9 Experimental Case No.
In this chapter the experimental case no. 2 introduced in section 1.4.2 is described.
This experimental case concerns a vibration based inspection (VBI) of a lattice steel
test mast. The ideas behind VBI have been explained in section 1.2.2. VBI is divided
into four levels. In this experimental case only the use of level 1 methods will be
illustrated. In other words, it will only be shown how damage can be detected on the
basis of modal parameter estimates. In section 9.1, the measurement procedure and
the signal processing applied prior to the system identification of the mast are
explained. Section 9.2 concerns the system identification of the dynamic properties
of the mast in its virgin state, i.e. before the damage is introduced. The system
identification of the mast in its damaged states is described in section 9.3, whereas
the actual damage detection is given in section 9.4.
9.1 Data Acquisition and Signal Processing
As explained in section 1.4.2, the mast has been equipped with 6 accelerometers.
However, as also explained, one of these was unfortunately destroyed during the test
because of heavy rain. This section explains how the acquisition of the measurements
of the five remaining accelerometers has been performed. It will also be explained
how these measurements have been processed prior to the actual system identifica-
tion.
The data have been sampled at 500 Hz using a 16 bit A/D converter, as well as analog
antialias filtered. After sampling the data were decimated 10 times, i.e. lowpass
filtered and resampled at a 10 times lower frequency. The reason for this
oversampling is that it improves the signal to noise ratio in the data, since every data
point is constructed from 10 samples. Thus, the resulting sampling interval becomes
T = 0.02 sec. Finally, the filtered data have been detrended and calibrated to
accelerations described in the unit m/sec2. Each measurement session lasted 160 sec,
which implies that the final record length has N = 8000 points. This record length is
assumed to be adequate in order to obtain an accurate system identification. In figure
9.1, the data of one of the measurement sessions are presented for all five channels.
Experimental Case No. 2 191
Filtered and Calibrated Measurements − Channels #1 to #5
0.5
m/sec²
−0.5
0 20 40 60 80 100 120 140 160
0.5
m/sec²
−0.5
0 20 40 60 80 100 120 140 160
0.5
m/sec²
−0.5
0 20 40 60 80 100 120 140 160
0.2
m/sec²
−0.2
0 20 40 60 80 100 120 140 160
0.2
m/sec²
−0.2
0 20 40 60 80 100 120 140 160
Time [sec]
Figure 9.1: Example of the measured accelerations of the five channels.
9.2 Virgin State Analysis
In the VBI approach used in this chapter a virgin state of the structure has to be
defined. In this state the dynamic properties of the structure have to be determined
with as much accuracy as possible. The damage indicators of the structure identified
at a later state will then be compared with the damage indicators of the virgin state.
The virgin state tests of the structure were initiated at August 30, 1996 and lasted
until November 4, 1996. During this period 96 measurement sessions were
performed. In this section, it is described how the dynamic properties of the mast
have been modelled and identified on the basis of these measurements.
9.2.1 Adequate Modelling of the Dynamic Properties
At the present state, it is not possible to determine which of the modal parameters
will be sensitive to damage. However, it is assumed that some of the estimated
natural eigenfrequencies or the associated damping ratios are applicable as damage
indicators. Since the mode shapes are only described in a few points of the structure
these will not be used in the VBI analysis as a damage indicator. This means that the
system identification in principle can be performed by a univariate model structure.
192 Virgin State Analysis
However, as seen in figure 9.2 there are two closely spaced modes around 2 Hz and
two closely spaced modes around 11.5 Hz.
Average of All Spectral Densities of All Virgin State Measurements
103
102
101
100
Magnitude of Spectral Density
10-1
10-2
Zoomed Region Zoomed Region
300
20
10 -3
250
15
200
10-4
10
150
10 -5
100
5
1.8 1.9 2 2.1 2.2 11.4 11.6 11.8
10-6
0 5 10 15 20 25
Frequency [Hz]
Figure 9.2: FFT-based spectral density of the response using all virgin state
measurements. The auto-spectral densities and magnitudes of the cross-
spectral densities of all measurement records have been averaged. All
records have been normalized to unit standard deviation. The frequency
resolution is 2048 between 0 and 25 Hz. The zoomed regions are the
areas around the peaks at 2 Hz and 11.5 Hz.
The two modes located around 2 Hz are the first modes of bending in the two
perpendicular directions, see figure 1.4. The mode located around 8 Hz is the first
torsional mode of the structure, and the two modes located around 11.5 Hz are the
second modes of bending. These observations are in accordance with prior research
reported in Kirkegaard et al. [58], where the dynamic behaviour of the structure has
been investigated through experimental studies as well as FEM analysis. Because of
the closely spaced modes the most sensible choice of model structure is an ARMAV
model with as many channels as possible, i.e. 5 channels. In this way information
from all sensors is used. Assuming that all 5 modes are underdamped, it is necessary
to use at least the ARMAV(2,1) model. However, besides modelling the five
structural modes there is a significant filter characteristic present just below the
Nyquist frequency as seen in figure 9.2 which also has to be modelled. This filter
characteristic originates from the lowpass filter applied in the decimation procedure
before the data were resampled. Further, in some of the measurements the signal-to-
noise ratio is low due to a limited presence of wind.
Experimental Case No. 2 193
This implies that the model structure in general must be of the ARMAV(n,n) type.
The order n has been selected on the basis of the Akaike FPE criterion as n = 3.
According to the results of section 3.3.1, this is also the lowest model order to use
when the measurements are accelerations. This implies that the number of parameters
needing to be estimated is 150. This model structure has been used in all identifica-
tions. The actual system identifications have been performed using the non-linear
PEM algorithm described in chapter 5 and implemented in the routine armav.m of
the STDI toolbox.
In all identifications the maximum number of iterations used in the PEM algorithm
is 40. 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 × 10-6 the iterations are stopped. The
robustification parameter D defined in section 7.1.4 is set to 1.6. In figure 9.3, the
adequacy of the model is illustrated by plotting the FFT-based auto-spectral densities
together with the auto-spectral densities obtained from the identified ARMAV model
for one of the virgin state measurement sessions.
Autospectral Density − Channel #1 Autospectral Density − Channel #2
0 0
10 10
−5 −5
10 10
0 5 10 15 20 25 0 5 10 15 20 25
Frequency [Hz] Frequency [Hz]
Autospectral Density − Channel #3 Autospectral Density − Channel #4
0 0
10 10
−5 −5
10 10
0 5 10 15 20 25 0 5 10 15 20 25
Frequency [Hz] Frequency [Hz]
Autospectral Density − Channel #5
0
10
−5
10
0 5 10 15 20 25
Frequency [Hz]
Figure 9.3: Comparison of the auto-spectral densities obtained using FFT and the
estimated ARMAV(3,3) model using data from one of the virgin state
measurement session.
194 Virgin State Analysis
In the damaged states of the structure, it is the intention to make use of the estimated
uncertainties of the estimated modal parameters. However, as in the experimental
case no. 1 the prediction errors must be Gaussian white noise for these estimates to
be accurate enough. It is therefore necessary to check, whether the prediction errors
are Gaussian distributed or not. This can be checked using a normal probability plot,
which is illustrated in figure 9.4 for the prediction errors of an ARMAV model of one
of the virgin state measurement sessions.
Normal Probability Plot − Channel #1 Normal Probability Plot − Channel #2
0.999 0.999
0.99 0.99
Probability
Probability
0.95 0.95
0.75 0.75
0.25 0.25
0.05 0.05
0.01 0.01
0.001 0.001
−4 −2 0 2 4 −4 −2 0 2 4
Prediction Errors Normalized to Unit Standard Deviation Prediction Errors Normalized to Unit Standard Deviation
Normal Probability Plot − Channel #3 Normal Probability Plot − Channel #4
0.999 0.999
0.99 0.99
Probability
Probability
0.95 0.95
0.75 0.75
0.25 0.25
0.05 0.05
0.01 0.01
0.001 0.001
−4 −2 0 2 4 −4 −2 0 2 4
Prediction Errors Normalized to Unit Standard Deviation Prediction Errors Normalized to Unit Standard Deviation
Normal Probability Plot − Channel #5
0.999
0.99
Probability
0.95
0.75
0.25
0.05
0.01
0.001
−4 −2 0 2 4
Prediction Errors Normalized to Unit Standard Deviation
Figure 9.4: Normal probability plot of the prediction errors of an estimated
ARMAV(3,3) model obtained using data from one of the virgin state
measurement sessions.
In the present case the normal probability plot in figure 9.3 indicates a 95%
confidence in that the prediction errors of at least the first three channels of this
particular model are Gaussian distributed. Besides checking whether the prediction
errors are Gaussian or not, it is also necessary to check whether the prediction errors
are white noise or not. This is checked by plotting their auto-correlation functions or
their spectral densities as illustrated in section 7.3.3. These two checks must be
performed in all the cases where the estimated uncertainties of the modal parameters
are used. If the confidence is low, the estimated uncertainties should not be used. In
the following a 95% confidence is assumed adequate and any estimated models that
do not have prediction errors satisfying this requirement are rejected in the damaged
states analysis.
As mentioned above, 96 identifications were performed. This section describes how
the structural modes have been identified from these identifications. It is the intention
Experimental Case No. 2 195
to average the 96 estimates of the modal parameters and use these averaged values
to describe the virgin state dynamic properties of the structure. In the same way, it is
possible to describe the uncertainties of the dynamic properties of the virgin state by
means of the sampled standard deviations of the modal parameter estimates.
The structural modes of the identified models have been determined by the following
two stability [Link] stability diagram shown in figure 9.5 utilizes the MAC
criterion. If the mode shape of a mode of a currently estimated model correlates with
a mode shape belonging to the previously estimated model in such a sense that the
MAC between these mode shapes are at least 0.95, then a circle “o” will be marked
in the stability diagram. The location of this circle will be at the coordinate indicating
the number of the current model and the natural eigenfrequency that corresponds to
the mode shape. All identified modes of the models are marked with dots “·” and
located in the same way in the diagram.
In figure 9.6, a frequency stabilization diagram is shown. Again all identified modes
are marked with dots “·”. If the damping ratio of a particular mode is less than 1% the
mode is marked with a circle “o”. If the natural eigenfrequencies of a mode of the
current and previous models have an absolute and maximum deviation of 5%, then
this mode is marked with a “+”. The modes that originate from the structural system
will be underdamped. This implies that their damping ratios will be small, which can
be verified by the spectral density plot in figure 9.2. Therefore, it makes sense to
assume that the damping ratios of the identified structural modes are less that 1%.
The modes of the noise will typically have a broad-banded nature, i.e. be more
heavily damped. The natural eigenfrequencies of these modes will typically deviate
from model to model, since the noise properties of the measurement records are never
the same. Further, the “mode shapes” of these modes will typically deviate from
model to model. This is why the MAC criterion is an efficient way to locate structural
modes. See also section 6.4 for further details on identification of structural modes.
196 Virgin State Analysis
MAC Stability Diagram − Threshold=0.95
90
80
70
ARMAV(3,3) − Model Number 60
50
40
30
20
10
0
0 5 10 15 20 25
Frequency [Hz]
Figure 9.5: MAC stability diagram. A dot “·” indicates an identified mode. If the
mode shape of a mode of an estimated model has a MAC value not less
than 0.95 with a mode shape of the previously estimated model, then the
mode is marked with a circle “o".
Frequency Stability Diagram − Damping Threshold is 1% and Maximum Frequency Deviation is 0.5%
90
80
70
ARMAV(3,3) − Model Number
60
50
40
30
20
10
0
0 5 10 15 20 25
Frequency [Hz]
Figure 9.6: Frequency stability diagram. A dot “·” indicates an identified mode. A
circle “o” indicates that the damping ratio of the mode is less 1%,
whereas a “+” indicates that a particular natural eigenfrequency of a
model has been recovered in the previously estimated model with a
maximum absolute deviation of 5%.
Experimental Case No. 2 197
From the MAC stability diagram, it is seen that two circles are marked around 2 Hz
and around 11.5 Hz for almost all estimated models. It is also observed that the
modes around 8.1 Hz are marked with a circle. Although it might be difficult to
observe, the frequency stability diagram reveals that the two modes around 2 Hz have
damping ratios less than 1% for almost all models. This is also observed for the two
modes located around 11.5 Hz and the one around 8.1 Hz. Finally, it should be
observed that the modes at the same locations are marked with a “+”, which indicates
that the maximum and absolute deviation of these is less than 5%, between two
consequetive model estimates. The average values and sampled standard deviations
of all the marked modal parameter estimates of the five structural modes have been
calculated. From these values it is also possible to estimate the coefficient of
variation, that turns out to be an important parameter in the analysis of the estimated
modal parameters in the damaged states. The results are seen in tables 9.1 and 9.2.
Mode # fi [Hz] Ff [Hz] × 10-3 <f ×10-3
i i
1 (Bending) 2.014 2.49 1.24
2 (Bending) 2.044 2.63 1.28
3 (Torsion) 8.166 3.91 0.48
4 (Bending) 11.504 8.66 0.75
5 (Bending) 11.642 5.44 0.46
Table 9.1: Mean values fi, sampled standard deviations Ff and coefficients of
i
variation <f of the natural eigenfrequencies of the five structural modes.
i
Mode # .i [%] F. [%] <.
i i
1 (Bending) 0.31 0.14 0.46
2 (Bending) 0.31 0.15 0.48
3 (Torsion) 0.10 0.05 0.48
4 (Bending) 0.18 0.20 1.07
5 (Bending) 0.13 0.06 0.45
Table 9.2: Mean values .i, sampled standard deviations F. and coefficients of
i
variation <. of the natural eigenfrequencies of the five structural modes.
i
As seen the coefficients of variation of the estimated natural eigenfrequencies are
very small compared to the coefficients of variation of the estimated damping ratios.
Since the coefficients of variation of the estimated damping ratios have values
between 0.45 and 1.07, it implies that the estimates are very poor. It is therefore not
worthwhile to use these estimates in a VBI.
198 Virgin State Analysis
So in conclusion :
L The natural eigenfrequency estimates of the virgin state of five structural
modes are very accurately estimated and can thus be used in a VBI if they
are unbiased and sensitive to the actual damage.
On the basis of the results of the previous chapter, concerning the relation between
record length and bias, it will be assumed that the estimated natural eigenfrequencies
are unbiased.
9.3 Damaged States Analysis
The location of the introduced crack was described in section 1.4.2. In this section,
the influence of this growing crack is described. The accuracy of the natural
eigenfrequency estimates will be compared with the accuracy of the natural
eigenfrequency estimates of the virgin state, and it will be determined which of the
natural eigenfrequencies are sensitive to the introduced crack. Finally, a simple
damage detection will be performed.
9.3.1 Accuracy Check of Estimated Eigenfrequencies
It is assumed that the structure shows linear and time-invariant behaviour in each
damage state. This is a resonable assumption since the size of the crack is extended
between the last and the first measurement session of two consequetive damage
states. If the size of the crack was increased significantly during a measurement
session the system would perhaps behave time-variant. In such a case, it would be
necessary to apply an on-line system identification approach. The result of the virgin
state analysis was an accurate determination of the natural eigenfrequencies of the
five structural modes. From table 9.1, it is seen that the estimation inaccuracies result
in coefficients of variation around 1.0 × 10-3. As described in section 9.2.1, any model
that does not have prediction errors that with a resonable confidence form a Gaussian
white noise will be rejected. But in addition, natural eigenfrequency estimates with
estimated coefficients of variation larger than 1.0 × 10-3 will result in a rejection of
the model. In this way the uncertainties of the natural eigenfrequency estimates of
the damaged states will qualitatively correspond to the estimates of the virgin state.
9.3.2 Determination of Sensitive Eigenfrequencies
At the present state, it has been established that the natural eigenfrequencies of the
five structural modes can be estimated with a high degree of accuracy. However, this
does not mean that they can be used as damage indicators. They might simply not be
sensitive to the damage. The easiest way to determine whether they are suitable as
damage indicators or not is by plotting the estimated natural eigenfrequencies that
have passed the rejection criteria described in the previous section.
Experimental Case No. 2 199
In the following five figures these estimates will be plotted together with their
estimated 95% confidence limits. The confidence limits are determined as ±2F̂f i ,
where F̂f i is the standard deviation estimated using the technique described in section
6.2.1. The damage state that corresponds to the individual estimate is indicated in the
lower plot.
Estimated Natural Eigenfrequencies and 95% Confidence Interval of Mode #1
2.025
2.02
Frequency [Hz]
2.015
2.01
2.005
0 2 4 6 8 10 12
Number of Estimates
Relation Between Estimation Number and Damage State
5
4
Damage State
0
0 2 4 6 8 10 12
Number of Estimates
Figure 9.7: Estimated natural eigenfrequencies of the first mode that have passed the
rejection criteria. The estimates are plotted together with their estimated
95% confidence interval. The relation between the estimates and the
damage states is plotted below.
200 Damaged States Analysis
Estimated Natural Eigenfrequencies and 95% Confidence Interval of Mode #2
2.055
2.05
Frequency [Hz]
2.045
2.04
2.035
2.03
0 2 4 6 8 10 12
Number of Estimates
Relation Between Estimation Number and Damage State
5
4
Damage State
0
0 2 4 6 8 10 12
Number of Estimates
Figure 9.8: Estimated natural eigenfrequencies of the second mode that have passed
the rejection criteria. The estimates are plotted together with their
estimated 95% confidence interval. The relation between the estimates
and the damage states is plotted below.
Estimated Natural Eigenfrequencies and 95% Confidence Interval of Mode #3
8.19
8.18
Frequency [Hz]
8.17
8.16
8.15
8.14
8.13
8.12
0 10 20 30 40 50 60 70 80 90 100
Number of Estimates
Relation Between Estimation Number and Damage State
5
4
Damage State
0
0 10 20 30 40 50 60 70 80 90 100
Number of Estimates
Figure 9.9: Estimated natural eigenfrequencies of the third mode that have passed the
rejection criteria. The estimates are plotted together with their estimated
95% confidence interval. The relation between the estimates and the
damage states is plotted below.
Experimental Case No. 2 201
Estimated Natural Eigenfrequencies and 95% Confidence Interval of Mode #4
11.56
11.54
Frequency [Hz] 11.52
11.5
11.48
11.46
11.44
0 20 40 60 80 100 120
Number of Estimates
Relation Between Estimation Number and Damage State
5
4
Damage State
0
0 20 40 60 80 100 120
Number of Estimates
Figure 9.10: Estimated natural eigenfrequencies of the fourth mode that have passed
the rejection criteria. The estimates are plotted together with their
estimated 95% confidence interval. The relation between the estimates
and the damage states is plotted below.
Estimated Natural Eigenfrequencies and 95% Confidence Interval of Mode #5
11.68
11.66
Frequency [Hz]
11.64
11.62
11.6
11.58
11.56
0 10 20 30 40 50 60 70 80 90 100
Number of Estimates
Relation Between Estimation Number and Damage State
5
4
Damage State
0
0 10 20 30 40 50 60 70 80 90 100
Number of Estimates
Figure 9.11: Estimated natural eigenfrequencies of the fifth mode that have passed
the rejection criteria. The estimates are plotted together with their
estimated 95% confidence interval. The relation between the estimates
and the damage states is plotted below.
202 Damaged States Analysis
From the five figures, it is seen that the third and fifth mode changes significantly.
The changes especially occur when the third damage state is entered. It is seen that
in both cases the change from the first to the fourth damage state is so significant that
the 95% confidence intervals are seen to be non-overlapping.
So in conclusion :
L It can be concluded that the third and fifth natural eigen-frequency have
changed significantly when the final damage state is reached.
The detected changes of the natural eigenfrequencies are so significant that they
probably are caused by a structural change. However, the modal parameters can also
exhibit small changes due to fluctuations in the ambient environment. When e.g. the
ambient temperature changes, thermal expansion effects and changes of the stiffness
will occur. However, it is beyond the scope of this experimental case to consider this
influence.
In the context of influence of the ambient environment the effects of fluctuating
ambient temperatures on this particular mast have been investigated in Kirkegaard
et al. [58], and a technique based on Kalman filtering for removal of such an
influence has been proposed in Andersen et al. [7].
9.4 Damage Detection
The above observations imply that the the third and fifth natural eigenfrequencies can
be used as damage indicators for this particular damage. The question is at what
damage state it is possible to detect the damage with a significant confidence. It will
in the following be assumed that a damage has been detected if the confidence
intervals of all the natural eigenfrequency estimates of a mode at some damage state
are non-overlapping with the 95% confidence interval of the natural eigenfrequency
of the same mode in the virgin state. This detection approach can be utilized by
plotting the estimates of the third and fifth natural eigenfrequencies and their
estimated 95% confidence intervals of the damage states together with the averaged
natural eigenfrequency estimates and the estimated confidence intervals of the virgin
state. In figure 9.12, this is done for the third natural eigenfrequency estimates, and
in figure 9.13 for the fifth natural eigenfrequency estimates.
Experimental Case No. 2 203
Estimated Natural Eigenfrequencies and 95% Confidence Interval of Mode #3
8.19
8.18
Frequency [Hz]
8.17
8.16
8.15
8.14
8.13
8.12
0 10 20 30 40 50 60 70 80 90 100
Number of Estimates
Relation Between Estimation Number and Damage State
5
4
Damage State
0
0 10 20 30 40 50 60 70 80 90 100
Number of Estimates
Figure 9.12: Estimated natural eigenfrequencies of the third mode and their
estimated 95% confidence intervals.
Estimated Natural Eigenfrequencies and 95% Confidence Interval of Mode #5
11.68
11.66
Frequency [Hz]
11.64
11.62
11.6
11.58
11.56
0 10 20 30 40 50 60 70 80 90 100
Number of Estimates
Relation Between Estimation Number and Damage State
5
4
Damage State
0
0 10 20 30 40 50 60 70 80 90 100
Number of Estimates
Figure 9.13: Estimated natural eigenfrequencies of the fifth mode and their
estimated 95% confidence intervals.
204 Damage Detection
In figure 9.12, it is seen that the confidence intervals become non-overlapping at the
moment damage state four is entered. However, already at damage state three the
confidence intervals of the natural eigenfrequency estimates of the fifth mode are
completely non-overlapping with the virgin state confidence interval, see figure 9.13.
So in conclusion :
L Due to the above definition of a significant damage, it can be concluded that
the actual damage has been detected when it entered the third damage state.
By using the estimated standard deviations, it is actually possible to detect a damage
and provide a probabilistic measure of the confidence that should be put in the
detection.
9.5 Summary
This chapter has concerned how system identification using ARMAV models can be
used in applications such as VBI as a basis for damage detection. By using the
ARMAV models accurate natural eigenfrequency estimates have been obtained and
their uncertainties have been estimated. Due to the accuracy of the ARMAV
estimates, it has been possible to identify two sets of closely spaced modes and to
detect significant changes of some of the estimates due to an introduced damage
down to a few per cent. By using the estimated standard deviations, the confidence
in the significance of these changes has been estimated in a probabilistic sense. The
conclusion is that the damage can be detected with a 95% confidence when it enters
its third state.
Experimental Case No. 2 205
206
