Professional Documents
Culture Documents
Ijeei, Laximipriya
Ijeei, Laximipriya
Ijeei, Laximipriya
Abstract: This article presents a novel Vector Quantized Eigen technique for effective
classification of Power Quality (PQ) events. The Stockwell Transform (ST), and Empirical
Mode Decomposition (EMD) are used as baselines to develop individual feature matrices. The
MATLAB platform has been used to generate the PQ events synthetically. Several statistical
parameters along with the Vector Quantized Coefficients and Eigenvectors are computed from
each matrix for further analysis. Finally, a novel Vector Quantized Eigen Mapping algorithm
(VQEM) is proposed in a combinational platform for better efficacy. The lower-order Eigen-
coefficients (LECs) are more revealing as compared to the Principal Coefficients, hence are
contemplated. The reliability of the proposed VQEM is attributed to the vector dimensionality,
correlation, and Probability Density Function (PDF) of VQ to cluster irrelevant data and the
feature selection ability of eigenvectors. For classification, the Cost-Sensitive Decision Tree
(CSDT) algorithm is chosen as it assumes implicitly the cost of all the recognition errors to be different.
Results reveal an improved classification accuracy, the highest %Best cost savings, and a lower
Friedman Ranking (F-Rank) with the proposed technique as compared to other cited works of literature in
this field.
1. Introduction
Power quality (PQ) disturbances such as sag, swell, transients, etc. occurring in a signal during its
generation and distribution lead to malfunctioning of electrical devices and components. A survey
conducted by the Electrical power research Institute shows the loss due to faulty power systems
amounts to approximately $15 billion to $24 billion in the U.S. economy each year [1]. The last decade
has addressed the issue to some extent; however, the provisioning of clean power for domestic and
industrial use still concerns society [2-7]. The complexity involved in identifying the sources of
disturbances, characterization, and classifying them using appropriate signal processing models
makes the domain a challenging field of research in recent years.
The community has relied on several tools such as the Short-Time Fourier Transform (STFT), WT,
Hilbert Transform (HT), Stockwell Transform (ST), Curvelet transform, EMD, etc. to deal with PQ [4-
15]. However, these conventional algorithms often result in high-dimensional data comprising redundant
information. It leads to a decrease in classification accuracy and an increase in memory space and
classification time [5-6], [9-10], [12-14]. The application of feature selection and optimization
algorithms has indeed alleviated these issues. The claimed accuracy has been 82.2%, 98.2%,
97.4%, and 98.5% using KNN, expert system, PNN, and DT respectively with 50 dB noisy
data [8]. Similarly, the EMD and WT statistics have improved the accuracy to 97.57%,
95.28%, and 95.33% respectively with the PNN and SVM classifiers [7] [16]. Other feature
selection techniques such as Principal Component Analysis (PCA), Independent Component
Analysis (ICA), Vector Quantization (VQ), Sequential Forward Selection (SFS), maximum
relevance minimum redundancy algorithms, Genetic Algorithm (GA), and Competitive Swarm
Optimization (CSO), etc. Whale and wild-goat Optimization Algorithms (WOA) have
outperformed the conventional high-dimensional feature vectors of faulty signals [9, 11-14, 17-
856
Laxmipriya Samal, et al.
18]. These algorithms maximize the interclass division and minimize the intra-class
compaction, hence motivating the authors.
In this work, the authors illustrate a two-stage feature selection algorithm by combining the
VQ and Eigen Value Decomposition (ED). The objective is to extract appropriate parameters
from a few chosen disturbance signals, which can improve recognition accuracy. The VQ
algorithm obeys vector dimensionality, correlation, and Probability Density Function while
forming cluster centers, hence remaining more versatile than the conventional PCA in
analyzing the non-stationary signal [18]. On the contrary, the application of ED helps to select
the relevant features from the clustered VQ coefficients using a projected feature space.
Furthermore, the design of the VQ codebook depends on the number of levels, whereas, the ED
technique is dimensional-dependent. Thus, the simultaneous application of both the levels and
dimensions is likely to fetch more information that is relevant and hence considered. This study
conceptualizes the formation of the novel Vector Quantized Eigen Mapping algorithm
(VQEM) algorithm differently. It extracts the lower-order Eigen coefficients (LECs) for the
intended analysis, unlike PCA, which emphasizes the principal components (PCs). The basic
argument is to retain only those eigenvalues having the least distance among the extracted
coefficients, which can provide a greater number of correlated features for better utility and
efficiency. The authors have applied the Bhattacharyya distance norm and observed that the
LECs provide a minimum distance as compared to the PCs [17]. Finally, the extracted features
are considered in a combinational platform for more available information and enhanced
performance.
The rest of the article is organized as follows. Section 2 elaborates on the proposed feature
extraction technique. Section 3 briefs the classification algorithms employed. Section 4
discusses the simulation results and section 5 concludes the work with new research directions.
A. Mapping of ST Matrix
The proposed mapping of the ST matrix [15] is accomplished as shown in Figure 1.
857
Power Quality Disturbance Classification using Vector Quantized Eigen
f. Continue performing steps (c) to (e) until the rows of at each are defined.
(3)
(4)
• where is a complex matrix of size in which each row and each column
correspond to a certain frequency and time respectively. The term denotes the
number of disturbance signals present in a PQ event.
• From the matrix , ten-statistical parameters such as the Mean (F1), Standard Deviation
(F2), range (F3), Skewness (F4), kurtosis (F5), crest factor (F6), the form factor (F7), The
Shannon entropy (F8), energy (F9), and RMS (F10) are computed as given in equation (4)
for the Sag-PQ event. The resultant matrix has a dimension , which is
very small as compared to the ST-matrix of size .
858
Laxmipriya Samal, et al.
(5)
• Retain the last thirty-one LECs per signal based on a threshold for further processing. For a
PQ event comprising signals, the size of the ST-LEC vectors thus
becomes .
The application of the VQ [13, 18] to the ST-matrix has been formulated as follows.
• Apply VQ to each column of the ST-matrix of dimension . The resultant ST-VQ
vector of a signal has the size .
The proposed ST-VQEM vector for the Sag PQ event is formulated as given below.
• The ST-VQ matrix of the Sag PQ event ( ) is expressed as
(6)
• Apply ED to each column of equation (4) with necessary zero paddings. The resultant
matrix has a size of out of which only 250 LECs or ST-VQEM features are retained
for further processing.
859
Power Quality Disturbance Classification using Vector Quantized Eigen
Ten chosen statistical, ED, and VQ techniques are applied to each IMF level as explained
earlier. The resultant sizes of the EMD-statistical, EMD-EM, EMD-VQ, and EMD-VQEM
feature matrices are , , , and 30 respectively.
where denotes the cost of misclassifying an event into another PQ event and is the
conditional probability that the feature value belongs to class .
The lowest cost of recognizing all the feature values of an event is computed as
(9)
In this equation, denotes the learning algorithm predicting all the feature values of the set as
belonging to the desired class . denotes the learning algorithm predicting all the feature elements of
the set as to the desired class . is the lowest cost of recognizing all the feature values of
an event. The cost of saving is expressed as in [19].
860
Laxmipriya Samal, et al.
signals, hence representing non-stationary PQ signals accurately. It is flexible and the mono-
components or narrow band of frequencies makes it more adaptive for aperiodic signals.
Table 2 provides the highest savings in cost (% Best) versus the F-rank corresponding to
the chosen classifiers using different derived feature vectors. The statistical parameter cannot
address the dynamism involved in a feature vector. On the contrary, the Eigenvectors signify
the optimum modes and estimate distinct correlated variables using the orthogonal linear
transformation. Thus, it improves the reliability and discrimination of the features with greater
visibility.
Table 2. Comparison of Cost of the Classification Algorithms (% Best) Versus F-Rank using
the Derived Feature Vectors
Classifier Method % Best F-Rank Method % Best F-Rank Method % Best F-Rank
CSDT EMD-EM 92.6 4.1 EMD-VQ 97.1 1.9 EMD-VQEM 97.7 1.1
Table 3. The % Best cost saving compared with the earlier methods to the proposed method is
provided
% %
Literature Patterns Feature Feature Selection Classifier
Best Accuracy
MDA 71.5
Jabeur, et al., 2020 Spain financial MLR 75.13
bond rating Fisher’s test -
[20] data SVM 67.50
DNN 68.35
Krawczyk, et al., SMOT SE: 75.98
imbalanced KEEL Repository datasets evolutionary algorithm -
2014 [21] CSDT SP: 92.98
Credit card fraud detection, DT 35.33
Bahnsen, et al, 2015 Real-world churn modeling, credit RF 77.37
sampling -
[19] example scoring, and direct LR 63.39
marketing CSDT 88.69
<50
CS-C4.5
Li, et al., 2015 [22] Bio-medical UCI Repository ASCP and ARA <70 -
CSDT
C5.0 85.0
Sahin, et al., 2013 Fraud credit
Statistical SVM 78.3 -
[23] detection card
CSDT 94.9
PQ EMD-VQ-EM+EMD-Stat VQ+ Eigen Value 96.9%
Proposed work CSDT
Disturbance EMD-VQ-EM Decomposition 99.7%
861
Power Quality Disturbance Classification using Vector Quantized Eigen
The proposed VQEM combines the nonlinearity phenomena of VQ and the linearity
phenomena of ED. It exhibits better outcomes among all the derived feature vectors as
observed in this Table.
The % Best cost saving as the performance measure of patterns has been successfully
employed in the field of fraud detection, credit card recognition, ECG beat classification, Medical
Diagnosis, bond rating prediction, etc. [19-23] The measured value however varies based on
the type of data, features extraction and selection methods, and the classification mechanism. A
summary of the % Best cost saving compared with the earlier methods to the proposed method
is provided in Table 3.
The cases correctly classified are considered zero cost, whereas equal costs have been
considered for both FN and FP.
Figure 3 and Figure 4 provide the CSDT training accuracy and the training loss Versus the
Number of Iteration for EMD-VQEM+ EMD-Stat feature vectors for convenience and
understanding.
862
Laxmipriya Samal, et al.
Table 5. Comparison of the Feature Extraction and the Classification Time of the Proposed
Techniques
Feature Classification Time (T2) Total Time (T1+T2)
Methods Extraction
Time (T1) DT CSDT DT CSDT
Statistical 4.18s 0.79s 0.92s 4.97s 5.1s
VQ 0.428s 0.24s 0.69s 0.668s 1.118s
EM 1.56s 0.76s 0.85s 2.32s 2.14s
VQEM 0.571s 0.19s 0.42s 0.761s 0.991s
ST-Stat 11.42s 0.87s 0.88s 12.29s 12.3s
ST-VQ 0.610s 0.56s 0.84s 1.17s 1.45s
ST-EM 1.58s 0.84s 0.91s 2.43s 2.49s
ST-VQEM 0.442s 0.29s 0.35s 0.732s 0.792s
EMD-Stat 12.54s 0.77s 0.79s 13.31s 12.43s
EMD-VQ 0.433s 0.54s 0.81s 0.973s 1.361s
EMD-EM 2.31s 0.75s 0.79s 3.06s 3.1s
EMD-VQEM 0.492s 0.21s 0.24s 0.702s 0.732s
ST-VQEM+ ST-Stat 11.89s 0.92s 0.94s 12.38s 12.39s
EMD-VQEM+ EMD-Stat 12.63s 0.80s 0.83s 13.40s 12.48s
863
Power Quality Disturbance Classification using Vector Quantized Eigen
Table 6. Impact of Feature Dimension on The Classification Accuracy, Highest %Best Cost
Savings, and the Computation Time for CSDT Classifier
Feature Vectors Dimension of a PQ Total Time for
Highest
Event Classification Feature
% Best cost
, Accuracy Extraction
Savings
and Classification
ST-Stat 85.1% 87.3% 12.3s
ST-EM 85.4% 87.9% 2.49s
ST-VQ 90.5% 93.2% 1.45s
ST-VQEM 92.7% 94.6% 0.792s
EMD-Stat 92.6% 94.8% 1.243s
EMD-EM 93.7% 95.2% 3.1s
EMD-VQ 97.1% 99.1% 1.243s
EMD-VQEM 97.7% 99.4% 0.732s
ST-VQEM+ ST-Stat 96.9% 95.9% 24.28s
EMD-VQEM+ EMD-
99.7% 99.7% 25.11s
Stat
Table 7 provides a comparison of the classification accuracy using the proposed techniques
with noisy data. The accuracy in Table 7 is slightly less than that of Table 6.
Table 8 provides a comparative table of the proposed work with that of the earlier work based
on accuracy.
864
Laxmipriya Samal, et al.
4. Conclusions
This article extracts a novel VQ Eigen Matrix (VQ-EM) from the PQ signals for improved
classification accuracy. The proposed technique has been compared with several feature
mapping techniques such as the statistical, ED, and VQ for the classification of thirteen-PQ
events. The VQ method of feature mapping remains more efficient than either the ED or the
statistical methods as observed in this work. While the statistical technique does not consider
the dynamics involved in a signal, the ED parameters are projected variables, and hence cannot
represent the signal in a true sense. On the contrary, the VQ method follows vector
dimensionality, correlation, non-linearity dependency, and Probability Density Function (PDF)
shape while clustering the features. It suits well when the feature set is closely matched or
densely clustered as revealed by these results. The two-state feature mapping method using
both the VQ and the ED has further improved the recognition accuracy. The proposed
combination of nonlinearity phenomena corresponding to VQ and linearity phenomena of ED
has a possible reason for improving the identification system modeling. A further improvement
in the recognition accuracy has been manifested when the proposed VQ-EM is considered
along with the statistical parameters due to an increase in available information as revealed
from this investigation. The proposed algorithms are giving superior classification accuracy
with both ideal and noisy environments including noise levels from 20dB to 40dB. If the
proposed system is embedded in real-time data, it can provide better accuracy as compared to
the state-of-art methods in both hardware and real-time environment. However, due to the
unavailability of adequate laboratory facilities, it could not be tested at this instant.
5. References
[1]. D. Burpee, H. Dabaghi, L. Jackson, F. Kwamena, J. Richter, T. Rusnov, and D. Meyer,
“US-Canada Power System Outage Task Force,” Final report on the implementation of
task force recommendations, 2006.
[2]. C. K. Hemapriya, M. V.Suganyadevi, and C. Krishnakumar, “Detection and classification
of multi-complex power quality events in a smart grid using Hilbert–Huang transform and
support vector machine,” Electr. Eng. Vol. 102, no. 3, pp. 1681-1706, Sep. 2020.
[3]. I. S.Samanta, P. K. Rout, and S.Mishra, “Feature extraction and power quality event
classification using Curvelet transform and optimized extreme learning machine,” Electr.
Eng. Vol. 103, no. 5, pp. 2431-2446, Oct. 2021.
[4]. M. K. Saini, and R. K.Beniwal, “Recognition of multiple PQ issues using modified EMD
and neural network classifier,”. Iran. J. Electr. Electron. Eng. vol. 14, no. 2, pp. 188-203,
Jun. 2018.
[5]. S., Samanta, P. K. Rout, and S. Mishra, “Power Quality Events Recognition Using S-
Transform and Wild Goat Optimization-Based Extreme Learning Machine,” Arab. J. Sci.
Eng. vol. 45, no. 3, pp. 1855-1870, Mar. 2020.
[6]. S. Dash, and U. Subudhi, “Multiple Power Quality Event Detection and Classification
using Modified S Transform and WOA tuned SVM Classifier,” Int. J. Electr. Power
Energy Syst. Vol. 12, no. 4, pp. 338-363, Aug. 2019.
[7]. M. Manjula, and A. V. Sarma, “Comparison of empirical mode decomposition and
wavelet-based classification of power quality events,” Energy Procedia, vol. 14. pp.
1156-1162, Jan. 2012.
[8]. P. R. Babu, P. K. Dash, S. K. Swain, and S. Sivanagaraju, “A new fast discrete
S‐transform and decision tree for the classification and monitoring of power quality
865
Power Quality Disturbance Classification using Vector Quantized Eigen
disturbance waveforms,” Int. Trans. Electr. Energy Syst. Vol. 24, no. 9, pp. 1279-1300,
Sep. 2014.
[9]. G. G. Pozzebon, G. G. Peña, A. F. Gonçalves, and R. Q.Machado, “Neural classification
of power quality disturbances: An application of the wavelet transform and principal
component analysis,”. In ninth IEEE/IAS Int. Conf. Industry Applications-INDUSCON,
pp. 1-6, Nov. 2010.
[10]. F. Hafiz, and C. Shahnaz, “A Machine Learning Based Classification Approach for
Power Quality Disturbances Exploiting Higher Order Statistics in the EMD
Domain,” Comp. Sc. 58922110, Aug. 2013.
[11]. J. C. Bravo-Rodríguez, F. J. Torres, and M.D. Borrás, “Hybrid machine learning models
for classifying power quality disturbances: A comparative study,” Energies, vol. 13, no.
11, p. 2761, Jan. 2020.
[12]. G. Liu, F. Li, G. Wen, S. Ning, and S. Zheng, “Classification of power quality
disturbances based on independent component analysis and support vector machine,”
IEEE Int. Conf. Wavelet Analysis and Pattern Recognition, pp. 115-123, Jul. 2013.
[13]. T. K. Abdel-Galil, E. F. El-Saadany, A. M. Youssef, M. M. Salama, “Disturbance
classification using hidden Markov models and vector quantization,” IEEE Trans. Power
Deliv. vol. 20, no. 3, pp. 2129-2135, Jun. 2005.
[14]. J. Chung, E. J. Powers, W. M. Grady, S. C. Bhatt, “Power disturbance classifier using a
rule-based method and wavelet packet-based hidden Markov model,” IEEE Trans. Power
Deliv. vol. 17, no. 1, pp. 233-241, Jan. 2002.
[15]. C. N. Bhende, S.Mishra, B. K. Panigrahi, “Detection and classification of power quality
disturbances using S-transform and modular neural network,” Electr. Power Syst.
Res. Vol. 78, no. 1, pp. 122-128, Jan. 2008.
[16]. N. R. Babu, and B. J.Mohan, “Fault classification in power systems using EMD and
SVM,” Ain Shams Eng. J. vol. 8, no. 2, pp. 103-111, Jun. 2017.
[17]. M. N. Mohanty, and H. K.Palo, “Child emotion recognition using pra probabilistic neural
network with effective features,” J. Int. Meas. Confed. vol. 152, p. 107369, Feb. 2020.
[18]. M., N. Mohanty, and H. K. Palo, “Segment based emotion recognition using combined
reduced features,”. Int. J. Speech Technol. Vol. 22, no. 4, pp. 865-884, Dec. 2019.
[19]. Bahnsen AC, Aouada D, Ottersten B. (2015) Example-dependent cost-sensitive decision
trees. Expert Systems with Applications, 42(19): 6609-6619.
[20]. S. B. Jabeur, A. Sadaaoui, A. Sghaier, and R. Aloui, “Machine learning models and cost-
sensitive decision trees for bond rating prediction,” J. Oper. Res. Soc. Vol. 71, no. 8, pp.
1161-1179, Aug. 2020.
[21]. Krawczyk B, Woźniak M, Schaefer G. Cost-sensitive decision tree ensembles for
effective imbalanced classification. Applied Soft Computing. 2014 Jan 1;14:554-62.
[22]. Li X, Zhao H, Zhu W. A cost sensitive decision tree algorithm with two adaptive
mechanisms. Knowledge-Based Systems. 2015 Nov 1;88:24-33.
866
Laxmipriya Samal, et al.
[23]. Sahin Y, Bulkan S, Duman E. A cost-sensitive decision tree approach for fraud detection.
Expert Systems with Applications. 2013 Nov 1;40(15):5916-23.
[24]. J. Demšar, “Statistical comparisons of classifiers over multiple data sets,” J. Mach. Learn.
Res. Vol. 7, pp. 1-30, Dec. 2006.
[25]. C. Y. Lee, and Y. X. Shen, “Optimal feature selection for power-quality disturbances
classification,” IEEE Trans. Power Deliv. vol. 26, no. 4, pp. 2342-2351, Jun. 2011.
Laxmipriya Samal has received her B.E. and M. Tech. degree in Electronics
and Communication Engineering from the BPUT, Odisha, and the Siksha ‘O’
Anusandhan Deemed to be University, Odisha in 2003 and 2009 respectively.
She is currently working towards a Ph.D. degree at the Siksha ‘O’
Anusandhan Deemed to be University, Odisha, India. Her research interests
include Power quality disturbance and applied signal processing.
Hemanta Kumar Palo completed his Master of Engineering from “The Birla
Institute of Technology”, Mesra, Ranchi in 2011 and his Ph.D. in 2018 from
the Siksha ‘O’ Anusandhan Deemed to be University. He is having 20 years of
experience in the field of Electronics and Communication Engineering.
Currently, he is working as an Assoc. Prof. in the Dept. of ECE, SOA
University, Bhubaneswar, Odisha, India. He has published around 70 papers in
international journals and conferences. His research interests include signal
processing, speech processing, and power electronics.
Badrinarayan Sahu completed his M-Tech in the year 2004 from UCE Burla,
Sambalpur, and his Ph.D. from Siksha ‘O’ Anusandhan University. Currently,
he is serving as a Professor in the Department of Electronics and
Communication Engineering at Siksha ‘O’ Anusandhan University,
Bhubaneswar, Odisha, India. His areas of interest are signal processing, image
processing, and power quality recognition.
867