Annals of Nuclear Energy 90 (2016) 433–440

Contents lists available at ScienceDirect

Annals of Nuclear Energy

journal homepage: www.elsevier.com/locate/anucene

Fault diagnosis of Pakistan Research Reactor-2 with data-driven

techniques
Farhan Jamil a, Muhammad Abid a,⇑, Inamul Haq a, Abdul Qayyum Khan a, Masood Iqbal b
a
Pakistan Institute of Engineering & Applied Sciences, P.O. Nilore, Islamabad, Pakistan
b
NED, PINSTECH, P.O. Nilore, Islamabad, Pakistan

a r t i c l e i n f o a b s t r a c t

Article history: In nuclear reactors, safety is of prime importance in their operation. Fault detection and isolation (FDI)
Received 13 August 2015 methods are making their applications to improve safety, reliability and availability of nuclear reactors.
Received in revised form 3 December 2015 Among various FDI techniques, data-driven techniques are best suited for fault diagnosis of nuclear reac-
Accepted 11 December 2015
tors because process data is available to sensors, both in normal operation and under faulty conditions.
Available online 5 January 2016
Among data-driven techniques, principle component analysis (PCA) and Fisher discriminant analysis
(FDA) have been successfully applied to many industrial processes. In this paper, PCA and FDA are applied
Keywords:
for fault detection and fault isolation in Pakistan Research Reactor-2 (PARR-2) for known faults of control
Research reactors
Fault diagnosis
rod withdrawal and external reactivity insertion. PCA model is developed using training data set obtained
Principle component analysis during normal operation of PARR-2. It is then applied to test data set collected from the reactor during
Fisher discriminant analysis control rod withdrawal fault and external reactivity insertion fault. Likewise, FDA model is constructed
for the above mentioned faults using the training data set and applied to the test data for fault isolation.
The results demonstrate that PCA is successful in detection of both the faults. Additionally, FDA not only
detects faults, but it is also successful in isolation/localization of the two faults in PARR-2.
Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction conflicting indications. The operators were unable to recognize and

With depletion in conventional energy resources, scientists and
engineers are looking for alternate energy resources. Among these,
nuclear energy has become a major contributor at 10.9 percent,
fourth behind coal at 40.4 percent, natural gas at 22.5 percent,
and hydro-power at 16.2 percent worldwide (International
Energy Agency, 2014). Nuclear energy is not only cheap but also
environment friendly. Compared to conventional energy resources,
where major cost in energy production is fuel cost, in nuclear
energy 73 percent of cost is incurred on operation and mainte-
nance (Hashemian, 2011). It has been shown that the operation
and maintenance cost, and therefore the per unit energy cost,
can be significantly reduced by introducing modern fault diagnosis
methods to the nuclear power plants. In addition to economics,
safety of nuclear reactors is of utmost importance. The safety and
reliability of nuclear reactor got national concern during the Three
Mile Island incident of 1979 (Zhao, 2005). During that incident,
operators were faced with a large number of alarms coupled with
⇑ Corresponding author.
E-mail addresses: msne1324@pieas.edu.pk (F. Jamil), mabid@pieas.edu.pk
(M. Abid), fac185@pieas.edu.pk (I. Haq), aqkhan@pieas.edu.pk (A.Q. Khan),
masood@pinstech.org.pk (M. Iqbal).
prioritize the alarms and indications efficiently, thus complicating
the recovery procedures. So, this has drawn the attention of
researchers towards the development of fault diagnosis methods
for safe and reliable operation of nuclear reactors (Park and
Seong, 2002; Bartlett and Uhrig, 1992; Ohga and Seki, 1993).
The classical approach used for fault diagnosis in nuclear reac-
tors is by hardware redundancy, that is, multiple sensors are uti-
lized to measure the same quantity and a voting scheme gives
the information of faulty sensor. However, hardware redundancy
can not be added to other components of nuclear reactor.
Therefore, to detect faults in these components, limit checking is
performed, the measurements are monitored to see if these are
within pre-defined limits, in case the measurements exceed the
bounds, that implies the presence of some fault. Limit checking
technique has the disadvantage that fault is only detected if it
grows enough to cause some variable to exceed its bound. This dis-
advantage may be overcome by incorporating modern and more
advanced fault diagnosis techniques which are based on software/
analytical redundancy. In many industrial processes, the hardware
redundancy is replaced by analytical redundancy which not only
increases efficiency of fault diagnosis scheme but also reduces
the cost. However, for safety critical processes, like nuclear reac-
tors, software redundancy based fault diagnosis techniques should

http://dx.doi.org/10.1016/j.anucene.2015.12.023
0306-4549/Ó 2015 Elsevier Ltd. All rights reserved.
434 F. Jamil et al. / Annals of Nuclear Energy 90 (2016) 433–440

be used in complementary with hardware redundancy based which is a low power miniature neutron source research reactor.
techniques. Detailed study is carried out to analyze possible faults in PARR-2.
Major and well-established software redundancy methods that Real data is collected from available sensors in the reactor and
are used for fault detection and isolation (FDI) include analytical two most commonly studied data driven techniques, that is, prin-
model-based, knowledge-based and data driven techniques. In ciple component analysis (PCA) and Fisher discriminant analysis
analytical-model based technique, a mathematical model is (FDA) are applied for fault diagnosis of the reactor. Control rod
obtained, usually by using the first principles, to represent normal (CR) withdrawal fault and external reactivity (ER) insertion fault
behavior of system to be monitored. A residual is then generated are successfully detected by PCA. Furthermore, Fisher discriminant
by comparing observed behavior of actual system with the pre- analysis (FDA) is successfully applied to detect and isolate the
dicted behavior from the mathematical model. A nonzero residual above mentioned faults in PARR-2. A very important feature of this
signal gives the information about occurrence of fault(s) in the sys- manuscript is that fault diagnosis techniques are applied to data
tem (Abid, 2010). These techniques are only applicable to small collected directly from PARR-2, compared to many other papers
systems which can be accurately modeled. In knowledge-based which are either based on simulations or on data collected from
techniques, model is constructed from qualitative information reactor simulators.
about the system. Like, analytical model-based techniques, The paper is organized as follows. Section 2 describes Pakistan
knowledge-based FDI techniques have limited applications to large Research Reactor-2 (PARR-2), its characteristics and possible faults
industrial systems because extensive information is required to that may occur in the reactor. Review of FDI methods including
construct qualitative model which is cumbersome (Chiang et al., analytical model-based approach, knowledge-based and data dri-
2001). In contrast, data-driven techniques are more convenient ven approach with particular focus on principle component analy-
and efficient because there we have to monitor the input–output sis (PCA) and Fisher discriminant analysis (FDA) is presented in
information from sensors without going into details of the system. Section 3. In Section 4, PCA and FDA techniques are applied for
Therefore, data-driven techniques are well suited for fault diagno- fault diagnosis of PARR-2 and results are discussed. Section 5 pre-
sis of large industrial systems. Among the most studied data driven sents a brief summary and concluding remarks for the paper.
methods are artificial neural networks (ANN), genetic algorithms
(GA), multivariate statistical techniques including support vector
2. Pakistan Research Reactor-2
machines (SVM), principle component analysis (PCA) and Fisher
discriminant analysis (Elshenawy et al., 2009; Isermann, 2006;
Pakistan Research Reactor-2 (PARR-2) is a low power (30 kW)
Chiang et al., 2001; Ding et al., 2014).
miniature neutron source reactor (MNSR), tank-in-pool type
Although, a lot of work has been carried out for fault diagnosis
research facility. It is cooled, moderated and shielded by deminer-
of many industrial processes, including aerospace industry, chem-
alized light water. The core is surrounded by beryllium (Be) reflec-
ical industry, automobile industry, etc., and modern fault diagnosis
tor in order to minimize neutron leakage. It uses ninety percent
techniques are successfully applied (see, for example, Isermann
enriched U-235 based fuel. It has 344 fuel rods and one central oval
(2011), Liu et al. (2014), and Schwung et al. (2015)), only a little
shaped rod made of cadmium which serves the purpose of control
attention has been paid to diagnosis of nuclear reactors. There
rod. The core is enclosed in aluminum vessel suspended in an
are only a few manuscripts addressing the application of modern
underground pool as shown in Fig. 1. Long term reactivity compen-
fault diagnosis techniques for nuclear reactor. Since it is hard to
sation is achieved by increasing the thickness of top beryllium
obtain a detailed and accurate model for nuclear power systems,
reflector. Reactor has ten irradiation sites, five of which are present
a limited number of publications are available on model-based
inside the beryllium annulus while the others surround the reac-
FDI for nuclear reactors (Ablay and Aldemir, 2013; Pang and Xia,
tor. These sites are assessed through pneumatic sample transfer
2014; Singer et al., 1997). Data-driven techniques are more suit-
tubes. Fission heat generated in the core is first removed by natural
able for fault diagnosis of nuclear reactors because heavy instru-
convection and then transferred to pool water, which serves as
mentation is there to obtain input–output data. There are a few
heat sink (Bokhari and Pervez, 2010; Mahmood et al., 2008).
articles addressing the application of data-driven techniques for
The reactor has inherent safe characteristics of under moder-
monitoring of nuclear reactors. Artificial neural networks (ANN)
ated core array and low excess reactivity. Under-moderated design
are applied for detection of a total instantaneous blockage fault
feature leads to negative temperature coefficient of reactivity, that
of a sodium-cooled fast reactor in Martinez-Martinez et al.
is, reactivity decreases with increase in temperature (Iqbal et al.,
(2015) and to propose accident diagnosis advisory system for
nuclear power plants in Lee et al. (2005). Distributed fault diagno-
sis system in nuclear power plants, based on fuzzy neural network
and data fusion, is proposed in Yong-kuo et al. (2013) and its effi-
ciency is improved in Liu et al. (2014). Three types of sensor fault
signals were simulated using modified ensemble empirical mode
decomposition and probabilistic neural network in Yu et al. (2015).
Application of genetic algorithms for fault diagnosis of nuclear
power plants is proposed in Yangping et al. (2000), the simulations
are carried out on nuclear power plant simulator data to detect sev-
eral faults by observing 62 sensors. Detection of sensor faults and
component faults in nuclear power plants using SVM is presented
in Zavaljevski and Gross (2000) and Liu et al., 2013). Application of
PCA is applied to detect faults in nuclear power plants is studied
in Dunia et al. (1996), Upadhyaya et al. (2003), Kaistha and
Upadhyaya (2001), and Lu and Upadhyaya (2005). A comprehensive
study of these approaches with relevance to nuclear power plants is
presented in review paper (Ma and Jiang, 2011).
The contribution of this paper is application of data driven tech-
niques for fault diagnosis of Pakistan Research Reactor-2 (PARR-2) Fig. 1. PARR-2 core suspended in an underground pool (Iqbal et al., 2002).
 F. Jamil et al. / Annals of Nuclear Energy 90 (2016) 433–440
2002). The negative temperature coefficient of reactivity limits the
power excursion due to impulsive addition of reactivity. The
impulsive addition of reactivity leads to power surge and hence
increases the temperature of fuel. Heat transfer from fuel to cool-
ant will increase the coolant temperature. Negative moderator
coefficient of reactivity then plays its role and decreases the excess
reactivity. Thus the reactor power is reduced which limits its fuel
temperature. The excess reactivity of the cold clean core is less
than the effective delayed neutron fraction. Therefore, the possibil-
ity of prompt critical accident is eliminated.
There are several possible faults that may appear in the reactor,
a detailed study can be found in final safety analysis report for the
reactor. In this manuscript, due to safety limitations, only below
mentioned faults have been addressed.
2.1. Full withdrawal of control rod
Control rod is used in positive reactivity control of the system.
Control rod remains at critical position during the safe operation
of the reactor. If, due to operator error or control system malfunc-
tion, the control rod is fully withdrawn and stayed there during
reactor operation, then reactivity will be inserted into the system.
This reactivity will cause an increase in reactor power but due to
inherent safety feature, power will be limited to 87 kW. This power
rise will lead to increase in temperature difference across the core.
2.2. Accidental external reactivity insertion
In this accident, it is considered that if a neutron absorbing sam-
ple is present in the core and accidentally removed from reactor
core through rabbit system, then reactivity will be inserted into
the system. This reactivity will bring a power increase and flux will
continue to rise depending upon the amount of fissile material.
Fuel and clad temperature will also increase due to which coolant
temperature will rise.
3. Review of fault diagnosis techniques
Fault is any un-permitted deviation of a characteristic property
or parameter of a system/component from its acceptable/usual/
standard condition. Fault detection is determining whether a fault
has occurred in the system under consideration. Following fault
detection, fault isolation is determining the location of the fault.
Fault identification is identifying the size and the time of onset
of the fault (Isermann and Ballé, 1997). Fault diagnosis contains
both fault detection and fault isolation. Fault detection and isola-
tion (FDI) methods continuously monitor a system during opera-
tion, which is called online monitoring (OLM).
The earliest approach, which is still most commonly used, for
fault diagnosis is by utilizing hardware redundancy, that is, multi-
ple sensors are utilized to measure the same quantity and a voting
scheme can give information about faulty sensor. This technique,
however, suffers from many shortcomings; hardware redundancy
is associated with additional cost, weight and size of the system
and the most important is that although redundancy can be added
to sensors, but it cannot be added to process components, which
means that component faults cannot be directly detected. These
shortcomings are overcome by soft/analytical redundancy
techniques.
There are a number of analytical redundancy techniques that
have been successfully applied for fault detection of industrial pro-
cesses, a detailed study of these techniques may be found in Ding
(2013) and Isermann (2006). In analytical model-based techniques,
mathematical models often made from first principles are utilized
435
to describe the behavior of system to be monitored. Then a symp-
tom signal, called residual signal, is generated by comparing the
measured output of the process with the estimations generated
by the mathematical model, a non-zero residual gives indication
about the occurrence of fault. The analytical model-based
approaches have an advantage in the sense that these give physical
understanding of the process in process monitoring procedures.
These approaches are applicable to systems in which enough infor-
mation and satisfactory models are available. For systems, where
detailed analytical model is not available, knowledge based tech-
niques which are built upon qualitative models can be utilized.
Both of these model-based techniques (analytical model and qual-
itative model) are mostly applicable to systems which have a small
number of inputs, outputs and states and difficult to apply for large
scale system which have large number of inputs, outputs and/or
states. For large scale systems, a huge amount of time and efforts
are required to build detailed models in order to be effective.
In contrast to analytical model and qualitative model based
techniques, data driven techniques are modeled directly from pro-
cess data that is collected from sensors. Most industrial systems,
whether they are nuclear plants or chemical industries, are large
scale and complex systems with heavy instrumentation. They pro-
duced high dimensional correlated data which contains much
information about the processes. The idea of data driven tech-
niques is based on transformation of high dimensional correlated
data into low dimensional uncorrelated form by capturing signifi-
cant information. A test statistics is then applied on the low dimen-
sional data to determine the existence of faults. Among data-driven
techniques, principal component analysis and Fisher discriminant
analysis are most widely studied and successfully applied to many
industrial systems (see for example Yin et al. (2015) and references
therein). These techniques are briefly described in below.
3.1. Principal component analysis
Principal component analysis (PCA) is a basic method of multi-
variate analysis which serves as a linear dimensionality reduction
technique while retaining maximum possible variation in original
data set. It considers correlation among variables and helps in
identifying the most affected variables and the variables most
responsible for the occurrence of faults. In addition, it separates
the observation space into information rich subspace and a sub-
space having random noise. A fault in a process may affect either
of the above two mentioned subspaces, so sensitivity of fault
monitoring process can be increased by observing these subspaces
separately. In lower dimensional space representation, various sta-
tistical plots, for example, T 2 or Q charts are used for visualization
of process trends. PCA has been widely used in many research
areas including pattern classification, feature extraction, data
compression and process monitoring (Yin et al., 2015). Due to its
simplicity and successful application in processing large amount
of process data, it has become a powerful tool for fault detection.
The procedure to apply principal component analysis (PCA) for
fault detection consists of three steps, firstly, offline computations
are performed on training data collected from the process to be
monitored to determine a set of transformation vectors, called
loading vectors; secondly, the computed loading vectors are used
to transform online data into a lower dimensional space and,
thirdly, test statistics is used to determine the occurrence of fault.
To understand the procedure, consider a training set of n observa-
tions for each of m process variables is normalized to zero mean
and unit variance, by subtracting each variable by its sample mean
and dividing by standard deviation, and is represented in the form
of input data matrix X
Rn
m ,
436 F. Jamil et al. / Annals of Nuclear Energy 90 (2016) 433–440

2 3 Pn
x11 x12 : : : x1m where h0 ¼ 1  2h3h1 2h3 ; hi ¼ r2ij and ca is normal deviation cor-
j¼aþ1
6x x22 : : : x2m 7
2
6 21 7 responding to ð1  aÞ percentile. a is the level of significance. A fault
6 7
6 : : : : : : 7 will be detected whenever Q-statistic exceeds its threshold Q a , that
X¼ 6
6 :
7
6 : : : : : 77 is,
6 7
4 : : : : : : 5 Q 6 Q a Fault  free ð9Þ
xn1 xn2 : : : xnm Q > Q a Fault ð10Þ
The loading vectors are computed by singular value decomposi-
tion (SVD) of input data matrix X,
3.2. Fisher discriminant analysis
1
pffiffiffiffiffiffiffiffiffiffiffiffi X ¼ U RV T ð1Þ
n1 PCA is optimal dimensionality reduction technique in the sense
that it captures maximum information (variance) in the original
where U
R and V
R
n
n
are unitary matrices and R
R
m
m
is the n
m
data. However, while computation of transformation vectors, it
diagonal matrix having non-negative real singular values in does not consider the information among data classes in the obser-
decreasing magnitude ðr1 P r2 P    P rm P 0Þ. The loading vec- vation space. This problem is incorporated by Fisher discriminant
tors are the orthonormal column vectors of matrix V. The variance analysis (FDA) in which data collected under various fault condi-
of the training set projected along the ith column of V is equal to tions is classified into different classes, where each class represents
r2i . The loading vector corresponding to a largest singular values a particular fault. Fisher discriminant analysis (FDA) reduces the
are retained in order to capture the data variation while reducing dimensions of all data classes simultaneously, such that the
the effect of noise in it. If P
Rm
a is the matrix having first a col- reduced space is favorable in discriminating among these data
umns of loading vectors corresponding to a largest singular values, classes. It calculates a set of linear transformation vectors, called
the projections of observation space X into lower dimensional space FDA vectors, which maximize the distance among different classes
called score matrix T is given as, while minimizing the distance within a class in the projected
T ¼ XP ð2Þ space. Hence, this method is superior to PCA for fault isolation.
The procedure to apply FDA for fault isolation is quite similar to
After projecting the data in lower dimensional space, multivari- that for PCA. A set of transformation vectors, called FDA vectors, is
ate Hotteling’s T 2 -statistics and square prediction error (SPE) or computed from offline data. Compared to PCA, where only normal
Q-statistics are used for detection of faults in a process. Hotteling’s operation data is generally required, for FDA training data of all
T 2 -statistics can be computed as, classes including data associated with normal operation as well
as data associated with each fault is needed. FDA vectors are uti-
T 2 ¼ xT PR1 T
a P x ð3Þ lized to project online data into a lower dimensional space and
where P is the matrix of loading vectors, related with first a singular finally a discriminant function is used to isolate fault. Consider a
values, containing significant amount of variance and Ra is the process with m variables and n observations for each variable
corresponding singular value diagonal matrix. Given a level of stacked in the matrix X
Rn
m ; p as the number of classes under
significance a, the appropriate threshold values for the different faults and nj number of observation in jth class. Let xi rep-
T 2 -statistics can be computed as, resent the transpose of ith row of stacked matrix X. From this train-
ing data, the transformation vector v is computed such that the
mðn  1Þðn þ 1Þ following optimization problem is solved.
T 2a ¼ F a ðm; n  mÞ ð4Þ
nðn  mÞ
j v T Sb v j
where F a ðm; n  mÞ is the F-distribution with m and ðn  mÞ degree J FDA ðv Þ ¼ arg max ; ð11Þ
v –0 jv T Sw v j
of freedom (MacGregor and Statistical, 1995). A fault will be
detected if the T 2 -statistics exceeds a threshold, that is, where Sw is the within class scatter matrix obtained by

T 2 6 T 2a Fault-free ð5Þ X
p
Sw ¼ Sj ; ð12Þ
T 2 > T 2a Fault ð6Þ j¼1

P
The T 2 -statistics is based on loading vectors corresponding to with Sj ¼ nxi 2vj ðxi  xj Þðxi  xj ÞT , and the mean of jth class
first a largest singular values present in Ra . So, it is oversensitive P
xj ¼ n1 x 2v xi . Furthermore, Sb is the between class scatter matrix
j i j
to smaller singular values when inverse is used in calculating
2
defined by
T -statistics model. Consequently, Q-statistics is used to overcome
this problem by considering the observation space corresponding X
p
Sb ¼ ðxj  xÞðxj  xÞT ; ð13Þ
to ðm  aÞ singular values.
i¼1
 
Q ¼ rT r; r ¼ I  PP T x ð7Þ P
with 
x representing the total mean vector defined by x ¼ 1n ni¼1 xi . It
has been shown that solution to the above optimization problem is
where r is projection of observation x into the residual space. So, Q-
equivalent to following generalized eigenvalue decomposition
statistics measures the total sum of variations in the residual space
problem (Chiang et al., 2001),
and it does not face the problem of over sensitivity to inaccuracies
in smaller singular values. The distribution for Q-statistic, also Sb v k ¼ k k Sw v k ; ð14Þ
called the squared prediction error (SPE), can be computed by:
where kk is generalized eigenvalue indicating degree of separability
pffiffiffiffiffiffiffiffi !1=2
h0 ca 2h2 h2 h0 ðh0  1Þ between classes and v k are corresponding eigen-vectors. The
Q a ¼ h1 þ1þ ð8Þ optimization problem (11) ensures that there is maximum scatter
h1 h21
between classes while the within class scatter is minimum, this
 F. Jamil et al. / Annals of Nuclear Energy 90 (2016) 433–440 437

feature helps to easily isolate the faults. For projecting online data 4.1. Monitoring of PARR-2 using PCA
into a lower dimensional space, a matrix V p
Rm
p1 with p  1
FDA vectors as columns is defined, such that the projected data PCA implementation is illustrated using input data set of PARR-
zi
Rp1
is described by; 2. For computation of PCA loading vectors, the data set consists of
70 observations were recorded for each variable during normal
zi ¼ V Tp xi : operating conditions with a sampling time of 1 s. PCA algorithm
as described in Section 3 is applied to this data set. The SVD of
Finally, for fault detection/isolation, following discriminant the data set gives the singular values and their corresponding sin-
function is most commonly used (Chiang et al., 2001). gular vectors.
 1 2 3
1 1 1:90 0 0 0 0 0
g j ðxÞ ¼  ðx  xj ÞT V p V Tp Sj V p V Tp ðx  xj Þ þ lnðpi Þ 6
2 nj  1 6 0 1:13 0 0 0 0 7
7
   6 7
1 1 6 0 0 1:09 0 0 0 7
 ln det V Tp Sj V p : ð15Þ K¼6
6
7
2 nj  1 6 0 0 0 0:86 0 0 7
7
6 7
4 0 0 0 0 0:55 0 5
where g j ðxÞ is the discriminant function associated with class j
0 0 0 0 0 0:47
given a data vector x
Rm and the online data belongs to class i if
the discriminant function corresponding to ith class is maximum, 2 3
that is, 0:23 0:37 0:74 0:16 0:30 0:37
6 0:46 0:37 0:16 0:53 0:37 0:45 7
6 7
g i ðxÞ > g j ðxÞ class i fault has occurred ð16Þ 6 7
6 0:52 0:13 0:46 0:00 0:24 0:67 7
V ¼6
6 0:47 0:16
7
PCA, being optimal dimensionality reduction technique, gener- 6 0:24 0:64 0:27 0:46 7
7
6 7
ates best results for fault detection, however, it is not suitable for 4 0:41 0:53 0:01 0:43 0:60 0:00 5
fault isolation, as stated earlier. FDA can not only detect faults,
0:29 0:63 0:38 0:31 0:53 0:02
but is also superior to PCA for fault isolation. The only problem
with FDA is that, for training of FDA, data corresponding to normal We have retained four principal components, that is a ¼ 4, then
4:98

operation as well as under faulty conditions is required. This limits 6:00

¼ 83% of total variance is captured by first four principal
the usability of FDA. Therefore, the two techniques may be used to components. The loading vector P contains score space correspond-
complement each other, that is, PCA may be used for fault detec- ing to these four singular values.
tion and FDA may be used for isolation of the faults for which train- 2 3
ing data is available. Furthermore, for processes where only small 0:23 0:37 0:74 0:16
6 0:46 0:37 0:16 0:53 7
set of fault training data is available, recently proposed techniques 6 7
6 7
like exponential discriminant analysis (Adil et al., 2016) may be 6 0:52 0:13 0:46 0:00 7
P¼ 6
6 0:47 0:16
7
utilized for fault isolation.
6 0:24 0:64 77
In next section, PCA and FDA are applied for fault diagnosis of 6 7
4 0:41 0:53 0:01 0:43 5
Pakistan Research Reactor-2.
0:29 0:63 0:38 0:31

4. Applications of FDI methods in PARR-2 After computation of loading vectors, online data obtained from
sensors was utilized for computation of T 2 -statistics. Control rod
As stated earlier, the first step in application of principal com- withdrawal fault and external reactivity insertion fault were intro-
ponent analysis and Fisher discriminant analysis for fault diagnosis duced for 70 s as described at the start of this section. Fig. 2 shows
is offline computation of transformation vectors from training data
collected from sensors attached to process to be monitored. For
application of PCA, only normal operation training data is needed.
Whereas, for FDA to isolate faults, fault-free data as well as data 60
after occurrence of fault is required for computation of FDA vec-
tors. In PARR-2, there is heavy instrumentation and a large number 50
of measurements are available and accessible through computer
Inlet Temp (°C)
interface. For our study, data is collected from sensors for inlet
40 Outlet Temp (°C)
temperature, outlet temperature, pool temperature, coolant con-
Pool Temp (°C)
ductivity, reactivity and neutron flux. Reactor was operated at full
Measurements

Pool Conductivity (uS/cm)

power of 30 kW and after steady state condition was achieved, 70 30 Reactivity (mk)
observations were recorded for each variable during normal oper- Flux (#/1E11−cm2−s)
ating conditions with a sampling time of 1 s. This data represents
20
the class 1 data (fault free data). Then control rod withdrawal fault
was introduced by uplifting the control rod up to 15 percent from
its critical position and 70 measurements were taken. This data 10
represents the class 2 data (control rod withdrawal fault). After
CR Withdrawal Fault Data
that, this fault was cleared and steady state conditions were
0
achieved then external reactivity insertion fault was introduced
by inserting 1mk reactivity through pneumatic system and again Normal Data ER Insertion Fault Data
70 samples were recorded that represent class 3 data (external −10
0 50 100 150 200
reactivity insertion fault). The reactor remains within its permissi- Number of Observations
ble limits of power and temperature by insertion of these two
faults. Fig. 2. PARR-2 sensors output data.
438 F. Jamil et al. / Annals of Nuclear Energy 90 (2016) 433–440

150 150
2 Q for faulty case
T for faulty case
2
Ta Threshold Q Threshold

100 100
2

Q
T

50 50

0 0
0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140
Samples Samples

Fig. 3. T 2 -statistics for control rod withdrawal fault detection. Fig. 5. Q -statistics for control rod withdrawal fault detection.

detected as it crosses its threshold value. It can also be noted that

a plot of online data under different fault conditions before the
application of PCA. The randomness of the data made it impossible
to predict the fault just by looking at it.
Figs. 3 and 4 illustrate T 2 charts for control rod withdrawal fault
and external reactivity insertion fault respectively with their
threshold. For a ¼ 4 and a ¼ 98%, threshold comes out to be
T 2a ¼ 13:82 as defined in Section 3. It can be seen from Fig. 3 that
as soon as control rod withdrawal fault appears in the system, it
is detected. This fault appears after 70 samples, so large variations
occur in T 2 chart at that moment, which indicates that a fault has
occurred in the system and its value crosses the threshold.
Similarly, when external reactivity insertion fault appears in the
system after 70 samples as in Fig. 4, T 2 charts remains above the
threshold showing that fault is present in the system.
Figs. 5 and 6 illustrate Q charts for control rod withdrawal fault
and external reactivity insertion fault respectively with their
threshold. For a ¼ 4 and a ¼ 98%, threshold comes out to be
Q a ¼ 3:98 as defined in Section 3. The Q chart in these figures
shows that something went wrong after 70 samples and fault is
some points of Q-statistic, called false alarm, exceed the threshold
value during normal operation due to low value of level of
significance (a ¼ 98%). Hence, the lower threshold leads to 2 per-
cent false alarm. Both T 2 and Q charts are efficient in detection of
faults as soon as they appear in the system but cannot identify
the type of fault which has appeared, so we move towards FDA
technique which serves both of the purpose of fault detection
and isolation.
4.2. Monitoring of PARR-2 using FDA
Fisher discriminant analysis is applied for fault detection and
isolation in PARR-2. For computation of FDA vectors, the training
data consist of data related to each class, that is, the fault-free data,
full control rod withdrawal fault and accidental external reactivity
insertion fault. The data set collected from the reactor, as described
in the beginning of this section, is utilized for that purpose. The
within-class scatter and between class scatter matrices are com-
puted as:

150 150
2 Q for faulty case
T for faulty case
2
Ta Threshold Q Threshold

100 100
2

Q
T

50 50

0 0
0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140
Samples Samples

Fig. 4. T 2 -statistics for external reactivity insertion fault detection. Fig. 6. Q -statistics for external reactivity insertion fault detection.
 F. Jamil et al. / Annals of Nuclear Energy 90 (2016) 433–440 439

2 3
128:22 70:58 59:04 13:87 57:04 22:43 2
6 70:58 75:93 59:62 13:59 60:55 27:34 7 Normal Data
6 7
6 7 1.5 CR Withdrawal Fault Data
6 59:04 59:62 100:68 11:93 62:98 22:40 7
Sw ¼ 6
6 13:87 13:59 11:93 20:85
7 External Reactivity Insertion Fault Data
6 14:82 6:24 7
7
6 7 1
4 57:04 60:55 62:98 14:82 82:12 24:88 5
22:43 27:34 22:40 6:24 24:88 12:32 0.5

z2
2 3 0
20:78 23:96 12:87 18:89 17:16 20:97
6 23:96 73:07 27:98 92:79 68:73 96:63 7
6 7 −0.5
6 7
6 12:87 27:98 48:32 55:23 35:48 55:29 7
Sb ¼ 6
6 18:89 92:79 55:22
7
92:08 132:29 7
−1
6 128:15 7
6 7
4 17:16 68:73 35:48 92:08 66:88 95:34 5 −1.5
20:97 96:63 55:29 132:29 95:34 136:68
−2
Solving Eq. (14), we obtain the FDA vectors that are eigen- −2 −1.5 −1 −0.5 0 0.5 1 1.5
z1
vectors wk of this eigenvalue problem.
2 3 Fig. 7. Projection of training data onto the first two FDA vectors.
0:03 1:00 0:21 0:18 0:73 0:20
6 0:22 0:36 0:53 0:45 1:00 0:27 7
6 7
6 7
6 0:01 0:58 0:79 0:02 0:10 0:17 7 100
Wk ¼ 6
6 0:08
7
6 0:00 0:02 0:63 0:22 0:03 7
7
Normal Operation
6 7 50 CR Withdrawal Fault
4 0:19 0:77 1:00 0:14 0:46 1:00 5 External Reactivity Insertion Fault
1:00 0:40 0:75 1:00 0:66 0:58 0
which correspond to respective eigenvalues;
−50
2 3
57:21 0 0 0 0 0
g(x)

6 0 0:00 0 0 0 0 7 −100
6 7
6 7
6 0 0 0:76 0 0 0 7
kk ¼ 6
6
7 −150
6 0 0 0 0:00 0 0 7
7
6 7
4 0 0 0 0 0:00 0 5 −200
0 0 0 0 0 0:0
−250
The eigenvalues matrix shows that only two eigenvalues are
nonzero i.e. k1 ¼ 57:21 and k3 ¼ 0:76 as we have three class of data.
−300
So, the eigen-vectors associated to these nonzero eigenvalues are: 0 50 100 150 200
2 3 Number of Observations
0:03 0:21
6 0:22 0:53 7 Fig. 8. Evaluation of discriminant function under CR withdrawal & external
6 7
6 7 reactivity insertion fault.
6 0:01 0:79 7
Wa ¼ 6
6 0:08
7
6 0:02 7
7
6 7 These results have demonstrated that PCA has successfully
4 0:19 1:00 5
detected faults and FDA has successfully isolated faults in
1:00 0:75
PARR-2. Therefore, these techniques are very well suited for mon-
Fig. 7 shows the transformation of three training data classes itoring of research reactors.
onto first two discriminant vectors. It is obvious from the figure
that the training data classes, when projected using FDA vectors,
are very well separated from one another in the projected space, 5. Summary
so, fault can easily be isolated. There is some overlap between nor-
mal class data and control rod withdrawal data that is called miss- Safety and reliability of a nuclear reactor can be increased by
classification. application of software redundancy based fault diagnosis tech-
After computation of FDA vectors, the online data is tested by niques in complementary to conventional hardware redundancy
computation of the discriminant function for each class, the plot based techniques. Nuclear reactors are large complex engineering
of these discriminant functions are shown in Fig. 8. It can be seen systems and it is difficult to obtain their accurate mathematical
from Fig. 8 that up to 70 samples, discriminant function corre- model, therefore, data driven techniques are best suited software
sponding to the normal class has the maximum value relative to redundancy techniques for fault diagnosis of nuclear reactors.
others, which implies that there is no fault in the system. After Principal component analysis is optimal dimensionality reduction
70 samples, discriminant function of control rod withdrawal data technique and very efficient in fault detection. However, it does
class has maximum value which indicates that now control rod not take into account the information between various classes of
withdrawal fault has occurred in the system. After 140 samples, faults, therefore, it does not generate good result for fault isolation.
discriminant function of external reactivity insertion class data Fisher discriminant analysis is also a dimensionality reduction
has maximum value which means that now this fault has hap- technique and it considers the information between classes while
pened in the system. calculating its transformation matrix and generates maximum
440 F. Jamil et al. / Annals of Nuclear Energy 90 (2016) 433–440
scatter between classes and minimum scatter within class data.
Thus, it has advantages over PCA in fault isolation.
In this paper, PCA and FDA techniques are applied for fault diag-
nosis of Pakistan Research Reactor-2 (PARR-2) which is a low
power miniature neutron source reactor. Two known faults, that
is, control rod withdrawal fault and external reactivity insertion
fault are considered. A training data, comprised of 6 parameters
and 70 measurements for each fault, is obtained from sensors of
the reactor and processed for computation of PCA and FDA models.
Then test data obtained from the reactor, under normal operating
conditions and in the presence of the two faults, is tested with
PCA and FDA. The results demonstrate that the two faults are
successfully detected with PCA and are very well isolated with
FDA. Therefore, PCA and FDA are well suited for monitoring of
PARR-2.
References
Abid, M., July 2010. Fault detection in nonlinear systems: an observer-based
approach (Ph.D. thesis), University of Duisburg-Essen, Duisburg, Germany.
Ablay, G., Aldemir, T., 2013. Fault detection in nuclear systems using sliding mode
observers. Nucl. Sci. Eng. 173 (1), 82–98.
Adil, M., Abid, M., Khan, A.Q., Mustafa, G., Ahmed, N., 2016. Exponential
discriminant analysis for fault diagnosis. Neurocompuring 171, 1344–1353.
Bartlett, E.B., Uhrig, R.E., 1992. Nuclear power plant status diagnostics using an
artificial neural network. Nucl. Technol. 97 (3), 272–281.
Bokhari, I.H., Pervez, S., 2010. Safety analysis for core conversion (from HEU to LEU)
of Pakistan Research Reactor-2 (PARR-2). Nucl. Eng. Des. 240 (1), 123–128.
Chiang, L.H., Braatz, R.D., Russell, E.L., 2001. Fault Detection and Diagnosis in
Industrial Systems. Springer Science & Business Media.
Ding, S.X., 2013. Model-based Fault Diagnosis Techniques: Design Schemes,
Algorithms, and Tools. Springer Science & Business Media.
Ding, S.X., 2014. Data-driven Design of Fault Diagnosis and Fault-tolerant Control
Systems. Springer.
Dunia, R., Qin, S.J., Edgar, T.F., McAvoy, T.J., 1996. Identification of faulty sensors
using principal component analysis. AIChE J. 42 (10), 2797–2812.
Elshenawy, L.M., Yin, S., Naik, A.S., Ding, S.X., 2009. Efficient recursive principal
component analysis algorithms for process monitoring. Ind. Eng. Chem. Res. 49
(1), 252–259.
Hashemian, H.M., 2011. On-line monitoring applications in nuclear power plants.
Prog. Nucl. Energy 53 (2), 167–181.
International Energy Agency, Key world energy statistics, International Energy
Agency, 2014.
Iqbal, M., Abdullah, M., Pervez, S., 2002. Parametric tests and measurements after
shimming of a beryllium reflector in a miniature neutron source reactor
(MNSR). Ann. Nucl. Energy 29 (13), 1609–1624.
Isermann, R., 2006. Fault-diagnosis Systems: An Introduction from Fault Detection
to Fault Tolerance. Springer Science & Business Media.
Isermann, R., 2011. Fault-diagnosis Applications: Model-based Condition
Monitoring: Actuators, Drives, Machinery, Plants, Sensors, and Fault-tolerant
Systems. Springer Science & Business Media.
Isermann, R., Ballé, P., 1997. Trends in the application of model-based fault
detection and diagnosis of technical processes. Control Eng. Pract. 5 (5), 709–
719.
Kaistha, N., Upadhyaya, B.R., 2001. Incipient fault detection and isolation of field
devices in nuclear power systems using principal component analysis. Nucl.
Technol. 136 (2), 221–230.
Lee, S.J., Seong, P.H., 2005. A dynamic neural network based accident diagnosis
advisory system for nuclear power plants. Prog. Nucl. Energy 46 (3), 268–281.
Liu, J., Seraoui, R., Vitelli, V., Zio, E., 2013. Nuclear power plant components
condition monitoring by probabilistic support vector machine. Ann. Nucl.
Energy 56, 23–33.
Liu, Z., Ahmed, Q., Rizzoni, G., He, H., 2014. Fault detection and isolation for lithium-
ion battery system using structural analysis and sequential residual generation.
In: ASME 2014 Dynamic Systems and Control Conference. American Society of
Mechanical Engineers.
Liu, Y.-K., Xie, C.-L., Peng, M.-J., Ling, S.-H., 2014. Improvement of fault diagnosis
efficiency in nuclear power plants using hybrid intelligence approach. Prog.
Nucl. Energy 76, 122–136.
Lu, B., Upadhyaya, B., 2005. Monitoring and fault diagnosis of the steam generator
system of a nuclear power plant using data-driven modeling and residual space
analysis. Ann. Nucl. Energy 32 (9), 897–912.
Ma, J., Jiang, J., 2011. Applications of fault detection and diagnosis methods in
nuclear power plants: a review. Prog. Nucl. Energy 53 (3), 255–266.
MacGregor, J., Kourti, T., 1995. Statistical process control of multivariate processes.
Control Eng. Pract. 3 (3), 403–414.
Mahmood, T., Pervez, S., Iqbal, M., 2008. Neutronic analysis for core conversion
(HEU–LEU) of Pakistan research reactor-2 (PARR-2). Ann. Nucl. Energy 35 (8),
1440–1446.
Martinez-Martinez, S., Messai, N., Jeannot, J.-P., Nuzillard, D., 2015. Two neural
network based strategies for the detection of a total instantaneous blockage of a
sodium-cooled fast reactor. Reliab. Eng. Syst. Saf. 137, 50–57.
Ohga, Y., Seki, H., 1993. Abnormal event identification in nuclear power plants using
a neural network and knowledge processing. Nucl. Technol. 101 (2), 159–167.
Pang, Y., Xia, H., 2014. Fault detection of nuclear reactors by estimation of unknown
input for systems with disturbances. Nucl. Eng. Des. 275, 91–95.
Park, J.H., Seong, P.H., 2002. An integrated knowledge base development tool for
knowledge acquisition and verification for NPP dynamic alarm processing
systems. Ann. Nucl. Energy 29 (4), 447–463.
Schwung, A., Beck, M., Adamy, J., 2015. Fault diagnosis of dynamical systems using
recurrent fuzzy systems with application to an electrohydraulic servo axis,
Fuzzy Sets and Systems.
Singer, R.M., Gross, K.C., Herzog, J.P., King, R.W., Wegerich, S., 1997. Model-based
nuclear power plant monitoring and fault detection: Theoretical foundations,
Tech. rep., Argonne National Lab., IL (United States).
Upadhyaya, B., Zhao, K., Lu, B., 2003. Fault monitoring of nuclear power plant
sensors and field devices. Prog. Nucl. Energy 43 (1), 337–342.
Yangping, Z., Bingquan, Z., DongXin, W., 2000. Application of genetic algorithms to
fault diagnosis in nuclear power plants. Reliab. Eng. Syst. Saf. 67 (2), 153–160.
Yin, S., Li, X., Gao, H., Kaynak, O., 2015. Data-based techniques focused on modern
industry: an overview. Ind. Electron., IEEE Trans. 62 (1), 657–667.
Yong-kuo, L., Min-jun, P., Chun-li, X., Ya-xin, D., 2013. Research and design of
distributed fault diagnosis system in nuclear power plant. Prog. Nucl. Energy 68,
97–110.
Yu, Y., Li, W., Sheng, D., Chen, J., 2015. A novel sensor fault diagnosis method based
on modified ensemble empirical mode decomposition and probabilistic neural
network. Measurement 68, 328–336.
Zavaljevski, N., Gross, K.C., 2000. Sensor fault detection in nuclear power plants
using multivariate state estimation technique and support vector machines. In:
Third International Conference of the Yugoslav Nuclear Society, sn, 2000,
pp. 1–8.
Zhao, K., 2005. An integrated approach to performance monitoring and fault
diagnosis of nuclear power systems (Ph.D. thesis), University of Tennessee,
Knoxville.