Gaussian process-based data augmentable deep Bayesian neural network: Application to estimate the stress transfer length in pretensioned concrete beams

Abstract

Accurately estimating the stress transfer length (TL) in pretensioned concrete beams is critical for ensuring structural safety and performance. This study proposes a novel Gaussian Process-based Data Augmentable Deep Bayesian Neural Network (GPDA-DBNN) to address the limitations of conventional methods, especially under small datasets. The proposed GPDA-DBNN model achieved a test R2 of 0.86, outperforming DBNN, long shortterm memory, transformer, and tree-based machine learning models. Notably, in the estimation of a missing TL value, the GPDA-DBNN provided predictions (mean: 725.84 mm) closely matching the reference range of similar cases (min: 645 mm, max: 870 mm), while alternative models showed significant deviations. The augmented dataset proved essential in enhancing prediction reliability, as evidenced by narrower uncertainty intervals and superior alignment with empirical data. These results demonstrate that this combination of data augmentation and a Bayesian approach offer a robust and uncertainty-aware approach for TL estimation in pretensioned concrete, supporting safer and more efficient structural design.