Detection of Incipient Faults in Power Transformer

Using Wavelet Transform: Lessons Learned

Essam Abu El-Zahab (1), Dalal Helmi(2), Maha Salah (2) , Ahmad Alokabi (3)*

(1)
Faculty of Engineering, Cairo University
(2)
Egyptian Electricity Transmission Company (EETC)
(3)
Cairo Electricity Production Company (CEPC)
Egypt

These include the following, but not limited to:

• Most of the available commercial programs treat
SUMMARY
As initial inter-turn insulation failure in Power the transformer as one box without any
Transformers does not draw sufficient current, capability of simulating transformer internal
traditional protection is unable to detect such faults winding faults. The only way to simulate such
that develop slowly, often in form of gradual faults is to treat the transformer as a mutually
deterioration of insulation. As a result, the coupled R-L circuit which requires complete
transformer will only be disconnected from the line
• In some cases more sophisticated test data is
transformer data.
when the fault extends to such a degree to affect
considerable portion of the winding and produces required particularly homopolar or zero sequence
high current. Although differential protection measurements for core type transformers having
provides the best protection for phase and ground only star earthed winding This test data is not
faults, it does have its problems. One of these normally performed by manufacturers as it
problems is its inability to detect incipient faults requires special tests. Without such data the
because there is no sufficient current drawn from results obtained are inaccurate in case of star
the line to operate such relays. These faults can
• The available commercial software does not have
earthed winding mainly.
only be detected by Buchhloz relay but this relay is
slow and detects only faults below oil level. Using the capability of simulating actual turn-to-turn
these protection devices, the transformer will only faults, only disc-to-disc faults (each disc contains
be disconnected as the fault being developed to a multiple of turns) as there is no capability of
severe one with high amount of fault current. getting external taps available from separate
turns.
As field recordings of transformer abnormal
The results obtained from simulation are then used
conditions, especially for internal faults, are
as an input data to Wavelet Technique to detect the
seldom available, the information needed for
transformer incipient fault. The proposed technique
investigation of protective systems may be
also succeeded in discriminating between internal
achieved by accurate simulation work. This paper
incipient faults and other conditions (load
introduces an accurate model in simulating
switching, transformer inrush, short circuits).The
transformer incipient faults using ATP (Alternative
proposed Wavelet Technique could be used as
Transient Program) that overcomes many
compliment to differential protection to detect
obstacles; it also provides the lessons learned
transformer incipient faults before being developed
during the simulation that should be taken into
into more severe one.
consideration in new S/W packages.
KEYWORD Power transformers - incipient faults-
modelling - differential protection - wavelet packet

* [email protected]
1. INTRODUCTION A protection technique was proposed to detect the
incipient faults in the transformer, based on
The worldwide international changes in the analysis of the differential current of the simulated
electricity market put the electric utilities under faulty transformer by the developed WPT
severe stress to reduce the running (operating) algorithm.
cost, enhance the availability of the power system
and prevent forced outage while work with the
existing equipment harder and longer. Power 2. COMPUTER MODELING PRINCIPLES
transformers are considered as the most expensive
infrastructure in the substations, and hence the The basic model used in this work is the one
outage and/or damage of such critical equipment supplied by the BCTRAN routine of the ATP
is costly to any electric utility. software (Alternative Transient Program). The
transformer is handled as mutually coupled R, L
A high voltage transformer connected to an branches as shown in figure (1). Based on data of
overhead transmission system is very likely to be excitation and short circuit tests, this routine
subjected to steep fronted impulse voltages. A line computes two matrices [R] and [L] modeling the
surge, which may be of several times the rated transformer. In the case of a three-phase
system voltage, will concentrate on the end turns transformer with two windings, these matrices are
of the winding because of the high equivalent of order 6 as in equations (1-a) and (1b), where Ri
frequency of the surge front. The interturn and Li are the resistance and the self inductance of
insulation of the end turns is reinforced, but coil i, and Mij is the mutual inductance between
cannot be increased in proportion to the insulation the coils i and j [3, 4] .
to earth, which is relatively great. The risk of a
partial winding flashover compared with that of a The principle used to model a fault between a coil
breakdown to earth is therefore comparatively turn and the earth or between any two turns is to
high. A study of the records of modern divide the faulty coil into sub coils. In case of
transformer breakdowns, which have occurred turn-to-earth fault, the faulty coil will be divided
over a period of years, shows very conclusively into two coils (a, b), then, the transformer can be
that between 70 and 80% of the number of failures described with two 7x7 matrices [R], [L]. In case
are finally traced to short-circuit between turns of turn-to-turn fault, the faulty coil will be divided
[1]. The subsequent progress of the fault, if not into three coils (a, b, c), consequently, the
detected in the earliest stage, may well destroy the transformer is divided into two 8x8 matrices are
evidence of the true cause. A short circuit of a few required. This principle is used such that the
turns of the winding will give rise to a heavy fault number and location of the shorted turns can be
current in the short-circuited loop, but the terminal variable. The divided coil, no. 3, for the two faulty
currents will be very small, because of the high cases is shown in Figures 2(a, b). In the following
ratio of transformation between the whole two sections, the analysis will be described in
winding and the short-circuited turns [2]. details [5, 6].

This paper presents a method for modeling

internal faults in a three-phase oil immersed
power transformer. The approach taken in this
paper utilizes the electrical indicators for internal
winding fault, and is mainly based on the
BCTRAN procedure of ATP program. The
terminal current behavior is discussed when the
transformer is in several typical faulty situations
(turn-to-earth and turn to turn) on the primary and
secondary sides of the transformer, and for
different number of shorted turns.

An algorithm based on the Wavelet Packet

Transform (WPT) was developed to measure the Figure 1:
RMS value of the harmonics content in a signal The transformer is handled as mutually
with a high accuracy. coupled R, L branches
 where: na, nb number of turns in the sub-coils a
(1-a) and b respectively, na + nb = n3
[R]=
The shadow area in the matrix [L] represents the
unknown values which are the self and mutual
inductances between the sub-coils a, and b.
(1-b) Reference [5] introduces the three rules to
[L]= calculate these values, which are consistency,
leakage and proportionality.

1. Consistency : [La + 2Mab + Lb ] = L (4)

2. Leakage: δab=[1 – (Mab2) / (LaLb)] (5)
3. Proportionality : [La/ Lb ]=[na / nb]2 (6)

The leakage factor δab can be calculated using the

core and winding geometrical data. Alternatively,
as an approximation, it can be treated as
parameter. In our approach and as a good
approximation, it considered equal zero.

The previous three equations (4, 5, and 6) are for

three unknown quantities, the resolution leads to
(a) turn to ground fault (b) turn to turn fault the following relations:
Figure 2: The divided faulty coil
La = (na /n1)2*L1 (7)
Lb = (nb /n1)2*L1 (8)
2.1 Winding-to-Earth Fault
Mab = (na /n1)* (nb /n1)* L1 (9)
Assume the winding number (3) suffers an
internal winding-to-ground fault (Fig.2a). A single To determine the mutual inductances between the
inner node is thus created, which divides the coil coil a and any other coil i except for b (as well as
(3) into two sub-coils, say (a) and (b), and the those between the coil b and any other coil i
matrices [R] and [L] becomes the following new except for a. General speaking, the consistency
(7x7) matrices: and The proportionality rules lead to:

1. Consistency :
M3i = Mai + Mbi (10)
2. Proportionality :
[R]= M ai / M bi = na / nb = k (11)
(2-a)

From equations (10 and 11) we can get M ai

and M bi :

Mai = [K / (1+K)] M3i (12)

[L]= Mbi = [1/ (1+K)] M3i (13)
(2-b)
2.2 Turn – to – Turn Fault

Let us consider the turn-to-turn fault shown in

The matrix [R] can be easily determined using the
figure (2-b). As in part (2.1), the matrix [R] (8x8),
following equations:
can be easily determined with the help of the
following relations:
Ra = (na /n3)*R3 , Rb = (nb /n3)*R3 (3)
 Ra = (na/n1) *R1 , transform) would seemingly be suitable for
Rb = (nb/n1) * R1 , transient analysis. That is, one would make use of
and Rc = (nc/n1) * R1 (14) a transient to analyze a transient, as opposed to
using a stationary function (e.g., a sine or a
Matrix [L] (8x8), can be computed in a similar cosine) for such analysis [8].
way depending on the three rules (the
consistency, the leakage factor and the
proportionality). 3.1 Wavelet Packet Transform Filtering
Process

In wavelet analysis, we often speak of

approximations and details. The approximations
are the high-scale, low-frequency components of
[R]= the signal. The details are the low-scale, high-
frequency components [7] In the WPT, a signal is
(15-a) split into an approximation and a detail. The
approximation and the detail are then themselves
split into a second-level approximations and
details, and the process is repeated. WPT provides
uniform frequency bands.

In Figure (3), we can see the wavelet packet

transform process, where h (n), h1(n) are the low
and high pass filter respectively. If the original
[L]= signal frequency band is (0 – f), it will be divided
(15-b)
in the first resolution level to low frequency band
(0 – f/2) and high frequency band (f/2 – f), then
each frequency band will be divided into two
other smaller bands. P02 is the wavelet
coefficients at level 2 and node 0 ,and so on [9].

3.2 Harmonics RMS Measurement Using WPT

3. WAVELET TRANSFORM
The RMS value of the current (i RMS) is [10, 11]:
Wavelet Transform is a mathematical tool that
cuts up data or functions into different frequency
components, and then studies each component
with a resolution matched to its scale. For
example, in signal analysis, the wavelet transform
allows us to view a time history in terms of its
frequency components, which means it maps a
one-dimensional signal of time, f (t), into a two
dimensional signal function of time and (15)
frequency. The wavelet transform represents the
signal as a sum of wavelets at different locations
(positions) and scales (frequency bands) [7].

A wavelet is defined as a waveform of effectively Where,

limited duration that has an average value of zero
(i.e., no dc component) and decays quickly to zero 2N the number of the original signal
at both ends and thus are more suitable for short samples
duration disturbances. Thus, by definition, a k the wavelet coefficient number
wavelet is a transient signal, and methods that Iij the rms value of the current for the
make use of wavelet functions (e.g., the wavelet frequency band at node i and level j
 Figure (3): Two Levels wavelet packet transformation

Let the original signal contains 2N sampling This will provide sharper cut-off frequency to the
points. The wavelet coefficients at the level (j) and selected mother wavelet and
node (i) is (dij).The maximum resolution level can reduces the amount of leakage energy to the
be achieved is N. adjacent resolution levels [12, 13]

Where
i = 0 ,1 , 2 , 3,……… 2 j -1 ,
Capture Signal
j = 1 , 2 , 3 , 4,……….N
Samples
Figure (4) shows a flow chart for the proposed
algorithm.
Apply Wavelet
Packet Analysis
From the previous analysis, it's clear by choosing
the appropriate sampling frequency and the WPT
resolution level; we can allocate the required Get the Wavelet
harmonics in different frequency bands (nodes). Coefficients

RMS of each frequency band (node i) can be Apply the Proposed

estimated as follows: Algorithm

I = ∑ (d
Get the RMS Values of the
i 1 i 2 Harmonics
) (16)
k =0
j
2N j

Figure (4): Flow chart for the use of WPT for

Different disturbances can be modeled and Harmonic Analysis
presented using different wavelet Ψ (t) and scaling
Φ (t) functions. The accuracy of this presentation As mentioned before, the selection of the mother
depends on the smoothness of the selected mother wavelet affects on the accuracy of the results, so
wavelet function. As the number of vanishing after testing many mother wavelets we found that
moments of the selected wavelet function (Daubechie 40) is the most appropriate one to get
increases, more smoothness can be achieved and the most accurate results, the MATLAB software
more accurate representation of the distorted and WAVELAB toolbox were used to achieve the
signal is obtained. required analysis.
The signals we applied the analysis on are 50 Hz 4.2 Terminal Currents Behavior of Different
signals and have sampling frequency (6.4 kHz, Simulated Incipient Fault Types
128 sample/cycle), i.e. (2N=128 =27). The
maximum resolution level can be achieved is Three simulated incipient fault cases will be
(N=7), at this resolution level we will have 128 presented and studied here; turn-to-turn fault on
nodes (frequency bands). the primary side, turn-to-turn fault on the
From the Nequist theory, the highest frequency secondary side, and turn-to-earth fault on the
that can be accurately represented is less than the secondary side near the neutral point. Fault on the
half of the sampling frequency. Therefore, in our primary side was done on the winding has voltage
study the highest frequency that can be Vac, and the faults on the secondary side was
represented is less than (3.2 KHz). At resolution done on the winding has voltage Va. The shorted
level (7), the number of the frequency bands is turns percentage to the whole winding number is
128, so the frequency band width = 3200 Hz / 128 0.01%, and the transformer loading percentage is
= 25 Hz, to make this decomposition compatible 25% with 0.8 PF load.
with the harmonics frequencies, we need to
calculate the rms values of each two adjacent Figures (6, 7, and 8) show the phase (A) terminal
frequency bands then get their square root currents behavior of the three cases. It can be
summation [14]. For example, the 4th band has noticed that the results of the three cases are the
frequency range (75 – 100) Hz and the 5th band same. At the inception fault instant (t=0.04sec),
has frequency range (100 – 125) Hz, to get the the primary current increased slightly with no
rms value of the 2nd harmonic (100 Hz), the rms change in the secondary current. The differential
value of the 4th and 5th bands will be calculated current between the primary and secondary
separately then the 2nd harmonic rms can be currents (referred to the primary side) increased
estimated as follow: with a small magnitude, which cannot be detected

RMS2 ndHarm = ( RMS4thBand ) 2 + ( RMS5thBand ) 2

by the traditional differential relay.

(17)
and so on for the other harmonics.

4. SIMULATION WORK

4.1 System under Study

Simulation work was carried out on three phase

72/13.8 kV, 66 MVA, 50 Hz power transformer.
Incipient turn-to-turn and turn-to-ground faults
were carried out on both primary and secondary Figure (6): Primary Currents of the
windings. Different values of the ratio between the three simulated Incipient Fault cases
shorted turns to the complete winding were taken.
Incipient faults simulations were done under
different loading conditions (different loading
percentages, different load power factors). Figure
(5) shows the circuit of the system under the study
in the ATP program environment.

Figure (5): The circuit of the system under the study Figure (7): Secondary Currents of the
in the ATP program environment three simulated Incipient Fault cases
 Figure (8): Differential currents of the
three simulated incipient fault cases

4.3 Harmonic Analysis of the Differential Figure (9): The harmonic analysis results of the
Currents of the Simulated Incipient Fault differential currents of the simulated incipient faults
Cases

Harmonic analysis was done for the differential

currents of the previous simulated cases using the
proposed algorithm based on the WPT. The
results are shown in figure (9).
The harmonic analysis showed that DC
component, 2nd and 3rd harmonics increased at
the fault instant (0.04 sec) to values greater than
60% then decreased again to the normal operating
values in one cycle (0.02 sec).

4.4 Load Switching Simulation

In the previous section, harmonic analysis was Figure (10): The harmonic analysis results of the
done for the differential current in cases of differential currents in case of switching 25% load
transformer incipient faults and it was found that
the change in the harmonics value has some
characteristics, which could be used to detect the
incipient faults. To make sure from these results
we need to make analysis of other transient
conditions, which may occur to the transformer,
from these transient conditions is load switching.

Harmonic analysis of the differential current in

cases of load switching of 25 and 100 percentage
load of the transformer rating with 0.8 PF is
shown in figures (11) and (10).

The analysis showed that the DC component, 2nd

and 3rd harmonics rise at the load switching
moment and decrease again to the normal values,
but the rise value is very small compared with the Figure (11): The harmonic analysis results of the
value in the cases of the incipient faults. differential currents in case of switching 100% load
5. Summery of the Transformer Simulation estimated and observed if any considerable change
Results is going on. If the change between a sample (k)
and 2 samples before (k-2) is larger than 3%, a
From the previous simulations results of the considerable change is detected (this value was
incipient faults (turn-to-turn and turn-to-ground specified after studying the change characteristics
faults) and the transformer load switching, the in harmonics in many different simulation cases).
results can be summarized in table (1). It shows After the change is detected, its period (T1) and
for each case the characteristics of the harmonics the period of the maximum value greater than
content and their change behavior. 60% (T2) are recorded.

From this summery it can be shown that the If T1 equals one cycle time and T2 equal zero, a
incipient internal faults can be detected and load switching condition is detected. If T1 equals
discriminated by observing and monitoring the one cycle time, T2 does not equal zero and the
characteristics of the harmonics content and its primary current in time T1 changed with no
change behavior. change in the secondary current, an incipient
internal fault is detected.

6. THE PROPOSED TECHNIQUE

7. CONCLUSIONS
After the study of the incipient internal faults
characteristics, we can design a technique, which This paper presents a technique dealing with one
can be used to detect the incipient internal faults of the most difficult problems facing protection
and discriminate it from the other transient engineers which is transformer incipient fault
conditions. detection.

Figure (12) shows a flow chart for the designed As initial inter-turn insulation failure does not
technique. The results shown in table (I) were draw sufficient current from the line, traditional
used to make this technique. The differential differential relays are unable to detect such faults
currents of the three phases are computed as a result the transformer will only be
continuously. Harmonic analysis is done by the disconnected from the line when the fault has
WPT algorithm, the DC component, 2nd and 3rd extended to such a degree to embrace considerable
harmonics are portion of the affected winding.
Table (I)

Dominant Variable (Discriminator)

Case Type
DC Component% 2nd Harmonic% 3rd Harmonic%
Incipient Fault:
- Turn to Turn Fault Max > 60 Max > 60 Max > 60
- Turn to Ground Fault
Tmax<1/2cycle Tmax<1/2cycle Tmax<1/2cycle

Comment - The Transient Change in the Harmonic values starts and ends
in one cycle.
- The Max value remains for less than 1/2 cycle
- The primary current increase with no change in the secondary
current.
Load Switching
Max <<60 Max <<60 Max << 60
Comment - The Transient Change in the Harmonic values starts and ends
in one cycle.
- The Transient Change is accompanied by a large change in
the secondary current.
 Start
T1=0, T2=0

Input: Read new data

Primary &Secondary
Currents Samples

Calculate the Differential Currents of

the 3 Phases
IDA, IDB, IDC

Using The Wavelet Algorithm

Calculate the Harmonic Percentage
in IDA, IDB and IDC
Idc , I2nd , I3rd %

NO | Idc(k) - Idc(k-2) | ≥ 3%
and | I2nd (k) -I2nd(k-2) | ≥ 3%
and | I3rd (k) -I3rd (k-2) | ≥
YES
T1 =0
YES
NO
The End of the Transient Period Transient Change Detected
Transient Period = T1 Compute the Transient Period
T1=T1 + sample time

NO
T1 =1 Cycle?
Idc(k) > 60% NO
YES Or I2nd (k) > 60%
Or I3rd (k) > 60%
T2=zero?
YES
NO YES
Compute the Transient Peak Period
NO T2=T2 + sample time
In T1 the Prim Current
changed with no
change in the Sec

YES
Detect an Incipient Fault Detect a Load Switching
Condition Condition

Fig (12): The flow chart of the proposed discriminator technique

 In this work, a transformer with a real data was [4] ATP Rule Book, part (XIX-C), BCTRAN to
accurately modeled using BCTRAN subroutine of derive [A], or[R], [ωL] of multi-phase
ATP program. In this model the transformer is transformer.
treated as mutually coupled R-L circuit, from the [5] Patrick Bastard, Pierre Bertrand, Michel
matrices of resistance and inductances computed Meunier, “A Transformer Model for Winding
by BCTRAN, a new matrices were computed Fault Studies”, IEEE Transactions on Power
which allows the simulation of any kind of Delivery, Vol. 9, No.2 April 1994, pp. 690-
internal fault. Hence, using the proposed model, 699.
turn-to-turn and turn-to-earth faults were
simulated with different fault and operating [6] Mladen Kezunovic, Bogdan Kasztenny, Zijad
conditions. This power transformer modeling Galijasevic, Donna Williams, "A New Add-
technique can be used in representing power ON For Modeling Internal Faults In Power
transformer transients, which is very important in Transformer", American Power Conference,
testing and evaluating protective relays. It allows Chicago, Illinois, April, 2000.
the generation of some difficult fault cases for [7] Wavelet Matlab Toolbox User Guide, Version
power transformer protection evaluation, which 3
may not be possible using other known physical [8] Anthony Wayne Galli, Oli Moller Nielsen,
means. "Wavelet Analysis for Power System
Transients", IEE Computer Application in
Wavelet packet (WPT) is introduced showing that Power, January 1999
WPT provides decomposition with uniform [9] F. Vatansever, A. Ozdemir, "A New
frequency bands. An accurate algorithm was Approach for Measuring RMS Value and
developed to estimate the RMS value of the Phase Angle of Fundamental Harmonic Based
harmonics content in a signal using the wavelet on Wavelet Packet Transform", Science
coefficients of the WPT analysis. Direct, Electric Power System Research 78
(2008) 74-79
The developed WPT algorithm was used to [10] Effrina Yanti Hamid, Zen-Ichiro Kawasaki,
analysis the differential current of all studied "Wavelet Packet Transform for RMS Values
cases. The harmonics content changing and Power Measurements", IEEE Power
characteristics were studied in each case. It was Engineering Review, September 2001
found that this change has characteristic signature, [11] Ahmed Mohamed Gaouda, "Wavelet-
which can be used to detect the incipient faults Automated Recognition System for Power
and discriminate between them and other Quality Monitoring", A thesis presented to the
conditions. University of Waterloo in fulfillment of the
thesis requirement for the degree of Doctor of
The proposed technique could be used as a Philosophy in Electrical and Cornputer
complement to traditional transformer differential Engineering Waterloo, Ontario, Canada,
protection to detect incipient fault at an early stage 2001.
(increase its sensitivity). It could be applied to any [12] Chethan Parameswariah, Mickey Cox,
power transformer without hardware, addition "Frequency Characteristics of Wavelets",
only software add-in. IEEE Transactions on Power Delivery,
VOL.17, no.3, July 2002
[13] Ajay Karthik Hari, "A Novel Index To
BIBLIOGRAPHY Analyze Power Quality Phenomena Using
Discrete Wavelet Packet Transform", Thesis
[1] William H. Bartley, "Investigating Presented to the Graduate School of the
Transformer Failures," The Hartford Steam University of Florida on Partial Fulfillment of
Boiler inspection & Insurance Co, ACTI Inc. the requirements for the Degree of Master of
Second Annual Conference, New Diagnostic Science, 2003
Concepts for Better Asset Management, [14] Julio Barros, Ramón I. Diego, "Application of
November 2003. the Wavelet-Packet Transform to the
[2] AREVA Transmission Distribution - Network Estimation of Harmonic Groups in Current
Protection & Automation Guide,1st Edition 2002 and Voltage Waveforms", IEEE Transactions
[3] EMTP Theory Book, 1st Edition 1987 on Power Delivery, VOL. 21, No. 1, January
2006