Pitfalls in the Analysis of

Machinery Vibration Measurements

John L. Frarey, Schenectady, New York

Strictly speaking, the FFT analyzer is intended for use with frequency signal by an acceptable amount. The random noise
harmonically related signals that are stationary in time. In level in the same bin as a discrete signal may add or subtract
other words, all signals are in the center of the FFT bin and from the amplitude of the signal. (This actually takes place in
never change with the analysis time. We have learned how to the real and imaginary spectra before combining them into the
live with non-bin centered data by the use of windows. We still magnitude spectrum.) The number of averages smooths the
have a way to go in handling very noisy data. That path is even noise level to its average value but averages the magnitude of
longer in being able to extract useful data from transient vi- the discrete signal to a value nearer its actual amplitude. There-
bration data. The secret to obtaining useful data in these cases fore, the variability in the discrete signal amplitude will ap-
is an understanding of the FFT process and how it may be used proach zero if enough averages are taken.
to provide good approximations of transient data. This article Figure 1 shows the time history of a noise signal including
will discuss several cases where great care must be taken in one discrete frequency component. The spectrum shows a 400
setting up the analyzer or very misleading data will be pro- line analysis that has been averaged by summation eight times.
duced. The spectrum appears quite clean. Figure 2 shows the ampli-
tude of the discrete signal after repeating the test ten times. The
The knowledge and understanding of the FFT process by actual level of the signal is shown, as are the ten amplitudes
vibration specialists have greatly expanded since I gave my first measured, the average of the ten amplitudes and the 3 val-
talk on the FFT at the 1980 Annual Meeting of the Vibration ues. About 99% of the data will fall between these 3 values.
Institute in New Orleans. Then, data acquisition times, win- One can see that in two successive tests, the data for the con-
dows and the relationship between the A/D sample rate and stant amplitude discrete signal varied from 8.5 to 9.7. This is
the F max of the 400 line spectrum were not widely understood. a 14% increase in the measured amplitude when none existed
Today, anyone certified as a Level 1 Vibration Specialist has a in the actual values. Figures 3 and 4 show the same test except
pretty good idea of how the FFT vibration analyzer works. In that the number of averages was 64. Of the ten trials, the mini-
fact they probably use one daily in their machinery condition mum amplitude was 9.15 and the maximum amplitude was
monitoring program. 9.41. This is an increase in the measured amplitude of 2.8%
In spite of this improved understanding of the FFT process, when none existed in the actual value of the discrete signal.
there are still some areas where the unsuspecting can get into Suppose the above variations in reading the amplitude of the
serious trouble. In some monitoring programs the number of discrete signal were coupled with a machine speed that could
averages to take for accelerometer data seems to be a constant vary. Assuming a Hanning window is used, an additional 14%
of nature. We always take 8 averages for example. The number increase could occur if a speed change moved the discrete sig-
of averages should be determined after a review of the charac- nal frequency from the edge of the bin to the center of the bin.
ter of the data. For extremely noisy data, the combination of The lesson from the above data is that there is a relationship
too few averages coupled with very tight vibration limit specs between averaging time or data acquisition time for higher line
may lead to an unnecessarily high number of alarms. spectra and how tight one can set the alert limits for a changed
Strictly speaking, the FFT process should be used only on vibration reading. Eight averages of a 400 line spectrum for all
stationary data. In other words, the data must not be varying acceleration data is not a wise rule for data collection. Rather
in frequency or statistical characteristics. Additionally, it was we must make the decision based on consideration of the time
intended to only treat harmonic data where each signal is pre- to take the data and the amount of variability we will accept
cisely a harmonic of the center of the first bin. If these rules in the amplitude of discrete signals.
were strictly followed, then the FFT would not be the popular
processing tool it has become. One must realize however, that Beware of Fat Peaks
when one uses the process in nonstationary data situations, one In an ideal situation, all signals would reside in the middle
has to be very careful in setting up the analyzer or serious er- of a given FFT bin and not vary from this location during the
rors can occur. The FFT has even been used to process tran- time to acquire data. Sometimes in real life this condition can-
sient data. The idea that the peak hold feature can produce the not be fulfilled. Figure 5 is a comparison of two time histories
amplitude portion of a Bode plot during startups or shutdowns and their respective spectra. Note that the peak amplitude of
can result in getting bad data for the response at the critical both traces is approximately 250 mV. The top spectrum, asso-
speed, depending on the setup of the analyzer. Some of these ciated with the top time trace only shows a peak amplitude of
situations will be analyzed in this article. 180 mV. The bottom spectrum shows the correct amplitude of
250 mV. Besides the different amplitude readings, the major dif-
The Number of Averages is a Constant of Nature ference between the two spectra is the width of the peak. Very
The statement that When looking at accelerometer data from close inspection of the top trace would show that the time per
a new machine, a 400 line spectrum and 8 averages is a rea- cycle is varying. In other words, the signal is frequency modu-
sonable first choice has been interpreted in some plants as lated and thus moving through several bins during the data ac-
Use a 400 line spectrum and 8 averages for all accelerometer quisition time. In a real case this means that the machine speed
data. The first statement assumed that one would produce a has varied during the data acquisition time or in the case of an
spectrum and then depending on the character of the data, instability such as oil whip or steam whirl, the frequency of
modify the number of lines and the number of averages to pro- the signal has varied during the data acquisition period.
duce a spectrum where the noise level will impact the discrete The true amplitude of a fat peak can be partially corrected
by taking the square root of the sum of the squares of the data
Based on a paper presented at the National Technical Training Sym-
in each bin. While this correction will partially correct the
posium and 26th Annual Meeting of the Vibration Institute, Pittsburgh, amplitude of the varying signal, it may not give us the true
PA, June 2002. value of 250 mV. The signal resides for some time at the edge

18 SOUND AND VIBRATION/SEPTEMBER 2002

 4 200

Eng. Units
2 0
0 200
2
200
4 0
0.5 1 1.5
Time, msec 200

Amplitude, mV
0 100 200 300
9
Time, msec
8
Eng. Units

6 160
4 80
2 0
0
0 100 200 300 200
Frequency, Hz 100
0
Figure 1. The top trace is the time history of a signal containing one 0 250 500 750 1000 1250
discrete frequency component and a high level of noise. The bottom Frequency, Hz
trace is the spectrum of these data after 8 averages.
Figure 5. Two signals of equal amplitude are shown in the top two traces.
10.5 The spectra for these two traces are shown below. Note that the spectra
of the data in the top trace show a reduced amplitude and a widened
Engineering Units

10.0 peak. The bottom spectrum reads the amplitude correctly. Careful ex-
9.5 amination of the top time trace shows that the frequency is not constant
but varying.
9.0
8.5 within its bandwidth.
8.0
1 2 3 4 5 6 7 8 9 10
Trial
We Can Make the Vibration Disappear
This example is taken from Case Histories Power Indus-
Measured Data Input Amplitude Average try by Kevin Guy and published by the Vibration Institute. Un-
Ave. +3 Ave. 3 stable vibrations began occurring during the operation of a 650
MW turbine generator. Operation would continue for some time
Figure 2. The amplitude of the discrete signal as measured after 8 av- with the vibration levels on the turbine supervisory system
erages. The test is repeated ten times. The average of the peak read- showing 1 to 2 mils but then growing to 8.5 mils for about one
ings, the actual input data and the 3 levels are also shown. sec. As time went on, the levels became higher and in order to
prevent an alarm, the power level was decreased by 20 MW
4 which caused the vibration variations to be temporarily elimi-
Eng. Units

2 nated. After a period of time the high vibration excursions

0 would return and the power would again be reduced.
2 Vibration spectra were produced from the shaft riders that
4 were installed on the unit. When using only one average, the
0.5 1 1.5
Time, msec vibration could be higher on a few spectra while the majority
of the single spectra showed low vibration. A sub synchronous
8
vibration at 1575 CPM was observed. The peak was very wide
Eng. Units

6
and was exciting the shaft critical at 1575 CPM. If several av-
4
erages of the spectral data were taken, the vibration levels were
2
low and indicated a smooth running machine. In order to in-
0
0 100 200 300 vestigate the spectral width, a zoom spectrum was produced.
Frequency, Hz
This showed even lower levels and when several zoom spec-
tra were averaged, the vibration levels were even lower.
Figure 3. The top trace is the time history of a signal containing one
discrete frequency component and a high level of noise. The bottom The utility management decided that since this vibration
trace is the spectrum of these data after 64 averages. dropped below the alert level when averaged and was even
lower with a zoom spectrum, this was not the problem. Their
9.6 solution was to filter the data below 3000 CPM, which pro-
Engineering Units

9.5 duced uniformly low vibration readings.

9.4 Understanding the FFT process can explain the disappear-
9.3 ance of the vibration averaging and using a zoom spectrum.
9.2 First, the high vibration would occur on only a few of the non-
9.1 averaged spectra. This is consistent with the turbine supervi-
9.0
1 2 3 4 5 6 7 8 9 10
sory system. Averaging the data simply averaged one high vi-
Trial bration spectrum with several low vibration spectra resulting
in a low level of vibration in the averaged spectrum. The exci-
Measured Data Input Amplitude Average
tation occurred over a wide band so that when the bandwidth
Ave. +3 Ave. 3
was narrowed in the zoom spectrum, the levels became even
lower. The referenced case history did not describe what hap-
Figure 4. The amplitude of the discrete signal as measured after 64 av- pened next but one can be sure this was not the end of that
erages. The test is repeated ten times. The average of the peak read-
ings, the actual input data and the 3 levels are also shown. story.

of the bin, which introduces a picket fence error (reduces the What Answer Do You Want?
amplitude by 14% for a Hanning window) in the signal recon- The peak hold option of an FFT analyzer will store the maxi-
struction. A solution to the problem would be to reduce the mum amplitude that was achieved in a bin during the input
resolution of the spectrum by either selecting a higher fmax or time. It is touted as the way to get the amplitude portion of the
reducing the number of lines of resolution. This will increase bode plot during coast-downs or startups. This situation is the
the bin width, allowing for better containment of the signal epitome of non-stationary data. Using the FFT analyzer in this

SOUND AND VIBRATION/SEPTEMBER 2002 19

 20 20 20 20 20 20
Peak-Peak Peak-Peak Value:
1600 Lines 200 Lines 22.4 mils
Value:

Displacement, mils
Displacement, mils

22.3 mils

0 10 0 0 10 0

20 0 20 20 0 20
0 16 0 200 0 8 0 16 0 200 0 1
Time, sec Frequency, Hz Time, sec Time, sec Frequency, Hz Time, sec

Figure 6. The complete time trace of a coast down is shown in the left Figure 7. The complete time trace of a coast down is shown in the left
panel. The amount of this transient that is required to fill the input panel. The amount of this transient that is required to fill the input
buffer for a 1600 line spectrum is shown in the right panel. The middle buffer for a 200 line spectrum is shown in the right panel. The middle
panel shows the peak hold representation of the transient using 90% panel shows the peak hold representation of the transient using 90%
overlap. overlap.

case violates the rule that the data must be stationary and leads
to the question, Is it possible to get a good amplitude versus 5 4
speed plot during a coast down? The answer is yes, if one is
careful to set up the analyzer in a manner that will give it a 0 2
chance to get something close to accurate data. If the analyzer
is not set up properly, inaccurate or misleading results can be 5 0
4 4
obtained. Unfortunately, the results of an expensive FFT ana-
lyzer are usually trusted regardless of how it was set up. 2 2
The data presented here are from a classroom demonstration

Magnitude

Magnitude
given by Nelson Baxter in the Level 3 Vibration course, and is 0 0
used here with his permission. The first advantage we have 4 1
today is that many analyzers allow us to capture the complete
2 0.5
coast-down in the input buffer and then decide how best to set
up the analyzer to handle the data. Not so long ago, one had to
0 0
set up the analyzer correctly before the data were seen or use 4 1
an analog tape recorder to replay the data until the result was
closer to the actual case. 2 0.5
One of the advantages of modern analyzers is that they al-
low us to select the number of lines of resolution in the spec- 0 0
0 500 1000 0 100 200
trum. We can select from 200 to 1600 lines of resolution in most Time, msec Frequency, Hz
analyzers and have even more choices in some analyzers. One
would think that the highest resolution would be the proper Figure 8. Spectrum energy distribution as a function of pulse shape.
choice to exactly determine the critical speed.
Figure 6 shows a complete coast-down and peak-hold spec- tween the 1600 line and the 200 line spectrum. The 200 line
trum. The right panel shows the data needed to fill the input spectrum is very close to the actual peak to peak value as mea-
buffer for a 1600 line spectrum. The spectrum shows that the sured from the time trace at the critical speed. The actual RPM
peak to peak response at the critical speed is about 2.3 mils. If may now be determined by measuring the period of the data at
we did not observe the data that filled the input buffer and the critical speed.
remained confident with our firm belief that the analyzer can I would guess that there has been a lot of incorrect data re-
not lie to us, we might report this as the response of the ma- ported on a machines response as it passed through the criti-
chine while it coasted down through the critical speed. We look cal by analysts that only looked at the peak hold spectrum of
at the data in the input buffer that was used to produce the the data. This was particularly true before analyzers allowed
spectrum and we see two things. For 1600 lines it takes 8 sec one to store the complete coast down in the input buffer.
to acquire the data needed to produce the spectrum. During this
time both the amplitude and the frequency have changed dra- Bad Bearing Detection? Now it is Much Easier
matically. The peak hold value saved in the spectrum is low Back in the dark ages of vibration analysis in the 1960s and
because the data were within the bandwidth of the bin for a early 1970s, detecting a defective rolling element bearing
very short time. The peak to peak amplitude of the data in the sounded simple but unfortunately turned out to be very diffi-
input buffer also changed during the time taken to fill the cult. At the time, equations for the frequency of signals gener-
buffer. Clearly we have done something wrong. ated by a defective bearing depended on the location of the
We fell into the trap of thinking that higher resolution is bet- defect, the geometry of the bearing and the speed. Conceptu-
ter. To obtain high resolution requires a lot of data to fill the ally the bearing defect frequencies would be calculated and
input buffer, which takes time. We should have minimized the then the spectrum would be examined for a signal at these fre-
time to fill the buffer and not maximized the resolution. quencies. If a signal were there, then the bearing would be de-
Figure 7 shows the same coast down with a 200 line spec- fective. If no signal were detected, then the bearing must be
trum instead of the 1600 line spectrum. Instead of 8 sec to fill ok. The problem with this approach was that in many situa-
the input buffer it now takes only 1 sec. The panel at the right tions where no signal was detected, the bearing was known to
shows the data in the input buffer that is used to compute the be defective or worse yet, the bearing soon failed. In cases
FFT. Now looking at the peak hold coast-down shown in the where defective bearings were detected using this procedure,
middle panel, one can see that the peak to peak amplitude as the background noise in the spectrum was extremely low or the
the machine coasted down through the critical is 22.3 mils. bearing defect was very large.
Next, looking at the time trace of the data as it traverses the The reason for the poor success of this method lies in the
critical, one can measure the actual peak to peak value of the nature of the signal generated by the defective bearing. Figure
maximum amplitude which is 22.4 mils. Notice that there is 8 shows the spectra generated for several types of signals. The
almost a ten to one difference in the measured response be- top pair shows the spectrum generated by a sine wave, a single

20 SOUND AND VIBRATION/SEPTEMBER 2002

 2 1 cos sin
Impact Rate Hilbert Filter
Impact Rate Does (58 Hz)
Not Show Up in
Spectrum cos

Eng. Units
Eng. Units

0
Figure 11. Hilbert filter and analytic signal.

160

Eng. Units
80
0
0 1
0 1k 0 200 80
Frequency, Hz Time, msec 160
160

Eng. Units
Figure 9. A simulation of the time trace of a defective bearing and the 80
resultant spectrum of this signal. 0
80
160
0 0.5 1.0 1.5
Displacement, Mils

1.0
Time, sec
0.5 160

Eng. Units
120
0
80
0.5 40
Impact Envelope 0
1.0 0 50 100 150
0 Time, msec 1 Frequency, Hz
0.5
Acceleration, g

Figure 12. The top trace is a high frequency modulated by a sine wave.
The second trace is the envelope of the top trace using the Hilbert trans-
form. The bottom trace is the spectrum of the envelope.

0 400

Eng. Units
1 Frequency, Hz 2000 200
0.5 0
Acceleration, g

200
400
0 30 60 90
Time, msec
400
Eng. Units

200
0 0
1 Frequency, Hz 500 200
400
0 250 500 750 1000
Figure 10. The top trace shows the simulated impact and ring down of Time, Sec
a defective bearing. The middle trace is the spectrum of this impact and 150
ring down. The second signal in the top trace is the envelope of the
Eng.Units

impact and ring down. The bottom trace is the spectrum of the enve- 75
lope.
0
0 100 200 300 400 500
peak. The next pair of signals shows the spectrum generated Frequency, Hz
by a square wave. The energy of the signal is shared by all the
odd harmonics of the repetition rate of the square wave. The Figure 13. The top trace is the signal from a defective bearing. The sec-
ond trace is the envelope of the top trace using the Hilbert transform.
fundamental is still the largest signal in the spectrum. The third The bottom trace is the spectrum of the defective bearing.
pair shows the spectrum for a much shorter duration pulse than
the square wave. The fundamental is still the largest signal in amplified by a resonance. It is quite common to see a bearing
the spectrum but much more energy is shared with all the defect pulse excite a structural resonance but this usually hap-
higher harmonics. The fourth pair of signals shows a very short pens at much higher frequencies where very small motions can
impact in the time trace. This spectrum shows that there are produce very high g levels.
many harmonics of the fundamental and all are very low in Figure 10 shows three traces. In the top trace a simulated
amplitude compared with the time trace amplitude. This is the bearing defect impact train is shown. In the middle trace the
typical spectrum generated by a hammer in modal analysis spectrum of the impact train is shown. Note that there are very
testing. low amplitude signals at the defect frequency in the lower fre-
If the shape of the defect time trace were a sine wave, then quency portion of the spectrum. The data at the high end of
one should have very little trouble in detecting the defect sig- the spectrum are the actual ring down frequency of the train
nal in the spectrum. On the other hand, if the time trace is more of impacts shown at the top. This resonance is very strongly
like a hammer impact, then we can see that the amplitude of modulated by the defect signal. One could determine the pres-
the fundamental is very low and shared almost equally with ence of the defect signal from the sideband spacing of the high
all harmonics. We can imagine that a ball impacting a pit in frequency data. In real life cases, machinery noise obscures this
the raceway would produce a signal more like the impact ham- sideband structure or possibly more than one resonance is
mer than the sine wave. Figure 9 shows the spectrum and the excited and overlaps the first resonance. In any event, detect-
time trace for a defective bearing. This figure was furnished ing the low frequency defect frequency or detecting the defect
by Nelson Baxter from the Level 3 Vibration course. Note that frequency by the resonant sidebands is very difficult. One
the time trace shows a very sharp impact, meaning that the en- could measure the time between the impacts in the top trace
ergy will be spread among many harmonics, as the spectrum and convert this period to a frequency to show that it is the
shows. The fundamental defect frequency and its harmonics defect frequency of the bearing. This may often be done when
are barely detectable. If a little machinery background noise viewing a bearing defect in the time domain. Again, sometimes
were added to the time trace, one can see that the evidence of machine noise and other resonances preclude making an ac-
a defect could easily be obscured. Some early reports on the curate measurement of the impact spacing.
use of this technique claimed that the defect showed up in the Suppose one could beef up the thickness of the impact pulse
third harmonic while others reported seeing evidence in the and also remove the ring down frequency. This is done in the
seventh harmonic. What was happening in these cases was second trace in the top chart. The line drawn is called the en-
much like the impact hammer modal test the harmonic was velope of the train of pulses and may be produced by a simple

SOUND AND VIBRATION/SEPTEMBER 2002 21

diode, capacitor and resistor network. The circuit must have a 6
very fast rise time while the decay time is set by the time con- Raw Signal
stant of the RC circuit. The spectrum of the envelope is shown
in the bottom trace. The defect signal is now clearly identified.
One might wonder what the harmonics of the defect frequency

Eng. Units
mean. The answer is that they have no physical significance 0
but only define the decaying pulse shape which is controlled
by the R and C, not by the condition of the bearing. This tech-
nique has dramatically improved the ability to detect defec-
tive bearings and has become one of the standard methods.
6
Software Envelope Detector 6
We all know that a function such as an envelope detector, Signal After 48 Averages
implemented in hardware, costs money to originally design and
test. In addition, it costs money for parts and labor every time

Eng. Units
we make one. If we could duplicate the function in software,
we would still have to design and test it, but it would not cost 0
anything to build. We can charge enough for the software func-
tion to pay the nonrecurring costs. Since there are no recurring
costs, we should increase our profit. In addition, a software
solution does not soak up any power, weighs nothing and uses
no space. Therefore, all other things being equal, a software 6
6
solution is preferred. The manufacturers make more money and
Signal After 256 Averages
the user has a smaller, lighter unit.
It is possible to make a software envelope detector utilizing
a technique called the Hilbert transform. The Hilbert transform

Eng. Units
shifts the phase of every signal in the input by 90 degrees,
0
shown schematically in Figure 11. By utilizing the original and
shifted data one can see that we have a situation analogous to
real and imaginary signals. The amplitude of this signal may
be obtained by taking the square root of the sum of the squares
at each sample of the data. For example, if the signal had a fixed 6
amplitude, this technique would produce a fixed amplitude DC 0 40 80 127
value. If the original signal varies with time or has an enve- Time, msec
lope varying with time, the output of the combination will be
Figure 14. Shows the effect of time series averaging for 1, 48 and 256
the envelope of the input signal. averages.
Figure 12 shows a high frequency signal modulated by a low
frequency sine wave. The second trace is the result of shifting the result of 128 synchronous averages. This trace clearly
the phase of the signal by the Hilbert transform and then com- shows that one pulse is different from the others, such as one
bining the real and imaginary signals to get the amplitude. tooth mesh in the signal generated in a gearbox. The bottom
This amplitude is the envelope of the signal or the modulat- trace is the spectrum of the averaged signal. It is almost impos-
ing signal. The bottom trace is the spectrum of the envelope. sible from this spectrum to determine that one tooth mesh is
Figure 13 shows the same sequence of signals. The top trace different from the others.
shows raw data from a defective bearing. The second is the en- Figure 16 is a plot of the averaged noise amplitude versus
velope of the defect signal and the bottom is the spectrum of the number of averages. Notice that this is a smooth curve
the envelope. The fmax of the spectrum is high since the reso- where the noise level is inversely proportional to the square
nance excited by the defect impacts is a high frequency. To root of the number of averages. For example, the fourth aver-
examine the spectrum of the defect signal, a zoom Hilbert trans- age is one over the square root of 4 times the first average or a
form could be performed. This has the effect of moving the ratio of 1:0.5 It is thus easy to predict the signal to noise ratio
defect signal to the right in the spectrum. The disadvantage of from the number of averages.
the zoom spectrum is that it destroys the typical defect type Figure 17 is the same type of plot but here the signal is an
signals in the top two traces. unsynchronized sine wave instead of random noise. For the
first few averages, the amplitude drops dramatically as the
Time Series Averaging Extracting Signals from Noise unsynchronized signal drifts out of phase with the first signal.
What is required for time series averaging is both the noisy Then the amplitude increases as the signals drift back in phase.
signal and a pulse synchronized to the data. One possibility is To ensure a good ratio between a synchronized and a non-syn-
to use a key phasor marker on the input shaft of a gear box and chronized signal one must either take a great number of aver-
a signal from an accelerometer mounted on the gearbox. Al- ages or observe the running average and stop it when the de-
though most analyzers perform time series averaging in a more sired ratio is achieved or the non-synchronized signal has
sophisticated manner, it is easy to picture if one envisions con- disappeared. Figure 18 shows the results of 16 trials of 8 aver-
tinually adding the synchronized signal to itself and then stop- ages of a non-synchronized signal. Note that the resulting
ping the process and dividing the sum by the number of addi- amplitude could vary from 30 to 160.
tions made to arrive at the average value. One can also see that
adding a random noise signal to itself (random phase and sign) Time Series Averaging Can Also Fool You
and eventually dividing the sum by the number of additions To illustrate that there is no one magic data processing tech-
will result in a noise signal that will approach zero as the num- nique, Figure 19 shows data from a single mesh gear box. The
ber of averages becomes large. Figure 14 shows this process for input shaft is bent so that the contact point of teeth varies
1 average (raw signal), 48 averages and 256 averages. The syn- around the pitch line circle in a sinusoidal manner. This modu-
chronized signal emerges from the noisy signal as more aver- lation is clearly shown in both the top time trace and the top
ages are taken. spectrum. Note that if the time series averaging were synchro-
Figure 15 dramatically shows the value of time series aver- nized to the output shaft, all evidence of the modulation dis-
aging. The top trace is the original signal. The second trace is appears. This illustrates the fact that one should not get locked

22 SOUND AND VIBRATION/SEPTEMBER 2002

 200

Unsynchronized Signal Amplitude

6
Raw Signal 180
160
140
120
Eng. Units

100
0 80
60
40
20
0
0 10 20 30 40 50 60 70
6
Number of Averages

4
Signal After 128 Averages
Figure 17. Averaged amplitude of unsynchronized signal vs the num-
ber of averages

160
Eng. Units

Unsynchronized Signal Amplitude

0 140

120

100

80
4
0 80 160 255
60
Time, msec
40
1
Spectrum of Averaged Data 20
0 2 4 6 8 10 12 14 16
Number of Trials
Eng. Units

Figure 18. Sixteen trials of 8 averages of unsynchronized signal.

2
a

0
0 40 80 128
2
Frequency, Hz 2
b
Figure 15. Time series average is used to recover a signal from noise
and display it in the time domain. 0
Acceleration, g

1.0
2
0.9 0 200 400 600 800 1000
Averaged Noise Amplitude

0.8 Time, msec

0.7 2
c
0.6
0.5 1
0.4
0.3
0
0.2 2
0.1 d
0 10 20 30 40 50 60 70
Number of Averages 1

Figure 16. Averaged amplitude of noise versus number of averages.

0
into one data processing technique. One must look at time trace 0 100 200 300 400 500
data and spectral plots. If time series averaging is used, it is Frequency, Hz
helpful to obtain plots where the data are synchronized once
Figure 19. 128 time series averages of a modulated signal when syn-
to the input shaft and once to the output shaft. chronized to the input and output shaft.

The Inverse FFT This is the displayed spectrum. IFFT the 512 real spectral val-
The inverse FFT (IFFT) is a very powerful tool that is not of- ues and 512 imaginary spectral values to get the 1024 time data
ten used. For example, consider a Fourier transform process samples from the original. This would just be an interesting
starting with 1024 input samples. The FFT computes a 512 line process except that the spectral data may be modified before
spectrum, but to reduce the danger of aliasing, generally only the IFFT is performed.
400 lines are kept and displayed. If all 512 lines and the 512 Figure 20 (top trace) shows the original time trace of the
phase angles are kept, then the inverse FFT may be performed. combination of three discrete signals and noise. The next trace
The process is as follows: is the 512 line spectrum. In the third trace, all data are set to
FFT the 1024 time domain data samples to get 512 real spec- zero except in the immediate vicinity of the first discrete sig-
tral values and 512 imaginary spectral values. Then process nal. The final trace is the IFFT of the modified signal and shows
that data to get 512 spectral amplitudes and 512 phase angles. a relatively clean sine wave.

SOUND AND VIBRATION/SEPTEMBER 2002 23

 4 3
a a

4 3
0 time 0 time

1 b 1.5
b
Engineering Units

Engineering Units
0 0
0 fmax 0 fmax

1 c 1.5 c

0 0
0 fmax 0 fmax
2.5 3
d d

0 0

2.5 3
0 time 0 time

Figure 20. Using the FFT and IFFT as a filter: (a) original time trace; Figure 22. Removing noise from a square wave signal: (a) mixture of a
(b) 512 line spectrum; (c) all data set to zero except near the first dis- square wave and noise; (b) spectrum of the top trace; (c) all bins set to
crete signal; (d) IFFT of the modified signal. zero except the three harmonics; (d) IFFT.

4
Figure 21 is the same process except that all FFT bins have
a been set to zero except for the single bin containing the dis-
crete signal. The IFFT now produces a pure sine wave. This
process may be used to filter the data using very narrow band
0 pass filters. Realize however that if an original signal is modu-
lated and/or is a shape other than a sine wave, then additional
bins must be included to obtain the original signal. One must
4 include sideband bins for modulated data and harmonics for
0 time
non-sinusoidal wave shapes.
1
b Figure 22 illustrates this point for a non-sinusoidal wave
shape. The top trace is a mixture of a square wave (harmonics
1,3 and 5) and noise. The next trace is the spectrum of the top
time trace. In the next spectrum, all bins have been set to zero
except the three harmonics. The final trace is the IFFT, which
Engineering Units

now clearly shows square wave character.

0
0 fmax

1
Conclusion
c
The Fourier transform was originally conceived to determine
the harmonic content of a given repetitive wave shape. Noisy
signals and transient phenomena were far from Fouriers mind
in his original formulation. The Fast Fourier Transform algo-
rithm is now in common use in the vibration analysis of rotat-
ing machinery vibration. We commonly encounter signals that
0
0 fmax are not bin centered, very noisy or exhibit transient phenom-
2.5 ena. The FFT may still be used in these cases to gain insight
d
into the condition of the machinery, but it must be realized that
one is skating on very thin ice and serious errors can easily
0 occur for the careless. The cases discussed in this article show
both sides of the question. Many errors can be encountered by
carelessly following rules of thumb and not understanding the
FFT process. On the other hand, careful, knowledgeable use
2.5
0 time of the FFT can provide very valuable insight about the condi-
Figure 21. Using the FFT and IFFT as a filter: (a) original time trace; tion of machinery in noisy and transient conditions.
(b) 512 line spectrum; (c) all bins set to zero except for discrete signal;
(d) IFFT of the modified signal. The author can be contacted at: [email protected].

24 SOUND AND VIBRATION/SEPTEMBER 2002