A squeeze-and-excitation network for SNR estimation of communication signals

Abstract

Accurate signal-to-noise ratio (SNR) estimation is critical in wireless communication systems as it directly impacts system performance and the assessment of signal quality. Recent advances in deep learning-based SNR estimation have significantly improved estimation accuracy in low SNR conditions. This paper presents a novel deep learning approach that uses a power spectrum generated through overlapping segmentation as input to a neural network for SNR estimation. The performance of SNR estimation has been enhanced by integrating an augmented squeeze-and-excitation (SE) attention mechanism with a residual block fusion module, employing multiple residual structures, and deepening the network architecture. To validate the efficacy of this method, extensive simulation experiments were conducted under various scenarios, including additive white Gaussian noise (AWGN), Rayleigh, and Rician channel conditions. The results demonstrate that this method outperforms state-of-the-art techniques in high SNR environments and across diverse channel conditions. Furthermore, there is only minimal performance degradation under low signal-to-noise ratio conditions.