Development of Neural Network Based Hysteretic Models for Steel Beam-Column Connections Through Self-Learning Simulation

Abstract

Beam-column connections are zones of highly complex actions and deformations interaction that often lead to failure under the effect of earthquake ground motion. Modeling of the beam-column connections is important both in understanding the behavior and in design. In this article, a framework for developing a neural network (NN) based steel beam-column connection model through structural testing is proposed. Neural network based inelastic hysteretic model for beam-column connections is combined with a new component based model under self-learning simulation framework. Self-learning simulation has the unique advantage in that it can use structural response to extract material models. Self-learning simulation is based on auto-progressive algorithm that employs the principles of equilibrium and compatibility, and the self-organizing nature of artificial neural network material models. The component based model is an assemblage of rigid body elements and spring elements which represent smeared constitutive behaviors of components; either nonlinear elastic or nonlinear inelastic behavior of components. The component based model is verified by a 3-D finite element analysis. The proposed methodology is illustrated through a self-learning simulation for a welded steel beam-column connection. In addition to presenting the first application of self-learning simulation to steel beam-column connections, a framework is outlined for applying the proposed methodology to other types of connections.