Displacement Prediction by Acceleration of Bridge Pier Using Long Short-Term Memory

After a large earthquake, prompt assessment of damage to seismic isolation bearings, especially in bridges, is critical. Although accelerometers are commonly used in structural health monitoring, displacement response calculations based on time integration of acceleration data are often inaccurate due to numerical errors. In this paper, using dynamic analysis results on a bridge model with seismic isolation bearings, a long short-term memory (LSTM) model was employed to predict displacement more accurately. Results showed that the LSTM model reduced the maximum response displacement error from 62.88% to 19.44% compared to a conventional neural network without LSTM. However, reproducing residual displacement at the pier top due to pier plasticization remained challenging.