Convolutional Autoencoders for Non-Geostationary Satellite Interference Detection

The new wave of mega-constellations in low-earth orbits (LEOs), which occupy frequency bands that are also used by legacy geostationary orbit satellites (GSOs), will eventually result in increased interference events. In this paper, we explore the application of unsupervised deep learning models, specifically convolutional autoencoders (CAEs), for detecting non-Geostationary Orbit Satellite (NGSO) interference in GSO ground stations (GGS). We propose and evaluate two DL models, the CAE1D model handling amplitude values as 1D data, and the CAE2D model handling In-phase/Quadrature (IQ) samples as 2D data. Through rigorous experimentation, we examine the models' performance against traditional energy detector (ED) methods, employing single-model (SM) and multi-models (MMs) approaches for training the models. Our findings reveal that DL models, particularly under the MMs approach, significantly outperform conventional methods with up to 11% in the probability of detecting interference, demonstrating the potential of advanced machine learning techniques to improve the reliability of satellite communication systems, especially for such fastvarying interference environments (LEO mobility), where only GSO interference-free signals are needed for training the models.