Research On End-to-End MIMO System Based On Deep Learning

With the development of society,people put forward higher and higher requirements for communication quality,and the design of wireless communication system has played a vital role.Traditional communication systems are generally designed based on modules,and the modules are individually optimized,which cannot guarantee the optimal performance of the entire system.The end-to-end communication system based on deep learning is used to jointly optimize the entire system through the back propagation algorithm of the neural network,so that the error between the input signal and the output signal of the system is minimized,thereby improving the reliability of the entire end-to-end communication system.In order to improve the reliability of the communication system based on the autoencoder,this thesis combines the wireless communication system with deep learning,and conducts the following research on the multi-input multi-output(MIMO)system based on deep learning.First of all,this thesis analyzes the principle of a communication system based on deep learning,and verifies the feasibility of implementing a communication system through an autoencoder.In order to improve the reliability of the communication system based on the autoencoder,the space-time coding technology is introduced into the communication system of the autoencoder,and the network of the autoencoder is designed.Thus,the scheme of the end-to-end MIMO communication system based on the autoencoder is proposed.This scheme converts the binary bit stream into a one-hot vector,and then inputs it into the autoencoder communication system.The multi-layer neural network of the system transmitter modulates the signal.The modulated signal is transmitted by multiple transmitting antennas through a symbol transmission scheme that conforms to the space-time coding rules,The receiver receives the signal through multiple receiving antennas and restores the signal to a binary bit stream.This scheme enables communication system based on the autoencoder obtain antenna diversity gain.Multiple sets of simulation experiments show that compared with the traditional MIMO communication system and related work of predecessors,the MIMO communication system based on the autoencoder proposed in this thesis has better symbol error rate.Based on this,the limitations of the end-to-end MIMO communication system based on deep learning are analyzed.In the scenario where each symbol carries a large number of binary information bits,the symbol error rate of the system will decrease under a high signal-tonoise ratio.The reason for this problem is that when each symbol carries a large number of binary bits,the dimension of the one-hot vector input to the autoencoder is too large.In response to this problem,a scheme optimized for one-hot encoding is proposed,which groups binary information bits for one-hot encoding.This scheme greatly reduces the dimension of the one-hot vector.Simulation experiments show that symbol error rate of the end-to-end MIMO system with optimized one-hot coding scheme has been significantly improved,and the training speed of the system has also been greatly improved.However,its symbol error rate still has a gap compared with the traditional MIMO system.In order to more effectively improve the reliability of the system,the network structure of the system is optimized.At the same time,a residual network is introduced to avoid gradient disappearance caused by network deepening,and an end-to-end MIMO system based on a deep residual autoencoder is proposed.The simulation results show that symbol error rate of the system is significantly improved compared to the end-to-end MIMO system that only optimizes one-hot encoding,and it has better symbol error rate than the traditional MIMO system.