| Since its commercial launch in 2018,the 5th generation mobile communication systems(5G)has continued to evolve and has entered the second evolution stage,namely,5G-Advanced.Meanwhile,the vision and technological pre-research for the 6th generation mobile communication systems(6G)are also being carried out worldwide.Artificial intelligence(AI)is widely regarded as a key enabling technology for 5G-Advanced and 6G,and will play an essential role in considerably improving the performance of mobile communication systems further.AIbased massive multiple input multiple output(MIMO)channel state information(CSI)feedback has become a hot topic in academia and industry,and has gained tremendous attention.This thesis focuses on four aspects of AI-based CSI feedback: expert knowledge exploitation,multimodule joint design,integration with new technologies,and practical deployment challenges.First,an intelligent CSI feedback scheme is proposed on the basis of the correlation between nearby users’ CSI.To reduce the feedback overhead,the CSI magnitude information of nearby users is divided into shared and individual information.Nearby users collaborate to reconstruct shared information through NNs and separately reconstruct individual information.This strategy can reduce the additional overhead caused by repeated feedback of shared information.Subsequently,the digital CSI feedback is considered,and two bit generation strategies,including quantization and binarization are proposed and compared.Also,for the feedback of CSI phase information,a CSI phase feedback strategy based on statistical or instantaneous CSI magnitudes is introduced.Simulation results show that the expert knowledge exploitation of CSI correlation of nearby users can reduce the CSI magnitude feedback overhead by about60%,and the proposed phase feedback strategy can reduce the phase feedback overhead by more than 85% given the feedback accuracy of-10 dB compared with the basic autoencoder architecture.Second,a domain adaptation-based intelligent CSI feedback framework is investigated based on the fact that many users may stay in a relatively stable environment for a long time.A new encoder is generated to adapt to the new CSI distribution in the single-user scenario.Then,given the limited CSI datasets at a single user,two data augmentation methods(including random phase shifting and random erasure)based on the physical characteristics of CSI are introduced.The proposed method is extended to a multi-user scenario,where crowd intelligence is fully utilized through decentralized gossip learning to enable online training of NN models without the participation of base stations.Simulation results show that the proposed domain adaptation-based method can improve the reconstruction accuracy by more than 1.2 dB given the compression ratio of 16,while the gossip learning supported by crowd intelligence can improve the reconstruction accuracy by more than 2 dB further.Afterwards,an intelligent CSI acquisition framework based on partial bidirectional reciprocity and joint design is introduced to reduce pilot and feedback overheads.Considering the pilot design and channel estimation,we propose an NN that generates the downlink pilot based on the uplink CSI amplitude.Given that the downlink CSI feedback is only required to feed information not contained in the uplink CSI,thus the uplink CSI amplitude information is also fed into the decoder-based reconstruction module,thereby forcing the encoder to only feed the desired information through end-to-end learning.To achieve overall rather than local optimality in CSI acquisition,the above three modules are jointly designed,and we propose to feed back the obtained pilot signal directly to the base station without extra channel estimation.Simulation results show that the uplink-aided joint design can considerably reduce the overhead of downlink CSI acquisition,with a reduction of over 25% in pilot overhead and about 20% in feedback overhead.Subsequently,the joint intelligent CSI feedback and beamforming design is investigated based on the task-driven approach given that the feedback CSI is mainly used for beamforming design at the base station.In the single-cell scenario,the encoder at the user compresses the downlink CSI,and the decoder at the base station designs the beamforming vector according to the received feedback codeword via NNs.The NNs at the encoder and decoder are trained together via an unsupervised approach,whose optimization objective is to maximize the achievable rate Once obtaining the CSI feedback information,the CSI reconstruction is no longer required,and the NN at the base station directly designs the downlink beamforming vector based on the feedback information.Then,for a soft hand-off multi-cell scenario,where the user receives the signals not only form the local but also the nearby cells,two NN encoders are proposed to feed the above two channels separately,and the NN at the base station performs beamforming design based on the feedback information of the two channels.Numerical simulations show that the joint design reduces the feedback overhead by 40% and 50% for single and multiple cells,respectively,without performance drop in system rate.Next,inspired by the new technology,namely,semantic communication,an intelligent CSI feedback based on deep data hiding is proposed to eliminate the feedback overhead.The CSI features are extracted by the user through NNs,and then hidden into the image to be transmitted through deep data hiding given that the semantics of the original image remains unchanged.The base station reconstructs the downlink CSI directly from the image,thus eliminating the original feedback link and its corresponding overhead.Then,considering the impacts of source coding on the proposed CSI hiding,a two-stage training strategy is provided to ensure that the base station can correctly recover the hidden CSI from the coded image,and a block with few/no impacts on the code length of the coded image is selected by an entropy-assisted strategy for CSI hiding.The numerical results show that the proposed method can feed back downlink CSI with zero(or very low)overhead and high accuracy without affecting the semantics and visual quality of transmitted images.Later,the CSI feedback in the reconfigurable intelligent surface(RIS)assisted mobile communication system is studied,which makes full use of the features of RIS-assisted communication systems to reduce feedback overhead.Based on the two-timescale characteristics of CSI in RIS communication systems,a two-timescale feedback strategy is proposed,where the feedback periods of the CSI between base station-RIS and RIS-user are different.Then,the above two CSI matrices are compressed by two NN-based encoders at the user,while the base station directly designs the beamforming vector and RIS discrete phase shift on the basis of the received feedback information.To avoid repetitive input of the feedback information of the long-term constant base station-RIS channel,this feedback information is introduced into the beamforming and RIS phase shift design via a hypernetwork structure.Simulations show that the proposed intelligent CSI feedback scheme considerably outperforms other algorithms in achievable rate and its inference time is less than half of that of the conventional algorithm.Then,an intelligent CSI feedback scheme based on codebook enhancement is introduced to improve the performance of CSI reconstruction at the base station through AI without changing the existing feedback framework.The user still uses the traditional codebook feedback architecture for the codebook index feedback,and improves the reconstruction accuracy of CSI by introducing additional NNs at the base station to ”refine” the results of the codebook feedback and introduce the environmental knowledge.Then,on the basis of the above framework,the uplink CSI magnitude is also fed into the NN,thereby making full use of the partial reciprocity between the uplink and downlink to extend the refinement gain.Next,the impacts of imperfect channel estimation on the proposed framework are considered,and the estimated CSI is denoised by NNs at the user and the base station,respectively.Simulation results show that the proposed method improves the system rate of the baseline codebook by more than 100%without changing the standard,and is robust to channel estimation errors.Finally,deployment-oriented lightweight intelligent CSI feedback methods are investigated to reduce the storage and computing power requirements of CSI feedback NNs.Multiplerate compression of CSI is considered,and to avoid storing multiple NN encoders,the fully connected layer at the encoder is reused to achieve multiple-rate compression multiples by a single NN.Next,the complexity challenges of the feedback NNs are considered,and several representative NN compression and acceleration techniques,such as NN pruning and quantization,are adopted to reduce the redundancy of CSI feedback NNs and the complexity overhead.Numerical simulations on public datasets show that the proposed multiple-rate compression scheme can reduce the NN storage overhead by more than 50% without a drop in the feedback accuracy,and NN pruning and quantization can also reduce the complexity by about 80%. |
