See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/318931411

Identification of nonlinear hysteretic parameters by enhanced response

sensitivity approach

Article  in  International Journal of Non-Linear Mechanics · August 2017

DOI: 10.1016/j.ijnonlinmec.2017.07.012

CITATIONS READS

27 611

4 authors:

Z.R. Lu Yao Renzhi

Sun Yat-Sen University Sun Yat-Sen University
151 PUBLICATIONS   2,144 CITATIONS    4 PUBLICATIONS   52 CITATIONS   

SEE PROFILE SEE PROFILE

Li Wang J.K. Liu

Sun Yat-Sen University Sun Yat-Sen University
64 PUBLICATIONS   414 CITATIONS    217 PUBLICATIONS   2,293 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Parameters identification based on response sensitivity analysis View project

Research on damage identification of wind turbine blade based on time-domain response analysis View project

All content following this page was uploaded by Li Wang on 17 November 2020.

The user has requested enhancement of the downloaded file.

 Identification of nonlinear hysteretic parameters by enhanced
response sensitivity approach
Zhong-Rong Lu1 , Renzhi Yao1 , Li Wang∗,1 , and Jike Liu1
1
Department of applied mechanics and engineering, Sun Yat-sen University, Guangzhou,
P.R. China

Abstract
Hysteresis is a ubiquitous phenomenon describing the special nonlinear memory-based
relation between the input and the output in many physical systems. Identifying the hys-
teretic parameters is the first step towards practical application of hysteretic models. In
this paper, a general framework for parameter identification of nonlinear hysteretic models
is developed based on the enhanced response sensitivity approach. To do so, three typical
hysteretic models—Bouc-Wen model, bilinear model with kinematic hardening and bilinear
model with equal yielding force are analyzed at first and the general way to model a struc-
ture with such hysteretic components is established thereafter. Then, the enhanced response
sensitivity approach is presented for inverse parameter identification where the key lies in
the sensitivity analysis and the trust-region constraint. Particularly, smoothing procedure is
introduced to overcome the non-differentiability of bilinear hysteretic functions for sensitiv-
ity analysis of bilinear models. Numerical examples are studied to testify the feasibility and
performance of the proposed approach.
Keywords: enhanced response sensitivity approach, parameter identification, Bouc-Wen
hysteresis, bilinear hysteresis, smoothing procedure

1 Introduction
The hysteretic behavior is frequently encountered in many physical systems, such as mechanical
systems [1], structural dampers/seismic isolation systems [2, 3], friction models [4], oscillato-
ry circuits [1, 5] and so on. In definition, hysteresis is a special nonlinear property describing
the memory-based relation between the input and the output, that is, the output at a given
time instant depends not only on the instantaneous input but also on its past history. Var-
ious mathematical models [1, 6–9], including Preisach, Ishlinskii, Bouc-Wen, bilinear models,
have been developed to describe the hysteretic behavior. Among these models, the Bouc-Wen
models are quite general since they can represent a wide variety of softening/hardening, degrad-
ing/pinching, smoothly-varying/nearly-bilinear hysteretic behaviors [5,10], but they are often of
complex forms; while the bilinear models with kinematic hardening or equal yielding force are
the ever simplest hysteretic models and are quite practical for applications in friction modeling
and seismic isolation [2] etc, but they suffer from non-smoothness or non-differentiability. In the
present paper, the focus is typically on the Bouc-Wen and bilinear hysteretic models and other
hysteretic models can be dealt with likewise.
Hysteretic parameter identification or calibration is the first step towards practical applica-
tion of the nonlinear hysteretic models and it often turns out to be a nontrivial task due to the
∗
Correspondence to: Li Wang, Email: wangli75@mail.sysu.edu.cn

1
inherent nonlinearity and memory nature. Two main classes of identification methods have been
developed in the literature. In the first class, the identification procedure is transformed into a
state estimation problem after discretizing the differential equations into discrete state equations
through numerical integrations and treating the parameters as state variables. Then, various
techniques are proposed to deal with the state estimation problem and some representative
work is: Lin and Corigliano et al [11, 12] proposed to use the extended Kalman filter to identify
the state variables, for which, linearization is adopted in order to fit with the general linear
Kalman filter framework; Wu and Smyth [10] developed an unscented Kalman filter approach
to estimate the parameter by resorting to the unscented transformaiton; Chang and Shi [13]
identified the state variables as well as hysteretic parameters through wavelet multiresolution
analysis. The whole procedure for the state estimation seems simple, however, in order to gain
good identification, the discrete state equations should well approximate the original differential
equations implying that the sampling time step should be as small as possible, and the duration
for measurement data should be long enough.
In the second class, identification of the hysteretic parameters from the measured data is
formulated as an inverse problem and is often fulfilled by solving an optimization problem. The
objective function is typically defined as the weighted least-squares of the error between the
measured data and the derived theoretic data—function of parameters, and then, actual model
parameters should minimize the objective function. Different kinds of the measured data have
led to different objective functions so well as different identification procedures. Ni et al [14]
proposed to use the frequency-domain displacement data from periodic vibration experiment
and then, the Levenberg-Marquardt algorithm was adopted to solve the nonlinear least squares
optimization problem. However, getting the experimental frequency-domain data of nonlinear
hysteretic systems is not so straightforward as of linear systems and considerable efforts should
be taken for large amount of experiment data. In contrast, the time-domain data is much
more easily accessed. Yar and Sues et al [15–18] successfully identified the Bouc-Wen hysteretic
parameters by resorting to the time-domain displacement and restoring force data and the
multi-stage estimate schemes. The multi-stage therein is mainly invoked to tackle the complex
exponential coefficient in the Bouc-Wen hysteretic model. Though satisfactory accuracy as well
as convergence was gained, the multi-stage approaches are usually time-consuming.
As is noteworthy, the above two classes of approaches are almost for parameter identification
of the Bouc-Wen models and few are devoted to identify bilinear hysteretic parameters. The
reason is two-fold. On the one hand, if time-domain displacement and restoring force data of
a bilinear hysteretic component is measured, the parameters are explicitly obtainable from the
hysteretic loop graph of the bilinear hysteretic model. On the other hand, if other kinds of
time-domain data are available, sensitivity analysis is often required for parameter identification
process which is possibly infeasible due to the non-smoothness and non-differentiability of the
bilinear hysteretic functions. Practically, for structures with bilinear hysteretic components,
getting all the displacement and restoring force data may be costly and sometimes impossible;
this indicates that new approach is still in demand for practical parameter identification of
bilinear hysteretic models.
In the present work, the enhanced response sensitivity approach [19–21], which has been
shown successful for structural damage identification, is utilized to establish a general framework
for parameter identification of various hysteretic models with arbitrary kind of time-domain
measured data. There are several remarkable features of the enhanced response sensitivity
approach which are helpful to reach the goal. Firstly, the present work is within the framework
of the least squares optimization and therefore, there is no strict constraint on sampling time
step. Secondly, the easily available time-domain data is adopted and arbitrarily single kind or
combinatory of the data—the acceleration, the velocity, the displacement or even the restoring
force can be used to identify the parameters as along as corresponding response sensitivity

2
Figure 1 Nonlinear hysteretic models: (a) Bouc-Wen model, (b) bilinear model with kinematic
hardening, (c) bilinear model with equal yielding force

analysis is performed. Thirdly, the enhanced response sensitivity approach has already been
proved to be weakly convergent [21] and this can to some extend guarantee the performance for
parameter identification of hysteretic models in this work.
The remainder of this paper is organized as follows. Section 2 introduces three usual hys-
teretic models and establishes a general model procedure for dynamic motion of structures with
hysteretic components. Then, identifying hysteretic parameters is formulated as a nonlinear
least-squares optimization problem herein. In Section 3, how to solve the nonlinear optimiza-
tion problem by the enhanced response sensitivity approach is elaborated. Numerical examples
are conducted in Section 4 and final conclusions are drawn in Section 5.

2 Identification problem statement

2.1 Nonlinear hysteretic models
Herein, the Bouc-Wen hysteretic model and two bilinear hysteretic models with kinematic hard-
ening and equal yielding force, respectively are considered (see Figure 1). Generally, hysteresis
describes the memory-based relation between the displacement u and the restoring force r, or
mathematically, there is
ṙ = f (r, u, u̇, p) (1)
where p represents the effective model parameters, the notation  ˙ := d is adopted for time t
dt
and f (·) is the nonlinear hysteretic function governing the hysteretic behavior. In this paper,
hysteretic functions of the three hysteretic models in Figure 1 are of the following forms:
• Bouc-Wen model [5, 22] (see Figure 1(a))

3
 1
ṙ = [Au̇ − ν(βr|r|n−1 |u̇| + γ|r|n u̇)]
η
A νβ νγ (2)
:= f BW (r, u, u̇, p); p = [p1 = , p2 = , p3 = , p4 = n]T
η η η
= p1 u̇ − (p2 r|r|p4 −1 |u̇| + p3 |r|p4 u̇)
where η, A, ν, β, γ and n are design parameters for the hysteretic behavior and p are ef-
fective parameters. In order for the equivalence between the design parameters and the
effective parameters, the parameters η, ν, which are typically used to control the degrad-
ing and pinching behavior [10], respectively, are both set to unity herein without loss of
generality. Physically, the exponential coefficient verifies n ≥ 1 and satisfaction of the
thermodynamic admissibility [23] requires |γ| ≤ β, A > 0; in other words, the admissible
parametric space should be

p ∈ P BW := {p ∈ R4 : p1 > 0, |p3 | ≤ p2 , p4 ≥ 1}. (3)

• Bilinear model with kinematic hardening [2] (see Figure 1(b))

k1 + k2 k1 − k2
ṙ = [ − sgn((r − k2 u)sgn(u̇) − F0 )]u̇
2 2
k1 − k2 (4)
:= f B1 (r, u, u̇, p); p = [p1 = , p2 = k2 , p3 = F0 ]T
2
= [p1 + p2 − p1 sgn((r − p2 u)sgn(u̇) − p3 )]u̇
or in a simpler form

r = r̃ + p2 u
(5)
r̃˙ := f˜B1 (r̃, u, u̇, p) = p1 [1 − sgn(r̃sgn(u̇) − p3 )]u̇

where k1 , k2 are respective primal and degraded stiffness of the model (see Figure 1(b)),
F0 is the yielding force under zero displacement and sgn(·), throughout this paper, denoting
the following sign function 
−1, if x < 0,
sgn(x) = (6)
1, if x ≥ 0.
Obviously, the parameters of this model should lie in the following space,

p ∈ P B1 := {p ∈ R3 : p1 > 0, p2 ≥ 0, p3 > 0}. (7)

• Bilinear model with equal yielding force [8] (see Figure 1(c))

k1 + k2 k1 − k2
ṙ = [ − sgn(r sgn(u̇) − Fy )]u̇
2 2
k1 + k2 k1 − k2 (8)
:= f B2 (r, u, u̇, p); p = [p1 = , p2 = , p3 = Fy ]T
2 2
= [p1 − p2 sgn(r sgn(u̇) − p3 )]u̇
where k1 , k2 are stiffness parameters before and after yielding, respectively and Fy is the
yielding force. Likewise, the parametric space should satisfy

p ∈ P B2 := {p ∈ R3 : p1 ≥ p2 , p2 > 0, p3 > 0}. (9)

4
As is seen from the above models, the Bouc-Wen model (2) is smooth but the parameters are not
so obviously derived from the hysteretic loop graph in Figure 1(a); while the two bilinear models
(5) and (8) is non-differentiable due to the existence of the discontinuous function sgn(·), but
the three parameters can be explicitly observed from the hysteretic loop graphs in Figure 1(b)
and (c). In what follows, these hysteretic models will be incorporated into a specific structure
as hysteretic components and how to systematically model such a structure with hysteretic
components is elaborated.

2.2 Modeling structures with hysteresis

Consider a structure with time-invariant mass matrix M, stiffness matrix K and linear damping
matrix C. The displacements are denoted by u = [u1 , u2 , ..., un ]T which are of n DOFs and
the external loading vector is F (t). Assume that there are m hysteretic components in this
structure and let y(t) ∈ Rm collect the displacements of the m hysteretic components. Usually,
the displacements of the hysteretic components are linearly dependent on the displacements of
the whole structure through a connection matrix T ∈ Rm×n , i.e.,

y = Tu. (10)

Conversely, with the connection matrix T, one can gain the information about where and how the
hysteretic components are connected to a specific structure. For illustrative purpose, turning
to the structure in Figure 2 where u1 , ..., u4 are n = 4 structural DOFs and y1 , y2 , y3 are
displacements of m = 3 hysteretic components, the connection matrix is found to be,
 
  u1  
y1  u2  −1 1 0 0
 y2  = T   −1 0 1 0  .
 u3  ; T = (11)

y3 0 0 −1 1
u4

The restoring forces of the m hysteretic components are denoted by r = [r1 , r2 , ..., rm ]T with rk
being the restoring force of the kth hysteretic component and they are of the following relation
to the displacements y,

 ṙ1 = f1 (r1 , y1 , ẏ1 , p1 ); p1 ∈ P1

ṙ2 = f2 (r2 , y2 , ẏ2 , p2 ); p2 ∈ P2

⇔ ṙ = f (r, y, ẏ, p); p = [p1 ; p2 ; ...; pm ] ∈ P (12)

 ...
ṙm = fm (rm , ym , ẏm , pm ); pm ∈ Pm


where fk (·) ∈ {f BW (·), f B1 (·), f B2 (·)}, k = 1, 2, ..., m denotes an arbitrary hysteretic function of
the three models in Section 2.1 and Pk ∈ {P BW , P B1 , P B2 } the corresponding parametric space.
If the jth component pertains to the bilinear hysteretic model with kinematic hardening, the
jth equation in (12) should be replaced by the simpler model r̃˙j = f˜B1 (r̃j , yj , ẏj , pj ) in Eq. (5).

To establish the dynamical equations for the above structure with hysteresis, the kinetic as
well as potential energies of the structure is firstly introduced as
1 1
E = u̇T Mu̇, V = uT Ku. (13)
2 2
While the damping forces −Cu̇, restoring forces −r and external forces F (t) are treated as
non-conservative forces and can contribute to the non-conservative work Wnc as follows

δWnc = (δu)T F (t) + (δu)T (−Cu̇) + (δy)T (−r) (14)

5
 Figure 2 Illustrative example for description of the connection matrix

where δ is the variation operator and δy = Tδu holds by referring to Eq. (10). Then, application
of the Hamilton’s principle gives rise to
Z t2
δ(E − V ) + δWnc dt = 0, ∀t1 < t2 ⇒ Mü + Cu̇ + Ku + TT r = F (t) (15)
t1

where if fj = f B1 , there should replace rj by r̃j + pj2 tj u with tj the jth row of T and pj2 the
2nd parameter of the jth hysteretic component. At last, combining (12) with (15), complete
description for motion of the structure with hysteresis is established as follows,
Mü + Cu̇ + Ku + TT r = F (t), t ≥ 0


ṙ = f (r, Tu, Tu̇, p); p ∈ P (16)

u(0) = u0 , u̇(0) = u̇0 , r(0) = r0
where u0 , u̇0 , r0 are initial displacements, initial velocities and initial restoring forces. The final
governing equations (16) are nonlinear and for given parameters p, one can get the solution
u(p, t) by the adaptive Runge-Kutta method [24] (or refer to the ’ode45’ function in MATLAB).

2.3 Inverse problem

For identification of the hysteretic parameters, the response data of the structure should be
measured. Now assume that the quantities d are observed at a series of l time points 0 ≤ t1 <
t2 < ... < tl and the measured data at time tk is denoted by d(t ˆ k ). Usually, d are linearly
dependent on the displacements, i.e., d = Lu with L a linear matrix operator. For instance, if
d2
the first acceleration is measured, there is d = ü1 and L = [ dt 2 , 0, ..., 0]. For simplicity, designate
ˆ ˆ ˆ
all the measured data as R̂ = [d(t1 ); d(t2 ); ...; d(tl )] and let R(p) = [d(p, t1 ); d(p, t2 ); ...; d(p, tl )]
be the theoretical solution of the measured quantities with d(p, tk ) = Lu(p, tk ). Then, inverse
identification of the hysteretic parameters can be formulated in a general way as: find parameters
p∗ ∈ P such that the following objective function g(p) is minimized,
g(p) = ||R̂ − R(p)||2 (17)
√
where the norm || · || is defined as ||v|| = v T v. Obviously, the inverse problem (17) is posed
to be a typical nonlinear least-squares problem and in the next section, how to solve such a
problem through the enhanced response sensitivity approach will be elaborated.

3 Enhanced response sensitivity for parameter identification

3.1 Response sensitivity analysis
For an optimization problem with a nonlinear and non-quadratic objective function such as the
one defined in (17), the solution is often inevitably obtained in an iterative manner and the key

6
lies in how to get a proper update δp from known parameters p̄ such that g(p := p̄ + δp) is as
small as possible. To do so, a usual strategy is to linearize the nonlinear function R̂ − R(p) at
p̄ [25, 26], that is,

R̂ − R(p) = δR(p̄) − S(p̄)δp;

δR(p̄) := R̂ − R(p̄),
 
∇p1 d(p, t1 ) ∇p2 d(p, t1 ) · · · ∇pm d(p, t1 )
 ∇p d(p, t2 ) ∇p d(p, t2 ) · · · ∇p d(p, t2 ) 
1 2 m
S(p) = ∇p R(p) :=  ,
 
.. .. . . .
.
 . . . . 
∇p1 d(p, tl ) ∇p2 d(p, tl ) · · · ∇pm d(p, tl ) (18)
 ∂u1 (p,ti ) ∂u1 (p,ti )
· · · ∂u∂p1 (p,ti)

∂pk1 ∂pk2 kNk
 ∂u2 (p,ti ) ∂u2 (p,ti ) ∂u2 (p,ti ) 

 ∂pk1 ∂pk2 · · · ∂pkNk

∇pk d(p, ti ) = L 

.. .. .. .. 

 . . . . 

∂un (p,ti ) ∂un (p,ti ) ∂un (p,ti )
∂pk1 ∂pk2 ··· ∂pkN k

where the key term S(p) is called the response sensitivity matrix and pk = [pk1 , pk2 , ..., pkNk ]T ∈
Pk represent the parameters of the kth hysteretic component. In order to get the specific
expression of the response sensitivity matrix S(p), the following response sensitivity analysis of
the motion equations (16) to the parameter pkj , k ∈ {1, 2, ..., m}, j ∈ {1, 2, ..., Nk } , is necessarily
invoked,
d2 ∂u
(
d ∂u ∂u T ∂r
M dt 2 ∂p
kj
+ C dt ∂pkj + K ∂pkj + T ∂pkj = 0, t ≥ 0
d ∂r ∂r ∂u d ∂u
dt ∂pkj = (∇r f ) ∂pkj + (∇y f )T ∂pkj + (∇ẏ f )T dt ∂pkj + (∇pkj f ) (19)
∂u d ∂u ∂r
(0) = 0, (0) = 0, (0) = 0
∂pkj dt ∂pkj ∂pkj
where ∇r f , ∇y f , ∇ẏ f , ∇pkj f are of the following specific forms with the help of the definition
of f in Eq. (12),
 ∂f   ∂f   ∂f 
1 0 ··· 0 1 0 ··· 0 1 0 ··· 0
∂r1 ∂y1 ∂ ẏ1
∂f2 ∂f2 ∂f2
 0
∂r2
··· 0   0
∂y2
··· 0   0
∂ ẏ2
··· 0
∇r f =  . , ∇y f =  . , ∇ẏ f =  . (20)
     
 . .. . . ..   .. .. . . ..   .. .. . . .. 
. . . .  . . .  . . . 
∂fm ∂fm ∂fm
0 0 ··· ∂rm
0 0 ··· ∂ym
0 0 ··· ∂ ẏm

and
∂fk
∇pkj f = [ 0, ..., 0 , , 0, .., 0]T . (21)
| {z } ∂pkj
of length k−1
∂fk ∂fk ∂fk ∂fk ∂fk ∂fk
Note that the basic terms ∂rk , ∂yk , ∂ ẏk and ∂pk = [ ∂p ,
k1 ∂pk2
, ..., ∂p∂fkNk ] require individual sen-
k
sitivity analysis of the kth hysteretic component/model fk (·) ∈ {f BW (·), f B1 (·), f B2 (·)}, k =
1, 2, ..., m. To do so, the three hysteretic models used in this paper—one Bouc-Wen model (2)
and two bilinear models (5) and (8)— are called for sensitivity analysis as follows,
• Bouc-Wen model (2)
( BW BW
∂f
∂r = −p4 |r|p4 −1 (p2 |u̇| + p3 u̇ sgn(r)), ∂f∂u =0
∂f BW
∂ u̇ = p1 − (p2 r|r|p4 −1 sgn(u̇) + p3 |r|p4 )
(
∂f BW ∂f BW
(22)
p4 −1 |u̇|
∂p1 = u̇, ∂p2 = −r|r|
∂f BW BW
∂p3 = −|r|p4 u̇, ∂f∂p4 = (p2 r|r|p4 −1 |u̇| + p3 |r|p4 u̇) ln(|r| + κ)

7
 1 30

sgn(x) 25
0.5 smoothing, " = 1 7h(x) with " = 1
smoothing, " = 0:1 20 7h(x) with " = 0:1

derivative
function

smoothing, " = 0:001 7h(x) with " = 0:001

0 15

10
-0.5
5

-1 0
-1 -0.5 0 0.5 1 (b) -1 -0.5 0 0.5 1
(a)
x x

Figure 3 Smoothing (a) and derivative (b) of the non-differentiable function sgn(x)

where κ is a small positive constant, e.g., = 10−16 to guarantee the positiveness of |r| + κ
for function ln(·).

• Bilinear model with kinematic hardening (5)

∂r ∂r ∂r ∂ r̃ ∂u ∂ r̃ ∂u ∂ r̃ ∂u
[ , , ]=[ + p2 , + p2 + u, + p2 ],
∂p1 ∂p2 ∂p3 ∂p1 ∂p1 ∂p2 ∂p2 ∂p3 ∂p3
( ˜B1
∂f ∂ f˜B1
∂ r̃ = −p1 ~(r̃sgn(u̇) − p3 )|u̇|, ∂u = 0
∂ f˜B1 (23)
∂ u̇ = p1 − p1 sgn(r̃sgn(u̇) − p3 )
∂ f˜B1 ∂ f˜B1
(
∂p1 = (1 − sgn(r̃sgn(u̇) − p3 ))u̇, ∂p2 =0
∂ f˜B1
∂p3 = p1 ~(r̃sgn(u̇) − p3 )u̇

where ~(·) denotes an asymptotic derivative of sgn(·). It is noteworthy that sgn(·) is non-
differentiable at 0 and to overcome this difficulty, one can reasonably use the smoothing
x+ε x+ε
procedure sgn(x) ≈ |x|+ε for a small positive number ε since limε→0+ |x|+ε = sgn(x). In
doing so, the following form of ~(·) is adopted as derivative of sgn(·) throughout this paper,

ε(1 − sgn(x))
~(x) = (24)
(|x| + ε)2
with ε being a proper and small positive constant. ε is also called the smoothing factor
since it controls the smoothing procedure of the non-differentiable function sgn(·) and
it is noteworthy that the smoothing procedure is only involved in the sensitivity analy-
x+ε
sis. To have a schematic view, the smoothing function |x|+ε and its derivative ~(x) are
shown in Figure 3. Obviously, as  > 0 decreases, the smoothing function gives better
approximation of sgn(·) but its derivative becomes sharper.

• Bilinear model with equal yielding force (8)

( B2
∂f ∂f B2
∂r = −p2 ~(r sgn(u̇) − p3 )|u̇|, ∂u = 0
∂f B2
∂ u̇ = p1 − p2 sgn(r sgn(u̇) − p3 )
( B2
∂f ∂f B2
(25)
∂p1 = u̇, ∂p2 = −sgn(r sgn(u̇) − p3 )u̇
∂f B2
∂p3 = p2 ~(r sgn(u̇) − p3 )u̇

where ~(·) is defined in Eq. (24).

To conclude, the nonlinear sensitivity analysis equation (19) is closely related to the solutions
r, u, u̇ of the original motion equation (16) and therefore, Eq. (19) is often solved in conjunction

8
with Eq. (16) by the Runge-Kutta method [24]. In what follows, how to efficiently and robustly
get the update δp based on the linearization or sensitivity analysis (18) and the Tikhonov
regularization is addressed.

3.2 Enhanced solution strategy through Tikhonov regularization

With the linearization (18), an approximate linear least-squares form of the original objective
function (17) is obtained,
ĝ(δp, p̄) = ||δR(p̄) − S(p̄)δp||2 . (26)
Such a linear least-squares problem (26) may be ill-posed and therefore, the Tikhonov regular-
ization is often introduced to get the update,

δpλ = arg min ||δR(p̄) − S(p̄)δp||2 + λ||δp||2 ,


δp
(27)
= [S (p̄)S(p̄) + λI]−1 ST (p̄)δR(p̄)
T

where I denotes the identity matrix and λ ≥ 0 is the regularization parameter. Still, the value
of the regularization parameter λ remains to be determined.
One usual way to determine the regularization parameter is to use the L-curve method [27]
for which the regularization parameter is designated as λL (p̄); this leads to the general response
sensitivity approach [19]. The L-curve method tries to balance the update norm ||δp|| and the
residual norm ||δR(p̄) − S(p̄)δp|| and for specific tools on the calculation of λL (p̄), refer to the
work of Hansen [28]. In actual, such a choice of λ = λL (p̄) has been shown to work well in
identification of small and some moderate damages in frames and plates [19, 20]. However, the
L-curve method only concerns the approximate linear least-squares problem (26) and does not
account for the original nonlinear problem (18) at all. This can to some extend explain why the
general response sensitivity approach may not work for identification of large structural damage
and needs to be enhanced for better performance and convergence.
To attain better performance, the nonlinearity of the original problem should be additionally
taken into accounted. Note that the key of the above solution procedure lies in the linearization
(18) and to make the approximate linear form agree well with the original nonlinear form,
the update norm ||δp|| should be reasonably small. To measure how well the approximate
objective function ĝ(δp, p̄) agrees with the original objective function g(δp + p̄), an agreement
indicator [21, 25] in introduced as follows,

g(p̄) − g(δp + p̄) ||δR(p̄)||2 − ||δR(p̄ + δp)||2

ρ(δp, p̄) = = (28)
ĝ(0, p̄) − ĝ(δp, p̄) ||δR(p̄)||2 − ||δR(p̄) − S(p̄)δp||2
Good agreement requires satisfaction of the following agreement condition

ρ(δp, p̄) ≥ ρcr ∈ [0.25, 0.75]. (29)

To conclude, the update δp should be reasonably small such that the agreement condition (29)
is satisfied; this is also termed the trust-region constraint.
Now, turn to the Tikhonov regularization (27) and recall the following properties which has
been strictly proved in [21],

lim ||δpλ || = 0, lim ρ(δpλ , p̄) = 1. (30)

λ→+∞ λ→+∞

With (30), it is easily deduced that the trust-region constraint can always be fulfilled by simply
increasing λ to a certain value in the Tikhonov regularization (27); in this way, the final λ is called
the enhanced regularization parameter. Practically, one can get the enhanced regularization
parameter as well as the enhanced update thereof by the following iterative procedure,

9
 Table 1 Algorithmic details for enhanced response sensitivity approach

- set initial parameters p(1) ∈ P,

- define error tolerance tol (e.g., = 10−6 ) for convergence criterion,
- fix the maximum number of iterations N max (e.g., =1000),
- fix trust-region parameters ρcr ∈ [0.25, 075] (e.g., =0.5) and γ > 1 (e.g., =2),
- set the maximum number of steps for trust-region procedure N tr, (e.g.,=20),
- load the measured response data R̂,
- for k = 1 : N max
- solve (16) and (19) to get response R(p(k) ) and response sensitivity S(p(k) ),
- compute response residual δR = R̂ − R(p(k) ),
- use L-curve method to get the initial regularization parameter λL (p(k) ),
- for i = 1 : N tr % Trust-region procedure
- λ = γ i−1 λL (p(k) ),
- compute the update δp = [ST (p(k) )S(p(k) ) + λI]−1 ST (p(k) )δR,
- if p(k) + δp ∈/ P continue,
- solve (16) to get response R(p(k) + δp),
- compute new response residual δRnew = R̂ − R(p(k) + δp),
||δR||2 −||δRnew ||2
- calculate the agreement indicator ρ = ||δR|| 2 −||δR−S(p(k) )δp||2 ,

- if ρ ≥ ρcr break,
- end for
- update stiffness parameters p(k+1) = p(k) + δp,
- if ||δp||/||p(k+1) || ≤ tol break.
- end for

1. Set the initial value as λ = λL (p̄) from the L-curve method,

2. Get δpλ from (27),

3. Calculate the agreement indicator ρ(δpλ , p̄) as in (28),

4. If the agreement condition (29) is satisfied, terminate the iteration and acquire the en-
hanced update δpλ ; otherwise, increase λ up to a factor γ > 1, i.e., λ ⇐ γλ and go to step
2.
Finally, in conjunction with the sensitivity analysis, the enhanced response sensitivity approach
is established for which the convergence proof has already been presented in reference [21] and
algorithmic details are given in Table 1.

4 Numerical tests
Two examples—a hysteretic single-degree-of-freedom (SDOF) system and a hysteretic multiple-
degree-of-freedom (MDOF) system—are concerned in this section. The algorithm in Table 1 is
utilized for hysteretic parameter identification with the algorithmic parameters tol, N max, ρcr , γ
and N tr taking the default values in the following brackets. The measured data is acquired
from numerical simulation along with the consideration of measurement noises. Let dk =
[dk (t1 ), dk (t2 ), ..., dk (tl )]T denote the simulated data corresponding to an arbitrary quantity dk
(e.g., acceleration or displacement) and then, considering the measurement noise of level enoise ,
the measured data dˆk is calculated as

dˆk = dk + enoise ∗ std(dk ) ∗ Rand (31)

10
 Figure 4 Model of SDOF system with hysteresis

Table 2 Values of parameters for the three hysteretic models

Hysteretic model Values of parameters

Bouc-Wen model (2) (p1 = 2, p2 = 1, p3 = 0.5, p4 = 2)
Bilinear model with kinematic hardening (4) (p1 = 1, p2 = 2, p3 = 1)
Bilinear model with equal yielding force (8) (p1 = 3, p2 = 1, p3 = 1)

where std(v) denotes the standard deviation of vector v and Rand is a l × 1 random vector
pertaining to the standard normal distribution.

4.1 SDOF system with hysteresis

Consider a SDOF system with mass m = 1, stiffness k = 1 and damping c = 0.2 over the time
interval [0,10] in Figure 4. The external loading is F (t) = 2 cos(t) and the restoring force r is
provided by one of the three hysteretic models as addressed in Section 2.1. Initial displacement,
velocity and restoring force of the system are all zero. For later analysis, parameters of the three
hysteretic models are given in Table 2.
1. Numerical response and sensitivity analysis
For the nonlinear SDOF system connected by different kinds of hysteretic models, the gov-
erning motion equation can be established by referring to Eq. (16) with m = n = 1, T = 1.
Then, the responses caused by different hysteretic models can be numerically obtained by the
Runge-Kutta method or the ’ode45’ function of MATLAB and before doing so, the second-order
differential equations (16) are simply transformed into the corresponding first-order differential
equations as follows

T u̇ = v,
Mü + Cu̇ + Ku + T r = F (t) ⇔
v̇ = M−1 (F (t) − Ku − Cv − TT r)

and so is the treatment of equation (19). Final numerical results on the accelerations, displace-
ments and restoring forces for different hysteretic models are plotted in Figure 5 where from
here on, ’bilinear 1’ refers to the bilinear model with kinematic hardening and ’bilinear 2’ the
bilinear model with equal yielding force. The two bilinear hysteretic models can give similar
acceleration, displacement and restoring force results; while the restoring force of the Bouc-Wen
model is bounded which is obvious from the hysteretic loop graph (see Figure 1(a)).
Next, response sensitivity is analyzed by resorting to Eq. (19). Herein, for brevity, only
the sensitivity of the acceleration to parameters are concerned. For the SDOF system with
Bouc-Wen hysteresis, the sensitivity results are graphically shown in Figure 6. As is seen, the
sensitivity is smoothly varying with time and the sensitivity to p4 or the exponential coefficient
n is smaller than that to other three parameters, indicating that the acceleration is often less
sensitivity to the exponential coefficient. On the other hand, Figure 7 and Figure 8 present

11
 3 3 3

2 2 2

restoring force
displacement
1 1 1
acceleration

Bouc-Wen
0 0 0 Bilinear 1
Bilinear 2
-1 -1 -1

-2 -2 -2

-3 -3 -3
0 5 10 0 5 10 0 5 10
t t t

Figure 5 Acceleration (left), displacement (middle) and restoring force (right) responses of
SDOF system with different hysteresis

2 1

0
@p1

@p2
@u

@u
B

B
0
-1

-2 -2
0 2 4
t 6 8 10 0 2 4 t 6 8 10
1 0.2

0 0
@p4
@p3

@u
B
@u
B

-1 -0.2

-2 -0.4
0 2 4 6 8 10 0 2 4 t 6 8 10
t

Figure 6 Sensitivity of acceleration to four parameters for SDOF system with Bouc-Wen hys-
teresis

2 4 2

1 2 1
@p1

@p2

@p3

"=1
@u

@u

@u
B

0 0 0
" = 1e ! 01
" = 1e ! 03
-1 -2 -1 " = 1e ! 05

-2 -4 -2
0 5 10 0 5 10 0 5 10
t t t

Figure 7 Sensitivity of acceleration to three parameters for SDOF system with kinematic-
hardening bilinear hysteresis under different smoothing factors ε

2 1 1

1 0.5
0.5
"=1
@p1

@p2

@p3

" = 1e ! 01
@u

@u

@u
B

0 0
" = 1e ! 03
0 " = 1e ! 05
-1 -0.5

-2 -1 -0.5
0 5 10 0 5 10 0 5 10
t t t

Figure 8 Sensitivity of acceleration to three parameters for SDOF system with equal-yielding-
force bilinear hysteresis under different smoothing factors ε

12
 Table 3 Scenarios for parameter identification of the SDOF system with different hysteresis

Scenario hysteretic model measurements noise level

1-1 / 2-1 / 3-1 Bouc-Wen / Bilinear 1 /Bilinear 2 ü 0
1-2 / 2-2 / 3-2 Bouc-Wen / Bilinear 1 /Bilinear 2 ü 5%
1-3 / 2-3 / 3-3 Bouc-Wen / Bilinear 1 /Bilinear 2 ü 10%
1-4 / 2-4 / 3-4 Bouc-Wen / Bilinear 1 /Bilinear 2 u 0
1-5 / 2-5 / 3-5 Bouc-Wen / Bilinear 1 /Bilinear 2 u 5%
1-6 / 2-6 / 3-6 Bouc-Wen / Bilinear 1 /Bilinear 2 ü 10%
1-7 / 2-7 / 3-7 Bouc-Wen / Bilinear 1 /Bilinear 2 u, r 10%

4 4 4

3
2 2
2

restoring force
displacement
Acceleration

Bouc-Wen
1 Bilinear 1
0 0 Bilinear 2
0

-1
-2 -2
-2

-4 -3 -4
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
t t t

Figure 9 Measurement data for SDOF system with 10% measurement noise

the results for the two bilinear hysteretic models. Note that for sensitivity analysis of bilinear
hysteresis, the smoothing procedure as illustrated around Eq. (24) is inevitably invoked and
therefore, different smoothing factors are further studied to see the effects; these are also shown
in Figures 7 and 8. For different smoothing factors, different sensitivity results are obtained
and smoothness of the sensitivity degrades with decrease of smoothing factor ε; particularly, the
sensitivity to the third parameter p3 for both bilinear models is more sensitive to the smoothing
factor and when ε ≤ 0.001, jump will almost arise in the sensitivity results. It is also observed
that for ε ≤ 0.001, the sensitivity results almost keep unchanged with ε. Generally, larger ε leads
to smoother derivative ~(·) and therefore, often corresponds to faster solution of the sensitivity
equation (24) but worse approximation of the sgn(·) function. Thus, for this SDOF system, one
can reasonably set the smoothing factor to be 0.001 ≤ ε ≤ 0.1.
2. Hysteretic parameter identification under different measurements and various noise levels
Herein, the present enhanced response sensitivity approach will be used to inverse identifi-
cation of the hysteretic parameters from the measured data and seven identification scenarios
for each hysteretic SDOF system, that is to say, 21 scenarios in total are considered as specified
in Table 3. The measurement data is obtained from numerical simulation at the sampling rate
of 500Hz during time interval [0,10] and measurement noise is enforced as in Eq. (31). For
illustrative purpose, the measured data under noise level enoise = 10% regarding acceleration,
displacement and restoring force for the SDOF system with three different hysteretic models is
depicted in Figure 9.
To identify the hysteretic parameters from the measured data, the initial values are set to
(1) (1) (1) (1) (1) (1) (1)
be: (p1 = 3, p2 = 2, p3 = 1.5, p4 = 3) for the Bouc-Wen model, (p1 = 2, p2 = 3, p3 = 2)
(1) (1)
for the bilinear model with kinematic hardening (or ’Bilinear 1’ model) and (p1 = 5, p2 =
(1)
2, p3 = 2) for the bilinear model with equal yielding force (or ’Bilinear 2’ model). Specifically,
for the two bilinear models, the smoothing factor is taken as ε = 0.1 at first. After carrying
out the algorithmic procedure in Table 1, the hysteretic parameters can be identified and the
results are listed in Table 4 in detail where #iter represents the number of iterations and the

13
 Table 4 Identified results for the 21 scenarios of the SDOF hysteretic system

Scenario hysteresis identified parameters #iter eREL

1-1 Bouc-Wen (1.9999,1.0001,0.4998,2.0002) 19 0.04%
1-2 Bouc-Wen (2.0167,0.9942,0.5262,1.9205) 18 5.24%
1-3 Bouc-Wen (2.0559,1.0025,0.5758,1.8598) 21 16.60%
1-4 Bouc-Wen (1.9997,1.0003,0.4993,2.0009) 22 0.14%
1-5 Bouc-Wen (2.0309,0.9519,0.5891,1.9655) 21 17.82%
1-6 Bouc-Wen (2.0749,0.9677,0.6067,1.9153) 20 21.34%
1-7 Bouc-Wen (2.0129,0.9961,0.5346,1.8884) 19 6.92%
2-1 Bilinear 1 (1.0002,1.9996,1.0003) 14 0.03%
2-2 Bilinear 1 (1.0017,1.9943,1.0041) 24 0.41%
2-3 Bilinear 1 (1.0070,1.9874,1.0089) 24 0.89%
2-4 Bilinear 1 (1.0000,2.0000,1.0000) 14 0.00%
2-5 Bilinear 1 (1.0011,1.9981,1.0011) 15 0.11%
2-6 Bilinear 1 (0.9985,2.0040,0.9964) 17 0.36%
2-7 Bilinear 1 (0.9994,1.9992,1.0002) 30 0.02%
3-1 Bilinear 2 (3.0001,0.9999,0.9999) 13 0.01%
3-2 Bilinear 2 (2.9992,1.0008,1.0033) 14 0.33%
3-3 Bilinear 2 (2.9935,0.9962,1.0049) 23 0.49%
3-4 Bilinear 2 (2.9997,1.0001,1.0008) 17 0.08%
3-5 Bilinear 2 (2.9989,0.9977,1.0011) 18 0.23%
3-6 Bilinear 2 (2.9958,1.0091,1.0059) 22 0.91%
3-7 Bilinear 2 (2.9952,1.0004,0.9999) 15 0.04%

relative error eREL is computed as follows,

|p̂i − pex
i |
eREL = max ex × 100%
i pi
with p̂i denoting the identified parameter and pex i corresponding exact value (see Table 2).
As is seen, all scenarios, with or without measurement noise, using acceleration or displace-
ment data are well identified within 30 steps, indicating that the present enhanced response
sensitivity approach is very efficient for hysteretic parameter identification. More specifically,
for the Bouc-Wen model, the relative error or identification precision is greatly affected by the
measurement noise and usage of the acceleration data can lead to better accuracy than usage
of the displacement data. In case of 10% measurement noise, the relative error arises up to
16.60% for the acceleration data and 21.34% for the displacement data, nevertheless, these are
acceptable since 10% measurement noise has led to severe perturbation in comparison with the
exact data in Figure 9. Also noteworthy is that, usage of both the displacement and the restor-
ing force data can give rise to more satisfactory results with relative error being 6.92%; this
is reasonable since in principle, more data would always result in more robust identification.
While for the two bilinear models, the identification results are shown to be insensitive to the
measurement noise since even at the noise level of 10%, the relative identification error does not
exceed 1% and all identified results are quite satisfactory; robustness of the enhanced response
sensitivity approach along with the time-domain data for parameter identification of bilinear
hysteretic models is perfectly exhibited.
3. Effect of initial parameters and smoothing factors
To further see whether the present identification approach is sensitive to initial choice of
parameters p(1) , different initial parameters are considered for scenarios 1-2 / 2-2 /3-2 involving
the three hysteretic models and using the acceleration data with 5% measurement noise. The

14
Table 5 Identified results using 5%-noised acceleration data under different initial parameters

Scenario initial parameters identified parameters #iter

1-2 (0.1,0.1,0.05,1.3) (2.0168,0.9942,0.5263,1.9204) 23
1-2 (0.5,0.5,0.25,1.3) (2.0168,0.9942,0.5263,1.9204) 21
1-2 (5,5,3,4) (2.0167,0.9942,0.5262,1.9205) 25
1-2 (10,10,8,4) — > 30
2-2 (0.25,0.25,0.25) (1.0006,1.9946,1.0040) 30
2-2 (0.5,0.5,0.5) (1.0017,1.9943,1.0041) 29
2-2 (4,2,1) (1.0014,1.9948,1.0039) 18
2-2 (10,10,10) — >30
3-2 (0.05,0.02,0.02) (2.9991,1.0008,1.0033) 17
3-2 (0.5,0.2,0.2) (2.9992,1.0008,1.0033) 17
3-2 (4,2,3) (2.9991,1.0008,1.0033) 18
3-2 (10,8,4) — >30

Table 6 Identified results for bilinear models using 5%-noised acceleration data with different
smoothing factors

scenario 2-2 scenario 3-2

ε
identified parameters #iter identified parameters #iter
1e+01 — >50 — >50
1e-03 (1.0019,1.9939,1.0043) 32 (2.9991,1.0009,1.0033) 14
1e-05 (1.0017,1.9943,1.0041) 33 (2.9992,1.0008,1.0033) 23
1e-07 (1.0017,1.9943,1.0041) 29 (2.9991,1.0008,1.0033) 26
1e-09 — >50 — >50

smoothing factor is still set to ε = 0.1 for the two bilinear models. Details on the initial
parameters as well as corresponding identified results are presented in Table 5 where ’—’
means unusual termination or the solution with relative error exceeding 200%. As is observed,
almost all initial parameters can lead to convergent and satisfactory identified results, except
those are too large with respect to exact values; while using small or too small initial parameters
can always give good identification. This can be intuitively explained as follows: for too large
initial parameters, the effect of the initial hysteretic model tends to be a linear elastic spring
model which is substantial different from the actual hysteretic behavior, but for small and even
very small initial parameters, the hysteretic behavior is completely posed. As a consequence,
smaller initial parameters are preferred for hysteretic parameter identification.
In addition, the effect of the smoothing factor for bilinear hysteretic parameter identification
is also investigated. Herein, scenarios 2-2 and 3-2 are again considered with initial parameters
being (0.5,0.5,0.5) for the ’bilinear 1’ model and (0.5,0.2,0.2) for the ’bilinear 2’ model. Different
smoothing factors including ε =1e+01, 1e-03, 1e-05, 1e-07 and 1e-09 —the case of 1e-01 is
already presented in Table 5— are considered for parameter identification and the results are
listed in Table 6 where ’—’ means unusual termination or wrong results. Evidently, too large
ε ≥1e+01 or too small ε ≤1e-09 would make the parameters unidentifiable, nevertheless, the
present approach has shown considerably good identification of bilinear hysteretic parameters
for a wide range of smoothing factors 1e+01> ε ≥1e-07.

15
 Figure 10 Simplified model of seven-storey building with bilinear hysteresis

4.2 MDOF system with bilinear hysteresis

Herein, a simplified n−storey building (n=7) as presented in Figure 10 is concerned. Basic
structural properties are: concentrated mass of each storey m = 1, stiffness k = 1 and linear
damping c = 0.1, and then, mass, stiffness and damping matrices of the MDOF system are
obtained as follows,
   
2 −1 2 −1
 −1 . . . . . .  −1 . . . . . .
   
 
M = mIn , K = k   ..

 , C = c 
 ..


 . 2 −1   . 2 −1 
−1 1 n×n
−1 1 n×n

where In denotes the identity matrix of order n. Usually, the hysteretic components are used for
the purpose of structural seismic isolation [2, 3]. More often, the bilinear hysteretic components
with kinematic hardening are preferred due to its relatively simple realization by spring and
friction models. To illustrate, the equivalent rheology description [29] of the bilinear hysteretic
model with kinematic hardening is shown in Figure 11 where the parameters k1 , k2 , F0 are the
same with those in Figure 1(b).
Now, turn to the n-storey building and three bilinear hysteretic components are connected to
the building. For later analysis, two connection cases (see Figure 10 for details) are considered:
for case I, the three hysteretic components are placed at the 1st, 3rd and 5th storeys, or precisely,
the connection matrix is  
1 0 0 0 0 0 0
TI =  0 −1 1 0 0 0 0  ;
0 0 0 −1 1 0 0
for case II, the hysteretic components are located at the 2nd, 4th and 6th storeys with the

16
 Figure 11 Equivalent rheology model for bilinear hysteresis with kinematic hardening

Table 7 Scenarios for parameter identification of the MDOF system

Scenario connection case measurements noise level

M1 Case I ü6 , ü7 0
M2 Case I ü6 , ü7 5%
M3 Case I ü6 , ü7 10%
M4 Case II ü4 , ü7 0
M5 Case II ü4 , ü7 5%
M6 Case II ü4 , ü7 10%

connection matrix being

 
−1 1 0 0 0 0 0
TII =  0 0 −1 1 0 0 0  .
0 0 0 0 −1 1 0

Actual parameters of the hysteretic components are given in Figure 10. For dynamic modeling
of the n-storey building with hysteresis over time interval [0,30], initial states for displacements,
velocities and restoring forces are all zero, and the building pertains to the following ground
motion acceleration (see Figure 12(a) for a schematic view),

 (t/t0 )2 , t ≤ t0


ag (t) = 2 cos(t)w(t; t0 , tn ); w(t; t0 , tn ) = 1, t0 < t ≤ tn

 −0.2(t−tn )
e , t > tn

where w(t; t0 , tn ) is a window function such that the excitation mainly occurs during [t0 , tn ] and
in this example, it is set that t0 = 2, tn = 6. Then, external forces are F (t) = −ag (t)M · 1n×1
with 1n×1 a n × 1 vector of 1. After all the preparation, the governing equations for motion of
the building considering the two hysteretic connection cases can be established by referring to
Eq. (16). For later parameter identification of p = [p1 , p2 , p3 ], accelerations are measured at
the sampling rate of 100Hz and it is noteworthy that the measured acceleration in this example
of the kth storey is the absolute acceleration, i.e., âk = û¨k + ag or equivalently, the measured
¨
relative acceleration should be ûk = âk − ag .
Next, six identification scenarios as shown in Table 7 are investigated for the building/MDOF
system. The measured data with 10% noise for the two cases are plotted in Figure 12(b) and
(c) for visual purpose. For practical identification of the bilinear hysteretic parameters in this
example, the smoothing factor is set to be ε = 0.1 for both cases. Then, the enhanced response
sensitivity approach in Table 1 is applied for parameter identification with the initial parame-
ters being the same p(1) =[0.25,0.25,...,0.25] (of length 9) for the two hysteresis connection cases.
Thereafter, parameters can be identified and specific results are listed in Table 8. Obviously,
the parameters are perfectly identified for all scenarios: the identification relative errors for

17
 2

-2
0 5 10 15 20 25 30
(a) ground motion acceleration t
5 u
B6
u
B7

-5
0 5 10 15 20 25 30
(b) acceleration data for case I t
5 u
B4
u
B7

-5
0 5 10 15 20 25 30
(c) acceleration data for case II t

Figure 12 Plots for (a) ground acceleration, (b) measured data for case I with 10% noise and
(c) measured data for case II with 10% noise

Table 8 Identified results for MDOF system with bilinear hysteresis

Scenario identified parameters #iter eREL

M1 (0.9998,2.0003,0.9998,0.5000,0.4999,0.5000,0.5001,0.9999,0.5001) 53 0.02%
M2 (0.9891,2.0132,0.9956,0.5147,0.4965,0.5157,0.4956,1.0002,0.4948) 60 3.14%
M3 (1.0084,2.0570,0.9854,0.4853,0.4743,0.5033,0.4800,1.0086,0.4804) 26 5.14%
M4 (1.0005,0.9997,1.0004,0.9999,1.0001,0.9999,0.5000,0.5000,0.5000) 40 0.05%
M5 (1.0070,1.0007,1.0075,1.0037,0.9973,1.0030,0.5085,0.4944,0.5055) 57 1.70%
M6 (0.9738,1.0145,1.0022,1.0025,0.9968,1.0013,0.4939,0.5012,0.4987) 46 2.62%

18
 1.5
parameter value parameter value
3

1
2

1 0.5

0
30 0
50
25
40
20

15 30
iteration no. iteration no.
10
20
5
1
32 10
54
0 76
98
parameter no. 1
5432
(a) (b)
0 9876
parameter no.

Figure 13 Convergence procedure for parameter identification of (a) scenario M3 and (b)
scenario M6

both hysteresis cases are no more than 0.05% in case of no noise (scenarios M1 and M4) and are
less than 3.4% at the noise level of 5% (scenarios M2 and M5); even under 10% measurement
noise (scenarios M3 and M6), the parameters are identified with relative errors not exceeding
6%. As regards the number of iterations, parameters of all six scenarios are identified within
60 steps. Specifically, to have a schematic impression on the convergence procedure, hysteretic
parameters at each iteration step for scenarios M3 and M6 are plotted in Figure 13. Indeed,
good identifiability and satisfactory efficiency of the proposed approach have been verified for
hysteretic parameter identification of the MDOF system.

5 Conclusions
A new framework for parameter identification of various hysteretic models as well as struc-
tures with hysteresis by the enhanced response sensitivity approach has been established in this
work. Particular attention is paid to the bilinear hysteretic models due to the inherent non-
differentiability of bilinear hysteretic functions and smoothing procedure is introduced to enable
sensitivity analysis of the bilinear models. Numerical examples on a SDOF system with different
kinds of hysteretic models and a MDOF system with bilinear hysteretic components are studied
and results show that

• hysteretic parameters of various hysteretic models including Bouc-Wen models, bilinear

model with kinematic hardening and bilinear model with equal yielding force are well
identified by the present enhanced response sensitivity approach;

• the present approach works well even for 10% measurement noise;

19
 • the present approach can admit a wide range of initial parameters and smaller initial
parameters are shown to be better choice;

• for bilinear models, the smoothing factor should be properly selected, i.e., its value should
be neither too large nor too small;

• parameters of MDOF hysteretic structures are also well identifiabe by the present ap-
proach.

Thus, it is believed that the present enhanced response sensitivity approach would be a viable
and efficient tool for general hysteretic parameter identification.

Acknowledgement
The present investigation was performed under the support of National Natural Science Foun-
dation of China (No. 11572356) and Guangdong Province Natural Science Foundation (No.
2015A030313126).

References
[1] Macki JW, Nistri P, Zecca P. Mathematical models for hysteresis. SIAM Review 1993,
35(1): 94-123.

[2] Katsaras CP, Panagiotakos TB, Kolias B. Restoring capability of bilinear hysteretic seismic
isolation systems. Earthquake engineering and structural dynamics 2008, 37: 557-575.

[3] Fragiacomo M, Rajgelj S, Cimadom F. Design of bilinear hysteretic isolation systems. Earth-
quake engineering and structural mechanics 2003, 32: 1333-1352.

[4] Berger EJ, Krousgrill CM. On friction damping modeling using bilinear hysteresis elements.
ASME journal of vibration and acoustics 2002, 124: 367-375.

[5] Ismail M, Ikhouane F, Rodellar J. The hysteresis Bouc-Wen model, a survey. Archives of
computational methods in engineering 2009, 16: 161-188.

[6] Wen YK. Method of random vibration of hysteretic systems. ASCE Journal of the Engi-
neering Mechanics Division 1976, 102: 249-263.

[7] Belbas SA, Mayergoyz ID. Optimal control of dynamical systems with Preisach hysteresis.
International journal of non-linear mechanics 2002, 37: 1351-1361.

[8] Yar M, Hammond JK. Modeling and response of bilinear hysteretic systems. Journal of
engineering mechanics 1987, 113(7): 1000-1013.

[9] Ikhouane F., Rodellar J. On the hysteretic Bouc-Wen model. Part I: Forced limit cycle
characterization. Nonlinear dynamics 2005, 42: 63-78.

[10] Wu M, Smyth A. Real-time parameter estimation for degrading and pinching hysteretic
models. International journal of non-linear mechanics 2008, 43: 822-833.

[11] Lin JS, Zhang Y. Nonlinear structural identification using extended Kalman filter. Com-
puters and structures 1994, 52: 757-764.

20
 [12] Corigliano A, Mariani S. Parameter identification in explicit structural dynamics: performa-
ce of the extended Kalman filter. Computer methods in applied mechanics and engineering
2004, 193: 3807-3835.
[13] Chang CC, Shi Y. Identification of time-varying hysteretic structures using wavelet mul-
tiresolution analysis. International journal of non-linear mechanics 2010, 45: 21-34.
[14] Ni YQ, Ko JM, Wong CW. Identification of nonlinear hysteretic isolators from periodic
vibration tests. Journal of sound and vibration 1998, 217(4): 737-756.
[15] Yar M, Hammond JK. Parameter estimation for hysteretic systems. Journal of sound and
vibration 1987, 117: 161-172.
[16] Sues RH, Mau ST, Wen YK. Systems identification of degrading hysteretic restoring forces.
ASCE Journal of the Engineering Mechanics 1988, 114: 833-846.
[17] Roberts JB, Sadeghi AH. Sequential parameter identification and response of hysteretic
oscillators with random excitation. Structural safety 1990, 8: 45-68.
[18] Loh CH, Chung ST. A three-stage identification approach for hysteretic systems. Earthquake
engineering and structural dynamics 1993, 22: 129-150.
[19] Lu ZR, Law SS. Features of dynamic response sensitivity and its application in damage
detection. Journal of sound and vibration 2007, 303: 305-329.
[20] Fu YZ, Lu ZR, Liu JK. Damage identification in plates using finite element model updating
in time domain. Journal of sound and vibration 2013, 332: 7018-7032.
[21] Lu ZR, Wang L. An enhanced response sensitivity approach for structural damage iden-
tification: convergence and performance. International journal for numerical methods in
engineering 2017, DOI: 10.1002/nme.5502.
[22] Brewick PT, Masri SF. An evaluation of data-driven identificaton strategies for complex
nonlinear dynamic systems. Nonlinear dynamics 2016, 85: 1297-1318.
[23] Erlicher S, Point N. Thermodynamic admissibility of Bouc-Wen type hysteresis models.
Comptes rendus mecanique 2004, 332: 51-57.
[24] Cash JR, Karp AH. A variable order Runge-Kutta method for initial value problems with
rapid varying right-hand sides. ACM Transactions on mathematical software 1990, 16: 201-
222.
[25] More JJ. The Levenberg-Marquardt algorithm: implementation and theory. Chapter Nu-
merical analysis, volume 630 of the series Lecture Notes in Mathematics 1978, pp: 105-116.
[26] Wang L, Liu JK, Lu ZR. Incremental response sensitivity approach for parameter identifica-
tion of chaotic and hyperchaotic systems. Nonlinear dynamics 2017, DOI: 10.1007/s11071-
017-3442-0.
[27] Hansen PC. Analysis of discrete ill-posed problems by means of the L-curves. SIAM Review
1992, 34(4): 561-580.
[28] Hansen PC. Regularization tools—A matlab package for analysis and solution of discrete
ill-posed problem. Numerical algorithms 1994, 6(1): 1-35.
[29] Sivaselvan MV. Hysteretic models with stiffness and strength degradation in a mathematical
programming format. International journal of non-linear mchanics 2013, 51: 10-27.

21

View publication stats