MFE-YOLO for Overlapping LPI Radar Signal Modulation Recognition and Parameter Extraction

Accurate modulation recognition and parameter extraction of overlapping Low Probability of Intercept (LPI) radar signals in dense electromagnetic environments remain significant challenges. Traditional methods and existing machine learning approaches based on multi-label classification or semantic segmentation often lack the capability to characterize individual components, especially multiple instances of the same modulation type. This paper proposes a Modulation Feature Enhancement-YOLO (MFE-YOLO) model for overlapping signal modulation recognition and parameter extraction in dense electromagnetic environments. By utilizing the Smoothed Pseudo Wigner-Ville Distribution (SPWVD) and treating each signal as a detectable object, MFE-YOLO simultaneously predicts modulation types and bounding boxes, facilitating the extraction of time width and bandwidth for each component. Built upon YOLOv11, the proposed method integrates a Gated Enhancement of Modulation Feature (GEMF) module to refine modulation characteristics. Extensive simulations involving eight modulation types across diverse signal-to-noise ratios (SNRs) and power ratios (PRs) demonstrate MFE-YOLO’s superiority over baseline methods. The proposed framework achieves significantly higher accuracy in both recognition and parameter estimation, showing strong robustness, especially in low-SNR scenarios.