Damage localization using modal submatrix scaling factor and neural network

Abstract

Damage identification and localization is a major issue in aging civil engineering structures, even more so when the structures involved are made of reinforced concrete. Mathematical models that show solid results and good performance in laboratory are often unreliable on real structures because of the additional complexity and additional uncertainties involved. In this study, the existing submatrix scaling factors model for damage localization is modified and enhanced. A submatrix scaling factor identifies the damage within a structure at element level through the element stiffness matrix. The proposed modal submatrix scaling factor identifies the damage at sub element level. It determines not only which element is damaged but also which mode is affected: Axial load stiffness, bending moment and shear stiffness, torsional stiffness. This gives useful insight about the damage; it can identify if a crack is involved rather than a loss of pretension. The new method requires a larger and deeper neural network framework, deep learning, that was prohibitively expensive up to now. Nowadays however, the rise of GPGPU parallel computing makes the method affordable. The neural network is the center piece of a larger framework that includes: Structural Health Monitoring (SHM) system and a parametric finite element (FE) analysis. The SHM system is installed on a real reinforced concrete bridge and it is used to gather the input data for the neural network. The parametric FE analysis is used to compare the output of the neural network with the real structure and to train the neural network.