Automated Prediction Models for the Seismic Vulnerability of Masonry Structures Considering Intelligence and Learning Algorithms

The assessment of the seismic vulnerability of large portfolios of existing structures is the core indicator for developing reliable risk and mitigation plans at regional and urban scales. Masonry structures are widespread in different regions worldwide, and assessing their seismic vulnerability can contribute positively to the definition of large-scale seismic zoning and risk distribution. However, traditional empirical and experimental testing approaches present several drawbacks in practice, such as that they require the analysis of a large set of data with high computational effort. This poses new challenges in terms of quickly predicting the seismic vulnerability of masonry structures by reducing manual estimations in favour of efficient approaches based on machine learning and artificial intelligence. This paper innovatively combines machine learning algorithms with probabilistic seismic hazard models, considering eight characteristic factors affecting the seismic vulnerability of masonry structures, to develop an automated model for predicting the seismic vulnerability of masonry structures. In detail, using artificial intelligence and data-driven technology, data collection and analysis were performed on 2559 masonry structures and 1913,934 acceleration records monitored by 12 seismic stations in Dujiangyan (DJY) city affected by the Wenchuan earthquake in Sichuan (SC) Province, China, on May 12, 2008. Using four developed automated learning models (K-nearest neighbor (KNN), eXtreme Gradient Boosting (XGB), decision tree (DT), and random forest (RF)), confusion matrices and receiver operating curves (ROCs) were defined, with the aim of predicting the seismic vulnerability grades of masonry structures based on different intensity zones. The results of the proposed approaches, in terms of vulnerability curves, were compared with the analogous outputs obtained by adopting existing empirical approaches and applied to the collected seismic damage dataset of masonry structures. A comparison among the four algorithms and empirical models revealed that the random forest algorithm presented the best generalizability and the highest prediction accuracy.