Abstract
In this paper we present a model-based approach to the on-line diagnosis of dynamic systems. We model the system to be diagnosed as a discrete, synchronous transition system and capture temporal phenomena such as the change of the system inputs, the evolution of the system internal status and the evolution of the health conditions of the system components.
The on-line diagnostic task consists of three subtasks: estimating the potentially highly ambiguous belief state (i.e. the set of possible system states), detecting significant changes in the belief state (in particular, changes in the set of preferred diagnoses) and presenting the preferred diagnoses to the user. We present a backtrack-free algorithm that keeps track of the complete belief state even when such a set is very large; we then introduce efficient algorithms that perform the detection of changes in the set of preferred diagnoses and the presentation of preferred diagnoses. The selection of preferred diagnoses is based on the adoption of ranks for representing the probabilities of occurrence of faults.
In order to achieve completeness and efficiency, we exploit symbolic techniques (in particular, Ordered Binary Decision Diagrams) to encode and manipulate the system model and the belief state. The approach is tested on two real-world models, taken from the automotive and aerospace domains.