Objective Rail fasteners play a vital role in railway infrastructure by securing rails to sleepers and preventing misalignment. Prolonged usage of these fasteners can lead to different types of defects, including visual defects such as missing, fractured, and misplaced fasteners, as well as structural defects like overly loose or tight fasteners. These defects can range from minor issues affecting passenger comfort to serious risks of derailment, posing significant safety concerns for railway operations. The use of two-dimensional visual imaging techniques allows for quick identification of visual fastener defects, while three-dimensional vision sensors capture color and depth images simultaneously. Implementing multi-modal image fusion methods helps mitigate environmental and illumination effects to improve the accuracy of visual defect detection. Three-dimensional structured light imaging aids in accurately capturing the 3D point cloud of the railway track, enabling the detection of structural defects using the fastener's spatial structure. However, further improvements are needed to enhance the accuracy and reliability of structural defect detection. As a result, a new detection approach for structural defects in railway clip fasteners based on 3D line laser sensors is proposed.
Methods Initially, a 3D line laser sensor is employed to capture the point cloud of the railway track. Subsequently, the point cloud corresponding to the fastener area is swiftly identified based on the fastener's height, and the metal clip point cloud is separated from this region using the PointNet++ network. The clip point cloud is then projected onto a 2D image, from which the clip skeleton is derived. This 2D skeleton is then transformed back into the 3D point cloud to establish the initial clip skeleton, with each point being approximated by a circular cross-section to determine the clip skeleton's center representing the clip's outline and spatial arrangement. Following this, feature points of the clip's 3D skeleton are extracted, aiding in the construction of the fastener's pressing plane and calculation of the clip gap to identify structural defects.
Results and Discussions This paper conducts experiments from five aspects to verify the effectiveness of the method: 1) the measurement error of the imaging system ranges between 0.019 mm to 0.054 mm by measuring the standard parts (Fig.12), indicating that the constructed imaging system is capable of accurately capturing the point cloud of track fasteners. 2) The trained PointNet++ network achieves nearly perfect accuracy in segmenting the components of fasteners, thereby providing precise data source for extracting clip skeleton and other point cloud computations (Fig.13). 3) By measuring the clip gap of three types of fasteners, WJ-8, WJ-7, and WJ-2, with different degrees of tightness, the measurement error does not exceed 0.1 mm (Fig.15). Furthermore, showcasing the method's resilience to railway conditions, rust, and contamination on the fasteners (Fig.18-19, Tab.2). 4) For bulk defect detection, with a permissible measurement error of ±0.1 mm, the defect detection accuracy is consistently above 95% (Fig.20-22, Tab.3). 5) Compared with other methods, the proposed method is more precise but is more time-consuming (Tab.5).
Conclusions A visual imaging system has been developed for rail fasteners using 3D line laser sensors. The system accurately captures point cloud data of rail fasteners. The measured clip gap of the fastener using the proposed method shows a measurement error within 0.1 mm when compared with ground-truth data. The proposed approach demonstrates strong resilience against environmental factors such as lighting, rust, and contamination on the fasteners. With a permissible measurement error of ±0.1 mm, the proposed method achieves over 95% accuracy in detecting fastener tightness defects. It is applicable for detecting structural defects in WJ-2, WJ-7, and WJ-8 types of fasteners. The computation time for analyzing a single fastener clip gap is close to 3 s, making the system suitable for offline clip gap analysis and nearly 36 times faster than manual measurement. Future work will involve utilizing point cloud of fastener components segmented with PointNet++ for precise measurements of fastener component, establishing a comprehensive database for "one-pillar-one-file" fastener classification.
DownLoad: