Limited Feedback Beamforming Design Based On Deep Learning For 6G Extra Large-scale Massive MIMO

With the development and progress of science and technology,people’s demand for the quality of wireless communication is gradually rising.Scholars from all over the world in communication field have started the exploration and research on the 6th Generation Mobile Communication System(6G)successively.As one of the key technologies in 6G,the extra large-scale massive Multiple-Input Multiple-Output(MIMO)technology can achieve higher communication efficiency,wider coverage,lower delay and more reliable transmission by deploying a large number of antennas at the base station.However,in extra large-scale massive MIMO scenarios,problems such as heavy computation requirements,storage content and transmission overhead bring huge challenges to the engineering realization.Moreover,the improvement of user experience depends on the reasonable allocation of wireless resources and the acquisition of high throughput,which requires accurate channel precoding at the base station to achieve the above conditions.The realization of precoding technology depends on accurate Channel State Information(CSI).However,in frequency division duplexing(FDD)systems,there is no channel reciprocity between the uplink and downlink,the base station(BS)cannot obtain the downlink CSI through the uplink CSI directly.Consequently,downlink CSI in FDD systems should be fed back through feedback link from the user equipment(UE)to the BS so that the wireless communication system can perform its functions normally.Therefore,based on this background,this paper designs a limited feedback beamforming scheme based on deep learning for 6G extra large-scale massive MIMO system to solve the problems of high computation cost,high quantization noise and low feedback accuracy.Firstly,this paper proposes a quantizer KMQBlock based on k-means clustering algorithm.The quantizer can find the appropriate quantization level according to the distribution of CSI codewords elements by iterating and updating to reduce the quantization noise.Then,a CSI feedback network based on deep learning named MRNet,which contains multiple residual blocks is proposed to further demonstrate the performance of KMQBlock,and KMQBlock is used to quantify the feedback codewords.In order to further improve the accuracy of CSI feedback,vector quantization scheme is introduced in KMQBlock.Compared with scalar quantization,vector quantization scheme can quantify the feedback codewords at a higher dimension,which can help neural network learn the distribution characteristics of codeword elements better.Experimental results show that KMQBlock can provide more reconstruction accuracy for MRNet than conventional quantization methods.Secondly,for extra large-scale massive MIMO scenarios,this paper proposes a data dimension reduction scheme to deal with the problem of oversized datasets.In this scenario,the limited feedback beamforming network FBFNet based on deep learning is designed so that the precoding matrix can be obtained directly at the base station.RDBlock,a data dimension reduction module is added in this network.This module can cut the training data according to the numerical distribution characteristics of the training dataset.In order to solve the problems such as excessive network overhead and long computing time in the extra large-scale massive MIMO scenario.In addition,KMQBlock is added in the feedback process to reduce the feedback overhead.