1-s2.0-S026322411400195X-main.pdf

Measurement 55 (2014) 15–24

Contents lists available at ScienceDirect

Measurement

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

Fault diagnosis on material handling system using featureselection and data mining techniques

http://dx.doi.org/10.1016/j.measurement.2014.04.0370263-2241/� 2014 Elsevier Ltd. All rights reserved.

⇑ Corresponding author. Tel.: +90 216 336 57 70.E-mail address: [email protected] (M. Demetgul).

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, Turkeyb Department of Computer and Control Education, Technical Education Faculty, Marmara University, Turkeyc Department of Electrical and Electronics Engineering, Engineering Faculty, Celal Bayar University, Turkeyd Department of Materials and Mechanical Engineering, Engineering Faculty, Florida International University, FL, USAe Department of Industrial Engineering, Design and Engineering Faculty, Istanbul Commerce University, Turkey

a r t i c l e i n f o

Article history:Received 4 February 2014Received in revised form 21 April 2014Accepted 23 April 2014Available online 9 May 2014

Keywords:Servo-pneumaticMaterial handling systemFault diagnosisFeature selectionData miningDimension reductionGustafson–Kesselk-Medoids

a b s t r a c t

The material handling systems are one of the key components of the most modern manu-facturing systems. The sensory signals of material handling systems are nonlinear and haveunique characteristics. It is very difficult to encode and classify these signals by using mul-tipurpose methods. In this study, performances of multiple generic methods were studiedfor the diagnostic of the pneumatic systems of the material handling systems. DiffusionMap (DM), Local Linear Embedding (LLE) and AutoEncoder (AE) algorithms were used forfuture extraction. Encoded signals were classified by using the Gustafson–Kessel (GK)and k-medoids algorithms. The accuracy of the estimations was better than 90% whenthe LLE was used with GK algorithm.

� 2014 Elsevier Ltd. All rights reserved.

1. Introduction

The modern manufacturing facilities have to detect theproblems, identify their sources and fix them very quicklywith very limited man power. Researchers have starteddevelopment of computational diagnostic tools for theindustrial applications in early 1970s by considering thisneed. Although, various diagnostic tools have been devel-oped by research community and successfully used inindustrial applications in last two decades [1,2] still theircapabilities are limited. In this study, feasibility of a multi-purpose fault detection approach was investigated. Theproposed approach used the combinations of the generic

dimension reduction methods for feature extraction andclassified the encoded data with clustering algorithms.

One of the key components of the automated manufac-turing is material handling systems. Pneumatic andhydraulic systems are widely used for material handling.These systems may have hundreds of actuators and sen-sors. Identification of faulty components and their loca-tions in a very short time is very difficult. Several studieswere performed for development of fault diagnostic toolsfor these systems in the last decade [3]. The studies mainlyaimed evaluation of the condition of the cylinders [4] anddigitally controlled valves [5]. Other studies focused ondetection of leakage of the seals [6–9], friction increase[4,10] and malfunctions [11–14].

Most of the fault diagnostic tools have two compo-nents: feature extractor (encoder) and classifier. Someresearchers have used the intelligent data analysis

16 M. Demetgul et al. / Measurement 55 (2014) 15–24

techniques for fault diagnostic [15–17]. Support vectormachines [18], self-organizing feature maps (SOM) [19],expert systems, neural networks, rough sets and fuzzylogic have been used for classification of data. The comput-ing complexity of feature extraction and learning processhave been the main disadvantages of these approaches.

The data of the material handling systems for the faultdiagnosis comes from multiple sensors. The data is high-dimensional and nonlinear. While the large number of datafrom different sensors provide more information, at thesame time feature extraction and classification becomesmore complex. The dimension reduction methods com-press the data automatically, reduce the noise, may extractfeatures for fault diagnostic and minimize required storage.

Clustering algorithms have been used for classification.Fuzzy c-means (FCM) [20] and its variants Gustafson–Kes-sel (GK) [21] algorithm are popular pattern classificationmethods. They have been used for fault detection and iso-lation [22–24], k-medoids [25] is a partitional clusteringalgorithm and may be used for classification purposes.DM method [26–29] used diffusion semigroups for learn-ing the global characteristics of the data-set. The complexstructures were represented at different scales by the helpof these semi groups. The eigenfunctions of Markov matri-ces were effectively used with this purpose. LLE [30]method used an unsupervised learning algorithm tochange the dimensions of the data. The algorithm foundthe neighbors in X space, calculated weights for recon-struction and calculated the embedding coordinates in Yspace by using the calculated weights. AutoEncoder (AE)[31] methods use an artificial neural network (ANN) tolearn the compact representation of data set. The dimen-sionality of the data set is reduced by using this ANN. Var-ious multilayer architectures [32,33] and optimizationmethods [31,34] have been proposed to improve the per-formance of the ANN.

GK and k-medoids algorithms were used to create thedesired number of clusters to partition or classify the dataafter it was compacted. GK algorithm [21] calculates thecenter and covariance matrix to represent the clusters[35,36]. They are used during the optimization process.This approach allows identification of ellipsoidal clustersand improves the performance of the method relative toother approaches. k-medoids [37] is another clusteringalgorithm. The algorithm divides the data into the groups,chooses the data points as medoids. In this study, Cluster-ing and Data Analysis Toolbox [38] was used for classifica-tion of the compressed data.

Diffusion map, AutoEncoder, and Local Lineer Embed-ding techniques were used for dimension reduction pro-cess respectively. In classification process two algorithmswere used namely; k-medoids and GK. These algorithmswere given in detailed below.

2. Classification and feature extraction

2.1. k-Medoids

It is a standard clustering algorithm [37] where theupdate rule always moves the cluster center to the nearestdata point in the cluster. k-Medoids is a partitioning

technique of clustering that clusters the data set of nobjects into k clusters with k known a priori. t could bemore robust to noise and outliers as compared to k-meansbecause it minimizes a sum of general pairwise dissimilar-ities instead of a sum of squared Euclidean distances.

2.2. Gustafson–Kessel algorithm (GK)

Gustafson–Kessel algorithm (GK), providing a degree ofmembership of each data point to a particular cluster cre-ates a fuzzy partition [21]. One set of data to detect clustersof different geometrical shapes, this method is an adaptivedistance norm for each cluster was introduced. Each clus-ter has its own standard distance norm affects the Aimatrix Mi, inducing which statement is the followingequation [36].

dðxj; LiÞM ¼ IIxj � LIII ¼ ðxj � LiÞTMiðxj � LiÞ ð1Þ

possible use of the Z matrix of each cluster, every step ofthe data fit the geometric structure of the distance norm.Based on the norm-inducing matrices, the aim of the GKmethod, is obtained by minimizing the function J as Eq. (2).

JðP; L;MÞ ¼Xc

i¼1

Xn

j¼1

ðAiðxjÞÞ2d2ðxj; LiÞ0M ð2Þ

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

2.3. Diffusion maps

Diffusion maps as a system of eigenfunctions of Markovmatrices consider effective representation of data geomet-ric descriptions of the original data set to obtain the coor-dinates [27–29]. A given data set X = (x, i = 1, � � �, N) is a d-dimensional data space, said N nodes can be built over X isa finite graph corresponding to the N data. Usually, theform of a Gaussian kernel as in Eq. (3)

wðxi; xjÞ ¼ expð� IIxi � xjII2

2a2 Þ ð3Þ

where r is the kernel width parameter, is used to constructthe similarity matrix. The kernel reflects the degree of sim-ilarity between xi and xj, and II, II is the Euclidean norm inRd [39].

2.4. AutoEncoder

Multilayer encoders are hidden layer feed forward neu-ral networks with an odd number [32,33]. The input andthe output layer have D nodes and the middle hidden layerhas d nodes.

AutoEncoders usually high number of multi-layer con-nection. Therefore, the back propagation approaches islikely to get stuck in slow convergence and local minimum.In [32] this disadvantage is overcome by performing aRestricted Boltzmann Machines using pretraining (RBMs)[34]. The mean square error between the input and outputof the network is trained to minimize (Ideally, the inputand output equal). Linear activation functions in the useof neural network, PCA is a very similar AutoEncoder [40].

Table 1Faults in the system and output values of ANN algorithm.

Faults Experiments Symbol

Normal operation (System operationpressure 6 bar)

1 Normal

Low pressure (LP) (System operationpressure 4 bar)

2 Fault1

Low Low pressure (LLP) (Systemoperation pressure 2 bar)

3 Fault2

x Axis error positioning for back motion 4 Fault3x Axis error positioning for forward

motion5 Fault4

y Axis error positioning for up motion 6 Fault5y Axis error positioning for down motion 7 Fault6Pick faults for gripper (one pneumatic

hose release)8 Fault7

Place faults for gripper (other pneumatichose release)

9 Fault8

Gripper close fault (digital magneticproximity sensor which send gripperclose information cable disconnectfault)

10 Fault9

After gripper receive material to dropfailure.

11 Fault10

Fig. 1. Experimental set-up.

M. Demetgul et al. / Measurement 55 (2014) 15–24 17

2.5. Local Linear Embedding (LLE)

Local Linear Embedding (LLE) [30] is a local techniquefor dimensionality reduction that is similar to Isomap inthat it constructs a graph representation of the datapoints.In contrast to Isomap, it attempts to preserve solely localproperties of the data, making LLE less sensitive to short-circuiting than Isomap. Furthermore, the preservation oflocal properties allows for successful embedding of non-convex manifolds. In LLE, the local properties of the datamanifold are constructed by writing the datapoints as alinear combination of their nearest neighbors. In the low-dimensional representation of the data, LLE attempts toretain the reconstruction weights in the linear combina-tions as well as possible.

3. Experimental set-up

In this study, servo-pneumatic positioning experimen-tal set-up built by the Festo Didactic Company was used.The experimental setup is presented in Fig. 3. The gripper’smotions along the X and Y axes were controlled with pneu-matic dual action rodless cylinders. For the position mea-surements in the X and Y axes a linear potentiometer anda contactless absolute magnetostrictive linear displace-ment sensor were used respectively. The pneumatic grip-per of the system was installed at the Y axis actuator.

The proportional directional control valves controlledthe air flow to the cylinders. Four analog transducers mea-sured the pressures of the entire system and the actuatorscontrolling the motions and the gripper. Experimental datawere collected by using the National Instrument (NI) com-pact Data Acquisition (NI-cDAQ) system and control mod-ules. LabVIEW™ program was used for management andanalysis of the collected data.

The trainer was programmed to pick up the parts from aconveyor and to put on a pallet in a systematic fashion. So,the programmed motion was very similar but slightly

different to pile the parts next to each other. The LabVIEWuser interface is prepared. Four analog sensors showed thepressures of the system at the compressor output, and cyl-inders. One cylinder opened and closed the gripper whilethe other two moved it along the X and Y axes. Data wascollected with 50 ms sampling interval. Each experimenttook 27 s.

4. Experimental procedure and data collection

In this study, data was collected while the pneumaticsystem was operated at the normal and additional 10 dif-ferent faulty conditions. The imposed problems are listedin Table 1. During the experiments, the data were collectedfrom 4 analog and 2 digital sensors for a period of 27 s. Thedata was collected 3 times at each experimental condition.The pneumatic system’s main pressure; the x and y axispneumatic cylinders’ pressures; pressure of the gripper’scylinder; two proximity sensors detecting when the grip-per is closed and opened were monitored during the exper-iments. The signals of 4 analog pressure sensors providedimportant information about the system. They had signifi-cant noise. The compressor pressure was selected at theborderline to increase the difficulty of condition estima-tion. When the valves were activated the signals of allthe pressure sensors fluctuated. The each sensor’s dataare presented in Figs. 1–6 for 11 different conditions. Thepressure variation of the gripper actuation cylinder ispresented in Fig. 1. The signals of the proximity sensorsdetecting when the gripper closed and opened are

0 36 9 12 15 18 21 24

27

01234567891011120

1

2

3

4

5

6

Time (s)Experiments

Volts

(V)

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

03

69

1215

1821

2427

01

23

45

67

8910

11120

1

2

3

4

5

6

Time (s)Experiments

Volts

(V)

Fig. 3. Proximity switch detecting gripper is closed.

18 M. Demetgul et al. / Measurement 55 (2014) 15–24

presented in Figs. 2 and 3 respectively. The system’s mainpressure is presented in Fig. 4. The pressures of thecylinder’s creating the motion in the X and Y axes are pre-sented in Figs. 5 and 6 respectively.

In Fig. 2 gripper to take the material to the specifiedlocation until you release the gripper during the periodshowing the change in pressure is observed. Gripperaround 4 bar pressure, in general, are monitored with

pressure fluctuations, until the system pressure is reducedto 2 bar indicates the status of the fault with the experi-ment number 3 gripper closure failure of the pressurevalue for the duration of the experiment created 9 num-bered falls naturally observed. The gripper activation timecan be followed by the Fig. 2. It is activated from 6th to21st seconds. The fluctuation in the figure occurs betweenthe 12th and 15th seconds.

0 36 9 12 15 18 21 24

27

01234567891011120

0.2

0.4

0.6

0.8

1

Time (s)Experiments

Volts

(V)

Fig. 4. Proximity switch detecting gripper is opened.

03

69

1215

1821

2427

01

23

45

67

8910

1112

1.5

2

2.5

3

3.5

4

4.5

5

5.5

Time (s)Experiments

Volts

(V)

Fig. 5. The main pressure of the system.

M. Demetgul et al. / Measurement 55 (2014) 15–24 19

Exchange of closed gripper shown in Fig. 3, the sensorgives only ON/OFF (1 and 0) information which is differ-ent from other conditions. The information is generatedin experiment number 9 and 11, for a period of less than1 signals due to malfunction. Gripper shown, the openexchange of information in Fig. 4 that the sensor is only1 and 0 on the state of knowledge of the changesbecause it shows the opposite of the situation describedin Fig. 6. As shown in Fig. 4, the gripper deactivation

time from 6th to 18th seconds. In experiment 3, proxim-ity switch gives continuously ON signal because of thefault.

Fig. 5 shows the main system pressure changes duringall the experiments. Here, up to 3 of the experiment, reduc-ing the failure pressure 2 bar chart was created monitoredthe situation. From the above scenario, the system capacitycompressor to normal operation as a result of the use ofactive and passive valves, the pressure variation in the sig-

0 36 9 12 15 18 21 24

27

01234567891011120

1

2

3

4

5

6

Time (s)Experiments

Volts

(V)

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

03

69

1215

1821

2427

01

23

45

67

8910

11120

1

2

3

4

5

6

Time (s)Experiments

Volts

(V)

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

20 M. Demetgul et al. / Measurement 55 (2014) 15–24

nal chart are monitored during take. However, the systemin general, except for 3 of the experiment, defines lowpressure fault, show that a change around 4 bar.

Figs. 6 and 7 show the pressure of the cylinder movingthe gripper along the X and Y axis respectively. Accordingto the Fig. 6, X axes position valve pressure changesbetween 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 position-ing for forward and back motion. Also, as seen in Fig. 7, Yaxes position valve pressure changes at the two points(6–9th s and 18–21st s) where the cylinder is activatedand deactivated. Actually this figure shows the scenarioof the designed automation system for the gripper actionswhich attached at the bottom of Y axes. Especially in theexperiments 6th and 7th pressure of the valve is over the

Fig. 8. Flowchart of experimental design.

0 10 20 30 40 50 60 700

50

100

150

200

250

Iteration Number

Cos

t Val

ue

DMLLEAE

Fig. 9. Cost versus iteration when the compressed data was classifiedwith the GK algorithm.

M. Demetgul et al. / Measurement 55 (2014) 15–24 21

3 bars because of the Y axis error positioning for up anddown positions.

5. Dimension reduction and classification of theexperimental data

The two step process for the analysis of the experimen-tal data is presented in Fig. 8. First the dimension of the

Fig. 10. The classification results after the data

data was reduced by using the DM, LLE and AE methods.Thus the useful properties, defining the signals adequately,have been obtained from the collected signals for usingfurther classification process. Then the compressed datawas classified by using the GK and k-medoids algorithms.

6. Results

In this study, the data that was taken from the systemhas been embedded to five dimension. The classificationprocess has been performed with this new feature space.The experimental results are shown in Figs. 9–13.

The variation of the cost value with the iterations of theGK algorithm is presented in Fig. 9. The compressed data ofthe DM, LLE and AE dimension reduction algorithms wereused.

The classification performance values for all algorithmswere obtained by comparing the accurate classificationvalues and the values getting from algorithms. Figs. 10–13 show the randomly chosen values obtaining from thedimensionally reduced data for X and Y axis. X and Y axisillustrates only two dimensions among five dimensionsrespectively.

The obtained results after the GK algorithm classifiedthe encoded data of the DM method were given in

was compressed with the DM method.

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

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

22 M. Demetgul et al. / Measurement 55 (2014) 15–24

Fig. 10. According to the figure the classification perfor-mance was found as 88.05%.

Fig. 11 shows the classification results after the datacompressed by the LLE method was partitioned by the GKalgorithm. The accurate classification performance was91.07%.

The worst estimations were observed when the GKalgorithm classified the data compressed by the AE method(Fig. 12). The result of the classification performance is78.95%.

The classification results after the dimension reductionprocess with using LLE algorithm and classified with k-medoid is seen Fig. 13. The results performance is 90.75%.

The performance of dimension reduction (compressionor encoding) methods and classification algorithms areoutlined in Fig. 14. The best results were obtained whenthe GK algorithm classified the data compressed by theLLE method. Only two cases were classified wrong in thiscase.

In this study, according to the results obtained whenusing a method of dimensionality reduction algorithm Gkbetter results were attained. Because gk method is amethod of fuzzy based classification has increased resultsperformance. Also in applications gk produces fasterresults during the classification of faults. Studiesconducted by method of both classifications have been

GKclust k-medoids0

20

40

60

80

100

Pero

frman

ce %

Classification performance

DMLLEAE

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

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

M. Demetgul et al. / Measurement 55 (2014) 15–24 23

found better results with LLE algorithm. LLE algorithmreduction in the size of the data point for which rearrangethe data points are all possibilities to nearest neighborpoints and weights to calculate the best way.

7. Conclusion

The typical material handling system of automatedmanufacturing facilities was simulated by using a trainer.The gripper of the trainer picked up objects, moved andpiled up. Three pneumatic cylinders moved the gripperalong the x and y axis in addition to opened and closedit. The system was operated at the normal and 10 faultymodes to collect data. Four pressure and two digital grip-per mode (open/close) signals were monitored in the timedomain. The pressures of the entire system and three cyl-inders were monitored. These cylinders activated the grip-per and moved it along the x and y axis. Proximity sensorswere used to detect when the gripper opened and closed.

Experiments were repeated three times at each test condi-tion. The characteristics of the signals of six sensors werereasonably repetitive at the identical operating conditions.

The performances of three dimension reduction (encod-ing) methods and two classification algorithms were eval-uated when they worked together. For dimensionreduction DM, LLE and AE methods were used. GK and k-medoids algorithms classified the encoded data. The bestclassification performance was observed when the GKalgorithm classified the data encoded by the LLE method.The performance of the GK and k-medoids were not onpart with the AE algorithm for the considered cases.

Acknowledgements

The authors present their special thanks to the CelalBayar University Scientific Research Projects Commissionfor the supports of the study under project number 2012-52.

24 M. Demetgul et al. / Measurement 55 (2014) 15–24

The authors also thanks to the Marmara University Sci-entific Research Projects Commission for the supports ofthe study under project number FEN-A-080410-0081.

References

[1] L. Mendonça, J. Sousa, J. Sá da Costa, An architecture for faultdetection and isolation based on fuzzy methods, Expert Syst. Appl.36 (2009) 1092–1104.

[2] L. Shi, N. Sepehri, Adaptive fuzzy-neural-based multiple models forfault diagnosis of a pneumatic actuator, in: Proceedings of the 2004American Control Conference, 2004, pp. 3753–3758.

[3] Z. Nakutis, P. Kaškonas, An approach to pneumatic cylinder on-lineconditions monitoring, Mechanika.-Kaunas: Technologija, 2008, p.72.

[4] J. Wang, J. Wang, N. Daw, Q. Wu, Identification of pneumatic cylinderfriction parameters using genetic algorithms, IEEE/ASME Trans.Mech. 9 (2004) 100–107.

[5] M. Karpenko, N. Sepehri, D. Scuse, Diagnosis of process valveactuator faults using a multilayer neural network, Control Eng.Pract. 11 (2003) 1289–1299.

[6] Z. Nakutis, P. Kaskonas, Pneumatic cylinder diagnostics usingclassification methods, in: Instrumentation and MeasurementTechnology Conference Proceedings, 2007. IMTC 2007. IEEE, 2007,pp. 1–4.

[7] Z. Nakutis, P. Kaškonas, Application of ANN for pneumatic cylinderleakage diagnostics, Matavimai 36 (2005) 16–21.

[8] W. Yang, Establishment of the mathematical model for diagnosingthe engine valve faults by genetic programming, J. Sound Vib. 293(2006) 213–226.

[9] M. Sepasi, F. Sassani, On-line fault diagnosis of hydraulic systemsusing unscented Kalman filter, Int. J. Control Autom. Syst. 8 (2010)149–156.

[10] T. Nogami, Y. Yokoi, M. Kasai, K. Kawai, K. Takaura, Failure diagnosissystem on pneumatic control valves by neural network, IEEE Int.Conf. Neural Networks 1993 (1993) 1876–1881.

[11] B. Ould Bouamama, K. Medjaher, M. Bayart, A. Samantaray, B.Conrard, Fault detection and isolation of smart actuators using bondgraphs and external models, Control Eng. Pract. 13 (2005) 159–175.

[12] M. Demetgul, I. Tansel, S. Taskin, Fault diagnosis of pneumaticsystems with artificial neural network algorithms, Expert Syst. Appl.36 (2009) 10512–10519.

[13] M. Demetgul, S. Taskin, I.N. Tansel, Conditioning Monitoring andFault Diagnosis for a Servo-Pneumatic System with Artificial NeuralNetwork Algorithms, 2011.

[14] M. Demetgul, M. Unal, I. Tansel, O. Yazıcıoglu, Fault diagnosis onbottle filling plant using genetic-based neural network, Adv. Eng.Softw. 42 (2011) 1051–1058.

[15] F.J. Uppal, R.J. Patton, Fault diagnosis of an electro-pneumatic valveactuator using neural networks with fuzzy capabilities, Proc. Eur.Symp. Art. Neural Networks (2002) 501–506.

[16] F.J. Uppal, R.J. Patton, V. Palade, Neuro-fuzzy based fault diagnosisapplied to an electro-pneumatic valve, in: Proceedings of the 15thIFAC World Congress, 2002, pp. 21–26.

[17] S.O. Song, G. Lee, E. Yoon, Process monitoring of an electro-pneumatic valve actuator using kernel principal componentanalysis, Proc. Int. Symp. Adv. Control Chem. Process. (2003).

[18] S.F. Yuan, F.L. Chu, Fault diagnostics based on particle swarmoptimisation and support vector machines, Mech. Syst. SignalProcess. 21 (2007) 1787–1798.

[19] T. Kohonen, Self-organization and associative memory, Self-Organization and Associative Memory, 100 figs. 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 Classification, vol. 2, Wiley,New York, 2003.

[21] D.E. Gustafson, W.C. Kessel, Fuzzy clustering with a fuzzy covariancematrix, in: Conference on Decision and Control, San Diego,California, 1979, pp. 761–766.

[22] P. Teppola, P. Minkkinen, Possibilistic and fuzzy C-means clusteringfor process monitoring in an activated sludge waste-watertreatment plant, J. Chemom. 13 (1999) 445–459.

[23] S.W. Choi, C.K. Yoo, I.B. Lee, Overall statistical monitoring of staticand dynamic patterns, Ind. Eng. Chem. Res. 42 (2003) 108–117.

[24] Q.P. He, S.J. Qin, J. Wang, A new fault diagnosis method using faultdirections in Fisher discriminant analysis, AIChE J. 51 (2005) 555–571.

[25] Q. Zhang, I. Couloigner, A new and efficient k-medoid algorithm forspatial 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 andstructure definition of data: diffusion maps, Proc. Nat. Acad. Sci. USA102 (2005) 7426–7431.

[27] R.R. Coifman, S. Lafon, Diffusion maps, Appl. Comput. Harm. Anal. 21(2006) 53–94.

[28] S. Lafon, Y. Keller, R.R. Coifman, Data fusion and multicue datamatching by diffusion maps, IEEE Trans. Pattern Anal. Mach. Intell.28 (2006) 1784–1797.

[29] S. Lafon, A.B. Lee, Diffusion maps and coarse-graining: a unifiedframework for dimensionality reduction, graph partitioning, anddata set parameterization, IEEE Trans. Pattern Anal. Mach. Intell. 28(2006) 1393–1403.

[30] S.T. Roweis, L.K. Saul, Nonlinear dimensionality reduction by locallylinear embedding, Science 290 (2000) 2323.

[31] L.J.P. Van der Maaten, An introduction to dimensionality reductionusing matlab, 2007.

[32] D. DeMers, G. Cottrell, Non-linear dimensionality reduction, Adv.Neural Inform. Process. Syst. (1993) 580–587.

[33] G.E. Hinton, R.R. Salakhutdinov, Reducing the dimensionality of datawith neural networks, Science 313 (2006) 504–507.

[34] G.E. Hinton, S. Osindero, Y.W. Teh, A fast learning algorithm for deepbelief nets, Neural Comput. 18 (2006) 1527–1554.

[35] J.C. Bezdek, J. Keller, R. Krisnapuram, Fuzzy models and algorithmsfor pattern recognition and image processing, vol. 4, KluwerAcademic Publishers, 1999.

[36] C. Sarbu, K. Zehl, J.W. Einax, Fuzzy divisive hierarchical clustering ofsoil data using Gustafson–Kessel algorithm, Chem. Intell. Lab. Syst.86 (2007) 121–129.

[37] P.J. Rousseeuw, L. Kaufman, Finding Groups in Data: An Introductionto Cluster Analysis, John Wiley & Sons, 1990.

[38] B. Balasko, J. Abonyi, B. Feil, Fuzzy Clustering and Data AnalysisToolbox, Department of Process Engineering, University ofVeszprem, Veszprem, 2005.

[39] R. Xu, S. Damelin, B. Nadler, I. Wunsch, C. Donald, Clustering of high-dimensional gene expression data with feature filtering methodsand diffusion maps, Artif. Intell. Med. 48 (2010) 91–98.

[40] S.Y. Kung, K.I. Diamantaras, J.S. Taur, Adaptive principal componentextraction (APEX) and applications, IEEE Trans. Signal Process. 42(1994) 1202–1217.