Seismic fragility analysis with artificial neural networks: Application to nuclear power plant equipment. (1st May 2018)

• Record Type: Journal Article
• Authors: Wang, Zhiyi; Pedroni, Nicola; Zentner, Irmela; Zio, Enrico
• Is Part Of: Engineering structures. Volume 162(2018)
• Page Start: 213
• Page End: 225
• Publication Date: 2018-05-01
• Subjects: Seismic probabilistic risk assessment -- Fragility curve -- Artificial neural network -- Feature selection -- Prediction uncertainty; Structural engineering -- Periodicals; Structural analysis (Engineering) -- Periodicals; Construction, Technique de la -- Périodiques; Génie parasismique -- Périodiques; Pression du vent -- Périodiques; Earthquake engineering; Structural engineering; Wind-pressure; Periodicals; 624.105
• DOI: 10.1016/j.engstruct.2018.02.024
• Languages: English
• ISSNs: 0141-0296

Abstract

Highlights: A methodology of computation of fragility curves with ANNs is presented. An efficient way of selecting the most relevant seismic intensity measures is presented. ANN prediction uncertainties are decomposed into aleatory and epistemic components. The impact of ANN prediction uncertainties on the fragility curves is investigated. Abstract: The fragility curve is defined as the conditional probability of failure of a structure, or its critical components, at given values of seismic intensity measures (IMs). The conditional probability of failure is usually computed adopting a log-normal assumption to reduce the computational cost. In this paper, an artificial neural network (ANN) is constructed to improve the computational efficiency for the calculation of structural outputs. The following aspects are addressed in this paper: (a) Implementation of an efficient algorithm to select IMs as inputs of the ANN. The most relevant IMs are selected with a forward selection approach based on semi-partial correlation coefficients; (b) quantification and investigation of the ANN prediction uncertainty computed with the delta method. It consists of an aleatory component from the simplification of the seismic inputs and an epistemic model uncertainty from the limited size of the training data. The aleatory component is integrated in the computation of fragility curves, whereas the epistemic component provides the confidence intervals; (c) computation of fragility curves with Monte Carlo method and verification of the validity of the log-normal assumption. This methodology is applied to estimate the probability of failure of an electrical cabinet in a reactor building studied in the framework of the KARISMA benchmark.