An Optical Mode Shape-Based Damage Detection Using Convolutional Neural Networks

Abstract

One of the most widely used methods for structural health monitoring (SHM) is vibration-based, in which changes of the dynamics of a structure are collected and associated to damages. While the change of the natural frequencies may be influenced by many different factors, mode shapes provide a more informative feature set in identifying damages. However, most of the sensing hardware, such as accelerometers, and even some of the noncontact techniques, such as laser Doppler vibrometer, are only able to take measurements from specific locations, leading to a limited spatial resolution of deformation field. The use of camera-based sensing allows the extraction of information at much denser locations on a structure. Among the current techniques for motion sensing, phase-based motion estimation (PME) and phase-based motion magnification (PMM) aim to extract and amplify subtle motions, allowing the extraction of full-field mode shapes with an enhanced visibility. Convolutional neural network (CNN) is applied to metamodel the amplified mode shapes extracted from both undamaged baseline and damaged conditions. A lab-scale testbed is employed to provide data to validate the damage identification approach via the CNN algorithm under different operational conditions.KeywordsStructural health monitoringConvolutional neural networkPhase-based motion magnificationMachine learningOptical measurement