Co-authors Yu (Eve) Zhang, Miguel Martínez, Emanuil Mladenov, PhD, Shaheryar (Shery) Khan.
Variational Autoencoders (VAEs) are a major kind of deep generative model. In a nutshell, a VAE is an autoencoder whose encodings distribution is regularised during the training in order to ensure that its latent space has good properties allowing us to generate some new data.
The paper demonstrates a use case for detecting anomalies in an unsupervised, black-box approach in various subsystems of aircraft by training it on a state-of-the-art flight dataset. The hypothesis tested is whether the latent space extracted from a VAE can be effectively used to forecast faults and assist in scheduling the maintenance of conventional aircraft. The algorithm is
validated by matching regions in the latent space to different flight modes.
Download the full Paper on Unsupervised Flight Fault Propagation Analysis Using a Variational Autoencoder.