Research On Nonlinear System Identification Based On Deep Neural Network

How to characterize nonlinear system characteristics is an important field in scientific research and engineering applications.At present,some achievements,such as differential geometry method,Hammerstein-Wiener method,iterative method and so on,have been achieved in the research and characterization of nonlinear system characteristics.However,some studies shown that these methods have deviations for analyzing the nonlinear systems.And these methods cannot accurate the model and identify nonlinear systems.Because these methods are approximated as a linear system to carry out research by sacrificing the partial nonlinear characteristics of the nonlinear system.If the methods applied to identify complex nonlinear systems,estimation error will become bigger.For improving the complex nonlinear systems accurately model and identify,the thesis researched the property of complex nonlinear systems and propose a method for accurately modeling and identifying complex nonlinear systems.In this thesis,firstly,nonlinear systems modeled and identified has been reviewed in the aspects of signal pre-process,nonlinear system model and nonlinear system parameter identification and much.And then study and analyze the shortages in their research.So we propose a method that combine deep neural network(DNN)with Fourier series to complete the accurate modeling and multi-parameter identification of complex nonlinear systems.First,the thesis introduced the DNN theory.In the specific research,this thesis analyses the model function equation and establishes the relationship between the model and Fourier series theory.So that we can realizes the modeling of interference model.Based on the theoretical and experimental analysis of the DNN,the hyper-parameters of the DNN is determined in thesis.On the basis of signal denoising by using the wavelet threshold denoising method,the identification scheme is applied to solve the interference model.The experimental results show that the proposed scheme can accurately model and identify complex nonlinear systems,so as to minimize the error between the identification value and the real value.