Sie sind auf Seite 1von 14

44th AIAA Aerospace Sciences Meeting and Exhibit AIAA 2006-341

9 - 12 January 2006, Reno, Nevada

Support Vector Regression-driven Multidisciplinary Design


Optimization for Multi-Stage Space Launch Vehicle
Considering Throttling Effect

Saqlain Akhtar* and He Linshu†


School of Astronautics, Beihang University, Beijing, 100083

The design of new Space Launch Vehicle (SLV) involves a full set of disciplines –
propulsion, structural sizing, aerodynamics, mission analysis, flight control, stages layout –
with strong interaction between each other. The optimal design of launch vehicles based on
liquid rocket engines is critically dependent on ascent trajectories and thrust throttling.
Recently many authors have incorporated trajectory optimization at conceptual design level
Downloaded by Stanford University on October 5, 2012 | http://arc.aiaa.org | DOI: 10.2514/6.2006-341

but still have not gauged the potential of using thrust throttling at conceptual design phase.
Since multidisciplinary design optimization of multistage launch vehicles is a complex and
computationally expensive, an efficient Least Square Support Vector Regression (LS-SVR)
technique is used to approximate the current problem. In this study we proposed not only
the trajectory optimization but also the thrust throttling at the conceptual design phase. The
whole problem with 26 parameters, linked to, the architecture and the command (trajectory
optimization), 8 constraints - is solved using well known population based Genetic Algorithm
as optimizer. The objective is to find minimum gross take-off mass, optimal trajectory and
thrust throttling profile for liquid fueled space launch vehicle (SLV). The improvement in
system design using thrust throttling with metamodeling for Space Launch Vehicle is studied
in detail.

Nomenclature
CD = coefficient of drag
Cf = thrust coefficient
CLα = coefficient of lift
c* = characteristic velocity
dt = diameter of throat
ew = weld efficiency
F = thrust
fs = safety factor
g = gravity
Hp = pump’s head pressure rise
h = altitude above Earth surface
k = equivalent wall thickness of ellipsoidal tank end
Lc = length of combustion chamber
L = lift force
Ln = length of nozzle
lc = length of cylindrical part of propellant tanks
m = instantaneous mass of rocket
m0i = liftoff weight of ith stage rocket
mpi = mass of propellant of ith stage rocket
mki = mass of structure of ith stage rocket
m0 (i+1) = pay load of the ith stage rocket
Ns = stage specific speed
*
Ph.D. Student, Beihang University, 37 Xueyuan Lu 100083 Beijing, China.

Professor, Beihang University, 37 Xueyuan Lu 100083 Beijing, China.

1
American Institute of Aeronautics and Astronautics

Copyright © 2006 by the American Institute of Aeronautics and Astronautics, Inc. All rights reserved.
n = number of pump stages
pc = combustion chamber pressure
pe = exit plane pressure
RE = radius of Earth
r = mixture ratio of the propellant.
rc = radius of combustion chamber
sw = allowable operating stress
µp = propellant mass ratio
V = velocity
χ = relative propellant mass ratio
υo = weight to thrust ratio
λ = fineness ratio for propellant tanks
ϑ = velocity inclination angle w.r.t. local horizon
αm = maximum angle of attack at subsonic speed
η = polar angle
ϕ = pitch angle
Downloaded by Stanford University on October 5, 2012 | http://arc.aiaa.org | DOI: 10.2514/6.2006-341

θ = amplitude angle of inclination of velocity to launch horizon


θcn = nozzle cone half angle
ε = nozzle expansion ratio
ηp = pump efficiency

I. Introduction

T HE design of a new Space Launch Vehicle (SLV) involves, from the beginning of conceptual design
architecture to the working out of a detailed definition, a full set of disciplines: propulsion, structural sizing,
aerodynamics, mission analysis, flight control and stages layout. All these disciplines have more or less strong
interactions and coupling effects between each other. They all interact at system level or sub-system level, or both.
They manage numerous data and constraints, either shared or local. The overall project is therefore a typical Multi-
Disciplinary Analysis and Optimization (MDAO) design process.
Conceptual design refers to systems studies conducted early in the design process and intended to reveal trends and
allow relative comparisons among alternatives. Such conceptual design studies provide quantitative data that can be
used by decision makers while the design is still flexible and before the greatest share of life cycle costs are
committed. At the beginning of conceptual design, often only the mission requirements are known, but, in some
cases, additional information regarding vehicle concept, operational approach, and subsystem technologies may also
be available. The conceptual design is crucial to the success of the overall design process and resulting vehicle
system. It has been estimated that at least 80% of a vehicle’s life-cycle cost is locked in by the concept that is chosen
and also conceptual design decision have a 100:1 leverage on end product quality and cost.
A brief overview for conceptual design of launch vehicle has accounted for in Ref. 1. Conceptual design of two-
stage-to-orbit for minimum total dry mass of the first booster and the second stage orbiter with trajectory
optimization has been discussed in Ref.2. The conceptual design of two-stage reusable rocket vehicle also includes
trajectory optimization in Ref. 3. The two vehicles, vertical launch and horizontal takeoff vehicles with cryogenic
rocket engine were presented for minimum gross weight comparison. A first stage reusable design with attached
expendable space transportation system with nominal thrust and fixed mixture ratio has been discussed in Ref. 4.
Stable static throttling capability was demonstrated from 30 to 100 percent of its maximum thrust level in Ref. 5.
The concept of selection, refinement and verification phases of conceptual design for the Space Launch Vehicle
capabilities of the Peacekeeper ICBM has been given by Ref. 6. This reference has taken input trajectory and
constant thrust for each stage.
Simulation-based analysis tools are finding increased use during preliminary design to explore design
alternatives at the system level. In spite of advances in computer capacity and speed, the enormous computational
cost of complex, high fidelity scientific and engineering simulations makes it impractical to rely exclusively on
simulation codes for the purpose of multidisciplinary design optimization. A preferable strategy is to utilize
approximation models which are often referred to as metamodels, replacing the expensive simulation model during
the design and optimization process. Metamodeling techniques have been widely used for design evaluation and
optimization in many engineering applications; a comprehensive review of metamodeling applications in
mechanical and aerospace systems can be found in Ref. 7. A review of metamodeling applications in

2
American Institute of Aeronautics and Astronautics
multidisciplinary design optimization is given in Ref. 8. The survey reveals that most of the references have
included trajectory optimization at conceptual design level but thrust throttling capability and metamodeling has not
been considered simultaneously. The authors have introduced trajectory optimization and thrust throttling capability
with efficient metamodeling technique (LS-SVR) for SLV at conceptual design level.
Almost every discipline in aerospace from Guidance, Navigation, Control, Propulsion and Structures has yielded
itself to the power of computational intelligence9. Population based, non-gradient, stochastic direct search
optimization methods are therefore attractive choices for such problem as they are easy to implement and effective
for highly nonlinear problems. In this analysis, design of an expendable three-stage liquid fueled low earth orbit
launch vehicle includes specification of the ascent trajectory, as well as determination of the subsystem weights and
sizes. Commercially available LOX/RP-1 is selected as propellant for expendable liquid fueled SLV. The design
problem with throttling and non-throttling thrust, posed within numerous constrained optimization architecture is
presented and discussed.

II. Launch Vehicle Design Problem

A. Objective Function and Constraints


Downloaded by Stanford University on October 5, 2012 | http://arc.aiaa.org | DOI: 10.2514/6.2006-341

In launch vehicle design minimum gross take-off mass concepts have traditionally been sought. Since vehicle
development costs tend to vary as a function of gross take-off mass, this minimum take-off mass vehicle may be
considered a minimum development cost concept10. Minimum take-off mass is taken as objective function. Dynamic
penalty function is used to handle in flight and terminal constraints. A symbolic problem statement can be expressed
as follows.

m
min f (x) = f(x) +h (k) ∑ max{0, g ( x)}
i =1
i (1)

Where f(x) is the objective function, h (k) is a dynamically modified penalty value; k is the current iteration
number of the algorithm. The function gi(x) is a relative violated function of the constraints11.
The liquid fueled SLV is sized to deliver a 450kg payload to 500km altitude low earth circular orbit with the
correct inclination. SLV Mission velocity and corresponding altitude are formulated as trajectory constraints. Axial
overload constraint is implemented to restrict it below 12g. The parameter angle of attack times dynamic pressure
(q-α)max was constrained below 15000 kPa-deg structural design limit to ensure that the aerodynamic loads do not
exceed the structural capability of the vehicle. During launch maneuver; the maximum angle of attack is constrained
to be below 8 deg and to ensure that it is zero during transonic phase. Weight to thrust ratios ν0 and propellant mass
ratio µp was constrained to be within allowable ranges. Nozzle exit diameters are constrained to be less than stage
diameters. The lateral overload constraint is also implemented to restrict it below 0.3.

B. Design Variables
For conceptual design of liquid fueled Space Launch Vehicle 26 design variables with upper and lower bounds are
presented in Table 1.
Table 1. Limits of design variables
STAGE 1 STAGE II STAGE III
Design Variables Lower Upper Lower Upper Lower Upper
Bound Bound Bound Bound Bound Bound
Chamber pressure pc, bar 50 100 35 90 28 80
Nozzle exit pressure pe, bar 0.45 0.70 0.15 0.50 0.05 0.20
Stage diameter d, m 1.50 2.20 0.9 2.20 0.54 2.20
Fineness ratio, λ 4 8 3 7 1.5 4
Throat diameter dt, m 0.25 0.80 0.15 0.70 0.10 0.63
Throttling factor kt, % 70 100 70 100 70 100
Maximum angle of attack αm, deg 3 8 - - - -
Flight program design factor, a 0.05 0.5 - - - -

3
American Institute of Aeronautics and Astronautics
C. Throttling Thrust
One of the important performance parameter of rocket is the thrust delivered. This parameter determines the
accelerations achieved. Some missions also require that the thrust is controllable for example to allow reducing
acceleration loads towards the end of the flight, when the propellant tanks are almost empty. Throttling capability
control/change the thrust of an individual rocket motor given as a ‘percentage’ (%) of nominal thrust. For example, a
throttling capability of 70% means that the thrust can be reduced to 70% of its nominal value. Controlled variation
of thrust is of interest in manned aircraft power plants, missile-guidance programming, and ducted rocket engines.
Simple thrust control, smooth combustion, and high specific impulse throughout the thrust range are desirable
features in such applications. High specific impulse requires the maintenance of good mixture preparation and high
nozzle pressure ratios throughout the thrust range. To achieve both of these, throttling of liquid flow at the injector
and of gas flow at the engine throat are required. A simple approach has been implemented with mass flow rate
control at the injector to get 70% throttling thrust in this study. A throttling design factor kt is introduced for each
stage. The lower and upper bound is defined at three intervals during burn out time of each stage. The throttling
profile through optimized points is achieved through Matlab function of curve fitting.
Variable and constant thrust propulsion is compared by evaluating its performance by means of the 3-D trajectory
optimization of multistage space launch vehicle. The main purpose of this paper is to evaluate the advantage given
Downloaded by Stanford University on October 5, 2012 | http://arc.aiaa.org | DOI: 10.2514/6.2006-341

by throttling thrust in comparison with non-throttling thrust using Support Vector Machine based metamodeling
along with Genetic Algorithm as optimizer.

D. Mass Module
The total mass of a multistage rocket includes the masses of propellants and their tanks, engines, feeding system
and payload mass. The mass equation for a multi-stage rocket can be written as:

m0i = m pi + mki + m0(i +1) (2)

The engine weight is a difficult but very important parameter necessary for accurately performing trades on the
engine design. Using a combination of physics-based methods and empirical data, the weight of the major engine
components can be determined. The expression for throat area of nozzle may be expressed as follows:

F
At = (3)
pc .C f

γ −1
γ +1
⎡ ⎤
2γ 2 ⎡ 2 ⎤ γ −1 ⎢ ⎛ p e ⎞ γ ⎥ ⎡ p − pa ⎤
Cf = 1 − ⎜ ⎟ +ε⎢ e ⎥
γ − 1 ⎢⎣ γ + 1⎥⎦ ⎢ ⎜⎝ p c ⎟⎠ ⎥ ⎣ pc ⎦
⎢⎣ ⎥⎦
(4)

The computational formula for the mass of combustion chamber is:

⎡ ⎛ r 2 − rt2 ⎞⎤
mc = πρtw ⎢rc Lc + ⎜ c ⎟⎥ (5)
⎢⎣ ⎜ tan θc ⎟⎥
⎝ ⎠⎦

The wall thickness of combustion chamber can be determined as:

f s prc
tw = (6)
Ftu

Mass of Nozzle-constant thickness is given by the following expression:

mn = πρt w Ln (re + rt ) (7)

De − Dt
Ln = (8)
2 tan θcn

4
American Institute of Aeronautics and Astronautics
Propellant mass calculation expression is shown as below:

m pi = µ pi moi (9)

Mass of fuel and oxidizer can be calculated using the following expression:

mp
mf = (10)
1+ r

mox = r.m f (11)

Mass flow rate of propellant can be found with following formula.

. pc At
m= (12)
c*
Downloaded by Stanford University on October 5, 2012 | http://arc.aiaa.org | DOI: 10.2514/6.2006-341

Burn out time calculation for each stage is calculated using the expression below.

mp
tk = •
(13)
m

Mass of propellant tanks calculation expression is given as:

πr 2te E / ρ tan k 2πVcyl pt ρ tan k


M tank = +
k πswew
(14)

Following expression is given for Feed pump mass calculation:

0.6
⎛ ⎛•⎞
3/ 2 ⎞
⎜ ⎜ ⎟ ⎟
⎜ g H
o p⎜ ⎟m ⎟
mtp = 1.5⎜⎜ ⎝ ⎠ ⎟
0.75 ⎟
(15)
⎜ ⎛ Hp ⎞ ⎟
⎜ ⎟
⎜⎜ η p N s ρ ⎜ n ⎟ ⎟⎟
⎝ ⎝ ⎠ ⎠

To estimate the support-structural mass, we assumed a 10% increase in the overall inert mass of the system12.

E. Aerodynamic Module
The ascent control analysis requires complete data sets of the aerodynamic coefficients CD (Drag) and CL (Lift)
as function of the angle of attack and Mach number for the complete ascent configuration. These sets can not be
obtained by sophisticated calculations in this early phase of preliminary design. For the aerodynamic characteristics
assessment, the vehicle is decomposed in basic geometric elements. The aerodynamic coefficients are evaluated for
each element by using semi-empirical methods.

F. Trajectory Module
To analyze the ascent flight-path, a three degree-of-freedom trajectory model was developed and simulated in
SIMULINK. Within the model, the equations of motion are numerically integrated from an initial to a terminal set
of state conditions. Within the present investigation, the vehicle is treated as a point-mass, flight in 2D over
spherical non-rotating earth was assumed, which implied that the Coriolis and centrifugal pseudo forces were
negligible. The 1976 standard atmosphere (no winds) was used. The trajectory was optimized during maneuvering
phase, after transonic phase SLV turning was considered under gravity. A vehicle is modeled as a group of stages,
with each stage having its own aerodynamic and propulsion inputs. After two power stages there was a coasting

5
American Institute of Aeronautics and Astronautics
stage until mission height was achieved and finally last stage engine was ignited to achieve mission orbit velocity.
Trajectory inputs includes vehicle (e.g., gross liftoff weight, liftoff thrust-to-weight ratio, propellant ass flow rate,
aerodynamic coefficients and reference area).

• T cos α − D
V= − g sin ϑ
m ( 16)
• T sin α + L g cos ϑ V
ϑ= − + cos ϑ
mV V Re + h

h = V sin ϑ
• R
l = ( e )V cos ϑ
Re + h
α = η + ϕ −ϑ
l
η=
Re
Downloaded by Stanford University on October 5, 2012 | http://arc.aiaa.org | DOI: 10.2514/6.2006-341

ϕ = ϕ pro (t)
1 2
L = C Lα ρ v Aref
2
1
D = C D ρv 2 Aref
2

Figure 1. Forces acting on a rocket body during flight.

Terminal constraints on altitude, velocity, and flight-path angle as well as maximum in flight dynamic pressure,
angle of attack during maneuvering phase, pitch rate, and normal force limits are enforced. Interference forces are
not considered in this simple, but fast and efficient approach.

III. Methodology

A. Latin Hypercube Sampling (LHS)


Design of Experiment based Space-filling designs are used when there is little or no information about the
underlying effects of factors on responses, as in the case under study. Latin hypercube sampling (LHS), a stratified-
random procedure, provides an efficient way of sampling variables from their distributions. The LHS involves
sampling ns values from the prescribed distribution of each of k variables X1, X2,…, Xk. The cumulative distribution
for each variable is divided into N equiprobable intervals. A value is selected randomly from each interval. The N
values obtained for each variable are paired randomly with the other variables. Unlike simple random sampling, this
method ensures a full coverage of the range of each variable by maximally stratifying each marginal distribution13.
For this study built-in MATLAB function from its toolbox is used to generate two sets of design vectors from
Latin Hyper Sampling. These two sets of design vectors are then passed to calculate actual weight, propulsion,
aerodynamics and trajectory parameters for training and testing data for metamodel module.

B. Metamodeling
Much of today’s engineering analysis requires running complex and computationally expensive analysis and
simulation codes. Despite continuing increases in computer processor speeds and capabilities, the huge time and
computational costs of running complex engineering codes maintains pace. A way to overcome this problem is to
generate an approximation of the complex analysis code that describes the process accurately enough, but at a much
lower cost. Such approximations are often called “metamodels” in that they provide a “model of the model”14.
Mathematically, if the inputs to the actual computer analysis are supplied in vector x and the outputs from the
analysis in vector y, then the true computer analysis code evaluates:

y = f (x) (17)

6
American Institute of Aeronautics and Astronautics
Where f(x) is a complex engineering analysis function. The computationally efficient metamodel approximation is:

ŷ = g (x) (18)

Such that: y= ŷ+ε (19)

Where ε includes both approximation and random errors.


There currently exist a number of metamodeling techniques to approximate f(x) with g(x), such as polynomial
response surface models, multivariate adaptive regression splines, radial basis functions, kriging, and neural
networks, and a recent comparison of the first four of these metamodeling techniques can be found in Ref. 15.
All of these techniques are capable of function approximation. In particular, although neural networks are able to
approximate very complex models well, they have the two disadvantages of (i) being a “black box” approach and (ii)
having a computationally expensive training process. “Black box” means that little can be seen and understood
about the model, because an exact function is not generated, only a trained “box” that accepts inputs and returns
outputs.
1. Support Vector Machine Based Metamodeling
Downloaded by Stanford University on October 5, 2012 | http://arc.aiaa.org | DOI: 10.2514/6.2006-341

Support Vector Machines is a powerful methodology for solving problems in nonlinear classification, function
estimation and density estimation which has also led to many other recent developments in kernel based methods in
general. Originally, it has been introduced within the context of statistical learning theory and structural risk
minimization. In the methods one solves convex optimization problems, typically quadratic programs. The
computationally efficient theory behind Support Vector Regression is presented, and Support Vector Regression
approximations are compared against the aforementioned four metamodeling techniques using a test bed of 22
engineering analysis functions in Ref. 16 and it was concluded that Support Vector Regression achieves more
accurate and more robust function approximations than these four metamodeling techniques and shows great
promise for future metamodeling applications.
Least Squares Support Vector Machines (LS-SVM) are reformulations to the standard SVMs which lead to
solving linear Karush-Kuhn-Tucker (KKT) systems. LS-SVMs are closely related to regularization networks and
Gaussian processes but additionally emphasize and exploit primal-dual interpretations. Links between kernel
versions of classical pattern recognition algorithms such as kernel Fisher discriminant analysis and extensions to
unsupervised learning. Robustness, sparseness and weightings can be imposed to LS-SVMs where needed17. To cut
down the computational cost, surrogate models (LS-SVR), are constructed from and then used in lieu of the actual
simulation models.
2. .Architecture of Least Square Support Vector Machine
Some typical choices of positive definite kernels are linear, polynomial, Radial Basis Function (RBF) and Multi
Layer Perceptrons (MLP) kernel. In classification problems one frequently employs the linear, polynomial and RBF
kernel. In nonlinear function estimation and nonlinear modeling problems one often uses RBF kernel18.
In order to construct a LS-SVM model with RBF kernel, we need two important parameters: γ (gamma) and
σ2 (sigma2). The γ (gamma) is the regularization parameter, determining the trade-off between the fitting error
minimization and smoothness. In the case of RBF kernel, σ2 (sigma2) is the bandwidth. The tuning of these two
parameters is utmost important to get fine results from metamodel module. For the metamodeling of Space Launch
Vehicle trajectory and propulsion module, a cascade of four LS-SVMs is proposed with multiple output regression
for each of two optimization algorithms. Each cascade represents first two power stages then a coasting stage and
last power stage of Space Launch Vehicle.
3. Training of Least Square Support Vector Machine
The Least Square Support Vector Machines are trained from a set of 1800 design points obtained from Latin
Hypercube Sampling. The performance of trained LS-SVM is evaluated by comparing its simulated output with
predirected values from test data of 550 Latin Hypercube Sampled design points. The training and evaluation
procedure is repeated by different values of σ2 ( sigma2) and γ (gamma) combinations to find out the best pair with
least Mean Absolute Error. After training, the support value and bias term for each LS-SVM is saved for LS-SVM
cascade being used in optimization routine. The values of sig2 and gam parameters and model capability to map the
input & output by Mean Absolute Error (MAS) are given in Table 2.
4. Trajectory Metamodel
Optimization techniques based on Genetic Algorithm are very attractive in terms of robustness and finding
global solution but at the expense of large number of fitness evaluation and substantial amount of computational
resources in each of exact analysis. To circumvent this adverse behavior it is proposed to incorporate Support Vector
Machine as a metamodel to replace the exact trajectory analysis module.

7
American Institute of Aeronautics and Astronautics
Table 2. Least Square Support Vector Machine architecture and Regression performance

S.No RBF Kernal γ σ2 Mean Absolute


Least Square Support Vector (gam) (sig2) Error (MAE)
Machine
Performance Parameters Case 1 Case 2 Case 1 Case 2 Case 1 Case 2
1 1st stage velocity, km/s 20 15 30 25 0.0190 0.0181
st
2 1 stage height, km 20 15 30 25 0.0015 0.0018
st
3 1 stage thrust, KN 20 15 30 25 0.0020 0.0092
st
4 1 stage axial overload, g 20 15 30 25 0.060 0.060
nd
5 2 stage velocity, km/s 30 50 20 30 0.037 0.037
nd
6 2 stage height, km 30 50 20 30 0.0020 0.0023
nd
7 2 stage axial overload, g 30 50 20 30 0.1923 0.194
Downloaded by Stanford University on October 5, 2012 | http://arc.aiaa.org | DOI: 10.2514/6.2006-341

8 Coasting stage velocity, km/s 15 20 15 20 0.1361 0.1180


9 Final stage velocity, km/s 15 25 10 20 0.0396 0.0420
10 Final stage height, km 15 25 10 20 0.002 0.011
11 Final stage thrust, KN 15 25 10 20 0.0324 0.0299
12 Final stage axial overload, g 15 25 10 20 0.247 0.246
Case 1 ---- Non -Throttling thrust Case 2 ---- Throttling thrust

C. Genetic Algorithm
One of the better-known stochastic methods is Genetic Algorithm, which is designed to mimic evolutionary
selection. Genetic Algorithms apply Darwin’s idea to problems where the solution can be expressed as an individual
and the problem is to minimize/maximize the fitness of individuals. Like evolution, Genetic Algorithms proceed by
having fitter individuals propagate their genetic material to succeeding generations. Less fit individuals and their
generation materials do not survive. In Genetic Algorithm, the search space of a problem is represented as a
collection of individuals. These individuals are represented by character strings or matrices, which are often referred
to as chromosomes. The purpose of using a Genetic Algorithm is to find the individual from the search space with
the best genetic material. The quality of an individual is measured with an evolution function. The part of the search
space to be examined is called the population. The genetic operations that define the child production process and
the mutation process are called crossover operators and the mutation operators respectively.
Mutation and crossover play different roles in a Genetic Algorithm. Random mutation provides background
variation and occasionally introduces beneficial material into chromosomes. Mutation is also needed in the process
to explore new states and helps an algorithm avoid local optima. Crossover is the key to Genetic Algorithms.
Crossover exchanges corresponding genetic material from two parent chromosomes, allowing beneficial genes from
different parents to be combined in offspring. There can be various criteria for stopping the Genetic Algorithm, e.g.,
the number of iterations needed (if it can be determined priori). However, the stopping criteria should normally take
into account the uniformity of the population, and the relationship between the average objective function with
respect to the objective function of the best individual19.
There are five basic steps that form a Genetic Algorithm.
1) Start with a randomly generated population of n chromosomes (candidate solutions to a problem)
2) Calculate the fitness f(x) of each chromosome x in the population.
3) Repeat the following steps until n off springs have been created.
• Randomly select a pair of parent chromosomes from the current population.
• Crossover (mate) the pair at a randomly chosen point to form two offspring
• Randomly mutate the two off springs and add the resulting chromosomes to the new population.
• Calculate the fitness of the resulting chromosomes.
4) Let the n fittest chromosomes survive to the next generation.
5) Go to Step 3 (repeat for a pre-specified number of generations).
First, the initial population is chosen, and the quality of this population is determined. Next, parents are selected
from the population in each iteration. These parents produce children and the children are added to the population.
For all newly created individuals of the resulting population, a probability near zero exists that they will mutate, i.e.,

8
American Institute of Aeronautics and Astronautics
that they will change their hereditary distinctions. After that, some individuals are removed from the population
according to selection criteria in order to reduce the population to its initial size. The basis of Genetic Algorithm is
the continual improvement of the fitness of the population by means of selection, crossover, and mutation as genes
are passed from one generation to the next. After a certain number of generations (typically several dozens or
hundred), the population evolves to a near-optimal solution.
In this study real coded GA is used, because it is better suited for the optimization problem in view of the
efficiency of search performance. Before specifying Genetic Algorithm parameters, an extensive parametric study
was conducted for the problem in hand by varying one design parameter at a time. Selecting an optimal combination
of GA parameters is very difficult because each of the GA parameters is varied individually and the number of
combinations of the GA parameters is infinite. Although the values used here are not necessarily optimal but based
partly on practices. A population size of 80 with two point crossover function and uniform mutation with rate of
0.025 is used. A tournament selection was used to allow for minimization and to avoid potential scaling concerns
Both the population size and the string length influence the choice of the mutation probability, so this also varied
with each problem. The Genetic Algorithm stops when the stall generation limit of 50 exceeds, or after 200
generations, whichever occurs first. This allows for a comparison of computational cost for the different strategies.
Downloaded by Stanford University on October 5, 2012 | http://arc.aiaa.org | DOI: 10.2514/6.2006-341

D. Optimization Process
The optimization problem as stated in section II.A is solved by Least Square Support Vector based
metamodeling technique using evolutionary Genetic Algorithm. To elaborate the performance potential of proposed
metamodel based optimization, first the problem is solved by simple Genetic Algorithm using exact function counts
and second by proposed Support Vector Machine as a metamodel tool for constant and throttling thrust. Both
optimization algorithms are shown in Fig. 2 and Fig. 3. In the first case a set of design variables with upper and
lower bounds is passed to Genetic Algorithm which create initial random population and perform its further
operations using exact analysis until stopping criteria of optimal solution is reached. While in second case a set of
design variables with upper and lower bounds is passed to Latin Hypercube Sampling to generate a set of design
vectors based on design of experiments. These candidate design vectors are then passed to trained Support Vector
Machine cascade with multiple output regression. The Genetic Algorithm then instead of performing exact analysis
will get approximate output from metamodel module. The constraints are calculated and handled by external penalty
function. The proposed metamodel optimization method is much faster than ordinary optimization method using
Genetic Algorithm with exact analysis.

Design Variables X Design Variables X


Mass Mass
Module Module Latin Hypercube
Sampling

Propulsion Propulsion
Module Module Least Square Support
Global Optimizer Vector Machine
Aero- GA Aero-
dynamic dynamic
Module Module Global Optimizer
GA
Trajectory Trajectory
Module Module Optimal Launch Vehicle
Optimal Launch Vehicle

Figure 2. Objective Function computation Figure 3. Objective Function computation


process (Exact Analysis). process (Metamodel).

9
American Institute of Aeronautics and Astronautics
IV. Results and Discussion
The performance of LS-SVM cascade for two optimization algorithms are shown in Fig. 4 and Fig. 5.Each
optimization algorithm architecture is executed five times with exact function count and metamodel to find
performance index for conceptual design problem of Space Launch Vehicle. By using five trial runs, it is hoped to
reduce some of the stochastic error that can normally be present in each individual run. The best feasible fitness
value is the lowest fitness i.e., minimum Space Launch Vehicle take-off mass ever encountered that does not violate

1st stage height 1st stage velocity 1st stage axial overload
80 3 10

Observed velocity, km/s

Observed overload, g
Observed height, km

2.5
60 8

2
40 6
1.5

20 4
1
Downloaded by Stanford University on October 5, 2012 | http://arc.aiaa.org | DOI: 10.2514/6.2006-341

0 0.5 2
0 20 40 60 80 0.5 1 1.5 2 2.5 3 2 4 6 8 10
Predirected height, km Predirected velocity, km/s Predirected overload, g

2nd stage height 2nd stage velocity 2nd stage axial overload
300 7 20
Observed velocity, km/s

Observed overload, g
Observed height, km

250 6
15
200 5

150 4 10

100 3
5
50 2

0 1 0
0 100 200 300 1 2 3 4 5 6 7 0 5 10 15 20
Predirected height, km Predirected velocity, km/s Predirected overload, g

Last stage height Last stage velocity Last stage axial overload
2000 12 35
Observed velocity, km/s

Observed overload, g
Observed height, km

30
10
1500
25
8
1000 20
6
15
500
4
10

0 2 5
0 500 1000 1500 2000 2 4 6 8 10 12 5 10 15 20 25 30 35
Predirected height, km Predirected velocity, km/s Predirected overload, g

1st stage thrust Coasting stage velocity Last stage thrust


3000 7 1000
Observed velocity, km/s

6
Observed thrust, KN

Observed thrust, KN

2500
800
5
2000
4 600
1500
3 400
1000
2
500 200
1

0 0 0
0 1000 2000 3000 0 2 4 6 8 0 200 400 600 800 1000
Predirected thrust, KN Predirected velocity, km/s Predirected thrust, KN

Figure 4. Performance of trained Least Square Support Vector Machine(Case 1 Non- throttling thrust).

10
American Institute of Aeronautics and Astronautics
the constraints. The total number of function counts required to reach the stopping criterion was also computed.
Both of these values were used to compare the performance of the computational algorithm. Most significant
contribution is the drastic reduction in number of exact analysis required for convergence by using metamodel based
on Support Vector Machine. The number of function counts needed to converge reflected the performance of these
two architectures shown below. Obviously metamodel architecture was observed computationally more attractive
choice than exact analysis architecture.

1st stage height 1st stage velocity 1st stage axial overload
80 3.5 10

Observd velocity, km/s

Observed overload, g
Observed height, km

3
60 8
2.5

2
40 6
1.5

1
20 4
0.5
Downloaded by Stanford University on October 5, 2012 | http://arc.aiaa.org | DOI: 10.2514/6.2006-341

0 0 2
0 20 40 60 80 0 1 2 3 2 4 6 8 10
Predirected height, km Predirected velocity, km/s Predirected overload, g

2nd stage height 2nd stage velocity 2nd stage axial overload
350 7 20
Observed velocity, km/s

Observed overload, g
Observed height, km

300 6
15
250
5
200
4 10
150
3
100
5
50 2

0 1 0
0 100 200 300 1 2 3 4 5 6 7 0 5 10 15 20
Predirected height, km Predirected velocity, km/s Predirected overload, g

Last stage height Last stage velocity Last stage axial overload
1200 12 35
Observed velocity, km/s

Observed overload, g
Observed height, km

1000 10 30

800 8 25

600 20
6
400 15
4
200 10
2
200 400 600 800 1000 1200 2 4 6 8 10 12 10 15 20 25 30 35
Predirected height, km Predirected velocity, km/s Predirected overload, g

1st stage thrust Coasting stage velocity Last stage thrust


2500 8 800
Obsrved velocity, km/s
Observed thrust, KN

Observed thrust, KN

2000
6 600

1500
4 400
1000

2 200
500

0 0 0
0 500 1000 1500 2000 2500 0 2 4 6 8 0 200 400 600 800
Predirected thrust, KN Predirected velocity, km/s Predirected thrust, KN

Figure 5. Performance of trained Least Square Support Vector Machine (Case 2 Throttling thrust).

11
American Institute of Aeronautics and Astronautics
100 100
Constant Thrust Constant Thrust

Vehicle take-off mass, Mg


Vehicle take-off mass, Mg

80 Throttling Thrust 80 Throttling Thrust

60 60

40
40

20
20
0 0 50 100 150 200
0 50 100 150 200
Downloaded by Stanford University on October 5, 2012 | http://arc.aiaa.org | DOI: 10.2514/6.2006-341

Number of generations Number of generations

Figure 6. Genetic Algorithm evolution Figure 7. Genetic Algorithm evolution


(Exact analysis). (Metamodel).

Table 3. Solution from Genetic Algorithm using exact analysis

Test Solution from GA Solution from GA Total Exact


Run Non-Throttling Thrust Throttling Thrust Analysis

Generations Function Mass Generations Function Mass Non


Count [Mg] Count [Mg] Throttling Throttling
1 191 15280 34.30 141 11280 30.35 15280 11280
2 132 10560 31.75 132 10560 29.65 10560 10560
3 195 15600 32.64 130 10400 29.59 15600 10400
4 197 15760 33.12 143 11440 28.82 15760 11440
5 185 14800 32.73 180 14400 30.36 14800 14400

Table 4. Solution from Genetic Algorithm using Least Square Support Vector Regression
Function
Test Count in Solution from GA Solution from GA Total Exact
Run LS-SVM Non-Throttling Thrust Throttling Thrust Analysis
Training
Gener- Function Mass Gener- Function Mass Non
ations Count [Mg] ations Count [Mg] Throttling Throttling
1 192 15280 34.29 182 14560 29.85 1800 1800
2 1800 174 13920 32.58 195 15600 27.60 1800 1800
3 171 13680 33.69 160 12800 29.44 1800 1800
4 196 15680 33.80 91 7280 30.54 1800 1800
5 185 14800 32.76 102 8160 30.40 1800 1800

The proposed scheme is especially beneficial for complex, nonlinear and expensive engineering analysis required
for optimization. Convergence performance of Genetic Algorithm for two optimization algorithms is presented in
Fig. 6 and Fig. 7

12
American Institute of Aeronautics and Astronautics
Analysis results given in Table 3 and Table 4., reflects the computational advantage gained through Support
Vector Regression as well as throttling thrust advantage in terms of least Space Launch Vehicle take-off mass. The
first achievement in terms of computational cost is nearly 87% cut down in exact analysis function count during
Space Launch Vehicle conceptual design for non throttling
thrust case and 85% for throttling thrust case. The second 5
x 10
achievement gained through throttling thrust is nearly 9% 7
reduction in launch mass of SLV is observed in exact
analysis algorithm and 11% reduction is observed in 6.5
algorithm using metamodel. Table 5 details the optimum

Thrust, KN
design parameters for the vehicle. The throttling thrust 6
routine is linked with trajectory optimization module to get
optimized thrust profile upto 70% throttling. Figure. 8 shows 5.5
optimum throttling thrust profile for first stage. Figure 9
reflects the trajectory performance using optimum design
5
parameters. Given this performance advantage, second
algorithm using throttling thrust with metalmodeling is
Downloaded by Stanford University on October 5, 2012 | http://arc.aiaa.org | DOI: 10.2514/6.2006-341

recommended for such multidisciplinary optimization 4.5


0 20 40 60
problem. The proposed scheme is especially beneficial for Time, s
resource hungry expensive engineering analysis required for
multidisciplinary design optimization. Figure. 8 First stage throttling thrust profile

Table 5. Optimum values of design variables


Design Variables STAGE 1 STAGE II STAGE III
Combustion chamber pressure pc, bar 51 38 31.74
Nozzle exit pressure pe, bar 0.45 0.20 0.09
Stage diameter d, m 1.56 1.18 1
Fine ness ratio λ 4.54 5.99 1.94
Throttling design factor kt , % 75 | 94 | 77 88 | 84 | 81 96 | 95 | 94
Throat diameter dt, m 0.33 0.219 0.135
Maximum angle of attack αm, deg 5.4 0 0
Flight program design factor a 0.30 - -

2 600 30
Angle of Attack, deg

500 25
Vehicle Mass, Mg

0
Altitude, km

400 20

-2 300 15

200 10
-4
100 5

-6 0 0
0 5 10 15 20 25 30 0 100 200 300 400 500 0 100 200 300 400 500 600
Time, s Time, s Time, s
5
x 10
7 12 8000

6 10
Axial Overlaod, g

6000
8
Velocity, m/s

5
Thrust, N

4 6
4000
3 4

2 2
2000
1 0

0 -2 0
0 100 200 300 400 500 600 0 100 200 300 400 500 600 0 200 400 600
Time, s Time, s Time, s]

Figure 9. Trajectory performance with optimum design parameters


13
American Institute of Aeronautics and Astronautics
Conclusion
An integrated methodology allowing to perform multidisciplinary optimization of an expendable multi-stage
Space Launch Vehicle was successfully experimented using Least Square Support Vector Regression. It proved able
to provide a conceptual design considering propulsion and mass features for each stage and trajectory/performance
objectives and constraints. The most important accomplishment of this research is the simultaneous trajectory
optimization and thrust throttling profile optimization with metamodeling at conceptual design phase for a liquid
fueled Space Launch Vehicle. The results have proved that proposed idea has tremendous effect on launch mass of
SLV. The most significant achievement gained through this analysis is nearly 85% reduction in computational cost
by using metalmodel during conceptual design level. Such a design strategy will allow vehicle designers to rapidly
consider a number of fully converged design alternatives in a very short time without sacrificing design detail, thus
improving the quality of conceptual space vehicle design. Due to ease of implementation and their potential to solve
aerospace problems, non gradient methods are more attractive choice than traditional design and analysis tools when
complicated non-linear phenomena dominate the optimization space.

References
Downloaded by Stanford University on October 5, 2012 | http://arc.aiaa.org | DOI: 10.2514/6.2006-341

1
Lawrence F. Rowell, John J. Korte, “Launch Vehicle Design and Optimization Methods and Priority for the Advanced”,
Engineering Environment NASA/TM-2003-212654.
2
Takeshi Tsuchiya, Takashige Mori, “Optimal Design of Two-Stage-to-Orbit Space Planes with Air breathing Engines”,
Journal of Spacecrafts and Rockets vol.42, No.1, 2005.
3
Takeshi Tsuchiya, Takashige Mori, “Optimal Conceptual Design of Two-Stage Reusable Rocket Vehicles including
Trajectory Optimization”, Journal of Spacecrafts and Rockets vol.41, No.5, 2004.
4
Martin Sippel, Josef Klevanski, Holger Burkhardt, Philip Langholf, Andreas Rittweger “Progress in the Design of a
Reusable Launch Vehicle Stage”, AIAA 2002-5220.
5
Naruo, Yoshihiro, Inatani, Yoshifumi, Morita, Yasuhiro, Nakai, Shunichiro, Mori and Hatsuo “Throttling dynamic response
of LH2 rocket engine for vertical landing rocket vehicle”,Space cooperation into the 21st century; Proceedings of the 7th
International Space Conference of Pacific-Basin Societies, Nagasaki, Japan; United States; 15-18 July1997.pp.229-240.1997.
6
A. Kumpel, P. Barros, C. Burg, F. Villeneuve, and D. Mavris, “A Conceptual Design for the Space Launch Capability of
the Peacekeeper ICBM”,Georgia Institute of Technology Atlanta, GA AIAA 2002-5854.
7
Simpson, T. W., Peplinski, J., Koch, P. N. and Allen, J. K., 1997, "On the Use of Statistics in Design and the Implications
for Deterministic Computer Experiments," Design Theory and Methodology - DTM'97, Sacramento, CA, ASME, Paper No.
DETC97/DTM-3881
8
Sobieszczanski-Sobieski, J. and Haftka, R. T., "Multidisciplinary Aerospace Design Optimization: Survey of Recent
Developments," Structural Optimization, Vol. 14, No. 1, 1997, pp. 1-23.
9
Murray B. Anderson, “Genetic Algorithms in Aerospace Design: Substantial Progress, Tremendous Potential”, Sverdrup
Technology Inc. /TEAS Group Eglin Air Force Base, FL 32542, USA.
10
J. L. Walsh,J. C. Townsend, A. O. Salas,J. A. Samareh, V. Mukhopadhyay, and J.-F. Barthelemy, “Multidisciplinary High-
Fidelity Analysis and Optimization of Aerospace Vehicles, Part 1: Formulation”, AIAA-2000-0418
11
Olds, J., “System Sensitivity Analysis Applied to the Conceptual Design of a Dual- Fuel Rocket SSTO." AIAA Paper 94-
4339, Sept. 1994.
12
William A. Crossley and Edwin A. Williams, “A Study of Adaptive Penalty Functions For Constrained Genetic
Algorithmbased Optimization”, Purdue University West Lafayette, IN 47907-1282.
13
M. Stein, “Large sample properties of simulations using Latin Hypercube Sampling”, Technometrics, 29:143-151, 1987.
14
Kleijnen, J. P. C., Statistical Tools for Simulation Practitioners, Marcel Dekker, New York, NY, 1987.
15
Jin, R., Chen, W., and Simpson, T. W., 2001, “Comparative Studies of Metamodeling Techniques under Multiple
Modelling Criteria,” Journal of Structural Optimization, 23(1), pp. 1-13.
16
Stella M. Clarke, Jan H. Griebsch and Timothy W. Simpson,” Analysis of Support Vector Regression for Approximation
of Complex Engineering Analyses”, ASME 2003 Design Engineering Technical Conferences and Computers and Information in
Engineering Conference Chicago, Illinois USA, September 2-6, 2003
17
J.A.K. Suykens, T. Van Gestel, J. De Brabanter, B. De Moor, J. Vandewalle, Least Squares Support Vector Machines,
World Scientific, Singapore, 2002 (ISBN 981-238-151-1)
18
Johan A K Suykens, Gestel T Van, Brabanter J De, Moor Bob de, Vandewalle J, “Least Squares Support Vector
Machines”2002.
19
Goldberg D.E., Genetic Algorithms in Search, Optimization, and Machine Learning, 1st edition., Addison Wesley
Longman, Reading, MA, 1989.

14
American Institute of Aeronautics and Astronautics

Das könnte Ihnen auch gefallen