Diagnosis of Class using Swarm Intelligence applied to problem of Identification of Transient Nuclear

2011 International Nuclear Atlantic Conference - INAC 2011
Belo Horizonte,MG, Brazil, October 24-28, 2011
ASSOCIAÇÃO BRASILEIRA DE ENERGIA NUCLEAR - ABEN
ISBN: 978-85-99141-04-5

DIAGNOSIS OF CLASS USING SWARM INTELLIGENCE APPLIED
TO PROBLEM OF IDENTIFICATION OF TRANSIENT NUCLEAR

Manoel Villas Bôas Junior1, Andressa dos Santos Nicolau2, Edilberto Strauss1,3 , Flávio
Luis de Mello3 and Roberto Schirru2
1
Mestrado Integrado em Computação Aplicada – Instituto Federal de Educação, Ciência e Tecnologia do Ceará
Universidade do Estado do Ceará
Avenida Paranjana, 1700
60740-903 Itaperi, CE
junior@lmp.ufrj.br
2
COPPE/UFRJ – Nuclear
Universidade Federal do Rio de Janeiro
Ilha do Fundão, s/n
21945-970 Rio de Janeiro, RJ
andressa@lmp.ufrj.br
3
POLI/UFRJ – Escola Politécnica - Departamento de Engenharia Eletrônica e Computação
Universidade Federal do Rio de Janeiro
Ilha do Fundão, s/n
21945-970 Rio de Janeiro, RJ
edilberto.strauss@gmail.com
flavioluis.mello@gmail.com

ABSTRACT

This article presents a computational model of the Diagnostic System of Transient. The model makes use of
segmentation techniques applied to support decision making, based on identification of classes and optimized
by Particle Swarm Optimization algorithm (PSO). The method proposed aims to classify an anomalous event in
the signatures of three classes of the design basis transients postulated for the Angra 2 nuclear plant, where the
PSO algorithm is used as a method of separation of classes, being responsible for finding the best centroid
prototype vector of each accident/transient, ie equivalent to Voronoi vector that maximizes the number of
correct classifications. To make the calculation of similarity between the set of the variables anomalous event in
a given time t, and the prototype vector of variables of accident/transients, the metrics of Manhattan, Euclidean
and Minkowski were used. The results obtained by the method proposed were compatible with others methods
reported in the literature, allowing a solution that approximates the ideal solution, ie the Voronoi vectors.

1. INTRODUCTION

An important and natural human activity is the process of grouping people, objects and
events in similar classes [1]. In many computing systems make use of this human concept,
such as Systems, Database Systems, Data Mining and Decision Support Systems. Still in this
line of Systems we can mention the Diagnostic Systems based on classes, these Systems are
important tools that help measure the degree of collaboration between data problems,

identifying which class a given sample must be associated. The degree of similarity between
data of a class usually is related to a measure of distance between them.

In the area of Nuclear Engineering can be found diversity of these problems, such as the case
that in order to diagnose a possible accident, the operator in his decision-making process,
needs to analyze and classify different information of the event in progress. This is a real
problem of the Nuclear area known as Nuclear Accident Identification Problem and is
considered complex due to the large number of instruments and the dynamics of each
measured quantity. The process of analysis, identification and decision-making requires a
great cognitive load by the operator and the Decision Support Systems, such as diagnostic
systems and classification of accident/transient of a nuclear plant, the operator can help to
reduce or minimize their cognitive load at the time of an adverse condition. These support
tools can be incorporated into the operating system of the plant, in order to assist the operator
in making a decision quickly and correctly, minimizing the risk of an action, or
misidentification of the event in progress.

In the last years, several Identification and Diagnosis Systems of accident/transients have
been proposed based on artificial neural networks [2], genetic algorithms [3], swarm
intelligence [4] and quantum-inspired algorithms [5]. In this context, this article proposes a
prototype model of Diagnosis System of accident/transient able to classify an anomalous
event signatures within three design-basis accident/transient, where the algorithm Particle
Swarm Optimization (PSO) is used as a optimization tool to find the best position of the
prototype vectors (vectors of Voronoi [6]). Unlike the models proposed by [4] and [5], where
the Euclidean metric is used to measure the similarity distance between the prototype vectors
of accident/transients and the anomalous event, this System uses the Manhattan, Euclidean
and Minkowski metrics to measure such a distance. It also proposes a new classification
hits/error methods, which only considers a hit if the similarity distance selected is less than all
the timelines of the other two accidents. The PSO is different of the method proposed by [4],
does not effect the partitioning of the search space which increases the complexity of the
problem.

The remainder of the paper is organized as follows: Section 2 presents the proposed
Diagnostic System, Section 3 presents the metrics used to calculate the similarity distances.
Section 4 presents the main features of the algorithm of Particle Swarm Optimization (PSO).
Section 5 presents the implementation of the proposed method, along with the settings used.
Section 6 presents the tests performed and the results provided by the proposed method and
Section 7 presents the conclusions.

2. DIAGNOSTIC SYSTEM OF TRANSIENT

The Diagnostic System of Transient proposed is based on distance and uses the Manhattan,
Euclidean and Minkowski metrics to make the similarity distance between the vector of
variables of the anomalous event in a given time t, and the prototype vector of design basis
accident/transient selected. The PSO, as an optimization tool, is used to find the best position
of the prototypes vectors of each of the three accident/transient. As there is no requirement
that the classification prototype vectors be the centroids resulting from any function of
partitioning the space search, the search algorithm was used in the prototype vectors that

INAC 2011, Belo Horizonte, MG, Brazil.

maximize the classification hits number of all accidents. With this approach, it was
established a model solution that matches your search of the Voronoi vectors for the
identification of classes of accident/transient.

In the proposed Diagnostic System of accident/transient was represented by the temporal
evolution of 18 state variables, necessary and sufficient for the recognition of each transient
in the range 0 to 60 seconds, in a second one, where the first corresponds to the second drop
bars (TRIP) reactor. The anomalous event classification was done by measuring the shortest
distance between the set of variables (signature) of the anomalous event in a given time t, and
the prototype vector of each accident/transient, where the hit is only considered when the
similarity distance selected is smaller than all the timelines of the other two accidents.

In the previous works reported in literature some models of the Diagnostic System of
accident/transient was developed to partition the search space ([3] and [4]) and as calculated
using the Euclidean distance metric ([3] [4] and [5] ). In this paper, the PSO does not affect
the partition of the search space, which increases the complexity of the problem. It is used for
the first time for the proposed problem, such as calculating similarity distance the metrics of
the Manhattan and Minkowski.

3. METRICS USED FOR THE CALCULATION OF SIMILARITY DISTANCE

With the purpose of evaluating the performance of the Diagnostic System proposed, with
different metric similarity distance, were chosen the metrics of the Manhattan, Euclidean and
Minkowski.

The Minkowski metric is used to calculate the similarity or dissimilarity between two points,
being a generalization of Euclidean and Manhattan distances. The Minkowski distance is
calculated by equation (1), where D is the number of dimensions.


D
d ( xi , x j )  q k 1
( | xik  x jk | )q , q  1 (1)

The variation of the parameter q defines the three most common variations of the Minkowski
distance, being calculated by equations 2, 3 and 4 [6].

 Manhattan Distance,


D
d ( xi , x j )  k 1
( | xik  x jk | ) (2)

 Euclidiana Distance,


D
d ( xi , x j )  2 k 1
( | xik  x jk | )2 (3)


 “sup” Distance,

d ( xi , x j )  max1 k  D | xik  x jk | (4)

4. PARTICLE SWARM OPTIMIZATION (PSO)

In 1995 it was introduced the concept of Particle Swarm Optimization (PSO) of continuous
nonlinear functions [7], and was based on modeling the simulation of simplified social
groups. It is an optimization technique based on population, which makes its implementation
through the metaphor of social behavior and interaction of groups of birds or schools of
fishes.

In PSO members of the population are regarded as particles. The individual particle can not
assess whether their current position is good or bad. It is necessary to send these coordinates
to find a function that evaluates quantitatively, providing a number as a result. This is called
the objective function or fitness, which is a measure of fitness, to show how good a position
relative to other historically found.

In search of better results in the PSO, a new position found should be compared both with the
individual best position obtained so far by the particle itself (pBest), and with the best
position of the entire cluster (gBest). As described in [8], the PSO is a swarm of particles
simulated candidates for solving a given problem, where the search space in a n-dimensional
the particles are attracted to regions of high adaptive value, as Figure 1.

Figure 1. Update of a particle.

The position of the particle represents a candidate solution, while the topology of the search
space is given by the objective function of the problem. Each particle is assigned a speed,
where the information of direction and rate of change of position versus time, and attribute


performance or fitness, obtained by the evaluation of the objective function in the position of
the particle.

Changing the particle position and its speed is guided by its historical information of the
regions in which good and bad has passed, and by observing their neighbors successful.
 
X i ( t )  { xi ,1 ( t ),..., xi ,n ( t )} and Vi ( t )  { vi ,1 ( t ),...,vi ,n ( t )} , are respectively, the position (the
actual candidate solution vector) and speed (rate of change) of the particle i at time t in a

space of n-dimensional search. Considering also pBest i ( t )  { pBesti ,1 ( t ),..., pBesti ,n ( t )} , the

best position ever found by the particle i to time t and gBest i ( t )  { gBesti ( t ),...,gBesti ( t )} the
best position ever found by the swarm up to time t. The update rules for the PSO velocity and
position in the canonical PSO, is given by:

vi , j ( t  1 )  vi , j ( t )  c1 .r1 ( pBesti , j ( t )  xi , j ( t ))  c2 .r2 ( gBest j ( t )  xi , j ( t )) (5)

xi , j ( t  1 )  xi , j ( t )  vi , j ( t  1 ) (6)

Where i  { 1,..., p } (number of particles), j  { 1,..., n } (size of the search space), r1 and r2
are uniformly distributed random number between 0 and 1. The coefficients c1 and c2 are the

constant acceleration (called cognitive and social acceleration, respectively) for pBest

and gBest respectively, and n while indicating the components of the concerned individual, in
equations (5) and (6) above.

When the PSO finds the optimum region, within this region, it may face difficulties in
adjusting the speed of its growth to continue in a more refined search, resulting in a loss of
efficiency of his method. To resolve this issue introduce a weight for the previous speed of
the particle, called inertial weight (w) whose rule is considered critical for the convergence of
PSO. High values of w promote global exploration and exploitation, while low values lead
to a local search. With this, the equation of particle velocity, described by equation (5), which
can be rewritten as:

vi , j ( t  1 )  w.vi , j ( t )  c1 .r1 ( pBesti , j ( t )  xi , j ( t ))  c2 .r2 ( gBest j ( t )  xi , j ( t )) (7)

By observing the equations (5) and (6), there is no mechanism that limits the speed of a given
particle. This fault can result in a low efficiency for PSO, when compared to other
evolutionary computation techniques. Very high speed values can cause the particle out of the
search space, is the maximum or minimum limit can be solved in many ways, in this work
such speed limit has been resolved as follows:

1. If the particle's position extrapolate the maximum search space, the particle is
returned at random for the search space and speed is divided by 10;


2. If the particle's position extrapolate the minimum limit of the search space, the
particle is returned at random for the search space and speed is equal to 0.

In PSO algorithm, the swarm is initialized randomly (positions and velocities) and while the
stopping criterion (for a maximum number of iterations) is not reached, is executed in a loop,
Figure 2 shows the pseudo-code of the PSO.

Figure 2. Pseudo-code of PSO.

5. IMPLEMENTATION OF THE METHOD PROPOSED

5.1. Selection of Accidents Project-Based Transient

For the purpose of this study, the number of accident/transient selected to test the prototype
Diagnostic System of was based on the design-basis list of accident/transient of the Nuclear
Power Plant Angra 2, contained in Chapter 15 of the Final Report of Accident Analysis
(FSAR), required by the National Commission of Nuclear Energy. The accident/transient
were selected:

 BLACKOUT: consists in the loss of external power.
 LOCA: Consists of the loss of reactor cooling.
 SGTR: Consists of ruptures in steam generator tubes.

The time of 59 seconds was considered sufficient for the accident/transient could be detached
from each other, because the evolution of one or more state variables of the system,
considered as those that contribute most to the characterization of accident/transient in


question. Samples of these variables were made at intervals of one second. Thus, the
maximum score is 59 * 3 (accidents) = 177 (maximum score value to be obtained).

The process variables selected for the event in recognition of the course were the same used
in the work [3] [4] and [5] and are shown in Table 1.

Table 1. State Variables Accident/Transient

Variable Measure
1. Time s
2. Reactor water flow %
3. Hot leg temperature ºC
4. Cold leg temperature ºC
5. Primary water flow kg/s
6. Steam generator water level – large range %
Steam generator water level – narrow
7. %
range
8. Steam generator pressure MPa
9. Feed water flow kg/s
10 Steam flow kg/s
11. Flow in the rupture kg/s
12. Primary system flow kg/s
13. Primary system pressure MPa
14. Thermal power %
15. Nuclear power %
16. Subcooling power ºC
17. Pressurizer water level %
18. Primary mean temperature ºC

5.2. Setting the PSO

The PSO model, adopted to be implemented in the Nuclear Accident Identification Problem
of accident/transient is basically the same model of PSO described in Chapter 4. Each
individual in PSO consists of an array of 18 * 3 (18 variables of Table 1, and 3
accident/transient), each individual's position is represented by a real value in the interval
[0,1], representing the range of normalized variables .


In the implementation of PSO, the objective function (fitness) was implemented using the
following method:

 If the calculated value of the accident/transient each time is less than the lowest value
of the other two accident/transient, at all times, so it is said to be a hit, otherwise it is
considered an error.

5.3. Data Normalization

The state variables of accident/transient selected for this study, as described in Table 1, have
values in different scales (units). The need to harmonize these scales was supplied by
normalizing the values of attributes.

The normalization method chosen for this work was the MAX-MIN equalizer [9], which uses
the maximum and minimum value to normalize the data linearly between [0,1], using the
following equation:

x  min( x )
N( x )  (8)
max(x )  min( x )

Where N ( x ) is the normalized value of the variable x , min( x ) is the variable x smaller
variable and max( x ) is the variable x larger value.

6. RESULTS

The purpose of the tests performed in this work is to evaluate the performance of the
Diagnostic System proposed with different metrics of the similarity distance. For this were
chosen the Manhattan, Euclidean and Minkowski metrics.

For the test the method proposed in this work, 10 experiments were realized for each metric,
with different seeds, in 100 generations for a population of 500 individuals and 500
generations for a population of 100 individuals. In PSO has been used in the parameter w
varies between 0.5 and 0.9 with an increase of 0,005, the constants c1 and c 2 were fixed and
the values 2.3 and 1.7 respectively. Were also used for the Minkowski metric values of the
parameter q = 3, 6 and 9. On Table 2 presents the results of tests performed and the number of
correct classifications found by PSO.


Table 2. Results of the metrics for 177 hits.

Manhattan Euclidean Minkowski
Correct classifications
q=1 q=2 q=3 q=6 q=9
Best 177 172 177 177 177
Worse 175 165 175 177 177

The results in Table 2 show that the values 3, 6 and 9 the parameter q produced very similar
solutions, there is no indication that a discrepancy of these values should be selected as
preferred for the problem at hand.

In order to show the robustness of the method proposed in this paper, Table 3 presents the
results of 10 experiments, with the average. The configuration of the PSO was the parameter
used w = 0.5, the constants c1 and c2 with the values 2.3 and 1.7 respectively, and was used
for the Minkowski metric parameter q = 3, 6 and 9. The population used was 500 individuals
and the number of generations was 100.

Table 3. Results of the System proposed with three metrics with population = 500.

Hits Errors
Seed Manhattan Euclidean Minkowski Manhattan Euclidean Minkowski
q=1 q=2 q=3 q=6 q=9 q=1 q=2 q=3 q=6 q=9
12345 175 168 177 177 177 2 9 0 0 0
4948 177 167 177 177 177 0 10 0 0 0
7777 175 168 177 177 177 2 9 0 0 0
1111 177 167 176 177 177 0 10 1 0 0
4444 177 166 174 177 177 0 11 3 0 0
1254 177 168 176 177 177 0 9 1 0 0
4945 177 167 175 177 177 0 10 2 0 0
7845 177 171 177 177 177 0 6 0 0 0
5455 177 168 177 177 177 0 9 0 0 0
7878 177 165 177 177 177 0 12 0 0 0
Average 176,6 168 176,3 177 177 0,4 9,5 0,7 0 0

It appears that in Table 3, Manhattan’s and Minkowski’s metrics with parameter q=3,
obtained almost the same result, while using the Minkowski metric q=6 and 9 with the PSO
parameter reaches the global optimum (177) in all experiments.

In the last test performed, Table 4, the PSO has been maintained with the same configuration
of Table 3 tests, only the population ranging from 500 to 100 individuals and the number of
generations from 100 to 1000 generations.


Table 4. Results of the System proposed with three metrics with population = 100.

Hits Errors
Seed Manhattan Euclidean Minkowski Manhattan Euclidean Minkowski
q=1 q=2 q=3 q=6 q=9 q=1 q=2 q=3 q=6 q=9
12345 177 171 177 177 177 0 6 0 0 0
4948 177 171 177 177 177 0 6 0 0 0
7777 177 168 177 177 177 0 9 0 0 0
1111 177 171 177 177 177 0 6 0 0 0
4444 177 170 177 177 177 0 7 0 0 0
1254 177 173 177 177 177 0 4 0 0 0
4945 177 172 177 177 177 0 5 0 0 0
7845 177 171 177 177 177 0 6 0 0 0
5455 177 170 176 177 177 0 7 1 0 0
7878 177 168 177 177 177 0 9 0 0 0
Average 177 170,5 176,9 177 177 0 6,5 0,1 0 0

The results presented in Table 4, show consistency with Table 3 demonstrating the robustness
of the method, including the number of individuals in the population.

Figures 3, 4 and 5 show the graph convergence of the algorithm for the best solutions found
with the Manhattan and Minkowski metrics. This result was obtained with the configuration
of the PSO with the parameter w=0.5, the constants c1 and c2 with the values 2.3 and 1.7,
respectively, and was used for the Minkowski metric parameter q=6 and 9. The population
used was 100 individuals and the number of generations was 1000.


Figure 3. Graphic Convergence of Manhattan metric.

Figure 4. Convergence Graph of the Minkowski metric q = 6.

Figure 5. Convergence Graph of the Minkowski metric q = 9.

It is observed that the model presented several steady states, or remains from a few
generations without significant development of learning. The largest of these steady states is
observed in the Minkowski metric, with the parameter q = 6, between generation 5 and 16,
which the algorithm converges to the global optimum (177). In addition, the results show that


the Manhattan and Minkowski metrics are significant and almost equal in the search for a
solution, and the Minkowski metric converges faster in obtaining the global optimum.

7. CONCLUSION

In this paper we explore the Swarm Intelligence, through the ability of Particle Swarm
Optimization (PSO) algorithm, demonstrating that it is feasible as a method of classifying
data based on classes. The same was responsible for finding the best prototype vector, ie,
equivalent to the Voronoi vector that maximizes the number of correct answers for each of
the three accident/transient design basis, demonstrating robust and efficient solution to
Problem Identification accident/transient of a PWR nuclear plant.

Initially we must emphasize the simplicity of the algorithm with inspiration in the metaphor
of birds social behavior, whose implementation is easily translated from the equations that
define the canonical model of the algorithm. Moreover, we note that in most studies in the
literature the methods used for diagnosis of accident/transient was implemented to partition
the search space [3] and [4] and as calculated using the Euclidean metric of similarity [3], [4]
and [5]. In this paper, the PSO does not affect the partition of the search space, which
increases the complexity of the problem, and uses the first time for the proposed problem,
measurement and calculation of similarity between classes, Manhattan and Minkowski
metrics. After the tests, it can be said that the proposed system is robust and effective in the
identification problem of accident/transient of a nuclear plant and that the Manhattan and
Minkowski metrics for the problem at hand, are effective and superior to metric Euclidean.

The Manhattan’s and Minkowski’s metrics get better results than the results obtained by the
Euclidean metric. The results obtained by the Manhattan’s and Minkowski’s metrics are
significant and almost equal in the search for a solution, but the Minkowski metric stands out
to converge faster in obtaining the global optimum, thereby being regarded as more effective.

Thus we can conclude that the algorithm Particle Swarm Optimization has been successfully
applied to the solution of the Nuclear Accident Identification Problem and was effective in
finding solutions in search spaces of continuous high-dimensional without prior knowledge
about the complexity of the search spaces involved, allowing a solution that can be seen as
the ideal solution, ie, the prototype vectors obtained by PSO behave as vectors of Voronoi
classes of accident/transient selected always getting the maximum score possible (177).

ACKNOWLEDGMENTS

The authors R. S. and A. S. N. acknowledge CNPq (Conselho Nacional de Desenvolvimento
Científico e Tecnológico) and FAPERJ (Fundação de Amparo à Pesquisa do Estado do Rio de
Janeiro).

REFERENCES

1. VIANA, J. R. F., "Multivariate Analysis of groupings of data using fuzzy techniques and
rigid." Dissertation M. Sc., UECE, Ceará, CE, Brazil (2008).


2. BARTLETT, E.B., UHRIG, R.E., “Nuclear Power Plant Status Diagnostics Using an
Artificial Neural Network, Nuclear Tecnology”, vol. 97, Mar, pp. 272 – 281, 1992.

3. ALMEIDA, J.C.S., "Method of Approach to Transient Identification of possibilistic,
Optimized by Genetic Algorithm", M.Sc. Thesis, COPPE/UFRJ, Rio de Janeiro, RJ,
Brazil, 2001.

4. MEDEIROS, J. A. C. C., "Particle Swarm Optimization as a Tool for Complex Problems
of Nuclear Engineering," D.Sc. Thesis, COPPE/UFRJ, Rio de Janeiro, RJ, Brazil, 2005.

5. NICOLAU, A. S., "Quantum Computing and Swarm Intelligence applied in Accident
Identification of a PWR Nuclear Plant," M.Sc. Thesis, COPPE/UFRJ, Rio de Janeiro, RJ,
Brazil, 2010.

6. HAYKIN, S., “Neural networks – A comprehensive foundation”, 1 ed., New Jersey,
Prentice-Hall, 1994.

7. KENNEDY J., EBERHART R.C., “Particle Swarm Optimization”, Proc IEEE International
Conference on Neural Net Works, pp. IV: 1942-1948, Perth, Australia, 1995.

8. WAINTRAUB, M., M. D. Schirru, R., "Parallel Algorithms for Particle Swarm
Optimization on Nuclear Issues", In Press, Corrected Proof, 2009.

9. "MLP model," http://equipe.nce.ufrj.br/thome/p_grad/nn_ic/transp/T5_mlp.pdf (2008).


Diagnosis of Class using Swarm Intelligence applied to problem of Identification of Transient Nuclear

More Related Content

Similar to Diagnosis of Class using Swarm Intelligence applied to problem of Identification of Transient Nuclear (20)

Recently uploaded (20)

Diagnosis of Class using Swarm Intelligence applied to problem of Identification of Transient Nuclear