You are on page 1of 12

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/317589785

Cepstral removal of resonance effects to improve the use of traditional gear


diagnostic indicators at different speeds

Conference Paper · June 2017

CITATIONS READS

0 416

4 authors, including:

Wade A Smith R.B. Randall


UNSW Sydney UNSW Sydney
98 PUBLICATIONS   1,585 CITATIONS    286 PUBLICATIONS   11,551 CITATIONS   

SEE PROFILE SEE PROFILE

Zhongxiao Peng
The University of New South Wales (UNSW Sydney)
182 PUBLICATIONS   3,775 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Wear monitoring View project

Machine condition monitoring based on vibration simulation View project

All content following this page was uploaded by Wade A Smith on 14 June 2017.

The user has requested enhancement of the downloaded file.


Cepstral removal of resonance effects to improve the use of traditional
gear diagnostic indicators at different speeds

Wade A Smith, Robert B Randall, Chongqing Hu and Zhongxiao Peng


School of Mechanical and Manufacturing Engineering,
University of New South Wales, Australia, wade.smith@unsw.edu.au

Abstract

Most traditional gear diagnostic indicators were developed for constant speed operation
and are based on the spectral amplitudes of the gearmesh component and its harmonics,
as well as those of the sidebands spaced at integer multiples of shaft speed around the
meshing harmonics. In variable speed operation, it is common to perform order tracking
to remove the frequency modulation so as to restore the smeared gearmeshing
components to discrete frequency lines. However, as well as frequency modulation,
there are usually amplitude modulation effects associated with the various meshing
components passing through system resonances in accordance with the changing speed.

This paper presents a method to remove the amplitude modulation effects of these
resonances in applications involving variable speed or a range of different constant
speeds. The approach employs a cepstral method, developed originally for operational
modal analysis, which extracts modal information from continuous response signals,
even when they are non-stationary (since the modal information is constant in the
normal frequency domain). The technique is applied prior to order tracking, after which
the traditional indicators can be applied to give an improved diagnosis. In this study the
method is demonstrated on signals from a laboratory gearbox running at a range of
different constant speeds, but will be extended in future to variable speed cases. It is
shown that although the traditional indicators do not reliably diagnose the local gear
faults in this instance, the cepstral method is quite successful in reducing the effects of
speed.

1. Introduction
Most traditional gear diagnostic techniques were developed for constant speed and load
operation. In such conditions, the effects of the transfer path between the main
excitation source (gearmesh) and some fixed measurement point remain constant (in the
absence of damage in the structure), meaning that any changes in response amplitude, in
particular of components such as gearmesh harmonics and modulation sidebands, can be
attributed to variations in the forcing function, and are thus indicative of a change in
gear condition. Under variable speed, or even at different constant speeds, this no longer
holds. At different speeds the signal components move relative to the system’s fixed
frequency resonances, and so both their amplitude and phase will change in the
measured response.
It is common to apply order tracking and time synchronous averaging (TSA) in gear
diagnostics, and while the former removes the frequency modulation effects from
variable speed, it does not remove amplitude modulation from passage of the gearmesh
components through resonances. The resulting TSA signal would then be of dubious
validity, because it would consist of some sort of time average of the various
gearmeshing components, and much of the deterministic signal would leak through to
the ‘random’ residual signal.

It was proposed by Randall et al. that compensation for these resonance effects could be
made using cepstral editing (liftering), leaving a signal dominated by the discrete
frequency forcing function components(1,2). It is thought that such an approach would
produce a more reliable TSA signal and an improvement in the performance of TSA-
based diagnostic techniques. This resonance compensation method is applied here to a
number of different constant speed cases to see whether it can improve the use of
traditional gear diagnostic indicators, such as those outlined by Samuel and Pines(3),
most of which are based on modified versions of the TSA signal. Future work will
include a number of variable speed applications.

2. Proposed method
2.1 Cepstrum defined

Various definitions for the cepstrum exist, but here we use the ‘real cepstrum’ x̂ and
‘complex cepstrum’ x̂c , defined for some time signal x(t) as:

xˆ (t ) = ℑ−1 [log{ℑ[x(t )]}] and xˆc (t ) = ℑ−1 [log{ℑ[x(t )]}] (1)

where ℑ represents the Fourier transform operator.

For a linear, single-input (SIMO) system (e.g., single-stage gearbox), the system
response y is obtained from the convolution of the forcing function f with the impulse
response function h:


y (τ ) = h(τ ) ∗ f (τ ) = ∫ h(τ ) ⋅ f (τ − τ ) ⋅ dτ (2)
−∞

In the frequency domain this convolutive relationship becomes multiplicative, and, due
to the logarithmic operation, in the cepstrum the relationship is simply additive:

yˆ (τ ) = hˆ(τ ) + fˆ (τ ) (3)

2.2 Cepstral liftering and resonance compensation

This additive property has been used extensively in operational modal analysis(4-6) to
separate excitation and transfer path effects, one or both of which often have reasonably
well-known characteristics in the cepstrum. Frequency response functions, for example,

2
generally occupy a broad quefrency range (quefrency, with units of ‘seconds’, being the
independent variable in the cepstrum domain) but are dominated by low quefrency
content. On the other hand, discrete frequency components, as produced by gears, have
sharp spectral characteristics and would thus occupy a higher quefrency range. These
properties allow for their (partial) separation through careful liftering. It was proposed
in Refs.(1,7,8) that an appropriate lifter to achieve this would be an exponential lifter,
e −σt , designed to retain modal information and discard higher quefrency (sharp spectral)
content. σ, the damping constant, represents the decay of the exponential window and
should be chosen carefully. As explained in Ref.(8), applying an exponential window
amounts to adding damping of σ to the damping level of every pole and zero in the
frequency response function, which can in principle be compensated for. Subtracting
this modal information then leaves a signal dominated by deterministic content suitable
for gear diagnostics, while the modal portion could be valuable for example for
operational modal analysis or bearing diagnostics(7). This approach for separating the
signal types can be expressed as a modified version of Eq. (3):

( )
yˆ (τ ) = hˆ(τ ) + fˆ (τ ) ≈ e−στ yˆ (τ ) + 1 − e−στ yˆ (τ ) (4)

( )
where the 1 − e −στ yˆ (τ ) portion is used to form the TSA signal for further analysis.

In many cases, including the current application, the edited time waveform is required
for further processing after liftering. This presents a challenge in that the real cepstrum
of Eq. (1) does not include phase information, and so the edited time waveform is not
recoverable. The complex cepstrum does include phase information, but then to
calculate it requires that the phase be unwrappable, which is not the case for general
machine response signals. However, Randall and Sawalhi(9) showed that cepstral editing
can be performed on the real cepstrum, obtained from spectral amplitudes only, and
then recombined with the original phase after returning to the log spectrum, from which
the edited time signal can be obtained with the usual exponential and inverse Fourier
transform operations. This means that cepstral liftering can be applied to general
response signals without the requirement that the phase be unwrappable.

2.3 Traditional gear diagnostic indicators

The diagnostic indicators investigated here are very well-established and have been
explained thoroughly in the literature, for example in Refs.(3,10). A brief outline is given
in this section. All indicators are based on the TSA signal, x(t), calculated over one
cycle of the gear under investigation. In the proposed method, this TSA signal was
obtained after order tracking the liftered signal processed in accordance with the
techniques outlined in the preceding sections.

Use is also made of the signal yd(t), comprising the mesh frequency component, its
harmonics and its first-order sidebands, and signal yr(t), the mesh frequency component
and its harmonics (i.e., yd(t) without sidebands). From these signals, the difference d and
residual r signals are formed as follows:

d (t ) = x(t ) − y d (t ) and r (t ) = x(t ) − y r (t ) (5)

3
The final signal used in the traditional indicators is the envelope e of signal xb, where xb
is obtained by bandpass filtering x around the gearmesh frequency. A description of the
seven dimensionless indicators is given in Table 1. In the table, x refers to the mean of
x, and N represents the number of points in the signal.

It should perhaps be pointed out that the so-called energy ratio (ER) is a misnomer in
that it is a root-mean-square ratio, whose square is a power (not ‘energy’) ratio, and in
fact its origins are questionable. Samuel and Pines(3) and Decker and Lewicki(10) both
cite Swansson(11) as its source, but the term does not appear in the listed reference. They
also both state that ER is useful for detecting “heavy uniform wear”, but this is at odds
with the basic physical understanding that uniform effects should by definition produce
periodic signals (at gearmesh frequency), suggesting only the gearmesh components
should be affected, at least until the wear becomes non-uniform. The energy ratio
should not be confused with the Sideband Energy Ratio™, a patented technique
developed by Bentley Nevada(12) based on the ratio of the sum of the amplitudes of the
six sidebands on either side of the gearmesh frequency to the amplitude of the gearmesh
component itself (calculated separately for a small number of gearmesh harmonics).

Table 1. Indicators used in this study

Indicator Formula Description


Energy RMS d
Ratio of root-mean-square values of d and yd.
ratio (ER) RMS yd

∑ (x − x )
N 4 Normalised fourth central moment. Measure of
N i
i =1 impulsiveness in the TSA signal. Should be sensitive to
Kurtosis
[∑ (x − x ) ]
N
i =1 i
2
2
sharp modulations as would occur with a localised gear
fault.

∑ (d − d )
N 4 Measure of impulsiveness in the difference signal. Since
N i
i =1 the meshing harmonics and first-order sidebands are
FM4
[∑ (d − d ) ]
N
i =1 i
2 2 removed in d, this should be more sensitive than x to
modulations from localised gear faults.

N ∑ (d − d )
2 N 6
i =1 i As with FM4, although more sensitive to peaks in the
M6A
[∑ (d − d ) ]
N
i =1 i
2 3 difference signal.

∑ (d − d )
N 8
N3 i
i =1 As with M6A, although even more sensitive to peaks in the
M8A
[∑ (d − d ) ]
N
i =1 i
2 4 difference signal, meaning it is more prone to false alarms.

∑ (r − r )
N 4
N i Nearly identical to FM4. A measure of impulsiveness in the
i =1
NA4
[∑ (r − r ) ]
N
i =1 i
2
2 residual signal. Can also be used for trending over time (not
applied here).

∑ (e − e )
N 4
N i Measure of impulsiveness in the envelope of the bandpass
i =1
NB4
[∑ (e − e ) ]
N
i =1 i
2
2 filtered TSA signal, and so should also be sensitive to
modulations from localised gear faults.

4
3. Experimental setup
The experimental data used in this study was obtained from the UNSW gearbox test rig,
shown in Figure 1, by Dr Hiroaki Endo(13,14). In the rig, a three-phase induction motor
drives a hydraulic system to provide a resistance torque to the gearbox, which consists
of a basic 32:32 parallel spur gear stage. The speed of the input shaft is controlled by a
variable frequency drive.

Figure 1. UNSW gearbox test rig showing accelerometer positions(13)

Table 2. Description of seeded gear faults(13,14)

Gear state Description


Good gear Gear in good condition
Pitchline spall of length 0.5 mm (along tooth profile), width 6.25 mm (25% face
25% spall
width) and depth 0.5 mm
Pitchline spall of length 1.0 mm (along tooth profile), width 12.5 mm (50% face
50% spall
width) and depth 0.5 mm
Pitchline spall of length 1.5 mm (along tooth profile), width 25.0 mm (100%
100% spall
face width) and depth 0.01 mm (note: bottom of spall in meshing contact here)
Small crack Root crack of length 3 mm
Large crack Root crack of length 7 mm

In each test the vibration data was acquired for 5 s at a sampling rate of 65,536 Hz from
two B&K 4384 accelerometers located at positions P1 (vertical) and P2 (horizontal), as
shown in Figure 1. A once-per-rev tachometer signal was obtained from a Heidenhain
tacho on the free end of the input shaft. Testing was performed with a torque load of
120 Nm and at three different speeds: 2.8, 8.4 and 14 Hz shaft speed. Gears with various
localised faults (spalls and root cracks of different severities) were used to test the
performance of the proposed method, as outlined in Table 2. In each case, one gear was
in good condition and the other was as described in the table. Note that along with the
three different speeds, the use of two measurement points provides another test of the
capability of the liftering approach to remove transfer path effects from the response
signals.

5
4. Results
4.1 Cepstral editing preprocessing

Figure 2 shows the effect of applying an exponential lifter to one of the measured
responses. Through brief trial and error it was found that a damping constant of
σ = 251 rad/s (corresponding to a 3 dB bandwidth of 80 Hz) provided a suitable
separation of the modal information. This can be seen in the left plot, which shows the
spectra of the original signal and the liftered signal, showing a roughly 50 dB dynamic
range in the modal part of the signal. Subtraction of this signal (in the cepstrum or log
amplitude spectrum) from the original results in the spectrum on the right, showing
vastly reduced differences in the amplitude of the discrete frequency components. All of
the measured datasets were processed with these lifter settings (i.e., a window function
( )
of w(t ) = 1 − e −251t ) before order tracking and averaging.

150 150

100 100
Amplitude (dB)

Amplitude (dB)

50 50

0 0

0 1 2 3 4 5 0 1 2 3 4 5

Frequency (kHz) Frequency (kHz)

Figure 2. Spectra of original (left) and ‘liftered’ signal (right). Black line (left)
represents the modal content obtained via an exponential lifter in the cepstrum.
Subraction of the liftered cepstrum from the original gives the signal on the right

4.2 Indicator values for raw and liftered data

Figure 3 shows the values of indicators ER, Kurtosis, FM4 and M6A (see Table 1) for
the raw and liftered data from all test cases, and Figure 4 shows the same for indicators
M8A, NA4 and NB4 (note that some y-axes have a log scale).

One striking feature of the plots is that all the indicators perform very poorly in virtually
all the fault cases, with the only exception being the 100% spall in the raw datasets, and
even this has a large range that overlaps with the good gear case for most indicators.
This suggests that the dataset as a whole is not giving typical symptoms of a local gear
fault, i.e., amplitude modulation of the gearmeshing component. Nevertheless, the
effectiveness or otherwise of the indicators is not the central point here, but rather the
effect of the liftering on reducing transfer path effects arising from both different
running speeds and the use of different measurement points. A better test of the
method’s success is whether for each combination of indicator and fault case the results
for the six different speed/measurement point combinations cluster around the same
value.

6
P1, 2.8 Hz P1, 8.4 Hz P1, 14 Hz P2, 2.8 Hz P2, 8.4 Hz P2, 14 Hz
2.5 2.5

2 2

1.5 1.5
ER

ER
1 1

0.5 0.5

0 0
ll

ll

ll

ll
ck

ck
ar

ar
l

l
k

k
al

al
a

a
ac

ac
sp

sp

sp

sp
ge

ge
ra

ra
sp

sp
cr

cr
lc

lc
%

%
d

d
0%

0%
oo

oo
al

al
e

e
25

50

25

50
rg

rg
Sm

Sm
10

10
G

G
La

La
P1, 2.8 Hz P1, 8.4 Hz P1, 14 Hz P2, 2.8 Hz P2, 8.4 Hz P2, 14 Hz
12 12

10 10

8 8
Kurtosis

6 Kurtosis 6

4 4

2 2

0 0
l

l
ck

ck
ar

ar
l

l
al

al

al

al
k

k
al

al
ac

ac
sp

sp

sp

sp
ge

ge
ra

ra
sp

sp
cr

cr
lc

lc
%

%
d

d
0%

0%
oo

oo
al

al
e

e
25

50

25

50
rg

rg
Sm

Sm
10

10
G

G
La

La
P1, 2.8 Hz P1, 8.4 Hz P1, 14 Hz P2, 2.8 Hz P2, 8.4 Hz P2, 14 Hz

1 1
10 10
FM4

FM4

0 0
10 10
l

l
k

k
ar

ar
l

l
al

al

al

al
k

k
al

al
ac

ac
ac

ac
sp

sp

sp

sp
ge

ge
sp

sp
cr

cr
cr

cr
%

%
d

d
0%

0%
l

l
oo

oo
al

al
e

e
25

50

25

50
rg

rg
Sm

Sm
10

10
G

G
La

La

P1, 2.8 Hz P1, 8.4 Hz P1, 14 Hz P2, 2.8 Hz P2, 8.4 Hz P2, 14 Hz
4 4
10 10

3 3
10 10
M6A

M6A

2 2
10 10

1 1
10 10
l

l
ck

ck
ar

ar
l

l
al

al

al

al
k

k
al

al
ac

ac
sp

sp

sp

sp
ge

ge
ra

ra
sp

sp
cr

cr
lc

lc
%

%
d

d
0%

0%
oo

oo
al

al
e

e
25

50

25

50
rg

rg
Sm

Sm
10

10
G

G
La

La

Figure 3. Results from traditional indicators for all speed/fault cases. From top to
bottom: ER, kurtosis, FM4 and M6A. Left: raw data; right: liftered data

7
P1, 2.8 Hz P1, 8.4 Hz P1, 14 Hz P2, 2.8 Hz P2, 8.4 Hz P2, 14 Hz
2 2
10 10

0 0
10 10
M8A

M8A
-2 -2
10 10
l

l
k

k
ar

ar
l

l
al

al

al

al
k

k
al

al
ac

ac
ac

ac
sp

sp

sp

sp
ge

ge
sp

sp
cr

cr
cr

cr
%

%
d

d
0%

0%
l

l
oo

oo
al

al
e

e
25

50

25

50
rg

rg
Sm

Sm
10

10
G

G
La

La
P1, 2.8 Hz P1, 8.4 Hz P1, 14 Hz P2, 2.8 Hz P2, 8.4 Hz P2, 14 Hz

1 1
10 10
NA4

NA4

0 0
10 10
l

l
ck

ck
ar

ar
l

l
al

al

al

al
k

k
al

al
ac

ac
sp

sp

sp

sp
ge

ge
ra

ra
sp

sp
cr

cr
lc

lc
%

%
d

d
0%

0%
oo

oo
al

al
e

e
25

50

25

50
rg

rg
Sm

Sm
10

10
G

G
La

La
P1, 2.8 Hz P1, 8.4 Hz P1, 14 Hz P2, 2.8 Hz P2, 8.4 Hz P2, 14 Hz
5 5

4 4

3 3
NB4

NB4

2 2

1 1

0 0
l

l
ck

ck
ar

ar
l

l
al

al

al

al
k

ck
al

al
ac
sp

sp

sp

sp
ge

ge
ra

ra
sp

sp

a
cr

cr
lc

lc
%

%
d

d
0%

0%
oo

oo
al

al
e

e
25

50

25

50
rg

rg
Sm

Sm
10

10
G

G
La

La

Figure 4. Results from traditional indicators for all speed/fault cases. From top to
bottom: M8A, NA4 and NB4. Left: raw data; right: liftered data

This improved clustering can certainly be seen in most cases, the main exceptions being
ER and NB4. Interestingly, these are also the indicators that do not seem sensitive to the
100% spall fault even in the raw datasets (these giving a much stronger indication of the
spall fault with the other indicators, despite a very large range in values depending on
speed and measurement point).

To examine this improved clustering, the average range of the indicator values for each
fault case was calculated as a percentage of the mean value for the particular class
(calculated separately to account for the effect of both speed and measurement point).
The results are given in Table 3. It can be seen that the liftering gives a reduction in the

8
range of values in most cases, in some quite substantially. When averaged out, the
method gives a reduction in the range of indicator values of just over 25% (46% versus
64%) when correcting for speed and just under 50% (19% versus 35%) when correcting
for measurement point.

Table 3. Range of indicator values (expressed as percentage of mean in each class)


showing performance of cepstral liftering in correcting for the effect of speed and
measurement point

Speed Measurement Point


Indicator Raw Liftered Raw Liftered
ER 79.7 96.1 50.3 46.3
Kurtosis 44.6 24.2 28.2 12.9
FM4 38.6 10.9 19.0 4.3
M6A 88.5 32.4 44.9 14.4
M8A 122.1 119.2 71.2 29.8
NA4 41.7 9.5 16.4 4.5
NB4 28.5 30.4 12.5 19.8
Average 63.4 46.1 34.6 18.8

4.3 Discussion

While these results do exhibit improved clustering and suggest the effects of different
speeds and measurement points have been at least partly removed, there is clearly room
for improvement. It does appear that the dataset analysed here is atypical of local gear
faults, with little sign of the typical amplitude modulation patterns expected of such a
condition. Future testing is planned to gather more typical data, which will also include
tests run under variable speed. It should also be noted that the five-second records used
here gave a suboptimal number of averages for the TSA process (12 in the low speed
case), and this is something that will be addressed in future work.

Other possible areas for further development concern the fitting of a complete modal
model to account for phase changes in the response, and not merely the amplitude
effects corrected for here. This would of course require not just the transfer function
poles, but also the zeros, as both involve phase jumps of π radians.

5. Conclusions
This paper proposed the use of cepstral liftering to remove the amplitude effects of the
transfer path in gearbox response signals so that traditional gear diagnostic indicators
can be applied more effectively in cases involving variable speed or different constant
speeds. Measurements from a gearbox test rig running at different constant speeds and
with a number of different local gear faults were used to demonstrate the approach,
which involves first applying an exponential lifter to the response signals to obtain the
modal information. This liftered signal is then subtracted from the original (in the
cepstrum domain) to obtain the part of the signal dominated by the forcing function (in
this case the gearmesh), prior to order tracking and time synchronous averaging.

9
It was found that in this application the traditional indicators performed poorly in
detecting most of the local faults (spalls and root cracks of varying severity), but that the
cepstral liftering at least improved the variability of these indicators with different
running speeds and measurement points. It is thought that the dataset used for the
analysis was quite atypical of local gear faults, and so a better illustration of the method
might be possible with improved data and variable speed records.

Acknowledgements

The authors would like to thank Dr Hiroaki Endo for the use of his experimental data.

References

1. R Randall, M Coats and W Smith, 'Gear diagnostics under widely varying speed
conditions', 4th International Conference on Condition Monitoring of Machinery in
Non-Stationary Operations, Lyon, France, 15-16 December, 2014.
2. R B Randall, M D Coats and W A Smith, 'Improvement in the diagnostics of
gearbox vibration signals under varying speed conditions', 8th Australasian
Congress on Applied Mechanics, Melbourne, Australia, 23-26 November, 2014.
3. P D Samuel and D J Pines, 'A review of vibration-based techniques for helicopter
transmission diagnostics', Journal of Sound and Vibration, Vol 282, pp 475-508,
2005.
4. R B Randall, 'Cepstral Methods of Operational Modal Analysis', in Encyclopedia of
Structural Health Monitoring, Wiley, 2009.
5. R B Randall and W A Smith, 'New cepstral techniques for operational modal
analysis', First World Congress on Condition Monitoring, London, UK, 13-16 June,
2017.
6. W A Smith and R B Randall, 'Cepstrum-based operational modal analysis revisited:
a discussion on pole-zero models and the regeneration of frequency response
functions', Mechanical Systems and Signal Processing, Vol 79, pp 30-46, 2016.
7. R Randall, W Smith and M Coats, 'Bearing diagnostics under widely varying speed
conditions', 4th International Conference on Condition Monitoring of Machinery in
Non-Stationary Operations, Lyon, France, 15-16 December, 2014.
8. R B Randall and W A Smith, 'New cepstral methods for the diagnosis of gear and
bearing faults under variable speed conditions', 23rd International Congress on
Sound and Vibration, Athens, Greece, 10-14 July, 2016.
9. R B Randall and N Sawalhi, 'A New Method for Separating Discrete Components
from a Signal', Sound and Vibration, May 2011, pp 6-9, 2011.
10. H J Decker and D G Lewicki. Spiral Bevel Pinion Crack Detection in a Helicopter
Gearbox. 2003.
11. N Swansson, 'Application of Vibration Signal Analysis Techniques to Condition
Monitoring', Conference on Lubrication, Friction and Wear in Engineering,
Melbourne, Australia, pp 262-267, 1980.
12. C T Hatch, D D Hess, A A Weiss, S Woodson and M B Kalb. Sideband Energy
Ratio Method For Gear Mesh Fault Detection. Google Patents, 2012.

10
13. H Endo, R B Randall and C Gosselin, 'Differential diagnosis of spall vs. cracks in
the gear tooth fillet region: Experimental validation', Mechanical Systems and
Signal Processing, Vol 23, pp 636-651, 2009.
14. H Endo, 'Simulation of Gear Faults and Its Application to the Development of
Differential Diagnostic Technique', PhD Thesis, University of New South Wales,
2005.

11

View publication stats

You might also like