地震研究

利用日本KiK-net台网提供的407个台站的30 952条地震动记录,提出了一种基于BP神经网络的场地等效剪切波速比变化预测模型。模型采用了均方误差函数及Adam优化算法,由3个输入参数、5个隐藏神经元及1个输出参数组成。输入参数为地面峰值加速度PGA、Arias烈度Ia及场地剪切波速VS30,输出为场地等效剪切波速比(VSr)。研究结果表明:该神经网络模型残差对于各输入变量整体呈现出无偏的特性,在大部分的软硬场地中均有较好的预测性能,该模型预测得到的PGA关于VS30的相关系数曲线与用传统的最小二乘法回归得到的函数曲线相比,其相关系数有更好的表现。该模型预测曲线显示,B类场地在PGA达到175 cm/s2时,场地剪切波速下降5%,D、E类场地在PGA达到140 cm/s2时,场地剪切波速下降5%,多数场地的非线性阈值为50～100 cm/s2。PGA在该网络模型中占据着较高的权重,为场地等效剪切波速变化的最主要控制参数。该网络模型捕捉到场地等效剪切波速比随PGA的增大有下降的趋势,而较为松软的D、E类场地受PGA影响更大,下降幅度更大。

In this paper,a total of 30952 records from 407 stations of the Japanese KiK-net is used to propose a prediction model for the change of equivalent shear wave velocity ratio based on BP neural network.The model adopts the mean square error function and the Adam optimization algorithm,consists of three inputs,five hidden neurons and one output.The input parameters are Peak Ground Acceleration(PGA),Arias intensity(Ia)and site VS30.The output parameter is site equivalent shear wave velocity ratio(VSr).The research shows that the residual error of the network model is unbiased for each input variable,and has good prediction performance in many kinds of sites.Compared with the function curve of the traditional least-square method,the neural network model has a relatively better performance.In the prediction curve of the network model,the shear wave velocity of the site of Class B decreases by 5% when the PGA reaches about 175 cm/s2,and the shear wave velocity of the sites of Class D and E decreases by 5% when the PGA reaches about 140 cm/s2.The nonlinear threshold of most sites is between 50～100 cm/s2.PGA occupies a high weight in the network model and is the main controlling parameter of the site equivalent shear wave velocity change.The network model captures that the equivalent shear wave velocity ratio of the site has a downward trend with the increase of PGA.At the same time,it shows that the Class D and E sites are greatly affected by PGA,and the declineing range is larger.

引言
1 数据库创建
2 BP神经网络模型搭建
3 网络模型性能分析
4 结论

表1 NEHRP规范分类<br/>Tab.1 Site classification by NEHRP specification

表1 NEHRP规范分类
Tab.1 Site classification by NEHRP specification

图1 KiK-net台网PGA>20 cm/s2记录分布<br/>Fig.1 The distribution of PGA >20 cm/s2 records

图1 KiK-net台网PGA>20 cm/s2记录分布
Fig.1 The distribution of PGA >20 cm/s2 records

图2 本文选取地震动记录的震级与震中距散点分布<br/>Fig.2 Distribution of magnitude and epicentral distance

图2 本文选取地震动记录的震级与震中距散点分布
Fig.2 Distribution of magnitude and epicentral distance

图3 根据NEHRP规范划分的场地分类统计各场地地震动记录所占比例<br/>Fig.3 Proportion of PGA at each site according to NEHRP specification

图3 根据NEHRP规范划分的场地分类统计各场地地震动记录所占比例
Fig.3 Proportion of PGA at each site according to NEHRP specification

图4 BP神经网络结构<br/>Fig.4 Multilayer perceptron of BP Neural Network

图4 BP神经网络结构
Fig.4 Multilayer perceptron of BP Neural Network

表2 本文所选6个台站场地情况<br/>Tab.2 Site conditions of the selected six stations

表2 本文所选6个台站场地情况
Tab.2 Site conditions of the selected six stations

表3 使用模型1得到VSr关于PGA变化的回归参数<br/>Tab.3 Regression parameters of shear wave velocity ratios with respect to PGA variation by Model 1

表3 使用模型1得到VSr关于PGA变化的回归参数
Tab.3 Regression parameters of shear wave velocity ratios with respect to PGA variation by Model 1

图5 本文所选6个台站的等效剪切波速比随PGA变化的预测曲线<br/>Fig.5 Predicted curves of the equivalent shear wave velocity ratio of six stations changing with PGA

图5 本文所选6个台站的等效剪切波速比随PGA变化的预测曲线
Fig.5 Predicted curves of the equivalent shear wave velocity ratio of six stations changing with PGA

表4 输入层-隐藏层神经元参数<br/>Tab.4 Neuron parameters of the input layer and the hidden layer

表4 输入层-隐藏层神经元参数
Tab.4 Neuron parameters of the input layer and the hidden layer

表5 隐藏层-输出层神经元参数<br/>Tab.5 Neuron parameters of the hidden layer and the output layer

表5 隐藏层-输出层神经元参数
Tab.5 Neuron parameters of the hidden layer and the output layer

表6 用BP神经网络模型和最小二乘法计算得到的相关系数对比<br/>Tab.6 Comparison of correlation coefficients of different models

表6 用BP神经网络模型和最小二乘法计算得到的相关系数对比
Tab.6 Comparison of correlation coefficients of different models

表7 PGA、ln(Ia)和VS30在各区间的残差分布统计<br/>Tab.7 Statistics of residual error distribution in each interval of PGA,ln(Ia)and VS30

表7 PGA、ln(Ia)和VS30在各区间的残差分布统计
Tab.7 Statistics of residual error distribution in each interval of PGA,ln(Ia)and VS30

图6 BP神经网络模型输入变量PGA(a)、ln(Ia)(b)和VS30(c)相对于输出变量VSr的残差<br/>Fig.6 Residuals of the input variables PGA(a),ln(Ia)(b)and VS30(c)of the BP neural network model corresponding to the output variable VSr

图6 BP神经网络模型输入变量PGA(a)、ln(Ia)(b)和VS30(c)相对于输出变量VSr的残差
Fig.6 Residuals of the input variables PGA(a),ln(Ia)(b)and VS30(c)of the BP neural network model corresponding to the output variable VSr

图7 BP神经网络模型输入变量PGA、ln(Ia)和VS30 相对于等效剪切波速比VSr的相关权重<br/>Fig.7 Relative importance of PGA,ln(Ia)and VS30 relative to equivalent shear wave velocity ratio VSr

图7 BP神经网络模型输入变量PGA、ln(Ia)和VS30 相对于等效剪切波速比VSr的相关权重
Fig.7 Relative importance of PGA,ln(Ia)and VS30 relative to equivalent shear wave velocity ratio VSr

图8 不同场地等效剪切波速比VSr随PGA的变化<br/>Fig.8 Variation of equivalent shear wave velocity ratio VSr with PGA in different sites

图8 不同场地等效剪切波速比VSr随PGA的变化
Fig.8 Variation of equivalent shear wave velocity ratio VSr with PGA in different sites

董凯月.2020.基于深度神经网络的地下地震动参数预测研究[D].哈尔滨:哈尔滨工业大学.
Dong K Y.2020.Study on the prediction of underground motion parameters based on deep neural network[D].Harbin:Harbin Institute of Technology.(in Chinese)
刘启方,陈长龙.2022.基于移动窗解卷积法的2011年日本东北MW9.0大地震场地非线性时变识别[J].地震学报,44(1):96-110.
Liu Q F,Chen C L.2022.Identification of the temporal changes of site nonlinearity during 2011 MW9.0 Tohoku earthquake by moving time window deconvolution method[J].Acta Seismologica Sinica,44(1):96-110.(in Chinese)
苗雨,施洋,王苏阳,等.2018.基于竖向台阵地震记录的非线性场地反应研究[J].自然灾害学报,27(6):51-58.
Miao Y,Shi Y,Wang S Y,et al.2018.Assessing nonlinear soil behavior using vertical array data:A case at TCGH16 station from KiK-Net in Japan[J].Journal of Natural Disasters,27(6):51-58.(in Chinese)
苗雨,施洋,张昊,等.2021.基于地震观测记录的场地竖向非线性性质[J].北京工业大学学报,47(7):669-679.
Miao Y,Shi Y,Zhang H,et al.2021.Investigating vertical nonlinearity of site using KiK-net seismic data[J].Journal of Beijing University of Technology,47(7):669-679.(in Chinese)
万永革,李鸿吉.1995.人工神经网络在地球物理中的应用综述[J].国际地震动态,(1):9-14.
Wan Y G,Li H J.1995.Overview of the application of artificial neural networks in Geophysics[J].Progress in Earthquake Sciences,(1):9-14.(in Chinese)
张亚沛.2019.基于KIK-NET数据库的土层特性识别及地下地震动幅值特性研究[D].哈尔滨:哈尔滨工业大学.
Zhang Y P.2019.Research on identification of soil characteristics and amplitude characteristics of underground motion based on KiK-Net database[D].Harbin:Harbin Institute of Technology.(in Chinese)
Alavi A H,Gandomi A H.2011.Prediction of principal ground-motion parameters using a hybrid method coupling artificial neural networks and simulated annealing[J].Computers & Structures,89(23-24):2176-2194.
Benito B,Herraiz M.1997.An approach to the measurement of the potential structural damage of earthquake ground motions[J].Earthquake Engineering and Structural Dynamics,26:79-92.
Beresnev I A,Wen K L.1996.Nonlinear soil response—A reality?[J].Bulletin of the Seismological Society of America,86(6):1964-1978.
Chin B H,Aki K.1991.Simultaneous study of the source path and site effects on strong ground motion during the 1989 Loma Prieta earthquake:a preliminary result on pervasive nonlinear site effects[J].Bulletin of the Seismological Society of America,81(5):1859-1884.
Chousianitis K,Del G V,Kalogeras I,et al.2014.Predictive model of Arias intensity and Newmark displacement for regional scale evaluation of earthquake-induced landslide hazard in Greece[J].Soil Dynamics and Earthquake Engineering,65(4):11-29.
Chousianitis K,Del G V,Sabatakakis N,et al.2016.Assessment of earthquake-induced landslide hazard in Greece:From Arias intensity to spatial distribution of slope resistance demand[J].Bulletin of the Seismological Society of America,106(1):174-188.
Derras B,Bard P Y,Cotton F.2014.Towards fully data driven ground-motion prediction models for Europe[J].Bulletin Earthquake Engineering,12(1):495-516.
Dhanya J,Raghukanth S T G.2017.Ground motion prediction model using artificial neural network[J].Pure and Applied Geophysics,175(3):1035-1064.
Dhanya J,Raghukanth S T G.2020.Neural network-based hybrid ground motion prediction equations for Western Himalayas and North-Eastern India[J].Acta Geophysica,172(3):762-781.
Garson G D.1991.Interpreting neural-network connection weights[J].AI Expert,6:47-51.
Gevrey M,Dimopoulos I,Lek S.2003.Review and comparison of methods to study the contribution of variables in artifificial neural network models[J].Ecological Modelling,160(3):249-264.
Ghofrani H,Atkinson G M,Goda K.2013.Implications of the 2011 M9.0 Tohoku Japan earthquake for the treatment of site effects in large earthquakes[J].Bulletin of Earthquake Engineering,11(1):171-23.
Kerb T,Ting S B.2005.Neural network estimation of ground peak acceleration at stations along Taiwan high-speed rail system[J].Engineering Applications of Artificial Intelligence,18(7):857-866.
Kingma D P,Ba J.2014.Adam:A method for stochastic optimization[C].ICLA,arXiv preprint arXiv:1412.6980.
Regnier J,Cadet H,Bonilla L F,et al.2013.Assessing nonlinear behavior of soils in seismic site response:Statistical analysis on Kik-net strong-motion data[J].Bulletin of the Seismological Society of America,103(3):1750-1770.
Rubinstein J L.2011.Nonlinear site response in medium magnitude earthquakes near Parkfield California[J].Bulletin of the Seismological Society of America,101(1):275-286.
Wang H Y,Jiang W P,Wang S Y,et al.2019.In situ assessment of soil dynamic parameters for characterizing nonlinear seismic site response using KiK-net vertical array data[J].Bulletin of Earthquake Engineering,17(5):2331-2360.
Wu C,Peng Z,Benzion Y.2010.Refined thresholds for non-linear ground motion and temporalchanges of siteresponse associated with medium-size earthquakes[J].Geophysical JournalInternational,182(3):1567-1576.

备注

引言

1 数据库创建

2 BP神经网络模型搭建

3 网络模型性能分析

4 结论

期刊信息