| In recent years,deep learning has made great breakthroughs in the field of wireless communication.Deep learning for Massive MIMO signal detection can achieve high computational parallelism.Deep learning methods applied in the field of communication can be divided into two categories: deep learning methods based on data-driven and deep learning methods based on model-driven.In this paper,the two methods are respectively introduced into the Massive MIMO signal detection problem.Ms Net(Multisegment Mapping Network)detection based on data-driven and RGNet(RIGS-based Deep Learning Network)detection based on model-driven are proposed.Firstly,Ms Net detection is proposed by combining data-driven deep learning method with Massive MIMO signal detection.The network structure of Ms Net detection is formed by the network iteration of the same layer.Each layer of network is divided into linear and nonlinear modules.The linear module cascaded the known information and added trainable variables to perform linear estimation of the sent signal.In order to improve the flexibility and detection performance of the network in high-order modulation scenarios,a Multisegment mapping activation function sig S is designed for nonlinear modules.In addition,in order to accelerate the convergence speed of the network under high SNR,the overall structure of the network adopts the residual structure with variable residual coefficient.Simulation results show that the constructed of Ms Net detection has the characteristics of simple structure,fast convergence speed and low complexity.Secondly,RGNet detection is proposed by combining the model-driven deep learning method with Massive MIMO signal detection.Firstly,a Hybrid method of Richardson and Gauss-Seidel Algorithm was proposed based on Richardson and Gauss-Seidel Algorithm.The RIGS algorithm combines these two algorithms to achieve faster convergence.However,the performance of this detection algorithm is limited in related channel scenarios.To improve robustness,this paper extended the RIGS algorithm to networks,adding learnable parameters in each iteration and introducing sig S activation functions to significantly improve detection performance.Simulation results show that the RGNet proposed in this paper has low detection complexity,simple and fast training process,and can achieve excellent detection performance in Rayleigh fading channel and spatial correlation channel. |
