Fundamentals of dsp

The Fundamentals of
Signal Analysis
Application Note 243

3
Table of Contents
Chapter 1 Introduction 4
Chapter 2 The Time, Frequency and Modal Domains:
A matter of Perspective 5
Section 1: The Time Domain 5
Section 2: The Frequency Domain 7
Section 3: Instrumentation for the Frequency Domain 17
Section 4: The Modal Domain 20
Section 5: Instrumentation for the Modal Domain 23
Section 6: Summary 24
Chapter 3 Understanding Dynamic Signal Analysis 25
Section 1: FFT Properties 25
Section 2: Sampling and Digitizing 29
Section 3: Aliasing 29
Section 4: Band Selectable Analysis 33
Section 5: Windowing 34
Section 6: Network Stimulus 40
Section 7: Averaging 43
Section 8: Real Time Bandwidth 45
Section 9: Overlap Processing 47
Chapter 4 Using Dynamic Signal Analyzers 49
Section 1: Frequency Domain Measurements 49
Section 2: Time Domain Measurements 56
Section 3: Modal Domain Measurements 60
Appendix A The Fourier Transform: A Mathematical Background 63
Appendix B Bibliography 66
Index 67

4
The analysis of electrical signals is
a fundamental problem for many
engineers and scientists. Even if the
immediate problem is not electrical,
the basic parameters of interest are
often changed into electrical signals
by means of transducers. Common
transducers include accelerometers
and load cells in mechanical work,
EEG electrodes and blood pressure
probes in biology and medicine, and
pH and conductivity probes in chem-
istry. The rewards for transforming
physical parameters to electrical sig-
nals are great, as many instruments
are available for the analysis of elec-
trical signals in the time, frequency
and modal domains. The powerful
measurement and analysis capabili-
ties of these instruments can lead to
rapid understanding of the system
under study.
This note is a primer for those who
are unfamiliar with the advantages of
analysis in the frequency and modal
domains and with the class of analyz-
ers we call Dynamic Signal Analyzers.
In Chapter 2 we develop the concepts
of the time, frequency and modal
domains and show why these differ-
ent ways of looking at a problem
often lend their own unique insights.
We then introduce classes of instru-
mentation available for analysis in
these domains.
In Chapter 3 we develop the proper-
ties of one of these classes of analyz-
ers, Dynamic Signal Analyzers. These
instruments are particularly appropri-
ate for the analysis of signals in the
range of a few millihertz to about a
hundred kilohertz.
Chapter 4 shows the benefits of
Dynamic Signal Analysis in a wide
range of measurement situations. The
powerful analysis tools of Dynamic
Signal Analysis are introduced as
needed in each measurement
situation.
This note avoids the use of rigorous
mathematics and instead depends on
heuristic arguments. We have found
in over a decade of teaching this
material that such arguments lead to
a better understanding of the basic
processes involved in the various
domains and in Dynamic Signal
Analysis. Equally important, this
heuristic instruction leads to better
instrument operators who can intelli-
gently use these analyzers to solve
complicated measurement problems
with accuracy and ease*.
Because of the tutorial nature of this
note, we will not attempt to show
detailed solutions for the multitude of
measurement problems which can be
solved by Dynamic Signal Analysis.
Instead, we will concentrate on the
features of Dynamic Signal Analysis,
how these features are used in a wide
range of applications and the benefits
to be gained from using Dynamic
Signal Analysis.
Those who desire more details
on specific applications should look
to Appendix B. It contains abstracts
of Agilent Technologies Application
Notes on a wide range of related
subjects. These can be obtained free
of charge from your local Agilent
field engineer or representative.
Chapter 1
Introduction
* A more rigorous mathematical justification for the
arguments developed in the main text can be found
in Appendix A.

5
A Matter of Perspective
In this chapter we introduce the
concepts of the time, frequency and
modal domains. These three ways of
looking at a problem are interchange-
able; that is, no information is lost
in changing from one domain to
another. The advantage in introducing
these three domains is that of a
change of perspective. By changing
perspective from the time domain, the
solution to difficult problems
can often become quite clear in the
frequency or modal domains.
After developing the concepts of each
domain, we will introduce the types
of instrumentation available. The
merits of each generic instrument
type are discussed to give the reader
an appreciation of the advantages and
disadvantages of each approach.
Section 1: The Time Domain
The traditional way of observing
signals is to view them in the time
domain. The time domain is a record
of what happened to a parameter of
the system versus time. For instance,
Figure 2.1 shows a simple spring-
mass system where we have attached
a pen to the mass and pulled a piece
of paper past the pen at a constant
rate. The resulting graph is a record
of the displacement of the mass
versus time, a time domain view of
displacement.
Such direct recording schemes are
sometimes used, but it usually is
much more practical to convert
the parameter of interest to an
electrical signal using a transducer.
Transducers are commonly available
to change a wide variety of parame-
ters to electrical signals. Micro-
phones, accelerometers, load cells,
conductivity and pressure probes are
just a few examples.
This electrical signal, which repre-
sents a parameter of the system, can
be recorded on a strip chart recorder
as in Figure 2.2. We can adjust the
gain of the system to calibrate our
measurement. Then we can repro-
duce exactly the results of our simple
direct recording system in Figure 2.1.
Why should we use this indirect
approach? One reason is that we are
not always measuring displacement.
We then must convert the desired
parameter to the displacement of the
recorder pen. Usually, the easiest way
to do this is through the intermediary
of electronics. However, even when
measuring displacement we would
normally use an indirect approach.
Why? Primarily because the system in
Figure 2.1 is hopelessly ideal. The
mass must be large enough and the
spring stiff enough so that the pen’s
mass and drag on the paper will not
Chapter 2
The Time, Frequency and
Modal Domains:
Figure 2.2
Indirect
recording of
displacement.
Figure 2.1
Direct
recording of
displacement -
a time domain
view.

6
affect the results appreciably. Also
the deflection of the mass must be
large enough to give a usable result,
otherwise a mechanical lever system
to amplify the motion would have to
be added with its attendant mass
and friction.
With the indirect system a transducer
can usually be selected which will not
significantly affect the measurement.
This can go to the extreme of com-
mercially available displacement
transducers which do not even con-
tact the mass. The pen deflection can
be easily set to any desired value by
controlling the gain of the electronic
amplifiers.
This indirect system works well
until our measured parameter begins
to change rapidly. Because of the
mass of the pen and recorder mecha-
nism and the power limitations of
its drive, the pen can only move
at finite velocity. If the measured
parameter changes faster, the output
of the recorder will be in error. A
common way to reduce this problem
is to eliminate the pen and record on
a photosensitive paper by deflecting
a light beam. Such a device is
called an oscillograph. Since it is
only necessary to move a small,
light-weight mirror through a very
small angle, the oscillograph can
respond much faster than a strip
chart recorder.
Another common device for display-
ing signals in the time domain is the
oscilloscope. Here an electron beam is
moved using electric fields. The elec-
tron beam is made visible by a screen
of phosphorescent material.
It is capable of accurately displaying
signals that vary even more rapidly
than the oscillograph can handle.
This is because it is only necessary to
move an electron beam, not a mirror.
The strip chart, oscillograph and
oscilloscope all show displacement
versus time. We say that changes
in this displacement represent the
variation of some parameter versus
time. We will now look at another
way of representing the variation of
a parameter.
Figure 2.3
Simplified
oscillograph
operation.
Figure 2.4
Simplified
oscilloscope
operation
(Horizontal
deflection
circuits
omitted for
clarity).

7
Section 2: The Frequency
Domain
It was shown over one hundred years
ago by Baron Jean Baptiste Fourier
that any waveform that exists in the
real world can be generated by
adding up sine waves. We have illus-
trated this in Figure 2.5 for a simple
waveform composed of two sine
waves. By picking the amplitudes,
frequencies and phases of these sine
waves correctly, we can generate a
waveform identical to our
desired signal.
Conversely, we can break down our
real world signal into these same sine
waves. It can be shown that this com-
bination of sine waves is unique; any
real world signal can be represented
by only one combination of sine
waves.
Figure 2.6a is a three dimensional
graph of this addition of sine waves.
Two of the axes are time and ampli-
tude, familiar from the time domain.
The third axis is frequency which
allows us to visually separate the
sine waves which add to give us our
complex waveform. If we view this
three-dimensional graph along the
frequency axis we get the view in
Figure 2.6b. This is the time domain
view of the sine waves. Adding them
together at each instant of time gives
the original waveform.
However, if we view our graph along
the time axis as in Figure 2.6c, we
get a totally different picture. Here
we have axes of amplitude versus
frequency, what is commonly called
the frequency domain. Every sine
wave we separated from the input
appears as a vertical line. Its height
represents its amplitude and its posi-
tion represents its frequency. Since
we know that each line represents a
sine wave, we have uniquely
characterized our input signal in the
frequency domain*. This frequency
domain representation of our signal
is called the spectrum of the signal.
Each sine wave line of the spectrum
is called a component of the
total signal.
Figure 2.6
The relationship
between the time
and frequency
domains.
a) Three-
dimensional
coordinates
showing time,
frequency
and amplitude
b) Time
domain view
c) Frequency
domain view.
Figure 2.5
Any real
waveform
can be
produced
by adding
sine waves
together.
* Actually, we have lost the phase information of the sine
waves. How we get this will be discussed in Chapter 3.

8
The Need for Decibels
Since one of the major uses of the frequency
domain is to resolve small signals in the
presence of large ones, let us now address
the problem of how we can see both large
and small signals on our display
simultaneously.
Suppose we wish to measure a distortion
component that is 0.1% of the signal. If we set
the fundamental to full scale on a four inch
(10 cm) screen, the harmonic would be only
four thousandths of an inch (0.1 mm) tall.
Obviously, we could barely see such a signal,
much less measure it accurately. Yet many
analyzers are available with the ability to
measure signals even smaller than this.
Since we want to be able to see all the
components easily at the same time, the
only answer is to change our amplitude scale.
A logarithmic scale would compress our large
signal amplitude and expand the small ones,
allowing all components to be displayed at the
same time.
Alexander Graham Bell discovered that the
human ear responded logarithmically to power
difference and invented a unit, the Bel, to help
him measure the ability of people to hear. One
tenth of a Bel, the deciBel (dB) is the most
common unit used in the frequency domain
today. A table of the relationship between
volts, power and dB is given in Figure 2.8.
From the table we can see that our 0.1%
distortion component example is 60 dB below
the fundamental. If we had an 80 dB display
as in Figure 2.9, the distortion component
would occupy 1/4 of the screen, not 1/1000
as in a linear display.
Figure 2.8
The relation-
ship between
decibels, power
and voltage.
Figure 2.9
Small signals
can be measured
with a logarithmic
amplitude scale.

9
It is very important to understand
that we have neither gained nor lost
information, we are just represent-
ing it differently. We are looking at
the same three-dimensional graph
from different angles. This different
perspective can be very useful.
Why the Frequency Domain?
Suppose we wish to measure the
level of distortion in an audio oscilla-
tor. Or we might be trying to detect
the first sounds of a bearing failing on
a noisy machine. In each case, we are
trying to detect a small sine wave in
the presence of large signals. Figure
2.7a shows a time domain waveform
which seems to be a single sine wave.
But Figure 2.7b shows in the frequen-
cy domain that the same signal is
composed of a large sine wave and
significant other sine wave compo-
nents (distortion components). When
these components are separated in
the frequency domain, the small
components are easy to see because
they are not masked by larger ones.
The frequency domain’s usefulness
is not restricted to electronics or
mechanics. All fields of science and
engineering have measurements like
these where large signals mask others
in the time domain. The frequency
domain provides a useful tool in
analyzing these small but important
effects.
The Frequency Domain:
A Natural Domain
At first the frequency domain may
seem strange and unfamiliar, yet it
is an important part of everyday life.
Your ear-brain combination is an
excellent frequency domain analyzer.
The ear-brain splits the audio spec-
trum into many narrow bands and
determines the power present in
each band. It can easily pick small
sounds out of loud background noise
thanks in part to its frequency
domain capability. A doctor listens
to your heart and breathing for any
unusual sounds. He is listening for
frequencies which will tell him
something is wrong. An experienced
mechanic can do the same thing with
a machine. Using a screwdriver as a
stethoscope, he can hear when a
bearing is failing because of the
frequencies it produces.
Figure 2.7
Small signals
are not hidden
in the frequency
domain.
a) Time Domain - small signal not visible
b) Frequency Domain - small signal easily resolved

10
So we see that the frequency domain
is not at all uncommon. We are just
not used to seeing it in graphical
form. But this graphical presentation
is really not any stranger than saying
that the temperature changed with
time like the displacement of a line
on a graph.
Spectrum Examples
Let us now look at a few common sig-
nals in both the time and frequency
domains. In Figure 2.10a, we see that
the spectrum of a sine wave is just a
single line. We expect this from the
way we constructed the frequency
domain. The square wave in Figure
2.10b is made up of an infinite num-
ber of sine waves, all harmonically
related. The lowest frequency present
is the reciprocal of the square wave
period. These two examples illustrate
a property of the frequency trans-
form: a signal which is periodic and
exists for all time has a discrete fre-
quency spectrum. This is in contrast
to the transient signal in Figure 2.10c
which has a continuous spectrum.
This means that the sine waves that
make up this signal are spaced
infinitesimally close together.
Another signal of interest is the
impulse shown in Figure 2.10d. The
frequency spectrum of an impulse is
flat, i.e., there is energy at all frequen-
cies. It would, therefore, require
infinite energy to generate a true
impulse. Nevertheless, it is possible
to generate an approximation to
an impulse which has a fairly flat
spectrum over the desired frequency
range of interest. We will find signals
with a flat spectrum useful in our
next subject, network analysis.
Figure 2.10
Frequency
spectrum
examples.

11
Network Analysis
If the frequency domain were
restricted to the analysis of signal
spectrums, it would certainly not be
such a common engineering tool.
However, the frequency domain is
also widely used in analyzing the
behavior of networks (network
analysis) and in design work.
Network analysis is the general
engineering problem of determining
how a network will respond to an
input*. For instance, we might wish
to determine how a structure will
behave in high winds. Or we might
want to know how effective a sound
absorbing wall we are planning on
purchasing would be in reducing
machinery noise. Or perhaps we are
interested in the effects of a tube of
saline solution on the transmission of
blood pressure waveforms from an
artery to a monitor.
All of these problems and many more
are examples of network analysis. As
you can see a “network” can be any
system at all. One-port network
analysis is the variation of one
parameter with respect to another,
both measured at the same point
(port) of the network. The impedance
or compliance of the electronic
or mechanical networks shown in
Figure 2.11 are typical examples of
one-port network analysis.
Figure 2.11
One-port
network
analysis
examples.
* Network Analysis is sometimes called Stimulus/Response
Testing. The input is then known as the stimulus or
excitation and the output is called the response.

12
Two-port analysis gives the response
at a second port due to an input at
the first port. We are generally inter-
ested in the transmission and rejec-
tion of signals and in insuring the
integrity of signal transmission. The
concept of two-port analysis can be
extended to any number of inputs
and outputs. This is called N-port
analysis, a subject we will use in
modal analysis later in this chapter.
We have deliberately defined network
analysis in a very general way. It
applies to all networks with no
limitations. If we place one condition
on our network, linearity, we find
that network analysis becomes a
very powerful tool.
Figure 2.12
Two-port
network
analysis.
Figure 2.14
Non-linear
system
example.
Figure 2.15
Examples of
non-linearities.
Figure 2.13
Linear network.
θ2
θ2
θ1
θ1

13
When we say a network is linear, we
mean it behaves like the network
in Figure 2.13. Suppose one input
causes an output A and a second
input applied at the same port causes
an output B. If we apply both inputs
at the same time to a linear network,
the output will be the sum of the
individual outputs, A + B.
At first glance it might seem that all
networks would behave in this fash-
ion. A counter example, a non-linear
network, is shown in Figure 2.14.
Suppose that the first input is a force
that varies in a sinusoidal manner. We
pick its amplitude to ensure that the
displacement is small enough so that
the oscillating mass does not quite hit
the stops. If we add a second identi-
cal input, the mass would now hit the
stops. Instead of a sine wave with
twice the amplitude, the output is
clipped as shown in Figure 2.14b.
This spring-mass system with stops
illustrates an important principal: no
real system is completely linear. A
system may be approximately linear
over a wide range of signals, but
eventually the assumption of linearity
breaks down. Our spring-mass system
is linear before it hits the stops.
Likewise a linear electronic amplifier
clips when the output voltage
approaches the internal supply
voltage. A spring may compress
linearly until the coils start pressing
against each other.
Other forms of non-linearities are
also often present. Hysteresis (or
backlash) is usually present in gear
trains, loosely riveted joints and in
magnetic devices. Sometimes the
non-linearities are less abrupt and are
smooth, but nonlinear, curves. The
torque versus rpm of an engine or the
operating curves of a transistor are
two examples that can be considered
linear over only small portions of
their operating regions.
The important point is not that all
systems are nonlinear; it is that
most systems can be approximated
as linear systems. Often a large
engineering effort is spent in making
the system as linear as practical. This
is done for two reasons. First, it is
often a design goal for the output of a
network to be a scaled, linear version
of the input. A strip chart recorder
is a good example. The electronic
amplifier and pen motor must both be
designed to ensure that the deflection
across the paper is linear with the
applied voltage.
The second reason why systems are
linearized is to reduce the problem
of nonlinear instability. One example
would be the positioning system
shown in Figure 2.16. The actual
position is compared to the desired
position and the error is integrated
and applied to the motor. If the gear
train has no backlash, it is a straight-
forward problem to design this
system to the desired specifications
of positioning accuracy and
response time.
However, if the gear train has exces-
sive backlash, the motor will “hunt,”
causing the positioning system to
oscillate around the desired position.
The solution is either to reduce the
loop gain and therefore reduce the
overall performance of the system,
or to reduce the backlash in the gear
train. Often, reducing the backlash
is the only way to meet the
performance specifications.
Figure 2.16
A positioning
system.
∑

14
Analysis of Linear Networks
As we have seen, many systems are
designed to be reasonably linear to
meet design specifications. This
has a fortuitous side benefit when
attempting to analyze networks*.
Recall that an real signal can be
considered to be a sum of sine waves.
Also, recall that the response of a
linear network is the sum of the
responses to each component of the
input. Therefore, if we knew the
response of the network to each of
the sine wave components of the
input spectrum, we could predict
the output.
It is easy to show that the steady-
state response of a linear network
to a sine wave input is a sine wave
of the same frequency. As shown in
Figure 2.17, the amplitude of the
output sine wave is proportional to
the input amplitude. Its phase is
shifted by an amount which depends
only on the frequency of the sine
wave. As we vary the frequency of
the sine wave input, the amplitude
proportionality factor (gain) changes
as does the phase of the output.
If we divide the output of the
network by the input, we get a
Figure 2.17
Linear network
response to a
sine wave input.
Figure 2.18
The frequency
response of
a network.
* We will discuss the analysis of networks which
have not been linearized in Chapter 3, Section 6.

15
normalized result called the frequen-
cy response of the network. As
shown in Figure 2.18, the frequency
response is the gain (or loss) and
phase shift of the network as a
function of frequency. Because the
network is linear, the frequency
response is independent of the input
amplitude; the frequency response is
a property of a linear network, not
dependent on the stimulus.
The frequency response of a network
will generally fall into one of three
categories; low pass, high pass,
bandpass or a combination of these.
As the names suggest, their frequency
responses have relatively high gain in
a band of frequencies, allowing these
frequencies to pass through the
network. Other frequencies suffer a
relatively high loss and are rejected
by the network. To see what this
means in terms of the response of a
filter to an input, let us look at the
bandpass filter case.
Figure 2.19
Three classes
of frequency
response.

16
In Figure 2.20, we put a square wave
into a bandpass filter. We recall from
Figure 2.10 that a square wave is
composed of harmonically related
sine waves. The frequency response
of our example network is shown in
Figure 2.20b. Because the filter is
narrow, it will pass only one compo-
nent of the square wave. Therefore,
the steady-state response of this
bandpass filter is a sine wave.
Notice how easy it is to predict
the output of any network from its
frequency response. The spectrum of
the input signal is multiplied by the
frequency response of the network
to determine the components that
appear in the output spectrum. This
frequency domain output can then
be transformed back to the time
domain.
In contrast, it is very difficult to
compute in the time domain the out-
put of any but the simplest networks.
A complicated integral must be evalu-
ated which often can only be done
numerically on a digital computer*. If
we computed the network response
by both evaluating the time domain
integral and by transforming to the
frequency domain and back, we
would get the same results. However,
it is usually easier to compute the
output by transforming to the
frequency domain.
Transient Response
Up to this point we have only dis-
cussed the steady-state response to a
signal. By steady-state we mean the
output after any transient responses
caused by applying the input have
died out. However, the frequency
response of a network also contains
all the information necessary to
predict the transient response of the
network to any signal.
Figure 2.20
Bandpass filter
response to a
square wave
input.
Figure 2.21
Time response
of bandpass
filters.
* This operation is called convolution.

17
Let us look qualitatively at the tran-
sient response of a bandpass filter. If
a resonance is narrow compared to
its frequency, then it is said to be a
high “Q” resonance*. Figure 2.21a
shows a high Q filter frequency
response. It has a transient response
which dies out very slowly. A time
response which decays slowly is said
to be lightly damped. Figure 2.21b
shows a low Q resonance. It has a
transient response which dies out
quickly. This illustrates a general
principle: signals which are broad in
one domain are narrow in the other.
Narrow, selective filters have very
long response times, a fact we will
find important in the next section.
Section 3:
Instrumentation for the
Frequency Domain
Just as the time domain can be
measured with strip chart recorders,
oscillographs or oscilloscopes,
the frequency domain is usually
measured with spectrum and
network analyzers.
Spectrum analyzers are instruments
which are optimized to characterize
signals. They introduce very little
distortion and few spurious signals.
This insures that the signals on the
display are truly part of the input
signal spectrum, not signals
introduced by the analyzer.
Network analyzers are optimized to
give accurate amplitude and phase
measurements over a wide range of
network gains and losses. This design
difference means that these two
traditional instrument families are
not interchangeable.** A spectrum
analyzer can not be used as a net-
work analyzer because it does not
measure amplitude accurately and
cannot measure phase. A network
analyzer would make a very poor
spectrum analyzer because spurious
responses limit its dynamic range.
In this section we will develop the
properties of several types of
analyzers in these two categories.
The Parallel-Filter
Spectrum Analyzer
As we developed in Section 2 of
this chapter, electronic filters can be
built which pass a narrow band of
frequencies. If we were to add a
meter to the output of such a band-
pass filter, we could measure the
power in the portion of the spectrum
passed by the filter. In Figure 2.22a
we have done this for a bank of
filters, each tuned to a different
frequency. If the center frequencies
of these filters are chosen so that
the filters overlap properly, the
spectrum covered by the filters can
be completely characterized as in
Figure 2.22b.
Figure 2.22
Parallel filter
analyzer.
* Q is usually defined as:
Q =
Center Frequency of Resonance
Frequency Width of -3 dB Points
** Dynamic Signal Analyzers are an exception to this rule,
they can act as both network and spectrum analyzers.

18
How many filters should we use to
cover the desired spectrum? Here we
have a trade-off. We would like to be
able to see closely spaced spectral
lines, so we should have a large
number of filters. However, each
filter is expensive and becomes more
expensive as it becomes narrower,
so the cost of the analyzer goes up
as we improve its resolution. Typical
audio parallel-filter analyzers balance
these demands with 32 filters, each
covering 1/3 of an octave.
Swept Spectrum Analyzer
One way to avoid the need for such
a large number of expensive filters is
to use only one filter and sweep it
slowly through the frequency range
of interest. If, as in Figure 2.23, we
display the output of the filter versus
the frequency to which it is tuned,
we have the spectrum of the input
signal. This swept analysis technique
is commonly used in rf and
microwave spectrum analysis.
We have, however, assumed the input
signal hasn’t changed in the time it
takes to complete a sweep of our
analyzer. If energy appears at some
frequency at a moment when our
filter is not tuned to that frequency,
then we will not measure it.
One way to reduce this problem
would be to speed up the sweep
time of our analyzer. We could still
miss an event, but the time in which
this could happen would be shorter.
Unfortunately though, we cannot
make the sweep arbitrarily fast
because of the response time of
our filter.
To understand this problem,
recall from Section 2 that a filter
takes a finite time to respond to
changes in its input. The narrower the
filter, the longer it takes to respond.
If we sweep the filter past a signal
too quickly, the filter output will not
have a chance to respond fully to the
signal. As we show in Figure 2.24,
the spectrum display will then be in
error; our estimate of the signal level
will be too low.
In a parallel-filter spectrum analyzer
we do not have this problem. All the
filters are connected to the input
signal all the time. Once we have
waited the initial settling time of a
single filter, all the filters will be
settled and the spectrum will be valid
and not miss any transient events.
So there is a basic trade-off between
parallel-filter and swept spectrum
analyzers. The parallel-filter analyzer
is fast, but has limited resolution and
is expensive. The swept analyzer
can be cheaper and have higher
resolution but the measurement
takes longer (especially at high
resolution) and it can not analyze
transient events*.
Dynamic Signal Analyzer
In recent years another kind of
analyzer has been developed
which offers the best features of the
parallel-filter and swept spectrum
analyzers. Dynamic Signal Analyzers
are based on a high speed calculation
routine which acts like a parallel
filter analyzer with hundreds of
filters and yet are cost-competitive
with swept spectrum analyzers. In
* More information on the performance of swept
spectrum analyzers can be found in Agilent
Application Note Series 150.
Figure 2.24
Amplitude
error form
sweeping
too fast.
Figure 2.23
Simplified
swept spectrum
analyzer.

19
addition, two channel Dynamic Signal
Analyzers are in many ways better
network analyzers than the ones we
will introduce next.
Network Analyzers
Since in network analysis it is
required to measure both the input
and output, network analyzers are
generally two channel devices with
the capability of measuring the ampli-
tude ratio (gain or loss) and phase
difference between the channels.
All of the analyzers discussed here
measure frequency response by using
a sinusoidal input to the network
and slowly changing its frequency.
Dynamic Signal Analyzers use a
different, much faster technique for
network analysis which we discuss
in the next chapter.
Gain-phase meters are broadband
devices which measure the amplitude
and phase of the input and output
sine waves of the network. A sinu-
soidal source must be supplied to
stimulate the network when using a
gain-phase meter as in Figure 2.25.
The source can be tuned manually
and the gain-phase plots done by
hand or a sweeping source, and an
x-y plotter can be used for automatic
frequency response plots.
The primary attraction of gain-phase
meters is their low price. If a
sinusoidal source and a plotter are
already available, frequency response
measurements can be made for a very
low investment. However, because
gain-phase meters are broadband,
they measure all the noise of the
network as well as the desired sine
wave. As the network attenuates the
input, this noise eventually becomes a
floor below which the meter cannot
measure. This typically becomes a
problem with attenuations of about
60 dB (1,000:1).
Tuned network analyzers minimize
the noise floor problems of gain-
phase meters by including a bandpass
filter which tracks the source fre-
quency. Figure 2.26 shows how this
tracking filter virtually eliminates the
noise and any harmonics to allow
measurements of attenuation to
100 dB (100,000:1).
By minimizing the noise, it is also
possible for tuned network analyzers
to make more accurate measure-
ments of amplitude and phase. These
improvements do not come without
their price, however, as tracking
filters and a dedicated source must
be added to the simpler and less
costly gain-phase meter.
Figure 2.26
Tuned net-
work analyzer
operation.
Figure 2.25
Gain-phase
meter
operation.

20
Tuned analyzers are available in the
frequency range of a few Hertz to
many Gigahertz (109 Hertz). If lower
frequency analysis is desired, a
frequency response analyzer is often
used. To the operator, it behaves
exactly like a tuned network analyzer.
However, it is quite different inside.
It integrates the signals in the time
domain to effectively filter the signals
at very low frequencies where it is
not practical to make filters by more
conventional techniques. Frequency
response analyzers are generally lim-
ited to from 1 mHz to about 10 kHz.
Section 4:
The Modal Domain
In the preceding sections we have
developed the properties of the time
and frequency domains and the
instrumentation used in these
domains. In this section we will
develop the properties of another
domain, the modal domain. This
change in perspective to a new
domain is particularly useful if we are
interested in analyzing the behavior
of mechanical structures.
To understand the modal domain let
us begin by analyzing a simple
mechanical structure, a tuning fork.
If we strike a tuning fork, we easily
conclude from its tone that it is pri-
marily vibrating at a single frequency.
We see that we have excited a
network (tuning fork) with a force
impulse (hitting the fork). The time
domain view of the sound caused by
the deformation of the fork is a
lightly damped sine wave shown
in Figure 2.27b.
In Figure 2.27c, we see in the
frequency domain that the frequency
response of the tuning fork has a
major peak that is very lightly
damped, which is the tone we hear.
There are also several smaller peaks.
Figure 2.27
The vibration
of a tuning fork.
Figure 2.28
Example
vibration modes
of a tuning fork.

21
Each of these peaks, large and small,
corresponds to a “vibration mode”
of the tuning fork. For instance, we
might expect for this simple example
that the major tone is caused by the
vibration mode shown in Figure
2.28a. The second harmonic might
be caused by a vibration like
Figure 2.28b
We can express the vibration of any
structure as a sum of its vibration
modes. Just as we can represent an
real waveform as a sum of much sim-
pler sine waves, we can represent any
vibration as a sum of much simpler
vibration modes. The task of “modal”
analysis is to determine the shape
and the magnitude of the structural
deformation in each vibration mode.
Once these are known, it usually
becomes apparent how to change
the overall vibration.
For instance, let us look again at our
tuning fork example. Suppose that we
decided that the second harmonic
tone was too loud. How should we
change our tuning fork to reduce the
harmonic? If we had measured the
vibration of the fork and determined
that the modes of vibration were
those shown in Figure 2.28, the
answer becomes clear. We might
apply damping material at the center
of the tines of the fork. This would
greatly affect the second mode which
has maximum deflection at the center
while only slightly affecting the
desired vibration of the first mode.
Other solutions are possible, but all
depend on knowing the geometry of
each mode.
The Relationship Between the Time,
Frequency and Modal Domain
To determine the total vibration
of our tuning fork or any other
structure, we have to measure the
vibration at several points on the
structure. Figure 2.30a shows some
points we might pick. If we
transformed this time domain data to
the frequency domain, we would get
results like Figure 2.30b. We measure
frequency response because we want
to measure the properties of the
structure independent of the
stimulus*.
Figure 2.29
Reducing the
second harmonic
by damping the
second vibration
mode.
Figure 2.30
Modal analysis
of a tuning fork.
* Those who are more familiar with electronics might
note that we have measured the frequency response of
a network (structure) at N points and thus have performed
an N-port Analysis.

22
We see that the sharp peaks
(resonances) all occur at the same
frequencies independent of where
they are measured on the structure.
Likewise we would find by measuring
the width of each resonance that the
damping (or Q) of each resonance
is independent of position. The
only parameter that varies as we
move from point to point along the
structure is the relative height of
resonances.* By connecting the
peaks of the resonances of a given
mode, we trace out the mode shape
of that mode.
Experimentally we have to measure
only a few points on the structure to
determine the mode shape. However,
to clearly show the mode shape in
our figure, we have drawn in the
frequency response at many more
points in Figure 2.31a. If we view this
three-dimensional graph along the
distance axis, as in Figure 2.31b, we
get a combined frequency response.
Each resonance has a peak value cor-
responding to the peak displacement
in that mode. If we view the graph
along the frequency axis, as in Figure
2.31c, we can see the mode shapes of
the structure.
We have not lost any information by
this change of perspective. Each
vibration mode is characterized by its
mode shape, frequency and damping
from which we can reconstruct the
frequency domain view.
However, the equivalence between
the modal, time and frequency
domains is not quite as strong as
that between the time and frequency
domains. Because the modal domain
portrays the properties of the net-
work independent of the stimulus,
transforming back to the time domain
gives the impulse response of the
structure, no matter what the stimu-
lus. A more important limitation of
this equivalence is that curve fitting
is used in transforming from our
frequency response measurements to
the modal domain to minimize the
effects of noise and small experimen-
tal errors. No information is lost in
this curve fitting, so all three domains
contain the same information, but not
the same noise. Therefore, transform-
ing from the frequency domain to the
modal domain and back again will
give results like those in Figure 2.32.
The results are not exactly the same,
yet in all the important features, the
frequency responses are the same.
This is also true of time domain data
derived from the modal domain.
Figure 2.31
The relationship
between the
frequency and
the modal
domains.
* The phase of each resonance is not shown for clarity of
the figures but it too is important in the mode shape. The
magnitude of the frequency response gives the magnitude
of the mode shape while the phase gives the direction of
the deflection.

23
Section 5:
Instrumentation for
the Modal Domain
There are many ways that the modes
of vibration can be determined. In our
simple tuning fork example we could
guess what the modes were. In simple
structures like drums and plates it is
possible to write an equation for the
modes of vibration. However, in
almost any real problem, the solution
can neither be guessed nor solved
analytically because the structure is
too complicated. In these cases it is
necessary to measure the response
of the structure and determine
the modes.
There are two basic techniques for
determining the modes of vibration in
complicated structures: 1) exciting
only one mode at a time, and 2)
computing the modes of vibration
from the total vibration.
Single Mode Excitation
Modal Analysis
To illustrate single mode excitation,
let us look once again at our simple
tuning fork example. To excite just
the first mode we need two shakers,
driven by a sine wave and attached
to the ends of the tines as in Figure
2.33a. Varying the frequency of the
generator near the first mode reso-
nance frequency would then give us
its frequency, damping and mode
shape.
In the second mode, the ends of the
tines do not move, so to excite the
second mode we must move the
shakers to the center of the tines. If
we anchor the ends of the tines, we
will constrain the vibration to the
second mode alone.
Figure 2.32
Curve fitting
removes
measurement
noise.
Figure 2.33
Single mode
excitation
modal analysis.

24
In more realistic, three dimensional
problems, it is necessary to add many
more shakers to ensure that only one
mode is excited. The difficulties and
expense of testing with many shakers
has limited the application of this
traditional modal analysis technique.
Modal Analysis From Total Vibration
To determine the modes of vibration
from the total vibration of the
structure, we use the techniques
developed in the previous section.
Basically, we determine the frequency
response of the structure at several
points and compute at each reso-
nance the frequency, damping and
what is called the residue (which
represents the height of the reso-
nance). This is done by a curve-fitting
routine to smooth out any noise or
small experimental errors. From
these measurements and the geome-
try of the structure, the mode shapes
are computed and drawn on a CRT
display or a plotter. If drawn on a
CRT, these displays may be animated
to help the user understand the
vibration mode.
From the above description, it is
apparent that a modal analyzer
requires some type of network
analyzer to measure the frequency
response of the structure and a
computer to convert the frequency
response to mode shapes. This can
be accomplished by connecting a
Dynamic Signal Analyzer through
a digital interface* to a computer
furnished with the appropriate soft-
ware. This capability is also available
in a single instrument called a Struc-
tural Dynamics Analyzer. In general,
computer systems offer more versa-
tile performance since they can be
programmed to solve other problems.
However, Structural Dynamics
Analyzers generally are much easier
to use than computer systems.
Section 6: Summary
In this chapter we have developed
the concept of looking at problems
from different perspectives. These
perspectives are the time, frequency
and modal domains. Phenomena that
are confusing in the time domain are
often clarified by changing perspec-
tive to another domain. Small signals
are easily resolved in the presence of
large ones in the frequency domain.
The frequency domain is also valu-
able for predicting the output of any
kind of linear network. A change to
the modal domain breaks down
complicated structural vibration
problems into simple vibration
modes.
No one domain is always the best
answer, so the ability to easily change
domains is quite valuable. Of all the
instrumentation available today, only
Dynamic Signal Analyzers can work
in all three domains. In the next
chapter we develop the properties
of this important class of analyzers.
Figure 2.34
Measured
mode shape.
* GPIB, Agilent’s implementation of
IEEE-488-1975 is ideal for this application.

25
We saw in the previous chapter that
the Dynamic Signal Analyzer has the
speed advantages of parallel-filter
analyzers without their low resolution
limitations. In addition, it is the only
type of analyzer that works in all
three domains. In this chapter we will
develop a fuller understanding of this
important analyzer family, Dynamic
Signal Analyzers. We begin by pre-
senting the properties of the Fast
Fourier Transform (FFT) upon which
Dynamic Signal Analyzers are based.
No proof of these properties is given,
but heuristic arguments as to their va-
lidity are used where appropriate. We
then show how these FFT properties
cause some undesirable characteris-
tics in spectrum analysis like aliasing
and leakage. Having demonstrated a
potential difficulty with the FFT, we
then show what solutions are used
to make practical Dynamic Signal
Analyzers. Developing this basic
knowledge of FFT characteristics
makes it simple to get good results
with a Dynamic Signal Analyzer in a
wide range of measurement problems.
Section 1: FFT Properties
The Fast Fourier Transform (FFT)
is an algorithm* for transforming
data from the time domain to the fre-
quency domain. Since this is exactly
what we want a spectrum analyzer to
do, it would seem easy to implement
a Dynamic Signal Analyzer based
on the FFT. However, we will see
that there are many factors which
complicate this seemingly
straightforward task.
First, because of the many calcula-
tions involved in transforming
domains, the transform must be
implemented on a digital computer if
the results are to be sufficiently accu-
rate. Fortunately, with the advent of
microprocessors, it is easy and inex-
pensive to incorporate all the needed
computing power in a small instru-
ment package. Note, however, that
we cannot now transform to the
frequency domain in a continuous
manner, but instead must sample and
digitize the time domain input. This
means that our algorithm transforms
digitized samples from the time do-
main to samples in the frequency
domain as shown in Figure 3.1.**
Because we have sampled, we no
longer have an exact representation
in either domain. However, a sampled
representation can be as close to
ideal as we desire by placing our
samples closer together. Later in
this chapter, we will consider what
sample spacing is necessary to
guarantee accurate results.
Chapter 3
Understanding Dynamic
Signal Analysis
Figure 3.1
The FFT samples
in both the time
and frequency
domains.
Figure 3.2
A time record
is N equally
spaced samples
of the input.
* An algorithm is any special mathematical method of
solving a certain kind of problem; e.g., the technique
you use to balance your checkbook.
** To reduce confusion about which domain we are in,
samples in the frequency domain are called lines.

26
Time Records
A time record is defined to be N
consecutive, equally spaced samples
of the input. Because it makes our
transform algorithm simpler and
much faster, N is restricted to be
a multiple of 2, for instance 1024.
As shown in Figure 3.3, this time
record is transformed as a complete
block into a complete block of
frequency lines. All the samples of
the time record are needed to
compute each and every line in the
frequency domain. This is in contrast
to what one might expect, namely
that a single time domain sample
transforms to exactly one frequency
domain line. Understanding this block
processing property of the FFT is
crucial to understanding many of
the properties of the Dynamic
Signal Analyzer.
For instance, because the FFT
transforms the entire time record
block as a total, there cannot be
valid frequency domain results until
a complete time record has been
gathered. However, once completed,
the oldest sample could be discarded,
all the samples shifted in the time
record, and a new sample added to
the end of the time record as in
Figure 3.4. Thus, once the time record
is initially filled, we have a new time
record at every time domain sample
and therefore could have new valid
results in the frequency domain at
every time domain sample.
This is very similar to the behavior of
the parallel-filter analyzers described
in the previous chapter. When a signal
is first applied to a parallel-filter ana-
lyzer, we must wait for the filters to
respond, then we can see very rapid
changes in the frequency domain.
With a Dynamic Signal Analyzer we
do not get a valid result until a full
time record has been gathered. Then
rapid changes in the spectra can
be seen.
It should be noted here that a new
spectrum every sample is usually too
much information, too fast. This
would often give you thousands of
transforms per second. Just how fast
a Dynamic Signal Analyzer should
transform is a subject better left to
the sections in this chapter on real
time bandwidth and overlap
processing.
Figure 3.3
The FFT works
on blocks
of data.
Figure 3.4
A new time
record every
sample after
the time record
is filled.

27
How Many Lines are There?
We stated earlier that the time
record has N equally spaced samples.
Another property of the FFT is that
it transforms these time domain
samples to N/2 equally spaced lines
in the frequency domain. We only
get half as many lines because each
frequency line actually contains two
pieces of information, amplitude and
phase. The meaning of this is most
easily seen if we look again at the
relationship between the time and
frequency domain.
Figure 3.5 reproduces from Chapter 2
our three-dimensional graph of this
relationship. Up to now we have
implied that the amplitude and
frequency of the sine waves contains
all the information necessary to
reconstruct the input. But it should
be obvious that the phase of each
of these sine waves is important too.
For instance, in Figure 3.6, we have
shifted the phase of the higher
frequency sine wave components
of this signal. The result is a severe
distortion of the original wave form.
We have not discussed the phase
information contained in the
spectrum of signals until now
because none of the traditional
spectrum analyzers are capable of
measuring phase. When we discuss
measurements in Chapter 4, we shall
find that phase contains valuable
information in determining the
cause of performance problems.
What is the Spacing of the Lines?
Now that we know that we have N/2
equally spaced lines in the frequency
domain, what is their spacing? The
lowest frequency that we can resolve
with our FFT spectrum analyzer must
be based on the length of the time
record. We can see in Figure 3.7 that
if the period of the input signal is
longer than the time record, we have
no way of determining the period (or
frequency, its reciprocal). Therefore,
the lowest frequency line of the FFT
must occur at frequency equal to the
reciprocal of the time record length.
Figure 3.5
The relationship
between the time
and frequency
domains.
Figure 3.6
Phase of
frequency domain
components is
important.

28
In addition, there is a frequency line
at zero Hertz, DC. This is merely the
average of the input over the time
record. It is rarely used in spectrum
or network analysis. But, we have
now established the spacing between
these two lines and hence every line;
it is the reciprocal of the time record.
What is the Frequency Range
of the FFT?
We can now quickly determine that
the highest frequency we can
measure is:
fmax = q
because we have N/2 lines spaced by
the reciprocal of the time record
starting at zero Hertz *.
Since we would like to adjust the fre-
quency range of our measurement,
we must vary fmax. The number of
time samples N is fixed by the imple-
mentation of the FFT algorithm.
Therefore, we must vary the period of
the time record to vary fmax. To do
this, we must vary the sample rate so
that we always have N samples in our
variable time record period. This is
illustrated in Figure 3.9. Notice that
to cover higher frequencies, we must
sample faster.
* The usefulness of this frequency range can be limited by
the problem of aliasing. Aliasing is discussed in Section 3.
Figure 3.7
Lowest frequency
resolvable by
the FFT.
Time
Time
AmplitudeAmplitude
Period of input signal longer than the time record.
Frequency of the input signal is unknown..
b)
Time Record
Time Record
Period of input signal equals time record.
Lowest resolvable frequency.
a)
??
Figure 3.8
Frequencies of
all the spectral
lines of the FFT.
Figure 3.9
Frequency range
of Dynamic Signal
Analyzers is
determined by
sample rate.
N 1
2 Period of Time Record

29
Section 2*:
Sampling and Digitizing
Recall that the input to our Dynamic
Signal Analyzer is a continuous
analog voltage. This voltage might
be from an electronic circuit or could
be the output of a transducer and be
proportional to current, power,
pressure, acceleration or any number
of other inputs. Recall also that the
FFT requires digitized samples of the
input for its digital calculations.
Therefore, we need to add a sampler
and analog to digital converter (ADC)
to our FFT processor to make a spec-
trum analyzer. We show this basic
block diagram in Figure 3.10.
For the analyzer to have the high
accuracy needed for many measure-
ments, the sampler and ADC must be
quite good. The sampler must sample
the input at exactly the correct time
and must accurately hold the input
voltage measured at this time until
the ADC has finished its conversion.
The ADC must have high resolution
and linearity. For 70 dB of dynamic
range the ADC must have at least
12 bits of resolution and one half
least significant bit linearity.
A good Digital Voltmeter (DVM) will
typically exceed these specifications,
but the ADC for a Dynamic Signal
Analyzer must be much faster than
typical fast DVM’s. A fast DVM might
take a thousand readings per second,
but in a typical Dynamic Signal
Analyzer the ADC must take at
least a hundred thousand readings
per second.
Section 3: Aliasing
The reason an FFT spectrum
analyzer needs so many samples per
second is to avoid a problem called
aliasing. Aliasing is a potential prob-
lem in any sampled data system. It is
often overlooked, sometimes with
disastrous results.
A Simple Data Logging
Example of Aliasing
Let us look at a simple data logging
example to see what aliasing is and
how it can be avoided. Consider the
example for recording temperature
shown in Figure 3.12. A thermocouple
is connected to a digital voltmeter
which is in turn connected to a print-
er. The system is set up to print the
temperature every second. What
would we expect for an output?
If we were measuring the tempera-
ture of a room which only changes
slowly, we would expect every
reading to be almost the same as the
previous one. In fact, we are sampling
much more often than necessary to
determine the temperature of the
room with time. If we plotted the
results of this “thought experiment”,
we would expect to see results like
Figure 3.13.
Figure 3.10
Block diagram
of dynamic
Signal Analyzer.
Figure 3.11
The Sampler
and ADC must
not introduce
errors.
Figure 3.13
Plot of
temperature
variation
of a room.
Figure 3.12
A simple
sampled
data system.
* This section and the next can be skipped by those not
interested in the internal operation of a Dynamic Signal
Analyzer. However, those who specify the purchase of
Dynamic Signal Analyzers are especially encouraged to
read these sections. The basic knowledge to be gained
from these sections can insure specifying the best analyzer
for your requirements.

30
The Case of the
Missing Temperature
If, on the other hand, we were
measuring the temperature of a
small part which could heat and cool
rapidly, what would the output be?
Suppose that the temperature of
our part cycled exactly once every
second. As shown in Figure 3.14, our
printout says that the temperature
never changes.
What has happened is that we have
sampled at exactly the same point on
our periodic temperature cycle with
every sample. We have not sampled
fast enough to see the temperature
fluctuations.
Aliasing in the Frequency Domain
This completely erroneous result is
due to a phenomena called aliasing.*
Aliasing is shown in the frequency
domain in Figure 3.15. Two signals
are said to alias if the difference of
their frequencies falls in the frequen-
cy range of interest. This difference
frequency is always generated in the
process of sampling. In Figure 3.15,
the input frequency is slightly higher
than the sampling frequency so a low
frequency alias term is generated. If
the input frequency equals the sam-
pling frequency as in our small part
example, then the alias term falls at
DC (zero Hertz) and we get the
constant output that we saw above.
Aliasing is not always bad. It is
called mixing or heterodyning in
analog electronics, and is commonly
used for tuning household radios and
televisions as well as many other
communication products. However,
in the case of the missing tempera-
ture variation of our small part, we
definitely have a problem. How can
we guarantee that we will avoid this
problem in a measurement situation?
Figure 3.16 shows that if we sample
at greater than twice the highest
frequency of our input, the alias
products will not fall within the
frequency range of our input.
Therefore, a filter (or our FFT
processor which acts like a filter)
after the sampler will remove the
alias products while passing the
desired input signals if the sample
rate is greater than twice the highest
frequency of the input. If the sample
rate is lower, the alias products will
fall in the frequency range of the
input and no amount of filtering
will be able to remove them from
the signal.
Figure 3.14
Plot of temperature
variation of a
small part.
Figure 3.15
The problem
of aliasing
viewed in the
frequency
domain.
* Aliasing is also known as fold-over or mixing.

31
This minimum sample rate
requirement is known as the Nyquist
Criterion. It is easy to see in the time
domain that a sampling frequency
exactly twice the input frequency
would not always be enough. It is less
obvious that slightly more than two
samples in each period is sufficient
information. It certainly would not
be enough to give a high quality time
display. Yet we saw in Figure 3.16 that
meeting the Nyquist Criterion of a
sample rate greater than twice the
maximum input frequency is suffi-
cient to avoid aliasing and preserve
all the information in the input signal.
The Need for an Anti-Alias Filter
Unfortunately, the real world rarely
restricts the frequency range of its
signals. In the case of the room
temperature, we can be reasonably
sure of the maximum rate at which
the temperature could change, but
we still can not rule out stray signals.
Signals induced at the powerline
frequency or even local radio stations
could alias into the desired frequency
range. The only way to be really
certain that the input frequency range
is limited is to add a low pass filter
before the sampler and ADC. Such a
filter is called an anti-alias filter.
An ideal anti-alias filter would look
like Figure 3.18a. It would pass all
the desired input frequencies with no
loss and completely reject any higher
frequencies which otherwise could
alias into the input frequency range.
However, it is not even theoretically
possible to build such a filter, much
less practical. Instead, all real filters
look something like Figure 3.18b with
a gradual roll off and finite rejection
of undesired signals. Large input
signals which are not well attenuated
in the transition band could still alias
into the desired input frequency
Figure 3.16
A frequency
domain view
of how to avoid
aliasing - sample
at greater than
twice the highest
input frequency.
Figure 3.18
Actual anti-alias
filters require
higher sampling
frequencies.
Figure 3.17
Nyquist
Criterion in the
time domain.

32
range. To avoid this, the sampling fre-
quency is raised to twice the highest
frequency of the transition band. This
guarantees that any signals which
could alias are well attentuated by
the stop band of the filter. Typically,
this means that the sample rate is
now two and a half to four times the
maximum desired input frequency.
Therefore, a 25 kHz FFT Spectrum
Analyzer can require an ADC that
runs at 100 kHz as we stated without
proof in Section 2 of this Chapter*.
The Need for More Than One
Anti-Alias Filter
Recall from Section 1 of this Chapter,
that due to the properties of the FFT
we must vary the sample rate to vary
the frequency span of our analyzer.
To reduce the frequency span, we
must reduce the sample rate. From
our considerations of aliasing, we
now realize that we must also reduce
the anti-alias filter frequency by the
same amount.
Since a Dynamic Signal Analyzer is
a very versatile instrument used in
a wide range of applications, it is
desirable to have a wide range of
frequency spans available. Typical
instruments have a minimum span of
1 Hertz and a maximum of tens to
hundreds of kilohertz. This four
decade range typically needs to be
covered with at least three spans per
decade. This would mean at least
twelve anti-alias filters would be
required for each channel.
Each of these filters must have
very good performance. It is desirable
that their transition bands be as
narrow as possible so that as many
lines as possible are free from alias
products. Additionally, in a two
channel analyzer, each filter pair
must be well matched for accurate
network analysis measurements.
These two points unfortunately mean
that each of the filters is expensive.
Taken together they can add signifi-
cantly to the price of the analyzer.
Some manufacturers don’t have a
low enough frequency anti-alias filter
on the lowest frequency spans to save
some of this expense. (The lowest
frequency filters cost the most of all.)
But as we have seen, this can lead to
problems like our “case of the
missing temperature”.
Digital Filtering
Fortunately, there is an alternative
which is cheaper and when used in
conjunction with a single analog anti-
alias filter, always provides aliasing
protection. It is called digital filtering
because it filters the input signal after
we have sampled and digitized it. To
see how this works, let us look at
Figure 3.19.
In the analog case we already
discussed, we had to use a new
filter every time we changed the
sample rate of the Analog to Digital
Converter (ADC). When using digital
filtering, the ADC sample rate is left
constant at the rate needed for the
highest frequency span of the analyz-
er. This means we need not change
our anti-alias filter. To get the
reduced sample rate and filtering
we need for the narrower frequency
spans, we follow the ADC with a
digital filter.
This digital filter is known as a
decimating filter. It not only filters
the digital representation of the signal
to the desired frequency span, it also
reduces the sample rate at its output
to the rate needed for that frequency
span. Because this filter is digital,
there are no manufacturing varia-
tions, aging or drift in the filter.
Therefore, in a two channel analyzer
the filters in each channel are identi-
cal. It is easy to design a single digital
filter to work on many frequency
spans so the need for multiple filters
per channel is avoided. All these
factors taken together mean that
digital filtering is much less expen-
sive than analog anti-aliasing filtering.
Figure 3.19
Block diagrams
of analog and
digital filtering.
* Unfortunately, because the spacing of the FFT lines
depends on the sample rate, increasing the sample rate
decreases the number of lines that are in the desired
frequency range. Therefore, to avoid aliasing problems
Dynamic Signal Analyzers have only .25N to .4N lines
instead of N/2 lines.

33
Section 4:
Band Selectable Analysis
Suppose we need to measure a small
signal that is very close in frequency
to a large one. We might be measur-
ing the powerline sidebands (50 or
60 Hz) on a 20 kHz oscillator. Or we
might want to distinguish between
the stator vibration and the shaft
imbalance in the spectrum of a
motor.*
Recall from our discussion of
the properties of the Fast Fourier
Transform that it is equivalent to a
set of filters, starting at zero Hertz,
equally spaced up to some maximum
frequency. Therefore, our frequency
resolution is limited to the maximum
frequency divided by the number
of filters.
To just resolve the 60 Hz sidebands
on a 20 kHz oscillator signal would
require 333 lines (or filters) of the
FFT. Two or three times more lines
would be required to accurately
measure the sidebands. But typical
Dynamic Signal Analyzers only have
200 to 400 lines, not enough for
accurate measurements. To increase
the number of lines would greatly
increase the cost of the analyzer. If
we chose to pay the extra cost, we
would still have trouble seeing the
results. With a 4 inch (10 cm) screen,
the sidebands would be only 0.01 inch
(.25 mm) from the carrier.
A better way to solve this problem
is to concentrate the filters into the
frequency range of interest as in
Figure 3.20. If we select the minimum
frequency as well as the maximum
frequency of our filters we can “zoom
in” for a high resolution close-up shot
of our frequency spectrum. We now
have the capability of looking at the
entire spectrum at once with low
resolution as well as the ability to
look at what interests us with much
higher resolution.
This capability of increased
resolution is called Band Selectable
Analysis (BSA).** It is done by mixing
or heterodyning the input signal
down into the range of the FFT span
selected. This technique, familiar to
electronic engineers, is the process
by which radios and televisions
tune in stations.
The primary difference between the
implementation of BSA in Dynamic
Signal Analyzers and heterodyne
radios is shown in Figure 3.21. In a
radio, the sine wave used for mixing
is an analog voltage. In a Dynamic
Signal Analyzer, the mixing is done
after the input has been digitized, so
the “sine wave” is a series of digital
numbers into a digital multiplier.
This means that the mixing will be
done with a very accurate and stable
digital signal so our high resolution
display will likewise be very stable
and accurate.
* The shaft of an ac induction motor always runs at a rate
slightly lower than a multiple of the driven frequency, an
effect called slippage.
** Also sometimes called “zoom”.
Figure 3.20
High resolution
measurements
with Band
Selectable
Analysis.
Figure 3.21
Analyzer block
diagram.

34
Section 5: Windowing
The Need for Windowing
There is another property of the Fast
Fourier Transform which affects its
use in frequency domain analysis.
We recall that the FFT computes the
frequency spectrum from a block of
samples of the input called a time
record. In addition, the FFT algorithm
is based upon the assumption that
this time record is repeated through-
out time as illustrated in Figure 3.22.
This does not cause a problem with
the transient case shown. But what
happens if we are measuring a contin-
uous signal like a sine wave? If the
time record contains an integral
number of cycles of the input sine
wave, then this assumption exactly
matches the actual input waveform
as shown in Figure 3.23. In this case,
the input waveform is said to be
periodic in the time record.
Figure 3.24 demonstrates the
difficulty with this assumption
when the input is not periodic in the
time record. The FFT algorithm is
computed on the basis of the highly
distorted waveform in Figure 3.24c.
We know from Chapter 2 that
the actual sine wave input has a
frequency spectrum of single line.
The spectrum of the input assumed
by the FFT in Figure 3.24c should be
Figure 3.24
Input signal
not periodic
in time record.
Figure 3.22
FFT assumption -
time record
repeated
throughout
all time.
Figure 3.23
Input signal
periodic in time
record.

35
very different. Since sharp phenome-
na in one domain are spread out in
the other domain, we would expect
the spectrum of our sine wave to be
spread out through the frequency
domain.
In Figure 3.25 we see in an actual
measurement that our expectations
are correct. In Figures 3.25 a & b, we
see a sine wave that is periodic in the
time record. Its frequency spectrum
is a single line whose width is deter-
mined only by the resolution of our
Dynamic Signal Analyzer.* On the
other hand, Figures 3.25c & d show
a sine wave that is not periodic in
the time record. Its power has been
spread throughout the spectrum as
we predicted.
This smearing of energy throughout
the frequency domains is a phenome-
na known as leakage. We are seeing
energy leak out of one resolution line
of the FFT into all the other lines.
It is important to realize that leakage
is due to the fact that we have taken
a finite time record. For a sine wave
to have a single line spectrum, it must
exist for all time, from minus infinity
to plus infinity. If we were to have
an infinite time record, the FFT
would compute the correct single
line spectrum exactly. However, since
we are not willing to wait forever to
measure its spectrum, we only look
at a finite time record of the sine
wave. This can cause leakage if the
continuous input is not periodic in
the time record.
It is obvious from Figure 3.25 that the
problem of leakage is severe enough
to entirely mask small signals close
to our sine waves. As such, the FFT
would not be a very useful spectrum
analyzer. The solution to this problem
is known as windowing. The prob-
lems of leakage and how to solve
them with windowing can be the
most confusing concepts of Dynamic
Signal Analysis. Therefore, we will
now carefully develop the problem
and its solution in several representa-
tive cases.
* The additional two components in the photo are the
harmonic distortion of the sine wave source.
Figure 3.25
Actual FFT results.
b)
a) & b) Sine wave periodic in time record
d)
c) & d) Sine wave not periodic in time record
a)
c)

36
What is Windowing?
In Figure 3.26 we have again repro-
duced the assumed input wave form
of a sine wave that is not periodic in
the time record. Notice that most of
the problem seems to be at the edges
of the time record, the center is a
good sine wave. If the FFT could
be made to ignore the ends and con-
centrate on the middle of the time
record, we would expect to get much
closer to the correct single line
spectrum in the frequency domain.
If we multiply our time record by
a function that is zero at the ends
of the time record and large in the
middle, we would concentrate the
FFT on the middle of the time record.
One such function is shown in Figure
3.26c. Such functions are called
window functions because they
force us to look at data through a
narrow window.
Figure 3.27 shows us the vast
improvement we get by windowing
data that is not periodic in the time
record. However, it is important to
realize that we have tampered with
the input data and cannot expect
perfect results. The FFT assumes the
input looks like Figure 3.26d, some-
thing like an amplitude-modulated
sine wave. This has a frequency
spectrum which is closer to the
correct single line of the input sine
wave than Figure 3.26b, but it still is
not correct. Figure 3.28 demonstrates
that the windowed data does not
have as narrow a spectrum as an
unwindowed function which is
periodic in the time record.
Figure 3.26
The effect of
windowing in
the time domain.
Figure 3.27
Leakage reduction
with windowing.
a) Sine wave not periodic in time record b) FFT results with no window function
c) FFT results with a window function

37
The Hanning Window
Any number of functions can be used
to window the data, but the most
common one is called Hanning. We
actually used the Hanning window in
Figure 3.27 as our example of leakage
reduction with windowing. The
Hanning window is also commonly
used when measuring random noise.
The Uniform Window*
We have seen that the Hanning
window does an acceptably good
job on our sine wave examples, both
periodic and non-periodic in the time
record. If this is true, why should we
want any other windows?
Suppose that instead of wanting the
frequency spectrum of a continuous
signal, we would like the spectrum
of a transient event. A typical tran-
sient is shown in Figure 3.29a. If we
multiplied it by the window function
in Figure 3.29b we would get the
highly distorted signal shown in
Figure 3.29c. The frequency spectrum
of an actual transient with and with-
out the Hanning window is shown in
Figure 3.30. The Hanning window has
taken our transient, which naturally
has energy spread widely through the
frequency domain and made it look
more like a sine wave.
Therefore, we can see that for
transients we do not want to use
the Hanning window. We would like
to use all the data in the time record
equally or uniformly. Hence we will
use the Uniform window which
weights all of the time record
uniformly.
The case we made for the Uniform
window by looking at transients
can be generalized. Notice that our
transient has the property that it is
zero at the beginning and end of the
time record. Remember that we intro-
duced windowing to force the input
to be zero at the ends of the time
record. In this case, there is no need
for windowing the input. Any func-
tion like this which does not require a
window because it occurs completely
within the time record is called a self-
windowing function. Self-windowing
functions generate no leakage in the
FFT and so need no window.
* The Uniform Window is sometimes referred to as a
“Rectangular Window”.
Figure 3.28
Windowing reduces
leakage but does
not eliminate it.
b) Windowed measurement - input not periodic
in time record
a) Leakage-free measurement - input periodic
in time record
Figure 3.29
Windowing loses
information from
transient events.
Figure 3.30
Spectrums
of transients.
b) Hanning windowed transientsa) Unwindowed trainsients

38
There are many examples of self-
windowing functions, some of which
are shown in Figure 3.31. Impacts,
impulses, shock responses, sine
bursts, noise bursts, chirp bursts and
pseudo-random noise can all be made
to be self-windowing. Self-windowing
functions are often used as the exci-
tation in measuring the frequency
response of networks, particularly
if the network has lightly-damped
resonances (high Q). This is because
the self-windowing functions gener-
ate no leakage in the FFT. Recall that
even with the Hanning window, some
leakage was present when the signal
was not periodic in the time record.
This means that without a self-win-
dowing excitation, energy could leak
from a lightly damped resonance into
adjacent lines (filters). The resulting
spectrum would show greater
damping than actually exists.*
The Flat-top Window
We have shown that we need a
uniform window for analyzing self-
windowing functions like transients.
In addition, we need a Hanning
window for measuring noise and
periodic signals like sine waves.
We now need to introduce a third
window function, the flat-top window,
to avoid a subtle effect of the
Hanning window. To understand
this effect, we need to look at the
Hanning window in the frequency
domain. We recall that the FFT acts
like a set of parallel filters. Figure
3.32 shows the shape of those filters
when the Hanning window is used.
Notice that the Hanning function
gives the filter a very rounded top.
If a component of the input signal
is centered in the filter it will be
measured accurately**. Otherwise,
the filter shape will attenuate the
component by up to 1.5 dB (16%)
when it falls midway between the
filters.
This error is unacceptably large
if we are trying to measure a signal’s
amplitude accurately. The solution is
to choose a window function which
gives the filter a flatter passband.
Such a flat-top passband shape is
shown in Figure 3.33. The amplitude
error from this window function does
not exceed .1 dB (1%), a 1.4 dB
improvement.
Figure 3.33
Flat-top
passband
shapes.
* There is another way to avoid this problem using Band
Selectable Analysis. We will illustrate this in the next
chapter.
** It will, in fact, be periodic in the time record
Figure 3.31
Self-windowing
function examples.
Figure 3.32
Hanning
passband
shapes.
Figure 3.34
Reduced
resolution
of the flat-top
window.
Flat-top
Hanning

39
The accuracy improvement does
not come without its price, however.
Figure 3.34 shows that we have flat-
tened the top of the passband at the
expense of widening the skirts of the
filter. We therefore lose some ability
to resolve a small component, closely
spaced to a large one. Some Dynamic
Signal Analyzers offer both Hanning
and flat-top window functions so that
the operator can choose between
increased accuracy or improved
frequency resolution.
Other Window Functions
Many other window functions are
possible but the three listed above
are by far the most common for
general measurements. For special
measurement situations other groups
of window functions may be useful.
We will discuss two windows which
are particularly useful when doing
network analysis on mechanical
structures by impact testing.
The Force and Response Windows
A hammer equipped with a force
transducer is commonly used to
stimulate a structure for response
measurements. Typically the force
input is connected to one channel
of the analyzer and the response of
the structure from another transducer
is connected to the second channel.
This force impact is obviously a
self-windowing function. The
response of the structure is also
self-windowing if it dies out within
the time record of the analyzer. To
guarantee that the response does go
to zero by the end of the time record,
an exponential-weighted window
called a response window is some-
times added. Figure 3.35 shows a
response window acting on the re-
sponse of a lightly damped structure
which did not fully decay by the end
of the time record. Notice that unlike
the Hanning window, the response
window is not zero at both ends of
the time record. We know that the
response of the structure will be zero
at the beginning of the time record
(before the hammer blow) so there
is no need for the window function
to be zero there. In addition, most of
the information about the structural
response is contained at the begin-
ning of the time record so we make
sure that this is weighted most heavi-
ly by our response window function.
The time record of the exciting force
should be just the impact with the
structure. However, movement of the
hammer before and after hitting the
structure can cause stray signals in
the time record. One way to avoid
this is to use a force window shown
in Figure 3.36. The force window is
unity where the impact data is valid
and zero everywhere else so that the
analyzer does not measure any stray
noise that might be present.
Passband Shapes or
Window Functions?
In the proceeding discussion we
sometimes talked about window
functions in the time domain. At
other times we talked about the filter
passband shape in the frequency
domain caused by these windows. We
change our perspective freely to
whichever domain yields the simplest
explanation. Likewise, some Dynamic
Signal Analyzers call the uniform,
Hanning and flat-top functions “win-
dows” and other analyzers call those
Figure 3.36
Using the
force window.
Figure 3.35
Using the
response
window.

40
functions “pass-band shapes”. Use
whichever terminology is easier
for the problem at hand as they are
completely interchangeable, just as
the time and frequency domains are
completely equivalent.
Section 6:
Network Stimulus
Recall from Chapter 2 that we can
measure the frequency response at
one frequency by stimulating the
network with a single sine wave and
measuring the gain and phase shift at
that frequency. The frequency of the
stimulus is then changed and the
measurement repeated until all
desired frequencies have been
measured. Every time the frequency
is changed, the network response
must settle to its steady-state value
before a new measurement can be
taken, making this measurement
process a slow task.
Many network analyzers operate in
this manner and we can make the
measurement this way with a two
channel Dynamic Signal Analyzer. We
set the sine wave source to the center
of the first filter as in Figure 3.37.
The analyzer then measures the
gain and phase of the network at
this frequency while the rest of the
analyzer’s filters measure only noise.
We then increase the source frequen-
cy to the next filter center, wait for
the network to settle and then meas-
ure the gain and phase. We continue
this procedure until we have
measured the gain and phase of
the network at all the frequencies
of the filters in our analyzer.
This procedure would, within
experimental error, give us the same
results as we would get with any of
the network analyzers described in
Chapter 2 with any network, linear
or nonlinear.
Noise as a Stimulus
A single sine wave stimulus does
not take advantage of the possible
speed the parallel filters of a
Dynamic Signal Analyzer provide. If
we had a source that put out multiple
sine waves, each one centered in a
filter, then we could measure the
frequency response at all frequencies
at one time. Such a source, shown in
Figure 3.38, acts like hundreds of sine
wave generators connected together.
Although this sounds very expensive,
Figure 3.37
Frequency
response
measurements
with a sine
wave stimulus.
Figure 3.38
Pseudo-random
noise as a
stimulus.

41
just such a source can be easily
generated digitally. It is called a
pseudo-random noise or periodic
random noise source.
From the names used for this source
it is apparent that it acts somewhat
like a true noise generator, except
that it has periodicity. If we add
together a large number of sine
waves, the result is very much like
white noise. A good analogy is the
sound of rain. A single drop of water
makes a quite distinctive splashing
sound, but a rain storm sounds like
white noise. However, if we add
together a large number of sine
waves, our noise-like signal will
periodically repeat its sequence.
Hence, the name periodic random
noise (PRN) source.
A truly random noise source has a
spectrum shown in Figure 3.39. It is
apparent that a random noise source
would also stimulate all the filters at
one time and so could be used as a
network stimulus. Which is a better
stimulus? The answer depends upon
the measurement situation.
Linear Network Analysis
If the network is reasonably linear,
PRN and random noise both give the
same results as the swept-sine test of
other analyzers. But PRN gives the
frequency response much faster. PRN
can be used to measure the frequency
response in a single time record.
Because the random source is true
noise, it must be averaged for several
time records before an accurate fre-
quency response can be determined.
Therefore, PRN is the best stimulus
to use with fairly linear networks
because it gives the fastest results*.
Non-Linear Network Analysis
If the network is severely non-linear,
the situation is quite different. In this
case, PRN is a very poor test signal
and random noise is much better. To
see why, let us look at just two of the
sine waves that compose the PRN
source. We see in Figure 3.40 that
if two sine waves are put through
a nonlinear network, distortion
products will be generated equally
spaced from the signals**. Unfortu-
nately, these products will fall exactly
on the frequencies of the other sine
waves in the PRN. So the distortion
products add to the output and there-
fore interfere with the measurement
Figure 3.39
Random noise
as a stimulus.
Figure 3.40
Pseudo-random
noise distortion.
∆
∆
∆∆
* There is another reason why PRN is a better test signal
than random or linear networks. Recall from the last
section that PRN is self-windowing. This means that
unlike random noise, pseudo-random noise has no leakage.
Therefore, with PRN, we can measure lightly damped (high
Q) resonances more easily than with random noise.
** This distortion is called intermodulation distortion.

42
of the frequency response. Figure
3.41a shows the jagged response of
a nonlinear network measured with
PRN. Because the PRN source
repeats itself exactly every time
record, this noisy looking trace never
changes and will not average to the
desired frequency response.
With random noise, the distortion
components are also random and will
average out. Therefore, the frequency
response does not include the distor-
tion and we get the more reasonable
results shown in Figure 3.41b.
This points out a fundamental
problem with measuring non-linear
networks; the frequency response is
not a property of the network alone,
it also depends on the stimulus.
Each stimulus, swept-sine, PRN and
random noise will, in general, give a
different result. Also, if the amplitude
of the stimulus is changed, you will
get a different result.
To illustrate this, consider the
mass-spring system with stops that
we used in Chapter 2. If the mass
does not hit the stops, the system
is linear and the frequency response
is given by Figure 3.42a.
If the mass does hit the stops,
the output is clipped and a large
number of distortion components are
generated. As the output approaches
a square wave, the fundamental com-
ponent becomes constant. Therefore,
as we increase the input amplitude,
the gain of the network drops. We
get a frequency response like Figure
3.42b, where the gain is dependent
on the input signal amplitude.
So as we have seen, the frequency
response of a nonlinear network is
not well defined, i.e., it depends on
the stimulus. Yet it is often used in
spite of this. The frequency response
of linear networks has proven to be a
very powerful tool and so naturally
people have tried to extend it to
non-linear analysis, particularly since
other nonlinear analysis tools have
proved intractable.
If every stimulus yields a different
frequency response, which one
should we use? The “best” stimulus
could be considered to be one which
approximates the kind of signals you
would expect to have as normal
inputs to the network. Since any large
collection of signals begins to look
like noise, noise is a good test signal*.
As we have already explained, noise
is also a good test signal because it
speeds the analysis by exciting all the
filters of our analyzer simultaneously.
But many other test signals can be
used with Dynamic Signal Analyzers
and are “best” (optimum) in other
senses. As explained in the beginning
of this section, sine waves can be
used to give the same results as other
types of network analyzers although
the speed advantage of the Dynamic
Signal Analyzer is lost. A fast sine
sweep (chirp) will give very similar
results with all the speed of Dynamic
Signal Analysis and so is a better
test signal. An impulse is a good test
signal for acoustical testing if the net-
work is linear. It is good for acoustics
because reflections from surfaces
at different distances can easily be
isolated or eliminated if desired. For
instance, by using the “force” window
described earlier, it is easy to get the
free field response of a speaker by
eliminating the room reflections from
the windowed time record.
Band-Limited Noise
Before leaving the subject of network
stimulus, it is appropriate to discuss
the need to band limit the stimulus.
We want all the power of the stimulus
to be concentrated in the frequency
region we are analyzing. Any power
* This is a consequence of the central limit theorem. As an
example, the telephone companies have found that when
many conversations are transmitted together, the result is
like white noise. The same effect is found more commonly
at a crowded cocktail party.
Figure 3.42
Nonlinear
system.
Figure 3.41
Nonlinear transfer function.
a) Pseudo-random noise stimulus b) Random noise stimulus

43
outside this region does not
contribute to the measurement
and could excite non-linearities.
This can be a particularly severe
problem when testing with random
noise since it theoretically has the
same power at all frequencies (white
noise). To eliminate this problem,
Dynamic Signal Analyzers often limit
the frequency range of their built-in
noise stimulus to the frequency span
selected. This could be done with an
external noise source and filters, but
every time the analyzer span changed,
the noise power and filter would have
to be readjusted. This is done auto-
matically with a built-in noise source
so transfer function measurements
are easier and faster.
Section 7: Averaging
To make it as easy as possible to
develop an understanding of Dynamic
Signal Analyzers we have almost
exclusively used examples with deter-
ministic signals, i.e., signals with no
noise. However, as the real world is
rarely so obliging, the desired signal
often must be measured in the pres-
ence of significant noise. At other
times the “signals” we are trying to
measure are more like noise them-
selves. Common examples that are
somewhat noise-like include speech,
music, digital data, seismic data and
mechanical vibrations. Because of
these two common conditions, we
must develop techniques both to
measure signals in the presence of
noise and to measure the noise itself.
The standard technique in statistics
to improve the estimates of a value
is to average. When we watch a
noisy reading on a Dynamic Signal
Analyzer, we can guess the average
value. But because the Dynamic
Signal Analyzer contains digital
computation capability we can have
it compute this average value for us.
Two kinds of averaging are available,
RMS (or “power” averaging) and
linear averaging.
RMS Averaging
When we watch the magnitude of the
spectrum and attempt to guess the
average value of the spectrum com-
ponent, we are doing a crude RMS*
average. We are trying to determine
the average magnitude of the signal,
ignoring any phase difference that
may exist between the spectra. This
averaging technique is very valuable
for determining the average power
in any of the filters of our Dynamic
Signal Analyzers. The more averages
we take, the better our estimate of
the power level.
In Figure 3.43, we show RMS aver-
aged spectra of random noise, digital
data and human voices. Each of these
examples is a fairly random process,
but when averaged we can see the
basic properties of its spectrum.
If we want to measure a small signal
in the presence of noise, RMS averag-
ing will give us a good estimate of the
signal plus noise. We can not improve
the signal to noise ratio with RMS
averaging; we can only make more
accurate estimates of the total signal
plus noise power.
Figure 3.43
RMS averaged spectra.
a) Random noise b) Digital data
c) Voices
Traces were separated 30 dB for clarity
Upper trace: female speaker
Lower trace: male speaker
* RMS stands for “root-mean-square” and is calculated
by squaring all the values, adding the squares together,
dividing by the number of measurements (mean) and
taking the square root of the result.

44
Linear Averaging
However, there is a technique for
improving the signal to noise ratio
of a measurement, called linear aver-
aging. It can be used if a trigger sig-
nal which is synchronous with the
periodic part of the spectrum is
available. Of course, the need for a
synchronizing signal is somewhat
restrictive, although there are numer-
ous situations in which one is avail-
able. In network analysis problems
the stimulus signal itself can often be
used as a synchronizing signal.
Linear averaging can be implemented
many ways, but perhaps the easiest to
understand is where the averaging is
done in the time domain. In this case,
the synchronizing signal is used to
trigger the start of a time record.
Therefore, the periodic part of the
input will always be exactly the same
in each time record we take, whereas
the noise will, of course, vary. If we
add together a series of these trig-
gered time records and divide by the
number of records we have taken we
will compute what we call a linear
average.
Since the periodic signal will have
repeated itself exactly in each time
record, it will average to its exact
value. But since the noise is different
in each time record, it will tend to
average to zero. The more averages
we take, the closer the noise comes
to zero and we continue to improve
the signal to noise ratio of our meas-
urement. Figure 3.44 shows a time
record of a square wave buried in
noise. The resulting time record after
128 averages shows a marked im-
provement in the signal to noise ratio.
Transforming both results to the
frequency domain shows how many
of the harmonics can now be accu-
rately measured because of the
reduced noise floor.
Figure 3.44
Linear averaging.
b) Single record, no averaginga) Single record, no averaging
d) 128 linear averagesc) 128 linear averages

45
Section 8:
Real Time Bandwidth
Until now we have ignored the fact
that it will take a finite time to com-
pute the FFT of our time record. In
fact, if we could compute the trans-
form in less time than our sampling
period we could continue to ignore
this computational time. Figure 3.45
shows that under this condition we
could get a new frequency spectrum
with every sample. As we have seen
from the section on aliasing, this
could result in far more spectrums
every second than we could possibly
comprehend. Worse, because of the
complexity of the FFT algorithm, it
would take a very fast and very
expensive computer to generate
spectrums this rapidly.
A reasonable alternative is to add a
time record buffer to the block dia-
gram of our analyzer. In Figure 3.47
we can see that this allows us to
compute the frequency spectrum of
the previous time record while gath-
ering the current time record. If we
can compute the transform before
the time record buffer fills, then we
are said to be operating in real time.
To see what this means, let us look at
the case where the FFT computation
takes longer than the time to fill the
time record. The case is illustrated in
Figure 3.48. Although the buffer is
full, we have not finished the last
transform, so we will have to stop
taking data. When the transform is
finished, we can transfer the time
record to the FFT and begin to take
another time record. This means that
we missed some input data and so we
are said to be not operating in real
time.
Recall that the time record is not
constant but deliberately varied to
change the frequency span of the ana-
lyzer. For wide frequency spans the
time record is shorter. Therefore, as
we increase the frequency span of the
analyzer, we eventually reach a span
where the time record is equal to the
FFT computation time. This frequen-
cy span is called the real time band-
width. For frequency spans at and
below the real time bandwidth, the
analyzer does not miss any data.
Real Time Bandwidth Requirements
How wide a real time bandwidth is
needed in a Dynamic Signal Analyzer?
Let us examine a few typical meas-
urements to get a feeling for the
considerations involved.
Adjusting Devices
If we are measuring the spectrum or
frequency response of a device which
we are adjusting, we need to watch
the spectrum change in what might
be called psychological real time. A
new spectrum every few tenths of a
second is sufficiently fast to allow an
operator to watch adjustments in
what he would consider to be real
time. However, if the response time
of the device under test is long, the
speed of the analyzer is immaterial.
We will have to wait for the device
to respond to the changes before the
spectrum will be valid, no matter
how many spectrums we generate
in that time. This is what makes
adjusting lightly damped (high Q)
resonances tedious.
Figure 3.45
A new
transform
every sample.
Figure 3.46
Time buffer
added to
block diagram.
Figure 3.48
Non-real time
operation.
Figure 3.47
Real time
operation.

46
RMS Averaging
A second case of interest in determin-
ing real time bandwidth requirements
is measurements that require RMS
averaging. We might be interested in
determining the spectrum distribution
of the noise itself or in reducing the
variation of a signal contaminated
by noise. There is no requirement in
averaging that the records must be
consecutive with no gaps*. Therefore,
a small real time bandwidth will not
affect the accuracy of the results.
However, the real time bandwidth
will affect the speed with which an
RMS averaged measurement can be
made. Figure 3.49 shows that for
frequency spans above the real time
bandwidth, the time to complete the
average of N records is dependent
only on the time to compute the
N transforms. Rather than continually
reducing the time to compute the
RMS average as we increase our
span, we reach a fixed time to
compute N averages.
Therefore, a small real time band-
width is only a problem in RMS aver-
aging when large spans are used with
a large number of averages. Under
these conditions we must wait longer
for the answer. Since wider real time
bandwidths require faster computa-
tions and therefore a more expensive
processor, there is a straightforward
trade-off of time versus money. In the
case of RMS averaging, higher real
time bandwidth gives you somewhat
faster measurements at increased
analyzer cost.
Transients
The last case of interest in determin-
ing the needed real time bandwidth
is the analysis of transient events. If
the entire transient fits within the
time record, the FFT computation
time is of little interest. The analyzer
can be triggered by the transient and
the event stored in the time record
buffer. The time to compute its
spectrum is not important.
However, if a transient event contains
high frequency energy and lasts
longer than the time record necessary
to measure the high frequency energy,
then the processing speed of the ana-
lyzer is critical. As shown in Figure
3.50b, some of the transient will not
be analyzed if the computation time
exceeds the time record length.
In the case of transients longer than
the time record, it is also imperative
that there is some way to rapidly
record the spectrum. Otherwise, the
Figure 3.49
RMS averaging
time.
Figure 3.50
Transient
analysis.
* This is because to average at all the signal must be
periodic and the noise stationary.

47
information will be lost as the
analyzer updates the display with
the spectrum of the latest time
record. A special display which
can show more than one spectrum
(“waterfall” display), mass memory,
a high speed link to a computer or a
high speed facsimile recorder is need-
ed. The output device must be able to
record a spectrum every time record
or information will be lost.
Fortunately, there is an easy way to
avoid the need for an expensive wide
real time bandwidth analyzer and an
expensive, fast spectrum recorder.
One-time transient events like explo-
sions and pass-by noise are usually
tape recorded for later analysis
because of the expense of repeating
the test. If this tape is played back at
reduced speed, the speed demands on
the analyzer and spectrum recorder
are reduced. Timing markers could
also be recorded at one time record
intervals. This would allow the analy-
sis of one record at a time and plot-
ting with a very slow (and commonly
available) X-Y plotter.
So we see that there is no clear-cut
answer to what real time bandwidth
is necessary in a Dynamic Signal
Analyzer. Except in analyzing long
transient events, the added expense
of a wide real time bandwidth gives
little advantage. It is possible to ana-
lyze long transient events with a nar-
row real time bandwidth analyzer, but
it does require the recording of the
input signal. This method is slow and
requires some operator care, but one
can avoid purchasing an expensive
analyzer and fast spectrum recorder.
It is a clear case of speed of analysis
versus dollars of capital equipment.
Section 9:
Overlap Processing
In Section 8 we considered the case
where the computation of the FFT
took longer than the collecting of the
time record. In this section we will
look at a technique, overlap process-
ing, which can be used when the FFT
computation takes less time than
gathering the time record.
To understand overlap processing, let
us look at Figure 3.51a. We see a low
frequency analysis where the gather-
ing of a time record takes much
longer than the FFT computation
time. Our FFT processor is sitting
idle much of the time. If instead of
waiting for an entirely new time
record we overlapped the new time
record with some of the old data,
we would get a new spectrum as
often as we computed the FFT. This
overlap processing is illustrated in
Figure 3.51b. To understand the
benefits of overlap processing, let
us look at the same cases we used
in the last section.
Adjusting Devices
We saw in the last section that we
need a new spectrum every few
tenths of a second when adjusting
devices. Without overlap processing
this limits our resolution to a few
Hertz. With overlap processing our
resolution is unlimited. But we are
not getting something for nothing.
Because our overlapped time record
contains old data from before the
device adjustment, it is not complete-
ly correct. It does indicate the direc-
tion and the amount of change, but
we must wait a full time record after
the change for the new spectrum to
be accurately displayed.
Nonetheless, by indicating the
direction and magnitude of the
changes every few tenths of a
second, overlap processing does
help in the adjustment of devices.
Figure 3.51
Understanding
overlap
processing.

48
RMS Averaging
Overlap processing can give dramatic
reductions in the time to compute
RMS averages with a given variance.
Recall that window functions reduce
the effects of leakage by weighting
the ends of the time record to zero.
Overlapping eliminates most or all of
the time that would be wasted taking
this data. Because some overlapped
data is used twice, more averages
must be taken to get a given variance
than in the non-overlapped case.
Figure 3.52 shows the improvements
that can be expected by overlapping.
Transients
For transients shorter than the time
record, overlap processing is useless.
For transients longer than the time
record the real time bandwidth of
the analyzer and spectrum recorder
is usually a limitation. If it is not,
overlap processing allows more
spectra to be generated from the
transient, usually improving
resolution of resulting plots.
Section 10: Summary
In this chapter we have developed
the basic properties of Dynamic
Signal Analyzers. We found that
many properties could be understood
by considering what happens when
we transform a finite, sampled time
record. The length of this record
determines how closely our filters
can be spaced in the frequency
domain and the number of samples
determines the number of filters in
the frequency domain. We also found
that unless we filtered the input we
could have errors due to aliasing and
that finite time records could cause
a problem called leakage which we
minimized by windowing.
We then added several features to
our basic Dynamic Signal Analyzer
to enhance its capabilities. Band
Selectable Analysis allows us to make
high resolution measurements even
at high frequencies. Averaging gives
more accurate measurements when
noise is present and even allows us
to improve the signal to noise ratio
when we can use linear averaging.
Finally, we incorporated a noise
source in our analyzer to act as a
stimulus for transfer function
measurements.
Figure 3.52
RMS averaging
speed improvements
with overlap
processing.

49
In Chapters 2 & 3, we developed an
understanding of the time, frequency
and modal domains and how
Dynamic Signal Analyzers operate.
In this chapter we show how to use
Dynamic Signal Analyzers in a wide
variety of measurement situations.
We introduce the measurement
functions of Dynamic Signal
Analyzers as we need them for
each measurement situation.
We begin with some common elec-
tronic and mechanical measurements
in the frequency domain. Later in the
chapter we introduce time and modal
domain measurements.
Section 1: Frequency Domain
Measurements
Oscillator Characterization
Let us begin by measuring the charac-
teristics of an electronic oscillator.
An important specification of an
oscillator is its harmonic distortion.
In Figure 4.1, we show the fundamen-
tal through fifth harmonic of a 1 KHz
oscillator. Because the frequency is
not necessarily exactly 1 KHz, win-
dowing should be used to reduce the
leakage. We have chosen the flat-top
window so that we can accurately
measure the amplitudes.
Notice that we have selected the
input sensitivity of the analyzer so
that the fundamental is near the top
of the display. In general, we set the
input sensitivity to the most sensitive
range which does not overload the
analyzer. Severe distortion of the
input signal will occur if its peak
voltage exceeds the range of the
analog to digital converter. Therefore,
all dynamic signal analyzers warn the
user of this condition by some kind of
overload indicator.
It is also important to make sure the
analyzer is not underloaded. If the
signal going into the analog to digital
converter is too small, much of the
useful information of the spectrum
may be below the noise level of the
analyzer. Therefore, setting the input
sensitivity to the most sensitive range
that does not cause an overload gives
the best possible results.
In Figure 4.1a we chose to display
the spectrum amplitude in logarith-
mic form to insure that we could see
distortion products far below the
fundamental. All signal amplitudes
on this display are in dBV, decibels
below 1 Volt RMS. However, since
most Dynamic Signal Analyzers have
very versatile display capabilities,
we could also display this spectrum
linearly as in Figure 4.1b. Here the
units of amplitude are volts.
Power-Line Sidebands
Another important measure of an
oscillator’s performance is the level
of its power-line sidebands. In Figure
4.2, we use Band Selectable Analysis
to “zoom in” on the signal so that we
can easily resolve and measure the
sidebands which are only 60 Hz away
from our 1 KHz signal. With some
analyzers it is possible to measure
signals only millihertz away from the
fundamental if desired.
Phase Noise
The short-term stability of a high
frequency oscillator is very important
in communications and radar. One
measure of this is called phase noise.
It is often measured by the technique
shown in Figure 4.3a. This mixes
down and cancels the oscillator
Figure 4.1
Harmonic distortion
of an Audio Oscillator -
Flat-top window used.
a) Logarithmic amplitude scale b) Linear amplitude scale
Figure 4.2
Powerline
sidebands of an
Audio Oscillator -
Band Selectable
Analysis and
Hanning window
used for maximum
resolution.
Chapter 4
Using Dynamic
Signal Analyzers

50
carrier leaving only the phase noise
sidebands. It is therefore possible to
measure the phase noise far below
the carrier level since the carrier does
not limit the range of our measure-
ment. Figure 4.3b shows the close-in
phase noise of a 20 MHz synthesizer.
Here, since we are measuring noise,
we use RMS averaging and the
Hanning window.
Dynamic Signal Analyzers offer
two main advantages over swept
signal analyzers in this application.
First, the phase noise can be meas-
ured much closer to the carrier. This
is because a good swept analyzer
can only resolve signals down to
about 1 Hz, while a Dynamic Signal
Analyzer can resolve signals to a few
millihertz. Secondly, the Dynamic
Signal Analyzer can determine the
complete phase noise spectrum in
a few minutes whereas a swept
analyzer would take hours.
Spectra-like phase noise are usually
displayed against the logarithm of fre-
quency instead of the linear frequen-
cy scale. This is done in Figure 4.3c.
Because the FFT generates linearly
spaced filters, the filters are not
equally spaced on the display. It is
important to realize that no informa-
tion is missed by these seemingly
widely spaced filters. We recall on
a linear frequency scale that all the
filters overlapped so that no part of
the spectrum was missed. All we have
done here is to change the presenta-
tion of the same measurement.
Figure 4.3
Phase Noise
Measurement.
a) Block diagram of phase noise measurement
b) Phase noise of a frequency synthesizer -
RMS averaging and Hanning window used for noise measurements
c) Logarithmic frequency axis presentation of phase noise normalized to a
1 Hz bandwidth (power spectral density)

51
In addition, phase noise and other
noise measurements are often nor-
malized to the power that would be
measured in a 1 Hz wide square filter.
This measurement is called a power
spectral density and is often provided
on Dynamic Signal Analyzers. It sim-
ply changes the presentation on the
display to this desired form; the data
is exactly the same in Figures 4.3b
and 4.3c, but the latter is in the more
conventional presentation.
Rotating Machinery Characterization
A rotating machine can be thought
of as a mechanical oscillator.* There-
fore, many of the measurements we
made for an electronic oscillator are
also important in characterizing
rotating machinery.
To characterize a rotating machine
we must first change its mechanical
vibration into an electrical signal.
This is often done by mounting an
accelerometer on a bearing housing
where the vibration generated by
shaft imbalance and bearing imper-
fections will be the highest. A typical
spectrum might look like Figure 4.4.
It is obviously much more complicat-
ed than the relatively clean spectrum
of the electronic oscillator we looked
at previously. There is also a great
deal of random noise; stray vibrations
from sources other than our motor
that the accelerometer picks up. The
effects of this stray vibration have
been minimized in Figure 4.4b RMS
averaging.
In Figure 4.5, we have used the Band
Selectable Analysis capability of our
analyzer to “zoom-in” and separate
the vibration of the stator at 120 Hz
from the vibration caused by the
rotor imbalance only a few tenths of
a Hertz lower in frequency.** This
ability to resolve closely spaced spec-
trum lines is crucial to our capability
to diagnose why the vibration levels
of a rotating machine are excessive.
The actions we would take to correct
an excessive vibration at 120 Hz are
quite different if it is caused by a
loose stator pole rather than an
imbalanced rotor.
Since the bearings are the most
unreliable part of most rotating
machines, we would also like to
check our spectrum for indications
of bearing failure. Any defect in a
bearing, say a spalling on the outer
face of a ball bearing, will cause a
small vibration to occur each time
a ball passes it. This will produce
a characteristic frequency in the
vibration called the passing frequen-
cy. The frequency domain is ideal for
Figure 4.6
Vibration caused by
small defect in
the bearing.
Figure 4.5
Stator vibration
and rotor imbalance
measurement with
Band Selectable
Analysis.
* Or, if you prefer, electronic oscillators can be viewed as
rotating machines which can go at millions of RPM’s.
** The rotor in an AC induction motor always runs at a
slightly lower frequency than the excitation, an effect
called slippage.
Figure 4.4
Spectrum of
electrical motor
vibration.

52
separating this small vibration from
all the other frequencies present. This
means that we can detect impending
bearing failures and schedule a shut-
down long before they become the
loudly squealing problem that signals
an immediate shutdown is necessary.
In most rotating machinery monitor-
ing situations, the absolute level of
each vibration component is not of
interest, just how they change with
time. The machine is measured when
new and throughout its life and these
successive spectra are compared.
If no catastrophic failures develop,
the spectrum components will
increase gradually as the machine
wears out. However, if an impending
bearing failure develops, the passing
frequency component corresponding
to the defect will increase suddenly
and dramatically.
An excellent way to store and com-
pare these spectra is by using a small
desktop computer. The spectra can
be easily entered into the computer
by an instrument interface like GPIB*
and compared with previous results
by a trend analysis program. This
avoids the tedious and error-prone
task of generating trend graphs by
hand. In addition, the computer can
easily check the trends against limits,
pointing out where vibration limits
are exceeded or where the trend is
for the limit to be exceeded in
the near future.
Desktop computers are also useful
when analyzing machinery that
normally operates over a wide range
of speeds. Severe vibration modes
can be excited when the machine
runs at critical speeds. A quick way
to determine if these vibrations are
a problem is to take a succession of
spectra as the machine runs up to
speed or coasts down. Each spectrum
shows the vibration components
of the machine as it passes through
an rpm range. If each spectrum is
transferred to the computer via
GPIB, the results can be processed
and displayed as in Figure 4.8. From
such a display it is easy to see shaft
imbalances, constant frequency vibra-
tions (from sources other than the
variable speed shaft) and structural
vibrations excited by the rotating
shaft. The computer gives the
capability of changing the display
presentation to other forms for
greater clarity. Because all the values
of the spectra are stored in memory,
precise values of the vibration com-
ponents can easily be determined.
In addition, signal processing can
be used to clarify the display. For
instance, in Figure 4.8 all signals
below -70 dB were ignored. This
eliminates meaningless noise from
the plot, clarifying the presentation.
So far in this chapter we have been
discussing only single channel fre-
quency domain measurements. Let us
now look at some measurements we
can make with a two channel
Dynamic Signal Analyzer.
Figure 4.7
Desktop
computer
system for
monitoring
rotating
machinery
vibration.
Motor Computer
Dynamic
Signal
Analyzer
Accelerometer
GPIB
Figure 4.8
Run up test
from the system
in Figure 4.7.
* General Purpose Interface Bus, Agilent’s
implementation of IEEE-488-1975.

53
Electronic Filter Characterization
In Section 6 of the last chapter, we
developed most of the principles we
need to characterize a low frequency
electronic filter. We show the test
setup we might use in Figure 4.9.
Because the filter is linear we can use
pseudo-random noise as the stimulus
for very fast test times. The uniform
window is used because the pseudo-
random noise is periodic in the time
record.* No averaging is needed
since the signal is periodic and rea-
sonably large. We should be careful,
as in the single channel case, to set
the input sensitivity for both channels
to the most sensitive position which
does not overload the analog to
digital converters.
With these considerations in mind,
we get a frequency response magni-
tude shown in Figure 4.10a and the
phase shown in Figure 4.10b. The
primary advantage of this measure-
ment over traditional swept analysis
techniques is speed. This measure-
ment can be made in 1/8 second with
a Dynamic Signal Analyzer, but would
take over 30 seconds with a swept
network analyzer. This speed
improvement is particularly impor-
tant when the filter under test is
being adjusted or when large volumes
are tested on a production line.
Structural Frequency Response
The network under test does not have
to be electronic. In Figure 4.11, we
are measuring the frequency response
of a single structure, in this case a
printed circuit board. Because this
structure behaves in a linear fashion,
Figure 4.10
Frequency response
of electronic filter using
PRN and uniform window.
Figure 4.9
Test setup
to measure
frequency
response
of filter.
* See the uniform window discussion in Section 6
of the previous chapter for details.
Figure 4.11
Frequency
response test
of a mechanical
structure.
a) Frequency response magnitude b) Frequency response magnitude and phase

54
we can use pseudo-random noise as
a test stimulus. But we might also
desire to use true random noise,
swept-sine or an impulse (hammer
blow) as the stimulus. In Figure 4.12
we show each of these measurements
and the frequency responses. As we
can see, the results are all the same.
The frequency response of a linear
network is a property solely of the
network, independent of the
stimulus used.
Since all the stimulus techniques
in Figure 4.12 give the same results,
we can use whichever one is fastest
and easiest. Usually this is the impact
stimulus, since a shaker is not
required.
In Figure 4.11 and 4.12, we have
been measuring the acceleration of
the structure divided by the force
applied. This quality is called
mechanical accelerance. To properly
scale the displays to the required
g’s/lb, we have entered the sensitivi-
ties of each transducer into the
analyzer by a feature called engineer-
ing units. Engineering units simply
changes the gain of each channel
of the analyzer so that the display
corresponds to the physical parame-
ter that the transducer is measuring.
Other frequency response measure-
ments besides mechanical acceler-
ance are often made on mechanical
structures. Figure 4.14 lists these
measurements. By changing transduc-
ers we could measure any of these
parameters. Or we can use the com-
putational capability of the Dynamic
Signal Analyzer to compute these
measurements from the mechanical
impedance measurement we have
already made.
Figure 4.12
Frequency
response
of a linear
network
is independent
of the stimulus
used.
a) Impact stimulus
Figure 4.13
Engineering
units set input
sensitivities to
properly scale
results.
b) Random noise stimulus
c) Swept sine stimulus

55
For instance, we can compute
velocity by integrating our accelera-
tion measurement. Displacement is
a double integration of acceleration.
Many Dynamic Signal Analyzers have
the capability of integrating a trace
by simply pushing a button. There-
fore, we can easily generate all the
common mechanical measurements
without the need of many expensive
transducers.
Coherence
Up to this point, we have been
measuring networks which we have
been able to isolate from the rest of
the world. That is, the only stimulus
to the network is what we apply and
the only response is that caused
by this controlled stimulus. This
situation is often encountered in test-
ing components, e.g., electric filters
or parts of a mechanical structure.
However, there are times when the
components we wish to test can not
be isolated from other disturbances.
For instance, in electronics we might
be trying to measure the frequency
response of a switching power supply
which has a very large component
at the switching frequency. Or we
might try to measure the frequency
response of part of a machine while
other machines are creating severe
vibration.
In Figure 4.15 we have simulated
these situations by adding noise and
a 1 KHz signal to the output of an
electronic filter. The measured
frequency response is shown in
Figure 4.16. RMS averaging has
reduced the noise contribution, but
has not completely eliminated the
1 KHz interference.* If we did not
know of the interference, we would
think that this filter has an additional
resonance at 1 KHz. But Dynamic
Signal Analyzers can often make an
additional measurement that is not
available with traditional network an-
alyzers called coherence. Coherence
measures the power in the response
channel that is caused by the power
in the reference channel. It is the out-
put power that is coherent with the
input power.
Figure 4.17 shows the same frequency
response magnitude from Figure 4.16
and its coherence. The coherence
goes from 1 (all the output power at
Figure 4.14
Mechanical
frequency
response
measurements.
Figure 4.15
Simulation
of frequency
response
measurement
in the presence
of noise.
Figure 4.17
Magnitude and
coherence of
frequency
response.
Figure 4.16
Magnitude of
frequency
response.
* Additional averaging would further reduce this
interference.

56
that frequency is caused by the input)
to 0 (none of the output power at that
frequency is caused by the input). We
can easily see from the coherence
function that the response at 1 KHz is
not caused by the input but by inter-
ference. However, our filter response
near 500 Hz has excellent coherence
and so the measurement here
is good.
Section 2: Time Domain
Measurements
A Dynamic Signal Analyzer usually
has the capability of displaying the
time record on its screen. This is the
same waveform we would see with
an oscilloscope, a time domain view
of the input. For very low frequency
or single-shot phenomena the digital
time record storage eliminates the
need for storage oscilloscope. But
there are other time domain measure-
ments that a Dynamic Signal Analyzer
can make as well. These are called
correlation measurements. We will
begin this section by defining correla-
tion and then we will show how to
make these measurements with a
Dynamic Signal Analyzer.
Correlation is a measure of the
similarity between two quantities. To
understand the correlation between
two waveforms, let us start by multi-
plying these waveforms together at
each instant in time and adding up all
the products. If, as in Figure 4.18, the
waveforms are identical, every prod-
uct is positive and the resulting sum
is large. If however, as in Figure 4.19,
the two records are dissimilar, then
some of the products would be posi-
tive and some would be negative.
There would be a tendency for the
products to cancel, so the final sum
would be smaller.
Now consider the waveform in
Figure 4.20a, and the same waveform
shifted in time, Figure 4.20b. If the
time shift were zero, then we would
have the same conditions as before,
that is, the waveforms would be in
phase and the final sum of the prod-
ucts would be large. If the time shift
between the two waveforms is made
large however, the waveforms appear
dissimilar and the final sum is small.
Figure 4.18
Correlation of
two identical
signals.
Figure 4.19
Correlation of
two different
signals.
Figure 4.20
Correlation of
time displaced
signals.

57
Going one step farther, we can find
the average product for each time
shift by dividing each final sum by
the number of products contributing
to it. If we now plot the average prod-
uct as a function of time shift, the
resulting curve will be largest when
the time shift is zero and will dimin-
ish to zero as the time shift increases.
This curve is called the auto-correla-
tion function of the waveform. It is
a graph of the similarity (or correla-
tion) between a waveform and itself,
as a function of the time shift.
The auto-correlation function is easi-
est to understand if we look at a few
examples. The random noise shown
in Figure 4.21 is not similar to itself
with any amount of time shift (after
all, it is random) so its auto-correla-
tion has only a single spike at the
point of 0 time shift. Pseudo-random
noise, however, repeats itself periodi-
cally, so when the time shift equals a
multiple of the period, the auto-corre-
lation repeats itself exactly as in
Figure 4.22. These are both special
cases of a more general statement;
the auto-correlation of any periodic
waveform is periodic and has the
same period as the waveform itself.
Figure 4.21
Auto correlation
of random noise.
a) Time record of random noise
Figure 4.22
Auto correlation
of pseudo-random
noise.
τ
Ν∆
∆
∆
Ν∆
b) Auto correlation of random noise

58
This can be useful when trying to
extract a signal hidden by noise.
Figure 4.24a shows what looks like
random noise, but there is actually a
low level sine wave buried in it. We
can see this in Figure 4.24b where
we have taken 100 averages of the
auto-correlation of this signal. The
noise has become the spike around
a time shift of zero whereas the
auto-correlation of the sine wave is
clearly visible, repeating itself with
the period of the sine wave.
If a trigger signal that is synchronous
with the sine wave is available, we
can extract the signal from the noise
by linear averaging as in the last
section. But the important point
about the auto-correlation function
is that no synchronizing trigger is
needed. In signal identification prob-
lems like radio astronomy and pas-
sive sonar, a synchronizing signal is
not available and so auto-correlation
is an important tool. The disadvan-
tage of auto-correlation is that the
input waveform is not preserved as it
is in linear averaging.
Since we can transform any time
domain waveform into the frequency
domain, the reader may wonder what
is the frequency transform of the
auto-correlation function? It turns out
to be the magnitude squared of the
spectrum of the input. Thus, there is
really no new information in the auto-
correlation function, we had the same
Figure 4.23
Auto-correlation
of periodic
waveforms.
Figure 4.24
Auto-correlation
of a sine wave
buried by noise.

59
information in the spectrum of the
signal. But as always, a change in
perspective between these two
domains often clarifies problems.
In general, impulsive type signals
like pulse trains, bearing ping or
gear chatter show up better in corre-
lation measurements, while signals
with several sine waves of different
frequencies like structural vibrations
and rotating machinery are clearer in
the frequency domain.
Cross Correlation
If auto-correlation is concerned with
the similarity between a signal and a
time shifted version of itself, then it
is reasonable to suppose that the
same technique could be used to
measure the similarity between two
non-identical waveforms. This is
called the cross correlation function.
If the same signal is present in both
waveforms, it will be reinforced in
the cross correlation function, while
any uncorrelated noise will be
reduced. In many network analysis
problems, the stimulus can be cross
correlated with the response to
reduce the effects of noise. Radar,
active sonar, room acoustics and
transmission path delays all are net-
work analysis problems where the
stimulus can be measured and used
to remove contaminating noise from
the response by cross correlation.*
Figure 4.25
Simulated radar
cross correlation.
a) ‘Transmitted’ signal, a swept-frequency sine wave
Figure 4.26
Cross correlation
shows multiple
transmission
paths.
b) ‘Received’ signal, the swept sine wave plus noise
c) Result of cross correlation of the transmitted and received signals.
Distance from left edge to peak represents transmission delay.
* The frequency transform of the cross correlation
function is the cross power spectrum, a function
discussed in Appendix A.

60
Section 3: Modal Domain
Measurements
In Section 1 we learned how to make
frequency domain measurements of
mechanical structures with Dynamic
Signal Analyzers. Let us now analyze
the behavior of a simple mechanical
structure to understand how to make
measurements in the modal domain.
We will test a simple metal plate
shown in Figure 4.27. The plate
is freely suspended using rubber
cords in order to isolate it from
any object which would alter its
properties.
The first decision we must make in
analyzing this structure is how many
measurements to make and where to
make them on the structure. There
are no firm rules for this decision;
good engineering judgment must be
exercised instead. Measuring too
many points make the calculations
unnecessarily complex and time
consuming. Measuring too few points
can cause spatial aliasing; i.e., the
measurement points are so far apart
that high frequency bending modes
in the structure can not be measured
accurately. To decide on a reasonable
number of measurement points, take
a few trial frequency response mea-
surements of the structure to deter-
mine the highest significant resonant
frequencies present. The wave length
can be determined empirically by
changing the distance between the
stimulus and the sensor until a full
360° phase shift has occurred from
the original measurement point.
Measurement point spacing should
be approximately one-quarter or less
of this wavelength.
Measurement points can be spaced
uniformly over the structure using
this guideline, but it may be desirable
to modify this procedure slightly. Few
structures are as uniform as this sim-
ple plate example,* but complicated
structures are made of simpler, more
uniform parts. The behavior of the
structure at the junction of these
parts is often of great interest, so
measurements should be made in
these critical areas as well.
Once we have decided on where the
measurements should be taken, we
number these measurement points
(the order can be arbitrary) and enter
the coordinates of each point into our
modal analyzer. This is necessary so
that the analyzer can correlate the
measurements we make with a
position on the structure to compute
the mode shapes.
The next decision we must make is
what signal we should use for a stim-
ulus. Our plate example is a linear
structure as it has no loose rivet
joints, non-linear damping materials,
or other non-linearities. Therefore,
we know that we can use any of the
stimuli described in Chapter 3,
Section 6. In this case, an impulse
would be a particularly good test
signal. We could supply the impulse
by hitting the structure with a ham-
mer equipped with a force transducer.
* If all structures were this simple, there would be
no need for modal analysis.
Figure 4.28
Spacial Aliasing -
Too few
measurement
points lead to
inaccurate
analysis of
high frequency
bending mode.
Figure 4.27
Modal analysis
example -
Determine
the modes in
this simple
plate.

61
This is probably the easiest way to
excite the structure as a shaker and
its associated driver are not required.
As we saw in the last chapter, howev-
er, if the structure were non-linear,
then random noise would be a good
test signal. To supply random noise to
the structure we would need to use a
shaker. To keep our example more
general, we will use random noise as
a stimulus.
The shaker is connected firmly to the
plate via a load cell (force transduc-
er) and excited by the band-limited
noise source of the analyzer. Since
this force is the network stimulus, the
load cell output is connected through
a suitable amplifier to the reference
channel of the analyzer. To begin
the experiment, we connect an
accelerometer* to the plate at the
same point as the load cell. The
accelerometer measures the struc-
ture’s response and its output is con-
nected to the other analyzer channel.
Because we are using random noise,
we will use a Hanning window and
RMS averaging just as we did in the
previous section.
The resulting frequency response
of this measurement is shown in
Figure 4.29. The ratio of acceleration
to force in g’s/lb is plotted on the
vertical axis by the use of engineering
units, and the data shows a number
of distinct peaks and valleys at partic-
ular frequencies. We conclude that
the plate moves more freely when
subjected to energy at certain specific
frequencies than it does in response
to energy at other frequencies. We
recall that each of the resonant peaks
correspond to a mode of vibration of
the structure.
Our simple plate supports a number
of different modes of vibration, all of
which are well separated in frequen-
cy. Structures with widely separated
modes of vibration are relatively
straightforward to analyze since
each mode can be treated as if it is
the only one present. Tightly spaced,
but lightly damped vibration modes
can also be easily analyzed if the
Band Selectable Analysis capability
is used to narrow the analyzer’s filter
sufficiently to resolve these resonanc-
es. Tightly spaced modes whose
damping is high enough to cause the
responses to overlap create computa-
tional difficulties in trying to separate
the effects of the vibration modes.
Fortunately, many structures fall into
the first two categories and so can be
easily analyzed.
Having inspected the measurement
and deciding that it met all the above
criteria, we can store it away. We
store similar measurements at each
point by moving our accelerometer to
each numbered point. We will then
have all the measurement data we
need to fully characterize the struc-
ture in the modal domain.
Recall from Chapter 2 that each
frequency response will have the
same number of peaks, with the same
resonant frequencies and dampings.
The next task is to determine these
resonant frequency and damping
values for each resonance of interest.
We do this by retrieving our stored
frequency responses and, using a
curve-fitting routine, we calculate
the frequency and damping of each
resonance of interest.
With the structural information we
entered earlier, and the frequency and
damping of each vibration mode
which we have just determined, the
analyzer can calculate the mode
Figure 4.29
A frequency
response of
the plate.
* Displacement, velocity or strain transducers could also be
used, but accelerometers are often used because they are
small and light, and therefore do not affect the response of
the structure. In addition, they are easy to mount on the
structure, reducing the total measurement time.

62
shapes by curve fitting the responses
of each point with the measured
resonances. In Figure 4.30 we show
several mode shapes of our simple
rectangular plate. These mode shapes
can be animated on the display to
show the relative motion of the vari-
ous parts of the structure. The graphs
in Figure 4.30, however, only show
the maximum deflection.
Section 4: Summary
This note has attempted to demon-
strate the advantages of expanding
one’s analysis capabilities from the
time domain to the frequency and
modal domains. Problems that are
difficult in one domain are often
clarified by a change in perspective
to another domain. The Dynamic
Signal Analyzer is a particularly
good analysis tool at low frequencies.
Not only can it work in all three
domains, it is also very fast.
We have developed heuristic argu-
ments as to why Dynamic Signal
Analyzers have certain properties
because understanding the principles
of these analyzers is important in
making good measurements. Finally,
we have shown how Dynamic Signal
Analyzers can be used in a wide
range of measurement situations
using relatively simple examples.
We have used simple examples
throughout this text to develop
understanding of the analyzer and
its measurements, but it is by no
means limited to such cases. It is
a powerful instrument that, in the
hands of an operator who under-
stands the principles developed in
this note, can lead to new insights
and analysis of problems.
Figure 4.30
Mode shapes
of a rectangular
plate.

63
The Fourier Transform
The transformation from the time
domain to the frequency domain and
back again is based on the Fourrier
Transform and its inverse. This Fourier
Transform pair is defined as:
Sx(f) = x (t) e-j2πftdt (Forward Transform) A.l
x(t) = Sx(f) e j2πftdf (Inverse Transform) A.2
where
x(t) = time domain representation of the signal x
Sx(f) = frequency domain representation of the
signal x
j = √

-1
The Fourier Transform is valid for both
periodic* and non-periodic x(t) that
satisfy certain minimum conditions.
All signals encountered in the real
world easily satisfy these requirements.
The Discrete Fourier Transform
To compute the Fourier Transform
digitally, we must perform a numerical
integration. This will give us an
approximation to a true Fourier
Transform called the Discrete
Fourier Transform.
There are three distinct difficulties
with computing the Fourier Transform.
First, the desired result is a continuous
function. We will only be able to
calculate its value at discrete points.
With this constraint our transform
becomes,
Sx(m∆f) = x (t) e-j2πm∆ftdt A.3
where m = 0, ±1, ±2
and ∆f = frequency spacing of our lines
The second problem is that we must
evaluate an integral. This is equivalent
to computing the area under a curve.
We will do this by adding together the
areas of narrow rectangles under the
curve as in Figure A.l.
Our transform now becomes:
Sx(m∆f) ≈ ∆t x (n∆t) e-j2πm∆fn∆t A.4
where ∆t = time interval between samples
The last problem is that even with this
summation approximation to the integral,
we must sum samples over all time from
minus to plus infinity. We would have to
wait forever to get a result. Clearly then,
we must limit the transform to a finite
time interval.
Sx(m∆f) ≈ ∆t x (n∆t) e-j2πm∆fn∆t A.5
As developed in Chapter 3, the frequency
spacing between the lines must be the
reciprocal of the time record length.
Therefore, we can simplify A.5 to our
formula for the Discrete Fourier
Transform, S'x.
S'x(m∆f) ≈ x(n∆t) e-j2πmn/Ν A.6
Figure A.l
Numerical
integration
used in the
Fourier
Transform
∆
π ∆
⌠⌠ ∞∞
⌡⌡−−∞∞
⌠⌠ ∞∞
⌡⌡−−∞∞
* The Fourier Series is a special case of the
Fourier Transform.
⌠⌠ ∞∞
⌡⌡−−∞∞
n=-∞
∑
n-1
∑
n-0
T
N
n-1
∑
n-0
∞
Appendix A
The Fourier Transform:
A Mathematical Background

64
The Fast Fourier Transform
The Fast Fourier Transform (FFT) is
an algorithm for computing this
Discrete Fourier Transform (DFT).
Before the development of the FFT
the DFT required excessive amounts
of computation time, particularly
when high resolution was required
(large N). The FFT forces one further
assumption, that N is a multiple of 2.
This allows certain symmetries to
occur reducing the number of calcula-
tions (specifically multiplications)
which have to be done.
It is important to recall here that the
Fast Fourier Transform is only an
approximation to the desired Fourier
Transform. First, the FFT only gives
samples of the Fourier Transform.
Second and more important, it is only
a transform of a finite time record of
the input.
Two Channel Frequency Domain
Measurements
As was pointed out in the main
text, two channel measurements are
often needed with a Dynamic Signal
Analyzer. In this section we will
mathematically define the two channel
transfer function and coherence
measurements introduced in
Chapter 4 and prove their more
important properties.
However, before we do this, we wish
to introduce one other function, the
Cross Power Spectrum, Gxy . This
function is not often used in measure-
ment situations, but is used internally
by Dynamic Signal Analyzers to
compute transfer functions and
coherence.
The Cross Power Spectrum, Gxy, is defined as taking
the Fourier Transform of two signals separately and
multiplying the result together as follows:
Gxy (f) = Sx (f) S*y(f)
where * indicates the complex conjugate of the function.
With this function, we can define the
Transfer Function, H(f), using the cross
power spectrum and the spectrum of the
input channel as follows:
H(f) =
where

denotes the average of the function.
At first glance it may seem more
appropriate to compute the transfer
function as follows:
|H(f)|2 =
This is the ratio of two single channel,
averaged measurements. Not only does
this measurement not give any phase
information, it also will be in error when
there is noise in the measurement. To see
why let us solve the equations for the
special case where noise is injected into
the output as in Figure A.2. The
output is:
Sy(f) = Sx(f)H(f) + Sn(f)
So
Gyy=SySy*= Gxx|H|2+SxHSn+Sx*H*Sn+|Sn|2
Figure A.2
Transfer
function
measurments
with noise
present.
Gyy
Gxx
Gyy(f)
Gxx(f)

65
If we RMS average this result to try to
eliminate the noise, we find the SxSn
terms approach zero because Sx, and Sn,
are uncorrelated. However, the |Sn|2
term remains as an error and so we get
= |H|2 +
Therefore if we try to measure |H|2 by
this single channel techniques, our value
will be high by the noise to signal ratio.
If instead we average the cross power
spectrum we will eliminate this noise
error. Using the same example,
Gyx=SySx*=(SxH+Sn)Sx*= GxxH +SnSx*
so
=H(f)+SnSx*
Because Sn, and Sx, are uncorrelated,
the second term will average to zero,
making this function a much better
estimate of the transfer function.
The Coherence Function, γ2, is also
derived from the cross power
spectrum by:
γ2(f) =
As stated in the main text, the coherence
function is a measure of the power in
the output signal caused by the input.
If the coherence is 1, then all the output
power is caused by the input. If the
coherence is 0, then none of the output
is caused by the input. Let us now look
at the mathematics of the coherence
function to see why this is so.
As before, we will assume a measurement
condition like Figure A.2. Then, as we
have shown before,
Gxy= Gxx|H|2+SxHSn*+ Sx*H*Sn+|S|2
Gxy=GxxH+SnSx*
As we average, the cross terms SnSx,
approach zero, assuming that the signal
and the noise are not related. So the
coherence becomes
γ2(f) =
γ2(f) =
We see that if there is no noise, the
coherence function is unity. If there is
noise, then the coherence will be reduced.
Note also that the coherence is a function
of frequency. The coherence can be unity
at frequencies where there is no interference
and low where the noise is high.
Time Domain Measurements
Because it is sometimes easier to under-
stand measurement problems from the
perspective of the time domain, Dynamic
Signal Analyzers often include several time
domain measurements. These include auto
and cross correlation and impulse response.
Auto Correlation, Rxx(τ), is a comparison
of a signal with itself as a function of
time shift. It is defined as;
Rxx(τ)= x(t)x(t+τ)dt
Gyy
Gxx
|Sn|2
Gxx
Gyx
Gxx
Gyx(f)
Gxx(f)
Gxy*(f)
Gyy(f)
|H|2Gxx
|H|2Gxx+Sn
2
(HGyx)2
Gxx(|H|2Gxx+|Sn|2)
lim
T→∞
1
T
⌠⌠
⌡⌡ΤΤ

66
That is, the auto correlation can be
found by taking a signal and multiply-
ing it by the same signal displaced by
a time τ and averaging the product
over all time. However, most Dynamic
Signal Analyzers compute this quanti-
ty by taking advantage of its dual in
the frequency domain. It can be
shown that
Rxx(τ)=F -1[Sx(f)Sx*(f)]
where F -1 is the inverse Fourier
Transform and Sx is the Fourier
Transform of x(t)
Since both techniques yield the same
answer, the latter is usually chosen for
Dynamic Signal Anlyzer since the
Frequency Transform algorithm is
already in the instrument and the
results can be computed faster
because fewer multiplications are
required.
Cross Correlation, Rxy(τ), is a
comparison of two signals as a
function of a time shift between them.
It is defined as:
Rxy(τ)= x(t)y(t+τ)dt
As in auto correlation, a Dynamic
Signal Analyzer computes this quanti-
ty indirectly, in this case from the
cross power spectrum.
Rxy(τ)=F -1[Gxy]
Lastly, the Impulse Response, h(t), is
the dual of the transfer function,
h(t) = F -1[H(f)]
Note that because the transfer func-
tion normalized the stimulus, the
impulse response can be computed no
matter what stimulus is actually used
on the network.
Bendat, Julius S. and Piersol, Allan G.,
“Random Data: Analysis and
Measurement Procedures”,
Wiley-Interscience,
New York, 1971.
Bendat, Julius S. and Piersol, Allan G.,
“Engineering Applications of
Correlation and Spectral Analysis”,
Wiley-lnterscience, New York, 1980.
Bracewell, R., “The Fourier
Transform and its Applications”,
McGraw-Hill, 1965.
Cooley, J.W. and Tukey, J.W.,
“An Algorithm for the Machine
Calculation of Complex Fourier
Series”, Mathematics of Computation,
Vo. 19, No. 90, p. 297,
April 1965.
McKinney, W., “Band Selectable
Fourier Analysis”, Hewlett-Packard
Journal, April 1975,
pp. 20-24.
Otnes, R.K. and Enochson, L.,
“Digital Time Series Analysis”,
John Wiley, 1972.
Potter, R. and Lortscher, J.,
“What in the World is Overlap
Processing”, Hewlett-Packard
Santa Clara DSA/Laser Division
“Update”, Sept. 1978.
Ramse , K.A., “Effective Measure-
ments for Structural Dynamics
Testing”, Part 1, Sound and
Vibration Magazine,
November 1975, pp. 24-35.
Roth, P., “Effective Measurements
Using Digital Signal Analysis”,
IEEE Spectrum, April 1971,
pp. 62-70.
Roth, P., “Digital Fourier Analysis”,
Hewlett-Packard Journal, June 1970.
Welch, Peter D., “The Use of
Fast Fourier Transform for the
Estimation of Power Spectra:
A Method Based on Time
Averaging Over Short, Modified
Periodograms”, IEEE Transactions
on Audio and Electro-acoustics,
Vol. AU-15, No. 2,
June 1967, pp. 70-73.
lim
T→∞
1
T
⌠⌠
⌡⌡ΤΤ
Appendix B
Bibliography

67
Aliasing 29
Anti-Alias Filter 31
Auto-Correlation 57, 65
Band Selectable Analysis 33, 51, 61
Coherence 55, 65
Correlation 56, 65
Cross Correlation 59, 66
Cross Power Spectrum 64
Damping 17, 61
Decibels (dB) 8, 49
Digital Filter 32
Discrete Fourier Transform 63
Engineering Units 55
Fast Fourier Transform (FFT) 25, 64
Flat-top Window or Passband 38, 49
Force Window 39
Fourier Transform 63
Frequency Response 15, 61
Frequency Response Analyzer 19
Gain-Phase Meter 19
Hanning Window or Passband 37, 49, 60
Impulse Response 66
Leakage 36
Linear Averaging 44
Linearity 12
Lines 25
Logarithmic Frequency Display 50
Network Analysis 11
Network Analyzer 19
Nyquist Criterion 31
Oscillograph 6
Oscilloscope 6
Parallel Filter Spectrum Analyzer 17
Periodic Random Noise 41, 53
Phase Noise 49
Power Spectral Density 51
Pseudo Random Noise 41, 53
Q (of resonance) 17
Random Noise 41
Real Time 45
Rectangular Window 37
Response Window 39
RMS Averaging 43, 46, 48, 50
Self-Windowing Functions 37
Spectrum 7
Spectrum Analyzer 17
Spectrum Component 7
Stimulus/Response Testing 11
Strip Chart Recorders 5
Swept Filter Spectrum Analyzer 18
Time Record 26
Transfer Function 42, 64
Transient Response 16
Tuned Network Analyzer 19
Uniform Window or Passband 37
Vibration Mode 20
Windowing 34
Zoom 33
Index

Product specifications and descriptions in this
document subject to change without notice.
Copyright © 1994, 1995, 1999,
2000 Agilent Technologies
Printed in U.S.A. 6/00
5952-8898E
Agilent Technologies' Test and Measurement
Support, Services, and Assistance
Agilent Technologies aims to maximize the
value you receive, while minimizing your risk
and problems. We strive to ensure that you get the
test and measurement capabilities you paid for
and obtain the support you need. Our extensive
support resources and services can help
you choose the right Agilent products for your
applications and apply them successfully. Every
instrument and system we sell has a global war-
ranty. Support is available for at least five years
beyond the production life of the product. Two
concepts underlie Agilent's overall support
policy: "Our Promise" and "Your Advantage."
Our Promise
Our Promise means your Agilent test and
measurement equipment will meet its advertised
performance and functionality. When you are
choosing new equipment, we will help you with
product information, including realistic perform-
ance specifications and practical recommenda-
tions from experienced test engineers. When
you use Agilent equipment, we can verify that it
works properly, help with product operation, and
provide basic measurement assistance for the use
of specified capabilities, at no extra cost upon
request. Many self-help tools are available.
Your Advantage
Your Advantage means that Agilent offers a wide
range of additional expert test and measurement
services, which you can purchase according to
your unique technical and business needs. Solve
problems efficiently and gain a competitive edge
by contracting with us for calibration, extra-cost
upgrades, out-of-warranty repairs, and on-site
education and training, as well as design, system
integration, project management, and other
professional engineering services. Experienced
Agilent engineers and technicians worldwide can
help you maximize your productivity, optimize the
return on investment of your Agilent instruments
and systems, and obtain dependable measurement
accuracy for the life of those products.
For More Assistance with Your
Test & Measurement Needs go to
www.agilent.com/find/assist
Or contact the test and measurement experts
at Agilent Technologies
(During normal business hours)
United States:
(tel) 1 800 452 4844
Canada:
(tel) 1 877 894 4414
(fax) (905) 206 4120
Europe:
(tel) (31 20) 547 2323
(fax) (31 20) 547 2390
Japan:
(tel) (81) 426 56 7832
(fax) (81) 426 56 7840
Latin America:
(tel) (305) 267 4245
(fax) (305) 267 4286
Australia:
(tel) 1 800 629 485
(fax) (61 3) 9272 0749
New Zealand:
(tel) 0 800 738 378
(fax) 64 4 495 8950
Asia Pacific:
(tel) (852) 3197 7777
(fax) (852) 2506 9284

Fundamentals of dsp

Recommended

More Related Content

What's hot (19)

Similar to Fundamentals of dsp (20)

Recently uploaded (20)

Fundamentals of dsp