Early AI Diagnostics at Westinghouse
Figure 2: Combined High-Pressure and Intermediate-Pressure steam turbines, ca 1985. As the steam expands, the turbine blades
become longer. At the end of the Low-Pressure turbine, they were typically 31 inches long. 1
expanded as the square of the number of
conditions diagnosed. This fact precluded
very large diagnostic systems based on this
method.
T HE B EGINNINGS
In the mid 1970’s, Robert Osborne, Manager
of
Controls
Development
in
the
Westinghouse Steam Turbine-Generator
Division, was convinced that predictive
diagnosis was the future of controls
development.
He
supported
the
development of a predictive diagnostic
algorithm
based
on
conditional
1
probabilities. Lacking accurate values for
the conditional probabilities, the prototype
developers had control experts estimate the
probabilities. Thus, as implemented, it was a
precursor expert system. It became clear
that the conditional probability matrix
One of the problems with developing a
diagnostic program was that the incidents to
be diagnosed were infrequent. With
reasonable maintenance, the equipment
was extremely reliable. Forced outage rates
(time out of service due to an unplanned
outage) were below 1% with 0.05% believed
to be attainable. 2 In the entire fleet, a
frequent incident might occur several times
per year. There were not many examples
and even fewer examples with extensive
data. The paucity of examples made neural
1
Osborne, Robert L., Paul H. Haley, and Stephen J. Jennings, "Method and Apparatus for the Automatic Diagnosis of System
Malfunctions" US patent 4,402,054, 8/30/1983.
2
Meador, John T., Steam Turbines
States)).
IIC Journal of Innovation
No. ANL/CES/TE-78-7. ANL (Argonne National Laboratory (ANL), Argonne, IL (United
- 8 -