A comprehensive approach for model-based SHM of bridges using Bayesian neural networks

Structural Health Monitoring (SHM) is essential for safeguarding infrastructure by providing continuous insights into structural performance and degradation over time. However, converting monitoring data into actionable decisions remains challenging due to uncertainties arising from incomplete structural knowledge, measurement noise, and modelling simplifications. This study addresses these challenges by proposing a model-based probabilistic framework that couples a finite element model with a Bayesian Neural Network (BNN) to classify structural states and track their evolution. The Bayesian formulation enables the integration of prior engineering knowledge, sensor measurements, and new information as it becomes available, allowing the model to continuously refine its predictions and quantify associated uncertainties. Unlike conventional neural networks, the Bayesian approach is particularly suited to small, noisy datasets, producing probabilistic outputs that support risk-informed decisions. The methodology is validated through a real-world application to a prestressed concrete box-girder bridge with vertically prestressed internal joints under multiple damage scenarios. Results confirm the framework’s capability to detect and localise damage, adapt to evolving conditions, and enhance reliability in long-term monitoring. This approach enhances SHM reliability and supports informed, data-driven decisions for structural assessment and long-term infrastructure management.