Multi-feature driven seismic damage state identification for reinforced concrete shear walls using computer vision and machine learning

In this paper, an image-based methodology using machine learning algorithms is developed for earthquake-induced damage state prediction in rectangular reinforced concrete shear walls. The machine learning models are developed using a database including experimental data points of 285 surface crack maps of damaged reinforced concrete shear walls collected from the literature. Eight different machine learning algorithms are utilized to train the damage-level classification models. The damage levels are defined according to the FEMA P-58 damage categories. In addition to the structural and geometric properties of the reinforced concrete shear walls with rectangular cross-section, three image-based indices including Succolarity, Lacunarity, and generalized fractal dimensions are measured as input features of the predictive models. Nine groups of features are selected as input for the machine learning algorithms. Using the GridsearchCV function, the hyperparameters resulting in the best algorithmic performance are chosen from a set of possible parameters. A five-fold cross-validation technique is applied to evaluate the models by resampling procedure. According to the results, the predictive model that uses the Extreme Gradient Boosting (XGB) algorithm with inputs that include both structural parameters and image indices performs best in terms of both overfitting prevention and classification accuracy. The outcomes of the damage state identification can be employed for safety assessment of the reinforced concrete buildings as well as repair/demolish decision-making after an earthquake.