You are on page 1of 8

Expert Systems with Applications 36 (2009) 44234430

Contents lists available at ScienceDirect

Expert Systems with Applications


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

A decision support system for the production control of a semiconductor packaging assembly line
P.Y. Mok *
Institute of Textiles and Clothing, The Hong Kong Polytechnic University Hunghom, Hong Kong

a r t i c l e

i n f o

a b s t r a c t
In this paper, production control of an actual industrial semiconductor packaging assembly line is studied. An evolutionary optimization methodology is developed to synthesize decision support systems for providing both crisp-logic and fuzzy-logic production control for the assembly line. The performance of such decision support systems is compared with that of the system originally implemented by the manufacturer. It is thus demonstrated that the proposed evolutionary methodology is effective in synthesizing decision support systems for both crisp-logic and fuzzy-logic control of the semiconductor packaging assembly production. It indicates that the proposed methodology can also synthesize decision support systems for productivity improvement in other complex manufacturing environment. 2008 Elsevier Ltd. All rights reserved.

Keywords: Production control Semiconductor packaging Evolutionary computation Fuzzy-logic control

1. Introduction The development of effective and efcient manufacturing systems is becoming increasingly necessary in order to cope with cut-throat competition in the modern industrial environment. It is therefore imperative that manufacturing systems be controlled optimally so as to minimize production costs and to satisfy exogenous demands in a timely fashion. However, true optimal production control may not be practically feasible because there is seldom perfect system on earth. Moreover, the actual manufacturing environment like that in apparel production or semiconductor assembly is usually very complex and inevitably involves uncertainties. Machine failures, setup changes, and variations in demand and supply, as discussed in the following paragraphs, are the major uncertainties against which manufacturing is always ghting (Gershwin, 1994; Sharifnia, 1995). Buffer storage is often placed between machines to prevent or alleviated negative effects of machine failures. Hence, when the downstream machine fails, the upstream machine can continue working if the buffer is not full; or when the upstream fails, the downstream machine can continue working if the buffer is not empty. However, the challenge of production control is to decide what material storage (buffer size) should be allowed. In more complicated manufacturing systems, where there are many product-types and alternative production paths, a failure can be mitigated by using the remaining capacity for the production of products that do not require the machine that is down. Nevertheless, it is not trivial to answer how production should be shifted
* Tel.: +852 2766 4442; fax: +852 2773 1432. E-mail address: tracy.mok@inet.polyu.edu.hk 0957-4174/$ - see front matter 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2008.05.021

when a failure occurs, so that the operational machines are well utilized without making unnecessary inventory. In cases where machines change over from producing one product-type to producing another, some time delay or cost must be incurred for setting up or changing tools. If setup changes occur too often, excessive production time is lost and costs are incurred. The consequence is that costs go up and a factory may not be able to meet demands protably. On the other hand, if setup changes occur too infrequently, inventory and delays increase with the consequence that other costs go up. In such case, the production control challenge should be how to schedule setup changes so as to minimize the cost incurred. Manufacturing converts raw material into nished products. The input to the system is material and the output from the system needs to satisfy customer demand. However, the customer demand usually varies with time and the material supply to the system may not be reliable. It is thus important to decide what adjusting action to take in order to deal with the variations in demand and supply. In addition, due to the trend of globalization in recent years, the actual physical systems tend to increase in size. Taking the case of semiconductor fabrication as an example, dozens of product-types are produced simultaneously by hundreds of workers on dozens of machines in a factory. Each product-type follows a predened process that consists of hundreds of operations. Machines are subject to random failures, and need setup changes for different producttypes. Maintenance and rework must be considered. Workers are absent at random. Because of the difculties of the problem, research on production control remained less fruitful until late 80s. Kimemia and Gershwin (1983) published his seminal work and introduced an important concept of hedging point control. In such control policy,

4424

P.Y. Mok / Expert Systems with Applications 36 (2009) 44234430

Nomenclature uj(t) production-rate of the jth machine at time t processing time for each IC strip on the jth machine cumulative output of the production line at time t the level of the jth buffer which is located between the jth machine (upstream) and the j+1th machine (downstream) at time t upper limit of the jth buffer level lower limit of the jth buffer level nished-product inventory level at time t constant demand rate for IC strips over a nite planning horizon T D(T) production target over a planning horizon T machine-condition variable denoting the machine condition at time t, which is such that a(t) = 1 if the machine is operative and a(t) = 0 if it is under repair total production cost of the production line over a planning horizon T efciency of the production line work-in-progress in the production line batch throughput time of the production line fuzzy rule; fuzzy set;

sj
W(t) bj(t)

a(t)
J E WIP TPT FR e A

Bupper j Blower j x(t) d

production is controlled such that manufacturing systems maintain inventories at specied optimal threshold values (i.e., the so-call hedging points) to compensate for intrinsic system uncertainties. The implementation of such optimal production-control rules clearly requires that the actual optimal hedging points be available in each particular application. However, the determination of such hedging points needs not only to consider the complicated system uncertainty statistical characteristics, but also to solve intricate non-linear partial differential equations representing productionrate and inventory cost simultaneously. Since the work of Kimemia and Gershwin (1983), different control theories including dynamic programming (Akella & Kumar, 1986; Buzacott & Shanthikumar, 1980; Sethi & Zhang, 1999; Sethi, Zhang, & Zhang, 1998), perturbation analysis (Bremaud, Malhame, & Massoulie, 1997; Caramanis & Liberopoulos, 1992; Ho, 1987; Ho, Eyler, & Chien, 1979; Liberopoulos & Caramanis, 1995; Veatch & Caramanis, 1999), queuing theory (Chen & Yao, 1992; Hu, 1995; Hu & Xiang, 1993), Pontryagin maximum and minimum principles (Boukas, Yang, & Zhang, 1995; Egilmez & Sharifnia, 1994; Sousa & Pereira, 1994) and other numerical techniques (Caramanis & Sharifnia, 1991; Wu & Perkins, 1996) have been extensively used to solve the production-control problems analytically or numerically. It is proved optimal control is analytically intractable for complex systems. In fact, complete results have been obtained only for single-machine single-product-type manufacturing systems (Akella & Kumar, 1986; Bielecki & Kumar, 1988). In the case of multiple-product-type systems, optimal results have been obtained only for special cases (Srivatsan & Dallery, 1998; Sethi & Zhang, 1999) with rather restrictive assumptions regarding the production priorities for different product-types. In the case of manufacturing systems with multiple machines and inter-station buffers, only heuristic results are available (Yan, Yin, & Lou, 1994; Yan, Zhou, & Yin, 1999). These results also apply only to manufacturing systems subjected to constant demands for innite time periods. In addition, all the machine failure and repair times are assumed to be exponentially-distributed random variables even though such exponential distributions may not be good models of actual manufacturing behavior (Yan et al., 1994). Moreover, these results involve intensive computation and are rather complex for implementation in industrial practice. However, Mok and Porter (2006) recently proposed a general evolutionary optimization methodology that can be used to estimate the optimal hedging points for manufacturing systems. This methodology estimates the optimal hedging points by using evolutionary algorithms, and is not restricted to the previously mentioned special classes of manufacturing system for which theoretical results are available. Indeed, this evolutionary methodology has produced some promising results in the cases of singlemachine single-product-type systems (Porter & Mok, 1999, 2000a), of multiple-product-type systems (Porter & Mok, 2000b; Mok &

Porter, 2006), and of multiple-machine systems with inter-station buffers (Mok, 2001). It has also been shown in Mok (2001) that the applicability of the evolutionary methodology is not restricted to exponentially-distributed machine failure and repair times. In this paper, it is shown through a case study of an actual semiconductor packaging assembly line that the evolutionary optimization methodology can synthesize decision support systems (DSSs) employing in addition to the previously considered hedgingpoint production-control policies different kinds of parametric crisp-logic production-control policies (such as maximal production-rate control policies). Furthermore, evolutionary algorithms are also shown to be effective in synthesizing DSS for fuzzy-logic control for this semiconductor packaging assembly line. The assembly line investigated in this paper was selected for study from a well-known international semiconductor manufacturers production plant in Hong Kong. The manufacturer prefers to remain anonymous so that all the system data presented in this paper are for demonstrative purpose only and have therefore been modied to protect the condentiality of the manufacturers information. The performance of DSSs synthesized by the proposed evolutionary methodology for both crisp-logic and fuzzy-logic control is compared with that of the system originally implemented by the manufacturer. The manufacturer found the results obtained useful, and the DSSs for crisp-logic control are immediately applicable to their existing production systems with promising productivity improvement. The rest of the paper is organized in the following way. In Section 2, the system conguration, dynamics, and control strategies of the packaging assembly line are rst described. The performance measures for DSSs are given in Section 3. Then, the development of DSS for crisp-logic control using evolutionary optimization procedure is described in Section 4. This followed by the explanation of an evolutionary rule-induction methodology for synthesizing a DSS for fuzzy-logic control in Section 5. Finally, a performance comparison is provided in Section 6, where the crisp-logic control system originally adopted by the manufacturer is compared with the evolutionarily synthesized DSSs.

2. System description 2.1. System conguration Semiconductor manufacturing includes many operations such as wafer fabrication, IC packaging process, etc., which are classied into different levels, namely, zeroth-level of wafer fabrication, rst-level of IC packaging, second-level of printed circuit board assembly, and third-level of nal assembly of circuit boards. The selected line is responsible for part of the front-end processes in

P.Y. Mok / Expert Systems with Applications 36 (2009) 44234430

4425

Fig. 1. Strips of lead frame.

the rst-level packaging, including operations of die bonding, wire bonding, moulding, deashing, cleaning, and tin-plating. IC chips are assembled on strip-formed lead frames as shown in Fig. 1. Each single strip occupies 40 IC chip sets. The production output is measured by the number of strips that have been processed by the assembly line in a specic period of time. The selected line consists of six machines in total: two die-bond machines, DB 1 and DB 2, where dies are attached to the strips; two wire-bond machines, WB 1 and WB 2, where the dies and strip-formed lead frame are wired; one moulding machine, MD, which is responsible for the moulding procedure; and one tin-plating line, TP, where the deashing, cleaning and tin-plating processes are completed at different stages of the plating line. These six machines are connected in series, and each single strip must be processed by all these machines. Strips are temporarily stored in cassette holders and transferred between machines in the line. Therefore, such cassette holders
Table 1 Machine production-rates j 1 2 3 4 5 6 Machine DB 1 DB 2 WB 1 WB 2 MD TP Production-rate u 6 strips/min 6 strips/min 6 strips/min 6 strips/min 66.75 strips/min 6.25 strips/min Processing time s 0.1667 min/strip 0.1667 min/strip 0.1667 min/strip 0.1667 min/strip 0.1480.167 min/strip 0.16 min/strip

can be viewed as buffers between machines. It is noted that there are many possible setup for semiconductor packaging assembly line, while the manufacturer assumes a unique lean line or modular approach that different machines are grouped together to form serial processing and minimize material handling between operations. There are dozens of such lean lines with different congurations on the manufacturers shopoor, and the selected line presented in the paper (Fig. 1) is one of the examples to demonstrate the proposed methodology. The die-bond and wire-bond machines in the selected line have constant production-rates, u(t). The processing time of each strip, s, on these machines are constants. The moulding machine has two possible production-rates. The production-rate of the tin-plating line is adjustable between 0 and maximum rate by controlling the electric current for plating. The production-rates for these six machines are listed in Table 1. The machines in this assembly line are subject to both major and minor random failures. The associated major and minor machine failure and repair times follow various random distributions, as indicated in Table 2. These statistical distribution characteristics were obtained by analyzing the actual historical records of machine failure and repair times using the ARENA software package (Arena URL). The packaging assembly line can produce different types of IC chips; but the same kinds of chips are grouped into lots before processing: when the assembly line nishes a lot of chips, the line changes lot. In this sense, a random setup time is initiated when changing lots. In addition to the setup time for changing lots, the die-bond machines also experience random setup disruptions for changing wafers. It is noted that die separation of wafers is completed before they are supplied to the die-bond machines for die attaching. Setup times are random variables following different distributions as listed in Table 3. In addition to random setup durations, the intervals between setup also vary because lot size and wafer size are not constants but variables as summarized in Table 4. These statistical distribution characteristics for lot sizes, wafer sizes, and setup times were obtained by analyzing the historical records using again the ARENA software (Arena URL). Since the six machines are subjected to different kinds of random failures, it is desired to keep appropriate amounts of strips as work-in-progress buffers in the line so as to compensate for the disruptive effects of machine failures and setups. In fact, no physical buffers exist in the assembly line. However, strips are temporarily stored in cassette holders and transferred between machines for processing. These cassette holders thus serve as buffers with adjustable sizes in the assembly line. There are accordingly ve work-in-progress buffers and one nishedproduct inventory store in the line. Fig. 2 shows the schematic conguration of the assembly line. In this line, the nished prod-

Table 2 Distributions of machine failure and repair times j Machine Major failure and repairs (h) Up times 1 2 3 4 5 6
a b c d

Minor failure and repairs (min) Down times Logn(0.387,0.238)a Logn(0.387,0.238)a Logn(0.519,0.317)a Logn(0.519,0.317)a 0.22 + Weib(0.174,2.0)b 0.14 + 0.41 Beta(2.23,2.03)d Up times Logn(60,50)a Logn(60,50)a Logn(60,50)a Logn(60,50)a Logn(60,50)a N/A Down times Tria(1,2,5)c Tria(1,2,5)c Tria(1,2,5)c Tria(1,2,5)c Tria(1,2,5)c N/A

DB 1 DB 2 WB 1 WB 2 MD TP

6 + Logn(13.2,21.5)a 6 + Logn(22.6,42.0)a 4 + Weib(17.6,0.75)b 6 + Logn(18.8,33.4)a 4 + Logn(2.69,2.67)a 7 + Weib(12.2,0.82)b

Logn(x, y) is the lognormal distribution with mean, x, and standard deviation, y, of the return lognormal random variable. Weib(x, y) is the weibull distribution with scale parameter b, x, and shape parameter a, y. Tria(x,y,z) is the triangular distribution of minimum value, x, modal value, y, and maximum value, z. Beta(x, y) is the beta distribution with shape parameter a1, x, and a2, y

4426 Table 3 Distributions of random setup times Machine DB 1 & DB 2 DB 1 & DB 2 WB 1 & WB 2
e

P.Y. Mok / Expert Systems with Applications 36 (2009) 44234430

Disruption Change lot Change wafer Change lot

Setup times (min) 6 + Expo(15.8)e 6 + Weib(14.5,1.03)b Tria(2,4,7)c

The nished-product inventory, x(t), is the difference between the cumulative production output and the cumulative demand at time t, i.e.,

xt Wt Dt:

Expo(x) is the exponential distribution of random variable with mean, x.

Table 4 Distributions of random lot sizes and wafer sizes Distribution (strips) Lot size Wafer size 34583 + Logn(18.0, 21.0)a Disc(0.18,Tria(225,1225,1500), 0.5, Tria(1500,1850,2100), 0.75, Tria(2100,2825,2827.5), 1.0, Tria(2827.5,3875,4250))f For DB and WB For WB

The objective of production control is to meet the production target, D(T), as closely as possible at the end of the time period, T, or equivalently, to make x(T) as close as possible to zero. The instantaneous production-rate must satisfy the capacity constraints

sj uj t 6 aj t
and

5a

uj t P 0

5b

f Disc(prob1, value1, prob2, value2, . . ..) is an empirical discrete distribution, where probi is a cumulative probability, and valuei is the value associated with probi.

for each individual machine j (j = 1, 2, . . ., 6), where sj is the processing time of each strip on the jth machine as dened in Table 1. In Eq. (5a), aj(t) (j = 1, 2, . . ., 6) is the machine condition variable such that

aj t

0; if the jth machine is down; 1; if the jth machine is up:

ucts are strips that have completed the tin-plating process and would then be transferred to another production line for further processing. 2.2. System dynamics The assembly line operates continually 24 h a day and 7 days a week. The production target,

The level of the jth buffer, which is located between the jth machine and the (j + 1)th machine, is bj(t). The system dynamics of this packaging assembly line are then governed by the following ow differential equations:

_ bj t uj t uj1 tj 1; 2; . . . ; 5;
and

7a

DT d T;

_ xt u6 t d:

7b

is the total quantity of strips required in the production period of one week (i.e., T = 7 24 60 = 10,080 min), where d is the production demand rate. If the production objective of the assembly line is to produce as much as possible, then

d umax
j

It implies that the change of buffer level is the difference between the production-rate of upstream machine and the consumption rate of downstream machine. Similarly, the change of nished production inventory level is obtained by deducing the demand rate of downstream production line from the output rate of the sixth machine, i.e., tin-plating line. 2.3. Control strategy The manufacturer regulates the assembly operations by a maximal production-rate control policy when machines are not under repair: if the downstream buffer levels do not exceed the corresponding upper buffer limits, and the upstream buffer levels (if they exist) are not below the corresponding lower buffer limits,

assuming the jth machine is the bottleneck machine in the line with maximum production-rate umaxj , as listed in Table 1. If W(t) is the cumulative production output, i.e., the total quantity of strips produced at time t, then

Wt

Z
0

u6 tdt:

M/c DB 1

Buffer DD

M/c DB 2

Buffer DW

M/c WB 1

u1(t)

q1(t)=b1(t)

u2(t)

q2(t)=b2(t)

u3(t)
Buffer WW

q3(t)=b3(t)
Finishedproduct inventory

M/c TP

Buffer MT

M/c MD

Buffer WM

M/c WB 2

x(t)=b6(t)

u6(t)

q5(t)=b5(t)
Flow of strips

u5(t)

q4(t)=b4(t)
Input of wafer

u4(t)

Fig. 2. Schematic conguration of packaging assembly line.

P.Y. Mok / Expert Systems with Applications 36 (2009) 44234430 Table 5 Original buffer level settings Buffer j Buffer limits (strips) Blower DD DW WW WM MT 1 2 3 4 5 12.5 12.5 12.5 75 75
j

4427

Bupper 31.25 31.25 31.25 175 300

then the machines operate at their full capacities; otherwise, the machines stop producing. It is obvious that no production is possible when machines are under repair. In this control policy, buffers have both upper limits, Bupper, and lower limits, Blower. Thus, according to the maximal production-rate policy, the productionrates are controlled as follows:

8 > u2 ta1 t; if b1 t Bupper1 ; > < u1 t umax a1 t; if b1 t < Bupper1 ; > 1 > : 0; otherwise: 8 > uj1 taj t; if bj t Bupperj > > > > > and bj1 t > Blowerj ; > > < if bj t < Bupperj uj t umax aj t; > j > 26j65 > > > and bj1 t > Blowerj ; > > > : 0; otherwise: ( umax a6 t; if b5 t > Blower5 ; 6 u6 t 0; otherwise:

8a

where T = 10,080 min for a time horizon of one week, and x+(t) = max(x(t), 0) and x(t)= max(x(t), 0) are the respective surplus and backlog of nished product inventory at time t. In addition, hj (j = 1, 2, . . ., 5), g+ and g are the cost weightings for work-in-progress (buffer contents), nished-product surplus, and nishedproduct backlog, respectively. The cost weightings indicate the managements preferences on keeping WIP at different stages along the production line, and in turn reveal the companys inventory policy. Traditionally, WIP at downstream processes is less preferable to that at upstream processes because WIP at downstream are of higher value and inventory hides quality problems. The later the quality problem is identied, the more it costs. However, in this paper, since the manufacturer assumed a special line conguration so that material will come out, once enters, from the line only after nishing all six machine operations. Therefore, all WIPs along the line have equal value, and thus the manufacturer decided hj = 1 for j = 1, 2, . . ., 5. The manufacturer hopes to keep similar amount of nished-product inventory and WIP at different stages of the line, but try to avoid nished-product backlog. Therefore, cost weightings for nished-product surplus and backlog are set to g+ = 1 and g = 10 according to managements preference. In addition to reducing the production cost, it is also desirable to maximize the efciency of the assembly line. This efciency is usually dened as

8b

WT WT 100% 100% W desirable T dT

10

8c

Because the nished-product inventory does not have a lower buffer limit, the sixth machine (the tin-plating line) produces at its full capacity provided that the upstream buffer does not reach its lower limit. With such maximal production-rate policy (8), the governed machine production-rate can automatically satisfy the capacity constraints (5). Crisp-logic DSS employing control rule (8) is designed to regulate the assembly operations. The design of such DSSs is done by dening the buffer limits, Bupper and Blower. The buffer lower limits, Blower, are essential for material handling (cassette transportation) between machines, and therefore their values are xed constants. However, the values of the upper buffer limits, Bupper, signicantly inuence the performance of the assembly line. The manufacturer originally dened upper and lower limits of the ve buffers are shown in Table 5. The manufacturer rst dened an inventory policy that suits the company strategy, and based upon such policy, the numerical values of buffer limits as shown in Table 5 are obtained by trial and error. In this paper, it is shown that the optimal values of the buffer limits can be estimated by evolutionary optimisation procedure proposed (Mok, 2001; Mok & Porter, 2006). 3. Performance measures The effectiveness of different DSSs in regulating the packaging assembly process can be assessed in various ways. However, one of the most important ways is by evaluating the production cost of the assembly line when under the control of each particular DSS. The lower the production cost, the better is the DSS performance. The weekly production cost in this case is dened as

and is of primary importance: it is the actual production output of IC chips expressed as a percentage of the maximum achievable production output if absence of uncertainty (i.e., the production target, D(T), with reference to Eqs. (1) and (2)). The higher the efciency, the better is the DSS performance. The average work-in-progress, WIPave, is also of interest. If WIPtotal is the number of strips that enter the system minus the number of strips that leave the system during the time period, T, then

WIPtotal
and

5 XZ j1 0

bj tdt

11a

WIPave

WIPtotal : T

11b

The smaller the WIPave, the better is the DSS performance. It is nally desirable to minimize the batch throughput time, which is the time for the complete movement of a batch through the assembly line. In this example, the manufacturer evaluates the throughput time for a batch of 950 strips by the formula

TPTbatch t strip950 ;out t strip1 ;in ;

12

Z
0

"

5 X j1

# hj bj t g x t g x t dt

where tstrip950 ;out is the time when the 950th strip leaves the assembly line, and t strip1 ;in is the time when the 1st strip enters the assembly line. The production control problem is to decide desired buffer levels so as to smooth out adverse effects caused by machine breakdown and/or setup. Keeping large quantity of WIP buffer, by raising the upper buffer limits, Bupper, in control policy (8), can continue the line operation even when presence of machine breakdown or setup, and thus improve the overall assembly line efciency, E, but it raises up the inventory cost, J, and average work-in-progress, WIPave, and prolong the throughput time, TPTbatch. It is a trade-off among different inventory control objectives. Therefore, the ultimate objective of the production control of this assembly line is to minimize the total cost

4428

P.Y. Mok / Expert Systems with Applications 36 (2009) 44234430

C q1 J

q2
E

q3 WIPave q4 TPTbatch ;

13

Table 6 Evolutionarily synthesized upper buffer settings Buffer j Buffer limits (strips) Blower DD DW WW WM MT 1 2 3 4 5 12.5 12.5 12.5 75 75
j

where q1, q2, q3, and q4 are the weighting parameters of the four performance measures. The manufacturer carefully decides the values of these weighting parameters in order to achieve a desired balance among inventory cost, efciency, average work-in-progress contents, and batch throughput time. In this case, the manufacturer prefers the weighting parameters are set so that the four performance measures are of equal importance. 4. Evolutionary synthesis of DSS for crisp-logic control It was mentioned previously that the selection of the upper buffer limits directly affects the production cost, J, and also affects the other performance measures indirectly. However, using the evolutionary optimization procedure (Mok, 2001; Mok & Porter, 2006), the optimal upper buffer limits, which are analytically intractable, can be estimated without difculty. Indeed, in the evolutionary optimization procedure, these upper buffer limits, Bupper = [Bupper1, Bupper2, Bupper3, Bupper4, Bupper5], in the maximal production-rate policies (8) can be readily represented by the string of binary substrings illustrated in Fig. 3. It is assumed that the production time period of one week, T = 10,080 min is long enough for the effect of different machinecondition vectors on the production cost, J, to be ignored and for the initial buffer states to be unimportant. A machine-condition vector, a(t) = [a1(t),a2(t),a3(t), a4(t),a5(t),a6(t)], represents an entire set of failure and repair histories of the six machines in the line. The times for the transitions between machines ups and downs follow the various random distributions dened in Table 2. Then, for each candidate upper buffer limit vector, Bupper = [Bupper1, Bupper2, Bupper3, Bupper4, Bupper5], the cost function, J, the efciency, E, the average work-in-progress, WIPave, and the batch throughput time, TPTbatch, are obtained. These quantities are determined for a particular machine-condition vector, a (t), and also for a particular set of disruptions caused by setups with stochastic characteristics dened in Tables 3 and 4, and when the line has initial states equal to twice the lower buffer limits, i.e., b(0) = 2 Blower. The total cost, C, associated with this candidate upper buffer limit vector is then calculated by using Eq. (13). The evolutionary procedure for the optimisation of buffer levels represents the elements of each candidate vector of upper buffer limits, Bupper, in the control rules (8) as the previously mentioned binary substrings shown in Fig. 3. These substrings are concatenated to form a complete binary string whose Darwinian tness, U, is inversely proportional the total cost, i.e.,

Bupper 62.5 48.2 62.5 287.9 162.1

combined population of 50 parents and 50 offspring. This evolutionary process is repeated, and thus yields successive generations of binary strings with increasing mean tness. The evolutionary process is terminated when there is little further increase in the tness of the ttest binary string. The ttest binary string is then decoded to provide the best upper limit settings for the ve buffers. Fig. 5 illustrates the general structure of such (l + k)-evolution strategies. Table 6 lists the evolutionary estimates of the optimal upper buffer limits which were thus obtained after evolution over 50 generations.

Randomly generate initial population of chromosomes Evaluate each chromosome in the current population

Parents

Mutation Generate new chromosomes

Offspring

Fitness Evaluation of offspring

No

14
parents and offspring Ranking

where k > 0 is an appropriate constant. The optimal upper buffer level limits were estimated using a (50 + 50)-evolution strategy (Rechenberg, 1973; Mok, 2001) with mutation probability 0.03. The evolutionary process begins by randomly generating an initial population, which consists of 50 binary strings. This population of 50 binary strings is mutated according to the specied mutation probability to create 50 offspring in each generation. The ttest 50 binary strings are then selected from the

Selection

Meet stopping criteria?

binary binary substring 1 substring 2 Bupper 1 Bupper 2

binary binary binary substring 3 substring 4 substring 5 Bupper 3 Bupper 4 Bupper 5

Yes
Return the best chromosome as the optimal solution
Fig. 4. Flow diagram of (l + k)-evolution strategy.

Fig. 3. Binary string representation of Bupper.

P.Y. Mok / Expert Systems with Applications 36 (2009) 44234430 Table 7 Evolutionarily synthesized fuzzy-logic controllers Controller 1 Fuzzy sets Q1: [6.262, 33.752] U1: [1.111, 7.395] Q2: [16.314, 74.161] U2: [ 0.857, 6.975] Q3: [14.510, 56.604] U3: [0.667, 7.226] Q4: [96.246, 391.118] U4: [0.603, 6.997] Q5: [98.827, 336.922] U5: [0.159, 7.592] X: [68.539, 428.288] U6: [1.302, 7.144] Fuzzy rules

4429

Genotype Phenotype

binary binary substring 1 substring 2 FLC 1 FLC2

... ...

binary substring 6 FLC6

Fuzzy-set FS Part
Configuration of FLC design Fuzzy Set Input Range Fuzzy Set Output Range FR 1

Fuzzy-rule FR Part
FR 2
2

1. If q1(t) is Z, then u1(t) is PB 2. If q1(t) is PS, then u1(t) is PB 3. If q1(t) is PB, then u1(t) is PB 1. If q2(t) is Z, then u2(t) is PB 2. If q2(t) is PS, then u2(t) is PB 3. If q2(t) is PB, then u2(t) is PB 1. If q3(t) is Z, then u3(t) is PB 2. If q3(t) is PS, then u3(t) is PB 3. If q3(t) is PB, then u3(t) is PB 1. If q4(t) is Z, then u4(t) is PB 2. If q4(t) is PS, then u4(t) is PB 3. If q4(t) is PB, then u4(t) is PS 1. If q5(t) is Z, then u5(t) is PB 2. If q5(t) is PS, then u5(t) is Z 3. If q5(t) is PB, then u5(t) is PB 1. If x(t) is N, then u6(t) is PB 2. If x(t) is Z, then u6(t) is PS 3. If x(t) is P, then u6(t) is PS

FR 3

FLC j with 1 input, 1 output, and total 3 fuzzy rules (FR). e.g., If input or output has decoded FS range [a b], then the domain settings of its 3 fuzzy-sets are as follow
FS 1

FS 2

FS 3

a Fuzzy Set
FS
FS
FS

(a+b)/2 Domain Setting

b Linguistic value e.g. N=Negative Z=Zero P=Positive

The rule-part consists of 3 FR sections, the position represents the "antecedents" (if-part), the decoded value at different positions are the corresponding "consequents" (thenpart). FR k (k=1, 2, 3) decodes a output real value ck of a range [0 v], where v is the number of fuzzy-sets of the output, i.e. v=3.
ck value the output (o/p) consequent
o FS1 / p

(-Inf, a, (a+b)/2) (a, (a+b)/2, b) ((a+b)/2, b, Inf)

[0 1) [1 2) [2 3]

e.g. Z PS PB

FSo / p 2
o FS3 / p

Fig. 5. Binary representation of fuzzy-logic control system.

5. Evolutionary synthesis of DSS for fuzzy-logic control Since the various successful applications of fuzzy control in 1980s, this form of control has attracted much research attention. Fuzzy control is often good at handling complex, nonlinear, and illdened systems. In this paper, in addition to crisp-logic control, it is now of interest to use evolution strategies to synthesize a DSS to provide fuzzy-logic control of the assembly process for the semiconductor packaging line. Machines under fuzzy-logic control are no longer constrained to produce at either zero or maximum rates (as described in Section 2.2), but at any appropriate rates determined by the DSS. The DSS embody fuzzy rules of the following generic forms: e GFR1: If buffer level bj(t) is Q j , then the machine productione j j 1; 2; . . . ; 5, and rate uj(t) is U e GFR2: If inventory level x(t) is X, then the machine productione rate u6(t) is U 6 . e e e e e e e In these generic fuzzy rules, Q 1 ; Q 2 ; . . . ; Q 5 ; X ; U 1 ; U 2 ; . . . ; U 6 are fuzzy sets that can assume linguistic values such as PS (positive small), N (negative), or Z (zero) (see Fig. 4). Evolution strategies can be used to synthesize DSS by representing each complete candidate system as a binary string. Each such complete binary string is formed by concatenating six substrings, which each represents an individual fuzzy-logic controller in the system. Evolutionary algorithms have been used in recent years for optimising fuzzy logic controller design (Tong, 2002). With reference to the above generic form of fuzzy rules, the design of FLC e includes fuzzy set domain setting of both antecedents (e.g., Q j e e and X and consequents (e.g., U j ), and fuzzy rule denition. The

binary representation for each fuzzy-logic controller in this paper is composed of fuzzy set part (for the fuzzy-set domains in the antecedents and in the consequents) and fuzzy rule part which details the consequents of individual fuzzy rules. The binary representation of a DSS for fuzzy-logic control of the assembly line is illustrated in Fig. 5. The tness of each such binary string is dened by Eq. (14). In case the DSS cannot automatically regulate the production-rates in accordance with the production constraints (5a) and (5b), then the tness of the corresponding binary string is set to zero. Table 7 shows the DSS synthesized for fuzzy-logic control for the semiconductor packaging assembly line. This system was found using a (200 + 200)-evolution strategy with mutation probability 0.03 for evolution over 300 generations. 6. Performance comparison It was discussed in Section 3 that four performance measures, namely, production cost, J, efciency, E, average work-in-progress, WIPave, and batch throughput time, TPTbatch, are conventionally used to evaluate the effectiveness of DSSs for production control. It is therefore interesting to compare the performance of the evolutionarily synthesized DSSs for crisp-logic control (Table 7), for fuzzy logic control (Table 7), and the crisp-logic control system originally adopted by the manufacturer (Table 5). It is noted that the manufacturer decided the buffer limits for crisp-logic control (Table 5) based on experiences and by trial-and-error approach as any systematic approach to achieve optimal inventory control is still absent prior the proposed evolutionary optimisation methods. It is important to note that the weighting parameters, q1, q2, q3, and q4, of the four performance measures in the total cost function (13) represent the managements preference on inventory control. With the proposed evolutionary optimisation methods, once

Table 8 Performance comparison of control systems Control system Performance measures J Original crisp-logic control system Evolutionarily synthesized DSS for crisp-logic control Evolutionarily synthesized DSS for fuzzy-logic control 4.75 108 4.02 108 3.19 108 E 84.9% 87.2% 90.9% WIPave 245.4 strips 290.5 strips 664.4 strips TPTbatch 210 min 210 min 226 min 4.76 108 4.03 108 3.20 108 Total cost C

4430

P.Y. Mok / Expert Systems with Applications 36 (2009) 44234430 Buzacott, J. A., & Shanthikumar, J. G. (1980). Models for understanding exible manufacturing systems. AIIE Transactions, 12(4), 339350. Caramanis, M., & Liberopoulos, G. (1992). Perturbation analysis for the design of exible manufacturing system ow controllers. Operations Research, 40(6), 11071125 (NovemberDecember). Caramanis, M. C., & Sharifnia, A. (1991). Near optimal manufacturing ow controller design. International Journal of Flexible Manufacturing Systems, 3, 321336. Chen, H., & Yao, D. D. (1992). A uid model for systems with random disruptions. Operations Research(Supplement), S239S247. Egilmez, K. & Sharifnia, A. (1994). Optimal control of a manufacturing system based on a novel continuous-ow model with minimal WIP requirement. In Proceedings of the fourth international conference on computer integrated manufacturing and automation technology (pp. 113118). Kimemia, J., & Gershwin, S. B. (1983). An algorithm for the computer control of production in exible manufacturing systems. IIE Transactions, 15, 353362. Gershwin, S. B. (1994). Manufacturing systems engineering. Prentice Hall. Ho, Y. C. (1987). Performance evaluation and perturbation analysis of discrete event dynamic systems. IEEE Transactions on Automatic Control, AC-32(7), 563572. Ho, Y. C., Eyler, M. A., & Chien, T. T. (1979). A gradient technique for general buffer storage design in a production line. International Journal of Production Research, 17(6), 557580. Hu, J.Q. (1995). A queueing approach to manufacturing ow control models. In Proceedings of the 34th IEEE conference on decision and control (Vol. 4, pp. 3372 3375. Hu, J. Q., & Xiang, D. (1993). The queueing equivalence to a manufacturing system with failures. IEEE Transactions on Automatic Control, 38(3), 499502. Liberopoulos, G., & Caramanis, M. (1995). Dynamics and design of a class of parameterized manufacturing ow controllers. IEEE Transactions on Automatic Control, 40(6), 10181028. Mok, P. Y., & Porter, B. (2006). Evolutionary optimization of hedging points for unreliable manufacturing systems. International Journal of Advanced Manufacturing Technology, 28(12), 205214. Mok, P.Y. (2001). Evolutionary optimisation of production-control systems. PhD Thesis, The University of Hong Kong. Porter, B., & Mok, P. Y. (1999). Gain-scheduled adaptive control of unreliable manufacturing systems with variable demands. In Proceedings of the second IEEE international conference on recent advances in mechatronics (pp. 269274). Porter, B., Mok, P. Y. (2000a). Evolutionary optimisation of hedging points in the control of unreliable manufacturing systems. In Proceedings of the second CIRP international seminar on intelligent computation in manufacturing engineering (pp. 5560). Porter, B., & Mok, P. Y. (2000b). Evolutionary optimisation of hedging points for unreliable manufacturing systems with multiple products. In Proceedings of international ICSC congress on intelligent systems & applications, symposium on computational intelligence (Vol. 2, pp. 118124). Rechenberg, I. (1973). Evolutionsstrategie: Optimierung technischer Systeme Nach Prinzipien der Biologischen evolution. Frommann-Holzboog: Verlag. Sethi, S. P., & Zhang, H. Q. (1999). Average-cost optimal policies for an unreliable exible multiproduct machine. International Journal of Flexible Manufacturing Systems, 11(2), 147157. Sethi, S. P., Zhang, H. Q., & Zhang, Q. (1998). Minimum average cost production planning in stochastic manufacturing systems. Mathematical Models and Methods in Applied Sciences, 8(7), 12511276. Sharifnia, A. (1995). Optimal production control based on continuous ow models. In P. Brandimarte & A. Villa (Eds.), Optimization models and concepts in production management (pp. 153185). Gordon and Breach Science Publishers. Sousa, J. B. & Pereira, F. L. (1994). A receding horizon strategy for the hierarchical control of manufacturing systems. In Proceedings of fourth international conference on computer integrated manufacturing and automation technology (pp. 443450). Srivatsan, N., & Dallery, Y. (1998). Partial characterization of optimal hedging point policies in unreliable two-part-type manufacturing systems. Operations Research, 46(1), 3645. Tong, C. M. (2002). Evolutionary design of fuzzy-logic controllers with minimal rule sets for manufacturing systems. PhD Thesis, The University of Hong Kong. URL: ARENA software information. <http://www.arenasimulation.com>. Veatch, M. H., & Caramanis, M. C. (1999). Optimal average cost manufacturing ow controllers: Convexity and differentiability. IEEE Transactions on Automatic Control, 44(4), 779783. Wu, F. X. & Perkins, J. R. (1996). Control of unreliable manufacturing systems: A mesh of analytical and numerical techniques. In Proceedings of the Japan-USA symposium on exible automation ASME (Vol. 2, pp. 12831286. Yan, H., Yin, G., & Lou, S. (1994). Using stochastic optimization to determine threshold values for the control of unreliable manufacturing systems. Journal of Optimization Theory and Applications, 83(3), 511539. Yan, H., Zhou, X. Y., & Yin, G. (1999). Approximating an optimal production policy in a continuous ow line: Recurrence and asymptotic properties. Operations Research, 47(4), 535549.

management decides the relative importance of these four performance measures, DSSs for both crisp-logic and fuzzy-logic control of the assembly line can be readily synthesized. However, without the use of the method, the managements desired performance is not easy to achieve. Table 8 accordingly compares the four performance measures, together with the total cost dened in Eq. (13), for these three systems. It is evident from this table that the evolutionarily synthesized DSS for crisp-logic control outperforms the control system originally adopted by the manufacturer, in that it has a lower production cost, higher efciency, lower average work-in-progress, and a similar batch throughput time. It is further indicated in the table that the synthesized DSS for fuzzy-logic control provides the best overall performance as it has the lowest production cost and the highest efciency. However, the work-in-progress in the fuzzy-logic controlled assembly line is very high and is more than twice that in the crisp-logic controlled cases. Evidently, large workin-progress is essential to ensure high efciency and hedge against line uncertainties such as machine breakdowns and random setups. Although the selected assembly line involves six machine operations, the uncertainties encountered in production operation are representative. The proposed evolutionary optimisation methodologies can be readily applied to other assembly lines with different system characteristics in the factory oor. In fact, the method has been used by the manufacturer to decide buffer limits for crisp-logic inventory control, and positive results are obtained so far. 7. Conclusion In this paper, production control of an actual industrial semiconductor packaging assembly line has been studied. An evolutionary optimization methodology has been developed to synthesize DSSs for providing both crisp-logic and fuzzy-logic production control for the assembly line. The methodology was developed in order to provide practically realizable policies for the control of complex manufacturing systems. The performance of the DSSs for both crisp-logic and fuzzy-logic control using the proposed methodology has been compared with that of the system originally implemented by the manufacturer. It has thus been demonstrated that the proposed evolutionary methodology is effective, and such method is applicable to the design of DSSs for productivity improvement in complex manufacturing environment. Acknowledgement The authors are grateful to the University of Hong Kong for the provision of nancial support and computational facilities for this study. References
Akella, R., & Kumar, P. R. (1986). Optimal control of production rate in a failure prone manufacturing system. IEEE Transactions on Automatic Control, 31, 116126. Bielecki, T., & Kumar, P. R. (1988). Optimality of zero-inventory policies for unreliable manufacturing systems. Operations Research, 36, 532541. Bremaud, P., Malhame, R. P., & Massoulie, L. (1997). A manufacturing system with general stationary failure process: Stability and IPA of hedging control policies. IEEE Transactions on Automatic Control, 42(2), 155170. Boukas, E.K., Yang, H. & Zhang, Q. (1995). Minimax production planning in failure prone manufacturing systems. In Proceedings the 34th IEEE conference on decision and control (Vol. 3, pp. 31343139).

You might also like