Research On Millimeter Wave Channel Estimation And Precoding Based On Deep Learning

Millimeter wave(mmwave)communication and large-scale multiple input multiple output(MIMO)technology are the key technologies of the next generation wireless communication network.The combination of them can effectively improve the system capacity,spectrum resources and transmission rate.Because the wavelength of millimeter wave signal is short and there is serious path loss,large-scale MIMO technology can effectively overcome the path loss.At present,the low-frequency communication system usually adopts the method of full digital precoding,and needs to configure a radio frequency(RF)link for each antenna.For millimeter wave large-scale MIMO system,if full digital precoding method is adopted,it will face a lot of hardware overhead.The hybrid precoding method can not only guarantee the achievable rate,but also reduce the cost of hardware.Therefore,this paper studies channel estimation and hybrid precoding for millimeter wave massive MIMO communication systems.The accuracy of channel estimation directly affects the performance of millimeter wave massive MIMO communication system.The traditional algorithm is affected by the large-scale array antenna,and the computational complexity is high.In order to solve this problem,we focus on the estimation of the path angle and path complex gain of the channel matrix,and propose a channel estimation algorithm based on deep learning to achieve end-to-end channel estimation.Specifically,the estimation of the path angle is modeled as a multi-label classification problem based on deep learning,and the path complex gain is solved directly using the least square method.The simulation experiments show that the proposed channel estimation algorithm based on deep learning has similar performance to the channel estimation algorithm based on OMP(Orthogonal Matching Pursuit),but its calculation time complexity is lower.To solve the problem of hybrid precoding in millimeter wave large-scale MIMO communication system,this paper proposes a novel hybrid precoding algorithm based on deep learning.The algorithm can directly predict the RF beamforming/combining matrix based on the beam scanning signal.Its network structure mainly includes two parts: encoder and hybrid beam prediction.Specifically,the encoder part takes the beam scanning signal as input,and obtains structural information from the input data through the two-layer convolutional neural network;while the hybrid beam prediction part learns how to map from the output of the encoder to the RF beamforming/combining matrix through the four-layer full connection network.Therefore,the neural network is an end-to-end neural network model that combines encoder and hybrid beam prediction.Compared with the traditional solution,this method can directly predict the RF beamforming/combining matrix,reducing the estimation of the channel matrix,and its achievable rate is close to the optimal achievable rate.