Module 3

Syllabus
 Data Acquisition & Signal Processing

 Classification of signals, Signal analysis, Fast Fourier Transform (FFT),

 Essential Settings in Data Acquisition System

 Plot Formats, Frequency Span and Frequency Resolution,

 Average Types and Number of Averages,

 Windowing, Spectrum Scaling,

 Signal conditioning.

Classification of signals
 Signals essentially convey information.
 Signals are often distinguished by the repetition frequencies of periodic events and
so one of the most fundamental ways of evaluating signals is in terms of their
‘frequency spectrum’, showing how their constitutive components are distributed with
frequency.
 Mathematically, this is done with various forms of Fourier analysis, but at this stage it
is sufficient to see how the various signal types manifest themselves in the time and
frequency domains.
 Features of signal can be constant over a period of time or vary with time. Based on
this condition, signals can be classified as stationary or nonstationary.
 Figure3.1 shows the basic breakdown into different signal types.
 Figure 3.1: Types of Signal in Vibration Monitoring

Figure 3.2: Types of signals with examples

Stationary Signal
 The signals whose statistical features do not change with time and are known as
stationary signals. i.e. statistical properties are invariant with time.

Stationary Deterministic Signal

 Stationary deterministic signals have specific distinct frequency components.
 Signals from rotating machines operating at a constant rotational speed are examples
of stationary deterministic signals.
 Deterministic signals basically means that they are composed entirely of discrete
frequency sinusoids and thus their frequency spectrum consists of discrete lines at
the frequencies of those sinusoids.
  Once the frequency, amplitude and initial phase (i.e. at time zero) of these components
is known, the value of the signal can be predicted at any time in the future or past;
hence the term ‘deterministic’.

Stationary Random Signal

 Random signals are somewhat more complex, as their value at any time cannot be
predicted, but for stationary random signals their statistical properties are unchanging
with time.
 Random deterministic signals are well characterized by their statistical features, such
as mean, standard deviation, variance, and so forth.
 Many real-world signals are random, for example, noise produced by rain drops on a
roof.
 Unlike deterministic signals, the spectral content of random signals is continuous.
 Individual random signals must be considered as realizations of a ‘random
process’, where all realizations vary randomly, but are equally valid.

Figure 3.3: Ensemble Averaging

 The statistical properties can be obtained by averaging across an ‘ensemble’ of

realizations, as illustrated in Figure3.3.
 The conditions for stationarity, for measurements on a machine, are typically that the
latter is operating at constant speed and load.
 If the function being averaged using the expectation operator E [·] is the signal itself,
that is fx(t) = x(t), then the result of the average will be the mean value. If fx(t) = x 2(t),
the result will be the mean square value.
 One rarely has a large number of realizations of a process and never an infinite number,
so it is convenient to be able to perform the averaging along the record.
  This is valid if the signals are not only stationary but also ‘ergodic’. The fundamental
meaning of this is that all realizations are statistically equivalent.
 The signals depicted in Figure3.3 might for example be vibration signals measured on
a number of vehicles driven at constant speed around a uniform test track.
 If the vehicles varied from small cars to large trucks, it is quite possible for the process
to be stationary (i.e. the mean value at any time t to be constant), but for it to be ergodic,
all the vehicles would have to be of the same type. It is then clear that averages along
the record would have equal validity to averages across the ensemble.

Non Stationary Signal

 ‘Non-stationary’ means anything which does not satisfy the conditions for stationarity
i.e. the time domain statistical features of nonstationary signals change with time.
 The most common nonstationary signal is a transient type that is due to an impact like
the sound of something falling from a height or a slamming door. In some situations,
these signals can be of the continuous repetitive type, like the noise produced by a
jackhammer.
 It can be divided into two main classes, ‘continuous’ and ‘transient’. There is no hard
and fast rule for distinguishing between these two types, but in general it can be said
that transient signals only exist for a finite length of time and are typically analysed as
an entity. Once again, this requires clarification, since a decaying exponential function,
for example, theoretically decays to infinity, but in practical terms it only has a
measurable value for a finite time.
 The terms ‘energy’ and ‘power’ are used to distinguish between transient and
continuous (stationary or non-stationary) signals.
 An analogy can be drawn with electrical signals in a resistive circuit, where the power
W = EI, and E and I are the voltage and current, respectively.
 Since E = IR, where R is the resistance, the power is proportional to the square of the
voltage or current, that is
W = I R = E R
 Similarly, the true power associated with a vibration signal is related to the square of
its amplitude through some sort of impedance or admittance function, and it is common
simply to call the squared value the ‘power’.
 A transient signal has an instantaneous squared value or power at each point in time,
but is characterized by the integral of this ‘power’ over its whole length in time, this
being called its ‘energy’.

Cyclostationary Signal

 They have statistical properties which vary periodically.

  A typical example is the combustion signal in an internal combustion (IC) engine,
where there is a combustion event in each cylinder each cycle (thus happening
periodically), but with significant random variations from one cycle to another.
 As mentioned above, the different types of signals have different characteristics in the
time and frequency domains and these are summarized in Figure 3.4 for continuous
signals (i.e. stationary and cyclostationary).

Figure 3.4: Typical signals in the time and frequency domains

Signals Generated by Rotating Machines

 In condition monitoring, changes in vibration signals are ascribed to changes in condition, so
it is important that other factors which cause changes in vibration signals are considerably
reduced or eliminated.
 Vibrations tend to change with the speed and load of a machine, but signals generated by a
rotating machine operating at constant speed and load, for which the signals will typically be
stationary and/or cyclostationary.
 Occasionally, use can be made of non-stationary signals, such as those generated by a machine
under run-up or coast-down conditions, but such signals should be processed with the
appropriate analysis techniques, such as the time/frequency techniques.
Computer Aided Data Acquisition
 Data acquisition (commonly abbreviated as DAQ or DAS) is the process of sampling signals
that measure real-world physical phenomena and converting them into a digital form that can
be manipulated by a computer and software.
 Modern digital data acquisition systems (Figure 3.5, 3.6) consist of four essential components
that form the entire measurement chain of physics phenomena:
(i) Sensors
(ii) Signal Conditioning
(iii) Analog-to-Digital Converter
(iv) Computer with DAQ software for data logging and analysis

Figure 3.5: The elements of the modern digital data acquisition system (Source: DEWESOFT)

 To perform digital signal analysis (DSA) on a signal measured by a transducer mounted on

machinery, signal data needs to be collected.
 The signal measured by the transducer is analog in nature and continuous in time.
 The data is collected by a device known as the analog-to-digital convertor (ADC) and the
digital samples xi are stored in a memory space in the device for subsequent digital
computation.
 The purpose of the data acquisition system is to accurately represent the measured analog
signal to its corresponding digital values.

Figure 3.6: Configuration of data acquisition system for high-frequency signals

 The typical data acquisition system has multiple channels of signal conditioning circuitry
which provide the interface between external sensors and the A/D conversion subsystem.
 The following two important aspects of the data acquisition system need to be considered,
among others:
(i) Sampling frequency
(ii) Digital bit size

Sampling Frequency
 The data has to be sampled at an adequate sampling frequency. In Figure3.7, the original
sample adequately sampled by a sampling frequency of f = is represented by the dark line.
In the same figure, the dashed line indicates the same signal sampled at a slower sampling rate,
∗
f = ∗
.

Figure 3.7: Signal Aliasing Error

 Due to the slower sampling rate, the original signal appears to be a low-frequency signal. In
other words, the signal has been aliased as a low-frequency signal.
 This is a serious error in the data acquisition system, and is known as the aliasing error. To
prevent signal aliasing, the signal has to be sampled at a rate at least two times higher than the
maximum frequency of the signal present in the system. This fact is stated in Shannon’s
sampling theory.
 When data acquisition is done for a signal whose maximum frequency is not known, a low-
pass analog filter with a cut off frequency of is used to prevent signal aliasing, as shown in
Figure3.7.
 This is a very important fact to consider when using data acquisition for dynamic signals like
noise and vibration that change very quickly. Data acquisition devices without the low-pass
antialiasing filters are available for acquiring static signals over a period of time, for example,
the temperature signals from thermocouples.
 When there are multiple inputs to the ADC, each known as an input channel, the sampling
frequency is expressed as the number of data points per second per channel. A digital switch
known as a multiplexer is used to routinely scan the channels in a sequence for data acquisition
by an ADC.

Digital Bit Size

 Another important point is the corresponding digital value assigned to a digitized analog signal
by the ADC. The ADCs store the digital data in binary bits as a digital value corresponding
to 2n, where n is the bit size of the ADC. ADCs are available in many bit sizes of n = 3, 10, 12,
24, etc.
 The maximum analog voltage to a digital convertor is ±5 V or a range of 10V. Thus, the
minimum analog that can be detected by an ADC is given by Equation as

Range
Amplitude Resolution =
2

 For example, for a 3-bit ADC, there are a maximum of 8 digital values from 000 to 111, that
can be used to digitally represent the input analog voltage of 10V range. This corresponds to
an amplitude resolution of 1.25 V as per Equation.
 The problem arises when the analog voltage at a particular instance is at an amplitude
resolution less than 1.25 V.
 Thus the small voltage variations in the signal less than the amplitude resolution of the ADC
cannot be captured. This is shown in Figure3.8 and is known as the digitization error.
 To overcome the digitization error in the ADC process, it is advisable to have an ADC of
a higher bit size. For example, for the same input voltage range of 10 V, with a bit size of 12,
the amplitude resolution would correspond to 2.49 mV.
 Thus, very small deviations in the analog signal can be captured by the ADC conversion
process.
 Figure 3.8: Effect of bit size on digitization

 Many times during data acquisition, analog amplifiers are used before the ADC process to
amplify the signal so that the ADC device can capture the small changes in the analog signal.
 The input analog signal to the data acquisition system can be unipolar, where the signal is
referenced to a ground voltage using a single wire system, or it can be bipolar with reference
to a high and low value of the analog signal.
 The noise associated with the data acquisition process reduces with the bipolar input.

Data Storage
 The digital data thus obtained by the ADC process need to be stored in the digital memory for
further computations.
 Depending upon the number of lines of FFT required, the data size, N, will change. Some ADC
devices have on-board random access memory (RAM) to store the digitized data.
 The data acquisition process is controlled by driver software that is resident on the host
computer wherein the triggering of the data acquisition process can be initiated, based on a
certain input voltage level.
 The software can also control the rate at which the data is stored in the on-board memory, the
mode and time of data acquisition, and whether it should be in a continuous mode or
intermittent.
 Many standard commercial hardware systems are available for data acquisition along with their
driver software. The digital data thus acquired, needs to be transferred to the computer system
through a data transfer protocol, based on the architecture of the computer.
 Thus the compatibility of the ADC hardware with the computer system must be ensured. Some
of the standard computer architectures over the years that have been used for interfacing with
the ADC are ISA (Industry Standard Architecture), EISA (Extended Industry Standard
Architecture), PCI (Peripheral Component Interconnect), PCMCIA (Personal Computer
Memory Card International Association), and USB (Universal Serial Bus).
Signal Analysis

Fast Fourier Transform

 As all signals are acquired in the time domain, a waveform must be transformed to analyze
the data in the frequency domain.
 There are several standard methods to do so, but perhaps the most popular method is the
fast Fourier transform (FFT).
 The FFT is an algorithm that converts a digitized signal from the time domain to the
frequency domain. The time data are divided into frames of equal time length, and an FFT
is calculated for each frame.
 Following this linear transformation, engineers can observe the frequency content of the
time history waveform and gather information, such as excitation frequencies, peak
acceleration and distribution, and harmonic content.
 This principle states that any periodic signal (what we measure with vibration) can be
broken down into a series of simple sinusoids that, when combined, will generate the
periodic signal we have just analyzed. In practical terms that mean this process can generate
the spectrum we see here from a time domain signal it has analyzed.

Figure 3.9: Time signal consisting of sinusoidal components, and how it can be represented in the frequency domain
Development of FFT

Joseph Fourier (1768 – 1830)

Basics of FFT
1. FFT assumes time domain continues forever.
2. Number of points in time domain equals number of points in FFT.
3. The alias region in normally hidden.
• Mirror image about f /2 “Aliasing” (Where f is the sampling frequency)
4. Frequency Resolution = ∆f = f /N

Figure 3.10: Basics of FFT

Time Record
 A time record is the amount of time-domain data the analyzer needs to perform one FFT
operation. Essentially, the time record is a block of time-domain sample points.
 The time record and its FFT are the building blocks the analyzer needs for all subsequent
measurements.
 The actual Fourier Transform does not have explicit time or frequency references (it simply
operates on a sequential collection of points), FFT analysers must assign arbitrary start and
finish times for data to be transformed. These blocks of input data are called time records.
 For example, with the default display resolution of 401 frequency points, the analyzer takes
up to 1024 samples of time data to produce 512 points of frequency domain data.
 The analyzer usually displays the first 401 points of this data and discards the rest.
 The time record can be described by both a length and a size.
 The time record length is the amount of time required to acquire a time record and is altered
by changing resolution bandwidth, window, main length, or gate length.
 The time record size is the number of time points in the time record and is dictated by the
time record length in combination with the sample rate (and sample rate, in turn, is directly
related to span).

Figure 3.11: Time Record

Essential Settings in Data Acquisition System

Frequency Span and Frequency Resolution

(i) Frequency Span: It is the overall width of the FFT spectrum that we can see on the FFT
display (left to right).
 The maximum frequency of the FFT is half of the signal sampling frequency, but in the
upper region the results are never reliable, so the sampling result should be set to:

Sample Rate = Maximum Signal Frequency × 2 × 1.25

  1.25 is the absolute minimum factor for also getting the right values in the upper region of
the FFT.
 A factor of 1.28 is commonly used in signal analysis in order to obtain a 'nice' Analysis
Bandwidth (also referred to as Frequency Span).
 For example having a sample rate given by:

Sample Rate = 2 Hz = 32768 Hz

 Then,

Sample Rate 32768 Hz

f = = = 12800 Hz
2 × 1.28 2.56

 The factor of 2 comes from the famous Nyquist criteria (or more correctly from the
Nyquist–Shannon sampling theorem), which says that maximal signal frequency
adequately presented in the digitized wave is the half of the sampling rate.

(ii) Frequency Resolution: It is the minimum change in frequency that FFT can detect
 The result of FFT is a set of amplitudes of certain frequencies.
 The amount of amplitudes in the set is given by the Number of Lines parameter for the
FFT.
 The Number of Lines parameter is user-selectable, and it determines the resolution of the
FFT.
 Line resolution is a change in frequency between two frequency lines, which are extracted
from the signal and is calculated with the equation:

Sample Rate
Line Resolution = 2
Number of Lines

 So the question is, why not always use the maximum number of available frequency lines,
which gives more exact results?
 The answer is simple, because, with more frequency lines it takes more time to calculate
FFT spectra.
Number of lines × 2
Time to calculate =
Sample Rate
 If we combine above equations then we get,

1
Line Resolution =
Time to calculate
  The number of lines combined with the sample rate also defines the speed of the FFT
when non-stationary signals are applied.
 With more lines, FFT will appear slower and changes in signal will not be shown that
rapidly.

Frequency Bin
 Frequency bins are intervals between samples in frequency domain.
 For example, if your sample rate is 100 Hz and your FFT size is 100, then you have 100
points between [0 to 100] Hz.
 Therefore, you divide the entire 100 Hz range into 100 intervals, like 0-1 Hz, 1-2 Hz, and
so on.
 Each such small interval, say 0-1 Hz, is a frequency bin.

Windowing
• Let us first understand concepts of Spectral Leakage.

Spectral Leakage/ Spectral Energy Leakage

 Let us assume that we set ∆f = 1 Hz
o Condition 1: We want to observe a simple sine wave signal of 5V amplitude
and 3Hz frequency
 It will be represented as shown in Figure3.12

Figure 3.12: Representation of 5V; 3Hz signal

o Condition 2: We want to observe a signal of 5 V amplitude and 2.5 Hz

frequency
 One can say that we observe anything in spectrum as 2.5 Hz LOR is not
available as shown in Figure3.13 (a). But that essentially means there is
no signal; which is a wrong interpretation.
 Another possibility is that the amplitude will be distributed to adjacent
frequencies of 2 Hz and 3Hz respectively, as shown in Figure3.13 (b).
 (a) (b)
Figure 3.13: Assumptions in the signal representation of 5V; 2.5 Hz

 But, the actual FFT spectrum is something different from the above mentioned assumptions as
shown in figure 3.14

Figure 3.14: Smearing of spectral content

 This phenomenon is referred to as Spectral Leakage.

 Spectral leakage is takes place when the energy of frequency components in a signal is
spread out to a broad range of spectral lines instead of being represented by single
spectral lines.
 Figure 3.15: Leakage phenomena, where energy is spread to multiple spectral lines

 It happens when the spectral content of the signal does not correspond to an available
spectral line i.e. when analyzed signals contain energy at frequencies not described by
the spectral lines of the FFT spectrum.
 For example, configuring an FFT analyzer to a line spacing of 2 Hz; if the analyzed
signal contains energy at an uneven frequency like 10.5 Hz, for example, leakage will
occur. (Figure 3.15)
 Why does this happen? Because no single spectral line can describe the energy at 10.5
Hz when the line spacing is 2 Hz.
 This leakage phenomenon arises since FFT algorithms describe blocks of time data
with periodic sinusoidal components. Such a representation requires that time signals
are periodized into time blocks that are continuous at the ends where the blocks are
effectively joined into a loop.
 Remember that the FFT time block length T is defined by the reciprocal of the line
spacing:

1
T=
∆f

 Given a spectral line spacing of 2 Hz, the time block length is 0.5 sec. This causes all
non-even frequency components to have discontinuities at the looped block ends.
Adding to the example two paragraphs ago, 10.5 Hz will have a 90 o phase difference
between the block ends.
 If time signals are periodized to have a block length T that causes the time blocks to
have discontinuities at the ends, FFT algorithms will try to represent such
 discontinuities by leaking a portion of the energy to a broad range of sinusoidal
components.
 For most signal types it is hard or impossible to find block lengths with no
discontinuities at the looped time block ends, and therefore time weighting window
functions are used to help with solving this problem.

Main Lobe
 The centre of the main lobe of a window occurs at each frequency component of the time-
domain signal.
 By convention, to characterize the shape of the main lobe, the widths of the main lobe at –3
dB and –6 dB below the main lobe peak describe the width of the main lobe. The unit of
measure for the main lobe width is FFT bins or frequency lines. (Refer Figure 3.16)
 The width of the main lobe of the window spectrum limits the frequency resolution of the
windowed signal. Therefore, the ability to distinguish two closely spaced frequency
components increases as the main lobe of the smoothing window narrows.
 As the main lobe narrows and spectral resolution improves, the window energy spreads into
its side lobes, increasing spectral leakage and decreasing amplitude accuracy. A trade-off
occurs between amplitude accuracy and spectral resolution.

Side Lobes

 Side lobes occur on each side of the main lobe and approach zero at multiples of fs/N from the
main lobe.
 The side lobe characteristics of the smoothing window directly affect the extent to which
adjacent frequency components leak into adjacent frequency bins. (Refer Figure )
 The side lobe response of a strong sinusoidal signal can overpower the main lobe response of
a nearby weak sinusoidal signal.
 Maximum side lobe level and side lobe roll-off rate characterize the side lobes of a smoothing
window. The maximum side lobe level is the largest side lobe level in decibels relative to the
main lobe peak gain.
 Ideally, we would like a very narrow main lobe and very deep attenuation in side lobe.

Effect of Leakage
 As a result of the amplitude errors caused by spectral leakage, small frequency peaks will
occur close to larger ones.
 If there are two sinusoids, with different frequencies, leakage can interfere with the ability to
distinguish them spectrally.
 If their frequencies are dissimilar, then the leakage interferes when one sinusoid is much
smaller in amplitude than the other. That is, its spectral component can be hidden or masked
by the leakage from the larger component.
 But when the frequencies are near each other, the leakage can be sufficient to interfere even
when the sinusoids are equal strength; that is, they become undetectable.

Why leakage occurs during vibration analysis?

 The FFT is performed on a block of samples called the time record.
 One assumption made in the FFT calculation is that the time record is continuous. That is, the
signal just before the captured time record, and the block immediately after our time record are
identical.

Figure 3.17: Time record start and end at zero

 In this example, although we are performing the FFT on the block of data with the
black background, the FFT calculation "assumes" that the data continues endlessly
before and after this block of data - as shown with the data with a gray background.
 In this example it is true that the single frequency sine-wave begins and ends at zero
amplitude. Four complete cycles live within the time record.
 If we are analysing a pure sine wave, i.e. just one frequency, and there is an integer
number of cycles in the time record, then this assumption is correct.
 However it is seldom true that the time record starts and ends at zero. More commonly
they are similar to Figure 13.18.

Figure 1.18: Time record neither start nor end at zero

 When the FFT calculation is performed the signal is discontinuous. It seems to have a step
increase in level and looks similar to an impact to the FFT calculation.
  It generates a peak that is spread over a wide frequency band similar to an impact as shown
in Figure3.18. That is not what we want to see as a result.
 The real data in Figure 2 shows that the ends of each sample do not have the starting and
ending amplitude at zero.

Figure 2: Real data example where the ends of the sample blocks do not end at zero amplitude

Figure 3.20: A non-periodic signal resulting from sampling

When does leakage not happen?

 There are two possible scenarios that leakage does not occur.
(i) The first is that when the whole time capture is long enough to cover the complete
duration of the signals. This can occur with short transient signals. For example in a
hammer test, if the time capture is long enough it may extend to the point where the
signal decays to zero. In this case, data window is not needed.
 (ii) The second case is when a periodic signal is sampled at such a sampling rate that
is perfectly synchronized with the signal period, so that with a block of capture, an
integer number of cycles of the signal are always acquired.
a. For example, if a sine wave has a frequency of 1000Hz and the sampling rate
is set to 8000Hz. Each sine cycle would have 8 integer points.
b. If 1024 data points are acquired then 128 complete cycles of the signal are
captured. In this case, with no window applied you still can get a leakage-free
spectrum.

Concept of Window Functions

 Time weighting techniques add a “window” with individual weighting coefficients to each
time sample in an FFT time block.
 This primarily reduces those samples that cause spectral leakage. In effect, samples at the time
block ends are reduced to zero (or heavily-reduced), so that the discontinuities in the
periodized time signal are removed.
 Examples of some window function shapes are illustrated in the picture below:

Figure 3.21: Shapes of some commonly used window functions called Flat top, Hanning, and Rectangular

 The parameters characterizing window functions are listed below:

(i) NBW
(ii) Maximum amplitude error (also referred to as Ripple)
(iii) Selectivity
 Window functions do not completely remove spectral leakage, and side lobes will still
exist, but they are more attenuated.
 The selectivity parameters that describe such side lobe characteristics are illustrated in the
picture below:
 Figure 3.22: Side lobe Characteristics in Hanning filter

 Figure illustrates the parameters defining the filter Selectivity. The frequency axis is linear
- giving a curved shape for the Side lobe fall-off rate.

Types of Window Functions in Software

Types of Windowing Functions

Uniform Window Hamming Blackman Kaiser Bessel Exponential

Hann Window Flattop Window
(Rectangular) Window Window Window Window
Figure 3.23: Simulations of commonly used window functions, indicating their different characteristics which are used to
select which window to use for a certain analysis scenario (Source: DEWESOFT)

How to choose data window

 If a measurement can be made so that no leakage effect will occur, then do not apply
any window (in the software, select Uniform.). This only occurs when the time capture
is long enough to cover the whole transient range, or when the signal is exactly periodic
in the time frame.
 If the goal of the analysis is to discriminate two or multiple sine waves in the frequency
domain, spectral resolution is very critical. For such application, choose a data window
with very narrow main slope. Hanning is a good choice.
 If the goal of the analysis is to determine the amplitude reading of a periodic signal,
i.e., to read Apk, Apkpk, Arms or Arms2 , the amplitude accuracy of a single frequency
component is more important than the exact location of the component in a given
frequency bin, choose a window with a wide main lobe. Flattop window is often used.
 If you are analyzing transient signals such as impact and response signals, it is better
not to use the spectral windows because these windows attenuate important information
at the beginning of the sample block. Instead, use the Force and Exponential
windows.
 A Force window is useful in analyzing shock stimuli because it removes stray signals
at the end of the signal.
 The Exponential window is useful for analyzing transient response signals because it
damps the end of the signal, ensuring that the signal fully decays by the end of the
sample block.
  If the nature of the data is has a random nature or unknown, choose Hanning window.

https://www.youtube.com/watch?v=Q8N8pZ8P3f4
https://www.youtube.com/watch?v=pD7f6X9-_Kg

Averaging

 In an ideal world, the data collector would collect a single time record free of noise from a
never changing vibration signal, then produce the FFT and store it.
 But the vibration is constantly changing slightly and there is noise in the signal.
 Changes occur as rotating elements go through cycles and there is random noise from
inside and outside the machine.
 There is a way to minimize the effects of the noise and keep more of the changes due to
cycles inside the machine. The process used to correct this is called Averaging.
 Averaging can be performed in the time domain or in the frequency domain. But, in this
section, we will focus mainly on averaging in the frequency domain, which is the primary
type of averaging used with FFT analysers.
 Averaging in the frequency domain is sometimes referred to as Spectrum Averaging.
 FFT analyzers often have different options for setting up the spectrum averaging process.
 The most common averaging modes are described in the following sections

Figure 3: Types of averaging

RMS Averaging
 RMS averaging (also referred to as power-spectrum averaging or energy averaging) is
typically the default spectrum averaging mode in FFT analyzers.
  RMS averaging is used to reduce the fluctuation of spectral noise levels.
 With RMS spectrum averaging, the individual spectral lines are averaged over multiple
instantaneous power spectra or cross power spectra.
 In the picture below a non-averaged instantaneous power spectrum (red) is compared to an
averaged spectrum averaged over 100 power spectra (blue).

Figure 3.25: Comparison of an instantaneous spectrum with no averaging (red) and an RMS averaged spectrum over 100
instantaneous spectra (blue).

 When performing RMS averaging, the noise in the signal is averaged in the same way as
the pure/consistent signal. As a result, the noise is not reduced or averaged away by
spectrum averaging, but the spectral noise levels will become more and more steady
(averaged) with increasing numbers of averaged spectra.
 The standard deviation of the random noise in RMS averaged spectra will be reduced by a
factor of 1 , where N is the number of averaged instantaneous spectra. This is a
√N
reduction of the standard deviation of -5 dB each time the number of averages is 10 times
greater. Conversely, the measurement time will increase as the number of averages
increase.
 RMS averaging calculates the mean power sum which relates to the mean energy. The
square root of the mean power sum is calculated to output the same unit as the input signal:
Root-Mean-Square. As for all spectrum averaging modes, the averaging is done for all
spectral lines individually.

Complex Spectrum Averaging

 Some FFT analyzers support complex spectrum averaging.
 This method calculates the mean over the complex instantaneous FFT spectra, for all
spectral lines.
  Complex spectrum averaging can be used to reduce noise levels and other inconsistent
components in spectra. Its performance is similar to synchronous time-domain averaging.
 When performing averaging of instantaneous complex spectra, the amplitude components
with varying phase characteristics across spectra will be averaged out.
 Random noise can be reduced, but the relevant signal can also be reduced if the FFT
analyzer is not configured properly. This should be avoided.
 If the relevant signal components have varying phase characteristics like noise components
do, then the resulting spectrum will also get those relevant components averaged out.
 In order to avoid reducing the relevant signal components, an FFT block trigger can be
used. In this way, the phase characteristics of the relevant signal components can be
maintained such that only the noise is reduced.
 Another way to make complex FFT spectrum averaging performs as desired in multi-
channel structural testing is to define a reference channel and measure its instantaneous
phase reference spectrum. Adding the same phase shifts to all channel spectra as required
for the phase reference spectrum to keep a constant phase over time (for all spectral
components), will make the relevant frequency components consistent for all channels. In
this way, complex averaging will only average out the noise components.
 Using complex spectrum averaging, inconsistent random noise components in the averaged
spectrum can be reduced by 1/N, where N is the number of averaged complex instantaneous
FFT spectra. This is -10 dB each time the number of averages is 10 times greater.

Linear or Exponential Spectral Averaging

 When averaging is performed with equal weights for all instantaneous spectra, it is
referred to as Linear Averaging.
 For example, overall linear RMS averaging is performed when the energy content is
equally important over the full measurement time.
 Linear averaging can often be set to cover a certain time duration, or a number of FFT
spectra, where all data are equally weighted. For example, in a user scenario, it might be
desired to average the vibration energy or to determine the equivalent spectral sound
pressure levels over a certain time period.
 If newer instantaneous spectra should be weighted greater than older spectra that are
involved in the averaging process, then exponential averaging can be used instead of linear
averaging.
 With exponential averaging, the influence of spectra decays exponentially over time.
Normally a parameter relating to the exponential time decay constant tau can be adjusted
by the operator of the FFT analyzer. In this way, the weighting of the different involved
spectra is set.
 Exponential averaging is especially used for monitoring non-stationary signals with
varying amplitude levels.
  Since the newest spectrum is weighted greater than the previous spectra, sudden events
will be indicated stronger in the averaged spectrum when using exponential averaging than
when using linear averaging.

Figure 3.26: Linear averaging with 4 samples

Overlap Averaging
 Within the context of FFT analysis, the parameter Overlap refers to overlapping FFT time
blocks.
 Overlap can be used to calculate FFT spectra more consistently when window functions
are applied, and to increase the rate of produced spectra.

Figure 3.27: Overlap- 0 % overlap and 50 % overlap of FFT time blocks, having a window function applied
  Because FFT analysers produce a spectrum for every FFT time block, when these blocks
are overlapped, the analysis will produce spectra with an increased rate compared to when
using no overlap (0 % overlap). This increases the update rate of spectral displays, but
conversely, spectra will include overlapping signal content.
 Window weighted FFT blocks typically have very small (or zero) values near the block
boundaries, as shown in the Figure 3.27 above.
 The reduced values near the boundaries affect a significant portion of the time signal to be
effectively ignored in the analysis process. In measurement situations where data is
gathered at great expense, this situation should be avoided and hence overlapping FFT
blocks can be used to improve this situation.
 When using rectangular windows, all block values will already be equally weighted, and
overlapping will only help to increase the rate of produced spectra.
 Overlapping FFT blocks can be adjusted to obtain equal weighting for all time samples
over multiple overlapping spectra, giving a frequency representation of a flat (equally
weighted) time signal. This is used to obtain results equivalent to a real-time analysis,
where the overall weighting function must be uniform, for example when using Hanning
weighting. The overlap has to be at least ⅔ to obtain this.
 As the overlap is increased, FFT spectra will also become more and more correlated to
subsequent spectra. Correlated spectra are in many cases unnecessary, and therefore not
much is gained after an overlap fraction is reached which provides near equal overall
weights for the time samples. Therefore, the ideal overlap fraction is often determined to
balance equal total weights of samples and small correlation.
 Even though the ideal overlap depends on the window function and the measured signal
type, a reasonable overlap fraction to use is typically ⅔ or ¾.

Figure 4.28: Illustration of overlap processing

Maximum Hold
 Even though maximum hold does not perform averaging, it is sometimes listed under
available averaging modes for FFT analysers.
  This might be due to the fact that multiple instantaneous spectra are involved in the process,
as when performing averaging.
 Maximum hold keeps the maximum value of individual spectral lines over the specified
averaging time.
 As a result, the resulting maximum hold spectrum might have some spectral lines holding
values from some instantaneous spectra, and other lines holding values from other
instantaneous spectra.

Figure 3.29: Peak Hold Averaging

 Peak hold averaging is normally not used in routine data collection. Instead, it is used for
special tests such as Run-up, Coast Down, and Bump Tests.
 Maximum spectral hold can be used for the inspection of worst-case scenarios, by
obtaining a spectrum indicating maximum amplitudes for all frequencies over a determined
test time.

Time Synchronous Averaging

 Time Synchronous Averaging (TSA) is a mean of greatly increasing the information that
can be extracted from the time-domain vibration waveform.
 TSA is a fundamentally different process than the usual spectrum averaging that is
generally done in FFT analysis.
 It is used to greatly reduce the effects of unwanted noise in the measurement.
 The waveform itself is averaged in a time buffer before the FFT is calculated, and the
sampling of the signal is initiated by a trigger pulse input to the analyzer.
 If the trigger pulse is synchronized with the repetition rate of the signal in question, the
averaging process will gradually eliminate the random noise because it is not synchronized
with the trigger.
 However, the signal that is synchronous with the trigger will be emphasized, as shown in
Figure :

Figure 3.30: Time synchronous averaging

 When the time domain averaging is performed on the vibration signal from a real machine,
the averaged time record gradually accumulates those portions of the signal that are
synchronized with the trigger, and other parts of the signal, such as noise and any other
components such as other rotating parts of the machine, etc., are effectively averaged out.
 This is the only type of averaging that actually does reduce noise.
 Another important application of time synchronous averaging is in the waveform analysis
of machine vibration, especially in the case of gear drives.
 In this case, the trigger is derived from a tachometer that provides one pulse per revolution
of a gear in a machine.
 This way, the time samples are synchronized in that they all begin at the same exact point
in the angular position of the gear.
 After performing an enough number of averages, spectrum peaks that are harmonics of
RPM will remain when non-synchronous peaks will be averaged out from the spectrum.
 Consider a gearbox containing a pinion with 13 teeth and a driven gear with 31 teeth as
shown in Figure3.31.
 If a tachometer is connected to the pinion shaft, and its output is used to trigger an analyzer
capable of time synchronous averaging, the averaged waveform will gradually exclude
vibration components from everything except the events related to the pinion revolution.
 Any vibration caused by the driven gear will be averaged out, and the resulting waveform
will show the vibration caused by each individual tooth on the pinion.
 Figure 3.32: Application of TSA in Gearbox monitoring

 Note that in the Figure3.32, the lower averaged waveform indicates one damaged tooth on
the pinion.

Figure 3.33: Simplified exponential/linear average block diagram

 The simplified time synchronous exponential/linear average block diagram is shown in

3.34.

Figure 3.34: The simplified time synchronous average block diagram

Number of Averages
 Selection of number of averages for the analysis depends on a number of factors.
 (i) If the frequency of the rotation of the machine (speed) changes; averaging from an FFT
would not work well. In this case use of Order Spectrum will suit our requirement.
(ii) If loads are changing (and speed is constant or you are using order spectrum); a smaller
number of averages are required in order to avoid averaging out of the effect of fast load
changes.
(iii) Do few experiments with different number of averages and note when the noise floor
appears stable.
(iv) If you want faster results; smaller number of averages will produce an "average complete"
result more often.

Plots
 In order to analyse the data in a simple way it is represented in the form of plots.
 Vibration data can be represented with the help of number of plot formats.

Trend Plot
 A trend plot is simply a number of amplitude values, snapshots of the total vibration
(vibration at all frequencies) – over a period.
 The interval between readings will be the time elapsed between those readings.
 That time interval could be anything from months to milliseconds depending on the
specifics of the vibration program and system(s) involved.
 Trend graphs provide a quick visual view to the changes that are occurring.
 A trend plot offers limited analysis tools (there is no identification of specific frequencies,
for instance) but can be an important indicator of developing problems.

Figure 3.35: Trend Plot

 There are many different trend plots available in most software packages.
Time Domain Plot

Figure 3.36: Time domain plots

 Figure is a "time domain" plot.

 Typically, the length of a time domain plot will be very short - commonly in milliseconds.
 A 'Time Domain' plot displays amplitude vs. time.
 However, unlike a trend plot, the time plot is a continuous representation of the amplitude
value.
 For instance, if the amplitude unit for the above plot were displacement, the line would
represent the actual bearing location as it moves back and forth.
 Also unlike a trend plot, the values can be negative or positive since, for instance, the
displacement can be on either side of a neutral, or 'at-rest' position, and velocity or
acceleration amplitudes can be in one direction or the other (defined as the '+' and '-'
directions depending on the direction the transducer is pointing).
 The time domain is more difficult to analyze than the next plot we will discuss - the
"Spectrum" - but under certain conditions it can provide insights and information not
available on the spectrum plot.
Frequency Domain Plot (Spectrum)

Figure 3.37: Frequency domain plot

 A "Spectrum" is plot of amplitude vs. frequency.

 The above plot is a spectrum that was created from a time domain plot using FFT. This
plot is often simply referred to as an "FFT".
 By plotting amplitude versus frequency (instead of time), it becomes far easier to analyze.
 By relying on complex mathematical processes, however, it also becomes susceptible to
generating what can be misleading information. The plot displays a certain number of
amplitude values (400, 800, 1600, etc.) over a range of frequencies. The plot seen here tells
the analyst that there is:
 'A' amplitude at a frequency of approximately 3534 cpm (58.9 Hz)
 'B' amplitude at approximately 7084 cpm (118.07 Hz)
 'C' amplitude at approximately 10,633 cpm (177.22 Hz) and so on.
 This plot is the most commonly used analysis tool since, by enabling frequency
identification, it allows for preliminary identification of the source of the vibration.

Waterfall
 In addition to two-dimensional plots, common display formats include orbit plots, waterfall
plots and spectrographs.
 An orbit plot shows one time trace on the x axis and a second time trace on the y axis.
 A waterfall is a three dimensional plot made by stacking up consecutive two-dimensional
plots. Waterfall plots show how a signal changes over time, or how a signal measured from
 a rotating machine changes with variations in the RPM. They are also useful for Order
Analysis.
 Figure 3.38 shows a typical waterfall plot of the spectrum of the vibration measured on a
rotating machine during a run-up and coast down. Often the waterfall plot includes an
option to display one slice, and record of the waterfall in separate panes.

Figure 3.35: Time waterfall plot of PSD measured from a rotating machine during run-up with spectrum slice on top.

Spectrograph
 Waterfalls can also be presented as a spectrogram as shown in Figure 63.39, a two
dimensional format using colour to represent amplitude.

Figure 6.39: Spectrograph of rotating machine run-up shown in Figure 3.38

https://www.youtube.com/watch?v=PHOwO2ZoDXs

https://www.youtube.com/watch?v=LlyH6YciDhw

Signal Conditioning
 The signal from the transducer on a machine may require additional processing, like signal
amplification, noise reduction, filtering, linearization, and so on.
 These functions are usually done through standalone analog signal conditioners, and
sometimes some of these functions are done in the digital domain through dedicated digital
signal processing software after the analog-to-digital conversion.
 Some of the transducers require an external power supply, which could be provided by the
signal conditioners.
 A common requirement is to supply 4-mA current to many of the integrated charge–type noise
and vibration transducers

Signal Filtering
 During signal processing, a requirement arises to analyse the acquired or measured signals in
a particular frequency band of interest. This is achieved by filtering the signals.

Figure 3.40: Signal Filtering (Low pass filter to remove high frequency 'noise' on signal)

 Signal filtering can be done both in the analog domain and the digital domain.
 Following are the common analog filters used in signal processing:
(i) high-pass filter
(ii) low-pass filter: Pass Frequencies below a limit
(iii) band-pass filter
(iv) notch filter
High Pass Filter
 A high-pass filter allows signals with frequencies beyond a cut-on frequency to be passed
through.
 Usually in machinery condition monitoring, high-pass filters are used to remove near-mean or
DC values of the signal, and cut-on frequencies of 0.1 Hz or 1 Hz are quite common.

Low Pass Filter

 Another very important filter is the low-pass filter, and its significance in preventing
signal aliasing during data acquisition was mentioned earlier.
 Low-pass filters allow only signals to pass up to a cut-off frequency.

Band Pass Filter

 A combination of a high-pass and low-pass filter can be used as a band-pass filter, which allows
only signals in a particular frequency band to be passed through.

Notch Filter
 Many times, due to a ground loop with the electrical supply frequency, the electrical supply
frequency (50 Hz or 60 Hz) shows up in the acquired machinery signals.
 This single frequency can be removed by using a notch filter.
 The electrical supply frequency in some European and Asian countries is 50 Hz, whereas in
the Americas it is 60 Hz. At the cut off and cut-on frequencies, the filters are not sharp and
some roll off occurs, which depends on the order of the filter.

Figure 3.41: Types of filters (Source: DEWESOFT)