In this paper reliability based design optimization by using radial basis function networks (RBFN) as surrogate models is presented. The RBFN are treated as regression models. By taking the center points equal to the sampling points an interpolation is obtained. The bias of the network is taken to be known a priori or posteriori. In the latter case, the well-known orthogonality constraint between the weights of the RBFN and the polynomial basis functions of the bias is adopted. The optimization is performed by using a first order reliability method (FORM)-based sequential linear programming (SLP) approach, where the Taylor expansions are generated in intermediate variables defined by the iso-probabilistic transformation. In addition, the reliability constraints are expanded at the most probable points which are found by using Newton’s method. The Newton algorithm is derived by proposing an in-exact Jacobian. In such manner, a FORM-based LP-formulation in the standard normal space of problems with non-Gaussian variables is suggested. The solution from the LP-problem is mapped back to the physical space and the suggested procedure continues in a sequence until convergence is reached. This is implemented for five different distributions: normal, lognormal, Gumbel, gamma and Weibull. It is also presented how the FORM-based SLP approach can be corrected by using second order reliability methods (SORM) and Monte Carlo simulations. In particular, the SORM approach of Hohenbichler is studied. The outlined methodology is both efficient and robust. This is demonstrated by solving established benchmarks as well as finite element problems.

This content is only available via PDF.
You do not currently have access to this content.