Chapter 2

From linear to nonlinear

2.1 Introduction
It is probable that all practical engineering structures are nonlinear to some extent,
the nonlinearity being caused by one, or a combination of, several factors such
as structural joints in which looseness or friction characteristics are present,
boundary conditions which impose variable stiffness constraints, materials that
are amplitude dependent or components such as shock absorbers, vibration
isolators, bearings, linkages or actuators whose dynamics are input dependent.
There is no unique approach to dealing with the problem of nonlinearity either
analytically or experimentally and thus we must be prepared to experiment with
several approaches in order to ascertain whether the structure can be classified as
linear or nonlinear. It would be particularly helpful if the techniques employed
in modal testing could be used to test nonlinear structures and it is certainly
essential that some form of test for linearity is carried out at the beginning of
any dynamic test as the majority of analysis procedures currently available are
based on linearity. If this principle is violated, errors may be introduced by the
data analysis. Thus the first step is to consider simple procedures that can be
employed to establish if the structure or component under test is linear. In the
following it is assumed that the structure is time invariant and stable.

2.2 Symptoms of nonlinearity

As stated at the end of the last chapter, many of the properties which hold for
linear structures or systems break down for nonlinear. This section discusses
some of the more important ones.

2.2.1 Definition of linearity—the principle of superposition

The principle of superposition discussed briefly in the first chapter is more than a
property of linear systems; in mathematical terms it actually defines what is linear

41
42 From linear to nonlinear

and what is not.

The principle of superposition can be applied statically or dynamically and
simply states that the total response of a linear structure to a set of simultaneous
inputs can be broken down into several experiments where each input is applied
individually and the output to each of these separate inputs can be summed to give
the total response.
This can be stated precisely as follows. If a system in an initial condition
S1 = fy1(0); y_1 (0)g responds to an input x 1 (t) with an output y1 (t) and in a
separate test an input x 2 (t) to the system initially in state S2 = fy2 (0); y_ 2 (0)g
produces an output y 2 (t) then superposition holds if and only if the input x 1 (t)+
x2 (t) to the system in initial state S3 = f y1 (0) + y2 (0); y_1 (0) + y_2 (0)g
results in the output y 1 (t) + y2 (t) for all constants ; , and all pairs of inputs
x1 (t); x2 (t).
Despite its fundamental nature, the principle offers limited prospects as a
test of linearity. The reason being that in order to establish linearity beyond
doubt, an infinity of tests is required spanning all , , x 1 (t) and x2 (t). This
is clearly impossible. However, to show nonlinearity without doubt, only one set
of ; ; x1 (t); x2 (t) which violate superposition are needed. In general practice
it may be more or less straightforward to establish such a set.
Figure 2.1 shows an example of the static application of the principle of
superposition to a uniform beam rigidly clamped at both ends subject to static
loading at its centre. It can be seen that superposition holds to a high degree of
approximation when the static deflections are small, i.e. less than the thickness
of the beam; however, as the applied load is increased, producing deflections
greater than the beam thickness, the principle of superposition is violated since the
applied loads F1 + F2 do not result in the sum of the deflections y 1 + y2 . What
is observed is a stiffness nonlinearity called a hardening stiffness which occurs
because the boundary conditions restrict the axial straining of the middle surface
(the neutral axis) of the beam as the lateral amplitude is increased. It is seen that
the rate of increase of the deflection begins to reduce as the load continues to
increase. The symmetry of the situation dictates that if the applied load direction
is reversed, the deflection characteristic will follow the same pattern resulting in
an odd nonlinear stiffness characteristic as shown in figure 2.2. (The defining
property of an odd function is that F ( y ) = F (y ).)
If the beam were pre-loaded, the static equilibrium point would not be
centred at (0; 0) as in figure 2.2 and the resulting force-deflection characteristic
would become a general function lacking symmetry as shown in figure 2.3.
This is a common example of a stiffness nonlinearity, occurring whenever
clamped beams or plates are subjected to flexural displacements which can be
considered large, i.e. well in excess of their thickness. The static analysis is
fairly straightforward and will be given here; a discussion of the dynamic case
is postponed until chapter 9.
Consider an encastré beam (a beam with fully clamped boundary conditions)
under a centrally applied static load (figure 2.4). The deflection shape, with
 Symptoms of nonlinearity 43
F
t

F y<t y>t
l
ea
r Id
F3 = F1+ F2 n ea
F3 Li

F2

y3 = y + y2
1
F1

y y y
1 2 3 Y
Figure 2.1. Example of the static application of the principle of superposition to a uniform
clamped–clamped beam showing that for static deflections in excess of the beam thickness
a ‘hardening’ stiffness is induced which violates the principle.

y >> t

-y y
y << t
F = k1 y + k3 y3

-F

Figure 2.2. The effect of reversing the applied load on the beam of figure 2.1: a symmetric
‘hardening’ stiffness nonlinearity.

the coordinates located at the mid-point of the beam, can be assumed to be

a polynomial which satisfies all the boundary conditions and the eigenvalue
44 From linear to nonlinear

F F

F
F = k1 y + k2 y 2+ k3 y 3
y y y
F, y

Figure 2.3. The result of pre-loading the beam in figure 2.1 is a general cubic form for the
stiffness, lacking the symmetry of figure 2.2.

L F
x

y
L1
dx
dy
∆ L1

Figure 2.4. An encastré (clamped–clamped) beam under a centrally applied static load
resulting in a change of length from L to L1 . The elemental length represents the axial
extension.

problem, i.e. an admissible function

!
ax2 bx4 cx6
y(x) = Y 1  L 2 +  L 4  L 6 +


: (2.1)
2 2 2
Using this assumed shape and by deriving the axial and flexural strain
energies, an expression for the lateral stiffness at the centre of the beam can be
found. If only the first three terms in the series are used with the appropriate
values for the constants, the expression for the deflection is
!
x2 x4 x6
y(x) = Y 1 2:15  L 2 + 1:30  L 4 0:15  L 6


(2.2)
2 2 2
and the flexural strain energy V F is found from
Z L=2  2 2
M 2 EI L=2
Z
dy
VF = dx = dx
x2
(2.3)
L=2 2 EI 2 L=2 d
 Symptoms of nonlinearity 45

to be
 4  2 !2
EIY 2 L=2
Z
L L 2
=  L 12 dx 4:3 + 15:6 x 4:5x4 (2.4)
2 0 2 2
so finally
Y2
VF = 98:9EI 3 : (2.5)
L
The strain energy due to the in-plane axial load is found from the expression
governing the axial extension,
" 2 # 21

d y

L1 = (dx2 + dy2 ) = dx 1 +
1
2 (2.6)
dx
i.e.
" # " #
1 dy 2 1 dy 4 1 dy 2
    

L1 = dx 1 + +


 dx 1 + :
2 dx 8 dx 2 dx
(2.7)
Therefore,
"  #
1 dy 2 dy 2
Z L=2
1 L=2
Z  
L1 = dx 1 + =L+ dx (2.8)
L=2 2 dx 2 L=2 dx
and
L, the change in axial length of the beam, is given by
 2
1 L=2
Z
dy
L1 L = dx : (2.9)
2 L=2 dx
Substituting for y (x) from equation (2.1) gives

Y2

L = 2:44 : (2.10)
L
Thus, the axial strain energy is

1 EA Y4
VA = (
L)2 = 2:98EA 3 : (2.11)
2 L L
From Lagrange’s equations, the stiffness terms are given by

@ EIY EAY 3
(VF + VA ) = 197:8 3 + 11:92 3 (2.12)
@Y L L
i.e. the linear elastic stiffness term is k 1 = 197:8EI=L3 and the nonlinear
hardening-stiffness term is k 3 = 11:92EAY 2 =L3 . (Note that the linear elastic
46 From linear to nonlinear

stiffness term k1 should be 192EI=L3 from simple bending theory. The small
error is due to limiting the assumed deflection polynomial to only three terms.)
In practise, because it is not possible to fully implement the principle of
superposition, i.e. spanning all the possibilities of inputs, simpler procedures
are employed. Since best practice in dynamic testing should always include
some check of linearity, it is important that easy-to-use procedures for detecting
nonlinearity are available. The most commonly used procedures are based on
harmonic distortion, homogeneity and reciprocity.

2.2.2 Harmonic distortion

Harmonic or waveform distortion is one of the clearest indicators of the
presence of nonlinearity. It is a straightforward consequence of the principle of
superposition. If the excitation to a linear system is a monoharmonic signal, i.e.
a sine or cosine wave of frequency ! , the response will be monoharmonic at the
same frequency (after any transients have died out). The proof is elementary 1 and
proceeds as follows.
Suppose x(t) = sin(!t) is the input to a linear system. First of all, it is
observed that x(t) ! y (t) implies that x_ (t) ! y_ (t) and x(t) ! y(t). This
is because superposition demands that

x(t +
t) x(t)

t
! y(t +

t)t y(t) (2.13)

and x_ (t) ! y_ (t) follows in the limit as
t ! 0. (Note that there is also an
implicit assumption of time invariance here, namely that x(t) ! y (t) implies
x(t +  ) ! y(t +  ) for any  .) Again, by superposition,
x1 (t) + !2 x2 (t) ! y1 (t) + !2 y2 (t) (2.14)

so taking x1 (t) = x
(t) and x2 (t) = x(t) gives
x(t) + !2 x(t) ! y(t) + !2 y(t): (2.15)

Now, as x(t) = sin(!t),

x(t) + !2 x(t) = 0: (2.16)
In the steady state, a zero input to a linear system results in a zero output. It
therefore follows from (2.15) that

y(t) + !2 y(t) = 0 (2.17)

and the general solution of this differential equation is

y(t) = A sin(!t ) (2.18)

1 The authors learnt this proof from Dr Hugh Goyder.
 Symptoms of nonlinearity 47

Displacement (m)

-1.5
Velocity (m/s)

-1.5
0.0
Acceleration (m/s/s)

-1.5
100.0
0.0
200.0
300.0
400.0
500.0
600.0
700.0
800.0
900.0
1000.0
1100.0
1200.0
1300.0
1400.0
1500.0
1600.0
1700.0
1800.0
1900.0
2000.0
2100.0
2200.0
2300.0
2400.0
2500.0
2600.0
2700.0
2800.0
2900.0
3000.0
3100.0
3200.0
3300.0
3400.0
3500.0
3600.0
3700.0
3800.0
3900.0
4000.0
4100.0
4200.0
4300.0
4400.0
4500.0
4600.0
4700.0
4800.0
4900.0
5000.0
5100.0
5200.0
5300.0
5400.0
5500.0
5600.0
5700.0
5800.0
5900.0
6000.0
6100.0
6200.0
6300.0
6400.0
6500.0
6600.0
6700.0
6800.0
6900.0
7000.0
7100.0
7200.0
7300.0
7400.0
7500.0
7600.0
7700.0
7800.0
7900.0
8000.0
Time (s)

Figure 2.5. Response signals from a nonlinear system showing clear distortion only on the
acceleration signal.

and this establishes the result. This proof is rather interesting as it only uses
the fact that x(t) satisfies a homogeneous linear differential equation to prove
the result. The implication is that any such function will not suffer distortion in
passing through a linear system.
It is not a corollary of this result that a sine-wave input to a nonlinear system
will not generally produce a sine-wave output; however, this is usually the case
and this is the basis of a simple and powerful test for nonlinearity as sine waves are
simple signals to generate in practice. The form of the distortion will be discussed
in chapter 3, it will be revealed that the change in form is due to the appearance
of higher harmonics in the response such as sin(3!t), sin(5! ) etc.
Distortion can be easily detected on an oscilloscope by observing the
input and output time response signals. Figures 2.5 and 2.6 show examples of
48 From linear to nonlinear

Input
Force
Signal

t
Figure 2.6. Distortion on the input force signal arising from vibration exciter misalignment
(the severe distortion is due to the exciter coil rubbing against the magnet).

harmonic waveform distortion where a sinusoidal excitation signal is warped due

to nonlinearity.
In figure 2.5 the output response from a nonlinear system is shown in terms
of the displacement, velocity and acceleration. The reason that the acceleration
is more distorted compared with the corresponding velocity and displacement is
easily explained. Let x(t) = sin(!t) be the input to the nonlinear system. As
previously stated, the output will generally (at least for weak nonlinear systems 2 )
be represented as a Fourier series composed of harmonics written as

y(t) = A1 sin(!t + 1 ) + A2 sin(2!t + 2 ) + A3 sin(3!t + 3 ) +


(2.19)

and the corresponding acceleration is

y(t) = !2 A1 sin(!t + 1 ) 4!2 B2 sin(2!t + 2 )

9!2B3 sin(3!t + 2 )


: (2.20)

Thus the nth output acceleration term is weighted by the factor n 2 compared
to the fundamental.
In figure 2.6 the signal represents the output of a force transducer during
a modal test. The distortion is due to shaker misalignment resulting in friction
between the armature of the shaker and the internal magnet—a nonlinearity.
If non-sinusoidal waveforms are used, such as band-limited random signals,
waveform distortion is generally impossible to detect and additional procedures
are required such as the coherence function described in section 2.5.2.

2 There are a number of opinions as to what constitutes weak nonlinearity. What it means here is
simply that the system does not undergo transition to chaos or show subharmonic generation.
 Symptoms of nonlinearity 49

2.2.3 Homogeneity and FRF distortion

This represents a restricted form of the principle of superposition. It is
undoubtedly the most common method in use for detecting the presence of
nonlinearity in dynamic testing. Homogeneity is said to hold if x(t) ! y (t)
implies x(t) ! y (t) for all . In essence, homogeneity is an indicator of
the system’s insensitivity to the magnitude of the input signal. For example, if
an input x1 (t) always produces an output y 1 (t), the ratio of output to input is
independent of the constant . The most striking consequence of this is in the
frequency domain. First, note that x(t) ! y (t) implies X (! ) ! Y (! ).
This means that if x(t) ! x(t),

Y (! )
H (! ) =
X (!)
! XY ((!!)) = H (!) (2.21)

and the FRF is invariant under changes of or effectively of the level of

excitation.
Because of this, the homogeneity test is usually applied in dynamic testing
to FRFs where the input levels are usually mapped over a range encompassing
typical operating levels. If the FRFs for different levels overlay, linearity is
assumed to hold. This is not infallible as there are some systems which are
nonlinear which nonetheless show homogeneity; the bilinear system discussed
in the next chapter is an example. The reason for this is that homogeneity is a
weaker condition than superposition.
An example of the application of a homogeneity test is shown in figure 2.7.
In this case band-limited random excitation has been used but, in principle, any
type of excitation signal may be employed. Although a visual check is often
sufficient to see if there are significant differences between FRFs, other metrics
can be used such as a measure of the mean-square error between the FRFs. The
exact form of the distortion in the FRF depends on the type of the nonlinearity,
some common types of FRF distortion produced by varying the level of excitation
are discussed in the following section.
One possible problem with the homogeneity test is caused by force ‘drop-
out’. Drop-out is a common phenomenon which occurs when forced vibration
tests are carried out during dynamic testing. As its description implies, this is
a reduction in the magnitude of the input force spectrum measured by the force
transducer and occurs in the vicinity of the resonant frequencies of the structure
under test. It is a result of the interaction between an electrodynamic exciter and
the structure [251]. A typical experimental force drop-out characteristic is shown
in figure 2.8.
If homogeneity is being used as a detection method for nonlinearity, force
drop-out can create misleading results. This is because the test for homogeneity
assumes that the input is persistently exciting, i.e. exercises the system equally
across the whole excitation bandwidth, whereas the effect of force drop-out is to
effectively notch-filter the input at the resonant frequency. This results in less
50 From linear to nonlinear

Figure 2.7. Application of a homogeneity test on a real structure. The close agreement of
the results is an indicator that the structure is linear within the excitation bounds used.

..
y

F
Accelerance FRF (log)

Force spectrum

FRF (accelerance)

Linear Frequency

Figure 2.8. A typical force ‘drop-out’ characteristic overlayed on the FRF of a cantilever
beam. Note the correspondence between the force spectrum minima and the FRF maxima.
 Symptoms of nonlinearity 51

Figure 2.9. Application of a reciprocity test on a real structure. The close agreement of
the results is an indicator that the structure is linear within the test bounds.

force communicated to the structure near resonance and the response may be
linearized. If a control system is employed to maintain a constant excitation force
spectrum, nonlinearity can easily be detected using homogeneity.

2.2.4 Reciprocity

Reciprocity is another important property which, if violated, can be used to detect

the presence of nonlinearity. For linearity to hold reciprocity is a necessary but
not a sufficient condition since some symmetrical nonlinear systems may exhibit
reciprocity but will not satisfy the principle of superposition. Reciprocity holds
if an output y B at a point B due to an input x A at a point A, gives a ratio
yB =xA numerically equal to that when the input and output points are reversed
giving yA =xB . It follows that if this condition holds, the FRFs for the processes
xA ! yB and xB ! yA are equal. This is the basis of the experimental test.
Figure 2.9 shows the results of a reciprocity test on a structure using band-
limited random excitation and the FRFs between two different points, A and B.
As in the homogeneity test, the difference is usually assessed by eye.
When employing reciprocity it is important to note that all the response
parameters must be the same, e.g. displacements or accelerations and all the inputs
must be forces. If reciprocity holds, then by definition the stiffness matrix of a
structure will be symmetric as will the FRF matrix.
52 From linear to nonlinear

2.3 Common types of nonlinearity

The most common types of nonlinearity encountered in dynamic testing are
those due to polynomial stiffness and damping, clearances, impacts, friction
and saturation effects. As one would expect, these nonlinearities are usually
amplitude, velocity and frequency dependent. However, it is usual to simplify
and idealize these in order that they can be incorporated into analysis, simulation
and prediction capabilities. Consider an SDOF oscillator with nonlinear damping
and stiffness terms:
my + fd(y_ ) + fs (y) = x(t): (2.22)
Figure 2.10 summarizes the most common types of nonlinearity in terms of their
idealized force against displacement or force against velocity characteristics.
Some examples of the effects of several of the nonlinearities shown in
figure 2.10 on the vibration characteristics of an isolated mode of vibration (in
this case considered as an SDOF) in the FRF subject to sinusoidal excitation
can be seen in figure 2.11. Here, the frequency response characteristics are
shown in terms of the Argand plane in the Nyquist plot) and the modulus of
the receptance FRF. Distortions are clearly seen which, if not recognized and
understood, may produce errors in the parameters which are extracted from
these FRFs if curve-fitting is used. A detailed discussion of the origin of these
distortions is postponed until chapter 3, only brief observations will be made here.
If a structure incorporates actuators, bearings, linkages or elastomeric elements,
these can act as localized nonlinearities whose characteristics may be represented
by one or more of those shown in figure 2.10.
It is instructive to consider each nonlinearity briefly in turn.

2.3.1 Cubic stiffness

In this case, the force displacement characteristic has the form,
fs (y) = ky + k3 y3 (2.23)
and k3 may be positive or negative. If k 3 > 0, one can see that at high
levels of excitation the restoring force will be greater than that expected from
the linear term alone. The extent of this excess will increase as the forcing
level increases and for this reason such systems are referred to as having a
hardening characteristic. Examples of such systems are clamped plates and
beams as discussed earlier. If k 3 < 0, the effective stiffness decreases as the
level of excitation increases and such systems are referred to as softening. Note
that softening cubic systems are unphysical in the sense that the restoring force
changes sign at a certain distance from equilibrium and begins to drive the system
to infinity. Systems with such characteristics are always found to have higher-
order polynomial terms in the stiffness with positive coefficients which dominate
at high levels and restore stability. Systems which appear to show softening cubic
behaviour over limited ranges include buckling beams plates.
 Common types of nonlinearity 53

Force

Force

Force
Displacement Displacement Displacement

Hardening Softening
Cubic Stiffness Bilinear Stiffness
Force

Force
Displacement Displacement

Saturation (or limiter) Clearance (or backlash)

Force

Force

Velocity Velocity

Coulomb Friction Nonlinear Damping

Figure 2.10. Idealized forms of simple structural nonlinearities.

The equation of motion of the SDOF oscillator with linear damping and
stiffness (2.23) is called Duffing’s equation [80],

my + cy_ + ky + k3 y3 = x(t) (2.24)

and this is the single most-studied equation in nonlinear science and engineering.
The reason for its ubiquity is that it is the simplest nonlinear oscillator which
possesses the odd symmetry which is characteristic of many physical systems.
Despite its simple structure, it is capable of showing almost all of the interesting
behaviours characteristic of general nonlinear systems. This equation will re-
occur many times in the following chapters.
The FRF distortion characteristic of these systems is shown in figures 2.11(b)
and (c). The most important point is that the resonant frequency shifts up for the
hardening system as the level of excitation is raised, this is consistent with the
54 From linear to nonlinear

Figure 2.11. SDOF system Nyquist and FRF (Bode) plot distortions for five types of
nonlinear element excited with a constant amplitude sinusoidal force; —— low level, – – –
high level.
 Common types of nonlinearity 55

increase in effective stiffness. As one might expect, the resonant frequency for
the softening system shifts down.

2.3.2 Bilinear stiffness or damping

In this case, the stiffness characteristic has the form,


fs (y) = kk1 y; y>0 (2.25)
2 y<0
y;
with a similar definition for bilinear damping. The most extreme example of a
bilinear system is the impact oscillator for which k 1 = 0 and k2 = 1; this
corresponds to a ball bouncing against a hard wall. Such systems can display
extremely complex behaviour indeed (see chapter 15 of [248]). One system
which approximates to a bilinear damping system is the standard automotive
damper or shock absorber which is designed to have different damping constants
in compression and rebound. Such systems are discussed in detail in chapters 7
and 9.
Figure 2.11 does not show the FRF distortion characteristic of this system
because it is one of the rare nonlinear systems which display homogeneity. (This
last remark is only true if the position of the change in stiffness is at the origin, if
it is offset by any degree, the system will fail to show homogeneity if the level of
excitation is taken sufficiently high.)

2.3.3 Piecewise linear stiffness

The form of the stiffness function in this case is

8
< k2 y + (k1 k2 )d; y > d
fs (y) =
:
k1 y; jyj < d (2.26)
k2 y (k1 k2 )d; y < d.
Two of the nonlinearities in figure 2.10 are special cases of this form. The
saturation or limiter nonlinearity has k 2 = 0 and the clearance or backlash
nonlinearity has k 1 = 0.
In aircraft ground vibration tests, nonlinearities of this type can arise from
assemblies such as pylon–store–wing assemblies or pre-loading bearing locations.
Figure 2.12 shows typical results from tests on an aircraft tail-fin where the
resonant frequency of the first two modes reduces as the input force level is
increased and then asymptotes to a constant value. Such results are typical of
pre-loaded backlash or clearance nonlinearities.
Typical FRF distortion is shown in figure 2.11(f ) for a hardening piecewise
linear characteristic (k 2 > k1 ).
56 From linear to nonlinear

21
.

Response Amplitude (g)

..

5
.. Frequency

Frequency (Hz)
.
.

20

4
.

Accel 04
.
.

3
.
Amplitude

19

2
..
. . .. .
. ..

1
18

0
0 6 12 18 24 30 36
Input Force (N)
79

8
.

Response Amplitude (g)

Frequency (Hz)

6
Frequency

Accel 00
.
78

.
4
.
. . Amplitude
. .
2
. .
.
77

0 5 10 15 20 25
Input Force (N)

Figure 2.12. Results from ground vibration tests on the tail-fin of an aircraft showing
significant variation in the resonant frequency with increasing excitation level. This was
traced to clearances in the mounting brackets.

2.3.4 Nonlinear damping

The most common form of polynomial damping is quadratic:

fd (y_ ) = c2 y_ jy_ j (2.27)

(where the absolute value term is to ensure that the force is always opposed to
the velocity). This type of damping occurs when fluid flows through an orifice or
around a slender member. The former situation is common in automotive dampers
and hydromounts, the latter occurs in the fluid loading of offshore structures. The
fundamental equation of fluid loading is Morison’s equation [192],

F (t) = c1 u_ (t) + c2 u(t)ju(t)j (2.28)

where F is the force on the member and u is the velocity of the flow. This system
will be considered in some detail in later chapters.
 Nonlinearity in the measurement chain 57

The effect of increasing excitation level is to increase the effective damping

as shown in figure 2.11(d).

2.3.5 Coulomb friction

This type of damping has characteristic,

fd (y_ ) = cF sgn(y_ ) (2.29)

as shown in figure 2.10. This type of nonlinearity is common in any situation with
interfacial motion. It is particularly prevalent in demountable structures such as
grandstands. The conditions of constant assembly and disassembly are suitable
for creating interfaces which allow motion. In this sort of structure friction will
often occur in tandem with clearance nonlinearities. It is unusual here in the sense
that it is most evident at low levels of excitation, where in extreme cases, stick–
slip motion can occur. At higher levels of excitation, the friction ‘breaks out’
and the system will behave nominally linearly. The characteristic FRF distortion
(figure 2.11(e)) is the reverse of the quadratic damping case, with the higher
damping at low excitation.

2.4 Nonlinearity in the measurement chain

It is not uncommon for nonlinearity to be unintentionally introduced in the
test programme through insufficient checks on the test set-up and/or the
instrumentation used. There are several common sources of nonlinearity whose
effects can be minimized at the outset of a test programme and consideration
should be given to simple visual and acoustic inspection procedures (listening for
rattles etc) before the full test commences.
The principal sources of nonlinearity arising from insufficient care in the test
set-up are:
 misalignment
 exciter problems
 looseness
 pre-loads
 cable rattle
 overloads/offset loads
 temperature effects
 impedance mismatching
 poor transducer mounting

Most of these problems are detectable in the sense that they nearly all cause
waveform distortion of some form or other. Unless one observes the actual
input and output signals periodically during testing it is impossible to know
whether or not any problems are occurring. Although tests frequently involve the
58 From linear to nonlinear

measurement of FRFs or spectra it is strongly recommended that a visual check

is maintained of the individual drive/excitation and response voltage signals. This
can be done very simply by the use of an oscilloscope.
In modal testing it is usual to use a force transducer (or transducers in
the case of multi-point testing) as the reference input signal. Under such
circumstances it is strongly recommended that this signal is continuously (or at
least periodically) monitored on an oscilloscope. This is particularly important
as harmonic distortion of the force excitation signal is not uncommon, often due
to shaker misalignment or ‘force drop-out’ at resonance. Distortion can create
errors in the measured FRF which may not be immediately apparent and it is very
important to ensure that the force input signal is not distorted.
Usually in dynamic testing one may have the choice of observing the
waveform in terms of displacement, velocity or acceleration. For a linear system
in which no distortion of the signal occurs it makes little difference which variable
is used. However, when nonlinearity is present this generally results in harmonic
distortion. As discussed earlier in this chapter, under sinusoidal excitation,
harmonic distortion is much easier to observe when acceleration is measured.
Thus it is recommended that during testing with a sine wave, a simple test of the
quality of the output waveform is to observe it on an oscilloscope in terms of the
acceleration response. Any distortion or noise present will be more easily visible.
Due to their nature, waveform distortion in random signals is more difficult
to observe using an oscilloscope than with a sine-wave input. However, it is still
recommended that such signals are observed on an oscilloscope during testing
since the effect of extreme nonlinearities such as clipping of the waveforms can
easily be seen.
The first two problems previously itemized will be discussed in a little more
detail.

2.4.1 Misalignment

This problem often occurs when electrodynamic exciters are used to excite
structures in modal testing. If an exciter is connected directly to a structure
then the motion of the structure can impose bending moments and side loads
on the exciter armature and coil assembly resulting in misalignment, i.e. the coil
rubbing against the internal magnet of the exciter. Misalignment can be detected
by using a force transducer between the exciter and the test structure, the output
of which should be observed on an oscilloscope. If a sine wave is injected into
the structure, misalignment will produce a distorted force signal which, if severe,
may appear as shown in figure 2.6. If neglected, this can create significant damage
to the vibration exciter coil, resulting in a reduction in the quality of the FRFs and
eventual failure of the exciter. To minimize this effect it is recommended that a
‘stinger’ or ‘drive-rod’ is used between the exciter and the test structure described
in [87].
 Two classical means of indicating nonlinearity 59

2.4.2 Vibration exciter problems

Force drop-out was briefly mentioned in section 2.2.3. When electrodynamic

vibration exciters are employed to excite structures, the actual force that is
applied is the reaction force between the exciter and the structure under test. The
magnitude and phase of the reaction force depends upon the characteristics of the
structure and the exciter. It is frequently (but mistakenly) thought that if a force
transducer is located between the exciter and the structure then one can forget
about the exciter, i.e. it is outside the measurement chain. In fact, the quality
of the actual force applied to the structure, namely the reaction force, is very
dependent upon the relationship between the exciter and the structure under test.
Detailed theory shows that, in order to apply a constant-magnitude force to a
structure as the frequency is varied, it would be necessary to use an exciter whose
armature mass and spider stiffness are negligible. This can only be achieved using
special exciters such as non-contact electromagnetic devices or electrodynamic
exciters based on magnets which are aligned with lightweight armatures that are
connected to the structure, there then being no spider stiffness involved.
When a sine wave is used as the excitation signal and the force transducer
signal is observed on an oscilloscope, within the resonance region the waveform
may appear harmonically distorted and very small in magnitude. This is
particularly evident when testing lightly damped structures. The harmonic
distortion in the force signal is due to the fact that at resonance the force supplied
by the exciter has merely to overcome the structural damping. If this is small
(as is often the case), the voltage level representing the force signal becomes
very small in relation to the magnitude of the nonlinear harmonics present in the
exciter. These nonlinearities are created when the structure and hence armature
of the exciter undergoes large amplitudes of vibration (at resonance) and begins
to move into the non-uniform flux field in the exciter. This non-uniform flux field
produces strong second harmonics of the excitation frequency which distorts the
fundamental force signal.

2.5 Two classical means of indicating nonlinearity

It is perhaps facetious to use the term ‘classical’ here as the two techniques
discussed are certainly very recent in historical terms. The reason for the
terminology is that they were both devised early in the development of modal
testing, many years before most of the techniques discussed in this book
were developed. This is not to say that their time is past—coherence, in
particular, is arguably the simplest test for nonlinearity available via mass-
produced instrumentation.
60 From linear to nonlinear

2.5.1 Use of FRF inspections—Nyquist plot distortions

FRFs can be visually inspected for the characteristic distortions which are
indicative of nonlinearity. In particular, the resonant regions of the FRFs will
be the most sensitive. In order to examine these regions in detail, the use of the
the Nyquist plot (i.e. imaginary versus real part of the FRF) is commonly used.
(If anti-resonances are present, they can also prove very sensitive to nonlinearity.)
The FRF is a complex quantity, i.e. it has both magnitude and phase, both of
which can be affected by nonlinearity. In some cases it is found that the magnitude
of the FRF is the most sensitive to the nonlinearity and in other cases it is the
phase. Although inspecting the FRF in terms of the gain and phase characteristics
separately embodies all the information, combining these into one plot, namely
the Nyquist plot, offers the quickest and most effective way of inspecting the FRF
for distortions.
The type of distortion which is introduced in the Nyquist plot depends upon
the type of nonlinearity present in the structure and on the excitation used, as
discussed elsewhere in this chapter. However, a simple rule to follow is that if
the FRF characteristics in the Nyquist plane differ significantly from a circular or
near-circular locus in the vicinity of the resonances then nonlinearity is a suspect.
Examples of common forms of Nyquist plot distortion as a result of structural
nonlinearity, obtained from numerical simulation using sinusoidal excitation, are
shown in figure 2.11. It is interesting to note that in the case of the non-dissipative
nonlinearities under low levels of excitation, e.g. the polynomial and piecewise
nonlinear responses, the Nyquist plot appears as a circular locus. However, by
inspecting the
! spacings (proportional to the change in phase) it is possible
to detect a phase distortion. When the input excitation level is increased to the
point at which the effect of the nonlinearity becomes severe enough to create the
‘jump’ phenomenon (discussed in more detail in the next chapter), the Nyquist
plot clearly shows this.
In the case of dissipative nonlinearities and also friction, the distortion in the
Nyquist plot is easily detected with appropriate excitation levels via the unique
characteristic shapes appearing which have been referred to as the ‘apples and
pears’ of FRFs.
An example of nonlinearity from an attached element is shown in figure 2.13
where a dynamic test was carried out on a cantilever beam structure which had a
hydraulic, passive, actuator connected between the beam and ground. Under low-
level sinusoidal excitation the friction in the actuator seals dominates the response
producing a distorted ‘pear-shaped’ FRF as shown in figure 2.13.
When the excitation level was increased by a factor of three (from a 2N to
a 6N peak), the FRF distortion changed to an oval shape. These changes in the
FRF can be attributed to the nonlinearity changing from a friction characteristic at
low input excitation levels to a nonlinear velocity-dependent characteristic such
as a quadratic damping effect.
It is relatively straightforward to demonstrate that such distortions occur
 Two classical means of indicating nonlinearity 61

Im (Fy )
-3 -2 -1 0 1 2 3
y
26.0
27.0 23.5 Hz Re (F )
23.4 Hz
24.3
25.0 24.1 23.9 22.0 Hz
26.0
-2
24.0 A 23.6
23.9
-3
24.5 23.7
23.8 22.5
-4 B

-5
24.0

-6 C

-7 22.9
23.5

23.0
23.2 23.1
..
y
Curve A, F = 1.5N
Curve B, F = 2 N
Curve C, F = 5 N F cos ωt
Hydraulic passive
actuator

Figure 2.13. Nyquist plot distortions arising from a combination of seal friction
nonlinearity in the passive hydraulic actuator at low excitation levels and a velocity-squared
nonlinearity at higher excitation levels.

in the Argand plane when nonlinearity is present. Anticipating the theme of

the next chapter a little, consider the case of a simple oscillator, with structural
damping constant Æ and Coulomb friction of magnitude c F , given by the equation
of motion,
my + k(1 + iÆ)y + cF sgn(y_ ) = P ei!t: (2.30)
By using the method of harmonic balance (see chapter 3) the Coulomb
friction function can be represented by an equivalent structural damping constant
h , where
4cF
h = jY j (2.31)

where Y is the peak displacement. Thus equation (2.30) can be written as

my + k(1 + iÆ )y = P ei!t (2.32)

with
4cF
Æ = Æ + jY j: (2.33)

62 From linear to nonlinear

The solution to equation (2.32) can be written as

y(t) = Y ei!t with Y = jY jei (2.34)

i.e.  
Æ
jY j = Pk [(1 2 )2 + Æ2 ] ; tan  =
1
2
(1 2) (2.35)

where = !=!n. Substituting (2.33) in (2.35) gives the magnitude of the

response as

2 )2 + Æ2 g r2 (1 2 )2 
Ær + Pk f(1
1
2

jY j = (1 2 )2 + Æ2
(2.36)

and the phase as

r
[Æ + jY j ]
 = tan 1
(1 2 )
(2.37)

where r = 4cF =k .

A solution for jY j is only possible when r < P=k . If this condition is
violated, stick–slip motion occurs and the solution is invalid. When the vector
response is plotted in the Argand plane the loci change from a circular response
for r = 0, i.e. a linear system, to a distorted, pear-shaped response as r increases.
In the case of viscously damped systems, the substitution Æ = 2 can generally
be made without incurring any significant differences in the predicted results.

2.5.2 Coherence function

The coherence function is a spectrum and is usually used with random or impulse
excitation. It can provide a quick visual inspection of the quality of an FRF and,
in many cases, is a rapid indicator of the presence of nonlinearity in specific
frequency bands or resonance regions. It is arguably the most often-used test of
nonlinearity, by virtue of the fact that almost all commercial spectrum analysers
allow its calculation.
Before discussing nonlinearity, the coherence function will be derived for
linear systems subject to measurement noise on the output (figure 2.14). Such

x S y

Figure 2.14. Block diagram of a linear system with noise on the output signal.
 Two classical means of indicating nonlinearity 63

systems have time-domain equations of motion,

y(t) = S [x(t)] + m(t) (2.38)

where m(t) is the measurement noise. In the frequency domain,

Y (!) = H (!)X (!) + M (!): (2.39)

Multiplying this equation by its complex conjugate yields

Y Y = HXHX + HXM + HXM + MM (2.40)

and taking expectations gives 3

Syy (!) = jH (!)j2 Sxx (!) + H (!)Sxm(!) + H (!)Smx (!) + Smm(!): (2.41)

Now, if x and m are uncorrelated signals (unpredictable from each other), then
Swx(!) = Sxw (!) = 0 and equation (2.41) reduces to
Syy (!) = jH (!)j2 Sxx(!) + Smm (!) (2.42)

and a simple rearrangement gives

jH (!)j2 Sxx(!) = 1 Smm (!) : (2.43)

Syy (!) Syy (!)
The quantity on the right-hand side is the fraction of the output power, which
can be linearly correlated with the input. It is called the coherence function and
denoted 2 (! ). Now, as 2 (! ) and Smm (! )=Syy (! ) are both positive quantities,
it follows that
0 21 (2.44)
with 2 = 1 only if Smm (! ) = 0, i.e. if there is no measurement noise. The
coherence function therefore detects if there is noise in the output. In fact, it will
be shown later that 2 < 1 if there is noise anywhere in the measurement chain.
If the coherence is plotted as a function of ! , any departures from unity will be
readily identifiable. The coherence is usually expressed as

2 (! ) = jSyx (!)j2 : (2.45)

S (!)S (!)
yy xx
Note that all these quantities are easily computed by commercial spectrum
analysers designed to estimate H (! ); this is why coherence facilities are so
readily available in standard instrumentation.
3 It is assumed that the reader is familiar with the standard definitions of auto-spectra and cross-
spectra, e.g.
Syx (!) = E [Y X ]:
64 From linear to nonlinear

The coherence function also detects nonlinearity as previously promised.

The relationship between input and output spectra for nonlinear systems will be
shown in later chapters to have the form (for many systems)

Y (!) = H (!)X (!) + F [X (!)] (2.46)

where F is a rather complicated function, dependent on the nonlinearity.

Multiplying by Y and taking expectations gives

Syy (!) = jH (!)j2 Sxx (!) + H (!)Sxf (!) + H (!)Sfx (!) + Sff (!) (2.47)

where this time the cross-spectra S fx and Sxf will not necessarily vanish; in terms
of the coherence,
 
2 (!) = 1 2 Re H (!) Sxf (!) Sff (!)
(2.48)
Syy (!) Syy (!)
and the coherence will generally only be unity if f = 0, i.e. the system is linear.
The test is not infallible as unit coherence will also be observed for a nonlinear
system which satisfies

2 Re H (!)Sxf (!) = Sff (!) (2.49)

However, this is very unlikely.

Consider the Duffing oscillator of equation (2.24). If the level of excitation
is low, the response y will be small and y 3 will be negligible in comparison. In
this regime, the system will behave as a linear system and the coherence function
for input and output will be unity (figure 2.15). As the excitation is increased, the
nonlinear terms will begin to play a part and the coherence will drop (figure 2.16).
This type of situation will occur for all polynomial nonlinearities. However, if
one considers Coulomb friction, the opposite occurs. At high excitation, the
friction breaks out and a nominally linear response will be obtained and hence
unit coherence.
Note that the coherence is only meaningful if averages are taken. For a one-
shot measurement, a value of unity will always occur, i.e.

2 = Y XXY = 1: (2.50)
Y Y XX
Finally, it is important to stress again that in order to use the coherence
function for detecting nonlinearity it is necessary to realize that a reduction in
the level of coherency can be caused by a range of problems, such as noise on the
output and/or input signals which may in turn be due to incorrect gain settings on
amplifiers. Such obvious causes should be checked before structural nonlinearity
is suspected.
 Use of different types of excitation 65
20

|FRF|
dB

-60
0 Frequency 1kHz
1.0
Coherence

0
0 Frequency 1kHz

Figure 2.15. FRF gain and coherence plots for Duffing oscillator system given
by equation (2.24) subject to low-level random excitation showing almost ideal unit
coherence.

20

|FRF|
dB

-60
0 Frequency 1kHz
1.0
Coherence

0
0 Frequency 1kHz

Figure 2.16. The effect of increasing the excitation level for the Duffing oscillator of
figure 2.15, the coherence drops well below unity in the resonant region.

2.6 Use of different types of excitation

Nonlinear systems and structures respond in different ways to different types

of input excitation. This is an important observation in terms of detecting the
presence of nonlinearity or characterizing or quantifying it, some excitations will
be superior to others. In order to fully discuss this, it will be useful to consider a
concrete example of a nonlinear system. The one chosen is the Duffing oscillator
66 From linear to nonlinear

(with fairly arbitrary choices of parameter here),

y + 0:377y_ + 39:489y + 0:4y3 = x(t): (2.51)

The excitation, x(t) will be chosen to represent four common types used in
dynamic testing namely steady-state sine, impact, rapid sine sweep (chirp) and
random excitation.

2.6.1 Steady-state sine excitation

It is well known that the use of sinusoidal excitation usually produces the most
vivid effects from nonlinear systems. For example, a system governed by a
polynomial stiffness function can exhibit strong nonlinear effects in the FRF
such as bifurcations (the jump phenomenon) where the magnitude of the FRF
can suddenly reduce or increase. With stepped sinusoidal excitation, all the input
energy is concentrated at the frequency of excitation and it is relatively simple, via
integration, to eliminate noise and harmonics in the response signal (a standard
feature on commercial frequency response function analysers).
As such, the signal-to-noise ratio is very good compared with random or
transient excitation methods, an important requirement in all dynamic testing
scenarios, and the result is a well-defined FRF with distortions arising from
nonlinearity being very clear, particularly when a constant magnitude force
excitation is used.
It should be remembered that one of the drawbacks of using stepped sine
excitation methods is that they are slow compared with transient or random input
excitation methods. This is because at each stepped frequency increment, time is
required for the response to attain a steady-state condition (typically 1–2 s) before
the FRF at that frequency is determined. However, this is usually a secondary
factor compared with the importance of obtaining high-quality FRFs.
Consider figure 2.17(a). This FRF was obtained using steady-state
sinusoidal excitation. At each frequency step a force was applied consisting
of a constant amplitude sinewave. The displacement response was allowed to
reach a steady-state condition and the amplitude and phase at the excitation
frequency in the response were determined. The modulus of the ratio of the
amplitude to the force at each frequency increment constitutes the modulus of
the FRF (see chapter 1) shown in figure 2.17(a). The same (constant) amplitude
of force was chosen for each frequency and this amplitude was selected so that
the displacement of the system would be similar for all the excitation methods
studied here. The FRF was obtained by stepping the frequency of excitation
from 0.4 to 1.6 Hz (curve a–b–c–d) and then down from 1.6 Hz (curve d–c–e–
a). As previously discussed, the distortion of the FRF from the usual linear form
is considerable. The discontinuity observable in the curve will be discussed in
considerable detail in chapter 3.
 Use of different types of excitation 67
0.5 0.5
(a) e b (b)

c
a d

0.0 0.0
0.4 1.0 1.6 0.4 1.0 1.6
Frequency (Hz) Frequency (Hz)

0.5 0.5
(c) (d)

0.0 0.0
0.4 1.0 1.6 0.4 1.0 1.6
Frequency (Hz) Frequency (Hz)

Figure 2.17. Measurement of the FRF of a single degree-of-freedom nonlinear oscillator

with polynomial stiffness subject to different types of oscillation signals: (a) sinusoidal
input; (b) pulse input; (c) rapid sweep (chirp) input; (d) random input.

2.6.2 Impact excitation

The most well-known excitation method for measuring FRFs is the impact
method. Its popularity lies in its simplicity and speed. Impact testing produces
responses with high crest factors (ratio of the peak to the rms value). This property
can assist in nonlinearity being excited and hence observed in the FRFs and their
corresponding coherence functions, usually producing distortions in the FRFs
opposite to those obtained from sinusoidal excitation. The use of impact testing
methods however, suffers from the same problems as those of random excitation,
namely that the input is a broad spectrum and the energy associated with an
individual frequency is small, thus it is much more difficult to excite structural
nonlinearity. Impact is a form of transient excitation.
The FRF in figure 2.17(b) was obtained by applying the force as a very
short impact (a pulse). In practice pulses or impacts of the type chosen are
often obtained by using an instrumented hammer to excite the structure. This
makes the method extremely attractive for in situ testing. The FRF is obtained
68 From linear to nonlinear

by dividing the Fourier transform of the response by the Fourier transform of the
force. Averaging is usually carried out and this means that a coherence function
can be estimated. The pulse used here was selected so that the maximum value
of the response in the time domain was similar to the resonant amplitude from
the sine-wave test of the last section. The results in figure 2.17(b) confirm the
earlier remarks in that a completely different FRF is obtained to that using sine
excitation.

2.6.3 Chirp excitation

A second form of transient excitation commonly used for measuring FRFs is chirp
excitation. This form of excitation can be effective in detecting nonlinearity and
combines the attraction of being relatively fast with an equal level of input power
across a defined frequency range. Chirp excitation can be linear or nonlinear
where the nonlinear chirp signal can be designed to have a specific input power
spectrum that can vary within a given frequency range [265]. The simplest form
of chirp has a linear sweep characteristic so the signal takes the form

x(t) = X sin( t + t2 ) (2.52)

where and are chosen to give appropriate start and end frequencies. At any
given time, the instantaneous frequency of the signal is
d
!(t) = ( t + t2 ) = + 2 t: (2.53)
dt
As one might imagine, the response of a nonlinear system to such
a comparatively complex input may be quite complicated. The FRF in
figure 2.17(c) was obtained using a force consisting of a frequency sweep between
0 and 2 Hz in 50 s. (This sweep is rapid compared with the decay time of the
structure.) The FRF was once again determined from the ratio of the Fourier
transforms. The excitation level was selected so that the maximum displacement
in the time-domain was the same as before. The ‘split’ response in figure 2.17(c)
is due to the presence of the nonlinearity.

2.6.4 Random excitation

The FRF of a nonlinear structure obtained from random (usually band-limited)
excitation often appears undistorted due to the randomness of the amplitude and
phase of the excitation signal creating a ‘linearized’ or ‘averaged’ FRF.
Due to this linearization, the only way in which random excitation can assist
in detecting nonlinearity is for several tests to be carried out at different rms
levels of the input excitation (auto-spectrum of the input) and the resulting FRFs
overlaid to test for homogeneity. A word of warning here. Since the total power
in the input spectrum is spread over the band-limited frequency range used, the
ability to excite nonlinearities is significantly reduced compared with sinusoidal
 FRF estimators 69

excitation. In fact, experience has shown that it is often difficult to drive structures
into their nonlinear regimes with random excitation unless narrower-band signals
are used. This effect is also compounded by the fact that if an electrodynamic
exciter is being used to generate the FRFs in an open-loop configuration (no
feedback control for the force input) the force spectrum will suffer from force
drop-out in the resonant regions. This makes it even more difficult to drive
a structure into its nonlinear regimes and the measured FRFs corresponding to
different input spectrum levels may not show a marked difference. However, the
speed at which FRFs can be measured with random excitation and the combined
use of the coherence function makes random excitation a useful tool in many
practical situations for detecting nonlinearity.
Note that pseudo-random excitation is not recommended for use in
nonlinearity detection via FRF measurements. Pseudo-random excitation is
periodic and contains harmonically related discrete frequency components. These
discrete components can be converted (via the nonlinearity) into frequencies
which coincide with the harmonics in the input frequency. These will not average
out due to their periodic nature and hence the coherence function may appear
acceptable (close to unity) even though the FRF looks very ‘noisy’.
The FRF in figure 2.17(d) was obtained by using a random force and
determining spectral density functions associated with the force and response.
These were then used to estimate the FRF using

S (!)
H (!) = yx : (2.54)
Sxx(!)

2.6.5 Conclusions
These examples have been chosen to demonstrate how different answers can be
obtained from the same nonlinear model when the input excitation is changed. It
is interesting to note that the only FRF which one would recognize as ‘linear’ in
terms of its shape is the one shown in figure 2.17(d), due to a random excitation
input. This is because random excitation introduces a form of ‘linearization’ as
discussed in later chapters. As opposed to linear systems, the importance of
the type of excitation employed in numerical simulation or practical testing of
nonlinear systems has been demonstrated. Many of the detection and parameter
extraction methods for nonlinear systems, described later in this book, are
dependent upon the type of input used and will only provide reliable answers
under the correct excitation conditions.

2.7 FRF estimators

In the section on coherence, a linear system subject to measurement noise on the
output was studied. It was shown that the coherence dips below unity if such noise
is present. This is unfortunately not the only consequence of noise. The object of
70 From linear to nonlinear
x m

n y

u S v

Figure 2.18. Block diagram of a linear system with input and output measurement noise.

the current section is to show that noise also leads to erroneous or biased estimates
of the FRF when random excitation is used via equation (2.54).
This time a general system will be assumed which has noise on both input
and output (figure 2.18). The (unknown) clean input is denoted u(t) and after
the addition of (unknown) noise n(t), gives the measured input x(t). Similarly,
the unknown clean output v (t) is corrupted by noise m(t) to give the measured
output y (t). It is assumed that m(t), n(t) and x(t) are pairwise uncorrelated. The
basic equations in the frequency domain are

X (!) = U (!) + N (!) (2.55)

and
Y (!) = H (!)U (!) + M (!): (2.56)
Multiplying (2.55) by X and taking expectations gives

Sxx(!) = Suu (!) + Snn (!): (2.57)

Multiplying (2.56) by X and taking expectations gives

Syx(!) = H (!)Suu (!) (2.58)

as Smx (! ) = 0. Taking the ratio of (2.58) and (2.57) yields

Syx(!) H (!)Suu (!) H (! )

= =  S (!)  : (2.59)
Sxx(!) Suu (!) + Snn (!) 1 + Snn
uu (! )

This means that the estimator S yx =Sxx —denoted H1 (! )—is only equal to
the correct FRF H (! ) if there is no noise on the input (S nn = 0). Further, as
Snn =Suu > 0, the estimator is always an underestimate, i.e. H 1 (!) < H (!) if
input noise is present. Note that the estimator is completely insensitive to noise
on the output.
Now, multiply (2.56) by Y and take expectations, the result is

Syy (!) = jH (!)j2 Suu (!) + Smm (!): (2.60)

 FRF estimators 71

Multiplying (2.55) by Y and averaging yields

Sxy (!) = H (!)Suu (!) (2.61)

and taking the ratio of (2.60) and (2.61) gives

 
Syy (!) S (!)
= H (!) 1 + mm (2.62)
Sxy (!) Suu (!)
and this means that the estimator S yy =Sxy —denoted by H 2 (! )—is only equal to
H (!) if there is no noise on the output (S mm = 0). Also, as Smm =Suu > 0,
the estimator is always an overestimate, i.e. H 2 (! ) > H (! ) if output noise is
present. The estimator is insensitive to noise on the input.
So if there is noise on the input only, one should always use H 2 : if there is
noise only on the output, one should use H 1 . If there is noise on both signals a
compromise is clearly needed. In fact, as H 1 is an underestimate and H 2 is an
overestimate, the sensible estimator would be somewhere in between. As one can
always interpolate between two numbers by taking the mean, a new estimator H 3
can be defined by taking the geometric mean of H 1 and H2 ,
s
p Smm (!) + Suu (!)
H3 (!) = H1 (!)H2 (!) = H (!) (2.63)
Snn (!) + Suu (!)
and this is the estimator of choice if both input and output are corrupted.
Note that a byproduct of this analysis is a general expression for the
coherence,

2 (!) = jSyx (!)j2 = 

1
 
Smm (!) (!) (2.64)
Syy (!)Sxx (!) 1+ 1 + SSnn
Svv (!) vv (! )

from which it follows that 2 < 1 if either input or output noise is present. It also
follows from (2.64), (2.62) and (2.59) that 2 = H1 =H2 or

H (!)
H2 (!) = 21 (2.65)
(! )
so the three quantities are not independent.
As the effect of nonlinearity on the FRF is different to that of input noise or
output noise acting alone, one might suspect that H 3 is the best estimator for use
with nonlinear systems. In fact it is shown in [232] that H 3 is the best estimator
for nonlinear systems in the sense that, of the three estimators, given an input
density Sxx , H3 gives the best estimate of S yy via Syy = jH j2 Sxx . This is
a useful property if the object of estimating the FRF is to produce an effective
linearized model by curve-fitting.
72 From linear to nonlinear

2.8 Equivalent linearization

As observed in the last chapter, modal analysis is an extremely powerful theory of
linear systems. It is so effective in that restricted area that one might be tempted
to apply the procedures of modal analysis directly to nonlinear systems without
modification. In this situation, the curve-fitting algorithms used will associate a
linear system with each FRF—in some sense the linear system which explains it
best. In the case of a SDOF system, one might find the equivalent linear FRF

1
Heq(!) =
meq!2 + iceq! + keq
(2.66)

which approximates most closely that of the nonlinear system. In the time domain
this implies a best linear model of the form

meqy + ceq y_ + keq y = x(t) (2.67)

and such a model is called a linearization. As the nonlinear system FRF will
usually change its shape as the level of excitation is changed, any linearization
is only valid for a given excitation level. Also, because the form of the FRF is
a function of the type of excitation as discussed in section 2.6, different forcing
types of nominally the same amplitude will require different linearizations. These
are clear limitations.
In the next chapter, linearizations based on FRFs from harmonic forcing
will be derived. In this section, linearizations based on random excitation will
be discussed. These are arguably more fundamental because, as discussed in
section 2.6, random excitation is the only excitation which generates nonlinear
systems FRFs which look like linear system FRFs.

2.8.1 Theory
The basic theory presented here does not proceed via the FRFs, one operates
directly on the equations of motion. The technique—equivalent or more
accurately statistical linearization—dates back to the fundamental work of
Caughey [54]. The following discussion is limited to SDOF systems; however,
this is not a fundamental restriction of the method 4.
Given a general SDOF nonlinear system,

my + f (y; y_ ) = x(t) (2.68)

one seeks an equivalent linear system of the form (2.67). As the excitation
is random, an apparently sensible strategy would be to minimize the average
difference between the nonlinear force and the linear system (it will be assumed
4 The following analysis makes rather extensive use of basic probability theory, the reader who is
unfamiliar with this can consult appendix A.
 Equivalent linearization 73

that the apparent mass is unchanged, i.e. m eq = m), i.e. find the ceq and keq
which minimize

J1 (y; ceq; keq ) = E [f (y; y_ ) ceq y_ keqy]: (2.69)

In fact this is not sensible as the differences will generally be a mixture of negative
and positive and could still average to zero for a wildly inappropriate system. The
correct strategy is to minimize the expectation of the squared differences, i.e.

J2 (y; ceq; keq ) = E [(f (y; y_ ) ceqy_ keqy)2 ] (2.70)

or

J2 (y; ceq ; keq) = E [(f (y; y_ )2 + c2eqy_ 2 + keq

2 y2 2f (y; y_ )ceqy_
2f (y; y_ )keq y + 2ceqkeqyy_ ]: (2.71)

Now, using elementary calculus, the values of c eq and keq which minimize
(2.71) are those which satisfy the equations

@J2 @J2
= = 0: (2.72)
@ceq @keq
The first of these yields

E [ceqy_ 2 yf
_ (y; y_ ) + keq yy_ ] = ceqE [y_ 2 ] E [yf
_ (y; y_ )] + keqE [yy_ ] = 0 (2.73)

and the second

E [keqy2 yf (y; y_ ) + ceqyy_ ] = keq E [y2 ] E [yf (y; y_ )] + ceqE [yy_ ] = 0 (2.74)

after using the linearity of the expectation operator. Now, it is a basic theorem
of stochastic processes E [y y_ ] = 0 for a wide range of processes 5 . With this
assumption, (2.73) and that (2.74) become

E [yf
_ (y; y_ )]
ceq =
E [y_ 2 ]
(2.75)

and
E [yf (y; y_ )]
keq =
E [y2 ]
(2.76)
5 The proof is elementary and depends on the processes being stationary, i.e. that the statistical
()
moments of x t , mean, variance etc do not vary with time. With this assumption

dy2 = 0 = d E [y2 ] = E  dy2  = 2E [yy_ ]:

dt dt dt
74 From linear to nonlinear

and all that remains is to evaluate the expectations. Unfortunately this turns out
to be non-trivial. The expectation of a function of random variables like f (y; y_ )
is given by Z 1 1
Z
E [f (y; y_ )] = dy dy_ p(y; y_ )f (y; y_ ) (2.77)
1 1
where p(y; y_ ) is the probability density function (PDF) for the processes y and y_ .
The problem is that as the PDF of the response is not known for general nonlinear
systems, estimating it presents formidable problems of its own. The solution to
this problem is to approximate p(y; y_ ) by p eq (y; y_ )—the PDF of the equivalent
linear system (2.67); this still requires a little thought. The fact that comes to the
rescue is a basic theorem of random vibrations of linear systems [76], namely:
if the excitation to a linear system is a zero-mean Gaussian signal, then so is the
response. To say that x(t) is Gaussian zero-mean is to say that it has the PDF

x2
 
1
p(x) = p exp
2x2
(2.78)
2x
where x2 is the variance of the process x(t).
The theorem states that the PDFs of the responses are Gaussian also, so
!
1 2
yeq
peq(yeq ) = p exp
2y2eq
(2.79)
2yeq
and !
1 2
y_eq
peq(y_eq ) = p exp
2y2_eq
(2.80)
2y_eq
so the joint PDF is
!
1 2
yeq 2
y_eq
peq(yeq ; y_ eq) = peq(yeq )peq(y_eq ) = p exp :
2yeq y_eq 2y2eq 2y2_eq
(2.81)
In order to make use of these results it will be assumed from now on that
x(t) is zero-mean Gaussian.
Matters can be simplified further by assuming that the nonlinearity is
separable, i.e. the equation of motion takes the form

my + cy_ + ky + (y_ ) + (y) = x(t) (2.82)

in this case, f (y; y_ ) = cy_ + ky + (y_ ) + (y).

Equation (2.75) becomes

E [y_ (cy_ + ky + (y_ ) + (y))]

ceq =
E [y_ 2 ]
(2.83)
 Equivalent linearization 75

or, using the linearity of E ,

cE [y_ 2 ] + kE [yy
_ ] + E [y
_ (y_ )] + E [y_ (y))]
ceq = 2 (2.84)
E [y_ ]
which reduces to
E [y
_ (y_ )] + E [y_ (y))]
ceq = c +
E [y_ 2 ]
(2.85)

and a similar analysis based on (2.75) gives

E [y(y_ )] + E [y (y))]
keq = k + :
E [y2 ]
(2.86)

Now, consider the term E [y(y_ )] in (2.86). This is given by

Z 1Z 1
E [y(y_ )] = dy dy_ peq(y; y_ )y(y_ ) (2.87)
1 1
and because the PDF factors, i.e. p eq(yeq; y_eq ) = peq(yeq )peq(y_eq ), so does the
integral, hence,
Z 1 Z 1 
E [y(y_ )] = dy peq (y)y dy_ peq(y_ )(y_ ) = E [y]E [(y_ )] (2.88)
1 1
but the response is zero-mean Gaussian and therefore E [y ] = 0. It follows that
E [y(y_ )] = 0 and therefore (2.86) becomes
E [y (y))]
keq = k +
E [y2 ]
(2.89)

and a similar analysis for (2.85) yields

E [y
_ (y_ )]
ceq = c + :
E [y_ 2 ]
(2.90)

Now, assuming that the expectations are taken with respect to the linear
system PDFs ((2.79) and (2.80)), equation (2.90) becomes

Z 1 !
1 y_ 2
ceq = c + p 3 dy_ y
_ (y_ ) exp
2y2_eq
(2.91)
2y_eq 1

and (2.89) becomes

Z 1 !
1 y2
keq = k + p 3 dy y (y) exp
2y2eq
(2.92)
2yeq 1
76 From linear to nonlinear

which are the final forms required. Although it may now appear that the problem
has been reduced to the evaluation of integrals, unfortunately things are not quite
that simple. It remains to estimate the variances in the integrals. Now standard
theory (see [198]) gives
Z 1 Z 1 Sxx(!)
y2eq = d! jHeq (!)j2 Sxx (!) = d!
m!2 )2 + c2eq!2
(2.93)
1 1 (keq

and Z 1 !2 Sxx (!)

y2_eq = d!
1 (keq m!2 )2 + c2eq!2
(2.94)

and here lies the problem. Equation (2.92) expresses k eq in terms of the variance
y2eq and (2.93) expresses  y2eq in terms of keq . The result is a rather nasty pair of
coupled nonlinear algebraic equations which must be solved for k eq . The same is
true of ceq . In order to see how progress can be made, it is useful to consider a
concrete example.

2.8.2 Application to Duffing’s equation

The equation of interest is (2.24), so

(y) = k3 y3 (2.95)

and the expression for the effective stiffness, from (2.92) is

Z 1 !
k3 y2
keq = k + p 3 dy y4 exp :
2y2
(2.96)
2yeq 1 eq

In order to obtain a tractable expression for the variance from (2.93) it will
be assumed that x(t) is a white zero-mean Gaussian signal, i.e. S xx (! ) = P a
constant. It is a standard result then that [198]
Z 1 1 P
y2eq = P d! 2 2 2 2 = : (2.97)
1 (keq m! ) + ceq ! ckeq

This gives

k
Z 1 
ckeqy2

keq = k + p  3  3 dy y4 exp : (2.98)
1 2P
2 ckPeq
2
 Equivalent linearization 77

Now, making use of the result 6 ,

Z 1 3 2
1

dy y4 exp( ay2 ) = (2.99)

1 4a 52
gives
3k3 P
keq = k + (2.100)
ckeq
and the required k eq satisfies the quadratic equation

2 ckkeq 3k3 P = 0:
ckeq (2.101)

The desired root is (after a little algebra)

r
k k 12k3 P
keq = + 1+
ck2
(2.102)
2 2
which shows the expected behaviour, i.e. k eq increases if P or k3 increase. If k3 P
is small, the binomial approximation gives

3k3 P
keq = k + + O(k32 P 2 ): (2.103)
ck

6 Z
Integrals of the type
1
dy yn exp( ay2 )
1
occur fairly often in the equivalent linearization of polynomial nonlinearities. Fortunately, they are
fairly straightforward to evaluate. The following trick is used: it is well known that
Z 1 1

I= dy exp( ay2 ) =  12 :
1 a2
Differentiating with respect to the parameter a yields

dI Z 1 2
1

da = dy y2 exp( ay2 ) =
1 2a 32
and differentiating again, gives the result in (2.99)

d2 I = Z 1 dy y4 exp( ay2 ) =
3 1
2
:
da2 1 4a 5
2

Continuing this operation will give results for all integrals with n even. If n is odd, the sequence is
started with Z 1
I= dy y exp( ay2 )
1
but this is the integral of an odd function from 1 to 1 and it therefore vanishes. This means the
integrals for all odd n vanish.
78 From linear to nonlinear
0.0006

P=0 (Linear)
P=0.01
0.0005
P=0.02

0.0004
Magnitude FRF

0.0003

0.0002

0.0001

0.0000
50.0 70.0 90.0 110.0 130.0 150.0
Frequency (rad/s)

Figure 2.19. Linearized FRF of a Duffing oscillator for different levels of excitation.

To illustrate (2.102), the parameters m = 1, c = 20, k = 10 4 and

k3 = 5  109 were chosen for the Duffing oscillator. Figure 2.19 shows the
linear FRF with keq given by (2.102) with P = 0, 0.01 and 0.02. The values of k eq
found are respectively 10 000.0, 11 968.6 and 13 492.5, giving natural frequencies
of !n = 100:0, 109.4 and 116.2.
In order to validate this result, the linearized FRF for P = 0:02 is compared
to the FRF estimated from the full nonlinear system in figure 2.20. The agreement
is good, the underestimate of the FRF from the simulation is probably due to the
fact that the H1 estimator was used (see section 2.7).

2.8.3 Experimental approach

The problem with using (2.75) and (2.76) as the basis for an experimental method
is that they require one to know what f (y; y_ ) is. In practice it will be useful to
extract a linear model without knowing the details of the nonlinearity. Hagedorn
and Wallaschek [127, 262] have developed an effective experimental procedure
for doing precisely this.
Suppose the linear system (2.67) (with m eq = m) is assumed for the
 Equivalent linearization 79
0.0006
P=0.0 (Linear)
P=0.02 (Analytical)
P=0.02 (Numerical)
0.0005

0.0004
Magnitude FRF

0.0003

0.0002

0.0001

0.0000
50.0 70.0 90.0 110.0 130.0 150.0
Frequency (rad/s)

Figure 2.20. Comparison between the nonlinear system FRF and the theoretical FRF for
the linearized system.

experimental system. Multiplying (2.67) by y_ and taking expectations yields

mE [y_ y] + ceq E [y_ 2 ] + keq E [yy
_ ] = E [xy_ ]: (2.104)
Stationarity implies that E [y y_ ] = E [y_ y] = 0, so
E [xy_ ]
ceq = :
E [y_ 2 ]
(2.105)

(All processes are assumed zero-mean, the modification if they are not is fairly
trivial.) Similarly, multiply (2.67) by y and take expectations
mE [yy] + ceq E [yy_ ] + keq E [y2 ] = E [xy]: (2.106)
Now using stationarity and E [y y] = E [y_ 2 ] which follows from
d
E [yy_ ] = 0 = E [y_ 2 ] + E [yy] (2.107)
dt
yields
E [xy] + E [y_ 2 ]
keq =
E [y2 ]
(2.108)
80 From linear to nonlinear

and it follows that the equivalent stiffnesses and dampings can be obtained
experimentally if the signals x(t), y (t) and y_ (t) are measured. In fact, the
experimental approach to linearization is superior in the sense that the equivalent
damping and stiffness are unbiased. The theoretical procedure yields biased
values simply because the statistics of the linearized process are used in the
calculation in place of the true statistics of the nonlinear process.
This analysis concludes the chapter, rather neatly reversing the title by going
from nonlinear to linear.