optimization of process plans using a constraint-based tabu search approach

This article was downloaded by: [University of California, San Francisco]On: 19 November 2014, At: 04:33Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

International Journal of ProductionResearchPublication details, including instructions for authors andsubscription information:http://www.tandfonline.com/loi/tprs20

Optimization of process plans usinga constraint-based tabu searchapproachW. D. Li a , S. K. Ong b & A. Y. C. Nee ba Singapore Institute of Manufacturing Technology , 71Nanyang Drive, Singapore 638075b Department of Mechanical Engineering , National Universityof Singapore , 9 Engineering Drive 1, Singapore 117576Published online: 21 Feb 2007.

To cite this article: W. D. Li , S. K. Ong & A. Y. C. Nee (2004) Optimization of process plansusing a constraint-based tabu search approach, International Journal of Production Research,42:10, 1955-1985

To link to this article: http://dx.doi.org/10.1080/00207540310001652897

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the“Content”) contained in the publications on our platform. However, Taylor & Francis,our agents, and our licensors make no representations or warranties whatsoeveras to the accuracy, completeness, or suitability for any purpose of the Content. Anyopinions and views expressed in this publication are the opinions and views of theauthors, and are not the views of or endorsed by Taylor & Francis. The accuracyof the Content should not be relied upon and should be independently verifiedwith primary sources of information. Taylor and Francis shall not be liable for anylosses, actions, claims, proceedings, demands, costs, expenses, damages, and otherliabilities whatsoever or howsoever caused arising directly or indirectly in connectionwith, in relation to or arising out of the use of the Content.

This article may be used for research, teaching, and private study purposes. Anysubstantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,systematic supply, or distribution in any form to anyone is expressly forbidden.

http://www.tandfonline.com/loi/tprs20

http://dx.doi.org/10.1080/00207540310001652897

Terms & Conditions of access and use can be found at http://www.tandfonline.com/page/terms-and-conditions

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

http://www.tandfonline.com/page/terms-and-conditions

http://www.tandfonline.com/page/terms-and-conditions

int. j. prod. res., 15 may 2004, vol. 42, no. 10, 1955–1985

Optimization of process plans using a constraint-based tabu search

approach

W. D. LIy, S. K. ONGz and A. Y. C. NEEz*

A computer-aided process planning system should ideally generate and opti-mize process plans to ensure the application of good manufacturing practicesand maintain the consistency of the desired functional specifications of a partduring its production processes. Crucial processes, such as selecting machiningresources, determining set-up plans and sequencing operations of a part shouldbe considered simultaneously to achieve global optimal solutions. In this paper,these processes are integrated and modelled as a constraint-based optimizationproblem, and a tabu search-based approach is proposed to solve it effectively. Inthe optimization model, costs of the utilized machines and cutting tools, machinechanges, tool changes, set-ups and departure from good manufacturing practices(penalty function) are the optimization evaluation criteria. Precedence constraintsfrom the geometric and manufacturing interactions between features and theirrelated operations in a part are defined and classified according to their effects onthe plan feasibility and processing quality. A hybrid constraint-handling methodis developed and embedded in the optimization algorithm to conduct the searchefficiently in a large-size constraint-based space. Case studies, which are used forcomparing this approach with the genetic algorithm and simulated annealingapproaches, and the proposed constraint-handling method and other constraintmethods, are discussed to highlight the performance of this approach in terms ofthe solution quality and computational efficiency of the algorithm.

1. Introduction

Computer-aided process planning (CAPP) is an essential component for linkingthe various models and processes in a computer-integrated manufacturing system(CIMS). Process plans generated by a CAPP system should ensure the consistencyof the desired functional specifications of a part during its production processesand the observation of good practices to manufacture the part economically andcompetitively. The initial ‘variant’ CAPP systems are based on the group technology(GT) coding and classification system to select a baseline process plan for a partfamily. In such systems, approximately 90% of the part attributes and parameterscan be yielded automatically while the remaining 10% is achieved through modifyingor fine-turning the process plans manually. Current systems are mostly generativeand artificial intelligence (AI) techniques are used to facilitate the generation of acomplete process plan. An important trend in generative CAPP systems is to makethem more adaptable to dynamically varying manufacturing conditions such asthe availability of machines and tools. Conventionally, a generative CAPP system

Revision received September 2003.ySingapore Institute of Manufacturing Technology, 71 Nanyang Drive, Singapore 638075.zDepartment of Mechanical Engineering, National University of Singapore, 9

Engineering Drive 1, Singapore 117576.*To whom correspondence should be addressed. e-mail: [email protected]

International Journal of Production Research ISSN 0020–7543 print/ISSN 1366–588X online # 2004 Taylor & Francis Ltd

http://www.tandf.co.uk/journals

DOI: 10.1080/00207540310001652897

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

consists of three main consecutive activities: (1) recognizing manufacturing features

from a designed part; (2) determining machining operation types and enumerating

alternative set-up plans as well as applicable machining resources in a dynamic

workshop environment; and (3) selecting suitable set-up plans and machining

resources, and sequencing machining operations to seek the lowest machining cost

of the part. The workflow of a generative CAPP system is shown in figure 1. A global

optimum process plan can be achieved from an optimization of each individual

activity. The third activity, which is the focus of this paper, can be modelled as an

optimization problem and solved using AI techniques. However, such a problem is

well known to be an intractable reasoning and decision-making process considering

the inter-related geometric relationships between features, the complex technologi-

cal requirements and the multiple evaluation criteria. To address this problem

effectively, effort should be made to design a more apt optimization model and

develop a more efficient method for handling precedence constraints of a part.

Determination of machining operations for the manufacturing features

Selection of suitable set-ups and machining resources

Sequencing machining operations considering the precedence constraints

Machining resources

AI reasoningengines

Enumeration of candidate set-up plans and applicable machining resources for

the machining operations

Determination of precedence constraints among the machining operations

Optimizationalgorithms

A designed part Recognized

manufacturing features

Geometric and AIreasoning engines

Phase 1:Recognition ofmanufacturingfeatures

Phase 2: Identification of machiningoperations and resources

Phase 3:Determination ofprocess plans and machining cost

Optimized process plansand machining cost

Figure 1. The workflow of a generative CAPP system.

1956 W.D. Li et al.

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

In this paper, selecting machining resources, determining set-up plans andsequencing operations for the process planning of a prismatic part are incorporatedas a constraint-based optimization model. In this model, six criteria are evaluated:(1) cost of machines utilization; (2) cost of cutting tools utilization; (3) number ofmachine changes; (4) number of tool changes; (5) number of set-ups; and (6) numberof violated constraints (a penalty function). Some precedence constraints based onthe geometric and manufacturing interactions between features and their relatedoperations in a part are defined and classified from the viewpoints of plan feasibilityand processing quality. A tabu search-based approach embedded with a hybridconstraint-handling method is developed to effectively search for optimized processplans in a large-size multiple-dimensional space with complex constraints. Throughcomprehensive comparisons of this approach with a genetic algorithm and asimulated annealing approach for two prismatic parts under different workingconditions, and the proposed constraint-handling method in this research withother constraint methods, the merits can be illustrated clearly.

The rest of this paper consists of four sections. In section 2, previous related workis investigated and summarized. In section 3, representation models of process plans,precedence constraints and optimization criteria are defined. Based on these models,a constraint-based tabu search approach is described in detail. In section 4, compu-tational experiments are made to compare the approach with other methods. Finally,the present study is concluded and future research is outlined in section 5.

2. Previous related work

Recent related work can be categorized as the knowledge-based reasoningapproach (Chang 1990, Wong and Siu 1995, Chu and Gadh 1996, Wu and Chang1998, Tseng and Liu 2001), graph manipulation approach (Chen and LeClair 1994,Irani et al. 1995, Lin et al. 1998), Petri-nets-based approach (Kruth and Detand1992), fuzzy logic reasoning approach (Ong and Nee 1994, Zhang and Huang1994, Gu et al. 1997), evolutional algorithm and heuristic reasoning-based approach(Vancza and Markus 1991, Chen and LeClair 1994, Yip-Hoi and Dutta 1996, Zhanget al. 1997, Chen et al. 1998, Reddy et al. 1999, Ma et al. 2000, Qiao et al. 2000, Leeet al. 2001, Li et al. 2002a), etc. In the reasoning processes, manipulating preliminaryprecedence constraints between operations effectively using ‘if-then’ rules, graphsand matrices, is an important issue.

The QTC system developed by Chang (1990) uses a knowledge-based reasoningapproach, and aggregates machining operations with the same tool approachdirection (TAD) as a set-up. The sequence of the machining operations andset-ups is reasoned according to the precedence constraints, and an optimal sequenceis selected from several feasible sequences based on the minimum number of set-ups.In this system, several geometric and technological constraints of a part, stemmingfrom geometric interactions between operations, location tolerance requirements,reference or datum requirements, and good manufacturing practices, are consideredin the reasoning process. The aggregation concept was enhanced by Chu and Gadh(1996) to cluster operations that are machined with the same cutting tool in a set-upso as to reduce the number of tool changes. In the APSS system reported byWong and Siu (1995), the operations sequencing algorithm consists of threeconsecutive algorithms, viz., the transformation algorithm, the refinement algorithm,and the linearization algorithm. After preliminary precedence constraints have beengenerated in the transformation algorithm, the refinement algorithm creates the

1957Optimization of process plans

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

details of the operations using a ‘refinement’ knowledge base considering goodmanufacturing practices, and represents them in a tree structure. In the linearizationalgorithm, the tree structure will be linearized into the final required operationsequence.

Irani et al. (1995) developed a graph manipulation approach for operationssequencing, in which the Hamiltonian path (HP) analogy is utilized to representthe process plans and the Latin multiplication method (LMM) is implemented toenumerate all the feasible HPs under constraints. The optimal process plan is an HPthat corresponds to the least number of set-up disruptions required from start tofinish to process each feature once and only once. In the work reported by Lin et al.(1998), a graph-search strategy was designed for operations sequence planning forprismatic parts with interacting features. The machining of a feature might affect thesurface quality of other interacting features, and several heuristic rules concerningmachining practices were developed to help the graph-search process generatehigh-quality process plans with lowest machining cost. A hybrid method combiningan unsupervised learning algorithm and a graph manipulation algorithm wasproposed by Chen and LeClair (1994). The sequence of the features in each set-upis constrained and determined by the interacting feature relationships, which arestored in an episodal associative memory. After the process of feature sequencing,several graph manipulation algorithms are proposed to obtain the optimal toolssequence for creating the features in a set-up by minimizing the tool changes.

Kruth and Detand (1992) proposed generic Petri-nets to represent parametricfeatures and their related machining operations. After being evaluated usingmanufacturing knowledge bases, such as general machine data and manufacturingcapability data, the separated Petri-nets for compound features or features withidentical TADs are joined together. The same procedure is then applied to thefeatures with different TADs, and a large Petri-net is finally formed, in which allvalid alternatives to machine a part are described.

The common characteristics of the above approaches include: (1) difficulty infinding global and optimal plans using these reasoning approaches although theseapproaches can generate feasible solutions; and (2) low reasoning efficiency in acomplex machining environment.

Process planning objectives are often imprecise and can even be conflicting due toinherent differences in the feature geometry and technological requirements. Fuzzylogic is suitable for presenting such imprecise knowledge and several approachesusing this technique have been developed to address process planning problems.Based on fuzzy memberships, the objective (Zhang and Huang 1994, Tiwari andVidyarthi 1998) is to minimize the dissimilarity among the process plans selected fora family of parts, and optimal process plans can be generated for each part family. Afuzzy logic-based approach has been reported by Ong and Nee (1994) to identify andprioritize important features based on the geometric and technological informationof a part. Important features and their operations correlate well with the manu-facturing cost than the less important features and operations. Hence, operationssequencing of important features is first carried out within a much smaller searchspace. The operations of the less important features can then be arranged easily dueto reduced constraints.

Recently, evolutional and heuristic algorithms have been applied to processplanning research, and multiple objectives, such as the minimum use of expensivemachines and tools, minimum number of set-ups, machine changes and tool changes,

1958 W.D. Li et al.

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

and achieving good manufacturing practice, have been incorporated and consideredas a unified model to achieve a global optimal target (Zhang et al. 1997, Li et al.2002a). The related works are summarized in table 1. The popular strategies forgenerating neighbourhood or next-generation process plans in tabu search (TS),simulated annealing (SA), and genetic algorithm (GA) techniques are illustrated intable 2. However, the following two issues are still outstanding and require carefulconsiderations.

The first issue is that the multiple objectives in the unified optimization modelsare often conflicting, and a prismatic part with alternative TADs and cuttingmachines/tools creates a large searching space. Much effort is needed to design asuitable optimization approach to determine the optimal results with short reason-ing iterations to meet the practical workshop situations with dynamically varyingresources and workloads. It is also imperative to conduct comprehensive studieson the performance of the various optimization algorithms to highlight theircharacteristics.

The second issue is the processing of precedence constraints in the variousoptimization approaches. The efficiencies of the graph-based heuristic algorithm(Vancza and Markus 1991) and the tree traversal algorithm (Yip-Hoi and Dutta1996) are not high and the search is not global such that optimal plans might be lostduring the inferencing and reasoning processes. The test and generation method usedby Zhang et al. (1997), Reddy et al. (1999), Ma et al. (2000) and Lee et al. (2001) is togenerate process plans randomly, and then test and select some feasible plans forfurther manipulations. The fundamental problems of this approach include the lowefficiency and difficulty to generate reasonable initial plans for a complex part. Thetwo-step manipulation (Zhang et al. 1997, Reddy et al. 1999, Qiao et al. 2000) isspecific for the crossover operations, and it is not applicable to SA and TS, or otheroperations in GA such as creating initial generations and mutations. In comparison,the penalty method (Chen et al. 1998, Reddy et al. 1999) and the constraintadjustment method (Li et al. 2002a) performs better in terms of computationalefficiency and extensible search space. However, the former method, which penalizesmore values to the evaluation function of invalid plans, might cause more time tobe spent on evaluating these invalid plans, and the final optimal plan might notbe usable. The latter method, which adjusts a plan from an invalid arrangementto a valid plan through a set of manipulations, does not work well with conflictingconstraints and evaluation criteria. New methods should be studied to handle thepossibility of conflicting constraints, which can be classified according to their impacton the feasible manufacturability of process plans, to achieve good computationalefficiency and robustness.

3. Process planning optimization using the tabu search algorithm

TS, which utilizes some selected concepts of GA and SA and is characterized bythe use of a flexible memory strategy, is a meta-heuristic algorithm that can guidesearch processes to overcome local optimal solutions in combinatorial optimizationproblems. The fundamental of TS is to avoid entrapping in cycles by forbiddingmoves that take the solution to points in the solution space previously visited (hence‘taboo’). Although TS is still in its infancy stage, during the last few years it hasbeen reported as a satisfactory solution approach for a variety of problems suchas scheduling, parallel computing, transportation, routing and network design.However, relatively few works have been reported in using TS to address process


Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

Related workOptimizationstrategies

Optimizationcriteria

Adjustmentstrategies

Constrainttypes

Constrainthandling

Vancza and Markus(1991)

Genetic algorithm (1) Number of set-ups(2) Number of tool changes(3) Total cost of individual

operations

(1) Crossover(2) Mutation

(1) Fixed order ofoperations

(2) Reference precedence(3) Feature interactions

Graph-based heuristicalgorithm

Yip-Hoi and Dutta(1996)

Genetic algorithm Part machining time (1) Crossover(2) Mutation

(1) Fixed order ofoperations

(2) Resource constraints

Tree traversalalgorithm

Zhang et al. (1997) Genetic algorithm (1) Machine costs(2) Cutting tool costs(3) Number of machine

changes(4) Number of tool changes(5) Number of set-ups


(1) Fixture constraints(2) Datum dependency(3) Fixed order of

operations

(1) Test and generation(2) Two-step manipulation

for crossover

Chen et al. (1998) Hybrid Hopfieldneural networkand SimulatedAnnealing

(1) Number of set-ups(2) Number of tool changes(3) Number of constraint

violation

Exchange of therows in NeuralNetworks

(1) Fixture constraints(2) Feature interactions(3) Tolerance(4) Tool approach

directions

Penalty algorithm

Reddy et al. (1999) Genetic algorithm (1) Number of machinechanges

(2) Number of tool changes(3) Number of set-ups(4) Number of constraint

violation


(1) Tool approachdirections

(2) Datum dependency(3) Tolerance(4) Feature interactions(5) Good machining

practice

(1) Test and generation(2) Two-step manipulation

for crossover(3) Penalty algorithm

1960

W.D

.Liet

al.

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

Qiao et al. (2000) Genetic algorithm (1) Number of constraintviolation

(2) Number of set-ups(3) Number of tool changes


Good machiningpractice

Two-step manipulationfor crossover

Ma et al. (2000) Simulated annealingalgorithm

(1) Machine costs(2) Cutting tool costs(3) Number of machine


(1) Swapping(2) Random

exchange(3) Mutation

(1) Fixture constraints(2) Datum dependency(3) Good machining

practice

Test and generation

Lee et al. (2001) (1) Simulated annealingalgorithm

(2) Simplified tabu searchalgorithm



(1) Swapping(2) Random

exchange(3) Mutation

(1) Fixture constraints(2) Datum dependency(3) Good machining

practice

Test and generation

Li et al. (2002a) Hybrid geneticalgorithm andsimulated annealingalgorithm



(1) Crossover(2) Mutation(3) Shift(4) Adjacency

swapping

(1) Fixture constraints(2) Tool interactions(3) Datum interactions(4) Thin-wall interactions(5) Feature priorities(6) Material-removal

interactions(7) Fixed order of operations

Constraint adjustmentmethod

The adjustment strategies is the next generation strategies for the genetic algorithm, neighbourhood strategies for simulated annealing, or optimized strategies for neuralnetworks.

Table 1. Summary of the related work for evolutional and heuristic algorithm-based process plan optimization.

1961

Optim

izatio

nofprocess

plans

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

planning optimization in a multiple-objective and complex constraint-based searchspace.

This section consists of two parts. First, the process planning problem is repre-sented as a constraint-based optimization model. Next, a TS-based optimizationalgorithm is presented and a constraint-handling method that can address complexprecedence constraints effectively is elaborated.

3.1. Knowledge modelling of process planning optimizationA process plan for a part consists of machining operations, applicable alter-

native machining resources (machines and cutting tools), set-up plans, machiningparameters, operation sequence, etc. A set-up is usually defined as a group ofoperations that are machined on a single machine with the same fixture configuration.Here, a set-up is a group of features with the same TADmanufactured on a machine.

Strategies Descriptions and examples Applicable algorithms

Two process plans* are chosen randomly as parent plans. For them, a cutting point is randomly determined, and each parent plan is separated as left and right parts from the cutting point.

A two-step manipulation is applied to generate two children plans that can satisfy the precedence constraints. (i) Copy the left part of parent 1 to the left part of child 1; (ii) In parent 2, find the operations in the right part of parent 1 and copy them to the right part of child 1 according to their sequences in parent 2. Child 2 can be obtained in a similar procedure.

Crossover

Parent 2 Oper[1] Oper[4] Oper[10] Oper[5] Oper[8] Oper[3] Oper[2] Oper[7] Oper[9] Oper[6]

Parent 1 Oper[9] Oper[2] Oper[4] Oper[1] Oper[10] Oper[7] Oper[8] Oper[6] Oper[5] Oper[3]

Child 1 Oper[9] Oper[2] Oper[4] Oper[1] Oper[10] Oper[5] Oper[8] Oper[3] Oper[7] Oper[6]

A cutting point

Genetic algorithm

For a process plan, an operation in the plan is randomly chosen and an alternative operation is used to replace this operation.

Mutation An alternativeoperation

Replace

A plan Oper[1] Oper[5] Oper[3] Oper[10] Oper[9] Oper[2] Oper[7] Oper[4] Oper[6] Oper[8]

(1) Genetic algorithm

(2) Simulated annealing

(3) Tabu search

The shift strategy is to remove an operation from its position and insert it at another position in the current plan.

Shift



(2) Tabu search

The swapping strategy is to exchange two operations chosen randomly in a plan.

Swapping



(2) Tabu search

The adjacent swapping strategy is to exchange two adjacent operations in a plan. Adjacent Swapping



(2) Tabu search

* The example of a process plan here consists of ten operations. Each operation is represented as Oper and its ID (Oper[ID]), and the sequence of these operations in the string is the machining sequence of the operations.

Table 2. The popular neighbourhood or next-generation strategies in the optimizationmethods.

1962 W.D. Li et al.

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

For example, in figure 2, a hole with two TADs is considered to be related with two

set-ups. A valid TAD should satisfy the following conditions:

. Tool accessibility. If a cutting tool for machining a feature on a part along one

of its TADs is obstructed by other features on the part, or the cutting tool

cannot be positioned to machine the feature along the TAD correctly, the

TAD for this feature is considered to be inaccessible and invalid.

. Fixture. If there are no valid fixture elements for holding a part to machine a

feature along one of its TADs, the feature cannot be fixtured and machined

along the TAD, and the TAD is invalid.

. Availability of cutters. Along one of its TADs, if the shape of a feature is

beyond the scope of any cutting tools available, the feature cannot be

machined along the TAD, and the TAD is invalid.

. Tolerance and surface finish requirements. A feature should not violate the

tolerance and surface finish requirements of machines when it is machined

along one of its TADs. Otherwise, the TAD is invalid.

The operations and their relevant machines, cutting tools, TADs, and machining

parameters can be modelled as a process plan. If there are n operations for machin-

ing a part, the plan will be n bits, and each bit represents an operation. The sequence

of the bits is the machining operation sequence of the process plan. An operation

has a set of candidate machines, tools, and TADs under which the operation can be

executed. In an object-oriented description, a bit (an operation) in a process plan and

a process plan can be defined as shown in table 3. The method for determining the

machining costs of a process plan and the relevant variables defined in table 3 are

explained next.

Two methods are defined in the class process plan. In Cost_Comp(), the

contribution factors for the total machining cost in a process plan include all the

machine costs, tool costs, machine change costs, set-up costs, tool change costs, and

additional penalty costs due to the violation of precedence constraints. These costs

can be computed as below.

(1) Total Machine Cost (TMC)

TMC ¼Xni¼1

ðOper½i�:Machine id�MC½Oper½i�:Machine id�Þ ð1Þ

where MC is the Machine Cost of a machine.

First TAD Second TAD

Figure 2. A through hole with two TADs.


Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

(2) Total Tool Cost (TTC)

TTC ¼Xni¼1

ðOper½i�:Tool id�TC½Oper½i�:Tool id�Þ ð2Þ

where TC is the Tool Cost of a tool.

Class Process_Plan_Bit (Only containing a variable domain)

Data types Variables Descriptions

int Operation_id The id of the operationint Machine_id The id of a machine to

execute the operationint Tool_id The id of a cutting tool to

execute the operationint TAD_id The id of a TAD to apply

the operationint[ ] Machine_list[ ] The candidate machine list

for executing the operationint[ ] Tool_list[ ] The candidate tool list for

executing the operationint[ ] TAD_list[ ] The candidate TAD list for

applying the operationspecific_type Operation_parameters Other machining parameters

of the operations

Class Process_Plan (Containing a variable domain and a method domain)The Variable Domain

Data types Variables Descriptions

Process_Plan_Bit(an operation)

Oper[n] Define a process plan Oper[n]based on the above class –Process_Plan_Bit. n is thenumber of operations in the plan.

double TMC Total Machine Cost of the plandouble TTC Total Tool Cost of the plandouble TSC Total Set-up Cost of the plandouble TMCC Total Machine Change

Cost of the plandouble TTCC Total Tool Change Cost of

the plandouble APC Additional Penalty Cost of

violating constraints in the plandouble TWC Total Weighted Cost of the plan

The Method Domain

Return types Methods Descriptions

double[ ] Cost_Comp() The method is used to computethe individual costs and weightedtotal cost of the plan

Boolean Variant_Plan_Generation()

Two basic manipulations in themethod are used to generate variantplans from the plan (neighbourhoodstrategies)

Table 3. Class definitions of an operation (a process plan bit) and a process plan.

1964 W.D. Li et al.

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

(3) Number of Set-up Changes (NSC), Number of Set-up (NS), and TotalSet-up Cost (TSC)

NSC and NS are computed using equations (3) and (4).

NSC ¼Xn�1

i¼1

�2ð�1ðOper½i�:Machine id,

Oper½i þ 1�:Machine idÞ,�1ðOper½i�:TAD id,

Oper½i þ 1�:TAD idÞÞ ð3Þ

NS ¼ 1þNSC ð4Þ

The Set-up Cost (SC) is considered to be the same for each set-up. Hence,

TSC ¼XNS

i¼1

SC ð5Þ

where �1ðX ,YÞ ¼10X 6¼ YX ¼ Y

�, �2ðX ,YÞ ¼

0 X ¼ Y ¼ 01 otherwise

�

(4) Number of Machine Changes (NMC) and Total Machine Change Cost(TMCC)

NMC ¼Xn�1

i¼1

�1ðOper½i�:Machine id,Oper½i þ 1�:Machine idÞ ð6Þ

The Machine Change Cost (MCC) is considered to be the same for each machinechange. Hence,

TMCC ¼XNMC

i¼1

MCC ð7Þ

(5) Number of Tool Changes (NTC) and Total Tool Change Cost (TTCC)

NTC is computed as:

NTC ¼Xn�1

i¼1

�2ð�1ðOper½i�:Machine id,Oper½i þ 1�: ð8Þ

Machine idÞ,�1ðOper½i�:Tool id,Oper½i þ 1�:Tool idÞÞ

Similarly, the Tool Change Cost (TCC) is considered to be the same for each toolchange. Thus

TTCC ¼XNTC

i¼1

TCC ð9Þ

(6) Number of Violating Constraints (NVC) and Additional Penalty Cost(APC)

NVC ¼Xn�1

i¼1

Xnj¼iþ1

�3ðOper½i�:Operation id,Oper½ j�:Operation idÞÞ ð10Þ


Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

A fixed Penalty Cost (PC) is applied to each violated constraint. Thus

APC ¼XNVC

i¼2

PC ð11Þ

where �3ðX ,Y Þ ¼

1 The sequence of X and Y operations violates constraints0 The sequence of X and Y operations is in accordance

to constraints

8<: :

(7) The total weighed cost (TWC)

TWC ¼ w1�TMC þ w2

�TTC þ w3�TSC þ w4

�TMCC þ w5�TTCC þ w6

�APC ð12Þ

where w1�w6 are the weights.In equation (12), the functions of w1�w5 are twofold. They serve as switch

functions for users to select the cost factors to be considered. For example, insome workshops, the machine costs, the numbers of set-ups and the number ofmachine changes are considered to contribute mainly to the total machining cost.Therefore, in this case, w2 and w5 for the utilization of tools and the tool changescan be assigned as 0, while the rest can be assigned as 1. Another function ofthese weights is to provide the flexibility to customize the optimization algorithmin some situations. For example, if a user considers the utilization of machines to beimportant, its weight can be increased. w6 can be used to switch the penalty functionof the ‘soft’ constraints on or off. If no ‘soft’ constraints are considered, it is assignedas 0. Otherwise, it is 1. Detailed explanations of the constraints and the algorithm aregiven in the subsequent sections.

MC, TC, SC, MCC, TCC and PC are stored in several tables of a relationaldatabase as illustrated in table 4.

Another method that is defined in the class process plan is theVariant_Plan_Generation() method. It can be used to provide neighbourhoodstrategies in the proposed TS algorithm. The details will be explained in section 3.2.

The geometric and manufacturing interactions between features as well astechnological requirements in a part are considered to generate some preliminary

Fields Descriptions

Information about machinesID Id of the machine (keyword)Type Type of the machineMC Machine Cost of the machine

Information about cutting toolsID Id of the cutting tool (keyword)Type Type of the toolTC Tool Cost of the tool

Cost information for set-ups, machine changes, tool changes and penaltiesSC Set-up CostMCC Machine Change CostTCC Tool Change CostPC Penalty Cost

Table 4. Definitions of several tables for storing machining index costs in a database.

1966 W.D. Li et al.

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

precedence constraints between operations. Constraints affect the generation ofprocess plans and can be classified as ‘hard’ or ‘soft’ constraints (Faheem et al.1998). Hard constraints affect the manufacturing feasibility and a process planshould be consistent with these constraints. Compared to hard constraints, softconstraints only affect the quality, cost or efficiency of a process plan. These softconstraints can be violated at certain times in cases of contradictions to some hardconstraints to achieve the lowest total machining cost. The classifications, definitionsand illustrative examples of precedence constraints are given in table 5.

Based on these concepts, the constrained process planning optimization modelcan be represented as follows:

Minx TWCðxÞ

s:t: NVCðxÞ ¼ 0 for all hard constraints

x 2 fTrial process plansg

3.2. TS algorithmA typical TS is used here and it consists of three main strategies, viz., the

forbidding strategy, freeing strategy and aspiration strategy (Glover 1997). Theforbidding strategy controls the solution that enters the tabu list. The freeingstrategy is used to manage the solution that exits the tabu list and when. The aspira-tion strategy is the interplay between the forbidding and freeing strategies for selectingtrial solutions. The workflow of the algorithm is shown in figure 3. Some basicelements in the algorithm are briefly stated as follows.

(1) Initial plan, current plan and elite plan. An initial plan is generated with noperations. The sequence of the n operations is randomly arranged, and themachine, tool and TAD for the execution of an operation in the plan arerandomly determined from the corresponding candidate lists. The currentplan is the optimized solution in each iteration and is used for generating theneighbourhood trial solutions. An elite plan records the best solution foundso far. A current plan might not be an elite plan since the current plan is onlythe best move in its neighbourhood trials at a particular iteration. Thus, itmight be worse than the elite plan recorded thus far.

(2) Neighbourhood strategies. A set of variant plans can be generated from acurrent plan for trials using neighbourhood strategies. The neighbourhoodstrategies, which are illustrated in figure 4, include two basic manipulations.The first mutation manipulation randomly chooses a set of machines,tools and TADs from the alternative lists to replace the current ones inthe operations of a plan. The second manipulation changes the sequenceof two operations in a plan using shifting, swapping or adjacent swappingoperations. The size of the variant plans and four probabilities of applyingthe mutation, shifting, swapping and adjacent swapping operations arerepresented as SN, Pmu, Psh, Psw and Pas respectively.

(3) Forbidding, aspiration and freeing strategies. The forbidding strategy canavoid cycling and local minima by forbidding certain moves that havebeen attempted during the most recent computational iterations. A tabu list,which is stored as a linked list and managed as a ‘first-in-first-out’ (FIFO)queue, is employed to store recently visited current plans to realize thisstrategy. An aspiration strategy can enable a plan that has been forbidden


Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

Constraints Definitions Examples Explanations

Hard constraints

Fixture interactions The clamping or supporting faces formachining a feature are destroyed bymachining another feature earlier.

Vise jaw

Vise jaw

Hole

Chamfer

The hole should be machined before thechamfer, otherwise it cannot be fixtured.

Tool interactions The positioning faces required by a cuttingtool to machine a feature are removed bythe machining of another feature earlier.

Chamfer

Hole

In order to position a drilling tool correctly,the drilling of the hole should precede themachining of the chamfer.

Datum interaction In order to locate a part for machining orinspection, some datum faces in the partare used as reference planes. A datuminteraction occurs when machining afeature destroys the datum required foranother feature.

A

0.01 A

Datum feature (top face)

Machiningface

The top face (the datum feature) should bemachined prior to the base face.

Feature priorities A feature should be machined before itsassociated features. Another case is thata feature should be machined first toprovide an entrance face for machiningan interacting feature.

Hole

Countersunk The countersink is an associated feature andshould be machined after the primaryhole.

Pocket 1

Pocket 2Slot The two pockets should be machined first to

expose the entrance faces of the slot.

1968

W.D

.Liet

al.

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

Fixed order ofmachining operations

This case includes some explicit precedenceconstraints, for example, turning-grooving-chamfering prior to threadcutting.

Operations for a hole:(1) Drilling(2) Boring and(3) Reaming

A typical sequence of machining a hole isdrilling-boring and reaming.

Soft constraintsThin-wall interactions A thin-wall interaction occurs when the

distance between features is very smalland causes precedence constraints inmachining. Hole

Slot Thin wall Good practice should be drilling the hole,then machining the slot to avoid thedeformation of the thin wall.

Material-removal interactions For two features with geometricinteractions, if the different materialremoval sequences of features influencethe cost or the quality of machining andcause precedence constraints betweenthese features, a material-removalinteraction occurs.

Step

Hole

The step should be machined prior to thehole for achieving high machiningefficiency (milling is faster than drilling)and surface quality.

Table 5. Definitions and classifications of precedence constraints.

1969

Optim

izatio

nofprocess

plans

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

by the tabu list to become acceptable if it satisfies a certain criterion. Thisstrategy can provide some flexibility to the tabu restrictions to lead the movein a desirable direction. A common criterion is to override a tabooed plan ifits machining cost is lower than that of the elite plan. The freeing strategy isused to control the plan that can be freed from the tabu list and when thisshould occur. This strategy is applied in one of the following two cases: (a)when the tabu list is full and a new plan needs to join the list, the earliest

Initialize a process plan and a tabu list to record the

recently visited plans

Create a set of variant process plans from the current plan using neighbourhood strategies

1. Evaluate the initial process plan 2. Assign this plan as the elite plan 3. Assign this plan as the current plan 4. Activate the tabu list to store the current plan

1. Evaluate these variant plans 2. Choose the best plan from

these variant plans

Update the tabu list

Is the chosen plan in the tabu list ?

Is the chosen plan better than the elite

plan ?

Y

Y

1. Replace the current plan with the chosen plan

2. Replace the elite plan with the chosen plan

Replace the current plan with the chosen plan

N

N

Is the stopping criterion satisfied ?

N

Y

Forbidding strategy

FreeingstrategyAspiration

strategy

END

Figure 3. The workflow of the TS algorithm for process planning optimization.

1970 W.D. Li et al.

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

forbidden plan in the tabu list should be freed so that they can be recon-sidered in future search; and (b) when an evaluated current plan satisfies theabove forbidding strategy but passes the aspiration criterion test, this planshould be considered as admissible (in this case, the aspiration strategy isequivalent to the freeing strategy). Some details for applying these threestrategies are given below (illustrated and highlighted in figure 3 as well):

(i) After applying the neighbourhood strategies to a current plan, somevariant plans are generated. Select the best plan from these variantplans as the chosen plan.

(ii) If the chosen plan is not better than the current elite plan that recordsthe best solution found until the last iteration, the current plan is putinto the tabu list to avoid this plan being revisited (forbidding strategy).Once the tabu list is full, and a new plan needs to join the list, someplans stored in the tabu list will be released according to the FIFO rule( freeing strategy).

(iii) However, if the chosen plan is better than the current elite plan, this plancan be assigned as the new current plan to continue a new iteration ofthe search process (aspiration strategy).

The size of the tabu list, which is represented as Ts, plays a crucial role in the

search for high-quality solutions and should be well-defined. A Ts that is too

Machine_id

Tool_id

TAD_id

Chosen

Replace

Operation_id

Operation_parameter

Machine_id

Tool_id

TAD_id

Machine_list[ ]

Tool_list[ ]

TAD_list[ ]

An operation

Operation_id

Operation_parameter

Machine_id

Tool_id

TAD_id

Machine_list[ ]

Tool_list[ ]

TAD_list[ ]

An operation

…………….

A process plan

Operation_id

Operation_parameter

Machine_id

Tool_id

TAD_id

Machine_list[ ]

Tool_list[ ]

TAD_list[ ]

An operation

Operation_id

Operation_parameter

Machine_id

Tool_id

TAD_id

Machine_list[ ]

Tool_list[ ]

TAD_list[ ]

An operation

…………….

A process plan

(a)

(b)

The positions of operations are changed

Figure 4. Two basic manipulations to generate variant plans from a current plan; (a) Anoperation is changed by mutating the determined machining resources from the candidatelist; (b) The sequence of operations is changed by shifting, swapping or adjacent swappingmanipulations.


Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

small might cause a high occurrence of cycling, and a Ts that is too large

might forbid too many moves and deteriorate the solution quality. Using aHamming distance (HD) defined in equation (13) below, a current plan Operican be compared with a plan Operj stored in the tabu list to determine theirsimilarity, as well as when to apply the forbidding or freeing strategies:

HD ¼Xnl¼1

�1ðOperi½l�,Operj ½l�Þ

¼0 The two plans are considered the same

others The two plans are considered to be different

�ð13Þ

(4) Stopping criteria. Termination conditions for the searching algorithm can beset using one of the following criteria: (a) the iterations reach a predefinednumber; (b) the computation is carried out continuously for a predefinednumber of iterations after the ratio between the machining cost of thegenerated elite plan and a prespecified value (an optimal or near-optimalvalue determined using other methods such as a GA or SA) falls into a certainrange (e.g. 1%); and (c) the elite plan is kept unchanged for a predefinednumber of iterations.

(5) Hybrid constraint handling method. In the TS, the precedence constraints ina part should be considered and handled effectively to ensure the manufac-turing feasibility of process plans. However, the sequences of an initial planor a variant plan generated using neighbourhood strategies might be incon-sistent with these constraints. A two-step hybrid method has been developedto handle hard and soft constraints respectively.

[STEP 1] An adjustment algorithm is imposed on a plan (an initial plan orvariant plan generated by neighbourhood strategies) to ensure its

consistency with the hard constraints.[STEP 2] After STEP 1, the optimization model is transformed from a con-

strained problem to an unconstrained problem where the weightedsum of the various machining costs and the additional penalty

costs caused by soft constraints is minimized using the above pro-posed TS process. Under this scheme, soft constraints can be vio-

lated and compromised to achieve the lowest total machining costfrom the multiple criteria.

The workflow of the constraint adjustment algorithm in STEP 1 is described infigure 5. An example for the proposed constraint adjustment algorithm is illustrated

in figure 6. For a 14-bit chromosome (n¼ 14), the bit sequence and hard constraints

are listed in table 6. Six bits�Oper[1], Oper[4], Oper[6], Oper[11], Oper[13] andOper[14]� have no constraint relationships with other operations (n_1¼ 6).

Hence, their positions are kept the same and a linked list (LL) is formed for theother eight bits (n� n_1¼ 8). The first current bit is Oper[8], and Oper[3], Oper[9],

Oper[5] should be posterior to it according to the constraints. The updating processof LL is illustrated in figure 6. After Oper[8] has been handled, the reference to the

current bit is moved to the tail and the same procedure is continued until all bits areassigned as Handled. The final updated process plan satisfies all the hard constraints.

1972 W.D. Li et al.

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

4. Experimental results and system implementation

Three experiments have been conducted to validate and illustrate the compu-tational results of the proposed approach. In the first experiment, the crucialparameters of the approach are determined. The second experiment is used tocompare this approach with typical GA and SA methods to demonstrate their

A process plan PP with n bits (operations)

Select the bits (the number is assumed to be n_1) that do not have hard constraint relationships with other bits

in PP and keep their positions unchanged in PP .

The remaining ( 1_nn − ) bits, which are constrained to be prior to or posterior to other bits in PP , are used to form a double-

linked list ( LL ) according to their relative positions in the PP .

Traverse a bit (the current bit) in LL from the tail.

Is the current bit labeled as Handled?

Traverse the prior bit of the current bit in LL as the new current bit.

Y

N

Is there one or more bits (violated bits), which are prior to the current bit in LL , that should

be posterior to the current bit according to the hard

constraints?

The violated bits are deleted from the LL and used to form another double-linked list ( 1_LL ), which is

initially set as void, according to their relative positions in LL . 1_LL is inserted into LL just after the current bit.

Y

Set the current bit as Handled.

If all bits are labeled as Handled?

N

N

Fill the bits in the LL one-by-one back to the ( 1_nn − )

positions of PP according to their orders in LL .

The new PP satisfies all hard constraints while some initial randomness of the operation

sequence can be kept.

Y

Figure 5. The workflow of a constraint-adjustment algorithm for hard constraints.


Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

different characteristics. The third experiment compares the proposed constraint-

handling method with other methods to highlight its advantages.

Two prismatic parts are employed for the case studies. The first prismatic

part (Part 1) used by Zhang et al. (1997) is illustrated in figure 7. It consists of 14

STEP-defined manufacturing features and machining operations (n¼ 14). The rele-

vant information for machining resources, features and operations, and precedence

constraints are given in tables 7–9. The second prismatic part (Part 2 shown in figure 8)

The current bit

^ Oper[7] Oper[2] Oper[10] Oper[12] Oper[3] Oper[5] Oper[8] ^ Oper[9]

Head TailThe Initially formed LL

The formed LL_1 and updated LL for the current bit - Oper[8]

^ Oper[9] Oper[3] Oper[5] ^LL_1The current

^ Oper[7] Oper[2] Oper[12] Oper[8] ^ Oper[10]

Head Tail

LL

The updated LL after LL_1 is inserted in it

The current


Head TailHandled

The current move to

…..

Oper[7]-Oper[14]-Oper[2]-Oper[10]-Oper[4]-Oper[11]-Oper[9]-Oper[12]-Oper[3]-Oper[13]-Oper[6]-Oper[5]-Oper[8]-Oper[1]

PP Unchanged Unchanged Unchanged Unchanged Unchanged Unchanged

Oper[10]-Oper[14]-Oper[12]-Oper[8]-Oper[4]-Oper[11]-Oper[9]-Oper[3]-Oper[5]-Oper[13]-Oper[6]-Oper[7]-Oper[2]-Oper[1]

PP Unchanged Unchanged Unchanged Unchanged Unchanged Unchanged

Handled


Head Tail

The finally formed LL

Figure 6. An example process of the constraint-adjustment algorithm.

1974 W.D. Li et al.

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

with more complex features and constraints (Li et al. 2002a) consists of 14 STEP-defined manufacturing features and 20 machining operations (n¼ 20). The relevantinformation for machining resources, features and operations, and precedenceconstraints are given in tables 10–12.

Original processplan

Oper[7]-Oper[14]-Oper[2]-Oper[10]-Oper[4]-Oper[11]-Oper[9]-Oper[12]--Oper[3]-Oper[13]-Oper[6]-Oper[5]-Oper[8]-Oper[1]

Constraint 1 Oper[5] and Oper[9] should be prior to Oper[2] and Oper[7]Constraint 2 Oper[12] and Oper[8] should be prior to Oper[3], Oper[5] and Oper[9]Constraint 3 Oper[3] should be prior to Oper[5]Constraint 4 Oper[10] should be prior to Oper[7]

Table 6. A process plan with four hard constraints.

Machines

No. Types MC

M1 Drill press 10M2 Milling machine 35M3 Three-axis vertical milling machine 60

Tools

No. Types TC

C1 Drill2 3C2 Drill12 3C3 Reamer 8C4 Boring tool 15C5 Milling cutter 1 10C6 Milling cutter 2 15C7 Slot cutter 10C8 Chamfer tool 10

MCC ¼ 300, SC ¼ 120, TCC ¼ 15, PC ¼ 100.

Table 7. Available machining resources and costs in a workshop environment for Part 1.

F1

F2F3F4

F5

F6

F7

F8

F9

F10

F11

F12

F13

F14

Figure 7. A sample part with 14 features – Part 1.


Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

4.1. Experiments

(1) Experiment 1 – determination of parameters

The main parameters that determine the performance of the TS include the size

of the tabu list, Ts, the size of the variant plans from a current plan, Ns, and the four

probabilities of applying the shifting, swapping, adjacent swapping and mutation

operations, Psh, Psw, Pas and Pmu. Ts plays an important role in the search for

Features Feature descriptions Operations(Oper_id)

TADcandidates

Machinecandidates

Toolcandidates

F1 Two holes arrangedas a replicated feature

Drilling(Oper1)

þ z, �z M1, M2, M3 C1

F2 A chamfer Milling(Oper2)

�x, þ y,�y, �z

M2, M3 C8

F3 A slot Milling(Oper3)

þ y M2, M3 C5, C6


þ y M2 C5, C6

F5 A step Milling(Oper5)

þ y, �z M2, M3 C5, C6

F6 Two holes arranged as areplicated feature

Drilling(Oper6)

þ z, �z M1, M2, M3 C2

F7 Four holes arranged as areplicated feature

Drilling(Oper7)

þ z, �z M1, M2, M3 C1


þ x M2, M3 C5, C6

F9 Two holes arranged as areplicated feature

Drilling(Oper9)

�z M1, M2, M3 C1


�y M2, M3 C5, C6


�y M2, M3 C5, C7

F12 Two holes arranged as areplicate feature

Drilling(Oper12)

þ z, �z M1, M2, M3 C1


�x, �y M2, M3 C5, C6

F14 Two holes arranged as areplicate feature

Drilling(Oper14)

�y M1, M2, M3 C1

Table 8. The features, operations and candidate machining information for Part 1.

Constraints Descriptions Hard or soft

Tool interactions Oper1 should be prior to Oper2. Hard

Datum interactions Oper6 should be prior to Oper7.Oper10 should be prior to Oper11.Oper13 should be prior to Oper14.

Hard

Thin-wall interactions Oper9 should be prior to Oper8.Oper12 should be prior to Oper10.

Soft

Material removal interactions Oper8 should be prior to Oper9.Oper10 should be prior to Oper12.Oper13 should be prior to Oper14.Oper3 should be prior to Oper4.

Soft

Table 9. The precedence constraints for Part 1.

1976 W.D. Li et al.

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

high-performance solutions. A small Ts might cause a high occurrence of cycling,

and a large Ts might deteriorate the solution quality. Through trials, Ts is chosen as

20 and the comparison results are illustrated in figure 9(a) and (b). A suitable Ns can

ensure good computational efficiency and algorithm stability. Through the observa-

tions from figure 9(c) and (d), Ns¼ 30 is a good choice for the algorithm to achieve

the above objectives. Several groups of four probabilities are chosen to determine

the suitable values. In table 13, ten trials are conducted for each group of param-

eters respectively, and the statistical machining costs of the mean, maximum

and minimum for the ten final results are listed. It shows that the fourth group of

parameters in the table (Psh¼ 0.85, Psw¼ 0.85, Pas¼ 0.5 and Pmu¼ 0.85) can yield

good performance of the algorithm.

The above comparisons are based on Part 1. Through trials for Part 2 using the

same chosen parameters, satisfactory results were obtained. Hence, these parameters

are generally applicable in other situations.

(2) Experiment 2 – comparison studies of TS, SA and GA

Part 1 and Part 2 are used to demonstrate the performances of TS, SA and GA

on the same optimization model to give a comprehensive understanding of their

characteristics. The SA and GA algorithms employed here are standard algorithms

(Pham and Karaboga 2000). For GA, in order that the algorithm converges elite

strategies are used during its optimization process. The parameters of the SA and

GA algorithms have been optimized based on several case studies.

The following two conditions are chosen for the studies on Part 1.

(a) All machines and tools are available, and w1�w5 in equation (12) are setas 1; and

(b) All machines and tools are available, and w2¼w5¼ 0, w1¼w3¼w4¼ 1.

Machines

No. Types MC

M1 Drilling press 10M2 3-axis vertical milling machine 40M3 CNC 3-axis vertical milling machine 100M4 Boring machine 60

Cutting tools

No. Types TC

C1 Drill 1 7C2 Drill 2 5C3 Drill 3 3C4 Drill 4 8C5 Tapping tool 7C6 Mill 1 10C7 Mill 2 15C8 Mill 3 30C9 Ream 15C10 Boring tool 20

MCC ¼ 160, SC ¼ 100, TCC ¼ 20, PC ¼ 100.

Table 10. The information for available machines and cutting tools for Part 2.


Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

For Part 2, besides the above two conditions, an additional condition for a

dynamic workshop environment is tested as follows:

(c) Machine M2 and Tool C7 are down, w2¼w5¼ 0, w1¼w3¼w4¼ 1.

The computations illustrated in figure 10 were made for the two parts

under condition (a). The current plans of TS and SA at each iteration are used forgenerating the neighbourhood and next solutions, while in GA, each of them refers

Features Featuredescriptions

Operations(Oper_id)

TADcandidates

Machinecandidates

Toolcandidates

F1 A planar surface Milling(Oper1)

þ z M2, M3 C6, C7, C8

F2 A planar surface Milling(Oper2)

�z M2, M3 C6, C7, C8

F3 Two pockets arrangedas a replicated feature

Milling(Oper3)

þ x M2, M3 C6, C7, C8

F4 Four holes arrangedas a replicated feature

Drilling(Oper4)

þ z, �z M1, M2, M3 C2


þ x, �z M2, M3 C6, C7

F6 A protrusion (rib) Milling(Oper6)

þ y, �z M2, M3 C7, C8

F7 A boss Milling(Oper7)

�a M2, M3 C7, C8

F8 A compound hole Drilling(Oper8)

Reaming(Oper9)

Boring(Oper10)

�a M1, M2, M3

M1, M2, M3

M3, M4

C2, C3, C4

C9

C10

F9 A protrusion (rib) Milling(Oper11)

�y, �z M2, M3 C7, C8

F10 A compound hole Drilling(Oper12)

Reaming(Oper13)

Boring(Oper14)

�z M1, M2, M3

M1, M2, M3

M3, M4

C2, C3, C4

C9

C10

F11 Nine holes arrangedin a replicated feature

Drilling(Oper15)

Tapping(Oper16)

�z M1, M2, M3

M1, M2, M3

C1

C5

F12 A pocket Milling(Oper17)

�x M2, M3 C7, C8


�x, �z M2, M3 C6, C7

F14 A compound hole Reaming(Oper19)

Boring(Oper20)

þ z M1, M2, M3

M3, M4

C9

C10

Table 11. The feature and operation information for Part 2.

1978 W.D. Li et al.

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

to the best plan chosen from a population in a generation (an iteration). Each of the

elite plans is the best plan at each iteration of the three algorithms. Figure 10(a) and

(b) are for Part 1, and (c) and (d) are for Part 2. It shows that the decreasing trends

of the curves for TS and GA are smoother than that of SA. For SA, there are some

Features Operations(Oper_id)

Precedence constraintdescriptions

Hard or softconstraints

F1 Milling(Oper1)

F1 (Oper1) is the datum and supporting face forthe part, hence it is machined prior to all featuresand operations.

Hard

F2 Milling(Oper2)

F2 (Oper2) is prior to F10 (Oper12, Oper13,Oper14) and F11 (Oper15 Oper16) for thematerial removal interactions.

Soft

F3 Milling(Oper3)

F4 Drilling(Oper4)

F5 Milling(Oper5)

F5 (Oper5) is prior to F4 (Oper4) and F7(Oper7)for the datum interactions.

Hard

F6 Milling(Oper6)

F6 (Oper6) is prior to F10 (Oper12, Oper13,Oper14) for the datum interaction.

Hard

F7 Milling(Oper7)

F7 (Oper7) is prior to F8 (Oper8, Oper9,Oper10) for the datum interactions.

Hard

F8 Drilling(Oper8)

Reaming(Oper9)

Boring(Oper10)

Oper8 is prior to Oper9 and Oper10,Oper9 is prior to Oper10 for the fixedorder of machining operations.

Hard

F9 Milling(Oper11)

F9 (Oper11) is prior to F10 (Oper12,Oper13, Oper14) for the datum interaction.

Hard

F10 Drilling(Oper12)

Reaming(Oper13)

Boring(Oper14)

Oper12 is prior to Oper13 and Oper14, Oper13is prior to Oper14 for the fixed order ofmachining operations.

Hard

F10 (Oper12, Oper13, Oper14) is prior to F11

(Oper15, Oper16), and Oper12 of F10 is priorto F14 (Oper19, Oper20) for the datuminteraction.

Hard

F11 Drilling(Oper15)

Tapping(Oper16)

Oper15 is prior to Oper16 for the fixed orderof operations.

Hard

F12 Milling(Oper17)

F13 Milling(Oper18)

F13 (Oper18) is prior to F4 (Oper4) andF12 (Oper17) for the material removalinteraction.

Soft

F14 Reaming(Oper19)

Boring(Oper20)

Oper19 is prior to Oper20 for the fixedorder of machining operations.

Hard

Table 12. Precedence constraints between machining operations for Part 2.


Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

‘abrupt’ decreasing points during its iteration process. TS and SA can achieve bettersolutions than GA in terms of lower mean, maximum and minimum machiningcosts of the best process plans obtained. From the observations of the case studies,generally, TS can achieve more stable performance than SA. These observations aregenerally valid in constraint-based optimization problems for typical GA, SA andTS algorithms. These observations are in accordance with the main characteristics ofGA, which is prone to ‘premature’ solutions (converges too early and difficult to find

1000

1500

2000

2500

3000

3500

4000

4500

5000

5500

6000

1 20 39 58 77 96 115 134 153 172 191 210 229 248 267 286

Iterations

Costs

30=ST10=ST 20=ST

1000

1500

2000

2500

3000

3500

4000

4500

5000

1 20 39 58 77 96 115 134 153 172 191 210 229 248 267 286

Iterations

Costs 30=ST10=ST 20=ST

(a)

(b)

1000

1500

2000

2500

3000

3500

4000

4500

1 39 77 115 153 191 229 267 305 343 381 419 457 495 533 571

IterationsIterations

Costs

Costs

1000

1500

2000

2500

3000

3500

4000

4500

5000

5500

1 19 37 55 73 91 109 127 145 163 181 199 217 235 253 271 289

10=SN

30=SN

60=SN10=SN

30=SN

60=SN

(c)

(d)

Figure 9. Determination of parameters of the TS algorithm; (a) Machining costs of currentplans under different TS; (b) Machining costs of elite plans under different TS; (c)Machining costs of current plans under different NS; (d) Machining costs of elite plansunder different NS.

F1

F2

F3

F4

F5

F6

F7

F8

F10 F11

F9

F12

F13

x

zy

a

F14

Figure 8. ANC 101 sample part with 14 features – Part 2.

1980 W.D. Li et al.

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

the optimal or near-optimal solutions), and SA, which is vigilant to parameters andproblems (Pham and Karaboga 2000).

In tables 14 and 15, more thorough comparisons for the three algorithms onPart 1 under two conditions, and on Part 2 under three conditions are made.The computations are based on ten trials for each algorithm under each conditionrespectively (each computation was conducted for 30 s in a PIII-800 computer with256 Mb memory). Similar observations on the characteristics of the three algorithmscan be obtained. Comparing the present approach to the optimal or near-optimalresults obtained previously by the authors using a hybrid GA-SA approach (Li et al.2002a), the lowest machining costs under conditions (a) and (b) are the same, whilein condition (c), a lower machining cost (2580.0) has been found using the currentapproach (2590.0 in Li et al. 2002a). For the computation efficiency of the threealgorithms, some observations can be obtained from figure 10. In the initial optimi-zation stage (before 30 iterations), GA drops faster than SA and TS. In the middlestage (from 30 to 150 iterations), GA converges while SA and TS continue to dropgradually. In the final stage (after 150 iterations), the lowest point was found a littleearlier in the TS algorithm than the SA algorithm. The computing time for aniteration in the GA, SA and TS algorithms is usually different considering that

2800

3800

4800

5800

6800

1 26 51 76 101 126 151 176 201 226 251 276

Iterations

Costs

GA

SA TS

GA SA TS

2800

3300

3800

4300

4800

5300

5800

6300

6800

1 20 39 58 77 96 115 134 153 172 191 210 229 248 267 286

Iterations

Costs

GA

SATS

GA SA TS

2400

3400

4400

5400

6400

1 27 53 79 105 131 157 183 209 235 261 287 313 339 365 391

Iterations

Costs

2400

2900

3400

3900

4400

4900

5400

5900

6400

1 27 53 79 105 131 157 183 209 235 261 287 313 339 365 391

Iterations

Costs

GA

SA

TS GA SA TS

(a) (b)

(c) (d)

GA

SA

TS GA SA TS

Figure 10. Comparison studies of three algorithms for the two parts; (a) Machining costs ofcurrent plans for Part 1; (b) Machining costs of elite plans for Part 1; (c) Machining costsof current plans for Part 2; (d) Machining costs of elite plans for Part 2.

Psh (shifting), Psw (swapping), Pas(adjacent swapping), and Pmu (mutation)

(1.0, 1.0,1.0, 0.75)

(0.85, 1.0,0.85, 1.0)

(0.5, 0.5,0.5, 0.5)

(0.85, 0.85,0.5, 0.85)

(0.5, 0.85,0.85, 0.85)

(0.85, 0.85,0.5, 0.85)

Mean 1368.0 1366.0 1589.5 1340.5 1478.5 1419.0Maximum 1463.0 1508.0 1758.0 1363.0 1918.0 1703.0Minimum 1343.0 1343.0 1343.0 1328.0 1328.0 1328.0

Table 13. Determination of four probabilities in the TS algorithm.


Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

different manipulation operations are employed in these algorithms. In figure 10, thetime for one iteration of each of the three algorithms is normalized so that thesealgorithms can be compared directly in terms of computation efficiency and solutionquality.

(3) Experiment 3 – comparisons of constraint-handling methods

The graph-based heuristic algorithm (Vancza andMarkus 1991), the tree traversalalgorithm (Yip-Hoi andDutta 1996), and the test and generationmethod (Zhang et al.1997, Reddy et al. 1999, Ma et al. 2000, Lee et al. 2001) cannot work well on Part 1with conflicting constraints and Part 2 with complex constraints. The proposed hybridconstraint handling method in this research and the penalty method (Chen et al.1998, Reddy et al. 1999), which are applicable to complex situations with conflic-ting constraints, are used for comparisons of computation performance. The resultsare shown in figure 11. For the two parts, it can be concluded that the proposedconstraint-handling method ensures that the computational process is conducted ina smoother and more efficient way.

4.2. System implementationThe TS-based approach has been implemented using Java 1.4 under the JDK

environment. This approach, together with other functional modules developedpreviously in manufacturing feature recognition (Li et al. 2002b, Li et al. 2003)and process planning optimization (Li et al. 2002a), has been integrated to sup-port a generative CAPP system. A prismatic part can be created usingUnigraphics V15.0 and input to a feature recognition prototype system developedusing Cþþ to extract the manufacturing features and the relevant information of apart. Information about machining resources and machining operations for the

Tabu search Simulated annealing Genetic algorithm

Condition (a) Mean 1342.0 1373.5 1611.0Maximum 1378.0 1518.0 1778.0Minimum 1328.0 1328.0 1478.0

Condition (b) Mean 1194.0 1217.0 1482.0Maximum 1290.0 1345.0 1650.0Minimum 1170.0 1170.0 1410.0

Table 14. Comparison studies of three algorithms for Part 1 under two conditions.

Tabu search Simulated annealing Genetic algorithm

Condition (a) Mean 2609.6 2668.5 2796.0Maximum 2690.0 2829.0 2885.0Minimum 2527.0 2535.0 2667.0

Condition (b) Mean 2208.0 2287.0 2370.0Maximum 2390.0 2380.0 2580.0Minimum 2120.0 2120.0 2220.0

Condition (c) Mean 2630.0 2630.0 2705.0Maximum 2740.0 2740.0 2840.0Minimum 2580.0 2590.0 2600.0

Table 15. Comparison studies of three algorithms for Part 2 under three conditions.

1982 W.D. Li et al.

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

generated features are stored in an MS ACCESS database. Through JDBC and

SQL, the related machining and operation information from the database can be

retrieved and provided for the process planning optimization module, which can sup-

port four alternative methods, namely, the GA, SA, hybrid GA-SA, and TS-based

approaches, for generating optimized process plans. A sketch of the information

flow is shown in figure 12.

5. Conclusions

In this paper, a TS-based approach has been developed to optimize the

processes of selecting machining resources, determining set-up plans and sequencing

operations for process planning of a prismatic part. Through several case studies,

the characteristics and advantages of the approach have been highlighted. The

contributions of the approach can be summarized as follows:

A part file The feature recognition

prototype system

ACCESSdatabase

Machiningresource information

Featureinformation

1. GA-based approach 2. SA-based approach 3. Hybrid GA-SA approach 4. TS-based approach

JDBC& SQL

Input

Feature and machiningresource information

ODBC&SQL

Process planning optimization

Figure 12. Information flow in the system.

2400

3400

4400

5400

6400

7400

1 27 53 79 105 131 157 183 209 235 261 287

Iterations

Costs

2400

3400

4400

5400

6400

7400

1 20 39 58 77 96 115 134 153 172 191 210 229 248 267 286

Iterations

Costs

1200

2200

3200

4200

5200

6200

1 20 39 58 77 96 115 134 153 172 191 210 229 248 267 286

Iterations

Costs

Hybrid method

Penalty method

1200

2200

3200

4200

5200

6200

1 20 39 58 77 96 115 134 153 172 191 210 229 248 267 286

Iterations

Costs

Hybrid method

Penalty method

(a) (b)

Hybrid method

Penalty method

Hybrid method

Penalty method

(c) (d)

Figure 11. Comparison studies of two constraint-handling methods for two sample parts; (a)Machining costs of the intermediate current plans for Part 1; (b) Machining costs of theintermediate elite plans for Part 1; (c) Machining costs of the intermediate current plansfor Part 2; (d) Machining costs of the intermediate elite plans for Part 2.


Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

(1) The proposed approach can generate optimal or near-optimal process plansfor a prismatic part with good computation efficiency based on a com-bined machining cost criterion with weights. The optimization modelcan conveniently simulate a practical dynamic workshop through changingthe strategy of cost evaluation and considering substitution or breakdownof machines or tools. Through comprehensive comparison studies, thecharacteristics and advantages of the approach are highlighted.

(2) According to the effects on the plan feasibility, precedence constraints aredefined and classified. The developed hybrid constraint-handling methodcan address a complex situation with conflicting constraints and achievegood computation performance.

In future, the optimization of the machining parameters for each operation,which is not considered here, will be studied. In addition, hybrid strategies ofTS, SA and GA will be attempted to achieve good computational efficiency insearching for alternative optimal or near-optimal process plans with predeterminedobjectives.

References

CHANG, T. C., 1990, Expert Process Planning for Manufacturing (New York: Addison-Wesley).CHEN, C. L. P. and LECLAIR, S. R., 1994, Integration of design and manufacturing: solving

set-up generation and feature sequencing using an unsupervised-learning approach.Computer-Aided Design, 26, 59–75.

CHEN, J., ZHANG, Y. F. and NEE, A. Y. C., 1998, Set-up planning using Hopfield net andsimulated annealing. International Journal of Production Research, 36, 981–1000.

CHU, C. C. P. and GADH, R., 1996, Feature-based approach for set-up minimization of processdesign from product design. Computer-Aided Design, 28, 321–332.

FAHEEM, W., CASTANO, J. F., HAYES, C. C. and GAINES, D. M., 1998, What is a manufacturinginteraction? Proceedings of 1998 ASME Design Engineering Technical Conferences,Atlanta, GA, pp. 1–6.

GLOVER, F., 1997, Tabu Search (Boston: Kluwer Academic Publishers).GU, Z., ZHANG, Y. F. and NEE, A. Y. C., 1997, Identification of important features for

machining operations sequence generation. International Journal of ProductionResearch, 35, 2285–2307.

IRANI, S. A., KOO, H. Y. and RAMAN, S., 1995, Feature-based operation sequence generationin CAPP. International Journal of Production Research, 33, 17–39.

KRUTH, J. R. and DETAND, J., 1992, A CAPP system for nonlinear process plans. Annals of theCIRP, 41, 489–492.

LEE, D. H., KIRITSIS, D. and XIROUCHAKIS, P., 2001, Search heuristics for operation sequencingin process planning. International Journal of Production Research, 39, 3771–3788.

LI, W. D., ONG, S. K. and NEE, A. Y. C., 2002(a), Hybrid genetic algorithm and simula-ted annealing approach for the optimization of process plans for prismatic parts.International Journal of Production Research, 40, 1899–1922.

LI, W. D., ONG, S. K. and NEE, A. Y. C., 2002(b), Recognizing manufacturing features froma design-by-feature model. Computer-Aided Design, 34, 849–868.

LI, W. D., ONG, S. K. and NEE, A. Y. C., 2003, A hybrid method for recognizing interactingmachining features. International Journal of Production Research, 41, 1887–1908.

LIN, A. C., LIN, S. Y., DIGANTA, D. and LU, W. F., 1998, An integrated approach to determin-ing the sequence of machining operations for prismatic parts with interacting features.Journal of Materials Processing Technology, 73, 234–250.

MA, G. H., ZHANG, Y. F. and NEE, A. Y. C., 2000, A simulated annealing-based optimizationalgorithm for process planning. International Journal of Production Research, 38,2671–2687.

ONG, S. K. and NEE, A. Y. C., 1994, Application of fuzzy set theory to set-up planning.Annals of the CIRP, 43, 137–144.

1984 W.D. Li et al.

Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

PHAM, D. T. and KARABOGA, D., 2000, Intelligent Optimization Techniques: GeneticAlgorithms, Tabu Search, Simulated Annealing and Neural Networks (Berlin: Springer).

QIAO, L., WANG, X. Y. and WANG, S. C., 2000, A GA-based approach to machining operationsequencing for prismatic parts. International Journal of Production Research, 38,3283–3303.

REDDY, S. V. B., SHUNMUGAM, M. S. and NARENDRAN, T. T., 1999, Operation sequencingin CAPP using genetic algorithm. International Journal of Production Research, 37,1063–1074.

TIWARI, M. K. and VIDYARTHI, N. K., 1998, An integrated approach to solving the processplan selection problem in an automated manufacturing system. International Journal ofProduction Research, 36, 2167–2184.

TSENG, Y. J. and LIU, C. C., 2001, Concurrent analysis of machining sequences and fixturingset-ups for minimizing set-up changes for machining mill-turn parts. InternationalJournal of Production Research, 39, 4197–4214.

VANCZA, J. and MARKUS, A., 1991, Genetic algorithms in process planning. Computers inIndustry, 17, 181–194.

WONG, T. N. and SIU, S. L., 1995, A knowledge-based approach to automated machiningprocess selection and sequencing. International Journal of Production Research, 33,3465–3484.

WU, H. C. and CHANG, T. C., 1998, Automated set-up selection in feature-based processplanning. International Journal of Production Research, 36, 695–712.

YIP-HOI, D. and DUTTA, D., 1996, A genetic algorithm application for sequencing operationsin process planning for parallel machining. IIE Transactions, 28, 55–68.

ZHANG, F., ZHANG, Y. F. and NEE, A. Y. C., 1997, Using genetic algorithms in processplanning for job shop machining. IEEE Transactions on Evolutional Computation, 1,278–289.

ZHANG, H. C. and HUANG, S. H., 1994, A fuzzy approach to process plan selection.International Journal of Production Research, 32, 1265–1279.


Dow

nloa

ded

by [

Uni

vers

ity o

f C

alif

orni

a, S

an F

ranc

isco

] at

04:

33 1

9 N

ovem

ber

2014

optimization of process plans using a constraint-based tabu search approach

Documents