Chapter 8

Moving Force Identification with

Generalized Orthogonal
Function Expansion
8.1 Introducti on
Existing methods of moving load identification can be broadly classified into two
categories, with one based on a continuous bridge model and modal superposition
technique to decouple the equation of motion and the subsequent solution using opti-
mization schemes as in Chapters 5 and 6. The second category is based on discrete
bridge model with the finite element method to decouple the equation of motion, such
as the state space approach in Chapter 7 and the finite element method (FEM) in
Chapter 9. The modal superposition technique has good accuracy for identification
but it demands extensive computation when multiple vehicles cross a multi-span bridge
structure. The FEM approach is flexible when dealing with vehicular axle-loads mov-
ing on top of a bridge system with complex boundary conditions. However, a great
deal of care must be used in transforming the displacement or strain to velocities and
accelerations (OConnor and Chan, 1988). Numerical differentiation may lead to large
error in the identified results.
This chapter introduces the moving load identification through the generalized
orthogonal function expansion to overcome the above computation problem. The
method has efficient computational performance and good identification accuracy,
especially with the orthogonal function smoothing technique to obtain the veloci-
ties and accelerations from the measured strains (Zhu and Law, 2001a). Orthogonal
functions, such as the generalized orthogonal functions and wavelets, are introduced
in Section 8.2. The moving load identification with these functions is presented in
Section 8.3. The applications by numerical simulations and laboratory experiments
are discussed in Sections 8.4 and 8.5.
8.2 Orthogonal Functi ons
8.2.1 Seri es Expansi on
In order to improve the accuracy of the computed velocities and accelerations from
strains or displacement measurements, the measurement is approximated with an
analytical function. The velocity and acceleration are then obtained by differentiation
of the function. The function is represented as a series expansion:
f (t) =
m

i=1
a
i

i
(t) (8.1)
210 Movi ng Loads Dynami c Anal ysi s and I dent i f i cat i on Techni ques
where
i
(t) is the set of functions. There are many series expansions, such as polynomial
functions, Fourier series and orthogonal functions, such as the Legendre polyno-
mials, Chebyshev polynomials, Bessel functions and B-spline functions. Generalized
orthogonal functions and wavelets will be discussed in the following sections.
8.2.2 General i zed Orthogonal Functi on
The Chebyshev polynomial has been used widely in numerical analysis. This section
introduces the generalized orthogonal function from the first kind of Chebyshev poly-
nomial T
n
(x). It is a polynomial in x of degree n, defined by the relation (Mason and
Handscomb, 2003):
T
n
(x) = cos n when x = cos (8.2)
If the range of the variable x is in the interval [1,1], then the range of the cor-
responding variable can be taken as [0, ]. These ranges are traversed in opposite
directions, since x=1 corresponds to = and x=1 corresponds to =0. We have
the fundamental recurrence relation from Equation (8.2):
T
n
(x) = 2xT
n1
(x) T
n2
(x), (n = 2, 3, . . .) (8.3)
with the initial conditions:
T
0
(x) = 1, T
1
(x) = x (8.4)
All the polynomials {T
n
(x)} can be generated recursively from Equation (8.3). The
orthonormal polynomials can be obtained as follows, with scaling of the polynomials:
1

T
0
(x),
_
2

T
i
(x), i = 1, 2, . . .
_
(8.5)
For an independent variable t in a general range [0, T], we can map the independent
variable t to the variable x with the transformation
x =
2t
T
1 (8.6)
and this leads to a shifted Chebyshev polynomial (of the first kind) T

n
(t) of degree n
in variable t in the interval [0, T] which is a generalized orthogonal function given as:
T

1
=
1

2
=
_
2

_
2
T
t 1
_
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

Movi ng Force I dent i f i cat i on 211
T

3
=
_
2

_
2
_
2
T
t 1
_
2
1
_

T

j+1
= 2
_
2
T
t 1
_
T

j
T

j1
(8.7)
8.2.3 Wavel et Deconvol uti on
The Daubechies wavelets and associated scaling functions
j,k
(t) are obtained by
translation and dilation of functions (t) and (t) respectively (Law et al., 2008).

J,k
(t) = 2
J/2
(2
J
t k) J, k Z (8.8)

J,k
(t) = 2
J/2
(2
J
t k) J, k Z (8.9)
where J is the resolution. The scaling function (t) and wavelet function (t) can be
derived from the dilation equation as:
(t) =

k
a
k
(2t k) (8.10)
(t) =

k
(1)
k
a
1k
(2t k) (8.11)
where a
k
, a
1k
are the filter coefficients and they are fixed for specific wavelet or
scaling function basis. It is noted that only a finite number of a
k
, a
1k
are nonzero for
compactly supported wavelets.
The scaling function (t) and wavelet function (t) have the following properties:

(t)dt = 1 (8.12)

(t j)(t k)dt =
j,k
, j, k Z (8.13)

t
m
(t)dt = 0, (m = 0, 1, . . . , L/2 1) (8.14)
where m denotes the number of vanish moments and L is the order of Daubechies
wavelet with L=2m.
The translation of the scaling and wavelet functions on each fixed scale forms the
orthogonal subspaces:
V
J
= {2
J/2
(2
J
t k), J Z} (8.15)
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

212 Movi ng Loads Dynami c Anal ysi s and I dent i f i cat i on Techni ques
W
J
= {2
J/2
(2
J
t k), J Z} (8.16)
such that V
J
forms a sequence of embedded subspaces:
{0}, . . . , V
1
V
0
V
1
, . . . , L
2
(R) and V
J+1
= V
J
W
J
(8.17)
where is the operator for the addition of two subspaces. At a certain resolution J, the
approximation of a function f (t) in L
2
(R) space using
J,k
(t) as basis can be denoted as:
P
J
(f ) =

k

J,k

J,k
(t), J, k Z (8.18)
where P
J
(f ) is the approximation of f (t) and
J,k
is the approximation coefficient.
Let Q
J
(f ) be the detail of the function using
J,k
(t) as basis at the same level J, and
Q
J
(f ) =

J,k

J,k
(t), J, k Z (8.19)
where

J,k
is the detail coefficient. The approximation P
J+1
(f ) of the next level (J +1)
of resolution is given by:
P
J+1
(f ) = P
J
(f ) +Q
J
(f ) (8.20)
This forms the basis of multi-resolution analysis associated with wavelet
approximation.
The Wavelet-Galerkin approximation to the signal f (t) at a certain resolution J can
be expressed as:
h(t) =

k

J,k
2
J/2
(2
J
t k), J, k Z (8.21)
from Equations (8.8) and (8.18). Substituting y =2
J
t into Equation (8.21), we obtain:
h(y) =

J,k
(y k);
J,k
= 2
J/2

J,k
, J, k Z (8.22)
If y takes up only integer values, the approximation is discretized at all dyadic points
with t =2
J
y as:
h(i) = h(iy) = h
i
, (i = 0, 1, 2, . . . , N
T
) (8.23)
where N
T
is the number of time instances. Equation (8.22) can be rewritten as:
h
i
=

ik
=

ik

k
(8.24)
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

Movi ng Force I dent i f i cat i on 213
with
k
=(k). In matrix form this becomes:
_

_
h
1
h
2
h
3
.
.
.
.
.
.
.
.
.
.
.
.
h
N
T
1
_

_
=
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
0 0 0
L2

2

1

1
0 0 0
3

2

2

1
0 0
4

3
.
.
.
.
.
.

L2

L3

L4
0 0
0
L2

L3
0 0
.
.
.
.
.
.
0 0 0
L3

1
0
_

_
_

3
.
.
.
.
.
.
.
.
.
.
.
.

N
T
1
_

_
(8.25)
The periodic boundary condition has been included in Equation (8.25) for the finite
domain analysis denoted as:
_

1
=
N
T
1

2
=
N
T
3
.
.
.

L+2
=
N
T
L+2
_

_
and
_

N
T
=
0

N
T
+1
=
1
.
.
.

N
T
+L2
=
L2
_

_
(8.26)
8.3 Movi ng Force Identi f i cati on
8.3.1 Beam Model
8.3.1.1 Gener al i zed Or t hogonal Funct i on Expansi on
The strain in the Euler-Bernoulli beam at a point x and time t can be written as:
(x, t) = h

2
w(x, t)
x
2
(8.27)
where h is the distance between the lower surface and the neutral plane of bending of
the beam. Substituting the transverse displacement w(x, t) in Equation (2.7) of Chapter
2 into Equation (8.27), and assuming there are N modes in the responses, we have:
(x, t) = Q (8.28)
where
= {h

1
(x), h

2
(x), . . . , h
N

(x)}; Q = {q
1
(t), q
2
(t), . . . , q
N
(t)}
T
.
and

i
(x) is the second derivative of
i
(x).
The strain is then approximated by a generalized orthogonal function T(t) as shown
in Equation (8.5) as:
(x, t) =
N
f

i=1
T
i
(t)C
i
(x) (8.29)
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

214 Movi ng Loads Dynami c Anal ysi s and I dent i f i cat i on Techni ques
where {T
i
(t), i =1, 2, . . . , N
f
} are the generalized orthogonal functions; {C
i
(x), i =1,
2, . . . , N
f
} is the vector of coefficients in the expanded expression. Note that the wavelet
form of the orthogonal function in Equation (8.21) can also be used. The strains at
the N
s
measuring points can be expressed as:
= CT (8.30)
where
T = {T
0
(t), T
1
(t), . . . , T
N
f
(t)}
T
;
= {(x
1
, t), (x
2
, t), . . . , (x
N
s
, t)}
T
;
C =
_
_
_
_
_
_
C
10
(x
1
) C
11
(x
1
) C
1N
f
(x
1
)
C
20
(x
2
) C
21
(x
2
) C
2N
f
(x
2
)
.
.
.
.
.
.
.
.
.
.
.
.
C
N
s
0
(x
N
s
) C
N
s
1
(x
N
s
) C
N
s
N
f
(x
N
s
)
_

_
and {x
1
, x
2
, . . . , x
N
s
} is the vector of the location of the strain measurements. By the
least-squares method, the coefficient matrix can be obtained as:
C = T
T
(TT
T
)
1
(8.31)
Substitute Equation (8.28) into Equation (8.30), we have:
Q = (
T
)
1

T
CT (8.32)
where
=
_
_
_
_
_
_
_
h

1
(x
1
) h

2
(x
1
) h

N
(x
1
)
h

1
(x
2
) h

2
(x
2
) h

N
(x
2
)
.
.
.
.
.
.
.
.
.
.
.
.
h

1
(x
N
s
) h

2
(x
N
s
) h

N
(x
N
s
)
_

_
and it can be obtained from Equation (2.4) in Chapter 2.
8.3.1.2 Movi ng For ce I dent i f i cat i on Theor y
The vector of generalized coordinates obtained fromEquation (8.32) can be substituted
into Equation (2.8) for the beam, and rewrite it in matrix form to become:
I

Q+C
d

Q+KQ = BP (8.33)
where
C
d
= diag(2
i

i
);
K = diag(
2
i
)
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

Movi ng Force I dent i f i cat i on 215
B =
_
_
_
_
_
_

1
( x
1
(t))/M
1

1
( x
2
(t))/M
1

1
( x
N
p
(t))/M
1

2
( x
1
(t))/M
2

2
( x
2
(t))/M
2

2
( x
N
p
(t))/M
2
.
.
.
.
.
.
.
.
.
.
.
.

N
( x
1
(t))/M
N

N
( x
2
(t))/M
N

N
( x
N
p
(t))/M
N
_

_
The required

Q and

Q can be obtained by directly differentiating Equation (8.32) to
have:

Q = (
T
)
1

T
C

Q = (
T
)
1

T
C

T
The moving forces obtained from Equation (8.33) using a straight forward least-
squares solution would be unbound. Let the left-hand-side of Equation (8.33) be
represented by U. Regularization technique is used to solve the ill-posed problem in
the form of minimizing the function:
J(P, ) = BP U
2
+ P
2
(8.34)
where is the non-negative regularization parameter.
The success of solving Equation (8.34) lies in how to determine the regularization
parameter . Two methods are used in this chapter. If the true forces are known, the
parameter can be determined by minimizing the error between the true forces and the
predicting values as:
S =

P P (8.35)
In the practical case when the true forces are not known, the method of generalized
cross-validation (GCV) is used to determine the optimal regularization parameter. The
GCV function to be minimized in this work is defined by (Golub et al., 1979):
g() =
B

P U
2
2
{trace[I B((B
T
B + I)
1
B
T
)
1
]}
2
(8.36)
where

P is the vector of estimated forces.
8.3.2 Pl ate Model
The displacement w(x
s
, y
s
, t) at location (x
s
, y
s
) and at time t is rewritten in matrix
form from Equation (3.16) in Chapter 3 as:
w(x
s
, y
s
, t) = W
s
Q (s = 1, 2, . . . , N
s
) (8.37)
where N
s
is the number of measuring points, and Q is a matrix of q
ij
(t) from
Equation (3.16).
W
s
= {W
11
(x
s
, y
s
), W
12
(x
s
, y
s
), . . . , W
1n
(x
s
, y
s
), W
21
(x
s
, y
s
), . . . , W
mn
(x
s
, y
s
)}
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

216 Movi ng Loads Dynami c Anal ysi s and I dent i f i cat i on Techni ques
The modal strains in x-direction can be written as:
W
ij
(x
s
, y
s
) = z
t
_
i
a
_
2
sin
_
i
a
x
s
_
Y
ij
(y
s
), (i = 1, 2, . . . , m; j = 1, 2, . . . , n).
where z
t
is the distance from the measuring point at the outer surface to the neutral
surface of bending. For N
s
measuring points:
w
ns
= W
ns
Q (8.38)
where
w
ns
= [w(x
1
, y
1
, t), w(x
2
, y
2
, t), . . . , w(x
N
s
, y
N
s
, t)]
T
W
ns
=
_
_
_
_
_
_
_
W
11
(x
1
, y
1
) W
12
(x
1
, y
1
) W
mn
(x
1
, y
1
)
W
11
(x
2
, y
2
) W
12
(x
2
, y
2
) W
mn
(x
2
, y
2
)
.
.
.
.
.
.
.
.
.
W
11
(x
N
s
, y
N
s
) W
12
(x
N
s
, y
N
s
) W
mn
(x
N
s
, y
N
s
)
_

_
N
s
mn
The modal displacement can be obtained from Equation (8.38) by least-squares
method as:
Q = (W
T
ns
W
ns
)
1
W
T
ns
w
ns
(8.39)
Since the displacements or strains are measured, the velocities and accelerations can
be obtained by dynamic programming filter (Trujillo and Busby, 1983) or orthogonal
polynomial method described in Section 8.2, and the modal velocities and accelera-
tions are calculated by the least-squares method from Equation (8.39). They are then
substituted into Equation (3.17) for the plate to form the matrix equation:
B = SP (8.40)
where
S =
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
2sin
_

a
x
1
(t)
_
Y
11
( y
1
(t))
ha
_
b
0
Y
2
11
(y)dy
2sin
_

a
x
2
(t)
_
Y
11
( y
2
(t))
ha
_
b
0
Y
2
11
(y)dy

2sin
_

a
x
N
p
(t)
_
Y
11
( y
N
p
(t))
ha
_
b
0
Y
2
11
(y)dy
2sin
_

a
x
1
(t)
_
Y
12
(

y
l
(t))
ha
_
b
0
Y
2
12
(y)dy
2sin
_

a
x
2
(t)
_
Y
12
( y
2
(t))
ha
_
b
0
Y
2
12
(y)dy

2sin
_

a
x
N
p
(t)
_
Y
12
( y
N
p
(t))
ha
_
b
0
Y
2
12
(y)dy
.
.
.
.
.
.
.
.
.
2sin
_
m
a
x
1
(t)
_
Y
mn
( y
l
(t))
ha
_
b
0
Y
2
mn
(y)dy
2sin
_
m
a
x
2
(t)
_
Y
mn
( y
2
(t))
ha
_
b
0
Y
2
mn
(y)dy

2sin
_
m
a
x
N
p
(t)
_
Y
mn
( y
N
p
(t))
ha
_
b
0
Y
2
mn
(y)dy
_

_
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

Movi ng Force I dent i f i cat i on 217
B =
_

_
q
11
(t) +2
11

11
q
11
(t) +
2
11
q
11
(t)
q
12
(t) +2
12

12
q
12
(t) +
2
12
q
12
(t)
.
.
.
q
mn
(t) +2
mn

mn
q
mn
(t) +
2
mn
q
mn
(t)
_

_
; P =
_
p
1
(t), p
2
(t), . . . , p
N
p
(t)
_
T
(8.41)
The moving load P can be obtained by the straightforward least-squares method
from Equation (8.40). But the solutions are frequently unstable in the sense that small
noises in the responses would result in large changes in the predicted moving force.
Regularization technique is utilized to improve the conditioning. The load identifica-
tion is formulated as a nonlinear least-squares problem.
min J(P, ) = (B SP, R(B SP)) + (P, P) (8.42)
where is an optimal regularization parameter or a vector. R is a weight matrix
and it can be determined from the measured information (Santantamarina and Fratta,
1998). Generalized cross-validation method (Golub et al., 1979) and L-Curve method
(Hansen, 1992) are then used to determine the optimal regularization parameter in
this study.
8.4 Appl i cati ons
8.4.1 I denti f i cati on wi th a Beam Model
The method, described in previous sections, is illustrated in the following simulation
studies. The effect of discarding some of the information contained in the measured
responses on the error of identification is studied.
8.4.1.1 Si ngl e- Span Beam
A single span simply supported beam is studied with two varying forces moving on
top at a constant spacing of 4.27m.
f
1
(t) = 9.9152 10
4
[1 +0.1sin(10t) +0.05sin(40t)] N;
f
2
(t) = 9.9152 10
4
[1 0.1sin(10t) +0.05sin(50t)] N.
(8.43)
The parameters of the beam are as follow: EI =2.510
10
Nm
2
, A=5000kg/m,
L=30m, h=1m. The first eight natural frequencies of the beamare 3.9, 15.61, 35.13,
62.48, 97.58, 140.51, 191.25 and 249.8Hz, and they are used in the computation of
the analytical mode shapes from Equation (2.4) in Chapter 2. The forces are moving
at a speed of 30m/s. Random noise is added to the calculated strains to simulate the
polluted measurement and 1, 5 and 10 percent noise levels are studied with:
=
calculated
+E
p
N
iose
var(
calculated
) (8.44)
where is the vector of strains; E
P
is the noise level; N
iose
is a standard normal dis-
tribution vector with zero mean and unit standard deviation;
calculated
is the vector of
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

218 Movi ng Loads Dynami c Anal ysi s and I dent i f i cat i on Techni ques
Table 8.1 Error of Identification for single span beam
Number of Noise level
mode shapes
1% 5% 10%
First Second First Second First Second
force force force force force force
2 31.51 31.53 45.11 45.05 174.27 112.50
3 11.56 12.00 13.04 13.24 16.31 16.08
4 6.78 6.18 7.05 6.95 8.47 8.69
5 5.16 3.62 5.07 3.89 5.46 4.80
6 3.90 3.10 4.02 3.28 4.14 3.66
7 3.43 2.99 3.45 3.13 3.66 3.44
8 3.15 2.86 3.19 3.01 3.42 3.29
9 9.52 8.96 9.48 8.94 9.48 8.94
10 18.02 17.30 18.51 17.99 18.42 17.88
calculated strains; var(
calculated
) is the standard deviation of
calculated
. The errors in the
identified forces are calculated as:
Error =

P P
True

P
True

100% (8.45)
Table 8.1 shows the errors of identification from using different number of mode
shapes in the identification. The time step is 0.001s in the calculation. The strain
consists of responses from the first eight mode shapes polluted with 5 percent noise
level. Ten measuring points are available in the identification and they are evenly
distributed along the beam length. The different combination of number of mode
shapes used in the identification and the number of measuring points are studied.
Figure 8.1 shows the identified results using three and six mode shapes. The following
observations were made:
1. Results in Table 8.1 show that the errors in the identified forces are insensitive to
the noise level in the responses. This is because orthogonal functions have been
used to approximate the strains in the identification, and this approximation
suppresses the errors due to high frequency measurement noise.
2. When the number of mode shapes used in the identification is the same as
the number of mode shapes in the responses, i.e. eight mode shapes, the
errors of identification are the smallest. The errors become large when the number
of mode shapes used in identification is either larger or smaller than the number
of mode shapes in the responses. This indicates that the pairing of the number of
mode shapes in both the responses and the identified forces has a large effect on
the errors in the identification. The correct pairing can be determined from an
inspection of the frequency content in the measured responses.
3. Figure 8.1 shows that there are large discrepancies in the identified forces near the
beginning and the end of the moving forces when only three modes are used in the
identification. These discrepancies are much less when six modes are used. This is
because of the sudden appearance and disappearance of the forces at these points,
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

Movi ng Force I dent i f i cat i on 219
15
10
4
The first axle force
10
5
5
N
0
0 0.2 0.4 0.8 1.2 0.6 1
15
10
4
The second axle force
Time (s)
10
0
N
5
0 0.2 0.4 0.8 1.2 0.6 1
Figure 8.1 Identified results with different number of mode shapes ( True loads; - - from 3
modes; from 6 modes)
which can be represented by an equivalent impulsive force. These impulsive forces
excite the beam with a broad-band vibration that covers a large number of modal
frequencies. Therefore, more mode shapes should be used in the identification to
take advantage of the information of the forces at higher modal frequencies in the
responses at the beginning and the end of the time histories.
8.4.1.2 Two- Span Cont i nuous Beam
Table 8.2 shows the errors in the identified moving forces on a two-span continuous
beam with different numbers of mode shapes and numbers of measuring points. The
parameters of the beam are the same as for the single-span beam except that each span
measures 30mlong. The first eight natural frequencies of the beamare 3.9, 6.1, 15.61,
19.75, 35.12, 41.22, 62.43 and 70.48Hz. Figure 8.2 shows the identified forces from
using strains polluted with 5 percent noise level at six measuring points. Inspection of
the results in Table 8.2 and Figure 8.2 gives the following observations:
1. Results in Table 8.2 showthat the errors increase as the noise level in the response
increase. The errors are more than twice of that under similar conditions for the
single-span beam. Therefore, moving load identification in a multi-span beam
would be less accurate than that in a single-span beam.
2. When the number of mode shapes used in the identification equals to that in the
responses as shown in the first two rows and the lower part of Table 8.2, the
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

Table 8.2 Identified Errors for Two-span Beam
No. of mode No. of mode No. of Noise level
shape in shapes in measuring
responses identification points 1% 5% 10%
N
1
N
2
N
s
First Second First Second First Second
force force force force force force
10 10 14 7.85 9.07 18.48 19.44 27.64 28.99
10 10 10 7.78 8.98 16.67 17.98 26.28 27.42
10 9 10 8.90 10.61 15.95 17.23 24.24 25.12
10 8 10 11.54 13.71 23.05 24.12 31.37 32.59
10 7 10 13.81 16.19 17.02 19.01 21.99 23.70
10 6 10 17.05 19.55 20.76 22.99 26.61 28.54
6 6 6 15.99 18.15 20.15 21.91 26.26 27.78
6 6 8 16.06 18.21 20.93 22.69 27.74 29.24
6 6 10 15.98 18.15 19.98 21.77 26.11 27.69
6 6 12 15.90 18.08 18.90 20.73 24.02 25.65
6 6 14 15.85 18.04 18.19 20.06 22.47 24.14
20
10
4
The first axle force
15
10
5
N
0
0 0.5 1 2.5 1.5 2
5
20
10
4
The second axle force
Time (s)
15
5
0
N
5
0 0.5 1 2.5 1.5 2
10
Figure 8.2 Identified forces on continuous beam from different number of mode shapes (True
loads; - - from with 3 modes; from with 6 modes)
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

Movi ng Force I dent i f i cat i on 221
13.715m
0.2m
2.743m
4.865m
24.325m
Figure 8.3 A typical single span bridge deck
errors in the identified forces varies only slightly with more measuring points.
The number of the measuring points is best selected to be equal to the number of
mode shapes.
3. Results from the upper part of Table 8.2 also show that the errors would be
smallest when the number of mode shapes in the identification is the same as that
in the responses. This confirms the observation made in the case of the single
span beam.
4. The identified forces in Figure 8.2 have large fluctuations close to the intermediate
support at 1.0s. This is due to the presence of the small responses generated at
this time instance with subsequently a small signal to noise ratio in the measured
data.
8.4.2 I denti f i cati on wi th a Pl ate Model
A simply supported prototype bridge composed of five I-section steel girders and a
concrete deck as shown in Figure 8.3. It is noted that the model is similar to the one
used by Fafard and Mallikarjuna (1993) in their study of bridgevehicle interaction.
It is also similar to the continuous bridge deck shown in Figure 4.12 of this book. It is
wide enough to accommodate four lanes of traffic. The parameters of the bridge deck
are listed as follow: a =24.325m, b=13.715m, h=0.2m, E
x
=4.168210
10
N/m
2
E
y
=2.973310
10
N/m
2
, =3000kg/m
3
,
xy
=0.3. For the steel I-beam: web
thickness =0.01111m, web height =1.490m, flange width =0.405m, flange
thickness =0.018m. For the diaphragms, the distance between two diaphragms is
4.865m, cross-sectional area =0.001548m
2
, I
y
=0.707 10
6
m
4
, I
z
=210
6
m
4
,
J =1.210
7
m
4
. The rigidities in the x-direction of the equivalent orthotropic plate
can be calculated according to Bakht and Jaeger (1985), as D
x
=2.41510
9
Nm,
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

222 Movi ng Loads Dynami c Anal ysi s and I dent i f i cat i on Techni ques
Table 8.3 Natural Frequencies of the equivalent orthotropic plate (Hz)

n 1 2 3 4 5 6 7 8
m
1 4.96* 6.31 10.01 16.07 24.81 36.46 51.08 68.70
2 19.84* 21.26 25.41 32.17 41.51 53.49 68.22 85.82
3 44.65* 46.07 50.32 57.33 67.06 79.48 94.60 112.49
4 79.37* 80.80 85.07 92.17 102.06 102.50 114.74 115.58
5 124.01* 124.02 125.42 125.44 125.46 129.40 129.43 129.51
* Longitudinal bending modes.
D
y
=2.181310
7
Nm, D
xy
=2.219510
8
Nm. The natural frequencies of the bridge
deck are listed in Table 8.3. It should be noted that this structure is similar to, but not
the same as, that in Chapter 4.
A two-axle vehicle model is used in the simulation. The axle spacing and wheel
spacing are 4.26mand 1.829mrespectively. The four wheel loads are listed as follows:
_

_
P
1
(t) = 3134.

(1 +0.1sin(10t) 0.1sin(20t) +0.05sin(40t)) kg

P
2
(t) = 6166.

(1 0.1sin(10t) 0.1sin(20t) +0.05sin(40t)) kg

P
3
(t) = 3134.

(1 +0.1sin(10t) +0.1sin(20t) +0.05sin(40t)) kg

P
4
(t) = 6166.

(1 +0.1sin(10t) +0.1sin(20t) +0.05sin(40t)) kg

(8.46)
where P
1
and P
3
are the front wheels and P
2
and P
4
are the rear wheels with P
2
and P
4
after P
1
and P
3
respectively. The total vehicle load is 18.6 Tonnes and the proportion
of axle loads follows the pattern of vehicle type H20-44 from AASHTO (2002). The
vehicle moving speed is 20m/s, and the time step of analysis is 0.001s in the simulation.
White noise is added to the calculated displacements or strains to simulate the polluted
measurements.
8.4.2.1 St udy on t he Noi se Ef f ect
The vehicle is moving along the centerline of the deck. The measured responses from
25 modes (m=5, n =5) in Table 8.3 are used in the calculation, and the number
of modes in the identification Equation (8.40) is the same as that in the responses.
According to discussions in Chapter 6, the number of measuring points should not be
less than the number of vibration modes in the measured information. And therefore
25 measuring points are selected evenly distributed on the five I-beams. The identified
individual wheel loads fromusing displacement responses with 1 percent and 5 percent
noise levels are shown in Figure 8.4. The following observations are made:
1. The beginning or end of the identified results is under-estimated when there is
noise in the responses. This is due to the small responses at the beginning or the
end of the time duration, and the fact that the regularization parameter has been
optimized over the total time duration of the event.
2. Errors in the identified results increase with the noise level. Hence when the noise
level is high, a data treatment process (such as filtering or smoothing) should be
used to reduce the noise in the responses before the computation.
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

Movi ng Force I dent i f i cat i on 223
5
(a) Wheel load1
Time (s)
10
4
4
3
N
2
1
0
0 0.5 1.5 1
8
(b) Wheel load2
Time (s)
10
4
6
4
N
2
0
0 0.5 1.5 1
5
(c) Wheel load3
Time (s)
10
4
4
3
N
2
1
0
0 0.5 1.5 1
8
(d) Wheel load4
Time (s)
10
4
6
4
N
2
0
0 0.5 1.5 1
Figure 8.4 Identified results with different noise levels ( true load; - - - 1% noise; 5% noise)
8.4.2.2 I dent i f i cat i on wi t h I ncompl et e Modal I nf or mat i on
In practice, the vibration modes selected for identification are not the same as that in
the responses. In general the lower modes of the structure are dominating the measured
responses, and they are used in the identification. The moving loads are identified again
with this incomplete modal information with fewer modes in the identification than
those in the responses. The parameters of the system used in the simulation are the
same as those in the last study. The vehicle is moving along the centerline of the bridge
deck. Table 8.4 shows the errors in the identified results, and Figure 8.5 shows the
identified results using 20 modes (m=4, n =5) or (m=5, n =4) with 1 percent noise
in the responses. The following observations are made from the results:
1. When m4, n 4, an acceptable result can be obtained with most of the errors
less than 10 percent at 1 percent noise level. This is because the natural frequencies
of these modes (shown in Table 8.3) have covered most of the excitation frequency
range of the car as shown in Equation (8.46). In practice, the frequency range
required in the identification can be obtained from the spectrum of the responses.
2. The more vibration modes used in the identification, the fewer errors are found in
the identified results (Table 8.4). However, large errors still exist at the beginning
and the end of the load time histories as seen in Figure 8.5. This is due to the fact
that impulses are generated by the moving loads at the beginning and the end of
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

224 Movi ng Loads Dynami c Anal ysi s and I dent i f i cat i on Techni ques
Table 8.4 Errors (percent) in the identified loads from different measured information
Noise Level 1% 5% 10%
m n Total Axle-1 Axle-2 Axle-1 Axle-2 Axle-1 Axle-2
mode no.
5 5 25 6.97 6.00 17.90 14.46 28.68 22.16
6.51 5.98 15.58 13.38 24.77 20.28
5 4 20 9.05 7.74 17.95 14.80 28.07 21.94
9.23 6.60 16.50 12.78 26.31 20.41
5 3 15 30.56 18.62 32.25 20.32 36.11 23.99
31.27 18.29 32.64 19.75 36.02 22.91
4 5 20 7.06 6.10 18.01 14.59 28.79 22.33
6.59 6.06 15.69 13.51 24.89 20.44
4 4 16 9.13 7.82 18.05 14.93 28.17 22.10
9.34 6.68 16.63 12.91 26.44 20.57
4 3 12 30.86 18.82 32.53 20.51 36.32 24.15
31.57 18.43 32.90 19.88 36.23 23.04
3 5 15 10.82 9.83 23.71 21.93 34.64 32.22
9.90 9.37 20.68 20.12 30.20 29.67
3 4 12 12.57 12.03 23.23 21.97 33.66 31.59
14.67 10.29 23.12 20.03 32.58 30.45
3 3 9 34.69 23.48 36.19 25.91 40.10 32.10
38.30 23.86 38.86 25.21 40.69 30.02
2 5 10 22.91 18.95 35.41 37.72 44.28 49.50
19.71 17.13 32.53 35.17 40.23 46.76
Note:The errors in table correspond to each wheel load as
wheel 1|wheel 2
wheel 3|wheel 4
the time duration, and a lot of higher modes of the structure are excited which
are not covered by the selected vibration modes in the identification.
3. There are large errors in the identified results with m=5, n =3. This shows that
the torsional modes are also very important in the moving load identification on
bridge decks even if the vehicle is moving along the centerline.
8.4.2.3 Ef f ect s of Tr avel Pat h Eccent r i ci t y
There are four lanes on the bridge deck. Normally the vehicle is not moving exactly
along the centerline. Table 8.5 shows the errors in the identified results with the car
moving at different eccentricities and using different number of vibration modes in
the identification. The identified results for different eccentricities with 25 modes
(m=5, n =5) are shown in Figure 8.6. The parameters are the same as for the above
studies, and the responses are calculated with 25 modes (m=5, n =5). The following
intermediate conclusions can be drawn from Table 8.5 and Figure 8.6.
1. When m3, n =5, an acceptable result can be obtained with most of the errors
less than 10 percent at 1 percent noise level. This shows that the method proposed
in the chapter is also effective to identify the eccentric moving loads on the bridge
deck.
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

5
(a) Wheel load1
Time (s)
10
4
4
3
N
2
1
0
0 0.5 1.5 1
8
(b) Wheel load2
Time (s)
10
4
6
4
N
2
0
0 0.5 1.5 1
5
(c) Wheel load3
Time (s)
10
4
4
3
N
2
1
0
0 0.5 1.5 1
8
(d) Wheel load4
Time (s)
10
4
6
4
N
2
0
0 0.5 1.5 1
Figure 8.5 Identified results with different mode combinations (1% noise) ( true load;
- - - m=4, n =5; m=5, n =4)
Table 8.5 Errors (percent) in the identified load moving at different eccentricities
Eccentricity m n 1% 5% 10%
Axle-1 Axle-2 Axle-1 Axle-2 Axle-1 Axle-2
3/8b 5 5 7.20 6.35 14.33 11.33 23.16 17.93
6.16 4.37 19.06 12.26 34.71 21.75
4 5 6.72 6.41 15.25 13.26 24.41 20.99
10.16 9.95 25.20 21.58 41.72 33.84
3 5 11.31 11.56 21.06 22.25 29.91 33.50
13.56 13.25 29.46 27.37 45.00 40.97
5 4 33.78 33.35 38.48 37.06 46.14 42.97
16.83 20.04 26.33 21.33 41.11 26.60
1/8b 5 5 6.46 5.56 15.88 12.86 25.22 19.60
7.73 7.07 15.42 12.35 22.31 20.31
4 5 6.54 5.64 15.96 12.97 25.31 19.72
7.80 7.16 15.51 14.48 22.41 20.47
3 5 10.18 9.17 21.23 19.25 30.81 27.89
11.22 10.93 20.08 21.19 27.42 29.28
5 4 21.16 20.38 23.83 22.02 30.45 26.05
22.03 18.21 23.46 17.44 26.87 18.61
Note:The errors in table correspond to each wheel load as
wheel 1|wheel 2
wheel 3|wheel 4
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

226 Movi ng Loads Dynami c Anal ysi s and I dent i f i cat i on Techni ques
5
(a) Wheel load1
Time (s)
10
4
4
3
N
2
1
0
0 0.5 1.5 1
8
(b) Wheel load2
Time (s)
10
4
6
4
N
2
0
0 0.5 1.5 1
5
(c) Wheel load3
Time (s)
10
4
4
3
N
2
1
0
0 0.5 1.5 1
8
(d) Wheel load4
Time (s)
10
4
6
4
N
2
0
0 0.5 1.5 1
Figure 8.6 Identified results for different eccentricities (1 percent noise) ( true load; -.-.- e =0;
- - - e =1/8b; e =3/8b)
2. In the cases with n <5, there are large errors in the identified eccentric load.
This shows that the torsional modes are more important in the eccentric load
identification than that for the case with the car moving along the centerline.
3. When the eccentricity increases, the errors in the identified results also increase as
seen in Table 8.5. Since the same measured information is used in all sets of iden-
tification, eccentric moving loads are more difficult to be identified accurately.
There is a need for an optimum selection of measuring locations for different
moving load configuration, and it is a subject of further research.
8.5 Laboratory Studi es
8.5.1 Beam Model
8.5.1.1 Exper i ment al Set up and Measur ement s
The measurements from the experimental system described in Section 6.3.4.1 are used
for this study.
8.5.1.2 For ce I dent i f i cat i on
Strains at 1/8L, 1/4L, 3/8L, 1/2L, 3/4L, 7/8L are used in the identification. Table 8.6
shows the correlation coefficients between the measured and the reconstructed strains
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

Movi ng Force I dent i f i cat i on 227
Table 8.6 The correlation coefficients between measured and reconstructed
responses at 5/8L
Case Number of Measuring locations Correlation
Mode Shapes Coefficient
A 3 1/4L,1/2L,3/4L 0.9809
B 4 1/8L,1/4L,1/2L,3/4L 0.9470
C 5 1/8L,1/4L,1/2L,3/4L,7/8L 0.9752
D 3 1/8L,1/4L,3/8L,1/2L,3/4L,7/8L 0.9853
E 4 1/8L,1/4L,3/8L,1/2L,3/4L,7/8L 0.9837
F 5 1/8L,1/4L,3/8L,1/2L,3/4L,7/8L 0.9822
G 6 1/8L,1/4L,3/8L,1/2L,3/4L,7/8L 0.9716
at 5/8L obtained from the identified forces with different number of mode shapes in
the identification. The number of measuring points is taken equal to the number of
mode shapes in the identification. Figure 8.7 shows that the identified forces from
Cases (A) and (G) of the study using 3 and 6 sensors respectively. The combined force
is also presented in Figure 8.7(c). The following observations are made:
1. Table 8.6 shows that the correlation coefficients are all larger than 0.9 for different
combination of modes and measuring points. It shows that the method based
on generalized orthogonal function is effective to identify the moving forces in
practice.
2. There is a low frequency component in the identified individual forces in
Figure 8.7. This is the pitching motion of the moving car.
3. The identified forces from using six modes are closer to the static forces at the
beginning and the end of the time histories than those obtained from using
three modes. This gives experimental evidence that more mode shapes in the
computation should be used to identify the moving forces near these locations.
8.5.2 Pl ate Model
8.5.2.1 Exper i ment al Set - up
The model vehiclebridge system fabricated in the laboratory as described in Section
6.3.4.2 is used for this study. The same model car as described in Section 6.3.4.2 is
used in the experiment. Details on the experimental setup are referred to Figure 6.28.
The layout plan of the sensors is reproduced in Figure 8.8 for easy reference in this
study.
Twenty-five strain gauges were located at the bottomof the ribs to measure the strain
of the bridge deck as shown in Figure 8.8. Six B&K model 4370 accelerometers were
placed at the bottom of beams 4 and 5 at 1/4, 1/2 and 3/4 span for the acceleration
measurements. p
1
(t) and p
3
(t) are the left and right wheel loads at the front looking
in the direction of the traveling path; p
2
(t) andp
4
(t) are the left and right wheel loads
at the back following p
1
(t) and p
3
(t).
The rigidities of the equivalent orthotropic plate are calculated as D
x
=7.3677
10
4
Nm, D
y
=4.269610
3
Nm and D
k
=8.601810
3
Nm. The first ten measured
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

228 Movi ng Loads Dynami c Anal ysi s and I dent i f i cat i on Techni ques
200
100
0
100
200
0 0.5 1.5 2.5 3.5 3 2
(a) The first axle force
Time (s)
N
1
200
100
0
100
200
0 0.5 1.5 2.5 3.5 3 2
(b) The second axle force
Time (s)
N
1
300
200
100
0
100
200
0 0.5 1.5 2.5 3.5 3 2
(c) The resultant force
Time (s)
N
1
Figure 8.7 Identified results from using different modes ( Static load; - - with 3 modes
(Case (A)); with 6 modes (Case (G))
and calculated natural frequencies of the bridge deck are listed in Table 6.14 in
Chapter 6. The response is sampled at 1000Hz, and the number of data in each
record segment is 7680. The average speed of the vehicle on the whole bridge deck is
used for the identification of the moving loads in this study.
8.5.2.2 Wheel Load I dent i f i cat i on
The measured strains are re-sampled to have a time interval of 0.003s. When the
model car is moving along the centerline (Rail 3), Rail 1 or Rail 2 in turn, the strains
at 1/4a, 1/2a and 3/4a of each beam are measured to identify the moving wheel loads.
The average speed of the model car is 0.54m/s. Figures 8.98.11 show the identified
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

Movi ng Force I dent i f i cat i on 229
610 305 305
Nine Photoelectric Sensors in Equal Spacing
Rail 3
Rail 2
Rail 1
(a) Top face of the bridge deck
(b) Bottom face of the bridge deck
(c) Section A-A
(Dimensions are in millimetres)
12.5
Rail
Gauges
Photoelectric
Sensors
6.35
25
A
A
610 610
122
244 1 6 11 16
2 7 12 17
3 8 13 18
4 9 14 19
5 10 15 20
152 305 152
244
610
1
2
2
0
Figure 8.8 Layout of the bridge deck
wheel loads, axle loads and the combined load from different vibration modes used
in the identification with different number of measuring points when the model car
moves along the centerline. Table 8.7 shows the correlation coefficients between the
reconstructed and measured strains at 3/8a of each beam for different moving paths
of the car. The following observations are made from the Figures and Table 8.7.
1. The method based on generalized orthogonal function is effective to identify
individual moving wheel loads and acceptable results can be obtained.
2. The correlation coefficients between the reconstructed and measured strains on
the beams adjacent to the moving path of the car are larger than 0.8. This shows
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

230 Movi ng Loads Dynami c Anal ysi s and I dent i f i cat i on Techni ques
100
50
50
0
N
0 1 2 3
(a) Wheel load1
4 5 6
100
50
50
0
N
0 1 2 3
(c) Wheel load3
4 5 6
200
100
100
0
N
0 1 2 3
(b) Wheel load2
4 5 6
200
100
100
0
N
0 1 2 3
(d) Wheel load4
Time (s)
4 5 6
Figure 8.9 Identified wheel loads for different combinations of measured information ( static
loads; - - m=3, n =2(9); - - - m=3, n =3(15); m=3, n =2(15))
that the method is effective to identify the wheel loads moving with or without
an eccentricity.
3. When the distance between the measuring point and the path of moving car is
large, the correlation coefficient is small, as seen from Beam #1 for e =3/8b. This
is because of the small responses at the measuring points, and the reconstructed
response is very sensitive to error in the identified loads.
4. The identified loads from the case with 15 vibration modes (m=3, n =3) is
always smaller in all the results shown in Figures 8.98.11. The reason is due to
an unequal number of modes used in the responses and in the identification, and
it will be discussed in next section.
8.5.2.3 Ef f ect of Unequal Number of Modes i n t he Response and
i n t he I dent i f i cat i on
Figures 8.98.11 show that the identified loads from (m=3 and n =3) is less than
the loads identified from (m=3 and n =2). This difference cannot be the result of
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

150
100
50
50
0
N
0 1 2 3
The front axle load
s
s
4 5 6
300
200
100
100
0
N
0 1 2 3
The back axle load
4 5 6
Figure 8.10 Identified axle loads for different combinations of measured information ( static
loads; - - m=3, n =2(9); - - - m=3, n =3(15); m=3, n =2(15))
300
250
200
150
N
100
150
50
0
0 1 2 3
s
4 5 6
Figure 8.11 Identified total loads from using different modes ( static loads; - - m=3, n =2(9);
- - - m=3, n =3(15); m=3, n =2(15))
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

232 Movi ng Loads Dynami c Anal ysi s and I dent i f i cat i on Techni ques
Table 8.7 Correlation coefficient between reconstructed and measured strains at
3/8a
Eccentricity Modes Beam 1 Beam 2 Beam 3 Beam 4 Beam 5
0 m=3; n =4(15) 0.783 0.897 0.931 0.909 0.799
(Rail 3) m=3; n =3(15) 0.935 0.944 0.931 0.951 0.941
m=3; n =2(15) 0.901 0.935 0.929 0.944 0.922
m=3; n =2(9) 0.932 0.949 0.933 0.953 0.947
3/8b m=3; n =4(15) 0.112 0.772 0.897 0.939 0.936
(Rail 1) m=3; n =3(15) 0.166 0.793 0.915 0.951 0.948
m=3; n =2(15) 0.039 0.813 0.914 0.948 0.947
m=3; n =2(9) 0.044 0.692 0.839 0.849 0.837
1/8b m=3; n =4(15) 0.550 0.794 0.848 0.790 0.758
(Rail 2) m=3; n =3(15) 0.859 0.937 0.945 0.949 0.974
m=3; n =2(15) 0.896 0.947 0.953 0.948 0.966
m=3; n =2(9) 0.880 0.922 0.920 0.929 0.945
Note: (15) denotes 15 measuring points located evenly on the five beams; (9) denotes nine
measuring points located evenly on the three beams near the moving path of the car.
any calibration error. An inspection of Equations (8.38) to (8.41) gives the following
reasons for the existence of this error.
Equation (8.38) is valid for both the measured responses and for the identification.
Let N
R
=m
R
n
R
and N
I
=m
I
n
I
be the number of the modes in the responses and
in identification respectively. Equation (8.38) can be rewritten as follows:
w
ns
= W
N
s
N
R
Q
N
R
(8.47)
where
W
N
s
N
R
=
_
_
_
_
_
W
1
(x
1
, y
1
) W
2
(x
1
, y
1
) W
N
R
(x
1
, y
1
)
W
1
(x
2
, y
2
) W
2
(x
2
, y
2
) W
N
R
(x
2
, y
2
)
.
.
.
.
.
.
.
.
.
W
1
(x
N
s
, y
N
s
) W
2
(x
N
s
, y
N
s
) W
N
R
(x
N
s
, y
N
s
)
_

_
N
s
N
R
Q
N
R
= {q
1
(t), q
2
(t), . . . , q
N
R
(t)}
T
We have two possible cases:
Case (a): When N
I
<N
R
, Q
N
I
can be obtained from partitions of Equation (8.47) as
Q
N
I
= (W
T
N
S
N
I
W
N
S
N
I
)
1
W
T
N
S
N
I
(w
ns
W
N
S
N
R
N
I
Q
N
R
N
I
) (8.48)
with
W
N
s
N
R
N
I
=
_
_
_
_
_
_
_
W
N
I
+1
(x
1
, y
1
) W
N
I
+2
(x
1
, y
1
) W
N
R
(x
1
, y
1
)
W
N
I
+1
(x
2
, y
2
) W
N
I
+2
(x
2
, y
2
) W
N
R
(x
2
, y
2
)
.
.
.
.
.
.
.
.
.
W
N
I
+1
(x
N
s
, y
N
s
) W
N
I
+2
(x
N
s
, y
N
s
) W
N
R
(x
N
s
, y
N
s
)
_

_
N
s
(N
R
N
I
)
Q
N
R
N
I
= {q
N
I
+1
(t), q
N
I
+2
(t), . . . , q
N
R
(t)}
T
(8.49)
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4

Movi ng Force I dent i f i cat i on 233
The terms in last bracket in Equation (8.48) represents the responses from the lower
sets of N
I
modes in identification. But in practice, the total measured responses are
used instead, leading to an over-estimation of the forces when substituting Q
N
I
and its
derivatives into Equation (3.17) in Chapter 3.
Case (b): When N
R
<N
I
, Q
N
I
can be obtained in a similar way as:
Q
N
I
= (W
T
N
S
N
I
W
N
S
N
I
)
1
W
T
N
S
N
I
(w
ns
+W
N
S
N
R
+N
I
Q
N
R
+N
I
) (8.50)
where W
N
S
N
R
+N
I
and Q
N
R
+N
I
are similarly defined as in Equation (8.49). The last
term in bracket in Equation (8.50) represents the total responses corresponding to the
modes used in identification. But, in practice, only the measured responses w
ns
is used
in the equation leading to under-estimation in the forces.
In the simulation studies, the measured responses are computed from 25 modes
(m=n =5). In the cases with N
I
<N
R
, the modes using in the identification covered
the excitation frequency range of the moving loads. 200 terms in the orthogonal poly-
nomial have been used in obtaining the derivatives of Q
N
I
. The small magnitudes of
modes higher than N
I
are further reduced by the low pass filtering effect. The term
Q
N
I
N
R
becomes insignificant small and the over-estimation from Equation (8.48) is
therefore not noticeable.
In the experiment, the excitation or natural frequencies of the car are not small. But
only 50 terms in the orthogonal polynomial have been used because of a very small
signal to noise ratio in the higher mode responses. The number of vibration modes
left after filtering is greatly reduced leading to N
R
<N
I
. The term W
N
S
N
R
+N
I
Q
N
R
+N
I
does not exist in the measured responses, and hence the modal coordinates Q
N
I
and its
derivatives are under-estimated giving smaller than true loads when substituting into
Equation (3.17) in Chapter 3. There is a difference of three modes between the cases of
(m=3, n =3) and (m=3, n =2), and yet the final results differs by a great percentage.
This is because of the large responses in these three modes which should contribute
greatly to the final identified results if they are included in the identification.
8.6 Summary
The moving load identification with the generalized orthogonal functions has been
presented in this chapter. A generalized orthogonal function approach is proposed
to obtain the modal velocity and acceleration from measured strain response. This
reduces the error due to measurement noise, and the moving forces are identified
with bounds in the errors using regularization method in the solution. The moving
force identification method is illustrated with numerical simulations and experimental
studies of a beam and plate structures. The significance of having enough structural
vibration modes in the identification is shown with error studies in the structures.
D
o
w
n
l
o
a
d
e
d

b
y

[
I
S
T
A
N
B
U
L

T
E
K
N
I
K

U
N
I
V
E
R
S
I
T
E
S
I
]

a
t

0
6
:
5
2

1
2

J
a
n
u
a
r
y

2
0
1
4