Digital prediction model of temperature-induced deflection for cable-stayed bridges based on learning of response-only data

Time-varying behavior of deflection field under effects of non-uniform temperature field is a focus of structural health monitoring for long-span cable-stayed bridges. In this paper, LSTM neural network is used to obtain the timing characteristics of deflection field and explore the correlation between each measuring point by using memory characteristics of neural unit, so as to predict deflection field under multiple working conditions. By using LSTM neural network, it can realize response-only deflection prediction. At the same time, LSTM neural network can also be applicalbe to solve the problem of recovery and regression of deflection field when temperature data is missing. Through the case study, we can find that the following: (1) the deflection field shows good linear correlation at each measuring point within main span, while the deflection between main span and side span has obvious nonlinear distribution characteristics; (2) compared with SVM method, the LSTM model can obtain better prediction results, and it can realize response-only deflection field prediction; (3) through the improved prediction model, historical deflection data of measured points are introduced to improve prediction accuracy. Because historical data are introduced to improve the input data set, the improved model can be used to predict the temperature-induced deflection in whole range of bridge under various working conditions; (4) the size of input data set can affect prediction model accuracy. The prediction effect of 1 min input data set is better than that of 10-min and 1-h input data set, which can effectively reduce the offset items of predicted data.