Seismic Damage Forecasting for Steel Moment Frames using Neural Networks

In this research a neural network algorithm is used to process information of numerous nonlinear dynamic analyses data so that the damage induced by an earthquake (similar to earthquakes of this study) in a structure can be obtained in an acceptable range of accuracy by spending much less time than the computation time for actual nonlinear analysis of the structure [1]. To this end, about 800 nonlinear dynamic analyses of steel moment frames under Tabas, Chichi and Kobe earthquakes has been conducted spending a very long computation time. Then the damage of structures with five different damage indices has been calculated. To design a neural network 70% of responses were randomly chosen for training data and the remaining 30% were used for the test and verification of the neural network software. The designed neural network is a multiple layer Perceptron network that has a hidden layer and is trained by Error Back-Propagation algorithm. With the aid of this neural network software if a structure (similar in characteristics to the frames of this study) is excited by an earthquake similar to the abovementioned earthquakes, its damage indices with five definitions will be calculated in a few seconds with an acceptable accuracy. In the paper the five damage indices including the damage index that is suggested by the authors are briefly introduced. Finally, the accuracy of the results of neural networks software for the five different damage indices are compared to each other.