Identification of nonlinear aerodynamic damping based on LSTM-modified unscented Kalman filter

Accurate identification of amplitude-dependent aerodynamic damping in highly flexible structures is severely challenged by strong nonlinearities and unpredictable noise. To improve the identification performance, this study proposes a deep fusion adaptive identification algorithm (LSTM-UKF) based on a long short-term memory (LSTM) network-modified unscented Kalman filter (UKF). Integrating data-driven deep learning with a physics-driven state-space model, the proposed method utilizes an embedded LSTM gain correction module to extract time-series features of prediction residuals and observation innovations in real time, thereby dynamically adjusting the Kalman gain. Expanding on this, this method establishes an unsupervised online training framework driven by measurable data. Numerical validation based on simulated vortex-induced vibration (VIV) responses demonstrates that, under the tested initial noise covariance settings, the proposed LSTM-UKF is less sensitive to covariance mismatch and achieves more accurate state tracking and parameter convergence than the traditional UKF. The gain correction based on LSTM reduces the difficulty of manually tuning the noise covariance parameters. Validation using wind tunnel experimental data reveals that the aerodynamic damping curves identified by the LSTM-UKF are in high agreement with those obtained from the Hilbert Transform method. The algorithm simplifies the difficulty of parameter tuning in engineering practice and provides a new technical pathway for the real-time online identification of nonlinear system parameters in complex wind fields.