Large scale neural network for damage detection in civil engineering structures

Abstract

This study presents a new neural network algorithm to detect damage in structural systems. The focus is on civil engineering buildings and infrastructures but it can be applied to other systems, mechanical engineering, and nuclear engineering. The input of the neural network is the autocorrelation coefficient computed from a data set of synchronized acceleration signals, which are recorded by a structural health monitoring system. Modern SHM systems installed on large structures, such as bridges, have hundreds of sensors. Even using a small model order when computing the autocorrelation parameters leads to thousands of coefficients. To handle such a large neural network General Purpose Graphic Processing Unit (GPGPU) technology is used. GPGPU is already used in other fields to handle large neural network, such as: face recognition technology. However, its application to structural engineering is a new field of research. The output is the Submatrix Scaling Factors (SSF) for each structural element at sub element level. The sub element level is also a new feature of this study. The new algorithm identifies with a good accuracy the damaged elements. Instead, while the extension and type of damage is also computed, further experiments are necessary to increase their accuracy.