A Knowledge Transfer LSTM Model to Predict the Seismic Response of Structures

A novel deep learning framework is proposed to rapidly and accurately predict the seismic response of a structure. By adopting transfer learning and unsupervised learning, only two temporal ground motion records and the corresponding structural displacements are required to train and test a deep learning model. The proposed framework consists of four parts: 1) The seismic history is employed to build a database for a specific structure, consisting of important values extracted from earthquake ground motions. 2) The Structural Seismic Response (SSR) net is established on previously recorded earthquake ground motions and displacements. Instead of training the full set of ground motions on a single LSTM model, the SSR net is composed of multiple LSTM models trained independently on a specific earthquake and the corresponding displacements. Each LSTM model tries to understand how to predict the time history of dynamic displacements based on a single earthquake ground motion. 3) The unsupervised nearest neighbor algorithm aims to identify the most relevant previous earthquake when a new earthquake occurs. 4) In the knowledge transfer strategy, knowledge acquired from the most relevant previous earthquake will be transferred to predict the structural displacement caused by a new earthquake. To validate the novel framework proposed in this study, ground motion data and field structural response data measured from a building and bridge is utilized. The results show that the proposed framework can predict reliable seismic structural responses without excessive training procedures and offer significant potential in advancing seismic fragility analysis and reliability assessment.