Performance of neural networks for simulation and prediction of temperature-induced modal variability

Vibration-based damage detection methods use changes in modal parameters to diagnose structural degradation or damage. Structures in reality are subject to varying environmental effects which also cause changes in modal parameters. The well-defined nature of the environmental effects on modal properties is essential for reliable damage diagnosis based on vibration measurement. In this paper, the performance of artificial neural networks (ANNs) for simulation and prediction of temperature-caused variability of modal frequencies is investigated. Making use of one-year measurement data of modal frequencies and temperatures from an instrumented cable-stayed bridge, three- layer back-propagation (BP) neural networks are configured to model the correlation between the temperatures and frequencies. Two approaches are adopted in defining the training samples to train the neural networks and the testing samples to verify the prediction capability of the neural networks. It is shown that when using appropriate training data covering a wide range of temperature variations, the trained neural networks exhibit satisfactory performance in both reproduction (simulation) and prediction (generalization). A good mapping between the temperatures and frequencies is obtained by the neural network models for all measured modes.