Probabilistic Modelling Approach to Corrosion in Steel Bridges Using Bayesian Networks
Please login to view abstract download link
Computational modelling has become a cornerstone of modern structural analysis, enabling the prediction of infrastructure behavior under various degradation processes. However, traditional deterministic approaches exhibit critical limitations, such as the lack of uncertainty treatment and the use of simplifications that can compromise the reliability of the results. A prominent example is found in the macro-modelling of steel bridges, where corrosion is frequently characterized as a uniform phenomenon; this is primarily because the experimental characterization and local modelling of pitting corrosion entail prohibitive computational and experimental costs. Nevertheless, localized pitting corrosion is critical, as it induces stress concentrations and creates zones of vulnerability that can prematurely compromise structural integrity. To address these challenges, probabilistic approaches such as Random Fields (RF) have been employed, which are common in the study of continuous materials. However, in truss structures composed of multiple discrete elements, random fields present significant limitations, as their correlation depends on Euclidean distance rather than the topological connectivity of the structural members. Consequently, this paper employs a new spatial modelling approach using Bayesian Networks (BN) [1]. Unlike conventional methods, this model allows for the integration of visual inspection data to perform spatial inference of corrosion based on the structure's topology. Starting from a model calibrated via natural frequencies, the points of most severe localized degradation are identified to infer the condition of the remaining elements. The study performs a detailed comparison of how the deterministic versus the probabilistic approach affects the structural response. While the deterministic model provides a single value based on averaged degradation, the proposed approach generates confidence intervals that reflect real uncertainty. The results demonstrate that ignoring the localized nature of corrosion can lead to an overestimation of safety, highlighting the critical importance of this new method for infrastructure management.