【作者】 陈鑫；

【导师】 杨笑梅；

【作者基本信息】 广东工业大学 ， 土木工程， 2023， 硕士

【摘要】 国内外的多次震害调查已表明,由于特殊的场地条件影响,地震动在盆地区域内会出现异常放大现象。沉积盆地的表层软土会过滤高频分量,放大长周期地震动增加持时,盆地特殊的基底几何会使地震波在盆地边缘及中央区域产生显著放大。局部场地地表放大特征的预测一直是地震工程方向研究的关键问题,通过对盆地放大特征的合理预测,可为建筑结构的抗震设防标准的制定、震害评估及地震小区划等工作提供重要的依据。强震观测受场地条件限制,依赖大范围台网数据建立的地面运动预测方程,将复杂物理过程理想化,但对于局部区域刻画不够细致,无法反映局部场地的影响。确定场地的数值模拟结果能体现详细的内在物理联系,参数选取灵活,但由于地震机理的复杂性,存在参数不确定性的影响及计算过程复杂的困难。本文基于人工智能的深度学习方法和考虑盆地局部特征的二维标准模拟数据,对局部场地条件影响下的盆地放大效应预测方法进行初步探索,通过深度学习中的卷积神经网络(CNN)建立盆地地表放大特征预测模型,并对比分析所建立的不同CNN预测模型的差异,讨论以模拟结果作为数据基础的盆地放大特征预测模型的合理性,同时分析预测结果的精度及适用性。本文的主要研究内容及结论如下:1、通过ABAQUS/Python二次开发进行二维盆地的参数化建模,批量自动生成模型的INP文件。对盆地进行场地响应模拟时遇到的问题,如土体本构、介质阻尼、网格划分、人工边界和地震动输入等问题进行了详细介绍,并给出了在ABAQUS软件中实现的具体方法。通过有关文献算例对比验证本文采用的显式有限元数值模拟方法的可靠性。2、利用“天河二号”的强大算力,批量计算盆地有限元模型的地面运动响应。提取盆地放大特征构建CNN预测模型所需的大样本数据集。对采集的数据研究了盆地形状比、波速比和盆地倾角等因素对放大特征的影响,为CNN模型输入参数的确定提供了理论依据。3、基于单层土盆地模拟数据建立多个CNN预测模型,对其结果分别进行对比分析,包括专项放大系数峰值预测模型CNN-AF与传统BP神经网络的对比,8参数的联合预测模型CNN-AP-8和9参数联合预测模型CNN-AP-9的对比,联合预测模型与专项预测模型效果的对比,结果显示CNN-AP-8的效果最佳。最后,选择一真实盆地利用CNN-AP-8进行放大特征预测,结果表明基于单层土盆地训练的CNN预测模型会低估盆地的放大系数峰值,但能较为准确的预测危险位置。4、基于多层土盆地的模拟数据建立了CNN-APM预测模型,对比真实盆地与预测模型的结果,发现在弥补了CNN-AP-8放大系数低估问题的同时会错判危险位置。因此,基于简化模型的数值模拟结果,分析预测模型结果与真实盆地存在差异的原因。针对表层薄土可能引起的误差,扩增样本模型并新建改进模型CNN-APMP,结果表明改进模型进一步降低了放大系数峰值的相对预测误差,又更正了危险位置的错判。更多还原

【Abstract】 Multiple investigations of earthquake damage at home and abroad have shown that there is an abnormal amplification phenomenon of seismic motion in the basin area due to the influence of special site conditions.The shallow soft soil in sedimentary basins filters high-frequency components,amplifies long-period seismic motions,and increases their duration.The unique basement geometry of the basin causes significant amplification of seismic waves at the basin edges and central regions.Predicting local site surface amplification characteristics has always been a key issue in the research of earthquake engineering.A reasonable prediction of basin amplification characteristics can provide an important basis for the formulation of seismic design standards for building structures,seismic hazard assessment,and earthquake zoning.Strong-motion observation is limited by site conditions,and ground-motion prediction equations established based on large-scale network data simplify complex physical processes.However,they are not detailed enough for the characterization of local areas and cannot reflect the impact of local sites.The numerical simulation results of site determination can reflect the detailed internal physical relationships,with flexible parameter selection.However,due to the complexity of the earthquake mechanism,there are influences from parameter uncertainty and the complexity of the calculation process.This thesis makes a preliminary exploration on the prediction method of basin amplification effect under the influence of local site conditions,based on deep learning methods of artificial intelligence and two-dimensional standard simulation data considering local features of basins.It compares and analyzes the differences of different CNN prediction models,and discusses the rationality of using neural networks base on simulation results as data basis to predict basin amplification characteristics,and the accuracy and applicability of the prediction results are analyzed.The main contents and conclusions are as follows:1.ABAQUS/Python secondary development is used for parametric modeling of two-dimensional basins,and INP files of the models are automatically generated in batches.Some problems about the site response simulation of basins,such as the soil constitutive models,the material damping,the mesh generation,the artificial boundaries and the ground motion input,are introduced in detail,and the specific methods to implement in ABAQUS software are given.The reliability of the explicit finite element numerical simulation method used in this article is verified by comparing with relevant literature examples.2.The powerful computing power of"Tianhe-2"is used to calculate the ground motion response in the finite element model of the basins in batches.Basin amplification features are extracted to construct large sample data set required by CNN prediction models.The effects of factors such as basin shape ratio,velocity ratio and basin inclination angle on the amplification characteristics are studied base on the collected data,which provides a theoretical basis for determining the input parameters of the CNN models.3.Multiple CNN prediction models are established based on simulation data of single-layer soil basins.The results are compared and analyzed,including the comparison between the specialized amplification factor peak prediction CNN-AF model and the traditional BP neural network,the comparison between the joint prediction models CNN-AP-8 with 8parameters and CNN-AP-9 with 9 parameters,and the comparison between the joint prediction models and the specialized prediction models.The results are showed the CNN-AP-8 model has the best effect.Finally,a real basin is selected to use CNN-AP-8 for amplification feature prediction.The results are showed that the CNN prediction model based on the training of single-layer soil basin will underestimate the peak value of the amplification coefficient of the basin,but can predict the dangerous position more accurately.4.The CNN-APM prediction model is established based on simulation data of multi-layer soil basins,and the results of the real basin are compared with the predicted model.It is found that dangerous position would be misjudged while CNN-AP-8 magnification coefficient was underestimated.Based on the numerical simulation results of the simplified model,the reasons for the differences between the predicted model results and the real basin are analyzed.A new improved model CNN-APMP is established to analyze the causes of errors.The results are showed that the improved model further reduced the relative prediction error of amplification coefficient peak and corrected the misjudgment of dangerous position.更多还原