Demand-based algorithms have been widely studied in the satellite community, where the satellite’s radio resources are allocated according to the on-ground users’ demands. However, these algorithms have high computational time because they are required to optimise many parameters, which hinders the practical implementation of the algorithms. In this work, we propose a methodology to alleviate the computational complexity of demand-aware bandwidth and power allocation algorithm by combining conventional optimisation and deep learning (DL). Hence, conventional optimisation allocates the radio resources, while DL accelerates the computation.

Flying satellite communication systems often have to deal with intended and unintended radio-frequency interference, especially now with the advent of non-geostationary orbit (NGSO) systems causing orbital crowding. In this work, we investigate the use of machine learning (ML) for interference detection and classification. In particular, we investigate the effects of datasets representations on the performance of convolutional neural network (CNN) when deployed on-board of a geostationary orbit (GSO) satellite, to detect interference and classify the spectrum of interest. Focusing on the frequency representation of the observed signal, we consider different input datasets depending on the fast Fourier Transform (FFT) size and their transformation. In particular, we considered complex and magnitude values of full dimension and reduced dimension FFT. Our analysis shows that the magnitude of the reduced dimension FFT, attains the best results in terms of accuracy in detecting the presence of the interference signal, and its location in the spectrum of interest.

Recently, non-geostationary orbit (NGSO) satellite communication constellations have regained popularity due to their ability to provide global coverage and lower-latency connectivity. However, the new wave of Low Earth Orbit (LEO) satellite constellations operate on the same spectral bands as legacy satellites in geosynchronous orbit (GSO), which concurrently access the electromagnetic spectrum. Even if international regulations are in place, such increased spectral congestion will result in interference events. Therefore, both regulator entities and GSO operators have a high interest in detecting illegal or unlicensed NGSO interference sources. In this work, we simulate a realistic downlink interference scenario by emulating an actual commercial NGSO orbit whose signal is eventually received in a GSO receiver that is pointed toward a specific GSO satellite. We design an autoencoder deep neural network and we evaluate its performance considering both time-series and frequency-domain series of the overall received samples.

Coexisting satellite and terrestrial networks present a unique set of challenges and opportunities when the two networks share the same spectrum. One of these challenges is the signal recovery. In this work, we formulate
an optimization problem and propose a diffusion model to perform signal recovery. The proposed diffusion model leverages the denoising mechanism to recover the signals from noisy and distorted signals. The proposed diffusion model consists of encoder to encode the input to the latent space, U-Net for denoising, attention block to integrate different relevant feature to create better context for signal recovery, and decoder to deliver the recovered signal.