Separation of Intertwined ADS-B Signals Using a Contrastive Learning Method

The transmission of Automatic Dependent Surveillance-Broadcast (ADS-B) signals has the characteristic of random broadcasting, and the overlapping of signals is inevitable, which will limit the surveillance performance of the ADS-B system. In response to the problem that traditional signal processing-based ADS-B signal separation algorithms have high requirements for parameters such as signal power difference, carrier frequency difference, and relative delay, ADS-B signal separation based on deep learning has become a new approach. Different from other deep learning-based ADS-B signal separation algorithms, this paper proposes a contrastive learning-based ADS-B interwoven signal separation method for three interwoven ADS-B signals. In addition to ADS-B signals, the proposed algorithm regards background noise as an additional output and simultaneously separates ADS-B signals and background noise in a masked manner. Experiments show that the proposed algorithm outperforms dual-channel Transformer, Convtasnet, recurrent neural network (RNN), etc., and the correct rate of interwoven signal separation can reach $88.25 \%$, thereby demonstrating the effectiveness of the proposed algorithm.