Formulation of an uncertainty model relating modal parameters and environmental factors by using long-term monitoring data

From the point of view of structural health monitoring, it is extremely important to discriminate alteration in structural behavior/response attribute due to damage from that due to environmental and operational fluctuation. In this paper, the correlation between natural frequencies and temperature is investigated for the cable-stayed Ting Kau Bridge by using measurement data from a long-term monitoring system installed on this bridge. One-year continuously acquired data from 45 accelerometers (a total of 67 channels) and 83 temperature sensors are used for this study. The data from 20 temperature sensors at the locations susceptible to temperature are first selected for the correlation analysis. Natural frequencies of the first 10 modes are identified by spectral analysis of the acceleration data at one-hour intervals. In order to ensure the identification accuracy, the natural frequency for a specific mode is determined using only the data from the accelerometers which produce large spectral peaks at that mode. The identified natural frequencies for each hour are used to correlate with the one-hour average temperatures measured from the 20 sensors during the same time. Based on the one-year measurement data which cover a full cycle of varying environmental and operational conditions, a four-layer perceptron neural network with 20 input nodes and 1 output node is trained for each mode to represent the relation between the measured temperatures (input) and the corresponding natural frequency (output). The configured neural networks for the 10 modes show excellent capabilities for mapping between the temperatures and natural frequencies for all the one-year measurement data.