed FDIR system shall respect the constraints of a satellite system, such as limited energy and computational power. In this work, a deep learning model, based on recurrent neural networks, in particular Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), was trained to forecast housekeeping signals coming from Mars Advanced Radar for Subsurface And Ionosphere Sounding (MARSIS), which was launched with ESA's mission Mars Express. This approach allows to implement Model Based FDIR without any knowledge of the underlying physics of the system, substituting computationally expensive equations with a deep learning model, which can be efficiently executed using modern low power hardware accelerators. Finally, thanks to the exploitation of a general approach as the regression, the proposed solution can be applied to multiple functional blocks and also extended to other application domains.
File