Applicability of meta-model assisted reliability assessment for dynamic problems: a comparison between regression-based methods

Abstract

There is a growing intent among engineers, stakeholders, and decision makers to use probabilistic methods for infrastructure assessment or design objectives. However, the corresponding limit state for such problems usually requires the construction of complex computational models, usually using commercial software without parallelization capability. Such a requirement makes performing reliability analysis computationally prohibitive, which is even more challenging for dynamic problems, since a very short time step is required to obtain sufficiently accurate predictions. This concern has led to several methods being proposed to surrogate the limit state function with a generally black box called a meta-model. A variety of them, such as Kriging, Polynomial Chaos Expansion (PCE), Artificial Neural Networks (ANN), and response surfaces (e.g., polynomial, spline, or radial-base functions), have been adopted for this purpose. These meta-models are typically trained on a limited data set collected by computing the true responses of carefully selected input variables. Their applicability for assessing the probability of failure has been studied individually in the literature for both benchmark and practical problems. However, as far as the authors are aware, no comparison has been made between them for dynamic problems. This comparison needs to be made from the point of view of both accuracy and performance (number of calls to the limit state function). In this context, this paper takes a systematic approach to evaluate their performance under identical conditions, i.e., with similar training datasets. For this purpose, the dynamic response of railway bridges with different spans excited by the passage of trains with a wide range of speeds is used as a reference problem.