Predicting Rail Corrugation Based on Convolutional Neural Networks Using Vehicle's Acceleration Measurements

• Published in: Sensors (Basel, Switzerland) Vol. 24; no. 14; p. 4627
• Main Authors: Haghbin, Masoud, Chiachío, Juan, Muñoz, Sergio, Escalona Franco, Jose Luis, Guillén, Antonio J, Crespo Marquez, Adolfo, Cantero-Chinchilla, Sergio
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI AG 17.07.2024
• Subjects: convolutional neural networks; Deep learning; Experiments; Fourier transforms; Friction; Grad-CAM; Neural networks; rail corrugation; Research methodology; deep learning; convolutional neural networks; rail corrugation; Grad-CAM
• ISSN: 1424-8220; 1424-8220
• DOI: 10.3390/s24144627

This paper presents a deep learning approach for predicting rail corrugation based on on-board rolling-stock vertical acceleration and forward velocity measurements using One-Dimensional Convolutional Neural Networks (CNN-1D). The model's performance is examined in a 1:10 scale railway system at two different forward velocities. During both the training and test stages, the CNN-1D produced results with mean absolute percentage errors of less than 5% for both forward velocities, confirming its ability to reproduce the corrugation profile based on real-time acceleration and forward velocity measurements. Moreover, by using a Gradient-weighted Class Activation Mapping (Grad-CAM) technique, it is shown that the CNN-1D can distinguish various regions, including the transition from damaged to undamaged regions and one-sided or two-sided corrugated regions, while predicting corrugation. In summary, the results of this study reveal the potential of data-driven techniques such as CNN-1D in predicting rails' corrugation using online data from the dynamics of the rolling-stock, which can lead to more reliable and efficient maintenance and repair of railways.