Radar pulse repetition interval modulation recognition with combined net and domain-adaptive few-shot learning
Han Cong Feng, Bin Tang, Tao Wan
Abstract
Automatically identifying pulse repetition interval (PRI) patterns is an important task for electronic support measure (ESM) systems in a complex electromagnetic environment. Existing learning-based approaches need handcrafted features to increase recognition accuracy and are unable to deal with the potential mismatch between the training set and testing samples. In this paper, we design a multiscale convolution block with a vectorized embedding and squeeze-and-excitation mechanism for feature extraction. Accordingly, we also consider the problem of domain shift, proposing a nonlinear, metric-based relation module to make the network quickly adapt to the new environment with an extremely low number of target training samples. Experiments on the PRI simulation datasets not only show the superiority of our network architecture but also prove the effectiveness of the relation module and our training method on domain generalization.