Identification of nonlinear aerodynamic forces of bridges based on LSTM neural network and Hilbert–Huang transform

Abstract

Research is conducted on the identification of nonlinear aerodynamic forces in the process of bridge vortex-induced vibration. An identification method based on neural network and Hilbert–Huang Transform (HHT) is proposed. First, the time-history data of the bridge is identified through the Hilbert–Huang transform, the aerodynamic force polynomial is fitted, and the corresponding coefficients are obtained to establish the original aerodynamic force model. Then, by utilizing the Long Short-Term Memory (LSTM) neural network, an initial state prediction model is established. When changing the initial conditions, on the one hand, the original aerodynamic force model can be directly used to predict the aerodynamic forces under the new initial conditions. On the other hand, through transfer learning in the neural network, the initial state prediction model is transferred to the new initial conditions. After updating and adjusting the network data parameters of the model and retraining it, a new prediction model is obtained to predict the aerodynamic forces under the new initial conditions. Numerical examples and wind tunnel tests show that when the bridge is within the vortex-induced vibration lock-in range, the method proposed in this paper can accurately identify the nonlinear aerodynamic forces of the bridge.