Rail transport is a critical component of production and logistics in coal mines, where safe and efficient operation directly impacts profitability and personnel safety. Traditional inspection methods, including manual patrols and simple sensors, are often inadequate in the complex subterranean environment, which is characterized by variable lighting, high dust concentrations, and humidity. These factors contribute to low efficiency, poor real-time performance, and high rates of missed detections and false alarms. This paper presents a comprehensive review of deep learning applications in rail safety, focusing on tasks such as foreign object intrusion detection and component defect and wear analysis. We systematically examine core algorithms, including the YOLO series for object detection, semantic segmentation networks like U-Net, and the growing use of unsupervised and weakly supervised models to overcome data limitations. The review analyzes specific application cases and technical enhancements, such as model lightweighting for edge deployment and the integration of attention mechanisms to improve accuracy. Furthermore, the paper identifies key challenges, including model adaptability to harsh working conditions and the need for effective few-shot learning strategies. Finally, we discuss future trends, including multi-sensor fusion and edge computing, to provide a reference for the development of robust, intelligent safety assurance systems for rail transport.