Applied Thermal Engineering 61 (2013) 220e227

Contents lists available at ScienceDirect

Applied Thermal Engineering

journal homepage: www.elsevier.com/locate/apthermeng

Application of infrared thermography for predictive/preventive

maintenance of thermal defect in electrical equipment
A.S. Nazmul Huda*, Soib Taib
School of Electrical and Electronic Engineering, Engineering Campus, Universiti Sains Malaysia, 14300 Nibong Tebal, Pulau Pinang, Malaysia

h i g h l i g h t s

Intelligent thermal defect detection using infrared thermography.

Use of statistical features as inputs of classier.

Applications of MLP network and discriminant analysis as classier.

Selection suitable features using discriminant analysis.

Classication accuracy up to 82.40%.

a r t i c l e i n f o a b s t r a c t

Article history: Internal heat of electrical equipment rises due to various reasons such as contact problems, unbalance
Received 19 January 2013 loading, cracks in insulation, defective relays or terminal blocks, etc until eventual failure which can
Accepted 25 July 2013 cause unplanned outages and possible injuries. Therefore, early prevention of thermal abnormalities in
Available online 2 August 2013
electrical equipment is necessary to avoid failure. Nowadays a fast, reliable, non-contact and cost
effective infrared thermographic inspection system is being utilized widely for detecting where and how
Keywords:
internal and external defects occur in equipment. The system collects data for investigating thermal
Infrared thermography
condition through the analysis of thermal image of equipment. The image is captured by infrared camera
Electrical equipment
Preventive/predictive maintenance
without interrupting the running operation of system. Both preventive and predictive maintenance
Intelligent system programs are adopted for defect diagnosis. Actually, the frequency and sequence of thermographic in-
Discriminant analysis spection depend on various factors including thermal condition, expense of a failure, type of production
and age of equipment, etc. The objectives of this article are to discuss about preventive and predictive
maintenance, thermal defects in electrical equipment and also an intelligent thermal defect identica-
tion system for accelerating the predictive defect diagnosis technique. The intelligent system proposes
the application of articial neural network and statistical features to detect the existence of defect within
equipment by monitoring its thermal condition. Using discriminant analysis, the optimum features were
chosen as the inputs of the neural network. The performances of the neural network were compared
with the performances of discriminant analysis classier. The comparison results showed that
discriminant analysis classier produced better performance with accuracy 82.40%.
2013 Elsevier Ltd. All rights reserved.

1. Introduction picture and measures temperature variation of the object surface

All objects with a temperature above absolute zero emit
infrared radiation and cause an increase in temperature. The naked
eyes cannot see the infrared energy because it is emitted as heat
energy. Infrared thermography is a technique for converting
invisible heat into a visual image which shows the heat picture of
the object surface. The colour of the object surface in the heat
picture varies with the surface temperature. An infrared camera or
sometimes called thermal imager is used to capture the heat
* Corresponding author. Tel.: 60 645996012.
E-mail address: nhudaeee0450@gmail.com (A.S.N. Huda).
[1,2]. For several years, thermographic inspection has become an
important tool for preventive and predictive maintenance of sur-
face defect in various materials due to its non-invasiveness, safety
and relatively low cost approach. Therefore, numerous studies
were executed to prove thermography as a useful technique in
various applications [3e11].
Heat energy is an important factor in electrical equipment for
increasing operational reliability. Electrical current passes through
a resistive component and generates heat. Thermal energy gener-
ated from an electrical component is directly proportional to the
square of the current passing through it and resistance (I2R Loss).
Therefore, an increase in electrical resistance results an increase in

1359-4311/$ e see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.applthermaleng.2013.07.028
 A.S.N. Huda, S. Taib / Applied Thermal Engineering 61 (2013) 220e227
heat. With the time and age, the condition of the electrical
component gets deterioration due to various reasons such as loose
connection, corrosion and insulators crack, etc. As component de-
teriorates, its resistance increases and also generates more heat.
The increase in heat energy can cause the failure of electrical
equipment and also re may break out. By utilizing thermography
for inspecting the equipment under loading condition, the defec-
tive component can be identied and classied by its level of
severity [12].
Infrared thermography senses the heat which is produced in
the electrical equipment. The thermal proles of electrical
equipment and connectors are captured by using a thermal
imager. Thermal prole which is also called thermogram consists
of a thermal picture and a temperature scale. The different colours
of temperature scale represent the different temperature spots of
the equipment. Heating components are generally detected as red
or brighter coloured areas in an infrared image. A circuit breaker
(CB) thermogram obtained in the present study is shown in Fig. 1.
In Fig. 1a, the brighter area (R phase) indicates the higher resis-
tance and results overheating of that phase. The probable reason
of this high resistance could be a bad or dirty connection. Fig. 1b
shows the temperature prole across the line L0 showing the
variation of temperature in the equipment surface. Thus defects in
components are identied and then recorded. Finally, the main-
tenance programs are planned and carried out before breakdown
of equipment.
Maintenance cost of electrical equipment plays an important
role for reducing the total operating cost of electrical power system.
The maintenance cost can be increased due to the lack of infor-
mation about the maintenance scheduling and kind of mainte-
nance needed for the equipment. Typically, the operating
performance of electrical system is not monitored periodically
which can play an important role in early prevention of a fault and
for increasing the life of equipment. Instead, the schedule for
maintenance is determined by using the failure history of equip-
ment or according to the experiences of maintenance personnel.
Infrared thermography is used to monitor the performance of
equipment which offers a complete information regarding the
operating status of equipment for increasing production,
decreasing or removing unnecessary repairs, preventing eventual
failure and decreasing the maintenance cost of equipment.
Usually, applying infrared thermography technology, the elec-
trical thermal defect and level of defect are identied by moni-
toring its Delta T (DT) criteria [13]. The technique is widely known
as qualitative based temperature measurement system [14]. The
Delta T criteria of a component is dened as an increase in tem-
perature value above the temperature of a reference value, which is
typically the ambient temperature, temperature of a similar
Fig. 1. a. Thermogram of R phase overheating in CB b. Temperature prole across line L0.
221
component under the same condition or the maximum allowable
temperature of the component [15]. Several standards for DT value
are found such as Electrical Testing Association (NETA) [15], Na-
tional Fire Protection Association (NFPA) e NFPA 70-B [16] and
InterNational American Society for Testing & Materials (ASTM) e E
[17], etc. Sometimes thermographers prefer to classify thermal
condition based on their own inspection experiences instead of
using the available standards.
Articial neural networks (ANNs) are potentially powerful,
robust and adaptive mathematical model for pattern recognition
and classication [18,19]. Nowadays ANNs are being utilized to
solve complex non-linear real world problems. Several articles
about the application of articial neural networks (ANN) for elec-
trical fault diagnosis system using infrared thermography have
been proposed in the past few years. In these articles, different
features were used as inputs of ANN. Almeida et al. [20] used a
neuro-fuzzy approach for fault diagnosis of lighting arrester while
the inputs of the ANN were thermographic and identication var-
iables. The system achieved resulting index of wrong diagnosis
lower than 10% for classifying thermal condition into normal, sus-
picious and faulty classes. Shai and Hamzah [21] proposed the
application of multilayered perceptron network. The RGB colour
scale data and temperature data were used to classify internal faults
into four classes namely low, intermediate, medium and high. The
result obtained 99.38% of accuracy. Rahmani et al. [22] developed
an intelligent system applying support vector machine (SVM) as a
classier and 22 image features of Zernike moments for diagnosing
of ground substation faults. The diagnosis was carried out on 20
thermal images and was able to obtain 68.42% of accuracy. Smed-
berg [23] and Wretman [24] proposed an intelligent classication
system based on ANN for diagnosing three-phase fuses and
different forms of connections problems. The four input parameters
of ANN were used which are absolute maximum temperature,
relative maximum temperature, mean temperature difference
compared to the other regions of the image and histogram distance
to the other regions of the image. The test error rate using all four
feature parameters as input of ANN was 9.5% and the error rate
using only the histogram distance as input was 31.2%. The set of
data was 74 infrared images. One of the disadvantages of working
with a small dataset is that the reliability of the results can some-
times be questionable.
This paper will investigate the application of infrared ther-
mography technology as predictive/preventive maintenance pro-
gram to identify the presence of thermal defect in electrical
equipment. The technique uses multilayered perceptron network,
statistical features and discriminant analysis classier to charac-
terize the thermal status of hotspot into defect and no defect
categories. Discriminant analysis classier was utilized to select
222 A.S.N. Huda, S. Taib / Applied Thermal Engineering 61 (2013) 220e227
suitable features and also to build the linear regression model of
each category using suitable features.
2. Preventive/predictive maintenance
Typically, for maintaining the equipment, two types of ap-
proaches, either run-to-failure or preventive maintenance are used.
The run-to-failure approach is simple and straightforward, where
there is no outow of money on maintenance purpose before
eventual failure of equipment. The approach waits for equipment
failure before any maintenance action is taken. Therefore, the
method is more expensive. Thermography based diagnosis system
allows the preventive/predictive maintenance for early prevention
of equipment failure without interrupting the running operation
and thus saves money.
According to the historical data in USA [25], effective use of
preventive/predictive maintenance will reduce about 33%e50% of
maintenance cost that are wasted by most of the manufacturing
and production plants and also introduced the following benets
from various aspects:
1. Reduced unscheduled downtime or system ultimate failure
2. Increased the proper utilization of manpower
3. Increased the production capacity of the plant
4. Reduced the maintenance expenditure of equipment
5. Increased the life time of equipment
2.1. Preventive maintenance
Maintenance of electrical equipment according to the statistical
or historical information of operating capacity, failure history,
mean-time-to-failure (MTTF) instead of tracking equipment per-
formance is referred as preventive maintenance [25]. Preventive
maintenance program schedules repairing and rebuilding activities
for equipment. Suppose, an electrical equipment normally runs 10
months before needed any repairing. Using preventive technique,
the equipment will be removed from service and rebuilt after 10
months of operation. In this example, the schedule of repairing of
equipment is predened which is determined by maintenance
personnel. Now, if there is no need to rebuild after 10 months, then,
there would be wasted of labour and material. Again, if the equip-
ment fails before 10 months, then, it is necessary to x problem
after failure which is usually more expensive than schedule main-
tenance. This is a just random decision which is not supported by
any kind of realistic data.
2.2. Predictive maintenance
Maintenance of electrical equipment based on the direct mo-
nitoring of the actual operating condition, regularly collecting
measurements, efciency, heat distribution and other indicators
instead of depending on statistical or historical data is called pre-
dictive maintenance. The program schedules all maintenance ac-
tivities based on factual data and if necessary, repairs before the
system failure.
Early detection of imminent equipment failures and preventing
them before critical condition signicantly reduce downtime, cost
for maintenance, maximize uptime.
3. Defect in electrical equipment
Electrical installations such as transformer, circuit breaker, fuse
and capacitor, etc can cause faults when its internal temperature
reaches at an abnormal level. Most of the anomalies are due to
various reasons such as contact problems, unbalance current dis-
tributions, cracks in insulators, defective relays or terminal blocks,
etc. Some of the examples of the defects in electrical system which
can be revealed by thermographic technology are mentioned here.
3.1. Transformer
Heat is generated due to the losses in transformer and temper-
ature increases within the core and windings. The transformer is
generally cooled by either oil or air and the operating temperature is
higher than ambient. The cooling system helps to maintain the rise
of temperature within tolerable limits. Usually, the thermographic
inspection is utilized for thermal defects diagnosis of oil-lled
transformers instead of dry transformers. Since, normally, the coil
temperatures of the dry transformers are so much higher than
ambient, it is not easy to distinguish the internal defect using ther-
mography before it becomes serious. Therefore, other monitoring
tools including built-in temperature and pressure gauge can
perform more accurate measurement for the dry transformer. In the
oil-lled transformers, thermographic inspection usually detects
the defects occurring at the primary and secondary bushing con-
nections, cooling ns, cooling fans and internal bushing connec-
tions, etc. The problems are detected visually by the following ways:
U Primary and secondary bushing connections e The loose or dirty
connections, unbalance and overloading problems can cause
overheating in the connections. Thermographic inspections
show hotspots at the defected phase connection compared to
other phases.
U Cooling ns e In an oil-lled transformer, cooling is provided by
the circulation of oil. The oil gets heated due to the heat generated
in the core and windings. Therefore, cooling ns normally appear
warm compared to the inactive cooling n. From the thermal
image, if it is seen that one or more ns have comparatively low
temperature or cool compared to other warm ns, then it can be
concluded that the circulation of oil is being limited in those ns,
probably due to the low oil level.
U Cooling fans e In higher rating transformer, where more heat
will be generated, cooling fans are used to increase the cooling
rate. In general, the fan in action is normally warm and thus
thermogram will show whether it is working or not.
U Internal bushing connections e If there is any defect in the con-
nections, it will be higher temperature than surface temperature.
3.2. Defect in circuit breaker
A circuit breaker (CB) always senses the heat produced by the
current. If it nds the abnormality of heat, instantly opens the cir-
cuit automatically and protects the electrical equipment from
damage. Thermographic inspection has the advantages to nd the
abnormalities by analysis the surface temperature variation of cir-
cuit breaker and can prevent these abnormalities earlier.
3.3. Fuse connection
Fuses are used to protect the system against overloading con-
dition. Overloading condition in a system can increase the fuse
connection heating and moreover, the fuse clips can be overheated
due to the loose and dirty connections.
3.4. Defect in contact
Electrical connection switches, breakers and other reclosing
devices operate with contacts. If these contacts become loose,
 A.S.N. Huda, S. Taib / Applied Thermal Engineering 61 (2013) 220e227
oxidized, dirty or anyhow increase in resistance, then the defect can
be occurred in the contacts.
3.5. Insulation problem
Insulation failure can cause short circuit current between two
conductors. Overcurrent generates overheating and allows the circuit
breaker or fuse to open. Poor insulation can also cause overheating.
3.6. Utility connection box break elbows
In residential areas, many underground connection and
switching locations are in utility box enclosures. Normally, break
elbows provide a point of disconnection and if these break elbow
connections are not installed properly or connection resistance
increases, then faults can occur.
3.7. Surge arrestors problem
Problems with surge protection and lightning arrestors leaking
to ground and current tracking over insulators can also be detected
using thermography. However, nding such problems requires the
capture of subtle temperature differences which are often too
difcult to be monitored.
4. Intelligent defect identication system
4.1. Image acquisition and Delta T criteria analysis
The infrared camera manufactured by Fluke Ti25 was used to
produce the thermogram of equipment. Its scale of temperature
measurement was 20  Ce350  C. The thermal imager consists of a
160  120 focal plane array, uncooled microbolometer detector and
operates in the infrared spectral band of 7.5 mme14 mm. The ther-
mal lens capture images of 320  240 pixels while the ordinary lens
produce 640  480 pixels (visual images). For capturing the image,
the thermal imager orientation was directly facing to the target
equipment in order to get an accurate measurement. The distance
between the target electrical equipment and the thermal imager
was in the range of 0.5e1.0 m. The emissivity value was set at 0.95
as recommended generally for electrical installations thermog-
raphy [26]. It was noted that the ambient temperature around the
equipments was about 30e33  C during the inspection. A total of
368 thermal images of circuit breakers operating at the same
loading condition having 500 hotspots were captured.
In this study, the difference of maximum value between hotspot
and normal (reference) temperature of an equipment was deter-
mined as DT criteria. The method is widely known as qualitative
temperature measurement system. The main benet of this method
is that there is a minor effect of emissivity on temperature mea-
surement. On the other hand, main demerit is that qualitative
method cannot describe the defect, if failures of all three phases
occur at the same time due to the absence of reference component
[11].
In this research, depending on the equipment operating tem-
perature, initially the thermal conditions of equipments were
manually classied into three categories based on priority level
(priority level 1, 2 and 3) using DT specications in Table 1. From these
three, the statuses can be further divided into two categories, defect
(priority 1 and 2) and no defect (priority 3) class. The DT values and
necessary steps regarding the classes have been described in Table 1.
Figs. 1e3 show the thermal status of these thermograms. In Fig. 1a
the maximum temperature value of hotspot A0 is 42.9  C, while
36.9  C is the maximum temperature value of A1 which is assigned as
the reference area. The temperature difference (6  C) between these
223
Table 1
DT specications of defect and no defect classes with recommended actions.
Class DT ( C) Recommended actions
Defect DT  15 Major discrepancy; repair immediately
(Priority 1)
5 < DT < 15 Probable deciency; repair as time permits
(Priority 2)
No defect DT  5 Minor overheating; warrants investigation
(Priority 3)
two phases reveals that A0 is in defect class. Fig. 2a shows the
maximum temperature value of A1 and A2 is 72.7  C and 64.4  C
respectively, while 57.5  C is the maximum temperature value of A0
or reference area. The temperature difference between these two
phases reveals that both A1 and A2 are in defect class. Similarly from
Fig. 3, it is seen that both A0 and A1 areas are in no defect class with
respect to A2 region. Figs. 1b, 2b and 3b show the temperature prole
across line L0 for their respective thermogram.
4.2. Features extraction
Initially, the original thermal image is converted into greyscale
image which consists of black and white regions. The regions of
interest (ROI) for both the reference and hotspot were selected
manually in a rectangular shape. Fig. 4 shows the grey image of
thermogram (Fig. 1a), while white or brighter region indicates
defect.
Total fteen features of rst, second order statistical features
from both ROI and also mutual features (i.e. numerical difference
between the feature value of reference and hotspot of same
equipment) were evaluated [27,28]. The two rst order features
were kurtosis, standard deviation and four second order features
were entropy, contrast, homogeneity and energy. On the other
hand, a total of nine mutual features were extracted. In the case of
rst order, seven were average grey intensity entropy highest grey
value, kurtosis, skewness, standard deviation, and variance. From
second order, two features were entropy and energy. A list of fea-
tures is displayed in Table 2.
5. Suitable features selection
After completing the feature extraction process, all the 15 fea-
tures for 500 hotspots could be used as the inputs of the MLP
network. But the insignicant feature reduces the performance of
the network. Therefore, before using as inputs, suitable feature
selection is an important task to eliminate the noisy features and
enhance network performance. In the present study, discriminant
analysis (DA) has been proposed to select optimum features,
determine the performance of each class and also to calculate the
overall performance [29]. The 15 features were denoted as X1, X2, Y1,
Y2, Y3, Y4, X11, X12, X13, X14, X15, X16, X17, Y11 and Y12. The univariate
and multivariate tests were applied to analyze the performance of
features. The features with the p-value distribution less than 0.05
were selected as suitable features. The univariate test is a test of
equality of each class mean of individual feature which is used to
investigate the p-value distribution of each feature individually.
Table 3 shows the results of univariate test for the original 15 fea-
tures. Based on the results, the features Y1 and Y3 have low impact
for identifying thermal defect with p-value distribution more than
5%. The other features showed high impact for identifying defect
with p-value distribution less than 5%. Hence, null hypothesis is
rejected for Y1 and Y3 and accepted for other 13 features.
Here, the 13 features which give high impact have been pro-
cessed using multivariate test. Table 4 tabulates the p-value
224 A.S.N. Huda, S. Taib / Applied Thermal Engineering 61 (2013) 220e227
Fig. 2. a. Thermogram of phase overheating (Y phase) b. Temperature prole.
distribution of each feature using multivariate test. From the 13
features, 10 optimum features have been chosen using the multi-
variate test as suitable features for defect recognition. The ignored
features which have p-value distribution more than 0.05 are
highlighted in Table 4. The 10 optimum features are X2, Y2, Y4, X11,
X13, X15, X16, X17, Y11 and Y12.
Table 5 presents the suitable features with its discriminant co-
efcients which will be used during the formation of discriminant
function.
The discriminant function Zp value is obtained by adding the
sum of multiplication between discriminant coefcient and domi-
nant features values. A function has a multiple root (i.e. a root with
multiplicity greater than one) in the complex numbers, if and only
if its discriminant is zero [30]. From Table 5, the discriminant
function is derived as follows:
Zp 0:021X2 0:017X11 0:006X13  0:0321X15  0:068X16
0:001X17  1:551
(1)
Based on the Zp value, the class mean called centroid is deter-
mined. Table 6 shows the centroid (mid-point) for each thermal
class. Then, the type and range of each class are found by deter-
mining the separation points between the classes. The separation
point must be established, since the value can be used to identify
the range for each class of the thermal status.
From Table 6, it is found that the centroid of no defect class is
less than defect class. So, the separation point between the no
defect and defect class is found [(1.122 0.642)/2] 0.24. The
no defect class is located at the separation point which is less than
Fig. 3. a. Thermogram of phase overheating (both R and Y) b. Temperature prole.
0.24 and defect class is located at the separation point which is
greater than 0.24.
Apart from determining the suitable features, the discriminant
analysis is also able to predict the linear regression model of each
class taking the account of classication function coefcients and to
classify the thermal class (i.e. determine the number of correctly
classied data, the number of misclassied data and the accuracy
for each class). Table 7 shows the classication function coefcients
of each class for all selected features.
Now the linear predictive model of each class is obtained by
adding the sum of multiplication between classication function
coefcient and dominant features values. The predictive model for
defect class, Dp can be written as
Dp 1:806X2 0:153X11 0:02X13  0:0342X15 1:998X16
 0:034X17  4:02  106 Y2  7:28  1010 Y4  1:09
 105 Y11  3:01  107 Y12  46:08
(2)
and the predictive model for no defect class, Np can be written as
Np 1:768X2 0:123X11 0:01X13  0:908X15 2:118X16
 0:035X17  4:84  106 Y2  6:69  1010 Y4  1
 105 Y11  3:20  107 Y12  43:797
(3)
Finally, the discriminant analysis is used to classify the thermal
status of hotspots into two classes namely no defect and defect.
 A.S.N. Huda, S. Taib / Applied Thermal Engineering 61 (2013) 220e227 225

Table 3
Results for univariate test.

Feature p-value Feature p-value

distribution distribution

X1 0.006 X13 0.00

X2 0.00 X14 0.00
Y1 0.289 X15 0.001
Y2 0.02 X16 0.00
Y3 0.129 X17 0.00
Y4 0.00 Y11 0.002
X11 0.00 Y12 0.00
X12 0.03 e e

Insignicant features are highlighted in bold.

weights between hidden and output layer respectively. The

Fig. 4. Grey image of thermogram (Fig. 1a).
Table 8 shows the classication accuracy for thermal condition
classication.
The overall data are 500; each data has 10 optimum features.
The classication accuracy is 81.1% for defect class (258 thermal
conditions correctly classied out of 318) and 84.6% for no
network is fully connected in this sense that every neuron in each
layer of the network is connected to every other neuron in the
adjacent forward layer by the weighted connection. zi and bj denote
the inputs that are supplied to the input layer and thresholds in the
hidden nodes respectively. ni, m and nh are the number of input,
output and hidden nodes respectively. Then, the predicted output
of the k-th node in the output layer of the MLP network can be
expressed as
!
X
nh X
ni

defect (154 out of 182). The overall accuracy is calculated as

b
y k t w2jk F w1ij zi tbj for 1k<m and 1j<nh (5)
j1 i1
follows:

 In this experiment, ten input nodes and one output node were
Overall accuracy Number of correct classified data  100%
Total number of data being used. The target outputs are classied into two categories i.e.
258154 (4) either defect or no defect. The F(.) is the sigmoid activation
500  100% function in the MLP network. The weights w1ij , w2jk and threshold bj
82:4% are unidentied and should be chosen for reducing the prediction
error. The prediction error is dened as follows:

6. Multilayered perceptron network k t yk t  b

y k t (6)

where yk t and by k t are the actual (i.e. target) and predicted

One of the popular networks for object classication application
output respectively.
is multilayered perceptron neural network. In MLP architecture, the
network is consisted by three layers namely input, hidden and
output layer. Fig. 5 shows a standard MLP architecture [31]. In the 7. Results and discussion
input layer, the numbers of nodes correspond to the number of
input features and the numbers of nodes in the output layer In order to obtain the performance of the MLP network
correspond to the number of target classes. In the present study, regarding electrical equipment thermal defect detection (i.e. no
single hidden layer neural network was used. defect or defect class), the 500 samples from the thermal images
Consider, a standard MLP network with inputs z1, z2, ., zn, were divided into training and testing dataset. For arranging the
predicted outputs y1, y2, ., ym and nh hidden nodes. w1ij and w2jk training and testing dataset, the 5-fold cross validation analysis
denote the connecting weights between input and hidden layer and method was used [32]. In this method, entire data are partitioned
into 5 segments or folds. The summary of data samples in each fold
is tabulated in Table 9. Five iterations of training and testing are
Table 2
List of features used in this study.
performed. In each iteration, a single fold data is selected for testing
and the remaining four folds are used for training the network.
Feature type Feature Abbreviation
Finally, the average value of each fold performance is assumed as
Histogram based 1st order Kurtosis X1
features Standard deviation X2
GLCM features Homogeneity Y1 Table 4
Entropy Y2 Results for multivariate test.
Energy Y3
Contrast Y4 Feature p-value Feature p-value
Mutual feature (1st order) Average intensity X11 distribution distribution
Entropy X12 X1 0.984 X13 0.000
Maximum intensity X13 X2 0.000 X14 0.143
Kurtosis X14 Y2 0.000 X15 0.027
Skewness X15 Y4 0.000 X16 0.000
Standard deviation X16 X11 0.000 X17 0.000
Variance X17 X12 0.264 Y11 0.000
Mutual features (GLCM) Energy Y11 Y12 0.000 e e
Entropy Y12
Insignicant features are highlighted in bold.
226 A.S.N. Huda, S. Taib / Applied Thermal Engineering 61 (2013) 220e227

Table 5
Discriminant function coefcients.
w 1ij
Feature Coefcient Feature Coefcient
z1 w2jk
X2 0.021 Y2 0.000
X11 0.017 Y4 0.000
X13 0.006 Y11 0.000 y1
X15 0.321 Y12 0.000
X16 0.068 Constant 1.551 z2
X17 0.001

Table 6 ym
Centroid for each thermal class.

Class Class centroid zn

Defect 0.642
No defect 1.122
Fig. 5. Standard MLP architecture.

the network performance. The MLP network was trained varying

Table 9
hidden nodes from 1 to 20, keeping epochs constant at 300 and
Summary of data samples.
learning rate at 0.005. For the MLP performance analysis, the study
introduces the accuracy, specicity and sensitivity percentages 5 e fold Number of data samples
analysis. The denition of accuracy, specicity and sensitivity are Defect class No defect class Total
written as follows: 1st fold 64 36 100
2nd fold 64 36 100

pq  3rd fold 64 36 100
Accuracy  100% (7) 4th fold 63 37 100
Number of total data 5th fold 63 37 100
Total 318 182 500
 
p
Sensitivity  100% (8)
py

Table 10
  Training phase performance.
q
Specificity  100% (9)
qx Training data folds Training phase

Sensitivity Specicity Accuracy

where, p total number of defect condition correctly classied
2nd, 3rd, 4th, 5th 90.55 77.39 85.75
1st, 3rd, 4th, 5th 98.88 95.89 97.75
q total number of no defect condition correctly classied 1st, 2nd, 4th, 5th 91.73 82.87 88.50
1st, 2nd, 3rd, 5th 83.52 83.44 83.50
1st, 2nd, 3rd, 4th 84.31 87.50 85.50
Table 7 Average 89.798 85.418 88.20
Classication function coefcients.

Feature Defect No defect

(coefcient) (coefcient)

X2 1.806 1.768
X11 0.153 0.123 Table 11
X13 0.02 0.01 Testing phase performance.
X15 0.342 0.908 Testing data fold Testing phase
X16 1.998 2.118
X17 0.034 0.035 Sensitivity Specicity Accuracy
Y2 4.02E-06 4.84E-06
1st fold 87.50 72.22 82.0
Y4 7.28E-10 6.69E-10
2nd fold 84.37 85.55 83.0
Y11 1.09E-05 1.00E-05
3rd fold 73.43 91.66 80.0
Y12 3.01E-07 3.20E-07
4th fold 85.71 70.27 80.0
(Constant) 46.08 43.767
5th fold 88.88 56.75 77.0
Average 83.98 75.29 80.40

Table 8
Classication using discriminant analysis.

Class Predicted class membership Total Table 12

Performance comparison between DA and MLP classier.
Defect No defect Overall

Count Defect 258 60 412 318 Performances MLP DA

No defect 28 154 182 Sensitivity 83.98 81.10
In percentage Defect 81.1 18.9 82.4 100 Specicity 75.29 84.60
No defect 15.4 84.6 100 Accuracy 80.40 82.40
 A.S.N. Huda, S. Taib / Applied Thermal Engineering 61 (2013) 220e227
y total number of defect condition classied as no defect
condition
x total number of no defect condition classied as defect
condition
Tables 10 and 11 show the training and testing phase performance
of MLP network trained using LM algorithm. In the training phase,
MLP network produced 88.20%, 89.789% and 85.418% of accuracy,
sensitivity and specicity respectively. From all the folds, the results
using second fold as training dataset provided the best performance
regarding accuracy, sensitivity and specicity which are 97.75%,
98.88% and 95.89% respectively.
Similarly in the testing phase, the same network provided ac-
curacy 80.40%, sensitivity 83.98% and specicity 75.29%. Therefore,
it is stated that the MLP network trained using LM algorithm and
also the proposed 10 features have high capacity to detect the
defect in equipment.
Table 12 shows the performance comparison between discrim-
inant analysis classier and multilayered perceptron network. The
DA produced better specicity and accuracy compared to MLP
network, while MLP network achieved higher sensitivity. The value
of sensitivity, specicity and accuracy difference between two
classier are 2.88%, 9.31% and 2% respectively.
8. Conclusion
In this paper, an intelligent predictive diagnosis system for
detecting invisible thermal defects in electrical equipment by
infrared thermography technology is proposed. The diagnosis sys-
tem employed the multilayered perceptron network with Leven-
bergeMarquardt training algorithm to classify the thermal
conditions of equipments into two classes namely no defect and
defect. Total fteen features including rst, second order statistical
features and also mutual features (i.e. the numerical difference
between the value of hot and reference features) were calculated
from sample images. Using discriminant analysis feature selection
method, 10 optimum features were chosen as the inputs of the
neural network. The performance of the neural network was
compared with the discriminant analysis classier. The comparison
results showed that the discriminant analysis classier produced
better performance with 82.40% of accuracy. By developing the
system, the technical department can schedule the recommended
actions. This will play a signicant role for fast and reliable in-
spection to prevent equipment failure potentially.
Acknowledgements
The authors gratefully acknowledge the contribution of Dr.
Kamarul and Mr. Shawal, Universiti Malaysia Pahang for helping in
thermal image acquisition. The work is supported by Fundamental
Research Grant Scheme (203/PELECT/6071211), Universiti Sains
Malaysia.
References
[1] M.R. Clark, D.M. McCann, M.C. Forde, Application of infrared thermography to
the non destructive testing of concrete and masonry bridges, NDT&E Inter-
national 36 (2003) 265e275.
[2] X.P.V. Maldague, et al., Theory and Practice of Infrared Technology for Non-
destructive Testing, Wiley, New York, 2001.
227
[3] Z. Ge, X. Du, L. Yang, Y. Yang, Y. Li, Y. Jin, Performance monitoring of direct air-
cooled power generating unit with infrared thermography, Applied Thermal
Engineering 31 (2011) 418e424.
[4] B.B. Lahiri, S. Bagavathiappan, T. Jayakumar, J. Philip, Medical applications of
infrared thermography: a review, Infrared Physics & Technology 55 (2012)
221e235.
[5] C. Bougriou, R. Bessa, R. Le Gall, J.C. Solecki, Measurement of the temperature
distribution on a circular plane n by infrared thermography technique,
Applied Thermal Engineering 24 (2004) 813e825.
[6] C.A. Balaras, A.A. Argiriou, Infrared thermography for buildings diagnostics,
Energy and Buildings 34 (2002) 171e183.
[7] R. Royo, M.A. Albertos-Arranz, J.A. Crcel-Cubas, J. Paya, Thermographic
study of the preheating plugs in diesel engines, Applied Thermal Engineering
37 (2012) 412e419.
[8] M. Manana, A. Arroyo, A. Ortiz, C.J. Renedo, S. Perez, F. Delgado, Field winding
fault diagnosis in DC motors during manufacturing using thermal monitoring,
Applied Thermal Engineering 31 (2011) 978e983.
[9] D.J. Titman, Applications of thermography in non-destructive testing of
structures, NDT&E International 34 (2001) 149e154.
[10] A.R. Al-Kassir, J. Fernandez, F.V. Tinaut, F. Castro, Thermographic study of
energetic equipments, Applied Thermal Engineering 25 (2005) 183e190.
[11] M.S. Jadin, Soib Taib, Recent progress in diagnosing the reliability of electrical
equipment by using infrared thermography, Infrared Physics & Technology 55
(2012) 236e245.
[12] Improving Electrical System Reliability with Infrared Thermography, www.
infraredelectrical.com/index.htm.
[13] P. Gill, Electrical Power Equipment Maintenance and Testing, Marcel Dekker
Inc., New York, 1998, pp. 385e391.
[14] Y.C. Chou, L. Yao, Automatic diagnosis system of electrical equipment using
infrared thermography, in: International Conference of Soft Computing and
Pattern Recognition, Malacca, December 4e7, 2009, 155e160.
[15] Standard for Infrared Inspection of Electrical Systems & Rotating Equipment,
Infraspection Institute, 2008, 2011. http://www.armco-inspections.com/les/
ir/Electrical%20Rotating%20Std.pdf (accessed 11.06.11).
[17] ASTM, ASTM E 1934: Standard Guide for Examining Electrical and Mechanical
Equipment with Infrared Thermography, ASTM International, West Con-
shohocken, Pennsylvania, 2005.
[16] NFPA, NFPA 70B: Recommended Practice for Electrical Equipment Mainte-
nance, National Fire Protection Association, Quincy, Massachusetts, 2006.
[18] A. Darabi, X. Maldague, Neural network based defect detection and depth
estimation in TNDE, NDT&E International 35 (2002) 165e175.
[19] G. Manduchi, S. Marinetti, P. Bison, E. Grinzato, Application of neural network
computing to thermal non-destructive evaluation, Neural Computing & Ap-
plications 9 (1997) 148e157.
[20] C.A.L. Almeida, A.P. Braga, S. Nascimento, V. Paiva, H.J.A. Martins, R. Torres,
W.M. Caminhas, Intelligent thermographic diagnostic applied to surge ar-
resters: a new approach, IEEE Transactions on Power Delivery 24 (2) (2009).
[21] M.A. Shai, N. Hamzah, Internal fault classication using articial neural
network, in: Proceedings of 4th International Power Engineering and Opti-
mization Conference, Shah Alam, June 23e24, 2010, 352e357.
[22] A. Rahmani, J. Haddadnia, O. Seryasat, Intelligent fault detection of electrical
equipment in ground substations using thermo vision technique, in: Pro-
ceedings of 2nd International Asia Conference on Mechanical and Electronics
Engineering, vol. 2, August 1e3, 2010, pp. 150e154. Kyoto.
[23] M. Smedberg, Thermographic Decision Support e Detecting and Classifying
Faults in Infrared Images. Master thesis, Royal Institute of Technology,
Stockholm, Sweden, 2006.
[24] D. Wretman, Finding Regions of Interest in a Decision Support System for
Analysis of Infrared Images. Master thesis, Royal Institute of Technology,
Stockholm, Sweden, 2006.
[25] The basics of predictive/preventive maintenance. Web: http://www.cedesa.
com.mx/pdf/uke/ukeTi30_serie_nota.pdf (accessed 02.08.11).
[26] Why thermography is good for your business. Web: http://www.newark.com/
pdfs/techarticles/Why_Thermography_is_Good_for_Business.pdf (accessed
01.08.12).
[27] S. Kabir, Imaging-based detection of AAR induced map-crack damage in
concrete structure, NDT&E International 43 (2010) 461e469.
[28] A.S.N. Huda, S. Taib, M.S. Jadin, D. Ishak, A semi-automatic approach for
thermographic inspection of electrical installations within buildings, Energy
and Buildings 55 (2012) 585e591.
[29] D.J. Hand, Discrimination and Classication, John Wiley and Sons, New York,
1981.
[30] Discriminant. Web: http://en.wikipedia.org/wiki/Discriminant (accessed
25.03.12).
[31] S. Haykin, Neural Networks: a Comprehensive Foundation, second ed., Pren-
tice Hall, New Jersey, 1988.
[32] C. Schaffer, Selecting a classication method by cross validation, Machine
Learning 13 (1993) 135e143.