Nested surrogate model for discrete parameter optimization of structural reliability analysis

The dynamic Bayesian network which is widely used for efficient reliability analysis, requires random variable discretization. To estimate the accuracy of the discretization results, the Kullback-Leibler (KL) divergence and reliability index errors are used as metrics. Since both errors have a trade-off relationship in a combined discretization method, determining the optimal discrete parameters is crucial to accurate and efficient reliability analysis. However, the optimization process is significantly time-consuming, so we proposed the nested surrogate model for the discrete parameter optimization in structural reliability analysis. The first-stage surrogate model based on the Gaussian process regression (GPR) method calculated the stress intensity factor (SIF). The sample points for GPR were obtained by using the finite element analysis (FEA) method. The artificial neural network for the nested surrogate model is then trained using the data generated from the first-stage surrogate model, to determine the SIF range for the crack growth behavior. The nested surrogate model was implemented to optimize the discrete parameters of the crack length distribution. The proposed method considerably reduced the computation time in obtaining the optimal discrete parameters that capture the features of regions with low probability.