【作者】 艾永明；

【导师】 黄平明；

【作者基本信息】 长安大学 ， 桥梁与隧道工程， 2012， 博士

【摘要】 由于年代的久远和环境的影响，许多工程结构呈现出老化现象，个别的构件出现损伤甚至断裂，结构在损伤积聚下一旦发生溃塌，会对国家和人民的生命财产造成极大的损失。因此，对工程结构进行损伤识别具有重大的学术、经济和社会意义。本文利用小波变换对时间序列的时间和频率定位能力、小波函数具有的空间域缩放属性，以及神经网络具有的捕捉非确定性性质和复杂非线性的能力，结合模糊聚类技术，提出了结构非参数系统识别的动态时延模糊WNN模型，然后基于伪谱方法，提出了环境激励下大型复杂结构的损伤识别方法，并结合理论分析、数值模拟及试验方法，对结构损伤识别进行了系统和深入的研究。本文主要研究内容如下：1）基于非正交小波——Mexican hat小波，构造了动态系统函数逼近的动态时延WNN模型，该模型可以任意逼近瞬时非线性函数，并利用动态时延WNN模型在小波设计中的灵活性，添加了额外功能，如函数逼近中理想的平移参数调整。2）基于混沌理论的重构状态空间概念构造动态时延WNN的输入向量，该向量保留了时间序列数据的动态特性，可以识别结构系统的非物理参数（如结构的动态响应）。然后利用FNN方法确定动态时延WNN输入向量的最佳嵌入维数，以及修正Gram-Schmidt算法和AFPE准则确定WNN隐层节点的最佳数量，减小了WNN模型的规模，提高了模型的计算效率和识别精度。3）将具有空间域缩放属性的小波函数应用于动态时延WNN模型，它仅在输入时间序列的有限范围影响模型输出，这种属性减少了WNN节点之间的不良影响，提高了函数逼近的精确度，加快了WNN的训练收敛过程。4）基于NARMAX方法，将重构状态空间向量的模糊聚类与非正交WNN相结合，构造了结构非参数系统识别的动态时延模糊WNN模型。该模型可有效而精确地捕捉时间序列传感器数据的动态特征，并可克服传统WNN模型在处理存在局部不准确的训练数据时，由于相同的空间域缩放属性易导致的较大局部输出误差问题。5）采用自适应LM-LS混合学习算法训练动态时延模糊WNN模型，训练过程为两步迭代法，首先采用LS算法确定模型的线性参数，然后采用LM算法调整模型的非线性参数。该算法避免了Gauss-Newton算法的二阶微分问题，克服了最速下降算法的数值不稳定问题，并可显著提高WNN模型的训练收敛速度。6）小波在动态时延模糊WNN模型中应用于两个方面：○1采用离散小波包变换（DWPT）技术对传感器数据去噪，加快了结构系统识别模型的训练收敛速度，并明显提高了模型的识别精度。○2将小波函数作为神经网络的激活函数，结合模糊聚类技术构造模糊WNN模型，该模型可以有效而精确地捕捉时间序列传感器数据的动态特征。7）基于结构非参数系统识别的动态时延模糊WNN模型、以及伪谱方法，提出了环境激励下大型复杂结构的损伤识别方法。该方法的识别结果为选定子结构内的结构局部损伤，通过选定一系列的子结构可识别出结构的所有损伤，具有结构全局和局部损伤识别能力；损伤指标以不同频段的形式感知损伤，包含了丰富的损伤信息；不需对损伤指标凭经验设定阀值判断结构是否发生损伤，只要损伤指标大于零，就可确定结构发生了损伤，且具有很高的损伤灵敏度；外部动态荷载只是用来对结构进行充分激励，而不需对荷载谱进行量测，因此可对环境激励下的工程结构进行损伤识别，具有较高的工程适应性。更多还原

【Abstract】 Long-term use and environmental factor will cause various defects (damages) in civilstructures, such as bridges. Damage accumulated in civil structures could lead to catastrophiccollapse of the structure, causing huge loss to the society. Thus, structural damageidentification of civil structures is of great academic, economic and social significance. In thispaper, the positioning capabilities on the time and frequency of the time series of the wavelettransform, the scaling properties in the spatial domain of the wavelet function, and capture thenon-deterministic nature and complex non-linear capabilities of the neural network, combinedwith the fuzzy clustering technique, the dynamic time delay fuzzy WNN model of thestructure non-parametric system identification is proposed, then based on the pseudo-spectralmethod, the damage identification method for large complex structures under ambientexcitation is proposed. Finally, based on a combined theoretical analysis, numericalsimulation and physical model experiment approach, the in-depth study on Structural dynamicnondestructive detection methods is conducted. In sum, the main contents are concluded asfollows.1) Based on the non-orthogonal wavelet-Mexican hat wavelet, the dynamic time delayWNN model for the dynamic system function approximation is developed, the model canapproximate the instantaneous nonlinear function at will. The flexibility of dynamic timedelay WNN model in the wavelet design is empolyed to add the additional functionality, suchas the ideal transform parameter adjustment in the function approximation.2) On the basis of the reconstruction state space concept of chaos theory, the input vectorof the dynamic time delay WNN is constructed, it retains the dynamic characteristics of thetime series data, can identify the non-physical parameters of the structural system (such as thedynamic response of the structure). Then, FNN method is empolyed to find the optimalembedding dimension of the input vectors, and using the modified Gram-Schmidt algorithmand AFPE criteria for selecting the number of the WNN hidden layer nodes, reducing thescale of the network model, and improving the computational efficiency and recognitionaccuracy.3) The wavelet function with the spatial-domain scaling properties is empolyed in thedynamic WNN model, it is only change the model output in the limited range of the inputtime series, such the property reduce the adverse effects between the nodes in the neural network, improve the accurate of the function approximation, and speed up the trainingprocess of the neural network.4) Be based upon the NARMAX method, the fuzzy clustering of the reconstruction statespace vector and the non-orthogonal wavelet neural network is combined to construct thedynamic time delay fuzzy WNN model of the structure non-parametric identification. Themodel can capture the dynamic characteristics of time-series sensor data effectively andaccurately, and can overcome the large local output error caused by the same spatial domainscaling property when the traditional WNN model dealing with the local accurate trainingdata.5) The adaptive LM-LS hybrid learning algorithm is empolyed to train the dynamicdelay fuzzy WNN model, the learning process is divided into the two-step iterative, Firstusing the LS algorithm to determine the linear parameters of the model, and then using theLM algorithm to adjust the nonlinear parameters of the model. The algorithm avoids thesecond derivative of the Gauss-Newton algorithm, overcomes the numerical instabilityproblems of the steepest descent algorithm, and improves the learning convergence rate of theWNN model significantly.6) The wavelet is empolyed in two aspects of the dynamic time delay fuzzy WNN model.○1Using discrete wavelet packet transform (DWPT) techniques to denoise the sensor data, itcan speed up the training convergence rat of the structural system identification model, andimprove the recognition accuracy of the system significantly.○2The wavelet function is usedas the activation function of the neural networks, and combine with the fuzzy clusteringtechniques to construct the fuzzy WNN model, The model can capture the dynamiccharacteristics of the time-series sensor data effectively and accurately.7) Ground on the dynamic time delay fuzzy WNN model of the structure non-parametricsystem identification, and the pseudo-spectral method, the damage identification method forlarge complex structures under ambient excitation is proposed. The damage identificationresults obtained by this method is the local damage of the structure within the selectedsub-structure, the damage all of the structure can be identified by selecting a series ofsubstructure, this method has the global and local damage identification capability of thestructure; The pseudo-spectral method cognitives damage in the form of the differentfrequency bands, contains the damage information wealthly; The pseudo-spectral method forthe structural damage identification without set the threshold of the damage index byexperience to determine whether the structure occur damage, as long as the damage index is greater than zero, the structure is determined to damage, and has a high damage sensitivity;The external dynamic loads is only used to incentive the structure sufficiently, without tomeasure the load spectrum, thus, it can identificate the damage of the engineering structuresunder ambient excitation, and with a strong engineering adaptability.更多还原