Partial Discharge Source Classification Based on

Machine Learning and PRPD

João Victor Jales de Melo George Rossany Soares de Lira Edson Guedes da Costa
Post-graduate Program Electrical Engineering Department Electrical Engineering Department
Federal University of Campina Grande Federal University of Campina Grande Federal University of Campina Grande
Campina Grande, Brazil Campina Grande, Brazil Campina Grande, Brazil
[email protected] [email protected] [email protected]

Pablo Bezerra Vilar Antonio Francisco Leite Neto Lenilson Andrade Barbosa
Electrical Engineering Department Post-graduate Program Post-graduate Program
Federal University of Campina Grande Federal University of Campina Grande Federal University of Campina Grande
Campina Grande, Brazil Campina Grande, Brazil Campina Grande, Brazil
[email protected] [email protected] [email protected]

Allan David da Costa Silva Ana Cristina de Freitas Marotti André Irani Costa
Post-graduate Program Eletrobras S.A. Eletrobras S.A.
Federal University of Campina Grande Rio de Janeiro, Brazil Rio de Janeiro, Brazil
Campina Grande, Brazil [email protected] [email protected]
[email protected]

Abstract— A method for analyzing partial discharges is discharges. Corona discharges manifest in pointed
proposed to detect faults and evaluate the operational status of conductive materials or at locations with a high electric field
electrical system assets. Initially, the effectiveness of the method between the conductor and the surrounding air. These
was hindered by electromagnetic interferences affecting the discharges, being external, are not indicative of internal
sensor, including corona, surface discharges, and radio signals. insulation degradation. Surface discharges, on the other hand,
To address this challenge, multiple layers of signal processing occur at the interface of an insulating material with the air and
were introduced. This paper proposes a new methodology for are often triggered by adverse environmental conditions,
noise removal and classification of partial discharge sources mechanical wear, or existing deterioration of the insulating
based on clustering pulses by attributes derived from their
material. Lastly, internal discharges take place within cavities
shape. Consequently, it became feasible to distinguish partial
discharges from the various sources of interference present in
embedded within the insulating material. These cavities
the measurements. The methodology was tested using a dataset facilitate the formation of regions with concentrated electric
of partial discharge measurements obtained from Current fields, which can progressively degrade the material until
Transformers at an Eletrobras group substation in Brazil. severe failures occur. Understanding these distinct sources of
Through this testing, the methodology proved capable of partial discharges is crucial for effective fault detection and
assessing the status of the analysed equipment. insulation system assessment.
Both the source of the discharge and the insulating
Keywords—partial discharge, classification, feature
extraction
medium in which they occur can impact the waveform
characteristics of the resulting pulse. Hence, numerous
I. INTRODUCTION studies have proposed techniques to identify and differentiate
discharge sources by analyzing pulse attributes extracted
Partial Discharges (PD) is often considered a significant from the waveforms. Typically, these studies focus on
indicator of non-conformities in electrical equipment, extracting attributes related to frequency and time [10], [11].
consistently associated with insulation issues [1], [2], [3]. For However, there are also investigations that explore
this reason, several studies propose PD analysis to detect alternative attributes, such as energy distribution across
faults and assess the operating condition of electrical system different frequency ranges [12].
assets [4], [5], [6].
In this scenario, a new methodology for data processing
While partial discharge measurements offer a potential and separation of discharge sources based on pulse shape is
means of estimating insulation system degradation, the propose. The proposed attributes are used to differentiate the
literature widely acknowledges several limitations to their groups of pulses, and an unsupervised model is employed to
applicability [7], [8]. Sensors tasked with measuring partial form clusters. The proposed methodology is tested with real
discharge are susceptible to interference from radio signals, measurements obtained at an Eletrobras substation, with a
TV transmissions, WiFi, corona, and other environmental particular focus on the noise removal capability.
factors. Moreover, the presence of multiple partial discharge
sources further complicates fault detection, presenting a II. FEATURE EXTRATION
significant challenge in this field of research. In this context,
distinguishing between different types of partial discharge As the waveform of the pulse serves is the primary source
and isolating them from other sources of noise emerges as a of information for determining its origin, this paper proposes
critical challenge within this research domain. characterizing the pulse based on the following attributes:

According to [9], PD typically originate from three main • Pulse duration:

sources: corona discharges, surface discharges, and internal
= + . (1)

979-8-3503-7498-8/24/$31.00 ©2024 IEEE

 • Fall time divided by rise time: A. Data Base
As depicted in Fig. 3, the measurement campaign
t= / . (2) involves acquiring partial discharges and leakage current on
the grounding cable of current transformers. Measurements
• Highest frequency component. were conducted on a total of 30 current transformers of 345
kV. The analysis of partial discharge pulses requires the
• Energy of the first half of the signal divided by the reference voltage applied to the current transformer, enabling
energy of the second half of the signal. identification of the phase in which the discharge occurred.
• Number of pulse oscillations. However, since the reference voltage is not always available,
the leakage current sensor is a viable alternative, as the
The attributes are depicted in Fig. 1. voltage will be approximately 90° out of phase with the
current. The sensor used for PD measurement is the HFCT
with a bandwidth between 1 and 80 MHz and a sensitivity of
19 mV/mA.

Fig. 1. Example of PD pulse.

The proposed attributes are related to time, frequency and

shape of the pulses. The purpose is to allow the differentiation Fig. 3. Experimental setup in (a) and leakage current sensor and HFCT in
(b).
of patterns. Although it is not known how these attributes are
affected by the type of discharge, location or insulating The signals obtained with the leakage current sensor and
material, the existence of changes is well known. For this the HFCT were fed to a data acquisition system (DAQ),
reason, these attributes can be used to differentiate different which performs sampling at 125 MSa/s. Each of the
types of discharges. measurements was conducted with a window of 33.4 ms (two
The dimensionality of the attributes was then reduced periods of 60 Hz).
using PCA so that it's possible to analyse the results in up to B. Signal Preprocessing
3 dimensions. Thus, the three first principal components
The preprocessing stage consists of filtering the
(PC1, PC2, and PC3) of the proposed variables are utilized.
measurements and extracting pulses in standardized
III. METHODOLOGY windows.
Partial discharge measurements were carried out on Initially, a band-pass filter with a lower limit of 1.5 MHz
current transformers installed at a high-voltage substation in and an upper limit of 62.5 MHz is used to reject frequency
Brazil. The PD signals were utilized to evaluate the bands outside the sensitivity region of the HFCT sensor and
effectiveness of the proposed attributes in distinguishing the Nyquist criterion, respectively. Subsequently, a filtering
based on the wavelet decomposition technique was used [13],
between partial discharge sources and eliminating noise. The
[14]. The selection of the mother wavelet was made based on
obtained data is pre-processed and subsequently evaluated the preservation of signal energy, as described in [15].
following the methodology described in this session. The
flowchart presented in Fig. 2 illustrates the methodology. To initiate the separation process for partial discharge
measurements, it's necessary to extract and catalog pulses
within standardized time windows. To achieve this, a Python
algorithm is devised to systematically navigate through the
entire database and identify these pulses. A threshold is
established based on the background noise level of the
measurement, with each pulse allotted a window of 8.0 µs.
The selection of threshold and window size is calibrated to
ensure accurate representation of pulses from all discharges
while disregarding residual noise post-filtering as valid
pulses.
C. Partial Discharge Clusterization
The separation process initiates with the computation of
parameters as outlined in Section II for every measured pulse.
Subsequently, these attributes are stored, followed by the

Fig. 2. Methodology flowchart.

application of a clustering algorithm to discern potential The preprocessing step is essential in partial discharge
clusters within the attribute-formed database. measurements in substation. There are various sources of low-
frequency noise that hinder the identification of partial
The clustering model used in this work was HDBSCAN, discharge pulses. As seen in Fig. 4, it was observed that TC-
as proposed by [16], which enables the formation of groups 14 has a PD source with an amplitude greater then 100 mV,
based on density. Moreover, this type of model automatically which cannot be visualized without filtering.
identifies the number of clusters, which minimizes
interference in the process of identifying discharge sources. After filtering and pulse extraction, the attributes
presented in Section II were calculated for all pulses measured
D. Results Evaluation in each of the 30 current transformers at the substation. The
The assessment of the status of the internal insulation of HDBSCAN model was used to identify clusters in the set of
the equipment under analysis is conducted through the measurements. An example of this result is presented in Fig.
analysis of the PRPD graph of the field measurements. 5, again for TC-14.
Typically, analysing the PRPD of real measurements would
not be feasible due to multiple sources of discharges.
However, by segregating the pulses into distinct clusters, each
cluster can be individually assessed based on its PRPD profile.
The objective is to pinpoint characteristic patterns associated
with noise, corona discharge, surface discharge, or internal
discharge.
All measurements conducted at the power substation are
evaluated following the methodology described in this
section. The measurements are particularly used to verify the
capability of the proposed attributes to identify and separate
sources of noise present in substation measurements.
IV. RESULTS
Initially, the measurements performed on the substation
were filtered and the pulses were extracted according to the
proposed methodology. Due to the large amount of
information, it is not possible to present the results for all 30
current transformers, therefore, an example of the filtering
process is presented in Fig. 4 for an acquisition performed on
TC-14.

Fig. 5. Separation of the TC-14 pulses using the features proposed in this
paper.

As observed in Fig. 5, the proposed attributes allowed the

separation of the measured pulses into different clusters. Then,
the PRPD of each of these clusters was built and analysed to
ascertain the nature of each cluster. The PRPD plot of the
obtained clusters in Fig. 5 is presented in Fig. 6.

Fig. 6. PRPD graph of the partial discharge measurement of TC-14.

Fig. 4. Example of partial discharge measurement before and after filtering
process in TC-14. The analysis of Fig. 6 demonstrates that cluster 6 exhibits
a typical PRPD pattern of internal PD occurrence. The
separation technique was even able to distinguish between [4] E. C. T. Macêdo, “Metodologia para a Classificação de Descargas
different sources of noise, as observed in the formation of Parciais Utilizando Redes Neurais Artificiais”, Ph.D. dissertation,
Electrical Engineering Depart., Federal University of Campina Grande,
clusters 1 to 5. Campina Grande, 2014.
The same procedure described for TC-14 was conducted [5] I. F. Carvalho et al., “PRPD analysis in current transformers using UHF
for the entire database. The analyses facilitated the sensors and signal conditioning system”, presented at the 23rd
Internacional Symposium on High Voltage Engineering, Glasgow,
identification of at least two other current transformers 2023.
possibly experiencing partial discharges. These transformers [6] G. R. S. Lira et al., “Monitoramento inteligente das condições
are currently under analysis by the maintenance team at operacionais de transformadores de corrente”, presented at the XXVII
Eletrobras. Seminário Nacional de Produção e Transmissão de Energia Elétrica,
Brasilia, 2023.
V. CONCLUSIONS [7] H. Karami and G. B. Gharehpetian, “Limitations of Partial Discharge
De-noising of Power Transformer Using Adaptive Singular Value
A novel methodology for separating sources of partial Decomposition”, in 34th International Power System Conference,
discharges and noise based on the waveform of the pulse was 2019, pp. 777–781. doi: 10.1109/PSC49016.2019.9081523.
proposed. To evaluate the applicability of this technique, a [8] X. Li, W. Liu, e Y. Xu, “Partial Discharge Detection of insultors in
dataset from a power substation of the Eletrobras Group in GIS: Effectiveness and Limitation”, in 7th IEEE International
Brazil was utilized. Conference on High Voltage Engineering and Application, 2020. doi:
10.1109/ICHVE49031.2020.9279483.
The proposed separation technique demonstrated its [9] G. C. Stone, “Partial Discharge - Part VII: Practical Techniques for
effectiveness in discerning sources of discharges and noise, Measuring PD in Operating Equipment”, IEEE Electrical Insulation
and it even enabled the identification of three current Magazine, vol. 7, no 4, pp. 9–19, 1991, doi: 10.1109/57.87656.
transformers operating with a notable level of internal partial [10] A. Contin, A. Cavallini, G. C. Montanari, G. Pasini, and F. Puletti,
discharges. “Digital detection and fuzzy classification of partial discharge signals”,
IEEE Transactions on Dielectrics and Electrical Insulation, vol. 9, no
The results from the clustering model showed that 3, pp. 335–348, jun. 2002, doi: 10.1109/TDEI.2002.1007695.
automatic identification of discharge sources is feasible in [11] L. Hao et al., “Discrimination of multiple PD sources using wavelet
online measurements. Employing a supervised model trained decomposition and principal component analysis”, IEEE Transactions
on Dielectrics and Electrical Insulation, out. 2011, pp. 1702–1711. doi:
to recognize partial discharge patterns would facilitate 10.1109/TDEI.2011.6032842.
continuous and autonomous monitoring of electrical [12] J. A. Ardila-Rey, J. M. Martinez-Tarifa, M. Mejino, R. Albarracin, M.
equipment. V. Rojas-Moreno, and G. Robles, “Chromatic classification of RF
signals for partial discharges and noise characterization”, in
ACKNOWLEDGMENT Proceedings of IEEE International Conference on Solid Dielectrics,
ICSD, 2013, pp. 67–70 , doi: 10.1109/ICSD.2013.6619863.
The authors would like to thank Eletrobras for the
technical and financial support within the scope of the R&D [13] X. Zhou, C. Zhou, and I. J. Kemp, “An improved methodology for
application of wavelet transform to partial discharge measurement
project "Development of an Intelligent Online Monitoring denoising”, IEEE Transactions on Dielectrics and Electrical
System for Current Transformers " financed by the Brazilian Insulation, vol. 12, no 3, pp. 586–594, jun. 2005, doi:
Electricity Sector R&D Program, regulated by the National 10.1109/TDEI.2005.1453464.
Electric Energy Agency (ANEEL). The authors would also [14] A. T. Carvalho, A. C. S. Lima, C. F. F. C. Cunha, and M. Petraglia,
like to thank the financial support from the Coordenação de “Identification of partial discharges immersed in noise in large hydro-
generators based on improved wavelet selection methods”,
Aperfeiçoamento de Pessoal de Nível Superior (CAPES), Measurement, vol. 75, pp. 122–133, nov. 2015, doi:
great number: 88887.641227/2021-00 and CNPq great 10.1016/J.MEASUREMENT.2015.07.050.
number: 309743/2023-0. [15] X. Zhou, C. Zhou, e I. J. Kemp, “An improved methodology for
application of wavelet transform to partial discharge measurement
REFERENCES denoising”, IEEE Transactions on Dielectrics and Electrical
[1] H. J. Ju, J. G. Lee, K. S. Han, J. W. Kang, and W. Choi, “An Analysis Insulation, vol. 12, no 3, pp. 586–594, jun. 2005, doi:
of Partial Discharge Characteristics due to Transformer Bushing 10.1109/TDEI.2005.1453464.
Failure”, in ICEPE-ST 2022, 2022, pp. 245–248. doi: 10.1109/ICEPE- [16] R. J. G. B. Campello, D. Moulavi, e J. Sander, “Density-Based
ST51904.2022.9757090. Clustering Based on Hierarchical Density Estimates”, in PAKDD 2013,
[2] T. Tanmaneeprasert, P. L. Lewin, and G. Callender, “Analysis of 2013, pp. 160–172, doi: 10.1007/978-3-642-37456-2_14.
degradation mechanisms of silicone insulation containing a spherical
cavity using partial discharge detection”, in 2017 IEEE Electrical
Insulation Conference, 2017, pp. 233–236. doi:
10.1109/EIC.2017.8004633.
[3] W. McDermid and T. Black, “Failure of service aged 230 kV current
transformers”, in Conference Record of IEEE International Symposium
on Electrical Insulation, 2012, p. 235–236. doi:
10.1109/ELINSL.2012.6251464.