Linearization and state estimation of unknown discrete-timenonlinear dynamic systems using recurrent neurofuzzy networks

摘要

Model-based methods for the state estimation and control of linearnsystems have been well developed and widely applied. In practice, thenunderlying systems are often unknown and nonlinear. Therefore, datanbased model identification and associated linearization techniques arenvery important. Local linearization and feedback linearization havendrawn considerable attention in recent years. In this paper,nlinearization techniques using neural networks are reviewed, togethernwith theoretical difficulties associated with the application ofnfeedback linearization. A recurrent neurofuzzy network with an analysisnof variance (ANOVA) decomposition structure and its learning algorithmnare proposed for linearizing unknown discrete-time nonlinear dynamicnsystems. It can be viewed as a method for approximate feedbacknlinearization, as such it enlarges the class of nonlinear systems thatncan be feedback linearized using neural networks. Applications of thisnnew method to state estimation are investigated with realisticnsimulation examples, which shows that the new method has usefulnpractical properties such as model parametric parsimony and learningnconvergence, and is effective in dealing with complex unknown nonlinearnsystems