App Note 38 www.vibetech.

com 30-Mar-21

MEscope Application Note 38

Digital Signal Processing

The steps in this Application Note can be carried out using any MEscope package that includes the VES-3600 Ad-
vanced Signal Processing option. Without this option, you can still carry out the steps in this App Note using the
AppNote38 project file. These steps might also require a more recent release date of MEscope.

APP NOTE 38 PROJECT FILE

• To retrieve the Project for this App Note, click here to download AppNote38.zip
This Project file contains numbered Hotkeys & Scripts for carrying out the steps of this App Note.
• Hold down the Ctrl key and click on a Hotkey to open its Script window
THE FFT AND THE DFT
The FFT is a computer algorithm that calculates the Digital Fourier Transform (DFT) of a uniformly sampled time
waveform. Three equations govern the FFT algorithm.
1. SAMPLED TIME WAVEFORM EQUATION
The FFT assumes that the time waveform contains N uniformly spaced samples
The spacing (or resolution) between time samples is denoted as Δt (in seconds)
The sampling time period (also called the sampling window), spans the time period (t ➔ 0 to T) (in seconds)
The time waveform parameters are related by the equation,
T = N (Δt) (in seconds)
2. DIGITAL FOURIER TRANSFORM (DFT) EQUATION
The DFT contains (N/2) uniformly spaced samples of complex (magnitude & phase) data
The spacing (or resolution) between frequency samples is denoted as Δf (in Hz)
The DFT is calculated over a frequency span (f ➔ 0 to Fmax) (in Hz)
The DFT parameters are related by the equation,
Fmax = (N/2) Δf (in Hz)
3. SHANNON’S (NYQUIST) SAMPLING CRITERION
Shannon's Sampling Criterion says that to calculate an accurate DFT over the span (f ➔ 0 to Fmax),
The time waveform must be sampled at no less than twice the frequency Fmax
The minimum sampling rate is called the Nyquist sampling rate
The sampling criterion relates Fmax and the Nyquist sampling rate by the equation,
Nyquist sampling rate ➔ 1/Δt = 2 Fmax (in Hz)
This formula states that to obtain a valid DFT, a digital time waveform must be sampled at twice the expected value of
Fmax.

Page 1 of 13
App Note 38 www.vibetech.com 30-Mar-21

FUNDAMENTAL SAMPLING RULE

Another important equation is derived from the three equations above.
Δf = 1/T (in Hz)
This equation says that the frequency resolution (Δf) of the DFT is the inverse of the time length (T) of the time do-
main sampling window.

SAMPLING RATE VERSUS FREQUENCY RESOLUTION

To increase the frequency resolution (reduce Δf) of a DFT, the time domain signal must be sampled over a longer
time period (T).
Increasing the sampling rate (1/Δt) of the time waveform does not increase the frequency resolution (reduce Δf) of
its DFT.

ANTI-ALIASING FILTERS

When a continuous analog time domain signal is sampled, frequencies higher than Fmax in the signal will fold back
and appear as lower frequencies in the DFT.
These aliased high frequency components are not part of the DFT at frequencies below Fmax in the original signal.
To ensure that no frequencies higher than Fmax are contained in a DFT, higher frequencies must be removed from the
analog time waveform before it is sampled.
Frequencies higher than Fmax are removed using an analog low pass filter called an anti-aliasing filter.

Passing a time waveform through an anti-aliasing filter before sampling it ensures that all frequency components higher
than Fmax are removed from the frequency span ➔ 0 to Fmax of the DFT.

All anti-aliasing filters have a finite roll off frequency band.

If the cutoff frequency (start of the filter roll off) is set to 80% of Fmax, or 40% of the sampling frequency, then 80%
of a frequency span ➔ 0 to Fmax will be alias-free.
Most FFT analyzers have anti-aliasing filters with a cutoff frequency set to 80% of Fmax, or 40% of the sampling
frequency.

FOURIER SPECTRUM (DFT)

Several types of frequency spectra can be calculated in MEscope.
The Fourier spectrum is the DFT of a time waveform
The FFT algorithm is used to calculate the DFT of a time waveform
The DFT is complex valued, with Real & Imaginary parts, or Magnitude & Phase
AUTO SPECTRUM
Each Auto spectrum estimate is calculated by multiplying a DFT by its own complex conjugate
An average Auto spectrum is calculated by averaging together multiple Auto spectra
An Auto spectrum is real-valued, with Magnitude only

Page 2 of 13
App Note 38 www.vibetech.com 30-Mar-21

CROSS SPECTRUM
Each Cross spectrum estimate is calculated by multiplying the DFT of one signal by the complex conjugate of
the DFT of a different signal
An average Cross spectrum is calculated by averaging together multiple Cross spectra
The Cross spectrum is complex valued, with Real & Imaginary parts, or Magnitude & Phase
POWER SPECTRAL DENSITY (PSD)
A PSD is an Auto spectrum divided by the frequency resolution of the Auto spectrum
If the units of an Auto spectrum are (g^2), the units of its corresponding PSD are (g^2 / Hz)
ENERGY SPECTRAL DENSITY (ESD)
An ESD is a PSD multiplied by the time length (T) of the time waveform used to create the spectrum
If the units of a PSD are (g^2 / Hz), the units of its corresponding ESD are (g^2 - sec / Hz)
ESDs are used mostly to characterize transient signals
TIME DOMAIN WINDOWS

The FFT algorithm assumes that the time waveform to be transformed is periodic in its sampling window.

A signal is periodic in its sampling window if it satisfies one of the following criteria,
1. An integer number of cycles of the signal are contained within its sampling window
2. The signal has no discontinuity between its beginning & end in its sampling window
3. The signal is completely contained within its sampling window
NON-PERIODIC SIGNAL
Many signals are non- periodic in their sampling window.

A purely random signal is is non-periodic (never completely contained) within a finite length sampling window.

WHAT IS LEAKAGE?
If a time waveform is non-periodic in its sampling window, a smearing of its spectrum (called leakage) will occur
when it is transformed to the frequency domain as a DFT.
Leakage distorts the spectrum, especially around resonance peaks.
Leakage spreads the spectrum surrounding resonance peaks, which is detrimental for modal parameter estima-
tion (curve fitting).
Leakage is reduced by multiplying the sampled time waveform by a special weighting function (called a time
domain window), before the FFT is applied to the time waveform.

Page 3 of 13
App Note 38 www.vibetech.com 30-Mar-21

HANNING WINDOW FOR BROADBAND SIGNALS

If a time waveform is non-periodic in its sampling window, leakage cannot be eliminated, but it can be reduced.

A Hanning window reduces the leakage in the spectrum of a broad band signal such as a random signal.

Hanning Window.
FLAT TOP WINDOW FOR NARROW BAND SIGNALS
A Flat Top window makes the magnitudes of peaks more accurate in the DFT of a narrow band signal such as a sinus-
oidal signal.

A Flat Top window also reduces leakage in the spectrum of a narrow band signal.
A Flat Top window also makes the peaks wider in the spectrum of a narrow band signal.

Flat Top Window.

Page 4 of 13
App Note 38 www.vibetech.com 30-Mar-21

EXPONENTIAL WINDOW FOR TRANSIENT SIGNALS

A decreasing Exponential window should be applied to transient (or impulse response) signals that do not decay com-
pletely within their sampling window
A decreasing Exponential window artificially damps the signal toward zero before the end of its sampling win-
dow, thus making it nearly periodic in its sampling window.
An Exponential window adds a fixed amount of damping to all the decay waveforms in an impulse response.

Following curve fitting in MEscope, the artificial damping added by an Exponential window is subtracted from the
damping estimates of all modes.

Exponential Window.
RECTANGULAR WINDOW FOR PERIODIC SIGNALS
A Rectangular window is used on a signal that is periodic (or nearly periodic), in its sampling window
All values of a rectangular window ➔ "1"
This window is also called a Box Car window or No window

Rectangular Window.

Page 5 of 13
App Note 38 www.vibetech.com 30-Mar-21

SPECTRUM AVERAGING
Spectrum averaging is used for two important reasons,
1. To remove extraneous random noise from the DFT of a signal
2. To remove randomly excited non-linearities, which appear as random noise in the DFT
The time waveform Block Size (number of samples) is twice the Block Size (number of samples) in its corresponding
DFT
Time Waveform Block Size = 2 DFT Block Size.
The following steps are carried out during spectrum averaging by the Transform | Spectra command in the Data Block
window.
1. Each time waveform is divided into several smaller sampling windows
2. Each sampling window is windowed (multiplied by a time domain window) to reduce leakage in its spectral es-
timate
3. Each windowed time waveform is transformed into its Digital Fourier Transform (DFT) using the FFT
4. An Auto spectrum estimate is calculated from each DFT
5. Multiple Auto spectrum estimates are averaged together to yield a single Auto spectrum for each M# in the
original Data Block

Spectrum Averaging Calculation Loop

Page 6 of 13
App Note 38 www.vibetech.com 30-Mar-21

NUMBER OF AVERAGES
Depending on the Block Size of the time waveforms in a time domain Data Block, two cases can occur,
DFT Block Size ➔ 1/2 (Time Waveform Block Size)
In this case, only one spectrum estimate can be calculated using all the time waveform samples.
DFT Block Size ➔ less than 1/2 (Time Waveform Block Size)
In this case, a large time domain waveform can be divided into many smaller sampling windows, and spectrum averag-
ing can be performed.
OVERLAP PROCESSING

Overlap processing divides each time waveform into a series of smaller overlapping sampling windows.

The percentage of overlap of the sampling windows depends on three parameters,

1. The time waveform Block Size (the total number of time waveform samples)
2. The spectrum Block Size
3. The Number of Spectrum Averages
Increasing the Number of Spectrum Averages increases the percentage of overlap processing
50 % Overlap means that half of the time waveform samples are used over again in each successive sampling
window
0% Overlap means that unique time waveform samples are used for each new sampling window
LINEAR (OR STABLE) AVERAGING

Linear averaging is the same as summing together all the spectral estimates and dividing by the number of averages.

A stable averaging formula is used for linear spectrum averaging. Each stable average is calculated using a weighted
sum of the current spectrum estimate and the preceding stable average.
The Nth stable average is calculated with the following formula,
Stable Average (N) = (1/N) Spectrum (N) + (1 - (1/N)) Stable Average(N-1)
PEAK HOLD AVERAGING

Peak Hold averaging retains the maximum value at each sample from all spectral estimates.

The Ith sample of the Nth average is determined with the formula,
Peak Average (N,I) =Maximum (Spectrum (I) , Peak Average(N-1,I))
STEP 1 - FOURIER SPECTRUM OF PERIODIC SINE WAVES
• Press Hotkey 1 Fourier Spectrum of Periodic Sine Waves
To illustrate the calculation of a Fourier spectrum, a Data Block file with a time waveform containing three periodic sine
waves was created using the File | New | Data Block command. The waveform was saved in BLK: 20 30 50 Hz Sine
Waves.
When Hotkey 1 is pressed, two Data Block windows will open. The Data Block BLK: 20 30 50 Hz Sine Waves is dis-
played on the left and contains a time waveform with 20,000 samples of sinusoidal data in it.
BLK: 20 30 50 Hz Sine Waves contains data for a T ➔ 50 seconds, but only 0 to 1 seconds of data is displayed
The Data Block BLK: Fourier Spectrum (DFT) is displayed on the right and contains the Fourier spectrum (DFT) of
the time waveform in BLK: 20 30 50 Hz Sine Waves. The DFT has three peaks at 20, 30 & 50 Hz with magnitude = 1
and phase = 0.
BLK: Fourier Spectrum (DFT) has a frequency span ➔ 0 to 200 Hz but only 0 to 60 Hz is displayed

Page 7 of 13
App Note 38 www.vibetech.com 30-Mar-21

The Peak cursor in the Fourier spectrum shows a magnitude = 1g for the 30 Hz peak

The frequencies of the three sine waves (20, 30, 50) Hz divide evenly into 200 Hz, hence they are periodic in the 20,000
sample time domain window and there is no leakage in their spectrum

Fourier Spectrum (DFT) of a Periodic Signal Containing Three Sine Waves.

ONE-SIDED VERSUS TWO-SIDED FFT

The Fourier Transform is defined as an integral over all frequencies, positive & negative.
The DFT is also defined over all frequencies, positive & negative.
The spectrum for the negative frequencies has the same information in it as the spectrum for the positive frequencies.

Therefore, only the DFT for positive frequencies is displayed in MEscope.

A One-Sided FFT assigns all the energy from the time waveform to the positive frequencies of its DFT (the
part that is displayed)
A Two-Sided FFT assigns half of the energy to the positive frequencies and half of the energy to the negative
frequencies of its DFT
DFT values from a One-Sided FFT are twice as large as the DFT values from the Two-Sided FFT
STEP 2 - SPECTRUM AVERAGING USING A FLAT TOP WINDOW
• Press Hotkey 2 Auto Spectrum with Flat Top
When Hotkey 2 is pressed, two Data Block windows will open. The Data Block BLK: 20 30 50 Hz Sine Waves is dis-
played on the left and contains a time waveform with 20,000 samples of sinusoidal data in it.
BLK: 20 30 50 Hz Sine Waves contains data for a T ➔ 50 seconds, but only 0 to 1 seconds of data is displayed
The Data Block BLK: Auto Spectrum is displayed on the right contain the Auto spectrum which has three peaks at 20,
30 & 50 Hz with magnitude = 1 & phase = 0.
BLK: Auto Spectrum has a frequency span ➔ 0 to 200 Hz but only 0 to 60 Hz is displayed

Page 8 of 13
App Note 38 www.vibetech.com 30-Mar-21

Auto Spectrum of a Periodic Signal Containing Three Sine Waves With a Flat Top Window Applied.
The Peak cursor value in the Auto spectrum shows the magnitude of 1 g*g for the 30 Hz sine wave.
OVERLAP PROCESSING
Now the calculations done when Hotkey 2 was pressed will be done manually
• Right click in Data Block BLK: 20 30 50 Hz Sine Waves and execute Transform | Spectra
The following dialog box will open.

Transform | Spectra Dialog Box.

• Verify that Spectrum Block Size ➔ 1000, Number of Averages ➔ 11, and Percent Overlap ➔ 10%
This means that to calculate 11 Auto spectra and average them together, 10 percent of the time waveform samples will
be used over again in each successive sampling window.
TIME DOMAIN WINDOW

When spectrum averaging is used, data that is periodic for all samples in a time waveform window might not be period-
ic in each sampling window.

Therefore, to preserve the magnitudes of the three sine waves in the Auto Spectrum, a Flat Top window will be applied
to each sampling window before the FFT is used to calculate its DFT.
Time Domain Window ➔ Flat Top is also listed in the dialog box above

Page 9 of 13
App Note 38 www.vibetech.com 30-Mar-21

Auto Spectrum With a Flat Top Window Applied and 11 Spectrum Averages & Overlap Processing.
• Press Calculate in the Transform | Spectra dialog box
• Save the Auto Spectrum in BLK: Auto Spectrum with Flat Top
Because the Flat Top window was used, the three sine wave peaks appear at their respective frequencies (20, 30, 50) Hz,
with magnitudes ➔ "1 g*g" in the Auto spectrum. However, compared to the Fourier spectrum, the sine wave peaks
now have "width" to them.

A Flat Top window preserves peak magnitudes but increases peak widths.

TWO-SIDED FFT
• Double click on the FFT column in the M#s spreadsheet in Data Block BLK: Auto Spectrum with Flat Top
• In the dialog box that opens, choose No, click on OK
• Choose Yes in the next dialog box to re-scale the data

Page 10 of 13
App Note 38 www.vibetech.com 30-Mar-21

Auto Spectrum from Two-Sided FFT.

The magnitude of each spectrum peak in the Auto spectrum is now 0.25 g*g because half of the sine wave energy has
been assigned to the negative frequency peaks.

STEP 3 - PSD USING A FLAT TOP WINDOW

• Press Hotkey 3 PSD with Flat Top
When Hotkey 3 is pressed, the same sinusoidal signal that was used to calculate the Auto spectrum will now be used to
calculate a PSD and add it to the BLK: Auto Spectrum with Flat Top Data Block.
Two Data Block windows will open. The Data Block on the left shows the same time waveform with 20,000 samples of
sinusoidal time waveform data in it.
The Data Block on the right shows both the Auto spectrum & PSD of the waveform on the left, with three peaks at (20,
30, 50) Hz.

Auto spectrum and PSD of the Sinusoidal Waveform.

Page 11 of 13
App Note 38 www.vibetech.com 30-Mar-21

The Peak cursor in the Auto spectrum shows a magnitude of 1 g*g for the 30 Hz sine wave
The units of the PSD are (g*g/Hz) which is a power (mean squared) quantity.
In the Window Correction column of BLK: Auto Spectrum with Flat Top, choose Narrow for M# 2
Answer Yes in the dialog that opens to rescale the PSD
The Peak cursor in the PDS now shows a magnitude of 5 g*g/Hz for the 30 Hz sine wave
The magnitudes of the three sine waves are 1 g of the original time waveform. Therefore, the three Auto spectrum peaks
have magnitude ➔ 1 g*g.
A PSD is an Auto spectrum "normalized by" (divided by) the frequency resolution (Δf) of the spectrum.
• Execute File | Properties in the BLK: Auto Spectrum with Flat Top window
Frequency Resolution ➔ 0.2 Hz
Therefore, the PSD peaks should be 5 (g*g/Hz), which is confirmed by the cursor value on the PSD (M#2) in the Data
Block window shown above.
STEP 4 - SPECTRUM AVERAGING USING A RECTANGULAR WINDOW
• Press Hotkey 4 Auto Spectrum with Rectangular
When Hotkey 4 is pressed, the same sinusoidal signal that was used to calculate the Auto spectrum will now be used to
calculate a PSD and add it to the BLK: Auto Spectrum with Flat Top Data Block.
Two Data Block windows will open. The Data Block on the left shows the same time waveform with 20,000 samples of
sinusoidal time waveform data in it.
The Data Block on the right again shows both the Auto spectrum & PSD of the time waveform on the left, with three
peaks at (20, 30, 50) Hz.

Auto spectrum & PSD of a Periodic Signal Using a Rectangular Window.

The narrow peaks at the three frequencies at (20, 30, 50) Hz in both the Auto spectrum & PSD verify that the time
waveform remained periodic in each sampling window when 11 averages were calculated with 10% overlap pro-
cessing.

Page 12 of 13
App Note 38 www.vibetech.com 30-Mar-21

CONCLUSION
In Steps 2 & 3, a Flat Top window was applied to the time waveforms in each sampling window to obtain accurate
magnitudes in each spectral estimate. An Auto spectrum and PSD were calculated from the time waveform in BLK: 20
30 50 Hz Sine Waves using the following parameters
Spectrum Block Size ➔ 1000 samples
Number of Averages ➔ 11
Overlap processing ➔ 10%
Time Domain Window ➔ Flat Top
The time waveform in BLK: 20 30 50 Hz Sine Waves contains 20000 samples over a time of T ➔ 50 seconds.

With a Spectrum Block Size ➔ 1000 samples, each time domain sampling window contains 2000 samples over a time
of T ➔ 5 seconds.
With these sampling parameters, each sinusoidal waveform is still periodic in its sampling window.
• The three sine waves complete exactly 100 (20 Hz), 150 (30 Hz), 250 (50 Hz) cycles in T ➔ 5 seconds
• With overlap processing, the next sampling window starts after 10 (20 Hz), 15 (30 Hz), 25 (50 Hz) cycles of
each sine wave
• No leakage will occur in the calculated DFT, and therefore the Auto spectrum & PSD have no leakage

Because each sinusoidal waveform is periodic in its sampling window a Rectangular window can also be used instead
of a Flat Top window.

STEP 5 - REVIEW STEPS

To review the steps of this App Note,
• Press Hotkey 5 Review Steps

Page 13 of 13