You are on page 1of 9

C623/048/2004

Evaluation of damping ratio estimation techniques


for rotordynamic stability measurements
C H CLOUD, E H MASLEN, and L E BARRETI
Rotating Machinery and Controls (ROMAC) Laboratories, University of Virginia, Charlottesville,
Virginia, USA

ABSTRACT
With the recent addition of analytical stability level requirements into industriaJ
turbomachine.r y standards, the verification of stability level through shop or field testing wilJ
continue to grow in practice. Published cases of rotor stability measurements have mainly
incorporated only single degree of freedom (SDOF) damping ratio estimation techniques.
Simulation results indicate that popuJar SDOF methods can produce relatively poor damping
ratio estimates of the rotor's first forward whirling mode. A widely used time domain
technique, backward autoregression, is shown to accurately estimate damping ratios while
providing many other attractive benefits for measurement of rotor stability.

1. INTRODUCTION
Thirty years after Jorgen Lund's landmark paper on stability analysis (1), avoiding instability
problems is still a primary concern for turbomachinery manufacturers and end-users. Often
requiring time-intensive design solutions, self-excited instabilities can affect users in a variety
of ways, from costly failures to schedule delays to production losses. Shop testing has long
been utilized to help identify rotordyna.rnic problems before reaching the field. While such
testing can help verify unbalance .r esponse performance and predictions, even full-speed, load
and pressure testing {ASME PTC 10 Type l test) does not provide a clear measurement of a

machine's stability level or margin versus that predicted.


To measure stability level, an external excitation must be applied during operation and the
response measured. Using these measurements, the problem becomes one of how to estimate
the damping ratio of the primary mode of interest, typically a rotor's first forward whirling
mode. Two single degree of freedom (SDOF) estimates for a mode's critical damping ratio,~.
are the following:

C623/048/2004 C> !MechE 2004

541

Material protegido por derechos de auto1

modes' damping levels. are clearly divergent, their natural frequencies remain identical for
this system without gyroscopics. This stability sensitivity performance is similar to a rotor
mounted on tilting pad journal bearings with between-pad loading. Our problem is to measure
this stability performance using simulated test data in the time or frequency domain .

..
t

,
I

,'
---;---,
,

I ,'

~:

,,

,,

'

'

--- ---

A0.7S., I \.2! .. - -

____ ...

A0.1~ 81.~5

'
...

4eM

...

C...-..plo4S-NQ(Nfmnll

.....

ttM

...

...

C........,..piOdS-NO(NIYnm)

Figure I: Rigid rotor modal parameters versus Q

'

1'

--'

, " '

I
I

..

'

..,....

f'... ....

--

C11 11 _.... fS..._ Q.....

Hx,.x

J[

=--

=
- -- .....

ht . ....
,~

01ut ........

I ' "' -J
""-.
~

-4

HY_'I'

CUOJ~Mmt

.,...,.,.,,

...

-___
- M,..._.

,., - - Ill', ..._

-- --.....
.......... ,_

OJ'tU~

Figure 2: Symmetric rigid rotor with


directional forcing AF estimates

To begin, we examine frequency domain data where the amplification factor is the popular
damping ratio estimate. Shaking the rotor horizontally and vertically as in a sine sweep test,
four frequency response functions (FRF) are obtained, two for each measurement probe.
Figure 2 displays the FRFs at several levels of stability margin. For brevity, only the FRFs
measured by probe Xp due to horizontal forcing are shown. Since this system is symmetric,
the others are identical. Plotted also is the magnitude contributions of the forward and
backward modes. One clearly sees that even though a single peak is present in the
"measured" FRF, there are actually two modes contributing to the response. The forward
m.ode's contribution becomes increasingly dominant as the stability margin declines.
The damping ratio esti.mates using AF (4 for each level of Q) are shown on Figure 2's
stability curve. In general, good agreement is obtained, especially near the stability threshold,
where the SDOF assumption is more valid. However, the accuracy of the estimates is reduced
at high stability margin levels.
Many rotor systems are not symmetric, and Figure I displays tbe modal characteristics of one
such rigid rotor system with slight asymmetry (A-0.75, B=1.25). The challenges for this
slightly asymmetric system become cJearly apparent when observing the FRFs in Figure 3.
Once again, for brevity purposes, only the Hxpx frequency response functions are displayed
for each level of cross-coupling Q. Even though the modes are separated, only one peak is
still measured. This interaction between forward and backward modes causes large
inaccuracies in the AF estimates of damping. The stability curve shows the four AF damping
estimates for each set of level of Q. Once again, near the stability threshold, the AF estimates
are accurate because the forward mode dominates the response. However, the stability
measurements can be very inaccurate at higher stability margins.
A popular stability measurement technique is to apply a forward whirling, circular excitation
to the rotor in an effort to excite the forward mode (5). As shown in Figure 4, the circular
FRFs contradict the commonly held perception that forward circular excitation results in only

C623/048/2004

e IMech.E 2004

543

Material protegido per derechos de aulor

=::-

HJr.x =-=--=- ::;

Hllpll

II
~

Hx.c

.............

'12 ._ ..........

J
r

- '

,, ._._

Figure 3: Asymmetric rigid rotor with


directional rorc:ing AF estimates

HXpc - -- - -

'

I'

'

=== ,. .,.......

,
,

' - ,' -

J'

v r=--

t ...... ,.....,. -

Ot.l1t . . . .

. ' ........

c a

" $

/
,,

o I

'

e.c
Hx.a _ ==
___ ,'"'
_.....
_

,L

~._..._.,...............

=
...
-- - = = WsssmC

7 ......

,,

O JMJ.._

___ ,__
==

..

oo.l.llt - -

, =-.... . . \

\
! :\

c-

Figure 4: Asymmetric: rigid rotor with c:irc:ular


rorc:ina: AF estimates

forward mode response. For this slightly asymmetric condition, the backward mode does
participate and creates inaccuracies for the AF damping estimates (two for each level of Q).
Since Newkirk (6), the estimation of damping from response outputs' free-decay time signals
has been a popular option. This can be done using impulse, step, or resonance excitation. Log
decrement has been the damping estimator of choice. Here, we examine such a time-transient
measurement on the first rigid rotor case (A=B= l).
With steady-state vibration amplitude of roughly 25 J..lffi p-p and rotating at I 0,000 rpm, the
rotor is subjected to a horizontal impulse large enough to generate a transient amplitude of
four times the synchronous amplitude (- I 00 J..lffi p-p). Figure 5 displays the transient response
observed by the probes. Without filtering the synchronous response contribution, log
decrement measurements for the two probes result in poor damping estimates for the forward
mode (actual ~ F = 4.27% (l> = 0.2685)).

Ar

II\

r- y probe

V1

llifll

"

_A

'\

.J

-'

c; = 5.33% <o = o.33s)

--

t; = 4.75% (S = 0.299)
"

--

r- rv
- r-
J~ - , . _'" " "_IJ\_v_l~---~~-~~____o J~r-\-ll---"'-------c; = 2.8Wo (o "' 0.180)

--

1-----..-----.....------

c; =4.72" (S =0.297)

Figure 5: Rigid rotor impulse response waveforms (A B=l) and log decrement damping estimates

Figure 5 also shows the transient response with synchronous notch fi ltering applied. While
the damping measurement accuracy has greatly improved, there still exists some error. This
relatively small error is of concern because this is an ideal case without noise and with a
strong enough impulse excitation to excite the forward mode. Furthennore, our ideal system

544

C623/04812004 'J IMechE 2004

Material protegido par derechos de autor

is without other excitations outside the notch filter range. If steady-state excitations exist near
the frequencies of the concerned modes, they cannot be filtered. Therefore, any error
experienced in this ideal situation will likely be amplified in tests of real machines. The same
can be said for the AF errors shown earlier.
Even for the simplest rotor systems, then, the interaction between a rotor's forward modes
and its sister backward modes can be significant Such interactions can cause large
inaccuracies when SDOF damping estimates are employed. Excitation direction and filtering
may improve the accuracy of such SDOF techniques directionally, but the response remains
multi-modal in nature. Therefore, a multiple degree of freedom (MDOF) parameter
estimation technique is needed to accurately measure the stability of rotors. Such a technique
needs to be robust enough to handle close/repeated modes, which may or may not be well
excited by testing, and which will be surrounded by noise and other internal excitations.
Some time domain techniques provide such capabilities.

3. TIME DOMAIN PARAMETER ESTIMATION


In the structural dynamics field, frequency domain parameter estimation techniques, requiring
Fourier transfonn spectrum analysis of both inputs and outputs, continue to be used with
wide success. All frequency domain techniques, however, have performance limitations
regarding spectral resolution and leakage. Spectral resolution defines the ability to distinguish
the responses of two signals, which is clearly an issue for a rotor's close/repeated modes.
Leakage results in higher-than-actual damping estimates (7,8).
These limitations become very problematic when only finite, limited data records are
available (one impact test instead of l 0). Such limited data sets are of primary concern jn
fields like radar, sonar, speech processing and seismology. These fields led the development
of modem parameter estimation technjques based in the time domain. Sucb techniques are
often called parametric or model-based or direct methods and provide improved accuracy and
frequency resolution (9) with solid theoretical and methodological foundations ( t 0).
Ironically, the work of another group of rotordynamicists, those concerned with the dynamics
of helicopter rotors, led this investigation toward time domain methods. Helicopter rotor
stability measurements face challenges very similar to those faced by turbomachinery rotors:
both systems feature very close modes which are lightly damped, som etimes difficult to
excite, and in the presence of substantial ampUtudes of rotational speed harmonics (t 1).
Likewise, stability testing time is a critical factor and can be extremely costly. For helicopter
rotor stability measurements, damping estimation is done using measured output response
only. While measuring the input forcing has been relatively common for turbomachinery
stability testing (especially with the recent popularity of magnetic actuators), eliminating the
need to do so is certainly attractive, especially for field testing.
Looking at techniques to estimate the damping of helicopter rotor blade modes, Tasker and
found that a derivative ofthe Sparse Time Domain technique was very effective
Chopra.(ll)

at estimating the damping of close modes. They performed extensive evaluations using noisy
data as well as close modes at different damping levels. Developed to achieve more accurate
damping estimates than the Ibrahim Time Domain technique, the Sparse Time Domain
technique is simply an autoregression (AR) method. Autoregression is one of several direct
time domain methods (autoregressive-moving average, ARMA, is another) that are applicable

C623/048/2004

!MechE 2004

545

Material protegido por derechos de autor

to a wide variety of systems. A simple autoregressive model of a noisy, discrete sampled


output process takes the form:
x(n) = -a1x(n - 1) - a2 x(n - 2) -a px(n- p) + w(n) = - 'f,a,x(n- i) + w(n)

{l)

11

where w(n) is assumed to be a zero mean, white Gaussian noise sequence and p is the order
of the model's linear regressors, aj. The term "autoregression" implies that x(n) is predicted
using its own previous values (12). Here we will focus on backward AR techniques. Using
subsequent values instead of prior values (equation 1). a backward AR model takes the form:
x(n) =- b,x(n + 1) - b2 x(n + 2) - bPx(n + p) + w(n) = - fAx(n + i) + w(n)

(2)

11

where b1 are backward regression parameters.


A discretely sampled response consisting of M exponentially damped modes can be
represented by
2M

y(n) =

L f/1 1exp(A.1 11~t) + w(n)


11

where ~t is the sample time interval, 'Vi are the complex modal amplitudes, and AJ are the
modes' eigenvalues (poles) of the form ~ = l;iCOi + j0) 1 (l-~, 2) 112. Using the entire data
sequence (N samples), the following backward AR model can be written:
y (t)

y(2)

y (p)

b1

y(O)

y(2)

y(3)

y(p + I)
.
..

b2

y (l)
.
.

..

.. = -

y(N- p)y(N- p + 1) y(N -1) bP

(3a)

y(N - p - l)

(3b)

Using this linear prediction equation, the unknown vector b can be calculated, yielding the
individual backward regression coefficients. These coefficients form the characteristic
polynomial of the AR model
A(z) = 1+ h.z- 1 + b2.z -2 + + bPzP
Solving for the roots of this polynomial, one can obtain the estimated system poles ~ , which
are participating in the measured response. The modal damping estimates are calculated using
the real and imaginary parts of each pole.
.

Measurement noise is the primary factor affecting the accuracy of time domain estimation
methods. Frequency domain techniques rely on the FRF averaging process to eliminate noise
influence.s. Time domain techniques handle noise by increasing the model 's order, p, beyond
what is needed. For example, a noise-free sine wave only requires an AR model order of2.
However, in the presence of noise, the order may need to be increased to 20 or 30, depending
on the noise level and number of samples. This 'overspecification' or 'overdetennination'
results in some poles which actually model the noise participation. Sometimes called
"computational" or "extraneous" poles, these noise-related poles must be distinguished from
the true physical, or signal-related, poles in the response. Backward AR accomplishes this by
forcing all the roots of A.(z) associated with noise to fall inside the z-domain unit circle. All
of the roots associated with physical system modes are located outside of the unit circle. This

546

C623/048/2004

IMechE 2004

MatenaJ protegido por derechos de au!or

is the primary reason for using the backward AR data format instead of the forward AR
version expressed in equation 1 (13).
Noise also affects the solution accuracy of the regression parameter vector, b. Thus, principal
component analysis (9), using truncated singular value decomposition (SVD) (13, 14), is
applied to improve the accuracy. Such a principal component analysis can improve tbe
signal-to-noise ratio (SNR) by decomposing the data into a signal subspace and a noise
subspace. The key becomes determining which principal components (associated with the
largest singular values) should be kept so that only the noise contributions are eliminated.
The rotordynamicist needs to be aware of several issues regarding using such a principal
component solution approach. In some cases, where the SNR is low or a mode is poorly
excited, the SVD is unable to clearly separate a signal and a noise subspace. A large number
of principal components or the full decomposition may be necessary to accurately identify the
modes (15). Also, the truncation process was found to add some bias to damping estimates.
Bias compensation improves the accuracy of the damping estimates (11 ) and bas been
implemented in this investigation.
The single output version of the backward AR method may be extended to include the
measurements of multiple outputs simultaneously (16). For this technique, we define the
following vector of C output channels:
y(n) =

{y1 (n)

y 2 (n)

Yc (n)V

Using N time samples, the multiple output linear prediction equation becomes
ji(l) ji(2) ji(N- p)

]]
[ Lo.1 Lo2] ... r"
lP p

ji(2) ji(3)

= [y(O)

y(l) y (N - p - I)]

(4a)

ji(p) ji(N - 1)

(4b)

The unknown AR parameter matrix ~ is solved for using principal component analysis once
again. The remainder of the technique follows that of the single output version.
For damping estimation, the multiple output backward AR (MOBAR) method has been
shown to be more accurate, with less bias and deviation due to noise, than the single output
version. Possessing a data set that is richer in modal infonnation, the necessary time sample
block needed for analysis is also reduced ( 17) resulting in faster computation time and less
supervision from the analyst. MOBAR also provides the flexibility of utilizing different
response signals (acceleration, velocity and displacement) at the same time (16). These
attractive features will be examined for rotor stability measurement applications using
simulated test data due to an impulse.
S. THREE MASS ROTOR EXAMPLE

With six modes contributing to the response, the three mass rotor mounted on identical
flexible journal bearings represents a simplified turbomachine (Figure 6). Simulated stability
measurements were conducted for three different bearing designs with increasing
levels of stiffuess (A) and damping (B) asymmetry (constant Kyy = 10,507 N/mm and Cyy =

C6231048/2004 C !MechE 2004

547

MatenaJ potegido por derechos de au tor

4.2 N-sec/rnm). Cross-coupled stiffness Q is only applied at the midspan, mass 2. Generating
roughly 25 .,un p-p at the bearing probes, equal unbalances at 0, 90 and 180 are
respectively placed on the masses. rotating at 10,000 rpm. The impulse is applied at the mass
1 journal. This nonideal location mimics the external excitation restrictions on an actual
machine. The magnitude of the impulse is varied to investigate how the level of excitation
influences the estimates. A transient response strength ratio is defined which equates to the
maximum transient response amplitude over the steady-state amplitude (25 .,un p-p). A ratio
of 4 indicates that the impulse caused the response (at any of the probes) to be four times the
steady-state vibration amplitude due to unbalance. To examine the MOBAR's performance
under noisy conditions, white noise was added to the four simulated bearing probe responses.
The noise level was set at a percentage of the steady-state RMS vibration amplitude.
SHAFT CHAAACTERISTICS:
O=a&.9mm
L=1219.2mm
E"' 207 GPA
>

W.SSES;
M1 : M3 " 11.34 kg

M2 =45.35kg
Y,

,,

' Three mass rotor oo journal bearin&s


Figure 6:
For the three bearing design cases, Figures 7 through 9 present the actual stability
characteristics and the MOBAR damping estimates from the simulated impulse testing at
three stability levels. The highly asymmetric bearing (Figure 9) results .in a much greater
stability threshold, with stability characteristics similar to that of a load-on-pad .tilting pad
bearing. Results indicate bow noise and low response strength ratio can degrade the overnll
accuracy of the estimates. The symmetric bearing design at 5% noise presented the greatest
inaccuracy in. the MOBAR estimates, except for strong transient response levels and low
stability margin conditions. An advanced AR order selection technique, in combination with
different SVD truncation, would most likely improve their accuracy.

In some cases, the MOBAR technique was unable to identify the first forward mode (Figure
8, 5% noise, X impulse) and, therefore, no data points are presented. This occurred for the
asymmetric bearing designs at high noise levels. However, changing the impulse direction to
vertical produced very accurate results because the forward mode is more vertically oriented.

No Noise

'
-

'

~-

____ .,.

'\ -------

'\.

"~

1% Noise


--------------'1)

-...-- ----------o ---------------<>

_..,..__. _________
0

5% Noise

..,.

'

'

I
-

~ -

-....

---------

----6..-~---- - - A

__. __... _________

..

'
I

--- --- ---------~

_____..., _________ ,

Figure 7: Three mass rotor case I (A B=l) with multiple out.put ba.c kward AR damping estimates

548

C623/04812004 C !MechE 2004

Matonal protegido por dorechos de autor

-e

!'

'

t
a

1!

---

----- _._..
\-

!'..

4-- -~~ -------- A ----r---------.


..
t

..-'

S~ Nolooe

-...--- ----------.0 _...__&---------'b ----.----------9

11/,. No&..

\.

'

_ ;! _ _.,_ ___ _ __ _ _ _ ..

---t.---------

---...---------

Figure 8: Tbree mass rotor case l (A--6.75, B=J.l 5) witb multiple output backward AR damping
estimates (5% noise bas additional estimates witb Y 1mpulse direction = solid symbols)

'

- -- ~

-1-.
\

- .

No Noise

1% Noise

6% Noise

0

---------------

Y lmpuiM

- ~---- - - ------------------'
(
)

..
---- - -- - - ----~
-4--- --------- , --------------

\
..

..

-~-

_,

.....

-- -- - - --- ~

'

____
,_, _________

.,

-~ --~-------- -~

Figure 9: Three mass rotor case 3 (A=0.25, B-'3) with multiple output backward AR damping
esti mates (5% noise has only Y impulse dir ection estimates = solid symbols)

It must be emphasized that no FFT analysis or data averaging was required to obtain these
fairly accurate damping estimates. Typically, only 0.15 second of response time (N=300) was
required to analyze these example systems. Furthermore, no notch fi ltering of rotational
frequency was necessary. The backward AR technique identifies rotational speed response
(and any steady-state response) as simply an undamped mode. Therefore, for this rotor
example, seven complex conjugate pole pairs are identified; six rotor modes and one for the
rotational speed undamped mode.
6. CONCLUSIONS

While popul.ar, the accuracy of SDOF damping estimates is severely limited when applied to
rotor systems where forward and backward modes closely exist. In most cases, this makes
them unsuitable. Multiple output backward autoregression, a time domain technique, has
been shown to handle such close/repeated rotor modes with relative ease. While investigation
into the damping estimation performance of other techniques continues, the MOBAR
technique appears very promising for rotor stability measurements. Simulations have shown
that it can provide accurate damping ratio estimation even in the presence of noise. Requiring
only short data records, MOBAR has a variety of benefits (free-decay output only response
needed; no measure of input required; no filtering of steady-state vibrations required). Such
capabilities, and the widespread use of such time domain techniques in other fields, make the
MOBAR technique worthy of closer study.

C623/048/2004

IMechE 2004

549

Material protegido por derachos de autor

7. REFERENCES
1. Lund, J. W. (1974) "Stability and Damped Critical Speeds of a Flexible Rotor in FluidFilm Bearings," ASME Journal of Engineering/or Industry, Vol. 96, No.2, pp. 509-517.
2. American Petroleum Institute (2002) API 617: Axial and Centrifugal Compressors and
Turboexpanders for Petroleum, Chemical and Gas !ndust1y Services, 71h Edition,
Washington, D.C.
3. Nordmann, R. ( 1982) "Modal Parameter Identification and Sensitivity Analysis in Rotating
Machinery," JFToMM Conference on Rotordynamic Problems in Power Plams, Rome,
Italy, pp. 95- 102.
4. Nordmann, R. ( 1984) "Identification of Modal Parameters of an Elastic Rotor with Oil
Film Bearings," ASME Journal of Vibrations, Acoustics. Stress and Refiability in Design,
Vol. 106, pp. 107-112.
5. Bently, D. E. and Muszynska, A. ( 1982) "A Stability Evaluation of Rotor/Bearing Systern
by Perturbation Tests," NASA CP-2250, pp. 307-322.
6. Newkirk, B. L. ( 1924) ''Shaft Whipping", General Electric Re view, Vol. 27, No. 3, pp.
169- 178.
7. Kay, S.M. and Marple, S. L. (1981) "Spectrum Analysis - A Modem Perspective,"
Proceedings ofthe iEEE, Vol. 69, pp. 1380-1419.
8. Maia, N. M. M. and Si lva, J. M. M. (1997) Theoretical and Experimental Modal Analysis,
Research Studies Press LTD, Hertfordshire, UK.
9. Zhong, P. ( 1997) Rotor Bearing System Identification Using Time Domain Methods, Ph.D.
Dissertation, School of Engineering & Applied Science, University of Virginia.
10. Juang, J . N. (1994) Applied System identification, Prentice HaU, NJ.
11. Tasker, F. A. and Chopra, 1. (1990) "Assessment ofTransient Analysis Techniques for
Rotor Stability Testing," Journal ofthe American Helicopter Society, Vol. 35, No. l , pp.

39-50.
12. Ljung, L. ( 1999) System Identification: Theory for the User, Prentice Hall, N J .

13. Kumaresan, R. and Tufts, D. W. ( 1982) "Estimating the Parameters of Exponentially


Damped Sinusoids and Pole-Zero Modeling in Noise," IEEE Tra11sactions on Acoustics.
Speech. and Signal Processing, Vol. ASSP-30, No. 6, pp. 833-840.
14. Cloud, C. H., Foiles, W. C., Li, G., Maslen, E. H. and Banett, L. E. (2002) "Practical
Applications of Singular Value Decomposition in Rotordynamics," JFToMM 6'h
international Conference on Rotor Dy namics, Sydney, Australia, pp. 429-438.
15. Hollkamp, J. J. and Batill, S.M. (1992) "Structural Identification Using Order
Overspecified Time-Series Models," ASME Journal of Dynamic Systems. Measurement,
and Comrol, Vol. 114, pp. 27-33.
16. Hung, C. F. and Ko, W. J. (2002) "Identification ofModal Parameters From Measured
Output Data Using Vector Backward Autoregressive Model," Journal ofSmmd and
Vibration , Vol. 256, No. 2, pp. 249-270.
17. Tasker, F. A. and Chopra, I. ( 1990) "Multi-output Implementation of a Modified Sparse
Time Domain Technique for Rotor Stability Testing," AIAAIASMEIASCEIAHS
Strucwres, Structural Dynamics and Marerials Conference, pp. 1871 - 1884.

C623/048/2004 C !MechE 2004


550

Material protegido por derechos de au tor

You might also like