Time Waveform Analysis 

  
Many fault conditions reveal themselves via the time waveform, yet many vibration analysts
do not even save the time waveform.  I hope this article will convince you to start saving
and analyzing the time waveform.  You might be surprised how useful it is! 

Time waveform analysis is an important analysis tool.  Time waveform data should be
collected on every machine on every route. Sadly most analysts do not collect time
waveform data, and even those who do collect the data, many believe the waveform
patterns are too complicated, or that the information is available in the spectrum.

In reality, time waveform data is easy to collect, easy to interpret, and can contain
information that is not available in the spectrum.

What are time waveforms?

When you collect vibration data from a bearing, the data collector first digitizes the analog
signal from the accelerometer, and then performs the Fast Fourier Transform (FFT) to
produce the spectrum.  The data collector will repeat that process a number of times,
averaging the spectra together, resulting in one averaged spectrum.  That spectrum is
stored in the collector for later analysis.  Many people believe that the spectrum contains all
of the information that you could possibly need from the machine.  That is not true.

The time waveform is the digitized signal that is used by the data collector to perform the
FFT calculation.  For example, if you collect an 800 line spectrum, the data collector will
collect a time waveform with 2048 samples (numbers).  Normally this data is discarded by
the data collector after the FFT calculation has been performed.

A time waveform

If you opt to save the time waveform for later analysis, the block of 2048 numbers are also
saved in the data collector.  That data will be saved in the software database and can be
viewed on screen.

Once upon a time data collectors had limited memory, so analysts would opt not to save the
time waveform in order to conserve memory.  But most modern data collectors have ample
memory, and a time waveform will only a few kilobytes – which is next to nothing in this
day and age.
Some people are also under the impression that it takes longer to collect time waveform
measurements.  However, given that the data collector has to collect the time waveform
anyway, it does not take any extra time at all.  (Later we will discuss reasons why you
might like to collect a separate time waveform, but it is a very fast measurement in any
case.)

What is time waveform analysis?

Time waveform analysis is the process of studying the waveform data in order to look for
patterns that may indicate a fault condition.  Time waveforms show precisely how the
vibration changed from one split-second to the next.  Every time gear teeth mesh together,
or balls roll across cracks in bearing raceways, or bubbles implode during cavitation, the
vibration will change – just for a split-second.  The time waveform captures those
instantaneous changes.

Why perform time waveform analysis?

This is not the place to explain the FFT process; however it is worth saying that the FFT
process is excellent for capturing linear (smooth) periodic events.  Events in the data that
are linear and periodic will be represented by peaks in the spectrum.  For example, if a
machine is out-of-balance, or there is a misalignment problem, or a rotor is eccentric, the
vibration amplitude rises and falls periodically.

When we see ‘clean and strong’ peaks in the spectrum we can be sure that that there were
sources of periodic vibration.  If we see sidebands, it is likely that the period source of
vibration itself periodically rose and fell in amplitude.  In this instance, time waveform
analysis is useful, but we can do without it.

 
 Figure x: A simple spectrum computed from the smooth time waveform

However, when we see harmonics, a raised noise floor, broad peaks, ‘haystacks’ and/or ‘ski-
slopes’ in the spectrum, the chances are that the vibration was not entirely smooth (linear)
and periodic.  Instead there may have been impacts, transients, random bursts of energy,
varying machine speed and/or other sources of non-periodic or non-linear vibration.

Spectrum with harmonics and a slightly raised noise floor

 Time waveform with a periodic series of impacts

Most analysts will attempt to interpret these patterns from the spectrum alone.  For
example, a raised noise floor may be interpreted as looseness.  A ‘ski-slope’ may be
interpreted as a bad transducer. A raised area around the pump vane rate frequency may
be interpreted as cavitation.  And ‘haystacks’ may be interpreted as a bearing fault.

Now, the analyst may well be correct.  From past experience and other evidence the analyst
may be able to diagnose the condition from the spectrum.  But every analyst should be
willing to consider methods that increase the likelihood of an accurate diagnosis, and time
waveform analysis is one such method.

Assuming the measurement setup is satisfactory, the time waveform will provide clear
indications of impacts, transients, modulation, beating, and bursts of energy.  Time
waveforms can help you to diagnose a wide range of fault conditions including bearing
damage, looseness, gear damage, cavitation, and the reason for measurement errors.

Measurement setup
The way in which the time waveform is measured has a large bearing on how useful that
data will be.  Of course, you can always go back out to the machine to collect more data,
but it would be best to collect everything you are likely to need during the routine test.

There are three ways to collect time waveform data:

1. The easiest option is to simply save the waveform that was used to compute the FFT
(spectrum).  The data collector has the waveform in memory, so you could just save
it for later analysis.  As I will discuss momentarily, the ‘default’ time waveform may
or may not be useful.

2. If the ‘default’ waveform is not ideal for analysis purposes, some data collectors
allow you to add a separate waveform test to your route setup.  Once the spectrum
measurement is complete the data collector will automatically acquire a time
waveform with different settings.  It should not add very much time to your route.

3. For completeness, I should mention that some data collectors allow you to perform
special tests which capture a time waveform that may represent many minutes of
vibration data.  This test may be used to capture the vibration from a machine as it
is starting up or shutting down, or during some other transient event.  The software
should then allow you to analyze the time waveform data in its entirety and extract
 small ‘blocks’ of data for detailed spectrum or time waveform analysis.  This option
will not be discussed any further in this article.

The settings used to acquire the spectrum dictate the settings used to collect the time
waveform.  Those settings in turn dictate what you will learn from the time waveform.  Let’s
have a look at a few examples.  The first time waveform was acquired when the spectrum
was set up for 800 lines and an Fmax of 500 Hz.  There are 2048 samples over 1.6 seconds.

The settings were changed to 6400 lines, with the same Fmax.  Now there are 16384
samples over 12.8 seconds.  The box overlaid on the graph indicates just 1.6 seconds and
2048 samples – i.e. the previous waveform.

So you can see that the second waveform reveals something that is very important (the
beating), however it did require an additional 13 seconds of data collection time.

Understanding the settings

Now, this is where time waveform analysis becomes a little tricky…

When you select an Fmax setting for your spectrum, the data collector determines the
appropriate ‘sample rate’ – the number of samples collected per second.  For example,
when you select an Fmax of 1000 Hz the sample rate is 2560 samples per second.  All you
have to do is multiply the Fmax by 2.56.

When you select the lines of resolution (LOR) for your spectrum, the data collector
determines the number of samples that must be acquired in order to compute the
spectrum.  For example, an 800 line spectrum is computed from a time waveform with 2048
samples.  All you have to do is multiply the number of lines by 2.56.

So, now we know the number of samples that need to be collected, and we know the
sample rate (how many samples are acquired per second), we can therefore determine the
length of time that it takes to acquire the time waveform.  Without boring you with all of the
equations, the solution is to simply divide the lines of resolution by the Fmax.  So, if we
choose a resolution of 800 lines, and an Fmax of 1000 Hz, the time waveform will take 0.8
seconds (800 ÷ 1000).

Length of measurement = LOR/Fmax

As you can see (and you probably already knew this), if the Fmax is higher, the
measurement takes less time.  If you increase the resolution the measurement takes
longer.  But why is this important?

We will use the time waveform to study the events that occur while the shaft is turning, the
gears are meshing, and the balls are rolling around the race way.  We want to collect
enough data to see at least ten rotations of the shaft, and we wish to collect enough data to
see the events in detail.

One unfortunate situation is that the time span and resolution of the time waveform is
dictated by the frequency span and resolution of the spectrum.  Most people set their
collection parameters based on the desired frequency span (based on the speed of the
machine), and not on the time waveform.

The frequency range and the span of the time waveform are inversely proportional.  Here is
the actual relationship.

Time span = Number of samples/(Frequency span * 2.56)

Frequency span = Number of samples(Time span * 2.56)

For example, for an 800 line spectrum (2048 samples in the time waveform), if you have a
high frequency range (3000 Hz), then the time span will be small (0.267 seconds).  If you
have a small frequency space (300 Hz) then you will have a longer time span (2.67
seconds).

Let's look at two examples – samples taken from the same machine at different rates.  In
this first case we have a full 6 seconds of time waveform data, and we have a 3200 lines
spectrum.  We can see the low frequency beating or modulation in the time waveform.
 Black liquor DIL #2

At the other extreme, if we set up for a 1.6 kHz frequency span and just 800 lines, the data
looks quite different.

 
 

To look at the time waveform you would have no idea that beating or modulation was
occurring; however we can see random looking transients that may indicate looseness.

Resolution and time span are everything in time waveform analysis.  If we wish to analyze
high frequency events such as transients then we need a very short time span.  If we wish
to be able to see modulation, then we need a wider time span.  And if we wish to see long
term changes in the vibration levels, we need an even wider time span.

We can attempt to quantify this (Catlin, 1987).  If we consider a machine with a rotation
rate of period T, then we need a time span of T/100 in order to view the details of high
frequency events (transients).  For example, a machine running at 1800 RPM (1500 RPM)
we would want a frequency span of .000333 seconds (0.0004 seconds).

If we wish to see multiple high frequency events in the time waveform, we need a span of
T/10.  For example, a machine running at 1800 RPM (1500 RPM) we would want a
frequency span of .00333 seconds (0.004 seconds).

If we wish to capture one cycle of the shaft, ideal for performing balancing and phase
analysis, we would have a time span of T.  For example, a machine running at 1800 RPM
(1500 RPM) we would want a frequency span of .0333 seconds (0.04 seconds).

To capture multiple cycles of rotation, we would sample at 10T.  For example, a machine
running at 1800 RPM (1500 RPM) we would want a frequency span of .333 seconds (0.4
seconds).  We have a sample of data to illustrate this point.  The machine runs at 1776 CPM
(29.6 Hz) and the period is therefore 1/29.6 or 0.034 seconds.

 
 Black liquor DIL #1

To capture a larger number of cycles in order to see changes over time – in particular to be
able to note phase relationships within the waveform, you need approximately 10T. 
Although this sample is not a full 10T, it does illustrate some of the patterns you can see
when you have a larger number of cycles.

Blower M&D Motor 7/30/99

And finally, to view longer term trends and low frequency changes as a result of modulation
or beating, you should set up a time span of approximately 100T.  Although this sample is
not a full 100T, it does show the benefit of having a longer time record.

Black liquor DIL #2 high res

 
Measurement directions and storage
The other question you will have to ask yourself is whether you will collect time waveform
data with every spectrum you collect, or just in a single axis on any one component.

In the author's opinion, the "cost" associated with collecting time waveform data is so small
that you should always collect time waveforms.  The time waveform has to be collected in
order to compute the FFT (the spectrum), so you may as well just keep it.

Even if you routinely perform a special test in order to acquire a waveform with a longer or
shorter time record, the time required is minimal.

Although time waveform data requires more memory than a spectrum, most data collectors
and computers have more than enough memory, and the transfer time (data collector to
computer) is not great.

Performing the analysis

There are basically three reasons why we analyze time waveform data.  First, you see
phenomenon in the time waveform that cannot be seen in the spectrum.  Second, there are
many conditions that are well supported by viewing the time waveform.  And third, we can
use the time waveform to perform useful calculations.

Detecting and diagnosing a fault

We will often see evidence of a fault condition in the time waveform that we cannot see
from the spectrum.  Non-periodic events (like transients and pulses) will not appear in the
spectrum – that is, you would not know that there have been pulses by looking at the
spectrum alone. 

Also, broken teeth in gearboxes stand out clearly in the time waveform (especially if it has
been triggered) whereas they are almost impossible to detect in the spectrum.

For this reason, with components like gearboxes, it is important to review time waveforms
even if review of the spectra reveals little.

Diagnosing a fault 
Likewise, there are many fault conditions can are indicated in the spectrum, but are well
supported by review of the time waveform.

For example, looseness will appear as harmonics in a spectrum.  There are other reasons
why one may view harmonics though.  A review of the time waveform may reveal random
impacts/transients, a sure sign that there is looseness.
 Black liquor DIL #2

Wear in gears and misalignment is also visible in the spectrum as a peak at the gear mesh
frequency with sidebands of rotation rate (of the offending gear).  It is evident in the
spectrum, but the extent of the damage may be best assessed through review of the time
waveform.

 
 

If the inner race of a bearing is pitted or cracked, there will be impacts at the inner race
rotation rate, but there will also be modulation as the crack moves in and out of the load
zone.  You will often see the sidebands in the spectrum, however the time waveform will
also give a clear indication.

Air washer #1

In this example you can see the bursts of energy due to cavitation or air ingestion into the
pump.  The humps in the spectrum are the result. 
 Effect 3rd transfer pump

In this example, taken from iLearnHandsOn, we have a belt fault.  Although there are
harmonics in the spectrum, the fault is clearly visible in the time waveform.

 
Analysis techniques
We have now learned two reasons why it is important to look at the time waveform, but we
also need to actually analyze it and extract information that we can use in a diagnosis.  The
key feature we will cover for now is simply the time between two events, and thus the
frequency.

Let's take a simple example first.  Here we can clearly see a strong sinusoidal signal.

Ash hopper sluice pump

If we place the cursor on the time waveform we will get a time relative to the beginning of
the record.  We don't really care what that time is, however it does allow us to measure the
time difference between two events.
Let's take the time at the top of two cycles.  We can see that the samples are at 0.0141
seconds and 0.0308 samples.  The delta is 0.0167 seconds.  We know that the frequency is
the inverse of the period, so the frequency must be 1/0.0167 or 59.88 Hz or 3593 CPM.

The machine in questions rotates at 3593 CPM.  If we now look at the spectrum, we can see
that there is a clear peak at this frequency, just as we would expect.

Let's take a look at a slightly more complicated waveform.  If we revisit a sample we looked
at a moment ago, we can see that there are pulses in the time waveform.  If we look at the
time of any two pulses, and then compute the difference, we therefore have the time
between the two events.

 
In this case the difference is 0.0109 seconds.  This would suggest that the period of this
signal is 0.0109 seconds.    The frequency is therefore 1/0.0109 or 91.74 Hz (5504 CPM). 
The running speed of this machine is 1776 CPM, so this signal is 3.099X the running speed. 

In the spectrum we also find a peak at this frequency.  In this case we had verification –
there was a peak in the spectrum at the same frequency.  This is not always the case, and
that's why this method is so powerful.