On distribution-based global sensitivity analysis by polynomial chaos expansion

Abstract

This paper presents a novel distribution-based global sensitivity analysis based on the Kullback–Leibler divergence derived directly from generalized polynomial chaos expansion (PCE). The synergy between PCE and Gram–Charlier expansion is utilized for derivation of novel and computationally efficient sensitivity indices. In contrast to a standard procedure for estimation of higher statistical moments, this paper reviews standard linearization problem of Hermite and Jacobi polynomials in order to efficiently estimate skewness and kurtosis direclty from PCE. Higher statistical moments are used for an estimation of probability distribution by Gram–Charlier expansion, which is represented by derived explicit cumulative distribution function. The proposed sensitivity indices taking the whole probability distribution into account are calculated for several numerical examples of increasing complexity in order to present their possibilities. It is shown, that the proposed sensitivity indices are obtained without any additional computational demands together with Sobol indices, and thus can be easily used as complementary information for a complex sensitivity analysis or any decision making in industrial applications. Application of the proposed approach on engineering structure is presented in case of prestressed concrete roof girders failing shear. Moreover, the potential of the proposed approach for reliability-oriented sensitivity analysis is investigated in pilot numerical example.