FUNCTION-BASED CANDIDATE DISCRIMINATION DURING MODEL-BASED DIAGNOSIS

We propose function for candidate discrimination, i.e., suspect orderi ng during model-based diagnosis. Function offers advantages over struc ture and fault probabilities currently being used for candidate discri mination. It is readily available from device design, unlike fault pro babilities, which are hard to obtain. Function-based discrimination is not dependent on the topology of the device, unlike structure-based d iscrimination. We propose classes as a scheme for representation of fu nction. As part of classes, we define a set of function primitives and provide a framework for identifying the functions of components and s ubsystems of a device. The representation scheme is domain independent . We propose a function-based technique for candidate discrimination c alled the default order technique, and outline a diagnosis algorithm t hat applies the technique to the class model of a device. Function-bas ed diagnosis is in addition to and as a supplement for model-based dia gnosis based on behavior and structure. We demonstrate by qualitative analysis that function-based discrimination is at least as effective a s fault probabilities for candidate discrimination of simple devices. In complex devices, function facilitates explanation generation based on causality, which is a desirable feature of diagnosis systems. Our d iscrimination technique provides a functional basis for partitioning c omponents in the practicable version of the minimum entropy technique proposed by deKleer.