You are on page 1of 25

Received: 18 May 2021 Revised: 16 February 2022 Accepted: 26 February 2022

DOI: 10.1002/stc.2957

RESEARCH ARTICLE

Modal analysis and tension estimation of stay cables using


noncontact vision-based motion magnification method

Samten Wangchuk1 | Dionysius M. Siringoringo2 | Yozo Fujino2

1
Department of Urban Innovation,
Yokohama National University,
Summary
Yokohama, Japan This paper describes a study on modal analysis and tension estimation of stay
2
Institute of Advanced Sciences, cables by vision-based measurement under ambient condition. Microvibration
Yokohama National University,
of stay cable is captured by video camera, and the amplitude is amplified using
Yokohama, Japan
phase-based video motion magnification method. From the sequences of cable
Correspondence images, spatial displacements of the cable are extracted via discretized centroid
Dionysius M. Siringoringo, IAS Assoc.
Professor Institute of Advanced Sciences,
searching method. Modal parameters of the cable, namely, natural frequency,
Yokohama National University, 79-1 damping ratio, and mode shape, are identified by dynamic mode decomposi-
Tokiwadai, Hodogaya-ku, Yokohama, tion method from cable displacement responses. Furthermore, tension of the
Kanagawa 240-8501, Japan.
Email: dion@ynu.ac.jp
stay cable is estimated based on the identified natural frequencies by minimiz-
ing an error function of an approximate relationship between frequency and
Funding information
tension iteratively. The technique is verified in the laboratory-scale experi-
Ministry of Education, Culture, Sports,
Science and Technology (MEXT), ments and implemented to full-scale measurement of medium size and long-
Government of Japan span cable-stayed bridges. To compare the accuracy and efficacy of the vision-
based method, noncontact vibration measurement using laser Doppler
vibrometer was also conducted. The results demonstrate that the proposed
vision-based vibration measurement techniques can estimate modal parame-
ters and tension of the stay cables accurately in a noncontact and distant mea-
surement under ambient condition. The proposed method offers an alternative
of effective and accurate vibration measurement of stay cables using only
motion of the cables recorded by video camera.

KEYWORDS
cable tension estimation, dynamic mode decomposition, image processing, modal analysis,
motion magnification, vision-based vibration measurement

1 | INTRODUCTION

Stay cable is an integral and important structural component of a cable-stayed bridge. Structural assessment of existing
stay cables is essential to ensure that their conditions are within the acceptable limits for safety and serviceability. Ten-
sion estimation of in-service stay cables is among the primary concerns of structural health monitoring of cable-stayed
bridge.1 Currently, the available techniques for cable monitoring can be broadly classified into three categories: (1) static
and direct method that directly measures the tension using pre-installed load cells, fiber-optics,2 or by the lift-off
hydraulic jacks; (2) the magnetic method based on the magnetic permeability measurement; and (3) the dynamic
method using vibration measurement.3 In the first category, the pre-installed calibrated load cell is used to monitor the
tension. The method, however, requires the load cell to have lifespan comparable with that of the cables for a

Struct Control Health Monit. 2022;e2957. wileyonlinelibrary.com/journal/stc © 2022 John Wiley & Sons, Ltd. 1 of 25
https://doi.org/10.1002/stc.2957
2 of 25 WANGCHUK ET AL.

permanent deployment. Moreover, the initial cost of load cells with such lifetime requirement might be prohibitively
expensive. The lift-off technique using hydraulic jack is also difficult to implement since it requires removal of parts of
the anchorage system, which is very expensive for larger cables because of the high strength needed. The second tech-
nique involves magnetic-based method developed for inspection of stress and nondestructive test. Electromagnetic tech-
nology is used to monitor forces in the prestressed structures and stay cables, while the magnetic-flux leakage method
was applied to diagnose the cuts, corrosions, and compression damages of wires inside the cables.4 The magnetic-based
techniques, however, require specific measurement device such as robot, extensive investment, and special expertise to
be implemented in the monitoring and diagnostic of stay cables.
Dynamic method by vibration measurement is the most widely used method for stay cable monitoring. Cable
tension can be estimated using dynamic parameters obtained from vibration measurement. The essential part of the
method is identification of cable natural frequency because it significantly affects the precision of cable force
estimation. The conventional vibration measurement usually utilizes accelerometer mounted on the cable surface and
vibrating the cable to obtain its natural frequencies. The technique has been widely implemented in the measurements
of stay cables of large cable-stayed bridges for examples by.3,5–7 This setup requires sensors to be wired to external
devices, which is costly for long-span cable-stayed bridges with numerous cables and time-consuming because of
sensors installation and data acquisition system. Recent developments in wireless sensing have enabled more efficient
measurement system without complicated cabling and installation issues. The wireless sensor network has provided a
portable low-cost measurement system that is more practical than the wired system.8–10 However, the contact-type
vibration measurement still requires physical access to the structure which is often difficult or in some cases even
impossible.
Noncontact vibration measurement system has been implemented on stay cables as an alternative to contact accel-
erometer. Notable examples of such system are the laser Doppler vibrometer (LDV)11,12 and the radar system.13,14 These
noncontact measurement devices, while enabling high accuracy measurement system, are relatively expensive and
require special skill to operate. Compared with traditional contact displacement, vision-based measurement system
offers more advantages in terms of instrumentation cost, efforts, capacity of frequency range, and spatial resolution.15
Furthermore, vibration measurement by video camera is less expensive, easier to operate, and deployable for frequent
measurement compared with the abovementioned noncontact measurement systems. Full-field structural dynamic
response of structures can be obtained accurately from video camera measurements with advanced image-processing
algorithms.16–19 Moreover, difficulty in capturing microvibration of civil structures in ambient condition can now be
overcome by recent advancements of computer vision such as using the phase-based video motion magnification
(PVMM) technique.20 By this method, microvibration of an object is amplified within the frequency of interest, and the
spatial displacement response can be extracted using available advanced image-processing techniques such as centroid
detection,21,22 edge detection,23 subpixel orientation code matching,24 principal component analysis, and blind source
separation.25–28
Various vision-based methods for stay cable measurement have been proposed recently such as videogrammetric
technique,29 remotely controllable pan-tilt drive,16 handheld shooting by smartphones,30 and unmanned aerial vehi-
cle.31 These studies have demonstrated feasibility of the vision-based measurement for effective modal analysis and ten-
sion estimation of stay cable. As for the image-processing tools used in the process to obtain vibration responses of the
cables, several techniques such as edge detection32 and subpixel orientation code matching24 have been proposed. In
this study, we describe an implementation of vision-based motion magnification technique for noncontact modal analy-
sis of stay cables on a cable-stayed bridge. The objective of the study is to realize a vision-based measurement technique
for modal analysis and tension estimation of stay cables under ambient condition by implementing the PVMM to stay
cable's microvibration. To construct vibration response from the sequences of cable images, we propose the centroid
detection method because it can effectively trace the movement of discretized cable elements. With this technique,
modal parameters are estimated using video measurement of stay cable under ambient condition. The modal parame-
ters include natural frequency, damping ratio, and mode shapes. Using information of natural frequencies, tension of
the cable is estimated by minimizing an error function in an iterative algorithm. Information on other modal parame-
ters, namely, damping and mode shapes, which are not included in the tension estimation, can be utilized for structural
assessment of the stay cables. Damping information, for example, can be used to evaluate effectiveness of stay cable's
damper system, while mode shapes can be used as an indicator for damage detection of the cable using mode shape-
based or curvature-based damage detection method.
In this study, a measurement system that utilizes a video camera as an alternative for a simpler, cost-effective, and
convenient vibration measurement method for stay cables' tension estimation of cable-stayed bridge under
WANGCHUK ET AL. 3 of 25

microvibration ambient condition is proposed. The main contribution and objective of the study are on the application
of the PVMM method for stay cables tension estimation of medium- and large-scale cable-stayed bridges. The objective
is achieved by implementing the following procedure: identification of the stay cable's spatial displacement responses
from video measurement, modal parameters identification using the amplified microvibration of stay cables obtained
from video measurement, and stay cable tension forces estimation using identfied natural frequencies in a normal non-
contact operational condition.
The paper is organized as follows. Section 2 describes briefly methodologies used in the study to obtain modal
parameters of the stay cable from video measurement, namely, PVMM, discretized centroid searching method (DCSM)
and dynamic mode decomposition (DMD). This is followed by the methodology to estimate cable tension force using
relationship between cable tension and the natural frequency. Section 3 presents the laboratory experiments to demon-
strate the feasibility and verify accuracy of the methods. Section 4 describes full-scale implementations of the method
on a pedestrian cable-stayed bridge and long-span cable-stayed bridge. Finally, important findings and conclusions are
summarized in Section 5.

2 | METHODOLOGIES

2.1 | Vibration measurement by PVMM

In the following section, a brief description of the method based on the work by Wadhwa et al.20 is provided. Video
recordings of a vibrating object consist of temporally displaced frames with image intensity, I ðx þ δðx,t ÞÞ, where x is the
pixel coordinate, and δðx,t Þ is the spatially and temporally varying local motion. The local phase and amplitude can be
extracted by employing the complex steerable pyramid filters such as the complex Gabor-type filter with sinusoids win-
dow using Gaussian envelope.33 With this, summation of the single frequency of complex sinusoids (ω) representing
the displaced frames can be obtained as follows:

X∞ X∞
i2πωðxþδðx,t ÞÞ
I ðx þ δðx, t ÞÞ ¼ B ðx, t Þ ¼
ω¼∞ ω
P e
ω¼∞ ω
ð1Þ

In Equation 1 Bω ðx,t Þ represents the sinusoid sub band, while ωðx þ δðx, t ÞÞ denotes the phase that contains motion
information. One can amplify the motion by altering the phase as in the Fourier shift theorem using magnification fac-
tor α as I ðx þ ð1 þ αÞδðx, t ÞÞ, which results in

eω ðx, t Þ ¼ Bω ðx, t Þeiα2πωδðx,tÞ ¼ Pω ei2πωðxþð1þαÞδðx,tÞÞ


B ð2Þ

In the above equation, Beω ðx, t Þ is a complex sinusoid that has ð1 þ αÞ scale of the magnified motion. It should be noted
that introduction of amplification factor α influences noise characteristics of the image. Amplitude-weighted spatial
Gaussian blur on the phases is introduced to deal with this problem. Assuming that the response of noise-contaminated
image for a particular scale and orientation I þ σ n n can be written as follows:

N ω ¼ Bω ðx, t Þ þ σ n M ω ðx, t Þ ð3Þ

where M ω ðx,t Þ is the response of n to the complex steerable pyramid filter ω, then the response due to magnification is
shifted by eiα2πωδðx,tÞ . As the result, the motion-magnified band becomes:

e ω ðx, t Þ ¼ B
N eω ðx,t Þ þ σ n eiα2πωδðx,tÞ Mω ðx, t Þ ð4Þ

This phase shift corresponds to the translation of noise that could result in incorrect magnification. To deal with this
problem, an amplitude-weighted spatial Gaussian kernel and 2-D Gaussian smoothing kernel with standard deviation
4 of 25 WANGCHUK ET AL.

specified by sigma σ can be utilized.20 Finally, one can obtain the video of amplified motion of the object by combining
all phases of the magnified frames.

2.2 | Spatial displacement identification by DCSM

In a flexible structure like stay cable, total response of the structure is a summation of multimode modal displacements
where many modes have relatively equal modal contribution. Therefore, spatial displacements information on numer-
ous locations along the cable is essential for identification of modal parameters. To obtain the spatial displacement
time-history along the stay cable, an identification method named discretized centroid searching method (DCSM) is
adopted in this study. In this method, the cable image is discretized into smaller grids, whose centroids movements are
traced and then later reconstructed to give information on the spatial displacement time-histories of the discretized
cable structure. The phase magnified video explained in the previous section consists of four-dimensional array frames
of still images, denoted as a -by- b -by- 3 -by- k, where k represents the instantaneous time. Each still image frame is
divided into multigrid of cropped regions. In each grid, centroid of the discretized object C xy(t) is computed using the
Otsu Threshold Segmentation method (OTSM).21 The centroid information is retained for the consecutive frame t, t
+ 1, until the last frame. Coordinates of the centroids represent movement of the discretized cable, and they are inde-
pendently extracted to obtain time-histories of the object's spatial displacements as illustrated by Figure 1.
Consider the pixels in a grid of a still image frame represented by L gray levels ½1, 2,…L, where ni denotes the num-
ber of pixels at level i and N, ðN ¼ n1 þ n2 þ …nL Þ is the total number of pixels. After normalizing the gray-level histo-
PL
gram as a probability distribution (i.e., pi ¼ ni =N, pi ≥ 0, i¼1 pi ¼ 1), one can divide the pixels into two classes C0 and
C1 (background and objects, or vice versa) by setting a threshold at level k so that C0 denotes the pixels with levels
[1,2.., k], and C1 denotes the pixels with levels ½k þ 1, …L. A criterion that measures the separability of threshold at level
k can be defined as a function of probabilities of class occurrence and the class mean levels, defined as follows:

ηðk Þ ¼ σ 2B ðk Þ=σ 2T ð5Þ

where σ B and σ T are defined as follows:

½μT Ωðk Þ  μðk Þ2


σ 2B ðk Þ ¼ ð6Þ
Ωðk Þ½1  Ωðk Þ

XL
σ 2T ðk Þ ¼ i¼1
ði  μ T Þ2 p i ð7Þ

FIGURE 1 Schematic figure showing the process of tracing centroid movement of vibrating cable
WANGCHUK ET AL. 5 of 25

The quantities Ωðk Þ and μðk ) are the probabilities of class occurrence and class means, respectively. The threshold value
was selected using Equations 5–7 to minimize intraclass variance of the threshold of black and white pixels. Binary
images are created from two-dimensional grayscale images by setting all values above the threshold to ones and the
values below the threshold to zeros. In this step, bin of image histogram is employed to compute the threshold. Next,
the original frame of R rows and C columns is divided into n rows and m of columns of grids. The binary image num-
bers corresponding to the i th row and the j th column can be defined as follows:

Sði, jÞ ¼ binðði  1ÞR=n þ 1 : iR=n, ð j  1ÞC=m þ 1 : jC=mÞ ð8Þ

where bin denotes the binary image matrix. Note that the original number of rows and columns are extracted from the
magnified binary image. By employing this method, centroids of all grids in both x and y directions are obtained in the
pixel resolution and later transformed to the physical coordinate. Spatial movement of the cable is obtained by combin-
ing the centroids sequentially over the length of the video. It is important to note that a better quality of displacement
along the cable can be obtained by dividing the frame to larger number of grids thus providing finer resolution of spa-
tial displacement.

2.3 | Extraction of modal parameters from displacement responses

The cable displacements obtained by DCSM consist of multimode responses with spatial properties along the cable. To
extract modal parameters from the cable spatial displacements, multi-output system identification is needed, and in this
paper, the Dynamic Mode Decomposition (DMD) is adopted. The DMD was originally used as a low-rank diagnostic
tool for decomposing fluid flow data into dominant spatiotemporal modes to provide a valuable interpretation of coher-
ent structures in complex systems.34 The DMD can be considered as the family of subspace identification techniques. It
has been benchmarked and compared with several other leading techniques and was found robust and versatile in
dimensionality reduction and in reduced order modeling.35,36 In this study, the DMD is adopted for modal identifica-
tion of the cable using the spatial displacement responses obtained by the DCSM. Note that outputs of DCSM are the
magnified displacement responses at a specific frequency range ωi . They are expressed as zωi ðχ, tÞ, where χ denotes the
spatial coordinate of the cable χ ¼ 1, 2…, m and t is the time notation. Assuming that a sequence of images can be
divided by equal time interval Δt, the continuous time variable can be represented as a discrete variable k, where
k ¼ 0, 1, 2…, ðp þ qÞ. Therefore, the time-history of magnified spatial deformation can be rewritten in a matrix form as
follows:

Z ωi ðk Þ ¼ ½zωi ð1,k Þ zωi ð2, kÞ… zωi ðm,k ÞT ð9Þ

One can construct an input matrix X that consists of consecutive time-sampled data with the size of p and q, where
p and q are the user-defined numbers selected so that matrix X is full rank and has enough amount of data.
2 3
Z ωi ð0Þ Z ωi ð1Þ … Z ωi ðq  1Þ
6 Z ωi ð1Þ Z ωi ð2Þ … Z ωi ðqÞ 7
6 7
X ≜6
6 .. .. .. ..
7
7 ð10Þ
4 . . . . 5
Z ωi ðp  1Þ Z ωi ðpÞ … Z ωi ðp þ q  2Þ

Similarly, a new time-shifted block matrix Y of the same size is constructed using the shifted time displacements data.
2 3
Z ωi ð1Þ Z ωi ð2Þ … Z ω i ð qÞ
6 Z ω ð2Þ Z ωi ð3Þ … Z ωi ðq þ 1Þ 7
6 i 7
Y ≜6
6 .. .. .. ..
7
7 ð11Þ
4 . . . . 5
Z ωi ðpÞ Z ωi ðp þ 1Þ … Z ωi ðp þ qÞ
6 of 25 WANGCHUK ET AL.

The procedure for system matrix realization starts by taking the singular value decomposition (SVD) of block matrix X,
such that X ¼ UΣV  , where U's size is pm  n, Σ is a diagonal matrix with the size n  n, V is a q  n matrix, and *
denotes the conjugate transpose. Here, n is the rank of the reduced SVD approximation of X. Next, matrix A e that con-
sists of low-dimensional linear model of the dynamical system is defined as follows:

e ≜ U  YV Σ1
A ð12Þ

e are computed as AW
Then, the eigenvalues and eigenvectors of matrix A e ¼ W Λ, where the columns of matrix W are
the eigenvectors and the diagonal matrix Λ contains the corresponding eigenvalues λi . The eigen-decomposition of
e can be reconstructed from matrix W and Λ by the following equation:
matrix A

Φ ¼ YV Σ1 W ð13Þ

Note that the eigenvectors of matrix A e are the dynamic decomposition mode given by the columns of Φ; (ϕi ) and the
eigenvalues λi are expressed in z-domain. They are related to the modal characteristics by the following trans-
formation: λi ¼ ln e
λi =Δt. Finally, using the following relationship, the natural frequency (ωi ) and damping ratio (ξi ) of
the cable can be obtained as follows:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Reðλi Þ
ω0i ¼ Reðλi Þ2 þ Imðλi Þ2 , ξi ¼ ð14Þ
ω0i

2.4 | Tension estimation using frequency information

Vibration-based tension estimation of stay cable depends on the relationship between the natural frequencies and the
tension forces of the cables. In the literature, there are numerous formulations available on this relationship, which
were derived based on the taut string theory with consideration of sag-extensibility and/or bending stiffness. They can
be broadly divided into four categories. The first category is based on the flat taut string theory without considering sag-
extensibility and bending stiffness of the cable.37 The second category considers the bending effect but not the sag-
extensibility,38,39 the third category considers sag-extensibility but not the bending stiffness,40 and the fourth category
considers both sag-extensibility and bending stiffness.41 While tension forces are normally considered constant during
measurement, there is also an attempt to identify the time-varying cable tension forces based on time-frequency analy-
sis of the acceleration.42,43
To obtain cable tension experimentally, more information in addition to the natural frequencies are normally
required. The additional information varies depending on categories of the formulation, but mostly, it includes the sag-
to-span ratio, flexural stiffness, axial stiffness, mass density, and effective length of the cable. In practice, however, this
information is not always available. Sag-to-span ratio requires measurement of the cable static shape, which is not
always possible for frequent measurement. The flexural stiffness information is not easily available and in most cases is
difficult to measure since it may vary along the cable. In practice, perhaps only the information of effective length, mass
density, and sectional area of the cables can be measured or available in addition to the natural frequencies. Therefore,
in selecting the method for stay cable tension force estimation, it is important to consider the complexity and availabil-
ity of the cable information. In this study, we adopt an approximate relationship proposed by Mehrabi and Tabatabai,44
which was developed using the finite-difference cable model by considering the cable bending stiffness and sag extensi-
bility. The k th natural frequency of the cable ωk is defined as follows:
8 2   39
< 4 þ k 2π θ =
2 2

2
ωk ¼ kω1s ð1 þ aθÞ41 þ þ þb 5 ð15Þ
: ξ ξ2 ξ ;

where k is the mode number, a ¼ 0:039, and b ¼ 0:24. θ is a nondimensional parameter defined as follows:
WANGCHUK ET AL. 7 of 25

ðmgL=T Þ2 LEA
θ¼   ð16Þ
Þ2
TL 1 þ ðmgL=T
8

for the first in-plane mode (k = 1) and is zero for the other higher modes. The quantities ω1s and ξ are the idealized first
angular frequency of a taut string and the nondimensional parameter of bending stiffness, respectively. They are
defined as follows:
rffiffiffiffi
π T
ω1s ¼ ð17Þ
L m
rffiffiffiffiffi
T
ξ¼L ð18Þ
EI

Note that in the above equations (Equations 15–18), T is the stay cable tension force, m is the mass per unit length, L is
the effective length of the cable, g is the gravitational constant, and E is the Young's modulus, while I and A are the
cable's moment of inertia and cross-sectional area, respectively.
For natural frequencies of the modes higher than the first one, Equation 15 can be simplified to the following:

ωk ¼ kβk ω1s ð19Þ

where βk is a parameter relating the k th natural frequency with the idealized first angular frequency of a taut string;
and it is defined as follows:
   
2 k2 π 2 1
βk ¼ 1 þ þ 4 þ ð20Þ
ξ 2 ξ2

The similar approximate relationship was proposed by Morse and Ingard,45 Geier et al.,6 and Zarbaf et al.46 Further-
more, an error function E k ðω1s ,ξÞ, which is a summation of the squared of discrepancies between the natural frequen-
cies computed by Equation 19 and the ones obtained experimentally for p number of modes, can be defined as follows:
Xp
2
E k ðω1s , ξÞ ¼ k¼1
ωk ðω1s , ξÞ  ωexp
k ð21Þ

The estimated ω1s and ξ values can be obtained numerically by solving the partial derivation of error function E k ðω1s , ξÞ
∂E k ðω1s ,ξÞ ∂E k ðω1s ,ξÞ
with respect to the two variables ω1s and ξ, that is, ∂ω1s ¼0 and ∂ξ ¼ 0. By minimizing the error function using
iterative algorithm such as Newton–Raphson, the ω1s can be obtained, from which the cable tension T is estimated. It
should be mentioned that as per Equation 17, the accuracy of cable tension estimation requires accurate information of
the mass per unit length and the effective free length of the cable. The two quantities are normally known and available
from design drawing with relatively high accuracy.

2.5 | Main procedure of analysis

Figure 2 illustrates the four main steps employed for modal parameters identification and tension estimation of stay
cable using noncontact vision-based measurement. They can be described as follows.

• Step 1: Video measurement and implementation of PVMM. Vibration of stay cable is recorded by a video camera for a
certain period of time using a specified frame-per-second measurement sampling rate. The measurement is
8 of 25 WANGCHUK ET AL.

FIGURE 2 Flowchart of four main procedures for implementation of phase-based motion magnification for tension force estimation of
stay cable

performed in the direction perpendicular to the cable's vibration plane. After conducting Fourier transformation and
multiscale image decomposition, the phase-based magnified motion of the cable is obtained.
• Step 2: Displacement extraction by DCSM. The procedure starts by dividing the video into consecutive frames of still
images, then grayscale and binary transformations are performed. Next, each frame is divided into n grids, and cen-
troid of the object on each grid is computed by OTSM. The centroid information is retrieved in pixel coordinate and
then transformed to physical coordinate to obtain spatial displacement of the cable. The procedure is repeated for all
frames of images to obtain spatial time-histories displacement of the cable.
• Step 3: Modal analysis by DMD. Displacement responses are collected for certain duration to construct input block
matrix as described by Equations 9–11. Afterwards, using DMD, SVD and realization of the system matrix are per-
formed. Modal parameters of the cable, namely, natural frequency, damping ratio, and mode shape, are estimated
from the realized system matrix as explained by Equations 13 and 14.
• Step 4: Tension estimation. Using the identified natural frequencies from the previous step, information of mass per
unit length, and effective length of the cable, error function E k ðω1s , ξÞ is computed as described in Equation 21. The
error function is minimized iteratively by Newton–Raphson method to obtain the estimate of ω1s . The cable tension
can be estimated from ω1s by employing Equation 18.

The above procedure is implemented in the following section. There are three types of implementations or experiments
with different objectives described in the paper. The first one is a laboratory experiment as a proof of concept to verify
the proposed method and investigate the accuracy under a controlled environment. The second one is to test efficacy of
the method on a medium-scale cable-stayed bridge. In this case, a pedestrian cable-stayed bridge is used as a case study.
The experiment is important to investigate whether mode shapes of the whole or some portion of the cable can be identi-
fied. Another important fact is that in this experiment, the cables are still accessible; thus, they can be excited manually
if necessary. The third experiment is to test the capability of the method for implementation on a cable-stayed bridge
with very long span. In this test, the whole length of cable cannot be captured in one frame because of its length, so it is
important to test capability of the mode shape identification under limited length condition. In addition, stay cables are
inaccessible in the third experiment so manual excitation of the cable cannot be conducted. Each of the experiments has
the purposes and challenges, and they are presented in this paper to describe the capability and limitation of the method.

3 | LABORATORY EXPERIMENTAL VERIFICATION FOR MODAL


P A R A M E T E R S AC C U R A C Y

To verify the methodologies explained above, experimental verification was conducted on a tensioned cable specimen.
The cable was attached on a steel frame with one end fixed and the other loaded by a weight to induce tension. The
WANGCHUK ET AL. 9 of 25

cable specimen was made of tubular rubber with series of springs inserted inside. Cable effective lengths were 1530 and
1640 mm; with diameter and thickness of 40 and 2 mm, respectively (Figure 3). Three cable sets with different cable
arrangement and effective length were tested as listed in Table 1. It should be mentioned that the main objective of the
experiment is to verify the accuracy of vision-based measurement in obtaining modal parameters of the cable. It is not
designed to verify the cable tension estimation using formulation explained in Section 2.4, because the formulation has
been verified extensively in the previous studies (7,44,46, and6).
A Sony 4K NXCAM video camera was utilized to record cable vibration. The motion was recorded in a natural light-
ing condition without special illumination to represent an ambient measurement condition. For verification, cable
vibration was also recorded by LDV Polytec's RSV-150 with field sampling rate of 4000 samples/s. The cable was excited
vertically by inducing a small initial displacement to measure the in-plane free-vibration of cable was measured. The
video camera recorded the cable vibration with a pixel resolution of 3840  2160 at a frame rate of 30 fps. The pixel res-
olution was later reduced to 1200  420  3600 for further processing, cropping only the region of interest containing
the cable and its background. Note that the notation 1200  420 represents resolution, while 3600 represents the num-
ber of frames in the video that was sampled at the frequency 30 Hz. The magnified frames were extracted, and the true-
color (RGB) image was converted to the grayscale image and the binary image consecutively. Figure 4 illustrates an
example of transformation from an original, converted grayscale and binary image of a motion-magnified image frame.
A detailed example of analysis is provided for cable set 1 with dimension listed in Table 1. Initially, the PVMM is
performed for cable set 1 with the frequency range of 0.5–14.5 Hz to identify all possible vibration frequencies of the
cable. For amplification factor α = 1, relatively small amplitude of displacement was obtained from a point in the mid-
dle of the cable. The displacement was in the range of ±0.5 mm as shown in Figure 5. From the Fourier spectra peaks
of displacement time-history, natural frequencies are identified at 4.028 Hz for the first mode, 8.071 Hz for the second

F I G U R E 3 (a) Schematic figure of setup for experimental verification of modal analysis using LDV and video camera, (b) photo of
experiment setup, and (c) cable specimen

TABLE 1 Results of experimental verification using cable specimen by LDV and video measurement

Identified by LDV Identified by video Differencea (%)


Mass/ Added Freq [Hz]/
Cable Length length weight damp. Mode Mode Mode Mode Mode Mode Mode Mode Mode
set no (m) (kg/m) (kg) ratio (%) 1 2 3 1 2 3 1 2 3
1 1.53 0.29 0.50 ω [Hz] 4.029 8.080 12.080 4.030 8.070 12.060 0.025 0.124 0.166
ξ [%] 0.101 0.529 0.508 0.190 0.120 0.270 0.089 0.409 0.238
2 1.64 0.29 0.50 ω [Hz] 4.740 9.472 14.160 4.785 9.580 14.365 0.949 1.140 1.446
ξ [%] 0.234 0.184 0.147 0.125 0.080 0.067 0.109 0.104 0.080
3 1.64 0.83 1.00 ω [Hz] 3.027 6.006 8.936 3.010 5.960 8.980 0.562 0.764 0.498
ξ [%] 0.126 0.211 0.189 0.396 0.157 0.259 0.270 0.054 0.070

Abbreviation: LDV, laser Doppler vibrometer.


a
Note that the differences here refer to the relative differences between the results
from video measurement with respect to the ones obtained by
ωLDV −ωvideo
LDV; freq. difference = Δωi ¼ i ωLDV i  100%; damping difference = Δξi ¼ ξLDV i − ξvideo
i  100%.
i
10 of 25 WANGCHUK ET AL.

F I G U R E 4 (a) The 20th RGB frame of magnified motion, (b) converted grayscale image, (c) converted binary image and the centroid,
and (d) discretization of binary image for DCSM

F I G U R E 5 Example of cable in-plane displacement time-history and Fourier spectra measured from the middle point for (a) video after
phase-based video motion magnification (PVMM), (b) laser Doppler vibrometer (LDV), (c) Fourier spectra of video after PVMM, and
(d) Fourier spectra of LDV

mode, and 12.06 Hz for the third mode. Figure 5 also shows comparison of cable free vibration response measured by
LDV. It is noted that the time-history response is slightly different compared with the one obtained by video measure-
ment because the laser beam is not perpendicular to the direction of cable's in-plane vibration. Despite this difference,
the first three modes of the cable can be clearly identified at 4.053Hz, 8.057Hz, and 12.06 Hz, showing a good agree-
ment with the results from the video measurement.
In the next process, the video motion was magnified using α = 10 and the response was isolated in their respective
frequency band of interest to implement PVMM for modal identification. This is to ensure that the reconstructed video
contains only the mode of interest. Afterwards, the DCSM was employed to obtain displacement information along the
WANGCHUK ET AL. 11 of 25

length of the cable specimen. After converting the temporal frame of magnified motion from grayscale to binary image,
the image was divided into several grids for implementation of DCSM as explained in Equation 8. The centroid infor-
mation of each grid in the pixel resolution was estimated by OTSM and then transformed from pixel resolution to phys-
ical coordinate. This procedure was repeated for all frames of temporal image to obtain spatial displacement of the
cable for the entire length of video. Figure 4d shows a frame of binary image of the first mode and the corresponding
discretization into 30 grids. The number of rows and columns extracted from the magnified binary image in the first
mode was 214  1260  2100, which corresponds to the row number R = 214, column C = 1260, and number of
frames 2100. Using DCSM, the centroids information was obtained for all grids in all frames and the time-history of the
cable displacement was extracted.
The displacement responses were used as input for DMD method for system identification. The size of input matrix
in Equation 10 was selected using p = 500 and q = 200, which equals to 21 s of free-vibration displacement responses.
The same size of shifted input matrix Y was then constructed using the shifted time displacement response as noted by
Equation 11. The system identification was continued by conducting the SVD of the input matrix X and estimating the
system matrix A.e This involved a truncation of singular values in the diagonal matrix of singular value Σ that represents
the double of the number of selected system order. Afterwards, the eigenvalues of matrix A e were estimated, from which
the natural frequency (ωi ) and damping ratio (ξi ) were obtained. The eigenvectors were estimated by Equation 13 to
obtain the corresponding mode shapes. The DMD method successfully identified the first three modes as listed in
Table 1. Their natural frequencies are very close to the results obtained by LDV, and all modes generally have relatively
low damping (less than 1%). The mode shapes illustrated in Figure 6 clearly shows the typical shape of the taut cable's
modes with the modal assurance criteria (MAC) values all higher than 90%. The identified mode shapes closely resem-
ble the operational deflection shape extracted directly from the camera measurement after the PVMM using edge detec-
tion by the Laplacian of Gaussian (LOG) method. Note that videos of cable vibration in its natural frequency for the
first, second, and third modes are available in depository (Video S1). The same identification procedure was
implemented on other cable sets, to ensure the effectiveness of identifying the frequencies using the proposed method.
The maximum discrepancy in frequency estimation is 1.45%, which is within the acceptable range of approximation as
tabulated in Table 1.
The experimental verification demonstrates that modal parameters estimated by video motion using PVMM, DCSM,
OTSM, and DMD methods agree well with reference values measured by LDV. It should be mentioned that depending
on the pixel size of the frame, the PVMM and DCSM procedures require quite extensive computation. For example, the
calculation time for 1 min video with pixel resolution 2160  1080 and frame rate 30 fps for one grid in the DCSM pro-
cess was about 4 h using the PC with the processor-Intel® Core™ i7-8650U, CPU1.902.11 GHz, and 16-GB RAM of
memory. Better spatial resolution of modal displacement can be obtained with more grids, but this would require lon-
ger computation. The heavy computation sometimes limits the attainable length of displacement response extracted
from the video measurement.

F I G U R E 6 Results of laboratory experiment using the proposed vision-based modal analysis for cable setup 1: (a) mode shapes and
(b) comparison with edge detection by Laplacian of Gaussian (LOG) method Note: Video S1 showing vibration of the cable in its first,
second, and third modes is available in depository
12 of 25 WANGCHUK ET AL.

4 | FULL-SCALE IMPLEMENTATIONS ON CABLE-STAYED BRIDGES

This section describes full-scale implementation of the proposed measurement method on two cable-stayed bridges.
The first implementation is on a moderate length pedestrian cable-stayed bridge, and the second one is on a long-span
cable-stayed bridge. There are some practical considerations and important aspects to this method that require atten-
tion especially during video recording. One important aspect of measurement is the angle and position of the camera
with respect to the measured object. The best angle for camera is at the perpendicular angle to the plane on which the
object is vibrating. Another important aspect of measurement is weather condition. The most favorable weather condi-
tion for video measurement based on practical application is on a clear day without any clouds and wind gust. The pres-
ence of clouds in the background hinders the process of separating the target structure from background, and may
result in multiple centroids which ultimately reduce the accuracy when implementing the DCSM algorithm. This can
be avoided by selecting only parts of the video that do not include the rich background, although it may limit the spatial
information of the measured object. The presence of wind gust would shake the camera and introduces unwanted noise
to the actual object vibration. In a case where vibration of the camera is substantial and affects quality of the video, an
additional countermeasure should be considered. One way to solve the problem is by identifying the characteristics of
camera vibration and removing the frequency content associated with the camera vibration using digital filtering. For
this purpose, characteristics of camera vibration need to be identified by recording the camera vibration using small
wireless sensor mounted on the tripod. The other aspect that needs special attention during video recording is the frame
rate. In this study, the field measurement was conducted using frame rate 30 fps. After considering the Nyquist princi-
ple, this frame rate should be adequate to capture vibration of an object with the natural frequency less than 15 Hz
accurately. Predominant natural frequencies of flexible structures such as stay cables are normally less than 3 Hz.
Therefore, the current video camera with 30-fps sampling rate is adequate to capture vibration of the structure in its
predominant vibration modes.

4.1 | Implementation on pedestrian cable-stayed bridge

4.1.1 | Measurements of K-Bridge stay cables

The first example of full-scale implementation is the K-bridge. The bridge is a single pylon double-plane pedestrian
cable-stayed bridge located in Tokyo city, Japan. The bridge has an A-shaped pylon of 40 m high that divides the girder
to two continuous spans, the main span of 90 m and the side span of 49.4 m. The girder is made of steel box and
supported by five pairs of stay cables on each cable plane. The cable is made of high strength PC wires composed of
37 strands with the diameter of 7 mm. The cross-sectional area of cable is 1.424 mm2 and is protected by standard
polyethylene (PE) with the diameter of 87 mm. The total unit weight of PC cable including grouting is 18.8 kg/m.
The longest cable is 75.468 m located on the main span, and the shortest one is 29.630 m located on the side span.
Figure 7 shows the bridge dimension, cable locations and numbers, and the experimental setup using LDV and video
camera.

F I G U R E 7 Full-scale implementation. (a) Photo of pedestrian cable-stayed K-Bridge, bridge dimensions, and cable numbering
(b) experimental setup using laser Doppler vibrometer (LDV) and video camera
WANGCHUK ET AL. 13 of 25

Cable vibration was captured by Sony NXCAM camera with pixel resolution 3840  2160 at frame rate of 30 fps.
The video camera was placed about 30 m away from the cable plane as shown in Figure 7b. Measurement was carried
out for one cable plane before moved to the other plane. Simultaneously, LDV (Polytec's RSV-150) was also used to
record in-plane cable vibration for the purpose of comparison. The LDV was placed under the cable as shown in
Figure 7b. The bridge is located in relatively secluded area without significant number of traffic and pedestrian, so the
cables vibrations were very small. To initiate cable vibration, a simple manual pull-and-release excitation mechanism
was implemented. The vibration record was taken for about 10 min on each cable. The total measurement time for all
cables was about 4 h. An example of analysis for in-plane displacement of stay cable no. 1 and no. 4 of the upstream
side is explained in detail in this section. The same procedure was repeated for all cables. At first, the video measure-
ment recorded cable vibration at a pixel resolution 3840  2160. It was later reduced to 246  230 by cropping only the
region of interest consisting of the cable under consideration for further processing. PVMM was applied with amplifica-
tion factor α = 100 for a wide frequency band of 0.5–14.5 Hz as a preliminary survey of vibration modes that are present
in the cable vibration. This frequency range was selected considering the Nyquist frequency, and the frame rate of
30 fps was used for recording. The magnified frames were kept separated for implementing the proposed displacement
extraction technique. The RGB image was converted to a grayscale image and binary image. In this case, the binary
image transformation separated the background (black pixel “0”) and the target cable (white pixel “1”). Afterwards,
center of mass of the region (centroid) in pixel resolution was computed and tracked for full length in the video. The
pixel coordinate was converted to a physical displacement, and the displacement time-history response was extracted
from video measurement.
Figure 8a,e illustrates the displacement time-histories of cable no.1 and no. 4 on the upstream side, respectively.
The corresponding displacement responses recorded by LDV are shown by Figure 8e,g. Note that the measurement
point for video analysis was taken from the grid where the LDV measurement point was located so that the results
are comparable. Comparison of displacement time-histories clearly reveals similar characteristics of vibration with the
maximum amplitude within 3 mm for both cables. Fourier spectra of displacement time-histories of cable no.1 and
no. 4 on the upstream side obtained from the video measurement are shown by Figure 8b,f, respectively. One can see
clearly on the figures the three peaks of the cable's first three natural frequencies. For cable no.1 the first three fre-
quencies were identified at 2.37 Hz, 4.7 Hz, and 7.03 Hz, while for cable no. 4 the first three frequencies are 1.22 Hz,
2.43 Hz, and 3.65 Hz, respectively. The corresponding peaks are also observed on the spectra of displacement
recorded by LDV as shown by the bottom Figure 8d,h. The results show that the frequency peaks obtained from cable
vibration measurement by the video measurement agree well with those obtained using LDV with the maximum fre-
quency difference less than 0.5%. It should be mentioned, however, that, in general, video measurement can identify
up to the third modes, while LDV that has a better accuracy can identify many more of the higher modes. Neverthe-
less, as shall be seen later, the information of natural frequency up to the third modes is sufficient to estimate
cable tension.

F I G U R E 8 In-plane displacement time-histories and frequency spectra of K-Bridge stay cables. Stay cable no.1 (downstream) measured
by proposed vision-based method: (a) displacement time-history, (b) Fourier spectra. Stay cable no.1 (downstream) measured by laser
Doppler vibrometer (LDV): (c) displacement time-history, (d) Fourier spectra. Stay cable no.4 (upstream) measured by proposed vision-based
method: (e) displacement time-history, (f) Fourier spectra. Stay cable no.4 (upstream) measured by LDV: (g) displacement time-history,
(h) Fourier spectra
14 of 25 WANGCHUK ET AL.

4.1.2 | Identification of modal parameters of K-Bridge stay cables

In the foregoing analysis, frequency information of the cable was obtained from the spectra by taking the Fourier
transform of displacement from one centroid only. While the frequency spectra provide information of natural
frequencies, other modal parameters were not identified. To obtain a complete modal displacement along the cable,
DCSM and DMD were implemented on the motion-magnified video frame. The same process of displacement extrac-
tion was conducted, namely, performing the PVMM, converting the RGB frame to grayscale image then to binary
image, and finding the centroid for all grids in the segmented image series. The frequencies identified from the spectra
in the previous section were used as a reference to isolate the modes and to magnify the motion within the frequency
band of interest, which was the frequency band for each peak in the spectra. After motion magnifying the video within
the frequency band, the centroids of all grids were extracted by DCSM and the spatial displacements along the cable
were obtained.
An example of analysis for DCSM and modal identification by DMD is explained for cable no. 9 of the downstream
side, which is the second longest cable of the side-span with the effective length of 49.784 m. Figure 9 shows that the
video preprocessing procedure requires reduction of the region of interest of the original video measurement from
3840  2160 to 2130  190 containing only the cable no. 9 with a clear background. Initially, as in the case of other
cables, PVMM was applied in a wide frequency range (0.5–14.5 Hz) to identify all the frequencies present in the cable.
The response obtained by video measurement is shown by Figure 10a, and the responses measured simultaneously by
LDV are depicted by Figure 10b. The video measurement and postprocessing by PVMM successfully identified up to
the fifth mode, with the following frequencies: first mode (1.458 Hz), second mode (2.885 Hz), third mode (4.364 Hz),
forth mode (5.779 Hz), and fifth mode (7.222 Hz). These frequencies are very close with the corresponding frequencies
identified by LDV measurement as shown in Figure 10c,d with maximum frequency discrepancy less than 1%. Next,
PVMM was applied to isolate the individual modes and obtain the magnified video with the cable vibrating in that
mode only. The magnified video was then divided into several grids as shown in Figure 9c to obtain displacement infor-
mation along full length of the cable by applying DCSM.
It should be mentioned that the cable spatial displacements estimated within its natural frequency range have ran-
dom amplitude. To obtain damping by DMD method, the free-vibration responses should be extracted. Random decre-
ment technique (RDT)47 or Natural Excitation Technique (NExT)48 can be employed for this purpose. In this study, the
NExT was utilized because the method can accommodate multimode free-vibration responses. To implement the tech-
nique, cross-correlation functions between displacements of all grids and a selected reference displacement were com-
puted as follows:

Rkir ðT Þ ¼ E yki ðt þ T Þ  ykr ðt Þ ð22Þ

F I G U R E 9 Procedure for modal identification of stay cable no.9 (downstream). (a) Original recorded video RGB frame (3840  2160).
(b) Top: reduced size (2310  190) RGB frame, middle: grayscale transformed frame, and bottom: converted binary image and its centroid.
(c) Discretization of binary image for discretized centroid searching method (DCSM)
WANGCHUK ET AL. 15 of 25

F I G U R E 1 0 In-plane displacement time-histories of K-Bridge stay cable no.9 (downstream) measured by (a) proposed vision-based
method and (b) laser Doppler vibrometer (LDV) and their Fourier spectra measured by (c) proposed vision-based method and (d) LDV

where yi denotes the spatial displacement and i and yr the reference displacement. The reference displacement was
selected as the one that has the maximum modal displacement for each mode. The resulted cross-correlation responses
Rkir ðT Þ were used as inputs of X and Y matrices defined in Equations 10 and 11 using p = 500 and q = 200. This is
equivalent to constructing an input matrix of 21 s of free-vibration displacement data. The system identification contin-
ued by taking SVD of input matrix X and estimating the matrix A. e Afterwards, the eigenvalues of matrix A e were esti-
mated from which the natural frequency (ωi ) and damping ratio (ξi ) were obtained using Equation 14. The eigenvectors
were estimated by Equation 13 to obtain the corresponding mode shapes. The first three mode shapes of cable no.9
identified by DMD are shown in Figure 11. It should be mentioned that since the original recorded video's RGB frame
has limited size, it does not cover the entire cable; therefore, only half portion of the cable's mode shape can be identi-
fied as shown in this figure. After normalization to the maximum value, it is evident that the identified mode shapes
agree well with the reference mode shapes of taut cable. The MAC values of all mode shapes are higher than 90%. Video
S2 depicting the vibration of stay cable no. 9 in its natural frequency for the first, second, and third modes is available
in depository.
The same procedure was repeated for all cables, and the results of identification are listed in Tables 2 and 3 for the
natural frequencies and damping ratios, respectively. It should be mentioned that the number of identified modes is
not equal for all cables. In general, the proposed vision-based identification can identify up to the third mode. There-
fore, in these tables, only results from the first three modes are shown. For comparison, the displacement responses
obtained by LDV were also analyzed, and the natural frequency and damping ratio were estimated by random decre-
ment and logarithmic damping method. As revealed in Table 2, generally, the natural frequencies estimated by the pro-
posed vision-based identification agree well with the estimates obtained by LDV with the maximum frequency
difference of 2.28%. Table 3 shows that all cables have relatively low damping. They were all identified with damping
ratio less than 1% by both LDV and vision measurements. Damping ratios obtained by LDV and vision measurement
have quite large variation. Obtaining accurate estimates of damping can be problematic because they are estimated
from random vibration response especially for such a low-damped structure. In addition, unlike natural frequency,
accuracy of damping estimate depends on duration of the response. In this case, the length of displacement obtained by
vision measurement is limited up to 100 s.

4.1.3 | Tension forces estimation of K-Bridge stay cables

After estimating the natural frequencies of the stay cables, the next step is to use a cable model that relates the identi-
fied natural frequencies to the tension of the stay cables. Cable tension was estimated using the advanced approach
16 of 25 WANGCHUK ET AL.

F I G U R E 1 1 Mode shapes of K-Bridge stay cable no.9 (downstream) measured by the proposed vision-based method, (a) first mode and
the modal assurance criteria (MAC) values, (b) second mode, and (c) third mode. Note: Video S2. Showing vibration of stay cable no.9
(downstream) in its natural frequency for the first, second and third modes is available in depository

described in Section 2.4. The mass of the cable per unit length is m = 18.8 kg/m for all cables. The free length of
vibration L was taken from the static design, construction drawings, and the prestressing records. Their values vary
from the shortest one 23.973 m to the longest one 74.323 m. In the implementation of the method, only the second and
third natural frequencies are used, which is the minimum number of modes that should be used to solve Equation 21.
Natural frequency of the first mode was not included in solving Equation 21 because it would require the
nondimensional parameter μ whose values depend on the estimate of axial stiffness EA whose information was not
available at the time of analysis.
By minimizing the error function Equation 21 iteratively using Newton–Raphson, the ω1s and the dimensionless
stiffness ξ can be obtained. The initial values were first selected, and at each iteration, the ω1s and ξ were updated until
the difference between two consecutive values was smaller than the predefined threshold. It should be noted that the
initial values were selected by trial and error, but the final estimated values were independent of the selection of initial
values. In this case, the following initial values were selected: 5 < ω1s < 10 rad/s and 100 < ξ < 200: Results of calcula-
tion are presented in Table 4 along with the estimated ω1s from which the cable forces were estimated. Note that only
the natural frequencies obtained from vision measurement were used to calculate the tension forces. The results show
that tension forces of cables vary between 337 kN and 498.478 kN, where the shortest cable on the side span of each
cable plane (cable 6) has the largest tension force.

4.2 | Implementation on Tatara cable-stayed bridge

4.2.1 | Measurements of Tatara Bridge stay cables

The second case of full-scale implementation is the Tatara Bridge. The bridge is currently the longest cable-stayed
bridge in Japan and the eighth longest cable stayed bridge in the world. It connects two islands: Ikuchijima and
Ohmishima on the Onomichi-Imabari route of the Honshu-Shikoku Bridge, Japan. The bridge was opened to traffic on
WANGCHUK ET AL. 17 of 25

TABLE 2 Results of full‐scale implementation on K‐Bridge

Frequency by vision (Hz) Frequency by LDV (Hz) Freq. differencea (%)


Cable code
Upstream ω1 ω2 ω3 ω1 ω2 ω3 ω1 ω2 ω3
1 2.402 4.775 7.119 2.393 4.785 7.129 0.389 0.209 0.137
2 1.992 3.955 5.900 2.002 4.004 5.859 0.490 1.224 0.688
3 1.435 2.813 4.277 1.426 2.832 4.199 0.631 0.689 1.860
4 1.216 2.432 3.647 1.220 2.383 3.565 0.344 2.056 2.314
5 1.025 2.021 3.047 1.016 1.968 2.978 0.925 2.693 2.317
6 2.754 5.508 8.232 2.734 5.566 8.398 0.732 1.046 1.977
7 2.256 4.541 6.797 2.246 4.492 6.738 0.441 1.086 0.871
8 1.622 3.339 4.951 1.660 3.272 4.883 2.283 2.063 1.397
9 1.406 2.842 4.160 1.416 2.789 4.150 0.706 1.893 0.246
10 1.406 2.805 4.160 1.401 2.783 4.150 0.378 0.791 0.246
a
Frequency by vision (Hz) Frequency by LDV (Hz) Freq. difference (%)
Cable code
Downstream ω1 ω2 ω3 ω1 ω2 ω3 ω1 ω2 ω3
1 2.373 4.717 7.119 2.393 4.688 7.129 0.819 0.629 0.139
2 1.875 3.779 5.647 1.904 3.711 5.566 1.539 1.843 1.448
3 1.348 2.695 3.984 1.318 2.637 3.955 2.222 2.211 0.733
4 1.187 2.333 3.486 1.179 2.325 3.467 0.679 0.344 0.554
5 1.025 1.934 3.047 1.025 1.953 2.978 0.000 0.993 2.314
6 2.813 5.713 8.525 2.832 5.615 8.496 0.689 1.742 0.340
7 2.256 4.541 6.709 2.246 4.492 6.738 0.441 1.086 0.435
8 1.641 3.252 4.863 1.660 3.272 4.883 1.169 0.596 0.399
9 1.458 2.885 4.364 1.465 2.881 4.346 0.464 0.142 0.421
10 1.348 2.666 4.072 1.367 2.686 4.004 1.412 0.726 1.706

Note: Comparison of stay cables natural frequencies obtained by LDV measurement and the proposed vision‐based measurement. Note that frequency
estimates by LDV are obtained via spectra and random decrement; frequency estimates by vision‐based measurement are obtained by system identification.
Abbreviation: LDV, laser Doppler vibrometer.
a
The frequency differences here refer to the relative differences between the results from video measurement with respect to the ones obtained by LDV.

May 1, 1999. Total bridge length is 1.480 m with the main span of 890 m; the side spans are 270 m long towards the
Ikuchijima Island and 320 m long towards the Ohmishima Island. This bridge utilizes a composite structure of steel
and prestressed concrete. The dead-load imbalance between the center span and the side spans, caused by the side
spans being shorter than the center span, is compensated by installing concrete girders at the end of the side spans. The
bridge has two inverted Y-shaped pylons made of steel box with the total height 226 m above the sea level. The bridge
has a multifan-type stay cables with two planes (168 cables in total). Each cable consists of galvanized steel wires and is
coated with PE. Cable diameters vary between 110 and 170 mm with the weight between 94.8 kg/m and 121.8 kg/m.
The cable lengths vary between the shortest 108 m and the longest 470 m, which give the frequency of the first mode
between 0.26 Hz and 1.05 Hz. As measures for vibration control, the stay cables are equipped with dampers on the
cable support at the girder and cables surfaces are indented to provide better resistance against wind-rain-induced
vibrations.
Full-scale implementation on Tatara Bridge was conducted to test the capability of the method on a cable-stayed
bridge with very long span. In this test, the whole cable cannot be captured in one frame because of its length, so it is
important to test the capability of the mode shape identification under limited length condition. Also, because of its size
and location, stay cables were inaccessible for manual excitation, resulting in the measurement relied entirely on ambi-
ent vibration. Figure 12 provides general information of Tatara Bridge dimensions. Due to access limitation, only stay
cables on the side span of Ikuchijima side were measured in this study as shown in Figure 12b. The cable vibration was
captured by the same Sony NXCAM camera with pixel resolution 3840  2160 at frame rate of 30 fps. The video camera
18 of 25 WANGCHUK ET AL.

TABLE 3 Results of full‐scale implementation on K‐Bridge

Damp. ratio by vision (%) Damp. ratio by LDV (%) Damp. ratio differencea (%)
Cable code
Upstream ξ1 ξ2 ξ3 ξ1 ξ2 ξ3 ξ1 ξ2 ξ3
1 0.085 0.240 0.447 0.117 0.338 0.179 0.033 0.098 0.267
2 0.154 0.185 0.185 0.241 0.630 0.038 0.087 0.445 0.148
3 0.274 0.217 0.368 0.088 0.073 0.073 0.186 0.144 0.295
4 0.456 0.161 0.512 0.087 0.070 0.104 0.369 0.091 0.408
5 0.691 0.196 0.511 0.115 0.444 0.139 0.576 0.248 0.372
6 0.446 0.213 0.164 0.155 0.901 0.584 0.446 0.213 0.164
7 0.104 0.056 0.183 0.098 0.106 0.051 0.006 0.050 0.133
8 0.597 0.747 0.389 0.147 0.090 0.128 0.450 0.657 0.261
9 0.166 0.208 0.338 0.063 0.080 0.193 0.103 0.128 0.145
10 0.124 0.285 0.358 0.138 0.130 0.128 0.124 0.285 0.358
Damp. ratio by vision (%) Damp. ratio by LDV (%) Damp. ratio difference (%)
Cable code
Downstream ξ1 ξ2 ξ3 ξ1 ξ2 ξ3 ξ1 ξ2 ξ3
1 0.080 0.096 0.369 0.084 0.077 0.334 0.004 0.019 0.035
2 0.069 0.287 0.180 0.066 0.111 0.333 0.003 0.176 0.153
3 0.438 0.274 0.419 0.073 0.099 0.093 0.366 0.176 0.327
4 0.275 0.442 0.237 0.086 0.698 0.122 0.189 0.256 0.116
5 0.192 0.317 0.375 0.115 0.445 0.139 0.077 0.128 0.235
6 0.330 0.143 0.125 0.021 0.230 0.683 0.309 0.087 0.559
7 0.133 0.061 0.182 0.045 0.076 0.044 0.089 0.015 0.138
8 0.179 0.196 0.379 0.493 0.247 0.142 0.314 0.052 0.237
9 0.480 0.413 0.217 0.116 0.110 0.236 0.364 0.303 0.019
10 0.210 0.173 0.277 0.165 0.077 0.315 0.210 0.173 0.277

Note: Comparison of stay cables damping ratios obtained by LDV measurement and the proposed vision‐based measurement. Note that damping ratio
estimates by LDV are obtained via logarithmic decrement method; damping estimates by vision‐based measurement are obtained by system identification.
Abbreviation: LDV, laser Doppler vibrometer.
a
The damping differences here refer to the relative differences between the results from video measurement with respect to the ones obtained by LDV.

TABLE 4 Tension force estimation of stay cables of K‐Bridge obtained by the proposed vision‐based measurement

Free Free
Cable code length Identified Tension Cable code length Identified Tension
Upstream (m) ω1s (Hz) estimates (kN) Downstream (m) ω1s (Hz) estimates (kN)
1 29.732 2.319 357.593 1 29.732 2.279 345.195
2 40.354 1.920 451.448 2 40.354 1.834 412.011
3 51.478 1.410 396.216 3 51.478 1.301 337.491
4 62.837 1.190 420.174 4 62.837 1.145 388.959
5 74.323 0.993 409.620 5 74.323 0.969 389.711
6 29.732 2.654 468.158 6 29.732 2.738 498.478
7 36.311 2.199 479.578 7 36.311 2.198 479.163
8 43.011 1.616 363.210 8 43.011 1.595 353.908
9 49.784 1.372 350.773 9 49.784 1.415 372.902
10 56.603 1.370 452.313 10 56.603 1.316 417.331
WANGCHUK ET AL. 19 of 25

FIGURE 12 (a) Photo of Tatara cable-stayed bridge, (b) the stay cables measured in this study and its numbering, and (c) dimensions of
the bridge

F I G U R E 1 3 Experimental setup for cable measurement at Tatara Bridge. (a) Setup of video camera and (b) setup for laser Doppler
Vibrometer. (c) The RGB image of cables on north side of the side span: 3840  2160(original) and 572  432 (reduced size). (d) Grayscale
image. (e) Binary image with centroid. (Note: Video S3. Showing vibration of stay cable C1 in its natural frequency for the first, second and
third modes is available in depository)

was placed under the girder at a distance about 50–100 m away from the cable plane as shown in Figure 13. Like in the
previous example, LDV (Polytec's RSV-150) was placed under the cable to record in-plane cable vibration for validation.
Both devices measured the target on the same cable at the same measurement time. Since there was no excitation given
to the cable to initiate vibration, the measurement was conducted under full ambient vibration condition. The general
placement of the recording device and the target cables is depicted in Figure 13. The original video data which had a
pixel resolution of 3840  2160 was reduced to region of interest of 572  432 by cropping only the cable's region of
interest for analysis. The ambient vibration present in the cable was motion magnified by implementing PVMM. This
20 of 25 WANGCHUK ET AL.

process was conducted by amplifying the phase using amplification factor α = 10 for a cable frequency range while con-
sidering the Nyquist theorem. The reconstructed magnified frames produce a magnified video which was the amplified
ambient vibration of the cable. These magnified frames were then used to extract displacement using image-processing
technique. Figure 13c–e shows the original magnified frames in true-color (RGB) image, the converted grayscale image,
and then the binary image.
The same procedure of DCSM was implemented to the binary images to obtain the centroid in pixel resolution and
tracked the centroids for full length of the video. The pixel coordinate was then converted to physical displacement to
obtain time-history response of the cable. Figure 14a,b illustrates the time-history and Fourier amplitude spectra of in-
plane cable displacement obtained by DCSM and PVMM on cables C1–C4, respectively. The corresponding displace-
ment responses recorded by LDV are shown by Figure 15a,b. The figures demonstrate that displacement time-histories
have the maximum amplitude of 2 mm for all cables. The Fourier spectra of displacement obtained from the video mea-
surement generally result in three peaks of the cable's first three natural frequencies. On the other hand, the spectra of
displacement from LDV measurement have more peaks up to the fifth or the sixth frequencies. Figures 14b and 15b
depict the comparisons of natural frequency obtained by video measurement and LDV. It is evident from the figure that
the frequencies obtained by both measurements generally have good agreement, with the maximum frequency differ-
ence of 2%.

4.2.2 | Identification of modal parameters of Tatara Bridge stay cables

Next, the DCSM and modal identification by DMD were implemented to obtain the complete set of modal parameters.
As the first step, the PVMM was applied to isolate the individual modes and obtain the magnified video with the cable
vibrating in that mode only. The magnified video was then divided into 12 grids to obtain displacement information
along cable by applying DCSM. Similar to the first case above, the NExT was employed to derive free-vibration displace-
ment response of the cable. One of the 12 centroids was selected as a reference, upon which the cross-correlation func-
tions between displacements of all grids were computed. The reference displacement was selected as the one that has

F I G U R E 1 4 In-plane displacement of Tatara Bridge stay cables C1–C4 measured by the proposed vision-based method. (a) Time-
histories of displacement. (b) Fourier spectra of time-histories
WANGCHUK ET AL. 21 of 25

F I G U R E 1 5 In-plane displacement of Tatara Bridge stay cables C1–C4 measured by the laser Doppler vibrometer (LDV). (a) Time-
histories of displacement. (b) Fourier spectra of time-histories-

the maximum modal displacement for each mode. The resulted cross-correlation responses Rkir ðT Þ were used as inputs
of X and Y matrices defined in Equations 10 and 11 using p = 500 and q = 200 similar to the example in the previous
bridge. The system identification continued by taking SVD of input matrix X and estimating the matrix A e from which
the eigenvalues were estimated to obtain natural frequency (ωi ) and damping ratio (ξi ), while the mode shape was
derived from the corresponding eigenvectors.
Results of identification for the natural frequencies and damping ratios are listed in Table 5 for cables C1–C5. In
general, the proposed vision-based identification can identify up to the third mode. Therefore, only results from the first
three modes are shown in the table. The natural frequencies and damping ratios estimated by random decrement and
logarithmic damping method using the displacements measured by LDV are also presented in the table. The results are
presented by the mean and standard deviation values. Note that variations in the estimated natural frequencies are
smaller compared with that of damping ratios, signifying the higher accuracy and repetitiveness of frequency identifica-
tion. The comparisons show that generally, the natural frequencies estimated by the proposed vision-based identifica-
tion agree well with the estimates obtained by LDV. The maximum frequency difference between the two methods is
2.35%. Table 5 also shows the estimated damping ratios of the cables. The damping ratios were identified between
0.75% and 2.5% by both LDV and video measurements, with quite large variations. The cable mode shapes were esti-
mated from eigenvectors computed by Equation 13. However, since the portion of the cable captured by video measure-
ment is significantly smaller compared with the total length of the cable, mode shapes estimated by the system
identification do not represent the overall characteristics of taut-cable's mode shape as in the case of K-Bridge. Examples
of vibration of stay cable C1 in its natural frequency for the first, second, and third modes are shown by Video S3 that is
available in depository.

4.2.3 | Tension forces estimation of Tatara Bridge stay cables

Following the procedure described in Section 2.4, tension of cables C1–C5 was estimated using the additional
information of mass of the cable per unit length and the free length of vibration L that was taken from the static
22 of 25 WANGCHUK ET AL.

TABLE 5 Results of full‐scale implementation on Tatara Bridge

Mean frequency by vision (Hz)a [standard Mean frequency by LDV (Hz)a [standard
deviation (%)] deviation (%)] Freq. differencea (%)

Cable code ω1 ω2 ω3 ω1 ω2 ω3 ω1 ω2 ω3
C‐1 0.375 [0.41] 0.710 [1.35] 1.069 [2.56] 0.366 [0.28] 0.708 [0.43] 1.074 [1.18] 2.347 0.345 0.458
C‐2 0.381 [0.19] 0.732 [0.58] 1.084 [1.84] 0.378 [0.23] 0.732 [0.48] 1.098 [2.26] 0.761 0.055 1.292
C‐3 0.388 [0.72] 0.740 [0.49] 1.113 [8.72] 0.391 [0.61] 0.739 [0.55] 1.147 [8.34] 0.670 0.115 3.055
C‐4 0.396 [0.18] 0.762 [0.97] 1.143 [8.29] 0.396 [0.26] 0.757 [0.61] 1.148 [3.74] 0.126 0.643 0.429
C‐5 0.417 [0.98] 0.791 [1.21] 1.201[11.61] 0.415 [0.38] 0.806 [1.28] 1.209 [4.35] 0.480 1.896 0.608
a a
Mean damping ratio by vision (%) Mean damping ratio by LDV (%) Damp. ratio difference
[standard deviation (%)] [standard deviation (%)] (%)

Cable code ξ1 ξ2 ξ3 ξ1 ξ2 ξ3 ξ1 ξ2 ξ3
C‐1 1.022 [52.5] 1.134 [54.7] 0.872 [32.9] 1.001 [24.9] 1.286 [44.3] 0.889 [32.2] 0.020 0.152 0.018
C‐2 2.162 [51.1] 0.790 [20.1] 1.540 [30.0] 1.259 [46.3] 0.749 [20.4] 1.062 [14.8] 0.903 0.042 0.178
C‐3 1.466 [42.8] 0.855 [32.2] 0.815 [38.6] 0.947 [18.7] 1.262 [28.5] 1.116 [19.9] 0.519 0.407 0.301
C‐4 1.365 [49.5] 0.893 [36.2] 1.695 [60.5] 1.043 [26.1] 0.688 [28.1] 0.889 [24.5] 0.322 0.205 0.806
C‐5 2.491 [66.2] 1.089 [41.1] 2.318 [42.1] 1.389 [51.2] 0.544 [17.5] 0.818 [27.1] 1.103 0.548 1.500

Note: Comparison of stay cables natural frequencies and damping ratios obtained by LDV measurement and the proposed vision‐based measurement. Damping
ratio estimates by LDV are obtained via logarithmic decrement method; damping estimates by vision‐based measurement are obtained by system identification.
The differences here refer to the relative differences between the mean results from video measurement with respect to the ones obtained by LDV.
a
The values are the mean, and inside the brackets [.] are the standard deviation (in %).

TABLE 6 Tension force estimation of stay cables of Tatara‐Bridge obtained by the proposed vision‐based measurement

Cable Mass per length Free length Identified w1s Tension Design tension Difference in tension
code (kg/m) (m) (Hz) estimates (kN) (kN) estimatea (%)
C1 121.8 317.722 0.3553 6170.1 6210.2 0.94
C2 112.6 309.943 0.3633 5676.0 5712.1 3.04
C3 108.2 301.202 0.3702 5347.0 5380.4 3.84
C4 100.7 292.927 0.3807 4978.2 5008.4 7.92
C5 94.8 284.683 0.3976 4829.7 4857.7 6.80

T design −T vision_est
Difference in tension = ΔT i ¼ i designi
a ñ100%:

Ti

design as listed in Table 6. Like the case of K-Bridge, only the second and third natural frequencies estimated
from the vision measurement were used for tension estimation. Next, the error function in Equation 21 was mini-
mized iteratively using Newton–Raphson to obtain the ω1s . The initial values were first selected, and at each itera-
tion, the values of ω1s and ξ were updated until the difference between two consecutive values is smaller than the
threshold. In this case, the following initial values were selected: 5 < ω1s < 10 rad/s and 8000 < ξ < 10,000: Results of
calculation are presented in Table 6 along with estimated ω1s from which the cable forces were estimated. The esti-
mated cable tension forces vary between 4830 kN and 6170 kN, where the longest cable (C1) has the largest tension
force. The table reveals comparisons of the estimated tension forces and the design values provided by the design docu-
ment. In general, the estimation values are in good agreement with the design values, with the maximum difference
of 8%.
The results demonstrate that the proposed vision-based video measurement can identify low-order modal parame-
ters of stay cables accurately in ambient condition. It has also been demonstrated in the earlier measurements on K-
Bridge that the pull and release mechanism of cables can excite the cable higher modes and that they can be identified
WANGCHUK ET AL. 23 of 25

correctly by vision-based method. The fact that modal parameters can be identified through a measurement without
intentionally exciting the cable is an advantage for routine inspection because manual excitation of a long and heavy
cable is complicated. Therefore, the proposed vision-based vibration measurement offers a more effective and efficient
way of identifying cable frequencies with good accuracy. There are some practical considerations and important aspects
to this method that require attention especially during video recording. The most favorable weather condition for video
measurement based on practical application is on a clear day without any clouds. We found several times during mea-
surement that the presence of clouds in the background complicated the process of separating the target structure from
the background. In such condition, multiple centroids appeared when conducting the DCSM, which ultimately reduced
accuracy of the measurement. Further research works are needed to alleviate such effect to make measurement more
reliable and robust.

5 | C ON C L U S I ON S

In this study, vision-based modal identification and tension force estimates of stay cables are described. The
vision-based method is based on the phase-based motion magnification method that amplifies small-amplitude
cable motion. Spatial displacements of the cable are extracted by DCSM that is based on the OTSM. Afterwards,
modal parameters of stay cables are estimated using the spatial displacements by the DMD method to obtain
the cable's natural frequency, damping ratio, and mode shape. Information obtained from identified natural
frequency is used to estimate cable tension forces by minimizing an error function of approximate relationship
between frequency and tension iteratively. The proposed modal identification and tension estimation procedure
are verified in the laboratory-scale experiments and implemented on full-scale measurement of two cable-stayed
bridges.
Comparisons of results of the proposed vision-based measurement with the measurement using high-precision LDV
have demonstrated that modal parameters of stay cables can be extracted by vision measurement using PVMM, DCSM,
and DMD accurately with the maximum frequency difference of 2%. Results of stay cable tension estimation from full-
scale implementation on Tatara Bridge have shown that the vision-based measurement can provide accurate estimates
of tension when compared with the tension design values. The practical importance of the proposed method is that it
enables modal identification and frequency-based tension estimation of stay cable in a noncontact manner using a rela-
tively simple, less expensive video measurement from a distance under small amplitude of ambient excitation with good
accuracy. This would provide a promising alternative to the current contact-sensor-based measurement system and
frequency-based tension estimation of stay cables.

A C K N O WL E D G M E N T S
The first author would like to acknowledge the support of scholarship from the Government of Japan's Ministry of
Education, Culture, Sports, Science and Technology (MEXT), Government of Japan. We also gratefully acknowledge
the Honshu-Shikoku Bridge Authority (HSBA) for providing access to Tatara Bridge and relevant documents for this
research and Bridge & Structures Laboratory, The University of Tokyo (Dr. Di Su and Dr. Tomonori Nagayama) for
the LDV measurements. Opinions and findings in this paper are of the authors and not of those mentioned
institutions.

A U T H O R C ON T R I B U T I O NS
Samten Wangchuk: Field measurements, laboratory experiments, data analysis, structural modeling, and manuscript
writing. Dionysius M. Siringoringo: conceptualization, structural modeling, programming, manuscript preparation,
supervision, and academic advisor. Yozo Fujino: conceptualization, supervision, and academic advisor.

DATA AVAILABILITY STATEMENT


The data that support the findings of this study are available from the corresponding author upon reasonable request.

ORCID
Dionysius M. Siringoringo https://orcid.org/0000-0003-2267-0673
Yozo Fujino https://orcid.org/0000-0001-8993-3779
24 of 25 WANGCHUK ET AL.

R EF E RE N C E S
1. Fujino Y, Siringoringo DM, Abe M. Japan's experience on long-span bridges monitoring. Struct Monit Maint. 2016;3(3):233-257.
2. Brönnimann R, Nellen PM, Sennhauser U. Application and reliability of a fiber optical surveillance system for a stay cable bridge. Smart
Mater Struct. 1998;7(2):229-236.
3. Mehrabi AB. In-service evaluation of cable-stayed bridges, overview of available methods and findings. J Bridge Eng. 2006;11(6):716-724.
4. Christen R, Bergamini A, Motavalli M. Three-dimensional localization of defects in stay cables using magnetic flux leakage methods. J
Nondestr Eval. 2003;22(3):93-101.
5. Cunha A, Caetano E, Delgado R. Dynamic tests on large cable-stayed bridge. J of Bridge Eng. 2001;6(1):54-62.
6. Geier R, De Roeck G, Flesch R. Accurate cable force determination using ambient vibration measurements. Struct Infrastruct Eng. 2006;
2(1):43-52.
7. Zarbaf HAMSE, Norouzi M, Allemang RJ, Hunt VJ, Helmicki A, Venkatesh C. Ironton-Russell bridge: Application of vibration-based
cable tension estimation. J Struct Eng. 2018;144(6):04018066.
8. Feltrin G, Meyer J, Bischoff R, Motavalli M. Long-term monitoring of cable stays with a wireless sensor network. Struct Infrastruct Eng.
2010;6(5):535-548.
9. Morgenthal G, Sebastian R, Jakob T, Tajammal A. Determination of stay-cable forces using highly mobile vibration measurement
devices. J Bridge Eng. 2018;23(2):04017136.
10. Sim SH, Li J, Jo H, et al. A wireless smart sensor network for automated monitoring of cable tension. Smart Mater Struct. 2013;23(2):
025006.
11. Cunha A, Caetano E. Dynamic measurements on stay cables of cable-stayed bridges using an interferometry laser system. Exp Tech.
1999;23(3):38-43.
12. Mehrabi AB, Farhangdoust S. A laser-based noncontact vibration technique for health monitoring of structural cables: Background, suc-
cess, and new developments. Adv Acoust Vib. 2018;2018:1-13.
13. Gentile C, Cabboi A. Vibration-based structural health monitoring of stay cables by microwave remote sensing. Smart Struct Syst. 2015;
16(2):263-280.
14. Zhao W, Zhang G, Zhang J. Cable force estimation of a long-span cable-stayed bridge with microwave interferometric radar. Comput.-
Aided Civ Infrastruct Eng. 2020;35(12):1419-1433.
15. Xu Y, Brownjohn JM. Review of machine-vision based methodologies for displacement measurement in civil structures. J Civ Struct
Health Monit. 2018;8(1):91-110.
16. Kim SW, Jeon BG, Kim NS, Park JC. Vision-based monitoring system for evaluating cable tensile forces on a cable-stayed bridge. Struct
Health Monit. 2013;12(5–6):440-456.
17. Dong C-Z, Catbas FN. A review of computer vision–based structural health monitoring at local and global levels. Struct Health Monit.
2020;20(2):692–743.
18. Jana D, Nagarajaiah S. Computer vision-based real-time cable tension estimation in Dubrovnik cable-stayed bridge using moving hand-
held video camera. Struct Control and Health Monit. 2021;28(5):e2713.
19. Kim SW, Cheung JH, Park JB, Na SO. Image-based back analysis for tension estimation of suspension bridge hanger cables. Struct
Control Health Monit. 2020;27(4):e2508.
20. Wadhwa N, Rubinstein M, Durand F, Freeman WT. Phase-based video motion processing. ACM Trans Graphics. 2013;32(4):1-10.
21. Otsu N. A threshold selection method from gray-level histograms. IEEE Trans Sys, Man, Cyber. 1979;9(1):62-66.
22. Thiyagarajan JS, Siringoringo DM, Wangchuk S, Fujino Y. Implementation of video motion magnification technique for noncontact
operational modal analysis of light poles. Smart Struct & Syst. 2021;27(2):227-239.
23. Bhowmick S, Nagarajaiah S, Zhilu L. Measurement of full-field displacement time history of a vibrating continuous edge from video.
Mech Syst Signal Process. 2020;144:106847.
24. Feng D, Scarangello T, Feng MQ, Ye Q. Cable tension force estimate using novel noncontact vision-based sensor. Measurement. 2017;99:
44-52.
25. Chen JG, Wadhwa N, Cha YJ, Durand F, Freeman WT, Buyukozturk O. Modal identification of simple structures with high-speed video
using motion magnification. J Sound Vib. 2015;345:58-71.
26. Chen JG, Adams TM, Sun H, Bell ES, Büyüköztürk O. Camera-based vibration measurement of the world war I memorial bridge in
portsmouth, New Hampshire. J Struct Eng. 2018;144(11):04018207.
27. Yang Y, Dorn C, Mancini T, et al. Blind identification of full-field vibration modes from video measurements with phase-based video
motion magnification. Mech Sys Signal Proc. 2017;85:567-590.
28. Yang Y, Sanchez L, Zhang H, et al. Estimation of full-field, full-order experimental modal model of cable vibration from digital video
measurements with physics-guided unsupervised machine learning and computer vision. Struct Control Health Monit. 2019;26(6):e2358.
29. Zhou X, Xia Y, Wei Z, Wu QA. Videogrammetric technique for measuring the vibration displacement of stay cables. Geo-Spatial Inform
Sci. 2012;15(2):135-141.
30. Zhao X, Ri K, Wang N. Experimental verification for cable force estimation using handheld shooting of smartphones. J Sens. 2017;2017:
1-13.
31. Tian Y, Zhang C, Jiang S, Zhang J, Duan W. Noncontact cable force estimation with unmanned aerial vehicle and computer vision.
Comput.-Aided Civ Infrastruct Eng. 2021;36(1):73-88.
WANGCHUK ET AL. 25 of 25

32. Bhowmick S, Nagarajaiah S. Identification of full-field dynamic modes using continuous displacement response estimated from vibrat-
ing edge video. J Sound Vib. 2020;489:115657.
33. Portilla J, Simoncelli EP. A parametric texture model based on joint statistics of complex wavelet coefficients. Int J Compt Vision. 2000;
40(1):49-70.
34. Schmid PJ. Dynamic mode decomposition of numerical and experimental data. J Fluid Mechanics. 2010;656:5-28.
35. Kutz JN, Brunton SL, Brunton BW, Proctor JL. Dynamic mode decomposition: Data-driven modeling of complex systems. In: Society for
Industrial and Applied Mathematics; 2016.
36. Tu JH, Rowley CW, Luchtenburg DM, Brunton SL, Kutz JN. On dynamic mode decomposition: Theory and applications. J Comput Dyn.
2014;1(2):391-421.
37. Irvine HM. Cable Structures. New York: Dover Publications; 1992.
38. Morse PM, Ingard KU. Theoretical Acoustics. Princeton university press; 1986.
39. Shimada T. Estimating method of cable tension from natural frequency of high mode. Doboku Gakkai Ronbunshu JSCE. 1994;501(501):
163-171.
40. Triantafyllou MS, Grinfogel L. Natural frequencies and modes of inclined cables. J Struct Eng. 1986;112(1):139-148.
41. Zui H, Shinke T, Namita Y. Practical formulas for estimation of cable tension by vibration method. J Struct Eng. 1996;122(6):651-656.
42. Bao Y, Shi Z, Beck JL, Li H, Hou TY. Identification of time-varying cable tension forces based on adaptive sparse time-frequency analysis
of cable vibrations. Struct Cont Health Monit. 2017;24(3):e1889.
43. Yang Y, Li S, Nagarajaiah S, Li H, Zhou P. Real-time output-only identification of time-varying cable tension from accelerations via com-
plexity pursuit. J Struct Eng. 2015;142(1):4015083.
44. Mehrabi AB, Tabatabai H. Unified finite difference formulation for free vibration of cables. J Struct Eng 1998;124(11):1313-1322.
45. Morse PM, Ingard KU. Theoretical acoustics. Princeton university press; 1987.
46. Zarbaf HAMSE, Norouzi M, Allemang RJ, Hunt VJ, Helmicki A, Nims DK. Stay force estimation in cable-stayed bridges using stochastic
subspace identification methods. J Bridge Eng. 2017;22(9):04017055.
47. Vandiver JK, Dunwoody AB, Campbell RB, Cook MF. A mathematical basis for the random decrement vibration signature analysis
technique. J Mech Des. 1982;104(2):307-313.
48. James GH, Carne TG, Lauffer JP. The natural excitation technique (NExT) for modal parameter extraction from operating structures.
J Anal Exp Modal Anal. 1995;10(4):260-277.

S UP PO RT ING IN FOR MAT ION


Additional supporting information may be found in the online version of the article at the publisher's website.

How to cite this article: Wangchuk S, Siringoringo DM, Fujino Y. Modal analysis and tension estimation of
stay cables using noncontact vision-based motion magnification method. Struct Control Health Monit. 2022;
e2957. doi:10.1002/stc.2957

You might also like