Unsupervised Deep Learning methods for Analysing Acoustic Emission data

Abstract

The foundation of track construction in Germany and many other countries is typically based on a gravel superstructure with railway sleepers made of prestressed concrete. Concrete has been approved as a material for sleeper construction for several decades. However, sleepers are exposed to weather conditions and changing temperatures, which can lead to erosion processes. Artificial intelligence methods have been successfully applied to locate cracking events in concrete structures. Typically, supervised learning is used to train deep neural networks on events identified by traditional triangulation methods. In this work, self-supervised methods are applied to raw acoustic emission data to evaluate information without prior labeling. We compare autoencoders that create rich latent space embeddings of the events with results achieved by using Self-Organizing Maps and traditional clustering on vectors created through feature engineering of the signals. This study analyzes data from over 19,000 events with 16 channels of sound recordings obtained in a static three-point bending test with railway sleepers. The high dimensional latent space can be visualized with dimension reduction techniques. This evaluation of the manifold of the embedded vectors indicates that location information is preserved by this embedding process.