Detection of Spark Erosion on Insulated Rail Joints by Deep Learning

Abstract

Insulated Rail Joints (IRJs) are among the critical elements of railway infrastructure, providing structural continuity and electrical isolation within track systems. These joints play a very important role in the detection of trains and in operational safety by segmenting rail networks into electrically isolated sections. Although being a critical part, IRJs suffer from degradation due to high mechanical loads, electrical discharges, and repeated stress. One significant form of localized damage is spark erosion. If left unaddressed, such defects may lead to severe operational failures and safety risks, making effective monitoring systems essential. This paper presents a deep learning-based approach for automatic detection of spark erosion in IRJs. The proposed methodology is based on a two-stage framework: in the first stage, the SqueezeNet convolutional neural network is used for the classification of rail images and the detection of IRJs. In the second stage, a semantic segmentation network is applied to accurately localize and detect spark erosion regions in the detected IRJ images. The study uses two different data sets: one is a high-resolution rail image data set for IRJ classification, and the other labeled images of spark erosion. In addition, advanced image preprocessing and augmentation techniques are adopted to deal with class imbalance and enhance model robustness. Experimental results prove the effectiveness of the proposed method, which achieves 98.1% accuracy in IRJ classification and accurate segmentation of spark erosion areas. These findings underline the potential of two-stage deep learning frameworks to provide real-time, automated, and highly accurate monitoring solutions for railway maintenance. The proposed methodology contributes to enhancing the safety, efficiency, and reliability of railway operations by allowing the early detection of critical defects.