Research On New Algorithms For Channel Estimation In Massive MIMO

In massive Multiple Input Multiple Output(MIMO)systems,the pilot overhead required by traditional linear channel estimation algorithms is quite large,and the complexity of algorithms based on Least Sqaure(LS),Minimize Mean Square Error(MMSE),and compression perception techniques will also increase.To overcome the above problems,the sparse characteristics of the massive MIMO channel can be used to reduce the effective dimension of the channel matrix,and the deep learning technology can be combined with the compression perception technology to solve the channel estimation problem in the massive MIMO system.Firstly,an channel estimation algorithm for Convolutional Reconstruction Network(CSNet)is proposed to combine deep learning technology with compression perception technology.The main idea of the algorithm is to combine a data-driven approach in deep learning with a compression-aware approach,and the CSNet network will replace the traditional refactoring method,and the final layer of the network adds judgment to make the conformity system sparse feature.At the same time,discard the imaginary part of the estimated value of the final channel to obtain the final estimated channel.Simulation results show that the algorithm has good performance in mean square error(MSE)performance and false bit rate performance.Secondly,in data-driven deep learning channel estimation,it is difficult to generate a more realistic channel matrix due to the loss of information in the continuous layer of the network,which does not significantly improve the estimation performance.To improve performance,an improved generative Adversarial Networks(GAN)channel estimation algorithm is proposed.In this algorithm,a random quantization method is introduced to improve the input of the GAN network and make the input data more realistic.By introducing penalty terms in the generator and discriminator,a new optimization objective function is generated,so that the network optimization is oriented correctly.Adversarial training of generators and discriminators to predict more realistic channels.Numerical simulation results show that from the perspective of Normalized Mean Squared Error(NMSE),the method significantly improves the accuracy of channel estimation of massive multiple input multiple output systems.Finally,a new deep convolutional learning network(LDAMP-Net)based on Denoising-based Approximate Message Passing(DAMP)is proposed by combining model-driven and data-driven.The Network is an iterative algorithm,and each iteration’s LDAMP-Net block consists of a learnable LDAMP(Learned-DAMP)Network and a deep Convolutional learning Network(CNN).Firstly,a new denoising function is proposed by changing the denoising device in the original LDAMP network,and the initial approximate solution is obtained by using the sparsity of the channel matrix in the beam space.The deep convolutional learning network is added to reduce the influence of channel noise,and the estimated value of the LDAMP network is further optimized.To sum up,data-driven methods in deep learning and model-driven methods are used for channel estimation in this paper.Firstly,data-driven methods in deep learning are combined with traditional compressed sensing technology to obtain a CSNet network for channel estimation.To solve the problem that it is difficult to generate a more realistic channel matrix due to the information loss of continuous layer in a data-driven network,an improved GAN network is used for channel estimation.Finally,a model-driven and data-driven LDAMP-Net network is proposed to estimate the channel.