Abduction and Abstraction in Diagnosis: A Schema Based

Account

Carl R. Stern
University of New Mexico
and
George F. Luger
University of New Mexico

Address correspondence to:
Carl R. Stern
Department of Computer Science
University of New Mexico
Albuquerque, New Mexico 87131
(505) 345-4147
[email protected]

Abstract

We present a schema-based architecture for semiconductor failure analysis. Schemas serve to guide the application of scientific domain knowledge in the construction of causal explanations. They organize data gathering and provide a road map for the detailed investigation of causal processes. They shape the form of explanations, ensuring that explanations satisfy not only theoretical requirements, e.g. consistency with applicable causal laws, but also practical requirements, in particular the need to support a specific range of remedial practices.

Our work is based on the observation and analysis of expert performance in the area of semiconductor component failure analysis. We have worked with five failure analysts over nearly half a decade in the process of constructing a failure analysis expert system. Our schema-based architecture represents an attempt to model the way in which experts combine heuristic reasoning with reasoning from first principles in order to produce both correct and useful explanations.

(1) EXPERTISE IN CONTEXT: COMPONENT LEVEL FAILURE ANALYSIS

Semiconductor component failure analysis offers an important example of expertise in context (Luger & Stern, 1992). The failure analyst is presented with an initial set of signs, for example the abnormal behavior of a diode after burn in (fn 1) and is required to organize an investigation based on an interpretation of those signs. The analyst begins the analysis by gathering information about the history and vulnerabilities of the device as well as the particular circumstances of the current failure. The initial visual and electrical examinations is conducted against the background of this historical information. Based on the initial examination, the analyst adopts a prioritized list of hypotheses, the failure mechanisms which could account for the abnormal device behavior. Data gathering then proceeds, focused by the active hypothesis set.

As new information is acquired, some hypotheses are dropped while others are modified. In the