Chapter 8

The Volterra series and higher-order

frequency response functions

8.1 The Volterra series

In the first chapter it was shown that linear systems admit dual time- and
frequency-domain characterizations 1:
Z 1
y(t) = d h( )x(t  ) (8.1)
1
and
Y (!) = H (!)X (!): (8.2)

All information about a single-input–single-output (SISO) system is encoded

in either the impulse response function h(t) or the frequency response function
(FRF) H (! ). The representation to be used in a given problem will usually
be dictated by the form of the answer required. In vibration problems, the
frequency-domain approach is usually adopted; displaying the FRF H (! ) shows
immediately those frequencies at which large outputs can be expected, i.e. peaks
in H (! ) corresponding to the system resonances.
Equations (8.1) and (8.2) are manifestly linear and therefore cannot hold for
arbitrary nonlinear systems; however, both admit a generalization. The extended
form of equation (8.1) was obtained in the early part of this century by Volterra
1 There are of course other characterizations. The set m; c; k fixes the behaviour of a linear SDOF
f g
() ()
system in just the same way as the functional forms h t and H ! do, and arguably provides a more
() ()
parsimonious means of doing so. However, the h t and H ! can provide a visual representation
that communicates the likely behaviour of the system in a way that the set of numbers does not.
f g
A more meaningful combination of the parameters, say m; ; !n conveys better understanding
to the average structural dynamicist. In the case of a SISO (single-input–single-output) continuous
system, all the representations involve infinite-dimensional sets and the distinction becomes otiose.
The authors would like to thank Dr Steve Gifford for discussion on this point.

377
378 The Volterra series and higher-order frequency response functions

[261]. It takes the form of an infinite series 2

y(t) = y1 (t) + y2 (t) + y3 (t) +


(8.3)

where
Z +1
y1 (t) = d h1 ( )x(t ) (8.4)
1
Z +1 Z +1
y2 (t) = d1 d2 h2 (1 ; 2 )x(t 1 )x(t 2 ) (8.5)
1 1
Z +1Z +1Z +1
y3 (t) = d1 d2 d3 h2 (1 ; 2 ; 3 )x(t 1 )x(t 2 )x(t 3 ):
1 1 1
(8.6)

The form of the general term is obvious from the previous statements. The
functions h1 ( ); h2 (1 ; 2 ); h3 (1 ; 2 ; 3 ); : : : hn (1 ; : : : ; n ); : : : are generaliza-
tions of the linear impulse response function and are usually referred to as Volterra
kernels. The use of the Volterra series in dynamics stems from the seminal paper
of Barrett [20], in which the series was applied to nonlinear differential equations
for the first time. One can think of the series as a generalization of the Taylor
series from functions to functionals. The expression (8.1) simply represents the
lowest-order truncation which is, of course, exact only for linear systems.
The derivation of the series is beyond the scope of this book, but heuristic
arguments can be found in [261, 25, 221]. Note that these kernels are not
forced to be symmetric in their arguments. In fact, any non-symmetric kernel
can be replaced by a symmetric version with impunity so that h 2 (1 ; 2 ) =
h2 (2 ; 1 ) etc. A formal proof is fairly straightforward; for simplicity, consider
the expression for y 2 (t):
Z +1 Z +1
y2 (t) = d1 d2 h2 (1 ; 2 )2 (1 ; 2 ; t) (8.7)
1 1
with the newly-defined

2 (1 ; 2 ; t) = x(t 1 )x(t 2 ) (8.8)

and note that  2 is manifestly symmetric in its arguments  1 and 2 .

2 The term weak nonlinearity has occasionally appeared in this book without a convincing definition.
The Volterra series allows at least a mathematically precise characterization if one defines a weak
nonlinear system as one that admits a representation in terms of a Volterra expansion. Because the
Volterra series is essentially a polynomial representation it cannot describe systems with multi-valued
responses. As a result, this definition of weak agrees with the more imprecise idea that strongly
nonlinear systems are those that exhibit the sort of bifurcations that result in subharmonic or chaotic
behaviour.
 The Volterra series 379

Assuming that h2 has no particular symmetries, it still has a canonical

decomposition into symmetric and antisymmetric parts:

h2 (1 ; 2 ) = hsym asym

2 (1 ; 2 ) + h2 (1 ; 2 ) (8.9)

where

hsym 1
2 (1 ; 2 ) = 2 (h2 (1 ; 2 ) + h2 (2 ; 1 ))
asym
h2 (1 ; 2 ) = 12 (h2 (1 ; 2 ) h2 (2 ; 1 )): (8.10)

Now, consider the contribution to y 2 (t) from the antisymmetric component

of the kernel:
Z +1 Z +1 asym
h2 (1 ; 2 )2 (1 ; 2 ; t) d1 d2 : (8.11)
1 1
Any (infinitesimal) contribution to this ‘summation’, say at  1 = v; 2 = w will
cancel with the corresponding contribution at  2 = v; 1 = w, as

hasym asym
2 (v; w)2 (v; w; t) = h2 (w; v)2 (w; v; t) (8.12)

and the overall integral will vanish. This is purely because of the ‘contraction’ or
summation against the symmetric quantity  2 (1 ; 2 ; t). Because hasym2 makes
no contribution to the quantity y 2 (t) it may be disregarded and the kernel h 2 can
be assumed to be symmetric. Essentially, the h 2 picks up all the symmetries of
the quantity 2 . This argument may be generalized to the kernel h n (1 ; : : : ; n ).
In general, for a symmetric kernel, h sym n is obtained by summing all of
the possible permutations of the argument, weighted by an inverse factor which
counts the terms. The following section describes a method of extracting the
kernel transforms directly, which automatically selects the symmetric kernel. This
method will be adopted throughout the remainder of the book. For this reason,
the identifying label ‘sym’ will be omitted on the understanding that all kernels
and kernel transforms are symmetric. For information about other conventions
for kernels, mainly the triangular form, the reader can consult [217].
As previously stated, there exists a dual frequency-domain representation
for nonlinear systems. The higher-order FRFs or Volterra kernel transforms
Hn (!1 ; : : : ; !n ), n = 1; : : : ; 1 are defined as the multi-dimensional Fourier
transforms of the kernels, i.e.
Z +1 Z +1
Hn (!1 ; : : : ; !n ) = ::: d1 : : : dn hn (1 ; : : : ; n )
1 1
 e i(Z!1 1 +


+!Zn n) (8.13)
1 + 1 + 1
hn (1 ; : : : ; n ) = ::: d!1 : : : d!n Hn (!1 ; : : : ; !n )
(2)n 1 1
 e+i(!1 1 +


+!n n) : (8.14)
380 The Volterra series and higher-order frequency response functions

It is a simple matter to show that symmetry of the kernels implies symmetry

of the kernel transforms so, for example, H 2 (!1 ; !2 ) = H2 (!2 ; !1 ).
It is then a straightforward matter to obtain the frequency-domain dual of the
expression (8.3)

Y (!) = Y1 (!) + Y2 (!) + Y3 (!) +


(8.15)

where

Y1 (!) = H1 (!)X (!) (8.16)

1 +1
Z
Y2 (!) = d! H (! ; ! !1 )X (!1 )X (! !1 )
2 1 1 2 1
(8.17)
Z +1 Z +1
1
Y3 (!) = d! d! H (! ; ! ; ! !1 !2 )
(2)2 1 1 1 2 3 1 2
 X (!1 )X (!2 )X (! !1 !2 ): (8.18)

The fundamental problem associated with the Volterra series is the

determination of either the kernels or the kernel transforms. This must be done
analytically if the equations of motion are known or numerically if time series are
given for the input and output processes. Section 8.3 will consider what happens
if the equations of motion are known, but first some motivation for use of the
series will be given.

8.2 An illustrative case study: characterization of a shock

absorber
Before proceeding to the main body of the theory of functional series, it is useful
to pause and consider what sort of problems one might apply them to. This section
illustrates their use on a real engineering system, namely a shock absorber. The
system considered will be a Monroe–McPherson strut; this is simply a coil spring
mounted over an automotive damper of the sort briefly discussed in the previous
chapter. It is characterized by a linear stiffness and a nonlinear damper. The work
described in this section was carried out by Dr Steve Cafferty and a much more
detailed discussion can be found in [50].
The experimental arrangement is shown in figure 7.51. The higher-order
FRFs are obtained by a harmonic testing approach. First the system is tested
without the coil spring and then with.
There are one or two interesting features of this problem: first, the force
in the shock absorber without the spring is a function of the velocity, not the
displacement, i.e. assuming linear viscous damping

f (t) = c1 y:_ (8.19)

 An illustrative case study: characterization of a shock absorber 381

The first-order FRF of interest is for the process y ! F and this is termed the
dynamic stiffness. A simple calculation yields
F (!)
= H1 (!) = ic1 ! (8.20)
Y (! )
and it follows that the dynamic stiffness varies linearly with ! and the gradient is
the linear damping coefficient. The presence of the imaginary term simply shows
that the displacement and force are in quadrature (90 Æ out of phase).
The first task is to establish H1 (! ). The experimental procedure is a standard
stepped-sine test. The system is excited by a displacement signal, a sinusoid
Y cos(!t) at a given frequency and the amplitude and phase of the force response
F cos(!t ) are recorded. The gain and phase of H 1 (!) are simply F=Y and
 as discussed in chapter 1.
In reality it is not quite as simple as this because the damper is nonlinear.
Assuming a polynomial expansion up to third order gives

f (t) = c1 y_ + c2 y_ 2 + c3 y_ 3 : (8.21)

Just as the first-order FRF is completely specified by c 1 , the higher-order

coefficients are encoded in the higher-order FRFs. Anticipating equation (8.32)
(in the form appropriate for velocity nonlinearity) gives, for a harmonic input
y(t) = ei t , y_ = i ei t
f (t) = H1 ( )i ei t H2 ( ; ) 2 ei2 t H3 ( ; ; )i 3 ei3 t +


(8.22)

and the higher-order FRFs are read off from (8.21):

H2 ( ; ) = c2 2 (8.23)
H3 ( ; ; ) = ic3 3 : (8.24)

The necessary experimental testing program follows from these formulae. In

order to find H 1 (! ) the standard linear stepped-sine procedure is used. In order
to find H2 (!; ! ), the amplitude and phase of the second harmonic is extracted,
i.e. the amplitude and phase of the component at 2! , to find H 3 (!; !; ! ), take the
amplitude and phase of the component at 3! . Note that this procedure only gives
values on the diagonal line in the frequency space where ! 1 = !2 =


= .
For this reason, the quantities are called the diagonal HFRFs.
The second subtlety alluded to earlier comes into play here. The earlier
argument assumes that the excitation is a pure harmonic e i t and this is impossible
in practice as it is a complex quantity. In reality, a cosinusoid is used which is
the sum of two harmonics, cos( ) = (e i t + e i t )=2. It will be shown later
that this means that the quantities measured in the test are not the pure FRFs
Hi ( ; : : : ; ). For example, the amplitude and phase of the component at 2 is
equal to H2 ( ; )+ higher-order terms involving H 4 , H6 etc. Fortunately, it can
be shown that the contamination of H 2 by the higher-order terms can be ignored
382 The Volterra series and higher-order frequency response functions

Figure 8.1. Principal diagonals of the first-, second- and third-order composite HFRFs for
the shock absorber.

if the amplitude of excitation is small enough. However, in order to eliminate

confusion, the measured FRFs will be termed composite FRFs and will be denoted
by s1 ( ), s2 ( ; ) etc. The s-subscript denotes that the FRFs are the result of
a sine test.
Figure 8.1 shows the first three measured diagonal HFRFs in terms of
amplitude and phase over the testing range 2–50 Hz for a low displacement
amplitude. The assumption of linear growth of the  s1 appears well-justified,
also the s2 and s3 curves have the required polynomial forms. Dividing  s1
by ! , s2 by ! 2 etc should yield constant values by the previous arguments
and figure 8.2 shows the results of these operations. At higher frequencies,
the HFRFs tend to the required constants; however, there is some distortion at
lower frequencies. The estimated coefficients are given in table 8.1. They show
the ‘softening’ behaviour in the damping which might well be expected from
 An illustrative case study: characterization of a shock absorber 383

Figure 8.2. Nonlinear damping values for the shock absorber estimated from the principal
diagonals.

Table 8.1. Parameter estimates for damping coefficients.

Coefficient Estimate Units
c1 1 600.0 Nsm 1
c2 832.0 N s2 m 2
c3 38 500.0 N s3 m 3

characteristic diagrams of the absorber like that in figure 7.50.

The testing procedure is not restricted to producing the diagonal elements
of the HFRF. For example, if a two-tone signal is used for the excitation by
combining frequencies 1 and 2 , then the amplitude and phase of the output
384 The Volterra series and higher-order frequency response functions

Figure 8.3. Principal quadrant of the second-order composite HFRF 2 (!1 ; !2 ) for the
shock absorber.

component at frequency 1 + 2 approximates the values for 2H 2 ( 1 ; 2 ).

Again it is assumed that the level of excitation is low enough for contributions
from H4 etc to be ignored. Strictly, the measured quantity is the composite FRF
s2 ( 1 ; 2 ). Similarly, if three frequencies are used to excite the system, the
amplitude and phase at the sum frequency approximates H 3 . Figures 8.3 and 8.4
show s2 (!1 ; !2 ) and s3 (!1 ; !2 ; !1 ) over the so-called ‘principal quadrants’.
(Note that it is not possible to plot  s3 in its full generality as it would require a
representation of four-dimensional space.) There is very little structure in these
plots, a very smooth variation of the HFRFs is observed with no resonances; this
is to be expected of course as there is no stiffness in the system. The theory
developed later in this chapter gives

H2 (!1 ; !2) = c2 !1 !2 (8.25)

 An illustrative case study: characterization of a shock absorber 385

Figure 8.4. Principal quadrant of the third-order composite HFRF 3 (!1 ; !2 ; !3 ) for the
shock absorber.

and
H3 (!1 ; !2; !3 ) = c3 !1 !2 !3 : (8.26)
The second series of tests were with the coil spring in place. These produced
slightly more structured HFRFs due to the internal resonances of the spring.
Using basic elasticity theory, a dynamic stiffness FRF for the spring alone was
estimated and is shown in figure 8.5, the resonances are clear. A monofrequency
test gave the results shown in figure 8.6 for the diagonal composite HFRFs, the
polynomial rise from the damping is combined with the spring resonances. A
bifrequency test yielded the  s2 and s3 shown in figures 8.7 and 8.8.
This section has shown how the HFRFs can be estimated using sine-testing
and also how they allow a parametric identification of the damping characteristics
of a shock absorber (although there are easier ways of obtaining estimates of the
c1 , c2 and c3 as discussed in the previous chapter). The figures showing the
386 The Volterra series and higher-order frequency response functions

Figure 8.5. Simulated FRF showing the coil spring’s first four resonant frequencies
calculated from spring theory.

HFRFs themselves actually yield important non-parametric information about

the system and the interpretation of the HFRFs is an important subject which
will be returned to later. In the meantime, it is important to show how the
theoretical HFRFs described earlier were obtained, and this forms the subject
of the following section.

8.3 Harmonic probing of the Volterra series

The subject of this section is a direct method of determining the higher-order FRFs
for a system given the equations of motion. The method of harmonic probing
was introduced in [22] specifically for systems with continuous-time equations of
motion. The method was extended to discrete-time systems in [30] and [256] An
alternative, recursive approach to probing is presented in [205].
In order to explain the harmonic probing procedure, it is necessary to
determine how a system responds to a harmonic input in terms of its Volterra
series.
First consider a periodic excitation composed of a single harmonic

x(t) = ei t : (8.27)

The spectral representation of this function follows immediately from the

well-known representation of the Æ -function (appendix D):
1 1
Z
Æ ( t) = d! ei!t (8.28)
2 1
so that
X (!) = 2Æ(! ): (8.29)
 Harmonic probing of the Volterra series 387

Figure 8.6. Principal diagonal of the first-, second- and third-order composite HFRFs for
the shock absorber and coil spring at an input voltage of 0.5 V.

Substituting this expression into equations (8.16)–(8.18) and forming the

total response as in (8.15) yields, up to third order,

1 +1
Z
Y (!) = H1 (!)2Æ(! )+ d! H (! ; ! ! 1 )
2 1 1 2 1
 2Æ(!1 )2Æ(! !1 )
Z +1 Z +1
1
+ d! d! H (! ; ! ; ! !1 !2 )
(2)2 1 1 1 2 3 1 2
 2Æ(!1 )2Æ(!2 )2Æ(! !1 !2 ) +



(8.30)

using the argument-changing property of the Æ -function and carrying out the
388 The Volterra series and higher-order frequency response functions

Figure 8.7. Principal quadrant of the second-order composite HFRF 2 (!1 ; !2 ) for the
shock absorber and coil spring at an input voltage of 0.5 V.

integrals gives

Y (!) = 2fH1 ( )Æ(! ) + H2 ( ; )Æ(! 2 )

+ H3 ( ; ; )Æ(! 3 ) +


g: (8.31)

Taking the inverse Fourier transform yields the required response:

y(t) = H1 ( )ei t + H2 ( ; )ei2 t + H3 ( ; ; )ei3 t +


: (8.32)

This shows clearly that components in the output at multiples of the

excitation frequency are expected, i.e. harmonics. The important point here is
that the component in the output at the forcing frequency is H 1 ( ).
Probing the system with a single harmonic only yields information about the
values of the FRFs on the diagonal line in the frequency spaces. In order to obtain
 Harmonic probing of the Volterra series 389

Figure 8.8. Principal quadrant of the third-order composite HFRF 3 (!1 ; !2 ; !3 ) for the
shock absorber and coil spring at an input voltage of 0.5 V.

further information, multi-frequency excitations must be used. With this in mind,

consider the ‘two-tone’ input

x(t) = ei 1 t + ei 2 t (8.33)

which has spectral representation

X (!) = 2Æ(! 1 ) + 2Æ(! 2) (8.34)

substituting into (8.16)–(8.18) and thence into (8.15) yields

Y (!) = H1 (!)2Æ(! 1 ) + H1 (!)2Æ(! 2 )

Z +1
1
+ d! H (! ; ! !1 )f2Æ(!1 1 ) + 2Æ(!1 2 )g
2 1 1 2 1
390 The Volterra series and higher-order frequency response functions

 f2Æ(! !1 1 ) + 2Æ(! !1 2 )g
1 +1
Z Z +1
+ d! d! H (! ; ! ; ! !1 !2 )
(2)2 1 1 1 2 3 1 2
 f2Æ(!1 1 )+2Æ(!1 2 )gf2Æ(!2 1 )+2Æ(!2 2 )g
 f2Æ(! !1 !2 1 ) + 2Æ(! !1 !2 2 )g +


:
(8.35)

It is a straightforward but tedious matter to expand this expression and

perform the integrals. After making use of the symmetry properties of the higher-
order FRFs, namely H (! 1 ; !2 ) = H (!2 ; !1 ) and H ( !1 ; !2 ) = H  (!1 ; !2 ),
one obtains
Y (!)
= H1 ( 1 )Æ(! 1 ) + H1 ( 2 )Æ(! 2 ) + H2 ( 1 ; 1 )Æ(! 2 1)
2
+ 2H2 ( 1 ; 2 )Æ(! 1 2 ) + H2 ( 2 ; 2 )Æ(! 2 2 )
+ H3 ( 1 ; 1 ; 1 )Æ(! 3 1 ) + 3H3( 1 ; 1 ; 2 )Æ(! 2 1 2)
+ 3H3 ( 1 ; 2 ; 2 )Æ(! 1 2 2 ) + H3 ( 2 ; 2 ; 2 )Æ(! 3 2)
+


: (8.36)

On taking the inverse Fourier transform, one obtains the response up to third
order:

y(t) = H1 ( 1 )eit 1 + H1 ( 2 )eit 2

+ H2 ( 1 ; 1 )eit2 1 + 2H2( 1 ; 2 )eit( 1 + 2 ) + H2 ( 2 ; 2 )eit2 2

+ H3 ( 1 ; 1 ; 1 )eit3 1 + 3H3 ( 1 ; 1 ; 2 )eit(2 1 + 2 )

+ 3H3 ( 1 ; 2 ; 2 )eit( 1 +2 2 ) + H3 ( 2 ; 2 ; 2 )eit3 2 +


:
(8.37)

The important thing to note here is that the amplitude of the component at
the sum frequency for the excitation, i.e. at 1 + 2 is twice the second-order
FRF H2 ( 1 ; 2 ). In fact, if a general periodic excitation is used, i.e.

x(t) = ei t +


+ ei n (8.38)

it can be shown that the amplitude of the output component at the frequency
1 +


+ n is n!Hn ( 1 ; : : : ; n ). This single fact is the basis of the harmonic
probing algorithm. In order to find the second-order FRF of a system for example,
one substitutes the expressions for the input (8.33) and general output (8.37) into
the system equation of motion and extracts the coefficient of e i( 1 + 2 )t ; this
yields an algebraic expression for H 2 .
The procedure is best illustrated by choosing an example. Consider the
continuous-time system
Dy + y + y2 = x(t) (8.39)
 Harmonic probing of the Volterra series 391

where D = ddt . In order to find H 1 , the probing expressions

x(t) = xp1 (t) = ei t (8.40)

and
y(t) = y1p (t) = H1 ( )ei t (8.41)
are substituted into the equation (8.39), the result being

(i + 1)H1 ( )ei t + H1 ( )2 ei2 t = ei t (8.42)

equating the coefficients of e i t on each side of this expression yields an equation

for H1
(i + 1)H1 ( ) = 1 (8.43)
which is trivially solved, yielding the expression
1
H1 ( ) = : (8.44)
i +1
Evaluation of H 2 is only a little more complicated. The probing expressions

x(t) = xp2 (t) = ei 1 t + ei 2 t (8.45)

and

y(t) = y2p(t) = H1 ( 1 )ei 1 t + H1 ( 2 )ei 2 t + 2H2 ( 1 ; 2 )ei( 1 + 2 )t (8.46)

are used. Note that in passing from the general output (8.37) to the probing
expression (8.46), all second-order terms except that at the sum frequency have
been deleted. This is a very useful simplification and is allowed because no
combination of the missing terms can produce a component at the sum frequency
and therefore they cannot appear in the final expression for H 2 . Substituting
(8.45) and (8.46) into (8.39), and extracting the coefficients of e i( 1 + 2 )t yields

(i( 1 + 2 ) + 1)H2 ( 1 ; 2 ) + H1 ( 1 )H1 ( 2 ) = 0: (8.47)

So that
H1 ( 1 )H1 ( 2 )
H2 ( 1 ; 2 ) = = H1 ( 1 )H1 ( 2 )H1 ( 1 + 2 )
i( 1 + 2 ) + 1
1
= (8.48)
(i 1 + 1)(i 2 + 1)(i[ 1 + 2 ] + 1)
on using the previously obtained expression for H 1 .
The next example is a little more interesting. Consider the asymmetric
Duffing equation

mD2 y + cDy + ky + k2 y2 + k3 y3 = x(t) (8.49)

392 The Volterra series and higher-order frequency response functions

this time with the D notation.

H1 and H2 for this system can be evaluated by exactly the same procedure
as used on the previous example. The results are
1
H1 (!) =
m!2 + ic! + k
(8.50)
H2 (!1 ; !2 ) = k2 H1 (!1 )H1 (!2 )H1 (!1 + !2 ): (8.51)

Note that the constant k 2 multiplies the whole expression for H 2 , so that if
the square-law term is absent from the equation of motion, H 2 vanishes. This
reflects a quite general property of the Volterra series; if all nonlinear terms in
the equation of motion for a system are odd powers of x or y , then the associated
Volterra series has no even-order kernels. As a consequence it will possess no
even-order kernel transforms.
In order to obtain H 3 , the required probing expressions are

x(t) = xp3 (t) = ei!1 t + ei!2t + ei!3 t (8.52)

and

y(t) = y3p (t) = H1 (!1 )ei!1 t + H1 (!2 )ei!2 t + H1 (!3 )ei!3 t

+ 2H2(!1 ; !2)ei(!1 +!2 )t + 2H2 (!1 ; !3 )ei(!1 +!3 )t
+ 2H2(!2 ; !3)ei(!2 +!3 )t + 6H3 (!1 ; !2 ; !3 )ei(!1 +!2 +!3 )t (8.53)
which are sufficiently general to obtain H 3 for any system. Substituting into the
Duffing equation and extracting the coefficient of e i(!1 +!2 +!3 )t yields

H3 (!1 ; !2 ; !3) = 61 H1 (!1 + !2 + !3 )

 f4k2(H1 (!1 )H2 (!2 ; !3 ) + H1 (!2 )H2 (!3 ; !1)
+ H1 (!3 )H2 (!1 ; !2 ))
+ 6k3 H1 (!1 )H1 (!2 )H1 (!3 )g: (8.54)

A discussion of the interpretation of these functions is deferred until a little later.

It is property of many systems that all higher-order FRFs can be expressed
in terms of H1 for the system. The exact form of the expression will depend on
the particular system.
The harmonic probing algorithm has been established for continuous-time
systems, i.e. those whose evolution is governed by differential equations of
motion. The NARMAX models discussed in chapter 6 are difference equations so
the probing algorithm requires a little modification as in [32] and [256]. Consider
the difference equation analogue of equation (8.39):

4y + y + y2 = x(t): (8.55)

where 4 is the backward shift operator, defined by 4y (t) = y (t 1).

(Throughout this chapter it is assumed, except where indicated, that the sampling
 Harmonic probing of the Volterra series 393

interval for a discrete-time system is scaled to unity. This yields a unit sampling
frequency and Nyquist frequency of 0.5.) In the usual notation for difference
equations, (8.55) becomes

yi 1 + yi + yi2 = xi : (8.56)

However, the form containing 4 allows the most direct comparison with
the continuous-time case. It is clear from the previous argument that the only
differences for harmonic probing of discrete-time systems will be generated by
the fact that the operator 4 has a different action on functions e i!t to the operator
D. This action is very simple to compute, as shown in chapter 1 3,
4ei!t = ei!(t 1) = e i!
ei!t : (8.57)

It is now clear that one can carry out the harmonic probing algorithm for
(8.55) exactly as for the continuous-time (8.39); the only difference will be that
the 4 operator will generate a multiplier e i! wherever D generated a factor i! .
As a consequence H 1 and H2 for (8.55) are easily computed.

1
H1 (!) =
e i! + 1
(8.58)
H1 (!1 )H1 (!2 )
H2 (!1 ; !2) = = H1 (!1 )H1 (!1 )H1 (!1 + !2 ):
e i(! +! ) + 1
(8.59)
1 2

Note that the form of H 2 as a function of H 1 is identical to that for the

continuous-time system.
It is possible at this point to make a quite general statement. Given
a continuous-time system with linear or nonlinear equation of motion
f (D; y; x) = 0 and HFRFs Hnc (!1 ; : : : ; !n), n = 1; : : : ; 1, the corresponding
discrete-time system f (4; y; x) = 0 has HFRFs Hnd (!1 ; : : : ; !n ) =
Hnc ( ie i!1 ; : : : ; ie i!n ), n = 1; : : : ; 1. Further the functional relationships
between the Hn and H1 will be identical in both cases.
The system in equation (8.56) is not an NARMAX system as it is a nonlinear
function of the most recent sampled value y i . As discussed in chapter 6, an
NARMAX, or more strictly NARX, model has the general form

yi = F (yi 1 ; : : : ; yiny ; xi 1 ; : : : ; xinx ) (8.60)

with appropriate noise modelling if necessary.

The relevant existence theorems obtained in [161, 162] show that this form
is general enough to represent almost all input–output systems.
3 4=e
It is amusing to note that this action follows from the fact that D as an operator equation;
as e i!t is an eigenfunction of D with eigenvalue !, it is also an eigenfunction of with eigenvalue
i 4
e i! .
394 The Volterra series and higher-order frequency response functions

8.4 Validation and interpretation of the higher-order FRFs

In order to justify studying the higher-order FRFs it is necessary to show that
they contain useful information about whatever system is under examination. In
fact, as time- and frequency-domain representations are completely equivalent,
the higher-order FRFs contain all system information; later in this section it is
demonstrated that important facts can be conveyed in a very direct and visible
way.
Before discussing matters of interpretation it is important to address the
question of uniqueness of the higher-order FRFs as it is critical to any analysis
that the non-uniqueness of the time-domain NARMAX representation of a system
does not affect the frequency-domain representation.
The first thing which must be established is the correspondence between
the FRFs of the continuous system and the FRFs of the discrete approximations.
Consider the Duffing oscillator of equation (8.49), a discrete-time representation
for this system could be obtained by adopting discrete approximations to the
derivatives. The coarsest approximation available is the backward-difference
approximation

y y
y_i  i i 1 (8.61)

t
yi+1 2yi + yi 1
yi 

t2
(8.62)

which gives the discrete-time representation

2m + c
t + k
t2
   
m
yi = yi 1 y
m + c
t m + c
t i 2
k2
t2 k3
t2
t2
     
y 2 3
y + x :
m + c
t i 1 m + c
t i 1 m + c
t i 1
(8.63)

In fact, because this is based on the coarse approximations (8.61) and

(8.62), it does not yield good representations of the higher-order FRFs. In order
to demonstrate accurate FRFs from a NARX model, the following numerical
simulation was carried out. A fourth-order Runge–Kutta scheme [209], was used
to obtain the response of the system (8.49) under excitation by a Gaussian noise
sequence x(t) with rms 10.0 and frequency range 0–90 Hz. The coefficient values
adopted were: m = 1, c = 20, k = 10 4 , k2 = 107 , k3 = 5  109 . This system
has a resonant frequency of ! r = 99 rad s 1 or fr = 2!r = 15:75 Hz. The data
were generated with a sampling interval of 0.005 s, giving a Nyquist frequency of
100 Hz.
A NARX model was fitted to 1000 points of the resulting discrete x and
y data using the estimation and validation methods described in the previous
 Validation and interpretation of the higher-order FRFs 395

Figure 8.9. Comparison between simulated Duffing oscillator data and the prediction by a
NARX model.

section. The result was

yi = 1:6696yi 1 0:903 48yi 2
2:1830  102yi2 1 1:0665  105yi3 1
6
+ 3:0027  10 xi + 1:8040  10 5xi 1 (8.64)
6
+ 2:7676  10 xi 2 :
Figure 8.9 shows a comparison between the original y data from the
simulation, and that predicted by the NARX model (8.64), when excited by the
same input data x; the NARX model clearly gives a good representation of the
system in the time domain. The fitted model was then used to generate the higher-
order FRFs, H1 , H2 and H3 , by the method of harmonic probing. As the exact
results could also be obtained by harmonic probing of (8.49), direct comparisons
could be made. In all cases, the exact FRFs are given with the frequency scale
in Hz; the FRFs for the discrete model are given with corresponding normalized
frequency scales f n = f=fs where fs is the sampling frequency; the Nyquist
frequency is 0.5 in these units.
Figure 8.10 shows a comparison between the exact H 1 and that obtained
from the model; the agreement looks excellent. However, an important point
must be raised here. H 1 for the discrete system is only an approximation to
H1 for the continuous system up to the Nyquist frequency of 0.5 (100 Hz); it is
only plotted up to this frequency in figures 8.10(c) and 8.10(d) because it simply
repeats beyond this point and is therefore meaningless.
396 The Volterra series and higher-order frequency response functions
Frequency (Hz)

(a)
Gain (dB)

Frequency (Hz)
Phase (degrees)

(b)

Normalised Frequency

(c)
Gain (dB)

Normalised Frequency

(d)
Phase (degrees)

Figure 8.10. H1 (f ) for the Duffing oscillator system: (a) exact magnitude; (b) exact
phase; (c) NARX model magnitude; (d) NARX model phase.
 Validation and interpretation of the higher-order FRFs 397

(a) (b)

(d)
(c)

Figure 8.11. H2 (f1 ; f2 ) surface for the Duffing oscillator system: (a) exact magnitude;
(b) exact phase; (c) NARX model magnitude; (d) NARX model phase.

The comparison between the exact H 2 and that from the NARMAX model
is given in figure 8.11. The same comparison using the contour maps for the
functions is shown in figure 8.12; again the agreement is very good. Note that
because H2 contains factors H1 (2f1 ) and H2 (2f2 ) it would be meaningless
to plot it outside the ranges corresponding to f 1  100; f2  100. Further,
H2 also contains a factor H1 (2(f1 + f2 )) so that the plots should not extend
past the area specified by f 1 + f2  100. Rather than plot irregularly shaped
regions, the H2 figures presented in this book include information beyond this last
bound, which is indicated by the full line in the model contour maps in figure 8.12;
information presented outside this region on any H 2 plot should not be regarded
as meaningful.
The comparison between the exact H 3 and model H 3 is given in figure 8.13,
and in contour map form in figure 8.14. Unfortunately, the whole H 3 surface
cannot be plotted as it exists as a three-dimensional manifold embedded in a
four-dimensional space over the (! 1 ; !2 ; !3 )-‘plane’. However, one can plot
two-dimensional submanifolds of H 3 , and this is the approach which is usually
adopted. Figures 8.13 and 8.14 show H 3 (!1 ; !2 ; !1 ) plotted over the (! 1 ; !2 )-
plane. The region of validity of the H 3 surface is a little more complicated in
398 The Volterra series and higher-order frequency response functions
Gain (dB)

(a)
f1

f2
Phase

(b)
f1

f2

Gain (dB)

(c)
f1

f2
Phase

(d)
f1

f2

Figure 8.12. H2 (f1 ; f2 ) contours for the Duffing oscillator system: (a) exact magnitude;
(b) exact phase; (c) NARX model magnitude; (d) NARX model phase.
 Validation and interpretation of the higher-order FRFs 399

(a) (b)

(d)
(c)

Figure 8.13. H3 (f1 ; f2 ; f1 ) surface for the Duffing oscillator system: (a) exact magnitude;
(b) exact phase; (c) NARX model magnitude; (d) NARX model phase.

this situation. In all cases, agreement between the exact H n and those obtained
from the NARMAX model appears impressive. For a less passive comparison,
figure 8.15 shows the gain and phase of the output components y 1 , y2 and y3
obtained from the systems defined by the exact and model FRFs when excited by
a unit sinusoid at various frequencies. Again, agreement looks excellent. Note
that the plot for second harmonic in figure 8.15 contains a peak at f r =2. This
is due to the fact that the diagonal HFRF contains a factor H 1 (2! ) as shown by
equation (8.51).
Having established that a NARX model can yield good representations of
the FRFs from a continuous system, the next question which must be addressed
concerns the correspondence between frequency-domain representations of
different yet exactly equivalent NARX models. (Non-uniqueness is actually a
problem with most methods of modelling, it is not specific to NARX). Suppose
one has obtained as an accurate discretization of a continuous system, the ARX
model,
yi = a1 yi 1 + a2 yi 2 + b1 xi 1 : (8.65)

As this expression holds for all values of i (away from the initial points), it
400 The Volterra series and higher-order frequency response functions
Gain (dB)

(a)
f1

Phase
f2

(b)
f1

f2

Gain (dB)

(c)
f1

f2
Phase

(d)
f1

f2

Figure 8.14. H3 (f1 ; f2 ; f1 ) contours for the Duffing oscillator system: (a) exact
magnitude; (b) exact phase; (c) NARX model magnitude; (d) NARX model phase.
 Validation and interpretation of the higher-order FRFs 401
Frequency (Hz)

(a)
Amplitude (dB)
Phase (deg)

(b)

Frequency (Hz)

Normalised Frequency

(c)
Amplitude (dB)
Phase (deg)

(d)

Normalised Frequency

Figure 8.15. H1 , H2 and H3 components for the Duffing oscillator response excited by
a unit sinusoid: (a) exact magnitude; (b) exact phase; (c) NARX model magnitude; (d)
NARX model phase.
402 The Volterra series and higher-order frequency response functions

can just as well be written as

yi 1 = a1 yi 2 + a2 yi 3 + b1 xi 2 (8.66)

and substituting (8.66) into (8.65) yields the ARX model

yi = (a21 + a2 )yi 2 + a1 a2 yi 3 + b1 xi 1 + a1 b1 xi 2 (8.67)

which is exactly equivalent to (8.65) yet contains different terms. This type of
ambiguity will occur for any system which regresses the present output onto
past values of output. It is a reflection of a type of ambiguity for continuous-
time systems; one can always differentiate the equation of motion to obtain a
completely equivalent system. The only thing which changes is the set of objects
for which initial conditions are required. Harmonic probing of (8.65) yields (in
symbolic notation where 4 = e i! )

b1 4
H1(8:65) =
1 a1 4 a2 42
(8.68)

while probing of (8.67) gives the superficially different

b1 4 +a1 b1 42
H1(8:67) = :
1 (a21 + a2 ) 42 a1 a2 43
(8.69)

However, the latter expression factors:

b1 4 (a1 4 +1) b1 4
H1(8:67) = = = H1(8:65):
(a1 4 +1)(1 a1 4 a24 ) 1 a1 4 a2 42
2
(8.70)
The final type of non-uniqueness is generated by the fact that NARMAX
models can be approximately equivalent. As an illustration consider the simple
system
yi = yi 1 + xi 1 : (8.71)
If is small, a simple application of the binomial theorem gives

(1 4)yi = xi 1 =) yi = (1 4) 1xi 1
=) yi = (1 + 4)xi 1 + O( 2 ) (8.72)

So the system
yi = xi 1 + xi 2 (8.73)
is equivalent to the system in (8.71) up to O( 2 ). Now, harmonic probing of
system (8.71) yields the FRF
1
H1(8:71)(!) =
e i!
(8.74)
1
 Validation and interpretation of the higher-order FRFs 403

and a similar analysis for (8.73) gives

1
H1(8:73)(!) = 1+ e i! = +O( 2 ) = H1(8:71) (!)+O( 2 ):
1 e i!
(8.75)

Note that by retaining n terms in the binomial expansion, the model

yi = xi 1 + xi 2 +


+ n 1 xi n (8.76)

is obtained which is equivalent to (8.71) up to O( n ). As a result, the system

(8.71) can be represented with arbitrary accuracy by the binomial expansion if
n is large enough. However, note that one representation has only three model
terms while the other has n with n possibly large. This serves to illustrate why
it is important to correctly detect the model structure or which terms are in the
model in order to yield a parsimonious model [32].
One must be careful not to regard these simple arguments as generating
a general principle; however, it would seem likely that equivalence of two
NARX models up to a given order of accuracy would imply equivalence of the
corresponding HFRFs up to the same order of accuracy. This is easy to establish
in the case of a general linear system by an extension of the previous argument.
The various cases discussed earlier exhaust all possibilities for obtaining
different NARX representations of a given system.
This discussion is simply intended as an argument that all NARX models
which are equivalent in the sense that they furnish a discrete approximation to
a continuous system will have higher-order FRFs which not only approximate
to each other but also to those of the underlying continuous system. It does
not constitute a rigorous proof in any sense; however, it is difficult to imagine
a situation under which this condition would not hold.
Having established some confidence in their reliability, the interpretation of
the higher-order FRFs can be discussed. The Duffing oscillator system (8.49)
serves well as an illustration. The magnitude and phase of the expression (8.50)
for H1 (! ) = H1 (2f ) is given in figures 8.10(a) and (b) on the frequency interval
0–100 Hz. The interpretation of these figures, traditionally given together and
universally called the Bode plot, has been described in earlier chapters, notably
chapter 1. The peak in the magnitude at f = f r = 15:75 Hz shows that for
this frequency of excitation the amplitude of the linear part of the response y 1 (t)
is a maximum. The Bode plot thus allows the immediate identification of those
excitation frequencies at which the vibration level of the system is likely to be
high.
Interpretation of the second-order FRF is also straightforward. The
magnitude and phase of H 2 for the Duffing system given earlier are given in
figures 8.11(a) and (b) as surfaces, or in figures 8.12(a) and (b) as contour maps,
over the (f1 ; f2 ) = ( 2!1 ; !
22 ) plane. The frequency ranges for the plot are the same
as for H1 in figure 8.10. A number of ridges are observed. These are in direct
correspondence with the peak in H 1 as follows. According to equation (8.51),
404 The Volterra series and higher-order frequency response functions

H2 is a constant multiple of H1 (!1 )H1 (!2 )H1 (!1 + !2 ). As a consequence H2

possesses local maxima at positions where the H 1 factors have local maxima.
Consequently there are two ridges in the H 2 surface corresponding to the lines
!1 = !r = 2fr and !2 = !r . These are along lines parallel to the frequency
axes. In addition, H 2 has local maxima generated by the H 1 (!1 + !2 ) factor
along the line ! 1 + !2 = !r . This ridge has an important implication; it indicates
that one can expect a maximum in the second-order output y 2 (t) if the system is
excited by two sinusoids whose sum frequency is the linear resonant frequency.
This shows clearly why estimation of a transfer function by linear methods is
inadequate for nonlinear systems; such a transfer function would usually indicate
a maximum in the output for a harmonic excitation close to the linear resonant
frequency. However, it would fail to predict that one could excite a large nonlinear
component in the output by exciting at ! = !2r ; this is a consequence of the
trivial decomposition 2e i 2 t = ei 2 t + ei 2 t which means that the signal can be
!r !r !r

regarded as a ‘two-tone’ input with a sum frequency at the linear resonance ! r .

The importance of the second-order FRF is now clear. It reveals those pairs of
excitation frequencies which will conspire to produce large levels of vibration as
a result of second-order nonlinear effects.
The interpretation of H 3 for the system is very similar. Consideration of
equation (8.54) shows that for a three-tone input of the form (8.52) one should
expect maxima in the third-order output y 3 (t) if the following conditions are
satisfied: !1 = !r , !2 = !r , !3 = !r , !1 +!2 = !r , !2 +!3 = !r , !3 +!1 = !r ,
!1 + !2 + !3 = !r . The presence of these ‘combination resonances’ would be
indicated by the presence of ridges in the H 3 surface. Although figures 8.13 and
8.14 only show the ‘projections’ of H 3 over the (!1 ; !2 )-plane, they are sufficient
to indicate the presence of the ‘combination resonances’ ! 1 = !r , !2 = !r ,
!1 + !2 = !r, 2!1 = !r, 2!1 + !2 = !r. It is clear that the local maximum
distributions become more and more complex as the order of the HFRF increases.
These arguments show that the higher-order FRFs provide directly visible
information about the possible excitation of large nonlinear vibrations through
the cooperation of certain frequencies.

8.5 An application to wave forces

The power of the NARX and higher-order FRF approaches can be demonstrated
by the following example used in chapter 6 where force and velocity data were
obtained from a circular cylinder placed in a planar oscillating fluid flow in a large
U-tube [199]. The standard means of predicting forces on cylinders used by the
offshore industry is to use Morison’s equation (6.121) which expresses the force
as a simple nonlinear function of the instantaneous flow velocity and acceleration.
For one particular frequency of flow oscillation, Morison’s equation gave the force
prediction shown in figure 8.16(a) compared with the measured force. Morison’s
equation is inadequate at representing the higher-frequency components of the
 FRFs and Hilbert transforms: sine excitation 405

force. The model inadequacy is shown clearly by the correlation-based validity

tests (section 6.8.3) in figure 8.17(b) 4.
A NARX fit to the force–velocity data gave the model prediction shown in
figure 8.17(a). This model also passes the correlation tests (figure 8.17(b)).
A similar analysis has been carried out on fluid-loading data encompassing
a broad range of flow conditions ranging from U-tube data to that from a
unidirectional wave in a large flume to data from a random directional sea. In all
cases, the NARX analysis produced a better model than Morison’s equation [276].
Unfortunately the model structures varied. In order to examine the possibility
that this was simply due to the non-uniqueness of the NARX representations,
the higher-order FRFs were obtained by harmonic probing. The results were
very interesting, as an example, H 3 for the U-tube data of figure 8.17 is given
in figure 8.18. The pronounced ridges were shown to appear in the third-order
FRFs for all of the flow conditions examined; this is in direct contradiction to
Morison’s equation which forces a constant H 3 .
The higher-order FRFs can often throw light onto a problem in this way;
the direct visualization of the system properties which they provide is appealing.
They have actually been used in wave loading studies for some time now;
however, the computational burden imposed by traditional methods of estimation
has prohibited the use of functions higher than second order [85]

8.6 FRFs and Hilbert transforms: sine excitation

8.6.1 The FRF
It was shown earlier that the Volterra series provides a convenient means for
calculating the nonlinear system response to a single harmonic; this forms the
basis of the harmonic probing method. It is only slightly more complicated to
calculate the response to multiple harmonics. The benefit is that one can then
determine the response to a sinusoid and this, in turn, will allow us to develop an
expression for the stepped-sine FRF of the system. Suppose the excitation is a
two-tone signal
x(t) = Aei a t + B ei b t (8.77)
which translates into the frequency domain as
X (!) = 2fAÆ(! a ) + BÆ(! b )g: (8.78)
Substituting this into (8.16)–(8.18) and thence into (8.15) leads, after a long
but straightforward calculation, to
y(t) = AH ( )ei a t + BH ( )ei b t + A2 H ( ; )e2i a t
1 a 1 b 2 a a
4 Of course, with enough parameters, one can fit a model to an arbitrary level of accuracy on a given
estimation set of data. The modeller should always carry out appropriate levels of model validity
testing in order to ensure that the model is genuine and does not simply represent an isolated data set.
This is particularly pressing in the situation where one might abandon a physical model like Morison’s
equation in favour of a non-physical model on the grounds of model accuracy.
406 The Volterra series and higher-order frequency response functions

(a)

(b)

Figure 8.16. Morison equation fit to experimental U-tube data: (a) model-predicted
output; (b) correlation tests.

+ 2ABH2 ( a ; b )ei( a + b )t + B 2 H2 ( b ; b)e2i b t

+ A3 H3 ( a ; a ; a )e3i a t + A2 BH3 ( a ; a ; b )ei(2 a + b )t
 FRFs and Hilbert transforms: sine excitation 407

(a)

(b)

Figure 8.17. NARX model fit to experimental U-tube data: (a) model-predicted output;
(b) correlation tests.

+ AB 2 H3 ( a ; b ; b)ei( a +2 b )t + B 3 H3 ( b ; b ; b )e3i b t +



(8.79)
408 The Volterra series and higher-order frequency response functions

(a) (b)

Gain (dB)

(c)
f1

f2
Phase

(d)
f1

f2

Figure 8.18. H3 (f1 ; f2 ; f1 ) from NARX fit to U-tube data: (a) magnitude; (b) phase; (c)
magnitude contours; (d) phase contours.

to third order.
Now, for the response to a cosinusoid

x(t) = cos( t) = 21 X (ei t + e i t ) (8.80)

one simply substitutes A = B = X=2, a = and b = . To third order

 FRFs and Hilbert transforms: sine excitation 409

again, the result is

X X X2
y(t) = H1 ( )ei t + H1 ( )e i t + H2 ( ; )e2i t
2 2 4
X2 X2 X3
+ H2 ( ; ) + H2 ( ; )e 2i t + H3 ( ; ; )e3i t
2 4 8
3X 3 3 X 3
+ H ( ; ; )ei t + H ( ; ; )e i t
8 3 8 3
X3
+ H3 ( ; ; )e 3i t +


: (8.81)
8
Making use of the reflection properties H 1 ( ) = H  ( ) etc, and applying
de Moivre’s theorem in the form

z ei t + z  e i t = jz jei( t+\z) + jz je i( +\z) = 2jz j cos( t + \z ) (8.82)

yields

y(t) = X jH1( )j cos( t + \H1 ( ))

X2 X2
+ jH2 ( ; )j cos(2 t + \H2 ( ; )) + H( ; )
2 2 2
X3
+ jH3 ( ; ; )j cos(3 t + \H3 ( ; ; ))
4
3X 3
+
4
jH3 ( ; ; )j cos( t + \H3 ( ; ; )) +


(8.83)

which shows again that the response contains all odd and even harmonics. The
component of the response at the forcing frequency is

y(t) = X jH1 ( )j cos( t + \H1 ( ))

3X 3
+
4
jH3 ( ; ; )j cos( t + \H3 ( ; ; )) +


(8.84)

and this immediately identifies the composite FRF  s ( ) as

3X 2
s ( ) = H1 ( ) + H ( ; ; ) +



4 3
(8.85)

or
3X 2 5X 4
s ( ) = H1 ( ) + H( ; ; )+ H( ; ; ; ; )+



4 3 8 5
(8.86)
to the next highest order. Again, it is useful to take the Duffing oscillator (8.49)
as an example. Equation (8.54) with k 2 = 0, gives

H3 (!1 ; !2 ; !3) = k3 H1 (!1 )H1 (!2 )H1 (!3 )H1 (!1 + !2 + !3 ) (8.87)
410 The Volterra series and higher-order frequency response functions

(adopting lower-case ! from now on) or

H3 (!; !; !) = k3 H1 (!)3 H1 (!): (8.88)

Harmonic balance gives for (8.49)

1
H5 (!1 ; !2 ; !3 ; !4 ; !5) = H1 (!1 + !2 + !3 + !4 + !5 )
5!
 ( 3!3!k3 fH3(!1 ; !2 ; !3)H1 (!4 )H1 (!5 )
+ H3 (!1 ; !2 ; !4)H1 (!3 )H1 (!5 )
+ H3 (!1 ; !3 ; !4)H1 (!2 )H1 (!5 ) + H3 (!2 ; !3 ; !4 )H1 (!1 )H1 (!5 )
+ H3 (!1 ; !2 ; !5)H1 (!3 )H1 (!4 ) + H3 (!1 ; !3 ; !5 )H1 (!2 )H1 (!4 )
+ H3 (!2 ; !3 ; !5)H1 (!1 )H1 (!4 ) + H3 (!1 ; !4 ; !5 )H1 (!2 )H1 (!3 )
+ H3 (!2 ; !4 ; !5)H1 (!1 )H1 (!3 ) + H3 (!3 ; !4 ; !5 )H1 (!1 )H1 (!2 )g)
(8.89)

and thence

H5 (!; !; !; !; !) = 103 k32 (3H1 (!)4 H1 (!)3 + 6H1 (!)5 H1(!)2
+ H1 (!)4 H1 (!)2 H1 (3!)): (8.90)

Substituting (8.90) and (8.88) into (8.86) gives the FRF up to O(X 4 )

3X 2 3X 4 2
s ( ) = H1 ( ) k3 H1 (!)3 H1 (!) + k (3H (!)4 H1 (!)3
4 16 3 1
+ 6H1 (!)5 H1 (!)2 + H1 (!)4 H1 (!)2 H1 (3!)) + O(X 6 ): (8.91)
(Amongst other places, this equation has been discussed in [236], where it
was used to draw some conclusions regarding the amplitude dependence of the
stepped-sine composite FRF.)
In order to illustrate these expressions, the system

y + 20y_ + 104y + 5  109y3 = X cos(!t) (8.92)

was chosen. Figure 8.19 shows the FRF magnitude plots obtained from (8.91)
for X = 0:01 (near linear), X = 0:5 and X = 0:75. At the higher amplitudes,
the expected FRF distortion is obtained, namely the resonant frequency shifts
up and the magnitude at resonance falls. Figure 8.20 shows the corresponding
Nyquist plots. (Note the unequal scales in the Real and Imaginary axes; the plots
are effectively circular.) Figure 8.21 shows the O(X 4 ) FRF compared with the
‘exact’ result from numerical simulation. There is a small degree of error near
resonance which is the result of premature truncation of the Volterra series.
 FRFs and Hilbert transforms: sine excitation 411

0.00050

0.00040
FRF magnitude

0.00030

0.00020

0.00010

0.00000
70.0 80.0 90.0 100.0 110.0 120.0 130.0
Frequency (rad/s)

Figure 8.19. Distortion in the magnitude plot of s1 (! ) computed from the Volterra series
for different levels of excitation.

8.6.2 Hilbert transform

Recall from chapters 4 and 5 that the Hilbert transform provides a mean of
diagnosing structural nonlinearity on the basis of FRF data. The mapping on
the FRF (! ) reduces to the identity on those functions corresponding to linear
systems. For nonlinear systems, the Hilbert transform results in a distorted version
~ , of the original FRF.
From chapter 5, if (! ) is decomposed so

(!) = +(!) +  (!) (8.93)

where + (! ) (respectively  (! )) has poles only in the upper (respectively

lower) half of the complex ! -plane. It is shown in chapter 5 that

H[ (!)] = (!) (8.94)

and the distortion suffered in passing from the FRF to the Hilbert transform is
given by the simple relation

(!) = H[(!)] (!) = 2 (!): (8.95)

412 The Volterra series and higher-order frequency response functions
0.00000

-0.00020
FRF Imaginary Part

-0.00040

-0.00060
-0.00040 -0.00020 0.00000 0.00020 0.00040
FRF Real Part

Figure 8.20. Distortion in the Nyquist plot of s1 (! ) computed from the Volterra series
for different levels of excitation. (Note that the Real and Imaginary axes do not have equal
scales.)
This section presents a technique which allows the Hilbert transform
distortion to be derived term by term from a Volterra series expansion of the
system FRF, the expansion parameter being X , the magnitude of the applied
sinusoidal excitation. It is illustrated on the Duffing oscillator (8.49), and the
basic form of FRF used is the O(X 4 ) approximation given in (8.91). If the FRF
is known, the Hilbert transform follows from the distortion (8.95). In order to
obtain the distortion, the pole–zero form of the FRF is needed.

8.6.2.1 Pole–zero form of the Duffing oscillator FRF

As the approximate nonlinear FRF has been expressed in terms of the linear FRF
in (8.91), it is necessary to find the pole–zero form of H 1 (! ); this will then yield
the pole–zero form of (8.91). The poles of H 1 (! ) are well known:
p1 ; p2 = !d + i!n (8.96)
where !d = !n (1  ) 2 1=2 is the damped natural frequency. In terms of these
quantities H1 (! ) may now be expressed as
1
H1 (!) = (8.97)
m(! p1 )(! p2 )
 FRFs and Hilbert transforms: sine excitation 413

0.00050

0.00040
FRF Magnitude

0.00030

0.00020 FRF (linear)

FRF (nonlinear-analytical)
FRF (nonlinear-numerical)
0.00010

0.00000
70.0 80.0 90.0 100.0 110.0 120.0 130.0
Frequency (rad/s)

Figure 8.21. Comparison between FRFs s1 (! ) computed from the Volterra series and
from numerical simulation.

and this is the required ‘pole–zero’ expansion. Note that p 1 and p2 are both in
the upper half-plane so H 1 (! ) = H1+ (! ) and the Hilbert transform is therefore
the identity on H 1 (! ) as required. However, the expression for  s (! ) in (8.91)
contains terms of the form H 1 (! ) with poles p1 and p2 ; these are in the lower
half-plane and are the cause of the Hilbert transform distortion for  s (! ). In
pole–zero form (8.91) becomes

1
s (! ) =
m(! p1 )(!
p2 )
3X 2 1
k3 4
4 m (! p1 ) (! p2 )3 (! p1 )(! p2 )
3

3 2 4 3
k3 X
16 m (! p1 ) (! p2 )4 (! p1 )3 (! p2 )3
7 4
6
+ 7
m (! p1 )5 (! p2 )5 (! p1 )2 (! p2 )2

1
+ 7
m (! p1 )4 (! p2 )4 (! p1 )2 (! p2 )2 (3! p1 )(3! p2 )
(8.98)

up to O(X 4 ). This is the appropriate form for calculating the distortion.

414 The Volterra series and higher-order frequency response functions

8.6.2.2 Partial fraction expansion

The method of effecting the decomposition (8.93) for the nonlinear FRF (8.91)
is to find the partial fraction expansion. Due to the complexity of the task, this
is accomplished using computer algebra. The O(X 2 ) and O(X 4 ) terms in the
transfer function may be considered separately.
The partial fraction expansion of the O(X 2 ) is easily found to have the form

A1 A2 A3 A4
+ 2 + 3 +
(! p1 ) (! p1 ) (! p1 ) (! p1 )
B1 B2 B3 B4
+ + + +
(! p2 ) (! p2 )2 (! p2 )3 (! p2 )
(8.99)

where
1
A4 =
( p1 + p1 ) ( p1 + p2 )3 ( p1 + p2 )
 3  (8.100)
1
A3 =
( p1 + p1 )( p1 + p2 )3 (p1 p2 )
(8.101)

5p21 + 4p1 p1 + 2p1 p2 p1 p2 + 4p1p2 3p1 p2 p2 p2

A2 =
( p1 + p1 )2 ( p1 + p2 )4 ( p1 + p2 )2
(8.102)

and finally
N1
A1 = (8.103)
D1
with

N1 = 15p41 + 24p31p1 10p21p12 + 12p31 p2 15p21p1 p2 + 5p1 p12 p2

3p21 p22 + 3p1 p1 p22 p12 p22 + 24p31 p2 37p21p1 p2 + 15p1p12 p2
15p21p2 p2 + 14p1p1 p2 p2 3p12p2 p2 + 3p1p22 p2 p1 p22 p2 10p21p22
+ 15p1p1 p22 6p12 p22 + 5p1p2 p22 3p1 p2 p22 p22 p22 (8.104)
 3 5
D1 = ( p1 + p1 ) ( p1 + p2 ) ( p1 + p2 ) :  3 (8.105)

The B coefficients are obtained simply by interchanging the 1 and 2

subscripts throughout.
Given the formula for the distortion, it is sufficient to consider only those
terms in (8.99) with poles in the lower half-plane. Further, it is sufficient to
concentrate on the pole at p 1 as the expression for p 2 will follow on interchanging
the subscripts 1 and 2. Hence

3X 2k3
s (!) =
2m4(p1 p1 )3 (p2 p1 )3 (p1 p2 )(! p1 )
+ (p1 ! p2 ):
(8.106)
 FRFs and Hilbert transforms: sine excitation 415

On substituting for p 1 and p2 , in terms of the physical parameters, the O(X 2 )

distortion (denoted here by
(2) s (! )) finally emerges as
3X 2k3
 

(2) s (!) =
2m4
2 2 2 2 2 2
 64!3  3 (4!!2n+ (!2!d2 )+3 (!!n  !) ++i!i(!!d )(!3!+n! )+ i!  ) :
n d n d n d n
(8.107)
A similar but more involved analysis for the O(X 4 ) distortion yields the
following six terms which generate
(4) s (! )—all other terms lie in +
s (!).
"
3X 4k32
 
3

(4) s (!) =
16m7 2048!d3!n4  4 (! !d + i!n  )3 (i!d + !n )4
3
+
2048!d3!n4  4 (! + !d + i!n  )3 ( i!d + !n  )4
12i!d3 + 89!d2!n  150i!d!n2  2 36!n3  3
+
8192!d4!n5  5 (! !d + i!n  )2 (i!d + !n  )5 (i!d + 2!n )
12i!d3 + 89!d2 !n + 150i!d!n2  2 36!n3  3
+
8192!d4!n5  5 (!+!d+i!n  )2 ( i!d+!n )5 ( i!d +2!n  )
+ T5 + T6 (8.108)
where T5 is the quotient N5 =D5 with
N5 = 280i!d5 + 1995!d4!n  5080i!d3!n2  2 5344!d2!n3  3
+ 2016i!d!n4  4 + 288!n5  5 (8.109)
5 5 5 6 2
D5 = 32 768!d!n  (! !d + i!n  )(i!d + !n  ) (i!d + 2 !n ) (8.110)
and T6 is given by N6 =D6 where
N6 = 280i!d5 1995!d4!n  5080i!d3!n2  2 + 5344!d2!n3  3
+ 2016i!d!n4  4 288!n5  5 (8.111)
and
D6 = 32 768!d5!n5  5 (! + !d + i!n )( i!d + !n  )6 ( i !d + 2!n )2 ]:
(8.112)
Using the O(X 2 ) and O(X 4 ) distortion terms,
(2) s (! ) and
(4) s (! ),
the Hilbert transform of the Duffing oscillator FRF  s (! ) (represented by a three-
term Volterra series) may be expressed as
H[s (!)] = ~s (!) = s (!) +
(2) s (!) +
(4) s (!): (8.113)
This relationship may be used to calculate numerical values for the Hilbert
transform as a function of frequency, forcing and level of nonlinearity.
416 The Volterra series and higher-order frequency response functions
0.00020
FRF (analytical)
HT (numerical)
HT (analytical)
FRF Imaginary Part 0.00000

-0.00020

-0.00040

-0.00060
-0.00040 -0.00020 0.00000 0.00020 0.00040
FRF Real Part

Figure 8.22. Comparison between the numerical estimate of the Hilbert transform and the
O(X 2 ) Volterra series estimate for the Duffing oscillator under sine excitation. (Note that
the Real and Imaginary axes do not have equal scales.)

8.6.2.3 Numerical example

Using the expressions for the O(X 2 ) and O(X 4 ) contributions to the nonlinear
FRF (equation (8.91)), and the
(2) s (! ) and
(4) s (! ) distortion terms,
a FORTRAN program was used to evaluate the FRF and Hilbert transform
numerically for the particular Duffing oscillator given in (8.92). The expressions
were obtained for 1024 spectral lines from 0 to 200 rad s 1 .
The FRF and HT expressions were evaluated for two levels of excitation,
specified by X = 0:5 and 0.75.
Figure 8.22 shows an overlay of the Volterra series FRF (full line), and
the associated analytical Hilbert transform (broken), as obtained from the
(2)
distortion term; this result was obtained using the excitation with X = 0:75N .
The rotation of the Hilbert transform towards the left and the increase in amplitude
over that of the FRF are both established features of the Hilbert transform of
a Duffing oscillator (see chapter 4). The broken trace in figure 8.22 shows the
Hilbert transform evaluated from the FRF by numerical means. Even using only
the
(2) distortion, the theory gives excellent agreement. With the fourth-order
distortion included (figure 8.23), agreement is almost perfect. Note that the plots
are effectively circular but that in the figures the Real and Imaginary axes are not
of equal scales.

8.7 FRFs and Hilbert transforms: random excitation

The object of this section is to derive the composite FRF for a Duffing Oscillator
under random excitation. Although the FRF mirrors the sine-excitation FRF
 FRFs and Hilbert transforms: random excitation 417
0.00020
FRF (analytical)
HT (numerical)
HT (analytical)

0.00000
FRF Imaginary Part

-0.00020

-0.00040

-0.00060
-0.00040 -0.00020 0.00000 0.00020 0.00040
FRF Real Part

Figure 8.23. Comparison between the numerical estimate of the Hilbert transform and the
O(X 4 ) Volterra series estimate for the Duffing oscillator under sine excitation. (Note that
the Real and Imaginary axes do not have equal scales.)

in many respects, there are important differences. This section is this book’s
only real foray into the realm of random vibration. If the reader would like to
study the subject in more depth, [198] is an excellent example of an introductory
textbook. A considerably more advanced treatment can be found in [52], which
treats nonlinear random vibration amongst other topics.
There have been a number of related calculations over the years. The
simplest method of approximating an FRF for a nonlinear system is based on
equivalent linearization [54]. This approach estimates the parameters of the linear
system which is closest (in a statistical sense) to the original nonlinear system.
The FRF of the linearized system is computed. In [75], statistical linearization
was combined with perturbation analysis [68], in order to calculate the spectral
response of a Duffing oscillator to white noise excitation. (This is equivalent
to the FRF calculation up to a multiplicative constant.) It was shown that the
FRF exhibits a secondary peak at three times the natural frequency, a result
which is unavailable from statistical linearization alone. An approach based on
perturbation theory alone is described in [147] and the calculation is carried to
first order in the perturbation parameter. A number of studies of spectra have
appeared based on the use of the Fokker–Planck–Kolmogorov equation (FPK)
[55, 15, 137, 138, 284]. The latter two references actually examine the Duffing
oscillator system which is studied in the current work. Good representations of
the spectra were obtained; however, to the order of approximation pursued, the
approach was unable to explain the presence of the secondary peak described
418 The Volterra series and higher-order frequency response functions

earlier. An interesting approach to approximating the spectral response of a

Duffing oscillator is adopted in [184]. There, the expected response of an
equivalent linear system was calculated where the natural frequency of the
linear system was a random variable. The results compared favourably with
numerical simulation, but the secondary peak could not be obtained. The Volterra
series approach given here has been applied in [53] and [250], amongst others;
however, the calculation was not carried far enough to allow a description of
FRF distortions or the occurrence of the secondary peak. Using a Volterra
series approach also allows the definition of higher(polynomial)-order equivalent
systems, for example, the method of statistical quadratization is discussed in [78].

8.7.1 Volterra system response to a white Gaussian input

The problem of nonlinear system response to a generic random input is
completely intractable. In order to make progress, it is usually assumed that
the noise is white Gaussian. The power spectrum of such an input is constant
over all frequencies and, as a consequence, Gaussian white noise is a physically
unrealizable signal since it has infinite power. In practice, Gaussian white
noise is approximated by Gaussian random processes that have sufficiently broad
frequency bandwidth for the application of interest.
The definition of the FRF of a linear system based on the input/output
cross-spectrum, S yx (! ), and input auto-spectrum, S xx (! ), is well known (and
is repeated here for convenience)
S (!)
H (!) = H1 (!) = yx : (8.114)
Sxx(!)
The composite FRF,  r (! ), of a nonlinear system under random excitation
is defined similarly
S (! )
r (!) = yx : (8.115)
Sxx(!)
The term composite FRF is used again because  r (! ), for a nonlinear
system, will not in general be equal to H 1 (! ) but will receive contributions from
all Hn . It will be shown that random excitation leads to a different composite
FRF than sine excitation, hence the identifying subscript. The FRF also depends
on the power spectral density of the input. However,  r (! ) tends to the linear
FRF as the power spectral density of the excitation tends to zero.
In order to obtain a more detailed expression for  r (! ), an expression for
Syx (!) must be derived. Using the Volterra series representation given in (8.3)
results in the expression
S (!) + Sy2 x (!) +


+ Synx (!) +



r(!) = y1 x : (8.116)
Sxx (!)
r (!) will be approximated here by obtaining expressions for the various cross-
spectra between the input and the individual output components. First, consider
 FRFs and Hilbert transforms: random excitation 419

the cross-correlation function  y1 x ( ); this is defined by y1 x ( ) = E [y1 (t)x(t

 )] where E [ ] is the expected value operator. Substituting in the expression for
the first-order component of the Volterra series response from (8.4) gives
Z +1 
y1 x ( ) = E d1 h1 (1 )x(t 1 )x(t  ) : (8.117)
1
It is known that the operations of taking the expected value and integrating
commute, thus
Z +1
y1 x ( ) = d1 h1 (1 )E [x(t 1 )x(t  )]
1
Z +1
= d1 h1 (1 )xx ( 1 ) (8.118)
1
where xx ( ) is the input autocorrelation function defined by  xx ( ) =
E [x(t)x(t  )].
Taking Fourier transforms of both sides of this equation gives
Z +1 Z +1
Sy1 x(!) = d e i! d1 h1 (1 )xx ( 1 ) (8.119)
1 1
and, changing the order of integration, gives
Z +1 Z +1
Sy1 x (!) = d1 h1 (1 ) d e i! xx ( 1 ): (8.120)
1 1
Using the Fourier transform shift theorem yields
Z +1
Sy1 x(!) = d1 h1 (1 )e i!1 Sxx(!) = H1 (!)Sxx (!): (8.121)
1
The result is no more than the expression for the linear FRF as stated in (8.2).
However, the example serves to illustrate the methods used to obtain expressions
for the cross-spectra between the input and higher-order output components.
To obtain the S y2 x (! ) term the expression for the second-order component
of the Volterra series response is substituted into the equation  y2 x ( ) =
E [y2 (t)x(t  )]. Following a similar procedure as before gives
Z +1 Z +1 
y2 x ( ) = E d1 d2 h2 (1 ; 2 )x(t 1 )x(t 2 )x(t  )
1 1
Z +1 Z +1
= d1 d2 h2 (1 ; 2 )E [x(t 1 )x(t 2 )x(t  )]:
1 1
(8.122)
420 The Volterra series and higher-order frequency response functions

It can be shown [158] that for zero-mean Gaussian variables x 1 ; x2 ; : : : ;

xn ; : : :
E [x1 x2 : : : xn ] = 0 (8.123)
if n is odd and if n is even
XY
E [x1 x2 : : : xn ] = E [xi xj ] (8.124)
PQ
where means the sum of the products of E [x i xj ], the pairs xi xj being taken
from x1 ; x2 ; : : : ; xn in all the possible distinct ways.
It follows from (8.123) that all cross-correlation functions, and hence all
cross-spectra, between the input and the even-order output components will be
zero, i.e. Sy2n x (! ) = y2n x ( ) = 0, 8n.
Moving on to the S y3 x (! ) term, this is given by
Z +1 Z +1 Z +1
y3 x ( ) = d1 d2 d3 h3 (1 ; 2 ; 3 )
1 1 1
 E [x(t 1 )x(t 2 )x(t 3 )x(t  )]: (8.125)

From (8.124) the expected value of the product of inputs, i.e. the fourth-order
moment of the input, reduces to the following product of second-order moments,

E [x(t 1 )x(t 2 )x(t 3 )x(t  )]

= E [x(t 1 )x(t 2 )]E [x(t 3 )x(t  )]
+ E [x(t 1 )x(t 3 )]E [x(t 2 )x(t  )]
+ E [x(t 1 )x(t  )]E [x(t 2 )x(t 3 )]: (8.126)

Using this equation and taking advantage of the symmetry of the Volterra
kernels leads to
Z +1 Z +1 Z +1
y3 x ( ) = 3 d1 d2 d3 h3 (1 ; 2 ; 3 )
1 1 1
 E [x(t 1 )x(t 2 )]E [x(t 3 )x(t  )]
Z +1 Z +1 Z +1
=3 d1 d2 d3 h3 (1 ; 2 ; 3 )
1 1 1
 xx (2 1 )xx ( 3 ): (8.127)

Fourier transforming this equation and manipulating the result eventually

yields

3Sxx(!) +1
Z
Sy3 x (!) = d!1 H3 (!1 ; !1; !)Sxx(!1 ): (8.128)
2 1
This result is already available in the literature [25]. Its presence here is
justified by the fact that the derivation of the general term is a simple modification.
 FRFs and Hilbert transforms: random excitation 421

The general term is

(2n)!Sxx(!) +1
Z Z +1
Sy2n 1 x (!) = : : : d!1 : : : d!n 1
n! 2n (2)n 1 1 1
 H2n 1 (!1 ; !1 ; : : : ; !n 1 ; !n 1; !)
 Sxx(!1 ) : : : Sxx (!n 1 ): (8.129)

Now, given that the input autospectrum is constant over all frequencies for
a Gaussian white noise input (i.e. S xx (! ) = P ), the composite FRF for random
excitation follows. Substituting (8.129) into (8.116) gives
=1
nX
(2n)!P n 1 +1
Z Z +1
r (!) = n n 1 1 : : : 1 d!1 : : : d!n 1
n=1 n! 2 (2 )
 H2n 1 (!1 ; !1; : : : ; !n 1 ; !n 1; !): (8.130)

This equation will be used to analyse the effect of a Gaussian white noise input
on the SDOF Duffing oscillator system.

8.7.2 Random excitation of a classical Duffing oscillator

Using the theory developed in the last section, an expression for  r (!) up to
O(P 2 ) will be calculated for the standard system (8.49) with k 2 = 0. From
(8.130) the first three terms are given by
Sy1 x (!)
= H1 (!)
Sxx(!)
Sy3 x (!) 3P +1
Z
= d! H (! ; ! ; ! )
Sxx (!) 2 1 1 3 1 1
Sy5 x (!) 15P 2 +1 +1
Z Z
= d! d! H (! ; ! ; ! ; ! ; !):
Sxx(!) (2)2 1 1 1 2 5 1 1 2 2
(8.131)

The first term of this equation needs no further work but the others require
expressions for the HFRF terms as functions of the H 1 s and k3 . The results for
H3 and H5 are given in (8.87) and (8.89) respectively, the specific forms needed
for (8.131) are

H3 (!1 ; !1; !) = k3 H1 (!)2 H1 (!1 )H1 ( !1 ) = k3 H1 (!)2 jH1 (!1 )j2

(8.132)
and
3k2
H5 (!1 ; !1 ; !2 ; !2; !) = 3 H1 (!)2 H1 (!1 )H1 ( !1 )H1 (!2 )H1 ( !2 )
10
 f2H1(!) + H1 (!1 ) + H1 ( !1 ) + H1 (!2 ) + H1 ( !2 )
+ H1 (!1 + !2 + !) + H1 (!1 !2 + !)
422 The Volterra series and higher-order frequency response functions

+ H1 ( !1 + !2 + !) + H1 ( !1 !2 + !)g
3k2
= 3 H1 (!)2 jH1 (!1 )j2 jH1 (!2 )j2 f2H1 (!) + H1 (!1 ) + H1 ( !1)
10
+ H1 (!2 ) + H1 ( !2 ) + H1 (!1 + !2 + !) + H1 (!1 !2 + !)
+ H1 ( !1 + !2 + !) + H1 ( !1 !2 + !)g (8.133)
S 3 x (!)
So only one integral needs to be evaluated for Syxx (!) compared to nine for
Sy5 x (!)
Sxx (!) .
S 3 x (!)
Substituting (8.132) into the Syxx (!) term of (8.131) gives
3P k3H1 (!)2 +1
Z
Sy3 x (!)
= d!1 jH1 (!1 )j2 : (8.134)
Sxx (!) 2 1
This integral may be found in standard tables of integrals used for the
calculation of mean-square response, e.g. [198]. However, the analysis is
instructive and it will allow the definition of notation for the integrals which
follow.
Consider the common expression for the linear FRF (8.97). In terms of this,
2
the integral in (8.134), jH 1 (!1 )j may be written
1
jH1 (!1 )j2 = m2 (! p1 )(!1 p2 )(!1 p1 )(!1 p2 )
(8.135)
1
where p1 (respectively p2 ) is the complex conjugate of p 1 (respectively p2 ).
The integral is straightforwardly evaluated using the calculus of residues 5 . The
approach is well known and details can be found in numerous textbooks, e.g. [6].
The appropriate contour is given in figure 8.24. (In the calculation, the radius of
the semicircle is allowed to go to infinity.) The result of the calculation is the
standard
Z +1  
d! jH (!)j2 = = :
2m2 !n (!d2 +  2 !n2 ) ck1
(8.136)
1
Substituting this expression into (8.134) gives

Sy3 x(!) 3P k3 H1 (!)2

= (8.137)
Sxx(!) 2ck1
for this system. It can be seen that the third-order component of the response does
not change the position of the poles of  r (! ) from those of the linear FRF. In fact
it creates double poles at the positions where simple poles were previously found.
The effect of the next non-zero term will now be analysed.
5 Although the approach is straightforward, the algebraic manipulations involve rather complicated
expressions. In order to facilitate the analysis, a computer algebra package was used throughout this
work.
 FRFs and Hilbert transforms: random excitation 423

3 i ζ ωn

2 i ζ ωn

p p1
2
i ζ ωn α1
α2

-3 ω d -2 ω d - ωd ωd 2 ωd 3 ωd

- i ζ ωn
p* p1*
2

-2 i ζ ω n

-3 i ζ ω n

Figure 8.24. The pole structure of jH1 (!1 )j2 with the integration contour shown closed in
the upper-half of the !1 -plane.

S 5 x (!)
Substituting (8.133) into the Syxx (!) term of (8.131) gives the integrals
Z +1 Z +1
Sy5 x(!) 9P 2 k32 H1 (!)2

= 2 H1 (! ) d!1 d!2 jH1 (!1 )j2 jH1 (!2 )j2
Sxx(!) 8 2 1 1
Z +1 Z +1
+ d!1 d!2 H1 (!1 )jH1 (!1 )j2 jH1 (!2 )j2
1 1
Z +1 Z +1
+ d!1 d!2 H1 ( !1 )jH1 (!1 )j2 jH1 (!2 )j2
1 1
Z +1 Z +1
+ d!1 d!2 H1 (!2 )jH1 (!1 )j2 jH1 (!2 )j2
1 1
Z +1 Z +1
+ d!1 d!2 H1 ( !2 )jH1 (!1 )j2 jH1 (!2 )j2
1 1
Z +1 Z +1
+ d!1 d!2 H1 (!1 + !2 + !)jH1 (!1 )j2 jH1 (!2 )j2
1 1
Z +1 Z +1
+ d!1 d!2 H1 (!1 !2 + !)jH1 (!1 )j2 jH1 (!2 )j2
1 1
Z +1 Z +1
+ d!1 d!2 H1 ( !1 + !2 + !)jH1 (!1 )j2 jH1 (!2 )j2
1 1
Z +1 Z +1 
+ 2
d!1 d!2 H1 ( !1 !2 + !)jH1 (!1 )j jH1 (!2 )j : 2
1 1
(8.138)
424 The Volterra series and higher-order frequency response functions

However, this is not as daunting as it first appears; exchanging ! 1 for !1

in the second integral results in the expression
Z 1 Z +1
( d!1 ) d!2 H1 ( !1 )jH1 ( !1 )j2 jH1 (!2 )j2 : (8.139)
+1 1
Noting that jH1 ( !1 )j2 = jH1 (!1 )j2 results in the expression
Z +1 Z +1
d!1 d!2 H1 ( !1 )jH1 (!1 )j2 jH1 (!2 )j2 (8.140)
1 1
which is identical to the third integral. A similar argument shows that the fourth
and fifth integrals are identical to the second and third. It is also straightforward
to show that the last four integrals are identical to each other. This means that
(8.138) reduces to only three integrals, i.e.

Sy5 x (!) 9P 2 k32 H1 (!)3 +1 +1

Z Z
= d!1 d!2 jH1 (!1 )j2 jH1 (!2 )j2
Sxx(!) 42 1 1
9P 2k32 H1 (!)2
+

22
Z +1 Z +1
 Re d!1 d!2 H1 (!1 )jH1 (!1 )j2 jH1 (!2 )j2
1 1
Z +1 Z +1 
+ 2 2
d!1 d!2 H1 (!1 + !2 + !)jH1 (!1 )j jH1 (!2 )j :
1 1
(8.141)

The first integral may be evaluated straightforwardly as

+1 Z +1 +1 2
2
Z Z
d!1 d!2 jH1 (!1 )j2 jH1 (!2 )j2 = d!1 jH1 (!1 )j2 = 2 2
1 1 1 c k1
(8.142)
on making use of (8.136).
The second integral can also be factorized
Z +1 Z +1
Re d!1 d!2 H1 (!1 )jH1 (!1 )j2 jH1 (!2 )j2
1 1
Z +1 Z +1
= Re d!1 H1 (!1 )jH1 (!1 )j2 d!2 jH1 (!2 )j2 (8.143)
1 1
and the second integral is given by (8.136). Contour integration can again be used
to evaluate the first part. The integrand may be written
1
H1 (!1 )jH1 (!1 )j2 =
m3 (!1 p1 )2 (!1 p2 )2 (!1 p1 )(!1 p2 )
(8.144)
 FRFs and Hilbert transforms: random excitation 425

3 i ζ ωn

2 i ζ ωn
p p1
2
i ζ ωn Double pole

-3 ω d -2 ω d - ωd ωd 2 ωd 3 ωd
α2 α1
- i ζ ωn
p* p1*
2

-2 i ζ ω n

-3 i ζ ω n

Figure 8.25. The pole structure of H1 (!1 ) H1 (!1 ) 2 with the integration contour shown
j j
closed in the lower-half of the !1 -plane.

S 3 x (!)
and this has a similar pole structure to the Syxx (!) integrand, except the poles at
p1 and p2 are now double poles (figure 8.25). The contour should be closed in the
lower half of the ! 1 -plane as there are only simple poles there. The result is
Z +1  
Re d!1 H1 (!1 )jH1 (!1 )j2 = 2 =
3 2 2 2 2ck12
(8.145)
1 4m !n (!d +  !n )
and this combined with (8.136) yields
Z +1 Z +1 2
Re d!1 d!2 H1 (!1 )jH1 (!1 )j2 jH1 (!2 )j2 = :
2c2 k13
(8.146)
1 1
S 5 x (!)
The third and final integral of the Syxx (!) expression is slightly more
complicated as the integrand does not factorize.
Z +1 Z +1
I (!) = d!1 d!2 H1 (!1 + !2 + !)jH1 (!1 )j2 jH1 (!2 )j2
1 1
Z +1 Z +1
= d!2 jH1 (!2 )j 2 d!1 H1 (!1 + !2 + !)jH1 (!1 )j2 :
1 1
(8.147)
The second integral must be solved first. In terms of poles,
H1 (!1 + !2 + !)jH1 (!1 )j2
1
= 3
m (!1 p1 )(!1 p2 )(!1 p1 )(!1 p2 )(!1 q1 )(!1 q2 )
(8.148)
426 The Volterra series and higher-order frequency response functions

3 i ζ ωn

2 i ζ ωn

q2 p2 q1 p1
i ζ ωn
ω + ω2 ω + ω2

-3 ω d -2 ω d - ωd ωd 2 ωd 3 ωd
α2 α1
- i ζ ωn
p* p1*
2

-2 i ζ ω n

-3 i ζ ω n

Figure 8.26. The pole structure of H1 (!1 + !2 + !3 ) H1 (!1 ) 2 with the integration
j j
contour shown closed in the lower-half of the !1 -plane.

where p1 , p2 , p1 and p2 are the same as before and q1 and q2 are the poles of
H1 (!1 + !2 + !)
q1 = p1 ! !2 = (!d ! !2 ) + i!n
q2 = p2 ! !2 = ( !d ! !2 ) + i!n : (8.149)
For simplicity, the contour is again closed on the lower half of the ! 1 -plane
(figure 8.26). The result is
Z +1
d!1 H1 (!1 + !2 + !)jH1 (!1 )j2
1
= [ (! + !2 4i!n )][2m3 !n (!d2 +  2 !n2 )(! + !2 2i!n )
 (! + !2 + 2!d 2i!n )(! + !2 2!d 2i!n )] 1
= [ (! + !2 4i!n )][mck1 (! + !2 2i!n )
 (! + !2 + 2!d 2i!n )(! + !2 2!d 2i!n )] 1 : (8.150)
The expression above is then substituted into (8.147) and the integral over
!2 evaluated. The integrand is this expression multiplied by jH 1 (!2 )j2 . In terms
of poles it is
(! + !2 4i!n )
3
m ck1 (!2 p1 )(!2 p2 )(!2 p1 )(!2 p2 )(!2 q1 )(!2 q2 )(!2 r)
(8.151)
where p1 , p2 , p1 and p2 are as before and
q1 = ! + 2!d + 2i!n ; q2 = ! 2!d + 2i!n
 FRFs and Hilbert transforms: random excitation 427

3 i ζ ωn
q2 q1
r 2 i ζ ωn

ω ω p1 ω
p2
i ζ ωn

-3 ω d -2 ω d - ωd ωd 2 ωd 3 ωd
α2 α1
- i ζ ωn
p* p1*
2

-2 i ζ ω n

-3 i ζ ω n

Figure 8.27. The pole structure of equation (8.151) with the integration contour shown
closed in the lower-half of the !1 -plane.

r = ! + 2i!n : (8.152)

The contour is again closed on the lower half of the ! 2 -plane (figure 8.27).
Finally

I (!) = [ 2 (!2 3!d2 10i!n ! 27 2 !n2 )]

 [mc2 k12 (! !d 3i!n )(! + !d 3i!n )
 (! 3!d 3i!n )(! + 3!d 3i!n)] 1 : (8.153)

Substituting (8.153), (8.146) and (8.142) into (8.141) gives the overall expression
S 5 x (!)
for Syxx (!) :
Sy5 x (!) 9P 2k32 H1 (!)3 9P 2 k32 H1 (!)2 9P 2 k32 H1 (!)2
= + + I (! )
4c2 k12 4c2 k13 22
(8.154)
Sxx (!)
for the classical Duffing oscillator. This equation shows that the first two terms
do not affect the position of the poles of the linear system. The first term does,
however, introduce triple poles at the position of the linear system poles. The term
of greatest interest, though, is the final one which has introduced four new poles
at

!d + 3i!n ; !d + 3i!n ; 3!d + 3i!n ; 3!d + 3i!n : (8.155)

The pole structure to this order is shown in figure 8.28. The poles at 3! d +3i!n
explain the secondary observed peak at three times the resonant frequency in the
output spectra of nonlinear oscillators [285].
428 The Volterra series and higher-order frequency response functions

3 i ζ ωn

2 i ζ ωn

i ζ ωn

Figure 8.28. The pole structure of the first three terms of r (! ) for the classical Duffing
oscillator.

FRF Distortion (k2=0,k3=5x10^9)

0.0006

P = 0.0 (Linear System)

P = 0.01
P = 0.02
0.0005

0.0004
FRF Magnitude

0.0003

0.0002

0.0001

0.0000
50 60 70 80 90 100 110 120 130 140 150
Circular Frequency (rad/s)

Figure 8.29. Composite FRF r (! ) to order O(P 2 ) for the classical Duffing oscillator.

Combining (8.154) and (8.137) into (8.131) yields an expression for the
composite FRF r (! ) up to O(P 2 ). The magnitude of the composite FRF is
plotted in figure 8.29 for values of P equal to 0 (linear system), 0.01 and 0.02.
The Duffing oscillator parameters are m = 1, c = 20, k 1 = 104 and k3 = 5  109.
 FRFs and Hilbert transforms: random excitation 429

As observed in practice, the resonant frequency shifts upwards with increasing P ,

while the peak magnitude reduces. This is an encouraging level of agreement
with experiment given that only three terms are taken in the expansion (8.116).
Although the frequency shifts, there is no real evidence of the sort of distortions
observed during stepped-sine testing. This lends support to the view that random
excitation produces a ‘linearized’ FRF.
It is significant that all the poles are located in the upper half of the ! -plane.
It is known (e.g. see chapter 5) that applying the Hilbert transform test to a system
with all its poles in the upper half of the complex plane results in the system
being labelled linear. If this behaviour continues for higher terms this then shows
agreement with apparent linearization of FRFs obtained under random excitation.
In order to determine whether or not the inclusion of further terms in
the r (! ) approximation results in further poles arising at new locations,
S 7 x (!)
the fourth non-zero term (i.e. Syxx (!) ) for this system was considered. The
H7 (!1 ; !1; !2 ; !2; !3 ; !3; !) expression consists of 280 integrals when the
problem is expressed in H 1 terms. However repeating the procedure of combining
terms which yield identical integrals results in 13 integrals. The expression for
Sy7 x (!)
Sxx (!) is

27P 3k33 H1 (!)4 +1 +1 +1

Z Z Z
Sy7 x(!)
= d!1 d!2 d!3
Sxx(!) 83 1 1 1
 jH1 (!1 )j2 jH1 (!Z
2 )j2 jH1 (!3 )j2
27P 3k33 H1 (!)3 +1 Z +1 Z +1
d!1 d!2 d!3
23 1 1 1
 H1 (!1 )jH1 (!1 )j2 jH1 (!2 )j2 jH1 (!3 )j2
Z +1 Z +1 Z +1
+ d!1 d!2 d!3
1 1 1 
 H1 (! + !1 + !2 )jH1 (!1 )j2 jH1 (!2 )j2 jH1 (!3 )j2
 Z +1 Z +1 Z +1
9P 3k33 H1 (!)2
6 d!1 d!2 d!3
83 1 1 1
 H1 (!1 )2 jH1 (!1 )j2 jH1 (!2 )j2 jH1 (!3 )j2
Z +1 Z +1 Z +1
+3 d!1 d!2 d!3 jH1 (!1 )j4 jH1 (!2 )j2 jH1 (!3 )j2
1 1 1
Z +1 Z +1 Z +1
+ 12 d!1 d!2 d!3 H1 (!1 )H1 (!2 )
1 1 1
 jH1 (!1 )j2 jH1 (!2 )j2 jH1 (!3 )j2
Z +1 Z +1 Z +1
+ 24 d!1 d!2 d!3 H1 (!1 )H1 (! + !1 + !2 )
1 1 1
430 The Volterra series and higher-order frequency response functions

 jH1 (!1 )j2 jH1 (!2 )j2 jH1 (!3 )j2

Z +1 Z +1 Z +1
+ 12 d!1 d!2 d!3 H1 (!1 )H1 (!1 + !2 + !3)
1 1 1
 jH1 (!1 )j2 jH1 (!2 )j2 jH1 (!3 )j2
Z +1 Z +1 Z +1
+6 d!1 d!2 d!3 H1 (! + !1 + !2 )2
1 1 1
 jH1 (!1 )j2 jH1 (!2 )j2 jH1 (!3 )j2
Z +1 Z +1 Z +1
+ 24 d!1 d!2 d!3 H1 (! + !1 + !2)
1 1 1
 H1 (! + !1 + !3 )jH1 (!1 )j2 jH1 (!2 )j2 jH1 (!3 )j2
Z +1 Z +1 Z +1
+ 12 d!1 d!2 d!3 H1 (! + !1 + !2)
1 1 1
 H1 (!1 + !2 + !3 )jH1 (!1 )j2 jH1 (!2 )j2 jH1 (!3 )j2
Z +1 Z +1 Z +1
+ 12 d!1 d!2 d!3 H1 (! + !1 + !2)
1 1 1
 H1 ( !1 !2 + !3 )jH1 (!1 )j2 jH1 (!2 )j2 jH1 (!3 )j2
Z +1 Z +1 Z +1
+2 d!1 d!2 d!3 jH1 (!1 + !2 + !3 )j2
1 1 1 
 jH1 (!1 )j2 jH1 (!2 )j2 jH1 (!3 )j2 : (8.156)

The evaluation of these integrals was carried out as before. However, the
results will not be included here. The important point is whether the calculation
introduces new poles. By extrapolating from the poles in the first three terms of
r (!), a fourth term might be expected to introduce poles at
!d + 5i!n ; !d + 5i!n ; 3!d + 5i!n ;
3!d + 5i!n ; 5!d + 5i!n ; 5!d + 5i!n: (8.157)
However, after evaluating the integrals in (8.156), again by contour integration,
it was found that no new poles arose. Instead, three of the integrals resulted
in simple poles at the locations given in equation (8.155), whilst another three
integrals resulted in double poles at these locations.
Due to the rapidly increasing level of difficulty associated with the addition
S 9 x (!)
of further terms to  r (! ) it was not possible to completely examine the Syxx (!)
term. It was possible, however, to consider one integral which would be included
S 9 x (!)
in the overall expression. The expression for Syxx (!) is given by
Sy9 x (!) 945P 4 +1 +1 +1 +1
Z Z Z Z
= d! d! d! d!
Sxx(!) (2)4 1 1 1 1 1 2 3 4
 H9 (!1 ; !1; !2 ; !2; !3 ; !3; !4 ; !4 ; !): (8.158)
 Validity of the Volterra series 431

One of the possible HFRF products which make up H 9 (!1 ; !1 ; !2 ; !2 ; !3 ;

!3 ; !4; !4 ; !) is H5 (!1 ; !2 ; !3; !4 ; !)H3 ( !1 ; !2 ; !3)H1 ( !4). This
in turn results in several integrals, one of which is given by

27P 4k34 H1 (!)2 +1 +1 +1 +1

Z Z Z Z
d!1 d!2 d!3 d!4
1284 1 1 1 1
 H1 (! + !1 + !2 + !3 + !4 )H1 (! + !1 + !2 )H1 ( !1 !2 !3 )
 jH1 (!1 )j2 jH1 (!2 )j2 jH1 (!3 )j2 jH1 (!4 )j2 : (8.159)

This integral was evaluated as before and was found to have triple poles at
the locations given in equation (8.155), simple poles at the locations given in
equation (8.157) and also at the following locations

!d + 7i!n ; !d + 7i!n ; 3!d + 7i!n ; 3!d + 7i!n ;

5!d + 7i!n ; 5!d + 7i!n ; 7!d + 7i!n ; 7!d + 7i!n : (8.160)

Although it is possible that these contributions cancel when combined with other
S 9 x (!)
integrals from Syxx (!) , it can be conjectured that including all further terms would
result in FRF poles for this system being witnessed at all locations a! d + bi!n
where a  b are both odd integers. Note that this implies the existence of an
infinite sequence of FRF peaks, each one associated with an odd multiple of the
natural frequency.
The restriction of a and b to the odd integers might be expected from the fact
that only odd-order HFRFs exist for systems which only contain odd polynomial
nonlinearities. In that case the introduction of an even nonlinearity results in both
odd- and even-order HFRFs. This is discussed in appendix K where a Duffing
oscillator with additional quadratic spring stiffness is considered. Appendix K
also shows an extension of the analysis to a simple MDOF system.
An interesting feature of this analysis is that the multiplicity of the poles
increases with order P . The implication is that the poles will become isolated
essential singularities in the limit. This implies rather interesting behaviour in the
!-plane near the poles as Picard’s theorem [6] asserts that the FRF will take all
complex values in any neighbourhood of the poles regardless of the size of the
neighbourhood. This behaviour need not, of course, be visible from the real line.

8.8 Validity of the Volterra series

The question of validity is often ignored when the Volterra series is used. Not
only should the series exist, but it should converge. Both of these requirements
have been the subject of intense study over the years.
The first condition of importance for the Volterra series is the existence
condition. When does an input–output functional S , as in y (t) = S [x(t)], admit a
Volterra representation of the form (8.3)? The definitive work on this question can
432 The Volterra series and higher-order frequency response functions

be found in [200], and requires fairly advanced techniques of functional analysis

which are beyond the scope of this book. The important conclusion from [200] is
that ‘the class of Volterra-like representable systems is only restricted by a kind
of smoothness condition’. ‘Smoothness’ here means sufficiently differentiable.
Now strictly speaking, the derivatives referred to are Fréchet derivatives of
the functional S , but this translates easily onto a smoothness condition on the
nonlinear functions in the equation of motion. For the purposes of this book,
which is almost exclusively concerned with polynomial nonlinearities (which are,
of course, infinitely differentiable), the Volterra series will always exist. The
review [201] is a less rigorous but more readable introduction to questions of
existence.
As the Volterra series is an infinite series, establishing existence immediately
raises the vexed question of convergence. Namely, for what range of inputs x(t)
does the series converge. A concrete example for illustration is available now:
the calculation of the composite FRF in section 8.8 assumed values for the input
spectral density P . The convergence of the series for these P will be examined
here.
General convergence results are few and far between. Ku and Wolf [156]
established some useful results. For a bounded-input deterministic system, i.e. if

jx(t)j  K; for all t (8.161)

they showed that

1
X 1
X
yi (t)  an K n (8.162)
i=1 i=1
where Z 1 Z 1
an = ::: d1 : : : dn h(1 ; : : : ; n ) (8.163)
1 1
and the radius of convergence of the Volterra series is given by
  1
!1 jai j
R = ilim :
1
i (8.164)

For the random case,Pthey established that if x(t) is stationary with bounded
1
statistical moments then i=1 yi converges in the mean if
n
X
lim ak < 1
n!1
(8.165)
k=1
and
lim a = 0:
k!1 k
(8.166)

Fortunately, in order to validate the results of section 8.8, it is not necessary

to use the general theory as results can be obtained for the classical Duffing
 Validity of the Volterra series 433

oscillator using the criteria developed by Barrett [21] 6 . The first step is to convert
the equation of motion (8.49) with k 2 = 0, to the normalized form
0
y0 + 2 y_ 0 + y0 + y 3 = x0 (t0 ) (8.167)

and this is accomplished by the transformation

x
y0 = !n2 y; x0 = ; t0 = !n t (8.168)
m
so that
k3
=
m!n6
(8.169)

and  has the usual definition. Once in this coordinate system, convergence of the
Volterra series is ensured as long as [21]

ky0k < yb0 = p 1 (8.170)

3H
where

H = coth p :
1 2
(8.171)

The norm ky 0 k on an interval of time is simply the maximum value of y 0 over

that interval. Using the values of section 8.4, m = 1, c = 20 and k 1 = 104 ,
k3 = 5  109. The value of yb0 obtained is 4:514. This translates into a physical
bound yb = 4:514  10 4 .
Now, the mean-square response,  yl , of the underlying linear system (k 3 =
0) is given by the standard formula
Z 1 Z 1 P
y2l = d! jH (!)j2 Sxx(!) = P d! jH (!)j2 = (8.172)
1 1 ck1
and in this case, yl = 3:963 32  10 4 if P = 0:01 and yl = 5:604 99  10 4
if P = 0:02. However, these results will be conservative if a non-zero k 3 is
assumed. In fact for the nonlinear system [54, 68]

y2nl = y2l 3 y2l (8.173)

to first order in = k3 =k1 . If this result is assumed valid ( is by no

means small), the mean-square response of the cubic system can be found. It
is ynl = 3:465  10 4 when P = 0:01 and ynl = 4:075  10 4 when
P = 0:02. In the first case, the Barrett bound is 1.3 standard deviations and
in the second case it is 1.11 standard deviations. Thus, using standard tables for
Gaussian statistics [140], it is found that the Volterra series is valid with 80.6%
6 Barrett derives his convergence criterion using a recursion relation; an alternative proof using
functional analysis can be found in [63].
434 The Volterra series and higher-order frequency response functions

yb
mb

fd ks

yt
mt

kt
y0

Figure 8.30. 2DOF automotive suspension lumped-mass model.

confidence if P = 0:01 and with 73:3 confidence if P = 0:02. As the Barrett

bound is known to be conservative [173], these results were considered to lend
support to the assumption of validity for the Volterra series.
For practical purposes, establishing convergence may not actually help, as
the important question is how many terms of the series need to be included in
order to obtain a required accuracy in the representation?. Not surprisingly, there
has been little success in establishing an answer to this question. Some discussion
can be found in [282] and [254].

8.9 Harmonic probing for a MDOF system

Having established the harmonic probing techniques needed for SDOF systems,
the study can now proceed to more realistic MDOF models. A lumped-mass
model of an automotive suspension (figure 8.30) [115] will be used to illustrate
the techniques. Note that when a MDOF system is considered, there are two
possible positions for the nonlinearities, between two of the masses and between
a mass and ground; both cases will be discussed here.
The equations of motion of the system are

mtyt + fd(y_t y_b) + (kt + ks )yt ks yb = kt y0

mb yb fd(y_t y_b) + ks (yb yt ) = 0 (8.174)

where mt , kt and yt are the mass, stiffness and displacement of the tyre—the
unsprung mass. m b and yb are the mass and displacement of the body or the
sprung mass (this is usually taken as one-quarter of the total car body mass).
fd is the characteristic of the nonlinear damper and k s is the (linear) stiffness
 Harmonic probing for a MDOF system 435

of the suspension. A cubic characteristic will be assumed for the damper, i.e.
fd (z ) = c1 z + c2 z 2 + c3 z 3 . y0 is the displacement at the road surface and acts
here as the excitation.
Each of the processes y 0 ! yt and y0 ! yb has its own Volterra series
and its own HFRFs and it will be shown later that the responses will depend on
the kernel transforms from both series. The notation is defined by

1 1
Z
Yt (!) = H1t (!)Y0 (!) + d! H t (! ; ! !1 )Y0 (!1 )Y0 (! !1 )
(2) 1 1 2 1
Z 1Z 1
1
+ d! d! H t (! ; ! ; ! !1 !2 )
(2)2 1 1 1 2 3 1 2
 Y0 (!1 )Y0 (!2 )Y0 (!Z !1 !2 ) +


(8.175)
1 1
Yb (!) = H1b(!)Y0 (!) + d! H b (! ; ! !1 )Y0 (!1 )Y0 (! !1 )
(2) 1 1 2 1
Z 1Z 1
1
+ d! d! H b (! ; ! ; ! !1 !2 )
(2)2 1 1 1 2 3 1 2
 Y0 (!1 )Y0 (!2 )Y0 (! !1 !2 ) +


: (8.176)

The harmonic probing procedure for MDOF systems is a straightforward

extension of that for SDOF systems. In order to obtain the first-order (linear)
kernel transforms, the probing expressions

y0p = ei t (8.177)
ytp = H1t ( )ei t +


(8.178)

and
ybp = H1b( )ei t +


(8.179)
are used. These expressions are substituted into the equations of motion (8.174),
and the coefficients of e i t are extracted as before. The resulting equations are, in
matrix form
    
mt 2 + ic1 + kt + ks i c1 ks H1t ( ) = kt :
ic1 ks mb 2 + ic1 + ks H1b( ) 0
(8.180)
This 2  2 system has a straightforward solution:

kt ( mb!2 + ic1 ! + ks )
H1t (!) =
( mt !2 + ic1 ! + kt + ks )( mb !2 + ic1 ! + ks ) (ic1 ! + ks )2
(8.181)
kt (ic1 ! + ks )
H1b (!) = :
( mt ! + ic1 ! + kt + ks )( mb !2 + ic1 ! + ks ) (ic1 ! + ks )2
2
(8.182)
436 The Volterra series and higher-order frequency response functions

It will prove useful later to establish a little notation:


mt !2 + ic1 ! + kt + ks ic1 ! ks
 1 1 
(!) = ic1 ! ks mb !2 + ic1 ! + ks 1 :
(8.183)
The second-order kernel transforms are obtained using the probing
expressions

y0p = ei 1 t + ei 2 t (8.184)
ytp = H1t ( 1 )ei 1 t + H1t ( 2 )ei 2 t + 2H2t( 1 ; 2 )ei( 1 + 2 )t +


(8.185)

and

ybp = H1b( 1 )ei 1 t + H1b ( 2 )ei 2 t + 2H2b( 1 ; 2 )ei( 1 + 2 )t +


: (8.186)

These expressions are substituted into the equations of motion (8.174) and
the coefficients of e i( 1 + 2 )t are extracted. The resulting equations are, in matrix
form
 
mt( 1 + 2 )2 + ic1 ( 1 + 2 ) + kt + ks ic1( 1 + 2 ) ks
ic1 ( 1 + 2 ) ks mb( 1 + 2 )2 + ic1 ( 1 + 2 ) + ks
 t 
 HH2b(( 11;; 22))

1

= 11 c2 1 2 [H1t ( 1 ) H1b( 1 )][H1t ( 2 ) H1b( 2 )] (8.187)

so
 
H2t (!1 ; !2 ) = (! + ! )c ! ! [H t (! ) H b (! )][H t (! ) H b (! )]:
H1b (!1 ; !2 ) 1 2 2 1 2 1 1 1 1 1 2 1 2
(8.188)
The calculation of the third-order kernel transforms proceeds as before,
except a three-tone probing expression is used:

y0p = ei 1 t + ei 2 t + ei 3 t : (8.189)

The result of the computation is

 
H3t (!1 ; !2 ; !3 ) = (! + ! + ! )(F (! ; ! ; ! ) + G(! ; ! ; ! ))
H3b (!1 ; !2 ; !3 ) 1 2 3 1 2 3 1 2 3
(8.190)
where
2 X
F (!1 ; !2 ; !3 ) = c2 !1(!2 + !3 )[H1t (!1 ) H1b(!1 )]
3 C
 [H2t (!2 ; !3 ) H2b(!2 ; !3 )] (8.191)
 Harmonic probing for a MDOF system 437

yb
mb

fd
ks

yt
mt

kt
y0

Figure 8.31. 2DOF skyhook model of the automotive suspension.

P
where C denotes a sum over cyclic permutations of ! 1 , !2 and !3 . Also
3
Y
G(!1 ; !2; !3 ) = ic3 !n[H1t (!n ) H1b(!n )]: (8.192)
n=1
Note that to obtain Y t (! ) (respectively Yb (! )) requires a specification of H 3t
(respectively H3b ) and H1t , H1b , H2t and H2b .
This theory is easily adapted to the case of the ‘sky-hook’ suspension model
(figure 8.31) [128]. The equations of motion are
mt yt + (kt + ks )yt ks yb = kt y0
mb yb + fd (y_b) + ks (yb yt ) = 0: (8.193)
The H1 functions for this system are given by
kt ( mb!2 + ic1 ! + ks )
H1t (!) =
( mt !2 + kt + ks )( mb!2 + ic1 ! + ks ) ks2
(8.194)

kt + ks
H1b(!) = :
( mt !2 + kt + ks )( mb !2 + ic1 ! + ks ) ks2
(8.195)

The H2 functions are given by

 t 
H2 (!1 ; !2 ) = (! + ! )c ! ! H b(! )H b (! )
H1b(!1 ; !2) 1 2 2 1 2 1 1 1 2 (8.196)

where

mt !2 + kt + ks ks
 1 0
(!) = ks mb!2 + ic1 ! + ks 1 : (8.197)
438 The Volterra series and higher-order frequency response functions

Finally, the H3 s are given by

 
H3t (!1 ; !2; !3 ) = (! + ! + ! )(J (! ; ! ; ! ) + K (! ; ! ; ! ))
H3b (!1 ; !2 ; !3) 1 2 3 1 2 3 1 2 3
(8.198)
where
2 X
J (!1 ; !2; !3 ) = c2 !1 (!2 + !3 )H1b (!1 )H2b (!2 ; !3 ) (8.199)
3 C
and
3
Y
K (!1 ; !2 ; !3 ) = ic3 !n H1b(!n ): (8.200)
n=1

8.10 Higher-order modal analysis: hypercurve fitting

Linear modal analysis was discussed briefly in chapter 1. The philosophy of the
approach was to extract the modal parameters for a given linear system by curve-
fitting in the time or frequency domain. For a given linear system, this is an
exercise in nonlinear optimization as the parameters of the system do not enter
into the FRF expression in a linear manner, i.e.
N
X Ai
H1 (!) = 2 2 : (8.201)
i=1 (! !ni ) + 2ii !ni ! + k
A remarkable observation due to Gifford [110] is that for a nonlinear system,
the nonlinear parameters are much easier to obtain from the HFRFs than the linear
parameters are from the H 1 . To illustrate this, consider an asymmetrical Duffing
oscillator as in (8.49), and assume for the moment that the FRFs are measurable.
The linear parameters m, c and k can be obtained from the H 1 :
1
H1 (!) =
m!2 + ic! + k
(8.202)

using whatever curve-fitter is available. Now, if one considers H 2 :

H2 (!1 ; !2 ) = k2 H1 (!1 )H1 (!2 )H1 (!1 + !2 ): (8.203)

The quadratic stiffness coefficient k 2 enters as a linear multiplier for the product
of H1 s. Most importantly, H 1 is now known as the linear parameters have been
identified at the first stage. This useful property also holds for the H 3 , which has
the form

H3 (!1 ; !2 ; !3 ) = 1 H1 (!1 + !2 + !3 )f4k2 (H1 (!1 )H2 (!2 ; !3 )

6
+ H1 (!2 )H2 (!3 ; !1 ) + H1 (!3 )H2 (!1 ; !2 ))
+ 6k3 H1 (!1 )H1 (!2 )H1 (!3 )g (8.204)
 Higher-order modal analysis: hypercurve fitting 439

or
H3 (!1 ; !2; !3 ) = k2 F1 [H1 ; H2 ] + k3 F2 [H1 ] (8.205)

so the H3 function is also linear in the parameters k 2 and k3 . The importance of

these observations is clear. Once H 1 and the linear parameters have been obtained
from a nonlinear optimization step, the nonlinear parameters for the system can
be obtained by linear least-squares analysis of the HFRFs.
The basis for Gifford’s identification procedure is the general formulae for
H2 and H3 for a general N DOF lumped-mass system. These are stated here
without proof, the derivation can be found in [110].

N
X
H2rs (!1 ; !2 ) = (!1 !2 c2mm k2mm )H1sm (!1 + !2 )H1rm (!1 )H1rm(!2 )
m=1
NX1 X
N
(!1 !2c2mn k2mn )[H1sm (!1 + !2 ) H1sn (!1 + !2 )]
m=1n=m+1
 [H1rm(!1 ) H1rn (!1 )][H1rm (!2 ) H1rn(!2 )] (8.206)

where c2mn is the quadratic velocity coefficient for the connection between mass
m and mass n (c2mm is the connection to ground). k 2mn is the corresponding
stiffness coefficient. H2rs etc. refer to the FRFs between DOF r and DOF s.
Note that despite the considerable increase in complexity, the expressions are still
linear in the parameters of interest. H 3 is more complex still:

N
2X
H3rs (!1 ; !2 ; !3 ) = H sm (! + !2 + !3 )
3 m=1 1 1
 [[!1 (!2 + !3 )c2mm k2mm ]H1rm (!1 )H2rm (!2 ; !3 )
+ [!2 (!3 + !1 )c2mm k2mm ]H1rm (!2 )H2rm(!3 ; !1)
+ [!3 (!1 + !2 )c2mm k2mm ]H1rm (!3 )H2rm(!1 ; !2)]
2 NX1 X N
[H sm (! + !2 + !3 ) H1sn (!1 + !2 + !3 )]
3 m=1n=m+1 1 1
 [[!1 (!2 + !3 )c2mn k2mn ][H1rm (!1 ) H1rn(!1 )]
 [H2rm (!2 ; !3 ) H2rn(!2 ; !3 )][!2 (!3 + !1 )c2mn k2mn ]
 [H1rm (!2 ) H1rn(!2 )][H2rm(!3 ; !1 ) H2rn (!3 ; !1 )]
 [!3 (!1 + !2 )c2mn k2mn ][H1rm (!3 ) H1rn(!3 )]
 [H2rm (!1 ; !2 ) H2rn(!1 ; !2 )]]
N
X
+ (i!1 !2 !3 c3mm k3mm )
m=1
 H sm (!
1 1 + !2 + !3 )H1rm(!1 )H1rm (!2 )H1rm (!3 )
440 The Volterra series and higher-order frequency response functions
NX1 X
N
(i!1 !2 !3 c3mn k3mn )
m=1n=m+1
 [H1sm (!1 + !2 + !3 ) H1sn (!1 + !2 + !3 )]
 [H rm (! ) H rn(! )][H rm(! ) H rn (! )][H rm (!
1 1 1 1 1 2 1 2 1 3 ) H1rn (!3 )]
(8.207)

and again, despite the complexity of the expression, the coefficients c 2mn , k2mn ,
c3mn and k3mn enter the equation in a linear manner.
In principle then, the availability of measurements of the HFRFs up to order
n would allow the identification of the parameters in the differential equations of
motion up to the nth-power terms. The problem is to obtain the HFRFs accurately
and without bias. Gifford used a random excitation test which is described in the
next section along with an illustrative example.

8.10.1 Random excitation

The basic premise of Gifford’s testing procedure is that at low excitations, the
HFRFs can be approximated by certain correlation functions 7. For the first-order
FRF, this is confirmed by (8.130) where
Syx (!)
= r1 (!) = H1 (!) + O(P ): (8.208)
Sxx(!)
This type of relation holds true for the higher-order FRFs also. It can be
shown by the methods of section 8.7 that
Sy0 xx (!1 ; !2 )
= r2(!1 ; !2)
Sxx (!1 )Sxx(!2 )
 n2 1 Z 1
P
d!(1) : : : d!( 2 1)
n
= n
2 1
 Hn (!1 ; !2 ; ! ; !(1); : : : ; !( n2 1) ; !( n2 1) )
(1)
(8.209)

where y 0 is the output signal with the mean removed and


n(n 1)(n 3) : : : 1; n even
n = 0; n odd. (8.210)

Recall, that (8.208) assumes that the input x(t) is white Gaussian. This
is also an assumption used in the derivation of (8.209). The most important
7 The arguments of this section do not directly follow [110]. The analysis there makes extensive
use of the Wiener series, which is essentially an orthogonalized version of the Volterra series. The
important facts which are needed here are simply that the correlations used approach the HFRFs in
the limit of vanishing P and these can be demonstrated without recourse to the Wiener series.
 Higher-order modal analysis: hypercurve fitting 441

Charge
Amplifier
Accelerometer

m1 m2 m3
Nonlinear
Circuit

Power
Amplifier
Excitation Damper

Figure 8.32. General arrangement of experimental nonlinear beam rig.

consequence of (8.209) is that

Sy0 xx (!1 ; !2 )
= H2 (!1 ; !2 ) + O(P ) (8.211)
Sxx (!1 )Sxx (!2 )
so as P ! 0, the correlation function tends towards H 2 .
The theory described here was validated by experiment in [110]. The
structure used for the experiment was a forerunner of the beam used in
section 7.5.3 and is shown in figure 8.32. The main difference between this and
the rig used to validate the restoring force surface approach is that Gifford’s beam
was fixed–fixed and the nonlinear damping force was applied to the third lumped
mass instead of the second. Figure 8.33 shows a first-order FRF for the system
(the feedback circuit which provides the nonlinear damping force is switched off),
the fourth mode is sufficiently far removed from the first three to make this a
credible 3DOF system as the excitation force is band-limited appropriately.
Figure 8.34 shows the form of the feedback circuit used to provide the cubic
velocity force on mass m 3 . After collecting input v 1 , and output voltage v 2 , data
from the nonlinear circuit alone, a polynomial curve-fit established the circuit
characteristics
v2 = 1:34v1 + 1:25v12 + 0:713v13: (8.212)
The overall gain of the feedback loop was obtained by measuring the FRF
between input v 1 and output v 4 when the circuit was in linear mode; this is
shown in figure 8.35. This FRF is very flat at all the frequencies of interest.
Using the gain from the plot and all the appropriate calibration factors from
the instrumentation, the linear force–velocity characteristics of the nonlinear
feedback loop were obtained as

Ff = 120y_ 3 (8.213)
442 The Volterra series and higher-order frequency response functions

Figure 8.33. First-order frequency response function for experimental nonlinear beam
structure under low-level random excitation.

so from (8.213), it follows that the nonlinear characteristic of the loop is

Ff = 120y_ 3 3540y_ 32 638 000y_ 33: (8.214)

When the experiment was carried out, the level of excitation was set low
enough for distortion effects on the H 1 and H2 to be minimal. Figure 8.36 shows
a driving-point FRF for the system with and without the nonlinear part of the
circuit switched in; the effect of the nonlinearity is apparently invisible.
In order to estimate the H 2 functions, the autocorrelation function

y0 xx(1 ; 2 ) = E [y0 (t)x(t + 1 )x(t + 2 )] (8.215)

was estimated by an averaging process and then Fourier transformed. The

resulting HFRFs are shown in figure 8.37 for the processes x ! y 1 , x ! y2
and x ! y3 . Altogether, 286 000 time samples were used to form these
estimates. This is because a substantial number of averages are needed to
adequately smooth the HFRFs. Having said this it may be that ‘smoothness’ from
a visual point of view may not be critical; it depends on the noise tolerance of the
curve-fitting procedure. Also if a sufficient number of H 2 points are sampled, it
may be possible to tolerate a higher degree of noise.
Before the LS procedure for fitting the quadratic terms can be carried out,
the H1 functions for the system must be determined by the nonlinear curve-fitting
procedure. Figure 8.38 shows the three first-order FRFs (in real and imaginary
 Higher-order modal analysis: hypercurve fitting 443

Accelerometer
& v1 = 31.6 y
Charge Amplifier

v1

Nonlinear
Circuit

v2

Power
Amplifier

v3

Shaker

F
Force Link
& v4 = .316 F
Charge Amplifier

v4

Figure 8.34. Block diagram of all the elements that make up the linear damping system
with all their calibration factors in the configuration in which the results were measured.

form), together with the results of a standard modal analysis global curve-fitter;
the results are very close indeed, even by linear system standards.
In order to carry out the LS step which fits the model (8.206), it is necessary
to select the data from the (! 1 ; !2 )-plane. By analogy with common H 1 curve-
fitters, the data are chosen from around the peaks in order to maximize the signal-
to-noise ratio. Figure 8.39 shows how the data is selected to fit H 2 .
Gifford fitted three models to the H 2 measurements: one assuming a
quadratic stiffness characteristic, one assuming quadratic damping and one
assuming both. Figure 8.40 shows the H 2 reconstruction from the two ‘pure’
444 The Volterra series and higher-order frequency response functions

Figure 8.35. A measured FRF of v4 over v1 , see figure 8.34.

models compared to the measured data. It is clear that the damping model (on the
right) is the appropriate one. The coefficients obtained from a local curve-fit to
the second FRF H212 were as given in table 8.2.
By far the dominant contribution comes from the c 233 term as expected. The
estimated value compares well with the calibration value of 3540 ( 15.8% error).
When a global LS fit was carried out using data from all three H 2 curves at once,
the coefficient estimate was refined to 3807, giving a percentage error of 10.3%.
The results of this exercise validate Gifford’s method and offer a potentially
powerful means of using the Volterra series and HFRFs for parametric
identification.

8.10.2 Sine excitation

Gifford’s thesis inspired Storer [237] to make a number of useful observations.

First and foremost, he noted that all the structural parameters of interest appear in
the diagonal FRFs and it is therefore sufficient to curve-fit to them 8 . To illustrate,
consider the Duffing oscillator again, the diagonal forms of (8.203) and (8.204)
8 There is an implicit assumption throughout here that the FRFs contain enough modal information
to completely characterize the system, given an assumed model order. The methods may fail if the
FRFs are too severely truncated.
 Higher-order modal analysis: hypercurve fitting 445

Figure 8.36. First-order FRF from point 1 of the nonlinear rig in both its linear and
nonlinear modes.

can easily be found:

H2 (!; !) = k2 H1 (!)2 H1 (2!) (8.216)

H3 (!; !; !) = 61 H1 (3!)(12k2H1 (!)H2 (!; !) + 6k3 H1 (!)3 ): (8.217)
446 The Volterra series and higher-order frequency response functions

Figure 8.37. Contour plots of the moduli of the three second-order HFRFs of the nonlinear
rig operating in its nonlinear mode.

The equivalent forms of (8.206) and (8.207) for a general N DOF system
form the basis of Storer’s approach to identification discussed in detail in
[237, 238, 253] 9 .
Storer’s experimental procedures were based on sinusoidal testing and he
made essentially the same assumptions as Gifford regarding the need for low
excitation levels. In the case of a SDOF system, equation (8.86) confirms that

s (!) = H1 (!) + O(X 2) (8.218)

where X is the amplitude of excitation.

The sine-testing techniques were exploited in [50, 51] in order to identify
automotive shock absorbers as described in an earlier section here.
One way of overcoming the distortion problems on the measured FRFs is
given by the interpolation method [90, 45]. The approach is as follows: if sine
excitation is used, the response components at multiples of the forcing frequency
are:

Y ( ) = XH1 ( ) + 43 X 3H3 ( ; ; ) + 85 X 5 H5 ( ; ; ; ; )+

(8.219)
9 In Gifford’s later work, he observed that it was only necessary to measure the FRF at a limited
number of points and that this number was only a linear function of the spectral densities and not the
quadratic function expected if the whole second-order FRF were to be measured. This allowed him to
reduce the size of the data sets involved to the same order as those considered by Storer. See [113].
 Higher-order modal analysis: hypercurve fitting 447

Figure 8.38. The three first-order FRFs of the system formed from time-domain
correlation measurements (dotted).
448 The Volterra series and higher-order frequency response functions

Frequency f2

f1 = fr
000
111
111
000
11
00
000
11100
11
000
111
00
11
11
00 000
111
0
1
000
111
1
0 11
00
000
11100
11
000
11111
00
00
11 00
11
f2 = fr

00
11
00
110
1 0
1
000
111
0
1
00
1100
110
1
0
1
0
1
00
11 0
1
00
11
000
111
11
00
11
00
00
11
11
00 11
00
00
11
000
111 00
11
000
11100
11
000
111
Frequency f1

00
11
000
111 00
11
000
111
00
11
000
11100
11
000
111
00
11
11
00 11
00
011
11
11
0011
0000
11
000
111
0111
000
000
111
00
11 00
11
00
0
1
000
111000
111
00
11
000
111
000
111
f2 = -f

11
00 10111
000
000
111
r

000
1111111
0000
00
11
000
111 11
00
000
111
000
111
000
111
000
111000
111
000
111
000
111
000
111
000
11100
11
000
111
11
00
000
111
000
111
00
11
11
00
f1
+
f2
=
fr

Figure 8.39. Curve-fitting to 2 (!1 ; !2 ) takes place in the black area where the bands
containing the poles overlap.

Y (2 ) = 12 X 2 H2 ( ; ) + 12 X 4 H4 ( ; ; ; )
+ 15 6
32 X H6 ( ; ; ; ; ; ) +


(8.220)
Y (3 ) = 14 X 3 H3 ( ; ; ) + 165 X 5 H5 ( ; ; ; ; ) +


: (8.221)

If the system is harmonically excited at a series of amplitude levels

X1 ; X2 ; : : : ; XN, an overdetermined system can be constructed (in this case up
to sixth order) as follows:
0
X1 3 2 5 X13 1
0
Y1 ( ) 1 4 X1 8 0 1
B Y2 ( ) C BX 3 X22 5 X23 C H1 ( )
B 2 4 8 C@ C
B
@ .. C
A =B H3 ( ; ; ) A (8.222)
@ ... .. .. A
. . . H5 ( ; ; ; ; )
YN ( ) XN 43 XN2 58 XN3
 Higher-order modal analysis: hypercurve fitting 449

Figure 8.40. Comparison of contour plots of the modulus of the second-order HFRF for
point 1 and both stiffness (left) and damping (right) curve-fits.

Table 8.2. Parameter estimates for quadratic damping model.

Coefficient Estimate
c211 246:24  37:53
c222 84:24  44:28
c233 4099:68  84:78
c212 62:1  44:28
c213 188:19  84:78
c223 67:23  14:31

1 X12
0 1 4 15 X16 1
0
Y1 (2 ) 1 2 2 X1 32 0 1
B Y2 (2 ) C B 1 X2 1 4 15 X26 C H2 ( ; )
B .. C=B
B
2 2 2 X2 32 C C@ H4 ( ; ; ; ) A
@ A @ .. .. .. A
. . . . H6 ( ; ; ; ; ; )
YN (2 ) 1 XN2 1 XN4 15 XN6
2 2 32
(8.223)
450 The Volterra series and higher-order frequency response functions

0.0005

Magnitude (m/N)
0.0004

0.0003

0.0002

0.0001

0.0000
0.0
Phase (deg)

-50.0
-100.0
-150.0
-200.0
0.0 100.0
Frequency (rad/s)

Figure 8.41. Comparison of the theoretical (dotted) against calculated (full) H1 (! ) from
the Duffing oscillator system sinusoidally excited with amplitudes of 0.1, 0.15 and 0.2 N.

0
Y1 (3 ) 1 1 X13 5 X15 1
0
4 16
B Y2 (3 ) C B 1 X3 5 5 C 
B 4 2 16 X2 C H3 ( ; ; )
H5 ( ; ; ; ; ) :
B .. C=B C (8.224)
@ A @ .. .. A
. . .
YN (3 ) 1 X3 5 X5
4 N 16 N
These systems of equations can be solved by standard LS methods.
The illustration for this method will be taken from [56]. A Duffing oscillator
(8.49) was chosen with coefficients m = 1, c = 20, k = 10 4 , k2 = 0
and k3 = 5  109 . The response to a stepped-sine excitation was computed
using fourth-order Runge–Kutta. The magnitudes and phases of the harmonic
components were extracted using an FFT after the transients had died down
at each frequency. Data for equations (8.222) to (8.224) were assembled for
up to H4 at amplitudes 0.1, 0.15 and 0.2 N and the first three diagonal FRFs
were calculated. The results are shown in figures 8.41–8.43 compared with the
analytical results; the agreement is impressive.

8.11 Higher-order FRFs from neural network models

The NARX class of neural networks was introduced in chapter 6 as a useful
non-parametric model structure for system identification. It is not immediately
obvious how the models relate to the Volterra series and to higher-order FRFs;
however, Wray and Green [280] have shown that there is a rather close
connection. That study demonstrated that the Volterra kernels of a given time-
 Higher-order FRFs from neural network models 451
0.00010

0.00008
Magnitude (m/N^2)
0.00006

0.00004

0.00002

0.00000
200.0
Phase (deg)

100.0
0.0
-100.0
-200.0
0.0 100.0
Frequency (rad/s)

Figure 8.42. Comparison of the theoretical (dotted) against calculated (full) principal
diagonal of H2 (!1 ; !2 ) from the Duffing oscillator system sinusoidally excited with
amplitudes of 0.1, 0.15 and 0.2 N.

Figure 8.43. Comparison of the theoretical (dotted) against calculated (full) principal
diagonal of H2 (!1 ; !2 ; !3 ) from the Duffing oscillator system sinusoidally excited with
amplitudes of 0.1, 0.15 and 0.2 N.

delay neural network (TDNN) are rather simple functions of the network weights.
The work is discussed here together with the extension to NARX networks and
452 The Volterra series and higher-order frequency response functions

u01
xi

w1

xi-1

yi

wn k

xi-N
uNn k

Figure 8.44. Example of a time-delay neural network (TDNN).

loosely follows the more detailed discussion in [56].

As in the case of NARMAX or NARX models generally, the structure
and parameters of the neural network NARX model will not necessarily be
unique. However, the HFRFs completely characterize the network at each order
of nonlinearity and therefore offer a means of validating neural network models
used in identification and control.

8.11.1 The Wray–Green method

Recently a method of directly calculating a system’s Volterra kernels was
presented by Wray and Green [280]. As physiologists, Wray and Green were
primarily interested in the time-domain Volterra kernels or higher-order impulse
responses. They established that if a TDNN could be trained to model a given
system, the Volterra kernels of the network could be calculated directly from its
weights. Consequently, if the network was an accurate model of the system,
its Volterra kernels would approximate closely to those of the system for an
appropriate set of input functions.
The basic TDNN is very similar to the NARX network, the main difference
being that only lagged inputs are used to form the model (figure 8.44). The
mathematical form is
nh nx 1 
X X
yi = s + wj tanh ujm xi m + bj (8.225)
j =1 m=0
where wj is the weight from the j th hidden unit to the output unit and n h is the
 Higher-order FRFs from neural network models 453

number of hidden layer units.

The method is based around the fact that the equation of a TDNN can be
shown equivalent to the discrete form of the Volterra series which is given by
1
X
y(t) = h0 +
1 h1 (1 )x(t 1 )
1
1X
X 1
+
1
2 h2 (1 ; 2 )x(t 1 )x(t 2 ) +



1 1
1 X
X 1
+ : : :
1 : : :
n hn (1 ; : : : ; n )x(t 1 ) : : : x(t n ) +



1 1
(8.226)

where hn is the usual nth-order Volterra kernel and the
 i are the sampling
intervals which can all be taken equal if desired. The requirement of causality
allows the lower index to be replaced by zero and the effect of damping is to
impose a finite memory T on the system, so the discrete series becomes
T
X
y(t) = h0 +
h1 (1 )x(t 1 )
1 =0
T X
X T
+
 2 h2 (1 ; 2 )x(t 1 )x(t 2 ) +



1 =02 =0
T
X T
X
+ :::
 n hn (1 ; : : : ; n )x(t 1 ) : : : x(t n ) +


:
1 =0 n =0
(8.227)

The first step of the Wray–Green method is to expand the activation function
of the neural network—the hyperbolic tangent as a Taylor series, then
1(
X 1)n+1 Bn (24n 22n ) 2n 1
tanh(z ) = z (8.228)
n=1 (2n)!
where the Bn are the Bernoulli numbers defined by

2(2n)! X1 1
Bn = :
(2)2n h=1 h2n
(8.229)

The first three non-zero coefficients for this series are 1 = 1, 3 = 1=3
and 5 = 2=15. The function is odd so all the even coefficients vanish. In
practice, in order to deal with the bias b j for each of the hidden nodes, Wray
and Green expand the activation function not around zero, but around b j of each
454 The Volterra series and higher-order frequency response functions

individual hidden node j . This effectively yields a different function f j (dj ) at

each node where
nXx 1
dj = ujm xi m : (8.230)
m=0
The coefficients in each expansion are labelled a pj following the definition
1
X
fj (dj ) = apj dpj (8.231)
p=0
and for the hyperbolic tangent function
1
apj = tanh(p) (bj ) (8.232)
p
where tanh(p) (bj ) is the pth derivative of tanh(b j ).
Adopting the definitions in (8.225), (8.231) and (8.232), the equation of the
TDNN network may be written:

yi = s + w1 (a01 + a11 d1 + a21 d21 +


)

+ w2 (a02 + a12 d2 + a22 d22 +


) +



+ wnh (a0nh + a1nh dnh + a2nh d2nh +


) +


: (8.233)

Now, collecting the coefficients of each power of xi gives, after a little

algebra

yi = w1 a01 + w2 a02 +


+ wnh a0nh

N
X
+ (w1 a11 um1 + w2 a12 um2 +


+ wnh a1nh umnh )xi m
m=0
XN XN
+ (w1 a21 um1 uk1 + w2 a22 um2 uk2 +


+ wnh a2nh umnh uknh )
m=0k=0
 xi m xi k +


(8.234)

and equating the coefficients of (8.234) and (8.226) yields a series of expressions
for the Volterra kernels
Xnh
h0 = wj a0j (8.235)
j =1
Xnh
h1 (m) =
 1 wj a1j umj (8.236)
j =1
Xnh
h2 (m; l) =
 2 wj a2j umj ulj (8.237)
j =1
 Higher-order FRFs from neural network models 455
0.00010

0.00005
Amplitude

0.00000

-0.00005

-0.00010
0.00 0.10 0.20 0.30 0.40 0.50
Time (s)

Figure 8.45. First-order time-domain Volterra kernel from the Duffing oscillator obtained
from a neural network model by the Wray–Green method.

and, in general,
nh
X
hn (m1 : : : mn ) =
 n wj anj um1 j um2 j : : : umnj : (8.238)
j =1
In order to illustrate this approach, time data for an asymmetric Duffing
oscillator (8.49) were simulated with the usual coefficients m = 1, c = 20,
k = 104, k2 = 107 and k3 = 5  109. A TDNN with 20 hidden units was
fitted using an MLP package and the resulting predictions of the first- and second-
order impulse responses are shown in figures 8.45 and 8.46. Wray and Green
consider that this method produces considerably better kernel estimates than the
Toepliz matrix inversion method which they cite as the previous best method of
kernel estimation [150] and their results are confirmed by an independent study
by Marmarelis and Zhao [172].

8.11.2 Harmonic probing of NARX models: the multi-layer perceptron

The object of this section is to extend the method of Wray and Green to NARX
neural networks with the form
Xnh X ny x 1
nX 
yi = s + wj tanh vjk yi k + ujm xi m + bj : (8.239)
j =1 k=1 m=0
(In practice, the situation is slightly more complicated than equation
(8.239)—and for that matter (8.253)—implies. During training of the network,
456 The Volterra series and higher-order frequency response functions

Figure 8.46. Second-order time-domain Volterra kernel from the Duffing oscillator
obtained from a neural network model by the Wray–Green method.

all network inputs are normalized to lie in the interval [ 1; 1] and the output is
normalized onto the interval [ 0:8; 0:8]. Equation (8.239) actually holds between
the normalized quantities. The transformation back to the physical quantities,
once the HFRFs have been calculated, is derived a little later.)
The NARX models considered here arise from models of dynamical systems
and a further simplification can be made. It is assumed that the effects of all
the bias terms cancel overall, as the systems being modelled will not contain
constant terms in their equations of motion. In dynamical systems this can always
be accomplished with an appropriate choice of origin for y (i.e. the equilibrium
position of the motion) if the excitation x is also adjusted to remove its d.c. term.
Because of the presence of the lagged outputs in (8.239), the correspondence
between the network and Volterra series does not hold as it does for the TDNN,
and an alternative approach is needed. It will be shown here that the Volterra
kernels are no longer accessible; however, the HFRFs are.
The method of deriving the HFRFs is to apply the standard method of
harmonic probing. Because of the complex structure of the model however, the
algebra is not trivial. In order to identify H 1 (! ), for example, the system is probed
with a single harmonic
xpi = ei t (8.240)
and the expected response is as usual

yip = H1 ( )ei t + H2 ( ; )e2i t + H3 ( ; ; )e3i t +


: (8.241)

 Higher-order FRFs from neural network models 457

Before proceeding, note that (8.239) is in an inappropriate form for this

operation as it stands. The reason is that the term of order n in the expansion
of the tanh function will contain harmonics of all orders up to n, so extracting
the coefficient of the fundamental requires the summation of an infinite series.
The way around this problem is to use the same trick as Wray and Green and
expand the tanh around the bias; this yields
Xnh X1  tanh(t) (b )  Xny x 1
nX t 
j
yi = s + wj vjk yi k + ujm xi m (8.242)
j =1 t=0 t! k=1 m=0
so each term in the expansion is now a homogeneous polynomial in the lagged
x’s and y’s. tanh(t) is the tth derivative of tanh at b j .
The only term in this expansion which can affect the coefficient of the
fundamental harmonic is the linear one; therefore take
x 1
tanh(1) (bj ) X
nh  ny nX 
X
yi = wj vjk yi k + ujm xi m : (8.243)
j =1 1 k=1 m=0
Following the discrete-time harmonic probing algorithm, substituting x p and
yp and extracting the coefficient of e i
t , yields

Xnh Xny Xnh x 1

nX
H1 ( ) = wj tanh (bj )
H1 ( ) + wj tanh(1) (bj )
(1) k ujm
m
j =1 k=1 j =1 m=0
(8.244)
which can be rearranged to give
Pnh (1) Pnx 1 i mÆt
j =1 wj tanh (bj ) m=0 ujm e
H1 ( ) = Pnh (1) P ny i kÆt : (8.245)
1 j =1 wj tanh (bj ) k=1 vjk e
Extraction of H 2 requires a probe with two independent harmonics:
xpi = ei 1 t + ei 1 t (8.246)
and
yip = H1 ( 1 )ei 1 t + H1 ( 2 )ei 2 t + 2H2( 1 ; 2 )ei( 1 + 2 )t +


: (8.247)
The argument proceeds as for H 1 ; if these expressions are substituted into
the network function (8.242), the only HFRFs to appear in the coefficient of
the sum harmonic e i( 1 + 2 )t are H1 and H2 , where H1 is already known from
equation (8.245). So as before, the coefficient can be rearranged to give an
expression for H 2 in terms of the network weights and H 1 . The only terms in
(8.242) which are relevant for the calculation are those at first and second order.
The calculation is straightforward but lengthy:
1 X nh
tanh(2) (bj )
H2 ( 1 ; 2 ) =
2!D j=1
wj
2!
fAj + Bj + Cj g (8.248)
458 The Volterra series and higher-order frequency response functions

where
ny X
X ny
Aj = vjk vjl H1 ( 1 )H1 ( 2 )(e i 1 kÆt e i 2 lÆt + e i 2 kÆt e i 1 lÆt )
k=1 l=1
(8.249)
x 1 nX
nX x 1
Bj = ujk ujl (e i 1 kÆt e i 2 lÆt + e i 2 kÆt e i 1 lÆt ) (8.250)
k=0 l=0
x 1
ny nX
X
Cj = 2 vjk ujl (H1 ( 1 )e i 1 kÆt e i 2 lÆt + H1 ( 2 )e i 2 kÆt e i 1 lÆt )
k=1 l=0
(8.251)

and
nh
X ny
X
D=1 wj tanh(1) (bj ) vjk e i( 1 + 2 )kÆt :
(8.252)
j =1 k=1
Derivation of H 3 is considerably more lengthy and requires probing with
three harmonics. The expression can be found in [57].

8.11.3 Radial basis function networks

Much of the recent work on system identification has abandoned the MLP
structure in favour of the radial basis function networks introduced by Broomhead
and Lowe [47]. The essential differences between the two approaches are in
the computation of the hidden node activation and in the form of the nonlinear
activation function. At each hidden node in the MLP network, the activation z is
obtained as a weighted sum of incoming signals from the input layer:
X
zi = wij xj : (8.253)
j
This is then passed through a nonlinear activation function which is
sigmoidal in shape, the important features of the function are its continuity,
its monotonicity and its asymptotic approach to constant values. The resulting
hidden node response is global in the sense that it can take non-zero values at all
points in the space spanned by the network input vectors.
In contrast, the RBF network has local hidden nodes. The activation is
obtained by taking the Euclidean distance squared from the input vector to a point
defined independently for each hidden node—its centre c i (which is of course a
vector of the same dimension as the input layer):

zi = kxi ci k: (8.254)

This is then passed through a basis function which decays rapidly with its
argument, i.e. it is significantly non-zero only for inputs close to c i . The overall
 Higher-order FRFs from neural network models 459

output of the RBF network is therefore the summed response from several locally-
tuned units. It is this ability to cover selectively connected regions of the input
space which makes the RBF so effective for pattern recognition and classification
problems. The RBF structure also allows an effective means of implementing the
NARX model for control and identification [61, 62].
For the calculation given here, a Gaussian basis function is assumed as this
is by far the most commonly used to date. Also, following Poggio and Girosi
[206], the network is modified by the inclusion of direct linear connections from
the input layer to the output. The resulting NARX model is summarized by

nh  y n x 1
nX 
X 1 X
yi = s + wj exp (y vjk )2 + (xi m ujm )2
j =1 2j2 k=1 i k m=0
n
X y x 1
nX
+ aj yi j + bj xi j (8.255)
j =1 j =0
| {z }
from linear connections

where the quantities v jk and ujm are the hidden node centres and the  i is the
standard deviation or radius of the Gaussian at hidden node i. The first part of
this expression is the standard RBF network.
As with the MLP network the appearance of constant terms in the exponent
will lead to difficulties when this is expanded as a Taylor series. A trivial
rearrangement yields the more useful form
nh   n
y
X 1 X
yi = s + wj j exp (y2 2vjk yi k )
j =1 2j2 k=1 i k
x 1
nX 
+ (x2i m 2ujm xi m )
m=0
ny
X x 1
nX
+ aj yi j + bj xi j (8.256)
j =1 j =0

where
  n
y nx 1 
1 X 2 + X u2
= exp v :
j
2j2 k=1 jk m=0 jm
(8.257)

Now, expanding the exponential and retaining only the linear terms leads to
the required expression for obtaining H 1 :

nh  ny x 1
nX 
X wj j X
yi = vjk yi k + ujm xi m : (8.258)
j =1 j k=1 m=0
460 The Volterra series and higher-order frequency response functions

Substituting the first-order probing expressions (8.240) and (8.241) yields

Pnh 1 Pnx 1 i mÆt + Pnx 1 bj e i jÆt
j =1 j wj j2 m=0 ujm e j =0
H1 ( ) = Pny
i jÆt Pnh j wj 12 Pny vjk e i kÆt : (8.259)
1 j =1 aj e j =1 j k=1

The second-order FRF H 2 is obtained as before:

nh
1 X
H2 ( 1 ; 2 ) = w
2!D j=1 j j
 nXx 1 nX
x 1
 21 2 (e ik 1 Æt e il 2 Æt + e ik 2 Æt e il 1 Æt )
j k=0 l=0
x 1 nXx 1
1 nX
+ 4 u u (e ik 1 Æt e il 2 Æt + e ik 2 Æt e il 1 Æt )
j k=0 l=0 jk jl
ny Xny
1 X
H ( )H ( )
2j2 k=1 l=1 1 1 1 2
 (e ik 1 Æt e il 2 Æt + e ik 2 Æt e il 1 Æt )
ny X ny
1 X
+ 4 v v H ( )H ( )
j k=1 l=1 jk jl 1 1 1 2
 (e ik 1 Æt e il 2 Æt + e ik 2 Æt e il 1 Æt )
ny nX x 1
1 X
+ 4 v u (H ( )e ik 1 Æt e il 2 Æt
j k=1 l=0 jk jl 1 1

+ H1 ( 2 )e ik 2 Æt e il 1 Æt ) (8.260)

where
ny nh n y
X X 1 X
D=1 aj e i( 1 + 2 )jÆt j wj e i( 1 + 2 )kÆt :
j2 k=1
(8.261)
j =1 j =1

8.11.4 Scaling the HFRFs

The network input vectors must be scaled prior to presentation to avoid saturation
of the processing units. In the MLP program used with tanh activation functions,
the y and x values are scaled between the interval [0:8; 0:8]. If a linear network
is used then both are scaled between [1; 1]. The resulting HFRFs must therefore
be scaled accordingly.
The scaling process gives the following, where the superscript ‘s’ denotes
scaled values and and a are the x and y scaling factors respectively, and and
 Higher-order FRFs from neural network models 461

b are the origin shifts.

xs = x + and y s = ay + b: (8.262)

The neural network program used here, MLP, calculates the scaled x values
as
 
x xmin 1
xs = 1:6
xmax xmin 2
1:6x 0:8(xmin + xmax )
= (8.263)
xmax xmin xmax xmin
which gives
1:6
= (8.264)
xmax xmin
and similarly
1:6
a= : (8.265)
ymax ymin
The dimensions of the nth kernel transform follow from

sF [y(t)jn ] sF [ay + b]
Hn ( ) =
F [x(t)j ]n = F [ x + ]n (8.266)

which holds from a dimensional point of view with s a constant.

As the constant offsets only affect the d.c. lines, this reduces to

saF [y(t)]
= n
F [x(t)]n (8.267)

so the scaling relation is

a
Hns ( 1 ; : : : ; n ) = n Hn ( 1 ; : : : ; n ): (8.268)

Therefore the true HFRF is given by

(y y )
Hn ( 1 ; : : : ; n ) = 1:6n 1 max min Hns ( 1 ; : : : ; n ): (8.269)
(xmax xmin)n

The network is therefore scale specific and may not perform well with data
originating from very different excitation levels. The solution to this problem is
for the training set to contain excitation levels spanning the range of interest.
462 The Volterra series and higher-order frequency response functions

0.0005

Magnitude (m/N)
0.0004

0.0003

0.0002

0.0001

0.0000
0.0
Phase (deg)

-50.0
-100.0
-150.0
-200.0
0.0 100.0
Frequency (rad/s)

Figure 8.47. Exact (dashed) and estimated (full) H1 from the Duffing oscillator, estimate
from probing of 10:2:1 NARX MLP network.

8.11.5 Illustration of the theory

The asymmetric Duffing oscillator specified in (8.49) was chosen for the
simulations with m = 1 kg, c = 20 N s m 1 , k = 104 N m 1 , k2 = 107 N m 2
and k3 = 5  109 N m 3 . The differential equation of motion stepped forward
using fourth-order Runge–Kutta. The excitation x(t) was a white Gaussian
sequence with zero mean and rms of 0.5, band-limited within the range 0–100 Hz.
A time-step of 0.001 s was adopted and the data were subsampled by a factor of
5. This gave a final
t of 0.005 s—corresponding to a sampling frequency of
200 Hz. For network training 1000 points of sampled input and displacement
data were taken.

8.11.5.1 Results from MLP network

Using the time data described earlier, various networks were trained and tested.
Once the weights were obtained, the previous formulae were used to compute the
first- and second-order FRFs of the network.
Figure 8.47 shows the H 1 ( ) (full in the figure) that best approximated the
theoretical result (shown dotted in the figure). The almost perfect overlay was
obtained from a network with ten input units and two hidden units (i.e. 10:2:1).
The network had converged to a model with an MPO error (see chapter 6) of 0.35.
The comparison between measured data and MPO is given in figure 8.48.
The second-order FRF proved a little more difficult to estimate accurately.
The best H2 ( 1 ; 2 ) estimation is compared to the theoretical kernel transform
 Higher-order FRFs from neural network models 463

Figure 8.48. Comparison between the measured data and the NARX MLP network
estimate for figure 8.47.

in figure 8.49 along its leading diagonal ( 1 = 2 ). This was calculated from a
10:4:1 network trained to an MPO error of 0.27. The corresponding H 1 ( 1 ) from
the same network, shown in figure 8.50, shows a little discrepancy from theory.
This is unfortunate from a system identification point of view as one would like a
correct representation of the system at all orders of nonlinearity. That this happens
appears to be due to the fact that the network can reproduce the signals with some
confusion between the different ordered components y n . This is discussed a little
later.

8.11.5.2 Results from radial basis function networks

The RBF networks were trained and tested using the force and displacement data.
A spread of results was observed, the best H 1 ( ) was given by a 6:2:1 network,
trained to a 0.48 MPO error as shown in figure 8.51. A 4:4:1 network produced the
best H2 ( 1 ; 2 ) after training to a 0.72 MPO error; this is shown in figure 8.52.
464 The Volterra series and higher-order frequency response functions
0.00010

0.00008
Magnitude (m/N^2)
0.00006

0.00004

0.00002

0.00000
200.0
Phase (deg)

100.0
0.0
-100.0
-200.0
0.0 100.0
Frequency (rad/s)

Figure 8.49. Exact (dashed) and estimated (full) principal diagonal for H2 from the
Duffing oscillator; an estimate from probing of a 10:4:1 NARX MLP network.

0.0005
Magnitude (m/N)

0.0004

0.0003

0.0002

0.0001

0.0000
0.0
Phase (deg)

-50.0
-100.0
-150.0
-200.0
0.0 100.0
Frequency (rad/s)

Figure 8.50. Exact (dashed) and estimated (full) H1 from the Duffing oscillator; the
estimate from probing of a 10:4:1 NARX MLP network.

8.11.5.3 Discussion

The results indicate possible over-parametrization by the network in modelling

the time data rather than accurately modelling the system in question. In system
 Higher-order FRFs from neural network models 465

0.0005

Magnitude (m/N)
0.0004

0.0003

0.0002

0.0001

0.0000
0.0
Phase (deg)

-100.0

-200.0
0.0 100.0
Frequency (rad/s)

Figure 8.51. Exact (dashed) and estimated (full) H1 from the Duffing oscillator; the
estimate from probing of a 6:2:1 NARX RBF network.

0.00010

0.00008
Magnitude (m/N^2)

0.00006

0.00004

0.00002

0.00000
200.0
Phase (deg)

100.0
0.0
-100.0
-200.0
0.0 100.0
Frequency (rad/s)

Figure 8.52. Exact (dashed) and estimated (full) principal diagonal for H2 from the
Duffing oscillator; the estimate from probing of a 4:4:1 NARX RBF network.

identification, over-parametrization often causes misleading results [231]. Over-

parametrization is caused by the model having many more degrees of freedom
than the system it is modelling. As the complexity of the network increases,
its data-modelling abilities generally improve, up to a point. Beyond that point
466 The Volterra series and higher-order frequency response functions

the error on the training data continues to decrease but the error over a testing
set begins to increase. Neural network users have long known this and always
use a training and testing set for validation, and often use a training, testing and
validation set. This may be necessary in this approach to system identification.
Often the MPO error is taken as conclusive.
A principled approach to model testing may be able to inform the modeller
that the model is indeed over-parametrized. The problem remaining is to remove
the over-parametrization. It may be possible to accomplish this in two ways.
First, there is a possibility of regularization [40], which seeks to ensure that the
network weights do not evolve in an unduly correlated manner. The ideal situation
would be if an analogue of the orthogonal LS estimator discussed in appendix E
could be derived for neural networks. This appears to be a difficult problem and
there is no current solution. The second possibility is pruning, where the network
nodes and weights are tested for their significance and removed if they are found
to contribute nothing to the model. This reduces the numbers of weights to be
estimated for a given network structure.

8.12 The multi-input Volterra series

The single-input–single-output (SISO) Volterra series is by now established as

a powerful tool in nonlinear system theory as previously discussed. In contrast,
the multi-input version of the series appears to have received little attention since
its inception (in the guise of the closely related Wiener series) in the work of
Marmarelis and Naka [171]. This section aims to show that the theory is no more
complicated and this is achieved by use of a case study.
When a nonlinear system is excited with more than one input,
intermodulation terms arise in the response. In fact, the Volterra series still
applies, however, a generalization is necessary. A short discussion is given here,
details can be found in [279]. First of all, an extension of the notation is needed.
A superscript is added to each Volterra kernel denoting the response point and
the number of occurrences of each particular input relevant to the construction of
(j:aabbb) (1 ; : : : ; 5 ) represents a fifth-order kernel
that kernel is indicated; e.g. h 5
measured at response point j and having two inputs at point a and three at point b.
For example, consider a nonlinear system excited at locations a and b with inputs
x(a) (t) and x(b) (t). The expression for the response at a point j is the same as
equation (8.3) in the single-input case. However, in the single-input case each
(j)
nonlinear component y n (t) in (8.3) is expressed in terms of a single Volterra
kernel; in the multi-input case, several kernels are needed. For the two-input case
the first two components are given by:

Z +1 Z +1
( j )
y1 (t) = (j : a) ( a)
d h1 ( )x (t  ) + d h(1j:b) ( )x(b) (t  ) (8.270)
1 1
 The multi-input Volterra series 467

and
Z +1 Z +1
( j )
y2 (t) = d1 d2 h(2j:aa) (1 ; 2 )x(a) (t 1 )x(a) (t 2 )
1 1
Z +1 Z +1
+ d1 d2 fh(2j:ab) (1 ; 2 ) + h(2j:ba) (2 ; 1 )g
1 1
 x(a) (t 1 )x(b) (t 2 )
Z +1 Z +1
+ d1 d2 h(2j:bb) (1 ; 2 )x(b) (t 1 )x(b) (t 2 ):
1 1
(8.271)

For convenience, the sum of cross-kernels is absorbed in a redefinition:

h2(j:ab) + h(2j:ba) ! 2h(2j:ab) . The corresponding frequency-domain quantities
are defined as before. It can be shown that there is no longer total symmetry under
permutations of the n symbols of h n and Hn ; a more restricted symmetry group
applies. For example, taking the series to second order as before, h 2
(j:ba) (1 ; 2 ) 6=
h(2j:ba) (2 ; 1 ).

8.12.1 HFRFs for a continuous-time MIMO system

For the purposes of illustrating the general multi-input–multi-output (MIMO)
theory, it is sufficient to consider a nonlinear 2DOF system with two inputs. The
system here has equations of motion

m1 y(1) + c1 y_ (1) + k11 y(1) + k12 y(2) + k0 y(1)2 + k00 y(1)3 = x(1) (t) (8.272)
m2 y(2) + c2 y_ (2) + k21 y(1) + k22 y(2) = x(2) (t): (8.273)

The harmonic-probing algorithm extends straightforwardly to this case. The

first-order FRFs are extracted as follows: substitute the probing expressions

x(1) (1:1) (2:1)

p (t) = ei!t ; x(2)
p (t) = 0; yp(1) (t) = H1 ei!t ; yp(2) (t) = H1 ei!t
(8.274)
into equations (8.272) and (8.273) and, in each case, equate the coefficients of
ei!t on the left- and right-hand sides. Next, substitute the probing expressions

x(1)
p (t) = 0; x(2) p (t) = ei!t
yp(1) (t) = H1(1:2) (!)ei!t ; yp(2) (t) = H1(2:2) (!)ei!t (8.275)

and equate coefficients as before. This procedure results in four simultaneous

(1:1) (1:2) (2:1) and H (2:2) . The H1
equations for the unknown quantities H 1 , H1 , H1 1
468 The Volterra series and higher-order frequency response functions

matrix is obtained from

H1(1:1) (!) H1(1:2) (!)
 

H1(2:1) (!) H1(2:2) (!)


!2 m1 + i!c1 + k11 + k12 k12
 1
= k21 !2 m2 + i!c2 + k21 + k22 :
(8.276)

In order to display the FRFs, the linear parameters were chosen as m 1 =

m2 = 1, c1 = c2 = 20, k11 = k22 = k12 = k21 = 10 000. For later, the
nonlinear parameters were chosen as k 0 = 1  107 and k 00 = 5  109 . The H1
magnitudes are displayed in figure 8.53. Only two of the functions are shown as
the system is symmetrical and so H 1
(1:1) = H (2:2) and H (1:2) = H (2:1) because
1 1 1
of reciprocity. The continuous time H 1 ’s are given by the full line.
The extraction of the second-order FRFs is a little more involved. First, the
probing expressions
x(1) (t) = ei!1 t + ei!2 t
p
x(2)
p (t) = 0
yp (t) = H1 (!1 )e + H1 (!2 )ei!2 t + 2H2(1:11) (!1 ; !2 )ei(!1 +!2 )t
(1) (1:1) i ! 1t (1:1)
yp(2)(t) = H1(2:1)(!1 )ei!1 t + H1(2:1)(!2 )ei!2 t + 2H2(2:11) (!1 ; !2 )ei(!1 +!2 )t
(8.277)
are substituted into the equations (8.272) and (8.273) and the coefficients of the
sum harmonic e i(!1 +!2 )t are extracted. Next, the exercise is repeated using the
probing expressions
x(1)
p (t) = 0
xp (t) = ei!1 t + ei!2 t
(2)
yp(1)(t) = H1(1:2)(!1 )ei!1 t + H1(1:2)(!2 )ei!2 t + 2H2(1:22) (!1 ; !2 )ei(!1 +!2 )t
yp(2)(t) = H1(2:2)(!1 )ei!1 t + H1(2:2)(!2 )ei!2 t + 2H2(2:22) (!1 ; !2 )ei(!1 +!2 )t
(8.278)
and
x(1)
p (t) = ei!1 t
x(2)
p (t) = ei!2 t
yp(1)(t) = H1(1:1)(!1 )ei!1 t + H1(1:2)(!2 )ei!2 t + 2H2(1:12) (!1 ; !2 )ei(!1 +!2 )t
yp(2) (t) = H1(2:1) (!1 )ei!1 t + H1(2:2) (!2 )ei!2 t + 2H2(2:12)(!1 ; !2)ei(!1 +!2 )t :
(8.279)
 The multi-input Volterra series 469

(a)

Frequency (rad/s)

(b)

Frequency (rad/s)

Figure 8.53. Comparison between HFRFs from a continuous-time system (full) and a
(1:1) (1:2)
discrete-time model (dashed): (a) H1 (! ); (b) H1 (! ).

(j:ab) (!1 ; !2 ). The solution is

This results in six equations for the variables H 2

H2(1:11) (!1 ; !2 ) H2(1:12) (!1 ; !2 ) H2(1:22)(!1 ; !2)

 

H2(2:11) (!1 ; !2 ) H2(2:12) (!1 ; !2 ) H2(2:22)(!1 ; !2)

 (1:1) (1:2) (! + ! ) 
= H 1 (!1 + ! 2 ) H 1 1 2
H1(2:1) (!1 + !2) H1(2:2) (!1 + !2 )
 (1:11) (1:12) (! ; ! ) S (1:22) (! ; ! ) 
 S 2 (! 1 ; ! 2 ) S 2 1 2 2 1 2 (8.280)
0 0 0
where
S2(1:ij) (!1 ; !2 ) = k0 H1(1:i) (!1 )H1(1:j) (!2 ): (8.281)
470 The Volterra series and higher-order frequency response functions

(a)

(b)

(c)

(1:11)
Figure 8.54. H2 magnitude surface: (a) from the continuous-time equation of motion;
(b) from a NARX model; (c) difference surface.

The magnitude surface for H 2

(1:11)
(!1 ; !2 ) is given in figure 8.54(a). To
display the symmetry properties of the surface better, a contour map is given as
 The multi-input Volterra series 471

(a)

(b)

Figure 8.55. Contour map of H2

(1:11)magnitude surface: (a) from the continuous-time
equation of motion; (b) from a NARX model.

figure 8.55(a). The magnitude of the cross-kernel HFRF H 2

(1:12)
(!1 ; !2 ) is given
in figure 8.56(a). Note from the contour map in figure 8.57(a) that the symmetry
about !1 = !2 is now absent.
472 The Volterra series and higher-order frequency response functions

(a)

(b)

(c)

(1:12)
Figure 8.56. H2 magnitude surface: (a) from the continuous-time equation of motion;
(b) from a NARX model; (c) difference surface.
 The multi-input Volterra series 473

(a)

(b)

Figure 8.57. Contour map of H2

(1:12)magnitude surface: (a) from the continuous-time
equation of motion; (b) from the NARX model.

8.12.2 HFRFs for a discrete-time MIMO system

In order to obtain a discrete-time system which compared with the continuous-
time (8.272) and (8.273), these equations were integrated in time using a fourth-
474 The Volterra series and higher-order frequency response functions

order Runge–Kutta scheme. The two excitations x (1) (t) and x(2) (t) were two
independent Gaussian noise sequences with rms 5.0, band-limited on the interval
(0; 200) Hz using a Butterworth filter. The time step was 0.001 corresponding
to a sampling frequency of 1000 Hz. Using the resulting discrete force and
displacement data, a NARX model was fitted. The model structure was easily
guessed as

yi(1) = 1 yi(1)1 + 2 yi(1)2 + 3 yi(1)3 + 4 yi(2)1 + 5 yi(2)2

+ 6 yi(1)21 + 7 yi(1)31 + 8 x(1) (1)
i + 9 xi 1 + 10 xi 2
(1) (8.282)
yi(2) = 11 yi(2)1 + 12 yi(2)2 + 13 yi(2)3 + 14 yi(1)1 + 15 yi(1)2
+ 16 x(2) (2)
i + 17 xi 1 + 18 xi 2
(2) (8.283)

and the parameters were fitted using a linear LS algorithm. The actual values of
the i coefficients will not be given here as they have little meaning. The model
was tested by stepping it forward in time using the same excitations x (1) and x(2)
and comparing the results with those from the original system. Figure 8.58 shows
the comparisons for each DOF; it is impossible to distinguish the two curves (the
NARX prediction is the broken line, the original data are the full line).
The HFRF extraction algorithm is almost identical to that for the continuous-
time system, the same probing expressions and coefficient extraction procedures
apply; the only difference is that where the derivative operator produces a
prefactor i! when it meets the harmonic e i!t , the lag operator extracts a
prefactor ei!
t , where
t is the sampling interval.
The H1 matrix for the discrete-time model is found to be

H1(1:1) (!) H1(1:2) (!)

 

H1(2:1) (!) H1(2:2) (!)

 i!
t i2!
t 3 e i3!
t
= 1 1 e e i!
2te
14 15 e i2!
t
 1
4 e i!
t 5 e i2!
t
1 11 e i!
t 12 e i2!
t 13 e i3!
t
 i!
t + 10 e i2!
t 
 8 + 9 e 0 :
0 16 + 17 e i!
t + 18 e i2!
t
(8.284)

The H1 s are shown in figure 8.53 as the broken lines, in comparison with
the H1 s from the original continuous-time model. As would be expected from the
accuracy of the NARX models, the two sets of H 1 s agree almost perfectly.
 The multi-input Volterra series 475

(a)

(b)

Figure 8.58. Comparison between the measured data and the NARX model-predicted
output.
476 The Volterra series and higher-order frequency response functions

The H2 matrix is found to be

H2(1:11) (!1 ; !2 ) H2(1:12)(!1 ; !2) H2(1:22) (!1 ; !2 )

 

H2(2:11) (!1 ; !2 ) H2(2:12)(!1 ; !2) H2(2:22) (!1 ; !2 )

 (1:1) (1:2)  1
= H1(2:1) (!1 + !2 ) H1(2:2) (!1 + !2 )
H (!1 + !2 ) H1 (!1 + !2 )
1 i(!1 +!2 )
t + 10 e i2(!1 +!2 )
t
 8 + 9e 0

0
16 + 17 e i(!1 +!2 )
t + 18 e i2(!1 +!2 )
t
 (1:11) (1:12) (1:22) 
 T2 (0!1 ; !2 ) T2 (0!1 ; !2 ) T2 (0!1 ; !2 ) (8.285)

where

T2(1:ij) (!1 ; !2 ) = 6 H1(1:i) (!1 )H1(1:j) (!2 )e i(!1 +!2 )t : (8.286)

Examples of some of the H 2 surfaces for the discrete-time system can be

found in figures 8.54–8.55 where they are compared with the corresponding
continuous-time objects. As with the H 1 functions, the agreement is excellent.
The MIMO version of the Volterra series is no more difficult to apply than the
standard SISO expansion. If multiple inputs are applied to a MDOF system, the
Volterra cross-kernels or cross-HFRFs must be taken into account in any analysis,
as they encode important information about intermodulations between the input
signals.