Monitoring, Prediction, and Fault Isolation in Dynamic Physical Systems

Abstract

Diagnosis of dynamic physical systems is complex and requires close interaction of monitoring, fault generation and refinement, and prediction. We establish a methodology for model-based diagnosis of continuous systems in a qualitative reasoning framework. A temporal causal model capturing dynamic system behavior identifies faults from deviant measurements and predicts future system behavior expressed as signatures, i.e., qualitative magnitude changes and higher order time-derivative effects. A comparison of the transient characteristics of the observed variables with the predicted effects helps refine initial fault hypotheses. This allows for quick fault isolation, and circumvents difficulties that arise when interactions caused by feedback and dependent faults. This methodology has been successfully applied to the secondary cooling loop of fast breeder reactors.