Research On Multicarrier Waveform Classification And Physical Layer Transmission System Based On Deep Learning

The traditional communication physical layer technology involves the design of modules such as channel modeling,optimal sampling scheme,and signal detection to ensure error-free transmission of data.Because channel models are difficult to describe with precise mathematical models in scenarios such as underwater communication and molecular communication,systems such as millimeter-wave communication and ultra-densification networks need to process massive amounts of data in real time,the limitations of traditional physical layer technologies have emerged.Therefore,the application of machine learning and deep learning tools with massive data processing,deep feature mining,and general function fitting capabilities to the communication physical layer has become an important research topic.Based on this,this thesis studies the multi-carrier waveform recognition technology based on deep learning and the physical layer transmission system designed using deep learning.For the multicarrier waveform recognition problem of orthogonal frequency division multiplexing(OFDM),universal filter multicarrier technology(UFMC)and the filter bank multicarrier with offset quadrature amplitude modulation(FBMC/OQAM),this thesis proposes a multicarrier recognition system based on convolutional neural networks(CNN).In this system,principal components analysis(PCA)is introduced to pre-process the received signal to suppress the additive white Gaussian noise(AWGN),and the one-dimensional amplitude data of the signal is used as the input of the CNN.Through theoretical analysis and data simulation,it is proved that the system proposed in this thesis obtains a high multicarrier signal recognition rate under the channel conditions of AWGN,attenuation,frequency offset and timing offset.When the signal-to-noise ratio is greater than 9d B,the recognition rate reaches 90%,and the recognition rate performance is better than the CNN multicarrier recognition system with in-phase / quadrature(I / Q)data as input.On the other hand,For the physical layer transmission system designed using deep learning,the existing physical layer transmission system based on autoencoders has poor block error rate performance in multipath fading channels.This thesis proposes an autoencoder network system combined with auxiliary sequences.In this system,the transmission signals after passing through the channels are spliced together with the auxiliary sequence and sent to the receiving end,so that the receiving network automatically extracts channel information from the auxiliary sequence and performs signal reconstruction operations.Simulation results show that under the conditions of Rayleigh fading channel and Rice fading channel,the system can achieve better block error rate performance than the existing autoencoder-based communication systems.When the channel conditions are worse,the system is more robust and has a more flexible network structure.