Durability Prediction And Assessment Using Deep Learning:A Comprehensive Review

Abstract

In recent decades, the durability of concrete has become a significant area of research and continues to be a key concern in the construction field. Issues such as cracking and spalling are widespread and often stem from environmental influences, substandard construction practices, insufficient oversight, design flaws, and other contributing factors. This paper explores the latest developments in concrete durability research, addressing common issues like alkali-aggregate reactions, sulfate attacks, corrosion of steel reinforcement, and freeze-thaw cycles. These problems can lead to structural degradation or a loss of strength within just a few years. Accurately identifying the location and size of cracks can also be quite difficult. Recent advancements in deep learning have introduced highly accurate methods for crack detection. In this study, over 60 research papers published in top-tier journals and conferences within the past three years were collected through a systematic literature review. These studies were then categorized into 10 key topics based on the accuracy of their crack prediction results: trial-and-error methods, Transfer Learning (TL), Encoder-Decoder (ED), Generative Adversarial Networks (GAN), YOLO V5, LeNet-5, Mask R-CNN, Artificial Neural Networks (ANN), Support Vector Machines (SVM), Binarization, YOLO V3, 3D-SM, IPZ, and VGG-16. This survey aims to analyze the strengths and weaknesses of the models within each category, with a particular focus on the latest advancements in Convolutional Neural Networks (CNNs) and YOLO V5. Third, the study identifies the commonly used evaluation metrics and loss functions applied to CNN and YOLO V5 datasets. Finally, it examines several recurring challenges in the fields of CNN and YOLO V5, analyzes existing solutions, and offers recommendations for future research directions.