Use of Relevance Vector Machine in Structural Reliability Analysis

The applicability of a relevance vector machine to structural reliability analysis is introduced in this work. Samples covering the important domains in the input space are first generated and selected by a modified Metropolis algorithm. Then, the samples are employed to build a surrogate model with the help of a relevance vector machine to approximate the real performance function. The surrogate model is further used for reliability analysis as a substitution of the real one, as calls to the real performance function may be time consuming or cumbersome, especially in engineering cases. In this work, the relevance vector machine and the Metropolis algorithm are combined in an iterated learning process, constantly updating the samples used according to the convergence trend of the failure probabilities. Both numerical and engineering examples have been analyzed and discussed with the proposed method, as well as comparisons to some other classical reliability methods. The research in this work shows that the proposed method based on the relevance vector machine is a viable alternative for structural reliability analysis.