Processing for Improved Spectral Analysis

Eric Bechhoefer1, Brandon Van Hecke and David He 2

1
Green Power Monitoring Systems, LLC, Vermont, 05753, USA
[email protected]
2
Dept of Mechanical and Industrial Engineering, The University of Illinois at Chicago, Illinois, 60607 USA

ABSTRACT of that energy will result in variance in the trend, or just an

inaccurate estimate of component damage.
The Fast Fourier Transform (FFT) is the work hoarse of
condition monitoring analysis. The FFTs’ assumption of Variance in shaft speed in rotating equipment is always
stationarity is often violated in rotating machinery. Even in present to varying degrees. The variation is due to:
a six second acquisition on a wind turbine, the shaft speed
can change by 5%. For Shaft and Gear analysis, this is • Limits in the control system bandwidth,
mitigated through the use of the time synchronous average. • Varying loads associated with the work the
For general spectrum analysis, or bearing envelope analysis, machine is producing,
there is no such mitigation: one hopes that the effect of • Or in the case of wind turbines, varying wind
variation in shaft speed is small. Presented is a time speed and torque ripple
synchronous resampling algorithm which corrects for Wind turbines pose a particularly difficult environment. The
variation in shaft speed, preserving the assumption of wind, as noted, is time varying. Additionally, because the
stationarity. This allows for improved spectral analysis, flow of wind is stalled in front of the tower, the lift on a
such as used in bearing fault detection. This is demonstrated blade as it passes in front the tower is reduced. This causes a
on a real world-bearing fault. 3/revolution torque ripple/change in shaft rate. Finally,
because of wind shear, the wind speed at the top of the rotor
1. INTRODUCTION arc is greater than the bottom of the arc. Figure 1 shows the
It would be hard to imagine the condition monitoring (CM) variance of a high speed shaft on a wind turbine, over a 6
of rotating equipment without the use of the Fast Fourier second acquisition. The instantaneous speed of the shaft is
Transform (FFT). Everything from simple spectrums (such seen to go from 30.9 Hz to a maximum of 32.01 Hz, or a
as Welch’s method for power spectral density), to more change in speed of 3.6%.
advanced analysis (amplitude modulation and frequency
modulation analysis (McFadden, 1985)) are dependent on
the FFT.
In using the FFT, the CM practitioner must understand the
base assumptions of continually differentiable (Gibbs
Effect), and stationarity. In general, window functions
(Hann, Hamming, etc.) are used to control or mitigate Gibbs
effect, while the time synchronous average (TSA,
McFadden 1987, Bechhoefer, 2009), is used to mitigate the
effect of non-stationarity in rotating machinery for shaft and
gear analysis.
The issue of non-stationarity is not well addressed for the
power spectral density (such as Welch’s method), or for
bearing analysis (envelope/heterodyne methods included). It
is assumed that the smearing of energy due to changing
shaft speed is small. For bearing analysis, the energy
Figure 1. Variation in High Speed Shaft over 6 Second
associated with a fault frequency is trended. A poor measure
Acquisition
 Annual Conference of Prognostics and Health Management Society 2013

Consider the effect of this variation in speed on the spectral • Xi is the amplitude of the kth mesh harmonic
content of a bearing fault frequency, where the cage, ball,
inner and outer race rates are: [0.42, 2.87, 9.46, 6.72]. The • fm(t) is the average mesh frequency
range of fault frequencies during this acquisition are: • ai(t) is the amplitude modulation function of the ith
Table 1. Bearing Rates feature harmonic.
Bearing\shaft Low: 30.9 High 32.01 • φi(t) is the phase modulation function of the ith
Cage 12.98 13.44 feature harmonic.
Ball 88.69 91.88
Inner Race 292.3 302.8 • Φi is the initial phase of harmonic k, and
Outer Race 207.7 215.1 • b(t) is additive background noise.

For the higher frequency bearing components, (inner/outer The mesh frequency is a function of the shaft rotational
race), the frequency difference is significant: 10 Hz. For the speed: fm = Nf(t), where N is the number of teeth on the
FFT, it poses a problem. Not only is there the issue of the gear and f(t) is the shaft speed as a function of time. For
spectral content smeared across a number of bins, but also bearings, N is the component rate, which is a non-integer
which shaft rate do you use (the mean shaft rate over the value based on the bearing geometry. As noted, because of
acquisition)? the finite bandwidth of the feedback control, or due to the
environment, there is some wander in the shaft speed. This
This issue of spectral spreading in the FFT is not academic. change in speed will result in smearing of amplitude energy
Consider the trend of an inner race fault on a high speed in the frequency domain.
shaft (Figure 2). The variance in the condition indicator is If a tachometer signal is present (such as a key phasor) and
proportional to the inner race energy. This is likely not due the ratio from the key phasor to the shaft under analysis, the
to some stochastic process associated with vibration, but vibration data can be resampled such that number of data
more a function measurement error due to variance in shaft points between one revolution and the next is the same. In
speed. the case of time synchronous averaging (TSA), the
ensemble average of EQ(1) is calculated summing each
revolution of resampled data, then dividing by the number
of revolutions during the acquisition. .
Since the radix-2 FFT is most commonly used, the number
of data points in one shaft revolution (rn) are interpolated
into m number of data points, such that:
• For all shaft revolutions n, m is larger than r, and
• m = 2ceiling (log2 (r)) (again assuming Radix 2 FFT)
However, this is not the case for the proposed synchronous
resample algorithm.
Since some other analysis process will be done on the
resampled signal (envelop analysis, for example) – a radix-2
length is not necessary. Instead, the largest time between
Figure 2. Trend of an inner race fault key phasor zero crossing is used (corresponds to the lowest
shaft rate during an acquisition) to calculate the resample
This increased variance affects both the threshold setting data length, l. Then, for each revolution, the current set of
process and alerting. Clearly if one does not have the luxury data points over one revolution are resampled to length l.
of sampling under steady state, a process is needed to
mitigate the change in shaft RPM. 2.1. Synchronous Resampling Algorithm

2. SYNCHRONOUS RESAMPLING For example, say the sample rate was 1000 samples per
second, and the lowest shaft rate was 10 Hz, for a .5 second
The model for vibration in a shaft in a gear box was given in acquisitions. The resample length, l, is 100. The number of
(McFadden 1987) as: data points between each key phasor is: 87, 92, 100, 95, 89,
37. For each shaft revolution, the data is resampled to length
x(t) = Σi=1:K Xi(1+ ai(t))cos(2πi fm(t)+ Φi)+b(t) (1)
l.: Rev 1: 87->100, Rev 2: 92->100, Rev 3: 100->100, Rev
where: 4: 95->100, Rev 5: 89->100. Note that for half of a second
of data, there is 500 data points – the remaining 37 data

2
 Annual Conference of Prognostics and Health Management Society 2013

were in the next, incomplete revolution, so the last 37 data

points are dropped.
Because of interpolation, the sample rate for each revolution
is now changed. To accurately determine the frequency
associated with a FFT bin, an apparent sample rate is
needed. The apparent sample rate is the original sample rate
* length of the resampled data / length of the original data:
1000 * 500/(463), or 1082.

2.2. TSA for Bearing Analysis

For shaft and gear analysis, existing TSA algorithms control
for changes in shaft speed. For bearing, because they do not
have integer number of shaft for a rate, the TSA is felt to be
inappropriate for three reasons. Frist, bearings are quasi-
stationary – there is always some slippage such that even
with correct geometry, the rates are not exact. This will Figure 3. Raw, resampled, and TSA spectrum
make the bearing component non-synchronous with the
TSA algorithm and in fact may separate the bearing signal In Figure 4, a detailed view of the spectrum is given from
out of the TSA. 2200 to 4000 Hz, showing that the resampled spectrum has
more spectral content than the raw spectrum or the TSA
Second, a bearing has rates for each component: cage, ball, spectrum. In the 9000 to 10,500 Hz view, spectral peaks are
inner and outer race. This would require the TSA to be run visible 9220, 9510 9800 and 10,090 Hz. The 290 Hz
four separate times for each bearing. While this may not be difference is close to the modulation rate of an inner race
a problem for off line analysis, it may exceed the resources fault, which was 292 Hz. The high, broadband spectrum is
of an on-line analysis system. Considering that any given indicative of bearing resonance. Because resonance is non-
shaft is supported by 2 to 3 bearings, which would require 8 synchronous, the TSA does not capture this bearing
to 12 TSA analyses. resonance.
Finally, in the evaluation of bearing health, it is important to
be able to see the relationship between the shaft, cage, ball,
inner and outer race fault features. For example, an inner
race fault that is modulated by shaft (e.g. side bands that are
1 shaft rate off of the inner race fault) is a more serious fault
than an inner race fault, as it indicates wear and clearance
issues on the shaft.

3. HIGH SPEED SHAFT BEARING FAULT

A commercial wind turbine with a 2 MW power output was
installed with a condition monitoring system. Data was
collected at 10-minute intervals. Increased inner race energy
on the high speed shaft bearing indicated a fault (Figure 1).
An inspection of the bearing showed that the inner race was
cracked. Using this data, the raw spectrum was compared to
the resampled spectrum and the TSA (Figure 3). The raw Figure 4. Detail of raw, resampled and TSA spectrum
and resampled spectrum was calculated using Welch’s
algorithm, with a 50% overlap. The spectrum length was the As an aside, plotting the TSA against the raw spectrum is a
same as the TSA length, so that the plots have similar bin good way to identify bearing faults: Tone not present in the
widths. TSA which are present in the raw spectrum can only be gear
mesh frequencies from other shafts in the gearbox, or a
The average shaft rate is 30.9 Hz. This shaft is driven by a
bearing fault.
20-tooth pinion. The first 6 harmonics of the pinion are
The envelope of the raw and resampled data was taken with
clearly visible at 620, 1240, 1860, 2480, 3100 and 3720 Hz.
a window from 9KHz to 11KHz, which captures the bearing
The TSA ran 188 revolutions, so that the noise floor of the
resonance (Bechhoefer, 2010), as seen in Figure 2. The
TSA is approximately 1/sqrt(188), or 0.073x that of the raw
envelope spectrum is seen in Figure 5.
or resampled spectrums (Bechhoefer, 2009).

3
 Annual Conference of Prognostics and Health Management Society 2013

envelope spectrum. This type of information gives a

maintainer additional diagnostics, which are missing in the
raw spectrum.

4. CONCLUSION
Condition monitoring of rotating machinery is complicated
by the fact that machines under analysis do not always turn
at a constant rate. While the time synchronous average can
be used to control variance in machine speed for shafts and
gears – there is not such standard practice or algorithm to
control variance in shaft speed for bearing or other non-
synchronous analysis.
In general, it is assumed that the effect of spectral smearing
due to variance in shaft speed is small. However, variation
in shaft speed is commonly observed in the field. This
Figure 5. Envelope spectrum of raw and resampled data problem is especially great for wind turbines, in which there
is variation due to: changing wind speed, a 3/revolution
Note that the cage (13 Hz), shaft (31 Hz) and inner race torque ripple due to tower shadow, and a 1/revolution effect
fault feature (292 Hz) are overlaid on the spectrum. Clearly, from wind shear. It is not surprising to see a 4% change in
the resampled envelope spectrum fault features have greater shaft speed in a 6 second acquisition.
energy. This is not a scale issue, as the noise floor for both
spectrums are the same. The increased energy is because In this paper, a resampling process is proposed in which raw
there is less spreading of energy into neighboring FFT bins. data is synchronized by a key phasor to a shaft under
Since it is known that the bearing has a cracked inner race, analysis. The resampling process changes the effective
Figure 5 gives a detail view of the spectrum. sample rate and normalizes the data by removing the effect
of changes in shaft speed. It allows both synchronous
(shaft/gear) analysis and non-synchronous (bearing, bearing
resonance) analysis.
This is demonstrated on a wind turbine high speed shaft
bearing with an inner race fault. It is shown that by
resampling, the frequency content of the envelop spectrum,
which is spread over a frequency of 292 to 303 Hz (14 FFT
bin), with raw envelope spectrum of 0.17 Gs. For the
resampled spectrum, the true energy value is closer to .32
Gs. It is shown that for this example, the raw energy
spectrum had an error of 47% when compared to the
resampled spectrum.
This is a significant improvement in performance. It is
likely that the variance in speed accounts for a large
variance in condition indicator values. It is anticipated that
Figure 6. Detail view of inner race fault: envelope spectrum by fielding this improved analysis algorithm, that:
• Bearing faults will be easier to identify,
This view highlights the improvement in resolution of the
resampled data over the raw data. Note that from Table 1, • that threshold setting will be simplified,
the fault frequency range for the inner race was 292 to 303
Hz, which is validated in Figure 6 in the raw envelope • that trend analysis will be improved and finally,
spectrum. Note that maximum value of the raw envelope • that this will facilitate an improved prognostics
spectrum was 0.17 Gs. As seen in the resampled spectrum, capability.
the true energy value is closer to .32% or a 47% error in the
original measurement. This smearing of measurement data, The resampling algorithm used linear interpolation, but
as seen in Figure 2, could well be caused by the large spline or cubic interpolation could be used.
variance in shaft speed.
Both inner race modulated by cage and inner race
modulated by shaft are also clearly present in the resampled

4
 Annual Conference of Prognostics and Health Management Society 2013

REFERENCES in electronic systems. Dr. Bechhoefer is a board member of

the Prognostics Health Management Society, and a member
McFadden, P., Smith, J., (1985), A Signal Processing
Technique for detecting local defects in a gear from a of the IEEE Reliability Society.
signal average of the vibration. Proc Instn Mech Engrs.
McFadden, P., (McFadden 1987) “A revised model for the Brandon Van Hecke received his B.S. in Industrial
extraction of periodic waveforms by time domain Engineering from the University of Illinois at Chicago in
averaging”, Mechanical Systems and Signal Processing 2010. He is a Ph.D. candidate in the Department of
1 (1) 1987, pages 83-95 Mechanical and Industrial Engineering. His research
Bechhoefer, E., Kingsley, M. (Bechhoefer 2009). “A interests include digital signal processing, machinery health
Review of Time Synchronous Average Algorithms”. monitoring and fault diagnostics based on the evaluation of
Annual Conference of the Prognostics and Health vibration and acoustic emission signals, and condition based
Management Society maintenance.
Christian, K., Mureithi, N, Lakis, A., Thomas, M.,
(Christian, 2007), “On the use of Time Synchronous
Averaging, Independent Component Analysis and
Support Vector Machines for Bearing Fault Diagnosis”, David He received his B.S. degree in metallurgical
First International Conference on Industrial Risk engineering from Shanghai University of Technology,
Engineering, Montreal, Dec 17-19 Montreal. China, MBA degree from The University of Northern Iowa,
Bechhoefer, E., Kingsley, M., Menon, P., (Bechhoefer, and Ph.D. degree in industrial engineering from The
2009), “Bearing Envelope Analysis Window Selection University of Iowa in 1994. Dr. He is a Professor and
Using Spectral Kurtosis Techniques”, IEEE PHM Director of the Intelligent Systems Modeling &
Conference, 2011. Development Laboratory in the Department of Mechanical
and Industrial Engineering at The University of Illinois-
BIOGRAPHIES Chicago. Dr. He’s research areas include: machinery health
monitoring, diagnosis and prognosis, complex systems
Eric Bechhoefer received his B.S. in Biology from the
failure analysis, quality and reliability engineering, and
University of Michigan, his M.S. in Operations Research
manufacturing systems design, modeling, scheduling and
from the Naval Postgraduate School, and a Ph.D. in General
planning.
Engineering from Kennedy Western University. His is a
former Naval Aviator who has worked extensively on
condition based maintenance, rotor track and balance,
vibration analysis of rotating machinery and fault detection