Pulse Deinterleaving for Multifunction Radars with Hierarchical Deep Neural Networks

Abstract

Multi-function radars (MFRs) work with pulse groups, and pulses in different groups are weakly correlated, which greatly increases the difficulty for MFR pulse deinterleaving, especially in cases of significant data noises. At present, no relevant research results have been reported to address this problem. In this article, a hierarchical deep learning model will be established to describe the sequential patterns of pulse trains from MFRs. The bottom layer of the model represents discrete pulse groups with continuous vectors, which describes the temporal pattern of pulse groups and makes them machine readable. The top layer uses a recursive neural network to mine the sequential pattern between consecutive pulse groups. The two layers are then synthesized to form a hierarchical model, which is able to describe the semantic correlations between different pulses within MFR pulse trains, and the parameters of subsequent pulse groups and their inner pulses can be predicted based on preceding pulses. Based on the hierarchical model, this article proposes an iterative pulse deinterleaving method and a parallel one for MFRs. Simulation results demonstrate that, the proposed methods perform satisfyingly in separating interleaved pulses from radars of the same or different types, and they are robust to significant pulse noises.