See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/336430864

Hybrid Technique for the Analysis of Non-Linear and Non-Stationary Signals

focused on Power Quality

Conference Paper · December 2019

DOI: 10.1109/FISECIGRE48012.2019.8984975

CITATIONS READS
5 704

4 authors:

Johinner Mauricio Sanabria Villamizar Maximiliano Bueno López

Universidad de La Salle Universidad del Cauca
10 PUBLICATIONS 34 CITATIONS 102 PUBLICATIONS 438 CITATIONS

SEE PROFILE SEE PROFILE

Marta Molinas Efrain Bernal A.

Norwegian University of Science and Technology Würth Elektronik eiSos GmbH & Co. KG
549 PUBLICATIONS 10,820 CITATIONS 36 PUBLICATIONS 125 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Maximiliano Bueno López on 11 October 2019.

The user has requested enhancement of the downloaded file.

 Hybrid Technique for the Analysis of Non-Linear
and Non-Stationary Signals focused on Power
Quality
Mauricio Sanabria-Villamizar ∗ , Maximiliano Bueno-López∗ , Marta Molinas † and Efrain Bernal ‡
∗ Department of Electrical Engineering, Universidad de la Salle, Bogotá, Colombia
Email: [email protected], [email protected]
† Department of Engineering Cybernetics

Norwegian University of Science and Technology

Email: [email protected]
‡ Department of Automation Engineering, Universidad de la Salle, Bogotá, Colombia

Email: [email protected]

Resumen—This paper aims to develop a combination method The basic idea of a time-frequency analysis is the design
for the evaluation of power quality disturbances. First, we apply of a strategy that can describe the characteristics of a signal
the Fast Fourier Transform, Wavelet Transform and Hilbert in both domains, so that it is not only possible to detect the
Huang Transform on a synthetic signal that represents typical
behavior in a power system with high penetration of Renewable presence of any frequency component but also the moment of
Energies. Then, we combine the methods to extract the best of time in which it is presented [8].
each of these and achieve a better signal decomposition. The In this paper, we present an analysis methodology capable
paper seeks to generate decision criteria on the method of analysis of extract the instantaneous frequency of a typical signal that
of signals to be used according to the application. appears in a power system with high penetration of renewable
Index Terms—Hilbert-Huang Transform, Hilbert Spectrum,
Power Quality, Fourier Fast Transform, Wavelet Transform, energy. Based on the combination of Hilbert Huang Transfor
Instantaneous frequency. (HHT), Wavelet Transform (WT) and Fast Fourier Transform
(FFT), the proposed methodology not only can display the
I. I NTRODUCTION estimation of the instantaneous frequency-amplitude, but also
Increased energy demand has given the need to generate the fundamental frequency. Furthermore, with the advantage
energy from unconventional methods, producing high varia- of requiring less computational time with greater efficiency
tions in the network due to its nonlinear nature, caused by compared to the other conventional methods of analysis for
nonlinear devices [1]. Quality in power systems is focused on this type of signals.
ensuring a voltage supply with an excellent waveform accor- II. N ON -L INEAR AND N ON -S TATIONARY S IGNAL
ding to the parameters set by the IEC 61000, to protect the A NALYSIS
operation of the equipment, which handles critical processes of
great importance to customers [2] [3]. Due to the penetration II-A. Hilbert-Huang Transform (HHT)
of the new technologies, such as non-conventional sources In order to solve the time-frequency-energy paradigm of
of renewable energies (NCSRE) in distribution networks, non-stationary and non-linear signals, Norden E. Huang in
the harmonic distortion of current and voltage waveform is [9], presented the Hilbert-Huang Transform. The method is
becoming an important issue [4]. Therefore, it is necessary based on the local characteristic timescale of the signal x(t),
to develop new methodologies for the analysis of signals that helped by the Empirical Mode Decomposition (EMD), finding
have a wide content of harmonics and noise from NCSRE the Intrinsic Mode Functions (IMFs). The IMF should be a
signals [5]. function that fulfils two conditions:
Alterations in the quality of the power can be classified into 1. The number of ends and the number of crosses by zero
two categories, these are: variations and events. Variations are should be different from just one.
disturbances that influence each cycle, such as harmonics or 2. Its local medial is zero.
voltage imbalances. Events are disturbances that last a while, For a signal x(t), the analytic signal z(t) is defined as:
from a fraction of a cycle to several of these, and then they
may not be repeated for several hours or days [6]. The most z (t) = x (t) + iy (t) = a (t) eiθ(t) (1)
common energy quality problems are: voltage drop, harmonics
where y(t) is the Hilbert Transform of x(t), so that:
and switching transients. For the most part, these are caused
Z ∞
by the intermittent nature of unconventional power generation 1 x (τ )
and large load feeders [7]. y (t) = P dτ (2)
π −∞ − τ
t
 Where P is the Cauchy Principal Value. a(t) and θ(t) are II-C. Fast Fourier Transform (FFT)
defined in (3) and (4) respectively. The Fast Fourier transform (FFT) advance an efficient
p computational algorithm for converting signals from the time
a (t) = x2 + y 2 (3) domain into the frequency domain [15]. One variation of the
FFT is the Discrete Transfer Function (DTF) and is defined
  according to the equation (8).
y (t)
θ (t) = arctan (4)
x (t) N −1
n
X
X (k) = x (n) e−j2πk N (8)
The instantaneous frequency is defined by (5).
n=0

1 dθ (t) II-D. Comparison between methods

f (t) = (5)
2π dt The FFT has the inability to break down a non-stationary
II-A1. Masking Signal: In order to find a solution to Mode and non-linear signal into a monotonic signal, this issue is
Mixing problem, Deering and Kaiser [10] defines the Masking called aliasing. The above-mentioned issue is not only present
Signal method, supported by the EMD. The algorithm is in the FFT, but also in the WT and WT, but with different
defined as: consequences. The most important weakness of the WT is
the selection of the type of mother wavelet (window), this
Algorithm 1 Masking Signal Method should have coupling characteristics and adaptation to the
Result: Descomposición Empı́rica en Modo con Masking signal to be analyzed. The WT has better resolution in different
Se construye una señal de máscara s(t) basada en G. applications, but just like the FFT has the issue of aliasing,
Rilling y Fosso, respaldada por la información de señal therefore, it does not generate an adequate frequency spectrum.
original, x(t). However, HHT is not free of problems, its weakness is in
El EM D se obtiene de las señales: x(t) y s(t). the heuristic way of selecting EMD type, like the WT it
depends on a mother function; therefore, a hybrid technique
x+ (t) = x (t) + s (t) is performed in this document using the above-mentioned
analysis methods in order to extract the advantages of each
x− (t) = x (t) − s (t)
and generate a robust method of analysis.
Luego, se obtienen IM F + y IM F − .
Calcule el valor principal, IM F se define como: III. C ASE S TUDY

IM F + (t) + IM F − (t) III-A. Synthetic Signal

IM F (t) =
2 To verify the methodologies presented above, the analysis
Realizando el proceso para cada residuo. based on the EMD Masking, EMD Masking+FFT, EMD
Masking+WT and EMD Masking+WT+FFT for a synthetic
signal is presented. The signals present a phenomenon called
Some recent variations to the method can be seen in [11] Mode Mixing, to solve this applies the methodology of the
and [12]. signal mask in [10], [11] and [16]. The Mode Mixing Problem
appears frequently in modern power systems where different
II-B. Wavelet Transform (WT) frequency components are detected. The synthetic signal is
In 1980 was developed the wavelet (WT) transformed by described in equation (9) and (10) and also shown in Fig. 2,
Morlet and Grossmam, defined as: with its corresponding frequency modes.
The following conditions are considered for the design
∞  
t−u
Z
1 and selection of the synthetic signal: High harmonic and
W (u, s) = √ x (t) ψ dt (6)
s −∞ s inter-harmonic content and a significant amount of noise,
characteristic of modern unconventional energy sources, as
The WT breaks down a signal in frequency bands, which the following shows Fig. 1 and Fig. 2. This signal simulates
are generated by means of a function called wavelets mother the ensuing phenomena: voltage dips, voltage swells, voltage
(7) and function translation operations in the time domain: fluctuations, rapid voltage changes and transients [17].
 
1 t−u
ψus (t) = √ ψ (7) x (t) = 3 · sen (2π · 21 · t)
s s
+ 4 · sen (2π · 5 · t)
So, it responds to the optimal resolution needs in a time-
frequency analysis for all signal ranges [13]. Currently the + 3 · sen (2π · 2 · t) (9)
WT is a popular method in the analysis of sound signals and + S3
vibration in engineering [14]. + N OISE
 
2 · sen(2π · 12 · t) → 1<t<2
S3 = (10)
0 other wise


Figura 3. Separation performance measurement for two-mode signals, a 2-D

representation.

Figura 1. Synthetic signal to be analyzed.

The other two IMFs show a severe problem of Mode Mixing.
Similarly, Fig. 5 shows the above-mentioned, evidenced in its
instantaneous frequencies.

Figura 2. Synthetic signal with their respective frequencies.

By knowing the characteristics of the signal to be analyzed,

it is expected that the frequency separation will be visible Figura 4. Decomposition using EMD with Masking Signal.
when applying EMD Masking to the signal. For this reason,
the mask signal is applied with the form s(t) = A0 · sen(2π ·
f · t), where a0 and f are obtained according [11], [11],
[16] and is explained as follows: the mask signal parameters
are chosen from a map of the ratio between amplitudes and
frequencies shown in Fig. 3, where the blue region represents
the separation of the mixed modes from the original signal.
In terms of frequency, this occurs when the ratio f¯/f < 0,67
(this number may change depending on the resolution used in
the map construction), where f¯ is the highest frequency of the
signal. For the analysis of this case, the ratio is 0.3 and the
amplitude ratio is -0.398. In this case the mask signal used
was s(t) = 11 · sen(2π · 70 · t).
IV. R ESULTS
IV-A. Masking signal
The EMD with masking signal is shown in Fig. 4, where Figura 5. Instantaneous frequency using EMD with Masking Signal.
the first two IMFs show the characteristic noise of the signal,
IV-B. Masking signal + FFT
Looking to find an efficient technique for frequency se-
paration, avoiding what is shown in Fig. 4 and Fig. 5, the
combination of FFT and HHT is proposed. Fig. 6 shows the
procedure performed, in the first part appears the original
signal, the following shows the signal after applying FFT and
the frequency spectrum.

Figura 8. Instantaneous frequency using EMD with Masking Signal+FFT

original signal, the approximation of the WT is obtained Fig.

9.

Figura 6. Original signal, signal after applying FFT and frequency spectrum.

The final result after applying making signal and FFT is

shown in Fig. 7. The first IMF show any remaining noise after
applying the FFT. The second IMF presents a Mode Mixing in
t = 1s and t = 2s, mixed frequencies correspond to 21Hz and
12Hz. The third and fourth IMFs show frequencies of 5Hz
and 2Hz, respectively. Fig. 8 corresponds to the instantaneous
frequency of this method, with the frequencies said above. In
addition, an acceptable accuracy in the separation of modal
frequencies is observed.
Figura 9. Original signal, signal after applying WT

The result of applying this methodology is shown in Fig.

10, frequency separation is similar to that obtained when
performing Masking+FFT, for example, the form of frequency
separation performs in the same way, a visible advantage of
this strategy is that the noise is more accurately separated.
The other characteristics continue to behave the same, the first
IMF show any remaining noise after applying the FFT. The
second IMF present a Mode Mixing in t = 1s and t = 2s,
the third and fourth IMFs show frequencies of 5Hz and 2Hz,
respectively, shown in Fig.11.

IV-D. Masking signal + WT + FFT

Figura 7. Decomposition using EMD with Masking Signal+FFT
In order to seek greater accuracy in the decomposition of
frequencies, we propose the combination of EMD with mas-
IV-C. Masking signal + WT king signal + WT + FFT. First, we applied WT (Daubechies 6,
In order to join the WT with the HHT, the selection of db6), and then perform the FFT. In Fig. 12 appears the result
the Wavelet (mother’s function) most suitable for the original of this process.
signal had to be carried out. The selection of this function is The execution of this methodology produces an excellent
done heuristically, due to the different characteristics of the decomposition, shows in Fig. 13 and Fig. 14. Fig. 13 in its first
signal to be analyzed. For this case, the Wavelet Daubechies part has an IMF only with noise components, the second IMF
6 (db6) was selected. When this process is carried out at the presents a Mode Mixing between the frequencies of 21Hz
 Figura 10. Decomposition using EMD with Masking Signal+WT
Figura 13. Decomposition using EMD with Masking Signal+WT+FFT

Figura 11. Instantaneous frequency using EMD with Masking Signal+WT

Figura 14. Instantaneous frequency using EMD with Masking Sig-
nal+WT+FFT

Cuadro I
C OMPARISON OF TIME - FREQUENCY ANALYSIS METHODOLOGIES

Masking+ Masking+ Masking+

Masking
FFT WT FFT+WT
Noise
X ≈ ≈ X
Reduction
Sampling
X X X X
Frequency
Feature
X ≈ ≈ X
Extraction

Figura 12. Original signal, signal after applying WT, finally the signal with
WT+FFT
the Masking+FFT and Masking+WT have a similar result
but with different difficulties: both present difficulties in ef-
and 12Hz, in this case, in less proportion and greater stability. ficient feature extraction proportional to their partial noise
Finally, the latest IMF shows the frequency of 5Hz isolation capacity, in-contrast the sampling rate is acceptable
in Masking-FFT and in Masking-WT this frequency is reduced
V. D ISCUSSION losing details of signal behavior. However, these problems are
In order to better understand the qualities of each met- solved with Masking+FFT+WT, breaking the paradigms that
hodology, table 1 brief the advantages and disadvantages of were had for signal analysis. This technique has the ability
the different transforms analyzed. The masking signal does to separate noise completely with the dexterity of keeping
not allow adequate noise extraction, making impossible to the sampling rate, improving signal details. For this reason,
efficiently extract signal characteristics. On the other hand, this novel technique has an excellent extraction of the signal
characteristics. R EFERENCIAS
[1] L. Peng, G. Jing, X. Duo, W. Chang, and Y. Xavier, “Hilbert-huang
VI. C ONCLUSIONS transform with adaptive waveform matching extension and its applica-
tion in power quality disturbance detection for microgrid,” Journal of
In this paper we have discussed different methods of sig- Modern Power Systems and Clean Energy, vol. 4, no. 1, pp. 19–27,
nal analysis, in order to find the most efficient strategy in 2016.
[2] W. Pei, K. Sheng, L. Kong, and Z.-p. QI, “Impact and improve-
terms of frequency separation for signals with high noise ment of distributed generation on distribution network voltage quality,”
content, characteristic of signals with high penetration of Non- Proceedings-Chinese Society of Electrical Engineering, vol. 28, no. 13,
Conventional Sources of Renewable Energies [5]. In Fig. 4, p. 152, 2008.
[3] D. Kai, L. Wei, H. Yuchuan, H. Pan, and Q. Yimin, “Power quality
it is possible to see the IMFs obtained with EMD Masking, comprehensive evaluation for low-voltage dc power distribution system,”
which do not adequately break down the signal, this is caused in 2019 IEEE 3rd Information Technology, Networking, Electronic and
by the high noise of the signal; this being evident in Fig. 5. Automation Control Conference (ITNEC), pp. 1072–1077, IEEE, 2019.
[4] M. Karimi, H. Mokhlis, K. Naidu, S. Uddin, and A. Bakar, “Photovoltaic
To solve this problem, three strategies have been proposed: penetration issues and impacts in distribution network–a review,” Rene-
1) EMD with Masking+FFT, 2) EMD with Masking+WT, 3) wable and Sustainable Energy Reviews, vol. 53, pp. 594–605, 2016.
EMD with Masking+WT+FFT. The results of applying EMD [5] G. Kulia, M. Molinas, L. Lundheim, and B. B. Larsen, “Towards a real-
time measurement platform for microgrids in isolated communities,”
with Masking+FFT and EMD with Masking+WT are similar Procedia engineering, vol. 159, pp. 94–103, 2016.
but very efficient, this process is a shown in Fig. 12 and Fig. 4, [6] N. R. Watson and A. Miller, “Power quality indices,” 2015.
both have the same characteristics: the first IMF represents the [7] A. Atputharajah, V. K. Ramachandaramurthy, and J. Pasupuleti, “Power
quality problems and solutions,” in IOP Conference Series: Earth and
excess noise after each process (FFT or WT), the second IMFs Environmental Science, vol. 16, p. 012153, IOP Publishing, 2013.
present a Mode Mixing in t = 1s and t = 2s, the third and [8] Q. Zhu, Y. Wang, and G. Shen, “Research and comparison of time-
fourth IMFs show frequencies of 5Hz and 2Hz, respectively, frequency techniques for nonstationary signals.,” JCP, vol. 7, no. 4,
pp. 954–958, 2012.
shown in Fig. 7 and Fig. 11. Achieving more accurate and [9] N. E. Huang, Z. Shen, S. R. Long, M. C. Wu, H. H. Shih, Q. Zheng, N.-
efficient frequency separation. C. Yen, C. C. Tung, and H. H. Liu, “The empirical mode decomposition
The limitation of applying the FFT together with EMD and and the hilbert spectrum for nonlinear and non-stationary time series
analysis,” Proceedings of the Royal Society of London. Series A:
Masking signal is the choice of the fundamental frequency of Mathematical, Physical and Engineering Sciences, vol. 454, no. 1971,
operation of the system, in this case, knowing beforehand the pp. 903–995, 1998.
characteristics of the signal the method is expected to work [10] R. Deering and J. F. Kaiser, “The use of a masking signal to improve
empirical mode decomposition,” in Proceedings.(ICASSP’05). IEEE
better. Applying the WT loses resolution due to the reduction International Conference on Acoustics, Speech, and Signal Processing,
of the original sample rate of the signal. In addition, it has the 2005., vol. 4, pp. iv–485, IEEE, 2005.
limitation of the choice of the Wavelet signal, this must have [11] O. B. Fosso and M. Molinas, “EMD mode mixing separation of signals
with close spectral proximity in smart grids,” in 2018 IEEE PES
peculiar characteristics depending on the signal to be analyzed. Innovative Smart Grid Technologies Conference Europe (ISGT-Europe),
These defects mentioned can be counteracted by joining the pp. 1–6, IEEE, 2018.
two methodologies, executing the EMD+WT+FFT results in [12] M. V. Gasca Segura, M. Bueno-Lopez, M. Molinas, and O. B. Fosso,
“Time-frequency analysis for nonlinear and non-stationary signals using
an excellent and accurate mode decomposition, and a smaller HHT: A mode mixing separation technique,” IEEE Latin America
magnitude of Mode Mixing, reducing the number of IMFs Transactions, vol. 16, pp. 1091–1098, April 2018.
needed to demonstrate better time-frequency resolution. [13] J. M. O’Toole, M. Mesbah, and B. Boashash, “A new discrete analytic
signal for reducing aliasing in the discrete wigner-ville distribution,”
The proposed strategies emerge as an alternative to imple- IEEE Transactions on Signal Processing, vol. 56, no. 11, pp. 5427–
menting HHT aided by different methods of analysis, taking 5434, 2008.
advantage of the qualities and strengths of each of them. This [14] S. G. Mallat, “A theory for multiresolution signal decomposition: the
wavelet representation,” IEEE Transactions on Pattern Analysis &
methodology results in a discovery of a robust and efficient Machine Intelligence, no. 7, pp. 674–693, 1989.
analysis technique for those signals where not only the ampli- [15] D. Donnelly, “The fast fourier and Hilbert-Huang Transforms: a com-
tude and frequency values are known previously, if not also, parison,” in The Proceedings of the Multiconference on Computational
Engineering in Systems Applications, vol. 1, pp. 84–88, IEEE, 2006.
those that have a high noise content. Common characteristics [16] G. Rilling and P. Flandrin, “One or two frequencies? the empirical
of signals from unconventional energy sources. These results, mode decomposition answers,” IEEE transactions on signal processing,
give an indication to the emergence of a technique of global vol. 56, no. 1, pp. 85–95, 2007.
[17] J. Barros, M. De Apráiz, and R. I. Diego, “Definition and measurement
analysis capable of breaking the limitations described, being of power quality indices in low voltage dc networks,” in 2018 IEEE 9th
a disruptive technique for the analysis of amplitude-time- International Workshop on Applied Measurements for Power Systems
frequency, nonlinear and non-stationary signals. (AMPS), pp. 1–5, IEEE, 2018.

VII. ACKNOWLEDGMENT
This paper is part of the project number 111077657914 and
contract number 031-2018, funded by the Colombian Admi-
nistrative Department of Science, Technology and Innovation
(COLCIENCIAS) and developed by the ICE3 Research Group
at Universidad Tecnologica de Pereira (UTP) and CALPOSA-
LLE Group at Universidad de La Salle.

View publication stats