基于深度神经网络的地表地震动幅值预测研究

苏闻浩; 刘启方

doi:10.11899/zzfy20240218

基于深度神经网络的地表地震动幅值预测研究

doi: 10.11899/zzfy20240218

苏闻浩,
刘启方^,

苏州科技大学, 江苏省结构工程重点实验室, 江苏苏州 215009

基金项目: 国家自然科学基金项目（51978434）；中国地震局工程力学研究所基本科研业务费专项资助项目（2021EEEVL0001）

详细信息

作者简介:
苏闻浩，男，生于1998年。硕士研究生。主要从事地震工程研究工作。E-mail：18862633233@163.com

通讯作者:
刘启方，男，生于1969年。研究员，博士生导师。主要从事地震工程研究工作。E-mail：Qifang_liu@126.com

计量
- 文章访问数: 44
- HTML全文浏览量: 30
- PDF下载量: 4
- 被引次数: 0
出版历程
- 收稿日期: 2022-12-15
- 刊出日期: 2024-06-30

Study on the Prediction of Ground Motion Amplitude Based on Deep Neural Network

Su Wenhao,
Liu Qifang^,

Suzhou University of Science and Technology, Key Laboratory of Structure Engineering of Jiangsu Province, Suzhou 215009, Jiangsu, China

摘要

摘要: 地震动预测模型是灾害分析和结构设计的重要组成部分，近年来神经网络技术愈发成熟地被应用在该预测模型的开发上，但多为使用地表以下30 m范围内土层等效剪切波速（V_S30）作为场地输入参数的单层全连接神经网络模型，忽略了完整土层厚度及剪切波速信息对地震动幅值的影响。本文采用卷积神经网络及全连接神经网络混合模型，选用日本KiK-net台网记录到的3174次地震共计39192条地震动记录，构建了一种基于深度神经网络框架的地表地震动幅值预测模型。该模型的输入参数为震级、震源深度、震中距、场地各土层厚度、剪切波速信息和井下地震动幅值，输出为相应的地表地震动幅值（PGA或PGV或PGD）。对模型进行训练并计算其各项评价指标予以评估，结果表明：(1)该混合神经网络模型的决定系数超过了0.85，模型残差服从正态分布，均值残差接近于0，模型表现出无偏的特性。(2)与现有经验公式相比，混合网络模型的PGA、PGV和PGD预测精度分别提升约26.9%、16.5%和11.6%。与使用V_S30作为场地参数的全连接神经网络模型相比,该框架下模型预测值与真实值的Person相关系数及各项评价指标均有所提升，模型残差的均值和标准差更小，PGA、PGV和PGD模型预测精度提升约6.3%、3.9%和3.4%，能更好地对地震动幅值进行预测。
- 地震动幅值 /
- 地震动预测模型 /
- 卷积神经网络 /
- 深度学习
Abstract: The ground motion prediction model is a critical component in disaster analysis and structural design. In recent years, neural network technology has increasingly been applied to develop such prediction models. However, most existing models are single-layer fully connected neural networks that use the equivalent shear wave velocity of soil layers within 30 meters below the surface (V_S30) as the site input parameter. This approach often overlooks the impact of the complete soil layer thickness and shear wave velocity information on ground motion amplitude. In this paper, a combined convolutional neural network (CNN) and fully connected neural network model is proposed. A dataset comprising 39192 ground motion records from 3174 earthquakes recorded by KiK-net is used to build a ground motion amplitude prediction model based on deep neural network technology. The model's input parameters include the magnitude, source depth, epicenter distance, thickness of each soil layer, shear wave velocity information, and underground ground motion amplitude. The output is the corresponding ground motion amplitude (PGA, PGV, or PGD). The model is trained and its evaluation indexes are calculated. The results indicate that: (1) The coefficient of determination (R²) of the mixed input neural network model exceeds 0.85. The model residual follows a normal distribution, and the mean residual is close to 0, demonstrating an unbiased characteristic. (2) Compared to the empirical formula, the prediction accuracy of the hybrid network model improves by 26.9%, 16.5%, and 11.6% for PGA, PGV, and PGD, respectively. (3) Compared to the traditional fully connected neural network model using V_S30 as the site parameter, the Pearson correlation coefficient and various evaluation indexes of the predicted values and actual values show improvement. The mean and standard deviation of the model residual are smaller, and the prediction accuracy of the PGA, PGV, and PGD models improves by about 6.3%, 3.9%, and 3.4%, respectively. This combined neural network model can better predict seismic amplitude, taking into account the complete soil layer information.
- Seismic amplitude /
- Ground motion prediction equations /
- Convolution neural network /
- Deep learning

HTML全文

图 1 各类场地数目分布

Figure 1. The histogram of site classification

下载: 全尺寸图片幻灯片

图 2 PGA 记录分布

Figure 2. PGA record distribution

下载: 全尺寸图片幻灯片

图 3 场地信息处理

Figure 3. Site information processing

下载: 全尺寸图片幻灯片

图 4 神经网络模型结构

Figure 4. Structure of neural network model

下载: 全尺寸图片幻灯片

图 5 混合输入网络模型与传统全连接网络模型目标值与预测值散点分布

Figure 5. Scatter distribution of target value and predicted value of hybrid input model and traditional model

下载: 全尺寸图片幻灯片

图 6 各神经网络模型的Person相关系数

Figure 6. Person correlation coefficient of neural network models

下载: 全尺寸图片幻灯片

图 7 网络模型残差分布

Figure 7. Residual distribution of neural network model

下载: 全尺寸图片幻灯片

图 8 混合神经网络框架下PGA、PGV、PGD模型在各类场地中残差分布

Figure 8. Residual distribution of hybrid neural network model in each site

下载: 全尺寸图片幻灯片

表 1 数据参数范围

Table 1. Data parameter range

输入参数	范围
震级M_W	1.9 ～ 9.0
震源深度/km	3 ～682
震中距/km	0.22～1217.50
地表峰值加速度/Gal	1～2430

下载: 导出CSV

表 2 NERPH场地分类标准

Table 2. NERPH site classification criteria

地表以下30 m范围内土层等效剪切波速 V_S30 /（m·s⁻¹）	场地类别
V_S30 >1500	A
760<V_S30≤1500	B
360<V_S30≤760	C
180<V_S30≤360	D
V_S30≤180	E

下载: 导出CSV

表 3 各神经网络模型散点拟合斜率

Table 3. Scatter fitting weight of neural network models

网络模型	PGA		PGV		PGD
网络模型	训练集	测试集	训练集	测试集	训练集	测试集
DNN	0.800	0.802	0.870	0.866	0.841	0.836
CNN+DNN	0.899	0.903	0.922	0.922	0.876	0.870

下载: 导出CSV

表 4 不同模型评价指标

Table 4. Evaluation indexes of different models

参数	模型	M_SE	M_AE	R²	误差/%
PGA	CNN+DNN	1.62×10⁻³	3.12×10⁻²	0.901	15.3%
	DNN	3.26×10⁻³	4.48×10⁻²	0.805	21.6%
	经验公式	4.82×10⁻³	5.91×10⁻²	0.490	42.2%
PGV	CNN+DNN	1.50×10⁻³	3.00×10⁻²	0.918	43.6%
	DNN	2.41×10⁻³	3.89×10⁻²	0.868	47.5%
	经验公式	4.93×10⁻³	6.01×10⁻²	0.421	60.1%
PGD	CNN+DNN	3.40×10⁻³	4.50×10⁻²	0.853	61.3%
	DNN	3.70×10⁻³	4.70×10⁻²	0.836	64.7%
	经验公式	5.22×10⁻³	6.31×10⁻²	0.396	72.9%

下载: 导出CSV

表 5 各网络模型残差的均值、方差、标准差统计

Table 5. Statistics of mean, variance and standard deviation of residuals of neural network models

网络模型	PGA			PGV			PGD
网络模型	均值	方差	标准差	均值	方差	标准差	均值	方差	标准差
DNN	−0.022	0.237	0.486	−0.021	0.231	0.480	−0.017	0.569	0.754
CNN+DNN	0.019	0.119	0.344	−0.018	0.139	0.373	0.015	0.518	0.720

下载: 导出CSV

表 6 混合网络模型在各场地中残差的均值、方差、标准差统计

Table 6. Statistics of mean, variance and standard deviation of residuals of mixed network models

场地类型	网络模型
	PGA			PGV			PGD
	均值	方差	标准差	均值	方差	标准差	均值	方差	标准差
B	0.023	0.123	0.351	0.005	0.138	0.372	0.009	0.489	0.699
C	0.027	0.123	0.350	−0.008	0.151	0.389	0.039	0.538	0.734
D	0.014	0.098	0.313	−0.019	0.121	0349	−0.002	0.486	0.697
E	0.323	0.181	0.425	−0.050	0.155	0.394	0.110	0.662	0.814

下载: 导出CSV

参考文献(24)

薄景山,李秀领,刘红帅,2003. 土层结构对地表加速度峰值的影响. 地震工程与工程振动,23(3):35−40. doi: 10.3969/j.issn.1000-1301.2003.03.006

Bo J. S., Li X. L., Liu H. S., 2003. Effects of soil layer construction on peak accelerations of ground motions. Earthquake Engineering and Engineering Vibration, 23(3): 35−40. (in Chinese) doi: 10.3969/j.issn.1000-1301.2003.03.006

董凯月,2020. 基于深度神经网络的地下地震动参数预测研究. 哈尔滨:哈尔滨工业大学.

Dong K. Y. ,2020. Study on the Prediction of underground motion parameters based on depth neural network. Harbin:Harbin University of Technology. (in Chinese)

胡安冬,张海明,2020. 机器学习在地震紧急预警系统震级预估中的应用. 地球物理学报,63(7):2617−2626. doi: 10.6038/cjg2020N0070

Hu A. D., Zhang H. M., 2020. Application of machine learning to magnitude estimation in earthquake emergency prediction system. Chinese Journal of Geophysics, 63(7): 2617−2626. (in Chinese) doi: 10.6038/cjg2020N0070

胡进军,谢礼立,2005. 地震动幅值沿深度变化研究. 地震学报,27(1):68−78. doi: 10.3321/j.issn:0253-3782.2005.01.008

Hu J. J., Xie L. L., 2005. Variation of earthquake ground motion with depth. Acta Seismologica Sinica, 27(1): 68−78. (in Chinese) doi: 10.3321/j.issn:0253-3782.2005.01.008

吕悦军,彭艳菊,兰景岩等,2008. 场地条件对地震动参数影响的关键问题. 震灾防御技术,3(2):126−135. doi: 10.3969/j.issn.1673-5722.2008.02.003

Lv Y. J., Peng Y. J., Lan J. Y., et al., 2008. Some key problems about site effects on seismic ground motion parameters. Technology for Earthquake Disaster Prevention, 3(2): 126−135. (in Chinese) doi: 10.3969/j.issn.1673-5722.2008.02.003

万永革,李鸿吉,1995. 人工神经网络在地球物理中的应用综述. 国际地震动态,16(1):9−14.

Wan Y. G., Li H. J., 1995. An overview of the application of artificial neural network in geophysics. Recent Developments in World Seismology, 16(1): 9−14. (in Chinese)

徐龙军,谢礼立,胡进军,2006. 地下地震动工程特性分析. 岩土工程学报,28(9):1106−1111. doi: 10.3321/j.issn:1000-4548.2006.09.011

Xu L. J., Xie L. L., Hu J. J., 2006. Analysis on engineering characteristics of sub-ground motions. Chinese Journal of Geotechnical Engineering, 28(9): 1106−1111. (in Chinese) doi: 10.3321/j.issn:1000-4548.2006.09.011

薛俊伟,刘伟庆,王曙光等,2013. 基于场地效应的地震动特性研究. 地震工程与工程振动,33(1):16−23.

Xue J. W., Liu W. Q., Wang S. G., et al., 2013. Research on ground motion characteristics considering site conditions. Earthquake Engineering and Engineering Vibration, 33(1): 16−23. (in Chinese)

于子叶,储日升,盛敏汉,2018. 深度神经网络拾取地震P和S波到时. 地球物理学报,61(12):4873−4886. doi: 10.6038/cjg2018L0725

Yu Z. Y., Chu R. S., Sheng M. H., 2018. Pick onset time of P and S phase by deep neural network. Chinese Journal of Geophysics, 61(12): 4873−4886. (in Chinese) doi: 10.6038/cjg2018L0725

Ahmad I., El Naggar M. H., Khan A. N., 2008. Neural network based attenuation of strong motion peaks in Europe. Journal of Earthquake Engineering, 12(5): 663−680. doi: 10.1080/13632460701758570

Atkinson G. M., 2001. An alternative to stochastic ground-motion relations for use in seismic hazard analysis in eastern North America. Seismological Research Letters, 72(2): 299−306. doi: 10.1785/gssrl.72.2.299

Borcherdt R. D. , 2012. VS30-A site-characterization parameter for use in building Codes, simplified earthquake resistant design, GMPEs, and ShakeMaps. In: The 15th World Conference on Earthquake Engineering. Lisbon: USGS.

Chicco D., Warrens M. J., Jurman G., 2021. The coefficient of determination R-squared is more informative than SMAPE, MAE, MAPE, MSE and RMSE in regression analysis evaluation. PeerJ Computer Science, 7: e623. doi: 10.7717/peerj-cs.623

Derras B., Bard P. Y., Cotton F., 2014. Towards fully data driven ground-motion prediction models for Europe. Bulletin of Earthquake Engineering, 12(1): 495−516. doi: 10.1007/s10518-013-9481-0

Dhanya J., Raghukanth S. T. G., 2018. Ground motion prediction model using artificial neural network. Pure and Applied Geophysics, 175(3): 1035−1064. doi: 10.1007/s00024-017-1751-3

Dokht R. M. H., Kao H., Visser R., et al., 2019. Seismic event and phase detection using time-frequency representation and convolutional neural networks. Seismological Research Letters, 90(2A): 481−490. doi: 10.1785/0220180308

Kerh T., Chu D., 2002. Neural networks approach and microtremor measurements in estimating peak ground acceleration due to strong motion. Advances in Engineering Software, 33(11-12): 733−742. doi: 10.1016/S0965-9978(02)00081-9

Khosravikia F. , Zeinali Y. , Nagy Z. , et al. , 2018. Neural network-based equations for predicting PGA and PGV in Texas, Oklahoma, and Kansas. In: Brandenberg S. J. , Manzari M. T. , eds. , Geotechnical Earthquake Engineering and Soil Dynamics V: Seismic Hazard Analysis, Earthquake Ground Motions, and Regional-Scale Assessment. Austin: American Society of Civil Engineers, 538−549.

Quinino R. C., Reis E. A., Bessegato L. F., 2013. Using the coefficient of determination R² to test the significance of multiple linear regression. Teaching Statistics, 35(2): 84−88. doi: 10.1111/j.1467-9639.2012.00525.x

Schober P., Boer C., Schwarte L. A., 2018. Correlation coefficients: appropriate use and interpretation. Anesthesia & Analgesia, 126(5): 1763−1768.

Shahriari B., Swersky K., Wang Z. Y., et al., 2016. Taking the human out of the loop: a review of Bayesian optimization. Proceedings of the IEEE, 104(1): 148−175. doi: 10.1109/JPROC.2015.2494218

Thomas S., Pillai G. N., Pal K., et al., 2016. Prediction of ground motion parameters using randomized ANFIS (RANFIS). Applied Soft Computing, 40: 624−634. doi: 10.1016/j.asoc.2015.12.013

Wang Y. W., Li X. J., Wang Z. F., et al., 2021. Deep learning for P-wave arrival picking in earthquake early warning. Earthquake Engineering and Engineering Vibration, 20(2): 391−402. doi: 10.1007/s11803-021-2027-6

Zhu W. Q., Beroza G. C., 2019. PhaseNet: a deep-neural-network-based seismic arrival-time picking method. Geophysical Journal International, 216(1): 261−273.

施引文献

资源附件(0)

访问统计