Wind turbine structure health monitoring through zero-shot learning with supervised variational autoencoders

Abstract. Real-world Structural Health Monitoring (SHM) applications can be classified to Zero-Shot Learning (ZSL) tasks, since structural data hardly ever reveal or label the full extent of damage that may be incurred to the system. This aligns with the definition of ZSL tasks, where part of the classes (or even all but one class) of the problem are hidden during the training time. Thus, the model must be able to generalize to classes that it has not been exposed to during the training. In SHM, this translates to a model of the studied structure generalizing the states/damages that are not present within the training data. This type of generalization is crucial to design resilient digital twins of infrastructure. Variational inference models are a prominent set of models for engineering features for ZSL tasks. This is because they learn latent features for the model with regularization for independence. Thus new clusters corresponding to unseen data would be more clearly identifiable in such a latent space. Specifically, Conditional Variational Autoencoders (CVAE) are a popular method used in ZSL tasks, as they allow the conditioning of the latent variables on external information pertaining to the data samples. Recent works have shown that Supervised VAEs (SVAE) can learn features that generalize just as well as CVAEs, yet they do not require the labels as an input during deployment. The SVAE conditions the latent space with an auxiliary tasks, in this case the GZSL task. Both the latent features and the GSZL task are learned jointly. Thus the labels are only needed during the training phase. We validate our method on a synthetic dataset consisting of simulated measurements on a wind turbine subject to stiffness degradation. In this case, the different classes correspond to different levels of erosion of the blades. Part of these damage levels are removed from the training set. We show that, when the unseen classes can be expressed by an interpolation of the seen classes, that the SVAE is able to learn global features. Moreover, our study demonstrates this generalization by tackling the damage detection task in a ZSL setting.