基于贝叶斯网络的地震液化风险分析模型研究

发布时间：2018-02-11 02:38

  本文关键词： 贝叶斯网络 地震液化 性能指标 概率预测 灾害评估 决策分析　出处：《大连理工大学》2016年博士论文　论文类型：学位论文

【摘要】：地震液化问题的深入研究对于抗震减灾工程而言是极其重要的。目前已有的研究成果绝大部分是关于地震液化发生的预测,而且已有的液化预测方法存在适用性窄,预测精度不高等缺陷,另外,对于液化后的灾害综合预测研究鲜有,特别是对于液化的减灾决策问题的研究。本文基于贝叶斯网络方法,通过统计计量手段筛选地震液化的重要影响因素,建立了地震液化的贝叶斯网络预测模型,随后引入液化灾害变量和抗液化措施及损失,分别建立了地震液化的贝叶斯网络液化灾害风险评估模型和液化减灾决策模型,扩展的模型不仅可以对地震液化的发生及液化后的灾害做出快速、准确的预测和评估,而且可以针对液化灾害程度及场地性质,分析得出最佳减灾决策方案,为工程抗震减灾提供科学依据。本文的主要研究有如下几个方面：(1)基于统计计量手段和地震液化重要因素筛选原则,从众多地震液化的影响因素中筛选出了12个重要因素：震中距、地震震级、峰值加速度、地震持续时间、土质类别、细粒或粘粒含量、颗粒级配、相对密度、可液化层厚度、可液化层埋深、上覆有效应力和地下水位,并采用解释结构模型方法对这12个重要影响因素建立了其层次结构网络,分析了各重要影响因素间的层次强弱关系,为建立地震液化的贝叶斯网络模型提供了基础准备。(2)选择地震等级、震中距、标准贯入锤击数、地下水位和砂土层埋深5个因素,基于解释结构模型法和K2算法分别建立了仅适用于自由场地的地震砂土液化判别的五因素主观贝叶斯网络模型、客观贝叶斯网络模型和混合贝叶斯网络模型,选用总体精度(OA)、准确率(Pre)、召回率(Rec)、F1值和ROC曲线下面积(AUC)五个性能指标,和其他常用的地震液化判别方法对比,验证了贝叶斯网络模型的有效性和准确性,其中地震砂土液化的混合贝叶斯网络模型的预测性能最好。随后探讨了训练样本的分类不均衡和抽样偏差对地震液化概率预测模型的影响,发现训练样本分类越不均衡、抽样偏差越大,模型的回判性能越好,而预测性能却越差,并给出了常用概率预测方法的最佳训练样本分类比例范围,以减小训练样本分类比例选择所带来的模型误差。另外,针对训练样本存在严重分类不均衡或抽样偏差的情况,探讨了过采样技术对在常用概率模型性能提升中的应用。(3)在大量数据不完备和数据完备两种情况下,基于筛选的12个重要影响因素,分别提出了基于解释结构模型和因果图法建立地震液化多因素主观贝叶斯网络预测模型的方法及融合解释结构模型和K2算法建立地震液化多因素混合贝叶斯网络预测模型的方法。采用K(K=5)折交叉试验,选用五个性能评估指标,与其他液化概率预测模型的性能进行综合对比,验证了两种方法所建的贝叶斯网路液化预测模型的准确性和鲁棒性,并对液化的12个影响因素进行了敏感性分析和反演推理。所建立的新贝叶斯网络模型扩展并改善了前面五因素模型的适用范围和预测精度,使其能适用于不同土质类别、不同细粒含量的自由场地液化和有上部结构物场地液化的更高精度预测。(4)在贝叶斯网路液化预测模型基础上,引入地震液化的灾害指标,如液化潜能指数、喷砂冒水、地面裂缝、地表沉降和侧向位移,建立了贝叶斯网路液化灾害风险评估模型,使其可以综合评估液化后场地的灾害程度。与神经网络方法对比,验证了模型的精准性和可靠性。随后,引入抗液化措施和其效用及成本,扩展了该模型,使其不仅可以预测液化及液化灾害,而且可以为液化减灾提供最佳决策支持。将地震液化减灾的决策模型应用到人工岛的抗液化决策中,采用数值模拟结果验证了模型的有效性,使其能为液化的防灾减灾提供科学依据。
[Abstract]:Study of earthquake liquefaction problems is very important for earthquake disaster mitigation projects. The existing research results mostly about prediction of liquefaction and liquefaction, prediction method has narrow applicability, the prediction accuracy is not high defect, in addition, for comprehensive prediction of post liquefaction disaster research rare, especially on the disaster reduction decision problem. In this paper the liquefaction method based on Bayesian network, by means of statistical measurement of seismic liquefaction of the important influence factors of screening, established a prediction model of Bayesian network for earthquake liquefaction, followed by liquefaction hazard variables and anti liquefaction measures and losses, the earthquake liquefaction disaster risk assessment model and Bayesian network liquefaction mitigation decision model of liquefaction established, extended model can not only make the earthquake liquefaction and post liquefaction disaster, accurately The prediction and assessment, and according to the nature of liquefaction hazard degree and site, analyze the best mitigation decision scheme, and provide scientific basis for earthquake disaster mitigation engineering. The main research of this paper are as follows: (1) statistical measurement method and seismic liquefaction is an important factor for screening based on the principle of selection from the many factors that influence the seismic liquefaction the 12 important factors: the earthquake magnitude, epicentral distance, peak acceleration, duration of earthquake, soil, or fine clay content, particle size, relative density, liquefiable layer thickness, liquefaction depth, effective overburden stress and underground water level, and the factors that influence the establishment of 12 important the hierarchical structure of network by using interpretative structural modeling method, analyses the factors influence the strength of the relationship between, for the establishment of a Bayesian network model of seismic liquefaction provides the basis for the choice (2). The earthquake epicenter, grade, SPT, underground water and sand layer depth of 5 factors, five factors of seismic liquefaction subjective Bias network model is only suitable for the free field established interpretive structure model and K2 algorithm respectively based on objective Bias network model and Bias mixed network model selection the overall accuracy (OA), accuracy (Pre), the recall rate (Rec), F1 value and the area under the ROC curve (AUC) five performance indexes, and other commonly used seismic liquefaction methods to verify the effectiveness of Bias network model and the accuracy of the prediction performance of hybrid Bias network model of sand liquefaction the best. Then it discusses the classification of training samples is not balanced and the sampling bias prediction model of influence on seismic liquefaction probability, we find that the training sample classification is more uneven, the greater the sampling bias, discriminant performance model The better, and the prediction performance is worse, and the best training sample classification ratio range prediction method commonly used to model the probability of error brought by reducing training samples classification scale selection. In addition, the training samples are seriously unbalanced classification or sampling bias conditions, discusses the application of over sampling technique to improve performance the commonly used probability model. (3) in a large number of incomplete data and complete data for two cases, 12 important factors are put forward based on screening, interpretative structural model and causal graph method to establish seismic liquefaction multi factor forecasting method and subjective Bayesian network fusion prediction model to explain the structure model and K2 based seismic liquefaction hybrid Bayesian network based on multi factor. Using K (K=5) fold cross test, using five performance evaluation index, and other prediction model of probability of liquefaction Can carry out comprehensive comparison, Bias network prediction model to verify the liquefaction of two methods the accuracy and robustness, and the 12 factors influencing liquefaction sensitivity analysis was carried out and the inverse reasoning. The new Bias network model to expand and improve the application range and the first five factors model prediction accuracy, the suitable for different categories of soil, the effect of fines content on the liquefaction free and have higher accuracy of site liquefaction prediction of upper structure. (4) based on the prediction model of liquefied Bias network, introducing the disaster index of seismic liquefaction, such as liquefaction potential index, sand at the water, ground crack, surface subsidence and lateral displacement. The risk of liquefaction hazard assessment model of network Bias, which can make the comprehensive evaluation after liquefaction site degree of the disaster. Compared with the neural network method to validate the model accuracy and Depending on the nature. Then, introducing the anti liquefaction measures and its utility and cost, the model is extended, so that it can not only predict the liquefaction and disaster, but also can provide the best decision support for disaster reduction. The application of liquefaction decision model of earthquake liquefaction mitigation to liquefaction decision artificial island, the numerical simulation results verify the model, which can provide scientific basis for disaster prevention and mitigation of liquefaction.

【学位授予单位】：大连理工大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TU435
，

本文编号：1502021