Documente Academic
Documente Profesional
Documente Cultură
School of Electrical and Electronic Engineering, Universiti Sains Malaysia, 14300 Nibong Tebal, Pulau Pinang, Malaysia
Faculty of Electrical & Electronic Engineering, Universiti Malaysia Pahang, 26600 Pekan, Pahang, Malaysia
a r t i c l e
i n f o
Article history:
Received 20 June 2012
Received in revised form 4 September 2012
Accepted 17 September 2012
Keywords:
Electrical installation
Infrared image
Multilayered perceptron
Statistical features
Conditions classication
Building
a b s t r a c t
Maintaining the reliability of electrical installation has become part of the energy efciency practices in
building. The degradation of electrical installations can cause overheating, which can lead to subsequent
failure of the equipments that can potentially result in unplanned power outages, possible injury and
re hazard. In addition, the efciency of an electrical system becomes low prior to failure, thus energy
is spent generating heat and causing unnecessary energy loses. Therefore, early prevention is required
to avoid this situation by monitoring the reliability of the electrical installations through energy audit
practices. This article proposes a semi-automatic approach for evaluating the thermal condition of electrical installations within the building in Malaysia by analyzing its infrared image. Initially the interest
regions of the images are manually segmented. Then the statistical features of rst order histogram and
gray level co-occurrence matrix features as well as the differences of feature parameters between hot
and reference regions are extracted from segmented regions. Principle component analysis is applied for
the best features selection and at the nal stage, the condition of electrical equipments will be classied
using multilayered perceptron neural network. The performances of multilayered perceptron networks
have been compared and tested with various training algorithms. The classication accuracy of multilayered perceptron networks are also compared with discriminant analysis classier and it is found
that the multilayered perceptron network using LevenbergMarquardt algorithm gives the best testing
performance. The result shows that the maximum testing accuracy 78.5% was obtained.
2012 Elsevier B.V. All rights reserved.
1. Introduction
According to the latest report [1] of Fire and Rescue Department
of Malaysia about the causes of re in buildings, 2317 re related
incidents have occurred between January 2012 and June 2012, making the average number of incidents around 387 a month. In this
report, 1049 of res were caused by electrical problems which was
about 46% of the total causes of re in building and were mainly
involved the electrical wiring problems (809 cases) and failure of
electrical equipments (240 cases). Failure in electrical distribution
equipments can potentially produce an ignition and re. One of
the causes of ignition is excessive ohmic heating in electrical distribution [2]. This condition can occur especially for old buildings
with outdated electrical wiring that is deteriorating, inappropriately amended, or insufcient for the electrical loads. However,
new constructions also are not protected from this condition. If
the quality of the connection degrades, in effect, more energy will
dissipate from the devices as its electrical resistance increases [3].
Therefore, periodical monitoring and diagnosis of equipments condition are essential for early fault detection and maintaining the
energy efciency in buildings.
Infrared thermography (IRT) is the best and useful tool to
monitor and determine the heat related problems [48]. Various problems can also be detected within the monitoring such as
poor connections, short-circuits, overloads, load imbalances, and
improperly installed electrical components [9]. The system allows
on-line maintenance process without interruption of service, minimizes downtime, reduces outage and manpower cost, avoids
sudden failure of the equipments that could be catastrophic, injuring and losing of life [1012].
Applying automatic condition monitoring system can improve
the fault detection technique and the level of abnormalities in
electrical equipments can be evaluated even when the expert
or experienced personnel are not present [13,14]. For instance,
Almeida et al. proposed an intelligent faults diagnosis system based
on thermography for surge arrestors using two kinds of variables
as inputs of neuro-fuzzy network [15]. This system was designed to
classify faults into two classes which are normal or faulty and the
index of wrong classication result is found to be lower than 10%.
Shai and Hamzah used RGB colour scale data and temperature
586
Table 1
Classication of conditions of electrical equipments.
Priority
Type of condition
Recommended action
1
2
3
Critical
Warning
Normal
T 15
5 < T < 15
T 5
587
Fig. 1. Thermal image of (a) abnormal condition, (b) warning condition, and (c) normal condition of electrical equipments. (For interpretation of the references to color in
text, the reader is referred to the web version of the article.)
Fig. 2. (a) Original infrared image of a three-phase cable, and (b) greyscale image with reference and hot region selections.
Formula
Maximum intensity
max
l1
Average intensity,
1
nx ny
ip(i)
i=0
l1
Variance, 2
1
nx ny
ip(i)
i=0
l1
(i ) p(i)
i=0
l1
Skewness
1
3 nx ny
Kurtosis
1
4 nx ny
(i ) p(i) 3
i=0
l1
(i ) p(i)
i=0
l1
Entropy
nx1ny
Standard deviation
i=0
l1
2
nx1ny
(i ) p(i)
i log(i)
i=0
588
Table 3
GLCM features.
Features
Formula
l1
l1
Homogeneity
i=0
l1
p(i,j)
1+|ij|
j=0
l1
Energy
p2 (i, j)
i=0
j=0
l1
l1
Entropy
i=0
l1
j=0
l
l
n2
Contrast
n=0
i=1
p(i, j), |i j| = n
j=1
Homogeneity measures the closeness of the distribution of elements in the co-occurrence matrix with respect to its diagonal.
Energy is the sum of square elements in the co-occurrence matrix.
Contrast is a measure of the intensity contrast between a pixel and
its neighbour over the whole image and entropy used to calculate statistical randomness. The homogeneity, energy, entropy and
contrast for the ROI of the hot component are denoted as p1 , p2 , p3 ,
p4 , respectively. Similarly for the reference component: g1 , g2 , g3 ,
g4 are represented for homogeneity, energy, entropy and contrast
of its region. The mutual parameters are calculated by taking the
difference between hot and reference parameters. These parameters
are given as d21 (p1 g1 ), d22 (p2 g2 ), d23 (p3 g3 ), and d24 (p4 g4 ).
But the conditions of equipments are mainly inuenced by the
characteristics of the hot component and difference of parameters
between hot and reference. Therefore, total 22 features are primarily selected neglecting 11 features of reference parameters which
are highlighted in bold as shown in Table 4.
4. Suitable features selection
Features selection is used for reducing data dimensionality. All
extracted features from 500 infrared images of electrical equipments in building are tabulated into three categories. Principle
component analysis (PCA) tool was employed for features selection
from the above mentioned of 22 features. The principle components are calculated using correlation matrix. The result of PCA is
shown in scree plot which is used to justify relative magnitude of
eigenvalues. Fig. 3 shows the scree plot for 22 features of all the
data set. The results of scree plot and eigenvalues as illustrated
in Table 5. Eigen analysis of the correlation matrix for electrical
equipments conditions classication data Table 5. It is found that
the most appropriate number of principle component is four for the
rst principle component (PC1) until fourth principle component
Table 4
Features extracted from both hot and reference region.
Feature type
List of parameters
Reference parameter
Hot parameter
Difference between
hot and reference
h1
h2
h3
h4
h5
h6
h7
x1
x2
x3
x4
x5
x6
x7
d11
d12
d13
d14
d15
d16
d17
Homogeneity
Entropy
Energy
Contrast
g1
g2
g3
g4
p1
p2
p3
p4
d21
d22
d23
d24
589
Table 6
Eigenvectors of PC1PC4 for electrical equipments conditions classication.
Variable
PC1
PC2
PC3
PC4
x1
x2
x3
x4
x5
x6
x7
d11
d12
d13
d14
d15
d16
d17
p1
p2
p3
p4
d21
d22
d23
d24
0.229
0.132
0.084
0.222
0.230
0.112
0.113
0.124
0.074
0.016
0.062
0.008
0.058
0.079
0.320
0.342
0.355
0.237
0.271
0.324
0.343
0.251
0.054
0.137
0.182
0.175
0.178
0.317
0.314
0.244
0.191
0.324
0.151
0.151
0.405
0.403
0.055
0.013
0.021
0.080
0.158
0.174
0.157
0.143
0.313
0.259
0.299
0.147
0.208
0.275
0.272
0.231
0.352
0.013
0.342
0.380
0.028
0.124
0.023
0.104
0.030
0.198
0.040
0.076
0.015
0.133
0.077
0.020
0.071
0.315
0.258
0.260
0.244
0.339
0.282
0.318
0.099
0.125
0.119
0.002
0.064
0.214
0.187
0.369
0.269
0.012
0.034
0.260
Table 7
Variables that have strong relationship for PC1PC4.
PC1
PC2
PC3
PC4
p1
p2
p3
d22
d23
x6
d13
d16
d17
d12
d14
d15
x4
d11
p4
Eigen value
Proportion
Cumulative
PC1
PC2
PC3
PC4
PC5
PC6
PC7
PC8
PC9
PC10
PC11
PC12
PC13
PC14
PC15
PC16
PC17
PC18
PC19
PC20
PC21
PC22
6.8786
4.9777
2.7349
1.9782
1.2864
0.9567
0.7587
0.6125
0.4627
0.3615
0.2719
0.2160
0.1319
0.1149
0.0958
0.0761
0.0350
0.0168
0.0155
0.0121
0.0033
0.0027
0.313
0.226
0.124
0.090
0.058
0.043
0.034
0.028
0.021
0.016
0.012
0.010
0.006
0.005
0.004
0.003
0.002
0.001
0.001
0.001
0.000
0.000
0.313
0.539
0.663
0.753
0.812
0.855
0.890
0.917
0.938
0.955
0.967
0.977
0.983
0.988
0.993
0.996
0.998
0.998
0.999
1.000
1.000
1.000
590
Table 8
Input and output parameters for MLP neural network.
Number of data
Training data
Testing data
Input features, zn = 15
Table 9
Classication performance for testing data set.
500
300
200
p1 , p2 , p3 ,d22 , d23 , x6 ,d13 , d16 , d17 , d12 ,
d14 , d15 , x4 , d11 , p4
Critical, warning and normal condition
Output classes, ym = 3
nh
2
wjk
F
j=1
n
wij1 zi (t) + bj
i=1
(1)
In this experiment, fteen input nodes and three output nodes
were being used. The target outputs are classied into three categories i.e. normal, critical, and warning condition. The F() is the
sigmoid activation function that is commonly used in the MLP net2 and
work which is also used in this study. The weights wij1 , wjk
threshold bj are unidentied and should be chosen for reducing
the error in prediction. The prediction error is dened as follows:
k (t) = yk (t) y k (t)
(2)
yk (t) are the actual (i.e. target) and predicted outwhere yk (t) and
put, respectively.
6. Classication performance analysis
The entire data set containing 500 images are divided into training data and testing data set. In this research, 60% of the data are
selected randomly as for data training and remaining 40% of the
data are used for testing. Table 8 summarizes the input and output
parameters for classication.
MLP network has been trained and compared by using twelve
different training algorithms namely LM, BFG, RP, SCG, CGB, CGF,
CGP, OSS, BR, GD, GDX and GDM to nd out the best training algorithm for classication. The numbers of hidden neuron are varied
from 1 to 50 for the networks to nd the best performance accuracy.
The number of epochs is kept at constant 2000 for all networks. For
comparing the performance of different classiers, the percentage
of true classication for the normal condition (Accnormal ), warning condition (Accwarning ), critical condition (Acccritical ) and overall
(Accoverall ) are evaluated, which can be dened as follows:
Training
algorithm
No. of
hidden
nodes
Accoverall
(%)
Accnormal
(%)
Accwarning
(%)
Acccritical
(%)
LM
BFG
RP
SCG
CGB
CGF
CGP
OSS
BR
GD
GDX
GDM
9
19
21
14
29
39
29
10
10
34
32
21
78.5
76.5
77.0
78.0
76.5
77.5
78.0
78.0
76.0
58.5
74.5
63.0
90.7
81.8
83.6
85.5
73.5
89.4
83.8
89.4
75.0
71.2
90.0
76.4
59.7
72.6
68.5
70.3
79.5
67.1
74.1
66.3
75.3
50.6
63.9
45.9
86.8
74.0
84.1
78.0
77.1
79.6
76.5
82.4
78.7
52.0
70.2
68.5
Table 10
Classication using discriminant analysis.
True group
Put into group
Critical
Warning
Normal
Total N
N correct
Accuracy (%)
Critical
93
30
2
125
93
74.4
(3)
Accnormal =
(4)
Acccritical =
Normal
Overall
36
117
40
193
117
60.6
1
33
148
182
148
81.3
130
180
190
500
358
71.6
Accoverall =
Accwarning =
Warning
(5)
(6)
(normal: 71.2%; warning: 50.6% and critical: 52.0%). The results also
show that the MLP network trained with LM need lower number of
hidden nodes which followed by OSS, BR, SCG, BFG, RP, GDM, CGP,
CGB, GDX, GD and CGF training algorithms.
Table 10 shows the classication performance of three conditions of electrical equipments using discriminant analysis (DA)
classier. Total numbers of data and input features are 500 and
15 respectively. The classication accuracy for critical condition is
74.4% (93 critical conditions correctly classied out of 125), 60.6%
DA
Accoverall (%)
Accnormal (%)
Accwarning (%)
Acccritical (%)
86.8
90.7
59.7
78.5
71.6
81.3
60.6
74.4
100%
(7)
591