State Assessment of Rail Surface Defects with the Fusion of Images and Vibration Signals

Accurate evaluation of rail surface defects is fundamental to condition-based track maintenance and the safe operation of trains. Although laser-line profile scanners can capture defect geometry, their high cost and the complexity of their data processing pipelines limit their applicability for real time, easily deployable dynamic monitoring of rail surface damage. To address the challenge that spalling assessment requires the simultaneous acquisition of the defect length and depth, this study proposes a multisource assessment method that integrates rail surface images with vibration signals. First, YOLOv5 is employed for object detection on the images, enabling precise localization of defects and extraction of their planar dimensions. Then, a nonlinear regression model based on multi-feature weighting is used to analyze the vibration signals and accurately estimate defect depth. Finally, the planar and depth information are fused to quantify the defect severity. Experiments conducted on two independent data sets demonstrate assessment accuracies of 97.5%, 96.4%, and 99.1% for light, serious, and spalling defects, respectively, validating the effectiveness and engineering potential of the proposed method.