一种基于卷积神经网络的谱匹配地震动选取方法

(1.江汉大学 精细爆破国家重点实验室,湖北 武汉 430056; 2.哈尔滨工业大学(威海)海洋工程学院,山东 威海 264209)

谱匹配地震动选取; 卷积神经网络; 加速度设计谱; 抗震规范; 时程反应分析

A spectrum-matched ground motion selection method based on convolutional neural networks
ZHAO Guochen1,XU Longjun1,ZHU Xingji2,LAI Qinghui1,XIE Lili1

(1.State Key Laboratory of Precision Blasting, Jianghan University, Wuhan 430056, China; 2.School of Ocean Engineering,Harbin Institute of Technology at Weihai,Weihai 264209,China)

spectrum-matched ground motion selection; convolutional neural networks; acceleration design spectrum; seismic design code; time-history responseanalysi

DOI: 10.15986/j.1006-7930.2023.02.011

备注

选取与目标设计谱匹配的地震动记录是建筑结构抗震分析与设计中的重要问题之一。鉴于卷积神经网络在图像识别与分类领域具有突出的性能,本文提出了一种基于卷积神经网络的谱匹配地震动选取方法以考虑反应谱图像的二维特征。首先阐述了通过卷积神经网络进行谱匹配地震动选取的基本原理和训练数据的生成方法,然后讨论了Keras深度学习框架中11个神经网络结构在谱匹配地震动选取中的特性,最后采用Xception和InceptionResNetV2两种性能最优的网络结构从200次地震的11462条地震动水平分量中选取与我国规范设计谱相匹配的数据。研究表明,卷积神经网络能够有效地提取反应谱图像的特征,所选地震动平均反应谱与目标设计谱之间的差异较小。本文的研究工作可为工程实践中选取与目标设计谱匹配的地震动提供方法参考和技术支撑。
Selecting the spectrum-matched ground motions is one of the critical problems in the seismic analysis and design of building structures. Because of the outstanding performance of convolutional neural networks in image recognition and classification, this paper proposes a spectrum-matched ground motionselection methodbased on convolutional neural networkto consider the two-dimensional characteristics of response spectrum images. Firstly, the basic principles of using convolutional neural networks to select spectrum-matched ground motions and the method of generating training data are described. Then, the characteristics of 11 neural network structures of the Keras deep learning framework in the spectrum-matched ground motions selection are discussed. Finally, two optimal networks of Xception and InceptionResNetV2 are used to select the records matching the design spectrum of the Chinese design code from 11462 horizontal components of 200 earthquakes.The results show that convolutional neural networks can efficiently extract the features of the response spectrum image, and the difference between the mean response spectrum of the selected motions and the target design spectrum is non-significant. The research work could provide method reference and technical support in engineering practice when the spectrum-matched ground motion selection is needed.