ANN-Based Prediction of Earthquake Damage in Steel Frame Structures

Abstract

Rapid, reliable, and precise evaluation of earthquake-induced damage in steel frame structures is essential for ensuring structural safety and aiding post-earthquake decision-making. Although dependable, traditional evaluation techniques such as finite element analysis and experimental testing can be labor-intensive and computationally demanding, which makes them unsuitable for rapid damage assessment. This paper highlights the development of an Artificial Neural Network (ANN) model for classifying seismic damage levels in steel frame structures using a data-driven method. A dataset of 500 numerically generated samples was used, which included important structural and seismic parameters such as building height, number of stories, peak ground acceleration, fundamental period, inter-story drift ratio, steel yield strength, damping ratio, beam depth, column depth, and ground motion duration. The damage index was treated as a categorical output with five damage classes ranging from no damage to collapse. To assess the performance of the model, subsets of the dataset were divided into training, testing, and validation. Supervised learning implemented using Orange Data Mining Software was employed to train the ANN model, and classification performance metrics such as Area Under Characteristic Curve (AUC), Matthews Correlation Coefficient (MCC), and F1-score were used to evaluate the model. The optimal ANN identified in this study had two hidden layers with 18 and 9 neurons (10–18–9–1). Testing and validation results showed F1-scores of 0.922 and 0.934, AUCs of 0.964 and 0.972, and MCCs of 0.937 and 0.949. This architecture was used for damage prediction and classification, demonstrating that ANNs are a fast and reliable tool for assessing seismic damage in steel frame structures.