Rail Surface Defect Detection Based on Spatio-Temporal Sequence

With the continuous increase in the operational mileage and service time of railways, issues such as weld defects, squatting of the tread, and joint damage on the track surface have become critical factors threatening the safety of train operations. Traditional detection methods based on CNN rely on a large amount of image data and have problems such as long training cycles, low utilization of one-dimensional signal features, and weak generalization ability for small samples. To break through these technical bottlenecks, this paper innovatively proposes a dual-branch defect classification architecture that integrates pseudo-spatiotemporal modeling, MobileViT v2, and BiGRU, which achieves efficient identification of railway metal surface defects. This architecture reshapes one-dimensional eddy current signals into two-dimensional pseudo-heat maps, uses MobileViT v2 to extract spatial texture features, and simultaneously captures the temporal dependencies of the signals through BiGRU. After feature fusion, defect classification is achieved. Experimental results show that the classification accuracy of this method for three types of defects reaches 99.6%, and the model parameter size is only 7 M, supporting real-time inference on edge devices, providing a lightweight and high-precision solution for railway defect detection.