QUANTIFICATION OF DAMAGE IN INFILLED REINFORCED CONCRETE FRAMES USING MACHINE AND DEEP LEARNING

This paper explores the use of classical and modern machine and deep learning techniques to quantify the damage in the columns of a reinforced concrete (RC) frame infilled with a masonry panel. The study focuses on a large-scale infilled RC frame subjected to in-plane gravity and seismic loads. Photographs of the structure at various damage levels of known inter-story drift ratio are used here to train four classical machine-learning techniques, i.e., Linear Regression, Support Vector Machines (SVM), Regression Trees, and Ensemble Trees, as well as four deep learning algorithms, i.e., Simple Convolutional Neural Networks (CNN), CNN with attention, Deep Learning with engineered predictors and a Hybrid algorithm that combines the engineered features with raw image data in a CNN. Once the algorithms are trained, the error analysis indicates that the best classical machine learning algorithm that applies to this dataset is the SVM. In terms of deep learning algorithms, the most suitable is the hybrid model. Overall, the latter provides the best estimations according to the error metrics. The potential use of these algorithms for assessing structures that deteriorated due to aging, extreme loading events, or both is further discussed.