基于时延神经网络的射频功放行为建模的数字预失真  

Digitalpre distortion modeling of RF mower amplifier based on time delay neural network

作　　者：徐佳伟 李天凯 杨毅民 XU Jiawei;LI Tiankai;YANG Yimin(School of Artificial Intelligence,Xidian University,Xi’an 710126,China)

机构地区：[1]西安电子科技大学人工智能学院,西安710126

出　　处：《空间电子技术》2024年第6期25-30,共6页Space Electronic Technology

基　　金：国家重点研发计划项目(编号:2021YFA003400)。

摘　　要：射频放大器作为通信系统中不可或缺的功率器件通常需要兼顾高效率和高线性度的信号传输,工作在邻近饱和点的射频放大器具有较高的效率和不可避免的较差的线性度。提高线性度的一个办法是对输入功放的信号进行预失真,引入和功放相反的预失真提高线性度。传统的模拟预失真不能充分考虑到功放的记忆效应,为了解决功放半导体电路在放大信号过程中不可避免地对信号造成高阶非线性失真问题,文章提出了一种基于时延神经网络模型,以此时刻的数据和前两个时刻的数据,并提取其中的高阶非线性用于功放的高阶非线性建模。该模型考虑到功放的记忆效应带来的动态高阶非线性效应。通过对功放的输入输出样本数据集进行非线性拟合,构建四层神经网络,对神经网络进行有监督的训练,然后利用训练后的模型对数据集进行预测。初始数据的归一化均方误差为-21.83 dB,经过预失真模型处理归一化均方误差达到了-37.77 dB,实验结果证明了所提出的模型具有高的线性化能力。As an essential power device in communication systems,RF amplifiers must typically balance high efficiency and high linearity in signal transmission.Operating near saturation,RF amplifiers achieve higher efficiency but suffer from reduced linearity.One approach to improve linearity is to pre-distort the input signal,introducing an opposite pre-distortion to counteract amplifier nonlinearity.Traditional analog pre-distortion cannot fully address the memory effects inherent in amplifiers.To mitigate the inevitable high-order nonlinear distortion in semiconductor amplifiers,this paper proposes a delay neural network model that leverages current and prior data points to capture higher-order nonlinearities for amplifier modeling.This model accounts for the dynamic high-order nonlinear effects arising from amplifier memory effects.A four-layer neural network is constructed and trained in a supervised manner by fitting the input-output sample dataset of the amplifier.After pre-distortion processing,the normalized mean square error(NMSE)improved from an initial-21.83 dB to-37.77 dB,demonstrating the model’s strong linearization capability.

关 键 词：功率放大器 高阶非线性 神经网络 数字预失真 记忆效应 

分 类 号：V443[航空宇航科学与技术—飞行器设计] TP391[自动化与计算机技术—计算机应用技术]