Machinery Messages
Sampling waveforms and
computing spectra
‘by Dom Southich
“Technical Training Engineer
Bently Nevada Corporation
here ate three “Laws of the Us
I verse" which deal with samp
waveforms and computing spec
tro with a Fast Fourier Transform
(FET):
The sample rate governs the fre=
queney response.
The wotal sample time governs the
resolution,
The number of samples governs
the numberof ines.
Since sample rate, total sample time
andthe numberof samples interact, you
red to only specify Wo ofthe thre.
Experiment 1
12 owit
Lawé1: The sample rate governs
the frequency response
‘The faster you sample, the higher the
frequency response you will get. Or,
‘conversely if you need to measure high
frequencies, the sample rate must alo
be high. Infact, to avoid a phenomenon
‘known as aliasing, the sample rate must
be atleast twice 8 fasta the highest
‘frequency component in the signal.
F,22Fya
Fy Its
where Fy= the sample rate ip Hertz
Puce the highest frequency
‘component in the signal
in Herte
tye the somple interval in
seconds
‘Thar seems easy enough, but how do
you know the highest frequency compo-
nent in a signal you haven't measured
yet? The answer that you don't know,
‘There is, however, a way around this
‘quandary first, specify the highest fre-
quency you want to measure, Fron
‘Then, remove all frequency compo
‘nentsin the signal higher than Fy, ith
‘low-pass filter (which iscalled an umti-
alias filter in this application). Finally,
the low-pass filtered signal can be sam-
pled aca rate faster than 2+P,y, (2.50
3 times is ypiea)
Since “real-world” filters must be
‘used (theoretically, ideal filters are not
yet commercially available), the
usualy a possibilty of aliasing occurs
Experiment 2
September 1903fing in the highest Frequencies of most
FFT spectra. To avoid plotting possibly
aliased values, these top ines are ust
ally discarded. Only the lowest 100 ines
of «128 line Spectrum are ploted, the
lowest 200 lines of a 256 line Speetrur,
the lawest 400 lines of @ S12 line Spec-
cum, ete
Examples
(1) To see up to 500 He in a Spee-
trum, the sample rate must be at
least 1000 samples per second,
2) For a sample rate of 32 syn-
chronous samples per revolution,
the highest frequency line in a
Speetcum will (theoretically) be
WGordersof running speed (16X).
[Note that this aw says nothing about
how many samples are taken or how
Tongit takes to collect allot the samples.
Consequently, it doeso't say anything
about how many Spectrum Fines there
willbe or what the spacing between the
Tines wil be either.
Law#2: The total sample time
‘governs the resolution
“The resolution you get out ofan FFT,
ice. the spacing between Spectrum lines,
i the reciprocal of the length of time
you sample, If you have to resolve or
Amplitude valu
Phase value = Arctangent
distinguish two separate frequencies
which are close together (small a) you
tnust sample fora lng time (large).
aur
where af= the spacing between fe-
«quency ines in Hoste
‘T= the total sample time in seconds
Lake at Experiment 1A sevefoem
hasbeen printed in yellow ink om a blue
background, Since yellow and blue ate
wwelleparated frequencies the wave
Torms can be quikly distinguished. In
Experiment 2, we hae changed the Boe
background io shade of yellow close in
hue tothe shade of yellow use forthe
veaveform, It is now more difficult to
{stinguish the yellow figure onthe yel-
Jow background, and you wll have to
look for longer period of time before
you ea resolve the waveform
Examples:
(1) For atoral sample time of 05 see
‘ons the frequency lines of the
Specirum willbe 2 He apart.
@)"To esol two frequency compo
nents which are 6 «pm (0.1 Hz)
part, the signal mist he sampled
for at east 1 seconds.
@) Fora tal sample time of 8 rx0-
ations, the frequency lines ofthe
Spectrum will be spaced 18 of on
"Direct valuc® + “Quadrature” value?
“Quadrature” value
“Direct” valu
“Direct” (cosine) Value =
Figure 1
(Complex Spectrum Lines
September 1993.
order of running speed apart
(18%).
[Note that this law says nothing about
what the sample rate is, oF how many
samples are collected. Correspondingly
itdoes not sy anything about frequency
response, or how many Spectrum lines
there will be either.
Law#3: The number of samples
‘govems the number of lines in the
‘Spectrum,
FFTs use complex numbers. The
numberof frequency domain valuesyoo
{get out of an FFT isthe same as the
rnumber of time domain simples you put
in. Think of it as a Conservation of
Numbers principle at work here. Esch
frequency line from an FFT has adieet
‘of cosine part and a quadrature or sine
part, These are called “real” and “imag-
inary respectively, in some math books.
Ifyou putin 2048 time domain samples,
you obtain 2048 frequency domain
walues — a 1024 line Spectrum where
cach line has two values, 2 direet value
and a quadrature value. Note that this is
‘equivalent to saying that each Spectrum
line has both an amplitude value and
phase value
For rotating machinery, we almost
always plot only the amplitude values
from the complex output of an FFT; the
phase values are not plotted as part of a
Spectrum. These amplitude spectra,
therefore, consist of half the number
lines as there are time domain samples,
‘The phase values are not discarded:
they ate simply used elsewhere. Typ
‘eal, only the phase of synchronous
frequency lines (integral multiples of
running speed, ie. IX, 2X, 3X, ete, oF
‘occasionally integral submuliples of
running speed, be. W2X, V3X, V4X,
2ISX, ete.) are used. The phase of all
non-synchronous frequency lines wil
vary as & function of exacty when the
signal is sampled; sampling cither a
fraction ofa second earlier oF later ean
produce willy different phase values.
[Note that this law says nothing about
the sample rate or the total imei takes
tocollect all of the samples. Correspon
ingly, it does not say anything about
frequency response of frequency res0-
lution either. >
omit 13Examples:
These three Laws of the Universe are
interdependent. Again, you must spe
ify any two of them to meet your
requirements; the third then becomes
fixed, and you can't do anything 40
change it
(1) Ifyou sample for 0.5 seconds and
you want a 3200 line $
(which is a truncated 4096 line
Spectrum):
Lawi: a
tet
Ha,
4006 line Spectrom
requires 8192 samples
Te F,= 8192 samples 5 seconds
‘= 16.384 samplesisecond
Fy = 16384 H2? = $192 He
nesam™ 3200s % 2 He
= 6100 He
2) Ifyou have to resolve frequencies
12 epm (0.2 Hz) apart, and you
want an SOK (really 1024) fine
Spectrum:
5 second
Lawes
Law?) T= vat
seconds
Law#3; a 1024 line Spectrum
equines 2048 samples
2048 samples seconds
409.6 samplesisecond
Fc™ 409.6 H2!2 = 214 8 He
800 lines * 0.2 Heine
10H
(3) Ifyou tive to reso frequencies
‘epm (0.1 Hz) apart and the high:
est frequency you need 10 see is
oa He
fF
180 He:
Lawifl: Fy22+F yy 2 24180 Hz
2360 Hz"
Lawe2: T= Wat = 0.1 He = 10
seconds
Therefore, taking 360 samples/second
(or more) for 10 seconds = 3600 (or
more) samples,
The nest higher power of 2 is 4096.
4096 samples will yield. a 2048 line Spec
tram which will be truncated to 1600
lines. A 1600 line Spectrum with lines
01 Hi apart only shows up wo 160 Hy,
which is not high enough to mect
requitements.
The next higher power of 2 is 8192.
{8192 samples ill yield. a 40% line Spee
trum which will he truncated to 3200
14 owit
nes A 3200 line Spectrum with lines
0.1 He apart will stow up 0320 He
(4) you ake 32 samples pertevolue
sionfor resolutions you wiltake Digital Signal Processing (DSP)
4 total of 256 samples and get a “Tricks”
128 line Spectrum, which will ‘There a
sruneate wo 100 Fines, and
100 lines % 0.125 Hevline
28x,
also some DSP techniques
‘which an be used to “enhance” an FFT.
T= 8 revolutions But be aware, there are always
IFT = US revolutions = tradeofts
18x Without resompling, the number of
Fux 32 samples per revolu- Spectrum lines eam be increased, and
ion? = 16X the line spacing ean be decreased, pro
Figures 263
For steady state applications, i.e. when the machine is running at a constant
speed, the signal can be sampled over a longer time period. Notice that the 3200
line Spectrum (Figute 1) required 32 seconds to acquire the waveform data. The
400 line Spectrum (Figure 2) only sampled for 4 seconds, and provides inadequate
resolution.
September 1993ducing an “enhanced” Spectrum. To
accomplish this, one major assumption
must be made: a set of samples taken
over a relatively short period of time
‘must accurately represent a longer term
behavior ofthe signal For example, you
‘assime that 'you were wosample 2,48,
cle., times longer, the signal wouldnt
change characteristics, Without getting
too technical, one technique (called
$9880 20m
zero-padding) and the rationale behind
itisas follows:
you sample a signal at a certain rate
for a certain period of time, say 2088
samples per second for 05 seconds, you
will get (24 samples, a 512 line Spee-
‘trum with a high frequency of 1024 Hz
and a resolution of 2 Hz, Assuming that
this 0.5 second period accurately epre-
sents the behavior of the signal aver a
‘rane
Figures 465
For transient applications,
startupishutdown, the signal should be sampled over as
short atime pesiod as possible to avoid “smearing.” Smearing occurs when the machine
speed changes while the datas being sampled, Lethe machine speed might be 2000 rpm
at the beginning of the sample period and, due to acceleration be 2200 rpm at the end of
the sample period. Comparing Figure 3 and Figure 4, natice thatthe 100 Ine Spectrum
_rovides beter information than the 400 line Spectrum (due tothe shorter sample period).
September 1903 ss
LD second period, scale the samples
(multiply each sample value by an
amplitude adjustment factor), put them
inthe first half of a 2088 sample buffer,
fil the last half ofthe buffer with zeros,
then perform the FFT. The resulting
Spectrum will have 1024 Fines, a high
Frequency of 1(24 Hz and a resolution of
1H,
‘This same technique will work with
‘ther larger bulers, such as ones four
times as large. In this example, fill the
first 1/4 of the buffer with scaled sam-
ples and the last 3/4 with zeros. If you
have a buffer which is eight times as big,
fl the frst 1/8 ofthe buffer with scaled
samples and the last 7/8 with eros, ete
‘These “enhanced” spectra will have
hoth high resolution and short sample
times.
Information accuracy and quality
‘Zero-pacding changes only the com-
putational line spacing: the effective
resolution bandwidth remains the same
‘no matter how many zeros are added
(Reference 1). Inother words, the infor-
‘mation content (the effective resolution
bandwidth) isthe same with or without
the zero padding.
‘The more accurately the samples
taken over short time period represent
what really happened over a longer (but
‘not totally sampled) time period, the
‘more accurate the enhanced Spectrum
willbe I the samples taken over a short
time period de not accurately eepresent
the signal over the longer period, the
cenhanced Spectrum will misrepresent
what really happened over the longer
period. The higher the degree of en
‘cement, for example, 8 t0 1,160 1,320
ete, the more likely iis for this mis-
representation to oecur.
Sometimes these DSP “tricks” are
valid and sometimes they aren‘. Uni-
versal application of any technique
which might he valid only within lim
ited scope of circumstances can produce
poor or misleading results.
Reference
1. Rober K Oinesand Loren Exec Appi
Time Senet dab, ohn Wiley & Sons, Now
ore 1m, p28.
obit 15