Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Optimal Control of Coning Motion of Spinning Missiles

Xuerui Mao 1 and Shuxing Yang 2

Abstract: Coning motion has been detected in flight experiments of many spinning projectiles for decades and the stability analysis of this
motion has been extensively studied recently. In this paper, a novel optimal control algorithm based on stability analyses is developed to
suppress the coning motion. A Lagrangian functional is built to compute the optimal control that minimizes the coning motion for a given
control cost. Then a linearized approach, which produces accurate enough control within the range of parameters considered, is proposed to
Downloaded from ascelibrary.org by Beihang University on 02/02/22. Copyright ASCE. For personal use only; all rights reserved.

reduce the computational cost. It is analytically presented that the optimal solution of this linearized optimization is unique at any given values
of control cost. The analytical analysis also indicates that there exists a direct solution of the control that approaches the linearized optimal
control when the control cost is small enough and that this direct solution can be obtained much more efficiently than the linearized optimal
control. The controllability problem associated with this optimal control algorithm is also discussed and the results shed lights on the choice
of numerical parameters in the optimization. These optimal control algorithms are implemented to calculate the optimal control that
minimizes the coning motion induced by initial disturbances to a wrap-around-fin missile. DOI: 10.1061/(ASCE)AS.1943-5525
.0000398. © 2014 American Society of Civil Engineers.
Author keywords: Nonlinear optimal control; Coning motion.

Introduction (converge or diverge) of the nutation angle when the projectile is


initially disturbed by an infinitesimally small perturbation.
Coning motion refers to the rotation of a projectile around its veloc- While the stability analysis is associated with calculating the
ity vector with a nutation angle or complex angle of attack (Fig. 1 largest eigenvalue of an operator and the corresponding most
shows a schematic illustration). This motion has been detected in unstable mode, the counterpart of this analysis is to suppress the
flight experiments and simulations of spinning rockets/missiles for development of the unstable motion, which has received limited
decades, and it was observed that this motion induces structure de- attention in either analytical or theoretical studies partially owing
structions, or significantly reduces the ranges and precisions of the to the complexity of the modeling of the control system. In this
projectiles. For example, in the United States, unstable coning paper, the detected coning motion to be controlled is treated as
motion was detected 20 times out of 50 flight experiments for the initial condition of the governing equation while the control
Nitehawk Rocket (Curry and Reed 1966), and this motion also ap- in the form of cross-jets can be treated as external forcing of the
peared in the subsonic and supersonic wind tunnel experiments for governing equation. Based on the modeling of the control and the
69.85 mm (2.75-in.) rockets (Nicolaides et al. 1969). In Spain, cata- coning motion, a computationally inexpensive optimal control
strophic coning motion made the velocity reduce 60% in just 1.5 s strategy is proposed to suppress the development of perturbations
in the flight trials of a 140-mm artillery rocket (Morote and Liano and maintain the desired state of the projectile.
2004). In China, serious coning motion has been detected in the The optimal control method, consisting of both direct and indi-
experiments for several types of unguided rockets (Zhang and Yang rect algorithms, has been well-developed and implemented in the
2004). Therefore the dynamic stability of the coning motion and the optimization of trajectories (Betts 1998). At the aim of increasing
control strategy to suppress this motion deserve a thorough accuracy, numerical efficiency, and robustness, some novel numeri-
investigation. cal techniques, such as multiresolution (Jain and Tsiotras 2008),
The stability of the conning motion stemming from side forces pseudospectral (Gong et al. 2008), and spectral element methods
has been theoretically investigated and a stability criterion has been (Darby et al. 2011) have been implemented in the discretization
proposed in the context of wrap-around-fin projectiles (Mao et al. of governing equations. Algorithms featuring inexpensive compu-
2006). This motion can be triggered by the control signals in spin- tation have been proposed (Najson and Mease 2006) to optimize
ning missiles, and the stability of this motion is therefore associated the trajectory for Mars landing vehicles. These optimizations of
with the rate loops (Yan et al. 2010) or attitude autopilots (Yan et al. trajectories can be interpreted as calculating the desired state of
2011). Most of these stability studies on coning motion are based the system while the current paper is concentrated on maintaining
on linearized eigenvalue stability analyses, which suggest the trend the desired state by suppressing the development of perturbations
in the form of initial perturbations or environmental disturbances.
To be consistent with the literature (Mao et al. 2006, 2007)
1
School of Engineering and Computing Sciences, Durham Univ., about stability analyses of the coning motion, a wrap-around-fin
Durham DH1 3LE, U.K. (corresponding author). E-mail: xuerui.mao@ missile with cross-jet attitude control is adopted as the physical
durham.ac.uk model (Fig. 2 shows a schematic). The wrap-around-fin generates
2
School of Aerospace Engineering, Beijing Institute of Technology,
a body-rotating moment which drives the missile to spin (Wu et al.
Beijing 100081, P. R. China.
Note. This manuscript was submitted on May 11, 2012; approved on 1995), and a lateral moment which activates the conning motion
November 12, 2013; published online on November 14, 2013. Discussion (Mao et al. 2006).
period open until December 7, 2014; separate discussions must be sub- The cross jet is provided by an air compressor which has four
mitted for individual papers. This paper is part of the Journal of Aerospace orifices around the cross section A. The orifices are located close
Engineering, © ASCE, ISSN 0893-1321/04014068(10)/$25.00. to the head of the missile and provides cross jets that effectively

© ASCE 04014068-1 J. Aerosp. Eng.

J. Aerosp. Eng., 2015, 28(2): 04014068


" #
control the attitude of the missile. This cross-jet control has been ωy
extensively adopted in trajectory control of projectiles. For exam- x¼ ωz ð1bÞ
ple, Burchett et al. (2002) used a cluster of lateral pulse jets to track θ
a specified trajectory of a dual-spin projectile. Wang et al. (2005) 2 3
studied performance of cross jets stemming from an attitude control M y =J 2 − ωx ωz ðJ 1 − J 2 Þ=J 2
thruster, which has been used successfully in a series of long-range fðxÞ ¼ 4 M z =J 2 þ ωx ωy ðJ 1 − J 2 Þ=J 2 5 ð1cÞ
rockets. Compared with aerodynamic surfaces, the cross-jet control ωz
has many advantages, as follows: (1) the actuators to generate
cross-jets is lighter because it has no moving parts, (2) the " #
uy
cross-jet control is more reliable because many errors such as lock- u¼ uz ð1dÞ
ing and wearing out of the moving parts are avoided automatically, 0
and (3) the cross-jet control responses faster than the aerodynamic
surfaces. where the overhead dot represents the derivative with respect to
time t; J 1 and J 2 = roll and lateral moments of inertia of the nearly
Downloaded from ascelibrary.org by Beihang University on 02/02/22. Copyright ASCE. For personal use only; all rights reserved.

axis-symmetric missile, respectively; uy and uz = thrusts induced


Governing Equations by cross-jet control in OY and OZ directions, respectively;
b ≡ sLc =J 2 , where Lc is the distance from the control jets to
Since the coning motion stemming from initial perturbations can be the center of gravity of the missile; s is a time-dependent factor
suppressed over a short time interval (as demonstrated in the sub- which models the responding speed of the control thrust and is
sequent sections), the velocity V is assumed to be constant. Because related with the structure of the control devices (in this paper,
2
the spinning rate ωx is related with the body-rotating moment gen- s ¼ 1 − e−t is adopted so that the control force starts from zero
erated by the wrap-around-fins and decoupled with the cross-jet and reaches a stable value smoothly); and M y and M z = aerody-
control thrust, ωx is also assumed to be constant. Therefore the gov- namic moments on OY and OZ directions, respectively
erning equations of coning motion with cross-jet control in terms of
angular speeds ωy , ωz and nutation angle θ (defined in Fig. 1) can be ω qSL2r qSL2r
M y ¼ my qSLr þ my y þ mm ωx ð2aÞ
expressed as (Mao et al. 2006) V V

ω qSL2r
M z ¼ mz qSLr þ mz z ð2bÞ
ẋ ¼ fðxÞ þ bu ð1aÞ V
where q, S, and Lr = dynamic pressure, reference area, and refer-
ω ω
ence length, respectively; and my , my y , mm , mz , and mz z = lateral
moment, lateral-damping moment, Magnus moment, pitching
moment, and pitch-damping moment coefficients, respectively.
These coefficients are obtained from wind-tunnel measurements of
wrap-around-fin missiles (Miao et al. 1996), and have been used
successfully in overall designs of missiles, trajectory simulations,
and stability analyses of the coning motion (Lei and Wu 2005; Mao
et al. 2006, 2007).
In the stability analysis, the state vector x is decomposed into a
summation of a base state X and a perturbation state x 0 , that is
x ¼ X þ x0 ð3Þ

Substitute Eq. (3) into the governing Eqs. (1a)–(1d) and linear-
ize around the base state X to reach the linearized governing Eq. (4)

Fig. 1. Schematic of the coning motion; O denotes the center of gravity ẋ 0 ¼ Ax 0 þ bu ð4Þ
of the projectile; OX represents the longitudinal axis; OV denotes the
where A is a linearized operator and a function of the base state X
velocity direction; OY is the vertical axis perpendicular to OX in the
(Mao et al. 2006). The eigenvalue stability analysis of the operator
VOX plane; OZ (which is not shown for clarity) is perpendicular to
A suggests that there exists a critical value for the spinning rate,
the XOY plane, and preserves the right-hand-rule with OX and OY;
denoted as wxc . For wx < wxc, all the eigenvalues of A have nega-
θ is the nutation angle between OV and OX; and ωy and ωz are angular
tive real parts and so this system is asymptotically stable to infini-
speeds in OY and OZ directions, respectively
tesimally small perturbations while for wx > wxc, at least one

Fig. 2. Schematic of a typical wrap-around-fin missile; the attitude of the missile is controlled by cross jets

© ASCE 04014068-2 J. Aerosp. Eng.

J. Aerosp. Eng., 2015, 28(2): 04014068


eigenvalue of A has positive real parts and so some form of per- δL½δxðTÞ ¼ 0; ⇒ x ðTÞ ¼ 0 ð8cÞ
turbations will grow asymptotically in the linear regime to unsta-
bilize the coning motion (Mao et al. 2006, 2007; Yan et al. 2010, δLðδλÞ ¼ 0; ⇒ ðu; uÞ ¼ C ð8dÞ
2011). After the asymptotical growth saturates, the system con-
Eqs. (8a)–(8d) are the recovery of the governing Eqs. (1a)–(1d),
verges to another stable state with a nonzero nutation angle.
which evolves the initial perturbation and control from t ¼ 0 to
In this paper a zero base state X ¼ ½0; 0; 0T , which is a solution
t ¼ T. Eq. (8b) is the adjoint equation, where ∂f=∂x is the Jacobian
of the governing Eqs. (1a)–(1d), is adopted. Therefore the state vec-
of the operator f and its entries are defined as
tor is equal to the perturbation vector x ¼ x 0, so in the subsequent
 
text the prime superscript is dropped for clarification. The geometry ∂f ∂f
and aerodynamic coefficients adopted in this paper were originally ¼ i ; with 1 ≤ i; j ≤ 3 ð9Þ
∂x ij ∂xj
obtained from wind tunnel experimental measurements for wrap-
around-fin rockets (Miao et al. 1996) and have been used success- where fi (1 ≤ i ≤ 3) is the ith component of f.
fully in the overall design of long-range rockets, the computation of Eq. (8c) provides the values of adjoint variables at time t ¼ T,
trajectories, and stability analyses of the coning motion (Lei and Wu and therefore indicates that Eq. (8b) should be integrated back-
Downloaded from ascelibrary.org by Beihang University on 02/02/22. Copyright ASCE. For personal use only; all rights reserved.

2005; Mao et al. 2006). The Mach number is set to M ¼ 4. At the wards from t ¼ T to t ¼ 0 and initialized by Eq. (8c). Eq. (8d)
combination of these parameters, the critical spinning rate is wxc ¼ is the recovery of the constraint on the control cost.
7.44 rad=s. To illustrate the control of an unstable coning motion, a In accordance with the definition of the Gateau differential dL
spinning rate in the unstable range ωx ¼ 10 rad=s is adopted. of the Lagrangian, the gradient of the Lagrangian with respect to
the control u, denoted as ∇u L, satisfies
Methodology of Optimal Control Lðu þ ϵδuÞ − LðuÞ
dLðδuÞ ¼ lim ¼ ð∇u L; δuÞ ð10Þ
ϵ→0 ϵ
Nonlinear Optimal Control of the Coning Motion
Therefore
The conning motion stemming from environmental disturbances is
initially excited by ωz since the nonzero value of θ is generated by ∇u L ¼ ½−2λuy − bx1 ; −2λuz − bx2 ; 0T ð11Þ
an integration of ωz and ωy ¼ 0 at θ ¼ 0. Therefore the basic strat-
egy is using control u generated by cross jets to suppress the coning In the optimization process, the control u is updated from step k
motion induced by the initial perturbation xð0Þ ¼ ½0; ωz ð0Þ; 0T . to step k þ 1 as
The objective function to be minimized, that is the magnitude of
the coning motion, is modeled as ukþ1 ¼ uk þ β k Pð∇u LÞk ð12Þ

L ¼ ðθ; θÞ ð5Þ where λ is chosen so that the new control satisfies the magnitude
constraint ðukþ1 ; ukþ1 Þ ¼ C; Pð∇u LÞk = search direction (a
where the scale product is defined as ða; bÞ ¼ ∫ T0 a · bdt, where Fletcher-Reeves conjugate gradient method is adopted to calculate
T = terminal time. The choice of T will be discussed in a sub- this direction); and β k = step length which satisfies the strong Wolf
sequent section. condition
The Lagrangian functional can be correspondingly defined as
Lkþ1 ≤ Lk þ c1 β k ð∇Lk ; P k Þ ð13aÞ
L ¼ L þ ðx ; ẋ − f − buÞ þ λ½C − ðu; uÞ ð6Þ
jð∇Lkþ1 ; P k Þj ≤ c2 jð∇Lk ; P k Þj ð13bÞ
where λ is a Lagrangian multiplier; x ¼ ½x1 ; x2 ; x3 T = adjoint
state; the first term on the right is the objective function; the second where c1 and c2 are constant, and satisfy 0 < c1 dt < c2 < 0.5, with
term is the constraint of the governing equations; and the last term dt denoting the discretized time step. A backtracking line search
is the constraint on the control cost, where C = prescribed control method is adopted to choose an acceptable value of β which is ini-
cost. There is a closely related optimization problem, that is to min- tially set to 1. It has been checked that the result is insensitive to the
imize the control cost with the constraint θðTÞ ¼ 0. This problem is choice of initial value of β.
not considered because even when the nutation angle is controlled Since θ is bounded (θ < 2π), L is also bounded, L ¼ ðθ; θÞ <
to reach zero at a fixed time horizon it could still diverge at a later ð2πÞ2 T. Because the adjoint equation is linear, x1 and x2 are linear
time. The coning motion is observed to diverge slowly and only one functions of x and subsequently the gradient ∇u L, as defined in
or two optimal control process is required during the flight, which Eq. (11), is also bounded. Therefore the Zoutendijk condition is
means the minimization of the coning motion is more meaningful satisfied and so the sequence of the gradient is bounded to zero
than the minimization of the control cost.
Apply integration by parts to the second term to obtain lim infð∇Lk ; ∇Lk Þ ¼ 0 ð14Þ
k→∞

L ¼ L þ x ðTÞ · xðTÞ − x ð0Þ · xð0Þ − ðẋ ; xÞ − ðx ; f þ buÞ which guarantees the convergence of the gradient (Nocedal and
þ λ½C − ðu; uÞ ð7Þ Wright 1999).
As a summary, the six-step procedure of this optimization of
At the minimizer of the Lagrangian functional, the first variation control can be outlined as follows:
of this function with respect to its variables x , x, xðTÞ, and λ are 1. Evolve a random guess of control u with magnitude ðu; uÞ ¼
zero, so C forwards from t ¼ 0 to t ¼ T in Eq. (8a), which is initialized
by the given initial state xð0Þ ¼ ½0; ωz ð0Þ; 0T , and record the
δLðδx Þ ¼ 0; ⇒ ẋ − f − bu ¼ 0 ð8aÞ full trajectory of the state xðtÞ;
 T 2. Initialize the adjoint Eq. (8b) with Eq. (8c) and evolve the
 ∂f adjoint state variables backwards from t ¼ T to t ¼ 0;xðtÞ
δLðδxÞ ¼ 0; ⇒ − ẋ − x þ ½0; 0; 2θT ¼ 0 ð8bÞ
∂x saved in Step 1 is required in Step 2;

© ASCE 04014068-3 J. Aerosp. Eng.

J. Aerosp. Eng., 2015, 28(2): 04014068


Downloaded from ascelibrary.org by Beihang University on 02/02/22. Copyright ASCE. For personal use only; all rights reserved.

Fig. 3. Sketch of the nonlinear optimization algorithm

3. Calculate the gradient of the Lagrangian with respect to u is obtained. Setting the first variation of the Lagrangian with respect
using Eq. (11); to xðTÞ, λ, and u to zero, one obtains the initial condition of the
4. Update the control and scale it to satisfy the constraint on adjoint equations, the constraint on control cost, and the gradient of
control cost using Eq. (12); the Lagrangian functional with respect to u. They are all in the
5. Evolve the updated control u in Eq. (8a); if the Wolf condition same form as discussed in the nonlinear optimization, as given
is satisfied, turn to Step 6, and otherwise reduce the step length in Eqs. (8c), (8d), and (11), respectively.
β, and turn to Step 4; and From the adjoint Eq. (16) only θ is required and the other two
6. Repeat Steps 2–5 until the control converges. state variables (ωy and ωz ) are absent. Therefore the memory and
Fig. 3 shows a flowchart of this nonlinear optimization pro- computational cost of the forward evolution can be significantly
cedure, where the two iterative loops (highlighted by circles) are reduced by only calculating and saving θðtÞ.
as follows: (1) to update the control, (2) to check the satisfaction The solution of linearized governing Eq. (4) is in the form
of the Wolf condition. Z t
xðtÞ ¼ eAt xð0Þ þ eðt−τ ÞA budτ
0
Linearized Optimal Control of the Coning Motion
In the nonlinear optimization presented previously, three governing Recall that x ¼ ½ωy ; ωz ; θT and u ¼ ½uy ; uz ; 0T . Owing to the
equations are solved in the forward integration (from t ¼ 0 to linear nature of Eq. (4), the contribution of the initial condition xð0Þ
t ¼ T) and the full trajectory of xðtÞ has to be saved for solving and control u to the nutation angle θ are decoupled. Therefore
the three adjoint equations. The constraint of the Wolf condi-
θ ¼ θ0 þ θu ð17aÞ
tion on the step length β results in multiple forward integrations
to obtain an acceptable step length. The requirements of computa-
θ0 ¼ D0 xð0Þ ð17bÞ
tional time and memory in this nonlinear optimization may prevent
obtaining an optimal control signal instantly when a perturbation is
θu ¼ Du u ð17cÞ
observed, or when the coning motion is detected.
Considering the perturbations are small in magnitude when the where θ0 = nutation angle induced by a fixed initial condition xð0Þ;
coning motion is observed in practical flight trials, the linearized θu = nutation angle generated by the control u; and D0 ðtÞ and
governing equations around the zero base state can predict the evo- Du ðtÞ are the development operators of the initial condition
lution of the state variables reasonably accurately. In this section, a xð0Þ and control u to θ, respectively.
linearized optimization algorithm is developed by using the linear- θ0 is constant during the optimization process and the time
ized governing Eq. (3), rather than the nonlinear governing sequence of θu can be obtained as
Eqs. (1a)–(1d) as a constraint. This linearized optimization signifi-
cantly reduces the computational cost and is capable to generate 2 3 2 32 3
θu ðdtÞ d1 0 ··· 0 uy ðdtÞ
optimal control signals instantly in flight experiments. 6 7 6 76 7
Replacing the nonlinear governing equations with the linearized 6 θu ð2dtÞ 7 6 d2 d1 ··· 0 76 uy ð2dtÞ 7
6 7 6 76 7
governing Eq. (6), the Lagrangian functional in the linearized op- 6 .. 7¼6 . .. .. .. 76 .. 7
6 . 7 6 .. . . . 76 . 7
timization can be expressed as 4 5 4 54 5
θu ðTÞ dN dN−1 · · · d1 uy ðTÞ
L ¼ L þ ðx ; ẋ − Ax − buÞ þ λ½C − ðu; uÞ ð15Þ 2 32 3
e1 0 ··· 0 uz ðdtÞ
Setting the first variation of the Lagrangian functional with re- 6 76 7
6 e2 e1 · · · 0 76 uz ð2dtÞ 7
spect to u and x to zero, the linearized governing Eq. (4) and the 6 76 7
þ6 . .. .. .. 7 6 .. 7 ð18Þ
corresponding adjoint Eq. (16) 6 .. . . 7
. 546 . 7
4 5
ẋ ¼ −AT x þ ½0; 0; 2θT ð16Þ eN eN−1 ··· e1 uz ðTÞ

© ASCE 04014068-4 J. Aerosp. Eng.

J. Aerosp. Eng., 2015, 28(2): 04014068


where di and ei (1 ≤ i ≤ N) are obtained by X
3
gi ¼ − 2Qj2 eiΛj dt Q−1
3j dt ð22bÞ
X
3 j¼1
di ¼ Q3j eiΛj dt Q−1
j1 bdt; 1≤i≤N ð19aÞ
j¼1 Because eΛj dt is constant, hi and gi can be calculated efficiently
in accordance with recursion relations before the first iteration, and
X
3 do not need to be updated afterwards. The other adjoint variable x3
ei ¼ Q3j eiΛj dt Q−1
j2 bdt; 1≤i≤N ð19bÞ is not solved since it is not used in the subsequent steps.
j¼1 The major difference between the linearized and nonlinear
optimizations is that in the linearized optimization there exists
where N ¼ T=dt is the dimension of the time discretized variables; an optimal step length that can be calculated efficiently before each
Λi (1 ≤ i ≤ 3) are the eigenvalues of A; and the columns of Q are update, and therefore the loop to compute an acceptable β that sat-
the corresponding eigenvectors. Because eΛj dt is constant, di and ei isfies the Wolf condition in the nonlinear optimization is removed.
can be calculated efficiently in accordance with recursion relations This difference results in a significant reduction of the computa-
and saved before the first iteration without being updated after- tional cost. The calculation of this optimal step length is presented
Downloaded from ascelibrary.org by Beihang University on 02/02/22. Copyright ASCE. For personal use only; all rights reserved.

wards. The other two state variables ωy and ωz are not solved since in Appendix.
they are not going to be used in the subsequent steps. As a summary, the procedure of the linearized optimization of
In the adjoint equations only x1 and x2 are used in the optimi- control can be outlined in nine steps, as follows:
zation process [Eq. (11)], and analogously researchers can reduce 1. Set the control to zero and evolve the initial disturbance xð0Þ
the computational cost of the adjoint integration by only calculating using Eq. (4) to obtain θ0 ¼ D0 xð0Þ;
x1 and x2 . 2. Set the initial disturbance to zero and evolve a random guess of
Starting from the solution of the adjoint Eq. (16) control u satisfying the constraint on control cost to obtain the
Z T time sequence of θu using Eq. (18);
x ðtÞ ¼ − eðτ −tÞA ½0; 0; 2θT dτ
T
ð20Þ 3. Calculate the adjoint variables through Eqs. (21a) and (21b);
t 4. Calculate the gradient ∇u L from Eq. (11);
5. Calculate the search direction P~ in accordance with the con-
the time sequence of the adjoint variables can be subsequently jugate gradient method and then compute the modified search
written as direction P~ from Eq. (39) in Appendix;
2  3 2 32 3 6. Substitute P~ as u in Eq. (18) to obtain Du P; ~
x1 ðdtÞ h1 h2 ··· hN 2θðdtÞ 7. Calculate the optimal step length β which minimizes L as
6 x ð2dtÞ 7 6 0 hN 7 6 7
6 1 7 6 h1 ··· 76 2θð2dtÞ 7 detailed in Appendix;
6 .. 7¼6 . .. .. 7
.. 766 .. 7 ð21aÞ
6 7 6 . 7 8. Update the control and scale it to satisfy the constraint on
4 . 5 4 . . . . 54 . 5 control cost using Eq. (40), in Appendix I; correspondingly
x1 ðTÞ 0 0 ··· h1 2θðTÞ update the time sequence of θu ; and
9. Repeat Steps 3–8 until the control converges.
2 3 2 32 3
x2 ðdtÞ g1 g2 ··· gN 2θðdtÞ Fig. 4 shows a flowchart of this linearized optimization pro-
6 x ð2dtÞ 7 6 0 gN 7 6 7 cedure. Compared with the flow chart of nonlinear optimization
6 2 7 6 g1 ··· 76 2θð2dtÞ 7
6 .. 7¼6 . .. .. 7
.. 766 .. 7 ð21bÞ (Fig. 3), there is only one iterative loop in this procedure, which
6 7 6 . 7
4 . 5 4 . . . . 54 . 5 is the update of the control. The loop to choose a step length to
satisfy the Wolf condition is removed since an optimal step length
x2 ðTÞ 0 0 ··· g1 2θðTÞ
can be calculated analytically.
where the entries hi and gi (1 ≤ i ≤ N) are
Uniqueness of the Optimal Control
X
3
hi ¼ − 2Qj1 eiΛj dt Q−1
3j dt ð22aÞ The optimization discussed previously computes the optimal con-
j¼1 trol as a minimizer of the Lagrangian functional. This minimizer

Fig. 4. Sketch of linearized optimization algorithm

© ASCE 04014068-5 J. Aerosp. Eng.

J. Aerosp. Eng., 2015, 28(2): 04014068


can be a global minimizer or a local minimizer. In this section, it is Projecting the discretized form of DTu θ0 and u onto the space vi
demonstrated that the Lagrangian functional has only one mini- (1 ≤ i ≤ 3N)
mizer, which is therefore the global minimizer even though this
function is not convex with respect to the control. X
3N
D̂Tu θ̂0 ¼ qi v i ð30aÞ
First express the Lagrangian as a function of u i¼1

L ¼ ðθ0 þ Du u; θ0 þ Du uÞ þ λ½C − ðu; uÞ ð23Þ


θ̂0 ¼ ½θ0 ðdtÞ; θ0 ð2dtÞ; : : : ; θ0 ðNdtÞT ð30bÞ
where the constraint of the governing equation is replaced by the X
3N
action of the evolution operator Du. û ¼ ri vi ð30cÞ
Then the first variation of the Lagrangian with respect to the i¼1
control can be formulated as
û ¼ ½uðdtÞ; uð2dtÞ; : : : ; uðNdtÞT ð30dÞ
δLðδuÞ ¼ ð2DTu θ0 þ 2DTu Du u − 2λu; δuÞ ð24Þ
Downloaded from ascelibrary.org by Beihang University on 02/02/22. Copyright ASCE. For personal use only; all rights reserved.

At the equilibrium point where the gradient of this function with Substitute these decomposed forms into Eqs. (25a) and (25b) to
respect to the control vanishes this variation is zero and therefore obtain

DTu θ0 þ DTu Du u ¼ λu ð25aÞ qi þ λi ri ¼ λri ; 1 ≤ i ≤ 3N ð31Þ


P3N 2
Considering the constraint on control, C ¼ ðu; uÞ ¼ i¼1 ri
DT θ þ DTu Du u; u
λ¼ u 0 ð25bÞ
ðu; uÞ X
3N
q2i
−C¼0 ð32Þ
i¼1
ðλ − λi Þ2
To verify if this point is an extreme point (local or global) or an
inflection point, inspect the second variation Then define a function
X
3N
q2i
δ 2 LðδuÞ ¼ 2ðDTu Du δu; δuÞ − 2λðδu; δuÞ ð26Þ FðσÞ ¼ −C ð33Þ
i¼1
ðσ − λi Þ2
In the subsequent text, the overhead symbol is used to denote the
time-discretized forms of operators or vectors. For example D̂u rep- Combing Eqs. (32) and (33), FðλÞ ¼ 0. This function FðσÞ is
resents the time discretized matrix form of Du . Then the joint monotonic in the range −∞ < σ ≤ λ1 , and Fð−∞Þ ¼ −C and
operator DTu Du can be discretized as a matrix D̂Tu D̂u . This matrix Fðλ1 Þ ¼ ∞. Therefore the function FðσÞ has only one root for
is self-adjoint and has 3N orthogonal eigenvectors, denoted −∞ < σ ≤ λ1 and this root is λ. From Eq. (31), a value of λ cor-
as vi , with vi · vi ¼ 1, 1 ≤ i ≤ 3N. These eigenvectors form a responds to a unique sequence of ri [Eq. (31)] and so a unique
complete base in the space R3N and the corresponding eigenvalues solution of the control u. Therefore it is proved that for each given
λi satisfy control cost C, there exists a unique optimal control u.
As C increases, r1 , the weight of v1 , increases faster in magni-
D̂Tu D̂u vi · vi ¼ λi vi · vi ¼ λi ; 1 ≤ i ≤ 3N ð27Þ tude than other ri and so the optimal solution approaches r1 v1 .
Because v1 is discretization dependent it can be expected that
D̂Tu D̂u vi · vi ¼ ðD̂u vi Þ · ðD̂u vi Þ ≥ 0. Therefore λi ≥ 0 the solution converges slow at large values of C. The coning motion
(1 ≤ i ≤ 3N). Without loss of generality, it is assumed that has been mostly suppressed at values of C lower than the poor-
0 ≤ λ1 ≤ λ2 ; : : : ; λ3N−1 ≤ λ3N . convergence threshold as discussed in a subsequent section.
The discretized variation of the control can be decomposed as a Correspondingly it can be verified that the Lagrangian also has a
summation of the eigenvectors vi unique maximizer. q1 ¼ 0 is a singular point but this singularity
does not change the statements made previously. Theoretically
X
3N
δ û ¼ pi v i ð28aÞ the optimization procedure introduced previously does not guaran-
i¼1
tee that the optimal control converges to the unique global mini-
mizer; the optimization could also terminate at an inflection
δ û ¼ ½δuðdtÞ; δuð2dtÞ; : : : ; δuðNdtÞT ð28bÞ point, which corresponds to λ1 < λ < λ3N . However, the inflection
points are sensitive to disturbances and the (global) minimizer is the
only robust solution. Therefore the fake solutions (inflection
Then
points) can be excluded by disturbing the solution after conver-
X
3N
gence and checking if the solution is robust. For the parameters
δ 2 LðδuÞ ¼ 2 ðλi − λÞp2i ð29Þ considered in this paper, the nonlinear optimal control solution
i¼1
does not deviate from the linearized optimal control solution sig-
If λ ≥ λ3N , δ 2 LðδuÞ ≤ 0 and the corresponding control is a nificantly and therefore the same convergence check is imple-
maximizer of the Lagrangian functional; if λ ≤ λ1 , δ2 LðδuÞ ≥ 0 mented to the nonlinear optimizations.
and the corresponding control is a minimizer of the Lagrangian If C is small enough, the term DTu Du u in Eqs. (25a) and (25b) is
functional; if λ1 < λ < λ3N , the sign of δ 2 LðδuÞ is undetermined negligible compared with DTu θ0 and therefore the discretized opti-
and the corresponding control is an inflection point of the Lagran- mal control u has opposite direction with DTu θ0 . This observation
gian functional. λ can be considered as a function of the control u suggests that at small values ofpcontrol cost the optimal control can
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
[Eqs. (25a) and (25b)]. In the subsequent text, it is demonstrated be obtained directed as −DTu θ0 C=ðDTu θ0 ; DTu θ0 Þ and no iterations
that for a given control cost there is a unique solution u satisfy- are required. This solution is denoted as the direct solution in the
ing λðuÞ ≤ λ1 . subsequent text.

© ASCE 04014068-6 J. Aerosp. Eng.

J. Aerosp. Eng., 2015, 28(2): 04014068


At the minimizer of Lun , the optimal control is obtained

Fig. 5. Sketch of the direct solution scheme ðθu ; θ0 Þ ðu; DT θ Þ


uable ¼ − u¼− T u 0 u ð36Þ
ðθu ; θu Þ ðDu Du u; uÞ

Fig. 5 shows a flowchart for this direct solution, where the and the corresponding control cost Cable
loops appearing in the flowcharts of linearized and nonlinear
ðu; DTu θ0 Þ2
optimization algorithms have been removed, and no iterative circles Cable ¼ ðuable ; uable Þ ¼ ðu; uÞ ð37Þ
are involved. ðDTu Du u; uÞ2
By monitoring the CPU time elapsed in the simulation, the lin-
earized optimization consumes less than 10% of the CPU time used The gradient of the Lagrangian functional with respect to u is
in the nonlinear optimizations while the direct solution costs less
2ðu; DTu θ0 Þ2 T 2ðu; DT θ Þ
than 20% computational time used in the linearized optimization. ∇u Lun ¼ 2
Du Du u − T u 0 DTu θ0 ð38Þ
ðDu Du u; uÞ
T ðDu Du u; uÞ
Downloaded from ascelibrary.org by Beihang University on 02/02/22. Copyright ASCE. For personal use only; all rights reserved.

Controllability Analysis The optimization procedures are similar to that discussed in a


subsequent section.
After demonstrating the optimal control strategy at fixed control
Adopting ωz ð0Þ ¼ 0.01 rad=s and all the other parameters as
costs, the controllability problem, that is how much of the coning
used previously, to minimize the coning motion over time T, the
motion can be suppressed by cross-jet control without any con-
control cost Cable decreases with T while Lun , denoting the coning
straint on the control cost in the linear framework, needs to be stud-
motion that cannot be controlled, is always in negligible level for
ied. In other words, this controllability problem is equivalent to
the range of parameters considered (Fig. 6). Therefore it can be
calculating the linearized optimal control across all values of the
concluded that a larger value of T reduces the control cost without
control cost.
sacrificing the control effect significantly but apparently requires
To remove the effects of control cost, one projects the time se-
more computational time.
quence of θ0 onto θu and define
At C ¼ Cable , the optimal control obtained from the linearized
optimization is the same as the global optimal control uable calcu-
ðθu ; θ0 Þ ðDu u; θ0 Þ lated from this controllability analysis. At this global optimal sol-
θable ¼ θ ¼ D u ð34Þ
ðθu ; θu Þ u ðDu u; Du uÞ u ution across all values of the control cost both the gradients of L
and Lun vanish. It can be demonstrated through standard algebraic
manipulations that this control cost Cable corresponds to λ ¼ 0.
which represents the nutation angle that can be suppressed by con- Therefore it suggests that any control cost exceeding Cable , corre-
trol with shape u. Owing to the linear nature of the operator Du, the sponding to 0 < λ ≤ λ1 , is wasted since according to the control-
magnitude of u or the control cost is irrelevant in this definition. lability analysis the extra control cost C − Cable does not help to
The coning motion that cannot be suppressed can be subsequently suppress the coning motion.
expressed as

ðu; DTu θ0 Þ2 Numerical Simulation of the Optimal Control


Lun ðuÞ ¼ ðθ0 − θable ; θ0 − θable Þ ¼ ðθ0 ; θ0 Þ − ð35Þ Algorithms
ðDTu Du u; uÞ
Fig. 7 shows the results of linearized and nonlinear optimizations at
Lun is only a function of the shape of u and independent of the various values of control costs. The terminal time is set to T ¼ 10 s.
magnitude of u, or the control cost. As the increase of the control cost C, L, which quantifies the
The minimum value of Lun is the coning motion that cannot be magnitude of the controlled coning motion, decreases mildly at
suppressed by the thrust control regardless of the shape and mag- the beginning and then drops sharply to negligible levels for all
nitude of the control. Therefore the controllability problem can be the initial perturbations considered. The linear and nonlinear
transferred to calculating u that minimizes the function Lun . results are indistinguishable until the initial disturbance reaches

10-7 1014

1012
10-8
Cable (N2s)
Lun (rad2s)

10
10
-9
10
108

10-10
106

10-11 104
10-1 100 101 10-1 10 0
101
(a) T (s) (b) T (s)

Fig. 6. (a) Remaining of the coning motion that cannot be suppressed; (b) optimal control cost Cable obtained at time interval T

© ASCE 04014068-7 J. Aerosp. Eng.

J. Aerosp. Eng., 2015, 28(2): 04014068


ωz(0)=0.005rad/s, linear
10-3
+ ωz(0)=0.005rad/s, nonlinear
ωz(0)=0.01rad/s, linear
10
-4 x x x ωz(0)=0.01rad/s, nonlinear

L(rad2s)
+ + x ωz(0)=0.02rad/s, linear
10-5
ωz(0)=0.02rad/s, nonlinear
+
x ωz(0)=0.05rad/s, linear
10-6
+ ωz(0)=0.05rad/s, nonlinear
x
-7
10
10-4 10-2 2
100 102
C(N s)
Downloaded from ascelibrary.org by Beihang University on 02/02/22. Copyright ASCE. For personal use only; all rights reserved.

Fig. 7. Minimization of the coning motion at fixed control cost in linear and nonlinear optimizations

1
0
direct solution
0.8 2
C= 6 (N s)
2
C=15 (N s) -0.1
C=20 (N2s)
0.6
uy (N)

uz (N) -0.2 direct solution


2
0.4 C= 6 (N s)
C=15 (N2s)
2
-0.3 C=20 (N s)
0.2

0 -0.4
0 2 4 6 8 10 0 2 4 6 8 10
(a) t (s) (b) t (s)

Fig. 8. Optimal control for ωz ð0Þ ¼ 0.01 rad=s [the control has been normalized so that ðu; uÞ ¼ 1]: (a) uy component; (b) uz component

ωz ð0Þ ¼ 0.05 rad=s. Since the environmental disturbance initializ- ðu; uÞ ¼ 1. The control thrust uz ðtÞ, which directly reduces the nu-
ing the coning motion is commonly small in magnitude, it suggests tation angle in the nutation surface VOX, is much smaller than
that the linearized optimization based on perturbation analyses is an uy ðtÞ, indicating that the intuitive control acting on the nutation
accurate alternative to the full nonlinear optimization. surface is not efficient. Since uz ðtÞ directly reduces the nuta-
Figs. 8(a and b) show two components of the optimal control tion angle and uy ðtÞ reduces the rotating speed of the missile
uy ðtÞ and uz ðtÞ. To compare the time variation of the control at around the velocity vector, these results suggest that reducing
various control costs, the control has been normalized so that the rotating speed ωy is more efficient than reducing the nuta-
tion angle θ directly. As expected, the optimal control at small
0.01 control costs [C ¼ 6 N2 · s in Figs. 8(a and b)] overlaps with
uncontrolled the direct solution. At large values of control costs the optimal
2
0.008 C=1 (N s) control becomes significantly wavy and deviates from the direct
2
C=6 (N s) solution.
2

0.006
C=15 (N s) To demonstrate the control effects of the optimal control u
C=20 (N2s) calculated previously, the evolution of the coning motion at three
θ (rad)

0.004
values of control cost is tested (Fig. 9). As the control cost increases
the divergence speed of the nutation angle is reduced significantly.
When the control cost reaches 15 N2 · s, the coning motion is
0.002
completely suppressed except the oscillation at the beginning
time. However, owing to the reaction time of the control devices
0 (modeled as s in a previous section) the oscillation of the nutation
0 2 4 6 8 10 angle at small time cannot be suppressed effectively. Based on
t (s) these numerical observations, a model-free control law can be
designed to generate effective control without solving the gov-
Fig. 9. Developments of the controlled and uncontrolled coning
erning equations or adjoint equations when the coning motion is
motions
detected.

© ASCE 04014068-8 J. Aerosp. Eng.

J. Aerosp. Eng., 2015, 28(2): 04014068


Conclusion function of ∇u L. This search direction can be partitioned into
two parts, (1) parallel with uk , and (2) normal to uk . The first part
Coning motion has been observed in several numerical and exper- does not change the shape of uk and can be removed from the
imental investigations of spinning missiles in recent years. Instead search direction. Therefore a new search direction can be obtained
of studying the stability of the coning motion as extensively re- as
ported in the literature, in this paper optimal control algorithms
are proposed to suppress this motion. Three optimization algo- ðP k ; uk Þ
P~ k ¼ P k − k k uk ð39Þ
rithms, including a (1) nonlinear optimization, (2) linearized opti- ðu ; u Þ
mization, and (3) direct approach, are developed to optimally
suppress the coning motion activated by initial perturbations using Then the control at step k þ 1 can be expressed as a function of
cross-jet thrusts, and numerical simulations are conducted to cal- the step length β
culate the optimal control and verify the control effects. These op- sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
timization algorithms can be extended to suppress the development C
of initial perturbations in order to maintain or resume the desired
kþ1
u ðβÞ ¼ ðuk þ β P~ k Þ ð40Þ
ðu ; u Þ þ ðP~ k ; P~ k Þβ 2
k k
Downloaded from ascelibrary.org by Beihang University on 02/02/22. Copyright ASCE. For personal use only; all rights reserved.

state in general asymptotically unstable systems.


From the nonlinear optimization, which uses the governing
equations of the coning motion as a constraint, an optimal control As discussed previously, the contributions of the initial condi-
in the form of thrusts generated by cross jets can be calculated when tion and control to the nutation angle are decoupled owing to the
the gradient of the Lagrangian functional with respect to the control linearization of the governing equations. Therefore the optimal step
vanishes. This algorithm produces an accurately optimal control length is the value of β that minimizes
but involves two iterative loops, (1) to update the control, and
Lkþ1 ¼ ðθkþ1 ; θkþ1 Þ ¼ ½θ0 þ Du ukþ1 ðβÞ; θ0 þ Du ukþ1 ðβÞ ð41Þ
(2) to calculate an acceptable step length.
The linearized optimization, which uses the linearized governing
The superscript k is omitted in the subsequent discussion for
equations as a constraint, is observed to produce accurate enough
clarification. Lkþ1 can be formulated as a function of β
control in the range of initial perturbations considered in this paper.
This linearized optimization involves only one loop and only one pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ða6 þ a2 βÞ 1 þ a1 β 2 þ a3 þ a4 β þ a5 β 2
linearized governing equation, and two adjoint equations are solved. L ðβÞ ¼
kþ1 þ ðθ0 ; θ0 Þ
Therefore the linearized optimization requires much less computa- 1 þ a1 β 2
tional time than the original nonlinear optimization. ð42Þ
Through inspecting the second variation of the Lagrangian func-
tional with respect to the control, it is analytically demonstrated that where
this function has only one minimizer even though it is not convex.
~ PÞ=C
a1 ¼ ðP; ~ ð43aÞ
Therefore the optimal solution obtained following the linearized
optimization process is ensured to be the global optimal control.
This analytical analysis of the second variation also suggests that ~
a2 ¼ 2ðθ0 ; Du PÞ ð43bÞ
there exists a direct solution that duplicates the optimal solution at
small values of control cost and no iteration is involved in the cal- a3 ¼ ðDu u; Du uÞ ð43cÞ
culation of this direct solution.
From the controllability analysis, the optimal control without ~
a4 ¼ 2ðDu u; Du PÞ ð43dÞ
any constraint on control cost can be obtained. The results of this
analysis indicate that there exists a global optimal control cost, ~ Du PÞ
a5 ¼ ðDu P; ~ ð43eÞ
which should not be exceeded in the design of control strategies
since any control energy exceeding this value does not contribute a6 ¼ 2ðθ0 ; Du uÞ ð43fÞ
to the suppression of the coning motion. At larger time intervals for
integrations the global optimal control cost reduces without deterio-
rating the control effects significantly. At the optimal value of β, ðdLkþ1 =dβÞ ¼ 0. Through standard
In the numerical simulation, the solutions of the three optimi- algebraic manipulations
zation algorithms are observed to be capable to suppress the coning
motion induced by initial perturbations significantly, the linearized c4 β 4 þ c 3 β 3 þ c 2 β 2 þ c 1 β þ c 0 ¼ 0 ð44Þ
solution is a good approximation of the nonlinear optimal solution
for the magnitude of initial perturbations considered, and the direct where
solution overlaps with the linearized solution when the control cost
c0 ¼ a24 − a22 ð45aÞ
is small enough. The distribution of the optimal control indicates
that the most efficient control is concentrated in the uy component
c1 ¼ −4a1 a3 a4 þ 4a4 a5 þ 2a1 a2 a6 ð45bÞ
rather than uz , which directly reduces the nutation angle. The lin-
earized optimization consumes less than 1=10 computational time
c2 ¼ ð2a1 a3 − 2a5 Þ2 − 2a1 a24 − a1 a22 − a21 a26 ð45cÞ
used in the nonlinear optimization while the direct approach is 5×
more efficient than the linearized approach.
c3 ¼ 2a1 ð2a1 a3 a4 − 2a4 a5 þ a1 a2 a6 Þ ð45dÞ

Appendix. Optimal Step Length in the Linearized c4 ¼ a21 ða24 − a1 a26 Þ ð45eÞ
Optimization

A Fletcher-Reeves conjugate gradient method (Nocedal and Wright Therefore there are five candidates for the optimal step length,
1999) is adopted to calculate the search direction P k , which is a β 1 ∼ β 5 , where β 1 ∼ β 4 are the solutions of Eq. (44) and β 5 ¼ ∞.

© ASCE 04014068-9 J. Aerosp. Eng.

J. Aerosp. Eng., 2015, 28(2): 04014068


Notation Superscripts

The following symbols are used in this paper: k = iteration step in optimization;
A = linearized state operator; · = derivative with respect to time; and
b = input operator; ^ = time discretized form of the time-continuous variables
C = constraint on control cost (N2 · s); or operators.
Cable = corresponding control cost of uable (N2 · s);
Du = evolution operator of the control thrust;
D0 = evolution operator of initial perturbations; References
dt = time step (s);
Betts, J. T. (1998). “Survey of numerical methods for trajectory optimiza-
f = state operator;
tion.” J. Guid. Contr. Dynam., 21(2), 193–207.
J 1 = roll moment of inertia (kg · m2 ); Burchett, B., Peterson, A., and Costello, M. (2002). “Prediction of swerv-
J 2 = lateral moment of inertia (kg · m2 ); ing motion of a dual-spin projectile with lateral pulse jets in atmospheric
L = time integration of the nutation angle (rad2 · s); flight.” Math. Comput. Model., 35(7–8), 821–834.
L = Lagrangian functional;
Downloaded from ascelibrary.org by Beihang University on 02/02/22. Copyright ASCE. For personal use only; all rights reserved.

Curry, W. H., and Reed, J. F. (1966). “Measurement of Magnus effects on a


Lc = longitudinal distance from the control thrust to the sounding rocket model in a supersonic wind tunnel.” Proc., Aerody-
center of gravity (m); namic Testing Conf., American Institute of Aeronautics and Astronau-
Lr = reference length (m); tics, Reston, VA.
Lun = integration of the uncontrollable nutation angle Darby, C. L., Hager, W. W., and Rao, A. V. (2011). “Direct trajectory
(rad2 · s); optimization using a variable low-order adaptive pseudospectral
method.” J. Spacecraft Rockets, 48(3), 433–445.
M = Mach number;
Gong, Q., Fahroo, F., and Ross, I. M. (2008). “Spectral algorithm for
M y , M z = aerodynamic moment (N · m); pseudospectral methods in optimal control.” J. Guid. Contr. Dynam.,
mm = Magnus moment coefficient; 31(3), 460–471.
my = lateral moment coefficient; Jain, S., and Tsiotras, P. (2008). “Trajectory optimization using multi-
mz = pitching moment coefficient; resolution techniques.” J. Guid. Contr. Dynam., 31(5), 1424–1436.
ω
my y = lateral-damping moment coefficient; Lei, J., and Wu, J. (2005). “Coning motion and restrain of large fineness
ω
mz z = pitch-damping moment coefficient; ratio unguided spinning rocket stabilized with tail fin.” Acta Aerodyn.
N = dimension of continuous-time states; Sinica, 23(4), 455–457.
O = center of gravity; Mao, X., Yang, S., and Xu, Y. (2006). “Research on the coning motion
P = search direction in the nonlinear optimization; of wrap-around fin projectiles.” Can. Aeronaut. Space J., 52(3),
P~ = modified search direction in the linearized
119–125.
Mao, X., Yang, S., and Xu, Y. (2007). “Coning motion stability of wrap
optimization; around fin rockets.” Sci. China E, 50(3), 343–350.
q = dynamic pressure (N=m2 ); Miao, R., Wu, J., Ju, X., and Xu, W. (1996). “Wind tunnel experiment
S = reference area (m2 ); of wrap around fin rockets.” Rep. Prepared for the Beijing Institute
s = time lag factor of the control thrust; of Technology, Beijing.
T = terminal time (s); Morote, J., and Liano, G. (2004). “Stability analysis and flight trials of
t = time (s); a clipped wrap around fin configuration.” Proc., Atmospheric Flight
u = control thrust; Mechanics Conf., American Institute of Aeronautics and Astronautics,
uable = optimal control without constraint on control cost; Reston, VA.
V = velocity (m=s); Najson, F., and Mease, K. D. (2006). “Computationally inexpensive
guidance algorithm for fuel-efficient terminal descent.” J. Guid. Contr.
v = eigenvectors of D̂u ;
Dynam., 29(4), 955–964.
wxc = critical spinning rate (rad=s); Nicolaides, J. D., Ingram, C. W., and Clare, T. A. (1969). “Investigation of
X = base state vector; the nonlinear flight dynamics of ordnance weapons.” Proc., Aerospace
x = state vector; Sciences Meeting, American Institute of Aeronautics and Astronautics,
x = adjoint state; Reston, VA.
x 0 = perturbation to the state; Nocedal, J., and Wright, S. (1999). Numerical optimization, Springer,
β = step length in optimization; London, 102–133.
λ = Lagrangian multiplier; Wang, X., Yang, S., Xu, Y., and Mao, X. (2005). “Effects of missile
λi = eigenvalues of D̂u ; rotation on supersonic fluidic element.” J. Solid Rocket Technol.,
θ = nutation angle (rad); 28(3).
Wu, J., Ju, X., and Miao, R. (1995). “Advances in the research for
θable = nutation angle that can be suppressed without
aerodynamic characteristics of wrap-around fins.” Adv. Mech., 25(1),
constraint on control cost (rad); 102–113.
θu = nutation angle induced by control thrust (rad); Yan, X., Yang, S., and Xiong, F. (2011). “Stability limits of spinning
θ0 = nutation angle induced by initial perturbations (rad); missiles with attitude autopilot.” J. Guid. Contr. Dynam., 34(1),
and 278–283.
ω = angular speed (rad=s). Yan, X., Yang, S., and Zhang, C. (2010). “Coning motion of spinning
missiles induced by the rate loop.” J. Guid. Contr. Dynam., 33(5),
Subscripts 1490–1499.
Zhang, C., and Yang, S. (2004). “Method to get the attitude of
x, y, z = component of vectors in OX, OY, OZ directions a rolling airframe missile.” Trans. Beijing Inst. Tech., 24(6),
(defined in Fig. 1), respectively. 481–485.

© ASCE 04014068-10 J. Aerosp. Eng.

J. Aerosp. Eng., 2015, 28(2): 04014068

You might also like