Skip to Main content Skip to Navigation
Theses

Etude de la tolérance aux aléas logiques des réseaux de neurones artificiels

Abstract : With the increasing complexity of treatments on satellite-borne and the utilisation of highly integrated circuits, the upset phenomenon becomes more and more crucial. Indeed, this phenomenon, caused by a heavy ion strike results in the modification of the information stored in a memory element. Upsets may perturb the operation of satellite-borne applications and can lead to serious consequences on the control of equipment operating in space. Artificial neural networks (ANN) constitute a new approach in information processing. They offer powerful and compact solutions to a wide range of problems, in particular those with real time constraints which is the case for most of current space applications. Among the main properties of neural networks we can mention their fault tolerance which measures their capability to perform the desired task under fault conditions (erroneous information) and to maintain their computing ability when a part of the network is damaged or removed. The goal of this thesis is to study the fault tolerance of neural networks against single event upsets in order to investigate the possibility of their utilisation in a radiative environment such as space. This work aims mainly at choosing, among the tested circuits, those that are accepted (or rejected) for space applications. Several networks and several circuits have been tested. Three kinds of experiments have been performed: software simulation of errors, hardware injection of faults and heavy ions tests. Obtained results show that Artificial Neural Networks are tolerant against upsets which make them a good candidate for use in space applications.
Document type :
Theses
Complete list of metadatas

Cited literature [163 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-00004913
Contributor : Thèses Imag <>
Submitted on : Friday, February 20, 2004 - 9:59:15 AM
Last modification on : Friday, November 6, 2020 - 4:03:26 AM
Long-term archiving on: : Thursday, September 13, 2012 - 2:00:29 PM

Identifiers

  • HAL Id : tel-00004913, version 1

Collections

IMAG | CNRS | UGA

Citation

Ammar Assoum. Etude de la tolérance aux aléas logiques des réseaux de neurones artificiels. Autre [cs.OH]. Institut National Polytechnique de Grenoble - INPG, 1997. Français. ⟨tel-00004913⟩

Share

Metrics

Record views

335

Files downloads

458