Modeling of temperature-frequency correlation using long-term monitoring data: Methods and comparison

A good understanding of environmental effects on structural modal properties is essential for reliable performance of vibration-based damage diagnosis methods. In this paper, the temperature-frequency correlation models are developed by means of the neural network (NN) technique and the combined principal component analysis and neural network (PCA-NN) technique. Then a comparative study of these two techniques for reproducing and predicting temperature-caused variability of modal frequencies is conducted with the use of long-term monitoring data from the cable-stayed Ting Kau Bridge. It is shown that perceptron neural networks with single hidden layer are sufficient for modeling the correlation and an appropriate number of hidden nodes are crucial to achieve good prediction performance. Using principal components of the measured temperatures as input to neural networks can achieve almost same simulation and prediction capabilities and also ensure stable regression estimates.