Reactive diagnosis of active systems

Automated diagnosis of discrete-event systems (DESs) is an interdisciplinary task being faced both by the Artificial Intelligence and the Control Theory communities. We consider a specific class of asynchronous DESs, called active systems, which are modeled as networks of communicating automata. A diagnostic technique for active systems, called reactive diagnosis, based on fragmented temporal observations, is presented. A fragmented observation is perceived one fragment at a time, under uncertain conditions. Each fragment carries some clues on how the system is evolving. Reactive diagnosis is based on three major requirements: parsimony, the unavailability of the global model of the system throughout the reasoning process; reactivity, the ability to output an updated group of candidate diagnoses at the occurrence of each observation fragment; and incrementality, the ability to reason in an incremental way, by exploiting previously generated knowledge-structures when dealing with each new fragment. Parsimony contrasts with the approaches in Control Theory, which need first generating the global model in order to make a diagnosis. Reactivity contrasts with those a posteriori diagnosis of Artificial Intelligence approaches, where the diagnostic task starts when the complete observation of the system has been collected. Experiments show that parsimony and incrementality make the technique workable when dealing with large-scale systems, such as telecommunication and power networks.