PSI - Issue 64

Christoph Brenner et al. / Procedia Structural Integrity 64 (2024) 1240–1247 Christoph Brenner et al./ Structural Integrity Procedia 00 (2019) 000 – 000

1241

2

1. Introduction Bridges are exposed to a variety of complex influences and damage mechanisms. At the same time, however, a high level of safety is required, which is commonly ensured by regular visual inspections. In order to support the existing inspection system and to receive early warnings of possible damage, Structural Health Monitoring (SHM) systems are installed on new and existing bridges. At the same time, the current condition of the structure should be captured in a simulation model in order to derive more targeted predictive maintenance actions. For this purpose, the SHM data has to be integrated into the bridges’ model. The current challenge is to transfer the mathematical methods and simulation techniques that have been developed at a small scale to real-world applications and thus to actual structures (Niederer et al. (2021)). The calibration of the simulation model can be treated as an inverse problem in which the model parameters are adjusted iteratively until the SHM data and the results of the simulation model match as closely as possible. To do this, the following optimization problem must be solved by minimizing the functional ℑ for a suitable objective function with respect to the parameters : ∗ = ℑ( ) (1) Especially for large-scale structures, a high number of parameters is involved to describe the condition of the structure as accurate as possible. Therefore, the minimization problem is generally ill-conditioned and solving the problem may be very computationally expensive. The use of conventional optimization algorithms also requires recalibration for each changed state of the structure. To address these issues, an alternative data-driven approach is proposed using a database of models with different damage states in combination with a suitable model selection algorithm. The model database is created during an offline training phase and serves as the foundation for training the model selector, streamlining the online application phase and enabling fast, efficient decision-making. The automatic update of Finite Element models with strain data for large-scale bridges was investigated by Okasha et al. (2012). The minimization problem was solved there using particle swarm optimization. Ritto and Rochinha (2021) used various machine learning classifiers alongside physics-based models to identify structural damage. Their investigations focused on the frequency response on a clamped bar structure modelled as a spring damper system. Fernandez-Navamuel et al. (2022) integrated Deep Neural Networks with Finite Element simulations to detect damage in bridges. Therefore, the dynamic response of the structure was analyzed. Svendsen et al. (2023) implemented a hybrid approach to monitor the condition of a large steel bridge. The support vector machine algorithm was used for structural diagnosis based on calibrated Finite Element models and modal properties of the bridge. Torzoni et al. (2024) developed a Digital Twin framework tailored for civil engineering structures, focusing on the use of probabilistic graphical models. They used reduced-order modelling together with artificial neural networks on the example of a railway bridge with six model parameters. Optimal classification trees (OCT), developed by Bertsimas and Dunn (2017), have demonstrated significant advantages in selecting the optimal model from a model database due to their high accuracy while the decisions remain interpretable. However the method has so far only been applied to an unmanned aircraft featuring fictitious, widespread damage with substantial stiffness degradations of up to 80% (Kapteyn et al. (2022)). The application to large-scale structures such as bridges and realistic damage scenarios, such as discrete cracks, has not yet been tested. This paper expands upon existing methods by not only adopting OCT for the predictive maintenance of bridges but also incorporating real-world data from a physical bridge with observed local cracks - damage scenarios that reflect actual conditions. This enhances the robustness and applicability of OCT to more realistic and complex cases. Additionally, unlike prior research that has primarily focused on OCTs for individual parameters, this study investigates multi-classification trees and the interdependencies among different parameters. Using a case study, the challenges and possibilities of the method with regard to the predictive maintenance of bridges are examined.