This action might not be possible to undo. Are you sure you want to continue?

**important for engineering and economy problems. Time series data
**

mining combines the fields of time series analysis and data mining

techniques. This method creates a set of methods that reveal hidden

temporal patterns that are characteristic and predictive of time series

events. Time series data mining examines the time series in a phase

space. In this paper, a prediction algorithm using time series data

mining based on fuzzy logic is proposed. Earthquake prediction has

been done from a synthetic earthquake time series by using

investigating method at first step ago. Time series has been

transformed to phase space by using nonlinear time series analysis

and then fuzzy logic has been used to prediction optimal values of

important parameters characterizing the time series events. Truth of

prediction algorithm based fuzzy logic has been proved by

application results.

Keywords—Event, fuzzy logic, time series analysis, time series

data mining

I. INTRODUCTION

time series data is obtained at determined time interval

from any system. Daily price change of a market, current

and voltage data of an induction motor and the population

distribution according to year in a country may be considered

as a time series. Such a time series contains events of interest

[1]. One innovate approach is to identify critical, time-ordered

structures, called temporal pattern, which are hidden in eq. (1)

and are characteristic of interesting events [2]. Time series

analysis is fundamental to engineering, scientific, and business

endeavors [3]. Investors are interested in predicting future

values from past time series data. A time series example has

been given in equation (1).

} ,......, 1 , { N t x X

t

= = (1)

In such a time series, t is time index and N is the total

number of observations. Important events are formed over the

time. For example, an earthquake from a synthetic seismic

time series, instantaneous fall at stock prices and incipient

faults in an induction motor are considered as an event for a

time series. Time series is analyzed for prediction of these

events. Traditional time series analysis methods such as Box-

I. Aydin is with Erzincan University, Erzincan, Zip code 24600 Turkey

(phone: 0090-533-428-56-92; fax: 0090-446-751-30-21; e-mail:

ilhanaydin@erzincan.edu.tr ).

M. Karakose is with the Computer Engineering Department, University of

Firat, Elazig, Zip code 23119 Turkey (e-mail: mkarakose@firat.edu.tr).

E. Akin is with the Computer Engineering Department, University of

Firat, Elazig, Zip code 23119 Turkey (E-mail: eakin@firat.edu.tr).

Jenkins or Autoregressive Integrated Moving Average

(ARIMA) method can be used to model such time series.

However, the ARIMA method is limited by the requirement of

stationary of time series and normality and independence of

residuals. An event characterization function g(t) can be

employed to characterize the event in equation (1) as shown in

equation (2).

) ,......, , ( ) (

1 1

x x x g t g

t t −

= (2)

This event characterization function changes according to

the prediction aim. For example, if x

t

represents today’s

closing price of a stock and our aim is to predict the

percentage change of tomorrow’s price, then the event

characterization function can be defined as shown in equation

(3). The event characterization function is defined such that its

value at t time index correlates highly with the occurrence of

an event at some specified time in the future [1]. If earthquake

prediction is wanted to do one step before, the event

characterization function in equation (4) can be selected.

t

t t

x

x x

t g

−

=

+1

) ( (3)

1

) (

+

=

t

x t g (4)

Event characterization functions at equation (3) and (4) can

be defined by different ways for event predictions at different

time series. Before the event characterization function is

determined, the aim is to select this function which predicts

important events. This is not trivial task due to nonlinear

behaviors of most time series in the real world applications.

An important disadvantage in time series analysis is that the

time series should be converted to stationary and periodic

series in order to analyze it [4]. As an emerging discipline,

data mining is the process of discovering hidden and useful

information from huge data. Data mining is defined as

extracting useful and meaningful information using statistic,

machine learning, artificial intelligence and pattern recognition

techniques from large data sets [5]. Data mining is the analysis

of data with the goal of uncovering hidden patterns. Weiss and

Indurkhya [6] defined data mining as “the search for valuable

information in large volumes of data”. Time series data

mining combines data mining and nonlinear time series

analysis to analyze a time series. When data mining is applied

to time series data, an event is considered as an interesting

pattern. Povinelli [1] defines time series data mining as

“combining of data mining, time series analysis and genetic

I. Aydin, M. Karakose, and E. Akin

The Prediction Algorithm Based on Fuzzy

Logic Using Time Series Data Mining Method

A

World Academy of Science, Engineering and Technology 51 2009

91

algorithm techniques”. He used the genetic algorithm to

discover interesting patterns in a time series by data mining.

Time series should be examined in a phase space in order to

get interesting pattern from it. The main goal in time series

data mining is to use time delay embedding and phase space

based on Taken theorem [7]. The phase space of a time series

in equation (1) is generated by using time delay embedding

and embedding dimension in order to get interesting patterns.

Elements of X

t

form time lagged shape of original time series

with embedding dimension b and time delay factor τ . The

phase space of the time series in equation (1) according to time

delay factor τ and embedding dimension b is presented in

equation (5). X is a matrix whose row vector is a point in the

phase space.

b

t b t b t t

R x x x X ∈ =

− − − −

) ,......., , (

) 2 ( ) 1 ( τ τ

(5)

Where τ =1, 2,….,k is time delay embedding which allows

the phase space to be spanned over nonconsecutive time-

lagged instances. For instance, τ =3 and b=3, the vector is

shown by X

t

=(x

t-6

,x

t-3

,x

t

). This transformation preserves the

nonlinear dynamics of the original time series and the temporal

patterns are represented by data points X

t

in the phase space

[2]. The temporal patterns representing the characteristic of

time series are used to prediction of the important events. This

method is successfully applied non-periodical, nonlinear, and

complex time series.

Data mining is the process of discovering useful pattern in

data that are hidden and important pattern. It comes from

several fields, including statistics, database and machine

learning [5]. It uses techniques such as clustering, association

rules, and classifications models to identify hidden and useful

information in large databases. Others who have applied data

mining concepts to finding patterns in time series include

Keogh and Smyth [8], [9]. Their approach uses a dynamic

programming method for aligning the time series and a

predefined set of templates. Moon et al. [10] proposed a new

time subsequence matching method. They used the dual match

for time series similarity. Time series data mining has been

used in a lot of fields such as clustering and event prediction in

literature. One of these methods is classification and fault

diagnosis in induction motors [11], [12]. Povinelli [1] was

suggested an application in order to event prediction. In his

study, he did different applications such as earthquake

prediction, sharp fall of stock price. At first step, he

transformed a time series to phase space. He selected different

event characterization functions for each application. Optimal

temporal pattern cluster with radius r and temporal pattern

center v was found by using genetic search algorithm. Despite

the novelty of this method, several problems limited its

applications and must to be addressed. First, temporal pattern

cluster in this study was defined by a rigid region with fixed

center and radius. This restriction made it hard to adjust to the

situation when there is noise in phase space and often

generated high degree false-positive prediction. The second

disadvantage is the computational complexity and stability.

The genetic algorithm yields heavy computing load. The

optimization results were often inconsistent. The choices of

embedding dimension b and time delay τ were always based

on user’s experience with trial-and-error. Feng et al. [2]

suggested new time series data mining for identifying temporal

patterns. In his study, temporal pattern cluster was chosen as a

fuzzy set with Gaussian shaped membership function. An

efficient two step optimization strategy was proposed to search

the optimal temporal cluster in the phase space. Time delay

embedding τ and embedding dimension b is chosen by mutual

information and false nearest neighbor method, respectively.

Gradient descent optimization algorithm is chosen in order to

find optimum temporal pattern cluster. Multilayer feed-

forward neural network was performed on the D-stock

prediction [13]. The goal was predicted next day’s closing

price change, with the input of today as well as four previous

day’s price change.

In this paper, proposed time series data mining method is

based on fuzzy logic. This method uses the fuzzy logic for

earthquake and Lorenz series prediction. Therefore, event

prediction doesn’t include complex mathematical equations

and the prediction is simplified. The prediction is flexible due

to property of fuzzy logic method. Because temporal pattern

cluster is selected as a fuzzy set, each of point in temporal

pattern cluster belong to cluster with a specific membership

degree. So the event prediction is predicted accurately. The

heavy load which the genetic algorithm yields is removed.

Because user experience about event prediction is shown in

phase space and augmented phase space easily, the fuzzy rules

is also determined, easily.

The paper is organized as follows: Section II provides brief

review of time series data mining using phase space strategy.

We introduce and discuss our application results in section III.

Conclusion and future study are discussed in section IV.

II. APPLICATIONS OF TIME SERIES DATA MINING

Data mining is a technique of discovering useful pattern in

data that are hidden and unknown in normal circumstance.

Data mining consist of machine learning, statistics and

database design [2]. It uses methods such as clustering,

classification, association rule mining and probabilistic

graphical dependency models to identify hidden and useful

information from large databases [2], [3].

Predictive data mining is a search for very strong patterns in

big data that can generalize to accurate future decisions [2].

Data mining refers to extracting or “mining” knowledge from

large amounts of data [5]. Keogh et al. [8] used piecewise

linear approximation representation of time series for

clustering, classification and association rule mining of time

series data. They have examined the first extensive review and

empirical comparison of time series segmentation algorithms

from a data mining perspective. They have developed an

efficient sequential pattern for identifying frequent temporal

patterns. Faloutus et al. [14] have developed efficient time

series similarity search methods namely subsequence

matching. Povinelli et al. [15] have proposed a new signal

analysis and classification method based on reconstructed

phase space. They have used statistical method to estimate of

phase space. Bayesian likelihood and artificial neural network

World Academy of Science, Engineering and Technology 51 2009

92

have been used for classification and compared two

techniques. Feng et al. [2] have proposed fuzzy set and the

Gaussian shaped membership function to define temporal

patterns in time delay embedding phase space. The resulting

objective function represents not only the overall value of the

event function, but also the weight of the vector in temporal

pattern cluster to which it contributes. Gradient descent

optimization method has been used to search temporal

patterns. Furuta et al. [16] have applied soft computing

techniques for earthquake acceleration data. Applicability and

efficiency of chaos prediction and neural network method have

been compared and examined using the time series data of

earthquake acceleration.

III. PROPOSED PREDICTION ALGORITHM BASED ON FUZZY

LOGIC USING TIME SERIES DATA MINING

Time series data mining method establishes a method that

uncovers hidden patterns in a time series. Methods based on

the time series data mining technique are able to successfully

characterize and predict complex, non-periodic, irregular and

chaotic time series [1]. The time series data mining methods

overcome limitations of traditional time series analysis

techniques by adapting data mining concepts for analyzing

time series. Method used here is based on time delay

embedding. If the embedding is performed correctly, the

theorems involved guarantee that the reconstructed dynamics

is topologically identical to the true dynamics of the system,

and therefore that the dynamical invariants are also identical

[17]. The time series which is applied data mining should be

examined in a phase space. This is an extremely powerful

correspondence; it implies that conclusions drawn from the

embedded or reconstruction-space dynamics are also true of

real dynamics. Time delay embedding τ and embedding

dimension b are important parameters in constituting a

reconstructed phase space or briefly called phase space.

The phase space of a time series given in equation (1) is a

matrix whose elements are time lagged version of the original

time series data as follows:

(

(

(

(

(

(

(

(

(

(

¸

(

¸

=

− − − −

+ − +

+ − +

τ τ

τ τ

τ τ

) 1 ( ) 2 (

2 2 ) 1 ( 2

1 1 ) 1 ( 1

...

.

.

.

.

.

.

.

.

...

...

b N b N N

b

b

x x x

x x x

x x x

X

(6)

Where, x

N

is the original time series data, N is number of

observations; b is embedding dimension and τ is time delay. It

is important to point out that each row vector of X is a single

point in the phase space. Here, the time lag τ is determined

using first minimum of mutual information technique or

autocorrelation function and the embedding dimension b can

be estimated using false nearest neighbor method.

Block diagram of proposed prediction method based on fuzzy

logic is given in Fig. 1. According to this diagram, before

target information is extracted from time series all data must

be preprocessed and initial parameters are selected. In this

stage, the embedding dimension b and time delay embedding

τ is determined using false nearest neighbor and mutual

information techniques, respectively. Then time series must be

transformed to the phase space. After event characterization

function g(X

t

) is defined, Euclidian distance from initial points

of center points v of temporal cluster and cluster radius r are

selected as inputs of prediction algorithm.

Fig. 1 The prediction algorithm scheme based on fuzzy logic using

time series data mining

A. Estimation of Time Delay Embedding

It must be selected an appropriate time delay τ for phase

space. If τ is small compared to the internal time scales of the

system, successive elements of the delay vector are strongly

correlated. All vector of X are then clustered around the

diagonal in the R

b

, unless b is very large [4]. The two methods

for estimating τ are the information of autocorrelation and

mutual information. First method is the autocorrelation

functions. For a time series given in equation (1), time delay is

estimated by calculating autocorrelation between

t

X and

τ − t

X as shown in equation (7).

C(

t

X ,

τ − t

X )=

} )] ( {[

) ( ) (

2

2

t t

t t t

X E X E

X E X X E

−

−

−τ

(7)

If we plot values

t

X versus the corresponding values a fixed

lag v earlier

τ − t

X , the autocorrelation C quantifies how these

points are distributed. If they spread out evenly over the plane,

then C is zero. If they tend to crowd along the diagonal, then C

becomes big from zero. The autocorrelation function takes the

estimate of time instance when the autocorrelation function

World Academy of Science, Engineering and Technology 51 2009

93

first crosses zero as the time delay. The mutual information is

computed by:

∑ ∑

−

+

j i

i i

j i

ij ij t

p p p p X

, ,

t

ln 2 ) ( ln ) ( = ) X , M( τ τ

τ

(8)

It is sometimes advocated that one look for the first minimum

of time delayed mutual information. This is the information we

already possess about the value X

t+1

if we know X

t

. The

expression we have to compute is based on Shannon’s entropy.

On the interval explored by the data, we create a histogram for

the probability distribution of the data. Denote by p

i

the

probability that the signal assumes a value inside the ith bin of

histogram, and let ) (τ

ij

p be the probability in bin i and

τ − t

X

is in bin j. The first minimum of mutual information function

marks as the time lag [4].

B. Estimation of the Embedding Dimension

When we start to analyze a nonlinear time series, we must

determine the embedding dimension. One way to estimate the

embedding dimension b is using the nearest neighbor method

[4]. This means that a b-dimensional phase space will have the

same topological properties as the original phase space [2]. A

phase space is a b dimensional metric space into which a time

series is unfolded. Takens [7] showed that if Q is large enough,

the phase space is homeomorphic to the state space that

generated the time series. Taken’s Theorem provides the

theoretical justification for reconstructing state spaces using

time-delay embedding. Taken’s proved that the state space of

an unknown system can be reconstructed. If the embedding is

performed correctly, Takens’ Theorem guarantees that the

reconstructed dynamics are topologically identical to the true

dynamics of the system. The algorithm of finding false nearest

neighbor is described as follows:

For each data point which is denoted by

b

i

X in R

b

, its

nearest neighbor is found by

|| || min arg

b

j

b

i

X

b

j

X X X

b

j

− =

, (9)

Where || ||

b

j

b

i

X X − is the Euclidean distance between the two

points. If the embedding dimension is changed from b to b+m,

distance between the two points is calculated by:

2

2 2

|| ||

|| || || ||

b

j

b

i

b

j

b

i

m b

j

m b

i

i

X X

X X X X

r

−

− − −

=

+ +

(10)

If

i

r exceed a given threshold v, then

b

i

X is marked as having

a false nearest neighbor. Therefore, the embedding dimension

is determined by the fact that the number of data points

i

r >v is

zero in R

b

. After time delay embedding τ and embedding

dimension b is determined, the phase space matrix is

constituted. Another preparation stage in the time series data

mining is determined the event characterization function which

changes according to the application. In this paper, because

our aim is the prediction of an earthquake from a synthetic

seismic a time series, the event characterization function is

selected as given in equation (4). To find the important event,

the target function must also be determined. Such a target

function is described in equation (11).

¦

¦

¹

¦

¦

´

¦

+ −

≥ < −

=

∑

∑

=

=

others g

N

M

g X g

M

N M if r v x X g

M

r v f

M

i

i

i

M

i

i

, ) ) (

1

((

|| || , ) (

1

) , (

0 0

1

1

β

β (11)

Where g

0

is the smallest event value, N is total number of data,

β is rate of the minimum cluster size, r is cluster radius, v is

center points of cluster, and M is total number of data inside

cluster. Disadvantage of this cluster is that it uses a crisp

clustering. Radius of the cluster is crisp and each point in the

cluster is own same membership degree. Alternatively, the

temporal pattern cluster can be defined as fuzzy set:

)} | ) ( , {(

b

t t F t

R X X X F ∈ = µ (12)

Where ) X (

t F

µ the membership is a function indicating the

degree of X

t

in F and is computed as in equation (12).

According to this new temporal pattern cluster, the target

function is changed as equation (14).

2

2

|| ||

) (

r

X v

t F

t

e X

−

−

= µ (13)

¦

¹

¦

´

¦

¦

)

¦

`

¹

=

∑

=

N

i

i i F

X g X r v f

1

) ( ) ( max )} , ( max{ µ (14)

Since the fuzzy logic theory was described by Prof. L. Zadeh,

it has been used many of field of engineering [18]. Fuzzy logic

is a problem-solving control system methodology that lends

itself to implementation in systems ranging from simple, small,

embedded micro-controllers to large, networked, multi-

channel PC or workstation-based data acquisition and control

systems. It can be implemented in hardware, software, or a

combination of both. It provides a simple way to arrive at a

definite conclusion based upon vague, ambiguous, imprecise,

noisy, or missing input information. Fuzzy logic’s approach to

control problems mimics how a person would make decisions,

only much faster. Generally, fuzzy logic is quite effective

method in systems that the mathematical model is unknown or

cannot be established. The problem solution in fuzzy logic is

three steps: fuzzifier, inference mechanism, and defuzzifier.

First, the parameters of the system to be examined are

measured. Then the inputs are exposed to fuzzifier.

Knowledge base to be obtained fuzzy rules extracts fuzzy

values in order to justification of the inspected system. In last

step, these fuzzy output values are made certain with one of

defuzzifier methods and justification of the system is made.

World Academy of Science, Engineering and Technology 51 2009

94

IV. APPLICATION RESULTS

A. The Seismic Synthetic Earthquake Example

The prediction algorithm based on fuzzy logic is applied to

a synthetic seismic time series. Matlab package program is

used in order to estimate an earthquake degree from synthetic

seismic time series. Graphics of synthetic seismic time series

is given in Fig. 2.

Fig. 2 Synthetic seismic time series

After synthetic seismic time series is obtained, embedding

dimension and time delay factor of phase space must be

determined. In this study, embedding dimension is selected as

two and time delay factor is selected as one. According to

these parameters, constructed phase space is given in Fig. 3.

Every pair of adjacent observations in the original time series

forms a single point in this phase space. The determined radius

and center points of temporal pattern cluster that predict the

earthquake one step ago is our target in this study using time

series data mining. Augmented phase space is given in Fig. 4.

Our aim is find the temporal pattern that will show to be an

earthquake. The augmented phase space is a three dimensional

phase space with (x

t

,x

t-1

,x

t+1

) dimensions.

Fig. 3 The phase space of synthetic seismic time series

Fig. 4 Augmented phase space of synthetic seismic time series

The first input of the fuzzy logic is radius of the temporal

pattern cluster and the second input is Euclidian distance of

cluster center points from initial points. The output of fuzzy is

shown the size of the earthquake. The membership of fuzzy

logic is given in Fig. 5.

a) Membership functions for r b) Membership functions for v

c) Membership functions of output D

Fig. 5 Input and output membership functions of fuzzy logic for

synthetic time series phase space

The constituted rule base according to these membership

functions is given in Fig. 6. The output graphic of prediction

method based on fuzzy logic and the application results are

given in Fig. 7 and Table I, respectively.

Fig. 6 Rule base of prediction algorithm based on fuzzy logic

World Academy of Science, Engineering and Technology 51 2009

95

Fig. 7 The earthquake prediction graphics according to fuzzy input

values

TABLE I

THE APPLICATION RESULTS OF PREDICTION ALGORITHM BASED ON FUZZY

LOGIC

v r D

0.49 0.35 5.5

0.28 3.83 2.25

3.66 1.23 3.09

6.08 0.50 0.832

1.41 5.70 2.47

Earthquake size according to radius and Euclidian distance

from initial points of center points of temporal pattern cluster

is given in Fig. 7. If v is big and r is small, then the degree of

earthquake is small. If v is small and r is small, then the

earthquake is big. If the center points of the temporal pattern

cluster is 0.5 and 1.3 and radius of the temporal pattern cluster

is 1.3, then the earthquake is maximum shown as in Fig 7.

Therefore, Fuzzy logic could predict v and r parameters which

will predict the earthquake one step ago. One of the

advantages of this method is to use data mining for analysis

of time series and such a system is modeled by fuzzy logic

without too many complex mathematical equations. As shown

in Table I, the optimum values of v and r is values of first row

and where size of earthquake is 5.5. These parameters predict

an earthquake one step after.

B. Lorenz Series Analysis

A well known nonlinear time series-Lorenz series is

generated by equation (15), with three state variables(x, y, z):

¦

¦

¦

¹

¦

¦

¦

´

¦

− =

− − =

− =

bz xy

dt

dz

xz y rx

dt

dy

x y

dt

dx

) ( σ

(15)

Where r , σ and b are model parameters of equation 14,

known as Prandtl number, Rayleigh number, and geometric

factor respectively, see [6] for details.

Fig. 8 plots the Lorenz series with 0 . 10 = σ , r=28.0 and

b=8/3. In this application the x component is chosen as

observed time series with 0.01 step size.

Fig. 8 x(t) component of Lorenz series

Time delay τ was determined as τ =1.2 s and the

embedding dimension b=3 by mutual information and false

nearest neighbor method, respectively. The estimated mutual

information for x(t) component of Lorenz series is shown in

Fig. 9 and the embedding dimension is shown in Fig. 10.

Fig. 9 Mutual information of x(t) component of Lorenz series

Fig. 10 Minimum Embedding dimension

World Academy of Science, Engineering and Technology 51 2009

96

The phase space of x(t) has three embedding dimension. 3D

dimensional phase space is as shown in Fig. 11.

Fig. 11 Phase space of x(t)

The importance occurrence or event for this series was

defined as to when the magnitude of the predicted next point is

greater than 15.

0 15 ) (

1

> − =

+ t t

x x g (16)

The goal of our approach is not identify all event points in the

Lorenz time series. The algorithm takes an approach to

identify those temporal patterns that are likely predictive of the

events. The target function f(v,r) was used to extract the

fuzzy rules from phase spaces. The membership function of

fuzzy logic is given in Fig. 12.

a) Input membership functions for v

b) Input membership function for z

A

axis

NB

NS Z PS

-20 -15 -10 -5 0 5 10

0.5

1

15 20

PB

c) Output membership function for T

Fig. 12 Input and output membership functions of fuzzy logic for

lorenz time series phase space

The rule base for Lorenz time series prediction is given in

Fig. 13. The output surface of fuzzy is demonstrated in Fig.

14.

Fig. 13 Rule base of prediction algorithm based on fuzzy logic fro

Lorenz time series phase space

Fig. 14 The event prediction graphics according to fuzzy input values

In Fig. 14, the event is predicted when distance of cluster

center points to initial points is big and the axis Z

A

is positive

small.

V. CONCLUSION

Time series data mining is one of important issues which

uses for time series analysis and extract hidden, useful and

interesting patterns from them. So this method uses data

mining for analyzing time series, it yields a new approach to

World Academy of Science, Engineering and Technology 51 2009

97

time series analysis. When data mining whose main aim is

extract useful, interesting and hidden patterns from large data

sets is used in time series data, an earthquake prediction and

the sudden falling of stock price is estimated one step before.

One of advantage our approach, it detect many events.

Membership functions and fuzzy rules change based on

problem. Gaussian shaped fuzzy membership function is used

in order to determine temporal pattern clusters without

dependent of applications. Thus the temporal pattern can be

defined at a smooth region. Genetic search algorithm catches

local minima and yield to heavy computing load. The fuzzy

temporal pattern cluster represents temporal patterns more

realistically in the time delay embedding.

There is a potential for further development. Although the

Gaussian shaped membership function is employed to

represent the temporal pattern and objective function, other

different functions could be applied. Also, some fuzzy

optimization algorithm could be applied to different time

series data.

REFERENCES

[1] R. J. Povinelli, “Time Series Data Mining: Identifying Temporal

Patterns for Characterization and Prediction of Time Series Events”,

Ph.D. Dissertation, Marquette University, 180 p., 1999.

[2] X. Feng and H. Huang, “A Fuzzy-Set-Based Reconstructed Phase Space

Method for Identification of Temporal Patterns in Complex Time

Series”, IEEE Trans. on Knowledge and Data Engineering, Vol. 17, No.

5, pp. 601-613, 2005.

[3] R. J. Povinelli and X. Feng, “Data Mining of Multiple Nonstationary

Time Series, in Proceedings of Artificial Neural Networks in

Engineering”, St. Louis, Missouri, pp. 511-516, 2005.

[4] H. Kantz and T. Schreiber, Nonlinear Time Series Analysis, Cambridge:

Cambridge University Press, 388 p., 1997.

[5] J. Han and M. Kamber , Data Mining: Concepts and Techniques, San

Francisco: Academic Press, 800 p.,2005

[6] S.M. Weiss and N. Indurkhya, Predictive Data Mining: A practical

Guide. San Fransisco: Morgan Kaufmann., 228 p., 1998.

[7] F. Takens, “Detecting Strange Attractors in Turbulence, presented at the

Dynamical Systems and Turbulence”, Warwick, U.K, 1981.

[8] E. Keogh and P. Smyth, “A Probabilistic Approach to Fast Pattern

Matching in Time series Databases”, in proc. Third Int’l Conf.

Knowledge Discovery and Data Mining, 1997.

[9] E. Keogh, “A fast and Robust Method for Pattern Matching in Time

Series Databases”, in proc. Ninth Int’l conf. Tools with Artificial

Intelligence (TAI’97), pp. 578-584, 1997.

[10] Y. S. Moon and K. Young Whang, W. Kee Loh, “Efficient time-series

subsequence matching using duality in constructing windows”,

Information Systems, Vol. 26, pp 279-293, 2001.

[11] R. Povinelli, J. Bangura, N. Demerdash, and R. Brown, “Diagnostics of

Bar and End-Ring Connector Breakage Faults in Polyphase Induction

Motors Through a Novel Dual Track of Time-Series Data Mining and

Time-Stepping Coupled FE-State Space Modeling,” IEEE Transactions

on Energy Conversion, Vol. 17, No. 1, pp. 39-46, 2002.

[12] C., Yeh, R. J. Povinelli, B. Mirafzal and N. A. O. Demerdash,

“Diagnosis of Stator Winding Inter-Turn Shorts in Induction Motors Fed

by PWM-Inverter Drive Systems Using a Time-Series Data Mining

Technique”, in proceedings of Power System Technology, pp.891-896,

2004.

[13] M. T. Hagan, H. B. Demuth and M. H. Bale, Neural Network Design,

Boston: PWS Publishing, 736 p., 1996.

[14] C. Faloutsos, M. Ranganathan, and Y. Manolopoulos, “Fast

Subsequence Matching in Time-Series Databases” Proc. Sigmod Record

(ACM Special Interest Group on Management of Data Conf.), pp. 419-

429, 1994.

[15] R. J., Povinelli and X., Feng, “A New Temporal Pattern Identification

Method for Characterization and Prediction of Complex Time Series

Events”, IEEE Trans. On Knowledge and Data Engineering, Vol. 15,

No. 2, pp. 339-352, 2003.

[16] H. Furuta and Y. Nomura “Time Series Prediction of Earthquake Input

by using Soft Computing”, proc of the Fourth International Symposium

on Uncertainty Modeling and Analysis (ISUMA’03), pp. 351-356, 2003

[17] E. Bradley, Analysis of Time Series- An Introduction to Intelligent Data

Analysis, M. Berthold and D. Hand, eds., pp. 167-194, New York:

Springer, 1999.

[18] L., Zadeh, “Fuzzy Sets”, Information and Control, Vol. 8, pp. 338-353,

1965.

I. Aydin (M’2006) was born in Elazig, Turkey, in 1981. He received the BS

and MS degree in computer engineering from University of Firat, Elazig,

Turkey, in 2001 and 2006, respectively. He is currently working toward PhD

degree in Department of Electrical-Electronic Engineering, Firat University,

Elazig, Turkey. From 2001 to 2006, he was with the University of Firat as a

lecturer. He is working as a lecturer in Erzincan University. His research

interests are soft computing, fault diagnosis, signal processing and data

mining.

M. Karakose was born in Elazig, Turkey, 1976. He received the BS and MS

degree in electrical–electronic engineering and computer engineering

departments from Firat University, Elazig, Turkey in 1998 and 2001,

respectively. He received PhD degree in electrical-electronic engineering from

University of Firat, Elazig, Turkey, 2005. He is currently an Assistant

professor with computer engineering department, firat university, Elazig,

Turkey. His main research interests are in fuzzy control and digital control of

variable-speed ac drives.

E. Akin was born in Erzincan, Turkey, 1963. He received the BS and MS

degrees in electrical engineering from Firat University, Elazig, Turkey in 1984

and 1987 respectively, and the PhD degree in the area of ac drives from Firat

University, Elazig, Turkey in 1994. He was with the Department of Electrical

Engineering, Firat University, first as Assistant professor and then as a

Associate profersor of electrical machines. He is currently a Full Professor of

computer engineering with university of firat, Elazig, Turkey. His main

research interests are in power electronics, digital control of variable-speed ac

drives, fuzzy control and soft computing techniques.

World Academy of Science, Engineering and Technology 51 2009

98

x t ) ∈ R b (5) Where τ =1. Time series should be examined in a phase space in order to get interesting pattern from it. Others who have applied data mining concepts to finding patterns in time series include Keogh and Smyth [8]. he did different applications such as earthquake prediction. sharp fall of stock price. and classifications models to identify hidden and useful information in large databases. APPLICATIONS OF TIME SERIES DATA MINING Data mining is a technique of discovering useful pattern in data that are hidden and unknown in normal circumstance.xt-3. [8] used piecewise linear approximation representation of time series for clustering. In his study. each of point in temporal pattern cluster belong to cluster with a specific membership degree. Gradient descent optimization algorithm is chosen in order to find optimum temporal pattern cluster. [3]. The optimization results were often inconsistent. he transformed a time series to phase space. Data mining refers to extracting or “mining” knowledge from large amounts of data [5].k is time delay embedding which allows the phase space to be spanned over nonconsecutive timelagged instances. Data mining is the process of discovering useful pattern in data that are hidden and important pattern. In this paper. with the input of today as well as four previous day’s price change. He selected different event characterization functions for each application. classification. It uses methods such as clustering. The second disadvantage is the computational complexity and stability. association rules.…. Their approach uses a dynamic programming method for aligning the time series and a predefined set of templates. [12]. Optimal temporal pattern cluster with radius r and temporal pattern center v was found by using genetic search algorithm. One of these methods is classification and fault diagnosis in induction motors [11]. This method uses the fuzzy logic for earthquake and Lorenz series prediction. τ =3 and b=3. Conclusion and future study are discussed in section IV. He used the genetic algorithm to discover interesting patterns in a time series by data mining. event prediction doesn’t include complex mathematical equations and the prediction is simplified. database and machine learning [5]. Engineering and Technology 51 2009 algorithm techniques”. [10] proposed a new time subsequence matching method. We introduce and discuss our application results in section III. and complex time series. [14] have developed efficient time series similarity search methods namely subsequence matching.. statistics and database design [2]. proposed time series data mining method is based on fuzzy logic. 2.. Despite the novelty of this method. The paper is organized as follows: Section II provides brief review of time series data mining using phase space strategy. Feng et al. [9]. classification and association rule mining of time series data.. The main goal in time series data mining is to use time delay embedding and phase space based on Taken theorem [7]. Because user experience about event prediction is shown in phase space and augmented phase space easily. Time delay embedding τ and embedding dimension b is chosen by mutual information and false nearest neighbor method.. including statistics. Bayesian likelihood and artificial neural network 92 . Povinelli [1] was suggested an application in order to event prediction. They have used statistical method to estimate of phase space. association rule mining and probabilistic graphical dependency models to identify hidden and useful information from large databases [2]. The prediction is flexible due to property of fuzzy logic method. Keogh et al. It uses techniques such as clustering. Elements of Xt form time lagged shape of original time series with embedding dimension b and time delay factor τ . They have developed an efficient sequential pattern for identifying frequent temporal patterns. Povinelli et al.xt). Therefore. It comes from several fields. x t −(b − 2)τ . II.. This method is successfully applied non-periodical. The goal was predicted next day’s closing price change. nonlinear. This restriction made it hard to adjust to the situation when there is noise in phase space and often generated high degree false-positive prediction. So the event prediction is predicted accurately. X t = ( x t −(b −1)τ . For instance. At first step.World Academy of Science. several problems limited its applications and must to be addressed. First... The choices of embedding dimension b and time delay τ were always based on user’s experience with trial-and-error. Predictive data mining is a search for very strong patterns in big data that can generalize to accurate future decisions [2]. respectively. An efficient two step optimization strategy was proposed to search the optimal temporal cluster in the phase space. The phase space of a time series in equation (1) is generated by using time delay embedding and embedding dimension in order to get interesting patterns. The heavy load which the genetic algorithm yields is removed. The temporal patterns representing the characteristic of time series are used to prediction of the important events. The phase space of the time series in equation (1) according to time delay factor τ and embedding dimension b is presented in equation (5). easily. Multilayer feedforward neural network was performed on the D-stock prediction [13]. They have examined the first extensive review and empirical comparison of time series segmentation algorithms from a data mining perspective. Time series data mining has been used in a lot of fields such as clustering and event prediction in literature. Because temporal pattern cluster is selected as a fuzzy set. [15] have proposed a new signal analysis and classification method based on reconstructed phase space. [2] suggested new time series data mining for identifying temporal patterns. X is a matrix whose row vector is a point in the phase space. temporal pattern cluster in this study was defined by a rigid region with fixed center and radius. The genetic algorithm yields heavy computing load.. In his study. temporal pattern cluster was chosen as a fuzzy set with Gaussian shaped membership function. Faloutus et al. This transformation preserves the nonlinear dynamics of the original time series and the temporal patterns are represented by data points Xt in the phase space [2]. Moon et al. the vector is shown by Xt=(xt-6. Data mining consist of machine learning.. They used the dual match for time series similarity. the fuzzy rules is also determined.

non-periodic. unless b is very large [4]. Time delay embedding τ and embedding dimension b are important parameters in constituting a reconstructed phase space or briefly called phase space.. irregular and chaotic time series [1]. the embedding dimension b and time delay embedding τ is determined using false nearest neighbor and mutual information techniques. but also the weight of the vector in temporal pattern cluster to which it contributes. b is embedding dimension and τ is time delay. This is an extremely powerful correspondence. . The two methods for estimating τ are the information of autocorrelation and mutual information. then C becomes big from zero. . x N − (b −1)τ x1 target information is extracted from time series all data must be preprocessed and initial parameters are selected. Block diagram of proposed prediction method based on fuzzy logic is given in Fig. then C is zero. 1 The prediction algorithm scheme based on fuzzy logic using time series data mining (6) A. According to this diagram. If the embedding is performed correctly.World Academy of Science. Fig. [2] have proposed fuzzy set and the Gaussian shaped membership function to define temporal patterns in time delay embedding phase space. X t −τ )= Where. Estimation of Time Delay Embedding It must be selected an appropriate time delay τ for phase space. After event characterization function g(Xt) is defined. before E ( X t X t −τ ) − E ( X t ) 2 E{[ X t − E ( X t )] 2 } (7) If we plot values X t versus the corresponding values a fixed lag v earlier X t −τ .. III. [16] have applied soft computing techniques for earthquake acceleration data. . 1. . . Engineering and Technology 51 2009 have been used for classification and compared two techniques. If they tend to crowd along the diagonal. The phase space of a time series given in equation (1) is a matrix whose elements are time lagged version of the original time series data as follows: x1+ (b −1)τ x 2 + (b −1)τ X = . The resulting objective function represents not only the overall value of the event function. C( X t . Feng et al. In this stage. Gradient descent optimization method has been used to search temporal patterns. successive elements of the delay vector are strongly correlated. For a time series given in equation (1). . The time series data mining methods overcome limitations of traditional time series analysis techniques by adapting data mining concepts for analyzing time series. x1+τ x2 +τ . time delay is estimated by calculating autocorrelation between X t and X t − τ as shown in equation (7). Euclidian distance from initial points of center points v of temporal cluster and cluster radius r are selected as inputs of prediction algorithm. If τ is small compared to the internal time scales of the system. The time series which is applied data mining should be examined in a phase space.. Methods based on the time series data mining technique are able to successfully characterize and predict complex.. . First method is the autocorrelation functions. N is number of observations. The autocorrelation function takes the estimate of time instance when the autocorrelation function 93 . the autocorrelation C quantifies how these points are distributed. If they spread out evenly over the plane. x N − (b − 2)τ x2 . respectively. the theorems involved guarantee that the reconstructed dynamics is topologically identical to the true dynamics of the system.. It is important to point out that each row vector of X is a single point in the phase space. Then time series must be transformed to the phase space. Here. Method used here is based on time delay embedding. All vector of X are then clustered around the diagonal in the Rb.. Furuta et al. PROPOSED PREDICTION ALGORITHM BASED ON FUZZY LOGIC USING TIME SERIES DATA MINING Time series data mining method establishes a method that uncovers hidden patterns in a time series. it implies that conclusions drawn from the embedded or reconstruction-space dynamics are also true of real dynamics. xN . and therefore that the dynamical invariants are also identical [17]. the time lag τ is determined using first minimum of mutual information technique or autocorrelation function and the embedding dimension b can be estimated using false nearest neighbor method. xN is the original time series data. Applicability and efficiency of chaos prediction and neural network method have been compared and examined using the time series data of earthquake acceleration.

its nearest neighbor is found by X b = arg min || X ib − X b || . The first minimum of mutual information function marks as the time lag [4]. it has been used many of field of engineering [18]. Engineering and Technology 51 2009 first crosses zero as the time delay.World Academy of Science. j i. Knowledge base to be obtained fuzzy rules extracts fuzzy values in order to justification of the inspected system. and let pij (τ ) be the probability in bin i and X t −τ is in bin j. Disadvantage of this cluster is that it uses a crisp clustering. inference mechanism. Taken’s proved that the state space of an unknown system can be reconstructed. the embedding dimension is determined by the fact that the number of data points ri >v is zero in R b . The algorithm of finding false nearest neighbor is described as follows: For each data point which is denoted by X ib in R b . because our aim is the prediction of an earthquake from a synthetic seismic a time series. and defuzzifier. we must determine the embedding dimension. It provides a simple way to arrive at a definite conclusion based upon vague. only much faster. the phase space is homeomorphic to the state space that generated the time series. Generally. 94 . v is center points of cluster. embedded micro-controllers to large. fuzzy logic is quite effective method in systems that the mathematical model is unknown or cannot be established. we create a histogram for the probability distribution of the data. A phase space is a b dimensional metric space into which a time series is unfolded. The expression we have to compute is based on Shannon’s entropy. If the embedding is performed correctly. Estimation of the Embedding Dimension When we start to analyze a nonlinear time series. The mutual information is computed by: M( X t . the parameters of the system to be examined are measured. Takens [7] showed that if Q is large enough. Denote by pi the probability that the signal assumes a value inside the ith bin of histogram. Therefore. Fuzzy logic’s approach to control problems mimics how a person would make decisions. B. or missing input information. i =1 βN M Where g0 is the smallest event value. r )} = max µ F ( X i ) g ( X i ) i =1 ∑ (14) (9) Where || X ib − X b || is the Euclidean distance between the two j points. In last step. µ F ( X t ) | X t ∈ R b )} (12) Where µ F (X t ) the membership is a function indicating the degree of Xt in F and is computed as in equation (12). Alternatively. j i ln pi (8) changes according to the application. Such a target function is described in equation (11). If the embedding dimension is changed from b to b+m. µF (X t ) = e − ||v − X t || 2r 2 (13) N max{ f (v. the target function must also be determined. To find the important event. N is total number of data. j j Xb j ∑ M i =1 g ( X i ). Radius of the cluster is crisp and each point in the cluster is own same membership degree. the target function is changed as equation (14). or a combination of both. This is the information we already possess about the value Xt+1 if we know Xt. small. imprecise. Takens’ Theorem guarantees that the reconstructed dynamics are topologically identical to the true dynamics of the system. the event characterization function is selected as given in equation (4). One way to estimate the embedding dimension b is using the nearest neighbor method [4]. Taken’s Theorem provides the theoretical justification for reconstructing state spaces using time-delay embedding. According to this new temporal pattern cluster. After time delay embedding τ and embedding dimension b is determined. distance between the two points is calculated by: ri = || X ib + m − X b + m || 2 − || X ib − X b || 2 j j || X ib − X b || 2 j (10) If ri exceed a given threshold v. On the interval explored by the data. noisy. Zadeh. β is rate of the minimum cluster size. It can be implemented in hardware. L. Then the inputs are exposed to fuzzifier. ambiguous. these fuzzy output values are made certain with one of defuzzifier methods and justification of the system is made. f (v. r ) = 1 (( M 1 M It is sometimes advocated that one look for the first minimum of time delayed mutual information. then X ib is marked as having a false nearest neighbor. the phase space matrix is constituted. Another preparation stage in the time series data mining is determined the event characterization function which Since the fuzzy logic theory was described by Prof. The problem solution in fuzzy logic is three steps: fuzzifier. the temporal pattern cluster can be defined as fuzzy set: F = {( X t . X t +τ ) = ∑ p (τ ) ln p (τ ) − 2∑ p ij ij i. First. || xi − v ||< r if M ≥ β N others (11) ∑ M g ( X i ) − g0 ) + g0 . r is cluster radius. software. Fuzzy logic is a problem-solving control system methodology that lends itself to implementation in systems ranging from simple. In this paper. and M is total number of data inside cluster. multichannel PC or workstation-based data acquisition and control systems. This means that a b-dimensional phase space will have the same topological properties as the original phase space [2]. networked.

3 The phase space of synthetic seismic time series Fig. embedding dimension and time delay factor of phase space must be determined. Graphics of synthetic seismic time series is given in Fig. The determined radius and center points of temporal pattern cluster that predict the earthquake one step ago is our target in this study using time series data mining. In this study. Augmented phase space is given in Fig. respectively. Every pair of adjacent observations in the original time series forms a single point in this phase space. constructed phase space is given in Fig.xt+1) dimensions. 4 Augmented phase space of synthetic seismic time series The first input of the fuzzy logic is radius of the temporal pattern cluster and the second input is Euclidian distance of cluster center points from initial points. 3.World Academy of Science. The augmented phase space is a three dimensional phase space with (xt. The Seismic Synthetic Earthquake Example The prediction algorithm based on fuzzy logic is applied to a synthetic seismic time series. a) Membership functions for r b) Membership functions for v c) Membership functions of output D Fig. Matlab package program is used in order to estimate an earthquake degree from synthetic seismic time series. The membership of fuzzy logic is given in Fig. 5 Input and output membership functions of fuzzy logic for synthetic time series phase space The constituted rule base according to these membership functions is given in Fig. The output graphic of prediction method based on fuzzy logic and the application results are given in Fig. 4. Engineering and Technology 51 2009 IV. 2 Synthetic seismic time series After synthetic seismic time series is obtained. Our aim is find the temporal pattern that will show to be an earthquake. According to these parameters. Fig. 6.xt-1. 5. Fig. Fig. 7 and Table I. APPLICATION RESULTS A. 2. The output of fuzzy is shown the size of the earthquake. embedding dimension is selected as two and time delay factor is selected as one. 6 Rule base of prediction algorithm based on fuzzy logic 95 .

and geometric factor respectively. 8 x(t) component of Lorenz series Time delay τ was determined as τ =1. Fuzzy logic could predict v and r parameters which will predict the earthquake one step ago. B. If v is big and r is small.25 3.2 s and the embedding dimension b=3 by mutual information and false nearest neighbor method.01 step size. z): Fig. Therefore. 9 Mutual information of x(t) component of Lorenz series dx dt = σ ( y − x) dy = rx − y − xz dt dz = xy − bz dt (15) Fig. Fig. As shown in Table I.09 0. then the earthquake is maximum shown as in Fig 7.0 .70 D 5. Earthquake size according to radius and Euclidian distance from initial points of center points of temporal pattern cluster is given in Fig. 9 and the embedding dimension is shown in Fig.08 1. Rayleigh number.50 5. One of the advantages of this method is to use data mining for analysis of time series and such a system is modeled by fuzzy logic without too many complex mathematical equations. known as Prandtl number. the optimum values of v and r is values of first row and where size of earthquake is 5. see [6] for details.41 r 0. with three state variables(x.66 6.0 and b=8/3. 8 plots the Lorenz series with σ = 10. These parameters predict an earthquake one step after.832 2. 7 The earthquake prediction graphics according to fuzzy input values TABLE I THE APPLICATION RESULTS OF PREDICTION ALGORITHM BASED ON FUZZY LOGIC v 0. Lorenz Series Analysis A well known nonlinear time series-Lorenz series is generated by equation (15).28 3.23 0. then the earthquake is big.5.3 and radius of the temporal pattern cluster is 1. If v is small and r is small. 10. Engineering and Technology 51 2009 Where σ . 10 Minimum Embedding dimension 96 . respectively. r and b are model parameters of equation 14. 7.47 Fig. The estimated mutual information for x(t) component of Lorenz series is shown in Fig. y.3.83 1. Fig. r=28.World Academy of Science.35 3. then the degree of earthquake is small.5 2.5 and 1.49 0. If the center points of the temporal pattern cluster is 0. In this application the x component is chosen as observed time series with 0.

11 Phase space of x(t) The importance occurrence or event for this series was defined as to when the magnitude of the predicted next point is greater than 15.5 -20 -15 -10 -5 0 5 10 15 20 c) Output membership function for T Fig. the event is predicted when distance of cluster center points to initial points is big and the axis ZA is positive small. 14. useful and interesting patterns from them. it yields a new approach to 97 . The membership function of fuzzy logic is given in Fig. 12 Input and output membership functions of fuzzy logic for lorenz time series phase space The rule base for Lorenz time series prediction is given in Fig.World Academy of Science. Fig. 12. 14 The event prediction graphics according to fuzzy input values b) Input membership function for zA axis In Fig. 11. 13. The output surface of fuzzy is demonstrated in Fig. V.r) was used to extract the fuzzy rules from phase spaces. 3D dimensional phase space is as shown in Fig. The algorithm takes an approach to identify those temporal patterns that are likely predictive of the events. g ( xt ) = xt +1 − 15 > 0 (16) The goal of our approach is not identify all event points in the Lorenz time series. 1 NB NS Z PS PB 0. The target function f(v. 14. Engineering and Technology 51 2009 The phase space of x(t) has three embedding dimension. CONCLUSION Time series data mining is one of important issues which uses for time series analysis and extract hidden. So this method uses data mining for analyzing time series. 13 Rule base of prediction algorithm based on fuzzy logic fro Lorenz time series phase space a) Input membership functions for v Fig. Fig.

it detect many events. Turkey. J.. Berthold and D. “Detecting Strange Attractors in Turbulence. respectively.” IEEE Transactions on Energy Conversion. 8. Turkey. The fuzzy temporal pattern cluster represents temporal patterns more realistically in the time delay embedding. 1963.D. 2. 1976. Takens. He received PhD degree in electrical-electronic engineering from University of Firat. No. Turkey. 167-194. M. “Fuzzy Sets”. pp. pp. F. 2005. Turkey. “Fast Subsequence Matching in Time-Series Databases” Proc. Turkey in 1994. 1999. Firat University. San Francisco: Academic Press. Feng. There is a potential for further development. I. Thus the temporal pattern can be defined at a smooth region. “Diagnosis of Stator Winding Inter-Turn Shorts in Induction Motors Fed by PWM-Inverter Drive Systems Using a Time-Series Data Mining Technique”.. Demerdash.. Analysis of Time Series. pp. J. T. Karakose was born in Elazig.. He was with the Department of Electrical Engineering. and R.. Keogh. Engineering and Technology 51 2009 time series analysis. Yeh. “Diagnostics of Bar and End-Ring Connector Breakage Faults in Polyphase Induction Motors Through a Novel Dual Track of Time-Series Data Mining and Time-Stepping Coupled FE-State Space Modeling. St. R.. Kantz and T. Ninth Int’l conf.2005 S. Elazig. “Data Mining of Multiple Nonstationary Time Series. Faloutsos. Feng. J. X. Keogh and P. Zadeh. in proc. 1997. “A Fuzzy-Set-Based Reconstructed Phase Space Method for Identification of Temporal Patterns in Complex Time Series”. REFERENCES [1] R. pp. 339-352. 1998. His main research interests are in fuzzy control and digital control of variable-speed ac drives.). H. Knowledge Discovery and Data Mining. Povinelli. [18] L. in proc. he was with the University of Firat as a lecturer. and the PhD degree in the area of ac drives from Firat University. Feng and H. He is currently working toward PhD degree in Department of Electrical-Electronic Engineering. Huang. 736 p. A. Turkey. R. Ph. Elazig. 338-353. S. Vol. other different functions could be applied. Hand. Third Int’l Conf. No. Neural Network Design. H. 601-613. 2002. Aydin (M’2006) was born in Elazig. pp. His main research interests are in power electronics. fuzzy control and soft computing techniques. 1981. M. and Y. When data mining whose main aim is extract useful. Manolopoulos. Furuta and Y. Brown. New York: Springer. “Efficient time-series subsequence matching using duality in constructing windows”. [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] 98 . Firat University. Bangura. Missouri. in 2001 and 2006.M. 1994. 351-356. C. Information and Control. digital control of variable-speed ac drives. on Knowledge and Data Engineering. some fuzzy optimization algorithm could be applied to different time series data. Nomura “Time Series Prediction of Earthquake Input by using Soft Computing”. 2003. Mirafzal and N. 15. in Proceedings of Artificial Neural Networks in Engineering”. Kee Loh. Kamber . 26. pp. pp. Turkey in 1998 and 2001. Vol. Vol. J. His research interests are soft computing. 800 p. Indurkhya. 578-584. Povinelli.World Academy of Science. fault diagnosis. N. He is currently an Assistant professor with computer engineering department. Warwick. “Time Series Data Mining: Identifying Temporal Patterns for Characterization and Prediction of Time Series Events”. Povinelli and X. Ranganathan. 180 p. IEEE Trans. 1965. Elazig. 5. Although the Gaussian shaped membership function is employed to represent the temporal pattern and objective function. Elazig. Turkey. No. M. 2004. Data Mining: Concepts and Techniques.K. 1. Bradley. R. One of advantage our approach. pp. 2001. Boston: PWS Publishing. Schreiber. Gaussian shaped fuzzy membership function is used in order to determine temporal pattern clusters without dependent of applications. Han and M. pp 279-293. Sigmod Record (ACM Special Interest Group on Management of Data Conf. Predictive Data Mining: A practical Guide. 1996. Elazig. San Fransisco: Morgan Kaufmann. Akin was born in Erzincan. C. B. He received the BS and MS degree in electrical–electronic engineering and computer engineering departments from Firat University.. Dissertation. Turkey. firat university. He received the BS and MS degree in computer engineering from University of Firat. signal processing and data mining. 388 p.891-896. 228 p. Povinelli and X. On Knowledge and Data Engineering. “A Probabilistic Approach to Fast Pattern Matching in Time series Databases”. Turkey. Also. Louis. pp. Young Whang. He is currently a Full Professor of computer engineering with university of firat. Marquette University. presented at the Dynamical Systems and Turbulence”. He is working as a lecturer in Erzincan University. 419429. W. Vol. 2003 [17] E. Demuth and M. E. O. Information Systems. Smyth... J. respectively. 511-516. Y. M. Vol. 17. Cambridge: Cambridge University Press. first as Assistant professor and then as a Associate profersor of electrical machines. R. Tools with Artificial Intelligence (TAI’97). 1997. Turkey in 1984 and 1987 respectively. Membership functions and fuzzy rules change based on problem.. Nonlinear Time Series Analysis. “A fast and Robust Method for Pattern Matching in Time Series Databases”. “A New Temporal Pattern Identification Method for Characterization and Prediction of Complex Time Series Events”. 2005.An Introduction to Intelligent Data Analysis.. U. Weiss and N. E. Elazig. IEEE Trans. [16] H. interesting and hidden patterns from large data sets is used in time series data. E. B. Elazig. 17. 1999. Bale. Elazig. Moon and K. in proceedings of Power System Technology. eds. proc of the Fourth International Symposium on Uncertainty Modeling and Analysis (ISUMA’03). Hagan. J. He received the BS and MS degrees in electrical engineering from Firat University. H. 39-46. an earthquake prediction and the sudden falling of stock price is estimated one step before. 2005. 1997. Demerdash. Genetic search algorithm catches local minima and yield to heavy computing load. in 1981. From 2001 to 2006. pp. Povinelli.

- Facts on WaPo Storyby Mitt Romney
- The Obama Record: Shipping Jobs & Money To Chinaby Republican National Committee
- As the 2012 Presidential Campaign Heats Up for Mitt Romney, Celestine Publishing Released its Latest Publication, Mormon Beliefs Exposedby PR.com 2
- No Miracle in Massachusetts: Why Governor Romney's Health Care Reform Won't Work, Cato Briefing Paper No. 97by Cato Institute

- A Chance for Every Child
- Dictionary of Literary Terms 2
- Mitt Romney Info
- Scan 001
- Romney’s Housing Plan
- CREW
- Wagstaff v. Romney, et al - Document No. 7
- null
- Facts on WaPo Story
- The Obama Record
- As the 2012 Presidential Campaign Heats Up for Mitt Romney, Celestine Publishing Released its Latest Publication, Mormon Beliefs Exposed
- No Miracle in Massachusetts
- Mitts GOP Plank 50 Bullet Points - @KLSouth
- Romney Santorum Deficit Memo
- Romney
- An American Century—A Strategy to Secure America’s Enduring Interests and Ideals
- ma-policy-2
- An Open Letter About Mitt Romney - RomneyExposed.com
- Romney Bain
- The Romney Scorecard RINO - @KLSouth
- Nightmare Ahead
- Mitt Romney Says Meg Will Bring 'Bright Future' To California
- McCain Notes on Romney 2008
- Women And The Obama Economy [Infographic]
- Believe In America
- Mitt & Ann Romney Tax Return
- The Romney Book
- Romney Energy Plan
- McCain 2008 Oppo File
- Bad English Translations
- Data Mining

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd