You are on page 1of 13

Int J Adv Manuf Technol (2013) 67:969981

DOI 10.1007/s00170-012-4541-1

ORIGINAL ARTICLE

Adaptive production control system for a flexible manufacturing


cell using support vector machine-based approach
V. K. Manupati & Rohit Anand & J. J. Thakkar &
Lyes Benyoucef & Fausto P. Garsia & M. K. Tiwari

Received: 10 September 2011 / Accepted: 2 October 2012 / Published online: 6 November 2012
# Springer-Verlag London 2012

Abstract Real-time adaptive production control in the flex- methodology to control the production offers better perfor-
ible manufacturing cell (FMC) is a complex issue that needs mance than the single-rule-based production control system.
to be addressed to realize good performance and high pro-
ductivity. In this paper, we have considered a support vector Keywords Flexible manufacturing cell . Support vector
machine (SVM)-based simulation approach to resolve a machine . Adaptive . Dynamic environment . Scheduling .
production control problem in an FMC that operates in a Sequencing . Simulation
dynamic environment. A SVM-based simulation approach
chooses the most relevant scheduling rule out of several
predefined ones on the basis of the current states of the
1 Introduction
system. This paper examines and compares the performance
of the SVM-based simulation approach with the competent
In todays business environment, products need to be man-
scheduling rules under two different operational environ-
ufactured with huge product variety and large volume of
ments which are characterized by the uncertainty of de-
small batches for achieving productivity and profitability.
mand. We have also developed a Visual Basic-based
However, for achieving productivity, quality, and profitabil-
simulation approach for scheduling of component parts in
ity, we have to follow the guidelines meant for advanced
the context of FMC under different situations. The SVM
manufacturing system. Flexible manufacturing cell (FMC)
is one such system that may bring us closure to realize the
desired goal.
V. K. Manupati : R. Anand : J. J. Thakkar : M. K. Tiwari (*)
An FMC consists of several machine tools, equipped
Department of Industrial Engineering and Management, Indian
Institute of Technology Kharagpur, with facilities for automatic changing of parts and cutting
Kharagpur 721302, West Bengal, India tools, interconnected by material handling and storage devi-
e-mail: mkt09@hotmail.com ces, all controlled by a central computer system [1]. It is an
V. K. Manupati evolving technology which is particularly suitable for mid-
e-mail: manupativijay@gmail.com volume and high variety production. Moreover, the FMC
J. J. Thakkar has flexibility to handle variety of part types and to provide
e-mail: jt@iem.iitkgp.ernet.in alternative processing routes. It also has the ability to handle
different production control systems operating in different
environments.
L. Benyoucef
The various kinds of operation environments have been
Facult de Saint Jerome, Aix-Marseille University, LSIS UMR 7296,
Avenue Escadrille Normandie Niemen, studied in [2]. Generally, FMC is operated under three
13397 Marseille Cedex 20, France different situations such as single-lot production, flow pro-
e-mail: lyes.benyoucef@inria.fr duction, and produce-to-order production. In single-lot pro-
duction, the known sets of parts are to be produced in a
given period of time. Whereas in flow production, the parts
F. P. Garsia
are produced continuously in fixed and known proportions.
ETSII, Universidad de Castilla-La Mancha,
Ciudad Real, Spain During produce-to-order situation, orders with different pro-
e-mail: FaustoPedro.Garcia@uclm.es cessing requirements are received randomly with different
970 Int J Adv Manuf Technol (2013) 67:969981

inter-arrival times and due dates [3]. However, for a single- al. [9] developed a neural network-based production control
lot production situation, the flexible manufacturing system system for a FMC and compared it with the simple sched-
(FMS) has produced parts within a specific period of time. uling rules to predict the FMS performance. However, sim-
Consequently, the simulation test has been conducted to ilar kind of work has been proposed by Piramuthu et al. [10]
assess the performance measures with various dispatching and Shinichi and Taketoshi [11]. In their research, authors
rules and the impact of the rules on FMS in a single-lot have considered a decision tree-based machine learning
production environment [4]. The impact of various sched- technique to production control system and found that it
uling schemes on scheduling and routing flexibility has been has been performing well in FMS. Yohanan Arzi et al.
investigated on different manufacturing systems in the con- [12] addressed the problem of decision tree-based adaptive
text of static environment. Furthermore, the authors of [5] production control system, and the system performance was
found that the routing flexibility enhances the system tested with various performance measures under two differ-
performance. ent environments. Further, they have simulated, validated,
Although there are bulk of literature available on single- and finally obtained the best performance measure. Shirazi
lot production and flow production, but not much literature and Babak [13] developed a simulation-based intelligent
is available on produce-to-order environment. In produce- decision support system for the adaptive real-time control
to-order environment, the variety of orders arrive dynami- of FMS. In their work, the authors have built a system
cally in the system. Each order consists of different process- design approach based on rule-based expert system for
ing requirements, different inter-arrival times, and different effective controllers performance. Zhibib and Richard [14]
due dates. Lin and Solberg [6] developed a market-type addressed an infrastructure-based adaptive production con-
environment which uses currency as a basis for decision trol in an agile manufacturing environment. Consequently, a
making in order to achieve optimal performance. Conse- virtual production system for job shops has been con-
quently, a framework has been proposed for distributed structed, and Petri nets were used as modeling tool for
production control system in a shop floor environment to detecting the deadlock in the system. Ottaway et al. [15]
control production using autonomous agents. The distribut- presented a conceptual and prototypical model of an adap-
ed production control system has the advantages of dealing tive production control based on agent technology. The
with the complexity in the context of FMS. Some of the resources and work pieces in the system are assumed as
advantages of distributed production control system are: agents. By introducing the supervisory agents, they have
greater reliability, modularity, effectiveness in handling de- reduced the production lead, and through varying simulation
vice failures, increased flexibility, improved performance, studies, they demonstrated the improvements for different
and cost-effectiveness [7]. The high-level uncertainty is performance measures. Li et al. [16] proposed a one-of-a-
found in produce-to-order environment especially when it kind production for adaptive production control in a flow
operates in real-time production control. General approach shop environment. Through this approach, the authors im-
for controlling the production in real-time scenario is proved the production efficiency and competitiveness. To
through simple dispatching rules. To our knowledge, no select essential system attributes based on various produc-
single scheduling rule has been performing better in differ- tion requirements is a complex task because of the existence
ent scenario. However, different operating strategies are in of large amount of shop floor information in a manufactur-
vogue in real-time production control of manufacturing ing system. To resolve the abovementioned issue, a hybrid
systems. Stecke and Solberg [1] and Montazeri and Van genetic algorithm/decision tree-based approach has been
Wassenhove [8] examined a real-time flow control system proposed and compared. Consequently, with the proposed
and in a FMS and tested alternatives with simulation experi- approach, better performance of the system has been
ments to improve the system performance. achieved by Yeou and Ruey [17]. Shiue and Su [18] have
In this paper, we have considered an adaptive production studied the selection system attributes in which a simplified
control system to resolve the conflict of the abovementioned model cannot guarantee selection of a near-optimal subset of
difficulties in real-time produce-to-order environment. The system features to build a knowledge base.
key feature of adaptive control system is that planned state The objective of this research is to introduce a support
variables are sporadically tested, and then examined with a vector machine-based simulation approach for adaptive pro-
decision module. Thereafter, it chooses a finite set of sched- duction control system. It has been examined in two differ-
uling/dispatching rules which would control the system till ent environments distinguished by their variability in
the next trail. Thus, in adaptive production control system, demand (random produce-to-order) and the proposed pro-
distinct rules are used at different periods of time, depending duction control system (highly random-produce-to-order)
upon the state of the manufacturing system. Extensive compared with each one of the simple scheduling rules.
researches have been performed with adaptive control sys- Our approach is based on supervised learning technique-
tem by implementing machine learning techniques. David et support vector machines (SVMs), and it performs better in
Int J Adv Manuf Technol (2013) 67:969981 971

different situations. The adaptive production control in gen- of a job into the system depends upon their individual due
eral and SVM-based approach for production control system date. In order to perform the scheduling task and make the
is also tried to handle random situations. Furthermore, the system more effective, we have introduced the system state
use of complex and expensive manufacturing is justified variables, dispatching rules, and performance measures.
only with random environments. The proposed SVM has The machines are equipped with local input buffers, where
the key features such as the flexibility to use variety of limited number of jobs is stored for processing on the machine.
kernels, the sparseness of the solution, and the production If the buffer capacity reaches the limit of the part types, jobs
control obtained by optimizing the margin. waiting outside the FMC are routed onto the machines with a
SVMs are a set of well-known supervised learning first-come-first-served (FCFS) basis. When a machine com-
methods that analyze data and recognize the patterns. pletes its processing, the completed part type stays in output
Through this analyzed data and various patterns, the buffer. Subsequently, the part type is routed and follows the
classification and regression analysis have been con- same pattern as explained above for consecutive operations. As
ducted. The SVM-based approach had shown better per- soon as the jobs enter in the output buffer, the computer-
formance than other traditional machine learning controlled material handling devices transfer the jobs for fur-
techniques such as neural network (NN) and decision ther processing. When more than one machine is available for
tree (DT) burgers. More information can be found in performing an operation, the machine on which the operation
Corts et al. [19]. In the past few years, the supervised can be finished at the earliest is chosen. This decision involves
learning methods such as NN and DT have been used as selecting the machine for processing an operation of a part in
the scheduling approaches in the FMC. These can relate order to optimize the given performance measure. The sched-
the system behavior and the system performance to the uling rules used for assigning parts on appropriate alternative
dispatching rules with the possible amount of simulation machine have been chosen by earliest finishing time with
runs for obtaining the training sets. Even though NN or alternatives (EFTA). This has been proved from the past liter-
DT has been used by many researchers, it is difficult to ature by Buzacott and Yao [20]. If an operation is performed by
guarantee that they can maintain high accuracy for unseen more than one machine, the EFTA priority rule is a proper one.
data for designing of the FMS scheduler. Consequently, the When a machine finishes its processing, it chooses the next job
problems like bias variance trade-off, capacity control, and from among the waiting jobs in the local buffer, according to
overfitting could be raised. Here, we have developed SVMs to production control system decisions.
overcome the above difficulties. In this problem, we have considered the following
The remainder of the paper is organized as follows. In assumptions for the operational configuration of the FMC:
Section 2, we introduce the problem definition. The sched-
Assumptions:
uling schemes and the logical frame work have been
depicted in Section 3. In Section 4, we describe the FMC 1. The FMC can process a variety of part types from a
system scheduling attributes. In Section 5, we introduced large but finite and known population.
production control system for support vector machine. The 2. The system is designed in such a way that it has to
experimentation and results are introduced in Section 6. assign the jobs in a given time.
Finally, the paper concludes with Section 7 which suggests 3. Processing times are assumed to be deterministic.
the direction to explore future work. 4. Transportation time is considered to be negligible.
5. The material handling system does not impose con-
straints (blocking and starving).
2 Problem definition 6. The operation times and setup activities are includ-
ed in operation times, and
FMC comprises several flexible computer-controlled machine 7. Machines are subjected to random failures.
tools capable of performing a wide range of operations. At an Given these data, our goal is to build schedules that
FMC, a stream of orders arrives randomly each with an maximize the total throughput time and minimize the
individual due date. An operation of a job can be processed weighted tardiness.
by more than one machine at varied efficiency for processing
different jobs with relatively few setup changeover activities
where the limited number of pallets (communication net-
works, pallet changers, etc.) is equipped for the machines. 3 The scheduling scheme
The challenge in this problem is to decide the most appropriate
dispatching rule for a respective part type whose arrival is Considering the above described problem, a scheduling
more random in nature for processing into the system. We scheme is presented and developed to realize the effective-
have considered each order as one job, and the arrival pattern ness of different performance measures. The presented
972 Int J Adv Manuf Technol (2013) 67:969981

scheme is divided into two kinds of phases named as in an FMC each operation can be assigned to each one
online generation phase and offline generation phase. of the alternative machines which are capable enough to
Figure 1 illustrates the logical framework of this sched- perform the assigned operations. Here, we have used
uling scheme. At the offline generation phase, we decid- average processing time of all alternative machines,poj,
ed the training period length followed with the arrival of as an indication for the weighting factors enforced on
parts and their inter-arrival time with respect to the the FMC for performing operation o of job j. Let Qoj is
probability distribution. We consider the exponential dis- a set of alternative machines for performing operation o
tribution for the arrival of parts for scheduling. During of job j, |Qoj| is the cardinality of this set, and pojm is
the process, we predicted the length of the warm-up the processing time of operation o of job j on machine
period. Consequently, the scheduling function generator m, where poj, is defined as:
has been implemented for generating the training func- X  
tion. However, the detail description of the offline and poj pojm Qoj  1
the online phase and their functions are mentioned in the m2Foj

following sections. A framework of the entire scheduling


scheme has been depicted in Fig. 1. The marginal contribution of job j to the weighted
throughput is obtained from:
X
pj poj 2
4 Scheduling attributes of the system o2j

4.1 The performance measures of FMC Let (t) be the set of jobs that have been completed
during the scheduling period which ends at time t, then the
The two different performance measures maximization throughput, Q (t), is measured as:
of total weighted throughput and minimization of X
Qt pj 3
weighted tardiness have been introduced in the previous j2t
section to optimize the SVM-based simulation approach
for production control system. The throughput time is Each operation in an FMC can be assigned to any one
measured as the total number of jobs that have been of the alternative machines which are equipped to per-
completed during the scheduling period. Subsequently, form the assigned operations; thus, we employed the

Fig. 1 The working procedure Off-line Phase On-line Phase


for scheduling scheme

Initialize
Training period length, No. of training periods,
Inter-arrival time, Sequence of arrival of parts, Support Vector Machine
Length of warm-up period,

A set of decision functions


Randomly select state variables at the end
of warm-up period
Scheduler
i=1
Select rule i Dispatching rules
System attributes

Run the simulation FMC

Record the performance measures

No
If i = n Extraction of attributes

Yes

Select the best rule for different performance


measures

Scheduling Function Generation


Int J Adv Manuf Technol (2013) 67:969981 973

average processing time of all the alternative machines 3. Stay variability in the system, SdZ (t): The standard
poj as the weighting factor. deviation of the length of stay of jobs within the
The second performance measure, total weighted tar- system at a given point in time, t. Where SdZ(t) is
diness, evaluates the systems ability to meet its respec- defined as:
tive due dates. A job is considered as tardy if at least v
  
u
one of the jobs operations has not been completed until u X t  Aj  MeZt 2
SdZt t 7
it completes its job due date. Consequently, the total
j2t
Njobt  1
tardiness can be measured by taking all the tardy jobs
into account and the jobs that have to complete the
processing during its scheduling period which ends at 4. Mean time to due date, MeS (t): The jobs time spent
time t. Let be the penalty per one unit time (minutes) within the FMC at a given point of time until due date.
of the marginal contribution to throughput of a tardy Let Dj be the due date of job j, then MeS(t) at time t is
job per one day tardiness (>0), dj is the number of obtained from:
tardy days of job j, (t) be the set of jobs within the
X Dj  t
FMC at time t, and (t) represents the set of jobs that MeSt 8
has been completed during the scheduling period which j2t
Njobt
ends at time t. Then, the total weighted tardiness, TI (t),
is obtained from:
5. Mean time to due date variability, MeV(t): In the system
X at a given point of time, the standard deviation for the
TIt pj gdj 4 jobs is obtained from:
j2t[t
v
  
u
u X Dj  t  MeSt 2
MeVt t 0:5em 9
4.2 The systems state variables Njobt  1
j2t

The FMC state variable is capable of measuring the


attribute of a manufacturing system which affects the
value of the performance measure, for example, the 6. Total workload, TWL(t): The combination of the set
number of tardy jobs in the system, the number of jobs of jobs processing on the system and the remaining
in the queue in front of a machine, the average length operations that should be performed on the system.
of stay of jobs in the system, and the current system Here, (t) and kj(t) are the set of remaining oper-
work load. The state of the system is defined by a ations of job j at time t and the set of jobs within
vector of all the state variable values at any given point the FMC at time t, respectively. Let o is the index
of time. of operation o of job j that has been already
The following state variables were selected for two oper- assigned to machine m.
ational environments. 0 1
X X
1. Number of jobs in the system, Njob (t): The number TWLt @ po0 jm poj A 10
of jobs in the system at a given point of time, t. Let j2t o 2 kj t; o6o0
(t) be the set of jobs within the system at time t
and is defined as:
7. Workload imbalance on the system, WLI(t): This state
variable is different with other state variables; a binary
Njob t jtj 5 variable has been suggested for finding out the imbal-
ance in workloads on the machines, at a given point in
2. Mean waiting time in the system, MeZ (t): The average time t.
length of stay of jobs within the FMC at a given point of
time, t. Let Aj is the arrival time of job j, MeZ (t) at time
t is obtained from:
where :
WLIt
1 if imbalance exists 11
0 otherwise
X t  Aj
MeZt 6 Let the workload Wm(t) imposed on machine m by jobs in
j2t
Njobt its local buffer. The Pojm processing time of an operation o
974 Int J Adv Manuf Technol (2013) 67:969981

!
of job j that is waiting in the local buffer of machine m is 1
ll0 T ll0 Xn
0

obtained from: Minimize w w C xlli 13


2 i1
X
Wm t po0 jm 12
j2t S:th:  i
0 T 0
yi wll y xi  bll xi  1  0 14

At a given point of time t, the workload imbalance


exists only when the subset of machines is at least
greater than percentage of (in this study we have xi  0;
15
taken the percentage of as 25) that was imposed on with i 1; . . . ; n; x x1 ; . . . ; xk ; and C > 0:
each one of the machines which belongs to a subset of
at least two other machines. where (wll) is the weight vector, C is the penalty weight, b is
a scalar, and data xi are mapped into a higher dimensional
4.3 The scheduling rules used in the scheduling of FMC space a nonlinear mapping function . Minimizing the
1
 ll0 by
T ll0
2 w w function is converted into maximization func-
For the systems smooth processing and to complete the tion 2/||wll|| for maximizing the distances from the decision
jobs in time, the scheduling rules are suggested. In this surface to the class boundaries. The slack variable in the
study, we have chosen different scheduling rules such Pn 0
form of the penalty term C xlli is added to the objective
as Dynamic Slack (DS), Earliest Due Date (EDD), i1
function. In this way, the optimizing function becomes a
Critical Ratio (CR), Shortest Processing Time (SPT), trade-off between the size of the margin and the size of the
Longest Processing Time (LPT), and Random Selection error. Furthermore, by considering set of nonnegative La-
0 0
(RS). In the assumptions, it has been mentioned that grangian multipliers alli and b lli the Lagrangian dual is
the system has to complete its jobs on time, and the considered.
jobs whose due date time is less than 1.5 times of their
total processing time are preferred over other jobs. !
X
n
0 1X n X n  
ll0 ll0
Hence, we have eliminated the SPT, LPT, and random Maximize alli  a a yi yj K xi ; xj 1:25em
rules. Thus, only the DS, EDD, and CR rules were i1
2 i1 j1 i j
selected among the set of competing rules given below: 16
1. DS: Select the job with minimum processing time.
2. EDD: Select the job with the earliest due date. S:t :
Pn 0
17
3. CR: Select the job with the minimum ratio between alli yi 0
i1
time now until due date and its remaining process-
ing times.
4. SPT: Select the job with the shortest processing
C  ai  0
time. 18
with i 1; 2; . . . ; n:
5. LPT: Select the job with the longest processing time.
6. RS: The job has been selected from any one of the
abovementioned scheduling rules. The values C and K are free parameters which represent
the cost and the kernel function, respectively. The kernel
function is defined as
5 Production control system using support vector      
machine K xi ; xj  K xj ; xi  y T xi y xj 19

5.1 Support vector machine In this study, we have used two different types of
kernel functionsGaussian kernel and polynomial func-
In this section, we have described the SVM-based tion, where is the width of the Gaussian kernel, and b
scheduling scheme and respective mathematical formu- is the power of the polynomial kernel. They are user-
lation. Subsequently, the hyperplane function which is defined parameters.
useful for scheduling the trained data set is defined.
Gaussian kernel; ekxyk =2
Consequently, the weighted vector and the scalar are 2 2
20
obtained. The SVM minimization problem is represented
in Eqs. (13), (14), and (15). The brief description of the
SVM formulation is given below: Polynomial kernel; 1 x  xi b 21
Int J Adv Manuf Technol (2013) 67:969981 975


T
0 0 0 0
alli wll y xi bll  1 xlli 0; i Rule l
-1
1; . . . ; n: 22

0

0

xlli C  alli 0; i 1; . . . ; n: 23
Rule l
At each iteration, the solution of the Lagrangian dual of C +1
and b is used to estimate the overall optimization. To ac-
complish this, we discard the points that are not support
vectors in C, and the remaining support vectors are used to
Fig. 3 Representation of clustered variables
construct a model. KuhnTucker complimentary conditions
play an important role in solving the bias b. Having the
support vectors whose data should satisfy the inequality, incoming part type. The warm-up period is followed by
then the optimal solution for the weight vector is given by scheduling period. At the beginning of scheduling period,
0
X 0 the state variables are noted down. The simulation is carried
wllo alli yi y xi 24 out with the same state variable values for all the competing
SVs
rules, and the corresponding performance measure is
The SVM-based production control system can be dem- recorded at the end of the scheduling period.
onstrated in two phases: offline phase and online phase. The
offline phase includes the formation of a SVM-based sched- 5.2.2 Scheduling function generation
uler based on training data, whereas the other is devoted to
the implementation of the scheduler for generating the Scheduling functions are generated for each pair of sched-
scheduling rule. This has been implemented during a sched- uling rules (l,l) based on SVM. Let (Zk,Yk) denote a training
uling period. The abovementioned phases are clearly dem- sample of SVM. Then,
onstrated in the next section.
ll0 1 rule l performs better for k
Yk
1 rule l0 performs better for k
5.2 Offline phase
the scheduling functions for the pair ll are generated by
5.2.1 Training data generation solving P for weight vector wll and scalar bll. The sched-
uling function for the pair ll is given by Eq. (25).
The training data are generated by carrying out a simulation ( )
run for N scheduling rules. Each training data consist of X 0
ll0 ll0
D x sign ai yi kx; xi b0
ll
25
initial state variable Xi and the performance measure P taken SVs
at the end of the scheduling period during simulation run. Xi
represents a vector of state variable values Xi 0(x1, x2xi Thus, N (N1)/2 scheduling functions are generated out
xI), and Pi represents the performance measure for each of of N rules. The hyperplane in Fig. 3 represents Eq. (25).
the competing rules. The generation of training data in each
simulation run is depicted in the following Fig. 2. 5.3 Online phase
Each simulation run consists of a warm-up period. Dur-
ing the warm-up period, the parts are distributed to In this phase, the scheduling functions are utilized for the
machines based on a RS of scheduling rules for each determination of the scheduling rule, and it is to be

Fig. 2 Representation of Scheduling Period


Warm up period
training data generation for a
single simulation

State Variable
Performance Measure
Measurement Measurement
Point Point
976 Int J Adv Manuf Technol (2013) 67:969981

implemented during a given scheduling period. At the start- point. At each decision point DPi (point at which a decision
ing of the scheduling period, the support vector scheduler needs to be taken for selection of scheduling rule), a voting-
outputs a dispatching rule for all the machines in this period. based selection strategy known as one against one method is
The time taken for dispatching rule is called the decision utilized for the selection of best rule to be implemented. The

Part-type 1 Part-type 2 Part-type 3


Machine
1.1 1.2 1.3 2.1 2.2 2.3 3.1 3.2 3.3

M1 40 46 65 * 33 * * 76 *
M2 78 55 64 30 * 72 * * *
M3 54 70 91 * * 64 * * 79
M4 29 34 18 * 70 87 * * 89
M5 * * * 29 45 16 29 * *
Average 50.25 51.25 59.5 29.5 49.33 59.75 50.25 76 84.5

Part-type 4 Part type 5


Machine
4.1 4.2 4.3 4.4 4.5 5.1 5.2 5.3 5.4 5.5

M1 70 79 * 61 69 * * * 91 39
M2 65 59 46 32 42 25 54 40 * *
M3 75 * 44 28 53 * * 54 * 51
M4 * * * * * * * * * *
M5 * * * * * * * * * *
Average 70 69 45 40.33 54.67 25 54 47 91 45
Part-type 6 Part-type 7 Part-type 8
Machine
6.1 6.2 6.3 7.1 7.2 7.3 7.4 8.1 8.2 8.3 8.4

M1 * * * * * * * * * * *
M2 10 15 17 * * * * * * * *
M3 * * 11 60 * * * * * * *
M4 17 * 19 * * * 10 28 * * 24
M5 * * * * 75 23 * 29 43 58 *
Average 13.5 15 15.67 60 75 23 10 28.5 43 58 24
Part-type 9 Part-type 10
Machine
9.1 9.2 9.3 9.4 10.1 10.2 10.3 10.4

M1 75 * * * * * 76 72
M2 68 * * * * * 80 91
M3 * * * * * * 74 63
M4 * 23 15 * * * * *
M5 59 * * * 12 57 * *
Average 67.33 23 15 44 12 57 76.67 75.33

Fig. 4 Part types and its processing time (minutes)


Int J Adv Manuf Technol (2013) 67:969981 977

voting strategy [21] is based on a comparison between each Table 1 Part type
probability distributions Part types Proportions
pair of rules. For each pair, the winner is decided on the
for the produce-to-order
basis of the evaluation of the scheduling function Dll as environment [2] 1 0.21
follows. 2 0.09
0 3 0.06
class l; if Dll x > 0
x2 0 26 4 0.20
class l0 ; if Dll x < 0
5 0.04
6 0.08
The candidate getting the maximum number of votes is 7 0.10
implemented for the scheduling period. Suppose in case of a 8 0.06
tie, any competing rule can be applied. Since no data are 9 0.09
available at the initial decision rule, then randomly chosen
10 0.07
FCFS decision rule is used as the starting rule for the
scheduling purpose.

The inter-arrival time between jobs is exponentially dis-


6 Experimentation tributed with a mean of 20 min and uniform distribution in
range from two to seven times of the marginal contribution
The FMC is proposed for an adaptive production con- of job j, which has been considered as the due date of each
trol system in order to test the performance measures of job. The jobs for the random produce-to-order and highly
the system under two environments (random produce-to- random-produce-to-order environments are randomly sam-
order and highly random produce-to-order) with each pled from part type probability distributions shown in
one of the four combinations. Here, we compared the Tables 1, 2, and 3.
two performance measures (maximum throughput time The purpose of these probability distributions for
and minimum tardiness) with the simple heuristic rules. high variance of part-mix was made in order to simu-
However, due to the behavior of the two configurations, late variable machine workloads in the shop floor. With
different sets of competing rules are selected. In this the help of the abovementioned specifications, the sim-
study, we have chosen simple heuristic rules as compet- ulation experiment has been carried out. The FMC
ing rules which we have mentioned in Section 4.3. model used in this paper is simulated by using the
Consequently, the simulation study was carried out with programming language Visual Basic 6.5 (VBA 6.5),
two different environments for calculating the perfor- which is an object-oriented programming tool that ena-
mance of the FMC. Moreover, we chose the simulation bles graphical user interface. The simulation model has
study due to the lack of analytical models for calculat- been run on a PC with Intel Core2 Duo CPU
ing the performance of the FMC under produce-to-order T7250 @2.00 GHz, 1.99 GB of RAM. Simulations
and highly random produce-to-order environments to with a series of 16 runs each with 27,082 min were
obtain the training set. performed. For each one of the 16 runs, a different
The objective is to compare the proposed production stream of job arrivals has been generated using a
control system with each one of the simple scheduling rules
that are in the set of competing rules. Thereafter, a compar- Table 2 Part type probability distributions for random produce-to-
ative approach is undertaken to assure the performance of order environment [2]
SVM-based approach and outperforms the rule-based one
Part type Mix ratio.1 Mix ratio.2 Mix ratio.3 Mix ratio.4
for different states of FMC.
1 0.10 0.10 0.07 0.09
6.1 The simulation model 2 0.10 0.09 0.10 0.07
3 0.10 0.20 0.09 0.10
The FMC comprises five machine tools, which can process 4 0.10 0.06 0.20 0.09
jobs from a population of ten part types. Different part types 5 0.10 0.04 0.06 0.20
have different processing time with various operations per- 6 0.10 0.08 0.04 0.06
formed on alternative machine tools. The part types and 7 0.10 0.21 0.08 0.04
their processing times defined in this study are portrayed 8 0.10 0.06 0.21 0.08
in Fig. 4. Here, the maximum number of jobs that can be 9 0.10 0.09 0.06 0.21
allowed into the FMC is 50 jobs due to the constraint of the 10 0.10 0.07 0.09 0.06
limited number of pallets within the cell.
978 Int J Adv Manuf Technol (2013) 67:969981

Table 3 Part type probability


distributions for highly random Part type Mix ratio.1 Mix ratio.2 Mix ratio.3 Mix ratio.4 Mix ratio.5
produce-to-order environment [2]
1 0.13 0.13 0.09 0.05 0.20
2 0.12 0.09 0.07 0.13 0.20
3 0.09 0.07 0.06 0.09 0.06
4 0.09 0.06 0.04 0.07 0.09
5 0.07 0.04 0.20 0.06 0.06
6 0.13 0.20 0.21 0.04 0.13
7 0.06 0.21 0.06 0.20 0.09
8 0.10 0.06 0.09 0.21 0.07
9 0.12 0.09 0.05 0.06 0.06
10 0.09 0.05 0.13 0.09 0.04

different set of random seeds. The simulation experiments 6.2 Training of multiclass support vector
have been carried out for each one of the four combinations
of the two environments and two performance measures, with The training of multiclass support vector has been
the best scheduling rules on an average. Another purpose of applied to schedule the trained data sets assorted into
the simulation model is to generate the training set, which has two phases. In the primary phase, the data scaling has
been explained in the following section. been used which is useful for data preprocessing. Here,
The training samples are obtained by executing a number we adopted a linear scaling method for data prepro-
of simulation runs under specific performance criteria. In the cessing. However, this process is used for data prepro-
training sample generation process, a total of 3,000 long cessing before training the SVM, and the scale for
runs were executed to obtain the optimal 530 training sam- each attribute is in the range of [1, 1] with the
ples. The optimal training sample indicates that under cer- following equation.
tain rule combination, the FMC has the maximum
throughput after long run. The program is designed to ac-
complish the process of generating the training samples X  Zi z i

f X ! 1; 1 2
Zi zi
27
automatically. 2

Table 4 Throughput time for


highly random produce-to-order Run number Throughput time (minutes)
and produce-to-order
Highly random produce-to-order Produce-to-order

Best rule Measure Best rule Measure

1 EDD 21,545 DS 22,309


2 DS 21,140 EDD 18,521
3 CR 29,934 EDD 28,184
4 CR 29,823 DS 17,173
5 CR 22,705 EDD 20,676
6 CR 22,330 EDD 20,209
7 DS 25,311 CR 14,946
8 CR 19,367 CR 18,565
9 CR 28,818 EDD 9,361
10 CR 25,517 EDD 16,739
11 EDD 14,402 CR 16,434
12 DS 22,321 DS 14,662
13 DS 16,879 EDD 16,596
14 DS 27,960 DS 18,380
15 CR 14,875 EDD 19,317
16 EDD 29,418 EDD 22,036
Int J Adv Manuf Technol (2013) 67:969981 979

Where X is the original ith attribute value and Zi and zi are Table 6 Comparison between support vector machine and best sched-
uling rule (minutes) for highly random produce-to-order and produce-
the largest and smallest values of the ith attribute.
to-order
In the second phase, the most appropriate kernel selection
decision has been made. However, the SVM performance Run number Throughput time (minutes)
depends upon the kernel functions, and in our case, we
Highly random produce-to-order Produce-to-order
chose RBF and polynomial kernels whose width () and
power (b) are chosen as two parameters to assess the per- SVM CR SVM EDD
formance of SVM. In this study, we have considered the
penalty weight C and to initialize the RBF kernel. Here, 1 31,000 29,998 25,772 23,247
by fixing the free penalty weight C to a constant and 2 27,875 27,674 21,201 20,125
changing the values of and b within large range, we have 3 28,751 28,818 24,866 22,979
established that the RBF kernel outperforms the polynomial 4 28,885 28,516 27,815 22,783
one in most of the cases. Moreover, grid search technique 5 20,562 21,201 22,322 21,855
has been implemented on the training set, where we found 6 29,645 29,823 26,824 24,073
C0100 and 00.5 achieves the better performance measure. 7 24,821 23,185 23,775 23,672
Consequently, we have performed the SVM training algo- 8 23,053 23,772 31,150 27,481
rithm using the tool MATLAB R 2010a to generate the 9 28,615 27,299 28,741 26,118
hyperplane equations and the scheduling functions. More- 10 27,347 27,106 26,672 26,091
over, we have used Visual Basic tool for scheduling to get 11 23,710 22,730 28,188 26,544
the best performance measures. 12 24,525 24,010 25,516 24,483
13 9,423 9,428 24,746 21,041
6.3 Simulation results 14 13,523 13,368 21,678 28,334
15 24,780 24,519 27,241 22,889
Since throughput time and tardiness are the most significant 16 21,842 21,281 24,746 21,103
performance measures in the FMC, the training set is acquired
under the optimization of these performance criteria. As stated
before, the performance of the SVM was compared to those of discussions are summarized in Tables 4, 5, 6, and 7. The
competing rules by two criteria: maximum weighted through- SVM-based simulation approach for production control system
put and minimum weighted tardiness. The upshots of the is compared with each one of the four combinations of the two

Table 5 Tardiness for highly


random produce-to-order and Run number Weighted tardiness (minutes)
produce-to-order with different
scheduling rules (minutes) Highly random produce-to-order Produce-to-order

Best rule Measure Best rule Measure

1 CR 982 EDD 0
2 DS 39,717 EDD, DS 0
3 DS 28,480 EDD 24,519
4 EDD 724 EDD 0
5 CR 751,266 EDD 52,895
6 EDD 0 DS 18,966
7 CR 52,145 CR 34,904
8 DS 15,846 EDD 12,207
9 EDD 23,654 EDD, DS 0
10 DS 82,594 DS 0
11 DS 678,542 CR 44,232
12 CR 0 DS 21,281
13 DS 0 EDD 34,856
14 DS 41,587 DS 33,766
15 EDD 22,851 ALL 0
16 EDD 38,954 EDD 0
980 Int J Adv Manuf Technol (2013) 67:969981

Table 7 Comparison between


support vector machine and best Run number Weighted tardiness (minutes)
scheduling rule (minutes) for
highly random-produce-to-order Highly random produce-to-order Produce-to-order
and produce-to-order
SVM DS SVM EDD

1 10,631 10,406 0 0
2 89,456 89,671 0 0
3 11,241 11,264 31,754 33,868
4 36,173 37,511 0 0
5 47,354 47,283 83,420 86,373
6 0 0 13,112 14,211
7 12,279 12,412 0 0
8 11,546 13,045 9,800 10,569
9 11,462 11,395 0 0
10 64,187 62,578 0 0
11 82,159 81,476 10,111 10,527
12 947,620 955,521 542,181 595,484
13 0 0 87,875 105,502
14 16,879 16,999 41,442 42,388
15 898 852 0 0
16 1,752,137 1,762,588 0 0

environments and the two performance measures to the best highly random-produce-to-order. We have also compared
single rule on average, and the results are summarized in the tardiness results between SVM and best dispatching rule
Tables 4 and 5. EDD for random produce-to-order. SVM outperformed sin-
The best rule column represents the dispatching rule gle dispatching rules in both cases.
which performs better out of the different rules used in The ratio of mean of performance measure (i.e.,
scheduling such as EDD, CR, and DS. Further, the average throughput and tardiness) for best scheduling rule and
best dispatching rule has been compared with the SVM mean of performance measure for SVM has been shown
under two environments. We compared the throughput in column 5 (i.e., performance compared to SVM) of
results between the SVM and the best dispatching rule under Table 8. Lower values of the ratio are better for tardiness
two environments (highly random produce-to-order and ran- measure, and higher values of ratio are better for
dom produce-to-order) in Table 6. From this table, it is found throughput measure for both kinds of environment. From
that SVM outperformed the best dispatching rule CR, for Table 8, it is observed that the SVM outperforms best
highly random produce-to-order. It is also established that scheduling rule-based production control system in re-
SVM is performing better than best dispatching rule EDD, spect of both performance measures. Statistical analysis
for random produce-to-order environment. of results (t test) has also been carried out to determine
Similarly in Table 7, we have compared the tardiness the significance of differences between the performances
results between SVM and best dispatching rule DS for of SVM and the best scheduling rule.

Table 8 P analysis and performance measure relative to SVM

Environment Performance measure Performance for SVM Best scheduling rule P value

Rule Performance compared to SVM

Highly random produce-to-order Throughput time 1 CR 0.987 0.865


Weighted tardiness 1 DS 1.006 0.994
Produce-to-order Throughput time 1 EDD 0.937 0.059
Weighted tardiness 1 EDD 1.078 0.918

Null hypothesis (H0)0there is no significant difference, alternate hypothesis (H1)0there is significant difference, significance level ()00.05
Int J Adv Manuf Technol (2013) 67:969981 981

7 Conclusion and future research 2. Arzi Y, Iaroslavitz L (2000) Operating an FMC by a decision-tree-
based adaptive production control system. Int J Prod Res 38
(3):675697
This paper presents an SVM-based approach to resolve a 3. Arzi Y, Iaroslavitz L (1999) Neural network-based adaptive pro-
problem in adaptive production control system of a FMC duction control system for a flexible manufacturing cell under a
which has never been investigated. The FMC operates under random environment. IIE Trans 31(3):217230
two random environments such as produce-to-order and 4. Slomp J, Gaalman GJC, Nawijn VM (1988) Quasi online sched-
uling procedures for flexible manufacturing systems. Int J Prod
highly random produce-to-order, uniquely characterized by Res 26(4):585598
their demand. We have simulated the FMC by mapping the 5. Hutchinson J, Leoong K, Snyder D, Ward P (1991) Scheduling
links between systems state variables and scheduling rules approaches for random job shop flexible manufacturing systems.
to get the best performance measures. It was demonstrated Int J Prod Res 29(5):10531067
6. Lin GY, Solberg JJ (1991) Integrated shop floor control using
that the SVM-based production control system could be autonomous agents. IIE Trans 24(3):5771
successfully used in controlling of random produce-to- 7. Shaw MJ, Park S, Raman N (1992) Intelligent scheduling with
order and highly random produce-to-order environments. machine learning capabilities: the induction of scheduling knowl-
We have used Visual Basic code for simulation due to its edge. IIE Trans 24(2):156169
8. Montazeri M, VanWassenhove LN (1991) Analysis of scheduling
flexibility and graphical user interface where manual vali- for an FMS. Int J Prod Res 28(4):785802
dation becomes possible at the initial phase of scheduling. 9. Arzi Y, David R (1998) Dispatching in a workshop belonging to a
SVM-based scheduling scheme was adopted for finding re-entrant production line under sequence dependent set-up times.
the best performance measures and its ability to yield ap- Prod Plan Contr 9(7):690699
10. Piramuthu S, Raman N, Shaw MJ (1994) Learning-based sched-
propriate dispatching rules based on the real-time system uling in a flexible manufacturing flow line. IEEE Trans Eng
attributes. However, the optimum performance of the sup- Manage 41(2):172182
port vector machine can reduce the variance of the total 11. Shinichi N, Taketoshi Y (1992) Dynamic scheduling system uti-
throughput and the time needed for simulation. Moreover, lizing matching learning as a knowledge acquisition tool. Int J Prod
Res 30(2):411431
we have compared the SVM-based approach for production 12. Arzi Y, Roll Y (1993) Dispatching procedures for a flexible man-
control system with each of the competing rules. The results ufacturing cell in constant production circumstance. Int J Oper
show that SVM performs better and faster than the other Prod Manage 13(2):3551
scheduling scheme. One can try with some more combina- 13. Babak S, Madhavi I, Solimanpur M (2010) Development of a
simulation-based intelligent decision support system for the adap-
tions of kernel function for checking the effectiveness and tive real-time control of flexible manufacturing systems. J Soft Eng
efficiency of different performance measures. As an exten- and App 3(7):661673
sion of this research, we plan to apply the proposed ap- 14. Zhibin J, Richard FYK (2003) An adaptive agile manufacturing
proach for more complex scenarios, and new variants of control infrastructure based on TOPNs-CS modeling. Int J Adv
Manuf Technol 22(34):191215
SVM can be tried to ensure the difference in performance 15. Ottaway TA, Burns JR (2000) An adaptive production control
measures. system utilizing agent technology. Int J Prod Res 38(4):721737
16. Li W, Luo X, Tu Y (2011) A heuristic for adaptive production
scheduling and control in flowshop production. Int J Prod Res 49
Acknowledgments The authors are very grateful to the Editor and (11):31513170
two anonymous reviewers for their critical comments and suggestions 17. Yeou R, Ruey S (2006) The optimization of attribute selection in
on the earlier version of our paper that helped us to modify the same so decision tree-based production control systems. Int J Adv Manuf
that it becomes more comprehensive and lucid. Technol 28(7):737746
18. Shiue YR, Su CT (2002) An enhanced knowledge representation
for decision-tree based learning adaptive scheduling. Int J Adv
Manu Technol 16(1):4860
19. Corts C, Vapnik VN (1995) Support vector networks. Mach Learn
References 20:273297
20. Buzacott JA, Yao DD (1986) Flexible manufacturing systems: a
review of analytical models. Manag Sci 32(7):890905
1. Stecke KE, Solberg JJ (1981) Loading and control policies for a 21. Friedman J (1996) Another approach to polychotomous classifica-
flexible manufacturing system. Int J Prod Res 19(5):481490 tion, Technical report, Stanford University

You might also like