Available online at www.sciencedirect.

com

ScienceDirect
Available online at www.sciencedirect.com
Procedia Computer Science 00 (2017) 000–000
www.elsevier.com/locate/procedia
ScienceDirect
Procedia Computer Science 114 (2017) 266–274

Complex Adaptive Systems Conference with Theme: Engineering Cyber Physical Systems, CAS
October 30 – November 1, 2017, Chicago, Illinois, USA

Crack Detection and Identification Using Vibration Signals and

Fuzzy Clustering
Issam Abu-Mahfouz*, Amit Banerjee
Pennsylvania State University Harrisburg, 77 West Harrisburg Pike, Middletown, PA 17055, USA

Abstract

Cracks in machine and structural components are among the most critical faults that require immediate attention by operations
and maintenance personnel. In many cases it is difficult to directly and accurately identify the fault type and its extent under
varying operating conditions. Therefore the ability to monitor and diagnose machine and structures health at early stages of fault
initiation is one of the fundamental requirements of the development of online machine monitoring systems. This work
demonstrates a novel procedure for crack detection and identification in an experimental set-up. Cracks of different lengths are
introduced at three different locations along a cantilevered steel beam. The beam is exposed to several forms of excitation at
different levels of amplitude and frequency. The excitation is introduced to the beam next to its fixed end using an
electromagnetic shaker controlled by an arbitrary function generator. The vibration signals are analyzed for the beam dynamic
signatures in the form of statistical moments, frequency spectra and wavelet coefficients. These are then used as features to create
a data-partition using fuzzy relational clustering. The objective is to correlate crack length and crack propagation to the statistical
moments, frequency spectra and wavelet coefficient data. The work will be later extended to supervise learning by training a
classifier in a two-class problem to identify signal features that correlate to critical faults

© 2017 The Authors. Published by Elsevier B.V.

Peer-review under responsibility of the scientific committee of the Complex Adaptive Systems Conference with Theme:
Engineering Cyber Physical Systems.

Keywords: crack detection; nondestructive testing; machine learning; fuzzy relational clustering; vibration; feature selection; pattern recognition

* Corresponding author. Tel.: +1-717-948-6361; fax: +1-717-948-6619.

E-mail address: iaa2@psu.edu

1877-0509 © 2017 The Authors. Published by Elsevier B.V.

Peer-review under responsibility of the scientific committee of the Complex Adaptive Systems Conference with Theme:
Engineering Cyber Physical Systems.

1877-0509 © 2017 The Authors. Published by Elsevier B.V.

Peer-review under responsibility of the scientific committee of the Complex Adaptive Systems Conference with Theme:
­Engineering Cyber Physical Systems.
10.1016/j.procs.2017.09.038
 Issam Abu-Mahfouz et al. / Procedia Computer Science 114 (2017) 266–274 267
Abu-Mahfouz / Procedia Computer Science 00 (2017) 000–000

1. Introduction

Cracks in machine component and structures are among the most encountered fault types. It is difficult to identify
these types of faults especially during their initiation stage using direct inspection procedures. The ability to monitor
and diagnose health of machines and structures at early stages of fault initiation is one of the fundamental
requirements of the development of monitoring systems. The machine or structural monitoring systems should be
able to accurately identify the fault type under varying operating or loading conditions.
The existence of a crack reduces the stiffness of the structure [1, 2] and therefore changes its natural frequencies.
This makes it possible to investigate the changes in the frequency spectrum features of the structural vibration
response in order to predict the location of crack damage. Many researchers have utilized vibration behavior of
structures to develop intelligent crack detection methods. A kurtosis-based prediction scheme for crack location and
size detection in Plexiglas cantilever beam based on the fundamental vibration mode has been proposed in Ref [3].
A hybrid neuro-genetic algorithm was proposed by Sahoo and Maity to detect damage in a cantilever beam [4]. A
method based on genetic algorithm (GA) was also used to monitor changes in the natural frequencies of a structure
in Ref [5]. A modified principle component analysis (PCA) method has been applied to selected frequency bands of
the frequency response function (FRF) in order to generate damage features for pattern recognition [6]. The location
and depth of the crack in a cantilever beam were determined using a particle swarm optimization (PSO) algorithm in
Ref [7]. A fuzzy adaptive PSO (APSO) for predicting both the size and location of the crack on a beam using
numerical and experimental studies was proposed in Ref [8]. Frequencies measured using single-input multi-output
(SIMO) based experimental modal analysis (EMA) of the clamped-clamped beam with crack were utilized to
identify crack location and its severity by the frequency-based approach as well as using a GA [9].
The objective of the present paper is to investigate the application of the fuzzy clustering method using
experimentally obtained vibration response data to predict the location of cracks in a cantilever beam. Signal
processing techniques such as Fast Hartley Transform (FHT) and Fast Fourier Transform (FFT) have been used as
features to for edge-detection using clustering with applications in infrastructure monitoring are presented in Ref
[10]. A principal component based unsupervised clustering algorithm for bridge crack detection is described in Ref
[11]. We extend the unsupervised clustering methodology to use features from signal processing of vibration signals
to identify whether a beam has a crack and if so discriminate between patterns based on distance of the crack from
the fixed end of the beam. In this paper, the fuzzy relative clustering (FRC) in the relational feature space is used.

2. Experimental Setup

Steel beams measuring 0.0625 in x 1 in x 16 in are used for this work. A TMS model 2007E electrodynamic
shaker is used to excite the beams at 3 inches away from the fixed end. Figure 1 shows a schematic of the
experimental setup. Seven excitation frequencies are chosen at 20Hz, 40Hz, 60Hz, 68Hz, 70Hz, 80 Hz, and 100Hz.
Beams with three different crack locations {5 in. (inner), 8 in. (central) and 11 in. (outer)} measured from the fixed
end are prepared. Each beam has only a single crack. This resulted in 28 different combinations of the data sets
including the beam with no crack. Vibration signals are collected using a Kistler accelerometer (model 8622, with
sensitivity of 9.34 mV/g and 500 g measuring range) mounted on the beam near its fixed end. All data are recorded
via a National Instruments USB-6211 data acquisition device at a sampling rate of 1000s/s.

Data Signal processing and

Accelerometer fuzzy clustering
acquisition
Beam

Amplifier Shaker

Signal
Figure 1. Experimental set-up
268 Issam Abu-Mahfouz et al. / Procedia Computer Science 114 (2017) 266–274
Abu-Mahfouz / Procedia Computer Science 00 (2017) 000–000

Figure 2 shows sample accelerometer signals for all four crack cases at 60 Hz excitation frequency. To reduce the
dimensionality of the feature space, raw vibration time histories are processed using several techniques, such as Fast
Fourier Transform (FFT) and continuous Wavelet analysis (CWT). The results of FFT signatures are averaged into
16 equal bands (each band is of 128Hz width over a total range of the first 2048Hz.) Figure 3 shows a sample bar
chart for the 16 FFT averaged features at three different inner crack depths. The Continuous Wavelet Transform
(CWT) [12] is used to extract more features from the vibration signals via the scalogram plots and the
approximations and details decomposition analysis. In this investigation, the Coiflet (coif4) and Mexican hat (mexh)
MATLAB functions are applied to a segment of 1000 data points for each beam crack case. Figure 4 shows
scalogram images for the four crack conditions at 60Hz excitation. Sample root-mean-square (rms) results of the
first ten approximations and details of the wavelet decomposition using the Coiflet wavelet are shown in Figure 5. In
addition, results are obtained for the kurtosis and skewness of the first 10 approximations and details. Kurtosis is the
fourth moment of the probability density curve about the mean and is calculated as

1 N
( xi  x ) 4
K
N
 i 1 4
(1)

where x is the signal mean and  2 is the variance. Skewness is a measure of symmetry (or asymmetry) of the data
distribution about the sample mean. The skewness of the normal distribution is zero. The skewness of a
distribution is defined as

1 N
( xi  x )3
S
N
 i 1 3
(2)

0.1 0.1
Crack location 5 inch at 60 Hz Crack location 8 inch at 60 Hz

0.05 0.05

0 0

-0.05 -0.05

-0.1 -0.1
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

0.1 0.1
Crack location 11 inch at 60 Hz No crack at 60 Hz

0.05 0.05

0 0

-0.05 -0.05

-0.1 -0.1
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

Figure 2. Time series signals for different crack locations run at 60 Hz excitation frequency.

The vibration feature sets used in this study along with their sizes (N) are: (1) average FFT spectrum (AFFT), Nfft =
16; (2) statistics of vibration time signal (mean, skewness, standard deviation, kurtosis, and variance) Nstcs = 5; (3)
average of the first 64 CWT coefficients, NCWT = 64; (4) kurtosis and skewness of 10 wavelet approximations and 10
wavelet details (KS), Nks = 40; (5) Root-mean-square (rms) of the 10 approximations and 10 details, NKSrms = 20.
 Issam Abu-Mahfouz et al. / Procedia Computer Science 114 (2017) 266–274 269
Abu-Mahfouz / Procedia Computer Science 00 (2017) 000–000

(a) (b)

Figure 3. Bar chart of the averaged FFT spectral bands for (a) 40 Hz and (b) 100 Hz.

61
57 No crack at 60 Hz
53 No crack
49
45
41
37
Scales a

33
29
25
21
17
13
9
5
1
0 50 100 150 200 250 300 350 400 450 500
61
57
53 Crack at 5”
49
45
41
37
Scales a

33
29
25
21
17
13
9
5
1
0 50 100 150 200 250 300 350 400 450 500
61
57
53 Crack at 8”
49
45
41
37
Scales a

33
29
25
21
17
13
9
5
1
0 50 100 150 200 250 300 350 400 450 500

61
57
53 Crack at 10”
49
45
41
37
Scales a

33
29
25
21
17
13
9
5
1
0 50 100 150 200 250 300 350 400 450 500
Time (or Space) b

Figure 4. CWT scalogram images for beams with different crack location at 60 Hz excitation
270 Issam Abu-Mahfouz
Abu-Mahfouz et al. /Computer
/ Procedia Procedia Science
Computer00 Science 114 (2017) 266–274
(2017) 000–000

0.05
No crack
0.04 5 inch
8 inch
0.03 11 chin

0.02

0.01

0
1 2 3 4 5 6 7 8 9 10
0.04
No crack
0.035
5 inch
0.03
8 inch
0.025
11 chin
0.02
0.015
0.01
0.005
0
1 2 3 4 5 6 7 8 9 10
Figure 5. Sample plots of the RMS values for the approximation (top) and details (bottom) for four crack locations at 80 Hz excitation.

3. Feature Extraction for Clustering

Clustering based on raw data implies working in a high dimensional space, especially for time series data
collected in our study at fast sampling rates. Several feature based clustering techniques for time series data has
been proposed. In a clustering study on fMRI time series data, two clustering algorithms: k-means and Wards
Agglomerative clustering were used on a cross-correlation feature set derived from the raw data [13]. In another
study, the delay and strength of activation were used a features to identify regions with significantly different
delays and activations in fMRI data [14]. Fu et al. [15] described the use of self-organizing maps for grouping data
sequences segmented from the numerical time series using a continuous sliding window with the aim to discover
similar temporal patterns along the time series. An algorithm called the Sequence Cluster Refinement Algorithm
(SCRA) is presented in [16] to group machine tool monitoring data into clusters represented as discrete hidden
Markov models (HMM). An incremental clustering approach for time series data using Haar wavelet transforms is
presented in [17]. The Haar Wavelet time decomposition is computed for every time series and the coefficients at
various resolutions are used as features. The cluster center produced at one level of resolution is used as the initial
center for the next level of resolution. In this section, a brief review of some high level representations of time
series, which can then be processed to be used as features for clustering are presented. The following feature sets
are used for clustering:

(A) Statistical features: each time series is represented by 5 features - mean, skewness, standard deviation, Kurtosis
and variance. (k = 5)
(B) PSD quantum: each time series is represented by 16 PSD coefficients. (k = 16)
(C) Kurtosis approximations: each time series is represented by Kurtosis of 10 approximations. (k = 10)
(D) Kurtosis details: each time series is represented by Kurtosis of 10 details. (k = 10)
(E) RMS approximations: each time series is represented by root mean square of 10 approximations. (k = 10)
 Issam Abu-Mahfouz et al. / Procedia Computer Science 114 (2017) 266–274 271
Abu-Mahfouz / Procedia Computer Science 00 (2017) 000–000

(F) RMS details: each time series is represented by root mean square of 10 details. (k = 10)

In this investigation, the problem of clustering is posed in the relational space by converting the feature sets into
relational data in the feature space. The Euclidean distance is used as a measure of dissimilarity between two
patterns in the feature space. The feature sets are first normalized using feature scaling so that all features are in the
range (0, 1). Also, city block distance measure was also considered instead of the normalized Euclidean distance,
but the performance of the algorithm was not significantly better than with Euclidean distance (this is not
investigated or presented further in this paper). For n patterns, the relational data R is an n x n reflexive matrix, so
only the lower (or upper) triangular matrix is required for computation purposes i.e.
Rij  0, Rii 0, Rij R ji (1  i, j  n). Fuzzy relational clustering (FRC) is used as the clustering algorithm [18]. FRC
relaxes the restriction by other fuzzy clustering algorithms on the relational data that they be based on Euclidean
distances. Although, in this paper, we use Euclidean distances as the dissimilarity measure, FRC allows us to change
the distance measure without any adverse effect on the procedure. The minimization function for FRC is,

n n

c  u m m
u R jk
ik ij c
J FRC   subject to the constraints  uik  1, 1  k  n ; uik  0, 1  i  c,1  k  n
j 1 
k 1
n
(3)
i 1
2 u m
it
i 1

t 1

The membership matrix U is a c  n matrix where uik is the membership of the kth pattern in the ith cluster. The
parameter m is called the fuzzifier or weighting exponent. It determines the fuzziness of the classification – with
higher values of m the boundaries between the clusters become softer, with lower values they get harder. For m = 1,
fuzzy clustering and hard clustering are the same. Usually m is chosen to be 2.0 but in this paper we study the effects
of m on the quality of clusters generated for the vibration signal-based feature sets. The memberships can be
obtained using,

1/ ( m 1)
 1  n n n
  m uitm Rtk m uitm uihm Rht
a
uik   ik  1/ ( m 1) 
where aik t 1
n
 h 1 t 1n (4)
c    
 
1
  m
uit 2   uitm 
j 1  a jk 
t 1  t 1 

The algorithm is initialized using an arbitrary membership matrix and the memberships are iteratively refined. A
successive substitution scheme is used to calculate the memberships within iteration, in which old values of uik to
obtain aik and then solve for new values of uik in the same step. This is repeated the membership matrix is
completely refreshed and this is continued until convergence. The algorithm is presented below,

Fix c, 2 ≤ c ≤ n, and m > 1, set iter = 0; fix MAX_ITER

Initialize fuzzy c-memberships U  uik , 1 ≤ i ≤ c, 1 ≤ k ≤ n;
do
for k = 1 to n
for i = 1 to c
Compute aik using eq. (4) using new memberships uij for j < k and old memberships uij for j ≥ k;
Update memberships uik using eq. (4)
end
end
iter = iter + 1;
until (|Uold – U| < ε or iter = MAX_ITER)

The preprocessing of the vibration signal features is therefore done independently of the clustering algorithm. For
272 Issam Abu-Mahfouz et al. / Procedia Computer Science 114 (2017) 266–274
Abu-Mahfouz / Procedia Computer Science 00 (2017) 000–000

every feature set, the input to FRC is the relational data R. The output of FRC is the membership matrix U which is
then decoded as k  i if uik  u jk ,1  i, j  c,1  k  n, j  i .

4. Results

The six feature sets (A) Statistical features, (B) PSD quantum features, (C) Kurtosis approximations, (D) Kurtosis
details, (E) RMS approximations, (F) RMS details, are first converted to relational data R of size n = 28. Euclidean
distance is used as a distance metric in the calculation of R. FRC was implemented for two distinct groupings, a 2-
cluster partition c = 2 and a 4-cluster partition c = 4, with three different values of the fuzzifier m = 1.0, 2.0 and
10.0; termination condition of ε = 0.001 and MAX_ITER = 100. In both cases (2-cluster and 4-cluster), the objective
is to compare the clusters generated by FRC to the two (or four) classes corresponding to crack/no crack (or distance
of crack) obtained experimentally. In this study the correlation is established using ROC analyses with sensitivity
and specificity used as measures. Results presented here are averaged over four independent runs of FRC with
membership matrix U randomly initialized for each run.
In the first set of results for c = 2, the two classes in the 2-cluster partition will be labeled “crack” and “no crack”.
The sensitivity of a 2-cluster partition is 1.0 if all the instances belonging to the class labeled “crack” have been
correctly identified (true positive rate of 1.0). There are 21 instances belonging to the class labeled “crack” and 7
instances belonging to the class labeled “no crack”. The specificity on the other hand is 1.0 if all instances that do
not belong to the class labeled “crack” have been correctly identified as not belonging in the same cluster (true
negative rate of 1.0). Across the board, FRC performs better for m = 2.0 with all feature sets followed by m = 10.0
and 1.0. For m = 1.0 FRC is the same as k-means based relational clustering. Among the feature sets, (C) performed
the best based on results from ROC analysis, followed by (E) and then (B). Feature set (A) produced clusters with
relatively high specificity but not sensitivity.

TABLE 1: SENSITIVITY AND SPECIFICITY OF FRC FOR 2-CLUSTER “CRACK” CLASS

FEATURE m = 1.0 m = 2.0 m = 10.0
SET Sensitivity Specificity Sensitivity Specificity Sensitivity Specificity
(A) 0.4286 0.5714 0.6667 0.7143 0.4762 0.7143
(B) 0.7619 0.7143 0.8095 0.7143 0.7143 0.8571
(C) 0.7619 0.8571 0.9048 1.0000 0.8095 0.5714
(D) 0.4762 0.2857 0.5714 0.8571 0.5238 0.4286
(E) 0.6667 0.8571 0.8571 0.8571 0.7619 0.7143
(F) 0.4762 0.5714 0.7143 0.7143 0.4762 0.4286

For the 4-cluster case, the four classes are “11-inch crack”, “8-inch crack”, “5-inch crack” and “no-crack”. All
classes have 7 instances. Again, FRC performed better for m = 2.0 with all feature sets followed by m = 10.0 and
1.0. However, among the feature sets, (E) obtained the best discrimination, followed by (B) and then (C). In both the
2-cluster and 4-cluster cases, feature sets (A), (D) and (F) did not obtain better discrimination when compared to
(B), (C) and (E).

TABLE 2: SENSITIVITY AND SPECIFICITY OF FRC FOR 4-CLUSTER “NO-CRACK” CLASS

FEATURE m = 1.0 m = 2.0 m = 10.0
SET Sensitivity Specificity Sensitivity Specificity Sensitivity Specificity
(A) 0.2857 0.7143 0.5714 0.8571 0.0000 0.5714
(B) 0.8571 0.8571 1.0000 0.8571 0.5714 1.0000
(C) 0.7143 0.8571 0.8571 0.7143 0.4286 0.7143
(D) 0.5714 0.5714 0.5714 0.7143 0.2857 0.5714
(E) 0.8571 1.0000 1.0000 1.0000 0.5714 0.8571
(F) 0.5714 0.5714 0.5714 0.5714 0.2857 0.5714
 Issam Abu-Mahfouz et al. / Procedia Computer Science 114 (2017) 266–274 273
Abu-Mahfouz / Procedia Computer Science 00 (2017) 000–000

The choice of m = 2.0 is better than m = 1.0 or m = 10.0. This is in accordance to the analysis presented in [19]
where it is shown that a higher value of m (such as 4.0) works well for very noisy data while the range [1.5,2.5] is
recommended for regular (smaller) datasets such as the one used here.

A possible reason why the kurtosis detail or the RMS detail feature set did not yield partitioning results compared
to the kurtosis approximations or the RMS approximations might be that the Euclidean distance measure is not
enough to capture the relative spread of signed features (for the sake of completeness, it should be mentioned that
relational data obtained using the city block distance measure did not yield significantly better results in this context
either).

5. Conclusion

This study presents a damage detection methodology that is based solely on the analysis of the vibration signals
of the harmonically excited structure at different frequencies. The vibration signals are analyzed to extract relevant
features to design a crack damage fuzzy clustering model for a steel cantilevered beam. The analyses are performed
using FFT averaged bands and statistical quantities applied to the time signals and wavelet approximations and
details. The obtained features show good sensitivity of the vibration signatures to the crack location. This is verified
by fuzzy clustering that clustered these vibration time series features into self-similar groups based on relational data
obtained from these features. The partitions provided by fuzzy clustering indicate that partitions obtained in the
signal feature space relate to either the existence of crack (or no crack) and to some extent are even able to
discriminate based on the distance of crack from the fixed end of the beam. The implication is that vibration signals
can be used as a predictive model to forecast the development and propagation of cracks. Since vibrations signals
can be processed during operation of machinery, a real time system can be developed to indirectly monitor the
development of cracks, and make necessary adjustments while the process in running. This relates to machine health
and process condition monitoring and to the best of our knowledge the use of vibration signal-based features in
online process condition monitoring has not been reported in literature.
However, one of the limitations of this work is the small sample size used for clustering study (28 instances) and
the subsequent ROC analyses. However, this investigation is the first step towards developing a prediction model
using FFT spectrum coefficients. The prediction model can be based on a classifier or a nonlinear regression model
to train an artificial neural network. Development of an online implementation of the prediction model is also
planned.

References

[1] Chondros TG, Dimarogonas AD. Identification of cracks in welded joints of complex structures. Journal of Sound and Vibration 1980;
69:531–8.
[2] Chondoros TG, Dimargonas AD. A continuous cracked beam vibration theory. Journal of Sound and Vibration 1988; 215(1):17–34.
[3] Hadjileontiadis LJ, Douka E, Trochidis A. Crack detection in beams using kurtosis. Computers and Structures 2005; 83:909–919.
[4] Sahoo B, Maity D. Damage assessment of structures using hybrid neurogenetic algorithm. Applied Soft Computing 2007; 7:89–104.
[5] Vakil-Baghmisheh M-T, Peimani M, Homayoun Sadeghi M, Ettefagh MM. Crack detection in beam-like structures using genetic algorithms.
Applied Soft Computing 2008; 8:1150–60.
[6] Trendafilovaa I, Cartmellb MP, Ostachowicz W. Vibration-based damage detection in an aircraft wing scaled model using principal
component analysis and pattern recognition. Journal of Sound and Vibration 2008; 313:560–566.
[7] Vakil Baghmisheh MT, Peimani M, Sadeghi MH, Ettefagh MM, Tabrizi AF. A hybrid particle swarm–Nelder–Mead optimization method for
crack detection in cantilever beams. Applied Soft Computing 2012; 12(8):2217–26.
[8] Jena PK, Thatoi DN, Parhi DR. Dynamically self-adaptive fuzzy PSO technique for the smart diagnosis of transverse crack. Applied Artificial
Intelligence 2015; 29:211–232.
[9] Munglaa MJ, Sharmab DS, Trivedi RR. Identification of a crack in clamped-clamped beam using frequency-based method and genetic
algorithm. Procedia Engineering 2016; 144: 1426–1434.
[10] Ikhlas A-Q, Abudayyeh O, Kelly ME. Analysis of edge-detection techniques for crack identification in bridges. Journal of Computing in
Civil Engineering 2003; 17: 255-263.
[11] Ikhlas A-Q, Pashaie-Rad S, Abudayyeh O, Yehia S. PCA-based algorithm for unsupervised bridge crack detection. Advances in Engineering
274 Issam Abu-Mahfouz et al. / Procedia Computer Science 114 (2017) 266–274
Abu-Mahfouz / Procedia Computer Science 00 (2017) 000–000

Software 2006; 37: 771-778.

[12] Blatter C. Wavelets: A Primer. Natick, MA: A. K. Peters Ltd., 1998.
[13] Goutte C, Toft P, Rostrup E. On clustering fMRI time series. Neuroimage 1999; 9: 298-310.
[14] Gouette C, Hansen LK, Liptot MG, Rostrup E. Feature-space clustering for fMRI meta-analysis. Human Brain Mapping 2001; 13: 165-183.
[15] Fu T-C, Chung F-L, Ng V, Luk R. Pattern discovery from stock time series using self-organizing maps. KDD 2001 Workshop on Temporal
Data Mining, 27-37; San Francisco, CA, 2001.
[16] Owsley LMD, Atlas LE, Bernard GD. Self-organizing feature maps and hidden Markov models for machine-tool monitoring. IEEE
Transactions on Signal Processing 1997; 45: 2787-2798.
[17] Vlachos M, Lin J, Keogh E, Gunopulos D. A wavelet-based anytime algorithm for k-means clustering of time series. In Proc. 3rd SIAM
International Conference on Data Mining, San Francisco, CA, 2003.
[18] Dave RN, Sen S. Robust fuzzy clustering of relational data. IEEE Transactions on Fuzzy Systems 2002; 10: 713-727.
[19] Wu, K-L. Analysis of parameter selections for fuzzy c-means. Pattern Recognition 2012; 45: 407-415.