Wavelet and Hilbert Huang Transforms Applied to

Park’s Vector for Fault Detection in a PMSG Wind

Turbine
1st Raúl Arturo Ortiz-Medina 2nd Mauricio Sanabria-Villamizar 3rd Irvin Lopez-Garcia
Dept. of Postgraduate and Research Dept. of Energy Dept. of Energy
Universidad Politécnica de Aguascalientes Universidad Autónoma Metropolitana Universidad Autónoma Metropolitana
Aguascalientes, México Ciudad de México, México Ciudad de México, México
[email protected] [email protected] [email protected]

4th Francisco Javier Villalobos-Piña 5th Francisco Beltran-Carbajal 6th Victor Arturo Maldonado-Ruelas
Dept. of Electronic Engineering Dept. of Energy Dept. of Postgraduate and Research
Instituto Tecnológico de Aguascalientes Universidad Autónoma Metropolitana Universidad Politécnica de Aguascalientes
Aguascalientes, México Ciudad de México, México Aguascalientes, México
[email protected] [email protected] [email protected]

Abstract—This article is a novel technique for the maintenance [7] or through magnetic field sensors in [8], but most of
of machines, specifically in wind turbines. The technique proposes the methodologies are applied to the current signal, as those
a fault detection method, identifying the pattern frequencies of that work with artificial intelligence based on deep learning
the disturbance mode. This methodology is carried out with the
Discrete Fourier Transform, Wavelet Transform and the Hilbert- [9], artificial neural networks [10] or SCADA data analysis
Huang Transform, making the comparison in the described case. with convolutional neural network and short term memory
On the other hand, the methodology shows us the frequencies network [5]. The Kalman filter is also used for fault detection
detected by each technique and the differences between them. in its extended and unscented forms in order to compare both
Index Terms—Oscillation, fault, detection, reliability techniques [11].
The interturn faults of the windings are common in PMSG
I. I NTRODUCTION and those are very important to detect in early stages because
Since wind turbines began to have greater integration into when they occur, the overcurret protections are not triggered
power systems one aspect of wind turbines that has gained until the fault becomes great enough to trigger a protection
importance in industry and research is maintenance through that stop the system. Interturn faults detection is given natu-
signal analysis. Wind turbines are exposed to faults in their rally through the current signal analysis because the signal
components that may be due to the climatic conditions of is directly modified because of the change in the winding
the installation site, the heating produced by the currents that parameters [12]. The fault detection through current signal
circulate in the electrical and electronic elements and the wear analysis in PMSG used in wind turbines has been studied and
and tear of the mechanical elements due to normal operation of developed with different detection tools in different studies,
the wind turbines. Many efforts have been made to detect faults that goes from an analysis and comparison of the principal
in these systems earlier and more efficiently. In [1], faults in electromechanical parameters current, speed and torque in [13]
the different subsystems of wind turbines are discussed, where or those that uses specialized techniques as [14], where the
it is indicated that the most common are mechanical failures in three-phase signal and the angular velocity is used for estimate
gear systems and electrical faults in the generator. The causes interturn shortcircuit faults in PMSG through attention based
of faults, their consequences and how to detect them in early recurrent neural network that estimate a fault indicator for
stages are studied in order to provide maintenance in a timely obtainig the existence and severity of the fault.
and scheduled manner. One of the most widely used wind The structure of the paper have a brief introduction in
turbine configurations is the one that uses a permanent magnet present section I. The section II contains information about
synchronous generator (PMSG) as an energy converter [2]. the discrete Fourier transform (DFT), the discrete Wavelet
In PMSG wind turbines there are investigations of faults transform (DWT) and the Hilbert-Huang transform (HHT) that
in specific parts, like the converter [3], the drive train [4], are used as signal analysis tools applied to the Park’s vector
gearbox [5], or the generator [6]. Also for PMSG wind also known as instantaneous spatial phasor (ISP). In section
turbines, the investigations for fault detection made have III the test bench and its parts is described as well the way
different methodologies as it is the parameter estimation in the PMSG was altered for the shortcircuit tests. The section
IV contains the results of the application of the fault detection B. Fourier transform
tools to the PMSG at 200 rpm with different severities of In DFT, the frequency range of discrete-time signals is
incipient fault. Finally, the conclusion for this investigation limited to the interval (−π, π) or (0, 2π). A discrete signal
work are given as section V. of fundamental period N can have components of separate
frequencies 2n/N radians at a frequency f = 1/N cycles.
II. FAULT DETECTION TOOLS The form of the FTD is described in (3).
Fault detection is based on signal analysis techniques that N −1
are used to compare the behavior of the healthy current signal 1 X
DF T (fn ) = xk e−i2πfn k∆T (3)
of the stator of electrical machines of small wind turbine N
k=0
systems with the current signal of the same faulty electrical
machine. The analyzed signals correspond to the stator cur- Where N = T /∆T is the number of samples, and fn =
rents, obtained through non-invasive sensors or sensors that are 0, 1, 2, ..., N − 1 are the discrete frequency components. To
an inherent part of the wind system. Four tools are used, the process a DFT of a signal with N samples, N2 arithmetic
ISP or Park’s vector as a pre-analysis tool for current signals, operations need to be performed to be completed. This causes
the DFT, DWT and HHT. the computation time of the DFT to increase rapidly with an
increase in the number of samples. The solution to the high
A. Park’s vector computational cost of DFT lies in an algorithm that recursively
divides the DFT known as fast Fourier transform or FFT that
The Park vector, also called instantaneous spatial phasor or reduces arithmetic operations to a total of N log(N ) [17].
ISP, is a tool used in the theory of electrical machines [15].
It is based on the magnetization currents corresponding to a C. Wavelet transform
balanced three-phase system with angular velocity ω = 2πf ,
The transform depends on a wavelet function that is defined
where f is the electric frequency.
as the square of an integrable function Ψ(t) that satisfies
A three-phase system, regardless of perturbation, satisfies
the admissibility condition which implies that the Fourier
Fortescue’s theorem, [16], which states that an unbalanced
transform of that function Ψ(t) vanishes at the zero frequency
three-phase signal can be generated from the positive, negative,
value [18]. One way to define the wavelet function is through
and zero sequence components. The ISP groups the currents
its processes, s-scale-dependent expansion and t-parameter
and represents them in a single equation, either in the abc
dependent translation. The wavelet function in these terms is
reference frame or by means of the sequence components
shown in (4)
mentioned. A direct way to determine the ISP in three-phase  
currents of electrical machines is the translation of the three- 1 t−τ
ψs,τ (t) = √ ψ , s > 0, τ ∈ R (4)
phase currents in reference frame abc to the fixed reference s s
frame with respect to the stator of the machine, known as
αβ. The change of reference frame is made by means of the based on the definition of the wavelet function, the continuous
Concordia transform shown in (1). wavelet transform of a signal x(t) is now defined by (5)
Z ∞  
1 ∗ t−τ
wt(s, τ ) = √
    
−2π

2π

x(t)ψ dt (5)
 iα   1 cos 3 cos 3   ia  s −∞ s
    
  1
=  −2π

2π

   where ψ ∗ (·) is the complex conjugate of the wavelet function
 iβ  2  1 −sen 3 −sen 3   ib 
  
     ψ(·) scaled and transferred.
     To discretize the scale s and link it to the steps of the values
i0 √1 √1 √1 ic
2 2 2 of the translation parameter τ , logarithmic discretization is
(1) used with s = sj0 and τ = kτ0 sj0 when s0 < 1, τ0 ̸= 0 and
The translation to the reference frame αβ of the currents j, k ∈ Z where Z is the set of integers.
is a direct way to obtain the representation of the current Normally in the discrete form of the wavelet the values of
in the components of positive and negative sequence, so the s0 = 2 and τ0 = 1 are adopted, which correspond to the
modulus of the currents in αβ of a three-phase signal is dyadic discretization and the wavelet takes the form shown in
equivalent to the module of the ISP in components of positive (6)
and negative sequence. Thus, the ISP can be written in terms t − k2j
 
1
of the reference frame αβ, as shown in (2) [15]. ψj,k (t) = √ ψ (6)
2j 2j
˜ 2 =iα 2 + iβ 2
|I| Then, the DWT is represented in (7)
or (2)
∞
t − k2j
Z  
q 1
ISP = iα 2 + iβ 2 wt(j, k) = ⟨x(t), ψj,k (t)⟩ = √ x(t)ψ dt
2j −∞ 2j
. (7)
where ⟨x(t), ψj,k (t)⟩ is the convolution of the data function Input signal to
analyze x(t)
and wavelet. DWT can be understood as a timescale represen-
tation of a digital signal that is obtained using digital filtering Local extrema of x(t)
techniques that is equivalent to performing the convolution of
the signal with the impulse response of certain filters [17]. Envelope of x(t)

D. Hilbert-Huang transform Average of envelope m(t)

The HHT consists of two methods: 1) Empirical mode
decomposition (EMD) and 2) Hilbert transform (HT). The h(t) = x(t) − m(t)

first method has the purpose of deconstructing the signal

into a set of functions that are easier to analyze and the
second method allows us to extract information from the No h(t) is to satisfy
two conditions
deconstruction carried out with the previous method, to carry
out a time vs frequency analysis [19].
Yes
1) Empirical mode decomposition: The method extracts
IM F (t) = h(t)
the frequency information from non-stationary and non-linear
signals. Sifting is the process of removing the highest fre- Get residual
quency or oscillation of information until only the lowest is r(t) = x(t) − IM F (t)

retained. In order to use the method accurately, the following

two requirements must be met:
The residual is No
• For a signal data set, it is assumed that the number of a monotonic?
zero crossings and extremes varies and must be greater
than or equal to one. Yes
• At any point in the signal, the mean value of the envelope
End
defined (interpolation) by the local maxima and that
defined by the local minima is zero. Fig. 1. Flowchart of EMD.
To obtain the signals that represent the different oscillation
modes (IMFs, intrinsic mode function) of the original signal,
the following steps shown in Figure 1 must be followed. In the instantaneous frequency is defined by (13).
the flowchart shown, all the extreme points (maximum and
minimum) of the original signal are calculated, separating the 1 dθ (t)
f (t) = (13)
signal into a set of IMFs with the separation process. Then, 2π dt
the final envelopes of each oscillation are calculated by cubic
III. E XPERIMENTAL PLATFORM
spline interpolation.
2) Hilbert transform: Each IMF obtained with the EMD A test bench for the detection of short-circuit faults in
method must be subjected to the HT to carry out the time- PMSG stators used in small wind turbines was designed and
frequency analysis. The result is a signal x(t), which after built for this propose.
applying the HT is obtained:
Z ∞ A. Test bench
1 x (s)
y (t) = H (x (t)) = P V ds (8)
π −∞ − s
t The test bench is basically made up of two parts: The
If s = t it is necessary to take the principal value of cauchy first part is a 180 V, 2238 W and 1800 rpm DC motor
to handle said singularity, solving it as follows (Eq (9)): controlled through a power electronics system designed to
Z simulate different wind speeds and its structure is shown in
1 x (s) Fig. 2. The control is based on a DSP that perform the control
H (x (t)) = lim ds (9)
ϵ→∞ π |s−t|>ϵ t − s actions of the power electronic system to drive the DC motor
and allows communication with the real-time resource with a
The analytic signal z(t) is defined as:
personal computer. For speed control, a discrete PI controller
was designed and implemented to control the speed of the DC
z (t) = x (t) + iy (t) = a (t) eiθ(t) (10)
motor, which is also embedded in the DSP program code. The
a(t) and θ(t) are defined in (11) and (12) respectively. second part of the test bench is a three-phase, 12 V, 400 W
p and 400 rpm PMSG, loaded with 130Ω resistors connected in
a (t) = x2 + y 2 (11) Y, with the stator winding altered to cause different levels of
short-circuit failure to interturn. The DC motor and the PMSG
are coupled to simulate the action of a small wind turbine at
 
y (t)
θ (t) = arctan (12) different speeds and with different levels of fault severity.
x (t)
 TABLE I
T ESTS TO BE CARRIED OUT WITH PERCENTAGES OF SHORTCIRCUIT
FAILURE IN THE COIL FOR EACH SPEED

200 rpm 400 rpm

healthy case healthy case
Fault 5% phase a Fault 5% phase a
Fault 20% phase a Fault 20% phase a
Fault 45% phase c Fault 45% phase c

Fig. 2. Structure of the power electronics characteristics suitable for the 1 kHz acquisition of current
signals for detect faults according to the characteristics of the
wind turbine PMSG to be tested. After the acquisition is made,
B. PMSG alteration the analysis of the signal take place. First, concordia transform
The PMSG is altered introducing taps to the stator winding is applied to the three-phase current signal, going from three to
through which intentional shortcircuits are caused. Under two phases and then the module of both signals is obtained,
normal conditions, the stator core of the PMSG has 36 thus obtaining the park vector which is a single signal that
slots, its winding is three-phase, Y-connected, wave double responds to the changes that occur in the three-phase signal.
layer winding and has 12 coils per phase. This three-phase
synchronous permanent magnet machine for application in IV. R ESULTS : S IGNAL ANALYSIS TOOLS APPLIED TO
wind turbines, has the three winding phases with the same FAULTY PMSG
characteristics. The winding has a 19 AWG, with 18 turns
and pitch 1-4, this conditions must be reproduced in the re- With the three-phase currents coming from the stator, the
winding of the PMSG but with taps to cause short circuit ISPs are obtained, whose single signal is analyzed through the
incipient faults. The incipient faults that will be introduced to discrete Fourier, Wavelet and Hilbert-Huang transforms.
the PMSG have approximate fault percentages of 5% and 20%,
corresponding to 4 and 7 short-circuited turns respectively of
a phase a coil. Derivations were also made in phase c with A. Fourier analysis
approximate fault percentages of 5% and 20%, corresponding
to 1 and 4 shortcircuited turns respectively. The three different These changes in the frequency components are located at
and possible faults analyzed are shown in Table I and Fig. 3 twice the operating frequency of the electrical machine when
shows the three-phase circuit by locating the leads in windings there is a short-circuit fault and it is through the DFT that
a and b. the frequency spectrum of the ISPs with and without fault is
obtained. The figures 4, 5 and 6 show the frequency spectra
of the ISP with the PMSG stator under the different failure
scenarios established in the table I for a speed of 200 rpm.
In the ISPs corresponding to 200 rpm, the DFT reveals that
the signals coincide in some frequency components, but there
is especially a significant difference in the value component
ff = 2fs . Since the operating frequency of the PMSG is
20Hz, the frequency component that appears in the presence
of a fault is 40Hz, which is considered as the fault signature
frequency.
It can also be observed that the more severe the fault
is that of 7 turns shortcircuited, the frequency component
that registers the fault also grows. This can be seen in the
magnitude of the 40Hz frequency from the figure 4 with 1
turn shorted in one coil, to the figure 6 representing the 4
shortcircuited turns.
Fig. 3. three-phase circuit showing the windings where 1, 2, 4 and 7 turns
are deliberately shorted The DFT is mainly used to show that the ISP shows the
presence of the PMSG operating frequency doubling compo-
nent increase as the incipient fault worsens. And even though
C. Current signal analysis the phase winding consists of 12 coils, the DFT can reliably
For acquiring the current signals from the PMSG stator to detect a fault from the 4th turn shorted or 20% of the coil
the resistive balanced three-phase fixed load three Honeywell failed. This is not the case with the faults presented with 1
brand hall effect sensors model CSNA111 were used, which turn or 5% which present a slight increase with respect to the
is a closed loop current sensor that has dynamic response healthy case that may not be decisive.
 occurs. In the case of the experiments carried out at 200
rpm, the frequency band corresponds, in position, to the 4th
band, which goes from 31.3 Hz to 62.5Hz. In the 200 rpm
experiments, significant differences in the rms value of the
wavelet coefficients occurred when the analysis was performed
with Daubechies 4 mother wavelet with differences from 0.21
to 0.38 for faults in one phase, having fault detection from the
first short-circuited turn. In the case of the experiments carried
out at 400 rpm, the frequency band corresponds, in position,
to the 3rd band, which goes from 62.5 Hz to 125Hz. In
the 400 rpm experiments, the greatest differences in the rms
Fig. 4. Comparison of frequency spectra of the ISP module healthy and value of the wavelet coefficients occurred when the analysis
faulty; PMSG with 5% short circuit of a stator coil at 200 rpm was performed with the Daubechies 4 wavelet with differences
from 0.28 to 0.38 for faults in one phase with fault detection
from the first shortcircuited turn.

TABLE II
D IFFERENCE IN RMS VALUES OF THE WAVELET BANDS BETWEEN THE
HEALTHY AND FAULTY CASES

Machine speed 200rpm 400rpm

Frequency band 4th band 3rd band
Frequency range 31.3-62.5 62.5-125
% of fault Wavelet RMS difference RMS difference
Harr -0.003585 0.005101
fault 5%
Daub 4 0.225957 0.286429
Harr -0.033681 0.011161
fault 20%
Daub 4 0.219905 0.359456
Harr -0.006964 0.023377
fault 45%
Daub 4 0.38521 0.38489
Fig. 5. Comparison of frequency spectra of the ISP module healthy and
faulty; PMSG with 10% short circuit of a stator coil at 200 rpm

C. Hilbert-Huang analysis
By using the Hilbert-Huang transform, the fault due to
short circuit between turns is detected. As in the case of the
WT and the DFT, the detection of the frequency ranges that
indicate the presence of the fault is carried out. The figures
7, 8 and 9 show the frequency spectra of the ISP with the
PMSG stator under the different failure scenarios for a speed
of 200 rpm In the ISPs corresponding to 200 rpm, the HHT
reveals oscillation modes between 35 and 55 Hz, delimiting the
frequency window supplied by the WT. Unlike the DFT, which
revealed little information, it only showed a frequency of 40
Fig. 6. Comparison of frequency spectra of the ISP module healthy and Hz. As an advantage, the HHT shows us the instantaneous
faulty; PMSG with 20% short circuit of a stator coil at 200 rpm frequency, that is, the behavior of the frequency of said failure
mode. Another advantage is that the detection of the failure
mode oscillation does not depend on a mother window as
B. Wavelet analysis the Wavelet transform does, but rather that the method is
As in the case of the DFT, for the detection of failure in the adaptive based on showing similar bands in the different failure
PMSG with the DWT, the healthy case and the faulty scenarios scenarios.
are compared applying the DWT analysis to the ISP.
The mother wavelets used to detect the short-circuit fault in V. C ONCLUSION
the DWT were Haar and Daubichies 4, the first because it is the In this work, we have demonstrated the ability of the differ-
basic wavelet for analysis and the second because it has shown ent techniques to detect the characteristic failure frequencies
good results for the detection of this type of fault in induction for a PMSG. On the other hand, it shows the advantages and
electric motors. Table II presents the comparison of the RMS disadvantages of each technique in said implementation. The
levels of the wavelet coefficients in the decomposition band limitation in stopping the frequency ranges using the DFT
where the change due to the failure presented for the failure is evident. The WT shows some more dynamic oscillation
scenarios of 5%, 20% and 45%. The comparisons in all modes, in it we can understand that said frequency is not
cases, show changes in the frequency band where the failure monotonic, but rather that the fault is represented in an
 R EFERENCES
[1] E. Artigao, S. Martı́n-Martı́nez, A. Honrubia-Escribano, and E. Gómez-
Lázaro, “Wind turbine reliability: A comprehensive review towards
effective condition monitoring development,” Applied Energy, vol. 228,
pp. 1569–1583, 2018.
[2] T. Yee Heng, T. Jian Ding, C. Choe Wei Chang, T. Jian Ping, H. Choon
Yian, and M. Dahari, “Permanent magnet synchronous generator design
optimization for wind energy conversion system: A review,” Energy
Reports, vol. 8, pp. 277–282, 2022. Selected papers from 2022 7th
International Conference on Advances on Clean Energy Research.
[3] S. Xu, S. Tao, W. Zheng, Y. Chai, M. Ma, and L. Ding, “Multiple
open-circuit fault diagnosis for back-to-back converter of pmsg wind
generation system based on instantaneous amplitude estimation,” IEEE
Transactions on Instrumentation and Measurement, vol. 70, pp. 1–13,
2021.
[4] W. Teng, X. Ding, S. Tang, J. Xu, B. Shi, and Y. Liu, “Vibration
Fig. 7. Instantaneous frequency of the faulty ISP module; PMSG with 5% analysis for fault detection of wind turbine drivetrains—a comprehensive
short circuit of a stator coil at 200 rpm. investigation,” Sensors, vol. 21, no. 5, 2021.
[5] L. Xiang, P. Wang, X. Yang, A. Hu, and H. Su, “Fault detection of wind
turbine based on scada data analysis using cnn and lstm with attention
mechanism,” Measurement, vol. 175, p. 109094, 2021.
[6] I. Bahloul, M. B. K. Bouzid, S. K. E. Khil, and G. Champenois, “Robust
novel indicator to distinguish between an inter-turn short circuit fault and
load unbalance in pmsg,” IEEE Transactions on Industry Applications,
vol. 59, no. 3, pp. 3200–3209, 2023.
[7] S. Bouslimani, S. Meradi, S. Drid, L. Chrifi-Alaoui, and A. Bezziane,
“State estimation mras and identification of stator winding phase fault
detection of the pmsg in wind energy based on the sliding mode control,”
Power Electronics and Drives, vol. 8, no. 1, pp. 109–127, 2023.
[8] X. Liu, W. Miao, Q. Xu, L. Cao, C. Liu, and P. W. T. Pong, “Inter-turn
short-circuit fault detection approach for permanent magnet synchronous
machines through stray magnetic field sensing,” IEEE Sensors Journal,
vol. 19, no. 18, pp. 7884–7895, 2019.
[9] C. Xiao, Z. Liu, T. Zhang, and X. Zhang, “Deep learning method for
fault detection of wind turbine converter,” Applied Sciences, vol. 11,
no. 3, 2021.
Fig. 8. Instantaneous frequency of the faulty ISP module; PMSG with 20% [10] Y. Tan, H. Zhang, and Y. Zhou, “Fault detection method for permanent
short circuit of a stator coil at 200 rpm. magnet synchronous generator wind energy converters using correlation
features among three-phase currents,” Journal of Modern Power Systems
and Clean Energy, vol. 8, no. 1, pp. 168–178, 2020.
[11] W. E. Sayed, A. Aboelhassan, A. Madi, A. Hebala, and M. Galea,
“Comparative analysis between unscented and extended kalman filters
for pmsg inter-turn fault detection,” in 2021 IEEE Workshop on Electri-
cal Machines Design, Control and Diagnosis (WEMDCD), pp. 243–248,
2021.
[12] N. M. A. Freire and A. J. M. Cardoso, “Fault detection and condition
monitoring of pmsgs in offshore wind turbines,” Machines, vol. 9, no. 11,
2021.
[13] S. Toumi, M. Benbouzid, and M. F. Mimouni, “Modeling and simulation
of a pmsg-based marine current turbine system with inter-turn faults,” in
Induction Motors - Recent Advances, New Perspectives and Applications
(D. A. El-Shahat, ed.), ch. 0, Rijeka: IntechOpen, 2022.
[14] H. Lee, H. Jeong, G. Koo, J. Ban, and S. W. Kim, “Attention re-
current neural network-based severity estimation method for interturn
short-circuit fault in permanent magnet synchronous machines,” IEEE
Transactions on Industrial Electronics, vol. 68, no. 4, pp. 3445–3453,
2021.
Fig. 9. Instantaneous frequency of the faulty ISP module; PMSG with 45% [15] D. L. Milanez and A. E. Emanuel, “The instantaneous-space-phasor:
short circuit of a stator coil at 200 rpm. a powerful diagnosis tool,” IEEE Transactions on Instrumentation and
Measurement, vol. 52, no. 1, pp. 143–148, 2003.
[16] J. J. Grainger and W. D. Stevenson Jr, Power system analysis. McGraw-
Hill series in electrical and computer engineering, 1994.
oscillation mode. This oscillation mode is evidenced with [17] M. Weeks, “Digital signals processing using matlab and wavelets,
greater efficiency when using the HHT, being an adaptive georgia state university,” 2007.
[18] J. S. Walker, A primer on wavelets and their scientific applications. CRC
method that better represents and separates the oscillation press, 2008.
modes. [19] I. O. Olalere and O. A. Olanrewaju, “Tool and workpiece condition
This method of analysis is a novel and disruptive technique classification using empirical mode decomposition (emd) with hilbert–
huang transform (hht) of vibration signals and machine learning models,”
for predictive and corrective maintenance of machines, im- Applied Sciences, vol. 13, no. 4, p. 2248, 2023.
proving the effectiveness of fault detection. Furthermore, it
could be a useful technique to implement in modern monitor-
ing systems, such as SCADA’s and PMU’s.