Rapid seismic fragility curves assessment of eccentrically braced frames through an output-only nonmodel-based procedure and machine learning techniques

Abstract

A nonmodel and output-only-based framework incorporating an Auto-Regressive model and roof absolute acceleration response are hired in this study to compute a robust wavelet-based damage-sensitive feature (rDSF) and estimate engineering demand parameters (EDPs) of buildings with eccentrically braced frames (EBFs). A large database is recruited to estimate the peak and residual story drift ratios and peak floor absolute accelerations as EDPs. The database comprises cmorfb-fc wavelet-based rDSF, number of stories, building heights, and EDPs of the 4- and 8-story EBF archetypes, resulting from 44 far-field ground motions through incremental dynamic analysis. The capability of both the closed-form equations and machine learning (ML) techniques is investigated to estimate the EDPs. The predictive equations are constructed using ordinary least squares linear regression, Symbolic and Bayesian regressions. To have a more accurate estimation of EDPs, 12 renowned ML-based regression algorithms are also implemented for training, validation, and test datasets. Sensitivity analysis using box-whisker plots on the validation dataset shows that Extreme Gradient Boosted Trees (ExGBT) ensemble has a remarkable efficiency to predict the EDPs and leads to the highest 10-fold correlation coefficient and the lowest root-mean-square error. Moreover, ExGBT has the lowest median absolute relative deviation in different damage states and consequently is selected as the best surrogate algorithm. The predicted EDPs can be further exploited to map rDSF-based fragility curves, a conditional probability at each defined damage state given the value of rDSF. The comparative results for the test data approve that the predicted fragility curves approximately match with those obtained from observed data and can be used for early decision-making in emergency operations.