Research On Autoencoder-based CSI Feedback In Massive MIMO Systems

With the rapid development of mobile communication technology,mankind has gradually entered the era of Internet of Everything,and a large number of new services and applications have appeared in front of people.However,the explosive growth of traffic demand has gradually approached the system capacity of the communication network,and communication requirements such as high-speed transmission and low delay have also brought huge challenges to the existing communication system.With the promotion and commercialization of 5G technology,the communication field is in a new stage of development,and innovative methods and technologies are urgently needed to promote its next development.In recent years,deep learning has become a focus of attention in academia and industry.Deep learning can learn the deep features of data samples through neural networks,and has great advantages in the face of unstructured information and massive data.This thesis studied the main application of deep learning in the field of wireless communication signal processing,and focuses on 5G key technology: frequency division duplexing(FDD)massive multiple-input multiple-output(MIMO)system downlink channel state information(CSI)feedback problem.In the Massive MIMO system,the base station will be equipped with a large number of antennas,which will cause a lot of overhead for the base station to obtain CSI information.In order to reduce the feedback overhead,this thesis studied the compression and reconstruction tasks of CSI information based on the deep learning autoencoder network.The specific research content is as follows:1.A CSI compression and reconstruction method based on convolutional layer receptive field transformation is designed.This method effectively displays the sparsity of the signal by expanding the convolution receptive field in the encoder of the network,and optimizes the feature extraction of the original signal.Simulation results show that this method is superior to the existing methods based on compressed sensing and Csi Net,which is a deep learning-based method.Under different compression ratios compared with Csi Net,the NMSE of the indoor scenario is reduced by about 1.13 dB on average,and the NMSE of the outdoor scenario is reduced by about 0.12 dB on average.2.Aiming at the problem that the existing deep-learning based CSI feedback method is difficult to adapt to multiple test scenarios,a CSI compression and reconstruction method based on multi-scale feature fusion is designed.This method optimizes the joint performance of CSI information compression and reconstruction tasks by designing correspondingly matched multi-scale feature fusion networks in the encoder and decoder.The simulation results show that this method has better robustness and provides a considerable performance gain with a small increase in network parameters.Under different compression ratios compared with Csi Net,the NMSE of the indoor scenario is reduced by about 3.66 dB on average,and the NMSE of the outdoor scenario is reduced by about 1.71 dB on average.