State of the Art in Monitoring

Rotating Machinery Part 2

Robert B. Randall, The University of New South Wales, Sydney, Australia
In the last thirty years there have been many developments
in the use of vibration measurement and analysis for monitoring the condition of rotating machinery while in operation.
These have been in all three areas of interest, namely fault
detection, diagnosis and prognosis. Of these areas, diagnosis
and prognosis still require an expert to determine what analyses to perform and to interpret the results. Currently much
effort is being put into automating fault diagnosis and prognosis. Major economic benefits come from being able to predict with reasonable certainty how much longer a machine can
safely operate (often a matter of several months from when
incipient faults are first detected). This article discusses the
different requirements for detecting and diagnosing faults, outlining a robust procedure for the former, and then goes on to
discuss a large number of signal processing techniques that
have been proposed for diagnosing both the type and severity
of the faults once detected. Change in the severity can of course
be used for prognostic purposes. Most procedures are illustrated using actual signals from case histories. Part 1 of this
article appeared in the March 2004 issue of S&V.
AR Models. AR or autoregressive models are more efficient
where there are sharp spectral peaks, and thus the required
transfer function has poles. This is the case with IIR (infinite
impulse response) filters where outputs are generated recursively from the previous outputs and the current input.
The relationship between input and output signals can be
expressed as:
N

y i = - ak y i - k + x i

(11)

k =1

vided a sufficient degree of spectrum interpolation is used. The

spectrum interpolation was achieved by padding the data
record with zeros to seven times its original length. Note that
the maximum entropy spectrum had to be represented on a
logarithmic amplitude scale because of the much wider range
of amplitude values than for the Fourier analysis cases.
AR modeling has recently been applied to the detection of
local faults on gears.13 A model is developed for the signal from
an undamaged section of a gear using linear prediction, and
when a damaged section of the gear enters the mesh, the actual value of the signal departs dramatically from the linearly
predicted value and the error gives a measure of the local
change. Fault indication using AR modeling gives better results
than the residual method used earlier.13 The latter is based
on removing the normal gearmesh signal obtained by synchronous averaging.
AR and MA models can be combined to give so-called ARMA
models with both poles and zeros, but these are used more for
system modeling than spectrum analysis.
Separation of Periodic and Random Signals. Machine signals are often very complex with mixtures of periodic and random signals, and mixtures of signals with different periodicity. It can be advantageous to separate the different components
from each other.
Synchronous Averaging. The most widely used technique for
separating different signals is time synchronous averaging,
which is useful to extract that part of a signal having the same
period as a trigger signal (e.g. a once-per-rev tacho signal from
a shaft in a rotating machine). In practice it is done by averaging together a series of signal segments each corresponding to
one period of the synchronizing signal. Thus:

which after Z-transformation gives:

N -1

Y (z )A(z ) = X (z )

(12)

from which comes the transfer function:

1
=
A(z )

1
N

1
N

ak z - k (1 - z -1pk )

k =0

(13)

k =1

which has no zeros and is an all-pole model.

There are a number of techniques that result in such an AR
model, one of which is the maximum entropy method. In
this, the coefficients are found by maximizing the entropy (disorder) of the signal, while ensuring that the autocorrelation
function is determined by the signal within the window. This
really means that the signal outside the window will be most
similar to the signal within the window because of the elimination of any biasing effect. Other AR techniques include linear prediction and statistical autoregression from which AR
modeling takes its name.
Figure 11 shows the results of applying maximum entropy
analysis to a very short record of envelope signal from a bearing with an inner race fault. 12 The record length comprised
only 1.29 revolutions of the shaft speed which determined the
spacing of modulation sidebands in the envelope spectrum.
The maximum entropy spectrum of Figure 10d appears to give
very good resolution of the sidebands, but Figure 10c shows
that Fourier analysis can give almost as much information proBased on a paper presented at ISMA 2002, the International Conference on Noise and Vibration Engineering, Leuven, Belgium, September, 2002.

10

y a (t ) = 1 / N

y (t + nT )

(14)

n =0

This can be modelled as the convolution of y(t) with a train of

N delta functions displaced by integer multiples of the periodic time T, which corresponds in the frequency domain to a
multiplication by the Fourier transform of this signal. This is
given by the expression:14
C ( f ) = 1 / N sin(N pTf ) sin(pTf )

(15)

The filter characteristic corresponding to this expression is

shown in Figure 12 for the case where N = 8, a comb filter selecting the harmonics of the periodic frequency. The greater the
value of N the more selective the filter, and the greater the rejection of nonharmonic components. The noise bandwidth of
the filter is 1/N , meaning that the improvement in signal/noise
ratio is 10 log10 N dB for additive random noise. For masking
by discrete frequency signals, it should be noted that the characteristic has zeros that move with the number of averages, so
it is often possible to choose a number of averages that completely eliminates a particular masking frequency. The above
characteristic is for an infinitely long time signal y(t). For the
practical situation of a finite length of signal with finite sampling frequency, it is possible to calculate an optimum number of averages to completely remove a discrete masking signal, in particular when the frequency is related by a rational
fraction to the synchronous frequency.14 This is always the case
for different shafts in gearboxes.
For good results the synchronizing signals should correspond exactly with samples of the signal to be averaged. One
sample spacing corresponds to 360 of phase of the sampling

SOUND AND VIBRATION/MAY 2004

0.01
A

0
0
0.01

100

200

300

400

500

600

100

200

300

400

500

600

100

200

300
400
Frequency (Hz)

500

600

0
0
0.01
C

00

Figure 13. Application of synchronous averaging to a mixture of gear

and bearing signals. A spectrum from low speed gear, B spectrum
from high speed gear, C spectrum of residual signal (bearing fault).
0.012
Before Tracking

0
0
0.012

500

After Tracking

Figure 11. Envelope spectra of a very short length of signal (1.29 periods of shaft rotation). A Envelope signal, B Direct FFT spectrum, C
Interpolated FFT spectrum, D Maximum entropy spectrum.

0
0

1000

500
Frequency (Hz)

1000

Figure 14. Use of tracking to avoid smearing of shaft speed related

components.

Figure 12. Filter characteristic for 8 averages.14

frequency, and thus to 144 of phase at 40%, a typical maximum signal frequency. Moreover, even a 0.1% speed fluctuation would cause a jitter of the same order of the last sample
in a (typical) 1k record with respect to the first, and thus an
even greater loss of information at the end of the record after
averaging.
Sampling the signal using a sampling frequency derived from
the synchronizing (tacho) signal as described below solves both
of these problems and is always to be recommended. Figure 13
shows the results of using synchronous averaging on speed
corrected data from a gearbox in a variable speed mining shovel
(see Figure 14). The order tracked data was arranged to have
an integer number of samples per period of the low speed gear,
which allowed determination of the harmonics of this gear
speed by synchronous averaging. The spectrum of this signal
is shown in Figure 14a. After a periodic repetition of this signal was subtracted from the overall tracked signal, the data
were resampled to have an integer number of samples per period of the high speed gear, after which its harmonics could be

SOUND AND VIBRATION/MAY 2004

determined in the same way (Figure 13b). Finally, after subtraction of this periodic signal from the data, the remaining
signal was dominated by the effects of an inner race bearing
fault (Figure 13c).
Order Tracking. In analyzing rotating machine vibrations,
an x-axis based on harmonics or orders of shaft speed is often desirable. This can avoid smearing due to speed fluctuations or show how the strength of the various harmonics
changes over a greater speed range, such as when they pass
through various resonances. For example, if a constant amplitude signal that is synchronous with the rotation of a shaft is
sampled a fixed number of times per revolution, the digital
samples are indistinguishable from those of a sinusoid and thus
give a line spectrum. On the other hand, if normal temporal
sampling is used, the spectrum spreads over a range corresponding to the variation in shaft speed. Thus, for order analysis it is necessary to generate a sampling signal from a tacho
signal synchronous with shaft speed. It is sometimes possible
to use a shaft encoder mounted on the shaft in question to provide a sampling signal, but more often the latter has to be generated electronically. Formerly, this was done using a phaselocked loop to track the tacho signal and then generate a
specified number of sampling pulses per period of the tracked
frequency. However, an analog phase-locked loop has a finite
response time and cannot necessarily keep up with the random

11

Primary Input

Gear I
+
Bearing I

Bearing I
S

Reference Input

Gear II

Adaptive
Filter

Gear I

0
B

fk

Transform size
Negative frequency
components moved
to left of sampling
frequency fs

Figure 15. ANC applied to separation of bearing and gear signals.

0
Zero and positive
frequency components

fs

Transform size
C

0
D

fs/2

Transform size

fs

Figure 17. Block shift procedure for selecting frequency band for demodulation. A original spectrum of one-sided bandpass section, B
Frequency shift by fk (center of passband), C frequency shift by amount
corresponding to lower passband limit (half size transform not recommended), D frequency shift by amount corresponding to lower passband limit (full size transform).
Figure 16. Separation of (additive) gear and bearing signals using DRS.
A combined signal, B deterministic part (gear signal), C stochastic part (bearing signal).

speed fluctuations that occur with an internal combustion engine from cycle to cycle. The best method is to digitally
resample each record based on the corresponding period of the
tacho signal. This can be done in a number of ways, based on
digital interpolation. 15
Quite apart from errors introduced by the interpolation,
when resampling at a lower frequency (for example as a machine speed reduces), it is necessary to ensure that the signal
is adequately lowpass filtered to prevent aliasing. Digital filtering can be useful here as the cutoff frequency varies directly
with the sampling frequency, but the initial analog lowpass filtration must be such that aliasing components do not enter the
measurement range. Digital oversampling can solve this problem, as the sampling frequency can be reduced by a large factor before overlap occurs.
Figure 14 illustrates the use of tracking to avoid smearing in
the spectrum of the vibration signal from a gearbox. The discrete frequency components in the spectrum after tracking
come mainly from gear-related components that were removed
using synchronous averaging as shown in Figure 13.
Adaptive Noise Cancellation. Adaptive noise cancellation
(ANC) is a method for separating two signal components (in a
primary signal) where there is access to another (reference)
signal containing only one of the two components. The reference signal does not have to be identical to the related component in the primary signal, just coherent with it so that they
are related by a linear transfer function. 16 Figure 15 illustrates
the basic principle applied to a situation where the primary
signal contains both gear and bearing (fault) components. These
could be measurements on a faulty bearing where the reference
signal contains only a gear signal, for example measured on a
more remote bearing. The adaptive filter adjusts its coefficients
so as to minimize the power of the error signal, the difference
between the primary and filtered reference inputs. When the
two components are statistically independent, this separates
them and the error signal becomes the bearing signal.
Another development of this, self adaptive noise cancellation (SANC), uses the difference in statistical properties of the

12

two components (as discussed in the section on bearing faults)

to separate them. The deterministic gear signal has a much
longer correlation length than the stochastic bearing signal,
hence the reference signal can be a delayed version of the primary mixture, so that only the deterministic gear signals are
coherent. 8
In separating gear and bearing signals, most often the signals
are stationary (order tracking may have to be used to ensure
this), and therefore there is no need to continually adapt the
filter. In such situations a new method called DRS (discrete/
random separation) has recently been proposed that achieves
virtually the same result much more efficiently.17 First a transfer function is found between the current signal and a delayed
version, which results in a comb filter corresponding to the
discrete frequency components (coherent even with a delay).
The amplitude of this transfer function is then used as a (zero
phase shift) filter to transmit the deterministic part and reject
the random part. Both the filter generation and the subsequent
filtering are done using FFT techniques, making the whole
procedure much more efficient than the SANC approach. The
difference compared with synchronous averaging is that the
discrete frequency components are not required to have harmonic relationships.
Figure 16 shows an application involving separation of a discrete frequency gear signal from an inner race bearing fault
signal, which despite its periodic appearance is actually random. After separation, envelope analysis gives a clear diagnosis of the bearing fault.
Demodulation. Modulation occurs when an otherwise sinusoidal signal, a so-called carrier signal, has its amplitude or
frequency made to vary with time. The first case is amplitude
modulation and the second can be considered frequency or
phase modulation. Phase modulation is the deviation in phase
(angular displacement) from the linearly increasing phase of
the carrier, while frequency modulation is the difference in
instantaneous frequency (angular velocity) from the constant
carrier frequency. Thus, frequency modulation is the derivative of phase modulation. A direct mechanical example of
phase/frequency modulation is shaft torsional vibration. When
expressed in terms of shaft angle, it is a phase modulation.
When expressed in terms of shaft speed, it is a frequency modu-

SOUND AND VIBRATION/MAY 2004

Raw Spectrum (complex)

0.4

Spectrum (complex)

Spectrum (complex)

B
B

C
B B

Zeros
0
fs/2 0
fs/2 0 f's/2 f's
Select frequency band (B)
Frequency Shift band (B)
Double amplitude and pad
for Envelope Analysis
with zeros to double the length

U2

ifft
Envelope Auto-Spectrum
(f)

0
0

Frequency, Hz

100

Ball Pass
Frequency

Envelope (real)

Analytic Signal (complex)

E

fft
5

B
f's/2
f's
Forward Fourier Transform
Obtain amplitude of
to obtain the envelope spectrum
Analytic Signal

U4

Inverse Fourier Transform

into an analytic signal

Figure 19. Procedure for envelope analysis using Hilbert Transform

technique.

0
0

Frequency, Hz

100

Figure 18. Advantage of analyzing the squared envelope (B) compared

with the envelope (A). Note the enhancement of the harmonic series at
15.5 Hz (BPFO).

lation. There is no modulation term for the angular acceleration obtained by further differentiation. A mechanical example
of amplitude modulation is the variation in vibration amplitude at the meshing frequency in a gearbox. The increase in
tooth deflection with load gives an increasing departure from
ideal involute profiles, and often tooth load varies periodically
with the rotation of the gears.
Thus, a generally modulated signal can be represented by:

Am (t )cos (2p fct + fm (t ))

(16)

where Am(t) represents the amplitude modulation function and

fm(t) represents the phase modulation function in radians. The
corresponding frequency modulating function (in Hz) is

1 dfm (t )
2p dt
Equation 16 is the real part of the rotating vector:

Am (t ) exp { j (2p fct + fm (t ))}

(17)

whose modulus is the amplitude modulating function and

whose phase is the phase modulating function plus the linear
carrier component. Thus, if it is desired to demodulate a real
signal such as Equation 16, it is desirable to find the corresponding imaginary part so as to form the complex Equation
17. Provided the fluctuating part of Equation 17,

Am (t ) exp { jfm (t )}

(18)

has a half bandwidth less than the carrier frequency f c , the

spectrum of Equation 17 will be one-sided, and Equation 17
will be an analytic function. In this case the required imaginary part is the Hilbert transform of the real part. As the spectra of the two parts are convolved, the total bandwidth is less
than the sum of the individual bandwidths. The bandwidth of
the amplitude part is directly that of A m(t). Though exp(jf m(t))
is theoretically infinite, if the maximum phase deviation is less
than 1 radian, the effective bandwidth (within the dynamic
range) is less than twice that of fm (t).
Note that a zoom processor (Figure 8) can be used directly
both to extract that part of the spectrum to be demodulated and
to remove the carrier component by zooming at the carrier frequency. Generally, the zoom process results in a considerable
reduction in the sampling rate to be more compatible with the
bandwidth of the modulating functions. The modulus of the

SOUND AND VIBRATION/MAY 2004

complex output from the zoom processor is the amplitude

modulating function, while the argument is the phase modulating function. This may have to be unwrapped to a continuous phase function (i.e., eliminating jumps over 2p), but in
general this is not a problem for well-behaved functions. Demodulating a larger bandwidth decreases the time step, and
thus phase jump, between samples and may facilitate unwrapping.
As illustrated in Figure 17, the same thing can be achieved
using FFT transforms, although the first one will have to be
large to accommodate the high carrier frequency while being
long enough to contain sufficient periods of the lower modulating frequencies. Where phase demodulation is required, the
center of the demodulation band will have to be shifted to zero
frequency (and negative frequency components shifted to the
other end of the frequency record). However, for amplitude
demodulation the result is unaffected by the frequency shift.
Further, it is more convenient to shift the left hand end of the
band to zero frequency and pad the negative frequency side
with zeros thus maintaining an analytic signal. In either case
there should be at least as many contiguous zeros in the spectrum as components since the modulus is the square root of the
amplitude squared and the latter corresponds to the convolution of the spectrum with its complex conjugate reversed endfor-end. The zeros prevent extraneous wrap-around errors.
Envelope Analysis. Figure 6 shows that for diagnostics of
bearing signals, it is advantageous to frequency analyze the envelope of the bearing signal because of the smearing of the raw
spectrum. The envelope signal has traditionally been obtained
by rectification and lowpass filtration, but demodulation of the
analytic signal (of Figure 17a) has an advantage the bandpass
filters used to extract the demodulation band are much sharper
than typical analog bandpass filters and can better isolate the
bearing signal from large adjacent masking components. Moreover, it is generally better to analyze the squared envelope
rather than the envelope, as this improves the signal to noise
ratio of the result.8
Figure 18 shows a typical example from a paper mill bearing with an outer race fault. The shaft speed is less than 2 Hz,
and BPFO only 15.4 Hz. But, the demodulated band was in the
range of 5.2-5.6 kHz where the largest spectral change occurred.
The spectrum of the squared envelope in Figure 18b is much
clearer than the normal envelope spectrum in Figure 18a.
Figure 19 shows the procedure for obtaining the envelope (or
envelope squared) spectrum for a typical bearing fault. The
band B is typically placed where the largest (dB) spectral
change has occurred as a result of the bearing fault, since this
is where the signal to noise ratio is the highest. After shifting
this band to zero frequency and padding with zeros, the new
sampling frequency is just 2B. B should thus be chosen to con-

13

50

Phase Angle, rad

50

100

150

200

Figure 20. Spectrum of ring gear tooth passage signal showing four harmonics and their sidebands.

tain a sufficiently wide range of frequency for diagnostic purposes, typically 4-5 harmonics of the highest fault frequency
(usually BPFI). If the analytic signal represented by the onesided spectrum of Figure 19c is denoted by f(t), its squared envelope can be obtained as f(t) f*(t), meaning that the corresponding spectrum is given by F ( f ) * F * (- f ), where F(f) is the
Fourier transform of f(t). To obtain the envelope, it is necessary to take the square root, which introduces distortion. 8
Phase and Frequency Demodulation. Phase and frequency
demodulation is illustrated by the example of detecting misfire in an internal combustion (IC) engine by the pattern of angular velocity in the torsional vibration of a spark ignition engine. This is obtained by frequency demodulation of a shaft
encoder signal, but the latter can be the pulses from a proximity probe detecting the passage of teeth on the ring gear.
Figure 20 shows the spectrum of such a signal. In principle,
any harmonic of the tooth-pass frequency can be demodulated,
as the speed variation gives rise to a fluctuation in the time
intervals between pulses, which gives rise to a phase modulation proportional to the order of the harmonic demodulated,
but otherwise of the same shape. To calibrate the result in terms
of shaft angle, it is necessary to divide the measured phase signal by the shaft order demodulated. In this case the first
toothpass harmonic was demodulated since it had the best signal to noise ratio.
Figure 21 shows the resulting phase signal where the slope
means that the correct carrier frequency has not been chosen.
If required, the slope can be removed by a detrend operation.
To obtain the angular velocity it is necessary to differentiate
the phase (angular displacement). This is best achieved by multiplying by jw in the frequency domain since a bandpass filtration can be performed at the same time. Figure 22 shows the
result of doing this with the phase signal of Figure 21.
In the angular velocity diagram of Figure 22, it is obvious that
there is a misfire on cylinder 6 (actually caused by detaching
the spark plug lead), as there is a rapid drop in speed where it
should fire, which is gradually built up by the firing of the remaining cylinders. In a number of student projects, different
faults were introduced and it became clear that a misfire could
always be detected; although the reason for the misfire could
not. Typical reasons for misfire that gave similar results were:
1. Failure of spark.
2. Faulty injection to one cylinder.
3. Simulated burnt valve by using an oversize pushrod. In this
case the misfire was noticeably different because not only
was the combustion pressure missing, but also the compression pressure.
4. A leaky valve simulated by using a loose spark plug. In this
case there was a partial misfire that could in fact be quanti-

14

0.2

0.4
Time, sec

0.6

0.8

Figure 21. Unwrapped phase from demodulating the first harmonic of

Figure 20.

fied by the smaller jump in angular velocity for firing on cylinder 6 compared with the other cylinders.
The transmission error (TE) signal for a gear pair (as defined
in gear faults section) can be obtained by demodulating the signals from a shaft encoder attached to each gear. 18 The TE signal can itself then be demodulated (with the toothmesh frequency as carrier) to highlight local faults. A similar
demodulation carried out on acceleration signals was able to
reveal a tooth root crack in a helicopter gearbox long before it
became apparent using other techniques.19

Cyclostationary Signal Analysis

Next, the subject of cyclo-stationarity is examined in more
detail, because of its relative novelty and its importance in
separating different components in machine vibration signals.
Cyclostationary signals are defined as those whose second
order statistics, such as their autocorrelation function, vary
periodically. Strictly speaking, this defines second order
cyclostationarity, as periodic signals can be considered as first
order cyclostationary. Thus the bivariate autocorrelation function R x(t,t), defined by:
(19)
Rx (t, t ) = E x t + t 2 x * t - t 2

has the property that:

) (

Rx (t, t ) = Rx (t + T , t )

(20)

where T is the periodic time.

Note that this definition is similar to that commonly used
for stationary signals, except that (by definition) the statistics
of the latter do not vary with time, and so it is expressed as a
function of time shift t only. Furthermore, it is common to displace only one of the signals by t rather than make the displacement symmetrical about time t. However when the function varies with t it is better to use the symmetrical displacement to
better assign the value to time t.
The autocorrelation function itself can be hard to interpret
and so it is often further processed. A Fourier transform of the
two-dimensional diagram in the t direction gives a result called
the instantaneous power spectrum, which is related to the
Wigner-Ville distribution (WVD), a series of spectra varying
with time.20 The WVD is typically applied to a single time
record and thus does not include the expected value or averaging operation in Equation 19. For cyclostationary signals,
the instantaneous spectra vary periodically with time, so it is
beneficial to perform a further Fourier transform (actually a
Fourier series) with respect to time t in order to quantify this
periodicity. The result is known as the spectral correlation,
and is used in some results given below. The formula is thus:

Sx (a , f ) =

( Rx (t, t ))

t a t f

(21)

SOUND AND VIBRATION/MAY 2004

Frequency f, Hz

500
400

300
200
100
0

50

100

150

200

250

300

350

400

450

500

Cyclic Frequency a, Hz

Integrated CSD

x 106
3
9.5 Hz

2.5

120 Hz

240 Hz

2
1.5
1
0.5
0

50

100

150

200

250

300

350

400

450

500

Cyclic Frequency a, Hz

where f represents normal frequency, and a represents so-called

cyclic frequency.
The integral of the spectral correlation over all f in fact gives
the same result as the Fourier analysis of the expected value
of the squared signal, and so is very closely related to envelope analysis.21 If x(t) is an analytic signal,8 the result is the
spectrum of the squared envelope, which as shown above
gives better results for bearing diagnostics than the simple envelope obtained by amplitude demodulation.
Machine signals that repeat cyclically but are not directly
phase-locked to shaft speeds are often cyclostationary. An example already mentioned is the signals associated with combustion in an IC engine. These can be decomposed into a periodic component (the expected value) and a second order
cyclostationary component that is like amplitude modulated
noise, so that its squared envelope is periodic. Rolling element
bearing signals were modelled as cyclostationary 21 because of
the random slip associated with the approximate periodicity.
The signals are only approximately cyclostationary (termed
pseudo-cyclostationary), but can usefully be treated as
cyclostationary as a first approximation. 6 This is because the
ballpass frequencies are not well defined because of the slip,
as opposed to engine signals, which are constrained to the
engine cycle frequency.
Figure 23 21 shows the spectral correlation for an inner race
fault in a ball bearing, demonstrating that the harmonics of
BPFI and sidebands spaced at shaft speed are distributed with
f but discrete with respect to a. The integral of this over all f
gives the same result as the spectrum of the squared envelope.
From this it would appear that the spectral correlation does
not give any benefit over (squared) envelope analysis, but there
is one situation where it does, illustrated in Figure 24. This is
the case of an inner race fault in a helicopter gearbox that was
detected so late that the spalling had extended about 1/3 of the
way around the race and had become smoothed. This meant
that high frequency pulses at BPFI were no longer generated,
but envelope analysis revealed strong modulation at shaft
speed. This is presumedly due to the varying support of the
gear as the fault passed through the load zone, but how to distinguish this from a gear fault? The answer lies in the fact that
a gear fault would give a deterministic modulation of the gear
mesh signal, which gives discrete points in the spectral correlation in both the f and a directions. The rough inner race fault,
on the other hand, gives a second order cyclostationary modulation since the rollers are in a different position on the spalled
surface for each revolution of the gear. The effect in the spectral correlation is distributed with f, just as in Figure 23. The
second order cyclostationary part would be combined with a
first order part (the mean value of the modulating signal) but
the effects from this can be removed, along with gear effects,

SOUND AND VIBRATION/MAY 2004

Figure 23. A spectral correlation for a localized inner race fault (actual span of frequency f [2800; 3300] Hz), B spectrum of squared
evelope. Shaft speed = 9.5 Hz, BPFI = 120 Hz.
60
Faulty
50

a=w

40

dB

Figure 22. Angular velocity signal obtained as the derivative of the

phase of Figure 21 (numbers along top denote cylinder firing cycle).

30

20
Good
10

0.05

0.1

0.15

0.2
0.25
0.3
0.35
Normalized Frequency, f

0.4

0.45

0.5

Figure 24. Spectral correlation for an extended inner race fault evaluated at a = w.

by using discrete/random separation before calculating the

spectral correlation. The latter can be evaluated for a specified
value of cyclic frequency where a result is expected, such as
the shaft speed W for an inner race fault. This eliminates stationary random noise, which only appears at zero cyclic frequency (it is constant with time). Thus, the only explanation
for the considerable increase as a result of the fault in Figure
24 is a cyclostationary signal with cyclic frequency equal to
the shaft speed, in other words a bearing inner race fault.

Prognosis
This is the least developed area in machine condition monitoring, but probably the one currently receiving the most attention.
For faults that develop gradually, it can be sufficient to simply carry out trend analysis of parameters that characterize the
fault severity, such as the amount of change from a baseline
value (e.g., exceedance spectra as in Figure 1). One system was
proposed with this as its basis. 22 Note that since equal changes
in severity are given by equal changes on a logarithmic scale,
the trend plots should be of logarithmic values (e.g., dBs) as a
linear trend of dB values represents a uniform change. Sometimes the fault severity feeds back on the rate of deterioration
(e.g., with gear wear) in which case it may be better to fit an
exponential curve to the data to make the best estimates of remaining life.
The best estimates of remaining life will be obtained from

15

models of the failure mechanism and much effort is currently

being put into developing such models. This is probably most
advanced in the application to rotor dynamics and it is expected that such models will soon be able to differentiate between unbalance, misalignment and rotor cracks, as well as
being able to make valid prognoses of future developments.
Such simulation models will also be very valuable to train
neural networks to recognize a wide range of faults, as in general it would not be economically viable to experience such
failures in sufficient numbers.

Conclusion
A large number of techniques are now available to use vibration analysis to detect and diagnose incipient faults in operating machines. Current developments will help in automating both the diagnosis and prognosis of such faults.

References
1. John S. Mitchell, Machinery Analysis and Monitoring, Penn Well,
1981.
2 E. Downham and R. Woods, ASME paper, Toronto, September 8-10,
1971.
3. P. Bradshaw and R. B. Randall, Early Fault Detection and Diagnosis on the Trans Alaska Pipeline, MSA Session, ASME Conf.,
Dearborn, pp 7-17, 1983.
4. J. Howard Maxwell, Induction Motor Magnetic Vibration, Proc.
Vibration Institute, Meeting, Houston, TX, April 19-21, 1983.
5. R. B. Randall, A New Method of Modeling Gear Faults, ASME J.
Mech. Design, 104, pp 259-267, 1982.
6. J. Antoni and R. B. Randall, Differential Diagnosis of Gear and Bearing Faults, ASME J. Vib. & Acoustics, 124, Apr 2002, pp. 165-171.
7. P. D. McFadden and J. D. Smith, Model for the Vibration Produced
by a Single Point Defect in a Rolling Element Bearing, J. Sound
Vib., 96 (1), pp 69-82, 1984.
8. D. Ho and R. B. Randall, Optimisation of Bearing Diagnostic Techniques Using Simulated and Actual Bearing Fault Signals, Me-

16

9.
10.
11.
12.
13.
14.
15.
16.
17.

18.
19.
20.
21.

22.

chanical Systems and Signal Processing, 14 (5), September 2000,

pp 763-788.
W. A. Gardner, Introduction to Random Processes with Applications
to Signals and Systems, Macmillan, 1986.
R. B. Randall, Frequency analysis, Brel & Kjr, Copenhagen,
1987.
S. Braun and J. K. Hammond, Parametric Methods, Mechanical
Signature Analysis, (Ed. S. Braun), Academic Press, London, 1986.
Y. Gao, R. B. Randall and R. Ford, Estimation of Envelope Spectra
Using Maximum Entropy Spectral Analysis and Spectral Interpolation, Int. J. of Comadem, 1 (3), pp 15-22, 1998.
W. Wang and A. K. Wong, Autoregressive Model-Based Gear Fault
Diagnosis, ASME J. Vib. & Acoustics, 124, Apr. 2002, pp 172-179.
P. D. McFadden, A Revised Model for the Extraction of Periodic
Waveforms by Time Domain Averaging, Mech. Systems & Signal
Processing, 1 (1), pp 83-95, 1987.
R. Potter and M. Gribler, Computed Order Tracking Obsoletes Older
Methods, SAE Paper 891131, 1989.
Chaturvedi, G. K. and Thomas, D. W., Bearing Fault Detection
Using Adaptive Noise Cancelling, Transactions of the ASME, 104,
Apr. 1982, pp 280-289.
J. Antoni and R. B. Randall, Unsupervised Noise Cancellation for
Vibration Signals Part II, A Novel Frequency Domain Algorithm.
Mechanical Systems and Signal Processing, Vol. 18, No. 1, pp 103118, 2004.
P. J. Sweeney and R. B. Randall, Gear Transmission Error Measurement using Phase Demodulation, Proc. I. Mech. E., Part C, J. Mech.
Eng. Sci. 210 (C3), pp 201-213, 1996.
P. D. McFadden. J. Vib., Acoust., Str., & Rel. in Des., 108, pp 165170.
L. Cohen, Time-Frequency Analysis, Prentice-Hall, NJ. 1995.
R. B. Randall, J. Antoni and S. Chobsaard, The Relationship Between Spectral Correlation and Envelope Analysis in the Diagnostics of Bearing Faults and Other Cyclostationary Machine Signals,
Mech. Systems & Signal Processing, 15(5), pp 945-962, 2001.
R. B. Randall, Computer Aided Vibration Spectrum Trend Analysis for Condition Monitoring, Maintenance Management International, 5, 1985, pp 161-167.

The author can be contacted at: b.randall@unsw.edu.au.

SOUND AND VIBRATION/MAY 2004