Integrating Surrogate Modeling and Bayesian Updating for Damage Assessment

Besides service loads, extreme loadings from the ambient environment, such as hurricane-induced strong winds and fallen trees, can cause severe damages to infrastructures. Damage assessment is crucial to infrastructure management and mitigation strategy planning. Due to many uncertainties of the civil infrastructure systems, such as those from loads, material properties, and empirical models used to describe the structural damages, systematic probabilistic damage assessment that include many damage conditions and scenarios could be very time-consuming. With detailed modeling of the infrastructure and their interactions with the ambient environment, each damage simulation of the complex structural system could also be computationally expensive. Nevertheless, the establishment of probabilistic and temporal load models can also be challenging because of data scarcity. Finally, the material deterioration in the natural coastal corrosive environment environments can further complex the damage assessment considering the time-dependent nature of the corrosion and degradation process. To address these challenges, data-driven-based surrogate modeling and data-informed dynamic Bayesian network (DBN) are integrated for damage modeling of critical civil infrastructures. The probabilistic damage assessment of two types of structures with material degradation issues is discussed. For the coastal slender bridge, the temporal sequences of vehicle, wind, and wave loads are simulated by Markov Chain Monte Carlo simulations based on the data collected by the in-field data-collecting systems. In a DBN framework, the evolution of the fatigue crack length in orthotropic steel deck (OSD) is evaluated with the facilitation of a surrogate model used to predict the fatigue crack length growth. For the overhead power distribution system (OPDS) composed of wooden poles, the risk model of fallen trees is developed using computer vision technique, and the risk model of hurricane-induced wind is established through a machine learning-based statistical model combined with historical hurricane track data. In the static analysis of OPDS subjected to wind loads, surrogate models are developed to predict the moments of poles using selected features as predictors. In the dynamic analysis of OPDS, a physics-informed long short-term memory algorithm is utilized to predict the dynamic response of poles. The uncertainties of the material properties are reduced within the DBN framework to improve the damage assessment accuracy. The optimal hardening strategy for OPDS is selected through the Bayesian decision network.