IEEE Transactions on Dielectrics and Electrical Insulation Vol. 15, No.

5; October 2008 1297

Rough-granular Approach for Impulse Fault Classification

of Transformers using Cross-wavelet Transform
D. Dey, B. Chatterjee, S. Chakravorti and S. Munshi
Jadavpur University
Department of Electrical Engineering
Kolkata 700032, India

ABSTRACT
A novel approach based on information granulation using Rough sets for impulse fault
identification of transformers has been proposed. It is found that the location and type
of fault within a transformer winding can be classified efficiently by the features
extracted from cross-wavelet spectra of current waveforms, obtained from impulse
test. Results show that the proposed methodology can localize the fault within 5% of
the winding length with a high degree of accuracy. The basic concepts of feature
extraction using cross-wavelet transform and the method of classification of those
features by rough-granular method are also explained.
Index Terms — Cross-wavelet transform, cross-wavelet spectrum, impulse fault
identification, information granulation, rough set.

1 INTRODUCTION Noise is an important issue for all types of measurements.

As the impulse testing of transformer is usually performed
IN case of failure during impulse testing of transformers within the shielded laboratory, severe noise contamination has
the oscillographic traces of applied voltage and corresponding
not been experienced by the authors. However, in real-life
current waveforms are investigated for judging the condition
experimentation it is commonly observed that the waveforms
of the transformers, which requires significant human
are superposed with very low level random noise. Cross-
expertise and knowledge. Impulse test is an acceptance test for
wavelet transform usually takes care of this type of random
power transformers. Guidelines for performing these tests are
un-correlated noise. As the cross-wavelet transform gives
described in standards [1]. Different computer-aided
correlation between two waves in time-frequency domain, low
methodologies, such as, Artificial Neural Network (ANN) [2],
level random noise has minimal effect on it. So, in this present
Fuzzy systems [3] and wavelet based analysis [4] have already
work no denoising scheme is applied for prior processing of
been reported for the classification and localization of fault
the waveforms. However, if the noise level is found to be high
within the transformer winding. In the present work an
then any suitable denoising technique may be applied before
information granulation based approach using Rough set
processing the recorded waveforms. Authors have already
analysis is proposed. In the proposed scheme transformer tank
reported different denoising schemes employing hybrid-
currents due to applied lightning impulse are processed by
filtering in real-time [5] as well as wavelet based
cross-wavelet transform and some significant features are
decomposition [4].
extracted from the cross-wavelet spectra. A Rough-granular
Rough set based analysis is suitable where knowledge is
classifier is used as a new paradigm to determine the fault
imprecise or the information system contains superfluous
location within the transformer winding and also to identify
information. The data about a system can be reduced using
the fault type.
Rough sets keeping all the information or features of the
It is worth mentioning here that during conventional
system intact [6]. Performance shows that the proposed
impulse testing, current waveforms for both standard full test
method can identify and localize the fault accurately within
voltage and reduced voltage are compared. In the present
5% of the winding length with a high degree of accuracy.
approach the nature of time variation of the current
waveforms for different fault conditions are investigated in
time-frequency domain. These waveforms are normalized
2 SIMULATION OF SERIES AND SHUNT
before processing. So, absolute magnitude of the applied FAULT IN ANALOG AND DIGITAL MODELS
voltage as well as that of the corresponding current waveform To analyze the performance of the proposed scheme,
is not an important parameter in this case. impulse faults are simulated on the analog model of a 3-MVA,
33/11 kV, three phase, 50 Hz, ON (i.e. oil natural cooling),
Manuscript received on 21 January 2008, in final form 18 March 2008. Dy11 (i.e. vector group delta-star-11) transformer. Results of

1070-9878/08/$25.00 © 2008 IEEE

1298 D. Dey et al.: Rough-granular Approach for Impulse Fault Classification of Transformers using Cross-wavelet Transform

analysis on Electromagnetic Transient Program (EMTP) based present problem, it is found that the performance of the
digital model are also shown in the paper. scheme is better with Morlet wavelet function. kψ is a
Details of the analog model of the transformer, used in this 2
work can be found in [4] and [7]. The model is having 89 discs. constant, defined as, k =
+∞
Ψ (ω )
ψ ∫dω < ∞ . The cross-wavelet
In this work the entire winding has been divided into 22 sections, −∞
ω
namely L1, L2,…, L22 each involving approximately 5% of the spectrum gives a measure of correlation between two
total length of the winding. Each section consists of 4 discs waveforms in time-frequency domain. So, cross-wavelet
sequentially. The terminal section L22 is having the last five spectrum shows regions in time-frequency space where two
discs. In the present case of study, two types of faults, series and waveforms are having high common power. The Cross-
shunt, are simulated. Series-fault (i.e. insulation failure between
the discs or between turns) is simulated by short-circuiting the
wavelet spectrum the magnitude of W xy and the phase angle,
ℑ{W xy } are plotted.
corresponding disc. On the other hand, for shunt-fault (i.e. failure φ = tan −1
of insulation between winding and earthed components) the disc ℜ{W xy }
is connected to ground. The photograph of analog model is
shown in Figure 1. Corresponding EMTP based digital model is
also shown in Figure 2. Implementation of the series and shunt
faults are also shown in the exploded views of the Figure 2.
The types of faults in transformer windings are
topologically of two types, namely, series and shunt faults.
These two types mostly cover the significant faults that occur
in transformers. These may characteristically be classified as
static or complex. The method of simulation and analysis of
partial or complex faults are different altogether from static
faults. So, in the present work static series and shunt faults are
analyzed. The simulation and analysis of complex faults lie
within the scope of future work.
Even if the transformer winding is symmetric in geometry,
while applying the impulse voltage to the line end of the
winding, the other end is earthed. The impulse voltage
distribution along the transformer winding is non-linear.
Therefore, from earthed end or from line end of the winding
no two points are symmetric from electric circuit point of
view under the application of impulse voltage.
Example of recorded currents waveforms obtained from the Figure 1. Photograph of the analog model of the transformer.
analog model, by tank-current method for no-fault condition
and for series fault at a certain disc position is shown in Figure
3. These waveforms have been normalized before processing.

3 FEATURE EXTRACTION FROM CROSS-

WAVELET SPECTRUM
Cross-wavelet transform may be considered as an extension
of wavelet based analysis, which has not been explored
extensively in fault analysis in the field of power engineering.
A very brief overview of cross-wavelet transform is given in
this section. However, the detailed mathematical background
can be found in [8-9].
The cross-wavelet transform of two signals, x(t) and y(t) is
defined as:
+∞+∞
1 * a b − τ dadb
W xy (s,τ ) = ∫ ∫W
x
(a, b)W y ( , ) 2 .
kψ −∞−∞
s s a
Here, W ( s, τ ) and W y ( s,τ ) are the wavelet transform of
x

x(t) and y(t) respectively with respect to a mother wavelet

ψ(t). The choice of mother wavelet depends on the nature of
the waveform being processed. In this case Morlet mother
wavelet is used. Comparing the results of other mother Figure 2. EMTP model of the transformer along with implementation of series
wavelets like Paul and Gaussian derivative functions for the and shunt fault.
IEEE Transactions on Dielectrics and Electrical Insulation Vol. 15, No. 5; October 2008 1299

The cross-wavelet spectra of different fault current

∑∑τ φ (s,τ ) ∑∑τ φ (s,τ )
waveforms with the impulse current at no-fault condition are 10. F10= s ; 11. F11= s

analyzed. Typical no-fault current waveform along with a φ ( s,τ ) peak ( s max − s min )(τ max − τ min )
fault-current waveform is shown in Figure 3. The
corresponding cross-wavelet spectrum is given in Figure 4. ∑∑τ ( F 11 − φ ( s, τ ) ) 2
12. F12= s
The ‘U’ shaped and black colored line shows the “cone of
( s max − s min )(τ max − τ min )
influence” (COI). COI indicates the region where edge effects
due to zero padding are significant. Similar to spectral
analysis, errors will occur at the edges (i.e. beginning or end)
of the waveform in the case of cross-wavelet because of the
finite length of time series. Padding with zeros introduces
discontinuities and the amplitude of the coefficients near the
edge decreases at larger scales as more zeroes are added. The
cone of influence is the region of the cross-wavelet spectrum
in which these effects become important. This is defined as
the e-folding time for the autocorrelation of wavelet power at
each scale. That means the power for a discontinuity drops by
a factor of e-2 (where, e= 2.7182) and the edge effects are
negligible beyond this point.
In Figure 4, W xy values at different ‘time’ and ‘scale’ are
plotted. The x-axis is considered as ‘time’ axis and y-axis
shows the ‘scale’, which is related to the inverse of frequency.
The color of the figure at a point shows the value of W xy at
Figure 3. Tank current waveforms for no-fault condition and series fault at
that time-frequency space. The color-bar given in the right disc no. 30.
side of Figure 4 indicates the value corresponding to a color.
Higher the value higher the common power at that time-
frequency point.
Black arrows show the phase angle. Arrows shown in
Figure 4, pointing towards right indicate “in-phase” (i.e. phase
difference in zero) and arrows pointing left indicate “anti-
phase” (i.e. phase difference is 180 degrees) conditions. For
classification and identification of fault some features are
extracted from W xy . The features are described below:

1. F1=
∑∑τ sτ W (s,τ )
s
xy

; 2. F2=
∑∑τ s τ 2 2
W xy ( s,τ )
s

∑∑τ W (s,τ )
s
xy
∑∑τ W
s
xy
( s,τ )

∑∑τ W xy
( s,τ ) ∑∑τ W xy
( s,τ )
Figure 4. Cross-wavelet spectrum of no-fault current and series fault current
3. F3= s ; 4. F4= s
for the fault at disc no. 30.
W ( s,τ )xy ( s max − s min )(τ max − τ min ) So, the total number of features taken from the cross-
peak
wavelet spectrum for the classification of fault is 12. These
∑∑τ ( F 4 − W xy ( s,τ ) ) 2 features are chosen because of the fact that they represent the
5. F5= s
salient features of the cross-wavelet spectrum. However, one
( s max − s min )(τ max − τ min )
may choose some other features like, location of local peaks
6. F6= “s” at peak of W xy ( s,τ ) i.e. W xy ( s,τ ) of W xy and φ ( s,τ ) surfaces, if any, or some more features
peak

7. F7= “τ” at peak of W ( s,τ ) i.e. W ( s,τ )

xy xy
from W xy and φ ( s,τ ) depending upon the nature of the
peak

Five more features, similar to F1 to F5, are also extracted from problem. In the present case the above-mentioned twelve
the phase angle φ ( s,τ ) data. They are: features are found to be sufficient, because using these
features the localization and identification of the faults can be
8. F8=
∑∑τ sτ φ (s,τ ) ;
s 9. F9=
∑∑τ s τ φ ( s,τ )
2 2
done with reasonable accuracy. To improve the accuracy or
s
for another problem some other features, as stated earlier, may
∑∑τ φ (s,τ )
s
∑∑τ φ ( s,τ )
s
be used.
1300 D. Dey et al.: Rough-granular Approach for Impulse Fault Classification of Transformers using Cross-wavelet Transform

4 INFORMATON GRANULATION USING the set of attributes. V = V , where Vq is the domain of the
U q
ROUGH SETS q∈Q

values of q and f denotes decision function

To classify the extracted features a Rough-Granular
as, f : U × Q → V . If the table is having a large number of
approach is used. It is evident that there is no apriori
knowledge regarding which features of the cross-wavelet attribute values i.e. card(Vq) is very high for some q ∈ Q , then
spectrum will be suitable and should be taken for there is a very low chance of classifying a new object by the
classification of impulse faults. So, the data table obtained rules generated directly from the table. Here, card() means
after the extraction of proposed features may contain cardinality operator, which means “number of elements of a
imprecise or superfluous information. A Rough-granular set”. Therefore, discretization of the decision table is required
classifier is used here as Rough sets are well-suited for in this for large real-valued decision table. Discretization of a data
kind of problems. For example, Rough Set Theory (RST) has table indicates some partitioning of the attribute values. In the
successfully been used for condition monitoring of present problem Maximal Discernible (MD) heuristic is
distribution feeder [10], for fraud detection in electrical followed which is discussed in details in [15]. The discretized
energy consumers [11], in data-mining for semiconductor decision table is shown in Table 2.
manufacturing [12] and also in case generation [13]. In RST, for different attributes, objects are called
Information granulation is basically representation of indiscernible, i.e. similar, if they are characterized by the same
information in the form of some aggregates, called granules information. If P ⊆ Q and xi , x j ∈ U , then xi and xj are
that hold a number of individual entities. Granulation implies
indiscernible wrt the set of attributes P,
that some subsets of the universe can be described
if f ( xi , q) = f ( x j , q), ∀q ∈ P . For example, let P= {F1}. The
approximately. In RST these approximations are termed as
lower and upper approximations. Rough sets granulate the part of the discretized decision table given in Table 2 shows
information by partition [13-14]. Details of the Rough set that, at least objects 1, 3, 4 are indiscernible with attribute F1,
theory can be found in [6, 10-15]. as all these objects are having same value (i.e. value=6) for
In RST, data is presented in a decision table in which each feature F1. Similarly objects 2,…, 84 and 87 are also
row represents an object (e.g. impulse fault information at indiscernible using F1 as they possesses same values (i.e.
different disc position) and each column represents an value =3) for these objects. An elementary set is the set of all
attribute. For example, in the present problem decision table indiscernible objects. So, for P ⊆ Q , an equivalence relation
contains extracted features (F1-F12) as the 12 condition on U, called P-indiscernibility relation is given by IP.
attributes and the fault location and type as the decision Considering the previous example of P= {F1}, it can be found
attribute. For decision attribute the following notation is that the P-elementary set includes elementary sets like
followed in the present problem. ‘SE1’ indicates the series {1,3,4,…}, {2,…,84,87}, {5,6,…,86,89} and {…,85,88} etc.,
fault in the section L1. Similarly ‘SH2’ implies shunt fault in because all the elements of each of these sets are having same
section L2 and so on. The normalized decision table is shown attribute values for attribute F1. The equivalence classes of
in Table 1. Mathematically, information system the partition induced by the P-indiscernibility relation are
T = 〈U , Q, V , f 〉 . Here, U is the finite set of objects and Q is called information granules. This is explained later.

Table 1. Decision table considered for Rule generation

Object Condition Attributes Decision
F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 Attribute

1 0.95 0.92 0.44 0.84 0.30 0.49 0.83 0.17 0.28 0.92 0.41 0.32 SE1
2 0.43 0.93 0.20 0.52 0.18 0.59 0.76 0.97 0.46 0.84 0.30 0.96 SE1
3 0.90 0.97 0.25 0.20 0.19 0.62 0.77 0.27 0.06 0.90 1.00 0.72 SH1
4 0.95 0.40 0.22 0.67 0.68 0.44 0.70 0.25 0.98 0.87 0.01 0.41 SH1
5 0.80 0.13 0.97 0.33 0.10 0.51 0.54 0.87 0.25 0.88 0.56 0.35 SE2
6 0.76 0.91 0.92 0.01 0.54 0.66 0.44 0.73 0.42 0.98 0.97 0.26 SE2
· · · · · · · · · · · · · ·
· · · · · · · · · · · · · ·
· · · · · · · · · · · · · ·

84 0.45 0.41 0.01 0.68 0.15 0.34 0.59 0.13 0.51 0.46 0.99 0.43 SH21
85 0.01 0.89 0.74 0.37 0.69 0.28 0.62 0.01 0.33 1.00 0.48 0.93 SE22
86 0.82 0.05 0.74 0.83 1.00 0.34 0.79 0.35 0.43 0.92 0.43 0.68 SE22
87 0.45 0.05 0.93 0.50 0.86 0.53 0.85 0.19 0.32 0.90 0.49 0.21 SE22
88 0.01 0.81 0.46 0.70 0.65 0.92 0.92 0.49 0.57 0.05 0.21 0.83 SH22
89 0.79 0.02 0.41 0.42 0.59 0.30 0.98 0.66 0.76 0.05 0.64 0.62 SH22
IEEE Transactions on Dielectrics and Electrical Insulation Vol. 15, No. 5; October 2008 1301
Table 2. Discretized Decision table for Rule generation
Object Condition Attributes Decision
F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 Attribute

1 6 5 1 3 2 1 2 0 0 1 1 1 SE1
2 3 5 0 2 1 1 2 2 1 1 1 4 SE1
3 6 5 0 0 1 1 2 0 0 1 3 3 SH1
4 6 2 0 2 5 0 2 0 2 1 0 2 SH1
5 5 0 4 1 0 1 1 2 0 1 1 1 SE2
6 5 5 4 0 3 1 1 2 1 1 3 0 SE2
· · · · · · · · · · · · · ·
· · · · · · · · · Table 3. ·Truncated Decision
· table
· for illustration
· ·
· · · · · · · · Object
· · Condition· Attributes· Decision
· ·
F1 F2 F3 Attribute
84 3 2 0 2 1 0 1 0 1 0 3 2 SH21
1 6 5 1 SE1
85 0 4 3 1 5 0 2 02 13 15 10 4 SE1 SE22
86 5 0 3 3 7 0 2 03 16 15 10 3 SH1 SE22
4 6 2 0 SH1
87 3 0 4 2 6 1 2 0 1 1 1 0 SE22
88 0 4 1 2 4 2 3 1 2 0 0 3 SH22
89 5 0 1 1 4 1 3 1 2 0 2 3 SH22

For any rough set Y, PY and PY are called P-lower and P- {3}, {4}, {5} and {6}; ID = {1,2}, {3,4} and {5,6}.
upper approximation of Y and defined as, Also, POS P ( D) = {1,2,3,4,5,6} . If the attribute F1 is removed
PY = {x ∈ Y | I P ( x) ⊆ Y } and PY = {x ∈ Y | I P ( x) ∩ Y ≠ φ } from P then, POS ( P −{F1 }) ( D) = {1,4,5,6} . Clearly,
respectively. This indiscernibility relation can reduce a POS ( P−{F1 }) ( D) ≠ POS P (D ) . Therefore the attribute F1 is D-
decision table. This can be done by keeping only one element indispensable in P. Similarly, removing attribute F2 gives,
of an equivalence class and also keeping those attributes
POS ( P −{F2 }) ( D) = {1,2,3,4,5,6} = POS P (D) . Therefore
which preserve the indiscernibility relation. In other words,
keeping all the information intact and removing the attribute F2 is D-dispensable in P. Again, for attribute F3 it is
superfluous attributes. Thus obtained minimal sets of
Table 3. Truncated Decision table for illustration
attributes are called Reduct. The CORE is the set of relations Object Condition Attributes Decision
occurring in every Reduct, i.e. CORE ( P ) = I RED ( P ) . Attribute
F1 F2 F3
Therefore, CORE represents the most important part of the 1 6 5 1 SE1
knowledge. From the CORE and Reducts one can generate the
2 3 5 0 SE1
decision rules. Usually these rules are considered in
“IF…THEN” formats. This is illustrated in the following 3 6 5 0 SH1
paragraphs. For better understanding of the method a part of 4 6 2 0 SH1
the decision table is considered as shown in grey color in 5 5 0 4 SE2
Table 2. The truncated decision table is shown in Table 3.
6 5 5 4 SE2
For a given subset P ⊆ Q , an attribute q ∈ P is dispensable
in P if and only if, I P = I ( P −{q}) ; otherwise q is indispensable. easy to observe that, POS( P −{F }) ( D) = {2,4,5,6} ≠ POSP ( D) .
3

If every element in P is indispensable then P is called Similarly, F3 is also D-indispensable in P. Thus, the set {F1,
independent otherwise dependent. Let P ⊆ Q and D ⊆ Q F3} is the D-reduct of P. Therefore, the simplified or reduced
have equivalence relations in U. The P-positive region of D is form of Table 3 is given in Table 4. ‘-’ indicates “don’t care”
indicated as, POS P ( D) = U PY . In other words, it denotes (i.e. dispensable) condition. It can be said that, attribute
Y ∈I D values, (F1=6 ∧ F3=1) ∨ (F1=3∧ F3=0) are the characteristic
the set of elements that can correctly be classified into D- for decision class ‘SE1’. ‘∧’ and ‘∨’ are logical “AND” and
elementary sets obtained from ID using the knowledge “OR” operators respectively. Similarly, (F1=6 ∧ F3=0) is the
described by IP. If q ∈ P and POS P ( D) = POS ( P −{q}) ( D) characteristic of decision class ‘SH1’ and (F1=5 ∧ F3=4) is the
characteristic of ‘SE2’. These are called information granules.
then q is D-dispensable in P, otherwise q is D-indispensable in Intersections of these reduct values for each of the decision
P. If the set of attributes G ( G ⊆ P ) is a D-independent in P class (i.e. SE1, SH1 and SE2) will give the CORE for the
and POS G ( D) = POS P ( D) , then G is called D-reduct of P or respective class. For the decision class ‘SE1’ no such CORE
in general Reduct of P. value is obtained from the Table 4, as the intersection of (F1=6
All these definitions can be explained using Table 3. For ∧ F3=1) and (F1=3 ∧ F3=0) is null. Similarly, for decision class
example, if P is taken as, P = {F1, F2, F3}, and D = {‘Location ‘SH1’ the intersection of (F1= 6 ∧ F3= 0) and (F1=6 ∧ F3=0)
and type of fault’} (i.e. decision attribute), then IP= {1}, {2}, gives CORE values F1=6 and F3=0. Again, for ‘SE2’ the
1302 D. Dey et al.: Rough-granular Approach for Impulse Fault Classification of Transformers using Cross-wavelet Transform

CORE values are F1=5 and F3=4. Furthermore, this reduced the present case no identical attribute or case is observed, the
Table 4 can be used to generate decision rules. The decision discretized table remains unchanged after this step for the
rules obtained from the Reduct and CORE values are given in present problem.
Table 5. Step 3: Dispensable attributes are removed. The method
Table 4. Reduced form of Table 3 described in the previous section is followed to obtain
Object Condition Attributes Decision dispensable attributes and it was found that attributes
Attribute (features) F4, F6, F9 and F10 are dispensable and the remaining
F1 F2 F3
attributes are indispensable. So the reduced decision table is
1 6 - 1 SE1
having 8 columns for condition attributes.
2 3 - 0 SE1 Step 4: Reducts and CORE are obtained. The sample
3 6 - 0 SH1 computation of Reduct and CORE is described earlier.
4 6 - 0 SH1 Following the same procedure the final form of the decision
5 5 - 4 SE2 table is obtained and is shown in Table 6.
6 5 - 4 SE2 Step 5: Decision rules are generated from the final table of
CORE and Reducts. The decision rules are constructed from
Table 5. Decision Rules obtained from CORE and Reducts
the granulated knowledge and some of the rules are shown in
Decision Statement of the Rule
Rule No. Table 7. Thus, 44 IF…THEN rules are obtained. The test
IF THEN
dataset is discretized and tested by the generated decision
1 (F1=6 ∧ F3=1) ∨ (F1=3∧ Location and type of rules to judge the validity of the rules.
F3=0) fault is ‘SE1’
2 (F1=6 ∧ F3=0) Location and type of Table 6. Simplified form of the complete Decision table
fault is ‘SH1’ Object Condition Attributes Decision
3 (F1=5 ∧ F3=4) Location and type of F1 F2 F3 F5 F7 F8 F11 F12 Attribute
fault is ‘SE2’ 1 5 2 1 SE1
- - - - -
2 - 5 - - 2 - 1 - SE1
It is worth mentioning here, that this derivation of decision 3 6 - 0 - 2 0 - - SH1
rules is a demonstrative example of the whole process. While 4 6 - 0 - 2 0 - - SH1
applying these procedures to the complete decision table the 5 5 - 4 - 1 2 - - SE2
rules may be different from these rules. 6 5 - 4 - 1 2 - - SE2

5 RESULTS AND DISCUSSIONS · · · · · · · · · ·

As mentioned earlier, series and shunt faults are simulated
· · · · · · · · · ·
at every disc position and corresponding tank-current · · · · · · · · · ·
84 3 - - 1 - 0 - 2 SH21
waveforms due to applied impulse are recorded. For each of
85 - - - - 2 0 1 - SE22
these waveforms a cross-wavelet spectrum is obtained from
86 - - - - 2 0 1 - SE22
cross-wavelet transform of the fault current data with the
87 - - - - 2 0 1 - SE22
current waveform at no-fault condition. Thereafter the feature
88 - - 1 4 3 1 - 3 SH22
vector for each of the fault condition is extracted. These
89 - - 1 4 3 1 - 3 SH22
feature vectors along with the corresponding fault condition
are tabulated in a table and forms the decision table. The
complete data set is having 178 cases of different fault
conditions, i.e. series and shunt faults at each of the 89 disc Among the 89 test cases the Rule set obtained from
locations. 50% of the data, randomly chosen, are used to proposed information granulation technique has correctly
develop the decision rules and the remaining 50% data are classified the fault 81 times. So, the percentage of success is
used for testing purpose. Therefore, the data table used for almost 90%, which is reasonable for localization of the fault
rule generation is having 89 rows (objects) and 13 columns within 5% of the winding length (i.e. within 4 discs out of
(excluding the column of object index), as evident from Table total 89 discs in the winding under investigation). It is to be
1. The first twelve columns consist of the features from F1-F12, mentioned here that earlier researchers have already reported
which are called condition attributes and the last column is for localization of fault within 20% [3] and 10% [4] of the
the decision attribute, i.e. fault location and type, as stated winding length.
earlier. The values of the condition attributes are obtained The results shown above are obtained from the simulation
from the equations for F1-F12. To obtain the decision rules for of faults on the analog model. The performance of the
classification of faults following steps are followed, proposed scheme is also verified with an EMTP based digital
Step 1: The data table is discretized. This means that each model as shown in Figure 2. Fault simulation, feature
condition attribute values are divided into ranges. In the extraction and decision rule generation are performed
present problem Maximal Discernible (MD) heuristic is similarly as described for the case of analog model. The final
followed which is discussed in details in [15], as stated in the rule set is shown in Table 8. Among 89 test cases, developed
previous section. The discretized table is shown in Table 2. rules can successfully classify the fault type and location for
Step 2: Identical Attributes and cases are eliminated. As in 83 times. So the percentage of success is 93%.
IEEE Transactions on Dielectrics and Electrical Insulation Vol. 15, No. 5; October 2008 1303
Table 7. Decision Rules obtained from the analysis of analog model [6] Z. Pawlak, Rough Sets: Theoretical Aspects of Reasoning about Data,
Decision Statement of the Rule Kluwer, Boston, USA, 1991.
Rule No. IF THEN [7] P. Purkait and S. Chakravorti, “Time and frequency domain analyzer
1 (F2=5∧ F7=1∧ F11=1) Location and type of based expert system for impulse fault diagnosis in transformers”, IEEE
fault is ‘SE1’ Trans. Dielectr. Elect. Insul., Vol. 9, pp. 433-445, 2002.
2 (F1=6 ∧ F3=0 ∧ F7=2 ∧ F8=0) Location and type of [8] B. G. Ruessink, G. Coco, R. Ranasinghe, and I. L. Turner, “A cross-
fault is ‘SH1’ wavelet study of alongshore non-uniform near shore sandbar behavior”,
3 (F1=5 ∧ F3=4 ∧ F7=1 ∧ F8=2) Location and type of Proc. Joint Conference on Neural Networks, Vancouver, Canada, pp.
fault is ‘SE2’ 4310-4317, 2006.
· · · [9] C. Torrence and G. P. Compo, “A Practical Guide to Wavelet Analysis”,
J. American Meteorological Soc., Vol. 79, pp. 61-79, 1998.
· · · [10] J.T. Peng, C.F. Chien and T.L.B. Tseng, “Rough set theory for data
· · · mining for fault diagnosis on distribution feeder”, IEE Proc.-Gener.
Transm. Distrib., Vol. 151, pp. 689-697 , 2004.
42 (F1=3 ∧ F5=1 ∧ F8=0 ∧ F12=2) Location and type of [11] J.E. Cabral, J.P. Pinto, E.M Gontijo and J.R .Filho, “Fraud Detection in
fault is ‘SH21’ Electrical Energy Consumers Using Rough Sets”, Proc. IEEE Conf.
43 (F7=2 ∧ F8=0 ∧ F11=1) Location and type of Systems, Man and Cybernetics, pp. 3625-3629, 2004.
fault is ‘SE22’ [12] A. Kusiak, “Rough Set Theory: A Data Mining Tool for Semiconductor
44 (F3=1 ∧ F5=4 ∧ F7= 3 ∧ F8=1 Location and type of Manufacturing”, IEEE Trans. Electronics Packaging Manufacturing,
∧ F12=3) fault is ‘SH22’ Vol. 24, pp. 44-50, 2001.
[13] S. K. Pal and P. Mitra, “Case Generation Using Rough Sets with Fuzzy
Table 8. Decision Rules obtained from the analysis of EMTP model Representation”, IEEE Trans. on Knowledge and Data Engg., Vol. 16,
Decision Statement of the Rule pp. 292-300, 2004.
Rule No. IF THEN [14] Y.Y. Yao, “Rough Sets, Neighborhood systems and Granular
1 (F2=4∧ F7=1∧ F11=1) Location and type of computing”, Proc. IEEE Conf. on Electr. and Comp. Engg., pp. 1553-
fault is ‘SE1’ 1558, Canada, May 1999.
2 (F1=5 ∧ F4=1∧ F7=2 ∧ F8=1) Location and type of [15] H. S. Nguyen, Discretization of Real Value Attributes: A Boolean
fault is ‘SH1’ Reasoning Approach, PhD Thesis, Dept. of Mathematics, Warsaw Univ.
3 (F1=4 ∧ F4=4 ∧ F7=1 ∧ F8=1) Location and type of Poland, 1997.
fault is ‘SE2’
· · · Debangshu Dey was born in Kolkata, West Bengal
in April 1980. He got his B.E.E and M.E.E. degrees
· · · from Jadavpur University, Kolkata, India in 2003 and
· · · 2005 respectively. Presently he is working as Senior
Research Fellow in High Tension Laboratory,
42 (F1=4 ∧ F4=2 ∧ F8=0 ∧ F12=2) Location and type of
Electrical Engineering Department, Jadavpur
fault is ‘SH21’
University, Kolkata, India. He has published ten
43 (F7=2 ∧ F4=3 ∧ F8=0 ∧ F11=1) Location and type of research papers. His areas of interest are sensor
fault is ‘SE22’ linearization, intelligent instrumentation, signal
44 (F3=1 ∧ F2=2 ∧ F7= 3 ∧ F8=2) Location and type of conditioning and application of optimization and
fault is ‘SH22’ computational intelligence in electrical measurements.

Biswendu Chatterjee was born in Kolkata, West

6 CONCLUSIONS Bengal, India, in February 1973. He obtained the
M.E.E. degree from Jadavpur University, Kolkata,
A novel approach based on information granulation using
India in 2004. Presently he is working as a Research
Rough sets for impulse fault identification of transformers is Associate in High Tension Laboratory, Electrical
proposed and investigated. The performance of the scheme is Engineering Department, Jadavpur University,
shown for both analog and EMTP based digital models of Kolkata, India. His areas of interest are partial
transformers. Results show that cross-wavelet transform may discharge (PD) measurements and instrumentation,
EMI reduction techniques and condition monitoring
be considered as an efficient tool for feature extraction and the of large electrical equipment.
Rough-granulation based classification approach is suitable
for transformer impulse fault classification and localization. Sivaji Chakravorti (M'90-SM'00) obtained his
B.E.E., M.E.E. and Ph.D. degrees from Jadavpur
University, Kolkata, India in 1983, 1985 and 1993
REFERENCES respectively. Since 1985 he has been a full-time
[1] Guide to the Lightning and Switching Impulse Testing of Power faculty member of the Electrical Engineering
Transformers and Reactors, IEC Std., Publ. 722, 1982. Department of Jadavpur University, where he is
[2] A. De and N. Chatterjee, “Recognition of impulse fault patterns in currently Professor in Electrical Engineering. In 1984
transformers using Kohonen’s self-organizing feature map”, IEEE Trans.r he worked at the Indian Institute of Science,
Power Del., Vol. 17, pp. 489–494, 2002. Bangalore as Indian National Science Academy
[3] A. De and N. Chartterjee, “A fuzzy ARTMAP fault classifier fo Visiting Fellow. He worked at the Technical University Munich as Humboldt
impulse testing of power transformer”, IEEE Trans. Dielectr. Electr.r Research Fellow in 1995-96, 1999 and 2007. He served as Development
Insul., Vol. 11, pp. 1026–1036, 2004. Engineer in Siemens AG in Berlin in 1998. He has also worked as Humboldt
[4] C. Koley, P. Purkait and S. Chakravorti, “Wavelet-Aided SVM Tool fo Research Fellow in ABB Corporate Research at Ladenburg, Germany in 2002.
Impulse Fault Identification in Transformers”, IEEE Trans. Power Del., In 2003 he worked as US-NSF guest scientist at the Virginia Tech, USA. He
Vol. 21, pp. 1283-1290, 2006. has published about 100 research papers and has authored a book. He is the
[5] D.Dey, B. Chatterjee, S. Chakravorti and S.Munshi, “Hybrid Filtering recipient of Technology Day Award of AICTE for best project work in 2003.
Scheme for Proper Denoising of Real-Time Data in Dielectric His current fields of interest are numerical field computation, computer aided
Spectroscopy”, IEEE Trans. Dielectr. Electr. Insul., Vol. 14, pp. 1323- design and optimization of insulation system, condition monitoring of large
1331, 2007. electrical equipment and signal conditioning in high voltage systems.
1304 D. Dey et al.: Rough-granular Approach for Impulse Fault Classification of Transformers using Cross-wavelet Transform

Sugata Munshi obtained the B.E.E and M.E.E.

degrees from Jadavpur University, Kolkata, India in
1980 and 1985, respectively. He worked as an
Engineer in the Plasma Physics Division of Saha
Institute of Nuclear Physics, India from 1985 to
1990. In 1986, he had a training on ‘Tokamak’
machine in the ‘Heavy Engineering Works’ of
Toshiba Corporation in Japan. In 1990 he joined the
Electrical Engineering Department of Jadavpur
University as a faculty member. At present, he is a Reader in this
department. He has published about 30 research papers in refereed journals.
He was the joint recipient of The President of India’s Prize (English) in
1989-90, The Pandit Madan Mohan Malaviya Memorial Prize in 1989-90,
The Sir Thomas Ward Memorial Prize in 1994-95, The Tata Rao Medal,
awarded by the Institution of Engineers (India) in 1996-97 and Certificate
of Merit from IE (India) in 1996-97.His current fields of interest are signal
processing, surge phenomena in power equipment and sensor systems.