Operational Modal Analysis 2012 (OMA Notes)

Operational
Modal Analysis
www.bksv.com
Brüel & Kjær Sound & Vibration Measurement A/S.
Copyright © Brüel & Kjær. All Rights Reserved.
www.bksv.com, 1
Outline
 Introduction
 Identification Techniques
 Frequency Domain Decomposition, FDD
 Enhanced Frequency Domain Decomposition, EFDD
 Comparison between FDD & EFDD
 Harmonic Detection, CFDD
 Automatic Mode Extraction
 Projection channels
 Stochastic Subspace Identification, SSI
 Measurement results & Validation
 Summary
www.bksv.com, 2
Operational Modal Analysis – (OMA)
 Determination of Modal Model by response testing only
– No measurement of input forces required
– Measurement procedure similar to
Operating Deflection Shapes (ODS)
 Determination of Modal Model under operational conditions

– In-situ testing
 Used successfully in Civil Engineering applications (Ambient Modal)

– Bridges and buildings
– Off-shore platforms etc.
 “Recently” introduced to Mechanical Engineering applications (OMA)

– Rotating Machinery
– On-road and in-flight testing etc.
www.bksv.com, 3 Car
Data Acquisition (measurement procedure)
Measurement procedure similar to Operating Deflection Shapes
 Measurements can be done in
a single data set (fast, data consistency) or
by using multiple data sets (limited number of transducers)
 Single data set: No reference transducers needed

 Multiple data sets: One or more fixed accelerometers are needed as
references for all data sets
Accelerometers are moved for each data set
Fixed Reference
Accelerometers
www.bksv.com, 4
Z24 Highway Bridge, near Zürich
9 Data Sets, 250 DOFs
5 references
The measurements and analysis presented here were performed and made public in order to
check the performance of various algorithms of Modal Analysis software.
3,8 Hz
www.bksv.com, 5
Heritage Court Tower, HCT
 HCT is located in downtown Vancouver, Canada

 Rectangular in shape
 Floor dimensions of the upper floors 25 m by 31 m
 Floor dimensions of the lower levels 36 m by 30 m.
 Parking structure 14 meters beyond the tower
1,3 Hz
The measurements and analysis presented here are performed by Professor Carlos Ventura,
University of British Columbia, Vancouver, Canada.
www.bksv.com, 6
Modal Analysis in the Presence of Harmonic Excitation
 Build at Flensburger Schiffbau Gesellschaft.

 Dimensions: length over all: 199.8 m
 Speed: 22.5 kn
 Main engine: 9L 60 MC-C (MAN B&W, 9 cylinder)
 Power: 20 070 kW
 Speed: 123 rpm (2.05 Hz) 1 Data Set, 16 DOFs
 Working process: two-stroke
 4-bladed propeller with controllable pitch
1,3 Hz
The measurements and analysis presented here are performed by

University of Rostock, Faculty of Mechanical engineering. Chair of Technical Mechanics.
www.bksv.com, 7
Erfttal Bridge, High Speed Railway Bridge, Aachen
3,7 Hz
5,2 Hz
13,2 Hz 13,7 Hz
www.bksv.com, 8
Wind Turbine Wing
 5 1 scaled down model
1 Data Set, 48 DOFs
14,3 Hz
Acoustic excitation
www.bksv.com, 9
Small Building Model
 Repeated Roots 2 Data Sets, 6 DOFs

1 Data Set, 8 DOFs
20 Hz
www.bksv.com, 10
Setup: Plate – Frontend - Accelerometers
1 Data Sets, 16 DOFs 4 Data Sets, 17 DOFs
www.bksv.com, 11
The Brüel & Kjær OMA Solution
Modal Test
Vibration Consultant Operational
Data PULSE Modal Analysis
File
LabShop Software
Transfer
PULSE, MTC, OMA
www.bksv.com, 12
Mobility Measurements (Traditional Modal Testing)
[m/s²] Time(R espo nse) - Input

Wo rking : Input : Input : FF T A nalyzer
 Input Force is measured
80
40
 Output Response is measured
0
-40  Output is related to Input by FRF estimators

-80 FFT
0 40m 80m 120m
[s]
160m 200m 240m
 FRF is independent of the input force
Output Signal
 Input/Output modal analysis
[(m/ Frequency
s²)/N ] Respo nse H 1(R espo nse,Excitatio n) - Input (M agnitude)
100
Output Spectrum
1
Input Input Spectrum

10m
0 400 800 1,2k 1,6k 2k 2,4k 2,8k 3,2k

[H z]
Structural System
Only the Modes of the System need to be identified

[N ]
200
Time(Excitatio n) - Input
FFT Even if the system is producing noise this is handled
100
0
by FRF estimators as H1, H2 , H3
-100
-200
0 40m 80m 120m 160m 200m 240m
[s]
Input Signal
www.bksv.com, 13
Limitations of traditional modal analysis
 If the structure is big

– it is difficult to excite it (Hammer/Shaker)
– it is difficult to measure the applied excitation
 In the experiment, it is difficult to reproduce

– object mounting (= boundary conditions)
– response magnitude
– static loads
– effect of the air (or fluid) stream
– associated mass (for underwater structures)
www.bksv.com, 14
Operational Modal – (The Combined Model)
If the system is excited by white noise
the output spectrum contains full information of the structure
as all modes are excited equally
 Output only modal analysis
Force Spectrum Structural System

Output Spectrum
But this is in general not the case!
www.bksv.com, 15
In general the excitation has a spectral distribution
Modes are weighted by the spectral distribution of the input force

Combined Spectrum
Both the “peaks” originating from the excitation signal and

the structural modes are observed as “modes” in the Response
www.bksv.com, 16
Noise also contributes to the Response

Computational Noise

Combined Spectrum
Measurement Noise
Computational as well as measurement noise may result in

what will be observed as “modes” in the Response
www.bksv.com, 17
Rotating parts creates Harmonic vibrations
Rotating Parts

Computational Noise

Combined Spectrum
The “Modes” in the combined spectrum contains information of

 The system under test (Physical Modes)
Measurement Noise  Input Force (Non-physical “Modes”)
 Noise (Non-physical “Modes”)
 Harmonics (Non-physical “Modes”)
www.bksv.com, 18
Combined System Model (analysis procedure)
Model of the combined system is estimated from measured responses
Combined System
Stationary
Zero Mean Measured
Gaussian Responses
White Noise Excitation Filter Structural System
(linear, time-invariant) (linear, time-invariant)
Unknown excitation forces
www.bksv.com
Modal Model of Structural System
extracted from estimated model of Combined System

www.bksv.com, 19
Assumptions
Mathematical
 Stationary input force signals can be approximated by
filtered zero mean Gaussian white noise
» Random in time and space
» Signals are completely described by their
correlation functions or auto- & cross-spectra
» Synthesized correlation functions or
auto- & cross-spectra are similar to
those obtained from experimental data
Practical Excitation {F}

 Force frequency spectra are
broadband and smooth
 Forces are un- or weakly correlated
 Forces act all over the entire structure
www.bksv.com, 21
How realistic the OMA assumptions are?
 Just to remind, the OMA assumptions about the input forces {F}:
– Force spectra are broadband and smooth
– Forces are un- or weakly correlated
– Forces act all over the entire structure
 Perfect excitation from OMA point of view 
 We typically consider that the excitation

due to
– Wind and waves
– Traffic
well satisfies OMA assumptions.
 Apparently, the quality of OMA results

directly depends on how good the
real excitation satisfies the assumptions
www.bksv.com, 22
Outline
 Introduction
 Summary
www.bksv.com, 23
Identification Techniques
Non-Parametric
 Modal parameters are estimated directly from curves,
functional relationships or tables
Experimental Data Signal Processing Modal Parameters
Parametric
 Modal parameters are estimated from a
parametric model fitted to the signal processed data
Experimental data Signal Processing

Parameter Estimation Modal Parameters
Parametric Model
www.bksv.com, 24
Non-Parametric Techniques
Example: Quadrature Picking in Mobility based Modal Analysis

 Modal Frequencies are estimated from peaks in FRFs
 Damping is estimated from bandwidth in FRFs
 Modeshapes are estimated from the quadrature peaks
Amplitude First
Mode Second
Mode Third
Mode
H(w0)
2
Beam
Force
w0 Frequency
www.bksv.com, 25
Plate (Reflex or MTC)
Non-Parametric Techniques
Example: Operating Deflection Shapes
Determine the forced dynamic deflection at the operating speed
 ODS frequencies are estimated from peaks in Autospectra
 Deflection patterns are estimated f.ex. from transmissibility or
Phase Assigned Spectra measurements at ODS frequencies
Accelerometer
Xi
1.2, 0 
 
1.1, 0 
0.6, 0 
  Reference
Relative 0.5, 180  Absolute

Accelerometer
 Xi   
  X 
   0.6, 180  Xi f0   i  Xref  Xi 
 Xref f0    Xref   
 0 . 4 , 180   Xref 
0.3, Xi Xref
0 

Structure
 

 1 . 0, 0 
 
1.3, 0  
www.bksv.com, 26
Identification Techniques
Measured Data fn, , 

Signal Processing Modal Fit
Time Histories
Natural Frequency, fn
Damping, 
Mode Shape, 
EFDD (Simplest fit)

FDD
(Pure Signal Processing) CFDD (SDOF fit)
SSI (MDOF, Advanced fit)
 Non-parametric method:
» Frequency Domain Decomposition, FDD
 Parametric methods:
» Enhanced Frequency Domain Decomposition, EFDD
» Curve-fit Frequency Domain Decomposition, CFDD
» Stochastic Subspace Identification, SSI
www.bksv.com, 27
Outline
 Introduction
 Summary
www.bksv.com, 28
Frequency Domain Decomposition (FDD)
Determination of “complete” Modal Model from Responses only
FDD procedure:
 Cross Spectral Matrix estimation for each data set, [Gyy(w)]
 Singular Value Decomposition (SVD) of Matrices
 Identification of Single Degree of Freedom (SDOF) models
from SVD curves
 Modal Parameter identification from SDOF models
SVD
www.bksv.com, 29
Why it works?
Let’s consider a linear, time invariant dynamic system.
The response spectrum can be expressed as a product:
Yw  Hw Fw

post-multiplying by its Hermitian
YwH  FwH HwH

yields:
w
Y Y
wH  Hw
 F
w F
wH HwH

[ G YY ( w)] [ GFF ( w)]
performing averaging:
GYY w  Hw GFF w HwH

where [GYY] и [GFF] are output and input cross-spectra matrices.
www.bksv.com, 30
Why it works? – cont.
Let’s make the following assumptions about the input forces {F} :
– their frequency spectra are broadband and smooth
– the forces are un- or weakly correlated
– the forces act all over the entire structure
{F}
Then
 G11 G12 ... G1n 

 G G22 ... G2n 
[GFF ]  {F} {F} 
H  21
 [I]
 ... ... ... ... 
 
Gn1 Gn2 ... Gnn 
www.bksv.com, 31
Why it works? – cont.
So if GFF w  I 
then GYY w  HwI  HwH
Conclusion: under the aforementioned assumptions about the loads, the

output cross-spectra matrix [GYY] holds all information about the
dynamic properties of the system.
The goal of OMA is to extract the modal parameters from measured

dynamic response of the structure
www.bksv.com, 32
Modal Behavior
The dynamic deflection y is a
linear combination of the Mode Shapes,
the coefficients being the Modal Displacements
= + + ++
y(t) = q1(t)1 + q2(t)2 + q3(t)3+    + qn(t) n
or
m
y   q   r qr
r 1
www.bksv.com, 35
 Modal behaviour: y(t) = []q(t)

=>
 Correlation: [Cyy()] = E{y(t+)y(t)T}
[Cyy()] = E{[]q(t+)q(t)H[]H} = [][Cqq ()] []H
 i.e. by Fourier Transform:
[G yy(w)] = [][Gqq(w)] []H
 Same form as Singular Value Decomposition, SVD of

Cross Spectrum Matrix:
[G yy(w)] = [V][S] [V]H
www.bksv.com, 36
Singular Value Decomposition of Hermitian matrices
[A] = [V] [S] [V]H = s1 v1 v1H+s2 v2 v2H +..

The Singular Value Decomposition
of the response matrices is
performed for each frequency
Singular values Singular vectors

 s1 0 0 0 . . 0.
0 s2 0 . . . 

0 0 . . .
[S] = 
0 . . s .

[V] = v1 v2  v3  . . . vn 
. . 0 .
 
. 0 .
. 0 .
 
 0 . . . . . . 0
 Orthogonal columns
 A real diagonal matrix  Unity length columns
 Identify Natural frequencies  Approximates the Mode shapes
from Singular Value spectra
www.bksv.com, 37
Singular Value Decomposition of Response Cross-spectrum Matrix
[A] = [V] [S] [V]H = s1 v1 v1H+s2 v2 v2H +..
 SVD performed for each frequency, response spectra of modes

 Frequency at peak found from decoupled modes i [G]
 The singular vectors approximates the mode shapes [G]
[G]
 No damping estimated [G]
[dB] Singular Values of Pow er Spectral Density
S1: At least-27one mode exists S2: At least two modes exist
-36 Mode 1
-45 Mode 2
-54
-63
-72
Gyy Mag. Example -81

SVD of [Gyy] Matrix Decoupled Modes
0 5 10 15 20 25
Frequency [Hz]
www.bksv.com, 38 Bridge z24

Where to fulfil the SVD assumption best
2 3 1 a1  1 a2  2
2 v1s1 v2s2
1
1
1
2
s1 2 a1  1
v1s1
a2  2
2 v2s2
1
s2 3 a2  2
v1s1
a1  1
v2s2
 2 is a good place for estimating shape 1 from singular vector v1

 3 is a good place for estimating shape 2 from singular vector v1
www.bksv.com, 39
Frequency Domain Decomposition
Peak-Picking on the average of the normalized singular

values of the Cross Spectrum Matrix for all datasets
dB | (1.0 m/s2)² / Hz Frequency Domain Decomposition - Peak Picking

Average of the Normalized Singular Values of
Spectral Density Matrices of all Data Sets.
9
-9
-18
-27
-36
0 7 14 21 28 35
Frequency [Hz]
Brüel & Kjaer, Operational Modal Analy sis Pro, Release 3.1
Project: Car 1.axp
 Resonance frequency, fn from Singular Values

 Modeshapes,  from Singular Vectors
 But no damping,  estimated
www.bksv.com, 40
HCT, Building Model
Outline
 Introduction
 Summary
www.bksv.com, 41
Enhanced FDD Method (EFDD)
 Estimates Frequency & Damping from each data sets
 Lists Average Values and Standard deviation (Multiple datasets)
Mode Avg. Freq. [Hz] Std. Freq. [Hz] Avg. Damp. [%] Std. Damp. [%]
352.3 Hz - Enhanced FDD 352,3 3.935 564.6m 146.3m
487.7 Hz - Enhanced FDD 487,7 0,822 635.3m 200.1m
971,5 Hz - Enhanced FDD 971,0 6.135 369.3m 197.1m
 Simple curvefitting is used (linear regression)
www.bksv.com, 42
Enhanced FDD (EFDD)
 Improved natural frequency and damping included from correlation functions
 Improved Mode shape estimation from weighted sum of singular vectors:
 weight    i  si
i
SVD of Spectral Density Matrices Normalized Correlation Function
Validation of Natural Frequency Estimate Validation of Damping Ratio Estimate
Natural frequency by counting zero-crossings Damping by linear regression (simple curvefit)
www.bksv.com, 43 Bridge
Improving and enhancing FDD -> EFDD
0  Select MAC rejection level
• • (default 0,8):
• • • 
•
• • i
• H
( 0  i )2
• MAC =  0,8
s1 • H
( 0  0 )  ( i  i )
H
•
 Improved shape estimation
from weighted sum:
s2
 weight    s
i
i i
 Mode shape from weighted sum of singular vectors

» H = Complex Conjugate transpose (Hermitian) of vector/shape, 
 IFFT performed to calculate Correlation Function of SVD function

» Frequency and Damping estimated from Correlation Function
www.bksv.com, 44
Frequency Domain Decomposition (FDD/EFDD)
Determination of complete Modal Model from Responses only
FDD/EFDD procedure:
 Cross Spectral Matrix estimation for each data set, [Gyy(w)]
 Singular Value Decomposition (SVD) of Matrices
 Identification of Single Degree of Freedom (SDOF) models
from SVD curves
 Modal Parameter identification from SDOF models
» FDD: Shape, Frequency
» EFDD: Shape, Frequency, Damping
www.bksv.com, 45
Outline
 Introduction
 Summary
www.bksv.com, 46
FDD vs. EFDD: Advantages & Disadvantages
 Advantages of FDD:
– Easy & Fast to use – snap feature to peaks
– Picking at any Frequency and any Singular Value is possible
» Better understanding of structural behaviour
 Disadvantage of FDD:
– Frequency resolution no better than linespacing
– No damping estimated
 Advantages of EFDD:
– Individual and average frequency and damping are estimated
– Standard deviation of frequency and damping are estimated
 Disadvantage of EFDD:
– Distinguished peaks required in all datasets for all modes
– Fine tuning of MAC rejction level, min/max correlation
may be required for each resonance in each dataset
www.bksv.com, 47 Plate
Outline
 Introduction
 Summary
www.bksv.com, 48
Consequences of Harmonic Components
Techniques Consequences of Harmonic Components
All techniques  Potentially mistaken for a structural mode

 Potentially bias the estimation of structural modes (freq, damp, shape)
 Potentially higher dynamic range required to extract “weak” modes
FDD  The picked FFT line may be biased by the harmonic component(s)
 Harmonics must be away from the resonances (no damping estimated)
EFDD  The identified SDOF used for modal parameter estimation may be
CFDD biased by harmonic component(s)
 Harmonics must be outside the SDOF bell thereby potentially
narrowing the SDOF and resulting in poorer identification (leakage)
 Or the Harmonics must be detected and removed
www.bksv.com, 49
OMA – Harmonic Detection Implementation
Eliminating harmonic components (deterministic signals) in EFDD & CFDD techniques
 Harmonic components indicated in SVD plots using kurtosis calculations
 Harmonic components removed in SDOF model before extracting the modal parameters
=> Makes OMA much more applicable for operating mechanical structures,
where the excitation are subject to deterministic forces due to the rotating parts
www.bksv.com, 51
Probability Density Function (PDF) - background
 Probability Density Function (PDF)

– The statistical properties of a narrowband stochastic response of
random data and a harmonic component are significantly different
PDF for Pure Random Data PDF for Pure Sinusoid
  x   2
y  f x  ,    y  f x a    cosarcsin x a 
1 2 2 1
e
 2
www.bksv.com, 53
Kurtosis - Definition
The kurtosis, γ* of a stochastic variable x provides a measure of

of how peaked or how flat the probability density function of x is

E x   
 x  ,   
4

 4
*i.e. 4th normalized moment
 > 3: Leptokurtic (”Peaked” and ”fat tails”)

 = 3: Mesokurtic (Normal Distribution)
 < 3: Platykurtic (Less ”peaked” and ”thinner tails”)
 = 1.5 for a sinusoidal component
NB! Often kurtosis is calculated as  - 3
www.bksv.com, 54
Using Kurtosis as Harmonic Indicator
Hamonic Indicators
 All responses are approximate
Gaussian distributed in case of
multiple broad-banded random inputs
Gaussian Sinusoidal
 Harmonic response has a totally Probability Probability Density
Density Function Function
different distribution
Procedure:
1. Bandpass filter
» At each frequency
2. Calculate Kurtosis

E x   
 x  ,   
4

 4
3. If γ < 3 then most

likely not Gaussian
www.bksv.com, 55
Summary - Extended / Fast Kurtosis Check
Extended Kurtosis Check and Fast Kurtosis Check
for detection of “rotating components”
Identification
 Extended Kurtosis Check
– Each frequency is investigated
 Fast Kurtosis Check
– Only investigate frequencies,
where peaks (abrupt changes)
are repeated in
multiple SVD curves
 Manual Selection
 Deterministic signals are indicated by
vertical lines
Suppression
 Deterministic signals removed by
interpolation (Linear/Polynomial)
 No damaging filtering applied
www.bksv.com, 59
Example – Ship Structure – Harmonic Detection
Harmonic Components with 2.05 Hz interval

Internal Excitation
 Fundamental frequency of main
engine at 2.05 Hz (123 rpm)
 4-bladed propeller fundamental
frequency at 8.2 Hz
Test
 90 min; 16 channels; fs = 128 Hz
Structural Mode: 2.02 Hz; Harmonic Component: 2.05 Hz  3 projection channels

 Extended Kurtosis Checking
 Curve-fitted FDD, CFDD
Acknowledgement:
Sven-Erik Rosenow, Santiago Uhlenbrock and Günther Schlottmann,
University of Rostock, Germany for providing the raw measurement data
www.bksv.com, 60 Ship alone

Example – Ship Structure – Kurtosis Spectrum
Indicator:
Normal distribution
Kurtosis
Definable
Threshold
Indicator:
Sinusoidal distribution
www.bksv.com, 61
Curve-fitted Frequency Domain Decomposition (CFDD)
 Polynomial Interpolation using SDOF curvefitting
 Can handle harmonics even when coinciding with resonances
www.bksv.com, 62
Curve-fitted Frequency Domain Decomposition (CFDD)
Response Matrix: G yy (f )  H f  G F F f  H H f   H f  H H f 
 Estimate MDOF spectrum Gyy(f) by SVD
» Select Frequency Interval (Peak picking on 1st SV, MAC rejection value)
 Calculate Complex Spectrum, Py(f) of Gyy(f) (using Hilbert Transform)

 Least Squares Estimates Curvefit of Py(f) using SDOF curvefitter
2 j f T
B (f ) B 0  B 1e  B 2e 4  j f T
FRF of SDOF system: H (f )  
A (f ) 1 A1e 2  j f T  A 2e 4  j f T
 Frequency and damping estimates are obtained from the roots of A(f)
 Mode shape estimated as EFDD:  weight   i  si

i
www.bksv.com, 63
Outline
 Introduction
 Summary
www.bksv.com, 66
Automatic Mode Extraction using FDD/EFDD/CFDD
Automatic Mode Extraction using the FDD/EFDD/CFDD Techniques
Simplifies and speeds up the modal analysis task
Reduces human errors and demands less modal analysis skills
Ease preliminary analysis of a new problem
Allows for unattended testing - Structural Health Monitoring
 Modal Coherence
(light blue graph with inverse fill)
– Defines whether you have a
» Modal Domain (High)
» Noise Domain (Low)
 Modal Domain
(light green graph)
– Property defined for all modes
– Defines the frequency region
dominated by the mode
 Excludes deterministic signals
from identification using
Fast Kurtosis Check (option)
www.bksv.com, 67 Bridge FDD

Indicator Function - Modal Coherence
1 N
Discriminator function: d1 ( f 0 )   u1 ( f n ) H u1 ( f 0 ) , N: No. of Freq. Lines Avg.
N n1
Number of frequency lines, N includes a moving average of singular vectors, u for the line in question.
For closely coupled modes, N should be small to avoid smearing.
If the frequency resolution is high (Δf small), N can be higher.
1
Low: Noise dominance
0.8 High: Modal dominance
Light blue graph with inverse fill

0
www.bksv.com, 68
Indicator Function - Modal Domain
For areas, where the Modal Coherence (d1(f0)) is higher than a certain user-
defined threshold, a Modal Domain can be established around peaks.
Peaks must fulfill user-defined criteria to be qualified as a mode.
Deterministic signals are automatically excluded (option).
www.bksv.com, 69
Procedure
Automatic Identification of Modes
For each Data Set specified, repeat step 1 to 3
1. Identify the highest peak among all 1st Singular Values
2. Characterize the peak as physical or non-physical mode. Criteria:
– 1) High modal coherence; 2) Not a deterministic signals; 3) Max damping
3. Establish a modal domain around a physical mode and a noise domain around
a non-physical mode and exclude from further searches
Until (specified number of peaks found) or (peaks below dynamic headroom)
4. Reject “modes” not fulfilling a certain MAC criteria (option)
Remark:
As only the upper singular curve
is searched, repeated roots
cannot be detected!
Closely-coupled modes can.
www.bksv.com, 70 HCT, 2048 lines, N=1

Conclusion
 Automatic modal parameter extraction for
easier Operational Modal Analysis
 Valuable indicators needed for automization:

– Harmonic Indicators
– Modal Coherence
– Modal Domain
 Handling of large channel count through

use of Projection Channels
– Work fast even for large channel counts
 This automatic modal parameter extraction works for

FDD, EFDD & CFDD methods
 Automization of the data driven SSI also available
www.bksv.com, 73
Outline
 Introduction
 Summary
www.bksv.com, 74
Projection channels
 Reducing redundant information

[Gyy]
– Only a few independent row/columns exist  G11 G12

G13 G14 . . . G1N 

G 21 G22 G23 . . . 
– Many row/columns are linear combinations G31 G32

. . . 

G 41 . . . 
of the others  . . . 
 
– Much unnecessary (redundant)  .
 .
. 
. 
 
information exist GN1 . . . . . . GNN 
 Reduction of linear dependent columns by  G1p1 G1p2 

proper choice of projection channels  
G G
 2p1 2p 2 
– [G(: , [p1 p2])] =    
– [p1 p2] : projection channels  
 Np1 Np 2 
G G
 Correlation Techniques are used to find the

measurement channels with maximum information
www.bksv.com, 75
Plate unprocessed
Determining proper projection channels
 The individual correlation coefficient of the measured data:
C ij( ) 
2

E ( y i ( t  )  y j ( t ))2 

Ey i ( t  )  y i ( t )  E y j ( t  )  y j ( t ) 
 The correlation coefficient matrix of the measured data:
 1 C12 C13 C14 . . . C1N 
 
C 21 1 C 23 . . . 
C31 C32 . . . 
 
C 41 . . . 
 . . . 
 
 . . 
 . . 
 
CN1 . . . . . . 1 
 These coefficients are used to determine the

projection channels
www.bksv.com, 77
Projection channels – Single data set
 Find the channel that correlates most with other channels

using Correlation Coefficient, Cij calculations. This channel
most likely contains maximum physical information:
N
Wi  C
j1
ij
ji, i=1,…,N
Find max ([W1, W2……..WN])
 Find as many channels needed that correlates the least with

all previous found projection channels.
These channels will most likely bring most new information
 If a channel has insignificant correlation with any other

channels, it will be disgarded since it might be dead
www.bksv.com, 78 Ship
Projection channels – Multiple data sets
 All reference channels are chosen as initial projection

channels, because the reference channels are present in
all data sets
 Find as many channels needed that correlates the least

with all previous found projection channels
These channels will most likely bring most new information
(as for single data set)
 If a channel has insignificant correlation with any other

channels, it will be disgarded since it might be dead
(as for single data set)
www.bksv.com, 79 Bridge 5 –> 6

Use of projection channels in FDD, EFDD & CFDD
 N: No of measurement channels
 M: No of projection channels
[G yy(: ,[p1 p2…])] = [V]NxM[S]MxM [V]HMxM
– a SVD of the [G yy(: ,[p1 p2…])] matrix is performed

– the modeshapes are found among the M columns of [V]NxM
– where [V]HMxM is not used for anything
 The use of projection channels reduces the calculation time

» Use SVD curves to find optimum number of projection channels, M
www.bksv.com, 80
Summary, Projection channels
 Advantages of using projection Channels:

– Increased computational speed
– Reduced redundant information
– Better stabilization of physical modes (SSI)
 Disvantages of using projection Channels:

– Missing modes if number of projection channels is
too small
– Missing modes not appearing in projection channels
 Use SVD curves to find optimum number of

projection channels, M
 Harmonic detection requires use of projection channels
www.bksv.com, 81
Outline
 Introduction
 Summary
www.bksv.com, 82
Stochastic Subspace Identification (SSI)
Classes of Identification (time domain method)
 Data Driven: Use of raw time data
 Covariance Driven: Use of Correlation functions
SSI procedure
 Generate compressed input format (Hankel Matrix)
– Select total number of modes (structural, harmonics, noise)
» Based on apriori knowledge
– Select Identification Class
» Unweighted Principal Components (UPC); Principal Components (PC);
Canonical Variate Analysis (CVA)
 Estimate Parameters from Stabilization diagram
» Select interval of model order candidates (use SVD diagram)
» Estimate models (adjust tolerance criteria)
» Select the optimal model for each data set (use validation (synthesis))
 Select and link modes across data sets
www.bksv.com, 83
Hankel matrices with outputs
Matrix with constant (positive sloping) skew-diagonals
 y( 0 ) y(1) y( 2 ) y( 3 ) ... y( j  1) 
 y(1) y( 2 ) y( 3 ) y( 4 ) ... y( j ) 
 ... ... ... ... ... ... 
 Y0|i 1  Y p   y( i  1) y( i ) ... ... ... y( i  j  2) 
Y   
 Y f   y( i )
Y     y( i  1) ... ... ... y( i  j  1) 
i | 2 i  1

 y( i  1) y( i  2) ... ... ... y( i  j ) 
 ... ... ... ... ... ... 
 y( 2i  1) y( 2i ) ... ... ... y( 2i  j  2)
i << j large
i = No of modes
2i = State Space Dimension Hermann Hankel
1839 - 1873
2i+j-2 = No of samples
www.bksv.com, 84
Combined System Model used in SSI
Combined System A
Xˆ i  [ xî xî 1 xî  j 1 ]
Stationary
zero mean Measured
Gaussian Responses
White Noise Excitation Filter Structural System
(linear, time-invariant) (linear, time-invariant) C xt    yt
wt

Unknown excitation forces vt
Discrete-time Stochastic State Space Model

State Equation xt 1  Axt  wt Model of the dynamics of the system
Observation
www.bksv.com (Output) equation
yt  Cxt  vt Model of the output of the system
wt: Process noise - vt: Measurement noise - Model order: Dimension of A
www.bksv.com, 85
Linear stochastic systems, State Space Model
wt vt
xt+1 xt yt
+ C +
State Equation x t 1  A x t  w t Model of system dynamics

Output equation yt  C xt  vt Model of system output
wt and vt are unobserved, white Gaussian noise sequences
www.bksv.com, 86
Modal parameter extraction from SSI
Discrete-time Stochastic xt 1  Axt  wt wt: Process noise

State Space Model yt  Cxt  vt vt: Measurement noise
Innovation form xˆt 1  Axˆt  Ket e t: Innovation (white noise)

yt  Cxˆt  et K: Kalman gain (noise model)
A  Vi V 1
zt 1   i  zt  et 
Modal decomposition 1
zt  V xt
yt  zt  et
Right hand mode shapes:
Eigenvalues:
Non-physical Modes
i Modal frequency 
Left hand mode shapes:

Physical Modes Modal distribution of e
and damping
Initial modal amplitudes
www.bksv.com, 87
xˆt 1  Axˆt  Ket
Least squares estimation of A and C
yt  Cxˆt  et
yt xˆt 1
x A, C Axˆt : xˆt 1  Axˆt  tI
x x
x x Cxˆt : yt  Cxˆt  tII
x Assuming xˆt , xˆt 1 , yt properties:
x̂t
 Zero mean Gaussian stochastic process
 Modeled by a state space formulation
Error
 t 
x
x
x
x
x
x x̂t Then a least squares estimation gives
Gaussian white noise residuals
Result : t  Ket & t  et
I II
Conclusion: If xˆt , xˆt 1 can be determined then A and C can be

optimally predicted using least squares estimation
www.bksv.com, 88
Estimation of state vectors: Xˆ i  [ xî xî 1 xî  j 1 ]

S1
 s1 0 0 0 . . 0.
0 s2 0 . . . 

0 .
V1T 
SVD: W1OW2  U1U 2 
0 . .
 
 T   U1S1V1
0 . . s . T
. . 0 .

.
.
0 .

V2 
0 .
 
 0 . . . . . . 0
O: Compressed input format matrix (Hankel)
State space dimension

W1 , W2: Weighting matrices Selected subspace
S1 : Subspace matrix
s6
s5
X̂ i is calculated from s4
s3
V  W1i  Xˆ iW2
s2
U S V U S
1 1 1
T
1
1/ 2
1 S 1/ 2
1 1
T s1
Singular value
www.bksv.com, 89
 Parametrical Modal estimation requiring apriory knowledge of Model Order
 Physical Modes as well as Non-physical Modes are estimated
How can we separate Physical Modes from Non-physical Modes?

Physical modes are repeated for multiple Model orders!
Stabilization Diagram Number of modes in the model
7 X X X X + X +
6 X X X + X X
5 X X + X X
4 X X + X
Frequency
Stable Modes
Stable Modes not fulfilling Damping apriori knowledge
+ Stable modes are repeated in two consecutive models fulfilling user defined criteria
www.bksv.com
X Estimated parameters not fulfilling apriori knowledge of damping
X Remaining modes are considered as unstable
www.bksv.com, 90 Erfttal scratch

Selection of State Space Dimension
 ”Error diagram”
 Stabilization
diagram
 Model vs. measurements
www.bksv.com, 91
Selecting proper model order for SSI
 Final Prediction Error

– Fitting error decreases with increasing model order
– Parameter uncertainty increases with increasing
model order
Parameter Uncertainty
Final Prediction Error
Fitting Error
Optimum Choice Model Order
www.bksv.com, 92
Select and Link Modes
 Select and link modes across all Data sets
or
 Autoselection: Selects all stable modes in all data sets
www.bksv.com, 93 Erfttal Bridge

SSI-UPC Merged Data Sets Technique
 A task merging all acquired data sets into a single data set
before performing SSI is available for the UPC technique
 Advantages
– Ease-of-use
» Only a single stabilization diagram is needed to identify modes
» No selection of model using individual stabilization diagrams
» No linking across the data sets
– Increased calculation speed
» The SSI technique is applied to one common data set
 Advantages of the un-Merged SSI-UPC technique

– Gives full control over the process in difficult situations
» Local modes, weak modes, strong harmonic components,
variable mass loading etc.
www.bksv.com, 94
Shape/Matrix merging – Data Set linking
SSI SSI
 Post Separate Re-scaling  Pre Global Estimation Re-

(PoSER) scaling (PreGER)
Merging partial mode shape Merging Hankel matrices of

estimates into a global mode each setup to obtain a global
shape estimate Hankel matrix and global mode
shape estimate
H Hankel Matrix
ϕ Mode Shape
www.bksv.com, 95 Erfttal Bridge

Select Identification Class
 Principal Components (PC)

» All modes of equal strength
» Good S/N relationship in data
» Low subspace model order can be used
 Canonical Variate Analysis (CVA)

» Modes of widely different strength
» Noisy data
» High subspace model order required
 Unweighted Principal Components (UPC)

» Compromise between PC and CVA
www.bksv.com, 96
Classes of Identification (time domain method)
 Data Driven: Use of raw time data
 Covariance Driven: Use of Correlation functions
SSI procedure
 Generate compressed input format (Hankel Matrix)
– Select total number of modes (structural, harmonics, noise)
» Based on apriori knowledge
– Select Identification Class
» Unweighted Principal Components (UPC); Principal Components (PC);
Canonical Variate Analysis (CVA)
 Estimate Parameters from Stabilization diagram (manual / automatic)
» Select interval of model order candidates (use SVD diagram)
» Estimate models (adjust tolerance criteria)
» Select the optimal model for each data set (use validation (synthesis))
 Select and link modes across data sets (or Merge Data Sets - UPC)
www.bksv.com, 97
Stochastic Subspace Identification – More Tools
 Decimation
 Filtering
 Projection Channels
 Crystal Clear Stabilization
www.bksv.com, 98
Use of decimation in SSI
 The use of decimation and in SSI
 Modes can be identified within a ”limited” frequency range
30 Hz 60 Hz 90 Hz 6 Hz 12 Hz 18 Hz 24 Hz 30 Hz
64 Hz range, modes >19 Hz 25,6 Hz range, modes > 8 Hz

(No modes found below 19 Hz) (5 modes found below 19 Hz)
 Example: 1:5 scaled model of a Wind Turbine Wing

– 1 data sets, 48 transducers, 6 projection channels
– Crystal Clear SSI used
www.bksv.com, 99
Use of filtering in SSI
 The use of high-pass, low-pass, band-pass and band-stop
filtering in SSI
 Modes can be identified with better stability and

at lower State Space Dimension, SSD
200 Hz, 7 modes, SSD=80 50 Hz, 11 modes, SSD=30

(only 50 Hz shown) (i.e. using a 50 Hz low-pass filter)
 Example: 1:5 scaled model of a Wind Turbine Wing

– 2 data sets, 26 transducers, 2 refs., 3 projection channels
www.bksv.com, 100
Use of projection channels in SSI
 Reduces the calculation time dramatically

 Reduces the scatter of noise modes in Stabilization Diagram
– Better Stabilization of the physical modes
With 6 projection channels Without projection channels
 Example: Z24 Highway Bridge at Zürich

– 9 data sets, 33 transducers, 5 references
www.bksv.com, 101
Bridge DS 1
Use of Crystal Clear Stabilization in SSI
 Removes the scatter of noise modes in Stabilization Diagram
– Even Better Stabilization of the physical modes
» Sort non-physical modes from physical modes
» Influence from non-physical modes minimized
With 5 projection channels With Crystal Clear SSI

 Example: Z24 Highway Bridge at Zürich
– 9 data sets, 33 transducers, 5 references
– 28 poles (i.e. 14 physical modes) specified
www.bksv.com, 102
Bridge DS 1
Estimation of the dynamics of a system
Fitting A and C is a linear least squares estimation problem:
Ax̂t : x̂t 1  Ax̂t  tI

Cx̂t : yt  Cx̂t  tII
Traditional SSI parameter fitting:

All eigenvalues of A (physical information and noise) are fittet using
unconditional least-squares estimation
 Disregarding of the energy content in the modes
 Noise modes are scattered over the frequencies between
DC and the Nyquist frequency
www.bksv.com, 103
Estimation of the dynamics of a system
Fitting A and C is a linear least squares estimation problem:
Ax̂t : x̂t 1  Ax̂t  tI

Cx̂t : yt  Cx̂t  tII
Crystal Clear SSI parameter fitting:

Eigenvalues of A (physical information and noise) are fittet using
conditional least-squares estimation
 Condition is a automatic (or user defined) number of
modes with largest energy content in the data
 Remaining eigenvalues are considered noise modes
– Are forced above the Nyquist frequency
www.bksv.com, 104
Automatic Mode Extraction using SSI
Automatic Mode Extraction using the SSI Techniques
Simplifies and speeds up the modal analysis task
Reduces human errors and demands less modal analysis skills
Ease preliminary analysis of a new problem
Allows for unattended testing - Structural Health Monitoring
All stable modes of all estimated models of all test setups are included
www.bksv.com, 105 Building Model

Crystal Clear SSI – Automatic Mode Extraction
 Crystal Clear Stabilization Diagrams

– Better overview of physics
– Easier mode extraction
– More robust mode estimates
» Very robust damping estimates across model orders,
so selecting the optimal model order is less critical
 Automatic Mode Estimation

– Less skills required
– Less uncertainties by using all stable modes across
all model orders in all data sets (Avg. and Std. Dev. Calculated)
– Easy to automate e.g. for Structural health monitoring
www.bksv.com, 106
Summary, SSI
 State Space Models:
– Efficient Matrix Calculations in pc (e.g. MatLab)
 Stochastic Data Identification using Least Squares Method:

– Subspace Identification using SVD (reduce problem size)
– Three Traditional Identification Classes, (PC, CVA & UPC)
 Stabilization Diagram:
– Identify Physical Modes from Poles of Combined Model
– Apriori Knowledge of Damping can be applied
– User definable tolerance criteria (fd, , MAC & Initial MAC)
 Validate Model vs. Measurements to select model order
 Automatic & Crystal Clear mode extraction
www.bksv.com, 108
Outline
 Introduction
 Summary
www.bksv.com, 109
Validation
 Common sense (does it look like a mode shape?)
 Complexity plots
» Normal vs. Complex shapes
 MAC (Modal Assurance Criterion)

» Check similarities of shapes
 Use several different methods

» Compare frequencies, damping, complexity and shapes (CrossMAC)
www.bksv.com, 110 Bridge/Erfttal export

Modal Phase Collinearity, – Modal Complexity
 Modal Phase Collinearity for mode r, MPCr expresses

– Consistency of linear relationship between the
real and the imaginary part of the mode shape
(Sxx  Syy ) 2  4S2 xy

MPCr 
(Sxx  Syy ) 2
where
Sxx  Rer Rer Syy  Imr Imr Sxy  Rer Imr

T T T
& &
and 0  MCF  1
 Modal Complexity Factor, MCFr
MCFr  1  MPCr
www.bksv.com, 111
Modal Assurance Criterion (MAC)
A measure of the consistency (degree of linearity) between
estimates of mode shape vectors:
cr  dr
T 2
*
MACcdr 
crT  crdrT  dr
* *
where cr and dr are estimates of the mode shape for mode r
from two different models c and d (CrossMAC)
Consistency between:
 Different experimentally determined mode shapes
– different projection channels (columns)
– different curvefitting algorithms, FDD vs. SSI
 Experimentally and analytically determined mode shapes

Modal Assurance Criterion (MAC)
A measure of the similarity between estimated modal vectors:
r  s
T 2
*
MACrs 
rT  rsT  s
* *
where r and s are the modal vectors for mode r and mode s
from the same model (AutoMAC)
High (Auto) MAC between modal vectors can be due to

 Insufficient spatial samling of the modes.
– not sufficient number of DOF’s in the measurements
 An estimated modal vector is caused by a forced vibration
(i.e. an ODS) and not a mode

Effects of time varying systems, 1
 Assumptions:
Combined System
Stochastic Deterministic
Stationary
Zero Mean Measured
Gaussian Responses
White Noise
Loading System Structural System
(linear, time-invariant) (linear, time-invariant)
Broad banded Narrow banded
 Four scenarios:
– Loading System: a) Time Varying or b) Time Invariant
– Structural System: a) Time Varying or b) Time Invariant
www.bksv.com, 114
Effects of time varying systems, 2
Stochastic Deterministic
Input Loading System Structural System Output
Structural System
Time Invariant Time Varying
~ Spectral Density Correct ~ Spectral Density Incorrect
Time Invariant
~ Clear Peaks ~ Incorrect Modal parameters

~ Good Results in FDD of all methods
EFDD and SSI ~ Bad validation in SSI
~ Good SSI validation no Covariance equivalence
Loading System
Covariance equivalance AVOID!

~ Spectral Density peaks
Time Varying
correct, valleys not

~ Good modal parameters
results in all methods AVOID!
~ Bad validation in SSI
no Covariance equivalence
 Run up/down tests gives good modal parameters,

but might have bad SSI validation
www.bksv.com, 115
Harmonic Excitation
 Mathematical Assumptions for OMA

– Input excitation forces (combined model) can be approximated
by stationary filtered zero mean Gaussian white noise
 Real-life Conditions – Harmonic Excitation

– Excitation forces of mechanical structures are typically a
combination of:
» Harmonic components from the rotating and reciprocating parts
» Broadband & Harmonic internal excitation from e.g. bearings and combustion
» Broadband ambient excitation from e.g. air turbulence, acoustic radiation and
road vibrations
– Civil engineering structures may have responses superimposed by
harmonics from e.g. ventilation systems, turbines and generators
www.bksv.com, 117
SVD – Harmonic Excitation
 Singular Value Decomposition (SVD) of Response Spectral Matrix
Sinusoidal components revealed:

Peaks in several Singular Values
 Short Time Fourier Transform (STFT) of responses
Broadband random and stationary

sinusoidal components
(1st, 3rd and 4th harmonic of rotational frequency)
 Fast or Extended Kurtosis Check

» Use EFDD & CFDD
www.bksv.com, 118
Consequences of Harmonic Components
Techniques Consequences of Harmonic Components
All techniques  Potentially mistaken for a structural mode

 Potentially bias the estimation of structural modes (freq, damp, shape)
 Potentially higher dynamic range required to extract “weak” modes
FDD  The picked FFT line may be biased by the harmonic component(s)
 Harmonics must be away from the resonances (no damping estimated)
EFDD & CFDD  The identified SDOF used for modal parameter estimation may be
biased by harmonic component(s), unless Harmonic Detection is used
 Harmonics must be outside the SDOF bell thereby potentially
narrowing the SDOF and resulting in poorer identification (leakage)
SSI  The SSI methods will estimate both harmonics and modes. The modes
(PC, UPC, CVA) are estimated correctly even for harmonics very close to the modes
 Information in the time signal is used both to extract the harmonics and
the modes, therefore the recording time should generally be longer
www.bksv.com, 119
SVD - Importance of proper excitation, 1
 Response from y(t) = []q(t)

Modal Model:
=>
 Correlation of
responses: [Cyy()] = E{y(t+)y(t)T}
[Cyy()] = E{[]q(t+)q(t)H[]H} = [] [Cqq ()] []H
i.e. by Fourier Transform:
[Gyy(w)] = [] [Gqq(w)] []H
 Same form as Singular Value Decomposition of Response Spectral Matrix:
[Gyy(w)] = [V] [S] [V]H

where singular vectors in [V] are orthogonal and [S] is diagonal
www.bksv.com, 120
SVD - Importance of proper excitation, 2
Validated using SVD of response spectral matrix
- Plate with heavy coupled modes
1: One fixed broadbanded

excitation source
2: Two fixed broadbanded

excitation sources
3: Two fixed broadbanded

excitation sources
(different points than above)
4: Two moving broadbanded

excitation sources
(correct modal estimates)
www.bksv.com, 121
Importance of proper excitation, 3
Validated using SVD of response spectral matrix
- Plate with heavy coupled modes
1: One fixed broadbanded

excitation source
2: Two fixed broadband

excitation sources
3: Two fixed broadband

excitation sources (different
points than above
www.bksv.com, 122
Importance of proper excitation, 4
4: Two moving broadbanded

excitation sources
 Coupled modes detected Bending mode:

in case 2 and 4 using - 186,6Hz, 3,4%
EFDD: - second SV
Torsion mode:
- 188,9Hz, 3,7%
- first SV
 Modes at 495Hz and 520Hz detected much better in case 4
www.bksv.com, 123
SVD – Defect transducer, cable break etc.
 Poor signal quality

– Transducer disconnected
– Defect Transducer
– Cable Break
– Etc.
Conclusion
 SVD plot should always be investigated
no matter which method is used for modal extraction
 SVD plot used for estimating

number of projection channels
www.bksv.com, 124
Conclusion - Overview and practical experiences
 SVD as validation tool

– Proper Excitation
– Harmonic Contents
– Signal quality
– Estimate number of projection channels
 Use of Projection Channels essential in case of larger ch. Counts
 Low-pass filtering and decimation essential for

modal identification using SSI over several octaves
 UPC, PC and CVA algorithms give almost identical results

– Trend on required SSD: PC lowest, and CVA highest – use UPC!
– Use “Crystal Clear” SSI
www.bksv.com, 130
Outline
 Introduction
 Summary
www.bksv.com, 131
OMA Conclusion
General Conclusions:
 Dedicated Technique for Operational Modal Analysis

offering reliable estimation of modal parameters
without known input force
 The method makes modal testing easier on large

structures as no elaborate excitation is needed
 Use data acquired during operation

for extraction of modal data
 The technique used in civil engineering since early 90’s

also used in mechanical applications this Millennium
www.bksv.com, 132
Conclusion (advantages)
 No elaborate fixturing of structures, shakers and force transducers
– No test rigs needed
– Short setup time
– No dynamic loading from shakers and stingers
– No crest factor problems as when using hammers
– No potential destruction of structure
 Modal model can represent real operating conditions

– True boundary conditions
– Actual force and vibration levels
 Only natural random or unmeasured artificial excitation required
 No interference or interruption of daily use
 Modal testing can be applied in parallel with other applications
www.bksv.com, 133
Conclusion (concerns)
 Unscaled (Non Calibrated) Modal Model

– No Forced Response and Modification Simulations
 More Operator Skills required!?

– Some apriori knowledge is advantageous (-)
– Pre-analysis is often needed (Time/Freq.) (-)
– New technique to most engineers (-)
– Automatic Mode Extraction exist for all methods (+)
 Large time histories might be required

– More data handling capacity needed
– Higher computational power needed
www.bksv.com, 134
Operational Modal Analysis – In brief
In-flight Testing or Taxi Testing Powertrain Testing Civil Structures
 Structures impossible to test using hammer or shaker excitation can be tested

– Accurate modal identification using natural excitation
 Modal model represents real-life structural behaviour
– Measurements done under real operating conditions
– Modal model represents true boundary conditions and actual force & vibration levels
 Simple, straightforward and fast setup, measurement and analysis by only
using accelerometers
www.bksv.com, 135
Operational Modal Analysis – Why?
 Structures impossible or difficult to excite can be tested
– Large or fragile structures
– Structures with boundary conditions difficult to simulate
– Uncontrolled and/or unmeasured internal or ambient excitation
 Modal model represents real operating conditions
– True boundary conditions
– Actual force and vibration levels
Testing of large structures
 Simple, straightforward and fast setup, measurement and like bridges, aircraft and ships
analysis by only using accelerometers
– No elaborate fixturing of structure, shakers and force
transducers
– Measurement procedure quite similar to ODS
 No interference or interruption of daily use
– Testing can be done in-situ using the natural excitation
– No costly downtime
 Modal testing can be applied in parallel with other applications Easy interpretation of results
e.g. order analysis and sound quality using animated geometries
 Inexpensive instrumentation
– No test rigs (in-situ measurements), shaker stands, shakers,
impact hammers, signal generators etc.
www.bksv.com, 136
Operational Modal Analysis – Application Cases
Automotive
– Engine, transmission, suspension and exhaust system
testing during operation in test cells
– On-road full-body testing of vehicles/motorcycles
– Trains and trams during operation
Aerospace & Defence
– Flight flutter certification of aircraft using ambient and artificial
unmeasured excitation
– In-flight testing of helicopters for various payloads Testing of ship at sea using natural
– Testing of satellites using acoustic excitation internal and ambient excitation
– Launch of missiles Fundamental frequency of main
Maritime engine close to 2nd mode of hull
– Ships at sea, submarine signature testing, diesel engines Fundamental frequency automatically
etc. removed and 2nd mode extracted
Rotating Machinery
– Engines, generators, turbines, compressors, etc. in operation
Civil Engineering
– Structures excited by ambient excitation like wind, waves,
traffic and seismic micro-tremors: bridges, buildings, towers,
wind turbines, dams, off-shore platforms, etc.
Consumer products
– White goods, power tools, disk drives, etc. Raw measurement data provided by S.E.
And many others … Rosenow et al., University of Rostock, Germany
www.bksv.com, 137
Operational Modal Analysis – Solution
 Patented frequency domain techniques
– Frequency Domain Decomposition FDD, EFDD and
CFDD
– Automated mode estimation
– Automatic suppression of harmonic components
– Quick and easy-to-use peak-picking technique
– For lightly to medium damped structures
 Time domain techniques
Easy spectral peak-picking
– Stochastic Subspace Identification (SSI): UPC, PC, CVA FDD, EFDD and CFDD
– Automated mode estimation
– Most accurate estimates of natural frequencies,
damping and mode shapes
– For lightly, medium and heavily damped structures
 Validation of results
– Synthesis of response spectra
– Modal Assurance Criterion (MAC) plots/tables
– Powerful animation capabilities
 Several projects can be run simultaneously
– Easy comparison of results Crystal Clear stabilization diagrams
 Part of complete solutions for structural dynamics (CC-SSI)
– Classical Modal Analysis, ODS Analysis, Model
Updating, Transducers and Acquisition hardware …
www.bksv.com, 138
Operational
Modal Analysis
www.bksv.com
www.bksv.com, 139

Operational Modal Analysis 2012 (OMA Notes)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Operational Modal Analysis 2012 (OMA Notes)

Uploaded by

Copyright:

Available Formats

Operational

 Frequency Domain Decomposition, FDD

 Enhanced Frequency Domain Decomposition, EFDD

 Comparison between FDD & EFDD

 Harmonic Detection, CFDD

 Automatic Mode Extraction

 Stochastic Subspace Identification, SSI

 Measurement results & Validation

 Determination of Modal Model under operational conditions

 Used successfully in Civil Engineering applications (Ambient Modal)

 “Recently” introduced to Mechanical Engineering applications (OMA)

 Single data set: No reference transducers needed

 HCT is located in downtown Vancouver, Canada

4 Data Sets, 24 DOFs

 Build at Flensburger Schiffbau Gesellschaft.

The measurements and analysis presented here are performed by

 5 1 scaled down model

1 Data Set, 48 DOFs

 Repeated Roots 2 Data Sets, 6 DOFs

PULSE, MTC, OMA

[m/s²] Time(R espo nse) - Input

-40  Output is related to Input by FRF estimators

Input Input Spectrum

0 400 800 1,2k 1,6k 2k 2,4k 2,8k 3,2k

Only the Modes of the System need to be identified

 If the structure is big

 In the experiment, it is difficult to reproduce

 Output only modal analysis

Force Spectrum Structural System

But this is in general not the case!

 Output only modal analysis

Force Spectrum Structural System

Both the “peaks” originating from the excitation signal and

 Output only modal analysis

Force Spectrum Structural System

Computational as well as measurement noise may result in

 Output only modal analysis

Force Spectrum Structural System

The “Modes” in the combined spectrum contains information of

Unknown excitation forces

extracted from estimated model of Combined System

Practical Excitation {F}

 Perfect excitation from OMA point of view 

 We typically consider that the excitation

 Apparently, the quality of OMA results

 Frequency Domain Decomposition, FDD

 Enhanced Frequency Domain Decomposition, EFDD

 Comparison between FDD & EFDD

 Harmonic Detection, CFDD

 Automatic Mode Extraction

 Stochastic Subspace Identification, SSI

 Measurement results & Validation

Experimental Data Signal Processing Modal Parameters

Experimental data Signal Processing

Example: Quadrature Picking in Mobility based Modal Analysis

Measured Data fn, , 

EFDD (Simplest fit)

 Frequency Domain Decomposition, FDD

 Enhanced Frequency Domain Decomposition, EFDD

 Comparison between FDD & EFDD

 Harmonic Detection, CFDD

 Automatic Mode Extraction

 Stochastic Subspace Identification, SSI