You are on page 1of 6

Fault Classification in Three Phase Self-Excited

Induction Generators using Deep Neural Networks


Sohom Mukherjeea , Arindam Duttab and Saradindu Ghoshc
Department of Electrical Engineering
National Institute of Technology Durgapur
Durgapur 713209, India
Email: sohom31011997@gmail.coma , arindamdutta1996@gmail.comb , sghosh.ee@gmail.comc

Abstract—In this paper, an algorithm is proposed for the of power at any instant of time. Proper detection and lo-
classification of faults in three phase self-excited induction calization of faults is very crucial to allow relays to act
generators using deep neural networks, on the basis of their accordingly, so as to prevent a grid failure. With microgrids
voltage and current waveforms. Three phase self-exited induction
generators, which are mostly used in wind power stations, are being employed on a larger scale, it becomes necessary
often connected to the national grid. Therefore, the transient to study fault classification of SEIG, which is an integral
stability analysis of this machine, prior and post symmetrical part of most microgrid systems. [8]–[10] present algorithms
and unsymmetrical short circuit faults is one of the main based on artificial neural networks (ANN), support vector
concerns in power system security and operation. In this study, machines (SVM) and adaptive neural fuzzy inference system
voltage and current waveforms of faults have been simulated
in the Simulink environment, for different conditions of fault. (ANFIS) for classification of power system faults. [11], [12]
Following this, visual time-frequency representations of the fault have employed various intelligent and hybrid fuzzy logic-
signals called scalograms are created, using continuous wavelet based techniques for fault classification. With the advent of
transform. Finally, a deep convolutional neural network is used high-voltage direct current (HVDC) transmission systems,
for classification of the fault signals. Experimental results show researchers have proposed various algorithms for classification
that a final accuracy of 82.14% as well as real-time inference is
achieved on the validation set, using the proposed scheme. of faults in them as well [13]. [14] presents a work on
Index Terms—Fault classification, Self-exited induction gen- fault classification in cascaded H-bridge converters and [15]
erator, Transient analysis, Continuous wavelet transform, Deep presents a paper on fault classification in microgrids. This
convolutional neural network clearly shows that fault classification has become an integral
domain of research in modern power systems. However, most
I. INTRODUCTION of the existing literature on SEIGs have confined themselves
The fact that conventional resources are depleting at alarm- to the transient analysis of the machine [16], [17], while fault
ing rates makes the world look up to renewable energy classification of SEIGs has been left unexplored.
resources to sustain the ever increasing energy demand. Con- One of the recent breakthroughs in the area of signal
sequently, solar and wind energy conversion systems have analysis and artificial intelligence has been the advent of
become one of the prime areas of attention. Solar energy con- deep neural networks (DNNs). DNNs have become popular
version systems employ photovoltaic (PV) cells for producing across multiple application domains, such as image recog-
electrical energy, while wind energy conversion systems use nition, speech recognition and natural language processing,
a turbine shaft connected to a generator. For wind energy on account of their state-of-the-art performances [18]. The
conversion systems, self-exited induction generators (SEIGs) success of DNNs can be attributed to the multiple levels of
are favoured over conventional synchronous generators due representation that can learn complex functions without the
to their inherent advantages, such as low cost and reliability. need for hand-crafted features. However, very little attention
Moreover, there are several schemes available in literature that has been devoted to employ the strong representation learning
allow them to be used for domestic power supply, which capability of DNNs to the classification of power system faults.
is generally adopted at remote locations [1], [2]. The fact The following major contributions have been made in this
that SEIGs are a vital component of wind energy conversion work to address the aforementioned issues:
systems makes their analysis necessary. [3], [4] present various • Continuous wavelet transform (CWT) is employed to
aspects of steady state analysis using different techniques. create visual time-frequency representation of fault sig-
Recently, soft computing algorithms have also been used to nals called scalograms. The scalograms for each fault
analyze the steady state behaviour of the machine [5]. Another class have their distinctive visual features, making them
aspect which is crucial to the working of the SEIG is its self- suitable for leveraging the learning capability of DNNs.
excitation, and considerable work has been done on this aspect • A deep convolutional neural network (DCNN) is used for
[6], [7]. fault classification, rather than conventional techniques
Fault classification is a major concern in power systems. such as ANN and SVM, thereby overcoming the bottle-
This is because, modern power systems deal with huge amount neck imposed by hand-crafted feature extraction.
where, r, y, b are the three phases of the stator winding, and
Vdr = Vqr = 0. (6)
Also, [R], [L] and [G] are 4×4 matrices defined by equations
(7), (8) and (11), respectively.
 
Fig. 1: Schematic circuit diagram for the SEIG considered Rs 0 0 0
for simulation. SCIM is a squirrel cage induction motor  0 Rs 0 0
[R] =   (7)
(with delta connected stator windings) connected to a delta 0 0 Rr 0 
connected resistive load. Capacitors have been used for the 0 0 0 Rr
purpose of self-excitation, to allow use of SCIM as an SEIG.
where, Rs is the stator resistance.
 
Lss 0 Lm 0
The remaining sections of the paper are organized as  0 Lss 0 Lm 
follows. In Section II, the transient analysis of self-exited [L] = 
Lm
 (8)
0 Lrr 0 
induction generators along with simulation of various types of
0 Lm 0 Lrr
faults in Simulink environment have been discussed. Section
III provides a detailed account of the proposed algorithm where,
including fault signal preprocessing and fault classification. Lss = Lls + Lm (9)
Experimental results and comparisons have been provided in
Section IV. Finally, Section V draws certain conclusions based Lrr = Llr + Lm (10)
on the results obtained and discusses the scope for future work. Here, Lss , Lm and Lrr refer to equivalent stator inductance,
II. TRANSIENT ANALYSIS OF SEIG magnetizing inductance and equivalent rotor inductance, re-
spectively. Whereas, Lls and Llr are the leakage inductance
Transient analysis of any machine is important since it of stator and rotor winding respectively.
allows us to study the behaviour of the machine under various  
conditions of faults and load changes. Under conditions of 0 0 0 0
unbalance, SEIGs incur additional power losses owing to the  0 0 0 0 
[G] =   (11)
flow of negative sequence current. Unbalanced conditions also  0 −Lm 0 −Lrr 
cause excessive vibrations and stress on insulation. Hence, val- Lm 0 Lrr 0
ues of transient voltages and currents are important quantities
The shaft torque may be given by:
and must be known for design of generator winding and related
protection system. Tshaf t = α − β × Wr (12)
The d-q model method as described in [16], has been used
to investigate the transient characteristics of SEIG. The d-q where, α and β are machine constants. Moreover, SEIG has
model axes can be either fixed or rotating at a synchronous a non-linear magnetizing characteristic and operates in the
speed. We have chosen the d-q stationary frame because the saturation region. Magnetizing current (Im ) is given as
frequency of generator voltage and current vary under transient √
Im = [(Ids + Idr )2 + (Iqs + Iqr )2 ]0.5 / 2 (13)
conditions.
The fundamental V-I equation for a SEIG is given by: The capacitors (with capacitance C), used for self excitation
may be modelled using:
[V ] = [R][I] + [L][I]0 + Wr [G][I] (1)
Vr0 = (Ibr − Iby )/3C (14)
where Wr is the rotor speed, and
 T Vy0 = (Ibr − 2Iby )/3C (15)
[V ] = Vds Vqs Vdr Vqr (2)
The phase currents can be calculated by using the following
 T equation:
[I] = Ids Iqs Idr Iqr (3) [Iphase ] = [H][Idq ] (16)
0
From equation (1), [I] , the time derivative of current may where,  
be obtained as: 1 √0
[H] = −0.5 √3/2 (17)
[I]0 = [L]−1 ([V ] − [R][I] − Wr [G][I])

(4)
−0.5 − 3/2
The fact that we have taken a three phase delta connected  T
squirrel cage induction machine implies: [Iphase ] = Ir Iy Ib (18)
 T
Vr + Vy + Vb = 0 (5) [Idq ] = Ids Iqs (19)
(a) LG Fault. (b) LL Fault.

(c) LLG Fault. (d) LLLG Fault.


Fig. 2: Current and voltage waveforms of simulated fault signals.

The above equations have been used to build a Simulink is precomputed, which is the preferred method when obtaining
model, similar to the one presented in [19]. The schematic the CWT of many signals using the same parameters.
circuit diagram of the SEIG used for simulation has been 1) Continuous Wavelet Transform: Most of the signals
shown in Fig. 1. Using the model, four shunt fault conditions encountered in the physical world are in the time domain.
on the machine, viz. single line to ground fault (LG), line However, the most distinguishing features of a particular signal
to line (LL) fault, double line to ground (LLG) fault and are often hidden in its corresponding frequency spectrum.
three phase to ground (LLLG) fault have been simulated. The Therefore, various transform techniques are widely used to
current and voltage waveforms for the machine under the four find the frequency spectrum of a given signal. One of the most
shunt fault conditions have been depicted in Fig. 2. It can be popular transform is the Fourier transform. However, it outputs
seen both current and voltage go to zero after the fault. This only the various frequency components present in a given
is because of the fact that the machine is immediately isolated signal, but does not cast any light upon at what time these
after the fault is simulated by action of protective relays. frequency components occur. Thus, Fourier transform does
not work for non-stationary signals. A modified approach to
III. FAULT CLASSIFICATION the problem is given by short-time Fourier transform (STFT),
which limits the part of the function which undergoes a Fourier
A. Fault Signal Preprocessing transform, by using a window function. In this case, however,
In the signal preprocessing step, time-frequency represen- a good time-frequency resolution is not obtained.
tations of the fault signals, called scalograms, are created. A This issue is well resolved by use of the wavelet transform
scalogram is a visual representation of the absolute value of [20], which provides accurate time-frequency plots by use
continuous wavelet transform (CWT) coefficients of a signal. of multiresolution analysis. The plot is actually three dimen-
There are three axes: x axis representing time, y axis repre- sional, with time, frequency and amplitude as the three axes.
senting frequency, and z axis representing coefficient values. The continuous wavelet transform of a function x(t) at a scale
The z axis is often shown by varying the color or brightness in s and translational value t is given by the following equation:
a two dimensional heat map representation of the scalogram. Z ∞  
1 u−t
A scalogram is the equivalent of a spectrogram for Fourier CW Tc (t, s) = x(u) √ ψ ∗ du (20)
−∞ s s
transform, and can be used for estimation of instantaneous
frequency (IF). To create the scalograms, a CWT filter bank where ψ(t) is the mother wavelet which satisfies the following
Input
Image

Conv 1 Conv 2 Conv 3 Conv 4 Conv 5 FC 1 FC 2 FC 3

Max Max Max

27 ✕ 27 ✕ 256

13 ✕ 13 ✕ 384

13 ✕ 13 ✕ 256
227 ✕ 227 ✕ 3

13 ✕ 13 ✕ 384
55 ✕ 55 ✕ 96
Pooling Pooling Pooling
2✕2 2✕2 2✕2

4096

256

4
Conv 11✕11 Conv 3✕3

Conv 5✕5 Dense

Fig. 3: Modified AlexNet Architecture

R∞
condition −∞ ψ(t)dt = 0. produces an output layer having size C × H × W , where C
2) Scalogram: The continuous wavelet transform of the is the number of channels, H is the height in pixels and W
discrete sequence x(nT ), sampled with a period T in τ is is the width in pixels. The last three FC layers have 4096,
defined as a convolution of the discrete sequence with a scaled 256 and 4 uints, respectively. The ReLU activation function
and translated version of ψ(nT ) is applied after every convolutional and fully connected layer.
X Deep convolutional neural networks with Rectified Linear Unit
CW Td (t, ω) = T x(t + nT )Ψ∗ (nT, ω) (21) (ReLU) [23] train several times faster than their equivalents
n with tanh units. Max pooling layers are introduced after the
first, second, and fifth convolutional layers to scale down
The time-frequency energy density representation obtained
spacial information. Dropout has been applied before the first
by using wavelet transform is called the scalogram and may be
and the second fully connected layer to reduce over-fitting.
be defined as square of the amplitude of the wavelet transform
Dropout [24] is a technique for training deep neural networks
[21]
that reduces complex co-adaptations of neurons by setting the
output of each hidden neuron with probability 0.5 to zero.
W (t, ω) = CW Td (t, ω)CW Td∗ (t, ω)
XX IV. RESULTS AND ANALYSIS
= T2 x(t + n1 T )x∗ (t + n2 T ) (22)
n1 n2 In this section, we present the experimental setup for
× Ψ∗ (n1 T, ω)Ψ(n2 T, ω) fault simulation and fault signal preprocessing, followed by
a detailed analysis of hyperparameter tuning for training the
B. Fault Signal Classification proposed DCNN. The experiments have been performed using
Training a deep convolutional neural network (CNN) from MATLAB R2018a on a machine equipped with an Intel Core
scratch is computationally expensive and requires a large i5 Processor and 8GB RAM.
amount of training data. Therefore, in applications where
sufficient amount of training data is not available, it is a A. Fault Simulation
common practice to leverage existing neural networks that For fault simulation in Simulink, the MATLAB implemen-
have been trained on large datasets for conceptually similar tation [25] of the ordinary differential equation (ODE) solver
tasks. This leveraging of existing neural networks is called algorithm proposed in [26] has been used to solve the related
transfer learning. In this work, the AlexNet architecture [22], differential equations, with an average time of 3 minutes per
pre-trained for image recognition, has been adapted to classify simulation. The specifications of the machine (Fig. 1) used
fault signal waveforms based on their time-frequency represen- for simulation are as follows: 2.2 kW, 230 Volts, 7.8 Amps,
tation. 4-pole, three phase delta connected stator winding, Lls = Llr
AlexNet is a deep CNN architecture that was originally = 0.0142 H, Rs = Rr = 2.88 Ω, C = C1 = C2 = C3 = 50
designed to classify images belonging to 1000 categories µF, α = 249.4, β = 0.788. Magnetizing characteristics are as
and famously won the ImageNet LSVRC-2012 competition. follows:
It contains five convolutional layers followed by three fully
connected layers (Fig. 3). The last two fully connected layers 
are modified to classify our 4 fault classes instead of 1000 0.32
 Im ≤ 0.7
2
= 0.35 − 0.035Im + 0.002Im
classes as in the original AlexNet architecture. The input Lm 0.7 ≤ Im ≤ 4.3
image size should be 227 × 227. Each convolution operation

0.177 Im ≥ 4.3

(a) LG Fault. (b) LL Fault. (c) LLG Fault. (d) LLLG Fault.

(e) LG Fault. (f) LL Fault. (g) LLG Fault. (h) LLLG Fault.
Fig. 4: Scalograms of fault signals. (a)-(d) Phase current. (e)-(h) Phase Voltage.

TABLE I: Variation of Accuracy with Learning Rate


Optimizer Batch Max Initial Learning Accuracy
Size Epochs Learning Rate
Rate Drop
Factor
SGD 15 20 0.0001 0.2 0.7500
SGD 15 20 0.0001 - 0.7857
SGD 15 20 0.0010 0.2 0.8214
SGD 15 20 0.0010 - 0.7500

(a) TABLE II: Variation of Accuracy with Batch Size


Optimizer Batch Max Initial Learning Accuracy
Size Epochs Learning Rate
Rate Drop
Factor
SGD 15 20 0.001 0.0001 0.8214
SGD 50 20 0.001 0.0001 0.6425
SGD 75 20 0.001 0.0001 0.3968
(b)
Fig. 5: Training and Validation results. (a) Accuracy. (b) Loss used for training and 280 samples (70 per class) are used for
validation.
For simplicity, the load has been assumed to be resistive
C. Fault Signal Classification
(Fig. 1), i.e., R1 = R2 = R3 = 800 Ω. Our assumption is
also justified by the observation that change of load resistance The model is trained using the stochastic gradient descent
changes only the amplitude of signals and not their frequency. (SGD) optimizer with a batch size of 15 examples and
momentum of 0.9. It is found that a certain amount of weight
B. Fault Signal Preprocessing decay is important for the model to learn and it reduces the
Scalograms are obtained from fault current and voltage models training error significantly. This has been demonstrated
signals using the ’Morse’ wavelet and with sampling frequency in Table I. It is found that an initial learning rate of 0.001 with
of 128 Hz (Fig. 4). Wavelet analysis is performed only for a learning rate drop factor of 0.2 in every 10 epochs gives
initial transient part of signals since the transients are rich the best accuracy of 82.14 %. Table II shows the effect of
in harmonics, making the fault signals distinguishable. The variation of batch size on the accuracy. It is clear a batch size
current and voltage scalograms of a particular class have sim- of 15 gives the best performance and the accuracy decreases
ilar appearance, which is different from that of another class. with increase in batch size. This is in confirmation with results
This makes the scalograms an ideal input for classification obtained in recent work showing that small batch sizes provide
by a CNN. The data is symmetric, with each fault class better training stability and generalization performance [27].
having 360 samples. In total, 1440 scalograms (360 per class) Fig. 5 shows a plot of the accuracy and loss of the model
are obtained, out of which 1160 samples (290 per class) are versus number of epochs on the training and validation sets.
TABLE III: Comparison of algorithms for fault classification [6] E. Levi and Y. Liao, “An experimental investigation of self-excitation
Classification Accuracy (%) Inference Time (ms) in capacitor excited induction generators,” Electric Power Systems
Algorithm Research, vol. 53, no. 1, pp. 59–65, 2000.
[7] D. Joshi, “Analysis and control of self-excited induction generators,”
Fuzzy Logic [28] 58.16 0.02
2018.
ANN [8] 69.03 0.82
[8] M. Jamil, S. K. Sharma, and R. Singh, “Fault detection and classification
ANFIS [10] 76.45 0.76
in electrical power transmission system using artificial neural network,”
SVM [9] 78.26 0.95 SpringerPlus, vol. 4, no. 1, p. 334, 2015.
Our (DCNN) 82.14 2.05 [9] N. R. Babu and B. J. Mohan, “Fault classification in power systems
using emd and svm,” Ain Shams Engineering Journal, vol. 8, no. 2, pp.
103–111, 2017.
[10] J. Zhang, Z. He, S. Lin, Y. Zhang, and Q. Qian, “An anfis-based
A comparison of the proposed approach (DCNN) with fault classification approach in power distribution system,” International
various classical techniques in terms of accuracy as well as Journal of Electrical Power & Energy Systems, vol. 49, pp. 243–252,
inference time of the algorithm has been shown in Table III. 2013.
[11] A. Yadav and A. Swetapadma, “Enhancing the performance of trans-
It is evident that the proposed DCNN algorithm outperforms mission line directional relaying, fault classification and fault location
all other approaches by large margins in terms of accuracy schemes using fuzzy inference system,” IET Generation, Transmission
(improvement of 3.88% compared to previous state-of-the-art & Distribution, vol. 9, no. 6, pp. 580–591, 2015.
[12] V. Ferreira, R. Zanghi, M. Fortes, G. Sotelo, R. Silva, J. Souza,
SVM [9]), at the cost of slight increase in inference time. The C. Guimarães, and S. Gomes Jr, “A survey on intelligent system
slightly higher inference time is negligible compared to the application to fault diagnosis in electric power system transmission
improvement in accuracy, and can be attributed to the large lines,” Electric Power Systems Research, vol. 136, pp. 135–153, 2016.
[13] H. Fathabadi, “Novel filter based ann approach for short-circuit faults
number of parameters involved in a deep neural network. The detection, classification and location in power transmission lines,” In-
performance can be improved by using a system equipped ternational Journal of Electrical Power & Energy Systems, vol. 74, pp.
with GPU. However, since systems equipped with GPU may 374–383, 2016.
[14] T. Wang, H. Xu, J. Han, E. Elbouchikhi, and M. E. H. Benbouzid,
not always be available, we show results on the CPU, thereby “Cascaded h-bridge multilevel inverter system fault diagnosis using a pca
providing a fair comparison with classical techniques. and multiclass relevance vector machine approach,” IEEE Transactions
on Power Electronics, vol. 30, no. 12, pp. 7006–7018, 2015.
V. CONCLUSION AND FUTURE SCOPE [15] A. Hooshyar, E. F. El-Saadany, and M. Sanaye-Pasand, “Fault type clas-
sification in microgrids including photovoltaic dgs,” IEEE Transactions
This work presented a new fault classification algorithm on Smart Grid, vol. 7, no. 5, pp. 2218–2229, 2016.
for self-excited induction generators using continuous wavelet [16] B. Singh, S. Murthy, and S. Gupta, “Transient analysis of self-excited
induction generator with electronic load controller (elc) supplying
transform for deriving visual time-frequency representations of static and dynamic loads,” IEEE Transactions on Industry Applications,
fault signals (scalograms) followed by a deep convolutional vol. 41, no. 5, pp. 1194–1204, 2005.
neural network for fault classification. The proposed fault [17] R. Dashti, A. Vahedi, B. Ranjbar, R. Dashti, and D. Babachahi, “Tran-
sient analysis of induction generator jointed to network at balanced and
classification algorithm is a generalized one, in the sense that it unbalanced short circuit faults,” in 2007 42nd International Universities
is not dependent on any hyperparameters that are specific to a Power Engineering Conference. IEEE, 2007, pp. 102–108.
particular machine. This is in contrary to traditional rule-based [18] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” nature, vol. 521,
no. 7553, p. 436, 2015.
algorithms or learning algorithms, which use hand-crafted [19] A. Kishore, R. Prasad, and B. Karan, “Matlab simulink based dq mod-
features. The generalizability of the proposed algorithm makes eling and dynamic characteristics of three phase self excited induction
it all the more elegant, since it can be trained for practically generator,” Piers Online, vol. 2, no. 3, pp. 312–316, 2006.
[20] I. Daubechies, “The wavelet transform, time-frequency localization and
any machine. Future work will aim to incorporate sequence signal analysis,” IEEE transactions on information theory, vol. 36, no. 5,
models such as Hidden Markov Models (HMMs) and recurrent pp. 961–1005, 1990.
neural networks (RNNs) into the proposed fault classification [21] E. Sejdic, I. Djurovic et al., “Quantitative performance analysis of
scalogram as instantaneous frequency estimator,” IEEE Transactions on
algorithm, enabling it to better capture temporal information Signal Processing, vol. 56, no. 8, pp. 3837–3845, 2008.
in fault signals. [22] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Advances in neural infor-
R EFERENCES mation processing systems, 2012, pp. 1097–1105.
[23] V. Nair and G. E. Hinton, “Rectified linear units improve restricted boltz-
[1] S. Murthy, B. Singh, S. Gupta, and B. Gulati, “General steady-state mann machines,” in Proceedings of the 27th international conference on
analysis of three-phase self-excited induction generator feeding three- machine learning (ICML-10), 2010, pp. 807–814.
phase unbalanced load/single-phase load for stand-alone applications,” [24] G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and R. R.
IEE Proceedings-Generation, Transmission and Distribution, vol. 150, Salakhutdinov, “Improving neural networks by preventing co-adaptation
no. 1, pp. 49–55, 2003. of feature detectors,” arXiv preprint arXiv:1207.0580, 2012.
[2] G. Singh, “Self-excited induction generator research–a survey,” Electric [25] L. F. Shampine and M. W. Reichelt, “The matlab ode suite,” SIAM
Power Systems Research, vol. 69, no. 2-3, pp. 107–114, 2004. journal on scientific computing, vol. 18, no. 1, pp. 1–22, 1997.
[3] A. Tandon, S. Murthy, and G. Berg, “Steady state analysis of capacitor [26] J. R. Dormand and P. J. Prince, “A family of embedded runge-kutta
self-excited induction generators,” IEEE Transactions on Power Appa- formulae,” Journal of computational and applied mathematics, vol. 6,
ratus and Systems, no. 3, pp. 612–618, 1984. no. 1, pp. 19–26, 1980.
[4] A. L. Alolah and M. A. Alkanhal, “Optimization-based steady state anal- [27] D. Masters and C. Luschi, “Revisiting small batch training for deep
ysis of three phase self-excited induction generator,” IEEE Transactions neural networks,” arXiv preprint arXiv:1804.07612, 2018.
on Energy conversion, vol. 15, no. 1, pp. 61–65, 2000. [28] R. Mahanty and P. D. Gupta, “A fuzzy logic based fault classification
[5] A. Chatterjee, K. Roy, and D. Chatterjee, “A gravitational search approach using current samples only,” Electric power systems research,
algorithm (gsa) based photo-voltaic (pv) excitation control strategy for vol. 77, no. 5-6, pp. 501–507, 2007.
single phase operation of three phase wind-turbine coupled induction
generator,” Energy, vol. 74, pp. 707–718, 2014.

You might also like