LA-UR- 99-4881Proceedings of Asia-Pacific Vibration Conference, Orchard Hotel, Singapore, December 13-15, 1999.

EXTRACTION OF RITZ VECTORS USING A

COMPLETE FLEXIBILITY MATRIX
Hoon Sohn and Scott W. Doebling
Engineering Sciences & Application Division
Engineering Analysis Group (ESA-EA), MS C926
Los Alamos National Laboratory
Los Alamos, New Mexico 87545, USA

ABSTRACT
Load-dependent Ritz vectors are proposed as a
potential alternative to modal vectors for damage
diagnosis problems. An experimental procedure to
extract Ritz vectors from vibration test data is proposed.
The extraction procedure is based on a measured
complete flexibility, and the proposed approach allows
the generation of Ritz vectors from arbitrary load
patterns. The new extraction procedure is demonstrated
using a grid-type bridge structure constructed and tested
at the Hyundai Institute of Construction Technology
(HICT), Korea.

INTRODUCTION
Load-dependent Ritz (or Lanczos) vectors have
shown potential advantages in structural dynamics over
modal parameters. For the reduction of linear dynamic
systems, Ritz vectors are shown to be more effective to
approximate the response quantities of interest using a
smaller number of independent basis vectors than the
best known modal vectors. The Ritz vectors also have
been widely used to find partial extremal solutions of
large eigenvalue problems and to reanalyze structural
system with a few localized modifications. In the
research field of structural monitoring and damage
diagnosis, Ritz vectors have been shown to offer
superior performance in identifying damage. However,
the Ritz vectors have existed only in numerical analyses
until Cao and Zimmerman [1] proposes a state-spacebased technique to extract Ritz vectors from measured
vibration tests.
In this paper, a new extraction procedure for Ritz
vectors is proposed. The procedure is based on a
complete flexibility matrix, which is obtained from the
dynamic test data. The proposed method enables to
generate Ritz vectors from any imaginary load patterns,
and the contribution of unmeasured modes to the
complete flexibility is explicitly considered in the
extraction procedure. The performance of the new

extraction procedure is demonstrated using an

experiment of a grid-type bridge structure.
This paper is organized as follows: First, the
proposed extraction procedure based on a measured
flexibility matrix is described. Next, we briefly describe
the grid-type bridge model employed in the
experimental study and the finite element model
corresponding to the test structure. The proposed
extraction procedure is then demonstrated and compared
with the state-space-based method using the test data.
Finally, this paper is concluded with a summary and
discussions.

EXPERIMENTAL EXTRACTION OF
RITZ VECTORS
In this section, we present a new extraction
procedure of Ritz vectors based on a measured
flexibility matrix. A close look at the analytical
generation procedure in Reference [1] reveals that the
generation of Ritz vectors uses the flexibility matrix G
(defined here as the inverse of the stiffness matrix)
rather than the stiffness matrix itself. The extraction of
Ritz vectors starts with the assumption that the dynamic
loading F(t) can be separated into a spatial load vector f
and a time function u(t):
F( t ) = f u ( t )

(1)

If the modal vectors are mass-normalized such that

V T KV = and V T MV = I , the flexibility matrix can
be represented with the modal parameters:

G = K 1 = V1 V T

(2)

where is the diagonal eigenvalue matrix and V is the

corresponding eigenvector (modal vector) matrix.
In most experimental modal analyses, only a few
lower modal frequencies and modal vectors are
identified. For this case, the flexibility matrix is divided
into the modal flexibility, which is formed from the
estimated frequencies and modal vectors, and the
residual flexibility formed from the residual modes [2]:

G = G m + G r = Vm -1m VmT + Vr -1r VrT

(3)

LA-UR- 99-4881Proceedings of Asia-Pacific Vibration Conference, Orchard Hotel, Singapore, December 13-15, 1999.

where the subscripts m and r denote the measured and

residual quantities, respectively. Here, the modal
flexibility G m is easily constructed using the measured
natural frequencies and modal vectors. The residual
flexibility G r is the contribution of the unmeasured
dynamic modes to the full flexibility matrix. In general,
the contribution of lower modes, which are normally
estimated in experimental modal analyses, are more
significant than those of higher modes because the
contribution of each mode is inversely proportional to
the magnitude of the corresponding natural frequencies.
However, as shown in later examples, the contribution
of the residual flexibility cannot be neglected in the
extraction procedure of Ritz vectors.
The residual flexibility is conducted using a residual
function R () , which is computed by subtracting the
reconstructed response of the identified modes from the
measured frequency response functions (FRFs) H() :

R () = H() + 2 Vm m 2 I

VmT

= G r 2

(4)

Curve-fitting R () over a set of frequency samples

( set ) yields an estimate of the residual flexibility:
Gr =

R ( )

set

(5)

In this paper, H( ) and R ( ) are assumed to be

squared matrices. That is, individual modal testing is
conducted by applying an impulse excitation force to a
single degree of freedom (DOF) and the excitation is
repeated for all measured DOFs. Reference [2]
addresses the extraction of a complete flexibility when
the number of excitation points are less than that of the
response points. While damping is not included in the
formulation of the residual flexibility, Equation (4) can
be easily modified to take into account the damping.
From the measured flexibility matrix G (= G m + G r )
and the analytical mass matrix M, the first Ritz vector
can be computed as:

r1 = G f

(6)

where f is the spatial load distribution vector defined in

Equation (1). The first Ritz vector is, then, massnormalized as:

r1 =

[r

T
1

r1
M r1

1
2

(7)

Then, the following Ritz vectors are recursively

generated.
Assuming the mass matrix times the
previous Ritz vector Mrs 1 as a load, the recurrence
relationship computes the next Ritz vector rs :

rs = GM rs 1

(8)

The linear independence of Ritz vectors is achieved

using the Gram-Schmidt orthogonalization:
~
rs = rs

(r
s 1

T
t

M rs rt

t =1

(9)

Finally, the current Ritz vector is mass-normalized:

~
rs
rs =
1
(10)
~
rsT M ~
rs 2

It is worthwhile to compare the proposed extraction

procedure with the state-space procedure proposed by
Cao and Zimmerman. Since the spatial load distribution
vector f in Equation (6) can be assigned arbitrarily, the
flexibility-based method is able to generate different
sets of Ritz vectors. On the other hand, the state-space
matrices estimated from the experimental modal
analysis retain the information of the actual load pattern
used in the modal test. Therefore, the state-space-based
method only identifies the Ritz vectors corresponding to
the specific excitation pattern used in the actual modal
testing. Furthermore, the contribution of the residual
flexibility is not included in the state-space method.
Note that both methods require an appropriate
approximation for the mass matrix. Damage often
causes stiffness deterioration without significant
changes of a system mass. Therefore, stiffness changes
are the main concern of damage detection, and the exact
estimation of the mass matrix is not necessary for
damage diagnosis problems.

Figure 1: An overview of a grid-type bridge structure

AN EXPERIMENTAL BRIDGE MODEL

For this study, a grid-type bridge model has been
constructed and tested at the Hyundai Institute of
Construction Technology (HICT), Korea (see Figure 1).
The steel bridge model consists of two parallel girders
and six evenly spaced cross beams connecting the two
girders. The girders are steel rectangular tubes and the
cross beams are C-shape members. Using impact
excitations, we extract Ritz and modal vectors from the
vibration response of this test structure.

LA-UR- 99-4881Proceedings of Asia-Pacific Vibration Conference, Orchard Hotel, Singapore, December 13-15, 1999.

A SA-390 signal analyzer with four channels is used

for the analog to digital conversion of accelerometer
signals and the Fast Fourier Transform (FFT)
calculation. Data acquisition parameters are specified
such that a frequency response function (FRF) in the
range of 0 to 100 Hz could be estimated. Each spectrum
is computed by averaging three 8 seconds long time
histories. A total of 2048 points are sampled for a 8
second time period and this sampling rate produces a
frequency resolution of 0.125 Hz. An exponential
window is applied to all measured time histories prior to
the FFT calculation.
For measurements, a Dytran 5801A4 impact
hammer and three Dytran 3100B accelerometers with a
normal sensitivity of 100 mV/g are used. The sensors
measure the vertical accelerations at the twelve nodes as
indicated in Figure 2. Impulse excitation is also applied
to every measurement point by an impact hammer.
Note that since the SA-390 data acquisition system has
only four channels and there are three accelerometers,
the first channel is always connected to the input
hammer and the remaining three channels are connected
to three accelerometers. To complete one set of modal
test, the hammer excitation is repeated twelve times at
one point and the three accelerometers are moved from
one set of three nodes to another set of three nodes after
every three excitations. Each FRF is computed by
averaging the three response time histories, and there
are
twelve
measurement
points
and
three
accelerometers.
The rational polynomial technique [3] is employed
to extract the first six natural frequencies and the
corresponding modal vectors from the recorded FRFs.
Node 8
Node 7

Y
Node 6
X

Node 5
Node 4

Z
Node 3

Node 16

Node 2

Node 15

Node 1

Node 14
Node 13
Node 12
Node 11

Accelerometer and impluse excitation points

Node 10
Node 9

50

7@ 700 mm = 4,900 mm

0m

Figure 2: Configuration of a grid-type bridge model

ANALYTICAL MODELING OF THE

TEST STRUCTURE
A finite element (FE) model for the grid type bridge
structure is constructed using twenty three-dimensional
beam elements. As shown in Figure 2, a girder segment
between two nodes or a cross beam is modeled as a
single element. An elastic modulus of 2.0 10 5 MPa, a
mass density of 7850 kg/m3, and a Poission ratio of 0.2
are specified for the model. Since the accelerometers
measure only the vertical movement of the structure, the

lateral DOFs are not included in the analytical model.

Therefore, each node of an element has two translational
DOFs and three rotational DOFs. The model has a total
of 64 DOFs including four rotational DOFs at the
boundary. Both ends of the beam are modeled as
simple pinned connections.
A pinned connection is modeled by a ball bearing
with a 35 mm diameter in the experimental setup.
Based on a preliminary vibration test, the boundary
conditions appear to be less accurately modeled. The
boundary conditions are then modified by introducing
rotational springs at the rotational DOFs of the
boundaries. Furthermore, additional springs are added
to the rotational DOFs at both ends of the cross beams
to simulate the bolted connection between the girders
and the cross beams. After these modifications, the
relative errors of the first six natural frequencies
between the analytical model and the test structure fall
within 4%.
Table 1: Comparison of the analytical and experimental
natural frequencies
Frequency (Hz)
Relative
error* (%)
Analytical
Experimental
1st Bending
5.4488
5.5780
2.32
1st Torsion
10.1494
10.0610
0.88
2nd Bending
19.1841
18.6441
2.90
2nd Torsion
30.6216
29.4503
3.98
3rd Bending
41.6086
42.6292
2.39
3rd Torsion
54.9704
57.2512
3.98
* error = 100 : and are analytical and
experimental frequencies, respectively.
Mode

Table 1 compares the natural frequencies of the

analytical model computed after model updating with
the experimental frequencies. Here, the experimental
frequency is a mean value of the twelve frequencies
estimated with an impact load applied at nodes 2-7, and
10-15, respectively. The damping ratios for the first six
modes are also estimated: 0.8725%, 0.9780%, 0.3420%,
0.6151%, 0.4231%, and 0.4981%.
Figure 3 displays the analytical and experimental
modal vectors of the first six modes. All figures are
plotted in the global X-Y plane of Figure 2, viewing the
structure from the side. As for the scaling of the modal
or Ritz vectors, a mass-normalization is conducted.
However, since the DOFs of the analytical model do not
coincide with the DOFs of the experimental modal
vectors, a reduced analytical mass matrix is first
computed using the Guyan (static) condensation
procedure. Both the analytical and experimental vectors
are normalized with respect to the reduced mass matrix.
Errors arisen from the model reduction are found to be
minimum since the inertial forces associated with the
omitted rotational and axial DOFs (slave DOFs) are
negligible in this example.

LA-UR- 99-4881Proceedings of Asia-Pacific Vibration Conference, Orchard Hotel, Singapore, December 13-15, 1999.

EXPERIMENTAL VERIFICATION

Analytical
Experimental

1
9

8
2
10

16

7
3

6
4

12

13

12

13

11

15

14

(a) Ritz vector 1

11

14

10

15

16

7
3

6
4

(b) Ritz vector 2

6
7
14
5

15

13

1
9

8
16

4
2

12
3

10
11

(c) Ritz vector 3

11

6
7

10
12

16
4

13

15
3

(d) Ritz vector 4

4

12

13
8

11

14

10

16

7
15

(e) Ritz vector 5

10

15
4

11

14

16

6
12

13

(f) Ritz vector 6

Figure 3: Comparison of analytical and experimental

modal vectors

MAC(i , j ) =

(r

T
i

(r

T
i

M r j

)(

M ri r Tj M r j

(11)

where ri and r j are the analytical and experimental Ritz

14

The new extraction procedure for Ritz vectors is

demonstrated using the vibration data obtained from the
test structure. Special focus is placed on the influence
of the residual flexibility on the accuracy of extracted
Ritz vectors. Figure 4 compares the first six Ritz
vectors estimated by the flexibility-based method with
the corresponding Ritz vectors computed from the FE
model. Imaginary point loads are simultaneously
applied to nodes 2 and 13 (upward point load at node 2
and downward point load at node 13), and the
corresponding Ritz vectors are generated. A complete
flexibility matrix, including the modal and residual
flexibility matrices estimated from the measured FRFs,
is employed for the extraction of Ritz vectors. Figure 4
shows a good agreement between the analytical and
experimental Ritz vectors. The first Ritz vector is
equivalent to a static deflection pattern observed when
the unit loads are applied at nodes 2 and 13. Note that
the discrepancy between the analytical and experimental
Ritz vectors increases for higher Ritz vectors.
To investigate influence of the residual flexibility,
Figure 5 shows the experimental Ritz vectors obtained
by only using the measured modal flexibility.
Comparison of Figure 4 and Figure 5 reveals that the
inclusion of the residual term significantly improves the
accuracy of higher Ritz vectors. More quantitative
analysis is presented in Table 2, where the Modal
Assurance Criterion (MAC) value is defined as follows:

vectors, respectively.
Although not presented here, the state-space-based
methods basically produces the same results as the
proposed method only using the modal flexibility.
Furthermore, as mentioned earlier, only the flexibilitybased method allows generating Ritz vectors from any
fictitious load patterns as well as the actual load pattern
applied during the tests.
In Table 3, the experimental Ritz vectors
corresponding to a point load at nodes 2 - 7 and 10 - 15
are extracted, and the comparison of MAC values with
the corresponding analytical Ritz vectors are presented.
Again, the measured complete flexibility is used for all
cases in the table. For brevity, only the diagonal
components of the MAC values are shown in the table.
The result indicates that the Ritz vectors can be
successfully generated from all the load patterns
imposed.

LA-UR- 99-4881Proceedings of Asia-Pacific Vibration Conference, Orchard Hotel, Singapore, December 13-15, 1999.

Analytical
Experimental

Analytical
Experimental
2

16

10

15
11

14
12

14
12

13

13

13

(a) Ritz vector 1

(a) Ritz vector 1

13

14

12

14

12

15

11

16

10

15
16

11

8
7

7
2

6
3

6
3

(b) Ritz vector 2

2

10

10

12

15

13

5
4

(b) Ritz vector 2

11

14

16

11

12
15
13

14

(c) Ritz vector 3

(c) Ritz vector 3

13

13
12
5
3

10

16

10

11

12
8

5
3

16
6

15

(d) Ritz vector 4

14

14
2

11

7
5

16

11

10

15

10

16

11

15

(d) Ritz vector 4

12

12
5

11

2
7
1

16

10

16
10

15

15
4

4
13

14

14
13

(e) Ritz vector 5

(e) Ritz vector 5

10

10

7
6

11

3
4
12
2

16

13

16

4
12
2

Figure 4: Comparison of analytical and experimental

Ritz vectors (using a complete flexibility, G = G m + G r )

15

14

(f) Ritz vector 6

15

11

13

9
1

14

(f) Ritz vector 6

Figure 5: Comparison of analytical and experimental

Ritz vectors (using only a modal flexibility, G = G m )

LA-UR- 99-4881Proceedings of Asia-Pacific Vibration Conference, Orchard Hotel, Singapore, December 13-15, 1999.

Next, the Ritz vectors are estimated using only the

modal flexibility, and the MAC values are presented in
Table 4. For all cases, the MAC values between the
analytical and experimental Ritz vectors fall below 0.8
after the third Ritz vectors. The inclusion of the
residual flexibility matrix into the extraction procedure
allows the contribution of higher residual modes into
experimentally estimated Ritz vectors significantly
improving the accuracy of the estimated Ritz vectors.

Table 2: MAC values between analytical and

experimental Ritz vectors

i for MAC(i , i )
1
2
3
4
5
6
G m + G r 0.9947 0.9930 0.9971 0.9956 0.9754 0.9720
Gm
0.9946 0.9930 0.9876 0.9099 0.3932 0.0211
G

Table 3: MAC values for different load patterns (using

the complete flexibility, G m + G r )
Node

2
3
4
5
6
7
10
11
12
13
14
15

1
0.9944
0.9983
0.9994
0.9994
0.9989
0.9954
0.9937
0.9984
0.9995
0.9996
0.9993
0.9963

2
0.9929
0.9964
0.9974
0.9966
0.9965
0.9952
0.9925
0.9956
0.9965
0.9970
0.9970
0.9953

i for MAC(i , i )
3
4
0.9980 0.9931
0.9955 0.9945
0.9966 0.9948
0.9903 0.9907
0.9967 0.9976
0.9917 0.9800
0.9950 0.9785
0.9978 0.9978
0.9932 0.9929
0.9968 0.9970
0.9951 0.9958
0.9956 0.9935

5
0.9888
0.9485
0.9857
0.9761
0.9755
0.9717
0.9697
0.9790
0.9825
0.9885
0.9542
0.9866

6
0.9805
0.9657
0.9900
0.9656
0.9824
0.9731
0.9737
0.9849
0.9764
0.9877
0.9635
0.9819

SUMMARY AND DISCUSSIONS

In this paper, a new procedure has been proposed to
extract load-dependent Ritz vectors using a statically
complete flexibility. First, modal flexibility and residual
flexibility matrices are estimated from measured FRFs.
Then, Ritz vectors are recursively generated using the
measured flexibility matrix.
The procedure is successfully demonstrated using an
experiment of a grid-type bridge structure. Particularly,
the influence of the residual flexibility on the extracted
Ritz vectors is investigated. The proposed method has
at least two advantages over the state-space method: (1)
the inclusion of the residual flexibility significantly
improves the accuracy of the estimated Ritz vectors, and
(2) the proposed method is able to generate Ritz vectors
from any arbitrary load patterns. In addition, the
proposed extraction procedure is computationally
efficient requiring only the identification of classical
modal vectors and natural frequencies. The increased
amount of information and better sensitivity to
structural parameter changes, which are achievable by
multiple loading and careful selection of load patterns,
could improve the results of damage detection, model
refinement, or component mode synthesis.

ACKNOWLEDGMENTS
This work is supported by National Science
Foundation under Grant No. CMS-95261-2.
The
authors wish to recognize the Hyundai Institute of
Construction Technology (HICT), Korea for providing
the laboratories and their personnel. The authors also
would like to thank Mr. Jeong Hwan Jang of Seoul
National University for his contributions to the
experimental portion of this paper.

REFERENCES
Table 4: MAC values for different load patterns (using
the modal flexibility, G m )
Node

2
3
4
5
6
7
10
11
12
13
14
15

1
0.9944
0.9982
0.9993
0.9994
0.9989
0.9952
0.9938
0.9984
0.9995
0.9996
0.9992
0.9962

2
0.9920
0.9958
0.9967
0.9958
0.9962
0.9933
0.9910
0.9952
0.9958
0.9964
0.9965
0.9938

i for MAC(i , i )
3
4
0.9913 0.9429
0.9887 0.9502
0.9678 0.9262
0.9472 0.8818
0.9909 0.9627
0.9779 0.9019
0.9827 0.9096
0.9925 0.9621
0.9569 0.8997
0.9630 0.9139
0.9895 0.9566
0.9847 0.9306

5
0.7541
0.4777
0.7802
0.6782
0.5722
0.6498
0.6753
0.5631
0.7208
0.7431
0.5213
0.7229

6
0.3560
0.0848
0.3623
0.2588
0.1649
0.2654
0.2651
0.1578
0.2973
0.3377
0.1140
0.3424

[1] T. T. Cao and D. C. Zimmerman, A procedure to

extract Ritz vectors from dynamic testing data,
Proceedings of the 15th International Modal
Analysis Conference, pp1036-1042, 1997.
[2] S. W. Doebling, L. D. Peterson, and K. F. Alvin,
"Estimation of Reciprocal Residual Flexibility
from Experimental Modal Data," AIAA Journal,
Vol. 34, No. 8, pp. 1678-1685, 1996.
[3] M. H. Richardson and D. L. Formenti, Parameter
estimation from frequency response measurements
using rational fraction polynomials, Proceedings
of the 1st International Modal Analysis
Conference, pp167-182, 1982.