论文部分内容阅读
Due to the quick growth in functionality and complexity of modern electronic devices,developing an efficient methodology of modeling complicated nonlinear circuits is becoming an attractive academic field.Deep neural networks are considered as an effective tool of modeling nonlinear systems.To precisely depict the behavior of complicated nonlinear circuits,a modeling methodology on the basis of deep feedforward neural networks is proposed in the thesis.The behavioral modeling procedure is clarified by creating the model of a push-pull PA,which is considered as a typical nonlinear electronic device.The behavioral model of the power amplifier is created on the basis of a deep feedforward neural network.The training and testing data sets are generated from Multisim simulation tool.The training and testing procedures are performed on Tensor Flow deep learning platform.Compared with the polynomial behavioral model,the proposed deep neural network model decreases the mean squared error by 13 decibel.The outcome of validation in time domain and frequency domain illustrates that the proposed deep neural network model precisely depicts the behavior of nonlinear circuits.Main contents of the thesis are organized as follows.Chapter 1 introduces the historic background and the research status.Chapter 2 describes the configuration of a deep feedforward neural network,which is used to model nonlinear circuits.The choice of activation function,loss function,and optimization algorithm are also discussed in this chapter.In chapter 3,a push-pull power amplifier is designed as a typical nonlinear circuit to be modeled.Training and testing data sets are attained from Multisim circuit simulation tool.Chapter 4describes the process of model training and validation.Chapter 5 summarizes the whole thesis and discusses the future work.The main innovation of the thesis is that a DFNN based modeling approach is proposed to characterize the behavior of nonlinear circuits.Compared with the frequently used polynomial behavioral model,the proposed DFNN model improves the fitting accuracy.