AI-Based Pre-coding Design For Multi-user Massive MIMO Systems

In Massive Multiple-Input Multiple-Output(MIMO)systems,downlink pre-coding can effectively eliminate inter-user interference and channel noise,thereby improving the system's communication performance.Traditional optimization-based pre-coding algorithms are slow to converge and have high computational complexity.Cannot meet the needs of real-time applications for fifth-generation(5G)and above systems,such as self-driving vehicles and mission-critical communications.Even in small-scale decaying non-real-time applications that vary in milliseconds,the delays introduced by the iterative process can make it difficult for beam-fitting solutions to meet communication needs.In addition,Massive MIMO system pre-coding requires the base station(BS)end to know the downlink channel status information(CSI)in advance,in FDD system,the downlink channel uses different frequency points,the channel is not reciprocal,the traditional downlink channel estimation method requires a lot of communication resources.In recent years,deep learning(DL)technology has become a research hotspot due to research in the field of wireless communications,and has made great progress in channel detection,channel estimation,CSI feedback and reconstruction,channel decoding,and even end-to-end wireless communication systems,providing a new way of thinking to the current traditional methods to solve the dilemma of communication problems.This paper focuses on two very important issues in FDD Massive MIMO systems,downlink pre-coding and downlink channel estimation.The main tasks are as follows.(1)For the problem of high-dimensional matrix inversal in the traditional linear pre-coding scheme,a low complexity R-WMMSE algorithm is proposed based on the classical WMMSE algorithm through mathematical analysis,so that the computational complexity of the matrix inversely involved is independent of the number of antennas at the transmitting end,which can reduce the computational consumption in the iterative process.(2)A low complexity pre-coding learning model was designed to.Various learning strategies such as supervised learning,unsupervised learning,supervised learning and unsupervised learning combined were tried,and various neural network models including MLP and CNN were used,while learning model design under different user weights was also considered.For the case where the pre-coded matrix is extremely high-dimensional with a large number of antennas at the transmitting end,we redesigned the input-output of the network so that the input of the neural network is independent of the number of antennas at the transmitting end,greatly reducing the training difficulty of the model.The experimental results show that our learning scheme can also achieve very good approximation accuracy under relatively simple network models.(3)A deep learning-based model for estimating the downlink channel of the FDD Massive MIMO system was designed.Based on the sufficiently similar characteristics of the scattering body and large-scale propagation environment of the FDD uplink and downlink,neural networks are able to understand the physical connections and constraints between two adjacent frequency bands.We use MLP and CNN networks for downlink channel estimation,train the CSI reconstruction network offline,and deploy the trained model online,and the simulation results show that our channel estimation model achieves relatively good estimation results.