The Earth-Moon system continuously experiences a high influx of meteoroids, often resulting in impacts. While Earth’s atmosphere acts as a shield against smaller objects, the growing number of manned and unmanned space missions active in the Earth-Moon system necessitate a deeper understanding of the meteoroid environment. In this context, the 12U spacecraft LUMIO is designed to characterise meteoroid impacts on the lunar farside by capturing the radiation emitted from these impacts, complementing Earth-based observations. However, processing the recorded data to accurately retrieve impactor properties presents a challenge, as no information about the velocity can be reliably measured. Association of the impacts with their originating source can help in constraining the properties of the meteoroid. The main goal of this study is to determine whether machine learning methods can effectively determine the impacting meteoroid sources in the context of the LUMIO mission. The effectiveness of several supervised algorithms in identifying meteoroid origins will be assessed and compared, supported by key performance metrics, leading to the identification of the best-performing algorithm for this application. The comparison of the methodologies investigated in the thesis highlights that decision tree and feedforward network architectures yield the highest performance in terms of accuracy. However, a limitation in the maximum achievable performance is observed due to the presence of sporadic source impacts, which introduce noise into the training process.
Il sistema Terra-Luna è costantemente bombardato da meteoroidi, che spesso risultano in impatti. Mentre l’atmosfera terrestre agisce da barriera contro gli oggetti più piccoli, il numero crescente di missioni spaziali in ambiente Terra-Luna richiede una conoscenza e comprensione più profonda dell’ambiente dei meteoroidi. In questo ambito, LUMIO si pone come obiettivo primario quello di caratterizzare gli impatti sul lato nascosto della luna, misurando la radiazione emessa a seguito dell'impatto. Queste osservazioni verranno integrate a quelle sulla terra, permettendo di ampliare la conoscenza su questi oggetti. Tuttavia, l’elaborazione dei dati per caratterizzare le proprietà rappresenta una sfida, poiché non viene rilevata alcuna informazione riguardo la velocità. L'associazione degli impatti con la loro fonte di origine può, però, aiutare in questo calcolo. L’obiettivo principale di questa tesi è di valutare se i metodi di machine learning possono effettivamente determinare l’origine del meteoroide nel contesto della missione LUMIO. Le prestazioni dei diversi algoritmi vengono valutati e confrontati, sulla base di alcuni parametri chiave, che porteranno all'identificazione del metodo più performante. Tra le metodologie considerate, l'albero decisionale e la rete neurale 'feedforward' risultano le architetture più performanti. Sebbene, le accuratezze ottenute siano limitate a causa della presenza di impatti dovuti a eventi sporadici.
Machine learning for lunar meteoroid impact classification: a study for the LUMIO mission
Vijayakumaran, Nishani
2023/2024
Abstract
The Earth-Moon system continuously experiences a high influx of meteoroids, often resulting in impacts. While Earth’s atmosphere acts as a shield against smaller objects, the growing number of manned and unmanned space missions active in the Earth-Moon system necessitate a deeper understanding of the meteoroid environment. In this context, the 12U spacecraft LUMIO is designed to characterise meteoroid impacts on the lunar farside by capturing the radiation emitted from these impacts, complementing Earth-based observations. However, processing the recorded data to accurately retrieve impactor properties presents a challenge, as no information about the velocity can be reliably measured. Association of the impacts with their originating source can help in constraining the properties of the meteoroid. The main goal of this study is to determine whether machine learning methods can effectively determine the impacting meteoroid sources in the context of the LUMIO mission. The effectiveness of several supervised algorithms in identifying meteoroid origins will be assessed and compared, supported by key performance metrics, leading to the identification of the best-performing algorithm for this application. The comparison of the methodologies investigated in the thesis highlights that decision tree and feedforward network architectures yield the highest performance in terms of accuracy. However, a limitation in the maximum achievable performance is observed due to the presence of sporadic source impacts, which introduce noise into the training process.File | Dimensione | Formato | |
---|---|---|---|
2024_12_Vijayakumaran_Executive_Summary.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Executive Summary
Dimensione
1.09 MB
Formato
Adobe PDF
|
1.09 MB | Adobe PDF | Visualizza/Apri |
2024_12_Vijayakumaran_Thesis.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Thesis Manuscript
Dimensione
6.86 MB
Formato
Adobe PDF
|
6.86 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/10589/230306