Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

Bridge model updating and health monitoring

using a single mobile actuator-sensor pair


Rajdip Nayek1 , Suparno Mukhopadhyay2 and Sriram Narasimhan1
1
Department of Civil and Environmental Engineering, University of Waterloo, ON N2L 3G1,
Canada
2
Department of Civil Engineering, Indian Institute of Technology, Kanpur, 208016, India
E-mail: rnayek@uwaterloo.ca

Abstract. A dense array of stationary sensors is usually needed to accurately capture the
distributed modal properties of a bridge structure. An appealing alternative is to use, if possible,
a single or a small set of mobile sensors, which can offer more flexibility by being able to move
from one location to another on the bridge. This study investigates the feasibility of using a
single mobile collocated actuator-sensor pair for modal identification and model updating of
a multi-span bridge modeled as a continuous beam. Acceleration responses measured by the
mobile actuator-sensor pair at different points of the bridge are used to identify an incomplete
set of natural frequencies, modal damping ratios and displacement mode shape components at
the measurement locations. These identified modal parameters are then used to update the
stiffness parameters of the bridge using a sensitivity analysis based approach. An application
of the method in structural damage detection is shown using numerical simulations.

1. Introduction
In vibration-based identification and health monitoring of bridges, a dense array of stationary
sensors is typically employed to accurately capture the vibration mode shapes. This can be cost
prohibitive in terms of installation and maintenance. For example, the cost of installing over
350 sensors on the Tsing Ma suspension bridge in Hong Kong exceeds $8 million [1]. The data
storage and processing costs associated with such an installation would also be accordingly high.
An economically appealing alternative is to use a single or a limited number of mobile sensors,
and make them move over the bridge to collect data at different locations. This alternative
strategy has already attracted some recent interest. For example, a mobile sensor along with a
fixed reference sensor has been used in [2] to estimate the mode shapes of a beam under ambient
excitation; a set of moving sensors and at least one fixed reference sensor has been used in [3]
for flexibility identification of a full-scale bridge. The issue of incomplete point measurements
when using mobile sensors has been addressed in [3] as the missing data problem in statistics.
In this paper, the feasibility of using only a single mobile actuator-sensor pair, without any
fixed reference sensor, for modal identification and model updating of bridges is studied. Given
a single mobile collocated actuator-sensor pair and an initial finite element (FE) model of a
bridge, the objectives are: (a) to determine the modal parameters of the bridge using a sequence
of input/output (I/O) point measurements, and (b) to update the element stiffness parameters
of the bridge. Acceleration measurements are considered. The following assumptions are made:
(i) the bridge is modeled as a beam, i.e. with bending modes only and no torsional modes,
(ii) the bridge is classically damped,
(iii) the mass matrix is not included in the updating procedure, and
(iv) all the active vertical degrees of freedom (DOFs) corresponding to the FE model, are
sequentially excited and measured.
Even while using mobile sensors, a fixed reference sensor is required to adjust the phase difference
between different point measurements [2, 3]. In contrast, here, the initial FE model is used
instead for phase information, as well as to account for the unmeasured rotational DOFs. The
details of the method is presented in the next section (Section 2). The performance of the
method is numerically tested in Section 3, in the context of an application to structural damage
detection.

2. Modal Identification and Model Updating Using a Single Collocated Mobile


Actuator-Sensor Pair
The equation of motion of an N -DOF continuous beam FE model of a multi-span bridge is:

Mÿ(t) + Cẏ(t) + Ky(t) = Lf (t) (1)

where M, C and K denote the global mass, damping and stiffness matrices, respectively; y(t)
denotes the displacement vector in physical coordinates; and the vector L of dimension N × 1
represents the instantaneous location of the force f (t) applied on the bridge by the actuator
of the moving actuator-sensor pair. Two models of the bridge is considered: (a) an initial
(baseline) FE model, and (b) a FE model representing the current state of the bridge at the
time of inspection. The initial baseline FE model is known a priori; the objective is to update
the parameters of this initial model to obtain the parameters of the FE model representing the
current state. We denote the stiffness and damping matrix of the baseline FE model as K0 and
C0 , and, that of the updated FE model as Kd and Cd , respectively. Assuming no significant
change in the mass, the mass matrix for both the FE models is denoted as M0 . The change in
stiffness between the initial and current FE models is written as:

∆K = Kd − K0 (2)

If the continuous beam model of the bridge contains ne Euler-Bernoulli beam elements, the
change in the global stiffness matrix ∆K can be expressed as:
ne
∆K = ∆αi ATi ki0 Ai (3)
X

i=1

where ki0 is the element stiffness matrix of the ith element of the initial FE model, Ai is the
FE assembly matrix for the ith element, and ∆αi denotes the change in the stiffness of the ith
element between the initial and current models. Here, we assume ∆αi ∈ [−1, 0] in the context of
applications to structural damage detection, with the current model denoting a possibly damaged
state of the bridge, while the initial model denotes a “healthy” state. ∆αi = 0 indicates no loss
of stiffness and ∆αi = −1 indicates complete loss of stiffness of the ith element in current state.

2.1. Modal Identification


We define a set A = {a1 , a2 , · · · , and } ⊆ {1, 2, . . . , N }, with nd ≤ N elements, as the set of all
active vertical DOFs of the bridge model. The mobile actuator-sensor pair excites and takes
acceleration measurements at all locations corresponding to the DOFs in A, and gathers a set
of single-input-single-output point measurement data, denoted by D, represented as:

D = {(fi (t), ÿi (t)) : i ∈ A} (4)


Typically, only a few low frequency modes may be identified from the measured data. We
define a set B = {b1 , b2 , · · · , bnm } ⊆ {1, 2, . . . , N }, with nm ≤ N elements, as the set of modes
identifiable from D. Then, using modal superposition, the acceleration response at the ai th
DOF, ÿai (t), can be written as:

q̃¨j (t)
2
ÿai (t) ≈ (5)
X
φdai j
j∈B

where φdai j is the ai th component, corresponding to the ai th DOF, of the jth mode shape of the
bridge in its current state; and q̃¨j (t) is the jth pseudo-modal acceleration response obtained by
solving the pseudo-modal equation of motion [5]:

q̃¨j (t) + 2ξjd ωjd q̃˙j (t) + (ωjd )2 q̃j (t) = fai (t) (6)

In Eq. (6), ξjd and ωjd respectively denote the jth modal damping ratio and natural frequency of
the bridge in its current state, which can be identified from D, for all j ∈ B, using any suitable
modal identification technique. Assuming the number of data points in any I/O time history to
be greater than nm , the squares of the mode shape components in Eq. (5) can then be estimated
as a least squared error solution of Eq. (5); taking the element-wise square root of this least
squares solution vector, we get the nm × 1 vector of absolute values of the nm identifiable mode
shapes at the ai th DOF:
iT
τ̂ ai = |φ̂dai b1 | |φ̂dai b2 | . . . |φ̂dai bnm | (7)
h

Performing the identification using all the point I/O measurements in D, we get the following
matrix of absolute values of mode shapes of the bridge in its current state:

|φ̂d | |φ̂da1 b2 | |φ̂da1 bnm |


 
...
 da1 b1
 |φ̂a b | |φ̂da2 b2 | |φ̂da2 bnm | 

...
τ̂ =  . .. .. (8)
2 1

.. ..
 
. . .

 
 
|φ̂dan b | |φ̂dan b | . . . |φ̂dan b |
d 1 d 2 d nm nd ×nm

If for any DOF ai , multiple sets of I/O data is available, the corresponding |φ̂dai j |’s in Eq. (8)
may be obtained as the average of the individual estimates from each data set.
To obtain the modes shape components (with sign) of the bridge in its current state, the
signs of the mode shape components of the initial FE model are used with the corresponding
identified absolute values of mode shapes in τ̂ :

φ̂dij = sgn φ0ij φ̂dij (9)


 
∀ i ∈ A, j ∈ B

where sgn denotes the signum function, and φ0ij is the is the ith component of the jth mode
shape of the initial FE model. The assumption that the mode shapes do not change sign from
the initial to the current model may be expected to be valid for the lower frequency modes, which
are the ones usually identified from data, and for moderate differences in stiffness between the
initial and current states of the bridge.

2.1.1. Sources of errors and improved parameter estimates: The accuracy of the estimated τ̂
in Eq. (8) depends on:
• the accuracy of the identified natural frequencies and modal damping ratios, as these are
used to compute the pseudo-modal acceleration responses in Eq.(6); and
• on the number of identified modes, as the approximation of modal superposition in Eq. (5)
improves with the number of modes identified.
In this paper, the sequence of I/O point measurements are individually used in the
Eigensystem Realization Algorithm - Observer Kalman Filter Identification (ERA-OKID) [6] to
identify the natural frequencies and modal damping ratios. These initial estimates are found to
contain measurement errors. To obtain improved estimates, the following optimization problem
is solved [5], starting with the estimates obtained using ERA-OKID:

Θ̂∗ = arg min ÿai (t) − ŷ¨ai (t; Θ̂) (10)




2

where k·k2 denotes the Euclidean norm, Θ̂ = ω̂bd1 , . . . , ω̂bdnm , ξˆbd1 , . . . , ξˆbdnm
is the parameter
n o

vector containing the natural frequencies and modal damping ratios to be identified, and ŷ¨a (t; Θ̂) i
is the reconstructed acceleration response (computedn via Eqs. (5) to (7)) for anyogiven Θ̂. Once
the optimal estimates of the modal parameters, ω̂bd∗1 , . . . , ω̂bd∗nm , ξˆbd∗
1
, . . . , ξˆbd∗
nm
are obtained,
these are then used to obtain the estimates of the mode shapes.
To address the second source of decreased accuracy in the identified mode shapes, we augment
the nm modal frequencies and damping ratios identified from the data, with m additional
frequencies and damping ratios obtained from the initial FE model. These set of m analytical
modes are termed as residual FE modes. Let the set B FE = {bFE 1 , b2 , · · · , bm }, denote the set
FE FE

of these residual FE modes, such that B ∩ B = ∅ and B ∪ B ⊆ {1, 2, . . . , N }. The absolute


FE FE

values of the mode shapes, including the residual FE modes, are then estimated following the
procedure discussed before:
 
|φ̂da1 b1 | ... |φ̂da1 bnm | |φ̂da FE | ... |φ̂da1 bFE |
 1 b1 m 
 |φ̂d | . . . |φ̂da2 bnm | |φ̂da bFE | . . . |φ̂da2 bFE | 
 a2 b1
τ̂ AUG =  2 1
(11)
m

.. .. .. .. .. ..

. . . . . .
 
 
 
|φ̂dan d d d
b1 | . . . |φ̂an bnm | |φ̂a bFE | . . . |φ̂an bFE |
d d nd 1 d m nd ×(nm +m)

From this estimated τ̂ AUG , the τ̂ of Eq. (8) is extracted as the matrix of identified absolute
values of the nm mode shapes, corresponding to the nm modes identified from data. The part of
τ̂ AUG corresponding to the residual FE modes is ignored. This idea of using additional residual
FE modes is to provide an improvement in the modal representation of the measured response in
Eq. (5). Based on numerical simulations, it is recommended to use 5 to 10 residual FE modes.

2.2. Model Updating


The nm natural frequencies and mode shapes identified using the procedure discussed in Section
2.1 are next used to update the element stiffness parameters of the initial FE model to get the
current FE model, using the sensitivity approach [7]. The updates in the stiffness parameters,
given by the ∆αi ’s in Eq. (3), are obtained as a solution to the following system of equations:

S ∆α = ∆Z (12)

where ∆α = {∆α1 ∆α2 · · · ∆αne }T ; ∆Z is the vector of differences between the identified modal
parameters and corresponding modal parameters of the initial model (initial minus identified);
and S = ∂α
∂Z
is the first-order sensitivity matrix, with αi = 1 + ∆αi . For nm identified modes
and nd observed vertical DOFs, the dimensions of S and ∆Z are, respectively, nm (nd + 1) × ne
and nm (nd + 1) × 1. The modal parameter derivatives necessary to construct S can be obtained
from [8, 9]:
" #
∂K0

∂λj
= (φ0j )T (φ0j )



∂αi ∂αi




" # 

 ∂φ ∂K0 ∂λj 0 ∀ j ∈ B, i ∈ {1, 2, . . . , ne } (13)

=− M (φ0j )

0 j
K − λ0j M0 −
∂αi ∂αi ∂αi 




∂φj 
(φ0j )T M0 =0




∂αi
where λj and φj respectively denote the eigenvalue (square of natural frequency ωj ) and mode
shape for the jth mode, and the superscript ‘0’ denotes parameters of the initial FE model.
Eq.(12) generally represents an overdetermined system of equations and a least squared error
solution to ∆α may be obtained. However, the sensitivity matrix S is often ill-conditioned.
Hence, we consider two possible regularization schemes: (i) Tikhonov (`2 ) regularization

∆α̂ = min kS∆α − ∆Zk22 + µ2 k∆αk22 (14)


 
∆α

where µ ≥ 0 is the Tikhonov regularization parameter obtained using the L-curve criterion [10];
and (ii) sparse (`1 ) regularization

∆α̂ = min kS∆α − ∆Zk22 + κ2 k∆αk1 (15)


 
∆α

where k.k1 is the `1 norm, and the parameter κ ≥ 0 is chosen using a `1 -reweighting procedure.
While the solution ∆α can be obtained explicitly when using Tikhonov regularization, the
sparse regularization has no closed-form solution. The primal-dual interior point algorithm [12]
is employed in this study to compute ∆α when using sparse regularization.

3. Numerical Example: Application to Damage Detection


In this section an application of the proposed modal identification and model updating method
is considered in the context of damage detection in a 40 m long three-span bridge, shown in
Figure 1. The bridge is modeled using 20 Euler-Bernoulli beam elements, with 21 nodes; the node
numbers (circled) and element numbers are shown in Figure 1. Each node has two DOFs, vertical
translation and in-plane rotation, except at the supports where only rotation is allowed; this
results in a total of 38 DOFs (17 translation + 21 rotation). The bridge material has an elastic
modulus of 30 GPa and mass density of 2400 kg/m3 . All three spans have the same width of 2
m. The spans have different depths: 0.4 m for the left span, 0.5 m for the center span, and 0.35
m for the right span. Modal damping is adopted, with the damping ratios for the first 10 modes
assumed to be: ξ1 = ξ2 = 0.02, ξ3 = ξ4 = ξ5 = 0.025 and ξ6 = ξ7 = ξ7 = ξ8 = ξ9 = ξ10 = 0.03,
and all the remaining damping ratios assumed to be 0.05. This known (initial) FE model is
assumed to represent the healthy state of the bridge. For the current (tested) state, multiple
local damages, simulated as reductions in element stiffness values, are introduced into the healthy
state (the damaged elements are shown in Figure 1):
• 25% damage in element 3 (middle of span 1)
• 15% damage in element 6 (near support 2 of span 2)
• 15% damage in element 10 (middle of span 2)
• 10% damage in element 11 (middle of span 2)
• 20% damage in element 16 (near support 3 of span 2)
The mobile actuator-sensor pair is assumed to excite the bridge in its current state with
vertical white noise inputs and record the vertical accelerations, sequentially at all unrestrained
nodes, collecting a total of 17 I/O point measurements. Each measurement is 30 s long, sampled
at 500 Hz. To simulate measurement noise, the true responses are corrupted by adding white
noise signals with root mean square (RMS) values equal to 15% of the RMS of the corresponding
true signals. Only the first six modes are assumed to be identified from the I/O measurements.
Table 1 compares the true natural frequencies and modal damping ratios with their estimates
obtained using ERA-OKID, and their optimized estimates obtained using Eq. (10). It is evident
that the optimization improves the estimates of the modal damping ratios.

Figure 1: Bridge modeled as continuous beam, used in the numerical example

Table 1: Comparison of true natural frequencies and damping ratios of the damaged bridge,
with estimates using ERA-OKID and optimized estimates from Eq. (10)

Mode Natural Frequency (in Hz) Damping Ratio (in %)


True ERA-OKID Optimized True ERA-OKID Optimized
1 2.5595 2.5929 2.5633 0.0200 0.0045 0.0173
2 5.6588 5.6616 5.6601 0.0200 0.0046 0.0201
3 8.7932 8.7927 8.7962 0.0250 0.0042 0.0252
4 12.5200 12.5430 12.5223 0.0250 0.0045 0.0254
5 18.4780 18.5060 18.4711 0.0250 0.0044 0.0285
6 22.0152 22.0484 22.0339 0.0300 0.0043 0.0378

The components, corresponding to the vertical DOFs, of the first six mode shapes of the
damaged bridge, are next computed using the method discussed in Section 2.1, using some
residual FE modes from the initial (healthy) FE model. The Modal Assurance Criterion (MAC)
values comparing these estimated mode shapes with the true mode shapes of the damaged bridge
are plotted in Figure 2, for different numbers of residual FE modes considered. It can be seen
that the use of residual FE modes, even though they are from the healthy bridge, leads to better
estimates of the damaged mode shapes.
The identified modal parameters, considering only five residual FE modes, are finally used
for detecting the changes in the element stiffness values (∆α) between the initial FE model and
the current (tested) state of the bridge. Figure 3 compares the true changes in the element
stiffness values with the estimated changes, obtained using the two regularization schemes
considered in Section 2.2. It is evident that the sparse regularization works better than the
Tikhonov regularization, both in detecting stiffness changes in the damaged elements, as well as
in detecting no change in stiffness of the undamaged elements. Moreover, the values of stiffness
changes obtained using sparse regularization agree better with true values, as compared to the
values obtained using Tikhonov regularization. Nonetheless, even in case of sparse regularization,
the existence of damage is missed in one element (element 16), while the presence of damage is
incorrectly detected in element 19.

0.99
MAC

0.98 Mode 1
Mode 2
Mode 3
Mode 4
0.97 Mode 5
Mode 6

0.96
0 5 10 15 20 30
Number of residual FE modes used

Figure 2: MAC values comparing identified and true mode shapes of damaged bridge, for
different numbers of residual FE modes

35 35 35

30 30 30

25 25 25
−∆αi × 100

−∆αi × 100
−∆αi × 100

20 20 20

15 15 15

10 10 10

5 5 5

0 0 0
6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 1
Element i Element i Element i Element i
(a) (b)

Figure 3: Comparison of true values of element stiffness changes with values estimated using
sensitivity-based model updating: (a) with Tikhonov regularization (mean squared error =
0.3319), and (b) with sparse regularization (mean squared error = 0.2768)

4. Conclusions
In this paper, the feasibility of using a single mobile collocated actuator-sensor pair for modal
identification and model updating of multi-span bridges has been investigated A method of
modal identification using a set of sequential point measurements of I/O data and an initial FE
model of the bridge has been proposed. The subsequent use of the identified modal parameters
in updating the initial FE model using a sensitivity-based framework has also been discussed.
A numerical example of structural damage detection using the proposed method shows that the
idea of employing a single mobile actuator-sensor pair for identification and health monitoring
of bridges is feasible, and has potential for success with further research.

References
[1] Ni Y Q, Wang B S, and Ko J M 2000 SPIE’s 5th Annual International Symposium on Nondestructive
Evaluation and Health Monitoring of Aging Infrastructure (Newport Beach, CA) (Bristol: IOP Publishing)
pp 312-323
[2] Marulanda J, Caicedo J M, and Thomson P 2016 J. Comput. Civ. Eng. 31
[3] Zhang J, Guo S L and Zhang Q Q 2015 Comput.-aided Civil Infrastruct. Eng. 30 pp 703-714
[4] Matarazzo T J, Horner M and Pakzad S N 2016 In Topics in Modal Analysis and Testing, 10 pp 255-260
[5] Mukhopadhyay S, Lus H and Betti R 2014 Mech Syst Signal Process. 45 pp 283-301
[6] Juang J N 1994 Applied System Identification (New Jersey: Prentice Hall)
[7] Mottershead J E, Link M and Friswell M I 2011 Mech Syst Signal Process. 25 pp 2275-2296
[8] Fox R L and Kapoor M P 1968 AIAA Journal 6 pp 2426-2429
[9] Nelson R B 1976 AIAA Journal 14 pp 1201-1205
[10] Hansen P C 1992 SIAM Review 34 pp 561-580
[11] Candes E J, Wakin M B and Boyd S P 2008 J. Fourier Anal. Appl. 14 pp 877-905
[12] Boyd S P and Vandenberghe L 2004 Convex Optimization (Cambridge: Cambridge University Press)

You might also like