A Pre-Trained Deep Learning Model for Fast Online Prediction of Structural Seismic Responses

Abstract

Deep learning techniques have gradually attracted considerable research interest in numerous application scenarios because of their capacity to simplify and accelerate calculations. Several researchers have adopted deep learning models, primarily end-to-end long short-term memory networks, to predict structural seismic responses in a data-driven manner and have achieved remarkable improvements. However, further research is required to reduce the training cost and complexity while maintaining the prediction accuracy of end-to-end models. In this study, a pre-training strategy was developed to meet these requirements. In brief, a generic base model was pre-trained by embedding the knowledge of calculating the response of single-degree-of-freedom systems, which was used as the basis for building tailored top models for the online prediction of seismic responses of different structures. This strategy modularized and simplified deep-learning-based seismic response prediction for training while maintaining the advantages of the end-to-end method. Proposed method was verified by using both simulation and actual datasets. The results demonstrated that a significantly lower training cost and complexity were required to achieve a prediction accuracy similar to that of the classic end-to-end method. Specifically, there was a reduction in FLOPs ranging from 20% to 80%, and a decrease in training time ranging from 44% to 82%. Moreover, the performance of deep learning with limited actual data was discussed.