Gaussian Process Regression-Based Structural Response Model and Its Application to Regional Damage Assessment

• Published in: ISPRS international journal of geo-information Vol. 10; no. 9; p. 574
• Main Authors: Park, Sangki, Jung, Kichul
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.09.2021
• Subjects: Buildings; Construction; Correlation analysis; Damage assessment; Disasters; Displacement; Earthquake damage; Earthquakes; Fragility; fragility curve; Gaussian process; Gaussian process regression; geophysics; Hazard assessment; Investigations; Learning algorithms; Machine learning; maximum displacement; model validation; Neural networks; Normal distribution; Normalizing; P-waves; Parameters; Performance enhancement; Performance evaluation; prediction; Property damage; Regional analysis; regional seismic damage assessment; Regression analysis; Regression models; Risk assessment; Sampling; Seismic activity; Seismic hazard; Seismic response; Seismic waves; Sensitivity analysis; South Korea; spatial data; Structural damage; Structural models; Structural response; Variables; Korean Peninsula; South Korea
• ISSN: 2220-9964; 2220-9964
• DOI: 10.3390/ijgi10090574

Seismic activities are serious disasters that induce natural hazards resulting in an incalculable amount of damage to properties and millions of deaths. Typically, seismic risk assessment can be performed by means of structural damage information computed based on the maximum displacement of the structure. In this study, machine learning models based on GPR are developed in order to estimate the maximum displacement of the structures from seismic activities and then used to construct fragility curves as an application. During construction of the models, 13 features of seismic waves are considered, and six wave features are selected to establish the seismic models with the correlation analysis normalizing the variables with the peak ground acceleration. Two models for six-floor and 13-floor buildings are developed, and a sensitivity analysis is performed to identify the relationship between prediction accuracy and sampling size. A 10-fold cross-validation method is used to evaluate the model performance, using the R-squared, root mean squared error, Nash criterion, and mean bias. Results of the six-parameter-based model apparently indicate a similar performance to that of the 13-parameter-based model for the two types of buildings. The model for the six-floor building affords a steadily enhanced performance by increasing the sampling size, while the model for the 13-floor building shows a significantly improved performance with a sampling size of over 200. The results indicate that the heighted structure requires a larger sampling size because it has more degrees of freedom that can influence the model performance. Finally, the proposed models are successfully constructed to estimate the maximum displacement, and applied to obtain fragility curves with various performance levels. Then, the regional seismic damage is assessed in Gyeonjgu city of South Korea as an application of the developed models. The damage assessment with the fragility curve provides the structural response from the seismic activities, which can assist in minimizing damage.