Delaunay-based Local Model Networks for Nonlinear System Identification

This paper presents an efficient methodology for nonlinear system identification based on Delaunay networks. These networks [1] share the learning capabilities of artificial neural networks but they are computationally much more efficient. In the approach discussed here, the interpolation nodes of a Delaunay network are interpreted as local linear models of a nonlinear plant. Hence, standard parameter estimation techniques can be applied to train the network. Combined with a heuristic strategy for structural optimization, the proposed method constructs nonlinear models which can be implemented on low-cost microcontrollers and can meet strict realtime requirements. Thus, the method is a promising approach in fields of application where computer power is a limited resource. Two examples of automotive applications are discussed and exemplify the approach