Deep learning-powered framework for rapid snow accumulation detection on railway switches in severe snowstorm conditions

Abstract Railway turnouts in cold regions are frequently impacted by the accumulation of snow during the winter season, which can result in delays to train services and disruptions to the rail network. It is therefore essential to monitor snow conditions in a timely and accurate manner. In order to address this challenge, this study proposes an innovative framework for railway turnout snow detection. This framework incorporates a diffusion model for dataset augmentation and the BCW-Fusion image enhancement technique, with the objective of constructing a comprehensive snow dataset. The EfficientNetV2-S deep learning model, integrated with a Parallel Small Kernel (PSK) attention module, is employed in the research to improve feature extraction and increase the accuracy of snow hazard classification. The experimental results demonstrate that this approach achieves exceptional performance in snow detection tasks, with a detection accuracy of 96.05%. These findings confirm that deep learning models enhanced with data augmentation and optimized algorithms can significantly improve the reliability and real-time performance of railway snow monitoring, thereby demonstrating strong potential for broader application in complex environments.