Cross-attention mechanism-based spectrum sensing in generalized Gaussian noise

Area of Science:

• Wireless Communications
• Signal Processing
• Machine Learning

Background:

• Spectrum sensing (SS) is crucial for cognitive radio (CR) networks to utilize idle spectrum efficiently.
• Accurate SS is challenging due to complex channel noise characteristics, especially non-Gaussian noise.
• Existing SS methods often rely on single features, limiting performance under diverse noise conditions.

Purpose of the Study:

• To propose a novel time-frequency cross fusion network (TFCFN) for enhanced spectrum sensing performance.
• To improve SS accuracy in non-Gaussian noise environments by overcoming single-feature limitations.
• To develop a robust feature extraction and fusion mechanism for reliable signal classification.

Main Methods:

• Utilized Gated Recurrent Units (GRU) for capturing long-term temporal dependencies in original signals.
• Employed Fast Fourier Transform (FFT) to extract frequency domain information.
• Applied Convolutional Neural Networks (CNN) for local spatial feature extraction in the frequency domain.
• Integrated time-domain and frequency-domain features using a cross-attention mechanism for dynamic fusion.

Main Results:

• The proposed TFCFN demonstrated superior detection ability compared to baseline methods in both Gaussian and non-Gaussian noise.
• TFCFN maintained lower computational complexity across different noise environments.
• Achieved a 10% probability of false alarm with over 90% probability of detection at -16dB SNR under GGD noise (shape parameter 0.5).

Conclusions:

• The TFCFN effectively fuses time-domain and frequency-domain features for robust signal classification in spectrum sensing.
• The proposed method significantly enhances SS performance, particularly under challenging non-Gaussian noise conditions.
• TFCFN offers a promising solution for reliable spectrum utilization in cognitive radio networks.