Sei sulla pagina 1di 17

1

Published in Mechanical Systems and Signal Processing 23 (2009) 1112-1122



Designing structures for dynamical properties via natural frequencies separation.
Application to tensegrity structures design

Cornel Sultan
Aerospace and Ocean Engineering, Randolph Hall 215,
Virginia Polytechnic Institute and State University, Blacksburg VA 24061, USA
csultan@vt.edu, Phone: 1-540-231-0047, Fax: 1-540-231-9632
Abstract
The design of structures for dynamic properties is addressed by placing conditions on the separation between natural
frequencies. Additional constraints, like lower and upper bounds on the natural frequencies, are also included. A fast
numerical algorithm that exploits the mathematical structure of the resulting problem is developed. Examples of the
algorithms application to tensegrity structures design are presented and the connection between natural frequencies
separation and proportional damping approximation is analyzed.
Keywords: dynamic design; natural frequencies separation; proportional damping approximation; tensegrity
structures

1. Introduction
Structures design using numerical methods has been initiated by Dorn [1] (see [2,3] for
detailed reviews) and most of the research has been focused on static requirements satisfaction
(e.g. constraints on displacement, stress, strain in equilibrium conditions). A limited number of
articles deal with dynamic requirements (see [4-7] and the references therein). There are two
major reasons for the limited interest in designing structures for dynamic properties. Firstly,
structures were traditionally designed for static operating conditions (arches, bridges, domes).
Secondly, because many classical structures are heavily damped, the dynamic transitory regime
decays rapidly and it is not considered important. However, with the advent of new technologies
in the area of controllable structures like morphing structures [8], adaptive buildings [9], flexible
manipulators [10], these facts will no longer hold true. Firstly, controllable structures will
operate in conditions for which the dynamic regime will play an important role in their design.
Secondly, for these applications, lightly damped structures will be preferred in order to dissipate
less energy and make their shape control cost effective and efficient.
This article approaches the design of structures from the dynamic perspective. The main
design requirement is represented by the natural frequencies location. Placing restrictions on the
natural frequencies is justified because of their importance in structures dynamics. Firstly, these
frequencies are crucial for dynamic response characteristics like the rise time, peak time, and
settling time [11]. Secondly, in many cases, separation between natural frequencies is crucial in
enabling accurate proportional damping approximation [12]. Thirdly, as remarked in [4], the
amount of degeneration in structures is reduced if constraints on the natural frequencies are
imposed.
A fast algorithm is presented which guarantees prescribed separation between natural
frequencies. The algorithm also considers constraints on the minimum and maximum natural
frequencies values. The algorithm is fast because it exploits the structure of the mass and
stiffness matrices, which are linear in the design parameters, employs active set methods in order


2
to reduce the number of computations, and the gradients used in the solution process are
analytically computed. Examples of the algorithms application to the dynamic design of
tensegrity structures are given. Analysis of the correlation between natural frequencies and
proportional damping approximation confirms the assertion that separation between natural
frequencies might be a misleading criterion for accurate approximation.

2. Dynamics
2.1. Physical and modal systems
The linearized dynamics of many structural systems is described by
0 , 0 , 0 , > > > = + + K C M f Kq q C q M (1)
where M, C, K are the mass, damping, stiffness matrices, respectively, q is the n-dimensional
vector of generalized coordinates, and f is the vector of external loads, respectively. A
transformation from the physical (q) to the modal (q
m
) coordinates is performed using the
modal matrix, U, which is constructed as follows:
0 . , , > = A = A = diag I U U U U M
M
T
M M
T
M M M

0 . , ,
2
1 1
> = O = O = A A

diag I U U U U KU U
T
K K
T
K K M M
T
M M

K M M
U U U
1
A = . (2)
Using
m
Uq q = in modal coordinates (1) becomes
f U q q C q
T
m m m m
= O + +
2
(3)
where CU U C
T
m
= , ) ( diag
2 2
l
e = O , and
l
e are the natural frequencies obtained by solving
0 ) det(
2
= M K
l
e . (4)
Because natural frequencies are crucial in the dynamic response, dynamic design problems
usually include constraints on their values [4]. The first condition imposed here is that the natural
frequencies are separated. There are several reasons for which this is considered an important
requirement as discussed next.

2.2. Natural frequencies separation and proportional damping approximation
System (1) is proportionally damped if
m
C is diagonal. There are big benefits if
m
C is
diagonal. For example, the equations of motion (3) decouple and they can be easily solved. This
is very important, because the number of equations used in structures dynamics, n, is usually
large. Proportional damping models are also desired because they lend themselves easily to
computationally efficient identification, model order reduction, and control design tools [12].
A general form of the damping matrix, which results in a diagonal modal damping matrix, is


3

=
1
0
1
) (
n
i
i
i
K M M a C (5)
where
i
a

are real numbers [13]. Expression (5) is a generalization of the Rayleigh damping
model, in which the damping matrix is a linear combination of the mass and stiffness matrices.
One can easily verify that if U simultaneously diagonalizes M and K it also diagonalizes C of (5).
If
m
C is not diagonal it is desired to determine under what conditions it can be
approximated by a diagonal matrix, i.e. when proportional damping approximation is possible.
The most popular approach to obtain such an approximation is to reduce the modal damping
matrix to a diagonal one by neglecting its off diagonal terms (see [12]), thus writing
). (
m p
C Diag C = (6)
The proportional and non-proportional damping models are then
f U q q C q
T
p p p p
= O + +
2
and (7)
0 ) ( , where , ) (
2
= = = O + + +
n p m n
T
m m n p m
C Diag C C C f U q q C C q , (8)
respectively.
In order to determine when accurate proportional damping approximation is possible,
various non-proportionality indices involving the modal damping matrix, natural frequencies,
and external excitations have been proposed. For example Tong et al. [14] proposed a measure of
damping non-proportionality based only on the damping matrix. Shahruz [15] showed through
an example that this measure is insufficient and indicated that the distribution of the systems
natural frequencies must be taken into account. Gawronski [12] and Gawronski and Sawicki [16]
showed that neglecting the off diagonal terms in the modal damping matrix yields good results in
terms of the relative error, which is negligible if the natural frequencies are sufficiently separated
and the damping is small. Adhikari [17] introduced a non-proportionality index based on the
normal modes of the system and not on the system matrices. He shows that for viscously damped
systems with small damping this index is inversely proportional to the separation between
natural frequencies, thus clustered natural frequencies should be avoided because they lead to
large values of the index. In [18] a measure of non-proportional damping which only depends on
the matrix of complex eigenvectors, and it is independent of the natural frequencies, has been
proposed. Over the years it has become generally accepted that sufficient separation of the
natural frequencies is crucial for accurate proportional damping approximation. However, when
the structure is subjected to harmonic excitation, this condition might not be sufficient, as
discussed next.
Park et al. [19] showed that if the input, f, is harmonic, the error between the response of
the system and the response of its proportional damping approximation depends strongly on the
excitation frequency and may be significant for excitation frequencies which are close to the
natural ones. Shahruz and Packard [20] showed that if the system is lightly damped and the
excitation frequency is close to some of the lightly damped natural frequencies, the error might
be big even if the off-diagonal terms of the modal damping matrix,
m
C , are small. However, if
the corresponding proportionally damped system is reasonably damped and the off-diagonal
terms of
m
C are small, then the error is small [21]. Park et al. [22] gave several examples which


4
show that, in the case of external harmonic input, neither the diagonal dominance of the modal
damping matrix nor the separation between natural frequencies is sufficient for accurate
proportional damping approximation; the location of the excitation frequency with respect to the
natural frequencies is an important factor in the error. Moreover, the approximation error
increases substantially when there are natural frequencies which are clustered and the excitation
frequency is close to these clustered natural frequencies (see also [23]).
It is remarkable that so far the discussion focused on accurate proportional damping
approximation in modal coordinates. Even more remarkable is the fact that accurate
approximation in the modal space is not sufficient for accurate approximation in the physical
space (see [24]). The following lemma connects the errors in modal and physical coordinates.
Lemma: The error in the physical space is smaller than the error in the modal space if
and only if the minimum eigenvalue of the mass matrix, ) (M , is greater than one.
Proof: Let ) (t
m
c and ) (t c denote the error in the modal and physical space, respectively:
) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( t U t t U t t q t q t
m m p m m
c o c c c c s = = (9)
where . is the Euclidean norm and ) (U o the maximum singular value of U. Since the upper
bound in Eq. (9) is tight, ) ( ) ( t t
m
c c s is equivalent to 1 ) ( s U o . From Eq. (2)
K M M
U U U
1
A = and, since
M
U and
K
U are unitary,
2 / 1
) ( ) (

= M U o . Thus 1 ) ( s U o

is
equivalent to 1 ) ( > M .

2.3. Natural frequencies separation and computational efficiency and accuracy
From the computational point of view, it is also advantageous to have separated natural
frequencies. Firstly, the sensitivities of repeated natural frequencies and of the associated
eigenvectors with respect to various parameters are difficult to compute, both analytically and
numerically [25]. Secondly, in the case of repeated natural frequencies numerical computations
might lead to unacceptable accumulation and propagation of numerical errors.
On the other hand, if the natural frequencies are adequately separated, the likelihood of
having repeated eigenvalues in the corresponding first order linear modal system,
, , ,
0
,
0
,
2
f u
q
q
x
U
B
C
I
A Bu Ax x
m
m
T
m
=
(

=
(

=
(

O
= + =

(10)
is reduced. This is important because, if A has repeated eigenvalues it may be defective, hence
not diagonalizable, and the response of Eq. (10) may include secular terms. This is definitely not
desirable. Note that in the proportional damping case, A is diagonalizable and secular terms do
not appear, even if there are repeated natural frequencies (see Appendix A). In general, defective
systems represent exceptions and have been rarely reported in practical structures.


5


3. The natural frequencies allocation problem
3.1. Problem formulation
As indicated before, a crucial dynamic design requirement for structures should be
sufficient separation between the natural frequencies. This requirement can be enforced in two
ways: through equality constraints, when all natural frequencies are prescribed fixed values to
achieve adequate separation, or through inequality constraints, when only lower bounds on the
separations are enforced via inequalities. These two options are discussed next.
The equality constraints category can be embedded in the larger class of ideal
dynamic designs, when the natural frequencies and the corresponding eigenvectors are
specified. Research conducted in this area indicated two major deficiencies. Firstly, solutions to
this problem, usually called the inverse spectral problem, require considerable freedom in the
structure of M, C, K, and, secondly, the solutions are restrictive with respect to the specification
of the modal data [5-7]. For example in [26], the inverse spectral problem was solved only for
lumped conservative systems (i.e. C=0) modeled using tri-diagonal matrices. Likewise, in [5] the
inverse spectral problem was solved when C and K are singular but the solution restricts the
eigenvalues to having complex values and does not even preserve the eigenvectors. The
interested reader may consult [7] for a review of results in the area.
A less constrained approach is to require that only the natural frequencies are allocated to
desired locations. In practice even this approach is not usually possible, because the design space
is limited [4]. In many cases the mass, damping, and stiffness matrices are linear combinations of
inertial, damping, and elastic characteristics of the individual elements (e.g. bars, dampers,
springs), and can be written in terms of free scalar design parameters 0 , 0 , 0 > > >
i i i
k c m

as
. , ,
1
0
1
0
1
0
= = =
+ = + = + =
E
i
i i
D
i
i i
I
i
i i
K k K K C c C C M m M M (11)
Here I, D, and E are the numbers of free design parameters and, in general, are small, because
many characteristics are fixed by other considerations (e.g. specifications on the materials). In
Eq. (11), matrices M
0
, C
0
, K
0
account for the fixed parameters.
In this article, the exact placement requirement is relaxed and prescribed separations,
lj
e ,
between natural frequencies are enforced through inequality constraints:
n l j n l
lj j l
,..., 1 , 1 ,..., 1 , + = = > e e e . (12)
Additional constraints on the natural frequencies are included:
n l
l
,..., 1 ,
max min
= < < e e e (13)
where
min
e and
max
e are such that

>
j l
lj
,
min max
e e e . These constraints can be easily justified
for controllable structures as follows. The structure should be designed to an upper bound on the
maximum natural frequency, because it is desired that the high frequency modes are measurable


6
and the structural sensing devices sampling rate, which must be at least twice the structures
maximum natural frequency, is limited (see [27]). The minimum natural frequency is lower
bounded in order to avoid slow modes. Low values of natural frequencies also correspond to a
soft (not sufficiently stiff) structure, which is not desirable.
To complete the problem formulation, limits on the design parameters are included:

E i k k k I i m m m
i i i i i i
,..., 1 , , ,..., 1 ,
max min max min
= < < = < < . (14)
Thus, the problem of interest is to find
i i
k m , subject to Eqs. (12) - (14).
The major advantage this inequality constraints approach has is that it guarantees
increased flexibility in the design process. The exact placement problem might not have any
solution, whereas the inequality constraints problem might have many solutions. This fact
facilitates the incorporation of the inequality constraints approach in more complex design
problems which include static constraints on stress, strain, displacement and optimization
requirements like the minimization of the structures mass.
The main disadvantage of this approach is that it leads to a nonlinear problem with
inequality constraints for which closed form solutions are not possible. However, an efficient
algorithm has been designed which exploits the mathematical structure of the problem. This
algorithm is described next.

3.2. An iterative solution algorithm
The algorithm proposed herein to solve Eqs. (12-14) is inspired by active set methods.
The idea underlying these methods is to partition inequality constraints into two groups: those
that are to be treated as active and those that are to be treated as inactive. The constraints treated
as inactive are ignored, decreasing the number of computations. This is especially useful for
problems with a large number of constraints, like structural design ones (see [28]).
The algorithm proceeds as follows: at the current iteration step, for a known value of the
vector of design parameters, called x, the natural frequencies are computed by solving Eq. (4)
and ordered:
n
e e e s s s ...
2 1
. Satisfaction of constraints (12-14) is evaluated. The constraints
which are violated are chosen as the active ones and a penalty function, P(x), is built:
), ( ) ( ) ( ) ( x P x P x P x P
b bx s e e
+ + = (15)
where
. ) (
2
1
) (
2
1
) (
, ) (
2
1
) (
2
1
) ( , ) (
2
1
) (
2
max
2
min
2
max
2
min
2
1 1


+ =
+ = =
+ +
l
l l
l
l l bx
l
l
l
l b
l
l l l l s
x x x x x P
x P x P e e e e e e e
e e
(16)
Here P
se
(x) is associated with the separation constraints (12), P
be
(x) with the constraints on the
boundaries of e, Eq. (13), and P
bx
(x) with the constraints on the boundaries of x, Eq. (14). In all
these sums, only the violating pairs appear. For example, in P
se
(x), only the indices l for


7
which
l l l l 1 1 + +
s e e e are considered. This particular choice of a quadratic penalty function is
advantageous because it leads to convexification of the problem, thus facilitating the use of
gradient or Newton based iterative procedures for fast convergence (see [28] for more details).
Next, the penalty function is driven to zero using a gradient method. The advantage of
using gradients is that, for this problem, they can be easily computed. Indeed:
) ( ) ( ) ( ) ( x P x P x P x P
b bx s e e
V + V + V = V
(17)
where

V + V = V
l
l l l
l
l l l b
x P e e e e e e
e
) ( ) ( ) (
max min
(18)

V V = V
+ + +
l
l l l l l l s
x P ) )( ( ) (
1 1 1
e e e e e
e
(19)
. ) ( ) ( ) (
max min
+ = V
l
l l l
l
l l l bx
e x x e x x x P (20)
Here e
l
is a vector with the l component equal to 1 and all the other components equal to 0.
The gradient of
l
e

is
T
w
l l l
l
x x x
(

c
c
c
c
c
c
= V
e e e
e ...
2 1
(21)
where w=I+E is the dimension of x. If
l
e is single,
j
l
x c
ce
can be easily computed using
( )
l
j
l
j
l
l
j
l
j
T
l
j
l
x x x
M
x
K
x
e
e e
e
e
2 /
2
2
2
c
c
=
c
c

(
(

c
c

c
c
=
c
c
(22)
where
l
is the corresponding mass-normalized eigenvector: I M
l
T
l
= . Using Eq. (11) the
following formulas are obtained:

=
=
=
c
c
. ,
,
2
2
j j l j
T
l l
j j l j
T
l
j
l
m x if M
k x if K
x
e

e
(23)
Note that zero natural frequencies cannot appear because of the condition that K > 0 (see Eq.
(1)). In the case of repeated natural frequencies, complex formulas have been derived for the
computation of
j
l
x c
ce
(see [25]), but complicated formulas are not necessary here because the
solution will be such that there are no repeated natural frequencies. Thus, if
l
e is not single, the
corresponding value of x will be randomly perturbed by a small amount.
Next, a line search method is used, in which a change in x is made along 0 = V = P g ,
, g x x o + =
+
(24)


8
and the resulting equation in one variable P(x
+
)=P(o)=0 is linearly approximated and solved
with respect to o. At the next step, the penalty function is updated based on the current violating
constraints. Iterations are performed until convergence is obtained (constraints are not violated),
the number of iterations allowed is exceeded, or the norm of the variation in x between two
consecutive steps is smaller than a prescribed tolerance.
There are two major advantages associated with the proposed algorithm. Firstly, at each
step the number of constraints taken into account is reduced to the critical ones. Secondly, at
each step the gradients are analytically computed using Eqs. (17-23) and not numerically
approximated. The result is a fast algorithm which has proven very effective in applications.
One possible issue with this algorithm is that it might lead to zig-zagg motions due to
changes in the set of violating constraints. However, this set is expected to stabilize in the
neighborhood of the solution, and the zig-zag motions will disappear. Also observe that at the
solution some of the constraints (but not all) might be tight, due to the quadratic penalty function.
Lastly, note that normalization of the quantities involved can be performed, as it is
customary for improved numerical accuracy. In the interest of clarity normalization is ignored in
the presentation of the algorithm. Actually, it can be easily ascertained that the algebraic
operations associated with normalization do not affect the main characteristics of the algorithm
because the gradients are still analytically computed.
The algorithm is described next; | |
T
E I
k k m m x ... ...
1 1
=

is the vector of design parameters.
Description of the Algorithm

1. Initialization: Select the initial values of the design parameters,
0
x x = .
2. Constraints evaluation: Solve 0 ) det(
2
= M K
l
e for
l
e where

=
+ =
I
i
i i
M m M M
1
0
,

=
+ =
E
i
i i
K k K K
1
0
and evaluate Eqs. (12-14). If these constraints are not violated, exit.
3. Calculation of the direction of movement, g: If there are no repeated eigenvalues,
l
e ,
compute the gradient of the penalty function ) (x P g V = using Eqs. (17-23). If there are
repeated eigenvalues or 0 = g slightly perturb x randomly and return to step 2.
4. Prediction: Predict g x x o + =
+
, where
2
) (
g
P
x
= o , with ) (x P given by Eq. (15) where
only the violating constraints are considered.
5. Return: If x g o o < || || , where x o is the minimum allowed variation of x, or the number of
iterations is greater than the maximum allowed, exit, else set
+
= x x and return to step 2.
The process terminates when all natural frequencies are sufficiently separated (step 2), the
variation in x is too small, or the number of iterations allowed is exceeded (step 5).
The algorithm can be easily extended to include an upper bound constraint on the
maximum singular value of the modal matrix,
max
) ( o o s U , since the gradient of this constraint


9
can also be computed analytically. The penalty function just has to be modified accordingly.
Some of the examples shown next will include this constraint.

4. Examples: tensegrity structures design
4.1. Tensegrity structures description
Tensegrity structures are assemblies of soft elements which can carry only tensile
forces (e.g. elastic cables), and hard (e.g. rigid bodies), and which are capable of yielding
equilibrium configurations under no external forces and torques and with all soft members in
tension. These configurations are called prestressable configurations [29]. Tensegrity
structures feasibility in controllable structures applications has been demonstrated [30-33]. In the
following, the previous algorithm will be applied to a tensegrity structure.
Consider a tensegrity structure composed of six bars, labeled A
ij
B
ij
, a top (B
12
B
22
B
32
), a
base (A
11
A
21
A
31
), and 18 tendons (Fig. 1). For mathematical modeling the tendons are
considered massless, viscoelastic Voigt elements, which consist of a linear elastic spring in
parallel with a linear viscous damper. The base is fixed and the top and the bars are rigid. The
bars are axially symmetric. For each bar the rotational degree of freedom around the longitudinal
axis of symmetry is ignored. No external forces act on the structure (see [29] for details).

Fig. 1. Tensegrity structure.
Linearized dynamics models around certain equilibria called symmetrical prestressable
configurations (see [29] for details on these configurations), have been derived. If the bars are
identical and the tendons have the same damping coefficients matrices M, C, K are given by
6 4
1 1 0
1 1
, ,
i i i i
i i
M mM C c C K K k K
= =
= = = +

(25)
where m
1
represents the mass of the top, m
2-4
its principal moments of inertia, m
5,6
the mass and
longitudinal moment of inertia of a bar, k
1-3
the stiffness of three classes of tendons called S
(A
i2
B
j1
), V (A
i1
B
j1
and B
i2
A
j2
), D (A
i1
A
j2
and B
i2
B
j1
), k
4
the pretension coefficient (see [29]


10
for details), and c
1
the damping coefficient for all tendons. These will be the design parameters.
The symmetrical prestressable configuration analyzed here is characterized by
1, 0.67, 0.75, 60 l b H o o

= = = = = (26)
where l is the length of a bar, b the length of the side of the base and top equilateral triangles, H
the height of the structure (all in meters), o the angle made by each bar with the vertical
symmetry axis (OO
t
) and o the angle made by the projection of A
11
B
11
on the horizontal plane
(A
11
A
21
A
31
) with the fixed direction
1

b . The corresponding matrices M


i
, C
1
, K
i
, which depend
only on l, b,o , ando , have been computed using the general formulas presented in [34].

4.2. Natural frequencies separation and proportional damping approximation
Consider the following ad-hoc values for the design parameters (the arbitrary design):
. 1 , 1 , 1 , 5 , 4 , 3 , 1 , 1
1 6 5 4 3 2 1 4 1
= = = = = = = =

c m m m m m m k (27)
All quantities are given in SI units. The natural frequencies distribution, shown in Fig. 2,
indicates regions in which these frequencies are clustered: for 8 pairs of neighboring natural
frequencies the separation is less than 0.02. This is not a good dynamic design. For example if
the responses to initial conditions of the proportionally and non-proportionally damped models -
(7) and (8) with f=0 - are computed, the approximation error is unacceptably large. Fig. 2 shows
the Euclidean norms of the modal, ) (t
m
c , and physical, ) (t c , error, for unity initial conditions:
1
0 0 0 0
= = = =
p p m m
q q q q . Hence proportional damping approximation cannot be used. Redesign
of the structure to achieve separation of the natural frequencies must be pursued.
The algorithm presented before has been applied for various prescribed separations,
lj
e ,
to solve (12)-(14). In some cases an upper bound constraint on the singular value norm of the
modal matrix,
max
) ( o o s U , has also been considered, as indicated next. In all cases only lower
bounds on the design parameters were enforced: l x
l
= , 0
min
. Convergence of the algorithm,
implemented in Matlab, was very fast (miliseconds to seconds on a standard desktop computer).
Fig. 3 shows the results obtained for 2 . 0 =
lj
e , 10 , 8 . 0
max min
= = e e (rad/s), and a hard
constraint on the maximum singular value of the modal matrix: 15 . 0 ) ( s U o . Analysis of
responses to initial conditions confirms that the proportional damping approximation can be
applied. Fig. 3 gives the errors for 1
0 0 0 0
= = = =
p p m m
q q q q

(similar patterns were observed for
other initial conditions) and shows that transformation in the physical space dramatically reduces
the error due to the hard constraint on the modal matrix norm, whose final value is 12 . 0 ) ( = U o .


11

Fig. 2. Natural frequencies distribution and initial conditions response errors
for the arbitrary design

Fig. 3. Natural frequencies distribution and initial conditions response errors
for the hard modal constraint design ( 12 . 0 ) ( = U o , 0.2
lj
e = )


12

Fig. 4. Natural frequencies distribution and initial conditions response errors
for the soft modal constraint design ( 8 . 0 ) ( = U o , 0.2
lj
e = )
Fig. 4 corresponds to a design obtained for 2 . 0 =
lj
e , 18 , 8 . 0
max min
= = e e and a soft
constraint on the modal matrix norm: 8 . 0 ) ( s U o . The error norms, shown in Fig.4 for
1
0 0 0 0
= = = =
p p m m
q q q q , increase and the modal and physical errors are noticeable closer
because of the soft constraint. Similar results were obtained for other initial conditions. The
design is still good and proportional damping approximation can be used.
The next set of results reveals very interesting features. Firstly, even if the prescribed
separation is increased it may so happen that the results are worse than the ones obtained for a
smaller separation. Fig. 5 corresponds to such a design, in which 4 . 0 =
lj
e , 18 , 2
max min
= = e e
and the constraint on ) (U o was eliminated. It can be easily ascertained that the natural
frequencies range is similar to the one in Fig. 4, but the error in modal coordinates is much
larger, even though the minimum separation between natural frequencies doubled (the initial
conditions considered in Fig. 5 are the same as before). Secondly, because the constraint on
) (U o has been removed, the error in the physical space is hugely amplified (the maximum
singular value of the modal matrix is 55 . 1 ) ( = U o ). This is not a good design if proportional
damping approximation is thought after. However it is a good dynamic design for other purposes
because, for example, sufficient separation between the natural frequencies is achieved.


13

Fig. 5. Natural frequencies distribution and initial conditions response errors
for the no modal constraint design ( 55 . 1 ) ( = U o , 0.4
lj
e = )

Fig. 6. Natural frequencies distribution and initial conditions response errors
for the no modal constraint design with small damping (c
1
=0.1)


14
These results indicate that separation between natural frequencies may not be sufficient
for accurate proportional damping approximation. They complement similar results obtained
when harmonic excitations, rather than nonzero initial conditions, were considered (see [19-24]).
Another interesting result is obtained if the damping is substantially reduced. Fig. 6
corresponds to the same design as before (i.e. same design parameters) but when c
1
=0.1 (a very
lightly damped structure). It is remarkable that the errors are substantially smaller. This is in
agreement with Gawronskis [12] observation, that for lightly damped structures separation of
natural frequencies is sufficient for accurate proportional damping approximation in modal space.
Nevertheless, the constraint on the modal matrix should be introduced in the design for accurate
approximation in the physical space.
The design parameters associated with Figs. 2-6 are given in Table 1. All the results
presented in this article correspond to situations in which the algorithm converged to solutions at
which some of the separation constraints (12) are tight hence the minimum prescribed separation
is achieved. Except for the solution corresponding to Fig. 4, when the modal matrix norm
constraint was also tight ( 8 . 0 ) ( = U o ), none of the other constraints were tight.
k
1
k
2
k
3
k
4
m
1
m
2
m
3
m
4
m
5
m
6
c
1

lj
e

Fig.
1 1 1 1 1 3 4 5 1 1 1 0.0 2
640.40 747.40 2055.5 1974.4 706.5 63 320.3 258.3 7.2 104.6 1 0.2 3
104.41 88.86 0.23 127.88 48.04 1.61 23.45 22.08 0.49 1.77 1 0.2 4
13.14 31.45 51.85 42.77 1.43 0.36 1.11 1 0.65 0.49 1 0.4 5
13.14 31.45 51.85 42.77 1.43 0.36 1.11 1 0.65 0.49 0.1 0.4 6
Table 1. Design parameters values for Figs. 2-6.
These results indicate that if proportional damping approximation is the main goal of the
design, one should not count on obtaining increasingly accurate approximations just by
increasing the separation between natural frequencies. This is due to the fact that the approximation
errors dependence on the natural frequencies separation is nonlinear. Also other factors, except for the
natural frequencies, like the modal damping matrix, play a role in the approximation error.
A better approach to the design of structures for proportional damping approximation is
to consider design requirements directly related to the approximation error like properties of the
transfer matrices between the initial conditions or the inputs and the approximation error. These
matrices can be easily obtained using the Laplace transform. For nonzero initial conditions,
0
0 0
= =
p m
q q

and 0
0 0
= =
p m
q q , and non-zero external input, f(t), Eqs. (7) and (8) yield
2 2
0 0 0
( ) ( ( ) ) ( ) ( )
T
m m m m m m m
s q s sq q C sq s q q s U f s + + O = (28)
2 2
0
( ) ( ) ( ( ) ) ( ) 0
m p m n m m m
s s C s s C sq s q s c c c + + +O = (29)
where e j s = . From Eqs. (28) and (29), it easily follows that


15
2
0 0
1
( ) ( ) ( )
T
m m m
s G s U f s q q
s
c
| |
= + O
|
\ .

(30)
where
( ) ( ) s s C I s C s C I s s G
m n p
1
2 2
1
2 2
) (

O + + O + + = . (31)
One approach is to design the structure such that the error norm is minimized when
inputs or initial conditions with certain properties are considered. This is a topic of future
research.

5. Conclusions
For future controllable structures dynamic requirements will play an important role,
hence constraints on their dynamic characteristics should be considered in the design process.
Specifically, constraints on the natural frequencies should be imposed due to these frequencies
influence on the dynamic response. A key dynamic design requirement is that the natural
frequencies are sufficiently separated, which is crucial for simple and exact computations, and
accurate proportional damping approximation. Other requirements, which are especially
important for controllable structures, are that the natural frequencies are lower and upper
bounded. Thus the dynamic design problem formulated in this article includes separation
constraints on the natural frequencies and lower and upper bounds on their values.
A numerical algorithm for the solution of this problem was proposed. The algorithm is
very fast because of two key features: it relies on active set methods and the gradients used in the
iterative solving process are analytically computed. The algorithm can be easily extended to
solve problems which include other constraints whose gradients can be analytically computed; an
upper bound constraint on the maximum singular value of the modal matrix is particularly
important for accurate proportional damping approximation in the physical space. The algorithm
was evaluated on tensegrity structures design and in all cases convergence was obtained very fast
(milliseconds to seconds).
Analysis of the relation between natural frequencies separation and the accuracy of
proportional damping approximation indicated that separation of these frequencies is essential
but it is may be a misleading design criterion. For example the approximation error might
increase when the separation between natural frequencies increases, even when only the response
to nonzero initial conditions is considered. Hence, if the main goal of the design is accurate
proportional damping approximation, a problem which includes constraints directly related to the
approximation error norm should be formulated and solved.

Appendix A: Proportional damping yields non-defective systems.

Consider the case of a diagonal modal damping matrix which can be written as
1 0 ), ( diag , 2 < < = O = =
i i
T
m
Z Z CU U C . (32)


16
It can be shown using linear algebra that
(

O O
=
Z
I
A
2
0
2
is diagonalizable:
1 2 0.5
0
, , [ ]
0
I I
E AE E Z j I Z

A ( (
= = A = O+ O
( (
A A A

(33)
where the inverse of E is
(
(

A A A A A
A A A A A +
=

1 1
1 1
1
) ( ) (
) ( ) ( I
E . (34)
The state transition matrix is
2 2 0.5 2 1 2 0.5 2
2 0.5 2 2 2 0.5 2
[cos( ) [ ] sin( )] [ ] sin( )
[ ] sin( ) [cos( ) [ ] sin( )]
At
Z t Z t
Z t Z t
e
e I Z t Z I Z I Z t e I Z I Z t
e I Z I Z t e I Z t Z I Z I Z t
O O
O O
=
(
O + O O O
(
(
O O O O

(35)
and the response of the system does not include secular terms.

References
[1] W. Dorn, R. Gornery, M. Greenberg, Automatic design of optimal structures, Journal de Mecanique 3 (1964) 25-
52.
[2] M.P. Bendsoe, A. Ben-Tal, J. Zowe, Optimization methods for truss geometry and topology design, Structural
Optimization 7 (1994) 141-159.
[3] Special issue on multidisciplinary design optimization, Journal of Aircraft 36 (1) (1999).
[4] N.L. Pedersen, A.K. Nielsen, Optimization of practical trusses with constraints on eigenfrequencies,
displacements, stresses and buckling, Structural and Multidisciplinary Optimization 25 (2003) 436-445.
[5] L. Starek, D. J. Inman, A. Kress, A symmetric inverse vibration problem, Transactions of the ASME Journal of
Vibration and Acoustics 114 (1992) 564-568.
[6] L. Starek, D.J. Inman, A symmetric inverse eigenvalue vibration problem with overdamped modes, Journal of
Sound and Vibration 181 (5) (1995) 893-903.
[7] L. Starek, D.J. Inman, Symmetric inverse eigenvalue vibration problem and its application, Mechanical Systems
and Signal Processing 15 (1) (2001) 11-29.
[8] R.W. Wlezien, G.C. Horner, A.R. McGowan, S.L. Padula, M.A. Scott, R.J. Silcox, J.O. Simpson, The Aircraft
Morphing Program, AIAA paper AIAA-1998-1927, in Proceedings of the AIAA/ASME/ASCE/AHS/ASC
Structures, Structural Dynamics, and Materials Conference and Exhibit, Long Beach, CA, 1998.
[9] B.F. Spencer Jr., S. Nagarajaiah, State of the art of structural control, Journal of Structural Engineering 129 (7)
(2003) 845-856.
[10] S.K. Dwivedy, P. Eberhard, Dynamic analysis of flexible manipulators, a literature review, Mechanism and
Machine Theory 41 (7) (2006) 749-777.
[11] G. Franklin, J.D. Powell, A. Emami-Naeini, Feedback Control of Dynamic Systems, 5
th
ed., Prentice Hall,
Englewood Cliffs, NJ, 2005.
[12] W.K. Gawronski, Advanced Structural Dynamics and Active Control of Structures, Springer, New York, NY,
USA, 2004.
[13] T.K. Caughey, Classical normal modes in damped linear dynamic systems, Journal of Applied Mechanics 27
(1960) 269-271.
[14] M. Tong, Z. Liang, G.C. Lee, An index of damping non-proportionality for discrete vibrating systems, Journal
of Sound and Vibration 174 (1994) 37-55.
[15] S.M. Shahruz, Comments on an index of damping non-proportionality for discrete vibrating systems, Journal of
Sound and Vibration 186 (3) (1995) 535-542.


17
[16] W.K. Gawronski, J.T. Sawicki, Response errors of non-proportionally lightly damped structures, Journal of
Sound and Vibration 200 (4) (1997) 543-550.
[17] S. Adhikari, Optimal complex modes and an index of damping non-proportionality, Mechanical Systems and
Signal Processing 18 (2004) 1-24.
[18] U. Prells, M.I. Friswell, A measure of non-proportional damping, Mechanical Systems and Signal Processing
14 (2) (1999) 125-137.
[19] S. Park, I. Park, F. Ma, Decoupling approximation of non-classically damped structures, AIAA Journal 30 (9)
(1992) 2348-2351.
[20] S.M. Shahruz, A.K. Packard, Approximate decoupling of weakly damped linear second-order systems under
harmonic excitations, in: Proceedings of the IEEE Conference on Decision and Control, Tucson, AZ, USA, 1992.
[21] S. M. Shahruz, G. Langari, Closeness of the solutions of approximately decoupled damped linear systems to
their exact solutions, Transactions of ASME Journal of Dynamic Systems, Measurement, and Control 114 (1992)
369-374.
[22] S. Park, I. Kim, F. Ma, Characteristics of modal decoupling in non-classically damped systems under harmonic
excitation, Journal of Applied Mechanics 61 (1994) 77-83.
[23] S.M. Shahruz, A.K. Packard, Approximate decoupling of weakly non-classically damped linear second-order
systems under harmonic excitations, Journal of Dynamic Systems, Measurement, and Control - Transactions of
the ASME 115 (1993) 214-218.
[24] S.M. Shahruz, P.A. Srymatsia, Approximate solutions of non-classically damped linear systems in normalized
and physical coordinates, Journal of Sound and Vibration 201 (2) (1997) 262-271.
[25] J.S. Jensen, N.L. Pedersen, On maximal eigenfrequency separation in two material structures: the 1D and 2D
scalar cases, Journal of Sound and Vibration 289 (2006) 967-986.
[26] G.M.L. Gladwell, Inverse Problems Vibrations, Martinus Nijhoff, Boston, MA, USA, 1986.
[27] C. Sultan, Tensegrity structures: from avant-garde art to next generation controllable structures, in: Proceedings
of the World Conference on Structural Control and Monitoring, San Diego, CA, USA, 2006.
[28] D.G. Luenberger, Linear and Nonlinear Programming, Addison-Wesley, Reading, MA, USA, 1984.
[29] C. Sultan, M. Corless, R.E. Skelton, The prestressability problem of tensegrity structures. Some analytical
solutions, International Journal of Solids and Structures 38-39 (2001) 5223-5252.
[30] C. Sultan, M. Corless, R.E. Skelton, Peak to peak control of an adaptive tensegrity space telescope, in:
Proceedings of the SPIE International Symposium on Smart Structures and Materials, Newport Beach, CA, USA,
1999.
[31] C. Sultan, M. Corless, R.E. Skelton, Tensegrity flight simulator, Journal of Guidance, Control, and Dynamics
23 (6) (2000) 1055-1064.
[32] C. Sultan, R.E. Skelton, Deployment of tensegrity structures, International Journal of Solids and Structures 40
(18) (2003) 4637-4657.
[33] C. Sultan, R.E. Skelton, A force and torque tensegrity sensor, Sensors and Actuators: A - Physical 112/2-3
(2004) 220-231.
[34] C. Sultan, Modeling, design, and control of tensegrity structures with applications, Ph.D. Thesis, Purdue
University, West Lafayette, IN, USA, 1999.

Potrebbero piacerti anche