Artificial Neural Network Assisted Digital Image Processing to Determine the Hydrophobicity of Polymeric Materials

glibdoadingAI and Robotics

Oct 20, 2013 (3 years and 10 months ago)

83 views

Artificial Neural Network

Assisted Digital Image Processing to Determine the
Hydrophobicity of Polymeric Materials


Daniel Thomazini
a
,
Raphael Ulisses Costa de Resende
b
, Daniel Henrique Gueratto
c

and
Maria Virginia Gelfuso
d


Universidade Federal de Itaju
bá, Av. BPS

1303, Itajubá,
MG
, Brazil
.

a
thomazini@unifei.edu.br
,
b
raphaelres@gmail.com,

c
danielgueratto@ig.com.br,

d
mvgelfuso@unifei.edu.br


Keywords
:
Digital image processing, artificial neural network, hydrophobicity, polymer
.


Abstract
.
Hydrophobicity of polymeric insulators
of high power transmission line
is
an

important

property
because it

is a good monitor of aging of polymeric outdoor insulator.
An employee on the
high
-
voltage transmission line makes this procedure

during operation, which can lead to injuries and
incorrect estimation of the insulator integrity. In this way, digital image processing is a
promising

and
objective tool to
analyze

the polymer surface.
In this study,
two thousand
pictures were taken
from
the wetted insulator surface and
analyzed

by artificial neural network assisted digital image
processing. The neural network used is based on back
-
propagation method, and Haralick
’s

descriptors were used to
quantify

the hydrophobic aspect of various polyme
ric aged surfaces.


Introduction


Polymers have contributed to replace glass and ceramics insulators due to the lightweight, lower
susceptibility to breakage and excellent hydrophobic surface. They have been us
ed with success

on
high pollution environments and on transmission system with compact towers

[1]
.

The hydrophobic behavior of a material is directly related to the surface energy that acts as a driving
force between the phases. The surface energy is defined as the free
energy increased in a system on
creating a unit area of new surface at constant temperature, pressure and composition. The force
required to extend a liquid surface per unit length is
related to
the surface tension and is numerically
equal to the surface e
nergy. In a system with a
defined
number of phases, the equilibrium
distribution is determined by the condi
tion that interface energy is
minimum. This requirement fixes
the contact angle for a liquid
-
solid
-
vapor system with suitable limitations

[2]
.

The hy
drophobicity surface of polymeric insulators is widely desired because it prevents the
formation of a water path for electric current, preventing flashover and discharges on the insulation,
causing electric breakdown

[3]
. However, some outdoor conditions s
uch as, ultraviolet radiations,
snow or salt can accelerate their aging process as well as reduce the hydrophobicity of polymers

[4,5,6]
. Fact that requires improvements in maintenance processes of electrical systems that use
these insulators. In this cont
ext, the monitoring of the hydrophobicity of insulator surfaces directly
in the field has become an important issue in power electric utilities.

Some methodologies to determine the surface hydrophobicity of electrical insulators come from
STRI Guide

[7]

or

from IEC TS 62073

[8]
. The first technique basically consists of wetting the
surface of the insulator with water and then taking pictures. By comparing them with standard
photos taken on laboratory conditions it is possible to identify modifications on th
e surface
hydrophobicity. The second method is based on the measurement of the contact angle of droplets on
the surface; which is more difficult to be performed in the field since the surface must be placed
horizontally. Therefore the STRI method is genera
lly used for monitoring outdoor insulators,
although it has a rough hydrophobicity classification (HC), since it compares subjective images,
taken on poor illumination conditions, with seven pattern images, which are defined from HC01 to
HC07 (from fully h
ydrophobic to fully hydrophilic texture).


The Artificial Neural Networks (ANN) is based in a mathematical model inspired by the biological
structure of
neurons that

acquire knowledge through experience and repetition

[9]
.

Among the
various types of ANN models, in this study was used the Back
-
Propagation model, which operates
on a set of output network.

During the training of the ANN with the algorithm, the network operates
in a sequence of two
-
steps. In the first step, a p
attern is presented to the input layer of the network
.
The processing of the data goes across the network, layer by layer, until the answer is
available

in
the out
put

layer
. In the second step, the output data is compared to the desired output value for a
particular pattern. If this value is
different
, the error is calculated and the parameters of the network
are readjusted. So, the error is propagated from input layer to output layer.
Fig.

1 illustrates the
Back
-
Propagation Neural Network
Model.



Figur
e

1.
Example of Artificial Neural Network based on Backpropagation method
.


The use of digital image processing has become a technique to identify patterns and shapes in a fast
and accurate way

[10]
, bing suitable to this study. In this
study
,
artificial
neural network, based on
back
-
propagation assisted four

H
aralick’s textures descriptors

as entropy, energy, variance and
homogeneity

were evaluated
. The
se

processing methods were chosen due to their abilities to identify
similar textures

[11,12]

and the ph
otos were taken from samples obtained from the experimental
procedure to prepare hydrophobic surfaces, described in

[11,13]
.

In this study, the Input Layer was filled with Haralick’s descriptors values, and the Output Layer
delivered the values related to
the %WIA solutions, regarded to the
hydrophobic

behavior

of the
surface.


Experimental Procedure


Ten specimens of silicon rubber with 10x10 cm
2

and 5mm thickness were sprayed with a solution of
water and isopropyl alcohol (WIA) at different concentrations (from 0 to 100%)
[13]
. These
solutions were identified as percentage of alcohol (%WIA). The parameters of the process to obtain
the silicon rub
ber in industry are very well defined, and it is kwon that this material recovers the
hydrophobicity behavior after some shelf time

[14]
.
Im
ages from each WIA solution were taken
from each specimen to prevent surface modification an
d to guarantee the relia
bility
of the
measurements. This solution was adopted due to the excellent results from previous works
[13]
,
which showed high chemical stability of the materials employed and reproducibility of the results
due to the low dispersion.

One hundred p
hoto images
were taken under natural illumination condition from each sample
wetted wi
th different percentages of WIA
.
Each picture was divided into four, so t
his procedure
provided
4.4
0
0
grayscale
images, which were classified according to the STRI guide.

Six

of these
images are presented in
Fig. 2
. The computational procedure was performed in MatLab® v7.12 and
Image Processing Toolbox v7.2 for Mac. The hardware used was an iMac with Intel i3 3.2GHz and
12G
b

of RAM.

In order to avoid the influence of natur
al conditions of illumination,
White Top
-
Hat filter
[15]
,
which gives an enhancement of the images,

was

used.
This light adjustment

was

used according to
previous studies
[13]
, which show lower deviation and higher correlation in the results observed for
these analyses.




Figure
2
.

Polymeric surface insulator aspects of a) 0%WIA, b) 20%WIA, c) 40%WIA, d) 60%WIA,
e) 80%WIA and f) 100%WIA.


The
Artificial Neural Network
(
ANN
)

used in this study is based on Back
-
Propagation method

[10]
.
The number of characteristics to image classifica
tion is based on initial
values that

were related to
four Haralick’s descriptors: Entropy, Energy, Variance and Homogeneity.
These descriptors we
re
tested in previous works and show

a hig
h
coherence

with %WIA solutions.

To determine the
efficiency

of image recognition, the network for training was
configured with three
layers, first one is the Input Layer

that was related to the images and desired HC;

the next one is the
Hidden
, that calculates each Haralick’s descriptor;

and the last is the Output Layer
, that gives the
WIA value. This procedure is illustrated

as shown in
Fig.

3
. The best number of data pr
ocessing in
the network (Epochs) was
10
00

iterations
. All other configurations of the ANN were the
standard
values in the network creation.


a)

b)

c)

d)

e)

f)


Figure 3. ANN interface used to evaluate the hydrophobic classification of the polymers.


In Hidden Layer, each H
aralick’s descriptor
corresponds

to
one

neuron in the network
.
The
deviation

of

the results in Output Layer can be adjusted related to the desired value
, changing the
configuration of the network as number of layers, neurons and training parameters.


Results and Discussions


Fig.

4 shows the evaluation
of the ANN
.

It can be
seen
that the correlation
of the T
raining

(Fig. 4a)
,

Validation

(Fig. 4b)

and Test
(Fig. 4c)
were

lower than those o
btained in previous study

[13]
. It

indicates that the use of these four Haralick’
s descriptors in ANN c
ould interfere negatively in the
final evaluation of the images

classification.



Figure 4. Correlations of a) training, b) validation, c) test and d) all
evaluation
s

of the ANN.


a)

b
)

c
)

d
)


Figure 5.

Training steps of ANN.


The training steps of ANN are presented in Fig. 5. It can be seen that after about 1000 steps
(Epochs), the ANN
tends to convert,
ind
icating that the input values
in the ANN show a
considerable dispersion, which
could

also be obser
ved in the correlation values
shown in Fig.

4.

Previous study [16] shows that in
an

ANN with lower input dispersion values, the convergence of
the training was obtained with 200 Epochs.


C
onclusions


To evaluate the hydrophobicity of polymeric outdoor insulator surfaces, which are direct associated
with the aging of materials,
four

Haralick’s descriptors,
Entropy, Energy, V
ariance
and
H
omogeneity

were
associat
ed to Artificial Neural Network
. S
amples of silicon rubber sprayed with
alcohol aqueous solutions were prepared.
These s
olution
s with
in

different

proportions were used
to
simulate the transition of a hydrophobic polymer surface to a hydrophilic surface. This methodology
has shown to be

very convenient to produce standard and reproducible samples with dif
ferent levels
of hydrophobicity.
As observed in previous studies, Haralick’s descriptors allowed a best correlation
i
n a shorter time of processing
.

The association of Artificial Neural
Networks and Digital Image Processing to classify the
hydrophobicity

of polymeric
surfaces

is
a brand new procedure th
at allowed
obtains

the
hydrophobic
behavior

of these materials.

In this study, the high dispersion in the Input Layer,
promoted to the low
er correlation values the when compared to the linear fit of Entropy descriptor.


Acknowledgments


The authors thanks to Conselho Nacional Desenvolvimento Científico e Tecnológico (CNPq) and
Fundação de Amparo à Pesquisa do Estado de Minas Gerais

(FAPEMIG) for financial support.


References

[1]

R. S
.
Gorur, E. A. Burnham
:

Outdoor Insulation
.
Ravi S. Gorur, Inc (
1999
)
.

[2]

W. D.
Kingery
:

Fabrication Process
. John Wiley & Sons (
1958
)
.

[3]

J. G.

Wankowicz, S. M. Gubanski And W. D. Lamp
:
IEEE
Transactions on Dielectrics and
Electrical Insulation
, Vol. 1, n. 4 (1994) p.

604
-
614.

[4]

B

Venkatesulu, M. J. Thomas, A. M. Raichur
, in:
Proceedings of the Annual Report Conference
on Electrical Insulation Diel
ectric Phenomena. Quebec: IEEE (2008)

p. 67
-
70.

[5]

D.
Allan
et al.
:
IEEE Transactions on Electrical Insulation
, Vol
. 27, n. 3,
(1992) p. 578
-
585
.

[6]

N.

Yoshimura, S. Kumaga, S. Nishimura
:
IEEE Transactions on Dielectrics and Electrical
Insulation
,
Vol.
6, n. 5,
(
1999
) p.

632
-
650.

[8]

IEC TS
62073
:

Guidance on the measurement of wettability of insulator surfaces
. International
Ele
ctrotechnical Commission
. 2003.

[7]

STRI Guide 1/92
. Sweden Transmi
ssion Research institute
. 1992.

[9]

J. C.
Russ
:

The Image Processing Handbook
.
CRC Press (
2007
)

[10]

I.
Pitas
:

Digital Image Processing Algorithms and Applications
.
John Wiley & Sons (
2000
)
.

[11]

D.

Thomazini
, M. V.

Gelfuso
, R. A. C.
Altafim:
Materials Research
,
Vol.
44,
n.
4
(2008)

p.

415
-
419,

[12]

R. M.

Haralick,
K.

Shanmugam,
I.
Dinstein
:

IEEE
Trans. on System, Man and Cyb.
,

Vol.
3, n.
6

(
1973
) p
. 610
-
621.

[13]

D.

Thomazini,
M. V.

Gelfuso, R. A. C. Al
tafim:

Materials Research
, São Carlos,
Vol.
15, n. 3,
(
2012
) p
. 365
-
371.

[14]

M.

Berg,
R.

Thottappillil, V.
Scuka, in:

Annual Report Conference on Electrical Insulation and
Dielectric.
(
1999
)
.

[
15
]

R. C.

Gonzalez, R. E. Woods
:

Digital image processing
. Addison
-
Wesley Publishing Company,
(
1992
)
.

[16]

D.
Thomazini
, M.

V. Gelfuso, R.A.C. Altafim, S. C. Izidoro, F. M. Teixeira
, G. B. de Oliveira
,
in: Proceddings of 19
o

CBECIMAT
. Campos do Jordão

(2010).