【作者】 宋志宇；

【导师】 李俊杰；

【作者基本信息】 大连理工大学 ， 防灾减灾工程与防护工程， 2007， 博士

【摘要】 以大坝的安全监测为背景，根据智能计算在各学科领域内的广泛应用和大坝安全监测分析方法的发展特点，进行了安全监测中正反分析智能化分析的新方法以及安全监测点位置优化选择的研究。主要内容如下：引入基于统计学习理论的机器学习方法—支持向量机(SVM)到大坝安全监测领域，建立了大坝安全监测的SVM模型，经对实际数据的建模分析并和以往方法的比较表明，支持向量机在预测精度、泛化性能等方面明显好于以往方法。针对经典SVM在处理大规模数据问题时效率慢的不足，引入最小二乘支持向量机(LSSVM)，大大改善了其学习效率，同时结合和声搜索(HS)算法，建立了和声搜索最小二乘支持向量机(HSLSSVM)的大坝安全监测自适应建模方法，实测数据分析表明该模型同时具有建模简单、自适应、学习效率高、预测精度高和泛化能力强等特点。应用大坝材料参数反演的两种新方法：微粒群算法(PSO)和人工鱼群算法(AFSA)并对其进行改进。对PSO重新构建了惯性权重的非线性衰减函数，能够更好地协调算法的全局和局部收敛能力；同时还将模拟退火算法(SA)引入到微粒群算法中，增强了算法跳出局部极小的能力，提高了搜索效率。对AFSA，首先应用具有完全混沌特性的Logistic映射系统构造了AFSA的随机数发生系统，同时构建了感知距离的动态调整策略，不仅使算法能较快地到达全局极值附近，同时还加强了算法后期的局部搜索能力。计算表明，改进后建立的ISAPSO和ICAFSA反演算法是两个整体性能较优的新方法，丰富了大坝材料参数反演的智能化方法集。分析了大坝效应量、参数灵敏度和监测点位置优选的关系，提出了大坝安全监测中测点位置优选度的概念并建立了测点位置优选度定量计算的方法，以此可以指导在建大坝安全监测点的布设和已建大坝进行正反分析时监测资料的选取。然后通过对神经网络和支持向量机结构的分析和公式推导，得到网络输出对网络输入的灵敏度计算的一般公式。在使用正交实验定性地分析了对大坝应力应变起主要作用的控制性参数之后，分别通过重力坝和土石坝算例，应用有限元-神经网络方法计算了大坝控制性参数的灵敏度在坝体内的分布，然后进行了大坝位移监测点的位置优选度计算，最后通过算例对本文计算方法的正确性进行了验证。更多还原

【Abstract】 Based on the development character of dam safety monitoring and the wide applicationof intelligence computation, with the theories and methods of computational intelligence, thenew modeling methods of forward and inverse analysis and the computation of optimalselection degree of measuring points’ location were investigated in this paper. The majorcontributions are summarized as follows:A novel machine learning method called support vector machine (SVM) based onstatistical learning theory was introduced into the field of dam safety monitoring, and theSVM model of dam safety monitoring was established to forecast the dam deformation.Through the comparison with traditional statistical regression model and neural networkmodel, results indicate that SVM-based model possesses not only higher precision offorecasting, but also better generalization ability in long-interval forecasting. In view of thelow training speed of the standard SVM algorithm which is hard to solve large scale dataproblem, a least squares support vector machine (LSSVM) was put forward for improving thedisadvantage of standard SVM algorithm. The training speed and the ability to solve largescale data problem were much improved by using LSSVM. Appropriate parameters were verycrucial to learning and generalization ability of LSSVM, so we introduced a novel heuristicand stochastic global optimization algorithm-Harmony Search (HS) algorithm to realize theadaptive selection of LSSVM parameters, and brought forward a novel fusion monitoringmodel-harmony search least squares support vector machine (HSLSSVM). Simulation showsthat HSLSSVM-based monitoring model of dam safety has the advantages of simplemodeling technique, adaptive technique, higher learning speed, higher forecasting precision,better generalization ability etc al.Two novel intelligence inversion algorithms of dam parameters were modified and used.They were particle swarm optimization (PSO) and artificial fish swarm algorithm (AFSA).For PSO, a new non-linear attenuation strategy of inertia weight was established to effectivelyharmonize the ability of global convergence and local convergence; simultaneity simulatedannealing algorithm was introduced to PSO, so that the ability of getting away from localminimum was reinforced and the search efficiency of algorithm was improved. To theimproving of AFSA, we firstly applied the Logistic mapping system provided withcompletely chaotic characteristic to replace the random number system of AFSA, and then anew strategy of dynamically adjusting visual was reconstructed, so the improved CAFSApossessed not only quicker speed of getting to global optimum, but also the stronger ability ofsearching local optimum. Results of computation example show that ISAPSO and ICAFSA have the predominant integer performance for dam parameters’ inversion.On the basis of deep analysis of the relationship among effect-variable, parameterssensitivity and optimal layout of measuring points, the concept and computing way of theoptimal selection degree of measuring points’ location was built, based on the way, we couldlay measuring points in the projects of constructing dams and selection measuring data toforward and inverse analysis for the projects of existing dams. Through the structure analysesand formula deductions of neural network and support vector machine, the normal formulaeof parameters sensitivity were obtained by network’s output variables derivative about inputvariables. After the key parameters which control the strain and stress of dam were foundthough orthogonal test method, the sensitivity distribution of key parameters in gravity damand embankment dam was obtained by using FEM-NN method, then the optimal selectiondegree of measuring points’ location in dam was computed based on the proposedcomputation method. Finally, the result indicates that the proposed method of optimalselection degree of measuring points’ location in dam is reasonable and practicable.更多还原