Research On Improved Neural Network Algorithm In Nonlinear System Identification

In recent years,with the increasing complexity and nonlinearity of control objects,new requirements for identification of nonlinear systems are proposed.As an emerging discipline,artificial neural network has been widely applied in different fields with its own characteristics,and it also provides a new way for the identification of nonlinear systems.In this paper,the artificial neural network is applied to the identification of nonlinear system from two aspects of nonlinear time-invariant system and nonlinear time-varying system.As the most widely used model in artificial neural network,BP neural network is applied to the identification of nonlinear stationary system in this paper.Aiming at the shortcomings of the standard BP neural network with slow convergence speed and easy to get into local minimum,this paper proposes two improvement algorithms for improving the excitation function and searching dynamic learning rate.At the same time,numerical simulations of two improved methods are studied.By analyzing the simulation results of output curves and error cumulative curves,the improved algorithm speeds up the convergence speed of BP neural network and reduces the computation time,that means a good identification effect for nonlinear stationary systemsFor nonlinear time-varying systems with reproducible changes in finite time intervals,based on the iterative learning algorithm,a time-varying radial basis neural network model is established.Simultaneously,the dynamic forgetting factor least-squares algorithm is applied to adjust the weights of the hidden layer and the output layer to realize the identification of the nonlinear time-varying system.Taking a nonlinear time-varying system as an example,numerical simulations are performed under different input conditions.Through the analysis of the simulation results,the time-varying RBF(Radial Basis Function)neural network can accurately identify the nonlinear time-varying system,and the identification accuracy is higher than the steady RBF neural network model.