Disentangled representation learning with physics-informed variational autoencoder for structural health monitoring

The maintenance and replacement of aging infrastructure is becoming increasingly costly, as a large number of structures is approaching the end of its design lifetime. Labor intensive manual inspection and assessment on a large scale is often infeasible, while data-driven machine learning techniques can fail to identify relevant failure mechanisms and suffer from poor generalization to previously unseen conditions, particularly when data is sparse. We propose a multi-task learning approach based on a physics-informed variational autoencoder that leverages a dictionary of physics-based models to improve the predictive performance when limited data is available. The latent space of the autoencoder is augmented with a set of physics-based latent variables that are interpretable and allow for prior knowledge on the latent variables to be included in the autoencoder formulation. To prevent the data-driven components of the encoder and decoder from overriding the known physics, a regularization method is used to impose physical constraints on the latent space and the generative model prediction. The feasibility of the proposed approach is evaluated on a case study of quay walls in Amsterdam, demonstrating improved performance over purely data-driven methods.