Research On Signal Detection Based On Deep Learning In Massive MIMO Systems

Among the technologies included in the Fifth Generation Mobile Communication System(5G),massive Multi-Input Multi-Output(MIMO)is one of the most important technologies,which can greatly improve the performance of communication systems.However,when applying massive MIMO technology,in addition to paying attention to the advantages it brings,it is also necessary to solve the technical challenges brought about by the massive MIMO.Among these challenges,signal detection is an issue that urgently needs technological breakthroughs,which directly affects the transmission performance of communication systems.Since a large number of antennas are equipped at both transmitter and receiver,the signal detection algorithm in massive MIMO systems must not only ensure the detection performance,but also have low complexity,otherwise it is difficult to apply in practice.Deep learning has been gradually applied to many industries in recent years,and it has shown better performance than traditional methods.Generally speaking,deep learning incldes offline training and online prediction.Once deep learning algorithm adjusts the learnable parameters in the network to appropriate value through offline training,the processing of the input data by the deep learning algorithm in the online prediction is matrix multiplication and addition,which does not involve complex data processing operations.Therefore,designing signal detection algorithms in massive MIMO systems through deep learning is expected to achieve better detection performance with lower complexity.The main work of this thesis is as follows:(1)The uplink system model of massive MIMO is introduced,and two channel models commonly used in signal detection are summarized.Furthermore,the existing traditional signal detection algorithms are introduced,including maximum likelihood signal detection algorithm,linear signal detection algorithm and nonlinear signal detection algorithm,and the detection performance of each algorithm is analyzed by simulation experiments,which lays a foundation for the following research.(2)This paper studies the existing data-driven deep learning signal detection algorithms in massive MIMO systems,and then designs an improved signal detection algorithm using fully connected neural network(FCN)and convolution neural network(CNN).The improvements include: optimize network structure,use CNN to reduce training parameters,introduce learnable residual vectors and optimize the input and loss function.Simulation and analysis show that the proposed algorithm has higher detection performance,and the algorithm complexity and the number of training parameters are less than the algorithm before improvement.(3)Aiming that the number of parameters to be trained in data-driven deep learning algorithm is too large and the interpretability is not strong,this thesis proposes a model-driven deep learning signal detection algorithm.The design prototype of proposed algorithm is symmetric successive over-relaxation(SSOR)iterative signal detection algorithm.Specifically,each iteration in the SSOR signal detection algorithm is regarded as a layer in deep learning,and then the relaxation parameter in the SSOR algorithm is set to learnable relaxation vector,and the learnable piecewise decision(LPD)function is introduced to makes adaptive decision to improve detection performance.Simulation and analysis show that only a small amount of data and a small number of iterations are needed to complete the training of the proposed algorithm.In addition,without increasing the complexity of the algorithm,the proposed algorithm achieves better detection performance than the SSOR algorithm.