Detecting Railway Track Irregularities with Data-driven Uncertainty Quantification

This study addresses the critical challenge of assessing railway track irregularities using advanced machine learning techniques, specifically convolutional neural networks (CNNs) and conformal prediction. Leveraging high-fidelity sensor data from high-speed trains, we propose a novel CNN model that significantly outperforms state-of-the-art results in predicting track irregularities. Our CNN architecture, optimized through extensive hyperparameter tuning, comprises multiple convolutional layers with batch normalization, Exponential Linear Unit (ELU) activation functions, and dropout regularization. This design enables the model to capture complex spatial and temporal dependencies in the train’s dynamic responses, translating them into accurate predictions of track irregularities. The model achieves a mean unsigned error of 0.31 mm on the test set, surpassing the previous state-of-the-art performance and approaching industry-standard benchmarks for track measurement accuracy. This level of precision is crucial for the early detection of track defects that could compromise safety and ride quality. To quantify uncertainty in the model’s predictions, we implement conformal prediction techniques, specifically the CV+ and CV-minmax methods. These approaches provide prediction intervals with high reliability, achieving a 97.18% coverage rate for the CV-minmax method. The resulting prediction intervals have an average width of 2.33 mm, offering a balance between precision and confidence in the model’s outputs. Notably, our model exhibits impressive computational efficiency, capable of processing over 2000 kilometers of track data per hour. This speed makes it suitable for real-time applications in continuous monitoring systems, potentially revolutionizing the approach to railway maintenance. The integration of CNNs with conformal prediction represents a significant advancement in the field of predictive maintenance for railway infrastructure. By providing both accurate predictions and well-calibrated uncertainty estimates, our approach enables more informed decision-making in track maintenance planning and safety assessments.