Research Of Generalization And Feasibility Of Deep Learning-based Massive MIMO CSI Feedback Networks

Massive multiple-input multiple-output(MIMO)is one of the crucial technologies in the 5th generation(5G)and future mobile communications system.In Frequency Division Duplex(FDD)mode,the acquisition of the downlink channel state information(CSI)with low overhead and high precision is an important issue in the research field of massive MIMO.The development of Artificial Intelligence(AI)and its wide applications in many fields provide a new approach for solving key problems in the wireless communications system.The CSI feedback task can be completed based on deep learning(DL)scheme.The deep neural network is utilized to perform the CSI compression at the user equipment(UE)and the CSI reconstruction at the base station(BS)to achieve the high-precision CSI feedback,which has become a representative use case of AI technology applied in wireless communications.Most of the existing DL-based CSI feedback researches focus on the improvement of the CSI feedback accuracy through the designs of the network structure or the training method.With the deepening of the research,problems during the application of DL-based schemes in practical communications systems have obtained more attention of academia and industry.The feasibility and generalization capability,including model complexity,the requirement of training data,the self-adaptation ability of the model,have also become the key metrics to evaluate the performance of the model.The thesis mainly focuses on the research of feasibility and generalization capability of the CSI feedback neural network model during the application in real-world systems,and proposes the designs of feedback architecture,data classification module,and model updating scheme.According to the experimental results,the proposed designs can effectively reduce the model complexity,achieve the automatic model selection,and enhance the selfadaptivity of the network on the premise of ensuring feedback accuracy,which is helpful to solve the crucial problems and further improve the application value of the model in practical systems.The specific research includes the following three aspects:Firstly,the layer-shared CSI feedback architecture is proposed based on transfer learning and multi-task learning to solve the problem of excessive storage usage of the model at the UE.The total parameters of the network model can be reduced by reusing the model in different channel scenarios at the UE.The training procedure of the models in the feedback architecture is also designed in detail.Compared with the regular training approach which completes the model training in different scenarios independently,the proposed scheme is able to significantly reduce the parameters of the model at the UE with the guarantee of feedback accuracy.The problem of the high storage usage of the network model at the UE can be effectively solved.Secondly,the CSI data classification module is achieved by using three types of classifiers based on semi-supervised learning to solve the problems of model selection and excessive overhead caused by the large-size pre-labeled training data.The classification module is combined with the feedback network to construct the CSI sensing and feedback system,and the training procedure of the model in the system is designed.The proposed classification module can recognize the channel scenario precisely according to the features of the CSI data,which is helpful to improve the performance of the feedback model and achieve the selection of the model of encoder at the UE and decoder at the BS.Both the feedback accuracy and the feasibility of the AI scheme can be further improved.Thirdly,to solve the problem of poor adaptation ability of the model in the time-varying channel scenario,the online learning-based CSI feedback model updating approach is proposed to ensure the feedback performance of the model in unknown channel scenarios with low cost.During the application of the model in practical systems,the small-size CSI data collected at the UE is utilized to complete the local update of the parameters of the model,and the model at the BS can be replaced by the updated model through the transmission of the parameters between the UE and the BS.The real-time update of the feedback model can effectively reduce the impact of the performance degradation caused by the mismatch between the feedback network model and the feedback data,significantly improve the adaptation ability of the model under the time-varying channel environment,and further improve the generalization capability of the AI scheme.