Reflectometry is a diagnostic method to identify and characterize electrical faults on cables or more generally on transmission lines. The analysis of these faults makes it possible to improve the maintenance of the cables. The LFIC laboratory at DM2I has developed a technology based on compressive sampling in order to make possible a finer spatial localization of defects while limiting the amount of data to be stored. Fault detection relies on reconstructing a complete signal from compressed measurements and this has a considerable calculated cost.
In this thesis proposed in collaboration with the LS2D laboratory of DM2I, we explore the possibility of making efficient high-level inferences on the presence and nature of faults in cables, directly on compressed measurements, without reconstruction. This relies on machine learning techniques.
The novelty of the planned work is based on an integrated consideration of the constraint of frugality:
- energy sobriety at the inference stage, by eliminating the reconstruction phase and using models to lighten models, in particular with a view to on-board implementations;
- energy sobriety in the learning phase by studying the transferability of models in connection with understanding the impact of different learning factors such as the type of cable, the compressive sampling matrix, etc.
- frugality in operational data by using simulations and studying the problem of domain adaptation between simulated data and "real" data. A possible extension of the work will focus on lightening simulations.