Early AI Diagnostics at Westinghouse
diagnostic system was required to remove
the sensor from consideration. This was
done by creating a set of hypotheses that
were validated conclusions about the sensor
value. Figure 4 illustrates this process for the
sodium sensor on the Condensate sample.
There was a condensate-sodium-high (Cond-
Na-high) hypothesis, based simply on the
value of the sodium sensor on the
condensate sample, and a validated-
condensate-sodium-high
hypothesis.
Connecting them was a rule that normally
transferred all the confidence from the first
to the second hypothesis. However, there
was a diagnosis of the sodium sensor based
on other sodium sensors and other sensors
on the condensate sample. If the condensate
sodium sensor was deemed degraded, a
Parametric Alteration Rule (Paralt Rule)
would reduce the transferred confidence.
Paralt rules could modify the parameters in
another rule. In Figure 4, this is denoted by
the valve symbol in the middle of the rule
that transfers the confidence from the
unvalidated Cond-Na-high hypothesis to the
validated-Cond-Na-high hypothesis. This
process is explained in detail in the author’s
paper that first revealed the chemistry
diagnostic project to the power plant
chemistry community in 1984. 1
The chemistry monitoring system that
supported the diagnostics had multiple
sensors of the same type at various locations
within the power plant. PDS was augmented
to allow writing of a subsystem for each
sensor type and instantiating the system
multiple times with node and rule labels
modified during the instantiation process.
The chemistry system, which would
eventually become ChemAID ® for Chemistry
Artificial Intelligence Diagnostics, had four
iterations completely starting over before it
was implemented at a power plant with over
4000 rules in 1988. 8
One of the important goals of the diagnostic
system was to be able to explain why a
diagnosis occurred. The author used
relatively long, descriptive names for the
hypotheses and developed each diagnosis by
solving the problem step-by-step. This
approach makes the rulebase clear. One can
trace the source of confidence in a diagnosis.
These two techniques have the added value
that the rulebase remains understandable
even after years without maintenance.
8
Bellows, James C., Karen L. Weaver & James Gallatin, "On-Line Steam Cycle Chemistry Diagnosis," Proc. 2nd Fossil Cycle
Chemistry Conf., Seattle, Sept 1988, Electric Power Research Institute Report GS-6166.
- 11 -
June 2019