Research On Deep Learning Based Downlink CSI Limited Feedback For FDD Massive MIMO Systems

As the key technology in the fifth-generation wireless communication system,due to the high spectrum efficiency,large link capacity and strong robustness,massive multi-input multi-output(MIMO)technology has been a major enabler for future wireless communication systems.These potential benefits are based largely on the accurate acquisition of uplink and downlink channel state information(CSI).However,in frequency division duplexity massive MIMO system,the channel reciprocity between the uplink channel and downlink channel does not exist.Inferring the downlink CSI from the uplink CSI by the channel reciprocity is impossible.Therefore,the base station(BS)usually sends the pilot to the user equipment(UE),and the UE estimates the downlink CSI and feeds it back to the BS.However,the estimating process of the downlink CSI consumes the valuable computing resources of UE,and the feedback process of the downlink CSI consumes the valuable bandwidth resources.Massive antennas at the BS and UE for massive MIMO make the already bad situation worse in the beyond systems.In this paper,a deep learning based fully convolutional neural network is proposed to compress and decompress downlink CSI,so as to realize the limited feedback process of downlink CSI and reduce the feedback overhead.The detailed work is as follows:Firstly,under the premise of abandoning the fully connection layer and ensuring the reconstruction performance of downlink CSI,this paper designs a new deep neural network:FullyConv.FullyConv is based on encoder-decoder architecture,which is all composed of convolutional layers.Specifically,the encoder uses convolution layers for compression and the decoder uses deconvolution layers for decompression.Simulation results show that FullyConv can effectively improve the reconstruction accuracy of downlink CSI and reduce the training and storage overhead.Secondly,quantization module,dequantization module and noise are added to FullyConv to simulate the real feedback scenario of downlink CSI.In the quantization module,the uniform quantization method and non-uniform quantization method are introduced respectively,and end-to-end training strategy and step-by-step training strategy are proposed respectively;In order to offset the quantization error caused by the quantization module,an offset network based on convolutional layer is introduced in the dequantization module;Noise is added after the quantization module to simulate the impact of noise on the data-bearing bitstream in the transmission.In the simulation experiments,this paper first analyzes the performance of two quantization methods and two training strategies.Secondly,we make suggestions for the design of non-uniform quantization function from the perspective of probability density distribution of codewords and non-uniform quantization curve.Finally,this paper considers the quantization module,dequantization module and noise at the same time,and analyzes the influence of quantization and noise on the reconstruction of downlink CSI.Finally,considering that the communication system needs to dynamically adjust the compression rate according to the change of the communication scene,this paper designs two multiple-rate downlink CSI compression networks: MR-FullyConv-M1 network and MR-FullyConv-M2 network to realize the dynamic change of compression rate.Compared with the single-rate network,the simulation results show that the MR-FullyConv-M1 network reduces the training overhead and storage overhead,but as a price,the performance of downlink CSI reconstruction has a certain decline;The proposed MR-FullyConv-M2 network shares the decompression module on the basis of the MR-FullyConv-M1 network,which further reduces the training overhead and storage overhead,but also further reduces the performance of downlink CSI reconstruction.