Mechanical Systems and Signal Processing 41 (2013) 254267

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/ymssp

Sensor placement methods for an improved force

identification in state space
J. Wang a, S.S. Law b,n, Q.S. Yang a
a

School of Civil Engineering, Beijing JiaoTong University, Beijing, Peoples Republic of China
Civil and Environmental Engineering Department, Hong Kong Polytechnic University, Hunghom, Kowloon, Hong Kong,
Peoples Republic of China
b

a r t i c l e in f o

abstract

Article history:
Received 19 June 2012
Accepted 10 July 2013
Available online 7 August 2013

The accuracy and effectiveness of force identification in time domain based on state space
can be influenced by the conditioning of the structural system Markov parameter matrix.
Two different sensor placement methods based on the conditioning analysis of the
Markov parameter matrix for improving the identification of input force are presented in
this paper. The first one is based on direct computation of the condition number of the
matrix, and it would involve computation for many different combinations of candidate
sensor locations. It would be time consuming particularly when a large number of
candidate combinations of sensor locations is considered. The second approach is based
on the correlation analysis of the system Markov parameter matrix. A sensor correlation
matrix is defined and the correlation criterion, which can indicate the ill-conditioning of
the Markov parameter matrix, is introduced. The performances of these two methods are
compared in numerical simulations with respect to their efficiency and accuracy. It is
concluded that the performance of both methods is similar when the number of candidate
combination of sensors is small. However, when there are many candidate combinations
of sensor locations, the method based on correlation analysis of the Markov parameter
matrix performs better with consistently good sensor placement for force identification
and much less computation effort.
& 2013 Elsevier Ltd. All rights reserved.

Keywords:
Force identification
Sensor placement
Markov parameter matrix
Condition number
Correlation analysis
Ill-posed problem

1. Introduction
Knowledge of the external force on a structural system is of great importance in many structural dynamic problems.
However, its direct measurement with transducer is a difficult task when the locations cannot be accessed, while vibration
responses can be conveniently measured. This partly explained why indirect methods are often preferred to direct
measurement, in which forces can be identified based on the inverse structural analysis. Various indirect force identification
methods have been proposed in recent years. Techniques involved are either in the frequency domain or in the time domain.
Fourier Transform is often used in the frequency domain methods to establish the relationship between the force and
response via the frequency response function [1,2]. However, it is well known that the identification process requires the
inversion of the transfer function, and this suffers from inherent instability caused by the ill-posed inverse problem.

Corresponding author. Tel.: +852 27666062; fax: +852 2334 6389.

E-mail addresses: 05121303@bjtu.edu.cn (J. Wang), cesslaw@polyu.edu.hk (S.S. Law), qshyang@bjtu.edu.cn (Q.S. Yang).

0888-3270/$ - see front matter & 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.ymssp.2013.07.004

J. Wang et al. / Mechanical Systems and Signal Processing 41 (2013) 254267

255

In particular, measurements containing noise would have an adverse combination effect with the ill-posed problem leading
to computation instability.
The usual approach to the solution of an ill-posed problem is to transform it into a well-posed problem by using
additional information on the solution sought. The methods such as singular value decomposition (SVD) and Tikhonov
regularization method [35] have been used to solve the ill-posed problem. There were also developments in time domain
force identification. Kammer [6] presented a time domain method for estimating the discrete input force based on the
measured response and the structural system Markov parameters. The inverse system Markov parameters can be computed
from the forward Markov parameters using a linear prediction algorithm. Mao et al. [7] proposed a time domain force
identification approach for a linear system based on the Markov parameters with precise computation and regularization
technique. Law and Ding [8] developed a time domain force identification method for a substructure. The interface forces as
well as the external forces acting on the substructure can be identified through the Tikhonov regularization method in
state space.
The accuracy of the inverse analysis may vary significantly with different spatial location of the response measurements,
and several approaches have been developed for choosing the sensor locations. Blau [9] suggested a frequency dependent
criterion for locating sensors in the force spectra identification. Thite and Thompson [10] commented that this technique is
difficult to apply to the whole frequency range of interest, and they proposed a method to select the sensor locations to
improve the inverse force determination based on a composite condition number. The minimum composite condition
number of the transfer function matrix associates with a reduction in the ill-conditioning of the transfer function as well as
in the error of the reconstructed forces.
There are also other methods on the selection of sensor locations in modal testing and condition monitoring of
structures. Kammer [11] proposed an Effective Independence (EFI) method to quantify the contributions of response
measurements so that the modal states of targeted modes can be optimally observed. Lim [12] employed the generalized
Hankel matrix, a function of the system controllability and observability, to develop an approach which can determine
sensor locations based on a given rank for the system observability matrix while satisfying modal test constraints. Hemez
and Farhat [13] proposed an Energy Matrix Rank Optimization method which is to maximize the strain energy information
in instrumented structural members. This method requires both the target mode shapes for the candidate sensors as well as
the structural stiffness. Heo et al. [14] derived a Kinetic Energy Optimization Technique (EOT) with the formulation similar
to EFI, and the difference lies in the quantity that is optimized. The EFI method maximizes the Fishers information matrix
while the EOT optimizes the kinetic energy matrix. A limitation of the EFI method is that the sensor locations with low
energy content may be selected with a consequent possible loss of information. The EFI-DPR (Driving Point Residue) method
eliminates this problem by multiplying the candidate sensor contribution of the EFI method with the corresponding DPR
coefficient [15]. Papadimitriou et al. [16] introduced the information entropy norm as a measure that best corresponds to the
objective of structural testing which is to minimize the uncertainty in the model parameter estimates. The spatial
correlation is important to avoid redundant information provided by neighboring sensors with distance less than the
characteristic length of the highest contributing mode. Valuable insight was provided into the effect of spatial correlation of
the prediction error on the optimal placement of sensors for modal identification or parameter estimation in finite element
model updating problems encountered in structural dynamics [17].
The accuracy and effectiveness of force identification in time domain based on state space can be influenced by the
conditioning of the structural system Markov parameter matrix which may vary significantly with measurement location. In
this paper, two different sensor placement methods based on the conditioning analysis of Markov parameter matrix for
improving the identification of input force are presented. The first one is based on direct computation of the condition
number of the Markov parameter matrix, and it may involve computation with many different combinations of candidate
sensor locations. It would be time consuming particularly when a large number of candidate combinations of sensor
locations is considered. The second approach is based on the correlation analysis of the Markov parameter matrix. The autocorrelation factor and cross-correlation factor are introduced in a sensor correlation matrix. A sensor correlation criterion,
which can be used as a measure to select the sensor locations, is defined. A set of responses can be found to associate with a
minimum sensor correlation criterion. Numerical simulations with a plane truss structure and a three-dimensional frame
structure are used to validate these methods and compare their performances with respect to their efficiency and accuracy
with noisy measurements.

2. Force identication algorithm in state space

2.1. State space equation of motion
For a general finite element model of a linear damped elastic structure, the equation of motion can be written as,
M x C x_ Kx f

where matrices M, C, and K are the mass, damping and stiffness matrices of the structural system, respectively. x , x_ and x are
vectors of acceleration, velocity and displacement of the structural system, respectively. f is the vector of external excitation
with matrix L mapping these forces to the associated degree-of-freedoms (DOFs) of the structure. Rayleigh damping

256

J. Wang et al. / Mechanical Systems and Signal Processing 41 (2013) 254267

expressed in the following form can be adopted as,

C a1 M a2 K

where a1 and a2 are constants to be determined from two modal damping ratios.
Eq. (1) can be expressed in the state space form as,
z_ AC z BC f

where
z




x
0
; AC
x_
M 1 K

The superscript
expressed as,

M

I
1

and BC

0
M

1

denotes that the matrices are for the continuous structural system. The output equation can be

y Ra x Rv x_ Rd x

mdof sdof

are the mapping matrices associate with the DOFs of the measured acceleration, velocity and
where Ra, Rv and Rd
displacement, respectively. mdof is the number of sensor. sdof is the total number of DOFs in the structure. Eq. (4) can be
rewritten as,
y Rz Df

where R Rd Ra M 1 K Rv Ra M 1 C and D Ra M 1 .

Eqs. (3) and (5) can be converted into discrete equations using the exponential matrix into the final discrete forms as,
zj 1 AD zj BD f j

yj Rzj Df j j 1; 2; ; nt
D

C 1

7
D

denotes that the matrices are for the discrete structural

where A expA dt and B A A IB . Superscript
system. dt is the time interval between the state variables zj 1 and zj. nt is the total number of sampling points. This is
formulated with the Zeroth-Order-Hold (ZOH) discrete method assuming the external force f is constant in a sampling
time step.
The output yj can be solved from Eqs. (6) and (7) with zero initial conditions of responses in terms of the previous input
f k; k 0; 1; ; j as,
j

yj H k f jk

k0

D
where H 0 D and H k RAD
k1 B which are the system Markov parameters.
Eq. (8) can be rewritten as,

Y HF
where

2
3
y0
H0
6
6
7
6 H1
6 y1
7
7
;H 6
Y 6
6
6
7
4
4
5
ynt1
Hnt1
mdof nt1
2
3
f 0
6
7
f
6 1
7
7
F 6
6
7
4
5
f nt1
2

H0

H nt2

7
0 7
7
0 7
5
H0

mdof ntf dof nt

f dof nt1

2.2. Regularized solution

The Markov parameter matrix H is a lower-block triangular Toeplitz matrix and it consists of nt2 matrix blocks, each of
dimension mdof  f dof , where fdof is the number of DOFs with external forces. The requirement of mdof f dof is often
found in the time domain method to make sure that the set of equations is over-determined. The ordinary least-squares
solution of Eq. (9) corresponds to an algebraic minimization problem in the following form,
minHF LS Y22

10

J. Wang et al. / Mechanical Systems and Signal Processing 41 (2013) 254267

257

and the resulting solution has the form of,

F LS H Y

11

where H denotes the pseudo-inverse of matrix H. However, like many other inverse problems, the force identification
problem is ill-posed. Several numerical tools known as regularization methods have been developed to deal with this
computation problem. The two most general regularization methods are SVD technique and Tikhonov regularization
method. In the Tikhonov regularization method, a priori bound of F reg 22 is added to the least-squares residue norm, and the
new cost function is defined as [18],
minfjjHF reg Yjj22 jjF reg jj22 g

12

The minimization of Eq. (12) results in the regularized solution as,

F reg H T H 2 I1 H T Y

13

To select an optimum regularization parameter , a popular and convenient method, called the L-curve method, can be
applied. The L-curve method involves a plot of the norm F reg 22 versus the corresponding HF reg Y22 in a log-log scale. The
L-curve plot has an L-shaped appearance with a distinct corner separating the vertical and the horizontal parts of the curve
for most problems. The corner of this curve represents a compromise between these two norms at which the optimum
regularization parameter occurs. Hansen [19] proposed using the point of maximum curvature to determine the optimum
regularization parameter.
Based on the principle of the SVD technique, matrix H in Eq. (9) can be decomposed in the form,
H UV

f dof nt

i1

ui si vTi

14

where URmdof ntmdof nt and VRf dof ntf dof nt are left and right singular vector matrices of matrix H, and they are
orthogonal matrices satisfying U T U I mdof nt and V T V I f dof nt and matrix diags1 sf dof nt has non-negative
diagonal elements appearing in non-increasing order such that s1 sf dof nt 0. The elements si are the singular values of
matrix H.
The regularized solution of Eq. (13) can be written as,
F reg

f dof nt

i1

fi

uTi Y
v
si i

15

where f i s2i =s2i 2 i 1; 2; f dof  nt are referred to as filter factors. It is noted that the Tikhonov regularization
makes use of the filter factors to damp the effects associated with small singular values.
The above described methods are used to solve for the force vectors in the inverse identification problem.

3. Method IBased on the condition number of Markov parameter matrix

Lee and Park [20] analyzed that the conditioning of the system FRF matrix can be improved by a proper selection of
measurement positions. Zheng et al. [21] presented another way to optimize the selection of the sensor locations. The
coherence of transfer function matrix is analyzed and the coherence factor of the matrix is introduced for an optimized
response location with a much less computational effort.
Similar to frequency domain methods, the accuracy and effectiveness of force identification in time domain based on
state space can be influenced by the conditioning of the system Markov parameter matrix as noted in Eq. (9). Condition
number can be a measure of ill-conditioning of the matrix. Here, the condition number is the 2-norm condition number
which is the ratio between the largest and smallest singular values of matrix H. Detailed procedure of the sensor placement
method based on the condition number of Markov parameter matrix is as follows,
(a)
(b)
(c)
(d)

Determine the number of sensor mdof and the number of candidate sensor location cdof (cdof sdof ).
Compute all possible combinations of mdof responses taken from cdof, which is C mdof
.
cdof
Calculate the system Markov parameter matrix H as in Eq. (9) for all the above combinations.
The combination giving the minimum condition number is taken as the optimal combination of sensor locations.

The condition number of system Markov parameter matrix can be used as a measure for selecting the sensor locations to
reduce the ill-posedness of the force identification. However, the process would involve the estimation of condition number
of all possible sensor combinations. This task becomes impractical with a large number of candidate combinations of sensor
locations.

258

J. Wang et al. / Mechanical Systems and Signal Processing 41 (2013) 254267

4. Method IIBased on the correlation analysis of Markov parameter matrix

Fabunmi [22] commented that a large condition number of a modal matrix may be a consequence of strong modal
behaviour of the structure, where one or a small number of modes dominates, and the responses are not independent. It
is noted that the independence of sensor responses can reflect the conditioning of matrix H, and independent sensor
responses can result in a well-posed matrix H. In the following paragraphs, a new sensor placement method via
correlation analysis of the Markov parameter matrix is presented, which can be an alternative approach to Method I
described above.

4.1. Correlation analysis of the Markov parameter matrix

To analyze the correlation of sensors at each possible sensor location, matrix H with the size of cdof  nt  f dof  nt is
calculated. The original form of Markov parameter matrix H consists of nt2 matrix blocks, each of dimension cdof  f dof .
Matrix H is re-assembled into cdof blocks, each of dimensions nt  f dof  nt as,
h
^1
H H

^2
H

^p
H

^ cdof
H

iT
cdof ntf dof nt

16

For the pth candidate sensor location, we have,

h
^ p Hp;1
H

H p;2

H p;nt

iT

; p 1; 2; ; cdof

17

^ p consists of nt blocks each of dimension 1  f dof  nt, and matrices H and H

^ p are shown in Fig. 1.
Matrix H
^ in Eq. (16) is defined. Different responses are
When response from one sensor is selected, one row sub-matrix H
described by different row sub-matrices of H. The correlation between different responses and the correlation between
^ p p 1; 2; ; cdof , respectively. These two
different time instants of one response could be analyzed with matrices H and H
kinds of correlation can influence the condition of matrix H.
The sensor correlation matrix can be given by,
2

r 1;1

6r
6 2;1
6
6
6
R6
6 r p;1
6
6
4
r cdof ;1

r 1;2

r 1;p

r 1;cdof

r 2;2

r 2;p

r 2;cdof

r p;2

r p;p

r p;cdof

r cdof ;2

r cdof ;p

r cdof ;cdof

3
7
7
7
7
7
7
7
7
7
5

18

cdof cdof

Fig. 1. Three different kinds of correlations.

J. Wang et al. / Mechanical Systems and Signal Processing 41 (2013) 254267

259

^ We define
which is positive and symmetric. The elements in matrix R denote the correlation between two sub-matrices H.
the non-diagonal elements
0

p 1; 2; ; cdof

B
C
r p;q @ q 1; 2; ; cdof A
pq
as cross-correlation factors and the diagonal element r p;p as auto-correlation factors. r p;q indicates the correlation between
the pth and qth sensor locations and r p;p indicates the correlation between different time instants of the pth response. Fig. 1
shows the three different kinds of correlations.
To calculate r p;q and r p;p , the cross-correlation matrix for two sensors and the auto-correlation matrix are defined. The
cross-correlation matrix is defined as,
2

e1;1

6
6 e1;2
Ep; q 6
6
4
e1;nt

e1;2

e2;2

e2;nt

e1;nt

7
e2;nt 7
7
7

5
ent;nt

19
ntnt

^ q , e.g., the
^ p and H
The elements in the cross-correlation matrix give the correlations between different sub-matrices in H
^ p and H q;2 in H
^ q.
element e1;2 is the correlation between H p;1 in H
The form of auto-correlation matrix is similar to that of the cross-correlation matrix. The only difference is that the
diagonal elements of auto-correlation matrix are set equal to zero instead of the usual unity. This is because the elements of
^ p . The correlation between one subauto-correlation matrix indicate the correlation between different sub-matrices in H
^ p has no meaning to serve as reference on the conditioning of matrix H. Therefore the diagonal
matrix H and itself in H
elements are not considered and set equal to zero. The auto-correlation matrix can be written as,
2

6
6 s1;2
Sp 6
6
4
s1;nt

s1;2

s2;nt

s1;nt

7
s2;nt 7
7
7
5
0

20
ntnt

Since there are cdof sensor locations, there would be cof !=cdof mdof ! number of Ep; q matrices and cdof number of
Sp matrices. One cross-correlation factor r p;q associates with one Ep; q matrix and the value of r p;q equals to the maximum
element in the corresponding Ep; q matrix. Similarly, the auto-correlation factor r p;p equals to the maximum element in the
corresponding Sp matrix.
(

r p;q maxfel1 ;l2 p; qg l1 1; 2; ; nt

r p;p maxfsl1 ;l2 pg l2 1; 2; ; nt

21

The trend of variation of these correlation factors has been checked to be more distinct when compared to those defined
as the Fobenious norm of Ep; q and Sp matrices.
The element el1 ;l2 in the cross-correlation matrix and sl1 ;l2 in the auto-correlation matrix are calculated as,
8
T
jHp;l1 Hq;l j
>
>
2
q
>
> el1 ;l2 p; q q
>
T
T
>
jH p;l1 Hp;l j jH q;l2 H q;l j
<
1
2

22

T
>
jH p;l H p;l j
>
2
>
sl1 ;l2 p q1 q
>
>
T
T
>
:
jH p;l H j jH p;l H j
1

p;l1

p;l2

In practice, if the number of sensor, mdof, is determined, the total number of candidate combinations of sensor locations
can be determined as C mdof
. For the wth sensor location combination (w 1; 2; C mdof
), the corresponding sensor
cdof
cdof
correlation matrix Rmdof mdof can be extracted from matrix Rcdof cdof as a sub-matrix in Eq. (18). The sensor correlation
criterion, w, which will be used as a measure to select the combination of sensor locations, is defined as the Frobenius
norm (F-norm) of the sensor correlation matrix Rmdof mdof ,
w Rmdof mdof F

mdof mdof

r 2k1 k2

k1 1 k2 1

!1=2
w 1; 2; C mdof

cdof

23

A smaller w indicates more independency of the combination of sensor responses and the Markov parameter matrix would
be less ill-conditioned. Thus, an optimal or sub-optimal combination of sensor locations can be found from minimizing w.

260

J. Wang et al. / Mechanical Systems and Signal Processing 41 (2013) 254267

4.2. Implementation procedure

Detailed procedure for selecting an optimal combination of sensor locations from the correlation analysis of the Markov
parameter matrix is shown below:
(a) Determine the number of sensor, mdof , and the number of candidate sensor location, cdof.
(b) Calculate the system Markov parameter matrix H with the size of cdof  nt  f dof  nt. Re-assemble the matrix H
into cdof blocks, each of dimension nt  f dof  nt.
(c) Calculate the elements el1 ;l2 p; q of the cross-correlation matrix Ep; q and the elements sl1 ;l2 of the auto-correlation
matrix Sp from Eq. (22). The cross-correlation factor, r p;q , and the auto-correlation, r p;p , can be calculated from Eq. (21).
(d) Assemble the cross-correlation factors and the auto-correlation factors into the correlation matrix, Rcdof cdof , according
to Eq. (18).
(e) The total number of candidate combinations of sensor locations, C mdof
, is computed. The sensor correlation matrix,
cdof
Rmdof mdof , corresponding to each candidate combination can be extracted from matrix Rcdof cdof .
(f) Calculate the sensor correlation criterion, w; w 1; 2; C mdof
from Eq. (23) for each candidate combination. The
cdof
combination that corresponds to the minimum w is taken as the optimal combination of sensor locations.
5. Numerical simulations
5.1. Example IA plane truss structure
Numerical studies on a plane steel truss structure as shown in Fig. 2 are used to illustrate the performances of the
proposed approaches. The finite element model of the structure consists of 9 nodes and 15 elements. Each node has two
DOFs. The structure is on hinge supports at Nodes 1 and 5. Rayleigh damping is assumed and the damping ratios for the first
two modes are both taken as 0:05. The elastic modulus of material is 200 GPa and the material density is 7:8  103 kg/m3.
The length of each bar is 2.0 m and the area of the member cross-section is 0.002 m2.
Two multiple sine wave external excitation forces are applied along the horizontal and vertical directions at Node 9.
These two forces are,
F 1 t 80 sin 50t 0:5 50 sin 80t 10 sin 120t 0:6 N
F 2 t 60 sin 55t 40 sin 70t 1:4 15 sin 100t N

24

Acceleration responses in the horizontal and vertical directions at Nodes 2, 3, 4, 6, 7 and 8 with cdof 12 are taken as the
measured responses. The number of sensor should be equal or bigger than the number of external force [7], and the cases
with the number of sensor mdof 2; 3; 4; 5; 6 are studied. The total sampling duration is 0.5 s, and the sampling rate is
500 Hz. White noise is added to simulate the polluted measured responses as,
x m x EP Nnoise sx

25

where x m denote the polluted measured responses and x denotes the responses without noise. Ep is the percentage noise
level, Nnoise is a standard normal distribution vector with zero mean and unit standard deviation, sx is the standard
deviation of the calculated acceleration response. In this simulation, 5% noise level is considered.
The relative error of force identification is defined as,
Relative error

F id F real
 100%
F real

26

where F id is the vector of identified input forces and F real denotes the vector of real input forces.
5.1.1. Comparison of computation effort
To evaluate the accuracy and computational cost of the two methods described above, both methods are implemented in the
MATLAB environment with a computer that has a 3.07 GHz Intel(R) Core(TM)i7 CPU and 12.0 GB (11.0 GB is available) of RAM.

F2
9

F1
(1)
1

(4)

(5)

(3)

(8)
(7)

(9)

(12)
(11)

6
(13)

60o
(2)

(6)

(10)

(14)

(15)
5

Fig. 2. Finite element model of the plane steel truss structure, () denotes the element number.

J. Wang et al. / Mechanical Systems and Signal Processing 41 (2013) 254267

261

Table 1
Comparison of the computation requirement for the plane truss structure.
Number of sensors

Number of candidate combination of sensor locations

Computing time (s)

Condition number (CN) method

Correlation analysis (CA) method

C 212 66

65.14

124.09

C 312 220

268.40

126.26

C 412 495

693.17

124.66

C 512 792

1330.65

126.96

C 612 924

1653.96

126.74

Table 2
Location and condition number of sensor location combinations.
Type

Method

Number of sensor

Location

Best combinations

CN

2
3
4
5
6
2

N2(x),
N2(x),
N2(x),
N2(x),
N2(x),
N2(x),

N8(y)
N8(x),
N6(x),
N3(x),
N3(x),
N8(y)

N8(y)
N8(x), N8(y)
N6(y), N8(x), N8(y)
N4(y), N6(y), N8(x), N8(y)

1.35  1010
326.10
329.74
337.48
345.91

3
4
5
6
2
3
4
5
6
2
3
4
5
6

N2(x),
N2(x),
N2(x),
N2(x),
N7(y),
N6(y),
N4(y),
N3(x),
N2(y),
N3(x),
N3(x),
N3(y),
N2(y),
N2(y),

N8(x), N8(y)
N2(y), N8(x), N8(y)
N2(y), N3(y), N8(x), N8(y)
N2(y), N3(y),N4(x),N8(x), N8(y)
N8(y)
N7(y), N8(y)
N6(y), N7(y), N8(y)
N4(x), N6(x), N7(x), N8(x)
N3(y), N4(y), N6(y), N7(y), N8(y)
N4(x)
N4(x), N7(x)
N4(y), N6(y), N7(y)
N3(y), N4(y), N6(y), N7(y)
N3(y), N4(y), N6(y), N7(y), N8(y)

1:35  1010
326.10
389.76
536.90
563.18
1
1
1
1
1
1
1
1
1
1

CA

Worst combinations

CN

CA

Condition number

Table 1 shows the computation time required by the two methods. It is noted that when only two sensors are selected for
the force identification with C 212 66, the computing time of the condition number (CN) method is shorter than that of the
correlation analysis (CA) method. The CA method does not have large change in the computing time with an increase of
number of sensors, while the computing time of the CN method becomes much greater with a large increase of candidate
combination of sensor locations. When the number of required sensor is 6, there are 924 sensor combinations, and the
computing time of the CN method is more than 10 times larger than that of the CA method.
5.1.2. Comparison of the best and worst sensor combinations
The best and worst sensor location combinations selected by the two methods with different number of sensors are
compared. Table 2 lists the sensor location and condition number of the System Markov Parameter (SMP) matrix from these
selected sensor location combinations. N2(x) and N8(y) denote the horizontal direction of Node 2 and vertical direction of
Node 8, respectively. It is noted that when the number of sensor is less than or equal to 3, the best sensor location
combinations selected by the two methods are the same. The condition number increases slightly with an increase of
number of sensors for both methods. When the number of sensor is 4 or larger, the condition number from the best sensor
location combination selected by the CN method is slightly smaller than that of the CA method. The minimum correlation
criteria w in CA method does not correspond to the smallest condition number, but it is close to it in general. The
condition numbers of the SMP matrix from the worst sensor location combinations selected by both methods are all equal
to infinity for different number of sensors.
5.1.3. Comparison of condition number of the SMP matrix
To further compare the performances of the two methods, the plot of condition number, in Log-scale, versus the order of
the selected sensor location combination for 2 required sensors is shown in Fig. 3. There are 66 sensor combinations in this
case. The condition number of the SMP matrix jumps to infinity from the 28th selected sensor location combination, and
therefore, only the first 27 combinations are studied. Fig. 3 shows that the first three selected sensor location combinations

262

J. Wang et al. / Mechanical Systems and Signal Processing 41 (2013) 254267

Fig. 3. Condition number versus the order of the selected sensor location combination with 2 required sensors.

Fig. 4. Condition number versus the order of the sensor location combination with 3 required sensors from the CA method.

are the same for both methods. The overall trends of the two curves are the same although there are some fluctuations in
the curve from the CA method.
Fig. 4 shows the condition number (in Log-scale) versus the order of the sensor location combination for 3 required
sensors from the CA method. The plot is noted to consist of plenty of fluctuations along the curve. The first 20 combinations
correspond to small condition numbers. The condition number increases obviously from around 120th combination with
large variations, and it jumps to infinity from around the 170th combination. This pattern of variation is also observed in the
curves for 4, 5 and 6 required sensors.
Since the number of sensor combination increases greatly with an increase in the number of required sensor, a
comparison of the trend of variation of the above plot is made by plotting the mean value of the condition number for a
group of sensor combinations, which is taken to be 10 for the case with 3 sensors, 15 for the case with 4 sensors, 25 for the
case with 5 sensors and 30 for the case with 6 sensors as shown in Fig. 5(a)(d). It can be seen that the overall trends of the
curves from both methods are consistent with different number of sensors.
5.1.4. Comparison of error of force identification
Since the accuracy of force identification can also be affected by noise, 100 simulation tests are carried out for comparison
of the error of force identification derived from both methods. Table 3 lists the errors of force identification from the best
and worst sensor location combinations with different number of required sensors. When the number of required sensor
is less than or equal to 3, the best sensor location combinations selected by the two methods are identical, and the errors
are the same. In the case of 2 required sensors from the worst sensor location combinations, the standard deviation and
maximum value are too big to be acceptable. This is because the number of sensor is equal to the number of unknown force.
The inverse problem in Eq. (9) can be solved mathematically but with serious ill-conditioning with measurement noise. To
improve the force identification accuracy, the number of sensor should be larger than the number of unknown force in
practice.
From the results with 4 or more sensors, it is noted in general that the error of identification from the best sensor
location combinations decreases slightly with an increase in the number of required sensor for both methods. In all cases,
the error of identification from the CA method is slightly smaller than that from the CN method which has a smaller
condition number in the SMP matrix (Table 2). This would suggest that the force identification is influenced not only by the
conditioning of the SMP matrix but also by the measurement noise. The influences of noise effects can be reduced when the
correlation between the measured responses is smaller. For almost all the cases, the standard deviation and maximum value
of the error of identification from the worst sensor location combinations from both methods are too big to be acceptable
in practice, although in some cases, the mean errors are satisfactorily small.

J. Wang et al. / Mechanical Systems and Signal Processing 41 (2013) 254267

263

Fig. 5. Condition number versus the order of the selected sensor location combination with 3, 4, 5 and 6 required sensors.

Table 3
Errors of identified force with 5% noise level.
Type

Best combinations

Method Number of
sensor

CN

CA

Worst combinations CN

CA

2
3
4
5
6
2
3
4
5
6
2
3
4
5
6
2
3
4
5
6

Error for F1 (%)

Error for F2 (%)

Mean value Standard

deviation

Maximum
value

Mean value Standard

deviation

Maximum
value

16.64
7.27
8.21
2.37
8.21
2.19
8.87
2.15
8.74
2.35
16.64
7.27
8.21
2.37
8.09
2.16
7.37
1.83
6.29
1.65
8.09  1010 2.48  1010
14.59
4.71
21.71
9.03
18.05
8.83
25.79
13.74
14.31
8.96
13.04
6.35
16.69
6.60
14.96
6.19
25.79
13.74

39.43
17.11
13.62
17.19
15.59
39.43
17.11
13.77
13.96
10.97
2.48  1010
40.34
56.75
50.81
69.11
47.29
39.18
42.05
47.11
69.11

26.23
12.17
12.29
10.98
10.91
26.23
12.17
12.28
10.11
10.13
7.25  1010
10.35
10.37
6.06
6.98
83.53
22.92
10.96
5.90
6.98

71.72
25.98
20.72
21.68
19.24
71.72
25.98
20.62
19.39
17.14
2.48  1010
30.74
32.09
14.14
38.28
285.47
53.76
35.26
18.98
38.28

14.20
3.52
3.24
2.62
2.83
14.20
3.52
3.17
2.47
2.64
2.29  1010
4.77
5.32
2.15
12.60
53.29
9.94
5.94
2.87
12.60

The case with 4 sensors is further studied for comparison. Table 4 shows the error of identified forces from 100
simulations with the first, second and third best sensor location combinations. All the three sensor combinations give
similar error in terms of the mean, maximum value and the standard deviation in the study. This indicates that there would
be a large group of sub-optimal sensor combinations that performs similarly to the optimal combination in terms of the
error of force identification.

264

J. Wang et al. / Mechanical Systems and Signal Processing 41 (2013) 254267

Table 4
Errors of identified forces with the first three best sensor location combinations with 5% noise level.
Type

Method

Best combinations

CN

CA

Order

1
2
3
1
2
3

Error for F1 (%)

Error for F2 (%)

Mean value

Standard deviation

Maximum value

Mean value

Standard deviation

Maximum value

8.21
9.27
8.61
8.09
7.69
8.22

2.19
2.44
2.27
2.16
1.96
2.16

13.62
15.36
14.28
13.77
12.36
13.55

12.29
11.58
12.30
12.28
12.19
12.10

3.24
3.02
3.27
3.17
3.10
3.15

20.72
19.70
21.04
20.62
20.81
20.30

Fig. 6. The three-dimensional three-storey frame structure () denotes the element number.

It can also be concluded from Tables 3 and 4 for the force identification in time domain, the number of required sensor
should be larger than the number of unknown forces, and the force identification accuracy can be improved by increasing
the number of sensors in both methods. Both methods can provide consistently good sensor placement, and the correlation
analysis method involves much less computation efforts and with similar accuracy to the CN method when there is plenty of
candidate sensor locations.
5.2. Example IIA three-dimensional frame structure
A three-dimensional three-storey frame structure as shown in Fig. 6 is studied as another example to further illustrate
the performances of the proposed methods. The finite element model of the structure consists of 36 elements and 28 nodes
each with six DOFs. The frame is fixed to the ground at Nodes 1, 2, 3 and 4 with rigid supports. The elastic modulus of
material is 210 GPa, and the material density is 7:8  103 kg/m3. The area of the member cross-section is 0.04 m2. The
flexural moment of inertias of all members in the x- and y-directions are 1:33  104 m4 . Rayleigh damping is assumed and
the damping ratios for the first two modes are both taken as 0:05. The first five modal frequencies are 3.08, 3.67, 4.62,
10.61 and 11.45 Hz, respectively.
Two multiple sine wave external excitation forces are applied along the x-direction at Nodes 21 and 28, respectively, and
they are,
F 3 t 3500 sin 3t 0:5 2000 sin 8t N
F 4 t 3000 sin 4t 2000 sin 9t 1:4 N

27

The total number of candidate sensor locations is 24  3 72. Only a fraction of these locations are taken as accessible for
illustration of the sensor placement methods for a reduced computation. The translational DOFs in x-, y- and z-directions at
Nodes 22, 23, 24, 25, 26 and 27 are taken as the candidate sensor locations with cdof 18. Five sensors will be selected for

J. Wang et al. / Mechanical Systems and Signal Processing 41 (2013) 254267

265

Table 5
Location and condition number of the first three best and last three worst sensor location Combinations.
Type

Method

Order

Location

Condition number

Best combinations

CN

1
2
3
1
2
3
1
2
3
1
2
3

N22(x),
N22(x),
N22(x),
N22(x),
N22(x),
N22(x),
N22(x),
N25(z),
N22(x),
N25(z),
N22(x),
N26(x),

294.93
294.93
295.24
303.18
303.18
298.91
2.02  104
2.13  104
2.13  104
7.62  103
8.96  103
8.96  103

CA

Worst combinations

CN

CA

N22(y), N24(x), N27(x), N27(y)

N22(y), N25(x), N27(x), N27(y)
N22(y), N24(y) N27(x), N27(y)
N22(z), N27(x), N27(y), N27(z)
N22(y), N22(z), N27(x), N27(z)
N22(y), N23(z), N27(x), N27(z)
N22(z), N23(y), N24(z), N25(z)
N26(y), N26(z), N27(x), N27(z)
N22(z), N23(y), N23(z), N24(z)
N26(x), N26(y), N26(z), N27(z)
N22(z), N23(x), N23(y), N23(z)
N26(y), N26(z), N27(x), N27(z)

Fig. 7. Comparison of the identified forces from the best and worst sensor location combinations with 5% noise (CN method). (a) Force identification
results for F3 and (b) Force identification results for F4

Fig. 8. Comparison of the identified forces from the best and worst sensor location combinations with 5% noise (CA method).

identifying the two forces, that is mdof 5, and there are C 518 8568 candidate sensor combinations. The total sampling
duration is 10 s, while the sampling rate is 25 Hz. 5% noise level is considered in the measurements.
The computation effort required by the CN method is 116953.2 s which is more than 600 times greater than 182.27 s
required by the CA method. It should be noted that this difference would increase dramatically as the number of candidate
sensor combination increases.
Table 5 lists the location and condition number of the first three best and last three worst sensor location
combinations. The condition number from both the CN and CA methods are similar for the groups of worst and best
sensor location combinations due to a different configuration of the three-dimensional frame structure compared to that of
the plane truss structure. Those from the CN method are slightly smaller than those from the CA method for the group of

266

J. Wang et al. / Mechanical Systems and Signal Processing 41 (2013) 254267

Table 6
Errors of identified forces with 5% noise level for the three-dimensional frame structure.
Type

Best combinations

Method

CN

CA

Worst combinations

CN

CA

Order

1
2
3
1
2
3
1
2
3
1
2
3

Error for F3 (%)

Error for F4 (%)

Mean value

Standard deviation

Maximum value

Mean value

Standard deviation

Maximum value

14.36
14.10
14.24
11.54
11.70
13.61
22.46
128.31
23.14
82.73
20.62
64.82

3.40
3.34
3.42
2.73
2.81
3.27
13.87
76.98
14.52
36.57
12.94
25.64

25.44
24.96
25.06
19.47
19.90
23.84
66.35
367.81
69.03
202.77
60.91
145.80

11.40
11.76
11.40
9.51
12.03
11.82
131.01
22.59
121.31
21.94
81.70
38.98

2.73
2.81
2.77
2.27
2.94
2.87
67.81
15.26
43.22
13.65
38.26
23.65

20.07
20.68
19.88
15.71
20.65
20.62
308.29
69.66
236.27
63.19
193.18
111.48

best sensor location combinations, while it is just opposite with the group of worst sensor location combinations. This
observation is similar to those shown in Fig. 5(a)(d) for the plane truss structure.
Figs. 7 and 8 show the real and identified input forces from the best and worst sensor location combinations obtained
from the CN method and the CA method, respectively. The best sensor location combinations from both methods can
estimate the input forces with high accuracy even with 5% noise level in measurements, while the worst sensor location
combinations lead to identified input forces with large errors.
Table 6 lists the errors of identification from 100 simulations of the first three best and the last three worst sensor
location combinations selected by both methods. Similar to the results for the plane truss structure, the error of
identification from the best sensor locations combinations selected by the CA method is slightly smaller than that from
combinations selected by the CN method. The error of identification from the three worst sensor location combinations is
also too big to be acceptable.
6. Conclusions
The accuracy and performances of force identification in time domain based on state space can be influenced by the illposedness of the problem and measurement noise, and a combination of which would lead to computation instability of the
inverse problem. An improved condition of the inverse problem may help to improve the computation. Two different sensor
placement methods based on conditioning analysis of the system Markov parameter matrix are presented. The first method
is based on direct computation of the condition number of the matrix. Sensor location combination corresponding to the
minimum condition number can be considered as the optimal sensor placement. The second method is based on correlation
analysis of the system Markov parameter matrix. Sensor correlation criterion is used as a measure to select the sensor
locations. Numerical simulations with a plane truss structure and a three-dimensional frame structure are conducted to
study the performances of these methods. Results show that both methods can provide consistently good sensor
placements. If the sensor placement problem is small, either method can be adopted to yield satisfactory combinations
of sensor locations with acceptable accuracy and computing time. However, when there are many candidate sensor
combinations, the selection based on the correlation analysis has a great advantage with the computation efficiency and yet
with similar accuracy of identification. The selection may not always associate with the smallest condition number of the
system Markov parameter matrix, but it is close to it in general. Besides, the error of identification is in general slightly
smaller when the measured responses are more independent with a reduced effect from the measurement noise.

Acknowledgment
The work described in this paper was supported by the Niche Area Project Funding of the Hong Kong Polytechnic
University Project No. 1-BB6F. The authors would also like to thank the funding supported from National Science Foundation
of China (Grant No. 90815021).
References
[1] F.D. Barlett, W.G. Flannelly, Model verification of force determination for measuring vibratory loads, Journal of the American Helicopter Society 24 (2)
(1979) 1018.
[2] Y. Liu, W.S. Shepard, Dynamic force identification based on enhanced least squares and total least-squares schemes in the frequency domain, Journal of
Sound and Vibration 282 (1-2) (2005) 3760.
[3] E. Jacquelin, A. Bennani, P. Hamelin, Force reconstruction: analysis and regularization of a deconvolution problem, Journal of Sound and Vibration 265
(1) (2003) 81107.

J. Wang et al. / Mechanical Systems and Signal Processing 41 (2013) 254267

267

[4] A.N. Thite, D.J. Thompson, The qualification of structure-borne transmission paths by inverse methodsPart 1: Improved singular value rejection
methods, Journal of Sound and Vibration 264 (2) (2003) 411431.
[5] A.N. Thite, D.J. Thompson, The qualification of structure-borne transmission paths by inverse methodsPart 2: Use of regularization techniques,
Journal of Sound and Vibration 264 (2) (2003) 433451.
[6] D.C. Kammer, Effects of noise on sensor placement for on-orbit modal identification of large space structures, Journal of Dynamic Systems,
Measurement, and Control 114 (3) (1992) 436443.
[7] Y.M. Mao, X.L. Guo, Y. Zhao, A state space force identification method based on Markov parameters precise computation and regularization technique,
Journal of Sound and Vibration 329 (2010) 30083019.
[8] S.S. Law, Y. Ding, Substructure methods for structural condition assessment, Journal of Sound and Vibration 330 (15) (2011) 36063619.
[9] M. Blau, Force spectra identification by FRF matrix inversion: a sensor placement criterion, Journal of the Acoustical Society of America 105 (2) (1999)
970.
[10] A.N. Thite, D.J. Thompson, Selection of response measurement locations to improve inverse force determination, Applied Accoustics 67 (2006)
797818.
[11] D.C. Kammer, Sensor placement for on-orbit modal identification and correlation of large space structures, Journal of Guidance, Control, and Dynamics
14 (2) (1991) 251259.
[12] K.B. Lim, A Method for optimal actuator and sensor placement for large flexible structures, Journal of Guidance, Control, and Dynamics 19 (1) (1990)
13521360.
[13] F.M. Hemez, C. Farhat, An energy based optimum sensor placement criterion and its application to structural damage detection, in: Proceedings, 12th
International Modal Analysis Conference (IMAC), Society of Experimental Mechanics, Honolulu, 1994, 15681575.
[14] G. Heo, M.L. Wang, D. Satpathi, Optimal transducer placement for health monitoring of long span bridge, Soil Dynamics and Earthquake Engineering
16 (7-8) (1997) 495502.
[15] M. Meo, G. Zumpano, On the optimal sensor placement techniques for a bridge structure, Engineering Structures 27 (10) (2005) 14881497.
[16] C. Papadimitriou, J.L. Beck, S.K. Au, Entropy-based optimal sensor location for structural model updating, Journal of Vibration and Control 6 (2000)
781800.
[17] C. Papadimitriou, G. Lombaert., The effect of prediction error correlation on optimal sensor placement in structural dynamics, Mechanical Systems and
Signal Processing 28 (2012) 105127.
[18] A.N. Tikhonov, A. Goncharsky, V.V. Stepanov, A.G. Yagola, Numerical Methods for the Solution of Ill-Posed Problems, first ed. Kluwer Academic
Publishers, Boston, 1995.
[19] P.C. Hansen, Analysis of discrete ill-posed problems by means of the L-curve, SIAM Review 34 (1992) 561580.
[20] J.K. Lee, Y.S. Park, Response selection and dynamic damper application to improve the identification of multiple input forces of narrow frequency
band., Mechanical Systems and Signal Processing 8 (6) (1994) 649664.
[21] S.F. Zheng, L. Zhou, X.M. Lian, K.Q. Li, Technical note: coherence analysis of the transfer function for dynamic force identification, Mechanical Systems
and Signal Processing 25 (6) (2011) 22292240.
[22] J.A. Fabunmi, Effects of structural modes on vibratory force determination by pseudo-inverse technique, Journals: The American Institute of
Aeronautics and Astronautics 24 (3) (1986) 504509.