Neural Networks Based Channel Estimation And Signal Detection Of MIMO-OFDM System

Multiple input multiple output(MIMO)and orthogonal frequency division multiplexing(OFDM)technologies,as key technologies under 4G,have been fully applied and developed.And with the popularity of 5G technology,it is expected to continue to be applied to new technologies,but also faces new challenges.As an important branch of artificial intelligence,deep learning has been applied in many fields,such as object detection,text analysis,and image processing.This thesis focuses on combining wireless communication with deep learning,especially in channel estimation and signal detection in MIMO-OFDM systems.In the process of research,the corresponding designs for OFDM signal processing and discrete IQ signal detection are put forward in turn.In addition,the design of peak-to-average ratio reduction in massive MIMO-OFDM systems is also proposed,and the feasibility of applying compressed sensing(CS)technology in communication is preliminarily discussed.First,in order to reduce the complexity of channel estimation and improve the accuracy of signal detection,and at the same time make full use of the latest achievements in current scientific and technological research,a new convolutional neural network(CNN)-based OFDM system is proposed for channel decoding.The block-structured communication system,as the traditional structure of the OFDM system,usually includes processes such as channel estimation,equalization,and demodulation.Channel estimation algorithms include least square(LS)and minimum mean square error(MMSE).Among them,the LS algorithm is the simplest with low algorithm complexity but poor accuracy,and the estimation results are usually affected by noise.The MMSE method considers the influence of noise on the demodulation but has high accuracy,and the statistical characteristics of the channel are used as prior knowledge.Traditional methods are gradually unable to meet the requirements of today’s millimeter wave communication in terms of accuracy or complexity.This thesis adopts deep learning tools to treat the channel problem as an autoregressive problem and the received signal as a one-dimensional array.A convolutional network is used for feature extraction,and then a fully-connected(FC)layer is used for classification to obtain soft-bit signals.The simulation results show that the method is comparable to the LS method in complexity and better than the MMSE algorithm in performance.Secondly,as a common architecture in digital communication,the main advantage of IQ modulation and demodulation is that it can combine two signals into one at the transmitting end,and divide it into two signals at the receiving end for separate processing.We propose to complete channel estimation and signal detection in traditional communication systems based on deep learning networks,especially long short-term memory(LSTM)networks,in an end-to-end mode.The design combines advanced deep learning structures and data-driven models to process complex-valued signals.Then,the peak-to-average ratio(PAPR)problem,as one of the most unfavorable factors in multi-carrier transmission systems,severely limits the performance of the transceiver.In this thesis,we propose a design that combines traditional partial transfer sequence techniques with deep learning,namely the peak-to-average ratio reduction network(EPRNet).By incorporating traditional scrambling techniques,EPRNet is designed to be model-driven scheme.It outputs the optimal phase rotation factor with low computational complexity and high accuracy.The design outperforms the traditional method in the evaluation indicators of complementary cumulative distribution function(CCDF)and bit error rate(BER).The simulation results show that the EPRNet has an average performance improvement of 2dB compared with other methods.Finally,in order to reduce the problem of large pilot overhead in high-dimensional signal estimation in the uplink,we propose a MIMO-OFDM channel estimator based on compressed sensing technology and DNN.The design combines the advantages of low sampling frequency of CS technology and superior performance of LSTM technology,and can perform channel estimation and signal detection with low complexity and high accuracy.