Unsupervised Flight Fault Propagation Analysis Using a Variational Autoencoder

Assessing fault propagation is essential for flight safety and for optimal maintenance scheduling of passenger aircrafts. This paper studies such application by way of latent space representation with variational autoencoder. The proposed method learns a topological representation of the dynamics of a normal flight from real industrial data. The method is validated by showcasing that the trained model can separate among different flight modes, e.g., taxiing, landing, and cruise, etc. Then, through case studies of various anomalous flights, it is shown that the dynamics, and as a consequence the latent space representation, of those flights significantly differ from those of the normal flights. The proposed method can be applied subsequently to analyze fault propagation.