Automated Estimation of Exposed Temperature and Strength Changing Ratio for Fire-Damaged Concrete Using Deep Learning Method

Hacıefendioğlu K. , Akbulut Y. E. , Nayır S. , Başağa H. B. , Altunışık A. C.

Experimental Techniques, 2021 (Journal Indexed in SCI Expanded) identifier identifier

  • Publication Type: Article / Article
  • Publication Date: 2021
  • Doi Number: 10.1007/s40799-021-00503-y
  • Title of Journal : Experimental Techniques
  • Keywords: Deep learning method, Convolutional neural networks, Resnet-50, VGG-19, Xception, Fire, Concrete, CRACK DETECTION


The variation of concrete strength after exposure to high temperatures is one of the critical parameters for structural behavior. Comprehensive analysis with traditional methods is required to determine the fire temperature to which concrete is exposed. This study aims to move away from this long and expensive process to determine the high temperature that concrete is exposed to by considering the surface appearance with a faster, practical, economical method that requires no experience. A deep learning method, convolutional neural network, is used to predict the temperature to which concrete is exposed and the resulting strength reduction due to fire exposure. To achieve higher accuracy, instead of training a model from scratch, a transfer learning technique was used and pre-trained Resnet-50, VGG-19, and Xception models were used to customize and initialize weights. For training with the deep learning method, surface images were obtained after the concrete samples produced in the laboratory were exposed to the initial temperature (about 27 °C) and different fire temperatures. The temperature limit of each test specimen has been set as 200 °C, 400 °C, 600 °C and 800 °C, respectively. A total of 3201 images were used for all temperature classes. An algorithm called Gradient-weighted class activation mapping using the gradients of any target concept, flowing into the final convolutional layer to produce a coarse localization map highlighting important regions in the image for predicting the concept was applied in this study. The analyses with the deep learning models were compared in detail and the most appropriate model was decided. Analysis has shown that the Resnet-50 CNN model has strong potential to handle fire-exposed concretes of different strengths, including those not covered in the training. The results of this comprehensive analysis highlight the value of its use in rapid analysis and decision making in structural fire engineering applications given its exceptional ability to grasp multidimensional events with ease, post-fire emergency response, high predictability and potential for continuous improvement.