Adaptive neural networks for model updating of structures

Abstract

A model updating methodology based on an adaptive neural network (NN) model is proposed in this study. The NN model has a feedforward architecture and is first trained off-line using some training data that are obtained from finite-element analyses and contain modal parameters as inputs and structural parameters as outputs. To reduce the number of training data while maintaining the data completeness, the variation of structural parameters is arranged using an orthogonal array. This NN model is then adaptively retrained on-line during the model updating process in order to eliminate the difference between the measured and the predicted modal parameters. A modified back-propagation algorithm is developed, in which the learning rate is dynamically adjusted once every few iterations. A jump factor is introduced to overcome the numerical difficulty caused by the saturation of the sigmoid function in order to improve the convergence performance of the NN model. The current adaptive NN updating procedure is applied to a suspension bridge model and verified both numerically and experimentally. The results indicate that by adaptively training the NN model and iteratively adjusting the structural parameters, it is possible to reduce the differences between the measured and the predicted frequencies from a maximum of 17% to 7% for the first eight vertical modes.