A training strategy based on iterative self-training enhanced transfer learning for accurate structural time history response prediction

Abstract

Seismic time history responses are vital for structural performance evaluation and deep learning technologies have shown promise in accelerating the response calculation. However, effective training of deep neural networks generally necessitates extensive data, while their acquisition remains constrained by the time-consuming refined simulations and the high-cost experiments. Here a novel Iterative Self-training Enhanced Transfer Learning (ISTL) method is proposed for deep neural network training to improve seismic time history response prediction accuracy, especially in data scarcity scenarios. ISTL method leverages self-training method to augment abundant samples without additional experiments while integrating domain adaptation with novel output conditional distribution regularization to enhance learning through augmented knowledge. Three experiments validate the proposed method by predicting responses from numerical simulations of a nonlinear frame structure, shake-table testing of a linear frame structure, and field-sensing records of an instrumented shear-wall structure. The results show that ISTL method can eliminate the need for additional experiments and improve prediction performance by up to 60 % over conventional direct training process, underscoring its potential for developing robust predictive models for structural time history responses.