OSE801

Engineering System Identification

Spring 2010

Lecture 2: Elements of System Identification:

Identification Process and Examples

Instructors: K. C. Park (Division of Ocean Systems Engineering) !

Y. J. Park (Division of Mechanical Engineering)!
Why do we need a good analytical model?

  A good analytical model can provide the experimentalist

with instrumentation needs, e.g., how many sensors and
where to mount them, the location(s) and levels of
excitations, anticipated response records( frequency
ranges, and response magnitudes), among others.

  No one model can serve for different applications:

is it for control? For damage detection? Or design
optimizarion? Etc. .

  Eventually, it is meaningful to obtain experimentally

verified models for the engineers to utilize them for
design improvements, performance evaluation, cost
reduction, and certification for codes and standards.
Preliminary test runs (from a single to more than dozens runs)

1. Check each output to examine the sampling rates set from the
analytical model is adequate;
2.  Readjust the filter bands (roll-off ranges, etc.) if necessary;
3.  Readjust the signal conditioners for each output line if necessary;
4.  Check all the lines from the sensors and exciters to the A/D convertor
lines that are fed into the computer;
5.  Obtain the auto and cross correlation plots and examine carefully
whether the frequency contents as well as the anti-resonances are
adequately detected;
6.  Now, make sure to minimize the temperature variations and
environmental vibrations during experiments;
7.  Run test, but if at all possible, repeat several runs and obtain
preliminary (input-output ratio) transfer functions or preliminary
impulse response functions;
8. If necessary, adjust sensor locations and repeat Steps 1 - 7.
Common overlook by novice experimentalists
(including myself)

1. Inadequate pre-testing modeling;

2. Wrong choice of excitations;
3. Wrong choice of sensors and their locations;
4. Mal-attachments of sensors and actuators;
5. Assuming textbook boundary conditions;
6. Inconsistency of units employed;
7. Ignoring noises and other uncertainties;
Of a plethora of system identification techniques, which
one(s) should I choose?

1.  Know the system characteristics that you are trying to

identify (This is where a good analytical model and
subsequent analyses play key roles!)

2.  Begin with a technique that you are most familiar. If you
find that the one you are familiar with may not be
adequate, then search for other techniques

3.  Be mindful of the limitations of the chosen identification

technique; no technique is free from any deficiencies.

4.  Do not stretch too much to stick with a technique just

because you are familiar with. A good engineer remains
a good one by adapting himself/herself to new
techniques available.
The following examples are provided
to give you a glimpse of the levels of
your learning experience as to what
kind of problems you may be able to
tackle after a successful completion
of this course.

We may undertake similar problems

as part of our term projects, notably for
control problems.
Example: Damage Identification utilizing
Transmission zeros

Poles vs. Zeros

•  Poles used extensively for:
poles
–  System Identification
–  Modal Analysis
–  Control Design
–  Input Tailoring of
Resonators
•  Zeros ignored except for:
–  Control Applications:
•  Deadband avoidance
zeros
•  Closed-loop
Performance Indication
•  Sensor/Actuator Pair
Selection and
Placement
 Motivation

•  Most model-based DD methodologies

require a large number of sensors.
–  Closely tied to modern modal testing
practices.
–  Analysis uses measured modes/mode shapes
as starting point.
–  Large matrix size, large computational cost.
•  Present method may require fewer sensors at select locations.
–  Localized identification, separate processing and observation of
select locations
–  Smaller matrix size, smaller computational cost.
 DD/HM Related Research
Global Local
Flexibility/Modeshape Mayes [1992], Pandey [1991,1994], Toksoy [1994] Peterson, Doebling, Alvin [1998], Park, Reich, Alvin [1998]
Updating Zimmerman & Kaouk [1994, 1995] Doebling [1996], Denoyer [1996], Hemez [1993]
Control/Filter Lim [1995, 1996], Cobb & Leibst [1997], Slater & Shelley [1993] Kranock [1998]
Transmission Zero Afolabi [1987] This thesis

•  Flexibility/modeshape - direct comparison of undamaged/

damaged values related to sensor/DOF location
•  Updating - correlation of analytical DOF to model changes
•  Control/filter - minimize modal force error though fictitious
controller/filter
•  Transmission zero - comparison of FRF elements correlated
to sensor/DOF locations
 Motivation, cont’d.
Global vs. Local, cont’d.
•  In the global case stiffness and mass properties are
associated with nodes. globally coupled
⎡k1 + k2 −k2 0 ⎤
⎢ ⎥ A node may have
Kg = ⎢ −k2 k2 + k3 −k3 ⎥ contributions from
⎢⎣ 0 −k3 k3 ⎥⎦ many elements.

•  In the local case stiffness and mass properties are

associated with elements. uncoupled
⎡k1 0 0 0 0 ⎤
⎢ ⎥ A local node has
⎢ 0 k2 −k2 0 0 ⎥ contributions only
K = ⎢⎢ 0 −k2 k2 0 0 ⎥ from the corresp.
⎥ element.

⎢ 0 0 0 k3 −k3 ⎥
⎢ ⎥
⎣ 0 0 0 −k3 k3 ⎦
 Definition of Transmission Zeros

yg ( s) = H g ( s) fg (s ) yg ( s ) : displacement output

H g ( s) = ( Kg + iωDg − ω 2 Mg ) = K g −1
−1
fg (s ) : force input

cof K g i, j
hgi, j (s ) =
det K g

K gj,i is the submatrix of K g without

cof K gi, j = ( −1)i+j det K g j ,i
rows j and columns i

zeros of hgi, j ( s) ⇒ roots of cof K gi, j = 0

Transmission Zeros, cont’d.
Transmission zeros of
the MIMO transfer
function

Antiresonances of the
SISO element of the
transfer function matrix

Partitioned vs. global flexibility
⎡ K + iωD − ω 2 M
( ) Bφ ⎤ ⎧⎪ φˆ ⎫⎪ ⎧⎪ fˆφ ⎫⎪
⎢ φ φ φ
⎥ ⎨ ⎬ = ⎨ ⎬
⎢⎣ BφT 0 ⎥⎦ ⎪⎩λˆ ⎪ ⎪⎩ 0 ⎪⎭
b ⎭

(4) φˆ = H φ fˆφ

H φ = Zφ − Zφ Bφ ( BφT Zφ Bφ ) BφT Zφ
−1

−1
(
Zφ = Kφ + iω Dφ − ω Mφ 2
)

φ = Φφ−1Lug LT Φ−φ T fφ = fg

H φ = Φ−1
φ LH g L Φφ H g = ( K g + iωDg − ω 2 Mg )
T −T −1
Roadmap
 A Simple Example
Global TZ [Afolabi, 1987]

⎡ k1 + k2 −k2 0 ⎤ ⎡m1 0 0 ⎤
Kg = ⎢⎢ −k2 k2 + k3 −k3 ⎥⎥ Mg = ⎢⎢ 0 m2 0 ⎥⎥
⎢⎣ 0 −k3 k3 ⎥⎦ ⎢⎣ 0 0 m3 ⎥⎦

Independent of k1

f(k1, k2, k3)

zeros of node 1: z g1,1 = k2 k3 + (k3m2 + k2 m3 + k3 m3 )s 2 + m2 m3 s 4
f(k1, k2, k3)

zeros of node 2: z g2,2 = k1k3 + k2 k3 + (k3m1 + k1m3 + k2 m3 )s 2 + m1m3 s 4
zeros of node 3: z g3 ,3 = k1k2 + k1k3 + k2 k3 + (k2 m1 + k3m1 + k1m2 + k2 m2 )s 2 + m1m2 s 4
 Simple Example, cont’d.
Element-by-element (Localized) TZ

⎡ 1 0 0 ⎤
SL = ⎢⎢ −1 1 0 ⎥⎥
⎢⎣ 0 −1 1 ⎥⎦

so that the localized TF matrix elements are:

Independent of k1

Independent of k3

Independent of k2

zeros of element 1: zε 1,1 = k2 k3 + (k3m2 + k2 m3 + k3 m3 )s 2 + m2 m3 s 4
zeros of element 2: zε 2,2 = k1k3 + (k3m1 + k3m2 + k1m3 + k3 m3 )s 2 + (m1m3 + m2 m3 )s 4
zeros of element 3: zε 3 ,3 = k1k2 + ( k2 m1 + k1m2 + k2 m2 + k1m3 + k2 m3 )s 2 + ( m1m2 + m1m3 )s 4
 Simple Example, cont’d.
Comparison of Global FRFs
Collocated input/output sets

Damage in element 2

hg1 ,1

hg2,2

hg3 ,3
 Simple Example, cont’d.
Comparison of Localized FRFs

Both zeros are

invariant - damage
occurs here.

hε1 ,1

hε 2,2

hε 3 ,3
Indeterminate Plate
•  12 elements, 3 DOF per node
(1 displ. & 2 rotations)
•  Strains can be either Mindlin-
or Kirchhoff-type
•  Strain-displacement matrix
limitations:
nεe = ne − nRe
where
nεe → # strain DOF / element
ne → # elemental DOF / element
nRe → # RBM / element

Also, Se must have full row rank.

Indeterminate Plate DD Results

Global TZ

Strain-basis TZ

 Ladder “Limited Mode” Problem

•  16 elements, 3 DOF per node (2 displ. & 1 rotation)

•  Lowest 12 modes plus residual flexibility term identified due
to burst-random input
•  Cumulative TZ variation computed based on state-space
form of modal values
 Ladder Limited Mode Case, cont’d.
•  On the left, the full analytical spectrum is used to compute
the TZ variation. On the right, only the 12 identified modes
plus the residual is used to compute the TZ variation.

Full Spectrum
Limited Modes

 NUPEC RCCV Experimental Test

•  “Scale model” of a Reinforced Concrete Containment

Vessel (RCCV)
•  Vibration test histories due to base motion recorded before
and after an internal pressure test (used to create damage
typical for this application)
RCCV Experimental Test, cont’d.

•  22 accelerometer locations
chosen for system identification
•  8 modes identified below 120 Hz
for both tests
•  After realization, modes
transformed to beam model for
damage detection analysis
RCCV Experimental Test, cont’d.
References"

1.  Alvin, K. F., Robertson, A. N., Reich, G. W. and Park, K. C., Structural system identification: from reality to models,
Computers & Structures, 81(2003), 1149-1176. !
2.  Reich, G. W. and Park, K. C., A Theory for Strain-Based Structural System Identification, Journal of Applied Mechanics,
68(4), 521-527. !
3.  Alvin, K. F. and Park, K. C., Extraction of Substructural Flexibilities from Global Frequencies and Mode Shapes, AIAA
Journal, vol. 37, no.11, 1999, p. 1444-1451. !
4.  Park, K. C. and Reich, G. W., A Theory for Strain-Based Structural System Identification, in: Adaptive Structures and
Technologies, ed. by N. W. Hagood and M. J. Attala, Technomic Pub., 1999, 83-93. !
5.  Park, K. C. and Reich, G. W., A Procedure to Determine Accurate Rotations from Measured Strains and Displacements
for system Identification, Proc. 17th International Modal Analysis Conference, 8-11 February 1999, Kissimmee, FL. !
6.  Robertson, A. N. and Park, K. C., An Investigation of Time Efficiency in Wavelet-Based Markov Parameter Extraction
Methods, Proc. 1998 AIAA SDM Conference, Paper No. AIAA-98-1889, April 20-24 1998, Long Beach, CA. !
7.  Robertson, A. N., Park, K. C. and Alvin, K. F., Extraction of Impulse Response Data via Wavelet Transform for Structural
System Identification, ASME Journal of Vibrations and Acoustics, 120, No.1, January 1998, 252-260. !
8.  Robertson, A. N., Park, K. C. and Alvin, K. F., Identification of Structural Dynamics Models Using Wavelet-Generated
Impulse Response Data, ASME Journal of Vibrations and Acoustics, 120, No.1, January 1998, 261-266. !
9.  Alvin, K. F., Park, K. C. and Peterson, L. D., Extraction of Undamped Normal Modes and Full Modal Damping Matrix
from Complex Modal Parameters, AIAA Journal, Volume 35, Number 7, 1997, 1187-1194 !
10.  G. W. Reich and K. C. Park, Localized system identification and structural health monitoring from vibration test data,
Proc. 1997 AIAA SDM Conference, Paper No. AIAA 97-1318, April 7-10 1997, Kissimmee, FL. !
11.  Robertson, A. N., Park, K. C. and Alvin, K. F., MIMO System Structural System Identification: A Revisit via Wavelet
Transform, Proc. the 14th Intnl. Modal Analysis Conf., Dearborn, MI, 12-15 Feb. 1996. !
12.  Alvin, K. F. and Park, K. C., A Second-Order Structural Identification Procedure via System Theory-Based Realization,
AIAA Journal, 32(2), February 1994, 397-406. !
13.  Alvin, K. F., Park, K. C. and Peterson, L. D., A Minimal-Order Experimental Component Mode Synthesis: New Results
and Challenges, AIAA Journal, 33(8), August 1995, 1477- 1485. !
14.  Alvin, K. F., Peterson, L. D. and Park, K. C., Experimental Identification of Normal Modes and Damping in an Actively
Controlled Structure, Proc. 1994 AIAA SDM Conference, Paper No. AIAA 94-1686, April 18-21 1994, Hilton Head, SC. !
15. Alvin, K. F., Park, K. C. and Peterson, L. D., Consistent Model Reduction of Modal Parameters for Reduced-Order
Active Control, AIAA Journal of Guidance, Control and Dynamics, 18(4), July-August 1995, 748-755. !
16. Alvin, K. F., Peterson, L. D. and Park, K. C., A Method for Determining Minimum-Order Mass and Stiffness Matrices
from Modal Test Data, AIAA Journal, 33(1), January 1995, 128-135. !