On the application of variational auto encoders (VAE) for damage detection in rolling element bearings

In structural health monitoring (SHM) and condition monitoring (CM) applications, the expense of testing programmes may be too high to obtain adequate datasets. When limited by the number of available data samples, one may rely on dimensional reduction methods to proceed with a meaningful statistical and probabilistic analysis. In this work, some state-of-the-art dimensionality-reduction techniques were investigated as part of a simple ball-bearing damage detection problem. A variational auto-encoder (VAE) was compared to other methods, based on their capability to generate low-dimensional representations of the data. Unlike other common alternatives, such as principal component analysis (PCA) or auto-encoding (AE) networks, the VAE introduces a probabilistic framework via the latent embeddings. A well-defined distribution is thereby constructed on the latent variables, making the transformed dataset an optimal one for subsequent pattern recognition analysis. The results demonstrated an increase in classification performance given the low-dimensional representation generated by the VAE.