Out-of-roundness damage wheel idenficaon in railway vehicles using autoencoder models

Abstract

The present study investigates the use of unsupervised Machine Learning (ML) techniques applied to the Structural Health Monitoring (SHM) of railway wheels with different stages of polygonalization in the running band. The proposed methodology is based on a comparative study of different AutoEncoder (AE) architectures, including the Variational AutoEncoder (VAE), the Sparse AutoEncoder (SAE), and the Convolutional AutoEncoder (CAE). The models are trained using vertical acceleration data from a virtual track monitoring system, and their performance is evaluated in distinguishing between normal and abnormal structural conditions, as well as in the processing time of input data. The integration of AE models with Hotelling’s T² control charts is explored as a strategy to enhance anomaly detection. The study highlights the potential of these approaches for predictive maintenance applications, contributing to greater efficiency and safety in railway operations