Nonlinear Continuous Dynamic System Identification by Automatic Localized Modeling

Abstract

This paper describes an automated localized modeling method to identify continuous nonlinear dynamic systems from their operating data. Using a method similar to finite element method’s automatic mesh generation, the input space is partitioned into overlapped regions that are small enough that a local model, such as a simple neural network, can approximate the data well in each region. Subsequently, adjacent regions are inspected to see if they can be represented well by a single local model to minimize the number of regions and local models needed to approximate a system. A nonlinear oscillator is used to test the proposed method, and the method was able to generate models that can simulate the system well. [S0022-0434(00)01902-X]