Un point de vue unifié sur la diagnosticabilité

The problem of model-based fault diagnosis in complex systems has received an increasing interest over the past decades. Experience has proved that it needs to be taken into account during the system design stage, by means of diagnosability analysis. Diagnosability is the ability of a system to exhibit different symptoms for a set of anticipated fault situations. Several approaches for diagnosability have been developed using different modelling formalisms., yet the reasoning for diagnosability analysis is very similar in all these approaches. This thesis provides a comparison of these and a unified definition of diagnosability. An original approach for diagnosability analysis, based on partial fault modes, is described. It is implemented in the context of service oriented architecture, more precisely on web services, and benefited from the framework of the european project WS-DIAMOND. An original generalization of the definition of diagnosability to any set of system states is presented, that accounts for many kinds of properties, like repair preconditions or quality of service. This work opens perspectives for model independent diagnosability reasoning, diagnosability based on other types of models, and in integrating diagnosis into a general purpose supervision tool. Model-based diagnosis and diagnosability of software systems is still a young applicative domain, and opens many connections with the software safety engineering domain. ; Le problème du diagnostic de défaillance à base de modèle dans les systèmes complexes a reçu un intérêt croissant durant les dernières décennies. Ce problème doit être pris en compte dès la phase de conception du système, au moyen de l'analyse de la diagnosticabilité. La diagnosticabilité est la capacité d'un système à exhiber des symptômes différents pour un ensemble de situations de défaillances anticipées. Plusieurs approches ont été développées basées sur différents formalismes de modélisation, toutefois les raisonnements menant à la diagnosticabilité sont très semblables ...