Approaches to the rapid seismic damage prediction of r/c buildings using artificial neural networks

The present paper deals with the investigation of the ability of Artificial Neural Networks (ANN) to reliably predict the r/c buildings' seismic damage state. In this investigation, the problem was formulated as a problem of approximation of an unknown function as well as a pattern recognition problem. In both cases, Multilayer Feedforward Perceptron networks were used. For the creation of the ANNs' training data set, 30 r/c buildings with different structural characteristics, which were subjected to 65 actual ground motions, were selected. These buildings were subjected to Nonlinear Time History Analyses. These analyses led to the calculation of the buildings' damage indices expressed in terms of the Maximum Interstorey Drift Ratio. The influence of several configuration parameters of ANNs to the level of the predictions' reliability was also investigated. In order to investigate the generalization ability of the trained networks, three scenarios were considered. In the framework of these scenarios, the ANNs' seismic damage state predictions were evaluated for buildings subjected to earthquakes, neither of which are included to the training data set. The most significant conclusion of the investigation is that the ANNs can reliably approach the seismic damage state of r/c buildings in real time after an earthquake.