Diagnosing physical systems by integrating models of normal and abnormal functioning

Scott Charles Bublin, Purdue University

Abstract

Traditional experiential knowledge-based diagnostic systems have suffered in their application to real world problems due to their inability to handle unanticipated situations and the difficulty of their construction. Subsequently, model-based diagnostic reasoning approaches called have been developed which rely on a model of the device's normal operation to mitigate these shortcomings. Unfortunately, traditional model-based systems often require an excessive amount of computation and data to arrive at a solution, primarily because information about any object in the model can only be utilized to make additional inferences about the objects physical proximal to it in the device. A formalism is presented that incorporates a multi-level causal network which represents knowledge about possible abnormalities together with a functional model of the device. Knowledge represented in the causal network is used to focus the search in the model-based component of the system. The two representations are integrated by incorporating fault models into the functional device model. The incorporation of fault models is used to minimize search and data acquisition by eliminating components from consideration whose fault models cannot explain the known symptoms. Finally, an architecture is presented for implementing the system.

Degree

Ph.D.

Advisors

Kashyap, Purdue University.

Subject Area

Electrical engineering|Computer science|Artificial intelligence

COinS