You are on page 1of 10

International Journal of Solids and Structures 185–186 (2020) 212–221

Contents lists available at ScienceDirect

International Journal of Solids and Structures


journal homepage: www.elsevier.com/locate/ijsolstr

Experimental investigation on use of regularization techniques and


pre-post measurement changes for structural damage identification
Jia Guo a, Li Wang b,∗, Izuru Takewaki a
a
Department of architecture and architectural engineering, Kyoto University, Kyoto, 615-8540, Japan
b
Department of applied mechanics and engineering, Sun Yat-sen University, Guangzhou, 510006, PR China

a r t i c l e i n f o a b s t r a c t

Article history: Practical damage identification often suffers from the ill-posedness and the unavoidable model errors. On
Received 5 June 2019 the one hand, the regularization techniques including the Tikhonov regularization and the sparse regular-
Revised 30 July 2019
ization have been well recognized as effective tools to circumvent the ill-posedness. On the other hand,
Accepted 20 August 2019
the model errors render the established baseline/analytical model inconsistent with the real structural
Available online 27 August 2019
and thereof, would introduce unexpected errors into damage identification. To deal with the model errors,
Keywords: the two-step strategy is often adopted for damage identification in practice: firstly update the baseline
Damage identification model so as to minimize the inconsistence and secondly, identify the damage upon the updated base-
Experimental investigation line model; Another empirical and not-yet-widely-recognized way is to use the measurement changes
Ill-posedness and motivated by this, a new MC (Measurement Changes) strategy that simply revises the measured data
Model errors with measurement changes is developed in this paper for sensitivity-based damage identification. This
Measurement changes
work aims to present experimental investigations and comparative studies on the possible remedies for
the difficulties in damage identification: the Tikhonov or sparse regularization for the ill-posedness and
the two-step or MC strategy for the model errors. Final results show that the MC strategy is rather effec-
tive and efficient for damage identification and synthesis of the sparse regularization and the MC strategy
is found as the appreciated way to practical damage identification.
© 2019 Elsevier Ltd. All rights reserved.

1. Introduction lution of the objective function (Mottershead et al., 2011; Friswell


and Mottershead, 2013; Lu and Wang, 2017).
Structural damage identification has been widely recognized as Due to the convenience in the acquisition of the modal data
an important content in health monitoring of practical infrastruc- (natural frequencies and/or mode shapes), a great number of dam-
tural and architectural structures (Ko and Ni, 2005; Ge and Lui, age identification methods using the modal data have been devel-
2005; Farrar and Jauregui, 1998). Changes in structural dynamic oped over the years. Advantages of using the modal data for dam-
properties introducing changes in natural frequencies was the pri- age identification are: (a) the modal data is easily accessible, e.g.,
mary investigation for the development of damage identification through operational modal analysis (Reynders et al., 2012; Wang
techniques (Doebling et al., 1996). After that, numerous types of et al., 2019), even without the knowledge of working loads for
different identification methods emerge, empirically or theoreti- practical structures; (b) the modal data is also independent from
cally (Sohn et al., 2003). Among them, the model-based methods the external excitations and the small structural damping. A com-
that require establishing the baseline finite element model of the prehensive review on this point can be found in the recent publi-
structure are often invoked when both locations and extents of the cation (Yuen, 2012). Despite all of these admiring aspects, difficul-
damage are to be identified. In general, such methods treat dam- ties still arise from the applications of model-based identification
age identification as a constrained nonlinear optimization problem, methods for real engineering structures:
of which the objective function is usually set up by measuring the
discrepancy between the numerical model and the dynamic test
• Ill-posedness. It is often required to identify the structural
results. The sensitivity approaches can then be used to get the so-
or damage parameters based on the incomplete modal data
(Yuen et al., 2006) due to the limited amount of measurement
transducers. This essentially renders damage identification be-

Corresponding author. come an ill-posed inverse problem, implying that the identifi-
E-mail address: wangli75@mail.sysu.edu.cn (L. Wang). cation would be rather sensitive to the measurement noise.

https://doi.org/10.1016/j.ijsolstr.2019.08.026
0020-7683/© 2019 Elsevier Ltd. All rights reserved.
J. Guo, L. Wang and I. Takewaki / International Journal of Solids and Structures 185–186 (2020) 212–221 213

• Model errors. When a practical structure is analyzed by a tification. The main idea behind is that measurement changes
baseline finite-element/mathematical model, the model errors caused by the same damage on the baseline model and on the
would arise inevitably. The model errors would definitely have real structure are quite approaching to each other even if the
nagative effects on damage identification (Lu et al., 2019). In model errors exist. For evidences on the effectiveness on using
total, the model errors consist of the model-parameter errors measurement changes, refer to (Doebling et al., 1996; Hao and
and the model-structure errors. For instance, uncertainties of- Xia, 2002; Salawu, 1997). In this paper, motivated by the idea
ten exist in the material properties of the structure and this using measurement changes, the MC (Measurement Changes)
would lead to the model-parameter errors; while when a prac- strategy by simply revising the measured data with measure-
tical tall building is simplified into a shear model, the model- ment changes is developed within the sensitivity-based damage
structure errors are introduced. An exhaustive list of the model identification framework. As is noteworthy, there is no need to
errors arising in model updating or damage identification can update the baseline model when using the MC strategy in dam-
be found in Ref. Mottershead et al. (2011). age identification and hence, the MC strategy turns out to be
more efficient than the two-step strategy.
Various researches have been carried out to overcome the diffi-
culties and they are briefly summarized in the following, The purpose of this paper is to investigate the performance
of the above remedies for the difficulties in damage identification
• To circumvent the ill-posedness, the effect of the measure- over an experimental structure. In proceeding so, the regularization
ment noise is to be minimized. Two stages are often invoked techniques—including the Tikhonov regularization and the sparse
in reducing the effect of the measurement noise: a) selec- regularization— along with the optimal weight matrix are used
tion of a proper weight matrix (Silva et al., 2016; Govers and to circumvent the ill-posedness, while the MC strategy that uses
Link, 2010) and b) use of regularization techniques (Titurus and measurement changes to correct the measured data is applied to
Friswell, 2008). On the first stage, it is well known that the circumvent the model errors with comparison to the conventional
weight matrix constitutes a key ingredient of the objective two-step strategy. Consequently, the present work is a comprehen-
function. An optimal weight matrix shall render the expecta- sive synthesis of the existing techniques and the revised strate-
tion of the identification errors minimized and such a choice gies, and then, to find an effective and efficient way to practical
has been proved in Ref. Lu et al. (2019) to be proportional to damage identification. It is noteworthy that the MC strategy is re-
the reciprocal of the covariance of the measurement noise. The stricted to the deterministic approach in this paper where dam-
improvement brought up by the optimal selection of the weight age identification is directly formulated by minimizing the misfit
matrix shall be substantial especially when different types of between the measured and calculated data. Nevertheless, there is
the measurement data, e.g., both mode shapes and natural fre- no constraint to extend the MC strategy to the probabilistic ap-
quencies, are used. On the second stage, the ill-posedness is proach (Vanik et al., 20 0 0) and the reason is given as follows. The
mainly caused by the possibly insufficient structural informa- MC strategy turns out to merely correct the measured data and
tion behind the measurement data and therefore, additional then, the statistical information (e.g., the covariance) of the cor-
consideration of the a priori knowledge on the damage pa- rected data can be easily derived from that of the measured data.
rameters can help to improve the well-posedness of damage In this way, the MC strategy for the probabilistic approach is sim-
identification; this leads to the so-called regularization tech- ply posed to correct the measured data and the corresponding sta-
niques (Benning and Burger, 2018). In principle, different a pri- tistical information.
ori knowledge on the damage parameters corresponds to dif- The remainder of this paper is structured as follows. In
ferent regularization techniques. For instance, the continuity or Section 2, the sensitivity-based framework for damage identifica-
bounding property of the parameters is well implicated in the tion with the modal data is briefly revisited and the difficulties
Tikhonov regularization (Tikhonov, 1963; Titurus and Friswell, arising in this framework as the ill-posedness and the model er-
2008), while the sparsity constraint on the damage locations rors are then introduced. Section 3 is devoted to furnishing cer-
can be simply enforced by the sparse (or 1 -norm) regulariza- tified as well as empirical remedies for the difficulties. The well
tion (Guo et al., 2018; 2019; Hernandez, 2014; Hou et al., 2018). certified (Tikhonov and sparse) regularization techniques are used
• In contrast, the publications on dealing with the model to circumvent the ill-posedness, while the empirical and revised
errors are rather limited. In most engineering and experi- MC strategy is developed to deal with the model errors along with
mental tests (Jang et al., 2002), the model errors are often comparison to the two-step strategy. Section 4 presents an elab-
reduced in the first step by adjusting certain model parame- orate experimental investigation on the performance of the reme-
ters so that the discrepancy between the calculated data from dies and final conclusions are drawn in Section 5.
the baseline model and the measured data of the undamaged
structure is minimized; this is also known as model updat- 2. Sensitivity-based damage identification using the modal data
ing (Mottershead et al., 2011). Damage identification shall be
conducted in the second step after model updating of the un- Consider a baseline structural model with the mass matrix M
damaged structure. The above two steps constitute the two- and the stiffness matrix K, that can be obtained from finite ele-
step strategy for damage identification with the model errors. ment analysis of the practical structure. Usually, the damage would
Primarily, the two-step strategy would work well if the struc- always cause a degradation of the stiffness. To represent the dam-
tural model is fine enough so that the model-structure errors age, the damage parameters are introduced as denoted by θ =
become negligible and the model parameters of the baseline [θ1 ; θ2 ; . . . ; θm ] and then, the stiffness matrix under the damage
model are well updated in the first step. However, when the parameters θ is designated as K(θ ). Moreover, in case of the in-
model-structure errors become non-negligible or when the se- tact or undamaged structure, the damage parameters are set to θ0 .
lected correction parameters are not consistent with the real In this way, the damage information including the locations and
source and the location of the errors (Mottershead et al., 2011), extents is totally implicated in the change of damage parameters
the updated baseline model in the first step would become θ = θ − θ 0 .
inconsistent with the real undamaged structure and this may For damage identification using the modal data, the following
introduce unexpected errors into damage identification. In ad- governing equation should be invoked,
dition to the two-step strategy, use of measurement changes
K(θ )φ i = ωi2 Mφ i (1)
has been empirically shown to be effective for damage iden-
214 J. Guo, L. Wang and I. Takewaki / International Journal of Solids and Structures 185–186 (2020) 212–221

where ωi , φ i are the corresponding natural frequencies and mode inaccurate, the rigid connection in the baseline model may be
shapes. Now, assume that the first l modes are measured for dam- inconsistent with the practice and the finite element mesh may
age identification at the possibly incomplete DOF set S and are col- also be not fine enough so that the discretization error exists.
lected in a column vector Rˆ , i.e., The model errors indicate inaccurate estimations of the intact
S damage parameters θ 0 and the model function R( · ) and there-
Rˆ = [{ω
ˆ i2 }li=1 ; {φˆ i }li=1 ] (2) fore, would definitely give rise to unexpected errors into the
where ω ˆ i , φˆ i are the measured natural frequency and correspond- identification problem (4). As is noteworthy, in case when the
ing mode shape of the ith mode, and the superscript S means re- model errors exist in θ 0 , the intact structure may admit im-
striction onto the index subset S. In correspondence to the mea- perfections or damage, however, the damage θ = θ − θ 0 to be
sured data in (2), the calculated data from the model (1) is also identified in this paper only captures the (stiffness) changes
expressed in the column vector form as and does not contain the damage in the intact structure.
R(θ ) = [{ωi2 (θ )}li=1 ; {φ Si (θ )}li=1 ] (3) In what follows, some certified and empirical remedies in the
where ωi (θ ), φ i (θ ) are the ith calculated modal data under the pre- literature — the regularization techniques to circumvent the ill-
scribed parameters θ . Then, along the general lines of inverse prob- posedness and the strategy using measurement changes (the MC
lems, damage identification can be formulated as a nonlinear op- strategy) to deal with the model errors — are revisited and syn-
timization problem, whose objective function is just the weighted thesized to cope with the difficulties.
least-squares of the discrepancy between the measured data Rˆ and
the calculated data R(θ ), that is, 3. Synthesis of regularization techniques and measurement
changes
θ ∗ = arg min{||Rˆ − R(θ 0 + θ )||2W } (4)

 The regularization techniques have been well recognized as ef-
where ||(· )||W := (· )T W(· ) and W is the positive definite weight fective tools to circumvent the ill-posedness of ill-posed problems,
matrix. As is noteworthy, the optimal weight matrix has been including damage identification (Benning and Burger, 2018; Titu-
shown proportional to the reciprocal of the covariance of the mea- rus and Friswell, 2008). In principle, regularization techniques are
surement noise (Lu et al., 2019) and this will guide the selection used to introduce a priori knowledge and allow a robust approxi-
of the weight matrix in the later of this paper. The sensitivity ap- mation of ill-posed (pseudo-)inverses (Ko and Ni, 2005). Different
proaches (Mottershead et al., 2011; Lu and Wang, 2017; Khoda- a priori knowledge on the parameters θ would lead to different
parast et al., 2008; Esfandiari et al., 2009) are often called to get regularization techniques. Mathematically, the regularization tech-
the solution of the problem (4). The key of such kind of approaches niques usually involve addition of the  p −norm term
lies in the linear expansion of the nonlinear term Rˆ − R(θ 0 + θ )
so that an approximate linear least-squares problem can be ob-
m
||θ || pp = | θ i | p (6)
tained and easily solved Lu and Wang (2017). In doing so, the sen-
i=1
sitivity matrix
⎛ ⎞ as a penalty term (or called regularization term) to the objective
.. .. .. ..
. . . . function, leading to the regularized formulation for damage identi-
⎜ ⎟
⎜ ∂ωi2 (θ ) ∂ωi2 (θ ) ∂ωi2 (θ ) ⎟
fication,
⎜ ∂θ1 ··· ∂θm ⎟
⎜ ∂θ2 ⎟ θ ∗ = arg min{||Rˆ − R(θ 0 + θ )||2W + λ||θ || pp } (7)
⎜ ⎟ θ
S(θ ) = ∇ R(θ ) := ⎜ ... ..
.
..
.
..
. ⎟ (5)
⎜ ⎟ where λ ≥ 0 is the regularization parameter. A proper regular-
⎜ ∂ φSi (θ ) ∂ φ Si (θ ) ∂ φ i (θ ) ⎟
⎜ ∂θ ⎟
S
··· ization parameter λ can be determined by the L-curve method
⎝ 1 ∂θ2 ∂θm ⎠
(Hansen, 1992; Hou et al., 2018). It has been well certified in the
.. .. .. ..
. . . . literature (see Zhang and Xu (2016) for instance) that different
choices of p in (7) correspond to different a priori knowledge on
should be obtained and details on the computation of the sensitiv-
∂ωi2 (θ ) ∂ φ Si (θ ) the damage parameters θ :
ity on natural frequencies ∂θ j and mode shapes ∂θ j can refer
• p = 2 corresponds to the Tikhonov (or 2 -norm) regularization.
to (Nelson, 1976; Friswell and Mottershead, 2013).
This regularization is equivalent to considering the constraint
Once the sensitivity approaches have solved the optimization
||θ||22 ≤ c for some positive number (Lu and Wang, 2017) and
problem (4), the damage θ ∗ shall have been identified. How-
in this way, the 2 -bounding property of the damage parame-
ever, difficulties would appear so that damage identification by
ters θ is implicitly enforced. The Tikhonov regularization has
purely solving problem (4) is often not-so-satisfactory and they are
been well applied to model updating (Mottershead et al., 2011;
mainly two-fold:
Friswell and Mottershead, 2013).
• Ill-posedness. Analogous to general inverse problems, damage • 0 ≤ p ≤ 1 leads to the sparse regularization. Likewise, such a reg-
ularization is equivalent to enforcing the constraint ||θ || p ≤
p
identification with the limited modal data is often an ill-posed
problem which is inherent in the ill-conditioning (or large con- η, 0 ≤ p ≤ 1 for some positive number η. In case of p = 0, the
dition number) of the matrix ST (θ )WS(θ ) for admissible dam- constraint ||θ ||00 ≤ η strictly restricts the number of non-zeros
age parameters θ . As a result, the identification is not robust at in θ to be no more than η and in this way, such a con-
all, being very sensitive to the measurement noise. straint is also termed the sparsity constraint. Nevertheless, p =
• Model errors. The baseline model with the mass and stiffness 0 makes the problem be a discrete optimization problem and
matrices M, K is often established through finite element analy- difficult to be solved. Luckily, it is found that 0 < p ≤ 1 can still
sis and this model is parameterized to involve all possible dam- pose the effect of the sparsity constraint (Candes et al., 2008).
In particular, p = 1 would make the regularization term ||θ || p
p
ages. However, such a baseline model is often an approximation
of the real structure, meaning that the model errors always ex- convex and then, many methods including the interior point
ist. For instance, the damage parameters θ 0 of the intact struc- method and the quadratic programming method can be used to
ture may be inexact due to the existence of possible imperfec- get the solution. Thus, p = 1 is often fixed in the sparse regular-
tions in the intact structure, the mass of the structure may be ization and in this way, the sparse regularization is also called
J. Guo, L. Wang and I. Takewaki / International Journal of Solids and Structures 185–186 (2020) 212–221 215

the 1 -norm regularization. The performance of the sparse reg- Table 1


Natural frequencies (Hz): experimental data vs calculated data
ularization in damage identification has been well verified in
from baseline models.
a number of publications, e.g., see Refs. Guo et al. (2018),
Hernandez (2014), Hou et al. (2018), Zhang and Xu (2016) to Mode order 1st 2nd 3rd 4th
cite a few. Experimental UF 2.796 8.015 27.027 55.541
BM-A 2.881 8.246 34.893 61.130
Before discussing the model errors, it is assumed that the
BM-B 2.795 8.000 27.001 55.544
modal data is measured not only on the damaged structure, but Experimental DF 2.654 8.023 26.533 53.302
also on the intact (or before damage) structure. To go further,
the measured modal data before damage is denoted by Rˆ 0 =
S
[{ω
ˆ 2 }l ; {φˆ 0i }l ], while the modal data after damage has been
0i i=1 i=1
collected in Rˆ (see equation (2)). For the baseline model, the infor-
mation —the intact damage parameters θ 0 and the model function where all the mode shapes φˆ i , φˆ 0i , φ i (θ ) are normalized in the same
R( · ) is prescribed in modelling the structure, though being possi- way. Consequently, by referring the match in (11), the MC strategy
bly inexact, and due to the existence of the model errors, the cal- suggests the following damage identification procedure,
culated data R(θ 0 ) shall be different from the measured data Rˆ 0 . θ ∗ = arg min{||Rˆ MC − R(θ 0 + θ )||2W + λ||θ || pp };
In most experimental and engineering tests, the two-step strategy θ

is often called to tackle the model errors in damage identification, ω2 ( θ0 ) 2
that is, Rˆ MC = {ωi2 (θ 0 ) + i 2 (ω ˆi −ω ˆ 02i )}li=1 ;
ωˆ 0i
• Step 1: model updating. Update the baseline model based on S S
l 
the modal data Rˆ 0 measured from the intact structure. Such φ Si (θ 0 ) + (φˆ i − φˆ 0i ) . (12)
a procedure is often similar to that in (7), i=1

θˆ 0 = arg min{||Rˆ − R(θ )||2W + λ||θ − θ 0 || pp } (8) In Eq. (12), Rˆ MC is termed the measurement-changes-corrected
θ
data and it is easily available once the possibly inexact baseline
where p = 2 is often fixed for model updating. Herein, θ
model is established and the modal data before and after the dam-
may include other damage-independent parameters, e.g., the
age is measured. Obviously, by the MC strategy, the identification
mass.
procedure (12) is quite similar to that in the original formulation
• Step 2: damage identification. The damage is then identified
(7), except that the measurement-changes-corrected data Rˆ MC is
based on the updated baseline model, i.e.,
used rather than Rˆ . Thus, the MC strategy (12) would introduce lit-
θ∗ = arg min{||Rˆ − R( θˆ 0 + θ )||2W + λ||θ|| pp }. (9) tle computational cost with respect to the original formulation (7),

and would be less demanding and computationally cheaper than
The two-step strategy would be effective if the errors in the
the two-step strategy (8) and (9).
model function R( · ) (or called the model-structure errors) are neg-
In the next section, a shaking table test is conducted to testify
ligible and the model parameters θˆ 0 are well updated in the first
the performance after synthesis of the regularization techniques—
step. However, in the updating step, efforts have to be made
including the Tikhonov regularization and sparse regularization—
to determine which parameters should be updated or modified
with the MC strategy or the two-step strategy.
Weng et al. (2009) and improper/missing update of some param-
eters would lead to undesired results. Moreover, even when a
proper set of parameters is selected in model updating, the com- 4. Shaking table test and damage identification
putational cost by the two-step strategy (8)–(9) seems to be almost
twice of that by the original formulation (7). 4.1. Shaking table test
In this paper, an empirical and revised MC strategy that uses
measurement changes to correct the measured data to tackle the Two frame structures— one is the undamaged frame (UF) and
model errors is additionally elaborated within the sensitivity-based another is the respective damaged frame (DF) as shown in Fig. 1
damage identification framework. The main idea behind the MC were selected as the prototype structure to testify the performance
strategy Hao and Xia (2002) is that of damage identification using different regularization techniques
and different strategies. A substructure shaking table test was used
• though the model errors exist between the baseline model and the
to produce dynamic loading in the structure laboratory of Kyoto
real structure, the changes in the modal data caused by the same
University. As shown in Fig. 1(a), the test structure was a steel
damage —called (damage-induced) measurement changes— on the
frame of one bay and two stories with the total height 0.475m and
baseline model and on the real structure are quite approaching to
the width 0.2m. To control the natural frequency of the frame so
each other.
as to deliver a more practical case, cube mass blocks were attached
This motivates a new identification strategy (termed the MC to the beam-column connections. The frames were bolted on the
strategy) that it is better to match measurement changes Rˆ − Rˆ 0 shaking table. Stainless L-angle connection was used for the beam-
with R( θ0 + θ ) − R( θ0 ), rather than to simply match the mea- column connection and column-base connection. Detailed configu-
sured data Rˆ with R(θ 0 + θ ). More specifically, for the modal ration of the connections is shown in Fig. 1(b).
data, the MC strategy tries to do the match Hao and Xia (2002); Eight accelerometers were installed to measure the horizontal
Doebling et al. (1996) as follows: for i = 1, 2, . . . , l accelerations of the frame (four at the middle of the columns and

ωi2 (θ 0 +θ )−ωi2 (θ 0 ) ωˆ 2 −ωˆ 2 four at the beam-column connections), from which the first four
ωi2 (θ 0 )
= i ωˆ 2 0i
0i (10) natural frequencies and mode shapes could be obtained by the
S S
φ Si (θ 0 + θ ) − φ Si (θ 0 ) = φˆ i − φˆ 0i subspace method Reynders et al. (2012). Details on the measured
frequencies and mode shapes can be found in Table 1 and Fig. 2.
or equivalently,

The positions and weights of the mass blocks and accelerom-


ωi2 (θ 0 )
ωi2 (θ 0 + θ ) = ωi2 (θ 0 ) + ωˆ 02i
(ωˆ i2 − ωˆ 02i ) eters during the experiment are given in Fig. 1(c). The sampling
S S
(11) frequency is set to 200Hz and the measurement duration is about
φ Si (θ 0 + θ ) = φ Si (θ 0 ) + (φˆ i − φˆ 0i )
30s. White noise excitation was selected as the shaking table input.
216 J. Guo, L. Wang and I. Takewaki / International Journal of Solids and Structures 185–186 (2020) 212–221

0.20m Mass (kg)


m2-1 1.208
mass block A4 m2-1 m2-2 A5
m2-2 1.210
0.25m a2-1 0.035
L-angle A3 a2-1 a2-2 A6 a2-2 0.035
connection m1-1 1.178
(b)
m1-2 1.179
150mm a1-1 0.021

40mm
A2 m1-1 m1-2 A7 a1-2 0.021
0.225m

Undamaged (UF) mass block


Accelerometer

20mm
A1 a1-1 a1-2 A8
(+cable)
L-angle
Damaged (DF) connection

(a) (d) (c)

Fig. 1. Overview of the test model: (a) the test structure; (b) mass block and L-angle connection; (c) placement of accelerometers; (d) undamaged and damaged beams.

Mode 1 Mode 2 Mode 3 Mode 4

(a) (b)

Fig. 2. Finite element models and mode shapes: (a) finite element model; (b) experimental mode shapes vs. numerical mode shapes.

Compared to the undamaged frame, the damage in the dam- undamaged frame (UF), being respectively rough and fine, are es-
aged frame is introduced by cutting a half of the width at the left tablished:
1/2 part of the beam on the first story and for a schematic view,
see Fig. 1(d). As a result, the stiffness reduction or the damage ex- • Baseline model A (BM-A): a rough model. In this model, the
tent for the damaged area is estimated to be 0.5. stiffness and mass matrices are obtained by finite element anal-
ysis based on the prescribed model information in the above,
along with which the stiffness of the L-angle connections is
simply seen as the stiffness of a beam but with more height
4.2. Finite element baseline models in the beam section.
• Baseline model B (BM-B): a fine model. To get this model, the
The finite element method is used to model the frame struc- mass and rotational inertia of the block masses and accelerom-
ture. To do so, the frame is divided into 20 elements as displayed eters are additionally considered and modified at first. Then, fi-
in Fig. 2(a) and for each element, the two-node Euler-Bernoulli nite element model updating (8) is called to update the stiff-
beam element is adopted and then, each node has three degrees- ness of the L-angle connections, and the beam and column el-
of-freedom (DOFs): two translational displacements and one rota- ements, based on the measured data of the UF.
tional angle.
Specific parameters of the model are given as follows: the Establishing the rough model BM-A is relatively simple, while
height of the first story is H1 = 0.225m and of the second story to get the fine model BM-B, more analysis and computation cost
is H2 = 0.25m, the width of the span is B = 0.2m, all columns is involved for model updating. The comparison of the modal data
and beams are of rectangular sections for which the columns and from the baseline models with the measured data is presented in
beams of the first story are of size b1 × h1 = 40mm × 1mm and Table 1 as well as graphically shown in Fig. 2. Obviously, the dis-
those of the second story are of size b2 × h2 = 30mm × 1mm. Ma- crepancy between the measured data and the calculated data of
terial properties are: the Young’s modulus E = 200GPa and the the rough model BM-A is to some extent large, while the mea-
mass density ρ = 7700kg/m3 . The mass and rotational inertia of sured data fits well with the calculated data of the fine model BM-
the mass blocks and/or accelerometers are considered in establish- B. Thus, it is deduced that the fine model BM-B is to some extent
ing the model. Two different finite element baseline models for the consistent with the real structure UF, while the gap between the
J. Guo, L. Wang and I. Takewaki / International Journal of Solids and Structures 185–186 (2020) 212–221 217

rough model BM-A and the real structure is non-negligible. Next, Table 2
Objective functions for damage identification using dif-
damage identification will be carried out upon the two baseline
ferent regularization techniques and with/without the
models BM-A and BM-B, respectively. MC strategy. The expressions of g(θ ) and gMC (θ ) are
given respectively in equations (14) and (15).
4.3. Damage identification procedures
Scenario Objective function

For damage identification, the choice of the weight matrix is 2 g(θ ) + λ||θ||22 ,
discussed at first. Note that the measurement noise in different 2 +MC gMC (θ ) + λ||θ ||22 ,
frequencies and mode shapes is reasonably assumed to be inde- 1 g(θ ) + λ||θ ||11 ,
1 +MC gMC (θ ) + λ||θ ||11 .
pendent and recall that the optimal weight matrix shall be pro-
portional to the reciprocal of the covariance of the measurement
noise. Under this circumstance, the objective function in the origi- Table 3
nal formulation (4) under the optimal weight Lu et al. (2019) shall Identification errors d (θ id , θ ex ) on different baseline models.
become
Baseline model 2 2 + MC 1 1 + MC
g(θ ) = ||Rˆ − R(θ 0 + θ )||2W BM-A 11.7674 0.3943 2.6566 0.0915

 2 S
 BM-B 0.3844 0.3726 0.2344 0.0574

l
ωˆ i2 − ωi2 (θ 0 + θ ) φˆ i − φ Si (θ 0 + θ )
=c + || ||22
i=1
σωˆ i2 σφˆ i
(13)
Table 4
Time comparison (in second) for different damage identification proce-
dures.
where c > 0 is a constant, σωˆ 2 , σφˆ are the respective standard devi-
i i
ations of the measurement noise in frequencies and mode shapes. Model 2 2 + MC 1 1 + MC

Empirically, many operational modal analysis methods would ad- BM-A - 0.72 - 0.89
mit the following identification accuracy for frequencies and mode BM-B (1.03+)0.92 (1.03+)0.69 (1.03+)0.58 (1.03+)0.74
shapes by referring to the results in Wang et al. (2019): the relative
errors in identified frequencies are around rf ∈ [0.05%, 0.5%], while
the MAC values between the identified and exact mode shapes
shall approximately verify MAC ∈ [0.997, 0.9999]. Usually, the mode
4.4. Identification results and discussions
shapes are normalized in the sense of 2 -norm and under this
circumstance, the relative error between the identified and exact
When damage identification is fulfilled using the four ob-
mode shapes, denoted respectively by φ id and φ , shall be approxi-
jective functions in Table 2, proper regularization parameters λ
mately obtained as
 
should be selected. Herein, the L-curve method Hansen (1992);
||φ id − φ ||2 Hou et al. (2018) is adopted to determine the regularization pa-
rφ = = 2 − 2φ Tid φ = 2 − 2MAC ∈ [1.41%, 7.75%].
||φ ||2 rameters and detailed L-curve plots by the four objective functions
Above all, it is approximately obtained that σωˆ 2 ≈ 2r f ω
ˆ i2 , σφˆ ≈ rφ . are given in Fig. 3 for the rough model BM-A and in Fig. 4 for the
i i fine model BM-B. Once proper regularization parameters are ob-
Now turn to the objective function (13) with the optimal weight, tained, the sensitivity approaches Mottershead et al. (2011); Lu and
set c = (2r f )2 and μ = (2r f /rφ )2 , and then, the objective function Wang (2017) can be applied to get the solution of the objective
(13) is simplify to functions and thereof, to identify the damage in the structure. To

 2  have an impression on how the identification procedures converge,

l
ωˆ i2 − ωi2 (θ 0 + θ ) S
g(θ ) = + μ|| φˆ i − φ Si (θ 0 + θ )|| 2 the evolution of the damage extent of element 3 is shown in Fig. 5
i=1
ωˆ i2 2
where the scenarios 2 , 1 under BM-A are not presented because
they do not give convergent and satisfactory results. Clearly, the
(14)
convergence is all reached in no more than seven iterations.
and by referring to the aforementioned analysis, it is reasonable to Detailed identification results corresponding to the four objec-
fix μ = 0.01 for optimal weight in this paper. For the MC strategy tive functions in Table 2 are shown in Fig. 6 for the rough model
(12), the optimal weight matrix is selected likewise so that the ob- BM-A and Fig. 7 for the fine model BM-B. To further quantify the
jective function (without regularization yet) becomes identification accuracy, the distance between the identified dam-

 2 age θ id and the exact damage θ ex is measured to represent the

l
ωˆ i2 − ωˆ 02i ωi2 (θ 0 + θ ) − ωi2 (θ 0 )
gMC (θ ) = − identification errors, that is,
i=1
ωˆ 02i ωi2 (θ 0 )
 d (θ id , θ ex ) = ||θ id − θ ex ||2 . (16)
S S and detailed results on such identification error indicator (16) are
+ μ||( φˆ i − φˆ 0i ) − (φ Si (θ 0 + θ ) − φ Si (θ 0 ))||22 . (15)
listed in Table 3. To quantify the computation cost, the computa-
tion times are counted as given in Table 4 where ’-’ means acquire-
Consequently, the specific objective functions for damage identi- ment of the undesired solution and the number in the brackets
fication herein using the Tikhonov (2 -norm) or sparse (1 -norm) indicates the time to get the updated model BM-B. It is easily ob-
regularization with or without the MC strategy are summarized in served from the results in the figures and the tables that
Table 2, where 2 , 1 indicate the respective Tikhonov and sparse
regularization, and ’+MC’ means use of the MC strategy. It is note- • Under the rough model BM-A, 2 and 1 are almost unable
worthy that the fine model BM-B along with 2 , 1 just correspond to reproduce the actual damage in the structure because even
to the two-step strategy (8)-(9) with different regularization tech- damage locations are wrongly detected as in Fig. 6(a) and (c)
niques. Then, damage identification is independently proceeded for and the corresponding identification errors in Table 3 are large,
the four objective functions in Table 2 on the rough model BM-A while under the fine model BM-B, the damage at element 3 is
and on the fine model BM-B. reasonably identified by the four objective functions (see Fig. 7).
218 J. Guo, L. Wang and I. Takewaki / International Journal of Solids and Structures 185–186 (2020) 212–221

(a) (b)
Fig. 3. L-curve plots for damage identification upon BM-A by: (a) 2 +MC and (b) 1 +MC.

Fig. 4. L-curve plots for damage identification upon BM-B by: (a) 2 ; (b) 2 +MC; (c) 1 and (d) 1 +MC.

The reason why 2 and 1 perform bad under the rough model clearly less for 2 +MC and 1 +MC than respectively for 2 and
BM-A is that non-negligible model errors exist between the 1 . This indicates that when model updating is used to get a
baseline model BM-A and the real structure. fine baseline model (e.g., BM-A), the model errors though being
• Under the rough model BM-A, 2 +MC and 1 +MC have found small and often negligible still exist and using the MC strategy
to give satisfactory identification of the damage in Fig. 6(b) and can again give rise to better identification results.
(d): the damage location at element 3 is well detected and the • Using 2 and 1 under the fine model BM-B (refer to Fig. 7(a)
identified extents are also acceptable with respect to the refer- and (c)) just corresponds the usual two-step strategy: firstly up-
ence extent 0.5. The MC strategy has been found to work well dating and then identifying. While using 2 +MC and 1 +MC un-
under the non-negligible model errors. Even for the final model der the rough model BM-A (refer to Fig. 6(b) and (d)) reflects
BM-B, the MC strategy can also give rise to better identification the primary MC strategy. Comparing the results of the two-step
of the damage because the identification errors in Table 3 are strategy with those of the (primary) MC strategy, both shall
J. Guo, L. Wang and I. Takewaki / International Journal of Solids and Structures 185–186 (2020) 212–221 219

0.7 • Note from Table 4 that the computation time to get the fine
model BM-B is at the same level with that to conduct damage
0.6 identification and as a result, the computation cost by the MC
stiffness reduction
strategy is almost half of that by the two-step strategy; the MC
0.5
strategy is more efficient.
0.4 • The sparse regularization has been found to perform bet-
ter in damage identification than the Tikhonov regularization
0.3 and the same conclusion has been well drawn in many ref-
erences, e.g., Refs. Guo et al. (2019), Zhang and Xu (2016),
0.2 BM-A l2+MC BM-B l1 Hernandez (2014) to cite a few.
BM-A l1+MC BM-B l2+MC • When both the sparse regularization and the MC strategy (or
0.1 1 +MC) are used as in Fig. 6(d) and Fig. 7(d), the identification
BM-B l2 BM-B l1+MC
error under the rough model BM-A is at the same level with
0
1 2 3 4 5 6 7 that under the fine model BM-B. This means that when 1 +MC
iteration is used for damage identification, there is no need to conduct
the cumbersome model updating for a fine model, rather, a
Fig. 5. The convergence history of the damage extent at element 3 for different rough model is often enough.
damage identification procedures.

To conclude, when dealing with the model errors, the MC strat-


be acceptable and under the Tikhonov (2 -norm) regularization, egy rather than the two-step strategy is often preferred no matter
the identification is also of the same accuracy by referring to whether a fine model is obtained or not; while as regards the ill-
the corresponding results in Table 3. While under the sparse posedness, the sparse regularization often performs superior to the
(1 -norm) regularization, the MC strategy gives much better Tikhonov regularization for damage identification. Though the MC
identification results than the two-step strategy by which fic- strategy is empirical, its sound performance for damage identifica-
tious damage location at element 13 with the extent approach- tion has been well verified in this test and the mechanism behind
ing 0.2 is also detected. In total, even under the rough model, the MC strategy shall be exploited in the future work. Eventually, a
the MC strategy along with the sparse regularization can give practical way by using the MC strategy along with the sparse reg-
rise to better damage identification than the two-step strategy, ularization is advised for practical damage identification from this
let alone for the MC strategy under the fine model. experimental investigation.

Approximated reduction for


element 3

Element 3
Element 3
Approximated reduction for
element 3

(a) (b)

Approximated reduction for


element 3

Element 3
Approximated reduction for
element 3
Element 3

(c) (d)
Fig. 6. Identification results upon BM-A by: (a) 2 ; (b) 2 +MC; (c) 1 and (d) 1 +MC.
220 J. Guo, L. Wang and I. Takewaki / International Journal of Solids and Structures 185–186 (2020) 212–221

Approximated reduction for Approximated reduction for


element 3 element 3

Element 3 Element 3

(a) (b)

Approximated reduction for Approximated reduction for


element 3 element 3

Element 3 Element 3

(c) (d)
Fig. 7. Identification results upon BM-B by: (a) 2 ; (b) 2 +MC; (c) 1 and (d) 1 +MC.

5. Conclusions Acknowledgement

An experimental investigation on how to deal with the ill- The present research was performed under the support of
posedness and the model errors in damage identification of prac- National Natural Science Foundation of China (No. 11702336) and
tical structures has been presented in this paper. In proceeding so, the MEXT scholarship of Japan.
the well certified regularization techniques including the Tikhonov
regularization and the sparse regularization are used to circum-
vent the ill-posedness, while the default two-step strategy—firstly
updating and then identifying and the empirical, not-yet-widely- References
recognized and revised MC strategy— simply using measurement
Benning, M., Burger, M., 2018. Modern regularization methods for inverse problems.
changes to correct the measured data are recalled to cope with Acta numerica 27, 1–111.
the model errors. Consequently, synthesis of the regularization Candes, E.J., Wakin, M.B., Boyd, S.P., 2008. Enhancing sparsity by reweighted l1 min-
techniques and the MC (as well as two-step) strategy within the imization. J. Fourier Anal. Appl. 14 (5), 877–905.
Doebling, S.W., Farrar, C.R., Prime, M.B., Shevitz, D.W., 1996. Damage identification
conventional sensitivity-based framework is developed for damage and health monitoring of structural and mechanical systems from changes in
identification and the experimental results show that their vibration characteristics. a literature review.
Esfandiari, A., Bakhtiari-Nejad, F., Rahai, A., Sanayei, M., 2009. Structural model up-
dating using frequency response function and quasi-linear sensitivity equation.
J. Sound Vibr. 326 (3), 557–573.
• The empirical MC strategy is found more effective and ef- Farrar, C.R., Jauregui, D.A., 1998. Comparative study of damage identification algo-
ficient than the conventional two-step strategy. This is be- rithms applied to a bridge: i. experiment. Smart Mater. Struct. 7 (5), 704.
cause the MC strategy can works well under the rough base- Friswell, M.I., Mottershead, J.E., 2013. Finite element model updating in structural
dynamics. volume 38. Springer Science & Business Media.
line model, and even under the fine model, the MC strategy Ge, M., Lui, E.M., 2005. Structural damage identification using system dynamic prop-
can also give rise to better identification results. erties. Comput. Struct. 83 (27), 2185–2196.
• Both the sparse regularization and the Tikhonov regulariza- Govers, Y., Link, M., 2010. Stochastic model updating|covariance matrix adjustment
from uncertain experimental modal data. Mech. Syst. Signal Process. 24 (3),
tion can be well synthesized with the MC strategy, but for 696–706.
damage identification, the sparse regularization is often pre- Guo, J., Wang, L., Takewaki, I., 2018. Modal-based structural damage identification by
ferred. minimum constitutive relation error and sparse regularization. Struct. Control
Health Monitor. 25 (12), e2255.
Guo, J., Wang, L., Takewaki, I., 2019. Frequency response-based damage identification
in frames by minimum constitutive relation error and sparse regularization. J.
Thus, due to its no need for exhaustively updating the baseline Sound Vibr. 443, 270–292.
model and its highlighted performance in detecting the damage, Hansen, P.C., 1992. Analysis of discrete ill-posed problems by means of the l-curve.
SIAM Rev. 34 (4), 561–580.
synthesis of the MC strategy and the sparse regularization is found Hao, H., Xia, Y., 2002. Vibration-based damage detection of structures by genetic
as the appreciated way to practical damage identification. algorithm. J. Comput. Civil Eng. 16 (3), 222–229.
J. Guo, L. Wang and I. Takewaki / International Journal of Solids and Structures 185–186 (2020) 212–221 221

Hernandez, E.M., 2014. Identification of isolated structural damage from incomplete Silva, T.A.N., Maia, N.M.M., Link, M., Mottershead, J.E., 2016. Parameter selection and
spectrum information using l1-norm minimization. Mech. Syst. Signal Process. covariance updating. Mech. Syst. Signal Process. 70, 269–283.
46 (1), 59–69. Sohn, H., Farrar, C.R., Hemez, F.M., Shunk, D.D., Stinemates, D.W., Nadler, B.R., Czar-
Hou, R., Xia, Y., Bao, Y., Zhou, X., 2018. Selection of regularization parameter for necki, J.J., 2003. A review of structural health monitoring literature: 1996–2001.
l1-regularized damage detection. J. Sound Vibr. 423, 141–160. Los Alamos National Laboratory, USA.
Jang, J.H., Yeo, I., Shin, S., Chang, S.P., 2002. Experimental investigation of system-i- Tikhonov, A.N., 1963. On the solution of ill-posed problems and the method of reg-
dentification-based damage assessment on structures. J. Struct. Eng. 128 (5), ularization. In: In Doklady Akademii Nauk, volume 151. Russian Academy of Sci-
673–682. ences, pp. 501–504.
Khodaparast, H.H., Mottershead, J.E., Friswell, M.I., 2008. Perturbation methods for Titurus, B., Friswell, M.I., 2008. Regularization in model updating. Int. J. Numer.
the estimation of parameter variability in stochastic model updating. Mech. Methods Eng. 75 (4), 440–478.
Syst. Signal Process. 22 (8), 1751–1773. Vanik, M.W., Beck, J.L., Au, S.K., 20 0 0. Bayesian probabilistic approach to structural
Ko, J.M., Ni, Y.Q., 2005. Technology developments in structural health monitoring of health monitoring. J. Eng. Mech. 126 (7), 738–745.
large-scale bridges. Eng. Struct. 27 (12), 1715–1725. Wang, L., Huang, M., Lu, Z.R., 2019. Bling separation of structural modes by com-
Lu, Z.R., Wang, L., 2017. An enhanced response sensitivity approach for structural pact-bandwidth regularization. Mech. Syst. Signal Process. 131, 288–316.
damage identification: convergence and performance. Int. J. Numer. Methods Weng, J.H., Loh, C.H., Yang, J.N., 2009. Experimental study of damage detection by
Eng. 111 (13), 1231–1251. data-driven subspace identification and finite-element model updating. J. Struct.
Lu, Z.R., Zhou, J., Wang, L., 2019. On choice and effect of weight matrix for response Eng. 135 (12), 1533–1544.
sensitivity-based damage identification with measurement and model errors. Yuen, K.V., 2012. Updating large models for mechanical systems using incomplete
Mech. Syst. Signal Process. 114, 1–24. modal measurement. Mech. Syst. Signal Process. 28, 297–308.
Mottershead, J.E., Link, M., Friswell, M.I., 2011. The sensitivity method in finite ele- Yuen, K.-V., Beck, J.L., Katafygiotis, L.S., 2006. Efficient model updating and health
ment model updating: a tutorial. Mech. Syst. Signal Process. 25 (7), 2275–2296. monitoring methodology using incomplete modal data without mode matching.
Nelson, R.B., 1976. Simplified calculation of eigenvector derivatives. AIAA J. 14 (9), structural control and health monitoring. Off. J. Int. Assoc. Struct. Control Mon-
1201–1205. itor. Eur. Assoc. Control Struct. 13 (1), 91–107.
Reynders, E., Houbrechts, J., De, G., 2012. Roeck. fully automated (operational) modal Zhang, C.D., Xu, Y.L., 2016. Comparative studies on damage identification with
analysis. Mech. Syst. Signal Process. 29, 228–250. tikhonov regularization and sparse regularization. Struct. Control health Mon-
Salawu, O.S., 1997. Detection of structural damage through changes in frequency: a itor. 23, 560–579.
review. Eng. Struct. 19 (9), 718–723.

You might also like