DETECTION OF A GEAR COUPLING

MISALIGNMENT IN A GEAR TESTING DEVICE

François Combet*, Nadine Martin*, Pierre Jaussaud*, François Léonard**

Laboratoire des Images et des Signaux*, BP 46

38402 St Martin d’Hères Cedex, France
Hydro-Québec**, IREQ, 1800 Lionel-Boulet, Varennes,
J3X 1S1 Québec, Canada,
[email protected]

Abstract
On the gear testing device of the LIS laboratory (Grenoble, France), a total of 20
signals are synchronously recorded including shaft acceleration signals in several
positions, torque, rotation speed, optical encoder signal of both shafts and currents
and tensions of the non-synchronous training motor. After about 3500 hours of using
this device, the driving gear shaft broke at the gear coupling position with the training
motor shaft, which created an impressive helicoïdal crack . This failure could be due
to a flexion strain caused by a misalignment of the two shafts, but this fault was not
detected before. In this paper we compare results of different signal processing
methods for the detection of this fault. Particularly, we use the phase spectrogram
whose advantage is to be able to highlight a slight phase modulation in signals that is
not detectable on the amplitude spectrogram. We also attempt to detect the fault using
spectral analysis, a non-stationary modelling based on Prony’s model and comparison
of signals recorded at different times before the crack: the acceleration signals, the
torque, and the torsional vibration signal obtained from the shaft encoder signal. The
detection results are compared and discussed between the different methods.

1. INTRODUCTION

In rotating machinery rotor misalignment is known to generate vibration problems. It

is measured that about 60 per cent of faults was caused by misalignment [Li 01]. The
moment created by a misalignment in a gear coupling is quite complex but its
fluctuations occur at twice the shaft speed [Dewell 84]. Thus, the presence of the 2
times shaft speed vibration is known to be an indicator of a misalignment.
Furthermore, coupling misalignment does not significantly alter the 1x vibration
response but the 2x vibration response is clearly a characteristic signature of
misaligned shafts [Sekhar 95]. Misalignment initially affects the 1x response but in
the case of severe misalignment, a 2x component appears in the response [Jordan 93].
So an increase of the 2x shaft speed vibration corresponds to an increase of the
misalignment. The physical source of these effects may be modelled as a rotor bend
for the 1x response and as a rotor asymmetry for the 2x response [Sinha 04].
In this paper, different signals measured on our gear testing device are analysed
in order to attempt to find criteria for an early detection of faults that could appear in
the gear coupling. Two signal processing methods are used. The first method takes
advantage of the non-stationarity of the observed measured. It is based on a classical
spectrogram followed by a time-frequency tracking of one harmonic modulation.
Then a spectral analysis of the extracted modulation yields interesting parameters for
supervising a running system. The second method takes advantage of a known model
of the observation and proposes the tracking of local model parameters. In this paper,
we aim at studying time-frequency local approaches.
The proposed methods actually do not need data means or synchronous
measures. Results are presenting on the gear testing device of the lab (GOTIX bench
of LIS-Grenoble France).
In this paper, we focus on the results and references are given for going insight
the methods we do not describe. We first describe the experiment and we analyse the
torsional vibration and the torque signals after synchronous averaging. We then use
the frequency spectrogram for analysing slight frequency modulations in the
acceleration signals. A time-frequency Prony analysis is finally applied in order to
track the evolution of Prony’s parameters.

2. DESCRIPTION OF THE EXPERIMENT

Fig.1: Left: gear testing device, middle: tested gears, right: cracked gear coupling.
The gear coupling that links the training motor shaft to the gear shaft appears in the
lower right corner of the middle figure.

The gear testing device of the LIS laboratory (Grenoble, France) is used to
study and develop new signal processing methods adapted to the early detection of a
crack or the wear of the tested gears. Twenty signals are synchronously recorded on
this device: accelerations at the bearing positions of the shafts, torque, rotation
speeds, encoder signals and tachometer signals of both shafts and currents and
tensions of the non-synchronous training motor. The sampling frequency is 25 kHz
except for the encoder and tachometer signals (100 kHz). This device (fig.1) has been
running for 3500 hours when the gear coupling that links the motor shaft to the gear
shaft suddenly broke, creating an impressive helicoidal crack of the gear shaft at its
coupling position (fig.1 right). This appears while we were testing new gears for 300
hours under normal load conditions. The gear coupling has never been changed for
the 3500 hours. GOTIX testing device is described on our website
(http://www.lis.inpg.fr/english/research/projets/aproj21.htm) where all data can be
loaded. Only one personal identification is required.

3. FIRST OBSERVATIONS

3.1. Torsional vibration signal

The torsional vibration signal of the shaft is obtained by phase demodulating the shaft
encoder signal [Sweeney 96]. The encoder signal is first band-pass filtered around the
carrying frequency (1024x12.3=12.6kHz). The band-pass filter is chosen of width +/-
8kHz, which allows high modulation frequency analysis. The spectrum of the
torsional vibration is rich in harmonics: it contains harmonics of the driving shaft
speed (12.3Hz) of the driven shaft speed (46.7Hz, gear ratio is 57/15=3.8), and also
harmonics of the 50Hz (current frequency). The time waveform of the vibration is so
complex to analyse. In order to retain the driving shaft synchronous components only
and to filter other components, we perform a synchronous averaging based on the
tachometer signal of the driving shaft. Figure 2 shows the torsional vibration at time
3200h averaged on 125 rotations of the shaft (25 times the gear coincidence period).
The main harmonics are the 1x shaft speed, 57x (gear frequency) and also 18x and
78x. These two last harmonics are likely to be due to the motor constitution.
In order to detect a fault that could appear in the gear coupling, we computed
the averaged torsional vibration at different instants from 3200h (beginning of the
new test) to 3465h (a few hours before the crack), but we did not find particular
evolution of the time waveform and of the main harmonic amplitudes. Since a
misalignment fault is characterised by an increase of the second harmonic, we plotted
the evolution of the 2x frequency modulation amplitude and compared it to the
fundamental one (figure 3). The 2x modulation slightly increases at 3385h and 3455h,
which could be an indicator of a misalignment fault. We also plotted the 2x/1x
amplitude ratio that may be more valuable for the detection.

3.2. Torque signal

As for the torsional vibration signal, the torque is perturbed by harmonics of the
driven shaft frequency and of the 50Hz frequency. We so performed a synchronous
averaging in order to retain the driving shaft harmonics only. The averaged and zero-
mean torque at time 3200h is shown on figure 4: the 18x shaft speed harmonic is here
preponderant, which confirms that this component must be due to the training motor.
The 1x and 57x are also present.
 We computed the averaged torque at different instants from 3200h to 3465h,
and we plotted the evolution of the 1x and 2x shaft speed modulation amplitudes
(figure 5). The 2x modulation turns out to decrease at the end of the experiment,
unlike the torsional vibration. The torque seems here to be difficult to interpret for the
detection of a fault.

synchronous averaging of the torsional vibration

0.1

0.05

-0.05

-0.1
0 200 400 600 800 1000 1200 1400 1600 1800 2000

harmonics
50
57x 78x
1x
18x 114x
dB

-50
0 50 100 150 200
rank

Fig.2: Synchronous averaged torsional vibration signal at 3200h and its

harmonics: main harmonics are 1x shaft speed, 57x (gear frequency), 18x and 78x.
-3
x 10 frequency modulation amplitude
8

6
1x
Hz

2
2x
0
3200 3250 3300 3350 3400 3450
frequency modulation amplitude ratio of 2x / 1x shaft speed

0.4
0.3

0.2

0.1
0
3200 3250 3300 3350 3400 3450
time (hours)

Fig.3: Evolution of the 1x and 2x shaft speed frequency modulation

amplitudes estimated from the torsional vibration signal. The 2x modulation
slightly increases at 3385h and 3455h (crack is at 3485h).
 synchronous averaged torque
1000

500

-500

-1000
0 200 400 600 800 1000 1200 1400 1600 1800 2000
harmonics
140
120
1x 18x 57x 114x
100
dB

80
60
40

0 50 100 150 200

rank

Fig.4: Synchronous averaged torque at 3200h and its harmonics: main

harmonics are at 1x shaft speed, 18x, 57x (gear frequency).
5 torque modulation amplitude
x 10
3
1x

2x
1

0
3200 3250 3300 3350 3400 3450
torque modulation amplitude ratio of 2x / 1x shaft speed

0.4

0.2

0
3200 3250 3300 3350 3400 3450
time (hours)

Fig.5: Evolution of the 1x and 2x shaft speed modulations of the torque.

4. ACCELERATION ANALYSIS BY THE FREQUENCY

SPECTROGRAM

4.1. The frequency spectrogram

On the well known power spectrogram, only the amplitude of the short-time Fourier
transform (STFT) is shown. Thus, a slight frequency modulation may not be clearly
visible on the power time-frequency plane. The frequency spectrogram [Léonard 00]
was developed as a complement to the power spectrogram. It exploits the phase
information of the STFT.
The frequency spectrogram is simply computed by the phase difference of the
STFT between two consecutive segments divided by the time shift between the
segments. Furthermore, its sensibility can be increased by applying an integer gain to
the result. The advantage is to be able to highlight a slight frequency or phase
modulation that does not clearly appears on the amplitude spectrogram. Figure 6
shows the power and frequency spectrograms of a simulated frequency modulated
sinusoid. The slight modulation is not evident on the power spectrogram itself but is
better visible on the frequency spectrogram. A cut can then be applied on it at the
modulated frequency in order to recover the modulation signal. This modulation can
actually be recovered by tracking the maxima position of the power spectrogram, but
this needs to highly interpolate the frequency axis in order to limit quantization
[Léonard 00]. Thus, the frequency spectrogram permits to decrease the calculation
time for recovering a frequency modulation. It also has the advantage to be more
visual in a first analysis.

150 0

10 0 0

0 0 .0 5 0 .1 0 .15 0 .2 0 .2 5 0 .3

time (s)

150 0

10 0 0

0 0 .0 5 0 .1 0 .15 0 .2 0 .2 5 0 .3

time (s)

10

-5

0 0 .0 5 0 .1 0 .15 0 .2 0 .2 5 0 .3

time (s)

Fig.6: top: power spectrogram of a simulated frequency modulated sinusoid,

middle: frequency spectrogram, bottom: modulation signal recovered by a cut of
the frequency spectrogram. Signal to noise ratio is 10dB.

4.2. Analysis of the acceleration signals

We applied the frequency spectrogram to the two acceleration signals denoted by
CH5 and CH6 which are the closest to the gear coupling. CH5 measures vertical
acceleration and CH6 measures horizontal and radial acceleration, both at the bearing
of the gear shaft near the coupling. The acceleration signals are dominated by the
harmonics of the gear frequency (700Hz).
We measured the frequency modulation of the second gear harmonic at 1400Hz
which here stronger than the fundamental. The rotation period of the driving shaft
corresponds to around 2032 samples. In order to track a modulation within the
rotation period, a size of 512 samples was chosen for the spectrogram gliding
window. Once the modulation signal is recovered by a cut of the frequency
spectrogram at 1400Hz, a spectral analysis can be applied on it. Figure 7 shows the
spectrum of the modulation signal of acceleration CH6 at the beginning of the new
test (3200h) and before the crack (3465h). By comparison of these two spectrums, the
frequency modulation amplitudes of the 1x shaft speed (12.3Hz) and of the 2x clearly
evolved from 3200h to 3465h: the 2x becomes higher than the 1x at 3465h. Figure 8
shows evolution of these harmonics. A fault seems to appear at 3300-3350h due to an
increase of the 2x/1x amplitude ratio. It also increases just before the crack (3465h)
for signal CH6.
The originality of the approach is the extraction of the modulation signal in the
time-frequency plane followed by a spectral analysis in order to estimate the
evolution of the modulation amplitude versus time.
2 .5
5

2
4

1.5
3

1
2

1 0 .5

0 20 40 60 80 0 20 40 60 80

Hz Hz

Fig.7: Spectrum of the modulation signal obtained from the frequency spectrogram
of acceleration CH6 at time 3200h (left) and 3465h before the crack(right).
frequency modulation amplitude of the second gear harmonic of acceleration CH5 frequency modulation amplitude of the second harmonic of acceleration CH6
5 8

4 6
1x
3

Hz
4
Hz

2
2x 2
1

0 0
3200 3250 3300 3350 3400 3450 3500 3200 3250 3300 3350 3400 3450 3500

frequency modulation amplitude ratio 2x / 1x frequency modulation amplitude ratio 2x / 1x

2 2

1.5 1.5

1 1

0.5 0.5

0 0
3200 3250 3300 3350 3400 3450 3500 3200 3250 3300 3350 3400 3450 3500
time of use (hours) time of use (hours)

Fig.8: Frequency modulation amplitude at 1x and 2x shaft speed of the 2nd gear
frequency harmonic (1400Hz) for accelerations CH5 (left) and CH6 (right).

5. PRONY ANALYSIS

We finally performed a time-frequency Prony analysis [Martin 04] to the

synchronous averaged acceleration signals CH5 and CH6. We observed the evolution
of the Prony parameters (frequencies, dampings, amplitudes and phases) versus the
time of analysis. The main evolution was observed for the damping parameter. Figure
9 shows evolution of the mean value and of the standard deviation of the estimated
damping parameter. Damping is often higher for a cracked structure.
damping parameter of CH5 damping parameter of CH6
5000 9000
mean value mean value
4500 standard deviation 8000 standard deviation

4000 7000
3500
6000
3000
5000
2500
4000
2000
3000
1500

1000 2000

500 1000
3200 3250 3300 3350 3400 3450 3500 3200 3250 3300 3350 3400 3450 3500
time (hours) time (hours)

Fig.9: Evolution of the mean value and standard deviation of the damping parameter
of Prony analysis applied to the averaged acceleration signals CH5 and CH6.
 6. CONCLUSION

In a diagnostic and surveillance purpose, this paper presents two signal processing
approaches in order to attempt to find criteria for early fault detection. The frequency
spectrogram is interesting in order to demodulate frequency modulations when the
modulation frequency is not known and/or apt to change before the occurring of a
crack. The only a priori knowledge is the modulation bandwidth, whose choice is not
critical. The algorithm is simple with a low computation cost. The Prony analysis is a
local estimation of model parameters and is suitable for detecting the modification of
shocks induced in a system. In our experimentation, the tracking of the damping
factor seems to be decisive. Contrary to the first one, this method has a higher
computation cost.
The proposed methods have been tested on GOTIX bench of LIS-Grenoble. We
observed from time 3200h (beginning of experiment) to 3465h (before the crack) an
increase of the 2x shaft speed modulation amplitude in the torsional vibration signal
and in the acceleration signals close to the gear coupling. This could be an indication
that the misalignment increased in the gear coupling, and the friction between the
teeth due to the created moment finally led to the crack of the shaft in the coupling.
Both approaches are working in time-frequency space and can be applied to all
observations including harmonic modulations or shocks. They are adapted for
tracking changes in a great amount of data. The next step now is to deduce a detection
algorithm. Works are in progress in that way trying to define a detector, which will
not be sensitive to the choice of a threshold, as in [Hory 02].

REFERENCES
[Dewel 00] D.L. Dewell and L.D. Mitchell “Detection of a misaligned disk coupling using spectrum
analysis”, Trans. of the ASME –Journal of Vibration Acoustics Stress and Reliability in Design 106
(1984) 9-16.
[Hory 02] C. HORY, N. MARTIN, A. CHEHIKIAN. Spectrogram segmentation by Means of
Statistical Features for Non-Stationary Signal Interpretation. IEEE Transactions on Signal
Processing,Vol.50, n°12, pp 2915-2925, 2002.
[Jordan 93] M.A. Jordan, “What are orbit plots, anyway?”, Orbit, December (1993) 8-15.
[Léonard 00] François Léonard, “Phase spectrogram and frequency spectrogram”, Traitement du
Signal 2000, Vol. 17 n°4, 269-286.
[Li 01] M. Li and L. Yu, “Analysis of the coupled lateral torsional vibration of a rotor-bearing system
with a misaligned gear coupling”, Journal of Sound and Vibration (2001), 243(2), 283-300.
[Martin 04] N. Martin, P. Jaussaud and F. Combet, "Close shock detection using time-frequency Prony
modeling", Mechanical Systems and Signal Processing, Vol. 18, No. 2, Mar. 2004.
[Sekhar 95] A.S. Sekhar and B.S. Prabhu, “Effects of coupling misalignment on vibrations of rotating
machinery”, Journal of Sound and Vibration (1995), 185(4), 655-671.
[Sinha 04] J.K. Sinha, A.W. Lees, M.I. Friswell, “Estimating unbalance and misalignment of a flexible
rotating machine from a single run-down”, Journal of Sound and Vibration, in press.
[Sweeney 96] P.J. Sweeney, R.B. Randall, “Gear Transmission Error Measurement using Phase
Demodulation”, Proc. of the Institution of Mechanical Engineers (1996), Part C, Journal of Mechanical
Engineering Science, Vol. 210, Iss. 3, pp. 201-213.