Abstract
This study aims to show the applicability of sparse modeling, which is drawing attention in the field of machine learning in recent years, to the surrogate modeling of the finite element (FE) analysis for the use in the structural reliability analysis of aging exiting bridges. The target of surrogate modeling here was the Monte Carlo calculation of the FE model of a steel plate-girder bridge with corrosion at the ends of girders and bearings. The inputs were uncertain FE model parameters, such as material propertyies and boundary conditions, and the output was the maximum Mises stress at the ends of main girders under the loading of designed live load. The performance of some regression methods; least-square method (LSM), Ridge, and Lasso, which can give the sparse solution, were compared in applying them to the surrogate modeling. As a result, the Lasso regression can make a proper surrogate model only by using 50 training data that was the one-third of training data required in the LSM. The reliability index β derived by the Lasso surrogate model then showed good agreement with that derived by the Monte Carlo FE model calculation with 500 samples. Moreover, the solution of Lasso, which showed sparsity, indicated the FE model parameters that had significant sensitivities to the output, i.e., the maximum Mises stress, under the corroded condition. It was then concoluded that the surrogate modeling by Lasso was effective in the structural reliability analysis of existing bridges, which has high uncertainties in a number of FE model parameters due to detoriaration or damages.
