Back to search

National Technical University of Athens (NTUA)

Reliability analysis and Bayesian updating in structural engineering using active learning

Abstract

dc:description

Uncertainty quantification supports the assessment of structural safety and the interpretation of observations through computational models. Its application to complex engineering systems often requires a large number of expensive model evaluations, particularly when estimating rare event probabilities or inferring uncertain model parameters. This Thesis addresses these computational challenges through active learning surrogate modelling for structural reliability analysis and Bayesian updating. In structural reliability analysis, accurate estimation of small failure probabilities can require a large number of expensive structural model evaluations. To reduce this computational cost, the quantified active learning Kriging (qAK) method combines a Kriging surrogate with active learning to concentrate structural model evaluations near the limit-state surface. A probabilistic measure of misclassification and weighted clustering guide the selection of batches of informative training points. A probabilistic stopping criterion controls enrichment within Monte Carlo and Subset Simulation frameworks. For high-dimensional problems, a comparative study examines Kriging with Partial Least Squares and neural networks with MC-dropout predictive uncertainty, each combined with single-point and multiple-point enrichment. The comparison distinguishes the effects of surrogate choice and enrichment strategy through the model evaluations and training time required to reach a common accuracy. For systems governed by differential equations, a separable physics-informed neural network approach is developed to approximate the solution field and estimate failure probabilities without requiring training data from numerical simulations. A two-phase training strategy and a proximity-weighted loss concentrate learning near the limit-state surface. These approaches are assessed through numerical benchmarks and engineering applications, including beam and plate models. In Bayesian updating, observations are used to revise the distributions of uncertain model parameters, but repeated evaluations of high-fidelity models can make posterior inference computationally prohibitive. This Thesis develops a multi-fidelity active learning approach that combines high fidelity Kriging and co-Kriging to exploit information from models with different computational costs and levels of accuracy. The updating task is formulated as an equivalent reliability problem using Bayesian Updating with Structural reliability methods. Multi-fidelity Bayesian optimization determines the BUS constant, while an improved qAK-based Subset Simulation procedure solves the associated reliability problem and obtains posterior samples. The resulting qAK-BUS framework directs computational effort according to predictive uncertainty and model fidelity to reduce the number of expensive evaluations. Its performance is investigated through numerical examples and the updating of a reinforced concrete bridge, demonstrating the application of active learning to Bayesian inference in structural engineering.

Degree

thesis:*
Grantor dc:publisher
National Technical University of Athens (NTUA)
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Prentzas, Ioannis
  • Πρέντζας, Ιωάννης

Subjects

dc:subject × 12

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/62937

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Prentzas, Ioannis; Πρέντζας, Ιωάννης. Reliability analysis and Bayesian updating in structural engineering using active learning. National Technical University of Athens (NTUA), 2026. http://hdl.handle.net/10442/hedi/62937