experimental investigations on induction machine...
TRANSCRIPT
Experimental investigations on induction machine conditionmonitoring and fault diagnosis using digital signal
processing techniques
Sa’ad Ahmed Saleh Al Kazzaz a, G.K. Singh b,*a Department of Electrical Engineering, University of Mosul, Mosul, Iraq
b Department of Electrical Engineering, Indian Institute of Technology, Roorkee 247667, India
Received 15 February 2002; received in revised form 7 November 2002; accepted 25 November 2002
Abstract
Condition monitoring is used for increasing machinery availability and machinery performance, reducing consequential damage,
increasing machine life, reducing spare parts inventories, and reducing breakdown maintenance. An efficient condition monitoring
scheme is capable of providing warning and predicting the faults at early stages. The monitoring system obtains information about
the machine in the form of primary data and through the use of modern signal processing techniques; it is possible to give vital
information to equipment operator before it catastrophically fails. The suitability of a signal processing technique to be used
depends upon the nature of the signal and the required accuracy of the obtained information. Therefore, in this paper, signals
obtained from the monitoring system are treated with different processing techniques with suitably modified algorithms to extract
detailed information for machine health diagnosis. In this study, on-line analysis of the acquired signals has been performed using
C��, while MATLAB has been used to perform the off-line analysis.
# 2002 Elsevier Science B.V. All rights reserved.
Keywords: Induction machine; Fault; Condition monitoring; Diagnostic; Digital signal processing; Fourier transform
1. Introduction
Predictive maintenance by vibration monitoring of
electrical machine is a scientific approach that becomes
the new route to the maintenance management [1�/4].
Electrical machines, even new ones, generate some level
of vibration [5�/17]. Small levels of ambient vibrations
are acceptable. However, higher levels and increasing
trends are symptoms of abnormal machine perfor-
mance. Machine vibration analysis becomes one of the
important tools for machine faults identification. There
are two types of analysis, time domain and frequency
domain. The frequency domain analysis is more attrac-
tive one because it can give more detailed information
about the status of the machine whereas; the time
domain analysis can give qualitative information about
the machine condition. Generally, machine vibration
signal is composed of three parts, stationary vibration,
random vibration, and noise. Traditionally, Fourier
transform (FT) was used to perform such analysis. If
the level of random vibrations and the noise are high,
inaccurate information about the machine condition is
obtained. Noise and random vibrations may be sup-
pressed from the vibration signal using signal processing
techniques such as filtering, averaging, correlation,
convolution, etc. Sometimes random vibrations are
also important because they are related to some types
of machine faults hence; there is a need to observe these
vibrations also.
Signals obtained from the transducers are in the form
of continuous voltage or current signals. It is necessary
to define their values at certain instants of time to be
suitable for digital signal processing (DSP) applications.
The obtained digital signal is an adequate substitute for
the underlying continuous signal if the interval between
the successive samples is sufficiently small. The sampling
frequency must be twice the highest frequency compo-* Corresponding author. Fax: �/91-1332-73560.
E-mail address: [email protected] (G.K. Singh).
Electric Power Systems Research 65 (2003) 197�/221
www.elsevier.com/locate/epsr
0378-7796/03/$ - see front matter # 2002 Elsevier Science B.V. All rights reserved.
doi:10.1016/S0378-7796(02)00227-4
nents of the signal (according to Shannon’s theorem) to
avoid aliasing of high frequencies components in the low
frequency region of the spectrum. In the present work,
the sampling frequency has been selected to be fourtimes the highest frequency component of the signal to
prevent any possibility of aliasing and to ensure the
complete reconstruction of the signal.
In this paper, signals obtained from monitoring
system as shown in Fig. 1, are treated with different
processing techniques with suitably modified algorithms
to extract detailed information for induction machine
diagnosis. All the techniques used here for signalanalysis and processing have been implemented in
C�� and MATLAB software. On-line analysis of the
acquired signals is performed using C��, while MATLAB
is used to perform the off-line analysis.
2. Nature of electrical machine faults
The induction motor is considered as a robust and
fault tolerant machine and is a popular choice in
industrial drives. It is important that the measures are
taken to diagnose the state of the machine as and when
it enters into the fault mode. It is further necessary to do
so on-line by continuously monitoring the machine
variables. The reasons behind failures in rotatingelectrical machines have their origin in design, manu-
facturing tolerance, assembly, installation, working
environment, nature of load and schedule of mainte-
nance. Induction motor like other rotating electrical
machine is subjected to both electromagnetic and
mechanical forces. The design of motor is such that
the interaction between these forces under normal
condition leads to a stable operation with minimumnoise and vibrations. When the fault takes place, the
equilibrium between these forces is lost leading to
further enhancement of the fault.
The motor faults can be categorised into two types:
mechanical and electrical. The sources of motor faults
may be internal, external or due to environmental, as
presented in Fig. 2. Internal faults can be classified withreference to their origin i.e. electrical and mechanical or
to their location i.e. stator and rotor. Usually, other
types of fault i.e. bearing and cooling faults refer to the
rotor faults because they belong to the moving parts.
Fig. 3 presents the fault tree of induction machine where
the faults are classified according their location: rotor
and stator.
3. Simulation of induction machine under healthy and
fault conditions
Modelling and simulation of electrical machine dy-
namics has attracted many researchers since the early
days of electrical machine invention [18]. The fastadvances in computing facilities and the improvement
in numerical techniques have lead to improvement in
accuracy and simulation efficiency. Mathematical mod-
els have been developed to include the effect of core loss,
saturation effect, winding distribution, and inherent
machine faults [18�/24].
3.1. Dynamic analysis of induction motor
For simulation of dynamic state, the choice of model
is made on the basis of operating conditions as follows:
. Machine operating from balanced sinusoidal supply
under nominal voltages, and under/over voltages
(phase variable model).
. Machine operating from balanced non-sinusoidal
supply obtained from inverter (stationary referenceframe model).
. Machine operating from unbalanced sinusoidal sup-
ply (instantaneous symmetrical component model).
Fig. 1. Schematic diagram of the monitoring system.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221198
Fig. 2. Sources of induction machine faults.
Fig. 3. Popular induction machine faults and their causes.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221 199
. Machines operating from acquired (recorded voltages
from monitoring system) supply voltages (phase
variable model).
3.1.1. Various test conditions used for simulation purpose
In this work, three phase 5 hp induction motors were
chosen for the simulation purpose because the samemachines are used in the practical investigations. These
machines were simulated (dynamic simulation) for
different test conditions, which are as follows:
. Transient performance of induction motor under
nominal supply.
. Transient performance of induction motor under
unbalanced (including single phasing) supply.
. Transient performance of induction motor with
under-voltage and over-voltage supply.
. Transient performance of induction motor undervariable frequency sinusoidal supply.
. Transient performance of induction motor under
variable frequency non-sinusoidal supply.
. Transient performance of induction motor under
mechanical faults (rotor eccentricity, dry bearing,
and faulty bearing due to ball defects).
3.2. Induction motor health identification based on on-line
machine modelling
The developed monitoring system (Fig. 1) comprisesof three transformers for voltage measurement, three
Hall-Effect probes for current measurements, pulse-
tachometer for speed measurement and four thermo-
couples for temperature measurements. Signal condi-
tioners are used for thermocouples to provide
linearisation, amplifications and cold junction compen-
sation, and for vibration transducer to provide excita-
tion, filtering and amplifications. The data is transferredto the computer using 12-bit A/D converter that
provides less than 0.1% error in amplitude of the
acquired signal. The sampling frequency has been
selected so that complete reconstruction of the signals
can be achieved. The sampling frequency is adjusted to 2
kHz for electrical variables, and 4 kHz for vibration
signals. The speed transducer output is in order of 1000
pulse per revolution. To obtain the motor speed, thesepulses are counted for a specific period of time (200 ms)
using an on-board 16-bit down counter. The system
accuracy and performance is tested with the known
inputs.
The hardware along with the software allows the users
to effectively monitor, store, and analyse machine
variables. The system provides on-line display of
voltages, currents, and temperatures (using C�� gra-
phic facility) with simple data analysis or directly storing
the acquired data for off-line data analysis and proces-
sing. A sampling frequency of the order 15 kHz is
achieved with the on-line display of the acquired data,
and about 50 kHz with the direct storing of the acquired
data. Moreover, the change in sampling frequency;
number of samples, range of display and number ofmachine variables is possible.
The machine health identification can be obtained
with the aid of the on-line monitoring system discussed
above. In this system, three phase currents; three phase
voltages and speed are recorded on-line and stored in
computer memory. The recorded three phase voltages
are fed to the developed machine model in order to
calculate the machine currents and to predict themachine conditions. By comparing the actual recorded
machine currents (recorded simultaneously with ma-
chine voltage) with the simulated currents, the machine
conditions can be obtained qualitatively. One of the
effective methods, which have been adopted recently to
predict machine condition using machine currents, is
Park’s vector approach [25]. Here this method is
employed to obtain the machine behaviour due tovarious supply conditions.
4. Signal analysis
Signal analysis is used to extract some useful features
of the signal i.e. mean value, mean square, root mean
square (RMS) value and the crest factor. The signal
detectors have been implemented by software using
simple algorithms.
4.1. Implementation of root mean square
In this study, the RMS value of the vibration signal is
used for primary investigation of the machine health.
The RMS values of the machine voltages and currents
are used to detect the unbalanced supply conditions, and
to differentiate its effect from the effect of the other
types of faults. Table 1 represents the RMS values of the
machine voltages and currents and the average speed for
five identical machines running under same operatingconditions. These values are used as input to the neural
network based fault classifier. Table 2 represents the
RMS values of radial and axial machine vibrations for
five identical machines running under different condi-
tions.
4.2. Implementation of Crest factor
The Crest factor is the ratio of the peak value to the
RMS value. It is meaningful where the peak values are
reasonably uniform and repeatable from one signal cycleto another. The Crest factor yields a measure of the
spikiness of a signal and is used to characterise signals
containing repetitive impulses in addition to a lower
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221200
level continuous signal. Crest factor is often used to
indicate the rolling element bearing faults.
It may be noticed that both radial and axial vibrations
are affected by the machine condition, but it is difficultto recognise the type of the fault by using these values.
These values can give qualitative information about the
machine condition if it is compared with the vibration
standard. The values of Crest factor of the vibration
signal for healthy and faulty five symmetrical induction
machines running at no load are represented in Table 3.
The bearing faults clearly affect the value of the Crest
factor but the change in Crest factor is small with drybearing fault. Continuous monitoring of this factor
along with historical records can provides useful in-
formation about the bearing condition. From the above
tables, it may be noticed that some of the machine
conditions have a significant effect on the RMS value of
the vibration while the others do not, and the same is
true with the Crest factor. However, using both the
RMS value and the Crest factor for machine healthidentification may increase the system efficiency for
diagnostics.
5. Signal processing techniques
There are two approaches used for processing the
signal; time domain and frequency domain. In time
domain approach, the discrete time signal is directly
analysed by one of the DSP techniques [26�/29] such asfiltering [30,31], averaging [2], convolution, correlation
etc. In frequency domain approach, the signal is first
transformed to the frequency domain using FT then,
different methods of analysis such as averaging, con-
volution, power spectrum, cepestrum etc. can be ap-
plied. Different signal processing approaches have been
used here to extract the salient features of the signalsobtained from the machine.
6. Implementations of signal processing techniques in
time domain
Different aspects are available for time domain
analysis such as; time period of the signal, the peak
value reached by the signal, the average value of the
signal, RMS value of the signal etc. The choice of such
approaches depends on the nature of the signal and the
required information. In this section, some of the DSP
techniques are introduced with their implementationswith the data of vibration and electrical variables.
6.1. Implementation of signal averaging
Time domain averaging is effective in suppressingsignals that are not correlated within the averaging
period. To get good results, it is necessary to know the
repetition frequency precisely and sampling the signal
with an integer number of samples per period. Noise
reduction goes as 1=ffiffiffiffiffi
Np
; greatly increasing the signal to
noise ratio. In machinery diagnostics using the vibration
signal, this approach has some serious drawbacks.
The underlying assumption is that the shaft rotationrate is constant. This is not the case in rotating machines
where there is variability in the shaft speed, which
broaden the spectral peaks and gives poor results. This
Table 1
RMS value of machine voltages, currents and average speed
V12 (V) V23 (V) V31 (V) I1 (A) I2 (A) I3 (A) N (rpm)
Machine 1 417.324 415.566 418.940 4.222 4.176 4.156 1473
Machine 2 413.450 412.053 413.618 4.112 4.141 4.162 1481
Machine 3 419.968 417.243 421.083 4.297 4.247 4.285 1479
Machine 4 416.826 417.940 420.556 4.364 4.314 4.388 1470
Machine 5 418.431 418.563 416.910 4.100 4.117 4.090 1485
Table 2
RMS value of radial and axial vibration for different machine conditions
RMS vibration (radial) (g) RMS vibration (axial) (g)
M/C 1 M/C 2 M/C 3 M/C 4 M/C 5 M/C 1 M/C 2 M/C 3 M/C 4 M/C 5
Healthy condition 0.024 0.025 0.021 0.025 0.031 0.023 0.033 0.032 0.035 0.040
Unbalanced supply 0.054 0.046 0.050 0.043 0.055 0.103 0.115 0.111 0.099 0.112
Single phasing 0.275 0.275 0.278 0.287 0.285 0.239 0.281 0.266 0.248 0.241
Mechanical unbalanced 0.042 0.044 0.038 0.047 0.049 0.043 0.046 0.042 0.051 0.055
Faulty bearing (dry) 0.023 0.025 0.024 0.023 0.026 0.034 0.043 0.045 0.034 0.042
Faulty bearing (ball defect) 0.137 0.124 0.150 0.118 0.168 0.149 0.136 0.138 0.134 0.154
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221 201
approach assumes that the shaft rate is known with agood deal of precision. Some measurement error in the
shaft rate that serves to mistune the comb filter giving
poor results is always expected. Hence, it is required that
the data acquisition system should have adaptable
sampling rate according to the current shaft rate. This
requirement is difficult to meet in practice. Some of
these problems have been previously addressed and
various solutions have been proposed. One solution isto synchronise the vibration signal with a tachometer
signal, where the tachometer signal marks the beginning
of each averaging period. The data is then, ensemble
averaged over the periods marked by the beginning of
the tachometer pulse. This approach gets around the
problems associated with the ability to adaptively
change the sampling rate. Another alternative is to
take a single vibration measurement at an arbitrarysampling rate and do irregular resampling to interpolate
the signal.
For voltage and current signals, this approach is
successfully implemented to clean up the signal from
noise. The reference point is precisely specified and an
average of 18 runs gave good results. Fig. 4 shows the
voltage waveforms of different runs and their average
where the improvement in the signal can be observedclearly. Then the obtained signal is transformed to the
frequency domain and treated with different signal
processing and analysis techniques.
6.2. Implementation of correlation
Correlation between two signals is used to obtain the
similarity between them. Correlation function of the
baseline and real time signals is used to investigate theirrelationship. Vibration, current and voltage signals of a
healthy machine are considered as baseline signals. For
primary investigation, two segments of a long vibration
signal are correlated to obtain their similarity, as shown
in Fig. 5. The correlation function of the vibration
signals obtained from different runs with the same
conditions is shown in Fig. 6. The figure shows that
the correlation among them is high. The correlationbetween a baseline vibration signal and a faulty bearing
vibration signal is shown in Fig. 7. It can be concluded
that even the vibration signals obtained from the same
machine running at the same condition, the correlationamong them is not very high. This is due to the presence
of noise and random vibrations. The information
obtained from the correlation function play an impor-
tant role in selecting the input data of the neural
network. For electrical quantities such as voltages and
currents, better results are obtained by implementing the
correlation function. Fig. 8 shows a couple of current
signal and a couple of voltage signal obtained from thesame machine under healthy and faulty conditions and
the corresponding correlation functions.
6.3. Implementation of signal filtering
Filters are used for two purposes, to attenuate the
noise and undesired frequency components and to
separate some individual frequencies or band of fre-
quencies for their relation with the machine faults. TheRMS value of selected frequency components or band
of frequencies is used to obtain machine condition by
comparing the obtained values with the corresponding
reference values. As it is mentioned above, the filter
characteristic plays the key role of obtaining the
required resolution and accuracy of the analysis. For
example, it is desired to pick up a frequency component
related to a certain type of machine fault which is closeto a dominant frequency component, i.e. supply fre-
quency of 50 Hz and double rotational speed 48.5 Hz
(for rotor speed�/1455 rpm). Fig. 9 shows the vibration
signal before and after using smoothing filter for
removing the high frequency noise from the signal.
Due to the fluctuations in the vibration signal, bands of
frequencies rather than individual frequency compo-
nents are used to identify the machine conditions. Theband pass filter that met the above-mentioned require-
ments is achieved using FIR filter.
Fig. 10 presents the results of band pass filtering of
three frequency regions of the vibration signal. For the
purpose of diagnosis, variable tuned band pass filter is
used to separate four frequency regions related to some
common types of machine fault. The filter is first tuned
to pass the frequency band of 1�/200 Hz, this band isrelated to the first and second harmonics of the bearing
characteristic frequencies and shaft frequency, which is
related to the mechanical unbalance. A narrow band
Table 3
Crest factor of radial vibration for different machine conditions
Machine 1 Machine 2 Machine 3 Machine 4 Machine 5
Healthy condition 0.7487 0.6442 0.5993 0.6755 0.8143
Unbalanced supply 0.6792 0.7451 0.6598 0.7027 0.8569
Single phasing 0.9914 0.9813 0.9380 0.9017 0.9540
Mechanical unbalanced 0.4678 0.6519 0.5406 0.5581 0.7276
Faulty bearing (dry) 0.8802 0.8536 0.7677 0.7844 0.9028
Faulty bearing (ball defect) 1.1275 1.0579 1.0231 1.1282 1.3382
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221202
Fig. 4. Sample of machine terminal voltage with average of 18 runs.
Fig. 5. (a and b) Two segments of radial vibration signal; (c) correlation among them.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221 203
filter is tuned to pass a frequency band of 2f9/4 Hz (f�/
supply frequency) i.e. 96�/104 Hz for 50 Hz supply
frequency. This band indicates the supply conditions i.e.
unbalanced supply, turn-to-turn short, and single phas-
ing. The third frequency band is selected by the filter to
pass a frequency band of 220�/400 Hz, which covers the
high order harmonics of bearing characteristic fre-
quency. The filter is then tuned to pass a frequency
Fig. 6. (a and b) Two vibration signal obtained from healthy machine; (c) correlation among them.
Fig. 7. (a and b) Two vibration signal obtained from healthy and faulty machine; (c) correlation among them.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221204
Fig. 8. (a and b) Terminal voltages of healthy and faulty machine; (c) correlation among them. (d and e) Line currents of healthy and faulty machine;
(f) correlation among them.
Fig. 9. (a) Vibration signal of faulty machine; (b) filtered version.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221 205
band of 550�/950 Hz, this band is related to the
vibration of electromagnetic origin i.e. rotor and stator
slot harmonics. Later to filtering the mentioned bands,
RMS values of these bands are calculated and then used
in the diagnostic algorithm. Table 4 presents the RMS
values of the selected bands for different machine
conditions. It can be observed from the Table 4 that
the RMS value of some frequency bands is affected with
the supply condition while the others have relations with
machine conditions. The effect of changing the operat-
ing frequency from 25 to 50 Hz in steps of 5 Hz on the
RMS values of these frequency bands for healthy and
faulty machines is also included and presented in Table
5. The change in the supply frequency leads to change in
the machine speed, so far all speed dependent frequency
harmonics and supply frequency dependent harmonics
change their location in the spectrum. The RMS value
of different frequency bands change with changing the
supply frequency. For small frequency band such as thesecond band presented in Table 5, the change in supply
frequency must be considered for correct use of the
RMS value for diagnosis purpose.
7. Implementation of signal processing technique in
frequency domain
Frequency analysis of a signal highlights many
important hidden features and extracts some useful
information. The accuracy of information extraction
depends upon the nature of the signal and the method of
analysis. In the present work, FTs and short term FT
Fig. 10. (a) Vibration signal and filtered version; (b) (10�/200 Hz) band pass; (c) (98�/102 Hz) band pass; (d) (680�/850 Hz) band pass.
Table 4
RMS value (g) of selected frequency bands of radial vibration
Healthy condition Unbalanced supply Single phasing Mechanical unbalanced Faulty bearing (dry) Faulty bearing (ball defect)
1�/200 Hz 0.02966 0.06506 0.25115 0.01343 0.01582 0.18849
96�/104 Hz 0.00770 0.08155 0.32252 0.01607 0.03263 0.08412
220�/400 Hz 0.01617 0.01486 0.03942 0.01293 0.02255 0.13093
550�/950 Hz 0.00204 0.00250 0.00993 0.00176 0.00300 0.02211
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221206
(STFT) are used to analyse the vibration, current, and
voltage signal in frequency domain.
7.1. Implementation of spectrum averaging
The implementation of signal averaging for vibration
signal in time domain has serious drawbacks due to the
instability of the signal and the difficulties in getting a
reference point. However, spectrum averaging (or scan
spectrum averaging) is used to obtain the averaging of a
long signal in frequency domain with adequate results.
The long signal is divided into number of equal length
segments and then the spectrum of each segment is
obtained. The average spectrum is then obtained by
adding all the spectrums and dividing by the number of
segments. Fig. 11 shows the vibration spectrum of three
segments and the average spectrum of 35 segments. It is
noticed that the effect of random vibrations and noise
are eliminated in the average spectrum. The average
spectrum is considered for the implementation of
harmonics analysis method. Scan spectrum averaging
method is also applied for vibration signals obtained by
repeating the same experiment for number of times at
same conditions. Fig. 12 shows the vibration spectrum
of three runs and the average spectrum of nine runs of
an induction machine running under same conditions.
Table 5
Effect of varying the supply frequency on the RMS value (g) of selected frequency bands of the vibration signals
Supply frequency (Hz) Healthy machine Faulty machine
1�/200 Hz 96�/104 Hz 220�/400 Hz 550�/950 Hz 1�/200 Hz 96�/104 Hz 220�/400 Hz 550�/950 Hz
25 0.01003 0.05000 0.05345 0.00667 0.19683 0.06191 0.05210 0.00816
30 0.01228 0.08660 0.09856 0.00816 0.24361 0.11475 0.04855 0.00943
35 0.03751 0.09292 0.07221 0.00816 0.36574 0.14478 0.09103 0.01700
40 0.01228 0.06892 0.10690 0.00943 0.25539 0.14634 0.06655 0.01333
45 0.01418 0.10488 0.12101 0.01247 0.10959 0.14732 0.09449 0.01563
50 0.01585 0.10042 0.13283 0.01700 0.18849 0.10412 0.13093 0.02211
Fig. 11. (a, b and c) Spectrums of a windowed long vibration signal; (d) average of 35 spectrums.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221 207
This approach has one drawback that is the effect of
window transitions. However, due to the finite length of
the window function, the ends of the convoluted
vibration signal may be distorted. Overlapping between
the windows may reduce this effect.
7.2. Implementation of correlation
Correlation in frequency domain can be achieved by
direct multiplication of the two spectrums after taking
the conjugate of one of them. The spectrum of the same
signals used in the implementation of the correlation in
time domain is used here as in the Figs. 13�/16. It can be
noticed that the results of correlating two vibration
signals in frequency domain give more informationabout the similarity of the two spectrums, in comparison
with same signals in time domain. Although, implement-
ing correlation with vibration signal does not give
adequate results due to change in the signal spectrum.
For current and voltage signals correlation coefficient
are more beneficial.
7.3. Implementation of signal filtering
Signal filtering in frequency domain is much easier to
implement than in time domain. It is simply achieved by
multiplying the spectrum of the signal with that of the
rectangular window. The width of the window is equal
to the bandwidth of the filter, while the centre frequency
of the window specifies the location of the filter in the
band. One of the main advantages of this technique is
the possibility of filtering only one frequency compo-
nent. The accuracy of the filtering depends upon the
resolution of the spectrum. Fig. 17 shows the vibration
spectrum of a healthy induction machine before and
after filtering some frequency components using vari-
able length rectangular window. The spectrum of the
machine current before and after filtering the
fundamental frequency is shown in Fig. 18. The best
results are achieved when the filtered spectrum compo-
nents are equal to integer multiple of the frequency
resolution.
7.4. Spectrum analysis
In the present work, FFT algorithm is used to
perform discrete Fourier transform (DFT) for the
vibration, voltage, and current signals. The time domain
signals and their spectrums are shown in Fig. 19. For
each type of signal, there are different techniques for
extracting the important features for diagnosis.
Fig. 12. Vibration spectrums of different runs and average of nine runs.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221208
Fig. 13. (a and b) Spectrums of two vibration segments of healthy induction machine; (c) cross correlation among them.
Fig. 14. (a and b) Vibration spectrums of healthy induction machine obtained from different runs; (c) cross correlation among them.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221 209
Fig. 15. Vibration spectrums of (a) healthy machine, (b) faulty machine, (c) cross correlation among them.
Fig. 16. (a and b) Voltage spectrum of two machine; (d and e) the corresponding line currents; (c and f) the voltages and currents correlation,
respectively.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221210
Fig. 17. (a) Healthy machine vibration spectrum and the filtered version; (b) (10�/200 Hz) band pass; (c) (98�/102 Hz) band pass; (d) (680�/850 Hz)
band pass.
Fig. 18. (a) Machine current spectrum; (b) spectrum after filtering the fundamental harmonic.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221 211
7.4.1. Vibration analysis
In order to achieve earliest possible recognition of the
defect in a machine, a comparison of the spectrum of the
machine under study with the spectrum of a typical
(healthy) one must be performed. However, four
approaches of comparison are used.
7.4.1.1. Narrow band analysis. This type of analysis is
performed using high-resolution in frequency. By in-
creasing the resolution of the spectrum, more details
about the frequency contents of the signal can be
achieved, but the stability of the spectrum becomes
poorer. Frequency resolution of 0.1, 1.0, 2.0 Hz is used
here for the purpose of comparison. Fig. 20 shows the
vibration spectrums with different resolutions. It can be
noticed that, although high frequency resolution gives
more detail information, any small change in the signal
(i.e. small random vibration and noise) makes a
difference in the individual lines and thereby in the
spectrum as a whole. In addition, if there is a change in
the rotation speed, all the rotation dependent frequency
components will change in the spectrum (i.e. 1% change
in the speed cause the 1�/ components to shift its
location by 1 and so on), so that direct comparison is
difficult to achieve. Frequency resolution of 1 Hz for
vibration signal gave acceptable detail and stable
spectrum for most of the studied cases.
However, from the literature some frequency compo-
nents related to some types of faults, and from the
calculation of the machine vibration another set offrequency components are obtained. The amplitudes of
these components are used to specify the degree of fault
for the certain operating condition. The vibration
spectrums for mechanical unbalance, supply unbalance
and single phasing are presented in Fig. 21. It can be
observed that the first rotational harmonic has a
dominant value in the three spectrums, which is higher
than the baseline value. In this case, it is difficult torecognise the type of the fault, hence another frequency
components must be considered for comparison. In
addition, by including another machine quantities such
as currents and voltages and the expert’s knowledge,
enhancement of the system ability for diagnosis may be
achieved.
7.4.1.2. Variable band harmonic analysis. In this ap-
proach, selected frequency components from the vibra-
tion spectrum are used for comparison. It has been
noticed that the locations of these selected components
rapidly change. These changes are due to the random
vibrations and change in machine speed. Hence, narrow
band analysis may not be suitable for the comparisonbetween the two spectrums. In the present approach, a
small band of frequency is considered instead of
individual frequency components. The bands used are
Fig. 19. (a, b and c) Voltage, current and vibration signals, respectively; (a?, b? and c?) the corresponding spectrums.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221212
Fig. 20. (a, b and c) Vibration spectrums for different value of frequency resolution 0.1, 1 and 2 Hz, respectively; (a?, b? and c?) zoomed version.
Fig. 21. Vibration spectrum correspond in various machine conditions; (a) healthy machine; (b) mechanical unbalanced; (c) unbalanced supply; (d)
single phasing.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221 213
a percentage of the order of the frequency component.
Different percentages of the frequency components are
tried to select the optimum value that gives acceptable
results for the comparison between different types ofmachine fault. For example, if the selected frequency is
equal to 100 Hz and the percentage is 2%, the band will
include the frequency components from 98 to 102 Hz
and so on for other frequencies. After selecting the
width of the band, three procedures are used for
obtaining the compared value of the band. The first
one is achieved by calculating the average value of the
band components and then using this value for compar-ison with the corresponding value of other spectrums.
The second one is achieved using the highest peak in the
band. The third procedure uses the energy of the band
for comparison with the energy in the corresponding
band in the reference spectrum. It is observed that all the
procedures give good results in the case of medium and
high degree of faults. For small degree of fault,
uncertain information may be obtained, especially ifthe level of random noise is high and there is fluctuation
in the speed, load and supply voltage
7.4.1.3. Band frequency measurement. This approach
can be used to give a primary investigation about the
status of the machine. Bands of frequency in the
vibration spectrum are selected according to the origin
of the fault. The RMS value of these bands is used to
specify the degree and the origin of the faults bycomparing them with the corresponding bands in the
reference spectrum. From the literature and experimen-
tal observations, four frequency bands are selected to
cover the vibration harmonics of mechanical and
electromagnetic origin. The problem in using this
approach arises with some type of faults, where the
vibration harmonics and its multiples may cover wide
range of the spectrum especially at high degree of faultsand that may increase the uncertainty of the obtained
information.
7.4.1.4. Spectrum masking. As mentioned earlier, if
there is a level of random change in the signal or change
in the speed, narrow band analysis approach may fail to
provide accurate information. This problem may be
recovered with the help of Spectrum Masking technique.The technique involves two steps. In the first, a new
spectrum is formed by adding the energy present in a
number of narrow frequency bands of the original
spectrum. The width of narrow frequency band is equal
to a fixed width. Fig. 22 shows the original spectrum and
the new-formed spectrum. The bands of the new
spectrum must be wide enough to encompass the
variation in the signal. From the new generated spec-trum, another spectrum is generated to compensate the
change in the speed. Taking each segment of the
spectrum and pushing one bandwidth to either side as
shown in Fig. 23 obtain this spectrum. The minimum
level of the spectrum components is fixed to a threshold
level. The new spectrum is used successfully to compen-
sate up to 3.5% change in the speed. If the speed changesare more than the amount corresponding to the selected
bandwidth, the vibration harmonics will be lied outside
the limits of the mask. Fig. 24 presents the spectrum
masking for healthy machine and for the machine with
defected bearing. The comparison using these two
spectrums is much easier than using the original
spectrums.
7.4.2. Voltage and current signals
Frequency analysis of electrical variables of the
machine has been used to predict the machine condition.
Current harmonics can be related to most types of the
machine faults [32�/34]. In the present investigation,
current harmonics are examined to demonstrate the
relation between current harmonics and the machine
health. Fig. 25 shows the current waveform and its
spectrum for faulty and healthy induction machine. Itcan be noticed that the spectrum becomes smoother in
the case of a fault. This is due to the interaction (adding
and subtracting) between the mmf space harmonics and
the harmonics developed by some types of machine fault
[35]. The effect of feeding the machine from PWM
inverter with filtered output on the machine current
waveform and the spectrum for healthy and faulty
condition are given in Fig. 26. It can be noticed thatin the case of non-sinusoidal supply such as PWM
inverter it is difficult to distinguish between the current
harmonics due to the supply harmonics and the
harmonics originated by the machine conditions. It
may be noticed that a precise monitoring system is
needed to distinguish between the low-level frequency
components and the high-level supply frequency com-
ponent. This method needs to adjust the samplingfrequency so that the interest frequency component is
equal to integer multiple of the frequency resolution
[36,37].
Terminal voltage harmonics are examined to find the
exact value of the supply frequency, which is needed by
the diagnostic algorithm. In addition, the harmonic
analysis of the terminal voltage is used to predict the
harmonics of the machine. This is achieved by recordingthe machine terminal voltage directly after switching off
the machine i.e. induced machine voltage. The induced
voltage harmonics are same as air-gap flux harmonics
that can be related to the machine condition. Fig. 27
shows the induced voltage waveform of healthy and
faulty machine and their spectrum. For the same reason
mentioned above the faulty spectrum of the induced
voltage became smoother.The harmonics analysis of the machine vibration can
be used to detect wide range of the machine faults. The
change in the amplitude of current and voltage harmo-
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221214
Fig. 22. Spectrum masking of the vibration signal using (a) band energy; (b) peak value.
Fig. 23. Vibration spectrum with spectrum masking.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221 215
Fig. 24. (a) Vibration spectrum and the modified mask of healthy induction machine; (b) vibration spectrum and the modified mask of faulty
induction machine.
Fig. 25. Effect of machine condition on current harmonics (a) healthy machine; (b) machine with faulty bearing.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221216
Fig. 26. Effect of machine condition on the current harmonics under non-sinusoidal supply. (a) Healthy machine; (b) machine with faulty bearing.
Fig. 27. Effect of the machine condition on the induced voltage harmonics. (a and b) Induced voltage of healthy and faulty machine, respectively; (a?and b?) the corresponding spectrums.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221 217
nics due to the machine faults have a small level and the
harmonic are very close to other spectrum harmonics,
hence can not be easily detected. The harmonic distor-
tion in current and voltage is used to assist the
diagnostic task in obtaining the machine condition.
7.5. Implementation of short term Fourier transform
STFT is used to estimate the frequency contents of the
non-stationary signals. STFT is dividing the signal into
a small segment using a window function (W) and
perform FFT for each segment. However, the vibration
signal picked up from the machine has two components;
stationary and non-stationary. Non-stationary part has
no relation with most of the studied faults. STFT is
applied to eliminate the non-stationary part by adding
the spectrum of all windowed segments and dividing it
by the number of segments as in case of spectrum
averaging.
The window function used in this investigation is the
rectangular window, which is simple to implement and
has some useful characteristics such as narrow main lobe
4p /(2N�/ 1). The spectrum of the rectangular is shown
in Fig. 28. It can be noticed that there are several side
lobes at both ends of the window. These side lobes may
give uncertain information in the vibration spectrum
and cause Gibbs phenomena [29]. Making overlap
between the successive windows can reduce this effect.
In this investigation, the number of samples used is
50 000 and the width of window is equal to 4000 samples
with overlap equal to 1200 samples. Then an average
spectrum is obtained using the spectrums of the wind-owed data. Fig. 29 shows the vibration signal, the
spectrums of the signal of different window locations
and the average spectrum. Fig. 29(b) shows that the
effect of random vibration and the noise has reduced in
the average spectrum. This procedure is repeated for all
experimental data and the obtained spectrums are used
to extract the required information and to state the
health of the machine.
8. Isolation of random vibrations
Any vibration signal obtained from electromechanical
systems contains a level of random changes. These
random changes in the measured signal may be due tothe random vibrations. These random vibrations can be
related to the health of the machine for some faults such
as dry bearing fault or bearing ageing. If these random
vibrations could be isolated from the measured signal,
useful information about bearing health may be ob-
tained.
From experimental observations, it is noticed that
there are some changes in the level of RMS values ofvibration signals obtained from the successive segments
of a long record signal or from repeatable tests
Fig. 28. Frequency response of the rectangular window.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221218
performed on induction machine subjected to dry
bearing fault running under similar conditions. The
presence of dry bearing fault at an early stage introduces
a small level of randomness in the vibration signal.
Therefore, it is difficult to isolate these random vibra-
tions by traditional signal processing techniques. The
difficulty in discovering such fault is that neither thevariation in vibration level nor additional spectrum
components is detectable In order to isolate these
random vibrations, the algorithm presented in Fig. 30
is used with some modifications. In this algorithm, the
RMS value of each segment of the vibration signal is
calculated but not averaged as shown in Fig. 30. These
values are obtained from the same machine running
with the same condition. The fluctuation in the RMSvalues is nothing but the random vibrations.
9. Conclusion
In this paper, the treatment of raw data obtained from
physical machine parameters are presented. The imple-
mentations of various DSP and analysis techniques in
time and frequency domain with different machinevariables are given. Signal detectors such as mean
square value and crest factor are used with the voltages,
currents and vibration signals in time domain. In
addition, signal filtering, signal averaging and correla-
tion are used in time and frequency domain simulta-
neously. FTs and STFTs are used to present the time
domain signal in frequency domain. The obtained
vibration spectrum is analysed using narrow band,
variable band, selected band and spectrum masking
approaches. The data implemented with mentioned
techniques covers different operating conditions of the
machine under test. The obtained information is used as
input to the neural network.
Time domain analysis is used to obtain the machine
condition qualitatively using correlation coefficient,
RMS value, and crest factor. For primary investigation,
time domain analysis provides a rough figure, and is not
appropriate for fault classification and ranking. Fre-
quency domain analysis of the vibration signal provides
detailed information. The vibration harmonics are
related to types of machine faults. The machine condi-
tion can be obtained by comparing the amplitude of
these harmonics with those obtained from correspond-
ing ones in the healthy machine.
The traditional treatment of vibration spectrum
fluctuations is the averaging, which may lead to hide
some features of short duration. The alternative ap-
proach to such non-stationary vibration signal is the
Wavelet transform that can provide useful information
about any signal in time domain with different bands of
Fig. 29. (a) Implementation of STFT using moving rectangular window; (b) average spectrum.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221 219
frequencies. WT gives variable time resolution for
different frequency bands rather than STFT, which
gives constant resolution.
The area of condition monitoring and diagnostic is
very wide and includes many topics. It is suggested to
make some improvements in the monitoring system
through the use of the following:
. Inclusion of mmf harmonics in induction machine
model so that the relation between the machines
space harmonics and vibration harmonics can be
established. This can be made using machine per-
meance approach. An alternative to this is the
modelling of electromagnetic behaviour of the ma-
chine using finite element approach.
. Estimation of machine parameters (resistance and
inductance) through on-line modelling of induction
machine. The estimated values of machine para-
meters can give indication about the machine health
through a comparison with healthy one.
. Building a database for vibration harmonics using
experimental and theoretical investigations for var-
ious size and design of standard of three phase
induction motors. Through this data base, a new
standard for vibration can be established instead of
the traditional one, which depends upon RMS
velocity of vibration rather than harmonic amplitude.
. Employing expert system for fault diagnosis of
induction motor using rules obtained from the
connection weight of a supervised neural network
and rules extracted from the heuristic knowledge.
This combination of ANN knowledge and expert’s
knowledge may enhance the accuracy and efficiency
of the monitoring system for diagnosis.
Fig. 30. Isolation of random vibrations from different wavelet decompositions.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221220
References
[1] A. Ahmad, D.P. Kothari, A review of recent advances in
generator maintenance scheduling, Electrical Machines and
Power Systems 26 (1998) 373�/387.
[2] K.G. Boving, NDE Handbook for Condition Monitoring, Butter-
worths, Denmark, 1989.
[3] P.J. Tavner, J. Penman, Condition Monitoring of Electrical
Machines, Wiley, UK, 1987.
[4] Plant Vibration Monitoring, Monitran Ltd., Monitor House,
Hazlemere Road, Penn, Bucks, HP10 8AD, (Technical Note).
[5] W. Barry, S. Wade, M.W. Dunnigen, How to track rolling
element bearing health with vibration signal analysis, Sound and
Vibration 25 (11) (1991) 133�/139.
[6] V.I. Chaban, Y.V. Kharchenko, V.I. Kozubash, An algorithm for
calculating the vibration characteristics of an induction motor
rotor, Electric Technology 1 (1) (1990) 34�/43.
[7] A.J. Ellison, C.J. Moore, Acoustic noise and vibration of rotating
electric machines, Proceedings of IEE 115 (11) (1968) 1633�/1640.
[8] F. Ishibashi, S. Noda, M. Mochizuki, Numerical simulation of
electromagnetic vibration of small induction motors, IEE Pro-
ceedings, Electrical Power Applications 145 (6) (1998) 528�/534.
[9] O. Prakash, K.R. Pandu, Experimental Studies of Vibration and
Noise in Centrifugal Pumps, Indo-US Symposium on Emerging
Trends in Vibration and Noise Eng., Delhi, March 1996, pp. 197�/
204.
[10] J. Qiu, The Vibrations of Mechanical and Electromagnetic
Coupling of an AC Motors in the Starting Process, Institution
of Mechanical Engineers International Conference on Vibration
in Rotating Machinery, Bath, September 1992, pp. 523�/527.
[11] S.P. Verma, A. Balan, Experimental investigations on the stators
of electrical machines in relation to vibration and noise problems,
IEE Proceedings, Electrical Power Applications 145 (5) (1998)
455�/461.
[12] B.I. Zubrenkov, A.I. Kaplin, V.S. Malyshev, M.F. Manyukov,
Analysis of induction motor vibration during start-up, Electric
Technology 1 (4) (1990) 87�/94.
[13] J.R. Cameron, W.T. Thomson, A.B. Dow, Vibration and Current
Monitoring for Detecting Air-gap Eccentricity in Large Induction
Motors, Second International Conference on Electrical Machine
Design and Applications, September 1985, pp. 17�/19.
[14] G.M. Joksimovic, J. Penman, The detection of inter-turn short
circuit in the stator windings of operating motors, IEEE
Transactions on Industrial Electronics 47 (5) (2000) 1078�/1084.
[15] J. Penman, J.R. Smith, W.E. Bryan, Software Based Approach to
Machine Condition Monitoring System, Second International
Conference on Electrical Machine Design and Applications,
September 1985, pp. 1�/9.
[16] J. Penman, C.M. Yin, Feasibility of using unsupervised learning,
artificial neural networks for the condition monitoring of
electrical machines, IEE Proceedings, Electrical Power Applica-
tions 141 (6) (1994) 317�/322.
[17] J. Penman, M.N. Dey, A.J. Tait, W.E. Bryan, Condition
monitoring of electrical drives, IEE Proceedings, Electrical Power
Applications, Part B 133 (3) (1986) 142�/148.
[18] H.C. Stanley, An analysis of induction machine, AIEE Transac-
tions 57 (1938) 751�/757.
[19] P.C. Krause, Analysis of Electrical Machinery (Book), McGraw-
Hill, 1986.
[20] E. Levi, Magnetizing Flux Saturation in Variable Speed Inverter-
Fed induction Motor Drives’, IEE Conference on Power Electro-
nics and Variable Speed Drives, UK, October 1994, pp. 78�/83.
[21] T.A. Lipo, P.C. Krause, Stability analysis of a rectifier inverter
induction motor drive, IEEE Transactions on Power Apparatus
and System PAS-88 (1) (1969) 55�/66.
[22] A.K. Sarkar, G.J. Berg, Digital simulation of three phase
induction motors, IEEE Transactions on Power Apparatus and
System PAS-89 (6) (1970) 1031�/1036.
[23] G.R. Slemon, Modeling of induction machines for electric drives,
IEEE Transactions on Industry Applications 25 (6) (1989) 1126�/
1131.
[24] M. Sokola, E. Levi, Iron Losses in Current-Controlled PWM
Inverter Fed Induction Machines, IEEE MELECON96, Bari,
Italy, 1996, pp. 361�/364.
[25] A.J.M. Cardoso, S.M.A. Cruz, D.S.B. Fonseca, Inter-turn stator
winding fault diagnosis in three-phase induction motors by park’s
vector approach, IEEE Transactions on Energy Conversion 14 (3)
(1999) 595�/600.
[26] N. Auther, J. Penman, Induction machine condition monitoring
with higher order spectra, IEEE Transactions on Industrial
Electronics 47 (5) (2000) 1030�/1041.
[27] M.N. Bandyopadhyay, Y.R. Sood, Noise analysis of rotating
electrical machines utilizing tools of digital signal processing, IE
(I) Journal EL 80 (1999) 66�/69.
[28] M. Benbouzid, H. El, A review of induction motors signature
analysis as a medium for fault detection, IEEE Transactions on
Industrial Electronics 47 (5) (2000) 984�/993.
[29] M. Benbouzid, H. El, H. Nejjari, R. Beguenane, M. Vieira,
Induction motor asymmetrical faults detection using advanced
signal processing techniques, IEEE Transactions on Energy
Conversion 14 (2) (1999) 147�/152.
[30] N. Mohanti, Signal Processing Signal Filtering Detection (Book),
Van Nostrand Reinhold, New York, 1987.
[31] T.M. Romberg, J.L. Black, T.J. Ledwidge, Signal Processing for
Industrial Diagnostic (Book), Wiley, Chichester, 1996.
[32] G.B. Kliman, J. Stein, Methods of motor current signature
analysis, Electric Machines and Power Systems 20 (1992) 463�/
473.
[33] R.R. Schoen, B.K. Lin, T.G. Habelter, J.H. Schlag, S. Farag,
Unsupervised on-line system for induction motor fault detection
using stator current monitoring, IEEE Transactions on Industry
Applications 31 (6) (1995) 1280�/1286.
[34] R.R. Schoen, T.G. Habelter, F. Kamran, R.G. Bartheld, A new
method of current-based condition monitoring in induction
machines operating under arbitrary load conditions, Electric
Machines and Power Systems 25 (3) (1997) 141�/152.
[35] P.L. Alger, The Nature of Polyphase Induction Machine, Wiley,
London, 1951.
[36] Y.S.H. Sam, R.A. Langman, Stator current frequency analysis for
the condition monitoring of induction motors. Part I: fixed supply
frequency, Journal of Electrical and Electronics Engineering,
Australia 17 (1) (1997) 43�/56.
[37] Y.S.H. Sam, R.A. Langman, Stator current frequency analysis for
the condition monitoring of induction motors. Part I: variable
supply frequency, Journal of Electrical and Electronics Engineer-
ing, Australia 17 (1) (1997) 57�/70.
S.A.S. Al Kazzaz, G.K. Singh / Electric Power Systems Research 65 (2003) 197�/221 221