High Performance Precoding Based On Deep Leaning In Massive MIMO Systems

With the rapid development of the fifth generation(5^th Generation,5G)technology,mobile communication technology has entered a new era,and people also put forward higher requirements for system capacity and transmission rate.At the same time,the carbon emission of information and communication related industries will increase greatly in the future,which will be one of the main factors affecting the environment.Therefore,it is imperative to develop green communication.In 5G,the integration of multiple input multiple output（MIMO）technology and deep learning technology,as an effective solution,can greatly improve system performance,which has become a research hotspot in recent years.Therefore,this thesis studies the precoding algorithm based on deep learning in massive MIMO system and gives three algorithm schemes.The specific contents are as follows:In the massive MIMO system,traditional precoding algorithms based on optimization and iteration have many drawbacks,such as complex calculations,slow convergence,high bit error rates,and cannot support real-time applications in 5G systems.In view of the above problems,this thesis designs a precoding algorithm based on convolutional autoencoder,and improves the convolutional autoencoder,which can reduce the data dimension and network parameters,reduce the amount of calculation,and effectively reduce the training time.At the same time,a new loss function is designed to improve the network convergence speed.And use the PRe LU function to solve the problem of neuron death due to negative numbers in the input matrix.Finally,considering the influence of parameters such as different numbers of users and base station antennas on system performance,two quantization methods of1bit and Multibit digital-to-analog converters are adopted respectively.The simulation results show that the proposed algorithm has obvious improvement in energy efficiency and bit error rate performance.Under the millimeter-wave massive MIMO system,the hybrid precoding algorithm based on supervised learning needs more data sets in the process of training the neural network,so there is a problem of high training data demand.To solve this problem,this thesis designs a hybrid precoding algorithm based on unsupervised learning,which effectively reduces the number of datasets and improves the training efficiency.Conv＿net and Fully＿net network structures are built to preserve boundary information and improve feature extraction capabilities while avoiding overfitting and improving the generalization ability of the algorithm.And according to the system model and spectral efficiency expression,a loss function that is more in line with the actual situation is designed,and the Leaky＿Re LU function is used to solve the problem of gradient disappearance due to negative input.Finally,consider the influence of parameters such as different number of transmit antennas,number of radio frequency chains on system performance.The simulation results show that compared with the OMP algorithm and the PE＿Alt Min algorithm,the performance of the proposed algorithm is more obvious.In order to further improve the accuracy of precoding algorithm,reduce the training time,and solve the problem that the gradient disappears and the algorithm performance drops sharply due to the increase of network depth.A hybrid precoding algorithm based on residual learning is designed.The residual block structure is built,and every two layers of Conv＿net constitute a residual block.With the assistance of the residual block,the problem of gradient disappearance can be effectively alleviated.At the same time,a squeeze-and-excitation networks is designed,so as to learn the importance of each feature channel independently,and improve it according to this importance.Relevant features,constraints irrelevant features.The Simulation analysis show that compared with the algorithm in Chapter 4,the performance of the proposed algorithm has been further improved,and it is closer to the Optimal digital precoding performance.This thesis optimizes the accuracy,speed and complexity of the precoding algorithm,and uses deep learning technology to propose three precoding algorithm schemes based on convolutional autoencoder,based on unsupervised learning and based on improved residual network.The simulation results show that the proposed three algorithms can effectively improve the performance of the precoding algorithm,and have better performance in terms of bit error rate,energy efficiency and spectral efficiency.