Design Of Deep Learning Based Implicit Feedback For Massive MIMO Channel

At present,the time division duplex mode is mainly used in the commercial massive multiple-input multiple-output(MIMO)systems.Recently,the frequency division duplex(FDD)massive MIMO systems have attracted extensive attention from industry and academia,whose core challenge is that the base station(BS)needs to acquire the downlink channel state information(CSI)through feedback.With the increase of antenna dimension,the traditional feedback schemes have encountered a bottleneck,and the deep learning(DL)-based CSI feedback has gradually become a research hotspot.However,most of the existing DL-based feedback schemes are explicit feedback,and there are few researches on implicit feedback.To solve these issues,this thesis conducts in-depth research on the design of DL-based implicit feedback for massive MIMO channel.First,for the CSI feedback problem in the FDD massive MIMO systems,we summarize the DL-based implicit feedback schemes,and construct the explicit feedback architecture based on the autoencoder structure to realize the compression and reconstruction of the complete channel matrix.For explict feedback architecture design,the basic theory of DL is summarized,including three classical neural network structures and the autoencoder structure,which lays a theoretical foundation for the feedback architecture design.The principles of two traditional CSI feedback schemes,i.e.,codebook-based CSI feedback and compressive sensing-based CSI feedback,are described in detail,and the limitations of each scheme are pointed out,indicating the necessity of introducing DL technology into the CSI feedback design in the FDD massive MIMO systems.In view of the shortcomings of the traditional feedback schemes,several research directions of the DL-based explicit feedback schemes are summarized,and their design ideas and performance advantages are clarified to provide the theoretical support and technical direction for the research on DL-based implicit feedback schemes.Then,to solve the problem that the DL-based explicit feedback does not conform to the standard,we design a novel DL-based implicit feedback scheme,which retains the implicit feedback mechanism with low overhead and introduces DL technology to enhance the feedback performance.Following the exisiting standard,typical massive MIMO system models are established,and the basic theory of the Type I/II codebook is introduced as a performance comparison benchmark for the DL-based scheme.For the system with a single resource block(RB),an architecture composed of fully connected layers,called Im Csi Net-s,is designed for implicit feedback.Neural networks are used to replace the precoding matrix index(PMI)encoding and decoding modules,respectively,learning a more refined mapping relationship between the precoding matrix and the PMI.By adjusting the network width of Im Csi Net-s,an architecture named Im Csi Net-m is constructed for the system with multiple RBs.On this basis,by replacing the fully connected layers in the encoder with bidirectional long short-term memory(LSTM)networks,an architecture named bi-Im Csi Net is designed to extract the correlation between subbands as auxiliary information.According to the generated channel data and the configured simulation parameters,the overall performance of the above-mentioned networks is tested and analyzed from the perspectives of feedback overhead,reconstruction performance,and network complexity.Numerical results show that the proposed DL-based implicit feedback scheme significantly outperforms the Type I/II codebook-based feedback scheme,and extracting the subband correlation as auxiliary information can further reduce the feedback overhead when ensuring the reconstruction performance.Finally,for the time-varying channel scenarios in massive MIMO,we design a temporal correlation-assisted implicit feedback,which integrates the LSTM networks suitable for time series processing and the implicit feedback mechanism to enhance the feedback performance.The massive MIMO system model with a single RB is established,and the potential gain of the temporal correlation on CSI feedback is analyzed.The feedback process of the Type I codebook-based scheme is described,and the codebook selection schemes are elaborated to generate the codeword according to the channel as the performance benchmark.The LSTM networks are introduced into Im Csi Net-s to construct the architecture called Im Csi Net-LSTM,which can extract the temporal correlation between adjacent channels.The dataset generation method for different vehicle speeds in the time-varying channel scenario and the network parameter settings are expounded.Without and with the temporal correlation,the performances of the DL-based implicit feedback networks are analyzed,and a comparison is made with the Type I codebook-based feedback scheme.In addition,the reconstruction accuracy and the generalization of the architecture under various vehicle speeds are tested and analyzed.Numerical results show that extracting temporal correlation as auxiliary information can further enhance the implicit feedback performance,and the proposed architecture have a good generalization under different vehicle speeds by training with mixed vehicle speeds.