Fig 1.
Common type of fastener on China’s railway and fastener structure.
(a) Vossloh-300. (b) WJ-8. (c) WJ-7. (d) WJ-2; (e) fastener structure.
Fig 2.
(a) Skewed clip. (b) Bolt missing; (c) Fastener missing.; (d) Over-tight or looseness fastener.
Fig 3.
Flowchart of our method.
Fig 4.
Principle and physical configuration of railway 3D imaging system based on 3D line laser sensor.
(a) Field of view schematic of 3D line laser sensor; (b) Physical configuration of railway fastener 3D imaging system.
Fig 5.
RGB-point cloud bimodal data of railway track.
(a) RGB depth image showing fastener shape and presence; (b) Point cloud rendering showing spatial configuration and depth details. Black pixels in (a) represent invalid points due to sensor occlusions.
Fig 6.
Flowchart of fastener point cloud segmentation based on RGB-P bimodal data mapping.
Fig 7.
The architecture of pointnet++ network and schematic diagram of point cloud segmentation. Segmentation labels (①-⑧): ①: Upper surface of bolt; ②: Metal clip; ③: Insulated block; ④: Upper surface of iron plate; ⑤: Upper surface of gauge block; ⑥: Upper surface of rubber pad; ⑦: Rail edge; ⑧: Upper surface of sleeper.
Fig 8.
Geometric parameters measurement of WJ-8 fastener.
(a) Schematic diagram showing key components; (b) In-situ image of the fastener; (c) Point cloud representation. Component labels (A-K): A: Insulated block; B: Rubber pad; C: Metal clip; D: Gauge block; E: Buffer pad; F: Embedded brush; G: Bolt; H: Flat washer; I: HPuR; J: Iron plate; K: HPuIP.
Fig 9.
The installation and structure of the insulated blocks.
Fig 10.
The 3D line laser sensor scans the standard block for analyzing the measurement accuracy of the 3D imaging system.
(a) Schematic of standard step block; (b) Point cloud of step block; (c) Measurement error of block.
Fig 11.
Detection and localization results of visually normal fasteners using different methods.
(a)-(d) show the detection results of DETR, RT-DETR, YOLOv11s, and YOLOv8s models, respectively.
Table 1.
Comparison of fastener detection performance across different models.
Fig 12.
Comparison of visual defect detection results for defective fasteners.
(a)-(d) show the detection results of DETR, RT-DETR, YOLOv11s, and YOLOv8s models, respectively.
Fig 13.
Segmentation results of individual components in the fastener region point cloud.
(a) IoU values for segmentation of fastener components; (b)–(c) Point cloud of the fastener region and corresponding segmentation results.
Fig 14.
Measurement results for height adjustment pads.
(a) and (b) show HPuIP and HPuR thickness measurements across 300 fasteners, respectively.
Fig 15.
Measurement error analysis for insulated block specifications at varying sampling intervals.
(a) Measurement error distribution across different values. (b) Statistical count of fasteners with measurement errors exceeding the 0.5 mm tolerance.
Fig 16.
Bolt height measurements for different fastener categories.
Fig 17.
Over-loose fastener detection results based on bolt height.
(a) Bolt height distribution of normal and over-loose fasteners; (b) Precision and recall curves under different threshold parameters.