Applications Of Deep Learning In The Physical Layer Of Wireless Communication

The fifth generation of wireless communication technology(5G)needs to support different needs in various application scenarios,so a huge network capacity that is 1000 times of 4G,a millisecond-level delay,ultra-dense large-scale networking,and a lower energy efficiency ratio are required.To meet these needs,new technologies such as massive MIMO systems and millimeter waves have been core technologies of 5G.However,lots of limitations in complex application scenarios of 5G are revealed by the existing traditional communication theory,such as difficulty in modeling complex channel scenarios and difficulty in system optimization.In recent years,the rapid development of deep learning and artificial intelligence technology has gradually achieved successful applications in the field of wireless communication because of its high accuracy,robustness,and no need of complex feature extractions.This paper focuses on the physical layer of wireless communication systems,and discusses the applications of deep learning in the direction of arrival estimation,MIMO detection,and OFDM receiver.Firstly,the paper discusses the DOA estimation algorithm based on convolutional neural networks.This paper presents a DOA estimation algorithm for one-bit MIMO systems based on convolutional neural networks.The system model,the data processing scheme,the network structure and the network optimization algorithms are elaborated clearly.Numerical results show that the convolutional neural network is more accurate and robust than the traditional MUSIC algorithm and ESPRIT algorithm.Furthermore,the computation complexity of the convolutional neural network is much lower than the MUSIC algorithm,and the ability of parallel computing of the neural network can greatly speed up the calculation.Secondly,this paper studies the signal detection problem of the massive MIMO system.Based on the Det Net and Deep Unfolding theory,the convolution operation is introduced into the iterative expansion of maximum likelihood detection,and a new convolutional neural network detection model called 'Conv Det Net'is proposed.The paper derives the recursive formula and network structure of Conv Det Net.In addition,in view of saturation of accuracy and difficulty of convergence when the network are going deeper,several solutions such as residual structure and pre-training schemes are proposed.Simulation results show that the performance of the Conv Det Net is better than ZF detection,MMSE detection and Det Net,and the performance advantage is more obvious under large-scale antennas and random channels.And as the network going deeper,the BER performance of the Conv Det Net using residual structure and pre-training schemes also improves.Finally,this paper studies the applications of deep learning in OFDM system receivers,and proposes a novel convolutional receiver network called 'CRNET' based on the OFDM system model.The channel estimation module and the channel equalization module are re-designed using neural network models respectively.So the convolutional channel estimation network Est Net and the convolutional channel equalization network Equ Net are constructed.The two networks together form the convolutional receiver Network CRNET.The recursive formula and structure of the CRNET are introduced in detail.The numerical results show that CRNET has the same performance as the traditional scheme with MMSE estimation and single-tap equalizer when the cyclic prefix is sufficient.However,when the cyclic prefix is insufficient or there is no guard interval,the performance of CRNET is far beyond the traditional scheme as well as the dense neural network.CRNET has stronger robustness and generalization ability over the inter symbol interference.