Investigation Of Neural Network-Based Nonlinear Equalization And Performance For Optoelectronic Devices

In recent years,the rapid development of new multimedia services has raised higher requirements on the transmission rate and capacity of wireless communication,which has also exacerbated the shortage of spectrum resources in traditional radio frequency communications.Visible light communication(VLC)technology has gradually received more attention because of its rich spectrum resources,immunity to electromagnetic interference,and the combination of lighting and communication.However,in VLC systems,due to the 3dB bandwidth limitation of the light emitting diode(LED)sources and the nonlinear electro-optic response characteristics,serious memory nonlinear distortion is generated,which greatly limits the transmission performance of a high-speed VLC system.How to effectively compensate for the LED nonlinearity of memory produced by optoelectronic devices has become a key issue in the research of high-speed VLC.This thesis presents both theoretical and experimental research on how to design a low-complexity high-performance nonlinear equalizer,and how to verify the feasibility of the designed high-performance nonlinear equalizer in a real-time VLC system such as an FPGA-based transceivers.First of all,in view of the existing defects in the performance and complexity of the existing LED nonlinear compensation technical solutions,the author proposes a post-equalization solution based on a lightweight neural network.Through Taylor series expansion of the classical nonlinear activation function and the expansion of the network forward propagation model,the parameter relationship between the equalizer signal model based on the neural network and the Volterra series model.After that,the nonlinear compensation performance of three kinds of neural networks using Sigmoid,Parametric Rectified Linear Unit(PReLU)and polynomial activation function under different structural parameters are compared in experiment to select the optimal activation function and network structure.After that,the unstructured pruning strategies are applied to the three neural networks with different activation functions to further reduce the parameter volume of the network.In order to verify the nonlinear compensation effect of the designed neural network equalizer,the author builds a high-speed VLC offline experimental system based on carrierless amplitude-phase modulation(CAP),and uses an ordinary commercial phosphorescent white LED with its 3dB modulation bandwidth of only 10.8MHz to achieve a transmission rate of 1.5Gbps at a distance of 3m.Compared with the classical Volterra series and memory polynomial,the parameter volume of the designed equalizer based on Polynomial Activation Neural Network(PANN)are reduced by 40.7%and 7.9%,and the computation is reduced by 53.1%and 17.1%,whilst the Error Vector Magnitude(EVM)performance is successfully reduced by 0.68dB and 1.34dB respectively.Therefore,the advantages of the designed neural network-based equalizer for nonlinear equalization in high-speed VLC systems are verified.Based on the verified designed neural network-based equalizer,the huge delay caused by the need to retrain the neural network due to the change of VLC channel parameters is expected to be reduced.Given the characteristic that the nonlinearity in the VLC system mainly comes from the LED light source at the transmitter,the author designs a digital pre-distortion(DPD)scheme based on indirect learning structure and neural network.After the neural network effectively compensating the memory nonlinearity of the LED at the transmitter,a linear equalizer with extremely low complexity can be used at the receiver to track VLC channel parameter changes in real time.After that,in order to further verify the feasibility of the neural network-based DPD scheme in real-time systems,the author implements the deployment of neural network-based digital pre-distorter in a CAP VLC real-time system based on a Field Programmable Gate Array(FPGA).Comprehensive analysis of key indicators such as the real-time equalization performance,resource occupancy rate and signal delay of neural network-based DPD are made,based on which the neural network-based DPD in real-time system is explored.The experimental results with the built CAP-VLC real-time system show that the neural network-based DPD successfully reduced the error vector magnitude(EVM)by 6.6dB compared to the classic Wiener linear optimal equalizer with a low resource occupancy rate of about 3 to 4 times that of the linear equalizer.The results verify the feasibility of neural network-based DPD in real-time VLC systems.