Predicting the building damage due to earthquakes has persistently been a challenge. Various ground motion Intensity Measures (IMs) through correlation analysis with Engineering Demand Parameters (EDPs) have long been employed for structural response prediction. However, considering several Machine Learning (ML) techniques and a comprehensive set of scalar and vector-valued IMs has not been entirely undertaken, especially in the context of special concentrically braced frames. In addressing these limitations, this study employs a blending ensemble learning method along with new IMs extracted in the time-frequency domain and three EDPs, including peak story drift ratio (PSDR), peak floor acceleration (PFA), and peak floor velocity (PFV). Time history analysis is performed using a set of 1126 far-field ground motions and 57 IMs extracted from ground motions in the time domain, frequency domain, time-frequency domain, and spectral domain. Also, some special IMs are proposed to capture ground motion's time and frequency nonstationary characteristics. Four base models—Random forest, Bayesian ridge regression, classification and regression tree, and K-Nearest neighbors—have been used to build a blended ensemble model. The findings showed that the XGBoost meta-model better predicted PSDR, PFA, and PFV, obtaining a coefficient of determination (R2) of 0.901, 0.962, and 0.961, respectively. A prediction accuracy greater than 90 % was achieved using the proposed vector-valued IM (including the mean of instantaneous amplitude, the area under the square of acceleration, and peak ground acceleration) for PSDR (R2 = 90.02 %), PFA (R2 = 94.75 %), and PFV (R2 = 94.41 %). Furthermore, the results showed that the new IMs extracted from the instantaneous amplitude and frequency perform well in predicting EDPs.
Read full abstract