Sie sind auf Seite 1von 5

A Simple and Efficient Method for Fault Diagnosis

Using Time Series Data Mining


I. AYDIN', Student Member, IEEE, M. KARAKOSE2, E. AKIN3
'Computer Technology and Programming Education Program
Kemaliye H. A. AKIN Technical Vocational School of Higher Education
Erzincan University, 24600, Kemaliye, Erzincan, Turkey
2, 3Computer Engineering Department
Firat University, 23119, Elazig, Turkey
{ iaydin ,mkarakose, eakinI @firat.edu.tr
Abstract-Early detection and diagnosis of incipient faults is
desirable for online condition evaluation and improved
operational efficiency of induction motors. A classification
technique based on time series data mining is developed to detect
broken rotor bar faults in induction motors. The proposed
algorithm uses only stator phase currents as input without the
need for any other signals. The stator phase currents are
transformed to park's vector components and a new feature
vector is constituted by using these components. The phase space
of constituted feature vector is constructed according to
determined time delay and embedding dimension for each motor
conditions. Each motor condition is separated to two clusters by
using fuzzy c-means clustering algorithm. The center points of
these clusters are saved for test phase. A Gaussian membership
function is used for that a point is the degree of belonging to a
cluster. The current signals of a three phase induction motor are
derived an actual experimental setup. A healthy induction motor
and one, two and three broken rotor bar faults are classified
under four different operation speed. Experimental results show
the strength of the proposed method.

In literature, several monitoring techniques have been


developed for broken rotor bars faults. Many of these
techniques are based on motor current signature analysis.
These techniques use the power spectrum of the stator current
to detect broken rotor bar faults. They are based on the

magntud of certain freque componstofute stao


cets MoTor
currenvt
spectrut
Fast Fourier Transform (FFT) of the current. The left spectral

of line frequency is called the low sideband. The magnitude


of the low sideband is measured and compared to a threshold.
If this comparison is bigger than a threshold, motor has
broken rotor bars. This technique may fail to detect faults

because the sidebands can be masked due to the windowing


process used to compute the power spectrum of current

sina the motor


current siae
re
measuring
motor current and speed. Motor speed can be
[4]

determined via motor current rotor slot harmonics, but this

technique needs number of rotor slots and motor components

information [5]. Bangura at al. [6] have proposed time series


data mining to detect broken rotor bar faults and eccentricity
faults in three phase induction motors. They used torque data
for their fault detection algorithm, but the measuring of this
signal is cost and sensitive. Fault detection algorithm based
on
have [3].
beenMotor
broken
developed
rotorBayesian
bar and classification
currentforsignature
eccentricity faults

Key words: Time series data mining, fuzzy c-means clustering,


broken rotor bar faults, fault detection and diagnosis, induction
motors.

I. INTRODUCTION
Induction motors are the most used type of motors in
industrial applications. The use of induction motors iS
widespread. They are generally used in petrochemical
industries, chemical and domestic appliances industries. In
critical application

such

nuclear

analysis and park vector approach have been used for feature

extraction.

Feature extraction methods using park vector


approach have been extracted to detect open phase and

plants,aerospaceand

applications snduchias nulerspan aerosaen aed.


i
notorsae
militarlityapitbions
tiodrd
ton used.
Reliability must be high
of standard
in these a.lso
applications
[1].
Although these motors are reliable and robust, environmental,
duty and installation issues may combine to accelerate motor
failure far sooner than the designed motor lifetimes [2]. These

critcal

as

voltage unbalance faults [7, 8]. Benbouzza at al [9] proposed


park's
vectorndeprm
broken rotor bar faults using
approach to detect
siuaon
tldt.Afutdagssad
simonion mtd basedton ata. inteligen od as
beengdeve od for on rtoribar and staor
fas
stator faults
deen
ew
feature
vetor
detection [10]. A new feature vector called as envelope is
achieved
from three
stator time
current.
Phase
spacemethod.
of this
is formed
nonlinear
series
usingphase
analysis
vector
Healthy and different faulty conditions of induction motor

[10]lAp

issues~ falt
~ in inuto
fteefut
induction motors.Ifthesefls
are netey
issesect
faulystsgin
at an
detectedcause
early stage,
ninctiom t are comleel
cas

oos

broke down. The early detecton of incipient faults is thus of

have

15 of total faults [3].

cand

classified using Gaussian mixture

models-based

Bayesian classification algorithm.


In this study, broken rotor bar faults in three phase
induction motors are diagnosed using time series data mining

This work was supported by Firat University Scientific Research Unit


(FUBAP). The Project No: FUBAP-1 140.

1 -4244-0743-5/07/$20.OO 2007 IEEE

analysis

and fuzzy classification algorithm. As different from above


fault detection algorithm, our algorithm use only three phase

596

of the motor current. The new time series as feature vector is


extracted from park's vector components. The phase space of
this vector is constructed by selecting appropriate time delay
and embedding dimension. Each phase space is separated
using fuzzy c-means clustering algorithm.

Data acquisition and signal preprocessing step


I
A i p
c
f motor
b
a
c
Park's vector transfonration
Id
Iq

II. PROPOSED FAULT CLASSIFICATION ALGORITHM


Fuzzy classification algorithm based on time series data
mining is proposed to classify broken rotor bar faults in three
phase induction motors. One, two and three broken rotor bar
faults are recognized under four different operation speed.
The proposed method for classification of broken rotor bar
faults needs three phase motor currents. The scheme of
proposed fault classification algorithm is shown in Fig. 1.
Three phase motor currents is shown with a, b and c in Fig. 1.
Park's vector components are achieved from three phase
motor current and they are termed as Id and lq. After the
feature vector I is achieved using park's vector components,
phase space of this vector is constructed and faults are
classified using fuzzy c-means algorithm. The outputs are
shown motor fault conditions.

Construct feature vector


lo

Fault classification step

Reconstrcut
phase space of
featurevector

I-b--c

Imax sin(ot)

T
=-I6I
q 2
-

mx

sin(wt --)
2

B. Time Series Data Mining


Time series data mining is a method that discovers
interesting patterns in a time series. This method is able to
successfully predict complex, non-periodic, irregular and
chaotic time series [11]. The time series data mining methods
eliminate limitations of traditional time series analysis
techniques using data mining concepts for analyzing time

(1)

series.

Under ideal conditions, Park's vector transformation of


balanced three phase currents is shown in (2).
V

algorithm

B1
12

Fig. 1. Fault classification scheme.

Xj

Id =

Fault classification
by fuzzy c-means

13

A. Park's Vector Approach and Feature Vector


Sum of stator phase current is zero in a three phase
induction motor. Park's vector approach is used to achieve a
2D representation of phase currents of a three phase induction
motor. Park's vector approach is given in (1).

Id =A-a - -b-

The time series which is applied data mining should be


examined in a phase space. Time delay embedding x 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 in (3) is
generated using time delay x and embedding dimension b as
shown in (4).

(2

(2)

X={xt,

Where Imax is the supply phase current maximum value and


o is the supply frequency. On the contrary, the Park's vector
is a circular pattern centered on the origin of the coordinates.
Park's vector approach has been used in open phase and
voltage unbalance faults and fault conditions can be separated
easily in these faults.
Broken rotor bar faults have been detected by constructing
the spectrum of park's vector components, but this technique
needs motor speed or requires knowledge about motor
components [9]. The achieved new feature vector IT by using
two park's vector components is used for fault classification,
This vector iS shown in Fig. 2.

t=I. NJ

(3)

Xt =(Xt-(b-1)T' Xt-(b-2)T.'--Xt)

(4)

It must be selected an appropriate time delay t for phase


space. If t is small compared to the time scales of the time
series, successive elements of the delay vector are correlated.
All vector of X are then clustered around the diagonal in the
Rb, unless b is very large [11]. If t is very large, successive
elements are already almost independent. Mutual information
is used estimating the time delay and false nearest neighbor is
used determining the embedding dimension. The details of
these methods are given in [12].

597

\ 1/II

and feature vector is given in Fig. 4 for one


_
_ ;
component
broken
rotor
bar fault supplying with 300 V.
1\ l )1 1 { l (E) :1 ~KIo
f
After the feature vector is achieved, phase space of each
condition has been constructed. For phase space, embedding
Id, Iq
|)
"Ir
-u
dimension b is selected
two and time delay x is selected
lI
1
l
lll
ll
ll
1ll
1
1lll
l
1ll
1lfour
nearest
false
by
using
neighbor and mutual information
iUK|1
methods, respectively. The graphical results of proposed
-| | l
11ilI
Etlil1
El 11El Imethods to achieve embedding dimension and time delay are
'I 11l E
_1 v1!ll'i "I'd
l lil
ll
0shown in Fig. 5 and Fig. 6, respectively.
I]IJ
11j

0.50l| I \ \ { \ 11|
oIt1

as

6.65

6.1

6.1 5

(time)

6.2

TABLE I

Fig. 2. The new feature vector.

INDUCTION MOTOR CHARACTERISTIC USED

EXPERIMENT
Description

C. Fuzzy C-Means Clustering


Fuzzy C-means method is a clustering technique using
fuzzy membership degree for each data in a cluster. The fuzzy
c-means clustering employs fuzzy partitioning. A data in this
method can belong to all groups with different membership
degree between 0 and 1 [13]. The aim of fuzzy c-means
clustering is to find cluster centers that minimize a
dissimilarity function. The membership matrix (U) is
randomly initialized according to (5). The dissimilarity
function is given in (6).
C

uUij=1,Vj =.n
i~~~~~~~~~~1~~~c

=Ji

J(U' Cl, C2-- Cc

i=1

i=1 j=1

j mdij

Power
Input Voltage
Full Load Current
Supply Frequency
Number of Poles
Number of Rotor Slots
Full Load Speed

(6)

Induction
Motor
Data

Acquisition
~~~~~~~~~~~~~Card

dii

Value

0.37kW
380 V
1.2 A
50 Hz
4
22
1390rpm

(5)(5)

IN THE

Sourcc

Where uij is between 0 and 1, ci is the center point of cluster


i, dij is the Euclidian distance between ith center (ci) and j th
data point and m e [1,oo] is a weighting exponent. The
dissimilarity function is minimized solving (7) and (8).

jk=l

as

PC

-,2 /(m-1)(8

Fig. 3. A diagram of experimental prototype.

0.5-- --

lll. EXPERIMENTAL RESULTS


We performed experiments on an actual induction motor

for our analyses. The characteristics of the three-phase


induction motor used in our experiment are listed in Table I.
The motor was tested with a healthy rotor and with a faulty

05

)l I)1\

rotor that had one, two and three broken rotor bars. The
diagram of experimental prototype is depicted in Fig. 3. The

50

r>

05

jH

0.1
1.04 .
0.05 0.1
(time)
t (time)
Fig.4. Phase currents, Park's vector component and feature vector.

frequency for duration 3 seconds. The supply voltages of the


each motor condition are 380, 340, 300 and 260 V,

respectively. Three phase motor currents, park's vector

598

1.0

2~ ~ ~61

The phase space of each condition is constructed according


to selected parameters. Fuzzy c-means clustering is applied in
phase space for classification. Phase space of the feature
vector that had two broken rotor bar faults and fuzzy cluster
center are shown in Fig. 7. Each phase space is normalized in
[0, 1] range. The combined phase spaces are shown in Fig. 8.
This is performed to show how each phase space of motor
condition is separated.
In table II, the cluster center is determined using fuzzy cmeans clustering for each phase space. Each motor condition
is divided two clusters. The test results are shown in table III.
The Gaussian membership function has been used for
determine the degree of a point to a cluster. The average of
this function gives us a test phase space to belonging a
training phase space. This function is given in (9).
N

_IIV_Iot 112
2r2

,/ e

t=1

0. 8

0.7

0.6
0.

04_
0.3

0.2
01

7.2

7.0

C,

Condition

6.4

Time lag

40

0.8

4.1

0. 7

20.6

0.26300

0.97352

B1
B2
B3

0.63319
0.31117
0.30214

0.22554
0.58022

0.22554
0.58022

0.64643
0.30429
0.32332

1/
C 2

4medn

0.05046

TABLE III
TEST RESULTS

Training data

B3

0.0441

0.0001

0.0767

0.9156

|0 Healthy motor

Onle broken rotor bar faullb

Two broken rotor bar faults

02

0A

06

082

Fig.8. The combined phase space for each motor condition.

IV. CONCLUSION
Fuzzy c-means clustering algorithm and time series data
mining are utilized to distinguish between broken rotor bar
fault and healthy condition of an induction motor at four
different operation speeds. Stator phase currents are the only
inputs to this algorithm. Motor current signature analysis has
been used to detect broken rotor bar faults in earlier studies.
This technique needs motor slip. Therefore rotor speed should
be measured or slip value should be calculated using number

0.5

Oi/

0.05046

This study is compared with other studies in table IV. Our

Cy2

0.26300

method has an advantage so that uses only three phase motor


currents. Furthermore, the mathematical complexity of
proposed method is lower and accuracy rate is higher than
other methods.

i -

Cx2

0.90598

Fig. 5. Determine time delay.

0.9 k

Cyl

6.8D-

20

TABLE II
HEALTHY CONDITION

BTri daB3
B2
B2
B
CZ~~~~~~~~i
112
0.2085
0.1099 0.0008
0.7194
s
B1
0.0249
0.7650
0.2419 0.2171
B2
0.0351
0.0239
0.8840 0.1502
E

1101

6.6

0.8

CLUSTER POINTS OF TRAINING FOR EACH FAULTY AND

Where v is the center of the cluster and r is the radius. The


distance between lot and v, denoted by liv-JotH2 in (9),
represents the degree of similarity, or degree that lot belongs
to the cluster.

06

Fig.7. Phase space of the feature vector for two broken rotor bar.

(9)

0
0.4)
n2

ieso

of poles and rotor slots information.

Fig. 6. Determine embedding dimension.

599

TABLE IV
COMPARISON TO OTHER STUDIES
Used methods
Mathematical
complexity
Medium
Proposedfeature vector and
time series data mining

Paper Id

Used Motor Signals

Our study

Motor phase currents

[1]

Motor phase currents


and speed
Motor phase currents
and speed calculation
from slot harmonics
Motor torque

Motor current signature


analysis
Motor current signature
analysis and park's vector
approach
Time series data mining

[9]

Motor phase currents


and speed

Medium

[10]

Motor phase currents


and speed calculation
from slot harmonics
Motor phase currents
and speed calculation
from slot harmonics

Park's vector approach and


motor current signature
analysis
Time series data mining
Motor current signature
analysis, self-adaptive filter

[3]
[6]

[14]

High
Medium
Medium

Detected faults
One and more broken rotor bar
faults forfour different
operation speeds
One broken rotor bar for four
different load conditions
Four broken rotor bar faults

Average
accuracy
%99.0

%93.75
%97.9

Three, six and nine broken rotor


bar and broken connector faults
One, two, three and four broken
rotor bar faults

%99.0

High

One, two, three and four broken


rotor bar and stator faults

%99.0

Medium

One and two broken rotor bar


faults

%95.0

%95.0

Knowledge about motor components such as number of poles


[6] J. F. Bangura, R. J. Povinelli, N. A. 0. Demerdash and R. H. Brown,
"Diagnostics of Eccentricities and Bar/End-Ring Connector Breakages in
and rotor slots isn't needed in this study. The new feature vector
Polyphase
Induction Motors Through a Combination of Time-Series Data
is achieved park's vector components. The phase space of this
and Time-Stepping Coupled FE-State-Space Techniques", IEEE
vector is usdtodstnushtealcnitoMining
vector iS used to distinguish the fault conditions
and healthy
Transactions on Industry Applications, Vol. 39, No. 4, pp. 1005-1013,
motor condition. Fuzzy c-means clustering algorithm separates
July/August 2003.
each phase space to two clusters. Then each fault condition and
[7] F. Zidani, M. E. H. Benbouzid, D. Diallo and M. S. N. Said, "Induction
Motor Stator Faults Diagnosis by Current Concordia Pattern-Based Fuzzy
healthy motor condition is classified using center points of these
Decision System", IEEE Transactions on Energy Conversion, Vol. 18, No.
clusters. The studies based on time series data mining have used
4, pp. 469-475, December 2003.
the Gaussian mixture models for classification of phase spaces.
[8] H. Nejjari, M.E.H. Benbouzid, "Monitoring and Diagnosis of Induction
The constructed of these models requires complex mathematical
Motors Electrical Faults Using a Current Park's Vector Pattern Learning
computations and many mixtures are used to represent a phase
Approach", IEEE Transactions on Industry Applications, Vol. 36, No. 3,
pp. 730-735, May-June 2000.
space. These models are useful when phase spaces can not
[9] N. Benbouzza, A. Benyettou, A. Bendiabdellah, "An Advanced Park's
exactly separate. But the phase spaces of each fault condition
IEEE
are searte

Vectors Approach for Rotor Cage Diagnosis",

First International

hveusduzySymposium On Control, Communications And Signal Processing, pp 461-

sccssulyinthstuy.S w

c-means clustering with two clusters for each phase space.


Therefore each fault condition is classified with minimum
computational cost.
REFERENCES

464,2004.
[10] Aderiano M. da Silva, "Induction Motor Fault Diagnostic and Monitoring
Methods", M.S. Thesis, Marquette University, Milwaukee, WI, 2006.
[11] H. Kantz and T. Schreiber, "Nonlinear Time Series Analysis ", Cambridge:

Cambridge University Press, 1997.


[12] X. Feng, H. Huang, "A Fuzzy-Set-Based Reconstructed Phase Space
Method for Identification of Temporal Patterns in Complex Time Series",
IEEE Transactions on Knowledge and Data Engineering, Vol. 17, No. 5,
pp. 601-613, May 2005.
[13] J. C. Bezdek, "Pattern Recognition with Fuzzy Objective Function
Algoritms", Plenum Press, New York, 1981.
[14] X. Boqiang, L. Heming, S. Liling, "Sensitive and Reliable Detection of
Broken Rotor Bar Fault in Induction Motors, Industry Applications
Conference, Vol. 2, pp. 714-719, October 2004.

[1] B. Ayhan, M. Y. Chow and M. H. Song, "Multiple Signature ProcessingBased Fault Detection Schemes for Broken Rotor Bar in Induction
Motors", IEEE Transactions on Energy Conversion, Vol. 20, No. 2, pp.
336-343, June 2005.
[2] M. E. H. Benbouzid, "A Review of Induction Motors Signature Analysis
as a Medium for Faults Detection", IEEE Transactions on Industrial
Electronics, Vol. 47, No. 5, pp. 984-993, October 2000.
[3] M. Haji and H. A. Toliyat, "Pattern Recognition - A Technique for
Induction Machines Rotor Broken Bar Detection", IEEE Transaction
Energy on Conversions, vol. 16, no. 4, pp. 312-317, December 2001.
[4] B. Mirafzal and N.A.O. Demerdash, "Effects of Load Magnitude on
Diagnosing Broken Rotor Bar Faults in Induction Motors Using the
Pendulous Oscillation of the Rotor Magnetic Field Orientation", IEEE
Transactions on Industry Applications, Vol. 41, pp. 771-783, May- June
2005.
[5] K. Kim, A. G. Parlos, R.M. Bharadwaj, "Sensorless Fault Diagnosis of
Induction Motors", IEEE Transaction on Industrial Electronics, Vol. 50,
No. 5, pp. 1038-1051, October 2003.

600

Das könnte Ihnen auch gefallen