Parametric models averaging for optimized non-parametric fragility curve estimation based on intensity measure data clustering

Abstract

Seismic fragility curves give the probability of exceedance of the threshold of a damage state of a structure, or a non-structural component, conditioned on the intensity measure of the seismic motion. Typically, fragility curves are constructed parametrically assuming a lognormal shape. In some cases, which cannot be identified a priori, differences may be observed between non-parametric fragility curves, evaluated empirically based on a large number of seismic response analyses, and their estimations via the lognormal assumption. Here, we present an optimized Monte Carlo procedure for derivation of non-parametric fragility curves. This procedure uses clustering of the intensity measure data to construct the non-parametric curve and parametric models averaging for optimized assessment. In simplified case studies presented here as illustrative applications, the developed procedure leads to a fragility curve with reduced bias compared to the lognormal curve and to reduced confidence intervals compared to an un-optimized Monte Carlo-based approach. In the studied cases, this procedure proved to be efficient providing reasonable estimations even with as few as 100 seismic response analyses.