DeepTrackSecure: an integrated classification-detection system with predictive risk analytics for proactive railway safety management.

Railway track failures such as cracks, displacements, corrosion, and missing fasteners pose serious safety risks and require timely detection and prioritization. Manual inspection techniques are time-consuming, prone to mistakes, and unsuitable for extensive monitoring. This paper suggests DeepTrackSecure, a multi-stage deep learning framework for railway track fault detection and severity-aware risk prediction, to overcome these constraints. The suggested system uses YOLOv5 for accurate fault region localization in defective images after using ResNet50 as an initial screening model to categorize railway track images as defective or non-defective. By combining defect area, detection confidence, and defect frequency, a quantitative severity score is created that goes beyond traditional detection and allows for well-informed risk assessment. Random Forest and XGBoost models are assessed for classifying track conditions into low, medium, and high-risk levels based on this severity representation, exhibiting dependable and consistent decision-making. Ablation experiments to compare YOLO-only detection, classification-detection fusion, and the full severity-based risk prediction model confirm the role of each of the system components. Also, runtime and memory analysis are used to ensure that the performance is near to what is required in edge and cloud deployment and that the robustness is tested in harsh conditions like blur, noise, and low-light conditions. The findings of the experiments indicate that DeepTrackSecure functions as a decision-support and maintenance prioritization tool that produces severity-derived risk proxies (Low/Medium/High) validated against expert opinion. The system is intended to assist maintenance scheduling decisions rather than serve as a directly deployable operational safety system; broader deployment should be preceded by recall improvement and validation against real failure-event ground truth.