Professional Documents
Culture Documents
Abstract—A new computational optimization approach is programming (NLP) problem, which parameters to be
proposed to generate Reusable Launch Vehicle (RLV)’s re-entry optimized are discretized state variables and/or control
trajectory. With the application of the new method, all state variables at every node. The direct method is widely used by
variables are approximated by B-spline curves. After the many researchers to solve trajectory optimization problems for
trajectory optimization problem is transformed into a general many reasons. First, the convergence radius of direct methods
nonlinear programming (NLP) problem, which parameters to be is larger than that of indirect methods. Second, it does not
optimized are coefficients of the B-spline curves and control require an initial guess for the costate variables. Finally, it’s
variables at every node and interpolation point, a nonlinear much easier to solve NLP problem than to solve HBVP
programming solver is selected to determine the parameters.
because many well-known software programs exist (e.g.,
Simulation results show that the optimized re-entry trajectories
satisfy the path constraints, and can accomplish the desired
SNOPT, NPSOL.) to solve large sparse NLP problem. In the
terminal conditions. Compared with the Direct Collocation with Direct Collocation with Nonlinear Programming (DCNLP) [4],
Nonlinear Programming (DCNLP) method, the new approach is the most commonly used direct method [5, 6], the state
better in continuity and has fewer coefficients to be determined; variables are approximated to be the piecewise Hermite cubic
moreover, the new approach can achieve better performance polynomial between each interval. Consequently, DCNLP
index. method leads to poor continuously differentiability at every
node and too many parameters to be determined in the whole
Keywords-RLV; trajectory optimization; DCNLP; B-spline range.
curves
A new direct method is proposed in this paper to generate
RLV’s re-entry trajectory. With the application of the new
I. INTRODUCTION method, all state variables are represented by B-spline curves.
During re-entry segment, Reusable Launch Vehicle After transforming the trajectory optimization problem into a
(RLV)’s model is highly nonlinear and strong coupling. RLV general NLP problem, the parameters to be optimized are
also has a variety of nonlinear constraints, so re-entry is the unknown coefficients of the B-spline curves and control
most challenging flight phase. Many researchers focus on variables at every node. Compared with DCNLP method, the
RLV’s re-entry trajectory optimization in preliminary design. new approach is better in continuity and has fewer coefficients
to be determined.
Numerical methods for trajectory optimization are
classified into two methods [1]: indirect methods and direct
methods. In an indirect method, the calculus of variations and II. OPTIMIZATION BASED ON B-SPLINE THEORY
Pontryagin minimum Principle [2] are employed to obtain the
first-order optimality conditions, and then the solution can be A. The Feature of B-spline Curves
found through solving the resulting Hamiltonian boundary- Suppose the RLV’s dynamics model equations are
value problem (HBVP). If solvable, HBVP produces state, described by:
control, and costate variables, so we can know how close to
true optimal solution. However, indirect methods are generally
intractable due to the following reasons [3]: 1) the convergence X = F ( x(t ), u (t ), t ) . (1)
radius of indirect methods is small. 2) It’s difficult to solve
HBVP because the solution of HBVB is very sensitive to the Where, x , u , t are state, control and time variables
initial guesses for the costate variables. 3) Often, the optimality respectively.
conditions cannot even be derived.
The time interval of whole re-entry phase is adopted by
In a direct method, the state and/or control variables are [t0, t f ] , divide phase into segments:
approximated using piecewise constant parameterization. Then
the optimal problem is transformed to a normal nonlinear 0 = t 0 < " ti −1 < " < t N = t f .
B. Preliminary Analysis
1) Analysis on DCNLP method:
Between each interval in DCNLP method, the state
variables are approximated to be the piecewise Hermite cubic
polynomial: xi (ζ ) = ai 0 + ai1ζ i + ai 2ζ i 2 + ai 3ζ i 3 .The method
have the following two drawbacks:
a) Poor continuously differentiability at every node: In
the N segments, the curves generated from the piecewise
polynomials are discontinuous. In order to make the
Figure 1. A cubic uniform B-spline curve independent curves connect to each other to form one complete
curve on the whole time period, DCNLP method have two
In Fig.1, the marked black points represent CUBS’s N + 3 extra boundary constraints:
“control points” P1 , P2 ," , PN + 3 , and the solid line express the • The value of the end of the i -th curve must be equal to
N curves generated from the control points. In the the value of the start point of the i + 1 -th curve.
segment [ti −1 , ti ] , their expression in matrix form is written as
• The first order derivative of the end of the i -th curve
must be equal to first order derivative of the start point
of the i + 1 -th curve.
ª −1 3 −3 1 º ªζ 3 º
« 3 −6 0 « »
1 4 »» «ζ 2 » However, the two constraints can guarantee only one time
f i (ζ ) = [ Pi Pi +1 Pi + 2 Pi + 3 ] « . (2) continuously differentiability at every node.
6 « −3 3 3 1» « ζ »
« »« »
¬1 0 0 0 ¼ ¬« 1 ¼» b) Too many parameters to be determined in the whole
range: Each piecewise Hermite cubic polynomial has 4
The first order derivative of ζ is parameters ai 0 , ai1 , ai 2 , ai 3 , so there are 4 * N parameters to
be determined totally. Besides, the approach used for
dfi 1 1
parameters determination is complicated. DCNLP method first
(ζ ) = − Pi (1 − ζ ) 2 + Pi +1 (3ζ 2 − 4ζ ) explores the law between Hermite cubic polynomial’s
dζ 2 2
. (3) parameters and discretized state variables and/or control
1 1 variables at every node. After the state variables and/or control
+ Pi + 2 ( −3ζ + 2ζ + 1) + Pi + 3ζ 2
2
4943
coefficients greatly compared with Hermite cubic 2) Numerical integration constraints on each segment
polynomials. The states are integrated by the third order Simpson’s rule
on each segment. So, the defect yields:
C. Trajectory Optimization Algorithm
In order to ensure that the state variables approximated by τ
CUBS satisfy the RLV’s dynamics model (1), the following Δi = xi +1 − xi − [ F ( xi , ui , t ) + 4F ( xic , uic , t ) + F ( xi +1 , ui+1 , t )] = 0 . (13)
6
constraints are deduced.
1) At every node and interpolation point, the first order Where, i ∈ [1, N ] . Now, the state differential equations have
derivative of CUBS must be equal with (1). been parameterized by the new method. Equation (9), (12), (13)
According to (2), the values of state variables approximated transform the state differential equations into nonlinear equality
by CUBS on N + 1 nodes are: constraints. Then, the optimal control problems can be
formulated as NLP problems, while optimizing the
performance index and satisfying all kinds of constraints,
xi =1, N +1 = [ f1 (0), f 2 (0),", f N (0), f N (1)] . (6) which parameters to be determined are coefficients of the B-
spline curves and control variables at every node and
The first order derivative of ζ is adopted by x ′ . interpolation point.
Cx qS
τ ⋅ F ( xi , ui , t ) = xi′ , i ∈ [1, N + 1] . (9) °v = − m − g sin θT
°
°θ = CL qS cosυ − g cosθT + v cosθT
Where, ui denotes the values of control variables at a node ° T mv v r
°
point. Suppose the interpolation point locates at midpoint of °σ = − CL qS sin υ + v tan φ cosθT sin σ T
every segment [ti , ti +1 ] , namely ζ = 1/ 2 . The values of state ° T mv cosθT r
® . (15)
variables on N interpolation points are: ° v cos θ cos σ
°φ = T T
° r
xic ( i =1, N ) = [ f1 (0.5), f 2 (0.5),", f N (0.5)] . ° v cosθT sin σ T
(10) °λ = −
° r cos φ
°¯r = v sin θT
At interpolation points, the first order derivative of ζ is
adopted by xc′ . Where the state variables are velocity v (m/s), flight path
angle θT (rad), velocity azimuth angle σ T (rad), geodetic
df1 df df latitude φ (rad), longitude λ (rad), distance from the center of
xic′ ( i =1, N ) = [ (0.5), 1 (0.5),", 1 (0.5)] . (11) Earth to the vehicle r (m). The angle of attack α (rad) and the
dζ dζ dζ
bank angle υ (rad) are adopted as control variables.
The constraints are given by: Where the mass of the RLV is denoted by m ( kg ), and
S ( m2 ) represents reference area. C x , C L are the aerodynamic
τ ⋅ F ( xic , uic , t ) = xic′ ˈ i ∈ [1, N ] . (12) drag and lift force coefficients, and they are functions of Mach
number and angle of attack.
Where, uic denote the values of control variables at Where q = 1/ 2* ρ v 2 is the dynamic pressure. Air density
interpolation points. ρ is expressed by ρ = ρ 0 e − β h , here ρ0 represents atmospheric
The controls are approximated by piecewise linear density at sea level and h represents altitude. Acceleration of
interpolating polynomials, because there are no derivative gravity is modeled by g = g 0 r02 / r 2 , here g 0 represents
conditions at the node or interpolation points.
4944
acceleration of gravity at sea level. r0 represents mean radius The initial and final states specified as:
of the Earth. (v0 , θT 0 , σ T 0 , φT 0 , λT 0 , h0 ) = (1600m / s, −3°,90.0°, 0°, 0°, 40000.0m)
(v f , θTf , σ Tf , φTf , λTf , h f ) = (762 ± 50m / s, 0 ± 5°, free, 0 ± 0.15°, free,30000 ± 200m )
B. Performance Index
Because the RLV is heading west from ( 0°, 0° ), the
The purpose of the research is to find angle of attack and performance index can be formulated as
bank angle history to maximum downrange in longitude profile.
So, the cost function is:
J = max( downrange) = min(λ f ) . (22)
J = max(downrange) . (16) The path limits and control inequality constraints are
specified respectively as
C. Reentry Constraints
ª qs max º ª 250kw / m º
2
1 2
q= ρ v ≤ qmax . (19)
2
4945
Figure 4. Velocity Figure 6. Altitude trajectories between two methods
V. CONCLUSION
A new optimization approach is proposed to generate
RLV’s re-entry trajectory. Simulation results show that the new
method only cost little time and the optimized re-entry
trajectories satisfy the path and terminal state constraints
Figure 5. Altitude successfully. Compared with DCNLP method, the new
approach is better in continuity and has fewer coefficients to be
In Fig. 4~ Fig. 5, the solid line represents the state variable determined; moreover, the new approach can achieve better
curve generated from B-spline, while the dashed line shows the performance index.
state variable history from simulation in (15) with control
variables in Fig. 1 and Fig. 2. There is a slight difference REFERENCES
between B-Spline curves and simulation results owing to few
segments divided. The more segments we divide the whole
[1] Betts J. T., “Survey of Numerical Methods for Trajectory Optimization”,
phase, the more accurate simulation results we can obtain. Journal of Guidance, Control, and Dynamics, vol. 21, No. 2, pp. 193̄
However, too many segments can lead to too many parameters 207, 1998.
to be determined in NLP problem, which will cost too much [2] D. G. Hull, Optimal Control Theory for Aplications. New York:
time to be solved. The error result from the 10 segments in this Springer-Verlag, 2003.
paper is in an acceptable range. [3] Rao A. V., “A Survey of Numerical Methods for Optimal Control”,
AAS/AIAA Astrodynamics Specialist Conference, AAS Paper 2009-334,
The terminal state conditions are v f = 731.166m / s ˈ Pittsburgh, PA, August 10-13, 2009.
θTf = −2.56° ˈ φ f = −0.01° ˈ h f = 30160m ˈ and the [4] E. D. Dickmanns, K. H. Well. “Approximate Solution of Optimal
Control Problems Using Third Order Hermite Polynomial Functions”. In
performance index is λ f = −4.11° (455 km ). The peak values of Proceedings of Optimization Techniques. London: Springer-Verlag,
heating rate, dynamic pressure, and normal load are pp, 158-166, 1974.
18.38 Kw / m 2 , 7.38 Kpa , and 1.43 respectively. Simulation [5] Tu Lianghui, Yuan Jianping, Yue Xiaokui. “Improving Design of
Reentry Vehicle Trajectory Optimization Using Direct Collocation
results show that the optimized re-entry trajectories satisfy the Method”. Journal of Northwestern Polytechnical University, vol. 24,
path and terminal state constraints strictly. No.5, pp. 653-656, 2006.
[6] Tang S, Conway B A. “Optimization of low-thrust interplanetary
trajectories using collocation and nonlinear programming”, Guidance
TABLE I. PERFORMANCE INDEX BETWEEN TWO METHOD
Control and Dynamics, vol.18, No.3, pp. 599-604, 1995.
[7] Fazhong Shi, Computer-Aided Geometric Design and Non-Uniform
Method Segments Final time Performance index Range
Rational B-Spline, Beijing: Higher Education Press, 2001, pp.212-218.
B-spline 10 438.69s -4.1e 455 km [8] Hanyuan Zhao, Vehicle Dynamics and Guidance, Changsha: National
-3.7e 417 km University of Deffense Technology Press, 1997, pp. 454.
DCNLP 20 361.34s
4946