Research On MIMO Detection Algorithms Based On Deep Learning

With the development of the mobile Internet,people have put forward higher requirements on the bandwidth and speed of the wireless communication network.The upcoming 5G to meet these needs,one of the important technical support is the large-scale Multi Input Multi Output(MIMO)technology.An important part of MIMO systems is MIMO detection.In the MIMO detection algorithm,the Belief Propagation(BP)detection algorithm,i.e.the MIMO-BP detection algorithm,has attracted much attention due to its excellent performance.However,the MIMO-BP detection algorithm may not converge and have high complexity.In order to solve the problem of MIMO-BP detection algorithm,this paper proposes MIMO-BP detection algorithm model based on Back Propagation Network(BPN)model(BPN-BP detection model)and based on Recurrent Neural Networks(RNN)model(RNN-BP detection model).The Damped Belief Propagation detection algorithm completes the partial parameterization of the MIMO-BP detection algorithm by adding a fixed damping factor.The experimental results show that the partially parameterized MIMO-BP detection algorithm can improve the performance of the algorithm,but the traditional method is difficult to determine the optimal value of the added parameters.Therefore,this paper attempts to fully parameterize the MIMO-BP detection algorithm and use the powerful parameter optimization ability of the neural network to determine the optimal value of these parameters.This paper deploys a parameterized MIMO-BP detection algorithm based on the BPN model and the RNN model.When designing the BPN-BP detection model,the parameterized MIMO-BP detection algorithm is designed firstly,then the neural network topology of the BPN-BP detection model unit is designed.Then the output unit of the model is designed.Finally,the design is based on the maximum However,the criterion loss function,and added the L2-Regularization term to prevent over-fitting,and selected Mini-Batch Gradient Descent as the optimization method of the model.When designing the RNN-BP detection model,the RNN parameter sharing mechanism is used to further reduce the parameters in the detection model,making the model easier to train.In addition,new output units are designed so that the output of each model unit is more rationally involved in the operation of the loss function.At the same time,in order to accelerate the model training,accelerated training methods such as Batch Normalization algorithm,Adam optimization algorithm and Xavier parameter initialization method are introduced in the RNN-BP model.In this paper,two experiments were set up to verify the performance of the proposed model.The results of both experiments show that the proposed bit error ratio performance is better than the MIMO-BP detection algorithm and the Damped-BP detection algorithm,which also reduces the complexity of the MIMO-BP detection algorithm.And from the experimental results,it can be known that the RNN-BP detection model has better BER performance than the BPN-BP detection model compared with the BPN-BP detection model,and the RNN-BP detection model is also easier to train.