A Data-Driven Approach for Predicting Mud Pumping in Railway Tracks Using GIS and In-Service Train Vibration Data

The automatic and timely identification of mud pumping is crucial for maintaining the reliability and safety of railway systems. While current prediction models mainly focus on monitoring the dynamic responses of railway tracks, they often overlook vital geotechnical factors, such as hydrological conditions, due to the difficulty of quantifying such information. This limitation reduces the accuracy of predictions. To address this challenge, we propose a novel approach that integrates Geographic Information System (GIS) technology with in-service train vibration data to quantify hydrological variables along railway tracks. Key factors, including elevation, proximity to rivers, rainfall, sink depth, and soil types, are incorporated into a multi-channel neural network, which processes these multi-attribute data separately to enhance prediction accuracy. To improve model interpretability, we apply a Genetic Algorithm (GA) to assess the importance of hydrological factors and their correlation with the likelihood of mud pumping. Tested on real-world data from Chinese railway tracks, the model achieves balanced accuracy, demonstrating the effectiveness of combining GIS and monitoring data to reduce false positives and enhance prediction precision. Our analysis reveals that rainfall is the most influential factor, with groundwater-related variables having a greater impact than surface water. These findings offer valuable insights for infrastructure managers, enabling the identification of vulnerable track sections and facilitating more targeted maintenance and optimized substructure design at the network level.