【作者】 郭国会；

【导师】 易伟建；

【作者基本信息】 湖南大学 ， 结构工程， 2001， 博士

【摘要】 在众多对结构进行损伤诊断的方法中，大多以简单结构为例进行分析，而且一般不考 虑误差的影响，这都与实际工程结构的损伤诊断有一定距离。本文针对实际的工程结构， 完成了以下的研究工作： l、以高墩粱式桥为研究对象，讨论基子灵敏度分析进行损伤诊断方法的优缺点。为 克服灵敏度分析法不能对严重损伤进行识别的缺陷，本文提出用迭代法成功解决了严重 损伤的识别问题。当结构中同时损伤的单元数目较多，以及结构单元数目远大于实测的 模态阶数时，灵敏度分析法的识别结果严重下降，本文提出损伤识别的两步法，将结构 的损伤分为定位和定量两个步骤，有效地改进了识别效果。为克服灵敏度分析法一阶近 似的本质缺陷，基于正交条件灵敏度分析法，利用MSECR对结构的损伤进行粗略定位 后，给出了准确的识别结果，但这种方法的缺点是需要完备的实测振型。 2、本文详细分析了模型误差对损伤识别结果的影响，经过公式推导及实例分析说 明，损伤灵敏度矩阵用有模型误差的理论分析模型来建立，如果损伤前后都用无模型误 差的实测数据来识别，此时误差为二阶微量可以忽略；如果损伤前采用有模型误差的数 据而损伤后采用无模型误差的实测数据，此时误差为一阶量，可能引起较大的识别误差。 3、将测量误差的影响考虑为正态分布的高斯白噪声，通过理论分析和数值模拟， 说明用灵敏度分析法识别结构损伤时，损伤识别结果也服从高斯正态分布，其平均值保 持不变，而方差则与误差水平成线性关系。 4、本文以结构的低阶频率和少数节点的一阶振型分量，经预处理后作为神经网络 输入，准确实现了对一钢析粱桥的损伤位置和损伤程度识别。在选取少数节点的振型分 量用于网络训练时，提出根据单元应变能系数的大小来选择节点，这些节点一般来讲在 振动中振幅较大，易于实测和用于网络训练。误差分析的结论同样在基于神经网络对结 构进行损伤诊断的方法中得以体现。 5、以北川河钢管混凝土拱桥为对象，研究采用神经网络方法实现大型结构的损伤 监测。对于拱肋的损伤，本文采用拱肋单元在结构损伤前后一阶应变模态的相对改变量 来实现。构造了一个由四个子网组成的组合网络，每个子网分别针对拱肋中的部分单元 进行训练，然后由各子网的输出共同给出整个拱肋的损伤状态。本文还对吊杆损伤对拱 肋应变模态的影响进行了初步分析，认为少量吊杆的损伤不会对拱肋的应变模态产生很 大的影响，即在少数吊杆存在损伤时，网络仍能对拱肋的损伤进行识别。 6、对于吊杆的损伤监测，本文根据桥面板节点在吊杆损伤前后一阶振型的变化， 采用 Kohonen网络来实现吊杆的损伤位置识别。以 10％、90％的吊杆损伤对应的数据验 证网络，准确地识别了吊杆的损伤位置。经过对比说明，不同位置拱肋单元的损伤，对 吊杆损伤位置识别的影响不同。更多还原

【Abstract】 Among many methods that deal with structural damage assessment, most of them focus on simple structures and don抰 consider any errors, so there is a certain distance before they can be used to assess the damage of real structures. This dissertation carries on research of real engineering structures. The main contents include:1 Take high pile and beam bridge for example, this paper discusses the merits and defects of the damage assessment method based on sensitivity analysis. In order to overcome the defect that this method can抰 identify serious damage, this paper presents iteration method to solve this problem. When damaged elements increase at the same time and the number of element is much more than that of measured modes, the sensitivity analysis method can抰 give desired results. Therefore, a method which can identify damage though two steps, Location and Extent, is put forward. After probable damage location through MSECR, the method based on the sensitivity of orthogonality conditions gives good predictions of damage. This method overcomes the natural defect of the concept first order approximation of the ordinary sensitivity analysis method. But this method needs complete measured modes.2.The influence of model error on damage assessment results is analyzed. We get the sensitivity matrix using the model that contains model error in the following two cases. If the identification data are got using the model without model error before and after the damage, the error resulted from model error can be overlooked. If the data are got using model with model error before the damage and using the model without model error after the damage, error results from model error is first order and model error can bring more assessment errors. All above is proved by equations and examples.3..The effect of measurement error is considered as Gauss white noise. Through theoretical analysis and numerical simulation, damage assessment results are also subject to Gauss distribution. Furthermore, results indicate that the mean keeps intact, but the standard deviation is linear to error level.4.Using processed frequencies and first order modal DOFs in a few points as network inputs, trained neural network gives excellent assessment results of a steel truss bridge. A method is put forward which selects modal DOFs in a few points according to element strain energy coefficient. Therefore, the selected modal DOFs in a few points usually have large amplitude, which is conducive to be measured and the network training. The effect of model error on assessment results is also reflected in the ANN method.5.This paper uses ANN to realize the damage monitoring of a large structure, a concrete-filled steel tubular bridge of Beichuan River. For the damage of the main arch-ring elements, this paper uses the relative change of first order strain mode before and after thedamage. A combined network which consists of four subnets is put forward. Every subnet is trained using the data of one portion of arch-ring elements, and the outputs from all subnets give the whole damage information of the arch-ring. The influence of hanger damage on the archring strain mode is also considered and get the conclusion that there is no much change in strain mode when the damaged hangers is small. Therefore, network can still gives good results when some hangers are damaged lightly.6..For the damage of the hanger elements, the change of first order mode of deck nodes before and after the damage is used. The Kohonen network was put forward to realize the location of the damaged hangers. When the hangers are damaged as seriously as 10%. 90%, the network still gives good predictions. At last, comparisons indicate that the damage of arch-ring elements with different location have different influence on the location of the damaged hangers.更多还原