buildings

Article
Detection of Structural Damage in a Shaking Table Test Based
on an Auto-Regressive Model with Additive Noise
Quanmao Xiao 1 , Daopei Zhu 1, *, Jiazheng Li 2 and Cai Wu 2,3, *

1 School of Civil and Surveying and Mapping Engineering, Jiangxi University of Science and Technology,
Ganzhou 341000, China; 6720211199@mail.jxust.edu.cn
2 Changjiang River Scientific Research Institute, Changjiang Water Resources Commission,
Wuhan 430010, China; lijz@mail.crsri.cn
3 School of Civil Engineering, Hubei Engineering University, Xiaogan 432000, China
* Correspondence: daopei.zhu@jxust.edu.cn (D.Z.); wucai@whut.edu.cn (C.W.)

Abstract: Damage identification plays an important role in enhancing resilience by facilitating precise
detection and assessment of structural impairments, thereby strengthening the resilience of critical
infrastructure. A current challenge of vibration-based damage detection methods is the difficulty of
enhancing the precision of the detection results. This problem can be approached through improving
the noise reduction performance of algorithms. A novel method based partially on the errors-in-
variables (EIV) model and its total least-squares (LS) algorithm is proposed in this study. Compared
with a classical damage detection approach involving adoption of auto-regressive (AR) models and
the least-squares (LS) method, the proposed method accounts for all the observation errors as well
as the relationships between them, especially in an elevated level of noise, which leads to a better
accuracy. Accordingly, a shaking table test and its corresponding finite element simulation of a
full-scale web steel structure were conducted. The acceleration time-series output data of the model
after suffering from different seismic intensities were used to identify damage using the presented
detection method. The response and identification results of the experiment and the finite element
analysis are consistent. The finding of this paper indicated that the presented approach is capable of
detecting damage with a higher accuracy, especially when the signal noise is high.
Citation: Xiao, Q.; Zhu, D.; Li, J.; Wu,
C. Detection of Structural Damage in Keywords: structure damage identification; autoregressive model; shaking table test; total least-
a Shaking Table Test Based on an squares; web steel structure
Auto-Regressive Model with
Additive Noise. Buildings 2023, 13,
2480. https://doi.org/10.3390/
buildings13102480 1. Introduction
Academic Editor: Marco Structures are subjected to diverse types of adverse factors such as earthquakes,
Di Ludovico hurricanes, explosions, and overcapacity loads, leading to evitable damages [1]. Damage
identification is crucial for enhancing resilience through timely detection and assessment
Received: 25 August 2023
of impairments in systems [2]. By integrating advanced sensing technologies, data analysis,
Revised: 15 September 2023
Accepted: 27 September 2023
and decision support systems, people can make informed decisions, optimize resource
Published: 29 September 2023
allocation, and enhance the resilience of critical infrastructure in the face of unforeseen
events or operational challenges [3,4]. Aiming at enhancing structural resilience, many
techniques have been used to study structural performance after damage. Over the past
several decades, damage detection methods such as the finite difference method, finite
Copyright: © 2023 by the authors. element (FE) simulations and analysis, and experiments such as shaking table model
Licensee MDPI, Basel, Switzerland. tests [5–8] have been reported in the literature. Research on the shaking table test and the
This article is an open access article corresponding FE analysis has reported on the dynamic responses and seismic performance
distributed under the terms and of structures [9]. Based on these results, the engineering community has also explored
conditions of the Creative Commons damage detection methods [5], which have been used to deal with observational and output
Attribution (CC BY) license (https:// data such as displacements, accelerations, and the natural frequencies of shaking table tests
creativecommons.org/licenses/by/ and FE simulations.
4.0/).

Buildings 2023, 13, 2480. https://doi.org/10.3390/buildings13102480 https://www.mdpi.com/journal/buildings

Buildings 2023, 13, 2480 2 of 18

The vibration-based damage evaluation method has received considerable attention

for its noninvasive and uncomplicated characteristics [10,11], especially popular in the
shaking table test model. The mechanism of these approaches involves the observation of
the structural mode, frequencies, or time series output over time, the extraction of sensitive
damage features, and the evaluation of structural damages [12]. Vibration-based damage
detection can be divided into frequency-domain-based methods and time-domain-based
methods and their combination. As for frequency-domain based methods, He et al. [13]
carried out shaking table tests on two concrete shear wall model structures and the test
of complete uniaxial compressive constitutive curves of concrete. Moaveni et al. [14] con-
ducted a full-scale shaking table test of a reinforced concrete building section and found
that three damage identification methods used in the experiment did not exactly coincide.
As for time-domain-based methods, Li et al. [15] studied the performance of a high-rise
building through the shaking table test and its corresponding FE simulations. The damage
levels and distributions of the model building were successfully identified based on the
observations in the experiment. In addition, some research investigated effective damage
detection methods including vibration-based methods, and they have been successfully
applied to detecting and identifying the damages in structures constructed from vari-
ous types of materials, such as reinforced concrete [16], steel [17], and composites [18].
Morita et al. [19] detected and estimated the damage of steel frames through a shaking
table test by two identification methods.
One of the most commonly used time-domain-based techniques, a method based
on the auto-regressive (AR) model, has also been applied in a variety of engineering
applications [20]. The AR model can account for correlations between the current time
parameter with its predecessors in the time series [21]. In this model, the output variable
depends linearly on its previous values and a stochastic term, which means errors only exist
in the current observations. Damage levels in a structure can be identified by AR coefficients
between the undamaged structure and the structures to be detected [22]. The famous least-
squares (LS) method is one of the most commonly used methods for estimating parameters
such as the structural damage indicators in the AR model [23].
However, some research has pointed out that the classical AR model ignored the
errors of previous observations, which introduced errors in the evaluated AR parameters,
and led to biased results [24]. Zeng [25] proved and quantified the negligible differences
in the estimated parameters caused by the LS solution and the classical AR model. To
deal with this defect, a large variety of bias-compensated LS methods have recently been
proposed in the past few years. For instance, Diversi et al. [26] proposed a new and effective
identification method considering the AR model with additive noise. Esfandiari et al. [27]
presented four new methods for estimating the parameters of an autoregressive (AR)
process based on observations corrupted by white noise. One can notice that the AR model
with noise is actually the well-known errors-in-variables (EIV) model in the field of geodesic
surveys [28]. Among the various solutions for the EIV model, total least-squares (TLS)-
based algorithms indicate better accuracy compared with the classical LS solution [29].
Hence, new parameter estimation methods based on the AR model plus noise, based on
the extensive achievements in geodesic surveys, may potentially be applied to the field of
structural damage detection.
To overcome the current limitations of damage identification results based on AR
models, which are greatly affected by environmental interference and prove ineffective in
practical cases, this paper aims to propose a high anti-noise structural damage identification
method that considers all errors in observations. Furthermore, the application effects of
this method in real cases are thoroughly examined. In the proposed structural damage
detection method, a TLS solution based partially on EIV model theory is applied to identify
the parameters of the AR model with noise. A full-scale shaking table test of a two-story
web steel structure is conducted to study its seismic performance. The proposed damage
detection method is used to further identify the damage levels of the model under different
earthquake intensities. The structure of this paper can be listed as follows. In Section 2, an
Buildings 2023, 13, 2480 3 of 18

AR model with noise and its parameter estimation method are introduced. This method
considers not only the noise in current observations but also in past observations, and a
sensitive damage indicator is also adopted to represent damage levels. In Section 3, the
performance of the estimation method is analyzed based on a mathematical simulation.
Section 4 introduces the shaking table test and the corresponding FE simulation of the web
steel structure. The results of the experiment, FE simulation, and damage detection are
presented in Section 5. This study not only indicates that the light steel structure performs
well in earthquakes, but also proves that the proposed damage identification method is
effective in practice.

2. LS Adjustment for the AR Model with Additive Noise

2.1. AR Model with Additive Noise
A classical AR model of order m is described by:

xt = β 1 xt−1 + β 2 xt−2 + . . . + β m xt−m + ext , (1)

where xt is the discrete-time signal of the acceleration responses; ext is the possible ran-
dom noise in xt ; m is the model’s order, it varies from 0 to (t − 1); β i (i = 1, 2, · · · , m) is
unknown coefficient in the AR model and it is calculated by corresponding algorithms. In
a practice conditions, when a structure operates under normal conditions, the noise ext can
be assumed as the famous Gaussian white noise, this noise has a zero-mean and unknown
variance [30].
In fact, the left term xt of the model can be seen as the sum of the two terms on the right.
The first term considers xt−1 to xt−m with unknown coefficients, while the second term
represents the noise affection. Denote y = [ xt , xt−1 , · · · , xt−m+1 ]T , β= [ β 1 , β 2 , · · · , β m ]T ,
and ey = [ext , ex(t−1) , · · · , ex(t−m+1) ]T . The AR model can be represented as:

x t −1 , x t −2 , · · · , x t − m eyt
      
xt β1
 x t −1   x t −2 , x t −3 , · · · , x t − m −1  β2   e y ( t −1) 
y= = +  = Aβ + ey , (2)
      
.. ..  .. ..
 .   .  .   . 
x t − n +1 x t − n , x t − n −1 , · · · , x t − n +1− m βm e y ( t − n +1)
where ey is the error corresponding to y. Since xt is obtained by observation in practice,
then vector y as well as the matrix A should contain errors. In this regard, the existing
errors in matrix A are neglected as a simplification in the classical AR model. However,
errors should be added to matrix A in Equation (2), represented by E A . And the AR model
can be rewritten as:
y=(A−E A )β+ey . (3)
subject to
−1
 E(e) = 0, D (e) = σ02 W


Wy WyA , (4)
 W=
W Ay ω
where Wy and ω are diagonal weight matrices of y and a, respectively. Here, a = vec(A),
which is the vector of putting the elements of A into a vector one column after another;
ea = vec (E A ) and e = [ey , ea ]T .
Equation (4) is the well-known EIV model. Notice that in a special case when the matrix
A contains no errors (E A = 0), then the model is the classical AR model in Equation (2) with
only noise ey . By comparing Equations (2) and (3), we can observe that Equation (3) includes
an error term, EA , corresponding to the matrix A. EA represents all the errors present in the
elements of matrix A. Since the elements of A are the previous observations, it is certain that
they would contain errors caused by various factors such as manual operations, machine
precision, and environmental influences. However, Equation (2) dismisses these errors.
Consequently, these errors, combined with the algorithm based on Equation (2), can result
Buildings 2023, 13, 2480 4 of 18

in inaccuracies in the calculated coefficients. Particularly in situations with high noise levels,
neglecting EA can lead to significant errors in parameter estimation, causing substantial
deviations in the identification factor from the actual scenario.
As mentioned in the introduction, studies of the EIV model have been quite popular in
the field of geodesic surveys and a lot of valuable results have already been reported. An ex-
tended, more general, EIV model called the partial EIV model proposed by Xu et al. [31,32]
have been proposed. Not all the elements in the design matrix A are random in the partial
EIV model. This model is represented as follows:
  
 y = βT ⊗ In h + B¯

a + ey

, (5)
 a=¯

a + ea
 
¯
In Equation (5) h + Ba = vec(A − E A ); h is a nm × 1 deterministic constant vector
whose elements are composed of non-random elements of vec (A−E A ); ⊗ is the symbol of
¯
the Kronecker product; a is a vector with the size of t × 1, and with independent random
entries in the design matrix A, it is also the true value of a; B is a given matrix with the size
¯
of nt × m, it depends on random element numbers in the matrix A; and B a is a vector that
represents the random component.
It is clear that the AR model with additive noise is a special case of the partial EIV
model. When t = nm and h = 0, Equation (5) reduces to Equation (3), which is shown
as follows:  
y = βT ⊗ In Ba + ey = (A − E A )β + ey . (6)

The observations utilized for damage detection can be acquired from the same types
of time-series. For example, accelerations are obtained by the same equipment in the same
testing sites. Hence, the diagonal weight matrix W in Equation (4) is assumed as the unit
matrix. By evaluating the unknown vector β of the partial EIV model, the coefficients of
the AR model with additive noise can be obtained.

2.2. Existing TLS Solutions for the EIV Model

The solution for Equation (6) can be described as follows. Firstly, assuming that ea and
ey are stochastically independent, that is,

cov(ey , ea ) = 0, (7)

   
Wy WyA Wy 0
W= = . (8)
W Ay ω 0 ω
The TLS solution is,
( ! )
ˆ  −1 m
¯

a= ω + STβ WSβ ωa − STβ W ∑ hi β̂i + STβ Wy , (9)
i =1

(Nh + NB + NBh + NhB )β̂ = µh + µB , (10)

where µh = γh Wy, µB = γB Wy, and h = [h1 , h2 , . . . , hm ]T , B = [B1 , B2 , · · · , Bm ]T ,

#T
ˆ
" T
ˆ T ¯ˆ T T ˆT T
!
¯ ¯ m ¯ h iT
γB = a B1 , a B2 , · · · , a Bm , Sβ = ∑ Bi β̂i = β ⊗ In B, γh = hT1 , hT2 , · · · hTm .
i =1
Buildings 2023, 13, 2480 5 of 18

Nh ,NB ,NBh ,NhB are m × m matrices, for i, j = 1, 2, · · · , m, these matrices are, respectively,
given by the following:
Nh (i, j) = hT Wh j


ˆT

ˆ
 NB (i, j) = ¯ ¯


a BTi WB j a


ˆT T . (11)
¯
N ( i, j ) = a B Wh

j
 Bh


 i
 ˆ
¯
 T
NhB (i, j) = h j WB j a
Based on this solution, an optimistic solution for the partial EIV model can be consid-
ered as an optimization problem. Thus, the cost function is rewritten as follows:

T ¯ T ¯
¯ ¯ ¯
min : S( a, β) = ( a − a) ω( a − a) + (Aβ − y) W(Aβ − y). (12)

Finally, the solution is as follows:

  −1
ˆ T ˆ ˆ T
¯ ¯ ¯
β̂ = A WA A Wy, (13)

ˆ
¯
a = a + ω−1 STβ E−1 y − Aβ̂ .


(14)

where E = W−1 + Sβ ω−1 STβ . The final solution can be obtained through an iterative process.
The new solution may be more concise and straightforward compared to formulas (9) and
(10). Specifically, the alternative approach is simpler and can be processed more efficiently
when the number of independent random elements in the design matrix A is considerably
larger than the number of measurements.

2.3. TLS Solution for the Modified AR Model

It is clear that in the AR model plus noise, the same elements occur in E A and ey ,
which means that they should be not stochastically independent. Thus, simply assuming
W = diag(Wy , ω) is not reasonable. This may lead to the solution above (Equations (13)
and (14)) not being able to be used directly. In this work, this solution is still used, named
as TLSp . The errors of xi in the ey are not assumed to be the same as the errors of xi in
the E A , i = (t − n + 1, · · · , t). That is, eyi is not always the same as ei . Then, the solution
above (Equations (13) and (14)) can be used for solving the AR model plus noise. Although
eyi should be the same as ei in true cases, the actual value of errors in xi is unknown.
Thus eyi and ei may be closer to true errors. In this regard, it is more appropriate to
assume that eyi and ei are not the same, and we should estimate them independently rather
than disregarding these errors in matrix A. Under these assumptions, the TLSp parameter
estimation steps of the AR model with additive noise are shown as follows,
1. Given A and y, W = I(n+1)m , h = 0;
ˆ
¯
2. Initialize a = a;
3. Compute β̂ by Equation (13);
¯ˆ
4. Compute a by Equation (14) based on the obtained β̂ in Step 3.
ˆ
¯
5. Given a predetermined tolerable error value, if the errors between a and β̂ are within
the given value, terminate the estimation. Otherwise, go to Step 3.
The AIC criterion is used in this paper to determine the optimal order m of the AR
model, formulated as [33]:
AIC(n) = ln σ̂2a (n) + 2n/N (15)
Buildings 2023, 13, 2480 6 of 18

where σ̂2a is the estimated variance of the residual errors of order n. N is the total number
of samples.

2.4. Damage Detection Indicator

After obtaining the parameter β of the AR model, a damage indicator can be intro-
duced to indicate the damage degree of the structure. The differences in the indicators
between β in the healthy conditions with those of the damaged conditions cannot be mea-
sured simply by visual inspection, especially where numerous elements exist in β. In this
paper, the structural damages are represented by the ratio between the Euclidean distance
of the undamaged β with the damaged β. The calculation steps are clarified as follows:
1 The obtained response acceleration time-series data are divided into two parts, i.e.,
part A0 and part B0 , where A0 is used as the baseline data and B0 serves as the
unknown data to be estimated when there is damage to the structure.
2 Estimating βA0 = [β A0,1, β A0,2, · · · , β A0,m ]T and βB0 = [β B0,1, β B0,2, · · · , β B0,m ]T through
Equations (12) and (13). The square of the Euclidean distance between β A0 with βB0
is calculated as:

m
D0 = ∑ [( β B0,j − β A0,j )]2 . (16)
j =1

3 Estimating the βi = [αi1 , αi2 , · · · , αim ] T of ith output data of the damaged structure
through Equations (12) and (13). The Euclidean distance between β i and β B0 is
calculated as,

m
Di = ∑ [( βi,j − β A0,j )]2 . (17)
j =1

4 Finally, the damage indicator, named IF in this paper, is calculated as the ratio between
Di and D0 , as follows:

Di
IF = . (18)
D0
The IF should be close to 1 when the data to be estimated are acquired from an
undamaged structure. As the damage level of the structure increases, the differences
between the parameters of the undamaged structure and the damaged structure should
increase, resulting in an increase in the IF value.

3. Performance Analysis
In this section, a mathematical simulation is used to analyze the performance of the
damage detection method presented above. The estimation results of the adopted TLS
method in Equations (13) and (14), and AR model with noise are studied carefully. A
comparison between the traditional AR model with the corresponding LS solution and the
AR model and the addition of noise with its TLS solution is demonstrated.
Taking the following 4th-order AR model as an example [26],

xt = 2.4xt−1 − 3.03xt−2 + 1.986xt−3 − 0.6586xt−4 + et , (19)

where et is Gaussian white noise, i is the time sequence number. e(t) = [et , et−1 , · · · ]T ,
and E e2 (t) = 1. The number of samples is limited to 200, and the Gaussian white noise


series with different signal-to-noise ratios (SNR) are applied to the real time-series values
xi (i = 1, 2, · · · , 200). Six SNR conditions, 60 dB, 50 dB, 40 dB, 30 dB, 20 dB, and 10 dB, are
simulated. Specifically, for each SNR, the elements in the noise series are the same. The real
value of xi and xi with the error conditions of 30 dB and 10 dB are shown in Figure 1. The
obtained β for different SNR conditions is listed in Table 1.
 and E  e (t )  = 1 . The number of samples is limited to 200, and the Gaussian white noise
series with different signal-to-noise ratios (SNR) are applied to the real time-series values
xi (i = 1,2,,200) . Six SNR conditions, 60 dB, 50 dB, 40 dB, 30 dB, 20 dB, and 10 dB,
are simulated. Specifically, for each SNR, the elements in the noise series are the same.
Buildings 2023, 13, 2480 The real value of xi and xi with the error conditions of 30 dB and 10 dB are shown 7inof 18
Figure 1. The obtained β for different SNR conditions is listed in Table 1.

(a) (b) (c)

Figure 1. Output time series: (a) real value; (b) SNR is 30 dB; (c) SNR is 10 dB.
Figure 1. Output time series: (a) real value; (b) SNR is 30 dB; (c) SNR is 10 dB.
Table 1. Values of estimated parameters.
Table 1. Values of estimated parameters.
SNR/dB NO 60 50 40 30 20 10
β
SNR/dB
1 2.4 NO 2.4001 60 2.3966 50 2.3942 40 2.3714 30 2.335720 2.201710
β1 2.4 2.4001 2.3966 2.3942 2.3714 2.3357 2.2017
β2 β −3.03 −3.03 −3.0298−3.0298 −3.0233
−3.0233 −3.0224
−3.0224 −2.9820
−2.9820 −2.9456
−2.9456 −2.5086
−2.5086
2
LSLS
β3 1.986 1.9859 1.9793 1.9795 1.9452 1.8973 1.4482
β3 1.986−0.6586 1.9859−0.6584 1.9793
β4 −0.6557 1.9795
−0.6562 1.9452
−0.6486 1.8973
−0.6162 1.4482
−0.4064
β 4 β1 −0.6586 2.4 −0.65842.4001 −0.6557 2.3966 −0.6562
2.3944 −0.6486
2.3736 −0.6162
2.3563 −0.4064
2.2818
β2 −3.03 −3.0298 −3.0234 −3.0228 −2.9860 −2.9840 −2.6567
TLSp β 1 β3 2.4 1.986 2.40011.9859 2.39661.9793 2.3944 1.9799 2.3736
1.9491 2.3563
1.9340 2.2818
1.5899
β4 −0.6586 −0.6584 −0.6557 −0.6564 −0.6500 −0.6301 −0.4600
TLSp β2 −3.03 −3.0298 −3.0234 −3.0228 −2.9860 −2.9840 −2.6567

β3 1.986 When1.9859
n = 4, the 1.9793
AIC meets its1.9799 1.9491
minimum value. 1.9340
When there is no noise,1.5899
the identifi-
T
cation results are real values β = [2.4 − 3.03 1.986 − 0.6586]. More identification results
with different SNR conditions are shown in Table 1. It is indicated that as the SNR = 60 dB,
the solution of the LS estimation and the TLSp estimation are nearly the same and quite
close to the real value, indicating both methods perform well in this condition. However,
the differences between the two methods increase as the SNR rises. As for the condition
of SNR 20 dB, the estimation results of the proposed model and algorithm are closer to
the true value, while the LS solution differs more significantly. When the SNR is 10 dB,
the results from both of the two methods contain large errors, The identification vector of
TLSp , [2.2818, −2.6567, 1.5899, −0.4600], exhibits four elements that are relatively closer
to the actual values compared to the four elements of the LS solution, which are [2.2017,
−2.5084, 1.4482, −0.4064]. This comparison highlights the improved accuracy of the TLSp
method in estimating the true values of the elements. Hence, it can be concluded that the
parameter estimation method proposed in this paper performs better, even in the presence
of strong noises.
It is clear that, as the SNR increases, the gaps between parameters of healthy and
estimated stages become larger. Assuming that the changes in the parameters in Table 1 are
not caused by different levels of errors, but by the different levels of inner damage of the
system. For example, assume βT = [2.3563 − 2.9840 1.9340 − 0.6301] (SNR = 20 dB, TLSp )
is estimated by the output signals of a damaged condition of a structure by the proposed
method. Then, in order to measure the structural damage level, the damage indicator needs
to be calculated to quantify the parameter changes to this system. Therefore, different levels
of SNR represent different levels of gaps between the parameters of the healthy system,
which can be estimated. The damage levels represented by the IF values of structures are
expected to increase as the SNR values increase. The IFs of the example in Table 1 are
shown in Table 2.
Buildings 2023, 13, 2480 8 of 18

Table 2. IFs of different conditions.

SNR/dB 60 50 40 30 20 10
LS 2.22 2.20 × 10 2.79 × 102 9.80 × 102 4.18 × 103 1.33 × 105
TLSp 2.16 2.18 × 10 2.50 × 102 8.20 × 102 1.50 × 103 6.99 × 104

It is clear that the increases in the IFs values are related to the rising SNR, reflecting
the effectiveness of the damage indicator. As for the same SNR, the IFs of the LS solution
are larger than that of the TLSp solution. For example, when the SNR = 10 dB, the IFs of the
LS solution are approximately twice as large as the IFs of the TLSp solution. Therefore, the
TLSp solution outperforms the traditional LS solution.

4. Experimental and Numerical Techniques

A shaking table test of a two-story full-scale web steel model under seismic excitation
with increasing intensities and the corresponding FE simulation is presented in this section.

4.1. Shaking Table Model Test

A full-scale model of a two-story web steel structure was subjected to a shaking table
model
Buildings 2023, 13, x FOR PEER REVIEW test to investigate its seismic performance. The web steel originating from Canada
9 of 20
Buildings 2023, 13, x FOR PEER REVIEW 9 of 20
is a type of cold-formed steel with a thin wall which has already been utilized in various
countries. The model used in the experiment has dimensions of 2 m × 3 m and a total height
of 6 m. Accelerations
countries. The model used andindisplacements
the experimentwere measured using
has dimensions of 2 mthe× 3DASP2003
m and a totaldynamic
countries.
signal
height TheAccelerations
model
ofacquisition
6 m. andused in the
analysis
and experiment
system, has dimensions
developed
displacements were of 2 Orient
by the China
measured using mthe× DASP2003
3m and a dy-
Institute total
of Noise
height
and
namic of 6 m.
Vibration.
signal Accelerations
The dynamic
acquisition and displacements
strain
and analysis were using
was obtained
system, developed measured usingOrient
theChina
by the DH3817 thedynamic
DASP2003
Institute dy-
and
of static
namic
testing
Noise signal
and acquisition
instrument.
Vibration. The and
The data analysis system,
acquisition
dynamic developed
instrument
strain was obtainedcan by the
be seen
using China Orient
in Figure
the DH3817 Institute
2, and aand
dynamic of of
photo
Noise
static
the and Vibration.
testing
completed modelThe
instrument. dynamic
The strain was
dataisacquisition
structure shown obtained
3 using
instrument
in Figure can bethe DH3817
seen dynamic
in Figure 2, andand
a
staticof
photo testing instrument.
the completed Thestructure
model data acquisition
is showninstrument
in Figure 3can be seen in Figure 2, and a
photo of the completed model structure is shown in Figure 3

(a) (b)
(a) (b)
Figure
Figure2.2.Data
Dataacquisition instruments:
acquisition (a) accelerometer;
instruments: (b) signal
(a) accelerometer; acquisition
(b) signal device.
acquisition device.
Figure 2. Data acquisition instruments: (a) accelerometer; (b) signal acquisition device.

Figure 3. Photo of the completed model.

Figure 3.
Figure 3. Photo
Photoof
ofthe
thecompleted
completedmodel.
model.
The seismic excitation of the El Centro waves (1940, America), and the Qian’an waves
(1976, The seismic
China) withexcitation of the El conditions
different intensity Centro waves
are (1940, America),
simulated on theand the Qian’an
shaking waves
table model.
(1976,
The China)of
sequence with
thedifferent
experimentintensity conditions
is shown in Tableare
3.simulated
As can beon the there
seen, shaking
aretable model.
five levels
ofThe sequence
tested of the with
earthquakes experiment is shown inpeak
the corresponding Tableground
3. As can be seen, there
acceleration (PGA)are five levels
values. The
of tested
free earthquakes
vibration with the
and the swept corresponding
frequency methodpeak ground acceleration
are conducted (PGA) values.
before the seismic The
test. Fur-
free vibration
thermore, whiteand the excitation
noise swept frequency method
is utilized areand
before conducted
after allbefore
of thethe seismic
groups of test. Fur-
seismic
thermore, white noise excitation is utilized before and after all of the groups of seismic
Buildings 2023, 13, 2480 9 of 18

The seismic excitation of the El Centro waves (1940, America), and the Qian’an waves
(1976, China) with different intensity conditions are simulated on the shaking table model.
The sequence of the experiment is shown in Table 3. As can be seen, there are five levels
of tested earthquakes with the corresponding peak ground acceleration (PGA) values.
The free vibration and the swept frequency method are conducted before the seismic test.
Furthermore, white noise excitation is utilized before and after all of the groups of seismic
activities to capture the output shifts due to the seismic induced damage. The sampling
time interval is 3.905 ms. The seismic waves used in the experiments were unidirectional
input excitations, with the peak accelerations corresponding to the seismic intensity values
specified in the Chinese Code [34].

Table 3. Sequence of the shaking table test.

Sequence Number Input Seismic Wave PGA (g)

1 Free vibration
2 Sweeping (0–30 Hz) 0.02
3 Sweeping (30–0 Hz) 0.02
4 El Centro wave, Qian’an waves 0.1
Buildings 2023, 13, x FOR PEER REVIEW 5 El Centro wave 0.1510 of 20
6 El Centro wave, Qian’an waves 0.2
7 El Centro wave, 0.27
8 El Centro wave, Taft wave 0.3
8 El Centro wave, Taft wave 0.3

The
The displacement
displacement isisachieved
achievedbybypulling
pulling thethetoptop of the
of the structure
structure withwith a certain
a certain degreedegree
of
of deflection, and then fixing the deflection, followed by an abrupt cutting of the rope, for for
deflection, and then fixing the deflection, followed by an abrupt cutting of the rope,
the
thepurpose
purpose of of assessing
assessingthe thestructure’s
structure’s free
free vibrations.
vibrations. TheThe measured
measured response
response represents
represents
the
the characteristic free vibrations
characteristic free vibrationsofofthe thestructure.
structure. TheThe white
white noise
noise excitation
excitation had had a peak
a peak
acceleration
acceleration of of 0.2
0.2 gg and
andaaduration
durationofof4040s. s.TheThe sweep
sweep frequency
frequency excitation,
excitation, consisting
consisting of of
both
both upward
upward andand downward
downwardfrequencyfrequency sweeps,
sweeps, hadhad peak
peak accelerations
accelerations of 29ofg29forgeach
for each
direction
direction and
and a duration
durationof of40
40s.s.The
Thepeak
peak accelerations
accelerations of the
of the excitation
excitation waves
waves usedused
in thein the
experiment
experiment can be be referred
referredto toininTable
Table3.3.
Five
Five accelerometers
accelerometers were weredistributed
distributed around
around thethe model,
model, withwith
twotwo on top
on the the floor,
top floor,
two on
two on the
the second
second floor
floorandandoneoneon onthe
thefirst floor,
first shown
floor, shown in Figure 4. Three
in Figure displacement
4. Three displacement
sensors and
sensors and eight
eight strain
strain gauges
gaugeswere wereused.
used.The Theaccelerometer
accelerometer distribution is shown
distribution in in
is shown
Figure 4. Point
Figure 4. Point 1 is on the table surface, points 2 and 3 are on the second floor,
on the table surface, points 2 and 3 are on the second floor, and points 4 and points
4 and
and 5 are
5 are ononthe
thetop
topfloor.
floor.

Figure 4.
Figure 4. Accelerometer
Accelerometer distributions.
distributions.

4.2.
4.2. Finite
Finite Element Simulation
Element Simulation
An
An FE
FE simulation
simulation ofofthe
theshaking
shakingtabletabletest
testwaswas also
also conducted
conducted using
using ANSYS.
ANSYS. The The
properties of the steel
properties steelwere
weredetermined
determinedbyby the
theexperiment
experiment andand
the the
nonlinear performance
nonlinear performance
wasconsidered.
was considered. The
Thebeam
beamandandshell
shellelements
elements were
were used to simulate
used to simulatethe the
square steelsteel
square tube tube
andbamboo
and bamboo plywood.
plywood.Properties
Propertiesofofthe
themain
mainsteel components
steel components in the web-steel
in the structure
web-steel structure
areshown
are shown in in Table
Table 4.
4. The
Thesteel
steelstrip
stripandandVVfittings
fittingswere
weresimulated
simulated by bar elements.
by bar The The
elements.
steel strip only sustains tension. Other elements in the model are simulated
steel strip only sustains tension. Other elements in the model are simulated by solid by solid 70. 70.
There were approximately 11,500 nodes in total. The finite element model of the structure
is shown in Figure 5.
Buildings 2023, 13, 2480 10 of 18

There were approximately 11,500 nodes in total. The finite element model of the structure
Figureis4.shown
Accelerometer distributions.
in Figure 5.

4.2. Finite
TableElement Simulation
4. Properties of the main steel components in the web steel structure.
An FE simulation of the shaking table test was also conducted using ANSYS. The
properties of the steel wereElastic
determined by theYield
experiment Tensile performance
and the nonlinear
Modulus Strength
Components
was considered. The beam and shell 2elements were Strength
used2 )to simulate the square Tensibility
steel tube
(N/mm ) (N/mm
(N/mm2 )
and bamboo plywood. Properties of the main steel components in the web-steel structure
Square steel tube 5
× 10and
are shown in Table 4. The steel
2.02strip V fittings299.6
were simulated by330.1bar elements. The19%
Rectangular
steel strip tube tension.
only sustains × 105 elements in279.9
2.42 Other 334.5 by solid 70.19%
the model are simulated
There were Connector 2.45 ×
approximately 11,500 105 in total. The
nodes 361.1 374.3of the structure14%
finite element model
Steel band 1.75 × 10 5 228.4 333.1 35%
is shown in Figure 5.

(a) (b)

Figure 5. The FE model: (a) total model; (b) skeleton model.

The testing conditions were consistent with the experiment. The input waves were
simulated by the tested acceleration data recorded on the table surface. The input waves
were simulated by the tested acceleration data recorded on the table surface. The frame
foundation and nodes at the base of the columns were set as rigid joints. The bound-
ary constraint types for the bottom three nodes of the frame columns were symme-
try/antisymmetry/encastre. The encastre type boundary condition was selected to define
the boundary condition at the bottom of the frame columns. According to the specific
conditions of the structure, the grid size for the beam elements in the structure was 0.6 m,
while the grid size for the shell elements and the solid elements was 0.3 m. The weight
of the model was adjusted by the material density of the floor units to ensure that the
self-weight of the simulation model was the same as that of the shaking table test model.
Firstly, modal analysis was performed to determine the natural frequencies of the
structure based on the vibration tests. The dynamic response of the structure under
earthquake excitations, such as the El Centro wave and the Qian’an wave, was then
obtained by applying these excitations to the finite element model. The natural frequencies
of the structure were calculated through frequency analysis, while the dynamic response
under earthquake excitations was determined using the modal dynamics analysis module
within linear perturbation analysis.

5. Experimental, Numerical, and Identification Results

In this section, the seismic performance of the model structure is studied. Based on
the proposed damage detection method, the damage levels and distributions of the test
model are identified using the output accelerations induced by white noise after different
testing conditions.
Buildings 2023, 13, 2480 11 of 18

5.1. Experimental Results and Analysis

The dynamic characteristics, such as the natural frequency, and the damping ratio
are listed in Table 5. The web steel structure performed well when subjected to seismic
waves in the experiment. No components were cracked or destroyed after all the testing
earthquake waves were applied. All of the steel pipes, connections, and other construction
members appear to remain in good condition. When the accelerations were small, such as
0.1 g, there were no visible responses or obvious sounds occurring. However, the model
responds with squeaky sounds when accelerations increase to 0.3 g, which reflects tension
and occlusion appearing in some joints, such as the connections between the bamboo
plywood and steel pipes and inter-story connections. Therefore, it can be inferred that the
structure has satisfactory energy dissipation capacity when suffering from seismic waves.

Table 5. Dynamic characteristics of different excitations.

Condition First Order (Hz) Second Order (Hz) Damping Ratio

Free vibration 4.42 14.2 1.2%
White noise 4.26 14.4 1.7%
Sweeping (0–30 Hz) 3.62 15.4 1.4%
Sweeping (30–0 Hz) 3.83 13.9 1.5%

Representative peak acceleration and displacement responses of the model subjected

to EI Centro waves and Qian’an waves under different testing conditions are shown in
Table 6. The acceleration time–history curves of the top floor suffering from 0.1 g and 0.3 g
EI Centro waves are shown in Figure 6. Figure 7 shows the relationship between the bottom
shear and the second story displacement corresponding to EI Centro waves.

Table 6. Peak responses of the model under different test conditions.

Peak Displacement (mm) Peak Acceleration (m/s2 )

Test Conditions
Second Floor Top Floor Second Floor Top Floor
El Centro wave
0.41 0.67 0.47 0.61
(0.1 g)
El Centro wave
0.9 1.46 1.14 1.56
(0.2 g)
El Centro wave
1.9 4.32 2.95 2.94
(0.3 g)
Qian’an wave
0.39 0.14 0.32 0.62
(0.1 g)
Qian’an wave
4.1 0.46 1.64 1.05
(0.2 g)

It can be seen in Table 6 that as the earthquake intensities increase, both the peak
accelerations and displacements of the model increase significantly. For example, the peak
displacement of the model responding to the El Centro waves (0.1 g) on the top floor is
0.61 m/s2 , while the response under the El Centro wave (0.3 g) is 2.94 m/s2 . Furthermore,
the peak accelerations of the top floor are always larger than those of the second floor,
reflecting the damage of the top floor may be more significant than that of the second floor.
The influence of El Centro waves is consistently larger than that of the Qian’an waves.
Figure 6 indicates that the peaks and valleys of acceleration responses under different
waves with different intensities are mainly accumulated in the same time instants. It can
be concluded, in Figure 7, that as the earthquake intensity rises, the relationship between
bottom shear and the displacement in the second story is response to the El Centro waves
significantly, which may be due to inner damages in the model.
 Sweeping (30–0 Hz) 3.83 13.9 1.5%

Representative peak acceleration and displacement responses of the model subjected

to EI Centro waves and Qian’an waves under different testing conditions are shown in
Table 6. The acceleration time–history curves of the top floor suffering from 0.1 g and 0.3
Buildings 2023, 13, 2480 12 of 18
g EI Centro waves are shown in Figure 6. Figure 7 shows the relationship between the
bottom shear and the second story displacement corresponding to EI Centro waves.

(a)

Buildings 2023, 13, x FOR PEER REVIEW

(b) (c) 13 of 20

Figure 6. Acceleration time–history curves of the top floor in response to EI Centro waves: (a) 0.1g;
Figure 6. Acceleration time–history curves of the top floor in response to EI Centro waves: (a) 0.1g;
(b) 0.2 g; (c) 0.3 g.
(b) 0.2 g; (c) 0.3 g.

(a) (b)

(c)
Figure 7. Relationship between the bottom shear with the displacement in the second story in re-
Figure 7. Relationship between the bottom shear with the displacement in the second story in
sponse to the El Centro waves: (a) 0.1g; (b) 0.2 g; (b) 0.3 g.
response to the El Centro waves: (a) 0.1g; (b) 0.2 g; (b) 0.3 g.
Table 6. Peak responses of the model under different test conditions.

Peak Displacement (mm) Peak Acceleration (m/s2)

Test Conditions
Second Floor Top Floor Second Floor Top Floor
El Centro wave (0.1 g) 0.41 0.67 0.47 0.61
El Centro wave (0.2 g) 0.9 1.46 1.14 1.56
El Centro wave (0.3 g) 1.9 4.32 2.95 2.94
Qian’an wave (0.1 g) 0.39 0.14 0.32 0.62
Buildings 2023, 13, 2480 13 of 18

5.2. Finite Element Simulation Results

Buildings 2023, 13, x FOR PEER REVIEW The first and second vibration modes are shown in Figure 8. The comparison 14 of 20between

the dynamic characteristics in the experiment and the results in the FE simulation is shown
in Table 7. It can be seen that the experimental and the FE simulation results are consistent
with in
shown each other,
Table 7. Itindicating
can be seenthethat
FE the
simulation is reasonable.
experimental Therefore,
and the FE theresults
simulation FE model
are can be
consistent
used forwith eachstudy.
further other, indicating the FE simulation is reasonable. Therefore, the FE
model can be used for further study.

(a) (b)
Figure 8. Vibration modes of the FE model: (a) first mode; (b) second mode.
Figure 8. Vibration modes of the FE model: (a) first mode; (b) second mode.
Table 7. Dynamic characteristics of the model in the experiment and FE simulation.
Table 7. Dynamic characteristics of the model in the experiment and FE simulation.
Experiment FE Simulation
Vibration Mode
Frequency (Hz)Experiment
Period (s) Frequency (Hz) FE Simulation
Period (s)
VibrationFirst
Mode 4.42
Frequency 0.23 4.87
Frequency 0.21
Direction Period (s) Period (s)
Second 14.2 (Hz) 0.07 15.01 (Hz) 0.067
First 4.42 0.23 4.87 0.21
Direction
The comparison between
Second the peak14.2 displacement and
0.07 acceleration15.01
responses in the
0.067
experiment and the results in the FE simulation is shown in Table 8. The results of the FE
simulation are consistent with results in the experiment. All the results indicate that as the
The comparison between the peak displacement and acceleration responses in the
earthquake intensity rises, the peak displacement and acceleration responses increase.
experiment and the results in the FE simulation is shown in Table 8. The results of the FE
simulation
Table are consistent
8. Peak displacement and with resultsresponses
acceleration in the experiment.
of the model All theexperiment
in the results indicate
and the that
FE as the
earthquake intensity rises, the peak displacement and acceleration responses increase.
simulation.

Second
Table Floordisplacement and acceleration responses Top
8. Peak Floor
of the model in the experiment and the
Test Condi-
Displacement (m/s 2) Acceleration
FE simulation. (m/s 2) Displacement (m/s 2) Acceleration (m/s2)
tion
Experiment FE Experiment FE Experiment FE Experiment FE
0.1 g 0.41 0.5 0.47
Second Floor 0.56 0.67 0.63 Top0.61
Floor 0.69
0.2 g 0.9 Displacement
1.08 (m/s2 ) 1.14 Acceleration
0.92 (m/s2 ) 1.46 Displacement
1.43 (m/s2 ) 1.56 Acceleration
1.71 (m/s2 )
Test Condition
0.3 g 2.7 2.49 2.95 3.07 4.32 3.91 2.94 3.17
Experiment FE Experiment FE Experiment FE Experiment FE
0.1 g 0.41 0.5
The acceleration 0.47
amplification 0.56 represents
factor 0.67 the ratio between
0.63 0.61 accelera-0.69
the peak
0.2 g 0.9 tion values
1.08 1.14
of each story 0.92input accelerations
with the peak 1.46 1.43
at the 1.56As shown1.71
shaking table.
0.3 g 2.7 in Figure 9,
2.49 2.95 3.07 4.32 3.91
the acceleration amplification factors in the experiment and the FE2.94simulation3.17
are similar. Therefore, it can be concluded that the acceleration responses are modest and
indicateThe
the acceleration
satisfactory seismic performance
amplification of the model.
factor represents theThe
ratiostructure
betweenisthe
an peak
assembly
acceleration
structure
values of each story with the peak input accelerations at the shaking table.type
that combines components with self-tapping screw connections. This As of
shown in
connection
Figure 9,isthe
flexible and has aamplification
acceleration strong energyfactors
dissipation capacity,
in the whichand
experiment maythe
be FE
greatly
simulation
beneficial in providing earthquake resistance capacity.
are similar. Therefore, it can be concluded that the acceleration responses are modest and
indicate the satisfactory seismic performance of the model. The structure is an assembly
 Buildings 2023, 13, 2480 14 of 18

structure that combines components with self-tapping screw connections. This type of
Buildings 2023, 13, x FOR PEER REVIEW connectionis flexible and has a strong energy dissipation capacity, which 15
mayof 20
be greatly
beneficial in providing earthquake resistance capacity.

Figure 9. Envelope
Figure of theofacceleration
9. Envelope amplification
the acceleration factor in
amplification response
factor to El Centro
in response to Elwaves.
Centro waves.

5.3.5.3. Damage
Damage Detection
Detection Results
Results
As mentioned
As mentioned before, the damage
before, the damageidentification results results
identification can directly represent
can directly the
represent the
property
propertychanges
changesin the model
in the structure.
model To further
structure. study
To further the the
study inner damage
inner damageof the webweb steel
of the
steel structureunder
structure underdifferent
differentearthquake
earthquakeintensities,
intensities, the theproposed
proposed damage
damage detection
detection method
method in Section 2 is adopted to identify the damage
in Section 2 is adopted to identify the damage of the model. The study of the model. The study also
alsoin-includes
cludes comparisons between the AR model with additive noise and its
comparisons between the AR model with additive noise and its adopted total least-squares adopted total least-
squares
(TLS)(TLS) solution,
solution, as well
as well as the
as the classic
classic ARAR model
model and and least-squares(LS)
least-squares (LS)solution.
solution.
TheTheacceleration responses in the experiment excited by
acceleration responses in the experiment excited by Gaussian white Gaussian white noise are are used
noise
used for detecting the damage levels and distributions of the model
for detecting the damage levels and distributions of the model structure. The test model structure. The test
model
cancan be divided
be divided intointo
twotwo parts.
parts. TheThe lower
lower partis isthe
part thefirst
firststory,
story,which
whichisisbetween
betweenthe table
the table surface and the second floor. The upper part is the second story, which is between
surface and the second floor. The upper part is the second story, which is between the
the second floor and the top floor. The damage levels of the model structure are repre-
second floor and the top floor. The damage levels of the model structure are represented by
sented by the values of the detection indicator presented in Section 2. The structure is
the values of the detection indicator presented in Section 2. The structure is considered in
considered in healthy stage under the white noise excitation before the earthquake wave
healthy stage under the white noise excitation before the earthquake wave tests.
tests.
The IFs of each testing point under different intensities are shown in Figure 10. The
The IFs of each testing point under different intensities are shown in Figure 10. The
results indicate
results indicate thatintensity
that the the intensity
factorsfactors (IFs) increased
(IFs) increased as theintensity
as the seismic seismicrose
intensity
in all rose in
all testing points, indicating that damage to the structure increases
testing points, indicating that damage to the structure increases with the earthquake level. with the earthquake
level. Besides, the gaps between 0.27 g and 0.3 g are relatively
Besides, the gaps between 0.27 g and 0.3 g are relatively small compared to those of other small compared to those of
other
levels levels
(except (except
testing testing
point pointmay
3, which 3, which
be due may be due
to some to somemistakes).
unknown unknownWhichmistakes).re- Which
reflects
flects that thethat the damage
damage identification
identification method can method
be usedcan be used
to detect to detect
structural structural
damage levels damage
levels correctly.
correctly.
The IFs of some accelerometer testing points under the same earthquake intensities
are listed in Figure 11. Testing points 2 and 3 are on the second floor, and 4 and 5 are
on the top floor. It can be concluded that the IFs of the first story are consistently larger
than the IFs of the second story (except for seismic intensity of 0.2 g), indicating more
damage occurs in the top part of the model structure. The damage to the top part increases
more rapidly and easily as the seismic intensity rises. In addition, IF values of the two
symmetrical testing points of the same floor are not the same, the eastern part (testing
points 3 and 5) of the web structure model contains more damage, which shows that the
west part of the model may have performed better. This may be due to construction errors.
Moreover, all the IFs are relatively smaller (in comparison with the IFs in the mathematical
simulation in Section 3), this indicates that the web steel model structure performs well
under earthquake excitations.
 Buildings 2023, 13, x FOR PEER REVIEW 16 of 20

Buildings 2023, 13, 2480 15 of 18

(a) (b)

(c) (d)

Figure 10. IFs ofFigure 10. IFs of different testing points: (a) point 1; (b) point 2; (c) point 3; (d) point 4.
different testing points: (a) point 1; (b) point 2; (c) point 3; (d) point 4.

The differences Thebetween

IFs of some accelerometer
the results obtainedtesting points
by the under the
detection same presented
method earthquakeinintensities
Section 2 and the classical model and its LS solution are shown in Figure 12. It is and
are listed in Figure 11. Testing points 2 and 3 are on the second floor, clear4that
and 5 are on
the top floor. It can be concluded that the IFs of the first story
the IFs of the two methods are quantitatively different. For example, redarding testing are consistently larger than
point 5, the IF of the proposed method is larger than the IF of the LS solution at intensities damage
the IFs of the second story (except for seismic intensity of 0.2 g), indicating more
occurs in the top part of the model structure. The damage to the top part increases more
of 0.27 g and 0.3 g, while being smaller at 0.1 g. Furthermore, the variations in IFs from
rapidly and easily as the seismic intensity rises. In addition, IF values of the two symmet-
the method are greater than that from the LS solution. It is argued that the differences
rical testing points of the same floor are not the same, the eastern part (testing points 3
attributed to the LS solution ignore possible errors in the designed matrix. Regarding the
and 5) of the web structure model contains more damage, which shows that the west part
model of the proposed identification method TLSp , all the observation data in matrix A,
of the model may have performed better. This may be due to construction errors. Moreo-
which consisted of the acceleration time series output in this experiment, are considered
ver, all the IFs are relatively smaller (in comparison with the IFs in the mathematical sim-
as signals withulation
noise.inHowever,
Section 3),the
thisAR model that
indicates doesthenot account
web for thestructure
steel model presence of errors,
performs well under
despite the fact that these errors
earthquake excitations. do exist during experimental operation, originating from
manual operations, unstable sensor fixation, environmental noise, etc. Based on the TLSp
model and its corresponding parameter estimation algorithm, a more accurate parameter
solution should be obtained, thereby resulting in an IF that is closer to the true value.
Hence the proposed damage identification method in this paper can be more suitable in
practice use.
To summarize, the experiment, FE simulation, and damage detection indicate that
the web steel structure can perform well in response to seismic waves. As the earthquake
intensity increases, the response values such as the acceleration and displacement of the
model structure become large. Furthermore, by applying the proposed damage detection
methods, it can be concluded that damage to the model rises with increases in earthquake
intensity. However, none of the damage is significant, due to the good seismic performance
of the web steel structure.
Buildings 2023, 13, 2480
Buildings 2023, 13, x FOR PEER REVIEW 16 of 18 17 of 2

(a) (b)

(c) (d)
Buildings 2023, 13, x FOR PEER REVIEW Figure 11. IFs of different testing points after the same intensities: (a) 0.15 g; (b) 0.2 g; (c)
18 0.27
of 20g; (d
Figure 11. IFs of different testing points after the same intensities: (a) 0.15 g; (b) 0.2 g; (c) 0.27 g;
0.3 g.
(d) 0.3 g.
The differences between the results obtained by the detection method presented i
Section 2 and the classical model and its LS solution are shown in Figure 12. It is clear tha
the IFs of the two methods are quantitatively different. For example, redarding testin
point 5, the IF of the proposed method is larger than the IF of the LS solution at intensitie
of 0.27 g and 0.3 g, while being smaller at 0.1 g. Furthermore, the variations in IFs from
the method are greater than that from the LS solution. It is argued that the difference
attributed to the LS solution ignore possible errors in the designed matrix. Regarding th
model of the proposed identification method TLSp, all the observation data in matrix A
which consisted of the acceleration time series output in this experiment, are considere
as signals with noise. However, the AR model does not account for the presence of error
despite the fact that these errors do exist during experimental operation, originating from
manual
(a) operations, unstable sensor fixation, environmental
(b) noise, etc. Based on the TLS
model and its corresponding parameter estimation algorithm, a more accurate paramete
Figure 12. Comparison
solution
Figure 12. Comparison should the
between between
be thedetection
obtained,
damage damage
therebydetection
resulting
results: results:
(a) in of(a)IF
IFsan IFs of point
that
testing testing5;point
is closer (b) 5;
the(b)
toIFs IFs of
oftrue valu
seismic intensity 0.3 g.
seismic intensity Hence
0.3 g. the proposed damage identification method in this paper can be more suitable i
practice use.
6. Conclusions To summarize, the experiment, FE simulation, and damage detection indicate that
the web steel structure can perform well in response to seismic waves. As the earthquake
This paper presents a highly robust structural damage identification method that can
intensity increases, the response values such as the acceleration and displacement of the
maintain stability even in the presence of high levels of noise. The method overcomes
model structure become large. Furthermore, by applying the proposed damage detection
the current limitations of AR-based damage identification, which is greatly affected by
methods, it can be concluded that damage to the model rises with increases in earthquake
observationintensity.
errors. However,
This paper introduces a modified auto-regressive (AR) model that
none of the damage is significant, due to the good seismic perfor-
accounts for additive noise. Unlike traditional AR models, this modified model takes into
mance of the web steel structure.
consideration all the errors in observations. To solve the modified AR model, a total-least-
squares (TLS) solution is adopted for the partial EIV model. In a mathematical simulation
6. Conclusions
example, the solution of the modified AR model along with its TLS solution is compared
This paper presents a highly robust structural damage identification method that can
to the classic AR model witheven
maintain stability the least-squares
in the presence(LS) solution.
of high levels The results
of noise. Thedemonstrate that the
method overcomes
current limitations of AR-based damage identification, which is greatly affected by obser-
vation errors. This paper introduces a modified auto-regressive (AR) model that accounts
for additive noise. Unlike traditional AR models, this modified model takes into consid-
eration all the errors in observations. To solve the modified AR model, a total-least-squares
(TLS) solution is adopted for the partial EIV model. In a mathematical simulation exam-
Buildings 2023, 13, 2480 17 of 18

the proposed method outperforms the traditional approach, even when dealing with high
levels of noise. Additionally, an effective damage indicator is employed to measure the
differences in AR parameters and quantify the extent of damage to a system. Finally, a
shaking table test and the corresponding finite element (FE) simulation of a full-scale two-
story web steel structure was conducted. The damage identification method proposed in
this paper was also applied to the acceleration responses to further study the damage levels
and distributions of the model structure under different earthquake conditions. It can be
concluded from the results that the web steel structure studied in this paper performs well
and that the damage incurred increases as the earthquake intensity increases. The damage
detection method presented in this paper can detect both damage levels and distribution in
the structure. This damage detection method may be effective for practical applications in
civil engineering.
The proposed detection method shows promise in reducing identification errors
compared to classical detection methods using the AR models and its LS solution. It
exhibits relatively accurate results, even in conditions of high levels of noise. Furthermore,
its performance with the web steel structure affected by earthquake waves can provide a
valuable reference for further studies on similar structures.
However, structural information in a healthy state for comparison is always necessary
when applying this method in practical engineering. This makes the identification impossi-
ble for structures that do not have stored healthy information. It also poses difficulties in
installing sensors and monitoring signals for complex and hazardous structures. Future
research could focus on the development and utilization of long-term stable embedded
sensors for structures to be monitored. This would enable real-time monitoring of structural
health based on the signals from these sensors combined with the damage identification
method proposed in this study. Another thing to address is how to fix the sensors to struc-
tures to reduce the noise. These advancements would provide new avenues for tracking
and identifying techniques in civil engineering.

Author Contributions: Methodology, Q.X. and D.Z.; Validation, Q.X.; Formal analysis, Q.X.; Investi-
gation, J.L.; Resources, C.W.; Data curation, C.W.; Writing—original draft, D.Z.; Writing—review &
editing, D.Z.; Project administration, J.L. and C.W. All authors have read and agreed to the published
version of the manuscript.
Funding: The authors would like to express their gratitude to Natural Science Foundation of Jiangxi
Province, China (No. 20232BAB214074) and the Central Non-Profit Scientific Research Fund for
Institutes of China (No. CKSF 2021431/CL).
Data Availability Statement: All data supporting the findings of this study are available from the
corresponding author upon reasonable request, but all personally identifiable information will be
identified before transferal.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Sarah, J.; Hejazi, F.; Rashid, R.S.; Ostovar, N. A Review of Dynamic Analysis in Frequency Domain for Structural Health
Monitoring. IOP Conf. Series Earth Environ. Sci. 2019, 357, 012007. [CrossRef]
2. Xiao, F.; Sun, H.; Mao, Y.; Chen, G.S. Damage identification of large-scale space truss structures based on stiffness separation
method. Structures 2023, 53, 109–118. [CrossRef]
3. Huynh, N.T.; Nguyen, T.V.; Nguyen, Q.M. Optimum Design for the Magnification Mechanisms Employing Fuzzy Logic-ANFIS.
Comput. Mater. Contin. 2022, 12, 5961–5983.
4. Huynh, N.T.; Nguyen, T.V.; Tam, N.T.; Nguyen, Q.M. Optimizing Magnification Ratio for the Flexible Hinge Displacement
Amplifier Mechanism Design. In International Conference on Material, Machines and Methods for Sustainable Development MMMS
2020, Proceedings of the 2nd Annual International Conference on Material, Nha Trang, Vietnam, 12–15 November 2020; Springer: Cham,
Switzerland, 2020; pp. 769–778.
5. Ni, H.; Li, M.H.; Zuo, X. Review on Damage Identification and Diagnosis Research of Civil Engineering Structure. Adv. Mater.
Res. 2014, 1006–1007, 34–37. [CrossRef]
6. Wang, Y.; Gu, Y.; Liu, J. A domain-decomposition generalized finite difference method for stress analysis in three-dimensional
composite materials. Appl. Math. Lett. 2020, 104, 106226. [CrossRef]
Buildings 2023, 13, 2480 18 of 18

7. Kabir, H.; Aghdam, M.M. A generalized 2D Bézier-based solution for stress analysis of notched epoxy resin plates reinforced
with graphene nanoplatelets. Thin Walled Struct. 2021, 169, 108484. [CrossRef]
8. Bert, C.W.; Malik, M. Differential quadrature: A powerful new technique for analysis of composite structures. Compos. Struct.
1997, 39, 179–189. [CrossRef]
9. Ahn, S.; Park, G.; Yoon, H.; Han, J.-H.; Jung, J. Evaluation of Soil–Structure Interaction in Structure Models via Shaking Table Test.
Sustainability 2021, 13, 4995. [CrossRef]
10. Xiao, F.; Zhu, W.; Meng, X.; Chen, G.S. Parameter Identification of Frame Structures by considering Shear Deformation. Int. J.
Distrib. Sens. Netw. 2023, 2023, 6631716. [CrossRef]
11. Xiao, F.; Zhu, W.; Meng, X.; Chen, G.S. Parameter Identification of Structures with Different Connections Using Static Responses.
Appl. Sci. 2022, 12, 5896. [CrossRef]
12. Kopsaftopoulos, F.P.; Fassois, S.D. Vibration based health monitoring for a lightweight truss structure: Experimental assessment
of several statistical time series methods (conference paper). Mech. Syst. Signal Process. 2010, 24, 1977–1997. [CrossRef]
13. He, J.; Chen, J.; Ren, X.; Li, J. A shake table test study of reinforced concrete shear wall model structures exhibiting strong
non-linear behaviors. Eng. Struct. 2020, 212, 110481. [CrossRef]
14. Moaveni, B.; He, X.; Conte, J.P.; Restrepo, J.I. Damage identification study of a seven-story full-scale building slice tested on the
UCSD-NEES shake table. Struct. Saf. 2010, 32, 347–356. [CrossRef]
15. Li, S.; Wu, C.; Kong, F. Shaking Table Model Test and Seismic Performance Analysis of a High-Rise RC Shear Wall Structure.
Shock. Vib. 2019, 2019, 6189873. [CrossRef]
16. Wu, C.; Li, S.; Zhang, Y. Structural Damage Identification Based on AR Model with Additive Noises Using an Improved TLS
Solution. Sensors 2019, 19, 4341. [CrossRef] [PubMed]
17. Hakim SJ, S.; Mokhatar, S.; Shahidan, S.; Chik, T.; Jaini, Z.; Ghafar, N.A.; Kamarudin, A. An ensemble neural network for damage
identification in steel girder bridge structure using vibration data. Civ. Eng. Archit. 2021, 9, 523–532. [CrossRef]
18. Wickramasinghe, W.R.; Thambiratnam, D.P.; Chan, T.H.T.; Nguyen, T. Vibration characteristics and damage detection in a
suspension bridge. J. Sound Vib. 2016, 375, 254–274. [CrossRef]
19. Morita, K.; Teshigawara, M.; Hamamoto, T. Detection and estimation of damage to steel frames through shaking table tests.
Struct. Control. Health Monit. 2005, 12, 357–380. [CrossRef]
20. Chen, W. Auto-Regressive Model Estimation Theory and Its Application in Deformation Monitoring Data Processing. Ph.D.
Thesis, Wuhan University, Wuhan, China, 2013.
21. Binder, M.D.; Hirokawa, N.; Windhorst, U. Auto-Regressive Model; Springer: Berlin/Heidelberg, Germany, 2008.
22. Bernagozzi, G.; Achilli, A.; Betti, R.; Diotallevi, P.P.; Landi, L.; Quqa, S.; Tronci, E.M. On the use of multivariate autoregressive
models for vibration-based damage detection and localization. Smart Struct. Syst. 2021, 27, 335–350.
23. Zheng, W.X. A least-squares based method for autoregressive signals in the presence of noise. IEEE Trans. Circuits Syst. Part II
Analog. Digit. Signal Process. 1999, 46, 81–85. [CrossRef]
24. Diversi, R.; Guidorzi, R.; Soverini, U. Identification of autoregressive models in the presence of additive noise. Int. J. Adapt.
Control. Signal Process. 2008, 22, 465–481. [CrossRef]
25. Zeng, W. Effect of the Random Design Matrix on Adjustment of an EIV Model and Its Reliability Theory. Ph.D. Thesis, Wuhan
University, Wuhan, China, 2013.
26. Diversi, R.; Soverini, U.; Guidorzi, R. A new estimation approach for AR models in presence of noise. IFAC Proc. Vol. 2005, 38,
160–165. [CrossRef]
27. Esfandiari, M.; Vorobyov, S.A.; Karimi, M. New estimation methods for autoregressive process in the presence of white observation
noise. Signal Process. 2020, 171, 107480. [CrossRef]
28. Guidorzi, R.; Diversi, R. Structural health monitoring application of errors-in-variables identification. In Proceedings of the
2013 21st Mediterranean Conference on Control and Automatino (MED), Platanias, Greece, 25–28 June 2013; Volume 2013, pp.
1098–1103.
29. Mahboub, V.; Amiri-Simkooei, A.; Sharifi, M. Iteratively reweighted total least squares: A robust estimation in errors-invariables
models. Surv. Rev. 2013, 45, 92–99. [CrossRef]
30. Datteo, A.; Busca, G.; Quattromani, G.; Cigada, A. On the use of AR models for SHM: A global sensitivity and uncertainty
analysis framework. Reliab. Eng. Syst. Saf. 2018, 170, 99–115. [CrossRef]
31. Xu, P.; Liu, J.; Shi, C. Total least squares adjustment in partial errors-in-variables models: Algorithm and statistical analysis. J.
Geodesy 2012, 86, 661–675. [CrossRef]
32. Xu, P. The effect of errors-in-variables on variance component estimation. J. Geodesy 2016, 90, 681–701. [CrossRef]
33. Khorshidi, S.; Karimi, M. Finite Sample FPE and AIC Criteria for Autoregressive Model Order Selection Using Same-Realization
Predictions. EURASIP J. Adv. Signal Process. 2010, 2009, 475147. [CrossRef]
34. Code for Seismic Design of Buildings (GB50011-2010); Ministry of Housing and Urban Rural Development of the People’s Republic
of China; China Architecture & Building Press: Beijing, China, 2010.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.