A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering, School of Civil Engineering

Thesis (M.Eng.) - Asian Institute of Technology, 2000

Damage detection is a challenging problem that is under vigorous investigation by numerous research groups. When a structure suffers localized damage, its dynamic properties can change. Specially, damage can cause a stiffness reduction, with an inherent reduction in . natural frequencies, and increase in modal damping, and a change to the modal shapes. The development of experimental modal analysis techniques has facilitates the accurate measurement of modal parameters. Alongside this work, several methods have been developed to detect structural parameter change (structural damage) by using location-dependent changes in the modal data. This study focuses on the application of neural networks approach to the assessment of structural damage based on the modal test data. The procedure of using the neural networks to detect structural damage is outlined. A kind of training algorithms called as Back-propagation is used to recognize the modal data from numerical simulations, and the location and the extent of damage of a structure can be recognized by comparison of the outputs from the trained networks fed the modal test data obtained from the structure at the undamage and damage states. Some essential features of this algorithm that influences its searching efficiency are also discussed, especially, several practical concerns involving the structural damage detection are addressed, including the problem of incomplete mode shape measurements, the robustness of detection, and nonlinearity of system. The proposed approach is applied to measured mode data of a real 5 story steel frame. The comparison of numeric results with those from the traditional damage detection method indicates that the proposed method has the potential as a practical tool for a structure damage detection methodology.