A dynamic class boundary enhancement with neighbour metric domain adaptation method for open-set radar modulation signal recognition

Radar signal modulation recognition is crucial for determining strategies in electronic warfare. In order to identify the modulation patterns of unknown radar signals in non-cooperative scenarios, we propose a dynamic class boundary enhancement with Neighbour Metric Domain Adaptation method called DCBE-NMDAN. In this algorithm, we design a Multi-level Feature Fusion Network for feature extraction. To distinguishing known and unknown categories, we propose a Dynamic Class Boundary Enhancement strategy which can dynamically set the classification threshold for each category. In order to further enhance the classification boundary between different categories, we use KL-divergence to optimize the probability distribution of the model output. On this basis, Neighbour Metric Domain Adaptation Network is proposed to align the source domain with the target domain containing unknown categories. Experiments show that the proposed method can accurately distinguish a variety of known and unknown radar signals in different SNRs.