A machine learning approach based on multifractal features for crack assessment of reinforced concrete shells

Abstract

The geometric properties and spatial characteristics of crack patterns are significant indicators of the extent of damage on reinforced concrete structures. However, manual visual assessment is subjective and depends highly on the inspector's skills. The current study proposes an automated approach for the quantification of digitally documented crack patterns on reinforced concrete shell elements subjected to reversed cyclic shear loading. Multifractal analysis is proposed as a feature extractor for images depicting crack patterns and a set of artificial cracks is analyzed, to quantify how the properties of crack patterns vary as a function of cracking inclination. The results of the parametric study motivated the training of a multiclass classification model, which is used to provide damage level estimates for cracked reinforced concrete members. The training of the classifier is performed using experimental data of reinforced concrete shell elements under well-defined and idealized two-dimensional pure shear stress loading conditions. A dataset with 119 images from crack patterns of reinforced concrete shells is used for training. The multifractal features successfully translate the shape of the crack patterns into meaningful information about the extent of damage; achieving an overall test accuracy of 89.3%.