Radar signal deinterleaving in open-set environments based variational autoencoder with probabilistic ladder structure

Abstract

In the field of electronic reconnaissance, deinterleaving techniques for radar signals are crucial. Although a large number of studies have been devoted to the classification of known radar signals by recurrent neural networks under closed set conditions, this task remains challenging in open set environments. To this end, this paper introduces a novel variational autoencoder (LVAEGRU) based on gated recurrent units that incorporates a probabilistic ladder structure. This model aims at capturing higher level abstract features through probabilistic ladder structure, thus avoiding information loss at intermediate levels. By forcing the latent representation to approximate different multivariate Gaussian distributions and combining this with reconstructing the loss information, the method performs well in open-set deinterleaving tasks. Experimental results show that the method proposed in this paper exhibits excellent performance in open-set scenarios compared to multiple baseline methods.