Engineering Structures 69 (2014) 183–193

Contents lists available at ScienceDirect

Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct

Structural health monitoring with statistical methods during progressive

damage test of S101 Bridge
Michael Döhler a,⇑, Falk Hille a, Laurent Mevel b, Werner Rücker a
a
BAM Federal Institute for Materials Research and Testing, Division Buildings and Structures, Unter den Eichen 87, 12205 Berlin, Germany
b
Inria, Campus de Beaulieu, 35042 Rennes, France

a r t i c l e i n f o a b s t r a c t

Article history: For the last decades vibration based damage detection of engineering structures has become an impor-
Received 27 April 2013 tant issue for maintenance operations on transport infrastructure. Research in vibration based structural
Revised 10 March 2014 damage detection has been rapidly expanding from classic modal parameter estimation to modern oper-
Accepted 11 March 2014
ational monitoring. Since structures are subject to unknown ambient excitation in operation conditions,
Available online 13 April 2014
all estimates from the finite data measurements are of statistical nature. The intrinsic uncertainty due to
finite data length, colored noise, non-stationary excitations, model order reduction or other operational
Keywords:
influences needs to be considered for robust and automated structural health monitoring methods. In this
Subspace methods
Operational modal analysis
paper, two subspace-based methods are considered that take these statistical uncertainties into account,
Uncertainty bounds first modal parameter and their confidence interval estimation for a direct comparison of the structural
Damage detection states, and second a statistical null space based damage detection test that completely avoids the iden-
Prestressed concrete bridge tification step. The performance of both methods is evaluated on a large scale progressive damage test of
a prestressed concrete road bridge, the S101 Bridge in Austria. In an on-site test, ambient vibration data of
the S101 Bridge was recorded while different damage scenarios were introduced on the bridge as a
benchmark for damage identification. It is shown that the proposed damage detection methodology is
able to clearly indicate the presence of structural damage, if the damage leads to a change of the struc-
tural system.
Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction the subject as well as an overview of many methods can be found

in [3–5]. Many of the currently used vibration-based methods for
The main objective of research work on Structural Health Mon- early damage detection performs modal system identification
itoring (SHM) systems is to develop or improve methodologies for and compare the obtained modal parameter of the actual state
an effectual identification and handling of risk implying deficien- with those of the undamaged state [6–10]. However, the sensitivity
cies of dynamic systems on the base of automated monitoring. of modal damage indicators is often not satisfying for a lot of real
Damages lead to changes in the modal parameters of a structure applications in civil engineering with ambient excitation and a lim-
(natural frequencies, damping ratios, mode shapes), hence one ited number of sensors. Varying operational or environmental
eminent component of a decision support framework is the ade- influences may have a greater impact on the dynamic properties
quate processing of measurement data to reliably detect such of a system than structural damage in early stages [10,11]. Besides,
changes in the systems under observation [1]. The detection of civil structures differ from many other mechanical structures by
damages is the first and essential step for automated SHM meth- their size and their uniqueness and therefore cannot undergo any
ods, before further levels in the whole damage identification pro- kind of prototype testing, as it is common for products from e.g.
cess – consisting of damage detection, localization, quantification automotive and aeronautic industry [12]. Also, they are subject
and prediction of remaining service life time [2] – can be treated. to operational and environmental loading of mostly large variances
Within the last 30 years a vast literature on vibration-based and are designed for an exceptional life span compared to mechan-
damage identification methods has emerged. An introduction to ical structures. For these reasons, civil structures are provided with
a high safety margin, which implies in general a high rigidity.
⇑ Corresponding author. Present address: Inria, Campus de Beaulieu, 35042 These facts make the development of vibration based strategies
Rennes, France. Tel.: +33 299842225; fax: +33 299847171. for damage detection in an early stage a complex and challenging
E-mail address: [email protected] (M. Döhler). task for civil structures. Methods that are successfully run on

http://dx.doi.org/10.1016/j.engstruct.2014.03.010
0141-0296/Ó 2014 Elsevier Ltd. All rights reserved.
184 M. Döhler et al. / Engineering Structures 69 (2014) 183–193

mechanical structures in a certain deterministic environment cally in the last decades [26]. It is shown that the proposed
might not at all be applicable on civil structures. methodology is able to clearly indicate the presence of structural
For the detection of damages in an early stage it is important to damage, if the damage leads to a change of the structural system.
take the uncertainty of damage related parameter estimates into Small damages which do not result in a system change when not
account to distinguish small (physical) changes from statistical activated by loading, do not lead to a modification of the dynamic
variability. Since usually only unknown ambient excitation and response behavior and for that cannot be detected with the pro-
no artificial excitation is available and practical for continuous posed global monitoring method.
and automated monitoring of structures, all parameters that are This paper is organized as follows. In Section 2, the basics of sto-
estimated from the output-only vibration measurements are af- chastic subspace identification are given. In Section 3, the confi-
flicted with statistical uncertainty. This uncertainty results from dence interval computation of modal parameters is described
the operation conditions, such as finite number of data samples, and in Section 4, the damage detection strategy is summarized.
measurement noises, non-stationary excitations, nonlinear struc- These algorithms are applied to S101 Bridge in Section 5, before
ture, model order reduction and further sources. finishing with some concluding remarks in Section 6.
Subspace methods have been proven to be very well-suited
for the output-only vibration analysis of civil structures in oper-
ation due to their excellent theoretical properties [13–15]. In this 2. Stochastic Subspace Identification (SSI)
paper two subspace-based methods for vibration monitoring are
considered that take the statistical uncertainties into account. Stochastic subspace identification methods are among the state
First, we use the covariance-driven subspace-based system iden- of the art methods for modal parameter estimation [14,15]. In this
tification [16,17] together with their confidence interval estima- section, an overview of the covariance-driven identification algo-
tion [18,19] for the operational modal analysis. Thanks to this rithm is given, whose properties are also the foundation of the sub-
uncertainty quantification, the significance of the modal parame- sequent damage detection algorithm.
ters and their changes during the monitoring period can be eval-
uated. Second, we perform the actual damage detection with a 2.1. Models and parameters
null space based statistical damage detection test [11,21–23],
where new data from a possibly damaged state is directly com- The behavior of a mechanical system is assumed to be described
pared to a reference state in a test statistic. Like this, the entire by a stationary linear dynamical system
structural response is compared without computing any modal M€zðtÞ þ Cz_ ðtÞ þ KzðtÞ ¼ v ðtÞ; ð1Þ
parameters. Comparing the test value to a threshold indicates if dd
the structure is damaged. where t denotes continuous time, M; C; K 2 R are the mass,
Both approaches are dealing with the SHM problem from the damping and stiffness matrices, high-dimensional vector z 2 Rd col-
identification and detection points of view. They are nonetheless lects the displacements of the degrees of freedom of the structure
related by the principles of subspace methods and the same funda- and the external force v is unmeasured and considered as noise.
mental uncertainty quantification computation from the output- The eigenstructure of (1) with the modes li and mode shapes (ob-
only measurement data. served eigenvectors) wi 2 Rr is a solution of
Our contribution to these methods concerns their applicability: detðl2i M þ li C þ KÞ ¼ 0; ðl2i M þ li C þ KÞ/i ¼ 0; wi ¼ L/i ; ð2Þ
Being elaborate methods but lacking some feasibility in practice,
rd
both methods were recently enhanced with a strongly decreased where matrix L 2 R maps the r sensor locations to the d degrees
computational burden, feasibility of high model orders and a more of freedom of the structure. Observing model (1) at the r sensor
numerically robust computation [17,19,23]. The results are fast locations (e.g. by acceleration, velocity or displacement measure-
algorithms that can be applied easily to vibration data from civil, ments) and sampling it at some rate 1/s yields the discrete model
mechanical or aeronautical structures. In this paper, they are dem- in state-space form
onstrated on the large scale progressive damage test of a pre- xkþ1 ¼ Axk þ v k
stressed concrete road bridge, the S101 Bridge in Austria [25]. In ; ð3Þ
yk ¼ Cxk þ wk
an on-site test, ambient vibration data of the S101 Bridge (see
Fig. 1) was recorded while different damage scenarios were intro- where A 2 Rnn is the state transition matrix, C 2 Rrn is the observa-
duced on the bridge as a benchmark for damage identification. The tion matrix, xk 2 Rn are the states of the system and yk 2 Rr the out-
verification of the effectiveness of SHM is particularly important on put measurements at the discrete times t = ks, where n is the system
such structures, especially since the effort for maintenance and re- order. The vectors vk and wk are the unmeasured input and output dis-
pair of the prestressed concrete road bridges increased dramati- turbances. The eigenstructure of system (3) is given by

Fig. 1. Bridge S101 during damage test [25].

 M. Döhler et al. / Engineering Structures 69 (2014) 183–193 185

detðA  ki IÞ ¼ 0; ðA  ki IÞ/i ¼ 0; ui ¼ C/i : ð4Þ The system is identified truncating in (9) at multiple model orders,
and frequencies from this multi-order system identification are
Then, the eigenstructure of the continuous system (1) is related
plotted against the model order [16,17]. From the modes common
to the eigenstructure of the discrete system (3) by
to many models and using further stabilization criteria, such as
esli ¼ ki ; wi ¼ ui : ð5Þ threshold on damping values, low variation between modes and
mode shapes of successive orders, etc., the final estimated model
From the eigenvalues ki , the natural frequencies fi and damping
is obtained. Like this, stabilization diagrams provide a GUI where
ratios ni of the system are directly recovered from
the user is assisted in selecting the identified modes of an investi-
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 gated structure.
a2i þ bi 100bi
fi ¼ ; ni ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi where
2ps 2
a2i þ bi 3. Confidence intervals of modal parameters
 
 Imðki Þ
ai ¼ arctan ; bi ¼ ln jki j: ð6Þ
Reðk Þ  i
The uncertainty quantification of the estimated modal parame-
ters that are obtained from subspace identification in the previous
section is necessary to assess the confidence one can have in these
2.2. Covariance-driven SSI values, e.g. when comparing the modal parameters of different
states of a structure. Modal parameters with little confidence
To obtain the eigenstructure of system (3) from measurements (and hence large confidence intervals) are little useful for compar-
(yk)k=1,. . .,N, the stochastic subspace identification algorithm is used ing structural states.
in its covariance-driven form [13,16]. In the case of a high number The uncertainties of the modal parameters at a chosen system
of available sensors r, a subset of r0 6 r sensors can be chosen as order can be computed from the uncertainty of the Hankel matrix
reference sensors (or so-called projection channels) in the output H by doing a sensitivity analysis. Since the modal parameters are
ðrefÞ
vector yk 2 Rr , defining the reference output vectors yk 2 Rr0 . functions of the Hankel matrix, it holds
Selecting reference sensors increases the computation time and
improves the quality of the results [16]. Then, the output Dfi ¼ J fi Dðvec HÞ; Dni ¼ J ni Dðvec HÞ; Dui ¼ J ui Dðvec HÞ; ð11Þ
correlations for (theoretical) first-order perturbations of the frequencies fi,
damping ratios ni and mode shapes ui, where J denotes their sen-
1X N
ðrefÞT
Ri ¼ y y 2 Rrr0 ð7Þ sitivity with respect to the stacked Hankel matrix vec H, as detailed
N k¼1 kþi k
in [18–20]. The uncertainty of the stacked Hankel matrix is quanti-
are computed for the lags i = 1, ..., p + q, where the parameters p and fied by its covariance RH, which can be easily evaluated by splitting
q are chosen such that min {pr, qr0} P n, with in general p + 1 = q. the available sensor data into blocks on which instances of the Han-
With these correlations, the Hankel matrix kel matrix are obtained. Then, RH is computed from their sample
0 1 covariance. It follows
R1 R2  Rq
B R2 R3    Rqþ1 C covðfi Þ ¼ J fi RH J Tfi ; covðni Þ ¼ J ni RH J Tni ; covðui Þ ¼ J ui RH J Tui ; ð12Þ
B C
H¼B
B .. .. ..
C
C ð8Þ
@ . . . A from where the standard deviations of the modal parameters are
obtained. This offers a possibility to compute confidence intervals
Rpþ1 Rpþ2 Rpþq
on the modal parameters directly in a fully automated way, without
is filled, which possesses the factorization property H = O C into repeating the system identification step. In [19] an efficient and fast
observability matrix O and the stochastic controllability matrix C. computation scheme of the uncertainty computation laid out in
The observability matrix O is obtained from H by a singular value (12) is described in detail, where confidence bounds up to high
decomposition (SVD) and truncation at the desired model order n model orders can be computed within reasonable computation
0 1 times. The efficient computation at multiple model orders for stabil-
C ization diagrams is also addressed in [19].
  B CA C
D1 def B C
H ¼ ð U1 U0 Þ V T ; O ¼ U 1 D1=2
1 ¼B B .. C 2 Rðpþ1Þrn
C
D0 @ . A 4. Statistical subspace-based damage detection
p
CA
Instead of performing damage detection by comparing the
ð9Þ
modal parameters from a new data set to a reference state, the sta-
From the observability matrix O the matrices C in the first block tistical subspace-based damage detection algorithm [21,22] com-
row and A from a least squares solution of pares different structural states by using a v2 test on a residual
1 0 1 0 function. This residual confronts the left null space of a subspace
C CA
B CA C B CA2 C matrix H0 in the reference state with a subspace matrix H of a
B C B C new, possibly damaged state. This algorithm has the advantage
OA ¼ O with O ¼ B C B
B .. C; O ¼ B .. C
C ð10Þ
@ . A @ . A that no modal parameter extraction is necessary, while the v2 test
checks for changes implicitly. Furthermore, the whole system re-
CAp1 CAp
sponse is compared implicitly and not only the first modes as it
are obtained. The eigenstructure (ki, ui) of the system (3) and the is often the case when monitoring modal parameters.
modal parameters are finally obtained from (4)–(6). The algorithm consists of the following steps. In the reference
In Operational Modal Analysis, the model order n in (9) is un- state, a subspace matrix H0 is computed from the output data of
known and spurious modes appear due to the unknown noise the system as in (8). Note that reference scenarios with different
characteristics in the ambient excitation. Based on the observation environmental conditions can be mixed when computing H0, see
that physical modes remain quite constant when estimated at dif- [11], such that the damage detection test will be robust to these
ferent over-specified model orders, while spurious modes vary, environmental conditions. Then, the left null space S of H0 is com-
they can be distinguished using so-called stabilization diagrams. puted and the characteristic property of a system in the reference
186 M. Döhler et al. / Engineering Structures 69 (2014) 183–193

state writes STH = 0. The null space S can be obtained with an SVD Retrospectively, some of the major design assumptions proofed
of H0 as in (9) from S = U0. The associated residual vector [11,21,22] to be erroneous and after short periods of operation significant
writes and characteristic damage patterns occurred at the structure
pffiffiffiffi [26,27]. The load bearing capacity and especially the durability of
fN ¼ NvecðST HÞ; ð13Þ the bridges remained mostly limited despite costly retrofitting
where N is the number of samples, on which H is computed. Fur- activities [28,29]. In addition, an increase of heavy load vehicle
thermore, compute the residual’s covariance R = E[fNfTN] in the ref- traffic has been ongoing since the start of operation. Thus, pre-
erence state, where E denotes the expectation operator. This is done stressed bridge structures in central Europe and other countries
by cutting the data into blocks and computing the residual on each have been exposed to degradation processes initiated by poor de-
of the data blocks, analogous to RH in the previous section, and it sign, while the operational loading and the associated dynamic
holds R = N (I  ST) RH (I  S), where I is the identity matrix and stressing increases steadily in the same time [30,31].
 denotes the Kronecker product. The covariance of the null space
S is not computed, since it was neglected in the derivation of the 5.1.2. Damage description
original detection methods [21,22], where S is assumed to be mod- The progressive damage test took place from 10–13 December
el-based and not data-based. Note that while case studies (e.g. in 2008. During the test the highway beneath the bridge was open
[11,23]) performed well without using the covariance of S, it should in one direction. The second direction was closed for traffic be-
be taken into account in future research to increase the robustness cause of construction works which in addition took place near
of the detection method, requiring a non-trivial extension of the ap- the bridge.
proach (see e.g. [24]). Two major damage scenarios were artificially induced. First, a
The residual vector fN is asymptotically Gaussian with mean damage of one of the four columns was inserted by cutting through
zero in the reference state. It manifests itself to damage by a the column on its lower end. With this action a change in the glo-
change in its mean value, corresponding to an increase of the v2 bal structural system was implemented since an extra hinge was
test formed just above the foundation. During the cutting process no
signs of extra cracking (noise emission) were observed. After a sec-
v2 ¼ fTN R1 fN : ð14Þ
ond cut a 5 cm thick slice of the column was removed and the col-
The monitoring of the system is achieved by calculating the residual umn was lowered for altogether 3 cm, while moderate noising was
vector fN in (13) on the Hankel matrices estimated from newly re- noticed. At the neighboring column a horizontal crack occurred.
corded output data and subsequently the v2 test value in (14), Additional cracking in the bridge deck was not observed. The col-
which is compared it to a threshold. Such a threshold is obtained umn was settled until reaching elastic limits of the bridge deck. Be-
empirically from the v2 test values of several data sets in the refer- cause of the hydraulic jack there was no complete loss of support.
ence state of the structure for a desired type I error. A significant in- The column was secured in its location by inserted steel plates.
crease in the v2 value indicates that the system is no more in the Afterwards the column was uplifted again to its original position
reference state. and secured there again by steel plates.
The residual covariance matrix R is usually high-dimensional Second, prestressing tendons of one of the beams were cut suc-
and its inversion R–1 costly and numerically fragile. In [23] the cessively for a further damage scenario. All in all three and a quar-
authors proposed an efficient and numerically robust computation. ter of a wire bundle were cut through. Between each intersection
To further increase the robustness of the computation, while pauses of several hours were kept to let the structural system
decreasing the computation time, so-called reference sensors or pro- change into a new state of equilibrium, while no indication of
jection channels [16] are used in the computation of the subspace changing structural conditions like cracking was observed. Since
matrix as indicated in Section 2.2, reducing the size of the involved the bridge was not loaded by traffic, it is assumed that cutting
matrices significantly. With these improvements a fast and robust the limited number of tendons reduced only the extra prestressing
computation of the damage indicator (14) was made. margin.
Pictures of both scenarios of the progressive damaging are
shown in Fig. 2. In Table 1 all damage actions are sorted in chrono-
5. Monitoring of S101 Bridge during progressive damage test
logical order.
5.1. Progressive damage test of S101 Bridge
5.1.3. Measurement description
Within the European research project ‘‘Integrated European For the vibration measurements a BRIMOSÒ measurement sys-
Industrial Risk Reduction System (IRIS)’’ a prestressed concrete tem with a permanent sensor grid was used. The grid consisted of
bridge was artificially damaged [25]. The intention was to provide 15 sensor locations on the bridge deck (see Fig. 3), in each location
a complete set of monitoring data during a defined loss of struc- three sensors for measurements in the bridge deck’s vertical, longi-
tural integrity for testing and evaluation of various SHM methods tudinal and transversal direction. Altogether, 45 acceleration sen-
and applications. Therefore, the static and dynamic behavior of sors were applied. The sensor layout was motivated by the
the structure was measured permanently during the 3-day damag- symmetry of the bridge, so that 14 locations were set only on
ing process. The progressive damage campaign was planned and one side of the bridge deck, while the 15th location on the other
organized by the Austrian company VCE. The characteristics of side has the purpose to distinguish between bending and torsional
the structure under observation, the measurement campaign and modes.
the introduced damages are presented briefly. The data were recorded permanently with a sampling fre-
quency of 500 Hz. During the three days measurement campaign
5.1.1. The S101 Bridge 714 data files each containing 45 channels with 165 000 data
The S101 Bridge was a prestressed concrete bridge from the points were produced, corresponding to 5.5 min of measurements
early 1960s and therefore a characteristic representative of the each. Note that there were hardly any temperature changes during
partly invalid highway infrastructure asset in Europe. Despite a the measurement period, where misty winter weather just below
general lack of experience in the time of their design, prestressed freezing was dominant, so nuisance from temperature changes
concrete was a very popular construction type in those days. can be excluded in the measurements.
 M. Döhler et al. / Engineering Structures 69 (2014) 183–193 187

Fig. 2. Progressive damage of column and tendons on bridge S101 [25].

Table 1
Notation of consecutive damage actions and their effects with time index after start of the measurements.

Time index Damage action Damage effect

A 14 h 52 m Begin of cutting through column  No signs of extra cracking (noise) or increase of existing cracks
B 16 h 15 m End of second cut through column  Formation of an extra hinge just above the foundation, which itself is a
constructive fixed support
C 17 h 43 m Lowering of column – 1st step (10 mm)  Moderate noise
D 18 h 44 m Lowering of column – 2nd step (20 mm)  Horizontal crack in neighboring column
E 20 h 45 m Lowering of column – 3rd step (27 mm)  Settling of bridge deck until reaching the elastic limits, no complete loss of
support because of hydraulic jack
F 21 h 40 m Inserting steel plates
G 1d 15 h 15 m Uplifting column  Closing of occurred cracks
 Hinge from cutting remains
H 1d 19 h 45 m Exposing cables and cutting of 1st cable  Reduction of prestressing without indication of changing conditions
I 1d 22 h 25 m Cutting through 2nd cable  No influence of structural behavior since bridge is not loaded by traffic
J 2d 15 h 02 m Cutting through 3rd cable
K 2d 17 h 14 m Partly cutting of 4th cable  The extra prestressing reservoir is depleted

Fig. 3. Location of the sensors on the bridge deck.

5.2. Modal parameters and their confidence intervals in Section 3. For the covariance estimate on the Hankel matrix,
the data was cut into 100 blocks.
Being interested in the identification of the first modes in a fre-
quency range [0–18 Hz], the data was downasampled and deci-
mated from sampling rate 500 Hz by factor 5. System 5.2.1. System identification results in reference state
identification and the covariance computation was done with the First, a modal analysis of the structure in the reference state
covariance-driven subspace identification detailed in Section 2 with the computation of the uncertainty bounds of the modal
with parameters p + 1 = q = 35 at model orders n = 1, . . ., 100. All parameters as in Section 3 was performed. In Fig. 4(a), the stabil-
r = 45 sensors were used and r0 = 3 reference sensors were chosen. ization diagram of the natural frequencies vs. the model order is
The covariance computation on all identified modes at the differ- presented, where a confidence interval (±one standard deviation
ent model orders was accomplished with the strategy explained rf) of each frequency is plotted as a horizontal bar.
188 M. Döhler et al. / Engineering Structures 69 (2014) 183–193

(a) all modes (b) with threshold on uncertainty bounds

Fig. 4. Stabilization diagrams containing the natural frequencies of the first five modes with their ±rf confidence intervals (horizontal bars).

In Fig. 4(a) it can be seen that the (true) structural modes seem Table 2
to have much lower uncertainty bounds than spurious modes. Overview of the estimated first 5 modes with natural frequencies f, their variation
coefficient r ~ f ¼ rf =f  100, the damping ratios n and their variation coefficient
With this observation, a threshold of 1% was put on the variation
r~ n ¼ rn =n  100.
coefficient of the frequencies (standard deviation divided by fre-
quency), which leads to a much clearer diagram in Fig. 4(b). Such Mode f (Hz) r~ f n (%) r~ n
a cleaned diagram is also easier to evaluate for an automated mod- 1 4.036 0.12 0.78 15
al parameter extraction. 2 6.281 0.08 0.56 20
3 9.677 0.18 1.3 14
In Fig. 5, the frequency and damping ratio of first mode identi-
4 13.27 0.13 1.5 13
fied at the different model orders is magnified for a better visibility 5 15.72 0.37 1.3 17
of the confidence bounds. Using the information from these confi-
dence bounds, values for the frequency and the damping ratio of a
mode can be chosen that are optimal for different model orders, as
the 15th sensor was very close to the middle of the bridge, the dis-
the true model order is unknown.
tinction of the mode shapes 3 and 4 was not easy due to symmetry.
From the stabilization diagrams, the modal parameters are cho-
However there is a small but significant shape difference of more
sen. In Table 2, the system identification results with their varia-
than the two standard deviations at the 15th sensor between
tion coefficients are summarized. As expected, the variation
modes 3 and 4, where the former is negative and the latter is po-
coefficients of the frequency estimates are very low (lower than
sitive, while the mode shape at the other sensor in the middle of
0.5%), while the estimates of the damping ratios show much higher
the bridge deck is negative in both cases in Fig. 6.
variation coefficients (up to 20% in this case).
In Fig. 6, the real parts of the obtained mode shapes in the ver-
tical direction at the 14 sensors of one side of the bridge deck are 5.2.2. System identification results in maximal damaged state
displayed with their uncertainty bounds (±2 standard deviations). In the next step, the modal analysis with the confidence interval
From the 15th sensor on the other side of the bridge deck, whose computation was performed in the maximal damaged state on
contribution is marked as the red point in Fig. 6, information about an example record after the settling of the column and the
the kind of the mode is obtained: Modes 1, 3 and 5 are vertical insertion of the steel plates, corresponding to damage event F in
bending modes and modes 2 and 4 are torsional modes. Since Table 1. The analysis is performed with the confidence interval

Fig. 5. Zoom on first mode: estimated natural frequencies (left) and corresponding damping ratios (right) at different model orders.
 M. Döhler et al. / Engineering Structures 69 (2014) 183–193 189

mode 1, 4.036 Hz, 0.78% mode 2, 6.281 Hz, 0.56% mode 3, 9.677 Hz, 1.3%

mode 4, 13.27 Hz, 1.5% mode 5, 15.72 Hz, 1.3%

Fig. 6. First five mode shapes in vertical direction on one side of the span (line) and the sensor on the other side of the bridge deck (red point) with their uncertainty bounds.

computation analogously to the previous section and the results was built for each dataset. In the diagrams, only modes in accor-
are presented in Table 3 and Fig. 7. dance with reasonable thresholds for the damping estimates and
In comparison to the modes in the reference state in the previ- confidence interval bounds were preselected. The automatic mode
ous section, it can be noted that the mode shapes have changed selection was then performed based on a simple statistic clustering
significantly especially at the location of the bridge that was set- algorithm, where each cluster contains modes with small devia-
tled (see Fig. 3). These changes are higher than the associated con- tions of the frequency, damping and MAC values between succes-
fidence bounds. They are also confirmed by a comparison of the sive model orders. Clusters with a minimal number of modes were
modal assurance criterion (MAC) between the modes in Fig. 8. chosen and the cluster centers were used as the final results. Note
Mode 3, which was a vertical bending mode in the reference state, that it was not our purpose to develop refined automated Opera-
is not present anymore. Instead, two torsional modes appear that tional Modal Analysis algorithms (as e.g. in [10,32]), but to demon-
are denoted as 3a and 3b, where mode 3b has no similarity to strate the confidence interval computation of the modal
any of the modes from the reference state as can be seen in parameters during the progressive damage test. In this context,
Fig. 8. Note that modes 3 and 4 from the reference state are hardly no tracking of the individual modes was realized, which would also
distinguishable in the comparison of MAC values. This is because prove to be difficult given the strong changes in the modes and
they are bending and torsion modes that only show a difference even in the number of modes, as described in the previous section.
at the one sensor on the other side of the bridge deck, where how- The results of the frequency monitoring of all datasets during
ever the deflection is close to zero for these modes. Torsional mode the progressive damage test are shown in Fig. 9, where the respec-
4 and bending mode 5 in the damaged state show similar behavior tive damage scenarios were explained in Table 1. Especially the fre-
as in the reference state with changes around the settled column. quency change can be clearly seen when one column of the bridge
Concerning the natural frequencies in comparison to the reference was lowered before it was lifted up again (between A and G). This
state, it can be observed that modes 1 and 2 show a frequency drop affected mainly the second, third and fourth mode, while the fre-
(0.5% and 6%, respectively), while the frequencies of modes 4 and 5 quency changes in the first mode were less significant. An evalua-
show an increase (7% and 3%, respectively). The damping ratios of tion of the changes in the fifth mode is difficult due to its high
the modes 1, 2, 4 and 5 have increased. uncertainty in the automated evaluation, but its change was shown
Results on system identification during all damage scenarios are on a dataset in the damaged condition in the previous section. The
presented in the next section. change in the third mode is remarkable, as it was replaced by two
modes during the lowering of the column, with one lower and one
higher frequency than before.
5.2.3. Frequency monitoring during progressive damage test The change in the frequencies when cutting the tendons (be-
During the progressive damage test of the S101 Bridge, more tween G and K) is not significant. Only after the uplifting of the col-
than 700 datasets were available. Some of them contained errone- umn and before cutting the first tendon (between G and H), some
ous data due to destruction work on the bridge or other influences of the frequencies are dropping, probably due to the settling of the
that were excluded from the analysis. On the left 680 datasets, an structural system after the uplifting. However, no significant fre-
automated monitoring procedure was applied, which did the sys- quency change can be observed afterwards.
tem identification and confidence interval computation automati-
cally for each dataset. This means that a stabilization diagram
5.3. Damage detection

Table 3 5.3.1. Data analysis

Overview of the first modes in the damaged state F (notation see Table 2).
The vibration measurements were recorded throughout the
Mode f (Hz) r~ f n (%) r~ n whole three day long damage test, including the nights. The mea-
1 4.017 0.29 1.0 36
surement campaign started approximately 12 h before the damage
2 5.891 0.27 1.2 25 of the column was executed to monitor the undamaged state for an
3a 9.347 0.28 0.97 29 adequate time period.
3b 10.89 0.33 2.1 29 The reference state of the undamaged structure was set up by
4 14.25 0.56 1.9 22
computing and averaging the reference matrices from several data-
5 16.25 0.83 2.7 7.3
sets. With this approach [11], a possible disturbance by single
190 M. Döhler et al. / Engineering Structures 69 (2014) 183–193

mode 1, 4.017 Hz, 1.0% mode 2, 5.891 Hz, 1.2% mode 3a, 9.347 Hz, 0.97%

mode 3b, 10.89 Hz, 2.1% mode 4, 14.25 Hz, 1.9% mode 5, 16.25 Hz, 2.7%

Fig. 7. Mode shapes in range [0–18 Hz] in vertical direction on one side of the span (line) and the sensor on the other side of the bridge deck (red point) with their uncertainty
bounds.

excitation events or different environmental conditions is mini-

mized. The size of the Hankel matrix depends on the number of
chosen output channels as well as on the time lags, to be accounted
for computing the output covariance matrices.
The size of the analyzed null space S depends on the order of the
system and can be generally determined by the rank of the matrix.
Because of noise, the singular values will not drop to exactly zero
and the system order must be chosen. The measured response sig-
nals of S101 Bridge were dominated by around ten modes, which
led us to a choice of n = 20 as the system order. For calculation of
the residual covariance matrix R it is important to use as much
datasets as possible for a precise estimation. It is convenient to di-
vide available datasets into subsets to provide the necessary
amount of data sets, whereas the number of data points for one
residual vector should not be too small to be significant.
In the test stage, the v2 test statistics in (14) is computed for
every data set, which in real time means an indicator of damage
Fig. 8. MAC values between the mode shapes from the reference and damaged for every 5.5 min. Note that no modal parameter estimation or
states. tracking is necessary for this damage detection approach.

Fig. 9. Natural frequencies with confidence bounds during progressive damage test of S101 Bridge. The color bar indicates the confidence bound in percent of the obtained
frequency. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
 M. Döhler et al. / Engineering Structures 69 (2014) 183–193 191

5.3.2. Results of damage detection x 10

7
4
Fig. 10 shows a bar plot of values as damage indicators of all
5

consecutive tests within the three days campaign, where all 45 8

x 10

sensors (r = r0 = 45) were used. For the computation of the residual

Damage Indicator
6

3
covariance matrix 100 datasets of the undamaged state were used. 5

Damage Indicator
4

The abscissa of the plot describes the chronological sequence of the 3

damage activities as noted in Table 1 as well as the 6 am and 6 pm 1

2
points of time for orientation. 0
A B
Time

The reaction of the damage detection test at the main damage

events can be easily observed. From the measurements in the ref- 1
erence state (before damage event A) a threshold of the damage
indicator can be set easily, such that all the subsequent scenarios
after damage event B would be correctly classified as damaged. 0
This qualifies the proposed method for damage detection and in A B C D E
the following the obtained results are analyzed in more detail. Time
It can be seen that the damage indication is interfered by noise
Fig. 11. Damage indicator for cutting and settling of one bridge column, detail of
in the ambient excitation of the bridge. The v2 values periodically
damage indicator for cutting through one bridge column.
swell up in the morning and ebb away in the evening. It is assumed
that the traffic going underneath the bridge and/or the construc-
did not drop completely to its original value, which is certainly
tion work nearby are the source of the noise. Since the measure-
due to the fact that the lowering of one column has led to cracking
ments in the reference state were only taken during 12 h in the
within the concrete structure to some extent and hence to a change
night, an extension of the measurement time period of the undam-
of the dynamic signature of the system.
aged state to a whole day circle might reduce these disturbances
significantly when taken into account in the residual covariance
matrix. Influences by solar radiation and/or temperature alterna- 5.3.4. Detection of tendon damage
As one can see in Fig. 10 the cutting of the prestressing tendons
tion can be excluded, since during the 3-days campaign misty win-
ter weather with only moderate temperature changes just below did not lead to a significant change in the damage indicator after
the single cutting steps. Nevertheless, a distinctive increase of
freezing was dominant.
the indicator could be observed at the end of the measurement.
Fig. 12 shows the last time period in detail. It is possible that a
5.3.3. Detection of column damage
Fig. 11 shows the damage indicator during the several steps of change of the bridge’s structural system took place with a time de-
lay after cutting partly the fourth tendon.
the first damage scenario, the cutting and settling of one of the four
bridge columns. With exception of the time periods of the direct For a prestressed concrete structure the loss of prestressing is a
major damage which comes along with a significant loss of its load
mechanical destruction processes, the displayed sequence of dam-
age indicators has a consecutive course. bearing capacity.
One reason that the cutting of the tendons does not affect the
As can be seen in Fig. 11, the three steps of the column settle-
proposed damage indicator might be that the overall prestressing
ment action are very distinctive in their influence on the computed
is designed for combinations of dead and traffic loads. Since the
damage indicator. Obviously, the dynamic system has changed to
bridge deck of S101 has a quite slender cross section, the dead load
quite some extent and the elastic settlement of altogether
is not that high in comparison with the maximum design traffic
27 mm can be clearly detected. Although not that strongly visible,
loading. An additional dead load reduction comes from the re-
the cutting of the column (A + B) also caused an increase of the
moval of the asphalt surface before the damage test. Furthermore
indicator of approximately 75%. However it has to be mentioned
it has to be recalled that the prestressing was designed with ade-
that the absolute effect is superimposed by noise effects. The inlet
quate safety margins. Also, the cutting results only in local loss
in Fig. 11 shows a detail of that time period.
of prestressing because of the bound between tendons and con-
The column remained in the settled condition for approxi-
crete. For these specified reasons it is assumed that the cutting of
mately one day and was then uplifted again in its former position
the tendons during the damage test did not lead to a significant
(event G). The effect of the uplifting is again clearly visible in
change of the global structural system, since the loading of the
Fig. 10 by a drop of the damage indicator. However, the indicator

6
7
x 10
x 10
10
4
Damage Indicator
Damage Indicator

5
2

0
6am ABCD EF 6pm 6am G H I 6pm 6am J K J K
Time Time

Fig. 10. Course of the damage indicator over the 3-day damage test. Fig. 12. Detail of damage indicator at the end of the tendon cutting process.
192 M. Döhler et al. / Engineering Structures 69 (2014) 183–193

7
3.5 x 10
applied to output-only vibration data from the progressive damage
test of the S101 Bridge, where it was shown how confidence inter-
3 vals of the modal parameters improve the assessment of system
identification results. With the recent improvements made on
2.5
the method, confidence intervals can now be efficiently computed
Damage Indicator

2
up to high model orders. Together with an automated modal anal-
ysis (e.g. as in [10]) the confidence intervals provide a supplemen-
1.5 tary tool to analyze the structural condition individually. In the
progressive damage test, a clear link between the change in the
1 modal parameters and the destruction states can be made.
With the statistical damage detection test, the change in the
0.5
system response of the entire structure is evaluated without the
0 need for an individual modal analysis. The link between the struc-
6am ABCD EF 6pm 6am G H I 6pm 6am J K
tural changes due to the artificially introduced damage cutting/
Time lowering of one column and the behavior of the damage indicator
Fig. 13. Damage indicator on S101 with four projection channels.
at each test stage was clearly shown. For a second damage sce-
nario, the cutting of single prestressing tendons, an early stage
change of the dynamic response behavior could not be indicated.
structure was not high enough to activate the damage right after It is presumed that the locally acting damage did not activate a sig-
its insertion. nificant change of the overall structural system due to the absence
The increase of the indicator at the end of the test series might of operational loading during the test. Under the assumption of an
be the result of a delayed rearrangement of the structural system adequate preliminary monitoring of the undamaged structure and
by a reduction of stresses under generation of cracks in the con- taking into consideration the preceding assertion about damage
crete of the bridge deck. Though, an evidence of that assumption, activation, the statistical damage detection method proves to be
for instance by an increase of the measured bridge deck deflection, feasible for structural health monitoring of civil engineering
could not be found. structures.
In our case study no relevant temperature or other environmen-
tal changes were present, so the robustness of the presented meth-
5.3.5. Projection channels
ods to such changes could not be shown. Note that the rejection of
With the application of so-called reference sensors or projection
environmental influences in the modal parameters was e.g. done in
channels (see Section 2.2) the computation costs of the analysis of
[10], and the confidence interval computation in the context of
the damage indicators was reduced massively. The full analysis of
such a rejection approach would be an interesting extension of
all 45 channels required an extent of memory space which com-
the presented approach for health monitoring through operational
monly is not provided on desktop computers.
modal analysis. The rejection of environmental influences for the
On S101 several numbers and constellations of projection chan-
statistical damage detection approach was already described and
nels were analyzed. Fig. 13 shows a v2 value plot for only r0 = 4
successfully performed in [11]. Thus, our presented statistical
well-chosen projection channels (the fifth sensor from the right
methods provide a powerful basis for structural health monitoring
with all three directions and the 15th sensor on the other side of
tools that can be efficiently and reliably used on civil engineering
the bridge deck in the vertical direction, see also Fig. 3). As one
structures.
can see, almost equal information about the damage indication
could be achieved compared to Fig. 10, and at the same calculation
the computing time was cut to a fifth compared to a complete sen- Acknowledgements
sor analysis.
This work was partially supported by the European project FP7-
PEOPLE-2009-IAPP 251515 ISMS. We also thank VCE for providing
6. Conclusion the data from S101 Bridge.

In this work, two approaches have been presented for the mon-
References
itoring of dynamic systems. Both are based on stochastic subspace
identification and take the uncertainties in the output-only vibra- [1] Worden K, Farrar CR, Manson G, Park G. The fundamental axioms of structural
tion data into account. With the first approach, modal parameter health monitoring. Proc Roy Soc A: Math, Phys Eng Sci 2007;463(2082):
estimates and their confidence intervals are computed, which are 1639–64.
[2] Rytter, A. Vibrational based inspection of civil engineering structures. PhD
essential for knowing the significance of the obtained estimates. thesis, Aalborg University, Denmark; 1993.
With the second approach, a statistical damage indicator is com- [3] Farrar CR, Doebling SW, Nix DA. Vibration-based structural damage
puted to automatically evaluate changes in the structural response. identification. Philos Trans Roy Soc London. Ser A: Math, Phys Eng Sci
2001;359(1778):131–49.
Both approaches represent in fact the same paradigm, with the dif- [4] Carden EP, Fanning P. Vibration-based condition monitoring: a review. Struct
ference that the first approach evaluates the uncertainty in each Health Monitor 2004;3(4):355–77.
modal parameter individually to be able to compare them directly, [5] Fan W, Qiao P. Vibration-based damage identification methods: a review and
comparative study. Struct Health Monitor 2011;10(1):83–111.
and the comparison is still to be done by the engineer, while in the [6] Kullaa J. Damage detection of the Z24 bridge using control charts. Mech Syst
second approach the change in the modal parameters is automat- Signal Process 2003;17(1):163–70.
ically assessed and quantified by the implicit use of these uncer- [7] Magalhães F, Cunha Á, Caetano E. Dynamic monitoring of a long span arch
bridge. Eng Struct 2008;30(11):3034–44.
tainties in a damage detection index.
[8] Ramos LF, Marques L, Lourenço PB, De Roeck G, Campos-Costa A, Roque J.
Newly developed improvements and extensions of the basic Monitoring historical masonry structures with operational modal analysis:
methodologies are integrated to increase numerical robustness, two case studies. Mech Syst Signal Process 2010;24(5):1291–305.
decrease computation efforts as well as to refine results. With [9] Rainieri C, Fabbrocino G, Manfredi G, Dolce M. Robust output-only modal
identification and monitoring of buildings in the presence of dynamic
these improvements, both theoretically profound methods can be interactions for rapid post-earthquake emergency management. Eng Struct
easily used in practical applications. They were successfully 2012;34:436–46.
 M. Döhler et al. / Engineering Structures 69 (2014) 183–193 193

[10] Magalhães F, Cunha A, Caetano E. Vibration based structural health monitoring [22] Basseville M, Mevel L, Goursat M. Statistical model-based damage detection
of an arch bridge: from automated OMA to damage detection. Mech Syst Signal and localization: subspace-based residuals and damage-to-noise sensitivity
Proc 2012;28:212–28. ratios. J Sound Vib 2004;275(3–5):769–94.
[11] Balmès É, Basseville M, Bourquin F, Mevel L, Nasser H, Treyssède F. Merging [23] Döhler M, Mevel L, Hille F. Subspace-based damage detection under changes in
sensor data from multiple temperature scenarios for vibration monitoring of the ambient excitation statistics. Mech Syst Signal Process 2014;45(1):207–24.
civil structures. Struct Health Monitor 2008;7(2):129–42. [24] Dong J, Verhaegen M, Gustafsson F. Robust fault detection with statistical
[12] Brownjohn JMW. Structural health monitoring of civil infrastructure. Philos uncertainty in identified parameters. IEEE Trans Signal Process 2012;60(10):
Trans Roy Soc A: Math, Phys Eng Sci 2007;365(1851):589–622. 5064–76.
[13] Benveniste A, Fuchs JJ. Single sample modal identification of a non-stationary [25] VCE. Progressive Damage Test S101 Flyover Reibersdorf (draft), Tech. Report
stochastic process. IEEE Trans Autom Control 1985;AC-30(1):66–74. 08/2308; 2009.
[14] Van Overschee P, De Moor B. Subspace identification for linear systems: [26] Zilch K, Weiher H. Assessment of the condition of prestressed or post-
theory, implementation, applications. Kluwer; 1996. tensioned German concrete bridges (in German). Bauingenieur 2007;82(1):
[15] Benveniste A, Mevel L. Nonstationary consistency of subspace methods. IEEE 14–24.
Trans Autom Control 2007;52(6):974–84. [27] Polder RB, Peelen WHA, Courage WMG. Non-traditional assessment and
[16] Peeters B, De Roeck G. Reference-based stochastic subspace identification for maintenance methods for aging concrete structures – technical and non-
output-only modal analysis. Mech Syst Signal Process 1999;13(6):855–78. technical issues. Mater Corros 2012;63(12):1147–53.
[17] Döhler M, Mevel L. Fast multi-order stochastic subspace-based system [28] Grünberg J, Hansen M, Liebig JP. Fatigue life of concrete bridges under
identification. Control Eng Practice 2012;20(9):882–94. consideration of increasing traffic loads (in German). Beton-und
[18] Reynders E, Pintelon R, De Roeck G. Uncertainty bounds on modal parameters Stahlbetonbau 2007;102(9):596–606.
obtained from stochastic subspace identification. Mech Syst Signal Process [29] Khan M. Bridge and highway – structure, rehabilitation and repair. New
2008;22(4):948–69. York: McGraw-Hill; 2010.
[19] Döhler M, Mevel L. Efficient multi-order uncertainty computation for stochastic [30] Brady KC, O’Reilly M, Bevc L, Žnidarič A, O’Brien E, Jordan R. COST 345 –
subspace identification. Mech Syst Signal Process 2013;38(2):346–66. procedures required for the assessment of highway structures, Final Report;
[20] Döhler M, Lam XB, Mevel L. Uncertainty quantification for modal parameters 2004. <http://cost345.zag.si/Reports/COST_345_Summary_Document.pdf>.
from stochastic subspace identification on multi-setup measurements. Mech [31] Colditz B. Bridge strengthening – a necessary prerequisite for a reliable
Syst Signal Process 2013;36(2):562–81. highway network (in German). Bautechnik 2013;90(3):184–92.
[21] Basseville M, Abdelghani M, Benveniste A. Subspace-based fault detection [32] Reynders E, Houbrechts J, De Roeck G. Fully automated (operational) modal
algorithms for vibration monitoring. Automatica 2000;36(1):101–9. analysis. Mech Syst Signal Process 2012;29:228–50.