Dinamica dei Sistemi Meccanici– Prof.

Stefano Bruni

Lesson 6
Modal parameter identification

1 Introduction
When modelling a mechanical system, the extraction of information from measurements performed
on the system can be used to improve the model (a process known as “model updating”) and also to
verify that the model reproduces to a good degree of accuracy the actual system’s behaviour, this
latter procedure being referred to as “model validation”. Both model updating and validation require
however that appropriate parameter identification techniques are used to extract valuable
information from the available measurements. Particularly noteworthy to the topics treated in this
course are the so-called modal parameter identification techniques, aimed at extracting some of the
system’s modal parameters (e.g. natural frequencies, damping factors, modal shapes) from the
measured system’s behaviour. Note that the identification of modal parameters can be of great use
not only in the perspective of improving and/or validating a mathematical model, but can also be
used directly (i.e. without the need to define a model) to analyse the dynamic behaviour of the
system and, if needed, to identify changes in the system to improve its behaviour.
A simple example of modal parameter identification technique is the use of the “logarithmic
decrement” to estimate the non-dimensional damping factor for a 1-d.o.f. system (cf. Diana- Cheli
and Meirovitch). However, this method is limited in scope as it allows to identify one single modal
parameter and the application to multi-d.o.f. systems can be problematic due to the difficulty of
separating the different modal contributions in the free motion of the system.
In this lesson, we will introduce methods for the identification of modal parameters based on the
measure of the system’s frequency response function (FRF). To this aim, we first derive in
Section 2 an analytical expression for the FRF functions in a single-input-multi-output (SIMO)
perspective, and we show these are defined as function of the modal parameters (modal masses,
damping, stiffness and modes of vibration). Then, we shortly describe how the experimental FRF
function for the same system can be derived from an experiment by measuring the input force
applied on the system and the system’s response (Section 3). Finally, we introduce some simple
methods to derive an estimate of the modal parameters out of the experimental FRF, based on the
analytical expression derived in Section 2. This is done firstly using a more accurate approach,
requiring the use of a computer and of a suitable residual minimisation technique (Section 4.1), and
then according to a simpler approximation of the expression derived in Section 2, which allows an
analytical treatment, leading to simple formulae for the identification of parameters such as the
system’s natural frequencies, non-dimensional damping ratio, modes of vibration (Section 4.2).

2 Analytical expression of the FRF

We consider here any mechanical system, which could be either a multi-degree of freedom lumped
parameter system or a distributed parameter system, and assume the system is excited by a single
input force of harmonic nature:

Fk t   Re Fk 0 e it  (2.1)
We denote by xk the displacement of the system at the location where the input force is applied,
taken along the direction of the input force. Based on the modal superposition principle (cf. Lesson
5) we may write1:

1
Note that from here on we will use symbol i to denote both the imaginary unit (i=sqrt(-1)) and, as a subscript, to
denote the generic i-th mode of vibration. Based on the position of this symbol, it should be straightforward to identify
the correct meaning to this symbol in the following equations.

1/12
Dinamica dei Sistemi Meccanici– Prof. Stefano Bruni

n
x k    ik q i (2.2)
i 1

with n the number of modal components considered, qi the i-th modal coordinate,  ik the value of
the i-th mode of vibration at the position of the applied input force and along the direction of the
force.
The virtual work of the force Fk is:
n
L  x k Fk t   Fk t   ik qi (2.3)
i 1

and hence, the Lagrangian component of the force Fk along the i-th modal coordinate qi is:
Qi t    ik Fk t  (2.4)
We introduce here the additional assumption that by writing the equations of motion of the
considered system according to the modal superposition principle, we obtain decoupled equations
not only as far as the inertial and elastic terms are considered (which is always true, cf. Diana e
Cheli, Meirovitch, Lesson 5 for distributed parameter systems), but also for viscous damping terms.
Based on this additional assumption, the system’s equations of motion written according to the
modal superposition principle consist of n decoupled single-d.o.f. equations as follows:
m11 q1  c11 q1  k 11 q1  1k Fk t 

m ii qi  c ii q i  k ii q i   ik Fk t  (2.5)

m nn qn  c nn q n  k nn q n   nk Fk t 

where m ii , c ii , k ii are respectively the mass, damping and stiffness modal parameters for the i-th
mode of vibration.
The complex valued solution in complex form of the single-d.o.f. equation of motion for the generic
i-th modal coordinate reads:
q i t   q i 0 e it (2.6)
with:
 ik
qi 0  Fk 0 (2.7)
  2 m ii  i c ii  k ii
We now consider the steady-state response of the system evaluated at a given location and along a
given component of displacement, denoted by xj(t):
n
x j t     ij q i t  (2.8)
i 1

Introducing (2.6) and (2.7) in (2.8) we get:

x j t   x j 0 e it (2.9)
with:

2/12
Dinamica dei Sistemi Meccanici– Prof. Stefano Bruni

n  ij  ik
x j0   Fk 0 (2.10)
i 1   m ii  i c ii  k ii
2

and therefore the complex-valued FRF of the system evaluated at location j for a single unit input
force applied at location k is:
n  ij  ik
G jk i    (2.11)
i 1   m ii  i c ii  k ii
2

We note that (2.11) contains all the modal parameters (modal masses, damping, stiffness and modal
shapes) for the first n modes of vibration of the system. Hence, we recognise the possibility to
identify these modal parameters by fitting a set of experimental FRF values obtained from
measurements using expression (2.11) or a simplified version of this expression, under the
assumption that the set of modal parameters providing the best fit of the measured FRFs are the
“actual” modal parameters of the system.

3 Extraction of the experimental FRFs from measurements

The experimental FRFs of the system can be derived from physical tests performed using a SIMO
approach. To this aim, we apply a single excitation force Fk(t) on the system and we derive the
steady-state response of the system at multiple locations xj(t) j=1, …, nm, with nm the number of
measured outputs for the system.
In this section we shortly recall how the FRFs Gjk(i) can be extracted from nm+1 measurements,
i.e. the measure of the force applied at k (measured using a load cell) and the system’s vibration
measured at locations j, j=1, …, nm. Note that for practical reasons, the response of the system is
often measured using accelerometers instead of linear displacement transducers, in which case the
response of the system in terms of displacements can be derived by double integrartion in time of
the measured acceleration signals. This integration needs to be performed with particular care, to
avoid drift effects produced by spurious very low frequency components in the measured signals,
but this topic is not treated here and the reader is referred to classical books on mechanical
measurements for more information in this regard.
We consider in this section two different cases of excitation: harmonic and impulsive excitation.
Note that other types of excitations can be used in an experiment aiming at the extraction of the
system’s FRFs, but due to time limitations only the two cases named above are treated in this
course. For information about other possible excitation cases, the reader is referred to (Diana e
Cheli, Chapter 8) and to the references cited in that chapter.

3.1 Harmonic excitation

We consider here that the system is excited by a single harmonic force at location k. Without
loosing generality, we choose time t=0 corresponding to a positive maximum of the exciting force,
hence the exciting force can be written in complex for as:
Fk t   Fk 0 e it (3.1)
with Fk0 the real-valued constant parameter representing the amplitude of the force (cf. Figure 3.1),
and  the circular frequency of the force, which is related to the period T (also shown in Figure 3.1)
by:
2
  (3.2)
T

3/12
Dinamica dei Sistemi Meccanici– Prof. Stefano Bruni

Assuming the system to behave linearly, the steady-state response of the system2 evaluated at
location j is also harmonic, with the same circular frequency  as the input force, and takes the
form:

x j t   x j 0 e it  x j 0 e

i t  x j 0  (3.3)

with xj0 a complex-valued constant quantity defined in polar notation by modulus x j 0 and phase
x j 0 . As shown in Figure 3.1, the modulus x j 0 represents the amplitude of the harmonic
response, and the phase is related to the delay t of the first positive maximum of the response,
evaluated with respect to the positive maximum of the input force occurring at t=0. From the time
delay t the phase delay of the response can be evaluated as:
2t
x j 0    t (3.4)
T

Figure 3.1: Input and output time histories for harmonic excitation of the system.

The experimental FRF is then derived as the ratio of xj0 over Fk0:
x j0 i x j 0
jk i  
G exp e (3.5)
Fk 0

By changing in discrete steps the circular frequency  of the applied force, the experimental FRF of
the system can be derived for input applied at k and output evaluated at j with j=1, …, nm. To
simplify this time consuming procedure, sometimes harmonic excitation is replaced by the so-called
“swept-sine” excitation, which means the circular frequency of the exciting force is slowly changed
with time to cover a pre-defined range of frequencies in one single test. More details on this type of
excitation are found in (Diana e Cheli).

3.2 Impulsive excitation

Impulsive excitation is a practical approximation of the theoretical Dirac’s impulse. Since the
Fourier transform of a theoretical impulse has constant modulus for all frequencies, a ‘practical’
impulsive excitation is expected to apply more or less evenly excitation over some range of
2
i.e. the response evaluated after the transient generated by the application of the input force has decayed to negligible
amplitude.

4/12
Dinamica dei Sistemi Meccanici– Prof. Stefano Bruni

frequencies, allowing to derive the experimental FRF in this whole frequency range by means of a
single test.
The usual way to produce impulsive excitation is to use an excitation hammer. This consists of a
balanced hammer with a load cell fitted on its tip, which is used to measure the force generated by
the hammer. Generally, the
stiffness of the hammer tip impacting the system can be adjusted by choosing the tip used in the test
among a set of tips made in different shapes and materials. This allows to control the frequency
range in which excitation is applied, as mentioned below.
Figure 3.2 shows the time histories of the exciting force Fk(t) applied at location k and of the
system’s response evaluated at location j. The exciting force is applied for a short time duration,
which actually depends on the stiffness of the tip of the hammer (a stiffer tip producing a shorter
impact duration), and the waveform of the applied force is well approximated by a half-sine
function padded with zero values. The response of the system contains different non-stationary
harmonic contributions and, after the hammer looses contact with the system (i.e. no excitation is
applied), corresponds to the free damped motion of the system, which in turn (for low damped
systems) consists of the combination of a number of decaying oscillations.

Figure 3.2: Input and output time histories for impulsive excitation of the system.

In order to extract the experimental FRF out of the measured time histories, first of all a Fourier
transform needs to be applied. Because of the complex nature of the signals, this needs to be done
numerically, i.e.:
i) the signals are acquired digitally, using analog to digital (A/D) conversion with a chosen
sampling frequency fs and duration of acquisition T. In this way, each signal (force at k,
vibration at j, j=1, …, nm) is transformed into a discrete set of ns sampled values with
n s  Tf s (3.6)
ii) the digitalised signals are processed according to the “Discrete Fourier Transform”
(DFT) algorithm, which provides a discrete spectrum for each signal.
The qualitative appearance of the discrete spectra for the force applied at k and the vibration
measured at j is shown in Figure 3.3. For each measured quantity, the spectrum consists of a couple
of Bode diagrams, one for the modulus and one for the phase. It is observed that:
a.) the spectra are defined at discrete frequency values which are the integer multiples of the
“fundamental frequency” f0:
1
f0  (3.7)
T

5/12
Dinamica dei Sistemi Meccanici– Prof. Stefano Bruni

note that the duration of the acquisition T is not a characteristic feature of the excitation
applied (as was the case for the period T of the harmonic excitation treated in 3.1), and is
instead a parameter which can be freely defined while performing the data acquisition
process.
b.) The maximum frequency at which the discrete spectra are defined is defined by the
Shannon’s theorem and is one half of the sampling frequency.
c.) The spectrum of the impulsive force shows almost constant amplitude up to some maximum
frequency which depends on the duration of the ‘real’ impulse and thus, ultimately, on the
stiffness of the tip used. A stiffer tip will extend the pass band of the excitation toward
higher frequencies.
d.) The amplitude of the harmonic components of the system’s response is otherwise changing
with frequency, with larger amplitude values occurring at frequencies close to resonances of
the system.

Figure 3.3: Discrete spectra of the input force and output vibration for impulsive excitation.

Like in the case of harmonic excitation, the experimental FRF can be defined by taking the complex
ratio of the harmonic component of the response xj,n over the corresponding harmonic component of
the input force Fk,n:
x j ,  
jk i 2f 0  
i x j ,  Fk ,
G exp e (3.8)
Fk ,

with  the integer multiple of the fundamental frequency f0. Equation (3.8) provides the values of
the FRF at some discrete, equi-spaced frequency values, which are integer multiples of the
fundamental frequency. The maximum value of index n that can be introduced in (3.8) is bounded
by Shannon’s theorem, i.e.  f0 < ½ fs must be satisfied. Furthermore, the definition of the
experimental FRF according to (3.8) shall not be used if the amplitude of the input force harmonic
component is too low (note this might happen at frequencies much lower than the Shannon’s limit,
depending on the hammer tip used and on the sampling frequency), because that would imply a too
low signal-to-noise ratio, which would lead to large errors in the definition of the experimental
FRF.

6/12
Dinamica dei Sistemi Meccanici– Prof. Stefano Bruni

4 Identification of modal parameters

In order to perform the identification of the system’s modal parameters, we look for a set of modal
parameters that introduced in (2.11) produce the best fit of this expression to the measured FRF
jk i  over a given frequency range.
G exp
This procedure is not easy to carry out if the whole frequency range of interest is considered at
once. Therefore, we observe that, based on (2.11), the FRF of the system is defined as the (complex
valued) linear combination of n single-d.o.f. FRFs, one for each modal component considered.
If the system satisfies the following two conditions:
i) the natural frequencies of the system are well separated;
ii) the system is low damped
then in a relatively narrow frequency range centred on each resonance frequency of the system, the
single term in (2.11) corresponding to the resonating modal component, is prevailing over the other
terms, as shown schematically in Figure 4.1

Figure 4.1: Qualitative graphic representation of the FRF modulus. In solid line: modulus of the
overall FRF, in dashed line: modulus of the contribution from each single modal component.

This observation shows that the modal parameter identification process can be performed for one
modal component at a time, restricting the analysis to the frequency range in which the considered
modal component is dominant in defining the system’s FRF, taking advantage from the possibility
to simplify the mathematical expression of the FRFs in this frequency range.
In the two following sub-sections, two different levels of simplification are introduced. In Section
4.1 a simplified expression still considering (at least in first approximation) the contribution of the
non-resonating terms is introduced, whereas in 4.2 all terms in (2.11) are sharply neglected except
the one excited in resonance. Clearly, the approach presented in 4.1 is likely to provide a more
accurate identification of the system’s modal parameters, as it involves a less severe simplification
of the analytical FRF expression, however, the simplified FRF expression still requires to be treated
numerically, using a residual minimisation technique, to perform the modal parameter
identification, whereas the simpler approximated FRF expression introduced in 4.2 allows for an
analytical treatment.

4.1 Modal parameter identification using residual minimisation

As anticipated above, we select a frequency range containing one single resonance frequency of the
system, and we aim at identifying the parameters of the resonating modal component.
Thus, we approximate (2.11) as follows:

7/12
Dinamica dei Sistemi Meccanici– Prof. Stefano Bruni

A j  iB j E j  iF j
G jk i    C j  iD j  (4.1)


2
m i
ii c 
ii k
ii  


2
residual of
resonating term modes resonating residual of
at higher frequencie s modes resonating
at lower frequencie s

The first term of this expression represents (with no approximation) the contribution to the FRF
from the i-th modal component, which is here assumed to be the resonating one. The second term is
a residual accounting in first approximation for the overall effect of all modal components
resonating at frequencies lower than the selected frequency range, and the third term is a residual
accounting for the effect of all modal components resonating at frequencies above the frequency
range considered.
Each term is defined by a complex-valued “gain” parameter, namely A j  iB j for the first term,
C j  iD j for the second term and E j  iF j for the third term. The reason for considering complex-
valued gains is that in case the damping in the system is of such nature that the damping matrix is
not diagonalised by the modal matrix defined using the real-valued modes of vibration of the
undamped system, then it can be proved that a complex-valued modal superposition can be defined
using the complex-valued modes of vibration of the damped system, which will lead to a diagonal
modal damping matrix. Using this complex-valued modal superposition however, the ‘gain’
parameters ij ik in (2.11) will be complex valued instead of real-valued.
To perform the identification of modal parameters for the resonating i-th mode, we prescribe:
A j  iB j E j  iF j
 C j  iD j  jk i 
 G exp (4.2)
  m ii  i c ii  k ii
2
 2

for:
  n 0  n 2f 0 n  1,..., n f
(4.3)
j  1,..., n m
with nf the number of frequency values at which the experimental FRF is available within the
frequency range considered for parameter identification.
Using (4.2) and considering (4.3) we can write a system of nf x nm equations in 3+6 x nm unknowns,
the unknowns being:
m ii , c ii , k ii
(4.4)
A j , B j ,C j , D j , E j , F j j  1,..., n m

Provided that nf is sufficiently large, this results in an over-determined system of equations, which
can be solved using a least square approach, i.e. defining for each equation (4.2) a (complex-valued)
residual as the difference of the right hand side minus the left hand side, and looking for the set of
modal parameters (4.4) that corresponds to a minimum of an overall residual function defined as the
squares of the residuals summed over each equation.
Because the residual function depends non-linearly from the unknown modal parameters, this
problem can only be solved using an iterative minimisation procedure. No detail is provided here
about the numerical methods that can be used for this purpose, and about possible problems that can
be experienced with the convergence of this numerical procedure. For these topics, the reader is
referred to any textbook on numerical analysis.
Once solved this numerical problem, the modal mass, damping and stiffness parameters are directly
obtained from the solution, cf. (4.4) whereas, based on comparison with (2.11), the real part A j of
the ‘gain’ of the resonating term represents the modal shape  ij of the resonating i-th mode

8/12
Dinamica dei Sistemi Meccanici– Prof. Stefano Bruni

evaluated at location j, multiplied by a common scaling factor  ik that can be removed by taking an
appropriate normalisation of the mode. Of course, in order to rely on this result it is required to
check that:
i) the imaginary part B j of the ‘gain’ of the resonating term is small compared to A j , so
that the actual feasibility of using the real-valued modal matrix of the un-damped system
to perform the modal superposition is verified;
ii) the real and imaginary parts of the ‘gains’ of the two residual, i.e. terms C j , D j , E j , F j
are all small compared to A j .
if one (or both) of the checks above is not satified, this means that some of the assumptions
underlying the modal parameter identification process are not satisfied, and hence the result of the
identification shall be rejected or at least checked with care.
The process outlined above can be repeated for different frequency ranges, thereby leading to the
identification of the modal parameters for all modal components of interest.

4.2 Modal parameter identification using analyticalformulae

A more radical approximation of (2.11) consists of simply neglecting all terms except the
resonating one (denoted below as the i-th term):
 ij  ik
G jk i   (4.5)
  2 m ii  i c ii  k ii
In this way, the FRF of the system is locally approximated as a single-d.o.f. FRF. This allows to
identify the following modal parameters of the system:
- un- damped natural frequency;
- non-dimensional damping ratio;
- mode of vibration.
using simple formulae, as show below.

Identification of the un- damped natural frequency

We know that for a low damped 1-d.o.f. system the absolute value of the FRF is maximum at a
frequency falling very close to the un-damped natural frequency of the system, denoted here as i.
Hence, we identify here I as the frequency at which G jk i  is maximum within the considered
frequency range.

Identification of the non-dimensional damping ratio – “phase derivative”

method
For a 1-d.o.f. system, the phase of the FRF is approximately 0 at frequencies well below the
resonance, - at frequencies well above the resonance and shows a transition from 0 to - around
the resonance frequency; the lower the non-dimensional damping, the steepest the transition. Hence,
measuring the “steepness” of the transition of the phase angle is a way to estimate the non-
dimensional damping of the system. We take as a measure of this steepness the derivative of the
phase angle with respect to the circular frequency  evaluated at the resonant frequency  = i
and we show below that this derivative is inversely proportional to the non-dimensional damping hi
of the resonating i-th modal component:

9/12
Dinamica dei Sistemi Meccanici– Prof. Stefano Bruni

c ii
hi  (4.6)
c cri
To this aim, we observe that:

  c ii   2 a i hi 
G jk i    artg     artg 

   artg b  (4.7)
1 a 2  i
 k ii   m ii
2
  i 
with:
2a i hi 
bi   ; ai  (4.8)
1  ai
2 i

Using the chain rule for derivation, we get:

 G jk  G jk a i  G jk 1
 
  i
a i ai 1
 a i ai 1
i

furthermore, also using the chain rule for derivation, we get:

 G jk  G jk bi

a i bi a i
and, recalling that:

artg ( bi )

1

1  a  i
2 2

bi 1  bi 2 1  a   4a i
2 2
i
2
hi 2
we get:

 G jk

1  a  2 2 
 2 hi  2 a i hi

 2ai   
  
2 hi 1  a i 2  4 a i 2 hi
  
i
a i 1  a   4a i
2 2
i
2
hi 2
 1  ai 2
 1  ai 2 
2 
  2

1  a i 2  4 a i 2 hi 2

and, evaluating this derivative at ai=1:

 G jk 1

a i ai 1
hi

so that, finally:
 G jk  G jk 1 1
  (4.8)
  i
a i ai 1
i hi  i

This expression allows to estimate the damping factor hi of the considered resonating modal
component from the derivative of the FRF phase (for a single output j of choice). Clearly, repeating
the estimation for different FRFs (i.e. for different outputs j) and averaging the result is likely to
reduce the results entailed by this simple estimation procedure.

Identification of the non-dimensional damping ratio – “half-power points”

method
A second method which can be defined to estimate the non-dimensional damping factor hi starting
from (4.5) is the so-called “half power points” method. To introduce this method, we consider the

10/12
Dinamica dei Sistemi Meccanici– Prof. Stefano Bruni

plot of the modulus of the FRF vs. frequency for a single output j of choice, as shown in figure 4.2.
On this plot, we define the so-called “quality factor” Q as the maximum amplitude of the FRF.
Next, we identify the two frequencies 1 and 2 at which the modulus of the FRF drops down to
Q 22 . These points are called the “half-power points”, because often in engineering problems the
power of a given quantity is related with the squared value, so that if the amplitude drops down
from the maximum value by a factor 22 , then the “power” (in some sense) is decreased by a
factor ½.
Assuming the non-dimensional damping hi << 1, it is possible to show by trivial mathematics (not
reported here for the sake of brevity, the interested reader is referred to either Diana and Cheli or
Meirovitch for a proof) that:

 2 2  12
hi  (4.9)
4 i 2
providing another simple formula that can be used, besides (4.8), for identifying the non-
dimensional damping factor of the resonating modal component in the frequency range considered.

Figure 4.2: Half-power points for a 1-d.o.f. frequency response function.

Identification of the mode of vibration

Finally, we provide a simple formula that can be used to identify the shape of the mode of vibration
evaluated at the measuring locations j, j=1, …, nm.
To this aim, we evaluate (4.5) for =i.
Because:
k ii
i  ; k ii   i m ii
2
m ii
we get:
 ij  ik  ik
G jk i    i  ij (4.10)
 i i c ii  c ii
this equation shows that the shape of the i-th resonating mode of vibration evaluated at location j is
equal to the imaginary part of the Gjk frequency response function evaluated at the resonance

frequency i, multiplied by a common scaling factor  ik . This scaling factor can be removed
 c ii
by taking an appropriate normalisation of the mode. We note that the resonance frequency i was

11/12
Dinamica dei Sistemi Meccanici– Prof. Stefano Bruni

previously identified, and hence is here considered as a known parameter. Also, we note that based
on (4.10) it would appear that taking the imaginary part of the FRF at the resonance frequency is the
same as taking the absolute value of the same FRF in resonance. We recall however that (4.10) is
derived from (4.5), which in turn comes from (2.11) neglecting all the non-resonant terms. Because
in the frequency range considered the non-resonant terms are excited either in the quasi-static or
seismographic region, their contribution is approximately real valued (in other words, the phase
associated with these non-resonant terms is either close to zero or close to –). Hence, by taking the
imaginary part of the FRF instead of the absolute value, the “disturbance” effect from the non-
resonating modal components can be depurated more effectively.

12/12