Machine learning and digital image processing for non-contact modal parameters identification of structures

Abstract

This study introduces an innovative non-contact sensing technique for vision-based displacement measurement. Existing vision-based displacement measurement techniques utilizes physical target panels or physical features to compute relative displacement between the target and the observation point. Instead, the proposed method exploits the optical reference of a speckle pattern. A coherent light that is diffusely reflected on the surface of the target structure creates the speckle pattern. In this study, a camera records the changes in the speckle pattern in real time. Because the speckle pattern is sensitive to small changes of surface, the ambient vibration is enough to affect it. To estimate the displacement of the target from the raw speckle images, speckle contrast imaging (SCI), speckle flow imaging (SFI), and k-means clustering algorithm were used. After SCI and SFI quantifies the blurring effect in each image, the k-means clustering algorithm creates virtual sensing node from each image. The connection of virtual nodes from frame to frame highlights the displacements of the surface in time domain. Because the algorithms are time-consuming and computationally intensive, a GPU executes the entire post-processing operation in parallel and identifies the natural frequencies of the structure.