You are on page 1of 6

2nd IEEE International Conference on Power and Energy (PECon 08), December 1-3, 2008, Johor Baharu, Malaysia

Development of Artificial Neural Network Based


Fault Diagnosis of Induction Motor Bearing
Abd Kadir Mahamad* and Takashi Hiyama*
* Department of Computer Science and Electrical Engineering of Kumamoto University
2-39-1 Kurokami, Kumamoto 860-8555, Japan. Email: kadir@uthm.edu.my

Abstract The common component failure of induction Today, there are various forms of fault diagnosis
motor is bearing. Thus, timely detection and diagnosis of technique for rotating machineries [1-3]. The intelligence
induction motor bearing (IMB) is very crucial in order to techniques such as artificial neural network (ANN), fuzzy
prevent sudden damage. This paper proposes developing logic, evolutionary algorithm (EA) are the best solution
artificial neural network (ANN) model of IMB fault for this matter. The literature on this matter is wide and
diagnosis by using Elman Network. The vibration signal diverse due to a wide variety of systems, components and
obtained from Case Western Reserve University website are parts. Hundreds of papers in this area, including theories
been used as input signal. During preprocessing stage, and practical applications, appear every year in academic
vibration signal have been converted from time domain into journals, conference proceedings and technical reports.
frequency domain through Fast Fourier Transform (FFT). In this paper, the main objective is to develop ANN
Enveloping method was then, used to eliminate the high
model of IMB fault diagnosis using Elman Network
frequency components from vibration signal. Subsequently,
which as shown good performance in order to diagnosis
a set of 16 features from time and frequency domain were
IMB as in [4]. Vibration signal is used as input signal
extracted. Furthermore, the distance evaluation technique is
used in features selection in order to select only informative
which is normally have been used to detect initial fault as
features. In order to make the ANN model more flexible, the
in [2, 5-7]. The selection of ANN for modeling process in
sensitivity analysis of IMB is introduced. Lastly, a graphical this paper is due to its wide use in many applications
user interface (GUI) program is created as a tool for help especially in order to determine fault diagnosis [1-7]. This
users determines the situation of IMB conditions. method can be used in industries to determine either IMB
is still in good condition or their having faulty in inner,
Keywords Induction motor bearing; Vibration; Fault outer or ball from IMB. The success in determining faulty
diagnosis; Artificial Neural Network; Distance Evaluation. part from IMB help the industries to do further
investigation of the root cause of this particular problem.
II. METHODOLOGY
I. INTRODUCTION A. Vibration Signal
The manufacturing and industrial sectors of the world The raw data used in this work is obtained from the
are increasing rapidly, to meet demands to produce goods Case Western Reserve University website [8]. The
at higher output and better quality. Therefore, machine apparatus involved in these experiments consist of a 2 hp,
failure is crucial to avoid as it comes with negative 3 phase induction motor, accelerometer, dynamometer,
implications and difficult to identify and localize in a and torque transducer with self alignment coupling. An
timely manner. Scheduled maintenance practices today is accelerometer is mounted on the top motor housing at the
also contribute in reduction of machine lifetime and drive end of the induction motor. The type of bearing
increase downtime, which result in loss of productivity. tested in this work is deep groove ball bearings with
To overcome this problem, many researchers propose inside diameter: 0.9843 inches, outside diameter: 2.0472
methods by using fault diagnosis. inches, ball diameter: 0.3126 inches, pitch diameter:
Fault diagnosis systems can be defined as systems that 1.537 inches and rpm of approximately to 1800. The
have capability to detect and diagnose faults occurred [1]. vibration signal is collected at 12 kHz samples per second
The word diagnosis normally is defined as a judgment to from the experimental rig and these bearings are tested
discover the current faulty by making a careful under different bearing status condition and operating
examination of the state of a system. The degradation load.
sources may come from the components, process or The bearings are tested under four different loads (0, 1,
subsystems of a complex system. Basically, fault 2 and 3 hp). Each defective bearing is tested with defect
diagnostics can be divided into 3 categories; fault sizes of 0.007, 0.014 and 0.021 inches with single point
detection, isolation and identification. Fault detection is a faults are introduced to the test bearing using electro-
task to indicate whether something is going wrong in the discharge machining (EDM). There are four different
monitored system; fault isolation is a task to locate the operating conditions obtained from experimental rig
component that is faulty; and fault identification is a task which are at normal condition, outer race fault condition,
to determine the nature of the fault when it is detected. inner race fault condition and ball fault condition.

1-4244-2405-4/08/$20.00 2008 IEEE 1387


2nd IEEE International Conference on Power and Energy (PECon 08), December 1-3, 2008, Johhor Baharu, Malaysia

B. Pre-processing stage TABLEE I.


FEATURES EXTRACTION FROM TIME DOMAIN
During pre-processing stage, the vibrration signals from
bearing have been converted from time domain to Time domain features
frequency domain through the Fast Fourier
F Transform
(FFT). This signal then has been coonverted to power Features Equation
spectrum in order to measure the pow wer of a signal at
particular frequency as shown in Fig. 1.
Standard Deviation
Besides to understand the bearing daamaged condition,
it is also important to know where thhe damaged come

from. For instance the bearing componnents can damaged
Root Mean Square

in outer race, inner race, rolling elemments or from the Maximum


cage. Therefore, the enveloping methood have been used Amplitude
to eliminate the high frequency com mponents from the Minimum Amplitude 



vibration signal and the peak amplitudee will be created to Crest Factor


show the repetition of dominant ampplitude lies in the

frequency spectrum. Skewness
This method has 3 steps, which are baand pass filter, non std

linear rectifier and low pass filter. A band
b pass filter is
Kurtosis
used to eliminate a signal at high ressonance frequency std
from the signal. The second steps,, the signal will
demodulated by non-liner rectifier. A low pass filter is
Clearance Factor

placed after demodulated signal, which eliminated the low


frequency signal due to bearing defect associated
a with low
Shape Factor
frequency signal.


C. Features Extraction
Impulse Factor

wn as a method to
Generally, feature extraction is know

reduce dimension from vast data. By B doing feature.
Mean


TABLEE II.
FEATURES EXTRACTION FROM FREQUENCY DOMAIN

Frequency domaain features

Features Equation
Max. Power

Spectrum
Max. Envelope



Frequency Centre




RMS Frequency

Standard Deviation


Frequency

extraction the relevant inform mation can be extracted,


which represent the behavior of data. Normally, statistical
extraction features such as mean, variance, standard
deviation and kurtosis are the common
c features employed
for features extraction, but the approach
a in this work done
for time domain and frequenccy domain as shown as in
Table I and Table II respectivelly.
D. Features Selection
Some features are relevant but some others are not.
Therefore, before the features are fed to classify process,
it is important to select informmative features and discard
irrelevant data. Features selecction that used in this work
Figure 1. Vibration signal, Power Spectrum andd Envelope Spectrum; is known as Distance Evaluatiion technique [7]. Suppose
Sampling frequency =12kHHz the joint feature set of c conditiion patterns is as in (1).

1388
2nd IEEE International Conference on Power and Energy (PECon 08), December 1-3, 2008, Johor Baharu, Malaysia

(1)

where is the kth feature of i and Ni is the number of


the features in i.

The average distance of all the features in can be
defined as in (2);



(2)

But the average distance of Si (i=1,2,,c) given as Figure 2. Block diagram of Elman network

respectively, the input vector, the output vector and the


(3) state vector, at a discrete time k. In addition to the input

and the output units, the Elman network has a hidden
Then, by substituting (2) into (3), we can have new unit, . and are
equation as shown in (4). the interconnection matrices, respectively, for the
context-hidden layer, input-hidden layer and hidden-

output layer. Theoretically, an Elman network with n

 (4) hidden units is able to represent an nth order dynamic

where system. The dynamics of the Elman network is described
by the difference equations as in (11) (13).



(5)

is a mean of all features in .  (11)


Then, the average distance of c difference condition (12)
patterns is given as; (13)

  (6)
where is an intermediate variable.
where Note that the Elman network differs from conventional
two-layer networks in that the first layer has a recurrent




(7) connection. The delay in this connection stores values
from previous time step, which can be used in the current
time step. Because the network can store information for
is the mean of all the features in all c condition-patterns. future reference, it is able to learn temporal patterns as
Finally, we define the distance evaluation as; well as spatial patterns. The Elman network can be
trained to respond and to generate, both kinds of patterns.

(8) The configuration of ANN model, for Elman network

has used logsig (log-sigmoid) transfer function in its
Then the optimal features can be selected from the hidden layer and purelin (linear) transfer function in its
output layer. With this combination, the network can give
original features set according to the bigger distance
approximation to any function. Furthermore, the network
evaluation of training was using Levenberg Marquart Algorithm due to
E. Elman Network. it better performance among others network training.
Elman [9] has proposed a partially recurrent network, The input data for ANN training were selected from
where the feedforward connections are modifiable as the results of features selection. The input data have been
shown in Fig. 2. The Elman network is a two-layer divided into two sets; training and validation set. The
network with feedback from the first layer output to the training sets were selected from features selection which
first layer input. 9 salient features were selected but for validation data
This recurrent connection allows the Elman network to sets were perturbed with + 5% of the training set.
both detect and generate time-varying patterns. To Meanwhile the output of ANN training were
understand the feature offered by Elman network, corresponding to four conditions of bearing, which are
consider a multivariable plant with m inputs and q normal, inner, outer or ball bearing fault condition. In
outputs, describe by a general nonlinear input-output order to select an optimum network of this networks the
discrete time state space model. training process is setup from 2 hidden nodes until 25
hidden nodes until the optimum network obtained. The
selection corresponds based on the smallest cross-
(9) validation errors produced.
(10) The performance of network is measured by using
mean square error (MSE) in order to define the best
where and are non-linear network [10].
The MSE equation is state as in (14);
functions; , and are,

1389
2nd IEEE International Conference on Power and Energy (PECon 08), December 1-3, 2008, Johhor Baharu, Malaysia

TABLE IV.
PERFORMANCE OF TRAINING AND VALIDATION ERROR OF ELMAN

(14)

NETWORRK

No. of
where ei = Error ai =A
Actual value ANN Type Hidden T
Training error
Validation
ti= Desired value N = Number of data Node
error
Elman
5 0.000026 0.000762
Network
III RESULTS AND DISCUSSSIONS
The bar graph from Fig. 3 shows thee result of features
selection. In this work, the threshold vaalue is set as 1.0 of
Ja value. The value of Ja significance ata the important of
those features which the bigger valuue indicate more
BFC
important of those features. The threshhold value was set
as 1.0 due to the performance of classification shows the OFC
better result could be obtained from this selection as
shown in Table III. This returns to thee lowest validation IFC
error of 0.000762 compared with other Ja J values.
From the total of 16 features extractioon, only 9 features NC
were selected for ANN fault diagnosis. There were
maximum, minimum, standard deviation, RMS and shape
factor from time domain and the resst were maximum
power spectrum, envelope spectrum, frequency centre
and RMS frequency from frequency dom main.
Table IV shows the result of trainiing and validation BFC
error of Elman network. The resultt shows that the
OFC
smallest cross-validation error of Elmaan network can be
obtained with hidden nodes of 5. The validation
v error for IFC
this network was 0.000762 with training t error of
0.000026. Thus, this result was selecteed as an optimum NC
network for IMB fault diagnosis. Thee performances of
classification (fault diagnosis) output neetwork correspond
to normal (NC), inner (IFC), outer (OF FC) and ball fault Figure 4. Performance for training annd validation of Elman Network
condition (BFC) for training and vallidation of Elman
network are depicted as in Fig. 4. The + sign indicates the actual output and the sign
indicates the prediction outputt from ANN model. The
Distance evaluation techn
nique performances of training and validation error obviously
show the Elman Network can gives a better in order to
8
diagnosis the IMB failure as prroven in [4].
6 Even though the perform mance of validation and
Ja Value

Threshold training performed very well towards


t the input data, its
4
performance was not fully satisfactory especially when it
2 encountered the rapidly changiing at input that never been
0 seen before. Therefore theree was indeed a need to
increase the adaptability for thhe model to response to the
xmin

CLF

CF
xm

xrmsf
xmaxf
xstd
xrms

xmaxen
xmax

xskew

SF
IF
xkurto

xstdf
fc

input changing. In order to solve this problem, the


sensitivity analysis was intrroduced to improve the
No of features adaptability of the model to unppredictable variations input
that may resulting the misclassiification at ANN output.
Figure 3. Result of Distance Evaluationn Technique Table V shows the potential misclassification of
sensitivity analysis of IMB faailure. From Table V, it is
TABLE III. found that during normal conddition by increasing 20% of
PERFORMANCE OF VALIDATION ERROR IN ORDER TO SELECT input data will cause a wronng of output classification
THRESHOLD VALUE OF Ja
which classification output shows inner fault instead of
Validation normal condition. Meanwhile, the output classification
Ja Value No. of features
error didnt shows any changed condition if input data
Ja  0.25 16 0.0015 decreasing by 20%. For the innner type fault, it is found
Ja  0.5 15 0.0025
by increasing 5.8% of input daata the output classification
Ja  0.75 10 0.0021
Ja  1.0 9 0.000762
will gives wrong classification result. The result supposes
Ja 1.25 5 0.0013 to be inner fault but it shows thhe outer fault. On the other
hand by decreasing 28% of o input data the output
classification gives the results as normal condition. The

1390
2nd IEEE International Conference on Power and Energy (PECon 08), December 1-3, 2008, Johhor Baharu, Malaysia

misclassification output results byy increasing or


decreasing biases are shown in Fig. 5. The sign
indicates the results of output cllassification after
increasing magnitude biases, the * sign indicates the
results of output classification after decreasing magnitude
biases, while + sign indicates the boundary
b limit of
output classification. The fault patterns from both cases
by increasing or decreasing input data show
s the variety of
bias may result misclassification at outpput classification.
The output data for the fault is designned in such a way
that the data is represented by a linear l relationship (d)
classification between 1 and 4 correspoonding to NC, IFC, Figure 5. Potential misclassification of
o sensitivity analysis for (a) NC,
OFC and BFC respectively. (b) IFC, (c) OFC and (dd) BFC respectively

TABLE V. The interest of this work was


w motivated by retraining
POTENTIAL MISCLASSIFICATION OF SENSITTIVITY ANALYSIS the ANN after it give the wronng output classification as it
Increased or have ability to adapt the new change
c in input data. From
Classification
Types of Fault decreased of the investigation, it waas observed that the
output changed
input data misclassification of ANN model occurred when the input
20% Inner fault data exceeded the boundary limit l of magnitude biases.
Normal (NC)
-20% No changed
5.8% Outer fault
Thus, this condition was incluuded as the main criteria in
Inner (IFC) activating the retraining strateggy. Here, three sets of data
-28% Normal
13% Ball fault used for retraining process thatt were the original data, the
Outer (OFC)
-13% Inner fault increased data and decreased data. Therefore, the
6.4% Undefined observation data becomes 1444 data instead of 48 data
Ball (BFC)
-5% Outer fault
during the first ANN classsification. The results of
retraining process are shownn in Table VI and the
performances of this process arre shown in Fig. 6.
In Table VI, the obtained vaalidation error and training
error were 0.00075 and 0.000000725 respectively. It
means that the retraining prrocess performance much
better for validation and traininng compared with the first
training process. The improveement result maybe due to
increasing input data that fed innto the model which makes
retraining model process moore flexible and adequate
which its resulting the better peerformance.

(a) TABLE VI.


PERFORMANCE OF RETRAINING PROCESS FOR TRAINING AND
VALIDATION ERROR OF ELMAN NETWORK
No. of Hidden Training Validation
ANN Type
Node error error
Elman
6 0.00000725 0.00075
Network

(b)

(a)

(c)

1391
2nd IEEE International Conference on Power and Energy (PECon 08), December 1-3, 2008, Johhor Baharu, Malaysia

IV CONCLLUSION
In this paper, a method forr fault diagnosis of IMB is
proposed. The vibration data haveh been converted from
time domain to frequency domaind before enveloping
techniques were applied to itt. Then 16 features were
extracted from time domain, frequency
fr domain and time
frequency domain signal. In order
o to reduce the size of
data, features selection was appplied where 9 significant
features were selected as the data for training process.
During the development off ANN modeling Elman
(b) Network was used for IMB fault fa diagnosis. In order to
Figure 6. Performance of retraining process for
f (a) training and make ANN model more flexibble, the sensitivity analysis
(b) validation of Elman Netwoork of IMB is introduced. Lastly, the
t graphical user interface
(GUI) program is created to heelp the user to use as a tool
Lastly, the GUI program as shown inn Fig. 7 is created in order to determine the conndition of IMB condition
to help the operator to determine the IMMB conditions. By either in normal or faulty condiitions.
using this GUI, the users can easily determine
d whether
the vibration signal is in the normal condition
c or in the REFEREN
NCES
faulty condition (inner, outer or ball fauult). This program [1] S.Rajakarunakaran P.Venkumar, D.Devaraj and K.Surya P. Rao,
can assist users that have no know wledge about the Artificial neural network approoach for fault detection in rotary
vibration signal and ANN to determine the condition of system, Applied Soft Computingg, vol. 8, pp. 740-748, 2008.
IMB. [2] Y. Lei, Z. He, Y. Zi and Q. Hu, H Fault diagnosis of rotating
machinery based on multiple ANFISA combination with Gas,
The Plot pushbutton is used to show w result of the raw Mechanical Systems and Signal Processing, vol. 21, pp. 2280-
data, power spectrum and envvelope spectrum 2294, 2007.
respectively. At the bottom side of program is a warning [3] J. M. F. Salido and S. Murakami, A comparison of two learning
text to show the status of IMB conditioon. For instance at mechanisms for the automatic design of fuzzy diagnosis systems
Fig. 7 shows the current IMB status is in ball damage for rotating machinery, Appliedd Soft Computing, vol. 4, pp. 413-
condition. This program is design to be simple and easy 422, 2004.
[4] A.K Mahamad and T.Hiyama, Artificial
Neural Network Based
to use especially for those who have no knowledge of Bearing Failure Diagnosis for Indduction Motor, The International
neural network diagnose the IMB failure. Conference on Electrical Engineeering 2008.
[5] Q. Hu, Z.He, Z. Zhang and Y. Zi, Fault diagnosis of rotating
machinery based on improved wavelet package transform and
SVMs ensemble, Mechanical Systems and Signal Processing,
vol. 21, pp. 688-705, 2007.
[6] Y. Yang, D. Yu and J. Cheng, A fault diagnosis approach for
roller bearing based on IMF envelope spectrum ad SVM,
Measurement , vol. 40, pp. 943-950, 2007.
[7] Y. Lei, Z. He, Y. Zi and X. Chen,C New clustering algorithm
based fault diagnosis using coompensation distance evaluation
technique, Mechanical Systems and Signal Processing, vol. 22,
pp. 419-435, 2008.
[8] K.A Loparo: Bearing vibration data set, Case Western Reserve
University. January 2008
http://www.eecs.cwru.edu/laboratory/bearing/download.htm.
[9] J. Elman, Finding Structure in Time,
T Cognitive Science, vol.
14, pp. 189-211, 1990.
[10] A. Ahmad and M. Kamaruddin, Detection of Multiple Sensor
Faults in A Palm Oil Fractionatiion Plant Using Artificial Neural
Network, Regional Symposium on o Chemical Engineering 2002.

Figure 7. GUI for IMB diagnoosis

1392

You might also like