Fuzzy Sets and Systems 129 (2002) 371–386

www.elsevier.com/locate/fss

Combined fuzzy logic and genetic algorithm techniques—application

to an electromagnetic )eld problem
I.G. Damousis, K.J. Satsios, D.P. Labridis ∗ , P.S. Dokopoulos
Power Systems Laboratory, Department of Electrical and Computer Engineering, Aristotle University of Thessaloniki,
Thessaloniki, GR-54006, Greece

Received 20 September 1999; received in revised form 29 May 2001; accepted 31 May 2001

Abstract
The in7uence of a faulted electrical power transmission line on a buried pipeline is investigated. The induced electromagnetic
)eld depends on several parameters, such as the position of the phase conductors, the currents 7owing through conducting
materials, and the earth resistivity. A fuzzy logic system was used to simulate the problem. It was trained using data derived
from )nite element method calculations for di9erent con)guration cases (training set) of the above electromagnetic )eld
problem. After the training, the system was tested for several con)guration cases, di9ering signi)cantly from the training
cases, with satisfactory results. It is shown that the proposed method is very time e:cient and accurate in calculating
electromagnetic )elds compared to the time straining )nite element method. In order to create the rule base for the fuzzy
logic system a special incremental learning scheme is used during the training. The system is trained using genetic algorithms.
Binary and real genetic encoding were implemented and compared.
c 2002 Elsevier Science B.V. All rights reserved.

Keywords: Fuzzy logic; Genetic algorithms; Fuzzy modeling; Inductive interference

1. Introduction that have appeared during the last few years. Fuzzy
set theory as well as various applications are presented
The word fuzzy in its technical meaning appeared thoroughly in [22].
for the )rst time in the scienti)c community by On the other hand, the underlying principles of
Prof. Lot) Zadeh [21]. Zadeh [21] laid the foundation genetic algorithms (GAs hereafter) were )rst pub-
for many applications of the fuzzy logic systems (FLS lished by Holland [7]. The mathematical frame-
hereafter) in diverse areas like control systems, pat- work was developed in the late 1960s and is
tern recognition, forecasting, reliability engineering, presented by Holland’s pioneering book “Adap-
signal processing, monitoring, and medical diagnosis tation in Natural and Arti)cial Systems”, pub-
lished in 1975 [8]. GAs have been also used in
∗ Corresponding author. Tel.: +30-31-996-374; fax: +30-31-
many diverse areas that require parameter train-
996-302. ing such as function optimization, image process-
E-mail address: [email protected] (D.P. Labridis). ing, system identi)cations, etc. A good reference

0165-0114/02/$ - see front matter  c 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 5 - 0 1 1 4 ( 0 1 ) 0 0 1 3 7 - 3
372 I.G. Damousis et al. / Fuzzy Sets and Systems 129 (2002) 371–386

Fig. 1. Cross-section of the examined electromagnetic )eld problem.

on GAs and their implementation is the book of This paper suggests the following steps for reducing
Goldberg [4]. the computational e9ort: (a) the problem is solved for
The present paper presents a combination of fuzzy several sets of parameters using FEM and a database
logic and genetic algorithm techniques for the cre- (training set) is built, (b) a fuzzy logic system is built
ation of a system that calculates the electromagnetic and trained using the training set, and (c) for a new
)eld induced by a faulted transmission line to the sur- set of parameters (evaluation set) the solution is found
rounding area and the induced voltage on a nearby- in negligible small computing time using the trained
buried pipeline. The inductive interference problem fuzzy rules.
between a faulted overhead transmission line and a The fuzzy logic system is trained using genetic
nearby-buried pipeline is of growing practical inter- algorithms; the result is called genetic fuzzy sys-
est, due to restrictions currently imposed on public tem. Genetic fuzzy systems (GFS) are already in
utilities in the use of right-of-ways. These restrictions use in the last years [1 – 3,5,6,9 – 12,14,19] and have
have resulted in situations in which overhead trans- led to standard coding schemes and genetic opera-
mission lines, pipelines, railroads, telecommunication tors. Unlike FEM the GFS does not su9er in case
lines, etc., have to be laid in straight narrow corri- the solution space is non-convex and once it has
dors for several kilometers. This policy minimizes the been trained it can calculate the electromagnetic
amount of land used but a faulted overhead transmis- )eld in fractions of a second, which is very helpful
sion line in such a corridor causes signi)cant interfer- especially if the environmental parameters change
ence to nearby parallel conductors. The mentioned rapidly.
interference is governed by Maxwell’s electromag-
netic )eld equations and depends upon several param-
eters such as the geometry, the boundaries and the 2. Description of the problem
electromagnetic properties of the materials. Recently,
a )nite element method (FEM) approach has been pro- An overhead transmission line with a single-phase
posed [15 – 17], in order to solve this problem in two fault runs in parallel with a buried pipeline (Fig. 1).
dimensions. FEM is an accurate numerical method, More details about this con)guration are given in
but its main disadvantage is that the computing time [15 – 17]. The magnetic vector potential (MVP) is
may increase tremendously with the number of the )- sought. Having the MVP, it is easy to calculate in-
nite elements [18], resulting to a huge computational duced voltages across pipeline and earth, which is an
e9ort. important engineering task. The solution is governed
 I.G. Damousis et al. / Fuzzy Sets and Systems 129 (2002) 371–386 373

by the di9usion equation and two consequence inputs, which are the dis-

 tance l of point (x; y) from the faulted line and
1 @2 Az @ 2 Az the distance l
of point (x; y) from the buried
+ − j!Az + Jsz = 0;

@x2 @y2 (1) pipeline.
− j!Az + Jsz = Jz : The jth fuzzy rule (Rj ) may be described as follows:
where
is the permeability, Az the phasor of vector
potential, x; y are the point coordinates,  the conduc- Rj : IF d and  belong to the jth membership
tivity, ! the current frequency, and Jsz the current den- functions
dj and
j correspondingly
sity. The solution depends on the boundary conditions,
the geometry, and the material properties (Fig. 1). In (premise part of the jth rule)
this paper we only consider as variables the point co- 1 1
THEN Aj = 0j + lj 2 + lj

2
ordinates (x; y) where we calculate MVP, the sepa- l +c l +c
ration distance d, and the soil resistivity . Although (consequence part of the jth rule) (2)
there are many other parameters (variables) to con-
sider, we restricted the variables in order to show in a where
simple way the e9ectiveness of the proposed method. 

l= x2 + y2 ; l
= (x − d)2 + (y − dp )2 ;
3. The method
c = 10−10 ; (3)
The MVP is found by solving the di9usion equa-
tion for several sets of parameters d;  using the )nite j = (1; : : : ; m); m is the number of rules, c is a constant
elements procedure of [15 – 17]. The MVP at coordi- to prevent over7ow in case the point is located on
nates (x; y) for other sets of parameters d;  is found the pipeline or coordinates (0; 0); d;  are the premise
by extrapolation of the known results in the space of input variables of the FLS, dp is the depth at which
the parameters. This extrapolation is made using fuzzy the pipeline is buried, Aj is the MVP proposed by the
logic techniques. Fuzzy rules are trained using genetic jth rule and
dj ;
j are the membership functions that
algorithms. The aim of the training is to minimize the de)ne the jth fuzzy rule. The parameters 0j ; lj ; lj are
average rms error between the real MVP values and the consequence part coe:cients of the jth rule and
the fuzzy system’s outputs. The procedure is as de- de)ne its output. The membership functions used in
scribed below. order to create the fuzzy inputs were chosen to be
A database is built which contains the FEM solu- Gaussian as it is described below:
tions for di9erent sets of d;  at various points. This   2 
set of calculations is called training data set (TDS) 1  − Rj

j () = exp − ; (4)
and contains the training patterns of the fuzzy logic 2 j
system (FLS) to be trained.
The training patterns of the system consist of four where “” stands for the premise input and takes the
inputs and one output. The inputs are values d; . In addition Rdj ; Rj are the mean values and
(a) the separation distance d between the overhead dj ; j are the standard deviations of the membership
transmission line and the buried pipeline, functions (Fig. 2). Trapezoid and triangular member-
(b) the coordinate x, ship functions have also been used, leading to a less
(c) the coordinate y of a point in the cross-section accurate system.
of the TLS, and In order to produce the actual output of the FLS, the
(d) the soil resistivity . The single output is the weighted average interface [21], which is a commonly
MVP A(x; y) at point (x; y). The rules of each used method, has been selected. Therefore the output
FLS consist of two premise inputs (variables): of the FLS de)ned above, i.e. the MVP in a point with
the separation distance and the soil resistivity, coordinates x; y for separation distance d and earth
374 I.G. Damousis et al. / Fuzzy Sets and Systems 129 (2002) 371–386

Fig. 2. Representation of the fuzzy search space and corresponding membership functions (MF). Further division of premise input d with
the introduction of a new MF.

resistivity , is given by If q is the number of training patterns in the train-

m j j ing data set (TDS), the FLS is trained by introducing
j=1 A

A(d; ) = m ; (5) it with the set of q patterns (dp ; xp ; yp ; p =ApFEM ; p =
j
j=1
1; : : : ; q). The average rms error Jav of the FLS is de-
where )ned by

j =
dj (d)
j () (6)

q
1
p
is the )ring strength for rule Rj by the input vector Jav = J ; (7)
q
(d; ), while Aj is the output of rule Rj as de)ned in (2). p=1
 I.G. Damousis et al. / Fuzzy Sets and Systems 129 (2002) 371–386 375

Fig. 3. String representing the kth fuzzy logic system of the GFS.

where the rms error J p of the FLS for pattern p is FLS that constitute the GFS). In this paper s has been
given by chosen equal to 50.
In the developed GFS two parameter-coding
J p = 12 |ApFLS (d; x; y; ) − ApFEM (d; x; y; )|2 ; (8) schemes were used, binary and real. In the binary
in which ApFLS (d; x; y; ) and ApFEM (d; x; y; ) are the coding, 8 bits resolution for every parameter was
calculated values of MVP for pattern p obtained from used. A vector of bits or real numbers (chromosome)
FLS and FEM, respectively. We de)ne training to is constructed, which embodies the FLS parameters
be the minimization of the average rms error, which (9). The vectors of the kth FLS for m = 32 rules are
means that the )nally determined FLS corresponds to shown in Fig. 3.
the highest accuracy in MVP calculation that could be
obtained with this method. 3.1.2. Fitness function of the genetic algorithm
The )tness function of the kth FLS-chromosome
3.1. GA for the training of the fuzzy parameters has been selected to be
1
3.1.1. Chromosome structure fk = k
(k = 1; : : : ; l); (10)
Jav +a
The parameters of each FLS of the proposed GFS
that have to be adjusted through the training pro- where Jav is given by (7) and  = 0:0001 is a con-
cedure are the parameters of the membership func- stant used to prevent over7ow in case Jav becomes
tions Rnmf () ; nmf () (for  = d;  and nmf () is the very small. The GA maximizes the )tness function fk ,
number of membership functions used to partition leading to the minimization of Javk .
each premise input ) and j (for  = 0; l; l
and
j = 1; : : : ; m). 3.1.3. GA operators, rules optimization
The GA that has been developed for the adjustment The evolution, which leads from the initial popula-
of the FLS parameters seeks the optimum FLS that tion of FLS to the best FLS, is described as follows:
presents the minimum rms error Jav . Every FLS of the
GFS is represented by a vector of its parameters C k , 3.1.3.1. Selection. After the evaluation of the initial
given by randomly generated population, the GA begins the cre-
ation of the new FLS generation. FLS-chromosomes
Ck = (nmf()k ; nmf()k ; jk ); from the parent population are selected in pairs
(9)
 = d; ;  = 0; l; l
; j = 1; : : : ; m; to replicate and form o9spring FLS-chromosomes.
The FLS-chromosome selection for reproduction
where k (k = 1; : : : ; s) is the index number of the FLS is performed using the method of Roulette wheel
and s is the population size (i.e. the number of the selection [4].
376 I.G. Damousis et al. / Fuzzy Sets and Systems 129 (2002) 371–386

Fig. 4. Multi-point crossover operator.

3.1.3.2. Crossover. When two chromosomes are se- 3.1.3.3. Mutation. In the binary coding every bit of
lected, their vectors are combined in order to produce the o9spring chromosomes undergoes a probability
two new FLS using genetic operators. The main oper- test and if it is passed, the mutation operator shown
ators used are crossover and mutation and are applied in Fig. 5 alters that bit. In the real coding the same
with varying probabilities. So, if a probability test is probability test is performed with higher mutation
passed crossover takes place. If the probability test probability and if passed Michalewicz’s non-uniform
fails, the produced children are identical replications mutation operator is applied [14]. This operator is
of their parents. described below:
In the binary GA a multi-point crossover operator If Cvt = (c1 ; : : : ; ck ; : : : ; cL ) is an FLS-chromosome
has been used as shown in Fig. 4. vector and ck is an FLS parameter that is chosen to be
In the real GA the Max–min-arithmetical crossover mutated, the new parameter ckmut will be after mutation:
operator was used [5]. If Cvt and Cwt are to be crossed 
four possible children are created: mut
ck + $(t; ckr − ck ) if r = 0;
ck = (12)
C1t+1 = Cwt + (1 − )Cvt ; ck − $(t; ck − ckl ) if r = 1;
C2t+1 = (1 − )Cwt + Cvt ; where L is the chromosome length, (ckl ; ckr ) the do-
(11)
C3t+1 with t+1
c3k = min{ck ; ck
}; main of parameter ck (k ∈ 1; : : : ; L); r a random bit,
C4t+1 with t+1
c4k = max{ck ; ck
}: and function $(t; y) returns a value in the range [0; y]
such that the probability of the returning value being
Parameter  is a constant equal to 0.5 for our exper- close to 0 increases with t:
iments. The two children who have the higher )tness b
are chosen to replace the parents in the new population. $(t; y) = y(1 − n(1−t=T ) );
 I.G. Damousis et al. / Fuzzy Sets and Systems 129 (2002) 371–386 377

Fig. 5. Mutation operator.

where n is a random 7oating-point number in the in- FLS of the new generation. The best FLS of the par-
terval [0; 1], t the current generation, T the maximum ent population is copied to the next generation while
number of generations, and b a parameter chosen by the best FLS found in all the previous generations
the user, which determines the degree of dependency is stored, so that the probability of their destruction
with the number of generations. In that way the op- through a genetic operator is eliminated. According
erator makes a uniform search at the beginning of to the schemata theory [5] the new generation usually
the training and in later stages narrows the search provides a better average )tness.
around the local area of the parameter resembling a
hill-climbing operator. For our experiments b was cho-
sen equal to 5 [3]. 3.2. Fuzzy rule base incremental creation
mechanism
3.1.3.4. Varying operator probabilities. It should be The aim of a fuzzy logic system (or model) is the
mentioned that chromosome selection method and acquisition of a knowledge (rule) base that represents
crossover operator lead to population convergence, the input–output function of the real system or prob-
while mutation operator helps to maintain population lem that we want to model. The objective of the learn-
diversity. If premature convergence or excessive di- ing process is to create and then )ne tune the fuzzy
versity occurs, the training becomes ine:cient. In this sets and rules consisting the rule base so as to meet
system crossover probability ranges from 40% to 90% user speci)ed performance criteria of the system, in
per chromosome while mutation probability ranges our problem minimization of the error in MVP calcu-
from 0.04% to 0.24% per bit and 1% to 10% per lation. In this context the training=learning of the rule
real parameter. Premature convergence is monitored base can be considered as a parameter optimization
by extracting statistical information from the popula- problem. The parameters to be optimized are the cen-
tion. When premature convergence is observed, the ters and deviations of the fuzzy membership functions
crossover probability is lowered by 10% while mu- and the consequence part coe:cients of each fuzzy
tation probability is increased by 0.004% per bit and rule. The encoding of these parameters lengthens a
0.2% per real parameter. When excessive diversity chromosome by 56 bits per rule (8 bits per parame-
occurs, the crossover probability is increased by 10% ter) in case of binary coding and by 7 real numbers
while mutation probability is lowered by 0.004% per in case of real coding. It is obvious that a complex
bit and 0.2% per real parameter. fuzzy system with a large number of rules results in a
huge chromosome and the )nding of the optimal rules
3.1.3.5. Elitism. The procedure for the two FLS- becomes a search for the “needle in the haystack”.
chromosomes described previously is repeated until In case of two premise inputs the number of param-
all the FLS of the parent generation are replaced by the eters remains tractable, but it grows rapidly with an
378 I.G. Damousis et al. / Fuzzy Sets and Systems 129 (2002) 371–386

Fig. 6. Partitioning of the premise input d into two sub-domains linguistically expressed as “pipeline is near to the transmission line” and
“pipeline is far from the transmission line”.

increasing number of membership functions per d containing only one membership function, form-
premise input. The number of rules for each FLS of ing eight rules (Fig. 2(a)). That initial membership
the proposed GFS is not constant but gradually in- function must cover the entire range of the premise
creases in order to partition the overall optimization input for the following reason: Let us assume that the
problem to smaller, more feasible steps. The creation initial membership function covers only partially the
of the rule base takes place in the following two steps: premise space and the system encounters a pattern
1. The system scans the training database locating that is beyond the range on the membership function.
the number of n discrete values in the earth resistivity Eq. (6) gives that the )ring strength of every rule for
 premise input domain. The domain of the premise this pattern will be zero. The weighted average (5)
input “earth resistivity” is then partitioned with the that produces the output of the system encounters a
introduction of n membership functions centered on division by zero resulting in an over7ow. Of course,
these values: this means that the calculated error (8) is in)nite and
there can be no continuance of the training since all
Rq = q ; q = 1; : : : ; n;
the FLS individuals will have the same bad perfor-
max − min (13) mance hence there can be no selection of the )ttest.
q = :
n The system optimizes the consequence part coe:-
cients  of the existing eight rules for 100 generations.
This is possible because the “soil resistivity” vari-
If the desired accuracy is achieved (for our problem
able takes only a few values, in our problem eight,
2%) the training is complete. If not then the domain
so the system does not become very complex. These
of the input d is partitioned into two sub-domains, us-
membership functions are centered on the patterns so
ing two, partially overlapping fuzzy sets with the re-
they do not need training leading this way to a smaller
sult that the knowledge base now contains 16 rules
chromosome.
(Fig. 2(b)). The parameters of the two new fuzzy sets
2. Gradual partitioning of the “separation dis-
that derive from the original fuzzy set are (Fig. 6):
tance” premise input. Separation distance between
the pipeline and the transmission line varies from 0 to
more than 2 km so we cannot follow the technique we dmax − dmin
a1new = ; (14a)
used for earth resistivity. Training begins with input 3
 I.G. Damousis et al. / Fuzzy Sets and Systems 129 (2002) 371–386 379

Table 1
Training data set used for the training of the two GFS. Input variables are the earth resistivity , the separation distance d and the
coordinates x and y of points in the earth around the pipeline neighborhood, including pipeline itself. Output for the )rst GFS is the
amplitude and for the second GFS the phase of the MVP as they were calculated with the FEM

 (U m) d (m) x (m) y (m) MVP(amplitude) (Wb=m) MVP(phase) (deg)

30 70 70.00 −15:00 3:61E − 04 −22:80

30 70 81.66 −27:03 3:29E − 04 −25:57
30 100 100.00 −30:00 2:99E − 04 −31:23
30 800 770.00 −30:00 4:23E − 05 −82:64
30 800 785.00 0.00 4:27E − 05 −78:83
30 800 818.25 −13:50 3:88E − 05 −82:61
30 1000 1030.00 −15:00 2:48E − 05 −90:27
30 2000 1970.00 −22:50 4:76E − 06 −108:10
30 2000 2000.69 −8:61 4:65E − 06 −108:54
70 400 384.81 −7:82 1:72E − 04 −44:46
70 400 392.25 −25:56 1:67E − 04 −46:05
70 400 424.77 −6:93 1:58E − 04 −46:72
70 1000 970.00 −15:00 5:95E − 05 −73:04
70 1000 1007.50 0.00 5:68E − 05 −72:98
70 1000 1015.00 −30:00 5:47E − 05 −76:05
100 70 40.00 −30:00 5:09E − 04 −20:45
100 70 40.00 −15:00 5:38E − 04 −19:34
100 70 40.00 0.00 5:59E − 04 −18:53
100 100 92.25 −25:56 4:15E − 04 −23:98
100 800 770.00 0.00 1:04E − 04 −59:87
100 1000 980.55 −16:99 7:58E − 05 −67:10
100 1000 1015.00 −30:00 7:16E − 05 −69:22
100 1000 1022.50 0.00 7:23E − 05 −67:27
300 300 312.38 −8:10 3:17E − 04 −29:23
300 300 324.05 −23:53 3:10E − 04 −30:00
300 2000 2007.50 0.00 5:86E − 05 −72:55
500 200 215.00 −30:00 4:18E − 04 −23:83
500 300 281.66 −27:03 3:75E − 04 −25:93
500 300 290.36 −15:80 3:71E − 04 −26:01
500 300 322.50 0.00 3:55E − 04 −26:74
500 1000 1030.00 −15:00 1:70E − 04 −44:60
700 150 120.00 −15:00 5:46E − 04 −19:26
700 400 384.81 −7:82 3:52E − 04 −26:89
700 700 670.00 −22:50 2:60E − 04 −33:74
700 700 690.36 −15:80 2:56E − 04 −34:07
700 700 712.38 −8:10 2:51E − 04 −34:41
900 150 150.55 −16:99 5:30E − 04 −19:70
900 200 194.77 −6:93 4:88E − 04 −20:90
900 800 830.00 −30:00 2:46E − 04 −35:01
900 1500 1499.09 −17:48 1:56E − 04 −46:35
900 1500 1524.77 −6:93 1:54E − 04 −46:56
1000 70 54.81 −7:82 7:03E − 04 −15:94
1000 150 131.66 −27:03 5:58E − 04 −18:98
1000 500 524.05 −23:53 3:29E − 04 −28:27
1000 2000 2030.00 −15:00 1:22E − 04 −52:73
380 I.G. Damousis et al. / Fuzzy Sets and Systems 129 (2002) 371–386

2(dmax − dmin )
a2new = ; (14b)
3
1;2
new = 0:6old : (14c)

The consequence part coe:cients  of the newly

created rules (Fig. 2(b)) remain the same as the ones
of the original rules. In that way, the behavior of the
fuzzy system is not greatly disturbed by the fuzzy set’s
division. The partitioned premise input d can now be
expressed by two linguistic variables: (1) the pipeline
is near the faulted transmission line or (2) the pipeline
is far from the faulted transmission line. This allows
a more detailed modeling of the problem in the struc-
ture of the FLS. While before the partitioning the FLS
could only be trained on the in7uence of the soil’s re-
sistivity on the MVP distribution, it can now produce Fig. 7. Binary GA vs. real GA. Best out of 20 runs.
di9erent outputs depending on the distance between
and the addition of fuzzy rules will not ruin a possible
the pipeline and the transmission line. The knowledge
better solution that appeared in a previous expansion
that was acquired through the )rst step is now ex-
step. Experiments though showed that the )nal FLSs
tended and not discarded since the new fuzzy rules in-
are always more accurate than the simpler ones which
herit the input=output function ( coe:cients) of the
is something that can be expected since the more rules
original rules during initialization. During the training
the FLS has the more adapted to the patterns it is,
these coe:cients will change so that they produce a
which in turn leads to a smaller average error.
more specialized output than the generic one of step 1.
The chromosome is expanded in order to include the


parameters of the newly created rules (dnew ; dnew ; 0j ; 4. Creation of the training data set



lj ; lj
), for new = 1; 2 and j
= 1; : : : ; 16. The new
fuzzy sets can shift and dilate during training. In or- The MVP is a phasor quantity and it is de)ned
der to prevent complete overlapping, the overlapping by its amplitude and its phase. Since the FLS of the
between fuzzy sets of input d is restricted to a max- developed GFS have a single output, two di9erent GFS
imum of half the fuzzy set’s standard deviation . are required to calculate complex MVP nodal values,
Again training of the new parameters takes place for a GFS for MVP amplitude and a GFS for MVP phase.
200 generations and the performance is checked. A Therefore, the TDS must have two outputs, one for
similar incremental building of the fuzzy rule base can the amplitude and the other for the phase training,
be found in [9]. respectively. Using the optimum FLS, derived after
The gradual partitioning of the existing sub- training of the GFS, it is possible to calculate the MVP
domains continues until the performance criterion values in the area of the complex electromagnetic )eld
is met or until a maximum number of fuzzy sets is problem of Fig. 1.
reached. This number is chosen to be four (linguis- A training data set (TDS) for the GFS has been
tically “very near”, “near”, “far”, “very far”) so that calculated for the TLS shown in Fig. 1 for IF = 1000A
the system does not become too complex. If premise and di9erent sets of  and d using the FEM procedure
input d is partitioned in four parts the resulting fuzzy described in detail in [15 – 17].
rules are 32. Various (x; y) points have been chosen in the earth
As it was mentioned in Section 3.1.3.5 a part of the around the pipeline neighborhood, as well as on the
elitism operator is the storage of the best chromosome pipeline itself. For each of those points, di9erent
(FLS) that was found in all the previous generations. separation distances d and earth resistivities  have
This ensures that the expansion of the chromosome been selected. As shown in the TDS of Table 1, the
 I.G. Damousis et al. / Fuzzy Sets and Systems 129 (2002) 371–386 381

Table 2
MVP amplitude distribution in the earth around the pipeline neighborhood, including pipeline itself, for several new con)guration cases of
the examined electromagnetic )eld problem, obtained by the FEM and the best FLS of the GFS, respectively. The FLS calculation errors
= |[(AFEM − AB-FLS )AFEM ] · 100| are also reported

 (U m) d (m) x (m) y (m) AFEM (Wb=m) AB-FLS (Wb=m) Error (%)

70 100 124.77 −6:93 3:47E − 04 3:47E − 04 0.04

70 150 120.00 0:00 3:50E − 04 3:56E − 04 1.55
70 150 162.38 −8:10 2:86E − 04 3:04E − 04 6.06
70 200 199.81 −1:75 2:55E − 04 2:73E − 04 6.77
70 200 200.00 −30:00 2:53E − 04 2:66E − 04 4.92
70 300 281.66 −27:03 2:15E − 04 2:16E − 04 0.35
70 300 299.81 −1:75 2:12E − 04 2:10E − 04 0.75
70 300 322.50 0:00 2:07E − 04 1:99E − 04 4.24
70 500 485.00 0:00 1:48E − 04 1:41E − 04 4.63
70 500 499.81 −1:75 1:46E − 04 1:36E − 04 7.66
70 700 699.09 −17:48 9:80E − 05 9:30E − 05 4.62
70 700 670.00 −15:00 9:90E − 05 9:80E − 05 0.38
70 800 799.81 −1:75 8:50E − 05 7:90E − 05 6.86
70 800 822.50 0:00 8:40E − 05 7:70E − 05 9.95
150 170 169.81 −1:75 3:46E − 04 3:61E − 04 4.02
150 170 169.09 −17:48 3:46E − 04 3:59E − 04 3.58
150 250 260.11 −21:49 2:85E − 04 2:89E − 04 1.49
150 250 274.05 −23:53 2:80E − 04 2:80E − 04 0.10
150 600 599.81 −1:75 1:58E − 04 1:64E − 04 3.58
150 600 600.00 −30:00 1:58E − 04 1:62E − 04 2.45
150 800 799.81 −1:75 1:19E − 04 1:24E − 04 4.25
150 800 790.36 −15:80 1:19E − 04 1:25E − 04 4.79
150 800 830.00 −15:00 1:18E − 04 1:19E − 04 0.97
150 900 900.00 −30:00 1:08E − 04 1:08E − 04 0.17
150 900 899.81 −1:75 1:08E − 04 1:09E − 04 1.06
150 900 892.50 0:00 1:08E − 04 1:10E − 04 1.81
150 1500 1500.69 −8:61 5:40E − 05 5:40E − 05 0.02
400 170 169.81 −1:75 4:43E − 04 4:43E − 04 0.01
400 250 231.66 −27:03 3:89E − 04 3:88E − 04 0.19
400 250 249.09 −17:48 3:83E − 04 3:77E − 04 1.47
400 600 599.81 −1:75 2:39E − 04 2:36E − 04 1.20
400 800 770.00 −30:00 1:96E − 04 1:96E − 04 0.00
400 800 824.77 −6:93 1:94E − 04 1:87E − 04 3.89
400 900 900.00 −30:00 1:82E − 04 1:73E − 04 5.32
400 1500 1524.05 −23:53 9:60E − 05 1:03E − 04 6.64
400 1800 1781.66 −27:03 8:40E − 05 8:40E − 05 0.01
400 1800 1792.50 0:00 8:40E − 05 8:40E − 05 0.01
600 170 169.81 −1:75 4:89E − 04 4:78E − 04 2.33
600 250 220.00 −30:00 4:20E − 04 4:30E − 04 2.27
600 250 249.81 −1:75 4:06E − 04 4:12E − 04 1.46
600 600 570.00 0:00 2:62E − 04 2:76E − 04 4.92
600 600 599.81 −1:75 2:60E − 04 2:67E − 04 2.78
600 800 784.81 −7:82 2:17E − 04 2:24E − 04 3.31
600 800 799.81 −1:75 2:16E − 04 2:22E − 04 2.74
600 900 884.81 −7:82 2:03E − 04 2:06E − 04 1.45
600 900 918.25 −13:50 2:02E − 04 2:00E − 04 1.00
600 1500 1507.50 0:00 1:30E − 04 1:29E − 04 0.72
600 1500 1499.81 −1:75 1:30E − 04 1:30E − 04 0.04
600 1800 1790.36 −15:80 1:06E − 04 1:07E − 04 0.55
600 1800 1799.81 −1:75 1:06E − 04 1:06E − 04 0.35
900 250 235.00 0:00 4:56E − 04 4:57E − 04 0.28
900 250 230.55 −16:99 4:59E − 04 4:59E − 04 0.08
900 900 870.00 −15:00 2:38E − 04 2:38E − 04 0.03
900 900 900.69 −8:61 2:37E − 04 2:33E − 04 1.74
900 1800 1792.50 0:00 1:32E − 04 1:32E − 04 0.05
382 I.G. Damousis et al. / Fuzzy Sets and Systems 129 (2002) 371–386

Fig. 8. Frequency distribution of the best fuzzy logic system errors, concerning (a) the amplitude, and (b) the phase of the magnetic
vector potential distribution in the earth around the pipeline neighbourhood including pipeline itself.

separation distance d between the overhead transmis- 5. Performance analysis

sion line and the buried pipeline varies between 70
and 2000 m, the earth resistivity  varies between Real-coded GA proved to be more e:cient than
30 and 1000 U m, coordinate x takes values between the binary-coded one. While both implementations
40 and 2030 m and )nally coordinate y takes val- needed almost the same time to produce a generation,
ues between 0 and −30 m. This range of the input the real-coded GA required fewer generations in or-
variables d; x; y;  in the TDS leads to a FLS, which der to converge (Fig. 7). Someone would expect that
is capable to determine the MVP values in the earth real coding would be faster than binary since there
around the pipeline neighborhood, including pipeline is no need for breaking down the chromosomes and
itself. decoding the FLS parameters, but this advantage dis-
 I.G. Damousis et al. / Fuzzy Sets and Systems 129 (2002) 371–386 383

Table 3
Fuzzy rules’ parameters of the best FLS of the GFS including the centers and standard deviations of the membership functions

Rule no. d 4 p p 0 l l

1 200 247 30 100 0.000261 0 0

2 700 249 30 100 0 0 0.01955
3 1322 165 30 100 0.000018 12.02346 10.86999
4 1899 119 30 100 0.000298 10.43988 13.64614
5 200 247 70 100 0.00003 1.270772 0
6 700 249 70 100 0.00002 0.136852 0.039101
7 1322 165 70 100 0.000002 5.806452 0.273705
8 1899 119 70 100 0.000313 17.10655 6.549365
9 200 247 100 100 0.000187 3.460411 0
10 700 249 100 100 0.000116 17.49756 14.15445
11 1322 165 100 100 0.00004 6.529814 2.776149
12 1899 119 100 100 0.000504 7.781036 19.08113
13 200 247 300 150 0.000604 2.482893 0.078201
14 700 249 300 150 0.000192 11.12415 13.15738
15 1322 165 300 150 0.000124 19.98045 19.98045
16 1899 119 300 150 0.000062 2.776149 19.04203
17 200 247 500 150 0.000033 0.351906 0.938416
18 700 249 500 150 0.000126 3.773216 1.661779
19 1322 165 500 150 0.000015 4.496579 17.43891
20 1899 119 500 150 0.000076 6.744868 17.81036
21 200 247 700 150 0.000625 3.049853 0.097752
22 700 249 700 150 0.000187 19.98045 20
23 1322 165 700 150 0.00025 0.01955 7.44868
24 1899 119 700 150 0.000124 17.65396 9.442815
25 200 247 900 150 0.000064 2.502444 13.56794
26 700 249 900 150 0.000382 11.4565 0.449658
27 1322 165 900 150 0.000002 6.27566 0.54741
28 1899 119 900 150 0.000106 18.96383 17.43891
29 200 247 1000 100 0.000252 13.47019 1.505376
30 700 249 1000 100 0 1.642229 0.097752
31 1322 165 1000 100 0.000128 0.508309 17.1652
32 1899 119 1000 100 0.000155 0.195503 19.33529

appears due to the more complex crossover scheme ration distances d between the overhead transmission
that requires more computational e9ort. On the other line and the buried pipeline as well as various earth re-
hand, using 7oating-point numbers for the representa- sistivities  and di9er signi)cantly from the cases used
tion of the FLS parameters in the chromosome solves for training. The training of the GFS has produced a
the problem of how many bits should be used to rep- knowledge base consisting of m = 32 fuzzy rules.
resent a parameter accurately in a binary GA. Fig. 7 Table 2 summarizes test results where MVP cal-
is a representation of the training process of binary culations by the B-FLS and FEM have been com-
vs. real GA. The curves represent the best out of pared. Absolute errors have been computed as follows:
20 runs. Error = |[(AFEM − AB-FLS )=AFEM ] · 100|. The average
After the training of the GFS down to an average error in amplitude calculation is 2.5% and in phase
training error of 1.8%, the performance of the best calculation 2.06%. For a new con)guration case, the
fuzzy logic system (B-FLS) has been tested in several computing time using the B-FLS is negligibly small
new con)guration cases of the examined electromag- (10−8 smaller) compared to the time needed for FEM
netic )eld problem. These cases have various sepa- calculations.
384 I.G. Damousis et al. / Fuzzy Sets and Systems 129 (2002) 371–386

Fig. 9. Best fuzzy logic system errors, for various con)gurations of the examined electromagnetic )eld problem, concerning (a) the
amplitude, and (b) the phase of pipeline surface magnetic vector potential.

Considering the range of parameters used the fre- current, so the presented results may be easily used
quency distribution of the errors in MVP amplitude for any given fault current IF .
and phase are shown in Figs. 8(a) and (b). It can be Fig. 10 )nally shows the voltage per km induced
seen that 77% of the errors in amplitude and 88% in across pipeline and earth calculated with the proposed
phase are less than 3% (see Table 3). method.
Fig. 9 shows the errors for various parameter ; d
con)guration cases. From Table 2, and Figs. 8 and 9 it 6. Conclusions
is evident that the B-FLS results are practically equal
to those obtained by FEM. It should be mentioned The magnetic )eld and the voltage induced by a
that the MVP distribution is proportional to the fault faulted transmission line on a buried pipeline have
 I.G. Damousis et al. / Fuzzy Sets and Systems 129 (2002) 371–386 385

Fig. 10. Voltage per km induced across pipeline and earth as a function of distance d (cf. Fig. 1) for various soil resistivity values.

been calculated using )nite element method and fuzzy [8] J.H. Holland, Adaptation in Natural and Arti)cial Systems,
techniques. The use of genetic algorithms in determin- Mich. Univ. Press, Ann Arbor, MI, USA, 1975.
ing the optimal rules of fuzzy logic system has been [9] F. Ho9mann, Incremental tuning of fuzzy controllers by
means of an evolution strategy, GP’98 Conf., Madison,
e:cient in providing accurate results. It is shown that Wisconsin, 1998.
an expert system can be built by which interference [10] H. Ishibuchi, T. Nakashima, T. Murata, Genetic-
problems can easily and quickly be solved. The pro- algorithm-based approaches to the design of fuzzy systems
posed expert system was capable of determining the for multi-dimensional pattern classi)cation problems, Proc. of
induced voltage with an average error of less than 3%. IEEE Internat. Conf. on Evolutionary Computation, Nagoya,
Japan, 1996, pp. 229 –234.
[11] H. Ishibuchi, K. Nozaki, N. Yamamoto, H. Tanaka, Selecting
fuzzy if–then rules for classi)cation problems using genetic
References algorithms, IEEE Trans. Fuzzy Systems 3 (3) (1995) 260–
270.
[1] J.C. Bezdek, R.J. Hathaway, Optimization of fuzzy clustering [12] C.L. Karr, Design of a cart-pole balancing fuzzy logic
criteria using genetic algorithms, Proc. 1st IEEE Conf. on controller using a genetic algorithm, SPIE Conf. on
Evolutionary Computation (EC-IEEE’ 94), vol. 2, 1994, pp. Applications of Arti)cial Intelligence, Bellingham, WA, 1991.
589 –594. [13] C.C. Lee, Fuzzy logic in control systems: fuzzy logic
[2] O. Cordon, F. Herrera, A three-stage evolutionary process controller—Part I and II, IEEE Trans. Systems Man Cybernet.
for learning descriptive and approximate fuzzy- 20 (2) (1990) 404–435.
logic-controller knowledge bases from examples, Internat. J. [14] Z. Michalewicz, Genetic Algorithms+Data Structures=
Approx. Reason. 17 (4) (1995) 369–407. Evolution Programs, Springer, New York, USA, 1996.
[3] O. Cordon, F. Herrera, Hybridizing genetic algorithms [15] K.J. Satsios, D.P. Labridis, P.S. Dokopoulos, Voltages and
with sharing scheme and evolution strategies for designing currents induced in a system consisting of a transmission
approximate fuzzy rule-based systems, Fuzzy Sets and line and a parallel pipeline, European Trans. Electrical Power
Systems 118 (2001) 235–255. (ETEP) 8 (3) (1998) 193–199.
[4] D.E. Goldberg, Genetic Algorithms in Search, Optimization, [16] J. Satsios, D.P. Labridis, P.S. Dokopoulos, Finite element
and Machine Learning, Addison-Wesley, New York, USA, computation of )eld and eddy currents of a system consisting
1989. of a power transmission line above conductors buried in
[5] F. Herrera, M. Lozano, J.L. Verdegay, Tuning fuzzy nonhomogeneous earth, IEEE Trans. Power Delivery 13 (3)
controllers by genetic algorithms, Internat. J. Approx. Reason. (1998) 876–882.
12 (1995) 299–315. [17] J. Satsios, D.P. Labridis, P.S. Dokopoulos, Inductive
[6] F. Herrera, J.L. Verdegay (Eds.), Genetic Algorithms & Soft interference caused to telecommunication cables by nearby
Computing, Springer, New York, USA, 1996. AC electric traction lines. Measurements and FEM
[7] J.H. Holland, Outline for a logical theory of adaptive systems, calculations, IEEE Trans. Power Delivery 14 (2) (1999) 588–
J. ACM 3 (1962) 486–493. 594.
386 I.G. Damousis et al. / Fuzzy Sets and Systems 129 (2002) 371–386

[18] P. Silvester, R. Ferrari, Finite Elements for Electrical [21] L.A. Zadeh, Outline of a new approach to the analysis
Engineers, Cambridge University Press, Cambridge, 1983. of complex systems and decision processes, IEEE Trans.
[19] K. Shimojima, T. Fukuda, Y. Hasegawa, Self-tuning fuzzy Systems Man Cybernet 1 (1973) 28–44.
modeling with adaptive membership function, rules, and [22] H.J. Zimmermann, Fuzzy Set Theory and its Applications,
hierarchical structure based on genetic algorithm, Fuzzy Sets Kluwer, Boston, USA, 1996.
and Systems 71 (3) (1995) 295–309.
[20] T. Takagi, M. Sugeno, Fuzzy identi)cation of systems and its
applications to modeling and control, IEEE Trans. Systems
Man Cybernet. SMC-15 (1985) 116–132.