Research And Application Of Large-scale MIMO Channel State Information Feedback Based On Deep Learning

As a key technology of mobile communication system,large-scale multi input multi output(MIMO)will continue to be used in the sixth generation(6G)mobile communication system.As an extension of MIMO technology,large-scale MIMO technology reduces noise and channel fading through multiple antennas,and uses the spatial dimension to increase system capacity and make communication more stable.However,large-scale MIMO systems with multiple antennas also bring some problems,such as the huge overhead of channel state information(CSI)feedback in frequency division duplex(FDD)mode.At present,artificial intelligence based on Deep Learning(DL)has achieved great success in natural language processing,target detection and other directions,which also provides a new solution for the application in the physical layer of communication.Based on intelligent communication,CSI feedback optimization of FDD large-scale MIMO system is deeply discussed in this paper.It mainly covers the following aspects:1.Aiming at the problem of poor recovery ability of data-driven CSI feedback channel,a more efficient design framework was proposed,namely,the attention mechanism network CLPNet based on multi-scale and multi-channel fusion.The framework mainly redesigns the decoder part,introduced extra-large convolutional kernel,and carried out feature fusion through multiple parallel paths.Bar convolution is added to strengthen the correlation between the Angle domain and the Delay domain.The encoder introduces a spatial attention mechanism to improve the coding ability and combines real and imaginary parts through small convolution nuclei.The simulation results show that the channel recovery capability is further improved,the complexity is greatly reduced,and the reconstruction performance is also good when the deep separation convolution layer is introduced.2.The model-driven approach is explored more deeply,the model driven feedback network AMFISTANet with adaptive slope threshold is presented.Compared with the data driven network.the model driven method makes the whole network interpretable by neural network of some parameters.AMFISTANet encoder introduces the hybrid attention mechanism CBAM.and the decoder at the base station is iteratively expanded through the soft-threshold shrinkage algorithm.By adding parallel attention subnetwork in the contraction module,the slope can be changed adaptively to improve the sample independence.The simulation results indicate that the feedback performance of AMFISTA network has been further improved,especially in the indoor environment with good recovery performance.3.Finally,feedback network assisted by RIS is researched,and a neural network assisted refinement framework AIsp is raised.The framework is based on the traditional codebook feedback method.After receiving the matrix of codebook recovery,the base station input it into the neural network to further refine and improve the mapping accuracy.The refinement network is mainly composed of residual embedding module,which is based on super resolution recovery.Simulation results demonstrate that the performance of this method is significantly improved compared with the traditional method,and it has strong practicability in the future deployment process,especially in the deployment of Internet of Things devices.