CNN-DPC algorithm for hybrid precoding in millimeter-wave massive MIMO systems

The dynamic partially connected (DPC) structure can achieve the trade-off between hardware loss and precoding performance of millimeter-wave multi-input multi-ouput systems. Via convolutional neural network (CNN), a novel hybrid precoding approach is proposed for the DPC structure in this paper, namely CNN-DPC algorithm. Firstly, the Euler’s formula and identity matrix are used to construct a phase shift (PS) layer that handles diagonal constraints of the analog PS precoding matrix and constant modulus. Then, the state of the connection between the antennas the radio frequency (RF) chains is determined by the probability layer. On this basis, a lambda layer is established to output the vectorized hybrid precoding matrix, and a loss function is expressed as the minimization of the Euclidean distance between the optimal fully digital precoding matrix and the hybrid precoding matrix. Finally, after the CNN learns the mapping relationship between the hybrid precoding matrix and the channel characteristics, it can directly output the desired hybrid precoding matrix with the input of the channel matrix. The proposed CNN-DPC algorithm achieves higher energy efficiency and the spectral efficiency compared with the related algorithms is indicated in the simulations.