Seismic fragility prediction models for masonry buildings considering integration algorithms and deep learning frameworks

The seismic fragility of masonry buildings have been ubiquitously discussed in multiple zones worldwide. However, field inspection and data collection after earthquakes demonstrate the disadvantages of being time-consuming and labour-intensive, which poses a challenge for quickly obtaining the seismic fragility features of masonry structure clusters. To improve the efficiency of earthquake damage prediction and the rapid estimation of the seismic fragility of masonry structures, a seismic fragility prediction model for masonry buildings involving gradient boosting machine (GBM), categorical boosting (CatBoost), and feedforward neural network (FNN) methods was developed via intelligent ensemble algorithms and deep learning techniques. Using dynamic response analysis and earthquake risk assessment processes, dynamic responsiveness, hazard analysis, and intelligent model development were conducted on 2,598,972 accelerations monitored by the Wenchuan earthquake (May 12, 2008) array group in China. A dataset of 2211 masonry buildings (investigated by the author) in Dujiangyan city was thoroughly analysed, and a fragility matrix considering the optimised seismic vulnerability (damage) grade was proposed. The input variables of the learning model were the construction year, major story, foundation, function, and instrument intensity; moreover, a sensitivity analysis of the characteristic parameters was conducted. The optimised vulnerability (damage) grade of masonry structures was used as the output variable of the integrated and deep learning models. Using the established database, 20% and 80% of the structural damage instances were stratified and selected for testing and training model development, respectively. An analytical (precision-recall) curve was generated to evaluate the correlation between the predicted parameters of a learning model. To confirm the predictive precision of the developed learning model, a set of seismic fragility comparison curves for masonry buildings considering intelligent and empirical data-driven strategies was generated.