Professional Documents
Culture Documents
Assessment of Gear Damage Monitoring Techniques
Assessment of Gear Damage Monitoring Techniques
Each gear damage monitoring technique has its merits and limitations. This paper
experimentally investigates the sensitivity and robustness of the currently well-accepted
techniques: phase and amplitude demodulation, beta kurtosis and wavelet transform. Four
gear test cases were used: healthy gears, cracked, "led and chipped gears. The vibration
signal was measured on the gearbox housing and processed, online, under three "ltering
conditions: general signal average, overall residual and dominant meshing frequency resid-
ual. Test results show that beta kurtosis is a very reliable time-domain diagnostic technique.
Phase modulation is very sensitive to gear imperfections, but other information should be
used to con"rm its diagnostic results. Continuous wavelet transform provides a good visual
inspection especially when residual signals are used. The diagnosis based only on dominant
meshing frequency residual, however, should not be used independently for gear health
condition monitoring, it may give false alarms.
2001 Academic Press
1. INTRODUCTION
Typical applications of gearboxes include electric utilities, automotive industry, ships and
helicopters. A practical and robust monitoring system is critically needed to provide the
earliest warning of damage or malfunction in order to avoid sudden failures. Currently,
there are three types of approaches to the detection of faults in geared systems: acoustic
signal analysis, debris monitoring and vibration analysis. The vibration-based diagnosis has
been the most popular monitoring technique because of ease of measurement. When
vibration features of a component are obtained, its health condition can be determined by
comparing these patterns with those corresponding to its normal and failure conditions.
This pattern classi"cation process can be conducted by visual inspection or by inference
approaches [1, 2].
There are many vibration-based monitoring techniques currently available for the
detection of gear faults. According to the analysis domain, they can be classi"ed into
frequency/cepstrum analysis, time/statistical analysis and time}frequency analysis. A brief
review of each is given below.
signature in the spectrum caused by a gear fault could be recognised [4, 5]. For complicated
gear systems, however, it is di$cult to identify faults from the spectrum or the cepstrum
because of the large number of components involved.
Figure 1. Gear damage used in the tests: (a) healthy gear tooth, (b) tooth with a 20% crack at its root, (c) "lled
tooth in the middle with 50% tooth surface and (d) chipped tooth from pitch point to the top at 20% of its
thickness.
synchronous average, phase and amplitude demodulation, beta kurtosis and continuous
WT. For the sake of completeness, the basic formulation of each of these methods is
included in the appendix.
We begin our presentation with a description of the experimental set-up, the data
acquisition and preprocessing methods utilised. Four gear test cases will be used as shown
in Fig. 1: (a) a healthy gear, (b) one with a transverse cut of 20% of root thickness to simulate
a tooth crack, (c) a "led tooth around the pitch point (50% of tooth surface) to simulate
spalling damage and (d) a chipped tooth, from zero thickness at pitch point to 20%
thickness at the tooth top, to simulate serious wear and scoring fault. One typical processing
result from each gear test case will be discussed in detail. Three speeds at three torque levels
will be used for each gear condition. Moreover, three data preprocessing methods will be
utilised: using the original time signal average, using the data after removing gear meshing
frequency (MF) and its harmonics, and using the data after band-pass "ltering the averaged
signal around the dominant MF harmonic. Due to the very large number of plots generated
in the course of the present study, only results from a sample test case will be presented for
each gear condition. A summary table including a subjective assessment of success of
diagnostics, in each test case, will be provided. It will be shown that the beta kurtosis is
a very reliable time-domain diagnostic technique. Phase modulation is very sensitive to gear
imperfections, but other information should be used to con"rm its diagnostic results.
Continuous WT provides a good visual inspection specially when the residual signals are
used. The diagnosis based only on dominant MF residual, however, should not be used
independently for gear health condition monitoring; it may give false alarms.
2. EXPERIMENTAL SET-UP
The gear test apparatus used in this study is shown schematically in Fig. 2. It consists of
two 1 HP permanent magnet DC motors (PM9100ATF) and a single-stage gearbox. A pair
of spur gears, with the module of 1.5 mm, is tested. The driving gear has 16 teeth and the
driven gear has 14 teeth. The motors and gearbox are mounted onto a sti!ened I-beam
which is anchored to a massive concrete block. The speed controller allows gearbox
operation in the range of 200}1400 rpm. The load torque is provided by a network of power
resistors connected to the load motor. The power resistors o!er a quiet dissipation
mechanism, and thus they do not contribute any additional vibration to the test rig. The
speed of the drive motor and the resistor network can be adjusted continuously to
accommodate the range of speed/torque operating conditions. The vibration is measured
with an accelerometer (PCB308) with a sensitivity of 104 mV/g, mounted on the gearbox
housing in the direction of gear action. An optical sensor is mounted in proximity to
a slotted disc attached to the driving shaft, which provides a one-pulse-per-revolution signal
908 W. Q. WANG E¹ A¸.
to be used for the synchronous averaging process. The signals from both sensors are
properly ampli"ed, then fed to the anti-aliasing "lter sets, which are of Bessel type (KH3362)
as these provide minimum-phase distortion. The "lter cut-o! frequency for the vibration
signal was set high enough to capture the "fth harmonic of the gear MF, while the cut-o!
frequency for the reference signal was set su$ciently high (e.g. 12 kHz) to reduce pulse wave
distortion after "ltering.
where ¹ is the period of one revolution, and N is the number of teeth of the monitored gear.
4. SIGNAL PROCESSING
All diagnostic algorithms employed here were coded in Matlab environment. The
vibration signal was processed following three "ltering conditions: (a) using the original
TSA data without further "ltering, (b) using the overall residual obtained by digitally
"ltering out the gear MF and its harmonics and (c) using the dominant MF residual,
obtained by band-pass "ltering the signal average around the dominant MF harmonic, with
the bandwidth of 18 shaft orders, and setting its amplitude at that frequency (bin) to zero.
These "ltering approaches have been implemented in several works in the literature,
e.g. [7, 19].
5. RESULTS
Tests were conducted with torques of 3, 4.5 and 6 Nm, each at shaft speeds of 400, 800 and
1200 rpm. The maximum speed and torque were limited by the test rig itself. In each test, the
vibration was, online, monitored and the data was collected only after the signature pattern
stabilised. Sample results are discussed below.
Figure 5. Vibration of healthy gear: processed based on dominant meshing frequency residual signal.
[Fig. 4(f )] has changed signi"cantly through this "ltering process, and it is di$cult to draw
any conclusion from this map.
The signatures in Fig. 5 are the processing results based on the residual of the dominant
MF. It can be seen that there are big amplitude irregularities around 903 and 2703 in TSA
trace in Fig. 5(a). These were caused by gear out-of-roundness; the o!-the-shelf gears were
bored to "t the shaft, a process in which the gear rim thickness was reduced greatly. Some
eccentricity was inevitably induced in the machining process as well as in clamping the gear
to the shaft using set screws. It is interesting to see that the e!ect of gear run-out is also
manifested in all other "gures, 5(b)}5(f). On the one hand, the present "ltering technique can
give false alarms regarding tooth damage or the like, but on the other hand, a careful
examination of the obtained maps can help identify the eccentricity problem associated with
manufacturing or assembly processes.
TABLE 1
Test result comparison of diwerent techniques under diwerent operating conditionsR
RNote: CI: Clear (fault) indication; SI: some (fault) indication; NI: no (fault) indication; BK: beta kurtosis;
AM: amplitude modulation; PM: phase modulation; WA: wavelet amplitude; MF: meshing frequency.
di$cult to be utilised as an independent diagnosis tool, its results are not listed either. We
should emphasise here that the above classi"cation of gear conditions, CI, SI and NI, is
rather coarse and was created using a subjective assessment of the obtained maps. Further
research is needed to quantify the di!erent diagnostic measures, and ultimately develop
a metric that combines these measures in a way that increases the accuracy and robustness
of fault detection.
Before discussing the results in Table 1, let us "rst present sample maps obtained
for each damage type, all at the same speed of 1200 rpm and torque of 4.5 N m.
Similar to the case of healthy gears, the plots will be presented using three kinds of data sets,
TSA signal, overall residual and dominant MF residual, as explained earlier. These samples
will help illustrate our classi"cation of diagnosis of the gear health condition into NI, SI
and CI.
GEAR DAMAGE MONITORING TECHNIQUES 913
5.2.1. Cracked tooth
Results of the cracked tooth for the case of using TSA signal without further "ltering are
shown in Fig. 6. From the signal average trace in Fig. 6(a), there exists some indication (SI)
of irregularities around 1803 (9th tooth). The amplitude and the corresponding phase
signatures in Fig. 6(c) and (d) clearly show (CI) the existence of a tooth fault around 1803.
Beta kurtosis in Fig. 6(b) also provides a clear tooth damage indication (CI) but around the
10th tooth. The reason for a di!erent tooth location is that beta kurtosis analysis is based
on data blocks, each corresponding to a tooth period. Since the beginning of a tooth period
used in the calculation might not be the exact start position in its actual meshing period,
some data associated with this tooth block could be shared between the neighbouring
blocks. Accordingly, the real damaged tooth might be the one adjacent to that visible in the
kurtosis map. Wavelet phase bifurcation features in Fig. 6(f ) clearly identify (CI) the tooth
fault, as indicated by the arrow. At the same time, there is also some irregular expression
(SI) in wavelet amplitude map [Fig. 6(e)] around the 2nd MF harmonic, but about one
tooth period ahead of that in the phase signature. This is consistent with the observations in
[19]. The similar phase lag can also be identi"ed from the amplitude and phase modulation
traces from 1803 to 2003.
The plots in Fig. 7 show the results based on the overall residual signal of the cracked
gear. From this set of results, little irregularities (SI) can be recognised around 1803 in the
traces of signal average [Fig. 7(a)] and amplitude modulation [Fig. 7(c)]. The phase
modulation [Fig. 7(d)] does not show an indication (NI) of this tooth fault. The beta
kurtosis, however, indicates the existence of the gear fault on the 9th tooth clearly (CI). Even
though no distinctive signatures can be recognised from wavelet phase map [Fig. 7(f )],
wavelet amplitude [Fig. 7(e)] could vividly indicate (CI) this gear fault around the second
MF harmonic.
Figure 6. Vibration of gear with a cracked tooth: processed based on TSA signal.
914 W. Q. WANG E¹ A¸.
Figure 7. Vibration of gear with a cracked tooth: processed based on overall residual signal.
Figure 8 shows the results of the cracked gear processed based on the dominant
MF residual. The signal average [Fig. 8(a)] and amplitude modulation [Fig. 8(c)]
have some abnormal signatures in the middle of the traces (SI). This irregularity
can be recognised from the kurtosis in Fig. 8(b) on tooth 8 (SI). At the same time, this
tooth damage is also clearly indicated (CI) in wavelet amplitude map [Fig. 8(e)]. However,
there are no distinctive signatures (NI) for this damaged tooth in both phase plots in Fig.
8(d) and (f ).
Figure 8. Vibration of gear with a cracked tooth: processed based on dominant meshing frequency residual
signal.
Figure 9. Vibration of gear with a "lled tooth in the middle: processed based on TSA signal.
916 W. Q. WANG E¹ A¸.
Figure 10. Vibration of gear with a "lled tooth in the middle: processed based on overall residual signal.
E The beta kurtosis successfully identi"ed the presence of damage correctly in all cases,
from the original time synchronous average, from the overall residual and from the
dominant MF residual. In the majority of experiments, it gave a clear indication (CI) of
the presence of the damage while in the remaining much fewer experiments, it gave some
indication (SI) of the gear damage.
E Following the beta kurtosis, the phase modulation of the original TSA signal
was also successful in detecting the damage in all experiments. The phase modula-
tion, however, gave inconsistent results, and mostly failed, when signal residuals were
used.
GEAR DAMAGE MONITORING TECHNIQUES 917
Figure 11. Vibration of gear with a chipped tooth on the top: processed based on TSA signal.
Figure 12. Vibration of gear with a chipped tooth on the top: processed based on overall residual signal.
918 W. Q. WANG E¹ A¸.
E The amplitude modulation also identi"ed the presence of damage successfully in all
experiments when the original TSA signal was used. The clarity of identi"cation,
however, deteriorated when the overall residual was used, and mixed results were
obtained when the dominant MF residual was employed.
E Similar to the beta kurtosis, the wavelet amplitude map also indicated the presence of
damage in all experiments, and for all signal "ltering cases, successfully.
E If we concentrate on the results obtained with the original TSA data, we see that it is only
in very few cases that all monitoring techniques gave the same clear indication (CI) of the
presence of damage; typically, the assessment varied from some indication (SI) to clear
indication (CI). The assessment varied widely in the case of residuals, from no indication
(NI) altogether to clear indication (CI).
Perhaps other observations can be made from the above table regarding the e!ects of
operating conditions and damage types on the e!ectiveness of monitoring techniques and
so on. But what is most important is to realise that it is the beta kurtosis and wavelet
amplitude, for any of the signal processing approaches employed here, that yield more
consistent results. If we complement the capabilities of the beta kurtosis and wavelet
amplitude with the phase modulation from the original signal average, the e!ectiveness of
the monitoring system will enhance even further.
6. CONCLUSION
From this experimental investigation, it has been found that beta kurtosis and wavelet
transform amplitude are reliable techniques for gear health condition monitoring; they have
less sensitivity to the load, speed and frequency bandwidth utilised. Their e!ectiveness can
be complemented by the phase modulation of the TSA signal. Research is currently
underway to quantify the above measures and to establish an inference system for online,
automated, gear fault diagnostics.
7. ACKNOWLEDGEMENTS
The "nancial support provided for this work by Natural Sciences and Engineering
Research Council of Canada, and Sanden Machine Ltd. is greatly appreciated.
REFERENCES
1. V. JAMMU, K. DANAI and D. LEWICKI 1998 Journal of Mechanical Design 120, 100}105.
Structure-based connectionist network for fault diagnosis of helicopter gearboxes.
2. Q. SUN, X. LI and B. XU 1999 Proceedings of the 1999 ASME Design Engineering ¹echnical
Conferences, DETC99/VIB-8329. Gearbox fault detection using neural networks technology, Las
Vegas, NY.
3. S. BRAUN 1986 Mechanical Signature Analysis, ¹heory and Applications. New York: Academic
Press.
4. R. RANDALL 1980 2nd International Conference of <ibration in Rotating Machinery, 169}174.
Advances in the application of cepstrum analysis to gearbox diagnostics. Cambridge, U.K.
5. H. TANG, J. CHA and Y. WANG 1991 ASME Design ¹echnical Conferences, De-Vol. 38, 141}144.
The principle of cepstrum and its application in quantitative fault diagnostics of gears, Miami, FL.
6. P. D. MCFADDEN 1989 Mechanical Systems and Signal Processing 3, 87}97. Interpolation
techniques for time domain averaging of gear vibration.
7. P. D. MCFADDEN 1986 Journal of <ibration, Acoustics, Stress, and Reliability in Design 108,
165}170. Detecting fatigue cracks in gears by amplitude and phase demodulation of the meshing
vibration.
GEAR DAMAGE MONITORING TECHNIQUES 919
8. G. DALPIAZ and U. MENEGHETTI 1991 ND¹&E International 24, 303}306. Monitoring fatigue
cracks in gears.
9. F. ISMAIL, H. MARTIN and F. OMAR 1995 Proceedings of the ASME Biennial Conference on
<ibration and Noise, Vol. DE-84-1, 1413}1418. A statistical index for monitoring toothcracks in
a gearbox. Boston, MA.
10. F. GOLNARAGHI, D. LIN and P. FROMME 1995 ASME Biennial Conference on <ibration and
Noise, Vol. DE-84-1, 121}127. Gear damage detection using chaotic dynamic techniques: a pre-
liminary study. Boston, MA.
11. D. LIN, F. GOLNARAGHI and F. ISMAIL 1997 Journal of Sound and <ibration 208, 664}670. The
dimension of the gearbox signal.
12. W. J. WANG and P. D. MCFADDEN 1993 Mechanical Systems and Signal Processing 7, 193}203.
Early detection of gear failure by vibration analysis*1, calculation of the time}frequency
distribution.
13. A. MERTINS 1999 Signal Analysis. New York: John Wiley & Sons Ltd.
14. P. D. MCFADDEN and W. J. WANG 1992 Institute of Mechanical Engineers International
Conference on <ibrations in Rotating Machinery 387}393. Analysis of gear vibration signature by
the weighted Wigner}Ville distribution, Bath, U.K.
15. G. STRANG and T. NGUYEN 1996 =avelets and Filter Banks. Wellesley: Cambridge Press.
16. W. J. WANG and P. D. MCFADDEN 1995 Mechanical Systems and Signal Processing 9, 497}507.
Application of orthogonal wavelets to early gear damage detection.
17. W. J. WANG and P. D. MCFADDEN 1996 Journal of Sound and <ibration 192, 927}937. Application
of wavelets to gear box vibration signals for fault detection.
18. S. T. LIN and P. D. MCFADDEN 1995 Institute of Mechanical Engineers 2nd International
Conference on Gearbox, Noise, <ibration, and Diagnostics, 59}72. Vibration analysis of gearbox by
linear wavelet transform. London, U.K.
19. D. BOLLAHBAL, F. GOLNARAGHI and F. ISMAIL 1999 Mechanical Systems and Signal Processing
13, 423}436. Amplitude and phase wavelet maps for the detection of cracks in geared systems.
vibration will have a corresponding modulation. These changes can be de"ned by the
amplitude and phase modulation functions a (t) and b (t), respectively [7]. The modulated
K K
gear meshing vibration x(t) can be expressed as
+
x(t)" X (1#a (t)) cos (2nmNf t# #b (t)). (A.3)
K K P K K
K
After "ltering out the MF and its harmonics, the overall residual can be approximated by
+
x(t)+ X a (t) cos (2nmNf t# #b (t)). (A.4)
K K P K K
K
If nNf is the dominant MF harmonic, band-pass "ltering around this frequency, the
P
dominant MF residual can be approximately expressed as
Since the gear vibration signal is periodic, the above equations can be implemented in
the frequency domain by a straightforward "ltering process. As an example, let us calculate
the dominant MF residual at frequency nNf , using a bandwidth of k shaft orders. If
P
X( f ) is the spectrum of x(t), modify X( f ) by setting all its values to zero except
f"[(nN#1)!0.5k, (nN#1)#0.5k], then setting the amplitude at the centre frequency
f"nN#1 (or a small band around it) to zero. Here, the shaft order at f"nN#1 instead
of nN is used because f"1 corresponds to the fundamental shaft rotation rate.
The dominant MF residual can be obtained by
Factor 2 appears because only a one-sided spectrum is used here. F\ ( ) ) denotes the inverse
Fourier transform, and real [ ) ] is the real part function.
At the same time, the analytic signal can also be determined by
Beta kurtosis is the fourth moment of the beta function. Suppose x is a continuous
random variable, normalised between 0 and 1, then beta probability density function f (x) is
V
de"ned as
1
f (x)" x?\ (1!x)@\
V B(a, b)
where B(a, b)" x?\ (1!x)@\ dx, a and b are numerical positive constants.
GEAR DAMAGE MONITORING TECHNIQUES 921
The kth moment about the mean, k, can be expressed as,
> 1
m" (x!k)I f (x) dx" (x!k)I x?\ (1!x)@\ dx. (A.9)
I V B(a, b)
\
Using binomial expansion,
I k!
(x!k)I" (!1)P kP xI\P
r!(k!r) !
P
equation (A.9) becomes
1 I k!
m" (!1)P kP xI\Px?\ (1!x)@\ dx
I B (a, b) r! (k!r)!
P
1 I k!
" (!1)P kP x ?>I\P\ (1!x)@\ dx
B (a, b) r!(k!r)!
P
1 I k!
" (!1)P kP B (a#k!r, b). (A.10)
B (a, b) r!(k!r)!
P
According to the relationship between Beta and Gamma functions,
C (a)"(a!1) C (a!1)
a
m "k! "0 (The "rst moment about the mean is zero) (A.11)
a#b
ab
m "p" (The second moment is the variance) (A.12)
(a#b) (a#b#1)
3ab (2a!2ab#ab#ab#2b)
m " (A.13)
(a#b) (a#b#1) (a#b#2) (a#b#3)
m 3(a#b#1) (2a!2ab#ab#ab#2b)
BK" " . (A.14)
(p) ab (a#b#2) (a#b#3)
For tooth-based BK analysis, k and p are the mean and variance of the tooth data block
corresponding to each tooth meshing period. Sometimes BK reciprocal, 1/BK, is used for
visual inspection [9].
>
=¹ (t, s)" x(q) (s w (s(t!q)) dq (A.15)
V
\
922 W. Q. WANG E¹ A¸.
where s and t are the scale and time parameters that produce dilation and translation,
respectively.
The crucial aspect in calculating the WT is the choice of an appropriate wavelet basis,
which should be determined by the signal itself and the purpose of the analysis. For the gear
damage diagnosis, through a series of tests and comparisons, it has been found that the
Morlet wavelet gives superior results. Morlet wavelet is a modulated Gaussian, de"ned as
t
w(t)"exp ! exp (i2n f t) (A.16)
2b
f 2nb
=( f/s)"b(2n exp 2nb !f "b(2n exp ( f!sf ) (A.19)
s s
i.e. the centre frequency of the dilated wavelet is sf , proportional to s.
To prevent the possible phase aliasing e!ects [19], the magnitude of "(1/s) =( f/s)" should
be used in lieu of (1/s) = ( f/s) in equation (A.18), which becomes
=¹ (t, s)"(s F\ [X ( f ) "(1/s)=( f/s)"]. (A.20)
V
In the implementation of equation (A.20), X ( f ) is calculated only once, while "(1/s) =( f/s)"
is calculated for all values of the scale parameter s.
WT amplitude and phase maps can then be determined from =¹ (t, s) over the
V
time-scale domain. In this paper, the linear scale was used, and the WT maps were displayed
in the time}frequency plane using the relationship f"sf .