CBL - Campus del Baix Llobregat

Projecte llegit

Títol: Machine learning applied to spectral mass gauging

Estudiants que han llegit aquest projecte:

Director/a: GONZÁLEZ CINCA, RICARD

Departament: FIS

Títol: Machine learning applied to spectral mass gauging

Data inici oferta: 06-02-2026     Data finalització oferta: 06-10-2026


Estudis d'assignació del projecte:
    MU AEROSPACE S&T 21
Tipus: Individual
 
Lloc de realització: EETAC
 
Segon director/a (UPC): ABECIA HERNANZ, SARA CECILIA
 
Paraules clau:
Spectral Mass Gauging, Machine Learning
 
Descripció del contingut i pla d'activitats:
Machine learning applied to spectral mass gauging for orbital propellant tanks
 
Overview (resum en anglès):
In the last decade the space sector has been expanding and growing extremely fast, fueled by both space agencies and new private companies, placing thousands of satellites in Low Earth Orbit (LEO). Because of this, the need for long-term storage of fuel and refueling has arised, for which is necessary to measure, as accurately as possible, the volume of fuel inside a tank.
This poses a problem, as in microgravity conditions, fuel is not settled in the bottom but takes different shapes. Measuring the volume without knowing the liquid's layout inside the tank proved to be a difficult task. In recent years, UPC's Space Exploration Laboratory has been researching this topic within the framework of both national projects, funded by the Spanish State Research Agency (AEI), and international projects, funded by ESA, NASA, CNES, and
the European Commission. One of the many solutions to this problem is Spectral Mass Gauging (SMG), which relies on
H. Weyl's postulates to infer the volume of liquid by counting the eigenfrequencies of the liquid given a certain excitation, which manifest as peaks in the frequency spectrum. Then, it is of the utmost importance to correctly detect these peaks, task for which we propose a Machine
Learning (ML) approach.
By training a one-dimensional Convolutional Neural Network (CNN 1D) with thousands of synthetically created datasets that aim to mimic the real data obtained from experiments, we seek to create a model that only detects peaks associated with liquid modes, to then infer the volume. In addition to creating noisy spectra with gaussian distributions where liquid modes are to be located, a study has been conducted on the attenuation parameter of the time signal, intimately linked to the width of the liquid modes in the frequency domain. Here, satisfactory results have been obtained, calculating an approximate theoretical value of 0-1
Hz of Full Width at Half Maximum (FWHM) of liquid modes, compared to the 1-3 Hz that have been measured.
The results show how testing this model with another synthetic spectrum, it is able to properly detect all peaks, even when a high level of Additive White Gaussian Noise (AWGN) is applied. However, when using real experimental data, the model detects all peaks, regardless of them
being noise, shell modes or liquid modes. For this reason, we conclude that better characterisation and study of the different parameters affecting the synthetic maps is to be made in an effort to create synthetic maps that are more similar to those obtained experimentally.

© CBLTIC Campus del Baix Llobregat - UPC