2020 IEEE PES/IAS PowerAfrica

Detection of Inter-Turn Faults in Transformers

Using Continuous Wavelet Transform and
Convolutional Neural Network
Desmond Nii Ashitey Hammond, Emmanuel Asuming Frimpong, MIEEE and Joel Yeboah Ohene
Department of Electrical and Electronic Engineering
Kwame Nkrumah University of Science and Technology
Kumasi, Ghana.

Abstract— This paper presents an approach for detecting profiles [2]. Research has therefore gone into the development
and locating transformer inter-turn faults using continuous of improved methods. For example, an approach using winding
wavelet transform (CWT) and convolutional neural transfer functions and an Artificial Neural Network (ANN) has
network (CNN). Transformer three-phase voltages and been proposed in [3]. Another technique uses transformer
currents are used as input parameter. The input data is current and ANN [4]. However, the approaches are not robust
processed using the CWT, resulting in six scaleogram to transformer parameter changes. Another technique based on
images. The six scaleogram images are normalized and Discrete Wavelet Transform (DWT) and Fuzzy Logic has been
concatenated into a single image. The concatenated image is proposed, which enables the detection of faults at incipient
then fed into a trained CNN which indicates whether an stages and can predict fault severity [5]. However, it is highly
inter-turn fault has occurred. Where an inter-turn fault is dependent on the performance of the data acquisition system as
identified, the magnitude/severity of the fault as well as the discrete wavelet transforms are sensitive to poorly conditioned
affected phase is determined. The technique was tested via and noisy signals. Furthermore, other methods such as those
the simulation of a three-phase 630kVA, 10.5kV/0.4kV, based on frequency response analysis [6]-[8], and other
transformer. Test results show that the approach has artificial intelligence techniques [4], [9]-[13] either have (i)
remarkable potential for field application. high implementation cost; (ii) complex operation; (iii) long
operating time; (iv) low performance accuracy or (v) are unable
Index terms: Inter-turn fault; transformer; wavelet transform; to operate on-line. Thus, further work is needed.
neural network This paper contributes to on-going efforts to develop a near
perfect technique for detecting inter-turn faults in transformers.
The technique uses the three-phase currents and voltages of
I. INTRODUCTION transformer as input data. Data processing is done using
continuous wavelet transform which yields scaleogram images.
P ower transformers form integral parts of power systems.
They are responsible for transferring electrical energy from
the power generation stage through to the distribution of
The images are then normalized and concatenated, and fed into
a trained convolutional neural network which indicates
transformer status. The approach is effective in operation and
power. During the operation of transformers, a number of faults
simple to implement. It is also suitable for on-line operation.
could occur which are likely to have dire consequences on the
transformer, if they are not detected and corrected on time. Any
II. CONTINUOUS WAVELET TRANSFORM
fault in transformers leading to power outages or catastrophic
power systems failures causes huge loss of capital, property and Wavelet transform is a commonly used tool for analyzing
in some cases, human casualties. localized variations of signals. It decomposes a time series into
Inter-turn fault is a common fault that occurs in transformers. time–frequency space, allowing the realization of the
Statistics has it that inter-turn faults account for the highest pronounced modes of variability and how they vary in time.
percentage of transformer damages. Between 70% and 80% of There are essentially two types of wavelet transform. These are
transformer failures are due to transformer inter-turn short- Continuous Wavelet Transform (CWT) and Discrete Wavelet
circuits [1]. Inter-turn short-circuits occur between coil turns in Transform (DWT). DWT is more popular in the field of
windings. engineering, because the observed time series are discrete in
Early approaches to detect inter-turn faults relied on real world [14].
Buchhloz relay and Dissolved Gas Analysis. However, they On the other hand, CWT is an effective tool for recording the
have been found to be slow and their operation adversely changing characteristics of non-stationary signals. The CWT is
affected by changing oil levels and similarities between gas a convolution of a signal with a scaled (dilated) and translated

978-1-7281-6746-6/20/$31.00 ©2020 IEEE

Authorized licensed use limited to: University of Nantes. Downloaded on April 16,2022 at 19:31:40 UTC from IEEE Xplore. Restrictions apply.
2020 IEEE PES/IAS PowerAfrica
(shifted) versions of a mother wavelet. The application of CWT the section or winding to be shorted was readjusted to the
can be expressed mathematically as [15]: percentage required using the voltage/turn ratios. All other
1 ∞ *§ t − b ·
windings were adjusted such that the total number of turns
W (a, b ) = ³ x(t )ψ ¨ ¸dt (1) remained unchanged in the winding. Figure 2 shows a
a −∞ © a ¹
MATLAB Simulink model of the transformer. The procedure
where, W(a, b) is the wavelet coefficient and ȥ (t) is the mother for simulation is further outlined as follows:
wavelet. Parameters a and b are scaling and translation factors 1. Initialize transformer power, frequency, voltage ratio,
respectively. current ratio and turn ratios.
Scaling and shifting of a mother wavelet produces child
2. Calculate the turns per volt.
wavelets. Wavelet coefficients are calculated from the
3. Translate the percentage of turns to be shorted in the primary
convolution of child wavelets and the signal. In order to
preserve energy at each scale, the convolution sum is multiplied side to the corresponding number of turns and voltage.
4. For any section of a winding (say middle of winding of
by the factor 1 / a . Different types of mother wavelets have phase b) to be shorted, the number of turns is adjusted to the
been used in literature. Morlet is the most commonly used percentage to be shorted (say 1.5% of total turns) which
mother wavelet. The Morlet wavelet is given by: translates into corresponding voltage, resistance and leakage
t2
1 j 2πf c t − f b inductance. The remaining number of turns are equally
ψ (t ) = e e (2)
πf b distributed into sections a and c such that the total number
of turns of all the three sections in parallel remain
where, fb is the bandwidth parameter of the mother wavelet, fc is
unchanged to give the rated output voltage.
the center frequency, t is the time, and j is the imaginary unit.
5. The circuit breaker attached to the taps of the section to be
The bandwidth parameter controls the shape of the wavelet.
shorted is activated to create the short-circuit at a specific
III. CONVOLUTIONAL NEURAL NETWORK
time during the simulation period.
6. The output terminal voltage and current waveforms of all
Convolutional neural networks (CNNs), also known as shift or phases are then obtained as .csv files to be used for wavelet
space invariant neural networks (SIANNs) enable machines to
analysis.
perform visual tasks like humans [16, 17]. They can execute
This process repeats itself for any given section and percentage
tasks such as detection, segmentation, establishing correlation
and classification [18]. CNNs do not have a preconceived of winding to be shorted in any of the phases.
architecture, parameter selection methods or rules regarding the
number of convolutional layers [19]. They have three major
design features namely, weight sharing, spatial sub-sampling,
and local receptive fields. They have up to four layers. These
are convolutional, pooling, fully connected, and softmax layers.
These layers are composed of a set of neurons with biases,
weights, and transfer functions. CNNs also have two main
processing stages. These are feature extraction and
classification. The feature extraction stage includes the
convolutional layer and the pooling layer. While the
classification stage involves both fully connected layer and
softmax layer [19, 20]. Figure 1 shows the structure of a CNN

Fig. 1. A convolutional neural network

IV. MODELING AND SIMULATION OF TRANSFORMER

A 3-phase 630kVA, 10.5kV/0.4kV transformer was modelled
in Matlab-Simulink as a multi-winding transformer. Data for
modelling was obtained from [21]. For simulation purposes, the
primary and secondary windings of the transformer were
divided into three windings each. This was done for all three
phases of the transformer. Faults were introduced by shorting
out any of the sections (i.e. upper, middle or lower section).
Before each short is applied, the number of turns contained in Fig. 2 A 630kVA 3-phase transformer in MATLAB Simulink
978-1-7281-6746-6/20/$31.00 ©2020 IEEE

Authorized licensed use limited to: University of Nantes. Downloaded on April 16,2022 at 19:31:40 UTC from IEEE Xplore. Restrictions apply.
2020 IEEE PES/IAS PowerAfrica
V. APPROACH TO DATA PROCESSING VII. PROPOSED TECHNIQUE
The approach to data processing is outlined as follows: A flowchart of the proposed approach for detecting inter-turn
1. Obtain, in array data format, the voltage and current of all faults is presented as Fig. 3. The scheme starts by extracting
three phases (i.e. a total of six arrays will be produced). voltage and current samples for all phases, in arrays, and
2. Remove the mean of each signal to eliminate possible DC processing them to remove DC components. The processed data
components in them. is then compiled into a batch of six signals and CWT applied to
3. Compute sampling period, dt, by finding the time difference them to generate six coefficients. Six scaleogram images are
between any two adjacent time steps. then produced from the coefficients and later normalized. The
4. Initialize the mother wavelet specifying its bandwidth normalized images are then concatenated into a single image
parameter and center frequency. and fed into a trained CNN which indicates whether an inter-
5. Initialize the Fourier frequencies to be analyzed to reduce turn fault has occurred. Where an inter-turn fault is identified,
the computational burden on CWT. the magnitude/severity of the fault as well as the affected phase
6. Convert initialized Fourier frequencies into Fourier periods. is determined.
7. Determine the scale of mother wavelet by finding the
product of the period and center frequency. The initialized
Fourier periods correspond to scale values ranging from
approximately 0.0003 to 12. Due to the infinite nature of
scales possible in continuous wavelet transforms, a
logarithmic distribution of the equivalent Fourier
frequencies is used. Smaller scales represent higher
frequencies and larger scales represent lower frequencies.
8. Use computed scales to dilate or translate the mother
Morlet wavelet to generate child wavelets.
9. Generate, from child wavelets, six scaleograms
corresponding to the 6 waveforms (arrays) using the jet
colormap and a logarithmic color scale distribution.
10. Normalize child wavelets.

VI. TRAINING OF CONVOLUTIONAL NEURAL NETWORK

The following were done to train the CNN:
1. The three-phase voltage and current scaleograms (6 images
for each simulation) were loaded from directory in chunks
or batches with their corresponding fault magnitude in
percent (shorted percentage of total windings) and location
of faults (Phases A, B and C).
2. The images were then normalized. This was done to
improve the model’s performance.
3. The normalized images were then concatenated along
channels to produce a single images.
4. The concatenated images were then fed into the CNN.
5. Custom data generator was used to feed randomized training Fig. 3 Proposed fault detection scheme
data into the model in chunks, to prevent memory allocation
(GPU memory) issues.
VIII. RESULTS AND ANALYSIS
The CNN was trained to indicate fault and non-fault conditions.
For fault conditions, the CNN indicates the magnitude/severity Figure 4 shows scaleogram images for three phase voltages and
of the fault (i.e. percentage of winding turns affected) as well as currents for a 0.26% inter-turn fault on phase C (secondary
the phase on which the fault occurred. side). With 0.26% shorted turns, it is virtually impossible to
The processed data was randomized and split into training visually determine and localize the fault due to absence of
and testing data. Eighty percent of the data was used to train the significant transient information in the spectrum. However, as
CNN. Of the remaining 20%, 70% was used for validation and the percentage of turns shorted increases, variations caused by
the remaining 30% used for final testing. the transients and aftermath of fault inception can be easily
visualized and meaningfully interpreted. This human
visualization difficulty is not applicable to the trained CNN.
For the fault condition presented, the proposed technique
produced the following output:

978-1-7281-6746-6/20/$31.00 ©2020 IEEE

Authorized licensed use limited to: University of Nantes. Downloaded on April 16,2022 at 19:31:40 UTC from IEEE Xplore. Restrictions apply.
2020 IEEE PES/IAS PowerAfrica
Predicted Magnitude: [9.144731], Predicted Location: [3] M. Faridi, M. Kharezi, E. Rahimpour, H.R. Mirzaei, & A. Akbari,
“Localization of turn-to-turn fault in transformers using artificial neural
Phase C networks and winding transfer function,” International Conference on
True Magnitude:0.26, True Location: Phase C Solid Dielectrics, Potsdam, Germany, July 4-9, 2010.
[4] P.D. Anagha and R.A. Vikramsingh, “Novel Approach To Power
Transformer Fault Protection Using Artificial Neural Network.
International Journal of Current Engineering and Scientific Research,
vol. 4, no. 2, pp. 68-71, 2017.
[5] K. Ramesh and M. Sushama. “Inter-Turn Fault Detection in Power
Transformer Using Fuzzy Logic”, Journal of Electrical Engineering,
vol. 15, no. 1, pp. 221-226, 2015.
[6] A. Shintemirov, W.H. Tang and Q.H Wu, “Transformer winding
condition assessment using frequency response analysis and evidential
reasoning”, IET on Electircal Power Applications, vol. 4, no. 3, pp.
198-212, 2010.
[7] M. Florkowski and J. Fargul, “A high-frequency method for determining
winding faults in transformers and electrical machines”, Review of
Scientific Instruments, vol. 76, no. 11, pp. 114701–1, 114701–6, 2005.
[8] S.A. Ryder, “Diagnosing transformer faults using frequency response
analysis”, IEEE Electrical Insulation Magazine, vol. 19, no. 2, pp. 16-
22, 2003.
[9] C. P. Babu, M. Surya kalavathi, and B. P. Singh, “Use of wavelet and
neural network (BPFN) for transformer fault diagnosis”, IEEE Ann. Rep.
Fig. 4: Scaleogram images for voltages and currents Conf. Elect. Ins. Diel. Pheno.,93-96, 2006.
[10] R.A. Hooshmand, M. Parastegari and Z. Forghani, “Adaptive Nero-
Thus, the proposed scheme successfully identified the fault as Fuzzy Inference system approach for simultaneous diagnosis of the type
and location of faults in power transformers”. IEEE Electrical Insulation
occurring in phase C but was in error in terms of the magnitude Magazine, vol. 28, no. 5, pp. 32-34, 2012.
of percentage turns involved. The percentage turns was [11] H. Firoozi, M. Kharezi and H. Bakhshi, “Turn-to-Turn Fault
indicated as 9.144731% instead of 0.26%. Localization of Power Transformers Using Neural Network Techniques.
For the same percentage turns fault, but on phase B Int. Conf. Pro. Appl. Diel. Mat, B-11, pp. 249-252, 2009.
[12] M. R. Rao, B. P. Singh, "Detection and localization of inter-turn fault in
(secondary), the output of the CNN is as follows: the HV winding of a power transformer using wavelets", IEEE Tras.
Predicted Magnitude: [6.2556105], Predicted Location: Dielec. and Elec. Ins., vol. 8, no. 4, pp. 652-657, 2001.
Phase B [13] A. Ngaopitakkul and A. Kunakorn, “Internal fault classification in
transformer windings using combination of discrete wavelet transforms
True Magnitude:0.26, True Location: Phase B and back-propagation neural networks”,. Intern. J. Contr. Autom. and
Here too the fault location was successfully identified but the Sys., vol. 4, no. 3, pp. 365-371, 2006.
winding percentage was in error. The indicated value was [14] S. Pittner and S.V. Kamarthi, “Feature Extraction from Wavelet
Coefficients for Pattern Recognition Tasks”, IEEE Transactions on
6.2556105% instead of the expected 0.26%. For a 42.76% inter- Pattern Analysis and Machine Intelligence, vol. 21, no. 1, pp. 83-88.
turn fault on phase B (secondary), the location was properly [15] F. Shiri and B. Mohammadi-ivatloo, “Identification of interarea
identified but with a percentage turn of 47.91702%. Overall, the oscillations using wavelet transform and phasor measurement unit data”,
algorithm was 98% accurate on the determination of fault Int. Trans. Electr. Energ. Syst., 2014.
[16] S S. J. Russell and P. Norvig, “Artificial Intelligence: A Modern
location but not impressive on magnitude (percentage turn) Approach”, 3rd ed., Prentice Hall, 2016.
determination. Further work will seek to address this weakness. [17] R. Zhao, R. Yan, Z. Chen, K. Mao, P. Wang, and R. X. Gao, ‘‘Deep
learning and its applications to machine health monitoring,’’ Mech. Syst.
Signal Process, vol. 115, pp. 213–237, 2019.
[18] L. S. Maraaba, A. S. Milhem, I. A. Nemer, H. Al-Duwaish and M. A.
IX. CONCLUSION Abido, "Convolutional Neural Network-Based Inter-Turn Fault Diagnosis
A continuous wavelet transform and convolutional neural in LSPMSMs," IEEE Access, vol. 8, pp. 81960-81970, 2020
[19] M. Skowron, T. Orlowska-Kowalska, M. Wolkiewicz and C. T.
network approach to inter-turn fault detection has been Kowalski, “Convolutional Neural Network-Based Stator Current Data-
presented. The method uses readily accessible transformer data. Driven Incipient Stator Fault Diagnosis of Inverter-Fed Induction Motor”,
It is an on-line scheme and can be developed as an algorithm in Energies, vol. 13, no. 1475, pp. 1-21, 2020.
existing numerical relays without the need for additional [20] L. Wen, X. Li, L. Gao, and Y. Zhang, ‘‘A new convolutional neural
network-based data-driven fault diagnosis method,’’ IEEE Trans. Ind.
hardware. It is therefore simple and cost-effective to Electron, vol. 65, no. 7, pp. 5990–5998, 2018.
implement. The method effectively identifies and locate faults [21] Y. Wang, X. Zhao, J. Han, H. Li, Y. Guan, Q. Bao, L. Xiao, L. Lin, X.
but is ineffective in determining the percentage of turns Xu, N. Song, and F. Zhang, “Development of a 630 kVA Three-Phase
affected. Further work is therefore needed to address this HTS Transformer with Amorphous Alloy Cores”, IEEE Transactions on
Applied Superconductivity, vol. 17, no. 2, pp. 2051-2054, 2007.
deficiency.

REFERENCES
[1] H. Wan and K. Butler, “Finite Element Analysis of Internal Winding
Faults in Distribution Transformers”. IEEE Transactions on Power
Delivery, vol. 16, 2001.
[2] N.C. Joshi, R.Y. Sood, R.K Jarial and R. Thapliyal, “Transformer
Internal Winding Faults Diagnosis Methods: A Review, MIT
International Journal of Electrical and Instrumentation Engineering, vol.
2, no. 2, pp. 77-81, 2012.

978-1-7281-6746-6/20/$31.00 ©2020 IEEE

Authorized licensed use limited to: University of Nantes. Downloaded on April 16,2022 at 19:31:40 UTC from IEEE Xplore. Restrictions apply.