22 March 2025 in Engineering Structures
22 April 2026 in SSRN Electronic Journal
Bayesian failure probability inference (BFPI) provides a coherent framework for developing high-performing Bayesian active learning (BAL) reliability analysis methods. However, the resulting BAL methods primarily increase efficiency by reducing the number of model (usually high-fidelity (HF)) evaluations. When each HF evaluation becomes expensive, the overall cost can still be computationally prohibitive. To address this issue, this study extends BFPI to multi-fidelity BFPI (MF-BFPI). Furthermore, building on MF-BFPI, a new BAL method is developed, termed 'multi-fidelity partially Bayesian active learning cubature' (MF-PBALC). Thereby, inexpensive yet informative low-fidelity (LF) evaluations are leveraged effectively with reduced reliance on HF evaluations. MF-BFPI employs an MF Gaussian process (MF-GP) model to represent the unknown HF performance function, enabling Bayesian inference using dense LF information to capture the overall trend of the performance function and sparse HF data to correct the discrepancy. Thereafter, the MF-PBALC method is established by framing MF-BFPI into an active learning loop. The stopping criterion is developed by exploiting the structure of the posterior mean of the failure probability. The learning function is built on the same posterior mean, while additionally incorporating predictive statistics across fidelity levels, thereby jointly selecting the best next evaluation point and fidelity level by properly accounting for cross-fidelity correlation and model cost ratios. Three numerical examples and one engineering application are used to illustrate the performance of the proposed method. The results empirically demonstrate that MF-PBALC can estimate failure probabilities with desired accuracy and efficiency.
11 March 2026 in Structures