Similarity improvement using angular deviation in multimodel nonlinear system identification

In this work an unsupervised fuzzy learning method for the identification of nonlinear dynamical systems is designed. Accordingly, the learning process is featured by an incremental fuzzy clustering algorithm involving, in addition to the usual Euclidian distance, a new angular deviation. It turns out that: (i) the domain associated to each local model is better located compared to methods based on only Euclidian distance; (ii) the concentration phenomenon, observed when using standard metric classification, is highly reduced. These futures are confirmed by simulation.