Radar Signal Sorting via a Dual-Layer Multi-Head Attention Broad Learning System

Abstract

Radar signal sorting is a fundamental task in passive electromagnetic sensing that separates pulses emitted by multiple radars from densely interleaved streams. Existing supervised deep learning methods often require large labeled datasets and iterative optimization, limiting their applicability in non-cooperative, latency-sensitive electronic-support-measure scenarios. This article proposes a dual-layer multi-head attention broad learning system (DMHA-BLS). Conventional random feature mapping is replaced by two attention blocks with fixed random-orthogonal query, key, and value projections. A residual branch maps first-layer attention features to the input dimension through closed-form ridge regression, preserving physical pulse attributes. The output weights are also solved analytically, and accumulated Gram statistics enable batchwise updates. On an 8000-pulse simulated stream, DMHA-BLS reaches 83.17% macro F1 with 200 samples and 91.88% with 8000 samples. It provides a clear advantage over the temporal baselines in the small-sample condition, while their performance gradually approaches DMHA-BLS as the number of labeled pulses increases. In a second 1000-pulse scenario with eight emitters and strongly overlapping PDW ranges, DMHA-BLS remains the top-performing method across the evaluated sample sizes. Full-data training requires 1.57 s, compared with 100–109 s for the evaluated deep networks, indicating a favorable trade-off between sorting performance and computational efficiency.