Data-driven federated hierarchical Bayesian neural network framework for region-specific prediction of chloride content in bridge decks

Chloride ingress induces reinforcement corrosion and reduces bridge durability. Deicing-salt application during winter further accelerates deterioration, making reliable long-term prediction essential for maintenance planning. However, two limitations hinder data-driven approaches: chloride data are primarily obtained via destructive testing, yielding scarce long-term observations, and security constraints restrict predictive modeling across agencies. To address these limitations, we propose a federated hierarchical Bayesian neural network framework, in which a federated scheme, aligned with the small number of agencies managing bridges, allows each district to share only BNN–ANN model parameters rather than raw data, while a hierarchical structure separates region-specific patterns to capture inter-regional heterogeneity. A global Bayesian neural network captures deck-level aleatoric and epistemic uncertainties and common temporal features, while region-specific exponential trend parameters are retained locally. The proposed federated framework showed robust predictive performance, and prediction intervals narrowed in data-scarce service-year ranges, confirming effective knowledge transfer to data-poor regions.