Relationship modeling between vehicle‐induced girder vertical deflection and cable tension by BiLSTM using field monitoring data of a cable‐stayed bridge

Abstract

Structural health monitoring (SHM) systems have been widely implemented in long-span bridges to measure various structural responses. It is difficult to directly perform condition assessment from the variations in structural responses because of initial residual stress and defaults, coupling effects of structure damage and external loads, and others. To fix these problems, this study proposes global and partial bidirectional long short-term memory (BiLSTM) models to establish a relationship between girder vertical deflection (GVD) and cable tension (CT). First, a global BiLSTM model is built using all the measured signals of GVD and CT, and the test results show that the average root mean square error (RMSE) and relative RMSE (RRMSE) between the predicted and ground-truth CTs are 1.83 kN and 3.19%, respectively. Second, the effects of both traffic volume and noise level on the model prediction error are investigated, indicating that the proposed method is robust to different noise levels and traffic volumes under normal operational conditions. Finally, to customize a prediction model for a certain CT, a group of partial BiLSTM models is further constructed with only a few GVDs as inputs and a single CT as output. Sobol's sensitivity index is adopted as the indicator to select the most significant inputs among all the GVD sensors. The partial model is more efficient for training because of significantly fewer channels in each layer and model parameters. The prediction results show that the partial models can achieve an average RMSE and RRMSE of 1.86 kN and 3.24%, respectively, which are similar to the prediction accuracy of the global model. In addition, the partial model can be applied if certain sensors are out of order when compared to the global model.