energies
Article
Automatic PRPD Image Recognition of Multiple Simultaneous
Partial Discharge Sources in On-Line Hydro-Generator
Stator Bars
Ramon C. F. Araújo , Rodrigo M. S. de Oliveira






Citation: Araújo, R.C.F.; de Oliveira,
R.M.S.; Barros, F.J.B. Automatic
PRPD Image Recognition of Multiple
Simultaneous Partial Discharge
Sources in On-Line Hydro-Generator
Stator Bars. Energies 2022, 15, 326.
https://doi.org/10.3390/en15010326
Academic Editor: Ayman El-Hag
Received: 3 November 2021
Accepted: 2 December 2021
Published: 4 Janauary 2022
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Energies 2022, 15, 326. https://doi.org/10.3390/en15010326
and Fabrício J. B. Barros *
Electrical Engineering Graduate Program, Federal University of Pará, Belém 66075-110, Brazil;
ramon.araujo@itec.ufpa.br (R.C.F.A.); rmso@ufpa.br (R.M.S.d.O.)
* Correspondence: fjbbrito@gmail.com
Abstract: In this study, a methodology for automatic recognition of multiple simultaneous types of
partial discharges (PDs) in hydro-generator stator windings was proposed. All the seven PD sources
typical in rotating machines were considered, and up to three simultaneous sources could be identified.
The functionality of identifying samples with no valid PDs was also incorporated using a new
technique. The data set was composed of phase-resolved partial discharge (PRPD) patterns obtained
from on-line measurements of hydro-generators. From an input PRPD, noise and interference were
removed with an improved version of an image-based denoising algorithm previously proposed by
the authors. Then, a novel image-based algorithm that separates partially superposed PD clouds was
proposed, by decomposing the input pattern into two sub-PRPDs containing discharges of different
natures. From the sub-PRPDs, one extracts features quantifying the PD distribution over amplitudes
and the contour of PD clouds. Those features are fed as inputs to several artificial neural networks
(ANNs), each of which solves a part of the classification problem and acts as a block of a larger
system. Once trained, ANNs work collaboratively to identify an unknown sample. Good results
were obtained, with overall accuracies ranging from 88% to 94.8% for all the considered PD sources.
Keywords: partial discharge recognition; multiple simultaneous sources; identification of PRPDs
with invalid PD source; separation of multiple PD types; PRPD pattern; hydro-electric generators
1. Introduction
High-voltage electrical apparatuses are subjected to intense electrical, thermal, and
mechanical stresses during operation [1], resulting in equipment aging. In aged rotating
machines, including hydro-generators, partial discharges (PDs) occur in the stator insula-
tion. PDs are local electrical discharges that do not completely bridge phase bars and the
ground [2]. Partial discharges deteriorate the insulation progressively, and, in the absence
of proper maintenance, eventually result in disruptive failures.
To avoid the great financial and operational losses of such incidents, power utility
companies have resorted to partial discharge monitoring, deemed as the most effective
technique for assessing the condition of stator insulation [3,4]. The idea is to track the
actual health of machine insulation so that one can anticipate potential problems and plan
for maintenance measures accordingly. This is in the context of condition-based monitor-
ing, in which maintenance is performed only when necessary, based on the equipment’s
actual condition [5].
Partial discharges have different characteristics depending on the type of insulation
defect they occur in [2,3]. The task of determining the type of insulation defect based on
the characteristics of PD patterns is called PD recognition. Knowledge of the PD source is
very important in partial discharge monitoring systems [6] because the types of discharges
impose different relative risks to insulation [7]. Given that partial discharges are a complex
phenomenon, PD recognition is often done by experts [2,7]. Great research efforts have
https://www.mdpi.com/journal/energies
Energies 2022, 15, 326 2 of 26

been made to automate the task of PD recognition, which would certainly help leverage
condition-based maintenance of electrical equipment on a larger scale [8].
In rotating machines, partial discharges usually occur simultaneously at several insu-
lation defects throughout the stator. The characteristics of each PD source will superimpose
in the detected signal, resulting in a complex pattern. Another task in partial discharge
monitoring is the rejection of patterns containing no valid PD source. Such a task reduces
the rate of false alarms, which is important as it avoids the planning of unnecessary main-
tenance interventions. Patterns having only noise or very incipient discharges are too
common to be ignored; they composed the majority of measurements in [4], for example.
In [9], a methodology was proposed to recognize single-source PD patterns, including those
containing only noise, in gas-insulated switchgear (GIS) using deep convolutional neural
networks. Reference [10] proposed to identify noise by grouping discharges in different
clusters in time-frequency maps.
The classification of patterns containing multiple simultaneous PD sources is an
intricate problem [2,3]. It is usually solved with two main approaches.
In the first approach, one decomposes other types of input patterns (e.g., time-domain
signals) using clustering techniques. Most discrimination methods are based on clustering
features extracted from complex PD signals. Reference [6] proposed a new method to
separate multiple sources. The smoothed density and density-peak-clustering techniques
were combined to cluster discharges in the 2D space formed by the charge and energy
of discharges. Time-domain signals are required to perform this separation. Work [11]
describes a discrimination method of multiple simultaneous PD sources by clustering opti-
mized features of the cumulative energy function of partial discharge signals. The method
presented good performance when tested on signals from laboratory measurements and
simulations. In [12], a multi-source PD separation algorithm was proposed based on linear
prediction analysis and the isolation forest technique. Linear prediction is used to create
a feature space by approximating each data point of PD signals as a linear combination
of previous points. The isolation forest quantifies the degree of clustering by calculating
the height score metric. Multiple PD sources are separated by grouping PD signals with
fuzzy C-means clustering technique, in the 3D space formed by two features from the linear
prediction plus the height score. The method was tested on measurements from laboratory
models, and good results were achieved.
In the second approach, one applies artificial-intelligence algorithms to phase-resolved
partial discharge (PRPD) matrices or image patterns but without using clustering. The
work of Lalitha and Satish [13] was the first to propose multi-source PD recognition based
on PRPD images. Single-source PD measurements were taken in laboratory conditions
and were superposed onto one another to create an artificial data set of multiple-source
PRPDs. Those PRPDs were decomposed with 2D wavelet transform, and features were
extracted and fed to radial basis neural networks for classification. Recognition rates of
about 86% were obtained. Reference [14] presents a methodology for the PD recognition of
two simultaneous sources in medium-voltage motors. Fractal features were extracted from
the typical PRPDs of standards [15], and the resulting fractal map was used as a reference.
When presented to an unknown PRPD, the same fractal features were extracted, and they
were compared to the reference map using the Center Score algorithm, which estimates
the probabilities of belonging to each PD source. In [3], a method for recognition of up
to three simultaneous types of PD was proposed. Fractal features were extracted from
PRPD patterns using image-processing techniques and then inserted as inputs to a non-
linear multi-class support vector machine for classification. The database was composed of
measurements taken from artificial laboratory models. Work [16] proposes the use of an
improved convolutional neural network (CNN) to classify single and multiple PD sources
using only single-source patterns in the training set. The proposed architecture consists
of a CNN followed by multiple fully connected networks, each specialized in identifying
a different PD source. The input patterns were PRPDs obtained from measurements in
artificial models in laboratory. Six types of PDs were considered, and the proposed CNN
Energies 2022, 15, 326 3 of 26

was able to classify up to three simultaneous PD sources. The two works mentioned
represent the main options available for PD recognition: machine learning [3] or deep
learning [16]. Machine learning is the traditional method of first extracting from the raw
data manually predetermined features, which are then used as inputs to classifiers. In
deep learning, the raw data are directly passed to classifiers, which automatically learn a
subset of features and use it to identify the sample. Many works in the literature report
that deep-learning-based classifiers outperform their machine-learning counterparts only
when the database is relatively large [9,17]. This is expected by theory because, since deep
classifiers usually have more hyperparameters, they are more prone to overfitting, and, to
compensate for that, more training samples are necessary for adequate generalization of
the problem. Due to the complexity of partial discharge measurements and limited access
to real-world machines, it is hard to build large PD databases [17].
There are some shortcomings in the mentioned works, as well as in the overall litera-
ture on PD monitoring. First, most studies trained and tested their recognition methodolo-
gies with artificial models in laboratory conditions. Signals collected in these conditions
were less infected by noise and interference; hence, the statistics calculated overestimated
the accuracies that would be obtained in real-world online conditions [9]. The work [16],
for example, reported recognition rates greater than 96%, but this metric was calculated on
patterns obtained in laboratory conditions with very little noise. Such high recognition rates
might be unrealistic for online measurements in hydro-generators, which is a much harder
problem. Another disadvantage is that usually up to six PD sources are differentiated,
which does not cover all possible types. Additionally, most recent advancements in PD
denoising and separation of multiple sources rely on the time-domain PD signals, which
may not be available in storage-constrained applications due to the high sampling rate
required to record these signals. Finally, the majority of studies do not seem to have the
functionality of identification of patterns with no valid PD source.
This study mitigates many of those limitations. We propose a methodology that is
capable of classifying multiple simultaneous PD sources, including the functionality of
recognizing samples not having valid PD activity. The data set was formed of PRPDs
obtained from measurements in online hydro-generators containing multiple simultaneous
PD patterns. Noise and interference were removed with a modified version, tailored for the
multiple-source case, of our previous work’s denoising algorithm [18]. For this purpose,
partial discharges were clustered in different groups in each PRPD, which is a contribution.
Moreover, a new algorithm separates PD clouds, even if they are partially superposed, and
its output is the input sample decomposed into two sub-PRPDs containing discharges of
different natures. Features are extracted from the sub-PRPDs, including a novel attribute
that quantifies the shape of PD clouds. The classification task was decomposed into
smaller problems, each solved with an artificial neural network (ANN) trained with one
of those features. All the algorithms in the methodology treat PRPDs as images. To the
best of our knowledge, this is the first study to develop a method combining all these
functionalities. The proposed methodology achieved good performance on the tested
samples, with accuracies ranging from 88% to 94.8%.
The motivation of this work was to propose an algorithm combining, for the first time,
many functionalities that are necessary for the classification of multiple simultaneous PD
types in hydro-generators operating in real-world conditions, which is of great interest to
power utilities. For that, the contributions of this study are the following:
• a novel recognition method of multiple simultaneous PD sources;
• the development of a new input feature that quantifies the shape of PD clouds;
• the proposal of a new algorithm to detect invalid samples, which contain no readily
identifiable PD source;
• a new method of separation of multiple PD sources partially superposed in the
PRPD image.
Energies 2022, 15, 326 4 of 26

2. Data Set Obtained from Online PD Measurements in Hydro-Generators

The data set used in this work was built the same way as in our previous study [18]
but with some important differences. It is composed of online PD measurements from
hydro-generators of the Tucuruí and Coaracy Nunes power plants (Northern Brazil).
Measurements were performed as described in [18]. Capacitive couplers were used
as PD sensors. Signals picked up by the sensors were digitized with the NI-USB 5133
oscilloscope and then processed by the Instrumentation for Monitoring and Analysis of
Partial Discharges (IMA-DP) [19] acquisition setup. In each measurement, the signal was
acquired for 700 AC cycles; all PDs detected during this interval appeared as high-frequency
pulses in the signal. IMA-DP scans for all the peaks in the captured signal and registers
the number of peaks as a function of their amplitude and the phase angle of the AC
cycle at which they occurred in patterns called PRPDs. IMA-DP performs this mapping
by discretizing each of the amplitude and phase ranges in 256 equal bins, resulting in
256 × 256 PRPDs. Mathematically, PRPD is a matrix whose element at index (i, j) is the
quantity of PDs detected during acquisition whose peaks lie within the i-th amplitude and
j-th phase angle bins. PRPDs are what constitute the database.
The generators of the Tucuruí power plant have different power ratings and construc-
tion designs compared to those at Coaracy Nunes. Additionally, the data set was built of
measurements taken over many years, at different seasons of the year and at several times
of the day. Thus, measurements represent multiple conditions of operation of different
types of machines, captured at different stages of their aging process. The result is that the
database is statistically representative of the phenomenon of partial discharges in rotating
machines. Automatic PD-recognition algorithms that perform well on this data very likely
have generalized the task of classification (good accuracy on new samples).
The PRPDs’ true PD sources were manually labeled by a human expert among the
types of PD commonly found in rotating machines [15,20]: internal void (InV), internal
delamination (InD), delamination between conductors and insulation (DCI), slot, corona,
surface tracking (Trk), and gap discharges (Gap). Figure 1 shows clear PRPDs (little
contamination with interference and noise) associated with each one of these sources. The
manual labeling was performed considering the expert’s knowledge of the generators and
Energies 2021, 1, 0 5 of 26
the PRPD characteristics correlated with the underlying PD source, which are the symmetry
of PDs in positive and negative polarities and the shape of PD clouds [15,20].
Voltage / Amplitude

0º Phase 360º 0º Phase 360º 0º Phase 360º 0º Phase 360º

(a) (b) (c) (d)
1 Column 256
0 1
1-2
Voltage / Amplitude

3-4
PD count

5-8
Row

9-16 0
17-32
33-64
65-128
129-256 256
0º Phase 360º 0º Phase 360º 0º Phase 360º
(e) (f) (g)

Examplesof
Figure1.1. Examples
Figure ofclearly
clearlydefined
definedPRPDs
PRPDsfromfromthe
thedata
dataset,
set,each
eachassociated
associatedto tothe
thePDPDsource:
source:
(a)internal
(a) internal void,
void, (b)
(b) internal
internal delamination,
delamination, (c)
(c) DCI,
DCI, (d)
(d) slot,
slot, (e)
(e) corona,
corona, (f)
(f) surface
surface tracking,
tracking, and
and
(g) gap discharges. Figure adapted from [18] for convenience.
(g) gap discharges. Figure adapted from [18] for convenience.

Data Set
400
3356
r of PRPDs

256
 0º Phase 360º 0º Phase 360º 0º Phase 360º 0º Phase 360
(a) (b) (c) (d)
1 Column 256
0 1
Energies 2022, 15, 326 5 of 26
1-2

Voltage / Amplitude
3-4

PD count
Energies 2021, 1, 0 5-8
5 of 26

Row
Figure
9-16 2 shows0the PD source distribution in the data set. Unlike the data sets of [18],
17-32
there were patterns with multiple simultaneous PD sources and also a great number of
33-64
unidentified (Und) samples. Und patterns do not present valid PD sources; there is only
65-128
noise, or the PD clouds are so malformed (low height, density, etc) that they were not

Voltage / Amplitude
129-256 256
considered valid by the0ºspecialistPhaseduring labeling.
360º
Another
0º
remarkable
Phase
difference
360º 0º
fromPhase
[18] 360
is the absence of merged PD sources, (e) i.e., every one of the considered
(f) PD types can be
(g)
distinctively identified by the recognition system developed here, whereas, in [18], the
Figureof1.some
samples Examples
PD sourcesof clearly defined
were treated PRPDstofrom
as belonging the data
the same set, eachThis
class (merged). associated
is t
because
(a) of a
internalnew input
void, feature
(b) proposed
internal in this
delamination, study, which is
(c) 360º
DCI, able to differentiate
(d) slot, between
(e) corona, (f) surfa
slot
0º and corona
Phase and between
360º 0º internal
Phase void
360º 0ºand internal
Phase delamination
0º based
Phase on360º
the shape of
(g)clouds.
PD gap discharges.
(a) Figure(b)adapted from [18] (c) for convenience. (d)
1 Column 256
0 1
1-2
Data Set
Voltage / Amplitude

3-4
400
3356
PD count

5-8

Row
9-16 0
Number of PRPDs

17-32
33-64
65-128 256
129-256 256
0º Phase 360º 0º Phase 360º 0º Phase 360º
(e) (f) (g)
110 124
Figure 1. Examples of clearly defined PRPDs from the data set, each associated to the PD source:
70
48 void,
(a) internal 38(b) internal 47
delamination, (c) DCI, (d) slot, (e) corona, (f) surface tracking, and
(g) gap discharges. Figure adapted from [18] for convenience.

InV InD DCI Slot Corona Trk Gap Und

Data Set
PD Source 400
3356
Number of PRPDs

Figure
Figure2. Distribution of hydro-generator
2. Distribution PD sources in the
of hydro-generator PDdata set used
sources ininthe
thisdata
work. set used in this w
256
Because PRPDs are from online measurements, they are contaminated with many
types of interference and noise [21]. All of the disturbances in the data sets of [18]—intense
124 110 of PD clouds—were also present here. These
noise, crosstalk, and an ambiguous shape
70
disturbances,
48 combined
38 with the presence
47 of multiple simultaneous sources, are exempli-
fied by the PRPDs of the data set shown in Figure 3 and greatly increased the difficulty
of automated
InV PD recognition;
InD DCI such obstacles
Slot Corona Trk cannotUnd
Gap be avoided and are very common in
online measurements; therefore, any classification methodology to be of practical use must
PD Source
be robust enough to deal with them. The recognition methodology proposed in this study
Figure 2. Distribution of hydro-generator PD sources in the data set used in this work.
is able to tackle all these obstacles.

(a) (b) (c)

Figure 3. Cont.

(a) (b) (c)

Figure3.3.Cont.
Figure Cont.
 Energies 2021,
Energies 15,0326
2022, 1, 6 of 266 of 26

(d) (e) (f)

Examples
Figure3.3.Examples
Figure of of hard-to-identify
hard-to-identify PRPDs
PRPDs fromfrom
the the database:
database: (a) intense
(a) intense noise,noise, (b) additional
(b) additional PD PD
clouds(pointed
clouds (pointed
byby blue
blue arrows)
arrows) caused
caused by by crosstalk,
crosstalk, (c) cloud
(c) PD PD cloud of ambiguous
of ambiguous shape,shape, (d,e) multiple
(d,e) multiple
simultaneous
simultaneous PDPD sources,
sources, and
and (f) (f) unidentified
unidentified pattern.
pattern. Green
Green arrows
arrows and labels
and labels indicate
indicate the dominant
the dominant
PD
PDclouds
cloudsand the
and PDPD
the source(s) identified
source(s) by the
identified expert.
by the expert.

3.3.Proposed
ProposedRecognition
RecognitionMethodology
Methodologyof Multiple PD PD
of Multiple Sources
Sources
3.1.
3.1.Overview
Overview
Before
Beforepresenting
presentingthethe
methodology,
methodology, it isitimportant
is importantto review some
to review terminology
some defined
terminology defined
ininour previous study [18]. The noise gap is the horizontal band around
our previous study [18]. The noise gap is the horizontal band around amplitude zero amplitude zero
without
withoutPDs.
PDs.ANGPD
ANGPD discharges
discharges (an(an
acronym
acronym for for
“Adjacent
“Adjacentto Noise Gap Gap
to Noise PDs”) are close
PDs”) are close
totothe noise gap, and the other discharges are called N-ANGPD
the noise gap, and the other discharges are called N-ANGPD (Non-Adjacent (Non-Adjacent to Noise
to Noise
Gap PDs). ANGPD clouds are candidates to belong to PD sources InV, InD, DCI, slot, and
Gap PDs). ANGPD clouds are candidates to belong to PD sources InV, InD, DCI, slot, and
corona, whereas N-ANGPD discharges are potential surface tracking and gap sources. PDs
corona, whereas N-ANGPD discharges are potential surface tracking and gap sources. PDs
termed higher or lower have higher or lower absolute amplitudes than others. Those terms
termed higher or lower have higher or lower absolute amplitudes than others. Those terms
are illustrated in Figure 3.
are illustrated in Figure 3.
The common PD sources in rotating machines produce PRPDs typically characterized
The to
according common PD sources
Table 1 [15,20]. in rotating
Following machines produce
the categorization of TablePRPDs
1, it is typically
possible tocharacterized
classify
according to Table 1 [15,20]. Following the categorization
PD sources using a two-step process. First, a preliminary separation can of Table 1, it is be
possible to classify
performed
PD sources using a two-step process. First, a preliminary separation
based on the symmetry between positive and negative PDs, and then the PD clouds’ can be performed
basedis on
shape the symmetry
considered to resolvebetween
betweenpositive
sources and with negative PDs, and information.
the same symmetry then the PDThe clouds’
shape is considered
proposed to resolve
PD recognition between
methodology was sources withwith
structured the same symmetry
that same information. The
reasoning.
proposed PD recognition methodology was structured with that same reasoning.
Table 1. Characteristics of the typical PRPD associated to each PD source [15,20].
Table 1. Characteristics of the typical PRPD associated to each PD source [15,20].
Symmetry between
PD Source Shape of PD Clouds
Positive
Symmetryand between
Negative PDs
PD Source Shape of PD Clouds
InV Positive and Negative PDs Rounded
Symmetric
InV InD Triangular
Rounded
Symmetric
InDDCI Higher-amplitude negative PDs Triangular
Not applicable
DCISlot Higher-amplitude
Higher-amplitude negative PDs Triangular
Not applicable
Corona positive PDs Rounded
Slot Higher-amplitude Triangular
Tracking
Corona positive PDs Vertical N-ANGPD clouds
Rounded
Not applicable
Gap Pair of horizontal N-ANGPD clouds
Tracking Vertical N-ANGPD clouds
Not applicable
Gap Pair of horizontal N-ANGPD clouds
In this study, a PD-recognition methodology that is able to identify a single or multiple
typical sources in rotating machines is proposed. It is capable of identifying up to three
In this study,
simultaneous a PD-recognition
PD types: the dominantmethodology
ANGPD source thatplus
is able to identify
surface a single
tracking and/oror gap
multiple
typical sources
discharges. in rotating
The dominant machines
ANGPD is proposed.
source It is capable
is the one associated withof the
identifying up to three
pair of positive
and negative ANGPD clouds with the highest number and density of PDs. Gap-discharge gap
simultaneous PD types: the dominant ANGPD source plus surface tracking and/or
discharges.
clouds The dominant
are detected ANGPD
even if they source superposed
are partially is the one associated with theclouds.
onto the ANGPD pair ofThe
positive
and negative
absence of ANGPDANGPD clouds
and/or with thesources
N-ANGPD highestcan number
also beand densityThe
detected. of PDs. Gap-discharge
technique is an
clouds are detected even if they are partially superposed onto the ANGPD clouds. The
absence of ANGPD and/or N-ANGPD sources can also be detected. The technique is an
extension of those previously proposed by the authors in [18,22]; it operates entirely on
PRPD patterns, which are treated as images.
Energies 2022, 15, 326 7 of 26

Energies 2021, 1, 0 7 of 26

extension of those previously proposed by the authors in [18,22]; it operates entirely on

PRPD patterns, which are treated as images.
During
Duringthis thisresearch,
research, wewe diddid
not not
findfind
a feature that would
a feature single-handedly
that would be capable
single-handedly to
be ca-
accurately differentiate between all the typical PD sources mentioned
pable to accurately differentiate between all the typical PD sources mentioned in IEC in IEC 60034-27-2 [15].
Therefore,
60034-27-2the recognition
[15]. Therefore,task thewas divided task
recognition into smaller problems.
was divided This approach
into smaller usually
problems. This
results in shorter training times and better generalization due
approach usually results in shorter training times and better generalization due to theto the simpler individual
classifiers [23].
simpler individual classifiers [23].
The
The proposed methodologyisisbased
proposed methodology basedprimarily
primarilyon onfully-connected
fully-connectedfeedforward
feedforwardartifi-
artifi-
cial
cial neural networks. It consists of training, validation, and testing stages.In
neural networks. It consists of training, validation, and testing stages. Inthe
thefirst
firsttwo
two
phases,
phases,different
different ANNs
ANNs areare
trained
trainedso that eacheach
so that solves a smaller
solves problem
a smaller of theof
problem classification
the classifi-
task in an
cation isolated
task manner. manner.
in an isolated In the testing
In the stage, the stage,
testing previously trained ANNs
the previously are combined
trained ANNs are
hierarchically to perform PD recognition on unknown
combined hierarchically to perform PD recognition on unknown samples. samples.
The
Thetraining
trainingand andvalidation
validationstages
stagesare areillustrated
illustratedby bythe
theblock
blockdiagram
diagramof ofFigure
Figure4.4.
First,
First,the
theinput
inputpattern
patternisissubjected
subjectedto toaaPRPD
PRPDimage-denoising
image-denoisingalgorithm
algorithmfor forsuppressing
suppressing
spurious
spurious discharges that are not related to the underlying PD source(s). The algorithmisis
discharges that are not related to the underlying PD source(s). The algorithm
based
basedon onourourproposal
proposalof of[18]
[18]but
butwith
withsomesomemodifications
modificationsto toaddress
addressthe themultiple-source
multiple-source
case
case (Section 3.2). Next, a novel image-processing technique is applied toseparate
(Section 3.2). Next, a novel image-processing technique is applied to separatehorizon-
horizon-
tal
talN-ANGPD
N-ANGPDclouds cloudspartially
partiallysuperposed
superposedonto ontothetheANGPD
ANGPDdischarges
discharges(Section
(Section3.3).
3.3).The
The
image
imagedenoising
denoisingyieldsyieldsgroups
groups(clusters)
(clusters)of ofdenoised
denoisedN-ANGPD
N-ANGPDand andANGPD
ANGPDdischarges.
discharges.
From
Fromthese
thesegroups,
groups,two twosub-PRPDs
sub-PRPDsare aregenerated:
generated:one onecontaining
containingonly onlyN-ANGPD
N-ANGPDPDs PDs
and the other with only ANGPD
and the other with only ANGPD clusters. clusters.

Start ANGPD Training and

sub-PRPD validation of ANNs-0
PRPD image Pick subset
denoising with labeled
Data set and multiple- ANGPD source
(PRPDs) source
separation N-ANGPD Classification of
sub-PRPD N-ANGPD sources
(Figure 6)

5-fold cross validation

Data
Partitioning
Merge InV into InD
Pick slot and Pick InV and
and slot into
corona subsets InD subsets
corona subsets

Extraction of Extraction of Extraction of Feature

contour features contour features amplitude histograms [18] Extraction

Training and validation Training and validation Training and validation

of ANNs-3 of ANNs-2 of ANNs-1

Legend: Not new Modified from [18] Novel contribution

Flowchartof
Figure4.4.Flowchart
Figure ofthe
themultiple-source
multiple-sourcePD PDrecognition
recognitionmethodology—training
methodology—trainingand andvalidation
validation
stages. ANNs-0 are neural networks that detect whether the pattern’s ANGPD
stages. ANNs-0 are neural networks that detect whether the pattern’s ANGPD clouds are clouds arenoisy;
noisy;
ANNs-1 perform preliminary separation with histograms; and ANNs-2 and ANNs-3 resolve
ANNs-1 perform preliminary separation with histograms; and ANNs-2 and ANNs-3 resolve between between
InVand
InV andInDInDand
andbetween
betweenslot
slotand
andcorona
coronabased
basedononthe
thecontour
contourfeature,
feature,respectively.
respectively.

Fromthe
From thepatterns
patternsknown
knownto tohave
havesurface
surfacetracking
trackingand/or
and/orgap gapactivity
activity(as(aslabeled
labeled
by the specialist), the clouds of their N-ANGPD sub-PRPDs were
by the specialist), the clouds of their N-ANGPD sub-PRPDs were used to optimize the used to optimize the
thresholds of a manually predetermined set of rules (Section 3.5). In the testing
thresholds of a manually predetermined set of rules (Section 3.5). In the testing stage, those stage, those
rulesare
rules areused
usedtotoidentify
identify thethe presence
presence of such
of such PDPD sources
sources by analyzing
by analyzing the dimensions
the dimensions and
and positions of a PRPD’s N-ANGPD
positions of a PRPD’s N-ANGPD clouds. clouds.
Thedominant
The dominantANGPDANGPDPD PDsource
sourceisisidentified
identifiedby bythe
thecooperation
cooperationof offour
fourtypes
typesofof
neural networks: ANNs-0, ANNs-1, ANNs-2, and ANNs-3. Some
neural networks: ANNs-0, ANNs-1, ANNs-2, and ANNs-3. Some of their hyperparameters of their hyperparameters
arelisted
are listed in
in Table
Table 2;2; the
the topology
topology (number
(number of of hidden
hidden layers
layers andand hidden
hidden neurons)
neurons) isis
determined heuristically in Section 4.2. As detailed in Sections 3.8 and
determined heuristically in Section 4.2. As detailed in Sections 3.8 and 3.9, the networks 3.9, the networks
were trained and tested with several combinations of mutually exclusive
were trained and tested with several combinations of mutually exclusive subsets of training,subsets of training,
validation, and test data. The numbers of PRPDs forming those subsets
validation, and test data. The numbers of PRPDs forming those subsets are also listed in are also listed in
Table22for
Table foreach
eachnetwork
networktype.type.
Energies 2022, 15, 326 8 of 26

Energies 2021, 1, 0 8 of 26
Table 2. Information about the types of neural networks used.

Type of Neural Network

Property
Table 2. Information about the typesANNs-0
of neural networks used.
ANNs-1 ANNs-2 ANNs-3
Number of input neurons 6 32 of Neural Network
Type 32 32
Property
Number of output neurons 2 3 2 2
ANNs-0 ANNs-1 ANNs-2 ANNs-3
Number of input
Activation neurons
functions 6 Hyperbolic tangent
32 in the hidden
32 layers; 32
Number of output neurons 2 softmax3 in the output2layer 2
Size of training/ 2430/810/ Hyperbolic
195/66/
tangent in the51/18/
hidden layers; 102/35/
Activation functions
validation/test sets 809 66 in the output17
softmax layer 34
Size of training/valida- 2430/810/ 195/66/ 51/18/ 102/35/
Fromtion/test sets
the ANGPD 809
sub-PRPD, neural network66 17 whether the34pattern’s
ANN-0 detects
ANGPD clouds are noisy by analyzing simple cloud characteristics such as height and
density From the ANGPD
(Section sub-PRPD,
3.7). Next, neural network
the ANGPD sub-PRPDsANN-0 detects
with whether
a valid ANGPD the pattern’s
PD source
ANGPD
(as labeledclouds
by the arespecialist)
noisy by analyzing simpleinto
are partitioned cloud characteristics
different subsetssuch as height
(Section 3.8) to and train
density (Section 3.7). Next, the ANGPD sub-PRPDs with a valid
other three types of neural networks, each fed with one of the input features: amplitude ANGPD PD source
(as labeled [18]
histograms by the
andspecialist)
the novelare partitioned
contour into described
attribute, different subsets
in Section(Section 3.8) first
3.6. The to train
neural
other three types of neural networks, each fed with one of the
network (ANN-1) uses histograms to perform preliminary recognition, by merging InV input features: amplitude
histograms
with [18] and
InD patterns the novel
(class InV/InD) contour
andattribute,
slot withdescribed in Section 3.6. The
corona (Slot/Corona) intofirst
twoneural
different
network (ANN-1) uses histograms to perform preliminary recognition, by merging InV
groups of samples. Those mergers are necessary because the histogram does not accurately
with InD patterns (class InV/InD) and slot with corona (Slot/Corona) into two different
capture the differences in cloud shape that differentiate those classes [15,18]. The other
groups of samples. Those mergers are necessary because the histogram does not accurately
neural networks (ANN-2 and ANN-3) solve these ambiguities with contour features, which
capture the differences in cloud shape that differentiate those classes [15,18]. The other
are more sensitive to the cloud shape.
neural networks (ANN-2 and ANN-3) solve these ambiguities with contour features, which
The testing
are more sensitivestage
to theiscloud
illustrated
shape. in Figure 5. An input pattern is subjected to the
algorithm of noise removal
The testing stage is illustratedand multiple source
in Figure 5. separation (Sections
An input pattern 3.2 and 3.3).
is subjected In the
to the
N-ANGPD
algorithm of noise removal and multiple source separation (Sections 3.2 and 3.3). In the the
sub-PRPD, one looks for surface tracking and gap PDs by comparing
dimensions
N-ANGPD and positioning
sub-PRPD, one looks of itsfor
clouds
surfaceagainst
trackingpredetermined
and gap PDsthresholds.
by comparing Fromthe the
ANGPD
dimensions and positioning of its clouds against predetermined thresholds. From the to
sub-PRPDs, several metrics were extracted and used by the ANN-0 network
estimate
ANGPDwhethersub-PRPDs, ANGPDseveral clouds
metrics are noisy
were or valid.
extracted andIf used
both by
clouds are deemed
the ANN-0 network as noisy,
to
itestimate
is considered
whether that there isclouds
ANGPD no ANGPDare noisy PDorsource
valid. in the pattern.
If both clouds are If at least one
deemed cloud is
as noisy,
judged valid, thethat
it is considered pattern
thereisisclassified
no ANGPD by aPD cascade
sourceofintrained networks
the pattern. If at ANN-1,
least oneANN-2,
cloud isand
judged valid,
ANN-3. ANN-1 theidentifies
pattern isthe classified
samplebybased
a cascade
on itsofamplitude
trained networks
histogram.ANN-1, ANN-2, and
If ANN-1’s output
isANN-3.
DCI, thisANN-1
alreadyidentifies the sample
is the system based on its
classification foramplitude
the ANGPD histogram.
PD source,If ANN-1’s output
since this class is
is DCI, this already
characterized solelyisby thethesystem classification
symmetry between forpositive
the ANGPD and PD source,PDs
negative since this class
(Table is
1) [15,20].
Ifcharacterized
ANN-1’s output solelyisby the symmetry
InV/InD between positive
or slot/corona, contourand negative
features arePDs (Table 1)and
calculated [15,20].
passed
toIfthe
ANN-1’s output is InV/InD
corresponding auxiliaryor slot/corona,
neural network contour
(ANN-2 features are calculated
or ANN-3) to solve and passed
the ambiguity.
to the corresponding auxiliary neural network (ANN-2 or ANN-3) to solve the ambiguity.

Tracking Identification of Both clouds noisy

N-ANGPD and Gap noisy ANGPD
source(s) clouds (ANN-0)
Tracking
At least
only
No one valid
Start N-ANGPD Gap
source only
Extraction of
amplitude Trained
histograms [18] ANN-1
Classification
PRPD of N-ANGPD Slot / Corona InV / InD DCI
sources
Extraction of Extraction of
Contour Contour
PRPD Image N-ANGPD Features Features
Denoising sub-PRPD
and Multiple-
Source Trained Trained
Separation ANGPD ANN-3 ANN-2
(Figure 6) sub-PRPD

Legend:
Dominant Slot InD No
Not new ANGPD ANGPD
source source
Modified from [18] InV
Corona DCI
Novel contribution

Flowchart of the multiple-source

Figure5.5.Flowchart
Figure multiple-sourcePD
PDrecognition
recognitionmethodology—testing
methodology—testingstage.
stage.
Energies 2021, 1, 0 9 of 26
Energies 2022, 15, 326 9 of 26

Keepingthe
Keeping the practice
practice of [18]
of [18] of describing
of describing the algorithms
the algorithms independently
independently from thefrom the
PRPD dimensions, in this study, the parameter values of the denoising and
PRPD dimensions, in this study, the parameter values of the denoising and multiple-sourcemultiple-source
separationalgorithms
separation algorithms areare described
described as functions
as functions of theofnumber
the number of A
of pixels pixels A forming
forming the the
width/heightofof
width/height a PRPD
a PRPD pattern.
pattern. In case
In the the case ofstudy,
of this this study, A = 256.
A = 256.
The
Thefollowing
following subsections provide
subsections moremore
provide detailed descriptions
detailed of the stages
descriptions of theforming
stages forming
the methodology, especially the novel contributions (shown as boxes in
the methodology, especially the novel contributions (shown as boxes in blueblue in Figures 4–6).
in Figures 4–6)

Tailored for N-ANGPD clouds Legend:

Pixel M2 pattern ANGPD Not new
Noisy PRPD sub-PRPD
Start submatrix
PRPD decomposition Modified
filtering [22]
M1 pattern from [18]
N-ANGPD Novel
Separation of sub-PRPD contribution
Grid ANGPD Removal of superposed
filtering phase non-dominant N-ANGPDs
[18] delimiting ANGPDs Tailored for
(Figure 7) ANGPD clouds

Flowchart
Figure6.6.Flowchart
Figure of of
thethe PRPD
PRPD image
image denoising
denoising and multiple-source
and multiple-source separation
separation algorithms.
algorithms.

3.2.
3.2.Modifications
Modificationsto to
thethe
Denoising Method
Denoising of Our
Method of Previous Work Work
Our Previous
The
Thedenoising
denoising algorithm
algorithm used in this
used study
in this was based
study on theon
was based description of [18] but
the description of [18] bu
with some adaptations to better suit PRPDs with multiple simultaneous
with some adaptations to better suit PRPDs with multiple simultaneous types of discharges types of discharges.
The denoising algorithm is shown in Figure 6, highlighting which steps were modified or
The denoising algorithm is shown in Figure 6, highlighting which steps were modified o
added relative to the description of [18].
added relative to the description of [18].
The noisy PRPD was subjected to two independent filtering paths, one tailored for low-
densityThe noisy PRPD
N-ANGPD was
clouds subjected
(pixel submatrixto two independent
filtering) filtering
and the other paths, one
for ANGPD tailored for low
discharges
density
(four N-ANGPD
steps clouds
in this study). The(pixel
stepssubmatrix
of ANGPD filtering) and the other
phase delimiting for ANGPD
and removal of non-discharges
dominant ANGPD s rely on rough contours (CR ), which are functions of phase that estimate of non
(four steps in this study). The steps of ANGPD phase delimiting and removal
dominant
the amplitudeANGPDs rely on rough
of the boundary between contours
ANGPD(Cand R ), which
N-ANGPD are functions
dischargesofatphase that estimate
each phase
the amplitude
angle. The ANGPD of the boundary
phase delimitingbetween ANGPDon
stage operates and N-ANGPD
smooth contours discharges
(CS ), which at are
each phase
equal
angle.toThe
the ANGPD
filtering ofphase
roughdelimiting
contours with stage exponential
operates on moving
smooth averages
contours(EMA (C(SC),R )which
), are
followed by a linear transformation to change its values from row
equal to the filtering of rough contours with exponential moving averages (EMA(CR )) coordinates to absolute
amplitudes
followed by [18].
a linear transformation to change its values from row coordinates to absolute
Rough contours are functions of a free parameter g. Smaller g produces rough contours
amplitudes [18].
passing at lower amplitudes. The first modification is that, in this study, rough contours
Rough contours are functions of a free parameter g. Smaller g produces rough contours
were calculated with g = 0.0117A in ANGPD phase delimiting and with g = 0.0195A
passing
in at lower
the removal amplitudes. The
of non-dominant ANGPD first s,modification
whereas, in [18], is that,
g =in this study,
0.0273A rough
was used forcontours
were calculated with g 0.0117A in ANGPD phase delimiting
both of those stages. The values of g in both steps are smaller so that rough contours =
= and with g do 0.0195A
in the
not removallow-amplitude
encompass of non-dominant N-ANGPD ANGPDs, clouds, whereas,
a common in [18], g =in0.0273A
situation was used for
this database.
both of those
Different valuesstages.
of g wereTheused
values alsoofbecause
g in both steps
of the are smaller
different goals so thatsteps:
of the roughincontours
the do
not encompass
ANGPD low-amplitude
phase delimiting step, theN-ANGPD
rough contour clouds,
needsatocommon
follow the situation
high-densityin this dis-database
charges
Differentof ANGPD
values of clouds to obtain
g were usedaccurate phase bounds,
also because of the and, in the goals
different next step,
of thethe noise
steps: in the
of low- and
ANGPD medium-amplitudes
phase delimiting step, must
the be removed
rough contourto better
needs isolate the ANGPD
to follow clouds
the high-density dis
from spurious PDs.
charges of ANGPD clouds to obtain accurate phase bounds, and, in the next step, the noise
The second difference is that a new step of separation of multiple N-ANGPD dis-
of low- and medium-amplitudes must be removed to better isolate the ANGPD clouds
charges (Section 3.3) was added following the removal of non-dominant ANGPD s. The
from spurious PDs.
two filtering paths produced two partially denoised patterns M1 and M2 (Figure 6). The
The last
third and second difference
difference is that,isinthat
this a newthose
study, step patterns
of separation of multiple
were combined N-ANGPD
to form two dis
charges (Section
sub-PRPDs: 3.3) was added
one containing following the
only N-ANGPDs, andremoval
the otherofformed
non-dominant
only by ANGPD ANGPDs. The
two filtering paths produced two partially denoised
discharges (Section 3.4). In [18], a single PRPD containing only one dominant patterns M 1 andPD M 2 (Figure 6)
source
The third
was generated. and last difference is that, in this study, those patterns were combined to
form two sub-PRPDs: one containing only N-ANGPDs, and the other formed only by
3.3.
ANGPDSeparation of Superposed
discharges N-ANGPDs
(Section 3.4). In [18], a single PRPD containing only one dominan
In this stage, illustrated
PD source was generated. by the namesake block in the blue of Figure 6 and detailed in
Figure 7, the separation between partially superposed horizontal N-ANGPD (potential gap
discharges) and ANGPD
3.3. Separation cloudsN-ANGPDs
of Superposed in the PRPD image was performed, a common situation in
In this stage, illustrated by the namesake block in the blue of Figure 6 and detailed in
Figure 7, the separation between partially superposed horizontal N-ANGPD (potential gap
discharges) and ANGPD clouds in the PRPD image was performed, a common situation in
Energies 1, 0326
2021, 15,
Energies 2022, 10
10 of
of 26
26

the database. The separation was based on the fact that N-ANGPD clouds usually have PD
the database. The separation was based on the fact that N-ANGPD clouds usually have PD
counts higher than the ANGPD clouds at the intersection region.
counts higher than the ANGPD clouds at the intersection region.

Start PRPD formed Calculation

Set of horizontal
only of the of matrices Loop of
N-ANGPD
selected cluster’s M and D thresholds clusters Csep
discharges (dim. A/2 x A/2)
Intermediate
PRPD (IP) *
Case Iterative
Separation Identification Yes Any
1 expansion of cluster
of case-1 of superpo- each cluster
superposition sition type in Csep?
Nearest neighbor in Csep
clustering on IP No
Case 2 M1
Separation of pattern
horizontal Separation
Select an N-ANGPDs not of case-2
untreated ANGPD detected during the superposition Map separated clusters’
cluster of IP pixels back to original
loop of thresholds
PRPD dimensions (A x A)

Yes
Remove separated
N-ANGPDs from the Treated
Separated
ANGPD cluster and all IP’s ANGPD
assign them to new
clusters
clusters?
clusters
No

* Note: IP is the output of the stage “Removal of non-dominant ANGPDs”.

Figure7.
Figure Detailedflowchart
7. Detailed flowchartof
ofthe
themethodology
methodologystage
stage“separation
“separationof
ofsuperposed
superposedN-ANGPDs”.
N-ANGPDs”.

After the
After the removal
removal of of non-dominant
non-dominant ANGPDs ANGPDs (Figure (Figure 6),6), PDs
PDs were
were grouped
grouped by by means
means
of nearest-neighbor clustering [24] using a 0.05A × 0.05A pixel
of nearest-neighbor clustering [24] using a 0.05A × 0.05A pixel submatrix. Figure 8b shows submatrix. Figure 8b shows
the results of clustering applied to the pattern of Figure 8a,
the results of clustering applied to the pattern of Figure 8a, which in turn is the PRPD of which in turn is the PRPD of
Figure 3d after the removal of non-dominant ANGPDs.
Figure 3d after the removal of non-dominant ANGPDs. Each cluster is indicated with its Each cluster is indicated with its
PDs in
PDs in aa different
different color.
color.
The following
The following algorithmalgorithm was was applied
applied to to each
each ANGPD
ANGPD clustercluster separately.
separately. From
From thethe
PRPD, a matrix containing only the PDs of the cluster
PRPD, a matrix containing only the PDs of the cluster being analyzed was formed. This being analyzed was formed. This
matrix was
matrix was divided
divided into
into aa gridgridof
A/2 × A
of A
A/2 cells. One calculates matrix M, whose elements
2 × 2 cells. One calculates matrix M, whose elements
are equal to the sum of PD counts
are equal to the sum of PD counts within each within each cell
cell ofof the
the grid.
grid. Matrix
Matrix D was then
D was then obtained
obtained
by normalizing M by its maximum non-zero element
by normalizing M by its maximum non-zero element excluding outliers. Figure 8c, excluding outliers. Figure 8c, for
for
example, shows the matrix D relative to the positive
example, shows the matrix D relative to the positive ANGPD cluster of Figure 8b. ANGPD cluster of Figure 8b.
In order
In order to to remove
remove as as much
much superposition
superposition as as possible,
possible, itit is
is necessary
necessary to to estimate
estimate thethe
actual dimensions
actual dimensions of of the
the superposed
superposed N-ANGPD N-ANGPD clouds clouds (if (if any).
any). For
For that,
that, we
we exploited
exploited
the fact that PD clouds are usually formed by higher-count
the fact that PD clouds are usually formed by higher-count discharges in the central discharges in the central
region, surrounded by smaller-count discharges. If only the
region, surrounded by smaller-count discharges. If only the high-count PDs are considered, high-count PDs are considered,
one obtains
one obtains smaller
smaller versions
versions of of the
the superposed
superposed N-ANGPD N-ANGPD clouds clouds whose
whose boundaries
boundaries
(formed by the furthermost discharges from the cloud’s center)
(formed by the furthermost discharges from the cloud’s center) are farther from one another, are farther from one another,
facilitating their identification as separate groups. From the identified
facilitating their identification as separate groups. From the identified higher-count portion higher-count portion
of the superposed N-ANGPD clouds, it is easier to estimate
of the superposed N-ANGPD clouds, it is easier to estimate their actual dimensions. The their actual dimensions. The
separation algorithm
separation algorithm developed
developed in in this
this study
study follows
follows such such reasoning.
reasoning.
It is convenient to introduce the A/2 × A/2 matrix D L , given by
It is convenient to introduce the /2 × /2 matrix DL , given by
A A
(
D (D ) j) if if
i, (ji, DD(i,(ji,) j≥
) ≥L L , 1 ≤ i, j ≤A A/2 ,
DDL (Li,(i,
j)j)== , 1 ≤ i, j ≤ /2, (1)
(1)
0 0 otherwise
otherwise

where
whereLLisisa real-valued
a real-valuednumber
number from 0 to 01 that
from to 1 sets
thatthe threshold
sets of the minimum
the threshold normal-
of the minimum
ized PD count considered. For a given value of L, the matrix D is formed, and
normalized PD count considered. For a given value of L, the Lmatrix DL is formed, and its its non-zero
pixels of Dpixels
non-zero L are grouped
of DL arewith nearest-neighbor
grouped clustering [24]
with nearest-neighbor using a[24]
clustering 0.0195A
using×a 0.0195A
submatrix.
× 0.0195A One determines
submatrix. One the bounding
determines thebox (BB) of each
bounding cluster,
box (BB) which
of each is the which
cluster, smallest is
imaginary
the smallestrectangle
imaginary(height is H and
rectangle width
(height is and
is H W pixels)
widthcontaining
is W pixels) allcontaining
the cluster’s all dis-
the
charges.
cluster’sIn Figure 8d, for
discharges. example,
In Figure 8d,the
forBB of the cluster
example, the BBin of
redthe
is shown.
cluster D ’s N-ANGPD
inL red is shown.
clusters are scanned,
DL ’s N-ANGPD andare
clusters those meeting
scanned, and H ≥ 0.0195A
those meeting and
H ≥ W/H
0.0195A 3 are
≥ and W/Hconsidered
≥ 3 are
considered horizontal. This is repeated iteratively varying L in the sequence {0.70, 0.50,
Energies 2022, 15, 326 11 of 26
Energies 2021, 1, 0 11 of 26

horizontal. This is repeated iteratively varying L in the sequence {0.70, 0.50, 0.40, 0.30,
0.20,
0.40, 0.15,
0.30, 0.10,
0.20, 0.08,
0.15, 0.05, 0.02, 0.05,
0.10, 0.08, 0.00},0.02,
in a 0.00},
processin acalled
processthecalled
loop ofthethresholds. On the
loop of thresholds.
course of this loop, D gradually rebuilds matrix D, from the highest-count
On the course of this loop, DL gradually rebuilds matrix D, from the highest-count pixels
L pixels to those
of smallest-count, as seen in Figure 8d–f. In the initial iterations (high
to those of smallest-count, as seen in Figure 8d–f. In the initial iterations (high L), there L), there are only
the innermost high-count PDs of each cloud, and thus those groups
are only the innermost high-count PDs of each cloud, and thus those groups tend to be tend to be separated
(Figure
separated 8d).(Figure
As L decreases
8d). As L(Figure
decreases8e–f), the groups’
(Figure boundaries
8e–f), the progressively
groups’ boundaries approach
progressively
each other,each
approach eventually forming a forming
other, eventually single cluster again.
a single clusterForagain.
each horizontal N-ANGPD
For each horizontal N-
cloud
ANGPD detected
cloud during
detectedthe loop,the
during let C sep be
loop, letthe
Csepcorresponding cluster in
be the corresponding matrix
cluster inD L for the
matrix DL
smallest L at which
for the smallest L atthe N-ANGPD
which and ANGPD
the N-ANGPD clustersclusters
and ANGPD are still are
separated. Based on
still separated. the
Based
Con clusters,
septhe one separates N-ANGPD clouds from the ANGPD
Csep clusters, one separates N-ANGPD clouds from the ANGPD cloud with the cloud with the procedure
explained
procedurebelow, illustrated
explained below,inillustrated
Figures 9–11. If no horizontal
in Figures 9–11. If nocloud is found
horizontal at any
cloud of the at
is found L
thresholds, it is assumed that there are no superpositions, and no separation
any of the L thresholds, it is assumed that there are no superpositions, and no separation is is performed
for the current
performed forANGPD cloud.
the current The loop
ANGPD of thresholds
cloud. The loop is ofthen repeated
thresholds is for
thentherepeated
other ANGPD
for the
clusters of the PRPD,
other ANGPD one
clusters ofat
thea time.
PRPD, one at a time.

Matrix D

(a) (b) (c)

Matrix Matrix Matrix

Bounding
box

W
(d) (e) (f)

(a)PRPD
Figure8.8. (a)
Figure PRPDafter
afterremoval
removalof ofnon-dominant
non-dominantANGPDs
ANGPDsand and(b)
(b)pixel
pixelclustering
clusteringonon its
its pixels.
pixels.
(c)Matrix
(c) MatrixDDrelative
relativetotothe
thepositive
positive ANGPD
ANGPD cluster;
cluster; superposed
superposed horizontal
horizontal clouds
clouds are are indicated
indicated by
by arrows. Pixel clustering on D
arrows. Pixel clustering on D L matricesL matrices reconstructing D (loop of thresholds L) for (d) L =
reconstructing D (loop of thresholds L) for (d) L = 0.70, 0.70,
(e)LL==0.30
(e) 0.30andand(f)
(f)LL==0.20.
0.20.

The horizontal
The horizontal clusters
clusters CCsep detected in
sep detected in the
the loop
loop of of thresholds
thresholds underestimate
underestimate the the
dimensionsof
dimensions ofsuperposed
superposedN-ANGPD
N-ANGPDclouds. clouds. For Foraamore
moreaccurate
accurateestimation,
estimation,the thebounding
bounding
box of
box of each
each CCsep cluster was
sep cluster was “expanded”
“expanded” using using thethe following
following procedure,
procedure, illustrated
illustrated in in
Figure99for
Figure forthe
thepattern
patternof of Figure
Figure 8. 8. Such
Such expansion
expansion occurs
occurs iteratively
iteratively in in matrix
matrix D, in the
D, in the
verticaland
vertical andhorizontal
horizontaldirections,
directions,while
whilethe thePDsPDswithin
withinthe theexpanded
expanded BB BB are
are distributed
distributed
similarlyto
similarly tothe
thePDsPDsbounded
boundedby byCCsep ’soriginal
sep’s originalBB. BB.For
Foreach
eachcolumn
columnof of CCsep ’s original
sep’s original BB,
BB,
onecalculates
one calculatesthe theproportion
proportionof of the
the non-zero
non-zero elements
elements (p (pcc)) relative
relative toto the
the total
total number
number
ofpixels
of pixelsin inthe
thecolumn
columnand andthetheaverage
averageof ofthe
therow
rowcoordinates
coordinatesof of the
the non-zero
non-zero elements
elements
weighted by
weighted by its
its values
values inin matrix
matrix D D (y(ycc),), as
as shown
shown in in Figure
Figure 9b. 9b. The
The averages
averages of of those
those
metricsover
metrics overallallthe
thecolumns
columnsof ofthe
theoriginal
originalBB BBareare ppcc and
and yycc.. In
In the
the horizontal
horizontal direction,
direction,
theBB’s
the BB’sleft
leftbound
boundisisshifted
shiftedtotothe
theleft,
left,one
onecolumn
columnat ataatime,
time,while
whilethe the relative
relative differences
differences
between the p and y values in the new column and p and
between the pcc and ycc values in the new column and pcc and ycc are less than 75% and y are less than 75% and 40%,
40%,
respectively. The
respectively. Thesame
sameisisperformed
performedto tothe
theBB’sBB’sright
rightbound.
bound. An Ananalogous
analogous procedure
procedure was was
Energies 1, 0326
2021,15,
Energies2022, 12
12of
of26
26

then applied in the vertical direction. The pixels within the expanded BB were estimated to
then applied in the vertical direction. The pixels within the expanded BB were estimated to
form the actual N-ANGPD cluster (pixels in green in Figure 9a).
form the actual N-ANGPD cluster (pixels in green in Figure 9a).

Binarized matrix D

... Legend
Column-wise
Initial cluster (C sep) horizontal
expansion
Expanded cluster Averages: Pixels of initial
cluster (C sep)
(a) (b)

(a)PDs
Figure9.9.(a)
Figure PDsofofhorizontal
horizontalcloud
cloudsuperposed
superposedonto
ontoANGPD
ANGPDcloud
cloudbefore
before(red)
(red)and
andafter
after(green)
(green)
expansion. (b) Column-wise expansion of a generic N-ANGPD cloud along the horizontal direction.
expansion. (b) Column-wise expansion of a generic N-ANGPD cloud along the horizontal direction.

Theproposed
The proposedseparation
separationalgorithm
algorithmdeals dealswithwithtwo twotypes typesof ofsuperposition.
superposition. Case Case1,1,
common in the database and illustrated in Figure 10a (original
common in the database and illustrated in Figure 10a (original PRPD shown in Figure 3e), PRPD shown in Figure 3e),
consists of a wide horizontal group above the ANGPD
consists of a wide horizontal group above the ANGPD cloud, both connected by noisy cloud, both connected by noisy
pixelsof
pixels oflow/intermediate
low/intermediatedensity. density.In Incase
case22(Figure
(Figure8a), 8a),therethereisisan anactual
actualsuperposition
superposition
between the horizontal N-ANGPD
between the horizontal N-ANGPD and ANGPD clouds. and ANGPD clouds.
Eachcase
Each caseisistreated
treateddifferently.
differently.Thus, Thus,ititisisnecessary
necessaryto toidentify
identifyto towhich
whichcategory
categorythe the
superposition between ANGPD and N-ANGPD clouds
superposition between ANGPD and N-ANGPD clouds belongs to. In the region bounded belongs to. In the region bounded
bythe
by thesuperposed
superposedcluster’s cluster’sBB, BB,0.08A
0.08Aequally
equallyspacedspacedcolumns columnsof of elements
elements of of matrix
matrix D D
(vertical profiles) were sampled along the phase axis, as
(vertical profiles) were sampled along the phase axis, as shown in Figure 10b. The point- shown in Figure 10b. The point-
wiseaverage
wise averageofofthose those0.08A
0.08Acurvescurves resulted
resulted in in
thethe average
average vertical
vertical profile,
profile, illustrated
illustrated in
in Figure 10c. The average vertical profile can be seen
Figure 10c. The average vertical profile can be seen as a function ν describing the average as a function ν describing the
average distribution
distribution of PD counts of PD counts
across across amplitudes.
amplitudes. For case 1, For thiscase 1, this should
function functionhave, shouldin
have, in the direction from the ANGPD cloud to
the direction from the ANGPD cloud to N-ANGPD discharges, a sequence of a local N-ANGPD discharges, a sequence of a
local maximum,
maximum, minimum, minimum, and maximum,
and maximum, respectively,
respectively, associated associated to the ANGPD
to the ANGPD clouds,clouds,
noise,
noise, and horizontal PDs (Figure 10c). In this study,
and horizontal PDs (Figure 10c). In this study, the local minima and maxima were foundthe local minima and maxima were
found by iteratively sliding a 0.02A-element-wide window
by iteratively sliding a 0.02A-element-wide window along the average vertical profile. The along the average vertical
profile.point
central The central point
is a local is a localifmaximum
maximum if it is the
it is the largest elementlargest element
inside the inside
window, theor window,
it is a
local minimum if it is the smallest. Moreover, since horizontal PDs span a phase intervala
or it is a local minimum if it is the smallest. Moreover, since horizontal PDs span
phase interval
comparable comparable
to the ANGPD cloud’s to the ANGPD width incloud’scase 1,width in case 1,between
little variation little variation between
the individual
the individual vertical profiles was expected. This variation
vertical profiles was expected. This variation was quantified inversely by the similarity Svp , was quantified inversely
by the similarity
calculated Svp , calculated
as the average as the average between
of the cross-correlations of the cross-correlations between all the
all the pairwise combinations of
pairwise combinations of vertical profiles. Based
vertical profiles. Based on this information, the superposition was considered on this information, the superposition
to belong
was
to caseconsidered
1 if all the to following
belong to case criteria1 ifwere
all the following
met: (i) N-ANGPD criteria were cloudmet: (i) N-ANGPD
is horizontal; (ii)
cloud is horizontal;
Svp ≥ 0.85; and (iii) the average(ii) S vp ≥ 0.85; and (iii) the average vertical
vertical profile presents exactly three local extrema, profile presents exactly
one
three local
minimum extrema,
between two one minimum
maxima, and thatbetween two maxima,
the minimum has a andvaluethat lessthe minimum
than 1% of thehas firsta
value less than
maximum’s. 1% of thesuperposition
Otherwise, first maximum’s. was Otherwise,
considered superposition
to be of case 2.was considered to be
of case 2.
The following explanations were described based on the x and y axes (Figures 10
and 11).TheThe following
x-axis isexplanations
horizontal, oriented were described
left-to-rightbased in theon thePRPD.x and axes (Figures
They y-axis is vertical, 10
and 11). The x-axis is horizontal, oriented
oriented from the bottom to the top of the superposed cluster. left-to-right in the PRPD. The y-axis is vertical,
oriented
In thefrom case-1thesuperposition,
bottom to the top let ofyNU theandsuperposed
yND be the cluster.
y coordinates of the upper and
In the case-1 superposition,
lower bounds of the intermediate noisyNU let y and y be
region, respectively
ND the coordinates
y (Figure 10b). of Duethetoupper and
the small
lower bounds of the intermediate noisy region, respectively
number of discharges in this zone, the average vertical profile function ν(y) assumes lower (Figure 10b). Due to the
small number
images of discharges
in this region (rows [yin this zone, the average vertical profile function ν(y) assumes
ND , yNU ]) than in the other two regions, which are associated
lower images in this region
with the ANGPD and horizontal clouds (rows , yNU ])10c).
[yND(Figure thanItinisthe other two
considered thatregions,
yNU matcheswhich the are
associated with the ANGPD and horizontal clouds (Figure
lower bound of the horizontal cloud’s expanded BB. Moving in the direction of decreasingNU 10c). It is considered that y y,
matches
the first rowtheatlowerwhich bound
ν(y) ≥ of νthe
(yNU horizontal cloud’sas
) was estimated expanded
the coordinate BB. MovingyND . Thein the direction
coordinates
yof decreasing y, the first row at which ν(y) ≥ ν(yNU ) was estimated as the coordinate yND .
NU and yND are illustrated in Figure 10b,c. Additionally, let x L and x R be the x coordinates
The coordinates yNU and yND are illustrated in Figure 10b,c. Additionally, let x L and x R
Energies 2022, 15, 326 13 of 26

Energies 2021, 1, 0 13 of 26

of the left and right bounds of the ANGPD cloud, and let y B be the y coordinate of this
cluster’s lower bound (Figure 10b).
be the x coordinates of the left and right bounds of the ANGPD cloud, and let y B be the y
Next, one calculates the vectors FL and FR , containing the x coordinate of the cluster’s
coordinate of this cluster’s lower bound (Figure 10b).
leftmost and rightmost non-zero pixel in each row, respectively, (red and green curves in
Next, one calculates the vectors FL and FR , containing the x coordinate of the cluster’s
Figure
leftmost 10b).
andThe left and
rightmost right contours
non-zero of the
pixel in each ANGPD
row, cloud(red
respectively, were and estimated
green curves by fitting
in
parabolas to certain points of the
Figure 10b). The left and right contours of Lthe ANGPDcurves F and FR using the least squares
cloud were estimated by fitting method. Such
parabolas
parabolasmust be convergent—that
to certain points of the curves is, they must
FL and FRintersect
using theatleast a point whose
squares x coordinate
method. Such is
within [ x L , must
parabolas x R ]—and the intersectionis,point
be convergent—that theydetermines
must intersect theatANGPD
a point whose cloud’sx uppercoordinate bound. is It
iswithin
estimated[ x L , xthat
R ] the
—and ANGPD
the cloud’s
intersection peak
point lies next
determines to thethe central
ANGPD point (
cloud’sx M , y
upper
M ) of the
bound. noisy
region, where xthat
It is estimated M =the + x R )/2cloud’s
( x LANGPD and y Mpeak = (ylies
NU + yND
next to)/2.
the Parabolas
central point were( x Mfitted
, y M ) to
of the
points of FLregion,
the noisy and FRwhere corresponding
x M = ( x L to+ xthe
R ) set
/2 of
and rows
y M =[ y ( ,
y
B NUy T ]+, where
y ND ) y
/2. T was initially
Parabolas equal
were
tofitted
y M . to
If parabolas
the points are of Fnot
L and convergent, new fitting
FR corresponding to theattempts
set of rows are [performed
y B , y T ], where by iteratively
y T was
initially equal
decreasing to y M
y T , one . If parabolas
row are not
at a time, until theconvergent,
convergence new offitting
parabolas attempts are performed
is achieved. If there is
by iteratively decreasing
no y B < y T ≤ y M such that the y , one row at a time, until the convergence
T parabolas intersect at a point above the expanded horizontal of parabolas is
achieved. If there
cloud’s lower bound, the is no y y
B fitting
< T ≤ y
was such that the parabolas intersect
M not adequate, because it violates the hypothesis ofat a point above the
noexpanded
superpositionhorizontal cloud’sthe
between lower bound,
clouds. In the
thisfitting
case, was
newnot adequate,
parabolas are because
fitted to it violates
the points
the hypothesis of no superposition between the clouds.
of FL and FR corresponding to the set of rows [y B , yND ], and also to the point ( x M In this case, new parabolas are
, y M ).
fitted to the points
The addition of this point of FL and F
tends corresponding to the set
R to induce the expected convergence of rows [ y B NDof parabolas. the
, y ] , and also to In the
point ( xof
example M , Figure
y M ). The 10d, addition of this point
the parabolas tends to induce
were obtained in the firstthe expected
fitting attempt convergence of
to the points
parabolas. In the example of Figure 10d, the parabolas were obtained in the first fitting
of FL and FR of the rows [y B , y M ]. Separation was performed by assigning to the ANGPD
attempt to the points of FL and FR of the rows [y B , y M ]. Separation was performed by
cloud the PDs below the fitted parabolas, and the horizontal cloud was composed of the
assigning to the ANGPD cloud the PDs below the fitted parabolas, and the horizontal
PDs within its expanded bounding box (Figure 10e).
cloud was composed of the PDs within its expanded bounding box (Figure 10e).

Region of matrix D
Vertical profiles
y
... y

Horizontal
Expanded cloud
horizontal cloud
yNU FL (xM,yM) yNU
FR Noisy
x PDs
yND
ANGPD yND
yB cloud
x PD count
xL xR

(a) (b) (c)

Region of matrix D
Points considered for fitting
Horizontal bounding box

FL FR
Fitted
parabolas Fitted
parabolas

Separated N-ANGPDs
Separated ANGPDs
(d) (e)

Figure10.
Figure Separation of
10. Separation of case-1
case-1 superposition
superpositionbetween
betweenANGPD
ANGPD andandhorizontal clouds.
horizontal (a) PRPD
clouds. (a) PRPD
after the removal of non-dominant ANGPDs. (b) Region of matrix D relative to
after the removal of non-dominant ANGPDs. (b) Region of matrix D relative to the positive the positive cluster.
cluster.
(c) Average vertical profile. (d) Fitting of parabolas on the cluster’s left and right contours (F and FR
(c) Average vertical profile. (d) Fitting of parabolas on the cluster’s left and right contoursL (FL and FR
curves). (e) Separated ANGPD (green pixels) and horizontal (purple) clouds.
curves). (e) Separated ANGPD (green pixels) and horizontal (purple) clouds.
Energies
Energies 0 326
2022,1,15,
2021, 14 of 26 14 of 2

5-element wide window centered in T

B
Divergence
Trios of local
T extrema
Maximum within
the window T
centered in it
B B

Divergence
(a) (b)
T

Fitted
B
parabolas

T
T

B B

x Divergent points
Fitted parabolas T Separated N-ANGPDs
Separation regions
Separated ANGPDs
(c) (d)

Figure11.
Figure Separation
11.Separation of case-2
of case-2 superposition
superposition between
between N-ANGPDN-ANGPD and ANGPD
and ANGPD clouds.
clouds. (a) Region(a)
of Region o
matrixDDrelative
matrix relative to the
to the positive
positive cluster.
cluster. (b) Identification
(b) Identification of localofmaxima
local maxima (T) and
(T) and local local(B)
minima minima
in (B) in
the F and F curves. (c) Identification of divergent points. (d) Separated ANGPD (green pixels)
the FL and FR curves. (c) Identification of divergent points. (d) Separated ANGPD (green pixels) and
L R and
N-ANGPD
N-ANGPD clouds
clouds(purple).
(purple).
For case-2 superpositions, the FL and FR curves of the cluster were also calculated in
For case-2 superpositions, the FL and FR curves of the cluster were also calculated in
matrix D (Figure 11a). In the absence of disturbances, FL and FR tend to vary monotonically
matrix D (Figure 11a). In the absence of disturbances, FL and FR tend to vary monotonically
along the x-axis. When there are superposed horizontal clouds, however, those curves
along the x-axis. When there are superposed
experience sharp variations along x in the opposite direction to what horizontal clouds, however,
is expected those curve
(“diver-
experience
gences”), sharpinvariations
as shown Figure 11b.along x in the the
Disregarding opposite
points ofdirection
FL and Fto what is expected (“diver
R associated with the
gences”), as shown in Figure 11b. Disregarding
divergences caused by superposed horizontal clouds (diverging points), the points of F L and itFRis associated
possible to with th
divergences
estimate causedcloud’s
the ANGPD by superposed
left and right horizontal clouds (diverging
contours without superpositions. points), it is possible to
estimate
Figurethe 11 ANGPD
illustratescloud’s left and
the separation ofright
case-2contours without
superpositions forsuperpositions.
the positive ANGPD
clusterFigure
of the pattern of Figure
11 illustrates the8a. One obtains
separation the local
of case-2 maxima (tops) for
superpositions andthe minima (bot-ANGPD
positive
toms) of F and
cluster ofL the patternFR curves (Figure 11b). In the
of Figure 8a. One obtains FL and Fthecurves, one finds the local
R local maxima (tops) and minima (bo maxima
(tops)
toms)and of F minima
and (bottoms)
curves within
(Figure the11b).
2-point
In distance.
the andForF more robustness against noise,
L F R FL R curves, one finds the local maxim
one considers only the pairs of consecutive local minimum and local maximum whose x
(tops) and minima (bottoms) within the 2-point distance. For more robustness against noise
values differ by at least nine pixels. Each sequence of three local extrema (top-bottom-top
one considers only the pairs of consecutive local minimum and local maximum whose x
for FL , or bottom-top-bottom for FR ) was assumed to be related to the superposition with a
values differ
horizontal cloud.byTheat least nine pixels.
superposed Eachclouds
horizontal sequence of three
may have counts local extrema
different (top-bottom-top
or similar to
for F
the ANGPD
L , or bottom-top-bottom for F )
cloud’s in the region of intersection.
R was assumed
In Figure 11a, for example, the superposed with
to be related to the superposition
horizontalPDs
horizontal cloud.
to theThe superposed
right (purple arrow)horizontal
and toclouds
the leftmay
(yellowhave counts
arrow) different
were or similar to
of the first
the the
and ANGPD secondcloud’s in the region
type, respectively. ofdiverging
The intersection. In of
points Figure
FL and 11a,
FR for
wereexample,
determined the superposed
for
horizontal
each triplet ofPDs localtoextrema.
the right (purple arrow) and to the left (yellow arrow) were of the firs
andFor theeach horizontal
second cloud of distinctive
type, respectively. count (expanded
The diverging pointscluster
of FL Candexp ),Fdetected during
R were determined fo
the
eachloop of thresholds,
triplet one verifies to which sides it causes divergence in the ANGPD
of local extrema.
cloud.For Leteach and yAD becloud
yAU horizontal the yof coordinates
distinctiveofcountthe upper and lower
(expanded clusterbounds
Cexp ),ofdetected
Cexp ’s during
BB (Figure 11a). In FL and FR , it was verified if there is a triplet whose first and third
the loop of thresholds, one verifies to which sides it causes divergence in the ANGPD
local extrema have y coordinates less than yAD and greater than yAU , respectively. In
cloud. Let y and yAD be the y coordinates of the upper and lower bounds of Cexp ’
Figure 11c, for AU example, this applies to the triplet in FR . If such a triplet exists in FR , and if
BB (Figure 11a).
the element Pblw (yAD In−F1,L xand) FofR ,Fit immediately
was verified if there
below is a triplet
the horizontal whose
cloud has xfirst
valueand third
blw R
local
(x extrema have y coordinates less than y points
blw ) between the left and right bounds of Cexp , the AD
andof greater than yAUto, the
FR corresponding respectively.
rows In
Figure 11c, for example, this applies to the
[yAD , yAU ] are divergent (Figure 11c). Analogous logic applies triplet in F . If such
R to FL . a triplet exists in F R , and i
the element Pblw (yAD − 1, xblw ) of FR immediately below the horizontal cloud has x valu
(xblw ) between the left and right bounds of Cexp , the points of FR corresponding to the row
[yAD , yAU ] are divergent (Figure 11c). Analogous logic applies to FL .
The other triplets of local extrema may be related to superpositions of horizonta
clouds not detected during the loop of thresholds. The following procedure was performed
Energies 2022, 15, 326 15 of 26

The other triplets of local extrema may be related to superpositions of horizontal

clouds not detected during the loop of thresholds. The following procedure was performed
for each of those triplets. Let F be the curve FL or FR to which the triplet in question belongs,
and let (y E1 , x E1 ) be the coordinates of the first local extremum in F (Figure 11c). Starting
from this extremum and moving in the direction of increasing y, one looks for the first point
(y0 , x 0 ) of F such that x 0 > x E1 if F = FL , or x 0 < x E1 if F = FR . The point (y0 , x 0 ) is shown
in Figure 11c. The determination of this point follows the assumption that, in the direction
of increasing y, once one gets past the divergence, FL ’s x values are again greater than (or
less than if F = FR ) x E1 , an indication that the ANGPD cloud resumed its natural trend of
narrowing in the phase axis. In order to reduce the influence of noise, the points of F in
the rows ]y E1 , y0 [ were considered divergent only if all the following criteria were met: (i)
there are at least two points of F between the first and third local extrema of the triplet; and
(ii) the magnitude of the difference between x E1 and the median of the x values of F in the
interval ]y E1 , y0 [ is greater than or equal to 12. In Figure 11c, the triplet of local extrema of
FL meets these conditions.
Using least squares, parabolas were fitted to the FL and FR curves disregarding the
divergent points, as shown in Figure 11c. For each group of divergent points of the curve
F (FL or FR ), the PDs of y coordinates equal to the divergent points and not contained by
the parabola fitted to F (to the left of the parabola if F = FL , or to the right if F = FR ) were
removed from the ANGPD cluster and assigned to a new N-ANGPD cluster. The hatched
areas in Figure 11c indicate the regions in which PDs were separated from the ANGPD
cloud. Figure 11d shows the separation of superposed horizontal and ANGPD clouds.
At this moment, the N-ANGPD clouds superposed onto a single ANGPD cluster were
separated. The entire separation process (the calculation of matrix D, the loop of thresholds,
and so on) was repeated for the PRPD’s other ANGPD clusters, one at a time. Finally, the
pixel coordinates of all the separated clusters were mapped back to the PRPD’s original
dimensions A × A. The values of thresholds described in this subsection were defined
empirically, in order to separate most of the superposed patterns in the database.
Superposed tracking PDs could be separated in a way analogous to the case-2 super-
position. Instead of FL and FR , the procedure would be based on curves containing the
y coordinates of the cluster’s uppermost and lowermost non-zero pixel at each column.
However, due to the low occurrence of this type of superposition in the database, the
separation of tracking clouds was not implemented in this study.

3.4. PRPD Decomposition

The separation of multiple sources was the last of a sequence of four steps focused on
removing as much noise as possible from ANGPD clouds, producing a partially denoised
pattern M1 (Figure 6). There was a second filtering path tailored for preserving low-
density N-ANGPD discharges (pattern M2 ). In the current step of PRPD decomposition,
illustrated by the namesake block in blue of Figure 6, M1 and M2 were combined to form
two sub-PRPDs.
Nearest-neighbor clustering [24] was applied separately to M1 and M2 , using a subma-
trix of 0.05A×0.05A pixels, the same dimensions adopted in [18]. Figure 12a,b illustrates
the results of clustering on the partially denoised versions M1 and M2 of the PRPD of
Figure 3d. Discharges of each cluster are shown in a different color.
PDs of M1 and M2 were combined to produce two sub-PRPDs: one containing only
ANGPD and the other only N-ANGPD discharges (Figure 12c,d). The ANGPD sub-PRPD
was composed of M1 ’s ANGPD clusters. The N-ANGPD sub-PRPD was formed of the
union of PDs belonging to M1 ’s and M2 ’s N-ANGPD clusters.
 Energies2022,
Energies 1, 0326
2021,15, 1616
ofof
2626
Energies 2021, 1, 0 16 of 26

Figure 12.
Figure Pixel clustering
12. Pixel clustering onon aa PRPD
PRPD treated
treated with
with(a)(a)grid
gridfiltering
filteringand
andremoval
removalofofnon-dominant
non-dominant
Pixel clustering
Figure 12.(pattern on apixel
PRPD treated with (a) grid filtering and removal of non-dominant
ANGPDs (pattern M1 ) and (b) pixel submatrix filtering (M2 ). From M1and
ANGPDs M 1 ) and (b) submatrix filtering (M 2 ). From M 1 2 , ,the
andMM 2 thepattern
patternwas
was
ANGPDs (pattern
decomposed into the 1 ) and
M(c) (b) pixel
ANGPD and submatrix
(d) N-ANGPD filtering (M2 ). From M1 and M2 , the pattern was
sub-PRPDs.
decomposed into the (c) ANGPD and (d) N-ANGPD sub-PRPDs.
decomposed into the (c) ANGPD and (d) N-ANGPD sub-PRPDs.
Applying the
Applying the described
described denoising
denoising algorithms
algorithmson
onthe
thepatterns
patternsofofFigure
Figure3a,b,f
3a,b,fproduced
produced
Applying thein
described denoising algorithms on the patterns of Figure 3a,b,f produced
the results shown in Figure 13. It can be observed that most of noise was eliminated,and
the results shown Figure 13. It can be observed that most of noise was eliminated, and
the results shown
simultaneous in Figure
PD sources 13.separated.
were It can be observed that most of noise was eliminated, and
simultaneous PD sources were separated.
simultaneous PD sources were separated.
ANGPD sub-PRPDs
ANGPD sub-PRPDs
Amplitude
Amplitude

0º Phase 360º 0º Phase 360º 0º Phase 360º

0º Phase 360º 0ºN-ANGPDPhase

sub-PRPDs 360º 0º Phase 360º
N-ANGPD sub-PRPDs
Amplitude
Amplitude

0º Phase 360º 0º Phase 360º 0º Phase 360º

(a) (b) (c)
0º Phase 360º 0º Phase 360º 0º Phase 360º
Figure 13. Filtering
(a) results of the proposed
(b) image denoising algorithm
(c) applied to PRPDs of the
classes
Figure (a)
13.corona, (b) results
Filtering slot, and
of(c)
theunidentified.
proposed image denoising algorithm applied to
to PRPDs
PRPDs of
of the
the
Figure 13. Filtering results of the proposed image denoising algorithm applied
classes(a)
classes (a)corona,
corona,(b)
(b)slot,
slot,and
and(c)
(c)unidentified.
unidentified.
3.5. Classification of N-ANGPD PD Sources
3.5. Given the characteristic
3.5.Classification
Classification ofofN-ANGPD
N-ANGPD format
PD of gap and surface tracking sources [15], in this study,
PDSources
Sources
thoseGiven
types theof PD were identified
characteristicformat
formatofby directly
ofgap
gapand comparing
andsurface the dimensions
surfacetracking
tracking [15],and
sources[15], relative
inthis
this study,
Given the characteristic sources in study,
positions
those of
types the
of sub-PRPD’s
PD were N-ANGPD
identified by clouds
directly against predetermined
comparing the thresholds,
dimensions and without
relative
those
using types of intelligence
artificial PD were identified
(AI). by directly comparing the dimensions and relative
positions
positions ofthe
of thesub-PRPD’s
sub-PRPD’s N-ANGPDclouds
N-ANGPD cloudsagainst
againstpredetermined
predetermined thresholds,without without
using The BB of each
artificial cluster of
intelligence N-ANGPD sub-PRPD
(AI). was determined; itsthresholds,
width and height
using artificial
were The
W and Hintelligence
pixels, (AI).
respectively (Figuresub-PRPD
12c). Onewas considers the presence of and
tracking
BBof of each cluster of N-ANGPD determined; itswidth
width height
activity in a pattern if thereofisN-ANGPD
The BB each cluster any clustersub-PRPD
whose BBwas meetsdetermined;
H ≥ 0.0742A, its W ≥and height
0.0078A,
were
were W and H pixels, respectively (Figure 12c). One considers the presence of tracking
H/WW≥and H pixels,
2.7 and respectively
the proportion (Figure 12c).
of non-zero pixels One considers
within the BBthe presenceofofdensity)
(a measure tracking
activityin
activity inaa pattern
pattern ifif there
there is
is any
any cluster
cluster whose
whose BB BB meets
meets H H ≥
≥ 0.0742A,
0.0742A, W W ≥≥ 0.0078A,
0.078A,
is greater than 12.6%. The gap discharge source is detected if there is a pair of horizontal
H/W ≥≥ 2.7
H/W 2.7andandthe theproportion
proportionof ofnon-zero
non-zeropixelspixelswithin
withinthetheBB
BB(a (ameasure
measure of of density)
density)
clouds, one in each polarity, of approximately the same magnitude amplitude (maximum
isgreater
is greaterthan
than12.6%.
12.6%.The Thegap
gap dischargesource sourceisisdetected
detectedififthere
there isaapair
pairofofhorizontal
horizontal
16% relative difference betweendischarge
the clouds’ centers amplitudes) andisapproximately 180◦
clouds,
clouds, one in
one each each
in each polarity,
polarity, of approximately the same magnitude amplitude (maximum
apart from other in theofphase
approximately
axis (phasethe same magnitude
distance amplitude
between clouds’ centers(maximum
within ◦
16% relativedifference
16% difference betweenthe theclouds’
clouds’centers
centersamplitudes)
amplitudes)and and approximately180 180◦
180◦ relative
± 23%). A cluster isbetween deemed horizontal if its BB meets H ≥ 0.0391Aapproximately
and W/H ≥ 1.56.
apart
apart from
from eacheach other
otherofin in the phase
the phasewere axis (phase
axisoptimized distance
(phase distance between
between clouds’ centers
clouds’ centers within
within
The◦informed values thresholds to maximize the system’s classification
180◦ ±±23%).
180 23%).AAcluster
clusterisisdeemed
deemedhorizontal
horizontalififits itsBB
BBmeets
meets H H≥ 0.0391Aand
≥0.0391A andW/HW/H ≥ 1.56.
≥ 1.56.
rates of surface tracking and gap PD sources, as discussed in Section 4.
The informed values of thresholds were optimized to maximize the system’s classification
rates of surface tracking and gap PD sources, as discussed in Section 4.
Energies 2022, 15, 326 17 of 26

Energies 2021, 1, 0 17 of 26
The informed values of thresholds were optimized to maximize the system’s classification
rates of surface tracking and gap PD sources, as discussed in Section 4.

3.6. Extraction of Contour Features

3.6. Extraction of Contour Features
As already mentioned, the PD-recognition task was divided into smaller problems
usingAstwo already mentioned,
input features: the PD-recognition
amplitude histograms [18] task
and was
thedivided into smaller
novel contour problems
attribute, which
using two input features: amplitude histograms [18] and the novel contour
quantify the PD symmetry and cloud contours, respectively. Those characteristics, besides attribute, which
quantify the PD symmetry and cloud contours, respectively. Those characteristics,
being mentioned in the literature to describe typical PRPDs of different sources [15,20], are besides
being
also mentionedrobust
considered in theto
literature
variations to describe typical PRPDs
in measurement of different
conditions [25]. sources [15,20], are
also considered robust to variations in measurement conditions
The contour attribute is a novel contribution of this work. This [25].feature quantifies the
shape of PD clouds, which is the main difference between InV and InDfeature
The contour attribute is a novel contribution of this work. This classesquantifies
and between the
shape of PD clouds, which is the main difference between InV and InD classes
slot and corona [15,20]. It is considered the contour of only the positive ANGPD cloud. The and between
slot and corona
negative cloud is[15,20].
ignoredIt is considered
because the contourfor
it is redundant of InV,
onlyInD,
the positive
slot, andANGPD
corona cloud.
sources.The
negative
Figurecloud is ignored
14 shows idealbecause it is redundant
triangular and rounded for ANGPD
InV, InD, clouds,
slot, andalongside
corona sources.
the first
Figure 14 shows ideal triangular and rounded ANGPD clouds,
derivative of the mathematical function describing the cloud’s contour. The distinctive alongside the first
derivative of the mathematical function describing the cloud’s contour.
variations of the first derivatives demonstrate that these curves may be used to efficiently The distinctive
variations
quantify theofshape
the first
of derivatives
PD clouds. PRPDsdemonstrate thatfrom
obtained theseonline
curvesmeasurements
may be used to efficiently
have clouds
with irregular and noisy contours, whose derivatives differ to a certain degree fromclouds
quantify the shape of PD clouds. PRPDs obtained from online measurements have those
with irregular
shown in Figure and14.noisy contours,
In this study,whose
neuralderivatives differfrom
networks learn to a certain degree
real data from those
the differences
shown inthe
between Figure 14. In this study,
first derivatives of bothneural networks
types of curves. learn from real data the differences
between the first derivatives of both types of curves.
Amplitude

Amplitude

0 0

Phase Phase
(a) (b)

Figure 14.
Figure First derivative
14. First derivative of
of contour
contour of
of ideal
ideal (a)
(a) triangular
triangular and
and (b)
(b) rounded
rounded positive
positive ANGPD
ANGPD clouds.
clouds.
The functions f (·) and g (·) describing contour shape are shown as triangular and parabolic functions
The functions f (·) and g(·) describing contour shape are shown as triangular and parabolic functions
for illustrative
for illustrative purposes.
purposes.

In order
In order toto obtain
obtain the the contour
contour feature,
feature, one
one first
first calculates
calculates the
the preliminary
preliminary contour
contour (C(CPP))
of the positive ANGPD cloud. For each phase angle between
of the positive ANGPD cloud. For each phase angle between this cloud’s phase bounds, this cloud’s phase bounds,
is equal
CPP is
C equal to
to the
the average
average of of the
the row
row coordinates
coordinates of of the
the six
six highest
highest non-zero
non-zero PDs
PDs inin that
that
column. C was subjected to a low-pass Butterworth filter [26] of
column. CPP was subjected to a low-pass Butterworth filter [26] of order 5 and normalized order 5 and normalized
cutoff frequency
cutoff frequency 0.1. 0.1. The
The first
first derivative
derivative ofof the
the filtered
filtered C was calculated,
CPP was calculated, which
which waswas
then smoothed
then smoothed by by aa second
second Butterworth
Butterworth filter
filter of
of order
order 22 and
and normalized
normalized cutoff
cutoff frequency
frequency
0.04. After
0.04. After sampling
sampling 32 32 equally
equally spaced
spaced points
points from
from the
the smoothed
smoothed derivative
derivative and
and applying
applying
min-max scaling to the [− 1, 1 ] range, one obtains the feature vector.
min-max scaling to the [−1, 1] range, one obtains the feature vector. The chosen length of The chosen length of
the feature vector provides a balanced compromise between contour
the feature vector provides a balanced compromise between contour resolution and neural resolution and neural
network complexity.
network complexity.
Figure
Figure 15 15 shows
shows the the contour
contour feature
feature calculated
calculated for for one
one InV
InV and
and one
one InD
InD sample
sample ofof
the database.
the database. Despite
Despite the the irregular
irregular ANGPD
ANGPD clouds,
clouds, representative
representative contours
contours (in
(in purple)
purple)
were extracted.
were extracted. It It is
is important
important to to note
note that
that features
features also
also resemble
resemble thethe derivatives
derivatives of of the
the
corresponding ideal contours (Figure 14). The informed values
corresponding ideal contours (Figure 14). The informed values of parameters used to of parameters used to
calculate C , its derivative, and smoothing filters were determined
calculate CP , its derivative, and smoothing filters were determined by parametric tuning.
P by parametric tuning.
Energies
Energies 1, 326
2021,15,
2022, 0 18 of 18
26 of 26

Energies 2021, 1, 0 18 of 26

Normalized
Normalized feature
feature
Preliminary contour (CP) Preliminary contour (CP)
Smoothed CP Smoothed CP
Preliminary
Refined phasecontour (CP)
bounds Preliminary
Refined phasecontour (CP)
bounds
Smoothed CP Smoothed CP
Index
Refined phase
(a) bounds Refined phase
(b) bounds (c)
Index
15. Extraction
Figure 15.
Figure Extraction
(a) of
ofcontours
contoursfrom
from(a)(a)
InV and
InV and
(b) (b)(b)
InD PRPDs
InD andand
PRPDs their (c) feature
their (c)(c) vectors.
feature vectors.
Figure
3.7. 15. Extraction
Identification of contours
ofofNoisy from (a) InV and (b) InD PRPDs and their (c) feature vectors.
3.7. Identification NoisyANGPD
ANGPDCloudsClouds
This step aimed
3.7. Identification to
of Noisy
This step aimed identify patterns
Cloudswhose
ANGPDpatterns
to identify whose twotwodominant
dominant ANGPD
ANGPD clouds werewere
clouds noisy, a a
noisy,
common situation in the database (Figure 2). Such a stage adds to the system the important
common Thissituation
step aimed into
the database
identify (Figure
patterns 2). Such
whose a stage adds
two dominant ANGPDto the system
clouds werethenoisy,
important
a
capability of identifying patterns not having a valid PD source, increasing its applicability
capability of identifying
common situation patterns(Figure
in the database not having a valid
2). Such PDadds
a stage source, increasing
to the system the itsimportant
applicability
in practical situations due to the reduction in false positives. Without this functionality,
in practicalofsituations
capability identifyingdue to thenot
patterns reduction
having a in false
valid PDpositives. Withoutitsthis
source, increasing functionality,
applicability
patterns with noisy ANGPD clouds would always be assigned to one of the ANGPD PD
in practical
patterns with situations
noisy due to the
ANGPD reduction
clouds would inalways
false positives.
be Without
assigned to onethis
of functionality,
the ANGPD PD
sources by neural networks.
patternsby
sources with noisynetworks.
neural ANGPD clouds would always be assigned to one of the ANGPD PD
Figure 16 shows typical cases of patterns with no valid ANGPD clouds. The most
sources by neural
Figure 16 showsnetworks.
typicalofcases of patterns with
common situations are clouds low density and/or lowno validThe
height. ANGPD clouds.
other cases The most
are clouds
common Figure 16 shows
situations are typical
clouds cases
of lowof density
patternsand/or
with no lowvalid ANGPD
height. The clouds.
other Theare
cases most
clouds
whose height and/or density are very similar to those of PDs in other phases (it is expected
common
whose situations
height and/or are clouds
density of
arelow density
very and/or
similar to low
those height.
of PDs The
in other
other cases
phases are
(it clouds
is expected
that such characteristics of valid ANGPD clouds are higher than in other phases, which are
whose
that height
such and/or density
characteristics are very similar to those
areof PDs inthan
otherinphases (it is expected
presumably occupied by of valid
noise ANGPD
only). clouds higher other phases, which are
that such characteristics of valid ANGPD clouds are higher than in other phases, which are
presumably occupied by noise only).
presumably occupied by noise only).

Figure 16. Examples of patterns with invalid ANGPD clouds. Such clouds present (a) low height,
(b) low density, (c) low density compared to PDs in other phase angles, (d) low height compared to
16. Examples
Figure 16. Examples of
ofpatterns
patternswith invalid ANGPD clouds. Such clouds present (a) low height,
Figure
PDs in other phase angles. Arrowswith
pointinvalid ANGPD
to relevant ANGPD clouds. Such
clouds. clouds present (a) low height,
(b) low density, (c) low density compared to PDs in other phase angles, (d) low height
(b) low density, (c) low density compared to PDs in other phase angles, (d) low height comparedcompared to to
PDs
PDs in in other
These phase
othersituationsangles.
phase angles. Arrows
areArrows point
identified to relevant
bytocapturing
point ANGPD clouds.
the following
relevant ANGPD clouds. attributes of a pattern’s
ANGPD clouds: height (H), density (Dty ), the height–width ratio (H/W), contour variance
These situations are identified by capturing the following attributes of a pattern’s
(var),These situations
signal–noise ratiosare
of identified
height (SNR by ),capturing
and density the(SNRfollowing
Dty ). attributes of a pattern’s
ANGPD clouds: height (H), density (DtyH), the height–width ratio (H/W), contour variance
ANGPD The clouds: height (H),
characteristics weredensity
extracted (Dty ), therough
from height–width
contour Cratio (H/W),
R (Section 3.2)contour variance
[18], which
(var), signal–noise ratios of height (SNRH ), and density (SNRDty ).
(var), signal–noise
was calculated withratios of height
specific values (SNR
of g andH ), on
anddifferent
densityversions
(SNRDtyof ). the pattern. CR was
The characteristics were extracted from rough contour CR (Section 3.2) [18], which
calculated with g = 0.0117A
The characteristics wereforextracted
H, H/W, from and SNR H , and
rough with gC= (Section
contour 0.0391A for3.2)the other
[18], which
was calculated with specific values of g and on different versionsRof the pattern. CR was
attributes.
was Using
calculated the specific
with corresponding
values g,
ofthe
g rough
and on contour
different was calculated
versions of onpattern.
the PRPD after C R was
calculated with g = 0.0117A for H, H/W, and SNRH , and with g = 0.0391A for the other
grid filtering
calculated withforgSNR= H and SNR
0.0117A for H,
Dty . H/W,
For the other
and SNR attributes,
H , and C
with R was
g = calculated
0.0391A foron the
the other
attributes. Using the corresponding g, the rough contour was calculated on PRPD after
ANGPD
attributes. sub-PRPD.
Using
grid filtering for the
SNRcorresponding g, the rough contour was calculated on PRPD after grid
H and SNRDty . For the other attributes, CR was calculated on the
Before SNR
filtering explaining how attributes
. For theare calculated, it is
CRconvenient to present
on thesome
ANGPDfor sub-PRPD. H and SNR Dty other
∗ and P∗ be the columns of ANGPD cloud’s
attributes, was calculated ANGPD
definitions.
sub-PRPD. Let P L R left and right phase bounds
Before explaining how attributes are calculated, it is convenient to present some
(e.g., Before
verticalexplaining
dashed lines∗ in Figure 15), respectively. The height of PDs over atogiven phase
definitions. Let PL∗ and PRhow attributes
be the columnsare calculated,
of ANGPD cloud’sit isleft
convenient
and right phase present
bounds some
interval is the 80th ∗ percentile
∗ of the values of smooth contour CS (Section 3.2) within this
definitions.
(e.g., verticalLet PL and
dashed PRinbe
lines the columns
Figure of ANGPD
15), respectively. Thecloud’s
height ofleft and
PDs right
over phase
a given bounds
phase
interval. The PD density is the proportion of non-zero PDs within a given phase interval
(e.g., vertical
interval is thedashed lines in Figure
80th percentile 15), respectively.
of the values The height
of smooth contour of PDs over
CS (Section a given
3.2) within phase
this
and below the EMA-filtered rough contour (EMA(CR )–Section 3.2).
interval.isThe
interval thePD density
80th is the of
percentile proportion
the values of of
non-zero
smoothPDs withinCSa (Section
contour given phase
3.2)interval
within this
and below
interval. ThethePDEMA-filtered
density is therough contour (EMA
proportion (CR )–Section
of non-zero 3.2). a given phase interval
PDs within
and below the EMA-filtered rough contour (EMA(CR )–Section 3.2).
Energies 2022, 15, 326 19 of 26

Energies 2021, 1, 0 19 of 26

H and Dty are the height and density of PDs within the phase interval [ PL∗ , PR∗ ], respec-
tively. H/W is the cloud’s height divided by its width, which is equal to the smallest phase
H and Dty are the height and density of PDs within the phase interval [ P∗ , PR∗ ], respec-
distance between PL∗ and PR∗ . SNRH is the subtraction between the cloudL height and the
tively. H/W is the cloud’s height divided ∗by its ∗ ]).width, which is equal to the smallest phase
height of PDs in the∗ other phases (/
∈ [ P , P SNR
L subtraction
R Dty is the cloud density divided by the
distance between PL and PR∗ . SNRH is the between the cloud height and the
density of PDs in the other phases. var∗ is ∗the variance of CS values within phases [ PL∗ , PR∗ ]
height of PDs in the other phases (/ ∈ [ P , P ]). SNRDty is the cloud density divided by the
and min-max normalized to the rangeL[0, 1R].
density of PDs in the other phases. var is the variance of CS values within phases [ PL∗ , PR∗ ]
and From
min-maxthenormalized
ANGPD sub-PRPD,
to the rangea pair
[0, 1].of values was calculated for each attribute: one
from From
the positive
the ANGPD sub-PRPD, a pairthe
cloud and one from negative
of values wascloud. Thefor
calculated feature vector, which
each attribute: one is
presented to ANNs-0, is formed by the maximum of each of those pairs.
from the positive cloud and one from the negative cloud. The feature vector, which is
presented to ANNs-0, is formed by the maximum of each of those pairs.
3.8. Data Partitioning
3.8. Data
ANGPDPartitioning
sub-PRPDs of patterns with labeled ANGPD source (Figure 4) were randomly
divided into mutually exclusive
ANGPD sub-PRPDs of patternssubsets withANGPD
with labeled five-fold cross-validation
source (Figure 4) were(CV) [27]. CV
randomly
generates several
divided into data partitions,
mutually exclusive each of which
subsets is a selection
with five-fold of three folds
cross-validation for training:
(CV) [27]. CV one
generates
for several
validation anddata partitions,
the others each of All
for testing. which
the is
20apossible
selectionpartitions
of three folds
werefor training:
considered.
one for
CVvalidation
partitionsand thebyothers
used ANNs-1for testing. All the 20
were obtained bypossible
merging partitions wereofconsidered.
the subsets InV with InD
CVfold
of each partitions used byclass.
into a single ANNs-1Thewere
sameobtained
was done by for
merging the corona
slot and subsets patterns.
of InV with InD
Partitions
of each
for fold into
ANNs-2 a singleonly
consisted class.ofThe
thesame
InV was
and done for slot and
InD subsets corona
of the patterns. Partitions
corresponding ANNs-1’s
for ANNs-2
partitions, consisted
whereas only ofused
ANNs-3 the InV
onlyand
the InD
slot subsets
and corona of the corresponding
subsets. Figure 17ANNs-1’s
presents the
partitions, whereas ANNs-3 used only the slot and corona
data-partitioning schemes for the three types of neural networks. subsets. Figure 17 presents the
data-partitioning schemes for the three types of neural networks.

Database Training folds Validation fold Test fold

InV1 / InD1, InV2 / InD2, InV3 / InD3, InV4 / InD4, InV5 / InD5,
InV1, InV2, InV3, InV4, InV5, Slot1, Slot2, Slot3, Slot4, Slot5,
P1 Slot1 / Cor1, Slot2 / Cor2, Slot3 / Cor3, Slot4 / Cor4, Slot5 / Cor5,
InD1 InD2 InD3 InD4 InD5 Cor1 Cor2 Cor3 Cor4 Cor5
DCI1 DCI2 DCI3 DCI4 DCI5

InV1 / InD1, InV2 / InD2, InV3 / InD3, InV4 / InD4, InV5 / InD5,
InV1, InV2, InV3, InV4, InV5, Slot1, Slot2, Slot3, Slot4, Slot5,
P2 Slot1 / Cor1, Slot2 / Cor2, Slot3 / Cor3, Slot4 / Cor4, Slot5 / Cor5,
InD1 InD2 InD3 InD4 InD5 Cor1 Cor2 Cor3 Cor4 Cor5
DCI1 DCI2 DCI3 DCI4 DCI5
...
...

...
InV1 / InD1, InV2 / InD2, InV3 / InD3, InV4 / InD4, InV5 / InD5,
InV1, InV2, InV3, InV4, InV5, Slot1, Slot2, Slot3, Slot4, Slot5,
P20 Slot1 / Cor1, Slot2 / Cor2, Slot3 / Cor3, Slot4 / Cor4, Slot5 / Cor5, InD1 InD2 InD3 InD4 InD5 Cor1 Cor2 Cor3 Cor4 Cor5
DCI1 DCI2 DCI3 DCI4 DCI5
(a) (b) (c)
Legend: P: partition Cor: corona

Figure 17.
Figure Data-partitioning schemes
17. Data-partitioning schemesfor
for(a)(a)ANN-1,
ANN-1, (b)(b)
ANN-2,
ANN-2, andand
(c) (c)
ANN-3 neural
ANN-3 networks.
neural networks.
Labels within blocks indicate the PD sources of patterns allotted to each fold.
Labels within blocks indicate the PD sources of patterns allotted to each fold.
3.9. Training and Validation of Neural Networks
3.9. Training and Validation of Neural Networks
The four types of neural networks (ANN-0, ANN-1, ANN-2, and ANN-3) were
The four types of neural networks (ANN-0, ANN-1, ANN-2, and ANN-3) were
trained separately, using the procedure described in [22], which are briefly reviewed here.
trained separately, using the procedure described in [22], which are briefly reviewed here.
Different subsets of ANGPD sub-PRPDs were extracted from the database to train the
Different subsetsFor
neural networks. of ANGPD
ANNs-0, all sub-PRPDs
the ANGPD were extracted
sub-PRPDs werefrom the database
categorized in twotogroups
train the
neural networks. For ANNs-0, all the ANGPD sub-PRPDs were categorized
(labeled ANGPD source or not) and divided with five-fold cross-validation [27]. For in two groups
(labeled ANGPD source or not) and divided with five-fold cross-validation
ANNs-1 to ANNs-3, the ANGPD sub-PRPDs of patterns with labeled ANGPD source were [27]. For ANNs-
1divided
to ANNs-3, the ANGPD sub-PRPDs of patterns with labeled ANGPD source
according to Section 3.8. For each type of ANN, several networks were trained on were divided
according
the trainingtofold
Section 3.8. while
samples, For each
usingtype of ANN, several
the validation fold as annetworks were trained
early stopping criterionon to the
training
reduce overfitting [27]. The training algorithm Scaled Conjugate Gradient [28] was used. to
fold samples, while using the validation fold as an early stopping criterion
reduce overfitting
Once trained, [27]. The
networks’ training algorithm
generalization Scaled
was evaluated by Conjugate Gradient
the performance [28]
on the was
test used.
fold.
Once Several
trained,ANNs
networks’
were generalization
trained by varying wasthe
evaluated by the performance
hyperparameters’ initial randomon the test fold.
weights
and Several
topologyANNs wereoftrained
(numbers hiddenby varying
layers the hyperparameters’
and hidden neurons). The best initial random weights
combinations of
hyperparameters
and were those
topology (numbers with the
of hidden highest
layers andrecognition rates on the
hidden neurons). Thetest
bestfold.
combinations of
The computational
hyperparameters complexity
were those of the
with the proposed
highest methodology
recognition rates is
onofthe
interest. The stages
test fold.
of PRPD image denoising
The computational and multiple-source
complexity separation,
of the proposed classification
methodology of N-ANGPD
is of interest. The stages
sources,
of PRPDand extraction
image of one
denoising type
and of input feature separation,
multiple-source (histogram orclassification
contour) combined took
of N-ANGPD
98 milliseconds
sources, (ms), on of
and extraction average, for aoftypical
one type input PRPD.
featureA(histogram
single ANNorofcontour)
any of thecombined
four typestook
takes,
98 on average,
milliseconds 140on
(ms), msaverage,
to be trained
for a and 10 ms
typical to classify
PRPD. 100ANN
A single PRPDs of from their
any of thefeature
four types
vectors. Those metrics were measured on a low-end laptop (Intel® Core™ i5-4200U), and
takes, on average, 140 ms to be trained and 10 ms to classify 100 PRPDs from their feature
they are reasonable for most PD-recognition applications.
Energies 2021, 1, 0 20 of 26

Energies 2022, 15, 326 20 of 26

4. Results and Discussion

TheThose
vectors. trained neural
metrics networks
were measuredwereon aevaluated with the
low-end laptop metric
(Intel δ, proposed
® Core™ byand
i5-4200U), the
authors in [22]. A neural network classifies a given
they are reasonable for most PD-recognition applications. set of patterns, and the resulting
confusion matrix [27] was built. From the confusion matrix, one calculates the recalls,
which
4. Resultsareand
equal to the proportion of samples of each class that were correctly classified.
Discussion
We calculated the average
The trained neural µ Ts and
networks standard
were deviation
evaluated of recalls,
σTsmetric
with the which were
δ, proposed then
by the used
authors
to [22].
in calculate δ by network classifies a given set of patterns, and the resulting confusion
A neural
matrix [27] was built. From the confusion matrix, one 1 calculates the recalls, which are equal
to the proportion of samples of eachδclass = µ Tsthat .
+ were correctly classified. We calculated the (2)
1 + σTs
average µ Ts and standard deviation σTs of recalls, which were then used to calculate δ by
δ is a pattern-recognition metric that ranks best (the higher the better) those classifiers
with recalls that are high (large µ Ts ) and with low 1 variation (small σTs ). The definition
δ = µ Ts + (2)
based on recalls makes δ unbiased against the1database’s + σTs class distribution.
In this section, the performance statistics of the proposed multiple-source PD recog-
δ issystem
nition a pattern-recognition
are shown andmetric that ranks
discussed. best
In the (the higher
proposed the better)
system, four those
types classifiers
of neural
with recalls that are high (large µ Ts ) and with low variation (small
networks were used to decompose the classification task into smaller problems, σ Ts ). The definition
acting like
based
blockson ofrecalls makes
a larger δ unbiased
identification against the database’s class distribution.
system.
In this section, the performance statistics of the proposed multiple-source PD recog-
nition system areofshown
4.1. Optimization and for
Thresholds discussed. In theofproposed
the Recognition N-ANGPD system,
Sourcesfour types of neural
networks were used to decompose the classification task into
The first system block is the recognition of surface tracking and smaller problems, acting like
gap PD sources. A
blocks
simpleofclassification
a larger identification system.
was performed by comparing certain properties of N-ANGPD clouds
against fixed thresholds. The values of those thresholds were fine-tuned with the Particle
4.1. Optimization of Thresholds for the Recognition of N-ANGPD Sources
Swarm Optimization (PSO) technique [29].
The
The first system
detection block is the
thresholds recognition
of surface of surface
tracking and gaptracking and gap
clouds were PD sources.
optimized A
together,
simple classification was performed by comparing certain properties
in a single PSO simulation. The optimization ran on denoised N-ANGPD sub-PRPDs. The of N-ANGPD clouds
against
objectivefixed thresholds.
function The values
was calculated the of those thresholds
following way for eachwerePSO
fine-tuned
particle with the iteration.
in every Particle
Swarm Optimization (PSO) technique [29].
All clusters of each sub-PRPD were subjected to the comparison rules for vertical clouds of
The3.5
Section detection
but usingthresholds of surface
the threshold valuestracking
coded by andthegap clouds
particle in were optimized
question. together,
The sub-PRPDs
in a single PSO simulation. The optimization ran on denoised N-ANGPD
with at least one vertical cluster were deemed to have surface tracking, while the others sub-PRPDs. The
objective function was calculated the following way for each PSO particle
were not. The confusion matrix was calculated (Figure 18a). A second scan was performed in every iteration.
All clusters
over of each
sub-PRPDs sub-PRPD
using were subjected
an analogous procedure to the comparison
searching rules
for pairs offor vertical clouds
horizontal clouds of as
Section 3.5 but using the threshold values coded by the particle in question.
described in Section 3.5, resulting in the gap discharge confusion matrix (Figure 18b). The The sub-PRPDs
with at least
objective one vertical
function cluster were
was calculated deemed
as the to have
geometric mean surface
of the tracking, while
four recalls the others
obtained from
were not. The confusion
the two confusion matrices. matrix was calculated (Figure 18a). A second scan was performed
over sub-PRPDs
A PSO simulationusing anwith
analogous procedure
30 particles searching
and 300 for pairs
iterations of horizontalItclouds
was performed. took ap-as
described in Section 3.5, resulting in the gap discharge
® confusion
proximately 3 hours on a low-end laptop (Intel Core™ i5-4200U and 8 GB RAM). The matrix (Figure 18b). The
objective function
final optimized was calculated
detection as the
thresholds, thegeometric
same ones mean of theinfour
reported recalls
Section 3.5,obtained
produced from
the
the two confusion matrices.
optimal surface tracking and gap-recognition confusion matrices of Figure 18.

Surface Tracking Gap Discharges

Predicted Predicted
Class Class
No Yes No Yes
Recalls Recalls
Trk Trk Gap Gap
No No
Trk 471 12 0.975 Gap 316 21 0.938
Class
Class

True
True

Yes Yes
Trk 6 104 0.945 Gap 18 238 0.930

(a) (b)

Figure 18.
Figure Confusion matrices
18. Confusion matrices of
of isolated
isolated blocks
blocks of
of (a)
(a) surface
surface tracking
tracking and
and (b)
(b) gap
gap discharge
discharge
recognition.The
recognition. Thedetection
detectionthresholds
thresholdswere
wereoptimized
optimizedvia
viaPSO.
PSO.

4.2. Performance of Each Recognition

A PSO simulation Block inand
with 30 particles an Isolated Mannerwas performed. It took ap-
300 iterations
In this section,
proximately 3 hoursweonshow statistics
a low-end of the(Intel
laptop ®
other system
Core™blocks, based
i5-4200U andon8 neural networks
GB RAM). The
(ANN-0
final to -3). detection
optimized ANN topology is the the
thresholds, dash-separated sequence
same ones reported inof the numbers
Section of neurons
3.5, produced the
of each hidden
optimal surface layer.
trackingNHL
andisgap-recognition
the topology with no hidden
confusion layers,
matrices of i.e.,
Figureinputs
18. are directly
Energies 2022, 15, 326 21 of 26

4.2. Performance of Each Recognition Block in an Isolated Manner

In this section, we show statistics of the other system blocks, based on neural networks
(ANN-0 to -3). ANN topology is the dash-separated sequence of the numbers of neurons
of each hidden layer. NHL is the topology with no hidden layers, i.e., inputs are directly
connected to output neurons. Moreover, since initial weights are random parameters not
related to the classification problem [18,22], their influences are minimized by also showing
the results for the 25% best ANNs from each CV partition.
The results are the performances obtained when neural networks classify the samples
of the test fold of the CV partition they were trained on. Since test fold patterns belong
exclusively to the sources considered by the corresponding type of ANN, the results shown
here are isolated performance statistics, relative to that system block only. Having in
perspective that this block works in cooperation with others in the testing stage, results
estimate the performance that would be achieved by the block if all the previous blocks in
cascade—whose outputs determine which samples are passed to the block in question—
worked flawlessly.
Figure 19 shows the average µ(δ) and standard deviation σ (δ) of topology-wise δ
performances for each isolated block of the recognition system. Results consider all trained
ANNs (red curves) as well as the collection of 25% best networks from each CV partition
(black). The subset of the best ANNs was composed of the 13 highest-δ networks from each
CV partition (approximately 25% of the 50 ANNs trained per partition). In all blocks, one
observes larger and less-variable δ for the 25% best ANNs from each CV partition. This
Energies 2021, 1, 0 was due to inadequate initial weights, which result in poor convergence during 22 ofANNs
26

training [18,22].

Identification of noisy ANGPD clouds (ANN-0) Preliminary separation via histograms (ANN-1)
0.98 0.99

0.97 0.97
Performance δ

Performance δ

0.96 0.94
Legend:
µ(δ) + σ(δ)
0.95 Best topology 0.92
x µ(δ)
All ANNs All ANNs
µ(δ) - σ(δ) 25% best ANNs from each CV partition 25% best ANNs from each CV partition
0.94 0.90
N
5
10
20
35
50
64
75
10
5-
10
20 0
35 0
50 5

N
5
10
20
35
50
64
75
10
5-
10
20 0
35 0
50 5
H

H
5

5
0

-1
-2
-3
-5

-1
-2
-3
-5
L

L
0

0
Topology Topology

(a) (b)
Separation between InV and InD (ANN-2) Separation between Slot and Corona (ANN-3)
1.02 1.00

0.97 0.96
Performance δ

Performance δ

0.92 0.92

0.87 0.89
All ANNs All ANNs
25% best ANNs from each CV partition 25% best ANNs from each CV partition
0.82 0.85
N
5
10
20
35
50
64
75
10
5-
10
20 0
35 0
50 5

N
5
10
20
35
50
64
75
10
5-
10
20 0
35 0
50 5
H

H
5

5
0

-1
-2
-3
-5

-1
-2
-3
-5
L

L
0

Topology Topology

(c) (d)

Figure19.
Figure Topology-wise average
19. Topology-wise average(line)
(line)and
andstandard
standard deviation
deviation(vertical bars)
(vertical of δofperformances
bars) for for
δ performances
eachisolated
each isolated system
system block.
block. Statistics
Statisticsshown
shownforforallallnetworks
networks and
andforfor
thethe
set set
of 25%
of 25%bestbest
ANNs fromfrom
ANNs
eachCV
each CVpartition.
partition. Blocks
Blocks (a)
(a)identification
identificationofofnoisy
noisy ANGPD
ANGPD clouds;
clouds;(b)(b)
preliminary
preliminary separation via via
separation
histograms; (c) distinction between InV and InD; (d) distinction between slot and
histograms; (c) distinction between InV and InD; (d) distinction between slot and corona. The best corona. The best
topologyof
topology of each
each ANN
ANN typetype isisindicated
indicatedbybyarrows.
arrows.

The different
Identification of noisytopologies (free
Preliminary parameter ofSeparation
separation neural between
networks) were compared
Separation between with
ANGPD clouds (ANN-0) via histograms (ANN-1) InV and InD (ANN-2) Slot and Corona (ANN-3)
one anotherPredicted
regarding their generalization
Predicted Class
capabilities. Applying
Predicted
reasoning similar
Predicted
to
that used in the definition of InD
Class δ, theDCIbestCorona
InV/ topologies are those whose ANNs present high
Slot/ Class Class
Slot Corona
average (µNoisy
(δ)) and
Valid
low standard
InV/ deviation (σ(δ)) of δInVmetrics.InD
Following Occam’s razor
True Class

96.4% 1.3% 2.3%

True Class

InD
Noisy 94.8% 5.2% InV 95.0% 5.0% Slot 96.1% 3.9%
Class
Class

True
True

DCI 3.2% 96.3% 0.5%

Valid 1.8% 98.2% Slot/
2.5% 0.0% 97.5% InD 4.9% 95.1% Corona 2.4% 97.6%
Cor

(a) (b) (c) (d)

Figure 20. Element-wise average confusion matrices of the set of best topology’s 25% best ANNs
Energies 2022, 15, 326 22 of 26
Energies 2021, 1, 0 22 of 26

principle,Identification
it is selected the simplest topology (leastPreliminary
of noisy ANGPD clouds (ANN-0)
numberseparation
of neurons) among those with
via histograms (ANN-1)
good0.98 generalization. 0.99

Figure 19a shows δ statistics for the block of identification of the noisy ANGPD cloud,
0.97 0.97

Performance δ

Performance δ
performed by ANNs-0. It can be observed that δs started relatively low for NHL topology,
increasing
0.96 progressively until topology 100,0.94 and it was stagnant for the configurations
with twoµ(δ)hidden layers. Therefore, it was inferred that, for a single hidden layer, networks
Legend:
+ σ(δ)
0.95 Best topology 0.92
with more hidden neurons areAllcapable
x µ(δ)
ANNs of extracting more useful information All ANNs about the
µ(δ) - σ(δ) 25% best ANNs from each CV partition 25% best ANNs from each CV partition
underlying
0.94 relation between inputs and outputs; 0.90 also, the addition of a second hidden
N
5
10
20
35
50
64
75
10
5-
10
20 0
35 0
50 5

N
5
10
20
35
50
64
75
10
5-
10
20 0
35 0
50 5
layer does not increase
H
explanatory power for this application. Topology 35 (1 hidden layer

H
5

5
0

-1
-2
-3
-5

-1
-2
-3
-5
L

L
0

0
Topology Topology
with 35 neurons) was (a)
considered the best on average. (b)
The next block is the preliminary separation of patterns between classes InV/InD, DCI,
and1.02
slot/corona, performed by ANNs-1. δs of1.00
Separation between InV and InD (ANN-2)
this block follow in Figure 19b. In the curve
Separation between Slot and Corona (ANN-3)

of the 25% best ANNs, performance increased from topology NHL to 5 and saturated after
that.
0.97This indicates that, for this specific block, 0.96 more complex networks are not capable
Performance δ

Performance δ
of extracting
0.92
more useful information that could 0.92
improve performance. Topology 5 was
considered the best for this block.
A similar trend was observed
0.87
All ANNs
in δs of ANNs-2
0.89
and ANNs-3, shownAllin ANNs
Figure 19c,d,
respectively.
0.82
The best topology for
25% best ANNs from each CV partition
ANNs-2 was
0.85
20, and 5-5 was the best for ANNs-3.
25% best ANNs from each CV partition
For
the ANN-3 block, we considered that the higher performance of topology 5-5 justified the
N
5
10
20
35
50
64
75
10
5-
10
20 0
35 0
50 5

N
5
10
20
35
50
64
75
10
5-
10
20 0
35 0
50 5
H

H
5

5
0

-1
-2
-3
-5

-1
-2
-3
-5
L

L
0

0
Topology Topology
approximately 15% increase in the number of weights compared to topology 5 (Figure 19d).
In order to have (c) a more general picture of neural networks’ (d) performance, the next
results
Figure 19.areTopology-wise
shown as confusion average (line) matrices. For each
and standard isolated
deviation block,bars)
(vertical theofelement-wise
δ performancesaverage for
of theisolated
each collection systemofblock.
confusion
Statistics matrices
shown for (obtained
all networksonand
thefortest
thefold)
set ofof 25%the 25%
best ANNs bestfrom
ANNs
from
each CVeach CV partition
partition. Blocks (a) was calculated.ofEach
identification noisyelement
ANGPD of the resulting
clouds; (b) preliminary matrix was calculated
separation via
as a percentage
histograms; relativebetween
(c) distinction to the number InV and InD; of patterns that truly
(d) distinction between belongslot andto the
corona. corresponding
The best
PD source
topology of (as
eachlabeled
ANN type by isthe specialist).
indicated The results are shown in Figure 20.
by arrows.

Identification of noisy Preliminary separation Separation between Separation between

ANGPD clouds (ANN-0) via histograms (ANN-1) InV and InD (ANN-2) Slot and Corona (ANN-3)
Predicted Predicted Class Predicted Predicted
Class InV/ Slot/ Class Class
DCI
InD Corona
Noisy Valid InV InD Slot Corona
InV/
True Class

96.4% 1.3% 2.3%

True Class

InD
Noisy 94.8% 5.2% InV 95.0% 5.0% Slot 96.1% 3.9%
Class
Class

True
True

DCI 3.2% 96.3% 0.5%

Valid 1.8% 98.2% Slot/
2.5% 0.0% 97.5% InD 4.9% 95.1% Corona 2.4% 97.6%
Cor

(a) (b) (c) (d)

Figure 20.
Figure Element-wise average
20. Element-wise averageconfusion
confusionmatrices
matricesof of
thethe
setset
of best topology’s
of best 25%25%
topology’s bestbest
ANNsANNs
from each
from each CV partition, for each isolated system block. Matrices obtained from the test sets. Blocks
partition, for each isolated system block. Matrices obtained from the test sets. Blocks
(a)identification
(a) identification of
of noisy
noisyANGPD
ANGPDclouds;
clouds;(b) preliminary
(b) preliminary separation
separationbetween
betweenInV/InD, DCI,DCI,
InV/InD, and and
Slot/Corona; (c) distinction between InV and InD; (d) distinction between slot and
Slot/Corona; (c) distinction between InV and InD; (d) distinction between slot and corona. corona.

The average
The average confusion
confusionmatrix
matrixrelative
relativetoto
the identification
the of noisy
identification ANGPD
of noisy ANGPD clouds is is
clouds
shown in Figure 20a. In this matrix, class “Noisy” denotes patterns whose
shown in Figure 20a. In this matrix, class “Noisy” denotes patterns whose two ANGPD two ANGPD
clouds were considered noisy by the specialist, whereas “Valid” patterns had at least one
clouds were considered noisy by the specialist, whereas “Valid” patterns had at least one
valid cloud. Good results were obtained, with recalls greater than 94.8%. The higher recall
valid cloud. Good results were obtained, with recalls greater than 94.8%. The higher recall
for class “Valid” was likely due to lower variability of its patterns in feature space, which
for class “Valid” was likely due to lower variability of its patterns in feature space, which
facilitates classification.
facilitates classification.
The next block is the preliminary separation performed by ANNs-1 using the his-
togram feature.block
The next is the preliminary
The average separation
confusion matrix (Figureperformed
20b) showed byexcellent
ANNs-1 usingwith
results, the his-
togram feature. The average confusion matrix
recognition rates greater than 96% for all classes. (Figure 20b) showed excellent results, with
recognition
The lastrates greater
blocks than 96% for all
are differentiation classes.InV and InD patterns by ANNs-2 and
between
The last
separation blocksslot
between are and
differentiation betweenThe
corona by ANNs-3. InVcontour
and InD patterns
feature by ANNs-2
was used by both and
separation between
types of networks. In slot and corona
the testing by ANNs-3.
stage, these The contour
blocks resolve feature
the ambiguity was used
between by both
patterns
types of networks.
classified as InV/InD In the testing stage,
or slot/corona bythese
ANNs-1blocks resolve
(Figure 5). the ambiguity
Average between
confusion patterns
matrices
classified as InV/InD
follow in Figure 20c,d, or slot/corona by ANNs-1 (Figure 5). Average confusion matrices
respectively.
follow in Figure 20c,d, respectively.
Energies 2021, 1, 0 23 of 26

4.3. Overall Recognition Statistics

Energies 2022, 15, 326 The results presented so far are isolated performance figures of system blocks, consid- 23 of 26
ering that all patterns presented to the block were correctly classified by the other preceding
blocks in the cascade.
To calculate
4.3. Overall the system’s
Recognition Statisticsoverall recognition rates for each PD source, one multiplies
the isolated hit rates of the blocks that a given sample must pass through to be classified as
that The results
source. For presented
example, for so far are isolated
a pattern performance
to be recognized byfigures of system
the system as InV,blocks, consid-
the following
ering that all patterns presented to the block were correctly classified
sequence of events must occur: (i) ANGPD clouds must be considered valid by ANNs-0; by the other preceding
blocks in themust
(ii) ANN-1 cascade.
identify the pattern as InV/InD; and (iii) ANN-2 must classify the sample
as InV. Thus, the the
To calculate system’s
system’s overall
global recognition
recognition rates
rate for thefor eachinternal
source PD source,
voidone multiplies
is equal to the
the isolated hit rates of the blocks that a given sample must pass
product of hit rates of blocks ANNs-0, ANNs-1, and ANNs-2. After performing analogousthrough to be classified as
that source. For example, for a pattern to be recognized by the system
calculations for the other PD sources, the resulting global recognition rates are shown in as InV, the following
sequence
Figure 21.ofThis events must
figure occur:
also shows (i)the
ANGPD
isolatedclouds must be considered
performances validblock.
of each system by ANNs-0;
Global
(ii) ANN-1 must
recognition ratesidentify
greater the
thanpattern
88% were as InV/InD;
obtainedand for (iii) ANN-2 must
all considered PDclassify
sources. the sample
as InV.The Thus, the system’s
collection of theglobal recognition
best network of rate for the
the best source internal
topology from eachvoidANNis equal
typetowas
the
product of hit rates of blocks ANNs-0, ANNs-1, and ANNs-2. After performing
assembled to form an instance of the overall classification system. The system classified the analogous
calculations
hard-to-identify for the otherofPD
PRPDs sources,
Figure the resulting
3 according to theglobal
testingrecognition rates 5.
stage of Figure areThe
shown in
results
Figure 21. This figure also shows the isolated performances of each
are shown in Table 3. All samples were correctly classified, which further validates the system block. Global
recognition rates greater than 88% were obtained for all considered PD sources.
proposed methodology.

Tracking Identification of Both clouds noisy

N-ANGPD and Gap noisy ANGPD
source(s) clouds (ANN-0) 94.8%
Tracking

98.2%
only At least
No one valid
N-ANGPD Gap
source only Extraction of
Trained
amplitude
histograms [18] ANN-1
Start
91.5%

87.9%

88.6%

90.7%

Slot / Corona InV / InD DCI

97.5% 96.4%

96.3%
Classification Extraction of Extraction of
PRPD of N-ANGPD Contour Contour
source(s) Features Features

PRPD Image N-ANGPD Trained Trained

Denoising sub-PRPD ANN-3 ANN-2
and Multiple-
95.1%
96.1%

97.6%

95.0%

Source
Separation ANGPD
(Figure 6) sub-PRPD

Dominant InD No
Slot
ANGPD ANGPD
Legend:
source 92.0% 90.0% source
XY.Z%: classification rates of
the isolated block InV 94.8%
Corona DCI
XY.Z%: system’s global
classification rates 93.4% 89.9% 94.6%

Figure 21.
Figure Isolated and
21. Isolated and calculated
calculated global
global classification
classification rates
rates of
of the
the proposed
proposed multiple
multiple PD-source
PD-source
recognition system.
recognition system.

The collection of the best network of the best topology from each ANN type was
assembled to form an instance of the overall classification system. The system classified the
hard-to-identify PRPDs of Figure 3 according to the testing stage of Figure 5. The results
are shown in Table 3. All samples were correctly classified, which further validates the
proposed methodology.
Energies 2022, 15, 326 24 of 26

Table 3. System classification results on the patterns of Figure 3.

True PD System’s Numerical Outputs of Internal Networks

Pattern
Source(s) Predicted Class(es) ANN-0 * ANN-1 † 2ANN-2 ‡ 2ANN-3 §
Figure 3a Corona Corona 0.01|0.99 0.01|0|0.99 — 0.14|0.86
Figure 3b Slot Slot 0.01|0.99 0|0|1 — 0.99|0.01
Figure 3c InD InD 0|1 0.91|0.04|0.05 0.01|0.99 —
Figure 3d InV + Gap InV + Gap 0.01|0.99 0.9|0.05|0.05 1|0 —
Figure 3e DCI + Gap DCI + Gap 0.05|0.95 0.01|0.99|0 — —
Figure 3f Und Und 0.99|0.01 — — —
* Values of output neurons associated to Noisy|Valid classes. † Values of output neurons associated to
(InV/InD)|DCI|(Slot/Corona) classes. ‡ Values of output neurons associated to InV|InD classes. § Values
of output neurons associated to Slot|Corona classes.

5. Final Remarks
In this study, a novel multiple-source PD-recognition methodology in hydro-generator
stator bars was pesented. The functionality of identifying patterns not having valid PD
source (e.g., only noise) was also included, that is, the identification of patterns with no
valid PD sources. The technique is based on AI and statistics, and the data used to train
the algorithms were PRPD patterns obtained from online PD measurements in hydro-
generators operating in real-world conditions.
PRPD samples were treated as images in the methodology. First, they were subjected
to a modified and extended version of the image denoising algorithm proposed by the
authors in a previous study. The main adaptations were the addition of a novel image-based
algorithm that separates partially superposed PD clouds in the PRPD image and that two
denoised sub-PRPDs were generated as outputs, each formed by discharges of a different
nature. The result is a denoising algorithm that, as stated in the authors’ previous study,
eliminates pattern variations introduced by disturbances found in online measurements; in
this work, it was extended and fine-tuned to better cope with samples presenting multiple
simultaneous PD sources.
The recognition task was divided into multiple smaller problems, each solved by
specific AI classifiers and input features. From the denoised sub-PRPDs, three types of
input features were extracted, two of which are novel: variables for identifying the presence
of valid PD source(s) and features that quantify the shape of PD clouds. The extracted
features were provided as inputs to four types of ANNs. Each type of ANN was trained to
solve a part of the problem, acting like blocks of a larger recognition system. Once trained,
ANNs were used cooperatively in a hierarchical fashion to classify new patterns.
The performance of each classifier was assessed individually. The system blocks
presented isolated accuracies higher than 95%. From those statistics, the system’s overall
performances were calculated, and the global accuracy rates ranged from 88% to 94.8% for
all the investigated PD sources.
Despite the good performance, the methodology has some limitations, the main of
which are the separation of superposed PDs in restricted circumstances and not identifying
multiple simultaneous ANGPD PD sources. The authors intend to work on those limitations
in future studies, by means of adding time-domain information to the methodology and
using more-advanced clustering techniques.

Author Contributions: Conceptualization, R.M.S.d.O. and F.J.B.B.; methodology, R.M.S.d.O., R.C.F.A.

and F.J.B.B.; software, R.C.F.A.; validation, R.C.F.A., R.M.S.d.O. and F.J.B.B.; formal analysis, R.M.S.d.O.,
R.C.F.A. and F.J.B.B.; investigation, R.C.F.A.; resources, R.C.F.A.; data curation, R.C.F.A.; writing—
original draft preparation, R.C.F.A.; writing—review and editing, R.C.F.A., R.M.S.d.O. and F.J.B.B.;
visualization, R.C.F.A.; supervision, R.M.S.d.O. and F.J.B.B.; project administration, R.M.S.d.O. and
F.J.B.B.; funding acquisition, R.M.S.d.O. and F.J.B.B. All authors have read and agreed to the published
version of the manuscript.
Energies 2022, 15, 326 25 of 26

Funding: This study’s publication fee was funded by PROPESP/UFPA, under Edital PAPQ. During
the research that gave rise to this study, R.C.F.A. received a doctoral scholarship from CNPq/BRASIL.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The authors reserve the right to not disclose Eletronorte’s private data
set of on-line PD measurements used in this study.
Acknowledgments: Federal University of Pará and the Brazilian National Research Agency CNPq
are acknowledged for financial support. We are grateful to Eletrobras Eletronorte and to Fernando
Brasil for providing the data set used in this work.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Stone, G.C. Condition Monitoring and Diagnostics of Motor and Stator Windings—A Review. IEEE Trans. Dielectr. Electr. Insul.
2013, 20, 2073–2080. [CrossRef]
2. Florkowski, M. Classification of Partial Discharge Images Using Deep Convolutional Neural Networks. Energies 2020, 13, 5496.
[CrossRef]
3. Basharan, V.; Siluvairaj, W.I.M.; Velayutham, M.R. Recognition of multiple partial discharge patterns by multi-class support
vector machine using fractal image processing technique. IET Sci. Meas. Technol. 2018, 12, 1031–1038. [CrossRef]
4. Mor, A.R.; Muñoz, F.; Wu, J.; Heredia, L.C. Automatic partial discharge recognition using the cross wavelet transform in high
voltage cable joint measuring systems using two opposite polarity sensors. Int. J. Electr. Power Energy Syst. 2020, 117, 105695.
5. Kuppuswamy, R.; Rainey, S. Facilitating Proactive Stator Winding Maintenance Using Partial Discharge Patterns. In Proceedings
of the 2018 IEEE Electrical Insulation Conference (EIC), San Antonio, TX, USA, 17–20 June 2018; pp. 566–571.
6. Castro Heredia, L.C.; Rodrigo Mor, A. Density-based clustering methods for unsupervised separation of partial discharge sources.
Int. J. Electr. Power Energy Syst. 2019, 107, 224–230. [CrossRef]
7. Zemouri, R.; Lévesque, M.; Amyot, N.; Hudon, C.; Kokoko, O.; Tahan, S.A. Deep Convolutional Variational Autoencoder as a
2D-Visualization Tool for Partial Discharge Source Classification in Hydrogenerators. IEEE Access 2020, 8, 5438–5454. [CrossRef]
8. Pardauil, A.C.N.; Nascimento, T.P.; Siqueira, M.R.S.; Bezerra, U.H.; Oliveira, W.D. Combined Approach Using Clustering-Random
Forest to Evaluate Partial Discharge Patterns in Hydro Generators. Energies 2020, 13, 5992. [CrossRef]
9. Song, H.; Dai, J.; Sheng, G.; Jiang, X. GIS Partial Discharge Pattern Recognition via Deep Convolutional Neural Network under
Complex Data Sources. IEEE Trans. Dielectr. Electr. Insul. 2018, 25, 678–685. [CrossRef]
10. Montanari, G.C.; Seri, P.; Ghosh, R.; Cirioni, L. Noise rejection and partial discharge source identification in insulation system
under DC voltage supply. IEEE Trans. Dielectr. Electr. Insul. 2019, 26, 1894–1902. [CrossRef]
11. Zhu, M.X.; Wang, Y.B.; Chang, D.G.; Zhang, G.J.; Shao, X.J.; Zhan, J.Y.; Chen, J.M. Discrimination of three or more partial discharge
sources by multi-step clustering of cumulative energy features. IET Sci. Meas. Technol. 2019, 13, 149–159. [CrossRef]
12. Wang, Y.B.; Chang, D.G.; Qin, S.R.; Fan, Y.H.; Mu, H.B.; Zhang, G.J. Separating Multi-Source Partial Discharge Signals Using
Linear Prediction Analysis and Isolation Forest Algorithm. IEEE Trans. Instrum. Meas. 2020, 69, 2734–2742. [CrossRef]
13. Lalitha, E.; Satish, L. Wavelet analysis for classification of multi-source PD patterns. IEEE Trans. Dielectr. Electr. Insul. 2000, 7,
40–47. [CrossRef]
14. Ma, Z.; Yang, Y.; Kearns, M.; Cowan, K.; Yi, H.; Hepburn, D.M.; Zhou, C. Fractal-based autonomous partial discharge pattern
recognition method for MV motors. High Volt. 2018, 3, 103–114. [CrossRef]
15. IEC 60034-27-2. Rotating Electrical Machines-Part 27-2: On-Line Partial Discharge Measurements on the Stator Winding Insulation of
Rotating Electrical Machines; International Electrotechnical Commission: Geneva, Switzerland, 2012.
16. Mantach, S.; Ashraf, A.; Janani, H.; Kordi, B. A Convolutional Neural Network-Based Model for Multi-Source and Single-Source
Partial Discharge Pattern Classification Using Only Single-Source Training Set. Energies 2021, 14, 1355. [CrossRef]
17. Wang, Y.; Yan, J.; Yang, Z.; Zhao, Y.; Liu, T. GIS partial discharge pattern recognition via lightweight convolutional neural network
in the ubiquitous power internet of things context. IET Sci. Meas. Technol. 2020, 14, 864–871. [CrossRef]
18. Araújo, R.C.F.; Oliveira, R.M.S.; Brasil, F.S.; Barros, F.J.B. Novel Features and PRPD Image Denoising Method for Improved
Single-Source Partial Discharges Classification in On-Line Hydro-Generators. Energies 2021, 14, 3267. [CrossRef]
19. Amorim, H.; de Carvalho, A.; de Oliveira, O.; Levy, A.; Sans, J. Instrumentation for Monitoring and Analysis of Partial Discharges
Based on Modular Architecture. In Proceedings of the International Conference on High Voltage Engineering and Application
(ICHVE 2008), Chongqing, China, 9–12 November 2008; pp. 596–599.
20. Hudon, C.; Bélec, M. Partial Discharge Signal Interpretation for Generator Diagnostics. IEEE Trans. Dielectr. Electr. Insul. 2005, 12,
297–319. [CrossRef]
21. Stone, G.C. Partial Discharge Diagnostics and Electrical Equipment Insulation Condition Assessment. IEEE Trans. Dielectr. Electr.
Insul. 2005, 12, 891–904. [CrossRef]
Energies 2022, 15, 326 26 of 26

22. Oliveira, R.M.S.; Araújo, R.C.F.; Barros, F.J.; Segundo, A.P.; Zampolo, R.; Fonseca, W.; Dmitriev, V.; Brasil, F. A System Based on
Artificial Neural Networks for Automatic Classification of Hydro-generator Stator Windings Partial Discharges. J. Microwaves
Optoelectron. Electromagn. Appl. 2017, 16, 628–645. [CrossRef]
23. Guan, S.; Bao, C.; Neo, T. Reduced Pattern Training Based on Task Decomposition Using Pattern Distributor. IEEE Trans. Neural
Netw. 2007, 18, 1738–1749. [CrossRef]
24. Everitt, B.; Landau, S.; Leese, M.; Stahl, D. Cluster Analysis, 5th ed.; Wiley: Chichester, UK, 2011.
25. Contin, A. On the Identification of Insulation Defects Supporting Partial Discharges. In Proceedings of the 2014 AEIT Annual
Conference-From Research to Industry: The Need for a More Effective Technology Transfer (AEIT), Trieste, Italy, 18–19 September
2014; pp. 1–6.
26. Rojo-Álvarez, J.L.; Martínez-Ramón, M.; Muñoz-Marí, J.; Camps-Valls, G. Digital Signal Processing with Kernel Methods; Wiley:
Hoboken, NJ, USA, 2018.
27. Witten, I.; Frank, E.; Hall, M.; Pal, C. Data Mining: Practical Machine Learning Tools and Techniques, 4th ed.; Morgan Kaufmann:
Cambridge, MA, USA, 2016.
28. Møller, M.F. A Scaled Conjugate Gradient Algorithm for Fast Supervised Learning. Neural Netw. 1993, 6, 525–533. [CrossRef]
29. Kennedy, J. Particle Swarm Optimization. In Encyclopedia of Machine Learning; Springer: New York, NY, USA, 2010; pp. 760–766.