Research On MIMO-OFDM Signal Detection Technology Based On Deep Learning

As a key technology of 4G and 5G,MIMO-OFDM greatly improves spectrum usage without increasing system bandwidth,increasing system capacity and transmission rate effectively.In wireless communication system,channel estimation and signal detection have been considered as a long-standing issue as the wireless channel varies with time,especially with the increase of antenna dimension,and the complexity of traditional signal detection algorithm will deteriorates sharply.Nowdays,as the third wave of artificial intelligence has eroded most industries,intelligent communication combined with deep learning and wireless communication technology is considered by more and more scholars to be a research hotspot of beyond 5G wireless communication.This thesis is aimed at MIMO-OFDM communication system,using the modular application method of deep learning network to study the MIMO-OFDM signal detection algorithm.First of all,this thesis studies the key techniques of signal detection in MIMOOFDM systems.After analyzing the characteristics of MIMO-OFDM wireless channel,traditional channel estimation and signal detection algorithms are deduced and their advantages and disadvantages are analyzed.Then,the composition and optimization process of deep learning network are introduced,and the application of classical deep learning models in wireless communication systems is investigated,including signal detection,modulation identification and classification,and MIMO channel feedback.It also further explores a feasibility that replaces the functional modules in wireless communication systems with neural networks,which provides strong technical support for designing deep network models for MIMO-OFDM signal detection.Secondly,this thesis studies the MIMO-OFDM signal detection algorithm based on data-driven deep learning.For the low reliability of traditional MIMO system signal detection technology,a signal detection model based on multiple DNN networks is proposed: Multi-DNN.The model uses two fully connected neural networks cascaded that to replace the channel estimation module and signal detection module respectively,and adopt LS estimation to obtain the initial channel estimation,the signal is predicted by Multi-DNN,which further improves the accuracy of MIMO signal detection.Aiming at the problem of channel decoding in the signal detection process of OFDM system,a detection algorithm based on residual connected convolutional neural network is proposed: DCNet.The model uses one-dimensional convolutional layer and fully connected layer to perform feature extraction and signal classification,respectively,to optimize the receiver of the OFDM system globally,and output the bit signal directly.Finally,the effectiveness of the proposed two models is verified by comparative simulation experiments.Finally,this thesis studies the MIMO-OFDM signal detection algorithm based on model-driven deep learning.Aiming at large demand for labeled data and time cost during the training of data-driven signal detection model,a model-driven deep learning model CSNet for OFDM receiver is proposed.In this model,OFDM receiver is divided into channel estimation module and signal detection module.In order to improve the training speed of the deep learning network and the generalization ability of the model,linear least mean square error and forced zero detection are used to initialize the neural networks parameters of the two modules respectively.In addition,a DFT-based smoother is added to CSNet to enhance denoising,which further accelerates the training speed and detection accuracy of the DNN network.Then,the model is integrated and enlarged to the MIMOOFDM system,and comparative experiments are carried out.The simulation results show that the model-driven deep learning model CSNet has better bit error ratio performance regardless of linear distortion or nonlinear distortion.