Special Theme: Control and System Theory
ERCIM News No.40 - January 2000

MAGDA - Alarm Supervision in Telecommunication Networks

by Eric Fabre


MAGDA (Modélisation et Apprentissage pour une Gestion Distribuée des Alarmes) is a research project supported by the French National Research Network in Telecommunications programme RNRT (Réseau National de Recherche en Télécommunication). The challenge for the MAGDA project is twofold: 1) adapting and mixing different formalisms towards the general objective of fault diagnosis in telecommunication networks, using alarm correlation, and 2) building an experimental platform to validate this approach.

Today, there are no widely-recognized standards for tools that could help alarm correlation. Thus, it is worth mentioning some key points of this research that could be regarded as useful achievements in and of themselves:

All the above mentioned points are very likely to lead to innovations in alarm correlation and network monitoring. We present now only the part of the project devoted to diagnosis using discrete-event control theory.

Model-based online monitoring

We consider a telecommunication network as a network of asynchronously interacting finite-state machines. Such systems are subject to spontaneous faults, occurrences of which may trigger alarms. Also, network elements get services from other elements and, in turn, provide services to several alternative network elements. This causes both fault and alarm propagation throughout the network.

As a first idea, we have proposed modelling such a situation as follows:

Alarm interpretation is then regarded as the problem of inferring, from the observation of alarms, the hidden state history of the PN. As some of the events (in particular, spontaneous faults) are typically random in nature, we consider some kind of probabilistic form for our PN. To prepare for distributed diagnostics, our requirement was that stochastic independence should match concurrency, implying that if two transitions of the PN are concurrent, all interleavings of their reception by the supervisor should have equal likelihood.

We have proposed a new class of stochastic PN that satisfy the above property: Partially Stochastic Petri Nets (PSPN). The well known Hidden Markov Models (HMM) theory has been extended to them. HMM are stochastic automata for which is it desired to infer the most likely hidden state (or transition) sequence from an observed sequence of transition signatures. This machinery is very popular in pattern recognition and in speech recognition. The basic algorithm is the so-called Viterbi algorithm, which computes on-line the most likely state history. In our PSPN framework, transitions are associated with ‘tiles’ that describe local state changes. The Viterbi algorithm then reconstructs hidden trajectories by concatenating tiles that match the observations. Therefore it is renamed the ‘Viterbi puzzle’.

ALMAP, the Alcatel Management Platform.

The following topics require further development for MAGDA and are currently under investigation:

Currently, the ‘Viterbi puzzle’ group at IRISA is comprised of three permanent researchers (Albert Benveniste, Eric Fabre, and Claude Jard) and two post-doctoral fellows (Mark Smith from MIT and AT&T Labs Research and Laurie Ricker from Queen’s University in Canada). Laurie Ricker is supported by the ERCIM PhD fellowship programme. After this project, she will continue her research activity at CWI with Jan van Schuppen.

The MAGDA project is a collaboration between two academic research centres (IRISA/INRIA Rennes and LIPN Paris) and three industrial companies (France-Télécom/CNET, Alcatel/CRC and ILOG).

Links:

Magda web site: http://magda.elibel.tm.fr/

Please contact:

Eric Fabre - INRIA/IRISA
Tel: + 33 2 9984 7326
E-mail: Eric.Fabre@irisa.fr


return to the ERCIM News 40 contents page