Phd of DAMPEYROU Charles

Abstract :

The goal of this thesis is to develop a numeric solution based on learning algorithms for unsupervised anomaly detection in multi-variate time-series. This solution will be developed in the context of plane equipping sensor systems. The behavior of the sensors being physically linked to the plane, the originality of the approach stays in the integration of these physical constraints into the definition and/or training of a VAE architecture. This approach will be confronted to state-of-the-art source separation methods, and particularly on anomaly detection techniques used for mechanical systems. This should allow the identification of anomalies which do not respect the underlying physic (defaulting sensor for instance) or anomalies corresponding to an extreme mechanical behavior (extreme load, structural default). Finally, the proposed approach could be used to automatically consolidate data used for the training of prediction models (these models can be prediction models for the structural behavior of a plane given the on-board instrument informations, as developed during the Challenge AI for Industry 2020, or other models using multi-variate time series as input) or be used for decision support by detecting the appearing of defaults in a context of predictive maintenance.

Supervision :
Under supervision of Prof Jean-Luc DION (ISAE-Supméca) and
MCF Martin GHIENNE (ISAE-Supméca)