Optimized non-parametric fragility curve estimation based on intensity measure data clustering and parametric model averaging

Abstract : In the context of seismic risk analysis, the mean annual rate of failure of a structure is computed by integrating the seismic fragility curve with respect to the seismic hazard curve. Seismic fragility curves are used to describe the seismic vulnerability of structures and give the failure probability conditioned on ground motion intensity. Lognormal fragility curves are commonly used, while differences between lognormal and non-parametric fragility curves have been observed. Usually, the number of available real records is insufficient for estimating fragility curves with small uncertainty and synthetic ground motions are used. Here, a database of real ground motions is enriched with spectrally equivalent synthetic ground motions. Time-history analyses of a single degree of freedom structure are used to compute engineering demand parameter observations as a function of the ground motion intensity measure. Clustering of the intensity measure observations based on the combined real and synthetic ground motion database is used and an empirical distribution of the engineering demand parameter observations is estimated for every cluster and used to compute the probability of exceeding the damage state threshold. Thus, fragility curves are represented non-parametrically as failure probabilities conditioned on the intensity measure cluster centroids. It is shown that an optimization based on parametric model averaging of the models constituting the aggregate fragility curve may lead to a non-parametric fragility curve which is more precise than the un-optimized curve based on the same number of seismic response analyses.