Local Quadratic Model Tree with Orthogonal Matching Pursuit (LOQUMOTOMP) Method for Nonlinear System Identification

Abstract

AbstractIn order to accurately model the behavior of nonlinear systems under different operating conditions, complex structures, such as the Volterra and Wiener series', are often used. The main drawback of such structures is the need for a large number of parameters. In an alternative approach, the input‐output relation of such complex nonlinear systems is described using different local linear models under different operating conditions. Such models also require many local models to maintain the accuracy under different operating conditions. In this paper, both methods are combined and a new iterative method is presented to obtain an accurate model with fewer parameters. To create a more precise model, a choice is made at each iteration of the identification algorithm between increasing the number of local models and increasing the complexity of each local model. Furthermore, to reduce the number of parameters and local models, the most significant regressors are selected by an orthogonal matching pursuit method. In the method proposed, polynomial or linear structure is used for the operating condition of the nonlinear system with more or less complexity. The proposed method has been tested on the experimental data of a hydraulic actuator. The proposed identification method provides a more precise model with fewer parameters than some well‐known nonlinear system identification methods, such as neural networks and the Local Linear Model Tree (LOLIMOT).