Comparison of some initialisation methods for the identification of nonlinear state-space models

In many measurement applications, highly accurate models are needed. Linear models can fail to represent the measured phenomena in a satisfactory way. In the class of nonlinear dynamic models, a nonlinear state-space model is an attractive option due to its good modelling capabilities. In order to obtain optimal results with this model, good initial estimates for the nonconvex optimisation problem are crucial. We show that the classical approach (via the best linear approximation) can get trapped in local minima and we present some alternative, recently developed, initialisation methods (nonlinear models). A simulation example is used to compare their performance, and the results (including time-efficiency and flexibility) are discussed.