ARTICLE IN PRESS

Mechanical Systems
and
Signal Processing
Mechanical Systems and Signal Processing 21 (2007) 1143–1152
www.elsevier.com/locate/jnlabr/ymssp

Prediction of strain responses from the measurements

of displacement responses
Gun-Myung Lee
School of Mechanical and Aerospace Engineering, Gyeongsang National University, Jinju, Gyeongnam 660-701, Republic of Korea
Received 3 June 2005; received in revised form 13 February 2006; accepted 15 February 2006
Available online 31 March 2006

Abstract

A method to predict the strain responses from the measurements of displacement responses is considered. The method
uses a transformation matrix which is composed of a displacement modal matrix and a strain modal matrix. The method
can predict strains at points where displacements are not measured as well as at displacement measuring points. One of the
drawbacks of the strain prediction method is that the displacement responses must be measured at many points on a
structure simultaneously. This difficulty can be overcome by measuring the FRFs between displacements at a reference
point and other point in sequence with a two channel measuring equipment. This procedure is based on the assumption
that the characteristics of excitation applied to the structure do not vary with time.
r 2006 Elsevier Ltd. All rights reserved.

Keywords: Strain response prediction; Displacement; Transformation matrix

1. Introduction

Modern mechanical structures should meet the design requirements of compactness and light weight. Also
those structures should be proven to have sufficient dynamic and fatigue strengths. To assure dynamic and
fatigue strengths it is necessary to measure the dynamic strain distribution on structures. However, strain
measurements with conventional strain gauges are not always possible, and the gauges are not reusable and
cannot be moved from point to point when they have been attached to structures.
Some methods have been developed to predict the dynamic strain distribution from displacement
measurements on structures. Koss and Karczub [1] developed a procedure to measure dynamic strain in a
beam by the use of two accelerometers placed on the beam and frequency response functions (FRFs) between
them which relate measured accelerations at different positions to the fixed reference acceleration position.
Camden and Simmons [2] established the linear (one-dimensional) displacement–strain relationship for a
specific structure and application. Karczub and Norton [3] applied finite differencing methods, developed for

Tel.: 82 55 751 5313; fax: 82 55 757 5622.

E-mail address: gmlee@gsnu.ac.kr.

0888-3270/$ - see front matter r 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ymssp.2006.02.008
 ARTICLE IN PRESS
1144 G.-M. Lee / Mechanical Systems and Signal Processing 21 (2007) 1143–1152

structural intensity measurements, to the measurement of dynamic strain in randomly vibrating structures.
These measurement methods provide predictions of dynamic bending strain from discrete measurements of
displacement, velocity or acceleration. Cuschieri [4] proposed an experimental technique to measure the
structural intensity on an aircraft fuselage. The technique involves estimation of the curvature by means of
finite difference schemes and discrete measurement points. Okubo and Yamaguchi [5] predicted the dynamic
strain distribution under operating condition using the transformation matrix from displacements to strains.
In the method proposed by Sehlstedt [6], results from hybrid modal analysis are transformed from the
displacement space to the strain space by use of finite difference schemes. Inversely, techniques to determine
displacements at points of a vibrating body from measured strains have been developed [7,8]. These techniques
can be applied for active control of smart composite structures with embedded fibre optic strain sensors and
for end-point control of flexible manipulators.
In this paper, the method to predict the strain distribution using the transformation matrix is considered.
Investigated are characteristics of the method such as effects of noise and modal truncation, and the possibility
of strain prediction at points other than displacement measuring points. Also limitations of the method are
investigated and a procedure to overcome the limitations is proposed.

2. Theory

The displacements at points on a vibrating structure, fuðtÞg, can be expressed as a linear combination of
vibrational modes as follows:
fuðtÞg ¼ ½F
fqðtÞg, (1)

where ½F
is the modal matrix whose columns represent the mode shapes of vibrational modes and fqðtÞg
modal coordinates. The strains at points on a structure, f
ðtÞg, can be expressed by the spatial differentiation of
the displacement distribution as follows:
f
ðtÞg ¼ Dð½F
ÞfqðtÞg
¼ ½C
fqðtÞg, ð2Þ

where D represents a linear spatial differential operator and ½C
is the strain modal matrix whose columns
represent the strain mode shapes. From Eq. (1), fqðtÞg is obtained as
fqðtÞg ¼ ½F
1 fuðtÞg. (3)

Substituting the above equation into Eq. (2)

f
ðtÞg ¼ ½C
½F
1 fuðtÞg. (4)

Using the notation,

½T
¼ ½C
½F
1 , (5)

we obtain the following:

f
ðtÞg ¼ ½T
fuðtÞg. (6)

In the above equation, ½T
is the transformation matrix which converts displacements to strains. The equation
means that the strain distribution can be obtained from the displacement measurements if the transformation
matrix is known. Eq. (6) can be written in detail as follows:
8 9 2 38 u ðtÞ 9
>
>
1 ðtÞ >
> T 11 T 12    T 1n > > 1 >
>
>
> > > >
>
2 ðtÞ >
< > 6 T 21 T 22    T 2n 7>
= > u2 ðtÞ >
< >
=
6 7
.. ¼6
6 .. . . . 7 . . (7)
>
> . > > 4 . .. .. .. 7
5>> .. > >
>
> >
> >
> >
>
:
n ðtÞ >
> ; > >
T n1 T n2    T nn : un ðtÞ ;
 ARTICLE IN PRESS
G.-M. Lee / Mechanical Systems and Signal Processing 21 (2007) 1143–1152 1145

When the number of measuring points on a structure is n, and the number of the considered modes is m,
the size of matrix ½F
becomes n  m. If nam, the pseudo-inverse matrix, ½F
 , can be used instead of ½F
1 in
Eq. (5).
½F
 ¼ ð½f
T ½F
Þ1 ½F
T . (8)

As Eq. (5) shows, the transformation matrix is composed of the displacement modal matrix and the
strain modal matrix of a structure. The modal matrices can be obtained analytically, or numerically by using a
finite element method, or experimentally. To obtain the modal matrices experimentally, displacement
frequency response functions (DFRFs) between applied forces and displacement responses and strain
frequency response functions (SFRFs) between applied forces and strain responses should be measured
through modal testing. Then the displacement modal matrix is obtained by the modal analysis of the DFRFs,
and the strain modal matrix by the modal analysis of the SFRFs. A DFRF is expressed for viscous damping as
follows:
X r fj r fk
H jk ðoÞ ¼ . (9)
r
o2r  o2 þ i2zr or o

Similarly, a SFRF is expressed as follows [9,10]:

X r cj r fk
Sjk ðoÞ ¼ , (10)
r
o2r  o2 þ i2zr or o

where r cj and r fk represent the mode shape component of strain mode r at point j and the mode shape
component of displacement mode r at point k, respectively. Since SFRFs take similar forms as DFRFs, modal
analysis routines which extract mode shape components from DFRFs can be also used to extract strain mode
shape components from SFRFs.

3. Application

3.1. Strain prediction at displacement measuring points

The strain prediction method using the transformation matrix was applied to a cantilever beam. The size of
the beam was 300 mm  30 mm  2 mm and the density was 7857 kg/m3 . As Fig. 1 shows, there are 5
measuring points with equal spacing on the beam.
To obtain the transformation matrix, the displacement modal matrix and the strain modal matrix of the
beam were obtained analytically. The DFRF of the beam can be obtained from the solution of the equation of
motion when subjected to harmonic excitation and is given as follows [11]:
1 X Y r ðxj ÞY r ðxk Þ
H jk ðoÞ ¼ , (11)
rAL r o2r  o2 þ i2zr or o

1 2 3 4 5

Fig. 1. A cantilever beam with five measuring points.

 ARTICLE IN PRESS
1146 G.-M. Lee / Mechanical Systems and Signal Processing 21 (2007) 1143–1152

where r; A and L represent the density, the cross-sectional area and the length of the beam, respectively, and
Y r ðxÞ is given as follows:
sin br L  sinh br L
Y r ðxÞ ¼ cos br x  cosh br x þ ðsin br x  sinh br xÞ, (12)
cos br L þ cosh br L
where br is related to the eigenvalue of mode r. Comparing Eq. (11) with the general expression of a DFRF,
Eq. (9), it is found that
1
r fj ¼ pffiffiffiffiffiffiffiffiffiffi Y r ðxj Þ. (13)
rAL
Since the strain on the surface of a thin beam with thickness 2h is obtained as
q2 uðxÞ

ðxÞ ¼ h , (14)
qx2
the strain mode shape component is expressed as follows:
h
r cj ¼ pffiffiffiffiffiffiffiffiffiffi Y€ r ðxj Þ. (15)
rAL
From Eq. (5) with mode shape components expressed in Eqs. (13) and (15), the transformation matrix is
obtained.
Considering five lower-order modes the transformation matrix calculated from Eq. (5) becomes
2 3
1:0568 0:6297 0:1861 0:0786 0:0201
6 0:5916 0:9376 0:5735 0:1349 0:0292 7
6 7
6 7
½T
¼ 6
6 0:1559 0:5716 0:9282 0:5366 0:0737 7.
7 (16)
6 7
4 0:0790 0:1599 0:5753 0:8519 0:3870 5
8:7205e  9 1:2227e  8 9:1773e  9 8:3791e  9 6:8744e  9
Examining the above transformation matrix, it is found that the elements of the last row are very small
compared to other elements. As a result the strain at point 5 becomes very small compared to those at other
points and this agrees with the fact that the strain at the end of a cantilever beam is zero.
The strain prediction method was verified using simulation data for the above cantilever beam subject to a
normally distributed random force. The random force had frequency components between 0 and 500 Hz, and
the Nyquist frequency of the sampled force signal was 2000 Hz. The force was applied at point 2 on the beam
and the strain response was predicted at point 4. Each of five modes had a damping ratio of 1%. As explained
in Eq. (7), the displacement responses at all measuring points are needed to predict the strain response at one
point. The displacement responses were calculated through modal analysis. That is, the response of each mode
due to the applied force was calculated by solving a second-order ordinary differential equation and then the
displacement responses were obtained by the superposition of the considered modes as follows:
X
m
uðx; tÞ ¼ Y r ðxÞqr ðtÞ. (17)
r¼1

The strain response at point 4 was calculated using Eqs. (14) and (17), and this strain response agreed exactly
with the response predicted by the explained method as Fig. 2 shows.

3.2. Effects of noise and modal truncation

In all practical cases measured data are contaminated by noise to some extent. To simulate the situation
normally distributed random noise was added to calculated displacements at each point. The noise level was
defined as the ratio of the rms (root-mean-square) value of added noise to the rms value of displacement.
When 1% noise was added to displacements, the effect of noise was apparent as Fig. 3 shows. As noise level
 ARTICLE IN PRESS
G.-M. Lee / Mechanical Systems and Signal Processing 21 (2007) 1143–1152 1147

x 10-5
2.5
calculated
2 predicted

1.5

0.5
strain

-0.5

-1

-1.5

-2
0.25 0.255 0.26 0.265 0.27 0.275 0.28 0.285 0.29 0.295 0.3
time (s)

Fig. 2. Comparison of the calculated and predicted strains when the displacements are ‘‘measured’’ simultaneously.

x 10-5
3
calculated
predicted
2

1
strain

-1

-2

-3
0.25 0.255 0.26 0.265 0.27 0.275 0.28 0.285 0.29 0.295 0.3
time (s)

Fig. 3. Comparison of the calculated and predicted strains when 1% noise is added to the displacements.

was increased to 2%, the accuracy of the predicted strain degraded rapidly as Fig. 4 shows. These results imply
that the above strain prediction method is very sensitive to added noise.
The above method predicts strain responses from measured displacements only using the known
transformation matrix, and we do not need to consider the effects of noise added to excitation forces on the
prediction accuracy. However, if the transformation matrix is obtained from measured FRFs between
 ARTICLE IN PRESS
1148 G.-M. Lee / Mechanical Systems and Signal Processing 21 (2007) 1143–1152

x 10-4
4
calculated
predicted
3

1
strain

-1

-2

-3
0.25 0.255 0.26 0.265 0.27 0.275 0.28 0.285 0.29 0.295 0.3
time (s)

Fig. 4. Comparison of the calculated and predicted strains when 2% noise is added to the displacements.

x 10-5
2.5
calculated
2 predicted

1.5

0.5
strain

-0.5

-1

-1.5

-2
0.25 0.255 0.26 0.265 0.27 0.275 0.28 0.285 0.29 0.295 0.3
time (s)

Fig. 5. Comparison of the calculated and predicted strains when the 5th mode is truncated.

excitation forces and responses, noise added to excitation forces will affect FRFs, and, in turn, prediction
accuracy. Since a lot of research results on the effects of noise added to excitation forces have been reported
[12], the noise problem is not considered in this paper.
Next, the effect of modal truncation on the strain prediction accuracy was considered. The considered beam
was assumed to have five modes. The natural frequencies of these modes were 18.1, 113.5, 317.8, 622.8, and
1029.5 Hz, respectively. To study the effect of modal truncation, only four lower modes with the fifth mode
 ARTICLE IN PRESS
G.-M. Lee / Mechanical Systems and Signal Processing 21 (2007) 1143–1152 1149

omitted were included in calculating the transformation matrix. As Fig. 5 shows, the difference between the
predicted and calculated strains is negligible. Noting that the omitted fifth mode is far beyond the excitation
frequency range, 0–500 Hz, it may be concluded that the truncation of modes far beyond the excitation range
does not affect the accuracy of the predicted strain.

3.3. Strain prediction at points other than displacement measuring points

In the above examples, the point where the strain response was predicted coincided with a displacement
measuring point. The strain prediction method can be applied similarly to predict the strain responses at
points where the displacement responses are not measured. Assume that the displacement responses are
measured at n1 points, the strain responses are predicted at n2 points, and the number of included modes is m.
Then the previous equations can be rewritten with the appropriate dimensions as follows:
fuðtÞgn1 1 ¼ ½F
n1 m fqðtÞgm1 , (18)

f
ðtÞgn2 1 ¼ ½C
n2 m fqðtÞgm1 , (19)

fqðtÞgm1 ¼ ½F

mn1 fuðtÞgn1 1 , (20)

f
ðtÞgn2 1 ¼ ½C
n2 m ½F


mn1 fuðtÞgn1 1 ¼ ½T
n2 n1 fuðtÞgn1 1 . (21)
The above equations imply that the strains can be predicted at points where the displacements are not
measured, and that in this case the transformation matrix is composed of displacement mode shape
components at displacement measuring points and strain mode shape components at strain prediction points.
To verify the strain prediction method in this case, the same cantilever beam was considered and more points
were added on the beam as in Fig. 6. The displacement responses at points 2, 4, 6, 8, and 10 were calculated for
the same random force applying at point 4. The strain response was predicted at point 7, and this predicted
response agreed exactly with the calculated response. The figure comparing the two responses is omitted for
brevity. The result shows that the strain responses can be predicted even at points where the displacement
responses are not measured.

3.4. Strain prediction using a two-channel measuring equipment

One of the drawbacks of the above strain prediction method is that the displacement responses must be
measured at many points on a structure simultaneously. Therefore, if the number of measuring points is
limited by the measuring equipments available, the method cannot be applied to the cases where many
measuring points are needed. To overcome this difficulty the following procedure is proposed. From Eq. (7)
the strain response at point i can be written in detail as follows:

i ðtÞ ¼ T i1 u1 ðtÞ þ T i2 u2 ðtÞ þ    þ T in un ðtÞ. (22)
Taking the Fourier transform of the above equation, we obtain
E i ðoÞ ¼ T i1 U 1 ðoÞ þ T i2 U 2 ðoÞ þ    þ T in U n ðoÞ. (23)

1 2 3 4 5 6 7 8 9 10

Fig. 6. A cantilever beam with 10 measuring points.

 ARTICLE IN PRESS
1150 G.-M. Lee / Mechanical Systems and Signal Processing 21 (2007) 1143–1152

x 10-5
2
calculated
1.5 predicted

0.5

0
strain

-0.5

-1

-1.5

-2

-2.5
0.25 0.255 0.26 0.265 0.27 0.275 0.28 0.285 0.29 0.295 0.3
time (s)

Fig. 7. Comparison of the calculated and predicted strains when pairs of displacements are ‘‘measured’’ in sequence.

Dividing both sides of the above equation by one of U j ðoÞ’s, for example, U 1 ðoÞ, we obtain

E i ðoÞ U 2 ðoÞ U n ðoÞ

¼ T i1 þ T i2 þ    þ T in ¼ HðoÞ. (24)
U 1 ðoÞ U 1 ðoÞ U 1 ðOÞ

That is,

E i ðoÞ ¼ HðoÞ  U 1 ðoÞ. (25)

In measuring HðoÞ, the FRF between the displacement response at point 1 and the strain response at point i, it
is necessary to measure FRFs between the displacement responses at point 1 and all the other points
simultaneously. However, if we assume that the characteristics of excitation (for example, excitation point)
applied to the structure under consideration do not vary, these FRFs would not vary with time. Therefore, it is
possible to measure each FRF in sequence using a two-channel measuring equipment such as a conventional
spectrum analyser. These FRFs can be measured similarly as in modal testing where two measured data are
force and displacement (or velocity or acceleration). Thus, this procedure eliminates the necessity of a
measuring equipment with many channels. Once E i ðoÞ is obtained by Eq. (25), the strain response at the point
can be obtained by its inverse Fourier transform. Any point on the structure can be selected as a reference
point, which is point 1 in Eq. (24).
To evaluate the above procedure the same system in Fig. 1 was considered. A random force was applied at
point 2 and the strain response was predicted at point 4. In Eq. (24) each U j ð$Þ=U 1 ð$Þ was calculated
repeatedly from separate simulations and averaged. The predicted strain at point 4 showed strong high-
frequency components and consequently there were much differences between the calculated and predicted
responses as Fig. 7 shows. The reason can be attributed to the fact that the division, U j ð$Þ=U 1 ð$Þ can cause
large errors at frequencies where the magnitude of U 1 ð$Þ is small. To solve this problem the following
procedure was followed. Since it is usual except special cases such as nodes that U j ð$Þ is also small at
frequencies where U 1 ð$Þ is small, the ratio U j ð$Þ=U 1 ð$Þ was neglected at frequencies where U j ð$Þ is smaller
than a pre-determined value, for example, a times its largest value. This procedure results in neglect of the term
including U j ð$Þ in Eq. (23) when U j ð$Þ is small, which is rational. Following this procedure, the accuracy of
the predicted strain was much improved as Fig. 8 shows. In this case averaging was also performed and the
 ARTICLE IN PRESS
G.-M. Lee / Mechanical Systems and Signal Processing 21 (2007) 1143–1152 1151

x 10-5
1.5
calculated
predicted
1

0.5

0
strain

-0.5

-1

-1.5

-2
0.25 0.255 0.26 0.265 0.27 0.275 0.28 0.285 0.29 0.295 0.3
time (s)

Fig. 8. Increased accuracy of the predicted strain by neglecting the terms with small U j ð$Þ ’s.

number of averaging was 100. The used value of a was 0.002. The accuracy of the predicted strain did not
change much with the value of a.

4. Conclusion

A method to predict the strain responses from the measurements of displacement responses is considered in
this paper. The method uses a transformation matrix which is composed of a displacement modal matrix and a
strain modal matrix. The method can predict strains at points where displacements are not measured as well as
at displacement measuring points. Using the prediction method accurate strain responses were predicted for a
cantilever beam subject to a random force. It was found that the method is very sensitive to noise added to
displacements. If the transformation matrix is obtained from measured FRFs between excitation forces and
responses, noise added to excitation forces will affect FRFs, and, in turn, prediction accuracy. However, since
a lot of research results on the effects of noise added to excitation forces have been reported, the noise problem
is not considered in this paper. It was also found that the truncation of modes far beyond the excitation range
does not affect the accuracy of the predicted strain.
One of the drawbacks of the strain prediction method is that the displacement responses must be measured
at many points on a structure simultaneously. This difficulty can be overcome by measuring the FRFs between
displacements at a reference point and other point in sequence with a two channel measuring equipment.
This procedure is based on the assumption that the characteristics of excitation applied to the structure
under consideration do not vary with time. The procedure was verified using simulation data for the same
cantilever beam.

Acknowledgements

This work was supported by Grant No. RTI04-01-03 from the Regional Technology Innovation Program of
the Ministry of Commerce, Industry and Energy (MOCIE) and by the Engineering Research Institute in
Gyeongsang National University.
 ARTICLE IN PRESS
1152 G.-M. Lee / Mechanical Systems and Signal Processing 21 (2007) 1143–1152

References

[1] L.L. Koss, D. Karczub, Euler beam bending wave solution predictions of dynamic strain using frequency response functions, Journal
of Sound and Vibration 184 (1995) 229–244.
[2] M. Camden, L. Simmons, Using a laser vibrometer for monitoring dynamic strain, modal analysis and calculating damping, in:
Proceedings of SPIE Conference on Laser Interferometry IX: Applications, 1998, pp. 311–318.
[3] D.G. Karczub, M.P. Norton, Finite differencing methods for the measurement of dynamic bending strain, Journal of Sound and
Vibration 226 (1999) 675–700.
[4] J.M. Cuschieri, Experimental measurement of structural intensity on an aircraft fuselage, Noise Control Engineering Journal 37
(1991) 97–107.
[5] N. Okubo, K. Yamaguchi, Prediction of dynamic strain distribution under operating condition by use of modal analysis, in:
Proceedings of the 13th IMAC, 1995, pp. 91–96.
[6] N. Sehlstedt, Calculating the dynamic strain tensor field using modal analysis and numerical differentiation, Journal of Sound and
Vibration 244 (2001) 407–430.
[7] A.C. Pisoni, C. Santolini, D.E. Hauf, S. Dubowsky, Displacements in a vibrating body by strain gauge measurements, in: Proceedings
of the 13th IMAC, 1995, pp. 119–125.
[8] G.C. Foss, E.D. Haugse, Using modal test results to develop strain to displacement transformations, in: Proceedings of the 13th
IMAC, 1995, pp. 112–118.
[9] O. Bernasconi, D.J. Ewins, Modal strain/stress fields, International Journal of Analytical and Experimental Modal Analysis 4 (1989)
68–76.
[10] O. Bernasconi, D.J. Ewins, Application of strain modal testing to real structures, in: Proceedings of the Seventh IMAC, 1989,
pp. 1453–1464.
[11] L. Meirovitch, Elements of Vibration Analysis, McGraw-Hill Book, New York, 1986.
[12] D.J. Ewins, Modal Testing: Theory and Practice, Research Studies Press, Letchwor, 1984.