Virtual PRI Assisted Radar Signal Sorting with A Dual-Path Network for Fine-Grained Cluster Fusion

Pulse repetition interval (PRI) estimation remains critical yet challenging in radar signal sorting (RSS), particularly under complex electromagnetic environments and diverse radar operating modes. Traditional methods relying on direct PRI estimation often fail to handle complex PRI modulation patterns and unknown radar modes, while exhibiting sensitivity to missing pulses. In light of this challenge, we introduce the concept of virtual PRIs, a representation derived from fine-grained clustering of pulse descriptor words (PDWs). Virtual PRIs serve as statistically robust surrogates for true PRIs that preserve temporal patterns even under complex PRI modulation and missing pulses, while eliminating the need for explicit PRI estimation. This enables better handling of complex PRI modulation without prior knowledge of radar operating modes. To address the over-clustering issue introduced by fine-grained clustering, where the number of obtained pulse clusters far exceeds the actual number of radar emitters, we introduce a Dual-Path network to merge fine-grained pulse clusters by jointly processing stacked virtual PRI sequences and individual virtual PRI pulses for global and local pulse feature extraction. The accuracy and robustness of our virtual PRI-assisted RSS schemes are validated on datasets incorporating complex PRI modulation, missing pulses, and unknown radar modes.