Research On Transceiver Design Of Noncoherent Massive SIMO Communication System

This paper considers that a number of transmitters simultaneously send messages to a base station over a multipath fading channel in the Internet of Things scenario.Non-coherent communication,which does not need channel estimation at the receiver,has been a popular research topic in short packet communication and low signal-to-noise ratio communication in recent years.However,the traditional noncoherent transceiver design method has high complexity,which makes it difficult to be applied to the Internet of Things scenarios with the rapid growth of users.Therefore,noncoherent transceiver design in Massive Input Multiple Output(SIMO)system based on machine learning is proposed to improve the transmission efficiency and flexibility of wireless communication system in this paper.According to different types of channel state information,two data-driven design methods of multi-user noncoherent transceivers are studied in this paper.Firstly,for noncoherent communication system under Intersymbol Interference(ISI)channel,a design method based on autoencoder is proposed in this paper.In this paper,the deep neural network is used to represent the noncoherent transceiver,and each transmitter and the receiver are respectively represented by a sub-neural network.Under the ISI chan-nel represented by Toeplitz matrix,the system coding and decoding scheme is obtained through end-to-end training.Compared with the traditional method,the design process no longer needs the knowledge of channel distribution,but only needs a limited number of channel samples.At the same time,the influence of training samples in transceiver design is discussed,and a estimation method of transmission error probability based on confidence intervals is proposed for a given number of samples.Furthermore,this paper proposes a transceiver design method based on reinforcement learning when the channel model is unknown.In this paper,independent neural networks are used to represent each transmitter and the receiver,and the decoding scheme is obtained through supervised learning of the receiver.The loss function value after training is fed back to the transmitter as the reward signal,and the parameters of the transmitter neural network are updated by using the policy gradient algorithm.When the target transmitter is trained,the parameters of the other transmitters are kept unchanged,so that the joint design of the transceiver can be completed iteratively and in rotation.Based on the theoretical analysis of the training process,this paper points out that this method is also applicable to the communication system with non-differentiable components.Finally,simulation experiments are carried out to analyze the proposed method.The noncoherent transceiver based on autoencoder successfully achieves efficient transmis-sion in typical ISI channel scenarios and is more robust than the traditional method when the channel environment changes.Compared with the end-to-end training using neural networks directly,the reinforcement learning-based design method has a higher training convergence rate and a lower performance variance.The above results show that the pro-posed noncoherent design method has significant performance gains compared with the existing methods,which provides an important reference for the further integration of machine learning in wireless communication.