System Identification Method Study Based On Neural Networks

System identification based on neural networks is the hotspot of study at present. Identification of nonlinear system is a topic which has been troubling people identification work for a long time, but neural network provide a new, general and powerful tool. This thesis does systematic research on system identification based on BP network, RBF network and hopfield network. The main works are written as follows: The improved algorithms of BP network have better effect on system identification. How to select the input signal for identification to get the most sufficient system response is simulated. How to choose the network structure is researched, and uses the periodic function as the activation function to do the identification work. This thesis does the identification work on low order, high order, linear and nonlinear systems, and their generalization ability. Those prove that BP network has the comprehensive identification ability, powerful function, wide adaptation and good generalization ability. But it has the shortcomings of longer calculation and easily getting into local minimum. RBF network has the virtues of the best approximation property, rapid calculation and needn set the node number of hidden layer. This thesis tests its rapid calculation property, good approximation property, but its generalization ability is not satisfied. So we give the explanation and analysis. Hopfield network is a typical dynamic network. Nowadays people use it to identify the argument of state space. Whether or not it can identify the input-output model? We give the result through simulation. This thesis is about the system identification on neural network. Identification of system plays the important role in control, prediction, analysis, design and computer simulation, so it certainly has the wide foreground and strong physic significance.