Machine Learning Research Based On Massive MIMO Measured Channel Data

With the gradual maturity of massive MIMO,this physical layer technology,which can greatly increase the communication rate,has entered the 5G mobile communication standard and will soon be widely deployed and utilized in 5G commercial applications.massive MIMO is a key technology to achieve high spatial resolution by increasing the number of base station antennas and RF links.The beamforming of massive MIMO communication system relies on accurate channel state information(CSI).In time division duplex,the base station needs to estimate the uplink channel state information in real time.In this system,since the number of antennas and the number of users at the base station side are one to two orders of magnitude larger than those of the conventional MIMO system,Coupled with the number of orthogonal frequency division multiplexing subcarriers and the variation of the channel in the time dimension and the number of samples,the channel state information data in the massive MIMO system can be regarded as”Big Data” in 5G wireless communication.” These CSI data are usually discarded directly after baseband processing.However,these data contain the channel environment information of the base station and the cell where the user is located,and have certain potential utilization value.This paper envisages storing channel state information for data analysis and data mining,and using machine learning techniques to discover the structure and correlation of Channel State Information,in order to assist Massive MIMO communication and reduce the complexity of massive MIMO systems,or Form new applications based on massive MIMO systems.The research work in this paper is based on massive MIMO measured channel data.The experimental environment is a wireless channel in a suburban outdoor environment.The base station uses a cylindrical antenna array of 128 antennas.The user end uses a single vertically polarized antenna.The carrier frequency is 2.6 GHz,bandwidth is50 MHz,and there is a line of sight propagation path between the base station and the user.In the analysis of the channel data,not only the form of the matrix is used,but also the data is transformed into a tensor form to analysis.The study used classical principal component analysis(PCA)and tensor multilinear principal component analysis(MPCA),and eigenvector non-orthogonal sparse dictionary learning method to quantify the sparsity of massive MIMO channel data.Then,the preprocessed data is geographically classified and identified using a neural network(NN).We propose the idea of using massive MIMO channel data as physical layer big data.After the large-scale MIMO channel is known to have sparse characteristics,the measured channel data is quantized,and the tensor form and method are used to preserve the potential correlation between different dimensions.After preliminary analysis of the data,we found that the geographic location classification and identification based on channel data can achieve accuracy of more than 80%,especially using data with only channel gain for classification and identification,which not only greatly reduces the computational complexity,and the accuracy has also improved significantly,reaching more than 90%.