Presented at 3RD IMAC Conference, Orlando, FL January, 1985

Global Curve Fitting of Frequency Response Measurements using

the Rational Fraction Polynomial Method
by
Mark H. Richardson and David L. Formenti
Structural Measurement Systems
San Jose, California

ABSTRACT
In summary, then, each mode is a global property of the
The latest generation of FFT Analyzers contain still more structure. Each mode is defined by a natural (or modal)
and better features for excitation, measurement and frequency, a value of (modal) damping, and a mode shape.
recording of frequency response functions (FRF's) from
Since modes are properties of the structure, itself, and are
mechanical structures. As measurement quality continues to
independent of the type of excitation force used to excite it,
improve, a larger variety of curve fitting methods are being
they should be identified from measurements which are also
developed to handle a set of FRF measurements in a global
independent of the type of excitation. The Frequency
fashion. These approaches can potentially yield more
Response Function (FRF) is such a measurement for linear
consistent modal parameter values than curve fitting
systems.
individual measurements independently.
The FRF is essentially a “normalized” measure of structural
In this paper, a new formulation of the Rational Fraction
response. That is, it is the ratio of a response spectrum
Polynomial method is given which can globally curve fit a
divided by the spectrum of the excitation which causes the
set of FRF measurements. The pros and cons of this
response. Hence, the FRF is a measure of the dynamic
approach are discussed, and an example is included to
properties between two degrees-of-freedom (DOF's) of a
compare the results of this method with a local curve fitting
structure; the excitation point (and direction) and the
method.
response point (and direction). Again, the modes of the
INTRODUCTION structure are indicated by the peaks in the measurement,
with at least one mode defined by each peak.
Physically speaking, a mode of vibration of a structure is
characterized by a so called "natural" or "resonant' Figure 1 shows in simplified form how the first three modes
frequency at which the structure's predominant motion is a of a beam are identified from a set of FRF measurements
well defined waveform, called the "mode shape". A mode is made from the beam. The figure shows the imaginary part
the manifestation of energy which is trapped within the of each of the FRF measurements which were made
boundaries of the structure, and cannot readily escape. between some (arbitrary) excitation point, and each of the
When a structure is excited, its linear response can be shown
to be a function of the combined motions of its modes of
vibration. That is, the overall motion can be represented as
a linear combination of the motions of each of the modes.
Likewise, when the excitation source is removed from the
structure, the trapped energy within it will slowly decay out
until it no longer vibrates. The rate at which energy decays
out of the structure is controlled by the amount of damping
in the structure. Damping is also a modal property, each
mode having a certain value of damping associated with it.
That is, the motion comprised of heavily damped modes
will decay out more quickly than that part of the motion
comprised of more lightly damped modes.
Modes of vibration can be observed in practically any
vibrating structure. When we measure the vibration of a
structure and decompose the vibration signal into its
frequency spectrum, the modes of vibration are evidenced
by peaks in the spectrum. (Other peaks may be present in
the spectrum due to large cyclical excitation forces). The
modal peaks, however, will appear in practically any
measurement made from any point on the structure.

Page 1 of 8
Presented at 3RD IMAC Conference, Orlando, FL January, 1985

response points marked with X's. In this case, responses

CURVE FITTING FRF's
were measured only in the vertical direction and with a
transducer that measured either displacement or Curve fitting, or Parameter Estimation, is a numerical
acceleration. Alternatively, the FRF's could have been process that is typically used to represent a set of
measured by mounting a single response transducer in one experimentally measured data points by some assumed
(arbitrary) location and exciting the beam at each X, in the analytical function. The results of this curve fitting process
vertical direction. are the coefficients, or parameters, that are used in defining
the analytical function. With regard to the Frequency
The figure shows a modal peak at the same frequency in
Response Function, the parameters that are calculated are its
each measurement, indicating the global nature of modal
so-called modal parameters (i.e. modal frequency, damping,
frequency. The “width” of the modal peak for each mode
and residue). The curve fitting process can also be thought
should also be the same in each measurement, again
of as a data compression process since a large number of
indicating the global nature of modal damping. Lastly, the
experimental values (the FRF measurements) can be
mode shape which is defined by assembling the modal peak
represented by a much smaller number of modal parameters.
values from all the measurements, is global in the sense that
it is defined for the entire expanse of the structure. Various forms of the transfer function dynamic model are
used to curve fit FRF measurements. The transfer function
Mathematically speaking, modes of vibration are defined by
model is, in effect, evaluated along the frequency axis (i.e.
certain parameters of a linear dynamic model for a structure.
The dynamic properties of a structure can be written either s=jω) during the curve fitting process. The entire transfer
as a set of differential equations in the time domain, or as a function model is shown in Figure 3 , and it is well known
set of equations containing transfer functions in the Laplace [2] from examination of this model that curve fitting of one
(frequency) domain. row or one column of FRF's is sufficient to identify the
modal properties of the structure. However, selection of the
These equivalent models are shown in Figures 2 and 3. correct row or column may be very important, depending on
Regardless of which model is used, it can be shown that the modes of interest, the geometry of the structure, etc.
either model can be written in terms of the same parameters
(frequencies, damping, and mode shapes) that describe the
modes of vibration.

Page 2 of 8
Presented at 3RD IMAC Conference, Orlando, FL January, 1985

Nevertheless, once a set of FRF measurements has been formulation) but it offers a unique way of handling the
made on a structure, whether they comprise one row or residual effects of out-of-band modes.
column, or several, the most commonly used method of
Most other MDOF methods require that additional
curve fitting FRF's is to fit them one at a time using one of
"computational" modes be used in order to compensate for
the analytical forms of the FRF shown in Figure 4.
the residual effects of out-of-band modes in the curve fitting
The most commonly used form is the Partial Fraction Form. frequency band. With some methods, these computational
Most SDOF (or single mode) methods, and various iterative modes can often cause the parameter estimates of the modes
MOOF (multiple mode) methods, are based on this form of of interest to be in large error if the “right" number of
the model. computational modes is not used. Choosing the “right"
number of computational modes can be a trial and error
RATIONAL FRACTION POLYNOMIAL (RFP)
process.
METHOD
The RFP method, on the other hand, allows the use of
In a previous paper [1], a curve fitting method based on the
additional numerator polynomial terms as a means of
Rational Fraction Polynomial form of the FRF was
compensating for the effects of out-of-band modes. As
introduced. This MDOF method fits the analytical
shown in [1], the use of these extra terms still permits the
expression (1) to an FRF measurement in a least-squared
accurate estimation of the modal parameters of interest, and
error sense, and in the process, the coefficients of the
is, in general, a more foolproof means of compensation than
numerator and denominator polynomials are identified.
the use of computational modes.
Analytical Forms GLOBAL CURVE FITTING
of the Most curve fitting is done today in a “local" sense. That is,
Frequency Response Function each measurement is individually curve fit, and the modal
frequency, damping, and complex residue are estimated for
Rational Fraction Form each mode in the measurement. Hence, four parameters are
m estimated for each mode, (counting the complex residue as

∑ a k sk two parameters), and if an MDOF curve fitter is used to

estimate the parameters of, for example, five modes, then a
H( ω ) = k=0
n
(1) total of twenty unknown parameters must be simultaneously
identified during the curve fitting process. With such a large
∑ bk s k
number of unknowns, significant errors can occur in the
parameter estimates. Many times the accuracy of the modal
k=0 s = jω
parameters is sacrificed during the curve fitting process to
yield a good looking (re-synthesized) curve fit function. In
Partial Fraction Form other words, the values of the estimated parameters can
n trade off errors. This is especially true for the modal
 rk
2
rk 
∗ damping and residue estimates.
H( ω ) = ∑  + ∗
(2)
k =1  s − p k s − pk 
Accurate damping and residue estimates are, in general,
more difficult to obtain than accurate frequency estimates.
s = jω Damping is the most difficult parameter to estimate
accurately from FRF measurements, and the residue is often
p k = − σ k + jω k = kth pole tightly coupled to damping. That is, if damping is in large
error, the residue estimate will be in large error even though
rk = residue for kth pole the curve fitting function closely matches the measurement
data.
FIGURE 4
One approach that can reduce errors is to divide the curve
Once these coefficients are known, it is a straightforward fitting process up into two steps; (1) estimate the frequency
matter to obtain the poles, the zeroes, and the modal and damping parameters and (2) the residue (or mode shape)
properties (poles and residues) of the FRF. This method has parameters. This process of using the measurement data to
been implemented in a variety of commercially available obtain frequency and damping estimates first, and then with
modal software packages and has been used successfully on known frequency and damping values to obtain mode shape
a large variety of FRF measurements. estimates by a second estimation process is called Global
As pointed out in [1], not only can the RFP method handle Curve Fitting.
noisy measurements (because of its least-squared error

Page 3 of 8
Presented at 3RD IMAC Conference, Orlando, FL January, 1985

One advantage of Global Curve Fitting is that more accurate (also called the characteristic polynomial) is therefore
frequency and damping estimates can potentially be known. Hence, the only unknowns are the coefficients of
obtained by processing all of the measurements, than can be the numerator polynomial. Once these coefficients are
obtained from curve fitting a single measurement. Another determined, then the residues, and hence the mode shapes,
advantage is that because damping is already known and can be computed.
fixed as a result of the first step, the residues are, in general,
more accurately estimated during the second step. Global FRP Using Ordinary Polynomials
GLOBAL FREQUENCY AND DAMPING
FRF in Terms of polynomial coefficients, ak
Reference [1] contains a special formulation of the RFP
m
hi = ∑ t i, k a k
method which obtains global frequency and damping
estimates from a set of FRF measurements. The Complex i=1,…,L (3)
Exponential (or Prony) method [4] can also be formulated in k=0
a similar manner to obtain global estimates from a set of
impulse response functions. (These functions would be where
obtained by taking the inverse FFT of each FRF
( jω i ) k
measurement). t i .k = n
Another very straightforward method of obtaining global
frequency and damping is to average together the ∑ b k ( jω i ) k
magnitudes of all of the FRF measurements. The resulting k =0
magnitude function will then contain the resonance peaks of ( jω i ) k
all of the modes in the measurement set, and this function = mod es
(4)

∑
can be curve fit to obtain frequency and damping estimates.
( ω k − ω i + j2σ k ω i )
2 2
Probably the simplest method, though, is to select some
measurements where each particular mode has a large k =1
response and then use the frequency and damping estimates
from these measurements as the best approximations of the In matrix form
global estimates.
 h1   t1, 0 t1, m   a 0 
h    
GLOBAL MODE SHAPE
 2   a1 
 =
Once the modal frequencies and damping are known, the

      
mode shapes (which are eigenvectors) can be obtained by
solving an eigenvalue/eigenvector set of equations, (3).
h n   t L , 0 
This type of an approach generally involves the t L , m  a m 
manipulation of large matrices, however, and hence requires
a relatively large computer; one larger than is found in most
laboratory testing systems.
or
The global curve fitting method introduced in this paper is { H} = [T]{ A } (5)
much simpler and easier to implement on a small computer
than an eigenvalue/eigenvector solution approach. The Least Squared Error Equations
approach discussed here is “global” in the sense that global
frequency and damping estimates are used, but is "local" in
the sense that the FRF measurements are processed one at a
[ T∗ ]t [ T]{ A } = Re([ T∗ ]t { Y}) (6)
time in order to obtain modal residue estimates. These
residue estimates are then assembled from all the various
{ Y} = L-vector of measurement data.
measurements to obtain the mode shapes. FIGURE 5
THE GLOBAL RFP METHOD TO ESTIMATE MODE
SHAPES
The RFP method can be reformulated to take advantage of
the global nature of modal frequency and damping. If the
Rational Fraction Form of the FRF shown in Figure 4 is
used for curve fitting, and the modal frequencies and
damping are already known, the denominator polynomial

Page 4 of 8
Presented at 3RD IMAC Conference, Orlando, FL January, 1985

Global FRP Using Orthogonal Polynomials These simultaneous linear equations can be solved using a
standard equation solver, but our experience has been that
numerical problems occur frequently in attempting to solve
FRF in terms of polynomial coefficients, ck them in this form. Again, as with our previous
m implementation of the RFP method, the left-hand side
h i = ∑ zi , k c k i=1,…,L (7) matrix in equation (6) becomes ill-conditioned due to the
polynomial functions, and hence the equation solutions can
k =0 contain significant errors.
where Figure 6, on the other hand, contains a reformulation of the
φ i,k RFP method using orthogonal polynomials. Not only does
zi , k = mod es
this formulation eliminate the numerical problems of

∑
equation (6), but the solution equations (11) become greatly
( ω k − ω i + j2σ k ω i )
2 2
simplified and don't require a simultaneous linear equation
k =1 solution at all. Notice from equation (9) that the numerator
φ i, k orthogonal polynomials are generated using the
= (8) denominator terms (gi, i = 1,...L) as a weighting function.
gi Once the matrix [Z] is computed, it can be saved and used to
obtain a new set of polynomial coefficients {C} for each
L
φ i, k φ i, j .5 k = j new vector of measurement data {Y}. Of course, the
∑ | g i |2
=
0 k ≠ j
(9) coefficients {C} are for the orthogonal polynomials, and the
procedure described in [1] must be used to recover the
i =1
coefficients {A} of the ordinary polynomial coefficients.
The residues are then recovered by a partial fraction
In matrix form expansion process. These final calculations are very
straightforward, though.
 h1   z1, 0  z1, m   c 0 
h     TEST CASE WITH HEAVY MODAL COUPLING
 2    c1 
  =  
The Global RFP method was compared with the Local RFP

         method by curve fitting five measurements with three

h n   zL , 0  zL , m  c m 
heavily coupled modes. Plots of the log magnitudes of
these five measurements are shown in Figure 8 . These
FRF's were synthesized from known modal parameters
(listed in Figure 7), and then random noise was added to
or
them to simulate more realistic measurements. Synthesized
{ H} = [Z ]{C} (10) measurements were used so that the curve fitting results
could be compared with the correct modal parameter values.
Least Squared Error Equations The measurements were synthesized for a frequency range
of 0 Hz to 100 Hz, and the curve fitting was done between
{ C} = Re([ Z∗ ]t { Y}) (11) 40 Hz and 65 Hz.
{ Y} = L-vector of measurement data. Notice that measurement No.1 contains a zero (0) residue
for mode No.2, indicating that this measurement was taken
FIGURE 6 at a node point of Mode No. 2. Likewise, measurement
In equation (3) in Figure 5, the FRF (hi) is written for L data No.3 was taken at a node point of mode No. 1.
Measurement No.4 (or No. 5) could be a driving point
points (i = 1 .... L). This expression of the FRF model is
measurement since all of the residues have the same sign.
written in a form which isolates the unknown parameters
(ak, k=0,...,m). The least squared error curve fitting
problem can then be formulated in a manner similar to [1],
and the solution equations (6) shown in Figure 5 will result.
Notice that only the right hand side of equation (6) is
directly dependent on the FRF measurement data. The
coefficient matrix on the left-hand side only depends on the
frequency range and the number of data points used for
curve fitting.

Page 5 of 8
Presented at 3RD IMAC Conference, Orlando, FL January, 1985

MODAL DATA USED TO

SYNTHESIZE MEASUREMENTS
MODE NO.: 1 2 3
FREQUENCY (Hz): 50 52 55
DAMPING (%): 3 2.5 3
----- RESIDUES -----
MEASMT. NO.1: 1 0 -1
MEASMT. NO.2: .5 -.25 -.6
MEASMT. NO.3: 0 -.5 .6
MEASMT. NO.4: -.5 -.75 -.3
MEASMT. NO.5: -1 -1 -.8
FIGURE 7

FIGURE 8.b

FIGURE 8.a

FIGURE 8.c

Page 6 of 8
Presented at 3RD IMAC Conference, Orlando, FL January, 1985

curve fitting results of measurement No.3. Measurement

Nos. 2, 4, and 5 all indicate the same curve fitting problem;
too many unknown parameters (12 per measurement) are
being simultaneously estimated on measurements with
heavy modal coupling and some small amount of noise.
Figure 10 contains the results of using the Global RFP
method on the five measurements in Figure 8. The global
frequency and damping estimates could be obtained by any
of the previously described methods. In this case, the
known values were merely used.
The residue estimates were obtained by applying equations
(11) in Figure 6 to each of the measurements. Comparing
these estimates with the correct answers in Figure 7, one can
conclude that all of the residues were accurately estimated.
All of the estimates are in error by less than 1%.

*** Modal Fit Data ***

MEAS: 1
FIGURE 8.d
MODE FREQ(Hz) DAMP(%) AMPL PHS
1 49.99 3.03 1.998E+00 .33
2 55.01 3.00 1.918E+00 180.20
3 55.26 .95 7.495E-03 43.96

MEAS: 2
MODE FREQ(Hz) DAMP(%) AMPL PHS
1 50.09 2.38 4.197E-01 339.91
2 54.72 4.26 7.120E-01 146.28
3 55.60 .73 4.585E-02 185.47

MEAS: 3
MODE FREQ(Hz) DAMP(%) AMPL PHS
1 51.99 2.49 4.914E-01 100.73
2 55.00 2.92 5.873E-01 .73
3 57.77 .95 7.943E-04 7.43

MEAS: 4
FIGURE 8.e MODE FREQ(Hz) DAMP(%) AMPL PHS

Figure 9 contains the results of curve fitting each of the 1 49.92 1.36 1.174E-01 207.24
2 51.83 4.04 1.468E+00 178.87
measurements using the Local RFP method. The curve
3 56.14 1.07 8.045E-02 167.07
fitter was set up to fit 3 modes in all measurements since
measurement No.'s. 4 and 5 at least show the evidence of 3
MEAS: 5
modes (3 modal peaks).
MODE FREQ(Hz) DAMP(%) AMPL PHS
The curve fitting results for measurement No.1 indicate the
first problem encountered with fitting these measurements; 1 50.12 2.61 9.644E-01 109.27
even though the modal parameters of the two modes in this 2 52.16 2.99 1.448E+00 190.45
measurement were estimated quite accurately, the curve 3 55.66 2.28 4.929E-01 175.28
fitter assigned these parameters to Mode Nos. 1 and 2 when
they should have been assigned to Mode Nos. 1 and 3. This FIGURE 9 Local RFP Results
difficulty is always encountered when using a local fitting
method on measurements where some of the modes are at, CONCLUSIONS
or near, node points. The same problem occurred with the

Page 7 of 8
Presented at 3RD IMAC Conference, Orlando, FL January, 1985

In this paper we showed that the RPF method for curve does however suffer from a few disadvantages. If modal
fitting FRF measurements can be reformulated so that it can frequency or damping varies by any substantial amount
be used in a global curve fitting scheme. That is, global from one measurement to another, then the global method,
frequency and damping can be estimated as a first step, and which assumes that these parameters are unchanging, will
these estimates can then be used to estimate residues in a be in error. However, such variations in modal frequency or
second step. This global curve fitting approach was then damping are not characteristic of a linear system, and if this
shown by example to give more accurate parameter occurs, more care should be taken to obtain a consistent set
estimates than simultaneously estimating modal frequency, of measurements.
damping, and residue for each mode in each measurement,
Secondly, with the global method, the accuracy of the
which is most commonly done today.
residue estimates strongly depends on the accuracy of both
The global curve fitting method was compared with the the global frequency and damping estimates. Therefore, the
more conventional local fitting method on measurements overall success of the method depends on whether global
that contained heavy modal coupling and additive random frequency and damping can be accurately estimated first.
noise. In this case, the global method was clearly more
These questions and further questions regarding the
accurate than the local method. In cases of less modal
handling of out-of-band modes will be addressed in future
*** Modal Residues *** papers.

MODE NO. FREQ(Hz) DAMP(%) REFERENCES

1 49.999 2.998 [1] Richardson, M.H. & Formenti, D.L., "Parameter
Estimation from Frequency Response Measurements
MEAS NO. AMPL PHS
using Rational Fraction Polynomials", Proceedings of
1 9.917E-01 359.65 the 1st International Modal Analysis Conference,
2 4.982E-01 359.69 Orlando, Florida, November 8-10, 1982.
3 1.303E-03 186.34
4 5.046E-01 179.15 [2] Richardson, M.H., “Modal Analysis using Digital Test
5 1.001E+00 180.31 Systems", Proceedings of Seminar on Understanding
Digital Control and Analysis in Vibration Test Systems,
The Shock of Vibration Information Center, Naval
MODE NO. FREQ(Hz) DAMP(%) Research Lab, Washington, D.C., May, 1975.
2 52.000 2.502 [3] Crowley, S.M. & Allemang, R.J., “Application of the
MEAS NO. AMPL PHS PolyReference Technique in Experimental Modal
Analysis”, Proceedings of the 2nd International Modal
1 1.315E-02 323.66 Analysis Conference, Orlando, Florida, February 6-9,
2 2.491E-01 179.68
1984.
3 4.998E-01 180.45
4 7.530E-01 180.62 [4] Spitznogle, F.R. & Quazi, A.H., “Representation and
5 9.974E-01 177.35 Analysis of Time Limited Signals Using a Complex
Exponential Algorithm”, J. of Acoustical Society of
America, Vol. 47, No. 5 (Part I), May 1970, pp. 1150-
MODE NO. FREQ(Hz) DAMP(%) 1155.
3 55.000 2.999
MEAS NO. AMPL PHS
1 9.990E-01 179.65
2 6.022E-01 180.03
3 6.022E-01 .12
4 2.978E-01 180.00
5 7.970E-01 180.27

FIGURE 10 Global RFP Results

coupling and/or less noise, the advantage of the global
method over the local method may be less pronounced.
Nevertheless, the global method gives generally better
results than local methods, especially on measurements
where some modes are at node points. The global method

Page 8 of 8