Structural health monitoring (SHM) and condition monitoring are crucial processes that ensure reliability and safety of engineering systems in a variety of fields, including aerospace, civil engineering, and industry. These systems are often assessed using vibration-based methods, where damage is detected by analyzing changes in a structure’s vibration characteristics. Traditional methods typically employ contact-type sensors for this purpose. While effective, these methods face several limitations, including low spatial resolution, high costs, difficulties in sensor placement, and measurements that are restricted to small regions around each sensor.
Vision-based methods, where non-contact, full-field vibration measurement is conducted directly from video sequences, have emerged as promising alternatives. In addition to being simple and low cost, these methods offer high spatial resolutions and are suitable for structures with complex geometries or limited accessibility. Full-field motion estimation also enables assessment of the entire structure. However, many existing vision-based approaches struggle with large structural motions, low-texture surfaces, or changes in lighting. Although recently developed phase-based optical flow methods improve robustness by estimating motion from phase information, they still rely on pixel-level data, which, in addition to being computationally intensive and difficult to interpret, is inherently vulnerable to noise, lighting fluctuations, and distortion.
To address these challenges, a research team led by Professor Gyuhae Park from the Department of Mechanical Engineering at Chonnam National University in South Korea, has developed a novel superpixel-based virtual sensor framework. “Our approach utilizes superpixels, clusters of neighboring pixels with similar vibrational and structural behavior, as virtual sensors for motion estimation,” explains Prof. Park. “This creates a virtual sensor grid that can adapt to any structure and offers robust and accurate full-field vibration measurement without the need for physical markers or contact sensors. ” The study was made available online on September 30, 2025, and published in Volume 240 of Mechanical Systems and Signal Processing on November 01, 2025.
The proposed approach operates in three stages. In the first stage, pixel-level motion is estimated from video sequences using the phase nonlinearity-weighted optical flow (PNOF) algorithm, developed by the authors in a previous study. For each pixel, the algorithm extracts local motion from phase information and evaluates the reliability of the estimated displacements in different directions. Unreliable displacement components with high phase nonlinearity are then discarded, and the remaining reliable components are integrated to produce a marker-free full-field displacement map.
In the second stage, the overall confidence of the full displacement at each pixel is calculated, providing a built-in reliability assessment, a first among vision-based vibration measurement methods.
In the third stage, this overall confidence and the full field displacement map are used together to group pixels into superpixels, creating a virtual sensor grid. Depth information is also incorporated to improve alignment between the sensor grid and structure. Finally, full-field displacement is calculated at the sensor level for damage detection.
Experimental validation performed on an air compressor system showed that the proposed method achieves accuracy comparable to that of a laser Doppler vibrometer (LDV) while enabling effective structural damage detection without physical markers or contact sensors. While individual pixels showed some variability, the superpixel-based virtual sensors effectively mitigated these effects.
“Vibration-guided superpixel segmentation enhances both robustness and interpretability of structural diagnostics even in complex environments,” explains Prof. Park. “Our approach makes full-field structural monitoring accessible, low-cost, and deployable using ordinary cameras supporting applications in infrastructure monitoring, aerospace and mechanical equipment diagnostics, smart cities, robotics, and digital twins.”
Overall, this innovative method represents a major advancement for vision-based SHM and may help pave the way for its broader adoption.
About the institute
Chonnam National University (CNU), established in 1952 as Korea’s first national university, is a leading institution of higher learning located in Gwangju and South Jeolla Province. Building on its founding commitment to cultivating leaders of integrity and professional excellence, CNU contributes to national development and global progress through the pursuit of knowledge, ethical responsibility, and inclusive excellence. Guided by the core motto “Truth, Creativity, and Service,” the university advances research, education, and public engagement that strengthen resilient societies, foster sustainable development, and promote the well-being of future generations. As a trusted partner in the global community, CNU remains dedicated to addressing complex challenges in an increasingly interconnected world.
Website: https://global.jnu.ac.kr/jnumain_en.aspx
About the author
Prof. Gyuhae Park is a Professor of Mechanical Engineering at Chonnam National University, contributing to advanced sensing and structural diagnostics research. His research group focuses on smart systems, noise and vibration, AI-driven structural health monitoring, and non-contact, vision-based sensing technologies. He has published more than 500 technical works, including over 130 journal articles, more than 400 conference papers, and 10+ book chapters (Google Scholar citations: 20,550; h-index: 59 as of Nov. 2025). He has served as an Associate Editor for nine top-tier SCI(E) journals and has actively contributed to international conferences as an organizing and scientific committee member. His work has been widely recognized, including being listed in the Stanford–Elsevier Top 2% Scientists list (lifetime and single-year categories) since its inception and his election as a Fellow of the American Society of Mechanical Engineers (ASME) in 2017. He received his Ph.D. in Mechanical Engineering from Virginia Tech in 2000 and his B.S. from Chonnam National University in 1992.
END