Professional Documents
Culture Documents
G. Dimitriadis
Universit de Lige, 4000 Lige, Belgium
DOI: 10.2514/1.28759
In recent years the aeroelastic research community has carried out substantial work on the characterization and
prediction of nonlinear aeroelastic phenomena. Of particular interest is the calculation of limit cycle oscillations,
which cannot be accomplished using traditional linear methods. In this paper, the prediction of the bifurcation and
postbifurcation behavior of nonlinear subsonic aircraft is carried out using numerical continuation. The analysis
does not make use of continuation packages such as AUTO or MatCont. Two different continuation techniques are
Downloaded by UNIVERSITY OF CALIFORNIA IRVINE on January 30, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.28759
detailed, specically adapted for realistic aeroelastic models. The approaches are demonstrated on a model of a
simple pitchplunge airfoil with cubic stiffness and an aeroelastic model of a transport aircraft with two different
types of nonlinearity in the control surface. It is shown that one of the techniques yields highly accurate predictions
for limit cycle oscillation amplitudes and periods while the second method trades off some accuracy for
computational efciency.
I. Introduction
f x; 0 (1)
T HE prediction of the bifurcation behavior of aircraft is an area of
research that is receiving increasing amounts of interest,
especially the prediction of bifurcations resulting in limit cycle
where f is a vector containing nonlinear functions, x is a vector of
independent unknowns, and is a parameter whose value can vary.
oscillations (LCO). Bifurcations in aircraft can occur due to a number Given a known solution x0 at 0 , numerical continuation can be used
of nonlinearities that may be present, including structural, aero- to obtain a new solution x x at 0 , where is a small
dynamic, or control nonlinearities. Various methods have been change in parameter value.
proposed in the past for the prediction of the bifurcation conditions The numerical continuation of nonlinear dynamic systems deals
and postbifurcation behavior of nonlinear aircraft, such as the with the solution of the equations of motion of these systems. In this
harmonic balance method, normal form [1], center manifold [2], and work, it will be assumed that the equations of motion are ordinary
various types of linearization. Numerical continuation has also been differential equations (ODEs) of the form
used by some researchers. However, the treatment of numerical
continuation by aeroelasticians has always been through software x_ fx; (2)
packages such as AUTO [3] and MatCont [4]. Such studies include
the piecewise linear work by Roberts et al. [5], the store-induced limit where xt x1 t xn t T is a n 1 vector of system states
cycle oscillation investigation by Beran et al. [6], the panel utter and f f1 fn T is a n 1 vector of functions of x and the
work by Gee [7], and other LCO prediction work [8,9] parameter . In this context, solving for fx; 0 is equivalent to
Although complete packages such as AUTO and MatCont are x_ 0, that is, looking for the systems stationary points or
very comprehensive and can deal with many different types of equilibrium points. As the parameter is varied it can happen that the
bifurcation, they were developed for general nonlinear dynamic behavior of the system changes from decaying toward one or more
systems. As such they are not necessarily suitable for very large stationary solutions to admitting periodic solutions, called limit cycle
systems with dozens or even hundreds of states. In this paper oscillations. This phenomenon is usually called a Hopf bifurcation
numerical continuation is implemented from rst principles, with- [10]. In this work, the continuation of such periodic solutions will be
out recourse to dedicated packages. It is shown how the method can mainly considered.
be applied to aeroelastic models of full aircraft. The method is It is assumed that the system of Eq. (2) admits periodic solutions of
tailored to be as efcient as possible given the particularities of period T for a range of values. The time coordinate can be scaled
aeroelastic models. In essence, this paper offers a complete numeri- with respect to the period using
cal continuation methodology for nonlinear aircraft with linear t
aerodynamics. (3)
T
aircraft model with 10 structural modes will have 20 structural states treated by the shooting approach.
and 40 aerodynamic states if unsteady aerodynamics with four lags
are used. If 200 time steps are used in one period then the total A. Starting the Continuation Scheme
number of unknowns is 12,000. By default, AUTO 2000 restricts the Nonlinear systems do not always have periodic solutions.
problem size to 36 states. Therefore, alternative techniques must be Aeroelastic systems, in particular, will generally have decaying
sought for aeroelastic problems. responses which bifurcate to limit cycle oscillations at a critical ight
Such an alternative method for the transformation of the condition. At subcritical ight conditions there is no sense in using
differential equations to algebraic equations is the shooting method, numerical continuation as described in the previous section. The
which is essentially a boundary value approach. For a solution x starting algorithm for the continuation scheme depends on the type
to be periodic, the values of the states at the start of the period must be of bifurcation that occurs at the critical conditions. The vast majority
equal to those at the end of the period. Therefore, the numerical of studies into nonlinear aeroelastic systems deal with Hopf
continuation scheme becomes a boundary value problem solving bifurcations.
The rst step is to determine the critical value of the airspeed, or
x 0 x1 (8) other continuation parameter, at which Hopf bifurcation occurs. It
can be shown that, when looking for a Hopf bifurcation, the stability
of the full nonlinear system [Eq. (2)] depends on the stability of the
The rst-order system of differential equations becomes a Newton linearized system around the equilibrium point, xE . The equations of
system for motion of this linearized system are given by
F x; T; x0 x1 0 (9) @f
_ (12)
@x xE ;
The function F is evaluated by choosing an initial condition vector which are rst-order, linear, time-invariant ODEs. Therefore, the
x0 and calculating x1 using a numerical integration method. Any stability of the nonlinear system can be investigated by calculating
numerical integration technique, such as Newmark or RungeKutta the eigenvalues, p, of the matrix A @f=@xjxE ; . If all the
can be used. In essence, the shooting method is a clever application of eigenvalues have negative real parts then the system is stable and its
numerical integration. Starting from a known solution x0 and T0
response amplitude will decay with time. If at least one complex
at 0 , the full Newton system at 0 becomes conjugate pair of eigenvalues, pc , has zero real parts, a Hopf
bifurcation has occurred [14]. The parameter value at which this
@F @F x condition is met is the critical parameter value, c .p
The
Fx0 0; T0 ; critical
pair of
@x0 x0 ;T0 ; @T x0 ;T0 ; T eigenvalues will be of the form
|!c , where | 1 and !c is the
(10) radial frequency of the resulting limit cycle. Therefore, the
x1 0 x0 0 x oscillation period at the Hopf point is Tc 2=!c [14].
T1 T0 T Once the critical value c has been encountered, the LCO
continuation scheme can be started. At a condition c the
Equations (10) describe n equations with n 1 unknowns. They solution of the equations of motion are approximated by a sine wave
can be completed using a phase-xing equation. The simplest and with period !c and amplitude parallel to the eigenvectors
most robust technique is to set one of the elements of x0 to be corresponding to the critical eigenvalue of A at c , that is,
always equal to zero, for example, x1 0 0. Evidently, this can be
done without loss of generality. Consequently, the number of x ; c "v1 e2| v2 e 2| (13)
unknowns reduces to n.
where v1 is the eigenvector of A corresponding to the |!c
The derivative matrices, @F=@x0 and @F=@T, can be calculated
eigenvalue, v2 is the eigenvector corresponding to the p |!
c
analytically if all the nonlinear functions in Eq. (4) are analytic. If this
is not the case, two approaches can be followed. The nonanalytic eigenvalue, and " is a small positive real scalar. Choosing "
functions can be replaced by analytic approximations, for example, will, in general, yield an initial guess good enough to allow the
Coulomb friction can be replaced by a hyperbolic tangent function. Newton system to converge to the true periodic solution at c .
Specically for piecewise linear functions, Roberts et al. [5] used the
analytic solutions of the linear subsystems to describe the response of B. Arclength Continuation
the fully nonlinear system. The numerical continuation scheme described in the previous
The other approach is to calculate the derivative matrices sections is the most basic scheme available. It is usually called
numerically. This can be achieved using a simple forward differences natural parameter continuation because the continuation parameter
scheme. For example, the derivative @F=@x0jx0 ;T0 ; can be is also the system parameter. For nonlinear problems the number of
calculated as real solutions is not constant. Solution branches can appear,
DIMITRIADIS 895
disappear, or intersect at different values of the natural parameter. several such algorithms but here only the step length adaptation by
Additionally, solution branches may undergo folds, that is, changes asymptotic expansion [12] method will be considered. The method
of direction in parameter space. Bifurcation and fold points are attempts to x the rate at which the Newton system converges.
collectively known as singularities [15]. Natural parameter Consider the Newton system of Eq. (15) and denote by u the vector
continuation will fail at singularity points. u x0 T
p T
T
T T . The Newton iterations are stopped
A more robust continuation strategy is the so-called arclength when uT u < ", where " is a small positive real number. Denote by
continuation. In this case, the chosen continuation parameter is not u1 the value of u at the end of the rst iteration and by u2 its value at
but the distance s, traveled along the branch being followed. The the end of the second iteration. A contraction rate can be dened as
numerical continuation is preformed in small steps of s instead of p
small steps of . uT2 u2
Around the fold, s must be small to continue the solution p (18)
uT1 u1
successfully; but away from the fold, a small value of s is not
necessary and computationally inefcient. This is usually achieved so that it represents the speed with which the solution is approached
by means of pseudoarclength continuation [11]. The pseudoar- at the rst two iterations. It is desired to keep this speed to a constant
clength is dened in terms of the rate of change of x0, T, and with value . Then, once the Newton system has converged and new
s as direction vectors have been calculated, a new arclength increment,
is dened as
s
@x0T @T @
Downloaded by UNIVERSITY OF CALIFORNIA IRVINE on January 30, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.28759
x T s (14) r
@s x0 ;T0 ;0 @s x0 ;T0 ;0 @s x0 ;T0 ;0
s s (19)
and is now treated as an additional unknown so that the
pseudoarclength form of the Newton system becomes In this manner the contraction rate is kept approximately constant as
0 1 the solution branch is followed.
@F
@F @F
0 1
B @x0 @T @ C x
B x
0 0 0
;T ; x
0 0 0
;T ; x ;T ; C
0 0 0 C@ T A D. Stability Analysis
B
@ @x0T
@T
@ A Periodic solutions of nonlinear systems can be stable, unstable, or
@s @s @s
x0 ;T0 ;0 x ;T0 ;0
0 0
x ;T0 ;0
0 0
half-stable depending on whether they attract trajectories or not.
(15) Formally, for a periodic orbit x to be stable, the Poincar stability
Fx0 0; T0 ; 0 condition [16] states that
s ZT
0 1 0 1 0 1
x1 0 x0 0 x r fx dt < 0 (20)
@ T1 A @ T0 A @ T A 0
that is, the length of the approximate total direction vector obtained It is desired to calculate other solutions x in the neighborhood of 0 .
from Eq. (16). The main assumption underlying this calculation is that the
derivative of f with respect to is continuous and nite in the
C. Step Control neighborhood. Then, if 0 is increased by a small amount , the
The value of the arclength increment, s, is a crucial parameter. If changes in x and T are also small, so that
it is too high, then the singularity may be missed and continuation x x0 x; T T0 T (23)
may fail. If it is too low, then the continuation scheme becomes
unnecessarily slow. In general, as a solution branch is followed, the Substituting Eqs. (23) into Eq. (4) yields
direction vectors will change length and the solution will advance at a
variable rate. x 00 x0 T0 Tfx0 x; 0 (24)
Step control algorithms attempt to normalize the advance rate of
the solution so that the arclength increment becomes smaller at areas Expanding the nonlinear function as a rst-order Taylor series
of high curvature and larger at areas of low curvature. There are around 0 yields
896 DIMITRIADIS
@f
@f Y
j
x00 x0 T0 T fx0 ; 0 @x x @ xj1 I T0 Aj i x0
x0 ;0 x0 ;0
i0
T0 C T0 Ax T0 B CT (25) X Y
j 1 j i 1
I T0 Aj k T0 Bi Ci T
where the derivatives of f with respect to x (Jacobian) and are i0 k0
x0 x1 (28) which is a linear system of n algebraic equations for the n 1
unknowns x0 and T. To complete the system of equations a
phase-xing condition must be introduced. One such condition is the
This condition is the governing equation for the calculation of x
Poincar phase condition, given by
and T. There are many ways to complete this calculation; here a
rst-order nite difference scheme is chosen. The derivative of x
with respect to can be expressed as x0 x0 0T x00 0 0 (36)
continuation step approximates the exact solution and the accuracy From Eq. (34), xT
is estimated by x m , which can be obtained as
of this approximation increases with decreasing . As there are no
NewtonRaphson iterations the method is very fast but requires more x m Pm 1 x
0 Qm 1 T Rm 1 (44)
continuation steps because must be signicantly smaller than in
the case of the shooting method.
so that Eq. (43) can be rewritten as
A. Pseudoarclength Continuation
" T PTm 1 Pm 1 " < "T " (45)
The incremental numerical continuation scheme can be used in
conjunction with a pseudoarclength parameter in a manner very
similar to that demonstrated in Sec. III.B. In the case of the INC This calculation can be performed numerically, for example, by
method the unknowns are the increments in the values of x0 and the setting " 0:1x0 and checking whether Eq. (45) is true.
increment in the period. The increments x0 cannot be used directly A more thorough but computationally expensive way of assessing
to dene an arclength, only a change in arclength. Therefore, the the stability of the trajectory is to use the Floquet theory. According
initial values x0 are used to dene the pseudoarclength, as was done to this theory the trajectory is stable if the eigenvalues of the
for the shooting method. monodromy matrix, M, all lie inside the unit circle, that is, their
The pseudoarclength form of the INC equations becomes magnitudes are less than 1. For the discrete representation considered
here the monodromy matrix is given by
0 1
Downloaded by UNIVERSITY OF CALIFORNIA IRVINE on January 30, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.28759
I Pm 1 Qm 1 Rm 1 8 9 ( ) X
B C 0T
0 0 C< x0 = 0 m
B C T 0 M exp Ai (46)
@ @x0T @T
@
A: ; i1
@s @s @s s
x0 ;T0 ;0 x0 ;T0 ;0 x0 ;T0 ;0
(39) where exp denotes the matrix exponential. The advantage of using
the monodromy matrix as a stability criterion is that, by examining
The new direction vectors can be calculated from the increments the behavior of its eigenvalues, the changes in stability can be
x0 , T, , and s as explained.
@x0 x0 @T T @
(40)
@s x;T; s @s x;T; s @s x;T; s IV. Implementation
and then scaled by parameter a in Eq. (17) to ensure that the length of All the algorithms described previously were implemented using a
the total direction vector is 1. combination of Matlab and Matlab-compatible C code (mex les).
The Matlab interface was used to dene the system and to set the
continuation and simulation parameters. The numerical integrations
B. Stability Analysis were carried out using the mex le by means of a fth-order Runge
The Poincar stability condition of Eq. (20) can be readily Kutta scheme [18]. The continuation results were produced
evaluated from INC calculations. The Jacobian, A, of function effectively and efciently even for a realistic aeroelastic system with
fx0 is obtained from Eq. (26) as part of these calculations. The 42 states. The codes developed were run on a 1.25-GHz G4
divergence of fx0 can be easily evaluated from r fx0j Powerbook with 1.25 MB of RAM; in other words, the
traceAj where trace denotes the sum of the diagonal elements of a computational requirements are very light, even for the largest
square matrix. The Poincar condition approximately becomes aeroelastic model treated later in this paper.
X
m
traceAj < 0 (41)
j1 V. PitchPlunge Airfoil Model
In this section, the two continuation methods outlined previously
Notice that in previous calculations Aj is only available for the are applied to the pitchplunge airfoil aeroelastic model. This model
starting orbit x0 . Therefore, Aj must be recalculated for the current has been used for aeroelastic research since the Theodorsen era and
orbit x0 x. has found applications in the investigation of nonlinear aeroelasticity
A much faster approximate estimation of the stability of a periodic over the last two decades. Lee et al. [19,20] used the pitchplunge
orbit can be obtained by considering a small disturbance. For a airfoil with cubic and other nonlinear stiffnesses to demonstrate some
periodic orbit to be stable, trajectories starting from initial conditions basic nonlinear aeroelastic phenomena. Tang et al. [21] and Liu and
close to the orbit should always end up on the orbit itself as time tends Dowell [22] used a version of the same airfoil with a control surface
to innity. Dene a set of initial conditions x0 such that to demonstrate secondary bifurcations and to apply a higher order
harmonic balance method, respectively. More recently, Millman
x0
x0 " (42) et al. [23] used the pitchplunge airfoil to analyze chaotic
bifurcations and Baldelli et al. [24] demonstrated a robust solution
where " is a n 1 vector of magnitude much smaller than x0. method for aeroelastic systems containing uncertainties. Marsden
The initial conditions x0
will cause the system to follow a and Price [25] built a wind-tunnel model of the nonlinear airfoil and
trajectory xt.
If the periodic orbit xt is stable then compared the responses to simulation results. Lee et al. [26] returned
to the simulated model with cubic stiffness to investigate the
lim xt
xt 0 effectiveness of higher order harmonic balance methods and to
t!1
investigate bifurcations in the presence of two cubic springs [27].
An equivalent condition can be used to speed up calculations, The pitchplunge airfoil without control surface, as investigated
looking at the behavior of the trajectory xt
over a single period. by Lee et al. [26,27], will be the subject of the application of
For stability, xT
must be closer to xT than x0 is to continuation methods in this paper. It consists of a two-dimensional
x0. This can be written as airfoil with pitch and plunge degrees of freedom, restrained by
torsion and extension springs, respectively; both of these springs can
xT
xTT xT
xT be nonlinear. The aerodynamic forces are calculated using
incompressible, inviscid, thin airfoil, unsteady assumptions. The
< x0
x0T x0
x0 (43) equations of motion for the system are
898 DIMITRIADIS
! ! ! !
c0 c1 00 c2 c3 0 integration did not reveal the cause of this jump and neither did the
harmonic balance analysis they performed. The numerical
00 0
d0 d2 d2 d3 integration results, recalculated by this author, are shown in Fig. 1.
! ! The maximum plunge amplitude, pitch amplitude, and LCO
c4 c10 c5
frequency are plotted at several postcritical airspeeds. The jump at
d4 d5 d10 around U =UL 2 is clearly visible.
0 1 The power of numerical continuation lies not only in quickly
w1
!B C ! estimating LCO amplitude and frequency for many parameter values
c6 c7 c8 c9 B w2 C c10 3 but also in fully characterizing all the bifurcation phenomena
B C 0 (47)
B C occurring in the parameter space of interest. Here, the shooting
d6 d7 d8 d9 @ w3 A d10
3
method was applied to Eq. (47). The results from this application are
w4 plotted in Fig. 2, along with the numerical integration results of
Fig. 1. The method detected a supercritical Hopf bifurcation at
where is the nondimensional plunge displacement, is the pitch UL 6:29, followed by a fold bifurcation, which causes the jump
angle in rad, the (prime) denotes differentiation with respect to effect observed by Lee et al. The LCO branch reaches a maximum
nondimensional time, and and
are cubic stiffness coefcients. airspeed ratio of 2.35, at which point it becomes unstable and folds
The coefcients ci , di are given in several (if not all) of the over toward decreasing airspeed values. The branch becomes stable
Downloaded by UNIVERSITY OF CALIFORNIA IRVINE on January 30, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.28759
publications mentioned previously but are included as an Appendix again and folds toward positive airspeed values at U =UL 1:85.
to this paper for completeness. This means that between airspeed ratios of 1.85 and 2.35 the system
In Lee et al. [26], results from this model are presented for the case can undergo two possible stable limit cycles. Which one it will
where 100, r 0:5, ah 0:5, 0, x 0:25, eventually follow depends on initial conditions. Numerical continua-
! 0:2,
80, and 0. Lee et al. used numerical integration to tion has provided a complete explanation of the jump phenomenon
solve Eq. (47); they observed an initial Hopf bifurcation leading to observed by Liu et al. and left unexplained by numerical integration
limit cycle oscillations at the linear utter speed UL , followed by a and harmonic balance analysis. It should be noted, however, that a
jump in LCO amplitude and frequency at around U =UL 2. This higher order harmonic balance method could have yielded results
jump occurred at slightly different airspeeds, depending on the initial similar to those presented in Fig. 2 but only if it was implemented
conditions used for the numerical integration [26,27]. Numerical within a continuation framework [28].
2.5 20
18
2 16
Pitch amplitude (rad)
14
Plunge amplitude
1.5 12
10
1 8
0.5 4
0 0
1 1.5 2 2.5 1 1.5 2 2.5
U*/U*L U*/UL*
0.09
Fundamental frequency (rad/nondimensional time)
0.085
0.08
0.075
0.07
0.065
0.06
0.055
0.05
0.045
0.04
1 1.5 2 2.5
U*/U *L
c) LCO frequency
Fig. 1 Numerical integration results for pitchplunge airfoil with 80, 0 [26].
DIMITRIADIS 899
2.5 20
18
2 16
1.5 12
10
1 8
0.5 4
0 0
1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6
Downloaded by UNIVERSITY OF CALIFORNIA IRVINE on January 30, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.28759
U*/U*L U*/UL*
Numerical Integration
0.085
0.08
0.075
0.07
0.065
0.06
0.055
0.05
0.045
0.04
1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6
U*/U*L
c) LCO frequency
Fig. 2 Numerical continuation results for pitchplunge airfoil with 80, 0.
Lee and Liu [27] demonstrated using numerical integration that, rectangular wing and a T tail with a straight rectangular tailplane.
when 20 and all other parameters remain as before, a different Five aerodynamic control surfaces are included, two ailerons, two
jump phenomenon is observed, this time independent of initial elevators, and a rudder. No engines are modeled. Figure 5a shows the
conditions. They identied the cause of this jump as a secondary aerodynamic grid and Fig. 5b the structural grid. The structure is a
Hopf bifurcation. Numerical continuation can provide evidence that nite element stick model created using bar elements. The masses are
indeed this is the cause of the jump by looking at the eigenvalues of
the matrix @F=@x0 @F=@T in Eq. (10). When a pair of
eigenvalues of this matrix becomes purely imaginary, a secondary 0.5
Hopf bifurcation occurs. Figure 3 shows the variation of the real part
of some of the eigenvalues, obtained when applying the shooting
method. It can be seen that one of the real parts becomes zero at
U =UL 1:78. The LCO amplitude and frequency predictions
obtained from numerical continuation are presented in Fig. 4, along
Real parts of eigenvalues
0
with numerical integration results. It can be seen that the jump
coincides with the secondary Hopf bifurcation.
1 20
0.9 18
0.8 16
0.6 12
0.5 10
0.4 8
0.3 6
0.2 4
0.1 2
0 0
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
Downloaded by UNIVERSITY OF CALIFORNIA IRVINE on January 30, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.28759
U*/UL* U*/U*L
Numerical Continuation
0.09 Numerical Integration
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
U*/U*L
c) LCO frequency
Fig. 4 Numerical continuation results for pitchplunge airfoil with 80, 20.
where M is the modal mass matrix, C is the modal damping matrix, j 1:7kmax
j
(50)
K is the modal stiffness matrix, q are modal coordinates, q is the nl 12
dynamic pressure, and Q|k are the aerodynamic forces, which are
functions of the reduced frequency, k. The number of retained
modes, nm , is 7. The equations of motion are transformed to rst- with kmax being the maximum reduced frequency at which the
order form using Rogers approximation [31]. The aerodynamic aerodynamic matrix, Q|k, was obtained from NASTRAN and nl
forces are written in the form being chosen as equal to 4.
40
35
25
20
15
10
a) Aerodynamic grid 5
0 50 100 150 200 250
Downloaded by UNIVERSITY OF CALIFORNIA IRVINE on January 30, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.28759
Airspeed (m/s)
a) Modal frequencies
0.6
0.5
0.4
Damping ratio
0.3
0.2
0.1
0.1
b) Structural grid
Fig. 5 Aerodynamic and structural grids for the GTA model. 0.2
0 50 100 150 200 250
Airspeed (m/s)
where V is the airspeed and b) Damping ratios
2 Fig. 6 Modal frequencies and dampings for the GTA model.
M M q b A2 C q b A1
C K
M qA0
V V
A
3 qA3 A
4 qA4 A 5 qA5 A
6 qA6 elements 563 and 623 which are equal to 1, Kc is a cubic stiffness
coefcient, and is the right-hand aileron angle.
(52) In the case of freeplay stiffness a nonlinear force of the form
8
The state vector x is now of the form x q_ T qT xTa T , where xa < Kf if
are aerodynamic states. The total number of states in Eqs. (51) is f nl N 0 if < < (54)
6nm 42. : K if
f
Figure 6 shows the resulting modal frequencies and damping
ratios obtained from the GTA model for a range of airspeeds at sea
is added to the equations of motion where Kf is a stiffness coefcient
level. It can be clearly seen that one of the damping ratios crosses zero
and is equal to half the width of the freeplay region. In the present
at an airspeed of just under 200 m=s, causing utter.
study the stiffness Kf was chosen equal to that of the linear system
Structural nonlinearity was included in the GTA model by
whose natural frequencies and damping ratios are plotted in Fig. 6.
replacing the springs connecting the aerodynamic control surfaces by
The freeplay amplitude is chosen equal to 0.005 rad, that is, almost
nonlinear springs. In this study the case where the right-hand aileron
0.3 deg.
stiffness is nonlinear will be considered. Two nonlinearities are
Although a single nonlinearity is included in the two models for
treated, cubic and freeplay. The nite element grid contains ng
this demonstration, the number of nonlinearities can be increased for
678 nodes. The right-hand aileron angle is given by the difference
negligible computational cost. For both types of nonlinearity three
between the rotations at nodes 563 and 623. To implement
sets of results are presented, one obtained from the shooting method,
nonlinearity in the model the elements of the physical structural
one from incremental numerical continuation, and one from simple
stiffness matrix (563,563), (563,623), (623,563), and (623,623) are
numerical integration.
set to zero and the resulting matrix is transformed to modal
coordinates. In the case of cubic stiffness a nonlinear force of the
form A. Shooting Method Results
The shooting method results were obtained using the Runge
f nl NKc 3 (53) Kutta fth-order scheme mentioned earlier to evaluate Eq. (9) with
m 251. Pseudoarclength continuation was chosen and the
is added to the equations of motion, where is the nm ng modal contraction rate was chosen as 0:013, while the initial arclength
matrix, N is a ng 1 vector whose elements are all zero apart from increment was set to s 20. The initial periodic solution was
902 DIMITRIADIS
obtained using Eq. (13). The system Jacobian was updated using predictions as initial conditions, only as a check to verify that these
Broydens method. are truly attainable LCOs.
The bifurcation behavior of the GTA model with cubic stiffness The bifurcation behavior of the GTA model with freeplay is
can be observed in Fig. 7. The gure plots the LCO amplitude and dictated by the stability of two underlying linear systems: one with
period variation with airspeed obtained from the shooting approach zero stiffness in the right-hand aileron (here called the inner linear
(line) and numerical integration (circles). It can be seen that the initial system) and one with stiffness Kf (outer linear system). The utter
Hopf bifurcation is followed by a slow increase in LCO amplitude speed of the inner linear system at sea level is 191:7 m=s. At this
and decrease in LCO period. The changes in these two quantities airspeed the nonlinear system starts to undergo LCOs with amplitude
accelerate briey after 206 m=s until a turning point is reached at equal to . The outer linear system utters at 204:1 m=s. At this
214 m=s. The turning point is a fold, which is easily negotiated by the airspeed the nonlinear system also undergoes divergent oscillations.
pseudoarclength technique. The limit cycle becomes unstable and Figure 8 plots the LCO amplitude and period against airspeed
moves toward lower airspeeds until it folds again toward higher obtained from the shooting approach (line) together with the
airspeeds and becomes stable again. The numerical continuation numerical integration results (circles). It can be seen that the two sets
results were obtained for another three occurrences of this phenom- of results agree quite closely.
enon and then stopped. Numerical integration results were calculated The LCO amplitude plot shows that, as expected, LCOs rst
only for the rst two folds. appear at the utter speed of the inner linear system and their
Figure 7 shows that, although there is only one possible low amplitude is equal to . The amplitude then increases with airspeed
Downloaded by UNIVERSITY OF CALIFORNIA IRVINE on January 30, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.28759
amplitude limit cycle at airspeeds less than 202 m=s, at higher until it shoots off to innity at the utter speed of the outer linear
airspeeds there is a multitude of possible limit cycles with system. The period of the LCOs decreases monotonically with
progressively higher amplitudes. This phenomenon can only be fully airspeed. It should be noted that the existence of the discontinuity
observed in the numerical continuation results. Numerical integra- complicates numerical integration in that the time instances at which
tion will only calculate the stable parts of the branch and is much the nonlinear force switches between one of its three possible values
more computationally expensive as integrations must be carried out must be calculated to within a suitable tolerance. Various authors
using a matrix of combinations of initial conditions and airspeeds. have proposed iterative schemes to pinpoint these time instances. As
The higher number of states makes this an impossible task. In the shooting method also depends on numerical integration, the same
the present study the numerical integration results for the higher considerations should apply. In this authors opinion the exact
amplitude LCOs were obtained using the numerical continuation location of switching time instances is not necessary if the time step is
small enough. To test this opinion, the numerical integration results
0.1
Continuation 0.045
Integration
0.09 Continuation
0.04 Integration
0.08
LCO amplitude in (rad)
0.035
LCO amplitude in (rad)
0.07
0.03
0.06
0.025
0.05
0.02
0.04
0.015
0.03
0.01
0.02
0.005
0.01
0 0
190 195 200 205 210 215
192 194 196 198 200 202 204
Airspeed (m/s) Airspeed (m/s)
a) LCO amplitude a) LCO amplitude
0.085 0.086
Continuation Continuation
Integration Integration
0.084
0.08
0.082
LCO period (s)
LCO period (s)
0.075
0.08
0.07
0.078
0.065
0.076
0.06 0.074
190 195 200 205 210 215 190 195 200 205
Airspeed (m/s) Airspeed (m/s)
in Fig. 8 were obtained with switch time location but the numerical the GTA with freeplay compared to the numerical integration results.
continuation results without. The good agreement between the two It can be clearly seen that increasing values of m yield signicantly
sets of results seems to support this assertion. improved LCO estimates. The m 501 case fails completely, while
the m 1001 case fails to follow the sharp rise in LCO amplitude as
B. Incremental Numerical Continuation Results the utter speed of the outer linear system is approached. The
m 2001 case is qualitatively correct but inaccurate at higher
Incremental numerical continuation does not solve the full airspeeds, while the m 4001 case is quite accurate at all airspeeds.
nonlinear equations of motion. It solves Eq. (27), which is a Figure 10 presents the incremental numerical continuation results
tangential linearization of the nonlinear equations of motion. for LCO amplitude and frequency for the GTA with cubic stiffness.
Equation (27) is only usable when x, T, and are small, for Three sets of results are shown, a continuation result obtained using
example, jj 1. For an aeroelastic system, it can be expected m 4001, a continuation result calculated from m 10; 001, and a
that the increments in modal displacements and the period with numerical integration result. It can be seen that the m 4001
airspeed are small. However, continuing a solution from, say, 100 to predictions are accurate for the part of the LCO branch before the rst
200 m=s in increments of 0:01 m=s is tremendously wasteful in fold. After the fold the predictions deteriorate. The m 10; 001
computational resources. Therefore, cannot be chosen to be the curve shows that the section of the branch between the rst two folds
airspeed. To speed up the continuation process, can be chosen as a is predicted quite accurately but the section after the second fold is
scaled airspeed, for example, V=r, where r is a scale factor. slightly less accurate. Because of the nature of the incremental
Therefore, V=r ensures that Eq. (27) is valid while allowing
Downloaded by UNIVERSITY OF CALIFORNIA IRVINE on January 30, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.28759
0.12
INC4001
INC10001
0.06 Integration
0.06
0.03
0.04
0.02
0.02
0.01
0 0
190 195 200 205 210 215
192 194 196 198 200 202
Airspeed (m/s) Airspeed (m/s)
0.085
0.085 INC4001
INC10001
INC 4001 Integration
0.084 INC 2001
INC 1001
0.083 INC 501 0.08
Integration
0.082
LCO period (s)
LCO period (s)
0.081 0.075
0.08
0.079 0.07
0.078
0.077 0.065
0.076
0.075
0.06
190 192 194 196 198 200 202 204 190 195 200 205 210 215
Airspeed (m/s) Airspeed (m/s)
VII. Conclusions d2 1 2ah 1 1 2 = r2
A complete methodology has been presented for the prediction of
the bifurcation and postbifurcation behavior of nonlinear subsonic
aircraft using numerical continuation. Two efcient methods have
been described, tailored to the needs of practical aeroelastic systems d3 1 2ah = 2r2
and coded using simple tools and low computational requirements.
The power of numerical continuation was demonstrated on a pitch 1 2ah 1 2ah 1 1 2 = 2r2 2 =U
plunge airfoil where it was shown that the approach can completely
characterize the whole spectrum of system responses, unlike
numerical integration. Both continuation methods described in this d4 1 2ah "1 "2 r2
1 2 =
paper have been demonstrated on a realistic aeroelastic model of a
transport aircraft with two types of nonlinearity. The shooting tech-
nique is very accurate while the incremental continuation approach
offers a reasonable degree of accuracy at lower computational cost. d5 1 2ah 1 1 2 = r2
It should be noted that, as presented here, continuation methods
can be applied only to aircraft with linear aerodynamics that can be 1 2ah 1 2ah "1 1 "2 2 = 2r2
written in a rational fraction approximation form, because they are
Downloaded by UNIVERSITY OF CALIFORNIA IRVINE on January 30, 2015 | http://arc.aiaa.org | DOI: 10.2514/1.28759
Vol. 23, No. 7, 2007, pp. 9831011. [23] Millman, D. R., King, P. I., and Beran, P. S., Airfoil Pitch-and-Plunge
doi:10.1016/j.juidstructs.2007.03.006 Bifurcation Behavior with Fourier Chaos Expansions, Journal of
[10] Griewank, A., and Reddien, G. W., Characterization and Computation Aircraft, Vol. 42, No. 2, 2005, pp. 376384.
of Generalized Turning Points, SIAM Journal on Numerical Analysis, [24] Baldelli, D. H., Lind, R. C., and Brenner, M., Robust Aeroelastic
Vol. 21, No. 1, 1984, pp. 176185. Match-Point Solutions Using Describing Function Method, Journal of
doi:10.1137/0721012 Aircraft, Vol. 42, No. 6, 2005, pp. 15971605.
[11] Doedel, E., Keller, H. B., and Kernevez, J. P., Numerical Analysis and [25] Marsden, C. C., and Price, S. J., Transient and Limit Cycle Simulation
Control of Bifurcation Problems (1) Bifurcation in Finite Dimensions, of a Nonlinear Aeroelastic System, Journal of Aircraft, Vol. 44, No. 1,
International Journal of Bifurcation and Chaos in Applied Sciences 2007, pp. 6070.
and Engineering, Vol. 1, No. 3, 1991, pp. 493520. doi:10.2514/1.21367
doi:10.1142/S0218127491000397 [26] Lee, B. H. K., Liu, L., and Chung, K. W., Airfoil Motion in Subsonic
[12] Allgower, E. L., and Georg, K., Numerical Continuation Methods: An Flow with Strong Cubic Nonlinear Restoring Forces, Journal of Sound
Introduction, SpringerVerlag, New York, 1990. and Vibration, Vol. 281, Nos. 35, 2005, pp. 699717.
[13] Broyden, C. G., A Class of Methods for Solving Nonlinear doi:10.1016/j.jsv.2004.01.034
Simultaneous Equations, Mathematics of Computation, Vol. 19, [27] Lee, B. H. K., and Liu, L., Bifurcation Analysis of Airfoil in Subsonic
No. 92, 1965, pp. 577593. Flow with Coupled Cubic Restoring Forces, Journal of Aircraft,
doi:10.2307/2003941 Vol. 43, No. 3, 2006, pp. 652659.
[14] Govaerts, W., Numerical Bifurcation Analysis for ODEs, Journal of [28] Dimitriadis, G., Continuation of Higher Order Harmonic Balance
Computational and Applied Mathematics, Vol. 125, Nos. 12, 2000, Solutions for Nonlinear Aeroelastic Systems, Journal of Aircraft (to be
pp. 5768. published).
doi:10.1016/S0377-0427(00)00458-1 [29] ZAERO Ver. 7.2: Theoretical Manual, ZONA Technology,
[15] Rheinboldt, W. C., Numerical Continuation Methods: A Perspective, Scottsdale, AZ, 2004.
Journal of Computational and Applied Mathematics, Vol. 124, Nos. 1 [30] Karpel, M., Moulin, B., and Chen, P. C., Dynamic Response of
2, 2000, pp. 229244. Aeroservoelastic Systems to Gust Excitation, Journal of Aircraft,
doi:10.1016/S0377-0427(00)00428-3 Vol. 42, No. 5, 2005, pp. 12641272.
[16] Li, M. Y., and Muldowney, J. S., Phase Asymptotic Semiows, [31] Roger, K. L., Airplane Math Modelling Methods for Active Control
Poincaes Condition, and the Existence of Stable Limit Cycles, Design, AGARD, Rept. AGARD-CP-228, 1977.
Journal of Differential Equations, Vol. 124, No. 2, 1996, pp. 425448. [32] Davidson, B., Large Scale Continuation and Numerical Bifurcation for
doi:10.1006/jdeq.1996.0018 Partial Differential Equations, SIAM Journal on Numerical Analysis,
[17] Lara, M., and Peaez, J., On the Numerical Continuation of Vol. 34, No. 5, 1997, pp. 20082027.
Periodic Orbits, Astronomy and Astrophysics, Vol. 389, No. 2, 2002, doi:10.1137/S0036142994273288