This study investigated the structural response of a 24-year-old， 160 m span concrete-filled steel tubular （CFST） arch bridge after the fracture of a single tie rod. By comparing monitoring data before and after the tie rod fracture， this study systematically analyzed the variation characteristics of key mechanical indicators， including arch foot displacement， suspender cable force， and the alignment of the arch ribs and the bridge deck. Furthermore， it evaluated the influence pattern of the tie rod fracture on the bridge structure. The results showed that after the fracture of a single tie rod， the arch feet underwent outward displacement， and both the arch ribs and the bridge deck experienced downward deflection， with the deformation values on the fractured side being significantly larger than those on the non-fractured side. The suspender cable force showed no obvious change. The deformations of the arch ribs and the deck exhibited an asymmetric distribution： the vertical deflection of the free-end half-span on the fractured side was larger than that of the fixed-end half-span， while the opposite was true for the non-fractured side. Furthermore， a full-bridge finite element model was established to simulate the structural response of the bridge induced by the tie rod fracture， and the calculated results reproduced the monitoring data of the actual bridge well. Based on this model， a dynamic amplification factor was introduced to evaluate the stress state of the remaining tie rods and the arch ribs upon the sudden fracture of multiple tie rods.

Efficient and reliable structural health monitoring is essential for ensuring the safety and extending the service life of steel structures. Owing to the advantages of non-contact nature， high efficiency， and a high degree of automation， computer vision （CV） has gradually become an important technology for the inspection and maintenance of steel structures. Focusing on surface cracks and corrosion damage of steel structures， this review systematically summarizes the recent research progress in CV-based damage detection and outlines the major approaches， including image classification， object detection， and image segmentation. Particular attention is paid to key optimization strategies for small object detection， robustness under complex backgrounds， few-shot learning， and on-site deployment. Existing studies indicate that CV has significantly improved the automation， intelligence， and precision of damage detection for steel structures. However， further advances are still required in dataset standardization， model robustness to interference， generalization capability across scenarios， and lightweight real-time inference.

To address the low efficiency， high risk， and limited quantitative capability of manual inspection for fatigue cracks in steel box girders， a study was conducted on the force analysis and motion control of a magnetic wall-climbing robot for crack inspection. According to the crack inspection requirements for deck plates and diaphragms， a wall-climbing robot equipped with an eddy current testing device was designed， and key parameters including overall dimensions， payload capacity， and operating speed were determined. Static and dynamic models of the robot were established to analyze the minimum magnetic adhesion force and driving torque required on steel plate surfaces. Based on the Webots platform， simulations were carried out to investigate the robot’s motion under different payload and weld obstacle conditions. The results showed that the robot could move continuously between the diaphragm and deck plate. As the payload increased， the start-up time on the diaphragm became longer and speed fluctuation became more pronounced， while the motion on the deck plate remained relatively stable. When crossing a weld， short-term speed fluctuations occurred， and the pitch angle increased significantly with the rising payload. Overall， the robot still maintained good obstacle-crossing capability and motion stability.

A vision-based detection method for bolt missing and loosening using the optimal hexagon rule is proposed. High-quality images are obtained through preprocessing operations such as perspective transformation and distortion correction. Based on the optimal hexagon rule， combined with a regional attention mechanism and a residual efficient layer aggregation network， a keypoint detection algorithm based on YOLOv26 is developed to precisely locate the six vertex coordinates of the head of standard hexagonal-head bolts. The Graham scan algorithm is adopted to obtain the convex hull of the circumcircle of the six bolt vertices. Taking the centroid of the convex hull as the center， the sum of deviations between the actual and estimated values of the six vertices is minimized. The Broyden-Fletcher-Goldfarb-Shanno （BFGS） quasi-Newton optimization algorithm is used to adjust the vertex coordinates and generate optimized regular hexagon vertex coordinates. By comparing the coordinate states of bolts， the rotation angle of the bolt can be calculated. Bolt joint rotation tests using 16-hole three-color bolt heads （gray， blue， and red） showed that within the loosening angle range of 0–60°， the absolute error of the identified bolt loosening angle ranged from -2° to 2°， with a maximum relative error of 8%， demonstrating that the proposed optimal hexagon method provided high detection accuracy and stability.

To enhance the intelligent supervision of photovoltaic （PV） project progress and address the issue of low recognition accuracy caused by component occlusion in complex scenarios， this study proposed an automated recognition approach for key PV components that integrates UAV images with an improved object detection algorithm. In response to common challenges such as small-scale targets and occlusion interference in UAV images， this study developed an optimized non-maximum suppression mechanism and a dynamic screening strategy based on target size and category features. Experimental results showed that the improved model achieved stable convergence of the loss function during training. Its key performance indicators， including detection precision， recall， mAP50， and mAP50-95， reached 94.8%， 93.2%， 94.8%， and 96.5%， respectively. In the practical application of the Anduo PV project in Nagqu， Tibet， the average recognition accuracy for core components such as pile foundations， PV supports， and PV modules exceeded 95%， significantly outperforming traditional manual inspection methods. These findings demonstrated that the proposed approach effectively reduces target omission under occlusion through dynamic detection optimization， providing a feasible technical solution for intelligent progress monitoring in complex PV construction environments and possessing considerable practical engineering value.

Local deformation is a common damage to steel structures， and 3D reconstruction is an important approach to evaluate the bearing capacity of locally deformed angle steels. To accurately evaluate the accuracy of the SfM （Structure from Motion）-MVS （Multi-View Stereo） 3D reconstruction algorithm in establishing point cloud models of locally deformed steel components， this study proposed an accuracy evaluation method for angle steel point cloud models based on spatial Euclidean distance. Four types of locally deformed angle steels with two thicknesses were selected as the research objects， and an accuracy verification test was carried out on the digital image models of locally deformed steel components. The iterative closest point algorithm was used to align the test point cloud with the reference point cloud in space， and the accuracy of the point cloud model was quantified according to the Euclidean distance between corresponding points， so as to verify the accuracy of the angle steel point cloud model from a 3D perspective. The results showed that through the accuracy evaluation of the angle steel point cloud model， in the height and width directions， the points within the allowable range of dimensional deviation in the 3D point cloud models of the four types of locally deformed angle steels accounted for approximately 98% of the total number of point clouds in the models； in the thickness direction， the average error of the 5-mm-thick angle steel point cloud model was approximately 1.09 mm， while the average error of the 10-mm-thick angle steel point cloud model was approximately 1.02 mm； in the local deformation area， the average error of the four angle steel point cloud models was approximately 1.00 mm.