Measurement 55 (2014) 1524

Contents lists available at ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Fault diagnosis on material handling system using feature

selection and data mining techniques
M. Demetgul a,, K. Yildiz b, S. Taskin c, I.N. Tansel d, O. Yazicioglu e
a

Department of Mechatronics Engineering, Technology Faculty, Marmara University, Turkey

Department of Computer and Control Education, Technical Education Faculty, Marmara University, Turkey
c
Department of Electrical and Electronics Engineering, Engineering Faculty, Celal Bayar University, Turkey
d
Department of Materials and Mechanical Engineering, Engineering Faculty, Florida International University, FL, USA
e
Department of Industrial Engineering, Design and Engineering Faculty, Istanbul Commerce University, Turkey
b

a r t i c l e

i n f o

Article history:
Received 4 February 2014
Received in revised form 21 April 2014
Accepted 23 April 2014
Available online 9 May 2014
Keywords:
Servo-pneumatic
Material handling system
Fault diagnosis
Feature selection
Data mining
Dimension reduction
GustafsonKessel
k-Medoids

a b s t r a c t
The material handling systems are one of the key components of the most modern manufacturing systems. The sensory signals of material handling systems are nonlinear and have
unique characteristics. It is very difcult to encode and classify these signals by using multipurpose methods. In this study, performances of multiple generic methods were studied
for the diagnostic of the pneumatic systems of the material handling systems. Diffusion
Map (DM), Local Linear Embedding (LLE) and AutoEncoder (AE) algorithms were used for
future extraction. Encoded signals were classied by using the GustafsonKessel (GK)
and k-medoids algorithms. The accuracy of the estimations was better than 90% when
the LLE was used with GK algorithm.
2014 Elsevier Ltd. All rights reserved.

1. Introduction
The modern manufacturing facilities have to detect the
problems, identify their sources and x them very quickly
with very limited man power. Researchers have started
development of computational diagnostic tools for the
industrial applications in early 1970s by considering this
need. Although, various diagnostic tools have been developed by research community and successfully used in
industrial applications in last two decades [1,2] still their
capabilities are limited. In this study, feasibility of a multipurpose fault detection approach was investigated. The
proposed approach used the combinations of the generic

Corresponding author. Tel.: +90 216 336 57 70.

E-mail address: mdemetgul@marmara.edu.tr (M. Demetgul).
http://dx.doi.org/10.1016/j.measurement.2014.04.037
0263-2241/ 2014 Elsevier Ltd. All rights reserved.

dimension reduction methods for feature extraction and

classied the encoded data with clustering algorithms.
One of the key components of the automated manufacturing is material handling systems. Pneumatic and
hydraulic systems are widely used for material handling.
These systems may have hundreds of actuators and sensors. Identication of faulty components and their locations in a very short time is very difcult. Several studies
were performed for development of fault diagnostic tools
for these systems in the last decade [3]. The studies mainly
aimed evaluation of the condition of the cylinders [4] and
digitally controlled valves [5]. Other studies focused on
detection of leakage of the seals [69], friction increase
[4,10] and malfunctions [1114].
Most of the fault diagnostic tools have two components: feature extractor (encoder) and classier. Some
researchers have used the intelligent data analysis

16

M. Demetgul et al. / Measurement 55 (2014) 1524

techniques for fault diagnostic [1517]. Support vector

machines [18], self-organizing feature maps (SOM) [19],
expert systems, neural networks, rough sets and fuzzy
logic have been used for classication of data. The computing complexity of feature extraction and learning process
have been the main disadvantages of these approaches.
The data of the material handling systems for the fault
diagnosis comes from multiple sensors. The data is highdimensional and nonlinear. While the large number of data
from different sensors provide more information, at the
same time feature extraction and classication becomes
more complex. The dimension reduction methods compress the data automatically, reduce the noise, may extract
features for fault diagnostic and minimize required storage.
Clustering algorithms have been used for classication.
Fuzzy c-means (FCM) [20] and its variants GustafsonKessel (GK) [21] algorithm are popular pattern classication
methods. They have been used for fault detection and isolation [2224], k-medoids [25] is a partitional clustering
algorithm and may be used for classication purposes.
DM method [2629] used diffusion semigroups for learning the global characteristics of the data-set. The complex
structures were represented at different scales by the help
of these semi groups. The eigenfunctions of Markov matrices were effectively used with this purpose. LLE [30]
method used an unsupervised learning algorithm to
change the dimensions of the data. The algorithm found
the neighbors in X space, calculated weights for reconstruction and calculated the embedding coordinates in Y
space by using the calculated weights. AutoEncoder (AE)
[31] methods use an articial neural network (ANN) to
learn the compact representation of data set. The dimensionality of the data set is reduced by using this ANN. Various multilayer architectures [32,33] and optimization
methods [31,34] have been proposed to improve the performance of the ANN.
GK and k-medoids algorithms were used to create the
desired number of clusters to partition or classify the data
after it was compacted. GK algorithm [21] calculates the
center and covariance matrix to represent the clusters
[35,36]. They are used during the optimization process.
This approach allows identication of ellipsoidal clusters
and improves the performance of the method relative to
other approaches. k-medoids [37] is another clustering
algorithm. The algorithm divides the data into the groups,
chooses the data points as medoids. In this study, Clustering and Data Analysis Toolbox [38] was used for classication of the compressed data.
Diffusion map, AutoEncoder, and Local Lineer Embedding techniques were used for dimension reduction process respectively. In classication process two algorithms
were used namely; k-medoids and GK. These algorithms
were given in detailed below.
2. Classication and feature extraction
2.1. k-Medoids
It is a standard clustering algorithm [37] where the
update rule always moves the cluster center to the nearest
data point in the cluster. k-Medoids is a partitioning

technique of clustering that clusters the data set of n

objects into k clusters with k known a priori. t could be
more robust to noise and outliers as compared to k-means
because it minimizes a sum of general pairwise dissimilarities instead of a sum of squared Euclidean distances.
2.2. GustafsonKessel algorithm (GK)
GustafsonKessel algorithm (GK), providing a degree of
membership of each data point to a particular cluster creates a fuzzy partition [21]. One set of data to detect clusters
of different geometrical shapes, this method is an adaptive
distance norm for each cluster was introduced. Each cluster has its own standard distance norm affects the Ai
matrix Mi, inducing which statement is the following
equation [36].
T

dxj ; Li M IIxj  LI II xj  Li Mi xj  Li

possible use of the Z matrix of each cluster, every step of

the data t the geometric structure of the distance norm.
Based on the norm-inducing matrices, the aim of the GK
method, is obtained by minimizing the function J as Eq. (2).

JP; L; M

c X
n
X
0
2 2
Ai xj d xj ; Li M

i1 j1

M = (M1, M,   , Mc) is a c-tuple, where the norm-inducing matrices.

2.3. Diffusion maps
Diffusion maps as a system of eigenfunctions of Markov
matrices consider effective representation of data geometric descriptions of the original data set to obtain the coordinates [2729]. A given data set X = (x, i = 1,   , N) is a ddimensional data space, said N nodes can be built over X is
a nite graph corresponding to the N data. Usually, the
form of a Gaussian kernel as in Eq. (3)

wxi ; xj exp

IIxi  xj II2

2a2

where r is the kernel width parameter, is used to construct

the similarity matrix. The kernel reects the degree of similarity between xi and xj, and II, II is the Euclidean norm in
Rd [39].
2.4. AutoEncoder
Multilayer encoders are hidden layer feed forward neural networks with an odd number [32,33]. The input and
the output layer have D nodes and the middle hidden layer
has d nodes.
AutoEncoders usually high number of multi-layer connection. Therefore, the back propagation approaches is
likely to get stuck in slow convergence and local minimum.
In [32] this disadvantage is overcome by performing a
Restricted Boltzmann Machines using pretraining (RBMs)
[34]. The mean square error between the input and output
of the network is trained to minimize (Ideally, the input
and output equal). Linear activation functions in the use
of neural network, PCA is a very similar AutoEncoder [40].

17

M. Demetgul et al. / Measurement 55 (2014) 1524

Table 1
Faults in the system and output values of ANN algorithm.
Faults
Normal operation (System operation
pressure 6 bar)
Low pressure (LP) (System operation
pressure 4 bar)
Low Low pressure (LLP) (System
operation pressure 2 bar)
x Axis error positioning for back motion
x Axis error positioning for forward
motion
y Axis error positioning for up motion
y Axis error positioning for down motion
Pick faults for gripper (one pneumatic
hose release)
Place faults for gripper (other pneumatic
hose release)
Gripper close fault (digital magnetic
proximity sensor which send gripper
close information cable disconnect
fault)
After gripper receive material to drop
failure.

Experiments

Symbol

Normal

Fault1

Fault2

4
5

Fault3
Fault4

6
7
8

Fault5
Fault6
Fault7

Fault8

10

Fault9

11

Fault10

2.5. Local Linear Embedding (LLE)

Local Linear Embedding (LLE) [30] is a local technique
for dimensionality reduction that is similar to Isomap in
that it constructs a graph representation of the datapoints.
In contrast to Isomap, it attempts to preserve solely local
properties of the data, making LLE less sensitive to shortcircuiting than Isomap. Furthermore, the preservation of
local properties allows for successful embedding of nonconvex manifolds. In LLE, the local properties of the data
manifold are constructed by writing the datapoints as a
linear combination of their nearest neighbors. In the lowdimensional representation of the data, LLE attempts to
retain the reconstruction weights in the linear combinations as well as possible.

different to pile the parts next to each other. The LabVIEW

user interface is prepared. Four analog sensors showed the
pressures of the system at the compressor output, and cylinders. One cylinder opened and closed the gripper while
the other two moved it along the X and Y axes. Data was
collected with 50 ms sampling interval. Each experiment
took 27 s.

3. Experimental set-up

4. Experimental procedure and data collection

In this study, servo-pneumatic positioning experimental set-up built by the Festo Didactic Company was used.
The experimental setup is presented in Fig. 3. The grippers
motions along the X and Y axes were controlled with pneumatic dual action rodless cylinders. For the position measurements in the X and Y axes a linear potentiometer and
a contactless absolute magnetostrictive linear displacement sensor were used respectively. The pneumatic gripper of the system was installed at the Y axis actuator.
The proportional directional control valves controlled
the air ow to the cylinders. Four analog transducers measured the pressures of the entire system and the actuators
controlling the motions and the gripper. Experimental data
were collected by using the National Instrument (NI) compact Data Acquisition (NI-cDAQ) system and control modules. LabVIEW program was used for management and
analysis of the collected data.
The trainer was programmed to pick up the parts from a
conveyor and to put on a pallet in a systematic fashion. So,
the programmed motion was very similar but slightly

In this study, data was collected while the pneumatic

system was operated at the normal and additional 10 different faulty conditions. The imposed problems are listed
in Table 1. During the experiments, the data were collected
from 4 analog and 2 digital sensors for a period of 27 s. The
data was collected 3 times at each experimental condition.
The pneumatic systems main pressure; the x and y axis
pneumatic cylinders pressures; pressure of the grippers
cylinder; two proximity sensors detecting when the gripper is closed and opened were monitored during the experiments. The signals of 4 analog pressure sensors provided
important information about the system. They had signicant noise. The compressor pressure was selected at the
borderline to increase the difculty of condition estimation. When the valves were activated the signals of all
the pressure sensors uctuated. The each sensors data
are presented in Figs. 16 for 11 different conditions. The
pressure variation of the gripper actuation cylinder is
presented in Fig. 1. The signals of the proximity sensors
detecting when the gripper closed and opened are

Fig. 1. Experimental set-up.

18

M. Demetgul et al. / Measurement 55 (2014) 1524

6
5

Volts (V)

4
3
2
1
0
12

11

10

Experiments

15

12

18

21

24

27

Time (s)

Fig. 2. The pressure of the gripper activation cylinder.

6
5

Volts (V)

4
3
2
1
0
12

11

10

Experiments

15

12

18

21

24

27

Time (s)

Fig. 3. Proximity switch detecting gripper is closed.

presented in Figs. 2 and 3 respectively. The systems main

pressure is presented in Fig. 4. The pressures of the
cylinders creating the motion in the X and Y axes are presented in Figs. 5 and 6 respectively.
In Fig. 2 gripper to take the material to the specied
location until you release the gripper during the period
showing the change in pressure is observed. Gripper
around 4 bar pressure, in general, are monitored with

pressure uctuations, until the system pressure is reduced

to 2 bar indicates the status of the fault with the experiment number 3 gripper closure failure of the pressure
value for the duration of the experiment created 9 numbered falls naturally observed. The gripper activation time
can be followed by the Fig. 2. It is activated from 6th to
21st seconds. The uctuation in the gure occurs between
the 12th and 15th seconds.

19

M. Demetgul et al. / Measurement 55 (2014) 1524

Volts (V)

0.8
0.6
0.4
0.2
0
12

11

10

Experiments

12

15

18

21

24

27

Time (s)

Fig. 4. Proximity switch detecting gripper is opened.

5.5
5

Volts (V)

4.5
4
3.5
3
2.5
2
1.5
12

11

10

Experiments

12

15

18

21

24

27

Time (s)

Fig. 5. The main pressure of the system.

Exchange of closed gripper shown in Fig. 3, the sensor

gives only ON/OFF (1 and 0) information which is different from other conditions. The information is generated
in experiment number 9 and 11, for a period of less than
1 signals due to malfunction. Gripper shown, the open
exchange of information in Fig. 4 that the sensor is only
1 and 0 on the state of knowledge of the changes
because it shows the opposite of the situation described
in Fig. 6. As shown in Fig. 4, the gripper deactivation

time from 6th to 18th seconds. In experiment 3, proximity switch gives continuously ON signal because of the
fault.
Fig. 5 shows the main system pressure changes during
all the experiments. Here, up to 3 of the experiment, reducing the failure pressure 2 bar chart was created monitored
the situation. From the above scenario, the system capacity
compressor to normal operation as a result of the use of
active and passive valves, the pressure variation in the sig-

20

M. Demetgul et al. / Measurement 55 (2014) 1524

6
5

Volts (V)

4
3
2
1
0
12

11

10

Experiments

12

15

18

21

24

27

Time (s)

Fig. 6. The pressure of the cylinder moving the gripper along the X axis.

6
5

Volts (V)

4
3
2
1
0
12

11

10

Experiments

12

15

18

21

24

27

Time (s)

Fig. 7. The pressure of the cylinder moving the gripper along the Y axis.

nal chart are monitored during take. However, the system

in general, except for 3 of the experiment, denes low
pressure fault, show that a change around 4 bar.
Figs. 6 and 7 show the pressure of the cylinder moving
the gripper along the X and Y axis respectively. According
to the Fig. 6, X axes position valve pressure changes
between the 12 and 15 s where the cylinder is activated.
In the 4th and 5th experiments pressure of the valve is

higher than the others because of the X axis error positioning for forward and back motion. Also, as seen in Fig. 7, Y
axes position valve pressure changes at the two points
(69th s and 1821st s) where the cylinder is activated
and deactivated. Actually this gure shows the scenario
of the designed automation system for the gripper actions
which attached at the bottom of Y axes. Especially in the
experiments 6th and 7th pressure of the valve is over the

M. Demetgul et al. / Measurement 55 (2014) 1524

21

Fig. 8. Flowchart of experimental design.

data was reduced by using the DM, LLE and AE methods.

Thus the useful properties, dening the signals adequately,
have been obtained from the collected signals for using
further classication process. Then the compressed data
was classied by using the GK and k-medoids algorithms.

250
DM
LLE
AE

Cost Value

200

150

6. Results

100

50

10

20

30

40

50

60

70

Iteration Number
Fig. 9. Cost versus iteration when the compressed data was classied
with the GK algorithm.

3 bars because of the Y axis error positioning for up and

down positions.
5. Dimension reduction and classication of the
experimental data
The two step process for the analysis of the experimental data is presented in Fig. 8. First the dimension of the

In this study, the data that was taken from the system
has been embedded to ve dimension. The classication
process has been performed with this new feature space.
The experimental results are shown in Figs. 913.
The variation of the cost value with the iterations of the
GK algorithm is presented in Fig. 9. The compressed data of
the DM, LLE and AE dimension reduction algorithms were
used.
The classication performance values for all algorithms
were obtained by comparing the accurate classication
values and the values getting from algorithms. Figs. 10
13 show the randomly chosen values obtaining from the
dimensionally reduced data for X and Y axis. X and Y axis
illustrates only two dimensions among ve dimensions
respectively.
The obtained results after the GK algorithm classied
the encoded data of the DM method were given in

Fig. 10. The classication results after the data was compressed with the DM method.

22

M. Demetgul et al. / Measurement 55 (2014) 1524

Fig. 11. The classication results after the data was compressed with the LLE method.

Fig. 12. The classication results after the data was compressed with the AE method.

Fig. 10. According to the gure the classication performance was found as 88.05%.
Fig. 11 shows the classication results after the data
compressed by the LLE method was partitioned by the GK
algorithm. The accurate classication performance was
91.07%.
The worst estimations were observed when the GK
algorithm classied the data compressed by the AE method
(Fig. 12). The result of the classication performance is
78.95%.
The classication results after the dimension reduction
process with using LLE algorithm and classied with kmedoid is seen Fig. 13. The results performance is 90.75%.

The performance of dimension reduction (compression

or encoding) methods and classication algorithms are
outlined in Fig. 14. The best results were obtained when
the GK algorithm classied the data compressed by the
LLE method. Only two cases were classied wrong in this
case.
In this study, according to the results obtained when
using a method of dimensionality reduction algorithm Gk
better results were attained. Because gk method is a
method of fuzzy based classication has increased results
performance. Also in applications gk produces faster
results during the classication of faults. Studies
conducted by method of both classications have been

23

M. Demetgul et al. / Measurement 55 (2014) 1524

Fig. 13. The results after the k-medoids algorithm classied the data compressed by the LLE method.

Classification performance

100

DM
LLE
AE

Perofrmance %

80

60

40

20

0
GKclust

k-medoids

Fig. 14. The classication performance of each method to the dataset with GK and k-medoids.

found better results with LLE algorithm. LLE algorithm

reduction in the size of the data point for which rearrange
the data points are all possibilities to nearest neighbor
points and weights to calculate the best way.
7. Conclusion
The typical material handling system of automated
manufacturing facilities was simulated by using a trainer.
The gripper of the trainer picked up objects, moved and
piled up. Three pneumatic cylinders moved the gripper
along the x and y axis in addition to opened and closed
it. The system was operated at the normal and 10 faulty
modes to collect data. Four pressure and two digital gripper mode (open/close) signals were monitored in the time
domain. The pressures of the entire system and three cylinders were monitored. These cylinders activated the gripper and moved it along the x and y axis. Proximity sensors
were used to detect when the gripper opened and closed.

Experiments were repeated three times at each test condition. The characteristics of the signals of six sensors were
reasonably repetitive at the identical operating conditions.
The performances of three dimension reduction (encoding) methods and two classication algorithms were evaluated when they worked together. For dimension
reduction DM, LLE and AE methods were used. GK and kmedoids algorithms classied the encoded data. The best
classication performance was observed when the GK
algorithm classied the data encoded by the LLE method.
The performance of the GK and k-medoids were not on
part with the AE algorithm for the considered cases.
Acknowledgements
The authors present their special thanks to the Celal
Bayar University Scientic Research Projects Commission
for the supports of the study under project number 201252.

24

M. Demetgul et al. / Measurement 55 (2014) 1524

The authors also thanks to the Marmara University Scientic Research Projects Commission for the supports of
the study under project number FEN-A-080410-0081.
References
[1] L. Mendona, J. Sousa, J. S da Costa, An architecture for fault
detection and isolation based on fuzzy methods, Expert Syst. Appl.
36 (2009) 10921104.
[2] L. Shi, N. Sepehri, Adaptive fuzzy-neural-based multiple models for
fault diagnosis of a pneumatic actuator, in: Proceedings of the 2004
American Control Conference, 2004, pp. 37533758.
[3] Z. Nakutis, P. Kakonas, An approach to pneumatic cylinder on-line
conditions monitoring, Mechanika.-Kaunas: Technologija, 2008, p.
72.
[4] J. Wang, J. Wang, N. Daw, Q. Wu, Identication of pneumatic cylinder
friction parameters using genetic algorithms, IEEE/ASME Trans.
Mech. 9 (2004) 100107.
[5] M. Karpenko, N. Sepehri, D. Scuse, Diagnosis of process valve
actuator faults using a multilayer neural network, Control Eng.
Pract. 11 (2003) 12891299.
[6] Z. Nakutis, P. Kaskonas, Pneumatic cylinder diagnostics using
classication methods, in: Instrumentation and Measurement
Technology Conference Proceedings, 2007. IMTC 2007. IEEE, 2007,
pp. 14.
[7] Z. Nakutis, P. Kakonas, Application of ANN for pneumatic cylinder
leakage diagnostics, Matavimai 36 (2005) 1621.
[8] W. Yang, Establishment of the mathematical model for diagnosing
the engine valve faults by genetic programming, J. Sound Vib. 293
(2006) 213226.
[9] M. Sepasi, F. Sassani, On-line fault diagnosis of hydraulic systems
using unscented Kalman lter, Int. J. Control Autom. Syst. 8 (2010)
149156.
[10] T. Nogami, Y. Yokoi, M. Kasai, K. Kawai, K. Takaura, Failure diagnosis
system on pneumatic control valves by neural network, IEEE Int.
Conf. Neural Networks 1993 (1993) 18761881.
[11] B. Ould Bouamama, K. Medjaher, M. Bayart, A. Samantaray, B.
Conrard, Fault detection and isolation of smart actuators using bond
graphs and external models, Control Eng. Pract. 13 (2005) 159175.
[12] M. Demetgul, I. Tansel, S. Taskin, Fault diagnosis of pneumatic
systems with articial neural network algorithms, Expert Syst. Appl.
36 (2009) 1051210519.
[13] M. Demetgul, S. Taskin, I.N. Tansel, Conditioning Monitoring and
Fault Diagnosis for a Servo-Pneumatic System with Articial Neural
Network Algorithms, 2011.
[14] M. Demetgul, M. Unal, I. Tansel, O. Yazcoglu, Fault diagnosis on
bottle lling plant using genetic-based neural network, Adv. Eng.
Softw. 42 (2011) 10511058.
[15] F.J. Uppal, R.J. Patton, Fault diagnosis of an electro-pneumatic valve
actuator using neural networks with fuzzy capabilities, Proc. Eur.
Symp. Art. Neural Networks (2002) 501506.
[16] F.J. Uppal, R.J. Patton, V. Palade, Neuro-fuzzy based fault diagnosis
applied to an electro-pneumatic valve, in: Proceedings of the 15th
IFAC World Congress, 2002, pp. 2126.
[17] S.O. Song, G. Lee, E. Yoon, Process monitoring of an electropneumatic valve actuator using kernel principal component
analysis, Proc. Int. Symp. Adv. Control Chem. Process. (2003).
[18] S.F. Yuan, F.L. Chu, Fault diagnostics based on particle swarm
optimisation and support vector machines, Mech. Syst. Signal
Process. 21 (2007) 17871798.

[19] T. Kohonen, Self-organization and associative memory, SelfOrganization and Associative Memory, 100 gs. XV, 312 pages.
Springer-Verlag, Berlin Heidelberg New York. Also Springer Series in,
Information Sciences, volume 8, vol. 1, 1988.
[20] R.O. Duda, P.E. Hart, D.G. Stork, Pattern Classication, vol. 2, Wiley,
New York, 2003.
[21] D.E. Gustafson, W.C. Kessel, Fuzzy clustering with a fuzzy covariance
matrix, in: Conference on Decision and Control, San Diego,
California, 1979, pp. 761766.
[22] P. Teppola, P. Minkkinen, Possibilistic and fuzzy C-means clustering
for process monitoring in an activated sludge waste-water
treatment plant, J. Chemom. 13 (1999) 445459.
[23] S.W. Choi, C.K. Yoo, I.B. Lee, Overall statistical monitoring of static
and dynamic patterns, Ind. Eng. Chem. Res. 42 (2003) 108117.
[24] Q.P. He, S.J. Qin, J. Wang, A new fault diagnosis method using fault
directions in Fisher discriminant analysis, AIChE J. 51 (2005) 555
571.
[25] Q. Zhang, I. Couloigner, A new and efcient k-medoid algorithm for
spatial clustering, Comput. Sci. Its Appl. ICCSA 2005 (2005) 207
224.
[26] R.R. Coifman, S. Lafon, A.B. Lee, M. Maggioni, B. Nadler, F. Warner,
et al., Geometric diffusions as a tool for harmonic analysis and
structure denition of data: diffusion maps, Proc. Nat. Acad. Sci. USA
102 (2005) 74267431.
[27] R.R. Coifman, S. Lafon, Diffusion maps, Appl. Comput. Harm. Anal. 21
(2006) 5394.
[28] S. Lafon, Y. Keller, R.R. Coifman, Data fusion and multicue data
matching by diffusion maps, IEEE Trans. Pattern Anal. Mach. Intell.
28 (2006) 17841797.
[29] S. Lafon, A.B. Lee, Diffusion maps and coarse-graining: a unied
framework for dimensionality reduction, graph partitioning, and
data set parameterization, IEEE Trans. Pattern Anal. Mach. Intell. 28
(2006) 13931403.
[30] S.T. Roweis, L.K. Saul, Nonlinear dimensionality reduction by locally
linear embedding, Science 290 (2000) 2323.
[31] L.J.P. Van der Maaten, An introduction to dimensionality reduction
using matlab, 2007.
[32] D. DeMers, G. Cottrell, Non-linear dimensionality reduction, Adv.
Neural Inform. Process. Syst. (1993) 580587.
[33] G.E. Hinton, R.R. Salakhutdinov, Reducing the dimensionality of data
with neural networks, Science 313 (2006) 504507.
[34] G.E. Hinton, S. Osindero, Y.W. Teh, A fast learning algorithm for deep
belief nets, Neural Comput. 18 (2006) 15271554.
[35] J.C. Bezdek, J. Keller, R. Krisnapuram, Fuzzy models and algorithms
for pattern recognition and image processing, vol. 4, Kluwer
Academic Publishers, 1999.
[36] C. Sarbu, K. Zehl, J.W. Einax, Fuzzy divisive hierarchical clustering of
soil data using GustafsonKessel algorithm, Chem. Intell. Lab. Syst.
86 (2007) 121129.
[37] P.J. Rousseeuw, L. Kaufman, Finding Groups in Data: An Introduction
to Cluster Analysis, John Wiley & Sons, 1990.
[38] B. Balasko, J. Abonyi, B. Feil, Fuzzy Clustering and Data Analysis
Toolbox, Department of Process Engineering, University of
Veszprem, Veszprem, 2005.
[39] R. Xu, S. Damelin, B. Nadler, I. Wunsch, C. Donald, Clustering of highdimensional gene expression data with feature ltering methods
and diffusion maps, Artif. Intell. Med. 48 (2010) 9198.
[40] S.Y. Kung, K.I. Diamantaras, J.S. Taur, Adaptive principal component
extraction (APEX) and applications, IEEE Trans. Signal Process. 42
(1994) 12021217.