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

net/publication/317589934

Use of Cyclostationarity to Detect Changes in

Gear Surface Roughness using Vibration
Measurements

Conference Paper · August 2017

CITATIONS READS

0 133

5 authors, including:

Wade A Smith P. Borghesani

UNSW Sydney Queensland University of Technology
50 PUBLICATIONS 227 CITATIONS 52 PUBLICATIONS 338 CITATIONS

SEE PROFILE SEE PROFILE

Zhongxiao Peng R.B. Randall

UNSW Sydney UNSW Sydney
126 PUBLICATIONS 1,505 CITATIONS 230 PUBLICATIONS 4,982 CITATIONS

SEE PROFILE SEE PROFILE

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

FRC Bridge View project

Hydraulically Interconnected Suspension 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.

Use of Cyclostationarity to Detect Changes
in Gear Surface Roughness using
Vibration Measurements

Xihao Zhang1, Wade A. Smith2, Pietro Borghesani3, Zhongxiao Peng4, Robert

B. Randall5

Abstract Wear in gears can usually be detected from the vibration signal once the
wear has reached a ‘macro’ level – millimetre-scale variations from the original
involute profile. Macro level wear is often preceded and accompanied by micro-
level surface roughness changes (micrometre scale), arising from either abrasive
wear or contact fatigue pitting. These micro- and macro-level phenomena interact
with one another, and so knowledge of surface roughness is needed to be able to
predict macro-level wear. It was recently suggested that it may be possible to use
the cyclostationary properties of the vibration signal as an indicator of gear surface
roughness, and the present paper examines this possibility further. It is thought
that roughness information is carried by random high frequency vibrations that are
modulated by the gearmesh cycle, and so any speed changes in the system should
change both the carrier and modulating frequencies. The paper tests this hypothe-
sis by studying laboratory measurements from a spur gearbox running at different
speeds and with gears of different roughnesses. The findings will be very im-
portant for the further development of gear prognostics methods.

1X. Zhang
School of Mechanical and Manufacturing Engineering, University of New South Wales
(UNSW), Sydney, Australia
2W. A. Smith ()
School of Mechanical and Manufacturing Engineering, UNSW, Sydney, Australia
Email: [email protected]
3P. Borghesani
Science and Engineering Faculty, Queensland University of Technology, Brisbane, Australia
4Z. Peng
School of Mechanical and Manufacturing Engineering, UNSW, Sydney, Australia
5R. B. Randall
School of Mechanical and Manufacturing Engineering, UNSW, Sydney, Australia
2 X. Zhang - Use of Cyclostationarity to Detect Changes in Gear Surface Roughness using Vibration Measurements

1 Introduction
Wear is one of the most common failure modes of gears, with abrasive wear
and fatigue being the main wear mechanisms. Gear wear is a very complex pro-
cess that is influenced by many factors, such as lubrication, temperature, operating
conditions, material properties and tooth geometry. It has been established that vi-
bration-based techniques can be used to detect macro-level wear (millimetre scale
variations from the original involute profile), through changes in the amplitudes of
gearmesh harmonics, in particular the second harmonic (Mark 1978; Randall
1982). However, most of the techniques were developed for fault detection and di-
agnosis, rather than for prognosis and remaining useful life (RUL) prediction.
Predicting RUL is the least developed aspect of gear condition monitoring, but it
carries perhaps the greatest potential benefits, both economic and safety related.
The wear and lifespan of gears is related to tooth surface roughness (Krantz 2005);
however, information about tooth surface roughness is very difficult to obtain
without stopping the machine and taking detailed measurements. As such, estab-
lishing a relationship between surface roughness and vibration would provide a
very valuable tool for the diagnosis of the gear wear state and for the calculation
of remaining useful life.

It was recently proposed that such a relationship could be established through

the use of cyclostationary (CS) signal analysis (Yang et al. 2015), and the present
paper explores this possibility further. It is widely known that most of the power
in gear vibration signals resides in the periodic components, such as the gearmesh
frequency harmonics, but it is thought that important information is also carried in
the second-order cyclostationary (CS2) components – in this case random vibra-
tions that are modulated by the gear meshing process (Antoni et al. 2004;
Capdessus, Sidahmed, and Lacoume 2000). One source of these random vibra-
tions in gears is from the friction and asperity contacts that occur between the slid-
ing surfaces of the meshing teeth, and it is thought that the strength of these vibra-
tions would be closely related to tooth surface roughness. This paper investigates
the relationship between gear surface roughness and cyclostationarity using data
obtained from a laboratory spur gearbox test rig. In particular, the effect of run-
ning speed on the frequency distribution of the cyclostationary content is exam-
ined for a number of different surface roughness levels. This will allow for more
targeted metrics of surface roughness to be developed in the future.

2 Gear Surface Roughness and Cyclostationarity

2.1 Cyclostationary signals and their indicators

A signal is defined to be cyclostationary at the nth-order if its nth-order statisti-

cal properties are periodic with respect to time (Antoni et al. 2004). The most rel-
 X. Zhang - Use of Cyclostationarity to Detect Changes in Gear Surface Roughness using Vibra-
tion Measurements 3

evant case here is that of a second-order CS signal, which has a periodic autocor-
relation function (or variance), a typical example of which is a cyclic repetition of
random bursts. Note that CS2 signals are random signals, such that their cyclic
structure (i.e. periodic variance) is not apparent in the ordinary frequency spec-
trum. For such cases, cyclostationary signal processing techniques must be applied
to uncover the underlying cyclic nature of the signal.

The degree of cyclostationarity in a signal can be measured by established indi-

cators (Raad, Antoni, and Sidahmed 2008). The indicator of second-order cy-
clostationarity (ICS2) is defined as:

𝑘𝛼 2
|𝐶2𝑥 0 (0)|
ICS2 = ∑𝑘𝜖𝑍,𝑘≠0 0 (0)| 2 (1)
|𝐶2𝑥

0 (0)
where 𝐶2𝑥 is the mean square power of the ‘centred signal’ 𝑥𝑐 , obtained by
𝑘𝛼
subtracting the synchronous average from the raw signal. 𝐶2𝑥 0 is the second-order
cyclic cumulant for the set of all cyclic frequencies α (=𝑘𝛼0 ). According to Raad
𝛼 (0)
et al. (Raad, Antoni, and Sidahmed 2008), a consistent estimator of 𝐶2𝑥 for a
discrete signal 𝑥(𝑛) with length N is given by the components at frequencies 𝛼 in
its envelope spectrum, in this case using the squared signal as an approximation
for the squared envelope:

𝛼 (0)
𝐶2𝑥 ≈ 𝑁 −1 𝐷𝐹𝑇{𝑥𝑐2 (𝑛)}(𝛼) (2)

where 𝐷𝐹𝑇{𝑥(𝑛)}(𝛼) stands for the N-point discrete Fourier transform of sig-
nal 𝑥(𝑛) calculated at frequencies 𝛼.

While ICS2 gives an indicator of the overall level of second-order CS content

in a signal, it does not indicate its spectral frequency distribution. For this, one
must employ more comprehensive tools such as the Spectral Correlation (SC),
𝑆𝑥 (𝛼, 𝑓), defined as the double discrete Fourier transform of the instantaneous au-
tocorrelation function (itself a function of both time and time lag for CS signals)
(Antoni, Xin, and Hamzaoui 2017). The Spectral Correlation indicates the power
distribution of the signal with respect to both the spectral frequency f and the cy-
clic frequency  (Antoni 2007). In this paper we employ a normalised version of
the SC, the Spectral Coherence, defined as (Antoni, Xin, and Hamzaoui 2017):

𝑆𝑥 (𝛼,𝑓)
𝛾𝑥 (𝛼, 𝑓) = (3)
√𝑆𝑥 (0,𝑓)𝑆𝑥 (0,𝑓−𝛼)

where 𝑆𝑥 (0, 𝑓) represents the ordinary power spectral density at frequency f.

I.e., the CS content at frequency f is normalised by the power at frequencies f and
f– in the stationary part of the signal. The Spectral Coherence can also be inter-
4 X. Zhang - Use of Cyclostationarity to Detect Changes in Gear Surface Roughness using Vibration Measurements

preted as the SC of the whitened signal, which tends to magnify weak cyclosta-
tionary signals (Antoni, Xin, and Hamzaoui 2017).

2.2 Effect of gear surface roughness on vibration signals

As explained in (Yang et al. 2015), the rationale for employing cyclostationary

analysis in this context arises from the periodic variation in sliding velocity (and
contact forces) inherent in meshing gears. This concept is outlined in Figure 1,
which shows, as a function of gear rotation angle, the variation in the number of
meshing tooth pairs, the approximate sliding velocity of the meshing pairs and the
possible CS2 signal generated from the random asperity contacts on the gear tooth
faces. It can be seen that a periodicity in the sliding velocity is produced corre-
sponding to the gearmesh frequency. It is likely that the strength of the vibrations
arising from random asperity contacts would be closely tied to sliding velocity
(and level of roughness), leading to the amplitude modulation effect illustrated in
the bottom plot of Figure 1. Note that even though this vibration (the carrier sig-
nal) is random, with unknown frequency content, it can be separated from other
parts of the signal using CS tools because the modulation signal is both known and
periodic, with a cyclic frequency equal to gearmesh frequency.

Figure 1 Possible CS signal generation from meshing gear teeth; top: number of tooth
pairs in contact; middle: approximate sliding velocity (N = number of gear teeth); bottom:
possible amplitude-modulated random signal (CS2) (Yang et al. 2015)
 X. Zhang - Use of Cyclostationarity to Detect Changes in Gear Surface Roughness using Vibra-
tion Measurements 5

3 Methodology
3.1 Experimental setup

Data was collected from a number of tests conducted on the UNSW gearbox
test rig, shown in Figure 2. The gearbox consists of a single parallel gear stage and
is powered by an induction motor connected to a variable frequency drive (VFD).
A water pump is connected to the output shaft to apply a torque load, and a tacho
(twice-per rev) is connected to the free end of the output shaft. The drive and driv-
en gears are KHK mild steel spur gears, with 46 and 25 teeth, respectively.

The vibration signal was collected using a B&K 4370 accelerometer stud-
mounted on the top of the gearbox casing. The signal was recorded using a Na-
tional Instruments CompactDAQ with an NI-9234 module. The sampling frequen-
cy was 51 kHz.

Figure 2 UNSW gearbox rig

3.2 Test program and surface roughness measurement

Three sets of tests were conducted, each with different gear pairs. Before some
tests the meshing surfaces of the gears were roughened manually (see below).
Each test consisted of a long running period (88, 24 and 45 hours for Tests 1, 2
and 3, respectively), in which the tooth surfaces were allowed to evolve (generally
smoothen) naturally over time, and periodically the vibration signals were record-
ed and the tooth surfaces measured again to obtain the roughness level corre-
sponding to each vibration signal. In total, 15 measurement points were used: two
in Test 1 (points 1A and 1B), four in Test 2 (2A-2D) and nine in Test 3 (3A-3I).
6 X. Zhang - Use of Cyclostationarity to Detect Changes in Gear Surface Roughness using Vibration Measurements

The rig was run with an input shaft speed fin of 23 Hz and a torque load of 14 Nm,
but at each measurement point signals were also recorded at 15 Hz / 7 Nm.

Sandpaper (grit 120, 220 and 320) was used to roughen the gears at the start of
Tests 2 and 3. (In Test 1 the gears were as-supplied.) The gears in Tests 1, 2 and 3
had initial surface roughnesses (Ra values) of 0.6, 1.5 and 3.2 µm, respectively,
and these roughnesses reduced to 0.5, 1.1 and 1.5 µm over the duration of the
tests.

Surface roughness was measured using a Mahr M1 Perthometer (cut-off wave-

length 0.8 mm), with measurements taken along a number of teeth (randomly se-
lected), and multiple measurements taken along the same tooth. The measured Ra
values were then averaged to give the reported value. At the start of Tests 2 and 3,
with manually roughened surfaces, a consistent roughness was achieved by ensur-
ing that the Ra values of all measurements fell within ±0.1 µm of the mean value.

3.3 Signal processing

As explained in Section 2.2, it is hypothesised here that surface roughness in-

formation would be carried in the random part of the vibration signal (and modu-
lated by a periodic function associated with gearmesh), so order tracking and time
synchronous averaging were applied to remove deterministic components syn-
chronous with both shafts. The residual signal (obtained by subtracting the two
TSA signals from the order-tracked signal) was then used to calculate ICS2 ac-
cording to Eqs. (1) and (2), with the cyclic frequency set to the gearmesh frequen-
cy (𝛼 = 𝑓gm ). The residual signal was also used to calculate the Spectral Coher-
ence (again at 𝛼 = 𝑓gm ), as defined in Eq. (3), using a new fast computation code
developed by Antoni et al. (Antoni 2016; Antoni, Xin, and Hamzaoui 2017).

4 Results and discussion

4.1 Correlation of CS indicator and surface roughness

Figure 3 shows the plots of ICS2 vs surface roughness for all the measurement
points for both input shaft speeds (15 and 23 Hz). The data from Tests 1 and 2 in
the high speed case was in fact presented in (Yang et al. 2015), where a very good
correlation between ICS2 and roughness was observed. It was explained in that
paper that point 2B was thought to be unreliable because looseness had developed
in the rig, so that point was discarded in the regression analysis. The plots shown
here, with a greater range of roughness levels and a new speed/load, suggest a
more complex relationship. Certainly, the monotonic trend previously observed in
the high speed case (right plot) seems to break down around Ra = 2 m. It is pos-
sible that for the outlying measurement points at higher roughness levels (the first
 X. Zhang - Use of Cyclostationarity to Detect Changes in Gear Surface Roughness using Vibra-
tion Measurements 7

four measurements from Test 3) there is a fundamental difference in the interac-

tion of the contacting surfaces. For example, a higher Ra value – obtained using
coarser grit sandpaper – may mean a larger average wavelength in the surface as-
perities, and beyond a certain point this could lead to a reduction in the rate of as-
perity breakage and deformation, and hence a lower ICS2 value. However, this
explanation is at odds with the fact that throughout Test 3 (45 hrs), the average
surface roughness of the gears did change considerably (from Ra = 3.2 to 1.5 m).
Implicit in this point is that Ra alone is not an adequate metric to completely char-
acterise a random surface, and this is one area that will be examined in future
analyses on the cyclostationarity/roughness relationship.

In the case of low speed/load (left plot), there is no observable trend, even from
the Test 1 and 2 data points. A likely reason for the poorer correlation in the low
speed case is due to the very low torque load (7 Nm), which could not be con-
trolled independently of speed with the rig setup at that time. Future testing will
address this issue.
ICS2

Average gear surface roughness, Ra (m)

Figure 3 ICS2 (calculated at 𝛼 = 𝑓𝑔𝑚 ) vs gear surface roughness; left: fin=15 Hz; right:
fin=23 Hz

4.2 Effect of running speed on cyclostationarity

One way of improving on ICS2 as a potential roughness indicator is to develop

a more targeted metric that examines the spectral frequency range carrying the
most roughness information. This section contributes to that by studying the effect
on CS content of the machine running speed. The rationale for using CS pro-
cessing tools to examine surface roughness (see Fig. 1) was based on variations in
sliding velocity between the contacting surfaces throughout the mesh cycle, and so
it seems likely that running speed (which dictates the range of sliding velocities)
8 X. Zhang - Use of Cyclostationarity to Detect Changes in Gear Surface Roughness using Vibration Measurements

would have a direct effect on the frequency of the CS content. This CS spectral
content was investigated using the Spectral Coherence defined in Eq. (3), ex-
2
pressed as |𝛾𝑥 | to give values ranging from 0 to 1. Four measurement points were
chosen for analysis – 1B, 2A, 3A and 3F – representing a range of roughness and
ICS2 levels. Figure 4 shows the Spectral Coherence for these points over the 0-
25 kHz spectral frequency range. This represents the power of the signal content
modulated at the gearmesh frequency normalised by the power in the stationary
part of the signal. To give a clearer indication of the spectral distribution, the first
order spectral moment (or ‘mean frequency’) for each coherence plot is shown as
a dotted line on the graphs, and the results for the two speed cases are plotted on
the same axes.

2
Figure 4 Spectral coherence |𝛾𝑥 (𝛼 = 𝑓𝑔𝑚 , 𝑓)| obtained at different operating speeds
corresponding to measurement points 1B (top-left), 2A (top-right), 3A (bottom-left) and 3F
(bottom-right). Black lines: fin=15 Hz; red lines: fin=23 Hz. Dashed lines represent first-
order spectral moments.
 X. Zhang - Use of Cyclostationarity to Detect Changes in Gear Surface Roughness using Vibra-
tion Measurements 9

It is clear in every plot that the average frequency of the CS content increases
with the running speed, almost in direct proportion, with the average change in
mean frequency found to be about 1.4, while the speed ratio was around 1.5. That
is, not only does the cyclic frequency change with running speed, but so does the
range of dominant carrier frequencies. This indicates that the main carrier signals
are indeed tied to sliding velocity and the rate of interaction of the asperities, and
not, for example, fixed frequency resonances such as might be excited in the pres-
ence of a bearing fault. Note that this analysis is complicated by the fact that the
machine speed only changes the range (or average) of the sliding velocity, i.e. the
sliding velocity still spans a range from zero to the maximum.

To develop a more effective roughness indicator further work is needed to es-

tablish the frequency range over which most roughness information is carried,
over a wider range of speeds, roughnesses and torque loads. However, the finding
that the ‘mean’ carrier frequency seems closely related to sliding velocity helps in
the development of more sophisticated tools, for example to isolate frequency
modulation effects.

5 Conclusion
This paper investigated the possibility of using the cyclostationary (CS) proper-
ties of vibration signals to detect changes in gear tooth surface roughness. It was
previously proposed that information about the surface roughness of meshing
gears would be carried in the random vibrations generated from asperity impacts,
and these vibrations would be modulated by the gear meshing cycle, which is en-
tirely deterministic. This creates a second-order cyclostationary (CS2) signal, and
a previous study found a strong correlation between CS2 content and surface
roughness. The present paper, using new data covering a much larger range of
roughness levels, found this relationship to be more complex than originally
thought, with a negative correlation between these two variables observed beyond
a certain roughness level (Ra = 2 m). More work is needed to understand the
physics underlying this observation.

The paper also used Spectral Coherence to study the frequency content of that
part of the vibration signals modulated at the gearmesh frequency, and how it is
affected by running speed. For the four measurement points studied, covering a
range of surface roughnesses, it was found that the mean frequency of the coher-
ence varied almost in direct proportion to the running speed, indicating that the
main carrier signals in the CS2 content were based on sliding velocity, such as
impacting asperities, and not on fixed frequency resonances. This provides useful
information for the development of more targeted roughness metrics in the future.
 10 X. Zhang - Use of Cyclostationarity to Detect Changes in Gear Surface Roughness using Vibration Measurements

Acknowledgements
The authors would like to thank Mr Yunzhe Yang and Dr Chongqing Hu for
their help in the experiments. This research was supported by the Australian Re-
search Council, through Discovery Project [DP160103501].

References
Antoni, Jérôme. 2007. 'Cyclic spectral analysis in practice', Mechanical Systems
and Signal Processing, 21: 597-630.

———. 2016. 'Fast_SC Program, Mathworks File Exchange', Accessed 3 March.

https://au.mathworks.com/matlabcentral/fileexchange/60561-fast-sc-x-nw-alpha-
max-fs-opt-?focused=7143132&tab=function.

Antoni, Jérôme, Frédéric Bonnardot, A Raad, and Mohamed El Badaoui. 2004.

'Cyclostationary modelling of rotating machine vibration signals', Mechanical
Systems and Signal Processing, 18: 1285-314.

Antoni, Jérôme, Ge Xin, and Nacer Hamzaoui. 2017. 'Fast computation of the
spectral correlation', Mechanical Systems and Signal Processing, 92: 248-77.

Capdessus, C., M. Sidahmed, and J. L. Lacoume. 2000. 'Cyclostationary

processes: application in gear faults early diagnosis', Mechanical Systems and
Signal Processing, 14: 371-85.

Krantz, Timothy. 2005. "The influence of roughness on gear surface fatigue." In

NASA Technical Report NASA/TM-2005-213958.

Mark, William D. 1978. 'Analysis of the vibratory excitation of gear systems:

Basic theory', The Journal of the Acoustical Society of America, 63: 1409-30.

Raad, Amani, Jérôme Antoni, and Ménad Sidahmed. 2008. 'Indicators of

cyclostationarity: Theory and application to gear fault monitoring', Mechanical
Systems and Signal Processing, 22: 574-87.

Randall, R. B. 1982. 'A New Method of Modeling Gear Faults', Journal of

Mechanical Design, 104: 259-67.

Yang, Y., W. A. Smith, P. Borghesani, Z. Peng, and R. B. Randall. 2015.

"Detecting changes in gear surface roughness using vibration signals." In
Acoustics 2015. Hunter Valley, Australia.

View publication stats