Machine Learning for Packet Detection in Satellite Communications

Abstract

Satellite-based IoT networks demand efficient and robust short-packet detection techniques, particularly in Low Earth Orbit (LEO) scenarios where devices operate with low power and transmit sporadically. This thesis explores and compares two approaches to address these challenges under realistic channel conditions. The first approach employs a traditional correlation-based detection method, widely regarded as optimal in noise-limited environments but subject to performance degradation under heavier traffic loads, collisions, and channel impairments. The second approach uses a supervised learning scheme based on convolutional neural networks (CNNs), designed to handle low signal-to-noise ratio (SNR) and diverse channel impairments. Initially, both methods are evaluated under ideal, noise-limited conditions, revealing similar detection rates. However, when multiple users transmit simultaneously and random phase shifts or Doppler effects arise, the CNN consistently outperforms correlation, demonstrating greater resilience. Correlation remains attractive due to its simplicity and lower computational overhead; it also offers an inherent Doppler estimation capability when implemented as a bank of correlators. By contrast, the CNN adapts more effectively to varying channel loads and unknown scenarios, maintaining good performance in general even under severe impairments. These results underscore the potential of machine learning for next-generation packet detection in satellite networks. Future work involves extending the CNN to estimate Doppler shifts, integrating detection and frequency estimation in a single neural framework, and further exploring hybrid solutions that combine neural networks and traditional methods for improved performance.