Preliminary Design System for Concrete Box Girder Bridges

Abstract

A neural network-based design system is presented in this paper for preliminary design of concrete box girder bridges. The system is based on a loose coupling model that integrates the artificial neural network and the fuzzy network to perform the task of noisy data filtering, knowledge extraction, and candidate synthesis. After a comparative study, the radial basis function neural network is chosen in the design knowledge generation instead of the commonly used back-propagation neural network. The fuzzy network is employed to determine the integer types of design parameters. The developed system provides a few feasible design configurations, and enables the user to overwrite some of the design parameters, so that that user can have a wide choice in his preliminary design. The accuracy of the neural network testing and the influence of the size of the design cases on the neural network prediction are discussed. A design example is included to illustrate the design procedure.