Recognition of Nonlinear Hysteretic Behavior by Neural Network using Deep Learning

Abstract

In Japan, large earthquakes have caused many damages to structures. Cracks in the reinforced concrete and plasticization of the steel bar and steel components resulted in increased damage. In the seismic design, the dynamic response analysis is carried out by using a mathematical model that appropriately evaluates the nonlinearity of the material and the components and the seismic performance is confirmed by performing a dynamic response analysis. However, when applying new materials and components, much time and much effort are required to select a mathematical model. In this study, we focused on the high pattern recognition capability of the neural network. We attempt to directly model using a neural network without replacing the complex nonlinear hysteretic behavior using the mathematical model. By improvement of learning data and introduction of deep learning, it was confirmed that the recognition ability for nonlinear hysteretic behavior of neural network was greatly improved, and its applicability as a numerical operation subroutine for time history response analysis was drastically improved.