A study on crack characterization, damage severity, and propagation prediction for reinforced concrete beams using deep learning and machine learning techniques

Abstract

Cracks in reinforced concrete (RC) beams pose serious implications on the structural safety, durability, and life cycle, making the detection and assessment of such defect’s indispensable for the purpose of structural health monitoring. The paper presents a unified deep learning approach for the purposes of assessing the crack severity, cracking types and predicting crack progression using the experimentally created image database. The RC beams specimens (100 × 100 × 500 mm) with GGBS (ground granulated blast furnace slag) and steel slag were subjected to progressive loading until failure to create images of flexural, shear and torsional cracks. The process of image processing was used to extract the properties of the cracks and to assess the crack severity index (SI). ResNet18, VGG16, MobileNet, and DenseNet are the four transfer learning models that were used for crack type classification, and ResNet18 was found to be the most efficient model for classification with an accuracy of 0.923 ± 0.006, which is better than all other pretrained models. For crack growth prediction, three algorithms are employed, including Random Forest Regression (RFR), Support Vector Regression (SVR), and Gradient Boosting Regression (GBR). Random Forest Regression (RFR) is proven to be the best one among all the three regression algorithms with MAE of 0.017 mm and R² score of 0.9988, better than SVR and GBR. The proposed system provides a comprehensive solution by integrating crack detection, severity assessment, classification, and growth prediction.