【作者】 邱颖；

【导师】 任青文；

【作者基本信息】 河海大学 ， 水工结构工程， 2005， 博士

【摘要】 利用结构的振动响应和系统动态特性参数进行结构损伤检测是目前国内外研究的热点和难点。虽然该方法已在实际工程损伤检测方面得到了较广泛的应用,但是随着损伤的积累以及老化结构的增加,为了更加准确的对损伤进行定位和定量,需要发展新的信号处理方法提取对损伤敏感的特征因子。本文在总结回顾以前工作的基础上,主要开展了如下4个方面的工作: (1) 当结构发生损伤时,在损伤处,其模态振型或其高阶导数不连续即存在奇异,这种奇异性可以通过寻找其小波变换模极大在细尺度下的收敛的横坐标来检测,奇异性位置即对应结构中的损伤位置;为了对损伤程度进行诊断,利用神经网络较强的非线性映射能力来模拟小波变换模极大和损伤程度之间的非线性关系,采用db2小波对模态振型进行离散小波变换,将多个尺度下小波变换模极大值进行归一处理后,作为BP网络的输入参数,输出为结构的损伤程度。通过数值算例验证了该方法简单易行、精度高,可以在实际检测中推广应用。 (2) 根据利用模态曲率进行损伤检测的原理,采用具有二阶消失矩的墨西哥草帽小波对结构在健康和损伤状态下的模态振型差进行连续小波变换,由此提出了表示损伤位置的指标。即相当于先用高斯函数对模态振型差进行平滑处理,然后对平滑后的模态振型差求解二阶导数,通过寻找小波变换模极大在细尺度下收敛的横坐标来检测损伤位置。根据Lipschitz指数与小波变换模极大的关系,讨论了损伤程度与Lipschitz指数的关系。数值模拟结果表明:该指标可以准确的对损伤进行定位。同时通过实验,利用该方法对实际结构进行了健康诊断。 (3) 以结构中存在两处损伤为例,研究了小波包节点能量用作结构多处损伤特征因子的可行性。并将某一尺度下特征因子进行归一处理后作为BP网络的输入参数,利用网络的分类识别能力,实现对结构损伤位置及程度的诊断。并在数值算例中讨论了噪声对该方法的影响,同时也讨论了网络的外推性能。 (4) 编制了小波神经网络程序,该网络兼具小波变换和神经网络两者的优点,为改善该网络的性能,使之能够最大限度地提取信号的特征,借助遗传算法来搜索最优网络参数,以提高小波神经网络的检测精度。更多还原

【Abstract】 The structural damage detection method based on structural vibration response and systemic dynamic parameters is a difficult and growing interest problem in nondestructive evaluation (NDE). Although the method have been widely applied in practical engineering. However along with damage accumulating and increase of aging structures, new signal analysis and interpretation algorithm need to be developed in order to accurately evaluate damage. After reviewing the past work, 4 aspects of researching is developed as following:(1) The mode shape or its certain derivative of a structure is discontinuous at the damage location, and the discontinuous property can be detected by the wavelet transform modulus maximum and the wavelet ridge. The discontinuous location is the damage location. The nonlinear relation between the modulus maximum and damage extent is simulated using an artificial neural network to detect the damage extent. The discrete wavelet transform is performed on the mode shape of a simple beam by db2 wavelet, and the modulus maximums of multi-scales are calculated and used as inputs into BP neural network for damage extent assessment. Numerical simulation results show the modulus maximums are sensitive to structural damage.(2) The continuous wavelet transform by a Mexican hat wavelet having two vanishing moments is applied to the difference in the first mode shapes of a structure under the health and damage state based on the theory of the mode curvature used in damage detection, and a damage index is presented. That is to say, the difference is first smoothed by Gauss function, then the second derivative of the smoothed difference is calculated, he damage location is detected by the horizontal ordinate convergence of the modulus maximum for low scales. The relation between Lipschitz exponent and damage extent is discussed based on the index maximum. Numerical simulation results show the index can exactly detect the location of the damage. Meanwhile, a correlative experiment is conducted.(3) Dynamic signals measured from a structure are decomposed into wavelet packet components. The feasibility of component energies taken as damage factor is studied when damage exits in two parts of a structure. The energies are calculated in an appropriate scale based on sensitivity study, and they are used as inputs into BP neural network for multi-location damage diagnosis. The effect of noises on the method is studied in the numerical simulations, and the extrapolation ability of the neural network is discussed.(4) The wavelet neural network(WNN) classifiers program is achieved using structural damage diagnosis, and the WNN has the virtues of the wavelet transform and the neural network. Genetic algorithm is used to improve the performance of the WNN.更多还原