Data-Driven Global Sensitivity Analysis Using the Arbitrary Polynomial Chaos Expansion Model

Considering the existence of stochastic uncertainty, quantifying the failure probability of structural systems is one of the important subjects of reliability research, and global sensitivity analysis (GSA) has been widely utilized to estimate the influence of the uncertainty on the system failure probability. Polynomial Chaos Expansion (PCE), a kind of surrogate model, is often utilized to replace the original model considering the heavy calculation cost of GSA. However, the traditional PCE and Sobol' method are not suitable for the situation that input variables of systems do not have complete probabilistic distribution functions. To solve this problem, a data-driven sensitivity analysis method is proposed in this paper. The surrogate model of the original model is constructed by arbitrary Polynomial Chaos Expansion (aPCE), and different order sensitivity indices of the actual systems are calculated by Sobol' combining with the Monte Carlo simulation. Compared with the traditional Sobol' method, the developed algorithm has a wider scope of application in practice. The experiment results show that the proposed method is an effective GSA tool when the distribution types of input variables are not completely known.