O P T I M A L DESIGN PROCEDURES

FOR A C T I V E STRUCTURES"

By J. Z. Cha, 1 J. M. Pitarresi,2 and T. T. Soong,3 Member, ASCE

ABSTRACT: The problem of optimal design of active structures has been shown
to be one of simultaneous structural and control design optimization. The basic
equations represent, in general, a large system of coupled nonlinear differential
equations. The purpose of this paper is to present a numerical design procedure
which optimizes the structural and controller parameters simultaneously. Two ex-
amples illustrating the nature of the optimization problem and the numerical pro-
cedure involved are presented.

INTRODUCTION

This work is a sequel to Soong and Manolis (1987) in which the general
concept of an active structure is presented. An active structure is defined as
one consisting of two types of load resisting members: the traditional static
(or passive) members that are designed to support basic design loads, and
the dynamic (or active) members whose function is to augment the struc-
ture's capability in resisting extraordinary dynamic loads. Their integration
in an optimal fashion produces a structure which is adaptive to changing
environmental loads and changing usages. As shown in Soong and Manolis
(1987) and Soong and Pitarresi (1987), the problem of active structure design
is one of simultaneous structural and control design optimization, which is
an application of the general theory of optimal control of parametric systems.
The general formulation of the problem of optimal design of active struc-
tures is presented in Soong and Manolis (1987). The major purpose of this
paper is to develop an efficient simultaneous optimization procedure for ac-
tive structural design. Two design examples illustrating the numerical pro-
cedure are presented.

BASIC EQUATIONS

The basic equations governing the optimal design of an active structure

can be stated as follows: Consider an n-degree-of-freedom active structure.
The equations of motion take the general form
z(r) = A(|)z(r) + B(|)u(f) + D({)f(f) (1)
where
"Parts of this work were presented at the August 17-20, 1987, ASCE Structures
Congress, held at Orlando, FL.
'Post-Doctoral Res. Assoc, Dept. of Civ. Engrg., State Univ. of N.Y. at Buffalo,
Buffalo, NY 14260.
2
Asst. Prof., Dept. of Mech. Engrg., State Univ. of N.Y. at Binghamton, Bing-
hamton, NY 13901.
3
Prof., Dept. of Civ. Engrg., State Univ. of N.Y. at Buffalo, Buffalo, NY 14260.
Note. Discussion open until May 1, 1989. To extend the closing date one month,
a written request must be filed with the ASCE Manager of Journals. The manuscript
for this paper was submitted for review and possible publication on July 20, 1987.
This paper is part of the Journal of Structural Engineering, Vol. 114, No. 12,
December, 1988. ©ASCE, ISSN 0733-9445/88/0012-2710/$!.00 + $.15 per page.
Paper No. 23009.

2710
wnloaded 22 Feb 2010 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]
 0
A(© = _-M-'(©K(© (2)
-M-*(©C(©
0
»(© = _-M-'(g)G(i)_ (3)

0
D© = (4)

and
At) = [xr(0,*r(01 (5)
In the above, x{i) is the n-dimensional position vector; M(|), C(|) and K(|)
are, respectively, the n X « mass, damping and stiffness matrices dependent
on a p-dimensional vector | of structural parameters; f(f) is the excitation
force vector; and the n X m matrix G(|) (m ^ «) determines the spatial
distribution of the active constituents which, together with the control force
m-vector u(t), represent the active characteristics of the structure. The initial
condition is z(0) = z0.
The basic problem of active structure design is to determine | and u(0
which minimize an objective functional of the form

/(*.€ u) = [zrQz + urQ2u + W(&\dt (6)

Jo
where
Q„ o (7)
Q =
0 Q,
In the above, W(g) represents a non-negative cost function depending on ij
but not on t in general. The matrices Q 0 , Qj and Q2 are the usual weighting
matrices with appropriate dimensions. The structural parameters, | , satisfy

g^fe (8)
where £, represents values of the structural parameters corresponding to the
base passive structure.
Extremization of the performance index function / given by Eq. 6 subject
to the constraining Eqs. 1 and 8 can be accomplished through the use of
variational calculus. The basic equations of optimal design of active struc-
tures are (see Soong and Manolis 1987)
Az + Bu + Df - z = 0, z(0) = 0 • (9)
2Qz + A'X + X = 0, \{tf) = 0 .. (10)
2Q2u + BrX = 0 (11)
r
X (V4Az + V4Bu + VeDf) + VSW = 0 (12)
over the time period tf, where X(0 is the 2«-dimensional vector Lagrange
multiplier and V| is a gradient operator with respect to the parameter vector
| . The equations above represent a system of differential equations with four
unknowns, namely, z(r), u(f), | and Kit). Their solution determines the op-
timal configuration for the active structure.
2711
wnloaded 22 Feb 2010 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]
 SIMULTANEOUS SEARCH PROCEDURE

Eqs. 9-12 represent, in general, a system of coupled nonlinear equations.

Note that this nonlinearity is present even when the equations of motion
represent linear elastic response, as is assumed in this paper. (Material and/
or geometric nonlinearities may be included, however, at the expense of
making the system of equations computationally more involved.) Because
of the complex nature of these equations, numerical techniques are used to
obtain a solution.
An iterative sequential search procedure is studied in Soong and Pitarresi
(1987) in which, at each iteration, the structural parameters are held constant
until the performance index defined by Eq. 6 is minimized. The structural
parameter values are then altered and the process is repeated until the global
optimum is achieved. It is clear that, in terms of efficiency and convergence
properties, a more desirable procedure is one in which the controller param-
eters and the structural parameters are optimized simultaneously. Simulta-
neous search procedures work in the full design space and update the struc-
tural and controller parameters simultaneously toward the final optimum design.
Hence, these procedures cannot only deal with high dimensional spaces with
fast convergence rates, but they may also avoid possible degeneracy into
sub-spaces caused by sequential search methods.
Starting from prescribed initial values, the structural and controller param-
eters are updated at the k-th iteration by (k = 0,1,2,...)
£*+1 = |* + <xkdl (13)

uf+1 = uf a*d* (14)

In the above, d| and d* denote search directions in the scalar space and the
functional space, respectively, whereas a* is a unified scalar amplitude for
both updating formulae and is selected so that the objective functional (Eq.
6) decreases with <xk along the search directions.
The search directions can be determined by, for example, the use of the
conjugate gradient method (Fletcher 1980). In this case, the search direction
calculations are based only upon the gradient information provided by
IIV/IP
dk = -VJ" + TT^— V/" 1 (15)

where dk is the search direction for the k-th iteration. V is the gradient vector
with respect to either u or (;, and J is the augmented Lagrangian function
formed by adjoining the constraining equations of motion to the performance
index via the vector Lagrange multiplier X. The algorithm following this
approach proceeds as follows:

1. Select initial values u° and |°.

2. Solve the state and costate equations (Eqs. 9 and 10) for z* and kk, k =
0, 1, ... .
3. Calculate the gradient with respect to n and g using
V„7? = VuH(uk,£,\zk,k\t) (16)

2712
Downloaded 22 Feb 2010 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]
 . f ....
V* = V{//(«Ms,z*,X*,T)rfT (17)
Jo
where
H = z r Qz + u r Q 2 u + W(© + X r [A(|)z + B(g)u + D(£)f] (18)
4. Calculate the search directions, for k > 1
df, = -V„J* + ^ ~ r 2 V.JT1 (19)
||W
4112

dI=-V5/ + ^ - ^ V / - (20)

and, for k = 0
d:=-V„y? (21)
A\ = - V (22)
where
\\Vjf = ( V ) * V + ||V„7||2 (23)
and
•'/
IMI1122 =
- I 1X1 T T\TT.
(VM VuJ7dr (24)
Jo
5. Update u and £ by using Eqs. 13 and 14.
6. Repeat steps 2 - 5 until Eqs. 9-12 are satisfied with prescribed accuracy.

The state equation, Eq. 9, and the costate equation, Eq. 10, are numer-
ically integrated using Newmark's beta method (Newmark 1959) and Gal-
erkin's two-point recurrence method (e.g., Zienkiewicz 1977), respectively.
The quadratic performance index, Eq. 6, is integrated using Simpson's 1/3
rule.
Unlike the sequential search procedure, where a structural optimization is
performed followed by a control force optimization, the proposed simulta-
neous procedure directs the solution toward a single final optimum design.
Consequently, only one global optimization is performed. Furthermore, this
procedure is not limited to linear structural behavior or linear control laws.

NUMERICAL EXAMPLES

Can substantial changes in the structural behavior be realized by allowing

some of the members of the structure to become active? To answer this
question, the following two examples are presented.

Example 1
Consider the design of a simple steel frame supporting a static vertical
load and subjected to a horizontal base acceleration xgr(i), as shown in Fig.

2713
Downloaded 22 Feb 2010 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]
 x(t)

FIG. 1. Simple Steel Frame

1. The frame consists of four identical columns with two sets of pretensioned
diagonal tendons which act as active members. The dimensions are: L =
48.0 in. (1.219 m) and h = 40.0 in. (1.106 m), with the mass of the dead
load equal to 16.52 slugs (2,890 kg). If the frame is modeled as a single-
degree-of-freedom shear building, and if it is assumed that the columns are
of rectangular cross section with the base b, equaling twice the depth, then
the stiffness K, mass M, and damping C can be expressed as
32Eb4 2
K= M = SpLb2 + 16.52
^l (25)

where E is the elastic modulus, taken as 29 X 106 psi (1.38 X 106 kPa) and
p is the density, taken as 7.35 x 10~4 lb s2/in.4 (2.58 X 10"4 kg/cm 4 ). The
damping ratio, £, is taken as 1.24%.
If one considers the base structure to be that which is just capable of
supporting the static loads without the assistance of the active members, then
it is readily verified that, for a buckling mode of failure, the minimum struc-
tural parameter b is 0.34 in. (0.86 cm).
For a support acceleration of
xgr(t) = 0.15g sin Ant, (26)
the equation of motion becomes
Mx + Cx + Kx = —Mxsr — Gu . . (27)
For convenience the control force, u, is transformed to a control displace-
2714
Downloaded 22 Feb 2010 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]
 300.0 •
290.0
2S0.0 -
270.0 •
260.0 •
250.0
240.0
230.0
220.0 -
210.0 •
200.0 •
190.0 •
180.0 •
170.0 •
160.0 -
150.0 -
140.0 •
130.0 •
120.0 •
110.0 •
100.0 •
I I 1 1 1 1 1
o.: 0.70 0.90 1.10 1.30

WIDTH OF COLUMNS [INCHES]

FIG. 2. Performance Index for One-Story Frame

ment by writing G as a function of the tendon stiffness and connection angle

a. Each tendon has a stiffness of 2,124 lb/in. (3,720 N/cm).
The time domain of interest is chosen as 2 sec with a step size of 0.005
sec and integration parameters of 1/6 for 6 and 1/3 for 0, where p and G
are the integration parameters for Newmark's beta method and Galerkin's

FIG. 3. Displacement Response of One-Story Frame

2715
wnloaded 22 Feb 2010 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]
 300.00

200.00 -

100.00 -

_i

LJ
0 0.00 -

-100.00 -

-200.00 -

-300.00 - | 1 1 1 1 1 1 1 1 1 ; 1 1 1 1 1 1 1 1 1
0.00 0.20 0.40 0.60 0.80 1.00 1.20 1.40 1.60 1.80 2.00

TIME [SEC]

FIG. 4. Required Control Force

method, respectively. The weighting matrices should be carefully chosen to

reflect the proper balance between the controlled response, input energy and
structural weight (cost). If at all possible, the weights should reflect some
physical parameter of the structure. In this example, the weight for the dis-
placement response and the required control displacement are chosen to be
proportional to the system stiffness. In particular, they are 3,967 and 7,934
for the displacement and control, respectively. For the velocity weight, a
value of 16.52 is used which is proportional to the structure's dead mass.
Note that these values have been chosen arbitrarily, consequently the min-
imum will be somewhat subjective.
Fig. 2 shows the performance index versus the structural parameter b, for
both the active and passive systems. The active system is minimized when
b = 1.05 in. (2.67 cm) and PI (performance index) = 124.15 while the
passive system is minimized when b = 1.10 in. (2.79 cm) and PI = 131.92.
For values of b in the range from 0.34 in. (0.86 cm) to 1.4 in. (3.56 cm),
it is clear from the reduced performance index that the active structure (ac-
tive members + passive members) is superior to an equivalent passive struc-
ture (passive members only). As b is increased above 1.40 in. (3.56 cm),
the required control force approaches zero. Clearly, no benefit is gained from
the active members in this range. Fig. 3 shows the displacement response
for the optimum b of 1.05 in. (2.67 cm). It is noted that the active members
successfully reduce the response. The required control force in the tendons
is shown in Fig. 4. Note that the stipulation on the final condition of the
costate equation requires that the control force be driven to zero at the end
of the time domain.
Example 2
In the preceding example, a relatively simple case with only one degree
of freedom was considered. In what follows, a somewhat more realistic
2716
Downloaded 22 Feb 2010 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]
 D/ v \
i. \ * k ;

V(t)

FIG. 5. King-Post Bridge Beam

structure is investigated. Consider the king-post bridge beam of Fig. 5. The

two king-posts, serving as active members, are located along the beam and
are capable of applying point forces directly to the beam. Sufficient preten-
sioning of the cables is assumed to allow both upward and downward control
forces. A moving load of constant magnitude P(x) and velocity V{t) is ap-
plied to the beam.
The governing partial differential equation of motion is taken as:
,d4Y(x,t) d2Y(x,t)
EI- + pA = P{x) + Gu(t) (28)
dx4 dt2
where EI is the flexural rigidity and pA is the mass per unit length. Assuming
that

Y{x,t) = 2 <t>,(*H(0 (29)

where <}>, are the mode shapes, Eq. 28 is transformed into an infinite number
of second-order ordinary differential equations. The !-th equation is ex-
pressed as
il, + mjt\i = Pi + GfU, (30)
where Pt and Gtut are the modal counterparts of P(x) and Gu(t). The natural
frequency co, and the mode shape (j), are given by

(o,. = W VE7/P~A (31)

/nrx\
<h = sin \ — \ (32)

Unfortunately, closed form solutions to the transformed equations of mo-

tion are not possible since the time variation of the control force is unknown
a priori. By considering linear elastic response, however, the solution can
be separated into two parts: (1) Moving load solution; and (2) control force
solution, with superposition of the two providing the total response. A closed
form solution to the moving load is given in Timoshenko and Young (1955).

2717
Downloaded 22 Feb 2010 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]
 TABLE 1. Performance Comparisons

Case PI b (in.) J W (in.) C/m„ (lb) XL XR Iterations

(1) (2) (3) (4) (5) (6) (7) (8)
1 145.23 1.45 -0.268 N/A N/A N/A 5
2 123.16 0.85 -0.108 608 0.30 0.70 27
3 102.44 0.93 -0.069 490 0.44 0.56 66
4 93.96 0.77 -0.036 496 Fig. 8 Fig. 8 87

So the problem is now reduced to finding the response due to the control
force.
In practice, it is common to use only a finite number of modes when
calculating the dynamic response. Letting n = the number of modeled modes
and m = the number of independent control forces, we can define three
classes of control: (1) n < m; (2) n > m; and (3) n = m. In the last case,
the equations of motion can be written as n independent modal equations,
each with its own modal control force. Transformation from modal forces
to physical forces is direct due to the one-to-one correspondence between n
and m. Such an approach is referred to as independent modal space control
(IMSC) (Meirovitch et al. 1983). Of the remaining two, case 1 is not con-
sidered here for essentially economic reasons; that is, the general trend is to
have less control forces than modeled modes. Therefore, case 2 is the focus
of this example. Since we are considering fewer control forces than modes,
the transformation from modal to physical control forces involves the pseudo-
inverse of an n X m matrix. This implies that the physical control forces
may contain errors that lead to system degradation (Meirovitch et al. 1983).
To avoid this problem the control force vector is not transformed into modal

~i i i i i i i i — i — i — i — i — i — i — i — i — i — i — i — i —
0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14 0.16 0.18 0.20

TIME [SEC]

FIG. 6. Displacement Response Comparison for King-Post Beam

2718
ownloaded 22 Feb 2010 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]
 -100 H 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 r—i 1-
0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14 0.16 0.18 0.20
TIME [SEC]

FIG. 7. Required Control Force Comparison for King-Post Beam

space. Instead, the m x 1 physical control vector is multiplied by the n x

m modal participation matrix G. The resulting modal differential equations
of motion are externally coupled by the physical control force vector. The
only requirement on the control force vector is that the system must be ren-
dered controllable.
For the numerical analysis, the first five modes are considered. The in-

0.00 0.02 0.04 0.06 0.08 0.10 0.14 0.16 0.20

TIME [SEC]

FIG. 8. Time-Dependent Location of Active Members for Case 4

2719
Downloaded 22 Feb 2010 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]
 tegration parameters, p and 8, are taken as 1/6 and 1/3, respectively. Four
hundred time steps each of duration 0.0005 sec are used. The moving load
has a constant magnitude of 1,000 lb (4.448 kN) and a horizontal velocity
of 500 in./sec (12.7 m/s). The material and cross-sectional properties of
the beam are the same as those used in example 1. The weights for the
performance index are chosen as follows: displacement, Q0, 5,000 (all modes);

f"Eld.,J>J u I , It V t = 0,02 sec.

t' " '

•piter'""" t = [Link] sec,

TT7TTT
t = 0.06 sec,

t = 0.08 sec.

""••>" t = 0,10 sec,

t = 0,12 sec,

t = 0,14 sec.

^frr"™™" t = 0.16 SeC .

t = 0,18 sec,

•VTT*|- t = 0,20 sec.

FIG. 9. King-Post Configuration during Loading for Case 4

2720
wnloaded 22 Feb 2010 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]
 velocity, Q,, zero (all modes); and control force, Q2> 1,000 (for eacli king-
post).
In an effort to demonstrate both the simultaneous search method and the
concept of an active structure, the king-post example is now investigated
through a number of different cases. Specifically, the following four cases
are identified.

Case 1
Passive structure. The only design parameter considered is the beam width,
b, subjected to the simple bounds: 0.10 in. (0.254 cm) < b < 10.0 in. (25.4
cm).

Case 2
Two active members (i.e., king-posts) located at x/L = 0.30 and 0.70,
as well as the beam width, are the design variables. The two active members
are capable of generating both up and down control forces. No bounds are
imposed on the magnitude of the control forces. The bounds on b are the
same as in Case 1.

Case 3
Identical to Case 2 except that, along with the two active members and
the width, the locations of the two active members are also considered as
design variables. Hence, five design variables are considered with the bounds
on the active member locations of: x/L = 0.0 < XL s 1.00; x/L = 0.0 £
XR ^ 1.00, which forces the two locations, XL for the left member and XR
for the right, to be along the beam.

Case 4
In this variation, the effect of allowing the two active members to change
position with time is considered. The simple bounds are identical to those
of Case 3. Note that while this is a logical extension of the active structure
concept, the realization of such a "fully" active structure may be difficult.
The solution to these four cases is obtained using the simultaneous search
procedure outlined in the previous section. The results for each are sum-
marized in Table 1. The quantities listed in Table 1 are the performance
index (PI), the beam width (b), the maximum centerline deflection of the
beam (ymax), the maximum control force (umax), the active member location
(XL and XR) and the number of iterations for convergence. Comparison of
the centerline displacement response and required control force in the left
king-post is shown in Figs. 6 and 7, respectively. Also, the moving locations
of the two king-posts for Case 4 are illustrated in Figs. 8 and 9.
It is seen from Table 1 and Figs. 6 and 7 that Case 4, with a PI of 93.96,
represents the optimum configuration of the king-post beam. This is ex-
pected since this case has the highest capacity to act as an active structure.

CONCLUDING REMARKS
A numerical design procedure based upon the simultaneous structural and
control force optimization has been proposed. Through two examples it has
been demonstrated that the simultaneous procedure is effective in the design
of the structural cross section, the required control force, as well as the
2721
Downloaded 22 Feb 2010 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]
 placement of the active members. Since the numerical procedure is general,
it is not limited to linear structural behavior or linear control laws.

ACKNOWLEDGMENT

This work was supported in part by the National Science Foundation under
Grant No. MSM 8604720.

APPENDIX I. REFERENCES

Fletcher, R. (1980). Practical methods of optimization. Vol. 1, John Wiley and Sons,
New York, N.Y.
Meirovitch, L., Barum, M., and Oz, M. (1983). "A comparison of control tech-
niques for large flexible systems." J. Guidance, AIAA, 6(4), 302-310.
Newmark, N. M. (1959). "A method of computation for structural dynamics." J.
Engrg. Mech. Div., ASCE, 85, 67-94.
Soong, T. T., and Manolis, G. A. (1987). "Active structures." J. Struct. Engrg.,
ASCE, 113(11), 2290-2302.
Soong, T. T., and Pitarresi, J. M. (1987). "Optimal design of active structures."
Computer Applications in Structural Engineering, D. R. Jenkins, ed., ASCE, 579-
591.
Timoshenko, S., and Young, D. H. (1955). Vibration problems in engineering. D.
Van Nostrand Company, Inc., Princeton, N.J.
Zienkiewicz, O. C. (1977). The finite element method, 3rd Ed., McGraw-Hill, New
York, N.Y.

APPENDIX II. NOTATION

The following symbols are used in this paper:

A = cross-sectional area;
A = system matrix;
B = control force vector coefficient matrix;
b = width of rectangular cross section;
C = damping matrix;
D = applied force vector coefficient matrix;
d = search direction vector;
E = Young's modulus;
f = applied force vector;
G = control force participation matrix;
H = Hamiltonian;
I = moment of inertia;
J = augmented functional;
K = stiffness matrix;
k = iteration number;
L = length of member;
M = mass matrix;
m = number of controllers;
n = number of degrees-of-freedom;
P = moving load;
P = number of structural parameters;
Q„ = displacement weighting matrix;
Qi = velocity weighting matrix;
2722
ownloaded 22 Feb 2010 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]
 Q.2- CGiicrGj. iorcc wcigiitnig iii&tnx;
s= conjugate gradient search direction;
t
= time;
u= control force vector;
w = structural weight function;
x,x,x = displacement, velocity and acceleration vectors;
XL = location of active member left;
XR = location of active member right;
Y = vertical displacement;
z = state vector;
a = one-dimensional search amplitude;
P = Newmark beta parameter;
7 = conjugate search direction;
in = generalized modal coordinate;
e = Galerkin parameter;
X = costate vector;
1 = structural parameter vector;
p = material density;
-e-

= mode shape; and

(i) = frequency.

2723
ownloaded 22 Feb 2010 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]