You are on page 1of 6

A Generalized Process for Optimal Threshold Setting in

HUMS
Eric Bechhoefer
Goodrich Fuels and Utility Systems
Vergennes, VT 05491
802-877-4875
Eric.Bechhoefer@Goodrich.com

Andreas P.F. Bernhard


Sikorsky Aircraft Corporation
Stratford CT 06615
203-386-4323
abernhard@sikorsky.com

Abstract— Monitoring the health of a helicopter drive train decision rule for detecting a component which is no longer
enhances flight safety and reduces operating costs. Health nominal.
and Usage Management Systems (HUMS) monitor the drive
train by using accelerometers to measure component The normalized distance distribution is a function of the
vibration. Algorithms process the time domain vibration component CI sample statistic. Procedures are developed to
data into various condition indicators (CI), which are used calculate the unbiased statistic: covariance for Rayleigh
to determine component health via thresholding. 1 2 based CIs and mean value/covariance for Gaussian based
CIs. In the cases where the population of components is not
For the rotating machinery, a standard set of CI are shaft nominal (e.g. mass imbalances which violate the Rayleigh
order one, two and three (i.e. 1, 2 or 3 times the shaft RPM). assumption) tools are presented to control this. For gear,
Shaft order one (SO1) is indicative of an unbalance, where normalizing transforms can be used to ensure the CIs are
as higher shaft order can be used to detect a bent shaft or more Gaussian. Example data from utility helicopters are
misalignment condition. In the case of bearings, CIs are given.
envelope spectrum or cepstrum analysis of the ball, cage,
inner race and outer race frequencies. There are a number TABLE OF CONTENTS
of standard CI used for gear analysis, such as line
1. INTRODUCTION ..................................................... 1
elimination and resynthesis, side band modulation, gear
2. THE NAKAGAMI DISTRIBUTION ........................... 3
misalignment, etc.
3. HI ALGORITHM FOR RAYLEIGH BASED CIS ....... 3
4. HI ALGORITHM FOR GAUSSIAN BASED CIS ........ 5
In general, some method is used to set thresholds for these
5. SYSTEM ENGINEERING CONSIDERATIONS ........... 6
CIs: when the threshold is exceeded, maintenance is
5. CONCLUSION ......................................................... 7
recommended. The HUMS system must balance the risk of
6. NOTES ON RAYLEIGH DISTRIBUTION .................. 7
setting the threshold too high such that a component may
REFERENCES............................................................. 8
fail in flight versus the risk of setting the threshold too low,
which results in additional maintenance cost. This paper BIOGRAPHY .............................................................. 8
covers a generalized process of optimally setting threshold
for CI and fusing the information into an Health Indicator. 1. INTRODUCTION

It can be shown that the distributions of CI for shaft Health and Usage Monitoring Systems (HUMS) have been
magnitude and bearing envelop energy are Rayleigh proven to be indispensable in improving aircraft readiness
distribution. The normalized distance functions for these and reduced maintenance cost. There is a consensus in the
CIs are a Nakagami distribution with µ (shape parameter) of helicopter community as what a number of functions do,
n (number of CI) and Ω (scale parameter) of 2 x 1/(2-π/2) x such as: Usage, Rotor Track and Balance, Exceedance
µ. For gear CIs, which are considered as Gaussian, the Monitoring and Regime Recognition. But in the case of
normalized distance function is again Nakagami, but with a mechanical diagnostics (MD), there are at least two
µ of n/2 and Ω of n. Given the theoretical µ and Ω, a paradigms for the evaluation of component condition. The
threshold for any set of CI can be generated resulting in first paradigm will be identified as CI based thresholding,
system probability of false alarm (PFA). This is an optimal and the second paradigm will be called Health Index (HI)
based thresholding.
1
In most HUMS systems, a vibration signature is processed
1
1-4244-0525-4/07/$20.00 V2007 IEEE.
2 into a number set of condition indicators (CI). For many
IEEEAC paper #1142, Version 1, Updated October 16, 2006
1
acquisition 75 progressed to an HI of 1. The shaft coupling, Here β was set to 1/(2-π/2), which corresponds to a standard
on removing, was found to be heavily fretted, cracked, and deviation of 1. The eccentricity varied from 0 (Rayleigh)
with evidence of heating. to 10. Note that the mean value of the Rice asymptotically
approaches the eccentricity, while the standard deviation
asymptotically approaches 1/(2-π/2).

This is again a significant result. Because the sample is not


dependent on the mean, it is relatively insensitive to
changes in eccentricity or model violations due to damage.
This model is dependent almost exclusively on the
assumption that the mass and stiffness of the component is
similar between aircraft (e.g β is constant). Due to the close
tolerance of aviation manufacturing, this is a good
assumption. Even in the presence of slightly damaged
components, the sample covariance is representative of
nominal components and the resulting HI algorithm will be
optimally suited to detect anomalous components.

4. HI ALGORITHM FOR GAUSSIAN BASED CIS

Figure 2 HI of AC Population vs. AC545 Many CIs, such as those used to infer the condition of gear
Effect of Model Violation on Rayleigh Assumption components, have Gaussian or near Gaussian distribution.
Under these conditions, the HI normalized distance can be
Through the development of this HI function, it was similarly developed. Taking advantage of Eq 1, a distance
assumed that the sample covariance would be sampled from function for the Gaussian case is given as:
nominal components. This may not be the case in that the
HUMs equipped aircraft are retrofitted on aircraft with an
R 2 = (CI − M) Σ−1 (CI − M)
T
(12)
unknown service history. Essentially, it is assumed that
components currently in the fleet are serviceable and the
CIs are generally representative of nominal components. Note that R2 is by definition a Chi-Square distribution with
n degrees of freedom, where n represents the number of CI
It was stated previously that damaged shaft and bearing used in for diagnostics. In this example case, n is six. From
components generate CI’s with a Rice distribution (e.g. no Theorem 1, the E[R2] = 6 and from Eq 4, Ω = 6. Using Eq
assumption of zero mean Gaussian). One can see the effect 5, it is seen that µ = Ω/2 = 3.
changing the eccentricity of a Rice distribution when β is
constant (Figure 3). The HI Algorithm
It is a simple matter to scale the distance function Eq (12)
by the critical threshold value derived from the inverse
Nakagami CDF (defined here as v)

HI = (CI − M )T Σ −1 (CI − M ) * 0.7 v (13)

In the case of the Gaussian distribution with six CI, one


finds:

• Ω = 6 , µ = 3, v = 6.183, and
• mean = 2.35, variance = 0.478

In this example, the mean nominal shaft would have a HI of


0.26, and the standard deviation would be (0.7/6.18) x
0.4780.5 = 0.078.

Example: Input Pinion Gear


Figure 3 Effect of Eccentricity on Mean and Standard
Deviation of a Rice PDF The most serious failure modes for gears are root bending
fatigue failures. Depending on the gear design, this type of

5
crack can propagate through the gear tooth causing tooth with heavily tailed distributions such as the Exponential, it
loss, or through the web causing catastrophic gear failures. takes relatively small number of additions to become
This type of gear tooth failure can be promulgated by Gaussian. It is possible to test the assumption of the HI
implanting an electronic discharge machine (EDM) notch in normality by using a Beta distribution (ref 10). If it is found
the gear tooth root. This creates a localized stress that the normality of the HI assumption fails, a normalizing
concentrator at the tooth root which can initiate a crack. transformation can by applied to the CI to make it more
Gaussian and thus satisfy the assumption.
This type of fault was tested by the Naval Air Warfare
Center Aircraft Division on a SH-60 intermediated gearbox
(IGB) input pinion on the 5. SYSTEM ENGINEERING CONSIDERATIONS
NAVAIRWARCENACDIVTRENTON test cell (ref 12).
The input pinion had two EDM notches (.25” length x .006” It must be noted that the HI algorithm presented in Eq 11
width x .04” depth) implanted along the length of the root. and 13 will never generate a HI of zero. This is because the
Data was collected by the Goodrich HUMS system, probably of R2 being zero is zero. However, the system
consisting of 37 acquisitions taken at 100% tail rotor torque. engineer may want to display new components with a HI of
Acquisitions where taken at approximately 15 minute zero. This is simply done using the Nakagami CDF.
intervals (figure 4). Consider the policy where the best 5% of the fleet will be
defined as having a health of zero. By using the appropriate
Ω, and µ, the lower bound threshold (e.g. HI offset o) is
generated. In the shaft example, one would have:
• Ω = 13.98 , µ = 3, o = 1.95, v = 9.44, and
• mean = 1.63, variance = 1.113

Eq 11 is now modified to:

HI = ( CI ΣT −1
)
CI − o * 0.7 (v − o ) (14)

The mean nominal shaft would have a HI of 0.2, and the


standard deviation would be (0.9/(9.44-1.95) x 1.1130.5 =
0.12, while the best 5% of components will have a health of
zero (e.g. if HI < 0, HI = 0) This affords the engineer a
number of scaling options in which to present information
to the operator.
Figure 4 Comparison of AC Population vs. Input Pinion The Importance of Treating Component CIs as a System
HI
The HI presented in Eq 11 and 13 make use of the inverse
Training data from 30 UH-60L aircraft was used to derive of the component CI covariance, sometimes called Fisher
nominal CI mean and covariance. Eq. 13 was then used to Information Matrix (ref 11). This is the information
generate the HI of the aircraft population, and the HI of the obtained from the observed data and represents the
IGB fault. Figure 4 shows the relationship between the importance (from an information theory point of view) of
aircraft population (mean approximately .26) and the fault the relationship between the measured CI’s. Consider the
HI. The IGB fault HI is nominal until approximately shaft’s a priori covariance and its associated correlation
acquisition 15, when the HI trends rapidly towards failure. (correlation is the covariance matrix normalized by
The HI saturates at 1, but if full scale where allowed, the HI variance) for a generator shaft:
is greater than 7 at the end of the test run. Upon post test
inspection, the pinion gear had a chipped tooth and had
⎡.142 .012 .014⎤ ⎡ 1 .92 .95⎤
cracked clear through the web.
⎢ ⎥ ⎢ ⎥
Σ = ⎢.012 .006 .002⎥, δ = ⎢.92 1 .75⎥ (15)
Effect of Model Violation on Gaussian Assumption
⎢⎣.014 .002 .006⎥⎦ ⎢⎣.95 .75 1 ⎥⎦
Through the development of this HI function, it was
assumed that the sample covariance would be sampled from Note that this shaft has a large, positive correlation between
nominal components with Gaussian distribution. The issue SO1 and SO2, SO3. This means that the when SO1 moved
of greatest concern is that the gear CIs may not have high, it is likely that SO2, and SO3 will similarly move.
Gaussian distributions. Due to the Central Limit Theorem From an information perspective, this suggest that a
(ref 5) it can be shown that sum of any normalized measured set of CI contains more information when SO1
distribution approaches the Gaussian distribution. Even
6
moves high and SO2 or SO3 does not move (shaft history of component failures, the HI is sensitive to
imbalance) or vices versa (SO2, SO3 moves, SO1 does not, individual CI and changes in groups of CI. The
mechanical looseness). As an example, consider the output contribution of the CI to the HI is based on the information
of the following operations where the normalized (e.g. inverse covariance) present in that CI. The HI
information matrix is operated on by a set of notional CIs: algorithm presents a number of attractive properties, such as
simplicity and low configuration overhead. Finally, in
⎡ 1 −.34 −.62⎤ testing against known faults, the algorithm performs well
⎢ ⎥ and indicates “Alarm” at a point when the component needs
δ −1 = ⎢−.34 1 −.52⎥, CIT δ −1CI (16) maintenance.
⎢⎣−.62 −.52 1 ⎥⎦
As with any algorithm, more examples and service history
will highlight any shortcoming in its performance. Finally,
Here are example cases were the vector of CI are varied
it must be remembered that the HI is just part of an overall
between Low (0.1) and High (1.0) for all possible
decision making strategy which will be driven by OEM and
combinations:
operator requirements.
CI Inner Product

[0.1 0.1 0.1] 0.0004 6. NOTES ON RAYLEIGH DISTRIBUTION

[1 0.1 0.1] 0.817 The Rayleigh distribution is frequently used to model the
statistics of signals transmitted through radio channel such
[0.1 1 0.1] 0.835 as cellular radio. The distribution is closely related to the
central chi-square distribution. The Rayleigh distribution is
[0.1 0.1 1] 0.785 defined as:

[1 1 .01] 1.102 R = X12 + X 22 , X1,2 ∈ N (0, β ) (a)

[0.1 1 1] 0.778 Where X1,2 are from the a Gaussian distribution with zero
mean and standard deviation of β (we define β as distinct
[1 0.1 1] 0.598 from the standard deviation of the distribution R, which is
σ). The PDF of R is:
[1 1 1] 0.040
f (R) = R β 2 exp(−R 2 2β 2 ) (b)

The inner product is relatively low when the CI move from which the expected value E[R]
together, while the inner product is large when the CI values
are different. The behavior suggests that when one or two E [R] = β π 2 (c)
of the individual CI move relative to the other, this is an
anomalous condition. Alternatively, if all three shaft CI and standard deviation
move high at once, this is probably a random event and one
does not want to generate an alert (as reflected in the a σ = 2 − π 2β (d)
priori data collected from 28 aircraft, 3000+ acquisitions).
This relationship between CIs is a discriminate that cannot
can be calculated.
be modeled or observed when looking at individual CIs.
This ability to access the information presents a system
level control of false alarm which is difficult, if not In regards to the normalized Z, where one assumes a
impossible to control, when performing diagnostics on Rayleigh distribution is Gaussian, one has:
individual CIs.
R − E [R] R−β π 2
Z= σ = (e)
β 2−π 2
5. CONCLUSION
The new distribution is by definition zero mean with a
The methodology presented for an HI based on identifying standard deviation of 1: it is an offset and scaled Rayleigh.
when a component is anomalous is powerful in that it The β for a Rayleigh with standard deviation of one is (2-
addresses a number of issues. First, with the limited service π/2)-1/2 or 1.5264, and the mean is (π/2)1/2 or 1.2533. The
7
offset is then simply (π/2)1/2 x (2-π/2)-1/2 or 1.9131. To BIOGRAPHY
evaluate the PFA at three standard deviation requires
evaluated the Rayleigh at 3 + the offset (4.9131) with a β of Dr. Bechhoefer is a retired Naval aviator with a M.S. in
1.5264. Rayleigh cumulative distribution function (CDF) is: Operation Research and a Ph.D. in General Engineering.
Dr Bechhoefer has focus on Statistics and
F(R) = 1− exp(−R 2 2β 2 ) (f) Optimization and Signal Processing. Dr. Bechhoefer has
worked at Goodrich Aerospace since 2000 on project
From which the PFA is calculated as: related to HUMS, wire fault detection and wireless sensor
exp(4.91312/(2x1.52642)) = 0.0056 vs. 0.0013 if the systems. Dr. Bechhoefer holds over 8 patents and has
distribution where Gaussian. published over 20 HUMS related papers.

Dr. Bernhard is a Senior Dynamicist at Sikorsky and is the


lead dynamics engineer for prognostics and health
REFERENCES
management and for active rotor control. Dr. Bernhard has a
Ph.D. from the University of Maryland (2000) focusing on
[1] Cyril M. Harris, Allan G. Piersol, Harris’ Shock and smart rotor technology.
Vibration Handbook 2002, McGraw-Hill, New York.
Page 16.1

[2] J. Derek Smith, Gear Noise and Vibration 1999, Marcel


Dekker, Inc. New York

[3] Eric Bechhoefer, Eric Mayhew, “Mechanical Diagnostics


System Engineering in IMD-HUMS” IEEE Aerospace
Conference, Big Sky, 2006.

[4] Eric Bechhoefer, Andreas Bernhard, “Use of Non-


Gaussian Distribution for Analysis of Shaft Components”
IEEE Aerospace Conference, Big Sky, 2006.

[5] Bechhoefer, E., Bernhard A., “HUMS Optimal Weighting


of Condition Indicators to Determine the Health of a
Component”, American Helicopter Society #60,
Baltimore, USA, 2004.

[6] Wackerly, D., Mendenhall, W., Scheaffer, R.,


Mathematical Statistics with Applications, Buxbury Press,
Belmont, 1996

[7] Bechhoefer, E et al. “Use of Hidden Semi-Markov Models


in he Prognostics of Shaft Failure”, American Helicopter
Society Forum 63, Phoenix, 2005

[8] Bechhoefer, E., “Function of Distribution” Patent


Application 2005.

[9]Proakis, John, G., Digital Communications, McGraw-


Hill, Boston MA, 1995, page 45-46

[10]Fukunaga, K., Introduction to Statistical Pattern


Recognition Academic Press, London, 1990, page 75.

[11]Van Tress, H., Detection, Estimation and Modulation


Theory, John Wiley & Sons, New York, 1968, page 80-
84.
[12] NAVAIRWARCENACDIVTRETON-PPE-271 of 3
Nov 95

8
9

You might also like