Présentation : Charbel-Pierre EL SOUEIDY, équipe TRUST
Invitation : Equipe MEO
M. Charbel-Pierre EL SOUEIDY, membre du GeM au sein de l’équipe TRUST donnera un séminaire le 30 mars 2017 à 14h00 à Centrale Nantes (Amphi A).
Ce séminaire portera sur la méta-modélisation en particulier pour les phénomènes dépendant du temps (dégradation, endommagement) et en présence d’incertitudes, intéressera, nous l’espérons, le plus grand nombre. Ce sujet a fait l’objet d’une présentation en 2016 par Lara Hawchar en finale inter-régionale de la thèse en 180 secondes MT180.
Recent Developments in Metamodeling Techniques for Time-Variant Reliability Analysis
Time-variant reliability analysis is a challenging task in practical engineering applications. It aims to evaluate the probability that a structural system successfully performs its intended function throughout its service life. Time dependent uncertainties, such as dynamic loadings and deterioration in material properties, are usually described as stochastic processes whose discretization may lead to a significant number of truncated terms (i.e. the input variables) leading to the curse of “dimensionality”. A second challenging issue in time-dependent reliability analysis is that it usually requires the computation of the extreme value of the limitstate function over the structure lifetime. The extreme value function may be highly nonlinear and follow multimodal distribution. Traditional reliability methods, such as FORM may produce large errors whereas simulation methods such as the Monte-Carlo simulation (MSC) can be prohibitively time-consuming.
In this talk, two Polynomial Chaos (PC)-based approaches for time-variant reliability analysis are proposed. The former method relies on the Principal Component Analysis (PCA). First, the time interval of study is discretized and an instantaneous performance function is associated to each time node. Then, a principal component analysis is performed in order to represent all these functions with a reduced number of components. A PC expansion is next used to approximate each of these components. The latter method is based on the Efficient Global Optimization (EGO) approach. Considering a reduced experimental design, the maximum value of the performance function for each realization is first obtained performing EGO. The obtained set of maxima is then used to approximate the extreme value function with PC by means of a non-intrusive regression scheme. In both cases, a surrogate model of the response is obtained on which classical simulation-based techniques such as the MCS can be implemented for an easy evaluation of the time-dependent probability of failure.
The proposed approaches are compared with some recent time-dependent reliability analysis methods. Non-linear time-dependent limit state functions of both non-Gaussian and nonstationary processes are considered. The accuracy and efficiency of the proposed approaches are demonstrated in all cases.