A Model-driven Deep Learning Signal Processing Scheme for OFDM System

Abstract

Deep learning (DL) has received more attention in physical layer communication. Channel estimation (CE) and signal detection (SD) constitute the key modules within the signal processing chain of orthogonal frequency-division multiplexing (OFDM) communication receivers. In this paper, we propose a model-driven deep learning scheme for the OFDM receiver, known as CSNet, which contains two modules, CE module and SD module. This structure maintains the block by block signal processing form in the OFDM system. In addition, we introduce the traditional linear algorithms minimum mean-squared error (LMMSE) and zero-forcing (ZF) to initialize the neural network of two modules, respectively, which improve the training speed of the sub-network and the generalization of the model. Compared with the traditional scheme and data-driven deep learning scheme, the model-driven DL receiver provides a more accurate channel estimation property. Simulation experiments demonstrate that the proposed scheme is robust in bit error ratio (BER) performance especially in the nonlinear case.