Annual Conference of the Prognostics and Health Management Society, 2009

Bearing Envelope Analysis Window Selection

Dr. Eric Bechhoefer1and Praneet Menon2
1,2
Goodrich – Sensors & Integrated Systems, Vergennes, VT, 05408, U.S.A

Eric.Bechhoefer@goodrich.com
Praneet.Menon@goodrich.com

ABSTRACT
Bearing envelope analysis (BEA) is a powerful technique in the detection of faults in bearings. However, the
improper selection of the envelope window frequency and window bandwidth can render the analysis ineffective. A
health and usage monitoring system’s (HUMS) ability to report bearing health can be degraded due to improper
window selection. In a recent teardown analysis (TDA) of a utility helicopter oil cooler fan housing, extensive
bearing damage was found; however, the damage was not detected by HUMS. Fortunately 38 raw vibration data
acquisitions were collected over an extended period of time, which allowed investigation into why HUMS did not
detect the fault and gave insight into envelope window selection.

In this paper, a description of the envelope analysis is given. Using the 38 acquisitions, a measure of effectiveness
for bearing envelop analysis was developed in order to determine the best envelope window for the detection of
bearing fault.

1. INTRODUCTION made, yet the bearing condition indicator values

where nominal. The damage seen in the teardown
Rotating equipment in general, and helicopters in analysis showed corrosion, and other spall/pitting
particular, are dependent on a transmission to condition the damage of the bearing ball elements, inner and outer
power for useful work. In the case of helicopters, the race. This type of damage occurs over an extended
power of a low toque, high speed input shaft is converted period of time. Fortunately, raw vibration data was
into a low speed, high torque main rotor shaft. Integral to collected over a period of time prior to the TDA.
the reliable operation of this transmission, are bearings. This would allow post processing of the bearing
Safety and readiness of the helicopter are improved if data and an investigation into why the system failed
monitoring techniques are developed, which can detect to detect the bearing damage.
when a degrading or faulty bearing. This is the essence of
HUMS bearing monitoring. Prior experience with BEA has shown that the
ability to detect a fault is based on the window
A number of bearing analysis techniques have been (band pass filter size) used (ref [3,4,5,6]). It is
developed. Because the vibration signals of a faulted hypothesized that the poor window selection was the
bearing are typically small compared to shaft order and cause of the missed detection. In this paper, we
gear mesh, detection of fault at the bearing rate show how the BEA functions, and a method of
frequencies using Fourier analysis is difficult. This is determining the optimal window for fault detection.
“stage 1” fault detection (ref[3]). Bearing faults detected
using these types of analysis are late stage and can be close
to catastrophic failure. Ultrasonic emission can detect
bearing inner and outer race roughness (a “stage 3” fault),
but the remaining useful life a bearing at this stage is
relatively long compared to the overall life of the bearing.
Bearing envelope analysis can typcially detect bearing
faults 10s if not 100s of hours prior to when it is
appropriate to do maintenance. It is for this reason that
many HUMS manufactures are using envelope analysis
techniques.

Given the known performance of the bearing envelope

analysis (BEA), there was some consternation when a
recent teardown analysis of the utility helicopter oil cooler
fan bearing housing showed extensive wear to a monitored
bearing (see figure 1 and 2. Ref[1]). The HUMS had not
indicated any bearing damage. The HUMS had been
installed for over a year; numerous acquisitions had been Figure 1. Outer Race Showing Damage

1
 Annual Conference of the Prognostics and Health Management Society, 2009

Because of the large difference in absolute values between ⎡1 1 ⎤ ⎡ CI1 ⎤

different window CIs, the 48 CIs generated by a given ⎢ ⎥ ⎢ ⎥
window where normalized between zero and one. See 1 2⎥ CI2 ⎥ ⎡intercept⎤
Figure 5 X=⎢ , Y=⎢ , b=⎢ ⎥ (7)
⎢.. .. ⎥ ⎢ .. ⎥ ⎣ slope ⎦
⎢ ⎥ ⎢ ⎥
⎣1 38⎦ ⎣CI38⎦
and the sum of square errors (SSE) is:

Prior to
Maintenance ˆ
Y = XT b
2
⎛ ˆ⎞
(8)
SSE = ∑ ⎜⎜ Y − Y⎟⎟
38
After Maintenance

⎝ ⎠
i=1

This MOE, slope/sse, quantifies how well a given

Figure 5. Comparison of Cage (CPF) CI using different
Envelope Spectrums
Figure 5 shows that if a different envelope window had
been used, the HUMS would have showed an increasing
trend in the CIs that would be used to trigger maintenance.
EW will facilitate detection of a fault. The MOE
was calculated from a EW low frequency of 13 KHz
to 25 KHz with a bandwidth of 1KHz to 7 KHz, in
0.5 KHz step size. This gives 325 experimental
ranging from an EW of 13 to 13.5 KHz up to an EW
of 25 to 32 KHz. As an example of the
appropriateness of this MOE, consider this example
Outer Race example (Figure 6). Here we plot the
highest MOE vs. the Lowest MOE.

4.2 Optimal Selection of Envelope Window (EW)

Now that it is shown that proper EW selection can result in

a CI that is sensitive at bearing fault, the question
becomes, what is the best window? In general, the window
will be different for different bearing sub elements, and
will be different across bearings. While for this study we
focus on the oil cooler bearing, we can make
generalization about window selection, and be specific
with regard to this bearing.

As noted, in comparing the CIs across windows it was

necessary to normalize the CIs. Even with normalized CI,
one would wish to have some metric or measure of
effectiveness (MOE) to describe, at a system level, the
ability of the CIs to be effective in triggering a
maintenance action when appropriate. Subjectively, a
good CI would trend smoothly as damage progressed. A
number of MOE functions come to mind. One that seemed
to quantify this the best was the slope of the line that
minimum mean square error divided by sum of square
errors of the line.
The minimum mean square error is calculated as:
b = (XT X) X T Y
−1
where
Figure 6. Outer Race MOE, Best vs. Poorest
4.3 Results
In an attempt to better visualize or map the MOE vs.
EW, contour plots of the 325 experiments where
made for ball, cage, inner and outer race and RMS
envelope energies where made (figures 7 through
11). NOTE: This research does not recommend the
use of envelope RMS as a indicator of bearing
health. While envelope RMS is sensitive to bearing
fault, it is also sensitive to a number of other
conditions, such as gear backlash, pump value
clatter, etc. Thus, relying on envelope RMS may
generate spurious maintenance actions.

4
 Annual Conference of the Prognostics and Health Management Society, 2009

Figure 7. Ball MOE Figure 10. Outer Race MOE

Figure 8. Cage MOE Figure 11. RMS MOE

As suggested, the “best” EW for each sub element is
different. Over all, what would be the best? If each
MOE is weighted equally, the overall MOE for ball,
cage, inner and outer rate is seen in figure 12. Table
1 gives the best 10 EW. Note the difference

Figure 9. Inner Race MOE

Figure 12. Overall MOE for Cage, Ball,

Inner and Outer Race

5
 Annual Conference of the Prognostics and Health Management Society, 2009

TABLE 1: TOP 10 ENVELOPE WINDOWS 5 CONCLUSION

Envelope analysis is a powerful tool for bearing
Window (KHz) MOE diagnostics, but as shown, improper EW selection
20 - 24 0.749 can render the algorithm almost useless. From a
19 - 23 0.745 certification perspective (for both the FAA in
applying for HUMS maintenance credits or the
22.5 – 29.5 0.744 Army aviation in applying for an Air Worthiness
24.5 – 28 0.734 Release) this suggests that either test stand or direct
24 – 27 0.733 evidence must be demonstrated to show the
effectiveness, at a system level, of detecting fault.
19.5 – 23.5 0.73 The FAA’s AC-29 MG 15 states as much, the Army
22.5 – 27.5 0.725 Engineering Depot’s ADS-79 does not.
20 – 25 0.725
24.5 – 27.5 0.724
22.5 – 28.5 0.722 ACKNOWLEDGMENT
Goodrich would like to acknowledge the support of
the Utility Helicopter program manager for access to
the data which allowed this analysis.
If an EW of 20-25 KHz had been used, one can calculate
the correlation of damage over time. Since the HI is a
weighted average of the CI, one can calculate the REFERENCES
correlation of the HI over time as well. As an example,
using the 38 data points prior to the replacement of the oil [1] McCain, B., “Tear Down Analysis (TDA) for
cooler, the correlation and variance of the CI/HI are the UH-60 Axial Fan,” CSTE-DTC-RT-
presented in table 2. E-CS, 4 Aug 2008.
[2] Kellar, J., Grabill, P., “Inserted Fault
TABLE 2: CORRELATION & VARIANCE OF Vibration Monitoring Tests for a CH-
CI/HI 47D Aft Swashplate Bearing,” AHS 61st
Annual Forum, 2005.
CI/HI R value Variance [3] Bechhoefer, E., He, D., “Bearing Damage
(correlation) Condition Indicator Correlation,” Center
HI 0.74 0.11 for Rotorcraft Innovation Project: 07-B-
Cage CI 0.84 0.11 6-59-S2.1
Ball CI 0.58 0.13 [4] Bechhoefer, E., “Data Analysis Trade Study
of the CH-47D Swash Plate Bearing,” E-
Inner Race CI 0.56 0.13
5823, Goodrich SIS, 2007.
Outer Race CI 0.70 0.12 [5] Bechhoefer, E., “Analysis of CH53E Tail
Rotor Hanger Bearings”, E-5581,
This shows that the HI, Cage CI and Outer Race CI are Goodrich SIS, 2007
highly correlated with time, and presumably damaged.
[6] Bechhoefer, E., “Analysis of CH53E Oil
What can be said about the detection of these faults had a Cooler Bearing Accelerometer Location
better EW been in use in HUMS? First, the acquisitions and Condition Indicator Development,”
within HUMS are 200K points of data vs. the 32K points SIS
used in this analysis. The relationship between the number [7] Ganeriawala, S., “Some Observation of the
of points and noise in a spectrum can be shown to be Detection of Rolling Element Bearing
proportional to 1/sqrt(n/2), where n is the number of data
points. Given 6.25 times as much data; one would expect Outer Race Faults,” SpectraQuest,
the system noise to be reduced by 2.5 times. If anything, www.spectraquest.com September, 2006
the trend data would have been significantly better then [8] Bechhoefer, E., Mayhew, E., “Mechanical
that shown in this analysis. Second, note the EW used was Diagnostics System Engineering in IMD-
13 to 18 KHz, which, as seen in this analysis, is almost the HUMS,” IEEE Aerospace Conference,
worst place to be for bearing analysis.
Big Sky, 2006.
In a number of studies, from test stand to on-aircraft, in [9] Bechhoefer, E., Berhnard, A., “A Generalized
practice this research has seen that EW works well at 20- Threshold Setting Process for HUMS”,
25 KHz. This analysis suggest that a better window is 20- IEEE Aerospace Conference, Big Sky,
24KHz. Given any extenuating circumstances or actual 2007.
data to show otherwise, this research recommends 20-24 [10] Bechhoefer, E., “A Method of Generalized
KHz as the default values for EW.

6
 Annual Conference of the Prognostics and Health Management Society, 2009

Prognostics of a Component Using Paris Law,”

AHS 64 Annual Forum, 2008.