Computer-Aided Civil and Infrastructure Engineering 30 (2015) 666–681

INDUSTRIAL APPLICATION

Online Simultaneous Reconstruction of Wind Load

and Structural Responses—Theory and Application to
Canton Tower

Yan Niu, Claus-Peter Fritzen*, Henning Jung & Inka Buethe

Institute of Mechanics and Control Engineering–Mechatronics, University of Siegen, Paul-Bonatz-Strasse 9-11,
Siegen, Germany

Yi-Qing Ni & You-Wu Wang

Department of Civil and Environmental Engineering, The Hong Kong Polytechnic University, Hong Kong, China

Abstract: The actual wind load information is helpful in recorded during Tyhoon Kai-tak in 2012. The recon-
evaluating the health status of high-rise structures. How- struction results are validated by comparing the simul-
ever, as a type of distributed load, the wind load is very taneously reconstructed structural acceleration with the
difficult to be measured directly. A possible solution is to corresponding sensor measurements. The mean compo-
reconstruct it from the structural response measurements. nent of the loads and moments are calculated using the
This is often an ill-posed inverse problem. In this article, real-time wind speed measurements and the available
such ill posedness is solved by using a stable input estima- aerodynamic force coefficients. It is noted here that the
tor. With the help of the proposed application-oriented focus of this article is not to develop a totally new the-
algorithm selection guidance, a type of state and input ory, but rather to explore the application of a state and
estimator is formulated. This type of estimator is de- input estimator in the foreground to a practical complex
signed based on the Kalman filter scheme, and is capable structure.
of estimating the unknown inputs and the system states
within one sampling time. This actually facilitates the on- 1 INTRODUCTION
line simultaneous reconstruction of the wind load and the
structural responses. The 600 m tall Canton Tower is sit- For high-rise structures, for example, offshore wind tur-
uated in a typhoon active area, and a structural health bines and tall buildings, the external wind load infor-
monitoring system has already been integrated onto this mation is very important in structural design, structural
tower. These two points make the Canton Tower an ideal control, and structural health monitoring (SHM). The
test bed for validating the above illustrated online recon- wind load distributes in space, whereas the widely used
struction strategy for the wind load and the structural re- anemometer can only provide the wind speed and di-
sponses. An operational modal analysis (OMA) is first rection information at a specified height. To get the in-
performed to identify the modal properties of the Canton formation of such a type of distributed load, a potential
Tower under the Typhoon Nanmadol in 2011. Then a solution is to reconstruct it from the structural response
reduced-order finite element model (FEM) of the Can- measurements, that is, displacement, strain, velocity, or
ton Tower is updated according the OMA results. Fi- acceleration. However, this often leads to an ill-posed
nally, the equivalent fluctuating lateral loads and mo- inverse problem, in the sense that a small amount of
ments, which act on the nodes of the reduced-order FEM measurement noise may cause a large deviation in the
are reconstructed using the acceleration measurements reconstructed input force (Stevens, 1987).
∗ To
Law et al. (2005) identified the time-varying wind
whom correspondence should be addressed. Email: claus-peter.
fritzen@[Link].
load on a 50 m guyed mast in a simulation study.


C 2015 Computer-Aided Civil and Infrastructure Engineering.

DOI: 10.1111/mice.12134
 Online reconstruction of wind load and structural responses for Canton Tower 667

Lourens et al. (2008) applied the dynamic programming Tower an ideal test bed for the wind load reconstruc-
algorithm (Trujillo, 1978) in a wind load reconstruction tion study using field measurement data from the SHM
study for an 80 m high tower structure. These two stud- system. In this article, the field measurements obtained
ies both used the Tikhonov regularization in convert- during Typhoon Nanmadol in 2011 and Typhoon Kai-
ing the ill-posed inverse problem to a well-posed one. tak in 2012 are analyzed according to the flowchart in
To find the optimal Tikhonov regularization parameter, Figure 1, in which the Canton Tower is taken as an
both studies applied the L-curve method. This process example for the high-rise structure. Similar flowcharts
needs the complete structural responses and is not suit- could also be found, for example, in the load estima-
able for real-time application. tion study for wind turbines (Klinkov and Fritzen, 2010;
Online wind load reconstruction is a study that adopts Hernandez et al., 2011).
real-time executable unknown input estimation algo- Following the flowchart in Figure 1, this article
rithms to solve the above illustrated ill-posed inverse is organized as follows. Section 2 reviews how the
problem. These methods transform the task of solving state-space structural model in the modal coordinates
the ill posedness to designing a stable estimator. Once is constructed. Section 3 presents the operational
a stable estimator is obtained from the design stage, it modal analysis (OMA) results using the measurements
can stably provide the estimate of the input in real time. recorded during Typhoon Nanmadol. In Section 4,
Klinkov and Fritzen (2010) adopted a robust observer the available reduced-order finite element model
technique in monitoring the wind load for a 5 MW wind (FEM) of the Canton Tower is first modified, and then
turbine. Berg and Keith Miller (2010) used a Kalman updated. According to the authors’ knowledge on the
filter method for estimating the forces acting on wind available unknown input estimation algorithms, an
turbine blades. Hwang et al. (2009) proposed a type application-oriented algorithm selection guidance is
of estimator based on the Kalman filter scheme, with proposed. Several practical aspects on the available
the unknown input being calculated using the general- sensor types from the SHM system are discussed. Such
ized inverse. This algorithm was also further validated practical aspects and the algorithm selection guidance
using wind tunnel test data from a rectangular shaped help the authors to select the input and state estimation
concrete chimney (Hwang et al., 2011). Niu and Fritzen algorithm proposed in Hwang et al. (2009). In Section 6,
(2012) evaluated the Kalman Filter with Unknown In- the main steps of the selected algorithm are reviewed,
put (KF-UI) algorithm and the unbiased minimum- and a modification to the algorithm is proposed to make
variance input and state estimator (Gillijns and De it more suitable for the wind load reconstruction study.
Moor, 2007b) on a laboratory two-storey structure. It In Section 7, the reconstruction results are provided. As
was noticed that the KF-UI method could successfully a validation, the reconstructed structural responses are
reconstruct the wind load generated by an electric fan, compared with those directly measured from sensors.
whereas a low frequency drift appeared in the estimated Finally, conclusions are given in Section 8.
force from the unbiased minimum-variance input and
state estimator. Such low frequency drift was also men-
tioned in Lourens et al. (2012b). Adding a high-pass fil- 2 STATE-SPACE MODEL FOR STRUCTURAL
ter to the force estimate could be a solution (Lourens DYNAMICS
et al., 2012a). While all the above mentioned algorithms
were applied to linear systems, the robust observer tech- The forced vibration of a linear structure with viscous
nique is also applicable for input estimation on a type of damping can be represented by the following nonhomo-
nonlinear systems (Klinkov, 2011; Fritzen et al., 2013). geneous equation of motion,
In this article, the scope is limited to the case that
the structure is modeled as linear. It is noted here that M n q̈ n (t) + C n q̇ n (t) + K n q n (t) = f (t) (1)
all the aforementioned online wind load reconstruction
methods can provide the estimates of the input and the In Equation (1), q n (t) ∈ RnDOF denotes the displace-
state within one sampling interval. This indicates that ment vector in the nodal coordinates; f (t) ∈ RnDOF is
the wind load and structural responses could be recon- the force input vector; M n , C n , and K n are the mass,
structed simultaneously in real time. damping, and stiffness matrices with appropriate di-
The 600 m tall Canton Tower (formerly named mensions.
Guangzhou New TV Tower) is located in a typhoon ac- Assuming proportional damping, Equation (1) can be
tive area, and a long-term SHM system has been de- transformed into the canonical form,
signed and integrated onto this tower (Ni et al., 2009; q̈ m (t) + C m q̇ m (t) + K m q m (t) = Ψ T f (t) (2)
Ni et al., 2011). These two points make the Canton
668 Niu et al.

field measurement data

Finite element model
(FEM)

Previous
Operational modal analysis Stage 1:
(OMA) Model updating Structural model construction
and classification

State-space model
(in modal coordinates)
field measurement data

Real-time executable
input and state estimation
algorithm Stage 2:
Current

Online reconstruction of
wind load and
Online Online structural responses
wind load structural response
reconstruction reconstruction

Fig. 1. Online simultaneous wind load and structural response reconstruction methodology for high-rise structures.

by using 1. If mDOF = nDOF, any force input vector f (t)

will be in the range space of M n Ψ . In other words,
q n (t) = Ψ q m (t) (3) it can be represented as
In Equations (2) and (3), q m (t) ∈ RmDOF (mDOF ≤
nDOF), represents the displacement vector in the f (t) = M n Ψ d(t), (d(t) ∈ RmDOF ) (7)
modal coordinates; Ψ ∈ RnDOF×mDOF denotes the mass
normalized modal matrix, which fulfills the orthogonal- 2. If mDOF < nDOF, only the projection of f (t)
ity condition in the range space of M n Ψ can be represented
Ψ T M n Ψ = I mDOF (4) as Equation (7). In other words, the projection
of f (t) in the null space of M n Ψ will be filtered
C m and K m are the diagonal modal damping and out according to the orthogonality in Equation (4).
modal stiffness matrices with the following expressions, This indicates that the distributed force f (t) can
only be partly reconstructed, if only limited modes
⎡ ⎤ (mDOF < nDOF) are included in the structural
2ζ1 ω1 0 ··· 0 model.
⎢ 0 2ζ2 ω2 ··· 0 ⎥
⎢ ⎥
Cm = ⎢ . .. .. .. ⎥ (5)
⎣ .. . . . ⎦ Defining a state vector,
0 0 ··· 2ζmDOF ωmDOF 
q m (t)
x(t) = (8)
⎡ ⎤ q̇ m (t)
ω12 0 ··· 0
⎢ 0 ω22 ··· 0 ⎥ the structural model in Equation (2) can be represented
⎢ ⎥
Km = ⎢ .. .. .. .. ⎥ (6) in a state-space form,
⎣ . . . . ⎦
0 0 ··· ωmDOF
2 ẋ(t) = Ax(t) + Gd(t) (9a)

In Equations (4)–(6), the ζi and ωi (1 ≤ i ≤ mDOF), y(t) = C x(t) + H d(t) (9b)

are the damping ratios and the undamped circular nat-
ural frequencies. where y(t) ∈ R p is the output vector; A is the state tran-
Since the column vectors in Ψ are linearly indepen- sition matrix; G is the input matrix; C is the output ma-
dent and the square matrix M n has full rank, trix; and H is the direct feedthrough.
 Online reconstruction of wind load and structural responses for Canton Tower 669

Accelerometer
direction and Data
channel label acquisition
y
19 18
20 # 02, 04, 06, 09,
Section 8, z=446.80 m
17 12, 14, 16, 19
16 # 08, 10
Section 7, z=384.24 m # 10, 20
15
14 x
Section 6, z=332.15 m # 01, 03, 05, 07,
13
12 11, 13, 15, 17
Section 5, z=275.30 m
11
09 08
10 Section 4, z=228.50 m
07
06
Section 3, z=171.10 m
05
04
Section 2, z=119.30 m
03

02
z Section 1, z=30.63 m
y x 01
Synchronization Data transmission

(a) (b)

Fig. 2. (a) Positions of accelerometers and data acquisition system. (b) Measurement directions of accelerometers.

For the case that the output considers only acceler- 3 OPERATIONAL MODAL ANALYSIS
ation, the matrices A, G, C, and H appear as follows,
The OMA technique uses only the structural response
  measurements for the identification of modal character-
0 I 0 istics of structures in their operational conditions sub-
A= ,G = (10) ject to the ambient or natural excitation (Zhang et al.,
−K m −C m I
2005). The available OMA studies using the measure-
ments from the Canton Tower under the normal ambi-
C = C a Ψ −K m −C m , H = C a Ψ (11) ent excitation (from 18:00 on January 19, 2010 to 18:00
on January 20, 2010) can be found in, for example,
where C a ∈ R p×nDOF denotes the selection matrix for Faravelli et al. (2010), Niu et al. (2012), Loh et al. (2012),
the acceleration output. and Ye et al. (2012); and under some earlier typhoon
In this study, the damping ratios ζi (1 ≤ i ≤ mDOF), events may be found in Chen et al. (2011) and Guo
are identified from the OMA. The mass matrix M n , the et al. (2012). Furthermore, Liu et al. (2013) conducted
stiffness matrix K n , the mass normalized modal matrix an OMA study for the Canton Tower using the mea-
Ψ and the undamped circular natural frequencies ωi surements caused by a long distanced earthquake.
(1 ≤ i ≤ mDOF), are all from the updated FEM. Sec- It is noted here that the height of the Canton Tower
tions 3 and 4 will present the results from the OMA and was reduced from 610 m to 600 m after January 20,
the model updating for the Canton Tower. 2010 by adjusting the length of the mast. Besides this, a
670 Niu et al.

30 360
N

25 300
60%

Wind Direction [degrees]

45%
Wind Speed [m/s]

20 240
30%
10−min mean wind speed 15%
15 180 W E
10−min mean wind direction 10 − 11 m/s
9 − 10 m/s
10 120 8 − 9 m/s
7 − 8 m/s
6 − 7 m/s
5 − 6 m/s
5 60 4 − 5 m/s
3 − 4 m/s
2 − 3 m/s
S
0 0
00:00 04:00 08:00 12:00 16:00 20:00 24:00
Time [hh:mm]
(a) (b)

Fig. 3. Wind measurements during Typhoon Nanmadol. (a) 10-minute mean wind speed and direction. (b) Rose diagram (from
[Link] on August 31, 2011 to [Link] on September 1, 2011).

1.5
2
Normal excitation in 2010
Identified natural frequency [Hz]

Normal excitation in 2010

Typhoon "Nanmadol" in 2011
Identified damping ratio [%]

Typhoon "Nanmadol" in 2011

1.5
1

0.5
0.5

0
0 1 2 3 4 5 6 7 8 9 10 11 12
1 2 3 4 5 6 7 8 9 10 11 12 Mode No.
Mode No.
(a) (b)

Fig. 4. (a) Identified natural frequencies. (b) Identified damping ratios.

Ferris wheel was installed on the top of the main tower in Figure 3b, where 0 degree denotes the due north
and open to the public on September 1, 2011. The OMA direction.
studies presented in the previous publications all used The algorithm adopted in this article for the OMA
the measurements obtained before September 1, 2011. study is the vector autoregressive models method. More
So it is worth checking if there is any change in the dy- details on the algorithm may be found in Kraemer and
namic property after such structural modifications. Fritzen (2010). This method has the assumption that
In this article, an OMA study is performed using the field measurements should be stationary. To check
24 1-hour field measurement data sets recorded dur- the stationarity of the available measurements, a
ing Typhoon Nanmadol on August 31, 2011. The sam- multivariate outlier detection technique based on the
pling frequency is 50 Hz for both acceleration and Mahalanobis distance with the threshold provided
wind measurements. Figure 2 shows the positions and by the F-distribution (Rencher, 2002) is applied on
the measurement directions of the accelerometers. The the measurements, and the undesired data sets are
sampling frequency for the accelerometers is set as excluded from the analysis. The identified natural
50 Hz. The time history of the measured wind speed frequencies and damping ratios of the first 12 modes
and wind direction at the height of 461 m are plotted in are shown in Figure 4, and the identified first 6 bending
Figure 3a. The corresponding rose diagram is shown modes of the main tower are plotted in Figure 5.
 Online reconstruction of wind load and structural responses for Canton Tower 671

Mode 1: the 1st short-axis bending mode Mode 2: the 1st long-axis bending mode
500 500 500 500

400 400 400 400

Height [m]
300 300 300 300

Height [m]

Height [m]

Height [m]
200 200 200 200

100 100 100 100

0 0 0 0
−1 0 1 −1 0 1 −1 0 1 −1 0 1
X Component Y Component X Component Y Component

Mode 6: the 2nd short-axis bending mode Mode 7: the 2nd long-axis bending mode
500 500 500 500

400 400 400 400

Height [m]

300 300 300 300

Height [m]

Height [m]

Height [m]
200 200 200 200

100 100 100 100

0 0 0 0
−1 0 1 −1 0 1 −1 0 1 −1 0 1
X Component Y Component X Component Y Component

Mode 8: the 3rd short-axis bending mode Mode 9: the 3rd long-axis bending mode
500 500 500 500

400 400 400 400

Height [m]

300 300 300 300

Height [m]

Height [m]

Height [m]

200 200 200 200

100 100 100 100

0 0 0 0
−1 0 1 −1 0 1 −1 0 1 −1 0 1
X Component Y Component X Component Y Component

Fig. 5. Identified first six bending modes of the main tower ( : Reference; : Mode shape).

Through this OMA study, the identified modal prop- obtained under Typhoon Nanmadol, the 2nd
erties (natural frequencies, damping ratios, and mode short-axis bending mode is identified as the 6th
shapes) are compared with those identified using the mode (0.5784 Hz), the 1st torsion mode as the 5th
measurements recorded under the normal ambient exci- mode (0.4989 Hz), and the 2nd torsion mode as the
tation in 2010 (Niu et al., 2012), and the following modal 11th mode (1.2522 Hz).
property changes of the Canton Tower are observed: 2. Natural frequency changes. The identified natu-
ral frequencies of the Canton Tower under the
1. Mode sequence changes. In the OMA results ob- Typhoon Nanmadol are larger, as illustrated in
tained under the normal ambient excitation, the Figure 4a.
2nd short-axis bending mode was identified as 3. Damping ratio changes, especially the damping ra-
the 5th mode (0.4747 Hz), the 1st torsion mode tio of the 2nd mode, which is also the 1st long-axis
as the 6th mode (0.5055 Hz), and the 2nd tor- bending mode. From Figures 2b and 3b, it can be
sion mode as the 12th mode (1.2507 Hz) (Niu found that the dominant wind direction is more or
et al., 2012). However, in the OMA results less parallel with the long axis of the main tower,
672 Niu et al.

which indicates the damping ratio increase in the Table 1

2nd mode may come from the aerodynamic damp- Nodal coordinates (z) for the adjusted reduced-order FEM
ing contribution. Such damping ratio increase in
Node z (m) Node z (m) Node z (m)
the along-wind direction is also observed and dis-
cussed in Guo et al. (2012). 1 −10.00 13 171.10 25 417.45
2 0.00 14 204.25 26 427.85
All the above listed modal property changes actually 3 12.00 15 228.50 27 438.25
show the necessity of this OMA study. 4 22.25 16 275.30 28 446.80
5 30.63 17 308.25 29 480.00
6 58.65 18 332.15 30 497.60
7 84.65 19 344.65 31 502.00
4 MODEL UPDATING
8 95.05 20 355.05 32 520.70
9 105.45 21 375.85 33 547.20
As the flowchart in Figure 1 shows, the applied wind 10 119.30 22 384.24 34 562.70
load reconstruction strategy is model based. In other 11 147.05 23 395.65 35 580.00
words, a reliable structural model, which can closely 12 157.45 24 407.05 36 600.00
represent the dynamics of the Canton Tower, will help
to provide more accurate load information.
Ni et al. (2012) provided a beam-like reduced-order
FEM of the Canton Tower specifically for a benchmark Node 36
study. This reduced-order FEM was developed based
on an elaborate three-dimensional full-scale FEM of the
Canton Tower with the height of 610 m. Node 29
As illustrated in Section 3, the height of the Canton
Tower was reduced from 610 m to 600 m after January
20, 2011, and a Ferris wheel was installed and open to
the public on September 1, 2011.
Before performing the model updating, a structural
model which is consistent with the sensor positions and
the current tower geometry is needed. In this article, the
available reduced-order FEM is first modified from the
following two points,
z
1. The nodal coordinates (z) are adjusted to coincide y x
with the positions of the accelerometers.
2. The change in the mast height is considered. Node 1

After these two modifications, the number of the Fig. 6. Adjusted reduced-order FEM.
nodes is reduced from 38 to 36. Table 1 lists the nodal
coordinates (z) of the modified reduced-order FEM. where
Figure 6 shows the beam-like reduced-order model
after the modification. The beam-like model in Figure 6 Ω = Km (13)
has in total 36 nodes and 35 beam elements, with 27
elements for the main tower and 8 elements for the an- and
tenna mast. Each node has two horizontally translated Ψ = ψ1 ψ2 · · · ψnDOF (14)
DOFs and three rotational DOFs. With the node 1 con-
strained, the modified reduced-order FEM has in total Introduce the residual vector
175 unconstrained DOFs. Now this modified reduced-  ω
r
order FEM is ready to be updated with respect to the r= (15)
OMA results. rψ
The undamped eigenvalue problem of Equation (1)
where the ith element of r ω and r ψ appear as
reads
ωiOMA − ωi
K nΨ = MnΨ Ω2 (12) r iω = (16)
ωiOMA
 Online reconstruction of wind load and structural responses for Canton Tower 673

and the model updating process, are listed in Table 2, from


OMA 2
which the following statements can be made,
ψ ψiT ψi
r i = 1 − MACi =  T
 (17)
ψiT ψi ψiOMA ψiOMA 1. After the model updating, the discrepancy be-
tween the natural frequencies calculated from the
In Equations (16) and (17), ωiOMA and ψiOMA are the modified reduced-order FEM and those from the
OMA counterpart of ωi and ψi ; and MAC is short for OMA are minimized.
modal assurance criteria. Define the objective function 2. After the model updating, the mode shapes cal-
culated from the modified reduced-order FEM
are more consistent with those identified from the
J = Jr + λ2 Jθ (18) OMA.
where
The above two observations show that the updated
Jr = r T W γ r (19) modified reduced-order FEM already can reasonably
represent the dynamics of the Canton Tower within the
and frequency range of the first 12 modes.
Jθ = θT θ (20)
In Equations (19) and (20), Jr is the measure- 5 APPLICATION-ORIENTED ALGORITHM
ment residual; Jθ is the side constraint; λ denotes the SELECTION GUIDANCE
regularization parameter; W r is a weighting matrix;
and θ represents the updating parameter change. The As stated in Section 1, there are already several avail-
model which corresponds to the minimum of J can best able online force reconstruction algorithms. A practical
represent the real structure. question confronting us now is which algorithm to

⎡ ⎤
α X K 1,1
e
α X αY K 1,2
e
α X K 1,3
e
α X K 1,4
e
α X K 1,5
e
α X K 1,6
e
α X αY K 1,7
e
α X K 1,8
e
α X K 1,9
e
α X K 1,10
e
⎢ αY K 2,2
e
αY K 2,3
e
αY K 2,4
e
αY K 2,5
e
α X αY K 2,6
e
αY K 2,7
e
αY K 2,8
e
αY K 2,9
e
αY K 2,10
e ⎥
⎢ ⎥
⎢ ⎥
⎢ e
K 3,3 e
K 3,4 e
K 3,5 α X K 3,6
e
αY K 3,7
e e
K 3,8 e
K 3,9 e
K 3,10 ⎥
⎢ ⎥
⎢ e
K 4,4 e
K 4,5 α X K 4,6
e
αY K 4,7
e e
K 4,8 e
K 4,9 e
K 4,10 ⎥
⎢ ⎥
⎢ e
α X K 5,6
e
αY K 5,7
e e e e ⎥
⎢ K 5,5 K 5,8 K 5,9 K 5,10 ⎥
K Ue =⎢ ⎥ (21)
⎢ α X K 6,6
e
α X αY K 6,7
e
α X K 6,8
e
α X K 6,9
e
α X K 6,10 ⎥
e
⎢ ⎥
⎢ αY K 7,7
e
αY K 7,8
e
αY K 7,9
e
αY K 7,10 ⎥
e
⎢ ⎥
⎢ e e e ⎥
⎢ Symmetric K 8,8 K 8,9 K 8,10 ⎥
⎢ ⎥
⎣ e
K 9,9 K 9,10 ⎦
e

e
K 10,10

Inspired by the idea in Chung et al. (2012), the model choose. In this section, an application-oriented algo-
updating in this article assigns two updating parameters rithm selection guidance is proposed, which may help
α X ( j) and αY ( j) (1 ≤ j ≤ 35), to each element stiffness to find a suitable algorithm for a specified practical ap-
matrix to represent the bending stiffness changes in the plication.
X and Y directions. The mass matrices are all kept con- According to the authors’ knowledge, the current
stant. A typical updated element stiffness matrix can be available online force reconstruction algorithms can be
expressed as in Equation (21), where K e is the initial divided into the following groups,
element stiffness matrix and K Ue is the updated one.
The first 12 modes identified from the OMA are set 1. Estimators, without considering the direct
as the reference in the model updating process. The feedthrough in the state-space model, are only
weighting matrix W r is set as 100I 12 , and the regulariza- applicable for the case that the displacement,
tion parameter λ is set as 0.1. The MATLAB function strain or velocity measurements are available,
fmincon is adopted to minimize the objective function for example, the algorithms given in Tuan et al.
J . The natural frequencies and MAC values, calculated (1996), Soeffker (1999), Gillijns and De Moor
from the modified reduced-order FEM before and after (2007a), and Pan et al. (2009).
674 Niu et al.

Table 2
Model updating results

Natural frequency [Hz] MAC value (%)

Mode OMA Before updating After updating Before updating After updating
1 0.0919 0.1129 (22.83) 0.0918 (0.09) 91.11 96.96
2 0.1356 0.1610 (18.73) 0.1355 (0.08) 90.62 99.12
3 0.3735 0.3670 (1.73) 0.3727 (0.21) 88.92 97.76
4 0.4640 0.4539 (2.18) 0.4647 (0.16) 90.43 97.07
5 0.4989 0.4777 (4.24) 0.4977 (0.23) 8.16 98.39
6 0.5780 0.5822 (0.73) 0.5817 (0.64) 87.38 96.63
7 0.6017 0.5875 (2.37) 0.5985 (0.53) 85.15 97.82
8 0.8031 0.7449 (7.25) 0.8036 (0.06) 82.22 94.01
9 0.9722 0.9220 (5.16) 0.9729 (0.07) 74.17 96.16
10 1.2132 1.0983 (9.47) 1.2096 (0.30) 18.17 95.98
11 1.2522 1.1628 (7.14) 1.2523 (0.01) 1.52 91.82
12 1.2881 1.2481 (3.10) 1.2921 (0.31) 53.60 95.10
Average (7.08) (0.2232) 64.29 96.40
Note: The values in column OMA are the identified natural frequencies (in Hz) using measurements from Typhoon Nanmadol; the values in the
parentheses are the absolute values of the differences (%) from the OMA results.

2. Estimators, considering the direct feedthrough in whose sufficient conditions can be fulfilled, are actually
the state-space model, are applicable for the case suitable for the specific application.
that the acceleration measurements are available, As shown in Figure 7, the Canton Tower has a tube-
for example, the algorithms given in Ji and Liang in-tube geometry, consisting of a reinforced concrete
(2000), Ha and Trinh (2004), Gillijns and De Moor inner structure and a steel lattice outer structure. Four
(2007b), Hwang et al. (2009), and Pan et al. (2011). levels of connection girders (at heights around 204 m,
3. Estimators, which require a direct feedthrough 230 m, 272 m, and 303 m) connect the inner structure
with full column rank, are only applicable for the with the outer structure. On the SHM system, all the
case that the acceleration measurements are avail- accelerometers are installed on the reinforced concrete
able, for example, the algorithms given in Ha and inner structure, with a sampling frequency of 50 Hz
Trinh (2004), Gillijns and De Moor (2007b), and (Ni et al., 2012). The strain sensors instrumented on
Pan et al. (2011). the reinforced concrete inner structure are vibrating
4. Estimators, which not only require the direct wire strain gauges, with the sampling frequency at
feedthrough having full column rank, but also as- 1/60 Hz (Xia et al., 2011). The strain sensors installed
sume an unbiased estimate of the initial condition on the steel lattice outer structure are fiber Bragg
or assume the input force with zero mean, may gratings sensors, with the sampling frequency of 50 Hz
work stably with only acceleration measurements, (Ni et al., 2009).
but the static force component still needs to be The sampling frequency of the strain sensors on the
calculated either by displacement (or strain) mea- inner structure is lower than the 1st natural frequency
surements or from a priori information, for exam- of the Canton Tower. The sampling frequency of the
ple, the algorithms given in Gillijns and De Moor strain sensors on the outer structure is higher, up to 50
(2007b) and Hwang et al. (2009). Hz, but it is difficult to model the strain on the steel lat-
tice outer structure to the reduced-order FEM. Consid-
To be a stable estimator, each algorithm listed ering these two constraints on the available strain sen-
above has to fulfill its own sufficient conditions. Being sors, only the acceleration measurements are chosen to
translated from the mathematics to the engineering be used for this wind load reconstruction study. This in-
practice, such sufficient conditions are actually some dicates the methods in the group (4) seem suitable. As
requirements on the sensor type, the sensor number stated in Section 1, with the help of a high-pass filter to
and the sensor position. For a specific engineering remove the low frequency drift the algorithm proposed
problem, the sensors may have been chosen in advance by Gillijns and De Moor (2007b) may work stably. How-
by the engineers for different applications. Given the ever, the first 12 modes of the Canton Tower are all
information on the available sensors, the algorithms, below 1.29 Hz, which makes the algorithm given in
 Online reconstruction of wind load and structural responses for Canton Tower 675

Connection
girders

Closed
function
areas

(a) (b) (c)

Fig. 7. (a) Canton Tower. (b) Steel lattice outer structure. (c) Reinforced concrete inner structure.

Gillijns and De Moor (2007b) not suitable for this case. Defining the state estimate error e(t) = x(t) −
The authors finally chose the algorithm proposed by x̂(t), the state estimate error dynamic equation
Hwang et al. (2009) for this wind load reconstruction can be obtained by subtracting Equation (24) from
study. Equation (9a),

ė(t) = ( A − K C)e(t) + (G − K H) d(t) − K υ(t) (25)

6 ADOPTED INPUT AND STATE ESTIMATION Assume the unknown input vector d(t) and the mea-
ALGORITHM surement noise υ(t) are uncorrelated with each other,
zero mean, and stationary with covariance matrices
In this section, the main steps of the algorithm, which is
selected from Section 5, are first reviewed. Then a mod- E d(t)d T (t) = Q d and E υ(t)υ T (t) = Q υ
ification to this algorithm is proposed, considering the
wind load characteristics. (26)

where E [·] is the expectation operator. Then the gain

6.1 Original algorithm matrix K can be calculated with

In practice, the measurements always include the re- K = (G Q d G T + P C T )(HQd H T )−1 (27)
sponses from higher modes and the measurement noise,
so Equation (9b) can be extended to include a measure- The matrix P in Equation (27) can be obtained by
ment noise vector υ(t) ∈ R p , solving the following Riccati equation,

y(t) = C x(t) + H d(t) + υ(t) (22) A P + PAT + G T Q d G − (C P + HQd G T )T

Consider the following state estimator, (HQd H T + Q υ )−1 × (C P + HQd G T ) = 0 (28)
˙
x̂(t) = A x̂(t) + K (y(t) − ŷ(t)) (23a)
The steps in Equations (27) and (28) are quite similar
ŷ(t) = C x̂(t) (23b) to the design of steady-state Kalman filter in Brown and
Hwang (1997).
where x̂(t) denotes the estimate of the state vector x(t)
Once the gain matrix K is available, the least squares
in Equation (8); and the gain matrix K is to be deter-
estimate of the unknown input vector d(t) can be calcu-
mined.
lated using Equation (29),
Substituting Equations (22) and (23b) into Equa-
tion (23a) leads to −1 T
d̂(t) = G T G G K ( y(t) − C x̂(t)) (29)
˙
x̂(t) = ( A − KC) x̂(t) + KCx(t)
In Hwang et al. (2009), the covariance matrix Q d
+ KHd(t) + K υ(t) (24) is simply set as an identity matrix. This indicates the
676 Niu et al.

30 360
10−min mean wind direction N
10−min mean wind speed
25 300

Wind Direction [degrees]

60%
Wind Speed [m/s]

20 240 40%
20%
15 180 W E

10 120 18 − 19 m/s
17 − 18 m/s
16 − 17 m/s
15 − 16 m/s
5 60 14 − 15 m/s
13 − 14 m/s
S
0 0
05:00 07:00 09:00 11:00 13:00 (b)
Time [hh:mm]
(a)

Fig. 8. Wind measurements during Typhoon Kai-tak. (a) 10-minute mean wind speed and direction. (b) Rose diagram
(from [Link] on August 17, 2012 to [Link] on August 17, 2012).

unknown inputs are assumed uncorrelated with the 7.1 Wind load reconstruction
same variance.
The time-varying wind speed u(t) at the height of z me-
ters can be represented by Equation (30),
u (z, t) = ū (z) + ũ (z, t) (30)
6.2 Proposed modification
where ū (z) denotes the mean wind speed and ũ (z, t) the
If the covariance matrix Q d is set as an identity matrix, it fluctuating wind speed. According to Simiu and Scanlan
means the unknown inputs are assumed as uncorrelated (1996), The time-varying wind load f (z, t) at the height
and with the same variance. This is actually in conflict z meters can be calculated by Equation (31),
with the wind characteristics. 1
According to Simiu and Scanlan (1996), the power ρμaw u 2 (z, t)
f (z, t) = (31)
2
spectral density (PSD) of wind speed varies as the
height increases. Furthermore, the wind speeds at two where ρ, μ, and aw denote respectively the air density,
different heights are also correlated. These two points the aerodynamic force coefficient, and the orthogonal
indicate the covariance matrix Q d is usually not an iden- exposed wind area for the considered section. Substitut-
tity matrix but with different diagonal elements and ing Equation (30) into Equation (31), the time-varying
nonzero off-diagonal elements. wind load f (z, t) can be decomposed into two compo-
In this article, the covariance matrix Q d is calculated nents, the mean wind load f¯ (z) and the fluctuating wind
according to a simulated fluctuating wind field. Such cal- load f˜ (z, t), as Equations (32)–(34) show.
culation of Q d can be performed in advance and saved f (z, t) = f¯ (z) + f˜ (z, t) (32)
in the databank, so it is not in conflict with the realiza-
tion of the online application. 1
f¯ (z) = ρμaw (z) ū 2 (z) + E ũ 2 (z, t) (33)
2

f˜(z, t) = ρμaw (z) ū (z) ũ (z, t)

7 WIND LOAD AND STRUCTURAL RESPONSE
RECONSTRUCTION 1
+ ρμaw (z) ũ 2 (z, t) − E ũ 2 (z, t) (34)
2
In this section, the field measurements recorded during
Typhoon Kai-tak in 2012 are analyzed to demonstrate From Figure 8a, it can be seen that the measured
the reconstruction methodology illustrated in Figure 1. wind speed is fluctuating around 15 m/s, so the mean
The measurement results of wind speed and wind di- wind load on the Canton Tower is nonzero according to
rection at the height of 461 m are plotted in Figure 8. Equation (33).
 Online reconstruction of wind load and structural responses for Canton Tower 677

x 10
−3 Acceleration Channel 07 wind direction of Typhoon Nanmadol and that of
Acceleration [m/s/s]

6 Measured Typhoon Kai-tak are more or less parallel with each

4 Recon. other. So the modified reduced-order FEM, which
2
0 has been updated according to the OMA results for
−2 Typhoon Nanmadol is also applicable for Typhoon
−4 Kai-tak. In other words, these two typhoon events can
[Link] [Link]
Time [hh:mm:ss] be classified into one group.
−4
x 10 If the MATLAB function care is used to solve the
1.5
Measured Riccati equation in Equation (18), the sufficient con-
1 Recon.
ditions in need to be fulfilled can be summarized as
PSD

0.5 follows,

0 1. The input matrix G should have full column rank;

0 0.2 0.4 0.6 0.8 1 1.2 1.4
Frequency [Hz] 2. The matrix pair ( A,C) is observable;
3. The matrix pair ( A,G Q d G T ) is controllable.
Fig. 9. Measured and reconstructed acceleration of
Channel 07. These sufficient conditions actually define the re-
quirements on the number and position of the ac-
−3
x 10 Acceleration Channel 20 celerometers.
Acceleration [m/s/s]

5 Measured In this article, the first 12 modes are included in

Recon. the state-space structural model. By fulfilling the
0 above sufficient conditions, it is concluded that at least
12 accelerometers are needed. The acceleration mea-
−5
[Link] [Link] surements from Channels 1, 2, 5, 6, 11, 12, 13, 14, 15,
Time [hh:mm:ss]
−4 16, 17, and 19 are selected for the reconstruction of the
x 10
1.5 fluctuating wind load in this study.
Measured
Recon. With the wind speed measurements at 461 m, the
1
mean wind speed can be calculated. A wind velocity
PSD

0.5 field is simulated using the algorithm proposed by Ding

et al. (2006), and the covariance matrix Q d can be es-
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 timated with the help of Equation (31). The aerody-
Frequency [Hz]
namics force coefficient μ and the orthogonal exposed
wind area aw for each tower section are taken from the
Fig. 10. Measured and reconstructed acceleration of
Channel 20.
wind tunnel test made by Gu et al. (2006) and Zhou
et al. (2010). The covariance matrix of the measurement
Table 3 noise, Q υ , is set as 10−8 I 12 . It is noted here, that the co-
Parameters used in mean wind load calculation variance matrix Q d is calculated in advance and saved in
the database. During the online wind load reconstruc-
Terrain Gradient Air density ρ
category∗ height z g ∗ (m) Exponent α ∗ (kg/m3 )
tion, the algorithm can select the corresponding Q d ac-
cording to the real-time wind speed measurements.
C 400 0.22 1.225 Because there is no direct measurement for the fluc-
∗ Accordingto the Building Structure Design Code of China (50009- tuating wind load, a reasonable question which may
2001; GB 50009-2001, 2006). come up is how to validate the reconstruction results.
Within this article, the following steps are adopted for
the validation.
As stated in Section 6, the adopted algorithm has the
assumption that the unknown input has zero mean. This 1. Reconstruct the fluctuating wind load;
indicates the algorithm can only reconstruct the fluc- 2. Apply the reconstructed fluctuating wind load to
tuating wind load if using acceleration measurements the updated modified reduced-order FEM to re-
only. In this article, the mean wind load is calculated construct the acceleration response for the ac-
with the help of the wind speed measurements and the celerometers, Channel 07 and Channel 20;
aerodynamic force coefficients identified from the wind 3. Compare the reconstructed acceleration with
tunnel test by Gu et al. (2006) and Zhou et al. (2010). those from the sensor measurements.

7.1.1 Fluctuating wind load reconstruction. From As mentioned earlier in this section, the acceleration
Figures 2b, 3b, and 8b it can be seen that the dominant measurements from Channels 1, 2, 5, 6, 11, 12, 13, 14,
678 Niu et al.

700 700
Mean Mean
Standard deviation Standard deviation
600 600

500 500

400 400
Height [m]

Height [m]
300 300

Closed
200 function 200
areas

100 100
z
y x
0 0

-120 -100 -80 -60 -40 -20 0  5 10 15 20 25 30 35 40 45

Force [kN] Force [kN]
(a) (b)

Fig. 11. Reconstructed wind load (a) in the x direction, (b) in the y direction.

15, 16, 17, and 19 are used by the algorithm for recon- According to Simiu and Scanlan (1996), the mean
structing the fluctuating wind load, whereas the mea- wind speeds at different heights follow the power law,
surements from Channels 07 and 20 are not used. If
the reconstructed acceleration for Channels 07 and 20  α
z
are consistent with their sensor measurement, it actually ū (z) = ū g (36)
zg
indirectly validates the fluctuating wind load reconstruc-
tion results. where z g denotes the gradient height; and ū g is the
Figures 9 and 10 show a part of the reconstructed mean wind speed at z g meters. Using the wind speed
acceleration of Channels 07 and 20. From these two measurements at the height of 461 m and the parame-
figures, it can be seen that the time history and PSD of ters information listed in Table 3, the mean wind speeds
the reconstructed acceleration fit quite well with their at different heights can be calculated. According to
corresponding sensor measurements. This shows that Table 1, Nodes 3–36 have positive z coordinates, which
the reconstructed fluctuating wind load can produce means these nodes are above the ground and under the
structural responses which are coincident with the sen- wind load effect. The orthogonal exposed wind area
sor measurement. aw (z) for each node is calculated according to the tower
dimension given in Gu et al. (2006) and Zhou et al.
(2010). The mean wind load in the x and the y direc-
7.1.2 Mean wind load reconstruction. Besides the for- tions and the mean moment around the z direction can
mula given in Equation (33), the mean wind load can be calculated by Equation (35).
also be calculated by Equation (35),
7.1.3 Some comments on the reconstructed wind load.
According to Equation (32), the reconstructed wind
1
f¯ (z) = ρ μ̄W T aw (z) ū 2 (z) (35) load can be obtained by summing up the reconstructed
2 fluctuating wind load and the reconstructed mean wind
load.
where μ̄W T denotes the mean aerodynamic force coef- Figure 11 shows the reconstructed wind load in the
ficient. In this study, the values for μ̄W T are taken from x and y directions. Figure 12 shows the reconstructed
the wind tunnel test made by Gu et al. (2006) and Zhou moment around the z direction. In these two figures,
et al. (2010). the red bars correspond to the reconstructed wind load
 Online reconstruction of wind load and structural responses for Canton Tower 679

700 8 CONCLUSIONS
Mean
Standard deviation In this article, an online simultaneous wind load and
600
structural response reconstruction methodology is ap-
plied to the 600 m tall Canton Tower. An OMA is
500 performed to identify the modal properties of the
Canton Tower under Typhoon Nanmadol in 2011. Then
a reduced-order FEM of the Canton Tower is up-
400
dated according to the OMA results. With the help
Height [m]

of the proposed application-oriented algorithm selec-

300 tion guidance, a type of state and input estimator is
adopted in this study. The equivalent lateral loads and
200
moments, acting on the nodes of the reduced-order
FEM, are reconstructed. The validation results show
that the reconstructed fluctuating wind load can pro-
100
z duce the acceleration responses which are coincident
y x with the sensor measurement. The mean component of
0 the loads and moments are calculated using the real-
time wind speed measurements and the available aero-
-45-40-35-30-25-20-15-10 -5 0 5 dynamic force coefficients. This methodology needs
Moment [kN*m] reliable structural models and correct sensor measure-
ments. In the next step, the measurements recorded
Fig. 12. Reconstructed moment around the z direction. under different wind events should be studied, so that
different conditions, for example, different wind direc-
tion, different wind loading mechanism (vortex shed-
(or moment), and the standard deviations are calculated ding), the fluid–structure interaction, could be consid-
from the reconstructed fluctuating wind load. Within ered. Furthermore, an online sensor fault detection
the tower section, from 200 m to 350 m, the nodal co- algorithm could also be included to make this method-
ordinates (z) are sparse. This results in relatively larger ology more robust.
values for the orthogonal exposed wind area aw , and
causes the reconstructed mean wind loads within this
ACKNOWLEDGMENTS
tower section to be larger.
The wind load reconstruction presented in this article
The authors would like to thank the editor and the
is only one example (or one case). There are still more
anonymous reviewers for their constructive comments
practical conditions in need to be considered, for ex-
and valuable suggestions on improving the quality of the
ample, different wind directions, fluid–structure inter-
article. The work published in this article was supported
action, wind loading mechanism (vortex shedding), and
by the Multi Modal Sensor Systems for Environmen-
so on. In the next step, the measurements recorded un-
tal Exploration and Safety (MOSES) program in NRW,
der different wind events should be studied, so that a
Germany, and by a grant from the Germany/Hong
comprehensive database for the structural model can be
Kong Joint Research Scheme sponsored by the
constructed.
German Academic Exchange Service (DAAD) and the
Furthermore, it would be interesting to use the real-
Research Grant Council (RGC) of Hong Kong (Refer-
time strain measurements to derive the equivalent mean
ence No. G˙HK035/11).
wind load on the nodes of the reduced-order FEM of
the Canton Tower (Xia et al., 2014).

REFERENCES
7.2 Sensor signal reconstruction
Berg, J. C. & Keith Miller, A. (2010), Force estimation via
As already validated in Subsection 7.1.1, the adopted kalman filtering for wind turbine blade control, presented
algorithm can reconstruct not only the fluctuating wind in IMAC XXVIII, Jacksonville, FL, USA.
load but also the acceleration response. Such structural Brown, R. G. & Hwang, P. Y. C. (1997), Introduction to
response reconstruction ability actually raises the possi- Random Signals and Applied Kalman Filtering: With Mat-
lab Exercises and Solutions, John Wiley & Sons, Inc.,
bility of reconstructing the sensor signals of faulty sen- New York.
sors or the structural responses at a position where sen- Chen, W. H., Lu, Z. R., Lin, W., Chen, S. H., Ni, Y. Q., Xia, Y.
sor installation is difficult. & Liao, W. Y. (2011), Theoretical and experimental modal
680 Niu et al.

analysis of the Guangzhou New TV Tower, Engineering Earthquake Engineering & Structural Dynamics, 42, 477–
Structures, 33, 3628–46. 98.
Chung, T. T., Cho, S., Yun, C. B. & Sohn, H. (2012), Finite ele- Loh, C. H., Liu, Y. C. & Ni, Y. Q. (2012), SSA-based stochas-
ment model updating of Canton Tower using regularization tic subspace identification of structures from output-only vi-
technique, Smart Structures and Systems, 10(4–5), 459–70. bration measurements, Smart Structures and Systems, 10(4–
Ding, Q., Zhu, L. D. & Xiang, H. (2006), Simulation of station- 5), 331–51.
ary gaussian stochastic wind velocity field, Wind and Struc- Lourens, E., Lombaert, G., De Roeck, G. & Degrande, G.
tures, 9(3), 231–43. (2008), Reconstructing time-varying wind loads from vi-
Faravelli, L., Ubertini, F. & Fuggini, C. (2010), System bration responses, Proceedings of ISMA 2008 International
identification toward FEM updating of a super high-rise Conference on Noise and Vibration Engineering, Leuven,
building, Proceedings of the Fifth European Workshop Belgium, pp. 639–48.
on Structural Health Monitoring, Sorrento, Naples, Italy, Lourens, E., Maes, K., Van Nimmen, K., Van den Broeck, P.,
pp. 1023–30. Guillaume, P., De Roeck, G. & Lombaert, G. (2012a), Mul-
Fritzen, C.-P., Klinkov, M. & Kraemer, P. (2013), Vibration- tiple localized force identification and response prediction
based damage diagnosis and monitoring of external loads, on a footbridge, Proceedings of ISMA 2012 International
in W. Ostachowicz and J. A. Guemes (eds.), New Trends in Conference on Noise and Vibration Engineering, Leuven,
Structural Health Monitoring, Springer, Wien, Austria. Belgium, pp. 1089–1102.
Gillijns, S. & De Moor, B. (2007a), Unbiased minimum- Lourens, E., Papadimitriou, C., Gillijns, S., Reynders, E., De
variance input and state estimation for linear discrete-time Roeck, G. & Lombaert, G. (2012b), Joint input-response
systems, Automatica, 43, 111–16. estimation for structural systems based on reduced-order
Gillijns, S. & De Moor, B. (2007b), Unbiased minimum- models and vibration data from a limited number of
variance input and state estimation for linear discrete-time sensors, Mechanical Systems and Signal Processing, 29,
systems with direct feedthrough, Automatica, 43, 934–37. 310–27.
Gu, M., Huang, P., Zhou, X. Y. & Zhu, L. D. (2006), Wind Ni, Y. Q., Wong, K. Y. & Xia, Y. (2011), Health checks
Tunnel Test on Sectional Models and Analysis of Wind In- through landmark bridges to sky-high structures, Advances
duced Vibration of the Guangzhou New TV Tower, Re- in Structural Engineering, 14(1), 103–19.
search report (in Chinese), Tongji University, China. Ni, Y. Q., Xia, Y., Liao, W. Y. & Ko, J. M. (2009), Technology
Guo, Y. L., Kareem, A., Ni, Y. Q. & Liao, W. Y. (2012), Per- innovation in developing the structural health monitoring
formance evaluation of Canton Tower under winds based system for Guangzhou New TV Tower, Structural Control
on full-scale data, Journal of Wind Engineering and Indus- and Health Monitoring, 16, 73–98.
trial Aerodynamics, 104–6, 116–28. Ni, Y. Q., Xia, Y., Lin, W., Chen, W. H. & Ko, J. M. (2012),
Ha, Q. P. & Trinh, H. (2004), State and input simultaneous SHM benchmark for high-rise structures: a reduced-order
estimation for a class of nonlinear systems, Automatica, 40, finite element model and field measurement data, Smart
1779–85. Structures and Systems, 10(4–5), 411–26.
Hernandez, A. V., Swartz, R. A. & Zimmerman, A. T. (2011), Niu, Y. & Fritzen, C.-P. (2012), Comparison and prac-
A framework for embedded load estimation from structural tical aspects of two approaches for online load re-
response of wind turbines, presented in IMAC XXIX, Jack- construction, Proceedings of the 6th European Work-
sonville, FL, USA. shop on Structural Health Monitoring, Dresden, Germany,
Hwang, J.-S., Kareem, A. & Kim, H. (2011), Wind load iden- pp. 651–58.
tification using wind tunnel test data by inverse analysis, Niu, Y., Kraemer, P. & Fritzen, C.-P. (2012), Operational
Journal of Wind Engineering and Industrial Aerodynamics, modal analysis for Canton Tower, Smart Structures and Sys-
99, 18–26. tems, 10(4–5), 393–410.
Hwang, J.-S., Lee, S.-G., Ji-hoon, P. & Eun-Jong, Y. Pan, S., Su, H., Wang, H. & Chu, J. (2011), The study of joint
(2009), Force identification from structural responses using input and state estimation with Kalman filtering, Transac-
Kalman filter, Materials Volume, 33, 257–66. tions of the Institute of Measurement and Control, 33(8),
Ji, C.-C. & Liang, C. (2000), A study on an estimation method 901–18.
for applied force on the rod, Computer Methods in Applied Pan, S., Su, H., Wang, H., Chu, J. & Lu, R. (2009), Input and
Mechanics and Engineering, 190, 1209–20. state estimation for linear systems: a least squares estima-
Klinkov, M. (2011), Identification of unknown structural loads tion approach, Proceedings of the 7th Asian Control Con-
from dynamic measurements using robust observers. Ph.D. ference, Hong Kong, China, pp. 378–383.
thesis, University of Siegen, Germany. Rencher, A. C. (2002), Methods of Multivariate Analysis, John
Klinkov, M. & Fritzen, C.-P. (2010), Wind load observer for a Wiley & Sons, Inc., Canada.
5 mw wind energy plant, presented in IMAC XXVIII, Jack- Simiu, E. & Scanlan, R.H. (1996), Wind Effects on Structures,
sonville, FL, USA. Wiley, New York.
Kraemer, P. & Fritzen, C.-P. (2010), Vibration analysis for Soeffker, D. (1999), Observer-based measurement of contact
structures of offshore wind energy plants, Proceedings of forces of the nonlinear rail-wheel contact as a base for ad-
the 10th German Wind Energy Conference (DEWEK 2010), vanced traction control, Mechatronics and Advanced Mo-
Bremen, Germany. tion Control, 49, 305–20.
Law, S. S., Bu, J. Q. & Zhu, X. Q. (2005), Time-varying wind Stevens, K. K. (1987), Force identification problems - an
load identification from structural responses, Engineering overview, Proceedings of Spring Meeting of SEM, Houston,
Structures, 27, 1586–98. Texas, USA, pp. 838–44.
Liu, Y. C., Loh, C. H. & Ni, Y. Q. (2013), Stochastic subspace Trujillo, D. M. (1978), Application of dynamic programming
identification for output-only modal analysis: application to to the general inverse problem, International Journal for
super high-rise tower under abnormal loading condition, Numerical Methods in Engineering, 12, 613–24.
 Online reconstruction of wind load and structural responses for Canton Tower 681

Tuan, P.-C., Ji, C.-C., Fong, L.-W. & Huang, W.-T. (1996), An Ye, X., Yan, Q., Wang, W. & Yu, X. (2012), Modal identifica-
input estimation approach to on-line two dimentional in- tion of Canton Tower under uncertain environmental con-
verse heat conduction problems, Numerical Heat Transfer, ditions, Smart Structures and Systems, 10(4-5), 353–73.
Part B: Fundamentals, 29(3), 345–63. Zhang, L., Brincker, R. & Andersen, P. (2005), An overview
Xia, Y., Ni, Y. Q., Zhang, P., Liao, W. Y. & Ko, J. M. (2011), of operational modal analysis: major development and is-
Stress development of a supertall structure during construc- sues, Proceedings of the 1st International Operational Modal
tion: field monitoring and numerical analysis, Computer- Analysis Conference, Copenhagen, Denmark, pp. 179–90.
Aided Civil and Infrastructure Engineering, 26, 1-18. Zhou, X. Y., Huang, P., Gu, M., Zhu, L. D. & Pan, H. (2010),
Xia, Y., Zhang, P., Ni, Y. Q. & Zhu, H. P. (2014), Deformation Wind loads and wind-induced responses of Guangzhou
monitoring of a super-tall structure using real-time strain New TV Tower, Advances in Structural Engineering, 13(4),
data, Engineering Structures, 67, 29–38. 707–26.