Studies On Structural Damage Detection By Finite Element Model Updating

Although FE modeling updating techniques have been widely studied and applied to different engineering structures over the past two decades, some problems, or in other words drawbacks, still exist. For example, for complex engineering structures having numerous parameters to be updated, the optimization problems are often ill-conditioned or require considerable computational effort resulting in slow convergence speeds. Due to it, this thesis aims to develop a few state-of-the-art model updating methods or algorithms for the purpose of improving the optimization performance during the updating process, which finally benefits the efficiency and accuracy of damage detection. Three model updating methods or algorithms have been proposed including a damage parameterization based method, a new artificial intelligence algorithm for multiobjective optimization and a response surface based model updating method. The first two methods still fall into the traditional model updating technique category, while the third one can be regarded as a great improvement on traditional methods since FE models are replaced by response surface models during the optimization process. Firstly in the aspect of damage parameterization, a bidimentional damage function is established to parameterize the damage distribution of a bidimentional FE model in the interest of reducing the number of unknown variables and thus generating a well-conditioned optimization problem. The correction factors of finite elements or substructures are linked by a global damage function and by this means, the real updating parameters turn to be the coefficients of the damage function resulting in considerable reduction of the unknown variables. Secondly for the employment of artificial intelligence algorithms, recently-developed particle swarm optimization algorithm is incorporated with the genetic algorithm in order to propose a new algorithm for rapidly searching Pareto minima in a multiobjective optimization problem. In specific application, particle swarm optimization shows its easy implementation and high convergence speed. But due to its single-point-centered characteristic, particle swarm optimization doesn't perform well in real-world complex problems like damage detection in which the search has to be made in multi-constrained solution spaces requiring global searching ability. On the other hand, genetic algorithms work well in searching global minima, but they are generally not cost-efficient for complex optimization problems since they require a relatively long time to obtain a Pareto front with high quality. Therefore, in this thesis above two algorithms are incorporated to obtain a fast and accurate convergence process in multiobjective optimization problems. Meanwhile, according to the literature inquiry, it is the first time that the proposed algorithm has been used in the topic of damage detection using multiobjective model updating strategies. Thirdly, to improve the traditional FE model updating strategy, a new model updating method is proposed to constitute an inverse problem in which response surface models are used as surrogates for FE models. The updating efficiency can be considerably improved without losing prediction accuracy since FE models are no longer required during the updating iterations. Compared with traditional FE model updating methods, the response surface based methods do not require the construction of sensitivity matrices and the computation of FE models. At the same time, the implementation is relatively easy with much higher computation efficiency. Compared with neural networks based model updating methods, the response surface based methods require much fewer samples and can provide explicit mathematical equations which may be easily connected to the modules of a damage detection system. In addition, based on the analysis of variance, the significance of updating parameters can be quantitatively analyzed for parameter screening purposes. And by this means, the subjectivity of experiential judgment or the limitations of sensitivity based methods can be avoided.Furthermore, for the first time, this thesis proposes a new parameter named power mode shape which utilizes the statistical properties of response signals. The power mode shape is proposed for the early damage detection of engineering structures. It possess similar geometric shapes to traditional mode shapes but are constructed based on different theory without any modal parameter extraction. The development of power mode shape aims to avoid the modal parameter extraction for structures having complex modal situations, and to enhance the robustness of damage indicators to random noises. Meanwhile, two parameters of power mode shapes curvature and power flexibility are derived from the concept of power mode shape and then two damage indicators are defined for damage localization.Lastly, the proposed methods in this thesis have been validated using different numerical and experimental structures including a tested reinforced concrete frame and an experimental full-scale bridge. The satisfactory damage detection results prove the feasibility and reliability of these methods and also show their potentials in relevant realms.