Leveraging Time-Frequency Distribution Priors and Structure-Aware Adaptivity for Wideband Signal Detection and...

Area of Science:

• Signal Processing
• Machine Learning
• Artificial Intelligence

Background:

• Wideband signal detection and recognition (WSDR) is crucial for spectrum monitoring.
• Current deep learning methods for WSDR often overlook time-frequency prior information and signal structural features.
• Optimization is needed to improve the accuracy and robustness of existing WSDR techniques.

Purpose of the Study:

• To propose a novel model, TFDP-SANet, for enhanced WSDR.
• To incorporate time-frequency distribution priors and structure-aware adaptivity into the WSDR model.
• To improve the detection and recognition accuracy of wideband signals.

Main Methods:

• Developed TFDP-SANet model incorporating Strip Pooling Module (SPM) and Coordinate Attention (CA) for feature extraction.
• Utilized an adaptive elliptical Gaussian encoding strategy for heatmap generation to improve center-point localization.
• Implemented a Time-Frequency Clustering Optimizer (TFCO) during inference to refine bounding box predictions using prior information.

Main Results:

• The proposed TFDP-SANet model demonstrated superior performance on the WidebandSig53 (WBSig53) dataset.
• Ablation and comparative experiments confirmed the effectiveness of the integrated modules (SPM, CA, TFCO).
• The model showed significant improvements in accuracy and robustness for WSDR tasks compared to existing methods.

Conclusions:

• TFDP-SANet effectively addresses limitations in current WSDR approaches by integrating time-frequency priors and structural awareness.
• The novel components significantly enhance the model's ability to capture signal characteristics and improve localization accuracy.
• The findings suggest a promising new direction for advanced wideband signal analysis and recognition.