Design And Implementation Of Massive MIMO Detection Algorithm Based On Deep Learning

As one of the core technologies of the 5th generation wireless systems,Massive Multiple-input Multiple-output(Massive MIMO)improves the transmission rate and reliability of mobile communication system significantly.However,the increasing number of antennas of base station and users makes the communication module design more difficult.Deep learning has been widely applied in many fields as its good robustness and parallelism,and it provides a solid theoretical basis for its development in the field of mobile communication.The design of Massive MIMO detection algorithm with high reliability and low complexity is the key problem to be solved in the design of large-scale MIMO system.At present,there are two solutions to solve the problem of large-scale MIMO detection with deep learning.One is to add training parameters to the classical algorithm to improve the performance of the algorithm.In this paper,taking the second-order Richardson(SORI)linear detection algorithm as an example,through training the parameters in the algorithm,the iterative process is expanded into a neural network,and the SORI net detection algorithm is designed to solve the problem of slow convergence of the linear iterative algorithm in the related channels.In order to further improve the performance of detection algorithm,the method of training parameters is extended to the non-linear detection algorithm.Based on the message passing detection(MPD),this paper designs a lower complexity MPD detection network(LCMPD-Net).The simulation results show that the performance of LCMPD-Net with 128×16 in MIMO system with 64-QAM modulation scheme is improved by 0.7d B when the bit error rate(BER)is 10^-5.On the other hand,the detection algorithm based on the deep neural network(DNN)has better generalization and adaptability.In this paper,the dimension of the network is reduced.To solve the problem that the network is hard to support high order quadrature amplitude modulation,a Gaussian de-noising active function based on radial basis function neural network is proposed.Simulation results show that the Denoising Sparsely connected detection network(DSNet)is friendly to higher-order modulation detection in AWGN channel and correlation channel with 2/15 of training parameters compared with similar detection networks.