Abstract
In a theoretical analysis of nonlinear vibratory systems, nonlinear normal modes are used to reduce the order of the system retaining the effect of the nonlinearity accurately. In experimental identification, it is expected similarly that more accurate result can be obtained by using nonlinear principal component analysis. In this report an identification technique that uses nonlinear principal component analysis by a neural network is proposed. This technique uses data in state space, and after determining the principal components by the sand-glass type of neural network, governing equations with respect to the principal components are determined by another neural network. The applicability of the technique is confirmed by numerical simulation.
