Life-cycle performance prediction and interpretation for coastal and marine prestressed concrete beams using active learning-enhanced Bayesian neural networks

In marine environments, prestressed concrete (PC) structures suffer from chloride-induced deterioration, impacting their serviceability and safety. Traditional deterministic and semi-probabilistic methods inadequately address the deteriorating mechanisms and uncertainties in environmental, material, and structural parameters, hindering accurate structural performance predictions. This study introduces an active learning-enhanced Bayesian neural network (BNN) framework for predicting the life-cycle performance and reliability of PC beams in coastal environments. The BNN is trained on a dataset generated via Latin Hypercube Sampling from a comprehensive model ensuring representative input. The active learning component strategically selects the most informative points, enhancing modeling accuracy and efficiency. The Guangdong-Hong Kong-Macao Greater Bay Area is chosen for a case study of PC hollow beams. A life-cycle prediction model for PC structures was developed, considering pitting effects on the geometry, mechanical, and bond properties of prestressing bars, etc. Finally, time-dependent reliability analysis is performed using the surrogate model and Monte Carlo simulation. Results indicate that the BNN achieves high accuracy with active learning. SHAP analysis identifies key factors affecting the behaviors of PC beams, highlighting the importance of material properties and environmental conditions. Also, reliability analysis emphasizes the impact of two-dimensional transport on structural reliability.