DFN-YOLO: Detecting Narrowband Signals in Broadband Spectrum.

Area of Science:

• Wireless communication
• Signal processing
• Machine learning for signal detection

Background:

• Efficient spectrum utilization is crucial for modern wireless communications.
• Detecting narrowband signals in broadband environments, especially at low signal-to-noise ratios (SNR), is challenging due to complex time-frequency features and noise.
• Existing object detection models struggle with the nuances of broadband spectrum sensing.

Purpose of the Study:

• To develop a robust signal detection model for blind signal detection in broadband scenarios.
• To enhance the extraction and integration of channel features for improved signal identification.
• To achieve higher detection accuracy under low-SNR conditions in complex wireless environments.

Main Methods:

• Introduction of the Deformable Feature-Enhanced Network-You Only Look Once (DFN-YOLO) model.
• Integration of a Deformable Channel Feature Fusion Network (DCFFN) with a deformable attention mechanism.
• Optimization of the loss function using Focal Scaled Intersection over Union (Focal_SIoU).
• Construction and utilization of a dedicated signal detection dataset for evaluation.

Main Results:

• DFN-YOLO achieved a mean average precision (mAP50-95) of 0.850 on broadband time-frequency spectrograms.
• The model significantly outperformed mainstream object detection models, including YOLOv8.
• Maintained an average time estimation error within 5.55×10-5 s and provided preliminary center frequency estimation.

Conclusions:

• DFN-YOLO demonstrates superior performance for blind signal detection in broadband environments.
• The model's ability to handle low-SNR conditions and complex features offers significant advantages.
• Findings have substantial implications for both civilian and military wireless communication applications.