Research On Superimposed CSI Feedback Based On Deep Learning

In massive MIMO system,the accurate downlink channel state information(CSI)is required by the beamforming and user selection at base station(BS).In time division duplex(TDD)mode,the downlink CSI can be estimated by the uplink channel for the reciprocity property.However,in frequency division duplex(FDD)mode,the reciprocity-based CSI is not available.Thus,the downlink CSI should be estimated by users and fed back to the BS.Due to the large number of BS antennas in massive MIMO system,the CSI feedback incurs significant amount of overhead and is extremely challenging.In order to reduce the feedback overhead and completely avoid the occupation of uplink bandwidth resources,the main work of this thesis is as follows.Firstly,to avoid the occupation of uplink bandwidth resources,a deep learning-based superimposed CSI feedback is studied by combining deep learning and superimposed feedback technology.The downlink CSI is spread and then superimposed on uplink user data sequences(UL-US)to feed back to BS.After BS receives the superimposed signals,an interference cancellation network,built by deep learning,is employed to recover the downlink CSI and detect UL-US,which can completely avoid the occupying of uplink bandwidth resources.The analysis and simulation results show that,compared with the traditional superimposed CSI feedback method,without losing the detection performance of UL-US,the deep learning-based superimposed CSI feedback significantly improves the reconstruction accuracy of downlink CSI.Secondly,considering the large training parameters and the complex parameter training process of the interference cancellation network on deep learning-based superimposed CSI feedback.An extreme learning machine-based superposition CSI feedback is studied.The downlink CSI is also spread and then superimposed on UL-US to feed back to BS.At BS,an interference cancellation network,built by ELM network,is employed to recover the downlink CSI and detect UL-US.Compared with the interference cancellation network based on deep learning,ELM is a forward neural network,whose parameters are optimized by a simple matrix operation process.And ELM can directly process complex data,which can greatly reduce the network parameters and space occupation.The analysis and simulation results show that,compared with the deep learning-based superimposed CSI feedback method,the ELM-based superimposed CSI feedback method also does not occupy the uplink bandwidth resources,and can significantly reduce the network parameters,storage space,offline training time and online running time while guaranteeing the UL-US and downlink CSI reconstruction performance.Finally,in order to minimize the superposition interference of deep learning superimposed CSI feedback method,a deep learning-based 1-bit superimposed CSI feedback is studied.At user side,the downlink CSI is quantized by 1-bit compression,and only its symbolic information is retained.Then,the downlink CSI is spread and superimposed on UL-US as feedback to BS.BS combines the traditional superimposed coding aided binary iterative hard thresholding algorithm(SCA-BIHT)and deep learning network to reconstruct the downlink CSI and UL-US.The analysis and simulation results show that,compared with the traditional1-bit superimposed CSI feedback method,deep learning-based 1-bit superimposed CSI feedback can significantly improve the reconstruction performance of UL-US and downlink CSI,and greatly reduce the online running time,which can reduce the processing delay.The deep learning-based superimposed CSI feedback method studied in this thesis can be applied to massive MIMO wireless communication systems,such as the fifth generation wireless communication system,the sixth generation wireless communication system,and the next generation Wi Fi(802.11ax/ay)system,etc.