Deep Unfolded Sparse Refinement Network Based Detection in Uplink Massive MIMO

Arijit Datta

doi:10.1109/TVT.2022.3166399

2022-06-01

PRE

Arijit Datta

Aneesh Nema

Vimal Bhatia *

DOI：10.1109/TVT.2022.3166399

原文链接

原文求助

摘要

En 中文

Massive multiple-input multiple-output (mMIMO) is a promising technique to realize the ever-increasing demand for high-speed data, quality of service (QoS), and energy efficiency for 5G and beyond wireless systems. However, the increased number of users in mMIMO systems significantly affects performance of the existing approximate matrix inversion based and matrix inversion less iterative symbol detection techniques. Conventional detection algorithms cannot learn the inter-relations of input-output parameters based on available data without having specific mathematical models of communication scenarios. Moreover, existing deep learning (DL) based symbol detection models lack in-network compression, resulting in large training time and high computational load while expected to be deployed in a low latency communication system. In this article, a sparse refinement architecture is proposed for symbol detection in uplink mMIMO. The proposed DL architecture requires less trainable parameters as compared to a conventional fully connected detection network and refines the estimated symbol vector in each layer. Convergence of the proposed symbol detection technique is analytically justified. An expression for the approximate upper bound on the BER is derived which is supported by simulations. The obtained results prove viability of the proposed symbol detection model as compared to the several existing state-of-art uplink mMIMO detection techniques, in terms of superior the error performance and low computational complexity.

Keyword:

Symbols

Computational modeling

Uplink

Sparse matrices

Mathematical models

Detectors

Training

Computational complexity

Deep learning

Maximum likelihood decoding

Massive MIMO