Parameter sensitivity of the earthquake damage prediction model for an RC frame structure

Abstract

In this paper, to resolve the problem of the redundant and inconvenient selection of the structural characteristic parameters for the seismic vulnerability assessment of reinforced concrete frame structures by machine learning algorithms, a neural network-based sensitivity analysis method is employed to investigate the effect of different input parameters on structural damage indicators at the structural and component levels, using two well-trained neural network models. The parameters used for sensitivity analysis include five geometric parameters (the number of structural layers, the height of the standard layer, X-direction span, the number of X-direction spans, and the number of Y-direction spans), two design parameters (seismic intensity and site category), and one ground motion parameter (peak ground acceleration). The results show that the plane geometry parameters of a structure are less sensitive to seismic damage indicators at the levels of structure and components. After the less sensitive parameters are excluded, the prediction accuracy for important seismic damage indicators remains high, providing a simpler parameter input basis for the seismic damage prediction of an RC frame structure.