【作者】 康飞；

【导师】 李俊杰；

【作者基本信息】 大连理工大学 ， 防灾减灾工程及防护工程， 2009， 博士

【摘要】 随着经济的高速发展,我国兴建了大批的重大土木工程项目,这些重大工程项目的使用期较长,影响力较大,一旦失事,会造成严重的生命财产损失。因此为了保障结构的安全性、完整性、适用性和耐久性,已经建成的许多重大工程结构和基础设施急需采用有效的手段检测和评定其安全状况、修复和控制损伤。许多新建的大型结构和基础设施,如大坝、桥梁、海洋平台等,增设了长期的安全/健康监测系统,以监测结构的服役安全状况,并为研究结构服役期间的损伤演化规律提供有效的、直接的方法。监测系统中数据采集与传感的一个基本假设是这些系统不是直接测量结构异常,而是测量系统在它的运作或环境载荷下的响应,或者是对嵌入传感系统中作动器输入的响应。传感器的读数或多或少的与结构异常的存在及其位置相关。数据处理程序对于结构健康监测系统来说是必须的,它们将传感器采集到的数据转化为结构状况的信息。计算智能是大坝等结构安全监测建立预报模型和进行反演分析的有力工具,已经取得了一些成果,但仍存在一些不足。计算智能目前仍处于快速发展阶段,将几种新型的计算智能方法引入大坝等结构的安全监测预报建模与反演分析领域,开展了一些有意义的工作。差分进化算法、微粒群优化算法和人工蜂群算法是几种具有较大发展潜力的新型智能优化方法,和传统的遗传算法相比具有实现简单、收敛性能好等优点。差分进化算法和微粒群优化算法在处理多维优化问题时具有较好的收敛性能,将他们用于损伤识别问题,并将几种人工免疫特性引入微粒群算法,提出了一种结构损伤识别的免疫加强微粒群算法。对人工蜂群算法进行了改进,针对其由搜索模式的单一性导致的参数较少时的“趋同”问题,将单纯形算子引入算法中,提出一种混合单纯形人工蜂群算法,改进算法不仅收敛速度明显加快,且由于搜索方式的增多,也很少陷入停滞现象。静动态反演分析算例表明,所提出的算法是高效的优化反演方法,为大坝参数的识别,进而进行结构响应预报建模与损伤评估提供了新的途径。径向基网络与BP网络相比,不仅具有生物学基础和数学基础,而且结构简单,学习速度快,隐节点具有局部特性,逼近能力更强。提出了一种处理复杂反演分析问题的蚁群聚类径向基网络模型。该模型避免了智能优化反演方法需要循环迭代,计算效率不高的问题;以及传统神经网络模型训练时间长、易陷入局部最优以及反演精度不高的问题。它可以直接用于三维土石坝双屈服面模型参数反演这样计算量巨大的大型非线性多参数反演问题。采用蚁群聚类选择径向基函数中心,克服了传统K-means聚类易陷入局部最优,和对初始聚类中心依赖强的缺点,能够获得更合理的聚类中心,得到满意的径向基网络模型。支持向量机是数据挖掘中的一项新技术,是借助于最优化方法解决机器学习问题的新工具。模型参数选择是采用支持向量机进行建模的关键影响因素,采用三种方法进行模型参数选择,分别是:基于网格平行搜索的交互验证法、遗传算法和粒子群算法。将所建立的模型用于金竹山电厂贮灰坝渗流测压管预测,表明支持向量机模型预测精度高,泛化能力强,是一种高效的系统建模方法。优化的传感器网络结构可以最小化所需要的传感器数量,节约投资,同时能够提高精度并提供一个鲁棒性的系统。在研究大坝安全监测中静动态传感器优化配置模型和准则的基础上,将单亲遗传算法用于求解该问题。传统遗传算法在求解组合优化问题时,交叉操作可能产生不可行解,需要借助一些复杂的操作算子,不仅效率不高且缺乏理论基础;单亲遗传算法遗传操作在一条染色体上进行,避免了该问题。同时为了进一步提高单亲遗传算法的性能,提出了两种改进算法,即自适应模拟退火单亲遗传算法和病毒协同进化单亲遗传算法。通过算例验证了所提出模型和算法的有效性。更多还原

【Abstract】 With the development of economy,a great deal large civil engineering projects,which are long used with great impact,are constructed in China.The wreckage of these important projects will cause serious loss of property and life.Therefore,many already constructed large engineering projects and infrastructures need effective measures to detect and evaluate their condition of safety,and then the damage can be repaired and controlled.Many new constructed structures,such as dam,bridge and ocean platform,are equipped with healthy monitoring system,which are applied to monitor the structural safety condition and provide direct effective methods to study damage evolve process during enlistment.A basic presume in health monitoring system is that these systems not measure structural abnormity directly,rather measure the responses under the operational and environmental loads or the responses to input of actors embedded in the sense system.The data collected by sensors are more or less correlated to the existence and position of structural abnormity.Data processing procedures are integrant to a health monitoring system;they translate the data collected by sensors into information of structural condition.Computational intelligence is a powerful tool for safety prediction modeling and inverse analysis,and many achievements have been obtained in this domain,but there are still some drawbacks in the traditional computational intelligence.Nowadays,computational intelligence is still in the rapid developing stage,several novel computational intelligent methods are introduced into the domain of safety prediction and inverse analysis of structures like dams,and some meaningful work is developed.Artificial bee colony algorithms(ABC),particle swarm optimization(PSO) and differential evolution(DE) are three novel intelligent optimization algorithms with tremendous developing potential.Compared to traditional genetic algorithms,the advantages of these optimization algorithms are easier to implementation and better convergence performance.Aim at the problem of "incline to be the same" phenomena when used to few parameters optimization caused by single search pattern,ABC are improved by introduce cultural frame,annealing operator and simplex operators,and cultural annealing ABC and hybrid simplex ABC are proposed.The convergence speed of the improved algorithms is accelerated,meanwhile the stagnation phenomena is reduced because of enriched search patterns.DE and PSO have good convergence performance whey they are applied to problems with many dimensions,and they are used to damage detection problems.Several immune properties are introduced into PSO,and an immunity enhanced PSO is proposed for damage detection problems.Static and dynamic inverse analysis indicates that the proposed algorithms are very efficient for inverse problem,so novel approaches are provided for structural parameter identification,which can be used to response prediction modeling and damage evaluation.Compare to BP neural networks,radial basis function neural networks(RBF) not only has biological foundation but also mathematical foundation,meanwhile it has simpler structure, faster training speed and higher accuracy of simulation which is because the hidden nodes have local tuned properties.An ant colony clustering RBF model for completed inverse analysis problems is proposed.The disadvantages of intelligent optimization based inverse need long time caused by repeating iteration and traditional neural network need long training time,easy trapped into local optimum and low inversion accuracy,are avoided in the new model.It can be applied to large-scale nonlinear inverse problems,such as inverse analysis of three dimensional rockfill dams.K-means clustering algorithms have the disadvantages of easily trapped into local optimal and dependence on initial clustering centers.However,ant colony clustering can avoid these disadvantages,more reasonable radial basis function centers and more satisfactory RBF model can be obtained.Support vector machines(SVM) are novel technology for data mine,and are novel tools for machine learning recurs to optimization methods.Several optimization algorithms are used to model parameter selection.Application example illustrated that SVM model have the advantage of high prediction accuracy,less over fitting and is an efficient prediction modeling method.Optimized sensor network configuration can minimize the sensor number needed and save investment,meanwhile can provide a robust system with high accuracy.Optimal sensor placement problem in health monitoring and testing is studied.A partheno-genetic algorithm (PGA) is used to solve this problem.The drawback of traditional genetic algorithm for this problem is avoided.In order to improve the performance of PGA,adaptive simulated annealing PGA and virus coevolution PGA are proposed.The efficiency of the proposed optimal sensor placement algorithms are illustrated by several numerical examples.更多还原