Detecting Railway Track Irregularities Using Conformal Prediction

Andreas Plesner, Allan P. Engsig-Karup, Hans True

Abstract: This study addresses the challenge of assessing railway track irregularities using convolutional neural networks (CNNs) and conformal prediction techniques. Using high-fidelity sensor data from high-speed trains, the study proposes a CNN model that outperforms state-of-the-art results, achieving a mean unsigned error of 0.31 mm on the test set. Incorporating conformal prediction with the CV-minmax method, the model delivers prediction intervals with 97.18% coverage, averaging 2.33 mm in width, ensuring reliable uncertainty estimation. The model also exhibits impressive computational efficiency, processing data at a rate suitable for real-time applications, with the capacity to evaluate over 2,000 km of track data per hour. These advances demonstrate the potential of the model for practical implementation in continuous monitoring systems, providing a contribution to the field of predictive maintenance within the railway industry.