Research On CSI Feedback Technology For Large Scale MIMO Systems Based On Deep Learning

With the rapid development of wireless communication and artificial intelligence technology,the combination of artificial intelligence technology and wireless communication technology has become one of the important directions of improving wireless communication’s ability.Among the existing communication tasks,how to use deep learning(DL)technology to improve the performance of traditional communication tasks or improve the system efficiency has become the key to the development of wireless communication technology.Recently,including channel feedback,channel estimation,channel tracking and other issues,they have tried to combine with deep learning technology and have achieved good performance compared with traditional methods.Channel State Information(CSI)feedback is a very important part of the multiple-input multiple-output(MIMO)technology.How to effectively improve the feedback accuracy while reducing the computational complexity and parameter quantity has become one of the important research directions,reducing the test delay is also the direction that needs to be concerned in wireless communication.In order to solve relevant problems,this paper uses DL technology to design reasonable modules to improve CSI feedback accuracy,reduce parameters and reduce the direction of delay.The main research contents include:(1)In view of the strong correlation between the channel state response of the same antenna and the same subcarrier of the CSI matrix,a feedback neural network with the criss-cross attention module as the core is designed to better extract the row-column correlation feature of the CSI matrix,so that the base station can better recover the completed channel response matrix,and at the same time,the neural network can’t input complex numbers,A module combining the real part and the imaginary part is designed,which together constitute the encoder part of the automatic encoder to better extract the feature information of the CSI matrix for decoding and recovery.Considering that the network needs to adapt to different computing resource platforms,a variable decoder is designed.By changing the channel number of the decoder,models of different sizes are designed.The simulation shows that the network with cross attention as the core module of feature extraction can more effectively extract the features of the matrix and achieve better feedback performance at lower computational complexity through feature map visualization,feedback performance comparison and ablation experiments.(2)For the feedback network of the above criss-cross attention design,there are four main problems to be solved:1.The criss-cross attention module does not perform well in outdoor conditions.2.The full connection layer parameters account for 85%of the overall network parameters,so it is difficult to reduce the network parameters by changing the feature extraction module and decoder module.3.There is no good measurement method to measure the relationship between different channel state responses.4.Because the existing network adopts the Transformer or multi-branch structure with the full connection layer as the core architecture,the test delay cannot be reduced.To solve the above problems,we first designed a new criss-cross attention mechanism,named criss-cross attention+,which can more fully extract the characteristics of the channel state response between the relevant antennas,and since more feature information is extracted,the subsequent training curve also proves that the newly designed attention module can make the training process more stable.Secondly,we designed a new data similarity measurement method,with cosine similarity as the core,which is mainly aimed at the similarity between different channel state responses in the same channel state information matrix.According to the measured similarity information,we designed a down-sampling layer to compress the feature map reasonably,so as to keep the feedback performance unchanged as much as possible under the condition of greatly reducing the parameters of the full connection layer,Subsequent performance comparison experiments show that the newly designed feedback network can achieve the best feedback performance after the parameter quantity is reduced by 20%to 40%compared with other advanced networks.For the problem of delay reduction,because most networks adopt the full-connection layer and multi-branch structure,it is very difficult to reduce the delay.This study has found a new way,referring to the RepVGG network,through re-parameterization,the multi-branch structure is used in training,and different convolution and BN layers are combined during testing.The single-branch model used in the testing phase can reduce the reasoning speed by 40%under the condition of nearly the same feedback performance as the multi-branch network.