Anomaly Detection in Rails Using Dimensionality Reduction

Abstract

As railroad rails are an important social infrastructure, monetary and human resources are spent on their maintenance and inspection for people’s safety and security. In particular, rails are uniformly ground periodically to reduce noise and vibration during railroad operations and avoid damage to the rail tops. However, due to cost reduction and lack of human resources, the damage mechanism of rails is being elucidated for more efficient grinding. Factors such as passing tonnage, transit speed, and weather conditions affect rail damage; however, the details of the damage mechanism are not clear. In this study, we measured a large amount of residual stresses and the full width at half maximum (FWHM) of the diffraction rings using a high-speed X-ray residual stress measurement system and explored the possibility of detecting abnormal areas of rails and diagnosing the signs of damage using statistical analysis methods and machine learning. For verification, 2D mapping measurements of the x-axis component of the vertical stress σx, shear stresses τxy, τxz, and τyz, and FWHM of the diffraction ring at the head of a cracked rail were used. The data were subjected to dimensionality reduction by principal component analysis, kernel principal component analysis, and an autoencoder. These normal models were built using the data of the normal areas without cracks. The anomaly score was defined as the differences from the normal models, and the detection accuracies of the models were compared using the area under the receiver operating characteristic curve; the autoencoder showed the best performance.