Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

Multi-Speed Gearbox Design

Using Multi-Objective
Kalyanmoy Deb
e-mail: deb@iitk.ac.ub Evolutionary Algorithms
Sachin Jain Optimal design of a multi-speed gearbox involves different types of decision variables and
objectives. Due to lack of efficient classical optimization techniques, such problems are
Kanpur Genetic Algorithms Laboratory usually decomposed into tractable subproblems and solved. Moreover, in most cases the
(KanGAL), explicit mathematical expressions of the problem formulation is exploited to arrive at the
Department of Mechanical Engineering, optimal solutions. In this paper, we demonstrate the use of a multi-objective evolutionary
Indian Institute of Technology Kanpur, algorithm, which is capable of solving the original problem involving mixed discrete and
Kanpur, PIN 208 016, India real-valued parameters and more than one objectives, and is capable of finding multiple
http://www.iitk.ac.in/kangal/pub.htm nondominated solutions in a single simulation run. On a number of instantiations of the
gearbox design problem having different complexities, the efficacy of NSGA-II in handling
different types of decision variables, constraints, and multiple objectives are demon-
strated. A highlight of the suggested procedure is that a post-optimal investigation of the
obtained solutions allows a designer to discover important design principles which are
otherwise difficult to obtain using other means. 关DOI: 10.1115/1.1596242兴

1 Introduction outcome of the optimization process is usually a single optimal


A multi-speed gearbox design optimization problem introduces solution. In order to obtain a set of so-called Pareto-optimal solu-
a number of challenges 关1兴, particularly if the design involves tions for multi-objective optimization, these methods must have to
optimization of both its kinematic structure and gear geometries. be applied many times with different weight or preference vectors.
The resulting optimization problem involves design variables Moreover, most classical approaches demand that all objectives
which can be both integer-valued 共for gear teeth兲, discrete-valued are of the same type 共either all are of minimization type or of
共for gear module兲, and real-valued 共for gear thickness兲. The fea- maximization type兲. One common way to convert a maximization
sibility of a design solution depends on the satisfaction of a num- problem into a minimization problem is to use the inverse func-
ber of equality and inequality constraints involving strength and tion. This conversion facilitates the use of a classical optimization
gear meshing restrictions. The resulting optimization problem be- method, but causes a difficulty in multi-objective optimization.
comes further complicated due to the presence of multiple con- Such conversions do not emphasize the complete range of the
flicting objectives, often desired to be optimized in such real- transformed objective uniformly. Thus, a number of well-
world problems. distributed 共or trade-off兲 solutions in the original objective space
While tackling such complex problems using classical optimi- may be difficult to obtain with a uniformly set of weight vectors
zation methods, researchers usually decompose the overall prob- used in the converted objective space. In this paper, we propose a
lem into tractable subproblems 关2兴. To tackle such problems, a multi-objective evolutionary optimization technique, which is a
multi-step optimization procedure is usually adopted. In each step, more flexible technique and alleviates most difficulties mentioned
a part of the design problem 共for example, involving only integer above. Particularly, the proposed technique—the modified non-
variables兲 is first solved by fixing the rest of the variables to dominated sorting GA or NSGA-II—is capable of handling the
reasonable values. Thereafter, another part of the problem 共say, following complexities often arises in an optimization problem:
involving real-valued decision variables兲 is solved by fixing pre-
vious variables to the values obtained in the first subproblem. This 1. NSGA-II is capable of handling discrete 共including integer兲
procedure is continued till all components of the overall problem and real-valued decision variables together without any spe-
is solved. For obvious reasons, although such a procedure enables cial consideration,
classical optimization techniques to be applied, the final obtained 2. NSGA-II is capable of handling constraints in a simple way
solutions may not necessarily correspond to the true optimal so- and without having to set any new parameter, such as pen-
lutions of the over-all problem. Moreover, such decomposition alty parameters.
technique may not be applicable to problems which cannot be
3. NSGA-II is capable of handling multiple objectives of
decomposed into meaningful subproblems. Some classical optimi-
mixed types easily.
zation methods also face difficulties in searching a constrained
search space having multiple equality and nonlinear inequality Here, the NSGA-II is applied increasingly to more complex ver-
constraints. Generic methods such as the penalty function based sions of the gearbox design problem than what has been attempted
approaches largely depends on the chosen penalty parameter val- in the literature 关2兴. Starting from the two-objective, 10-variable,
ues. real-parameter gearbox optimization problem, NSGA-II shows its
Classical optimization methods for handling multiple objectives ease of application and efficacy to more difficult problems, lead-
共weighted-sum approaches or goal-programming methods兲 require ing to a four-objective, 29-variable, 101-constraint, mixed-integer
the user to supply a weight vector or a preference vector, before
gearbox optimization problem. In addition to the optimal solu-
solving the problem 关3– 6兴. Since such a weight or preference
tions, the paper also suggests a procedure for the possibility of
vector scalarizes multiple objectives into a single objective, the
revealing important design principles for designing a gearbox. The
success of NSGA-II’s application in this problem suggests its im-
Contributed by the Design Automation Committee for publication in the JOUR-
NAL OF MECHANICAL DESIGN. Manuscript received March 2002; rev. Feb. 2003. mediate application to similar or other more complicated and
Associate Editor G. M. Fadel. challenging engineering design problems.

Journal of Mechanical Design Copyright © 2003 by ASME SEPTEMBER 2003, Vol. 125 Õ 609

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/27/2013 Terms of Use: http://asme.org/terms


2 Multi-Speed Gearbox Design optimal solutions can reveal important design principles. In the
following subsections, we outline the formulation of the optimal
In the multi-speed gearbox design problem, the purpose is to
gearbox design problem in a systematic manner.
achieve different speeds in the output shaft, whereas the input
shaft rotates at a fixed angular velocity. To facilitate multiple 2.1 Number of Teeth Fixed. When the number of teeth in
speed reductions, the gearbox contains a number of intermediate each gear is kept fixed, the decision variables are all gear-pair
shafts. Figure 1 shows a typical 18-speed gearbox having 18 thickness values t i 共for i⫽1,2, . . . ,G) and the power delivered p.
gears, placed in a two-dimensional array. Gears in shafts 1, 3 and For the gearbox layout shown in Fig. 1, there are nine gear-pairs,
5 can be translated to get meshed with corresponding gears in or G⫽9. The module is kept the same for all gearpairs. Earlier
shafts 2 and 4. With an input shaft speed of 1400 rpm and with a studies 关2,7兴 have considered two objectives—maximizing the
ratio between consecutive output speeds of 1.14, the i-th desired power delivered f 1 and minimizing the total volume of gear ma-
output shaft speed would be 1400/(1.14) i⫺1 . For the above gear- terial f 2 . A little thought will reveal that these two objectives are
box with 18 options in the output shaft speed, the desired lowest conflicting to each other. There also exist a number of constraints
output speed becomes 1400/(1.14) 17 or 150 rpm. associated with this problem. Each gear must satisfy the bending
The optimal design of such a gearbox then involves searching stress ␴ b and wear stress ␴ w constraints: 共i兲 ␴ b ⭐S b and 共ii兲 ␴ w
for a set of gears 共with adequate number of teeth, thickness, and ⭐S w , where S b and S w are the permissible bending and wear
module兲 which would achieve output shaft speeds closer to the stresses, respectively. The bending and wear stresses developed in
desired values. Moreover, the design should be such that the gear the i-th gear-pair can be calculated as follows 关2兴:
sizes make a feasible plan 共such as to maintain identical center
distance between all combination of mating gears between two 97500pk c k d 共 r i ⫹1 兲
␴ bi⫽ , (1)
shafts兲 and gears are strong from bending and wear consider- ␻ i a i t i mr i y i cos ␤


ations. Finding a feasible design satisfying all the above con-
straints is by itself a difficult task. However, it is often desired to 0.59共 r i ⫹1 兲 97500pk c k d E 共 r i ⫹1 兲
␴ wi⫽ , (2)
find a solution which is not only feasible but also optimizes an r ia i ␻ i t i sin 2 ␤
objective of design. In this problem, a number of such objectives
are possible: 共i兲 minimization of overall gear material used 共which where k c (⫽1.5) is the stress concentration factor and k d (⫽1.1) is
is directly related to the weight and cost of the gearbox兲, 共ii兲 the dynamic load factor. The parameter r i is the transmission ratio
maximization of power delivered by the gearbox, 共iii兲 minimiza- defined as the ratio of the number of teeth (n w i ) in wheel to the
tion of the center distance between input and output shafts, and number of teeth (n p i ) in pinion 共or, n w i /n p i ) for the i-th gear-pair.
共iv兲 minimization of error between achieved and desired output The parameter ␻ i is the speed of the wheel in rpm, a i is the center
shaft speeds. When multiple conflicting objectives are considered distance for the corresponding gear-pair 共calculated as m(n w i
共such as 共i兲 and 共ii兲兲, there are usually more than one optimal ⫹n p i )/2, where m is the module兲 in cm and t i is the thickness of
solution to the resulting problem. Such solutions are called Pareto- the gear-pair in cm. The parameter y i is the form factor defined as
optimal solutions or efficient solutions or non-inferior solutions follows: y i ⫽0.52(1⫹20/n w i ), where n w i is the number of teeth
关11兴. Although only one of the solutions have to be recommended
for fabrication, one approach to multi-objective problem solving for the wheel in the corresponding gear-pair. The pressure angle is
is to first find a widely distributed set of Pareto-optimal solutions ␤ 共⫽20 degrees兲 and the parameter E is the Young’s modulus of
before making any particular choice on the importance of one the gear material in kgf/cm2. The complete optimization problem
objective over the other. Thereafter, one solution can be picked is given below:


using some higher-level decision-making criteria. Here, we would Maximize f 1 ⫽p,
not like to belabor the discussion on what higher-level decision- G
␲m2
兺 共n
making criteria will be the best for a gearbox design problem, as
such a discussion would be subjective to the user and would de- Minimize f 2 ⫽ p i ⫹n w i 兲 t i
2 2
,
4 i⫽1
pend on many non-technical aspects which are beyond the scope
of the present discussion. Instead, we argue that such a decision- Subject to ␴ b i ⭐S b , for i⫽1,2, . . . ,G,
making becomes easier and more pragmatic than the usual prac-
tice of optimizing a single-objective function derived by prefixing ␴ w i ⭐S w , for i⫽1,2, . . . ,G,
a relative preference factor of each objective. As a by-product of 共L兲 共U兲
t ⭐t i ⭐t , for i⫽1,2, . . . ,G,
our approach, we shall demonstrate how an analysis of multiple
p 共 L 兲 ⭐p⭐p 共 U 兲 .
(3)
Thus, for the gear layout shown in Fig. 1, there are a total of 10
real-valued decision variables and 38 inequality constraints. The
problem can be made more complex by treating the module 共m兲
also as a decision variable. Since any arbitrary module is not
desired for difficulties in fabrication, the module can be treated as
a discrete-valued variable taking only a few selected values.
2.2 Number of Teeth Varying. When the number of teeth
in gears are also kept as decision variables, the resulting problem
must involve additional constraints considering the following
three aspects:
1. The center distance between two shafts must be maintained
by all corresponding gear-pairs,
2. No gear should interfere with any shaft. These constraints
arise because we have considered a two-dimensional layout of the
gears having extended shafts. Some of these constraints can be
eliminated by using a three-dimensional gearbox. However, here
Fig. 1 Schematic of a gear train we consider the two-dimensional gearbox shown in Fig. 1.

610 Õ Vol. 125, SEPTEMBER 2003 Transactions of the ASME

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/27/2013 Terms of Use: http://asme.org/terms


3. All obtained output shaft speeds ⍀ k (k⫽1,2, . . . ,K) must The above decomposition procedure of the original problem is a
be within a permissible error limit ⑀ from the ideal speeds ⍀ Ik , clever one, particularly because the first subproblem does not in-
and volve any mechanical strength considerations of the gears and is
4. The maximum gear-ratio in any gear-pair must not exceed a simply a kinematic layout optimization problem involving integer
limit r max. variables only. Two appropriate objectives involving number of
gear teeth are chosen to suit the optimization problem. Having
Moreover, the very nature of these additional decision variables obtained the optimal number of teeth for gears 共which incidentally
共number of teeth兲 requires them to be treated as non-zero positive also minimized the error in output speed deviation from ideal in
integers. For the gear-train layout shown in Fig. 1, we formulate 关2兴兲, the second subproblem focused on the mechanical strength
the above constraints: considerations and optimized two important objectives 共power de-
livered and the overall volume of the gearbox兲. It is important to
n 1 ⫹n 2 ⫽n 3 ⫹n 4 ⫽n 5 ⫹n 6 , 共 between shaft 1 and 2 兲 (4)
note that there are at least two difficulties with the above
n 7 ⫹n 8 ⫽n 9 ⫹n 10⫽n 11⫹n 12 , 共 between shaft 2 and 3 兲 decomposition-based approach:
(5) 1. First of all, such a clever decomposition is difficult to gen-
n 15⫹n 16⫽n 17⫹n 18 , 共 between shaft 4 and 5 兲 (6) eralize for any arbitrary problem. A great deal of specific
problem knowledge as well as expertise with the optimiza-
n 2 ⭐n 7 ⫹n 8 , n 4 ⭐n 7 ⫹n 8 , n 6 ⭐n 7 ⫹n 8 , n 7 ⭐n 1 ⫹n 2 , tion literature is needed to come up with a dichotomy of a
(7) problem for the use of suitable methods.
2. Secondly, even after cleverly decomposing a problem and
n 9 ⭐n 1 ⫹n 2 , n 11⭐n 1 ⫹n 2 , n 8 ⭐n 13⫹n 14 , n 10⭐n 13⫹n 14 , solving each subproblem, the obtained final solution need
(8) not to be the true optimal solution of the original problem.
This is because, in complex problems having interacting
n 12⭐n 13⫹n 14 , n 13⭐n 7 ⫹n 8 , n 14⭐n 15⫹n 16 ,
variables, the individual optimal solutions to parts of the
n 15⭐n 13⫹n 14 , (9) complete problem need not necessarily be combined to ob-
tain the true overall optimal solution.
n 17⭐n 13⫹n 14 , (10)
Having objectives of different types the classical approaches pose
兩 ⍀ k ⫺⍀ Ik 兩 ⭐ ⑀ ⍀ Ik , k⫽1,2, . . . ,K, (11) another difficulty. Since the second subproblem involves maximi-
zation of delivered power and minimization of volume of the
n wi gearbox, both past studies converted the first objective into the
⭐r max for i⫽1,2, . . . ,G, (12) minimization of the inverse of the delivered power. Although, in
n pi principle, optimization of such a converted objective should result
in the same optimal solution as that would be obtained by opti-
n w i ⭓n 共 L 兲 , for i⫽1,2, . . . ,G, (13) mizing the original objective in the case of a single objective
problem, the same may not be true for multiple objective optimi-
n p i ⭓n 共 L 兲 , for i⫽1,2, . . . ,G, (14) zation. In order to find a well-distributed set of Pareto-optimal
solutions using a classical generating approach 关5兴, such as a re-
n i 苸Z, for i⫽1,2, . . . ,2G. (15) petitive application of the weighted-sum approach as used in 关2兴,
In a mating gear-pair, n w i is assigned to be the one with the larger a uniform set of weight vectors in the converted case may not
produce a well-distributed set of Pareto-optimal solutions of the
number of teeth. Equations 共4兲 to 共6兲 take care of the center dis-
original problem. In the above case, the weighted objective
tance constraint, as mentioned above. Inequalities 7 to 10 take
care of the gear-shaft clearances. Inequality 11 must be satisfied
for all output shaft speeds 关2兴. There are a total of K 共18 in this
case兲 such constraints. The next set of constraints take care of the
F⫽w 1 共 volume兲 ⫹w 2 冉冊
1
p
maximum permissible gear ratio requirement in all gear-pairs. The will produce more dense solutions having smaller values of p with
next two set of constraints restrict the number of teeth in each gear a uniform set of weight vectors (w 1 ,w 2 ) T than that having a large
to have a reasonable lower bound and the final constraint enforces value of p.
that the number of teeth be integers. Thus, for the gearbox layout Some of the above difficulties associated with the classical ap-
shown in Fig. 1, there are five equality constraints and 58 inequal- proaches can be alleviated using an evolutionary optimization ap-
ity constraints involving the gear teeth variables. If the gear thick- proach, and a more systematic, flexible, and a combined optimi-
ness, teeth, power, and module are all kept as variables, there are zation task can be achieved. In the following section, we describe
10 real-valued, one discrete-valued, and 18 integer-valued vari- a specific multi-objective evolutionary optimization algorithm for
ables in the gearbox design problem. Moreover, there are five this purpose and present simulation results in the next section.
linear equality constraints and 96 nonlinear inequality constraints.
2.3 Past Studies and Their Shortcomings. Past efforts to 3 Multi-Objective Evolutionary Algorithms
solve this complex optimal gearbox design problem 关2,7兴 have
decomposed the overall problem into tractable subproblems to Over the past decade or so, there have been an increasing in-
enable them to be solved by classical methods. Essentially, the terest in using evolutionary algorithms 共EAs兲 in multi-objective
following decomposition was used: optimization. The main reason of their use is the ability of EAs to
find multiple optimal solutions simultaneously in a single simula-
1. The design problem having gear teeth as the only design tion run. With the advantage of inherent parallel processing of
variables was solved first in order to minimize two objec- EAs, multi-objective EAs 共MOEAs兲 can efficiently find a set of
tives: 共i兲 the error in output shaft speeds from ideal speeds well-distributed nondominated solutions close to the true Pareto-
and 共ii兲 the center distance between input and output shafts. optimal front. The first author’s recent book on the topic 关8兴 dis-
2. A gearbox with gear thicknesses as only decision variables is cusses a number of MOEAs and presents a number of application
optimized for two different objectives: 共i兲 minimization of case studies. Instead of using any weight information right in the
the volume of gear material and 共ii兲 maximization of the beginning of the design process, MOEAs first find a set of well-
power delivered. Here, the number of teeth in each gear is distributed set of trade-off solutions. Since an MOEA works with
kept fixed to that obtained in the first subproblem. a population of solutions in each iteration, a number of such so-

Journal of Mechanical Design SEPTEMBER 2003, Vol. 125 Õ 611

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/27/2013 Terms of Use: http://asme.org/terms


lutions are possible to obtain in one simulation run. Thereafter, the Mutation: Each variable in an offspring created by the recom-
obtained solutions are analyzed and a particular solution is cho- bination operator is perturbed in its vicinity with a mutation prob-
sen. Although some problem information may have to be used ability p m . For real-valued variables, the mutated value is calcu-
here, it is argued that the presence of a number of trade-off solu- lated as follows:
tions facilitates the use of the problem information better than the
absence of any such solution, as is the case for the classical y 共i 1,t⫹1 兲 ⫽x 共i 1,t⫹1 兲 ⫹ 共 x 共i U 兲 ⫺x 共i L 兲 兲¯␦ i , (19)
preference-based approaches to multi-criterion optimization. In where the parameter ¯␦ i is calculated from the zero-mean polyno-
the following subsection, we briefly describe the specific MOEA mial probability distribution using a uniform random number r i
used in this study. 苸 关 0,1兴 , as follows:

3.1 Nondominated Sorting Genetic Algorithm „NSGA-II….


NSGA-II begins it search with a population P of N randomly
¯␦ ⫽
i 再 共 2r i 兲 1/共 ␩ m ⫹1 兲 ⫺1,
1⫺ 关 2 共 1⫺r i 兲兴 1/共 ␩ m ⫹1 兲
,
if r i ⬍0.5,
if r i ⭓0.5.
(20)

created 共feasible or infeasible兲 solutions. The population members


are then divided into a number of subpopulations Fi according to The parameter ␩ m has a similar effect of that of ␩ c . For variables
an increasing level of nondomination 共so that P⫽艛 i Fi ). This coded in binary strings, a bit-wise mutation, which flips a bit with
procedure requires repetitive identification of nondominated solu- the probability p m , is used.
tions in the population and exclusion of identified nondominated Elite-preservation: After the offspring population Q is calcu-
solutions from the population. This procedure requires O(M N 2 ) lated by a serial application of the above three operators, both
computational complexity, where M is the number of objectives. parent P and offspring Q are combined to form a new population
Thereafter, each solution in a subpopulation Fi , is assigned a R of size 2N. The final operation of NSGA-II selects the best N
crowding distance value, equal to the sum of the difference in the solutions from this combined population R in a systematic man-
normalized objective function values between the two neighbor- ner. The population R is sorted according to different nondomina-
ing solutions in each objective. Thus, this distance value for a tion levels and each subpopulation from the best nondomination
solution captures the extent of crowding by other nondominated level is accepted one at a time, till no further subpopulation can be
solutions. If the neighboring solutions are close to a solution in the considered. If the final subpopulation is too big to be accepted in
objective space, the corresponding crowding distance is small. its totality, only those with the largest crowding distance values
Based on the nondomination level and crowding distance of each are chosen to fill up the population. This elite-preserving proce-
population member, three operators—tournament selection, re- dure allows preservation of good solutions from both parent and
combination and mutation—are applied to the population one af- offspring populations. This completes one generation of the
ter another. The details of NSGA-II can be found in 关9兴. We de- NSGA-II procedure. The overall computational complexity of one
scribe these operations in the following paragraphs. generation of NSGA-II to O(M N 2 ).
Tournament selection: The population is shuffled in a random The final accepted population of size N becomes the new parent
order and two solutions at a time are picked from the top of the population and the above procedure is continued for a number of
list for a tournament selection. The solution with a better non- generations till a termination criterion is satisfied. The continuous
domination level wins the tournament. In the case of a tie in the emphasis of nondominated solutions and preservation of less-
nondomination level, the solution with the larger crowding dis- crowded solutions in the population cause NSGA-II to converge
tance value wins. This way, better and well-dispersed nondomi- closer to the Pareto-optimal solutions with a well-distributed set
nated solutions are emphasized during the tournament selection of solutions.
procedure, thereby ensuring that a good spread of nondominated 3.2 Constraint Handling. For handling constraints in
solutions will be maintained at the end of a NSGA-II run. NSGA-II, we simply use a modified definition of the domination
Recombination: Two solutions 共called parents兲 are picked ran- principle:
domly from the mating pool at a time and two new solutions
共called offspring兲 are created using the following recombination Definition 1 A solution x( i ) is said to ‘‘constraint-dominate’’ a
operator. Each recombination is performed with a probability p c . solution x( j ) (or x( i ) ⭐ c x( j ) ), if any of the following conditions are
For handling real-valued decision variables, recombination is per- true:
formed variable-wise using the SBX operator 关10兴. The i-th vari- 1. Solution x( i ) is feasible and solution x( j ) is not.
able of two parents x i( 1,t ) and x i( 2,t ) is used to calculate two new 2. Solutions x( i ) and x( j ) are both infeasible, but solution x( i )
values for the offspring solution as follows: has a smaller constraint violation.
3. Solutions x( i ) and x( j ) are feasible and solution x( i ) domi-
x 共i 1,t⫹1 兲 ⫽0.5关共 1⫹ ␤ q i 兲 x 共i 1,t 兲 ⫹ 共 1⫺ ␤ q i 兲 x 共i 2,t 兲 兴 , (16) nates solution x( j ) in the usual sense.
The advantages of using the above definition in sorting a popu-
lation into different levels of nondomination are as follows: 共i兲
x 共i 2,t⫹1 兲 ⫽0.5关共 1⫺ ␤ q i 兲 x 共i 1,t 兲 ⫹ 共 1⫹ ␤ q i 兲 x 共i 2,t 兲 兴 , (17) any infeasible solution is dominated by any feasible solution, 共ii兲
between two infeasible solutions, the one with a smaller overall
where ␤ q i is computed so as to assign more probability for creat- constraint violation 共or in some sense, closer to a constraint
ing new solutions closer to the parent solutions using a uniform boundary兲 dominates the other. Here, constraints are first normal-
random number u i 苸 关 0,1兴 : ized and constraint violations, if any, are simply added together to

再冉
compute the overall constraint violation, and 共iii兲 between two
共 2u i 兲 1/␩ c ⫹1 , if u i ⭐0.5; feasible solutions, the usual domination principle is used.
␤ qi⫽ 1
2 共 1⫺u i 兲 冊 1/␩ c ⫹1
, otherwise.
(18)
4 Computational Results
In this section, we investigate the efficacy of NSGA-II in opti-
The parameter ␩ c is the distribution index of the SBX operator. mizing the gear-train design problem depicted in Fig. 1. We
For handling bounded variables, Eq. 共18兲 is slightly modified 关11兴, present a total of five cases, starting from a two-objective, 10-
so that there is zero probability of creating a solution outside the variable, 38-constraint real-parameter gearbox optimization prob-
variable bounds. For variables coded in binary strings, a single- lem and ending with a four-objective, 29-variable, 101-constraint,
point recombination, which swaps the bits on the right side of a mixed-integer gearbox optimization problem. In all cases, the
randomly chosen cross-site along the string 关12兴, is used here. same NSGA-II code 共A C code implementation which can be

612 Õ Vol. 125, SEPTEMBER 2003 Transactions of the ASME

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/27/2013 Terms of Use: http://asme.org/terms


objective evolutionary algorithm—NSGA-II—to the decomposed tion problem. This is due to the fact that in order to reveal any
problems has able to find the same or better Pareto-optimal front such properties, it is first necessary to find a set of well-distributed
as that obtained by classical methods. The other highlights of this Pareto-optimal solutions. Since the NSGA-II allows a tractable
paper are as follows: way of achieving a well-distributed set of nondominated solutions
close to the true Pareto-optimal front, such systematic applications
1. NSGA-II has found as many as 300 different trade-off opti- and an analysis of the resulting solutions are possible.
mal solutions in one single simulation run, as opposed to the need
of a repetitive use of any classical method to find multiple optimal References
solutions. 关1兴 Carroll, R. K., and Johnson, G. E., 1984, ‘‘Optimal Design of Compact Spur
2. NSGA-II has been able to solve different complexities of the Gear Sets,’’ ASME J. Mech., Transm., Autom. Des., 106, pp. 95–101.
gearbox design problem without any major change in its param- 关2兴 Jain, P., and Agogino, A. M., 1990, ‘‘Theory of Design: An Optimization
eter settings. Perspective,’’ Mech. Mach. Theory, 25共3兲, pp. 287–303.
关3兴 Chankong, V., and Haimes, Y. Y., 1983, Multiobjective Decision Making
3. The obtained solutions have been analyzed to discover im- Theory and Methodology, North-Holland, New York.
portant design principles related to an optimal design of a gear- 关4兴 Ehrgott, M., 2000, Multicriteria Optimization, Springer, Berlin.
box: 关5兴 Miettinen, K., 1999, Nonlinear Multiobjective Optimization, Boston, Kluwer.
关6兴 Rao, S. S., 1984, Optimization: Theory and Applications, Wiley, New York.
• For an identical output speed requirement, an optimal gear- 关7兴 Rao, S. S., and Eslampour, H. R., 1986, ‘‘Multistage Multiobjective Optimi-
box requires a larger module for larger delivered power. The zation of Gearboxes,’’ ASME J. Mech., Transm., Autom. Des., 108, pp. 461–
468.
optimal module varies proportional to the squared root of the 关8兴 Deb, K., 2001, Multi-Objective Optimization Using Evolutionary Algorithms,
desired delivered power. Wiley, Chichester, UK.
• However, in the event of impracticalities in using a large 关9兴 Deb, K., Agrawal, S., Pratap, A., and Meyarivan, T., 2002, ‘‘A Fast and Elitist
module for high-powered gears, it is better to fix the module Multi-Objective Genetic Algorithm: NSGA-II,’’ IEEE Transactions on Evolu-
tionary Computation, 6共2兲, pp. 182–197.
to be as large as possible and use thicker gears rather than a 关10兴 Deb, K., and Agrawal, R. B., 1995, ‘‘Simulated Binary Crossover for Continu-
gear with more teeth. ous Search Space,’’ Complex Syst., 9共2兲, pp. 115–148.
• For low-powered gearboxes, the wear stress failure is more 关11兴 Deb, K., and Goyal, M., 1998, ‘‘A Robust Optimization Procedure for Me-
critical than bending stress failure and for high-powered gear- chanical Component Design Based on Genetic Adaptive Search,’’ ASME J.
Mech. Des., 120共2兲, pp. 162–164.
boxes, the opposite is true. 关12兴 Goldberg, D. E., 1989, Genetic Algorithms for Search, Optimization, and Ma-
chine Learning, Addison-Wesley, Reading, MA.
Such important yet generic properties of a gearbox design are 关13兴 Krenzer, T. J., and Hotchkiss, R. G., 1986, ‘‘Bevel and Hypoid Gears,’’ J. E.
difficult to obtain by any other procedure, including an analysis of Shigley and C. R. Mishcke, eds., Standard Handbook of Machine Design,
the mathematical programming problem describing the optimiza- pages 34.1–34.61, McGraw-Hill, New York.

Journal of Mechanical Design SEPTEMBER 2003, Vol. 125 Õ 619

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/27/2013 Terms of Use: http://asme.org/terms


Fig. 4 Variation of gear thickness in obtained solutions
Fig. 3 NSGA-II solutions for the case of module being an ad-
ditional variable are shown against four Pareto-optimal solu-
tions of Case I „as reported in †2‡…

where A is a constant which depends on the critical gear-pair


4.2 Case II: Varying Gear Module and Thickness. We configuration. The parameter ␴ b,w represents either ␴ b or ␴ w ,
first allow the module of the gear train to vary in the range depending on the corresponding constraints which are active at
关0.10,0.41兴 cm in steps of 0.01 cm. This way, a discrete set of 32 the Pareto-optimal solutions. For the gear teeth configuration used
different modules are considered. We handle this decision variable here, it has been shown in 关2兴 that the bending stress constraints
using a five-bit binary substring. Thus, a complete solution to the are always more critical than the wear stress constraints. Figure 6
NSGA-II is a binary substring coding the gear module and 10 plots the bending stress ␴ b for all obtained nondominated solu-
real-parameter values representing nine gear-pair thicknesses and tions. It is clear from the plot that except for gear-pairs 2 and 3, all
the power delivered. Here, a bit-wise mutation with probability of other gear-pairs almost satisfy the bending stress constraint with
0.2 is used for the binary substring. Figure 3 shows the obtained the equality sign 共bending strength being almost equal to 2,500
solutions with a population size of 300. NSGA-II is run for 10,000 kgf/cm2兲. Thus, in those cases, equation 1 can be used to calculate
generations, taking an average of 660 seconds on the same ma- the gear thickness values. As also evident from Fig. 4, the allo-
chine as used in Case I. Four solutions of Case I 共in circles兲 cated thicknesses in gear-pairs 2 and 3 are equal to the chosen
reported in 关2兴 are also marked for comparison in the figure. It is lower limit 共or 0.5 cm兲. It is also obvious from Fig. 6 that the
clear that NSGA-II with the flexibility of choosing a suitable mod- critical constraints are the seven inequality bending stress con-
ule produces a better set of nondominated solutions, particularly straints 共1 and 4 to 9兲 satisfied with an equality sign. Thus, from
in the range of gearboxes with smaller delivered power. Here, we equation 1, we can write p⫽ ␬ m 2 t i for these seven gear-pairs,
have obtained solutions having the power varied in the range where ␬ is a constant. Now, in the expression for f 1 共volume兲, we
关1.00,18.28兴 kW, as compared to 关1.66,8.56兴 kW obtained in Case can replace t i 共for i⫽1,4, 5, 6, 7, 8, 9兲 by using p/( ␬ m 2 ). For i
I. It is also interesting to note that the entire nondominated front is ⫽2 and 3, t i is constant and is equal to t ( L ) ⫽0.5 cm. Thus, the
now arranged in an ascending order of magnitude of the module. relationship between f 1 and p is as follows:
The inset figure demonstrates this fact. Starting from a module
very close to the chosen lower bound, solutions with a gradual
increase in module are found. The module values are marked in
the inset figure. The top-right portion of the obtained front corre-
sponds to a module of 0.41 cm 共the chosen upper bound on mod-
ule兲 spanning the power delivered in the range 关17.30,18.28兴 kW.
These results provide an useful insight to the gearbox design
problem. If a large power transmission is desired, it is optimal to
use gears with a large module rather than to use gears with large
thickness and a small module. Figure 4 shows the variation of
thickness of each gear-pair in solutions obtained by NSGA-II. It is
clear from the plot that the gear thickness values do not change
much from one solution to another, but it is clear from Fig. 5 that
a monotonic increase in the module is needed for larger power
transmission. Another interesting aspect is that the rate of increase
in the value of module requirement reduces for large power trans-
missions. Such information are useful to designers and practitio-
ners and are often difficult to achieve by simply analyzing the
mathematical expressions describing the objective functions and
constraints.
However, by assuming that thickness of all gears are more or
less the same for all Pareto-optimal solutions, equation 1 or 2
reveals the following relationship between the power and module:
p⫽A ␴ b,w m 2 , (21) Fig. 5 Variation of gear module in obtained solutions

614 Õ Vol. 125, SEPTEMBER 2003 Transactions of the ASME

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/27/2013 Terms of Use: http://asme.org/terms


Fig. 6 The bending stress of NSGA-II solutions are shown
Fig. 7 NSGA-II solutions are shown with two limiting errors in
against the delivered power
output speeds. Gear teeth and thickness are varied here.

9 3
cm 2 p cm 2 n 3 ⫽n 1 ⫹⌬n 1,3 , n 5 ⫽n 1 ⫹⌬n 1,3⫹⌬n 3,5 .
兺 兺 共n
(23)
f 1⫽ 共 n w2 i ⫹n 2p i 兲 ⫹ w i ⫹n p i 兲 0.5,
2 2
4 i⫽1 ␬m 2 4 i⫽2 By restricting these incremental variables (⌬n i ) to be greater than
i⫽2,3
zero, we ensure that n 5 ⬎n 3 ⬎n 1 . With these considerations, the
⫽D 1 p⫹D 2 m 2 . total number of independent variables kept for representing num-
ber of gear teeth are 13, thereby making the total number of vari-
Since, there is only one module for the entire gear system, the ables to 23, causing a reduction of five integer variables. In the
module m at a particular power p must satisfy Eq. 共21兲, leading to NSGA-II implementation, we allow these incremental variables
p⬀m 2 or, m 2 ⬀p. The exponent of the polynomial variation of (⌬n i ) to vary in 关1,31兴 in steps of one and n i variables to vary in
module with power as observed in Fig. 5 is close to 0.5, thereby 关18,81兴 in steps of one. The recombination and mutation operators
supporting the above argument2. Therefore, the above equation treat these variables as real-valued at first. Before the objective
yields f 1 ⬀p. Figure 3 verifies that the relationship between ob- functions and constraints are evaluated, they are rounded to their
jectives for the Pareto-optimal solutions 共passing through the ori- nearest integers. For these integer variables, we use ␩ c ⫽2 and
gin, if extrapolated兲 is linear, thereby supporting our argument ␩ m ⫽50 in Eqs. 18 and 20, respectively. Figure 7 shows the ob-
above. tained nondominated solutions for ⑀⫽0.05 and 0.10. It is clear
It is important to mention that such mixed-integer programming from the figure that if a large error 共⑀兲 in the output shaft speeds
problems 共having both discrete and continuous decision variables兲 from the ideal is permitted, better nondominated solutions are
are difficult to handle using classical optimization methods. Since expected. The figure also marks the four Pareto-optimal solutions
evolutionary algorithms require only function information for any reported elsewhere 关2兴 obtained by using a fixed number of teeth
given solution and are flexible in handling mixed types of vari- and module in gears 共Case I兲. Since the number of teeth in gears
able, such problems are comparatively easier to handle. are also kept as variables here, a more flexible search is permitted,
4.3 Case III: Varying Gear Teeth and Thickness. Next, thereby obtaining a better set of nondominated solutions. Table 2
we vary the number of teeth in all 18 gears 共nine gear-pairs兲, in shows the gear teeth and thickness of four selected solutions 共uni-
addition to the nine gear-pair thicknesses and power delivered. We formly spaced across the obtained front兲. It is clear from the table
fix the module at 0.28 cm. In order to handle the constraints given that for gearboxes with a large delivered power, thickness and
in equations 4 to 6, we use following two strategies with NSGA- teeth in gears are also large. The table also shows the correspond-
II: ing gear details for ⑀⫽0.05. For the gear sizes, a trend similar to
Variable elimination: We use the first three linear equality con- that found in the ⑀⫽0.10 case is also observed here. Although the
straints to eliminate five decision variables. For example, the first latter solutions cause a maximum of ⑀⫽0.05 error in the output
equation can be used to eliminate two variables as follows: speeds from the desired, each of four solutions tabulated above
found a somewhat similar yet different gear teeth combination to
n 4 ⫽n 1 ⫹n 2 ⫺n 3 , n 6 ⫽n 1 ⫹n 2 ⫺n 5 . (22) achieve the same objective. In order to investigate the efficacy of
Redundancy reduction: The redundancy in solutions is reduced NSGA-II in handling constraints, we calculate the percentage
by prefixing an order of gear sizes in a gear-pair cluster. For maximum error in the output shaft speed for each obtained non-
example, the redundancy in solutions can be reduced by assuming dominated solution. Figure 8 shows the proportion of obtained
that n 5 ⬎n 3 ⬎n 1 . By this way several permutations of these gear nondominated solutions having different percentage maximum er-
sizes will not be reconsidered during optimization. ror. For clarity, all solutions with a percentage maximum error in
The latter condition is further made simpler by using a different a block of 0.25% are grouped together. The figure shows that a
set of decision variables. For example, instead of using n 1 , n 3 , large proportion of nondominated solutions maintain a percentage
and n 5 as decision variables, we may use n 1 , ⌬n 1,3 , and ⌬n 3,5 as maximum error close to the maximum permissible value. This
decision variables and derive n 3 and n 5 as follows: indicates that the constraint handling strategy used in NSGA-II is
effective in exploiting the chosen maximum error limit in the
2
It is also interesting to note that for a single bevel or hypoid gear-pair, Krenzer
output speeds.
and Hotchkiss 关13兴 have also suggested the use of gear diameters increasing polyno- With a specific number of gear teeth used in Case I, it was
mially with transmitted torque having an exponent of 0.34. observed that the bending stress constraints are always more criti-

Journal of Mechanical Design SEPTEMBER 2003, Vol. 125 Õ 615

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/27/2013 Terms of Use: http://asme.org/terms


0.50共0.50兲

1.29共1.29兲

1.78共1.65兲

1.82共1.82兲
9

23共23兲
28共30兲

24共18兲
37共47兲

24共27兲
49共58兲

31共36兲
72共72兲
0.51共0.50兲

1.88共1.80兲

2.46共2.30兲

2.50共2.49兲
8

18共18兲
33共35兲

18共22兲
43共30兲

18共20兲
55共65兲

23共27兲
80共81兲
Gear details of four selected solutions in Case III with a maximum of 10% error. The results with at most 5% error are shown in brackets.

0.50共0.50兲

0.92共1.10兲

1.06共1.25兲

1.08共1.30兲
7

18共18兲
32共28兲

19共22兲
38共29兲

20共22兲
39共35兲

30共36兲
47共45兲
0.50共0.50兲

0.50共0.55兲

0.66共0.67兲

0.72共0.91兲

Fig. 8 Proportion of obtained solutions having different maxi-


mum percentage error in speed ratios is shown for two cases
of optimization „with 5% and 10% limiting error…
6

20共22兲
21共24兲

25共29兲
32共32兲

26共28兲
39共36兲

33共39兲
42共42兲

cal than the wear stress constraints. However, in this case 共when
Gear-Pair Number

the number of gear teeth are also kept as decision variables兲, the
0.50共0.50兲

0.51共0.59兲

0.66共0.70兲

0.75共0.95兲

bending stress need not always be critical. To investigate which


constraint is critical in which scenario, we have plotted B i /C i for
5

19共21兲
22共25兲

26共30兲
31共31兲

27共30兲
38共34兲

35共40兲
40共41兲

all nine gear-pairs versus p in Fig. 9 for all obtained nondomi-


nated solutions 共where the bending constraints with the equality
sign are expressed as p⫽B i t i and the wear constraints with the
equality sign are expressed as p⫽C i t i 关2兴兲 It can be seen from the
0.50共0.50兲

0.55共0.64兲

0.68共0.73兲

0.79共1.07兲

figure that the ratio B i /C i is more than one for small values of p
indicating that wear stress constraints are more critical than bend-
4

18共19兲
23共27兲

28共28兲
29共33兲

29共32兲
36共32兲

37共37兲
38共44兲

ing stress constraints. Using the wear stress constraints, the para-
metric relationship of the Pareto-optimal front can be written as
follows:
9
n w i n 2p i
0.50共0.50兲

0.50共0.50兲

0.50共0.50兲

0.50共1.57兲

f 2⫽ ␥ 兺
i⫽1
共 n 2p i ⫹n w2 i 兲 t i , p⫽ ␣ i t i
n p i ⫹n w i
.
3

18共18兲
37共38兲

18共19兲
42共41兲

18共18兲
43共50兲

19共24兲
54共57兲

A dimensional study on the design variables reveal a linear rela-


tionship between the two objectives f 2 and p, as observed in Fig.
7. On the other hand, for solutions with a large value of p, bending
0.50共0.50兲

0.50共0.50兲

0.50共0.50兲

0.50共1.25兲
2

27共28兲
28共28兲

28共30兲
32共30兲

28共30兲
33共38兲

31共39兲
42共42兲
0.50共0.50兲

0.50共0.50兲

0.75共0.50兲

1.01共0.93兲
1

18共18兲
37共38兲

21共19兲
39共41兲

22共25兲
39共43兲

29共27兲
44共54兲
Table 2

n wi

n wi

n wi

n wi
n pi

n pi

n pi

n pi
ti

ti

ti
共381.95兲

共941.52兲

共1718.20兲

共3537.46兲
1443.68

2712.36
Volume

355.03

846.52
共cc兲

共1.050兲

共5.547兲

共9.094兲

共11.720兲
12.262
Power

1.040

5.503

9.008
共kW兲

Fig. 9 The ratio B i Õ C i for all nine gear-pairs observed in dif-


Sol.
No.

ferent obtained solutions


ti
1

616 Õ Vol. 125, SEPTEMBER 2003 Transactions of the ASME

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/27/2013 Terms of Use: http://asme.org/terms


Fig. 10 NSGA-II solutions for varying gear module, teeth, and
Fig. 11 Module m obtained in different nondominated solu-
thickness are shown. Nondominated fronts for two different
tions for two cases
ranges of modules are compared with that obtained in Case III
„with a fixed module of 0.28 cm…. Open circles indicate results
when the module is allowed to vary within †0.10,0.41‡ cm and
the triangles indicate results for modules varying in †0.10,0.73‡
cm.
increase monotonically with the delivered power. For these solu-
tions, the corresponding nondominated front is linear, as evident
from Fig. 10 and as explained in Case II. Thereafter, the module
remains fixed to the upper limit. For these solutions, the optimized
stress constraints are more critical. With a large power require- volume-power relationship is nonlinear in a manner similar to that
ment, the gear sizes also need to be large and are critical for observed in Case III. It is also interesting to note that the nonlin-
failure due to bending than wear. Using the bending stress con- ear portion of the nondominated front will get dominated, if larger
straints, the parametric relationship becomes modules are allowed. To demonstrate this fact we have rerun

冉 冊
NSGA-II with a relaxed upper bound 共module varying in the
9
n 2p i 20 range 共0.10,0.73兲 cm兲. The corresponding nondominated front ob-
f 2⫽ ␥ 兺
i⫽1
共 n 2p i ⫹n w2 i 兲 t i , p⫽ ␤ i t i
n wi
1⫹
n wi
.
tained after 10,000 generations are shown using triangles in Fig.
10. It is evident that now the obtained front remains linear till the
As can be seen from these equations, for an increase in the num- module of the gearbox becomes equal to the chosen upper limit
ber of gear teeth and thickness the rate of change of f 2 is more 共0.73 cm兲. Figure 11 also shows that the corresponding module
here than that in the former case. Figure 7 depicts this non-linear increases monotonically and in a very similar manner (m⬀ 冑p) as
relationship among the obtained solutions for large-powered gear- that obtained in the earlier case 共with m ( U ) ⫽0.41 cm).
boxes.
4.5 Case V: Three Objectives. To investigate further the
4.4 Case IV: Varying Gear Module, Teeth and Thickness. performance of NSGA-II on more than two objectives, we formu-
Next, we make the gearbox design problem even more flexible by late a third objective function of minimizing the center distance
allowing the module also as a decision variable. The range and between input and output shafts. This objective function can be
step used for this variable are kept the same as those used in Case written as follows:
II (m苸 关 0.10,0.41兴 cm in steps of 0.01 cm兲. This increases the m
total number of variables to 29. An ⑀⫽0.10 is used in this study. f 3⫽ 关共 n 1 ⫹n 2 兲 ⫹ 共 n 7 ⫹n 8 兲 ⫹ 共 n 13⫹n 14兲 ⫹ 共 n 15⫹n 16兲兴 .
Rest all parameters are kept the same as those used in the previous 2
case. Figure 10 shows the 300 nondominated solutions obtained (24)
using NSGA-II after 10,000 generations. On the same computer All constraints, decision variables, and other parameters as used in
mentioned earlier, an NSGA-II run took an average of 660 sec- the previous case are also adopted here. The module is varied in
onds. As in the Case II, a monotonically increasing magnitude of 关0.10,0.41兴 cm in steps of 0.01 cm. In section 2.3, we have dis-
module is observed for solutions having increasing requirement in cussed that previous studies on the gearbox design have consid-
delivered power. Figure 10 also shows the nondominated front ered a total of four different objectives. However, any particular
obtained in Case III 共optimized by keeping the module fixed兲. It is optimization did not consider more than two objectives. Here and
interesting to note that a better nondominated set of solutions is in the next section, for the first time, we show simulation results
possible to obtain by allowing the module to be a decision vari- for the gearbox optimization problem with three and four objec-
able. Two distinct phases can be observed among the obtained tives.
nondominated solutions. For gearboxes having a small delivered For the three-objective case, we have used ⑀⫽0.1. NSGA-II
power the volume-power relationship is linear and for gearboxes with 300 population members takes an average of 700 seconds to
having a large delivered power the relationship is nonlinear. These run 10,000 generations on the same machine mentioned before.
two phases have distinct properties. Although solutions found are satisfactory, in order to achieve a
Figure 11 shows that the variation of module in the obtained better convergence, we have continued the NSGA-II run for
solutions. It is interesting to note that module increases with 100,000 generations. Figure 12 shows 300 nondominated solu-
power requirement in a manner similar to that observed in Case II tions obtained in one such typical run of NSGA-II in triangles.
till the chosen upper limit (m ( U ) ⫽0.41 cm) on module is reached. The projection on the x-y plane shows a similar trade-off 共partly
An investigation on the gear thickness and the number of teeth for linear and partly nonlinear兲 in delivered power and volume as that
these solutions indicates that the same for gear-pairs 7, 8 and 9 observed in the two-objective case 共Case IV with m

Journal of Mechanical Design SEPTEMBER 2003, Vol. 125 Õ 617

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/27/2013 Terms of Use: http://asme.org/terms


Fig. 12 NSGA-II solutions for an additional objective of mini-
mizing center distance are compared with NSGA-II solutions
with two objectives „maximizing power and minimizing volume…

Fig. 14 Trade-off between ⑀ and center distance for two-


苸关0.10,0.41兴 cm). It is also interesting to note the dimension of objective NSGA-II and Jain and Agogino’s †2‡ study
the obtained nondominated front in the three-dimensional objec-
tive space. Although three objectives are considered, a two-
dimensional curve is obtained as the nondominated front. This
may be because the center distance and overall volume of gear objective space. It is observed that the projection of the nondomi-
material are not likely to be conflicting to each other, thereby nated solutions on the volume-power space has an identical rela-
loosing one dimension in the nondominated front. On this figure, tionship as that observed in Case V and the optimal volume-power
we also superimpose the nondominated solutions obtained in Case relationship is almost invariant to any level of ⑀. It is intuitive that
IV 共two-objective case of maximizing power and minimizing vol- gearboxes for a large ⑀ consideration can be made from a number
ume, run up to 100,000 generations兲. The center distance f 3 is of different gear-teeth combinations, thereby allowing a wide
computed for each obtained solution and is plotted on the three- variation in the center distance between input and output shafts.
objective space. It is clear that when all three objectives are con-
sidered in NSGA-II, smaller values of the third objective 共center On the other hand, a gearbox with a very small maximum output
distance兲 is found. This is because in Case IV no selection advan- speed error ⑀ may not correspond to many gear teeth combina-
tage was assigned to achieve a smaller center distance. tions, thereby reducing the range of allowable center distance be-
By investigating the variable values in the two cases 共two and tween the input and output shafts.
three objectives兲, it is observed that a smaller module and larger
gear thicknesses are obtained for the three-objective case. How- 4.7 Case VII: Output Speed Error Versus Center Distance
ever, the number of teeth in the gear-pairs are more or less the Figure 14 shows the obtained nondominated front 共shown in
same in both cases. These properties of the obtained solutions can circles兲, when the gearbox design problem is solved only for two
be explained by analyzing Eq. 共24兲. Since the center distance be- objectives: 共i兲 minimize the maximum error in output speeds and
tween the input and output shafts have no consideration for the 共ii兲 minimize the center distance between the input and output
gear thickness and is directly proportional to the module, shafts. All constraints and variables as used in the four-objective
NSGA-II has optimized the center distance by reducing the mod- case are used here. The trade-off between error and center dis-
ule. To maintain the volume-power optimality, NSGA-II finds tance is clear from the figure. An analysis of the obtained solu-
gearboxes with larger gear thickness values. Thus, the consider- tions reveals that all of them have a fixed module of 0.1 cm 共the
ation of the third objective is found to be beneficial in terms of lower bound兲 and a fixed delivered power of 1 kW 共the lower
reducing the overall size of the gearbox. bound兲. A gearbox with the smallest ⑀ of 0.15 percent and a cor-
responding center distance of 12.5 cm is found. Since both ⑀ and
4.6 Case VI: Four Objectives. Finally, we include the center distance are minimized, the optimal gearboxes are also ex-
fourth objective of minimizing the maximum error in the output pected to be small, requiring a small volume and delivered power.
speeds from the ideal speeds. The number of variables, con- The earlier study 关2兴 with these two objectives and without any
straints, and NSGA-II parameters are kept the same as in Case V. consideration of any mechanical strength constraints reported only
Figure 13 shows the obtained nondominated front on the three- one solution with a center distance of 40.04 cm and 6.28 percent
of error in output speed 共shown with a box兲. On the contrary, the
study here suggests that there exist multiple trade-off solutions
between these two objectives when all constraints are taken into
account and the NSGA-II is capable of finding a number of such
solutions.

5 Conclusions
The multi-speed gearbox design optimization problem consid-
ered in this paper provides an adequate challenge to almost every
aspect of optimization. The design problem involves 共i兲 more than
one 共and as many as four兲 conflicting objectives of design, 共ii兲 a
total of 29 mixed type of decision variables 共integer, discrete, and
real-valued兲, and 共iii兲 five equality and 96 nonlinear inequality
constraints. A couple of previous studies using classical optimiza-
tion methods had to decompose the overall problem into two trac-
table problems with at most 10 variables and 18 inequality con-
Fig. 13 Nondominated solutions for four objectives straints. This paper has shown that the application of a multi-

618 Õ Vol. 125, SEPTEMBER 2003 Transactions of the ASME

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/27/2013 Terms of Use: http://asme.org/terms


objective evolutionary algorithm—NSGA-II—to the decomposed tion problem. This is due to the fact that in order to reveal any
problems has able to find the same or better Pareto-optimal front such properties, it is first necessary to find a set of well-distributed
as that obtained by classical methods. The other highlights of this Pareto-optimal solutions. Since the NSGA-II allows a tractable
paper are as follows: way of achieving a well-distributed set of nondominated solutions
close to the true Pareto-optimal front, such systematic applications
1. NSGA-II has found as many as 300 different trade-off opti- and an analysis of the resulting solutions are possible.
mal solutions in one single simulation run, as opposed to the need
of a repetitive use of any classical method to find multiple optimal References
solutions. 关1兴 Carroll, R. K., and Johnson, G. E., 1984, ‘‘Optimal Design of Compact Spur
2. NSGA-II has been able to solve different complexities of the Gear Sets,’’ ASME J. Mech., Transm., Autom. Des., 106, pp. 95–101.
gearbox design problem without any major change in its param- 关2兴 Jain, P., and Agogino, A. M., 1990, ‘‘Theory of Design: An Optimization
eter settings. Perspective,’’ Mech. Mach. Theory, 25共3兲, pp. 287–303.
关3兴 Chankong, V., and Haimes, Y. Y., 1983, Multiobjective Decision Making
3. The obtained solutions have been analyzed to discover im- Theory and Methodology, North-Holland, New York.
portant design principles related to an optimal design of a gear- 关4兴 Ehrgott, M., 2000, Multicriteria Optimization, Springer, Berlin.
box: 关5兴 Miettinen, K., 1999, Nonlinear Multiobjective Optimization, Boston, Kluwer.
关6兴 Rao, S. S., 1984, Optimization: Theory and Applications, Wiley, New York.
• For an identical output speed requirement, an optimal gear- 关7兴 Rao, S. S., and Eslampour, H. R., 1986, ‘‘Multistage Multiobjective Optimi-
box requires a larger module for larger delivered power. The zation of Gearboxes,’’ ASME J. Mech., Transm., Autom. Des., 108, pp. 461–
468.
optimal module varies proportional to the squared root of the 关8兴 Deb, K., 2001, Multi-Objective Optimization Using Evolutionary Algorithms,
desired delivered power. Wiley, Chichester, UK.
• However, in the event of impracticalities in using a large 关9兴 Deb, K., Agrawal, S., Pratap, A., and Meyarivan, T., 2002, ‘‘A Fast and Elitist
module for high-powered gears, it is better to fix the module Multi-Objective Genetic Algorithm: NSGA-II,’’ IEEE Transactions on Evolu-
tionary Computation, 6共2兲, pp. 182–197.
to be as large as possible and use thicker gears rather than a 关10兴 Deb, K., and Agrawal, R. B., 1995, ‘‘Simulated Binary Crossover for Continu-
gear with more teeth. ous Search Space,’’ Complex Syst., 9共2兲, pp. 115–148.
• For low-powered gearboxes, the wear stress failure is more 关11兴 Deb, K., and Goyal, M., 1998, ‘‘A Robust Optimization Procedure for Me-
critical than bending stress failure and for high-powered gear- chanical Component Design Based on Genetic Adaptive Search,’’ ASME J.
Mech. Des., 120共2兲, pp. 162–164.
boxes, the opposite is true. 关12兴 Goldberg, D. E., 1989, Genetic Algorithms for Search, Optimization, and Ma-
chine Learning, Addison-Wesley, Reading, MA.
Such important yet generic properties of a gearbox design are 关13兴 Krenzer, T. J., and Hotchkiss, R. G., 1986, ‘‘Bevel and Hypoid Gears,’’ J. E.
difficult to obtain by any other procedure, including an analysis of Shigley and C. R. Mishcke, eds., Standard Handbook of Machine Design,
the mathematical programming problem describing the optimiza- pages 34.1–34.61, McGraw-Hill, New York.

Journal of Mechanical Design SEPTEMBER 2003, Vol. 125 Õ 619

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/27/2013 Terms of Use: http://asme.org/terms

You might also like