10.

3 Nuclear Power Plant Fault-Diagnosis Using Artificial Neural Net-

works

This paper was accepted for presentation at the Artificial .Veural ,Vetworks m En-
gineering (.4,V,VIE) 1992 conference and will be published irt the proceedings of the
conference. The confernce will be held November 16-18, 1992 at St. Louis.

CONF-921185--2

DE93 010314

!
i

507

,ml
'# I|ll Sv " =""

I
 NUCLEAR POWER PLANT
FAULT-DIAGNOSIS USING
ARTIFICIAL NEURAL NETWORKS

Keehoon Kim, Taher L. Aljundi, and Eric B. Bartlett

Nuclear Er_gmeemng Program
Department of Mechanzcal Engineering
Iowa State Unzverszty

ABSTRACT

Artificial neural networks (ANNs) have been applied to various fields

due to their fault and noise tolerance and generalization characteris-
tics. As an application to nuclear engineering, we apply neural net-
works to the early recognition of nuclear power plant operational tran-
sients. If a transient or accident occurs, the network will advise the
plant operators in a timely manner. More importantly, we investigate
the ability of the network to provide a measure of the confidence level
in its diagnosis. In this research an ANN is trained to diagnose the
status of the San Onofre Nuclear Generation Station using data ob-
tained from the plant's training simulator ill. Stacked generalization
is then applied to predict the error in the ANN diagnosis 121. The
data used consisted of 10 scenarios that include typical design basis
accidents as well as less severe transients. The results show that the
trained network is capable of diagnosing ali 10 instabilities as well as
providing a measure of the level of confidence in its diagnoses.

INTRODUCTION

Nuclear power plant safety is important to both the public and nuclear pro-
fessionals. Early transient diagnosis can give reactor operators extra time to
formulate and perform corrective actions that can prevent a transient from de-
veloping into a potentially serious accident. It is most beneficial to provide both
the diagnosis and a figure of merit describing the diagnosis accuracy. Estimat-
ing the uncertainties associated with the ANN diagnoses of nuclear power plant
transients is a vital step towards enhancing safety since it is unproductive for
operating personal to be misled by an incorrect automated diagnosis.
ANNs have many advantages over expert systems, since expert systems re-
quire explicit rules which is a timc consuming process for both programming and
excecuting. However, one disadvantage of ANNs has been the relative difficulty
of assigning error bounds to their outputs. This disadvantage has limited the
application of ANN techniques to areas where verification is not too important
or where other methods can be used as independent checks on the results.
In this work a fault diagnostic adviser was developed oy training a backprop-
agation neural network [4) to diagnose the status of the San Onofre Nuclear

509

!
 i

Generating Station using data obtained from the plant's training simulator ili.
The data simulate the plant's conditions through time during 10 transients sce-
narios. Table 1. lists these scenarios. We then apply stacked generalization
to provide a measure of the level of confidence in each transient's diagnosis
performed by the ANN.
The results of our research show that the ANN fault diagnostic adviser can
not only correctly diagnose each scenario but most importantly it can provide
error bounds on its diagnoses. The application of stacked genralization to nuclear
power plant fault diagnostics is the central contribution of this paper.

STACKED GENERALIZATION

Generalization is the ability to classify a given novel input based on its relation
with known stored knowledge. Mathematical algorithms that can interpolate
or extrapolate the behaviour of a given set of input-output examples are called
generalizers (5.61. Accordingly ANNs are generalizers.
Resolving the uncertainties associated with ANNs generalization is a major
concern for ANN researchers ;7]. Stacked generalization, however, is a tech-
nique that can be used to address this concern. There are several different
ways to implement stacked generalization. When applied to a single general-
izer. stacked generalization can provide estimates of classification error. In this
research we use a backpropagation neural network as the single generalizer to
diagnose nuclear power plant instabilities. We then apply the stacked gener-
alization technique to predict the error _ssociated with the diagnostic adviser
output. Predicting the error of the ANN adviser is accomplished by training
one network to learn the relationship between the inputs and the outputs of the
desired mapping (in our case the nuclear power plant fault diagnostics), and
another network is trained to learn the relationship between its inputs and the
classification errors of the first network. When presented with a novel input, the
first network will classify the input, while the second network will predict the
error in the output of the first network.
To employ stacked generalization we begin with a collection of r partitions of
the learning set L. The i-th partition Pi splits the learning set L into two disjoint
subsets Lij, where 1 < i < r, and je {1,2}. In this work, We arbitrarily chose a
partition set with r = m where m is the number of patterns in the training set L
18]. Therefore, for ali i, Li_ consists of a single pattern (x,y), the corresponding
Li1 consists of the remainder of the training set {L-(x,y)}. Wolpert f2l:defines
the original learning set L as the level 0 space. Any network applied directly to
L in the level 0 space is then called a level 0 generalizer, and the original learning
set L is called a level 0 learning set. The level 0 generalizer is then trained on the
subset {L-(x,y)} and asked the question x. The level 0 generalizer's output, g
and the vector (Euclidean distance) from x to its nearest neighbor in {L-(x.y)}.
k ere saved. Since the level 0 generalizer has not been trained with the pair (x,y),
g will. in general, differ from y. Therefore: when the question is x. the ANN
answer differs from the desired output y by (g - y). This information along with
the question x and its ve :tor to the nearest neighbor in the learning set can be
cast as input-output in a new l_arning set in level 1 space. The level 1 input is
the pair (x,k) and the output is ig-yi. Choosing other partitions of L gives other
510

!
such patterns. Taken together, these patterns constitute a level 1 learning set L'.
This level 1 learning set contains the relationship between the set of questions
{q} nd the level 0 ge,,:ralizer's error in guessing the outputs correspond to
the set {q}. We then train a level 1 generalizer to learn this relationship from
the level l training set. We then ask the level 0 generalizer a novel question
q. and feed the pair q and the vector from q to the nearest neighbor in L. as
a question to the level 1 generalizer. The output of the level 0 generalizer will
be the classification (answer) of the question q, while the output of the level
1 generalizer is an estimate of the level 0 generalizer's error in classif.ving the
question q.

METHOD

The nuclear power plant fault diagnostic adviser developed here uses a back-
propagation neural network with a 33 X 22 X 10 X 4 architecture as the level
0 generalizer. Thus. the network has 33 nodes in the input layer. '2'2 nodes in
the first hidden layer. 10 nodes in the second hidden layer, and 4 nodes in the
output layer. The 33 input nodes receive as an input a single time slice snapshot
of 33 of the plant variables, and the 4 output nodes are used to distinguish each
of the 10 transient conditions with a distinct 4-bit binary code. The two hidden
layer architecture was employed after attempting several different architectures.
Training the ANN adviser is accomplished iteratively 13,7,91. The initial
training set contained 20 patterns, two from each transient. These two patterns
were: the first pattern in each scenario, corresponding to normal operating con-
ditions at time = 1 second, and the last pattern in each scenario, corresponding
to a time when the transient is well established. Training on this data was per-
formed until a root mean square (RMS) error of .01 was obtained. The next
step in the iterative training approach is to recall the network on the entire data
set for each of the 10 transients. The error obtained from the recall set is then
plotted against time for each transient. Usually there are several peaks in these
plots where the error is very high. These peaks correspond to patterns thai the
ANN incorrectly classifies because they are very different from those chosen in
the initial training set. These incorrectly classified patterns are then included
in the training set of the network for the next training iteration. These steps
are repeated until all peaks in the recall set fell below .1 RMS error. The final
training set contained 113 patterns selected from all 10 scenarios.
The final training set. obtained by the procedure outlined above, is the level
0 training set. L. The level 1 training set was composed of the level 0 input
x and the vector k from x to its nearest neighbor. The level 1 desired output
ia the difference between the level 0 actual and desired outputs. The level 1
generalizer is another backpropagation neural network with a 66 X 30 X 20 X
10 X 4 architecture. Again, this architecture was chosen after several attempts
were made to find the optimal architecture.

511
RESULTS

Table I shows that the ANN fault diagnostic adviser is capable of classifying
each of the 10 transients as well as providing a measure of the confidence level
in its classification. The first column in Table 1. lists the 10 transients used
in this research. The second column shows the transients onset times. The
third column shows the time needed for the network to make a diagnosis. The
fourth column shows the additional time spent between the correct diagnosis
and the instant when the confidence in the classification reaches and maintains
an acceptable level. The acceptable confidence level was arbitrarily taken to be
a value below .1 in the estimated error of the diagnosis.
Figures 1 and 2 are one example of the 10 pairs of plots that were obtained
by diagnosing and predicting the error in the diagnoses of the 10 transients. The
small peaks at early times in Figure 1 indicate that the adviser has detected an
instability in the plant's conditions. After few seconds, the adviser tells us that
the Stuck Open Pressurizer Safety with High Pressure Injection Inhibited tran-
sient is responsible for the instability in the plant. Figure 2 however, indicates
that the level of confidence is low until 6.3 seconds after transient onset.
Another interesting example is the Turbine Trip From .50% Power. illust rated
in Figures 3 and 4. Figure 3 shows that the ANN adviser makes a correct diag-
nosis of this transient after only 1 seconds. Figure 4, however, shows that the
predicted error on this diagnosis is above 0.1 for the entire scenario. Although
the classification of this transient is correct, the network confidence in its classi-
fication is low since the estimated error is large for the entire time period. The
diagnosis of the Turbine Trip From .50% Power is unreliable even though it is
correct. The network is capable of diagnosing the transient because of the gen-
eralization capabilities of ANNs but it failed to assure the diagnosis. The cause
of this failure is that the data for this particular transient were collected from
50% power while the rest of the data were collected from 100% power. The level
of confidence in the diagnosis of this transient can be increased by trainin_ the
ANN adviser with more transients from this, lower, reactor power level.

CONCLUSION

This paper has demonstratedthefeasibility of usingANN technologycoupled

withthestackedgeneralization techniqueas a diagnostic toolfornuclearpower
planttransients. In thisresearchthestackedgeneralization techniquewas used
to predictthe levelof confidence in each diagnosis.Noticethat the network.
accordingtotableI.respondsveryrapidly tothechangesintheplantconditions.
The results ofthestackedgeneralization agreedwithwhat we would expect.The
diagnoses of9 ofthetransients was correctwitha highlevelofconfidence. While
theconfidence indiagnosingtheTurbineTripfrom 50% Power was low because
ofthedifferent power levelat whichthetransient occurred.
Futurework willfocuson thegoalof testinga prototypeANN adviserat a
nuclearpower plant.This willnecessitate thetrainingand developmentofan
integrated ANN advisor that is capable of classifying many more transients as
well as providing a prediction of the errors in its results. Implementing such an
adviser in a nuclear power plant will provide continuous and accurate monitoring

512
 Table 1: Summary of ANN diagnostic adviser results
Additional
time, after
Time. after correct
Name of transient the onset, diagnosis.
Transient needed to needed to
onset diagnose the assure the
time (s) transient (s) diagnosis (s)
1. Turbine Trip; Reactor Trip. 6 30 1
'2. Loss of Main Feedwater
Pumps. 47 3 0.
3. Closure of Both M_'n Steam
Isolation Valves. 7 28 0.
4. Trip of Ali Reactor
Coolant Pumps. 16 2 0.
5. Trip ofa Single Reactor
Coolant Pump. 14 62 0.
ft. Turbine Trip From 50% 50 1 Fail to make
power, an assured
diagnosis,
7. Loss of Coolant Accident
With Loss of Off-Site Power. 7 14 0.
8. _Iain Steam Line Break. 6 4 31
9. Stuck Open Pressurizer
Safety Valve With High Immediate
Pressure Injection Inhibited 15 diagnosis 63
10.Single Turbine Governor
Valve Closure. 7 16 0.

of the plant's integrity. It will also provide fast and reliable diagnosis of system
instabilities and therefore will significantly enhance the safety of nuclear power
plants.

0.5 - 0.4 "

: 0.4-

_ 0.3-

0.3"_ ._
_ _ o.2
o.2" _-
0.1-_ I
0.1-
.,
I
0 _-- _*-_ -- v. 0 ""
1 101 201 301 401 50I t 101 201 301 401 501
Time (s) Time Isl

Figures i and 2. Plots of time versus error for the level 0 generalizer tleft)
and time versus output for the level 1 generalizer (right) for the Stuck Open
Pressurizer Safety With High Pressure Injection Inhibited transient. The pre-
dicted confidence of the level 0 generalizer's output is one minus the level 1
generalizer's output.
513

,r, .......
I
 0.5 - 0.4 -

-- 0.4-

_: ,,,=
0.3- ili
-- 0.3- = _'--'i
I

_ 0.2-
0.2-,' _"
,=

I 0.1 -
0.1- '1
i
0 _ - -+ O
1 101 201 301 401 501 1 101 201 301 401 501
?_rnots) Time(s)

Figures 3 and 4. Plots of time versus error for the level 0 generalizer (left)
and time versus output for _,he level 1 generalizer (right) for the Turbine Trip
From 50% Power transient. The predicted confidence of the level 0 generalizer's
output is one minus the level 1 generalizer's output.

ACKNOWLEDGEMENT
This work was made possible by the gracious support of the United States Department of
Energy under Special Research Grant No. DE-FG02-g2ER75700, entitled "Neural Network
Recognition of Nuclear Power Plant Transients". Their support does not however constitute
an endorsement of the views expressed in this article.

REFERENCES
l. Data provided by T. James, S. Olmos, and D. Rogers of San Onofre Nuclear Generating
Station, San Clamente, CA.
2. D. H. Wolpert, "Stacked Generalization"' Neural Networks, Vol. 5, pp. 241-259, 1992.
3. E.B. Bartlett. 1990 "Nuclear Power Plant Status Diagnostics Using Simulated Condensa-
tion: An Auto-Adaptive Computer Learning Technique," Ph.D. Dissertation, The University
of Tennessee. Knoxville.
4. R. Hecht-Nielsen, 1989 "Theory of the Backpropagation neural Network," International
Joint Conference on Neural Networks ([JCNN), vol. I, pp. 593-605.
5. D. H. Wolpert. "A Mathematical Theory of Generalization Part [". Complex Systems. 4.
151-200, 1990.
6. D. H. Woipert, "A Mathematical Theory of Generalization Part II", Complex Systems. 4.
200-249, 1990d.
7. E. B. Bartlett and R. Uhrig, "Nuclear Power Plant Status Diagnostics Using an Artificial
Neural Networks"' Nuclear Technology, Vol. 97, March 1992.
8. Li, Ker-Chau(1985). "From Stein's Unbiased Risk Estimates to The Method of Generalized
Cross-Validation". The Annals of Statistics, 13, 1352-1377.
9. E.B. Bartlett and P,.E. Uhri$, 1991, "A Nuclear Power Plant Status Diagnostics Using Ar-
tificial Neural Networks," AI 91: Frontiers in Innovative Computing for the Nuclear Industry,
American Nuclear Society, September 199:1.

514
 r _1

=m