2010 IEEE International Energy Conference

Wind Turbines Condition Monitoring and Fault

Diagnosis Using Generator Current Amplitude
Demodulation
Y. Amirat, V. Choqueuse and M.E.H Benbouzid
University of Brest, EA 4325 LBMS
Rue de Kergoat, CS 93837, 29238 Brest Cedex 03, France
[email protected], [email protected] and [email protected]

Abstract—Wind energy conversion systems have become a real wind turbine failure data has shown important features of
focal point in the research of renewable energy sources. In order failure rate values and trends [1]. A failures number
to make wind turbines as competitive as the classical electric distribution check-off is reported in Figs. 1 and 2 for
power stations, it is important to reduce the operational and Swedish, Danish and German wind power plants that
maintenance costs. The most efficient way of reducing these
occurred between 1994 and 2004 [1].
costs would be to continuously monitor the condition of these
systems. This allows for early detection of the degradation of the
generator health, facilitating a proactive response, minimizing
downtime, and maximizing productivity. This paper provides
then an approach based on the generator stator current data
collection and attempts to highlight the use of Hilbert
transformation for failure detection in a Doubly-Fed Induction
Generator (DFIG) based wind turbine for stationary and
nonstationary cases.

Index Terms—Wind turbine, DFIG, fault detection,

diagnosis, amplitude modulation, Hilbert transform.

I. INTRODUCTION
Wind energy conversion systems are the fastest-growing
source of new electric generation in the world and it is
expected to remain so for some time. Classical power Fig. 1. Failures number distribution
generation systems, which are facing to constantly changing for Swedish wind power plants (2000-2004) [1].
operating parameters, such as fuel cost, multiple fuel
tradeoffs and maintaining older system, becomes more costly
and challenging with obsolescence of key components.
DFIG-based wind turbines offer an alternative and emerging
solution but due to geographical location of wind turbines Failures / Turbine x Year

particularly in the off-shore wind farms, it is important to

prevent failure and to reduce maintenance cost. A deep
knowledge about all the phenomena involved during the
occurrence of a failure constitutes an essential background for
the development of any failure diagnosis system. Regarding a
failure as a particular input acting on the generator, a
diagnosis system must be able to detect its occurrence, as
well as to isolate it from all other inputs such as disturbances
and controls affecting the behavior of the DFIG.
For the failure detection problem, it is important to know
if a failure exists or not in the generator via the processing of
available measurements. For the failure diagnosis problem, in
addition, to identify the element of the system which has
failed and to find the failure causes. A quantitative analysis of Fig. 2. Failure rates for Danish and German wind power plants [1].

978-1-4244-9380-7/10/$26.00 ©2010 IEEE 310

These figures show that approximately 45 % of failures were approach for data collection for real wind turbine. Section IV
linked to the electrical system, sensors and blades/pitch describes the data lab facility used to validate the proposed
components. The experience feedback of wind turbine approach. Finally, the performance of the method is reported
industries states that the major concern is on the electrical in section V.
system. Typical failures include: dynamic air gap
irregularities, generator bearing failure, stator and rotor
winding; insulation failures, inter-turn short circuits in stator II. SIGNAL PROCESSING TECHNIQUE
windings, broken rotor bar or cracked rotor end-rings and For amplitude modulated signals, the received current i(n)
harmonic derating. can be expressed as
Many techniques and tools are available for the condition
monitoring of DFIG-based wind turbines in order to extend
i (n) = a (n) cos(2πnf / Fe + φ) (1)
their life span. Some of the technology used for monitoring
includes pre-installed sensors, which may measure speed,
output torque, vibrations, temperature, flux densities, etc. where n = 0, ..., N – 1 is the sample index, N is the number of
These sensors are managed together in different architectures received samples, φ is the phase parameter and Fe is the
and coupled with algorithms to allow an efficient monitoring sampling frequency. The amplitude a(n) in (1) depends on the
of the system condition. Those methods are inspired from failure hypothesis: for a healthy generator, a(n) is constant,
electric motor condition monitoring. From the theoretical and and for faulty generator, it varies with time (Amplitude
experimental point of view, the well-established methods are: Modulation).
electrical quantities signature analysis (current, power...), For failure detection, a possible approach relies on the use
vibration monitoring, temperature monitoring and oil of an amplitude demodulation technique to estimate a(n).
monitoring. In the case of DFIG-based wind turbines, it has Then, statistical features can be extracted to detect if it is
been shown that failure in the drive train could be diagnosed time-varying or not.
from the electrical quantities of the generator [2]. This
principle has been used to diagnose unbalance and failure in A. Amplitude Demodulation
the blades of a small wind turbine by measuring the power
spectrum density at the turbine generator terminal [3]. The Popular amplitude demodulation techniques include
advantage of signature analysis of the generator electrical Hilbert Transform (HT) [8] and Teager energy operator [9].
quantities is that those quantities are easily extractible during Furthermore for three-phase system, it has been recently
operation i.e. the current can be acquired by current shown that the Concordia transform can be employed to
transformer, the voltage via a voltage transformer and the perform demodulation [10]. In this study, one phase current is
power by computation. Moreover, current and voltage considered. In this context, the Hilbert transform is chosen to
transducers are usually cheaper than vibration and torque estimate the envelope a(n) since it is usually more robust
transducers. Analysis of the generator electrical quantities against noise than the Teager energy operator.
usually involves the use of signal processing techniques. Let us consider a discrete sequence i(n). The Discrete
For steady state operations, the Fast Fourier Transform Hilbert Transform (DHT) of i(n) is given by [8].
(FFT) is the most popular algorithm; however it has a lack to
unfulfill the nonstationary quantities. In the case of variable Η [i (n)] = F −1 {F {i (n )} ⋅ u (n )} (2)
speed DFIG-based wind turbines, FFT is difficult to interpret
since the operation is predominately nonstationary due the -
where F{.} and F 1{.} correspond to the Fast Fourier
stochastic behavior of the wind speed. To overcome this
Transform (FFT) and Inverse FFT (IFFT), respectively, and
problem, electric machine conditions monitoring and failure
where u(n) is defined as
diagnosis procedures based on time-frequency representations
(Spectrogram, Quadratic TFR, etc...) or time-scale analysis
(wavelet) have been proposed in the literature of the electric ⎧ N
⎪1, n = 0, 2
machines community [4-6]. Nevertheless, theses techniques ⎪
have drawbacks such as high complexity, poor resolution ⎪ N
u n ⎨2, n = 1, 2,..... − 1
( ) = (3)
and/or may suffer from artifacts (cross-terms, etc.). ⎪ 2
This paper presents then a less complex failure detector ⎪ N
for DFIG-based wind turbines which is appropriate for ⎪0, n = 2 − 1,..., N − 1
⎩
nonstationary operations and transient behavior. It focuses on
mechanical failures that lead to stator current Amplitude
Modulation (AM). These include, for example, air gap Using (1), the estimated envelope, denoted â(n), is given
eccentricity, bearing wear and failure [7]. It is organized as by [8].
follows. Section II describes the signal processing technique
used for failure detection. Section III presents the adopted aˆ(n) = i 2 (n) + (H [ik (n)]) 2 (4)

311
B. Failure detector typically built in onshore or offshore wind farm
After applying Hilbert transform, we propose to determine configurations; there is a need for building up networks. The
if â(n) is time-varying or not. Let us compute the variance of proposed architecture is based on an industrial PC (EPC)
which is dedicated to collect data from the DFIG-based wind
â(n), denoted σ2, which is defined by
turbines via the extended I/O modules and transfers the data
N −1 to users through LAN network. The EPC is configured to
1
σ2 =
N
∑ (aˆ (n) − μ) 2
(5) transmit data in asynchronous mode such that all the data are
n=0 stored (buffered) in specific data blocs and no data are lost
during the processing. This allows investigation of data for
where μ is â(n) mean, i.e. further purposes. The EPC has also the task for managing
alarm and emergency shut down procedure. Figure 3 depicts
1 N −1 the data collection approach for a real wind turbine via an
μ=
N
∑ aˆ(n)
n=0
(6) industrial data bus. The data flow scheme is shown by Fig. 4.
This architecture incorporates a SCADA system and a
CMS system, where the SCADA system is devoted to
As â(n) is constant for healthy generator, it follows that μ
supervise and control the DFIG-based wind turbines through
= â(n) and then σ2 = 0. For faulty generator, the envelope the EPC. For laboratory experiments the wind turbine is often
â(n) is time-varying which implies that μ ≠ â(n) and then σ2 > replaced by a DC or an AC motor that acts as a prime mover.
0. These two properties lead us to propose a simple
hypothesis test for failure detection based on σ2:
– If σ2 < γ, the generator is stated healthy.
– If σ2 > γ, the generator is stated faulty. IV. TEST FACILITY DESCRIPTION
Where γ is a threshold which can be set subjectively
Figure 5 describes the experimental setup that is operated
depending on a false alarm probability.
in the motor configuration for experimental easiness. It is
composed of two parts: a mechanical part that has a tacho-
generator, a three-phase induction motor and an alternator.
III. DATA COLLECTION FOR REAL WIND TURBINE
The tacho-generator is a DC machine that generates 90 V at
Today, most turbines are fitted with equipment that makes 3000 rpm. It is used to measure the speed. It produces linear
it possible to collect condition monitoring data remotely via voltage between 2500 and 3000 rpm.
modem or internet. Moreover, since wind turbines are

Fig. 3. Test rig configuration.

312
 the fault detection process. For these reasons, the induction is
fed by an alternator. By this way, supply harmonics effects
are eliminated and only bearing failure effects could be
observed on the stator current. Figure 6 is then given to
illustrate the experimental test philosophy.
The tested induction motor has the following rated
parameters: 0.75 kW, 220/380 V, 1.95/3.4 A, 2780 rpm, 50
Hz, 2 poles, Y-connected. It has two 6204.2ZR type bearings.
Fig. 4. Data flow between DFIG-based wind turbines and embedded PC. From the bearing data sheet the following parameters are
obtained: The outside diameter is 47 mm and inside one is 20
The alternator is a three-phase synchronous machine with a mm. Assuming that the inner and the outer races have the
regulator and a rectifier circuit that stabilize the output same thickness gives the pitch diameter DP = 31.85 mm. The
voltage at 12 VDC. The advantage of using a car alternator bearing has eight balls (N = 8) with an approximate diameter
instead of DC generator is obtaining constant output voltage of DB = 12 mm and a contact angle θ = 0°. These bearings are
at various speeds. The induction motor could be identically made to fail by drilling holes of various radiuses with a
loaded at different speeds. Moreover, if the induction motor is diamond twist bit while controlling temperature by oil
supplied from the network, motor current will have time and circulation in experiments. Some of the artificially
space harmonic components as well as bearing fault sourced deteriorated bearings are shown in Figure 7.
harmonics. This makes it harder to determine the bearing
failure effect on the stator current and therefore complicates

Fig. 5. Experimental setup [11].

R
S
T

Tacho
DC Senkron DC Load Generator
Motor Generator Generator
Data Acq. Card

Induction Motor

Stator Current Current AM

PC

Fig. 6. Test facility [11].

313
 Fig. 7. Artificially deteriorated bearings: (a) outer race deterioration, (b) inner race deterioration, (c) cage deterioration, (d) ball deterioration [11].

V. FAILURE DETECTOR TEST

In this section, the result of the proposed approach is
presented with experimental signals corresponding to fault (a)
(Fig. 7).
After Hilbert transform, 10 samples have been removed at
the beginning and at the end of â(n) to avoid the edge effects
problem of Hilbert transform. Figures 8 and 9 display the
stator current i(n) and the envelope â(n), respectively, for a
healthy generator. As the system is not perfect, one could
note small variations on the envelope â(n). In the presence of
a bearing failure, the stator current and the envelope are
shown in Figs. 10 and 11, respectively. Compared to the
healthy case, stronger oscillations of â(n) can be observed.
Table I reports the value of σ2 for the faulty and healthy
generators. As previously discussed, σ2 is not strictly equal to
0 even if the generator is healthy (σ2 = 0.012). However
when a bearing failure occurs, this criteria is multiply by
Fig. 8. Stator current i(n) of a healthy generator.
4.333. In this condition, a failure can be detected by setting
the hypothesis-test threshold to σ2 = 0.032.

VI. CONCLUSION
This paper has presented an implementation of a low-
complexity signal processing technique for bearing faults
detection in DFIG-based wind turbines. It was found that the
proposed technique gives a significant criterion for failure
detection.

Table 1. Non-balanced system:

Fault detector for healthy and faulty generator.

Demodulation Healthy case Faulty case

Hilbert Transform σ2 = 0.012 σ2 = 0.052

Fig. 9. Envelope â(n) of a healthy generator.

REFERENCES
[1] Y. Amirat, M. Benbouzid and E. Al-Ahmar, “A brief status on condition [3] W.Q. Jeffries, J.A. Chambers, and D.G. Infield, “Experience with
monitoring and fault diagnosis in wind energy conversion systems,” bicoherence of electrical power for condition monitoring of wind
Renewable & Sustainable Energy Reviews, vol. 3, n°9, pp. 2629-2636, turbine blades,” IEE Proc Vision, Image & Signal Processing, vol. 145,
December 2009. n°3, pp. 141-148, June 1998.
[2] E. Al-Ahmar, M.E.H. Benbouzid and Y. Amirat, “DFIG-based wind [4] S. Rajagopalan, J. A. Restrepo, J. Aller, T. Habetler, and R. Harley,
turbine fault diagnosis using a specific discrete wavelet transform,” in “Non stationary motor fault detection using recent quadratic time
Proceedings of ICEM'08 (International Conference on Electrical frequency representations,” IEEE Trans. Industry Applications, vol. 44,
Machines), Vilamoura (Portugal), September 2008. n°3, pp. 735-744, May/June2008.

314
 Fig. 10. Stator current i(n) of a faulty generator. Fig. 11. Envelope â(n) of a faulty generator.

[8] A. Oppenheim, R. Schafer and W. Padgett. Discrete-Time Signal

[5] M. Blodt, D. Bonacci, J. Regnier, M. Chabert, and J. Faucher, “Online Processing, 3rd Ed. Prentice Hall, 2009.
monitoring of mechanical faults in variable-speed induction motor [9] P. Maragos, J. Kaiser, and T. Quartieri, “On amplitude and frequency
drives using the Wigner distribution,” IEEE Trans. Industrial demodulation using energy operators,” IEEE Trans. Signal Processing,
Electronics, vol. 55, n°2, pp., 522-533, February 2008. vol. 41, n°4, pp. 1532-1550, April 1993.
[6] M. Riera-Guasp, J. Antonio-Daviu, J. Roger-Folch, and M. P. M. [10] B. Trajin, M. Chabert, J. Regnier, and J. Faucher, “Hilbert versus
Palomares, “The use of the wavelet approximation signal as a tool for Concordia transform for three phase machine stator current time-
the diagnosis of rotor bar failure,” IEEE Trans. Industry Applications, frequency monitoring,” Mechanical Systems & Signal Processing, vol.
vol. 44, n°3, pp. 716-726, May/June2008. 23, pp. 2648–2657, 2009.
[7] M. Blodt, J. Regnier, and J. Faucher, “Distinguishing load torque [11] I. Önel and M. Benbouzid, “Induction motor bearing failure detection
oscillations and eccentricity faults in induction motors using stator and diagnosis: Park and Concordia transform approaches comparative
current Wigner distributions,” IEEE Trans. Industry Applications, vol. study,” IEEE/ASME Trans. Mechatronics, vol. 13, n°2, pp. 257–262,
45, n°6, pp. 1991-2000, November/December 2009. April 2008.

315