Research On Deep Learning Based CSI Feedback In Massive MIMO Systems

Massive MIMO technology has become one of the key technologies of 5Gowing to its high spectrum efficiency and energy efficiency.Due to the largenumber of base station antennas in the massive MIMO system,the acquisition of CSI by the base station will cause huge system overhead,which restricts the practical deployment of 5G networks in FDD mode.How to feed back high-dimensional CSI accurately with low overhead becomes one of the bottlenecks in designing FDD massive MIMO systems.The development of deep learning(DL)provides an alternative way to reduce the CSI feedback overhead in mas-sive MIMO systems.Intelligent communication has also become one of the promising directions in future 6G wireless communication researches.How-ever,how to achieve the efficient integration of wireless communication and DL remains an open research issue,as many existing DL-based works do not involve interpretability and the integration of domain knowledge in wireless communication despite apparent successes.Consequently,this thesis focuses on the DL-based low-overhead CSI feed-back in the massive MIMO system.In order to solve the low-overhead and high-precision feedback problem and interpretability problem faced by the ex-isting DL-based schemes,this thesis proposes the corresponding solutions by integrating the wireless channel characteristics and the advantages of traditional methods.In order to achieve high-precision CSI feedback with low feedback overhead,firstly,a bi-directional channel correlation-based deep CSI feedback scheme is proposed by exploiting correlated characteristics from the physical environment of massive MIMO uplink and downlink channels;Then,a Marko-vian time-varying channel model-based deep CSI differential feedback scheme is proposed by exploiting correlation characteristics of massive MIMO channelin the time domain;Next,to improve the encoding efficiency for feedback code-words,a j oint dimension compression and codeword quantization optimization-based CSI feedback scheme is proposed by utilizing the end-to-end optimiza-tion characteristics of DL.Finally,in order to improve the interpretability of the DL-based CSI feedback scheme,a deep unfolding based CSI feedback scheme is proposed by utilizing the interpretability of the compressive sensing and the advantages of neural networks in data-driven tuning.Specifically,the thesis mainly includes the following four aspects of innovative work:·The DL-based CSI feedback scheme always suffers from an obvious accu-racy loss in the complex scenario or when the compression ratio is small.Consequently,Chapter 2 of this thesis proposes a CSI feedback scheme DualNet based on the bi-directional channel correlation and DL.Specif-ically,DualNet exploits the correlation between the uplink and down-link CSI in the delay domain to construct the deep autoencoder architec-ture,and utilizes the available uplink CSI in the base station to assist the downlink CSI recovery in improving the feedback accuracy in the massive MIMO system.In order to further reduce the feedback overhead,this the-sis studies the downlink CSI prediction based on the relevant uplink CSI,and explores the influence of channel configurations on the bi-directional channel correlation,which provide the reference direction to reduce the DL-based CSI feedback overhead in massive MIMO systems.·In order to achieve low-overhead and high-precision CSI feedback while reducing deployment costs,this thesis uses the correlation of massive MIMO channels in the coherence time to further tap the potential of reducing CSI feedback overhead from the time dimension,and proposes a CSI differen-tial feedback scheme DiffNet based on the temporal correlation and DL.Specifically,Chapter 3 of this thesis first studies the temporal correlation characteristics of massive MIMO channels,and measures the prior infor-mation from adjacent CSI using conditional entropy.Although recurrent neural networks can exploit the temporal correlation to improve the CSI accuracy,the cost of computation and memory is staggeringly high for the massive MIMO deployment.Thus,DiffNet is designed as a differ-ential CSI feedback scheme based on the first-order Markovian model of the MIMO channel.By focusing on the prediction error based on CSI at adjacent moments,DiffNet can increase the CSI feedback accuracy while reducing the model size and computational complexity.Furthermore,ac-cording to CSI data characteristics,a spherical feedback framework and CSI feedback enhancement network structure are proposed in the DiffNet to provide high-precision CSI feedback at the initial moment as the prior information.To further reduce the CSI feedback model’s storage and cal-culation costs,a convolutional neural network-based dimension compres-sion module is designed to eliminate redundant connections in the neural network.The evaluation results show that DiffNet can significantly im-prove feedback performance while reducing the model size and computa-tional complexity.·Current CSI feedback works mainly focus on the dimensional compres-sion of CSI matrices,and how to efficiently encode the codewords after dimension compression is still to be solved.Consequently,Chapter 4 of this thesis designs an efficient modular compression framework CQNet to jointly optimize the CSI dimension compression,codeword quantiza-tion and recovery based on the end-to-end optimization characteristics of DL.CQNet contains an efficient quantization module by customizing the quantization interval for each element in the codeword vector after di-mension compression and the μ-law compander to improve quantization efficiency.CQNet also proposes an efficient quantization scheme in the polar coordinate by introducing the magnitude adaptive phase quantiza-tion.Then,this thesis integrates CQNet with two existing DL-based CSI feedback schemes to demonstrate the simplicity and efficacy of CQNet in improving bandwidth efficiency.The evaluation results show that CQNet can achieve better CSI feedback accuracy and robustness with low over-head.·Considering the lack of interpretability in the design of DL-based C SI feedback networks,Chapter 5 of this thesis j ointly utilizes the interpretabil-ity of compressive sensing and the advantages of neural network in data-driven tuning,and proposes a deep unfolding-based CSI feedback scheme TMM-ISTANet+.TMM-ISTANet+consists of an encoder composed of a measurement matrix for dimension compression and an interpretable de-coder for the iterative CSI reconstruction.Owing to the performance loss caused by randomly generated measurement matrix,TMM-ISTANet+in-troduces a trainable measurement matrix to mine the characteristics of CSI data and improve the CSI matrices’ compression efficiency.To improve the decoder’s interpretability,TMM-ISTANet+maps the ISTA compres-sive sensing reconstruction algorithm into the corresponding iterative mod-ules of the decoder.TMM-ISTANet+can optimize parameters of the en-coder and decoder according to the CSI data characteristics through the end-to-end gradient descent training.The evaluation results show that while TMM-ISTANet+improves the interpretability in the design of CSI feedback networks,it also guarantees the CSI reconstruction accuracy un-der low-overhead feedback for massive MIMO systems.