Orbits around the Earth have become extremely populated nowadays, hence, to assure feasibility and safety of future space missions, we have to take into account all space debris and spacecrafts populating these orbits. The aim of this thesis work is to exploit machine learning techniques in order to retrieve space debris angular velocity from the light curves collected from images, making possible to overcome classical techniques used in this field. These information could be useful to avoid crashes during missions or to start a program of removal of some of these debris from orbits. The first step is the creation of synthetic images, that simulates a night sky images captured with telescopes.These artificial images must contain a tracklet, that is a luminous curve on the image representing the trajectory of the object. Space objects crossing the sky, like space debris, typically are not light sources, but the trajectory followed is nevertheless illuminated in these images. This has been done because also in real images, coming from campaigns of observation, the passage of a space object would be visible. The path followed by these space objects is illuminated since they reflect light coming from other celestial bodies, usually the Sun for objects around the Earth, and reflect it towards the ground. Indeed these tracklets are also called light curves. The variation of luminous intensity along these curves can give an idea on the motion of these space objects. Since, up to now, space debris rotation rates are unknown, real images coming from campaigns of observation can’t be used directly in the present work and hence synthetic images are needed. During the creation of images some desired features are selected through the tuning of the parameters. This choice allows to have images and also the information about the light curve that would have been unavailable with real images. These information, combined with images, are fundamental in this work, since the neural network needs them to train itself. Information coming from artificial images need to be elaborated up to the final stage, in order to be then used as input for the neural network. Once trained, the neural network could be exploited to determine the motion of unseen space objects, also in real cases. Different neural network models have been implemented. The best results are obtained with the model having as output the lowest rotation rate. In this case the training loss goes under 0.01 while the validation loss below 0.1.
Lo scopo di questa tesi è quello di sfruttare le tecniche di intelligenza artificiale in modo da estrarre automaticamente la velocità angolare degli oggetti in orbita dalle loro curve di luce raccolte dalle immagini. Queste informazioni potranno essere utili al fine di evitare collisioni durante le missioni spaziali o per iniziare un programma di rimozione di alcuni di questi detriti spaziali dalle orbite. Il primo passo è la creazione di immagini sintetiche che simulino le immagini del cielo notturno catturate con dei telescopi. Queste immagini artificiali devono contenere una tracklet, ossia una curva luminosa, che rappresenti la traiettoria dell’oggetto. Gli oggetti spaziali che attraversano il cielo, come i detriti spaziali, tipicamente non sono delle sorgenti di luce, ma nonostante ciò la traiettoria seguita è illuminata in queste immagini. E’ stato fatto questo perchè anche nelle immagini reali, provenienti da campagne di osservazione, il passaggio di un oggetto sarebbe visibile. Il percorso seguito da questi oggetti spaziali è illuminato in quanto essi riflettono la luce proveniente da altri corpi celesti, tipicamente del Sole per detriti attorno alla Terra, verso il suolo. Perciò queste tracklet sono anche chiamate curve di luce. Le variazioni di intensità luminosa lungo queste curve possono dare un’idea del moto di questi oggetti. Dato che, per ora, le velocità di rotazione dei detriti spaziali sono sconosciute, le immagini reali provenienti da campagne di osservazione non sono utilizzabili direttamente in questo lavoro e quindi si rendono necessarie le immagini sintetiche. Durante la creazione delle immagini alcune caratteristiche sono selezionate attraverso il settaggio dei parametri. Questa scelta permette di avere le immagini e anche le informazioni riguardanti le curve di luce che non sarebbero state disponibili con l’impiego di immagini reali. Queste informazioni, combinate con le immagini, sono utili agli scopi di questo lavoro, in quanto la rete neurale le richiede per potersi allenare. Le informazioni provenienti dalle immagini artificali necessitano di essere elaborate fino alla forma finale utile in modo da essere date come input alla rete neurale. Una volta allenata, la rete neurale potrà essere sfruttata per determinare il moto di oggetti spaziali non ancora esaminati, anche in casi reali. Sono stati utilizzati diversi modelli di rete neurale. I migliori risultati sono stati ottenuti con il modello avente come output la velocità di rotazione minore. In questo caso la training loss scende oltre lo 0,01, mentre la validation loss raggiunge valori inferiori allo 0,1.
Space object tumbling motion reconstruction from light curves analysis with machine learning techniques
LA ROSA, MATTIA
2019/2020
Abstract
Orbits around the Earth have become extremely populated nowadays, hence, to assure feasibility and safety of future space missions, we have to take into account all space debris and spacecrafts populating these orbits. The aim of this thesis work is to exploit machine learning techniques in order to retrieve space debris angular velocity from the light curves collected from images, making possible to overcome classical techniques used in this field. These information could be useful to avoid crashes during missions or to start a program of removal of some of these debris from orbits. The first step is the creation of synthetic images, that simulates a night sky images captured with telescopes.These artificial images must contain a tracklet, that is a luminous curve on the image representing the trajectory of the object. Space objects crossing the sky, like space debris, typically are not light sources, but the trajectory followed is nevertheless illuminated in these images. This has been done because also in real images, coming from campaigns of observation, the passage of a space object would be visible. The path followed by these space objects is illuminated since they reflect light coming from other celestial bodies, usually the Sun for objects around the Earth, and reflect it towards the ground. Indeed these tracklets are also called light curves. The variation of luminous intensity along these curves can give an idea on the motion of these space objects. Since, up to now, space debris rotation rates are unknown, real images coming from campaigns of observation can’t be used directly in the present work and hence synthetic images are needed. During the creation of images some desired features are selected through the tuning of the parameters. This choice allows to have images and also the information about the light curve that would have been unavailable with real images. These information, combined with images, are fundamental in this work, since the neural network needs them to train itself. Information coming from artificial images need to be elaborated up to the final stage, in order to be then used as input for the neural network. Once trained, the neural network could be exploited to determine the motion of unseen space objects, also in real cases. Different neural network models have been implemented. The best results are obtained with the model having as output the lowest rotation rate. In this case the training loss goes under 0.01 while the validation loss below 0.1.File | Dimensione | Formato | |
---|---|---|---|
Tesi finale.pdf
Open Access dal 02/04/2022
Descrizione: La Rosa Mattia - Master Degree thesis - Space objects tumbling motion reconstruction from light curves analysis with machine learning techniques
Dimensione
6.01 MB
Formato
Adobe PDF
|
6.01 MB | 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/10589/173766