DL-Based Energy-Efficient Hybrid Precoding for mmWave Massive MIMO Systems

• Published in: IEEE transactions on vehicular technology Vol. 72; no. 5; pp. 6103 - 6112
• Main Authors: Liu, Fulai, Zhang, Lijie, Yang, Xianghuan, Li, Tiangui, Du, Ruiyan
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.05.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: adaptive architecture; Algorithms; Artificial neural networks; Complexity; Convolutional neural networks; deep learning; Energy efficiency; Hybrid power systems; Hybrid precoding; Hybrid systems; Machine learning; millimeter wave; Millimeter wave communication; Millimeter waves; Optimization; Phase shift; Power consumption; Power demand; Precoding; Radio frequency; Switches
• ISSN: 0018-9545; 1939-9359
• DOI: 10.1109/TVT.2022.3230931

Hybrid precoding based on an adaptive-connected structure is one of the promising technologies for millimeter-wave communications, which achieves a good tradeoff between spectral efficiency and power consumption by deploying a switch-controlled connection between the antennas and radio frequency chains. To maximally enhance the energy efficiency while maintaining superior spectral efficiency of hybrid precoding, a novel deep learning-based optimization algorithm is presented for an adaptive fully-connected (AFC) structure in this paper. Firstly, two special convolutional neural network (CNN) frameworks are designed to optimize the phase shift precoding matrix and switch precoding matrix with hardware constraints, namely CNN-Fps and CNN-Fs respectively. Furthermore, a CNN-based joint optimization network is developed, named as CNN-JO network, which can simultaneously optimize the CNN-Fps and CNN-Fs subnetworks, such that the optimal analog precoding matrix is obtained. Then, using the fully digital optimal precoder as the training label, the proposed CNN-JO network can be trained to maximize the energy efficiency of the AFC structure. Finally, the well-trained CNN-JO model can accept the estimated channel matrix as the input and directly output the phase shift precoding matrix, switch precoding matrix, and digital precoding matrix. Simulation results and complexity analysis show that the proposed algorithm can achieve better performance than the previous works in terms of energy efficiency and spectral efficiency with a lower complexity.