Toward an Improvement in the Identification

of Bridge Deck Flutter Derivatives

Gianni Bartoli1; Stefano Contri2; Claudio Mannini, Ph.D.3; and Michele Righi, Ph.D.4

Abstract: This paper presents a short review of the state-of-the-art methods to identify bridge deck flutter derivatives and proposes a new
algorithm to simultaneously extract the aeroelastic coefficients from free-vibration section-model tests, which is based on the improve-
ment of the unifying least-squares 共ULS兲 method and is therefore called modified unifying least-squares method. The advantages with
respect to ULS are the faster and better convergence and the improvement in accuracy due to the introduction of weighting factors in the
unifying error function. The method has been validated through numerically simulated noisy signals and experimental heaving and
pitching time histories for two different bridge deck cross sections: a single-box and a multiple-box girder section model. The analysis of
the artificial signals shows that a few system parameters are very difficult to be identified due to the fact that the problem is strongly
ill-conditioned. Nevertheless, all the diagonal and off-diagonal components of the stiffness and damping matrices which significantly
contribute to the output of the system are correctly estimated. The improvement with respect to other methods is extensively discussed.
For the wind-tunnel test cases the accuracy of the identification procedure is evaluated through the comparison between measured signals
and those simulated through the estimated mechanical and aerodynamic system parameters with very satisfactory results. With respect to
many previous attempts of validation, this approach clearly shows the degree of accuracy that can be expected from the identification
algorithm. Finally, for the considered test cases the linear model which stands behind the method seems to be an acceptable approximation
of the physics of the phenomenon.
DOI: 10.1061/共ASCE兲0733-9399共2009兲135:8共771兲
CE Database subject headings: Bridge decks; Aeroelasticity; Wind loads; Flutter; Wind tunnels; Identification.

Introduction theory proposed by Theodorsen 共1934兲, often used for airfoils in

subsonic flow, does not apply due to the nonstreamlined and com-
In the design of flexible bridges the aerodynamic and aeroelastic plex geometry of bridge decks and the unavoidable flow separa-
behaviors of the structure under wind are of primary importance. tion. Hence a semiempirical model was proposed by Scanlan and
In particular, the critical flutter wind speed has to be predicted Tomko 共1971兲, wherein the linear nature of theoretical lift and
with an acceptable degree of reliability and an appropriate safety moment 共Theodorsen 1934兲 is kept but the Theodorsen’s circula-
margin with respect to the maximum design wind speed has to be tory function is substituted by eight functions of the reduced wind
assured since flutter can lead to the collapse of the bridge. The speed, called flutter derivatives 共Simiu and Scanlan 1996兲, which
structural response to turbulent wind has to be accurately esti- have to be identified through wind-tunnel tests.
mated as well because the induced displacements, accelerations, According to Scanlan’s formulation of self-excited forces, lift
and stresses can be crucial for both ultimate and serviceability and moment are related to heaving and pitching motion and their
limit states. In order to accomplish these tasks it is necessary to first derivatives. Other authors also consider the drag force and
know the self-excited forces 共along with the buffeting forces兲, that sway motion, including in this way eighteen flutter derivatives
is, the load due to the structural motion itself and the consequent 共Singh et al. 1995; Jain et al. 1996; Chen et al. 2002; Chowdhury
interaction with the airflow. For this purpose the potential flow and Sarkar 2003兲. Bartoli and Mannini 共2005兲 showed that in
most cases the effect on the flutter instability limit of these addi-
1
Associate Professor, CRIACIV/Dept. Civil and Environmental tional functions is negligible. Nevertheless, Katsuchi et al. 共1999兲
Engineering, Univ. of Florence, Via S. Marta 3, 50139 Florence, Italy. found that the flutter derivatives relating drag and pitching motion
E-mail: [email protected]
2 共P*2 and P*3 兲 can have a significant effect on flutter for high-drag
MSc, A & I Progetti, Via Bolognese 48, 50139 Florence, Italy.
E-mail: [email protected] bridge decks such as Akashi Kaikyo Bridge, Japan. Further inves-
3
CRIACIV/Dept. Civil and Environmental Engineering, tigations are still needed to shed some light on this issue. In
Univ. of Florence, Via S. Marta 3, 50139 Florence, Italy. E-mail: contrast with thin airfoil theory, flutter derivatives are also known
[email protected] to be dependent on the mean angle of attack of the wind flow
4
Environmental Protection Officer HQ European Union FORce, 共Diana et al. 2004, Mannini 2006兲 and this nonlinearity must be
Butmir, 71000 Sarajevo, Bosnia Erzegovina. E-mail: michele.righi@ carefully accounted for in flutter calculations.
gmail.com From these considerations it is evident that the accuracy of
Note. This manuscript was submitted on June 5, 2007; approved on
flutter and buffeting calculations strongly depends on the actual
February 20, 2009; published online on July 15, 2009. Discussion period
open until January 1, 2010; separate discussions must be submitted for capability to identify flutter derivatives through wind-tunnel tests.
individual papers. This paper is part of the Journal of Engineering Me- Therefore a reliable and robust method of identification is neces-
chanics, Vol. 135, No. 8, August 1, 2009. ©ASCE, ISSN 0733-9399/ sary, particularly in order to extract off-diagonal terms, due to the
2009/8-771–785/$25.00. fact that the problem is ill-conditioned. As a matter of fact, the

JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009 / 771

Downloaded 24 Jul 2009 to 150.217.9.110. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
key role played in the flutter-instability mechanism by a few off- signals generated from systems with known parameters. More-
diagonal coefficients, such as A*1 and H*3 , has already been under- over, the possible advantages of MULS with respect to other al-
lined by Matsumoto 共1996兲 and Chen and Kareem 共2006兲. In case gorithms in terms of accuracy and efficiency are discussed.
of classical flutter, just simplified approaches 共Matsumoto 1996;
Bartoli and Mannini 2008兲 allow to calculate the critical wind
speed using diagonal terms only. Brief Overview of Free-Vibration Identification
Flutter derivatives can be measured with different techniques. Methods
A reliable one is the forced vibration method 共Li 1995; Matsu-
moto 1996; Diana et al. 2004兲, although it requires a sophisticated Generally speaking, the free-vibration system identification tech-
driving equipment, only rarely available to wind-tunnel laborato- niques briefly reviewed here can be classified as time-domain
ries. This approach allows us to identify flutter derivatives up to parametric methods. When flutter derivatives were first intro-
high reduced wind speed but the uncertainty of the measures is duced 共Scanlan and Tomko 1971兲, the writers showed that it is
usually very high at low reduced wind speed. In particular, Zasso possible to measure them from exponentially decaying signals
et al. 共1996兲 tried to identify the aeroelastic coefficients from a instead of steady-amplitude sinusoidal motion. They proposed to
full-aeroelastic bridge model excited by means of an electrody- extract the aeroelastic functions by means of three different test
namic shaker. setups: one-degree-of-freedom heaving and pitching elastic sys-
The most widely adopted technique is the free-vibration tems to obtain the “direct” flutter derivatives and a two-degree-of
method, wherein a section model is elastically suspended, usually freedom system with nearly coupled frequencies to complete the
allowing two degrees of freedom, and flutter derivatives are ex- set of functions. This approach requires to know the still-air fre-
tracted from the records of the heaving-pitching free-decaying quencies and damping coefficients but the technique to estimate
oscillatory motion from an imposed initial condition, comparing these parameters was not specified. Moreover, the efficiency and
the system matrices obtained with and without wind 共Scanlan and reliability of this procedure is questionable, as already remarked
Tomko 1971; Ibrahim and Mikulcik 1977; Sarkar 1992; Sarkar by Sarkar et al. 共1994兲.
et al. 1994; Gu et al. 2000; Chowdhury and Sarkar 2003兲. This Ibrahim and Mikulcik 共1977兲 proposed the so-called Ibrahim
technique requires more complicated identification algorithms but time domain 共ITD兲 method to identify stiffness and damping ma-
it is appealing for the relatively simple setup. It appears particu- trices of a linear-elastic system from n-degree-of-freedom free-
larly appropriate in case of smooth flow but in some instances decaying oscillatory motion. The method is based on complex
共Sarkar et al. 1994; Gu et al. 2001; Bartoli and Righi 2006兲, the eigenvalue analysis using the redundant number of available data
influence of grid-generated turbulence has also been investigated. for a least-squares estimation of the eigenvalues and eigenvectors
Other authors proposed to identify flutter derivatives from of the system. Pappa demonstrated that the estimation of the ei-
“ambient vibration” measurements, that is, from the response genvalues is affected by a positive bias and proposed a correction
of elastically suspended section models to turbulent wind, without in order to obtain a consistent estimator: the calculation of the
imposing any initial condition 共Bogunovic Jakobsen 1995; system eigenvalues should be repeated introducing a negative
Bogunovic Jakobsen and Hjorth-Hansen 1995; Ricciardelli and de time shift and then the positive and negative time-shift results
Grenet 2002; Fathi 2004; Gu and Qin 2004; Costa et al. 2007兲. should be averaged 共Sarkar 1992兲.
Finally, Caracoglia and Jones 共2003兲 tried to directly measure Another method to identify flutter derivatives simultaneously
indicial functions, which are the time-domain counterpart of flut- from coupled-motion displacement time histories of section-
ter derivatives. This is a more difficult task since it requires to model tests is the modified Ibrahim time domain 共MITD兲, pro-
measure the forces after the imposition of an as impulsive as posed by Sarkar 共1992兲 and Sarkar et al. 共1994兲. It is suggested to
possible stepwise motion of the model. use the ITD solution as an initial value and then an iterative
This paper concentrates on the free-vibration techniques and procedure is outlined, in which the measured time histories are
the state-of-the-art methods are shortly reviewed in the next substituted by the signals simulated through the estimated param-
section. Afterward the modified unifying least-squares 共MULS兲 eters, then performing a new estimation by means of ITD method.
method is presented in detail and the modifications introduced The loop is repeated until convergence. The accuracy of the iden-
with respect to the unifying least-squares 共ULS兲 method 共Gu et al. tification depends on two time shift parameters and an empirical
2000兲, in order to obtain a more efficient and accurate algorithm, formula to choose them is suggested.
are highlighted. This method has been referred to as the combined Gu et al. 共2000兲 proposed the ULS method wherein a unifying
system identification method 共CSIM兲 in some previous authors’ error function is minimized by a nonlinear least-squares proce-
publications 共Bartoli and Righi 2006兲. A remarkable effort has dure. At each iteration the error is given by the difference between
also been devoted to the validation of the method. As a matter of the measured signals and those simulated by means of the current
fact, given the importance of the issue, several algorithms for estimation of the system eigenvalues and eigenvectors. Errors
flutter-derivative identification are now available in the literature relative to heaving and pitching motions are simply summed. In
but their validation does not seem fully convincing, being based order to reach convergence this method requires a reasonable ini-
mostly on comparisons with Theodorsen’s theory results for the tial estimation of the system eigenvalues and in Gu et al. 共2000兲 it
thin flat plate. In particular, it is not possible to distinguish the is suggested to take the MITD solution. The same authors inves-
accuracy due to the identification procedure itself from that due to tigated in another paper 共Gu et al. 2001兲 the effect of turbulence
the underlined mechanical model. The writers believe that the and section-model dynamic parameters on the identification re-
most eloquent and reliable way to check the validity and accuracy sults. This method then was used by Chen et al. 共2002兲 to extract
of an identification method is the comparison between the mea- eighteen flutter derivatives from a three-degree-of-freedom sys-
sured signals and those numerically simulated by means of the tem. Li et al. 共2003兲 proposed to use the ULS technique regarding
estimated mechanical and aerodynamic system parameters. This several vibration records at the same wind speed as an ensemble
validation procedure, which is not common in the literature, is in order to reduce the effect of colored noise of few records on the
followed in this paper. Tests are also performed on artificial noisy convergence of the least-squares iteration process. Weighting fac-

772 / JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009

Downloaded 24 Jul 2009 to 150.217.9.110. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
 m关ḧ + 2␨h␻hḣ + ␻2hh兴 = Lh 共1兲

I关␣¨ + 2␨␣␻␣␣˙ + ␻2␣␣兴 = M ␣ 共2兲

where m and I = mass and mass moment of inertia per unit length;
␻h = 2␲f h and ␻␣ = 2␲f ␣ = circular frequencies of heaving and
pitching modes 共in still air兲; ␨h and ␨␣ = ratio-to-critical damping
coefficients; Lh and M ␣ = lift and moment per unit length; and the
dot denotes derivative with respect to time. Lift and moment can
be expressed as the sum of mean, buffeting, and self-excited
forces but, neglecting the load due to signature turbulence, only
the latter have to be retained when free-decaying motion around a
Fig. 1. Reference system for displacements and self-excited forces
mean position from an initial condition under smooth flow is
considered. Then, supposing perfectly coupled motion, self-
excited forces can be expressed in Scanlan’s form 共Scanlan and
tors are introduced to make each experiment record with the same Tomko 1971; Simiu and Scanlan 1996兲
weight in the total residual error analysis. They also suggest to

冋
take the identification results at the last wind speed as initial
values for the least-squares iteration at the next wind speed. ḣ B␣˙
Chowdhury and Sarkar 共2003兲 outlined the iterative least- Lse共t,K兲 = qB KH*1 共K兲 + KH*2 共K兲
U U

册
squares 共ILS兲 method and used it to extract eighteen flutter de-
rivatives from three-degree-of-freedom elastically suspended h
section models and study the sensitivity of the results to the num- + K2H*3 共K兲␣ + K2H*4 共K兲 共3兲
B
ber of considered degrees of freedom. The advantage of ILS with
respect to ITD, MITD, and ULS methods is the direct estimation

冋
of the state matrix of the system, avoiding the eigenvalue and
eigenvector calculations, thus simplifying the identification pro- ḣ B␣˙
cedure. In contrast, a possible drawback of the algorithm is M se共t,K兲 = qB2 KA*1 共K兲 + KA*2 共K兲
U U

册
the loss of a part of the signal 共“windowing”兲 after the zero-
phase digital filtering, necessary in order to numerically generate h
velocity and acceleration time histories by finite differencing. + K2A*3 共K兲␣ + K2A*4 共K兲 共4兲
B
The accuracy of the identification can be significantly penalized
when aerodynamic damping is high and only few cycles of oscil- where q = 共1/2兲␳U2 = dynamic pressure; U = undisturbed flow
lation are available for the estimation of flutter derivatives. This speed; ␳ = air density; B = bridge deck chord; K = ␻B / U = reduced
algorithm also does not seem to be accurate enough in case of frequency of oscillation; ␻ = circular frequency of the coupled
three-degree-of-freedom systems with high noise-to-signal ratio mode; and the functions H*j and A*j 共j = 1 , . . . , 4兲 are the flutter
共Chowdhury and Sarkar 2003兲. The same authors used the ILS derivatives, which have to be identified through wind-tunnel tests.
method to directly extract the rational function coefficients to be These aerodynamic coefficients are normally expressed as func-
used in time-domain analyses 共Chowdhury and Sarkar 2005兲. tions of the reduced wind speed UR = 2␲ / K. Discussion concern-
Furthermore, Shinozuka et al. 共1982兲 tried to use autoregres- ing the definition of UR and normalization of flutter derivatives
sive moving average models to identify flutter derivatives but the can be found in Iwamoto and Fujino 共1995兲 and Mannini 共2006兲.
result was not satisfactory for high noise signals. Yamada et al. The system of Eqs. 共1兲 and 共2兲 can be generalized to n degrees
共1992兲 proposed an algorithm based on Kalman filters which re- of freedom, which can be collected in the displacement vector x
quires the information about displacement and velocity time his- 苸 Rn. Then, moving the self-excited forces from the right to the
tories as well as initial condition values. This method has also left hand side, the equations of motion can be expressed in vec-
been used by Iwamoto and Fujino 共1995兲, who suggested to in- torial form as follows:
crease the model mass and mass moment of inertia in order to
make the observable decaying signals longer and therefore iden-
tification more accurate. Poulsen et al. 共1992兲 adopted a method Mẍ + Cẋ + Kx = 0 共5兲
which combines control theory and system identification tech- where M 苸 Rn⫻n=total mass; C 苸 Rn⫻n=damping; and K 苸 Rn⫻n
niques and used it to extract the flutter derivatives of the Great = stiffness matrices of the system, including structural and aero-
Belt East Bridge from section-model tests. dynamic contributions. It is worth noting that Eq. 共5兲 is more
general than Eqs. 共1兲 and 共2兲 because it accounts for possible
mechanical coupling between the degrees of freedom due to the
Mechanical Model fact that the section-model centroid, elastic center, and reference
center 共“O” in Fig. 1兲 do not necessarily coincide so that the
According to the simplest and most widespread model to ap- matrices M, C, and K can present nonzero off-diagonal compo-
proach flutter, the mechanical system can be described by a two- nents even in still air. According to complex eigenvalue analysis,
degree-of-freedom linear oscillator, free to vibrate in heaving h共t兲 this equation of order n defines n pairs of complex conjugate
and pitching ␣共t兲 modes 共Fig. 1兲. If the system is supposed to be eigenvalues ␭h 苸 C and eigenvectors ph 苸 Cn, h = 1 , . . . , 2n.
mechanically uncoupled, the equations of motion can be written The solution of Eq. 共5兲 in terms of the displacement vector x at
as follows: the instant t is given by

JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009 / 773

Downloaded 24 Jul 2009 to 150.217.9.110. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
 2n
K = − MG共2,1兲 共20兲
x共t兲 = 兺
h=1
␤ hp he ␭ht 共6兲
Finally, calling C mech
and K mech
the mechanical damping and
stiffness matrices in still air and remembering Eqs. 共3兲 and 共4兲,
where ␤h 苸 C normalizes the corresponding eigenvector in order the flutter derivatives corresponding to the reduced frequency of
to satisfy the initial condition. For a signal containing N data oscillation K, in case of a two-degree-of-freedom heaving-
points, sampled according to the time step ⌬t, k = 1 , . . . , N, and the pitching system, are given by
following matrices can be built:
2
⌳ ¯ 兴 = 关e␭h共k−1兲⌬t兴 苸 C2n⫻N
¯ = 关⌳ 共7兲 H*1 共K兲 = − 关C11共K兲 − Cmech
11 兴 共21兲
hk ␳B2␻

P = 关␤1p1, . . . ,␤2np2n兴 苸 Cn⫻2n 共8兲 2

H*2 共K兲 = − 关C12共K兲 − Cmech
12 兴 共22兲
␳B3␻

冤 冥
x1共0兲 ¯ x1关共N − 1兲⌬t兴
X= ] ] 苸 Rn⫻N 共9兲 2
H*3 共K兲 = − 关K12共K兲 − Kmech
12 兴 共23兲
xn共0兲 ¯ xn关共N − 1兲⌬t兴 ␳B3␻2
so that Eq. 共6兲 can be written in vectorial form
2
H*4 共K兲 = − 关K11共K兲 − Kmech
11 兴 共24兲
¯
X = P⌳ 共10兲 ␳B2␻2

If one defines the state vector z

再 冎
2
A*1 共K兲 = − 关C21共K兲 − Cmech
21 兴 共25兲
x共t兲 ␳B3␻
z共t兲 = 苸R 2n
共11兲
ẋ共t兲
2
the equation of motion 关Eq. 共5兲兴 can be rewritten in the state- A*2 共K兲 = − 关C22共K兲 − Cmech
22 兴 共26兲
␳B4␻
space domain as follows:

ż = Gz 共12兲 2
A*3 共K兲 = − 关K22共K兲 − Kmech
22 兴 共27兲
where ␳B4␻2

G= 冋 0n In
− M−1K − M−1C
苸 R2n⫻2n 册 共13兲 A*4 共K兲 = −
2
␳B3␻2
关K21共K兲 − Kmech
21 兴 共28兲

with 0n = n ⫻ n zero matrix and In = n ⫻ n identity matrix. Accord-

ing to complex eigenvalue analysis, it is possible to write
2n
Modified Unifying Least-Squares Algorithm
z共t兲 = 兺
h=1
␤ hv he ␭ th 共14兲
Given an estimate of the mechanical parameters, the output X of
the system can be simulated and an unifying error function can be
Comparing Eq. 共14兲 with Eq. 共6兲, the hth eigenvector is given by defined as the weighted sum of the errors ei for all degrees of

vh = 再 冎
ph
␭ hp h
苸 C2n 共15兲
freedom

J= 兺
n

w ie i = 兺
n

wi共X共i兲 − X̄共i兲兲 · 共X共i兲 − X̄共i兲兲T 共29兲

The modal matrix in the state-space domain can be expressed in i=1 i=1
the following way:
where 共 · 兲T = complex conjugate transposition. X共i兲 苸 R1⫻N = ith

V = 关v1, . . . ,v2n兴 = 冋 册P
P⌳
苸 C2n⫻2n 共16兲
row of the matrix X, whereas X̄共i兲 苸 R1⫻N = row-vector containing
the record of measured displacements for the ith degree of free-
dom. Therefore ei is defined as the sum of squares of the errors at
where P = 关␤1p1 , . . . , ␤2np2n兴 苸 Cn⫻2n = modal matrix and ⌳ each time instant for the ith degree of freedom. It is worth noting
= diag共␭1 , . . . , ␭2n兲 苸 C2n⫻2n = spectral matrix of the system. that in Gu et al. 共2000兲 the weights wi are assumed all equal to
Therefore the state matrix can be written as unity. This point will be discussed later.
The MULS method consists of the following iterative steps:
G = V⌳V−1 共17兲 1. A first estimation of the system eigenvalues ␭
Then, by defining = 兵␭1 , . . . , ␭2n其 苸 C1⫻2n is obtained through another method.

冋 册
2. Given the current estimation of the eigenvalues, the modal
G共1,1兲 G共1,2兲 matrix of the system P, which minimizes the objective func-
G= 苸 C2n⫻2n 共18兲
G共2,1兲 G共2,2兲 tion J, is determined:
the damping and stiffness matrices of the system can be calcu- ⳵J
=0 共30兲
lated as ⳵P
C = − MG共2,2兲 共19兲 From Eq. 共10兲 it is possible to write that

774 / JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009

Downloaded 24 Jul 2009 to 150.217.9.110. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
 n
X共i兲 = P共i兲⌳
¯ 共31兲 ⳵J ⳵e
where X共i兲 and P共i兲 are the ith rows of the matrices X and P. ⳵⌬
= 兺
i=1
wi
⳵⌬
i
=0 共41兲

This means that the error function associated with the ith
With simple mathematical manipulations, after substituting
degree of freedom depends on P共i兲 only. Hence
Eq. 共39兲, one obtains
n
⳵J ⳵e ⳵e
兺
n
r i
wr 共i兲 = 0 ⇒ 共i兲 = 0, i = 1, . . . ,n 共32兲 ⳵X共i兲
⳵P共i兲
=
r=1 ⳵ P ⳵ P 2 兺
i=1
wi
⳵⌬
· 共X共i兲 − X̄共i兲兲T = 0 共42兲

From the definition of ei one obtains

n
⳵ei ⳵X共i兲
⳵P 共i兲 = 2
⳵P共i兲
· 共X共i兲 − X̄共i兲兲T = 2⌳
¯ 共X共i兲 − X̄共i兲兲T = 0 共33兲 兺
i=1
ˆ 共X共i兲 − X̄共i兲兲T = 0
w i⌳ i 共43兲

Finally, with simple mathematical manipulations

n

¯ 共X共i兲 − X̄共i兲兲T = 0 ⇒ P共i兲⌳

2⌳ ¯⌳¯ T = X̄共i兲⌳
¯T 共34兲 兺
i=1
ˆ 共P共i兲⌳
w i⌳ i
ˆ − X̄共i兲兲T = 0
¯ + ⌬⌳
i 共44兲

and by transposing all terms of the equation

P共i兲 = X̄共i兲⌳
¯ T共⌳
¯⌳¯ T兲−1 共35兲
n n

¯ T共⌳
P = X̄⌳ ¯⌳¯ T兲−1 共36兲
⌬ 兺
i=1
ˆ ⌳
w i⌳ ˆT
i i = 兺
i=1
wi共X̄共i兲 − P共i兲⌳
¯ 兲⌳
ˆT
i 共45兲

¯ has been built from the current

Therefore, once the matrix ⌳ The previous equation can be written as
estimation of the eigenvalues, the calculation of the modal
matrix P is straightforward. ⌬ = BA−1 共46兲
3. Given the current estimation of P, the system eigenvalue
vector ␭ can be calculated by minimizing the objective func- where
tion J n

⳵J ⳵J A= 兺 ˆ ⌳
w i⌳ ˆT
i i 苸C
2n⫻2n
共47兲
=0⇒ = 0, h = 1, . . . ,2n 共37兲 i=1
⳵␭ ⳵␭h
Eq. 共37兲 defines a nonlinear system which should be solved n
through iterations. Being ␭0 the eigenvalue vector used to
estimate P, its updated value can be written as
B= 兺
i=1
wi共X̄共i兲 − P共i兲⌳ ˆ T 苸 C1⫻2n
¯ 兲⌳
i 共48兲

Due to the linearization, Eq. 共46兲 gives only an approximated

␭1 = ␭0 + ⌬ 共38兲
solution of Eq. 共37兲.
where the row-vector ⌬ 苸 C1⫻2n is now the equation un- 4. With the new estimate of the eigenvalues, the modal matrix
known. The problem can be linearized by expressing the dis- P has to be recalculated. In other words, Steps 2 and 3 have
placement time history relative to the ith degree of freedom to be iteratively repeated until converged values of the error
in Taylor series and neglecting all the nonlinear terms function J and of the eigenvalue vector ␭ are obtained.
The proposed method is a modified version of the ULS algo-
2n 2n rithm 共Gu et al. 2000兲. As extensively discussed by two of the
兺 兺
1 0
xi共t兲 = ␤h pihe␭ht = ␤h pihe共␭ +⌬ 兲th h writers in Righi 共2003兲 and Contri 共2003兲, the first difference
h=1 h=1 concerns the way the error function J is minimized, i.e., the way
2n 2n of solving the nonlinear system of equations
⳵x
兺 兺
0
␤h pihe␭ht +
i
= ⌬h + o共⌬兲

冦 冧
h=1 h=1 ⳵ ⌬ h ⳵J共P,␭ = ␭0兲
=0
2n 2n ⳵P
兺 ␤h pihe␭ t + 兺 ␤h piht⌬he␭ t 共49兲
0 0
⬵ h h
⳵J共P = P0,␭兲
h=1 h=1 =0
⳵␭
In matrix formulation this can be written as
As a matter of fact, the ULS method iteratively solves the second
共i兲 ˆ 共i兲 ¯ equation with a loop nested in another loop, wherein a solution
X = P ⌳ + ⌬⌳ i 共39兲 for the first equation is sought. This is not efficient because the
where exact solution ␭ of Eq. 共37兲 is not necessary for each iteration on
P. Conversely, the MULS algorithm suggests an iterative proce-
ˆ hk兴 = 关␤ p 共k − 1兲⌬te␭0h共k−1兲⌬t兴 苸 C2n⫻N
ˆ = 关⌳ dure in which ␭ and P, step by step, simultaneously approach the
⌳ 共40兲
i i h ih
solution of the system of equations. In other words, Eqs. 共30兲 and
and h = 1 , . . . , 2n and k = 1 , . . . , N. The problem consists now 共37兲 are solved at the same time by retaining only the first of the
in finding the vector ⌬ which minimizes the error function iterations which would be required by Step 3

JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009 / 775

Downloaded 24 Jul 2009 to 150.217.9.110. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
 Fig. 2. Flowcharts of ULS and MULS algorithms

776 / JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009

Downloaded 24 Jul 2009 to 150.217.9.110. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
Table 1. Exact Values of the Components of the Stiffness and Damping Matrices of the System and Average of the Absolute Values of Errors in Their
Identification with Different Methods
MITD ILS MULS
Component Unit Exact 共%兲 共%兲 共%兲
K11 kg m−1 s−2 1776.6185 0.69 0.63 0.29
K12 kg s−2 137.9103 2.32 1.43 0.86
K13 kg m−1 s−2 −161.6606 5.24 3.55 1.97
K21 kg s−2 −3.3999 25.19 21.25 16.58
K22 kg m s−2 31.1110 0.35 0.32 0.28
K23 kg s−2 −13.5995 1.52 2.14 1.29
K31 kg m−1 s−2 −40.4151 327.66 318.17 292.57
K32 kg s−2 −44.1476 36.68 38.84 35.40
K33 kg m−1 s−2 7106.4740 0.35 0.34 0.23
C11 kg m−1 s−1 8.7795 14.90 7.47 4.53
C12 kg s−1 2.5643 5.43 3.74 3.87
C13 kg m−1 s−1 −1.5188 11.39 12.48 8.87
C21 kg s−1 −0.7708 6.57 9.05 4.79
C22 kg m s−1 0.2452 1.78 1.64 1.65
C23 kg s−1 0.3417 2.15 2.70 1.48
C31 kg m−1 s−1 −1.3966 682.35 756.68 643.96
C32 kg s−1 −1.0474 103.36 101.99 94.26
C33 kg m−1 s−1 6.1340 5.71 5.19 3.50

冦 冧
⳵J共P,␭兲 B2
=0 w1 = wh = 1, w2 = w␣ = , w3 = w p = 1
⳵P 4
共50兲
⳵J共P,␭兲 where w3 = w p denotes the possible sway 共along-wind兲 degree of
=0
⳵␭ freedom. It is worth noting that the more natural choice for the
weighting factors, i.e., the inverse of the variance of the related
time history, is not optimal when a mode is much more damped
The flowcharts of ULS and MULS algorithms are compared in than the others, as it is for the heaving signal at high wind speed.
Fig. 2. This adjustment does not modify at all the accuracy of the Another possible choice of these factors is the inverse of the
identification but simplifies the implementation of the algorithm squares of the initial condition values, which can be useful when
by eliminating the internal loop in the main iterative loop and a few modes are excited in the experiments much less than the
significantly reduces the time cost of the computation, as it will others. The improvement in accuracy due to these weighting fac-
be shown later in the paper. The reduced complexity of the algo- tors is shown in the next section.
rithm cannot be completely negligible when a large number of Finally, the choice of the starting eigenvalue vector ␭0 needs
long signals have to be processed. to be briefly discussed. Gu et al. 共2000兲 proposed to use the so-
lution of the MITD algorithm. Nevertheless, such a refined esti-
The second modification introduced by the MULS method
mate of the system eigenvalues is not necessary in order to obtain
concerns the weighting factors in the unifying error function J. As
the convergence of the algorithm and any reasonable, although
a matter of fact, if it is assumed wi = 1 ∀i = 1 , . . . , n, as in Gu et al.
inaccurate solution is convenient. In this work the much simpler
共2000兲, J is given by the sum of nonhomogeneous quantities 共h
estimate of ␭0 obtained with the ITD method is employed.
and ␣, for instance兲 and the optimization result is dependent to a
certain extent on the unities chosen for vertical displacements and
rotations. In particular, if the amplitude of oscillation is much
smaller for a degree of freedom with respect to the others 共due to
the imposed initial condition or to the chosen unity兲, the errors on Table 2. Exact Values of the Modal Damping Ratios and Frequencies
that degree of freedom has little effect on the objective function J of the System and Average of the Absolute Values of Errors in Their
and that can imply a poor estimation of the related flutter deriva- Identification with Different Methods
tives. This shortcoming can be overcome by choosing comparable MITD ILS MULS
initial conditions for the system degrees of freedom during the Parameter Unit Exact 共%兲 共%兲 共%兲
wind-tunnel tests and introducing weighting factors in order to
f1 Hz 2.9683 0.40 0.33 0.05
make homogeneous the squares of the errors summed to obtain J.
f2 Hz 4.0302 0.19 0.18 0.05
The writers propose to express the pitching angles as the corre-
f3 Hz 5.9463 0.05 0.05 0.01
sponding vertical displacements of the leading/trailing edge
␨1 0.0867 6.74 4.58 0.60
h␣共t兲 ⬵ ⫿ 共B / 2兲␣共t兲 共for small angles兲, which seems to be a
␨2 0.0692 2.27 4.48 1.06
physically meaningful choice. This is equivalent to assume in
␨3 0.0165 2.08 2.18 0.87
case of pitching angles expressed in radians

JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009 / 777

Downloaded 24 Jul 2009 to 150.217.9.110. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
 Heaving motion Heaving motion
10 4
noisy signal
8
clean signal 3
6 MULS
2
4
1
2
h [mm]

h [mm]
0 0

−2
−1
−4
−2
−6
−3
−8

−10 −4
0 0.5 1 1.5 2 2.5 3 0.5 0.6 0.7 0.8 0.9 1
t [s] t [s]

Pitching motion Pitching motion

3 1
noisy signal
0.8
clean signal
2 MULS 0.6

0.4
1
0.2
α [°]

α [°]
0 0

−0.2
−1
−0.4

−0.6
−2
−0.8

−3 −1
0 0.5 1 1.5 2 2.5 3 0.5 0.6 0.7 0.8 0.9 1
t [s] t [s]

Sway motion Sway motion

20 15
noisy signal
15 clean signal
MULS 10
10
5
5
p [mm]

p [mm]

0 0

−5
−5
−10
−10
−15

−20 −15
0 0.5 1 1.5 2 2.5 3 0.5 0.6 0.7 0.8 0.9 1
t [s] t [s]

Fig. 3. MULS simulation of noisy coupled signals

Validation of the Method

The ULS method has been validated in Gu et al. 共2000兲 compar-

ing the flutter derivatives measured in the wind tunnel for a thin
flat plate with the theoretical solution 共Theodorsen 1934兲. This
procedure, also followed in many other publications, shows the
overall correctness of the algorithm but says little concerning its
actual accuracy. In particular, it is not possible to understand
whether or not the discrepancies between measured and theoreti-
cal values are due to the identification method or simply to the
imperfection of the theory. In this work a double procedure of Fig. 4. Single-box girder cross section

778 / JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009

Downloaded 24 Jul 2009 to 150.217.9.110. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
 Heaving motion Pitching motion
10 2.5

8 U = 0.0 m/s 2 U = 0.0 m/s

6 1.5

4 1

2 0.5

h [mm]

α [°]
0 0

−2 −0.5

−4 −1

−6 −1.5
measured measured
−8 MULS −2 MULS
−10 −2.5
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
t [s] t [s]

Heaving motion Pitching motion

12 2.5
10 U = 5.04 m/s 2 U = 5.04 m/s
8
1.5
6
1
4
2 0.5
h [mm]

α [°]
0 0
−2 −0.5
−4
−1
−6 measured measured
MULS −1.5
−8 MULS
−10 −2

−12 −2.5
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
t [s] t [s]

Heaving motion Pitching motion

15 3
U = 14.87 m/s U = 14.87 m/s
10
2
5
1
0
h [mm]

α [°]

−5 0

−10
−1
measured
−15
MULS
measured
−2
−20 MULS

−25 −3
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
t [s] t [s]

Heaving motion Pitching motion

10 3
U = 19.60 m/s U = 19.60 m/s
2
5

1
0
h [mm]

α [°]

−5
−1

measured
−10
MULS −2 measured
MULS

−15 −3
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
t [s] t [s]

Fig. 5. MULS simulation of measured heaving and pitching signals at different wind speeds for the single-box girder section model

JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009 / 779

Downloaded 24 Jul 2009 to 150.217.9.110. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
 Pitching motion Pitching motion
3 3
U = 14.87 m/s U = 14.87 m/s
2 2

α [°] 1 1

α [°]
0 0

−1 −1

measured measured
−2 −2
MITD MULS

−3 −3
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
t [s] t [s]

Fig. 6. Comparison between the results given by MITD 共left兲 and MULS 共right兲 methods

Heaving motion Heaving motion

0.01 0.01

0.008 U = 0.0 m/s 0.008 U = 0.0 m/s

0.006 0.006

0.004 0.004

0.002 0.002
h [m]

h [m]
0 0

−0.002 −0.002

−0.004 −0.004

−0.006 −0.006
measured measured
−0.008 −0.008
ULS − no weighting factors MULS
−0.01 −0.01
0 2 4 6 8 0 2 4 6 8
t [s] t [s]

Fig. 7. Comparison between the results given by ULS method without weighting factors 共left兲 and MULS method 共right兲

validation is followed. First, noisy signals are generated from an quasisteady theory, as shown for instance in Mannini 共2006兲,
arbitrary system and the identified parameters are compared with while it is assumed P*6 = P*4 = H*6 = H*4 and A*6 = A*4 . The resulting
the known exact values. Then flutter derivatives are identified stiffness and damping matrices of the system, as well as the
from time histories measured in the wind tunnel for two different modal frequencies and damping ratios, can be found in the third
bridge deck section models. In this case the exact mechanical and column of Tables 1 and 2. Heaving-, pitching-, and sway-motion
aerodynamic parameters of the system are unknown and therefore signals 关“clean signals” in Fig. 3兴 are generated through Eq. 共6兲.
the time histories numerically simulated through the estimated Then a Gaussian noise with a standard deviation equal to the 25%
parameters are compared directly with the measured signals. This of the mean fluctuation of the clean signal is superimposed in
procedure is strongly recommended in any attempt of measure- order to obtain the so called “noisy signals.” Ten different signals
ment of the flutter derivatives through free-vibration tests in order are processed in order to perform a statistical analysis of the
to have an idea of the accuracy of the identification. identification results. It is worth noting that the chosen noise-to-
signal ratio is higher than the one normally registered in the wind
tunnel for this type of tests and that this test case is challenging
Numerically Simulated Noisy Signals
for an identification algorithm due to strongly coupled ill-
A three-degree-of-freedom mechanical system, characterized in conditioned system.
still air by the natural frequencies f h = 3 Hz, f ␣ = 5 Hz, f p = 6 Hz, The results of the identification performed with MITD, ILS,
ratio-to-critical damping coefficients ␨h = ␨␣ = ␨ p = 0.01 共where and MULS methods are compared in Tables 1 and 2 in terms of
p = sway motion兲, and mass and mass moment of inertia, respec- the average of the absolute values of errors relative to the com-
tively, m = 5.114 kg/ m and I = 0.0485 kgm2 / m, is taken into ponents of the matrices K and C as well as modal frequencies and
account. Coupling is introduced in the system by self-excited damping ratios of heaving-, pitching-, and sway-branch modes
forces relative to a theoretical thin flat plate 共Theodorsen 1934兲 共f 1, ␨1, f 2, ␨2, and f 3, ␨3 respectively兲. The convergence criterion
at a reduced wind speed UR = 5.0. The aerodynamic coefficients adopted herein and in all the following analyses for the MULS
H*5 , A*5 , P*1 , P*2 , P*3 , and P*5 are calculated according to the method is

780 / JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009

Downloaded 24 Jul 2009 to 150.217.9.110. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
 Fig. 8. Multiple-box girder cross section

储⌬储 艋 10−10 and 冏 冏

⌬J
J
艋 10−12

Fig. 9. Multiple-box girder section model in the CRIACIV wind

tunnel
and ITD solution is used to start the iterative calculation.
From the analysis of Tables 1 and 2 it is evident that a few
terms related to the drag force 共particularly K31, C31, and C32, Experimental Signals for a Single-Box Girder
Section Model
respectively related to the flutter derivatives P*6 , P*5 , and P*2 兲 prac-
tically cannot be identified. In addition, the error in the estimation Realistic wind-tunnel signals are not simply affected by Gaussian
of K32 共that is, P*3 兲 is very large with any considered method. noise but several phenomena have influence on them, such as a
Among the basic eight aerodynamic coefficients 关see Eqs. 共3兲 and certain level of turbulence always present in the wind-tunnel flow,
共4兲兴, K21 共that is, A*4 兲 seems to be the most difficult to be accu- the nonperfect linearity of the mechanical system and so on.
rately identified. Concerning the other components, the errors Therefore the validity of the proposed method has been checked
seem to be larger for the damping matrix than for the stiffness also for experimental free-decaying signals measured in the Inter-
matrix. It is interesting to note that the MULS method gives re- university Research Center on Building Aerodynamics and Wind
sults more accurate than both MITD and ILS algorithms. This Engineering 共CRIACIV兲 boundary layer wind tunnel located in
improvement is particularly evident from Table 2, where the Prato, Italy.
modal properties of the system and therefore global features of The first test case is a two-degree-of-freedom section model
the signals are concerned. The capability of the MULS method to characterized by a single-box girder deck cross section with lat-
simulate the original signals through the identified parameters is eral cantilevers in the bare deck configuration 共Fig. 4兲. All details
clearly shown in Fig. 3. This demonstrates that the aforemen- about the experimental campaign on this common cross-sectional
tioned uncorrectly identified terms contribute very little to the geometry, very similar to the one of the Sunshine Skyway Bridge,
actual definition of the output of the system. Florida, can be found in Mannini 共2006兲, Bartoli et al. 共2006兲, and
The fact that in this particular case the ILS method does not Mannini and Bartoli 共2006兲.
significantly improve the accuracy of the identification with re- In Fig. 5 the measured heaving and pitching signals are com-
pared with those simulated by means of the eigenvalues and
spect to the MITD method is due to the fact that the aerodynamic
eigenvectors estimated through the MULS method. Oscillations
damping in the heaving and pitching signals is quite high so that
in still air 共U = 0.0 m / s兲, at low wind speed 共U = 5.04 m / s兲, at
few cycles of oscillation are actually available for the identifica-
high wind speed but far from flutter 共U = 14.87 m / s兲, and imme-
tion: the windowing step necessary in the ILS algorithm after the
diately before flutter 共U = 19.60 m / s兲 are taken into account. The
zero-phase digital filtering of the displacement time history
agreement between the signals is excellent, such that the mea-
causes the loss of a part of the signal and therefore reduces the sured and simulated records are almost indistinguishable. Only
accuracy of the identification. It is also worth noting that in this near flutter the measured signals show significant perturbations
analysis the standard windowing proposed by Chowdhury and which are not present in the simulated time histories. This result
Sarkar 共2003兲 has not been followed and the initial part of the confirms not only the capability of the proposed method to cor-
signal to be discarded has been limited as much as possible; oth- rectly identify the mechanical and aerodynamic parameters of the
erwise the result would have been significantly less accurate. Un- system but it also shows that the linear mechanical model which
fortunately this condition is quite common also in the wind-tunnel stands behind the method is in this case a good approximation of
practice, especially for the heaving signal 共Mannini 2006兲, as it the physics of the phenomenon, although small nonlinearities can
can be seen, for instance, in Fig. 5. be detected.
Although not reported here, the ITD results show that this The comparison with the results given by the MITD method,
method is clearly inadequate to estimate flutter derivatives in such for instance for the test case characterized by a flow speed of
a complicated test case. In contrast, tests on uncoupled systems 14.87 m / s 共Fig. 6兲, clearly shows the improvement in the identi-
show that all the considered methods give good 共MITD, ILS, and fication obtainable with the MULS method especially in the
MULS兲 or acceptable results 共ITD兲, even in the case of high case of short coupled signals. This fact had already been sug-
noise-to-signal ratio. A large number of this type of tests and gested by Tables 1 and 2 but seems to be even more evident
extensive discussions of the results can be found in Contri 共2003兲 for wind-tunnel records. As a matter of fact the MITD result
and Righi 共2003兲. slightly overestimates the damping in pitch, whereas the MULS

JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009 / 781

Downloaded 24 Jul 2009 to 150.217.9.110. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
 Heaving motion Pitching motion
10 3

8 U = 0.0 m/s U = 0.0 m/s

2
6

4
1
2

h [mm]

α [°]
0 0

−2
−1
−4

−6
−2
−8 measured measured
MULS MULS
−10 −3
0 2 4 6 8 10 0 2 4 6 8 10
t [s] t [s]

Heaving motion Pitching motion

10 3

8 U = 5.42 m/s U = 5.42 m/s

2
6

4
1
2
h [mm]

α [°]
0 0

−2
−1
−4
measured
−6 measured
MULS −2
MULS
−8

−10 −3
0 2 4 6 8 10 0 2 4 6 8 10
t [s] t [s]

Heaving motion Pitching motion

15 1

U = 21.81 m/s 0.8 U = 21.81 m/s

10
0.6

0.4
5
0.2
h [mm]

α [°]

0 0

−0.2
−5
−0.4
measured
measured −0.6
−10 MULS
MULS
−0.8

−15 −1
0 1 2 3 4 5 0 1 2 3 4 5
t [s] t [s]

Heaving motion Pitching motion

20 1.5
U = 24.57 m/s
15 U = 24.57 m/s
1
10
0.5
5
h [mm]

α [°]

0 0

−5
−0.5
−10
measured −1 measured
−15 MULS
MULS

−20 −1.5
0 2 4 6 8 10 0 2 4 6 8 10
t [s] t [s]

Fig. 10. MULS simulation of measured heaving and pitching signals at different wind speeds for the multiple-box girder section model

782 / JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009

Downloaded 24 Jul 2009 to 150.217.9.110. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
 0 0
10 10

MULS MULS
−2 ULS
10 ULS
−5
10
−4
10

|∆J/J| [−]
||∆|| [−]
−6 −10
10 10

−8
10
−15
10
−10
10

−12 −20
10 10
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
iteration iteration

Fig. 11. Convergence of MULS and ULS algorithms for the signals measured for the single-box girder section model at U = 19.60 m / s 共Fig. 5兲

iterations quickly cancel this error. Although less apparent and factors previously introduced are used also for the ULS method in
therefore not reported here for the sake of brevity, improvement order to make straightforward the comparison between the two
can be appreciated also in the simulation of the corresponding algorithms.
heaving signal.
The importance of the introduction of weighting factors in
Eq. 共29兲 is well exemplified by Fig. 7, where the heaving signal Conclusions
simulated through the original ULS method 共h in m and ␣ in
degrees兲 is compared with the one simulated through the MULS After a brief review of the existing free-vibration identification
method, in which the errors summed to obtain J are made homo- methods, a new algorithm to simultaneously extract all flutter
geneous by means of the previously mentioned weighting factors. derivatives from wind-tunnel records is presented. It is derived
In this case it is clear how a nonoptimal, though very common, from the ULS method introducing two basic modifications. The
choice of the units for the heaving and pitching displacements in iterative procedure to minimize the error function is made simpler
the original ULS method can lead to a completely wrong identi- and faster by eliminating the nested loop required by the original
fication result. In many other cases the improvement ascribable to procedure. In addition, weighting factors are defined in the error
the weighting factors is less astonishing but still significant. function in order to obtain comparable heaving and pitching time
histories and improve the accuracy of the identification.
Experimental Signals for a Multiple-Box Girder The second original contribution of this paper is the extensive
Section Model validation of the proposed method. First, tests are performed on
an artificial system, whose parameters are arbitrarily chosen and
The second test case is a multiple-box girder deck section model whose signals are affected by high levels of Gaussian noise,
with a profile similar to the cross section of the proposed Messina showing the good stability of the method. The fact that the prob-
Strait Bridge, Italy 共Figs. 8 and 9兲. The section model is free to lem is strongly ill-conditioned makes the identification of a few
vibrate in the heaving and pitching degrees of freedom. Further terms very difficult. Nevertheless, all diagonal and off-diagonal
details about the wind-tunnel tests can be found in Bartoli et al. components that significantly contribute to the output of the sys-
共2006, 2007兲. In particular, the signals shown in Fig. 10 refer to tem seem to be accurately estimated. Results are compared with
the so called “deck 2” in the configuration with lateral spoilers those given by other well known methods showing a nonnegli-
without any grid. Also in this case the signals simulated with the gible improvement in accuracy obtainable with the proposed al-
MULS method are in perfect agreement with measured time his- gorithm, especially in the challenging case of highly damped and
tories of heaving and pitching free-decaying oscillations. therefore short signals. The MULS method is also validated
through wind-tunnel test cases. In this instance, comparison is not
made on the flutter derivatives, which are unknown, but directly
Improved Convergence of the Algorithm between measured and simulated signals for two different bridge
deck section models. In both cases good results are obtained.
One of the advantages of the MULS algorithm with respect to Finally, the fact that with an approach based on complex modal
the original ULS is its improved efficiency in convergence and analysis it is possible to accurately reproduce the measured time
therefore its reduced computational burden, which can be non- histories, for the considered test cases, confirms the substantial
negligible when a large number of long records have to be pro- validity of a linear model to describe self-excited forces.
cessed 共Mannini 共2006兲. As a matter of fact, to find the exact
solution of Eq. 共37兲 for each iterative value of P is absolutely
unnecessary but can be time-consuming since it requires the in-
ternal loop shown in Fig. 2. By eliminating this loop, along with Acknowledgments
the simplification of the algorithm, the time required for the iden-
tification is significantly reduced and the convergence pattern be- This work has partially been supported by Italian Ministry of
comes more regular, as it is clearly shown by the example University and Research 共MIUR兲 as a part of the National Re-
reported in Fig. 11. It is worth noting that here the weighting search Programs WINDERFUL 共Wind and Infrastructures: Domi-

JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009 / 783

Downloaded 24 Jul 2009 to 150.217.9.110. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
nating Eolian Risk For Utilities and Lifelines兲 and PERBACCO z ⫽ system state-space vector;
共Life-Cycle Performance, Innovation and Design Criteria for ␣ ⫽ pitching rotation of the section model 共rad兲;
Structures and Infrastructures Facing Aeolian and Other Natural ␤h ⫽ multiplicative factor of the hth eigenvector
Hazards兲. accounting for initial conditions;
⌬ ⫽ updating vector of the eigenvalue estimate;
␨h ⫽ still-air ratio-to-critical damping coefficient in the
Notation heaving mode;
␨i ⫽ ratio-to-critical modal damping coefficient relative
The following symbols are used in this paper:
to the ith degree of freedom;
A ⫽ derived matrix of the identification method; ␨ p ⫽ still-air ratio-to-critical damping coefficient in the
A*j ⫽ flutter derivatives corresponding to torque; sway mode;
B ⫽ section model chord length 共m兲; ␨␣ ⫽ still-air ratio-to-critical damping coefficient in the
B ⫽ derived matrix of the identification method; pitching mode;
C ⫽ total damping matrix of the system; ⌳ ⫽ spectral matrix of the system;
Cmech ⫽ mechanical damping matrix of the system 共in still ¯ ⫽ matrix derived from the system spectral properties;
⌳
air兲; ˆ ⫽ matrix derived from the system spectral properties
ei ⫽ total error for the ith degree of freedom; ⌳ i
f h ⫽ still-air frequency of oscillation in the heaving 共ith degree of freedom兲;
mode 共Hz兲; ␭ ⫽ vector collecting the system eigenvalues;
f i ⫽ modal frequency of oscillation relative to the ith ␳ ⫽ air density 共kg/ m3兲;
degree of freedom 共Hz兲; ␻ ⫽ circular frequency of oscillation at flutter 共rad/s兲;
f p ⫽ still-air frequency of oscillation in the sway mode ␻h ⫽ still-air circular frequency of oscillation in the
共Hz兲; heaving mode 共rad/s兲; and
f ␣ ⫽ still-air frequency of oscillation in the pitching ␻␣ ⫽ still-air circular frequency of oscillation in the
mode 共Hz兲; pitching mode 共rad/s兲.
G ⫽ system state-space matrix;
H*j ⫽ flutter derivatives corresponding to lift;
h ⫽ heaving displacement of the section model 共m兲; References
h␣ ⫽ leading-/trailing-edge vertical displacement due to
pitching motion 共m兲; Bartoli, G., D’Asdia, P., Febo, S., Mannini, C., Pastò, S., and Procino, L.
I ⫽ section model mass moment of inertia per unit 共2007兲. “Innovative solutions for the design of long-span bridges:
length 共kgm2 / m兲; Investigation on the aeroelastic behavior of multiple-box girder deck
J ⫽ unifying error function 共m2兲; sections.” Proc., 12th Int. Conf. on Wind Engineering, Cairns, Austra-
K ⫽ reduced frequency of oscillation; lia, Australian Wind Engineering Society 共AWES兲.
K ⫽ total stiffness matrix of the system; Bartoli, G., and Mannini, C. 共2005兲. “From multimodal to bimodal ap-
Kmech ⫽ mechanical stiffness matrix of the system 共in still proach to flutter.” Proc., 6th European Conf. on Structural Dynamics,
air兲; Paris, C. Soize and G. I. Schuëller, eds., Millpress, 349–354.
Bartoli, G., and Mannini, C. 共2008兲. “A simplified approach to bridge
Lh ⫽ total lift force per unit length 共N/m兲;
deck flutter.” J. Wind. Eng. Ind. Aerodyn., 96共2兲, 229–256.
Lse ⫽ self-excited lift force per unit length 共N/m兲; Bartoli, G., Ricciardelli, F., Saetta, A., and Sepe, V. 共2006兲. “Performance
M ⫽ mass matrix of the system; of wind exposed structures.” Results of the PERBACCO project,
M se ⫽ self-excited torque per unit length 共Nm/m兲; Firenze University Press, Florence, Italy.
M ␣ ⫽ total torque per unit length 共Nm/m兲; Bartoli, G., and Righi, M. 共2006兲. “Flutter mechanism for rectangular
m ⫽ section model mass per unit length 共kg/m兲; prisms in smooth and turbulent flow.” J. Wind. Eng. Ind. Aerodyn.,
P ⫽ modal matrix of the system; 94共5兲, 275–291.
P*j ⫽ flutter derivatives corresponding to drag; Bogunovic Jakobsen, J. 共1995兲. “Fluctuating wind load and response of a
p ⫽ sway displacement of the section model 共m兲; line-like engineering structure with emphasis on motion-induced wind
ph ⫽ hth eigenvector of the system; forces.” Ph.D. thesis, Norwegian Institute of Technology, Trondheim,
q ⫽ undisturbed flow dynamic pressure 共N / m2兲; Norway.
t ⫽ time 共s兲; Bogunovic Jakobsen, J., and Hjorth-Hansen, E. 共1995兲. “Determination
of the aerodynamic derivatives by a system identification method.”
U ⫽ undisturbed flow velocity 共m/s兲;
J. Wind. Eng. Ind. Aerodyn., 57共2–3兲, 295–305.
UR ⫽ reduced wind speed; Caracoglia, L., and Jones, N. P. 共2003兲. “A methodology for the experi-
V ⫽ state-space modal matrix of the system; mental extraction of indicial functions for streamlined and bluff deck
vh ⫽ hth eigenvector of the state-space system; sections.” J. Wind. Eng. Ind. Aerodyn., 91共5兲, 609–636.
wh ⫽ weighting factor for the heaving degree of freedom; Chen, A., He, X., and Xiang, H. 共2002兲. “Identification of 18 flutter
wi ⫽ weighting factor for the ith degree of freedom; derivatives of bridge decks.” J. Wind. Eng. Ind. Aerodyn., 90共12–15兲,
w p ⫽ weighting factor for the sway degree of freedom; 2007–2022.
w␣ ⫽ weighting factor for the pitching degree of freedom Chen, X., and Kareem, A. 共2006兲. “Revisiting multimode coupled bridge
共m2兲; flutter: Some new insights.” J. Eng. Mech., 132共10兲, 1115–1123.
X ⫽ matrix of the estimated displacements; Chowdhury, A. G., and Sarkar, P. P. 共2003兲. “A new technique for iden-
tification of eighteen flutter derivatives using a three-degree-of-
X̄ ⫽ matrix of the measured displacements; freedom section model.” Eng. Struct., 25共14兲, 1763–1772.
x ⫽ system displacement vector; Chowdhury, A. G., and Sarkar, P. P. 共2005兲. “Experimental identification
x̄ ⫽ vector of the measured displacements at a certain of rational function coefficients for time-domain flutter analysis.”
time instant; Eng. Struct., 27共9兲, 1349–1364.

784 / JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009

Downloaded 24 Jul 2009 to 150.217.9.110. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright
Contri, S. 共2003兲. “Il flutter negli impalcati da ponte: Identificazione delle aeroelastico di un impalcato da ponte a cassone unicellulare.” Proc.,
derivate aeroelastiche in galleria del vento.” Master’s thesis, Univ. of 9th Italian National Conf. on Wind Engineering IN-VENTO,
Florence, Italy 共in Italian兲. P. D’Asdia, V. Sepe, and S. Febo, eds., Pescara, Italy, Litografia
Costa, C., Borri, C., Flamand, O., and Grillaud, G. 共2007兲. “Time-domain Botolini s.r.l. 共in Italian兲.
buffeting simulations for wind-bridge interaction.” J. Wind. Eng. Ind. Matsumoto, M. 共1996兲. “Aerodynamic damping of prisms.” J. Wind. Eng.
Aerodyn., 95共9–11兲, 991–1006. Ind. Aerodyn., 59共2–3兲, 159–175.
Diana, G., Resta, F., Zasso, A., Belloli, M., and Rocchi, D. 共2004兲. Poulsen, N. K., Damsgaard, A., and Reinhold, T. A. 共1992兲. “Determina-
“Forced motion and free motion aeroelastic tests on a new concept tion of flutter derivatives for the Great Belt Bridge.” J. Wind. Eng.
dynamometric section model of the Messina suspension bridge.” Ind. Aerodyn., 41共1–3兲, 153–164.
J. Wind. Eng. Ind. Aerodyn., 92共6兲, 441–462. Ricciardelli, F., and de Grenet, E. T. 共2002兲. “Evaluation of bridge flutter
Fathi, S. 共2004兲. “Méthodes d’identification des coefficients aéroélas- derivatives from wind excited vibration of section models.” Proc.,
tiques des tabliers de ponts: Application à la technique CBHM.” 5th European Conf. on Structural Dynamics, H. Grundmann, and
EN-CAPE 04.015 R, CSTB, Nantes, France 共in French兲. G. I. Schuëller, eds., Munich, Germany, Taylor and Francis, London,
Gu, M., and Qin, X.-R. 共2004兲. “Direct identification of flutter derivatives 587–592.
and aerodynamic admittances of bridge decks.” Eng. Struct., 26共14兲, Righi, M. 共2003兲. “Aeroelastic stability of long span suspended bridges:
2161–2172. Flutter mechanism on rectangular cylinders in smooth and turbulent
Gu, M., Zhang, R., and Xiang, H. 共2000兲. “Identification of flutter deriva- flow.” Ph.D. thesis, Univ. of Florence, Italy.
tives of bridge decks.” J. Wind. Eng. Ind. Aerodyn., 84共2兲, 151–162. Sarkar, P. P. 共1992兲. “New identification methods applied to the response
Gu, M., Zhang, R., and Xiang, H. 共2001兲. “Parametric study on flutter of flexible bridges to wind.” Ph.D. thesis, Johns Hopkins Univ., Bal-
derivatives of bridge decks.” Eng. Struct., 23共12兲, 1607–1613. timore.
Ibrahim, S. R., and Mikulcik, E. C. 共1977兲. “A method for the direct Sarkar, P. P., Jones, N. P., and Scanlan, R. H. 共1994兲. “Identification of
identification of vibration parameters from the free response.” Shock aeroelastic parameters of flexible bridges.” J. Eng. Mech., 120共8兲,
and Vibration Bull., 47共4兲, 183–198. 1718–1742.
Iwamoto, M., and Fujino, Y. 共1995兲. “Identification of flutter derivatives Scanlan, R. H., and Tomko, J. J. 共1971兲. “Airfoil and bridge deck flutter
of bridge deck from free vibration data.” J. Wind. Eng. Ind. Aerodyn., derivatives.” J. Engrg. Mech. Div., 97共6兲, 1717–1737.
54–55, 55–63. Shinozuka, M., Yun, C. B., and Imai, H. 共1982兲. “Identification of linear
Jain, A., Jones, N. P., and Scanlan, R. H. 共1996兲. “Coupled flutter structure dynamic system.” J. Engrg. Mech. Div., 108共6兲, 1371–1390.
and buffeting analysis of long-span bridges.” J. Struct. Eng., 122共7兲, Simiu, E., and Scanlan, R. H. 共1996兲. Wind effects on structures, 3rd Ed.,
716–725. Wiley, New York.
Katsuchi, H., Jones, N. P., and Scanlan, R. H. 共1999兲. “Multimode Singh, L., Jones, N. P., Scanlan, R. H., and Lorendeaux, O. 共1995兲. “Si-
coupled flutter and buffeting analysis of the Akashi Kaikyo Bridge.” multaneous identification of 3-dof aeroelastic parameters.” Proc., 9th
J. Struct. Eng., 125共1兲, 60–70. Int. Conf. on Wind Engineering, Wiley, New Delhi, India, 972–981.
Li, Q. C. 共1995兲. “Measuring flutter derivatives for bridge sectional mod- Theodorsen, T. 共1934兲. “General theory of aerodynamic instability and
els in water channel.” J. Eng. Mech., 121共1兲, 90–101. the mechanism of flutter.” NACA Technical Rep. No. 496, Annual
Li, Y., Liao, H., and Qiang, S. 共2003兲. “Weighting ensemble least-square Rep. No. 20, NACA, Langley Research Center, Hampton, Va.
method for flutter derivatives of bridge decks.” J. Wind. Eng. Ind. Yamada, H., Miyata, T., and Ichikawa, H. 共1992兲. “Measurement of aero-
Aerodyn., 91共6兲, 713–721. dynamic coefficients by system identification methods.” J. Wind. Eng.
Mannini, C. 共2006兲. “Flutter vulnerability assessment of flexible bridges.” Ind. Aerodyn., 42共1–3兲, 1255–1263.
Ph.D. thesis, Univ. of Florence, Italy–TU Braunschweig, Germany Zasso, A., Cigada, A., and Negri, S. 共1996兲. “Flutter derivatives identifi-
共Verlag Dr. Müller, Saarbrücken, 2008兲. cation through full bridge aeroelastic model transfer function analy-
Mannini, C., and Bartoli, G. 共2006兲. “Analisi del comportamento sis.” J. Wind. Eng. Ind. Aerodyn., 60, 17–33.

JOURNAL OF ENGINEERING MECHANICS © ASCE / AUGUST 2009 / 785

Downloaded 24 Jul 2009 to 150.217.9.110. Redistribution subject to ASCE license or copyright; see http://pubs.asce.org/copyright