Deep learning based nonlinear structural time history response prediction enhanced with multi-resolution convolution projection network

Abstract

Deep learning has shown potential for efficient seismic time history response prediction and performance evaluation of building structures. However, conventional network designs often omit initial projection layers, limiting input sequence utilization such as seismic acceleration data and generating suboptimal intermediate features for subsequent processing. In this study, a Multi-Resolution Convolution Enhanced Gate Recurrent Unit (MRCE-GRU) network is proposed, featuring a novel projection layer that integrates parallel convolutional layers with varying receptive fields. This architecture enables robust multi-resolution feature extraction during initial processing step. Three case studies are conducted to evaluate the proposed network: predicting displacement responses of a moment-resisting frame, analyzing field-sensing acceleration responses of an instrumented building, and modeling hysteresis curves of a nonlinear material. Extended evaluations are performed on network hyperparameters, intermediate projection outputs, and computational efficiency. The results demonstrate that the MRCE-GRU network achieves high accuracy with average coefficients of determination ${\text{R}}^2$ of 0.9590, 0.8784, and 0.9997 for three testing cases, respectively, while maintaining lightweight computational requirements compared to other methods. Moreover, the proposed projection network effectively captures critical features at an early step, transfers informative features to the subsequent layers, and ultimately enhances the response prediction performance.