Identification of bridge section flutter derivatives and numerical calculation of critical flutter wind speed based on deep learning

The current bridge flutter derivative identification method has difficulty in data acquisition, and its precision is affected by parameter settings, making it challenging to obtain flutter derivatives efficiently and accurately, which affects the calculation precision of the critical flutter wind speed. To address the problems, this paper explores the application of deep learning methods in bridge flutter derivative identification to reduce dependence on experiments and simulation calculations and improve identification precision and calculation efficiency. First, the computational fluid dynamics (CFD) method is used to generate aerodynamic data of different bridge sections, and flutter derivatives are extracted as training labels. Then, a model combining a one-dimensional convolutional neural network (1D-CNN) and a bidirectional long short-term memory network (Bi-LSTM) is constructed to extract the aerodynamic time series' local features and temporal dependencies and realize flutter derivative identification. 1D-CNN automatically captures the instantaneous fluctuation features in the aerodynamic time series through local convolution kernels. Bi-LSTM mines the long-term dependency of aerodynamic forces through bidirectional time series modeling. The identified flutter derivatives are interpolated by the parabola fitting method to construct the aerodynamic parameter variation curve under continuous wind speed. The critical flutter wind speed is calculated based on the improved Scanlan-Tomko flutter criterion combined with the numerical iteration method. The results show that the mean absolute error (MAE) of the flutter derivative of this method is ≤3.21% under various bridge sections and wind speed conditions. In the circular streamlined box girder, the relative error of the critical wind speed at a wind speed of 25 m/s is as low as 2.02%. The calculation efficiency is improved by 24.54% compared with the traditional method, and the error is reduced by 26% compared with the control group, which verifies its high efficiency and accuracy in the wind-resistant design of bridges.