A Data-Driven Building’s Seismic Response Estimation Method Using a Deep Convolutional Neural Network

Abstract

As a refined finite element model takes much effort to build and tune to simulate the building structure's response under seismic effects, many rapid estimation methods were proposed to predict the engineering parameters. These methods include simplified structure models, response spectrum, interstory drift spectrum, and machine learning method. This study proposes a method that combines the interstory drift spectrum and a deep learning method to estimate the maximum interstory drift ratio (MIDR). The proposed method includes two approximations. Firstly, use the interstory drift spectrum to estimate the MIDR as a first approximation. Since the differences exist between the interstory drift spectrum and the true responses, the interstory drift spectrum's adjustment is necessary. The second approximation uses a deep convolutional neural network (DCNN) to tune the first approximation to predict the MIDR under a new seismic event. In the training process of the DCNN, 30 reinforced concrete buildings' time history analyses results and 38 interstory drift spectrums were fed into the DCNN. The proposed method is also compared with four artificial neural network models and one support vector machine model to show its advantages. The results indicate that the DCNN could learn the relationship between the interstory drift spectrum and the time history analyses results and make a reasonable prediction of MIDR. Besides, the proposed method is used in MIDR estimation of 30 more detailed finite element models of steel moment-resisting frames. The results indicate that the methodology could give a reasonable estimation of the buildings' MIDR of new seismic events.