Journal of Mechanics

http://journals.cambridge.org/JOM

Additional services for Journal of Mechanics:

Email alerts: Click here

Subscriptions: Click here
Commercial reprints: Click here
Terms of use : Click here

The Health Monitoring of Bridges using Articial Neural Networks

Pei-Ling Liu and Shyh-Jang Sun

Journal of Mechanics / Volume 17 / Issue 03 / September 2001, pp 157 - 166

DOI: 10.1017/S1727719100004524, Published online: 05 May 2011

Link to this article: http://journals.cambridge.org/abstract_S1727719100004524

How to cite this article:

Pei-Ling Liu and Shyh-Jang Sun (2001). The Health Monitoring of Bridges using Articial Neural Networks. Journal of
Mechanics, 17, pp 157-166 doi:10.1017/S1727719100004524

Request Permissions : Click here

Downloaded from http://journals.cambridge.org/JOM, IP address: 130.237.29.138 on 01 Jul 2015

 THE HEALTH MONITORING OF BRIDGES USING
ARTIFICIAL NEURAL NETWORKS

Pei-Ling Liu * Shyh-Jang Sun **

Institute of Applied Mechanics
National Taiwan University
Taipei, Taiwan 10617, R.O.C.

ABSTRACT

This study develops a neural network system to monitor the safety of a bridge structure. A truck of
constant mass is driven at constant speed through the target bridge. Then, the maximal and minimal
values of the bridge elongations are processed by a monitoring system to evaluate the current condition
of the bridge. The monitoring system is composed of parallel backpropagation neural networks.
Each neural network monitors a part of the bridge. The neural networks are trained using simulation
data. The numerical example shows that the monitoring system is effective in the damage detection of
the bridge.
Keywords : Bridge, Monitoring, Artificial neural networks, Elongation, Backpropagation.

1. INTRODUCTION have developed a direct inverse method to identify the

axial rigidities of a truss using the static strains of the
Bridges play an indispensable role in land trans- truss. Transient data have also been adopted to
portation. The collapse of a bridge often results in identify the dynamic model of a structure. Shinozuka,
tremendous losses. Therefore, the safety assessment et al. [6] and Agbabian, et al. [7] have used the
of bridges is an important and challenging task for civil least-squares method and time domain response to
engineers. The conventional way of evaluating the identify a linear dynamic system. Yun and Shinozuka
integrity of a bridge is by visual inspection. However, [8], Hoshiya and Saito [9] have used the extended
bridge structures are usually large. Therefore, such Kalman filter to identify linear and nonlinear systems.
inspection is costly and time-consuming. Furthermore, On the other hand, some researchers have proposed
the inspection may not be complete because the the use of an artificial neural network (ANN) to detect
structure is not totally accessible. Besides, only quali- the damage levels and locations of structures.
tative evaluation is possible using such an approach. Different inputs have been proposed by various
System identification methods have been applied researchers. For instance, Pandey and Barai [10] have
widely to evaluate the global condition of existing used the nodal displacements to monitor a truss bridge.
structures. Many researchers have adopted modal data Wu, et al. [11] have used the acceleration spectrum to
in the identification of structures. For example, Chen detect the damage level and location of a three-story
and Garba [1] have developed a method to determine building. Elkordy and Chang [12] have used the
the changes in the structure stiffness matrix based on difference between the response of the original and the
the assumption that the mass matrix does not change damaged truss to estimate the reduction of element
with damage. The detection of damage is then rigidity. Tsou and Shen [13] have used the change of
attempted by examining the change of the stiffness natural frequencies to identify the rigidity reduction of a
matrix. Yao, et al. [2] have suggested the use of strain multi-degree-of-freedom mass-spring system.
modal shapes to evaluate structural damage. Liu [3] This paper proposes an intelligent inspection system
have incorporated the finite-element formulation in the for bridges. A truck of constant mass is driven at
least-squares inversion to determine the axial rigidities constant speed over the target bridge. The response of
of a truss. Sanayei and Onipede [4] have developed a the bridge is then processed by an artificial neural
method for computing stiffness and detecting damage at network system to evaluate the condition of the bridge
element level using static test data. Liu and Chian [5] automatically.

* Professor ** Assistant Scientist

The Chinese Journal of Mechanics, Vol. 17, No. 3, September 2001 157
 2. DYNAMIC ANALYSIS OF A BRIDGE 3. THE BRIDGE MODEL

A truck moving at constant speed V across a bridge Consider a simply supported, three-span bridge.
can be modeled as a mass moving on a continuous beam Each span is 60m long. The original flexural rigidity
(Blejaws, et al. [14]), as shown in Fig. 1. The EI of the bridge is 7.34×1010 N．m2 and the mass density
dynamic response of a beam subjected to a moving ρA is 1570kg/m. The mass of the test truck is
mass is as follows: 20,000kg, and the speed is 60km/hr. The distance
from the neutral axis of the cross-section to the bottom
∂2  ∂ 2 w( x, t )  ∂ 2 w( x, t ) of the bridge is 1,860mm. In the finite element
 EI  + ρA
∂x 2  ∂x 2  ∂t 2 analysis, the bridge is discretized into 30 uniform beam

elements.
∂ 2 w(η, t )
+ Mδ( x − η) = Mgδ( x − η) (1) Nine damage zones are considered for the bridge,
∂t 2 namely, 1L, 1M, 1R, 2L, 2M, 2R, 3L, 3M, and 3R,
where w is the transverse displacement of the beam, EI where 1, 2, and 3 denote the span number, and L, M,
is the flexural rigidity, ρA is the mass density per unit and R denote the left, middle, and right of the span, as
length, M is the mass of the truck, g is the gravitational shown in Fig. 1. Each damage zone consists of two
elements. For example, zone 1L constitutes of
acceleration, δ is the Dirac delta function, and η = Vt
elements 1 and 2, zone 2M constitutes of elements 15
is the position of the truck.
and 16, and zone 3R constitutes of elements 29 and 30.
In this study, damage of the bridge is modeled by
rigidity reduction. Define the rigidity ratio as
EI i′
ri = (6)
EI i

where EIi and EIi′ are the original and current flexural
rigidities of zone i. Take zone 2M for example. r2M =
0.8 means that the flexural rigidities of elements 15 and
16 are reduced to 80% of the original value.
Fig. 1 The example bridge The response of the example bridge is computed for
various damage conditions by integrating Eq. (2).
Finite element formulation of the above equation These responses will be used as a database to construct
leads to a monitoring system of the bridge based on artificial
neural networks.
&& = F
K U+M U (2)

where K, M, U, and F are the stiffness matrix, mass 4. ARTIFICIAL NEURAL NETWORK
matrix, nodal displacement vector, and external load
vector, respectively. The components of these The artificial neural networks adopted in this study
quantities are as follows: are multi-layered backpropagation networks (Zurada
[15]). The backpropagation network constitutes of an
d 2 ϕi d ϕ j
2
xB
K ij = ∫ EI dx (3) input layer, an output layer, and several hidden layers.
xA dx 2 dx 2 Each layer is composed of several neurons, and each
neuron is linked to every neuron of the adjacent layers,
 M g ϕ i (η) x A ≤ η < xB
(4) as shown in Fig. 2.
Fi = 
 0 otherwise

 xB


M ij = 
∫ xA
ρ A ϕ i ϕ j dx + M ϕ i (η) ϕ j (η) x A ≤ η < xB
xB
otherwise

 ∫ xA
ρ A ϕ i ϕ j dx

(5)
where xA and xB are the coordinates of the nodes, ϕi is
the shape function. Notice that M and F are dependent
on the truck position. The conventional beam element
is adopted in this study. Each element has two nodes, Fig. 2 Backpropagation nueral network
and each node has two degrees of freedom, namely, the
transverse displacement and rotation. The input layer reads the input data and passes it to
Equation (2) can be integrated by numerical schemes the first hidden layer. The neurons in the first hidden
to obtain the response of the bridge. layer take the weighted sums of the input data and

158 The Chinese Journal of Mechanics, Vol. 17, No. 3, September 2001
evaluate the activation function f :

tk = f (∑ p
j =1
Wkj s j ) (7)

where sj is the input data, tk is the output of the

activation function, and Wkj is the weight of a link.
Then, tk is passed forward to the next layer. The
neurons in the next layer again take the weighted sums
of the incoming data, evaluate the activation function,
and pass the function values forward. The same
process is repeated until the output layer sends out the
results.
The weights of the links, which characterize the
network, are determined by training. To train a
backpropagation network, a number of samples must be
provided. Suppose the target output of a sample is
d = [d1 d2 … d r ] T and the network output is
o = [o1 o2 … o r ] T . A sum-squared error can
Fig. 3 Elongation of zone 1L under various damage
be defined as follows:
conditions
1 2
∑ (d − ok )
r
E= k =1 k
2 (8)
In each training cycle, the weights are modified along
the direction of negative gradient of the error function.
The weight corrections are performed layer by layer,
starting from the output layer backward to the input
layer.

5. STRUCTURE OF NEURAL NETWORK

While designing an artificial neural network, the

most crucial step is the selection of input parameters.
The input parameters must be chosen such that they are
representative of the output. Hence, the input vector
must be able to reflect the current condition of the
bridge in this application.
In the beginning, the accelerations of the bridge were
used as input to the ANN monitoring system. After Fig. 4 Elongation of zone 1M under various damage
training, the system can detect the damage of the bridge conditions
successfully if only one zone is damage.
Unfortunately, it is not effective if there are multiple
damaged zones. Further investigations shows that
strain signals are excellent indicators of damages.
However, since strains are pointwise measurements, the
system fails if no strain gages are mounted on the
damaged zones.
To remedy the situation, the longitudinal elongations
at the bottom of the 9 zones are used as the input
features of the monitoring system. The longitudinal
elongation of a zone is equal to the integration of axial
strains in that zone. Therefore, it is expected to be
highly correlated to the condition of that zone.
Furthermore, there is no risk of missing a damage
because the longitudinal elongation can reflect the
change of strain at any point of the entire zone.
Figures 3 to 7 show the time histories of longitudinal
elongations of zones 1L to 2M for various damage
locations. In each case, only one zone is damaged, Fig. 5 Elongation of zone 1R under various damage
and the rigidity reduction of the damaged zone is 50%. conditions

The Chinese Journal of Mechanics, Vol. 17, No. 3, September 2001 159
Fig. 6 Elongation of zone 2L under various damage Fig. 8 The influence of damage level on elongation of
conditions zone 1M

Fig. 7 Elongation of zone 2M under various damage Fig. 9 Extreme elongation of zone 1L under various
conditions damage conditions

It is seen that as the bridge is damaged, the amplitude of

elongation of the damaged zone increases substantially.
The amplitude of elongation of some undamaged zones
also increases, but not as mush. The elongation curves
of zones 2R to 3R are not given here. However,
similar phenomena are observed.
Figure 8 shows the variation of longitudinal
elongation of zone 1M as zone 1M is damaged to
different levels. It is seen that the amplitude increases
with the damage level. The elongations of the other
zones possess similar trends. Obviously, the
longitudinal elongation not only reveals where the
damage is but also how serious the damage is.
The amplitude of a curve can by represented by the
maximum or minimum of the curve. Therefore, the
two extreme elongations are suitable input of the
monitoring system. Figures 9 to 13 respectively show
the change of the extreme elongations of zones 1L to Fig. 10 Extreme elongation of zone 1M under various
2M as the location and level of the damage vary. damage conditions

160 The Chinese Journal of Mechanics, Vol. 17, No. 3, September 2001
 These figures confirm that the extreme elongations are
indeed good indicators of damages. Because of
geometrical symmetry, the trend of the extreme
elongations of zones 2R is similar to that of zone 2L.
The similarity also holds between zones 1L and 3R, 1M
and 3M, and 1R and 3L.
Hence, one may develop an ANN system using the
18 extreme elongations as the input vector. The
system can be easily trained to detect the single-damage
cases. Suppose 5 damage levels are considered for
each zone. There are only 1 undamaged case and 45
(= 9 zones × 5 damage levels) damaged cases to be
analyzed to construct the training database. However,
if the system should also detect multiple-damage cases,
the training database will be extremely large. There
will be more than two million cases if all combinations
are included. Apparently, this architecture is not
feasible because it takes too much effort to prepare the
Fig. 11 Extreme elongation of zone 1R under various
database, not to mention training and testing.
damage conditions
Fortunately, further examination of Figures 9 to 13
shows that:
1. The extreme elongations of zone 1L change notably
only when the zone itself is damaged. Damage in
other zones has almost no influence on these values.
This is also true for zone 3R.
2. The minimal elongations of zones 1M, 2M, and 3M
are influenced by the damage of many zones.
However, the absolute values of the maximal
elongations increase noticeably only when the
observation zone is damaged.
3. The extreme elongations of zones 1R, 2L, 2R, and 3L
are influenced by the damage of all 9 zones.
However,
(a) only when zone 1R or 2M is damaged do both
w1Rmax and w1Rmin increase;
(b) only when zone 2L or 1M is damaged do both
w2Lmax and w2Lminincrease;
(c) only when zone 2R or 3M is damaged do both
Fig. 12 Extreme elongation of zone 2L under various w2Rmax and w2Rminincrease;
damage conditions (d) only when zone 3L or 2M is damaged do both
w3Lmax and w3Lminincrease,
where wimax and wimin are the maximal and minimal
elongations of zone i.
This phenomenon also holds when there are more
than one damaged location. Take zones 1L and 1R for
example. Figure 14 shows the elongations of zones 1R
as zones 1L and 1R are both 50% damaged. It is seen
that the elongation curve of zone 1R in the
1L-1R-damaged case almost coincides with that in the
1R-damaged case. That means the elongation of zone
1R is hardly influenced by the damage of zone 1L even
when two zones are damaged simultaneously.
Hence, depending on the location, the damage of a
zone may or may not influence the extreme elongations
of the other zones. When the middle of a span is
damaged, wmax and wmin of the zones in the
adjacent spans near the support will increase
simultaneously. Such interaction is depicted in Fig.
15(a). Simulations were also carried out for a
Fig. 13 Extreme elongation of zone 2M under various five-spanned bridge, and similar results were obtained,
damage conditions as shown in Fig. 15(b).

The Chinese Journal of Mechanics, Vol. 17, No. 3, September 2001 161
 Fig. 16 Structure of bridge monitoring neural network
Fig. 14 Elongation of zone 1R under various damage
conditions
The input features of the each network are the
normalized extreme elongations of the monitored zone,
as defined below:
wmax − wmax wmin − wmin
y max = , y min = (9)
wmax wmin
where wmax and wmin are the maximal and minimal
elongations of the current bridge, and wmax and wmin
are the maximal and minimal elongations of the
undamaged bridge.
The normalized elongation is adopted because there
is always discrepancy between the numerical model and
the real bridge. The ANN system is trained using
numerical results. Therefore, if the measured response
is input to the system directly, the bridge may be
Fig. 15 Interaction of damage and elongation diagnosed damaged even when it is perfectly healthy.
In order to overcome this, the extreme elongations of
Based on above observations, the 9 damage zones are the undamaged bridge, measured in advance, are used
divided into five groups: (1L), (1M, 2L), (1R, 2M, 3L), as a baseline for comparison.
(2R, 3M) and (3R). The extreme elongations of a zone Thus, networks 1M-2L and 2R-3M have four input
are dependent on the condition of the zones in the same neurons, and network 1R-2M-3L has 6 input neurons.
group, and insensitive to the condition of the zones in However, Networks 1L and 3R have only one input
other groups. neuron: the normalized maximal elongation. This is
According to such grouping, the bridge monitoring because the condition of zone 1L and 3R can be
system is constructed using five parallel neural evaluated based on the maximal elongation alone.
networks, namely, network 1L, network 1M-2L, In the beginning, the rigidity ratios of the monitored
network 1R-2M-3L, network 2R-3M, and network 3R. zones were adopted as the network output. However,
Network 1L processes the features of zone 1L to it was found that the training of the networks hardly
monitor the health of zone 1L; network 1M-2L converged. This difficulty arises because the rigidity
processes the features of zones 1M and 2L to monitor ratio of an undamaged zone is 1, but the activation
the health of zones 1M and 2L; and so on. function of output neuron yields 1 only when the input
The structure of the system is rather simple, as shown is ∞. To solve this problem, a linear transformation is
in Fig. 16. Only one hidden layer is used in each of adopted that maps the range of rigidity ratio from [0, 1]
the 5 networks. This is because each network to [0.02, 0.98]. Then, the scaled values are used as the
monitors no more than three zones, and only one hidden network output. The networks experienced no
layer is sufficient to simulate the mapping. The input convergence problem after the modification was made.
and hidden neurons use the hyperbolic tangent function Since the output vector no longer represents the real
as the activation function, and the output neurons use rigidity ratios, an inverse transformation is performed
the sigmoid function as the activation function. The before the results are printed. As such, the monitoring
type of activation function and the number of hidden system still outputs the rigidity ratios of the monitored
neurons are determined by trial runs. zones.

162 The Chinese Journal of Mechanics, Vol. 17, No. 3, September 2001
 6. TRAINING AND TEST OF THE Since noise always presents in practice, 20% random
MONITORING SYSTEM noise was superposed to the test data to see if the
system is vulnerable to noise. A thousand Monte
As mentioned before, one must provide a number of Carlo simulations were performed for each damaged
samples to train the neural network system. Hence, case. It turned out that the biases of the output ratios
finite element analysis was carried out for various are less than 15%, and the standard deviations are less
damage conditions. In the simulations, 5 damage than 19%. Both the biases and standard deviations do
levels were considered at each zone, the rigidity ratios not exceed the noise level. Apparently, this system is
being 0.1, 0.3, 0.5, 0.7 and 0.9. The training database insensitive to noise.
for the five networks contains the following samples:
1. Network 1L and 3R: 6.2 Two Zones Damaged
46 cases = the undamaged case
+ 45 single-zone-damaged cases Twelve combinations of damaged zones are
2. Networks 1M-2L and 2R-3M: randomly selected for use in the test, namely, 1L-1M,
71 cases = the undamaged case 1L-1R, 1L-3R, 1M-1R, 1M-2L, 1M-3M, 1R-2M, 1R-3L,
+ 45 single-zone-damaged cases 2L-2R, 2M-3L, 2R-3M and 2R-3L. In each
+ 25 1M-2L-damaged cases or combination, 9 rigidity ratios (0.9, 0.8, …, 0.2 and 0.1)
2R-3M-damaged cases are assumed for each zone. Hence, there are totally
3. Network 1R-2M-3L: 972 test samples, and only 75 cases have been used in
121 cases = the undamaged case the training stage. Notice that 9 of the 12
+ 45 single-zone-damaged cases combinations have never been analyzed by the system.
+ 25 1R-2M-damaged cases First, noiseless data were input to the system. It
+ 25 1R-3L-damaged cases turned out that most output errors are less than 10%.
+ 25 2M-3L-damaged cases The maximum error is 15.9%.
After the training of the neural networks is completed, Next, tests were performed using the noisy data.
more damage cases were simulated by finite element The noise level was 20%. The maximum bias of the
analysis. Then, the new samples are used to test the output ratios is 16%, and the maximum standard
effectiveness of the monitoring system. The test deviation is 19%. Again, both the biases and standard
results are as following: deviations do not exceed the noise level. Both the
maximum bias and standard deviation are about the
same as the single-zone-damaged case. This shows
6.1 Single Zone Damaged
that system is insensitive to noise in the detection of
There are 163 test cases, including the undamaged two damage zones. Even better, the system has very
case and 162 single-zone-damaged cases. Eighteen good extrapolation capability because most of the test
rigidity ratios are considered for each damaged zone, samples were not covered in the training database.
namely, 0.95, 0.9, 0.85, …, 0.15, 0.1. Using the
parallel networks to analyze these cases, the damaged 6.3 Multiple Zones Damaged
zone can be located accurately in all cases. In most
cases, the errors of the output ratios are less than 5%. Fifty three-zones-damaged cases were randomly
Take network 1M-2L for example, the output and exact chosen to examine the ability of the monitoring system
values r1M are very close, as shown in Fig. 17. The to detect multiple damage zones. Although the
maximum error among all cases is 8.2%. three-zones-damaged cases are not included in the
training database, the damaged locations still can be
identified successfully in all these cases. Furthermore,
the output errors are all less than 15%, as shown in
Table 1.

6.4 Zones Partially Damaged

In all the aforementioned cases, damage of a zone
means the rigidity reduction of both beam elements in
the zone. However, this is rarely the case for real
damages. Hence, the case in which a zone is partially
damaged is also tested. The partially damaged zones
considered here are 1M (elements 5 and 6), 1R
(elements 9 and 10), and 2M (elements 15 and 16).
The rigidity ratios for the damaged elements are
identically 0.5.
The output ratios are shown in Table 2. It is seen
Fig. 17 Test results of network 1M  noise free that the results are satisfactory. The damaged

The Chinese Journal of Mechanics, Vol. 17, No. 3, September 2001 163
 Table 1 System output  multiple zones damaged

damage ri 1L 1M 1R 2L 2M 2R 3L 3M 3R
1L, 1M, 2L 65%, 85%, 35% 0.617 0.859 1.008 0.343 0.995 0.998 0.992 0.990 0.994
1L, 1M, 2M 50%, 75%, 95% 0.506 0.756 1.000 0.991 0.977 0.997 0.993 0.990 0.993
1L, 1M, 3M 95%, 50%, 65% 0.970 0.522 1.017 1.003 0.993 0.996 1.004 0.642 0.998
1L, 1R, 2M 95%, 35%, 10% 0.962 0.960 0.338 1.020 0.073 1.017 1.012 0.983 0.984
1L, 1R, 2R 40%, 65%, 55% 0.410 0.993 0.657 0.990 0.995 0.531 1.000 0.991 0.996
1L, 1R, 3L 20%, 95%, 65% 0.166 0.994 0.974 0.976 0.986 0.997 0.627 0.990 0.995
1L, 1R, 3L 65%, 85%, 35% 0.629 0.988 0.880 1.006 0.986 1.001 0.344 1.005 0.992
1L, 1R, 3R 50%, 75%, 95% 0.508 0.995 0.751 0.996 0.992 0.990 0.998 0.994 0.956
1L, 2L, 3M 95%, 50%, 65% 0.963 0.994 1.007 0.460 0.996 1.004 1.001 0.639 0.992
1L, 2L, 3R 40%, 65%, 55% 0.411 0.992 0.999 0.648 0.994 0.972 0.999 1.005 0.577
1L, 2M, 2R 65%, 85%, 35% 0.630 0.993 0.989 0.995 0.921 0.370 1.014 0.970 0.991
1L, 2M, 2R 85%, 80%, 50% 0.879 0.991 0.997 1.008 0.859 0.524 1.010 0.989 0.992
1L, 2M, 3L 45%, 85%, 85% 0.465 0.997 0.979 0.994 0.888 0.995 0.881 1.002 0.993
1L, 2M, 3R 15%, 60%, 60% 0.122 0.994 0.994 0.984 0.491 0.992 0.978 0.987 0.605
1L, 2M, 3R 80%, 25%, 40% 0.805 0.978 1.006 1.020 0.301 1.011 1.003 0.999 0.416
1L, 2R, 3L 45%, 95%, 95% 0.466 0.998 1.000 0.989 0.991 0.966 0.963 0.995 0.987
1L, 2R, 3L 70%, 45%, 30% 0.676 0.989 1.009 0.999 0.997 0.404 0.376 1.008 0.988
1L, 2R, 3M 10%, 90%, 50% 0.101 1.001 1.001 0.894 0.969 0.856 0.988 0.480 0.978
1L, 2R, 3R 30%, 35%, 15% 0.284 0.992 0.999 0.972 0.994 0.322 1.011 0.982 0.120
1L, 3L, 3M 50%, 30%, 90% 0.506 0.987 0.985 1.007 0.982 0.997 0.327 0.932 0.995
1M, 1R, 2M 45%, 30%, 10% 0.986 0.596 0.411 1.021 0.062 1.017 1.011 0.980 0.983
1M, 1R, 3L 90%, 70%, 70% 0.992 0.926 0.706 0.998 0.983 0.998 0.675 0.997 0.992
1M, 1R, 3R 50%, 45%, 75% 0.988 0.425 0.545 0.995 0.980 1.000 0.996 1.000 0.861
1M, 2L, 2M 55%, 90%, 55% 0.992 0.595 1.018 0.986 0.497 1.005 0.994 0.996 1.003
1M, 2L, 2R 15%, 90%, 40% 1.008 0.134 1.021 0.923 0.983 0.435 1.007 0.981 0.982
1M, 2L, 2R 65%, 55%, 25% 0.984 0.605 1.017 0.543 0.988 0.264 1.009 0.971 0.983
1M, 2L, 3M 90%, 70%, 50% 0.993 0.944 0.992 0.676 0.992 0.986 1.012 0.568 1.003
1M, 2L, 3R 15%, 55%, 75% 1.008 0.113 1.021 0.561 0.993 1.004 0.989 1.001 0.738
1M, 2M, 2R 45%, 45%, 90% 0.994 0.548 1.020 1.019 0.437 0.979 1.008 1.002 0.997
1M, 2M, 3L 15%, 20%, 30% 1.005 0.235 1.021 1.021 0.076 1.017 0.407 0.994 0.989
1M, 3L, 3R 45%, 40%, 20% 0.991 0.452 1.015 1.004 0.989 0.998 0.464 1.005 0.161
1M, 3M, 3R 95%, 85%, 40% 0.993 0.974 0.994 0.999 0.995 0.990 1.005 0.905 0.428
1R, 2L, 2R 10%, 35%, 60% 0.975 0.936 0.100 0.412 0.952 0.617 1.019 0.974 0.985
1R, 2L, 3R 25%, 75%, 45% 0.986 0.981 0.243 0.840 0.990 0.999 0.998 0.997 0.478
1R, 2M, 3L 20%, 80%, 50% 0.985 0.980 0.195 0.991 0.942 1.002 0.494 1.004 0.987
1R, 2M, 3L 50%, 50%, 80% 0.989 0.988 0.512 1.011 0.527 1.004 0.837 0.998 1.001
1R, 2M, 3R 25%, 70%, 50% 0.988 0.985 0.253 0.994 0.702 1.000 0.985 0.999 0.540
1R, 2R, 3L 30%, 90%, 65% 0.989 0.987 0.295 0.991 0.994 0.964 0.639 1.004 0.990
1R, 2R, 3M 70%, 15%, 40% 0.982 0.993 0.726 0.997 0.992 0.140 1.010 0.388 0.991
1R, 2R, 3R 45%, 35%, 95% 0.986 0.989 0.536 0.995 0.986 0.372 1.014 0.986 0.958
1R, 3L, 3R 50%, 65%, 20% 0.989 0.988 0.480 0.995 0.998 0.991 0.634 0.998 0.154
1R, 3M, 3R 45%, 75%, 20% 0.990 0.989 0.413 0.989 0.995 0.974 0.988 0.804 0.160
2L, 2M, 3M 40%, 50%, 30% 0.983 1.009 1.019 0.296 0.566 0.995 1.017 0.318 1.003
2L, 2R, 3M 80%, 40%, 45% 0.984 1.001 1.008 0.840 0.997 0.521 1.003 0.346 0.870
2L, 3L, 3R 25%, 95%, 35% 0.977 0.980 1.012 0.244 0.981 0.989 0.966 0.999 0.348
2M, 2R, 3L 55%, 35%, 80% 0.990 0.980 1.007 1.016 0.637 0.350 0.935 0.996 0.992
2M, 2R, 3M 10%, 60%, 50% 0.989 0.970 0.931 1.021 0.098 0.680 1.021 0.548 1.000
2M, 3L, 3R 25%, 30%, 65% 0.986 0.963 1.002 1.020 0.203 1.015 0.274 0.996 0.643
2R, 3L, 3R 50%, 70%, 15% 0.991 0.990 0.997 0.993 1.001 0.467 0.734 1.005 0.119
3L, 3M, 3R 40%, 90%, 40% 0.991 0.992 0.986 0.998 0.992 1.000 0.415 0.941 0.428

164 The Chinese Journal of Mechanics, Vol. 17, No. 3, September 2001
 Table 2 System output  zones partially damaged

damaged
1L 1M 1R 2L 2M 2R 3L 3M 3R
elements
5 0.99 0.63 1.00 0.99 1.00 1.00 1.00 0.99 0.99
6 0.99 0.65 1.01 0.99 1.00 1.00 1.00 1.00 1.00
5, 6 0.99 0.50 1.01 0.98 0.99 1.00 1.00 1.00 1.00
9 0.99 1.00 0.66 0.98 0.99 0.99 1.00 1.00 1.00
10 0.99 0.99 0.63 1.01 0.99 1.00 0.99 1.00 0.99
9, 10 0.99 0.99 0.50 1.00 0.99 0.99 1.00 1.00 0.99
15 0.99 0.99 0.96 1.01 0.68 1.00 1.01 1.00 1.00
16 0.99 1.00 1.00 1.00 0.65 1.01 0.99 0.99 1.00
15, 16 0.99 0.99 1.00 1.01 0.51 1.01 0.99 1.00 1.00

locations can be identified without difficulty. Therefore, equipped with suitable measuring device, the
Furthermore, when only one element of the zone is proposed ANN has a good chance to become an
damaged, the output ratio lies between the true rigidity effective monitoring system of bridges.
ratios of the two elements. Therefore, even when the
zone is partially damaged, the monitoring system still
can identify the damage location and damage level ACKNOWLEDGMENT
successfully.
This work was supported by the National Science
Council of the Republic of China under Grant
7. CONCLUSIONS NSC85-2211-E-002-017.

A health monitoring system for a bridge is developed

in this paper. Different from previous approaches, the NOMENCLATURE
longitudinal elongations are used as an indicator to
detect the damage of the bridge. The backpropagation w transverse displacement of the beam
neural networks are adopted in the data processing. EI flexural rigidity
Several conclusions can be drawn from the numerical
ρA mass density per unit length
tests:
(1) The longitudinal elongations of a bridge subjected M mass of the truck
to a moving truck can reflect the damage condition g gravitational acceleration
of the bridge. δ Dirac delta function
(2) The damage zones of the bridge can be divided into
V velocity of truck
several groups. Each group contains at most three
zones: the middle of a span and the zones next to the η position of the truck
supports in the adjacent spans. The extreme K stiffness matrix
elongations of a zone are predominated by the M mass matrix
condition of zones in the same group. U nodal displacement vector
(3) According to the zone grouping, the bridge
monitoring system can be constructed using several F external load vector
neural networks. Each neural network monitors Kij element of stiffness matrix
the health of a certain group of zones. Such Mij element of mass matrix
parallel structure can reduce the number of training Fi external load vector
samples and training time.
xA , xB coordinates of element nodes
(4) The monitoring system developed in this study can
detect the locations and levels of damages ϕi shape function
accurately. The system is effective even in the le element length
case of multiple damages or partial damage. x local coordinate of element
(5) The monitoring system is insensitive to noise.
EIi original flexural rigidity of zone i
Since the monitoring system is developed based on
simulation data, its practicality needs to be justified by EIi′ current flexural rigidity of zone i
experiments. However, it is believed that the zonal ri rigidity ratio
elongation of a bridge will increase as damage occurs. f activation function

The Chinese Journal of Mechanics, Vol. 17, No. 3, September 2001 165
 sj input of activation function 6. Shinozuka, M., Yun, C.-B. and Imai, H., “Identification of
output of activation function Linear Structural Dynamic Systems,” Journal of
tk
Engineering Mechanics, ASCE, 108(6), pp. 1371–1390
Wkj weight of a link (1982).
d target ouput 7. Agbabian, M. S., Masri, S. F., Miller, R. K. and Caughey,
o network output T. K.,. “System Identification Approach to Detection of
E sum-squared error Structural Changes,” Journal of Engineering Mechanics,
ASCE, 117(2), pp. 370–390 (1991).
wmax, wmin maximal and minimal elongations of the
8. Yun, C.-B. and Shinozuka, M., “Identification of
current bridge Nonlinear Structural Dynamic Systems,” Journal of
wmax , wmin maximal and minimal elongations of the Structural Mechanics, 8(2), pp. 187–203 (1980).
original bridge 9. Hoshiya, M. and Saito, E., “Structural Identification by
ymax, ymin normalized maximal and minimal Extended Kalman Filter,” Journal of Engineering
elongations Mechanics, ASCE, 110(12), pp. 1757–1770 (1984).
10. Pandey, P. C. and Barai, S. V., “Multilayer Perceptron in
Damage Detection of Bridge Structures,” Computer and
REFERENCES Structures, 54(4), pp. 597–608 (1995).
11. Wu, X., Ghaboussi, J. and Garrett, J. H., “Use of Neural
1. Chen, J.-C. and Garba, J. A., “On Orbit Damage Networks in Detection of Structure Damage,” Computer
Assessment for Large Space Structures,” AIAA Journal, and Structures, 42(4), pp. 649–659 (1992).
26(9), pp. 1119–1126 (1988). 12. Elkordy, M. F. and Chang, K. C., “Neural Networks
2. Yao, G. C., Chang, K. C. and Lee, G. C., “Damage Trained by Analytical Simulated Damage State,” Journal
Diagnosis of Steel Frames Using Vibrational Signature of Computing in Civil Engineering, ASCE, 7(2), pp.
Analysis,” Journal of Engineering Mechanics, ASCE, 130–145 (1993).
118(9), pp. 1949–1961 (1992). 13. Tsou, P. and Shen, M.-H. H., “Structure Damage
Detection and Identification Using Neural Networks,”
3. Liu, P.-L., “Identification and Damage Detection of AIAA Journal, 32(1), pp. 176–183 (1994).
Trusses Using Modal Data,” Journal of Structural
14. Blejwas, T. E., Feng, C. C. and Ayre, R. S.,. “Dynamic
Engineering, ASCE, 121(4), pp. 599–608 (1995).
Interaction of Moving Vehicles and Structures,” Journal
4. Sanayei, M. and Onipede, O., “Damage Assessment of of Sound and Vibration, 67(4), pp. 513–521 (1979).
Structures Using Static Test Data,” AIAA Journal, 29(7), 15. Zurada, M. J., Introduction to Artificial Neural Systems,
pp. 1174–1179 (1991). Info Access & Distribution Pte Ltd., Singapore (1992).
5. Liu, P.-L. and Chian, C.-C., “Parametric Identification of
Truss Structures Using Static Strains,” Journal of
Structural Engineering, ASCE, 123(7), pp. 927–933 (Manuscript received July 31, 2001,
(1997). Accepted for publication Sept. 5, 2001.)

166 The Chinese Journal of Mechanics, Vol. 17, No. 3, September 2001