Research On Deep Learning-based MIMO Hybrid Beamforming Technology

Massive multiple-input multiple-output(MIMO)technology,which increases the number of antennas to improve spatial freedom,has tremendous potential to improve spectrum efficiency and signal coverage,is a key physical layer technology for meeting the growing demand for wireless data throughput and communication service quality.The ideal full digital chain massive MIMO systems suffer from high hardware costs and power consumption,which poses huge practical difficulties for its large-scale deployment.The analog-digital hybrid transmission architecture combines the digital and analog domains for joint transceiver processing of signals,reducing the number of RF chains,and achieving a good compromise between system performance and complexity and cost,which is widely adopted by practical massive MIMO systems.Beamforming of analogdigital hybrid systems is made more challenging by the inclusion of analog chains.The traditional optimization techniques-based hybrid beamforming algorithm designs have issues such as high computational complexity,large processing delays,and the need for expert knowledge,which pose difficulties for the practical deployment and application of the algorithm.In recent years,the development of deep learning(DL)technology has opened up new ways of solving analog-digital hybrid beamforming optimization problems.In this context,this article explores DL-based hybrid beamforming optimization methods for massive MIMO,focusing on unsupervised learning schemes and low-complexity designs.The main research content includes the following three parts:(1)For a massive multiple-input single-output(MISO)system,a problem-solving model for maximizing spectral efficiency under transmit power and constant modulus constraints is constructed,a DL-based low-complexity hybrid beamforming optimization algorithm is proposed.The algorithm constructs an unsupervised learning convolutional neural network(CNN)model and adopts a two-stage training scheme.The core idea is to take channel samples as inputs,offline train the neural network with a custom DL loss function guided by nonconvex optimization mathematical model,and then deploy the trained neural network for online optimization,shifting the computational complexity from online optimization to offline training,greatly reducing computational complexity and time delay.Meanwhile,the performance of the proposed algorithm is innovatively discussed in terms of the specific solving ability of a single channel state information(CSI)and the generalization ability of multiple CSIs,respectively.The algorithm achieves high spectral efficiency performance at low computational complexity,overcoming the issues of high computational complexity and poor spectral efficiency of existing DL methods.(2)Based on the above research,for a massive MIMO system,a problem-solving model with the goal of efficiently jointly optimizing digital and analog beamformers is constructed,an unsupervised learning CNN-based fully-connected hybrid beamforming optimization algorithm is proposed.On the one hand,the algorithm uses unsupervised learning strategy and improves the loss function to train the system using only channel samples,thus avoiding model errors caused by local optimal label data.On the other hand,it utilizes the super learning ability of CNN for complex MIMO channel features to adaptively generate feasible hybrid beamforming solutions based on CSI.The algorithm achieves the spectral efficiency performance of the benchmark algorithm with limited training samples,has generality for different channel models,is unaffected by different antenna configurations,has strong generalization capability,and meets the requirements of high transmission performance and low latency for future intelligent communication.(3)For a massive MIMO system,a problem-solving model with the optimization goal of minimizing the Euclidean distance between unconstrained beamformer and hybrid beamformer is constructed,a supervised graph neural network(GNN)-based partiallyconnected hybrid beamforming optimization algorithm is proposed.The algorithm uses the mean square error criterion to newly design the loss function,and training the model to make the hybrid beamformer infinitely approximate the sample label,thus solving the nonconvex hybrid optimization problem using the known structure of optimal solution directly.Compared with existing networks lacking interpretability and scalability,the proposed GNN network framework combines expert domain knowledge for algorithm optimization and proposes a theoretically guided design method for performanceenhancing neural network structures with the advantage of interpretability,and achieving higher performance gains in dense wireless networks.