Annals of Mathematics and Artificial Intelligence 11(1994)381-398 381 Diagnosing time-varying misbehavior: an approach based on model decomposition* Luca Console, Luigi Portinale, Daniele Theseider Dupr6 and Pietro Torasso Dipartimento di Informatica, Universith di Torino, Corso Svizzera 185, 10149 Torino, Italy The analysis of time-varyingsystemsis attracting a lot of attention in the model- based diagnosis community. In this paper we propose an approach to the diagnosis of such systems, relying on a component-orientedmodel; we provide separately a behav- ioral model, that is, knowledgeabout the consequencesof different behavioralmodes of the components, and a model of the possible temporal evolution of such modes (mode transition graphs). In the basic approach, we assume that the consequences of beha- vioral modes are instantaneous with respect to the transition between two modes; this allows us to decompose the solution of a temporal diagnostic problem into two subtasks: determining solutions of atemporal problems in different time points and assembling the solution of the temporal problem from those of the atemporal ones. Most of the definitions and machinery developed for static diagnosis can be re-used in such a framework. We then consider the consequences of some extensions. Even allowing for very simple temporal relations in the behavioral model leads to a more complex interferencebetween reasoning on the behavioral models and the consistency check with respect to possible temporal evolutions. We also brieflyanalyze the case of adding quantitative temporal knowledgeor probabilistic knowledgeto the mode tran- sition graphs. 1. Introduction The need of taking into account the temporal dimension in model-based diagnosis has been advocated by many researchers and becomes particularly relevant when diagnosis is integrated with monitoring, control and repair [16]. The temporal dimension in diagnosis, however, can be faced under different respects: 9 Associating a temporal label to observations; then the diagnostic hypotheses should also have one, in order to account for both the observations and for their temporal location. At least two phenomena can be taken into account in this way: (1) Temporal delays between faults and their (observable) consequences, that have been considered in [4, 17], where the problem of checking con- sistency between the temporal location of the observations and such delays is studied. * This work was partially supported by CNR under grants 91.00916.PF69 and 91.02351.CT12. 9 J.C. Baltzer A.G., SciencePublishers