Novel deep learning framework for wideband spectrum characterization at sub-Nyquist rate

Abstract

Limited availability and high auction cost of the sub-6 GHz spectrum led to the introduction of spectrum-sharing in 5G networks. This demands base stations with the capability of automatic wideband spectrum characterization (AWSC) to identify available vacant spectrum and parameters (carrier frequency, modulation scheme, etc.) of occupied bands. Since WSC at Nyquist sampling (NS) is area and power-hungry and conventional statistical AWSC performs poorly at a low signal-to-noise ratio (SNR), we propose a novel sub-Nyquist sampling (SNS) based deep-learning framework. It is a single unified pipeline that accomplishes two tasks: (1) Reconstruct the signal directly from the sub-Nyquist samples, and (2) Identify the occupancy status and modulation scheme of all bands. The proposed non-iterative approach based reconstruction provides the occupancy status of all bands in a single forward pass leading to significant improvement in execution time over state-of-the-art iterative methods. In addition, the proposed approach does not need complex signal conditioning between reconstruction and characterization. We extensively compare the performance of our framework for a wide range of modulation schemes, SNR and channel conditions. We show that the proposed framework outperforms existing SNS based characterization and its performance approaches NS based framework with an increase in SNR without compromising computation complexity. A single unified deep-learning framework makes the proposed method a good candidate for reconfigurable platforms.