Railway Track Fault Detection using AI

Abstract

Railway infrastructure is critical for the safe and efficient transportation of goods and people. However, defects in rail tracks can lead to severe accidents, resulting in significant human and financial losses. This thesis presents the development of an end-to-end railway track defect detection system utilizing advanced deep learning techniques. The proposed system is designed to detect various defects, including rail surface anomalies and component defects such as missing fasteners, bolts, and fishplates. The approach combines a supervised YOLOv8x Segment Model and a self-supervised U-Net model. The YOLOv8x-segment model, trained on a curated dataset, achieved a mean Average Precision (mAP) of 94% and demonstrated its capability for real-time detection with high frames per second of 30 FPS. To overcome the issue of scarcity of labeled rail surface defects dataset, the U-Net model was pre-trained on normal rail images and fine-tuned using a subset of Rail Surface Defect Dataset (RSDDs), resulting in significant improvements in defect segmentation accuracy. The system demonstrated strong performance across various metrics and datasets, providing a reliable tool for enhancing railway safety through automated defect detection. Future work includes expanding detection capabilities to include sleepers and ballast and real-time implementation with live video feeds.