What question did this study set out to answer?

The aim is to develop a dataset generation framework to improve RF signal detection and classification in complex environments with multiple emitters.

April 12, 2026

Blind Detection and Classification of RF Signals in Dense Multi-Emitter Environments: A Deep Learning Dataset Generation Framework

Key Points

The aim is to develop a dataset generation framework to improve RF signal detection and classification in complex environments with multiple emitters.
Utilized Hardware-in-the-Loop (HIL) techniques to create a single source dataset of characteristic RF signals.
Synthetically generated dense RF mixtures with randomized offsets to avoid interference.
Transformed RF mixtures into spectrogram images for signal detection tasks.
Trained YOLO-based detectors (YOLOv8, YOLOv9, YOLOv10) on the generated dataset.
Conducted Exploratory Data Analysis (EDA) to evaluate dataset distribution and balance.
YOLOv8 achieved the highest mean Average Precision (mAP50) of 86.45% and mAP50-95 of 68.56%.
YOLOv9 showed marginally better mid-epoch performance but was ultimately outperformed by YOLOv8.
YOLOv10 lacked inference stability in dense RF environments compared to YOLOv8.
Dataset demonstrated an equitable distribution across eleven signal types.

Abstract

In dense and contested Radio Frequency (RF) environments, conventional signal detection and classification techniques often fail due to overlapping transmissions, the absence of a known Signal of Interest (SoI), and the presence of unknown emitters. To address these challenges, this paper proposes a dataset generation framework that uses Hardware-in-the- Loop (HIL) based single source dataset of relevant characteristic signals to synthesize dense RF mixtures with randomized time and frequency offsets without co-channel interference or enforcing power dominance or structured interference ratios. These RF mixtures are transformed into spectrograms of composite RF scenes as images, and YOLO (You Only Look Once) based detectors (YOLOv8, YOLOv9, YOLOv10) are trained on the generated dataset to detect, localize, and classify multiple coexisting signals. The Exploratory Data Analysis (EDA) is used for evaluating the dataset's structure and balance, confirming an equitable distribution of different eleven signal types with averaging 9% per type, under different RF mixing situations, including single-source, dual-source, and triple-source signals, which achieved 33% for each mixture scenario in the dataset. Experimental analyses show that YOLOv8 achieves the highest mAP50 (86.45%) and mAP50-95 (68.56%), demonstrating superior performance in dense environments, while YOLOv10 has less inference stability across dense RF scenarios compared to YOLOv8. YOLOv9 exhibits robust overall performance, achieving marginally superior results compared to YOLOv8 at mid-epochs, however is ultimately outperformed by YOLOv8 in both mAP50 and mAP50-95 metrics. Comparative analysis among these model versions demonstrates that the equilibrium among model size, complexity, and computing cost under-scores the trade-offs between performance and efficiency. Depending on the application, these trade-offs will guide model selection, with YOLOv8 chosen for accuracy-sensitive tasks and YOLOv10 favored for real-time situations where speed is essential. The proposed dataset generation framework sets the foundation for future research in open-set RF detection, multi-sensor fusion, and adaptive spectrum dominance strategies.

Bookmark

Cite This Study

Aboelazm et al. (Fri,) studied this question.

synapsesocial.com/papers/69db37ca4fe01fead37c5d26 https://doi.org/https://doi.org/10.1177/18758967251411395

Bookmark