LIP6 1998/043
- Thesis
APPRENTISSAGE ET DIAGNOSTIC DE SYSTEMES COMPLEXES : RÉSEAUX DE NEURONES ET RÉSEAUX BAYÉSIENS.
Application à la gestion en temps réel du trafic téléphonique français - Ph. Leray
- 176 pages - 09/10/1998- document en - http://www.lip6.fr/lip6/reports/1998/lip6.1998.043.ps.tar.gz - 588 Ko
- Contact : Philippe.Leray (at) nulllip6.fr
- Ancien Thème : APA
- Keywords : Neural Networks, Bayesian Networks, Learning Diagnosis, Feature Selection, Confidence Measurement
- Publisher : Valerie.Mangin (at) nulllip6.fr
This work discusses about diagnosis of complex systems with pattern recognition methods. Industrial systems are more and more complex, they need to be monitored continuously to detect disruptions and maintain a good quality of service. These considerations led to important efforts in the diagnosis field. We have began to work with the CNET, the France Télécom research center, on the application of neural network techniques for real-time telephone network management.
The first section of this document consists of three chapters about the use of connectionist methods for complex system diagnosis. Chapter 1 briefly presents the main problematics of diagnosis, the application of pattern recognition for diagnosis, and more specially neural network and bayesian networks. Chapters 2 and 3 propose specific studies about the use of neural networks for diagnosis: confidence measurement for classification tasks and feature selection.
The second section concerns more precisly the application of connectionist tools to telephone network trafic management. Chapter 4 presents the task and the approach we developped: first, a local and statical generation of alarms corresponding to the different overloads, then different alarm filtering methods taking into account correlations existing in a complex system. Chapter 5 present the different experiments we proceed in the local level of our diagnosis architecture. Chapter 6 concerns the use of neural networks and bayesian networks for temporal alarm filtering and for spatial alarm filtering.
The first section of this document consists of three chapters about the use of connectionist methods for complex system diagnosis. Chapter 1 briefly presents the main problematics of diagnosis, the application of pattern recognition for diagnosis, and more specially neural network and bayesian networks. Chapters 2 and 3 propose specific studies about the use of neural networks for diagnosis: confidence measurement for classification tasks and feature selection.
The second section concerns more precisly the application of connectionist tools to telephone network trafic management. Chapter 4 presents the task and the approach we developped: first, a local and statical generation of alarms corresponding to the different overloads, then different alarm filtering methods taking into account correlations existing in a complex system. Chapter 5 present the different experiments we proceed in the local level of our diagnosis architecture. Chapter 6 concerns the use of neural networks and bayesian networks for temporal alarm filtering and for spatial alarm filtering.