An unsupervised structural health monitoring framework based on Variational Autoencoders and Hidden Markov Models

Vibration-based structural health monitoring (SHM) has traditionally been based on the variation of modal properties. Nowadays engineering structures require multiple sensors for reliable monitoring, which implies analyzing large volumes of data and demands the development of machine learning methods. As such, deep learning techniques applied to vibration have emerged recently, and successful applications have been reported for pattern identification from high-dimensional data. However, a lack of expert annotated labels related to damage conditions in real structures hinders the use of supervised techniques, which motivates the development of unsupervised methods. An unsupervised framework is here proposed that combines Variational Autoencoders (VAE) and a Hidden Markov Model (HMM), aiming to learn a degradation model and classify the state evolution from measured vibration signals. The proposed method is assessed using the IMS bearing dataset and an experimental dataset. A scaled prototype of a cable-stayed electrical transmission tower was monitored for eight weeks, while it has been subjected to progressive cable slack scenarios. Results show that the VAE successfully embedded the data to a feature space where it was possible to learn a degradation model, which indicates that the combination of VAE with HMM is a promising solution for SHM.