You are on page 1of 17

Structural Health Monitoring of Arch Dam from Dynamic Measurements

Jian Huang Weng1 and Chin Hsiung Loh2


1
Ph.D Candidate, National Taiwan University, Department of Civil Engineering,
Taipei, Taiwan; email: d94521002@ntu.edu.tw
2
Professor, National Taiwan University, Department of Civil Engineering, Taipei,
Taiwan; email: lohc0220@ccms.ntu.edu.tw

ABSTRACT

This paper presents the system identification of the Fei-Tsui arch dam using the
recorded seismic data and ambient vibration data. The modal properties of the dam
under different reservoir water level are identified using the recorded seismic data
from 84 earthquake events. Considering the spatial variability of input excitation,
both multi-input and single-input system models are employed in the input/output
subspace identification. The regression curves between the natural frequencies and
the reservoir water level are developed from the statistical analysis of identification
results. In order to compare the current behavior of the dam to the past, an ambient
vibration experiment is performed and the output-only stochastic subspace
identification method is used to identify the current modal properties of the dam.
Finally, a safety evaluation is made by pointing the current identification result on
the developed regression curve. Further more, the comparison between different
identification algorithms in this study is made. From the stability diagram of the
identification the output-only stochastic subspace identification (using ambient data)
provides more clear system characteristics than the input/output subspace
identification (using seismic data). Discussion on the single-input model and the
multi-input model for subspace identification is also made in this study.

1. INTRODUCTION

Monitoring technology plays an important role in securing integrity of structural


system and maintaining the longevity of the structure. It consists of three aspects: (1)
instrumentation with sensors, (2) methodologies for obtaining meaning information
concerning the structural health monitoring, (3) early warning from the measured
data. Various methods based on the dynamic and static test have been applied to
address the structural health monitoring and damage identification. One important
problem in the seismic safety analysis of dams is the evaluation of the hydrodynamic
forces induced reservoir-dam interaction during strong earthquakes. Generally, the
damage of structure (such as dam) may be detected from the variation of structural
features; however, these features may be affected by the changing environmental
conditions such as reservoir water level and temperature. Around the world, very few
major dams had enough seismic response data for developing the safety evaluation
guideline of dam. The problem is even more complicate for concrete arch dam
because of the geometry shape of the dam. Therefore, continuous monitoring as well
as the seismic monitoring of an arch dam becomes one of the important missions for
the safety assessment of the dam.
Application of system identification techniques to the recorded seismic response
of structures had been studied based on discrete-time linear filter approach [1, 2].
Validation of numerical model of concrete dam using ambient vibration data had
been studied [3-5]. Time- domain identification on the seismic response data of an
arch dam is even more difficult. In this study the discrete-time system identification
of the Fei-Tsui arch dam, located in Taiwan, using both the recorded seismic data and
the ambient vibration data are studied. The seismic records had been used to identify
the dynamic characteristics of the dam by using ARX-LS method in 1996 [6]. After
Chi-Chi earthquake a forced vibration test was conducted on the dam to identify the
resonant frequencies [7]. Discussions among all these study on the results of
identification of the dam are made.

2. DESCRIPTION OF THE FEI-TSUI ARCH DAM

The sketch of Fei-Tsui arch dam is shown in Figure 1. It is a 122.5 m high and
510 m long concrete arch dam which is located in the Taipei County, Taiwan. Built
from 1979 to 1987, it was constructed with a layout of three-centred double
curvature with variable thickness. The dam body is divided into 29 blocks and each
block is 17.5 m long. The level of the water in the reservoir normally varies between
elevation of 170 m and 120 m. The capacity of this reservoir is about 400 million m3
when full for a lake surface of 10.24 km2. The local temperature normally varies
between 50 ℉ to 86 ℉. According to the previous research, the reservoir level is the
most important factor to change the dynamic features of the arch dam. Therefore, the
relationship between reservoir water level and features must be clarified and
regressed before the structural health monitoring of the arch dam.
3. SYSTEM IDENTIFICATION FROM SEISMIC RESPONSE DATA

To monitor the dynamic properties of the dam during earthquake, eleven


triaxial accelerometers were deployed in the Fei-Tsui dam as shown in Figure 1(a).
Five of these instruments are installed along the abutment (SD1 ~ SD5), three of
them (SD6 ~ SD8) are installed in the 115 m - level gallery and the rest of them
(SDA, SDB and SDC) are deployed in the 150 m - level gallery. The strong motion
accelerometers in Fei-Tsui dam were upgraded in 1998. The strong motion records of
84 earthquakes collected after the upgrade in 1998 (from 1999 to 2008) were chosen
for this study. Figure 2 shows the relationship of the seismic events with respect to
the reservoir water level. The colour of the circle indicated different PGA level
recorded from accelerometer SD1. The most intense one of these earthquakes
occurred on March 31, 2002 and it was called as 331 Earthquake in Taiwan. The
recorded PGA of 331 Earthquake is up 0.028g which is greater than 921 Chi-Chi
Earthquake (0.025g). The strong motion instrumentation of Fei-Tsui arch dam
includes eleven GeoSIG AC-63 triaxial accelerometers on the dam body since 1998.
All of these instruments are connected together and share the same recording system
and all the collected data are synchronized. The analogue signals were digitized by
using a 16-bit analogue-to-digital converter and the corresponding sampling rate is
200Hz. The specification of the accelerometer is shown in Table 1.

3.1 Numerical model of the arch dam


The subspace identification algorithm [9] was applied to the seismic response
data of the dam to extract the dynamic properties of the dam during earthquakes. The
subspace identification algorithm is a time domain method that can effectively
extract the system natural frequencies and mode shapes through a multi-input and
multi-output system identification. According to the previous research of Loh et al.
[7], the input ground motion for the Fei-Tsui arch dam is not uniform along the
abutment of the dam and it is suggested that the dam should be considered as a
multi-input system for system identification. In fact, it is very important to define the
relationship between input and output when using the subspace identification
algorithm. As long as the ground motion was measured at finite locations
(SD1~SD5), the dam was naturally considered as a dynamic system with multiple
support excitation [8]. The equation of dynamic equilibrium for such a system can be
written in partitioned form:
 m m u
&&   c c u&   k k u   0 
t t t

+   +    =  
g g g

m m  u (1)
&&  c c u&  k k u  p 
T T T

 g gg g g gg g g gg g g

The total displacement vector now contains two parts: (1) u includes the DOFs of
g

the supports (such as SD1~SD5); and (2) u t includes all DOFs of the dam except
the DOFs of the supports. For earthquake loading it is observed that only the support

forces p are applied to the system. Eq.(1) then can be rewritten by focusing on the
g

dynamic displacements u on the DOFs of the dam:


&& + cu& + ku = p
mu eff
(2)
where the effective earthquake forces is:
p eff
= − (m u
&& + m u
&&
s

g g ) − ( cu& s
+ c u& g g ) (3)
where u = u − u is the quasi-static displacement. Eq.(3) can be further
s t

simplified by considering two “possible” assumptions for real application: (1) small
damped system; (2) lump mass system. Base on the first assumption the damping
force is much small relative to the inertia force so that it can be neglected. For the

second assumption, the undiagonal term m will be a null matrix and therefore can
g

be dropped. After the simplifications of the effective earthquake forces, Eq.(2) is


rewritten as:
&& + cu& + ku = −mιu
mu && g
(4)

where ι = −k k is the influence matrix because it describes the influence of


−1

support displacements on the structural displacements. Eq.(4) defines a numerical


model for a dynamic system with multiple support excitations. On the other hand, if
the input ground motion is assumed to be uniform and one of the records from
SD1~SD5 is selected to be the representative ground motion. The motion equation
will be simpler than Eq.(4) base on assumption of uniform input:

&& + cu& + ku = −m1u&&


mu g
(5)

where u is a scalar of the uniform input and 1 is a vector with each element equal
g

to unity. A multi-support excitation system it may use more input excitations to


model the non-uniform input ground motion but it also needs to discrete the
boundary of the dam in order to get the equivalent earthquake forces. And also, there
are two assumptions needed to be satisfied for the simplification of effective
earthquake forces in Eq.(3). As for the single-input system, there is only one
assumption should be satisfied but this assumption is not “completely” true for real
application. In this study, both of the two numerical models will be considered by
using subspace identification algorithm to identify the system. The comparison
between the results from multi-input and single-input may improve the accuracy of
the system identification results of the Fei-Tsui arch dam.

3.2 System identification using input/output subspace identification


In recent years, various versions of subspace identification algorithms have
been used to identify the modal properties of linear systems. A unified treatment of
most of these algorithms has been given by Van Overschee and De Moor [9]. The
input/output subspace identification algorithm starts from the continuous-time state
space model which is the different form of the motion equation:
x& = A x + B u
c
&& + wc g (6)
&& = Cx + v
u t
(7)
where
u   0 I  0
x =  ; A =   ; B =   ; C = [−m k −1
−m c ]
−1

 u&   −m k −m c  −ι 
c −1 −1 c

w is the process noise due to disturbances or modeling error and v is the


measurement noise due to disturbances or sensor noise. It is assumed that they are
zero mean and white vector sequences. In Eqs.(6) and (7), the inputs of this system

are the absolute accelerations of all supports && and the outputs are the absolute
ug

&& . For practical application, only few DOFs


accelerations of all DOFs of the dam u
t

of the dam were measured and Eq.(7) should be modified as:

um
&& = TCx + Tv = C ′x + v ′
&& = Tu
t t
(8)
Eq.(8) implies there are some thing lost if not all of DOFs are measured. For example,
we can only identify the mode shape on where we measured and the complete system
can not be reconstructed.
The basic concept of subspace algorithms is exploitation of the state as a
finite-dimensional interface between the past part and the future part. First, the input
and output data are arranged into the Hankel matrices. Then projection theorem is

employed to avoid the influence of noise and extract the observability matrix Γ: i
Y /f U f
W =ΓX
p i
d

f (8)
After extracting the observability matrix by using singular value decomposition, it is

easy to compute the system parameters A c


and C′ . Finally, we can identify the

natural frequencies and mode shapes of the dam. The detail procedure of the
subspace identification can be found in reference 6.
To identify the modal properties of Fei-Tsui arch dam, the seismic response
data from 331 Earthquake was firstly used to verify the input/output subspace
identification. The corresponding reservoir water level is 144.4 m during the
earthquake. As mentioned before, the theory of subspace identification comes from
the motion equations. It is important to make sure that all records are in the same
direction (orientation) so that the force equilibrium can be satisfied. Figure 3 shows
the setting directions of each accelerometer, and it is easy to find that the setting
directions of each accelerometer were all perpendicular to the dam surface. A global
coordinate system was defined to calibrate all the records to the same direction. The
input excitations after calibration are shown in Figure 4. It is observed that the
canyon phenomenon is confirmed from the non-uniform motion along the dam
abutment. The record of SD1 is similar to SD2 and the record of SD4 is similar to
SD5. But the seismic wave forms changed after passing through the canyon so that
SD2 is different from SD4. For the multi-input case, the absolute acceleration records
at the stations SD1, SD2, SD3, SD4 and SD5 were defined as the input excitations.
The absolute acceleration records at SD6, SD7, SD8, SDA, SDB and SDC were
defined as the outputs. In order to consider the canyon phenomenon we chose SD1 or
SD5 as the uniform input in the single-input case and the outputs are the same as
above. The identification results of 331 Earthquake are presented as the stability
diagrams and shown in Figure 5.
Follow the subspace identification the stability diagram describes the identified
modal frequencies under different choice of the number of block row. As shown in
the Figure 5, the multi-input case can identify more structural modes but the modal
frequencies are not completely consistent with the average of the Fourier spectrum of
the dam responses which is printed in the back ground. As for the single-input cases,
the results using SD1-input are very similar to the results using SD5-input.
According to the statistic results of the stability diagram, the modal frequencies
identified from SD1-input are 2.46, 2.58 and 3.40 Hz. Similarly, the frequencies from
SD5-input are equal to 2.47, 2.58 and 3.45 Hz. The frequencies of multi-input case
are equal to 2.59, 2.90 and 3.55 which were a little different from single-input cases.
Including 331 earthquake a total of 84 earthquake events (from 1999 to 2008) were
used to identify the modal frequencies under different reservoir water level. Finally,
the relationships between the reservoir water level and modal frequencies were
developed and regressed by using the curve fitting tool. As shown in the Figure 6, the
modal frequency can be calculated from the reservoir level according to the general
power-2 function:
f ( x ) = ax + c
b
(9)

where f is the modal frequency (in Hz) and x is the corresponding reservoir

level. It is observed that the modal frequencies of dam decreases when the reservoir
level increases.

4. SYSTEM IDENTIFICATION FROM AMBIENT VIBRATION DATA

System identification using seismic response data of the dam it is necessary to


determine whether single-input or multi-input should be used. From the study
described in the previous section, the identified results show a significant deviation.
But for ambient vibration measurement, the input information can be assumed as
white noise excitation. However, attention should be called to the fact that the small
level of input energy from ambient excitation does not always allow the induction of
measureable vibration in all of the frequency range of interest. For this reason, an
ambient vibration test was conducted to the dam on October 30, 2009, and the
corresponding reservoir water level is 164.0 m. There were seventeen positions
selected as the measurement locations of the ambient vibration test (AT1~AT17 as
shown in the Figure 1(a)). Measurement points AT1~AT7 were arranged on the top
abutment of the dam. Locations of the six strong motion instrumentation were also
included in this test (AT8~AT10 and AT12~AT14). Because of the limited number of
velocity sensors to measure all of the stations at the same time, it requires four steps
to complete the ambient vibration test to cover all the desirable measurement points.
Overlap on the deployment of sensor notes is required in order to extract the
vibration modes.
For ambient vibration test a total of eight VSE-15D velocity sensors (Tokyo
Sokushin Corporation) were used at the same time. Table 1 shows the specification
of the sensor. The ambient vibration data from Step2 is shown in Figure7. For data
from Step2 measurement, the peak velocity is 0.0054 cm/s at station AT4 and the
peak acceleration is 0.00052 g at station AT5. It is deserved to be mentioned, the
resolution of the sensor for ambient vibration test is 10-8 g and the resolution of
AC-63 accelerometer is about 10-5 g. Procedure to describe the number of sensors
and its covered range is shown below:

Step1 : AT1, AT2, AT3, AT4


Step2 : AT4, AT5, AT6, AT7, AT8, AT9, AT10, AT11
Step3 : AT8, AT9, AT10, AT11, AT12, AT13, AT14
Step4 : AT12, AT13, AT14, AT15, AT16, AT17

4.1 System identification using output-only stochastic subspace identification

Stochastic subspace identification algorithm was used to the ambient vibration


data. This method computes the state space model of the stochastic system using
output data. A stochastic system is similar to the Eq.(6) and (7) but its inputs are zero
mean, white vector sequences denoted as w and v :
x& = A x + w
c
(10)

&& = Cx + v
ut
(11)
The white noise excitation is the most important assumption of this theory which is
also the difficulty of the ambient vibration test. In order to ensure the quality of
signals, the measurement frequency range should be consistent with the assumption
(as white noise assumption). Interference from other source inputs must be
minimized (such as dam operation-induced vibration). The computation procedures
of stochastic subspace identification are very similar to the general subspace
identification expect for the application of orthogonal projection:

Y /Y = Γ X
f p
ˆ
i i
(12)

Following the procedure in subspace identification, the stability diagrams of the


system identification from each step can be attained, as shown in the Figure 8. These
stability diagrams from each test steps clearly mark out the locations of the dam’s
natural frequencies and the results are consistent with the dominant frequencies of
the Fourier spectrum of the responses.

Follow by the statistical analysis, Table 1 shows the mean and standard
deviation of the modal parameters for the first eight modes of the dam. High stability
for the identified modal frequencies was observed from the small value of the
calculated standard deviation. The global mode shapes can also be determined from
the measurement of each data set. Considering there is a normalization factor R
between each different experimental step [10]:

Rφ = R φ
i ik
n

j
n

jk
(13)

where i or j denotes the number of the experimental step, k is the number of

the sensor station and n is the number of mode. Eq.(13) shows that k station is

the one to be overlapped between step i and step j . Therefore, there are eight

equations available to determine the normalization factors in out experiment:

Rφ = R φ
1
n

14 2
n

24
, Rφ = Rφ
2
n

28 3
n

38
, Rφ = Rφ
2
n

29 3
n

39
, … (14)

Set R1 = 1, the above equations can be rewritten as the following form:

φ   φ 0
n n
0
R 
14 24

 0   −φ φ 0  
  
n n 2

 =
28 38
 R  (15)
 0   −φ φ 0  
n n 3

29 39

 M   M  R
M M 
4

The normalization factors can be solved by using pseudo-inverse and the global
mode shapes were determined. Figure 9 shows the global mode shapes from ambient
vibration test and the mode shapes estimated using seismic response data of 331
Earthquake are also shown.

5. CONCLUSION

In this study, three different approaches were used to identify the modal
properties of Fei-Tsui arch dam. For input/output subspace identification with
multi-input dynamic model, the earthquake force is idealized as a multiple support
excitation and the dam is considered as a lumped mass system with small damping.
Different from multi-input model, single-input model the uniform transmission of
earthquake force to the dam is assumed. The output-only stochastic subspace
identification uses the output data which should be generate from the white noise
excitation. From the study of the three approaches, it is concluded that the results
from the analysis of ambient vibration data will have a significant on the
improvement in the stability diagram by using the stochastic subspace identification.
In order to compare the accuracy of multi-input and single-input identifications, the
result from ambient vibration test was plotted as a circle in the Figure 6. Another
result from the forced vibration test was also plotted as the star in the same figure.
This force vibration data quoted from reference 4 which shows the first two modal
frequencies: 2.26 Hz and 3.02 Hz, respectively. The regression curves from
single-input identification accurately fit the results from both ambient vibration test
and force vibration test.

A brief conclusion can be made based on above comparisons. The output-only


stochastic subspace identification is the best approach to identify the modal
properties of dam in this study. It is suggested that an ambient vibration measurement
system may be a better solution for monitoring the dynamic characteristics of dam.
However, it needs continuous recording of ambient vibration to conform the
reliability of this approach. As for identifying the dam by using seismic data, further
study is required to prove the single-input identification is more accurate than
multi-input identification. Further more, this paper permit the identification of the
variation in modal frequencies with changes in reservoir level by using seismic data.
The influence of temperature should also be considered in the future since the
identification accuracy can be improved by using ambient data.

REFERENCES

1. Loh, C. H. and Tou, I. C. (1995) "A system identification approach to the


detection of changes in both linear and non-linear structural parameters,"
Earthq. Engrg. Struct. Dyn. 24,8597.
2. Safak, E. (1989) "Adaptive modeling, identification and control of dynamic
structural system, I: Theory," ASCE, J. Engq. Mech. 115, 2386-2405.
3. Alves, S.W., and Hall, J.F. (2006) “System identification of a concrete arch dam
and calibration of its finite element model.” Earthquake Engineering and
Structural Dynamics, 35, 1321-1337.
4. Daniell, W.E., and Taylor, C.A. (1999) “Effective ambient vibration testing for
validating numerical models of concrete dams.” Earthquake Engineering
and Structural Dynamics, 28, 1327-1344.
5. Darbre, G.R., de Smet, C.A.M., and Kraemer, C. (2000) “Natural frequencies
measured from ambient vibration response of the arch dam of Mauvoisin.”
Earthquake Engineering and Structural Dynamics, 29, 577-586.
6. Loh, C.H., and Wu, T.S. (1996). “Identification of Fei-Tsui arch dam from both
ambient and seismic response data.” Soil Dynamics and Earthquake
Engineering, 15, 465-483.
7. Loh, C.H., and Wu, T.C. (2000). “System identification of Fei-Tsui arch dam from
forced vibration and seismic response data.” J. Earthquake Engineering,
4(4),511-537.
8. Chopra, A.K. (2001) . Dynamics of structures – Theory and applications to
earthquake engineering, 2nd ed, Prentice Hall, New Jersey, 384-388.
9. Van Overschee, P., De Moor, B. (1996) Subspace identification for linear systems :
Theory - Implementation – Applications, Kluwer, Dordrecht.
10. Sim, S.H., Spencer, B.F., Jr., Zhang, M., and Xie, H. (2009) “Automated
decentralized smart sensor network for modal analysis.” Proc. SPIE, 7292,
72920W.

Table 1: Specification of tri-axial accelerometer and VSE-15D velocity sensor,


GeoSIG AC-63 tri-axial accelerometer VSE-15D velocity sensor
Full-scale ±2 g Full-scale ± 0.1 m/s
( 0.5, 1, 3, 4 g optional )
Type Force balanced Type Force balanced
accelerometer accelerometer
Sensitivity 10V/g Sensitivity 1000V/m/s
Dynamic > 120 dB Dynamic approximately 140 dB
Range Range
Bandwidth DC ~ 100 Hz Bandwidth 0.2 ~ 100 Hz
( 50 Hz or 200 Hz )
Damping 70 % of critical Damping 100.% of critical

Table 2. Natural frequencies and damping ratios from ambient vibration test
Figure 1. The Fei-Tsui arch dam; (a) Front view, (b) Photograph, (c)
Cross-sections at NPL1,NPL2 and NPL3.

Figure 2. Plot of seismic event triggered by the Fei-Tsui arch dam earthquake
monitoring system (from 1999 to 2008). The reservoir level is also
shown for each seismic event.
Figure 3. Setting direction of each station and the global coordinate system for system
identification.

Figure 4. Acceleration records (global Y-direction) and the corresponding Fourier


amplitude at stations SD1~SD5 from 331 Earthquake.
Figure 5. Stability diagrams of the multi-input case and signal-input cases (SD1 or
SD5) using seismic data of 331 Earthquake.
Figure 6. Regression analysis of the relationship between reservoir level and natural
frequencies of the Fei-Tsui arch dam. ( ☆ points the result from force vibration
test; ○ points the result from ambient vibration test)
Figure 7. Velocity records and the corresponding Fourier amplitude at stations AT4,
AT9, AT6 and AT11 from Step2 of ambient vibration test.

Figure 8. Stability diagrams from stochastic subspace identification.


Figure 9. Mode shapes of Fei-Tsui arch dam identified from seismic vibration data
(SD1-input case) and ambient vibration data.

You might also like