Online Deinterleaving of Frequency-Agile Radar Emitters via Jump-Set Representation and Adaptive Scoring

Passive electronic intelligence (ELINT) and electronic warfare receivers must deinterleave pulse descriptor word (PDW) streams in real time: each pulse is labelled on arrival, unsupervised and irrevocably, at bounded cost. Frequency- and PRI-agile emitters defeat the single-Gaussian models of mainstream online sorters: an agile emitter hops across (RF, PRI) cells, so, under a discrete-grid mixture model, its mean and covariance are not sufficient for membership, and moment-matched covariances of distinct emitters overlap. We cast sorting as membership inference under a mixture of probability mass functions (PMFs) on this grid, represent each emitter by its jump sets and joint PMF, and score each pulse by an approximate log-posterior at O(K^) cost with support-bounded storage. On the Turing Synthetic Radar Dataset, against HDBSCAN, leader clustering, Gaussian-mixture stream clustering and Bayesian merging, the strictly online configuration averages a V-measure of 0.833 over eight scenarios, +0.211 above the strongest baseline, and reaches 0.827 under pure agility, the only online method above 0.5; a guarded post hoc merge raises the average to 0.927, and a windowed optimal-transport sorter reaches 0.670 on agile scenarios. Given a correct partition and each emitter’s PRI, a shared-RF ablation credits 0.235–0.472 gains to the joint distribution term; no compared method separates such emitters from a blind cold start.