Hindawi

Shock and Vibration

Volume 2020, Article ID 3593018, 13 pages
https://doi.org/10.1155/2020/3593018

Research Article
Nonlinear Dynamics of a Duffing-Like Negative Stiffness
Oscillator: Modeling and Experimental Characterization

D. Anastasio , A. Fasana, L. Garibaldi, and S. Marchesiello

Dipartimento di Ingegneria Meccanica e Aerospaziale, Politecnico di Torino, Corso Duca Degli Abruzzi 24, 10129 Torino, Italy

Correspondence should be addressed to S. Marchesiello; [email protected]

Received 15 November 2019; Accepted 22 April 2020; Published 15 May 2020

Academic Editor: Alba Sofi

Copyright © 2020 D. Anastasio et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

In this paper, a negative stiffness oscillator is modelled and tested to exploit its nonlinear dynamical characteristics. The oscillator
is part of a device designed to improve the current collection quality in railway overhead contact lines, and it acts like an
asymmetric double-well Duffing system. Thus, it exhibits two stable equilibrium positions plus an unstable one, and the os-
cillations can either be bounded around one stable point (small oscillations) or include all the three positions (large oscillations).
Depending on the input amplitude, the oscillator can exhibit linear and nonlinear dynamics and chaotic motion as well.
Furthermore, its design is asymmetrical, and this plays a key role in its dynamic response, as the two natural frequencies associated
with the two stable positions differ from each other. The first purpose of this study is to understand the dynamical behavior of the
system in the case of linear and nonlinear oscillations around the two stable points and in the case of large oscillations associated
with a chaotic motion. To accomplish this task, the device is mounted on a shaking table and it is driven with several levels of
excitations and with both harmonic and random inputs. Finally, the nonlinear coefficients associated with the nonlinearities of the
system are identified from the measured data.

1. Introduction one, and the oscillations can either be bounded around one
stable point (in-well or intrawell, small oscillations) or in-
Devices and materials exhibiting a negative stiffness phase clude all the three positions (cross-well or infrawell, large
are often used as vibration isolators due to their amplified oscillations). In both cases, periodic oscillations can evolve
damping properties [1, 2]. In particular, in the case of en- to steady in-well or cross-well chaotic motions under ex-
gineering structures, such devices are usually designed ternal excitation [8, 9]. The occurrence of irregular motion,
adopting discrete macroscopic elements, such as post- consisting of random-like crossings from oscillations
buckled beams, plates, shells, and precompressed springs, around the two stable equilibrium positions, was first ob-
arranged in appropriate geometrical configurations. Ex- served experimentally in 1971 [10] and the motion was called
amples can be found in automotive suspensions [3, 4] or “snap-through”. This behavior is associated with an unstable
seismic isolation [5, 6]. phase generated by the negative stiffness effect; thus, the
While the practical applications of negative stiffness stability analysis plays a crucial role in the design process of
systems are relatively recent, the theoretical studies about these kinds of systems [11]. Many studies have been con-
their dynamical behavior have a much longer story. This is ducted starting from the 1970s to analyze the dynamics of
because of the wide kind of motions they can exhibit, this oscillator and the effects of the unstable paths, including
ranging from linear to highly nonlinear and chaotic [7]. In phenomena such as bifurcations, subharmonics, period
particular, when the negative stiffness effect is coupled to a doublings, and chaos. A comprehensive literature review can
nonlinear polynomial stiffness contribution, the so-called be found in [9].
double-well Duffing oscillator is retrieved. This oscillator In this work, a negative stiffness oscillator is studied and
exhibits two stable equilibrium positions plus an unstable tested. The oscillator is part of a device designed to improve
2 Shock and Vibration

the current collection quality in railway overhead contact Moving mass

lines, attempting to alter their damping distribution and
reducing the wave propagation [12].
Hinge U-shaped
A two-fold objective is pursued: first, the dynamical frame
properties of the oscillator are analyzed, replicating exper-
imentally the possible kind of motions it can exhibit (in-well,
cross-well, and chaotic); second, nonlinear system identi- Rods
fication is performed to extract the model parameters di-
rectly from the measurements. The latter in particular seems
to be quite a challenging task, given the rich nonlinear
dynamics the device is capable of showing and the strength
of the nonlinear response. To accomplish these tasks, the
oscillator is mounted on a shaking table and it is driven
Shaking table
through several levels of excitations with both harmonic and
random inputs. The nonlinear system identification is Figure 1: Schematic representation of the negative stiffness
performed by applying the nonlinear subspace identification oscillator.
(NSI) method [13–15]. Generally, NSI estimates a nonlinear
state-space model starting from the input-output data, by
separating the nonlinear part of the response from the so-
called underlying linear system. This is not a straightforward θ (t)
operation for the case considered here though. The bistable m y (t) mg
nature of the device implies that the underlying linear system
h p (θ) m p (θ)
is not stable, as it exhibits a negative linear stiffness.
Therefore, NSI cannot be directly applied and some prior
α my¨
data manipulation is required. Nonetheless, the identifica- b (t)
tion strategy proposed in this work is capable of estimating a
both the coefficients of the nonlinearities and the underlying
Figure 2: Model of the negative stiffness oscillator and free-body
linear systems associated with the stable equilibria starting diagram of m.
just from one cross-well measurement, with satisfying ac-
curacy. The nonexistence of a stable underlying linear system
and the strong nonlinear behavior make the test set really
interesting for the purposes of nonlinear system identifi- The equilibrium equation along the vertical coordinate is
cation, confirming the effectiveness of the nonlinear sub- € + 2p(θ(t))sin(θ(t)) + mg � 0.
my(t) (1)
space identification method. Eventually, the restoring force
surface (RFS) method [16, 17] is also applied as a Calling z(t) � y(t) − b(t), one obtains
comparison. €
€ + 2p(θ(t))sin(θ(t)) + mg � −mb(t).
mz(t) (2)

2. A Negative Stiffness Oscillator The system is a single-degree-of-freedom system; thus,

just one variable is needed to describe the motion. In
The device here studied consists in a U-shaped steel frame particular, since vertical displacements and accelerations are
connected through rods to a central moving mass. The frame measured in the experimental setup, the variable z(t) is
has the purpose to keep the rods under compression during taken as an independent variable; therefore, both θ and p(θ)
their movement so that a bistable mechanism is achieved. A should be written as a function of z. The elasticity of the
schematic representation of the device can be seen in frame can be analyzed to obtain the force that the frame
Figure 1. transmits to the rods. Just half of the frame is considered in
The lower surface of the frame is attached to a shaking the following, as depicted in Figure 3.
table so that a displacement b(t) can be imposed to the When the mass moves along the vertical axis, the frame
structure. It is also assumed that the inertia of the moving bends and deforms until a new equilibrium position is
parts can be concentrated into one central point with mass reached, corresponding to the resting (undeformed) posi-
m, comprising the mass of the central bushing and the tion of the frame. The analytic expression of p(θ) can be
equivalent inertia of the rods. The vertical movement of this found by studying the flexibility of the half-frame, consid-
point is described by the coordinate y(t), while the rotation ered as two connected cantilever beams under bending
of the rods is called θ. The elasticity of the frame is also taken stress. It is not the purpose of this work to analytically derive
into account, and it acts like a compression force p(θ) to the the expression of p(θ). Instead, a qualitative representation
rods. Since the flexural elasticity of the frame is much higher of its total vertical component pζ (θ) � 2p(θ)sin(θ) is
than the axial elasticity of the rods, the latter is considered depicted in Figure 4 as a function of z(θ).
infinitely rigid. A schematic representation of the functional It can be seen that pζ (z) has three roots and crosses the
model here described is reported in Figure 2, together with origin with a negative slope. Also, it can be approximated as
the free-body diagram of the mass m. a polynomial expansion of degree 3:
Shock and Vibration 3

θ0
θ1

Potential U (z)
)
p (θ 1 z∗0
0
z–∗ z+∗

0
z
Figure 5: Potential U(z). Orange dots: equilibrium positions.

When the oscillations are bounded around one equi-

librium position z±∗ , the system acts like a stable nonlinear
ζ oscillator, and the associated linear natural frequency ω± can
be computed by
􏽳�������
χ U″ z±∗ 􏼁
ω± � , (7)
m
with U″ (z±∗ ) being the second derivative of U(z) computed
Figure 3: Model of the half-frame. in z∗− or z∗+ . It is worth writing the equation of motion
considering the oscillations around one of the possible
equilibrium points. To ease the notation, the generic equi-
librium position is called z∗ . A new variable x(t) can be
defined as
pζ (z)

0 x(t) � z(t) − z∗ . (8)

Substituting equation (8) in equation (4) yields

0 €
mx€ + K x + z∗ 􏼁 � −mb, (9)
z (θ)
Figure 4: Qualitative graph of the force pζ (θ). with
2
K x + z∗ 􏼁 � k3 x3 + k2 + 3k3 z∗ 􏼁x2 + 􏼐3k3 z∗ 􏼁
(10)
pζ (z) � k3 z3 + k2 z2 − k1 z. (3) + 2k2 z∗ − k1 􏼁x � k3 x3 + k􏽥2 x2 + k􏽥1 x.
The equation of motion can eventually be written as Finally, equation (9) becomes
€
mz€ + k3 z3 + k2 z2 − k1 z + mg � −mb. (4) mx€ + k3 x3 + k􏽥2 x2 + k􏽥1 x � −mb.
€ (11)

Equation (4) has the form of a negative stiffness Duffing The advantage of using this formulation instead of
equation, whose coefficients k3 , k2 , and k1 have to be esti- equation (4) is that it allows the definition of a stable un-
mated. In particular, the final objective of this work is to derlying linear system. This is a crucial requirement for the
identify them directly from the experimental measurements, nonlinear subspace identification method, adopted in Sec-
in order to assure a good correspondence between the nu- tion 4 to perform the nonlinear system identification of the
merical model and the real measured behavior. Also, the structure under test.
restoring force K(z) and the potential U(z) can be defined as
K(z) � k3 z3 + k2 z2 − k1 z + mg, (5) 3. Experimental Characterization
1 1 1 Two photos of the experimental setup corresponding to the
U(z) � k3 z4 + k2 z3 − k1 z2 + mgz. (6) two stable equilibrium positions are reported in Figure 6.
4 3 2
The moving mass is instrumented with an accelerometer to
A qualitative representation of the potential is shown in € and a laser vibrometer
measure its absolute acceleration y(t)
Figure 5, where its double-well nature can be clearly to measure its absolute displacement y(t). The zero position
identified. The potential is not symmetric because of the of y(t) corresponds to the horizontal configuration of the
gravitational contribution and the asymmetry of the frame. rods (θ � 0). The acceleration of the base b(t) € is also
Also, the three equilibrium positions are depicted, obtained recorded through a second accelerometer, and the dis-
by setting K(z∗ ) � 0. Two out of three positions represent a placement z(t) is computed as the difference between the
stable equilibrium, namely, z∗− and z∗+ , while the central laser measure y(t) and the displacement of the base b(t),
position z∗0 is an unstable equilibrium point. obtained by integrating twice its measured acceleration.
4 Shock and Vibration

(a) (b)

Figure 6: Photos of the experimental setup: (a) negative equilibrium position z∗− ; (b) positive equilibrium position z∗+ .

The system is excited with both harmonic and random 0

inputs to characterize its linear and nonlinear behavior.

z (m)
–0.02
3.1. Random Tests. A first set of measurements is performed
with a random input over the frequency range 7–50 Hz. The –0.04
sampling frequency is fs � 512 Hz, and the duration of the 0 10 20 30 40 50 60
Time (s)
acquisition is t � 300 s. Several forcing levels are applied,
expressed as RMS values of the acceleration of the base b, € and (a)
∗
the starting position is z− . The results are depicted in Figure 7 –50
in terms of time series and transmissibility T, defined as the
T (dB)

ratio between the output acceleration z€ and the input b. € It can

be noted that the mean value of the displacements is not null,
–100
as in-well oscillations take place in the neighborhood of the 10 20 30 40 50
equilibrium position, which in this case is approximately at Frequency (Hz)
−0.03 m from the horizontal configuration of the rods
(b)
(θ � 0). Also, the asymmetry of the displacements increases
as the forcing level, as the system tries to cross the negative Figure 7: Random tests. Black line: b€ � 7 m/s2 RMS; orange line:
stiffness region and to reach the positive equilibrium position. b€ � 9 m/s2 RMS; blue line: b€ � 26 m/s2 RMS. (a) Time history of the
This results in a clear change in the transmissibility, where a displacement (first 60 seconds); (b) transmissibility T.
softening effect can be seen when increasing the excitation
level, in accordance with the theoretical studies that show a
because of the possibility of chaotic motion. A logarithmic
similar behavior in the case of in-well motion [9].
sine-sweep in the frequency range 7–21 Hz is considered first.
If the energy given to the system is high enough, cross-well
The sampling frequency is fs � 512 Hz, and the length of the
oscillations are retrieved and the moving mass oscillates in a
up-down sweep is t � 240 s. Two forcing levels are taken into
wider region, including the two stable equilibrium positions z∗−
account, expressed as the amplitude b0 of the base dis-
and z∗+ . This situation is represented in Figure 8, where the
placement, and the starting position is the negative equilib-
displacement clearly shows repeated crossings between nega-
rium z∗− . The up and down sweeps are shown in Figure 9.
tive and positive values. Also, the statistical distribution of z(t)
Both the excitation levels clearly produce a nonlinear
depicted in Figure 8(b) highlights the asymmetry of the
behavior for the in-well motion, as classic jump phenomena
structure so that oscillations around the negative position are
can be observed in the up and down sweeps. The softening
roughly the 66% of the total acquisition length. This result
effect can be seen in Figure 9(a), where an increase in the
agrees with the shape of the potential retrieved from the model
forcing amplitude corresponds to an earlier occurrence of
described in the previous section, which shows two wells with
the jump-up. Also, two distinct jumps can be noticed,
different heights. The final confirmation will be given in the
corresponding to the dominant frequency (10-11 Hz) and its
following section where the experimental data are processed so
second harmonic (20-21 Hz).
as to identify the actual potential of the system.
If the input is high enough, cross-well motion is retrieved
also in this case (Figure 9, orange line). It is interesting to
3.2. Sine-Sweep and Harmonic Tests. Harmonic excitations look at the harmonic contributions in this case by com-
are a powerful tool to understand the behavior of nonlinear puting the spectrogram of the relative displacement. The
systems. This is particularly true for the case considered here, result is reported in Figure 10, where the first two minutes
Shock and Vibration 5

0.04
0.02

Occurrences
z (m)

0 200
–0.02 100
–0.04 0
0 10 20 30 40 50 60 –0.04 –0.02 0 0.02 0.04
Time (s) z (m)
(a) (b)

Figure 8: Random test at the highest amplitude, b€ � 38 m/s2 RMS: (a) time history of the displacement (first 60 seconds); (b) statistical
distribution of the displacement.

0.04 0.04

0.02 0.02

0 0
z (m)

z (m)
–0.02 –0.02

–0.04 –0.04

10 15 20 20 15 10
Instantaneous frequency (Hz) Instantaneous frequency (Hz)
(a) (b)

Figure 9: Sine-sweep tests. Black line: b0 � 4.5 mm; orange line: b0 � 5 mm. (a) Sweep up; (b) sweep down.

100 –40

80 –60
Frequency (Hz)

–80
Power (dB)
60
–100
40
–120

20 –140

z+∗

z∗–
0.5 1 1.5 2 2.5 3 3.5
Time (min)
Figure 10: Spectrogram of the highest level sine-sweep test and corresponding time-domain response.

refer to the sweep up, while the second two minutes refer to performed to analyze these effects. The excitation frequency
the sweep down. is ] � 9 Hz, and three different amplitudes b0 are considered.
Both even and odd harmonics of the instantaneous The results are presented in the phase diagrams in Figure 11.
frequency are present along the whole acquisition, con- The different excitation amplitudes result in different
firming the asymmetrical behavior of the nonlinear system. kinds of motion of the moving mass, ranging from harmonic
Subharmonics are also visible in some regions, in particular oscillations to cross-well motion. In particular, when the
around 2 minutes. Generally, they are symptomatic of the amplitude is b0 � 2 mm (Figure 11(a)), the phase plane
possibility of bifurcations and chaotic motion [18, 19], and shows one closed orbit centered around the equilibrium
thus, a series of harmonic tests with constant frequency is position, i.e., one periodic solution [20]. When the
6 Shock and Vibration

0.1 1

z˙ (m /s)
z˙ (m /s)

0 0
–0.1
–1
–0.04 –0.02 0 0.02 0.04 –0.04 –0.02 0 0.02 0.04
z (m) z (m)
(a) (b)

z˙ (m /s) 0

–1
–0.04 –0.02 0 0.02 0.04
z (m)
(c)

Figure 11: Phase diagrams of the response under harmonic excitation: (a) b0 � 2 mm; (b) b0 � 4.7 mm; (c) b0 � 5 mm.

amplitude increases to b0 � 4.7 mm, some nested orbits can

be seen in the phase diagram (Figure 11(b)). Two paths in a –100
closed loop are generally representative of the period Z (dB)
doubling effect [7] so that a subharmonic with twice the –150
period of the dominant shows up. Such a behavior is re-
trieved in the experimental case of Figure 11(b), and this is
–200
clearer when looking at the power spectral density Z of the
displacement, represented in Figure 12. It can be seen that
the system responds at both integer multiples of the dom- 1/2ν ν 3/2ν 2ν 5/2ν 3ν 7/2ν 4ν
inant frequency ] (2], 3], . . .) and of its subharmonic Frequency
(1/2)v ((3/2)v, (5/2)v, . . .). Figure 12: Power spectral density of the harmonic response,
As for the highest excitation level in Figure 11(c), a cross- b0 � 4.7 mm.
well motion occurs, and the solution is not periodic any-
more. This behavior lasts for the whole acquisition (10
minutes), and a portion of the time response can be seen in
0.02
Figure 13.
This kind of persisting nonperiodic response to a peri-
odic excitation is a symptom of chaotic behavior. It should 0
z (m)

be recalled that no definition of chaos is universally accepted,

and this is particularly true when experimental data are –0.02
considered. The reason is that uncertainties and noise in the
data acquisition may interfere with the extreme sensitivity to
the initial conditions that characterize chaotic systems. An –0.04
420 430 440 450 460
useful way to check whether a system is behaving chaotically Time (s)
or not is to look at its largest Lyapunov exponent λ [20]. A
positive sign of λ means chaotic motion, while a negative Figure 13: Time response under harmonic excitation, b0 � 5 mm.
sign is a representative of a periodic orbit. Several methods
exist to compute λ from experimental time series, and the dynamics, as it covers all the possible positions of the moving
one proposed in [21] is adopted here. The results are shown mass. For this reason, cross-well measurements will be used
in Figure 14, where a positive λ is retrieved. in the following section to identify the nonlinear model
Eventually, the experimental Poincaré sections are parameters.
computed for different phase synchronizations ϕ of the data
with the forcing term [22]. The typical shape of a strange 4. Nonlinear System Identification
attractor is retrieved [8] and depicted in Figure 15(a) in a
polar plot, while one of its sections is represented in A first guess of the nonlinear restoring force K and the
Figure 15(b). potential U is estimated from the measured time series using
The considered structure is then capable of exhibiting a the restoring force surface method (RFS) [16, 17]. This
variety of motions, as expected from the theory. Among method is fairly simple and allows to visualize the nonlin-
them, the cross-well is surely the richest in terms of earity, providing a very useful insight to the user. On the
Shock and Vibration 7

20

15

10

5
0 500 1000 1500 2000 2500
Iterations
Figure 14: Estimation of the Lyapunov exponent. Red line: convergence mean value.

(a)

ż z ż z ż
z

(b) (c) (d)

Figure 15: Experimental Poincaré sections, b0 � 5 mm: (a) polar representation of the attractor surface; (b) Poincaré section, ϕ � 0° ; (c)
Poincaré section, ϕ � 170° ; (d) Poincaré section, ϕ � 320° .

contrary, its capabilities are very limited when trying to where f(t) is the forcing term and R(z, z) _ is the nonlinear
identify a nonlinear model structure, as it is essentially based restoring surface, generally a function of the displacement z
on raw data processing and simple fitting. For this reason, _ If the inertial term is shifted to the right-
and the velocity z.
the final model is identified using the nonlinear subspace hand side of the equation, the restoring surface can then
identification method (NSI) [13–15]. NSI gives as outcomes easily be computed and its features extracted. In particular, if
the fully nonlinear state-space representation of the system, small velocities are taken into account, such that |z| _ < εs , the
together with the FRF of the so-called underlying linear obtained slice of the restoring surface approximates the
system and the estimation of the coefficients defining the restoring force K(z). On the contrary, when small dis-
nonlinearities. It requires the input-output data and the placements around the equilibrium positions are considered,
knowledge of the nonlinear basis functions, whose selection such that |z − z∗ | < εd , an approximation of the damping
can take advantage from the preliminary RFS analysis. force, called D(z),
_ is retrieved.
For the case considered here, the cross-well measure-
ments are used to build R(z, z). _ The velocity z_ is obtained
4.1. First Step: RFS Method. The equation of motion de- by integrating the accelerations y€ and b€ and subtracting
scribing the dynamical system considered here can be them. The obtained restoring surface is reported in Fig-
written in the following form: ure 16, together its sections K(z) and D(z), _ computed
mz€ + R(z, z)
_ � f(t), (12) setting εs � εd � 0.1%.
8 Shock and Vibration

20

R (N)
0

–20

1
0 z+∗
˙z (m –1 0
/s) z–∗
z (m)

20
5
0

D (N)
K (N)

–20 –5

z∗– 0 z+∗
–1 0 1
z (m)
z˙ (m/s)

Figure 16: Experimental restoring surface. Blue dots: restoring force, εs � 0.1%; orange dots: damping force, εd � 0.1%.

The restoring force is eventually fitted to a polynomial 60

expansion of degree 3 as in equation (5), obtaining

K(z) � k3 z3 + k2 z2 − k1 z + k0 � 7.35 · 105 z3 40

(13)
+ 1.56 · 103 z2 − 550 z + 2.4 ,
20
K (N)

where k0 is the static component due to gravitational effects.

Once the expression of K(z) is obtained, the potential U(z) 0
can be computed from equation (6). Both K(z) and U(z)
are shown in Figure 17. As for the theoretical model pre- –20
sented in Section 2, the potential shows two wells with
different heights and three equilibrium positions are iden-
tified, two stable and one unstable. The final values of the –40
–0.05 0 0.05
coefficients defining the nonlinearity are estimated using z (m)
NSI, and a comparison with the ones obtained using RFS is
(a)
also reported in Section 4.2.
0.4
As for the damping plot in Figure 16, it is rather difficult
to estimate a proper damping model due to the very high
dispersion of D(z). _ Nevertheless, a possible and realistic
guess is that some friction is present between the bushing of
0.2
the moving mass and the vertical steel guide. For this reason,
U (Nm)

a nonlinear damping function is also added to the set of

nonlinear basis functions given to NSI in the following.
0

4.2. Second Step: NSI Method. The nonlinear subspace

identification method is based on a nonlinear state-space
representation of the system, obtained by considering the
–0.2
nonlinear terms as feedbacks to the underlying linear system –0.05 0 0.05
and processing the measured input-output data using the z (m)
subspace formulation. The existence of the underlying linear
(b)
system is an essential requirement for the method to work.
In the present study, this requirement is not fulfilled when Figure 17: RFS estimation of the restoring force K(z) (a) and the
the system goes through cross-well oscillations because of corresponding potential U(z) (b). Red dots: stable and unstable
the bistable nature of the device. A possible overcome to this _ < εs .
equilibrium positions; gray dots: restoring surface for |z|
Shock and Vibration 9

issue has been already shown in Section 2, and it consists of a 25 1

simple shift of the reference axis with respect to one stable
equilibrium position z∗ , called “reference value” hereafter. It 20 0.8

Inertance, G (dB)

Coherence, γ2
should be noted that the two stable equilibrium positions of 0.6
the system correspond to two underlying linear systems 15
mutually exclusive, each one existing only when the moving 0.4
mass oscillates around that equilibrium position. Thus, the 10
two underlying linear systems can be identified with NSI 0.2
5
from a single cross-well measurement by selecting the 0
corresponding equilibrium position as the reference value. 10 15 20 25 30 35 40 45 50
Whatever reference position is used, a new displacement Frequency (Hz)
variable x(t) � z(t) − z∗ can be defined, as in equation (11). Figure 18: Inertance of the system, random input with amplitude
The coefficients of the two nonlinear functions become k3 38 m/s2 RMS. Blue dotted line: coherence.
and k􏽥2 , and nonlinear damping is also considered in the
form x|_ x|.
_ Equation (11) can then be rewritten according to
the feedback formulation: 14

12
mx€ + cx_ + k􏽥1 x � −Rnl (x, x) €
_ − mb, (14)
10
where a linear viscous damping contribution is also con-

Model order
sidered and Rnl (x, x)_ is the assumed nonlinear restoring 8
surface, equal to
6
_ � k x3 + k􏽥 x2 + cnl x|
Rnl (x, x) 3 2 _ x|.
_ (15)
4
Referring to Section 3.1, the random test with excitation
2
amplitude 38 m/s2 RMS (Figure 8) is considered for the
nonlinear system identification with NSI. In particular, the 0
last 60 seconds are used as a validation set for the evaluation 5 10 15 20 25 30
Frequency (Hz)
of the residuals with the measured output, while the iden-
tification is performed on the remaining part of the ac- Figure 19: Stabilization diagram for NSI with reference position
quisition. The inertance of the system G is depicted in z∗ � z∗− . Stabilization thresholds for natural frequency, damping
Figure 18, together with the coherence estimation. As ex- ratio, MAC, and modal mass are 1%, 20%, 99.5%, and 20%, re-
pected, the FRF is very distorted due to the strong nonlinear spectively. Gray dot: new (not stable) pole; blue plus: pole stable in
contributions, and the coherence drops to very low values in frequency; red square: pole stable in frequency and MAC (modal
assurance criterion); orange circle: pole stable in frequency, MAC,
the frequency region up to the resonance peak.
and damping; green cross: pole stable in frequency, MAC,
NSI is performed considering first the negative equi- damping, and modal mass.
librium position as a reference value, meaning z∗ � z∗− .
Stability is checked varying the model order for frequencies,
damping ratios, MACs, and modal masses [15], and the
stabilization diagram is reported in Figure 19. Since the Table 1: Modal parameters of the two underlying linear systems
modal parameters show stability from a model of order 2, identified with NSI.
this is the selected value. Reference position Natural frequency (Hz) Damping ratio (%)
Once the model order has been selected, the fully z∗− 11.41 11.2
nonlinear state-space model is retrieved, together with the z∗+ 9.19 20.3
underlying linear system and the coefficients of the non-
linear basis functions. This is repeated also when the ref-
erence value is equal to the positive equilibrium position, and complex-valued quantities [13]. Assuming that the true
meaning z∗ � z∗+ . The stabilization diagram for this case is coefficients are real numbers, the imaginary part of the
not reported since it is similar to the one in Figure 19, and a identified counterparts should then be zero and the real part
model order equal to 2 is accomplished also in this case. The should not depend on the frequency. Since this happens only
identified modal parameters of the two underlying linear in complete absence of noise and nonlinear modeling errors,
systems related to the two reference values are reported in the ratio between real and imaginary parts can be taken as an
Table 1 in terms of natural frequencies and damping ratios. indicator of the goodness of the nonlinear basis functions
The FRFs of the two underlying linear systems are choice.
depicted in Figure 20 together with the measured (non- Running NSI for the two different reference values
linear) one, already shown in Figure 18. means that not only two independent underlying linear
The coefficients of the nonlinear basis functions can also systems are retrieved but also two nonlinear state-space
be computed from the nonlinear state-space model, recalling models. This result in a double estimation of the coefficients
that they are treated by the method as frequency-dependent of each nonlinearity. In particular, the estimation of the
10 Shock and Vibration

20

Inertance, G (dB)
15

10

5
10 15 20 25 30 35 40 45 50
Frequency (Hz)
Figure 20: Inertances of the two underlying linear systems. Gray dots: measured inertance of the nonlinear test; blue line: NSI estimation of
the linear inertance associated with the negative equilibrium z∗− ; orange line: NSI estimation of the linear inertance associated with the
positive equilibrium z∗+ .

×105
8

k3
7.5

0 5 10 15
Frequency (Hz)
(a)

×104
5
~
k2 0
–5
0 5 10 15
Frequency (Hz)
(b)

–4.4
cnl
–4.6

–4.8
0 5 10 15
Frequency (Hz)
(c)

Figure 21: Real parts of the identified coefficients as frequency-dependent quantities. Blue line: NSI with reference position z∗ � z∗− ; orange
line: NSI with reference position z∗ � z∗+ . (a) Coefficient k3 (N/m3 ); (b) coefficient k􏽥2 (N/m2 ); (c) coefficient cnl (Ns2 /m2 ).

coefficients k3 and cnl should be the same when NSI is It is worth highlighting that the imaginary part is always
applied to the two reference positions, as they are both much lower than the absolute value of the real part in the
invariant in the equation of motion (equations (14) and selected frequency range, which assesses the goodness of the
(15)). On the contrary, k􏽥2 depends on the choice of z∗ . The identification. Also, the real parts of k3 and cnl computed
identified coefficients are depicted in Figure 21 in their real with respect to the two reference values are very close to each
parts for the two reference positions. other, as expected. Since the system has a “favorite” equi-
A summary of the identified coefficients is also reported librium position, which is the negative one, the number of
in Table 2 for the two reference values. The final value of each samples associated with oscillations around this position are
coefficient is computed as the spectral mean of its real part higher than the other. For this reason, the estimation of the
over the selected frequency range, and the average ratio “negative” underlying linear system is considered more
between real and imaginary parts E[R/I] is also shown. reliable and so is the identification of the nonlinearity. The
Shock and Vibration 11

Table 2: Identified coefficients with NSI for the two reference positions.
Coefficient k3 Coefficient k􏽥2 Coefficient cnl
Reference position
Value (N/m3 ) E[R/I] Value (N/m2 ) E[R/I] Value (Ns2 /m2 ) E[R/I]
z∗− 7.35 · 105 32 −6.45 · 104 34 −4.76 49
z∗+ 7.86 · 105 29 5.81 · 104 29 −4.39 37

20

5
10

D (N)
K (N)

0
–10

–20 –5

–30 –1 0 1 2
–0.04 –0.02 0 0.02 0.04
z (m) ˙z (m/s)

(a) (b)

Figure 22: Estimation of the restoring force K(z) (a) and the damping force D(z)
_ (b). Gray dots: sliced restoring surface; orange line: RFS
estimation (only in a); blue line: NSI estimation.

0.06
0.04
x (m)

0.02
0
–0.02
0 30 60 24 25 26
Time (s) Time (s)
(a)

–100
X (dB)

–150

5 10 15 20 25 30 35 40
Frequency (Hz)
(b)

Figure 23: Validation of the nonlinear identification in the time domain (a) and frequency domain (b). Black line: measured output; blue
line: simulated output; red dots: residual with the measured output spectrum in dB scales (ref. 1 m2/Hz).

chosen final values of the coefficients are then the ones The agreement between RFS and NSI estimation of
related to the negative reference position. K(z) is very high so that the two curves in Figure 22(a)
These coefficients can then be compared with the ones appear to be almost overlapped, and the percentage devi-
estimated by the RFS method, by looking at the identified ation is around 1%. As for the damping force D(z), _ it is
restoring force. The comparison is reported in Figure 22, difficult to evaluate the goodness of the estimation due to the
where the damping force estimation is also shown. high dispersion of the points. However, at least for the linear
12 Shock and Vibration

_ it is possible to see that the slope of the

part (for small |z|), [3] C.-M. Lee, V. N. Goverdovskiy, and A. I. Temnikov, “Design
NSI estimation and the data points agree. of springs with “negative” stiffness to improve vehicle driver
Finally, the identified nonlinear state-space model can vibration isolation,” Journal of Sound and Vibration, vol. 302,
be used for reproducing the output time series, given the no. 4-5, pp. 865–874, 2007.
input force. Calling the simulated output xNSI , a com- [4] T. D. Le and K. K. Ahn, “A vibration isolation system in low
parison with the measured one can be carried out to frequency excitation region using negative stiffness structure
validate once more the identification. The validation set for vehicle seat,” Journal of Sound and Vibration, vol. 330,
corresponds to the last 60 seconds of the acquisition, and no. 26, pp. 6311–6335, 2011.
[5] H. Iemura and M. H. Pradono, “Advances in the development
the results are reported in Figure 23 as simulated time
of pseudo-negative-stiffness dampers for seismic response
history against its residual with the measured one in time
control,” Structural Control and Health Monitoring, vol. 16,
and frequency domains.
pp. 784–799, 2009.
The identified state-space model is capable of catching [6] A. A. Sarlis, D. T. R. Pasala, M. C. Constantinou,
the cross-well motion with a very good accuracy, providing a A. M. Reinhorn, S. Nagarajaiah, and D. P. Taylor, “Negative
percentage RMS deviation from the measurement of ap- stiffness device for seismic protection of structures,” Journal of
proximately 8%. Structural Engineering, vol. 139, no. 7, pp. 1124–1133, 2013.
[7] D. W. Jordan and P. Smith, Nonlinear Ordinary Differential
5. Conclusions Equations, Oxford University Press, New York, NY, USA, 4th
edition, 2007.
A negative stiffness oscillator has been characterized ex- [8] F. C. Moon and P. J. Holmes, “A magnetoelastic strange
perimentally to exploit its dynamical behavior. A variety of attractor,” Journal of Sound and Vibration, vol. 65, no. 2,
different kinds of motions can be obtained because of the pp. 275–296, 1979.
bistable nature of the device, from in-well to cross-well [9] I. Kovacic and M. J. Brennan, The Duffing Equation: Nonlinear
oscillations, including chaotic motion. These dynamical Oscillators and their Behaviour, John Wiley and Sons,
behaviors have been confirmed by the experimental ob- Hoboken, NJ, USA, 1st edition, 2011.
[10] W.-Y. Tseng and J. Dugundji, “Nonlinear vibrations of a
servations, retrieving linear, nonlinear, and chaotic motions.
buckled beam under harmonic excitation,” Journal of Applied
Eventually, the parameters defining the nonlinearity have
Mechanics, vol. 38, no. 2, pp. 467–476, 1971.
been recognized via nonlinear system identification, [11] C. S. Wojnar and D. M. Kochmann, “A negative-stiffness
adopting two different methods. A first guess has been phase in elastic composites can produce stable extreme ef-
obtained using the restoring force surface method, whose fective dynamic but not static stiffness,” Philosophical Mag-
implementation allows to easily visualize the nonlinear azine, vol. 94, no. 6, pp. 532–555, 2014.
behavior and the asymmetric double-well potential of the [12] D. Anastasio, A. Fasana, L. Garibaldi, and S. Marchesiello,
system. The final identification has been performed using the “Analytical investigation of railway overhead contact wire
nonlinear subspace identification (NSI) method with a dynamics and comparison with experimental results,” Me-
cross-well random measurement. NSI proved to be a robust chanical Systems and Signal Processing, vol. 116, pp. 277–292,
technique, capable of handling a very strong nonlinear 2019.
behavior. The nonlinear coefficients have been eventually [13] S. Marchesiello and L. Garibaldi, “A time domain approach
estimated, together with the linear modal parameters as- for identifying nonlinear vibrating structures by subspace
sociated with small oscillations around the two equilibrium methods,” Mechanical Systems and Signal Processing, vol. 22,
positions. no. 1, pp. 81–101, 2008.
[14] S. Marchesiello and E. Gandino, “Identification of a duffing
oscillator under different types of excitation,” Mathematical
Data Availability Problems in Engineering, vol. 2010, Article ID 695025,
15 pages, 2010.
The data used to support the findings of this study are
[15] S. Marchesiello, A. Fasana, and L. Garibaldi, “Modal con-
available from the corresponding author upon request. tributions and effects of spurious poles in nonlinear subspace
identification,” Mechanical Systems and Signal Processing,
Conflicts of Interest vol. 74, pp. 111–132, 2016.
[16] S. F. Masri and T. K. Caughey, “A nonparametric identifi-
The authors declare that there are no conflicts of interest cation technique for nonlinear dynamic problems,” Journal of
regarding the publication of this paper. Applied Mechanics, vol. 46, no. 2, pp. 433–447, 1979.
[17] E. F. Crawley and K. J. O’donnell, “Force-state mapping
identification of nonlinear joints,” AIAA Journal, vol. 25,
References no. 7, pp. 1003–1010, 1987.
[1] R. S. Lakes, “Extreme damping in composite materials with a [18] S. W. Shaw, “The dynamics of a harmonically excited system
negative stiffness phase,” Physical Review Letters, vol. 86, having rigid amplitude constraints, part 2: chaotic motions
no. 13, pp. 2897–2900, 2001. and global bifurcations,” Journal of Applied Mechanics,
[2] I. Antoniadis, D. Chronopoulos, V. Spitas, and vol. 52, no. 2, pp. 459–464, 1985.
D. Koulocheris, “Hyper-damping properties of a stiff and [19] J. H. B. Deane and D. C. Hamill, “Instability, subharmonics,
stable linear oscillator with a negative stiffness element,” and chaos in power electronic systems,” IEEE Transactions on
Journal of Sound and Vibration, vol. 346, pp. 37–52, 2015. Power Electronics, vol. 5, no. 3, pp. 260–268, 1990.
Shock and Vibration 13

[20] S. H. Strogatz and R. F. Fox, “Nonlinear dynamics and chaos:

with applications to physics, biology, chemistry and engi-
neering,” Physics Today, vol. 48, no. 3, pp. 93-94, 1995.
[21] A. Wolf, J. B. Swift, H. L. Swinney, and J. A. Vastano, “De-
termining Lyapunov exponents from a time series,” Physica D:
Nonlinear Phenomena, vol. 16, no. 3, pp. 285–317, 1985.
[22] F. C. Moon, “Experiments on chaotic motions of a forced
nonlinear oscillator: strange attractors,” Journal of Applied
Mechanics, vol. 47, no. 3, pp. 638–644, 1980.