Scholarlink

<p>The reliable estimation of time-variant failure probabilities in complex engineering systems poses a significant computational challenge, primarily due to the substantial number of high-fidelity model evaluations required by conventional Monte Carlo simulation. This work presents an enhanced multi-fidelity Gaussian process (EMFGP) active-learning framework that enables reliable estimation of time-variant failure probabilities while significantly reducing the number of required high-fidelity model evaluations. The main idea is to combine a low-fidelity model that captures general trends with a high-fidelity correction model that accounts for finer, more critical details. One of the key contributions of this work is a flexible kernel function (covariance structure) that allows the surrogate model to capture both broad spatial features and intricate time-dependent behavior within a single formulation. To improve reliability, particularly when the available data are sparse or nearly deterministic, this study adopts a variance-controlled, data-driven initialization strategy. This helps stabilize hyperparameter learning and prevents numerical instability during model training. Across five examples, EMFGP achieved similar reliability accuracy to Monte Carlo simulations while using around 20–30% fewer high-fidelity evaluations, and up to 40% fewer in cases where the low-fidelity model shows strong correlation. This highlights the method’s efficiency and suitability for time-dependent reliability analysis, especially when high-fidelity simulations are costly.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An enhanced multi-fidelity Gaussian process framework for efficient time-variant reliability analysis

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Tayyab Zafar,
  • Yi Guan,
  • Wei Zhang,
  • Zhonglai Wang

摘要

The reliable estimation of time-variant failure probabilities in complex engineering systems poses a significant computational challenge, primarily due to the substantial number of high-fidelity model evaluations required by conventional Monte Carlo simulation. This work presents an enhanced multi-fidelity Gaussian process (EMFGP) active-learning framework that enables reliable estimation of time-variant failure probabilities while significantly reducing the number of required high-fidelity model evaluations. The main idea is to combine a low-fidelity model that captures general trends with a high-fidelity correction model that accounts for finer, more critical details. One of the key contributions of this work is a flexible kernel function (covariance structure) that allows the surrogate model to capture both broad spatial features and intricate time-dependent behavior within a single formulation. To improve reliability, particularly when the available data are sparse or nearly deterministic, this study adopts a variance-controlled, data-driven initialization strategy. This helps stabilize hyperparameter learning and prevents numerical instability during model training. Across five examples, EMFGP achieved similar reliability accuracy to Monte Carlo simulations while using around 20–30% fewer high-fidelity evaluations, and up to 40% fewer in cases where the low-fidelity model shows strong correlation. This highlights the method’s efficiency and suitability for time-dependent reliability analysis, especially when high-fidelity simulations are costly.