In the recent years the number of space objects and debris has exponentially increased creating a serious risk of collisions in orbit. Space Safety Program is working to overcome this problem in several ways. One of the key task is the so called maneuver detection used to identify and predict whenever a satellite will perform a maneuver while in orbit. Thanks to the increasing amount of data collected through the years, new data driven techniques have been implemented thanks to machine learning. This approach is showing promising results even better than traditional methods based on statistical analysis. This thesis goes into detail of Long Short Term Memory, an alternative version of the recurrent neural network, developing different architectures in order to perform maneuver detection, maneuver characterization and prediction of orbital parameters. Several satellites are taken into consideration using the keplerian elements as input for the neural network. After the creation of the dataset, it is first preprocessed and divided into sequential samples and then associated to the output during the training. This method confirms the high performances obtained in previous works reaching values of precision and recall around 99% for all the satellites. The real strength of this method is that it demonstrated a good ability to generalize the process with different types of satellites, orbital patterns and tasks to perform. The results obtained in this thesis offer a valid possible solution to face the problem highlighted by safety programs and show the importance of continuing the research in this direction.
Negli ultimi anni il numero di oggetti e detriti spaziali è aumentato esponenzialmente creando un serio rischio di collisioni in orbita. Lo Space Safety Program sta lavorando per risolvere questo problema in vari modi. Uno degli aspetti fondamentali è l' identificazione delle manovre utile per trovare e predire se e quando un satellite manovrerà in orbita. Grazie al continuo aumento di dati raccolti negli anni, sono state ideate nuove tecniche basate sull'analisi di dati grazie al machine learning. Questa tesi entra nel dettaglio della Long Short Term Memory, una versione alternativa di rete neurale ricorsiva, sviluppando diverse architetture per svolgere identificazione e caratterizzazione delle manovre e predizione dei parametri orbitali. Sono stati considerati diversi satelliti di cui sono stati analizzati gli elementi kepleriani come input della rete neurale. Dopo la creazione, il dataset viene prima preprocessato e poi diviso in sample sequenziali per poi essere associati al rispettivo output durante l'allenamento. Questo metodo conferma le ottime prestazioni ottenute nei precedenti lavori raggiungendo valori di precisione e recall intorno al 99% per tutti i satelliti. Il grande punto di forza di questo metodo è la sua abilità nel generalizzare il processo con diversi tipi di satelliti, pattern orbitali e obiettivi da raggiungere. I risultati ottenuti in questa tesi offrono una possibile valida soluzione al problema sottolineato dai programmi di sicurezza spaziali e mostrano l'importanza di continuare le ricerche in questa direzione.
Space objects maneuver detection : characterization and prediction using recurrent neural networks
Leonzio, Italo
2022/2023
Abstract
In the recent years the number of space objects and debris has exponentially increased creating a serious risk of collisions in orbit. Space Safety Program is working to overcome this problem in several ways. One of the key task is the so called maneuver detection used to identify and predict whenever a satellite will perform a maneuver while in orbit. Thanks to the increasing amount of data collected through the years, new data driven techniques have been implemented thanks to machine learning. This approach is showing promising results even better than traditional methods based on statistical analysis. This thesis goes into detail of Long Short Term Memory, an alternative version of the recurrent neural network, developing different architectures in order to perform maneuver detection, maneuver characterization and prediction of orbital parameters. Several satellites are taken into consideration using the keplerian elements as input for the neural network. After the creation of the dataset, it is first preprocessed and divided into sequential samples and then associated to the output during the training. This method confirms the high performances obtained in previous works reaching values of precision and recall around 99% for all the satellites. The real strength of this method is that it demonstrated a good ability to generalize the process with different types of satellites, orbital patterns and tasks to perform. The results obtained in this thesis offer a valid possible solution to face the problem highlighted by safety programs and show the importance of continuing the research in this direction.File | Dimensione | Formato | |
---|---|---|---|
2023_05_Leonzio_Tesi_01.pdf
accessibile in internet per tutti
Descrizione: Tesi
Dimensione
4.12 MB
Formato
Adobe PDF
|
4.12 MB | Adobe PDF | Visualizza/Apri |
2023_05_Leonzio_ExecutiveSummary_02.pdf
accessibile in internet per tutti
Descrizione: Executive Summary
Dimensione
451.76 kB
Formato
Adobe PDF
|
451.76 kB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/209312