Abstract
Abstract
Maintaining the integrity of bolted joints is important in structural and aerospace applications, where loosening under dynamic loads can compromise safety. Manual torque checks require operator intervention, while sensor-based monitoring may require additional installation and calibration. This study presents a dual camera vision framework that uses YOLOv11-pose to detect semantic keypointson the bolt head and nut. Homography-based planar rectification supports relative-angle measurement, which is combined with a Turn-of-Nut model and torque-based initialization to estimate preload. Tests on a single instrumented M20 bolt–nut assembly yielded a mean error of 0.9◦ and a maximum error of 2.4◦ in the relative-angle measurement. The reported preload mean absoluteerror was 0.63 kN, with a maximum error of 1.8 kN under the baseline test conditions. Reduced lighting increased the preload error, indicating that marker visibility affects measurement performance. These results demonstrate the potential of combining vision based pose estimation with a mechanical preload model for non-contact bolt monitoring. Measurement uncertainty and validation across additional assemblies and operating conditions remain important considerations before field deployment.