Diagnosis of discrete-event systems from uncertain temporal observations

Observations play a major role in diagnosis. The nature of an observation varies according to the class of the considered system. In static systems, an observation is the value of a variable at a single time point. In dynamic continuous systems, such a value is observed over a time interval. In discrete-event systems, an observation consists of a sequence of temporally ordered events. In any case, what is observed is assumed not to be ambiguous. This certainty principle, whilst being a useful simplification for a variety of contexts, may become inappropriate for a wide range of real systems, where the communication between the system and the observer is either bound to generate spurious messages, to randomly lose messages, or to lose temporal constraints among them. Consequently, the observation may be underconstrained. To cope with this uncertainty, a number of principles affecting both the observations and the modeled behavior of a system are introduced, that are independent of any specific processing technique. Furthermore, the notion of an uncertain temporal observation for discrete-event systems is introduced and accommodated within a graph whose nodes are labeled by uncertain messages, while edges define a partial temporal ordering among messages. This way, an uncertain observation implicitly defines a finite set of observations in the traditional sense. Thus, solving an uncertain diagnostic problem amounts to solving at one time several traditional diagnostic problems. The notion of an uncertain observation is further generalized to that of a complex observation. Both notions can be exploited by any diagnostic approach pertinent to discrete-event systems. Complex observations are contextualized in the framework of diagnosis of active systems and substantiated by a sample application in the domain of power transmission networks.