Agile Diagnosis of Active Systems by Condensed Abduction

Automation of diagnosis has been addressed since the dawn of AI, both in medical and industrial domains. This paper focuses on model-based diagnosis of a class of asynchronous discrete-event systems (DESs), called active systems, which can also be an abstract representation of continuous dynamical systems, such as automobiles, power networks, or nuclear plants. In that context, diagnosis is abduction-based, where the behavior of a DES that conforms with a given temporal observation needs to be constructed ex ante in order to eventually distill the candidate diagnoses. Since the abduction knowledge-structure produced may be huge, the diagnosis engine may slow down to a point that the monitoring of the system could be jeopardized, especially in critical domains that require recovery actions in real-time as a reaction to the diagnosis response. To make the diagnosis task more agile, a novel abduction technique is presented, which is based on two requirements: reduction of the size of the abduction and, more to the point, reduction of the computation time by avoiding duplicating low-level model-based reasoning. A worst-case performance analysis indicates that a Condensed Abduction Engine may outperform abduction-based engines of active systems significantly.