NEURAL NETWORK MODELING OF STRUCTURAL RESPONSE UNDER PERSISTENT EXTERNAL EXCITATIONS

In this study a neural network modeling scheme is proposed where a neuro-model is trained in an supervised manner using a finite element model of the structure under consideration. This off-line training procedure allows one, not only to come up with a near optimal architecture for the neural network, but also to initialize its parameters at a desirable state. As a representative example, a geometrically nonlinear plate structure is utilized in time domain simulations where the behavior of the neuro-model is explored with respect to novel persistent external excitations in a desired frequency range. It is shown that the neuro-model performs very well in predicting the response of the nonlinear plate structure. Furthermore, the trained neural network would be ready for additional on-line adaptation in an unsupervised scheme.