Deep learning-based reconstruction of missing long-term girder-end displacement data for suspension bridge health monitoring

Abstract

The flexibility of a suspension bridge usually results in significant longitudinal displacement at the end of the main girder under the joint action of environmental factors and traffic loads. Fluctuation amplitude and accumulation of girder-end displacement (GED) are primary data for the performance evaluation of suspension bridge elements, including expansion joints, supports, and dampers. However, the performance of the bridge structural health monitoring (SHM) system is frequently deteriorated by the long-term continuous GED data loss or anomaly. In this study, this issue of the suspension bridge SHM system was resolved by proposing a deep learning-based framework for reconstructing the missing GED data. Under this framework, a long short-term memory (LSTM) network-based regression model was first built between the ambient temperature data and the thermal-induced low-frequency components of GED data. Next, a U-net-based spectral bandwidth expansion model was applied for the step-by-step generation of vehicle/wind-induced high-frequency GED data terms based on the available low-frequency ones. Finally, a statistical correction strategy was employed to improve the prediction accuracy for the high-frequency GED data fluctuation amplitudes. The trained model for the reconstruction of the missing GED data uses only the air temperature input values in the data loss period. This method is especially lucrative when GED data in all sensor channels are missing. The reliability and efficiency of the proposed model were demonstrated by a case study. Specifically, the problem of reconstructing the missing GED data in the SHM system of the Jiangyin Yangtze River Bridge in China was considered. Using the relative error of cumulative GED as the performance indicator, a reconstruction precision within 10% was obtained.