Displacement estimation of nonlinear MDOF systems under seismic excitation using a deep learning model

The present work proposes a method for estimating the seismic displacement response of structures to assess post-earthquake structural safety. For Multi-Degree-of-Freedom (MDOF) structures, the proposed method estimates displacement responses using a deep learning model based on time-history observations of ground motion acceleration and absolute acceleration responses at specific structural coordinates. We generated a dataset comprising ground acceleration, absolute acceleration responses, and relative displacement responses by conducting dynamic analysis simulations. These simulations utilized simulated ground motions with varying intensities and time histories applied to structures with randomly changing structural parameters and member hysteresis parameters. The deep learning model was trained using this dataset to extract features from ground motion and absolute acceleration inputs and to output relative displacement responses from ground motion observation point. Previous studies successfully estimated displacements for structures with unknown hysteresis dominated by the first mode, such as FEM models of bridge piers or full-scale bridge shaking table tests, by training models on Single-Degree-of-Freedom (SDOF) nonlinear simulation results. The results indicate that by extending the training data to MDOF systems, this framework can estimate seismic displacement responses for more complex MDOF structures without identifying specific structural or hysteretic properties, relying solely on training with simple simulation data.