Deep Learning Based Hybrid Precoding for Multiuser mmWave MIMO Assisted via Sub-6G

Hybrid precoding technology is an economically efficient solution for modern millimeter wave (mmWave) massive multiple-input multiple-output (MIMO) communications due to its low hardware complexity and high performance. However, the design of hybrid precoding heavily relies on accurate channel state information (CSI), which requires a large number of pilot signals, thereby degrading system performance. Additionally, the computational complexity of hybrid precoding based on optimization or greedy algorithms increases exponentially with the number of antennas and users. This paper proposes a neural network (NN)-based multiuser hybrid precoding (MHP) framework that leverages the spatial consistency between sub-6G and mmWave channels to reduce pilot overhead while enabling efficient hybrid precoding design. The proposed frame-work consists of two key components: HNet and PNet, each designed to achieve specific objectives. First, HNet employs a convolutional NN (CNN)-based U-Net architecture to extract sub-6G CSI and reconstruct high-frequency mmWave CSI. Then, PNet, specifically designed based on the computational process of hybrid precoding, learns to compute the hybrid precoder from the reconstructed CSI. Simulation results demonstrate that the proposed framework effectively performs channel reconstruction and precoding design, achieving near-optimal hybrid precoding performance while reducing pilot overhead by 87.5%.