Beruflich Dokumente
Kultur Dokumente
SUMMARY
A reassessment of the dynamic characteristics of the 542 m cable-stayed Bayview Bridge in Quincy,
Illinois, is presented using a newly developed output-only system identication technique. The tech-
nique is applied to an extensive set of ambient vibration response data acquired from the bridge in
1987. Vertical, torsional and transverse modal frequencies of the deck are identied, and uncertainty
in damping values are estimated using an automated procedure on several redundant measurements
at four locations. Important practical implementation issues associated with the implementation of the
procedure and selection of algorithm design parameters for stochastic subspace identication techniques
are discussed.
An overall mean and standard deviation of damping of 1:0 0:8% is estimated considering all
identied vertical, torsional and transverse modes in the 02 Hz band. The mean damping for the
fundamental vertical mode (0:37 Hz) is identied as 1:4 0:5%, and for the rst coupled torsion
transverse mode (0:56 Hz) is identied as 1:1 0:8%. Variability in the damping estimates is shown
to decrease as estimated modal RMS acceleration levels increase. Standard deviations on estimated
damping range from 0.05% to 2%. The results are shown to be a substantial improvement in the
evaluation of damping compared to earlier spectral analysis conducted on the same data set. Copyright
? 2005 John Wiley & Sons, Ltd.
KEY WORDS: ambient vibration; cable-stayed bridges; damping; Quincy Bayview Bridge; system
identication; stochastic subspace methods; expectation maximization algorithm; output-
only identication
1. INTRODUCTION
In 1987 Wilson and Liu [1] conducted an extensive ambient vibration survey (AVS) on the
newly constructed 542 m cable-stayed Bayview Bridge across the Mississippi River at Quincy,
Correspondence to: John C. Wilson, Department of Civil Engineering, McMaster University, Hamilton, Ontario,
Canada L8S 4L7.
E-mail: jcwilson@mcmaster.ca
Illinois. The purpose of the test program was to obtain full-scale measurements to determine
the dynamic properties of the bridgemodal frequencies, mode shapes and damping. The
results of analysis of the AVS data, reported in this journal in 1991 [1], clearly identied
frequencies and mode shapes associated with many closely-spaced low-frequency modes within
a bandwidth of 02 Hz. Less certain, however, were estimates of damping, which were made
using half-power bandwidth measurements for only two modes. Since that time, signicant
advances have been made in the development of techniques of system identication for the
analysis of dynamic response data from civil engineering structures. The emergence of output-
only techniques, and the authors current work in this area, have provided a motivation to
re-examine the original AVS data with emphasis on obtaining more information about the
damping of the bridge than was possible from the analysis conducted in the late 1980s.
In this paper, the previously developed EMSSI technique [2] is applied to the 1987 Quincy
AVS data. The method combines stochastic subspace identication (SSI), with the expectation
maximization (EM) algorithm, for estimation of modal parameters. This method provides a
more powerful way to look at the ambient vibration data than was available when it was rst
analysed by Wilson and Liu [1]. The redundancy in the measurement process (multiple mea-
surements at some locations), allows multiple estimates to be made of modal frequencies and
damping ratios for vertical (V), torsional (T) and transverse (Tr) modes of the superstructure
(deck) for quantication of their variability. In this study we do not examine mode shapes,
for reasons discussed later. Detailed mode shape information is available in Reference [1].
The paper is organized as follows: rst, details of the Bayview Bridge are presented along
with information regarding the dynamic measurement program. This is followed by a brief
review of the EMSSI technique, with focus placed on issues associated with its practical
implementation. Results from applying the technique to the Quincy AVS data are presented
together with a detailed discussion, placing emphasis on new information obtained about
damping.
Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:787805
REASSESSMENT OF DYNAMIC CHARACTERISTICS: QUINCY BAYVIEW BRIDGE 789
Figure 2. Sensor locations and labelling from ambient vibration survey (eastern half of Bayview Bridge).
(station congurations) were used for measurement of ambient responses: 36 set-ups along
the deck (main and side spans), 5 set-ups at the east tower, 5 set-ups involving decktower
combinations, 3 set-ups involving towerpier combinations and 6 calibration set-ups along
the deck and tower. Measurements included vertical (V), longitudinal (L) and transverse (Tr)
accelerations, with the majority being in the vertical direction along the deck. Each test set-up
involved four concurrent recording stations, with most measurements taken at a sampling rate
of 50 samples per second, per channel, for 10 minutes. Figure 2 illustrates the measurement
locations used in the survey and their station names. Arrows indicate the directions of motion
Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:787805
790 B. A. PRIDHAM AND J. C. WILSON
recorded by the accelerometers. Stations along the deck coincided with cable connection points,
and on the tower they were selected to capture the coupled motion of the deck and tower.
Wind speed was not recorded. Complete details are available in Reference [1].
3. IDENTIFICATION TECHNIQUE
In this section we present aspects of the EMSSI identication scheme introduced in Refer-
ence [2] that are relevant to the present study.
xk+1 = Axk + wk
(1)
yk = Cxk + vk
where A Rnn is the system matrix containing modal information and C Rln is the ob-
servation matrix. The l measured outputs are denoted by yk Rl1 and xk Rn1 is the state
vector (n is the dimension of the state space, i.e. theoretically, twice the number of modes in
the system). The initial state vector x0 is assumed to be normally distributed with mean 0
and covariance , i.e. x0 N( 0 ; ). The term wk Rn1 represents the process noise, which
accounts for modelling uncertainties as well as stochastic system input, and the measurement
noise vk Rl1 accounts for uncertainty in the measurement process as well as the measure-
ment model. Both noise quantities are assumed to be stationary Gaussian white-noise processes
with zero means and covariances E[wk wkT ] = Q, and E[vk vkT ] = R. For the EM algorithm it is
assumed that E[wk vkT ] = 0.
The objective of system identication is to estimate the parameter set = (A; C; Q; R; 0 ; ).
In this paper we are concerned primarily with the estimation of A since both the modal fre-
quencies (fr ) and damping ratios (r ) are recovered from its eigendecomposition.
Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:787805
REASSESSMENT OF DYNAMIC CHARACTERISTICS: QUINCY BAYVIEW BRIDGE 791
where represents the parameter set (A; C; Q; R; 0 ; ). This step requires a set of forward
and backward recursions employing the Kalman lter for calculation of the state and state
covariance estimates.
Once the expectation step has been computed, the maximization step of the EM algorithm
computes the maximization of Q over the parameter space:
M (i) = arg max Q ; (i) (3)
This approach ensures a monotonically increasing incomplete data likelihood L() at each
step. A detailed formulation can be found in Reference [2].
To ensure the migration of the solution space towards an optimal maximum, as well as fast
convergence of the estimation procedure, it is imperative that an accurate initialization model
be estimated for use in the EM algorithm. In this paper, the stochastic subspace identication
method is used for initialization. This approach is denoted as EMSSI, and has previously
been applied to seismic response data by the authors [2].
To obtain the initial subspace estimate, the output data are collected in a Hankel
matrix [6]:
j
y0 y1 yj1
yi2 yi1 yi+j3
(4)
def 1 yi1 yi yi+j2 Y0|i1 li def Yp
Y0|2i1 = def
= =
j
yi yi+1 yi+j1
Y i|2i1 li Yf
yi+1 yi+2 yi+j
y2i1 y2i y2i+j1
The dimension of this matrix is li j, where l is the number of output sensors (i.e. y0 is
l 1), i is the projection horizon and j is the length of the identied state sequence.
Following the construction of Y0|2i1 , an orthogonal projection is computed using the LQ
decomposition. This matrix is then decomposed using the singular value decomposition (SVD),
and an estimate of the observable subspace is obtained. Estimates of the system matrices may
then be computed using a least squares algorithm. It should be noted that for the analysis of
the Quincy data, the CVA weighting scheme [7] is used to compute the initial estimates.
4. ANALYSIS METHODOLOGY
Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:787805
792 B. A. PRIDHAM AND J. C. WILSON
statistics on the estimated damping ratios, only those signals for which three or more mea-
surements were made have been included in the analysis. For the vertical measurements on
the deck and longitudinal measurements of the towers, we used only those signals for which
concurrent measurements on both the north and south sides of the bridge were available.
This allows for separation of translational and torsional motions (by adding and subtracting
signals) and maintains the estimated model orders at manageable levels. With these factors
in mind, the six stations selected for inclusion in the analysis are: 5; 10; 11, TNL, TSL, TCB
(see Figure 2). The signals from these stations were classied as at one of four spatial lo-
cations on the structure: stations 5; 10; 11 and tower. This limited number of locations makes
it impractical to identify mode shapes. For ease of visualization all transverse data from the
tower, (i.e. signals from TNL, TSL and TCB) are grouped and statistics are computed for the
entire ensemble of results. Table I lists the number and length of signals considered at each
location for identications associated with the three directions of motion of the deck. Vertical
and torsional identications used vertical and torsional motions from the deck plus longitu-
dinal motions from the towers. Transverse identications used transverse measurements from
all four locations. Measurements from station 11 contain the greatest amount of information,
whereas those from station 10 are fewest in number. Note, however, that four of the signals
recorded at station 10 are 20 minutes in length, twice as long as the signals from station 11.
These longer signals generally provide for better estimates. During analysis we focus on a
02 Hz bandwidth. This is reasonable from both a wind and earthquake engineering perspective
since the response will typically be dominated by modes in this band.
Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:787805
REASSESSMENT OF DYNAMIC CHARACTERISTICS: QUINCY BAYVIEW BRIDGE 793
it is important that a preliminary analysis of each data set be conducted to determine the best
possible choices of each parameter.
From the denition of the data Hankel matrix in Equation (4) the values of i and j are
constrained by Ny = 2i + j 1, where Ny is the length of the data set. The value of j is
automatically determined by the choice of i given a specied length of data. An upper bound
on the order of the system n is pre-specied by li. Lower order values may be obtained by
retaining a selected number of singular values during the SVD of i .
Given the limitations on i, j, and n, a method for calibrating the SSI initialization model
is as follows:
1. For the p-th measurement signal yp of length Nyp , a series of ij pairings are specied.
For each ij pairing, a series of subspace estimates of the system are obtained, each
corresponding to a dierent model order, n.
2. Frequency and damping estimates at each order, for a given ij pairing, are tracked and
the number of estimated parameters exhibiting stability is monitored. The r-th pole is
considered to be stable at order if the percentage deviation in frequency and damping
computed by:
fr fr1 r r1
fr = 100%; r = 100% (5)
fr1 r1
are less than or equal to 1% and 5%, respectively.
3. For each ij pairing, the order value with the maximum number of stable poles, as well
as the number of stable poles, are retained.
4. Once all specied models have been estimated, the top ten {i; j; n} triples, sorted in
descending order according to the maximum number of stable poles, are considered the
ten optimal subspace estimates for use as initialization of the EM algorithm.
Each of the ten SSI models represents a dierent starting point in the parameter space for
the EM optimization. The selection of ten models has no real statistical signicance, however,
this value maintains the volume of information at a reasonable level during post-processing,
since several hundred parameter estimates are being tracked. Furthermore, the choice of ten
models allows the inclusion of a greater range of i and n values in the estimation process than
would the selection of a smaller number of models. As a result, parameter estimates from
the ten optimized models will incorporate uncertainty regarding knowledge of the dynamics
of the structure, the system order, and nite sample eects on the estimated projection Owi .
Each of these uncertainties are strong functions of i, j, n and Ny , and can never be exactly
quantied using experimental response data.
Once the ten best {i, j, n} triples have been selected for every data signal used in the
analysis, the EMSSI algorithm is invoked ten times using the corresponding estimated system
matrix A and observation matrix C from each SSI model. The state covariance matrix, ,
and initial state mean, 0 , are set to the identity matrix and a zero vector, respectively. The
initial noise covariances Q0 , R0 , are estimated from the subspace algorithm. The algorithm
is set to run until the value of the likelihood function L() converges to within 0.1%
of the previous estimate, or until the number of cycles of the algorithm reaches 400. For
most signals the algorithm converges in approximately 200 cycles, the exception being the
transverse data which are noisy. For these cases, convergence typically occurred in under
300 cycles.
Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:787805
794 B. A. PRIDHAM AND J. C. WILSON
4.3. Post-processing
Peeters [9] presented an inverse ltering technique for estimating RMS acceleration levels
for each mode within a measured signal. The technique requires the use of the whitening
lter formulation for the state space model of Equation (1). To arrive at this formulation,
the model is transformed to innovations form through application of the Kalman lter, thus
arriving at the following:
zk+1 = Azk + Kzk
(6)
yk = Czk + ek
where zk are the ltered states, K Rnl is the Kalman gain matrix and ek Rl1 is the
innovation (prediction error) white noise sequence. By invoking the eigendecomposition of
A = 1 and left multiplying the rst equation in (6) by 1 , the innovation model may
be placed in modal form, and rearranged to arrive at the whitening lter:
zm; k+1 = ( Km )zm; k + Km yk
(7)
ek = zm; k + yk
where the modal states are zm; k = 1 zk , the modal Kalman gain Km = 1 K and the mode
shape matrix = C. With this model, the responses yk act as inputs and the outputs are
the white sequence ek . The predicted outputs are yk = zm; k , which may then be segregated
into modal contributions using each mode shape vector:
n
n
yk = yrk = {r }zm;
r
k (8)
r=1 r=1
This inverse ltering operation is applied to each of the ten models estimated for each data
signal. The estimated RMS levels of modal acceleration (r ) are retained and processed along
with the modal frequencies and damping ratios to arrive at statistical information.
Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:787805
REASSESSMENT OF DYNAMIC CHARACTERISTICS: QUINCY BAYVIEW BRIDGE 795
frequency estimate is then used for scanning of the remaining estimates, and the process
is continued until all rows of have been either placed in a cluster (and retained in ),
or rejected. This scanning process ensures that only those modes consistently identied
are used for further processing.
3. Steps 1 and 2 are repeated for each measurement signal. The same matrix is updated
throughout the scanning procedure.
4. Once all signals have been identied, clustering analysis of is performed, and statistics
are computed for each modal cluster.
5. The results for this station are stored, and Steps 1 to 4 are then repeated for the next
station on the bridge.
Use of the clustering analysis of Step 4 has been motivated by successful applications to auto-
mated modal analysis procedures in the mechanical engineering literature [10]. The MATLAB
clustering algorithm is described in Reference [11]. The frequency estimates from all iden-
tication runs are used in the algorithm, and the clusters are determined according to the
frequency information. Damping ratios and modal RMS accelerations are assigned to clusters
associated with the frequency estimates for that mode, and the members of the cluster are
used to compute statistical information. The clustering analysis permits separation of close
modes, and rejects frequencies that were not consistently identied in all of the runs (e.g.
weakly excited modes). This ensures that the estimates used in the statistical computations
correspond to modes that were consistently identied.
Additional details on the calibration, post-processing and statistical analysis methodologies
are presented in Reference [12].
5. RESULTS
Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:787805
796 B. A. PRIDHAM AND J. C. WILSON
5 5 3
Station 5
4.5 4.5
Estimated Damping (%)
4 4
-4
2
s (%)
3 3
2.5 2.5
2 2
1
1.5 1.5
1 1
0.5 0.5
0 0 0
5 5 3
Station 10
4.5 4.5
Estimated Damping (%)
4 4
-4
2
s (%)
3 3
2.5
2.5
2 2
1
1.5 1.5
1 1
0.5 0.5
0 0 0
5 5 3
Station 11
4.5 4.5
Estimated Damping (%)
4 4
-4
2
s (%)
3 3
2.5
2.5
2 2
1
1.5 1.5
1 1
0.5 0.5
0 0 0
5 5 3
Tower
4.5 4.5
Estimated Damping (%)
4 4
Mean PSD (10-4 )
3.5 3.5
2
s (%)
3 3
2.5
2.5
2 2
1
1.5 1.5
1 1
0.5 0.5
0 0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Estimated Frequency(Hz) Mean Estimated Frequency (Hz) Frequency (Hz)
Figure 3. Vertical response: damping estimate scatter plots (left), mean () one standard deviation
(s) bounds (middle), and average power spectra for all signals (right), for stations 5; 10, and 11 vertical
motions, and tower longitudinal motions (top to bottom).
columns for each direction list the number of unique modes identied from all signals at that
station, and the total number of individual estimates, respectively. The third column provides
the mean and standard deviation of the estimates. For each direction, the mean values at the
four stations all fall within a narrow range, and with only one exception the standard deviations
Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:787805
REASSESSMENT OF DYNAMIC CHARACTERISTICS: QUINCY BAYVIEW BRIDGE 797
5 5 1.5
Station 5
4.5 4.5
Estimated Damping (%)
4 4
s (%)
3 3
2.5
2.5
2 2
1.5 1.5 0.5
1 1
0.5 0.5
0 0 0
5 5 1.5
Station 10
4.5 4.5
Estimated Damping (%)
4 4
-5
1
s (%)
3 3
2.5
2.5
2 2
1.5 1.5 0.5
1 1
0.5 0.5
0 0 0
5 5 1.5
Station 11
4.5 4.5
Estimated Damping (%)
4 4
3 3
2.5
2.5
2 2
1.5 1.5 0.5
1 1
0.5 0.5
0 0 0
5 5 0.5
Tower
4.5 4.5
Estimated Damping (%)
4 4 0.4
Mean PSD (10-6 )
3.5 3.5
s (%)
3 3 0.3
2.5 2.5
2 2 0.2
1.5 1.5
1 1 0.1
0.5 0.5
0 0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Estimated Frequency (Hz) Mean Estimated Frequency (Hz) Frequency (Hz)
Figure 4. Torsional response: damping estimate scatter plots (left), mean () one standard deviation
(s) bounds (middle), and average power spectra for all signals (right), for stations 5; 10; 11, and tower
torsional motions (top to bottom). Note the dierence in scale on the PSD plot for tower motions.
lie within the range of 0.70.9%. For these 12 cases, the maximum mean plus one standard
deviation is 2.1%. The last line of the table gives estimates averaged across all stations.
Although there are some slight dierences in the means and standard deviations amongst the
various cases in Table II, these dierences are negligible from a practical perspective. The
remarkable closeness in the overall estimates suggest that, statistically, the damping of the
Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:787805
798 B. A. PRIDHAM AND J. C. WILSON
5 5 1.5
Station 5
4.5 4.5
Estimated Damping (%)
4 4
-5
1
s (%)
3 3
2.5
2.5
2 2
1.5 1.5 0.5
1 1
0.5 0.5
0 0 0
5 5 1.5
Station 10
4.5 4.5
Estimated Damping (%)
4 4
-5
1
s (%)
3 3
2.5
2.5
2 2
1.5 1.5 0.5
1 1
0.5 0.5
0 0 0
5 5 1.5
Station 11
4.5 4.5
Estimated Damping (%)
4 4
-5
1
s (%)
3 3
2.5
2.5
2 2
1.5 1.5 0.5
1 1
0.5 0.5
0 0 0
5 5 1.5
Tower
4.5 4.5
Estimated Damping (%)
4 4
3.5
Mean PSD (10 )
3.5
-5
1
s (%)
3 3
2.5 2.5
2 2
1.5 1.5 0.5
1 1
0.5 0.5
0 0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Estimated Frequency (Hz) Mean Estimated Frequency (Hz) Frequency (Hz)
Figure 5. Transverse response: damping estimate scatter plots (left), mean () one standard deviation
(s) bounds (middle), and average power spectra for all signals (right), for stations 5; 10; 11, and tower
transverse motions (top to bottom).
bridge during ambient response can be adequately described by a mean damping value of
1% and a standard deviation of 0.8%. Examination of the middle columns of Figures 35
shows that the range of 0.21.8% damping includes virtually all mean damping values for the
individual frequencies. For station 11, where there is the greatest amount of data, virtually all
mean-plus-one standard deviation values are less than 2%. On a statistical basis, the means
Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:787805
Table II. Damping statistics across all modes for each measurement location.
Vertical motion Torsional motion Transverse motion
a
No. of No. of Damping (%) No. of No. of Damping (%) No. of No. of Damping (%)
Location modes estimates modes estimates modes estimates
s
s
s
Station 5 6 326 1:3 0:7 7 367 1:3 0:9 5 314 0:8 0:8
Station 10 8 219 1:1 0:7 7 172 0:9 0:8 8 242 1:0 0:5
Station 11 8 1699 0:9 0:7 10 1946 0:9 0:7 9 1040 0:9 0:9
Tower 7 287 1:3 0:9 7 288 1:1 0:8 4 443 0:8 0:7
All Stations 10 2531 1:0 0:7 13 2773 1:0 0:8 14 2039 0:9 0:8
Table III. Damping estimates for fundamental vertical, coupled torsionaltransverse modes.
Vertical (V) Torsional (T) Transverse (Tr) Combined T and Tr
(0:37 Hz) (0:56 Hz) (0:56 Hz) (0:56 Hz)
No. of Damping (%) No. of Damping (%) No. of Damping (%) No. of Damping (%)
Location estimates estimates estimates estimates
s
s
s
s
Station 5 55 1:8 0:5 55 1:4 1:1 68 1:5 1:3 123 1:5 1:2
Station 10 29 1:7 0:3 21 1:4 0:9 39 1:1 0:5 60 1:2 0:6
Station 11 240 1:2 0:4 213 0:8 0:7 160 1:0 0:6 373 0:9 0:7
REASSESSMENT OF DYNAMIC CHARACTERISTICS: QUINCY BAYVIEW BRIDGE
All Stations 373 1:4 0:5 336 1:0 0:8 266 1:2 0:8 602 1:1 0:8
and standard deviations, for the station 11 and for the all stations cases, imply that in at
least 90% of all identications using this data the estimated damping will be found to be less
than 2%.
5.1.1. Vertical response. A total of 10 frequencies, ranging from the fundamental vertical
mode at 0:37 Hz to a mode at 1:75 Hz were selected for damping identication in a 02 Hz
band at the four locations. All modes were identied at least 120 times across the 24 data
sets, with some modes being identied over 200 times. Mean estimated frequencies from the
vertical motions have been compared with the spectral analysis results of Wilson and Liu [1].
Frequencies that passed the tolerances of the selection procedure are in excellent agreement
with results from this earlier work. The only mode where there was a dierence in the two
studies is the 0.63 Hz mode identied at station 11. In Reference [1] this was identied
as a transverse mode, however, in the present study components of vertical, torsional and
transverse motions were identied, implying a more complex behaviour of the bridge at this
frequency than was possible to identify in the earlier study.
Mean damping estimates for all vertical modes shown in Figure 3 are within the range of
0.52%. From Table II, averaging across all mode identications at all stations (weighted by
the number of estimates at each station) gives a mean and standard deviation for damping
in the vertical direction of 1:0 0:7%. The rst vertical mode at 0:37 Hz is in agreement
with the spectral analysis results of Wilson and Liu [1]. The mean and standard deviation
of the damping estimates for this mode in Table III, based on 240 estimates at station 11,
is 1:2 0:4%. When information for the 0:37 Hz mode is averaged across all four locations
the result is 1:4 0:5%. These results are in close agreement with one another. Wilson and
Liu [1] report an average damping value for this mode in the range of 2.02.6%, based on
half-power bandwidth analysis. The largest ambient response (largest PSD peak) in Figure 3
is at 1:07 Hz at stations 10 and 11, corresponding to the fth vertical mode identied in
Reference [1]. The average damping identied for this mode across all stations in the vertical
direction is 0:9 0:7%.
The estimated frequencies in the rst two columns of Figure 3 are in excellent agreement
with the locations of the PSD peaks in the third column plots. Comparison of the amplitudes
of the PSD peaks with the damping statistics in the second column indicates that modes
having large spectral amplitudes tend to have relatively small standard deviations, and vice
versa. For example, damping estimates for the fundamental mode at station 11 have a standard
deviation of 0.4% whereas at the tower it is 0.9% (see Table III). This is a consequence of
a low signal-to-noise ratio for the mode at the tower, as indicated by the PSD amplitude.
Another example at 1:07 Hz shows that the standard deviation in damping is smallest (0.2%)
at station 10 where the PSD amplitude is the greatest, and is greater (1.0%) at station 11
where the PSD amplitude is lower than at station 10. The standard deviation is largest (1.6%)
for this mode at station 5 where the PSD amplitude is very small (barely perceptible on the
PSD plot).
Although some of the standard deviations presented in the plots extend to zero, and some
of the individual mean damping estimates have values near zero, realistically there is a non-
zero lower bound on the energy dissipation level for each mode. Had a oor been placed on
the acceptable damping levels during the analysis we would expect a slightly lower variation
in the estimates. Nevertheless, the approach taken in this paper accounts for variations in
the estimation process and supports the notion that although individual damping estimates can
Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:787805
REASSESSMENT OF DYNAMIC CHARACTERISTICS: QUINCY BAYVIEW BRIDGE 801
extend over a large range, multiple identications provide statistical condence in the damping
ratios based on the means and standard deviations.
5.1.2. Torsional and transverse response. Figure 4 illustrates a total of 13 modal frequencies
ranging from 0.561:92 Hz that were identied in the torsional response signals at the four
locations. A total of fourteen frequencies, ranging from 0.561:82 Hz, were identied from the
transverse signals as shown in Figure 5. With only one exception, the mean damping estimates
are all less than 2%. The standard deviations are generally similar to those obtained for the
vertical modes and the high density of the clusters at station 11 is once again noticeable.
Summary statistics on a station-by-station basis are given in Table II. The mean and standard
deviation of the estimated damping of torsional motions identied across all modes and all
stations is 1:0 0:8%. The value for transverse motions across all modes and all stations
is 0:9 0:8%. Note that for Table II the resulting statistics for modes that appear in both
torsional and transverse identications have not been combined.
The rst mode to occur with any signicant torsional component is the rst torsional
transverse mode at 0:56 Hz. Table III lists the estimated mean and standard deviation damping
of this mode for three cases: the identication results from the torsional signals (obtained
from vertical response measurements), the identication results from the transverse signals,
and the combined estimates for identications from both torsional and transverse signals.
Averaged across all stations, the identied mean and standard deviation damping for this
mode is 1:0 0:8%, 1:2 0:8%, and 1:1 0:8% for the torsional, transverse, and combined
identications, respectively. The estimates from torsional and transverse identications are
in strong agreement with one another. At station 11, where a total of 373 estimates were
identied, the result is 0:9 0:7%. Although the torsional component of this mode is barely
perceptible in the PSD in Figure 4, more than 300 identications were possible by the EM
SSI technique using the torsional signals (the mode is more evident in the PSD from torsional
motions at the centre of the bridge, but the data length at this location is short and it was not
used for identications). The transverse component of the 0:56 Hz coupled torsionaltransverse
mode is clearly evident in the PSD plots for stations 10 and 11 in Figure 5.
Comparison of the middle and right columns of Figure 4 again indicates that lower standard
deviations are generally associated with higher PSD levels. A similar result appears in Figure 5
for the transverse motion estimates. An exception to this observation occurs for the closely
spaced modes near 1:4 Hz in Figure 4. Although the EMSSI procedure is able to separate
closely spaced modes in this region, the estimated standard deviations are rather large (at
1:40 Hz the estimate is 1:5 0:9%). It appears that the close spacing of the modes in this
region has added to the uncertainty in the identication. At station 5 where the signal level
is low the standard deviation is 2%, the largest value found in the entire study. The smallest
standard deviation on damping estimated in the study is for the 1:71 Hz torsional mode at
station 10, having a value of 0.05%. The largest peak in the PSD for torsional motion occurs
at 1:40 Hz (station 10 in Figure 4), where the estimate is 0:7 0:6%. The largest peak in the
PSD for transverse motion occurs at 1:68 Hz (station 10 in Figure 5), where the estimate is
1:1 0:7%.
For those frequencies identied at multiple locations, their respective damping estimates
once again have mean values that are within one standard deviation of one another and
there are no signicant outliers. For several of the common mode identications, the damping
identications provide almost equal values at all deck locations. Two clear examples of this
Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:787805
802 B. A. PRIDHAM AND J. C. WILSON
are the modes at 0:89 Hz and 1:18 Hz in Figure 4 that have mean damping values of 0.9%
and 0.8%, and standard deviations less than 0.5% and 0.2%, respectively.
For the transverse motion identications in Figure 5, the noise level in the 1.62 Hz range
appears higher than observed in either the vertical or torsional signals. Only ve frequencies
passed the screening process for estimates at the tower (note the absence of the 0:56 Hz mode
in Table III at this location). As an experiment, when the acceptance tolerance (bandwidth
of frequency acceptance) was relaxed from a 50% to a 30% occurrence, several more modes
appeared in the scatter plots. This is indicative of the sporadic nature of estimation of these
very weak transverse signals from the tower.
6. CONCLUSIONS
Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:787805
REASSESSMENT OF DYNAMIC CHARACTERISTICS: QUINCY BAYVIEW BRIDGE 803
0.6
0.4
0.2
0.8
Station 10
RMS Acceleration (10-3 cm/s2)
0.6
0.4
0.2
0.8
Station 11
RMS Acceleration (10-3 cm/s2)
0.6
0.4
0.2
0.8
Tower
cm/s )
2
0.6
-3
RMS Acceleration (10
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Mean Estimated Frequency (Hz) Mean Estimated Frequency (Hz) Mean Estimated Frequency (Hz)
one standard deviation (s ) modal RMS acceleration bounds for stations (top to
Figure 6. Mean ()
bottom) 5; 10; 11 and the tower, (left to right) vertical, torsional and transverse motions.
Important results from the analysis of the Quincy bridge data include:
1. Mean damping ratios identied for all modes in the vertical, torsional and transverse
directions are 1.0%, 0.9% and 0.9% respectively, with a standard deviation of 0.8% in
each case. The statistics indicate that at least 90% of all EMSSI analyses using the
Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:787805
804 B. A. PRIDHAM AND J. C. WILSON
Quincy AVS data identify damping at less than 2%. The lowest mean damping level
identied was 0:1%.
2. Mean () one standard deviation (s ) damping estimates for the rst vertical (0:37 Hz),
and rst coupled torsionaltransverse (0:56 Hz) modes were estimated as 1:5 0:5%, and
1:1 0:8%, respectively.
3. Mean damping values for a mode estimated at multiple spatial locations for a given
direction of motion lie within the maximum standard deviation for that mode. This
indicates a degree of consistency in the estimated values even for those modes weakly
excited at a given location.
4. For those modes with signicant PSD peaks (i.e. the important modes), the mean
damping values are essentially the same at each location, and the variations are small.
This behaviour was observed for several of the modes at stations 10 and 11 on the main
span.
5. The standard deviation in estimated damping ranged from 0.05% (1:7 Hz torsional mode
at station 10) to 2% (1:4 Hz torsional mode at station 5).
6. Some of the frequencies originally described in Reference [1] as being associated with
purely torsional, vertical and transverse modes, appeared in the identication results in
all three directions of motion. This is indicative of a more complex behaviour of the
bridge in some modes than that explained during the original assessment.
7. The mean modal RMS accelerations obtained from the multiple identications match
closely with the relative spectral magnitudes for the spectra at a given location. In this
study, no signicant correlations were found to exist between modal RMS acceleration
levels and the mean damping estimates.
ACKNOWLEDGEMENTS
The research grant and scholarship support of the Natural Sciences and Engineering Research Council
of Canada is gratefully acknowledged as is the scholarship support of the Ontario Ministry of Training,
Colleges and Universities. The authors would also like to thank the reviewers of the paper for their
many suggestions, which aided the improvement of the nal manuscript.
REFERENCES
1. Wilson JC, Liu T. Ambient vibration measurements on a cable-stayed bridge. Earthquake Engineering and
Structural Dynamics 1991; 20:723748.
2. Pridham BA, Wilson JC. Identication of base-excited structures using output-only parameter estimation.
Earthquake Engineering and Structural Dynamics 2004; 33:133155. DOI: 10.1002=eqe.343.
3. Modjeski and Masters. Structural Drawings of Quincy Bayview Bridge. Modjeski and Masters Consulting
Engineers, Harrisburg, Pennsylvania, 1983.
4. Wilson JC, Gravelle W. Modelling of a cable-stayed bridge for dynamic analysis. Earthquake Engineering and
Structural Dynamics 1991; 20:707722.
5. Ghahramani Z, Hinton GE. Parameter estimation for linear dynamical systems. Technical Report CRG-TR-96-2,
University of Toronto: http://www.gatsby.ucl.ac.uk=zoubin=papers.html [10 March 2002].
6. Van Overshcee P, De Moor B. Subspace Identication for Linear Systems: TheoryImplementation
Application. Kluwer Academic Publishers: Dordrecht, 1996.
7. Larimore WE. System identication reduced order ltering and modeling via canonical variate analysis.
Proceedings of the American Control Conference, 1983.
8. Pridham BA, Wilson JC. An application example illustrating the practical issues of subspace identication.
Proceedings of the 21st International Modal Analysis Conference, Kissimmee, Florida, 2003.
9. Peeters B. System identication and damage detection in civil engineering. Ph.D. Dissertation, Katholieke
Universiteit Leuven, Belgium, 2000.
Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:787805
REASSESSMENT OF DYNAMIC CHARACTERISTICS: QUINCY BAYVIEW BRIDGE 805
10. Verboven P, Cauberghe B, Parloo E, Vanlanduit S, Guillaume P. User-assisting tools for a fast frequency-domain
modal parameter estimation method. Mechanical Systems and Signal Processing 2004; 18(4):759780. DOI.
10.1016=S0888-3270(03)00053-0.
11. Zadeh LA. Fuzzy logic toolbox for use with MATLAB. The Mathworks Inc., Nantick, MA, 1995.
12. Pridham BA. State space modeling and identication of stochastic linear structural systems. Ph.D. Thesis,
McMaster University, Hamilton, Ontario, Canada, 2004.
13. Zhang QW, Fan LC, Yuan WC. Trac-induced variability in dynamic properties of a cable-stayed bridge.
Earthquake Engineering and Structural Dynamics 2002; 31:20152021.
Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:787805