Flexibility Matrix And Neural Network-based Damage Identification Methods

In recent decades, the growing numbers of structures near or beyond their design lifetime, combined with the increasing cost of maintenance and repair, have accelerated the development of damage detection and health monitoring approaches. Many damage identification methods based upon analysis of structural nature characteristics use structural dynamic response information that seek to identify change in the modal parameters of the pre-damaged and the post-damaged state of the structure can exhibit the accuracy and efficiency. But the changes of structural parameters caused by damage are likely to result in the structural modes of the post-damaged not in the same order as those corresponding ones of the pre-damaged. In the term of theory, mode jumping may be occurred.The dissertation presents a novel damage identification index, namely flexibility diagonal curvature that only require the low order mode parameters of the post-damaged structure without those of the pre-damaged structure to avoid considering the ingredients of mode jumping and high order mode parameters that are difficult to measure to affect the accuracy of damage identification. A numerical simulation by the finite element model (FEM) of a cantilever beam is performed in the single-damage cases and the multi-damage cases of the structure. The two damage identification indexes as flexibility diagonal curvature and flexibility curvature are separately used to identify the existence, location, and extent of damage. The results show that the proposed index and the existed index are very simple and effective to identify the damage. In addition, the sensitivity of the two damage identification indexes to damage has been analyzed.The damage identification method based on damage identification index constructed by the modal parameters provides an approach for identifying the damage of structure but only evaluates qualitatively the location and extent of damage. Artificial neural networks (ANNs) can be identify quantitatively the location and extent of damage due to many strong capabilities, such as self-organization and learning, pattern recognition and classification, tolerance of incomplete and faulty data. Presently, the applications of ANNs to structural damage identification problems have attracted increasing attention. It's still a question that how to select the input parameters and design a high effective ANNs has a considerable influence on the performance and efficiency of neural networks for the identification of structure damage. In this paper, input vectors to neural networks, comprised of flexibility curvature and flexibility diagonal curvature respectively and to be suitable for identification of structural damage location and extent. Two improved back propagation (BP) network that are variable learning rate and momentum vector algorithm and conjugate gradient algorithm and radial basis function (RBF) network are applied. The feasibility and effectiveness of neural networks with the two kinds of damage identification indexes to assess the damage is verified by using the modal data obtained by a cantilever beam. The study shows that BP networks and RBF networks can be identify the damage locations and extents correctly in single-damaged cases. Comparing the performance of neural networks, the two improved BP algorithms, the latter has more efficiency and accuracy.