Autonomous Rail Surface Defect Identification Based on an Improved One-Stage Object Detection Algorithm

Abstract

The rail is an important component of track infrastructure, which withstands repeated wheel loading directly, and its condition is related to the safety of train operation. Thus, accurately identifying the size and location of surface defects in rails helps to optimize maintenance strategies, including adjusting regular monitoring and conducting timely repairs. This approach not only mitigates risks but also enhances work efficiency, which has real economic value and brings safety guarantees. This paper aims to build a data-driven model for rail surface defect identification using photos taken in real lines. Two modules, multidirection rectangular convolution (MRC) and cross-scale (CS) feature extraction, are proposed. The results indicate that the detection and classification of multiple rail surface defects can be automated simultaneously with greater accuracy. Among the defects, spalling sees the most significant boost, and its average detection precision increases from 44.1% to 67%. Moreover, the accuracy for detecting a bright contact band and corrugation exceeds 90%, with 0.995 and 0.915, respectively. Compared with the original You Only Look Once algorithm version 8, the mean of average precision (mAP) of the improved network increases from 85.3% to 88.1% when both models are trained for 300 epochs. Additionally, the precise location and size information of the rail surface defects are obtained through postprocessing, providing support for further intelligent track maintenance.