Deep Neural Network Aided Low-Complexity MPA Receivers for Uplink SCMA Systems

Abstract

Sparse code multiple access (SCMA) has exhibited superiority in spectrum efficiency, which is particularly essential in the Internet of Things (IoT) system where a steadily increasing number of device connections are accommodated. However, the computational complexity of the conventional message passing algorithm (MPA) for the multiuser SCMA detection increases exponentially with the degree of resource nodes (RNs). To address this issue, two low complexity MPA schemes are proposed by utilizing the sparse feature of codewords. First, a sorted MPA (SMPA) detector is introduced to reduce the message exchanging from RNs to variable nodes (VNs) by dropping the redundant superposed constellation points outside a belief interval. Next, in order to further speed up the sorting process of the Euclidean distances between the received signal and codeword combinations, a deep neural network aided MPA (DNNMPA) is proposed, in which, the DNN behaves as a function approximator to generate the belief interval and operates in parallel with the initialization procedure before iterative message passing. Simulation results illustrate that the proposed SMPA and DNNMPA detectors significantly reduce the computational complexity of the conventional MPA one, but with comparable decoding capabilities, for the uplink SCMA system.