Determination of some geometric parameters used in rapid seismic evaluation of existing reinforced concrete buildings with pre-trained convolutional neural networks

In our country, which is located in a seismic region, it is extremely important that the risk analyzes of existing buildings have been conducted and necessary precautions have been taken before possible earthquakes. Accordingly located in the city for a rapid seismic analysis of the existing residential buildings before the earthquake in Turkey, hundreds of thousands of buildings are required to be examined. Despite the rapid seismic assessment methods developed, determining the earthquake risk priority order of buildings requires a considerable cost and time due to the large number of structural inventories. In addition, the huge number of buildings to be examined will result in the appointment of a large number of technical experts with sufficient equipment. There may also be differences in the surveys made by the technical experts, depending on the experience of the expert. In addition to these, it is also an important problem that earthquake risk analysis is different from each other and depends on quite a lot of parameters. It is very important to make the evaluation as standard as possible in the decisions made on a vital issue such as earthquake risk analysis, and to take the decisions in a faster and measurable way. In this context, it is very important to reach accurate results as much as possible by making use of smart systems in performing earthquake risk analyzes of existing buildings, which is a critical issue for Turkey. Especially in recent years, smart systems have started to be used frequently in the solution of complex engineering problems related to multiple parameters. Based on this motivation, in this thesis; analyzes were carried out and analysis results were compared by making arrangements in some layers of Pre-Trained Convolutional Neural Networks structures (DarkNet-53, EfficientNet, Inception ResNetV2, NasNet Large, ResNet-101, ShuffleNet, SqueezeNet, VGG-19, Xception, ResNet-50), which can predict heavy overhang, adjacent regularity and sloping land presence, which are the parameters to be used in risk prioritization and will assist the expert engineer (or facilitate decision-making in the absence of an expert) in the pre-earthquake risk analysis of residential type reinforced concrete buildings. When the analysis results obtained in this context are evaluated; It has been observed that Pre-Trained CNN architectures can detect these parameters very quickly, with an accuracy rate of up to 97% according to the binary classification strategy and up to 79% according to the multi-class classification strategies, on the building facade images. In addition, it has been seen that the Pre-Trained CNN architectures used in the thesis study can be used for existing reinforced concrete buildings in terms of determining the selected geometric parameters.