This dissertation presents the development of structural identification procedure for the model updating and joint damage assessment in frame structures. It consists of three parts: i) the modal testing and identification techniques for civil engineering structures, ii) the structural joint modeling for model updating and joint damage assessment, and iii) the structural identification methods including neural networks and inverse perturbation techniques. Firstly, several modal testing methods applied to civil structures were summarized through literature survey. Particularly, the modal identification techniques without input information were introduced and applied to identify several benchmark structures. From the extensive analyses and comparisons, the mostly appropriate techniques for modal identifications are investigated considering the typical conditions of the tests. For modal parameter identification, it is highly recommended to use the frequency domain decomposition method during tests and the stochastic subspace identification method for more accurate estimation.
Secondly, the structural joint modeling was presented. For the simple and realistic structural joint model, a beam element with semi-rigid connections was proposed, and the joint damage severity was defined using the reduction ratio of the joint fixity factors before and after damage.
Lastly, two types of structural identification techniques were proposed for model updating and damage assessment. They are neural networks and inverse perturbation techniques. Several techniques were also utilized to improve the identification performance: such as substructural identification, noise injection learning for the neural networks, and statistical approach using measurement data perturbation scheme.
The proposed techniques were verified through a numerical simulation study for a 2-bay 10-story building and an experimental study for a 2-bay and 4-story frame structure. It was found that the inverse perturbation scheme is more efficient for baseline updating, while the neural networks technique is more effective for damage identification. It was also found that the noise injection learning could significantly reduce the effects of measurement noise and the combined usage of modal and static data can improve the identification performance. Through verification by experimental results, the substructural identification was found to be very feasible for damage assessment of large and complex structural systems.