Structured-Light 3-D Sensing for Quantitative Bolt Looseness Detection in Railway Elastic-Clip Fasteners

To address the lack of depth measurement in 2-D images, the sensitivity of conventional point-cloud geometric methods to complex field conditions and height-reference drift, and the high computational and annotation costs of deep learning methods, this article develops a lightweight, structure-prior-based 3-D geometric measurement framework for quantitative bolt-looseness assessment of Type-I elastic-clip fasteners. First, rail-region localization is performed using a top-view height map and strip-consistency constraints, and a fastener candidate search band is constructed from the detected rail boundaries. Then, a robust height profile and a gradient-energy envelope are combined with length, amplitude, and inner–outer consistency constraints to localize fastener regions. Within each region, 1-D height-distribution analysis, gap-tolerant connected-component extraction, and sector-wise high-percentile radial estimation are used to separate and identify the nut and screw top surfaces with incomplete and noisy point clouds. Finally, the unworn outer rail-head reference band is selected as the height datum, and the axial nut displacement is converted into the elastic-clip gap for looseness-state determination. Experiments show that the measurement unit achieves a mean absolute error of 0.024437 mm. The method was validated on 3108 Type-I elastic-clip fasteners installed on 60-kg/m rail, correctly localizing 3032 samples with a mean intersection over union (Mean IoU) of 0.933. On an independent 60-sample set, the elastic-clip gap measurement achieves a mean absolute error of 0.128 mm and a root-mean-square error (RMSE) of 0.160 mm, while the F1 -scores for the loose and severely loose states reach 91.1% and 93.3%, respectively.