Professional Documents
Culture Documents
Partial Discharge Spectrogram Data Augmentation Based On Generative Adversarial Networks
Partial Discharge Spectrogram Data Augmentation Based On Generative Adversarial Networks
Abstract—High voltage insulators are critical elements in power frequency cycle form distinct patterns according to its
power distribution and transmission systems, since their failure origin and magnitude. Since these signals can be acquired with
may represent losses in power quality and system reliability, proper sensors, an algorithm of pattern recognition could be
leading to faults which bring problems to both consumer and responsible for the classification of these signals, resulting in
electric utilities. Insulators in operation are subjected to all a diagnosis of the insulator condition, with respect to pollution
kinds of pollution which may cause the formation of dry bands and aging.
when in presence of moisture, which leads to the formation of
partial discharges which, in turn, causes degradation and aging Machine learning algorithms, especially artificial neural
of the surface material. Partial discharges origin can be networks, have been used for pattern recognition and
identified by the pattern of their pulses. Machine learning classification tasks. In later years, convolutional networks
algorithms are one of the most useful and efficient solutions for have been proved to be efficient for computer vision, as
pattern recognition and classification tasks. In this paper, a demonstrated by [1] with the application of the
methodology is proposed for data augmentation of the training backpropagation algorithm to the convolutional architecture.
set for an artificial neural network by means of generative This architecture has been used for speech recognition, by
adversarial networks. convolving filters over two-dimensional data and extracting
distinctive features which can be used for classification. Two-
Keywords—artificial intelligence, artificial neural networks,
dimensional speech data can be obtained through the short-
convolutional networks, generative adversarial networks,
insulators, machine learning, partial discharges
time Fourier transform (STFT) of voice signals, which convert
these time domain signals to the time-frequency domain [2].
I. INTRODUCTION Due to the randomness of all these phenomena and to the
When in operation, high voltage insulators undergo all high number of utility poles and towers, electric utilities
kinds of pollution deposition onto their surface. This cannot provide an effective inspection of the system. In
pollution, when in presence of moisture, may form a addition to that, an insulator itself is inexpensive and can be
conductive layer on the insulator surface, causing an increase easily replaced. However, as a whole, a great number of
of the leakage current, which may lead to the formation of dry failures start to represent an increase in the purchase of
bands which can further lead to flashover events. In coastal repairment material and substitution routines and, even worse,
cities, e.g., insulators are subjected to a heavy salt deposition may lead to faults that could represent costly penalties due to
due to salt spray and, during the months of low rainfall, this power supply interruption.
salt pollution is intensified, since rain helps to naturally wash
Here, a methodology is proposed for partial discharge data
insulators surface. Due to the continuous air humidity, there
augmentation by means of generative adversarial networks in
are areas where, even with no rain, critical events take place
order to obtain a sufficiently large training dataset. Scarcity of
over the so-called “dry season”, when pollution accumulates
data is one of the factors that may result in model overfitting
and moisture is still present, mainly in the dawn.
during training, being data augmentation one of the solutions
One of the phenomena that occur in the insulators are the proposed to tackle that. A PD dataset may be hard to obtain
partial discharges, i.e., a discharge that do not fully connect due to the randomness of their occurrence. The generative
the electrodes. Partial discharges (PD) can be measured by model was obtained by an adversarial process, being based on
means of its apparent charge, and its pulses with respect to the the two-dimensional convolutional architecture and trained on
two-dimensional PD spectrograms data. The generated data
This project is part of the Brazilian Electricity Regulatory Agency were then evaluated visually and used to train a convolutional
(ANEEL) Research & Development (R&D) program PD-00371-0036/2018 neural network for the classification of PD signals.
and was financed by Equatorial Energia and in part by the Coordination for
the Improvement of Higher Education Personnel – Brazil (CAPES) –
Finance Code 001.
978-1-7281-5532-6/19/$31.00
Authorized ©2019 IEEE
licensed use limited to: Centro Universitário Fei. Downloaded on September 09,2021 at 17:13:37 UTC from IEEE Xplore. Restrictions apply.
2019 International Conference on Electrical and Computing Technologies and Applications (ICECTA)
II. PARTIAL DISCHARGES IN HIGH VOLTAGE INSULATORS modes, thus PD activity can be displayed on a horizontal line,
Over the dry season, electric utilities must tackle a series on a sine wave and on an ellipse. These patterns can be used
of events due to the low precipitation. First, in the association in a variety of manners so that the PD can be classified. Some
with high air humidity, fog or dew, the pollution forms a statistical calculations analyze a few distributions obtained
conductive layer that may cause a flashover event, which may from the PRPD pulses and use the results as features to train
only lead to a system fault or may damage the insulator so that an artificial neural network. However, many methods today
it needs to be replaced. Second, electric utilities observe an present the raw PRPD signal for a neural network to be trained
increase in the purchase of new insulators during the months and the training algorithms itself can detect those features that
of lower precipitation. And third, electric utilities have to are distinctive and that allow the signal to be classified.
spend on insulators washing routines in order to prevent Partial discharges vary their features according to their
events as the described above. Washing insulators during the origin and severity of insulator degradation. If one can
dry season may be a normal part of the electric utilities’ correlate these features with defects in insulators, then it is
preventive maintenance schedule, however, to the case possible to obtain a neural network capable of diagnosing the
studied in this project, it is a critical point, because there are current condition of insulators under operation, making it
structures which demand washing every week, otherwise the possible to prevent failures and system faults.
pollution severity will certainly cause a system fault event.
III. MACHINE LEARNING AND THE DATASET SIZE ISSUE
Whenever moisture accumulates on the surface of a
polluted insulator, these two components may form a Artificial neural networks (ANN) evolved as computers
conductive layer which increases the leakage current across became more powerful and new network architectures were
the insulator surface. As electric current flows, it also heats being created. The convolutional networks, e.g., allowed
this layer, causing moisture to evaporate, forming dry bands computer vision to develop [4], since it employs filters that
which interrupt the current path. These bands are subjected to convolve over two-dimensional input data in order to create
an electric stress due to the strong electric field across them, feature maps, allowing the classification of images. One of the
which leads to the occurrence of arcing events across them. most recent advances in ANN is the generative adversarial
This arcing is the surface type of partial discharges, which can network (GAN), composed of two models, a discriminator and
also be internal (void) and external in the air (corona effect). a generator, which compete in a classification game and, in the
The corona effect happens when the air is ionized due to a high end, the generator becomes capable of generating fake but
electric field concentration and a free electron starts a process convincing data with features from real data [5]. With the
of an electron avalanche. Partial discharges are characterized power of today computers, the concept of deep learning
as a localized dielectric breakdown which only partially became widespread, but deep neural networks are not only
bridges the insulation between conductors [3]. deep in the sense of many layers stacked sequentially, but also
in their capability of interpreting raw data.
Partial discharges are accompanied by a series of physical
and chemical phenomena, such as ultraviolet radiation and Deep neural networks eliminate the need for a previous
visible light emissions, radiofrequency interference, audible handmade feature extraction step. Convolutional networks
and ultrasonic noise, heating and chemical reactions that may can learn to detect features in input data by updating the
result in toxic and corrosive substances that cause the aging of weights in its filters, so that each filter becomes responsible
insulator surface. All these phenomena described can be for detecting a specific feature. The deeper the layer, the
detected or measured by sensors or by laboratory tests. higher the complexity of features detected by filters.
Ultraviolet cameras are expensive alternatives which most of Traditionally, convolutional networks are used for
the electric utilities do not possess, thus radiofrequency classification tasks, thus, even though they are composed of
receptors with proper antennas and ultrasonic sensors are convolutional and pooling layers in its majority, the last few
preferred for online inspection, but these often depends on the layers are composed of the traditional fully connected layers.
knowledge of experienced personnel. Electrical methods are That is because the convolutional layers make the feature
specified for laboratory tests by specific standards and are extraction step, hence the features found to be relevant by the
based on the measurement of the apparent charge with a backpropagation algorithm are the input to a classification
coupling capacitor in parallel with the test object. It is not step. This idea leads to the conclusion that the last few
possible to measure the actual charge of the discharge, but convolutional layers before the classification section can
with the help of the coupling capacitor it is possible to measure present relevant features extracted from the training data set.
the charge supplied by it to the test object in order to suppress That is the main idea behind GANs, because, in the end,
the voltage variations caused by the PD. Many types of generators models are creating data which synthesize the main
equipment, such as transformer and surge arresters, have a features present in the real samples from the training data set.
maximum limit for the apparent charge, specified by Convolutional networks architecture implements filters
standards, for laboratory tests. However, these limits are not which convolve over a subsection of the input data called
specified for insulators and, due to its inexpensive cost, when receptive field, performing the dot product, and create feature
an insulator in operation presents a problem, it is simply maps from it. One of the benefits of this architecture is the
replaced. weight sharing, which reduces the number of trainable
Partial discharges present patterns which can be indicative parameters, i.e., the same weights in each filter kernel will be
of their origin and magnitude. One of the most used used in the convolution process over all receptive fields of the
representations of their pattern is the phase-resolved partial input data, in other words, a single weight can be connected to
discharge (PRPD) analysis, which exhibits the discharges several elements of the input data [1]. The exact position of a
pulses localization with respect to the phase angle of the feature in the input data is, therefore, irrelevant, since one
power frequency cycle and their magnitude, usually the filter detects the same feature at different locations [1].
apparent charge in picocoulombs. The PRPD has three view However, the relative position to other features in the input
Authorized licensed use limited to: Centro Universitário Fei. Downloaded on September 09,2021 at 17:13:37 UTC from IEEE Xplore. Restrictions apply.
2019 International Conference on Electrical and Computing Technologies and Applications (ICECTA)
Authorized licensed use limited to: Centro Universitário Fei. Downloaded on September 09,2021 at 17:13:37 UTC from IEEE Xplore. Restrictions apply.
2019 International Conference on Electrical and Computing Technologies and Applications (ICECTA)
Authorized licensed use limited to: Centro Universitário Fei. Downloaded on September 09,2021 at 17:13:37 UTC from IEEE Xplore. Restrictions apply.
2019 International Conference on Electrical and Computing Technologies and Applications (ICECTA)
exactly as the original data, it does contain their main features. urements of a same structure, which were not used in the
The results for each sample are shown in Table I. GANs training, were assigned the same label, proving that the
trained classification model is able to learn features from
Then, as expected, the signals used to train the generator different data sets and to identify their presence in totally new
models were successfully assigned the correct labels. But the data. The model trained on synthetic data did not overfit and,
performance of the convolutional model trained for above all, attained a consistent generalization performance.
classification should also be evaluated for other signals,
containing or not the features of the original data set. Thus, The GAN architecture can be used to generate synthetic
picking two sets of distinct patterns signals with four samples but consistent data containing the main features of real data,
each which were used neither for the GANs training nor for thus providing the augmentation of the training set and allow-
the convolutional classifier and presenting them to the ing the original data to be spared for the evaluation of the
classifier, it resulted in the classification shown in Table II. trained model. Therefore, this work sought to investigate the
The two patterns are shown in Fig. 3. GANs potential and to help partial discharge classification
studies. Future work involves fine tuning the GAN model in
It is noticeable that the classifier found relevant features in order to obtain even better results. In addition to that, data aug-
these signals, which caused high activations in some classes, mentation using the proposed framework can be applied to
and assigned these signals the respective labels. It is important
other areas, improving the results of data-based frameworks.
to call attention to the neuron which was activated by a set of
samples containing the same patterns, which, in this case, was ACKNOWLEDGMENT
the same neuron for all samples in the set, showing that it was
properly classified according to the logic and the knowledge The authors would like to thank Equatorial Energia,
learned by the classification convolutional network. ANEEL, CAPES, the High Voltage Laboratory of the Federal
University of Itajubá and every staff member of HVEX.
VI. CONCLUSION
REFERENCES
Data scarcity affects directly the performance of a model;
[1] Y. LeCun et al. “Backpropagation Applied to Handwritten Zip Code
therefore, the proposed framework can be employed to Recognition”, Neural Computation, 1(4), pp. 541-551, Winter 1989.
generate more data whose features are extracted directly from [2] Y. Hoshen, R. J. Weiss, K. W. Wilson. “Speech acoustic modeling
real acquired data. From the tests described in this work, it was from raw multichannel waveforms” Proceedings of 2015 IEEE
possible to notice that the trained model assigned the original International Conference on Acoustics, Speech and Signal Processing
data to the right classes, based on a training made only on syn- (ICASSP). Brisbane: IEEE, 2015.
thetic data from different generators. Spectrograms of meas- [3] International Electrotechnical Commission. “IEC 60270:2000 High-
voltage test techniques - Partial discharge measurements”.
International Standard. Dec. 2000.
TABLE I. CLASSIFICATION RESULTS FOR THE ORIGINAL DATA
[4] Y. LeCun and Y. Bengio. “Convolutional Networks for Images,
Sample Label Output Speech, and Time-Series”, in Arbib, M. A. (Eds), The Handbook of
1 0 [0.995 0.000 0.004 0.000] Brain Theory and Neural Networks, MIT Press, 1995.
2 0 [0.998 0.000 0.002 0.000] [5] I. J. Goodfellow et al. (2014, June 10). “Generative Adversarial
3 0 [0.964 0.000 0.034 0.003] Networks” [On-line]. Available: https://arxiv.org/abs/1406.2661 [July
4 0 [0.990 0.000 0.010 0.000] 27, 2019]
5 1 [0.000 0.925 0.001 0.074] [6] W. Yijiang et al. “Partial Discharge Data Augmentation of High
6 1 [0.000 0.667 0.012 0.321] Voltage Cables based on the Variable Noise Superposition and
7 1 [0.000 0.934 0.001 0.065] Generative Adversarial Network” Proceedings of 2018 International
8 1 [0.000 0.890 0.001 0.109] Conference on Power System Technology (POWERCON).
9 2 [0.044 0.001 0.824 0.132] Guangzhou: IEEE, 2018, pp. 3855-3859.
10 2 [0.023 0.001 0.752 0.224]
[7] X. Wang, H. Huang, Y. Hu, Y. Yang. “Partial Discharge Pattern
11 2 [0.046 0.001 0.808 0.146] Recognition with Data Augmentation based on Generative Adversarial
12 3 [0.000 0.086 0.210 0.704] Networks” Proceedings of 2018 Condition Monitoring and Diagnosis
13 3 [0.000 0.162 0.121 0.718] (CMD). Perth: IEEE, 2018.
14 3 [0.000 0.126 0.110 0.764]
[8] X. Peng et al. “A Convolutional Neural Network-Based Deep Learning
15 3 [0.000 0.255 0.047 0.698]
Methodology for Recognition of Partial Discharge Patterns from High-
Voltage Cables” IEEE Transactions on Power Deliver. vol. 34. n. 4. pp.
1460-1469. Mar. 2019.
[9] G. Li, M. Rong, X. Wang, X. Li, Y. Li. “Partial discharge patterns
recognition with deep Convolutional Neural Networks.” Proceedings
of the Condition Monitoring and Diagnosis, Xi’an, China, 25–28
September 2016. pp. 324–327
[10] G. Li, X. Wang, X. Li, A. Yang, M. Rong. “Partial Discharge
Recognition with a Multi-Resolution Convolutional Neural Network”
Fig. 3. Two signals presented as new data to the trained network.
Sensors 2018, 18(10), 3512.
[11] D. P. Kingma, J. Ba. (2014, Dec. 22). “Adam: A Method for Stochastic
TABLE II. CLASSIFICATION RESULTS FOR NEW DATA
Optimization” [On-line]. Available: https://arxiv.org/abs/1412.6980
Set Output Predicted Class [Dec. 22, 2014].
[0.000 0.340 0.010 0.650] [12] TensorFlow. “Deep Convolutional Generative Adversarial Network”
[0.000 0.167 0.025 0.808] Internet: https://www.tensorflow.org/tutorials/generative/dcgan [Sep.
1 3
[0.000 0.169 0.019 0.811] 23, 2019].
[0.000 0.075 0.063 0.863] [13] J. Brownlee. “How to Develop a GAN for Generating MNIST
[0.000 0.996 0.000 0.004] Handwritten Digits” Internet: https://machinelearningmastery.com/
[0.000 0.998 0.000 0.002] how-to-develop-a-generative-adversarial-network-for-an-mnist-hand
2 1
[0.000 1.000 0.000 0.000] written-digits-from-scratch-in-keras/ June 28, 2019 [Sep. 23, 2019].
[0.000 1.000 0.000 0.000]
Authorized licensed use limited to: Centro Universitário Fei. Downloaded on September 09,2021 at 17:13:37 UTC from IEEE Xplore. Restrictions apply.