Research On User Detection And Channel Compression Feedback Technology Based On Deep Learning

In intelligent communication technology,machine learning,deep learning and other algorithms are used as tools to solve the physical layer and link layer problems of wireless communication.At present,the deep learning and machine learning related software tools represented by Tensorflow,Pytorch and Mxnet have developed to a very mature stage.They all provide very rich and convenient interfaces for researchers and engineers.At the same time,deep learning technology has been proved to be more efficient than traditional algorithms in various fields.The first chapter introduces the basic principle and application of the deep neural network,and intro-duces the current application scenarios of deep learning in the communication system.In the communication technology,the traditional algorithm is usually based on some assumptions to derive the optimal solution in a certain situation,but in the actual situation,these assumptions are usually not tenable.Therefore,the data-driven method,represented by deep learning,can avoid the artificial construction of the model hypothesis of the actual situation,and directly start from the actual data to find the optimal solution.Such new methods also gradually show their superior performance in the field of communication.In the second chapter,we consider a standard sparse code division multiple access(SCMA)detection system.In the second chapter,we study the SCMA detection system in the AWGN channel.Using the traditional message propagation algorithm(MPA)as the reference,we propose a detection method based on the sparse connection deep neural network.In the second chapter,the network parameters are initialized to the same value to achieve the same performance as MPA,and then the gradient descent method is used to optimize the network parameters so that the network can get additional performance gains.When the number of iterations of MPA is 2 and the number of blocks of neural network is 2,and the bit error rate is 10^-2,the neuralnetworkmethodhasagainof0.3d B.Then,in the third chapter,the channel state information(CSI)compression feedback problem of a multi-input single-output(MISO)system is considered.The third chapter discusses the codebook based method and the performance of neural network based method in CSI compression feedback.In the third chapter,the multi-layer convolution neural network is used to extract the channel features,and then the recurrent neural network is used to capture the time series characteristics of the channel and compress the features.In the cost2100 indoor channel scenario,when 16 bits are allocated to each multipath,the method based on the recurrent neural network proposed in chapter 3 has more than 3d B mean square error(NMSE)performance gain compared with the codebook based method and the original convolutional neural network.In addition,the third chapter further studies the influence of different structure settings such as recurrent neural network,full connection network and residual network on the computation efficiency and performance of the network,and finally gives two different structures of residual recurrent neural network,which are applicable to high performance requirements and low complexity requirements,respectively,and provides multiple schemes for different communication scenarios technical reference.In the fourth chapter,aiming at the problem that the quantization process is not differentiable,the quan-tization network training method is introduced and the fixed value is proposed to replace its derivative,so that the designed bitquantization network can use the gradient descent algorithm in the CSI compression feedback task for end-to-end training.In chapter 4,we propose a bit level optimization network structure for chan-nel information compression feedback,which compresses the channel information directly into the final bit stream for feedback transmission.Compared with other methods based on compressed sensing algorithm,the method of bit information generated by bitquantization network has nearly 4 d B gain in NMSE and about 5d B gain in bit error rate.The fifth chapter summarizes the whole thesis and points out the possible research direction in the future.