Deinterleaving and Clustering Radar Pulse Trains via Gated Graph Convolution Deep Clustering

This paper proposes a novel graph convolutional deep clustering approach for radar signal deinterleaving. The algorithm features an innovative PRI Frequency-Driven Adjacency Matrix Generation (PF-AMG) mechanism that constructs weighted graph structures by statistically analyzing the frequency distribution of Pulse Repetition Intervals (PRIs) in Pulse Descriptor Word (PDW) trains. Leveraging a Graph Gated Convolutional Network, the proposed method adopts an inductive learning paradigm: initially training on PDW graphs with fully known node attributes, followed by generalization inference on new PDW graphs. Through optimization of a deep clustering loss function, the algorithm enforces dual constraints: ensuring compact intra-cluster cohesion for pulses from identical emitter while maintaining maximum inter-cluster separation between different emitters. Simulation results demonstrate that the proposed method maintains over 90% deinterleaving accuracy under challenging scenarios with 30% loss pulse rate and 30% spurious pulse rate, exhibiting significant advantages in complex electromagnetic environments.