Study On Channel Estimation Of MmWave Massive MIMO Based On Deep Learning

As a key technology of 5G,millimeter-wave(mm Wave)massive multiple-input multiple-output(MIMO)can address the severe path loss problem of millimeter wave communication and enable more antennas to be deployed in a limited space,by combining the advantages of mm Wave communication technology and massive MIMO technology.Efficient and accurate channel estimation algorithms are crucial to the design of mm Wave massive MIMO systems.However,in mm Wave massive MIMO systems,as the dimension of the channel increases,the impact of noise on the performance of channel estimation cannot be ignored.Therefore,removing noise to reduce estimation error after analyzing its impact is a challenge in enhancing the reliability of channel estimation.In addition,equipping a massive number of antennas on the base station in mm Wave massive MIMO systems will result in high hardware cost and power consumption if the number of radio-frequency(RF)chains is the same as that of antennas.Due to the limited scattering at mm Wave frequencies,the channel matrix has low rank and a few dominant beams can be selected to reduce the number of RF chains.However,in mm Wave massive MIMO systems with a large antenna array and a limited number of RF chains used,how to achieve low complexity and high reliability of channel estimation in the beam domain is a challenging problem that needs to be solved.With its powerful feature extraction ability and efficient data mining capability,the deep learning algorithm provides a new approach to reducing the complexity and improving the reliability of channel estimation algorithms.Therefore,this paper applies the deep learning algorithm to mm Wave massive MIMO channel estimation,and the specific research contents are as follows:1.Based on the analysis of the impact of noise on the performance of channel estimation in the beamspace mm Wave massive MIMO system,a noise feature denoising network(NFDNet)based channel estimation algorithm is proposed.The basic principle is to use the received signal and its noise features extracted through convolutional and pooling layers as the input of a denoising convolutional neural network(Dn CNN),and output the estimated noise matrix.Then,remove the estimated noise matrix from the received signal to obtain the estimated channel matrix.Simulation results show that the channel estimation algorithm based on NFDNet outperforms Dn CNN in terms of normalized mean squared error(NMSE)performance.2.Based on the analysis of the impact of noise on the performance of channel estimation,and the sparsity of mm Wave massive MIMO channel,this paper proposes a beamspace mm Wave massive MIMO channel estimation algorithm based on the two-step noise learning network(TNLNet).The basic principle is to use the received signal and noise features extracted through convolutional and pooling layers,and utilize the sparsity of the channel matrix to reconstruct the channel matrix into four sub-matrices using reversible downsampling,reducing complexity and improving training and testing efficiency.The results show that the TNLNet not only achieves better NMSE performance,but also has a wider range of signal-to-noise ratio applicability in a single training model than the fast flexible denoising convolutional neural network,enhancing its practicality.3.To further reduce the hardware cost and power consumption of mm Wave massive MIMO systems,this paper proposes a channel estimation algorithm based on the non-iterative reconstruction network(NIRNet)in the case where the antenna array is large but the number of RF chains is limited.The basic principle is that the proposed reconstruction network uses convolutional layers to perform non-iterative reconstruction of the channel matrix,which reduces complexity compared with existing learned approximate message passing(AMP)networks.On the basis of considering the prior information of the transmitted data,a selection network is trained to obtain a learning-based selection matrix(LSM),which is used to improve the accuracy of channel reconstruction.Compared with a random Bernoulli selection matrix,the LSM lays the groundwork for improving the accuracy of channel reconstruction.Finally,based on the analysis of the impact of noise on the performance of channel estimation,a denoising network is used to further improve the estimation accuracy.The results show that in the case of equipping only a small number of RF chains,compared with other algorithms such as learned denoising-based AMP algorithm and fully convolutional denoising AMP algorithm,NIRNet achieves better NMSE and achievable sum-rate(ASR)performance with a lower complexity,meeting the demand for low complexity and high reliability performance.