Adaptive Clustering With Similarity Learning for Enhanced Multi-Scenario Radar Signal Processing

In the complex and dynamic electromagnetic environment of modern electronic reconnaissance systems, high-density and highly intertleaving signals cause challenges to radar signal sorting (RSS). This paper introduces a novel RSS method based on similarity learning and pseudo-labeling to address the pulse deinterleaving problem in complex scenarios. The method is divided into two stages: first, unsupervised clustering is used to generate pseudo-labels necessary for calculating the pulse repetition interval (PRI); then, a two-branch network based on similarity learning uses the PRI and other signal features to determine whether two clusters correspond to the same radar. By combining pseudo-labeling with similarity learning, the method improves the performance in high-density and highly interleaving environments, provides better robustness and calculation time, and thus effectively meets the requirements of current electronic reconnaissance tasks.