A data-driven approach for linking models of large-scale bridges and monitoring data

Abstract

Predictive maintenance of large-scale structures such as bridges requires precise numerical models to describe their current condition. Typically, solving an inverse problem is necessary to determine model parameters from Structural Health Monitoring (SHM) data. However, conventional methods such as Finite Element Updating through optimization algorithms demand substantial computational resources, as numerous parameter combinations need to be assessed to identify the optimal model state in each calibration step. Consequently, these methods are only partially suitable for creating digital twins of bridges.This paper introduces an alternative approach by treating the inverse problem as a model parameter classification problem. This involves establishing a model database that covers a wide range of damage states. Notably, this method eliminates the need for multiple simulations during the application phase, as simulations are performed only once in an offline context. Subsequently, a classification algorithm is trained based on this database, enabling real-time selection of the best-fit model for practical applications using SHM data, without the necessity for additional simulations.Transparency in algorithm decisions is crucial for infrastructure maintenance, therefore, optimal classification trees from the field of interpretable machine learning are employed. Decision trees offer a balance between high accuracy and interpretability while providing additional advantages, such as sensor placement evaluation. In summary, this approach demonstrates the potential for linking numerical models with SHM data through the application of interpretable machine learning techniques, facilitating real-time decision-making for the preservation and management of critical infrastructure.