Modeling and linearizing broad-band power amplifier based on real and complex-valued hybrid time-delay neural network  

基于实复值混合时延神经网络的宽带功放的建模和线性化（英文）

作　　者：Hui Min Zhang Xingang Zhang Meng Yu Chao Zhu Xiaowei 惠明;张新刚;张萌;余超;朱晓维(东南大学毫米波国家重点实验室,南京211189;南阳师范学院物理与电子工程学院,南阳473061)

机构地区：[1]State Key Laboratory of Millimetre Waves,Southeast University,Nanjing 211189,China [2]College of Physics and Electronic Engineering,Nanyang Normal University,Nanyang 473061,China

出　　处：《Journal of Southeast University(English Edition)》2018年第2期139-146,共8页东南大学学报（英文版）

基　　金：The National Natural Science Foundation of China(No.61561052,61701262);the Science and Technology Foundation of Henan Province(No.182102410062,182102210114);the Science and Technology Foundation of Henan Educational Committee(No.17A510018)

摘　　要：A new real and complex-valued hybrid time-delay neural network(TDNN)is proposed for modeling and linearizing the broad-band power amplifier(BPA).The neural network includes the generalized memory effect of input signals,complex-valued input signals and the fractional order of a complex-valued input signal module,and,thus,the modeling accuracy is improved significantly.A comparative study of the normalized mean square error(NMSE)of the real and complex-valued hybrid TDNN for different spread constants,memory depths,node numbers,and order numbers is studied so as to establish an optimal TDNN as an effective baseband model,suitable for modeling strong nonlinearity of the BPA.A 51-dBm BPA with a 25-MHz bandwidth mixed test signal is used to verify the effectiveness of the proposed model.Compared with the memory polynomial(MP)model and the real-valued TDNN,the real and complex-valued hybrid TDNN is highly effective,leading to an improvement of 5 dB in the NMSE.In addition,the real and complex-valued hybrid TDNN has an improvement of 0.6 dB over the generalized MP model in the NMSE.Also,it has better numerical stability.Moreover,the proposed TDNN presents a significant improvement over the real-valued TDNN and the MP models in suppressing out-of-band spectral regrowth.提出了一种新型的实复值混合时延神经网络,用于建模和线性化宽带射频功放.该神经网络包含输入信号的广义记忆效应、复值输入信号和复值输入信号模值的分数阶次,因而其建模精度显著提高.对实复值混合时延神经网络在不同扩展常数、记忆深度、神经元数和阶数时的归一化均方误差(NMSE)进行了比较研究,以建立一个能够有效建模宽带功放强非线性的基带模型—最优时延神经网络(TDNN).采用51 dBm宽带功放和25 MHz带宽的混合测试信号用于模型的有效性验证.测试结果表明,实复值混合时延神经网络相比记忆多项式模型和实值时延神经网络具有更高的建模精度,NMSE提高5 dB.此外,实复值混合时延神经网络相比广义记忆多项式模型,NMSE提高0.6 dB,并具有更好的数值稳定性.相比实值时延神经网络和记忆多项式模型,所提出的时延神经网络能够更好地抑制带外的频谱再生.

关 键 词：power amplifier neural network LINEARIZATION MODELING 

分 类 号：TN925[电子电信—通信与信息系统]