The Support Vector Regression Method For Structural Damage Detection

As more and more high buildings are constructed, it is imperative to assess health states of the structures after catastrophic events, such as earthquake, typhoon and explosion. In the first chapter of this thesis, from the viewpoint of structural health monitoring, the state-of-art of the researches about structural health monitoring is depicted, some ongoing implementations of structural monitoring systems are introduced, the background of the research project and the content of this thesis are outlined also.Traditional statistical theory aims at the asymptotic theory when sample size is tend to infinity. However, in many practical cases, samples are limited. Most of existing method based on traditional statistical theory may not work well for the situation of limited samples. Statistical Learning Theory (SLT) is a new statistical theory framework established from finite samples. SLT provides a powerful theory fundament to solve machine learning problem with small samples. Support Vector Machine (SVM) is a novel powerful machine learning method based on SLT. SVM solves practical problems such as small samples, nonlinearity, over learning, high dimension and local minima, which exist in most of learning methods, and has high generalization. In this thesis, SLT, support vector regression and kernel methods are studied from the respective of integration of theory, algorithm and application.The standard SVM need to solve a quadratic problem, it is time-consuming for large data sets and not suitable to solve dynamic problems. Least Squares Support Vector Machines (LS-SVM) is an SVM version which involves equality instead of inequality constraints and works with a least squares cost function, the problem is simplified by this way. Based on the Least Squares Support Vector Machines method, Incremental LS-SVM and Online LS-SVM method is proposed for identification of structural systems in this paper. It efficiently updates a trained LS-SVM by means of incremental and decremental pruning algorithms whenever a sample is added to, or removed from, the training set. The method overcomes the drawback of sparsenesslost within the standard LS-SVM and makes online training for the LS-SVM possible. Examples of linear and nonlinear hysteretic structure parameters for online identification problems show the robustness, efficiency, and capability of damage tracing of the proposed method.To verify the practical performance of LS-SVM method, an experimental model under the excitation of EL Centra wave was studied. The LS-SVM method identified the stiffness of the model in the elastic phase only from acceleration records. Comparison between the actual records and identified records showed the efficiency of LS-SVM.At last, the main contributions and conclusions of this thesis are summarized and some problems to be addressed about the research are set forward.