Machine Learning Physical-Layer Receivers for DFT-s-OFDM

Abstract

In this work, we propose two alternative machine learning (ML) based physical-layer receivers for discrete Fourier transform (DFT) precoded multicarrier systems. The receivers are designed to provide reliable soft bit estimates under two coexisting transmitter impairments, namely, power amplifier nonlinear distortion and oscillator phase noise. Considering data modulations up to 256-QAM, realistic 5G NR uplink evaluations at 28 GHz demonstrate large performance gains compared to reference receivers, especially when the ML receiver has trainable layers both before and after the inverse DFT. Furthermore, compared to a baseline linear receiver, we show an uplink coverage increase of over 20% with the proposed fully learned receiver.