Research On Nonlinear Time-varying System Identification Based On Artificial Neural Network

The conventional system identification methods are mostly established on known model structures, which demand plethoric prior knowledge. However, most of the existing systems are nonlinear, time-varying, and lack of prior knowledge, which makes it difficult for identification. So neural network methods are analyzed and used here, which are advantageous in the'black box'characteristic and arbitrary precision in approximating.The system identification development and research status based on neural network theory is studied firstly. The commonly-used BP, RBF and GRNN network learning methods are analyzed and simulated. The GRNN network is highlighted and comprehensively analyzed. Simulation results show that the existing GRNN network has two defects. One is the GRNN pattern layer neurons are proportional to the number of the training samples, so the network structure becomes larger as the number of the training samples increase. Another is the approximation result would be inaccurate if all the pattern layer neurons take a same value, while the computation would be complex if every neuron takes a different value respectively. Accordingly, two solutions are presented. The FCM clustering algorithm is adopted to solve the first defect. A similarity index of input data is defined. Then the iterative clustering operation can be decided by comparing this index with the setting threshold, which solves the problem of complex iteration. To solve the second defect, a method to judge every input vector's contributing ratio to the output is brought up, which selectively optimizes the smoothing parameters. In the end, a system identification strategy based on the modified GRNN network is presented and applied to a gas hydrate resistance survey system. The experiment results show that this network model has rapid processing speed, high identification accuracy and generalization ability. So it is valuable in application.