Content uploaded by Antoine Tahan
Author content
All content in this area was uploaded by Antoine Tahan on Aug 21, 2014
Content may be subject to copyright.
278
International Journal of Fluid Machinery and Systems
Vol. 2, No. 4, October-December 2009
Original Paper (Invited)
Uncertainty in Operational Modal Analysis of
Hydraulic Turbine Components
Martin Gagnon1, S.-Antoine Tahan1 and André Coutu2
1Department of Mechanical Engineering, École de technologie supérieure
1100 Notre-Dame West, Montréal, QC, H3C 1K3, Canada
2Andritz-Hydro Ltd
6100 Trans Canada highway, Pointe-Claire QC, H9R 1B9, Canada
Abstract
Operational modal analysis (OMA) allows modal parameters, such as natural frequencies and damping, to be
estimated solely from data collected during operation. However, a main shortcoming of these methods resides in the
evaluation of the accuracy of the results. This paper will explore the uncertainty and possible variations in the estimates
of modal parameters for different operating conditions. Two algorithms based on the Least Square Complex Exponential
(LSCE) method will be used to estimate the modal parameters. The uncertainties will be calculated using a Monte-Carlo
approach with the hypothesis of constant modal parameters at a given operating condition. In collaboration with
Andritz-Hydro Ltd, data collected on two different stay vanes from an Andritz-Hydro Ltd Francis turbine will be used.
This paper will present an overview of the procedure and the results obtained.
Keywords: Flow induced vibration; modal analysis; system identification; uncertainty; modal parameters.
1. Introduction
Modal properties are needed to predict the dynamic behavior of a structure. For a hydraulic turbine component, this is a major
source of uncertainty because of the presence of water. The forces generated by the fluid flow modify the structure behavior. The
force components in phase with acceleration and displacement will act as added mass and the forces components in phase with
velocity will modify the damping. Even if analytical tools exist to predict these parameters, there is a need for validation with
experimental data and experimental modal analysis may not be sufficient. The main issue with this approach is the difference
between the experimental conditions and the real conditions occurring during operation. Extracting modal parameters directly
from operational data is a good option when measurements are available.
Operational Modal Analysis (OMA) allows the identification of modal parameters directly from the response of a system.
OMA relies upon the assumption that the system input could be assimilated to white noise. The white noise excitation assumption
however is not totally true. The input will almost always contain some harmonic excitations in addition to random input. But since
harmonic excitation can be regarded as virtual mode with no damping, the real mode can be differentiated from the virtual modes
generated by harmonic excitation. In this paper, the Natural Excitation Technique (NExT) combined with the Least Square
Complex Exponential (LSCE) algorithm will be used [1, 2]. The technique is fast and simple to use [3]. Furthermore, the LSCE
algorithm can be modified to explicitly include the harmonic excitations in the identification procedure, therefore leading to a
more robust approach [4].
With OMA algorithm, a change in the algorithm parameters may modify the results. Therefore, in this paper, a sensitivity
analysis will be performed to validate the chosen parameters. The results uncertainty will be estimated using a Monte-Carlo
approach with the hypothesis of constant modal parameters within a given operating condition. Then, using the validated set of
parameter, the results and their associated uncertainty will be calculated. In addition to an overview of the procedure, this paper
will also present results obtained from field test data.
This manuscript was presented at the 24th IAHR Symposium on Hydraulic Machinery and Systems, October 27-31, 2008, Foz do Iguassu-Brazil.
Accepted for publication May 27 2009: Paper number O09023S
Corresponding author: Martin Gagnon, Dr., martin.gagnon.8@ens.etsmtl.ca
279
2. Field Test Data
The data gathered on two different stay vanes from an Andritz-Hydro Ltd Francis turbine has been used. Each stay vane was
instrumented with four strain gauges, 2 on each side, located at each extremity as shown in Fig. 1.
Fig. 1 Strain gauges position
The turbine has 13 blades and 20 guide vanes with a rotating speed of 100 RPM. The #2 stay vanes, close to the inlet, and #18,
toward the end of the spiral case, have been instrumented. Their relative shapes are presented in Fig. 2.
Fig. 2 Instrumented stay vanes relative shapes
Data from 6 operating conditions, ranging from 0 % to 106 % of the Best Efficiency Point (BEP), have been analyzed. The
acquisition sampling rate was 2048 Hz. All the data presented here has been extracted from the same field test. The signal lengths
used for each operating condition are presented in Table 1.
Table 1 Signal length
% of BEP Length of signal [s]
0 250
24 150
39 250
57 250
84 250
106 250
Three harmonic excitations were expected in the stay vanes signal: the rotating speed, runner blade passing frequency from
rotor-stator interaction, and generator electrical interference. The calculated frequencies are shown in Table 2.
Table 2 Generated excitation
Frequency Equation Description
1.67 Hz RPM / (60 sec) Rotating speed
21.67 Hz (# Blades) x (Rotating speed) Rotor-stator interaction
60.00 Hz Generator
280
The expected natural frequencies in water obtained using analytical tools are presented in Table 3 for each stay vane.
Table 3 Expected natural frequency
Stay Vane Mode 1 Mode 2 Mode 3
#2 87 Hz 188 Hz 281 Hz
#18 82 Hz 187 Hz 214 Hz
3. Methodology
The underlying principle of the NExT method is that the correlation function ()
ij
R
t between the responses at positions i and
j is similar to the response at i due to an impulse at j.
/2
1
/2
1
() lim () ( ) sin( )
n
rr
TNtd
ri ri
ij i j r r
dd
Trrr
T
A
Rt q q td e t
Tm
ζω
φ
τ
ττ ωθ
ω
−
→∞ =
−
=−= +
∑
∫ (1)
2
1
dn
rr r
ω
ω
ζ
=−
(2)
Where
ri
φ
The ith component of the eigenmode number r
ri
A Constant associated to the jth response signal
r
m Modal mass
n
r
ω
Eigenfrequency
r
ζ
Modal damping ratio
r
θ
Phase angle associated with the rth modal response
The correlation between signals is the superposition of decaying oscillations having the same damping and natural frequencies
as the structural mode; they can thus be identified by a time domain method like the LSCE method. The correlation function can
also be written as eq. (3).
∑∑
==
ΔΔ +=Δ N
r
N
rrj
tks
rj
tks
ij CeCetkR rr
11
*
*
)( (3)
2
1
rrr r
r
si
ωζ
ω
ζ
=+ −
(4)
Where rj
C is a constant associated with the rth mode for the jth response signal. By numbering complex modes and
eigenvalues in sequence, the eq. (3) becomes eq. (5).
2'
1
() r
N
s
kt
ij rij
r
Rkt Ce
Δ
=
Δ=
∑ (5)
r
Sis in complex conjugate form for which a polynomial of order 2N exists with r
s
t
e
Δ
being the root as shown in eq. (6).
12 212
01 2 21
... 0
NN
rr Nr t
VV V V
ββ β β
−
−
++ ++ +=
(6)
With r
s
t
r
Ve
Δ
= and 21
N
β
=
Multiplying the impulse response for each sample k by k
β
and summing every sample, the following equation is obtained:
22222
''
00110
() 0
NNNNN
kk
k ij k rij r rij k r
kkrrk
Rkt CV C V
ββ β
=====
⎛⎞
⎛⎞
Δ===
⎜⎟
⎜⎟
⎝⎠
⎝⎠
∑∑∑∑∑
(7)
281
The coefficients k
β
can be found by solving eq. (8).
00 11 2121 2
... NN N
R
RRR
β
ββ
−−
+++ =− (8)
Rearranged as a linear system, giving
[
]
}
{
}
{
'
ij ij
RR
β
=− (9)
Since the parameters k
β
are related to modal parameters, they can be derived from any correlation or auto correlation within
the system. Using all correlation functions, eq. (9) becomes
[
]
[]
[]
{}
[
]
[]
[]
11 11
12 12
'
'
'
qp qp
RR
RR
RR
β
⎧
⎫
⎡⎤
⎢
⎥⎪⎪
⎪
⎪
⎢⎥
=−
⎨
⎬
⎢⎥
⎪
⎪
⎢⎥
⎪
⎪
⎢⎥
⎣
⎦⎩⎭
MM
(10)
Where q is the response and p is the reference. The solution for this system equation can be found using pseudo inverse
technique. The modal parameters are then extracted from the root of eq. (8).
A modified version of the LSCE method has been developed by Mohanty and Rixen [4]. This new method is able to deal with
the presence of harmonic excitations by including them explicitly in the identification algorithm using two new terms for each
harmonic excitation. These terms are eigenvalue with no damping as shown in eq. (11).
rr is
ω
±
=
(11)
)sin()cos( titeV rr
ti
rrΔ±Δ== Δ±
ωω
ω
(12)
Equation (12) being the root of eq. (6) the system of eq. (13) can be formed. Adding the linear eq. (13) to eq. (10), the solution
will include the exact frequency of the harmonic excitation.
{}
⎭
⎬
⎫
⎩
⎨
⎧
Δ
Δ
−=
⎥
⎦
⎤
⎢
⎣
⎡
Δ−Δ
Δ−Δ
tN
tN
tNt
tNt
r
r
rr
rr
)2(cos(
))2(sin(
))12(cos(...)cos(1
))12(sin(...)sin(0
ω
ω
β
ωω
ωω
(13)
On stay vane #2, the regular LSCE method failed to identify the first eigenvalue as shown in the stability diagrams Fig. 3. We
were not able to obtain stable results because of the presence of the 4th harmonic of the blade passing frequency (86.7 Hz) near
the first eigenvalue.
Fig. 3 Stay vane #2 – LSCE – 39% BEP
The modified LSCE method was used to deal with the 4th harmonic of the blade passing frequency on stay vane #2. The
results using the modified LSCE method with the fourth harmonic of the blade passing frequency (86.7 Hz) are shown in Fig. 4.
282
Fig. 4 Stay vane #2 – Modified LSCE – 39% BEP
The results obtained with the regular LSCE method were stable on stay vane #18. For this stay vane, the modified LSCE
method did not improve the results and therefore was not used.
4. Sensitivity analysis
To gain confidence in the results obtained, a sensitivity analysis was performed for each stay vane. The modified LSCE
algorithm was used on stay vane #2 and the regular LSCE algorithm was used on stay vanes #18. For stay vanes #2, the fourth
harmonic of the blade passing frequency 86.7 Hz was included in the algorithm. For each stay vane, the results from a defined
window length were bootstrapped using a Monte Carlos approach [5-7]. A section of the defined data length was therefore
randomly selected within the available signal for each estimate. The influence of each selected parameter was analyzed
individually while the others remained constant. No mixed effect was evaluated. The parameter reference values are shown in
Table 4.
Table 4 Parameter values
Parameter Value
Window length 50 000 samples
Model order 80
Over determination factor 10
The window length is the signal length used for the correlation function, the model order is the number of modes estimated
and the over determination factor is defined as the rank of the matrix divided by the number of values calculated in the pseudo
inverse solution. An example of the result distribution for 200 iterations using the reference values is shown in Fig. 5.
Fig. 5 Stay vane #2 – Modified LSCE – 39% BEP
Fig. 6 to Fig. 8 show the results of the sensitivity analysis. Only the results for stay vane #2 at 39% BEP are presented here. In
each figure, the mean value +/- the standard deviation of 200 iterations is shown.
283
Fig. 6 Data length effect – Stay vane #2 – Modified LSCE – 39% BEP
As shown in Fig. 6, a reduction of standard deviation can be obtained with more data in each window. On the other hand the
independence between these windows decreases with their length, especially at 24% BEP, where the available signal is only 150
seconds long.
Fig. 7 Model order effect – Stay vane #2 – Modified LSCE – 39% BEP
Furthermore, it was observed that the over determination factor tends to increase the standard deviation above a certain level,
and it can be observed in Fig. 8.
Fig. 8 Over determination effect – Stay vane #2 – Modified LSCE – 39% BEP
The sensitivity analysis results for all operating conditions on both stay vanes confirmed the reference values presented in
Table 4 as an acceptable choice.
284
5. Results
A common set of parameters have been used for all operating conditions to avoid change in results bias from one condition to
another. Thus, the comparison between results is done using the hypothesis of a constant bias. The added damping generated by
flow-induced force changes with fluid flow characteristics. Therefore the a priori assumptions were that the natural frequencies
will remain constant and the damping ratio will increase with load. The results obtained for the first mode on stay vane #2 are
presented in Fig. 9 and on stay vane #18 presented in Fig. 10. For each load condition, the results mean value +/- the standard
deviation for 200 iterations is shown. The results from stay vane #2 show a constant frequency and a significant increase in
damping ratio at 106% of BEP which matches the a priori assumption. On the other hand, the results on stay vane #18 differ from
stay vane #2 which was not expected. As for stay vane #2 the natural frequency is constant across the operating range but the
damping ratio does not increase. Rather, the damping ratio shows an almost negligible drop at 57% of BEP then comes back to
approximately the same value at 106% of BEP.
Fig. 9 Stay vane #2 – Modified LSCE
Fig. 10 Stay vane #18 – LSCE
The fluid flow characteristics are factors that can influence the added damping ratio [8, 9]. In this case, three main factors
which can influence damping behavior have been identified: the position within the distributor, the profile geometry, and the
presence of a harmonic excitation near the estimated frequency. Effects due to differences between the two identification
algorithms (regular LSCE and modified LSCE) have been neglected because of the constant bias hypothesis.
6. Conclusion
In this paper, field test data have been used to estimate the modal parameters on two stay vanes from a Francis turbine. On stay
vane #2, the modified LSCE method has been used because of the presence of a harmonic excitation near the first eigenvalue. On
stay vane #18, the harmonic excitations were not close enough to influence the results, therefore allowing the use of the regular
LSCE method. The variability of the results has been calculated using a Monte Carlos approach and the parameters used have
been validated by a sensitivity analysis. The hypothesis of a constant bias across all data sets is being used in order to neglect the
difference between the identification algorithms. Two different behaviors are observed: an increase in damping ratio with load is
observed on stay vane #2 and a relatively constant damping ratio is observed on stay vane #18. Three main differences were
identified as possible contributing factors: the position of the stay vane within the distributor, the stay vane profile geometry and
the presence of a harmonic excitation near the estimated frequency on stay vane #2. It was not possible to validate the hypothesis
285
or quantify the possible effect of each contributing factor with the data used here. More research needs to be done which might
include field test data and controlled laboratory experiments. The use of field test data is necessary to identify the component
behavior during operation but more controlled experiments will be needed to validate the assumptions and hypothesis.
Acknowledgments
The author would like to thank Andritz-Hydro Ltd, the National Sciences and Engineering Research Council of Canadian
(NSERC) and the École de technologie supérieure (ÉTS) for their support and financial contribution.
Nomenclature
ri
A Constant associated to the jth response signal ri
φ
The ith component of the eigenmode number r
rj
C Constant associated with the rth mode for the jth
response signal n
r
ω
Eigenfrequency
r
m Modal mass d
r
ω
Damped eigenfrequency
()
ij
R
t Correlation function r
ζ
Modal damping ratio
r
θ
Phase angle associated with the r th modal response
References
[1] James, G., Carne, T. & Lauffer, J., 1992, The Natural Excitation Technique (NExT) for Modal Parameter Extraction from
Operating Wind Turbines, SAND 92-1666, Sandia National Laboratories, Albuquerque, NM, 19921.
[2] James, G., Carne, T. & Lauffer, J, 1995, The Natural Excitation Technique (NExT) for Modal Parameter Extraction from
Operating Structures, Modal Analysis, 10, pp. 260-277.
[3] Andersen, P., Brincker, R., Peeters, B., De Roeck, G., Hermans, L. & Krämer, C., 1999, Comparison of system identification
methods using ambient bridge test data, In Proceedings of IMAC 17, the International Modal Analysis Conference, pp. 1035-1041,
Kissimmee, FL, USA.
[4] Mohanty, P. & Rixen, D. J., 2004, Operational modal analysis in the presence of harmonic excitation. Journal of Sound and
Vibration, 270, pp. 93-109.
[5] Politis, D. N., 2003, The Impact of Bootstrap Methods on Time Series Analysis, Statistical Science, 18, pp. 219-230
[6] Buhlmann, P. & Kunsch, H. R., 1999, Block length selection in the bootstrap for times series, Computational Statistics & Data
Analysis, 31, pp. 295-310.
[7] Vogel, R. M. & Shallcross, A. L., 1996, The moving blocks bootstrap versus parametric time series models, Water Resources
Research, 32, pp. 1875-1882.
[8] Naudascher, E. & Rockwell, D., 1994, Flow-Induced Vibrations: An Engineering Guide, Dover Publications.
[9] Blevins, R. D., 2001, Flow-Induced Vibration 2nd ed., Kieger Publishing Compagny.
