Global Reliability And Sensitivity Analysis Of Structures Using MLS-SVM Meta-Model

When studying the problem of computing structural reliability of complex structures, problem that the performance function is of implicit case is often faced. A usual approach of solving this problem is to use meta-models or surrogate models to take the place of the performance functions. Input-output data points are used in surrogate models as training points gained by some explicit maths model, and predict the responses of unkown points. Typical surrogate models include polynomial responsing surface model, artificial neural neatwork model, radial basis model and support vector machine model, Kriging model and so on. Surrogate models are being used widely in the field of engineering, including structural optimaization design, structural reliability analysis and so on. Howerver, computing structural global sensitivity and global reliability using surrogate models is still rare.1) On the basis of the general support vector machine model, moving least squares, referencing least squares support vector machine mode are put in conjuction for a new moving least squares support vector machine model. In principle, the training points and points to predict in the model can be self-adapted. Meanwhile, through the numerical analyses and simple structural reliability analyses, comparing to LS-SVM and SVR models, MLS-SVM model can be more accurate.2) In the non-linear support vector regression machine, we usually see the inner product as kernel function. More commonly, we usually set the kernel function as Gaussian kernel function. In this paper, the reproducing kernel function in a reproducing kernel space can be used as a support vector machine kernel function. So we build such a model, and SVR model based on Gaussian kernel were analyzed and compared. Meanwhile, LS-SVM and Kriging model based on reproducing kernel are also established. We find that SVR model based on reproducing kernel can be more accurate than the one based on Gaussian kernel, however LS-SVMmodel and Kriging model based on reproducing kernel don’t make it happen. Two ways of Kriging model, namely, UQLAB and DACE are compared in two numerical examples. The conclusion comes that the latter is better than the former.3) SVR, MLS-SVM and Kriging models are used to compute the structural global sensitivity and global reliability of RC frames. The result of them shows that methds based on surrogate models to compute the sensitivity and global reliability is correct. And results computed by several surrogate models are compared to show the accuracy of them.