Deep Denoising and Transformer-Based 3D Spectrum Sensing

Leveraging the locational flexibility of unmanned aerial vehicles (UAVs), this letter investigates the 3D spectrum sensing problem in scenarios with random arrival and departure of primary signals. We propose a spectrum sensing framework based on deep denoising and Transformer architecture. Specifically, a deep denoiser is developed by integrating soft thresholding techniques with deep learning, while employing depthwise separable convolution to enhance real-time denoising performance. Furthermore, a redesigned Transformer architecture is proposed to achieve classification. Empowered by the synergistic use of the multi-head self-attention and improved squeeze-and-excitation attention mechanism, the switch feature of primary signals can be effectively learned. Simulations demonstrate the superiority of the proposed sensing framework.