Detect, Extract, Recognize: Wideband RF Signal Detection and Modulation Recognition with Over-the-Air Validation

Abstract

Radio dynamic zones (RDZs) and other dynamic spectrum-sharing regimes require sensing systems that can detect and characterize previously unknown emitters. While automatic modulation recognition (AMR) is an important component of RF spectrum sensing capability, nearly all AMR methods assume the signal of interest has already been detected, isolated, and handed to the classifier which is specifically not a valid assumption for RDZs. In this paper, we present an integrated RF analysis framework that combines signal detection and modulation recognition, with over-the-air evaluation using software-defined radios. A YOLOv11 detector localizes signals in time-frequency spectrograms and estimates their temporal and spectral boundaries. These regions are mapped to the corresponding complex I/Q data, where individual signals are extracted and classified using a lightweight one-dimensional attention-based model. Using TorchSig-generated data with 57 waveform classes over 0–30 dB SNR, the detector achieves 99.2% mAP@0.5, while the classifier achieves 62.1% fine-grained accuracy and approximately 92% modulation-family accuracy. Over-the-air evaluation shows fine-grained accuracy increasing from 23.7% at 2 dB to 65.2% at 25 dB, with family-level accuracy reaching 98.0%. Overall, the proposed framework provides a unified pathway from RF signal detection to modulation recognition and supports more complete automated RF spectrum analysis.