Abstract
Space-based automatic dependent surveillance-broadcast (ADS-B) systems offer the potential for comprehensive global aircraft surveillance. However, they face substantial challenges due to severe signal collisions resulting from the simultaneous reception of asynchronous ADS-B transmissions from multiple aircraft within a satellite's expansive coverage area. Traditional collision mitigation approaches, such as serial interference cancellation and multichannel blind source separation, often have high computational costs, impose strict signal structure constraints, or rely on multiple-antenna configurations, all of which limit their practicality in satellite scenarios. To address these limitations, this paper proposes two novel deep learning-based models, designated SplitNet-2 and SplitNet-3. SplitNet-2 leverages a Transformer-inspired self-attention architecture specifically designed to separate two overlapping ADS-B signals, while SplitNet-3 employs a convolutional residual U-shaped network optimized for disentangling three simultaneous, colliding signals. Extensive simulations under realistic satellite reception conditions demonstrate that the proposed models significantly outperform conventional methods, achieving lower bit error rates (BERs) and improved demodulation accuracy. These advancements offer a promising solution to the critical problem of underdetermined signal separation in space-based ADS-B reception and significantly enhance the reliability and coverage of satellite-based ADS-B surveillance systems.