Research On Neural Network Based OFDM System Combating Multipath Fading

With the rapid development of wireless communication,Orthogonal Frequency Division Multiplexing(OFDM),as one of the key technologies in physical layer,has been studied in a lot of ways to combat multi-path fading.In recent years,Deep Learning(DL)technology has developed rapidly and achieved very good results in speech processing and image recognition.Meanwhile,the universality of DL enables it to be organically combined with the field of communication.This paper studies the problem of OFDM system combating multipath fading,and explores the possibility of deep learning technology to help improve the performance of communication systems.Main works are summarized as follows:Firstly,in the simulation of the spherical code index modulation system,aiming at the problems of numbers of idle subcarriers and low spectrum utilization rate,pilot assisted channel estimation technology is introduced to the spherical code index modulation system,and a scheme of channel estimation using neural network to detect pilot position is proposed.In this scheme,a simple fully connected neural network is used to extract the location of pilot frequency,and the extracted pilot frequency is used for channel estimation.This paper describes in detail the training of pilot detection network in the spherical code index modulation system and the concrete implementation of channel estimation.Through simulation analysis,the spherical code index modulation system can effectively reduce the bit error rate of system signal detection and improve the performance of anti-multipath fading after selecting appropriate pilot number and auxiliary channel estimation of pilot frequency.In view of the fact that traditional high-precision channel estimation algorithms always require the statistical characteristics of channels,this paper proposes a Rayleigh channel estimation scheme based on Condition Generative Adversarial Network(CGAN),which solves the problem that channel prior information is difficult to obtain,in which the initial channel response matrix obtained by the least squares algorithm is used as the initial input of CGAN,and the received signal is used as the conditional input of CGAN.The scheme proposed in this paper improves the structure of the CGAN generator,uses the time domain receiving signal as the conditional input,and uses the deep complex neural network to convert it into the frequency domain for secondary stitching,which can make up for the missing channel features during the training process and improve the accuracy of the model.After simulation analysis,estimation performance of the scheme adopted in this paper is better than the the traditional LMMSE algorithm,and combined with the spherical code index modulation system,the ability of the communication system combating multipath fading can be further improved.