Automatic Estimation of Post-fire Compressive Strength Reduction of Masonry Structures Using Deep Convolutional Neural Network

A deep learning-based image processing study was carried out to predict the post-fire safety of historical masonry structures. For this purpose, andesite stone and lime-based mortar, which are frequently used in historical structures, were selected as test samples and the samples were exposed to high temperatures (200°C, 400°C, 600°C and 800°C) at a heating rate of 2.5°C/min. The compressive strength values of andesite stone and lime mortar at different temperatures were determined by both pulse velocity tests and uniaxial compressive strength tests. Naturally, after the heating processes, chemical and physical changes occurred on the surfaces of stone and mortar samples at every temperature. Thus, the compressive strength values obtained as a result of different temperatures were associated with surface image changes. An image classification method based on deep learning, convolutional neural network, was used to predict the temperature to which materials are exposed and the resulting strength reduction due to fire exposure. Pre-trained models of Resnet-50, VGG-16, VGG-19, Inception-V3 and Xception, which are well-known deep learning approaches, are used to classify objects automatically. The Score-CAM visualization technique was also considered, depending on the deep learning method used to accurately predict the location of the common texture of the material to fire. A portable electronic microscope was utilized to take a large number of images of samples exposed to different temperatures. At the end of the study, the Xception model created by deep learning on the arch model built to scale with andesite stone and lime-based mortar was tested, and the strength loss of the arch model exposed to high temperature was tried to be estimated.