Wind-Induced Acceleration Prediction Of Cable-Stayed Bridges Based On LSTM Model

The proportion of cable-stayed bridges in bridge engineering is increasing,and they are more and more sensitive to the wind environmental effects.Structural response and bridge performance evaluation under long-term wind effects are of great significance for bridge operation and maintenance,and such evaluation depends on the establishment of structural wind-induced response models,and good prediction models can improve the accuracy and efficiency of structural response estimation and performance evaluation.Traditional windinduced acceleration response modeling methods are often based on more stringent assumptions and are difficult to calculate accurately and efficiently for complex structures.With the development of bridge structural health monitoring systems and the widespread use of deep learning algorithms,it is important to model and analyze the structural wind response of bridges based on monitoring data and deep learning algorithms.There have been studies on the prediction of jitter acceleration response using traditional machine learning methods from field measurement data,but there are few studies on the prediction of jitter acceleration response using deep learning methods.At the same time,there are more studies on the prediction of acceleration response under typhoon climate conditions but less studies on the prediction of acceleration response under benign climate conditions,and less studies on the prediction model of bridge structure response based on the meteorological data near the bridge.In this paper,the existing health monitoring data of Hedong Bridge are used to fully analyze the relationship between the data and the deep learning algorithm to predict the wind-induced acceleration response of the bridge under different working conditions,and the studied method can provide some reference for the bridge wind-induced acceleration response prediction modeling research.The main research of this paper is as follows:(1)Analyzed the distortion of wind field data and acceleration data from the health monitoring system of Hedong Bridge,and recover some of the data by applying the nonlinear relationship between sensor data at different points in the monitoring system.A long and short term memory network was established by selecting strongly correlated wind speed and acceleration response average value samples(10 min and 30 s respectively)from the processed data to build an acceleration response prediction model to validate the predictions for some samples.The squared correlation coefficients of most of the results are greater than 0.7,which is better than the literature results.(2)Based on the strongly correlated measured wind speed and synchronous acceleration response average value samples(sample length is 30 min),three conditions are considered,namely,only measured wind speed,only meteorological wind speed,and meteorological wind speed and meteorological temperature,and the corresponding acceleration prediction models are established.The results show that the prediction accuracy of more than half of the prediction conditions with meteorological wind speed and temperature is higher than that of the prediction conditions with only measured wind speed,indicating the feasibility of establishing acceleration response prediction models by meteorological data.(3)By processing the wind vibration data and using deep learning algorithms such as LSTM neural network,the wind induced vibration regression model was established to predict the root mean square values of the vertical and cross-bridge acceleration responses respectively.,and four algorithms,LSTM,CNN,DNN and Conv LSTM,were used to compare the models and add Bayesian parameters for optimization.The results show that the prediction results of the Bayesian optimized Conv LSTM model are better,and quared correlation coefficient SCC for all model predictions results is greater than 0.829,and the predicted response is more consistent with the measured response,which shows the effectiveness of the deep learning method in the wind-induced acceleration response prediction modeling.