DL-PC algorithm for partially-connected hybrid precoding in massive MIMO systems

Abstract

In hybrid precoding, a partially-connected (PC) architecture has aroused significant attention for its low hardware complexity. Unfortunately, the hybrid precoding with the PC architecture may suffer from a severe spectral efficiency loss. To overcome the loss, a deep learning-based hybrid precoding algorithm for the PC architecture is proposed in this paper. Firstly, the optimization problem of hybrid precoding is transformed into a neural network prediction problem. Furthermore, a convolutional neural network(CNN) is designed to predict the hybrid precoder, where the channel matrix is used as input. Especially, to meet the constraints of the hybrid precoder, two lambda layers are introduced into the CNN architecture. Specifically, a lambda1 layer is constructed to address the constant modulus constraint of the analog precoding matrix and the block diagonal constraint. Then, a lambda2 layer is constructed to optimize the digital precoding matrix and address the transmit power constraint. Finally, the vectorized hybrid precoding matrix is output by the well-trained network. Simulation results show that the proposed algorithm can enhance spectral efficiency and energy efficiency with lower computation time.