Research On Massive MIMO Signal Detection Algorithm Based On Deep Learning

With the continuous development of the fifth generation(5G)mobile communication,large-scale multiple input multiple output(MIMO)technology can improve the spectral efficiency of the system and ensure the reliable transmission of data.It has become a research hotspot and achieved remarkable results.However,the application of massive MIMO technology has brought great challenges to signal detection.With the increasing number of antennas,the traditional signal detection algorithms have the problems of high computational complexity or poor bit error rate performance,which restricts the development of massive MIMO technology to a certain extent.To solve the above problems,this thesis studies the signal detection problem in massive MIMO system.The main research contents are as follows:1.The traditional Conjugate Gradient(CG)iterative algorithm can get better performance in signal detection,but it requires a large number of iterations,which undoubtedly increases the complexity and cost.This thesis combines CG iterative detection algorithm with in-depth learning,and presents a model-driven CG-DNN signal detection algorithm.By selecting the appropriate trainable parameters,the CG iteration algorithm with trainable parameters is expanded into a deep neural network,and the optimal parameters at each level are found through network training.The convergence of the algorithm is improved by an initial value iteration strategy based on eigenvalue estimation.The numerical simulation results show that the complexity of CG-DNN is low and its performance is comparable to the existing massive MIMO signal detection algorithms.2.Traditional Successive Over Relaxation(SOR)iteration algorithm has performance and convergence problems related to relaxation parameters.This thesis combines SOR iteration algorithm with model-driven deep learning method,sets relaxation parameters as learnable parameters,and presents a SOR-Net deep learning network.Each layer of the network will find the optimal relaxation parameters through training.Appropriate relaxation parameters can effectively improve the SOR algorithm convergence rate and improve the performance of the algorithm.In order to solve the problem that gradients disappear during training,the residual structure is set and the residual coefficient is also set as a learning parameter to improve the performance of the algorithm.The simulation results show that the performance of the SOR-Net algorithm is better than that of the original SOR algorithm.