卫星通信功率放大器的预失真模型研究

发布时间：2018-04-27 20:04

本文选题：功率放大器 + 数字预失真　；参考：《北京交通大学》2017年硕士论文

【摘要】：随着通信行业的不断发展,卫星通信的重要性日益明显,可靠性是其最显著的特点。卫星通信系统中的功率放大器对信号质量有很大的影响,当输入信号功率过大时,会产生非线性失真,不仅效率低下还会影响相邻频域,因此功放的线性化技术势在必行。本文着重介绍了数字预失真技术对改善功放非线性失真的重要性,并对预失真模型进行了深入的研究。本文的主要工作与创新点如下:(1)以自适应理论为基础,建立Volterra级数预失真系统,使用最小均方算法进行参数提取,信号经过预失真后的邻道干扰有12dB左右的改善。建立记忆多项式预失真系统,使用递归最小二乘算法进行参数提取,信号经过预失真后的邻道干扰改善约14dB。(2)重点研究了人工神经网络类的BP神经网络和RBF神经网络。建立RBF神经网络的信号预测模型,使用K-means聚类的方法得到径向基中心,并利用LMS算法进行权值的更新,实验证明在对功放建模时,使用RBF神经网络比Volterra级数和记忆多项式具有更高的精确度。通过分析复数域函数的特性,阐述了神经网络中激励函数的局限性,提出了基于BP神经网络的DUAF结构并建立预失真系统。仿真结果表明,信号经过预失真后的邻道干扰改善约7dB,其预失真效果验证了 DUAF结构处理复数的局限性。(3)重点研究复数域的神经网络模型。建立全连接递归神经网络(FCRNN)预失真系统,使用RTRL算法进行参数提取。仿真结果表明,在处理复数信号时,FCRNN相比于DUAF结构具有更高的精确度。(4)提出了改进的短时记忆递归神经网络(STMRNN)。FCRNN模型的神经元反馈信号采用全连接的方式,其模型复杂度较大,而STMRNN模型将FCRNN模型中的输出层反馈信号改用短时记忆的方式,在保证精确度的同时能减少模型复杂度。(5)提出了改进的全反馈短时记忆递归神经网络(AFSMRNN)。AFSMRNN模型将STMRNN模型中的隐含层反馈信号改用短时记忆的方式,进一步地降低了模型复杂度,并且能达到与记忆多项式相同的精确度。
[Abstract]:With the development of communication industry, the importance of satellite communication is becoming more and more obvious. The power amplifier in the satellite communication system has a great influence on the signal quality. When the input signal power is too large, it will produce nonlinear distortion, which will not only affect the efficiency but also affect the adjacent frequency domain, so the linearization technology of power amplifier is imperative. In this paper, the importance of digital predistortion technology in improving nonlinear distortion of power amplifier is introduced, and the predistortion model is deeply studied. The main work and innovation of this paper are as follows: (1) based on the adaptive theory, the Volterra series predistortion system is established, and the parameters are extracted by using the least mean square algorithm. After the signal is predistorted, the adjacent channel interference is improved by 12dB or so. A memory polynomial predistortion system is established and the parameters are extracted by using the recursive least square algorithm. The BP neural network and the RBF neural network of the artificial neural network are studied emphatically after the signal is improved by the adjacent channel interference (about 14 dB.m-2) after the signal is predistorted. The signal prediction model of RBF neural network is established, the radial basis function center is obtained by K-means clustering method, and the weight is updated by using LMS algorithm. Using RBF neural network is more accurate than Volterra series and memory polynomial. By analyzing the characteristics of complex function, the limitation of excitation function in neural network is expounded, and the DUAF structure based on BP neural network is proposed and the predistortion system is established. The simulation results show that the signal is improved by about 7db after predistortion, and its predistortion effect verifies the limitation of DUAF structure in complex number processing. (3) the neural network model in complex domain is mainly studied. A fully connected recurrent neural network (RNN) predistortion system is established and the parameters are extracted by RTRL algorithm. The simulation results show that FCRNN has a higher accuracy than DUAF structure in processing complex signals.) an improved STMRNNN-. FCRNN model is proposed. The neural feedback signal of STMRNNU. FCRNN model is fully connected, and the complexity of the model is high. The STMRNN model changes the output layer feedback signal in the FCRNN model to short-term memory. This paper presents an improved full feedback short time memory recurrent neural network (AFSMRNN + AFS MRNN), which converts the hidden layer feedback signals in the STMRNN model to short time memory, and further reduces the complexity of the model. And can achieve the same accuracy as memory polynomial.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN722.75;TN927.2

【参考文献】