Design documents were used to choose relative parameter
values so the behaviors generated would be qualitatively accurate.
The monitoring sample rate corresponded to 20 seconds. Component failures
were modeled by changing model parameters by a factor of 5.
Capacitive and inductive failures produced
abrupt change of pressure and flow, respectively.
The margin of error was set at 
for
practical reasons, and signatures were generated up
to 
order derivatives. A straightforward implementation of the
diagnosis algorithms in Visual Basic Pro 3.0
performed fault hypothesis generation and prediction
in 9 seconds on a 200 MHz Pentium
Pro machine with 32 MB RAM.
Steady state was difficult to detect and not used.
The variables 
in Fig. 15 are typical measurements.
Simulation results (Table 1) showed that, except for 
, all
faults could be accurately diagnosed in a reasonable number of time steps, k.
To detect , flow of sodium through the overflow mechanism
would have to be monitored, which shows the importance of measurement
selection [9, 17].
, 
, 
, 
, 
, and 
were detected
but not uniquely isolated because
the configuration change as a result from the overflow mechanism
introduces uncertainties in the diagnosis.
Precision in diagnosis may improve by considering predicted effects
of order higher than 3,
but as noted before, they take longer to manifest which may then
cause cascading faults to appear. In real situations, cascading multiple
faults are more likely than independent multiple faults. Cascading
faults are best handled by quick analysis of transients to establish
root-causes and then suspend diagnosis when other faults begin to
influence the measured transients.
In spite of the
loss of precision, the results are more practical from a computational
viewpoint.
Table 1: Fault detection and isolation, k is the required number of
samples after detection.
