3 Continuous-Time Structural Systems

Structures can be regarded as distributed parameter systems characterized by the

distribution of the mass, damping and stiffness properties. However, parameter
identification of such systems is in general not easy. Thus, with a few exceptions, in
most of the literature on testing of structures, the data are analysed based on the
assumption that the system is described by one or more linear ordinary differential
equations. Because of their simplicity the linear time-invariant lumped parameter
models are the most widely used models in structural identification. More complex
models such as the linear continuous parameter models and nonlinear models are
used only when the lumped-parameter model cannot be used to provide an adequate
representation of the structural behaviour. In general, system identification concerns
the determination of modal parameters. If the excitation of a structure e.g. is the
wind, then the only available information about the dynamic behaviour of a structure
is the measured vibrations of it. As will be shown later, system identification using
ARMAV models is in these situations capable of providing good estimates of the
modal parameters of the structure. However, it is necessary to obtain some kind of
understanding of how the discrete-time ARMAV model relates to the modal
parameters, i.e. how it relates to the continuous-time lumped parameter model. It is
the purpose of this and the following two chapters to provide this understanding. This
chapter is restricted to the continuous-time modelling of civil engineering structures.

Section 3.1 concerns the modelling of a structural system using the second-order
lumped parameter model and how the modal parameters of this model are defined.
This model is then generalized in section 3.2, in order to cover situations where the
number of observed masses of the model are different from the total number of
modes in the system. It is also shown how the ambient excitation can be modelled.
In section 3.3, the model that describes the excitation is combined with the
generalized lumped parameter model of the structural system. Two examples will be
given as an illustration. Finally, in section 3.4, the modal decomposition of the
combined system is investigated. The results are illustrated by an example.

3.1 Modelling of Second-Order Structural Systems

Due to the complexity of structures, parameter estimation is usually simplified by
certain assumptions about the structures. Such simplifications can e.g. be that they
behave linearly, that they are time-invariant, and that they can be represented by a
mass-spring-dashpot model. Since a structure is a continuous system with distributed
mass, such a model should in principle have an infinite number of degrees of
freedom. However, since it is seldom more than a few dynamic modes that are of
interest, it is sufficient to construct a reduced model, capable of describing the
behaviour of the dynamic modes of interest in terms of modal parameters. Such a
model is termed a lumped parameter model.

Continuous-Time Structural Systems 43

3.1.1 Constructing a Mathematical Model

Experience has led to the following mathematical mass-spring-dashpot lumped

parameter model for a structure subjected to external loading, see e.g. Kozin et al.
[68] and Pi et al. [92]
Mz̈( t )  Cz(
t )  Kz( t )
f( t ) (3.1)

M, C and K are the mass, damping and stiffness matrices all of dimensions p × p. z(t)
and f(t) are the p × 1 displacement and force vectors at the mass points, respectively.
The differential equation represent a force equilibrium. The forces of inertia Mz̈(t)
are balanced by a set of linear-elastic restoring forces Kz(t), viscous damping
and the external forces f(t). This reduced model is capable of describing the
Cz(t)
motion of p fictitious mass points of the structure. The mass points are fictitious since
they represent a discretization of the distributed mass of the structure. It is thus
assumed that the distributed forces of inertia of the structure can be discretized into
p degrees-of-freedom (DOF). Since the motion of the system is observed at the mass
points, M will be diagonal, and due to the Maxwell theorem the stiffness matrix will
be symmetric and positive definite. It is the usual assumption in linear vibration
theory that the damping matrix is also symmetric, since any non-symmetric part does
not dissipate energy. Assuming a linear structure means that the response of the
structure, to any combination of forces simultaneously applied, is the sum of
individual responses to each of the forces acting alone. This is a good assumption for
a variety of structures. The time-invariant assumption implies that the parameters in
the model are constants.

3.1.2 Modal Analysis

One of the most important applications of system identification is that estimated

models can serve as a basis for a modal analysis of a structure. In this section, it is
shown how to obtain the modal parameters from the second-order structural system
(3.1). The modal parameters of the jth mode are the modal frequency, the modal
damping, the modal vector and the modal scaling, and below is listed how these four
modal parameters can be represented, see e.g. Hoen [38].

Modal frequency:
/ Eigenvalue,j.
/ Angular eigenfrequencies, 7j.
/ Natural eigenfrequencies, fj.

Modal damping:
/ Damping ratios, j.

Modal vector:
/ Eigenvectors, 5j.
/ Mode shapes, 0j.

44 Modelling of Second-Order Structural Systems

Modal scaling:
/ Modal masses, mj.
/ Residues, Rj.

Now, assume that the structural system is given a unit impulse and then left on its
own. The vibrations of the system will then solely be dependent upon the dynamic
characteristics of the structure, for which reason they are called free vibrations. Since
the vibrations are viscously damped, it is in general necessary to consider a complex
eigenvalue problem to determine the modal parameters. The solution of this
eigenvalue problem requires the construction of a 2p × 2p matrix, and correspond-
ingly a 2p order response vector on which the matrix operates. This vector is also
called the state vector of the system (3.1) and is defined in terms of the displacements
and velocities of the system

z( t )
x( t )
(3.2)
t)
z(

This state vector reduces the second-order differential equation system (3.1) to the
following first-order differential equation system
t )  B x( t )
u( t )
A x(

(3.3)
C M K 0 f( t )
A
, B
, u( t )

M 0 0 M 0

The first p rows of (3.3) are the same as the original equation of motion (3.1),
t ) Mx(
whereas the remaining rows are the identities Mx( t )
0 . The free vibrations
of the system (3.3) are then given by
t )  B x( t )
0
A x( (3.4)

and the solution form is assumed as

x( t )
5 e  t (3.5)

where 5 is a complex vector of dimension 2p × 1 and  is a complex constant.

Insertion of (3.5) into (3.4) shows that (3.5) is a solution if and only if 5 is a solution
to the first-order eigenvalue problem
A  B 5
0 (3.6)

which leads to the following characteristic polynomial

det  A  B
0 (3.7)

Continuous-Time Structural Systems 45

The order of this real-valued polynomial is 2p. The polynomial will as such have 2p
roots j, for j = 1 to 2p, which are the eigenvalues of (3.6). For each of these
eigenvalues a non-trivial solution 5j to (3.6) exists. This solution vector is called the
eigenvector. The eigenvalues and therefore also the eigenvectors can either be real
or complex. If all the eigenvalues are represented by complex conjugated pairs then
the system is underdamped. This system behaviour will be assumed here, since it is
the usual behaviour for a wide range of civil engineering structures. The complex
conjugated pairs of eigenvalues are then given by
2 2
j , j1
7j j ± i 7j 1 j
2 % fj j ± i 2 % f j 1 j
(3.8)
j < 1 , j
1, 3, ... , ( 2p 1 )

7j, f j and j are the angular and natural eigenfrequencies, and the damping ratio, of
the jth underdamped mode, respectively. From (3.2) and (3.3) it follows that the
eigenvector 5j must have the form
0j
5j
, j
1, 2, ... , 2p (3.9)
j 0j

and insertion of A and B and (3.9) into (3.6) yields

2
Mj  Cj  K 0j
0 , j
1, 2, ... , 2p (3.10)

This is a standard eigenvalue problem for the second-order system, with 0j being the
non-trivial solution vector of it. This vector is also called the damped complex mode
shape or simply the mode shape. Assembling all eigenvectors 5j, defines the complex
modal matrix
01 02 . . . 02p
4
(3.11)
1 01 1 02 . . . 2p 02p

This matrix has the orthogonality properties


diag { j }
4TA4
Md , 4TB4
 Md , (3.12)
Md
diag { mj }

with  and Md being diagonal matrices containing the 2p eigenvalues j and the
damped modal masses mj, respectively.

46 Modelling of Second-Order Structural Systems

For zero initial conditions, the solution of (3.1) is conveniently written in terms of the
convolution integral as
t
z(0)
0
z( t )
h( t - ) f( - ) d - , (3.13)
P
0
z(0)
0

where h(t) is the p × p impulse response function. This function fully describes the
dynamic behaviour of the structural system completely, and can conveniently be
expressed in terms of the modal decomposed system as
T
2p 0j 0j j t
2p
j t
h( t )
M e
M Rj e (3.14)
j
1 mj j
1

mj is the jth diagonal element of M d , and R j is the residue matrix that corresponds to
the jth eigenvalue. Since Rj is a normalized matrix, it is insensitive to any scaling of
the mode shapes.

3.1.3 A Canonical Realization of a Structural System

The system (3.3) is usually preferred in classical analytical vibration theory due to the
assumption of the symmetric M, C and K which also makes A and B symmetric.
Symmetric A and B can simultaneously be reduced to diagonal forms. However, in
the following sections the second order differential equation system is generalized to
a system of arbitrary order. In this case the representation (3.3) is not suitable, and
canonical realizations are preferred instead. Also, in the context of system identifica-
tion the system needs to be modelled canonically in order to carry out the numerical
manipulations. The system (3.3) can be converted to a canonical form by pre
multiplication of the state equation of (3.3) with A-1 to yield
t )
F x( t )  E f( t )
x(

0 I 0 (3.15)
F
A 1 B
, E

M 1 K M 1 C M 1

Notice that A-1 and B can be written as

1
A 1
4 Md 4T , B
4 T Md 4 1 (3.16)

From these properties and (3.15), it is simple to verify that 4 and its inverse
diagonalize F as

Continuous-Time Structural Systems 47

 F
A 1 B
1

4 Md 4T4 T Md 4 1
(3.17)
1

4 Md  Md 4 1

4  4 1

with  being the eigenvalues of F. As seen the damped modal masses are eliminated
from (3.17). This is because K and C in F have been normalised with M, see (3.15).
This implies that an arbitrary normalization of the eigenvectors can be applied
without changing F. Therefore, the damped modal mass is meaningless for this
realization. The solution of (3.15) may still conveniently be described in terms of the
convolution integral as in (3.13) by means of the impulse response function h(t),
which is now defined as, see e.g. Kailath [48]
2p 2p
j t j t
h( t )
M 0j 4 1jE e
M Rj e (3.18)
j
1 j
1

where 4-1j signifies the jth row of 4-1. Therefore, the only difference in the definition
of the impulse response function of the two realizations is the way the residue Rj is
defined. It is no longer dependent on any normalization, which underlines that the
modal mass is meaningless for this realization.

3.1.4 Spectrum Analysis

In this section, it is considered how to represent the dynamics of the structural system
if the applied excitation f(t) is a stationary zero mean Gaussian distributed stochastic
process. In this case the response z(t) of the system will also be a Gaussian distributed
stochastic process, and it will also be stationary when the effects of the initial
conditions have faded away. This section is primarily based on Bendat et al. [14], and
it will be shown how stochastically excited systems can be studied in frequency
domain. Since f(t) is assumed zero mean, it is fully described by its covariance
function $(-), defined as

$( - )
E f( t ) f T( t - ) (3.19)

Due to the linearity of the system, the system response z(t) will also be fully described
by its covariance function, which can be defined in terms of (3.13) and (3.19) as

48 Modelling of Second-Order Structural Systems

  

( - )
E z( t ) z ( t - )

T
h( t1 ) $( t2 t1  - ) hT( t1 ) d t1 d t2 (3.20)
P P
0 0

The covariance functions are time domain representations of the statistical properties
of the stochastic processes. The processes f(t) and z(t) can equally well be represented
in frequency domain by using the Wiener-Khinchine relation, see e.g. Bendat et al.
[14], to yield
 
1 1
Sff( 7 )
$( - ) e i 7- d - , S zz( 7 )

( - ) e i 7- d - (3.21)
2% P 2% P
 

Sff(7) and Szz(7) are the spectral density functions of f(t) and z(t) with 7 being an
arbitrary angular frequency. By introducing the frequency response function H(7) as
the Fourier transformed impulse response function

H( 7 )
h(  ) e i 7 d  (3.22)
P


it is possible to represent Szz(7) in terms of Sff(7) as

Szz( 7 )
H( 7 ) S ff( 7 ) HH( 7 ) (3.23)

where the superscript H signifies Hermitan or complex conjugate transpose. The

frequency response function is directly related to the impulse response function, and
can as such also be represented in terms of the residue Rj as
2p Rj
H( 7 )
M (3.24)
j
1 i7 j

which follows directly by inserting (3.14) into (3.22). In some cases the spectral
densities are easy to calculate. If f(t) is a Gaussian white noise, then its covariance
function is especially simple
$( - )
E f( t ) f T( t - )
F ( - ) (3.25)

F is a constant intensity matrix and (-) is the Dirac delta function. The spectral
density function of f(t) is constant and equal to
1
Sff
F (3.26)
2%

Continuous-Time Structural Systems 49

The spectral density of the response of a Gaussian white noise excited second-order
system is then obtained from (3.23) and (3.24) as
2p 2p H
1 1 Rj F Rk
S zz( 7 )
H( 7 ) F H ( 7 )

H
M M (3.27)
2% 2% j
1 k
1 ( i7 j ) ( i7 k )

The study of the spectral densities is also called spectrum analysis and is a helpful
tool for modal analysis of stochastically excited systems.

3.2 Modelling of General Structural Systems

The use of the second-order differential equation system, for description of the
dynamic behaviour of a structural system, only applies to systems where the number
of observed mass points is equal to the total number of vibrating masses of the
model. In the general case the number of observed mass points might be smaller than
the total number of vibrating masses. In this case the mathematical model will not be
the second order differential equation (3.1) but instead a differential equation of
higher order. If for example the vibrations of a single mass point are observed in a
system having p masses, then the order of the associated differential equation is 2p.

3.2.1 Constructing a General Mathematical Model

From a system identification point of view a generalization of the mathematical

model is therefore necessary, since the number of measurement channels is usually
less than the number of identified modes. A generalized multivariate model can be
formulated as in the following definition.

Definition 3.1 - Linear Time-Invariant Continuous-Time Structural System

Denote D the differential operator, defined as Dz(t)= z(  t ) . The p-variate continuous-time

differential equation of motion of the structural system of order s is then given by

D sz( t )  Az, s 1 D s 1z( t )  ...  Az, 1Dz( t )  Az, 0 z( t )

(3.28)
Bf, s 2 D s 2f( t )  ...  Bf, 1D f( t )  Bf, 0 f( t )

where the matrices Az,i and Bf,i are all of the dimension p × p. The displacement vector z(t) and
its derivatives are all of the dimension p × 1. The p × 1 vector f(t) describes the forces applied
to the system. a
The modes of a structural system will typically be underdamped, which implies that
each mode is described by a pair of complex conjugated eigenvalues. In this situation
the order s will be defined as s
2N with N being the number of underdamped modes.
p

50 Modelling of General Structural Systems

The number of output channels must be selected so that s becomes an integer value.
The matrices Az,i are now the system matrices that describe the vibration of the mass
points of the model. The matrices Bf,i describes the coupling to the unobserved
vibrating mass points of the model. For the second order structural differential system
(3.1) it can be verified that only Bf,0 exist and that it is equal to M-1. However, since
it is assumed that the system is subjected to ambient excitation, it still remains to
model this unknown excitation.

3.2.2 Modelling of Ambient Excitation

Instead of obtaining the ambient excitation by indirect measurements such as the sea
state, it is often assumed that the ambient excitation is given as the output of a linear
time-invariant shaping filter subjected to Gaussian white noise. Due to the Gaussian
assumption, it is implicitly assumed that the true ambient excitation is at least weakly
stationary. If the ambient excitation can be described by filtered white noise, then it
is possible to derive a model for it. The above assumptions lead to the following
definition on how to model the stochastic excitation f(t) applied to the structural
system (3.28).

Definition 3.2 - Excitation Generated by a Shaping Filter Subjected to White

Noise

Assume that the excitation f(t) of the structural system is obtained as the output of an mth-
order p-variate linear time-invariant continuous-time shaping filter

D mf( t )  Af, m 1 D m 1f( t )  ...  Af, 1 D f( t )  Af, 0 f( t )

w( t ) (3.29)

For simplicity, it is assumed that f(t) and w(t) have the same dimensions as z(t). This implies
that the matrices Af,i all have the dimension p × p. The stochastic process w(t) is a zero-mean
Gaussian white noise, fully described by its covariance function. This covariance function is
defined in terms of the p × p intensity matrix W as

E w( t )
0 , E w( t ) w T( t - )
( - ) W (3.30)

where (-) is the Dirac delta function. These statistical properties will in the following be
abbreviated NID(0,W). a
It is obvious that the response z(t) of the system will contain a mixture of the dynamic
behaviour of the structural system and of the excitation. It is also intuitively clear that
during a system identification the dynamic modes of this filter will also be estimated.
These modes are, together with any noise modes, called nonphysical modes. In this
way they can be distinguished from the physical modes of the structural system. In
figure 3.1, the generation of the excitation from Gaussian white noise is explained in
terms of spectral densities of a univariate system.

Continuous-Time Structural Systems 51

Figure 3.1: The constant spectral density of the univariate Gaussian white noise w(t)
with an intensity W is shaped into the desired univariate spectral density
of f(t). This is done by means of the SDOF shaping filter having 7f as
angular eigenfrequency.

If the structural system can be combined with the shaping filter of the excitation into
a resulting linear system subjected to a Gaussian white noise, then it is possible to
represent this system by a discrete-time Gaussian white noise excited state space
realization or an equivalent ARMAV model, according to section 2.3. The next
natural step is therefore to combine the structural system and the shaping filter of the
excitation.

3.3 Combined Continuous-time Systems

The response z(t) from the structural system can be expressed directly in terms of the
Gaussian white noise w(t) by a multivariate convolution of (3.28) into (3.29). This
convolution procedure can be represented in a simple manner using a state space
approach. The generalized structural system and the shaping filter can then
conveniently be represented by the following two coupled state space realizations
x 1( t )
A1 x1( t )  B1 f( t )
(3.31)
z( t )
H1 x1( t )

x 2( t )
A2 x2( t )  B2 w( t ) , w( t ) NID( 0 , W )
(3.32)
f( t )
H2 x2( t )

with the matrix triples {A1, B1, H1} and {A2, B2, H2} given by

52 Combined Continuous-time Systems

 0 I . . 0 0
0 0 . . 0 0
. . . . .
A1

. . . . .
0 0 . . I 0
Az, 0 Az, 1 . . Az, s 2 Az, s 1
1
I 0 . . 0 0 0 (3.33)
Az, s 1 I . . 0 0 B f, s 2
. . . . . . .
B1

. . . . . . .
Az, 2 Az, 3 . . I 0 B f, 1
Az, 1 Az, 2 . . Az, s 1 I B f, 0

H1
I 0 . . 0 0

0 I . . 0 0
0 0 . . 0 0
. . . . .
A2

. . . . .
0 0 . . I 0
A f, 0 A f, 1 . . A f, m 2 A f, m 1

0 (3.34)
0
.
B2

.
0
I
H2
I 0 . . 0 0

These realizations are in observability canonical form and their construction follows
the principles shown for the discrete-time systems in section 2.5, see also Kailath
[48]

Continuous-Time Structural Systems 53

Introduce the following augmented state vector
x1( t )
xc( t )
(3.35)
x2( t )

If f(t) in (3.31) is substituted by the observation equation of (3.32), and the resulting
two state equations of (3.31) and (3.32) are stacked, then the following coupled set
of equations is obtained
x 1( t )
A1 x1( t )  B1 H2 x2( t )
x 2( t )
A2 x2( t )  B2 w( t ) (3.36)

Combine this system with the augmented state vector and the observation equation
of (3.31) to yield
x c( t )
Fc xc( t )  Bc w( t ) , w( t ) NID( 0 , W )

z( t )
Hc xc( t )
(3.37)
A 1 B 1 H2 0
Fc
, Bc
, Hc
H1 0
0 A2 B2

As seen, this convolution follows the procedures described in section 2.1.2 for
nonwhite excited linear and time-invariant discrete-time systems. The dimensions of
Fc, Bc and Hc are:
dim Fc
( s  m ) p × ( s  m ) p

dim Bc
( s  m ) p × p (3.38)

dim Hc
p × ( s  m ) p

which implies that the order of the homogenous part of the corresponding
multivariate differential equation is n = s+m. This differential equation will have the
form

D nz( t )  Cz, n 1 D n 1z( t )  ...  Cz, 1 Dz( t )  Cz, 0 z( t )

Cw, s 2 D s 2w( t )  ...  Cw, 1 D w( t )  Cw, 0 w( t ) (3.39)

w( t ) NID( 0 , W )

54 Combined Continuous-time Systems

Based on theorem 2.2, which also applies for continuous-time state space systems,
see Gohberg et al. [30], the matrix coefficients Cz,i and Cw,i can be obtained by
n 1
Cz, 0 Cz, 1 . . Cz, n 1
Hc Fc Qo ( n )
(3.40)
Cw, 0 . . Cw, s 2 0 . . 0
Cz, 0 Cz, 1 . . Cz, n 1 I T( n1 )

with Qo(n) being the non-singular observability matrix, based on { Fc, Hc }. The
definition of the matrix T(n+1) follows theorem 2.2 as
0 0 . . 0 0
HcB c 0 . . 0 0
. . . . . .
T( n1 )
. . . . . . (3.41)
n 2 n 3
HcF c B c HcF c Hc . . 0 0
n 1 n 2
HcF c B c HcF c B c . . HcB c 0

If the following linear transformations are applied

x( t )
T xc( t ) , F
T Fc T 1 , B
T Bc , H
Hc T 1 (3.42)

where T = Qo(n), then (3.37), or equivalently (3.39), can be represented by the

following observability canonical state space realization
t )
Fx( t )  Bw( t ) ,
x( w( t ) NID( 0 , W )
(3.43)
z( t )
Hx( t )

The state vector x(t) of this realization will then be defined as

z( t )
Dz( t )
.
x( t )
(3.44)
.
D n 2z( t )
D n 1z( t )

and the matrix triple {F, B, H} as

Continuous-Time Structural Systems 55

 0 I . . 0 0
0 0 . . 0 0
. . . . . .
F

. . . . . .
0 0 . . 0 I
Cz, 0 Cz, 1 . . Cz, n 2 Cz, n 1

0
1
I 0 . . 0 0 . (3.45)
Cz, n 1 I . . 0 0 .
. . . . . . 0
B

. . . . . . C w, s 2
Cz, 2 Cz, 3 . . I 0 .
Cz, 1 Cz, 2 . . Cz, n 1 I .
C w, 0
H
I 0 . . 0 0

The combination of the shaping filter and the structural system is in figure 3.2
explained in terms of spectral densities for an univariate system.

Figure 3.2: The constant spectral density of the univariate Gaussian white noise
w(t) with an intensity W is shaped into the desired univariate spectral
density of f(t) by means of the SDOF shaping filter. The angular
eigenfrequency of the shaping filter is 7f. The spectral density of f(t) is
then reshaped by means of the SDOF structural system to yield the
spectral density of the output z(t). The angular eigenfrequency of the
structural system is 7z.

56 Combined Continuous-time Systems

A frequency domain interpretation can be found in Asmussen et al. [10] and Ibrahim
et al. [41], where the multivariate shaping filter is referred to as a pseudo force
system.

3.3.1 Generalization to an Arbitrary System Output

The realization (3.43) is an efficient way of representing the combined differential

equation system (3.39). However, in some cases it is desirable to observe other
characteristics than the displacements z(t) of the system. This can be accomplished
by a simple modification of the observation equation of (3.43). The matrix C is
defined as a generalized observation matrix and y(t) is a generalized output vector.
Assume that y(t) is obtained as a linear combination of the states of x(t), and that this
linear combination is controlled by the observation matrix C. A state space
realization of the generalized combined continuous-time system is then given by
t )
Fx( t )  Bw( t ) ,
x( w( t ) NID( 0 , W )
(3.46)
y( t )
Cx( t )

It should be noted that the state equation is unaffected, and that (3.43) can be
recovered from setting C = H and y(t) = z(t). Three special cases are worth
mentioning. These show the appearance of C in case the vector y(t) describes the
displacements, the velocities, or the accelerations of the structural system:
y( t )
C x( t )
I 0 0 . . 0 0 x( t )
z( t ) (3.47)

y( t )
C x( t )
0 I 0 . . 0 0 x( t )
z(
t) (3.48)

y( t )
C x( t )
0 0 I . . 0 0 x( t )
z̈( t ) (3.49)

Relation (3.49) implies that in order to extract the accelerations from the system, the
order of (3.39) must be at least three. A way to reduce the necessary order is to
include derivatives of the state vector x(t) in the observation equation in (3.45).
However, in the present context this is not desirable, since this will introduce a
Gaussian white noise component as a direct term in the observation equation. Such
a component will make the system impossible to sample, since the zero-lag
covariance of the output y(t) then will approach infinity, see e.g. Shats et al. [102].
In the following the construction of a univariate second-order structural system will
be illustrated in two examples. In the first example the applied excitation is assumed
to be Gaussian white noise. In this case the combined system resembles the second-
order structural system described in section 3.1. The next example illustrates what
happens when the excitation is no longer a Gaussian white noise.

Continuous-Time Structural Systems 57

3.3.2 Example 3.1: A White Noise Excited System

Consider a univariate structural system. The motion of this system is assumed to be described
by the following second-order differential equation

c k 1
z̈( t )  t) 
z( z( t )
f( t ) (3.50)
m m m

m, c and k are the mass, viscous damping, and stiffness respectively. z(t) is the displacement
of the system due the Gaussian distributed excitation f(t) = w(t). This excitation is assumed
to be a zero-mean Gaussian white noise described by the intensity W. Assume that the
observed response is the displacement. Then the matrix triple {F, B, C}of the combined
continuous-time system (3.46) is given by

0 1 0
F
, B
, C
1 0 (3.51)
k c 1
m m m

The observability matrix Qo(2) obtained from {F, C}is simply an 2 × 2 identity matrix. a
3.3.3 Example 3.2: A Nonwhite Excited System

Consider the univariate structural system in example 3.1. The motion of this system was
described by the second-order differential equation (3.50). Assume now that the excitation
f(t) is a nonwhite Gaussian process, obtained by filtering a zero-mean Gaussian white noise
w(t), described the intensity W, through the second-order shaping filter

¨ t )  a f(
f( 1 t )  a0 f( t )
w( t ) , w( t ) NID( 0 , W) (3.52)

From (3.33) and (3.34) the matrix triples {A1, B1, H1} and {A2, B2, H2} appear as

0 1 0
A1
, B1
, C1
1 0
k c 1
m m m
(3.53)
0 1 0
A2
, B2
, C2
1 0
a0 a1 1

and from (3.53), the matrix triple {Fc, Bc, Hc } of the combined continuous-time system (3.37),
is given by

0 1 0 0 0
k
c 1
0 0
Fc
m m m
, Bc
, Hc
1 0 0 0 (3.54)
0 0 0 1 0
0 0 a0 a1 1

58 Combined Continuous-time Systems

 The observability matrix Qo(4) is obtained from {Fc, Hc} to yield

1 0 0 0
0 1 0 0
Qo(4)
k
c 1
0 (3.55)
m m m

c2
ck
k c 1
m2 m m2 m m

Assuming that the output y(t) is the displacement of the system, the combined system can be
represented by the state space realization (3.46), defined by the matrices

0 1 0 0
0 0 1 0
F
0 0 0 1
a0 k a1 k  a0 c k  a1 c  a0 m c  a1 m

m m m m

0 (3.56)
0
B
0
1
m

C
1 0 0 0

3.4 Modal Decomposition of Combined Continuous-Time Systems

In this section the modal decomposition of the combined structural system is
considered. The modal decomposition of this system clearly shows the problems
exhibited when structural systems are identified from response measurements only.
In such systems there will always be a mixture of modes originating from the
structural system and the excitation. The question is whether this mixture affects the
structural modes in some ways or not. It might also be that the measured data are
either velocity or accelerations instead of displacements. Again the question is how
this affects the identification of the structural modes. This section starts by modal
decomposition of the combined continuous-time structural system. It is obvious that
the eigenvalues of the combined continuous-time system must be the eigenvalues of
the structural system and of the shaping filter. However, the question is whether or
not the mode shapes are affected by the choice of output and by the convolution of
the structural system and the shaping filter. Therefore, the emphasis is placed on
analysing how the mode shapes of the structural system are affected by way the
combined system is obtained.

Continuous-Time Structural Systems 59

3.4.1 Modal Decomposition of the Combined System

The free vibrations of the structural system are, in principle, obtained from the
homogeneous part of the p-variate differential equation (3.45) of the structural system

D sz( t )  As 1 D s 1z( t )  ...  A1Dz( t )  A0 z( t )

0 (3.57)

However, there is a problem since the dynamics of the structural system is mixed up
with the dynamics of the excitation. The modal decomposition can as such not be
applied to (3.57), but must be applied to the homogeneous part of the combined
continuous-time system (3.39) instead, i.e. to

D nz( t )  Cn 1 D n 1z( t )  ...  C1 Dz( t )  C0 z( t )

0 (3.58)

Since s  n there will be situations where only a subset of the modes of the combined
system originates from the structural dynamics. The modal decomposition of the
combined continuous-time system is obtained from the homogeneous part of the state
equation in (3.46)
t )
Fx( t )
x( (3.59)

with x(t) and F defined by (3.44) and (3.45). As in section 3.1.2, the solution form
is assumed as

x( t )
5 e  t (3.60)

where 5 is a complex vector of the dimension np × 1 and  is a complex constant.

Insertion of (3.60) into (3.59) shows that (3.60) is only a solution if and only if 5 is
a solution to the first-order eigenvalue problem
I F 5
0 (3.61)

This eigenvalue problem has only non-trivial solutions if the characteristic

polynomial, given by
det I  F
0 (3.62)

is satisfied. The order of this real-valued characteristic polynomial is np. Thus, there
will be np roots j, which are the eigenvalues of F. For each of these eigenvalues a
non-trivial solution vector 5j exists, which are the corresponding eigenvector. In view
of the structure of x(t) given by (3.44), it follows that 5j must have the form

60 Modal Decomposition of Combined Continuous-Time Systems

 0j
j 0j
.
5j
, j
1, 2, ... , np (3.63)
.
n 2
j 0j
n 1
j 0j

Insertion of F and (3.63) into (3.46) yields

n n 1
I j  Cn 1 j  ...  C1 j  C0 0j
0 (3.64)

which is an nth order eigenvalue problem, with 0j being a non-trivial solution vector
of it. Thus, for the jth of the sp eigenvalues of the underdamped structural system 0j
is actually the associated complex mode shape. By use of the modal matrix 4, and
its inverse, F can be diagonalized as

4 1F 4
 , 
diag { j } (3.65)

where  is a diagonal matrix of np eigenvalues j. The complex modal matrix is

formed by all np eigenvectors as
01 02 . . 0np 1 0np
1 01 2 02 . . np 1 0np 1 np 0np
. . . . . .
4
(3.66)
. . . . . .
n 2 n 2 n 2 n 2
1 01 2 02 . . np 1 0np 1 np 0np
n 1 n 1 n 1 n 1
1 01 2 02 . . np 1 0np 1 np 0np

The modal decomposition of F is then trivially obtained as

F
4  4 1 (3.67)

It should be noticed that the np eigenvectors, and as such also the sp mode shapes of
the structural modes, may be arbitrarily scaled without affecting the modal

Continuous-Time Structural Systems 61

decomposition. For this reason the mode shapes obtained from this modal decompo-
sition are called scaled mode shapes. It has been assumed that all np eigenvalues are
distinct, since this assumption is needed for F to be diagonalizable. When the
eigenvalues are not distinct,  will not be a diagonal matrix, but have diagonal blocks
for distinct eigenvalues and Jordan blocks for repeated ones. 4 will in this case
consist of eigenvectors and generalized eigenvectors, see e.g. Gohberg et al. [30] and
Metha et al. [78].

3.4.2 Effects of Arbitrarily Chosen Output

The free vibrations of the state space system (3.46) are given by
t )
Fx( t )
x(
(3.68)
y( t )
Cx( t )

The scaled mode shapes of the previous section were obtained under the assumption
that the output of this system was the displacement z(t) of the mass points. Thus, by
using the observation equation in (3.68) with C = H, the extraction of the scaled
mode shape 0j may be written as
0j
H 5j , j
1, 2, ... , sp (3.69)

However, from a system identification point of view, it is interesting to investigate

how the mode shapes are affected by the choice of output, since the measured
response of a civil engineering structure is often its accelerations. For a general
choice of observation matrix C, define temporarily the generalized mode shapes Qj.
These mode shapes are then obtained as
Qj
C5j , j
1, 2, ... , sp (3.70)

This relation clearly shows that the generalized mode shapes certainly depends on the
choice of output y(t), and that they in some cases will be scaled versions of 0j.
Consider the three special cases (3.47), (3.48) and (3.49). It follows from (3.70) that
if y(t) is the displacement of the system then
Qj
C 5j
I 0 0 . . 0 0 5j
0j , j
1, 2, ... , sp (3.71)

This result is of course equivalent to (3.69). If y(t) is the velocity of the system then
Qj
C 5j
0 I 0 . . 0 0 5j
0j j , j
1, 2, ... , sp (3.72)

62 Modal Decomposition of Combined Continuous-Time Systems

and finally, if y(t) is the acceleration then
2
Qj
C 5j
0 0 I . . 0 0 5j
0j j , j
1, 2, ... , sp (3.73)

It is seen that all mode shapes 0j are individually scaled due to change of C, but since
the mode shapes 0j are already arbitrarily scaled this additional scaling is immaterial
for the modal decomposition. Therefore, there is no point in distinguishing between
Qj and 0j, and (3.70) can just as well be written in terms of 0j as
0j
C5j , j
1, 2, ... , sp (3.74)

Finally, the question is whether or not the convolution of the structural system and
the shaping filter of the excitation have affected the scaled mode shapes of the
structural system or not.

3.4.3 Effects of the Shaping Filter Convolution

Assume that the observed output is the displacement z(t) of the system, and consider
the free vibrations of the combined system (3.37)
A 1 B 1 H2
x c( t )
xc( t )
0 A2 (3.75)
z( t )
H1 0 xc( t )

The structural system is described by the matrix triple { A1, B1, H1} see (3.33),
whereas {A2, H2} relates to the shaping filter of the excitation, see (3.34). Since this
realization is similar to (3.46) the np eigenvalues j obtained from (3.62) also
satisfies the characteristic polynomial
A 1 B 1 H2
det  I
0 (3.76)
0 A2

and it is thus possible to find np non-trivial eigenvectors 5j,c that satisfy the
eigenvalue problem
A 1 B 1 H2
j I 5c, j
0 , j
1, 2, ... , np (3.77)
0 A2

The characteristic polynomial (3.76) can also be expressed as

Continuous-Time Structural Systems 63

 det  I A1 det  I A2
0 (3.78)

which, not surprisingly, implies that the eigenvalues of the combined system are
exactly the eigenvalues of the structural system and of the shaping filter. If the
eigenvector of (3.77) is partitioned in two, the jth eigenvalue problem can be
formulated as
(1)
A 1 B 1 H2 5c, j
j I
0 , j
1, 2, ... , np (3.79)
0 A2 (2)
5c, j

or equivalently as
(1) (2)
j I A1 5c, j  B1H25c, j
0
j
1, 2, ... , np (3.80)
(2)
j I A2 5c, j
0

It is seen that if the eigenvalue j originates from the shaping filter, the second
(2)
equation of (3.80) is an eigenvalue problem with 5c, j as the corresponding
eigenvector. From the first equation of (3.80), the remaining part of the eigenvector
5c,j can be determined as

5c, j
j I A1 1B1 H25c, j ,
(1) (2)
j
1, 2, ... , np (3.81)

The observable part of the eigenvector, which is the “mode shape” of the shaping
filter, is then calculated by
(1)
5c, j

H1 j I A1 1B1 H25c, j ,
(2)
0j
H1 0 j
1, 2, ... , np (3.82)
(2)
5c, j

From (3.81) and (3.82) it is seen that the vector 0j corresponding to an eigenvalue of
the shaping filter, has changed due to the convolution process. It should also be
noticed from (3.81) that the eigenvalues of the structural system must be different
from the eigenvalues of the shaping filter, otherwise the matrix (jI - A1) will lose
rank and become impossible to invert. However, this does not create any problems
here, since all eigenvalue are assumed distinct. Now, assume that the eigenvalue j
originates from the structural system. In this case the matrix (jI - A2) in (3.80) will
be non-singular and positive definite. Therefore, the only way that the second

64 Modal Decomposition of Combined Continuous-Time Systems

 (2)
equation in (3.80) can be satisfied is if and only if 5c, j is zero, which implies that the
first equation in (3.80) is given by
(1)
j I A1 5c, j
0 , j
1, 2, ... , np (3.83)

This equation is a standard first-order eigenvalue problem for the structural system,
(1)
with 5c, j being a non-trivial eigenvector. Since this eigenvalue problem relates to the
structural system the observable part of the eigenvector will indeed be a true mode
shape. The associated scaled mode shape is therefore obtained from
(1)
5c, j
0j
H1 0
0j , j
1, 2, ... , np (3.84)
0

and it is seen to be equivalent to the original mode shape associated with the jth
eigenvalue of the structural system.

So in conclusion:

/ The scaled mode shapes of the structural system are not affected by the
convolution of the shaping filter of the excitation. The only possible scaling
they may exhibit is individual and depends on what the output is according
to (3.70).

3.4.4 Example 3.3: Modal Decomposition

In section 3.3.2 the second-order white noise excited structural system was represented by the
state space realization

t )
F x( t )  B w( t )
x(
(3.85)
y( t )
C x( t )

with the matrix triple {F, B, C}of the combined continuous-time system (3.46) is given by

0 1 0
F
, B
, C
1 0 (3.86)
k c 1
m m m

The eigenvalue problems can then be formulated as

j 0 0 1 1 0

, j
1, 2 (3.87)
0 j k
c
j 0
m m

Continuous-Time Structural Systems 65

 and the corresponding characteristic polynomial is

2 c k
j  j 
0 (3.88)
m m

The solution of this polynomial leads to two complex conjugated eigenvalues given by

c c 2 4 m k c c 2 4 m k
1
, 2
(3.89)
2m 2m

and the similarity transformation

1 1 1 1 0 2 1
F
(3.90)
2 1 1 2 0 2 1 1

3.5 Summary
This chapter has dealt with the continuous-time mathematical description of
structural systems. The concept of modal analysis has been introduced on a second-
order differential equation system. It has been shown how the dynamic behaviour of
such a system can be broken down into subproblems by a modal decomposition. Each
of the subproblems are described by modal parameters, which can be divided into
modal frequencies, modal damping, modal vectors, and modal scaling. Since the
number of measurement channels in general is lower than the identified number of
dynamic modes of a system, it is necessary to generalize the mathematical description
of the structural system. Further, to use the ARMAV model as a discrete-time
representation of an ambient excited structure, the ambient excitation is assumed to
be the output of a linear shaping filter subjected to Gaussian white noise. Since the
ARMAV model assumes a Gaussian white noise, the generalized structural system
and the shaping filter have been combined into one mathematical model, which is
excited by continuous-time Gaussian white noise. Further, to be able to extract other
characteristics than the displacements of the observed mass point of the structure, this
combined system has been generalized to make this possible. Through a modal
decomposition of the combined system it has been shown that it is still possible to
extract the modal frequencies, modal damping, and modal vectors of the structural
system. However, the modal scaling will be affected by the shaping filter.

66 Summary