Deep Learning-based Channel Estimation Approach for 3D Massive MIMO Millimeter-Wave System in Time-Varying Environments

Abstract

Channel estimation in the high-speed mobile environment is a hot issue for 3D massive multiple input multiple output (MIMO) millimeter-wave (mmWave) system. In this environment, the channel has fast time-varying and non-stationary characteristics. And its time-domain correlation coefficient is a time-varying parameter, which makes it difficult for traditional channel estimation methods to capture the channel variations over time and achieve ideal channel estimation performance. In order to track the variations of time-varying mmWave channel and perform channel estimation, a deep learning-based channel estimation network is proposed in this paper. In the design of the deep neural network, a CNN+RNN network structure is used to learn the characteristics of channel. Considering that the 3D mmWave channel has sparsity in both spatial and frequency domains, a convolutional neural network (CNN) is used to extract the spatial-frequency domain features of the channel response. Afterwards a maximum pooling network is used to reduce the training parameters of the network. Finally, a recurrent neural network (RNN) is used to extract the time-domain correlated features of the channel response to estimate the channel. In offline training phase, channel state information (CSI) at pilot is firstly initialized by least squared method, and an incomplete channel matrix with estimated values only at the pilot point is obtained. In order to estimate the values at the data points and complete the CSI matrix, the incomplete CSI is fed into the network and standard high-speed channel data is used to train the network. Simulation results show that the proposed method can track the variations of channel characteristics in a high-speed mobile environment and achieve better performance in the online prediction phase compared to conventional methods.