8 Experimental Case No.

In this chapter, the experimental case no. 1, which was introduced in section 1.7.1 is
described. This experimental case is based on the two-dimensional plane five storey
frame structure illustrated in figure 1.2, and it is described how this structure is
modelled and how the structural response is simulated. Finally, the simulation results
are presented and analyzed. The analysis will concern the asymptotic properties of
the PEM estimator armav.m of the STDI toolbox which is described in section 7.3.
The asymptotic properties are the bias and the standard deviations of the modal
parameters obtained from the estimated ARMAV models. These properties will be
analysed as a function of record length and noise level. Further, the sampled standard
deviations of the modal parameters will be compared with the estimated standard
deviations obtained on the basis of the PEM estimate.

8.1 A Simulation Study of a Five Storey Structure

In this section, it will be shown how the structure is modelled in continuous time and
converted to an equivalent discrete-time model. Finally, the modal properties of the
structure are presented.

8.1.1 Modelling of the Structure

All mass of the plane frame structure is assumed to be concentrated in the horizontal
beam elements of the structure, which implies that the centre of each of the beams is
a mass-point. These elements are also assumed to be completely stiff, which implies
that only horizontal displacements will be considered.

The horizontal motion of all mass-points is assumed to be observed and the five
bending modes will be investigated. This implies that the displacements of the five
storeys are described by the following second-order differential equation system, see
section 3.1

z̈( t ) % M&1 C z( &1 &1

0 t ) % M K z( t ) ' M f( t ) (1)

All system matrices have the same dimensions, and the structure is assumed to be
excited at all mass-points by forces assembled in the vector f(t). The mass matrix is
chosen as the identity matrix. A non-proportional damping matrix is assumed. This
damping matrix is selected in such a way that the modes of the structure will be
underdamped, with damping ratios around a few per cent. The mass-normalized
damping and stiffness matrices are given by

Experimental Case No. 1 151

 2.41 &2.40 0 0 0
&2.40 4.81 &2.40 0 0
M&1C ' 100 sec&1 × 0 &2.40 4.81 &2.40 0 (2)
0 0 &2.40 4.81 &2.40
0 0 0 &2.40 3.01

4.80 &4.80 0 0 0
&4.80 9.60 &4.80 0 0
M&1K ' 103
sec&2 × 0 &4.80 9.60 &4.80 0 (3)
0 0 &4.80 9.60 &4.80
0 0 0 &4.80 6.00

The excitation of the system is assumed to be a continuous-time Gaussian white

noise, described by the intensity matrix W = I.

8.1.2 Modal Properties of the Structure

The performance of the PEM routine armav.m will be tested on its ability to estimate
the modal parameters. Therefore, the true modal parameters of the model will be
presented in this section. These parameters are obtained by applying the techniques
given in section 3.1 to (8.1). The modal parameters used in the analysis are:

L Natural eigenfrequencies.
L Damping ratios.
L Scaled mode shapes.

The mode shapes will be scaled so that they are unity at the first floor. This is a
necessary step, in order to make sure that the numerical differentiation performed
works properly, when the standard deviations of the mode shape coordinates are
estimated. In this way the default scaling used by the MATLAB eigenvalue solver is
interrupted.

The natural eigenfrequencies and damping ratios are shown in table 8.1. In figure 8.1
the magnitudes of the mode shapes are plotted. In table 8.2 and table 8.3 the
magnitudes and phase angels are listed.

152 A Simulation Study of a Five Storey Structure

 Mode No. 1 2 3 4 5
fi [Hz] 2.141 7.590 13.292 17.970 21.003
.i [%] 1.24 1.34 2.61 2.71 3.31
Table 8.1: Natural eigenfrequencies and damping ratios.

Mode #1 Mode #2 Mode #3 Mode #4 Mode #5

5 5 5 5 5

4 4 4 4 4
Storey / Coordinate Number

3 3 3 3 3

2 2 2 2 2

1 1 1 1 1

0 0 0 0 0
0 1 2 -1 0 1 -2 0 2 -2 0 2 -5 0 5
f =2.141 Hz f =7.59 Hz f =13.29 Hz f =17.97 Hz f =21 Hz

Figure 8.1: Magnitudes of the normalized complex mode shapes.

Storey 1. mode 2. mode 3. mode 4. mode 5. mode

No. Magnitude Magnitude Magnitude Magnitude Magnitude
5 1.55 0.94 0.89 0.88 0.88
4 1.49 0.50 0.41 1.46 2.30
3 1.38 0.19 1.11 0.09 2.88
2 1.21 0.78 0.21 1.41 2.38
1 1 1 1 1 1
Table 8.2: Magnitudes of normalized complex mode shapes.

Experimental Case No. 1 153

 Storey 1. mode 2. mode 3. mode 4. mode 5. mode
No. Phase [E] Phase [E] Phase [E] Phase [E] Phase [E]
5 -0.96 -179.72 1.88 -177.01 4.55
4 -0.58 -176.91 176.67 -0.33 -178.02
3 -0.46 -8.01 -179.43 154.89 2.26
2 -0.16 -0.83 -170.98 -179.28 -179.32
1 0 0 0 0 0
Table 8.3: Phase angels of normalized complex mode shapes.

Since all the modes are well separated and since the phase angles are close to 0E and
±180E, the structural modes are close to being normal.

8.2 Simulating the System Response

This section describes how the system response is simulated and how disturbance is
added. It is also explained how the ambient excitation of the plane frame structure is
modelled and how to obtain a simulation model.

8.2.1 Discrete-Time Modelling and Addition of Disturbance

The process noise w(tk) is assumed to be zero, and four different levels of the
measurement noise v(tk) are applied. The covariance matrix of the measurement noise
is calculated as
p 2
R ' K
1
Eii(0) I (4)
p
j

i'1

where p is the number of output channels, and Eii(0) correponds to the ith diagonal
element of the sampled zero-lag covariance matrix of the noise-free output. K is a
constant that controls the noise level and can be considered as a signal-to-noise ratio.
In table 8.3 the values of K used in the analysis are listed.

No. 1 2 3 4
K 0.001 0.01 0.05 0.1
Table 8.3: The values of K to be used.

As seen the noise covariance matrix is defined as a diagonal matrix of the mean-
square of the sampled standard deviations of the noise-free output multiplied by the
signal-to-noise ratio K.

154 Simulating the System Response

To avoid numerical integration of the structural system (8.1), which could be
accomplished by a Runge-Kutta integration scheme, the continuous-time system is
converted to a covariance equivalent discrete-time ARMAV(2,1) model. This
conversion is performed using the technique described in theorem 4.2 in section
4.2.1. The sampling interval is T = 0.02 sec. The results of the conversion are two
auto-regressive coefficient matrices together with one of 252 solutions of the moving
average coefficient matrix and the associated covariance matrix of the zero-mean
Gaussian white noise input process.

The auto-regressive coefficient matrices are given by

&0.620 &1.073 &0.253 &0.016 &0.002

&1.137 0.284 &0.925 &0.213 &0.022
A1 ' &0.242 &0.889 0.269 &0.889 &0.237 (5)
&0.022 &0.214 &0.920 0.282 &1.071
&0.001 &0.020 &0.236 &1.070 &0.293

0.892 0.079 &0.011 0.004 &0.001

0.036 0.941 0.030 0.007 &0.001
A2 ' &0.001 0.056 0.884 0.056 &0.002 (6)
&0.001 0.005 0.033 0.933 0.044
0 0.001 0 0.046 0.962

The moving average coefficient matrix is given by

0.365 &0.138 0.080 &0.061 0.026

&0.194 0.712 &0.426 0.278 &0.112
B2 ' 0.111 &0.367 0.782 &0.379 0.130 (7)
&0.097 0.272 &0.430 0.724 &0.209
0.038 &0.104 0.142 &0.204 0.415

and its corresponding Gaussian white noise covariance matrix is given by

0.322 0.131 0.023 0.007 &0.001

0.131 0.229 0.120 0.009 0.009
)' 10&5 × 0.023 0.120 0.215 0.121 0.019 (8)
0.007 0.009 0.121 0.227 0.129
&0.001 0.009 0.019 0.129 0.306

Experimental Case No. 1 155

The noise can be incorporated into the system by using theorem 4.3 to yield a new
ARMAV(2,2) model. However, since only measurement noise is added, it is easier
to add the noise to the simulated noise-free output.

The final simulation model is obtained by conversion of the ARMAV(2,1) model to

a state space realization using e.g. the techniques described in theorem 2.4 in section
2.5.1. The measurement noise is then added to yield
x( tk%1 ) ' A x( tk ) % B u( tk ) , u( t k ) 0 NID( 0 , ) )
(9)
y( tk%1 ) ' C x( tk ) % v( tk ) , v( t k ) 0 NID( 0 , R )

The response is then obtained by simulating the outcome of the p-variate Gaussian
white noise processes u(tk) and v(tk), for k = 1 to N.

8.2.2 Organizing the Simulations

In addition to the use of four different values of the signal-to-noise parameter K four
different record lengths N are used. This means that the number of simulation cases
is 16. In each of these cases, 100 simulations and identifications have been
performed. This implies that the total number of simulations and identifications is
1600.

The organization of the simulation study is shown in table 8.4.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 run11i run12i run13i run14i
N = 2500 run21i run22i run23i run24i
N = 5000 run31i run32i run33i run34i
N = 10000 run41i run42i run43i run44i
Table 8.4: Organization of the simulations. The integer i takes values between 1 and
100.

The simulations and identifications are performed using the STDI toolbox, described
in section 7.3. The conversion from the continuous-time differential eqaution system
to the ARMAV(2,1) model is performed by the routine armav21.m. The noise-free
simulated system output is calculated by the routine ddssim.m. And finally, the PEM
identification of an adequate ARMAV model is performed by the routine armav.m.
This routine implements the robustified Gauss-Newton search scheme for the PEM
estimator. This Gauss-Newton search scheme is given in definition 5.1 in section
5.3.1, and the robustification procedure is described in section 7.1.4.

156 Simulating the System Response

8.3 The Results of the Simulation Study
In this section, the results of the simulation study will be presented. Since the
simulations are performed using an ARMAV(2,1) model and since Gaussian white
noise has been added, an adequate model structure is an ARMAV(2,2) model
according to section 4.3. In figure 8.2, this is verified by plotting the FPE criterion
for different identified ARMAV models of the simulation case run441, see table 8.4.

-8 Akaike´s Final Prediction Error Criterion

x 10
2.5

1.5
FPE

0.5

0
ARMAV(1,1) ARMAV(2,1) ARMAV(2,2) ARMAV(3,2) ARMAV(3,3)

Figure 8.2: Akaike’s FPE criterion plotted for the following choices of model
structure. ARMAV(1,1), ARMAV(2,1), ARMAV(2,2), ARMAV(3,2)
and ARMAV(3,3).

This model structure has been used in all 1600 simulations. In all cases the maximum
number of iterations used in the PEM algorithm is 20. If this number is exceeded or
if the RMS value of the search gradient, given by RN-1(2)FN(2) in definition 5.1, is
less than 1.0 × 10-5 the iterations are stopped. The robustification parameter D that is
used to remove outliers is set as 1.6, see section 7.1.4.

8.3.1 Adequacy of Chosen Number of Simulations

Besides checking the adequacy of the model structure, it is necessary to verify that
the statistical properties of the estimates converge within the 100 simulations of each
run. If it can be verified that the statistical properties converge in the worst case, then
it will be assumed that the statistical properties in all the other cases also converge.
The worst case is run14, since its record length is the shortest and the noise level is
the highest.

Experimental Case No. 1 157

The convergence check of the run14 simulations is performed by plotting the sampled
standard deviation F6(m) of the modal parameter 6 as a function of the number of
simulations m. In other words, plotting the function

m
1
F6( m ) ' 6̂( k ) & 60 2
, m ' 2, 3, ... , 100 (10)
m
j

k'1

The parameter 60 corresponds to the true modal parameter of the system being
simulated. By using this parameter the sampled standard deviation will be unbiased.
The modal parameter 6̂( k ) is either the estimate of the kth simulation of one of the
five natural eigenfrequencies, or one of the associated damping ratios, or one of the
normalized mode shape coordinates of one of the five modes.

The change of the unbiased standard deviation defined in (8.10) of the five estimated
natural eigenfrequencies of run14 is plotted in figure 8.3.

Mode #1 Mode #2
Standard Deviation

Standard Deviation

0.01 0.01

0.005 0.005

0 0
0 50 100 0 50 100
Simulation Number - m Simulation Number - m
Mode #3 Mode #4
Standard Deviation

Standard Deviation

0.01 0.01

0.005 0.005

0 0
0 50 100 0 50 100
Simulation Number - m Simulation Number - m
Mode #5
Standard Deviation

0.1

0.05

0
0 50 100
Simulation Number - m

Figure 8.3: Change of the standard deviation of the five natural eigenfrequency
estimates as a function of the number of simulations for run14.

After 50 simulations the changes of the standard deviations are beginning to saturate
and at 100 simulations they have almost converged to constant values. It is therefore
assumed that the chosen number of simulations with regard to the natural
eigenfrequencies is adequate. The estimated damping ratios of the five modes will be
analysed in a similar manner. The changes of the standard deviation of the five
estimated damping ratios of run14 are plotted in figure 8.4.

158 The Results of the Simulation Study

 Mode #1 Mode #2

Standard Deviation

Standard Deviation
1 1

0.5 0.5

0 0
0 50 100 0 50 100
Simulation Number - m Simulation Number - m
Mode #3 Mode #4
Standard Deviation

Standard Deviation
1 1

0.5 0.5

0 0
0 50 100 0 50 100
Simulation Number - m Simulation Number - m
Mode #5
Standard Deviation

0
0 50 100
Simulation Number - m

Figure 8.4: Changes of the standard deviation of the five damping ratio estimates
as a function of the number of simulations for run14.

Again, the standard deviations seem to converge to constant values, which implies
that the chosen number of simulations with regard to the damping ratios is adequate.

The next five plots show the absolute change of the standard deviation of the scaled
mode shapes of the five modes. Each of the mode shapes has been normalized with
respect to their first coordinate. Due to this normalization this coordinate will always
be one and therefore have the standard deviation zero.

In figures 8.5 to 8.9, the changes of the standard deviation of the remaining mode
shape coordinates are plotted for each of the five modes.

Experimental Case No. 1 159

 Mode #1, Coordinate #5 Mode #1, Coordinate #4
0.015 0.015
Standard Deviation

Standard Deviation
0.01 0.01

0.005 0.005

0 0
0 50 100 0 50 100
Simulation Number - m Simulation Number - m
Mode #1, Coordinate #3 Mode #1, Coordinate #2
0.015 0.015
Standard Deviation

Standard Deviation
0.01 0.01

0.005 0.005

0 0
0 50 100 0 50 100
Simulation Number - m Simulation Number - m

Figure 8.5: Changes of the standard deviation of the four mode shape coordinate
estimates of the first mode as a function of the number of simulations
for run14.

Mode #2, Coordinate #5 Mode #2, Coordinate #4

0.2 0.2
Standard Deviation

Standard Deviation

0.15 0.15

0.1 0.1

0.05 0.05

0 0
0 50 100 0 50 100
Simulation Number - m Simulation Number - m
Mode #2, Coordinate #3 Mode #2, Coordinate #2
0.2 0.2
Standard Deviation

Standard Deviation

0.15 0.15

0.1 0.1

0.05 0.05

0 0
0 50 100 0 50 100
Simulation Number - m Simulation Number - m

Figure 8.6: Changes of the standard deviation of the four mode shape coordinate
estimates of the second mode as a function of the number of simulations
for run14.

160 The Results of the Simulation Study

 Mode #3, Coordinate #5 Mode #3, Coordinate #4
1 1

Standard Deviation

Standard Deviation
0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 50 100 0 50 100
Simulation Number - m Simulation Number - m
Mode #3, Coordinate #3 Mode #3, Coordinate #2
1 1
Standard Deviation

Standard Deviation
0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 50 100 0 50 100
Simulation Number - m Simulation Number - m

Figure 8.7: Changes of the standard deviation of the four mode shape coordinate
estimates of the third mode as a function of the number of simulations
for run14.

Mode #4, Coordinate #5 Mode #4, Coordinate #4

4 4
Standard Deviation

Standard Deviation

3 3

2 2

1 1

0 0
0 50 100 0 50 100
Simulation Number - m Simulation Number - m
Mode #4, Coordinate #3 Mode #4, Coordinate #2
4 4
Standard Deviation

Standard Deviation

3 3

2 2

1 1

0 0
0 50 100 0 50 100
Simulation Number - m Simulation Number - m

Figure 8.8: Changes of the standard deviation of the four mode shape coordinate
estimates of the fourth mode as a function of the number of simulations
for run14.

Experimental Case No. 1 161

 Mode #5, Coordinate #5 Mode #5, Coordinate #4
2 2

Standard Deviation

Standard Deviation
1.5 1.5

1 1

0.5 0.5

0 0
0 50 100 0 50 100
Simulation Number - m Simulation Number - m
Mode #5, Coordinate #3 Mode #5, Coordinate #2
2 2
Standard Deviation

Standard Deviation
1.5 1.5

1 1

0.5 0.5

0 0
0 50 100 0 50 100
Simulation Number - m Simulation Number - m

Figure 8.9: Changes of the standard deviation of the four mode shape coordinate
estimates of the fifth mode as a function of the number of simulations
for run14.

The standard deviations of the mode shape coordinates also seem to converge to
constant values in all cases.

So in conclusion :

L In case of run14, which is assumed to be the worst of the 16 cases presented

in table 8.4, it has been shown that 100 simulations are enough to obtain
convergent statistical properties of the estimated modal parameters. On the
basis of this observation, it is therefore assumed that 100 simulation will
also be enough in all the other cases.

In the following the number of simulations m will therefore be a constant parameter,

defined as m = 100.

8.3.2 Bias of Estimated Modal Parameters

In section 5.5.2, it was stated that the estimated model parameters will be asymptoti-
cally unbiased if the true system is contained in the model structure and if the
prediction errors are Gaussian white noise. Due to the analytical relation between the
model and modal parameters these requirements also result in asymptotically
unbiased modal parameters.

162 The Results of the Simulation Study

Therefore, to be able to say that the estimated modal parameters are asymptotically
unbiased, the following remains to be verified :

L Adequacy of the model structure.

L Prediction errors are Gaussian white noise.

At the beginning of this section, it was verified that the applied ARMAV(2,2) model
was adequate, which means that the true system can be assumed contained in the
model. Since the excitation and the applied noise in all the simulations are Gaussian
distributed realizations, verification remains that the prediction errors of the
identified models are white noise realizations.

Since the unbiasness is an asymptotical property of the PEM estimator the following
analysis is performed for the simulation case run41, which has the maximum record
length. To prove whether the prediction errors are white noise or not, estimates of the
autocorrelation function can be plotted. The estimated autocorrelation function for
a white noise process should be approximately zero except at the zero lag. Further,
since the realizations of the excitation and the applied noise originate from the same
stochastic excitation and noise processes, any numerical inaccuracies can be reduced
by averaging of the estimated autocovariance of the prediction errors obtained from
each of the 100 simulations.

In figure 8.10 the averaged autocorrelation functions of the five channels obtained
from run41 are plotted.

Autocorrelation Channel #1 Autocorrelation Channel #2

1 1
Correlation

Correlation

0.5 0.5

0 0

0 5 10 15 20 0 5 10 15 20
Time Lag [Sec] Time Lag [Sec]
Autocorrelation Channel #3 Autocorrelation Channel #4
1 1
Correlation

Correlation

0.5 0.5

0 0

0 5 10 15 20 0 5 10 15 20
Time Lag [Sec] Time Lag [Sec]
Autocorrelation Channel #5
1
Correlation

0.5

0 5 10 15 20
Time Lag [Sec]

Figure 8.10: Averaged autocorrelation functions of run41.

Experimental Case No. 1 163

The autocorrelation at the non-zero time-lags are seen to be close to zero. This
implies that the prediction errors forms a realization of a multivariate white noise
process.

So in conclusion :

L The above analysis indicates that if the true system is contained in the
model, it is possible to obtain prediction errors that are white noise.

To visualize how the bias changes as a function of N and K the following bias
measures are plotted for all 16 simulation cases.

1
m 6̂( k ) & 6 6 ' fi for i ' 1, 2, ... , 5
$6 ' 0
,
m k'1
j

60 6 ' .i for i ' 1, 2, ... , 5

(11)
m i ' 1, 2, ... , 5
6 ' Mij for
1
$6 ' 6̂( k ) & 60 ,
j ' 2, ... , 5
/ /

m k'1
j
0 0

0 0

0 0

The first bias measure is applied for the natural eigenfrequency and the associated
damping ratio, whereas the last bias measure is applied for each of the mode shape
coordinates Mij where i indicate the mode and j the coordinate number according to
figure 8.1.

In figures 8.11 to 8.13 the first bias measure of (8.11) is plotted as a function of the
record length N and the noise level K.

164 The Results of the Simulation Study

 Bias of Eigenfrequency Estimates of Mode #1 Bias of Damping Ratio Estimates of Mode #1

-3
x 10

2 0.2

0.15
1
0.1

Bias

Bias
0.05
0
0.1 0.1
0

-1 0.05 -0.05 0.05

2000 4000 2000 4000
6000 8000 6000 8000
10000 0 Noise Level, K 10000 0 Noise Level, K
Record Length, N Record Length, N

Bias of Eigenfrequency Estimates of Mode #2 Bias of Damping Ratio Estimates of Mode #2

-3
x 10

2 0.2

0.15
1
0.1
Bias

Bias
0.05
0
0.1 0.1
0

-1 0.05 -0.05 0.05

2000 4000 2000 4000
6000 8000 6000 8000
10000 0 Noise Level, K 10000 0 Noise Level, K
Record Length, N Record Length, N

Figure 8.11: Estimated bias of the first and second natural eigenfrequencies and
associated damping ratios as a function of record length N and noise
level K.
Bias of Eigenfrequency Estimates of Mode #3 Bias of Damping Ratio Estimates of Mode #3

-3
x 10

2 0.2

0.15
1
0.1
Bias

Bias

0.05
0
0.1 0.1
0

-1 0.05 -0.05 0.05

2000 4000 2000 4000
6000 8000 6000 8000
10000 0 Noise Level, K 10000 0 Noise Level, K
Record Length, N Record Length, N

Bias of Eigenfrequency Estimates of Mode #4 Bias of Damping Ratio Estimates of Mode #4

-3
x 10

2 0.2

0.15
1
0.1
Bias

Bias

0.05
0
0.1 0.1
0

-1 0.05 -0.05 0.05

2000 4000 2000 4000
6000 8000 6000 8000
10000 0 Noise Level, K 10000 0 Noise Level, K
Record Length, N Record Length, N

Figure 8.12: Estimated bias of the third and fourth natural eigenfrequencies and
associated damping ratios as a function of record length N and noise
level K.

Experimental Case No. 1 165

 Bias of Eigenfrequency Estimates of Mode #5 Bias of Damping Ratio Estimates of Mode #5

0 0.8

0.6
-0.01
0.4
Bias

Bias
0.2
-0.02
0.1 0.1
0

-0.03 0.05 -0.2 0.05

2000 4000 2000 4000
6000 8000 6000 8000
10000 0 Noise Level, K 10000 0 Noise Level, K
Record Length, N Record Length, N

Figure 8.13: Estimated bias of the fifth natural eigenfrequency and associated
damping ratio as a function of record length N and noise level K.

In figures 8.14 to 8.18 the bias of each of the mode shapes is plotted coordinate-wise
as a function of the record length and the noise level.

Bias of Mode Shape Estimates of Mode #1 - Coor #5 Bias of Mode Shape Estimates of Mode #1 - Coor #4

-3 -3
x 10 x 10

1 1

0.8 0.8

0.6 0.6
Bias

Bias

0.4 0.4
0.1 0.1
0.2 0.2

0 0.05 0 0.05
2000 4000 2000 4000
6000 8000 6000 8000
10000 0 Noise Parameter, K 10000 0 Noise Parameter, K
Record Length, N Record Length, N

Bias of Mode Shape Estimates of Mode #1 - Coor #3 Bias of Mode Shape Estimates of Mode #1 - Coor #2

-3 -3
x 10 x 10

1 1

0.8 0.8

0.6 0.6
Bias

Bias

0.4 0.4
0.1 0.1
0.2 0.2

0 0.05 0 0.05
2000 4000 2000 4000
6000 8000 6000 8000
10000 0 Noise Parameter, K 10000 0 Noise Parameter, K
Record Length, N Record Length, N

Figure 8.14: Estimated bias of the mode shape coordinates of the first mode as a
function of record length N and noise level K.

166 The Results of the Simulation Study

 Bias of Mode Shape Estimates of Mode #2 - Coor #5 Bias of Mode Shape Estimates of Mode #2 - Coor #4

-3 -3
x 10 x 10

8 8

6 6

Bias

Bias
4 4

2 0.1 2 0.1

0 0.05 0 0.05
2000 4000 2000 4000
6000 8000 6000 8000
10000 0 Noise Parameter, K 10000 0 Noise Parameter, K
Record Length, N Record Length, N

Bias of Mode Shape Estimates of Mode #2 - Coor #3 Bias of Mode Shape Estimates of Mode #2 - Coor #2

-3 -3
x 10 x 10

8 8

6 6
Bias

Bias
4 4

2 0.1 2 0.1

0 0.05 0 0.05
2000 4000 2000 4000
6000 8000 6000 8000
10000 0 Noise Parameter, K 10000 0 Noise Parameter, K
Record Length, N Record Length, N

Figure 8.15: Estimated bias of the mode shape coordinates of the second mode as a
function of record length N and noise level K.

Bias of Mode Shape Estimates of Mode #3 - Coor #5 Bias of Mode Shape Estimates of Mode #3 - Coor #4

0.04 0.04

0.03 0.03
Bias

Bias

0.02 0.02

0.01 0.1 0.01 0.1

0 0.05 0 0.05
2000 4000 2000 4000
6000 8000 6000 8000
10000 0 Noise Parameter, K 10000 0 Noise Parameter, K
Record Length, N Record Length, N

Bias of Mode Shape Estimates of Mode #3 - Coor #3 Bias of Mode Shape Estimates of Mode #3 - Coor #2

0.04 0.04

0.03 0.03
Bias

Bias

0.02 0.02

0.01 0.1 0.01 0.1

0 0.05 0 0.05
2000 4000 2000 4000
6000 8000 6000 8000
10000 0 Noise Parameter, K 10000 0 Noise Parameter, K
Record Length, N Record Length, N

Figure 8.16: Estimated bias of the mode shape coordinates of the third mode as a
function of record length N and noise level K.

Experimental Case No. 1 167

 Bias of Mode Shape Estimates of Mode #4 - Coor #5 Bias of Mode Shape Estimates of Mode #4 - Coor #4

0.1 0.1

0.08 0.08

0.06 0.06
Bias

Bias
0.04 0.04
0.1 0.1
0.02 0.02

0 0.05 0 0.05
2000 4000 2000 4000
6000 8000 6000 8000
10000 0 Noise Parameter, K 10000 0 Noise Parameter, K
Record Length, N Record Length, N

Bias of Mode Shape Estimates of Mode #4 - Coor #3 Bias of Mode Shape Estimates of Mode #4 - Coor #2

0.1 0.1

0.08 0.08

0.06 0.06
Bias

Bias
0.04 0.04
0.1 0.1
0.02 0.02

0 0.05 0 0.05
2000 4000 2000 4000
6000 8000 6000 8000
10000 0 Noise Parameter, K 10000 0 Noise Parameter, K
Record Length, N Record Length, N

Figure 8.17: Estimated bias of the mode shape coordinates of the fourth mode as a
function of record length N and noise level K.

Bias of Mode Shape Estimates of Mode #5 - Coor #5 Bias of Mode Shape Estimates of Mode #5 - Coor #4

2 2

1.5 1.5
Bias

Bias

1 1

0.5 0.1 0.5 0.1

0 0.05 0 0.05
2000 4000 2000 4000
6000 8000 6000 8000
10000 0 Noise Parameter, K 10000 0 Noise Parameter, K
Record Length, N Record Length, N

Bias of Mode Shape Estimates of Mode #5 - Coor #3 Bias of Mode Shape Estimates of Mode #5 - Coor #2

2 2

1.5 1.5
Bias

Bias

1 1

0.5 0.1 0.5 0.1

0 0.05 0 0.05
2000 4000 2000 4000
6000 8000 6000 8000
10000 0 Noise Parameter, K 10000 0 Noise Parameter, K
Record Length, N Record Length, N

Figure 8.18: Estimated bias of the mode shape coordinates of the fifth mode as a
function of record length N and noise level K.

168 The Results of the Simulation Study

All figures presented in this section indicate that the bias increases rapidly as the
number of samples decreases. When the record length exceeds N = 5000 the decrease
of the bias of the natural eigenfrequencies and the damping ratios begin to saturate
for all levels of noise. However, the bias of the mode shapes is still descreasing at the
high noise levels. In general, the bias of the mode shapes is seen to be more
dependent on the presence of noise than on the record length as seen in figures 8.16,
8.17 and 8.18. This is in contrast to the bias of the natural eigenfrequencies and the
damping ratios, which in general is more dependent on the record length than on the
noise level as seen in figures 8.11 and 8.12.

In most of the cases, it is seen that the bias keeps descreasing as the number of
samples is increased. This indicates that the estimator in the present simulation study
will be asymptotically unbiased.

8.3.3 Sampled and Estimated Standard Deviations

In section 6.2, it was explained how the standard deviation of the estimated modal
parameters could be estimated on the basis of the approximative Hessian matrix. This
matrix could be obtained directly from a PEM algorithm that implements the Gauss-
Newton search procedure. In this section a comparison between the standard
deviations of the modal parameters obtained using this approach and the sampled
standard deviations will be made.

Since the true mean values 60 of the modal parameters are known in advance, the
following unbiased sampled coefficient of variation <k has been applied in the
comparison.

m
1
6̂( k ) & 60 2
m
j

k'1 (12)
<6 '
60

This coefficient of variation is preferred instead of the standard deviation itself. This
is because it is non-dimensional and as such makes comparison of the different modal
parameters easier. In the case of the mode shapes <k is complex. In this case the
magnitude of <k is used.

In each of the 16 simulation cases 100 simulations were performed. This implies that
the standard deviation of each of the modal parameters has been estimated 100 times.
To make use of all simulations these 100 estimates have been averaged in the
following analysis. These averaged standard deviations are also divided by the true
mean values in order to obtain the coefficient of variation. These estimated
coefficients of variation are denoted <ˆ 6 to distinguish these from the sampled
coefficients of variation <k.

Experimental Case No. 1 169

In the following figures 8.19 to 8.33 the sampled and estimated coefficients of
variation <k and <ˆ 6 of the estimated natural eigenfrequencies, damping ratios and
mode shape coordinates of the five modes will be presented. The left-hand sub-plots
will be the sampled coefficients of variation <k defined in (8.12), whereas the right-
hand sub-plots will be the coefficients of variation <ˆ 6 estimated from the Hessian
matrix.

Below each of the figures the normalized differences |<k- <ˆ 6 |/<k between the sampled
and estimated coefficients of variations of the 16 simulation cases will be listed in
tables.

170 The Results of the Simulation Study

 Sampled CV of Eigenfrequency Estimates of Mode #1 Estimated CV of Eigenfrequency Estimates of Mode #1

0.01 0.01

Coef. of Variation

Coef. of Variation
0.005 0.005

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Sampled CV of Damping Ratio Estimates of Mode #1 Estimated CV of Damping Ratio Estimates of Mode #1

1 1
Coef. of Variation

Coef. of Variation
0.5 0.5

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Figure 8.19: Sampled and estimated coefficients of variation for the estimates of the
natural eigenfrequency and associated damping ratios of the first mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 -1.0 × 10-1 -9.6 × 10-2 -1.2 × 10-1 -1.1 × 10-1
N = 2500 6.8 × 10-2 5.3 × 10-2 2.4 × 10-2 1.1 × 10-2
N = 5000 4.7 × 10-2 5.8 × 10-2 8.9 × 10-2 1.1 × 10-1
N = 10000 8.6 × 10-3 2.4 × 10-2 2.9 × 10-2 2.2 × 10-2
Table 8.5: Normalized differences of sampled and estimated coefficients of
variation of estimated natural eigenfrequencies of the first mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 6.7 × 10-2 6.5 × 10-2 6.2 × 10-2 3.1 × 10-2
N = 2500 1.4 × 10-1 1.5 × 10-1 1.5 × 10-1 1.5 × 10-1
N = 5000 3.6 × 10-2 2.8 × 10-2 3.9 × 10-2 7.6 × 10-2
N = 10000 -4.9 × 10-2 -4.6 × 10-2 -6.9 × 10-2 -6.4 × 10-2
Table 8.6: Normalized differences of sampled and estimated coefficients of
variation of estimated damping ratios of the first mode.

Experimental Case No. 1 171

 Sampled CV of Eigenfrequency Estimates of Mode #2 Estimated CV of Eigenfrequency Estimates of Mode #2

-3 -3
x 10 x 10

4 4
Coef. of Variation

Coef. of Variation
2 2

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Sampled CV of Damping Ratio Estimates of Mode #2 Estimated CV of Damping Ratio Estimates of Mode #2

0.4 0.4
Coef. of Variation

0.2 Coef. of Variation 0.2

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Figure 8.20: Sampled and estimated coefficients of variation for the estimates of the
natural eigenfrequency and associated damping ratios of the second
mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 1.3 × 10-2 1.4 × 10-2 -3.1 × 10-2 -1.5 × 10-2
N = 2500 -3.0 × 10-3 -2.0 × 10-4 9.6 × 10-3 -5.5 × 10-2
N = 5000 5.3 × 10-2 4.8 × 10-2 4.3 × 10-2 3.0 × 10-2
N = 10000 1.3 × 10-1 1.2 × 10-1 1.4 × 10-1 1.2 × 10-1
Table 8.7: Normalized differences of sampled and estimated coefficients of
variation of estimated natural eigenfrequencies of the second mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 5.5 × 10-2 6.3 × 10-2 9.6 × 10-2 9.3 × 10-2
N = 2500 -8.7 × 10-3 -1.0 × 10-2 6.2 × 10-2 -1.4 × 10-2
N = 5000 8.3 × 10-2 9.3 × 10-2 8.0 × 10-2 4.7 × 10-2
N = 10000 5.9 × 10-2 7.6 × 10-2 9.3 × 10-2 8.8 × 10-2
Table 8.8: Normalized differences of sampled and estimated coefficients of
variation of estimated damping ratios of the second mode.

172 The Results of the Simulation Study

 Sampled CV of Eigenfrequency Estimates of Mode #3 Estimated CV of Eigenfrequency Estimates of Mode #3

0.01 0.01

Coef. of Variation

Coef. of Variation
0.005 0.005

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Sampled CV of Damping Ratio Estimates of Mode #3 Estimated CV of Damping Ratio Estimates of Mode #3

0.4 0.4
Coef. of Variation

Coef. of Variation
0.2 0.2

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Figure 8.21: Sampled and estimated coefficients of variation for the estimates of the
natural eigenfrequency and associated damping ratios of the third mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 2.9 × 10-2 4.5 × 10-2 1.6 × 10-1 2.3 × 10-1
N = 2500 9.0 × 10-3 3.0 × 10-2 3.2 × 10-2 2.7 × 10-2
N = 5000 -2.9 × 10-2 -4.1 × 10-2 4.0 × 10-2 7.5 × 10-2
N = 10000 -3.0 × 10-3 -6.8 × 10-3 -1.2 × 10-2 7.0 × 10-3
Table 8.9: Normalized differences of sampled and estimated coefficients of
variation of estimated natural eigenfrequencies of the third mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 4.1 × 10-2 3.1 × 10-2 -1.1 × 10-2 2.9 × 10-2
N = 2500 -2.6 × 10-2 -3.4 × 10-2 5.3 × 10-2 3.1 × 10-2
N = 5000 -9.6 × 10-3 -5.0 × 10-4 -7.3 × 10-2 -4.6 × 10-2
N = 10000 1.2 × 10-1 1.5 × 10-1 1.4 × 10-1 1.5 × 10-1
Table 8.10: Normalized differences of sampled and estimated coefficients of
variation of estimated damping ratios of the third mode.

Experimental Case No. 1 173

 Sampled CV of Eigenfrequency Estimates of Mode #4 Estimated CV of Eigenfrequency Estimates of Mode #4

0.01 0.01
Coef. of Variation

Coef. of Variation
0.005 0.005

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Sampled CV of Damping Ratio Estimates of Mode #4 Estimated CV of Damping Ratio Estimates of Mode #4

0.4 0.4
Coef. of Variation

0.2 Coef. of Variation 0.2

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Figure 8.22: Sampled and estimated coefficients of variation for the estimates of the
natural eigenfrequency and associated damping ratios of the fourth
mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 9.0 × 10-3 5.4 × 10-2 1.1 × 10-2 1.1 × 10-1
N = 2500 3.8 × 10-2 4.6 × 10-2 -1.3 × 10-2 1.6 × 10-1
N = 5000 6.7 × 10-2 3.2 × 10-2 -3.3 × 10-2 5.7 × 10-2
N = 10000 2.5 × 10-2 4.1 × 10-2 7.6 × 10-2 9.4 × 10-2
Table 8.11: Normalized differences of sampled and estimated coefficients of
variation of estimated natural eigenfrequencies of the fourth mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 7.1 × 10-2 3.2 × 10-2 -7.2 × 10-2 1.8 × 10-1
N = 2500 9.8 × 10-2 1.2 × 10-1 1.7 × 10-1 2.8 × 10-1
N = 5000 -1.1 × 10-3 2.7 × 10-2 2.4 × 10-2 8.9 × 10-2
N = 10000 -3.5 × 10-3 1.5 × 10-3 -8.3 × 10-2 -6.7 × 10-2
Table 8.12: Normalized differences of sampled and estimated coefficients of
variation of estimated damping ratios of the fourth mode.

174 The Results of the Simulation Study

 Sampled CV of Eigenfrequency Estimates of Mode #5 Estimated CV of Eigenfrequency Estimates of Mode #5

0.08 0.08

Coef. of Variation

Coef. of Variation
0.06 0.06

0.04 0.04

0.02 0.1 0.02 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Sampled CV of Damping Ratio Estimates of Mode #5 Estimated CV of Damping Ratio Estimates of Mode #5

2 2
Coef. of Variation

Coef. of Variation
1 1

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Figure 8.23: Sampled and estimated coefficients of variation for the estimates of the
natural eigenfrequency and associated damping ratios of the fifth mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 5.9 × 10-2 1.1 × 10-1 1.6 × 10-1 8.1 × 10-1
N = 2500 -5.5 × 10-2 -5.2 × 10-2 1.3 × 10-1 8.0 × 10-1
N = 5000 9.8 × 10-2 1.2 × 10-1 1.3 × 10-1 1.5 × 10-1
N = 10000 5.2 × 10-2 3.4 × 10-2 9.2 × 10-2 2.2 × 10-1
Table 8.13: Normalized differences of sampled and estimated coefficients of
variation of estimated natural eigenfrequencies of the fifth mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 4.3 × 10-2 4.2 × 10-2 6.7 × 10-2 7.3 × 10-1
N = 2500 1.6 × 10-2 4.0 × 10-2 7.1 × 10-2 2.6 × 10-1
N = 5000 1.4 × 10-2 -1.2 × 10-3 5.6 × 10-2 2.0 × 10-1
N = 10000 2.4 × 10-3 -2.0 × 10-2 5.5 × 10-2 7.8 × 10-2
Table 8.14: Normalized differences of sampled and estimated coefficients of
variation of estimated damping ratios of the fifth mode.

Experimental Case No. 1 175

 Sampled CV of Mode Shape Estimates of Mode #1, Coor #5 Estimated CV of Mode Shape Estimates of Mode #1, Coor #5

0.015 0.015
Coef. of Variation

Coef. of Variation
0.01 0.01

0.005 0.005
0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Sampled CV of Mode Shape Estimates of Mode #1, Coor #4 Estimated CV of Mode Shape Estimates of Mode #1, Coor #4

0.015 0.015
Coef. of Variation

0.01 Coef. of Variation 0.01

0.005 0.005
0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Figure 8.24: Sampled and estimated coefficients of variation for the estimates of the
mode shape coordinates 5 and 4 of the first mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 -4.6 × 10-1 -4.6 × 10-1 -4.8 × 10-1 -5.2 × 10-1
N = 2500 -5.6 × 10-1 -5.2 × 10-1 -4.7 × 10-1 -5.0 × 10-1
N = 5000 -6.5 × 10-1 -6.5 × 10-1 -6.2 × 10-1 -6.2 × 10-1
N = 10000 -5.6 × 10-1 -5.7 × 10-1 -5.9 × 10-1 -6.0 × 10-1
Table 8.15: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 5 of the first mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 -4.3 × 10-1 -4.4 × 10-1 -4.9 × 10-1 -5.6 × 10-1
N = 2500 -5.3 × 10-1 -4.7 × 10-1 -3.7 × 10-1 -3.8 × 10-1
N = 5000 -5.5 × 10-1 -5.6 × 10-1 -5.0 × 10-1 -4.7 × 10-1
N = 10000 -4.9 × 10-1 -5.0 × 10-1 -5.5 × 10-1 -5.7 × 10-1
Table 8.16: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 4 of the first mode.

176 The Results of the Simulation Study

 Sampled CV of Mode Shape Estimates of Mode #1, Coor #3 Estimated CV of Mode Shape Estimates of Mode #1, Coor #3

0.015 0.015

Coef. of Variation

Coef. of Variation
0.01 0.01

0.005 0.005
0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Sampled CV of Mode Shape Estimates of Mode #1, Coor #2 Estimated CV of Mode Shape Estimates of Mode #1, Coor #2

0.01 0.01
Coef. of Variation

Coef. of Variation
0.005 0.005

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Figure 8.25: Sampled and estimated coefficients of variation for the estimates of the
mode shape coordinates 3 and 2 of the first mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 -3.5 × 10-1 -3.3 × 10-1 -3.6 × 10-1 -4.0 × 10-1
N = 2500 -4.0 × 10-1 -3.7 × 10-1 -3.7 × 10-1 -3.8 × 10-1
N = 5000 -4.0 × 10-1 -3.9 × 10-1 -3.2 × 10-1 -3.1 × 10-1
N = 10000 -3.8 × 10-1 -4.0 × 10-1 -5.2 × 10-1 -5.4 × 10-1
Table 8.17: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 3 of the first mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 -2.4 × 10-1 -2.6 × 10-1 -2.5 × 10-1 -2.1 × 10-1
N = 2500 -1.6 × 10-1 -1.0 × 10-1 -1.0 × 10-1 -1.4 × 10-1
N = 5000 -2.5 × 10-1 -2.1 × 10-1 -1.5 × 10-1 -1.4 × 10-1
N = 10000 -2.3 × 10-1 -2.4 × 10-1 -2.6 × 10-1 -2.3 × 10-1
Table 8.18: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 2 of the first mode.

Experimental Case No. 1 177

 Sampled CV of Mode Shape Estimates of Mode #2, Coor #5 Estimated CV of Mode Shape Estimates of Mode #2, Coor #5

0.04 0.04
Coef. of Variation

Coef. of Variation
0.02 0.02

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Sampled CV of Mode Shape Estimates of Mode #2, Coor #4 Estimated CV of Mode Shape Estimates of Mode #2, Coor #4

0.06 0.06
Coef. of Variation

Coef. of Variation
0.04 0.04

0.02 0.02
0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Figure 8.26: Sampled and estimated coefficients of variation for the estimates of the
mode shape coordinates 5 and 4 of the second mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 1.2 × 10-1 1.1 × 10-1 5.3 × 10-2 1.9 × 10-2
N = 2500 8.9 × 10-2 9.0 × 10-2 9.5 × 10-2 7.7 × 10-2
N = 5000 8.2 × 10-2 7.6 × 10-2 2.6 × 10-2 -4.0 × 10-4
N = 10000 2.8 × 10-2 3.9 × 10-2 7.1 × 10-2 9.3 × 10-2
Table 8.19: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 5 of the second mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 5.1 × 10-1 4.9 × 10-1 4.7 × 10-1 4.7 × 10-1
N = 2500 5.0 × 10-1 5.0 × 10-1 5.1 × 10-1 5.1 × 10-1
N = 5000 5.2 × 10-1 5.1 × 10-1 4.9 × 10-1 4.7 × 10-1
N = 10000 5.0 × 10-1 5.0 × 10-1 5.2 × 10-1 5.1 × 10-1
Table 8.20: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 4 of the second mode.

178 The Results of the Simulation Study

 Sampled CV of Mode Shape Estimates of Mode #2, Coor #3 Estimated CV of Mode Shape Estimates of Mode #2, Coor #3

0.2 0.2

Coef. of Variation

Coef. of Variation
0.1 0.1

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Sampled CV of Mode Shape Estimates of Mode #2, Coor #2 Estimated CV of Mode Shape Estimates of Mode #2, Coor #2

0.04 0.04
Coef. of Variation

Coef. of Variation
0.02 0.02

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Figure 8.27: Sampled and estimated coefficients of variation for the estimates of the
mode shape coordinates 3 and 2 of the second mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 8.2 × 10-1 8.2 × 10-1 8.2 × 10-1 8.2 × 10-1
N = 2500 8.1 × 10-1 8.1 × 10-1 8.3 × 10-1 8.3 × 10-1
N = 5000 8.2 × 10-1 8.1 × 10-1 8.1 × 10-1 8.1 × 10-1
N = 10000 8.1 × 10-1 8.0 × 10-1 7.9 × 10-1 8.0 × 10-1
Table 8.21: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 3 of the second mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 2.5 × 10-1 2.6 × 10-1 2.8 × 10-1 3.0 × 10-1
N = 2500 2.0 × 10-1 1.7 × 10-1 1.7 × 10-1 1.6 × 10-1
N = 5000 2.2 × 10-1 1.8 × 10-1 2.0 × 10-1 2.1 × 10-1
N = 10000 2.3 × 10-1 2.3 × 10-1 2.7 × 10-1 2.7 × 10-1
Table 8.22: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 2 of the second mode.

Experimental Case No. 1 179

 Sampled CV of Mode Shape Estimates of Mode #3, Coor #5 Estimated CV of Mode Shape Estimates of Mode #3, Coor #5

0.2 0.2
Coef. of Variation

Coef. of Variation
0.1 0.1

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Sampled CV of Mode Shape Estimates of Mode #3, Coor #4 Estimated CV of Mode Shape Estimates of Mode #3, Coor #4

0.4 0.4
Coef. of Variation

0.2 Coef. of Variation 0.2

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Figure 8.28: Sampled and estimated coefficients of variation for the estimates of the
mode shape coordinates 5 and 4 of the third mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 1.7 × 10-1 1.2 × 10-1 7.9 × 10-2 1.5 × 10-1
N = 2500 1.4 × 10-1 1.4 × 10-1 1.9 × 10-1 1.9 × 10-1
N = 5000 1.8 × 10-1 1.6 × 10-1 1.3 × 10-1 1.7 × 10-1
N = 10000 1.5 × 10-1 1.0 × 10-1 2.3 × 10-2 5.3 × 10-2
Table 8.23: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 5 of the third mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 6.0 × 10-1 6.1 × 10-1 6.2 × 10-1 6.2 × 10-1
N = 2500 6.2 × 10-1 6.1 × 10-1 6.2 × 10-1 6.5 × 10-1
N = 5000 6.2 × 10-1 6.1 × 10-1 5.7 × 10-1 5.8 × 10-1
N = 10000 6.1 × 10-1 6.0 × 10-1 5.7 × 10-1 5.8 × 10-1
Table 8.24: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 4 of the third mode.

180 The Results of the Simulation Study

 Sampled CV of Mode Shape Estimates of Mode #3, Coor #3 Estimated CV of Mode Shape Estimates of Mode #3, Coor #3

0.2 0.2

Coef. of Variation

Coef. of Variation
0.1 0.1

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Sampled CV of Mode Shape Estimates of Mode #3, Coor #2 Estimated CV of Mode Shape Estimates of Mode #3, Coor #2

1 1
Coef. of Variation

Coef. of Variation
0.5 0.5

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Figure 8.29: Sampled and estimated coefficients of variation for the estimates of the
mode shape coordinates 3 and 2 of the third mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 -3.1 × 10-2 -4.1 × 10-2 -1.2 × 10-1 -3.7 × 10-2
N = 2500 -1.1 × 10-1 -1.4 × 10-1 -1.5 × 10-2 -8.6 × 10-3
N = 5000 -2.4 × 10-2 -3.0 × 10-2 -6.5 × 10-2 -7.4 × 10-2
N = 10000 -5.4 × 10-2 -1.1 × 10-1 -1.4 × 10-1 -8.7 × 10-2
Table 8.25: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 3 of the third mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 8.0 × 10-1 8.0 × 10-1 8.2 × 10-1 8.3 × 10-1
N = 2500 8.2 × 10-1 8.2 × 10-1 8.0 × 10-1 7.9 × 10-1
N = 5000 7.8 × 10-1 7.9 × 10-1 7.8 × 10-1 7.9 × 10-1
N = 10000 7.8 × 10-1 7.9 × 10-1 8.0 × 10-1 8.0 × 10-1
Table 8.26: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 2 of the third mode.

Experimental Case No. 1 181

 Sampled CV of Mode Shape Estimates of Mode #4, Coor #5 Estimated CV of Mode Shape Estimates of Mode #4, Coor #5

1 1
Coef. of Variation

Coef. of Variation
0.5 0.5

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Sampled CV of Mode Shape Estimates of Mode #4, Coor #4 Estimated CV of Mode Shape Estimates of Mode #4, Coor #4

1 1
Coef. of Variation

0.5 Coef. of Variation 0.5

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Figure 8.30: Sampled and estimated coefficients of variation for the estimates of the
mode shape coordinates 5 and 4 of the fourth mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 1.6 × 10-1 1.5 × 10-1 1.2 × 10-1 1.7 × 10-1
N = 2500 1.6 × 10-1 1.5 × 10-1 1.7 × 10-1 2.4 × 10-1
N = 5000 1.7 × 10-1 1.2 × 10-1 9.9 × 10-2 1.2 × 10-1
N = 10000 1.6 × 10-1 1.5 × 10-1 1.4 × 10-1 1.5 × 10-1
Table 8.27: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 5 of the fourth mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 -4.1 × 10-1 -3.8 × 10-1 -3.6 × 10-1 -2.2 × 10-1
N = 2500 -4.7 × 10-1 -4.7 × 10-1 -3.9 × 10-1 -2.3 × 10-1
N = 5000 -3.5 × 10-1 -4.1 × 10-1 -4.6 × 10-1 -4.3 × 10-1
N = 10000 -3.8 × 10-1 -4.3 × 10-1 -5.5 × 10-1 -5.6 × 10-1
Table 8.28: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 4 of the fourth mode.

182 The Results of the Simulation Study

 Sampled CV of Mode Shape Estimates of Mode #4, Coor #3 Estimated CV of Mode Shape Estimates of Mode #4, Coor #3

4 4

Coef. of Variation

Coef. of Variation
2 2

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Sampled CV of Mode Shape Estimates of Mode #4, Coor #2 Estimated CV of Mode Shape Estimates of Mode #4, Coor #2

1 1
Coef. of Variation

Coef. of Variation
0.5 0.5

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Figure 8.31: Sampled and estimated coefficients of variation for the estimates of the
mode shape coordinates 3 and 2 of the fourth mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 9.2 × 10-1 9.2 × 10-1 9.1 × 10-1 9.2 × 10-1
N = 2500 9.2 × 10-1 9.2 × 10-1 9.1 × 10-1 9.2 × 10-1
N = 5000 9.1 × 10-1 9.1 × 10-1 9.1 × 10-1 9.1 × 10-1
N = 10000 9.2 × 10-1 9.2 × 10-1 9.1 × 10-1 9.2 × 10-1
Table 8.29: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 3 of the fourth mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 -4.3 × 10-1 -4.0 × 10-1 -3.8 × 10-1 -2.1 × 10-1
N = 2500 -2.9 × 10-1 -3.4 × 10-1 -3.7 × 10-1 -2.2 × 10-1
N = 5000 -4.4 × 10-1 -4.7 × 10-1 -4.2 × 10-1 -3.3 × 10-1
N = 10000 -3.7 × 10-1 -3.5 × 10-1 -3.7 × 10-1 -4.5 × 10-1
Table 8.30: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 2 of the fourth mode.

Experimental Case No. 1 183

 Sampled CV of Mode Shape Estimates of Mode #5, Coor #5 Estimated CV of Mode Shape Estimates of Mode #5, Coor #5

2 2
Coef. of Variation

Coef. of Variation
1 1

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Sampled CV of Mode Shape Estimates of Mode #5, Coor #4 Estimated CV of Mode Shape Estimates of Mode #5, Coor #4

3 3
Coef. of Variation

Coef. of Variation
2 2

1 1
0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Figure 8.32: Sampled and estimated coefficients of variation for the estimates of the
mode shape coordinates 5 and 4 of the fifth mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 1.7 × 10-1 1.3 × 10-1 1.3 × 10-1 3.6 × 10-1
N = 2500 1.3 × 10-1 1.8 × 10-1 2.4 × 10-1 2.8 × 10-1
N = 5000 2.0 × 10-1 1.6 × 10-1 2.3 × 10-1 2.0 × 10-1
N = 10000 1.5 × 10-1 1.5 × 10-1 2.0 × 10-1 2.0 × 10-1
Table 8.31: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 5 of the fifth mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 -1.2 × 10-0 -1.2 × 10-0 -1.1 × 10-0 -2.8 × 10-1
N = 2500 -1.3 × 10-0 -1.3 × 10-0 -1.2 × 10-0 -8.9 × 10-1
N = 5000 -1.1 × 10-0 -1.2 × 10-0 -1.1 × 10-0 -1.2 × 10-0
N = 10000 -1.2 × 10-0 -1.2 × 10-0 -9.5 × 10-1 -9.9 × 10-1
Table 8.32: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 4 of the fifth mode.

184 The Results of the Simulation Study

 Sampled CV of Mode Shape Estimates of Mode #5, Coor #3 Estimated CV of Mode Shape Estimates of Mode #5, Coor #3

3 3

Coef. of Variation

Coef. of Variation
2 2

1 1
0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Sampled CV of Mode Shape Estimates of Mode #5, Coor #2 Estimated CV of Mode Shape Estimates of Mode #5, Coor #2

2 2
Coef. of Variation

Coef. of Variation
1 1

0.1 0.1

0 0.05 0 0.05
2000 4000 6000 2000 4000 6000
8000 10000 0 Noise Level, K 8000 10000 0 Noise Level, K
Record Length, N Record Length, N

Figure 8.33: Sampled and estimated coefficients of variation for the estimates of the
mode shape coordinates 3 and 2 of the fifth mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 -1.8 × 10-0 -1.8 × 10-0 -1.6 × 10-0 -6.4 × 10-1
N = 2500 -1.9 × 10-0 -1.9 × 10-0 -1.8 × 10-0 -1.6 × 10-0
N = 5000 -1.8 × 10-0 -1.8 × 10-0 -1.7 × 10-0 -1.7 × 10-0
N = 10000 -1.8 × 10-0 -1.8 × 10-0 -1.5 × 10-0 -1.5 × 10-0
Table 8.33: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 3 of the fifth mode.

K = 0.001 K = 0.01 K = 0.05 K = 0.1

N = 1250 -1.3 × 10-0 -1.3 × 10-0 -1.2 × 10-0 -6.4 × 10-1
N = 2500 -1.4 × 10-0 -1.5 × 10-0 -1.2 × 10-0 -1.4 × 10-0
N = 5000 -1.3 × 10-0 -1.3 × 10-0 -1.2 × 10-0 -1.3 × 10-0
N = 10000 -1.3 × 10-0 -1.3 × 10-0 -1.2 × 10-0 -1.2 × 10-0
Table 8.34: Normalized differences of sampled and estimated coefficients of the
variation of estimated mode shape coordinate 2 of the fifth mode.

Experimental Case No. 1 185

Separate conclusions concerning the estimation of the uncertainties of the modal
parameters will be given in the following. These will primarily be based on the
normalized differences of the sampled and estimated coefficients of variation
presented in the previous tables.

When the number of simulations tends to infinity the sampled coefficient of variation
will approach the theoretical one. This is the reason why the difference |<k- <ˆ 6 | has
been normalized with respect to <k. Thus, the estimated coefficient of variation will
be compared to the sampled coefficient of variation.

If the normalized differences are positive it implies that the uncertainty of the modal
parameter is underestimated, whereas it is overestimated if the normalized difference
is negative. If the normalized difference is numerically less than 10-2-10-3, the
sampled and estimated coefficient of variation agrees with 2-3 decimals. An
agreement of this size will probably be adequate in applications such as VBI, see
chapter 9.

Conclusions on the Natural Eigenfrequency Estimates

The figures 8.19 to 8.23 as well as the tables 8.5, 8.7, 8.9, 8.11 and 8.13 reveal a good
agreement between the sampled and estimated coefficients of variation of the
estimated natural eigen-frequencies. The eigenfrequencies are all very accurately
determined, which is underlined by both the sampled and estimated coefficients of
variation which range from 1.0 × 10-4 to 1.0 × 10-3 for the long records.

If the record length is short and the noise level is high the normalized difference is
seen to be up to 10-1. However, if the record length is long and the noise level is small
the agreement between the sampled and the estimated coefficients of variation
improves up to 10-3. In this case the sampled and estimated standard deviation will
agree very well.

Only in 20% of the simulation cases the normalized difference is negative. This
implies that the estimated standard deviations of the natural eigenfrequency estimates
tend to be too small. However, this underestimation is very small for long record
lengths and moderate noise levels.

Finally, it is observed that the coefficient of variation seems to keep on descreasing

as the number of samples tends to infinity. This is observed for all levels of noise.

Conclusions on the Damping Ratio Estimates

Again, the figures 8.19 to 8.23 and the tables 8.6, 8.8, 8.10, 8.12 and 8.14 reveal a
good agreement between the sampled and estimated coefficients of variation of the
estimated damping ratios.

186 The Results of the Simulation Study

However, the damping ratios are not so accurately determined as the natural eigen-
frequencies. The range of both the sampled and the estimated coefficients of variation
is 0.1-0.2 for the long records.

Except for the very short record lengths and high noise levels where the normalized
difference is seen to be up to 10-1, the normalized difference ranges from 10-2 to 10-3.
It does not seem to help significantly if the record length is increased, nor if the noise
level is decreased. In other words, increasing the record length does not seem to
improve the estimated coefficients of variation.

Only in 26% of the simulation cases the normalized difference is negative. This
implies that the estimated standard deviations of the damping ratio estimates tend to
be too small.

Finally, when the record length becomes moderate the decrease of the coefficient of
variation seems to saturate. This is especially true for the low noise levels.

Conclusions on the Mode Shape Estimates

From the figures 8.24 to 8.33 and the tables 8.15 to 8.34 the coefficients of variation
are presented for the coordinates 5 to 2 of the five modes. The agreement between the
sampled and estimated coefficients of variation of the estimated mode shape
coordinates is not so good as for the natural eigenfrequencies and the damping ratios.

The numerical values of the normalized difference of the sampled and estimated
coefficients of variation are more or less constant with values around 10-1. This
indicates that it is impossible to improve the estimated standard deviations of the
mode shape estimated by increasing the record length. Further, in 50% of the
simulation cases the normalized difference is negative, which means that it is
impossible to conclude whether the standard deviations will be underestimated or not.
The conclusion is therefore, that the estimated standard deviation of the mode shapes
must be interpreted with care.

It might be claimed that the choice of normalization coordinate plays a role.

However, if this is the case the normalized differences should deviate more from
mode shape to mode shape. This is not the case. However, the influence of the choice
of normalization should be investigated in future work.

8.4 General Conclusions

The results of the analysis of the standard deviations of the modal parameter
estimates indicate that the estimator probably becomes efficient as the record length
tends to infinity. This is underlined by the bias analysis of the modal parameter
estimates.

Experimental Case No. 1 187

The bias of the natural eigenfrequency and damping ratio estimates decreased rapidly
until a record length of 5000 samples. This was experienced for all levels of noise.
The mode shapes decreased more slowly especially for high levels of noise. But in
any case, the bias keept on decreasing as the number of samples was increased.

The achievable accuracy of estimated ARMA models has been investigated in e.g.
Gersch [26], Jensen et al. [45] and Kirkegaard [52] and they have made the following
conclusions :

L The uncertainty of the modal parameters is relatively insensitive with

respect to the number of degrees of freedom, Gersch [26].

L In practice, for a given level of damping, only limited improvement in the

accuracy of the damping ratios estimates is obtained by increasing the
record length N when N has obtained a given magnitude, Jensen et al. [45].

L The optimal choice of N can in principle only be evaluated by a cost-benefit

analysis, Jensen et al. [45] and Kirkegaard [52].

L By choosing the record length large enough, the uncertaities of the natural
eigenfrequencies and the damping ratios can be quantified, Jensen et al.
[45].

L The accuracy of the estimated natural eigenfrequencies will increase but the
accuracy of the damping ratios will decrease for decreasing levels of
damping, Jensen et al. [45].

So in conclusion :

L For record lengths over 5000 samples in each channel and for a noise level
K below 1% the standard deviations of the natural eigenfrequencies and the
damping ratios can be quantified accurately and the bias will be insignifi-
cant.

The standard deviations of the mode shapes will be more inaccurately estimated and
care must be taken in interpreting these. The bias of the mode shapes will keep on
decreasing even for very long records.

188 General Conclusions

8.5 Summary
As described in section 1.7.1, the intention of this simulation study has been to assess
the statistical properties of the PEM estimate of an ARMAV model. If the true system
can be contained in the model structure and if the prediction error is Gaussian white
noise, then the PEM estimator will be efficient. In this case the estimate will be
asymptotically unbiased and the estimated standard deviations of the estimated modal
parameters will attain the Cramer-Rao lower bound.

At the beginning of section 8.3, it was established that the ARMAV(2,2) model used
in the estimation was an adequate model. This implies that the true system can be
contained in it. It has been verified that the prediction errors are realizations of a
multivariate Gaussian white noise process. It has also been verified, that the number
of simulations is adequate in order to obtain convergent statistical properties of the
estimated modal parameters.

It is concluded that for record lengths over 5000 samples in each channel and for a
noise level K below 1% the standard deviations of the natural eigenfrequencies and
the damping ratios can be quantified accurately and the bias will be insignificant.
However, the standard deviations of the mode shapes will be more inaccurately
estimated and care must be taken in interpreting these. The bias of the mode shapes
will keep on decreasing even for very long records.

Experimental Case No. 1 189

190