Research On Channel Estimation And Channel Feedback Based On Deep Learning

With the rapid development of wireless communication technology,5G is gradually commercialized and entered people's lives.In 5G systems,the use of massive multiple input multiple output(Massive MIMO)technology can effectively eliminate inter-user interference and greatly improve spectrum efficiency and system capacity.However,due to the large increase in the number of antennas in Massive MIMO systems,the complexity of traditional channel estimation,channel feedback,and signal detection technologies has been increased significantly,so there are some insurmountable problems,which lead researchers to study more effective solutions.In recent years,with the continuous development of deep learning technology,its application in the field of wireless communication has become more and more extensive,providing new ideas for better solving the above problems.Based on the existing research results,this thesis focuses on the application of deep learning in channel estimation and channel feedback.This thesis introduces the basic knowledge and relevant theory of deep learning technology firstly.Then based on in-depth study of existing algorithms,the pilot-assisted channel estimation algorithms in orthogonal frequency division multiplexing(OFDM)systems and the channel state information(CSI)compression feedback algorithms in frequency division duplex(FDD)Massive MIMO systems are studied by introducing the deep learning techniques.The main results obtained in this thesis are as follows:1.Regarding the channel estimation technology in OFDM system,this thesis proposes a deep learning-based channel estimation algorithm in comb pilot mode based on the analysis of traditional channel estimation algorithms and existing deep learning-based channel estimation algorithms.Considering the channel estimation part of the pilot position and channel interpolation part in the traditional algorithm jointly,the algorithm builds a network model named Comb-CEnet through fully connected layer neural networks to complete the work of both parts at the same time,thus completely replacing the traditional channel estimation module.Simulation results show that,compared with Comb-CEnet trained under low signal-to-noise ratio,Comb-CEnet trained under high signal-to-noise ratio has better performance in generalization of signal-to-noise ratio;then compared with traditional LS algorithm and LS-DFT algorithm,the Comb-CEnet trained under high signal-to-noise ratio can obtain preferable channel estimation quality and bit error rate performance when running online,and can also exhibit outstanding robustness when increasing the pilot interval.2.Regarding the downlink CSI feedback technology in FDD Massive MIMO system,this thesis proposes a CSI compression feedback algorithm based on deep autoencoder.Taking the existing convolutional autoencoder-based CSI compression feedback algorithm as the reference and the spatial correlation of the channel as the breakthrough point,and considering the difference in accuracy between image reconstruction and digital recovery,the algorithm uses fully connected layer neural networks to build a deep autoencoder(including encoder and decoder),in which the CSI is compressed by reducing the dimension of the encoder output layer.The algorithm places the trained encoder and decoder on the receiver and the transmitter respectively,where the receiver uses the encoder to compress the CSI and feedback it to transmitter,and the transmitter uses the decoder to reconstruct the CSI from the feedback information.The simulation results show that,compared with the existing convolutional autoencoders under different compression ratios in different channel scenarios,the deep autoencoder can obtain significantly improved CSI reconstruction quality and thus better beamforming gain.In terms of complexity,the training time for deep autoencoder is slightly higher than existing convolutional autoencoders,but there is almost no difference on running time online.