基于时序卷积神经网络的场地地震效应模拟

胡晓虎, 陈苏, 金立国, 傅磊, 王苏阳, 刘献伟

胡晓虎,陈苏,金立国,傅磊,王苏阳,刘献伟. 2024. 基于时序卷积神经网络的场地地震效应模拟. 地震学报,46(5):893−905. DOI: 10.11939/jass.20220226
引用本文: 胡晓虎,陈苏,金立国,傅磊,王苏阳,刘献伟. 2024. 基于时序卷积神经网络的场地地震效应模拟. 地震学报,46(5):893−905. DOI: 10.11939/jass.20220226
Hu X H,Chen S,Jin L G,Fu L,Wang S Y,Liu X W. 2024. Temporal convolution neural network model for simulation of site seismic effect. Acta Seismologica Sinica46(5):893−905. DOI: 10.11939/jass.20220226
Citation: Hu X H,Chen S,Jin L G,Fu L,Wang S Y,Liu X W. 2024. Temporal convolution neural network model for simulation of site seismic effect. Acta Seismologica Sinica46(5):893−905. DOI: 10.11939/jass.20220226

基于时序卷积神经网络的场地地震效应模拟

基金项目: 国家自然科学基金重大项目(52192675)和地震科学联合基金(U1839202)共同资助
详细信息
    作者简介:

    胡晓虎,在读硕士研究生,主要从事局部场地地震效应方面的研究,e-mail:huxiaohu5@163.com

    通讯作者:

    陈苏,博士,教授,主要从事地震工程与人工智能交叉方面的研究,e-mail:chensuchina@126.com

  • 中图分类号: TU435;P315.9

Temporal convolution neural network model for simulation of site seismic effect

  • 摘要:

    场地地震效应模拟作为岩土地震工程学的热点与难点,多基于数学物理方法或观测记录开展研究,需面对动力方程求解、建模不确定性、数据稀疏、泛化能力等问题。针对以上问题,本文构建了物理嵌入的时序卷积神经网络(Phy-TCN)模型,并验证了其与纯数据驱动的时序卷积网络(TCN)的性能差异。针对KiK-net数据库中场地井上/井下强震记录,采用Phy-TCN模型开展了场地地震效应模拟。结果表明:Phy-TCN 模型可有效模拟时序型数据;在KiK-net观测记录等含噪信号模拟中,以选取站点的地震事件特定周期点反应谱值为基准,Phy-TCN模型和等效线性化方法所得数据与实测记录的平均相对误差分别为0.067和0.379。基于上述结果认为,Phy-TCN模型可应用于土层剖面信息模糊条件的场地地震效应模拟。

    Abstract:

    The simulation of site seismic effects is a critical and challenging research area in earthquake engineering, providing a scientific basis for seismic safety evaluations of engineering sites, seismic fortification of buildings, and code revisions. Four primary research paradigms are typically employed: the empirical research paradigm, which relies on earthquake damage data; the theoretical research paradigm, which uses model experiments and mathematical tools to describe experimental phenomena; the computational research paradigm, which employs numerical methods to solve complex physical problems; and the data-driven paradigm, which utilizes machine learning tools to identify patterns in large datasets. Despite these approaches, challenges such as sparse data samples, weak generalization of results, and insufficient understanding of underlying laws persist. In this study, we introduce a fifth research paradigm, artificial intelligence for science, represented by physics-embedded deep learning. We investigate site seismic effects using strong motion records from the Japanese KiK-net array on-site/borehole stations.

    In this study, we primarily employ temporal convolution neural network (TCN) as the deep learning framework. Compared with traditional recurrent neural networks (RNNs, LSTMs, GRUs), TCN offers stronger parallelism and more flexible receptive fields. TCN uses a one-dimensional fully convolutional network architecture, with dilated causal convolutions to exponentially increase the receptive field, thus avoiding the loss of historical information when processing long sequences. Additionally, TCN uses residual blocks to prevent gradient vanishing issues. We detail how to impose physical constraints on the loss function of deep learning neural networks and develop a physics-embedded temporal convolution neural network (Phy-TCN) model. To validate the effectiveness of the Phy-TCN model, we generated a simple sparse sample dataset. Specifically, we used 30 sets of random white noise sequences with length 1000 as excitations for a single degree of freedom system to generate the sparse sample dataset, with 15 sets each for training and testing. Under sparse data conditions, we compared the performance of the Phy-TCN with a purely data-driven TCN and explored the limitations of the TCN. The results show that embedding physical information provides more information for the training process, constraining the simulation results within feasible spaces.

    Then, to further demonstrate the performance of the Phy-TCN in predicting soil layer seismic responses, we generated 30 sets of numerical simulation data, randomly dividing 20 sets for training and 10 sets for testing. These numerical simulation data were generated using a one-dimensional time-domain nonlinear site seismic response analysis method based on constructed site soil layer information. The seismic acceleration records input into the soil layer model were selected from the strong motion network database of the National Research Institute for Earth Science and Disaster Resilience in Japan, and the constitutive model used to describe the nonlinear behavior of the soil was a hybrid hyperbolic nonlinear soil model. The results show that the Phy-TCN can effectively simulate site seismic effects under seismic excitation. Comparing the predicted results with reference records, the coefficient of determination (R2) is generally greater than 0.97.

    Finally, in order to verify the application of the Phy-TCN model in practical engineering, we selected 50 seismic events with surface peak accelerations greater than or equal to 0.3 m/s2 from the KiK-net database at the IBRH11 and IBRH12 stations, randomly dividing 40 events for training and 10 events for testing, and conducted site seismic effect simulations using the Phy-TCN model. To illustrate the superiority of the Phy-TCN model, we used the equivalent linearization method, commonly used in engineering, to calculate the test set and compared the results with the Phy-TCN simulations. The results show that in the simulation of noisy signals such as KiK-net observation records, based on the response spectrum values of specific periodic points of seismic events at selected sites, the average relative errors of the Phy-TCN and the equivalent linearization method compared with the measured records are 0.067 and 0.379, respectively. As the intensity of seismic motion continues to increase, the measured surface peak acceleration gradually exceeds the surface peak acceleration simulated by the equivalent linearization method, while the surface peak acceleration simulated by the Phy-TCN method remains stable within a deviation range of ±20%. The simulation capability of the Phy-TCN remains strong, with the coefficient of determination (R2) generally remaining above 0.8. Under conditions of high uncertainty in shear wave velocity and soil dynamic parameters, or in the absence of suitable soil profile information, the simulation accuracy of the Phy-TCN model is higher than that of the equivalent linearization method.

  • 强震动观测和历史震害都表明局部场地条件会对地表地震动强度和震害产生重要影响。例如,局部软弱土层会造成地震动特定周期成分显著放大,从而导致位于软弱厚土层场地上的震害明显加重。场地条件对地震动参数的影响主要体现在地震动幅值(峰值、反应谱幅值)和频谱特性(反应谱特征周期)的变化上。对于场地地震效应模拟研究:Idriss和Seed (1968)提出了一维场地地震反应的等效线性方法,并编制出计算程序SHAKE系列;廖振鹏(1989)开发了具有自主知识产权的计算程序LSSRLI-1,并被广泛应用于重大工程地震安全性评价中。时域方法以土动力非线性为核心,例如Hashash和Park (2001)开发的DEEPSOIL计算程序等。

    近年来,人工智能方法已在地震工程多个领域实现交叉融合:Peng等(2021)采用深度学习方法对六层钢框架的动力响应进行分析,验证了深度学习方法的合理性;Li等(2022)开发了一种深度学习模型来快速预测高墩桥的抗震性能;Zhang等(2021a)提出使用长短期记忆(long short term memory,缩写为LSTM)的深度学习方法预测岩土应力-应变行为;Liu和Dai (2021)提出了一种实时预测地震动烈度趋势的深度学习模型;Roten和Olsen (2021)使用神经网络从离散剪切波速剖面中预测场地放大效应;刘献伟等(2023)采用动态聚类方法对日本相模湾海底台站的水平垂直谱比进行分类,获取了最优场地分类;任叶飞等(2023)基于K均值聚类(K-means)方法根据台阵真实场地放大因子对我国现行规范中Ⅱ类场地进行聚类分析,并给出了新的场地分类优化方案;Derakhshani和Foruzan (2019)基于深度学习,针对太平洋地震研究中心(Pacific Earthquake Engineering Research,缩写为PEER)的NGA-West2数据库,建立了峰值加速度(peak groud acceleration,缩写为PGA)和峰值速度(peak ground velocity,缩写为PGV)等地震动参数衰减模型;Zhang等(2019)针对数据驱动的结构地震响应建模,提出了两种LSTM网络,通过在有限数据集中训练,其模型能够准确预测建筑结构的弹性和非弹性响应。在人工智能的发展进程中,时序卷积神经网络已在各个领域得到应用,如能源燃料、声学、信息科技等,具体信息列于表1

    表  1  时序卷积神经网络模型在各领域的应用
    Table  1.  Application of temporal convolutional neural network model in various fields
    研究领域应用场景作者
    能源燃料可再生资源的超短期时空预测Liang,Tang (2 022
    电力系统暂态稳定评估模型刘聪等 (2 023
    电力系统短期负荷预测Yin,Xie (2 021
    分布式能源概率多周期预测Loschenbrand (2 021
    热负荷预测模型Song等 (2 020
    西班牙国家电力需求与电动汽车充电站电力需求模型Lara-Benítez等 (2 020
    声学高质量头部相关传递函数(HRTF)Gebru等 (2 021
    一种高效的端到端的句子级唇读模型Zhang等 (2 021b
    信息科技通用日志序列异常检测框架杨瑞朋等 (2 020
    社交物联网中情感识别Xiao等 (2 021
    医学一种用于识别胃旁路手术中手术阶段与手术步骤的模型Ramesh等 (2 021
    一种用于自动诊断脓毒症的自动化工具Kok等 (2 020
    机械工业工业设备剩余寿命预测模型刘丽等 (2 022
    金融融合情感特征的股价预测模型严冬梅等 (2 022
    地球科学复杂地层波阻抗反演模型王德涛,陈国雄 (2 022
    气象学高分辨的中短期区域天气预报模型Hewage等 (2 020
    下载: 导出CSV 
    | 显示表格

    传统深度学习多基于纯数据进行训练,缺乏物理可解释性。本文针对场地地震效应模拟,提出基于物理嵌入的时序卷积神经网络的深度学习框架,不仅可以进一步增强从数据中学习的鲁棒性和可靠性,而且具有一定的可解释性。同时,在稀疏数据情况下,嵌入式物理信息可以为训练学习过程提供更多信息,从而将模拟结果约束在可行空间内。本文将通过单自由度体系,对物理信息嵌入网络的可行性和必要性进行验证,并利用构建的场地地震效应模拟数据集,给出方法的适用性与可行性,最后使用KiK-net台站中典型井下/井上记录,实现基于物理嵌入的时序卷积神经网络(physics-embedded temporal convolutional neural network,缩写为Phy-TCN)的代理模型构建。

    时序卷积神经网络(TCN)(Bai et al,2018)具有比循环神经网络(recurrent neural network,缩写为RNN)、LSTM和门控循环单元(gated recurrent unit,缩写为GRU)更强的并行性和更为灵活的感受野,为了使得TCN实现输入序列与输出序列的步长相同,TCN采用一维全卷积网络架构(Long et al,2017),通过零填充方式使得每一个隐藏层中输入与输出长度相同。同时,我们也引入膨胀因果卷积(Yu,Koltun,2016),使感受野实现指数级增长,以此避免网络在处理长序列时增长得太深,同时保证历史信息不遗漏,TCN中膨胀因果卷积如图1a所示。对于超长序列数据,需要训练更深层次的网络以使模型足够强大。在网络模型构建中,TCN考虑使用残差块(He et al,2016)来避免梯度消失,TCN残差块如图1b所示。

    图  1  时序卷积神经网络(TCN)组构
    (a) 膨胀因果卷积;(b) TCN残差块
    Figure  1.  Architectural elements in temporal convolutional neural network (TCN)
    (a) Dilated causal convolution;(b) TCN residual block

    本文提出的网络结构是将物理过程嵌入神经网络损失函数中,从而实现物理-数据双驱动模式。以单自由度Phy-TCN架构及物理概念为例,系统控制方程为:

    $$ \ddot {\boldsymbol{u}} + 2\xi \omega \dot {\boldsymbol{u}} + {\omega ^2}{\boldsymbol{u}} = - {\ddot {\boldsymbol{u}}_{\rm{g}}}\text{,} $$ (1)

    令$2\xi \omega \dot {\boldsymbol{u}} + {\omega ^2}{\boldsymbol{u}} = g$,则

    $$ \ddot {\boldsymbol{u}} + g + {\ddot {\boldsymbol{u}}_{\rm{g}}} \to 0 \text{,} $$ (2)

    式中,ξ为阻尼比,ω为体系圆频率,g为质量归一化内力,${\boldsymbol{u}} $,$\dot {\boldsymbol{u }}$,$\ddot {\boldsymbol{u}} $,$\ddot {\boldsymbol{u}}_ {\rm{g}}$分别为相对位移、相对速度、相对加速度和输入基底加速度。

    针对此物理系统,本文采用Phy-TCN建模,该模型由TCN网络和差分求解器组成,如图2所示,包含输入层、输出层。其中,TCN残差块层和全连接层用于信息提取和特征学习。差分求解器用于神经网络学习过程中对输出层进行求导,如输出状态空间zt)中${\boldsymbol{u}} $和$\dot {{{\boldsymbol{u}} }}$计算导数得${\boldsymbol{u}}_{t} $和$\dot {\boldsymbol{u}} _t$,以此在损失函数中嵌入物理约束;模型将加速度激励时间序列信号作为输入,将相对位移、速度、加速度等状态变量作为输出,通过优化网络参数使得整个网络模型满足物理与数据同时约束。总体损失函数定义如下:

    图  2  Phy-TCN框架
    Figure  2.  Phy-TCN architecture
    $$ {J_{{\rm{T}}}} ( \theta ) = {J_{\rm{P}}} ( \theta ) + {J_{\rm{D}}} ( \theta ) \text{,} $$ (3)
    $$ {J_{\rm{P}}} ( \theta ) = \frac{1}{N}\left( {{\beta _1}\left\| {\dot {\boldsymbol{u}} - {{\boldsymbol{u}}_t}} \right\| + {\beta _2}\left\| {\ddot {\boldsymbol{u}} - {{\dot {\boldsymbol{u}}}_t}} \right\| + {\beta _3}\left\| {{{\dot {\boldsymbol{u}}}_t} + g + {{\ddot {\boldsymbol{u}}}_{\rm{g}}}} \right\|} \right) \text{,} $$ (4)
    $$ {J_{\rm{D}}} ( \theta ) = \frac{1}{N}\left( {{\alpha _1}\left\| {{\boldsymbol{u}} - {{\boldsymbol{u}}^ * }} \right\| + {\alpha _2}\left\| {\dot {\boldsymbol{u}} - {{\dot {\boldsymbol{u}}}^ * }} \right\| + {\alpha _3}\left\| {\ddot {\boldsymbol{u}} - {{\ddot {\boldsymbol{u}}}^ * }} \right\|} \right) \text{,} $$ (5)

    式中:JT表示总体损失函数,JP表示物理损失函数,JD表示数据损失函数,θ为网络中训练参数;${\boldsymbol{u}}^* $,$\dot {\boldsymbol{u}}^*$,${{\ddot {\boldsymbol{u}}}^ * } $表示实测相对位移、速度、加速度;β1β2β3α1α2α3为各损失函数的权重系数,可根据各损失函数量级进行估计。Phy-TCN的模型框架如图2所示,训练采用Adam (Kingma,Ba,2015)结合L-BFGS优化策略(Liu,Nocedal,1989)实现空间最优解。

    随机选取30组时序长度为1 000的白噪声作为单自由度体系激励,采用数值方法生成采样频率为20 Hz的加速度数据集,各取15组分别将其划分到训练集与测试集中。网络参数设置如下:$ [ $m×n×p$ ] $=$ [ $15×1000×1$ ] $,p1=10,q1=10,q=3,膨胀因子d=$ [ $1,2,4,8,16$ ] $,卷积核k=3。然后将嵌入物理信息的Phy-TCN与TCN模型进行对比。为了说明物理嵌入的必要性,本文从测试集中挑选两组工况进行阐明。图3a3b对应工况一下Phy-TCN模型与TCN模型模拟所得位移、加速度与参考值的对比结果,图3c3d对应工况二。从图中可以看出,TCN模型预测结果毛刺现象明显,而Phy-TCN模型模拟结果并无此现象,且能得到更合理的模拟结果。由此可见:嵌入物理信息的模型Phy-TCN相对TCN模型泛化能力更强。从图3c3d所示工况二中可以看出,在0—2 s时TCN模拟结果误差较大,这是因为TCN模型训练时没有0 s以前的前序时程作为样本,导致其对0 s后若干时序长度的模拟结果不准确,而采用物理嵌入时序卷积神经网络模型将进一步约束解空间,使得提出的网络结果的性能得到优化,模拟结果与参考记录对比更为准确。

    图  3  Phy-TCN与TCN在单自由度测试集下的比较结果
    图(a)和图(b)为工况一对应的位移和加速度比较;图(c)和图(d)为工况二对应的位移和加速度比较
    Figure  3.  Comparison of the Phy-TCN and TCN in single degree of freedom test set
    Figs. (a) and (b) are the comparisons of displacement and acceleration in case 1;Figs. (c) and (d) are those in case 2

    进一步将应用于非线性场地地震效应分析模型,通过构建设定钻孔,采用Shi和Asimaki (2017)提出的混合双曲(hybrid hyperbolic,缩写为HH)非线性土体本构模型来描述土体非线性动力行为,表达式如下:

    $$ {\tau }_{{\rm{HH}}} ( \gamma ) =w ( \gamma ) \dfrac{{G}_{\mathrm{max}}\gamma }{1+\beta {\left(\dfrac{\gamma }{{\gamma }_{{\rm{ref}}}}\right)}^{s}}+\left[1+w ( \gamma ) \right]\dfrac{{\gamma }^{d}\mu }{\dfrac{1}{{G}_{\mathrm{max}}+\dfrac{{\gamma }^{d}\mu }{{\tau }_{{\rm{f}}}}}} \text{,} $$ (6)
    $$ w ( \gamma ) = 1 -\frac{{ 1}}{{{1 + {{10}^{ - a ( {{{\lg }}{\frac{\gamma }{{{\gamma _{\rm{t}}}}}} - 4.039{a^{ - 1.036}}} ) }}} }} \text{,} $$ (7)

    式中,β=1,s=0.919,a=1,γ为剪切应变,γt为转变应变,μh为两个附加参数,Gmax为最大剪切模量,τf为抗剪强度,γref为参考剪应变。

    各土层具体参数及HH模型参数详见表2。从日本国家地球科学与灾害防御研究所(The National Research Institute for Earth Science and Disaster Resilience,缩写为NIED)的强震台网数据库中选择30组地震记录作为输入,包括20个样本的训练数据集和10个样本的测试数据集,来测试训练模型的泛化性能。其中网络参数设置为:$ [ $m×n×p$ ] $=$ [ $6×1200×1$ ] $,p1=30,q1=10,q=3,膨胀因子d=$ [ $1,2,4,8,16,32,64$ ] $,卷积核k=3。

    表  2  土层力学参数及混合双曲模型相关参数
    Table  2.  Mechanics parameters in soil layer and related parameters of hybrid hyperbolic model
    层号土层厚度/mvS/(m·s−1密度/(kg·m−3μ转变应变γth
    抗剪强度
    τf/kPa
    参考剪应变
    γref
    最大剪切模量
    Gmax/MPa
    1412016000.3080.048%0.93831.1700.034%23.760
    2436018000.0730.011%0.721143.1830.046%256.835
    3850018000.0610.010%0.699245.2760.047%503.375
    41262018000.0580.010%0.688358.8980.056%776.576
    522120020001.0000.010%0.941547.3180.073%3173.761
    下载: 导出CSV 
    | 显示表格

    由于Phy-TCN模型在测试集上模拟场地地震效应表现良好,而其与实际记录反应谱的对比显示,其决定系数大于0.97,故仅选两组对比结果进行展示(图4)。由图4a可知:决定系数R2=0.983时,模拟和实测加速度反应谱曲线、加速度、速度以及位移曲线基本匹配;由图4b可知:决定系数R2=0.992时,模拟和实测加速度反应谱曲线、加速度、速度以及位移结果匹配良好。由此可见,Phy-TCN在模拟一维非线性场地地震效应方面表现良好。

    图  4  场地模拟结果与实测结果对比图
    图(a)和图(b)分别为决定系数R2为0.983和0.992时所对应的加速度反应谱、加速度、速度及位移时程
    Figure  4.  Comparison of simulated results and measured results in simple site
    Figs. (a) and (b) are the comparison results of acceleration response spectra,acceleration,velocity and displacement time histories when determination coefficient R2 are 0.983 and 0.992,respectively

    采用KiK-net井上/井下台站对的强地震动记录作为数据。选择IBRH11和IBRH12站点,对应场地速度剖面如图5所示。选择地表峰值加速度大于等于0.3 m/s2的50个地震事件,随机将其中40个地震事件对应的两个水平方向井下地表加速度记录划分为训练集;将其余10个地震事件划分为测试集。整个网络模型参数设置为:$ [ $m×n×p$ ] $=$ [ $10×12 000×1$ ] $,p1=80,q1=10,q=3,膨胀因子d=$ [ $1,2,4,8,16,32,64,128$ ] $,卷积核k=3。为了说明Phy-TCN模型性能,采用等效线性法对测试集进行计算,并与Phy-TCN模拟结果对比。等效线性法采用SeismoSoil计算软件(Asimaki,Shi,2017),首先根据各土层剪切波速对现有土层重新离散化,如式8所示,基岩类型选择为刚性基岩;然后通过Darendeli (2001)提出的方法根据已知场地剪切波速合成不同深度处的土层非线性曲线(图6);最后将地震动输入形式选择为从井下输入。计算过程中当连续两次剪切模量G与阻尼比ξ之间的相对误差小于7.5%或进行10次迭代后,结束迭代。

    图  5  IBRH11与IBRH12场地剪切波速
    Figure  5.  Shear-wave-velocity of stations IBRH11 and IBRH12
    $$ \Delta h = \frac{{{v_{\rm{S}}}}}{{10 f}} \text{,} $$ (8)
    图  6  IBRH11 (a)与IBRH12 (b)场地不同土层土的归一化剪切模量和阻尼比的非线性曲线
    Figure  6.  G/Gmax and damping ratio nonliner curves of soil for different layer at the sites IBRH11 (a) and IBRH12 (b)

    式中,Δh为空间网格大小,vS为土层剪切波速,f为谐波分量频率。

    Phy-TCN与等效线性方法模拟地表峰值加速度对比结果如图7a所示,可见,随着地震动强度的增加,实测地表峰值加速度渐渐大于等效线性法模拟的地表峰值加速度,而Phy-TCN方法模拟的地表峰值加速度则稳定在偏差为±20%范围内不变。图7b展示了Phy-TCN与等效线性法基于IBRH11和IBRH12场地测试集共计40组样本所得的模拟反应谱的决定系数与实测地表峰值加速度之间的关系,可见,随着地震动强度的增加,对Phy-TCN的模拟能力干扰并不多,且决定系数R2基本保持在0.8以上。

    图  7  Phy-TCN与等效线性方法在IBRH11与IBRH12场地计算得到的地表峰值加速度(a)及决定系数(b)对比
    Figure  7.  Comparison of the peak ground acceleration (a) and coefficient of determination (b) for the sites IBRH11 and IBRH12 calculated by Phy-TCN and equivalent linear method

    为了说明Phy-TCN在场地地震效应的模拟能力,分别从两个场地中挑选两组地震事件的两个水平方向地震动进行模拟(图8)。在图8a的两个地震事件的模拟中,Phy-TCN相对等效线性法能够较合理地模拟出主要周期段0.1—0.4 s (上)及0.08—0.2 s (下)处的加速度反应谱值;在图8b中,两种方法基本上都能识别出主要周期段,如0.06—0.2 s (上)和0.09—0.2 s (下),Phy-TCN能够更精确地识别出主要周期段的加速度反应谱值。在图8b左上图中的周期0.5—0.7 s及右上图中的0.2—0.7 s,等效线性法过高地估计了加速度反应谱值,而Phy-TCN结果则较为合理。对每个地震事件上述方法相对误差进行比较(表3),可见Phy-TCN、等效线性法与实测结果平均相对误差分别为0.067和0.379。Phy-TCN与等效线性法模拟不确定性主要体现在参数不确定性方面,由于实际土层信息的不完备带来等效线性法土动力学参数的不确定性,相比之下,Phy-TCN模型参数不确定性主要为网络权重的不确定,本文物理嵌入和数据驱动联合训练可近一步将网络权重约束到最优状态。另外在建模不确定性方面,等效线性法建模假设在应用中存在一定的失真。Phy-TCN模型直接训练实际强震动记录,并未设置假设条件。从图7可知:在剪切波速以及土层力学参数不确定性较强的情况下或无合适土层剖面信息情况下,Phy-TCN模型模拟结果的精度较等效线性法更高。

    图  8  Phy-TCN与等效线性法基于IBRH11场地(a)和IBRH12场地(b)的两个地震事件的加速度反应谱模拟结果与实测结果的对比
    Figure  8.  Comparison of simulated and measured results of acceleration response spectra for two earthquake events in different horizontal directions at the sites IBRH11 (a) and IBRH12 (b) calculated by Phy-TCN and equivalent linear method
    表  3  IBRH11和IBRH12场地不同地震事件实测最大加速度反应谱下周期点地表加速度反应谱值比较
    Table  3.  Comparison of acceleration response spectra at periodic points under different measured maximum acceleration response spectra of earthquake events at the sites IBRH11and IBRH12
    场地地震事件方向周期/s$S_{{\rm{a}}}^{\rm{EQ}} $
    /(m·s−2
    $S_{{\rm{a}}}^{\rm{orig}} $
    /(m·s−2
    $S_{{\rm{a}}}^{\rm{Phy-TCN}} $
    /(m·s−2
    等效线性反应谱值相对误差
    (|$ S_{ {\rm{a} } }^{\rm{EQ} } $–$S_{ {\rm{a} } }^{\rm{orig} } $|/$S_{ {\rm{a} } }^{\rm{orig} } $)
    Phy-TCN反应谱值相对误差
    (|$ S_{ {\rm{a} } }^{\rm{Phy-TCN} } $–$S_{ {\rm{a} } }^{\rm{orig} } $|/$S_{ {\rm{a} } }^{\rm{orig} } $)
    IBRH111 104 021 656EW0.121.422.712.480.4760.085
    NS0.381.583.192.990.5050.063
    1 303 180 653EW0.200.972.962.740.6720.074
    NS0.143.042.962.760.0270.068
    IBRH121 104 111 758EW0.141.151.511.460.2380.033
    NS0.140.921.441.530.3610.063
    1 206 281 452EW0.142.783.983.770.3020.053
    NS0.151.863.393.710.4510.094
    注:$S_{ {\rm{a} } }^{\rm{EQ} } $为等效线性化方法求得的反应谱值,$S_{ {\rm{a} } }^{\rm{Phy-TCN} } $为Phy-TCN方法求得的反应谱值,$S_{ {\rm{a} } }^{\rm{orig} } $为实测反应谱值.
    下载: 导出CSV 
    | 显示表格

    本文提出了基于物理嵌入时序卷积神经网络(Phy-TCN)模型,并应用在场地地震效应模拟中,所得结论如下:

    1) 构建的Phy-TCN与TCN相比,嵌入物理约束使得时序卷积神经网络泛化能力更强,通过考虑将物理信息嵌入到网络损失函数中,以实现物理模型的数字孪生。

    2) 针对工程场地,通过构建的Phy-TCN模型来模拟场地效应,结果表明,此模型能够较好地模拟地震动激励下场地的地震效应。基于KiK-net典型台站井上/井下地震记录,针对不同场地建立Phy-TCN模型并与等效线性法进行对比,结果表明,Phy-TCN模型在土层剖面信息及土层力学参数不确定性较强的情况下,可较合理地模拟场地地震效应。

  • 图  1   时序卷积神经网络(TCN)组构

    (a) 膨胀因果卷积;(b) TCN残差块

    Figure  1.   Architectural elements in temporal convolutional neural network (TCN)

    (a) Dilated causal convolution;(b) TCN residual block

    图  2   Phy-TCN框架

    Figure  2.   Phy-TCN architecture

    图  3   Phy-TCN与TCN在单自由度测试集下的比较结果

    图(a)和图(b)为工况一对应的位移和加速度比较;图(c)和图(d)为工况二对应的位移和加速度比较

    Figure  3.   Comparison of the Phy-TCN and TCN in single degree of freedom test set

    Figs. (a) and (b) are the comparisons of displacement and acceleration in case 1;Figs. (c) and (d) are those in case 2

    图  4   场地模拟结果与实测结果对比图

    图(a)和图(b)分别为决定系数R2为0.983和0.992时所对应的加速度反应谱、加速度、速度及位移时程

    Figure  4.   Comparison of simulated results and measured results in simple site

    Figs. (a) and (b) are the comparison results of acceleration response spectra,acceleration,velocity and displacement time histories when determination coefficient R2 are 0.983 and 0.992,respectively

    图  5   IBRH11与IBRH12场地剪切波速

    Figure  5.   Shear-wave-velocity of stations IBRH11 and IBRH12

    图  6   IBRH11 (a)与IBRH12 (b)场地不同土层土的归一化剪切模量和阻尼比的非线性曲线

    Figure  6.   G/Gmax and damping ratio nonliner curves of soil for different layer at the sites IBRH11 (a) and IBRH12 (b)

    图  7   Phy-TCN与等效线性方法在IBRH11与IBRH12场地计算得到的地表峰值加速度(a)及决定系数(b)对比

    Figure  7.   Comparison of the peak ground acceleration (a) and coefficient of determination (b) for the sites IBRH11 and IBRH12 calculated by Phy-TCN and equivalent linear method

    图  8   Phy-TCN与等效线性法基于IBRH11场地(a)和IBRH12场地(b)的两个地震事件的加速度反应谱模拟结果与实测结果的对比

    Figure  8.   Comparison of simulated and measured results of acceleration response spectra for two earthquake events in different horizontal directions at the sites IBRH11 (a) and IBRH12 (b) calculated by Phy-TCN and equivalent linear method

    表  1   时序卷积神经网络模型在各领域的应用

    Table  1   Application of temporal convolutional neural network model in various fields

    研究领域应用场景作者
    能源燃料可再生资源的超短期时空预测Liang,Tang (2 022
    电力系统暂态稳定评估模型刘聪等 (2 023
    电力系统短期负荷预测Yin,Xie (2 021
    分布式能源概率多周期预测Loschenbrand (2 021
    热负荷预测模型Song等 (2 020
    西班牙国家电力需求与电动汽车充电站电力需求模型Lara-Benítez等 (2 020
    声学高质量头部相关传递函数(HRTF)Gebru等 (2 021
    一种高效的端到端的句子级唇读模型Zhang等 (2 021b
    信息科技通用日志序列异常检测框架杨瑞朋等 (2 020
    社交物联网中情感识别Xiao等 (2 021
    医学一种用于识别胃旁路手术中手术阶段与手术步骤的模型Ramesh等 (2 021
    一种用于自动诊断脓毒症的自动化工具Kok等 (2 020
    机械工业工业设备剩余寿命预测模型刘丽等 (2 022
    金融融合情感特征的股价预测模型严冬梅等 (2 022
    地球科学复杂地层波阻抗反演模型王德涛,陈国雄 (2 022
    气象学高分辨的中短期区域天气预报模型Hewage等 (2 020
    下载: 导出CSV

    表  2   土层力学参数及混合双曲模型相关参数

    Table  2   Mechanics parameters in soil layer and related parameters of hybrid hyperbolic model

    层号土层厚度/mvS/(m·s−1密度/(kg·m−3μ转变应变γth
    抗剪强度
    τf/kPa
    参考剪应变
    γref
    最大剪切模量
    Gmax/MPa
    1412016000.3080.048%0.93831.1700.034%23.760
    2436018000.0730.011%0.721143.1830.046%256.835
    3850018000.0610.010%0.699245.2760.047%503.375
    41262018000.0580.010%0.688358.8980.056%776.576
    522120020001.0000.010%0.941547.3180.073%3173.761
    下载: 导出CSV

    表  3   IBRH11和IBRH12场地不同地震事件实测最大加速度反应谱下周期点地表加速度反应谱值比较

    Table  3   Comparison of acceleration response spectra at periodic points under different measured maximum acceleration response spectra of earthquake events at the sites IBRH11and IBRH12

    场地地震事件方向周期/s$S_{{\rm{a}}}^{\rm{EQ}} $
    /(m·s−2
    $S_{{\rm{a}}}^{\rm{orig}} $
    /(m·s−2
    $S_{{\rm{a}}}^{\rm{Phy-TCN}} $
    /(m·s−2
    等效线性反应谱值相对误差
    (|$ S_{ {\rm{a} } }^{\rm{EQ} } $–$S_{ {\rm{a} } }^{\rm{orig} } $|/$S_{ {\rm{a} } }^{\rm{orig} } $)
    Phy-TCN反应谱值相对误差
    (|$ S_{ {\rm{a} } }^{\rm{Phy-TCN} } $–$S_{ {\rm{a} } }^{\rm{orig} } $|/$S_{ {\rm{a} } }^{\rm{orig} } $)
    IBRH111 104 021 656EW0.121.422.712.480.4760.085
    NS0.381.583.192.990.5050.063
    1 303 180 653EW0.200.972.962.740.6720.074
    NS0.143.042.962.760.0270.068
    IBRH121 104 111 758EW0.141.151.511.460.2380.033
    NS0.140.921.441.530.3610.063
    1 206 281 452EW0.142.783.983.770.3020.053
    NS0.151.863.393.710.4510.094
    注:$S_{ {\rm{a} } }^{\rm{EQ} } $为等效线性化方法求得的反应谱值,$S_{ {\rm{a} } }^{\rm{Phy-TCN} } $为Phy-TCN方法求得的反应谱值,$S_{ {\rm{a} } }^{\rm{orig} } $为实测反应谱值.
    下载: 导出CSV
  • 廖振鹏. 1989. 地震小区划: 理论与实践[M]. 北京: 地震出版社: 141–153.

    Liao Z P. 1989. Seismic Microzonation: Theory and Practice[M]. Beijing: Seismological Press: 141–153 (in Chinese).

    刘聪, 刘颂凯, 刘礼煌, 张磊, 谭瑞, 张雅婷. 2023. 基于TCN-CCRELMS的电力系统暂态稳定评估[J]. 电力系统及其自动化学报, 35(7): 36−44.

    Liu C, Liu S K, Liu L H, Zhang L, Tang R, Zhang Y T. 2023. Transient strability assessment of power system based on TCN-CCRELMS[J]. Proceedings of the CSU-EPSA, 35(7): 36−44 (in Chinese).

    刘丽,裴行智,雷雪梅. 2022. 基于时间卷积注意力网络的剩余寿命预测方法[J]. 计算机集成制造系统,28(8):2375–2386.

    Liu L,Pei X Z,Lei X M. 2022. Temporal convolutional attention network for remaining useful life estimation[J]. Computer Integrated Manufacturing Systems,28(8):2375–2386 (in Chinese).

    刘献伟,陈苏,李小军,傅磊,胡进军,孙浩. 2023. 基于HVSR谱比动态聚类的海域场地特性研究[J]. 岩土工程学报,45(1):213–220.

    Liu X W,Chen S,Li X J,Fu L,Hu J J,Sun H. 2023. Characteristics of marine site based on HVSR dynamic clustering method[J]. Chinese Journal of Geotechnical Engineering,45(1):213–220 (in Chinese).

    任叶飞,刘也,张鹏,冀昆,王宏伟,温瑞智,谢俊举. 2023. 基于聚类分析的我国工程场地分类方案优化研究[J]. 建筑结构学报,44(11):226–235. doi: 10.14006/j.jzjgxb.2022.0273

    Ren Y F,Liu Y,Zhang P,Ji K,Wang H W,Wen R Z,Xie J J. 2023. Research on optimization of site classification scheme in Chinese seismic code based on clustering analysis[J]. Journal of Building Structures,44(11):226–235 (in Chinese).

    王德涛,陈国雄. 2022. 基于时间卷积网络的地震波阻抗反演[J]. 地球科学,47(4):1492–1506. doi: 10.3321/j.issn.1000-2383.2022.4.dqkx202204021

    Wang D T,Chen G X. 2022. Seismic wave impedance inversion based on temporal convolutional network[J]. Earth Science,47(4):1492–1506 (in Chinese).

    严冬梅,何雯馨,陈智. 2022. 融合情感特征的基于RoBERTa-TCN的股价预测研究[J]. 数据分析与知识发现,6(12):123–134.

    Yan D M,He W X,Chen Z. 2022. Research on stock price prediction based on RoBERTa-TCN with emotional characteristics[J]. Data Analysis &Knowledge Discovery,6(12):123–134 (in Chinese).

    杨瑞朋,屈丹,朱少卫,钱叶魁,唐永旺. 2020. 基于改进时间卷积网络的日志序列异常检测[J]. 计算机工程,46(8):50–57.

    Yang R P,Qu D,Zhu S W,Qian Y K,Tang Y W. 2020. Anomaly detection for log sequence based on improved temporal convolutional network[J]. Computer Engineering,46(8):50–57 (in Chinese).

    Asimaki D, Shi J. 2017. SeismoSoil User Manual, V1.3[M]. California: GeoQuake Research Group, California Institute of Technology: 6−12.

    Bai S J, Kolter J Z, Koltun V. 2018. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling[DB/OL]. [2022-12-01]. https://arxiv.org/abs/1803.01271.

    Darendeli M B. 2001. Development of a New Family of Normalized Modulus Reduction and Material Damping Curves[D]. Austin: The University of Texas: 220−272.

    Derakhshani A,Foruzan A H. 2019. Predicting the principal strong ground motion parameters:A deep learning approach[J]. Appl Soft Comput,80:192–201. doi: 10.1016/j.asoc.2019.03.029

    Gebru I D, Marković D, Richard A, Krenn S, Butler G A, De la Torre F, Sheikh Y. 2021. Implicit HRTF modeling using temporal convolutional networks[C]//ICASSP 2021−2021 IEEE International Conference on Acoustics, Speech and Signal ProcessingICASSP). Toronto: IEEE: 3385–3389.

    Hashash Y M A,Park D. 2001. Non-linear one-dimensional seismic ground motion propagation in the Mississippi embayment[J]. Eng Geol,62(1/2/3):185–206.

    He K M, Zhang X Y, Ren S Q, Sun J. 2016. Deep residual learning for image recognition[C]//Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE: 770–778.

    Hewage P,Behera A,Trovati M,Pereira E,Ghahremani M,Palmieri F,Liu Y H. 2020. Temporal convolutional neural (TCN) network for an effective weather forecasting using time-series data from the local weather station[J]. Soft Comput,24(21):16453–16482. doi: 10.1007/s00500-020-04954-0

    Idriss I M,Seed H B. 1968. Seismic response of horizontal soil layers[J]. J Soil Mech Found Div,94(4):1003–1031. doi: 10.1061/JSFEAQ.0001163

    Kingma D P, Ba J. 2015. Adam: A method for stochastic optimization[C]//Proceedings of the 3rd International Conference on Learning Representations. San Diego: Institute for Catastrophic Loss Reduction: abs/1412.6980.

    Kok C,Jahmunah V,Oh S L,Zhou X J,Gururajan R,Tao X H,Cheong K H,Gururajan R,Molinari F,Acharya U R. 2020. Automated prediction of sepsis using temporal convolutional network[J]. Comput Biol Med,127:103957. doi: 10.1016/j.compbiomed.2020.103957

    Lara-Benítez P,Carranza-García M,Luna-Romera J M,Riquelme J C. 2020. Temporal convolutional networks applied to energy-related time series forecasting[J]. Appl Sci,10(7):2322. doi: 10.3390/app10072322

    Li C X,Li H,Chen X. 2022. Fast seismic response estimation of tall pier bridges based on deep learning techniques[J]. Eng Struct,266:114566. doi: 10.1016/j.engstruct.2022.114566

    Liang J K,Tang W Y. 2022. Ultra-short-term spatiotemporal forecasting of renewable resources:An attention temporal convolutional network-based approach[J]. IEEE Trans Smart Grid,13(5):3798–3812. doi: 10.1109/TSG.2022.3175451

    Liu D C,Nocedal J. 1989. On the limited memory BFGS method for large scale optimization[J]. Math Program,45(1):503–528.

    Liu T,Dai Z J. 2021. Real-time prediction of the trend of ground motion intensity based on deep learning[J]. Shock Vibrat,2021:5518204.

    Long J,Shelhamer E,Darrell T. 2017. Fully convolutional networks for semantic segmentation[J]. IEEE Trans Pattern Anal Mach Intell,39(4):640–651. doi: 10.1109/TPAMI.2016.2572683

    Loschenbrand M. 2021. A temporal neural network model for probabilistic multi-period forecasting of distributed energy resources[J]. IEEE Access,9:147029–147041. doi: 10.1109/ACCESS.2021.3121988

    Peng H,Yan J W,Yu Y,Luo Y Z. 2021. Time series estimation based on deep learning for structural dynamic nonlinear prediction[J]. Structures,29:1016–1031. doi: 10.1016/j.istruc.2020.11.049

    Ramesh S,Dall’Alba D,Gonzalez C,Yu T,Mascagni P,Mutter D,Marescaux J,Fiorini P,Padoy N. 2021. Multi-task temporal convolutional networks for joint recognition of surgical phases and steps in gastric bypass procedures[J]. Int J Comput Assist Radiol Surg,16(7):1111–1119. doi: 10.1007/s11548-021-02388-z

    Roten D,Olsen K B. 2021. Estimation of site amplification from geotechnical array data using neural networks[J]. Bull Seismol Soc Am,111(4):1784–1794. doi: 10.1785/0120200346

    Shi J,Asimaki D. 2017. From stiffness to strength:Formulation and validation of a hybrid hyperbolic nonlinear soil model for site-response analyses[J]. Bull Seismol Soc Am,107(3):1336–1355. doi: 10.1785/0120150287

    Song J C,Xue G X,Pan X H,Ma Y P,Li H. 2020. Hourly heat load prediction model based on temporal convolutional neural network[J]. IEEE Access,8:16726–16741. doi: 10.1109/ACCESS.2020.2968536

    Xiao G R,Tu G,Zheng L,Zhou T,Li X,Ahmed S H,Jiang D Z. 2021. Multimodality sentiment analysis in social internet of things based on hierarchical attentions and CSAT-TCN with MBM network[J]. IEEE Internet Things J,8(16):12748–12757. doi: 10.1109/JIOT.2020.3015381

    Yin L F,Xie J X. 2021. Multi-temporal-spatial-scale temporal convolution network for short-term load forecasting of power systems[J]. Appl Energy,283:116328. doi: 10.1016/j.apenergy.2020.116328

    Yu F, Koltun V. 2016. Multi-scale context aggregation by dilated convolutions[Z]. arXiv preprint arXiv: 1511.07122.

    Zhang N,Shen S L,Zhou A N,Jin Y F. 2021a. Application of LSTM approach for modelling stress-strain behaviour of soil[J]. Appl Soft Comput,100:106959. doi: 10.1016/j.asoc.2020.106959

    Zhang R Y,Chen Z,Chen S,Zheng J W,Büyüköztürk O,Sun H. 2019. Deep long short-term memory networks for nonlinear structural seismic response prediction[J]. Comput Struct,220:55–68. doi: 10.1016/j.compstruc.2019.05.006

    Zhang T,He L,Li X D,Feng G Q. 2021b. Efficient end-to-end sentence-level lipreading with temporal convolutional networks[J]. Appl Sci,11(15):6975. doi: 10.3390/app11156975

图(8)  /  表(3)
计量
  • 文章访问数:  412
  • HTML全文浏览量:  97
  • PDF下载量:  116
  • 被引次数: 0
出版历程
  • 收稿日期:  2022-12-18
  • 修回日期:  2023-04-25
  • 网络出版日期:  2023-07-02
  • 刊出日期:  2024-09-14

目录

/

返回文章
返回