Research On Digital Predistortion Of Power Amplifier Based On Complex-valued Neural Network

In the development of wireless communication technology,the demand for linearization of power amplifier is increasing.Linearization technology for power amplifier emerges in endlessly.As one of the most effective methods,it can deal with the problems caused by nonlinear characteristic and memory effect of power amplifier.This thesis studies the linearization of power amplifier based on digital predistortion technology.Firstly,the nonlinear characteristics and memory effects of power amplifier are analyzed.The nonlinearity of power amplifier is mainly reflected in the new spectrum components generated at the output end,which can be expressed by Adjacent Channel Power Ratio(ACPR).Then,several commonly used power amplifier behavior models and their advantages and disadvantages are introduced.Among them,the generalized memory polynomial(GMP)model is the most widely used because of its excellent performance.Three kinds of digital predistortion structures are introduced.Based on one of them,the verification of digital predistortion is carried out in Chapter 5.Amplifier behavior modeling is the basis of digital predistortion,and neural network has become an increasingly attractive solution because of its good approximation ability.In this thesis,the neural network model is used as the behavior model of power amplifier,and the complex-valued neural network is introduced on the basis of the traditional real-valued neural networks.The complex-valued neural network is applied to the field of digital predistortion for the first time.The structure adopted is Fully Connected Recurrent Neural Network(FCRNN),and the Complex Real-time Recurrent Learning(CRTRL)algorithm is deduced according to the characteristics of the complex-valued neural network.Finally,the simulation results are compared with the GMP model and the Focused Time-Delay Neural Network(FTDNN)model to prove the feasibility of the complex-valued neural networks model.Then an augmented CRTRL algorithm is proposed and verified.Then,on the basis of FCRNN model,the power amplifier modeling of Complex-Valued Pipelined Recurrent Neural Network(CPRNN)model is proposed,and then expands the CRTRL algorithm to adapt to the modular structure of CPRNN.In order to improve the learning performance,variable forgetting factor is introduced into the learning algorithm.Finally,the effect of the model is verified by simulation,and the performance is improved relative to the FCRNN model.Then an augmented CRTRL algorithm is proposed and its performance is verifiedFinally,two kinds of complex-valued neural network models are used as predistorters to correct the predistortion of class F and Doherty power amplifiers based on WCDMA signals.The experimental results show that the ACPR index of power amplifier output is significantly reduced after predistortion,it shows that the digital predistortion system based on complex neural network has good linearization ability,and further verifies the feasibility of the proposed model.