INSTITUTE OF TECHNOLOGY, NIRMA UNIVERSITY, AHMEDABAD – 382 481, 08-10 DECEMBER, 2011 1

Fault Detection and Diagnosis in Nuclear Power

Plant - A Brief Introduction
Manit D. Shah

Abstract--This paper reviews fault detection and diagnosis Before we begin further on discussing diagnostic
techniques and their application in nuclear power plant. Fault is approaches, let us see few of the terminologies which are
an unexpected change or malfunction in the system.Fault widely used in literature:
occurring at a point in the plant may propagate further in the -Failure: Permanent disruption of the system in performing its
system. It is necessary to arrest fault propagation for economy
and safety of plant. For this purpose quick detection of an
desired function
unexpected deviation of process variableis necessary. This -Disturbance: An unknown input adversely affecting
function is achieved with the help of Fault detection and operation
diagnosis unit. Diagnostic unit also assists operators in retaining - Irregularity: Intermittent departure from normal response of
normal and safe state of the system. In this paper,I have the system
mentionedthe characteristic requirements of Fault detection and -Abnormality: Fault or failure constitutes an abnormal event
diagnosis unit and the classification of diagnostic methods. With
the help of Fault detection and diagnosis, it is possible to realize
The rest of the paper is organized as follows: Section II
continuous monitoring of plant which enables diagnosis of describes fundamentals of FDD unit, Section III shows
incipient faults and thus prevents unexpected operational requirements of FDD in NPP and various diagnostic
upsets.Here, I have done case study for two approaches, and I approaches employed in NPP and Section IV concludes the
have observed that use of hybrid approaches for Fault detection paper.
and diagnosis yields quick and accurate assistant to the
operators compared to conventional approaches.
II. FAULT DETECTION AND DIAGNOSIS
Index Terms—Fault detection, Fault identification, Fault A. Classification of faults
isolation, Fault diagnosis, Nuclear power plant

I. INTRODUCTION We can classify types of faults based on their physical

location or their effect on system performance.

I n recent times, regulatory control could be automated with

the help of computers and thus, the continuous controller
actionsrequired from the human operators are reduced to
1) Classification based on physical location of fault:
- System faults
- Sensor faults
minimal. With this, a great progress in safety, product quality - Actuator faults
and economy of the plant was observed. 2) Classification based on fault characteristics:
In present era, current challenge for control engineers is the - Additive/Multiplicative faults
automation of abnormal event management (AEM) [1]. AEM - Abrupt/incipient faults
deals with detection and diagnosis of abnormal events and - Permanent/Transient/Intermittent faults
suggesting supervisory actions to operators, for bringing Fault occurring due to leakage in any one of the pipes,
system back in normal state. People in process industries damage in shaft of turbine generator, etc. are considered
view this development as next milestone in control systems. under system faults. Such faults beings changes in dynamic
Due to wide scope of this diagnostic system, various input/output properties of the system. When system under
techniques have been developed over the time. Fault observation faces malfunctioning of any one of the sensors in
detection and diagnosis (FDD) is the process to detect, isolate use, then these are called sensor faults. In this case, sensor
and identify faults in the system. Fault detection is the reading has substantial errors. Due to actuator faults,
determination of faults, present in the system. Fault isolation influence of the controller on plant gets affected. Actuator
is finding type and location of fault, followed by the faults may be jam of pump, stuck of control valve(s), etc.
faultdetection. Fault identification is analyzing the size and Multiplicative fault are observed when fault gets multiplied
time variant behavior of a fault [2], followed by the fault by system gain, while in additive fault system sees addition
isolation. FDD and AEM both have same functions but in of faults. In abrupt faults, sudden change in behavior of the
complex process industry, FDD nomenclature is widely used. system occurs, e.g. a step function; while incipient is gradual
change in fault, e.g. a ramp function. Permanent fault is the
We will also use, in this paper, FDD terminology. Though total failure of the equipment, transient is temporary
early attempts were made for fault diagnosis using diagraphs, malfunctioning while intermittent is the repeated occurrences
fault trees, hardware redundancy and analytical redundancy, of transient faults.
these methods are still used in practical application. In recent Automated FDD heavily relies on measurements taken
times more elegant methods such as expert system, from sensors, since they act as inputs to diagnostic unit.So,
knowledge based method and neural networks, due to two of the values which matters the most during system
advancement in computer technology, are being used. behavior analyses are: Reference value and dynamic reading
2 INTERNATIONAL CONFERENCE ON CURRENT TRENDS IN TECHNOLOGY, „NUiCONE – 2011‟

of the same variable. Actions of diagnosis unit are based on

reference value and reading taken in real-time. Errors in any In process history based method, diagnostic systems use
of above values may result in catastrophic consequences. A available large amount of historical data. These data are
great care has to be taken while selecting above values. transformed and presented to the diagnostic system as a priori
knowledge. The process of data transformation is known as
B. Characteristics requirements of FDD
feature extraction process. This extraction process can be
achievedeither as quantitative or qualitative extraction
For operator to rely on FDD result FDD should suffice few process(Black box or Gray box).
of the qualities, i.e. few of the characteristic requirements [1], Quantitative model based approach can be further
a FDD unit should fulfill are: subdivided into detailed and simplified physical models. For
Quick detection and diagnosis: Early diagnosis of a fault, modeling transient behavior of a dynamic system, detailed
as early as possible should be done. physics model is highly useful.In simplified physicalmodels
Isolation-ability:Ability to distinguish between different lumped parameter approach is generally employed, which is
types of fault. computationally simpler because a space-time coupled
Novel Identifiability: System should identify unknown, equations in partial differential form are transformed into
novel, fault occurring in the operation. ordinary differential and algebraic equations [3].Simplified
Robustness: System should be less prune to noise and physical models generally use explicit mathematical models.
system uncertainty. This enables detection of faulty behavior with an ease.
Adaptability: Diagnostic model should be adaptive to However, quantitative models are complex and difficult to
dynamic behavior of the system. develop[1], [3], [4].
Multiple fault identification: System should identify Unlike quantitative model based approach which uses
multiple faults. quantitative mathematical relationsto represent knowledge of
Explanation facility: Decisions and actions of a diagnostic the system, qualitative models use qualitative relations or
unit should be justified. knowledge bases to represent the state of the system.
Diagnostic method should be able to distinguish model Qualitative model based method can be further classified as
uncertainties, process disturbances and real faults. rule based and qualitative physics based model approaches.
Diagnostic unit in the case of abnormalities come up with Rule based method uses a priori knowledge to derive set of
set of faults that explains an abnormality. Completeness of a if-then-else rules. They are easy to develop and apply. Thus,
diagnostic unit would require the actual fault to be a subset of this technique enables transparent reasoning[3]. Qualitative
the proposed fault set. Resolution of a diagnostic unit would model involves deriving qualitative behavior from physical
demand fault set to be as minimal as possible. So, a trade-off behavior of the system[5]. This methodis useful in noncritical
between completeness and resolution exist. These two processes [3]. Qualitative model approach is specific to a
concepts would originate whenever modeling and designing system and we require complete list of rules to identify faults
of a diagnostic unit is performed. accurately.
Earlier we have mentioned classification of diagnostic Process history based method develops a relationship
system, now we understand them in more details. Diagnostic between measured inputs and measured outputs. Quantitative
methodscan be classified in number of ways due to similarity model develops a mathematical relation from available
in approaches in [1], [3]. We classify diagnostic methods as historical data.When model shows physical relationship
model based and process history (data-driven) based between available historical data, then such models are called
methods. Model based methods are developed based on some qualitative models[4]. In other words, when the model
fundamental understanding of the physics of the system. This parameters or features have no physical significance, they are
priori knowledge based approach can be realized either by referred to as black-box models[3].Modelparameters in a
quantitative or qualitative models. Quantitative models can be model, when designed based on first principle shows physical
developed from detailed physics models or simplified physics significance of the system. These models are known as gray
models, see fig. 1.1. While qualitative methods use either rule box models[3]. Process history based method find its
based models or qualitative physics based models. significance where training data are easy to create
orcollectand are implemented where no other methods
exist[3].
We have seen so far various standard methods for the
realization of a diagnostics unit. Now, we see general details
about NPP operation and requirements of FDD for NPP.

III. REQUIREMENT OF FDD FOR NPP

A. Description of NPP

Figure 1.1Classification of Diagnostic methods Based on the statistics from World nuclear association,
INSTITUTE OF TECHNOLOGY, NIRMA UNIVERSITY, AHMEDABAD – 382 481, 08-10 DECEMBER, 2011 3

nuclear power plants are delivering about 6% of world energy processes. Firstly linearization of nonlinear models is done.
and about 15% of world electricity. NPPcan be seen as a Secondly kalman filters are used on these linear models.
thermal power station in which the heat source is one or more General problem with model based methodology is the
nuclear reactors. As in a conventional thermal power station modeling uncertainties due to deviations in the
the heat is used to generate steam which drivessteam turbine parameters,linear approximations achieved for nonlinear
connected to a generator which produces electricity. models can be poor over the wide range andwe need to model
The heart of a NPP is the Nuclear reactor. In its central faults specifically for novelty identifiability.
part, the reactor core‟s heat is generated by controlled nuclear Diagraphs, fault tree, structural and functional methods are
fission. Coolant circuit removes heat energy from the reactor. used forqualitative model based approaches[10]-[15]. Signed
diagraphs represents the cause-effect relations. Arc is directed
Since nuclear fission creates radioactivity, the reactor core is
from cause node to effect node[1]. In case of fault tree, layers
surrounded by a protective shield. This containment absorbs
of nodes exist. Fault tree is constructed by using logic gate
radiation and prevents radioactive material from being
symbols[16]. This method is top down, deductive method
released into the environment. In addition, many reactors are
aimed at analyzing the effects of initiating faults and events
equipped with a dome of concrete to protect the reactor
on a complex process system. Structural approach specifies
against external impacts.
the connectivity information of a system and subsystem while
Apart from NPP and its facilities, health care facilities,
functional approach specifies the output of a unit as a
research institutes, natural sources like civil aircrew, mines,
function of its inputs. Disadvantage with use of qualitative
etc. also suffers from radiation exposure. Data from
models is the generation of spurious solutions. Considerable
International atomic energy agency shows that, for nuclear
work has been done so for reducing spurious solutions like
reactor annual average effective dose for an individual is
generation of latent constraints in signed directed graphs.
1.4mSv; while in civil aircrew effective dose is 3mSv and in
Process history method develops model in terms of
metal mines effective dose is 2.7mSv. Still in NPP,
quantitative or qualitative nature from available data[4], [8],
safety focuses on unintended conditions or events which may
[16]-[19]. Qualitative models are realized using expert
lead to radiological release from authorized activities. In
systems and Qualitative trend analysis (QTA) while
other words, operational safety is a prime concern for those
quantitative models are realized using principle component
working in nuclear plants. These are supported by continuous
analysis (PCA)/partial least square, neural networks,
monitoring of individual doses and of the work environment
etc.[10],[20]. Indevelopment of expert system, we have ease
to ensure very low radiation exposure compared with other
in development process, transparent reasoning and the ability
industries.So early detection and quick diagnosis of a fault is
to provide explanations for the solutions provided. With
of paramount importance for NPP.
QTA, we can achieve process monitoring, malfunction
The complexity and voluminous nature of NPP make it a
diagnosis and prediction of future states. This method
very difficult task to ascertain whether the plant is operating
provides valuable information that explains the process
within acceptable limits. Furthermore, when plant is not
behavior. PCA is based on an orthogonal decomposition of
operating within acceptable limits, it is difficult to determine
the covariance matrix, i.e. data set are reduced to lower
what has gone wrong. With use of FDD determination of,
dimensions which can still describe the major trends in the
part of the system which is failing and which kind of fault it
original data set.NN refers to the network of biological
is facing would be accurate.In addition, operator‟s task of
neurons. Modern usage refers to artificial neural networks
handling plant becomes easy.
(ANN). ANN can be realized using supervised learning
B. FDD in NPP strategy or unsupervised learning strategy. They mainly
consist of three layers. First layer is input, then hidden layer
All model based diagnostic methods rely on an explicit and finally an output layer. ANN being process historybased
model of the plant, under observation. Model based method approach has minimal modeling requirements. Once network
consists of two steps: is trained, on-line computation complexity is minimal. ANN
1) Residual generation: Generating inconsistencies between does not possess an explanation facility for its actions and
the actual and expected behavior also its underlying structure is difficult to understand.
2) Selection of decision rule We have seen in this sectionvarious approaches used for
Quantitative model based approaches can be realized by diagnosis. These methods individually satisfy few of the
analytical redundancy, i.e. observer,extended kalman filters, above listed characteristics requirements of FDD unit while
parity relation, etc., [6]-[9].In analytical redundancy unlike still lacks in few other qualities. Different approaches
physical redundancy (measurements from multiple sensors employed for diagnosis purpose has relative strengths and
are compared to each other), sensor measurements are weaknesses. Also some of these methods can complement
compared to values computed analytically. Observer based one another[1]-[3].Hybrid techniques combine two or more
method develops a set observers, each of which is sensitive to standard techniques with the aim to provide accurate
a subset of faults and insensitive to the remaining diagnosis to the system over the wide range of operational
faults.Kalman filters, a recursive algorithm for state conditions. In next section, we look upon few of the
estimation, in state space model is equivalent to an optimal hybridtechniques proposed or used in NPP for diagnostic
predictor for input output model. Extended kalman filters are purpose.
widely used in non-linear systems by employing two step
4 INTERNATIONAL CONFERENCE ON CURRENT TRENDS IN TECHNOLOGY, „NUiCONE – 2011‟

C. Case study: Diagnosis techniques used in NPP 2) Monitoring system for NPP

1)Identification of transients in NPP An on-line monitoring system for the NPP has been
developed using the neural networks and the expert system
Identification of transients for timely monitoring of nuclear [21]. With conventional monitoring methods in NPP,
plants operation is of fundamental importance. A fuzzy-logic detecting anomalies can be achieved. But detecting symptoms
based method for the identification of transients is being of anomalies is difficult because of wide error boundary
proposed[17]. This approach uses if-then rules, which covering power level operation from zero to full[21].
constitutes the heart of the approach, developed from the Neural network can detect deviation from the normal state
available input-output signal data. Marzio uses Quandry of operation, while interpretation of the deviation by an
based reactor kinetics code (QUARK) for simulation, to expert system is performed to diagnose the cause. Neural
obtain necessary data for the identification. networks lacks in interpreting the cause of deviation while
QUARK is being used to model the operations of the use of expert systems alone for the plant monitoring is too
Westinghouse Advanced Pressurized water reactor, AP600. much computational complexity.
Code allows simulating the neutronic and thermal-hydraulic Compact simulator of Surry-I, U.S.A. helps in
transient behavior of a 3-D Light water reactor. simulating many kinds of malfunctions caused by equipment
In the QUARK, analytical nodal method is used to solve failure during steady state and transient operation for the
two-group neutron diffusion equation and the upgraded testing purpose. Time interval of the simulation is 2 seconds.
version of the COBRA code is used to develop the core Nabeshima has considered the Borssele NPP, a two loop
thermal-hydraulic model.The output of the code is taken as pressurized water reactor, with normal electric power output
the correct (pseudo) experimental values of the process of 477 MWe. On-line data acquisition system sends 72 plant
parameters, which are used in the fuzzy model training and signals to the neuro-expert system every two seconds. Of
identification tasks. these 21 most significant signals are selected for the inputs of
Possible four forcing functions considered by Marzio neural network. Others signals are considered to be
which can originate transients are: unchanged during controlled operation of power rise or fall,
- Primary pressure in the core (Pp) so that conventional monitoring method can detect anomaly
- Primary core inlet temperature (Tin) of these signals. Anomaly detection patterns by NNs are
- Core inlet flow rate (Γin) created using PWR simulator. In anomaly detection principle,
- Boron concentration (CB) measured sensor value is compared with predicted value for
Selection of forcing functions allows simulation of various checking fault level.
transients listed as: Software of NN and expert system are programmed in
- Leakage for the pressurizer (PP) FORTRAN and executed on the PC. The advisory displays
- Pressure increase (PP) show the status of NPP diagnosed by neuro-expert system.
- Increase of coolant inlet temperature (Tin) Java language based program is used in the graphical
- Decrease of coolant inlet temperature (Tin) advisory displays [21].
- Loss of coolant accident (Tin) NNs are trained by the current and past system input and
- Rise in boron concentration (CB) output for predicting next output of the system. This process
- Decrease in boron concentration (CB) of one-step-ahead prediction can be implemented by ANN,
The proposed approach by Marzio begins with and applied for dynamic tracking. Expert system use sensory
automatically generating the Fuzzy rule base (FRB) starting signals and outputs of the NNS along with information
from plant input/output data pairs. Set of rules designed maps provided from human operators as the input.
measured data to system behavior required for the system Few of malfunction cases considered by Nabeshima are
identification. listed below:
Each signal is characterized by a classificability threshold, - Small reactor coolant system leak
given as: - Leakage of atmospheric steam dump valve
- Partial loss of feed water
- Turbine governor valves fails
- Volume control tank level control fails
- Steam generator level control fails
Where, smax, smin and savg are the maximum, minimum and Last two malfunctions mentioned above are some kind of
average value of the signal in the transients. controller failures. With conventional alarm system detecting
FRB algorithm takes care of possible ambiguities in the those failures is possible. But with these systems identifying
classification. During simulation, the responsible forcing the cause of these anomalies is difficult. The neuro-expert
function is readily classified with small percentage of errors system in next time step can identify the failures for the failed
from signals measured within first 7s after the beginning of signals because the failed signals show deviations which are
the transients. Result by Marzio shows that by increasing the much larger than the others.
signal variation threshold classification accuracy can be The simple expert system called DISKET is compared with
increased. neuro-expert system by Nabeshima. DISKET diagnosis
process utilizes information obtained from the conventional
alarm system for knowing plant status. Neuro-fuzzy expert
INSTITUTE OF TECHNOLOGY, NIRMA UNIVERSITY, AHMEDABAD – 382 481, 08-10 DECEMBER, 2011 5

system shows better result in identifying anomaly than [3] Srinivas Katipamula and Michael R. Brambley, “Methods for Fault
detection, diagnosis and prognostics for building systems- A review,
DISKET [21].
Part I”, HVAC&R Research, vol. 11, No. 1., July 2004.
Nabeshima applied neuro-expert system on Borssele NPP [4] Venkat Venkatasubramanian, Raghunathan Rengaswamy, Surya N.
and the on-line PWR simulator. Neural network successfully Kavuri and Kewen Yin, “A review of process fault detection and
detected the symptoms of anomalies. The expert system diagnosis Part III: Process history based methods”, Computers &
Chemical Engineering 27, pp. 327-346, Apr. 2002.
correctly recognized the plant operation mode and diagnosed [5] Venkat Venkatasubramanian, Raghunathan Rengaswamy and Surya N.
the plant status. Kavuri, “A review of process fault detection and diagnosis Part II:
Qualitative models and search strategies”, Computers & Chemical
IV. CONCLUSION Engineering 27, pp. 313-326, Apr. 2002.
[6] A. Ashokan and D. Sivakumar, “Fault detection and diagnosis for a
three-tank system using structured residual approach”, ICGST-ACSE
In this paper, we have shown types of faults which may Journal, Vol.7, Issue 2, Nov. 2007.
occur in system. For early detection of such faults, we have [7] Edward Y. Cho and Alan S. Willsky, “Analytical redundancy and the
design of robust failure detection systems”, Automatic Control, Vol.
explained the need of fault detection and diagnostic unit. To AC-29, No. 7, pp. 603-614, July 1984.
rely on such automatic process system we need fulfillment of [8] C. Angeli, “On-line fault detection techniques for technical systems: A
some necessary qualities, i.e. characteristic requirements of survey”, International Journal of Computer Science & Applications
FDD unit are expressed. Then classifications of diagnostic Vol. I, No. 1, pp. 12-30, 2004.
[9] Richard Dorr, Frederic Kratz, Jose Ragot, Francois Loisy and Jean-Luc
approach into model and process history based methods are Germain, “Detection, Isolation and Identification of sensor faults in
explained in brief. Ihave observed that diagnostic methods Nuclear Power Plants”, Control Systems Technology Vol. 5, No. 1, Jan.
realized with help of quantitative models are difficult to 1997.
develop. Qualitative models have transparent reasoning and [10] Chun-ling Xie, Jen-Yuan (James) Chang, Xiao-cheng Shi and Jing-min
Dai, “Fault diagnosis of nuclear power plant based on genetic-RBF
are well suited in data-rich environment. While process neural network”, 15th International conference on Mechatronics and
history based method are applicable for virtually any kind of Machine vision in Practice, Dec. 2008.
pattern recognition problem. [11] Irina Obreja, “Diagnosis of power plant faults using qualitative models
I have presented here two practical application of FDD in and heuristic rules”, IEA/AIE ’90 Proceedings of the 3 rd International
conference on Industrial and engineering applications of artificial
NPP. One is conventional approach of fuzzy logic while intelligence and expert systems Vol. 1, pp. 41-46, 1990.
second approach isneuro-expert system, where we integrate [12] Hiranmayee Vedam, “Signed Diagraph based multiple fault diagnosis”,
their individual strengths and compensates their Computers & Chemical Engineering Vol. 21, Issue 1-2, pp. 8655-8660,
weaknesses.Our observation on conventional and hybrid 1997.
[13] H. Chung, Z. Bien, J. Park, and P. Seong, “Incipient multiple fault
approach enables, me to decide our future course of action diagnosis in real time with application to large-scale systems”,Nuclear
i.e. future work would be aimed to elaborate more on Expert Science, vol. 33, no. 8, pp. 1943–1954, Aug. 1994.
systemand ANN approaches along with redundancy method, [14] Joon On Yang and Soon Heung Chang, “A Diagnostic expert system
so that we can prepare a better framework of diagnostic unit for the nuclear power plant based on the Hybrid Knowledge
Approach”, Nuclear Science Vol. 36, No. 6, pp. 2450-2458, Dec. 1989.
which accommodate all of these methods to yield wide [15] Chun-ling Xie, Jen-Yuan (James) Chang, Xiao-cheng Shi and Jing-min
operation range of diagnostic unit and still, retaining accuracy Dai, “Fault diagnosis of nuclear power plant based on genetic-RBF
of our result. neural network”, International Journal of Computer Applications in
Technology Vol. 39, No. 1/2/3, pp. 159-165, 2010.
[16] Enrico Zio, Piero Baraldi and Irina Crenguta Popescu, “A fuzzy
V. HELPFUL HINTS decision tree method for fault classification in the steam generator of a
pressurized water reactor”, Annals of Nuclear Energy 36, pp. 1159-
A. Abbreviations and Acronyms 1169, May 2009.
AEM: Abnormal event management [17] Marzio Marseguerra, Enrico Zio, Andrea Oldrini and Enrico Brega,
“Fuzzy identification of transients in nuclear power plants”, Nuclear
ANN: Artificial neural network
Engineering and Design 225, pp. 285-294, Jun 2003.
FDD: Fault detection and diagnosis [18] Dieter Wach, “Experience with knowledge based systems for
FRB: Fuzzy rule base maintenance diagnosis”, Control systems, Robotics and Automation
NN: Neural network Vol. XVI, No. 1/2/3, pp. 159-165, 2010.
[19] C. Angeli, “Diagnostic expert systems: From expert‟s knowledge to
NPP: Nuclear power plant real-time systems”,Advanced knowledge based systems: Model,
PCA: Principle component analysis Application & Research Vol. 1, pp. 50-73, 2010.
PWR: Pressurized water reactor [20] Shengwei Wang and Jingtan Cui, “Sensor- fault detection, diagnosis
QTA: Qualitative trend analysis and estimation for centrifugal chiller systems using principal
component analysis method”, Applied Engineering 82, pp. 197-213,
QUARK: Quandry based reactor kinetics code
Feb. 2005.
[21] K. Nabeshima, T. Suzudo, S. Seker, E. Ayaz, B. Barutcu, E. Turkcan,
VI. REFERENCES T. Ohno and K. Kudo, “On-line Neuro-Expert monitoring system for
Borssele Nuclear Power Plant”, Progress in Nuclear Energy Vol. 43,
No. 1-4, pp. 397-404, 2003.
[1] Venkat Venkatasubramanian, Raghunathan Rengaswamy, Kewen Yin
and Surya N. Kavuri, "A review of process fault detection and
diagnosis Part I: Quantitative model-based methods", Computers &
Chemical Engineering 27, pp. 293-311, Apr. 2002.
[2] Jianping Ma and Jin Jiang, “Application of Fault detection and
diagnosis methods in Nuclear power plants: A review”, Progress in
nuclear energy 53, pp. 255-266, Dec. 2010.