Research On Predistortion Modeling Of Power Amplifier Based On GAN Neural Network

In the coming 5G era,people place high hopes on 5G communication,and the communication network with high speed,low delay and high bandwidth has become an inevitable demand for the development of 5G wireless communication.However,the communication system with high broadband and high rate needs to use high-order modulation,which has strict requirements on the linear characteristics of the communication system.Due to the nonlinear characteristics of the communication system itself,the nonlinear distortion problem exists,and the nonlinear compensation is needed to improve the efficiency.Nonlinear characteristics are one of the main defects of power amplifier when it is operating near the saturated zone.The nonlinear problems will change the original signal characteristics,such as gain compression,spectrum regeneration of adjacent channels and in-band distortion,which will lead to the performance degradation of the communication system.The research and realization of the linearization and digital predistortion technology of power amplifier is of great significance to the further development of the green communication concept and the energy conservation and emission reduction.Due to 5G mobile broadband and speed will have a great promotion,if according to the traditional way to handle the predistortion,frequency band width of predistortion is several times the bandwidth of the signal,the need of AD sampling rate high,digits long,signal dynamic range is wide,such devices are expensive,power consumption is very high,device of energy consumption is higher than linearization of energy saving,predistortion,there is no practical significance.Due to the input and output of the power amplifier function can be used to express,and the neural network can take advantage of a large number of training data,approximate any function expressions,and neural network can be effectively compensated amplifier memory effect,so in order to compensate the nonlinear distortion of ultra wide band amplifier,using neural network to predistortion power amplifier modeling is a potential option.At present,many scholars use neural networks to conduct pre-distortion modeling,which has some advantages and disadvantages.For example,RVTDNN(Real Value Time Delay Neural Network)is simple in structure,but the update iteration speed is slow,which is prone to the risk of overfitting and gradient disappearance.RBFNN(Radial Basis Neural Network)has the advantages of simple structure,less computation,and is similar to RVTDNN in that it is difficult to accurately characterize the strong dynamic characteristics of high broadband rf power amplifier.LSTM(Long Short-term Memory)uses three door units for training,which determines what information should be discarded and what information should be kept for a Long time.It can transfer the short-term state and value of the unit,which has Memory effect and can be suitable for the pre-distortion modeling of power amplifier under various conditions.In order to achieve high broadband with strong dynamic nonlinear power amplifier predistortion modeling,this paper will research a new model,the model using the LSTM and GAN combined neural network,can effectively solve the gradient disappeared and fitting,and has the ability of infinite modeling,both has good versatility,and can represent memory effect,can be strong for dynamic modeling nonlinear power amplifier.Based on LSTM neural network and GAN neural network,the predistortion modeling of two kinds of power amplifier with different characteristics was carried out.The pre-distortion modeling effect was visually and qualitatively observed using AM/AM and AM/PM characteristic graphs,and the modeling accuracy was quantitatively measured using NMSE(Normalized Mean Square Error)and power spectral density.Predistortion modeling results show that the LSGAN(Least Squares GAN)and SNGAN(Spectral Normalized GAN)modeling effect is good,they are effectively inhibit out-of-band spectrum regeneration,improve the nonlinear distortion of power amplifier,the strong dynamic nonlinear power amplifier has very good effect,so using GAN neural network was used to model the predistortion has great potential.