Prediction On Wind-induced Displacement Of Cable-stayed Bridge Based On LSTM Deep Neural Network

Cable-stayed bridges are favoured by bridge designers for their reasonable force performance and large spanning capacity.However,the larger the span,the more flexible the structure and the more sensitive it is to wind loads.Wind-induced vibrations may cause excessive structural deformation or even structural damage to the bridge,so it is necessary to monitor the condition information of the structural elements over time.The wind displacement response is a good indicator of the deformation of important elements such as the main girders under wind loads,which can be used to determine the damage to the cable-stayed bridge structure in time for maintenance treatment.In this paper,a Long Short-Term Memory Neural Network(LSTM)model is developed to predict the wind-induced displacement response time series based on the health monitoring data of the Hedong Bridge,a cable-stayed bridge crossing the Pearl River in Guangzhou.The main research work and related findings of this paper are as follows:(1)The wind speed and GPS(Global Positioning System)electrical signals collected by the health monitoring system were converted into a time-series of wind speed and displacement to remove the outliers,and the wavelet packet coefficient correlation analysis was further used to reduce the noise of the wind-induced displacement series.The wind speed data were obtained from the propeller anemometer installed in the span of the main beam,and the displacement data included the cross-bridge and vertical displacements in the span of the main beam.The results show that the correlation coefficients of wind speed and displacement monitoring data are relatively high in the low frequency band,and reasonable wind speed and wind-induced displacement sequences can be obtained as the sample set for LSTM prediction by reconstructing the low frequency components of wavelet packets.(2)A bridge wind displacement prediction model based on the LSTM deep neural network model was established,and the influence of the input and output data sets of wind speed and lateral wind displacement sequences on the accuracy of the prediction results was investigated using a model with specified hyperparameter settings.The hyperparameters of the LSTM neural network were carefully analysed based on the 1min average input and output datasets,and reasonable hyperparameter settings for the LSTM model were determined.(3)A Bayesian optimization-based LSTM model was further developed,and the parameters such as the number of neurons and the learning rate were automatically tuned.The root-mean-square error,the mean absolute error and the coefficient of determination between the prediction results of wind-induced cross-bridge displacement and the measured results were0.29 mm,0.22 mm and 0.98 respectively.The prediction accuracy of vertical displacement is slightly lower than that of lateral displacement,and the prediction accuracy of lateral displacement is higher than that of literature results,random forest and support vector regression model.The non-linear relationship between meteorological wind speed and windinduced displacement was further fitted based on the Bayesian optimised LSTM model,indicating the feasibility of establishing displacement response predictions from meteorological data.Therefore,the method established in this paper can be applied to predict the wind-induced displacement time course of cable-stayed bridges more accurately,providing an important technical tool for the evaluation of wind safety of bridge projects.