Nonlinear Dyn (2019) 95:1165–1188

https://doi.org/10.1007/s11071-018-4622-2

ORIGINAL PAPER

On choosing state variables for piecewise-smooth dynamical

system simulations
Jin-Song Pei · Joseph P. Wright ·
François Gay-Balmaz · James L. Beck ·
Michael D. Todd

Received: 11 May 2018 / Accepted: 2 October 2018 / Published online: 14 November 2018
© Springer Nature B.V. 2018

Abstract Choosing state variables in a state-space order to develop accurate and efficient time integra-
representation of a nonlinear dynamical system is a tion, we incorporate MATLAB’s state event location
nonunique procedure for a given input–output rela- algorithm, which is a mathematically sound numerical
tionship and therefore a potentially challenging task. solver that deserves to be better known in the engi-
It can be even more challenging when there are neering mechanics community. We show that differ-
piecewise-defined restoring forces, as in bilinear hys- ent choices of state variables can affect state event
teresis or Bouc–Wen models, which are just two of implementation, which in turn can significantly affect
many such engineering mechanics models. Using vari- accuracy and efficiency, as judged by tolerance propor-
ous piecewise-smooth models, we make use of flow- tionality and work–accuracy diagrams. Programming
and effort-controlled system concepts, common to details of state event location are included to facili-
bond graph theory, to initiate our state variable selec- tate application to other models involving piecewise-
tion task, and we view numerical simulation as being defined restoring forces. In particular, one version of
within the framework of hybrid dynamical systems. In the Bouc–Wen–Baber–Noori (BWBN) class of mod-
els is implemented as a demonstration.
J.-S. Pei (B)
School of Civil Engineering and Environmental Science, Keywords Restoring force model · Bilinear hysteresis
University of Oklahoma, Norman, Oklahoma 73019-1024,
USA
model · Bouc–Wen model · BWBN model · Hybrid
e-mail: [email protected] dynamical system · State event location algorithm ·
J. P. Wright
Bond graph theory · Flow-controlled system ·
Division of Applied Science, Weidlinger Associates Inc., Effort-controlled system
New York, NY 10005, USA
F. Gay-Balmaz
CNRS LMD IPSL, Ecole Normale Supérieure de Paris, 24 1 Introduction
Rue Lhomond, 75005 Paris, France
J. L. Beck 1.1 Motivating example: state event location
Department of Computing and Mathematical Sciences, algorithm
California Institute of Technology, Mail Code 9-94,
Pasadena, CA 91125, USA
This study builds on previous work developing accurate
M. D. Todd
Department of Structural Engineering, University of
and efficient numerical methods for simulating nonlin-
California, San Diego, 9500 Gilman Drive, ear dynamical systems that include piecewise-defined
Mail Code 0085, La Jolla, CA 92093, USA restoring forces [34]. Throughout this paper, the fol-

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

1166 J.-S. Pei et al.

lowing equation of motion is used based on Newton’s Solved restoring force-displacement relation
1.5
second law:
1
m ẍ + r = u (1)
(xy , ry )

Restoring force
0.5
where x = x(t) is the displacement of the mass m at
time t, ẍ is the acceleration, r is the restoring force 0
generated by the elements connected to the mass that
tends to restore it to its neutral position and u = u(t) -0.5
is an arbitrary excitation force applied to the mass. For
example, Eq. (1) could describe steady-state or tran- (-xy ,- ry )
-1
sient dynamic response due to either force or base exci- Time selected by MATLAB ode45
Event Type #1 solved by MATLAB ode45

tation. For simplicity in this study, we set mass m = 1,

Event Type #2 solved by MATLAB ode45

-1.5
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x(0) = 0, ẋ(0) = 0 unless otherwise specified. Displacement
In this study, r = r (x) is a piecewise function of
x, making the dynamical system a hybrid dynamical Fig. 1 Trilinear elastic model: central piece has stiffness k = 1;
system, meaning a system with interactions of both the two other pieces have αk = 0.3; this result comes from SDOF
subjected to one cycle of sinusoidal excitation force u(t) (shown
continuous and discrete state/algebraic variables. We at top of Fig. 2); simulation used “event option” under MATLAB
assume that the displacement, velocity and restoring ode45 with RelTol = 10−3 and AbsTol = 10−12 . Event Types
force are continuous functions of time, meaning the #1 and 2 are defined in Eqs. (2) and (3), respectively
transition from one piece to the next is continuous, and
that they are differentiable within each piece. However,
the instantaneous stiffness, defined piecewise as dr/dx, quantities (displacement, velocity and restoring force
may change abruptly from piece to piece. As needed in this example). This algorithm was implemented by
herein, we will review ideas and supply the required combining the event option with ode45 in MATLAB.
definitions as they apply to a specific hybrid dynamical For the trilinear elastic model in Fig. 1, the event option
system (e.g., [11,30]). needs to implement two event functions:
In [34], we focused on the Ramberg–Osgood model
which has a piecewise-smooth algebraic form r = Event Type #1 : x(t) − x y = 0 (2)
r (x). In this study, more complicated models such
Event Type #2 : x(t) + x y = 0 (3)
as bilinear hysteresis or Bouc–Wen models [4,6,7,31]
will be investigated, where r is defined piecewise in
either algebraic or differential format. which are solved for the time instances when these
As a motivating example involving “state event loca- event functions are satisfied so that the numerical inte-
tion” [27], we present results of a single-degree-of- gration will stop and then restart at all time instances
freedom (SDOF) simulation with the trilinear elastic that are discovered (i.e., not known a priori) in the pro-
model shown in Fig. 1. The force–displacement rela- cess of numerical simulation. These events are called
tion, r versus x, is defined by three linear pieces joined state events as they are functions of the state variables
at two points (x y , r y ) and (− x y , − r y ), with x y = 0.75 (x in this case).
and r y = 0.75 where x y denotes yield displacement Figure 2 presents simulation results for the trilinear
and r y denotes yield force. Thus, the slope of the cen- elastic model. Time instances corresponding to the two
tral piece defines a spring stiffness k = 1. The two types of event are highlighted in both Figs. 1 and 2.
other pieces have equal stiffness values αk = 0.3. Although this is a simple model, with the excitation
We utilize the state event algorithm in [27] for accu- force u(t) being the smooth sinusoid shown at the top
racy and efficiency reasons. This algorithm ends time of Fig. 2, the seemingly simple idea of state event loca-
stepping “exactly” (within an approximation error) at tion requires care to be taken before, during and after
each time when yield displacement occurs, such a any event occurs. The procedure was discussed in [34]
time being called an “event,” and then restarts there using both the Ramberg–Osgood and Bingham models,
with accurately calculated values of all time-dependent and will be further elucidated in this study using other

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

On choosing state variables 1167

Time history of prescribed excitation force

chosen as state variables when formulating the state
Excitation Force
1
0.5
Prescribed excitation time history
equations corresponding to Eq. (1):
0
 
 
-0.5 y(1) x
y= = (5)
-1 y(2) ẋ
0 10 20 30 40 50 60 70 80 90 100      
Time ẏ(1) ẋ y(2)
ẏ = = = 1 = f (y, u, t)
ẏ(2) ẍ m (u − r (y(1)))
Time history of solved displacement
2
Displacement

1
Time selected by MATLAB ode45
Event Type #1 solved by MATLAB ode45
(6)
Event Type #2 solved by MATLAB ode45

0
-1 This choice of state variables and corresponding pair of
-2 first-order ODEs was used when implementing the state
0 10 20 30 40 50 60 70 80 90 100
Time
event location algorithm for the example just presented.
Time history of solved velocity In particular, the displacement x(t) was used in defining
0.4
0.2
the two event functions, which satisfies a MATLAB
requirement, namely that event functions need to be
Velocity

0
-0.2
defined as (simple) functions of the state variables.
-0.4 For piecewise restoring force models that are more
0 10 20 30 40 50 60 70 80 90 100
Time
complicated, appropriate choices of state variables are
Time history of solved restoring force not always this straightforward. Reference [33] vividly
2
Restoring force

acknowledges this fundamental challenge concerning

1
state-space formulation: “It is, however, impossible to
0
-1
deduce a priori, in physical terms, what will be the
-2
state. This, indeed, is a very difficult problem even for
0 10 20 30 40 50 60 70 80 90 100 relatively simple systems, and it appears to be the cause
Time
for much of the reluctance of introducing this concept
Fig. 2 Results from SDOF simulation with trilinear elastic in physics.”
model shown in Fig. 1; simulation used sinusoidal excitation Table 1 offers motivating examples of four ways of
force u(t) shown at top and “event option” under MATLAB choosing state variables for two different basic (lin-
ode45 with RelTol = 10−3 and AbsTol = 10−12 . Event Types
#1 and 2 are defined in Eqs. (2) and (3), respectively
ear) dynamical systems: (a) a mass in series with a
Kelvin model (also called Kelvin–Voigt) and (b) a mass
in series with a Maxwell model. A Kelvin model has
more complicated models. More information concern- a spring and dashpot in parallel, whereas a Maxwell
ing the algorithm itself is given in Sect. 2.1. model has a spring and dashpot in series.
Often when modeling a dynamical system, numer-
ous choices of state variables are possible, although
1.2 Motivating example: choices of state variables many are not straightforward, nor even physically or
computationally convenient. The text and “Appendix A”
State-space representation has been the cornerstone provide additional discussion and examples, particu-
for both theoretical and numerical studies of nonlinear larly Sect. 2.2 (flow- and effort-controlled systems) and
dynamical systems. Loosely speaking, state variables Sect. 4 (Bouc–Wen model), where the first and second
for a dynamical system can be used to model the sys- columns in Table 1 will be elaborated in Sect. 2.2.
tem as a set of first-order ordinary differential equations
(ODE) called state equations:
1.3 Contributions and structure of this paper
ẏ = f (y, u, t) (4)
This paper builds on previous work [34] promoting the
where u is an input vector (includes initial conditions) use of hybrid dynamical systems for nonlinear hys-
and y is a state vector containing all state variables. teresis modeling. The utilization of a state event loca-
For example, displacement and velocity are commonly tion algorithm is validated here for various piecewise-

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

1168 J.-S. Pei et al.

Table 1 Possible choices of state variables for basic (linear) dynamical systems
Choice State variable (a) Flow-controlled: mass + kelvin (b) Effort-controlled: Mass + Maxwell
k
u(t) k c u(t)
m m
c
           
x 0 1 0 0 1 0 0 
1 (F) y1 = ẏ1 = y1 + u ẏ1 = y1 + u+ udt
ẋ − mk − mc 1
m − mk − kc 1
m
1
m
           
p 0 1 0 0  0 1 0 
2 (E) y2 = ẏ2 = y2 + c u + udt ẏ2 = y2 + k udt
r − mk − mc k
− mk − kc
     1 m m
   1 m
x 0 − m1  0 −1 
3 (M) y3 = ẏ3 = y3 + mc udt ẏ3 = y3 + m udt
p k − mc m k − kc 0
         1 
ẋ 0 − m1 1 0 −1
4 (M) y4 = ẏ4 = y3 + mc u ẏ4 = y4 + m u
r k − mc m k − kc 0

defined constitutive models. Furthermore, bond graph tion is done by computing a mesh of time points that
theory is introduced to directly benefit the implementa- are generated adaptively, normally not at a fixed time
tion of the state event location algorithm. More specif- step. Great care and significant effort were taken to cor-
ically, flow- versus effort-controlled system classifi- rect discontinuity sticking, which is a numerical prob-
cation is applied to systematically generate different lem caused by algebraic variable(s) not being updated
choices of state variables, which are further tested properly at a state event location [34].
through programming and quantitative measures of Following [34], the accuracy of our numerical solu-
computational efficiency. Even though the primary tions is assessed by studying the global error (GE) of
focus is on the state event location algorithm, this study the displacement at a specific time tn :
not only exercises the flow- versus effort-controlled
systems classification, but also provides theoretical dis-
GE(tn ) = |x̂(tn ) − x(tn )| (7)
cussion of them. For both simulation and classifica-
tion, there exist nonunique choices of state variables
for piecewise-defined models. where x(tn ) is the exact displacement at tn and x̂(tn ) is
Section 1, where a simple trilinear elastic model was the approximated displacement at that time. Since the
presented to illustrate state event location, serves as exact displacement is unknown (in most cases), a highly
an introduction. Section 2 discusses quantitative mea- converged numerical solution is obtained by using a
sures of computational performance (tolerance propor- very small value of the relative tolerance parameter
tionality, work–accuracy diagrams), flow- and effort- called RelTol in MATLAB. In order to control approx-
controlled systems, and the essential elements in hybrid imation errors and assess accuracy in our numerical
dynamical system modeling that are relevant to this work, we fix the value of AbsTol to a very small value
study. Sections 3 and 4 focus on a comprehensive study while allowing RelTol to take on a wide range of values
of bilinear hysteresis and Bouc–Wen models, respec- down to the fixed value of AbsTol.
tively. Section 5 is dedicated to guidelines on choices The algorithm behind the event option under ode45
of state variables. Section 6 concludes this study. is considered quite robust according to some stud-
ies [9,21] and reviewed in [34]. For example, con-
2 Theoretical issues tinuing our discussion of the trilinear elastic model in
Fig. 1, Fig. 3 shows results for a wide range of RelTol
2.1 Discontinuity sticking correction and performance values, with algorithmic robustness being evident for
measures for state event location algorithm RelTol < 10−3 (approx.) based on the identified num-
ber of Event #1 and #2 defined earlier and convergence
MATLAB’s ode45 uses a pair of Runge–Kutta formu- of the identified time of each event. The results in Fig. 2
las due to Dormand and Prince [8]. Numerical integra- correspond to RelTol = 10−3 .

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

On choosing state variables 1169

Number of All Events

-2 0.5

log10 of Elapsed Time

6 -4

log10 of GE
0
4 -6

state event -8 -0.5

2 Event Type #1
Event Type #2 -10
0 -1
-14 -12 -10 -8 -6 -4 -2 -12
log10 of RelTol -14 -1.5
Time of First Event Type #1 Time of Last Event Type #1 -15 -10 -5 0 -15 -10 -5 0
14.7335 42
log10 of RelTol log10 of GE
14.7330 4 4.5
40

log10 of Number of
14.7325

log10 of Number of

Unsuccessful FE
Successful FE
38 3.5 4
14.7320

14.7315 36 3 3.5
-15 -10 -5 0 -15 -10 -5 0
log10 of RelTol log10 of RelTol
2.5 3
Time of First Event Type #2 Time of Last Event Type #2
61.695 87.69
2 2.5
61.690 87.68 -15 -10 -5 0 -15 -10 -5 0
61.685 log10 of GE log10 of GE
87.67
61.680
87.66
61.675 Fig. 4 Tolerance proportionality (upper left) and work–accuracy
61.670 87.65 diagrams (the rest) for the trilinear elastic model in Fig. 1, where
-15 -10 -5 0 -15 -10 -5 0
log10 of RelTol log10 of RelTol GE and FE stand for global error and function evaluation, respec-
tively
Fig. 3 Number of state events and their event time values plotted
against MATLAB ode45 RelTol for the trilinear elastic model in
Fig. 1. Event Types #1 and 2 are defined in Eqs. (2) and (3), advanced systems, e.g., [24,25]). Following the format
respectively
of Fig. 1 in [29], Fig. 5 depicts Paynter’s tetrahedron
of state [22,23] for (a) electrical circuits, which can
Following our work in [34], tolerance proportion- be translated to mechanical systems using (b) force–
ality (TP) and work–accuracy diagrams are also pro- voltage analogy, and (c) force–current analogy.
vided here to demonstrate the computational stability For mechanical systems, a commonly chosen pair
and efficiency of state event location. TP measures of state variables is velocity (called flow in bond graph
the relationship between GE and RelTol for explicit, theory) and displacement (the time integral of flow), as
adaptive Runge–Kutta time-stepping methods, while in Eq. (5) or Choice No. 1 in Table 1. An alternative pair
work–accuracy diagram measures computational sta- of state variables is restoring force r (t) (called effort in
bility. For example, above the upper left plot in Fig. 4, bond graph theory) and corresponding “momentum” p
the TP slope is 0.965, obtained by a straight line fit to (e.g., [16]) where
the TP data. This slope is quite close to one, which
 t
indicates very good computational stability for this
p(t) = r (τ )dτ (8)
series of trilinear elastic model simulations. Also, the −∞
work–accuracy diagrams for elapsed time, successful
and unsuccessful function evaluations (FE) are all quite as in Choice No. 2 in Table 1. The name momentum
smooth, which again affirms good computational sta- for p(t) is in accord with [16]. Not to be confused with
bility. momentum m ẋ, p will be renamed as “g-momentum”
herein, where “g” is short for “generalized” follow-
ing [22]. For the autonomous case (u = 0) of Eq. (1),
2.2 Flow- versus effort-controlled systems Eq. (8) holds with p = −m ẋ if the system is initially
at rest but at a displaced position.
The state variable alternatives in Table 1 are obtained, Kelvin and Maxwell models, the two systems in
first, by applying bond graph theory, especially recent Table 1, form a dual in the sense that the connectivity
research on flow- and effort-controlled system classi- of elements in parallel and series corresponds, respec-
fication (e.g., the memristor, memcapacitor and their tively, to flow- and effort-controlled classification of

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

1170 J.-S. Pei et al.

.
v r x

dx = x dt
dp = r dt
dj = v dt
Resistor Capacitor Damper Spring Damper . Mass
. . .
dv = R di dq = C dv dr = c d x dx = 1/k dr dx = 1/c dr dp = m dx

. .
i dq = i dt q x dx = x dt x r dp = r dt p

Memristor Memristor Memristor

Inductor Mass Spring
dj (q) = M(q) dq . dp(x) = M(x) dx dx(p) = M(p) dp
dj = L di dp = m d x dx = 1/k d r
dq(j) = W(j) dj dx(p) = W(p) dp dp(x) = W(x) dx

j p x

(a) (b) (c)

Fig. 5 Planar depiction of Paynter’s tetrahedron of state for the four fundamental a electrical circuit elements following Fig. 1 in [29],
b mechanical system elements using force–voltage analogy and c mechanical system elements using force–current analogy

systems. In (a) flow-controlled system, a mass plus a Choice No. 4, which is used, for example, in the Bouc–
Kelvin model, the natural state variables are displace- Wen model [3], does not fit naturally into bound graph
ment x and velocity ẋ. These state variables should theory in the form given, and is a mixed (M) sys-
be solved (or calculated) first from force equilibrium, tem. However, the Bouc–Wen model can be written
which is typically expressed as an ODE involving these in effort-controlled form, consistent with bond graph
state variables. In contrast, in (b) effort-controlled sys- theory [24].
tem, a mass plus a Maxwell model, the natural state The motivating examples in Table 1, albeit being
variables are g-momentum p and restoring force r . linear, reveal that various possibilities do exist to select
These state variables should be solved (or calculated) state variables and their corresponding state equations.
first from deformation compatibility, which is often In particular, there are alternatives to displacement and
expressed as another ODE involving the state variables. velocity which are often used in practice.
These two choices, which are labeled as (a) and (b) in To complement Table 1, additional examples of
Table 1, are from the perspective of flow- and effort- nonunique state-space representations are given in
controlled classification, and are so indicated by (F) Tables 2 and 3 in “Appendix A.” For brevity, derivations
for all these expressions are not provided in this paper;
and (E), respectively.
however, key technical details start with manipulating
There is a well-established procedure under bond Eq. (1) as follows:
graph theory on how to assign and then (if possi-
ble) reduce state variables in writing up system equa-  
m = a constant
tions for numerical simulation (or even skipping system m ẍ = u − r ⇒ m ẍdt = (u − r )dt ⇒
 
equations) [16,22,23]. The key idea in bond graph the-
m ẋ(t) = m ẋ(0) + udt − r dt (9)
ory is to identify energy storage elements, masses and 
momentum  
springs, in a system, and then assign energy storage impulse g-momentum
variables, namely g-momenta and displacements for
masses and springs, respectively, before other details
are carried out. This means that the first choices for state Equation (9) indicates that p, a kinetic quantity,
variables would be the integral of the internal forces and and ẋ, a kinematic quantity, are related (principle
displacements, a mixture of effort- and flow-controlled of impulse–momentum) and so are not independent,
variables. Hence, this choice of state variables leads to a which has two implications:
so-called mixed (M) system. Returning to the examples
1. A state variable vector y does not need to have
in Table 1, the state variables would be g-momentum p
all four vertices in Paynter’s tetrahedron of state,
and displacement x, which is Choice No. 3 in Table 1.
x, ẋ, p, r , the fact of which can be verified in

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

On choosing state variables 1171

Tables 1, 2 and 3. To obtain state equations with are the governing equation. This state equation
full rank, one cannot have ẋ and p coexisting in is called a flow map. Algebraic equation(s) may
one state variable vector as long as Eq. (1) holds. be needed as well so the flow map consists of
2. All state equations in Table 1 can be made equiva- differential–algebraic equations (DAEs).
lent. We have the following pairs: 4. Event An event is when the mode makes a transition
from one status to another. It is also called an edge,
Choice #1 (F) ↔ #3 (M) by replacing ẋ with p
or a jump. It is normally a state event defined by
Choice #2 (E) ↔ #4 (M) by replacing p with ẋ
state variables—sometimes, algebraic variable(s),
Choice #3 (M) ↔ #4 (M) by adjusting the order
as well.
of state variables
5. Event Function An event function defines a crite-
With the equivalency of these pairs in place, it can be rion for an event’s occurrence. It is also called a
seen that Choices #1 (F) and #2 (E) can be made equiva- guard. It is defined by using state variables—and
lent, leading ultimately to the equivalency of all choices sometimes, algebraic variable(s), as well.
in Table 1, meaning that the solution of any choice is 6. Reset Map A reset map defines mathematically how
mathematically the same as any other. However, for the process of reset takes place right after an event.
nonlinear models, one choice of state variables may be In general, a reset map can include both state and
superior to another, based on physical reasoning, inter- algebraic variables.
pretation of experimental results, and even on numeri-
cal convenience (including accuracy and efficiency). These components are applied to both bilinear hys-
teresis and Bouc–Wen models in Sects. 3 and 4, respec-
tively.
2.3 Proposed components for hybrid systems

Hybrid dynamical systems, or hybrid systems in short, 3 Bilinear hysteresis model

deal with the interaction between continuous and dis-
crete time variables. Different research communities 3.1 Bilinear hysteresis model: analysis
approach hybrid systems from different angles and
with different agendas. In [30], a hybrid automaton is Bilinear hysteresis models (e.g., [6,7]) are popular for
defined formally by a septuple and includes the use modeling plastic deformation. The four modes, I to IV,
of differential–algebraic equations (DAEs). The def- defined in [15] are illustrated first by using a simulated
inition in [11], on the other hand, seems simpler and example in Fig. 6a to replace the cartoon in Fig. 1(a)
involves six components with differential inclusion uti- in [15], and also by a mode transition diagram extended
lized. We will focus on the following components for a slightly from Fig. 1(b) in [15]—with details substanti-
typical hybrid dynamical system by applying the defi- ated in Table 4 in “Appendix B.” These details closely
nition in [11] but with some amendments: follow the hybrid system components elaborated above
so that the hybrid system concept directly facilitates the
1. Mode A piecewise-defined r = r (x) model corre- understanding and programming. The four algebraic
sponds to a collection of discrete characteristic sta- equations given in Eq. (2) in [15] are rewritten in a
tuses that the model has, called modes. For exam- slightly more general form as follows:
ple, in the trilinear elastic example, one can say that
there are two modes when either k = 1 or αk = 0.3 Mode I: r = k(x − x0 ) (10)
applies. Normally, a transition diagram is used to
Mode II: r = αkx − (1 − α)r y (11)
illustrate the relationships among multiple modes.
2. Domain Within a mode, there are state variable(s) Mode III: r = k(x − x0 ) (12)
and perhaps algebraic variable(s). A domain is a Mode IV: r = αkx + (1 − α)r y (13)
region where the time evolution of the state vari-
able(s) and perhaps algebraic variables’ values are where k, r y and α are system properties similar to the
given by a specified mode. trilinear example in Fig. 1, and x0 = x0 (t) is a quan-
3. Flow map Within a mode, the state variables are tity that changes as the simulation proceeds, starting
continuous and smooth when the state equation(s) with x0 (0) = 0. It is the reset map that updates the

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

1172 J.-S. Pei et al.

Fig. 6 Bilinear hysteretic

Restoring force
model following a II
flow-controlled formulation. IV
a Illustration of hysteresis 2 4
loops using a simulated
result. b Mode transition 6
-x y III I
diagram; see Table 4 for  1
3 I III 7
to 8 x of Mode III Displacement
0
x of Mode I 1
0
II
-r
y
8 5

IV

(a) (b)

value of x0 through the memory parameters Hl , Hu called “partially effort-controlled.” The two formula-
and O in programming, details of which are given in tions have their own state variables, event functions,
“Appendix B.” flow map and reset map, but the modes and events are
As shown in Fig. 6a, Modes II and IV are the low and the same for the same applied force u(t). The partially
high envelope as defined in Eqs. (11) and (13), respec- effort-controlled scheme is simpler to implement and
tively. Modes I and III, however, cannot be grasped more efficient in performance.
comprehensively from Fig. 6a, which does not cover Regarding the mixed (M) formulation, the bilinear
all possible behaviors of these two modes. hysteresis model can be treated as a “Jenkin’s ele-
The comprehensive depiction of Modes I and III is ment” and a linear spring connected in parallel, whose
in Fig. 6b in conjunction with Table 4. Mode I can both response is illustrated in blue in Fig. 7a and decom-
stem from () 8 and return to () 1 Mode IV; however, posed in Fig. 7b. The linear spring is used to capture
Mode I can only jump to () 2 Mode II but Mode II the hardening effect (whose response is plotted in red
cannot jump to Mode I. When Mode I has an unfulfilled in Fig. 7b), while the Jenkin’s element is another linear
return to Mode IV, an internal loop (still on the same spring and a Coulomb damper connected in series, an
straight line) forms ().3 Conversely, Mode III can both effort-controlled device (whose response is plotted in
stem from () 4 and return to () 6 Mode II; however, black in Fig. 7b). Thus, the bilinear hysteresis model
Mode III can only jump to () 5 Mode IV but Mode can be implemented as a mixed device (flow- and effort-
IV cannot jump to Mode III. When Mode III has an controlled mixture), called partially effort-controlled.
unfulfilled return to Mode II, an internal loop (still on When the bilinear model is treated as a “Jenkin’s
the same straight line) forms (). 7 Therefore, Modes I element” and a linear spring (for hardening effect) con-
and III are not the same – even when Eqs. (10) and (12) nected in parallel, we have:
appear the same (where x0 is actually being reset).
It is tempting to think that these four piecewise- r = r1 + r2 = r1 + αkx (14)
defined formulas for r (t) could be readily embedded in
a flow-controlled setting with displacement and veloc- where r1 and r2 represent the contribution to the restor-
ity as state variables, such as Eqs. (5) and (6), and imple- ing force from the Jenkin’s element and linear spring,
mented as was done for the Ramberg–Osgood model respectively.
in [34]. However, bilinear hysteresis has more com-
plicated events, event functions (for MATLAB imple-
mentation) and reset map, leading to an implementa- 3.2 Bilinear hysteresis model: programming
tion similar in kind, but not in detail, to the Ramberg–
Osgood model. Due to this complexity, we examined Equations (5) and (6) are rewritten into the following:
several formulations and ultimately implemented two:
(1) a flow-controlled formulation (F) and (2) a mixed
flow- and effort-controlled formulation (M), herein

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

On choosing state variables 1173

Restoring force

Restoring force
II
IV
2 4
IV
6
-x y III I
III I 3 I III 7
x of Mode III Displacement
0
x of Mode I Displacement 1
0
II
-r y
II 8 5

IV

(a) (b) (c)

Fig. 7 Bilinear hysteretic model following a partially effort-controlled formulation. a Illustration of hysteresis loops using a simulated
result. b An equivalent sum for hysteresis loops. c Mode transition diagram; see Table 4 for  1 to 8

 
  where tag2 is the same mode indicator explained in
yF (1) x Table 5 with the differential equations corresponding
yF = = , (15)
yF (2) ẋ to Eqs. (10)–(13).
     
ẏF (1) ẋ yF (2) For bilinear hysteresis, three types of events and
ẏF = = = 1 (16)
ẏF (2) ẍ m (u − r )
event functions are needed, two of which involve dis-
placement x(t) while the third type is a so-called
velocity-turning point, meaning a time when the veloc-
where r = r (x, x0 ) and the subscript F is for flow- ity is zero:
controlled. To complete the flow map, the algebraic
equations corresponding to Eqs. (10)–(13) are given
Event Type #3 : ẋ(t) = 0 (19)
in Table 5 in “Appendix B,” using tag2 , a mode indi-
cator established in [34], and O, a programming nota-
tion for one of the memory parameters in this study The F and M events (including Event Types #1 and #2
(see “Appendix B”). that are not presented here), event functions and reset
For the mixed (M) formulation, the flow map is given maps are given in Table 4 in “Appendix B” covering all
by: allowed mode transitions, from I to II, I to IV, II to III, III
to IV, III to II and IV to I. The columns “direction” con-
tain the information on domain for programming. The
⎧ ⎫ ⎧ ⎫
⎪
⎪ yM (1) ⎪⎪ ⎪
⎪ x ⎪
⎪ F and M reset maps are more complicated than the reset
⎪ yM (2) ⎪
⎪ ⎪ ⎪ ⎪ ⎪
⎪
⎨ ⎬ ⎨ ẋ ⎬ map for the trilinear elastic model. In the mixed (M)
yM = yM (3) = p = p1 + p2 , (17) formulation, the switching values of tag2 are fewer and
⎪
⎪ ⎪ ⎪ ⎪
⎪
⎪ y (4) ⎪ ⎪ ⎪ ⎪ r1 ⎪
⎪ the reset map is significantly simpler (without involv-
⎩ M ⎪ ⎭ ⎪ ⎩ ⎪
⎭
yM (5) r2 ing updating the memory parameters Hl , Hu and O
⎧ ⎫ ⎧ ⎫ as indicated in † in Table 4), in comparison with the
⎪ ẏM (1) ⎪
⎪
⎪ ⎪ ⎪
⎪ ⎪ ẋ
⎪
⎪
⎪
⎪
⎪ ⎬ ⎪
⎨ ẏM (2) ⎪ ⎨ ẍ ⎪
⎬
flow-controlled (F) formulation.
ẏM = ẏM (3) = r = r1 + r2 For both the F and M formulations (flow-controlled
⎪
⎪ ⎪ ⎪
⎪ ⎪ ṙ1 ⎪
⎪ and partially effort-controlled) and as illustrated in
⎪ ⎪ ⎪
⎪ ẏM (4) ⎪ ⎪ ⎪
⎪
⎩ ⎭ ⎩ ⎭ Fig. 8, main.m runs the ode45 solver, where the option
ẏM (5) ṙ2
⎧ ⎫ of “options” is activated so that Events.m and Out-
⎪
⎪ yM (2) ⎪
⎪ putFcn.m are invoked. The three event functions are
⎪
⎪ ⎪
⎨ m1 (u(t) − yM (4) − yM (5)) ⎪
⎬ defined inside Events.m, while all outputs are recorded
= yM (4) + yM (5) (18) by using OutputFcn.m whenever there is a successful
⎪
⎪ ⎪
⎪
⎪ (1 − tag2 )(1 − α)kyM (2) ⎪ ⎪
⎪
⎩ ⎭ time step under ode45. For the F formulation, the right-
αkyM (2) hand side of Eq. (16) including the piecewise algebraic

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

1174 J.-S. Pei et al.

2a 2b 3 4a 4b 2a 4b
1 state event location 1 state event location
OutputFcn.m Events.m OutputFcn.m Events.m
1 2a state variables
1 2a state variables

“[t, x, te, xe] = ... 2b algebraic variable r

“[t, x, te, xe] = ...
ode45 (at myfun options)” 3 memory parameters ode45 (at myfun options)”
main.m main.m
4a tag 1
2a 4a 2a 4b
2a 2b 3 4a 4b 4b tag 2
2a 4b 4b tag 2

myfun.m 3 myfun.m

(a) (b)

Fig. 8 Illustrations of the interactions among all MATLAB m-files. a Flow-controlled. b Partially effort-controlled

equations given in Table 5 in “Appendix B” is computed Two numerical examples with their performance
in myfun.m. For the M formulation, the right-hand side measures are presented in Figs. 9, 10, 11 and 12. It
of Eq. (18) is computed in myfun.m. can be seen that the partially effort-controlled formula-
As an adaptive time-stepping scheme, ode45 in tion (M) reduces computational time, while other per-
MATLAB does not have fixed time steps. Control of formance measures are similar, showing computational
time stepping is done through specifying RelTol— robustness in both cases.
and—MaxStep and InitialStep via odeset. While the
former has been investigated using tolerance propor-
tionality (TP), the effect of the latter two become
essential when simulating more complicated transient 4 Bouc–Wen model
response of the bilinear hysteresis model, e.g., Fig. 22
in [15]. 4.1 Bouc–Wen model: analysis
For both formulations, the system parameters are
m, k, α, x y and r y (with r y = kx y ). Additional pro- The formula for the Bouc–Wen model is given as fol-
gramming details are given in “Appendix B” using the lows:
programming jargon and notation established in [34].
ṙ = A ẋ − γ |ẋ||r |n−1r − β ẋ|r |n = h(ẋ, r ) (20)

where A, γ , β and n are model parameters (see

3.3 Bilinear hysteresis model: comparison between F
Eq.(5.24) in [13], which is based on [32]). Note that
and M formulations
any elastic term, such as kx, added to r just adds k to
A in Eq. (20) and so can be absorbed in A without any
The choice of state variables is straightforward for the
loss of generality. Because the absolute-value expres-
flow-controlled formulation (F), although enabling the
sions |ẋ| and |r | in Eq. (20) induce discontinuities, they
event functions and reset maps under the existing MAT-
are treated as state events ẋ(t) = 0 and r (t) = 0 in this
LAB ode45 setting turns out to be nontrivial, as illus-
study. For comparison purposes, we also present simu-
trated in Fig. 8a.
lation results where these discontinuities are ignored—
For the flow-controlled formulation (F), there is
by omitting state event coding—leading to reduced
strong coupling among all four MATLAB m-files, mak-
accuracy and computational efficiency.
ing the programming somewhat challenging. Figure 8a Using Eq. (9) and the sign function (sgn), Eq. (20)
uses arrows to show the data flow. In contrast, the pro- can be rewritten as follows:
gramming of the partially effort-controlled formulation    
(M) is simpler than the flow-controlled formulation (F). ṙ = A − γ sgn(ẋr ) + β |r |n ẋ,
 t   t
There is no strong coupling among all four MATLAB 1
ẋ = u(τ )dτ − p + ẋ(0), where p = r (τ )dτ
m-files, making the programming straightforward. Fig- m 0 0
ure 8b uses arrows to show the data flow. (21)

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

On choosing state variables 1175

Time history of prescribed excitation force Time history of prescribed excitation force
1 4

Excitation Force
Excitation Force

Prescribed excitation time history Prescribed excitation time history

0.5 2

0 0

-0.5 -2

-1 -4
0 10 20 30 40 50 60 70 80 90 100 0 50 100 150 200 250 300
Time Time
Time history of solved displacement Time history of solved displacement
2 10

Displacement
Displacement

1 5

0 0
Time selected by MATLAB ode45
Event Type #1 solved by MATLAB ode45
-1 Event Type #2 solved by MATLAB ode45 -5
Event Type #3 solved by MATLAB ode45
-2 -10
0 10 20 30 40 50 60 70 80 90 100 0 50 100 150 200 250 300
Time Time
Time history of solved velocity Time history of solved velocity
0.4 1.5
Time selected by MATLAB ode45
0.2 1 Event Type #1 solved by MATLAB ode45
Event Type #2 solved by MATLAB ode45

Velocity
0.5
Velocity

Event Type #3 solved by MATLAB ode45

0
0
-0.2 -0.5
-0.4 -1
0 10 20 30 40 50 60 70 80 90 100 0 50 100 150 200 250 300
Time Time
Time history of solved restoring force Time history of solved restoring force
2 4
Restoring force
Restoring force

1 2

0 0

-1 -2

-2 -4
0 10 20 30 40 50 60 70 80 90 100 0 50 100 150 200 250 300
Time Time

(a) (b)

Fig. 9 Bilinear hysteresis model with k = 1, α = 0.3 and (F) are presented here; the partially effort-controlled results
r y = 0.75 subject to different excitations and simulated using (M) are the same. Event Types #1 to 3 are defined in Table 4
the “event option” under MATLAB ode45, where RelTol = 10−3 in “Appendix B.” a One cycle of sinusoidal excitation. b An
and AbsTol = 10−12 : time histories. The flow-controlled results amplitude-modulated sine excitation

Thus, ṙ is now expressed as a function of p and r . Writ- where F+ (r ) and F− (r ) are the integrals with zero ini-
ten this way, it is evident that the Bouc–Wen element tial conditions. A qualitative discussion of the pair of
can be viewed as an effort-controlled device. Eqs. (22) and (23) is given as follows:
Following [25], we have the following pair of equa-
tions:
– A load versus displacement (r vs. x) plot of the
Bouc–Wen model can have a piecewise definition
 following the sign of ẋ. The stiffness is entirely a
1 function of r . This implies first the following fact:
x= dr + x+ (0)
A − γ |r |n−1r − β|r |n At any point on the loading path (x0 , r0 ), the stiff-
 F+ (r ) + x+ (0), when ẋ > 0 (22) ness is always calculated by either Eqs. (22) or (23)
 depending on the sign of ẋ. This means that there
1
x= dr + x− (0) are only two stiffness values for (x0 , r0 ): one for all
A + γ |r | r − β|r |n
n−1
loading paths and the other for all unloading paths.
 F− (r ) + x− (0), when ẋ < 0 (23) This also implies that parallel loading or unload-

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

1176 J.-S. Pei et al.

Solved hysteretic loops Solved hysteretic loops

3
1.5

1 2

Restoring force
Restoring force

0.5 1

0 0

-0.5 -1

-1 -2 Time selected by MATLAB ode45

Time selected by MATLAB ode45
Event Type #1 solved by MATLAB ode45 Event Type #1 solved by MATLAB ode45
Event Type #2 solved by MATLAB ode45 Event Type #2 solved by MATLAB ode45
Event Type #3 solved by MATLAB ode45 Event Type #3 solved by MATLAB ode45

-1.5 -3
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 -10 -8 -6 -4 -2 0 2 4 6 8
Displacement Displacement

(a) (b)

Fig. 10 Bilinear hysteresis model with k = 1, α = 0.3 and results (F) are presented here; the partially effort-controlled
r y = 0.75 subject to different excitations and simulated using the results (M) are the same. Event Types #1 to 3 are defined in
“event option” under MATLAB ode45, where RelTol = 10−3 Table 4 in “Appendix B.” a One cycle of sinusoidal excitation.
and AbsTol = 10−12 : hysteresis loops. The flow-controlled b An amplitude-modulated sine excitation
log10 of Number of Successful FE

log10 of Number of Successful FE

-2 0.5 3.6 -2 1 4.5
log10 of Elapsed Time

log10 of Elapsed Time

-4 -4
3.4
0.5 4
log10 of GE

log10 of GE

-6 0 -6
3.2
-8
-8 0 3.5
-10 3
-0.5 -10
-12 2.8 -0.5 3
-12
-14
-1 2.6 -14 -1 2.5
-15 -10 -5 0 -15 -10 -5 -15 -10 -5 -15 -10 -5 0 -14 -12 -10 -8 -6 -4 -2 -14 -12 -10 -8 -6 -4 -2
log10 of RelTol log10 of GE log10 of GE log10 of RelTol log10 of GE log10 of GE
flow-controlled (F) flow-controlled (F)
partially effort-controlled (M) partially effort-controlled (M)

(a) (b)
Fig. 11 Bilinear hysteresis model with k = 1, α = 0.3 and partially effort-controlled (M) simulations are compared here,
r y = 0.75 subject to different excitations and simulated using where GE and FE stand for global error and function evalua-
the “event option” under MATLAB ode45, where RelTol = 10−3 tion, respectively. a One cycle of sinusoidal excitation. b An
and AbsTol = 10−12 : tolerance proportionality (TP) and work– amplitude-modulated sine excitation
accuracy diagrams. The results from flow-controlled (F) and

ing paths are within the same range of r but for – There are no viscous damping terms, linear or non-
different ranges of x. linear. Having said this, the Bouc–Wen model is
– The original Bouc–Wen formula has now been rate-independent, which is often used in conjunc-
rewritten into a piecewise expression for the stiff- tion with viscous damping, i.e., to combine with
ness plus ẋ = 0 ⇒ ṙ = 0. This means that, rate-dependent components. Rate independency of
essentially, the Bouc–Wen model represents a non- the Bouc–Wen model is consistent with what is
linear spring whose stiffness is piecewise-defined claimed in [20].
with the velocity being the switching mechanism.

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

On choosing state variables 1177

50 50
Number of All Events

Number of All Events

state event state event
40 40
Event Type #1 Event Type #1
Event Type #2 Event Type #2
30 Event Type #3 30 Event Type #3

20 20

10 10

0 0
-14 -12 -10 -8 -6 -4 -2 -14 -12 -10 -8 -6 -4 -2
log10 of RelTol log10 of RelTol
Time of First Event Type #1 Time of Last Event Type #1 Time of First Event #1 Time of Last Event #1
111.390 290.769 111.390 290.769

290.768
290.768
111.385 111.385
290.767
290.767
290.766
111.380 111.380
290.766
290.765

111.375 290.764 111.375 290.765

-15 -10 -5 0 -15 -10 -5 0 -15 -10 -5 0 -15 -10 -5 0
log10 of RelTol log10 of RelTol log10 of RelTol log10 of RelTol
Time of First Event Type #2 Time of Last Event Type #2 Time of First Event #2 Time of Last Event #2
74.85 269.458 74.86 269.458

269.456 74.85 269.456

74.84
269.454 74.84 269.454
74.83
269.452 74.83 269.452
74.82
269.450 74.82 269.450

74.81 269.448 74.81 269.448

-15 -10 -5 0 -15 -10 -5 0 -15 -10 -5 0 -15 -10 -5 0
log10 of RelTol log10 of RelTol log10 of RelTol log10 of RelTol
Time of First Event Type #3 Time of Last Event Type #3 Time of First Event #3 Time of Last Event #3
31.96 287.196 31.96 287.196

287.194 31.95 287.194

31.95
287.192 31.94 287.192
31.94
287.190 31.93 287.190
31.93
287.188 31.92 287.188

31.92 287.186 31.91 287.186

-15 -10 -5 0 -15 -10 -5 0 -15 -10 -5 0 -15 -10 -5 0
log10 of RelTol log10 of RelTol log10 of RelTol log10 of RelTol

(a) (b)

Fig. 12 Bilinear hysteresis model with k = 1, α = 0.3 and time values plotted against MATLAB ode45 RelTol. Event Types
r y = 0.75 subject to the amplitude-modulated sine excitation #1 to 3 are defined in Table 4 in “Appendix B.” a Flow-controlled.
and simulated using the “event option” under MATLAB ode45, b Partially effort-controlled
where AbsTol = 10−12 : number of state events and their event

4.2 Bouc–Wen model: programming

ẋ > 0, r > 0
ẋ
One way of programming the Bouc–Wen model with
Event Type #1
no state events considered at all is as follows, a version Mode 1
Mode 2
of which is denoted by “A” herein:
Event Type #3
Event Type #3

⎧ ⎫ ⎧ ⎫
⎨ yA (1) ⎬ ⎨ x ⎬
y A = yA (2) = ẋ , r
⎩ ⎭ ⎩ ⎭
yA (3) r
⎧ ⎫ ⎧ ⎫
⎨ ẏA (1) ⎬ ⎨ ẋ ⎬
ẏA = ẏA (2) = ẍ
⎩ ⎭ ⎩ ⎭ Mode 1 Mode 2
ẏA (3) ṙ
⎧ ⎫
⎨ yA (2) ⎬ Event Type #2
= m1 (u(t) − yA (3))
⎩ ⎭ ẋ < 0, r < 0
A yA (2) − γ |yA (2)||yA (3)|n−1 yA (3) − β yA (2)|yA (3)|n
(24)
Fig. 13 Modes and domains for the Bouc–Wen model to be
When considering state events, the modes and associated with Eq. (26)
domain are illustrated in Fig. 13:

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

1178 J.-S. Pei et al.
⎧ ⎫ ⎧ ⎫
⎪
⎨ yC (1) ⎪ ⎨x⎪
⎬ ⎪ ⎬
Equation (21) indicates that only p and r are needed
yC = yC (2) = ẋ ,
as state variables. Assuming y(1) = p and y(2) = ⎩ y (3) ⎭ ⎩ r ⎪
⎪ ⎪ ⎪ ⎭
C
r , the state event associated with r can then be taken ⎧ ⎫ ⎧ ⎫
⎪
⎨ ẏC (1) ⎪ ⎨ ẋ ⎪
⎬ ⎪ ⎬
care of easily, whereas the state event associated with
ẏC = ẏC (2) = ẍ
ẋ would meet a challenge by utilizing Eq. (9): ⎪
⎩ ẏ (3) ⎪ ⎩ ṙ ⎪
⎭ ⎪ ⎭
C
⎧ ⎫
⎪
⎨ yC (2) ⎪
⎬
= m1 (u(t) − yC (3))
⎪
⎩ A y (2) − γ tag y (2)|y (3)|n−1 y (3) − β y (2)|y (3)|n ⎪
⎭
  C 2 C C C C C
1
ẋ = 0 ⇒ (u − r ) dt = 0 ⇒ y(1) = udt (27)
m
with an event function as follows:

yC (2) = 0 (28)
This, however, goes beyond the allowed forms of event
functions in MATLAB’s state event location algorithm
when yC (2) is either ascending or descending, belong-
according to [27]. Having said this, ẋ will be included
ing to the third type of state event, and tag2 = − tag2 .
as a state variable instead. The last state variable x
is simply added to obtain the displacement. The pro-
posed revision is given as follows, an option of which 4.3 Bouc–Wen model: results and comparisons
is denoted as “B” herein.
The flow map under Option B is given as follows: Figures 14 and 15 present a set of simulation results
under Option B. The same amplitude-modulated sine
excitation was used. All three types of state events were
⎧ ⎫ ⎧ ⎫ ⎧ ⎫ ⎧ ⎫
⎨ yB (1) ⎬ ⎨ x ⎬ ⎨ ẏB (1) ⎬ ⎨ ẋ ⎬ detected.
yB = yB (2) = ẋ , ẏB = ẏB (2) = ẍ A comparison is made among three options: A. no
⎩ ⎭ ⎩ ⎭ ⎩ ⎭ ⎩ ⎭
yB (3) r ẏB (3) ṙ state events considered at all, B. all three types of
⎧ ⎫ state events considered, and C. only one out of three
⎪
⎪ yB (2) ⎪
⎪
⎨ 1 ⎬ types of state events considered; the results are pre-
(u(t) − yB (3))
= m
⎪
⎪ A yB (2) − γ tag2 yB (2)(tag3 yB (3)) n−1 yB (3) ⎪
⎪ sented in Fig. 16 using tolerance proportionality and
⎩ ⎭
−β yB (2)(tag3 yB (3))n work–accuracy diagrams. It can be seen that, first,
(25) the utilization of state event location algorithm does
improve computational accuracy and reduce compu-
tational effort measured by both number of function
evaluation and computational time—when compared
with the events, event functions and reset map given as
to omitting the state event location algorithm. This is
follows:
the very motivation of our work. It can be seen that,

⎧
⎨ Event Type #1 : yB (3) = 0 when yB (3) is ascending tag3 = 1
Event Type #2 : yB (3) = 0 when yB (3) is descending tag3 = − 1 (26)
⎩
Event Type #3 : yB (2) = 0 when yB (2) is either ascending or descending tag2 = − tag2

One drawback in automation is to assign initial val- next, when not all types of state events are considered,
ues to tag2 and tag3 . Note that tag1 is used for correcting the TP and work–accuracy diagrams may not change
discontinuity sticking, which will not be used for state obviously from those—except with small GE values—
variables. when all types of state events are correctly considered.
To achieve a better understanding of the importance This indicates the importance of having a good under-
of all state events, we intentionally ignore the first two standing and doing a thorough job when implementing
restoring force events. When we do so, we have the state event location algorithm because mistakes made
following option denoted as “C”: may not be easy to detect.

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

On choosing state variables 1179

Time history of prescribed excitation force Solved hysteretic loops

Excitation Force
4 1
Prescribed excitation time history
2
0.8
0
0.6
-2 Time selected by MATLAB ode45
Event Type #1 solved by MATLAB ode45
0.4
-4 Event Type #2 solved by MATLAB ode45

Restoring force
Event Type #3 solved by MATLAB ode45
0 50 100 150 200 250 300
0.2
Time
Time history of solved displacement 0
1000
Displacement

Time selected by MATLAB ode45

Event Type #1 solved by MATLAB ode45 -0.2
500 Event Type #2 solved by MATLAB ode45
Event Type #3 solved by MATLAB ode45
-0.4
0
-0.6
-500
0 50 100 150 200 250 300 -0.8
Time
-1
Time history of solved velocity -400 -200 0 200 400 600 800
40
Displacement
20
Velocity

0
-20
Fig. 15 Bouc–Wen (BW) model with A = 0.5, γ = 0.25,
-40 β = 0.25 and n = 1 subject to the amplitude-modulated sine
-60 excitation and simulated under Option B using the “event option”
0 50 100 150 200 250 300 under MATLAB ode45, where AbsTol = 10−12 and RelTol =
Time 10−3 : hysteresis loops. Event Types #1 to 3 are defined in Eq. (26)
Time history of solved restoring force
Restoring force

1
0.5
0
  
-0.5 ẋ − ν β|ẋ||z|n−1 z + γ ẋ|z|n
ż = h(z) (30)
-1
0 50 100 150 200 250 300 η
Time

Fig. 14 Bouc–Wen (BW) model with A = 0.5, γ = 0.25,

β = 0.25 and n = 1 subject to identical amplitude-modulated
where the quantities h, η and ν are time varying and
sine and simulated under Option B using the “event option” under defined further using equations to be incorporated later
MATLAB ode45, where AbsTol = 10−12 and RelTol = 10−3 : on, while β, n and γ are model parameters with fixed
time histories. Event Types #1 to 3 are defined in Eq. (26) values. It should be noted that, for historical reasons,
the β and γ parameters in this subsection, starting at
Eq. (30), are reversed from those of the Bouc–Wen
4.4 Bouc–Wen–Baber–Noori Model model in Sect. 4.1, starting at Eq. (20).
Connecting a BWBN model to a mass to form a
The Bouc–Wen–Baber–Noori (BWBN) model has SDOF system, an elegant choice is to formulate the
been popular. There are slightly different versions used equations as differential–algebraic equations (DAEs)
and here we follow that in [10], where the model is specified below. Equation (31) is a state equation, while
considered to be a parallel connection between a linear Eq. (32) is a coupled algebraic equation—both are in
spring with stiffness αk and a hysteresis model. The vector form.
total restoring force is thus a sum of the linear spring
restoring force αkx and the nonlinear restoring force ⎧ ⎫ ⎧ ⎫ ⎧ ⎫ ⎧ ⎫
⎪
⎪ y(1) ⎪
⎪ ⎪x ⎪ ⎪ ẏ(1) ⎪ ⎪ ẋ ⎪
of the hysteresis model, the latter of which is scaled by ⎨ ⎬ ⎪ ⎨ ⎪ ⎬ ⎪
⎨ ⎪
⎬ ⎪ ⎨ ⎪ ⎬
y(2) ẋ ẏ(2) ẍ
using the coefficient (1 − α)k, as shown in Eq. (29). y=
⎪
= , ẏ = =
⎪ y(3) ⎪
⎩ ⎪ ⎪
⎭ ⎪z ⎪
⎩ ⎪
⎭
⎪
⎪
⎩
ẏ(3) ⎪
⎪
⎭ ⎪
⎪ ż ⎪
⎩ ⎪ ⎭
y(4) ε ẏ(4) ε̇
⎧ ⎫
r = αkx + (1 − α)kz (29) ⎪
⎪
y(2) ⎪
⎪
⎪
⎪ m1 (u(t) − w(1)) ⎪
⎪
⎨ ⎬
w(7)  
= w(2) y(2) − w(3) β tag2 y(2)(tag 3 y(3))
n−1 y(3)
⎪
⎪ ⎪
⎪
where α and k are model parameters with fixed values. ⎪
⎪
⎩ +γ y(2)(tag3 y(3)) n ⎪
⎪
⎭
z is a state variable having the same physical unit as the (1 − α)ky(2)y(3)
displacement: (31)

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

1180 J.-S. Pei et al.

log10 of Number of Successful FE

Fig. 16 Bouc–Wen (BW)
model subject to identical (a) (b) (c)
5 0.5 4

log10 of Elapsed Time

amplitude-modulated sine
excitation and simulated

log10 of GE
under Options A–C in terms 0 0 3.5
of tolerance proportionality
(a) and work–accuracy -5
diagrams (b, c), where GE -0.5 3
and FE stand for global
error and function -10
evaluation, respectively. -1 2.5
-15 -10 -5 0 -10 -5 0 5 -10 -5 0 5
Only time histories under
log10 of RelTol log10 of GE log10 of GE
Option B are presented in
Fig. 14 Option A without events
Option B with Event Types #1-3
Option C with Event Type #3 only

⎧ ⎫
⎧ ⎫ ⎧ ⎫ ⎪ αkx + (1 − α)kz ⎪
w(1) ⎪
⎪ ⎪
⎪
[t,y,te,ye] = ode45(@myfun,[tstart,tfinal],y0,
⎪
⎪ ⎪
⎪ ⎪
⎪ r ⎪
⎪ ⎪
⎪ 1 + δη ε ⎪
⎪
⎪
⎪ w(2) ⎪
⎪ ⎪
⎪ η ⎪
⎪ ⎪
⎪ ⎪
⎪
options,w)
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ 1 + δ ν ε ⎪
⎪
⎪
⎪ ⎪ ⎪
⎪ ⎪ ⎪
⎪ ⎪
⎪   ⎪
⎪
[value, isterminal, direction] = Events(t,y,w)
⎨ w(3) ⎬ ⎨ ν ⎬ ⎨ 1
1
n ⎬ ydot = myfun(t,y,w)
w = w(4) = z u = ν(β+γ
 )  status = OutputFcn(t,y,flag,w)
⎪
⎪ ⎪ ⎪ ⎪ ⎪ ⎪
⎪ w(5) ⎪
⎪ ⎪ ⎪
⎪ ⎪
⎪ ζ1 ⎪⎪
⎪
⎪
⎪
⎪ ζ10 1 − e− pε ⎪
⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ w(6) ⎪
⎪ ⎪
⎪ ζ 2⎪⎪ ⎪
⎪ ( 0 + δ ε) (λ + ζ )
1 ⎪⎪ Figures 17 and 18 present a set of simulation results
⎩ ⎭ ⎩ ⎭ ⎪ ⎪ 2 ⎪⎪
w(7) h ⎪
⎩
− [ zsgn( ẋ)−qz u ] ⎪
⎭ using the system parameter values in Fig. 10(b) of [10]
ζ22
1 − ζ1 e
⎧ ⎫ except for k. Their performance measures are presented
⎪ αky(1) + (1 − α)ky(3)
⎪
⎪
⎪
⎪
⎪ in Figs. 19 and 20. Defining MaxStep becomes essen-
⎪
⎪ 1 + δη y(4) ⎪
⎪
⎪
⎪ 1 + δ y(4) ⎪
⎪ tial in this numerical example.
⎪
⎪ ν ⎪
⎪
⎪
⎪ ⎪
⎪
⎨  1
n
⎬
1
= w(3)(β+γ ) (32)
⎪
⎪  − py(4)
 ⎪
⎪
⎪
⎪ ζ10 1 − e ⎪
⎪
⎪
⎪ ⎪
⎪ 5 Discussion
⎪
⎪ ( + δ y(4)) (λ + w(5)) ⎪
⎪
⎪
⎪
0 ⎪
⎪
⎪
⎩ −[ y(3)tag2 −qw(4)] 2
⎪
⎭
1 − w(5)e w(6) 2
5.1 Guidelines
where ε stands
t for hysteretic energy, defined as ε(t) =
Some guidelines concerning choices of state variables
(1 − α)k 0 z(τ )ẋ(τ )dτ ; r stands for restoring force; η
are summarized here and will be further validated in
and ν are time-varying stiffness and strength degrada-
future work:
tion parameters, respectively, and h is a pinching func-
tion, which requires time-varying quantities z u , ζ1 and 1. To apply Paynter’s tetrahedron of state using
ζ2 to define it. Overall, α, k, β, n, γ , δη , δν , q, ζ10 , p, energy/power variables, we have the following two
0 , δ , λ are system parameters. chains of kinematic and kinetic variables aligned
In terms of implementing this model under MAT- with each other:

LAB ode45 solver, all codes for the BW model under dt
Option “B” can be adopted as is, except for two updates: acceleration, ẍ −→ velocity,
 
First, Eq. (31) is keyed into myfun.m after Eq. (32). dt dt
ẋ −→ displacement, x −→ absement, a (33)
Discontinuity sticking is corrected as verified. Next,
following the recommendation in MATLAB, it is of rate of change in restoring force,
 
critical importance to put the algebraic variable w at dt dt
ṙ −→ restoring force, r −→ g-momentum, p
the end of the argument lists of the four m-files as 
dt
follows: −→ time integral of g-momentum, ρ (34)

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

On choosing state variables 1181

Time history of prescribed excitation force Solved hysteretic loops

4
Excitation Force

1
Prescribed excitation time history Time selected by MATLAB ode45
2 Event Type #1 solved by MATLAB ode45
0.8 Event Type #2 solved by MATLAB ode45
Event Type #3 solved by MATLAB ode45
0
0.6
-2
0.4
-4

Restoring force
0 50 100 150 200 250 300
0.2
Time
Time history of solved displacement 0
20
Displacement

10 -0.2
0
Time selected by MATLAB ode45
-0.4
-10 Event Type #1 solved by MATLAB ode45
Event Type #2 solved by MATLAB ode45
-20 Event Type #3 solved by MATLAB ode45 -0.6
-30
0 50 100 150 200 250 300 -0.8
Time
-1
Time history of solved velocity -30 -25 -20 -15 -10 -5 0 5 10 15 20
4
Displacement
2
Velocity

0 Fig. 18 Bouc–Wen–Baber–Noori (BWBN) models with α =

-2 0.1, k = 1, β = 1.5, n = 1, γ = −0.5, δη = 0.05, δν = 0.005,
q = 0.1, ζ10 = 0.97, p = 1, 0 = 0.2, δ = 0.002 and λ = 0.1
-4
0 50 100 150 200 250 300 subject to an amplitude-modulated sine and simulated using the
Time “event option” under MATLAB ode45, where AbsTol = 10−12 ,
Time history of solved restoring force RelTol = 10−3 and MaxStep = 10−1 : hysteresis loops. Event
1
Restoring force

Types #1 to 3 are defined in Eq. (26)

0.5
0
-0.5 in modeling should be included in the state vari-
-1
0 50 100 150 200 250 300
able derivative vector.
Time – When we utilize the state event location algo-
rithm, we need to make the variable that defines
Fig. 17 Bouc–Wen–Baber–Noori (BWBN) models with α = the event function a state variable—not a time
0.1, k = 1, β = 1.5, n = 1, γ = −0.5, δη = 0.05, δν = 0.005,
q = 0.1, ζ10 = 0.97, p = 1, 0 = 0.2, δ = 0.002 and λ = 0.1 derivative of a state variable. This means that
subject to an amplitude-modulated sine and simulated using the we cannot utilize the variable with the highest
“event option” under MATLAB ode45, where AbsTol = 10−12 , order in the system dynamics to define an event
RelTol = 10−3 and MaxStep = 10−1 : time histories. Event function.
Types #1 to 3 are defined in Eq. (26)
– Whenever ẋ is a state variable, then p does not
need to be a state variable, and vice versa. This
where absement a and time integral of g-momentum comes from Eq. (9) which shows that ẋ and p
ρ are not used in this study. Subsequently, the equa- are not independent state variables, based on
tion of motion can be integrated step-by-step as fol- Newton’s second law in Eq. (1).
lows:

          
... dt dt dt dt
m x = u̇ − ṙ −→ m ẍ = u − r −→ m ẋ = udt − p −→ mx = udtdt − ρ −→ ma = udtdtdt − ρdt (35)

We make the following additional comments:

2. It is important to reiterate the need to effectively
– The variable with the lowest order time deriva- address discontinuities in a model, whenever appli-
tive in the system dynamics that we care about cable. Velocity ẋ is often required for the state event
in modeling should be included in the state vari- location for piecewise-defined loading and unload-
able vector. ing branches in the r −x plane. Given Eq. (9), when-
– The variable with the highest order time deriva- ever there is only one mass connected in series with
tive in the system dynamics that we care about the rest of the assembly, ẋ and p cannot coexist in

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

1182 J.-S. Pei et al.
Number of All Events

log10 of Elapsed Time

-6 0.7
30
-7 0.6

log10 of GE
state event
Event Type #1
20 Event Type #2 -8 0.5
Event Type #3
-9 0.4
10
-10 0.3
0 -11 0.2
-14 -12 -10 -8 -6 -4 -2
-12 0.1
log10 of RelTol -15 -10 -5 0 -12 -10 -8 -6

Variation of First Event Type #1 Variation of Last Event Type #1

log10 of RelTol log10 of GE
Time = 98.9070 Time = 286.2068 4
-9 -8 4.7
× 10 × 10

log10 of Number of
log10 of Number of
2 5

Unsuccessful FE
3.9 4.6

Successful FE
0
0 3.8 4.5
-2
-5 3.7 4.4
-4
3.6 4.3
-6 -10
-15 -10 -5 0 -15 -10 -5 0 3.5 4.2
log10 of RelTol log10 of RelTol 3.4 4.1
-12 -10 -8 -6 -12 -10 -8 -6
Variation of First Event Type #2 Variation of Last Event Type #2 log10 of GE log10 of GE
Time = 49.7760 Time = 241.1848

× 10 -9 × 10 -8
10 10 Fig. 20 Tolerance proportionality (upper left) and work–
accuracy diagrams (the rest) for the BWBN model in Fig. 18,
5 5 where GE and FE stand for global error and function evaluation,
respectively
0 0
-15 -10 -5 0 -15 -10 -5 0
log10 of RelTol log10 of RelTol
force-state mapping alone due to the mismatch of
Variation of First Event Type #3 Variation of Last Event Type #3

-8
Time = 32.7039
-8
Time = 280.9137 the nature of the inputs and outputs.
3 × 10 6
× 10
4. Conversions of state variables can be carried out as
2 4
follows:
1 2
0 0     d    
x Eq. (9) x ẋ Eq. (9) p
−→ −→
dt
-1
-15 -10 -5 0
-2
-15 -10 -5 0 −→ (36)
ẋ p r r
log10 of RelTol log10 of RelTol

Fig. 19 Number of state events and the event time errors plotted implies that if ẋ and r can serve as state variables,
against MATLAB ode45 RelTol for the BWBN model in Fig. 18, then the flow-controlled system can be considered
where there are three, three and 24 Types #1, 2 and 3 events effort-controlled. To do so, as indicated in the equa-
throughout, respectively, to make a total of 30 events throughout.
Event Types #1 to 3 are defined in Eq. (26)
tion above, the rate of change seems essential for
this system, given the use of time differentiation for
state variables.
the state variable vector for the state equations to
        
be independent, so p will be replaced with ẋ and p Eq. (9) ẋ dt x Eq. (9) x
−→ −→ −→ (37)
thus will not appear in the state variable vector in r r p ẋ
models with discontinuity. This explains, to some
extent, why p is not popular for the choice of state implies that if x and p can serve as state variables,
variables. then the effort-controlled system can be considered
3. The force-state mapping comes in handy for flow- flow-controlled. To do so, as indicated in the equa-
controlled systems to map some kinematic quan- tion above, a memory effect seems essential for this
tities to some kinetic ones. When there is a need system, given the use of time integral for state vari-
for the inverse map (kinetic quantities to kinematic ables.
ones) as in effort-controlled systems, it would not 5. If an algebraic variable is needed to define an event
be rational if the force-state mapping technique is function, then the algebraic variable may be imple-
still adopted. Similarly, when dealing with mixed mented via a global variable under MATLAB. All
systems, mapping kinematic and kinetic quantities algebraic variables should be made into an alge-
into each other, it may not be proper to use the braic vector that is carried over through all MAT-

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

On choosing state variables 1183

LAB m-files as demonstrated in Sect. 4.4 for the Time history of prescribed excitation force
1

Excitation Force
BWBN model. 0.5
Prescribed excitation time history

0
-0.5
-1
5.2 The ṙ = h(ẋ, r ) formulation 0 10 20 30 40 50 60 70 80 90 100
Time
Time history of solved displacement
First, some review of the history is given here: [1,2] 2

Displacement
Time selected by MATLAB ode45
especially [3] using the Bouc–Wen and a “slip-lock” 1 Event Type #1 solved by MATLAB ode45
Event Type #2 solved by MATLAB ode45

model to capture pinching hysteresis are frequently 0

cited for the origin of this formulation. In fact, the -1

theoretical justification cited in [3] is [5], which actu- -2

0 10 20 30 40 50 60 70 80 90 100
ally gives a more general formulation making ṙ a Time
function of x, ẋ, r and a dynamic internal variable Time history of solved velocity
0.4
vector used to capture the path dependency of the 0.2
model.

Velocity
0
One major application of ṙ = h(ẋ, r ) is the Bouc– -0.2
Wen model. A review paper [17] on system identifica- -0.4
0 10 20 30 40 50 60 70 80 90 100
tion for nonlinear systems introduces this formulation Time
for the Bouc–Wen model. In this study, we showed that Time history of solved restoring force
2
the Bouc–Wen model is an effort-controlled model with
Restoring force

1
p and r as its state variables, but it has two types of dis-
0
continuities, leading to selecting ẋ as an additional state
-1
variable. Given the dependence between ẋ and p shown
-2
in Eq. (9), the necessary state variables for the Bouc– 0 10 20 30 40 50 60 70 80 90 100
Time
Wen model then become ẋ and r . Consequently, the
governing state equation is in the form of ṙ = h(ẋ, r ). Fig. 21 To reveal the insight to the peculiar point in Fig. 3 and
This is one rationale behind the ṙ = h(ẋ, r ) formu- as a counterpart for Fig. 2; simulation used sinusoidal excita-
lation of the Bouc–Wen model. Another explanation tion force u(t) shown at top and “event option” under MATLAB
ode45 with RelTol = 0.00077426 and AbsTol = 10−12 . Event
follows from bond graph theory, which says to use x
Types #1 and 2 are defined in Eqs. (2) and (3), respectively
and p. Instead, we differentiate them once and use ẋ
and r as state variables.
A more general application of ṙ = h(ẋ, r ) may be 5.3 Complications with state event location algorithm
deduced as follows using the pair of flow- and effort-
controlled formulations: Examination of Fig. 3 shows that a particular value
of RelTol caused the number of state events to reduce
by two. For this RelTol value (0.00077426), the time
Force-state mapping r = f (x, ẋ) (38)
histories are presented in Fig. 21 showing the loss
Formulation under discussion ṙ = h (ẋ, r ) (39) of two Type #1 events. One remedy is to introduce
MaxStep = 0.1. As Fig. 3 shows, this loss of two
     events did not happen with other selected values of
ẋ x RelTol, larger or smaller. To be studied in future work,
=f , t , flow-controlled (40)
ẍ ẋ this is an example of a numerical difficulty analogous
    
r p to grazing bifurcation, mentioned in [34].
=g , t , effort-controlled (41)
ṙ r
By using Eq. (1), it can be seen that Eq. (38) is con- 5.4 Physically consistent models
sistent with the flow-controlled formulation Eq. (40).
By using Eq. (9), it can be seen that Eq. (39) is consis- Nonlinear constitutive models are notorious for allow-
tent with Eq. (41). Equation (21) is such an example. ing physically inconsistent solutions to dynamic prob-

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

1184 J.-S. Pei et al.

lems, with nonuniqueness being just one of the under- – How can we select state variables that are not only
lying issues [26]. Various authors have addressed phys- theoretically sound but also computationally fruit-
ical inconsistencies in the Bouc–Wen family of mod- ful? Herein, a successful application of the state
els; e.g., [12–14,18,28]. Before embarking on compu- event location algorithm and the ode45 solver has
tations, with or without state event location techniques, been presented for some challenging piecewise-
it is crucial to start with a well-posed model since oth- defined restoring force models.
erwise the resulting simulations will have no practical – Does the choice of state variables for these models
predictive value. imply a classification of these models, or should it
be vice versa?

6 Conclusions There is no simple or one-size-fits-all answer to

these fundamental questions. Different facets of these
The need for a sufficient and efficient problem formu- questions have been demonstrated through case stud-
lation of a system’s dynamics is of critical importance ies provided in this work and general guidelines are
in many engineering mechanics applications, e.g., sys- also provided. The Bouc–Wen model and other models
tem identification, damage detection and earthquake discussed herein (and elsewhere) can be implemented
response simulation, among many others. Two char- without the state event location algorithm by choos-
acteristics of some commonly used models have not ing the algorithmic time steps to be sufficiently small.
been researched thoroughly: One is their flow-/effort- A key conclusion from our work is that the option
controlled nature and the other, discontinuity in their of implementing the state event location algorithm is
state questions. beneficial in reducing the computational burden while
These two characteristics are coupled in numeri- providing the same level of accuracy. In the spirit of
cal simulation of these models. Normally, one would reproducible computational research [19], the MAT-
decide the state variables based on the flow-/effort- LAB code used in this work will be available upon
controlled nature, when algebraic variables might exist request to the first author.
as well. The “switching rules” and “switching timing”
(i.e., the state event locations) can be decided efficiently Acknowledgements Dr. Pei would like to acknowledge the
when we properly treat the discontinuity in the sim- hospitality of California Institute of Technology during her sab-
batical leave for the completion of this study. Dr. Gay-Balmaz
ulation. Without a good understanding of these two
is partially supported by the ANR project GEOMFLUID, ANR-
characteristics, inefficiency, inaccuracy and incorrect- 14-CE23-0002-01.
ness in problem formulation of a system’s dynamics
and model class building can be experienced. This is
the motivation for us to study these two characteris-
tics. In particular, the proper treatment of discontinu- A Appendix for Sect. 2.2
ities has been explored by us starting with [34]. We
also explored the flow-/effort-controlled nature of some Tables 2 and 3 list nonunique state-space represen-
models in [24,25]. tations of three- and four-parameter linear models,
Choices for state variables are not unique as demon- respectively.
strated in this study. This is not often made clear in
applying state-space representations. Computationally,
some choices are better than others in order to imple-
ment the state event location algorithm. There are two
questions of particular interest to us:

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

On choosing state variables 1185

Table 2 State variables for the three-element models each connected with a mass

k2 k1
k1 u(t) u(t)
m m

(a)
c2 (b)
k2 c2

k2 c1
c1 u(t) u(t)
m m

(c)
c2 (d)
k2 c2

(a) Standard solid model #1 connected in series with a mass

⎡ ⎤ ⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
x ẋ 0 1 0 x 0
y = ⎣ ẋ ⎦, ẏ = ⎣ ẍ ⎦ = ⎣ 0 0 − m ⎦ ⎣ ẋ ⎦ + ⎣ m1 ⎦ u
1
k1 k2 k1 +k2
r ṙ c2 k1 c2
r 0
Alternatively by replacing ẋ with p equivalently
⎡ ⎤ ⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ 1 ⎤
x ẋ 0 − m1 0 x m 
⎣
y= p , ⎦ ⎣
ẏ = r = ⎦ ⎣ 0 0 1 ⎦ ⎣ p ⎦ + ⎣ 0 ⎦ udt
k1 k2 k1 k1 +k2
c2 − m
r ṙ r k1
c2 m
(b) Standard solid model #2 connected in series with a mass
⎡ ⎤ ⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
x ẋ 0 1 0 x 0
y = ⎣ ẋ ⎦, ẏ = ⎣ ẍ ⎦ = ⎣ 0 − m ⎦ ⎣ ẋ ⎦ + ⎣ m1 ⎦ u
1
0
k1 k2 k2
c2 k1 + k2 − c2
r ṙ r 0
Alternatively by replacing ẋ with p equivalently
⎡ ⎤ ⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
x ẋ 0 − m1 0 x 1
m 
y = ⎣ p ⎦, ẏ = ⎣ r ⎦ = ⎣ 0 0 1 ⎦⎣ p⎦ + ⎣ 0 ⎦ udt
k1 k2 k1 +k2 k1 +k2
r ṙ c2 − m − kc22 r m
(c) Standard fluid model #1 connected in series with a mass
    ! "  
1 c1 c2 
r ṙ − c1k+c
2
− m1 c1 c2 k2 c1
c1 +c2 c1 +c2 r m c1 +c2
y= , ẏ = = 2 + 1 u
ẋ ẍ −1 m 0 ẋ m
Alternatively by lowering the order of both y and ẏ
    ! "   
p r − c1k+c
2
− m1 c1 c2 k2 c1
p 1 c1 c2
m c1 +c2

y= , ẏ = = 2 c1 +c2 c1 +c2 + udt
1
x ẋ −1 m 0 x m
Alternatively in an entirely effort-controlled form
        
p r 0 1 p  0
y= , ẏ = = +
r ṙ − km2 c1c+c
1
2
− k2
c1 +c2 − 1 c1 c2
m c1 +c2 r k2 c1
m c1 +c2 udt + 1 c1 c2
m c1 +c2 u
Alternatively in an entirely flow-controlled form
    ! 0 1
"   
x ẋ   x 0
y= , ẏ = = + k2
ẋ ẍ − km2 c1c+c
1
2
− k2
c1 +c2 + 1 c1 c2
m c1 +c2 ẋ m(c1 +c2 ) udt + u
m
(d) Standard fluid model #2 connected in series with a mass
    ! k2 c1 k2
"  
c1 
r ṙ − c2 − m c2 (c1 + c2 ) r
y= , ẏ = = + m
1 u
ẋ ẍ − m1 0 ẋ m
Alternatively by lowering the order of both y and ẏ
    ! k2 "   
− c2 − c1 k2 c1 
y=
p
, ẏ =
r
= m c2 (c1 + c2 ) p
+ m udt
1
x ẋ −1 m 0 x m

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

1186 J.-S. Pei et al.

Table 2 continued

Alternatively in an entirely effort-controlled form

        
p r 0 1 p 0
y= , ẏ = = +
r ṙ − km2 c1c+c
2
2
− k2
c2 −
c1
m r k2 c1 +c2
m c2 udt + c1
mu
Alternatively in an entirely flow-controlled form
    ! 0
"   
x ẋ  1  x  0
y= , ẏ = = +
ẋ ẍ − km2 c1c+c
1
2
− k2
c2 +
c1
m ẋ k2
mc2 udt + u
m

Table 3 State variables for the four-element models each connected with a mass

k1
k2
k2 c2 u(t) k1 c1 u(t)
m m
c1
(a) (b)
c2

k1 c1 k1 k2
u(t) u(t)
m m

(c)
k2 c2 (d)
c1 c2
(a) A generalized Maxwell model connected in series with a mass
⎡ ⎤ ⎡ ⎤ ⎡ 0 1 0
⎤⎡ ⎤ ⎡ ⎤
x ẋ x 0
⎢ − m1 ⎥⎣ ⎦ ⎣ 1 ⎦
y = ⎣ ẋ ⎦, ẏ = ⎣ ẍ ⎦ = ⎣ 0  0  ⎦ ẋ + m u
k1 k2 k1 c1 c1
r ṙ c2 k2 1 + k2 + c2 − kc12 − c1
m
r m
Alternatively by replacing ẋ with p equivalently
⎡ ⎤ ⎡ ⎤ ⎡ ⎤⎡ ⎤
x ẋ 0 − m1 0 x
⎢ ⎥⎣ ⎦
y = ⎣ p ⎦, ẏ = ⎣ r ⎦ = ⎣ 0  0  1 ⎦ p
k1 k2 k2 k1 c1 k2 c1
r ṙ c2 − m 1 + k2 + c2 − c1 − m
r
⎡ 1
⎤
⎥
m
⎢
+⎣ 0  ⎦ udt
c1 k2 k1 c1
m u + m 1 + k2 + c2
(b) A generalized Kelvin chain connected in series with a mass
⎡ ⎤ ⎡ ⎤ ⎡ 0 1 0
⎤⎡ ⎤ ⎡ ⎤
x ẋ x 0
⎢ − m1 ⎥⎣ ⎦ ⎣
y = ⎣ ẋ ⎦, ẏ = ⎣ ẍ ⎦ = ⎣ 0 ⎦
0 1
  ⎦ ẋ + mu 
r ṙ k1 k2 k1 k2
c2 k1 + m c1 c2
k1
c2 1+ k2
k1 + cc21 r − kc11 kc22 udt
Alternatively by replacing ẋ with p equivalently
⎡ ⎤ ⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
x ẋ 0 − m1 0 x 1
⎢ ⎥ m 
y = ⎣ p ⎦, ẏ = ⎣ r ⎦ = ⎣ 0 0  1  ⎦ ⎣ p ⎦ + ⎣ 0 ⎦ udt
k1 k2 k1 k1 k2 k1 k2
r ṙ c2 − m − c1 c2 c2 1+ k1 + cc21 r k1
m
(c) Two Maxwell models connected in parallel—before being connected in series with a mass
⎡ ⎤ ⎡ ⎤ ⎡ 0 1 0
⎤⎡ ⎤ ⎡ ⎤
x ẋ x 0
⎢ − m1 ⎥⎣ ⎦ ⎣
y = ⎣ ẋ ⎦, ẏ = ⎣ ẍ ⎦ = ⎣ ⎦
1
0  0 ⎦ ẋ + mu
k1 k2
r ṙ k1 k2 c11 + c12 k1 + k2 − m kc11 kc22 − kc11 − k2
c2
r c1 c2 udt

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

On choosing state variables 1187

Table 3 continued

Alternatively by replacing ẋ with p equivalently

⎡ ⎤ ⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
x ẋ 0 − m1 0 x 1
⎢ ⎥⎣ ⎦ ⎣ m 
⎣
y= p , ⎦ ⎣ ⎦
ẏ = r = ⎣ 0  0 1 ⎦ p + 0 ⎦ udt
r ṙ k1 k2 c11 + 1
c2 − k1 +k
m
2
− k1 k2
c1 c2 − kc11 − k2
c2
r k1 +k2
m
(d) Two Kelvin models connected in series—before being connected in series with a mass
⎡ ⎤ ⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
x ẋ 0 1 0 x 0
y = ⎣ ẋ ⎦, ẏ = ⎣ ẍ ⎦ = ⎣ 0 − m ⎦ ⎣ ẋ ⎦ + ⎣ m1 ⎦ u
1
0
k1 k2 k1 c2 +k2 c1
r ṙ c1 +c2 c1 +c2 − kc11 +k
+c2
2 r 0
Alternatively by replacing ẋ with p equivalently
⎡ ⎤ ⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
x ẋ 0 − m1 0 x 1
m 
y = ⎣ p ⎦, ẏ = ⎣ r ⎦ = ⎣ 0 0 1 ⎦⎣ p⎦ + ⎣ 0 ⎦ udt
+k2 c1
r ṙ k1 k2
c1 +c2 − m1 k1 cc21 +c2
− kc11 +k
+c2
2 r 1 k1 c2 +k2 c1
m c1 +c2

B Appendix for Sect. 3 and updated by tagging the entire solution history H
in order to switch according to the correct logic and
For the F formulation (flow-controlled) and according to correct discontinuity sticking. Hl , Hu and O are for
to the programming framework in [34], our previous the lower (Mode II), upper (Mode IV) bound values in
work, global variables are utilized for passing values of the event functions as in Table 4, and a constant in the
restoring force r , memory parameters Hl , Hu and O, algebraic equations as in Table 5, respectively. tag1 is
and tags tag1 and tag2 among different MATLAB m- used to mark each state event for correcting disconti-
files. In particular, the memory parameters are obtained nuity sticking, while tag2 is an indicator for the mode.

Table 4 Events, event functions and reset maps for F and M formulations
Transition Event Type Flow-controlled formulation (F) Partially effort-controlled formulation (M)
Event function Direction Event function Direction


1 I → IV 1 yF (1) − Hu (end) = 0 +1 yM (4) − (1 − α)r y = 0 +1

2 I → II 2 yF (1) − Hl (end) = 0 −1 yM (4) + (1 − α)r y = 0 −1

3 I→I 3 yF (2) = 0 0 yM (2) = 0 0

4 II → III 3 yF (2) = 0 0† yM (2) = 0 0

5 III → IV 1 yF (1) − Hu (end) = 0 +1 yM (4) − (1 − α)r y = 0 +1

6 III → II 2 yF (1) − Hl (end) = 0 −1 yM (4) + (1 − α)r y = 0 −1

7 III → III 3 yF (2) = 0 0 yM (2) = 0 0

8 IV → I 3 yF (2) = 0 0† yM (2) = 0 0

† Need to reset Hl , Hu and O

Table 5 Algebraic and state equations in flow maps obtained from piecewise restoring force expressions for F and M formulations,
respectively
Mode Flow-controlled (F) Partially effort-controlled (M)

I r = k(yF (1) − O(end)), with tag2 = 1 ẏM (3) = (1 − tag2 )(1 − α)kyM (2)) + αkyM (2), with tag2 = 0
II r = −r y + αk(yF (1) + x y ), with tag2 = 2 ẏM (3) = (1 − tag2 )(1 − α)kyM (2)) + αkyM (2), with tag2 = 1
III r = k(yF (1) − O(end)), with tag2 = 3 ẏM (3) = (1 − tag2 )(1 − α)kyM (2)) + αkyM (2), with tag2 = 0
IV r = r y + αk(yF (1) − x y ), with tag2 = 4 ẏM (3) = (1 − tag2 )(1 − α)kyM (2)) + αkyM (2), with tag2 = 1

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

1188 J.-S. Pei et al.

In terms of variables, x(t), ẋ(t) and r (t) are continuous 16. Karnopp, D.C., Margolis, D.L., Rosenberg, R.C.: Sys-
variables, while Hl , Hu , O, tag1 and tag2 are discrete tem Dynamics: Modeling, Simulation, and Control of
Mechatronics Systems, 5th edn. Wiley, Hoboken (2012)
variables. 17. Kerschen, G., Worden, K., Vakakis, A.F., Golinval, J.C.:
For the M formulation (mixed, partially effort- Past, present and future of nonlinear system identification
controlled) and as in our previous work, a global vari- in structural dynamics. Mech. Syst. Signal Process. 20,
able is utilized for passing values of tag2 among differ- 505–592 (2006)
18. Kottari, A.K., Charalampakis, A.E., Koumousis, V.K.: A
ent MATLAB m-files. tag2 is an indicator for the mode consistent degrading Bouc-Wen model. Eng. Struct. 60,
and is the only discrete variable to be taken care of. 235–240 (2014)
19. LeVeque, R.J.: Wave propagation software, computational
science, and reproducible research. In: Proceedings of the
International Congress of Mathematicians, European Math-
References ematical Society, Madrid, Spain, pp. 1227–1253 (2006)
20. Oh, J., Bernstein, D.S.: Semilinear duhem model for rate-
1. Baber, T.T., Noori, M.N.: Random vibration of degrading, independent and rate-dependent hysteresis. IEEE Trans.
pinching systems. ASCE J. Eng. Mech. 111(8), 1010–1026 Autom. Control 50(5), 631–645 (2005)
(1985) 21. Park, T., Barton, P.I.: State event location in differential-
2. Baber, T.T., Noori, M.N.: Modeling general hysteresis algebraic models. ACM Trans. Model. Comput. Simul.
behavior and random vibration application. ASME J. Vib. 6(2), 147–165 (1996)
Acoust. Stress Reliab. Des. 108, 411–420 (1986) 22. Paynter, H.M.: Analysis and Design of Engineering Sys-
3. Benedettini, F., Capecchi, D., Vestroni, F.: Identification of tems: Class Notes for M.I.T. Course 2.751. M.I.T. Press,
hysteretic oscillators under earthquake loading by nonpara- Cambridge (1961)
metric models. ASCE J. Eng. Mech. 121(5), 606–612 (1995) 23. Paynter, H.M.: The gestation and birth of bond graphs.
4. Bouc, R.: Forced vibration of mechanical systems with https://www.me.utexas.edu/~longoria/paynter/hmp/Bond
hysteresis. In: Proceedings of 4th Conference on Nonlinear graphs.html (2000)
Oscillations (1967) 24. Pei, J.S., Wright, J.P., Todd, M.D., Masri, S.F., Gay-Balmaz,
5. Capecchi, D.: Accurate solutions and stability criterion for F.: Understanding memristors and memcapacitors for engi-
periodic oscillations in hysteretic systems. Meccanica 25, neering mechanical applications. Nonlinear Dyn. 80(1),
159–167 (1990) 457–489 (2015)
6. Caughey, T.K.: Random excitation of a system with bilinear 25. Pei, J.S., Gay-Balmaz, F., Wright, J.P., Todd, M.D.,
hysteresis. J. Appl. Mech. 27, 649–652 (1960a) Masri, S.F.: Dual input–output pairs for modeling hys-
7. Caughey, T.K.: Sinusoidal excitation of a system with teresis inspired by mem-models. Nonlinear Dyn. 88(4),
bilinear hysteresis. J. Appl. Mech. 27, 640–643 (1960b) 2435–2455 (2017)
8. Dormand, J.R., Prince, P.J.: A family of embedded Runge– 26. Sandler, I.S.: On the uniqueness and stability of endochronic
Kutta formulae. J. Comput. Appl. Math. 6(1), 19–26 (1980) theories of materials behavior. ASME J. Appl. Mech. 45,
9. Esposito, J.M., Kumar, V.: A state event detection algorithm 263–266 (1978)
for numerically simulating hybrid systems with model 27. Shampine, L., Gladwell, I., Brankin, R.: Reliable solution
singularities. ACM Trans. Model. Comput. Simul. 17(1), of special event location problems for ODEs. ACM Trans.
1–22 (2007) Math. Softw. 17(1), 11–25 (1991)
10. Foliente, G.C.: Hysteresis modeling of wood joints and 28. Sivaselvan, M.V., Reinhorn, A.M.: Hysteretic models for
structural systems. ASCE J. Struct. Eng. 121(6), 1013–1022 deteriorating inelastic structures. ASCE J. Eng. Mech.
(1995) 126(6), 633–640 (2000)
11. Goebel, R., Sanfelice, R., Teel, A.: Hybrid dynamical 29. Strukov, D.B., Snider, G.S., Stewart, D.R., Williams, R.S.:
systems. IEEE Control Syst. Mag. 28, 28–93 (2009) The missing memristor found. Nature 453, 80–83 (2008)
12. Ikhouane, F., Rodellar, J.: Physical consistency of the hys- 30. van der Schaft, A., Schumacher, H.: An Introduction to
teretic Bouc-Wen model. IFAC Proc. Vol. 38(1), 874–879 Hybrid Dynamical Systems. Lecture Notes in Control and
(2005) Information Sciences (Book 251). Springer, Berlin (1999)
13. Jayakumar, P.: Modeling and identification in structural 31. Wen, Y.K.: Method for random vibration of hysteretic
dynamics. Ph.D. thesis, California Institute of Technology, systems. ASCE J. Eng. Mech. 102(2), 249–263 (1976)
Pasadena, CA (1987) 32. Wen, Y.K.: Equivalent linearization for hysteretic systems
14. Jayakumar, P., Beck, J.L.: System identification using under random excitation. ASME J. Appl. Mech. 47,
nonlinear structural models. In: Nake, H.G., Yao, J.T.P. 150–154 (1980)
(eds.) Structural Safety Evaluation Based on System 33. Willems, J.C.: Dissipative dynamical systems part I: general
Identification Approaches, Friedr. Vieweg & Sohn Braun- theory. Arch. Ration. Mech. Anal. 45, 321–351 (1972)
schweig/Wiesbaden, Vieweg International Scientific Book 34. Wright, J.P., Pei, J.S.: Solving dynamical systems involving
Series, vol Proceedings of the Workshop at Lambrecht/Pfalz, piecewise restoring force using state event location. ASCE
pp 82–102 (1988) J. Eng. Mech. 138(8), 997–1020 (2012)
15. Kalmár-Nagy, T., Shekhawat, A.: Nonlinear dynamics of
oscillators with bilinear hysteresis and sinusoidal excitation.
Phys. D 238, 1768–1786 (2009)

123

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”),
for small-scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are
maintained. By accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use
(“Terms”). For these purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or
a personal subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or
a personal subscription (to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the
Creative Commons license used will apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data
internally within ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking,
analysis and reporting. We will not otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of
companies unless we have your permission as detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that
Users may not:

1. use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to
circumvent access control;
2. use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil
liability, or is otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by
Springer Nature in writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer
Nature journal content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates
revenue, royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain.
Springer Nature journal content cannot be used for inter-library loans and librarians may not upload Springer Nature journal
content on a large scale into their, or any other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any
information or content on this website and may remove it or features or functionality at our sole discretion, at any time with or
without notice. Springer Nature may revoke this licence to you at any time and remove access to any copies of the Springer Nature
journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express
or implied with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or
warranties imposed by law, including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be
licensed from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other
manner not expressly permitted by these Terms, please contact Springer Nature at

[email protected]