Parameter-Specific Uncertainty Partitioning in Seismic Fragility Assessment of Highway Bridges Using Heteroskedastic Metamodels

At present, substantial literature exists on vulnerability assessment of bridges involving metamodels that capture homoskedastic uncertainty; however, there is negligible research on heteroskedastic response evaluation as a function of bridge predictor variables. This study presents a novel framework that captures heteroskedasticity in bridge responses using the Gaussian process regression technique and performs parameter-specific uncertainty quantification utilizing statistical linearization. The framework is demonstrated using a case study multispan continuous concrete girder bridge located in the central and southeastern United States. The heteroskedasticity in bridge component seismic responses is first assessed using the Breusch–Pagan test that reveals multiple critical component responses exhibiting heteroskedastic behavior. This heteroskedastic uncertainty is then captured using the Gaussian process regression technique that is further validated with benchmark Monte Carlo trials. Furthermore, the parameterized fragility functions are improved by correcting the class imbalance by adopting the synthetic minority oversampling technique, resulting in confident fragility predictions. Finally, parameter-specific uncertainty quantification is performed for all bridge components and the overall bridge system using field-measurable critical bridge parameters by integrating improved parameterized fragility functions with statistical linearization. The proposed framework helps in reliable fragility predictions and informs bridge asset managers on the most effective monetary investments for measuring critical parameters that affect seismic response of highway bridges.