Fast and adaptive reliability analysis via kriging and partial least squares

This paper studies a framework for reliability analysis using Kriging with active learning and a dimensionality reduction technique based on partial least squares. The use of Kriging with partial least squares (KPLS) aims to accelerate the active learning process of computing the failure probability, which can be expensive due to the requirement to build several Kriging models as new samples are added. This speed-up is primarily achieved through the faster training time of the KPLS without any significant loss in the Kriging model’s accuracy. Experiments were performed on an analytical and heat-conduction problem with uncertain coefficients. Results show that KPLS with four principal components can significantly accelerate the computation of the failure probability while still yielding accurate values compared to the ordinary Kriging on high-dimensional problems. Moreover, in terms of the number of simulation calls, KPLS is also faster than the ordinary Kriging on the analytical problem. The result signifies that KPLS offers faster convergence in terms of overall time and the number of simulations.