Rapid prediction method of earthquake damage to masonry structures based on machine learning

Masonry structures are one of the most vulnerable to severe and extensive damage in terms of previous earthquakes. It is significant to quickly evaluate the seismic damage levels of masonry structures, to reduce casualties and economic losses caused by earthquakes. However, traditional methods based on manual judgment or finite element simulations tend to be relatively slower . In this paper, a machine learning-based rapid prediction method was proposed for assessing the seismic damage of masonry structures. By analysis of building data from several cities and combining ground motion with structural characteristics, 11 impact factors were identified as input variables. The LM-BP neural network model was developed by a backpropagation (BP) neural network with strong nonlinear modeling capabilities, and by the Levenberg-Marquardt (LM) algorithm. The accuracy and stability of the model were verified by comparing the predicted values with actual earthquake examples. The results show that the selected seismic damage impact factors can accurately reflect the structural damage level. By comparing methods using parameters on either the structure or ground motion, the predictive accuracy of the proposed method is significantly enhanced. It provides a basis for post-earthquake structural safety assessments and disaster prevention and mitigation work.