Professional Documents
Culture Documents
Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct
A R T I C L E I N F O A B S T R A C T
Keywords: As the automated modal analysis is crucial for a continuous monitoring system, this study proposes a framework
Automated modal identification for automated modal identification of bridge parameters based on the uncertainty of estimated frequencies and
Frequency uncertainty density-based clustering algorithm, which consists of the following three stages: First, the modal parameters and
DBSCAN algorithm
standard deviations of the estimated frequencies are calculated in a wide range of model orders to construct the
Spurious mode
Bridge modal analysis
stabilization diagram using the reference-based covariance-driven stochastic subspace identification algorithm.
Second, the criteria of frequency uncertainty and stabilization are adopted to eliminate the spurious modes.
Third, for present purpose, the modified version of an unsupervised density-based clustering algorithm is
The selling point of this paper is introduced to group physical modes and detect outliers to reach automated identification of bridge modal pa
automayed modal itervention rameters. From the analysis, it has shown that the proposed framework is powerful in eliminating the spurious
without relying on manual modes and robust in the presence of interference caused by spurious modes while a simple procedure for clus
interference. tering physical modes with desired statistical reliability is employed.
* Corresponding author.
E-mail address: jpyang@nycu.edu.tw (J.P. Yang).
https://doi.org/10.1016/j.engstruct.2022.113891
all stable modes as separate clusters and then group the adjacent two The conclusions of this study are drawn in Section 5.
clusters as a new cluster. Such a procedure is repeated until the distances
between the rest clusters are larger than a user-specified threshold, 2. Review of SSI-cov/ref and uncertainty computation of modal
which indicates that the core of the hierarchical clustering algorithm is parameters
to decide the threshold. On the other hand, the partitioning methods
such as the K-means and fuzzy C-means clustering algorithms [18] tend 2.1. SSI-cov/ref
to divide the stable modes into several clusters, which have the
following drawbacks: The number of clusters needs to be specified in The discrete-time state and observation equations of a linear, time-
advance; the results are sensitive to noise points and may not be invariant structure under the white noise excitation at time step k can
deterministic due to randomly generated clustering seeds. Recently, be respectively described as follows [10]:
some strategies have been proposed to determine the optimal cluster
xk+1 = Axk + wk (1a)
number in the modal analysis. Yi et al. proposed that the cluster number
in the sparse component analysis can be determined by using the sta
yk = Cxk + vk (1b)
tistical properties of normalized single-source-point vectors [19]. He
et al. suggested that the optimal cluster number of identified physical where xk ∈ Rn is the state vector of the structure with the model order n
modes can be determined through an iterative intelligent graph parti equal to two times of the number of degrees-of-freedom of the structure;
tioning algorithm [16]. The DBSCAN (density-based spatial clustering of A ∈ Rn×n is the state transition matrix; wk ∈ Rn is the process noise vector;
applications with noise) algorithm is a density-based data clustering yk ∈ Rr is the measurement vector obtained by r sensors; C ∈ Rr×n is the
algorithm, and it does not need to specify the number of clusters in observation matrix; vk ∈ Rr is the measurement noise vector.
advance and can discover clusters of different shapes and sizes from a In the operational modal parameter identification, a crucial step is to
large amount of data containing noises and outliers [20]. Thus, it might estimate the matrices A and C containing the dynamic properties of the
be suitable for grouping physical modes since the physical mode pa structure from the measured structural responses as they are related to
rameters naturally form several regions of high-density in the modal the mass and stiffness matrices of the structure [10]. To enhance
parameter space. computational efficiency and estimation accuracy, the r0 sensors are
Before conducting the clustering analysis, the removal of spurious selected from r sensors as the reference channels in the SSI-cov/ref al
modes from the stabilization diagram is the most important step in the gorithm [5]. The measurement data are arranged in matrix form as
automated modal identification since the presence of spurious modes ⎡ ⎤
could lead to undesired results. To this end, many criteria have been yq+1 yq+2 ⋮ yN+q
proposed to clean the stabilization diagrams [11,15], including the 1 ⎢ yq+2 yq+3 ⋮ yN+q+1 ⎥
y+ = √̅̅̅̅ ⎢ ⎥, y −
modal assurance criterion (MAC) [21], modal phase collinearity (MPC) N⎣ ⋮ ⋮ ⋮ ⋮ ⎦
yq+p+1 yq+p+2 ⋮ yN+p+q
[22,23], and modal transfer norm (MTN) [24]. Although these criteria ⎡ ⎤
can remove spurious modes to a certain extent, they need to be used (ref) (ref) (ref)
⎢ yq yq+1 ⋮ yN+q− 1 ⎥
jointly. On the other hand, some studies have pointed out that the ⎢ ⎥
1 ⎢ y(ref) ⋮ yN+q− 2 ⎥
(ref)
estimated modal parameters are asymptotically normally distributed in = √̅̅̅̅ ⎢ q− 1 y(ref)
q ⎥ (2)
⎢ ⎥
general [25], and the standard deviations of spurious eigenfrequencies N⎢ ⋮
⎣ (ref)
⋮ ⋮ ⋮ ⎥
⎦
are typically 10–100 times larger than those of physical eigen y1 y(ref)
2 ⋮ y(ref)
N
frequencies [26]. Therefore, the standard deviations of the estimated
frequencies can be served as an indicator to distinguish physical modes where y+ is a matrix composed of measurement data obtained by all
from spurious modes [16,27]. sensors; y− is a matrix composed of measurement data obtained by
In this study, a procedure for automatically identifying operational
reference sensors; N is the column number of y+ and y− ; q and p are the
modal parameters of bridges based on reference-based covariance-
user-specified quantities with N + p + q denoting the total length of the
driven stochastic subspace identification algorithm (SSI-cov/ref) is
measurement data. Theoretically, N + p + q should be infinite.
proposed with the following three stages: (1) modal parameter identi
The subspace matrix Hcov , a block Hankel matrix, is constructed from
fication and frequency uncertainty calculation using a range of model
the finite length of the measurement data given by
orders, (2) spurious modes elimination using frequency uncertainty and ⎡ ⎤
stabilization criteria, and (3) density-based clustering for distinguishing R1 R2 ⋯ Rq
physical modes. For the benchmark measurement data of bridges ⎢
Hcov = y+ (y− )T = ⎢
R2 R3 ⋯ Rq+1 ⎥ ⎥ (3)
⎣ ⋮ ⋮ ⎦
adopted in this study, the uncertainty sources might arise from mea ⋮ ⋱
surement noise, window length of samples, and nonstationary excita Rp+1 Rp+2 ⋯ Rp+q
tion. In practice, the preprocessing techniques can be adopted to ( )
partially reduce the adverse effects caused by the uncertainty sources, in which Rk = E yk yk− i ∈ Rr×r0 with k = 1, 2, ⋯, p +q are the
(ref)T
such as detrending, removing outliners, pre-filtering, resampling, etc. covariance matrices between all sensors and reference ones, and E( )
Nevertheless, due to limited research related to the application of denotes the expectation operator. It can be testified that Hcov has the
DBSCAN algorithm in automated modal parameter identification, as the following decomposition property [5,28]:
beginning of physical problem investigation, the present study considers
measurement data processing of bridge parameters, including fre (4)
quency, damping ratio, and mode shape. More theoretical study such as
the separation of uncertainty sources and evaluation of individual where Zq is the reversed extended controllability matrix; =
[ T T T
]
source influence on the final uncertainties of the estimated modal pa C (CA) T
⋯ (CA ) p is the extended observability matrix,
rameters requires further attention, which is not the focus of the present which can be determined by conducting the singular value decomposi
study. tion (SVD) on Hcov :
This study is organized as follows: In Section 2, the SSI-cov/ref and [ ][ T ]
uncertainty computation of modal parameters are briefly reviewed. In Σ1 0n×(qr0 − n) V1
Hcov = [ U1 U0 ] (5)
Section 3, the proposed automated modal identification framework is 0[(p+1)r− n ]×n Σ0 V T0
presented in detail with an illustrative numerical example. Two real
bridges with measurement data are adopted for validation in Section 4.
2
Y. He et al. Engineering Structures 255 (2022) 113891
are smaller than the thresholds. As the physical modes represent the
(6)
intrinsic nature of the structure, they are expected to appear in different
with the following definitions: model orders with similar characteristics. In contrast, the modal pa
rameters of spurious modes may vary significantly between two suc
U1 = [ u1 ⋯ un ] ∈ R(p+1)r×n (7a) cessive model orders; hence, they can be largely removed by comparing
the variations of modal parameters with the tolerable thresholds.
V 1 = [ v1 ⋯ vn ] ∈ Rqr0 ×n (7b) Thence, the physical modes can align vertically to form several stabili
zation axes on the frequency-model order plane, while the spurious
Σ1 = diag{σ1 , ⋯, σn } ∈ Rn×n (7c) modes are randomly distributed on this plane. The commonly adopted
measures of relative difference for estimating frequency, damping ratio,
where Σ1 is the diagonal matrix consisting of the first n non-zero singular
and mode shape are defined as follows:
values of the matrix Hcov ; the column vectors of U1 and V 1 are the
corresponding left and right singular vectors. After the SVD, the noise ( ) | fj − fk |
d fj , fk = ( ⃒ ⃒ ⃒ ⃒) < δf (13a)
terms in the measurement data can be partially removed since they are max ⃒ fj ⃒, ⃒ fk ⃒
often related to the small singular values in Σ0 and are discarded. From
the definition of the extended observability matrix, C is the first r rows of ( ) |ξj − ξk |
d ξj , ξk = ( ⃒ ⃒ ⃒ ⃒) < δ ξ (13b)
while A can be obtained by max ⃒ξj ⃒, ⃒ξk ⃒
(8) ⃒{ }
⃒ H
⃒2
⃒
( ) ( ) ⃒ ψ j {ψ k }⃒
d ψ j , ψ k = 1 − MAC ψ j , ψ k = 1 − ({ }H { } )( ) < δψ
where and denote the first and last p × r rows of ; ψj ψj {ψ k }H {ψ k }
3
Y. He et al. Engineering Structures 255 (2022) 113891
(19a)
(19b)
4
Y. He et al. Engineering Structures 255 (2022) 113891
Fig. 4. Cleared stabilization diagrams for simply-supported beam obtained by (a) frequency uncertainty criterion: c.v.f. = 0.02; (b) commonly-used criteria: c.v.f. =
0.02, δf = 0.02, δξ = 0.1, δψ = 0.02, and 0 < ξ < 0.1.
5
Y. He et al. Engineering Structures 255 (2022) 113891
β = 0.25 and γ = 0.5 for unconditional stability, and the time step is
0.001 sec.
The vertical accelerations of the left five nodes (nodes 2, 3, 4, 5, 6) of
the bridge are selected as output signals, in which the second node is
selected as the reference channel since it is not the node of the first five
mode shapes, and the responses of these five modes will show up at this
node. After the assignment, r = 5 and r0 = 1. To simulate the mea
surement noise, the calculated acceleration of the bridge is added by the
zero-mean white noise as
unoi = ucal + Ep Nw σ cal (20)
6
Y. He et al. Engineering Structures 255 (2022) 113891
Fig. 6. Clustered frequencies and damping ratios obtained by (a) K-means clustering algorithm; (b) fuzzy C-means clustering algorithm.
7
Y. He et al. Engineering Structures 255 (2022) 113891
Fig. 7. Comparison of the estimated and theoretical mode shapes of the simply-supported beam: (a)-(e) correspond to the 1st-5th modes.
Fig. 8. Dowling Hall Footbridge: (a) photo; (b) locations of accelerometers (adopted from Ref. [38]).
January 2010 to May 2010 for the research team to collect the data. The reference channels. In the SSI-cov/ref algorithm, p and q are assigned as
original sampling rate was 2048 Hz, but the measurement data were 50 and 51, respectively. To construct the stabilization diagram, the
downsampled to 128 Hz by the test personnel. After some post- model order is chosen to range from n = 2 to n = 100 with the increment
processing, the measurement data were made public on the website of of 2. To compute frequency uncertainty, nb is assigned as 100. Fig. 9(a) is
the School of Engineering [39]. More detailed information about the the stabilization diagram after using the frequency uncertainty criterion,
acquisition system and test data analysis can be found in Refs. [38,40]. i.e. removal of the identified modes having standard deviations larger
The data to be analyzed by this study were measured on January 7, than 2% of the identified frequencies. In this subfigure, the ANPSD is
2010. In the measurement duration of 267 sec., each channel has 34,144 also given for comparison. It is observed that almost all spurious modes
data points. The accelerometers 1, 2, 3, 5, 6, and 7 are selected as the are removed except one spurious mode, which confirms that the
8
Y. He et al. Engineering Structures 255 (2022) 113891
Fig. 9. Stabilization diagram of Dowling Hall Footbridge obtained by (a) frequency uncertainty criterion: c.v.f. = 0.02; (b) commonly-used criteria: c.v.f. = 0.02, δf =
0.02, δξ = 0.1, δψ = 0.02, and 0 < ξ < 0.1.
Fig. 10. Clustered physical modes of Dowling Hall Footbridge on the Fig. 11. Path of cluster numbers by iteration for the estimated frequencies
frequency-damping ratio plane. shown in Fig. 9(b).
9
Y. He et al. Engineering Structures 255 (2022) 113891
Fig. 12. The first six normalized mode shapes of Dowling Hall Footbridge: (a)-(f) correspond to the 1st-6th modes. The black and white circles denote the nodes of
the bridge deck at the supports and the positions of the accelerometers, respectively.
operation on the non-zero elements of the cumulative adjacency matrix [38], in which only small differences are observed between these two
until all of them become zeros. (3) Finding the optimal cluster number results. It is pointed out that no comparison is made for the estimated
from the iteration-cluster number graph, i.e. the cluster number appearing bridge damping ratios as the estimated bridge damping ratios vary
most frequently. For the estimated bridge frequencies as shown in Fig. 9 largely among different measurement data [38]. The identified mode
(b), the corresponding path of cluster numbers by iteration is displayed shapes are depicted in Fig. 12, where the first, second, fifth, and sixth
in Fig. 11. Obviously, this algorithm fails to give a unique optimal cluster modes are the vertical flexural modes while the third and fourth modes
number because both cluster numbers 6 and 14 appear most often (i.e. are the torsional modes.
four times) in this example. In contrast, the DBSCAN clustering algo
rithm generates a unique and deterministic result. 4.2. Example 2: Old ADA bridge
A representative mode of each cluster is chosen to be the one with the
median value of the identified damping ratios. In Table 2, the estimated The Old ADA bridge is a simply-supported steel truss bridge located
frequencies obtained by the present method are compared with the data in the Nara Prefecture, Japan, as shown in Fig. 13(a). Its main span
10
Y. He et al. Engineering Structures 255 (2022) 113891
Fig. 13. Old ADA bridge: (a) photo; (b) locations of accelerometers. (adopted from Ref. [44]).
Fig. 14. Stabilization diagram obtained by (a) frequency uncertainty criterion: c.v.f. = 0.05; (b) commonly-used criteria: c.v.f. = 0.05, δf = 0.02, δξ = 0.1, δψ =
0.02, and 0 < ξ < 0.1.
length and width are 59.2 m and 3.6 m, respectively. The bridge was
constructed in 1959 and demolished in 2012. Before the removal of this
bridge, a series of ambient and vehicle-induced vibration tests were
conducted to investigate the effects of artificially induced damages on
the changes of modal parameters. There were five scenarios subjected to
different damages, in which the first scenario had no artificially induced
damage and was regarded as the reference in comparison with the other
four damage-involved scenarios. The eight accelerometers were
installed on the bridge deck to record the dynamic responses of the
bridge. To reach better measurement resolution of damage-induced
modal parameters, five accelerometers were installed on the damaged
side while only three accelerometers were deployed on the non-
damaged side. The locations of accelerometers are shown in Fig. 13
Table 3
Comparison of the estimated frequencies of the Old ADA bridge.
Mode Ref. [43] (Hz) Proposed method (Hz) Relative error (%)
11
Y. He et al. Engineering Structures 255 (2022) 113891
Fig. 16. The normalized mode shapes of the Old ADA bridge: (a)-(e) correspond to the 1st-5th modes.
Fig. 17. (a) Stabilization diagram obtained by the criteria: c.v.f. = 0.02, δf = 0.02, δξ = 0.1, δψ = 0.02, and 0 < ξ < 0.1; (b) clustered physical modes on the
frequency-damping ratio plane.
(b). More details regarding the experimental process and test data The sampling rate is 200 Hz, and each channel has 75,964 data points.
analysis can be found in Refs. [43,44]. To construct the stabilization diagram, p and q are assigned as 59 and 60,
The measurement data adopted in this study were obtained from the respectively; the model orders in consideration range from n = 2 to n =
ambient vibration test (three-time tests), the undamaged scenario [44]. 60 with an increment of 2. The accelerometers 2, 3, and 4 are chosen as
12
Y. He et al. Engineering Structures 255 (2022) 113891
13
Y. He et al. Engineering Structures 255 (2022) 113891
[27] Döhler M, Mevel L. Efficient multi-order uncertainty computation for stochastic [35] Ren W-X, Peng X-L, Lin Y-Q. Experimental and analytical studies on dynamic
subspace identification. Mech Syst Signal Process 2013;38(2):346–66. characteristics of a large span cable-stayed bridge. Eng Struct 2005;27(4):535–48.
[28] Xie Y, Liu P, Cai G-P. Modal parameter identification of flexible spacecraft using [36] Cara FJ, Juan J, Alarcón E, Reynders E, De Roeck G. Modal contribution and state
the covariance-driven stochastic subspace identification (SSI-COV) method. Acta space order selection in operational modal analysis. Mech Syst Signal Process 2013;
Mech Sin 2016;32(4):710–9. 38(2):276–98.
[29] Yang X-M, Yi T-H, Qu C-X, Li H-N, Liu H. Modal identification of high-speed [37] Paz M, Kim YH. Structural dynamics: theory and computation. 6th ed. New York:
railway bridges through free-vibration detection. J Eng Mech 2020;146(9): Springer Science & Business Media; 2012.
04020107. https://doi.org/10.1061/(ASCE)EM.1943-7889.0001847. [38] Moser P, Moaveni B. Design and deployment of a continuous monitoring system for
[30] Döhler M, Hille F, Mevel L, Rücker W. Structural health monitoring with statistical the Dowling Hall Footbridge. Exp Tech 2013;37(1):15–26.
methods during progressive damage test of S101 Bridge. Eng Struct 2014;69: [39] https://engineering.tufts.edu/cee/shm/research_BM_continuousMonitoring.asp.
183–93. [40] Moser P, Moaveni B. Environmental effects on the identified natural frequencies of
[31] Su L, Huang X, Song M-L, Michael LaFave J. Automatic identification of modal the Dowling Hall Footbridge. Mech Syst Signal Process 2011;25(7):2336–57.
parameters for structures based on an uncertainty diagram and a convolutional [41] Hai-peng C, Xuan-Jing S, Ying-da Lv, Jian-Wu L. A novel automatic fuzzy
neural network. Structures 2020;28:369–79. clustering algorithm based on soft partition and membership information.
[32] Golub GH, Loan CFV. Matrix computations. 4th ed. Baltimore: The Johns Hopkins Neurocomputing 2017;236:104–12.
University Press; 2013. [42] Mok PY, Huang HQ, Kwok YL, Au JS. A robust adaptive clustering analysis method
[33] Pintelon R, Guillaume P, Schoukens J. Uncertainty calculation in (operational) for automatic identification of clusters. Pattern Recognit 2012;45(8):3017–33.
modal analysis. Mech Syst Signal Process 2007;21(6):2359–73. [43] Chang K-C, Kim C-W. Modal-parameter identification and vibration-based damage
[34] Ml B, Benveniste A, Goursat M, Hermans L, Mevel L, Van der Auweraer H. Output- detection of a damaged steel truss bridge. Eng Struct 2016;122:156–73.
only subspace-based structural identification: from theory to industrial testing [44] Kim C-W, Zhang F-L, Chang K-C, McGetrick Patrick J, Goi Y. Ambient and vehicle-
practice. J Dyn Syst Meas Control 2001;123:668–76. induced vibration data of a steel truss bridge subject to artificial damage. J Bridge
Eng 2021;26:04721002.
14