Structural safety evaluation using Bayesian ensemble neural networks

Jin-Ling Zheng

doi:10.1016/j.engstruct.2025.119709

2025-04-01

PRE

Jin-Ling Zheng

Sheng-En Fang *

DOI：10.1016/j.engstruct.2025.119709

原文链接

原文求助

摘要

En 中文

Safety evaluation is a pivotal issue for operational civil structures during their service lives. Recently, deep learning-based evaluation strategies have emerged, and such methods often require a substantial amount of training samples to prevent overfitting. However, this precondition is often difficult to satisfy in practice due to insufficient samples. Hence, a Bayesian ensemble neural network (BENN) has been proposed to overcome this drawback. Firstly, the network parameters of a Bayesian neural network (BNN) are established on probability distribution estimation to consider the uncertainties in a structure, which is divided into several substructures for evaluation. A multiple sampling strategy on the network parameter distributions yields different deterministic NNs. Secondly, the Bagging ensemble learning has been adopted to treat a BNN as a base learner, whose prediction will be used for an ensemble prediction of a substructure. A BENN is actually the ensemble of several BNNs (base learners). Specifically, the membership degree of each base learner's predictions is calculated and normalized to derive the corresponding weight. The ensemble prediction is obtained through the weighted summation of the predictions of all base learners. Meanwhile, the entropy that measures the structural uncertainty of each substructure, with corresponding weights calculated via the entropy weight method to construct an overarching structural state indicator. The effectiveness of the BENNs is validated through the numerical simulations and practical experiments conducted on a frame structure. As the structural degradation increased, the state indicator decreased from 16.88 to 14.59 for the numerical frame, as well as from 16.93 to 16.23 for the experimental frame.

Keyword:

Structural safety evaluation

Bayesian ensemble neural network

Bagging ensemble learning

Variational inference

Entropy weight method