You are on page 1of 13

JOURNAL OF AIRCRAFT

Vol. 41, No. 2, March–April 2004

Mode Acceleration Based Random Gust Stresses


in Aeroservoelastic Optimization
Frode Engelsen∗
The Boeing Company, Seattle, Washington 98124
and
Eli Livne†
University of Washington, Seattle, Washington 98195-2400

The present paper describes a method for obtaining accurate design-oriented stress and stress-sensitivity infor-
mation from reduced-order linear-time-invariant state-space models of integrated aeroservoelastic systems, using
Lyapunov’s Equation for calculating covariance matrices of the displacement and stress responses. A complete for-
mulation of the reduced-order stress gust response problem for aeroservoelastic design synthesis, tailored toward
integration with control-system design techniques based on modern control, is presented. It includes an adap-
Downloaded by BRISTOL UNIVERSITY on March 5, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.9329

tation of the mode-acceleration method, reduced-order analytic sensitivities of stress covariances, and efficient
approximations to be used in a nonlinear programming/approximation concepts approach to design optimization.
A realistic aeroservoelastic model of a typical passenger airplane is used as a test case, and the paper includes
results of convergence studies for the assessment of order-reduction effects on the accuracy of the integrated
structure/aerodynamic/control models.

Nomenclature n Ls = number of lag terms in a minimum-state


[A], [B], [C], [D] = state-space model matrices (with unsteady aerodynamic force rational
appropriate subscripts to designate approximation (structural and
sensors, actuators, control laws, control-surface motions)
gust filter, and overall system) nq = number of reduced-order structural
[Aiφ ] = Roger aerodynamic matrices (real) degrees of freedom
for fitting [A(s)][φ] [Eq. (8)] n u × n q nu = full-order number of structural
[A(s)] = Laplace transformed full-order influence degrees of freedom
coefficient matrix (associated with [Q] = intensity matrix of the white noise
full-order structural model) n u × n u driving the gust filter
[C] = viscous damping matrix (structural) {q} = vector of reduced-order generalized
g
cis , ci = aerodynamic lag terms for structural coordinates
and gust forces, respectively qD = dynamic pressure
[D] = aerodynamic matrix in a minimum-state [S] = matrix used to calculate stresses
rational approximation form [Eq. (9)] from full-order displacements
[E] = aerodynamic matrix in a minimum-state Sref = wing reference area
rational approximation form [Eq. (9)] s = Laplace variable
[I ] = identity matrix {sx x s yy sx y }T = vector of local stresses in a plane-stress
[K ] = stiffness matrix thin structural element
[M] = mass matrix U∞ = flight speed
NLs = number of aerodynamic lag terms {u} = full-order displacement vector
nA = order of the state-space model w = vertical speed of atmospheric gusts
of the actuators [X ] = covariance matrix for complete
n CO = order of state-space model of the aeroservoelastic system with structural
multi-input/multi-output control law block degrees of freedom reduced by mode
nc = number of active control surfaces displacement basis
ng = order of the state-space model {x} = state vector for complete
of the gust filter aeroservoelastic system
n Lg = number of lag terms in a minimum-state {xi } = state vector for subsystem i
unsteady aerodynamic gust force rational {ymeas } = true structural responses at the points
approximation of measurement
{δ} = vector of commands to the actuators
Received 30 July 2001; revision received 10 May 2003; accepted for [φ] = matrix containing (column by column)
publication 11 May 2003. Copyright  c 2003 by Frode Engelsen and Eli reduced-order modal basis n u × n q
Livne. Published by the American Institute of Aeronautics and Astronautics,
[0 ], [1 ], [2 ] = matrices defining [Eq. (17)] the true
Inc., with permission. Copies of this paper may be made for personal or
internal use, on condition that the copier pay the $10.00 per-copy fee to responses on the structure at the points
the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA of response measurement
01923; include the code 0021-8669/04 $10.00 in correspondence with the
CCC.
∗ Associate Technical Fellow, Loads and Dynamics; frode.engelsen@
Subscripts and Superscripts
boeing.com.
† Professor, Department of Aeronautics and Astronautics; eli@aa. A = actuator
washington.edu. Associate Fellow AIAA. CO = control law

335
336 ENGELSEN AND LIVNE

c = control degrees of freedom Formulation


cc = control-control partition Aeroelasticity
of system matrices The partitioned Laplace-transformed equations corresponding to
cs = control-structural coupling structural degrees of freedom of the linearized dynamic aeroelastic
in system matrices equations of motion for an actively controlled deformable airplane,
G = gust filter based on full-order structural dynamic degrees of freedom, are
g = gust  
s = structural degrees of freedom u s (s)
(s 2 [M ss M sc ] + s[C ss C sc ] + [K ss K sc ])
sc = structural-control coupling qc (s)
in system matrices
SE = sensor
 
u s (s)
sg = structural-gust partition of system matrices − q D Sref [A (s)
ss
A (s)]
sc
qc (s)
ss = structural-structural partition
of system matrices
= q D Sref {Asg (s)}[wg (s)/U∞ ] (1)
T = transpose
These equations can represent a free-free airplane in flight (the stiff-
ness matrix [K ss ] in this case will include rigid-body motions and
become singular) or a restrained vehicle, in which case the stiffness
Introduction matrix is nonsingular. General formulations of the free-free prob-
lem will further partition the system matrices into those associated
ANDOM stresses caused by flight in turbulence,1−4 super-
R
Downloaded by BRISTOL UNIVERSITY on March 5, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.9329

with rigid-body motion and those associated with elastic modes of


imposed on stresses as a result of steady maneuvers, must
vibration. However, to simplify the derivations in the following para-
be taken into account in any effort to synthesize a flexible flight
graphs it is assumed here that the stiffness matrix [K ss ] is always
vehicle with or without an active control system.5−8 Because the
nonsingular. This does not limit the generality of the formulation
mathematical models of integrated actively controlled flight ve-
used here significantly because free-free conditions can be always
hicles are large, some form of model order reduction becomes
simulated by connecting the vehicle “numerically” to the ground us-
necessary,9−11 whether frequency domain12 or time domain anal-
ing springs (in a statically determinate manner) that are soft enough
ysis techniques6,7,13,14 are used. For the structural dynamic part of
to separate rigid-body motions from elastic motions.
the model, various structural dynamic order-reduction techniques
Note that the full-order aerodynamic matrices [Ass (s) Asc (s)],
have been widely used over the years, including various mode dis-
{Asg (s)} are not usually available with large-scale finite element
placement (MD), Ritz vectors, modified mode displacement (Ficti-
structural models. They are calculated, using an aerodynamic mesh,
tious mass), mode acceleration (MA), and combined direct-adjoint
for a set of motion shapes in the form of some vibration modes, poly-
methods, to name a few.15−26
nomial functions, or Ritz vectors. In the case of modeling airplane
A number of problems arise when such order-reduction meth-
wings and control surfaces using the equivalent-plate approach,32−36
ods are used for design optimization. First, in order to provide
it is possible to create full-order aerodynamic matrices correspond-
gradient-based optimization algorithms the information they re-
ing to all polynomial Ritz functions used for the different wing
quire, derivatives of response functions (constraints and objectives),
and control surface zones.34,37 This difference has a bearing on the
must be calculated in addition to the analysis values of the response
method developed here. The derivation assumes the availability of
functions themselves. Additionally, it is well known8−11,25,26 that it
full-order aerodynamic matrices, corresponding to the full-order set
is not straightforward, with the widely used displacement-method
of structural degrees of freedom. Adaptation to the case of structural
based models of structural dynamics, to obtain accurate stress in-
finite element models is straightforward and is presented briefly in
formation from reduced-order models (based on finite element or
Appendix A.
any Rayleigh–Ritz discretization). With structural dynamic models
Stresses at any point in the thin-walled structure can be calculated
reduced-order stress design sensitivities are even more challenging
from the deformations using
to obtain.25,26 A significant effort over the last 20 years has been
aimed at formulation, order reduction, sensitivity, and approxima-
 
s x x 
tion techniques that would make it possible to include constraints
s yy = [S]T {u s } (2)
on combined maneuver/gust response stresses in integrated aeroser-  
voelastic optimization in a practical way. The work in this area is sx y
still evolving. The present paper describes a method for obtaining
accurate design-oriented stress and stress sensitivity information where different matrices [S] are used for different points on the
from reduced order linear-time-invariant state-space models of in- structure. The displacements can be approximated by a linear com-
tegrated aeroservoelastic systems, using Lyapunov’s equation for bination of lower-order mode shapes:
calculating covariance matrices of the displacement and stress re- {u s } = [φ]{qs } (3)
sponses. Lyapunov’s equations27 have been used in modern con-
trol to obtain covariance terms of the response to random inputs where each column of the transformation matrix represents a single
for many years. The utilization of Lyapunov’s equation for calcu- mode shape.
lating gust response of aeroservoelastic systems has already been The reduced-order equations of motion will be (after premulti-
explored in a number of studies.5−7,13,28,29 The emphasis, however, plication by [φ]T )
in most of those studies was on displacement and displacement-
rate responses (such as accelerations at points on the structure, mo- [ϕ]T s 2 [M ss ] + s[C ss ] + [K ss ] − q D Sref [Ass (s)] [ϕ]{qs (s)}
tion of control surfaces, and rate of motion of control surfaces).
Here we present a complete and detailed formulation of the stress + [ϕ]T s 2 [M sc ] + s[C sc ] + [K sc ] − q D Sref [Asc (s)] {qc (s)}
gust response problem for aeroservoelastic design synthesis, tai-
lored toward integration with control system design techniques = q D Sref [ϕ]T {Asg (s)}[wg (s)/U∞ ] (4)
based on modern control. It includes order reduction based on an Approximate stresses based on such a MD reduced-order model
adaptation of the MA method, reduced-order analytic sensitivity can, therefore, be calculated from [Eqs. (2) and (3)]
of stress covariances, and efficient approximations to be used in  
a nonlinear programming/approximation concepts30,31 approach to s x x 
design optimization. The paper includes results of thorough con- s yy = [S]T {u s } = [S]T [φ]{qs } (5)
vergence studies for assessment of the effect of order reduction  
on accuracy. sx y
ENGELSEN AND LIVNE 337

However, it is well known15−17 that the accuracy of MD stresses leading to


is poor. Unless special attention is paid to the capability to capture
local effects when vectors for the reduced-order basis are selected,25 s{rs } = [R s ]{rs } + [E s ]s{qs (s)} + [E c ]s{qc (s)}
conventional lower-order natural vibration modes cannot capture
local effects associated with concentrated loading, local stresses, or s{r g } = [R g ]{r g } + {E g }swg (s) (11)
sensitivities of stresses with respect to structural design variables.
These rational function expressions can now be inserted into the
Because improved stress and stress-sensitivity accuracy is ob-
load equations [Eq. (7)]
tained if the displacements used in Eq. (2) are calculated from a
full-order structural model, the MA method is based on rewriting
     
{Fs (s)} = K ϕss {qs (s)} − s 2 M̂ϕss + s Ĉφss + K̂ ϕss {qs (s)}
the full order Eq. (1) as
[K ss ]{u s (s)} = {Fs (s)} (6) − (s 2 [ M̂ sc ] + s[Ĉ sc ] + [ K̂ sc ]){qc (s)} + q D Sref Dϕs {rs (s)}
 g

where the force vector on the right-hand side is obtained using ap- + (q D Sref /U∞ )[D g ]{r g (s)} + (q D Sref /U∞ ) A0
proximate accelerations and velocities [Eq. (3)] obtained from the
MD solution of Eq. (4). Thus,  g

+ A1 s wg (s) (12)
{Fs (s)} = − s 2 [M ss ] + s[C ss ] − q D Sref [Ass (s)] [ϕ]{qs (s)}
where
− s [M ] + s[C ] + [K ] − q D Sref [A (s)] {qc (s)}
2 sc sc sc sc
K φss = [K ss ][φ], M̂φss = [M ss ][φ] − q D Sref Ass

Downloaded by BRISTOL UNIVERSITY on March 5, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.9329

+ q D Sref {Asg (s)}[wg (s)/U∞ ] (7) Ĉφss = [C ss ][φ] − q D Sref Ass



The unsteady aerodynamic forces for the full-order equations (when
the equivalent plate method is used), calculated for simple harmonic K̂ φss = [K ss ][φ] − q D Sref Ass
0φ (13)
motion at discrete reduced frequencies, can be approximated by a
rational function approximation as defined by Roger.38 Because the and
fitted matrices are used in both the load equations (7) and in the MD
equations of motion (4), the Roger fit is performed on the aerody-  M̂ sc  = [M sc ] − q D Sref Asc
2 , [Ĉ sc ] = [C sc ] − q D Sref Asc
1
namics for the load equations. This means that only the columns
of the structural degrees of freedom have been reduced to modal [ K̂ sc ] = [K sc ] − q D Sref Asc
0 (14)
coordinates, but the rows remain full order (see Appendix A for the
case of a finite element model). That is, a rational function approx- The vector of generalized displacements {qs (s)} on the right-hand
imation is first created for the frequency-dependent aerodynamic side of Eq. (12) is the mode displacement solution. The aeroser-
matrices [Ass (s)][φ], [Asc (s)], and [Asg (s)]. voelastic MA method approximates the full-order load distribution
by combining the full-order external forces with approximate in-
N Ls
s ertial and damping forces (both structural and aerodynamic) ob-
[Ass (s)][ϕ] = Ass
0ϕ + A 1ϕ s + A 2ϕ s +
ss ss 2
Ass
i =1
s + cis (i + 2)ϕ tained from the reduced-order MD solution. The coupling control
surface/structure stiffness and damping matrices (corresponding to
N Ls rigid control surface rotation caused by actuator commands) are
s zero by definition, that is, [K sc ] = [0] and [C sc ] = [0]. A possible
[Asc (s)] = Asc
0 + A1 s + A2 s +
sc sc 2
Asc
i =1
s + cis i + 2 definition of the full-order viscous damping matrix in the structural
dynamic equations is discussed in Appendix B.
The MD equations of motion [Eq. (4)] can now be written as
 g
  g
 N Lg
s  g       
{Asg (s)} = A0 + A1 s + g Ai + 2 (8)
s + ci s 2 [ϕ]T M̂ϕss + s[ϕ]T Ĉϕss + [ϕ]T K̂ ϕss {qs (s)}
i =1

It is assumed that the forces associated with structural and control + s 2 [ϕ]T [ M̂ sc ] + s[ϕ]T [Ĉ sc ] + [ϕ]T [ K̂ sc ] {qc (s)}
degrees of freedom have the same aerodynamic lag poles cis , but
g
that the lags associated with gust ci might be different. The gen- − q D Sref [ϕ]T Dϕs {rs (s)} − (q D Sref /U∞ )[ϕ]T [D g ]{r g (s)}
eralized aerodynamic matrices to be used in the MD equations of
motion [Eq. (4)] are obtained from the Roger matrices of Eq. (8) by g
= (q D Sref /U∞ ) [ϕ]T {A0 } + [ϕ]T {A1 }s wg (s)
g
(15)
premultiplying those matrices by [φ]T .
Alternatively, a minimum-state rational function approximation To obtain a desired frequency content of the gust excitation velocity,
(MS) can be used to fit the frequency dependent aerodynamic it is modeled as the output of a linear filter subjected to white noise
matrices39 : input {w(s)}.
   
[Ass (s)ϕ Asc (s)] = Ass
0ϕ Asc
0 + Ass
1ϕ Asc
1 s s{x g (s)} = [A g ]{x g (s)} + {Bg }w(s), wg (s) = C g {x g (s)}
s −1 (16)
+ Ass
2ϕ 2 s + s Dϕ (s[I ] − [R ]) [E
Asc 2 s s
E c]
   
{Asg (s)} = A0 + A1 s + s[D g ](s[I ] − [R g ])−1 {E g } Measurement, Actuation, and Control
g g
(9)
The measured responses on the structure can be displacements,
The matrices [Aiφ ] and [Dφ ] are then premultiplied by [φ]T for use velocities, or accelerations. Using Eq. (3), the “true” responses at
in the MD equations of motion [Eq. (4)]. The form of the MS fit the points where they are measured can be obtained from
[Eq. (9)] can also accommodate the Roger fit [Eq. (8)] and is used
in the following developments. {ymeas (s)} = [0 ] + [1 ]s + [2 ]s 2 {u s (s)}
Aerodynamic lag states are now introduced:
  = [0 ][φ] + [1 ][φ]s + [2 ][φ]s 2 {qs (s)} (17)
−1
qs (s)
{rs (s)} = (s[I ] − [R ]) [E s s c
E ]s
qc (s) Note that the matrices [i ] determine contribution of different de-
grees of freedom to the actual response, whereas the matrix [φ] is
{r g (s)} = (s[I ] − [R g ])−1 {E g }swg (s) (10) the matrix containing mode shape vectors.
338 ENGELSEN AND LIVNE

The measurement signals available to the control system are the When Eq. (22) is now used, we get
outputs of sensors modeled by    
 qc (s)  CA
s{xSE (s)} = [ASE ]{xSE (s)} + [BSE ]{ymeas (s)} sqc (s) =  C A A A  {x A (s)}
 2 
s qc (s) CA AA AA
{ySE (s)} = [CSE ]{xSE (s)} (18)
 
The actuator dynamics, from actuator command to control-surface 0
rotation, assuming irreversible controls, are + C A B A CCO  {xCO (s)}
s{x A (s)} = [A A ]{x A (s)} + [B A ]{δ(s)} C A A A B A CCO + C A B A CCO ACO
 
{qc (s)} = [C A ]{x A (s)} (19) 0
+ C A B A DCO CSE 
The control law (actuator commands caused by sensor measure-
C A A A B A DCO CSE + C A B A CCO BCO CSE + C A B A DCO CSE ASE
ments) is
 
s{xCO (s)} = [ACO ]{xCO (s)} + [BCO ]{ySE (s)} 0
× {xSE (s)} +  0 
{δ(s)} = [CCO ]{xCO (s)} + [DCO ]{ySE (s)} (20)
C A B A DCO CSE BSE
Note that the state-space models for the gust filter, sensors, and
× [0 φ] + [1 φ]s + [2 φ]s 2 {qs (s)}
Downloaded by BRISTOL UNIVERSITY on March 5, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.9329

actuators are all strictly proper. The control law model allows proper (26)
transfer functions, that is, in the control block of equations [Eq. (20)] Equations (21) and (26) can now be used in the load equation,
the D matrix can be nonzero. Eq. (12).
The vector of control surface motions {qc } is available from  
Eq. (19). Examination of Eqs. (12) and (15) shows that expres- {Fs (s)} = − M̂ϕss + M̂ sc C A B A DCO CSE BSE 2 ϕ s 2 {qs (s)}
sions for the rates swg , s{qc }, and s 2 {qc } are also required. When
the time derivative of the output in Eq. (16) is taken, + K φss − K̂ φss + M̂ sc C A B A DCO CSE BSE 0 ϕ {qs (s)}
     
wg (s) Cg 0
= {x g (s)} + w(s) (21) − Ĉφss + M̂ sc C A B A DCO CSE BSE 1 ϕ s{qs (s)}
swg (s) Cg Ag C g Bg

From Eqs. (17) and (18), − K̂ sc C A + Ĉ sc C A A A + M̂ sc C A A A A A {x A (s)}


   
ySE (s) CSE
= {xSE (s)} − Ĉ sc C A B A DCO CSE + M̂ sc C A A A B A DCO CSE
sySE (s) CSE ASE
  + M̂ sc C A B A CCO BCO CSE + M̂ sc C A B A DCO CSE ASE
0
+ [0 φ] + [1 φ]s + [2 φ]s {qs (s)}
2
(22)
CSE BSE × {xSE (s)} − Ĉ sc C A B A CCO + M̂ sc C A A A B A CCO
With the derivative of the output in Eq. (19), we get.
    g
 qc (s)  CA + M̂ sc C A B A CCO ACO {xCO (s)} + (q D Sref /U∞ ) A0 C g
sqc (s) =  C A A A  {x A (s)}
 2  g
+ A1 C g A g {x g (s)} + q D Sref Dφs {rs }
s qc (s) CA AA AA
   g

0 0   + (q D Sref /U∞ )[D g ]{r g } + (q D Sref /U∞ ) A1 C g Bg w(s) (27)
δ(s)
+  C A BA 0  (23) Equations (21) and (26) can also be substituted into the MD equa-
sδ(s) tions of motion, Eq. (15), to yield
C A A A BA C A BA  
And from Eq. (20) for the output of the control law equations, [ϕ]T M̂ϕss + M̂ sc C A B A DCO CSE BSE 2 ϕ s 2 {qs (s)}
      
δ(s) CCO DCO 0 ySE (s) = −[ϕ]T K̂ ϕss + M̂ sc C A B A DCO CSE BSE 0 ϕ {qs (s)}
= {xCO (s)} +
sδ(s) CCO ACO CCO BCO DCO sySE (s)
− [ϕ]T Ĉϕss + M̂ sc C A B A DCO CSE BSE 1 ϕ s{qs (s)}
(24)
Combining Eqs. (23) and (24), − [ϕ]T K̂ sc C A + Ĉ sc C A A A + M̂ sc C A A A A A {x A (s)}
   
 qc (s)  CA
− [ϕ]T Ĉ sc C A B A DCO CSE + M̂ sc C A A A B A DCO CSE
sqc (s) =  C A A A  {x A (s)}
 2 
s qc (s) CA AA AA + M̂ sc C A B A CCO BCO CSE + M̂ sc C A B A DCO CSE ASE
 
0 × {xSE (s)} − [ϕ]T Ĉ sc C A B A CCO + M̂ sc C A A A B A CCO
+  C A A C CO
B  {xCO (s)}
C A A A B A CCO + C A B A CCO ACO + M̂ sc C A B A CCO ACO {xCO (s)} + (q D Sref /U∞ )[ϕ]T
g
A0 C g
 
0 0   g
+ A1 C g A g {x g (s)} + q D Sref [ϕ]T Dϕs {rs }
+ C A B A DCO 0  ySE (s)
sySE (s) + (q D Sref /U∞ )[ϕ]T [D g ]{r g }
C A A A B A DCO + C A B A CCO BCO C A B A DCO
 g

(25) + (q D Sref /U∞ )[ϕ]T A1 C g Bg w(s) (28)
ENGELSEN AND LIVNE 339

g g
In the following steps the output equations [of Eqs. (18) and (20)] [V̄26 ] = (q D Sref /U∞ ) A0 C g + A1 C g A g
are inserted into the state-space equation (19). This leads to
[V̄27 ] = q D Sref Dφs , [V̄28 ] = (q D Sref /U∞ )[D g ]
s{x A (s)} = [A A ]{x A (s)}+[B A CCO ]{xCO (s)}+[B A DCO CSE ]{xSE (s)}
 g

(29) {W̄2 } = (q D Sref /U∞ ) A1 C g Bg (37)
Equation (17) is inserted into the state-space equation (18):
Equation (36) can be written in a more compact form as
−[BSE 2 φ]s 2 {qs (s)} + s{xSE (s)} = [BSE 0 φ]{qs (s)}
{Fs (s)} = −[Ū22 ]s{x2 (s)} + [V̄2 ]{x(s)} + {W̄2 }w(s) (38)
+ [BSE 1 φ]s{qs (s)} + [ASE ]{xSE (s)} (30)
where
Equations (18) and (20) are now used to obtain
[V̄2 ] = K φss + V̄21 V̄22 V̄23 V̄24 V̄25 V̄26 V̄27 V̄28 (39)
s{xCO (s)} = [ACO ]{xCO (s)} + [BCO CSE ]{xSE (s)} (31)

And the gust filter Eq. (16) Now, the MD aeroservoelastic equations of motion, Eq. (28), can
be written in a similar form:
s{x g (s)} = [A g ]{x g (s)} + {Bg }w(s) (32) [U22 ]s{x2 (s)} = [V21 ]{x1 (s)} + [V22 ]{x2 (s)} + [V23 ]{x A (s)}
Finally the state-space equations for the aerodynamic states are ob-
+ [V24 ]{xSE (s)} + [V25 ]{xCO (s)} + [V26 ]{x g (s)} + [V27 ]{rs }
tained from Eqs. (11) and (26):
+ [V28 ]{r g } + {W2 }w(s)
Downloaded by BRISTOL UNIVERSITY on March 5, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.9329

s{rs } = [E s ]s{qs (s)} + E c C A A A {x A (s)} + E c C A B A DCO CSE (40)


where the equation of motion corresponding to structural degrees
× {xSE (s)} + E c C A B A CCO {xCO (s)} + [R s ]{rs } (33) of freedom is reduced in order by premultiplication by [φ]T . Define
and the equations for the added gust aerodynamic states are obtained [U22 ] = [φ]T Ū22 
from Eqs. (11) and (21):
  [V2 j ] = [φ]T [V̄2 j ] j = 1, . . . , 8
s{r g } = E g C g A g {x g (s)} + [R g ]{r g } + E g C g Bg w(s) (34)
{W2 } = [φ]T {W̄2 } (41)
Complete Aeroservoelastic System
The state-space equations for the complete closed-loop system and the preceding equations (29–34) can be collected to create cou-
can now be assembled, after defining the system’s state-space pled aeroservoelastic equations of motion in the form:
vector:
s[U ]{x(s)} = [V ]{x(s)} + {W }w(s) (42)
{x}T =
  Examination of the matrices [U ] and [V ] reveals that they are of the
{x1 }T {x2 }T {x A }T {xSE }T {xCO }T {x g }T {rs }T {r g }T form:
(35a)
 
where I 0 0 0 0 0 0 0
0 U22 0 0 0 0 0 0
{x1 } = {qs }  
⇒ s{x1 } = {x2 } (35b) 0 0
{x2 } = s{qs }  0 I 0 0 0 0 
0 
 U42 0 I 0 0 0 0
In an effort to adopt a more concise notation, the load equation can [U ] =   (43)
0 0 0 0 I 0 0 0
now be written in the form:  
0 0 0 0 0 I 0 0
{Fs (s)} = −[Ū22 ]s{x2 (s)} + K φss + [V̄21 ] {x1 (s)}  
0 0 0 0 0 0 I 0
+ [V̄22 ]{x2 (s)} + [V̄23 ]{x A (s)} + [V̄24 ]{xSE (s)} 0 0 0 0 0 0 0 I
 
+ [V̄25 ]{xCO (s)} + [V̄26 ]{x g (s)} + [V̄27 ]{rs } 0 I 0 0 0 0 0 0
V21 V22 V23 V24 V25 V26 V27 V28 
+ [V̄28 ]{r g } + {W̄2 }w(s)  
(36) 0 0 
 0 V33 V34 V35 0 0 
where V 
   41 V42 0 V44 0 0 0 0 
[V ] =   (44)
[Ū22 ] = M̂φss + M̂ sc C A B A DCO CSE BSE 2 φ 0 0 0 V54 V55 0 0 0 
 
0 0 0 0 0 V66 0 0 
[V̄21 ] = − K̂ φss + M̂ sc C A B A DCO CSE BSE 0 φ  
0 V72 V73 V74 V75 0 V77 0 
[V̄22 ] = − Ĉφss + M̂ sc C A B A DCO CSE BSE 1 φ 0 0 0 0 0 V86 0 V88

       
{W }T = 0 W2T 0 0 0 W6T 0 W8T (45)
[V̄23 ] = − K̂ sc C A + Ĉ sc C A A A + M̂ sc C A A A A A
where
[V̄24 ] = − Ĉ sc C A B A DCO CSE + M̂ sc C A A A B A DCO CSE
[V33 ] = [A A ], [V34 ] = [B A DCO CSE ], [V35 ] = [B A CCO ]
+ M̂ sc C A B A CCO BCO CSE + M̂ sc C A B A DCO CSE ASE
[U42 ] = −[BSE 2 φ], [V41 ] = [BSE 0 φ]
[V̄25 ] = − Ĉ C A B A CCO + M̂ C A A A B A CCO
sc sc
[V42 ] = [BSE 1 φ], [V44 ] = [ASE ]

+ M̂ sc C A B A CCO ACO [V54 ] = [BCO CSE ], [V55 ] = [ACO ]


340 ENGELSEN AND LIVNE

[V66 ] = A g , {W6 } = {Bg } where the stresses and the force vector are dynamic. The covariance
matrix of the three stress components is
[V72 ] = [E ], s
[V73 ] = [E C A A A ]
c
   T 
 s x x  s x x 
[V74 ] = [E C A B A DCO CSE ],
c
[V75 ] = [E C A B A CCO ]
c
[Covs ] = E  s yy s yy 
  
[V77 ] = [R s ] sx y sx y

[V86 ] = E g C g A g , [V88 ] = [R g ] = [η]T E {Fs }{Fs }T [η] = [η]T [][η] (57)


 
{W8 } = E g C g Bg (46) Gust Filter—Modeling Considerations
Examination of the expressions for the load vector [Eqs. (38) and
The state-space equations of motion can also be written as (51)] reveals a direct dependency on the white noise input. Thus, the
covariance matrix of the full-order forces is only finite if {FL } = {0}.
s{x(s)} = [ Ã]{x(s)} + { F̃}w(s) (47) Inspection of equations (37) and (52) reveals that this is the case
 g

where {W̄2 } = (q D Sref /U∞ ) A1 C g Bg = {0} (58)

[ Ã] = [U ]−1 [V ], { F̃} = [U ]−1 {W } (48) g


Thus, we must require that either {A1 } = {0} or C g Bg = 0. In the
g
rational function approximation of the gust vector, {A1 } = {0} can
Downloaded by BRISTOL UNIVERSITY on March 5, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.9329

The second row partition of these equations [corresponding to be utilized without much loss in accuracy if one of the lag poles
Eq. (40) and the partitions in Eqs. (43–46)] gives an expression is large compared to the frequency range of interest. In this case
for s 2 {qs }: g g
s/(s + ci ) ≈ s/ci , that is, proportional to s, and the matrix corre-
g
sponding to this lag term will model the effect of the unused {A1 }.
s 2 {qs } = s{x2 } = [U22 ]−1 [V2 ]{x(s)} + [U22 ]−1 {W2 }w(s) (49) Alternatively, the matrix product C g Bg = 0 for the gust filter trans-
fer function if the order of the denominator is greater than the order
where of its numerator by at least two. In most common approximations
of gust filter transfer functions, there is only a first-order difference
[V2 ] = [V21 V22 V23 V24 V25 V26 V27 V28 ] (50) between the denominator and the numerator. This can be overcome
by either adding a low-pass filter to the gust filter13 or by converting
This means that the summation of forces for the right-hand side of the filter from the form:
Eq. (6) can be written in terms of the state-space vector [Eqs. (38)
and (49)]: wg cn − 1 s n − 1 + · · · + c1 s + c0
= n (59)
w s + dn − 1 s n − 1 + · · · + d1 s + d0
{Fs (s)} = [A L ]{x(s)} + {FL }w(s) (51)
to
where
wg cn − 1 s n − 1 + · · · + c1 s + c0
= n+1 (60)
[A L ] = [V̄2 ] − [Ū22 ][U22 ]−1 [V2 ] w εs + s n + dn − 1 s n − 1 + · · · + d1 s + d0

{FL } = {W̄2 } − [Ū22 ][U22 ]−1 {W2 } (52) by adding a (n + 1)th-order term with a small positive coefficient
ε. This introduces an additional gust filter state in the model with a
real negative lag pole close to −1/ε.
Stress Retrieval
Figure 1 shows a comparison between the exact von Kármán and
Having found a solution to the coupled aeroservoelastic problem the two third-order (n = 3) rational approximations determined by
(with MD-order reduction for the structural degrees of freedom) Eqs. (59) and (60) for a typical flight condition. The exact spectrum
Eq. (47), we can use Eq. (6) to retrieve stresses, using the full-order rolls off with a slope of − 53 , while the standard rational model40 will
stiffness matrix. In a displacement-based finite element or Rayleigh– have an integer slope of −2. The new rational model will be identical
Ritz method, stresses at a point (on a plane stress element) can be to the standard rational model up to a frequency of 1/ε and will then
calculated from
 
sx x (s)
s yy (s) = [S]T {u s (s)} = [S]T [K ss ]−1 {Fs (s)} (53)
 
sx y (s)

An adjoint matrix is defined as follows:

[η]T = [S]T [K ss ]−1 (54)

Because of the symmetry of the stiffness matrix, this leads to

[K ss ][η] = [S] (55)

The matrix [S] is static and is different for every point at which
stresses are calculated. A corresponding adjoint matrix [η] at each
stress recovery point can be found from Eq. (55). Dynamic stresses
can now be calculated using
 
sx x (s)
s yy (s) = [η]T {Fs (s)} (56)
  Fig. 1 Comparison of vertical spectra for the von Kármán and two
sx y (s) rational approximations.
ENGELSEN AND LIVNE 341

roll off with a slope of −4 above this frequency. The coefficient Then, from Eqs. (70) and (71):
ε must therefore be chosen small enough such that 1/ε is higher      
∂ AL ∂ V̄2 ∂ Ū22
than the highest frequency of interest. In the following example = − [U22 ]−1 [V2 ] + [Ū22 ][U22 ]−1
ε = 0.0001, which corresponds to a frequency of 1000 rad/s. ∂p ∂p ∂p
In the following it is assumed that  
∂[U22 ] ∂ V2
{FL } = {0} × [U22 ]−1 [V2 ] − [Ū22 ][U22 ]−1 (72)
(61) ∂p ∂p
and Eq. (51) becomes The sensitivity of the aeroservoelastic modally reduced state covari-
ance matrix [∂ X /∂ p] can be determined by differentiating Eq. (65)
{Fs (s)} = [A L ]{x(s)} (62) with respect to p:
       T
∂X ∂X ∂ F̃ ∂ F̃
Covariance Matrices of Full-Order Loads and Aeroservoelastic [ Ã] + [ Ã]T = − Q{ F̃}T − { F̃}Q
∂p ∂p ∂p ∂p
Mode-Displacement States
With Eq. (62) the covariance matrix of the full-order load elements    T
become ∂ Ã ∂ Ã
− [X ] − [X ] (73)
∂p ∂p
[] = E {Fs }{Fs } T
= [A L ]E({x}{x} )[A L ] = [A L ][X ][A L ]
T T T
This is also a Lyapunov equation where the right-hand side is known.
Following the derivations in Eq. (71),
(63)
     
Downloaded by BRISTOL UNIVERSITY on March 5, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.9329

∂ Ã ∂V ∂U
where = [U ]−1 + [U ]−1 [U ]−1 [V ] (74)
∂p ∂p ∂p
[X ] = E({x}{x}T ) (64) Note that a number of matrices whose derivatives are required in the
preceding equations depend on the set of modal vectors [φ] [Eqs. (3),
is the covariance of the states-space degrees of freedom found
(37), (41), and (46)]. Differentiation can be carried out assuming a
from the Lyapunov’s equation27 for the MD reduced-order coupled
fixed-mode41 approach or a variable-mode approach. In the latter,
aeroservoelastic system:
variations of design variables lead to variations of the modal matrix
used.
[ Ã][X ] + [X ][ Ã]T = −{ F̃}Q{ F̃}T (65)
Application
The intensity of the white noise input w is assumed to be Q.
The new methodology just derived has been implemented in
E[w(t)w(τ )T ] = Qδ(t − τ ) (66) an efficient, integrated aeroservoelastic design optimization capa-
bility, the Lifting Surface Augmented Structural Synthesis code
where δ(t) is the Dirac delta function. (LS-CLASS)6,7,37 and applied to a typical large, flexible commer-
cial transport configuration. The structural model is based on an
Behavior Sensitivities equivalent-plate formulation32−36 for wing, control surfaces, and
tail surfaces. Transverse shear effects are included. Horizontal-
The sensitivity of gust stresses to changes in the design variables and vertical-plate segments are used for modeling of the com-
can be determined from the sensitivity of [Covs ]. Equation (57) is plete configuration, with a fuselage modeled as a linked chain of
differentiated with respect to a design variable p: beam/narrow-plate segments. The equivalent-plate approach, al-
   T     though not as general and accurate as the finite element (FE) method
∂Covs ∂η ∂ ∂η for the modeling of real airplanes, has nevertheless been found to
= [][η]+[η]T [η]+[η]T [] (67)
∂p ∂p ∂p ∂p be remarkably accurate—at least for research/conceptual design
purposes—given its simplicity, ease of modeling, and special suit-
where ability for aeroelastic analysis. Its application to high-aspect-ratio
       T and low-aspect-ratio wings has been studied thoroughly. With just a
∂ ∂ AL ∂X ∂ AL small number of degrees of freedom (of the order of 200–400), com-
= [X ][A L ]T + [A L ] [A L ]T + [A L ][X ] plete quite complex configurations can be modeled using equivalent
∂p ∂p ∂p ∂p
plates, where FE models will require tens of thousands degrees of
(68) degrees of freedom. For the studies reported here, the doublet-lattice
unsteady aerodynamics module of the Elfini42 code was used to cre-
The sensitivity matrix [∂η/∂ p] can be determined by taking the ate full-order aerodynamic matrices for the Ritz functions used in
derivative of Eq. (55) with respect to the design variable p. LS-CLASS. These matrices were then imported into LS-CLASS and
    manipulated there. The equivalent plate model is shown in Fig. 2.
∂η ∂ K ss The model consists of equivalent-plate segments (zones) joined via
[K ss ] =− [η] (69) lumped springs, which model attachment and actuator stiffnesses.
∂p ∂p
Spars, stringers, ribs, and skins are included in the model. The aero-
Furthermore, from Eq. (52) dynamic mesh is presented in Fig. 3. The wings, engines, horizontal
      stabilizers, and vertical fin are incorporated into the unsteady cal-
∂ AL ∂ V̄2 ∂ Ū22 culations. The fuselage aerodynamic distribution is introduced by
= − [U22 ]−1 [V2 ] scaling rigid airplane empirical distribution based on the local de-
∂p ∂p ∂p
formation along the fuselage centerline. There are no interference
  effects between the fuselage and the remaining aerodynamic sur-
∂[U22 ]−1 ∂ V2
− [Ū22 ] [V2 ] − [Ū22 ][U22 ]−1 (70) faces. A block diagram of an active control system for symmetric
∂p ∂p motion (motion in the pitch plane) is shown in Fig. 4. The inte-
grated aeroservoelastic model of the passenger jet configuration was
We can differentiate [U22 ]−1 [U22 ] = [I ] to get checked against results obtained by standard industry codes, and the
accuracy of the model used here was found to be good in terms of
∂[U22 ]−1 ∂[U22 ]
= −[U22 ]−1 [U22 ]−1 (71) natural frequencies and mode shapes, deformation and local internal
∂p ∂p loads in maneuvers, as well as aeroservoelastic stability.
342 ENGELSEN AND LIVNE

Results of Convergence Studies


The stress covariance matrix and its sensitivity to a spar-cap-
area design variable on the outboard wing have been calculated for
a stress recovery point on the skin of the inboard wing. Figure 5
shows a comparison of the convergence characteristics between the
MA method and the MD method for one of the covariance terms.
These results are typical for all the matrix terms.
For the covariance matrix term the two methods will give the
same results for a full-order model (165 modes; the model has 165
Ritz degrees of freedom, but a similarity transformation has been
made to modal coordinates, i.e., no modal truncation). However, the
MA method converges with 30–50 modes, whereas the MD needs at
least 90 modes to converge. For the sensitivity term the MA method
will converge with about 30–40 modes, whereas the MD will need
about 120 modes for convergence.
The MA and MD methods do not converge to the same value
for the sensitivity terms. This is actually an artifact of ignoring the
eigenvector sensitivities in the analytical sensitivity calculations.
Because for simplicity and better computational speed the MA sen-
sitivity equations in the present formulation were based on a fixed-
Downloaded by BRISTOL UNIVERSITY on March 5, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.9329

mode approach, finite differences had to be used to clarify this issue.


A finite difference calculation of the sensitivities with a fixed modal
base (FIXMOD) verifies that the two methods will yield different
Fig. 2 Equivalent-plate structural model of a passenger airplane con-
figuration.

Fig. 3 Doublet-lattice aerodynamic mesh for the passenger airplane Fig. 5 Comparison of convergence of the MA method and the MD
configuration. method for a stress covariance term and its sensitivity.

Fig. 4 Pitch control system definition.


ENGELSEN AND LIVNE 343

results. Finite difference calculations with a variable (VARMOD)


modal base will give the same results. It is also interesting that
the finite difference calculations for the FIXMOD and VARMOD
are practically identical for the MA method. The mode-acceleration
results, then, converge to the correct value of sensitivity, and the
effect of ignoring the eigenvector sensitivities in the MA analytical
sensitivity calculations is basically negligible.

Taylor-Series Approximations
The key to the success of nonlinear programming in solving op-
timization problems is the use of approximation concepts. In each
stage of the optimization process, a detailed analysis and the as-
sociated behavior sensitivity analysis are used for constructing ap-
proximations of the objective and constraint functions in terms of
the design variables. The most common approximations are direct
or reciprocal Taylor-series approximations.31 These are local, linear
approximations based on the Taylor series and vary with the design
variable or the inverse of the design variable, respectively.
Taylor-series approximations are checked, using parametric stud-
ies, for a nominal flight case, with ample damping. Figure 6 shows
Downloaded by BRISTOL UNIVERSITY on March 5, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.9329

the results when the design variable is a spar cap area of the out-
board wing. Parametric studies for a closed-loop system, with a
control gain as the design variable, are shown in Fig. 7. A direct lin-
ear Taylor-series approximation appears to capture accurately the
variations of the stress covariance matrix terms with changes in the
design variables in most cases. The fact that the linear approximation

Fig. 8 Variation of covariance stresses with a spar-cap-area design


variable (flight case with lightly damped mode approaching instability).

is tangent to the exact variation curve at the baseline configuration


[design variable ratio = 1.0] indicates that the analytical sensitivity
calculations are correct.
However, in airplane configurations with lightly damped struc-
tural modes, when in the course of design variable variation certain
poles move toward instability, significant nonlinear behavior of the
response might result. Figure 8 shows a flight case with a lightly
damped mode close to instability. As the spar-cap-area design vari-
able is reduced, the damping approaches zero. This results in a rise
in the covariance matrix terms. It is apparent that a linear approx-
imation cannot capture this explosive rise. A new approximation
must be derived to capture these effects.

Fig. 6 Variation of covariance stresses with changes in a spar-cap-area New Approximation


design variable. In the new approximation presented here, a Taylor-series approxi-
mation for the stress covariance matrix at a point (linear in the deriva-
tion included here) is augmented by a correction term as follows:
Nk   
∂Covs
[Covs ] = [Covs ]o + ( pk − pko ) + [ Covs ({ p})]
k =1
∂ pk o

(75)

The summation is taken over Nk design variables, and the o denotes


the reference configuration for which analysis and sensitivity anal-
ysis are carried out. The linear portion of the approximation can of
course be replaced by a reciprocal approximation. To help guide
the derivation of the correction term [ Covs ({ p})], we search for
insight into the nature of the dependence of covariance terms on
design variables. How does a pole approaching instability lead to
explosive growth of the resulting response?
The stress covariance matrix [Eqs. (57) and (63–65)] can in gen-
eral be calculated from the following diagonalization of the system
matrix equations.29 The MD state covariance matrix can be obtained
from
Fig. 7 Variation of covariance stresses with changes in a pitch control
system gain design variable. [X ] = [ ][Y ][ ]T (76)
344 ENGELSEN AND LIVNE

where the i j term of the matrix [Y ] is given by away from this base point. Only the eigenvalues are allowed to vary,
and they are also evaluated using an approximation. In the case
−1 −T
(−[ ] {F}Q{F} [ ] T
)i j presented here, Rayleigh quotient approximations are used for the
Yi j = (77)
λi + λ j aeroservoelastic eigenvalues.

The covariance matrix for stresses at a point is then {θi }oT [V ({ p})]{ψi }o
λi ({p}) = (87)
{θi }oT [U ({ p})]{ψi }o
[Covs ] = [η]T [A L ][X ][A L ]T [η] (78)
where
The matrix of eigenvectors is [ ] = [ψ1 ψ2 · · · ψ N ], where the  
eigenvectors {ψi } satisfy ∂U
[U ({p})] = [U ]o + ( pk − pko )
∂ pk
[U ]{ψi }λi = [V ]{ψi } (79) k o

 
Or, alternatively ∂V
[V ({p})] = [V ]o + ( pk − pko ) (88)
∂ pk
{ψi }λi = [U ]−1 [V ]{ψi } = [ Ã]{ψi } (80) k o

and {θ}, {ψ} are left and right eigenvectors, respectively, of the
If we define
generalized eigenvalue problem [Eqs. (42) and (79)]. The correction
[]T = [ ]−1 (81) term (beyond Taylor series) of the [X ] matrix is now
Downloaded by BRISTOL UNIVERSITY on March 5, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.9329

where [] = [ξ1 ξ2 · · · ξ N ], then, [ X ({p})] = {ψi }o Yi j ({p}){ψ j }oT (89)


i j
−{ξi }T {F}Q{F}T {ξ j }
Yi j = (82) and, finally, the correction term for the stress covariance matrix is
λi + λ j

Because of the presence of aeroservoelastic eigenvalues in the de- [ Covs ({p})] = [η]oT [A L ]o [ X ({p})][A L ]oT [η]o (90)
nominator, the covariance of the response shoots to infinity when the
The derivatives of terms of [Y ] with respect to the aeroservoelastic
damping of any the aeroservoelastic modes approaches zero. This
eigenvalues [Eqs. (84)] are obtained by differentiating Eq. (86), and
physical phenomenon is well known as flutter and is characterized
evaluating the derivative at the reference point
by large response amplitudes often resulting in structural failure.
It is assumed that the explosive behavior near instability is solely    
∂Yi j ∂Yi j {ξi }oT {F}o Q{F}oT {ξ j }o
caused by the reduction of damping on some modes. The [Y ] ma- = = (91)
trix, then, depends on design variables through the dependence of ∂λi o
∂λ j o
(λi,o + λ j,o )2
aeroservoelastic eigenvalues λi on these design variables. Using a
Note that the correction term vanishes when the modal damping is
reference (baseline) design, the variation of [Y ] in the neighborhood
large. The correction term has significant contribution only from
of this design is described as the sum of a Taylor series in the design
lightly damped modes.
variables plus a correction term Yi j to account for whatever the
Approximate results (compared to “exact” parametric results) us-
Taylor series cannot capture.
! ing the new stress covariance approximation are shown in Fig. 9. The
  method, indeed, captures the explosive rise in the covariance matrix
∂Yi j ∂λi
Yi j ({p}) = (Yi j )o + terms, but because of inaccuracy of the eigenvalue approximation
k
∂λi o
∂ pk o used this rise (based on the approximation) happens at a smaller de-
sign variable perturbation than in the exact case. It is not surprising,
  " given the dependency of [Y ] on the eigenvalues λ, that the accuracy
∂Yi j ∂λ j
+ ( pk − pk,o ) + Yi j ({p}) (83) of the eigenvalue approximation used is extremely important.
∂λ j o
∂ pk o In the approximate results presented so far, Rayleigh quotient
eigenvalue approximations43,44 were based on a fixed-mode ap-
In Eq. (83) {p} is the vector of design variables, and pk is the kth proach. Changes in all eigenvectors were approximated. An im-
design variable. The correction factor is, thus, proved approximation can be created if a mixed fixed/variable mode
!   approach (where the right eigenvectors are allowed to vary but the
∂Yi j ∂λi
Yi j ({p}) = Yi j ({p}) − (Yi j )o − left eigenvectors are kept fixed) is used for the Rayleigh quotient
k
∂λi o
∂ pk o approximations (RQA).

  " {θi }oT [V ({p})]{ψi ({p})}


∂Yi j ∂λ j λi ({ p}) = (92)
+ ( pk − pk,o ) (84) {θi }oT [U ({p})]{ψi ({p})}
∂λ j o
∂ pk o
where
The nominal [Y ] matrix at the reference (base) point is (using the  
values of aeroservoelastic poles and eigenvectors at the base point) ∂ψi
{ψi ({p})} = {ψi }o + ( pk − pko ) (93)
∂ pk
−{ξi }oT {F}o Q{F}oT {ξ j }o k o
(Yi j )o = (85)
λi,o + λ j,o The eigenvector sensitivity can be calculated mode by mode31,42 by
a number of methods. Taking the derivative of Eq. (79) with respect
The [Y ] matrix based on reference eigenvectors but changing eigen- to design value pk results in
values (because of changes in the design variable vector {p}) is given       
by ∂ψi ∂V ∂U ∂λi
([U ]λi − [V ]) = − λi − [U ] {ψi }
−{ξi }oT {F}o Q{F}oT {ξ j }o ∂p ∂ pk ∂ pk ∂ pk
Yi j ({p}) = (86)
λi ({p}) + λ j ({p}) (94)

Equation (85) and the numerator of Eq. (86) are evaluated at the All terms here are evaluated at the reference configuration. The
base (reference) point once and are fixed for the approximations right-hand side of the equation is known. The matrix on the left side
ENGELSEN AND LIVNE 345
Downloaded by BRISTOL UNIVERSITY on March 5, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.9329

Fig. 11 New covariance approximation with standard RQA eigen-


Fig. 9 New covariance approximation with standard RQA eigenvalue value approximation for a control system gain design variable (flight
approximation for a spar-cap-area design variable (flight case with case with a lightly damped mode approaching instability).
lightly damped mode approaching instability).

is singular, but the normalization criterion for the eigenvector can be


used to eliminate this problem. When the eigenvector is normalized
so that, for example, the mth term is constant, the mth term in the
eigenvector sensitivity vector is zero. This criterion can be used to
solve the equations to obtain the eigenvector sensitivity {∂ψi /∂ p}o
for Eq. (93).
Clearly, the calculation of eigenvector sensitivities of all modes
with respect to all design variables adds considerable computational
cost to the preparation of the approximations. However, because
only very low damping in modes contributes significantly to the
correction term just derived we can limit the number of eigenvalues
for which a mixed fixed/variable mode RQA is calculated to only
those with damping below some very low threshold value.
Figure 10 shows significant improvement in stress covariance
response approximation when the improved eigenvalue approxima-
tions are used for the low damped modes. Figure 11 shows results
from the new approximation for a flight case where an increase in
control system gain makes the damping of the short period pitch
mode unstable. The damping on the short period pitch mode is also
shown in the figure. A mixed fixed/variable mode RQA is used for
the short period mode in the new stress covariance approximation,
and it is apparent that the new approximations capture the rise of
the response well.
Results of additional studies and utilization of the new covariance
matrix approximations for evaluating stress failure constraints for
the integrated aeroservoelastic vehicle are described in Refs. 45
and 46.

Conclusions
A complete detailed formulation of the integrated aeroservoelas-
tic gust stress response problem has been presented. The formulation
is design oriented. That is, it introduces reduced-order modeling (in
the form of a mode-acceleration method for random stresses), an-
Fig. 10 New covariance approximation with new RQA eigenvalue ap- alytic sensitivities, and fast approximate analysis techniques. The
proximation for spar-cap-area design variable (flight case with lightly new approximations are efficient and accurate. They are based on in-
damped mode approaching instability). sight regarding the mathematical and physical nature of the effects
346 ENGELSEN AND LIVNE

of design variable changes (structural or control) on the random The diagonal damping matrix can be defined as
stress response. Their importance is in the context of a nonlinear
programming/approximation concepts multidisciplinary design op- [φ]T [C ss ][φ] = [2ξi ωi ]([φ]T [M ss ][φ]) (B2)
timization strategy.
The paper contains results of convergence studies to evaluate the where [2ξi ωi ] is a diagonal matrix with 2ξi ωi along the diagonal.
new mode-acceleration adaptation. Convergence of both analysis With this expression
and sensitivity solutions is examined. Performance of the new ap- [C ss ][φ] = [M ss ][φ]([φ]T [M ss ][φ])−1 ([φ]T [C ss ][φ])
proximation is assessed using variation of both structural and control
design variables. An integrated aeroservoelastic model of a passen- = [M ss ][φ][2ξi ωi ] (B3)
ger airplane configuration is used. The results presented here, then,
apply to real configurations and realistic aerospace vehicle multi See also Ref. 47.
disciplinary optimization design challenges. The sensitivity of the damping matrix to a change in a design
variable p is on the form
Appendix A: Case of Finite Element Structural Models      
∂C ss ∂ M ss ∂ωi
When detailed large-scale finite element models are used for the [φ] = [φ][2ξi ωi ] + [M ss ][φ] 2ξi (B4)
structural dynamic behavior of a flight vehicle, full-order aerody- ∂p ∂p ∂p
namic matrices, corresponding to all degrees of freedom in the struc-
tural model, cannot be created using current lifting surface or panel Expression (B4) assumes that the sensitivities of the eigenvectors
methods. The aerodynamic matrices in the expression for full-order are zero. The eigenvectors and eigenvalues satisfy expression (B5).
load vector, [Ass (s)][φ], [Asc (s)], and [Asg (s)], are, in this case, ob-
[K ss ]{φi } = [M ss ]{φi }ωi2 (B5)
Downloaded by BRISTOL UNIVERSITY on March 5, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.9329

tained indirectly by what is known as the Summation of Forces and


Moments method.20,21 The number of rows in each of these matrices The sensitivity of expression (B5) with respect to design variables is
is the same as the number of degrees of freedom in the full-order shown in Eq. (B6), again with eigenvector sensitivities being zero.
finite element model, and they are obtained as follows.
On the aerodynamic mesh used to evaluate unsteady aerodynamic    
∂ K ss ∂ M ss ∂ω2
loads, the aerodynamic equation, for the case of lifting surface the- {φi } = {φi }ωi2 + [M ss ]{φi } i
ory, for example, is ∂p ∂p ∂p
 
[AI C(M∞ , s)]{ p} = (q D /U∞ ){wn (s)} (A1) ∂ M ss ∂ωi
= {φi }ωi2 + [M ss ]{φi }2ωi (B6)
∂p ∂p
where [AI C]is the aerodynamic influence coefficient matrix, p
the pressure distribution, and wn the normal velocity on the sur- By utilizing expression (B6), expression (B4) becomes
face. Equation (A1) is solved for right-hand sides {wn } j represent-        
ing normal velocity distributions caused by a set of motion-shape ∂C ss ∂ M ss ∂ K ss ξi
generalized coordinates (vibration modes, Ritz vectors, polynomial [φ] = [φ][ξi ωi ] + [φ] (B7)
∂p ∂p ∂p ωi
functions, etc.). For general motions, which are expressed as su-
perpositions of motions in the given generalized coordinates, the Here, [ξi ωi ] and [ξi /ωi ] are diagonal matrices with ξi ωi and ξi /ωi
normalwash is given by along on the diagonal, respectively.

{wn } = {wn } j q j (A2) References


1 Hoblit, F. M., Gust Loads on Aircraft: Concepts and Applications, AIAA,
The different pressure distributions { p} j can be integrated over Washington, DC, 1988, Chaps. 5 and 6.
2 Houbolt, J. C., Steiner, R., and Pratt, K. G., “Dynamic Response of
the surface of the vehicle, and the resulting loads distributed to the
nodes of the full-order structural mesh using a transformation Airplanes to Atmospheric Turbulence Including Flight Data on Input and
Response,” NASA TR-R-199, June 1964.
3 “Handbook for Aeroelastic Analysis,” MSC/NASTRAN Ver. 65,
{Fs } = [T1 ]{ p} = [T1 ]{ p} j
j
MacNeal–Schwendler Corp., Los Angeles, CA, Nov. 1987.
4 Manual on Flight of Flexible Aircraft in Turbulence, Advisory Group

qD for Aerospace Research and Development, AGARD-AG-317, Neuilly sur


= [T1 ][AI C]−1 {wn } j q j = [Ass ]{qs } (A3) Seine, France, May 1991.
U∞ j
5 Livne, E., “Integrated Aeroservoelastic Optimization: Status and Direc-
tion,” Journal of Aircraft, Vol. 36, No. 1, 1999, pp. 122–145.
The matrices [Asc (s)] and [Asg (s)] are obtained in a similar form, 6 Livne, E., Schmit, L. A., and Friedmann, P. P., “Towards an Integrated
representing forces on the full-order structural mesh caused by Approach to the Optimum Design of Actively Controlled Composite Wings,”
control-surface motions and gust excitations, respectively. Journal of Aircraft, Vol. 27, No. 12, 1990, pp. 979–992.
7 Livne, E., Schmit, L. A., and Friedmann, P. P., “Integrated Struc-
The rational function approximations (in the form of Roger or
minimum-state approximants38,39 can be used for the [Ass (s)][φ], ture/Control/Aerodynamic Synthesis of Actively Controlled Composite
[Asc (s)], and [Asg (s)] mastrices obtained from Eq. (A3). To cre- Wings,” Journal of Aircraft, Vol. 30, No. 3, 1993, pp. 387–394.
8 Bindolino, G., Ricci, S., and Mategazza, P., “Integrated Aeroservoelastic
ate the generalized aerodynamic matrices used for the mode- Optimization in the Design of Aerospace Systems,” Journal of Aircraft,
displacement equations (and their Roger or minimum-state approx- Vol. 36, No. 1, 1999, pp. 167–175.
imation), the aerodynamic forces on the full-order structural mesh 9 Karpel, M., “Reduced-Order Models for Integrated Aeroservoelastic Op-
[obtained from Eq. (A3)] are integrated over the configuration to timization,” Journal of Aircraft, Vol. 36, No. 1, 1999, pp. 146–155.
find the work they do when the configuration deforms in the motion- 10 Karpel, M., Moulin, B., and Love, M. H., “Modal-Based Structural

shape generalized coordinates one by one. Optimization with Static Aeroelastic and Stress Constraints,” Journal of
Aircraft, Vol. 34, No. 3, 1997, pp. 433–440.
11 Karpel, M., “Modal-Based Enhancement of Integrated Design Opti-
Appendix B: Viscous Damping Matrix
mization Schemes,” Journal of Aircraft, Vol. 35, No. 3, 1998, pp. 437–444.
If the damping in the structure is assumed to be viscous, Ref. 17 12 D’vari, R., and Baker, M., “Aeroelastic Loads and Sensitivity Analysis
shows that the nondiagonal damping matrix can be defined as for Structural Loads Optimization,” Journal of Aircraft, Vol. 36, No. 1, 1999,
pp. 156–166.
[C ss ] = [M ss ][φ]([φ]T [M ss ][φ])−1 13 Zole, A., and Karpel, M., “Continuous Gust Response and Sensitivity
Derivatives Using State-Space Models,” Journal of Aircraft, Vol. 31, No. 5,
× ([φ]T [C ss ][φ])([φ]T [M ss ][φ])−1 [φ]T [M ss ] (B1) 1994, pp. 1212–1214.
ENGELSEN AND LIVNE 347

14 Balis Crema, L., Mastrodi, F., and Coppotelli, G., “Aeroelastic Sensi- 32 Giles, G. L., “Equivalent Plate Modeling for Conceptual Design of
tivity Analyses for Flutter Speed and Gust Response,” Journal of Aircraft, Aircraft Wing Structures,” AIAA Paper 95-3945, Sept. 1995.
Vol. 37, No. 1, 2000, pp. 172–180. 33 Giles, G. L., “Design Oriented Analysis of Fuselage Structures Using
15 Bisplinghoff, R. L., and Ashley, H., Principles of Aeroelasticity, Dover, Equivalent Plate Methodology,” Journal of Aircraft, Vol. 36, No. 1, 1999,
New York, 1975, pp. 350, 417. pp. 21–28.
16 Hurty, W. C., and Rubinstein, M. F., Dynamics of Structures, Prentice– 34 Livne, E., Sels, R. A., and Bhatia, K. G., “Lessons from Application of
Hall, Upper Saddle River, NJ, 1964, pp. 299–307. Equivalent Plate Structural Modeling to an HSCT Wing,” Journal of Aircraft,
17 Craig, R. R., Structural Dynamics, Wiley, New York, 1981, Chap. 15. Vol. 31, No. 4, 1994, pp. 953–960.
18 Cornwell, R. E., Craig, R. R., Jr., and Johnson, C. P., “On the Applica- 35 Stone, S. C., Henderson, J. L., Nazari, M. M., Boyd, W. N., Becker, B. T.,
tion of the Mode-Acceleration Method to Structural Engineering Problems,” Bhatia, K. G., Giles, G. L., and Wrenn, G. A., “Evaluation of Equivalent
Earthquake Engineering and Structural Dynamics, Vol. 11, No. 5, 1983, Laminated Plate Solution (ELAPS) in HSCT Sizing,” AIAA Paper 2000-
pp. 679–688. 1452, April 2000.
19 Blleloch, Paul, “Calculation of Structural Dynamic Forces and Stresses 36 Livne, E., “Equivalent Plate Structural Modeling for Wing Shape Opti-
Using Mode Acceleration,” AIAA Journal, Vol. 12, No. 5, 1988, pp. 760–762. mization Including Transverse Shear,” AIAA Journal, Vol. 32, No. 6, 1994,
20 Perry, B., III, Kroll, R. I., Miller, R. D., and Goetz, R. C., “DYLOFLEX: pp. 1278–1288.
A Computer Program for Flexible Aircraft Flight Dynamic Loads Anal- 37 Livne, E., “Integrated Multidisciplinary Optimization of Actively Con-
ysis with Active Controls,” Journal of Aircraft, Vol. 17, No. 4, 1980, trolled Fiber Composite Wings,” Ph.D. Dissertation, Dept. of Mechanical,
pp. 275–282. Aerospace, and Nuclear Engineering, Univ. of California, Los Angeles, Sept.
21 Pototzky, A. S., and Perry, B., “Dynamic Loads Analyses of Flexible 1990.
Airplanes—New and Existing Techniques,” AIAA Paper 85-0808, 1985. 38 Roger, K. L., “Airplane Math Modeling Methods for Active Control
22 Arnold, R. R., Citerley, R. L., Chargin, M., and Galant, D., “Application Design,” Proceedings of the 44th AGARD Structures and Material Panel,
of Ritz Vectors for Dynamic Analysis of Large Structures,” Computers and Structural Aspects of Active Controls, AGARD-CP-228, Lissabon, Portugal,
Downloaded by BRISTOL UNIVERSITY on March 5, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.9329

Structures, Vol. 21, No. 5, 1985, pp. 901–908; also Vol. 21, No. 3, 1985, April 1977, pp. 4.1–4.11.
pp. 461–467. 39 Karpel, M., “Time Domain Aeroservoelastic Modeling Using Weighted
23 Kline, K. A., “Dynamic Analysis Using a Reduced Basis of Exact Unsteady Aerodynamic Forces,” Journal of Guidance, Control, and
Modes and Ritz Vectors,” AIAA Journal, Vol. 24, No. 12, 1986, pp. 2022– Dynamics, Vol. 13, No. 1, 1990, pp. 30–37.
2029. 40 Campbell, C. W., “Monte Carlo Turbulence Simulation Using Rational
24 Karpel, M., and Raveh, D., “The Fictitious Mass Element in Structural Approximations to Von Karman Spectra,” AIAA Journal, Vol. 24, No. 1,
Dynamics,” AIAA Journal, Vol. 34, No. 3, 1996, pp. 607–613. 1986, pp. 62–66.
25 Livne, E., and Blando, G. D., “Reduced Order Design-Oriented Stress 41 Haftka, R. T., and Yates, E. C., “Repetitive Flutter Calculations in Struc-
Analysis Using Combined Direct and Adjoint Solutions,” AIAA Journal, tural Design,” Journal of Aircraft, Vol. 13, No. 7, 1976, pp. 454–461.
Vol. 38, No. 5, 2000, pp. 898–909. 42 Nicot, P., and Petiau, C., “Aeroelastic Analysis Using Finite Ele-
26 Livne, E., and Blando, G. D., “Structural Dynamic Frequency Response ment Models,” DGLR/AAAF/RAES, European Forum on Aeroelasticity
Using Combined Direct and Adjoint Reduced-Order Approximations,” AIAA and Structural Dynamics, Aachen, Germany, April 1989.
Journal, Vol. 41, No. 7, 2003, pp. 1377–1385. 43 Canfield, R. A., “High Quality Approximation of Eigenvalues
27 Bryson, A. E., and Ho, Y-C., Applied Optimal Control, Ginn and Co., in Structural Optimization,” AIAA Journal, Vol. 28, No. 6, 1990,
Waltham, MA, 1969, Chap. 10 and 11. pp. 1116–1122.
28 Mukhopadhyay, V., Newsome, J. R., and Abel, I., “A Method for Ob- 44 Canfield, R. A., “Design of Frames Against Buckling Using a
taining Reduced-Order Control Laws for High Order Systems Using Opti- Rayleigh Quotient Approximation,” AIAA Journal, Vol. 31, No. 6, 1993,
mization Techniques,” NASA TP-1876, Aug. 1981. pp. 1143–1149.
29 Livne, E., “Alternative Approximations for Integrated Control/Structure 45 Engelsen, F., and Livne, E., “Design-Oriented Quadratic Stress Failure
Aeroservoelastic Synthesis,” AIAA Journal, Vol. 31, No. 6, 1993, pp. 1100– Constraints for Actively Controlled Structures under Combined Steady and
1108. Random Excitation,” AIAA Journal (to be published).
30 Schmit, L. A., “Structural Optimization—Some Key Ideas and In- 46 Engelsen, F., “Design-Oriented Gust Stress Constraints for Aeroser-
sights,” New Directions in Optimum Structural Design, edited by E. Atrek, voelastic Design Synthesis,” Ph.D. Dissertation, Dept. of Aeronautics and
R. H. Gallagher, K. M. Ragsdell, and O. C. Zienkiewicz, Wiley, New York, Astronautics, Univ. of Washington, Seattle, Aug. 2001.
1984. 47 Karpel, M., and Wieseman, C. D., “Modal Coordinates for Aeroelastic
31 Haftka, R. T., and Gurdal, Z., Elements of Structural Optimization, 3rd Analysis with Large Local Structural Variations,” Journal of Aircraft, Vol. 31,
ed., Kluwer Academic, Norwell, MA, 1992, Chap. 6. No. 2, 1994, pp. 396–403.

You might also like