The Research Of Recurrent Neural Network And Its Application On Nonlinear Dynamic System Identification

The identification of nonlinear dynamic system is always the difficulty and the focus in the fields of control research. Multilayer forward neural network belongs to static network, so there are many problems in nonlinear dynamic system application based on it. While with feedback behavior, the recurrent neural network can catch up with the dynamic response of the system, and it is proposed for nonlinear dynamic system identification. And in the identification based on the recurrent neural network, the model order need not to be selected, so the process of identification is simplified. Therefore the recurrent neural network is a potential network in the development of identification in the control system. And it has aroused the extensive attention of people in recent years.The structure of the recurrent neural network has a lot of kinds. The different structures will inevitably cause the different relations of input and output, so the recurrent neural networkshows the different dynamic performance. This paper has summarized the structures of the recurrent neural network. The recurrent neural network is classified into three sorts: global feedback recurrent neural network, global forward recurrent neural network and mixed recurrent neural network. Each kind of recurrent network has its subclasses. The structure diagrams of describing each network features were given. At the same time, the comparisons among variable different functions of network are done. The similarities and differences on these networks are also analyzed.In the development of the neural network, the research of learning algorithm has always been the core of the neural network research. The learning algorithm of the recurrent neural network has continued to use with that based on error back propagation, which has the obvious defects: slow convergence and easily to fall into local minima. Two kinds of learning algorithms of the recurrent neural network have been introduced, which put forward by domestic and international experts and scholars at present. They are the learning algorithms of the recurrent neural network based on recursive least-square (RLS) algorithm, and the learning algorithm of the recurrent neural network based on recursive prediction error (RPE) algorithm.Based on error-correct learning, looking for minimum value of objective function is optimization process. The optimalalgorithm is classified into two sorts: deterministic algorithm and stochastic algorithm. Looking from the viewpoint of the two kinds of optimization algorithm, two kinds of new learning algorithm of the recurrent neural network are proposed in this paper.The first algorithm utilizes Levenberg-Marquardt(LM) algorithm in optimization technique of standard numerical value to train the recurrent neural network. And this new algorithm provides the compromise of the fast speed of Newton's method and the convergence of gradient descent algorithm. Furthermore, in order to overcome the disadvantage of the centralized computing of LM algorithm, a parallel LM algorithm is derived. In the new parallel algorithm, the computation is distributed to each neuron in the network, which gives prominence to the nature of parallel handling information of the neural network. And fast speed and high accuracy of the convergence are obtained.The second algorithm is based on Alopex algorithm that is a stochastic algorithm on basis of simulated annealing idea. In this paper, improved Alopex algorithm is used to train the recurrent neural network. This new algorithm corrects the defect of unchangeable weight error in Alopex, and fast convergence is obtained. At the same time, it helps to escape from local minima, and approach global search.At the same time, a given model is identified by using the recurrent neural network trained with these algorithms, and themodel of a fuel heater is established. The simulation demonstrates the effectiveness of the proposed algorithms. A sensor model based on the recurrent neural network is set up, which aims at the sensor's complicated characteristic of nonlinear input and output that affected by temperature. The approach for sensor modeling compensates the effects of temperature and improves nonlinearity. The three examples show the superiority of the identification based on the recurrent neural network.