A Contextually-Aware Residual–Recurrent Framework for Radar Pulse Deinterleaving

Radar pulse deinterleaving in dense electromagnetic environments is hindered by overlapping emitters, pulse dropouts, and agile signal behaviors. We propose a unified deep learning framework that fuses spatial, temporal, and contextual cues for pulse-level classification. The architecture combines residual convolutions, bidirectional GRUs, and attention-based fusion, operating directly on quantized PRI–PW pulse descriptors. It achieves 98.1% accuracy for known emitters with reliable open-set detection at a 0.75 confidence threshold, and maintains above 92% accuracy even under 60% pulse dropout, demonstrating strong resilience to missed detections and dense spectral overlap—consistently outperforming RNN and ResRNN baselines.