Research On DOA Estimation Algorithm Based On Neural Network In Massive MIMO System

As a core technology in the 5G mobile communication systems,the performance of massive MIMO systems depends on accurate DOA estimation.Most widely used DOA estimation algorithms assume that the signal model is a point source.However,compared with point sources,the incoherently distributed source is more suitable for real communication environments.Therefore,this thesis investigates the DOA estimation algorithm for incoherently distributed source in massive MIMO systems.Since the huge number of antennas leads to an increase in the dimensionality of the covariance matrix of the received signals,which consequently raises the computational burden of the eigen decomposition,high-precision subspace-based algorithms have very high computational complexity in massive MIMO systems.In addition,the actual antenna arrays are usually with various imperfections,in which case the estimation performance of the subspace-based algorithms is greatly reduced.To address the aforesaid problems,neural network-based DOA estimation algorithm is proposed in this thesis.Firstly,an autoencoder-deep neural network multilayer classifier(AE-DNNMC)algorithm is proposed in this paper.The input of the network is first decomposed into different subregions by the autoencoder;then the DOA estimation is achieved by a multilayer classifier composed of multiple deep neural networks.The simulation experiments demonstrate that the algorithm can achieve DOA estimation in the presence of array imperfections;secondly,to address the problem of accuracy degradation caused by overfitting in the AE-DNNMC algorithm,the DNNMC is improved into RD-DNNMC by adding regularization and dropout method in the hidden layer to suppress overfitting;finally,the AE-DNNMC algorithm is improved into an autoencoder-deep convolutional network multilayer classifier(AE-RD-DCNMC)algorithm.The performance of the algorithm is further improved by adding a certain number of convolutional layers,and using the convolutional layers to process the relationship between the real and imaginary parts of the input elements of the multilayer classifier.Simulation experiments verify the effectiveness of the AE-RD-DCNMC algorithm in the presence of array imperfections.