Machine learning-based damage detection of RC wall using graph features of crack patterns

Concrete crack quantification is one of the challenges that has been investigated. In this article, a computer vision method is used to detect and quantify the cracks on a concrete surface. After processing the crack images, cracks are modeled as graphs for feature extraction. To study the proposed method, concrete surface crack images from a reinforced concrete shell under quasi-linear load at each load step are used. Having the graph and mechanical features, a PCA analysis is performed to study the dependency of the features. Using GPC as graph principal components and MPC as mechanical principal components, a linear Pearson correlation analysis is performed on the GPC and MPC data, results of which demonstrates more than 75% consistency. Finding the graph features in inherent relationship with the mechanical features, the paper continues with a machine learning study between the two features. Due to the low in-hand data, two different machine learning algorithms are used for the verification purpose. Results of the linear regression model and leave-one-out model showed a very close accuracy with 1% and 2% error, respectively. All findings attest the novel idea of presenting graph features. Graph features can be interpreted to use as a representative for mechanical features. Moreover, this method provides the opportunity of studying crack from a mathematical and fundamental viewpoint.