Research On Modulation And Demodulation Technology Of Visible Light Communication Based On Deep Learning

In recent years,with the increasing demands for bandwidth in RF communication,the limited spectrum resources are becoming more and more tense.Meanwhile,the power of communication equipment is also expanding.In order to alleviate these problems,researchers focus on developing emerging communication technologies with the advantages of low power consumption,low delay and high rate.Visible light communication(VLC)technology is one of them.It has the advantages of energy saving,environmental protection,no occupation of traditional RF spectrum resources,low delay,high transmission rate and no electromagnetic radiation.At present,the mainstream visible light communication technology adopts light emitting diode(LED)and photo diode(PD)as signal transmitter and receiver respectively.With a variety of new modulation and demodulation technologies being proposed,visible light communication has encountered the bottleneck of system capacity,in which the nonlinear characteristics of visible light channel is the key factor restricting its further improvement of spectrum efficiency.Based on the powerful deep learning technology,this paper studies the nonlinear modeling,nonlinear equalizer and detection technology of visible light channel,and designs the constellation with high spectral efficiency through probability and geometric shaping.The specific research contents of this paper are as follows:Firstly,the basic principle of visible light communication system is introduced,and the key devices and circuits are described in detail.Then,the basic model of visible light channel is briefly analyzed,and the time-domain response and amplitude frequency response of measured channel impulse based on traditional linear model are given.With the continuous improvement of visible light transmission rate,visible light communication is more and more restricted by bandwidth and nonlinearity.Therefore,this paper deeply studies the nonlinear model of visible light communication channel and proposes two modeling methods based on deep neural network.A large number of measured and simulated data show that the neural network channel model has better fitting performance and lower computational complexity than the traditional Volterra series model.Since the nonlinear distortion and inter-symbol interference of the visible light channel originate from multiple links in the entire system(including power amplifiers,LEDs and PDs,etc.),the visible light system channel no longer exhibits simple linear frequency selection characteristics.Therefore,the traditional channel Estimation and equalization methods are difficult to achieve better performance.This paper studies the equalization and detection technology based on deep learning,and designs three different types of neural networks based on fully connected neural network,convolutional neural network and recurrent neural network for equalization and detection.Simulation results obtained using measured data show that the deep neural network has far superior performance than traditional equalizers.In order to further improve the spectral utilization of modulation symbols in the visible light channel,this paper focuses on probability shaping and geometric shaping based on deep learning.A Bit-wise Autoencoder(Bit-AE)network is proposed to design the probability distribution and geometric shape of constellation diagrams with the optimization objective of maximizing bit-metric mutual information.The Bit-AE network consists of multiple parts,which simulate the source,modulator,channel and demodulator in the communication process.The constellation output by Bit-AE network optimization has the characteristics that the probability distribution is close to Gaussian-like distribution,the symbol mapping of constellation points conforms to the Gray mapping criterion,and the complete constellation diagram is copied from the sub-constellation diagram in a quadrant in an axisymmetric manner.Therefore,the constellation can be directly applied to the Bit-Interleaved Coded Modulation(BICM)system through the Probabilistic Constellation Shaping(PCS)scheme.The simulation results show that the constellation designed by the Bit-AE network has a significant shaping gain compared with the constellation without shaping optimization and the constellation optimized by traditional probability shaping.