Recent Research and Applications in Variational Autoencoders for Industrial Prognosis and Health Management: A Survey

Abstract

Whether in the industrial, medical, or real-world domains, more and more data are being collected. The common particularity of all these application domains is that a great part of this data is mostly unlabeled. Thus, designing a learning model with a minimum of labeled data represents a major challenge in the coming years. A particular emphasis has recently been put on unsupervised learning methods based on the idea of autoencoding. The objective of these methods is twofold: to reduce the dimensionality of the input space and to reconstruct the original observation from this lower dimensional representation space. The variational form of these autoencoders, called the Variational Autoencoders (VAEs), is particularly successful in almost all application areas. This enthusiasm comes from the fact that VAEs allow to take advantage of the theoretical foundations of the Variational Bayesian methods and the learning capabilities of artificial neural networks. This review paper gives to the PHM community a synthesis of the latest publications in the PHM domain using the VAEs related to four topics: 1) Data-Driven Soft Sensors for missing values and data outliers, 2) reconstruction error for fault detection, 3) resampling approach for imbalanced data generation and minority class and 4) the variational embedding as PHM preprocessing pipelines and data transformations. After a review of the theoretical foundations and some practical tricks to succeed the implementation of the VAEs in industrial applications, the four main topics used to exploit the VAEs in the PHM domain are detailed. Finally, a global view of the research done at the research institute of Hydro-Québec regarding the diagnosis and failure detection of hydro-generators with VAEs are presented.