Variational Autoencoder-Based Approach for Rail Defect Identification

Abstract

Nowadays, machine learning has been widely used in all fields of scientific research, and structural health monitoring (SHM) is no exception. However, for structural health monitoring, damage-relevant data is difficult to obtain in many cases. That is to say, an important branch of machine learning, supervised learning, is difficult to be applied for structural damage identification. On the other hand, generative models, a mechanism for unsupervised learning tasks, can be exploited to structure condition assessment, and autoencoder along with variational Bayes is an effective approach for implementation. This investigation aims to formulate an acoustic emission (AE)-based method for rail damage detection by using the properties of latent space under the autoencoding framework. The AE-based detection system is enabled by four piezoelectric ceramic transducer (PZT) sensors symmetrically attached to both sides of the railroad turnout. The collected signals are incorporated into variational autoencoder (VAE) [1] organization as the input, and the latent variables that extract from original information during encoding procedure are operated as the core indicators for the condition assessment of rail track. Signals collected from intact railway turnout and that containing damage information are both used in this investigation and they present different clusters. This result demonstrates the high potential of the latent space under the variational Bayesian framework for identification of rail defects.