Machine Learning Assisted Precoding Technique For Massive MIMO System

As one of the key technologies of the 5th Generation(5G)mobile communication system,massive multiple-input multiple-output(MIMO)technology is getting increasing attention.Due to the large number of antennas in the base station(BS),it can significantly increase spectrum efficiency while serving multiple users at the same time.In this paper,the low complexity robust precoding assisted by machine learning(ML)in massive MIMO systems is studied,focusing on the transmission performance improvement and complexity reduction of massive MIMO systems in the mobile environment.The low complexity ML-assisted robust precoding frameworks are proposed for three different communication scenarios.Firstly,we review the existing massive MIMO precoding methods,including Matched Filter(MF)precoding,zero-forcing(ZF)precoding,regularization zero-forcing(RZF)precoding,and robust precoding for imperfect channel state information(CSI).According to the simulation results,the MF precoding method has better rate performance in low SNR points,the ZF precoding method has better performance in high SNR points and the RZF precoding method has better performance in all SNR points.In mobile scenarios where the perfect CSI cannot be obtained,the sum-rate performance of the three typical linear precoding methods decreases significantly,while the robust precoding method can maintain the high sum-rate performance in all scenarios.Secondly,a ML-assisted weighted robust precoding method based on the ground communication scenarios with single-antenna terminals is proposed.First of all,the equivalent robust precoding design method is derived by transforming the traditional robust precoding method.Then,to match the DL method to reduce complexity,a weighted robust precoding framework assisted by ML is proposed.The precoder is divided into instantaneous precoding and statistical precoding.The instantaneous precoding is approximately regularized force-zero precoding,while the statistical precoding is designed by deep learning neural network.According to the simulation results,in the case of short-sounding SRS,the proposed ML-assisted weighted robust precoding method can achieve near-optimal performance with low complexity.With the extension of SRS,the prediction performance of the DL neural network deteriorates,thus leading to a performance decline.If the DL neural network with better performance is trained,the sum-rate performance of the proposed framework can be further improved.Thirdly,a joint transceiver beamforming method based on the ground communication scenarios with multiple-antennas terminals is proposed.First of all,based on the uplink-downlink duality of CSI,a lowcomplexity transceiver beamforming design with quality of service(Qo S)constraints is studied to minimize transmitting power.As the upper bound of the transmission power performance,the traditional iterationbased algorithm is used for reference,and a general DL framework and a general deep neural network(DNN)structure are introduced to reduce the complexity brought by iterations.The trained DNN can directly learn key features from CSI.In addition,to further reduce the complexity brought by the neural network itself,a heuristic algorithm based on the maximum eigenvalue-eigenvector of CSI is proposed.According to the simulation results,compared with the traditional iterative algorithms,the proposed joint transceiver beamforming design method has lower computational complexity and higher feasibility while keeping close to the optimal performance.Finally,a ML-based downlink precoding design method for satellite mobile communications scenarios with single-antenna terminals is proposed.For the downlink of a typical low earth orbit(LEO)satellite multiuser system,the ML-assisted precoding of satellite downlink was studied to maximize the sum rate performance considering the various Antenna Power Constraints(PAPCs).Firstly,an iterative downlink precoding method is derived by solving the sum-rate maximization problems under PAPCs.Secondly,a ML-based low complexity satellite downward precoding algorithm is proposed.A DL model is introduced to transform the iterative process into a neural network prediction process.According to the simulation results,the proposed ML-based low complexity satellite downward precoding method for satellite mobile communication has low computational complexity while keeping close to the optimal performance.