Integrating Deep Learning Into an Energy Framework for Rapid Regional Damage Assessment and Fragility Analysis Under Mainshock-Aftershock Sequences

Abstract

This study introduces an energy-based framework for rapidly assessing damage and fragility in regional buildings under mainshock-aftershock sequences. First, the ultimate energy of the structure is derived from the relationship between two time-varying parameters: effective intrinsic energy and input energy. An energy-based damage index is then defined, with uncertainties of the structure and earthquake quantified through Latin hypercube sampling. Subsequently, a Gaussian Process model, enhanced with K-Means clustering and Bayesian optimization, is employed to predict the structural ultimate energy. A Convolutional Neural Network-Long Short-Term Memory Network with an Attention mechanism model with a weighted loss function is developed to capture the structural energy time-history responses, integrating correlation analysis and hyperparameter optimization. A comparative analysis is performed with previous machine learning models. The framework's effectiveness is validated through a comparative study with the inter-story drift ratio (IDR) index. Finally, the framework is applied to Zeytinburnu in Istanbul, Turkey. The results indicate that reducing the dimensionality of the database through correlation analysis effectively decreases data dimensions while maintaining accuracy. In rapid damage assessment tasks, energy is a superior damage indicator compared to IDR, as it correlates positively with critical parameters such as building height and peak ground acceleration (PGA). It enables a tenfold reduction in response data, enhancing training efficiency by 7.4 times. PGA_as/PGA_ms of 1.0 is recommended for analyzing mainshock-aftershock effects, providing a more comprehensive perspective on structural performance and ensuring a conservative estimate of regional structural fragility.