Deep Learning Based Adaptive Hybrid Beamforming for mmWave MIMO Systems

Abstract

In the fifth-generation communication, the hybrid precoding technique is used in the massive multiple-input multiple-output (MIMO) system to reduce the RF chain number for power reduction. In recent years, deep learning techniques have been widely used in the hybrid precoding design to improve spectrum efficiency. This paper proposes an alternating minimization-based deep learning precoding technique for the hybrid precoding. This technique includes the phase information of the channel matrix in the deep learning model to improve the spectral efficiency. In addition, an on-line training method is also designed to track the channel features of the time-varying channel. Thus, the deep-learning neural network model can adaptively track the time-varying channel characteristics with a better performance than its counterpart deep-learning-based hybrid beamforming (DLHB) technique even if the initial network model is not good. The simulation experiments also analyze and compare the spectral efficiency with different hyperparameters of the deep-learning neural network model. The proposed adaptive hybrid precoding technique can further reduce 51.54% of the trainable parameters in the time-invariant environment and 76.14% of trainable parameters can be reduced in the time-varying environment compared to the benchmark technique of the DLHB. With the reduced parameter size, the proposed technique can be 1.6 ms faster than the DLHB with better spectrum efficiency.