Research On Deep Learning Based CSI Feedback In FDD Massive MIMO Systems

Massive Multi-Input Multi-Output(MIMO)can improve spectral efficiencies,energy efficiencies,and system capacities,and has become a key technology of the 5G communication system.Frequency Division Duplexing(FDD)is an important duplexing method for Massive MIMO systems because of its transmission capability and forward compatibility.However,the challenge of FDD mode is that the user equipment needs to feed the downlink channel status information(CSI)back to the base station through the uplink channel.With increasing the number of antennas and subcarriers,CSI feedback will consume a large number of time-frequency resources.Therefore,obtaining accurate downlink CSI with low feedback overhead is very important for FDD Massive MIMO systems.For FDD Massive MIMO systems,existing compressed sensing(CS)-based feedback schemes can reduce the feedback overhead by exploiting the channel sparsity.But this kind of method still has some shortcomings.First,it is difficult to obtain a strict sparse representation of the channel.When the angle of path does not fall on the sampling grid,the commonly used DFT transform has the phenomenon of energy leakage,which will affect the performance of the CSI reconstruction algorithm.Secondly,the commonly used convex relaxation and Bayesian learning algorithms usually suffer from high computational complexity and are not suitable for low delay scenarios.Compared with the traditional iterative algorithm,the deep learning(DL)-based feedback schemes have the advantage of computing time.However,most of the existing schemes regard the neural network as a black box,and the interpretability of its model construction is poor.In this thesis,model-based downlink CSI feedback networks are proposed,which combine compressed sensing and deep learning methods and utilize channel data features to improve the reconstruction performance of downlink CSI.To solve the two shortcomings of the traditional compressed sensing algorithm,this thesis proposes a CSI feedback scheme based on an iterative unfolding network.Specifically,an iterative unfolding network takes the place of the traditional iterative algorithm to reconstruct the downlink CSI in a shorter time.At the same time,to improve the performance of downlink CSI reconstruction,this thesis designed a sparse autoencoder to obtain a strict sparse representation of the channel.Experimental results show that the proposed scheme is superior to the traditional CS scheme and the existing DL-based feedback scheme when the compression rate is higher than 1/32.Moreover,the reconstruction speed of the proposed scheme is better than that of the traditional basis pursuit algorithm.To further improve the accuracy of downlink CSI reconstruction,this thesis proposes an uplink-aided downlink CSI feedback scheme by exploiting the reciprocity of angle and delay parameters in FDD UL and DL channels.The CSI reconstruction problem was modeled as a weighted L1 minimization problem,whose weighting coefficients depend on the support information of uplink CSI.According to the characteristics of uplink and downlink channels,a joint sparse autoencoder is designed to learn the joint representation of uplink and downlink channels.The experimental results show that the proposed scheme outperforms existing uplink-aided feedback networks and downlinkbased architecture.