IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL.

72, 2023 3511309

New Signal Processing-Based Methodology for

Optimal Feature Selection of Corona Discharges
Measurement in HVDC Systems
Gabriel Augusto David , Pedro Oliveira Conceição Junior , Senior Member, IEEE,
Fabio Romano Lofrano Dotto , and Benedito Roberto Dos Santos

Abstract— This article presents a new method based on the numerous advantages over high voltage alternating current
combination of digital signal processing parameters for the (HVAC) systems [1], [2], [3], [4]. In this article, the measure-
selection of optimal characteristics of corona discharges in ment of the corona discharge current is a fundamental concern
high voltage direct current (HVDC) systems, particularly for
linearization of the discharge model for applications that require in HVDC systems because of the important information it
a simplified computational approach. The proposed method provides, such as the occurrence of partial discharges (PDs),
implements a new metric from the coefficient of variation (CV), electric current intensity, and characteristics relevant to the
CVSTFT , based on the short-time Fourier transform (STFT) and design of transmission lines and equipment [5], [6], [7], [8],
the Hinkley criterion to measure the spectral variability and [9]. Corona discharge power losses, on the other hand, can
determine the corona discharge profile in different situations.
An experimental analysis was performed by applying voltages affect the safe and economical operation of HVDC transmis-
between ±30 and ±100 kV in a conductor, and electrical current sion lines [10]. HVDC transmission systems work by exceed-
signals proportional to the corona effect were collected through a ing the corona start voltage, which can lead to various effects
data acquisition system. The results indicated that the application such as audible noise, radio and television (TV) interference,
of the new method was successful in quantifying, in a simple way,
power losses and deterioration of insulation systems, and
the percentage of growth of corona discharges as a function of
the voltage applied within the range of 40–80 kHz. Moreover, environmental impacts from ozone formation and biological
it showed 90%, 91%, 92%, 97%, 89%, 92%, and 93% of problems that have serious negative impacts on the lives and
reliability in calculating the root-mean-square deviation (RMSD) health of residents in the surroundings [11]. Therefore, it is
based on approximation by a linear model. The frequency band essential to predict and monitor corona discharge conditions
resulting from this study proved to be favorable to establishing
to mitigate adverse effects caused by the increased magnitude
a threshold for the percentage of corona discharge growth
according to its profile or condition of application, indicating of the electric field around HVDC transmission lines.
this information may be useful in the construction of mobile Corona discharges have been the subject of several the-
devices with low consumption and computational performance, oretical and experimental investigations in the vast field of
meeting the demands of Industry 4.0 and the Internet of Things. PDs, aiming to obtain more in-depth knowledge and charac-
Index Terms— Corona discharge, high voltage direct current terization of the phenomenon. In this article, it has become
(HVDC), instrumentation and measurement, signal processing. popular to use digital processing techniques for electrical or
acoustic signals that are measured in the HV system and carry
I. I NTRODUCTION information from discharges in the form of pulses in the time
domain. Pattern recognition methods or machine learning algo-
H IGH voltage direct current (HVDC) electrical power
systems play an increasingly significant role in electrical
power transmission due to their economic superiority and
rithms are also used for feature extraction. Recent advances
in the scope of separation between PD and noise sources
are reported in [12]. With regard to HVDC systems, studies
Manuscript received 3 November 2022; revised 26 January 2023; accepted that stand out have used finite element methods (FEMs) [13],
20 February 2023. Date of publication 23 March 2023; date of current
version 4 April 2023. This work was supported in part by the Interli- fuzzy logic techniques [14], online PD monitoring using
gação Elétrica do Madeira S.A. (IE MADEIRA) and in part by the Agên- clusters based on time-frequency classification maps [15],
cia Nacional de Energia Elétrica [Brazilian Electricity Regulatory Agency and online source recognition of PD using neural networks
(ANEEL)] under Grant PD-06957-0005/2020. The Associate Editor coor-
dinating the review process was Dr. Yau Chung. (Corresponding author: and independent component analysis [16]. The work of [17]
Pedro Oliveira Conceição Junior.) proposes a new wavelet-first pulse (WFP) filtering algorithm
Gabriel Augusto David is with the Department of Electrical Engineering, based on the wavelet transform to measure the time of arrival
São Paulo State University, Bauru, Sao Paulo 17033-360, Brazil (e-mail:
[email protected]). (TOA) of electromagnetic signals for the localization and
Pedro Oliveira Conceição Junior is with the Escola de Engenharia estimation of corona discharge sources in ac and dc systems.
de São Carlos, University of São Paulo (EESC-USP), São Carlos, Al-geelani et al. [18] addresses the identification of acous-
Sao Paulo 13566-590, Brazil (e-mail: [email protected]).
Fabio Romano Lofrano Dotto is with FAROL Pesquisa, Desenvolvi- tic signals from corona discharges under different levels of
mento e Consultoria, Bauru, Sao Paulo 17012-225, Brazil (e-mail: contamination using the wavelet transform. The work of [19]
[email protected]). involves an analysis of the mathematical correlation of the
Benedito Roberto Dos Santos is with the Interligação Elétrica do Madeira
S.A., Cuiabá 78080-970, Brazil (e-mail: [email protected]). spectral characteristics of the corona-related electrical current
Digital Object Identifier 10.1109/TIM.2023.3260879 signal in the frequency band of 5–20 kHz for an ultrahigh
1557-9662 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
3511309 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 72, 2023

voltage direct current (UHVDC) transmission line through the on the corona effect and which is equipped with a data
use of the external corona cage method. The work of [20] acquisition system. To validate the proposed approach for
addresses the study of the theoretical relationships between a comprehensive situation, experiments were carried out to
the corona current energy spectrum and radio interference (RI) detect corona discharges under different conditions by testing
in UHVDC systems. In addition, the work of [21] addresses with and without load and varying the position and angle of
the prediction of RI from HVDC transmission lines based on the conductive metal rod relative to the voltage source.
corona cage tests, correlations between average corona current,
RI excitation current, and corona current pulse repetition II. S IGNAL P ROCESSING FOR F EATURE E XTRACTION
frequency. OF C ORONA D ISCHARGES
Despite the advances and contributions that have been
made so far, there are still several research gaps in the A. Time-Frequency Analysis Based on STFT
field of extraction of corona discharge features, mainly in The STFT can be used to evaluate changes in frequency
HVDC systems, and concerning methods for determining content over time in the electrical current signals studied here
the corona discharge profile at different voltage levels and to detect relevant information about the time instants of the fre-
operating conditions. This article contributes by proposing a quencies at which the occurrence of corona predominates. The
new strategy based on techniques for the digital processing STFT of a signal is obtained by sliding an analysis window of
of electric current signals aimed at extracting and selecting length M over the signal and calculating the discrete Fourier
optimal corona discharge characteristics in different situations. transform (DFT) of the signals in each segment. The window
The proposed method involves a new combination of digital cycles through the entire signal at intervals of R samples. If a
signal processing parameters not yet reported in the literature, non-zero overlap length L is specified, windowed segment
based on time–frequency and signal energy analysis using overlapping compensates for signal attenuation at window
the short-time Fourier transform (STFT) and the Hinkley edges. The DFT of each segment is added to a matrix that
criterion. The STFT analysis sheds light on the mechanism of contains the magnitude and phase for each instant of time and
discharges, more representative frequency bands with higher frequency. The number of rows in the STFT matrix is equal
signal intensity, duration of discharge pulses, and repeating to the number of points in the DFT [23], [24]. The number of
patterns. The Hinkley criterion is a representative metric columns is given by
that allows one to determine the start and end points of  
Nx − L
discharges based on the energy of signals over time. Thus, k= (1)
the key innovation of this study is the implementation of M−L
the coefficient of variation (CV) metric from the values of where N x is the length of the original signal x(n), and the
the STFT magnitude in the most representative frequencies symbols ⌊⌋ denote the floor function.
and in the instants of time in which the corona discharges The STFT matrix is given by X( f ) =
occurred, which are determined by the Hinkley criterion. The [X 1 ( f ) X 2 ( f ) X 3 ( f ) . . . X k ( f )] such that the math element
new CVSTFT metric enables the distribution in frequency and of this matrix is given by the following equation:
intensity to be delivered in percentage only within the time ∞
of occurrence of corona discharges, based on a reference
X
Xm( f ) = x(n)g(n − m R)e− j2π f n (2)
signal (baseline). The new combination then uses the root- n=−∞
mean-square deviation (RMSD) metric to determine the best
frequency band that shows the percentage growth of corona where g(n) is a window function of length M; X m ( f ) is
discharge as a function of the voltage applied to the HVDC the DFT of the windowed segments centered on time m R;
conductor. and R is the size of the jumps (intervals) between successive
One of the advantages of this proposal is its simplicity DFTs. The square of the magnitude of the STFT produces the
since it does not involve the use of any other complex noise spectrogram representation from the power spectral density
separation technique or machine learning algorithm, as it (PSD) function [23], [24]
addresses the problem of quantifying the voltage level through
a model of the linearization problem. Thus, the proposed B. Hinkley Criterion
approach aims to establish a safe threshold of protection
The Hinkley criterion is an advanced signal processing
around transmission lines or to provide useful information
technique that characterizes predominant variations in the
about field measurements of electric current intensity and
frequency content and energy of a random signal, which can
voltage, enabling its use in simpler microprocessor systems
detect a variety of relevant information such as TOA and
and in line with new technologies of boarded systems. In this
changes in the signal behavior pattern over time [25], [26].
work, experimental analysis was performed by applying a
The energy criterion S(k) of a discrete signal x is defined
voltage to a conductive metal rod using a high-voltage dc
as the cumulative sum of the amplitude values of this signal,
source (Hipot) that allows the application of different voltages
as expressed by
between ±30 and ±100 kV. The electrical current signals were
measured using an innovative HVDC sensor device developed i  
X SN
by Dotto et al. [22], whose measurement principle is based S(k) = xk2 − (3)
k=1
N
DAVID et al.: NEW SIGNAL PROCESSING-BASED METHODOLOGY FOR OPTIMAL FEATURE SELECTION 3511309

where i is the number of samples from a selected part of TABLE I

the signal, S N represents total energy, and N is the length E XPERIMENTAL S ETUP
of the signal. This metric will be used in the context of the
present study to detect the start and end points of corona
discharges in HVDC systems, which is a difficult task without
this calculation because of the numerous sources of noise
contained in the signals.

C. Proposed Coefficient of Variation-Based Method

The CV is a standardized measure of the dispersion of a
probability distribution or frequency distribution that shows
the ratio between the standard deviation σ of the data and its is expressed as follows:
mean value µ, i.e., it measures data variability and is defined v
as follows: XωF u
u CVSTFT,O (k) − CVSTFT,R (k) 2
 
RMSDCVSTFT = t (8)
σ 
CVSTFT,R (k)
2
CV = . (4) k=ω1
µ
where CVSTFT,R (k) and CVSTFT,O (k) represent the variability
Thus, a low CV indicates small data variations or scattering profiles and distribution of a frequency range from ω1 to ω F ,
among samples in the data. In the case of electrical current in the instants of time of occurrence of a corona discharge to
signals that carry corona discharge information around a an electrical current signal taken as reference R and the one
conductor, very different fluctuations in the spectra of the in which greater intensity O is observed.
signals may occur under two or more conditions of operation In this study, the reference signal adopted corresponds to
or applied voltage. This article, therefore, introduces the CV the mildest condition, i.e., the lowest value of applied voltage,
for the analysis of several spectral segments in the instants which indicates a little transition between noise and discharge
of time of the occurrence of electrical discharges, measuring due to the low intensity of current amplitude. Thus, when
the spectral variability at different voltage levels and operating the condition transitions to corona discharge corresponding
conditions to determine different profiles of corona discharge to the highest value of applied voltage, the RMSD index is
in HVDC systems. To this end, the proposed method uses the expected to increase, enabling the selection of a percentage
STFT matrix, X m ( f ), only for the elements m corresponding increase according to the operating conditions and level of
to a time window m R between the start (ns) and end (ne) applied voltage. The experimental evaluation is described in
times of the corona discharge, determined by the Hinkley Section III.
criterion, for a given electric current signal. Therefore, the
standard deviation rate for the mean, starting from the SFTF III. E XPERIMENTAL E VALUATION
matrix, is obtained as follows: A. Experimental Test Bench
 ne

1 X  An experimental analysis was carried out in a laboratory
µSTFT ( f ) =  X m ( f ) (5)
 
 N m=ns  with a controlled environment and conditions to measure
v electrical current signals carrying information about corona
u
u1 X ne discharges under different operating conditions in HVDC
σSTFT ( f ) = t (X m ( f ) − µSTFT ( f ))2 (6) systems, thereby validating the proposed method. Testing was
N m=ns
conducted by applying voltages ranging from ±30 to ±100 kV,
using a Baur PGK 110 HB variable dc source with half-
thereby generating the metric CVSTFT , as follows: wave rectifier output. The electrical current signal is mea-
σSTFT ( f ) sured using an electromechanical HVDC sensor designed
CVSTFT ( f ) = . (7) by Dotto et al. [22], which is equipped with a pointed end.
µSTFT ( f )
This sensor allows the occurrence of the corona phenomenon
In this case, CVSTFT will be more representative of the through the effect of the point, which is fixed to a contact
variation of the corona discharge profile in two or more rod that works as a high-voltage conductor. The experiment
operating conditions or voltages applied to a narrow frequency was reproduced in different operating conditions to obtain
range. To increase the robustness of the proposed method, different corona discharge profiles, i.e., with load and without
we propose to use the RMSD, which is based on the Euclidean load, with a fixed distance of 1.5 m between the positive and
norm, to measure variations between values observed from a negative electrodes, and at different distances and angles of
reference. This article introduces the RMSD calculation for the the measuring device relative to the source, as described in
CVSTFT matrix, aiming to determine the frequency band that Table I.
best expresses the percentage of corona discharge growth as a It should be noted that the voltages applied to the tests
function of applied voltage (following a linear fit) by adopting carried out with load did not exceed ±80 kV, due to the con-
a reference value. The RMSD calculated for the CVSTFT matrix structive limits of the resistive divider (safe physical distance
3511309 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 72, 2023

Fig. 2. Framework of the proposed methodology-based feature extraction

and selection.

First, in order to perform a time-frequency analysis using

spectrograms, an algorithm was implemented to calculate the
STFT in pure signals, using blocks varying from 1024 to
8192 samples, and a Kaiser window function with a Beta
parameter equal to 5, in order to obtain optimal parameters that
would enable a viable resolution in frequency and time scales.
Digital Butterworth filters were used to refine the analysis
and eliminate undesirable frequencies, such as noise, and
those above the region of interest in the corona phenomenon.
Fig. 1. (a) Experimental equipment. (b) Schematic of the setup. The proposed method was implemented according to the
framework illustrated in Fig. 2.
Concomitantly (see Fig. 2) an algorithm was implemented
of insulation). In other words, ±80 kV is the maximum voltage to calculate the Hinkley criterion in pure signals to determine
supported by the device. In addition, environmental variables the time instants, in milliseconds, of the start and end of
that can exert an influence, such as temperature, atmospheric corona discharges based on the signal energy and TOA. The
humidity, and pressure, were kept constant at 30 ◦ C, 78%, parameters obtained from the STFT and Hinkley criterion
and 101.6 kPa, respectively. The test bench is illustrated in served as a basis for the subsequent implementation of the
Fig. 1(a) and (b) (equipment) and (schematic), respectively. algorithm for calculating the new CVSTFT metric to study the
The HVDC electromechanical measuring system designed frequency distribution and variability and only during the time
by Dotto et al. [22] (Fig. 1) is equipped with an electronic of occurrence of corona discharges to determine the different
measuring system. The data acquisition system is equipped corona discharge profiles. As specified in Section II, the study
with an irDA2 infrared communication device (Fig. 1) that was aided by the calculation of the RMSD metric from the
provides an interface through an optical link duly connected CVSTFT matrix to determine the best frequency band that
to a PC via USB. The system enabled the authors to auto- indicates the percentage of discharge variability at each voltage
matically acquire data during the tests by adopting variable level that was applied in the experiments. The electrical
gain parameters based on self-tuning algorithms, as well as current signals corresponding to ±30 kV served as a reference
sampling rates in the order of 1875 MHz, with a band limited (baseline) for the observations and calculations, while the
to 950 kHz by internal filters. Electric current signals were RMSD metric (CVSTFT ) was calculated for different frequency
collected in each of the conditions tested in the experimental bands and sub-bands within the range obtained by the CVSTFT .
tests described in Table I. The most suitable frequency band was chosen according to the
tendency, considering linear adjustment and the R-square (R 2 )
B. Proposed Method-Based Feature Extraction and Selection function. The results are described in Section IV.

The proposed method for extracting and selecting charac-

IV. R ESULTS AND D ISCUSSION
teristics related to corona discharges consists of implementing
the CVSTFT metric defined in Section II, which depends on A. Time-Frequency Analysis Based on STFT
the calculation of the STFT and the use of the Hinkley The results of the time–frequency analysis using the STFT
criterion. Thus, all the electric current signals collected in are depicted in the spectrograms in Figs. 3 and 4. These
the different operating conditions tested here, as described in spectrograms reveal the time-localized frequency information
Table I, were subjected to digital processing using MATLAB for situations in which frequency components related to corona
software. In this context, it should be noted that the time discharges vary over time. A color scale represents the magni-
domain signals are periodic due to the characteristics of the tude in dB of the PSD. Although the STFT has been calculated
dc source used in the experiments (half-wave rectifier output). for all applied voltage levels, this section shows only the
According to the sampling frequency, each collected signal two most representative conditions that summarize the entire
has a duration of 80 ms and comprises 150 000 samples. behavior, i.e., the electrical current signals of ±30 kV (milder
DAVID et al.: NEW SIGNAL PROCESSING-BASED METHODOLOGY FOR OPTIMAL FEATURE SELECTION 3511309

Fig. 3. STFT spectrograms of the #h60g0s test. Fig. 4. STFT spectrograms of the #h30g0c test.

condition—reference) and the signals in the most severe Fig. 4 presents the results of the #h60g0s and #h60g90s
condition of ±100 kV (no-load tests) or ±80 kV (loaded). The tests. Like the previous no-load test (Fig. 3—#h30g90s),
frequency axis, in this case, was designed to display frequen- but unlike the loaded test, the signals of the #h60g0s test
cies of up to 800 kHz, albeit focusing on the bands of greatest indicate little differentiation of the corona discharge pulses
interest, in which the energies related to corona discharges are at positive voltages, lack of uniformity in the milder condi-
concentrated (below 600 kHz). Fig. 3 presents the results of tion (+30 kV), and occurrence of several intermediate peaks
the #h30g90s and #h30g0c tests. As can be seen from Fig. 3, between pulses corresponding to negative voltage discharges
the electrical current signals generally differ between positive (especially at −100 kV). The subtle difference between the
and negative voltages in both the #h30g90s and #h30g0c tests. results shown in Fig. 4 and those of the previous no-load test
In this case, for negative voltages (−30, −80, or −100 kV), (Fig. 3—#h30g90s) is achieved by varying the sensor angle
the magnitudes of the signals are more uniform throughout the and distance from the source (0◦ versus 90◦ and 30 versus
entire time scale and in the frequency range of approximately 60 cm). It is also worth noting, in Fig. 4, that the result of the
up to 600 kHz, when compared to those of the positive negative voltage in the most severe condition, i.e., −100 kV,
voltages (+30, +80, or +100 kV). In the signals, note the highlights a particular feature of the negative half-cycle sine
occurrence of five discharge pulses lasting approximately 5 ms wave of the dc source used in the experiments (in no-load
at different moments in time throughout the total duration of testing).
80 ms. Fig. 4 shows that, in general, the electrical current signals
Fig. 3 also reveals differences between the corona discharge differ between positive and negative voltages in both the
profiles in the no-load test signals (#h30g90s) and those in the #h30g90s and #h30g0c tests. At negative voltages (−30, −80,
loaded test (#h30g0c). In the case of the reference signal at or −100 kV), throughout the entire time scale and in the
positive voltages +30 kV, note that the behavior was similar frequency range of up to approximately 600 kHz, the magni-
in both the #h30g90s and #h30g0c tests. Conversely, a more tudes of the signals are more uniform than at positive voltages
marked difference is visible in the severe conditions of each (+30, +80, or +100 kV). The signals reveal the occurrence
test at positive voltages, i.e., +100 and +80 kV, respectively. of five corona discharge pulses lasting approximately 5 ms at
However, the most significant difference between the no-load different moments through the duration of 80 ms. In Fig. 3,
tests (#h30g90s, −30 kV, and −100 kV) and loaded tests note the differences in corona discharge profiles between the
(#h30g0s, −30 kV, and −80 kV) appears in the graphs of no-load (#h30g90s) test signals and those of the loaded test
the electrical current signals at the negative voltages. (#h30g0c). As can be seen, the behavior of the reference
3511309 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 72, 2023

Fig. 5. Hinkley criterion applied to the #h30g90s and #h30g0c tests.

Fig. 6. Hinkley criterion applied to the #h60g0s and #h60g90c tests.

signal at the positive voltage of +30 kV was similar in

both the #h30g90s and #h30g0c tests. On the other hand, each discharge pulse (blue signal), representing the onset of
a more marked difference is visible at positive voltages in corona discharges, thereafter, increasing to a maximum point
the severe conditions of each test, i.e., +100 and +80 kV, corresponding to the energy of each corona discharge pulse,
respectively. The most significant difference, however, appears representing the end of the event. Fig. 6 presents the results of
in the graphs of the electrical current signals at the nega- the #h60g0s and #h60g90c tests. As can be seen, the Hinkley
tive voltages between the no-load (#h30g90s, −30 kV, and criterion is robust for the detection of transients at the begin-
−100 kV) and loaded (#h30g0c, −30 kV, and −80 kV) ning and end of corona discharges, in the different operating
tests. The parameters obtained from the STFT served as a conditions tested in this study, including the discharge profile
basis for the subsequent implementation of the algorithm for in the no-load and loaded tests, variations in distance and
calculating the new CVSTFT metric to study further information angle, and negative and positive voltages. In Figs. 5 and 6,
regarding the corona discharges. The Hinkley criterion results one can detect the beginning and end of the signals even in
are described in Section IV-B. the mildest condition of ±30 kV, with a low signal-to-noise
ratio (SNR) and low transition of corona discharge events due
B. Hinkley Criterion to the low intensity of the electrical current.
This section presents the results obtained by applying the
Hinkley criterion to the electric current signals collected C. Coefficient of Variation-Based Optimal Feature Selection
during the experimental tests carried out in this work. These This section discusses the results obtained by applying
results are depicted in Figs. 5 and 6. the CVSTFT metric to the electric current signals described
It is important to note that the values shown in all the in this article. Based on the results obtained previously by
graphs were normalized between 0 and 1 (positive voltages) the STFT spectrograms, a frequency range was chosen that
and 0 and −1 (negative voltages), given that the magnitudes proved to be more representative, from 20 to 500 kHz, for
of the electrical current signals and the Hinkley criterion have the analysis of several spectral segments in the instants of
different proportions. Likewise, it should be emphasized that time of occurrence of the discharges from the matrix of STFT
only the results obtained in more representative conditions are magnitudes. The application of CVSTFT to the various spectral
presented, i.e., mild (±30 kV) and severe (±80 or ±100 kV), segments from 20 to 500 kHz in each corona discharge section
since the same result was found for other conditions. Fig. 5 yields the result shown in Fig. 7. Each graph in Fig. 7
illustrates the results of the #h30g90s and #h30g0c tests. represents a type of test conducted, and each one contains
In Fig. 5, note that the behavior of the signal of the a resulting vector that expresses the spectral variability and
Hinkley criterion (orange) follows the tendency of the original compares it with the different applied voltage levels. Note
signal (blue) over time. This indicates that the energy crite- that the ±30 kV reference signals always present a lower
rion curves decay periodically to a local minimum point for CV along the spectrum in all the conditions due to the low
DAVID et al.: NEW SIGNAL PROCESSING-BASED METHODOLOGY FOR OPTIMAL FEATURE SELECTION 3511309

TABLE II
M OST R EPRESENTATIVE F EATURES

SNR and low transition activity between discharge and noise.

Thus, it is clear that the variability of the corona discharge
profile is low in this condition. Similar results are visible in
the corona discharges of the next voltage levels, i.e., ±40
and ±50 kV, which, notwithstanding a subtle increase in CV
compared to ±30 kV, present an average profile of variability
along the spectrum, due to the low intensity of the electric
current. A transition begins to appear starting from the PDs
at ±60 kV, followed by the conditions in which the corona
discharge profile is more intense, i.e., at ±70, ±80, ±90, and
±100 kV (for no-load tests—#h30g90s and #h60g0s), or at
±70 and ±80 kV (for loaded tests—#h30g0c and #h60g90c).
Thus, the CVSTFT metric indicates that corona discharges have
significant spectral variability at voltages of or higher than
±60 kV when compared to that of the reference profile of
±30 kV.
In order to determine the frequency band that best expresses
the percent increase of corona discharges as a function of
applied voltage within the variability profile shown in Fig. 8,
the RMSD was calculated for the CVSTFT matrix, taking as ref-
erence (baseline) corona discharges at voltages of ±30 kV, and
varying sub-bands within the 20–500 kHz range of the CVSTFT
matrix. As a result, it was found that the 40–80 kHz frequency
Fig. 7. Spectral variability profile of corona discharges under different
band proved to be the most suitable one for representing the conditions.
corona discharge profile to express its percent increase as
a function of the level of applied voltage, as illustrated in in embedded electronic devices, given that the acquisition
Fig. 8. The criterion for choosing the best frequency band rate for the selected band and the use of a simple RMSD
was based on the values of the coefficient of determination metric would suffice to quantify the applied voltage. Table II
R 2 , i.e., when the increasing trend of the RMSD index best describes the most representative features. It should be noted
adjusted to a linear fit (45◦ ). This information is relevant that the confidence level in the 40–80 kHz frequency band
because the use of this technique facilitates the identification exceeded 90% in most tests, except for two cases that showed
of a band wherein the problem can be treated with a linear confidence levels of 89% (#h30g90s—negative voltages) and
approach. The chosen band proved feasible for implementation 82% (# h60g90c—negative voltages).
3511309 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 72, 2023

TABLE III
C OMPARATIVE A NALYSIS

V. C ONCLUSION
This work investigated the use of a new combination of
signal processing parameters based on the STFT and the
Hinkley criterion. A new metric was implemented based on
the CV, CVSTFT , for the analysis of several spectral segments
in the instants of time of occurrence of electrical discharges.
The spectral variability was measured at different applied volt-
age levels and in experimentally tested operating conditions
(loaded tests, no-load tests, and variations in distance and
angle of the measurement point), determining different corona
Fig. 8. Percent increase in corona discharges. discharge profiles in HVDC systems. The results indicate
the feasibility of the newly proposed method based on the
selection of the optimal frequency band in the range of
As can be seen in Table II, corona discharges generally 40–80 kHz, demonstrating reliability levels of 90%, 91%,
increased by more than 10% after reaching ±50 kV in the con- 92%, 97%, 89%, 92%, 93%, and 82%, by calculating the
ditions of #h30g90s, #h30g0c, and #h60g0s (positive voltages). RMSD in the CVSTFT matrix. Thus, although corona discharge
In the conditions of #h60g0s (positive voltages) and #h60g90c, models present a non-linear response, the problem clearly
the threshold of a 10% increase in corona discharges was begins to exhibit a linear behavior after the band of interest
established starting from ±40 kV, while from ±50 kV the has been extracted, allowing their implementation in low-cost
percent increases were 36.52%, 40.42%, and 16.28%, for embedded electronic systems, in line with the application of
example. The results indicate the feasibility of the new method new technologies for the Internet of Things and Industry 4.0.
proposed here to simply quantify the percentage of frequency The proposed method can be further investigated in future
variability of corona discharge profiles at different voltage research to validate its effectiveness for a wider range of
levels and operating conditions in HVDC systems. Thus, it was corona characteristics, including current pulse amplitude, rise
possible to establish a threshold for the percent increase in time, half-wave time, repetition frequency, as well as RI
a corona discharge according to its profile by applying a excitation current prediction from corona cage tests in HVDC
linearization approach to the problem. This information may lines.
be useful in the surroundings of transmission lines or to
provide information about field measurements of the intensity R EFERENCES
of electric current and voltage. [1] C. E. Ghislain Ogoulola et al., “An alternative proposal for HVDC
transmission systems using 24-pulse AC/DC converters based on
three-winding non-conventional transformers,” Electr. Power Syst. Res.,
vol. 182, May 2020, Art. no. 106230, doi: 10.1016/j.epsr.2020.106230.
D. Comparative Analysis With Other Literature Approaches [2] A. Alassi, S. Bañales, O. Ellabban, G. Adam, and C. MacIver, “HVDC
transmission: Technology review, market trends and future outlook,”
In order to highlight the main contribution of the present Renew. Sustain. Energy Rev., vol. 112, pp. 530–554, Sep. 2019, doi:
research work, Table III lists the comparison and sum- 10.1016/j.rser.2019.04.062.
marizes the key aspects that distinguish the proposed [3] A. Tosatto, T. Weckesser, and S. Chatzivasileiadis, “Market integra-
tion of HVDC lines: Internalizing HVDC losses in market clearing,”
approach from some of the main approaches described in the IEEE Trans. Power Syst., vol. 35, no. 1, pp. 451–461, Jan. 2020, doi:
literature. 10.1109/TPWRS.2019.2932184.
DAVID et al.: NEW SIGNAL PROCESSING-BASED METHODOLOGY FOR OPTIMAL FEATURE SELECTION 3511309

[4] Y. Cui, H. Yuan, X. Song, L. Zhao, Y. Liu, and L. Lin, “Model, design, [23] MATHWORKS. (2020). Short-Time Fourier Transform. [Online]. Avail-
and testing of field mill sensors for measuring electric fields under high- able: https://www.mathworks.com/help/signal/ref/stft.html
voltage direct-current power lines,” IEEE Trans. Ind. Electron., vol. 1, [24] J. O. Smith. (2020). Spectral Audio Signal Processing. Accessed:
no. 65, pp. 608–615, Jan. 2018. Dec. 9, 2021. [Online]. Available: https://ccrma.stanford.edu/~jos/sasp
[5] G. Wijeweera, B. Bahreyni, C. Shafai, A. Rajapakse, and D. R. Swatek, [25] B. A. de Castro, D. D. M. Brunini, F. G. Baptista, A. L. Andreoli,
“Micromachined electric-field sensor to measure AC and DC fields in J. Alfredo, and C. Ulson, “Assessment of macro fiber composite sensors
power systems,” IEEE Trans. Power Del., vol. 24, no. 3, pp. 988–995, for measurement of acoustic partial discharge signals in power trans-
Jul. 2009, doi: 10.1109/TPWRD.2008.2008427. formers,” IEEE Sensors J., vol. 17, no. 18, pp. 6090–6099, Sep. 2017.
[6] E. Xin and H. Yuan, “Development of a sensor for corona current [26] S. M. Markalous, Z. Ring, S. Tenbohlen, and K. Feser, “Detection and
measurement under high-voltage direct-current transmission lines,” Int. location of partial discharges in power transformers using acoustic and
J. Distrib. Sens. Netw., vol. 12, no. 8, 2016, Art. no. 1550147716664243. electromagnetic signals,” IEEE Trans. Dielectr. Elect. Insul., vol. 15,
[7] C. Suo, H. Sun, W. Zhang, N. Zhou, and W. Chen, “Adaptive safety no. 6, pp. 1576–1583, Dec. 2008.
early warning device for non-contact measurement of HVDC electric
field,” Electronics, vol. 9, no. 2, p. 329, 2020.
[8] A. J. Otto and H. C. Reader, “Wideband and narrowband HVDC
conductor corona test methods for radio noise prediction,” IEEE
Trans. Power Del., vol. 25, no. 4, pp. 2950–2957, Oct. 2010, doi:
10.1109/TPWRD.2010.2051689.
[9] H. Yuan, Q. Yang, Y. Liu, J. Lu, and S. A. Akhtar, “Development Gabriel Augusto David received the bachelor’s
and application of high-frequency sensor for corona current mea- degree in electrical engineering from Paulista
surement under ultra high-voltage direct-current environment,” IEEE University (UNIP), São Paulo, Brazil, in 2018.
Trans. Instrum. Meas., vol. 61, no. 4, pp. 1064–1071, Apr. 2012, doi: He is currently pursuing the master’s degree in
10.1109/TIM.2011.2179337. electrical engineering with São Paulo State Univer-
[10] X. Encheng, X. Rui, Y. Haiwen, and L. Jiayu, “Design and appli- sity (UNESP), Bauru, Brazil.
cation of corona current measurement system under high-voltage He is currently a Researcher with the Laboratory
direct-current environment,” in Proc. 5th Int. Conf. Instrum. Meas., of Data Acquisition and Signal Processing, UNESP.
Comput., Commun. Control (IMCCC), Sep. 2015, pp. 692–696, doi: He has experience in the field of digital signal
10.1109/IMCCC.2015.151. processing and the Internet of Things (IoT) systems.
[11] M. A. Abouelatta, S. A. Ward, A. M. Sayed, K. Mahmoud, M. Lehtonen,
and M. M. F. Darwish, “Measurement and assessment of corona current
density for HVDC bundle conductors by FDM integrated with full
multigrid technique,” Electr. Power Syst. Res., vol. 199, Oct. 2021,
Art. no. 107370, doi: 10.1016/j.epsr.2021.107370.
[12] J. A. Ardila-Rey, M. P. Cerda-Luna, R. A. Rozas-Valderrama,
B. A. de Castro, A. L. Andreoli, and F. Muhammad-Sukki, “Separation Pedro Oliveira Conceição Junior (Senior Member,
techniques of partial discharges and electrical noise sources: A review IEEE) received the M.Sc. and Ph.D. degrees in
of recent progress,” IEEE Access, vol. 8, pp. 199449–199461, 2020. electrical engineering from São Paulo State Uni-
[13] M. A. Fard, A. J. Reid, and D. M. Hepburn, “Analysis of HVDC versity (UNESP), Bauru, Brazil, in 2016 and 2020,
superimposed harmonic voltage effects on partial discharge behavior respectively.
in solid dielectric media,” IEEE Trans. Dielectr. Electr. Insul., vol. 24, He is currently an Assistant Professor with the
no. 1, pp. 7–16, Feb. 2017. University of São Paulo, São Carlos, Brazil, where
[14] A. Cavallini, G. C. Montanari, M. Tozzi, and X. Chen, “Diagnostic of he works in the areas of sensor systems and indus-
HVDC systems using partial discharges,” IEEE Trans. Dielectr. Electr. trial automation.
Insul., vol. 18, no. 1, pp. 275–284, Feb. 2011. Dr. Junior is an Associate Editor of the IEEE
[15] N. D. Jacob, W. M. Mcdermid, and B. Kordi, “On-line monitoring S ENSORS J OURNAL.
of partial discharges in a HVDC station environment,” IEEE Trans.
Dielectr. Electr. Insul., vol. 19, no. 3, pp. 925–935, Jun. 2012.
[16] C. S. Chang, J. Jin, C. Chang, T. Hoshino, M. Hanai, and N. Kobayashi,
“Online source recognition of partial discharge for gas insulated sub-
stations using independent component analysis,” IEEE Trans. Dielectr.
Electr. Insul., vol. 13, no. 4, pp. 892–902, Aug. 2006.
Fabio Romano Lofrano Dotto received the bache-
[17] M. Szirtes, R. Cselko, and I. Berta, “Developing a filtering algorithm lor’s, master’s, and Ph.D. degrees in electrical engi-
for partial discharge location approximation using the emitted electro- neering from São Paulo State University (UNESP),
magnetic signals of corona discharges,” IEEE Trans. Ind. Appl., vol. 57, Bauru, Brazil, in 2001, 2004, and 2019, respectively.
no. 1, pp. 932–940, Jan. 2021, doi: 10.1109/TIA.2020.3038627. He is currently the Director of FAROL Pesquisa,
[18] N. A. Al-geelani, M. A. M. Piah, and Z. Abdul-Malek, “Identification Desenvolvimento e Consultoria, Sao Paulo, Brazil.
of acoustic signals of corona discharges under different contamina- He has experience in the field of digital signal
tion levels using wavelet transform,” Electr. Eng., vol. 100, no. 2, processing and industrial automation.
pp. 1059–1067, 2018.
[19] Y. Liu, L. Zhou, H. Yuan, Y. Liu, and L. Ji, “Analysis of self-correlation
characteristics of corona current spectrum on UHVDC transmission
line,” IEEE Trans. Dielectr. Electr. Insul., vol. 25, no. 3, pp. 928–938,
Jun. 2018, doi: 10.1109/TDEI.2018.006933.
[20] Y. Liu, L. Zhou, Y. Liu, H. Yuan, L. Ji, and Q. Wang, “Analysis
on the spectrum characteristic of corona current and its relationship
with radio interference on UHVDC transmission line,” IEEE Trans.
Dielectr. Electr. Insul., vol. 23, no. 6, pp. 3336–3345, Dec. 2016, doi: Benedito Roberto Dos Santos was born in Taubaté,
10.1109/TDEI.2016.006108. São Paulo, Brazil, in 1967. He received the bache-
[21] C. Huang, H. Yin, P. Xu, B. Zhang, J. He, and J. Liu, “Prediction of radio lor’s degree in electrical engineering from the Uni-
interference from HVDC transmission lines based on corona discharge versity of Taubaté (UNITAU), Taubaté, in 1995.
characteristics,” High Voltage, vol. 5, no. 6, pp. 679–687, 2020. He is currently an Electrical Engineer with Interli-
[22] F. R. L. Dotto, P. O. Conceicao Jr., A. L. Andreoli, R. G. de Oliveira gação Elétrica do Madeira S.A., Cuiabá, Brazil.
Jr., and B. R. R. D. Santos, “Development of a portable corona current He has experience in the field of electrical main-
measuring device for high voltage detection in HVDC systems,” IEEE tenance and operation of high voltage direct current
Trans. Ind. Electron., vol. 70, no. 2, pp. 2048–2056, Feb. 2023, doi: (HVDC) transmission lines.
10.1109/TIE.2022.3165290.