ARTICLE IN PRESS

Robotics and Computer-Integrated Manufacturing 26 (2010) 212–222

Contents lists available at ScienceDirect

Robotics and Computer-Integrated Manufacturing

journal homepage: www.elsevier.com/locate/rcim

Chaotic particle swarm optimization for assembly sequence planning

Y. Wang, J.H. Liu 
School of Mechanical Engineering and Automation, Beihang University, No.37, Xueyuan Road, Haidian District, Beijing 100083, PR China

a r t i c l e in f o a b s t r a c t

Article history: Assembly sequence planning of complex products is difficult to be tackled, because the size of the
Received 12 May 2008 search space of assembly sequences is exponentially proportional to the number of parts or components
Received in revised form of the products. Contrasted with the conventional methods, the intelligent optimization algorithms
25 March 2009
display their predominance in escaping from the vexatious trap. This paper proposes a chaotic particle
Accepted 20 May 2009
swarm optimization (CPSO) approach to generate the optimal or near-optimal assembly sequences of
products. Six kinds of assembly process constraints affecting the assembly cost are concerned and
Keywords: clarified at first. Then, the optimization model of assembly sequences is presented. The mapping rules
Assembly sequence planning between the optimization model and the traditional PSO model are given. The variable velocity in the
Particle swarm optimization (PSO)
traditional PSO algorithm is changed to the velocity operator (vo) which is used to rearrange the parts in
Chaos method
the assembly sequences to generate the optimal or near-optimal assembly sequences. To improve the
quality of the optimal assembly sequence and increase the convergence rate of the traditional PSO
algorithm, the chaos method is proposed to provide the preferable assembly sequences of each particle
in the current optimization time step. Then, the preferable assembly sequences are considered as the
seeds to generate the optimal or near-optimal assembly sequences utilizing the traditional PSO
algorithm. The proposed method is validated with an illustrative example and the results are compared
with those obtained using the traditional PSO algorithm under the same assembly process constraints.
& 2009 Elsevier Ltd. All rights reserved.

1. Introduction hard to be enumerated for products composed of more than

20 parts even if the process constraints are considered [3].
Assembly sequence planning is significant to design for To meet the demands of product development, the efficient
assembly (DFA) which is one of the fundamental research methods are urgently called to tackle the hard problem to
branches of design for product lifecycle. Its aim is to detect generate the optimal or near-optimal assembly sequences.
feasible assembly sequences or further seek for the optimal Assembly sequences can be classified into sequential or non-
assembly sequence taking various assembly process constraints sequential, coherent or non-coherent, monotone or non-mono-
into consideration. The assembly sequence plays an important tone, linear or non-linear assembly sequences or any combination
role in the assembly plan. Many aspects of assembly process, such of those under special assembly process constraints [3]. The
as assembly line layout, assembly resources utilization, etc., are assembly process constraints include the topological relationships
designed and arranged referring to the assembly sequence. of assembly, the geometrical constrains and precedence relation-
In addition, the reasonable assembly sequences are prone to ships of parts or subassemblies and other assembly process
improve the assembly efficiency and reduce the assembly cost. constraints, such as the assembly directions and tools, accessi-
The generation and optimization of assembly sequences auto- bility, stability and safety [4]. These assembly process constraints
matically are essential to ensure the competitiveness of manu- severely limit the number of feasible assembly sequences and
factured goods and increase the profit margins [1]. many assembly sequences dissatisfying the process constraints
Assembly sequence planning is a typical of NP-complete are discarded. The optimal or near-optimal assembly sequences
problem under assembly process constraints [2]. It is well known belong to the assembly sequences satisfying the assembly process
that the size of the search space of assembly sequences is constraints and the assembly cost or assembly time of them is the
exponentially proportional to the number of parts or components minimum [5]. To generate the optimal assembly sequence, all the
composing the whole products. The total assembly sequences are assembly process constraints relevant to the assembly sequence
planning have to be properly modeled, and the completeness of
the assembly process constraints model has a great effect on the
 Corresponding author. Tel./fax: +86 10 82339174. correctness and quality of the results, while the complexity of
E-mail addresses: [email protected] (Y. Wang), the model directly impacts on the difficulty of the assembly
[email protected] (J.H. Liu). sequence planning problem to be solved. The assembly process

0736-5845/$ - see front matter & 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.rcim.2009.05.003
 ARTICLE IN PRESS
Y. Wang, J.H. Liu / Robotics and Computer-Integrated Manufacturing 26 (2010) 212–222 213

constraints model can be established based on the assembly The adverse situations will be lowered after the chaos method
structure or CAD model and the actual assembly process. is combined with the traditional PSO algorithm. Chaos method
Intelligent optimization algorithms have been viewed as one of [21] has the particular characteristics, such as the randomicity
the efficient means for coping with NP-complete problems. and ergodicity, which can enhance the diversity of the particles
Without exception, assembly sequence planning using intelligent and actuate the particles to move out from the local near-optimal
optimization algorithms has become a research focus. Assembly solutions. The premature convergence of the traditional PSO
sequence optimization using intelligent algorithms integrates algorithm is weakened and the convergence of it is also
modeling, generating and evaluating assembly sequences to- accelerated. Furthermore, the performance to accomplish the
gether, which can generate the optimal or near-optimal assembly global optimization is also improved subject to the chaos method.
sequences and the multiple assembly process constraints are The CPSO approach for assembly sequence planning and
considered simultaneously. Many intelligent optimization algo- optimization is proposed in the paper. First, the principal
rithms, e.g. artificial neural networks (ANN) [6], simulated assembly process constraints are specified and the optimization
annealing (SA) [7], genetic algorithm (GA) [8,9], ant colony model of assembly sequences taking these constraints into
optimization (ACO) [10,11] and artificial immune algorithm (AIA) account is presented. Second, the principle of the traditional
[12], have been applied to assembly sequence planning and PSO algorithm and the chaos method are outlined. The mapping
optimization, while the research and application using particle rules between the optimization model of assembly sequences and
swarm optimization (PSO) algorithm have not yet been investi- the traditional PSO model are given. The process to generate the
gated much at present. preferable assembly sequences utilizing the chaos method and
PSO algorithm was first presented by Kennedy and Eberhart in the optimal or near-optimal assembly sequences computed using
1995 [13]. Thenceforth, PSO algorithm has been successfully the CPSO algorithm is clarified in detail. Finally, the effectiveness
applied to many optimization problems, such as learning of of the method is verified with an example.
artificial neural networks, model predictive control [14]. A brief
and complete overview on the principle, technique and applica-
tion of PSO algorithm is introduced by both Kennedy and Eberhart 2. Optimization model of assembly sequences
[15] and Clerc [16]. Besides these research and application, PSO
algorithm has been broadly applied to the manufacturing Three assumptions are given to simplify the optimization
industry. Liao et al. [17] resolve the complex job-shop scheduling model of assembly sequences.
problem using an improved PSO algorithm in which local heuristic
information is introduced. Schutte et al. [18] implement PSO (1) All the parts are rigid, which means that the parts are not
algorithm for the biomechanical optimization and conclude that deformable in the assembly process. The effect of the
PSO algorithm is easier to be fulfilled than GA algorithm. The assembly tolerance is also neglected.
applications of PSO algorithm in integrated process planning and (2) The assembly directions are restricted to 7X, 7Y, 7Z
scheduling are studied by Guo et al. [19] to achieve manufacturing direction in the three orthogonal coordinate axes.
planning optimization. Several new operators including mutation, (3) Only one part is assembled along one direction and one tool is
crossover and shift are incorporated into the traditional PSO used in one assembly operation process.
algorithm. The results of two experiments show that the
performance of the modified PSO algorithm is improved.
Here, the chaos method and the PSO algorithm is combined
2.1. Assembly process constraints
to the chaotic particle swarm optimization (CPSO) approach
which is used to generate the optimal or near-optimal assembly
sequences. Assembly sequence optimization is to generate feasible
The mechanism of PSO algorithm generating the optimal or assembly sequences with the minimum assembly cost while
near-optimal solutions is a stochastic process. Though the satisfying the given assembly process constraints. First of all,
searching process is stochastic and the grads information is geometrical constraints must be considered to eliminate the
ignored, the ability to pursue the optimal solutions of the particles infeasible assembly sequences so that the search space is reduced
and the collaboration among them ensure that the PSO algorithm remarkably. The other assembly process constraints, such as the
can figure out the optimal or near-optimal solutions quickly. The local assembly precedence, the assembly stability, the change of
formulae of the traditional PSO algorithm are simpler than those the assembly directions, the change of the assembly tools and the
of the SA and the ACO algorithm, which means that the PSO change of the connection types, also have a great effect on the
algorithm can be more conveniently implemented. The popula- assembly cost. The assembly process constraints are specified and
tion size of the PSO algorithm is smaller than those of the SA and illustrated in Fig. 1.
GA algorithm, so the initialization of the populations is simpler
using the PSO algorithm than that of the other intelligent 2.2. Assembly cost model: the optimization model
optimization algorithms. In addition, the population size of the
PSO algorithm is constant and the amount of the solutions is apt The optimal or near-optimal assembly sequences not only
to be controlled in the optimization process. satisfy the geometrical constraints, but also have the minimum
Although the traditional PSO algorithm has many advantages number of parts violating the local assembly precedence np, the
to resolve the optimization problems, the performance of it, such minimum number of the unstable parts ns, the minimum number
as the premature convergence and the stochastic stagnation, is of the changes of the assembly directions nd, the assembly tools nt
heavily impacted by the principle and the parameters of the and the assembly connecters nc in the whole assembly process.
algorithm. The research [20] indicates that the traditional PSO The number of parts violating the geometrical constraints in an
algorithm will smoothly slip into the local near-optimal solutions assembly sequence is recorded with M and it can be resulted from
when the optimization problem is relatively complex and it the assembly sequence.
cannot jump over the obstruction. Therefore, the traditional PSO Assembly cost is mainly subject to the geometrical constraints
algorithm will be improved and modified to meet the applied and the other five assembly process constraints. The optimization
demands of the complex optimization problems. model of the assembly sequences actually is the assembly cost
 ARTICLE IN PRESS
214 Y. Wang, J.H. Liu / Robotics and Computer-Integrated Manufacturing 26 (2010) 212–222

where 7Iijk (kA[x, y, z]) is equal to 1 if part pi is blocked by part pj

Constraints Representation Optimization Information
during the assembling process along the k-axis direction (i, jA
The number of parts [1, n]). Otherwise, Iijk is equal to 0. Iiik is always equal to 0. If the
Interference matrix IM violating geometrical part pi can be assembled freely with the part pj in the direction of
constraints M the +k-axis, the part pj can also be assembled with the part pi in
the opposite direction along the k-axis.
The change of An assembly illustrated in Fig. 2 presented by De Fazio and
assembly directions nd
Whitney [3] is applied for explaining the geometrical constraints
The number of parts between the parts in the assembly process. For example, suppose
Local assembly violating local assembly that the part 1 has been assembled. In the next step, one of the
precedence PM precedence np parts 2, 3, 4, 7, 10, 11 can be assembled with part 1. If the part 4 is
The number of unstable assembled first along the X direction, then the parts 2 and 3 will
Assembly stability SM parts ns be blocked in the X direction in the following stage. It is marked as
I24x ¼ 1 and I34x ¼ 1. On the contrary, I42x ¼ 0 and I43x ¼ 0.
The changeof The parts to be assembled can be derived from the assembly
Assembly tools TM
assembly tools nt interference matrix, and the assembly directions of each part can
The change of be also acquired at the same time. Given that the temporary
Assembly connections
connections nc subassembly is represented as Sub ¼ {p1, p2, y, pm} (m is the
CM
number of parts already assembled), pi is the part to be assembled
Optimization model of at the current stage. The assembly directions of part pi can be
PSO algorithm assembly sequences defined as Di ¼ {+Iix, +Iiy, +Iiz, Iix, Iiy, Iiz}. The values of 7Iik
(kA[x, y, z]) can be computed with formula
The optimal sequences þIik ¼ Iip1 k _ Iip2 k _    _ Iipm k (3)

Fig. 1. The assembly process constraints related to assembly sequence optimiza- where ‘‘3’’ is the Boolean operator ‘‘OR’’. The result will be equal
tion. to 0 if all the elements Iipj k (1rjrm) are 0, which means that the
part pi can be assembled freely in the corresponding direction. If
function, which can be defined as the formulae any element involved in the operation is equal to 1, the result will
be equal to 1 and the part pi cannot be assembled in that direction.
f 1 ðsÞ ¼ cp np ðsÞ þ cs ns ðsÞ þ ct nt ðsÞ þ cd nd ðsÞ þ cc nc ðsÞ (1)
The elements belonging to Di, whose values are 0, compose the
feasible assembly directions sets FDi of the part pi. If FDi ¼ F, then
f 2 ðsÞ ¼ cg M (2)
the part pi cannot be assembled at the current state. For
cg, cp, cs, ct, cd and cc are the weights of the six assembly process one assembly sequence, if there exists one part that has no
constraints, respectively. For one assembly comprising n parts, feasible assembly directions, then the sequence is globally
s ¼ (p1, p2, y, pn) represents the assembly sequence and pi(1r- infeasible. The feasible assembly directions of each part will be
irn) are the parts of the assembly sequence. If s satisfies the verified and the number of parts violating the geometrical
geometrical constraints, f1(s) is utilized to calculate the assembly constraints M can be counted.
cost. Otherwise, f1(s)+f2(s) is for calculating the total assembly
cost. 2.4. Detection of parts violating local assembly precedence

2.3. Detection of parts violating geometrical constraints For the complex assemblies, there will be too many possible
assembly sequences, but most of them are actually not ensured to
There are several methods to dispose the assembly inter- be correct. Even if the assembly sequences satisfy the geometrical
ference. For example, the moving wedge matrix [9] and the constraints, some parts violating the local assembly precedence
disassembly matrix [11] are advocated to represent the inter- still exist. The local assembly precedence is strict to rearrange
ference between two parts/components in an assembly or a the assembly sequences based upon the local assembly
disassembly process. The detection of the global assembly
interference can also be realized with the sweeping
volume method introduced by Lin et al. [22] or the assembly
precedence relations (APRs) introduced by Su [23]. Here, the
interference matrix is suggested to represent the geometrical
constraints of an assembly, from which the assembly precedence
of parts can be derived.
The assembly interference matrix has the same formation as
the disassembly matrix. However, the elements of the matrix have
different meanings. Suppose that an assembly comprising n parts
is represented as AS ¼ {p1, p2, y, pn}, the assembly interference
matrix is given as follows:
p1 p2 ... pn
2 3
P1 I11x I11y I11z I12x I12y I12z I1nx I1ny I1nz
...
6I I I
IM ¼
P2 6 21x 21y 21z
6
I22x I22y I22z
... I2nx I2ny I2nz 7
7
.. 6 .. .. 7
6 ... ... 7
. 4 . . 5
Pn In1x In1y In1z In1x In1y In1z ... Innx Inny Innz
Fig. 2. An illustrated assembly [3].
 ARTICLE IN PRESS
Y. Wang, J.H. Liu / Robotics and Computer-Integrated Manufacturing 26 (2010) 212–222 215

precedence constraints. The accessibility of the assembly tools is sij ¼ 1, when the parts pi and pj are joined by the weak
another considerable constraint to restrict the assembly order of connections and there is no support between them.
some parts. Given one assembly comprising n parts, the local sij ¼ 0, when the part pi and pj are joined by surface mating
assembly precedence between any two parts of the assembly can without force and there is no support between them.
be represented as a precedence matrix PM ¼ [rij]n  n. The values of
elements rij can be defined as follows: Given that the temporary stable subassembly is represented as
Sub ¼ {p1, p2, y, pm}(m is the number of parts already assembled),
rij ¼ +1, when the part pi must be assembled prior to the the part pi is to be assembled in the next step. The assembly
part pj. stability of the part pi with Sub can be concluded depending on the
rij ¼ 1, when the part pi must be assembled later than the stability matrix, SM. The correlative stability value sij (1rjrm)
part pj. can be inferred as {si1, si2, y, sim}. If some value sij ¼ 2 (1rjrm),
rij ¼ 0, when there is no local assembly precedence between the part pi and Sub will be stable after they are assembled.
the part pi and pj. Otherwise, the part pi needs auxiliary devices in the assembly
process. The assembly stability of each part will be checked and
Referring to Fig. 2 again, suppose that the parts 1 and 7 have the number of parts ns that make the subassembly self-stable in
been assembled, and the part 9 or the part 10 can be assembled in the assembly sequence can be computed.
the next step. It is obvious that there is no geometrical constraint
between the part 9 and 10. However, the part 9 should be 2.6. Changes of assembly directions, tools and connections
assembled prior to the part 10 due to the local precedence
constraint (in general, the inner parts of a product is assembled
The changes of the assembly directions, the assembly tools and
before the outer parts, the upper parts is assembled later than the
the assembly connection types in the assembly process have a
bottom parts), so r9 10 ¼ 1.
great effect on the assembly efficiency and cost. To improve the
The parts violating the local assembly precedence in the
assembly efficiency and reduce the assembly cost, the parts
assembly sequences can be derived from the precedence matrix,
assembled in the same assembly direction, with the same type of
PM. Given that there are m pairs of precedence constraints
assembly connections and assembly tools are suggested to be
(one pair precedence is produced by two parts) in an assembly,
assembled in groups. For example, in Fig. 2, after the part 1 is
the number of parts violating the local precedence constraints in
P assembled, the parts 3, 2, 4, 5, 6 should be assembled in
an assembly sequence can be represented as np ¼ km¼ 1prk,
succession. The assembly directions of parts can be derived from
where prk ¼ 0, if the kth (1rjrm) pair of parts satisfy the local
the assembly interference matrix. Once the assembly sequence is
assembly precedence; otherwise, prk ¼ 1.
confirmed, the number of change (nd) of the assembly directions
of all parts can be counted.
2.5. Detection of unstable parts The assembly tools used to assemble one part may be multiple.
The selected tools to assemble the current part should be the
The subassemblies should keep stable in the assembly process. same as those used to assemble the former part so that the time
Auxiliary fixtures or jigs, i.e. the additional assembly equipments for changing the tools can be shortened. All the assembly tools can
invest, are needed once some parts are unsteady. The optimal or be represented by a tool matrix TM ¼ [tij]n  m. Where, n is the
near-optimal assembly sequences should always maintain the number of parts and m is the number of practicable tools to
stability of parts or subassemblies in the assembly process so that assemble the corresponding part. After the optimal or near-
the auxiliary devices can be reduced. The stability of parts or optimal assembly sequences have been generated, the corre-
subassemblies in the assembly process is ensured by the stable sponding tools are also confirmed at the same time and the
supports or the assembly connectors related to them. Given two number of change (nt) of the assembly tools can be obtained.
parts pi and pj, and assume that pj is on the top of pi. When It is assumed that the connection type for assembling each part
the gravity direction of the part pj passes through the contact is unique and the connection types of parts are determined in the
surface of the two parts, it can be concluded that the part pi is a design process. The assembly connections to assemble the parts
stable support of the part pj. In Fig. 2, the part 1 is a stable support can be represented as a vector of connection CM ¼ [cij]n  1. Where,
of the parts 2, 7, 10. If there is no stable support for the parts to be n is the number of parts. When the assembly sequences have been
assembled, the stability of the parts guaranteed by the assembly ascertained, the number of change (nc) of the assembly connec-
connectors will be considered. The stability strengthened by tion types can also be acquired.
the assembly connectors can be categorized into three levels
according to connection strength introduced by Lee [24]. The tight
fit, interference fit, screw, rivet and welding are the strong 3. Generating optimal assembly sequences
assembly connections. For example, the connections jointing the
parts 5 and 6 belong to this type in Fig. 2. The clearance fit, 3.1. Flowchart of assembly sequence optimization
sticking and attachment are the weak connections, e.g. the fit
between the parts 2 and 4 in Fig. 2. The surface mating and other The flowchart of the assembly sequence optimization is
fits without force are the unstable connections. The connectors illustrated in Fig. 3. After the assembly interference matrix IM,
are considered as the fasteners to assemble the other parts the local assembly precedence matrix PM, the assembly stability
together. They are viewed as unconstrained parts and excluded matrix SM, the assembly tools matrix TM and the assembly
from the assembly sequences. For an assembly comprising n parts, connections matrix CM are given, the assembly sequence
the assembly stability between two parts proved by the stable optimization using the CPSO algorithm can be executed until
support or the assembly connectors can be represented as the stability the optimal or near-optimal assembly sequences are generated.
matrix SM ¼ [sij]n  n, and the values of sij can be defined as follows: In the computation process, the chaos method is utilized to
produce the preferable assembly sequences of each particle to
sij ¼ 2, when the part pi is supported by pj or they are joined by accelerate the convergence rate of the CPSO algorithm. The
strong connections and the stable support is unilateral. assembly cost of each assembly sequence is evaluated by
 ARTICLE IN PRESS
216 Y. Wang, J.H. Liu / Robotics and Computer-Integrated Manufacturing 26 (2010) 212–222

3.3. PSO algorithm and chaos method

Input:
Interference matrix IM Generation of exchange
The traditional PSO algorithm comes from the simulation of
Local precedence matrix PM location u,w and
the social behavior of the birds seeking for food. The best strategy
Stability matrix SM preferable sequences
using chaos method for one bird to reach the destination is to pursue the bird that is
Tools matrix TM
Connection matrix CM nearest to the food. The solutions of the optimization problems
are viewed as the birds which are called particles. The optimiza-
tion process of the problems is regarded as the progressive paths
Update of vov and
sequences of the particles seeking for the food. Each particle has a fitness
Population size of particles N computed by the objective function. The flight direction and
Saving probabilities c0,c1,c2 Record of assembly
process information distance of each particle are subject to its velocity. All the particles
Cycles of computation NCmax
are pursuing the best particle owning the optimum fitness until
Coefficients of constraints:
the optimal or near-optimal solutions are obtained. The particles
cg, cp, cs, ct, cd and cc
are commenced with stochastic initializations and the optimal or
The base part BP Calculation of current near-optimal solutions are generated by iterative computations.
pbesti and gbesti Each particle renews its own solution based upon two best
solutions in each computation time step, i.e. the current
Generation of sequences individual best solution searched by itself and the global best
and process information solution searched by the particle colonies, which are represented
Calculation of initial pbest0 No Cycles as pbest and gbest, respectively. To improve the computation
and gbest0 i NCmax? efficiency, a portion of the whole particles are selected for
comparison and the global best solution can be replaced by the
Yes local best solution if the population size of the particles is too
large. For multivariable optimization problems, the best solutions
No pbest and gbest will be substituted with the vectors pbest and
Optimal Optimal
sequence? gbest, respectively.
assembly
The next velocity vector and solution vector of each particle are
sequences
updated according to formulae (4) and (5) [13] after the two
Yes current best solution vectors are calculated

Fig. 3. The flowchart of assembly sequence optimization.

vkþ1 ¼ c0 vk þ c1 ðpbest k  xk Þ þ c2 ðgbest k  xk Þ (4)

xkþ1 ¼ xk þ vkþ1 (5)

formulae (1) and (2), and the current individual and global near- where vk is the current velocity vector of some particle, xk the
optimal assembly sequences are selected. The fine fragments of current solution vector of the same particle, pbestk the current
the assembly sequences are maintained by each particle and the optimal solution vector of the particle, gbestk the current global
orders of the parts are altered referring to the current near- optimal solution vector and c0, c1, c2 the coefficients. In general,
optimal assembly sequences until all the particles converge at the c0A[0, 1] and c1, A[0, 2]. Each fraction of formula (4) has its own
optimal or near-optimal assembly sequences. particular significance. c0vk embodies the memorial behavior of
the particles and c1(pbestkxk) implies the cognitive behavior of
the particles, which illuminates that the behaviors of the particles
3.2. Initialization of assembly sequence optimization originate from their own experience. c2 (gbestkxk) indicates that
the behaviors of the particles are also intensively influenced by
The population size of the particles is mainly concluded by the the outstanding experience of the best particle in the colony.
number of parts. The more complex the assembly is, the larger The key issue to figure out the optimal or near-optimal
the number of particles is needed and the more cycles of iterative solutions of the traditional PSO algorithm is that the particles
computation will be required. For one assembly with n parts, the are able to find the preferable solutions progressively in the
assembly sequences are represented as the solution vectors optimization process. In reverse, if the best particle cannot detect
of particles and the elements of the vectors are symbolized as the better solutions, the PSO algorithm will stagnate at the local
the parts. The assembly sequence of the ith particle can near-optimal solutions for a long time or the optimization process
be represented as si ¼ (pi1, pi2, y, pin), which indicates that the becomes unsuccessful. In the traditional PSO algorithm, all the
assembly order of the parts is assembled from pi1 to pin. Where, other particles pursue the best particle with the minimum or
pi1, pi2, y, pin can be the codes of the parts. The first part to maximum fitness and the algorithm converges toward the optimal
be assembled is the base part. The part having the maximum or near-optimal solutions quickly at the early optimization stage.
assembly relations with the other parts is selected as the base part However, the stochastic principle to confirm the optimal solution
in this paper. of the traditional PSO algorithm will generate a large number of
The fitness of each particle or the cost of each assembly identical solutions in the subsequent optimization stage, which
sequence has to be evaluated in the optimization process. The prevents the PSO algorithm from skipping over the local near-
coefficients of the geometrical constraints cg and the other optimal solutions and the convergence rate becomes quite small
assembly process constraints cp, cs, ct, cd, cc should be given by or even zero, and a considerable quantity of computation resource
the designers. The correlative parameters of the PSO algorithm, is consumed. The reasonable methods ought to be given and
such as c0, c1, c2, are also appointed [13]. The maximum cycles combined with the traditional PSO algorithm to avoid the
NCmax of iterative computation of the CPSO program should also limitations of it.
be determined and the optimal or near-optimal assembly Chaos method is suggested to improve the performance of the
sequences can be generated in the controlled computation time traditional PSO algorithm, which generates the optimal or near-
step. optimal assembly sequences quickly. The two main characteristics
 ARTICLE IN PRESS
Y. Wang, J.H. Liu / Robotics and Computer-Integrated Manufacturing 26 (2010) 212–222 217

of the chaos method concentrated on will strengthen the For example, the initial assembly sequence is s ¼ (1, 3, 4, 2, 5)
performance of the traditional PSO algorithm. (1) Randomicity: and the vo ¼ vo(4, 2) is applied to s. Then, s0 ¼ s+vo ¼ (1, 3, 4, 2,
the Randomicity of the chaos method is concluded by the 5)+vo(4, 2) ¼ (1, 2, 4, 3, 5), where + is endowed with special
formulae of chaotic attractors and they are different from those signification which indicates that the vo acts on the assembly
of other methods (such as GA method, SA method) which sequence.
are appointed by the designers. It is prone to be applied once Velocity operator vector (vov): a velocity operator vector is
the chaotic attractor is chosen. The chaotic attractors generate the composed of two or more vos with formatted order. Given
stochastic variables conforming to the formulae in the domain. that there are n vos(1rirn), the vov can be represented as
The concrete values of the stochastic variables are unpredictable vov ¼ (vo1, vo2, y, von).
while the range of the variables is known. The Randomicity of the The vos in a vov (acting on the assembly sequence) do not meet
chaos method ensures the diversity of the particles, which the commutative law. For example, (vo1, vo2)a(vo2, vo1). The
prohibits the traditional PSO algorithm from being trapped into updated result of the assembly sequence due to the execution of a
the local optimal solutions. (2) Ergodicity: ergodicity is the other vov is equivalent to that due to the cumulate execution of the
significant characteristic of the chaotic attractors and each of the corresponding vos in the vector.
stochastic variables generated by them is unique in the whole Combination of vovs: the new vovs can be computed by
computation process, i.e. the ergodicity of chaos method ensures combining several vovs. For example, two vovs, vov1 and vov2
that the next solution generated is different from the previous act on the assembly sequence s one by one and a new assembly
solutions. When it is combined with the traditional PSO sequence s0 is generated. The sequence s0 can also be generated by
algorithm, the diversity of the particles is highly enhanced and a vov0 acting on the sequence s. Then, vov0 is the combination of
the performance of the global optimization of the traditional PSO vov1 and vov2. It can be formulated as
algorithm will be improved. The CPSO algorithm can work well in
vov0 ¼ vov1  vov2 (6)
seeking for the preferable solutions in the evolutionary process
and the premature convergence of the CPSO algorithm is 0
where the operator ‘‘’’ merges vov1 and vov2 into the vov , and
weakened. These characteristics of the CPSO algorithm are all the order of vos belonging to vov1 and vov2 is serialized. For
verified by the actual example in Section 4. example, (1, 2)(3, 4) ¼ ((1, 2), (3, 4)).
vovs can also be combined with the given probabilities which
are called the saving probabilities. That is to say, not all vos in
3.4. The mapping rules and definition of velocity operators (vos) the addend vectors are merged into the sum of the vovs. For
example, two vovs, vov1 and vov2 can be combined with saving
The optimization model of assembly sequences can be mapped the probabilities p1 and p2 (0rp1, p2r1), respectively. The
into the PSO model according to the five rules. combination formula is remodeled as

vov0 ¼ p1 vov  p2 vov2 (7)

(1) The assembly sequences are considered as the solution
vectors of the PSO algorithm. The number of parts is viewed where p1vov1 and p2vov2 indicate that the vos in the vov1 and vov2
as the dimension of the solution vectors. are remained with the probability p1 and p2, respectively. When
(2) The velocity of the particles is replaced by the velocity the probability p1 and p2 are changed, or the vos preserved from
operator, whose function is to change the assembly orders of the vovs are altered, the results of the combination are also
parts in assembly sequences for generating the optimal or different.
near-optimal sequences.
(3) The assembly cost is taken as the fitness of particles and
3.5. Generation of preferable sequences using the chaos method
formulae (1) and (2) are the objective functions.
(4) The population size of particles is modulated according to the
The exchange locations of parts are assigned randomly with
number of parts. The optimization results will be improved
the traditional PSO algorithm and a large number of identical
with the number increase of particles, but the computation
assembly sequences are generated in the iterative computation
time will also be augmented.
process. Therefore, the preferable assembly sequences of each
(5) The coefficients c0, c1 and c2 are considered as the saving
particle are hard to be sought in the next computation time step.
probabilities which maintain the reasonable velocity opera-
Most importantly, once the case of si ¼ pbesti ¼ gbestk emerges,
tors to generate the appropriate assembly sequences in the
the actualization to find the preferable assembly sequences
next time step.
mainly relies on the vos c0vi. All the particles will halt around
their current best solutions and the global preferable solutions
The function of the velocity in the PSO algorithm is used to alter cannot be detected in the next computation time step, and the
the positions of particles, while the vo for assembly sequence optimal or near-optimal assembly sequences will be lost in such
optimization is to change the assembly orders of parts till the situations.
optimal or near-optimal assembly sequences are sought. The The inacceptable situation will be weakened or ceased after the
crossover operator and mutation operator introduced in GA chaos method is accompanied. The special characteristic of
algorithm can be referred to define the vos. The operation of ergodicity of the chaos method can ensure that each of all the
interchanging two elements can avoid the operation complexity of solutions is unique in the whole optimization process and the
the crossover operator and the mutation operator, which is prone cases producing two or more identical solutions are excluded. In
to be applied and is considered as one of the efficient means to this paper, the chaos method is utilized to generate the preferable
solve the all difference problems [25]. The vo is defined as follows. assembly sequences pbesti,k of each particle in the current
Velocity operator (vo): given that the assembly contains n parts computation time step. Then, all the preferable assembly
and the assembly sequence is represented as s ¼ (p1, p2, y, pi, y, sequences participate in the optimization process of the tradi-
pj, y, pn). The function of vo ¼ vo(i, j) is to change the orders of the tional PSO algorithm. The process of discovering preferable
part pi and pj so that the new assembly sequence s0 ¼ (p1, assembly sequences using the chaos method is described in detail
p2, y, pj, y, pi, y, pn) is generated. as follows.
 ARTICLE IN PRESS
218 Y. Wang, J.H. Liu / Robotics and Computer-Integrated Manufacturing 26 (2010) 212–222

Step 1: selection of the chaotic attractor. The chaotic attractor In the next computation time step, the new assembly
of the logistic [21] is chosen as the actuator to generate the sequences are generated when the vectors vovs act on the current
chaotic variables. The attractor is represented as the formula (8). assembly sequences.
Where, m ¼ 4 and xke{0, 0.25, 0.5, 0.75, 1}. In the search space of
si;kþ1 ¼ si;k þ vsi;kþ1 (11)
optimization problem, once the next variable is changed a little in
contrast with the former variable, the following variables will be Before the optimal or near-optimal assembly sequences are
changed progressively and remarkably till they permeate the concluded, the assembly cost of each assembly sequence is
whole search space of solutions. calculated with formulae (1) and (2). The current optimal
assembly sequence of each particle and the current global optimal
xkþ1 ¼ m  xk  ð1  xk Þ; xk 2 ð0; 1Þ (8)
assembly sequence are obtained. The formulae (10) and (11) will
Step 2: conclusion of the exchange locations ui,k+1, wi,k+1 of two be used with many cycles till the satisfactory assembly sequences
parts in the assembly sequence. Two initial distinct number ti,k1, are found or the appointed cycles of iterative computation NCmax
ti,k2A(0, 1) (ti,k1ati,k2) are chosen, and the chaotic variables ti,k+1
1
, are over.
2
ti,k+1 in the successive cycle are computed with the formula (8).
The chaotic exchange locations ui,k+1, wi,k+1 of the two parts can be
computed by means of the formula (9). 4. Case studies and discussions

ui;kþ1 ¼ ½t 1i;kþ1  PartNum; wi;kþ1 ¼ ½t 2i;kþ1  PartNum (9)

4.1. Case study
where PartNum is the parts number of assembly and ‘‘[x]’’ is to
convert the number x into the integer. The CPSO algorithm for assembly sequence planning and
Step 3: the parts at the locations ui,k+1, wi,k+1 in the assembly optimization has been programmed with C++ language on a PC
sequence are exchanged referring to the appointed probability q with 1.8 GHz AMD3000+ processor and 512 M memory. By
and the renewed assembly sequence si,k+1 is generated. comparing the CPSO algorithm with the traditional PSO algorithm,
Step 4: the fitness of the assembly sequence si,k+1 is evaluated the validity of the method is confirmed with several assembly
by the formulae (1) and (2). If f(si,k+1)rf(si,k), the computation cases.
process will turn to the step 5, or else it turns to the step 6. The exploded solid model of a generator assembly is shown in
Step 5: the preferable assembly sequence s*i and the chaotic Fig. 4. The generator comprises 15 parts which are marked with
variables ti,k1, ti,k2 are updated as follows: numbers orderly. The assembly directions of all the parts are
along the X and Y axes. Part 7 is viewed as the base part which has
si ¼ si;kþ1 ; t 1i;k ¼ t 1i;kþ1 ; t 2i;k ¼ t 2i;kþ1 the most assembly relations with the other parts.
Step 6: the preferable assembly sequence s*i is replaced by Suppose that the assembly tool and assembly connector to
sk and the two chaotic variables ti,k1, ti,k2 are also substituted assemble the base part 7 are T1 and C1, respectively. The assembly
1
by ti,k+1 2
ti,k+1, respectively. interference matrix, local preference matrix, stability matrix,
Step 7: if all the preferable assembly sequences of each particle connection matrix and tool matrix are given based on the
at the current cycle of the traditional PSO algorithm are found, assembly structure and assembly process. In this case, the proper
then the process is finished and the process of the traditional PSO tools and the connections for assembling each part are illustrated
algorithm is started. Otherwise, the process turns to the step 2. in Table 1 and the other assembly process constraints are omitted.
After the chaos algorithm is completed in the procedure, each All the parameters of the assembly cost function and the CPSO
particle seeks for the preferable assembly sequence pbesti,k and algorithm are assigned as: cg ¼ 50, cp ¼ 0.15, cs ¼ 0.15, ct ¼ 0.15,
the convergence rate of the traditional PSO algorithm will be cd ¼ 0.4, cc ¼ 0.15, c0 ¼ 0.8, c1 ¼ 0.5, c2 ¼ 0.5. The initial assembly
accelerated. sequences and exchange locations of parts are generated with
random sample method.
The experiments of the assembly sequence optimization are
3.6. Update of the assembly sequences done four times using the CPSO algorithm by adjusting the
number of the particles from 20 to 50. One more 10 particles are
Considering the traditional PSO algorithm, the combination of added in the next experiment (if the number of the population
the vovs can be represented as

vsi;kþ1 ¼ c0 vsi;k  c1 ðPbest i;k  si;k Þ  c2 ðGbest k  si;k Þ (10)

where vsi,k is the vov of the ith particle in the kth time step, vsi,k+1
the vector in the next time step, si,k the assembly sequence
represented by the ith particle in the kth time step, Pbesti,k the
individual optimal assembly sequence of the ith particle, Gbestk
the current global optimal assembly sequence of the particle
swarms and c0, c1, c2 the saving probabilities of vsi,k, Pbesti,ksi,k
and Gbestksi,k, respectively.
The PSO algorithm can be used to generate the optimal
assembly sequence referring to the formula (6) and (10). c0vsi,k
indicates that the first improvement of the assembly sequences
originates from the former excellent vos. c1(Pbesti,ksi,k) implies
that the second improvement is derived from the current
preferable vos. The current global optimal vos c2(Gbestksi,k)
leads to the third improvement of the assembly sequences. The
bigger the value of ck (k ¼ 0, 1, 2) is, the more the function of the
corresponding vov is reinforced to optimize the assembly
sequences. Fig. 4. The assembly of a generator.
 ARTICLE IN PRESS
Y. Wang, J.H. Liu / Robotics and Computer-Integrated Manufacturing 26 (2010) 212–222 219

size of the particles is below 10, the program will be failed to find the generated optimal or near-optimal assembly sequences, the
the optimal or near-optimal assembly sequences). The initial assembly process constraints, the assembly cost f(s), the
assembly sequences are all distinct. The results of the experiments convergent cycles of iterative computation and the number of
are presented in Table 2. The first row of each experiment shows particles. The three following rows are the assembly tools, the
connection types and the assembly directions to assemble the
corresponding parts and the change numbers. On the other hand,
Table 1 the assembly sequence optimization using the traditional PSO
Assembly tools and connections for the parts.
algorithm is also executed four times for the comparisons. The
Particles Number Assembly tools Connections population size of the particles is adjusted from 600 to 900 (if the
population size is less than 400, the experiments will be failed or
1 T1, T2, T4 C2 last a long time). The optimization results of the traditional PSO
2 T1, T4, T5 C1 algorithm are shown in Table 3. The formation of the two tables is
3 T2, T3, T4 C3
4 T1, T2, T5 C1
identical.
5 T1, T2, T4 C1 Although all of the initial assembly sequences applied to the
6 T1, T2, T3 C1 experiments using the CPSO algorithm and the traditional PSO
7 T1 C1 algorithm are infeasible and different, the optimal or near-optimal
8 T3, T4, T5 C3
assembly sequences are found ultimately and the minimum
9 T1, T2, T4 C1
10 T1, T2, T5 C1 assembly costs are almost equal. It is revealed that the assembly
11 T2, T3, T4 C3 sequence planning and optimization can be accomplished with
12 T1, T4, T5 C1 the CPSO algorithm and the traditional PSO algorithm. With the
13 T1, T2, T4 C2 number increase of the population size of the particles, the
14 T3, T4, T6 C3
15 T4, T5, T6 C3
minimum assembly cost is decreased gradually. It is proved that
the performance of the global optimization of the CPSO algorithm
and the traditional PSO algorithm can be improved. In reverse, the

Table 2
The solutions of assembly sequence optimization using CPSO algorithm.

Experiment Assembly sequence (s) M np ns nt nc nd f(s) Convergent cycles Particles number

1 Sequence 7 6 5 4 9 10 11 12 3 2 1 13 14 15 8 0 0 0 3.10 118 20

Tools T1 T1 T1 T1 T1 T1 T4 T4 T4 T4 T4 T4 T4 T4 T4 1
Connections C1 C1 C1 C1 C1 C1 C3 C1 C3 C2 C2 C3 C3 C3 C3 5
Directions +X +X +X +X X X X X +X +X +X X X X +Y 4

2 Sequence 7 9 10 11 12 6 5 4 3 2 1 13 15 14 8 0 0 0 3.00 39 30
Tools T1 T1 T1 T4 T4 T2 T2 T2 T2 T4 T4 T4 T4 T4 T4 3
Connections C1 C1 C1 C3 C1 C1 C1 C1 C3 C1 C2 C2 C3 C3 C3 6
Directions X X X X X +X +X +X +X +X +X X X X +Y 4

3 Sequence 7 9 10 11 6 5 4 3 2 1 12 13 15 14 8 0 0 0 3.15 126 40

Tools T1 T1 T1 T2 T2 T2 T2 T2 T4 T4 T4 T4 T4 T4 T4 2
Connections C1 C1 C1 C3 C1 C1 C1 C3 C1 C2 C1 C2 C3 C3 C3 8
Directions X X X X +X +X +X +X +X +X X X X X +Y 3

4 Sequence 7 6 9 10 5 4 3 2 1 11 12 13 15 14 8 0 0 0 2.85 120 50

Tools T1 T1 T1 T1 T1 T1 T4 T4 T4 T4 T4 T4 T4 T4 T4 1
Connections C1 C1 C1 C1 C1 C1 C3 C1 C2 C3 C1 C2 C3 C3 C3 7
Directions X X X X +X +X +X +X +X X X X X X +Y 3

Table 3
The solutions of assembly sequence optimization using PSO algorithm.

Experiment Assembly sequence (s) M np ns nt nc nd f(s) Convergent cycles Particles number

1 Sequence 7 6 9 10 11 12 5 4 3 13 2 1 8 14 15 0 0 0 3.50 726 600

Tools T1 T1 T1 T1 T4 T4 T4 T2 T2 T2 T4 T4 T4 T4 T4 3
Connections C1 C1 C1 C1 C3 C1 C1 C1 C3 C2 C1 C2 C3 C3 C3 7
Directions X X X X X X +X +X +X X +X +X +Y Y –Y 5

2 Sequence 7 9 10 11 12 6 5 4 3 2 13 1 8 14 15 0 0 0 3.35 246 700

Tools T1 T1 T1 T4 T4 T2 T2 T2 T2 T4 T4 T4 T4 T4 T4 3
Connections C1 C1 C1 C3 C1 C1 C1 C1 C3 C1 C2 C2 C3 C3 C3 6
Directions X X X X X +X +X +X +X +X X +X +Y Y –Y 5

3 Sequence 7 9 10 6 5 4 11 12 13 3 2 1 8 14 15 0 0 0 3.20 568 800

Tools T1 T1 T1 T1 T1 T1 T4 T4 T4 T4 T4 T4 T4 T4 T4 1
Connections C1 C1 C1 C1 C1 C1 C3 C1 C2 C3 C1 C2 C3 C3 C3 7
Directions X X X +X +X +X X X X +X +X +X +Y Y –Y 5

4 Sequence 7 5 4 3 6 9 10 11 12 13 2 1 8 14 15 0 0 0 2.80 596 900

Tools T1 T1 T1 T2 T2 T2 T2 T2 T4 T4 T4 T4 T4 T4 T4 2
Connections C1 C1 C1 C3 C1 C1 C1 C3 C1 C1 C2 C2 C3 C3 C3 6
Directions +X +X +X +X X X X X X X +X +X +Y Y –Y 4
 ARTICLE IN PRESS
220 Y. Wang, J.H. Liu / Robotics and Computer-Integrated Manufacturing 26 (2010) 212–222

Fig. 5. The change of assembly costs in the evolutionary process using CPSO algorithm.

Fig. 6. The change of assembly costs in the evolutionary process using PSO algorithm.

computation time is ascending progressively with the number Table 4

increase of the particles. The comparisons of the computation results of the three algorithms.
In the process of the assembly sequence optimization, the fine
Algorithm The optimal The convergent Computation
subsequence fragments in the current assembly sequences are assembly cost cycles time (s)
remained constantly for the optimization in the next time step.
The parts in each assembly sequence are suitably rearranged and CPSO 2.85 120 10.65
the preferable assembly sequences are generated with the chaos PSO 2.80 596 149.35
SA 2.95 132 19.0
method. Furthermore, all the particles rearrange their parts
referring to the best particle and the assembly costs are decreased
gradually until they all converge at the optimal or near-optimal
assembly sequences with the minimum assembly cost. The PSO algorithm, respectively. The computation time is also
change of the assembly costs in the evolutionary processes using recorded for comparisons, for example, 10.656 s is consumed
the CPSO algorithm and the traditional PSO algorithm are by Exp. 4 in Table 2, whereas 149.359 s is spent by the Exp. 4 in
displayed in Figs. 5 and 6, respectively. It is discriminated that Table 3. Obviously, the number of the particles and cycles of
the initial assembly costs in Fig. 5 are higher than those in Fig. 6. It computation have a great effect on the efficiency of the two
is because more particles are used by the traditional PSO algorithms.
algorithm than the CPSO algorithm. The initial assembly To display the preferable performance of the CPSO algorithm in
sequences are totally infeasible, so the assembly costs of the the applications, the assembly sequence optimization using the
particles in the earlier computation time steps are higher. It is traditional SA algorithm is also implemented for the comparison
further observed that the convergence rate of the CPSO algorithm and analysis under the same constraints. The actual assembly
is faster than that of the traditional PSO algorithm. The assembly sequences and the other concerned assembly process information
sequences with one part violating the geometrical constraint are are excluded except for the optimal assembly cost, the convergent
sought before the 40th cycle in the computation process using the cycles and the computation time. The results of the two 4th
CPSO algorithm, whereas three experiments consume nearly 100 experiments of the CPSO algorithm and the traditional PSO
cycles of computation using the traditional PSO algorithm under algorithm are picked out for the comparison with those of using
the same constraints. With the progress of the optimization the traditional SA algorithm. The results for the comparisons of
process, the assembly sequences without geometrical constraints the three algorithms are listed in Table 4. The changes of the
are achieved before the 60th cycle of computation using the CPSO assembly costs of the experiments using the three algorithms are
algorithm, whereas most experiments consumes more than 200 displayed in Fig. 7. The assembly cost of the earlier 550
cycles of computation using the traditional PSO algorithm. Finally, computation time steps of the traditional PSO algorithm is lined
the optimal or near-optimal assembly sequences with the in Fig. 7.
minimum assembly cost are obtained, the convergent cycles of With respect to the datum in Table 4 and the changes of
the computation and the population size of the particles using the assembly costs in Fig. 7, it can be concluded that the performance
CPSO algorithm are nearly 15 and 20 1
of those using the traditional of the CPSO algorithm is better than that of the traditional SA
 ARTICLE IN PRESS
Y. Wang, J.H. Liu / Robotics and Computer-Integrated Manufacturing 26 (2010) 212–222 221

Fig. 7. The change of assembly costs in the evolutionary process using CPSO algorithm, PSO algorithm and SA algorithm.

algorithm, and the performance of the traditional PSO algorithm is process and the preferable solutions are generated progressively
the worst in the application of assembly sequence planning under until the optimal or near-optimal solutions are obtained.
the same constraints. The CPSO algorithm does need fewer particles to participate in
For the traditional SA algorithm, the computation time is the optimization process in contrast with the traditional PSO
mainly determined by the initial temperature, the computation algorithm. Generally, the traditional PSO algorithm produces
cycles of the inner iteration, the coefficient of updating tempera- hundreds of assembly sequences in each cycle of computation
ture and the terminal temperature. The four parameters are time step. A lot of assembly sequences will be generated and the
assigned as 200 1C, 1000, 0.95 and 0.001 1C in the program, optimal or near-optimal assembly sequences are chosen from
respectively. The cycles of the inner iterative computation are them and the mode is also adopted by the traditional SA
1000 means that there are 1000 assembly sequences generated in algorithm. Actually, the number of the particles is limited.
sequence in each stage of lowering the temperature process. The Otherwise, the initialization of the traditional PSO algorithm is
final displayed cycles of the iterative computation is not complex and the convergence rate of the traditional PSO
constrained by the designers but the program after the initial algorithm will be declined heavily. However, less than tens of
temperature, the coefficient of updating temperature and the particles are sufficient with the CPSO algorithm. Furthermore, the
terminal temperature are given. The total cycles of iterative cycles of the iterative computation are far less using the CPSO
computation are 417 in the experiment with the traditional SA algorithm than that of the traditional PSO algorithm.
algorithm. The vo operation is also used to change the orders of The above analysis and conclusions are proved by the
the parts in the assembly sequences. application of assembly sequence planning and optimization.
The CPSO algorithm concretely has revealed better performance in
solving the complex combinatorial problem of assembly sequence
4.2. Discussions planning than that of the traditional PSO algorithm.

It is observed that the quality of the initial assembly sequences

and the number of the particles have a great effect on the 5. Conclusions and future work
performance of the traditional PSO algorithm, but the CPSO
algorithm is not influenced much by such issues. If there exists The main goal of assembly sequence planning is to seek for the
one assembly sequence with low assembly cost in the initial assembly sequences with the minimum assembly cost while
assembly sequences, the traditional PSO algorithm converges at satisfying the assembly process constraints. The research of
the exact solutions quickly. On the other hand, if all the particles assembly sequence planning using the CPSO algorithm is
are initialized with inferior assembly sequences, more cycles of introduced and investigated. The six kinds of assembly process
iterative computation and more particles will be required to constraints are comprehensively considered and the optimization
generate the optimal or near-optimal assembly sequences. The model of assembly sequences is given referring to them. The
consequence for the high-quality initialization of assembly optimal or near-optimal assembly sequences conforming to the
sequences is predominated by the characteristics of the tradi- engineering demands can be obtained in view of the optimization
tional PSO algorithm. All the particles are encouraged to pursuit model of assembly sequences. The principle and application of the
the best particle in the optimization process until the optimal traditional PSO algorithm and chaos method are concisely
solutions are obtained. The preferable solutions are difficult to be introduced and merged into the CPSO algorithm which is used
identified by means of the stochastic process and many identical to generate the optimal or near-optimal assembly sequences. The
solutions generated in the optimization process prevent the PSO mapping rules between the optimization model of assembly
algorithm converging quickly. However, the ergodicity of chaos sequences and the PSO model bridge the gap of them and ensure
method ensures that each solution is unique in the optimization that the PSO algorithm can optimize the assembly sequence
 ARTICLE IN PRESS
222 Y. Wang, J.H. Liu / Robotics and Computer-Integrated Manufacturing 26 (2010) 212–222

planning problem. The vo is defined to exchange the orders of [7] Hong D, Cho H. Generation of robotic assembly sequences using a simulated
parts in the assembly sequences and drive the initial assembly annealing. In: Proceedings of the IEEE/RSJ international conference on
intelligent robotic and systems. 1999. p. 1247–52.
sequences to evolve into the optimal or near-optimal assembly [8] Guan Q, Liu JH, Zhong YF. A concurrent hierarchical evolution approach to
sequences according to the principle of the traditional PSO assembly process planning. International Journal of Production Research
algorithm. The chaos method is applied to enhance the unique- 2002;40(14):3357–74.
[9] Smith GC, Smith SSF. An enhanced genetic algorithm for automated assembly
ness of each particle in particular time step and the diversity of planning. Robotics and Computer Integrated Manufacturing 2002;18(5-6):
the particles is ensured in the optimization process, which 355–64.
improves the convergence rate of the traditional PSO algorithm [10] Failli F, Dini G. Ant colony systems in assembly planning: a new approach to
sequence detection and optimization. In: Proceedings of second CIRP
and the premature convergence of it is also weakened. The results
international seminar on intelligent computation in manufacturing engineer-
and analysis of the application validate that the CPSO algorithm ing. 2000. p. 227–32.
has better effectiveness than the traditional PSO and SA algorithm. [11] Wang JF, Liu JH. Intelligent selective disassembly using the ant colony
algorithm. Artificial intelligence for engineering design. Analysis and
However, the CPSO algorithm may be inadequate for solving the
Manufacturing 2003;17(4):325–33.
assembly sequence planning problem of large assemblies because [12] Cao P, Xiao R. Assembly planning using a novel immune approach.
the computation is explosively heavy. Some new methods and International Journal of Advanced Manufacturing Technology 2007;
technique will be probed and further researched to avail to tackle 31(7-8):770–82.
[13] Kennedy J, Eberhart R. Particle swarm optimization. In: Proceedings of the
assembly sequence planning problem. Future works will concen- IEEE international conference on neural networks. 1995. p. 1942–8.
trate on: (1) the methods of extracting subassemblies to reduce [14] Eberhart R, Shi Y. Particle swarm optimization: developments, applications
the search space of assembly sequences and (2) the automatic and resources. In: Proceedings of the IEEE International Conference On
Evolutionary Computation. 2001. p. 81–6.
acquisition of the assembly process constraints. [15] Kennedy J, Eberhart RC. Swarm intelligence. San Francisco: Morgan Kaufmann
Publishers; 2001.
[16] Clerc M. Particle swarm optimization. London: ISTE; 2006.
Acknowledgments [17] Liao C, Tseng C, Luarn P. A discrete version of particle swarm optimization for
flowshop scheduling problems. Computers & Operations Research 2007;
34(10):3099–111.
The authors gratefully acknowledge the fund support from the [18] Schutte J, Koh B, et al. Evaluation of a particle swarm algorithm for
National High-Tech Research and Development Program of China biomechanical optimization. Journal of Biomechanical Engineering 2005;
127(3):465–74.
(863 program), Grant no. 2006AA04Z138. [19] Guo YW, Li WD, Mileham AR, Owen GW. Applications of particle swarm
optimization in integrated process planning and scheduling. Robotics and
References Computer-Integrated Manufacturing 2009;25(2):280–8.
[20] Bergh F, Engelbrecht AP. A study of particle swarm optimization particle
trajectories. Information Sciences 2006;176(8):937–71.
[1] Romeo MM, Lee HSL, Kazem A. A genetic algorithm for the optimisation of [21] Yuan ZL, Yang LL, et al. Chaotic particle swarm optimization algorithm for
assembly sequences. Computers & Industrial Engineering 2006;50(4):503–27. traveling salesman problem. In: Proceedings of the IEEE international
[2] Laperriere L, Eimafaghy H. GAPP: a generative assembly process planner. conference on automation and logistics. 2007. p. 1121–4.
Journal of Manufacturing System 1996;15(4):282–93. [22] Lin A, Chang T. An integrated approach to automated assembly planning for
[3] Homem de Mello L, Lee S. Computer-aided mechanical assembly planning. three-dimensional mechanical products. International Journal of Production
London: Kluwer Academic Publishers; 1991. Research 1993;31(5):1201–27.
[4] Zha XF, Lim SY, Fok SC. Integrated intelligent design and assembly planning: a [23] Su Q. A hierarchical approach on assembly sequence planning and optimal
survey. International Journal of Advanced Manufacturing Technology 1998; sequences analyzing. Robotics and Computer-Integrated Manufacturing
14(10):1–15. 2008; online 28 January.
[5] Solnon C. Combining two pheromone structures for solving the car [24] Lee S. Subassembly identification and evaluation for assembly planning. IEEE
sequencing problem with ant colony optimization. European Journal of Transactions on System, Man and Cybernetics 1994;24(3):493–503.
Operational Research 2008;191(3):1043–55. [25] Rameshkumar K, Suresh RK, Mohanasundaram KM. Discrete particle swarm
[6] Chen W, Tai P, Deng W, Hsieh L. A three-stage integrated approach for optimization (DPSO) algorithm for permutation flowshop scheduling to
assembly sequence planning using neural networks. Expert Systems with minimize makespan. In: Proceedings of the international conference on
Applications 2008;34(3):1777–86. natural computation. 2005. p. 572–81.