Post-quake structural damage evaluation by neural networks: theory and calibration

Abstract

This paper deals with artificial neural networks (ANN) for post-quake structural damages evaluation. Two groups of governing parameters are considered: the structural components group and the secondary components group. The adequate ANN requests the analysis of: the best combination of the components’ damages which governs the global damage of a building; the neural network parameters, i.e. number of hidden layers and neurons as well as the activation functions. A set of 3614 damaged buildings is extracted from a database collected during a post-quake survey by trained staff (Boumerdes, Algeria: Mw = 6.8; May 21, 2003 earthquake). The comparison between predicted and observed damages shows that the best ANN corresponds to: one hidden layer with a number of neurons equal to the number of main building components (i.e. 4 or 8 components), and the “hyperbolic tangent” function as activation function. For the collected database, the ANN’s predictions and the observed global damages are in accordance for: 70% of the considered buildings when the global damage is assumed to be influenced only by the secondary (non-structural) elements group, 80% of the considered buildings when the global damage is assumed to be influenced only by the structural group, 87% of the considered buildings when the global damage is assumed to be influenced by both building components groups, i.e. structural and secondary components.