You are on page 1of 20

Received: 17 April 2018 Revised: 6 February 2019 Accepted: 10 April 2019

DOI: 10.1002/stc.2383

RESEARCH ARTICLE

Structural health monitoring of a 250‐m super‐tall building


and operational modal analysis using the fast Bayesian FFT
method

Feng‐Liang Zhang1,2 | Yan‐Ping Yang1 | Hai‐Bei Xiong1 | Jia‐Hua Yang1 | Zhi Yu1

1
Department of Disaster Mitigation for
Structures, Tongji University, Shanghai,
Summary
China This paper presents the work on the structural health monitoring design and
2
School of Civil and Environmental operational modal analysis of a 250‐m super‐tall building situated in Shanghai,
Engineering, Harbin Institute of
China. The building is a steel–concrete composite structure with a steel com-
Technology, Shenzhen, China
posite frame‐concrete core tube system. At the 21st and 36th–38th floors, out-
Correspondence rigger trusses and ring‐shaped trusses are set to strengthen this structure.
Hai‐Bei Xiong, Department of Disaster
Because the height of this structure is overlimited and its lateral stiffness in
Mitigation for Structures, Tongji
University, Shanghai, China. the vertical direction is nonuniform, a SHM system was designed to monitor
Email: xionghaibei@tongji.edu.cn the structural condition and evaluate its safety. The SHM system is presented
in this paper, and the instrumented equipment includes accelerometers and tilt
Funding information
Interdisciplinary Cooperation Foundation sensors. The locations of sensors were well arranged so that SHM could be con-
for the Peak Discipline of Civil Engineer- ducted using the least amount of sensors. On the basis of the system, an ambi-
ing at Tongji University; National Natural
Science Foundation of China, Grant/
ent vibration test was carried out to perform the analysis with four setups
Award Number: 51878484; China designed to investigate the modal parameters in the X and Y directions. The
National Science and Technology Major Fast Bayesian FFT method was employed to perform the operational modal
Project “Development of Energy Effi-
ciency Technologies for Tall Buildings in analysis. The first 10 modes were identified. The modal parameters obtained
Mega Cities Combined with Structural by the Bayesian method are studied and discussed, and they are compared with
Safety Solutions under Multi‐hazard
the results obtained by other methods. Long‐term monitoring of this super‐tall
Impacts”, Grant/Award Number:
2016YFE0105600 building was also carried out to investigate the change of modal parameters in
different stages. The results of this study are expected to provide a reference for
model updating, damage detection, and SHM of this high‐rise building.

KEYWORDS
Bayesian approach, modal parameters, operational modal analysis, structural health monitoring,
super‐tall building

1 | INTRODUCTION

Modal parameters including natural frequency, damping ratio, and mode shape are the basic dynamic characteristics of
structures that can reflect structural changes.1-13 Although these parameters can be obtained by numerical analysis or
scaled laboratory experiments, errors from measurement, modeling, and numerical calculation always exist.14,15 There-
fore, it is necessary to investigate these modal parameters based on field measurements. For super‐tall buildings, SHM
systems are usually designed to collect structural response such as acceleration, strain, and temperature.16,17 This can be
viewed as full‐scale experiments of structures, based on which local damage or overall condition of structures can be
Struct Control Health Monit. 2019;26:e2383. wileyonlinelibrary.com/journal/stc © 2019 John Wiley & Sons, Ltd. 1 of 20
https://doi.org/10.1002/stc.2383
2 of 20 ZHANG ET AL.

assessed to evaluate their safety or performance.18-20 How to design a good SHM system is a challenging issue, and it
should consider the balance between efficiency and economy.21
Operational modal analysis (OMA) allows one to identify modal properties of a structure using only output
vibration measurements (often acceleration) under operating environment.3 According to different excitation
manners, vibration tests can usually be classified into three types: forced vibration, free vibration, and ambient vibra-
tion tests. Compared with the two previous ones, which usually require artificial loading, ambient vibration tests are
more convenient and economical, because excitations of ambient vibration tests are mainly composed of micro
tremors, wind, and other ambient incentives, and lighter and less equipment is required.22,23 Furthermore, modal
parameters obtained from ambient vibration tests can well reflect operational conditions of structures. OMA can
be conducted to analyze ambient vibration data obtained from SHM systems to evaluate structural dynamic charac-
teristics. There are several kinds of methods that can be used for OMA,3,24,25 whose uncertainty quantification has
attracted increasing attention.26-29 Bayesian methods provide an important means to determine the most probable
values of parameters and quantify the associated posterior uncertainty.30-36 They have been applied to many kinds
of structures, for example, buildings, bridges, and towers,37-42 and extended to analyze free43,44 and forced45-47
vibration data.
Many researches have investigated structural conditions based on SHM systems and OMA. Aktan et al48 developed a
SHM method that offers solutions for measuring structural responses and provides real‐time information under extreme
situations. Yang et al49 presented a method of identifying natural frequencies and damping ratios of tall buildings based
on ambient wind vibration data. For high‐rise buildings, the vibration mechanism under strong winds and earthquakes
is of great significance. Yuen et al1 established a long‐term monitoring program for a 22‐storey reinforced concrete
building to monitor its structural dynamic responses and trace the variation of its modal frequencies for 1 year. Chen
et al50 presented the analytical and experimental modal analysis of the Guangzhou New TV Tower based on the ambi-
ent vibration measurement at different construction stages and under different excitation conditions. The world's tallest
building, Burj Dubai, was installed with accelerometers and displacement transducers to collect the dynamic response
under wind conditions. Kijiewski et al51 designed a unique prototype system, SmartSync, so that data streams from dis-
tributed heterogeneous sensors could be pushed through network interfaces in real time. Zhang et al52 presented their
work on the SHM of Shanghai Tower (the second tallest building in the world) and conducted OMA for this structure
using a Bayesian approach. The modal parameters in different constructed stages of this super‐tall building were inves-
tigated. The variation of modal parameter due to the change of environmental temperature, building mass, and excita-
tion level was studied.
This paper presents the work on the SHM of a 250‐m high‐rise structure located in Shanghai. Because the height of
the structure exceeds the limit in the existing Chinese code and its stiffness changed during the construction, it was nec-
essary to install a SHM system to continuously assess structural conditions. The SHM system has two kinds of sensors,
that is, tilt sensors and accelerometers. The sensors' locations were properly selected to meet the requirements of effi-
ciency and economy. On the basis of the system, ambient vibration tests incorporating multiple setups were carried
out to cover all the DOFs of interest. The fast Bayesian FFT method was used to analyze the ambient vibration data.
The first 10 modes were identified, including four translational modes in the X direction (weak direction) of the struc-
ture, two translational modes in the Y direction (strong direction), and three torsional modes. Two other OMA methods,
that is, the enhanced frequency domain decomposition (EFDD) and stochastic subspace identification (SSI), were also
applied. The results obtained in different methods were compared and investigated. This study will provide a reference
for model updating, damage detection, and SHM of this high‐rise building.

2 | D E S C R I P T I O N O F T H E 25 0‐M ‐TALL BUILDING

The target structure is a super‐tall building with a height of 250 m and 56 stories including four stories underground and
52 stories above ground. The height of a standard floor is 4.2 m. The first four floors serve as a shopping center and res-
taurants; the fifth floor is an equipment floor, and the floors above serve as office space. The structure is high and com-
paratively flexible. It is thus sensitive to the static and dynamic effects of wind loads. The building resists lateral load by
steel–concrete composite system that consists of a reinforced concrete core tube, a steel‐reinforced concrete column, and
a steel‐frame beam. An arm truss and a belt truss are arranged between the core cylinder and the frame column, and
between two frame columns. The arm truss and belt truss are set on the 21st floor, 21st mezzanine, 36th floor, 37th
floor, and 38th floor to form two strengthened floors. Figure 1 shows the overview of the whole structure.
ZHANG ET AL. 3 of 20

FIGURE 1 SHM system design (unit: m)

3 | STRUCTURAL HEALTH MONITORING SYSTEM

Because the height of the structure is over 250 m, a SHM system was suggested to collect the structural response sub-
jected to extreme loads (e.g., earthquakes and strong winds). A SHM system was designed as shown in Figure 1 such
that the conditions and safety of the building could be evaluated continuously based on the data acquired. This system
mainly includes two kinds of sensors, that is, accelerometers and tilt sensors. Due to the limited number of sensors, only
some typical floors that can well provide information of the structural properties were selected for sensor installation,
including the basement, ground floor, strengthened floors, and top floor. As shown in Table 1, along the height of
the main structure, the B4 floor (basement), first floor (ground floor), 21st floor (strengthened floor), 37th floor
(strengthened floor), and 51st floor (top floor) were selected to install accelerometers in the X and Y directions. To mon-
itor the structural response subjected to vertical earthquake excitation, one additional uniaxial accelerometer was
installed vertically on the B4 floor, first floor and 51st floor, respectively. The sensors in the basement were used to mea-
sure earthquake input whereas the others were used to collect structural output.
In addition to the accelerometers, tilt sensors were also installed to measure the story drift angles, based on which the
horizontal displacements in the instrumented floors of the structure can be calculated. On the 11th, 21st, 29th, 37th,
44th, and 51st floors, tilt sensors were installed to measure the story drift angles in the X and Y directions of the struc-
ture. The Layout of the tilt sensors are shown in Table 2. Note that the 21st and 37th floors are two strengthened floors.

TABLE 1 Layout of the accelerometers

Location Floor Height (m) The number of accelerometers Measured directions

LB4 B4 −18.8 3 X/Y/Z


L1 1 0 3 X/Y/Z
L21 21 89.4 2 X/Y
L37 37 159.0 2 X/Y
L51 51 219.1 3 X/Y/Z
4 of 20 ZHANG ET AL.

TABLE 2 Layout of the tilt sensors

Location Floor Height (m) The number of inclinometers Measured directions

L11 11 47.2 1 X/Y


L21 21 89.4 1 X/Y
L29 29 125.3 1 X/Y
L37 37 159.0 1 X/Y
L44 44 188.5 1 X/Y
L51 51 219.1 1 X/Y
LR Roof 249.8 1 X/Y

The story drift angles and horizontal displacements of this structural system are controlled by these strengthened floors.
The curve between the two strengthened floors forms a horseshoe. Therefore, the tilt sensors were installed on both the
strengthened (the 21st and 37th floors) and ordinary floors (the 11th, 29th, and 44th floors), such that the horizontal
displacements of the structure could be well captured.
To make the whole system stable, a wired data acquisition system was designed. The measured locations were
selected inside weak electricity rooms of the corresponding floors. These rooms are at the corners of the concrete core
and form a straight line in the vertical direction, making it possible to lay cables conveniently. Because the measured
locations are distributed widely in space from the top to the bottom of the building, it is difficult to use one single data
acquisition unit to collect all the data simultaneously. Therefore, subunits were used to acquire the structural responses.
These subunits were connected by cables to form a whole data acquisition system. According to the suggestion of the
equipment supplier, to ensure that the data collected by different subunits is completely synchronized, the length of
the cables between two neighboring substations should be less than or equal to 100 m. To meet this requirement, four
data acquisition units were put on the first, 21st, 37th, and 51st floors, respectively, and each one could provide eight
channels. The accelerometers and tilt sensors could be connected to the same data acquisition unit. However, the loca-
tion of the weak electricity room changes at the fourth floor, which made the length of the cable between the two sub-
units at the first and 21st floors larger than 100 m. To solve this problem, an additional data acquisition unit was added
at the seventh floor. In total, five data acquisitions units were used in this system. The controlling room was set on the
B1 floor.
pffiffiffiffiffiffi
Uniaxial piezoelectric accelerometers with the sensitivity 50 V/g, the noise floor about 1 μg= Hz, and the frequency
range from 0.05 to 500 Hz were used (see Figure 2b). A base was designed to help the alignment of the accelerometers in
three directions (see Figure 2). Three accelerometers were fixed on a base and the base was then fixed to a floor slab.
The tilt sensor is shown in Figure 3a. The resolution of these sensors is 0.008%, and each one can perform an auto-
matic temperature compensation for the tilt data. The accuracy of the sensors can reach ±0.05%. These sensors were
installed on the walls inside the weak electricity room by bolts and supports. The cables of tilt sensors on the 11th,
29th, and 44th floors were connected to the data acquisition (DAQ) unit on the first, 21st, and 37th floors, respectively.
The DAQ unit in the system as shown in Figure 3b is made by DEWESoft®. It has eight channels and can collect sig-
nals from both accelerometers and tilt sensors. In each unit, there is an input channel and an output channel to connect
its neighboring data acquisition unit by a cable. For the units at the first and 51st floors, power supplies are provided to
supply electricity to the whole system. During the construction of this building, the cables to connect the data acquisi-
tion units were first laid, which made it possible to carry out some field tests before the whole system went into
operation.

4 | AMB IENT VI B R AT I ON TES T BA SE D O N T HE S HM SY ST E M

On the basis of the SHM system, ambient vibration tests were carried out to investigate the modal properties of this
super‐tall building. If only the accelerometers in the SHM system are used, it is difficult to identify the higher order
modes and torsional modes. This is because the number of sensors along the height is too small, and on each floor, only
one location is measured. Therefore, to obtain detailed information for mode shapes, six additional uniaxial accelerom-
eters combined with the sensors in the SHM system were used to perform the ambient vibration tests. These additional
sensors were installed on the other three corners on the 51st floor, 37th floor, and 21st floor. Adding the three locations
ZHANG ET AL. 5 of 20

FIGURE 2 Accelerometer installation

on these floors was possible because DAQ units are installed on all these floors, and the channels used for the measure-
ment of story drift angle could be used to measure acceleration responses. Therefore, on each floor, at most four loca-
tions were measured biaxially, providing eight measured DOFs. Because on the seventh floor, one DAQ unit was also
set, using the same method, the four corners of the concrete core (see Figure 4) on this floor were also measured. To
provide information for identifying the high‐order modes, some locations with only tilt sensors installed were selected,
that is, the locations on the 29th and 44th floors. Because some cables had been laid to connect the tilt sensors, these
cables were used to connect the accelerometers by installing a port fitting the new sensor. With the above arrangement,
19 locations were measured including four locations on the 51st floor, one location on the 44th floor, four locations on
the 37th floor, one location on the 29th floor, four locations on the 21st floor, four locations on the seventh floor, and
one location on the first floor, providing 38 DOFs.
At the time of measurement, only 14 uniaxial accelerometers were available, so multiple setups were conducted to
cover all the measured locations. The setup plan can be found in Table 3 and Figure 5. In this table, the measured loca-
tions are denoted by numbers. The first two numbers denote the floor number, whereas the last two numbers denote the
locations on this floor. For example, “3703”denotes the third location on the 37th floor, as shown in Figure 4; “4401”
and “2901” denote the Location 1 on the 44th and 29th floors, respectively. The reference location was selected on
the 51st floor. In Setup 1, the sensors were put on the B4, first, 21st, 37th, and 51st floors. The locations of this setup
were the same as the one in the SHM system with the purpose of testing the whole system simultaneously. In Setup
2, the sensors were moved to the 29th, 44th and 51st floors. For the first two floors of this setup, only the locations inside
the weak electricity room were measured, whereas on the 51st floor, the four corners of the concrete core were all
6 of 20 ZHANG ET AL.

FIGURE 3 Tilt sensors and data acquisition system

FIGURE 4 Measurement locations on a typical floor


ZHANG ET AL. 7 of 20

TABLE 3 Setup plan

Setup\channel 1, 2 3, 4 5, 6 7, 8 9, 10 11, 12 13, 14

1 5101 3701 2101 0101 / / /


2 5101 5102 5103 5104 4401 2901 /
3 5101 3701 3702 3703 3704 / /
4 5101 2101 2102 2103 2104 0701 /
5 5101 5103 3701 3703 2101 2104 0701
6 5101 3701 2101 0701 0702 0703 0704

FIGURE 5 Setup information

measured. In Setup 3, the sensors near the weak electricity room were kept unchanged, with eight sensors moved to the
37th floor to measure the four corners of the concrete core. In Setup 4, the sensor on the 51st floor was kept unchanged,
with the eight sensors on the 37th floor moved to the 21st floor to measure the four corners of the concrete core.
8 of 20 ZHANG ET AL.

When carrying out the measurement, the site was quite complicated due to the ongoing construction work. In some
channels, due to the noise or equipment problem, the quality of the data was not good. This can be checked by the
power spectral density (PSD) spectra, in which no peaks can be found. Therefore, to obtain effective data, these chan-
nels were remeasured. Furthermore, considering the fact that the reference locations should provide common informa-
tion for all the setups, two additional setups, that is, Setups 5 and 6 were carried out. In the fifth setup, two locations
were measured for each of the 21st, 37th and 51st floors. On the seventh floor, one location inside the weak electricity
room was measured. In the sixth setup, three sensors in the weak electricity rooms of the 21st, 37th and 51st floors were
unchanged, and the other eight sensors were used to measure the four corners of the concrete core on the seventh floor.
The overview of the measurement can be seen in Figure 6.
Because the natural frequencies of the modes of interest were all lower than 5 Hz, the sampling rate was set to
100 Hz. As the fundamental natural frequency was about 0.2 Hz, at least 30 min of data were measured in each setup
to provide enough information for the modal identification later. It should be noted that at the time of measurement,
the main structure had been completed, whereas some other work was still under operation. Therefore, noise was inev-
itable during the measurement. To reduce the disturbance from the environment, in total, 12 persons participated in the
field test. Each one was in charge of one sensor to prevent the sensors and cables from being touched. In each setup,
about 90 min were needed, including 30 min for measurement and 60 min for sensor moving. The measured locations
were blocked from each other by the walls inside the building, so common tools such as lasers and rulers could not be
used for the sensor alignment. For the convenience of sensor alignment, the sensors were set with the shear walls of the
concrete core used as the references.

5 | METHODOLOG Y

To analyze the measured data and determine the modal properties of this super‐tall building, ambient modal identifi-
cation was carried out using a fast Bayesian FFT method.31 On the basis of this method, the most probable values
(MPVs) of the modal parameters and the associated posterior uncertainty could be determined. The method is briefly
introduced as follows.31,37
Let θ = {f, ζ, S, Se, Φ} denote the set of modal parameters to be identified, including natural frequencies f, damping
ratios ζ, PSDs of modal forces S, PSD of prediction error Se, and mode shapes Φ. The collected acceleration response,
xb
€ j ∈ Rn , consists of x€ j ∈ Rn (theoretical structural acceleration) and ej ∈ Rn (prediction error):

xb
€ j ¼ x€ j þ ej ; (1)

where j = 1, 2, …, N; N is the number of sampling points and n is the number of measured DOFs. The FFT of xb
€ j is
defined as follows:

FIGURE 6 Overview of the


measurement
ZHANG ET AL. 9 of 20

rffiffiffiffiffiffiffiffi  
2Δt N b ðj − 1Þðk − 1Þ
Fk ¼ ∑ x€ j exp −2πi ; (2)
N j¼1 N

where k = 1, …, Nq with Nq = int[N/2] + 1 (int[.] denotes the integer part) being the frequency index corresponding to
the Nyquist frequency; Δt denotes the sampling interval and i2 = −1.
Let Zk = [Re F k; Im F k] ∈ R2n denote a vector including the real and imaginary part of F k. Let {Zk} denote the FFT
data in a selected frequency band containing the modes of interest, given which, based on Bayes' Theorem, the posterior
probability density function (PDF) of θ, p(θ|{Zk}), can be expressed as follows:

pðθjfZk gÞ ∝ pðθÞpðfZk gjθÞ; (3)

where p(θ) is the prior PDF of θ and p({Zk}|θ) is the likelihood function. It is assumed that the prior PDF is a uniform
PDF, so the posterior PDF p(θ| {Zk}) is proportional to the likelihood function p({Zk}| θ). The MPVs of the modal param-
eters θ can then be obtained by maximizing p({Zk}| θ).
It is proved that {Zk} at different frequencies is asymptotically independent and follow Gaussian distributions53,54
when the number of sampling pointes N is large and the sampling interval Δt is small. On the basis of this, p({Zk}| θ)
can be obtained:
 
‐n ‐1=2 1 T ‐1
pðfZk gjθÞ ¼ ∏ð2π Þ ðjCk jÞ exp ‐ Zk Ck Zk ; (4)
k 2

where
  " #
1 Φ ReHk − ImHk ΦT Se
Ck ¼ þ I2n ; (5)
2 Φ ImHk ReHk Φ T 2

denotes the covariance matrix of Zk and contains the modal parameters to be identified; I2n ∈ R2n denotes an identity
matrix; Hk ∈ Rm × m denotes a transfer matrix, and its (i, j) component is given by:

  −1 h 2
i−1
Hk ði; jÞ ¼ Sij β2ik −1 þ 2iζ i βik βjk −1 −2iζ j βjk ; (6)

where Sij is the cross spectral density between the ith and jth modal excitation; ζi is the damping ratio of the ith mode;
βik = f i/fk is a frequency ratio at frequency fk, and f i is the natural frequency of the ith mode. For the convenience of
optimization, the posterior PDF is written in terms of the negative log‐likelihood function (NLLF) L(θ):

pðθjfZk gÞ ∝ exp½−LðθÞ; (7)

where

1  
LðθÞ ¼ ∑k ln detCk ðθÞ þ Zk T Ck ðθÞ−1 Zk ; (8)
2
Therefore, maximizing the posterior PDF p(θ| {Zk}) is equivalent to minimizing the NLLF L(θ).
Theoretically, the MPVs of modal parameters can be determined by minimizing L(θ). However, there will be compu-
tational problems if the optimization is directly performed, because the number of uncertain parameters to be identified

FIGURE 7 Root power spectral density of the measured data in setup 1


10 of 20 ZHANG ET AL.

TABLE 4 Modal parameters identified in Setup 1

Mode 1 2 3 4 5 6 7 8 9 10

f (Hz) 0.242 0.258 0.413 0.863 0.959 1.214 1.557 2.047 2.199 2.362
ζ (%) 0.60 0.42 0.76 1.01 0.65 0.54 0.86 2.11 1.23 2.25
S* 4.0 0.5 0.7 1.2 0.6 0.5 1.0 0.3 1.5 0.5
S*e 10.6 20.3 4.6 6.4 8.1 5.2 5.3 9.1 8.6 26.0

Note. *: unit: 10−12g2/Hz.

TABLE 5 Posterior uncertainties (coefficient of variation) of modal parameters in Setup 1

Mode 1 2 3 4 5 6 7 8 9 10

f (%) 0.18 0.14 0.14 0.12 0.09 0.07 0.08 0.25 0.09 0.67
ζ (%) 31.3 35.3 21.0 13.3 14.4 13.2 9.6 12.2 9.6 26.3
S 21.4 22.3 13.4 9.5 8.4 6.4 5.3 17.7 9.2 53.3
Se (%) 6.2 6.4 4.0 2.6 2.5 2.0 1.6 2.0 2.1 2.7

FIGURE 8 Mode shapes identified:


Modes 1 and 2
ZHANG ET AL. 11 of 20

are usually large in practical applications. The problem mainly comes from mode shape identification because the num-
ber of measured DOFs is large (e.g., larger than 10). The brutal‐force optimization with a large number of parameters
will be very difficult to converge. Moreover, there is an ill‐condition problem in the optimization when calculating
the inverse of the covariance matrix Ck. To solve these two problems, a fast algorithm was developed. With this algo-
rithm, the modal parameters can be identified in a few seconds. The associated posterior uncertainty can also be deter-
mined analytically. The details of the fast algorithm can be found in previous studies.32-34

6 | DATA ANALYSIS

6.1 | Field test for the building

The dynamic characteristics were identified by the fast Bayesian FFT method based on the collected acceleration data.
Figure 7 shows the root PSD spectra of the measured data in Setup 1 from 0 to 3 Hz. Below 3 Hz, there are 10 potential
modes with obvious peaks. Because the vibration of one building is mainly dominated by the first several modes, in this
study, only the first 10 potential modes were identified. The other modes above 3 Hz may also be investigated in the
future studies.
Table 4 shows the modal parameters identified in Setup 1, including the natural frequency, damping ratio, PSD of
modal force, and PSD of prediction error. The fundamental natural frequency is about 0.242 Hz. This coincides with
the fact that the height of this super‐tall building is over 250 m. The natural frequencies of the first two modes are close,

FIGURE 9 Mode shapes identified:


Modes 3 and 4
12 of 20 ZHANG ET AL.

indicating that the discrepancy of the stiffness between the X and Y direction is not very large. The damping ratios of
these modes range from 0.4% to 2.3%. The damping ratios of the last three modes tend to be larger than those of the
other seven modes, showing a better energy dissipation capability. This is consistent with the spectra in Figure 7, where
the peaks of these three modes are wider than those of other modes. The PSDs of modal force and prediction error are
all in the order of 10−12g2/Hz. The first mode has the largest PSD of modal force, which is also consistent with the spec-
tra where the magnitude of the peak in this mode is the largest.
Table 5 shows the posterior uncertainties of the modal parameters in Setup 1 that are represented by coefficient of
variation (COV) (=posterior standard deviation/MPV). The COVs of natural frequencies are all less than 1%, indicating
the identification of this quantity is quite accurate. For the damping ratios and PSDs of modal forces, the COVs range
from a few percent to a few tens' percent, whereas the COV of PSD of prediction error is in the order of a few percent.
The accuracy of these modal parameters are all acceptable. It is interesting to see that from Modes 1 to 7, the posterior
uncertainties of all the modal parameters tend to decrease with the increase of the mode order. This is because for the
same data length in the time domain, it provides more effective data for higher modes with larger natural frequencies,
that is, given the same data length, a larger number of cycles of modal oscillation is contained in the data for higher
modes than for lower modes. The more the effective data, the smaller the uncertainty of the modal parameters. For
Modes 8 to 10, it is seen from the PSD spectra that they have smaller signal to noise ratios compared with other modes,
especially for Modes 8 and 10, which results in relatively larger uncertainties for them.
Next, the mode shapes of these 10 modes were investigated. Figures 8–12 show the mode shapes from Modes 1 to 10,
respectively. Each figure shows two modes. The black solid line and red dashed line denote the deformed and

FIGURE 10 Mode shapes identified:


Modes 5 and 6
ZHANG ET AL. 13 of 20

FIGURE 11 Mode shapes identified:


Modes 7 and 8

undeformed shapes, respectively. The measured locations are represented by dots. In the second line of the title of each
figure, the first value denotes the averaged natural frequency over different setups; the second value denotes the aver-
aged damping ratio and the two values inside the parentheses denote the averaged posterior COVs of the natural fre-
quency and damping ratio, respectively. Mode 1 in Figure 8 with the natural frequency 0.242 Hz and damping ratio
0.54% is the first translational mode in the X direction. The whole building moves along the weak direction of this build-
ing. Mode 2 is the first translational mode in the Y direction (strong direction of the building) with a natural frequency
0.257 Hz and a damping ratio 0.57%. The damping ratios of these two modes are similar and less than 0.6%, indicating
that the energy dissipation capability of these two modes in strong and weak directions is similar. The third mode is the
first torsional mode with a natural frequency 0.412 Hz and damping ratio 0.62%. The torsional center is located at the
center of the concrete core. The torsional movement of the higher floors is more obvious than that of the lower floors.
Modes 4 to 6 are the second translational modes in the X and Y directions and the torsional mode, with the natural
frequencies 0.859, 0.958, and 1.211 Hz, and damping ratios 0.79%, 0.69%, and 0.56%, respectively. It is interesting to
observe that for Modes 4 and 5, the natural frequency of the mode in the strong direction is smaller than that of the
mode in the weak direction. This is opposite to the observation for Modes 1 and 2. Although the modal stiffness corre-
sponding to the strong direction is larger than that corresponding to the weak direction, the modal mass corresponding
to the strong direction may be also larger than that corresponding to the weak direction. As a result, it may happen that
the natural frequency corresponding to the strong direction is smaller than that corresponding to the weak direction.
This is also reflected in Modes 7 to 10 in Figures 11, 12, respectively, where Modes 7 and 10 are two higher translational
modes in the X direction, whereas Modes 8 and 9 are two higher torsional modes. The modes in the Y direction may
appear after 3 Hz, but they were not identified in this study.
14 of 20 ZHANG ET AL.

FIGURE 12 Mode shapes identified:


Modes 9 and 10

Next, the variation of modal parameters across different setups was investigated. Figures 13 and 14 show the identi-
fied natural frequencies and damping ratios in all the six setups, where the dot denotes the MPV and the error bar
denotes the ±2 standard deviation. For the natural frequencies, the variation of the MPVs across different setups is
small, showing that this quantity is relatively stable. It is also interesting to find that for Setups 5 and 6, the natural fre-
quencies tend to be smaller than those of other setups in all the modes. Note that the first four setups were carried out
on January 4, 2017, whereas the Setups 5 and 6 were conducted 2 months later. On January 4, the temperature in
Shanghai ranged from 11°C to 15°C, whereas on March 4, the temperature ranged from 10°C to 18°C. In the daytime,
the temperature in March tended to be higher than that in January. Furthermore, during these 2 months, the decoration
work was in progress. The curtain wall and some electronic device were being installed. At the same time, furniture was
also moved into the offices. As a result, the mass of the building increased. These reasons may have resulted in the
decrease of the natural frequencies. It is also interesting to see that the posterior standard deviations of the frequencies
are smaller than the corresponding sample standard deviations calculated over different setups. Note that the posterior
standard deviation (the posterior uncertainty) is used to quantify the remaining uncertainty of the identified modal
parameters given the data in one setup, whereas the sample standard deviations over different setups reflect the change
of modal parameters in different setups. The former is mainly affected by the measurement noise and modeling error. In
addition to the measurement noise and modeling error, the latter is also affected by environmental changes.
For the damping ratios, although the variation of the MPVs across different setups is a little large, it is still acceptable.
For Modes 8 and 10, it is found that the posterior uncertainty in the first setup is obviously larger than those in other
setups. This phenomenon can be observed for both the natural frequencies and damping ratios, indicating that the qual-
ity of the data in this setup is not good enough for these two modes, but for other modes, the posterior uncertainties of
the modal parameters identified using this set of data are similar with those in other setups.
ZHANG ET AL. 15 of 20

FIGURE 13 Natural frequencies and


±2 standard deviation across different
setups

FIGURE 14 Damping ratios and ±2


standard deviation across different setups

6.2 | Comparison with other methods

Besides the fast Bayesian FFT method, two other methods, that is, EFDD and SSI, were also employed for OMA. Table 6
shows the modal parameters of the first six modes identified by EFDD, SSI, and Fast Bayesian FFT method. It is seen
16 of 20 ZHANG ET AL.

TABLE 6 Natural frequency and damping ratio identified by EFDD, SSI and fast Bayesian FFT method

EFDD SSI Fast Bayesian FFT


Modes Frequency (Hz) Damping ratio (%) Frequency (Hz) Damping ratio (%) Frequency (Hz) Damping ratio (%)

1 0.242 2.487 0.246 4.784 0.242 0.60


2 0.258 2.420 0.262 2.217 0.258 0.42
3 0.409 1.741 0.464 3.140 0.413 0.76
4 0.850 1.300 0.853 1.445 0.863 1.01
5 0.954 1.148 0.955 0.688 0.959 0.65
6 1.209 0.869 1.211 0.999 1.214 0.54

Abbreviations: EFDD, enhanced frequency domain decomposition; SSI, stochastic subspace identification.

that the natural frequencies are close among the three methods. For Mode 3, the natural frequencies obtained by the
Bayesian method and EFDD method are close to each other, whereas the natural frequency obtained by the SSI method
has a relatively larger discrepancy to those obtained by the former two methods. The difference of the damping ratios
among the three methods is evident. The damping ratios obtained by the Bayesian method are the smallest, the ones

FIGURE 15 Most probable values


(MPVs) and ±2 standard deviations across
different stages: Modes 1 and 2

FIGURE 16 Most probable values


(MPVs) and ±2 standard deviations across
different stages: Modes 3 and 4
ZHANG ET AL. 17 of 20

obtained by EFDD are in the middle whereas the ones obtained by SSI are the largest. This may be caused by the dif-
ferent mechanism used to identify the damping ratios in different methods. The implication is that the damping ratios
are uncertain and it is recommended to assess their uncertainties in the future modal identification. For super‐tall build-
ings, usually the damping ratios are small, say, less than or around 1% under ambient conditions. From this point of
view, the Bayesian method is more reasonable than the other two methods because both the MPVs and the posterior
uncertainties can reflect the possible values of the damping ratios well.
From the identification results, it is found that the fundamental natural period of the structure is 4.132 s. Note that
the structure is located in the area with soft foundation, which means that the site predominant period is long. On one
hand, long‐term seismic waves in soft ground will be enlarged. On the other hand, during the earthquake propagation
process short‐term components decay rapidly, whereas long‐term components decay slowly. This means that the struc-
ture may be vulnerable to resonance with the long‐term components of earthquake motion. Therefore, it is necessary to
monitor the structural response subjected to earthquakes.

6.3 | Long‐term monitoring


To monitor the variation of dynamic characteristics in different construction stages, long‐term monitoring was carried
out from March 2016 to June 2017. In this process, acceleration data was collected on March 18, 2016; December 21,
2016; January 4, 2017; March 4, 2017; and June 3, 2017. The structure was measured biaxially and over 30 min of data
were collected in each measurement.
Figures 15–17 show the modal parameters of the first five modes in different stages, where the MPVs are denoted by
dots and the ±2 standard deviations are denoted by the error bars. It is seen that the natural frequencies of the five
modes decrease with time, and they tend to be stable starting from the fourth stage. This phenomenon can be explained
as follows. Note that since the first stage, the main building had been completed. In the past 1 year, due to the equip-
ment installation and decoration work, the mass of the structure became greater and greater. Although the stiffness also

FIGURE 17 Most probable values (MPVs) and ±2 standard deviations across different stages: Mode 5
18 of 20 ZHANG ET AL.

increased in this process, the increase of stiffness was smaller than the increase of mass. As a result, the natural frequen-
cies decreased gradually in the past year. The uncertainty magnitudes of natural frequencies in different stages are con-
sistent with each other and retain a small value. For the damping ratio, which has a larger uncertainty, it was difficult to
observe an obvious trend for the five modes. The damping ratios are all less than or around 1%, and this is consistent
with the properties of super‐tall buildings. These changes of modal parameters in different stages reflect the healthy
condition of the structure from this point of view.

7 | CONCLUSIONS
This paper presents the work on the structural health monitoring of a 250‐m high‐rise building. In the SHM system, two
kinds of sensors were installed, including accelerometers and tilt sensors. Five substations were set up, and each one
provided eight channels that can measure two kinds of structural responses. The balance between the economy and
the efficiency to reflect the structural properties having been considered, that is, using the fewest sensors to achieve
the best test results, the locations of sensors were properly selected according to the structural form. On the basis of
the SHM system and some additional accelerometers, a series of ambient tests was carried out. Due to the limitation
of the sensor number, multiple setups were designed to measure all the modes of interest.
On the basis of a recently developed Bayesian method, the MPVs of the modal parameters and the associated poste-
rior uncertainties were obtained and investigated. The first 10 modes were identified including four translational modes
in the X direction (weak direction) of the structure, two translational modes in the Y direction (strong direction), and
four torsional modes. The fundamental natural frequency of this building is about 0.242 Hz (natural period 4.132 s).
Note that the structure is located in Shanghai, where the foundations are soft and the site predominant period is long.
Therefore, the structure is vulnerable to the long‐term components of earthquake motion. Monitoring the structural
response subjected to earthquakes is necessary.
The variation of modal parameters across different setups was investigated and it was found that the natural frequen-
cies have a small variation whereas the variation of the damping ratios is a bit larger. This is also consistent with the fact
that the damping ratios have large posterior uncertainties. The modal parameters were also identified by two other
methods, that is, EFDD and SSI. It was found that the natural frequencies obtained by different methods are consistent
whereas a large discrepancy can be observed for the damping ratios. Long‐term monitoring of this building was carried
out and lasted for about 15 months. On the basis of the collected data in different stages, it was found that although the
main structure had been completed, the natural frequencies still tended to decrease due to the change of structural
mass. No obvious trend could be observed for damping ratios during this long‐term monitoring.

ACK NO WLE DGE MEN TS


This paper is funded by China National Science and Technology Major Project “Development of Energy Efficiency
Technologies for Tall Buildings in Mega Cities Combined with Structural Safety Solutions under Multi‐hazard Impacts”
through grant numbered 2016YFE0105600, the National Natural Science Foundation of China through grant 51878484,
Interdisciplinary Cooperation Foundation for the Peak Discipline of Civil Engineering at Tongji University and a
research starting funding from Harbin Institute of Technology, Shenzhen. The financial supports are gratefully
acknowledged. The authors also would like to thank the help and support of the building owner, Guanfeng (Shanghai)
Real Estate Development Co., Ltd., to support for the SHM program and the graduate students, Jixing Cao, Jiawei Chen,
Chao Guo, Hao Fu, Minghui Zhou, Zhifang Wang, Kai Sui, Xiangxiang Dou, Jie Wang, and Lu Ouyang from Tongji
University for participating in the test.

ORCID
Feng‐Liang Zhang https://orcid.org/0000-0001-6894-5842

R EF E RE N C E S
1. Yuen KV, Kuok SC. Ambient interference in long‐term monitoring of buildings. Eng Struct. 2010;32(8):2379‐2386.
2. Ng CT. Application of Bayesian designed artificial neural networks in phase II structural health monitoring benchmark studies. Aust J
Struct Eng. 2014;15(1):1‐10.
ZHANG ET AL. 19 of 20

3. Brincker R and Ventura CE. Introduction to operational modal analysis. John Wiley & Sons, Ltd, Chichester, UK. https://doi.org/10.1002/
9781118535141.ch11, 2015.
4. Ren WX, Chen HB. Finite element model updating in structural dynamics by using the response surface method. Eng Struct.
2010;32(8):2455‐2465.
5. Huang Y, Beck JL, Li H. Hierarchical sparse Bayesian learning for structural damage detection: theory, computation and application.
Struct Saf. 2017;64:37‐53.
6. Zhu HP, Mao L, Weng S. A sensitivity‐based structural damage identification method with unknown input excitation using transmissi-
bility concept. J Sound Vib. 2014;333(26):7135‐7150.
7. Lam HF, Peng H, Au SK. Development of a practical algorithm for Bayesian model updating of a coupled slab system utilizing field test
data. Eng Struct. 2014;79:182‐194.
8. Au SK, Zhang FL. Fundamental two‐stage formulation for Bayesian system identification, part I: general theory. Mech Syst Signal Process.
2016;66:31‐42.
9. Zhang FL, Au SK. Fundamental two‐stage formulation for Bayesian system identification, part II: application to ambient vibration data.
Mech Syst Signal Process. 2016;66:43‐61.
10. Yuen KV, Mu HQ. Real‐time system identification: an algorithm for simultaneous model class selection and parametric identification.
Comput‐Aided Civ Inf. 2015;30(10):785‐801.
11. Beck JL, Yuen KV. Model selection using response measurements: Bayesian probabilistic approach. J Eng Mech. 2004;130(2):192‐203.
12. Zhang FL, Ni YC, Lam HF. Bayesian structural model updating using ambient vibration data collected by multiple setups. Struct Control
Hlth. 2017;24(12):e2023. https://doi.org/10.1002/stc.2023
13. Cao L, Li C. Tuned tandem mass dampers‐inerters with broadband high effectiveness for structures under white noise base excitations.
Struct Control Hlth. 2019;26(4):e2319. https://doi.org/10.1002/stc.2319
14. Ni YC, Lu XL, Lu WS. Field dynamic test and Bayesian modal identification of a special structure—the palms together Dagoba. Struct
Control Hlth. 2016;23(5):838‐856.
15. Ni YC, Lu XL, Lu WS. Operational modal analysis of a high‐rise multi‐function building with dampers by a Bayesian approach. Mech Syst
Signal Process. 2017;86:286‐307.
16. Kuok SC, Yuen KV. Structural health monitoring of Canton tower using Bayesian framework. Smart Struct Syst. 2012;4:375‐391.
17. Brownjohn JMW, Carden EP, Goddard CR, Oudin G. Real‐time performance monitoring of tuned mass damper system for a 183 m rein-
forced concrete chimney. J Wind Eng Ind Aerod. 2010;98(3):169‐179.
18. Brownjohn JMW. Ambient vibration studies for system identification of tall buildings. Earthq Eng Struct D. 2003;32(1):71‐95.
19. Sohn H, Czarnecki J, Farrar C. Structural health monitoring using statistical process control. J Struct Eng. 2000;126(11):1356‐1363.
20. Brownjohn JMW, Pan TC. Identifying loading and response mechanisms from ten years of performance monitoring of a tall building. J
Perform Constr Fac. 2008;22(1):24‐34.
21. Ni YQ, Xia Y, Liao WX, Ko JM. Technology innovation in developing the structural health monitoring system for Guangzhou new TV
tower. Struct Control Hlth. 2009;16(1):73‐98.
22. Lam HF, Hu J, Yang JH. Bayesian operational modal analysis and Markov chain Monte Carlo‐based model updating of a factory building.
J Eng Struct. 2017;132:314‐336.
23. Gentile C, Saisi A. Ambient vibration testing of historic masonry towers for structural identification and damage assessment. Construct
Build Mater. 2007;21(6):1311‐1321.
24. Loh CH, Weng JH, Chen CH, Lu KC. System identification of mid‐story isolation building using both ambient and earthquake response
data. Struct Control Hlth. 2013;20(2):139‐155.
25. Yan BF, Miyamoto A, Bruhwiler E. Wavelet transform‐based modal parameter identification considering uncertainty. J Sound Vib.
2006;291(1‐2):285‐301.
26. Ciloglu K, Zhou Y, Moon F. Impacts of epistemic uncertainty in operational modal analysis. J Eng Mech. 2012;138(9):1059‐1070.
27. Döhler M, Lam XB, Mevel L. Uncertainty quantification for modal parameters from stochastic subspace identification on multi‐setup
measurements. Mech Syst Signal Process. 2013;36(2):562‐581.
28. Vu VH, Thomas M. Uncertainties on modal parameters by operational modal analysis. Mechanics & Industry. 2014;15(2):153‐158.
29. Papadimitriou C, Beck JL, Katafygiotis LS. Updating robust reliability using structural test data. Probab Eng Mech. 2001;16(2):103‐113.
30. Beck JL. Bayesian system identification based on probability logic. Struct Control Hlth. 2010;17(7):825‐847.
31. Au SK, Zhang FL, Ni YC. Bayesian operational modal analysis: theory, computation, practice. Comput Struct. 2013;126:3‐14.
32. Au SK. Fast Bayesian FFT method for ambient modal identification with separated modes. ASCE J Eng Mech. 2011;137(3):214‐226.
33. Zhang FL, Au SK. Erratum for fast Bayesian FFT method for ambient modal identification with separated modes by Siu‐Kui Au. ASCE J
Eng Mech. 2013;139(4):545‐545.
34. Au SK. Fast Bayesian ambient modal identification in the frequency domain, part II: posterior uncertainty. Mech Syst Signal Process.
2012;26:76‐90.
20 of 20 ZHANG ET AL.

35. Au SK, Zhang FL. Fast Bayesian ambient modal identification incorporating multiple setups. ASCE J Eng Mech. 2012;138(7):800‐815.
36. Zhang FL, Au SK, Lam HF. Assessing uncertainty in operational modal analysis incorporating multiple setups using a Bayesian approach.
Struct Control Hlth. 2015;22(3):395‐416.
37. Lam HF, Zhang FL, Ni YC, Hu J. Operational modal identification of a boat‐shaped building by a Bayesian approach. Eng Struct.
2017;138:381‐393.
38. Lam HF, Yang J, Au SK. Bayesian model updating of a coupled‐slab system using field test data utilizing an enhanced Markov chain
Monte Carlo simulation algorithm. Eng Struct. 2015;102:144‐155.
39. Au SK, Zhang FL, To P. Field observations on modal properties of two super tall buildings under strong wind. J Wind Eng Ind Aerod.
2012;101:12‐23.
40. Au SK, Zhang FL. Ambient modal identification of a primary‐secondary structure by fast Bayesian FFT method. Mech Syst Signal Process.
2012;28:280‐296.
41. Zhang FL, Ni YQ, Ni YC, Wang YW. Operational modal analysis of Canton tower by a fast frequency domain Bayesian method. Smart
Struct Syst. 2016;17(2):209‐230.
42. Zhang FL, Ventura CE, Xiong HB, Lu WS, Pan YX, Cao JX. Evaluation of the dynamic characteristics of a super tall building using data
from ambient vibration and shake table tests by a Bayesian approach. Struct Control Hlth. 2018;25(2):e2121.
43. Zhang FL, Ni YC, Au SK, Lam HF. Fast Bayesian approach for modal identification using free vibration data, part I–Most probable value.
Mech Syst Signal Process. 2016;70:209‐220.
44. Ni YC, Zhang FL, Lam HF, Au SK. Fast Bayesian approach for modal identification using free vibration data, part II—posterior uncer-
tainty and application. Mech Syst Signal Process. 2016;70:221‐244.
45. Au SK, Ni YC. Fast Bayesian modal identification of structures using known single‐input forced vibration data. Struct Control Hlth.
2014;21(3):381‐402.
46. Ni YC, Zhang FL. Fast Bayesian approach for modal identification using forced vibration data considering the ambient effect. Mech Syst
Signal Process. 2018;105:113‐128.
47. Ni YC, Zhang FL. Fast Bayesian approach for modal identification from seismic response data. Comput Struct. 2019;212:225‐235.
48. Aktan AE, Catbas FN, Grimmelsman KA, Tsikos CJ. Issues in infrastructure health monitoring for management. J Eng Mech.
2000;126(7):711‐724.
49. Yang JN, Lei Y, Lin S, Huang N. Identification of natural frequencies and dampings of in situ tall buildings using ambient wind vibration
data. J Eng Mech. 2004;130(5):570‐577.
50. Chen WH, Lu ZR, Lin W, et al. Theoretical and experimental modal analysis of the Guangzhou new TV tower. Eng Struct.
2011;33(12):3628‐3646.
51. Kijewski‐Correa T, Kwon D, Kareem A, et al. Smartsync: an integrated real‐time structural health monitoring and structural identifica-
tion system for tall buildings. J Sruct Eng. 2013;139(10):1675‐1687.
52. Zhang FL, Xiong HB, Shi WX, Ou X. Structural health monitoring of Shanghai tower during different stages using a Bayesian approach.
Struct Control Hlth. 2016;23(11):1366‐1384.
53. Yuen KV, Katafygiotis LS, Beck JL. Spectral density estimation of stochastic vector processes. Probab Eng Mech. 2002;17(3):265‐272.
54. Schoukens J, Pintelon R. Identification of linear systems: A practical guideline for accurate modelling. London: Pergamon Press; 1991.

How to cite this article: Zhang F‐L, Yang Y‐P, Xiong H‐B, Yang J‐H, Yu Z. Structural health monitoring of a
250‐m super‐tall building and operational modal analysis using the fast Bayesian FFT method. Struct Control
Health Monit. 2019;26:e2383. https://doi.org/10.1002/stc.2383

You might also like