Study of embedded prognostic strategies in wired networks based on temporal neural networks
Published : 8 February 2020
Whatever their fields of application, cables are very often victims of their operating environment. They often face aggressive conditions such as mechanical vibration, heat stress, moisture penetration, etc. These conditions favor the appearance of more or less serious defects ranging from a simple crack in the sheath to a cable break thus causing a malfunction of the system.
In this context, the CEA LIST studies methods of diagnosis and prognosis of defects in cable networks based on the reflectometry method. The idea is to inject a test signal into the cable. Whenever it encounters an impedance discontinuity (i.e. a fault), some of its energy is returned to the injection point. The processing of the reflected signal subsequently makes it possible to detect and locate this defect.
Despite the maturity of the reflectometry to detect a defect in a cable, it does not allow to determine the causes of the appearance of an incipient defect (ie damage of the shielding, radius of curvature, pinching, etc.) nor to predict its evolution in the future. The work of this thesis aims at developing new prognostic strategies for defects in wired networks. For this, the application of Machine Learning methods such as Artificial Neural Networks (ANN) on data from reflectometry sensors is a promising solution to solve this problem. It is in this context that the works of this thesis are inscribed.