PSI - Issue 44

Sergio Ruggieri et al. / Procedia Structural Integrity 44 (2023) 1964–1971 Sergio Ruggieri et al./ Structural Integrity Procedia 00 (2022) 000–000

1965

2

1. Introduction The structural behaviour of buildings in prone seismic areas is a topic widely studied by the scientific community within the last years, considering the disasters occurred in the near past in several geographic territories (e.g., the Mediterranean area). The main aim of this considerable interest is, on the one hand, to predict future effects on the existing stock at the occurrence of seismic events and, on the other hand, to propose reliable prioritization plans aimed to risk mitigation. With these goals in mind, risk quantification needs to be carried out through the vulnerability characterization of the entire existing stock, after a preliminary adequate screening of the buildings in the area under investigation. The analysis can be performed by opting for different approaches to provide a mathematical function that relates a seismic intensity with a damage condition, usually quantified in terms of possible losses. As reported in Silva et al. (2019), vulnerability models can be mainly classified as empirical and mechanical methods, where the first ones are obtained by correlating observed damages to a proper seismic event, while the second ones are derived by the correlation of specific analyses on numerical models representing a single building or a class of buildings. However, the success of the entire process is ruled by the quality and quantity of available data and the size of the area under investigation (and the building stock in it). In general, for large-scale vulnerability analysis, the primary input data source can be summarized as reported in Polese et al. (2019). One of the most critical problems in the process of data gathering from these sources regards the possible occurrence of human errors, considering the presence of typo, the low update frequency of census data and the subjectivity of the engineering judgments provided by interview based methods. There is a wide state of the art about data collection methodologies and several freely available web platforms providing information of the above-mentioned types. Instead, this study proposes a methodology to define the seismic vulnerability of existing buildings by using mechanical models developed on the information automatically extracted from a simple photo. In particular, we exploit a new data source from the tool proposed in Ruggieri et al. (2021a) and Cardellicchio et al. (2022), named VULMA , which is the acronym of VULnerability analysis using MAchine-learning . From the use of VULMA (using its first three modules) and the subsequent employment of the transfer learning technique, the available critical features of the photographed building are used for defining geometrically defined mechanical models that can be numerically investigated for vulnerability analysis. There are, of course, a number of unknown geometrical and mechanical parameters which are systematically varied with another automatic tool to account for the possible intra-building variability. The final results consist of a fuse of fragility curves, computed according to the seismicity of the zone where the photo was taken, and which can represent a reliable indication of the vulnerability of the building. The methodology has been tested on two case studies in Southern Italy, for which photographs were available. The results have been further improved by adding information for the investigated buildings, using additional data sources, to reduce uncertainties in the vulnerability identification. 2. State of the art on large scale analysis: input data, methodologies and area to investigate The definition of seismic vulnerability of existing buildings can be performed at different scales of analysis, based on the objectives of the study and on quality and quantity of available data. In general, the accuracy of investigation is classified according to three levels of analysis: 1 st , 2 nd and 3 rd levels. Regarding the second and third level of analysis for large-scale investigations, the success of the vulnerability analysis depends on the available level of information. The primary freely available source are census data, which contain information about all the buildings in an area: number of floors, construction material, year of construction. Additional sources are the Technical Regional Cartography (CTR), which report geo-spatially referenced information on polygons of buildings located in an area (e.g., height, area), and Cadastral maps, which report the contours of buildings. Employing these data, also by using a GIS environment, different simplified vulnerability analyses can be performed, such as, for example, the vulnerability index proposed in Uva et al. (2016) and Leggieri et al. (2022). Input information can be improved by using interviews based methods, in which experts are asked about the most recurrent features characterizing the building stock in a predefined, through a specific survey form. Recently, another way has been followed for extracting structural features of existing buildings, by using the transfer learning base tool proposed in Ruggieri et al. (2021a) and Cardellicchio et al. (2022), as mentioned in the introduction. Available data can be employed for vulnerability analysis according to different methodologies. The most popular approaches are mechanical (Silva et al., 2019, Aiello et al., 2017; Ruggieri