Research On Modeling Method Of Packaged Transistor Based On Neural Network

With the rapid development of electronic technology,research on device modeling methods has become an important development direction in the field of microwave design.Packaged transistors have become an irreplaceable device in contemporary communication systems.Efficient and accurate transistor models offer the potential to design larger,more complex circuits and systems.Transistors used in circuit systems include both active circuits and packaged circuits such as passive internal matching devices and bond wires.The existing packaged transistor modeling methods mainly build models for the packaging circuits,without the models of the nonlinear part.At present,there is no unified modeling method for the whole package transistors.In recent years,artificial neural network(ANN)techniques have been considered as effective alternatives to traditional modeling methods in microwave modeling.In this thesis,the neural network is used to model the DC,small signal and large signal characteristics of packaged transistors.Firstly,the DC model of the packaged transistor is built according to the existing Neuro-Space Mapping(Neuro-SM)structure and the trained DC model can be built in the simulation software easily.The modeling example demonstrates that the DC characteristic modeling of the power field effect transistor is feasible.Secondly,an improved mapping circuit structure is proposed for the small signal characteristics of the packaged transistor in this thesis.The packaged transistor is divided into three parts:the input package circuit,the nonlinear circuit and the output package circuit,and these three parts are modeled respectively.In addition,the conversion formula is derived to represent the relationship between the S-parameters of the input package circuit,nonlinear circuit,output package circuit and the overall packaged transistor.A new training method for adjusting the small signal characteristics of packaged transistors with different parameters is proposed,which make the proposed model match the measured data of the device effectively and accurately.Finally,an improved Neuro-SM structure is proposed and a new nonlinear function is used to improve the large-signal characteristics of the existing model,while ensuring DC and small signals features are not affected based on the good fitting of the DC and small signal characteristics of the packaged transistor.A step-by-step training method is developed for fast training of the proposed Neuro-SM model avoiding variables adjustment repeatedly.The modeling results prove that the large signal model established in this paper can accurately reflect the large signal characteristics of the packaged transistors.