Enhancing Spectrum Sensing For 5G and LTE With Improved U-Net Architecture

Spectrum sensing for wireless communication has a crucial role in next-generation networks, enabling devices to identify and adapt to the growing demand for wireless connectivity. In this paper, we address challenges in spectrum sensing and propose an innovative deep network, namely Spectrum Sensing Network (SpecSenseNet) for over-the-air communication sensing. SpecSenseNet is an architecture based on U-Net, incorporating advanced state-of-the-art deep learning techniques to optimize sensing performance. These approaches, including depth-wise separable convolution, recurrent residual blocks, and Atrous Spatial Pyramid Pooling, significantly reduce the trainable parameters of the network. Indeed, the network's size is optimized while maintaining its prediction performance. As a result, Spec-SenseNet reduces effectively its trainable parameters to 7.8M and maintain the same global accuracy at 97.23% compared to U-Net. In terms of evaluation, we generate a synthetic signal dataset containing two signal types, which are fifth-generation new radio (5G NR) and long-term evolution (LTE) adapted to various levels of additive noise in the range of [0, 30] dB. Regarding simulation results, SpecSenseNet demonstrates impressive prediction performance, even in high noise density environments. Notably, reducing the number of trainable parameters enhances model's efficiency, generalizability, and ease of deployment.