Machine Learning-Adapted Rapid Visual Screening Method For Prioritizing Seismic Risk States of Masonry Structures

Abstract

The majority of earthquake-related losses are associated with fully collapsed buildings. So, the determination of the seismic risk of buildings is essential for building occupants located in active earthquake zones. Unfortunately, the existing techniques employed to assess the risk status of extensive building inventories lack the requisite speed and precision for dependable decision-making. Furthermore, post-catastrophe risk categorizations of structures heavily rely on the expertise of engineering teams. Consequently, the decision on risk distributions of building stocks before and after hazards requires more sustainable and precise methodologies that include other means of technological advancement. In this study, using a database consisting of 4,356 masonry buildings provided by the Ministry of Environment, Urbanization and Climate Change (general directorate of geographical information systems), Engineering Firms and Gazi University, the building properties were determined, and detailed static analyzes were made. Then, for the first time in the literature, a new, fast and accurate seismic evaluation method has been proposed, which is associated with detailed evaluation results of structures with the help of machine learning algorithms. Within the scope of the study, the data set was subjected to data preprocessing methods (Synthetic Minority Oversampling Technique (SMOTE), Backward Feature Elimination and Forward Feature Selection, Feature Importance, and Feature Correlation methods). First, fifteen parameters obtained from detailed seismic analysis results, building drawings and building photographs were selected by applying data preprocessing and reduced to six parameters with the highest success impact. To achieve this, size reduction methods were used and considering some selected parameters from the street walking. In addition, the minority data classes were reproduced synthetically with the Synthetic Minority Oversampling Technique Method (SMOTE) during the training phase, and the success rate for test data was maximized. In this study, nine machine learning algorithms, namely; Logistic Regression, Decision Tree, Random Forest, Multivariate Adaptive Regression Spline, Support Vector Machine, K-Nearest Neighbor, Gradient Boosting Algorithm, Extreme Gradient Boosting Algorithm, LightGBM Algorithm and where all these algorithms work together with Voting Classifier Method are used. The risk layers of the buildings were estimated by creating risk classes according to the ratio of the floor shear force of the risky walls to the total floor shear force (Ve/Vr = RVS) or the damage detection level.. At the end of the study, this vulnerability assessment method that creates the risk layers of existing buildings in the literature and can determine the most dangerous or non-risk buildings class has been proposed. This is important for deciding the starting point of urban transformation and assessing the seismic vulnerability of buildings in different regions. As a result of the analysis of the algorithms in the study, the correct prediction rates of the three-tier risk class (RVS values) for the learning database (i.e., 3,484 buildings) and the test database (i.e., 872 buildings) of the proposed method were determined as approximately 99.19% and 86.58%, respectively. High success rates were also obtained in the estimation of RVS values with two and four layers. The parameter selections of the proposed method in the study were determined in a way that can be obtained from the photographs of the buildings with the Convolutional Neural Network structures. Therefore, without the need for technical personnel and without entering the building, with the automation methods of the structures, after the parameter selection, the estimations of the RVS values using machine learning methods can be made with high accuracy. This process is employed to identify, catalog, and prioritize the buildings at highest risk of sustaining damage in designated regions during an upcoming earthquake.. For this reason, this method is of great importance in order to determine and strengthen Turkey's weak structures and minimize the loss of life and property.