Signal Reconstruction Method for Lateral Displacement Response of Long-Span Suspension Bridges

Abstract

Lateral displacement is an important monitoring parameter that characterizes the degree of out-of-plane deformation of a long-span suspension bridge. Lateral displacement variation is mainly determined by the wind excitation and lateral temperature difference of the girder. Predicting structural responses using only these environmental measurements is new. Machine learning demonstrates a better generalization ability than conventional regression methods when analyzing big data and fitting multivariate nonlinear relationships. Furthermore, by using wind and temperature measurements as inputs, least-squares linear fitting, extreme gradient boosting (XGBoost), and long short-term memory (LSTM) neural networks are used to establish the prediction model of the lateral displacement of a 660-m suspension bridge. Compared with these nonlinear models, the linear model is not sufficiently accurate for signal reconstruction. Owing to the nonlinear time-delay consideration, the LSTM model can accurately fit the mapping relationship between the multipoint environmental effect sequence and a one-point displacement response. Moreover, the LSTM model shows better predictive performance than the XGBoost model. Consequently, the mean absolute error for the test set of the LSTM model was only 5 mm, and this statistical millimeter-level reconstruction error showed good accuracy for suspension-bridge deformation monitoring. Based on the proposed method, the suspension-bridge lateral displacement signal can be effectively reconstructed and restored when faced with a signal abnormality or loss owing to sensing and electrical problems.