What question did this study set out to answer?

To assess advanced deep learning frameworks for detecting railway defects and address class imbalance issues.

March 13, 2026Open Access

Multi-Class Rail Defect Detection using Advanced Artificial Intelligence

Key Points

To assess advanced deep learning frameworks for detecting railway defects and address class imbalance issues.
Evaluated YOLOv11, YOLO-NAS, and Roboflow 3.0 models for defect detection.
Utilized a dataset of 7,587 images with extensive annotations for training and testing.
Implemented targeted augmentation techniques to reduce class imbalance significantly.
YOLOv11 achieved the highest mean average precision of 89.8%.
Introduced a tiered augmentation approach reducing dataset imbalance by 88.6%.
Detected rare defects with over 95% average precision for specific categories.

Abstract

Railway infrastructure demands continuous monitoring to ensure operational safety and prevent catastrophic failures. Conventional inspection approaches, including manual visual examination and non-destructive testing, exhibit constraints in scalability, processing speed, and automation capability. This investigation presents a rigorous evaluation of three advanced deep learning frameworks—YOLOv11, YOLO-NAS, and Roboflow 3.0 (Robolow’s default model) on roboflow platform—for detecting multiple rail defect categories, with particular attention to addressing class imbalance challenges. A real-world dataset containing 7,587 images with 35,549 annotations spanning ten defect categories was assembled from North Western Railways, India. The research implements targeted augmentation methodologies and class-weighted training protocols to address severe imbalance (ranging from 0.4% for Broken-Rail to 17.6% for Shelling). Performance assessment across accuracy metrics, computational efficiency, and deployment viability demonstrates YOLOv11 achieving superior mean Average Precision (89.8%), while YOLO-NAS exhibits competitive speed-accuracy trade-offs. Critical examination of cross-regional generalization, temporal robustness, and real-time deployment constraints delivers practical insights for railway authorities implementing AI-based inspection frameworks. • Introduces a tiered augmentation framework that reduces dataset imbalance by 88.6% (from 45.3:1 to 5.15:1). • Evaluates three state-of-the-art models—YOLOv11, YOLO-NAS, and Roboflow 3.0 on Roboflow platform—on 10 railway defect classes. • YOLOv11 achieved the highest mAP (89.8%), while YOLO-NAS offered the best real-time inference speed (34.8 FPS). • Demonstrates robust detection of rare defects (Broken-Rail, Joints, Missing-Fasteners) with ≥95% AP. • Provides a deployment feasibility and ROI analysis for railway inspection systems. • Establishes a reproducible benchmark for AI-based defect detection in safety-critical infrastructure.

Read Full Paperexternally

AI से पूछें

Bookmark

View Full Paper