Early AI Diagnostics at Westinghouse
nets, which were already available at the
time, a poor choice for building a diagnostic
system. Neural nets also had the
disadvantage that the reason for a
conclusion was not explainable. Additionally,
at the time, data were largely on strip chart
recorders. To convert the strip chart
recorder to digital data was too laborious to
be practical.
Equipment VAX computers. PDS had an
interface for the knowledge engineer and
inference engine to execute the knowledge
base. PDS is described in reference. 6 The
design consisted of nodes (ideas) and rules
that connected them. The nodes were
classified
as
sensors,
hypotheses
(intermediate ideas) and malfunctions (final
diagnoses). Later a recommendation node
type would be added. Sensors were the
input, and the data could come from online
sensors (through the data center at the
power plant) or from manual entry of off-
line data. Rules took the confidence in the
input node and transferred it to an output
node. See Figure 4. The rules included AND
and OR functions of several types.
T HE D IAGNOSTIC S YSTEM
The solution was an expert system based on
the principles of MYCIN. 3 The result was an
expert system shell called Process Diagnosis
System (PDS). PDS was originally written in
LISP 4 by Mark Fox, then of Carnegie Mellon
University. 5 It was implemented on Digital
3
Shortliffe. E.H., MYCIN: A Rule-Based Computer Program for Advising Physicians Regarding Microbial Therapy Selection. PhD
dissertation, Stanford University (1974).
4
Winston, Patric Henry and Berthold Klaus Paul Horn, LISP, Addison Wesley (1981).
5
E.D. Thompson, E. Frolich, J.C. Bellows, B.E. Basford, E.I. Skiko, and M.S. Fox, “Process Diagnosis System (PDS) — A 30 Year
History,” Proc. 27th Conf. on Innovative Applications of Artificial Intelligence, 3928-3933, AAAI Publications (2015)
6
Bellows, James C., "An Artificial Intelligence Chemistry Diagnostic System," Proc. 45th International Water Conf. 15-23 (1984).
- 9 -
June 2019