Techniques of Identi Fication
Techniques of Identi Fication
Techniques of Identi Fication
Techniques of Identication
8.1 Introduction
The need to precisely and accurately predict the dynamic behaviour of increasingly
complex machines or, more generally, of installations and structures, and the need
to predict their reliability and life span, have led to the need to develop increasingly
rened and complex calculation methods which have become indispensable tools
for design and operation. Such methods, based on the use of suitable mathematical
models, such as multi-body methods and nite element methods for mechanical
systems, seek to reproduce the behaviour of the real system and make it possible not
only to predict the systems dynamic response to the action of different forms of
excitation but also its stability in different working conditions. During the design
phase, the advantage of having a sophisticated mathematical model offers the
possibility of avoiding partly, if not entirely, any experimental investigations on
prototypes or individual components, which would otherwise be necessary and
which is usually quite costly and requires often prohibitive implementation times.
The mathematical model associated with the real system is usually dened by:
assuming that the system in question has certain properties (related to the laws
of physics or to purely mathematical hypothesis);
including consolidated partial mathematical models.
So far in the discussion we have always assumed that the elastic, inertial and
damping characteristics of the mechanical system to be analysed are known (in
terms of stiffness, mass and concentrated and/or distributed damping), as is the type
of excitation that the system is subject to. With these assumptions, once a certain
mathematical model had been dened that was able to reproduce the true behaviour
of the system, the relevant equations of motion were dened and then they were
solved. Often, however, it is not possible to accurately dene the value of some of
the systems typical parameters (such as, for example, the stiffness and damping
coefcients of some components). In this case the mathematical models are subject
to uncertainties and the comparison between analytical and experimental results
may show signicant discrepancies, thus making the results obtained with the
mathematical model unreliable. Sometimes, given the complexity of the system, it
may not even be possible to dene a mathematical model a priori whose equations
make it possible to adequately reproduce the true behaviour of the system. To
overcome these difculties several methods have been developed over recent
decades, known as identication techniques, (Proceeding IFAC/IFOR,
Proceeding on Modal Analysis, [20]) that use experimental measurements on real
systems, making it possible to obtain:
the characteristics of several parameters of a pre-established mathematical
model of the considered system (techniques of parameter identication);
the equations relating to a mathematical model not corresponding to a physical
scheme, but reproducing the system as a black box (system identication
techniques);
the characteristics of the sources of excitation that it is subject to.
So the term identication techniques means the set of methodologies which,
starting from experimental measurements, aims to dene the characteristics of a
generic system: these methodologies were developed mainly as part of controls and
were only later extended to the study of systems of a mechanical, electrical, elec-
tronic, economic, biological nature etc., as well as to the study of generic processes.
In the specic issues of identication particular nomenclature and symbology are
traditionally used, usually by experts in the eld of controls, where these tech-
niques, as already mentioned, have undergone important developments. The gen-
eric term system is used to dene a set of several components (material objects)
which, being physically interconnected, act as a whole. In general, a mechanical
system is made up of a combination of interconnected rigid or flexible bodies with
elastic and damping elements interfaced with actuators or motors controlled in an
open or closed loop (e.g. a road or rail vehicle, a robot, a rotating machine,
including a rotor, bearings, case and foundation etc.). Anything that is not part of
the system is usually dened as the outside world separated from the system by a
physical (or ideal) surface. The objects that make up the system are known as
components or subsystems: the second term refers to the fact that, depending
on the type of investigation being carried out, a certain combination of objects may
be considered to be a stand-alone system or a subsystem (e.g. an articulated rod and
crank mechanism can itself be a system, while it can also become a subsystem of a
complete machine). A system that is subject to certain sources of excitation shows
certain behaviour dened as the system response: an unbalanced rotor shows
bending vibrations, a structure that is hit by turbulent wind oscillates around its
position of equilibrium etc. The dynamic behaviour of a generic system, or the
dynamic process that system is subject to, can often be described by relations
between:
8.1 Introduction 773
F i x; x_ ; x F s x; x_ F e x F x; x_ ; t 8:1a
where x is the vector of the independent variables, F i , F s , and F e are the forces of
inertia, damping and elastic (in general, non-linear functions of the independent
variables x). Lastly, F can group together all the forces not included to the left of the
equals sign, associated with eld actions, with the presence of a control system or,
more generally, with external forces that are explicit functions of time. In the case
where the system is linear, or it is linearized (see Chap. 5, regarding systems
surrounded by force elds), then (8.1a) becomes:
M x R_x K x F t 8:1b
where [M], [R] and [K] are, respectively, the equivalent matrices of mass, damping
and stiffness of the actual system. As already illustrated in Sect. 2.4.2.2, (8.1b)
dened with y the vector of velocity:
y x_ 8:2a
8.1 Introduction 775
Equation (8.1b) can be redened on the basis of the so-called state variables z:
x
z 8:2b
y
i.e. they can be rewritten in the generic form, also known as the state transition
function:
z_ f z; F t 8:2c
z_ Az ut 8:2d
(8.2c) or (8.2d) make it possible to determine the state of the system. In (8.2c) or
(8.2d) the vector F t or ut represents the input of the system it, while usually,
in mechanical systems, the output is identied with one or more of the state
quantities z (8.2d). More generally, the system outputs yu t are correlated to the
state variables z by a relation, also dened as the output transformation function, of
the form:
yu t yu zt; it 8:3
which denes the output yu t of the system in relation to the existing input it
and of the state of the system z. In the case of a generic vibrating mechanical
n-degree-of-freedom (d.o.f.) system, for example, the state variables may be rep-
resented by displacement x and by velocity x_ of the d.o.f. associated with the motion
of the individual bodies making up the actual system. One possible output trans-
formation can be represented by the stresses induced on a component of the system,
a function, in general, both of the inputs and the state of the actual system.
Identication of the frequency domain generally aims to identify the harmonic
transfer function of the system. The techniques used to identify the modal
parameters of a mechanical system that is considered linear can be traced back to
this category. A non-structured model, i.e. a black box model can, in actual fact, be
represented in the frequency domain (Sect. 2.4.1.4.1) by a relation of the form:
776 8 Techniques of Identication
X X H XF X 8:4
Introduced by Gauss in 1795 [1, 17] in relation to studies on the orbital motions of
the planets, the least squares method can be used as an identication technique.
Gauss realised that estimation of unknown parameters using experimental obser-
vations was conditioned by the number of measurements carried out and by the
errors that affect them. The studies by Gauss also led to the conclusion that esti-
mation was also influenced by the degree of approximation of the dynamic model.
The basic idea of the method is to minimise, in probabilistic terms, the difference
between the value predicted by the mathematical model and the observed value; in
other words it denes the best estimation of the most likely value of the unknown
parameters, assuming errors in the measurements that are independent of each
other. Let us assume that we wish to describe the behaviour of a certain real system
using the dynamic equation of motion (8.2a)(8.2f) where we consider as
unknowns the masses, stiffness and damping ranked in a vector p:
s
Suppose that we have n measured quantities xk xs tk at m discrete moments of
time t1 ; t2 ; . . .; tm :1 We use xk x p; tk to indicate the value of displacements x
obtained at the generic moment of time t tk to integrate the equation of motion
(8.1a, 8.1b) having introduced a set of trial parameters p. The least squares method
makes it possible to obtain the most likely values of the parameters p by minimising
the sum of the squared residuals which can be formally expressed as:
m T
1X s s
J xk xk p Wk xk xk p 8:5b
2 k1
where [Wk] is a matrix of weights [28]. This least squares approach can also be used
in the frequency domain. For example, when referring to a 1 d.o.f. system, the
transfer function hX is (Sect. 2.4.1.4.1):
1
hX 8:6a
X m iXr k
2
hs hs Xk 8:6b
1
In the following discussion we will use the superscript(s) to indicate the experimentally measured
generic quantity.
778 8 Techniques of Identication
derivatives with respect to the actual parameters; this corresponds to solve a system
of non-linear algebraic equations.
8.2.1 Introduction
The term modal identication refers to the set of methods used to describe the
dynamic properties (characterised by modal parameters) of a structure or machine
based on experimental data obtained by testing the actual structure.
Modal identication uses the modal approach as analytical support, meaning that
it reconstructs the analytical transfer function hjk X of the system considering it as
if it were made up of many one d.o.f. systems (Sect. 2.5), since the various d.o.f. are
dened by the modal variables. hjk X represents the harmonic transfer function for
an input at the generic point k and the output at the generic point j of the n d.o.f.
system considered. So this transfer function will have modal parameters as
unknown quantities:
hjk X hjk X; xi ; mi ; ki ; ri ; X i 8:7
being:
xi the natural frequencies;
mi the generalised masses;
ki the generalised stiffness;
ri the modal damping;
X i the modes of vibration.
With this approach the dynamic response of the structure subjected to a set known
excitation is measured at several points (the relative quantities will be shown below
with the subscript j); this dynamic response, expressed in terms of experimental
s
transfer functions hjk X, is then compared to the analytical response hjk X dened
beforehand by minimising the difference between the analytical values and the
experimental ones. It is thus possible to determine the set of modal parameters needed
to characterise the static and dynamic behaviour of the system being analysed.
Now, for ease of discussion, we will outline the main results obtained by analysing
a discrete generic n d.o.f. system using the modal approach (the full discussion of
8.2 Modal Identication Techniques 779
this issue has already been dealt with in Sect. 2.5.2). So, we will consider a generic
linear vibrating system, discretized with a n d.o.f. model (Fig. 8.2): the vibrating
system can be described, regardless of the method of discretization (with concen-
trated parameters, with nite elements etc.) by a system of second-order linear
differential equations with constant coefcients. In matrix form we have:
M x R_x K x F t 8:8a
where x represents the vector of the independent coordinates that represent the
displacements of some points of the structure:
[M], [R] and [K] are, respectively, the matrices of mass, damping and stiffness and
lastly, F is the vector of external forces. We can use transformation into principal
coordinates, see Sect. 2.5:
x Uq 8:8c
i
Xj is the generic jth component relative to the ith mode of vibration. So by
applying the transformation of coordinates (8.2c) to the Eq. (8.8a), the same
equation of motion of the system can be rewritten, if the damping matrix [R] is a
linear combination of those of mass and stiffness, such as:
780 8 Techniques of Identication
[m], [r] and [k] being diagonal. The equations of motion of the generic mode of
vibration can be rewritten as n distinct and uncoupled equations:
X
n
i
qi ri q_ i ki qi Qi t X iT F t
mi Xk Fk t i 1; p 8:10
k1
Fk t being the generic k-th generalised component of the forces vector F according
to the kth d.o.f. Xk . So to fully describe the dynamic behaviour of the discretized
system with a n d.o.f. model it is enough to know:
the natural pulsation of the system x2i mkii ;
the generalised mass mi and, from the previous denition, the generalised
stiffness ki ;
the generalised damping ri ;
the modes of vibration X i .
These modal parameters may be determined:
analytically, when the characteristics of the elements making up the structure in
question are known a priori, i.e. the matrices M , R and K of the Eq. (8.8a),
and by subsequently applying the transformation of coordinates (8.8c) to obtain
(8.10);
experimentally, by measuring the response of the real structure and assessing the
same modal parameters at a later stage.
The modal identication techniques described in this section are, as already men-
tioned, based on the experimental denition of the dynamic response of the system
s
using a harmonic transfer function hjk X (in this regard, see Sect. 2.4.1.4.1),
which can be obtained, for a linear system, by exciting the system with one or more
harmonic forces like Fk t Fko eiXt or with one or more assigned spectrum forces
(8.3).
Basically, there are two different approaches used to identify the modal
parameters of a mechanical system (as will be further explained and illustrated
below):
using a single external excitation (single-input; multi-output);
using different excitations simultaneously (multi-input; multi-output).
Here, for educational purposes, we will analyse several methods that require just
one input: so when adopting such an approach the Lagrangian component Qi t
relative to the generic ith mode of vibration, caused solely by the harmonic force
applied Fk Fko eiXt becomes:
8.2 Modal Identication Techniques 781
i
Qi t Fko eiXt Xk Qio eiXt 8:11
The equation relating to the generic mode of vibration (8.10) thus becomes (steady
state motion for a single harmonic force Fko eiXt ):
i
mi qi ri q_ i ki qi Fko eiXt Xk 8:12
we have:
i
Fko eiXt Xk
qio 8:12b
X mi iXri ki
2
The response of the system, in terms of independent variables and nodal dis-
placements x, will be of the form:
x X o eiXt 8:13
The response Xojk X of the system at the generic point j estimated analytically for
an applied force Fko eiXt at generic node k is, therefore, taking into account (8.8c)
and (8.12b):
X
n
i
X
n
i Fko Xk
i
Xojk X Xj qio Xj 8:15
i1 i1 X2 mi iXri ki
s
s Xojk X
hjk X 8:16
Fko
which, therefore, represents the response of the system for a unit force applied at
generic point k. So, similarly, the analytically estimated transfer function hjk X is
dened as:
s i i
Xojk X X
n
Xj Xk
hjk X 8:17
Fko i1 X2 mi iXri ki
i i i i
Xj Xk i Xj Xk
hjk X Xxi 8:18a
ixi ri xi ri
i i
i Xj Xk
hjk XXxi i 8:18b
2mi rrc x2i
8.2 Modal Identication Techniques 783
i i
i Xj Xk
hjk XXxi 8:20
2mi ri xi
(8.17) and (8.20) represent, in the case of pulsation of the generic force X, and in
the case of X xi , the response of the vibrating system in terms of harmonic
transfer function, i.e. the response of the jth generic d.o.f. for a harmonic force
applied at the kth point. As we can see, this function is uniquely dened as the
modal parameters of the system are known. Using the identication techniques to
minimise the difference between the measured experimental transfer function
s
hjk X (8.16) and the analytical transfer function hjk X, dened in relation to the
function of modal parameters (unknown) (8.17) and (8.20), it is possible to estimate
the actual modal parameters, i.e. natural pulsation xi , generalised mass mi , gen-
eralised damping ri and mode of vibration X i . In the denition of the analytical
transfer function hjk X, it is possible, about xi , to use the expression (8.17), while
neglecting the contribution of non-resonant modes in the summation: in this case a
dened modal approach with one d.o.f. is used. The use of (8.17) in complete form,
on the other hand, considers the contribution of the other modes and so, in this case,
a several d.o.f. approach is used (in this regard, see [8, 11, 22]).
The generic transfer function hjk X (both analytical and experimental) can be
plotted in different ways taking into account the difculty in representing, in a
plane, a complex quantity fully dened by 3 quantities:
frequency ;
two scalar quantities [magnitude hjk X and phase /jk X];
real part Re hjk X and imaginary part Im hjk X .
The three most common graphic representations are summarised below:
Bode plot,
which consists of two graphs that show, respectively, magnitude
hjk X and phase /jk X of the transfer function in relation to frequency X
(Fig. 8.4 shows this quantity for a 1 d.o.f. system);
784 8 Techniques of Identication
the graph of the real part Re hjk X and the imaginary part Im hjk X of the
transfer function in relation to the frequency
(Fig. 8.5);
the graph of the imaginary part Im hjk X in relation to the real part
Re hjk X upon transformation of X: this graph, known as the Nyquist plot or
plot in the Argand plane, does not explicitly contain information about the
frequency (Fig. 8.6).
This graph (in the Argand plane) represents a circumference: the vector P O
denes the magnitude and phase of the transfer function hjk X in correspondence
to a generic value of . The vector B O represents the transfer function in
conditions of resonance with phase /jk X equal to 90.
Often the Bode plot or the graph of the real and imaginary parts is represented in
logarithmic scale to avoid the issues linked with the graphic representation of
quantities, such as, in fact, the transfer function, which can attain values that are
very different to each other.
These different representations of the same quantity hjk X are used during the
identication phase in relation to the algorithm adopted. For example, Figs. 8.7, 8.8
and 8.9 show, respectively, the Bode plots of the real and imaginary part and the
Nyquist plots for a vibrating 2 d.o.f. system.
8.2 Modal Identication Techniques 785
In this section we will briefly deal with the main methods that can be used to
s
experimentally measure the transfer function hjk X in (8.16), while referring to
specialised texts (listed in the bibliography) and to other courses (e.g. courses on
786 8 Techniques of Identication
The structure may be analysed either when free or constrained. If the structure is
analysed while free (i.e. free from constraints with the outside world) it is possible
to dene its stiffness modes and therefore the overall characteristics of mass and
moments of inertia. In reality this condition cannot be achieved and, in practice, to
reproduce the behaviour of the free structure the actual structure must be con-
strained with interconnecting elastic elements (springs) that are extremely soft so
that the initial modes of the structure are rigid with frequencies close to zero and, in
any case, much lower than those relating to the vibration modes that cause the
788 8 Techniques of Identication
s
s Xojk X
hjk X 8:21
Fko
One advantage of this methodology is that it offers the possibility to set the pitch
DX and to therefore densify the points of analysis in proximity of the natural
frequencies xi and disperse them in areas further away from these.
Sinusoidal excitation with slow variable pulsation This methodology requires a
variable frequency oscillator X Xt that sends an analogue control signal to
the exciter: the frequency of excitation caused by the force Fk Fko eiXt is thus
made to vary slowly but constantly within the range of relevant frequencies.
Variation in pulsation (up or down) must take place quite slowly so that the
system can sweep constantly from one steady state to another: we should
remember that the harmonic transfer function is dened as being the relationship
s
between the response of the system in a steady state Xojk X and the harmonic
force introduced Fko . To verify the validity of the sweep speed a run-up is
usually performed (i.e. a sweep with increasing pulsation frequency ) and a
790 8 Techniques of Identication
run-down (with decreasing frequency ), to check that the two curves obtained
are coincident.
s
In this case too, the experimental transfer function hjk X is dened for each
pulsation of the force and for each point of measurement j on the structure by
(8.28).
Periodic excitation An extension of the previous method is to set periodic
excitation, i.e. a force that contains all the frequencies in the eld of analysis:
X
Fk t Fko ei n Xo t 8:22
n
Xo being the pulsation of the fundamental frequency. Unlike the method that
uses a harmonic excitation, in this case the force is obtained as an overlap of
different harmonics in the frequency range concerned. The denition of the
s
experimental harmonic transfer hjk X using a periodic excitation is based on
the fFourier transform: more specically, harmonic analysis is performed (i.e. the
Fourier transform) of both the force signal Fk t and the dynamic response of the
s
system measured at generic point Xojk nXo :
ZT
2
Fkon nXo Fk t ei n Xo t dt 8:23a
T
0
ZT
s 2 s
Xojk nXo Xjk t ei n Xo t dt 8:23b
T
0
s
s Xojk X
hjk X 8:24
Fkn
s
The experimental transfer function hjk Xn can be dened by simply dividing the
s
spectrum of the response Xojk nXo (8.23b) by the spectrum of the force applied
Fkon nXo , Eq. (8.23a).
8.2 Modal Identication Techniques 791
Random excitation With this type of excitation the structure is forced with a
white noise (corresponding, as seen in Sect. 7.3, to a spectral power density
function of constant amplitude). Thus the spectrum of the response corresponds
to the transfer function of the system.
Impulse excitation When using this methodology, the system is subjected to an
assigned impulse (for example, using an impact hammer). The Fourier transform
of the ideal impulse is characterised by a constant amplitude throughout the eld
of frequencies. However, considering that the impulse is not actually perfect,
transformation of the force signal and the displacement signal is, in any case,
performed at the different points of measurement: the ratio of the corresponding
s
harmonics at various frequencies Xn denes the transfer function hjk X:
s
s Xojk Xn
hjk Xn 8:25
Fkn Xn
Having dened in Sects. 8.2.2 and 8.2.3 the basic equations needed to dene the
a
link between analytic transfer function hjk X and unknown model parameters to
be used in the modal identication techniques and, having experimentally obtained
s
the transfer function hjk X by measuring the dynamic response of the system with
assigned forces (Sect. 8.2.4), we will now proceed to describe some of the most
basic algorithms used to dene the actual modal parameters, i.e.:
natural frequencies xi ;
non-dimensional damping r
rc i ;
vibration modes X i ;
generalised mass mi .
A simple method used to dene the natural pulsations of the system xi is based on
s
the consideration that the amplitudes of vibration Xojk X, and, therefore, the cor-
s
responding transfer functions hjk X of a resonance forced system, are amplied
dynamically: in the summation given in (8.17) the contributions made by non-
792 8 Techniques of Identication
s
Fig. 8.11 Trend of the experimental transfer function magnitude hjk X in relation to the
frequency X of the force
resonant modes are minimal, while that of the resonant mode is higher, being
X xi. After having plotted the trend of the experimental transfer function mag-
s
nitude hjk X in relation to the frequency X of the force (Fig. 8.11), the method
allows us to dene the natural frequencies as those frequencies to which there is a
corresponding maximum dynamic amplication,
that is, a peak in the magnitude of
s
the harmonic transfer function hjk X.
The lesser the damping of the system the more accurate the results obtained will
be, as the contribution of the non-resonant vibration modes is less signicant. The
method may cause errors in determining the natural frequencies if the experimental
transfer function analysed is that of a vibration node for the generic mode of
vibration: in actual fact, in this case even in conditions
of resonance, displacement
s
is very limited so a peak in the plot of the hjk X is unlikely to be highlighted. To
s
overcome this drawback, the transfer functions hjk X at different points of the
structure should be analysed: correlation of the various experimental harmonic
transfer functions also makes it possible to eliminate random errors in measurement
and to exclude the effect of any other external sources of excitation. The point of
application of the force could be in a nodal point for one or more modes: in this
case, these modes would not be excited. In these situations the point of application
of the force must be moved to a different position.
8.2 Modal Identication Techniques 793
Fig. 8.12 Determining damping with the half power peaks method
r
8.2.5.2 Calculating Generalised Damping rc i
The justication for such an approach is as follows: the response of the system, in
terms of analytical transfer function, has been dened in (8.17): obviously this
response is complex. Under conditions of resonance X xi , if we consider neg-
ligible the contribution of non-resonant modes, we have:
i i
Xj Xk i i X2 mi ki iXri
hjk X Xj Xk 2 2 8:27
X2 mi iXri ki X mi ki Xri 2
s
while the phase /jk X is dened as:
Xr 2ai hi
/jk X arctg arctg arctgb 8:28
X2 mi ki 1 a2i
794 8 Techniques of Identication
2ai hi
b
1 a2i
X
ai 8:28a
xi
i
ri r ri
hi
rci rc 2mi xi
s
The derivative of /jk X with regard to ai is:
@/jk X @arctgb @b 1 @b
8:29a
@ai @b @ai 1 b2 @ai
that is:
@/jk X 1 2hi 2ai hi 2hi 1 a2i 4ai hi
2 2ai
@ai 1 a2i 1 a2i 2 2
1 a2i 4a2i h2i
1 2ai hi
1a2i
8:29b
hence:
@/ X 1
jk 1
8:30b
@ai ai 1 hi ri=r
ci
The most commonly used expression, considering (8.30b) and (8.28), is the
following:
@/ X @/ X @a
jk jk i 1 1
8:31
@X Xxi @ai ai 1 @X ri=r xi ri
ci
ri h i x i 8:31a
The non-dimensional damping h of the i-th mode of vibration can thus be obtained
s
by plotting the progress of the phase /jk X of the experimental transfer function
8.2 Modal Identication Techniques 795
s
hjk X in relation to the pulsation X of the force and by estimating the derivative of
the phase compared to the actual pulsation X in resonance, that is, for X xi .
This method can be used in vibrating systems with several d.o.f. where the natural
frequencies are sufciently differentiated and damping is sufciently low, so that,
under resonance conditions the predominant contribution to the response of the
actual system is mainly due to the resonant mode, making it behave like a 1 d.o.f.
system. For a 1 d.o.f. system we have seen how in resonance the coefcient of
dynamic amplication (Sect. 2.4.1.4.1) is:
Xo 1 X
per 1 8:32
Xst 2h xo
h rrc being the non-dimensional damping coefcient. The two frequencies x1 and
x2 are dened as half power points (Fig. 8.12), about the resonance (i.e. in the
sidebands to the peak) which correspond to:
X0 1 p 1
2 for X x1 and X x2 8:32a
Xst 2 2h
2
The response of the system in correspondence to a generic pulsation of the force, in non-
dimensional form (Sect. 2.4.1.4.1) can be expressed as:
Xo 1
s
8:2:1
Xst 2
2 2
1 xXo 2hX xo
So assuming:
1 1 1 2
2 2 2 2 2h 8:2:2
1 xXo 2hX
xo
we obtain
4
X X 2
2 1 2h2 1 8h2 0 8:2:3
xo xo
hence
796 8 Techniques of Identication
x22 x21 x x1
4h 2 2 8:32b
x02 x0
Several methods can be used to determine the vibration mode. One simple approach
can be that which uses the imaginary part of the generic transfer function. As
mentioned, the unknown modal parameters are estimated to minimise the difference
s
between the experimental hjk X and the analytical hjk X transfer function (8.15)
expressed as a function of the same unknown parameters.
s
Both the analytical hjk X and the experimental hjk X transfer function are, as
already mentioned, complex and therefore, using Re. . . to indicate the real part of
the generic quantity and Im. . . we have
X i
X i
X i
hjk X Xj qio Xj Reqio i Xj Imqio 8:33
(Footnote 2 continued)
2 p
X
1 2h2 2h 1 h2 8:2:4
xo
Assuming that h 1 it is possible to neglect the higher order terms in (8.2.4), to obtain:
2
X
1 2h 8:2:5
xo
By assuming that the two corresponding frequencies at the roots of this equation are equal to x1
and x2 we have:
x22 x21 x x1
4h 2 2 8:2:6
x20 x0
8.2 Modal Identication Techniques 797
and
s s s
hjk X Re hjk X iIm hjk X 8:34a
i i
Xj Xk
s s s
hjk X Re hjk X iIm hjk X i 8:34b
Xri
analysing a single d.o.f. system. We will analyse the harmonic transfer function of a
1 d.o.f. system (see Sect. 2.4.1.4.1), represented on the Nyquist plot in Fig. 8.6:
k X2 m iXr
hX 2 8:35a
k X2 m Xr 2
k X2 m
RehX 2 8:35b
k X2 m Xr 2
Xr
ImhX 2 8:35c
k X m Xr 2
2
aX i X hX 8:36a
2
2
1 1
ReaX ImaX2 8:36b
2r 2r
This function is always shown in the Argand plane as in Fig. 8.14. The plot of the
imaginary part ImaX as a function of the real part ReaX by X which ranges
from zero to innity is thus represented by a circle of radius R equal to R 1=2r
and its centre has coordinates:
1
ReaX 8:37a
2r
ImaX 0 8:37b
It can also be shown that at the same pitch DX of the denition of the transfer
function, the arc of the circle dened by two successive positions of the vector aX
and aX DX in the Argand plane (see Fig. 8.14) is greater in the vicinity of the
natural frequency. Considering this property it is, therefore, easy to identify the
systems natural frequencies.
Taking into account (8.37a) it is also possible to obtain the damping of the
system. For this purpose it is possible (proceeding in a deterministic way or using a
process of minimization) to draw a circumference through the points in the Argand
plane, corresponding to aX about the natural frequency identied. Once the centre
and radius of the interpolating circumference has been dened, it is possible to
estimate the damping r of the system, using (8.44). By expanding to a n d.o.f.
system with a damping matrix [R] that typically may not even be a linear combi-
nation of mass and of stiffness matrices (thus non-proportional damping), in the
frequency range where the mode is predominant we have:
U
2 V
2 U 2 V 2
s j s j j j
Re hjk X Im hjk X 8:38
2ri 2ri 4r2i
i
having used xi to indicate the generic natural pulsation of the system, ri r
rc xi
the damping factor associated with it and also Uj iVj to dene the so-called residue
from which it is possible to obtain the generic mode of vibration. Equation (8.38)
represents the equation of a circumference passing through the origin of the axes,
with centre of coordinates:
Uj
xc
2ri
8:39a
Vj
yc
2ri
The residue Uj iVj associated with the ith vibration mode and relative to the jth
point of measurement is proportional to the diameter of the circumference.
Determining these circumferences allows us to dene the vibration modes of the
structure when the values of the systems natural frequencies have been dened
beforehand. So the residues may be dened, in magnitude and phase (or real part
and imaginary part), by imposing the passage of the circumference being sought
800 8 Techniques of Identication
through a suitable number of points of the transfer function close to the considered
resonance, then by applying a minimisation between the analytical function being
sought and the experimental points available. The accuracy of the results obtained
using this method is closely linked to the resolution with which the transfer
functions have been dened and to the number of points through which the passage
of the circumference sought is imposed.
s
In (8.47) hjk X is the experimental transfer function measured in correspondence
i i
to the resonance, Xj and Xk are the eigenvectors of the ith mode of vibration
measured using the method shown above (and therefore known) in correspondence
to the generic point j considered and point k where the force has been placed: the
generalised non-dimensional damping hi and natural pulsation xi have already
been dened previously. The only unknown in (8.40) is the generalised mass mi
which can be obtained by deriving f mi with respect to mi , to obtain:
PN i i
j1 Xj Xk
m i PN s
8:41
j1 i2x2i hi hjk xi
s
In (8.47) the transfer function hjk X calculated in resonance is purely imaginary
(if we neglect the contributions of the non-resonant adjacent modes) and thus the
mass mi is real.
8.2 Modal Identication Techniques 801
Remembering the denition of the natural frequency of the system xi , the gener-
alised stiffness can be easily dened using:
ki mi x2i 8:42a
r
ri 2mi x2i 8:42b
rc i
We will now consider one possible application of the results obtained using the
illustrated techniques of modal identication. We will consider the system in
Fig. 8.15 which outlines the foundation of a rotor.
This structure has been modelled with beam type nite elements; the transfer
function hjk X for a force Fk 1 eiXt applied to node 8 was calculated. Then we
assumed that we had to calculate the experimental transfer function of a real
structure by exciting it with a single excitation source. Figure 8.16 shows the
transfer function as a function of X, for the vertical and horizontal directions of
node 18.
From the peaks of these transfer functions it is possible to determine the natural
frequencies (see paragraph 8.2.5.1.1). Then the modes of vibration were identied
from the imaginary part of the transfer function (see paragraph 8.2.5.3.1).
Figure 8.17 shows the primary vibration modes identied compared to the exper-
imental ones. The derivative of the transfer function phase method was used to
estimate the damping factor hi of the ith vibration mode. Once identied, these
values make it possible to reconstruct the transfer functions which are shown
together with the experimental ones in Fig. 8.18.
8.3 Identication in the Time Domain 803
rail vehicles: in which case it is possible to obtain the response of the free system
using a specic technique known as the Random Decrement Technique [14].
This technique is based on the idea that, by extracting and averaging suitable
segments of measured time history Y1(t), associated with the response of a generic
mechanical system subjected to stochastic loads, it is possible to describe the
intrinsic properties of just the mechanical system being considered, thus eliminating
the effects caused by the actual random load applied.
The method is based on the concept that the response of the system at different
moments in time is made up of two components: the response to initial conditions
(general integral) and the response to forcing (particular integral). When operating
with numerous means on several segments Y1(t) characterised by the same initial
conditions, the part associated with random forcing which, as such, tends to
disappear, and the result of this averaging operation converges to the response
8.3 Identication in the Time Domain 805
(a) (b)
Fig. 8.19 Time histories (a) response of mechanical system to stochastic excitation; (b) free
motion of the system identied with R.D.T
of the system just at the initial conditions (Fig. 8.19). Another effect associated with
this averaging operation introduced by the R.D.T. is that of reducing the presence of
noise in the actual signal acquired.
Effectively [14] Y1(t) segments that satisfy a certain trigger condition that
determines the start T1 of each segment considered are extracted (in literature there
are several proposals for the choice of such triggers: the most commonly used
include those that consider, as a trigger, the passage of the signal measured by zero
with positive derivative, exceeding a given threshold value, peak values, etc.)
The R.D.T. technique subsequently foresees a time averaging operation on these
segments: since the time histories of the analysed signals are usually sampled at
discrete time with intervals of constant time Dt, this operation is described by the
following relation:
1X N
Xi mDt Yi mDt Ti 8:43
N i1
complex terms. The approach assumes a certain number of d.o.f. of the system
(number of modes considered in the reconstruction of the response) and considers a
certain number of measured experimental time histories, appropriately shifted over
time. After a specic set of operations has been carried out on the actual mea-
surements, it is possible to extract the systems unknown modal parameters by
tracing the problem back to a problem of eigenvalues-eigenvectors of a matrix
containing the same measurements. The method involves identifying the modes that
are actually present in the structure using an index of synthesis dened as the modal
condence factor.
So, in order to describe and justify Ibrahims method, we will consider the
equation of free motion of a n d.o.f. system:
M x R_x K x 0 8:44
X
2n
x X i eki t 8:45
i1
xs tk xtk 8:46
X
2n
x t k X i eki tk k 1. . .2r 8:47
i1
i.e. the response obtained analytically in terms of natural frequencies and vibration
modes. Assuming we have p measurement points equal to n d.o.f. and considering
the matrix:
h i h i
X s X s t1 X s t2 . . .X s t2r 8:48
With this equation parity is established between the experimental and analytical
responses of the systems free motion. The n d.o.f. system has 2n eigenvalues ki
and just as many vibration modes; as we only have, for the initial hypothesis, p = n,
measurement points, duplication is carried out over time. The individual
808 8 Techniques of Identication
experimental time histories are organised in the usual matrix form with the
instruction to phase shift them in time by one delay Dt3 . By rewriting (8.46) for a
tk Dt3 , we obtain:
X
2n X
2n X
2n
i
xs tk Dt3 xtk Dt3 X i eki tk Dt3 X i eki tk eki Dt3 X eki tk
i1 i1 i1
8:53a
s
X X i eki Dt3 8:53b
This leads to a second system of equations using the same operations described
above and which has the form:
h i
s X
X K 8:54
s
where in X the experimental time histories have been organised, shifted by a
time interval Dt3 compared to those shown in (8.48), in X the products X i eki Dt3
have been organised as in (8.51) and the matrix K has the form (8.52) (the
symbols are used to identify the quantity measured at time Dt3 ). By ordering
(8.50) and (8.54) in a single matrix equation, we obtain:
2h i3 " #
h i X s X
6 7
Us 4h i5 K Xd K 8:55
s
X
X
The method now involves accompanying (8.55) with a second equation, built in a
similar way, with the instruction to time shift the experimental time signals and the
analytical response by a time Dt1 . By rewriting the Eqs. (8.47) and (8.53a) for one
time tk Dt1 and tk Dt3 Dt1 , we obtain:
X
2n
xs tk Dt1 xtk Dt1 X i eki tk Dt1
i1
X
2n
X i eki tk eki Dt1 xs tk Dt3 Dt1 8:56
i1
X
2n X
2n
X i eki tk Dt3 Dt1 X i eki tk eki Dt1
i1 i1
8.3 Identication in the Time Domain 809
^
By organising in a single equation, as already done in (8.55), and identifying with X
the terms relative to the phase shift of Dt1 , we obtain:
2h i3 "
^ s #
h s i X X
^ 6 7
U 4h i5 aK Xd aK 8:57
^ s
X
X
where the appropriately time shifted experimentalh time i histories estimated at the
s ^ s
measurement points were organised in X ^
and X , the modes of vibration in
X (8.51) and the same multiplied by the term eki Dt3 in X (8.53b). Lastly, time
shift Dt1 is taken into account in the diagonal matrix []:
2 3
ek1 Dt1 0. . .0
4
a . . . 5 8:58
k2N Dt1
00 . . .e
Assuming that 2p = 2r demonstrates [14, 15] that the matrix Us is square and
invertible: obtaining the matrix K from the rst equation of (8.59) and replacing it
in the second we obtain:
h s ih s i1
^
U ^
U Xd Xd a 8:60a
h s i
^ ,
where the known experimental histories are organised in matrices Us and U
and matrices Xd and a contain the unknown eigenvectors and eigenvalues of the
system. A indicates the matrix containing the appropriately memorised
measurements:
h s ih s i1
^
A U ^
U 8:60b
(8.60a) becomes:
i
where X d is the generic column of the matrix Xd (8.57):
( )
i X i
Xd 8:62
i
X
i i i
AX d X d eki Dt1 ai X d 8:63b
where ai is the ith element of the diagonal of the matrix a. Equation (8.63b) is
the equation of an eigenvalue problem whose solution allows us to obtain the
i
eigensolutions ai and eigenvectors X d . The eigensolutions i are linked to the
eigenvalues ki of the original motion Eq. (8.44) by the relation:
ki ai ibi 8:64b
with:
1 c
bi tan1 i 8:64c
Dt1 bi
1
ai lnc2i b2i 8:64d
2Dt1
bi and ci being the real and imaginary part of i (8.64a). The approach described is
of a deterministic nature as it is assumed that we have 2r acquired points equal to
p points of measurement making it possible to have the square and invertible matrix
8.3 Identication in the Time Domain 811
U (8.60a). In general this does not occur and it is shown [14, 15], using a method
of minimisation, that the solution can be obtained from an eigenvalue problem of a
new matrix [A] dened as:
h ih s iT h ih iT 1
s
A ^
U ^
U Us
Us
8:65
X
2n
xs tk Dt2 xtk Dt2 X i eki tk Dt2
i1
8:66
X
2n X
2n
i ki tk ki Dt2 0i ki tk
X e e X e
i1 i1
(the superscript will be used to indicate the quantities relating to the latter time
delay Dt2 ). (8.66) is written in the form (see (8.50)):
h i
X 0s X 0 K 8:67a
The time shifts Dt2 are repeated until a number of measurements (real or pseudo-
measurements) greater or equal to the number of d.o.f. of the system are achieved.
The procedure is strictly that already described:
with the time shift Dt2 we can obtain a real or ctitious number of sensors that is
greater than n and write the Eq. (8.67a) in the form:
2h i3 " #
X s X
6 7
4h i5 K 8:68
X 0s X 0
812 8 Techniques of Identication
The time histories obtained in this manner are out of phase by an arbitrary Dt3
and organised with (8.55) in a single system:
2h i3
X s 2 3
6h i7 X
6 7
6 0s 7
h 0s i 6 X 7 6 X 0 7
^
U 6 h i 76
6
7
K Xd0 K 8:69
6 s 7 4 X
6 X 7 7 5
6 7
4h i5 X
X 0s
The solution is reduced again to an eigenvalue problem with a matrix [A] built
as in (8.66) for a deterministic approach or in (8.65) using a procedure of
minimisation depending on the number of points acquired.
The method must reconstruct a signal via the summation (8.45): if this signal is
particularly disturbed it is not possible to establish equality (8.46) by taking into
account just the n presumed vibration modes. Therefore, the meaning of n loses
importance as more ways are needed to take into account the presence of noise and,
furthermore, in general the mechanical system is a continuous system. So, the
eigenvalues and eigenvectors, real and related to noise (this describes the modes
that are indispensable for reconstructing the signal due to the presence of noise)
become 2m (of which only 2n will be those associated with the actual vibration
modes). The method does not distinguish between real and noise related modes and
proceeds by calculating 2m eigenvalues and eigenvectors exactly as described
above. In order to identify the real vibration modes of the system, it is necessary to
have an index of goodness of t. The model used is the Modal Condence Factor
(M.C.F.). Remembering (8.62):
( )
i X i
Xd 8:71
i
X
i
the method calculates the eigenvectors X d that contain, in the rst m positions, the
vibration modes (real or noise) of the system in question and in the subsequent
positions the same modes multiplied by the factor eki Dt3 . The control is performed
by exploiting this very characteristic: we calculate the complex ratio (M.C.F.)
i
between the identied eigenvector X i (rst m positions of the vector X d ) multi-
plied by eki Dt3 and that identied in the second m positions.
8.3 Identication in the Time Domain 813
i
i Xj eki Dt3
M:C:F:j 100 8:72
ji
X
having indicated Xj the jth element of the vector X i . The magnitude and phase of
the ratio (M.C.F.)j (for each component of each eigenvector) are calculated: thus the
vibration modes for which the M.C.F. has magnitude equal to 100 and zero phase
will be taken as the real vibration modes of the system.
8.3.1.1 An Example
We analysed the 4 d.o.f. system illustrated in Fig. 8.19, of which the inertial, elastic
and viscous characteristics are known.
While giving very general initial conditions, we estimated the displacement of
the four masses shown in Fig. 8.21 and this was taken as the experimental response
(Fig. 8.22).
Table 8.1 shows the comparisons between natural pulsations, vibration modes
and damping factor, between the results of the analyses with the Ibrahim method
and those calculated analytically, assumed to be accurate.
Noise was then added to the signal (Fig. 8.23) and the modal parameters were
identied. Table 8.2 again shows the accurate vibration modes compared to those
identied: note how the value of the magnitude and phase of the M.C.F. uniquely
identies the real mode of the system.
Summary The basic concepts related to parameter identication techniques are
summarized in this nal chapter. These methodologies allow for the identication
of some unknown parameters of numerical simulation models of machines and
structures, starting from experimental tests on prototypes or real structures. In
particular, the text describes modal identication techniques, both in the time and
frequency domain. Some examples are given.
8.3 Identication in the Time Domain 815
References
14. Ibrahim SR, Pappa R (1982) Large modal survey testing using the ibrahim time domain
identication technique. J Spacecr
15. Ibrahim SR, Mikulcik EC (1973) A time domain modal vibration test technique. Shock Vibrat
Bull
16. Jazwinski AH (1970) Stochastic process and ltering theory. Academic Press, London
17. Kalman RE (1960) A new approach to linear ltering and prediction problems. J Basic Eng
3543
18. Masri SF, Chassiokos AG, Chaughey TK (1988) Identication of non-linear dynamic systems
using neural networks. J Appl Mech
19. Mottershead JE, Stanway R (1986) A comparison of time and frequency domain lters for
parameter identication of vibrating systems. In: Proceedings of the international conference
on vibration problems in engineering Xian China, pp 580585
20. Natke HG (1988) A brief review on the identication on non linear systems. In: Proceedings
of the 6th international modal analysis conference, Kissimmee FI, Schenectady. pp 15691574
21. Natke HG (1988) Application of system identication in engineering. Springer, Berlin
22. Proceeding of the 9th IFAC/IFORS symposium on identication and systems parameter
estimation, Budapest, Hungarian, July 1991
23. Proceeding of the IFAC symposium on design methods of control systems, Zurich,
Switzerland, Sept 1991
24. Proceeding of the IMACS world congress, Dublin, Ireland, July 1991
25. Proceedings of the 10th international conference on modal analysis, Union College, S.E.M.,
San Diego, CA, 1992
26. Otnes RK, Enochson L (1982) Digital time series analysis. Wiley, New York
27. Randall RB (1988) Application of Bruel Kjaer equipment to frequency analysis. Research
Studies Press LTD, Bruel Kjaer, Letchworth, Hertfordshire
28. Sorenson HW (1985) Kalman ltering theory and applications. IEEE Press, New York
29. Tomlinson GR (1987) Vibration analysis and identication of non linear systems, short
course. Herriot-Watt University, Edinburgh
30. Young P (1984) Recursive estimation and time series analysis. Springer, Berlin