Journal of Structural Engineering · 2021 · 108 citations · 58 references
EngineeringCivil EngineeringData ScienceBridge EngineeringEarthquake EngineeringStructural EngineeringBridge Condition AssessmentBridge DesignGaussian Process ModelingKernel MethodsPost-earthquake Damage AssessmentStructural AssessmentBridge Damage AssessmentEarthquakesFragility AnalysisFragility AssessmentGaussian Process RegressionSeismic Damage AssessmentSeismic Fragility AssessmentStructural Damage Classification
The study proposes a uniform design–based Gaussian process regression approach to rapidly assess bridge damage and estimate fragility curves by optimizing training data selection. The method builds a GPR model linking seismic damage to bridge properties and earthquake attributes, uses uniform design to enrich the training set, and validates the model with empirical samples from the 2008 Wenchuan earthquake and other events while examining kernel choice, sample size, and data discrepancies to derive empirical fragility curves. The UD‑GPR approach was shown to be efficient and accurate, matching reference fragility curves and remaining robust even with small training sets, making it valuable for rapid post‑earthquake bridge assessment.
This paper proposes a uniform design (UD)-based Gaussian process regression (GPR) method for rapid damage assessment and fragility estimates of bridges. The core idea of the proposed method is that the GPR model was adopted to establish the relationship between the seismic damage and various parameters including bridge properties and earthquake attributes, and the UD method was implemented to enhance the training data set to improve the performance of the GPR model. An efficient method is proposed to improve the use of the UD method over empirical data to search the optimal training data set. Various empirical samples of damaged bridges from the 2008 Wenchuan earthquake in China were collected to validate the predictive ability of the proposed UD-GPR method. The ability and stability of the UD-GPR method for damage classification were evaluated using samples from other earthquakes that were not included in the training data set. The influence of different kernels for the GPR model, different sample sizes, and different discrepancies of training data sets on the predictions of the UD-GPR model was investigated. The empirical fragility curves for different bridge types were derived based on the results of damage classification. The efficiency and accuracy of the UD-GPR method for generating the fragility curves were validated by the actual fragility curves developed by the existing references. The effects of sample size and discrepancy on the median fragility and fragility dispersion were discussed. It can be concluded that the UD-GPR model is efficient and accurate for damage assessment and fragility analysis of bridges even when the training data set has a small number of samples, which can be helpful and beneficial for rapid postearthquake assessment of bridges.
58