【作者】 陈林；

【导师】 喻梦霞；

【作者基本信息】 电子科技大学 ， 无线电物理， 2021， 硕士

【摘要】 随着射频无线通信逐渐系统集成化、尺寸小型化、功能多样化,导致射频电路的结构复杂和工作频率升高。由于传统模型是基于Maxwell方程的全波分析法进行建模,电路模型消耗的计算资源增多,计算速度也明显变慢,使得设计人员对电路的仿真分析效率变低。因此,传统的射频电路建模分析方法已经逐渐无法满足电路高效快速的仿真分析设计需求。为了适应人们对高性能的射频电路模型日益迫切需求,基于人工神经网络建模方法逐渐受到人们重视且研究日益活跃。神经网络能够映射复杂的非线性输入输出关系,只要完成神经网络模型的学习,模型就能精确快速地输出响应,而无需了解其中的专业技术知识,是一种理想的建模分析方法。本文主要基于人工神经网络方法对建立高性能的射频电路模型技术进行相关研究,重点研究了数据的预处理技术、神经网络建模方法、神经网络过拟合方法以及智能建模方法。在数据预处理探究过程中,探究常用的数据转换、相似分析和数据降维方法。相似分析,结合了皮尔森相关系数,量化输入变量与输出响应之间的相关程度,能够排除与输出响应无关的输入变量。数据降维,探究了主成分分析(PCA)和基于核函数的主成分分析(KPCA)基本原理及其计算步骤,能够降低输入向量的维数,提高数据质量,也能够降低网络的复杂度。在神经网络建模探究中,深入学习了常用的反向传播神经网络(BP)、径向基神经网络(RBF)和支持向量机神经网络(SVM)的结构和计算步骤。此外,研究了神经网络过拟合原因以及应对办法,并提出一种融合随机激活(Dropout)和L2正则的新型正则化方法,该方法能够进一步消除模型过拟合现象,并结合实例验证了这种方法的有效性。最后,结合数据处理和人工神经网络的智能建模方法进行了研究。介绍了粒子群优化算法(PSO)的基本原理,利用该算法设置神经网络的初始值,能够加快神经网络的收敛。论文提出了一种新颖的智能建模方法,该方法利用数据预处理方法处理原始数据,提取原始数据中的有效信息,利用粒子群算法寻找全局最优的神经网络初值,网络只需略微调整就能获得最终解,从而实现神经网络的高效快速的学习过程。通过对数据预处理、网络模型和智能优化算法的分析研究,使得基于人工神经网络方法在射频电路建模有了更加深入的研究,为今后建立更加理想的射频电路模型奠定了基础。更多还原

【Abstract】 With the gradual development of wireless communication and RF technologies,the system of RF circuits becomes integrated,the size becomes miniaturized,and the functions become diversified,which leads to more complex structure and higher working frequency of RF circuit.According to Maxwell full wave analysis,the traditional modeling methods will consume more computing resources,increase long computing time and even cannot calculate,which makes the efficiency of circuit simulation analysis lower for designers.Therefore,the traditional RF circuit modeling and analysis methods have gradually failed to meet the requirements of efficient and fast simulation analysis and design of circuits.In order to meet the increasing demand for high-performance RF circuit models,artificial neural network(ANN)based modeling methods have been increasingly valued and studied.Neural networks can map complex non-linear inputoutput relationships.As long as the learning of the neural network model is completed,the model can output the response accurately and quickly without knowing the technical expertise.It is an ideal modeling analysis method.This thesis mainly studies the high performance RF circuit model technology based on the artificial neural network method,focusing on the data preprocessing technology,the neural network modeling method,the neural network overfitting method and the intelligent modeling method.In the process of data preprocessing exploration,the commonly used preprocessing methods such as data conversion,similarity analysis and data dimension reduction are analyzed.Similarity analysis,which combines Pearson correlation coefficient,quantifies the degree of correlation between input variables and output responses,and excludes input variables that are independent of output responses.To reduce the dimension of data,the basic principles and calculation steps of principal component analysis(PCA)and kernel principal component analysis(KPCA)are explored.These methods can not only reduce the dimension of the input vector,but also improve the data quality and reduce the complexity of the network.During the ANN modeling,the common BP,RBF and SVM network structure and calculation steps are deeply learned.In addition,the reasons for and solutions to over-fitting of the neural network are studied,and a new regularization method combining Dropout and L2 regularization is proposed.This method can further eliminate the over-fitting of the model,and its validity is verified by an example.Finally,the intelligent modeling method is proposed,which combine data processing and ANN methods.The PSO algorithm is used to set the initial value of the neural network,which can speed up the convergence of the neural network.A novel intelligent modeling method is presented,which uses data preprocessing to process the original data,extracts the valid information from the original data,and uses particle swarm algorithm to find the global optimal initial value of the neural network.The network only needs a slight adjustment to obtain the final solution,thus achieving an efficient and fast learning process of the neural network.Through the analysis and research of data preprocessing,network model and intelligent optimization algorithm,the artificial neural network method has been further studied in RF circuit modeling,which lays a foundation for establishing a better RF circuit model in the future.更多还原