Notes On Using the ABRAVIBE Toolbox for Experimental Modal

Analysis
Anders Brandt
University of Southern Denmark

June 19, 2013

Abstract
This technical note discusses some fundamental aspects of experimental modal analysis (EMA),
with an emphasis on the tools available in the ABRAVIBE toolbox. In addition to the theory
behind the modal parameter extraction techniques used in ABRAVIBE, this note also discusses
some important practical aspects of how to make a good EMA test. The importance of the
quality of measured data cannot be stressed enough, as the success of an EMA test largely lies in
the setup and the measurement of frequency response functions (FRFs). If the obtained FRFs
are of poor quality, the analysis results will be of poor quality.

Contents
1 Introduction 2

2 Experimental Setup 2
2.1 Points and DOFs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 General Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.3 Suspension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4 Pretest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.5 Measurement Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.6 Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.7 Data Acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.8 Checklist . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3 Modal Parameter Extraction Methods 9

3.1 Extraction Methods for Poles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.1.1 Pronys Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.1.2 The Least Squares Complex Exponential Method . . . . . . . . . . . . . . . . 11
3.1.3 Polyreference Time Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2 Methods for Computing and Scaling Mode Shapes . . . . . . . . . . . . . . . . . . . 16
3.2.1 The Least Squares Frequency Domain Method Single Reference Case . . . . 16
3.2.2 The Least Squares Frequency Domain Method Multiple Reference Case . . 17

1
4 Examples 19
4.1 Comparison of the frf2ptime Command Options . . . . . . . . . . . . . . . . . . . . 19
4.2 EstType Prony . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.3 EstType LSCE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.4 EstType PTD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

1 Introduction
Experimental modal analysis (EMA) is an important tool in many vibration analysis situations.
In this technical note we discuss some fundamental aspects of performing a successful test, from
planning, through data acquisition, to final modal parameter extraction. The emphasis in this note
is to provide some theory and information necessary to understand and use the ABRAVIBE toolbox
[1, 2] for this purpose, see more at http://www.abravibe.com. Knowledge of modal analysis theory
is assumed in this note. For a reference see chapter six in [3].
An EMA test is usually comprised of some key tasks:
1. The structure is suspended free-free (or sometimes measured installed in its operating envi-
ronment).
2. Measurement locations (nodes) are determined, and which degrees-of-freedom that should be
measured at each node are determined.
3. The structure is instrumented with sensors. If not all sensors can be measured at once, a test
scheme is made, describing in which sequence the nodes should be measured.
4. If a shaker test is performed (instead of an impact test), force sensors and shakers are installed.
5. Some preliminary measurements are made to determine optimal measurement settings.
6. The data are acquired.
7. The quality of the data is assessed.
8. The modal parameters are estimated based on the measured FRFs.
9. The modal analysis results are assessed for quality.

2 Experimental Setup
Let us state it right up front: the success of an EMA test lies in the measurements; if you do not
acquire top quality data, no matter how advanced your modal parameter extraction algorithm is,
you will not be able to obtain any reliable modal parameters. I often state that modal analysis is
80 % measurements (if we include the planning, pretest, or whatever you call it) and only (at most!)
20 % parameter extraction.
It is very important to realize the importance of good measurements, and the main reason we
need to have good measurements, is that errors typically occuring when measurements go bad are
bias errors and thus no algorithm can remove the effect of those errors. So, this section on how to
set up and check your measurements, is the most important section in this technical note.

2
2.1 Points and DOFs
In the following we will need some terms to specify the measurement degrees-of-freedom etc. We
will use the term node for a particular location on the structure (usually described by its x-, y-
, and z-coordinates in a Cartesian coordinate system). In each node, we can measure the three
translatory directions, x, y, and z. The combination of a node and a direction, we call a DOF.
Of course, we also have the rotational degrees-of-freedom in each node. These are, however, very
difficult to measure due to lack of good sensors for rotational degrees-of-freedom, so they are usually
not included in EMA. In most software for EMA, the nodes are simply numbered by an integer
number, and the directions are usually entered in the software as X+, Y+, Z+, or sometimes +X,
+Y, +Z. In ABRAVIBE both these label conventions can be used. If a sensor is mounted in a
negative direction, it is instead entered as X-, Y-, etc.

2.2 General Points

Before going into detail with the various considerations and practices, some words should be said
about the general test layout, and choice of equipment etc. To ensure top quality data, it is
important to use the right equipment. Today, this usually means IEPE sensors for both force and
accelerometer measurements. The dynamic range of IEPE sensors is a factor limiting the quality
of FRF measurements, so it must be ensured that the signal-to-noise ratio is as good as possible.
Since IEPE sensors have 5 V full scale voltage, a good indicator is that the maximum voltage from
the sensors is above, say, 0.5 to 1 V. Accelerometers should usually have sensitivity close to 1 V/g
(500 mV/g is ok, but usually not 100 mV/g).
I want to make an important point here; since IEPE sensors have fixed measurement range, it is
necessary to keep force sensors with several sensitivites in your lab equipment (if you do not always
measure the same type of structure, perhaps) so that you can select a sensor with proper sensitivity
for each test. This applies also to the impact hammer; it is necessary to have a few impact hammers
with different sensitivities (and also, with different weight).
Another choice is between using a few accelerometers which are roved over the structure, until
all FRFs are measured, or using enough accelerometers so that all accelerometers can stay on the
structure during the measurements. The choice here is, of course, often limited by the lab budget,
but it should be made clear that for top quality results, for many structures it is necessary to use
the latter approach. By using roving accelerometers, you have to interact with the structure which
is likely to produce some time invariance (you change the structure between the measurements),
and you are also more likely to encounter mass loading (also see Section 2.5).
Should you use impact testing or shaker testing? Again, for top quality results, for most struc-
tures it is required to use shaker testing (and I am then a strong advocate of using at least two
shakers simultaneously, see Section 2.5). For most trouble-shooting cases, it can be quite satisfactory
to use impact testing.
If you decide on impact testing, your choice is first to use roving impact strategy (accelerometer
reference sensors are fixed and you move the hammer around to each DOF and make a separate
measurement), or if you should use a fixed force location and apply accelerometers to all DOFs, or
if you should use a fixed force location and rove accelerometers around the structure. The most
practical choice here is definitely if you can use a roving force. You should then essentially always use

3
more than one reference accelerometer, the more the better (except for mass loading considerations,
see Section 2.5).
A reason for not using a roving hammer can be that there are points (degrees of freedom) which
can not (or at least not easily) be excited by an impact hammer. If there are many such DOFs, you
may have to revert to the fixed impact DOF approach. If there are only a few inaccessible DOFs, you
can, however, conveniently use the reciprocity principle, and for those points put an accelerometer
in the DOF, or DOFs, which you can not impact, and then sequentially measure the FRFs between
excitation in each of the reference locations (i.e. where you have your reference accelerometers) and
response in the DOFs you can not impact. You should remember to re-index these FRFs so that
they are indexed as if they were FRFs in the reverse order (force to acceleration location).
The measurement system (data acquisition system) should be a dedicated system for vibration
analysis. It is practical if it includes support for TEDS sensors, i.e. IEPE sensors with built-in
electronics so the measurement system can read for example the sensitivity of each sensor, provided,
of course, that you have TEDS in your accelerometers, see [3]. More important is to have support
for the book keeping part of the data acquisition. It is essential in modal testing that there are no
mistakes made regarding which DOF is measured in each measurement. Without support in the
measurement software, this becomes very difficult. When we are discussing this point, let us stress
once and for all: good EMA results are only obtained if you take it easy, do not stress, be very, very
cautious in every step discussed in this note. And most of all: keep good documenting procedures
so that you know everything that happened during the test.

2.3 Suspension
How to suspend the test structure is often the first thing to decide. Typically there are two possible
suspension options; (i) free-free conditions, or (ii) as supported in real-life (i.e. connected to some-
thing). The main reason for choosing the first option, free-free boundary conditions, is that it is
most likely to produce good measurements. When you support your structure freely, the vibration
energy you put into the structure by the excitation stays in the test structure until it has naturally
decayed due to the damping of the structure. On the other hand, when a structure is connected to a
world around it, a lot of the excitation energy gets lost to the surrounding, which usually means it is
much more difficult to obtain good results. So: if your aim is to make the best possible estimation of
the modal parameters of a structure, then support it free-free. Only for trouble-shooting purposes
should you consider performing an EMA test on a structure installed into its normal environment.
Let us now focus on free-free boundary conditions. Typically, this is achieved by suspending the
test structure on soft springs (or hanging in soft springs). Depending on the size and weight of the
test structure, this can be anything from rubber cords to steel springs, etc. The suspending springs
should be soft enough to produce the six rigid body modes due to the mass and geometry of the
test structure and the stiffness of the springs, at (as a rule of thumb) less than a tenth of the first
eigenfrequency of the structure. To obtain this, a couple of things need to be considered. First of
all, for the vertical rigid body mode, Equation (5.59) in [3] can be used to find how much the spring
will compress/extend for a particular rigid body frequency. As an example, to achieve a (vertical)
rigid body frequency of 1 Hz, then the springs should compress/extend approx. 25 cm.
The other rigid body modes also need some consideration. This is particularly important when
suspending long, slender objects. Such objects should usually be suspended hanging vertically, to

4
make it more likely that the rotational rigid body mode with rotation around the long axis of the
object, becomes sufficiently low. This is an often neglected fact. If long slender objects are hung
horizontally, there is a large risk that the rotational rigid body mode has a high frequency due to
the small mass moment of inertia.
A very important issue is to test the influence of the springs to ensure that the suspension springs
are not affecting the structure. This will be discussed in Section 2.5.
Rubber cords are very common for suspending smaller objects with free-free boundary conditions.
Some words about this are necessary. It should always be avoided to let the structure be touched
by the rubber cords, because this is likely to add damping to the structure. Instead, the structure
should be hung in for example fishing line (nowadays there are woven fishing lines which are very
strong and suitable for this purpose) or thin steel thread, which is then connected to the rubber
cords.

2.4 Pretest
We will use the term pretest for the selection of reference and response points. This is an important
consideration, and often results from finite element analysis (FEA) or other information of the mode
shapes and eigenfrequencies can be used to find suitable measurement DOFs. How the selection of
measurement points, hereafter referred to as nodes, is made depends on the use of the final modal
model. If the purpose of the modal test is to trouble-shoot some vibration problem, the nodes
can usually be determined by experience, with some thought of what the mode shapes will look
like. Approximately three nodes per wavelength of the mode shapes are usually sufficient for this
purpose.
If the purpose of the EMA is to compare the results to a computational model (for example
an FE model), there is usually higher demands on the selection of nodes. One reason for this is
that, in order to correlate the mode shapes, the off-diagonal values of the cross-MAC matrix should
be close to zero. Therefore, it is recommended, if possible, to extract nodes from the FE model
and compute an auto-MAC of the mode shapes defined by the selected nodes and make sure the
off-diagonal values are close to zero. Several methods exist to extract (sub)optimal nodes from FE
mode shapes, see e.g. [4].
Once the measurement nodes and DOFs are determined, a geometry description should be
made, by measuring the coordinates (usually x, y, and z in a rectangular coordinate system, or
corresponding coordinates for a cylindrical or spherical coordinate system in case of a cylindrical or
spherical test structure). The best practice is to enter this geometry into the software to be used,
to avoid mistakes. Having a picture of the geometry also helps during the acquisition of the data ,
to ensure the correct DOF information is allocated to each measurement channel.

2.5 Measurement Checks

How to best measure frequency response functions, regardless of whether by impact testing or by
shaker testing, is well described in [3] and will not be treated here. For impact testing, the method
described in [5] (also in [3]) is recommended. In this section we will summarize some tests which are
important to ensure the measurements will be possible to use for good modal parameter extraction.
It should perhaps be stressed here, that it is very important not to give up too early, with too poor

5
coherence functions as a result. In most cases with free-free structures, and with the correct selection
of accelerometers, force sensors, and measurement settings, it is possible to obtain coherence levels
very near unity.
For shaker excitation, it is necessary to check that the force sensor, stinger (Br. Eng.: rod), and
the shaker are correctly attached, particularly so that there is no transverse vibration caused by bad
alignment of the shaker. This is efficiently tested by assuring reciprocity between the force location
and some other point on the structure. This is definitely easiest to check by having (at least) two
shakers, and is a key reason why I always recommend using multiple shakers. If only one shaker is
used, the reciprocity can be checked by mounting an accelerometer where the force sensor is to be
installed, and using an impact hammer at an arbitrary point on the structure. A cross-FRF in the
backwards direction is then measured, before the force sensor is mounted. Then, after the force
sensor is mounted, the same FRF is measured in the opposite direction (now with an accelerometer
where the impact hammer was exciting before). If the two FRFs agree, it is a strong indication
that the force sensor and shaker are correctly installed. The reciprocity is, of course, much easier
to investigate if there are at least two shakers, and a reference accelerometer is kept in each force
DOF during the test. This is recommended so that one can keep track that the reciprocity stays
good during the entire test.
The next step, regardless of impact hammer or shaker excitation, is to determine the measure-
ment settings. For this, follow the procedures discussed in Chapter 13 in [3].
After the optimum measurement parameters are determined, it is necessary to make some checks
before starting to acquire data, to ensure everything with the measurement setup is correct. First,
mass loading should be checked. The easiest test is to measure an FRF between two points and
store this FRF. Then, an accelerometer of the same kind as the one used for the first measurement
is used as a dummy mass, and mounted right next to the first accelerometer. The same FRF is then
measured again, and is then compared with the first one. If the two FRFs are not identical, there
is likely to be mass loading already by the mass of the first accelerometer. Two considerations can
then be made; either a lighter kind of accelerometer is chosen for the measurements, or, if this is
not possible, all accelerometers can be added to the structure prior to the measurement. Although
this will mean some (perhaps quite substantial) mass loading of the structure, the measurement
will still be consistent, and the effect of the masses can be removed after the analysis. If not
enough accelerometers are available, and the test strategy is to rove the accelerometers between
measurement sets, dummy masses of approximately the same shape and with the same mass as the
accelerometers used can be used in the positions where accelerometers are not located, so that the
same mass distribution is achieved in each data set.
After checking the mass load effect and solving any issues with that, it is suitable to check the
suspension. This is essential, as it is very common to neglect significant effects from the suspension
that lead to erroneous results, particularly added damping. A way to ensure minimal effect of the
suspension on the modal parameters, is to measure some FRFs first with the structure suspended as
it should be suspended during the measurement. These FRFs are then used with some parameter
extraction technique, to estimate the natural frequencies and poles. A first check can now be made
by making sure that the highest peak in the FRF caused by the rigid body modes, is occuring at
a frequency less than a tenth of the first eigenfrequency of the structure (which is a good rule of
thumb, but no guarrantee that there is no influence on the modal parameters from the suspension).
If the rigid body modes are found at sufficiently low frequencies, the suspension is then changed

6
substantially, for example by shortening the springs (if the structure is hanging in rubber cords or
other type of suspension where the spring constant can be changed easily). FRFs between the same
nodes as before are then measured, and processed for modal parameters. The natural frequencies
and damping factors of the structure modes should be the same as in the first test. If not, there is
effect on the modal parameters from the suspension and this needs to be addressed.
The final check should be to measure some FRFs between excitation and response in different
directions, and between excitation and response far away from each other. In both cases it should
be ensured that the coherence is acceptable. In most cases this should mean that it is very near
unity.

2.6 Calibration
Correct calibration of all force sensors and accelerometers is very important in order to get top
quality results. It is not so much because a correctly scaled model is essential, but with multiple
references, without accurate calibration the data will not be consistent. This will inevitably lead
to errors in the multiple-reference parameter extraction routine used for estimating (scaled) mode
shapes, see Section 3.2.2. If shakers are used it is best to use the calibration factors from the latest
official calibration (remember to recalibrate your sensors regularly!). If, however, an impact test
with roving hammer is performed, a mass can be used for calibration, see Chapter 7 in [3].

2.7 Data Acquisition

Once all the checks mentioned in the previous sections have been performed, it is time to acquire the
data. This should be done in as short time as possible, keeping conditions (for example temperature)
as constant as possible, to avoid time invariance in the data. As mentioned in [3], the best strategy
is usually to store time data, so that the data processing can be redone at a later stage. With shaker
excitation, however, since (except if true random is used for excitation, but this should be avoided)
all non-windowing excitation signals necessitate fixing the measurement settings (predominantly the
blocksize) prior to the measurement, there is little need to store the time data. If time data are
not stored, it is good practice to store at least FRFs, coherence functions, and force spectra. All of
these functions may be necessary to use for assessment of the measurement quality.

2.8 Checklist
To summarize the above points, the following check list can be used to ensure that all the most
important checks are remembered during the measurement setup phase. This check list is also
available as a separate page on http:www.abravibe.com. No list like this can, unfortunately, be
entirely satisfactory. This is particularly true for EMA, which is often considered an artform more
than a technical application. Anyhow, this list should at least serve as some sort of guideline for
the inexperienced user, ensuring the most important things are not forgotten.

1. Plan the test by determining suitable measurement DOFs, reference points, etc.

2. Build the geometry description up in your software

3. Support the structure according to your decision (free-free or alternatively supported)

7
4. Mark all measurement nodes on the structure. If necessary measure the exact location.
5. For a shaker test
(a) Mount the accelerometers and possibly dummy masses, draw cables to the measurement
system
(b) Mount the force sensor(s), align the stinger/shaker(s) and attach it/them.
(c) Select your choice of excitation signal. The preferred excitation signal should be pseudo
random (single shaker) or periodic random (multiple shakers). Turn the amplifiers volume
knob all the way down and turn the signal generator(s) on. Then slowly turn the amplifier
volume up until you (barely) hear your excitation signal if you put your ear against the
structure (note! this is of course not always possible, for example if the frequency range
is below the audible range.). This is the preferred level, increasing the excitation level
increases the risk of nonlinearities!
(d) Look at the voltage levels of your accelerometer signals and ensure they are close enough
to 5 V, assuming you are using IEPE sensors. If you are using some other sensors, you
may need to change the input range of your measurement system for each channel.
(e) Try out measurement settings (frequency range, excitation signal, blocksize, perhaps other
settings like burst length for burst random excitation) until the quality of each FRF is
good (assessed by ensuring the coherence function is near unity).
(f) When the right measurement parameters are found, check reciprocity. If not in order, try
disconnecting all shakers, realigning, and connect again, until the reciprocity is good.
(g) Look at the imaginary part of each driving point FRF (if accelerances; real part if mobility)
and ensure that all peaks point in the correct direction (which depends on the individual
directions of the force sensor and accelerometer; if both point in the same direction, the
imaginary part should peak in positive direction).
(h) Next, investigate the correlation between the force signals, to avoid unreliable FRF es-
timates. There are different ways to this, depending on your measurement system. I
personally prefer to investigate the virtual coherences (see [3]).
6. For impact testing, instead do the following (for simplicity we assume a roving hammer test!)
(a) Mount the reference accelerometers.
(b) Test out the right hammer tip and measurement settings. Follow the instructions in [3].
(c) Check another excitation location far from the first one, to ensure the settings work also
for other impact locations.
7. You are now ready to start acquiring data. If you need to do this in several steps, be careful
when moving accelerometers so that you do not change the position of the structure on the
supporting springs etc. For each set of data, make sure to check all FRFs and coherence
functions before moving on, so that you are sure the data are good! Also be very careful and
ensure that you enter the correct measurement DOFs for all channels in each new measurement.
8. Once your data are acquired, proceed with the modal parameter extraction. This is the topic
for the next Section.

8
3 Modal Parameter Extraction Methods
In this section we will discuss some theory of the parameter extraction techniques that are im-
plemented in the command frf2ptime (FRFs to poles by time methods) in ABRAVIBE. We will
exclude the SDOF methods which are implemented in the command frf2msdof, because these are
well documented in the book [3]. There are thus three methods, or rather, one method in three dif-
ferent versions; the Prony method (Prony), the least squares complex exponential method (LSCE),
and the polyreference time domain method (PTD). All these methods compute the poles of the
system (natural frequencies and mode shapes). Prony is used for a single FRF, LSCE for many
FRFs for the same reference, and PTD for a test with several references (either a multiple shaker
test, or an impact test with several reference accelerometers).
Common to all the methods in this family is, that they produce the poles, and, in the case of the
PTD method, also the modal participation factors (see Section 3.1.3). After calling the command
frf2ptime, the command frfp2modes should therefore be used. This command implements the
least squares frequency domain method to estimate scaled mode shapes, based on the poles (and for
multiple references the MPFs).

3.1 Extraction Methods for Poles

All time domain parameter extraction methods for modal parameters rely on the modal decompo-
sition in the time domain
N

X
[h(t)] = [Ar ] esr t + [Ar ] esr t (1)
r=1
where [h(t)] is the matrix of impulse responses, [Ar ] the matrix of residues for pole sr , and N the
number of modes. The residue element between force in DOF q and response in DOF p, can be
divided into mode shape coefficients and modal scaling constants, by

Apqr = Qr p rq r. (2)

For more in-depth discussions of these and other methods for EMA parameter extraction, please see
the excellent books [6, 7, 8].

3.1.1 Pronys Method

The oldest method we have, which can be used as a local MDOF method, is the method due to
Baron de Prony in 1795, [9]. It is sometimes called the complex exponential method for reasons
obvious from the following. This method, and evolutions of it, are the most common methods for
extracting modal parameters. We will therefore go through this method in some detail, as it is
important to understand the basis of it, as this basis is essentially common for all time domain
methods.
Pronys method is based on the fact that for an impulse response, or a free decay, h(t), if we
assume we have N modes, in the SISO case we have that
2N
X
r h(ti+r ) = 0. (3)
r=0

9
because we have no force.
This equation is then rewritten by extracting the last value in the sum in Equation (3) and
by setting the coefficient 2N = 1, which does not change the result of Equation (3), because the
equation still sums to zero. The result of this trick is the Prony formulation,
2N
X 1
h(ti+2N ) = r h(ti+r ). (4)
r=0

Equation (4) can be formulated into a vector equation

0

1

h(ti ) h(ti+1 ) h(ti+1 ) . . . h(ti+2N 1 ) = h(ti+2N ). (5)

...

2N 1

The next trick is to formulate Equation (5) for a number of time values, usually overdetermining
the equation system. If we assume we have N modes, i.e. 2N poles, as in the previous equations,
we use L > 2N rows added to Equation (5) which then turns into

h(ti ) h(ti+1 ) h(ti+1 ) . . . h(ti+2N 1 )
0
h(ti+2N )
h(ti+1 ) h(ti+2 ) h(ti+3 ) . . . h(ti+2N ) 1
h(ti+2N +1 )

= . (6)
... ...

...

h(ti+L1 ) h(ti+L ) h(ti+1 ) . . . h(ti+2N +L1 ) 2N 1 h(ti+2N +L )

Note that the matrix in Equation (6) contains the same impulse response (or free-decay function),
shifted one sample forward for each column. (The same applies to the rows, but we will refer to the
columns in the following). Such a matrix is called a Hankel matrix. Sometimes in modal parameter
extraction it is referred to as a Toeplitz matrix, which works equally good; a Toeplitz matrix is a
Hankel matrix flipped upside down. The name Hankel matrix is, however, more common in current
parameter identification literature.
The coefficients r in Equation (6) are best solved by using a least squares approach or a pseudo
inverse. The final step is then to solve the roots of the polynomial, i.e.

2N 1 z 2N + 2N 1 z 2N 1 + . . . 0 = 0 (7)

which produces poles in the z-domain, zr , which are then converted to poles, sr in the Laplace-
domain by observing that
zr = esr t (8)
and thus
ln(zr )
sr = = fs ln(zr ). (9)
t
To implement the Prony method in a modern way, Equation (6) is solved for a number of assumed
orders (i.e. number of modes), 2N = 2, 4, . . . , 2Nmax , by using 2N of the columns in Equation (6)
which produces as many coefficients r . For each of these orders, the roots are computed and
converted into complex poles. A stabilization diagram as in Figure 1 is then produced, where poles

10
 Stabilization Diagram, green=stable; blue=stable freq.; red=unstable
60

50

40
Number of Poles

30

20

10

Select poles!
Then <RETURN>
0
200 300 400 500 600 700 800 900 1000
Frequency [Hz]

Figure 1: Example of a stabilization diagram for Pronys method.

are plotted along the frequency axis as x-axis, and the order as y-axis. If two poles at order n and
n + 1 are within some defined limits in frequency and damping estimates they are called frequency
stable or frequency and damping stable, respectively, and different symbols are used to indicate this.
From the stabilization diagram, the poles can be selected by the user by clicking once on each pole
which is to be stored into a pole table. It is practical to clean the stabilization diagram up by only
plotting poles which come in complex conjugate pairs, and poles that have positive damping.
The Prony method thus produces poles for one impulse response function. In a larger measure-
ment, the method can be extended by, for example, estimating the poles for each impulse response
function, and then averaging the pole estimates of each mode to obtain more accurate estimates.
However, it may be even better to use a more sophisticated method such as the LSCE or ITD
method if more than one impulse response functions are available.

3.1.2 The Least Squares Complex Exponential Method

The Least Squares Complex Exponential (LSCE) method is a global MDOF single-reference method,
which was first published in [10], and is an extension of Pronys method to a number of impulse
response functions and not just a single one. This is done by noting that the same coefficients, r ,
apply for each impulse response, and thus the matrix in Equation (6) and the right-hand side vector,
can be added with additional rows for each impulse response we take into account. The solution is
then similar to that for Pronys method. By solving the system using different numbers of columns
in the matrix (and thus coefficients r ), different orders are obtained, and a stabilization diagram
as described for Pronys method can be produced, see Figure 2.

11
3.1.3 Polyreference Time Domain
The Polyreference Time Domain method (PTD) developed by Vold, [11] is a natural extension of
the LSCE method, to multiple references. It is a MIMO method that can handle repeated or closely
spaced poles, and it extracts poles and modal participation factors. The PTD method is probably
the modal parameter extraction method most used in practice, because it has been popular in
commercial software.
To describe the PTD method, we will essentially follow the description in [12]. The PTD method
starts with impulse responses between an arbitrary response (output) point p and all reference
(input) points q = 1, 2, . . . , R, which can be written
2N
X
hp1 (t) = Ap1r esr t
r=1
2N
X
hp2 (t) = Ap2r esr t
r=1
(10)
...
2N
X
hpR (t) = ApRr esr t .
r=1

Each residue in this equation can be formulated relative to another residue by noting that (see
Chapter 6 in [3])
Apqr = Qr pr qr (11)
and thus, for another reference point, say m, we can write
mr
Apmr = Apqr . (12)
qr

We now introduce the modal participation factor, Lmqr ,

mr
Lmqr = (13)
qr

12
whereby we can rewrite Equation (10), arbitrarily using reference 1 as the relative reference, as

2N
X
hp1 (t) = Ap1r esr t
r=1

2N
X
hp2 (t) = L21r Ap1r esr t
(14)
r=1

...

2N
X
hpR (t) = LR1r Ap1r esr t .
r=1

Thus for time t = nt we can write Equation (14) in matrix form as

" #
I
{hp (n)} = n {Ap1 } (15)
[Lm1 ]

where I is a row vector with ones, the modal participation matrix [Lm1 ] is

L21,1 L21,2 . . . L21,2N
L31,1 L31,2 . . . L31,2N
[Lm1 ] =
...

(16)
LR1,1 LR1,2 . . . LR1,2N

the residue vector {Ap1 } is

Ap1,1

Ap1,2

{Ap1 } = (17)

...

Ap1,2N

and finally the diagonal matrix with the poles, is

s1 t
e 0 ... 0
s t

0 e 2 ... 0
=
...
.
(18)

0 0 0 es2N t

13
 Using Nt + 1 time samples we can repeat Equation (15) for n = 0, 1, . . . , Nt and get

{hp (0)} = [L] {Ap1 }

{hp (1)} = [L] {Ap1 }
{hp (2)} = [L] 2 {Ap1 } (19)
...
{hp (Nt )} = [L] Nt {Ap1 } .

The RNt poles and corresponding modal participation factors that we want to estimate can be
shown to be solutions to a matrix polynomial

[(0)] + [(1)] [L] + . . . + [(Nt )] [L] Nt = [0] (20)

where each [(n)] is size (R R). Each matrix coefficient comes from a multiplication in Equa-
tion (19)
[(0)] {hp (0)} = [(0)] [L] {Ap1 }
[(1)] {hp (1)} = [(1)] [L] {Ap1 }
[(2)] {hp (2)} = [(2)] [L] 2 {Ap1 } (21)
...
[(Nt )] {hp (Nt )} = [(Nt )] [L] Nt {Ap1 } .
Summing both sides of Equation (21) yields
Nt
X Nt
X
[(n)] {hp (n)} = [(n)] [L] n {Ap1 } (22)
n=0 n=0

from which it can be seen that the right-hand side equals Equation (20)times the factor {Ap1 } which
is independent of n, and thus the right-hand side equals zero. The solutions to Equation (22) do
not change if we scale it by letting the matrix coefficient [(Nt )] = I. Thus we can rewrite it as
N
X t 1

[(n)] {hp (n)} = {hp (Nt )} . (23)

n=0

To solve the coefficients of the matrix polynomial, we repeat Equation (23) for a number, m + 1,
of time points, each shifted t, and put each shift in a new column. We thus obtain the equation

{hp (0)} {hp (1)} . . . {hp (m)}
  {hp (1)} {hp (2)} . . . {hp (m + 1)}
[(0)] [(1)] . . . [(Nt 1)] .. =

.
{hp (N t 1)} {hp (Nt )} . . . {hp (m + Nt 1)}
 
{hp (N t)} {hp (N t + 1)} . . . {hp (N t + m)} .
(24)

14
 We rewrite Equation (24) for response p as
[A] [p HNt m (t)] = [p HNt m (t + tNt )] (25)
This matrix equation is now repeated for each response p = 1, 2, . . . , D which gives
[A] 1 HNt m (t) 2 HNt m (t) . . . D HNt m (t)
     
=
(26)
 1  2  D  
HNt 1 (t + tNt ) HNt 1 (t + tNt ) ... HNt 1 (t + tNt )
which can be written as  
[A] [H] = H (27)
This is the equation to solve for the matrix coefficients, usually by a least squares approach or
pseudoinverse.
The final step to obtain the poles and modal participation factors is to combine the coefficient
matrices into the companion matrix defined by

[(N t 1)] [(N t 2)] . . . [(0)] Nt N
zr {L}r zr t {L}r
I [0] ... [0]



z Nt 1 {L} z Nt 1 {L}

r r r r

[0] I ... [0]

.. = zr .. (28)
.. . .
.





{L}r {L}r

[0] ... I [0]
from which the eigenvalues zr give the poles as sr = fs log(zr ) and the corresponding modal partic-
ipation factors are the last R coefficients in the eigenvectors.
Although the mathematics of this method is somewhat intimidating, the process to implement
it is not that complicated. To simplify it, we shall summarize the procedure in a point list.
1. Select the impulse responses to be used for parameter extraction
2. Decide on the highest order (number of poles), N , and set Nt = N/R (Nt being an integer)
3. Build the block Hankel matrices [H] and [H ] using Equations (24) to (27)
4. For each mode order N = 1, 2, . . . , Nt
(a) Use N R rows of the Hankel matrices, to find the N coefficient matrices [(n)] by solving
Equation (27) in a least squares sense
(b) Build the companion matrix defined by Equation (28)
(c) Find the eigenvalues zr and eigenvectors of the companion matrix
(d) Convert the eigenvalues to poles, and extract the last R coefficients of the eigenvector as
the modal participation factors for the corresponding pole
5. Produce a stabilization diagram based on the estimates of the poles and modal participa-
tion factors and let the user extract appropriate poles. For each selected pole, extract the
accompanying participation factors.
Note that point 4 means that only pole orders R, 2R, . . . will be computed by the PTD method.

15
3.2 Methods for Computing and Scaling Mode Shapes
3.2.1 The Least Squares Frequency Domain Method Single Reference Case
The most common method used to estimate mode shapes is the least squares frequency domain
(LSFD) method. The main reason for this is that in the frequency domain, residual factors can be
estimated to account for the effects of modes outside the freqency band used for the mode shape
estimation. This means that, in most cases, better mode shapes are estimated using this method
than using the least squares time domain (LSTD) method.
We start with this method for a single reference. We then have that, for an arbitrary measured
frequency response, if we use N complex conjugate pair of poles
N
X Apqr Apqr
Hpq () = + . (29)
j r j r
r=1

In most cases we will only use a limited frequency band for the extraction of a limited number of
modes. The modes that lie outside of this frequency band thus influence the frequency response
inside the frequency band. To illustrate this, if we assume that the modes we have in the frequency
range 1 to 2 correspond to the modes N1 to N2 , we can rewrite Equation (29) as
NX
1 1 N2
Cpqr X Apqr Apqr
Hpq () = + + +
r2 2 + j2r r j r j r
r=1 r=N1
(30)
N
X Dpqr
+ 2 2
r + j2r r
r=N2 +1

where we assume it is only the middle sum that accounts for the modes in the frequency region we
use for extracting the mode shapes. We have also added the two terms together for each of the sums
for the outside modes. This is made to more easily see what these outside sums add up to. If we first
consider the modes lower than the frequency range of interest, the leftmost sum in Equation (30),
in the frequency region 1 to 2 , the frequencies are much larger than the highest natural frequency
N1 1 , and thus the sum is approximately
NX
1 1
Cpqr RI
2 (31)
r2 2
+ j2r r
r=1

where we call the constant RI the residual inertia. Similarly the upper sum, in the frequency range
above 2 can be approximated by
N
X Dpqr
RF (32)
2 2+ j2r r
r=N2 +1 r

where we denote the constant RF the residual flexibility. Note that these residual terms are indi-
vidual terms for each measured frequency response Hpq (). The use of residual terms can make the

16
estimates of the mode shape coefficients more accurate in cases where there is a strong influence
from out-of-band modes.
To estimate the mode shape coefficients plus the two residual terms for an arbitrary frequency
response, we can set up a matrix equation in the case of a single reference. In most cases it is
preferable to renumber the poles from r = 1, 2, . . . , 2N by treating each pole as a separate pole
number instead of keeping the complex conjugate pairs. We then define an eigenvalue matrix, by
1 1 1
. . . j1 1 12

j1 1 j1 2 2N 1
1 1 1

j2 1 j2 2 . . . j2 2N
1 12
[] = ..
2
(33)
.


1 1 1
jM 1 jM 2 . . . jM 2N 1 12
M

where we have denoted the total of M discrete frequencies by n = n2f for the frequency lines
n that are used. If no residuals are wanted, the last two columns are left out. We next define the
residue coefficient vector for point p by


Apq1
Apq2



..




{Ap } = . (34)


Apq2N



RF





RI

where, again, the last two coefficients are left out if no residuals are wanted. Using Equations (33) and (34)
we can now formulate the matrix equation corresponding to Equation (29), which becomes

[] {Ap } = {Hpq ()} (35)

where the vector on the right-hand side is the measured frequency response defined at the same
frequencies as was put in the matrix . Equation (35) is an overdetermined system assuming the
number of frequencies M is larger than the number of modes, which is, in principle, always the
case. It can therefore be solved, for each measured location p, by a least squares or pseudoinverse
approach.
It remains to scale the mode shapes. We refer to Chapter 6 in [3] where it is shown that leaving
the modal scale constant Qr to unity, the mode shapes are scaled to unity modal A. To find this
scaling, we normalize the residue vector using the coefficient Aqqr so that the final, scaled mode
shape for mode r is p
{}r = {A}r / Aqqr . (36)

3.2.2 The Least Squares Frequency Domain Method Multiple Reference Case
The LSFD method in the case of multiple references is based on the fact that we know the poles
and modal participation factors, and that we wish to find the scaled mode shapes. We thus must
have used a multiple-reference method such as the PTD method to extract the poles and the modal
participation factors.

17
 If we look at one response location, p, if we add residual terms similarly to the single reference
case in Section 3.2.1, we have that
1
{Hp ()} = [L]T {p } + {RF ()} {RI ()} (37)
2
where the diagonal eigenvalue matrix is defined by
1
0 ...

j1

1

0 0 ...
j2
= (38)

..

.
0 1
... j2N

and the measured column vector {Hp ()} contains the frequency responses for response DOF p
relative to all the references.
To obtain a single matrix equation that we can solve for the mode shapes, we need to define
some new matrices. First, we define an extended modal participation matrix, [L ] by
  h i
L = [L]T I I (39)

and an extended eigenvalue matrix [ ] by

  [0] [0]

= [0] I
[0] (40)
[0] [0] 12 I

and finally the vector with mode shape coefficients for response DOF p,


p1



p2



..




{p } = . (41)


p2N

{R }

F




{RI }

where each of the residual vectors have size (R 1). Using these definitions, Equation (37) can now
be rewritten as   
{Hp ()} = L {p } = [B()] {p } (42)
which is the base equation for the least squares frequency domain method. What remains is to
produce an overestimated system of equations. This is achieved by repeating Equation (42) for all
frequencies in the range of interest, [1 , 2 , . . . , M ]. We thus obtain a large equation


{Hp (1 )} [B(1 )]
{Hp (2 )}
[B(2 )]
.. = .. {p } (43)



.


.
{Hp (M )} [B(M )]

18
which is solved either in a least squares fashion, or by pseudoinverse, for each response location p,
for which the residues and residuals are saved. Note that the matrix [B] in Equation (43) only needs
to be built once, as it is independent on the response location p. Consequently, if the pseudoinverse
of [B] is used, it only needs to be computed once.
It remains to scale the mode shapes. After solving Equation (43), assume we have stored
the results in a residue matrix [A]r for each mode r. For multiple references, we use the modal
participation factors relative to one of the references to get the scaling. From Equations (11) to (13)
it follows that a modal participation factor Lqr is

Lqr = Qr qr . (44)

We also know from [3] that scaling modes to unity modal A corresponds to using the modal scale
constant Qr = 1. From one of the columns in the stored mode shape matrix obtained by solving
Equation (43) for all responses p and, if residuals were used, removing those from the end of the
vector, we can form a residue column vector for one of the references, q, and mode r, which is

{Apq }r = Lqr {}r = qr {}r . (45)

From this vector, a mode shape vector scaled to unity modal A is computed by dividing this vector
{Apq }r by the square root of the q-th coefficient, which obviously contains the mode shape coefficient
2 . It is usually best to choose the largest modal participation factor for each
squared, i.e. qqr = qr
mode for this scaling process. Particularly, using this factor will avoid using a zero, or close-to-zero
modal participation factor.

4 Examples
4.1 Comparison of the frf2ptime Command Options
To illustrate the differences between the implementations of the Prony, LSCE, and PTD meth-
ods, there is an example called PlexiModalComp available at www.abravibe.com, which we will go
through in this section.
Let us start by looking at the syntax of the command frf2ptime.

[p,L,fLimits] = frf2ptime(H,f,Nlines,MaxModes,MifType,EstType)

p Vector with complex poles with positive imaginary part

L This is a variable which varies with the EstType
for EstType=prony or lsce, L = 1.
for EstType=ptd, L contains the modal participation
factors
fLimits 2-by-1 vector with flo, fhi of selected frequency range

H FRF matrix with accelerances, N-by-D-by-R

f Frequency axis for H
Nlines Number of values in h(t) used for estimation, normally

19
 200-500
MaxModes Maximum number of modes, limit for stabilization
diagram
MifType Mode indicator function type for plot in stabilization
diagram. Can be a column vector with size(f), or a
string corresponding to the input of the AMIF command.
Empty vector or default uses the default mif type in
AMIF.
EstType Estimation method string
prony is a local method useful mostly for a single
FRF, but works also for several responses, although in
a tedious way, as one stabilization diagram is produced
for every response. Prony is also known as complex
exponential method.
lsce uses complex exponential single reference
method using all data in H (even if R>1) (Default)
ptd uses the polyreference time domain method to
produce poles and modal participation matrix

The input and output arguments should be explained by the text in the syntax description. It
should be especially noted, however, that the input to all parameter extraction commands should be
in the form of accelerance FRFs. Mobility or receptance FRFs can be differentiated in the frequency
domain by the fdiff command if necessary.
The PlexiTestPronyLscePtd.m example script in the PlexiModalComp example directory illus-
trates the use of each of the estimation types. First a few words about the beginning of the script.
There, we find the following lines.

DataDir='.\Data';
Prefix=strcat(DataDir,'\','PlexiH');
[H,f,Rdof,Rdir,Fdof,Fdir,FillMtrx]=data2hmtrx(Prefix,1,70,4);
modalchk(H,f,Rdof,Rdir,Fdof,Fdir,FillMtrx,'all');

These lines import and sort all data into the 3D matrix H. It should be especially noted that
the data are sorted as if a shaker test was made, using the reciprocity relations (i.e. if the test
data contain rows in the FRF matrix, this matrix will be transposed, to look as if it was, instead, a
shaker test being made). This way, you should always think as if the data come from a shaker test,
which makes the analysis more consistent.
The command modalchk finds the driving points and reciprocity points (if a multireference test
has been made), and plots the various FRFs overlaid, for the checks mentioned in Section 2.5. It
also produces the matrix FillMtrx, which contains one column for each reference, and as many rows
as there are measured responses. If all data points are measured, this matrix should contain only
the number one in each position. If there is a zero in a position, it means that DOF is not in the
available data.

20
4.2 EstType Prony
We now look at the line using the Prony method to estimate the poles of the system using one of
the measured FRFs.

[p1,L1,flim1]=frf2ptime(H(:,1,1),f,300,30,MIF,'prony');

The number 300 is the number of time values in the impulse response to be used. This number
could be selected differently depending on the blocksize of the FRFs, the damping of the system,
etc. The command produces first a plot where you have to select the lower and upper frequency
range to be used for the parameter extraction. Then it computes a stability diagram, in this case
going from 0 to 60 poles (since the fourth input parameter is number of modes, and each mode
corresponds to two poles). This stability diagram can look like the one in Figure 1. You should
select all the poles you want to use by clicking in the diagram, selecting one pole estimate for each
of the modes. It is usually best to select a stable pole (green plus sign) as high up in the diagram as
possible, but make sure there is no computational pole (red ring) near the pole you select, as this
may affect the stable pole a little. You do not have to select poles at the same model order (y-axis
level); you can choose freely for each pole.
In this example there is no mode shape computed using the Prony method. That is because
Prony should be used only for computing the poles from a single FRF, and not for a complete modal
analysis test. For this, we use the LSCE or PTD methods.

4.3 EstType LSCE

The least squares complex exponential method is called very similarly to the Prony method, with
the difference that we include all the FRFs with the first reference. The same plots are generated,
first to select the frequency range, and then the stability diagram. After selecting the poles in the
stability diagram, the mode shapes are computed using the least squares frequency domain method.
Although this is done further down in the example script file (line 79), we put the two lines next
after each other here for reference.

[p2,L2,flim2]=frf2ptime(H(:,:,1),f,300,30,MIF,'lsce');
fidx2=find(f flim2(1) & f flim2(2));
V2=frfp2modes(H(fidx2,:,1),f(fidx2),p2,1,.1);

The stability diagram produced by the LSCE method in the first line can look like the example
in Figure 2, which was produces after selecting approx. 50 to 1050 Hz. The second row in this code
finds the frequency indeces in the FRF matrix that were used to estimate the poles, to ensure that
the mode shapes are computed using the same frequency range, which is usually preferred. The last
line is computing the mode shapes scaled to the force in DOF 1, using the poles in the variable p2.
The method used is the least squares frequency domain technique outlined in Section 3.2.1. When
executing the third line, for each DOF, the measured FRF is plotted overlaid with the synthesized
function based on the estimated poles and mode shapes. This plot is held for 0.1 seconds in the
code above (the last input parameter to the command frfp2modes) which means that the plots will

21
 Stabilization Diagram, green=stable; blue=stable freq.; red=unstable
60

50

40
Number of Poles

30

20

10

Select poles!
Then <RETURN>
0
200 300 400 500 600 700 800 900 1000
Frequency [Hz]

Figure 2: Stabilization diagram for the LSCE method from the example in Section 4.3

flicker relatively fast on the screen. This is only done in order not to make it too tedious having to
watch all plots.

4.4 EstType PTD

The third method is the polyreference time domain method described in Section 3.1.3. This method
is called by using the commands in the following three lines.

[p3,L3,flim3]=frf2ptime(H(:,:,Ref),f,300,30,MIF,'ptd');
fidx3=find(f flim3(1) & f flim3(2));
V3=frfp2modes(H(fidx3,:,:),f(fidx3),p3,L3,.5,Fdof);

The call to frf2ptime is this time using both references in the FRF matrix in variable H (the
variable Ref is a vector equal to [1 2]). The main difference using the PTD method compared to the
LSCE, is that the modal participation factor matrix, L3 in the code, is important as input parameter
to the frfp2modes, for proper scaling of the mode shapes. The PTD method produces a stabilization
diagram similar to the one in 3. When executing the LSFD method, like for the single-reference
case described in Section 4.3, the command produces a plot for each DOF, with the measured FRFs
(in plural, because we now have several references) and the corresponding synthesized FRFs using
the modal model.
After all the methods for pole estimation have been called, a table similar to Table 1 is printed
in the MATLAB desktop. As can be seen in this table, both LSCE and PTD find approximately the
same poles. This means, that although the two first modes are relatively close in frequency, they

22
 Stabilization Diagram, green=stable; blue=stable freq.; red=unstable
60

50

40
Number of Poles

30

20

10

Select poles!
Then <RETURN>
0
100 200 300 400 500 600 700 800 900 1000
Frequency [Hz]

Figure 3: Stabilization diagram for the PTD method from the example in Section 4.4

Cross MAC matrix

0.9

0.8

0.8 0.7

0.6 0.6

0.4
0.5

0.2
0.4

0
145.4 0.3
146.9
332.3
409.3 990.2 0.2
421.4 831.5
746.2
518.8 608.6
608.7 519.5 0.1
421.8
745.1 409.9
832.2 332.7
993.2 147.4
145.7

Figure 4: Cross-MAC between the mode shapes based on the LSCE method versus the mode shapes
based on the PTD method

23
 Natural Frequencies [Hz] Damping [%] Difference [%]
Freq. Damp.
145.277 145.109 3.25 2.86 0.12 12.07
146.948 146.780 2.91 2.75 0.11 5.47
333.944 332.169 2.54 2.70 0.53 -6.18
409.322 409.379 2.46 2.48 -0.01 -0.60
421.404 421.263 2.38 2.40 0.03 -0.57
518.905 518.871 2.33 2.33 0.01 -0.15
608.203 607.863 2.32 2.34 0.06 -0.92
745.623 745.851 2.25 2.28 -0.03 -1.25
831.195 831.070 2.16 2.17 0.02 -0.54
993.457 990.408 2.16 1.93 0.31 10.65

Table 1: Comparison of pole estimates from the LSCE and PTD methods

are still reliably found by the LSCE method. As we will see next, it is the mode shape estimation
that does not work well for single-reference data.
Continuing with the example script, it produces a cross-MAC matrix of the two mode shape sets
from the LSCE poles and the poles and MPFs from the PTD method. This MAC matrix may look
like in Figure 4. It should be noted that the MAC numbers for the first two modes are less than
for the remaining modes. This is due to the fact that the two closely spaced modes around 146 Hz
do not separate very well using the LSFD method with only one reference, despite the fact that the
poles are correctly estimated by the LSCE method. For closely spaced modes like in this example,
a multiple-reference technique should be used for best mode shape estimation. After the script is
executed, there are two files generated in the directory, LSCEPlexiModal and PTDPlexiModal, both
with the extension .mat. Each of these files can be opened in the animation program animate,
for comparison. You will most likely see that the two first modes are badly separated in the mode
shapes based on the LSCE method, whereas they look much better from the PTD method.

References
[1] A. Brandt. ABRAVIBE A MATLAB toolbox for noise and vibration analysis and
teaching. Department of Technology and Innovation, University of Southern Denmark,
http://www.abravibe.com, July 2011.

[2] A. Brandt. The ABRAVIBE toolbox for teaching vibration analysis and structural dynamics.
In Proc. 31st International Modal Analysis Conference (IMAC), Garden Grove, CA, 2013.

[3] A. Brandt. Noise and Vibration Analysis Signal Analysis and Experimental Procedures. John
Wiley and Sons, 2011.

[4] A. Linderholt and T. Abrahamsson. Optimising the informativeness of test data used for
computational model updating. Mechanical Systems and Signal Processing, 19(4):736750,
July 2005.

24
 [5] A. Brandt and R. Brincker. Impact excitation processing for improved frequency response
quality. In Proc. 28th International Modal Analysis Conference, Jacksonville, FL. Society for
Experimental Mechanics, Feb 2010.

[6] N. Maia and J. Silva, editors. Theoretical and Experimental Modal Analysis. Research Studies
Press, Baldock, Hertforsdhire, England, 2003.

[7] D. J. Ewins. Modal Testing: Theory, Practice and Application. Research Studies Press, Baldock,
Hertfordshire, England., 2nd edition, 2000.

[8] W. Heylen, S. Lammens, and P. Sas. Modal Analysis Theory and Testing. Catholic University
Leuven, Leuven, Belgium, 2nd edition, 1997.

[9] Baron Gaspard Riche de Prony. Essai xperimental et analytique: sur les lois de la dilatabilit de
fluides lastique et sur celles de la force expansive de la vapeur de lalkool, diffrentes tempratures.
Journal de lcole Polytechnique, 1(22):24 76, 1795.

[10] D.L. Brown, R.J. Allemang, Ray Zimmerman, and M. Mergeay. Parameter estimation tech-
niques for modal analysis. SAE Tech. Paper 790221, 1979.

[11] H. Vold, John Kundrat, Thomas G. Rocklin, and Richard Russell. A multiple-input modal
estimation algorithm for mini-computer. SAE Tech. Paper 820194, 1982.

[12] F. Deblauwe, D. L. Brown, and R. J. Allemang. The polyreference time domain technique. In
Proc. 5th International Modal Analysis Conference (IMAC), London, England, 1987.

25