Fiber Fabry–Perot sensors for detection of partial
discharges in power transformers
Bing Yu, Dae Woong Kim, Jiangdong Deng, Hai Xiao, and Anbo Wang
A diaphragm-based interferometric fiber optic sensor that uses a low-coherence light source was designed
and tested for on-line detection of the acoustic waves generated by partial discharges inside high-voltage
power transformers. The sensor uses a fused-silica diaphragm and a single-mode optical fiber encap-
sulated in a fused-silica glass tube to form an extrinsic Fabry–Perot interferometer, which is interrogated
by low-coherence light. Test results indicate that these fiber optic acoustic sensors are capable of
faithfully detecting acoustic signals propagating inside transformer oil with high sensitivity and wide
bandwidth. © 2003 Optical Society of America
OCIS codes: 060.2370, 120.2230, 120.1880.
1. Introduction
Power transformers are the most critical and expen-
sive components in the power industry. The occur-
rence of partial discharges 共PDs兲 within transformers
may lead to insulation breakdown and catastrophic
failures whose effects may be large in both safety and
financial terms. The cost of each failure can easily
drive the total cost of a single transformer failure to
more than $10 million. Hence it is important that
PD activity be monitored and studied to detect incip-
ient insulation problems, to prevent catastrophic fail-
ures and to avoid extensive costs.
The use of electrical, chemical, and acoustic mea-
surements to detect PDs inside power transformers
has been extensively studied. The electrical method
may provide accurate recordings of PDs in laboratory
conditions but is difficult to apply in the field for
in-service transformers because of the high level of
electromagnetic interference 共EMI兲 and the impossi-
bility of achieving accurate calibrations. Current
chemical methods detect PDs in a transformer by
taking gas or oil samples from the transformer.
More-recent research includes the development of
gas sensors and their application for on-line gas
monitoring.
1–3
One major problem associated with
chemical methods is that there is a long delay be-
tween the initiation of a PD source and the accumu-
lation of enough gas to be detectable. For electrical
and chemical methods a further limitation is that it is
generally not possible to determine the exact location
of a detected PD source, information that is also im-
portant in PD study. In general, a PD results in a
localized, nearly instantaneous release of energy. It
produces ultrasonic waves that propagate through
the insulation oil of a transformer. An acoustic sen-
sor can detect ultrasonic waves and generate useful
information relevant to the PD sources.
4–6
One ob-
vious advantage of the acoustic methods is that by
using them one can locate the position of a PD by
studying the phase delay or the amplitude attenua-
tion of the acoustic waves. Furthermore, acoustic
methods have the potential advantage of better noise
immunity in on-line PD detection applications.
One can achieve acoustic PD detection by mounting
piezoelectric acoustic sensors externally on the walls
of a power transformer, and often a suitable coupler
is used to enhance the acoustic wave’s coupling to the
sensors. An externally mounted piezoelectric acous-
tic sensor offers the advantages of easy installation
and replacement. However, a piezoelectric sensor
may suffer from degeneration of the signal-to-noise
ratio caused by environmental noises such as EMI
and corona effects. Another possible disadvantage
associated with an externally mounted piezoelectric
sensor is that the multiple paths of acoustic wave
transmission make locating the exact positions of
partial discharges difficult.
7,8
It is thus desirable to
When this research was performed, the authors were with the
Center for Photonics Technology, Virginia Polytechnic Institute
and State University, Blacksburg, Virginia 24061-0287. J. Deng
is now with Nanoopto Corporation, Somerset, N.J. 08873. H. Xiao
is now with Agere Systems, Inc., Allentown, Pennsylvania 18109.
B. Yu’s e-mail address is biyu@vt.edu.
Received 16 September 2002; revised manuscript received 21
January 2003.
0003-6935兾03兾163241-10$15.00兾0
© 2003 Optical Society of America
1 June 2003 兾 Vol. 42, No. 16 兾 APPLIED OPTICS 3241
Optical Society of America. Bing Yu, Dae Woong Kim, Jiangdong Deng, Hai Xiao, and Anbo Wang, "Fiber Fabry-Perot Sensors
for Detection of Partial Discharges in Power Transformers," Appl. Opt. 42, 3241-3250 (2003). doi: 10.1364/ao.42.003241
have sensors that can function reliably inside a trans-
former, even deep within the transformer windings,
to pick up clean PD-induced acoustic signals. For
the sake of safety and easy installation these sensors
have to be chemically inert, electrically nonconduct-
ing, passive, and small in size.
Optical fiber-based sensors have been shown to be
attractive devices with which to measure a wide
range of physical and chemical parameters because
the sensors have a number of inherent advantages,
including small size, light weight, high sensitivity,
high frequency response, electrical nonconductivity,
and immunity to EMI noise.
9
These advantages
make optical fiber sensors excellent candidates for
PD detection. Fiber optical acoustic sensors have
been shown useful in many applications, such as un-
derwater hydrophones,
10,11
material property analy-
sis, civil structure nondestructive diagnosis,
12,13
vehicle detection and traffic monitoring,
14
and
partial-discharge detection.
15,16
Early fiber optic
sensors for acoustic signal detection were based
mostly on fiber optic intrinsic interferometers such as
all-fiber Michelson interferometers and Mach–
Zehnder interferometers.
These intrinsic fiber sensors usually use single-
mode fiber and coherent sources, such as lasers.
The light from a source is split into two fibers with
equal intensity by a 3-dB fiber coupler. One fiber,
referred to as the sensing arm, is exposed to the
acoustic signal, and the other, referred to as the ref-
erence arm, is shielded from the impact of the acous-
tic wave. Either the reflections 共Michelson兲 or the
transmissions 共Mach–Zehnder兲 of the light beams
propagating in the two arms are recombined to gen-
erate interference signals that are modulated by the
acoustic waves. The intrinsic fiber interferometric
sensors have shown high sensitivity when a long fiber
was used in the sensing arm. However, they suffer
from the fringe fading problems that result from ran-
dom polarization rotation. They are also unstable
because of drift in the source wavelength and
temperature-induced path-length changes.
More recently, fiber optic extrinsic Fabry–Perot in-
terferometric 共EFPI兲 sensors have been under devel-
opment for acoustic-signal detection.
17–20
Fiber
EFPI sensors are fabricated with a small sensing
element known as a Fabry–Perot cavity formed by
two parallel reflecting surfaces. Compared to Mich-
elson and Mach–Zehnder fiber sensors, EFPI sensors
are compact in size and therefore achieve virtually
single-point measurement. More importantly,
those random polarization rotation and phase
changes that are environmentally induced in the fi-
ber connecting the optical source, the sensor head,
and the detectors are common mode and therefore do
not affect the signal phases.
The Center for Photonics Technology of the Vir-
ginia Polytechnic Institute and State University
共CPT VT兲 has a research group experienced in the use
of fiber optic EFPI sensors in harsh environments
and has demonstrated a diaphragm-based EFPI pro-
totype sensor system that uses a distributed-
feedback laser source for on-line detection of partial
discharges in power transformers.
21
The PD sen-
sors are based on a Fabry–Perot 共F-P兲 interferometer
that comprises a diaphragm and the end face of a
single-mode fiber. The vibration of the diaphragm
caused by PD-generated acoustic waves operates the
interferometer in the linear range of one of its inter-
ference fringes. EFPI sensors operating in the lin-
ear range eliminate the common problems, such as
nonlinear transfer functions, complex signal process-
ing, and directional ambiguity, of sensors’ measuring
changes in air gap change by fringe counting. How-
ever, the PD detection system does suffer from source
power fluctuation caused by backreflection, initial
quadrature-point 共Q-point兲 offset because of fabrica-
tion tolerance, operating-point drift caused by static
pressure in the transformer oil, and temperature
drift. Although demodulation by multiwavelength
interrogation
14,22
or spectral interrogation
22
is suc-
cessful in solving the nonlinear transfer function, di-
rectional ambiguity, and signal fading, neither
approach is suitable for PD detection because it is
possible that all channels of the multiwavelength ap-
proach are not at the Q point and that neither method
may have enough frequency response for PD detec-
tion.
Low-coherence light sources, such as LEDs and
surface LEDs 共SLEDs兲, have found increasing appli-
cations in fiber optic interferometric systems because
of their advantages over laser sources in terms of
increased unambiguous measurement range, insen-
sitivity to environmental perturbations, high resolu-
tion, and large dynamic range.
23
The treatment of a
broadband source basically goes back to principles
related by Born and Wolf.
24
By use of split-
spectrum demodulation,
14,25,26
EFPI sensors with
broadband sources have been used for the measure-
ment of pressure, temperature, vibration, and acous-
tic waves. However, even this approach does not
help much to maintain the Q point in the center of the
linear region of a diaphragm-based F-P sensor sys-
tem.
A promising technique for controlling the Q point is
use of a tunable-bandpass filter for dynamic control of
the operating point. This approach takes full ad-
vantage of the broadband source and may also com-
pensate for the slow Q-point drift caused by initial
offset, static oil pressure, or temperature effects.
Because it is difficult to find any commercial tunable
bandpass filters at 1300 nm, the CPT-VT is develop-
ing two techniques for achieving tunable bandpass
filters. Although some positive initial results have
been obtained with both techniques, it will still be
some time before they can be used in our PD detection
system. In the meantime it is beneficial for us to
investigate the feasibility of the developing tech-
niques for Q-point control by studying PD detection
with a broadband source and a nontunable bandpass
filter. In this paper we describe and analyze the
performance of a PD detection system with a low-
coherence source and a bandpass filter.
3242 APPLIED OPTICS 兾 Vol. 42, No. 16 兾 1 June 2003
2. Method of Operation
The fiber optic acoustic sensor is illustrated schemat-
ically in Fig. 1. The system consists of a sensor
probe, a SLED, a bandpass filter, a low-noise wide-
band optical receiver, and single-mode fibers linking
the sensor probe and the optical receiver. The light
from a 1300-nm high-power superluminescent LED
is launched into an isolator and propagates along the
fiber to the sensor head through a filter and a 2 ⫻ 2
3-dB coupler. A 1310-nm optical isolator was in-
serted just after the SLED to reduce optical feedback
to the source, and an optic bandpass filter confined
the spectrum within a certain range. As shown in
the enlarged view of the sensor head 共Fig. 1, inset兲,
the lead in–out fiber and the thin silica glass dia-
phragm are bonded together in a cylindrical fused-
silica tube to form a F-P sensing element. The use of
fused silica for all parts ensures that the temperature
effects will be minimum. The incident light is first
partially reflected 共R
1
兲 at the end face of the lead
in–out fiber. The remainder of the light propagates
across the air gap to the inner surface of the dia-
phragm, where it is once again partially reflected
共R
2
兲. The diaphragm etalon formed by the two sur-
faces of the diaphragm had to be carefully treated in
the previous PD sensor system described earlier in
this paper. Three reasons make it negligible in the
new system that uses a SLED source and a bandpass:
共1兲 The coherence length of the light is much shorter
than that of the distributed-feedback laser, 共2兲 cou-
pling loss causes low fringe visibility and intensity;
and 共3兲 the refractive index of the environment is
close to that of the diaphragm 共⬃1.48 for transformer
oil and ⬃1.33 for water兲. The multiple reflections
travel back along the same lead-in fiber and through
the same fiber coupler to the optical receiver.
The spectral distribution of a SLED may be de-
scribed by a Gaussian function
23
:
f 共兲 ⫽
1
共2兲
1兾2
⌬
exp关⫺共 ⫺
0
兲
2
兾共2⌬
2
兲兴, (1)
where f 共兲 is the spectral density,
0
is the central
wavelength, and ⌬⫽⌬
FWHM
兾共8ln2兲
1兾2
. Assum-
ing that the fiber end and the diaphragm have the
same reflectance 共R
1
⫽ R
2
⫽ R兲, the optical intensity
I
r
共兲 received by the receiver at wavelength is
I
r
共兲 ⫽ 关␣I
0
f 共兲兴
R
2
关共1 ⫹ 兲 ⫺ 2
1兾2
cos共4n
0
L兾兲兴
1 ⫹ R
2
⫺ 2
1兾2
R cos共4n
0
L兾兲
,
(2)
where I
0
is the SLED output power after the pigtail,
␣ is the loss coefficient of the optical path 共from the
source to the sensor and from the sensor to the re-
ceiver, including the fiber, the isolator, the filter, and
the couplers兲, L is the air-gap length, n
0
⫽ 1 is the
refractive index of the air in the cavity, and is the
coupling efficiency for the round trip between the
fiber end and the diaphragm. Provided that the di-
aphragm is parallel to the fiber end, can be calcu-
lated by
27
⫽ 1兾兵1 ⫹ 关2
0
L兾共2n
0
w
2
兲兴
2
其, (3)
where w is the mode spot size of the single-mode fiber.
Therefore the total optical power received by the re-
ceiver can be expressed as
I ⫽
兰
0
⫺BW兾2
0
⫹BW兾2
I
r
共兲d
⫽
␣I
0
共2兲
1兾2
⌬
兰
0
⫺BW兾2
0
⫹BW兾2
⫻
R
2
关共1 ⫹ 兲 ⫺ 2
1兾2
cos共4n
0
L兾兲兴
1 ⫹ R
2
⫺ 2
1兾2
R cos共4n
0
L兾兲
⫻ exp关⫺共 ⫺
0
兲
2
兾共2⌬
2
兲兴d, (4)
where BW is the total spectrum width of the interro-
gating light wave. The SLED used is
SLED1300D20A from Opto Speed, which has a cen-
ter wavelength of 1300 nm and ⌬
FWHM
of 35 nm.
Its spectrum spreads from 1200 to 1400 nm, as shown
in Fig. 2.
From Eq. 共4兲, the interference of the multiple re-
flections produces almost sinusoidal intensity varia-
tions for a low-finesse F-P interferometer, defined as
interference fringes, as the air gap is continually
changed by acoustic wave pressure. Figure 3 gives
the theoretical results from Eq. 共4兲 normalized to the
maximal optical intensity reflected from the F-P cav-
ity, showing the change in air cavity length for a
system without 关Fig. 3共a兲兴 and with 关Fig. 3共b兲兴 a band-
pass filter, where the reflective surfaces are not
coated 共R ⫽ 3.5%兲. Obviously, the sensor with the
natural spectrum of the SLED has a short coherent
length and low visibility 共⬃25% for an air-gap length
of 15 m兲. These limitations require that the air
cavity be shorter than 10 m to produce higher fringe
visibility, which places more difficult manufacturing
constraints on the sensor head. In contrast, the in-
terference fringes have very good visibility 共⬃65% at
a cavity length of 15 m兲 if ⌬
FWHM
is reduced to 20
nm by use of coarse wavelength-division multiplexing
a bandpass filter. This use of a badnpass filter also
Fig. 1. Schematic of a fiber Fabry–Perot acoustic sensor system:
SMF, single-mode fiber; AR, antireflection.
1 June 2003 兾 Vol. 42, No. 16 兾 APPLIED OPTICS 3243
significantly increases the fabrication tolerance at
the expense of out-of-band optical power 共⬃50% for
the SLED1300D20A兲, which is acceptable in our sys-
tem’s power budget.
From Fig. 3, one period of fringe variation corre-
sponds to an air-gap change of one half of the optical
wavelength, which in our case is ⬃0.65 m. In prin-
ciple, continuous tracking of phase changes in the
interference fringes can yield information about air-
gap changes in the sensor element. The acoustic
signal generated by partial discharges causes deflec-
tion of the diaphragm and modulates the sealed air-
gap length. The sensor therefore yields outputs that
correspond to the applied acoustic signals. As in
regular interferometers, the measurement will have
ultrahigh sensitivity. However, the measurement
will suffer from the disadvantages of sensitivity re-
duction and ambiguity in fringe direction when the
sensor reaches peaks or valleys of the fringes. Sen-
sitivity is reduced at peaks or valleys of the fringes
because at those points the change in optical inten-
sity is nearly zero for a small air-gap change. Am-
biguity in fringe direction is defined as difficulty in
determining whether the air gap is increasing or de-
creasing by detecting the optical intensity only. If a
measurement starts with an air gap corresponding to
the peak of a fringe, the optical intensity will de-
crease, regardless of whether the gap increases or
decreases.
One approach to solving these problems is to design
the sensor head such that the maximum air-gap
change does not exceed the linear region of a half-
fringe. Figure 4 shows the interference fringes of
sensor heads normalized to the optical intensity in
the F-P cavity at reflectances R ⫽ 3.5%, 10%, 20% for
changes in air-gap length from 14.8 to 15.6 m.
Choosing as the operating point 共Q point or initial
air-gap length兲 the central point 共L
0
兲 between a peak
Fig. 2. Spontaneous-emission spectrum of SLED1300D20A.
Fig. 3. Interference fringes of a low-finesse sensor 共reflectivity R ⫽ 3.5%兲 with a SLED source: 共a兲 no filter, ⌬
FWHM
⫽ 35 nm; 共b兲 with
a bandpass filter, ⌬
FWHM
⫽ 20 nm.
Fig. 4. Illustration of the linear operating range of the sensor’s
response curve; R is the reflectivity.
3244 APPLIED OPTICS 兾 Vol. 42, No. 16 兾 1 June 2003
and its neighbor valley for each sensor head, we may
treat a region defined by 共L
0
⫾⌬L兲共between the
dotted–dashed vertical lines in Fig. 4兲 as a linear
region, which can be fitted with Matlab by the follow-
ing linear equation:
I共L兲 ⫽ S
0
L ⫹ C, (5)
where I共L兲 is the optical intensity at air gap L and S
0
and C are constants. Obviously S
0
represents the
fringe sensitivity of a sensor to a change in the air gap
near its Q point. For a norm of residuals of 10
⫺3
, the
fitting parameters were calculated and are listed in
Table 1. The increase in reflectivity R from 3.5% to
20% results in an increase in sensitivity of a factor of
4.27, but the decrease in linear operation ranges from
96 to 62 nm. In addition, the Q point of the linear
region drifts toward the valley of the fringes when the
reflectance is increased. However, visibility and ab-
solute amplitude increase with increased reflectivity.
In this linear region the ac electrical output of the
sensor V共P兲, is proportional to the air-gap change
y
0
共P兲 that is caused by acoustic pressure P and can be
expressed as
V共P兲 ⫽ Gᑬ␣S
0
I
ac
关L
0
⫹ y
0
共P兲兴 ⫽ S
oe
y
0
共P兲, (6)
where S
0
is the fringe sensitivity, i.e., the slope of the
fringes in the linear regions; ᑬ is the responsivity of
the InGaAs photodetector, ⬃0.9 at 1300 nm; G is the
total gain of the optical receiver; I
ac
关L
0
⫹ y
0
共P兲兴 is the
ac component of the received optical signal; y
0
共P兲 is
the diaphragm deflection caused by acoustic pressure
P; S
oe
is the total optical and electrical sensitivity
with respect to the air-gap change; and
S
oe
⫽ Gᑬ␣S
0
. (7)
3. Diaphragm-Design Considerations
Diaphragms and membranes have found extensive
applications in pressure and acoustic wave measure-
ments in the mechanical and microelectromechanical
system sensor industries. As shown in the inset of
Fig. 1, we fabricated a sensor head by thermally
bonding a fiber, a ferrule, a silica tube, and a silica
diaphragm together to form a sealed F-P interferom-
eter. The diaphragm vibrates in the presence of an
acoustic wave, which imposes a dynamic pressure on
it. The diaphragm’s design is probably the most
critical part of the sensor for efficient acoustic wave
detection. Sensitivity, linearity, frequency re-
sponse, temperature dependence, and size of the sen-
sor head are directly related to the properties of the
diaphragm. For the sake of extremely low depen-
dence on temperature, we selected fused silica, the
same material as used in the single-mode optical fi-
ber, as the material to be used for the ferrule, the
tube, and the diaphragm, as mentioned above.
However, the shape and the size of the diaphragm are
yet to be determined by the sensitivity and band-
width requirements of the system.
A. Sensitivity
The diaphragm will be deflected whenever there is a
differential pressure P between the inside and the
outside of the sealed cavity. The out-of-plane deflec-
tion of the diaphragm, y, is a function of the pressure
difference at any radial position, r. With the as-
sumption of uniform diaphragm thickness, small de-
flection, infinitely rigid clamping about the periphery
of the diaphragm, and perfectly elastic behavior,
which is almost true for the fused-silica diaphragm
and our bonding method, the deflection under pres-
sure difference P is can be expressed as
28
y共P兲 ⫽
3共1 ⫺
2
兲
16Eh
3
共a
2
⫺ r
2
兲
2
P, (8)
where is Poisson’s ratio 共 ⫽ 0.16 for fused-silica
glass 7940 at 25 °C兲, E is Young’s modulus of the
silica glass material 共E ⫽ 7.24 ⫻ 10
10
Pa or 73.5 ⫻ 10
8
kg兾m
2
at 25 °C兲, a is the radius of the diaphragm
defined by the inner diameter of the glass tube, and h
is the thickness of the diaphragm. In our sensor
configuration the fiber is positioned at the central
part of the diaphragm such that only the center de-
flection y
0
is of interest. The sensitivity of the dia-
phragm of a fused-silica diaphragm is given by
␦ ⫽
y
0
共P兲
P
⫽ 2.524 ⫻ 10
⫺6
a
4
h
3
关nm兾kPa兴, (9)
where y
0
is given in nanometers, a and h, in microme-
ters; and P, in kilopascals. Figure 5 shows a typical
Fig. 5. Sensor’s sensitivity and natural frequency versus dia-
phragm thickness for a ⫽ 1.25 mm.
Table 1. Fitting Parameters for Linear Regions with a Norm
of Residuals of 10
ⴚ3
Reflectivity,
R 共%兲 S
0
C
Q Point
L
0
共m兲
Linear Range,
2⌬L 共nm兲
3.5 0.3777 ⫺5.6425 15.108 96
10 0.96535 ⫺14.411 15.100 72
20 1.6124 ⫺24.036 15.092 62
1 June 2003 兾 Vol. 42, No. 16 兾 APPLIED OPTICS 3245