Predistortion Method Research Based On Volterra Series And Neural Network

In communication systems,the main non-linear generation and energy consumption come from RF power amplifiers.Therefore,the performance of a radio frequency power amplifier,which is an essential part of a communication system,is related to the transmission quality of the entire communication system.Because power amplifiers often have severe distortion when ensuring high-efficiency operation,digital predistortion technologies need to be adopted to linearize the signals passing through the power amplifier.Predistortion learning structures are mainly divided into indirect learning structures and direct learning structures.Predistortion models are mainly based on Volterra series and neural networks.This paper conducts an in-depth study of the direct learning structure based on Volterra series model and the networks,proposes an adaptive basis direct learning structure that automatically selects basis functions after each iteration,and a tailoring method of the neural network and a single output node neural network for predistortion.The main innovations of this article are as follows:1.The in-depth study of the negative feedback structure is proposed,and the necessary conditions for the convergence of the negative feedback structure are proposed.These conditions can ensure stable convergence to obtain the ideal predistortion signal.At the same time,based on the negative feedback structure,an adaptive basic direct learning structure is proposed.In the tests of AB and Doherty power amplifiers,it can be found that the adaptive basic direct learning structure has a more stable convergence process and the coefficients identification process is faster than the traditional direct learning structure.At the same time,the model finally obtained by the proposed method contains fewer coefficients than the original model,which saving the hardware resources.2.Apply the Group Lasso algorithm in linear regression to a neural network predistortion model which is ARVTDNN,and training this neural network by adding a penalty term to the loss function of the original ARVTDNN,so that the neuron output weights converge to 0,and the neurons that converge to 0 is the neurons that can be removed.This method can quickly screen out unnecessary neurons.Tests show that the proposed Group Lasso algorithm can reduce28%neurons without reducing performance.This method can save a lot of hardware resources,which is of great significance for the application of neural network predistortion.3.In-depth analysis of ARVTDNN,a neural network with a single output node is proposed.ARVTDNN needs to fit the I/Q signals separately,and this single-output node neural network only needs to fit one I/Q output of the original neural network to realize the modeling of the neural network predistorter.By adjusting the input vector,another output of the model can be obtained.This single output node neural network avoids redundant calculations and redundant storage in ARVTDNN,which not only improves performance,but also greatly reduces complexity.4.According to the analysis of ARVTDNN model,a parallel structure of ARVTDNN and GMP model is proposed.This parallel structure inherits the advantages of GMP model and ARVTDNN model.It not only has excellent predistortion performance,but also has a strong I/Q correction ability.