Research And Optimization Of MIMO Signal Detection Based On Deep Learning

With the development of the Internet in recent years,there are more and more mobile terminal devices,and the number of access required shows an exponential growth.At the same time,users also have a higher demand for communication systems.High reliability,low delay and high spectrum efficiency are all needed by users.The Multiple-Input Multiple-Output(MIMO)can meet user needs,so MIMO has become one of the key technologies in the 5G.However,the increasing number of antennas makes the signal detection more and more complex,and the combination of hundreds and thousands of antennas is a huge challenge to the signal detection of MIMO systems.Therefore,it is very important to study the signal detection algorithms with good detection performance under low complexity.Recently,Deep Learning(DL)has performed well in semantic recognition,AI intelligent cloud computing and behavior recognition,which fully demonstrates the advantages of DL.Inspired by this,we apply DL to the received end of MIMO wireless communication system and combine DL with MIMO detection to explore and find better signal detection algorithms.In addition,Tensorflow,an advanced DL framework,is used to complete network training and verification.The main content of this paper is the improvement and optimization of the signal detection of MIMO wireless communication systems.In this paper,the characteristics of MIMO system and the realization of detection algorithm at the detection end are introduced and analyzed in detail.In view of the high complexity of traditional detection algorithm and the insufficiency of detection effect,MIMO detection networks based on DL are proposed.The contributions of this paper are mainly composed of two parts.The first part is to propose a new signal detection network,named GS-Net,based on the deep unfolding and Model-Driven.This part is described in detail in the fourth part of the paper The second part improves and optimizes the traditional conjugate gradient detection algorithm model,and proposes a conjugate gradient detection network based on DL,called DLCGNET.This part is mainly focused on the fifth chapter of this paper.The main innovation points of this paper are as follows:(1)In order to achieve better detection performance,traditional Gauss-Seidel detection algorithm requires a large number of iterations,which needs a high detection complexity.In order to reduce the complexity of detection,this paper proposes a novel Gauss-Seidel detection network based on deep unfolding and Model-Driven,which is named GS-Net.GS-Net introduces two trainable parameters in the traditional Gauss-Seidel detection algorithm,which are used to adjust the proportional relation between the two items involved in optimization.The value of parameter changes with the change of network training times.The Gauss-Seidel detection model with introduced parameters is expanded into a deep neural network by using the technology of deep unfolding,and each layer network represents an iterative process of the traditional Gauss-Seidel detection algorithm.At the same time,the output function of the network adopts the soft decision method.Based on Tanh function,parameters are added to control the output value to gradually approximate to +1 and-1,so as to reduce the error rate of the decision.Finally,the detection performance of the proposed GS-Net is analyzed through simulation experiments,and the detection performance of GS-Net is compared with that of traditional detection algorithms.Simulation results show that the proposed GS-Net has better detection performance than the traditional detection algorithms.(2)Because the iterative algorithm of MIMO signal detection requires a large number of iterations to obtain a good detection effect,the complexity of signal detection is virtually increased.In theory,the conjugate gradient detection algorithm can find the real transmit signal in N steps,but it involves complex matrix operation in the calculation of step length,which increases the complexity of detection.In order to reduce the complexity,this paper proposes a conjugate gradient detection network based on DL,named DLCG-Net.The network sets the two steps that need to be calculated as trainable parameters,which no longer need complex matrix operations,but are directly obtained by network training.In this way,the complex matrix operation is avoided and the complexity of detection algorithm is reduced.In addition,in the design of DLCG-Net,the constraint of traditional conjugate gradient algorithm is fully considered,the dot product of each update direction is taken as a part of the loss function,and a parameter ? is introduced to compromise the detection accuracy and complexity.Finally,the detection performance of DLCG-Net is analyzed by computer simulation.Simulation results show that the proposed DLCG-Net detection network has better detection performance and lower detection complexity than traditional detection algorithms.At last,this paper summarizes the research contents,and points out the areas to be improved in this paper and the further research work to be carried out in the future.