You are on page 1of 58

Test Technology: DSP (Digital Signal Processing)

Fourier transform Aliasing & leakage Measurement functions


Hong Weng
Customer Service Engineer
LMS - A Siemens Business
Hong.weng@lmsintl.com
2 copyright LMS International - 2013
Lecture objectives
Understand the importance
of the Discrete Fourier
Transform (DFT)
Be able to explain aliasing
and leakage
See the advantages of
frequency-domain
measurement functions
By completing this lecture, you will:
0.00 800.00 Hz
0.00
7.70e- 3
A
m
p
lit
u
d
e
g
3 copyright LMS International - 2013
DSP in Test.Lab
Acquisition Time?
Frequency Resolution?
4 copyright LMS International - 2013
Signals and processing
Signal: measurable quantity carrying information on some physical phenomenon
Pressure, displacement, acceleration,
Temperature, voltage, biomedical potential (EKG, EEG, ...)

Information contained in the variation of the quantity over time (space, )
This signal is measured with a sensor
This signal is what you want to analyse in view of a particular problem

Analog Signal
5 copyright LMS International - 2013
Signals and processing
Signal Processing: specific manipulations of the measured signals to:
Extract the key information
Understand the physical problem
Provide input data for specific analysis or even simulations
Modify the signal for specific applications
Digital Signal Processing: doing all this using computer-based systems
Transform the sensor signal in a stream of digital words
Most sensors have an analog signal output
Computers are limited to analysing finite datasets
Discretisation in time and in amplitude
6 copyright LMS International - 2013
System
Transfer
System
Transfer
Receiver Receiver
Road
Wheel & Tire
Steering Wheel
Shake
Seat Vibration
Rearview mirror
vibration
Engine
Signals everywhere
X =
Gearbox and
Transmission
Turbomachinery
Accessories
Rotor
Cockpit vibration &
noise
Cabin comfort
Noise at Drivers &
Passengers Ears
Structural Integrity
Environmental
sources
Source Source
7 copyright LMS International - 2013
and they can look hmm interesting
Ariane 5 launch and
8 copyright LMS International - 2013
Joseph did help us a lot
Joseph Fourier (1768 - 1830)
Thorie analytique de la chaleur
(1822)
Fouriers law of heat conduction
Analyzed in terms of infinite
mathematical series

2
2
2
2
y
u
x
u
k
t
u
-2.5
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
-4
-3
-2
-1
0
1
2
3
4
Any signal can be described as a
combination of sine waves of different
frequencies
Useful by-product
9 copyright LMS International - 2013
Fourier transform
To go from time to frequency domain and back
Fourier integral:
Supported by modern signal analysers
Spectrum analysers
Basic function in all our software
( ) [ ] ( ) X t x F =
( ) [ ] ( ) t x X F =

1
( ) ( )
( ) ( )

=
=

+


+

d e X t x
dt e t x X
t j
t j
2
1
For mathematicians
For humans
-2.5
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
-4
-3
-2
-1
0
1
2
3
4
f [Hz]
10 20 40
Detect sine waves in signal Draw line at frequency of sine wave
10 copyright LMS International - 2013
Some definitions
t [s]
f [Hz]
[rad/s]
T
0
f
0

0
Time domain
Frequency domain
Period: T
0
[s]
Frequency: f
0
= 1/T
0
[Hz]
Pulsation / circular frequency:

0
= 2f
0
= 2/T
0
[rad/s]
1 rad

2
11 copyright LMS International - 2013
Frequency spectrum Time history
Selection of domain, depending on the application aims
Equivalence of time and frequency domain: no loss of information
Time Time Frequency Frequency
f
f
f
f
f
f t
t
t
t
t
t
12 copyright LMS International - 2013
Examples Fourier transform
13 copyright LMS International - 2013
Bridge Vibrations
t
t
f
t
Traffic
Shaker
Drop
weight
Time domain Frequency domain
14 copyright LMS International - 2013
There exist more domains
Representation of signals for analysis
t
A
f
A
2
/f
A
P
f
A
2
/f
A
P
t
A
Time domain:
The time history x(t)
Frequency domain:
The signal spectrum X( )
Amplitude domain:
The probability distribution P(A)
Gaussian
distribution
Uniform
distribution
15 copyright LMS International - 2013
Nice theory but we must do it on a computer
Sampled signals
Discrete time history
Discrete frequency spectrum
Finite signal segments
Limited number time samples
Limited number of frequency lines
Numerical representation
Discrete number of possible amplitude
values
( ) [ ] ( ) X t x F =
( ) [ ] ( ) t x X F =

1
( ) ( )
( ) ( )

=
=

+


+

d e X t x
dt e t x X
t j
t j
2
1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
-1.5
-1
-0.5
0
0.5
1
1.5
Consequences ?
16 copyright LMS International - 2013
Discretisation Effects:
Aliasing and Leakage
Two most frequently occurring problems using discretisation:
does not meet Shannons Theorem
Remedy
Use band-limited signals
Use low-pass filtering
The sampled function is not transient and not periodic
Remedy
Use periodic signals
Apply windowing (errors remain!)
( )
max
2 f f
s

s
f
ALIASING
LEAKAGE
17 copyright LMS International - 2013
Sampling
Sine wave of 10 Hz, sampled at 100 Hz
Digital representation looks like a perfect sine
Following slides:
Reducing sampling frequency
0.2 0.4 0.6 0.8 1
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
time - seconds
a
m
p
l
i
t
u
d
e
sampling frequency = 1000 Hz
10 Hz harmonic function
T=N

t
100 Hz
18 copyright LMS International - 2013
0 2 4 6 8 10
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
time - seconds
a
m
p
l
i
t
u
d
e
sampling frequency = 100 Hz
10 Hz harmonic function
4 4.2 4.4 4.6 4.8 5
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
time - seconds
a
m
p
l
i
t
u
d
e
sampling frequency = 100 Hz.
4 4.2 4.4 4.6 4.8 5
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
time - seconds
a
m
p
l
i
t
u
d
e
sampling frequency = 100 Hz.
t N T =
19 copyright LMS International - 2013
0 5 10 15 20 25
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
time - seconds
a
m
p
l
i
t
u
d
e
sampling frequency = 40 Hz.
10 Hz harmonic function
10 10.2 10.4 10.6 10.8 11
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
time - seconds
a
m
p
l
i
t
u
d
e
sampling frequency = 40 Hz.
10 10.2 10.4 10.6 10.8 11
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
time - seconds
a
m
p
l
i
t
u
d
e
sampling frequency = 40 Hz.
T N t =
20 copyright LMS International - 2013
0 10 20 30 40 50
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
time - seconds
a
m
p
l
i
t
u
d
e
sampling frequency = 20 Hz.
10 Hz harmonic function
T N t =
20 20.2 20.4 20.6 20.8 21
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
time - seconds
a
m
p
l
i
t
u
d
e
sampling frequency = 20 Hz.
20 20.2 20.4 20.6 20.8 21
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
time - seconds
a
m
p
l
i
t
u
d
e
sampling frequency = 20 Hz.
21 copyright LMS International - 2013
Sampling: exploring the limits
Sampling frequency = sine wave
frequency
f
s
= f
sine
Observed frequency = 0 Hz (DC)
Sampling frequency = 2 x sine wave
frequency
f
s
= 2 x f
sine
Observed frequency is correct, but it is
borderline (sampling frequency cannot be
lowered)
22 copyright LMS International - 2013
Sampling = only look from time to time
Different interpretations possible ???
-1.5
-1
-0.5
0
0.5
1
1.5
t
t
f
s

=
1
t
t
f
s

=
1
23 copyright LMS International - 2013
Sampling Potential source of trouble
20 Hz signal, sampled at 21.3 Hz, shows up as a 1.3 Hz signal Aliasing
f
s
2f
s
3f
s
f
s
/2 0
True
frequencies
Sampled
frequencies
f
s
/2
-1.5
-1
-0.5
0
0.5
1
1.5
-1.5
-1
-0.5
0
0.5
1
1.5
ff
ff
Correct
Observed
20
20 1.3
24 copyright LMS International - 2013
Aliasing Protection
Low-Pass Filter
Make sure the signal does not contain
frequencies above half the sample frequency fs
Do this by applying a sufficient performing low-
pass filter
Be aware that the amplitude of the last portion
of the spectrum is attenuated by the filter
Alias-free
Automatically done in good data acquisition hardware
25 copyright LMS International - 2013
Example
Alias-free
Frequency range
suffering from aliasing
26 copyright LMS International - 2013
Aliasing sometimes positive
Something strange?
Glass vibrates at 608
Hz, while we see it
vibrating at 2 Hz!
Sampling by
stroboscope at 101 Hz
(Operating range is 0
120 Hz)
6 x 101 Hz = 606 Hz
For the human eye: 101
Hz = analog (we dont
see the samples)
27 copyright LMS International - 2013
Discretisation Effects:
Aliasing and Leakage
Two most frequently occurring problems using discretisation:
does not meet Shannons Theorem
Remedy
Use band-limited signals
Use low-pass filtering
The sampled function is not transient and not periodic
Remedy
Use periodic signals
Apply windowing (errors remain!)
( )
max
2 f f
s

s
f
ALIASING
LEAKAGE
28 copyright LMS International - 2013
Finite Observation Length
Limited observation
Discrete Spectrum Periodicity Assumed
Complete original signal
We are NOT analysing
the original signal !!
29 copyright LMS International - 2013
Finite Observation Side Effect
Adverse effects
Wrong amplitudes
Smearing of the
spectrum
-2.5
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
Leakage
0.00 100.00 Hz
0.00
1.00
A
m
p
l
i
t
u
d
e
(
m
/
s
2
)
0.00 100.00 Hz
0.00
1.00
A
m
p
l
i
t
u
d
e
(
m
/
s
2
)
0.00 100.00 Hz
-60.00
0.00
d
B
(
m
/
s
2
)
0.00 100.00 Hz
-60.00
0.00
d
B
(
m
/
s
2
)
Linear scale
Log scale
Linear scale
Log scale
Expected spectrum of a
pure sine wave
30 copyright LMS International - 2013
Leakage Amplitude Uncertainty
Periodic observation
100% of amplitude
A-periodic observation
63% of amplitude
Boss, this 100.000$ system is giving me
something between 6 and 10g
31 copyright LMS International - 2013
Reducing Leakage by Applying Time Windows
Leakage originates from finite observation
(discontinuity-error at edges)
Original signal properties are best
represented in the middle of the observation
period : enhance information
Practical implementation : multiplication
by window-function (time domain) to reduce
discontinuities
Effects :
Improved amplitude estimate (flatten
central lobe)
Reduce frequency range of smearing
(lower side lobes)
Local smearing of spectral energy due
to wider central lobe effective
spectral resolution decreases
32 copyright LMS International - 2013
Window Types Specific Characteristics
T
i
m
e

d
o
m
a
i
n
F
r
e
q
.

d
o
m
a
i
n
Rectangular, uniform Hanning Flat top
33 copyright LMS International - 2013
Windowing Use Cases
Uniform (rectangular)
Only in leakage-free conditions
Hanning
Most commonly used for unknown signals
Compromise: amplitude relatively correct good frequency precision
High side lobes may mask neighbouring frequencies with low amplitude
Kaiser-Bessel
Good selectivity (low side lobes): measure close frequencies with large amplitude
differences
Flat top
Calibration: accurate amplitude measurement
Very bad effective frequency resolution
Impact testing windows
Exponential (response)
Force-window (input signal)
34 copyright LMS International - 2013
Example 1
Periodically observed sine
Rectangular window
Hanning window
Non-periodically observed sine
Rectangular window
Hanning window
0.00 100.00 Hz
-100.00
0.00
d
B
(
m
/
s
2
)
AutoPower_Per_Hann
AutoPower_Per_Rect
0.00 100.00 Hz
-100.00
0.00
d
B
(
m
/
s
2
)
AutoPower_Nonper_Hann
AutoPower_Nonper_Rect
35 copyright LMS International - 2013
Example 2
2 sines which are non-periodic within the measurement period. The amplitude of
the second sine is 100 lower than the amplitude of the dominant sine.
Alternatively: measure longer!
Rectangular
Flat top
Hanning
Kaiser-
Bessel
36 copyright LMS International - 2013
Discretisation Effects:
Aliasing and Leakage
Two most frequently occurring problems using discretisation:
does not meet Shannons Theorem
Remedy
Use band-limited signals
Use low-pass filtering
The sampled function is not transient and not periodic
Remedy
Use periodic signals
Apply windowing (errors remain!)
And perhaps a 3
rd
one:
Amplitude discretisation (e.g. 16/24 bit ADC)
( )
max
2 f f
s

s
f
ALIASING
LEAKAGE
37 copyright LMS International - 2013
Amplitude discretisation problem
Small variations are not
detected
Amplitudes are
approximated
Small signals look bad
6
7
5
4
3
2
1
0
38 copyright LMS International - 2013
Amplitude discretisation solution
Amplify signal to cover optimally
available input range
Many bits in ADC to provide many
possible values
So we can describe accurately
small variations
Currently 24 bit ADC
6
7
5
4
3
2
1
0
MAXIMUM VOLTAGE
MINIMUM VOLTAGE
39 copyright LMS International - 2013
So we need assistance for
Filtering
Several possible sample frequencies
Windowing
Amplification
Sufficient possible amplitude values
40 copyright LMS International - 2013
8

Analog sensor signal Fourier transform (infinite integral)


Sampled signal Discrete-time Fourier transform (DTFT)
Finite observation length Discrete Fourier transform (DFT)
Repetition of time blocks Sampled freq. domain (spectral lines)
Repetition of spectra Sampled time domain
Fourier & Co
41 copyright LMS International - 2013
DFT Parameters
Block size N
Sampling interval t = 1/f
s
Observation time T = N t
Sampling frequency f
s
= 1/t
Nyquist frequency (bandwidth) f
N
= f
s
/2
Spectral lines N
s
= N/2
Frequency resolution f = 1/T = f
s
/N
Time domain Frequency domain
t
f
f
N
f
s
0
f
t
T
42 copyright LMS International - 2013
DSP in Test.Lab
Spectral test specification:
Maximal signal frequency of
interest
Bandwidth (f
max
, f
N
)
Sampling (f
s
, t)
Frequency separation
requirement
Resolution (f)
Observation time
(T) and block size (N)
Aliasing prevention
Sample high enough +
filtering
Leakage prevention
Periodic signals,
transient signals, or
windowing
43 copyright LMS International - 2013
Some history
Fourier series - Joseph Fourier (1822)
Origin
Discrete Fourier Transform (DFT)
Sampling + finite time
Fast Fourier Transform (FFT) Cooley & Tukey (1965)
Efficient algorithm for DFT
Power of 2 number of samples (e.g. 512, 1024, 2048, 4096, )
Fastest Fourier Transform in the West (FFTW) Frigo & Johnson (1999)
Efficient algorithm for DFT for non-power-of-2 number of samples
44 copyright LMS International - 2013
Signal analysis measurement functions
Time domain and frequency domain calculations to
extract specific information from the test signals
Time history
Time data segment statistics
Auto/cross correlation function
Frequency spectrum, auto/cross power spectrum
Rotating machinery tracked spectrum analysis (See
Signature Testing lecture)
Coherence and Frequency Response Function (See
Structural Testing lecture)
The key issues to select a function are:
What information is needed? How is this information
best brought forward from the signal?
Averaging to enhance weak signal components
Absolute values
0.00 80.00 Hz
-140
-40
d
B
(
(
m
/
s
2
)
/
N
)
22.56 41.19
s
Time winr:61:+Z
Time winr:62:+Z
Averaging
45 copyright LMS International - 2013
23/11/2002: Bradford City Sheffield United: 0 5
Data acquisition: 4 h
Sampled at 80 Hz (down-sampled to 20 Hz)
Sliding RMS value ( )
1000 samples, 50% overlap
0.00 15000.00 s
-0.02
0.02
R
e
a
l
(
m
/
s
2
)
-0.20
0.20
R
e
a
l
(
m
/
s
2
)
F time_record roof :1:+X / Root Mean Square
B time_record roof :1:+X
Goal 1 Goal 2 Goal 3 Goal 4 Goal 5
Half time Empty Filling Seated Emptying
End of game
46 copyright LMS International - 2013
5 : Belgian blocks
1 : runups
2 : ramps
3 : asphalt
4 : ramps
Road load data analysis
47 copyright LMS International - 2013
To design representative test scenarios
Accelerated durability testing cycles
Meeting 1.2 million km durability
requirement
Real tests would take 3 years
Large-scale customer data collection
5000 km Turkish public road data
Ford Lommel proving ground
Development of accelerated rig test
Target setting
Test schedule definition
Resulting test schedule 8 weeks
Test acceleration of factor 100
LMS engineers performed dedicated data collection, applied extensive load
data processing techniques and developed a 6- to 8-week test track sequence
and 4-week accelerated rig test scenario that matched the fatigue damage
generated by 1.2 million km of road driving.
1
Damage based on strain gage signals, full truck
48 copyright LMS International - 2013
Electric motor powers machinery through
gear reduction drive units
Increased vibration level from wear
Gearbox geometry
Main shaft frequency: 59.7 Hz
Final shaft frequency
59.7*(17/55)*(20/68) = 5.43 Hz
Final gear mesh frequency
5.43*68 = 369 Hz
Fs = 1024 Hz
400.00 0.00 Hz
0.04
0.00
A
m
p
l
i
t
u
d
e
m
/
s
2
59.76 369.00
Applications: Electric Motor & Gear Mesh Analysis
49 copyright LMS International - 2013
0.00 400.00 Hz
10.0e-6
0.10
L
o
g
(
m
/
s
2
)
369 30 184 60
Main shaft frequency
Half of the main shaft frequency Harmonics of the main shaft frequency
Half of the gear mesh frequency Gear mesh frequency
Applications: Electric Motor & Gear Mesh Analysis
50 copyright LMS International - 2013
Monitor current drawn by electrical
motor
Spacing and asymmetry in the
sidebands related to defects in the
motor
Analysis
60 Hz running frequency of motor
Power line sidebands: 2.75
Hz/sideband away from 60 Hz
carrier
Motor slip sidebands: 1.25 Hz
away from 60 Hz carrier
35.00 85.00 Hz
-100.00
0.00
d
B
A
2
N = 1024, f = 1 Hz
N = 2048, f = 0.5 Hz
N = 8192, f = 0.125 Hz
Current probe power spectra
Hanning
Applications: Electric Motor & Gear Mesh Analysis
51 copyright LMS International - 2013
Power spectra N = 8192, f = 0.125 Hz
55.00 65.00 Hz
-100.00
0.00
d
B
A
2
35.00 85.00 Hz
-100.00
0.00
d
B
A
2
Zoom
Rectangular window
Hanning window
Kaiser-Bessel window
Applications: Electric Motor & Gear Mesh Analysis
52 copyright LMS International - 2013
0.00 1000.00 Hz
1.00e-6
10.0e-3
L
o
g
g
Autopower Example:
Pump Vibration Signatures
Misalignment between motor and pump
assemblies causes excessive bearing
wear
Good alignment shows up as reduced
harmonic content
Accelerometer measurement on the
motor bearing cap
Computation of vibration signatures
Power Spectra
Linear
RMS
Hanning
Amplitude correction
N = 1024
fs = 2048 Hz
Good alignment
Bad alignment
0.00 52.00 s
- 0.10
0.10
R
e
a
l
g
53 copyright LMS International - 2013
Harmonic cursor, display limited to 800 Hz, dB amplitude scale
0.00 800.00 Hz
1.00e-6
10.0e-3
L
o
g
g
29.73
Good alignment = reduced harmonic content
Bad alignment
Autopower Example:
Pump Vibration Signatures
54 copyright LMS International - 2013
0.00 800.00 Hz
0.00
7.70e-3
A
m
p
l
i
t
u
d
e
g
Good alignment = reduced harmonic content
Bad alignment
Linear amplitude scale
Autopower Example:
Pump Vibration Signatures
55 copyright LMS International - 2013
Industrial Printer Noise Problem
Story
Industrial printer
Excessive noise level
Measure effectiveness of noise
abatement shroud
0.00 33.00 s
-0.60
1.30
R
e
a
l
P
a
22.39 22387.21 Octave 1/3
Hz
20.00
70.00
d
B
P
a
20.00
70.00
d
B
P
a
A L
25.0 20000.0
Curve 25.0 20000.0 RMS Hz
28.1 46.4 69.2 dB dB
27.7 36.5 66.9 dB dB
1/3 octave band representation
56 copyright LMS International - 2013
Course summary
Good acquisition
system: aliasing
protection
Amplitude
discretisation
DFT = Discrete
Fourier
Transform
Measurement
functions
Skilled
experimentalist:
leakage
mitigation
57 copyright LMS International - 2013
Thank you
Questions ?
Do not hesitate to contact me or the LMS Test Support team
testsupport.us@lmsintl.com
Test Support Phone # : 248 502 2211
Or visit www.lmsintl.com
Please fill in the survey at the end of the WebEx
Thank you

You might also like