Fundamentals of Signal Analysis Series

Understanding
Dynamic Signal Analysis
Application Note 1405-2

Table of Contents

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3
Section 1: FFT Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4
Section 2: Sampling and Digitizing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8
Section 3: Aliasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9
Section 4: Band-Selectable Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .13
Section 5: Windowing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .14
Section 6: Network Stimulus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .20
Section 7: Averaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .23
Section 8: Real-Time Bandwidth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .25
Section 9: Overlap Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .28
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .29
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .29
Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .30

Introduction

This application note is a primer

for those who are unfamiliar with
the class of analyzers we call
dynamic signal analyzers. These
instruments are particularly
appropriate for the analysis
of signals in the range of a few
millihertz to about a hundred
kilohertz.

advantages and disadvantages of

each generic instrument type and
noted 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 note, we avoid using

rigorous mathematics and instead
depend 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 dynamic
signal analysis. Equally important,
this heuristic instruction leads to
better instrument operators who
can intelligently use these analyzers
to solve complicated measurement
problems with accuracy and
ease.*

In this application note, we will

develop a fuller understanding
of dynamic signal analyzers.
We begin by presenting the
properties of the Fast Fourier
Transform (FFT), upon which
dynamic signal analyzers are
based. We then show how these
FFT properties cause some
undesirable characteristics 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.

In Application Note 1405-1,

Introduction to Time, Frequency
and Modal Domains, we introduced
the concepts of the time, frequency
and modal domains and the
types of instrumentation available
for making measurements in
each domain. We discussed the

* A more rigorous mathematical justification for the

arguments developed in the main text can be found in
Application Note 1405-4

Section 1: FFT Properties

The Fast Fourier Transform (FFT)

is an algorithm* for transforming
data from the time domain to the
frequency 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 that complicate
this seemingly straightforward
task.
First, because of the many calculations involved in transforming
domains, the transform must be
performed on a computer if the
results are to be sufficiently
accurate. Fortunately, with the
advent of microprocessors, it is
easy and inexpensive to incorporate
all the needed computing power
in a small instrument 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 domain to samples in
the frequency domain as shown
in Figure 1.1.**
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, we will consider what
sample spacing is necessary
to guarantee accurate results.

* 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.

Figure 1.1. The FFT samples in both the time and frequency domains

Figure 1.2. A time record is N equally spaced samples of the input.

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 1.3, 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 1.4.
Thus, once the time record is

Figure 1.3. The FFT works on blocks of data.

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 AN1405-1, Section 3.
When a signal is first applied to
a parallel-filter analyzer, 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. In later sections on realtime bandwidth and overlap
processing, we discuss just how
fast a dynamic signal analyzer
should transform.

Figure 1.4. A new time record every sample after the time record is filled

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 1.5 shows a 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 1.6, we have shifted the
phase of the higher frequency sine
wave components of this signal.
The result is a severe distortion
of the original waveform.
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. In Application
Note 1405-3, you will see
that phase contains valuable
information in determining the
cause of performance problems.

Figure 1.5. The relationship between the time and frequency domains

Figure 1.6. Phase of frequency domain components is important.

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 1.7 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.
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.

Figure 1.7. Lowest frequency resolvable by the FFT

What is the
Frequency Range of the FFT?
We can now quickly determine
that the highest frequency we can
measure is:
f max = N
2

1
(Period of Time
Record )

Figure 1.8. Frequencies of all the spectral lines of the FFT

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
frequency range of our measurement, we must vary fmax. The
number of time samples N is fixed
by the implementation 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

Figure 1.9. Frequency range of dynamic signal analyzers is determined by sample rate.

variable time record period. This is

illustrated in Figure 1.9. 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.
7

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 it
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
spectrum analyzer. We show this
basic block diagram in Figure 2.1.
For the analyzer to have the
high accuracy needed for many
measurements, 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 multimeter (DMM)
will typically exceed these
specifications, but the ADC for
a dynamic signal analyzer must
be much faster than typical fast
DMMs. A fast DMM 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.

* You can skip this section and the next if you are not
interested in the internal operation of a dynamic signal
analyzer. However, if you specify the purchase of
dynamic signal analyzers, you are especially encouraged
to read these sections. The basic knowledge you gain
from these sections can insure you specify the best
analyzer for your requirements.
8

Figure 2.1. Block diagram of a dynamic signal analyzer

Figure 2.2. The sampler and ADC must not introduce errors.

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 problem in any sampled
data system. It is often overlooked,
sometimes with disastrous results.

Figure 3.1. A simple sampled data system

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 3.1. A
thermocouple is connected to a
digital voltmeter that is in turn
connected to a printer. The system
is set up to print the temperature
every second. What would we
expect for an output? If we were
measuring the temperature of a
room that 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 3.2.

The Case of the

Missing Temperature
If, on the other hand, we were
measuring the temperature of a
small part that could heat and
cool rapidly, what would the
output be? Suppose that the
temperature of our part cycled

Figure 3.2. Plot of temperature variation of a room

Figure 3.3. Plot of temperature variation of a small part

exactly once every second. As

shown in Figure 3.3, 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 3.4.
Two signals are said to alias if the
difference of their frequencies falls
in the frequency range of interest.
This difference frequency is always
generated in the process of sampling.
In Figure 3.4, the input frequency
is slightly higher than the sampling
frequency so a low frequency alias
term is generated. If the input
frequency equals the sampling
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
temperature 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 3.5 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.

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

Figure 3.4. The problem of aliasing viewed in the frequency domain

Figure 3.5. A frequency domain view of how to avoid aliasing - sample at greater than twice the highest input
frequency

* Aliasing is also known as fold-over or mixing.

10

sufficient information. It certainly

would not be enough to give a
high-quality time display. Yet we
saw in Figure 3.5 that meeting the
Nyquist Criterion of a sample rate
greater than twice the maximum
input frequency is sufficient to
avoid aliasing and preserve all the
information in the input signal.

Figure 3.6. Nyquist Criterion in the time domain

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 power-line 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 3.7a. 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 3.7b
with a gradual roll off and finite
rejection of undesired signals.
Large input signals that are not
well attenuated in the transition
band could still alias into the
desired input frequency range.
To avoid this, the sampling
frequency is raised to twice the
highest frequency of the transition
band. This guarantees that any
signals that could alias are well
attentuated by the stop band of the
filter. Typically, this means that
the sample rate is now two-and-ahalf to four times the maximum
desired input frequency. Therefore,
a 25 kHz FFT spectrum analyzer
can require an ADC that runs at
100 kHz*.
* 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.

Figure 3.7. Actual anti-alias filters require higher sampling frequencies.

The Need for More

Than One Anti-Alias Filter
You may recall from Section 1 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. Their
transition bands should 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 significantly
to the price of the analyzer. To cut
expenses, some manufacturers
dont use a low enough frequency
anti-alias filter on the lowestfrequency spans. (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.

11

Digital Filtering
Fortunately, there is an
alternative that 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 3.8.
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). For digital
filtering, the ADC sample rate is
left constant at the rate needed for
the highest-frequency span of the
analyzer. 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 variations,
aging or drift in the filter. Therefore,
in a two-channel analyzer, the
filters in each channel are identical.
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 expensive than
analog anti-aliasing filtering.

12

Figure 3.8. Block diagrams of analog and digital filtering

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 measuring the power-line 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 4.1. 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

Figure 4.1. High-resolution measurements with band-selectable analysis

Figure 4.2. Analyzer block diagram

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 4.2. 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.
13

Section 5: Windowing

The Need for Windowing

There is another property of
the Fast Fourier Transform that
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
throughout time, as illustrated
in Figure 5.1.

Figure 5.1. FFT assumption time record repeated throughout all time

This does not cause a problem

with the transient case shown.
But what happens if we are
measuring a continuous 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 5.2.
In this case, the input waveform
is said to be periodic in the time
record.
Figure 5.3 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 5.3c. 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 5.3c should be very different.
Since sharp phenomena 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.

Figure 5.2. Input signal periodic in time record

Figure 5.3. Input signal not periodic in time record

14

In Figure 5.4 we see in an actual

measurement that our expectations
are correct. In Figures 5.4a and b,
we see a sine wave that is periodic
in the time record. Its frequency
spectrum is a single line whose
width is determined only by the
resolution of our dynamic signal
analyzer. On the other hand,
Figures 5.4c and d show a sine
wave that is not periodic in the
time record. Its power has been
spread throughout the spectrum
as we predicted.

a)

b)

a) and b) Sine wave periodic in time record

c)

d)

This smearing of energy throughout the frequency domains is a

phenomena 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

c) and d) Sine wave not periodic in time record

Figure 5.4. Actual FFT results

leakage if the continuous input is

not periodic in the time record.
It is obvious from Figure 5.4 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 problems 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
representative cases.

15

What is Windowing?
In Figure 5.5 we have again
reproduced 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
concentrate 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 5.5c. Such functions are
called window functions because
they force us to look at data
through a narrow window.
Figure 5.6 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 5.5d, something 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 5.5b, but it still is not
correct. Figure 5.7 demonstrates
that the windowed data does not
have as narrow a spectrum as an
unwindowed function which is
periodic in the time record.

Figure 5.5. The effect of windowing in the time domain

a) Sine wave not periodic in time record

c) FFT results with a window function

Figure 5.6. Leakage reduction with windowing

16

b) FFT results with no window function

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 5.6 as
our example of leakage reduction
with windowing. The Hanning
window is also commonly used
when measuring random noise.
a) Leakage-free measurement input periodic in
time record

The Uniform Window*

b) Windowed measurement input not periodic in

time record

Figure 5.7. Windowing reduces leakage but does not eliminate it.

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 transient is shown in
Figure 5.8a. If we multiplied it by
the window function in Figure
5.8b we would get the highly
distorted signal shown in Figure
5.8c. The frequency spectrum of
an actual transient with and
without the Hanning window is
shown in Figure 5.9. 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 a 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
* The uniform window is sometimes referred to as a
rectangular window.

Figure 5.8. Windowing loses information from transient events.

a) Unwindowed trainsients

b) Hanning windowed transients

Figure 5.9. Spectrums of transients

is zero at the beginning and end of

the time record. Remember that
we introduced 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 function like this

which does not require a window

because it occurs completely
within the time record is called
a self-windowing function. Selfwindowing functions generate no
leakage in the FFT and so need no
window.
17

There are many examples of selfwindowing functions, some of

which are shown in Figure 5.10.
Impacts, impulses, shock responses,
sine bursts, noise bursts, chirp
bursts and pseudo-random
noise can all be made to be selfwindowing. Self-windowing
functions are often used as the
excitation in measuring the
frequency response of networks,
particularly if the network has
lightly-damped resonances
(high Q). This is because the selfwindowing functions generate
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-windowing 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 5.11 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 percent) when it falls midway
between the filters.
This error is unacceptably large if
we are trying to measure a signals
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 5.12. The
amplitude error from this window
function does not exceed .1 dB
(1%), a 1.4 dB improvement.
The accuracy improvement
does not come without its price,
however. Figure 5.13 shows that
we have flattened the top of the

Figure 5.11. Hanning passband shapes

Figure 5.12. Flat-top passband shapes

Figure 5.10. Self-windowing function examples

* There is another way to avoid this problem using bandselectable analysis. We illustrate this in Agilent
Application Note 1405-3.
** It will, in fact, be periodic in the time record.
18

Figure 5.13. Reduced resolution of the flat-top window

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 that 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
sometimes added. Figure 5.14
shows a response window acting
on the response 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 beginning of the
time record so we make sure that
this is weighted most heavily 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 5.15. 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 windows and other
analyzers call those 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.

Figure 5.15. Using the force window

Figure 5.14. Using the response window

19

Section 6: Network Stimulus

As we explained in Agilent
Application Note 1405-1, 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 6.1. The
analyzer then measures the gain
and phase of the network at this
frequency while the rest of the
analyzers filters measure only
noise. We then increase the source
frequency to the next filter center,
wait for the network to settle and
then measure 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.

Figure 6.1. Frequency response measurements with a sine wave stimulus

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
6.2, acts like hundreds of sine
wave generators connected
together. Although this sounds
very expensive, just such a source
can be easily generated digitally.
It is called a pseudo-random noise
or periodic random noise source.

This procedure would, within

experimental error, give us the
same results as we would get with
any of the network analyzers
described in Agilent Application
Note 1405-1, with any network,
linear or nonlinear.

20

Figure 6.2. Pseudo-random noise as a stimulus

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.

Figure 6.3. Random noise as a stimulus

Figure 6.4. Pseudo-random noise distortion

A truly random noise source has a

spectrum shown in Figure 6.3. 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 frequency response can
be determined. Therefore, PRN is
the best stimulus to use with fairly
linear networks because it gives
the fastest results.*

* 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.

a) Pseudo-random noise stimulus

b) Random noise stimulus

Figure 6.5. Nonlinear transfer function

Non-Linear Network Analysis

If the network is severely nonlinear, 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 6.4 that
if two sine waves are put through
a nonlinear network, distortion
products will be generated
equally spaced from the signals.**
Unfortunately, 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 therefore interfere
with the measurement of the
frequency response. Figure 6.5a
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 distortion and we
get the more reasonable results
shown in Figure 6.5b.
This points out a fundamental
problem with measuring nonlinear 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.
21

To illustrate this, consider the

mass-spring system with stops
that we used in Agilent Application
Note 1405-1. If the mass does not
hit the stops, the system is linear
and the frequency response is
given by Figure 6.6a.
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 component becomes
constant. Therefore, as we increase
the input amplitude, the gain
of the network drops. We get a
frequency response like Figure
6.6b, 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 too,
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 that
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

* 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.
22

Figure 6.6. Nonlinear system

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 network 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 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
automatically 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 deterministic signals, i.e.,
signals with no noise. However,
the real world is rarely so obliging.
The desired signal often must be
measured in the presence of
significant noise. At other times
the signals we are trying to
measure are more like noise
themselves. 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.

a) Random noise

b) Digital data

Figure 7.1. RMS averaged spectra

* 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.

RMS Averaging
When we watch the magnitude of
the spectrum and attempt to guess
the average value of the spectrum
component, 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 7.1, we show RMS
averaged 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 averaging will give us a good
estimate of the signal plus noise.
We can not improve the signal-tonoise ratio with RMS averaging;
we can only make more accurate
estimates of the total signal-plusnoise power.

c) Voices. Traces were separated 30 dB for clarity

Upper trace: female speaker
Lower trace: male speaker

23

Linear Averaging
However, there is a technique for
improving the signal-to-noise ratio
of a measurement, called linear
averaging. It can be used if a
trigger signal is available that is
synchronous with the periodic
part of the spectrum. Of course,
the need for a synchronizing
signal is somewhat restrictive,
although there are numerous
situations in which one is
available. 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 triggered 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-tonoise ratio of our measurement.

Figure 7.2 shows a time record

of a square wave buried in noise.
The resulting time record after
128 averages shows a marked
improvement in the signal to noise
ratio. Transforming both results to
the frequency domain shows how
many of the harmonics can now
be accurately measured because
of the reduced noise floor.

a) Single record, no averaging

b) Single record, no averaging

c) 128 linear averages

d) 128 linear averages

Figure 7.2. Linear averaging

24

Section 8: Real-Time Bandwidth

Until now we have ignored the

fact that it will take a finite time
to compute the FFT of our time
record. In fact, if we could compute
the transform in less time than
our sampling period we could
continue to ignore this computational time. Figure 8.1 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
diagram of our analyzer (Figure
8.2). In Figure 8.3 we can see that
this allows us to compute the
frequency spectrum of the previous
time record while gathering 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 8.4.
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
analyzer. 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 frequency
span is called the real-time
bandwidth. 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 measurements to get a
feeling for the considerations
involved.

Adjusting Devices
If we are measuring the spectrum
or frequency response of a device
that 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 or she
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 8.1. A new transform every sample

Figure 8.2. Time buffer added to block diagram

Figure 8.3. Real-time operation

Figure 8.4. Non-real-time operation

25

RMS Averaging
A second case of interest in
determining 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 8.5 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.

Figure 8.5. RMS averaging time

Transients
The last case of interest in
determining 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

Therefore, a small real-time

bandwidth is only a problem in
RMS averaging 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 computations 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.

Figure 8.6. Transient analysis

26

record necessary to measure the

high-frequency energy, then the
processing speed of the analyzer
is critical. As shown in Figure
8.6b, 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 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 highspeed link to a computer or a
high-speed facsimile recorder is
needed. 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 explosions
and pass-by noise are usually
digitally recorded for later
analysis because of the expense of
repeating the test. Continuously
sampled time data can be recorded
into a large time-capture memory
or a high-speed through-put disk.
This allows you to analyze the
data later with no information
loss.

So we see that there is no clearcut 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 realtime bandwidth gives little
advantage. It is possible to analyze
long transient events with a
narrow real-time bandwidth
analyzer, but it does require the
recording of the input signal. This
method is slow and requires some
operator care, but you can avoid
purchasing an expensive analyzer
and fast spectrum recorder. It is
a clear case of speed of analysis
versus dollars of capital
equipment.

27

Section 9: Overlap Processing

In Section 8 we considered the

case where the computation of the
FFT took longer than the time it
took to collect the time record.
In this section we will look at a
technique, overlap processing,
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 9.1a. We see
a low-frequency analysis where
gathering the 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 9.1b. 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 completely correct. It does
indicate the direction and the
amount of change, but we must
wait a full time record after the
change for the new spectrum to
be accurately displayed.

Figure 9.1. Understanding overlap processing

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.

must be taken to get a given

variance than in the nonoverlapped case. Figure 9.2
shows the improvements that
can be expected by overlapping.

RMS Averaging

Transients

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

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.

Figure 9.2. RMS averaging speed improvements with overlap processing

28

Summary

Bibliography

In this application note, 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 that we minimized by
windowing. We then added
several features to our basic
dynamic signal analyzer to
enhance its capabilities. Bandselectable 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.

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, HewlettPackard 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
Measurements for Structural
Dynamics Testing, Part 1,
Sound and Vibration Magazine,
November 1975, pp. 24-35.

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.

Related Agilent Literature

Agilent Application Note
Introduction to Time, Frequency
and Modal Domains, pub. no.
1405-1
Agilent Application Note
Using Dynamic Signal Analysers,
pub. no. 1405-3
Agilent Application Note
The Fourier Transform: A
Mathematical Background,
pub. no. 1405-4
Product Overview
Agilent 35670A Dynamic
Signal Analyzer,
pub. no. 5966-3063E
Product Overview
Agilent E1432/33/34
VXI Digitizers/Source,
pub. no. 5968-7086E
Product Overview
Agilent E9801B Data
Recorder/Logger,
pub. no. 5968-6132E

Roth, P., Effective Measurements

Using Digital Signal Analysis,
IEEE Spectrum, April 1971,
pp. 62-70.

29

Glossary
Aliasing a phenomenon that
can occur when a signal is not
sampled at greater than twice
the maximum frequency
component; high-frequency
signals appear as low-frequency
components; avoided by
filtering out signals greater than
1/2 the sample rate
Anti-alias filter a low pass filter
installed before the sampler
and analog-to-digital converter
to limit the input frequency
range of a signal to prevent
aliasing; designed to filter out
frequencies greater than 1/2 the
sample rate (typically 1/2.56 to
allow for filter rolloff)
Band-selectable analysis an
analysis capability that allows
you to zoom in for a highresolution close-up shot of the
frequency spectrum by
concentrating filters in the
frequency range of interest.
Digital filter a decimating filter
that filters the digital
representation of the input
signal (after it has been
sampled and digitized) to the
desired frequency span. It also
reduces the sample rate at its
output to the rate needed for
that frequency span.
Fast Fourier Transform (FFT)
an algorithm used in computers
and DSAs to compute discrete
frequency components from
sampled time data; invented by
Cooley and Tukey
Flat-top window a windowing
function that minimizes
amplitude error for off-center
input-signal components

30

Force window a windowing

function that eliminates stray
signals; used for the excitation
signal in impact test to improve
signal-to-noise ratio

Pseudo-random noise a
mathematically generated
random noise whose period is
matched to time record length,
thus eliminating leakage

Hanning window a windowing

function used to reduce leakage
when measuring noise and
periodic signals like sine waves

Random noise true noise

Leakage the spreading of

energy throughout the
frequency domain; energy leaks
out of one resolution line of an
FFT into other lines, sometimes
masking small signals close to a
sine wave. This happens when
the signal is not periodic within
a time record. Applying an
appropriate window function
will minimize the amplitude
error due to the leakage.
Linear averaging a technique
for improving the signal-tonoise ratio of a measurement;
can be used if a trigger signal is
available that is synchronous
with the periodic part of the
spectrum
Lines To reduce confusion
about which domain we are in,
samples in the frequency
domain are called lines.
Nyquist Criterion the minimum
theoretical sample rate for a
baseband signal to be
reproduced in sampled form,
equal to twice the highest
frequency of the input signal. In
most cases, you will want to use
a higher sample rate to
represent the signal accurately.
Periodic random noise (PRN)
a kind of pseudo-random noise
that periodically repeats its
sequence; the best stimulus for
testing linear networks.

Real time If we can compute the

transform (FFT) before the time
record buffer fills, we are said
to be operating in real time.
Rectangular window another
term for uniform window
Response window an
exponential-weighted window
that guarantees that the
response dies out (goes to zero)
within the time record; used in
impact test to avoid leakage
error; also called exponential
window
RMS averaging a technique for
measuring small signals in the
presence of noise. RMS
averaging is useful for
processing stationary signals.
Each data block can be
overlapped to achieve the
maximum number of averages.
RMS averaging produces
amplitude information only.
You lose phase information.
Self-windowing functions a
function that does not require a
window because it occurs
completely within the time
record or its period is matched
to time-record length (it
generates no leakage in the
FFT)
Time record a block of N
consecutive, equally spaced
samples of the input; this block
is the basic unit transformed by
an FFT.

Transfer function a ratio of the

output over the input, both in
the Laplace domain;
sometimes used
interchangeably with
frequency response function
Uniform window a windowing
function that weights all of the
time record uniformly; used for
transient signals
Windowing a way to reduce
leakage by forcing an FFT to
look at data through a narrow
window where the input is zero
at both ends of the time record.
Many different functions can be
used to window the data,
depending on the type of
measurement you are making.
Zoom another term for bandselectable analysis

31

www.agilent.com

Agilent Email Updates

www.agilent.com/find/emailupdates
Get the latest information on the products and applications you select.

By internet, phone, or fax, get assistance

with all your test & measurement needs
Online assistance:
www.agilent.com/find/assist
Phone or Fax
United States:
(tel) 1 800 452 4844
Canada:
(tel) 1 877 894 4414
(fax) (905) 282 6495
China:
(tel) 800 810 0189
(fax) 800 820 2816
Europe:
(tel) (31 20) 547 2323
(fax) (31 20) 547 2390
Japan:
(tel) (81) 426 56 7832
(fax) (81) 426 56 7840
Korea:
(tel) (82 2) 2004 5004
(fax) (82 2) 2004 5115
Latin America:
(tel) (305) 269 7500
(fax) (305) 269 7599
Taiwan:
(tel) 0800 047 866
(fax) 0800 286 331
Other Asia Pacific Countries:
(tel) (65) 6375 8100
(fax) (65) 6836 0252
(e-mail) [email protected]

Product specifications and descriptions in this

document subject to change without notice.
Agilent Technologies, Inc. 1994, 1995, 1999, 2000, 2002
Printed in the USA August 7, 2002
5988-6774EN

32