Systems Modeling Using Big Data Analysis Techniques and Evidence

Abstract

In the present contribution, the potentials of utilizing techniques of big data analysis as a means to improve the understanding of complex probabilistic system representations are investigated. It is assumed that a probabilistic model is available for the representation of the system performances and that an adequate Monte Carlo simulation technique is available and applied for the probabilistic analysis of these. Model-based clustering analysis is then applied to establish a visual representation of the Monte Carlo simulated scenarios of events leading to different performances of the considered system. Various conditioning events on the simulated scenarios, such as specific failure events, are readily introduced by sorting. Assuming that the Monte Carlo simulated scenarios of events are utilized to establish a surrogate representation of the considered system, variance based sensitivities are derived for both the case of independent and dependent random variables. To this end, so-called ANOVA and the very recently formulated ANCOVA decomposition s are applied. The proposed scheme is illustrated on a simple example in which the probabilistic characteristics of non-linear structural performances of a moment resisting frame structure are considered. It is seen from the example that big data techniques may readily be applied to provide significant insights on which scenarios of events govern the probabilistic characteristics of the performances of the system, and with respect to how uncertainties associated with the random variables used to model the system propagate in the system and affect its responses. The latter is especially useful when aiming to reduce model complexity, but also in the context of structural health monitoring where response characteristics that contain significant information about the state of the system must be identified.