2013 IEEE 1st International Conference on Condition Assessment Techniques in Electrical Systems

Virtual Instrument based Fault Classification in

Power Transformers using Artificial Neural Networks
Santosh Kumar Nanda

S. Gopalakrishna

Dept. of Electrical Engineering

National Institute of Technology Rourkela,
Rourkela, India
santosh24221@gmail.com

Dept. of Electrical Engineering

National Institute of Technology Rourkela,
Rourkela, India
gopal@nitrkl.ac.in

Abstract Inrush currents in power transformers are

detected based on magnitude of second harmonic component. To
avoid the harmful effects of inrush, amorphous core is widely
used in recent days. Transformers with amorphous core cause
low magnitude inrush current and hence the second harmonic of
inrush current is comparable with that during internal faults.
This increases the chances for relay mal operation when classical
techniques of discriminating inrush from other faults are used.
To overcome this, advanced signal processing techniques like
wavelets, S-transform, H-transform and pattern recognition tools
like fuzzy logic, neural network, support vector machine etc. are
being used in recent days. A combination of wavelets and neural
network is found to give satisfactory solution to the above
problem. In this paper, a comparative study using different
mother wavelets along with different activation function is made
to enhance the performance. Virtual instrument is used to
demonstrate the method of fault classification.
Keywords wavelets; neural network; virtual instrument;
faults; power transformer

I.
INTRODUCTION
Power transformers are a class of very expensive and vital
components of electric power systems. The crucial objective to
mitigate the frequency and duration of unwanted outages
related to power transformer puts a high pointed demand on
power transformer protective relays to operate immaculately
and capriciously. One of the major problems associated with
the power transformer relay is the false tripping during
magnetizing inrush current. It is found that the ratio of second
harmonic to fundamental component is more in the case of
inrush than other faults [1-2]. Considering this ratio, a method
known as harmonic restraint differential protection scheme was
developed to judge whether the current is inrush or due to
internal fault. Traditionally, frequency analysis of the currents
was performed using discrete Fourier transform (DFT). But it
is well known that DFT is not accurate if the current is
contaminated by harmonics that are not integer multiples of the
fundamental, especially when the computation window is very
short and DFT only accounts for frequency analysis and
provides no information in the time domain [3]. DFT assumes
the inrush and fault currents as periodic signal, whereas inrush
and fault currents are non-stationary and non-periodic
containing both high frequency oscillations and localized
impulses superimposed on the power frequency and its
harmonics [4].

The first few cycles of inrush current consists of high

magnitude peaks that cause huge mechanical forces in
transformer windings [4]. Also the presence of large quantity
of harmonics in the inrush current can cause damage to power
factor correction capacitor by exciting resonant overvoltage
[5]. Hence steps are taken to mitigate the transformer inrush
current by controlled switching and use of low loss amorphous
core materials in modern power transformer [5]. The inrush
current of these transformers produce low second harmonic
that is comparable with that of fault currents. Hence the
traditional methods are not reliable. New algorithm based on
advanced digital signal processing techniques like discrete
wavelet transform (DWT), wavelet packet transform (WPT), Stransform, H-transform etc. are being implemented for power
transformer protection [4-8]. Different pattern recognition tools
like fuzzy logic, artificial neural network (ANN), principal
component analysis, support vector machine (SVM) etc. are
also being used in recent days as classifiers to classify inrush
current from faults current [9-14].
Instead of using DWT and ANN separately to distinguish
between inrush and fault currents, the algorithm that uses
combination of both is proven to be more efficient than other
methods [1]. However, the procedure used in [1] is complex as
it involves four layered ANN and data extraction based on
Parsevals theorem. In this paper, a simplified method is
proposed by using statistical data extraction and two layered
feed forward ANN. This method requires less number of
iterations for training and found to classify inrush current
accurately.
II.

SIMULATION OF FAULTS AND INRUSH CURRENTS

A. Power System Studied

A simple block diagram model of the power system [1] as
shown in Fig. 1 is considered for study. The model consists of
3 components with ratings as shown in Table I. The power
system model with star-star transformer is simulated in
MATLAB/SIMULINK environment separately for normal
operation, internal faults (I), external faults (E) and inrush.
It is assumed that the power system model is not energized
at time t = 0+. To simulate different cases required for ANN
training, the instance of switching is varied so that initial angle
of voltage changes. To simulate different over excitation cases,
increased load is considered for study. To simulate internal and

CATCON2013

2013 IEEE 1st International Conference on Condition Assessment Techniques in Electrical Systems
external fault patterns, the fault resistance is varied from 0.1
to 100 in multiples of 10. To simulate inrush, secondary is
kept open and switching instances are varied.
Switching
circuits

3 power
transformer
Current
sensors

Transmission
line

R-L load

4
Current (A)

3 source

Table II. The wavelet analysis using db6 as the mother wavelet
is performed up to Detail 5 level for the differential current
signals of all the cases to extract the Detail coefficients. The
analysis is carried out for 20 ms during 0.1 s to 0.12 s.

Current
sensors
Differential
measurement

Display

Phase A
Phase B
Phase C

2
0
-2
-4
0

Transformer
(Y-Y connected)

Transmission Line
R-L Load
(Y connected)
Current transformers
( connected)

Current (A)

Rating
230 kV, 50Hz,
source resistance 0.8929
source inductance 16.58 mH
450 MVA, 230 kV/500 kV, 50 Hz
Winding resistance 0.02 pu
Winding reactance 0.08 pu
(for both Secondary and Primary)
Saturable core with no residual flux
100 Km
Resistance 0.01273 10-3 /m,
Inductance 0.933710-6 H/m
Capacitance 12.0410-6 F/m

0.2

Phase A
Phase B
Phase C

2
0
-2
0

0.05

0.1
Time (s)

0.15

0.2

Fig. 3. Differential inrush current of a three phase power transformer

20
Current (A)

Power Supply
(Y connected)

0.15

DETAILS OF POWER SYSTEM COMPONENTS

3 Component

0.1
Time (s)

Fig. 2. Differential current during normal operation

Fig. 1. Power system model

TABLE I.

0.05

10

450 MW, 530 MVAR

3000/5 (LV side), 800/5 (HV side)

Phase A
Phase B
Phase C

-10
-20
0

Initially, differential current measured during normal

operating condition, inrush and fault conditions are simulated
by closing the switches at 0.1 s, 0.1s and 0.05 s respectively.
The three phase current signals coming from the differential
measurement during normal operating condition and inrush are
shown in Fig. 2 and Fig. 3 respectively. Finite current observed
during normal condition is due to the difference in the current
transformer (CT) ratio.

0.05

0.1
Time (s)

0.15

0.2

Fig. 4. Differential current during internal L-L-L-G fault

Current (A)

10

Although different fault conditions (internal and external

faults) are considered for different patterns, only the severe
fault (L-L-L-G and L-L-L) currents are shown in Fig. 4, Fig. 5
and Fig. 6.

Phase A
Phase B
Phase C

-10
0

III. WAVELET ANALYSIS

The transient currents during fault, inrush and over
excitation conditions are fast decaying, oscillating and consists
of high frequency, so Daubichiess wavelet of level 6 (db6)
suits best for the analysis purpose. The performance of
different mother wavelets is given in Table V.

0.05

0.1
Time (s)

0.15

0.2

Fig. 5. Differential current during internal L-L-L fault

Current (A)

10

As the sampling frequency is 20 kHz, the highest

frequency that the signal could contain will be 10 kHz. This
frequency is observed at the output of high frequency filter
which gives the first Detail. Thus, the band frequencies
between 10 kHz to 5 kHz are captured in Detail-1. Similarly
the band frequencies are captured for other Details as given in

Phase A
Phase B
Phase C

-5
-10
0

0.05

0.1
Time (s)

0.15

0.2

Fig. 6. Differential current during external L-L-L-G fault

CATCON2013

2013 IEEE 1st International Conference on Condition Assessment Techniques in Electrical Systems
2
Current (A)

A. Wavelet Analysis in Inrush Case

Wavelet analysis for inrush case at the instant when
switching takes place at 00 angle of the source voltage is shown
in Fig. 7. Fig. 7(a) represents the original current signal of
phase A. It is seen that the current waveform is distorted in
shape. Abrupt changes in the slope of inrush current are
observed at the edges of the gaps. These sudden changes are
visible as ripples in the wavelet plots shown in Fig. 7(b) - 7(f).
These ripples of significant magnitude are the major
characteristics of inrush current.

1
0
-1
0.1

0.105

0.11
Time (s)

0.115

0.12

(f) Detail coefficient at level 5 decomposition.

Fig. 7. Wavelet analysis of phase-A differential current for inrush

Current (A)

B. Wavelet Analysis in Internal Fault Case

The power system is energized at 0.05s and fault is created
at 0.1s in the secondary side of the power transformer and lasts
for one cycle. Wavelet analysis for internal fault case at the
instant when switching takes place at 00 angle of the source
voltage and fault resistance 100 for the time 0.1s to 0.12 s is
shown in Fig. 8. Fig. 8(a) represents the original current signal
of phase-A during L-G fault. The fault occurs on the high
voltage side of power transformer between phase-A and
ground. A high frequency distortion is observed in the fault
current waveforms as shown in Fig. 8(a). This distortion is the
consequence of the effects of distributed inductance and
capacitance of the transmission line. This distributed
inductance and capacitance lead to significant second harmonic
in the internal fault during the transient period. Hence the
distributed inductance and capacitance poses difficulty in an
accurate discrimination between magnetizing inrush current
and internal fault currents by the conventional protection
method based on DFT.

4
2
0
-2
0.1

0.12

0.14
0.16
Time (s)

0.18

0.2

(a) Original current signal

Current (A)

0.02
0.01
0
-0.01
-0.02
0.1

0.105

0.11
Time (s)

0.115

0.12

(b) Detail coefficient at level 1 decomposition.

Current (A)

0.05

From the Detail coefficients (1 to 5) shown in Fig. 8 (b)-(f),

it is observed that there are several spikes immediately after
fault inception time in L-G fault. One important observation to
distinguish between the inrush and internal fault is that the
ripples during inrush sustain over long time than the case of
internal fault.

-0.05
0.1

0.105

0.11
Time (s)

0.115

0.12

Similarly other internal fault situations like L-L-G fault,

three phase symmetrical line to ground and line to line fault
and external fault are simulated by creating those in the
secondary side of the power transformer for the time same as
that of L-G fault.

(c) Detail coefficient at level 2 decomposition.

Current (A)

0.2
0.1
0

TABLE II.

-0.1

0.4
0.2

IV.

0.105

0.11
Time (s)

0.115

0.12

(d) Detail coefficient at level 3 decomposition.

Current (A)

FREQUENCY BAND OF DIFFERENT DETAIL COEFFICIENTS

Level of Detail coefficient

d1
d2
d3
d4
d5

-0.2
0.1

PERFORMANCE OF ANN AND IMPLEMENTATION IN

LABVIEW
The statistical data obtained from the decomposed signals
of wavelet analysis at level 1 to 5 are used to train and test the
ANN. The ANN consists of three layers with 9 nodes in the
input layer, 9 nodes in the output layer. The nodes of the
hidden layer are varied for optimization. Back propagation
algorithm is used to train the ANN. The activation function

0
-0.2
-0.4
0.1

0.105

0.11
Time (s)

0.115

Frequency band
5 kHz-2.5 kHz
2.5 kHz-1.25 kHz
1.25 kHz-0.625 kHz
0.625 kHz - 0.3125 kHz
0.3125 kHz-0.15625 kHz

0.12

(e) Detail coefficient at level 4 decomposition.

CATCON2013

2013 IEEE 1st International Conference on Condition Assessment Techniques in Electrical Systems
used in the ANN is of sigmoid type. The training and testing
data are statistical features obtained for the three phases. The
mean, standard deviation and norm (root mean square value) of
the decomposed Detail coefficients for the three phases are
used as input data as these statistical variables are distinct and
significant as observed from the analysis and acts like principal
components.

representing the different conditions. The input data are

normalized before training or testing by dividing the maximum
value of data of a row with other data.
Neural network training and testing programs are written
and simulated in Matlab. After performing analysis for
obtaining the optimal values for neural network it is found that,
learning rate of 0.2, the momentum factor of 0.9 and 16 nodes
in hidden layer gives least error and faster convergence. The
ANN is trained with 108 patterns and tested with 72 patterns
for each Detail coefficients separately. Percentage accuracy
with which ANN is able to classify the different cases during
testing is given in Table IV. From the table, Detail-3 (1.25
kHz-0.625 kHz) is found give better classification. This
indicates that for the above configuration of ANN, input and
output patterns formed using Detail-3 is well suited for
discrimination of inrush from faults.

10
Current (A)

5
0
-5
-10
0

0.05

0.1
Time (s)

0.15

0.2

(a) Original fault current

Current (A)

The weights obtained after training with Detail-3 input

and output pattern are used to construct a virtual instrument
ANN using LabVIEW. The fault waveforms and inrush
waveforms are fed as inputs to virtual instrument. To indicate
type of fault, LED indicators are used on the front panel. After
testing with several input waveforms, the virtual instrument is
found to classify the waveforms as expected. Fig. 9 shows the
front panel of virtual instrument with input data of inrush and a
glowing LED corresponding to inrush.

0.5
0
-0.5
0.1

0.105

0.11
Time (s)

0.115

0.12

(b) Detail coefficient at level 1 decomposition.

Current (A)

1
0.5
0
-0.5
0.1

TABLE III.
0.105

0.11
Time (s)

0.115

Name of the
event
Normal
Inrush
L-G fault (I)
L-L-G fault (I)
L-L-L-G fault (I)
L-L fault (I)
L-L-L fault (I)
L-L-L-G fault (E)
Over excitation

0.12

(c) Detail coefficient at level 2 decomposition.

Current (A)

0.5

-0.5
0.1

0.105

0.11
Time (s)

0.115

0.12

(d) Detail coefficient at level 3 decomposition.

TABLE IV.

Current (A)

-1
0.105

0.11
Time (s)

0.115

Normal
Inrush
L-G fault (I)
L-L-G fault (I)
L-L-L-G fault (I)
L-L fault (I)
L-L-L fault (I)
L-L-L-G fault (E)
Over excitation

0.12

(e) Detail coefficient at level 4 decomposition.

0.5
Current (A)

N9
0
0
0
0
0
0
0
0
1

COMPARISON OF TEST RESULTS OF ANN FOR DIFFERENT

Name of the
event

0
-0.5
-1
0.1

Target value of the ith nodes (Ni)

N2 N3
N4 N5 N6 N7 N8
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0

WAVELET DECOMPOSITION LEVEL

-2
0.1

N1
1
0
0
0
0
0
0
0
0

TARGET DATA SET OF THE ANN

0.105

0.11
Time (s)

0.115

% of correct discrimination using different

wavelet decomposition level data
d1
d2
d3
d4
d5
75
100
100
100
50
100
50
100
75
50
100
100
100
100
100
100
100
100
100
100
100
100
100
50
100
100
100
100
50
75
75
50
75
75
75
75
75
100
100
75
100
100
100
100
100

The detail analysis of all the cases are carried out for
different mother wavelets and statistical data obtained after the
analysis are used to train and test the ANN. The performance
of ANN for different mother wavelets is given in Table V.
From Table V it is clear that Daubichies mother wavelet when
used for analysis of the transient signals give high efficiency in
discrimination of simulated cases. The Haar wavelet gives the

0.12

(f) Detail coefficient at level 5 decomposition.

Fig. 8. Wavelet analysis of phase-A differential current for L-G fault

The output data is unity for the condition represented by the

input and zero for others. Table III shows the different nodes

CATCON2013

2013 IEEE 1st International Conference on Condition Assessment Techniques in Electrical Systems
least efficiency whereas Symlet and Coiflet wavelets
performance are better than Haar but not as good as
Daubichies.
TABLE V.

V. CONCLUSIONS
It is possible to distinguish between inrush currents
and fault currents by analyzing the ripple patterns obtained
from wavelet transforms. ANN with sigmoid activation
function is found to classify better with Detail-3 data. To show
the real time application of this method a virtual instrument in
LabVIEW is developed. The performance of trained ANN is
tested successfully for the classification of various cases. The
classification ability of the ANN in combination with advanced
signal processing technique opens the door for smart relays
power transformer protection with very less operating time and
with desirable accuracy.

COMPARISON OF TEST RESULTS OF ANN FOR DIFFERENT

MOTHER WAVELETS

Name of the
event
Normal
Inrush
L-G fault (I)
L-L-G fault (I)
L-L-L-G fault (I)
L-L fault (I)
L-L-L fault (I)
L-L-L-G fault (E)
Over excitation

% of correct discrimination using

different mother wavelets
Haar
Coiflet
Symlet
Daubichies
69.44
86.11
90.27
100
72.22
84.72
91.66
100
73.61
87.5
90.27
100
69.44
84.72
93.05
100
69.44
86.11
88.88
100
72.22
83.33
87.5
100
69.44
87.5
88.88
75
70.83
86.11
86.11
100
72.22
84.72
88.88
100

Fig. 9. Front panel of virtual instrument

[8]

REFERENCES
[1]

[2]

[3]

[4]

[5]

[6]

[7]

P.L.Mao and R.K.Aggarwal. A wavelet transform based decision

making logic method for discrimination between internal faults and
inrush currents in power transformers International journal of
Electric power and Energy systems, Vol.22, no.6, 389-395.Oct 2000.
X.N. Lin, P. Liu, O.P. Malik, Studies for identification of the inrush
based on improved correlation algorithm, IEEE Transactions on
Power Delivery. 17 (4), 901907, 2002.
M. Tripathy, R.P. Maheshwari, H.K. Verma, Advances in
transformer protection: a review, Electric Power Compo. Syst. 33
(11) 12031209 2005.
P. Arboleya, G. Diaz, J.G. Aleixandre, A solution to the dilemma
inrush/fault in transformer relaying using MRA and wavelets,
Electric Power Compo. Syst. 34 (3) 285301, 2006.
S.A. Saleh, M.A. Rahman, Modeling and protection of a three-phase
transformer using wavelet packet transform, IEEE Transactions on
Power Delivery. 20 (2),12731282, 2005.
H.K. Verma, G.C. Kakoti, Algorithm for harmonic restraint
differential relaying based on the discrete Hartley transform, Electric
Power Syst. Res. 18 (2) , 125129, 1990.
W. A. Wilkinson and M. D. Cox, Discrete wavelet analysis of power
system transients, IEEE Trans. Power System, vol. 11, no. 4, pp.
20382044, Nov. 1996.

[9]

[10]

[11]

[12]

[13]

[14]

Z. Moravej, A. A. Abdoos & M. Sanaye-Pasand Power Transformer

Protection Using Improved S-Transform, Electric Power
Components and Systems, 39:11, 1151-1174, 2011.
M.C. Shin, C.W. Park, J.H. Kim, Fuzzy logic based relaying for
large power transformer protection, IEEE Transactions on Power
Delivery. 18 (3),718724, 2003.
E. Vazquez, I.I. Mijares, O.L. Chacon, A. Conde, Transformer
differential protection using principal component analysis, IEEE
Transactions on Power Delivey.23 (1), 6772, 2008.
L. Yongli, H. Jiali, D. Yuqian, Application of neural network to
microprocessor based transformer protective relaying, IEEE Int.
Conf. on Energy Management and Power Delivery, vol. 2, 2123, pp.
680683, November 1995.
P. Bastard, M. Meunier, H. Regal, Neural network based algorithm
for power transformer differential relays, IEE Proceedings
Generation. Transmission and Distribution. 142 (4), 386392, 1995.
M.R. Zaman, M.A. Rahman, Experimental testing of the artificial
neural network based protection of power transformers, IEEE
Transactions on Power Delivery. 13 (2), 510517, 1998.
V. Thamarai Selvi and V. Malathi, A classification approach using
SVM to detect magnetic inrush in power transformers, International
Journal of Innovation and Applied Studies, Vol. 3, pp. 693-700, July
2013.

CATCON2013