System confidence-based design optimization using a multi-output Gaussian process accounting for epistemic uncertainty

초록

AbstractWhile system reliability-based design optimization (SRBDO) is critical for ensuring the safety of sophisticated engineering systems, the reliance on surrogate models to mitigate computational costs inevitably introduces epistemic uncertainty. This challenge is particularly exacerbated when multiple responses are correlated, a common scenario often overlooked in reliability analysis. Therefore, we propose a novel framework for system confidence-based design optimization (SCBDO) that rigorously accounts for such correlated surrogate model uncertainty. The proposed method utilizes a multi-output Gaussian process (MOGP) to simultaneously predict multiple system responses. By propagating the MOGP’s joint predictive distribution, the framework derives the entire probability distribution of the system reliability. The proposed approach can capture the correlation between epistemic uncertainties in component responses, which is often neglected in conventional confidence-based approaches. This enables the SCBDO formulation to find a cost-effective design that satisfies a target reliability with a predefined confidence level under surrogate model uncertainty. To ensure computational tractability for the nested calculations, the framework integrates efficient sampling of the high-dimensional predictive distribution with a variance-reduced importance sampling scheme. Numerical studies, including mathematical examples and an engineering application involving an electric multiple unit (EMU), demonstrate that the proposed SCBDO provides a more accurate representation of system-level uncertainty. © 2026 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.

키워드