Diagnosis and Sensor Validation through Knowledge of Structure and Function

Abstract

A broadly applicable algorithm can guide diagnostic reasoning using knowledge of structure and function. It is shown that, for a useful class of domains, a faulty sensor is no harder to diagnose than other system objects (and may even be easier), thereby offering a solution for at least a subset of "the sensor validation problem." A prototype called LES has been constructed which may be the first system to diagnose from sensor data using model-based knowledge of structure and function. While classical forward-chaining diagnostic systems need to find out whether or not their sensors are telling them the truth before they can draw inferences about the state of the system, LES uses both structural and functional knowledge to find the expected state of all system objects, including sensors. In such domains it has been found that it is even economical to use functional relationships in place of structural knowledge. The approach is generally algorithmic rather than heuristic and represents uncertainties as sets of possibilities. Functional relationships are inverted to determine hypothetical values for potentially faulty objects and may include conditional functions not normally considered to have inverses. LES' model-based generation of expectations and explanatory hypotheses and its compact and perspicuous knowledge-base design are likely to be useful for specifying component behaviors beyond the realm of diagnostic problem solving.