You are on page 1of 6

IEEE PES PowerAfrica 2007 Conference and Exposition

Johannesburg, South Africa, 16- 20 July 2007

Using Cross-Correlation Coefficients to Analyze Transformer Sweep Frequency


Response Analysis (SFRA) Traces
G.M. Kennedy, A.J. McGrail, and J.A. Lapworth
Doble Engineering Co 85 Walnut St, Watertown, MA 02472 USA, and Doble PowerTest Ltd, 5 Weyvern Park, Guildford,
Surrey, UK, GU3 ILZ, Phone, +44 1483 514123 Email:jlapworth@doble.com
Abstract Sweep Frequency Response Analysis (SFRA) is one of
the most powerful diagnostic tools for assessing mechanical
damage to a transformer winding. Analysis of the rsults, which
are in the form offrequency response traces can, however, be
daunting to new users. One approach is to attempt automated
analysis, and this paper presents an approach using crosscorrelation coefficients. This is apower signalprocessing tool that
can simplify result interpretation in some cases and allow limits to
be created.

INTRODUCTION

A common failure mode for power transformers is


consequent to mechanical deformation of the core or
windings. Core damage is more likely as a result of
transportation, while winding damage is more likely to be
caused by short circuit type forces. Consequently, the
requirements are for transformer to be checked before and
after a new delivery or re -location, and also after any major
fault. The latter might be after faults on a cable connected to
the low voltage winding, or on a measurement transformer on
the high voltage, or a tap changer or bushing failure. These
close-up faults may be so severe as to initiate protection and
then the need is to assess he damage to the transformer.
Alternatively, there could be lesser damage in the form of
deformation. This would reduce the capability of the winding
to withstand any further faults- i.e. the winding is bent but
not broken. Knowledge of such damage is, therefore, part of a
risk assessment process to be applied for critical units.
Within IEC and IEEE standards various diagnostic
techniques are described, and these are to be used after
routine or type testing. Two relevant to this discussion
include turns ratio to detect if the winding has faulted, and
leakage reactance to identify deformation. Over the years the
measurement of leakage reactance (or short-circuit
impedance) has proven valuable, particularly when used
during with-stand evaluation in high power laboratories. The
impedance change allowed post fault can be defined as being
less than a 2 percentage change (see IEC60076-5). Provided
the test is done correctly this provides a clear definition for
acceptability. However, for in-service assessment it is
considered by some that this method is insensitive. In the
1990s a group from the major utilities in northern Europe
evaluated an alternative method involving the injection of a
low voltage swept frequency sine wave into each winding in
turn. They met as a working group of EuroDoble and at the
end of the 1990s documented their experience in a Doble test

1-4244-1478-4/07/$25.00 2007 IEEE

guide and several papers at the annual Doble conferences [1],


[2]. Today, groups within IEEE and CIGRE are working to
introduce the method into IEEE and IEC standards, and
several companies now produce instruments to replace the
laboratory equipment used earlier. Over these recent years the
technique has gained acceptability with a much larger and
rapidly expanding group (and including the recent-entry
suppliers of equipment). But this has created some problems
for these new users and suppliers relating to the interpretation
of the results. The experience is that this is a transitional
phase and new users can acquire the skill within a short time.
But for that first phase there is an interest in having an
analytical method, ideally producing a number, ideally as
simple as that from leakage reactance.
II

THE NATURE OF SFRA RESULTS

The sweep frequency response analysis (SFRA) test


involves injecting a signal at one end of a winding and
measuring the response at the other end. Responses with
large variations in attenuation over the measured frequency
range are obtained as a result of variations in the impedance
of the complex L-C-R distributions of the windings. Since
capacitances and inductances depend on detailed winding
geometry, any movement results in changes in the
frequencies at which resonances occur. It is the identification
of changes in frequency response that is the essence of
analysis and diagnosis of mechanical integrity. This is
currently achieved with the expert eye - but the aim is to use
some processing to yield numerical evaluation of changes.
But differences in response also occur for reasons other than
deformation. These include differences between individual
phases, tap changer position and configurations of internal
leads between the bushings and windings- as well as whether
the oil was present or absent. Unsuitable test equipment, lead
connections and layout can produce repeatability issues.
These issues present considerable challenges when
contemplating any automation of the analysis.
Figure 1 shows a typical response for ahigh voltage star
connected winding. The frequency range of interest is
between 20Hz and 2MHz. Experience has shown that
different sub-bands are dominated by different internal
components of the transformer and are subsequently more
sensitive to different types of failures, as summarized in
Table I Measurements above 2MHz tend to be dominated
by variations in grounding practices for test leads.

F igure

I -

F requency Analysis banus

Table 1: Frequency Sub-Band Sensitivity

Region

I1

3
4

Frequency
Sub-Band
< 2kHz

Component

Failure Sensitivity

Main Core
Bulk
Winding
Inductance

Core Deformation,
Open
Circuits,
Shorted Turns and
Residual
Magnetism
Bulk
Winding
Movement between
and
windings
clamping structure
Deformation within
the Main or Tap

Bulk
Component
Shunt
Impedances
20 kHz to Main
400 kHz
Windings

2kHz to 20
kHz

400 kHz to
-1 MHz

Main
windings,
Tap
windings
and internal
leads

windings

Movement of the
main and tap

windings, Ground
impedances
variations

Normal practice is to separate the response into frequency


bands and relate the analysis to features in each band, as
indicated in Table 1. The first resonance, here at around 600
Hz, is where the change from inductive to capacitive
impedance occurs. The resonances at higher frequencies
relate to the winding configuration. The above regions are a
general rule of thumb dependent on transformer design and
test conditions.

Sub-bands should be modified and adjusted by the user


based on the observed waveform. For instance, the low
frequency Region 1 could be changed to include all
frequencies between the lowest and the first dominant pole,
to completely encapsulate the core effect, or a sub band
within region 3 could be chosen that includes a series of
dominant resonances.
In the analysis a response may be compared with one of
the followinga) An earlier result for the same phase tested with the
same tap changer position.
b) If no earlier result is available, then another phase of
the same transformer, tested at the same occasion.
c) The same phase, same tap changer position but on a
unit believed to be of the same design group and
made at the same factory.
The preference for selection of options is in the order as
given.

III

CROSS CORRELATION ANALYSIS

Several attempts have been made over the years at


creating automated or semi-automated SFRA analysis tools.
Nearly all have been unsuccessful. The reasons for this lie not
in the capability of the calculation or the application, but in
three main areas:
Large generic differences between responses of
different winding types. This means that setting
cross-correlation bands 'universally' is difficult, if
not impossible; hence the recommendation in this
paper to make sure the user is able to set the band
limits to match the transformer and winding under
-

test.

the fact that some differences between responses are


inevitable and must be allowed, e.g. due to
measurement limitations or manufacturing
tolerances. Unless appropriate allowances are made,
dependent on transformer designs and
manufacturers, there is a risk that there will be false
positives in the analysis.
the subtle nature of some failures, which mean that
the SFRA traces are 'almost acceptable', leading to
a possible false negative in the analysis.

The traditional view of the Doble team, some of whom


have been reviewing data for almost 20 years, is that where
unusual results are obtained, any decision must always be
made by the expert to evaluate the existence of interferences
from issues listed above. Where automated systems can help
is as a short term crutch for the new user. Most applications
will be for fingerprinting sound transformers Where there
should be good overlay between other phases, sister units etc.
Here some comfort factor will be a help to the new user since
it will be able to identify the normal low level differences
between traces being compared. The method can act as a
filter so that the engineer can use their limited resource of
time looking at the most important traces. The output should
not be "pass/fail" but "pass/investigate". The goals are to
express results as "pass" where variations are within normal
bounds for the configuration and a means to communicate
results to end-customers who are not so familiar with SFRA.
Cross correlation coefficients (CCFs) are already used in a
variety of industries, primarily telecom, where knowing the
quality of signals is important. They are already well
understood and provide the type of analysis needed without
blurring the analysis with "go no-go" type outputs.
In simplest terms, cross-correlation takes two sets of
numbers and looks at how similar they are. If two series of
numbers sach as an SFRA trace perfectly or nearly match,
they would have a CCF very close to 1.0. If two traces have
absolutely no correlation, in other words are completely
random, they would have a CCF of 0.0. If the two traces are
related by are diametrical, they would have a negative CCF.
In SFRA analysis negative CCF are not common, but they do
occur on occasion.
Regardless, negative correlation
coefficients are not considered acceptable when trying to
look for deviations between traces.
The CCF is defined [2] as:
n

CCF =

,(XI.

i=l

4E(Xi-X)2

XX)t -Y)
*7E(yj-y)2

Where Xi and are Yi are the two series (or traces in the case
of SFRA) being compared at each individual frequency 'i'
and X-bar and Y-bar are the means.

The equation (1) assumes two real series. In the case of


signal processing the math becomes a little more involved,
but the end results is still a coefficient between 1 and-1.
In simplest terms cross-correlation of two traces f and g is:

(f* g)(x) |f*(t)g(x +t)dt


=

(2)

This results in a new correlation trace that is then power


density normalized to the input and output signals. The large
* indicates convolution.
The mathematics of (2) and
subsequent CCF calculation are beyond the scope of this
paper and can be researched in a variety of mathematical
texts on the subject of signal processing [3].
Normalizing the results to the individual power spectrums
is what allows this resulting waveform to be expressed in a
simple single coefficient. Table 2 helps provide a rough
estimate of what the CCF means in simple language.
Table 2: Example CCFs explained

Good Match
Close Match
Poor Match
No or very poor
match
IV

CCF
0.95-1.0
0.90-0.94
<.89
<= 0.0

CALCULATING CROSS-CORRELATION
COEFFICIENTS:

Using CCFs to analyze SFRA data first requires an


understanding of what frequency sub-bands tell us about the
physical health of a transformer. Once the appropriate bands
are selected the CCFs can be evaluated in the context of the
individual parts of a transformer. The end results will be a
series of CCFs, evaluating different components of the
transformer.
The following traces were obtained using the Doble SFRA
software for M5200/M5300. By adjusting cursor positions to
frequency sub-bands, new auto-correlation constants can be
calculated. Figure 2 shows the comparis on of two phases of a
transformer where the cursors were set to the positions shown
in Table 1. Note that the first five cursor positions correlate
to the regions. The correlation coefficients are shown below
the cursor positions to indicate the correlation between the
two above frequency regions.

always be compared. In addition, the analysis should be done


with the transformer in the same tap changer positions. The
following table provides recommend correlation coefficients
limits for benchmark comparisons.
Table 4: Benchmark Limits

Good
Marginal

Investigate

Figure 2: Cross Correlation Analysis of the H2-HO Winding to


Benchmark Results

By observing the above two phases of the same unit, we can


see that there are possible problems noted with the lower
correlation coefficients of 0.9177 and 0.1906. The variations
can be seen in the overlay of the two phases. Additional
analysis and perhaps narrowing the sub-band could help tell
us more. The Region 4 correlation coefficient of 0.1906 is
most likely due to the mismatch of resonance in the 600 kHz
to 900 kHz region. It is important to note that this does not
necessarily indicate that the unit is bad as this highest
frequency region is known to be much more sensitive to tap
winding and internal leads. This type of variation could be
normal between phases of the same unit.
Phase to phase comparison can often be misleading and
must be done with care. There are expected variations
between phase and correlation coefficients limits should be
selected that draw attention to the analysis, and not
necessarily call for failure of the unit. Based on experience,
phase to phase analysis would need to have the most
forgiving limits. Phase to phase analysis of the outside core
legs will also be the most credible as these two windings
display the most geometry symmetry. Comparison to the
middle phase will be difficult with correlation coefficients
and should be analyzed using traditional SFRA methods.
Sister units can also be used to conduct correlation
coefficient analysis. In this case, we would expect some
variation between genuine sister units, but the analysis must
be done on the exact same phases of both units (A to A, B to
B, etc.) The following correlation coefficients limits are
recommended for phase to phase and sister unit comparison.
Table 3: Phase A/C

or

Good
Marginal
Investigate

Sister Unit (same phase) Limits

TCCF
0.95-1.0
0.90-0.94

<0O90

The most reliable method of conducting correlation


analysis is using a benchmark results. Similar to sister unit
analysis, the same phases of the same windings should

CCF
0.98-1.0
0.96-0.97
<0.96

These limits should be used as an aid to analysis and


reporting results but should not be used in isolation to judge
the mechanical condition of a transformer. A SFRA
technician should still analyze the results and use these limits
as an aid for additional investigation.
V

CASE STUDIES

The following case studies show various scenarios that a


SFRA tester may encounter in the field and demonstrates
how to use correlation coefficients in the analysis
Case 1: Benchmark Comparison
In this case we have factory and initial field results
available for a 675 MVA Generator Step Up transformer built
in 2002. The transformer suffered from a fire in the
connected isophase bus and was further tested using SFRA
and other electrical diagnostic tests. This case was presented
as a paper at the 2005 Doble Client Conference.
Variations at low frequencies relate to the magnetic state
of the core of the transformer; the variation is commonly
seen, is well understood and is acceptable. Minor variations
at the highest frequencies are due to minor differences in test
lead grounding - due to bushings which had been damaged
during the fire. The main responses overlay very well. This is
strong evidence that nothing has moved within the
transformer. SFRA was used in conjunction with Leakage
Reactance to propose that the transformer was mechanically
sound and was worthy of an internal inspection rather than
scrapping. The inspection confirmed the SFRA diagnosis and
the transformer was successfully returned to service.
Using the cross correlation coefficient analysis on the H2HO winding and default regions, we find that the unit was
acceptable with the following cross correlation coefficients.
As mentioned above, the slight dip in the Region 1 CCF is
due to the core magnetization and there were some minor
variations in the Region 3 but the CCF was still very good at
0.9882. This helps draw the technician's attention to these
two sub-bands.
The magnetization could be easily
recognized and the minor variation in Region 3 ruled out as
an issue. The end result is four figures that can be used to
quantity and report SFRA findings in an easy to interpret

format. All findings are above the recommended limit of


0.98 CCF for benchmark comparison.
Table 5: Case 1 CCF Results

Frequency Sub Band


0- 2kHz
I
2
2kHz- 20 kHz
3
20 kHz- 400 kHz -4
400 kHz- 1 MHz

---

CCF
0.9879
0.9964
0.9882
0.9988

Figure 4: Cross Correlation Analysis of Case 2 (Phase to Phase)Scan From 20Hz- 2MHz

Figure 3: Cross Correlation Analysis of Benchmark Results H2-H0


Winding

Case 2: Bent Transformer

The results here are from a 1960's vintage 50MVA


distribution transformer. The transformer had tripped out of
service on protection and was investigated; no reference
results were available for this unit. The HV results phase-tophase had typical variation for a HV delta winding.
An overall look at the LV windings shows a few obvious
variations, and a trained eye can see that this is particularly at
the resonances between 200 kHz and 2 MHz where there are
some shifts which are unexpected, Fig 4. If we look closer,
Fig 5, we can see that it is the X3-XO phase which is
consistently shifted to higher frequencies. This is an
indication of a problem which may relate to axial winding
movement. Note that we are also using the similarity of the
XI-XO and X2-XO phases to act as reference traces; if the
center and one outer phase are similar then, generally, the
other phase should also be similar! This transformer was
removed from service and scrapped.
Table 6: Case 2 CCF Results

Frequency Sub Band


0 - 2 kHz
2kHz- 20 kHz
20 kHz- 400 kHz
400 kHz- 1 MHz

CCF
0.9831
0.9868
0.8262
0.9567

Figure 5: Close-up view of LV Winding Frequency Shifts in 100kHz1MHz

Using CCF analysis methods, again the regions were set to


the default sub-bands. The Phase A and Phase B CCFs were
then calculated. We can see in Table 6 that there is a large
dip the Region 3 CCF down to 0.8262. Using the
recommended limits for analyzing phase to phase (Table 3),
we can see that Regions 1 and 2 are all within the good limit
and do not warrant continued investigation. Region 4 is
below 0.98 but above 0.9 and therefore is marginal and
deserves a closer look. Region 3 has dropped into the
investigate region (less than 0.9 CCF). Clearly the region did
not correlate as was discussed earlier.
Case 3: Two Large Sister Transformers

The following case involves the routine test of two large


transformers of 370 MVA 345/14 kV rating. SFRA tests
were used to ascertain the physical health of the units
individually then sister unit comparison was used to ascertain
the consistency of construction between the two units. By
using both of these techniques it was determined by both
phase to phase and by sister unit comparison that both units
were in good physical condition without indications of either
winding or core deformation.
Comparing sister units must be done with care to ensure
that the units are really sisters by design. Experience has

shown that it is not unusual for functionally identical units to


have significant design differences, particularly if some time
has elapsed between constructions. Manufacturers may take
the liberty to tweak the design of a transformer thus creating
physical deviations in the structures. In this case, they are not
sister units. Looking for successive serial numbers may not
guarantee two or more units are genuine sisters; more
attention should be paid to the date of manufacture. As a rule
of thumb, if more than six months has elapsed between
constructions, there may be some differences.
Table 7: Case 3 CCF Results - HV Phase A Sister Units
f

Frequency Sub Band


0 2 kHz
2kHz- 20 kHz
20 kHz- 400 kHz
400 kHz- 1 MHz
-

_ CCF

0.9898
0.9994
_ 0.9913
0.9923

VI

CONCLUSION

CCFs can be a useful tool to conduct specification based


analysis of SFRA traces under a variety of scenarios to
include benchmark results, phase to phase analysis and sister
unit comparison. Care must be taken to understand the
context of the trace and know that some variation is expected
based on tap changer position, test conditions and the
transformer design.
CCFs should be used with an understanding of traditional
SFRA analysis techniques. There is no replacing the trained
eye of an SFRA technician as they can recognize the features
and patterns associated with various faulted conditions. CCF
will not tell us the exact nature of the failure, only draw our
attention to regions of interest for continued analysis and
consideration. Users of this method should also be aware
that it is possible for the CCF to not indicate a problem when
there could be. Inappropriately assigned region boundaries,
or single resonance shifts for example, could cause only a
minor change to the CCF and in truth indicate a substantial
problem.
The suggested CCF limits should be used as general rules
of thumb and can be adjusted to fit the scenario in question.
For instance, comparing closely design sister units may
necessitate raising the CCF limits proposed in Table 3.
These "figures of merit" can help communicate SFRA
findings to end customers. Again, these should serve as
starting points for analysis and may change as more research
is done and CCFs are applied to more transformers.
REFERENCES

Figure 7: Cross Correlation Analysis of Case 3 (Phase to Phase


Comparison of HV Phase A- Sister Units)

Table 7 and Figure 7 show an example analysis of two


sister units. As can be seen from the CCFs, the results show
very good correlation. Using the recommended limits for
sister unit analysis (Table 3), we can see that all correlation
coefficients are well above the recommended limits and in
fact these two units are extremely close sister units in design.
The slight dip in the region 1 CCF is most likely due to a
very slight amount of reluctance variation in the core and is
normal.

[1] J.A. Lapworth and T.J. Noonan, "Mechanical condition


assessment of power transformers using frequency
response analysis" Proceedings of the 1995 International
client conference, Boston, MA, USA
[2] G.M. Kennedy, C.L. Sweetser and A.J. McGrail, "Field
Experiences with Sweep Frequency Response Analysis",
Proceedings of the 2006 International Conference of
Doble Clients. Sec T-6
[3] A.V. Oppenheim and R. Schafer, Discrete-Time Signal
Processing, Prentice-Hall Signal Processing Series, 1989

You might also like