Você está na página 1de 30

Submitted to IIE Transactions

Design of Flexible Plant Layouts

Saifallah Benjaafar and Mehdi Sheikhzadeh


Department of Mechanical Engineering
University of Minnesota
Minneapolis, MN 55455

Abstract

In this paper, we address the problem of designing flexible plant layouts for manufacturing

facilities where product demands are subject to variability. A flexible layout is one that maintains

low material handling costs despite fluctuations in the product demand levels. We extend existing

procedures for design of flexible layouts by (1) explicitly capturing the stochastic nature of

product demands and the resulting variability in material flows between different processing

departments, (2) allowing for the possibility of multiple processing departments of the same type

to exist in the same facility, and (3) letting material flows between pairs of individual departments

be determined simultaneously with the layout and as a function of demand scenarios. Optimal and

heuristic methods are presented for generating flexible layouts and determining flow allocations

under various design and operation assumptions.


1. Introduction
The ability to design and operate manufacturing facilities that can quickly and effectively

adapt to changing technological and market requirements is becoming increasingly important to the

success of any manufacturing organization. In the face of shorter product lifecycles, higher

product variety, increasingly unpredictable demand, and shorter delivery times, manufacturing

facilities dedicated to a single product line cannot be cost effective any longer. Investment

efficiency now requires that manufacturing facilities be able to shift quickly from one product line

to another without major retooling, resource reconfiguration, or replacement of equipment.

Investment efficiency also requires that manufacturing facilities be able to simultaneously make

several products so that smaller volume products can be combined in a single facility and that

fluctuations in product mixes and volumes can be more easily accommodated. In short,

manufacturing facilities must be able to exhibit high levels of flexibility and robustness despite

significant changes in their operating requirements.

Although there exists an abundant literature on manufacturing flexibility as it relates, for

example, to machines, material handling, product mix, part routing, and part sequencing [3] [27]

[31], very little of this literature deals with layout flexibility. Webster and Tyberghein [38] define

layout flexibility as the ability of a layout to respond to known and future product mixes. They

consider the most flexible layout to be the one with the lowest material handling cost over a

number of demand scenarios. Bullington and Webster [6] extend this definition to the multi-

period case and present a method for evaluating layout flexibility based on estimating the costs of

future relayouts. They recommend that these costs be used as an additional criterion in

determining the most flexible layout.

Gupta [11] presents a simulation approach for measuring layout flexibility. He obtains the

various flow matrices by random generation of flow volume instances between pairs of

departments. Using a heuristic, such as CRAFT. a layout for each generated flow matrix is

derived. For each such layout, the distance between all pairs of departments is computed. These

distances are, in turn, used to compute the average distance between departments over the set of

-2-
all generated layouts. A penalty function measuring the sum of absolute deviations from their

distance mean of all pairs of departments for a given layout is then calculated. A layout with the

smallest penalty is considered to be the most flexible layout.

Shore and Tompkins [29] also propose a penalty function as a criterion for choosing the

most flexible layout. Their penalty function measures the expected material handling inefficiency

of each layout over all possible production demand scenarios. Assuming the probability of each

demand scenario is known and the number of scenarios is finite, the layout with the least expected

inefficiency can be identified. This layout is considered to be the most flexible layout.

Rosenblatt and Lee [25] present a robustness approach to the stochastic plant layout

problem. They consider an uncertain environment in which the exact values of the probabilities of

the different possible scenarios are unknown. For such an environment, layout flexibility is

defined in terms of the robustness of the layout's performance under different scenarios. Thus,

the most flexible (robust) layout is the one whose cost performance remains close to the optimal

layout for the largest number of scenarios. A robustness approach to the single and multiple

period layout problem is also proposed by Kouvelis et al. [12].

Rosenblatt and Kropp [24] presented an optimal solution procedure for the single period

stochastic plant layout problem. They showed that their procedure only requires solving a

deterministic from-to flow matrix, where the deterministic matrix is a weighted average of all

possible flow matrices. They compared their results to the flexible layout measure developed by

Shore and Tompkins [29] and showed that their approach will always result in the most flexible

layout. The stochastic plant layout problem has also been addressed by Montreuil and Laforge

[18] and Palekar et al. [23], among others.

Recently, Drolet [8] introduced virtual cellular manufacturing systems (VCMS) as a more

flexible alternative to conventional cellular configurations for computer integrated manufacturing

systems. Instead of configuring a manufacturing facility into cells, each dedicated to a specific

part family, machines of various types are distributed throughout the shop floor and reconfigured

in real time in virtual cells in response to actual job orders. Upon completing the job order, the

-3-
virtual cell is disbanded and the associated machines are made again available to the system.

Figure 1 contrasts the differences between virtual cellular layouts, conventional cellular, and

process layouts. The author does not, however, provide a procedure for generating these layouts.

To support flexible system configurations, such as a VCMS, Montreuil et al. [20]

introduced the concept of holographic layouts as an alternative to process layouts for systems

operating in highly volatile environments. An holographic layout spreads the machines of each

type throughout the manufacturing facility. For each machine of a particular type, an attempt is

made to insure its proximity to machines of every other type so that routings that are flow efficient

can be created in real time by an intelligent shop floor control system. A heuristic design

procedure is proposed where the objective is to generate a layout such that each machine is as

centrally located, with respect to other machines of different type, as possible. The procedure

assumes, however, that no distinguishable flow patterns exist and that a maximally dispersed

layout is always desirable.

More recently, the same authors [21] [35] proposed the fractal layout as yet another

alternative for job shop environments. In a fractal layout, several almost-identical cells are

created, with each cell being allocated, whenever possible, an equal number of machines of each

type. This results in a set of flexible cells, where every product can be produced in almost every

cell. The authors propose a multi-stage design methodology for cell creation, cell layout, and cell

flow assignment. In particular, cell layouts are generated in an iterative fashion with flow

assignment decisions.

In this paper, we address the problem of designing flexible plant layouts for

manufacturing facilities where product demands and product mix composition are subject to

fluctuation. We define flexible layouts as those that can effectively cope with product demand

variability, where effectiveness is measured by expected material handling cost over the various

possible demand scenarios. A flexible layout is, thus, one that would maintain low material

handling costs despite fluctuations in the product demand levels and fluctuations in the resulting

material handling flows. In our solution method, we extend existing procedures for design of

-4-
(a) Process layout (b) Cellular layout

(c) Virtual cellular layout

Figure 1 Process layout, cellular layout, and virtual cellular layout

-5-
flexible layouts by (1) explicitly capturing the stochastic nature of product demands and the

resulting material flows between different processing departments, (2) allowing for the possibility

of multiple processing departments of the same type to exist in the same facility (e.g., two

separate machining centers not sharing a common processing location), and (3) letting material

flows between pairs of individual departments be determined simultaneously with the layout and

as a function of the demand scenarios.

This approach to layout design departs from conventional solution methods to both the

static and stochastic layout problem in several ways. Most important of which is the fact that we

account for the possibility of having multiple departments of the same type in the same facility.

This is significant since duplicating departments or disaggregating existing ones (e.g., not placing

all machines of a given type in the same plant floor area) is increasingly being recognized as an

effective mechanism for enhancing layout flexibility. By strategically locating duplicate

departments in different areas of the plant floor, a facility can hedge against future fluctuations in

job flow patterns and volumes. In fact, having alternative processing departments to which jobs

can be routed can reduce and simplify material handling requirements in a job shop even in the

absence of variability. This is evident in cellular manufacturing layouts, where distributing

individual machines throughout the manufacturing facility is found to significantly reduce material

handling effort. In such a layout, copies of the same machine type are allocated to different cells

so that jobs that are assigned to a cell can be completely processed within the cell by adjoining

machines.

Unfortunately, cellular manufacturing systems can be highly inflexible, since they are

generally designed with a fixed set of part families in mind whose demand levels are assumed to

be stable and their life cycles are considered to be sufficiently long. In fact, once a cell is formed,

it is usually dedicated to a single part family with limited allowance for intercell flows. While

such organization may be adequate when part families are clearly identifiable and demand volumes

are stable, they become inefficient in the presence of significant fluctuations in the demand of

-6-
existing products or with the frequent introduction of new ones. A discussion of the limitations of

cellular manufacturing systems can be found in [1], [2], [10], and [39].

These limitations resulted in recent calls for alternative and more flexible cellular

structures, such as overlapping cells [1], cells with machine sharing [2] [32], and virtual cellular

manufacturing systems [8]. For highly volatile environments, certain authors have even

suggested, as discussed earlier, a completely distributed layout, where copies of a given machine

type are dispersed as much as possible throughout the shop floor [20] or the creation of multiple

identical cells as in the fractal layout [21] [35].

The procedure proposed in this paper, in addition to generating cost effective layouts for

plants operating in stochastic environments, can serve as a tool for evaluating the desirability of

various types of distributed layouts and/or to assess the benefits of various degrees of

disaggregation of existing departments into smaller and geographically distributed sub-centers.

The procedure also offers, once a particular layout is selected, a method for determining optimal

flow volumes between pairs of departments and sub-departments for each demand scenario. This

provides a manufacturing facility with additional flexibility in dealing with demand variability.

We should note that letting both layout and flow volume allocation be decision variables is also

proposed by Montreuil et al. [21 [35] in the context of fractal cellular layouts.

2. Solution Procedure
The layout design procedure we describe in this section is for manufacturing facilities that

produce multiple product types whose demands may fluctuate from period to period according to a

known distribution. To illustrate the procedure, we consider the case where the demand for each

product is characterized by a finite discrete distribution, represented by a finite number of demand

realization scenarios and probabilities of occurrence of each scenario. Demand for products are

assumed to be independent. The procedure can be easily extended to continuous and correlated

demand distributions (see Appendix 1). Characterizing the demand distributions may be based on

historical data and/or forecasts. We assume the following information is known:

-7-
• the set of product types P = {p1, p2, …, pP} produced in the facility,

• the demand distribution for each product p i, as described by a set of demand realization
scenarios Di = {di1, di2, …, dini} and associated probabilities Πi = {πi1, πι2, …, πini},

• the process routing for each product type, consisting of the sequence of department types visited

by the product,

• the unit transfer load for each product between each pair of department types (i.e., size of the

transfer batch),

• number of copies of each department type,

• the set of available locations,

• the distance between each pair of locations,

• the average processing time per product per unit load of flow at both the originating and

destination departments, and

• the available operation time per department.

The basic steps of our procedure can be summarized as follows:

i) From the product demand distributions, the product process routings, and the product unit

transfer loads, we determine for each possible demand scenario the amount of product flow,
vijp, between each pair of department types i and j. This results in a multi-product from-to flow

matrix, m(s), for each demand scenario s.

ii) From the individual product demand distributions, we obtain the probability of occurrence of
each demand scenario s, π(s).

iii) For each demand scenario, we generate the corresponding optimal layout (when an optimal

layout is computationally difficult to obtain, we use the heuristic procedure outlined later) and

the corresponding optimal flow allocation between copies of the same department.

iv) Once all the layouts have been generated, we evaluate each layout over the entire set of

possible demand scenario and select the most flexible layout (for example, we choose the layout

with the lowest expected value of material handling cost over all demand scenarios).

-8-
The required calculations for generating flow matrices and their probability of occurrence

are illustrated for an example system in Tables 1-5. Note that the flow matrix is multi-dimensional

with each entry being a vector representing the individual flows from each product. It is also

important to note that the generated from-to flow matrices only give the amount of flow between

department types. The determination of the flow volume between individual departments is

determined by the optimization model, simultaneously with the layout. The number of possible

flow matrices can be large when either the number of products is high or the number of product

demand scenarios is large. An alternative to complete enumeration is to use Monte Carlo

simulation with a limited sample size (see Appendix 1).

2.1 Layout-Flow Allocation Model


For each demand scenario, the optimal layout and the corresponding optimal allocation of

flow between departments can be obtained by solving the following model:

P N N Ni Nj K K
Min z = ∑ ∑ ∑ ∑ ∑ ∑ ∑ v nim jpd klx nikx m jl
p = 1 i = 1 j = 1 ni = 1 mj = 1 k = 1 l = 1

subject to:
K
∑ x n ik = 1 ni = 1i, 2i, …, Ni; i = 1, 2, …, N (1)
k=1

N Ni
∑ ∑ x n ik = 1 k = 1, 2, …, K (2)
i = 1 ni = 1

P N Ni
∑ ∑ ∑ v n im jp t m jp ≤ c m j mj = 1j, 2j, …, Nj; j = 1, 2, …, N (3)
p = 1 i = 0 ni = 1

Ni Nj

∑ ∑ vnimjp = v ijp i, j = 0, 1, …, N; p = 1, 2, …, P (4)


ni = 1 mj = 1

N Ni N Nq

∑ ∑ v n im jp = ∑ ∑ vmjrqp m j = 1j, 2j, …, Nj; j = 1, 2, …, N; p = 1,2, …, P (5)


i = 0 ni = 1 q = 0 rq = 1

-9-
xnik = 0, 1 ni = 1i, 2i, …, Ni; k = 1, 2, …, K (6)

where:

1 if nth department of type i is assigned to location k


x n ik = {
0 otherwise
vnimjp: flow volume between nth department of type i and mth department of type j due to product

type p ,
vijp: flow volume between departments of type i and departments of type j due to product type p,

dkl: distance between location k and location l (known parameter),


tmjp: Processing time per unit load of product type p at department mj,

cni: capacity of department ni (available operation time,

Ni: number of departments of type i,

N: total number of department types,

K: total number of locations, and

P: total number of product types.

The above model solves simultaneously for department location and volume of flow

between individual departments, so that material handling costs are minimized. The decision
variables are the x n i k 's and the v n im jp 's. Constraints (1) and (2) ensure that each department is

assigned to exactly one location and each location is assigned to one department. When the

number of locations exceeds the number of departments, dummy departments with zero flows

may, without loss of optimality, be used to account for the difference. Constraint (3) ensures that

the flow volume allocated to a department does not exceed the capacity of that department.

Constraint (4) equates the amount of flow between multiple copies of departments of type i and j

to the amount of flow between department type i and department type j, as dictated by the from-to

flow matrix. Constraint (5) ensures that the amount of input and output flow (per product) to and

from a department are the same. Note that the index i = 0 is used to denote input/output

departments. This is necessary in order to capture both entering and exiting flows. However,

-10-
Table 1 Example part demands under different states

Parts Low Medium High


P1 50 100 150
P2 80 120 160
P3 150 200 250
P4 90 180 270

Table 2 Probability of occurrence for different demand states

Probability of
States occurrence
Low (L) 0.2
Medium (M) 0.5
High (H) 0.3

Table 3 Probability of occurrence for different demand combinations

Probability of
Scenarios occurrence
LLLL 0.0016
MMMM 0.0625
HHHH 0.008
LLLM 0.004
LLLH 0.0024
LLMH 0.006
• •
• •
• •

Table 4 Part Manufacturing Sequences

Parts Manufacturing sequence


P1 A→B→D→E→F
P2 B → C → A → D →Β → E
P3 B → A → C → F→ E → D
P4 C→B→D→A→F

Table 5 Multi-product Flow matrix for demand combination HHHH

A B C D E F
A (0,0,0,0) (150,0,0,0) (0,0,250,0) (0,160,0,0) (0,0,0,0) (0, 0,0,270)
B (0,0,250,0) (0,0,0,0) (0,160,0,0) (150,0,0,0) (0,160,0,0) (0,0,0,0)
C 0,160,0,270) (0,0,0,0) (0,0,0,0) (0,0,0,0) (0,0,0,0) (0,0,250,0)
D (0,0,0,270) (0,160,0,0) (0,0,0,0) (0,0,0,0) (150,0,0,0) (0,0,0,0)
E (0,0,0,0) (0,0,0,0) (0,0,0,0) (0,0,250,0) (0,0,0,0) (0,0,0,0)
F (0,0,0,0) (0,0,0,0) (0,0,0,0) (0,0,0,0) (0,0,250,0) (0,0,0,0)

-11-
this formulation is used only for modeling convenience, since we assume that a product can enter

or exit at any department.

The above model can be viewed as a variation on the classical quadratic assignment

problem (QAP) [9]. The differences between the two models are, however, important. In our

model, the objective function is polynomial, as department location and flow volume between

departments are both decision variables. This difference is due to the availability of multiple

departments of the same type. Obviously, when there is only a single copy of each department,

the model reduces to the classical QAP. As we discuss later on, the fact that flow volumes

between departments are not pre-determined allows us to optimize material handling costs even if

the layout is fixed.

The model assumes that all department copies are of the same size. In practice, this may

not always hold, especially if we consider duplicates of the same department not containing the

same numbers of machines. This problem could be addressed by disaggregating departments into

small grids with equal area and assigning artificially large flows between grids of the same

department so that they are always placed in adjoining locations. A more detailed discussion of

the general merits and limitations of this approach can be found in Liao [15], Kusiak and Heragu

[13] and Bozer and Meller [5], among others. Alternative methods for incorporating departments

of unequal size have been proposed (e.g., see Montreuil [17]). For a recent survey, the reader is

referred to Meller and Gau [16].

Finally, the model assumes that, as long as capacity constraints are not violated, there is

total flexibility in making workload assignments to departments. Such a strategy, while allowing

us to minimize material handling costs, may result in unbalanced workload distribution among

departments of the same type. In turn, this may negatively impact congestion levels and

throughput times at the more utilized departments. However, the model can easily be modified to

allow for a balanced workload assignment among all departments of the same type. For example,

exchanging constraint (3) with the following constraint:

-12-
P N Ni Nj P N Ni
∑ ∑ ∑ (vnimjptmjp)/cmj = ∑ [∑ ∑ ∑ (vnimjptmjp)/cmj]/N j , (7)
p = 1 i = 0 ni = 1 mj = 1 p = 1 i = 0 ni = 1

where mj = 1j , 2j , …, Nj and j = 1, 2, …, N, results in a solution that balances workload among

departments of the same type. Constraint (7) can be similarly modified to allow for alternative

workload assignment strategies and/or additional operating constraints.

2.2 A Heuristic Approach


Because the quadratic assignment problem has been shown elsewhere to be NP complete

[8], the model proposed here is also NP complete. This means that obtaining an optimal solution

for most problems in practice would require an excessive amount of computational effort.

Therefore, a heuristic method is provided below. The method is an extension of existing layout

heuristics, such as CRAFT [37], that use an iterative pairwise, or multi-step, exchange procedure

in generating a final layout. Our approach differs from these heuristics in that we are not only

solving for the layout but also for the flow volume allocation between departments.

Consequently, at each iteration step and for each new layout considered, the flow volume

allocation problem needs to be first solved before the layout cost can be calculated. Fortunately,

the problem can be formulated as a linear program and solved optimally in a reasonable amount of

time. The steps of the heuristic are described below:

Step 1: Set J = 1.

Step 2: Generate an initial layout.

Step 3: Solve optimally for flow volume allocation (a linear program).

Step 4: Calculate z(J), the resulting objective function value of the original layout problem.
Step 5: Set J = J + 1. If J > Jmax, go to step 9 (e.g., Jmax is the maximum number of feasible

pairwise interchanges).

Step 6: Generate the next layout (e.g., by a pairwise interchange).

Step 7: Solve optimally for flow volume allocation (a linear program).

Step 8: Calculate z(J), the resulting value for the objective function. Go back to step 5.

-13-
Step 9: Implement the minimum cost layout. If the minimum cost layout is the same as the

previous one, then go to step 10. Otherwise, set J = 1 and go back to step 5.

Step 10: Stop.

The linear program that must be solved at each iteration (steps 3 and 7) is given below:

P N N Ni Nj K K
Min z = ∑ ∑ ∑ ∑ ∑ ∑ ∑ v nim jpd klx nikx m jl
p = 1 i = 1 j = 1 ni = 1 mj = 1 k = 1 l = 1

subject to:

P N Ni
∑ ∑ ∑ v n im jp t m jp ≤ c m j mj = 1j, 2, …, Nj; j = 1, 2, …, N (1)
p = 1 i = 0 ni = 1

Ni Nj

∑ ∑ vnimjp = v ijp i, j = 0, 1, …, N; p = 1, 2, …, P (2)


ni = 1 mj = 1

N Ni N Nq

∑ ∑ v n im jp = ∑ ∑ vmjrqp m j = 1j, 2j, …, N j, j = 1, 2, …, N; p = 1,2, …, P (3)


i = 0 ni = 1 q = 0 rq = 1

Note that, in this case, the values of the variables x n ik and x m jl in the objective function are

already known. The only decision variables are therefore the flow volumes v n im jp . Once the

flow volumes between departments are known, the cost of the layout under consideration can be

calculated (steps 4 and 8).

The above heuristic, by requiring that at each iteration and for each potential pairwise

exchange the flow allocation problem be solved optimally, can be computationally demanding.

An alternative heuristic, found to be equally effective but computationally less intensive, is to

iteratively solve a layout problem with fixed flows followed by a flow volume allocation with a

fixed layout. This is similar to a heuristic used in facility location problems where both facility

location and customer allocation must be made simultaneously [7]. The heuristic can be

summarized as follows: (1) given a fixed layout, find a minimum cost flow allocation between

departments; (2) given fixed flow allocation between departments, find a minimum cost layout.

The heuristic alternates between steps (1) and (2) until convergence is achieved. An application of

-14-
the same principle to the design of fractal cellular layout can also be found in Montreuil et al. [21

[35].

The solution obtained, while not guaranteed to be optimal, satisfies the following

necessary condition for optimality [9]: (a) for the obtained layout, the solution cannot be improved

by changing flow allocations; (b) for the obtained flow allocations, the solution cannot be

improved by changing the layout. Computational experiments are inconclusive as to which of the

two heuristics performs best. However, the second heuristic appears to be computationally less

intensive.

Finally, we should note that alternatives to the pairwise exchange heuristic could be used

in layout generation. In fact, any heuristic/search method used to solve the quadratic assignment

problem can be applied. For a recent review of QAP problems and heuristics, the reader is

referred to Pardalos and Wolkowicz [22].

2.3 Flexible Layout Selection


The result of either the optimization model or the heuristic is a layout of the various

departments and an allocation of flow between these departments. Once such a layout has been

identified for every possible demand realization scenario, a decision must be made as to which of

these layouts is the most flexible. By definition, a flexible layout is one that maintains low

material handling costs despite changes in demand levels. Therefore, a possible measure of layout

flexibility is the layout's expected material handling cost over the range of feasible demand

scenarios. This is a similar definition to that proposed by Shore and Tompkins [29] and

Rosenblatt and Kropp [24]. A procedure for identifying the most flexible layout, can then be

obtained by (1) constructing the matrix z(L(I), J), where z(L(I), J) is the material handling cost

resulting from using a layout generated for demand scenario I when the actual demand

corresponds to that of scenario J, and (2) selecting the most flexible layout by choosing, for

example, the layout with the smallest expected material handling cost over all demand scenarios.

Note that in obtaining the material handling cost of operating a layout that was designed for

demand scenario I under scenario J, a linear program is solved in order to determine the optimal

-15-
flow volume allocation. The expected material handling cost of each layout, E[z(L(I)], is given

by:

E[z(L(I))] = ∑ π(J)z(L(I), J)
J∈S

where π(J) is the probability of occurrence of demand scenario J and S is the set of all feasible

scenarios.

The above definition of a flexible layout, as the one with the lowest average cost, is by no

means unique. Alternative criteria for selecting the most flexible layout can be used. For

example, a robustness criterion may, for certain facilities, be more appropriate. The layout whose

cost does not exceed the lowest cost layout by a certain percentage the most number of times is

chosen. This criterion is similar to the one used in Rosenblatt and Lee [25]. The robustness

criterion can be particularly useful when probabilities are difficult to assign to different demand

scenarios. A min-max criterion could also be used when the objective is to limit the cost of the

worst case scenario.

Both criteria, expected cost and robustness, do not, however, account for variance. A

layout with low expected cost may have high variance if its performance fluctuates significantly

from one demand scenario to another. This can be unacceptable if the material handling cost is

excessively high for one or more of these scenarios. Variance can be accounted for by using a

combined mean-variance criterion. Alternatively, certain layouts may be disqualified from further

consideration if their associated cost exceeds a certain level for one or more demand scenarios.

Finally, we should note that in identifying the most flexible layout, we have restricted in

our current procedure our attention to the set of the best layouts selected for different scenarios.

In doing so, we may have overlooked a layout with a lower overall expected material handling

cost that is not necessarily optimal for any scenario. This limitation can be overcome by directly

solving for the layout with the lowest expected material handling cost. A procedure that

implements this approach is described in Appendix 2. Extensive experimentation with both

procedures indicates that the second procedure yields layouts with slightly lower expected material

-16-
handling costs. The difference between the two procedures tends to diminish with increases in the

level of department duplication. Our original procedure is, however, more general since it allows

for alternative criteria (other than expected cost) for measuring layout flexibility to be used.

3. Computer Implementation and Example Problems


A computer software application, FLEX-LAYOUT, that implements the above flexible

layout design procedure was developed. The software application† is written in C and is

interfaced with the Math Programming software package LINDO [18]. LINDO is called when

generating a solution to the flow volume allocation problem (a linear program). The

implementation platform is a SUN Sparc 20 workstation. Several hundred problems of varying

sizes were evaluated. While no attempt was made to optimize the performance of the software at

this stage, fairly large scale problems were solved in less than 2 hours of CPU. For example, a

problem with ten department types, 35 departments, 243 demand scenarios, and 5 part types was

solved in 143 minutes of CPU. We should note that since plant layout decisions result in long

term commitments, few hours of CPU are generally affordable. For very large problems, a

simulation approach, as described in Appendix 1, becomes more appropriate.

A simple example that illustrates the various steps in generating a flexible layout is

described in Tables 6-11. The manufacturing facility consists of six different department types,

16 individual departments, and 16 department locations organized in a symmetric 4X4 matrix.

Rectilinear distances are used to measure distances between different locations. The workload

balancing constraint (constraint 7) is in effect. The processing times are assumed to be product

independent. Therefore, for a given department, the processing times are the same for all

products. There are two product types, with demand for each product type being either high (H),

medium (M), or low (L). This results in 9 possible demand scenarios. For each demand

scenario, the best layout and its corresponding cost are calculated. The results are listed in Table

10. The performance of each of these selected layouts is then evaluated under all possible

† A copy of the FLEX-LAYOUT software, along with the source code, is available upon request from the authors.

-17-
scenarios and the expected performance of each layout is calculated. The results are given in

Table 11. The layout with the lowest expected material handling cost is then selected. In this

case, the lowest expected cost layout, as shown in Figure 2, is the one initially selected for

demand scenario 1 (LL).

It is interesting to note that the performance differential between each of the selected

layouts for the different scenarios is not very significant. This is made evident by the similarity

between the different rows of Table 11. We believe that this is due to the highly distributed nature

of the generated layouts and to the existing high level of department duplication. Note that

departments of the same type are either evenly distributed throughout the layout (in case of

multiple copies) or centrally located (in case of a single copy). This high degree of department

dispersion results in layouts where it is always possible to find copies of any two departments in

relative close proximity.

Table 12 contrasts the performance of the distributed layouts we generated for each

demand scenario with equivalent functional layouts (i.e., layouts with the same number of

departments but where departments of the same type are restricted to be in adjacent locations). It

is easy to see that the distributed layouts outperform the functional layouts under all 9 demand

scenarios. More importantly, the functional layouts are less robust in the face of variability. A

functional layout that performs well under one scenario can, indeed, perform poorly under

another. This is illustrated in Table 13, where the values of E(z(L(I)) are shown for both the

functional and distributed layouts. Note that while for the distributed layout the values of

E(z(L(I)) vary very little, wider fluctuations are experienced by the functional layouts.

Extensive experimentation with both functional and distributed layouts for systems with

varying number of products and departments and for varying demand scenarios and levels of

department duplication can be found in Benjaafar and Sheikhzadeh [4]. The results largely

confirm the above observations. In particular, we found that the desirability for department

duplication and distribution tends to increase with increases in either product demand variability or

-18-
Table 6 Part demands under different states

Parts Low Medium High


P1 60 90 180
P2 80 150 180

Table 7 Probability of occurrence for different demand states

Probability of
States occurrence
Low (L) 0.2
Medium (M) 0.5
High (H) 0.3

Table 8 Part Manufacturing Sequences

Parts Manufacturing sequence


P1 C→A →E→B→D
P2 D→C→F→A→B

Table 9 Number of copies per department type

Department type Number of department


copies
A 4 (A1, A2, A3, A4)
B 3 (B1, B2, B3)
C 3 (C1, C2, C3)
D 2 (D1, D2)
E 3 (E1, E2, E3)
F 1 (F1)

-19-
Table 10 Generated layouts and material handling costs per scenario

Scenario Material
Scenario Flow matrix Selected layout
probability handling cost
VAB = 80, VAE = 60 [E1, B2, A1, E2]
LL (1) 0.04 VBD = 60, VCA = 60 [A4, E3, C3, B3] 620
VCF = 80, VDC = 80 [C1, A2, F1, A3]
[D2, B1, C2, D1]
VEB = 60, VFA = 80
VAB = 150, VAE = 60 [E1, B2, A1, E2]
LM (2) 0.10 VBD = 60, VCA = 60 [A4, E3, A3, B3] 935
VCF = 150, VDC = 150 [C1, A2, F1, C3]
[D2, B1, C2, D1]
VEB = 60, VFA = 150
VAB = 180, VAE = 60 [E1, B2, A1, E2]
LH (3) 0.06 VBD = 60, VCA = 60 [A4, E3, A3, B3] 1080
VCF = 180, VDC = 180 [C1, A2, F1, C3]
[D2, B1, C2, D1]
VEB = 60, VFA = 180
VAB = 80, VAE = 90 [E1, A2, B2, D1]
ML (4) 0.10 VBD = 90, VCA = 60 [C3, F1, A3, C1] 733
VCF = 80, VDC = 80 [B1, C2, D2, A1]
[E2, A4, B3, E3]
VEB = 80, VFA = 80
VAB = 150, VAE = 90 [D2, C3, B2, D1]
MM (5) 0.25 VBD = 90, VCA = 90 [A2, F1, A3, C1] 1050
VCF = 150, VDC = 150 [B1, C2, E1, A1]
[E2, A4, B3, E3]
VEB = 90, VFA = 150
VAB = 180, VAE = 90 [D2, C3, B2, D1]
MH (6) 0.15 VBD = 90, VCA = 90 [A2, F1, A3, C1] 1185
VCF = 180, VDC = 180 [B1, C2, E1, A1]
[E2, A4, B3, E3]
VEB = 90, VFA = 180
VAB = 80, VAE = 180 [B2, E1, A3, C2]
HL (7) 0.06 VBD = 180, VCA = 180 [E3, B3, E2, A2] 1140
VCF = 80, VDC = 80 [A1, D2, C3, F1]
[C1, D1, B1, A4]
VEB = 180, VFA = 80
VAB = 150, VAE = 180 [D1, B2, E2, C3]
HM (8) 0.15 VBD = 180, VCA = 180 [C1, A3, A1, F1] 1420
VCF = 150, VDC = 150 [A4, E3, B1, A2]
[E1, B3, D2, C2]
VEB = 180, VFA = 150
VAB = 180, VAE = 180 [D1, B2, E2, C3]
HH (9) 0.09 VBD = 180, VCA = 180 [C1, A3, A1, F1] 1560
VCF = 180, VDC = 180 [A4, E3, B1, A2]
[E1, B3, D2, C2]
VEB = 180, VFA = 180

-20-
Table 11 Expected Material handling costs for selected layouts

z(L(I),J)
Scenarios J=1 J=2 J=3 J=4 J=5 J=6 J=7 J=8 J = 9 E(z(L(I))
I=1 620 955 1120 770 1050 1185 1290 1500 1620 1153
I=2 637 935 1080 805 1070 1185 1343 1575 1680 1178
I=3 637 935 1080 805 1070 1185 1343 1575 1680 1178
I=4 627 1020 1200 733 1100 1260 1157 1420 1560 1159
I=5 620 955 1120 770 1050 1185 1290 1500 1620 1153
I=6 620 955 1120 770 1050 1185 1290 1500 1620 1153
I=7 693 1190 1420 783 1240 1440 1140 1510 1680 1282
I=8 627 1040 1240 733 1100 1260 1150 1420 1560 1163
I=9 627 1040 1240 733 1100 1260 1150 1420 1560 1163

Table 12 Material handling cost comparisons between selected functional and


distributed layouts for different demand scenarios

Distributed Functional
Scenario layout Layout
LL (I = 1) 620 1043
LM (I = 2) 935 1515
LH (I = 3) 1080 1740
ML (I = 4) 733 1333
MH (I = 5) 1050 1750
MM (I = 6) 1185 1935
HL (I = 7) 1140 1977
HM (I = 8) 1420 2015
HH (I = 9) 1560 2220

Table 13 Expected material handling cost comparisons between selected


functional and distributed layouts

E(z(L(I)) for E(z(L(I)) for


Scenario Distributed Functional
layout Layout
I=1 1153 2398
I=2 1178 2424
I=3 1178 2604
I=4 1159 2098
I=5 1153 2124
I=6 1153 2624
I=7 1282 2547
I=8 1163 2547

-21-
I=9 1163 2536

E1 B2 A1 E2

A4 E3 C3 B3

C1 A2 F1 A3

D2 B1 C2 D1

Figure 2 Selected flexible layout

-22-
material handling distances. These results are in line with those recently obtained by Venkatadri et

al. [36] in comparing holographic, fractal, and functional layouts.

Finally, since there seems to be little difference in performance between the distributed

layouts generated for different demand scenarios, it could be argued that there is no need to

optimize the layout for each demand scenario. Instead, provided a reasonable amount of

department duplication, any layout where departments are evenly distributed could be used. This

would, in fact, greatly simplify our design procedure since the layout would have to generated

only once. Only flow allocation would then be optimized for different demand scenarios. This is

similar to the approach proposed by Montreuil et al. [20] in the design of holographic layouts

where the objective is to simply generate a maximally dispersed layout. Further experimentation

is however needed before the general effectiveness of such an approach is established.

4. Discussion and Conclusion


In this paper, we addressed the problem of designing flexible plant layouts for

manufacturing facilities where product demands are subject to variability. New procedures for

designing these layouts were proposed. Both optimal and heuristic solution methods were

presented for various design and operation assumptions. The new procedures explicitly account

for the stochastic nature of product demands, allow for multiple separate departments of the same

type to exist in the same facility, and let material flows between pairs of individual departments be

determined simultaneously with department locations.

Numerical examples seem to indicate that a distributed layout generally results in a more

flexible and robust layout than the one resulting from a functional layout where departments of the

same type are restricted to be in neighboring locations. In fact, when department duplication is

high, the difference in performance between layouts with a reasonably even distribution of

departments is found to be minimal. On the other hand, in the absence of department duplication,

the need for carefully selecting the most flexible layout becomes more important and the

-23-
procedures presented here become more valuable. Additional experimentation is, however,

needed in order to generalize these observations.

As mentioned in the introduction, the disaggregation of functional layouts and the

duplication of department types is not a new idea in designing manufacturing layouts. In fact, this

forms the corner-stone principle in designing cellular layouts. However, in a truly distributed

layout, workload assignments to specific departments are made in real time based on actual

demand realizations. This contrasts with workload assignments in cellular layouts, where each

cell is dedicated to a specific part family. The dynamic workload assignment in a distributed

layout can be carried out, for example, by the creation of virtual cells, where a group of

departments is temporarily dedicated to a particular job [8]. This would, of course, necessitate a

real time shop floor control mechanism and a flexible material handling system.

Several issues still remain to be investigated. In this paper, we showed that some degree

of department duplication may be desirable. It is not clear, however, how much duplication is

generally needed and/or can be afforded. Methodologies for determining optimal levels of

duplication must thus be developed. The fact that certain departments are duplicated poses also a

challenge for workload assignments from period to period. The flow volume allocation procedure

described in this paper can certainly be used. However, the procedure must be enriched with

additional details regarding tooling, labor, and material handling requirements.

Acknowledgments

This research is supported by the National Science Foundation under grant No. DMII-

9309631 and the University of Minnesota Graduate School. The authors are grateful to Benoit

Montreuil and two anonymous referees for many constructive suggestions and comments.

-24-
Appendix 1

The procedure decribed below can be used when part demands are continuously

distributed. Monte Carlo simulation is used to generate instances of product demands. The

number of generated instances, or the sample size, must allow the simulation to reach a steady

state. Details on constructing Monte Carlo simulations can be found in [14]. An example facility

layout procedure using simulation is also given in [11].

Solution procedure:

Step 1: Read the distribution for the demand of each product.

Step 2: Read the process routing for each product and unit transfer loads.

Step 3: Read sample size S (number of scenarios to be evaluated).

Step 4: I = 0.

Step 5: I = I + 1. If I > S, then go to step 9.

Step 6: Generate demand values for each product by Monte Carlo simulation.

Step 7: Using demand values, process routings, and unit loads, generate the corresponding flow
matrix.

Step 8: Generate an optimal layout for the current flow matrix (see section 2 for details). Go back
to step 5.

Step 9: Evaluate all generated layouts and generate/choose the most flexible one (see section 2 for
details).

Stop 10. Stop.

-25-
Appendix 2

The following single optimal model may be solved directly for the most flexible layout, if the

objective function is to minimize expected material costs:

P N N Ni Nj K K
Min z = ∑ ∑ ∑ ∑ ∑ ∑ ∑ π (s)v nim jpsd klx nikx m jl
p = 1 i = 1 j = 1 ni = 1 mj = 1 k = 1 l = 1

subject to:
K
∑ x n ik = 1 ni = 1i, 2i, …, Ni; i = 1, 2, …, N (A.1)
k=1

N Ni
∑ ∑ x n ik = 1 k = 1, 2, …, K (A.2)
i = 1 ni = 1

P N Ni
∑ ∑ ∑ vnim jpstm jp ≤ cmj m j = 1j, 2j, …, N j; j = 1, 2, …, N, s = 1, 2, …, Smax (A.3)
p = 1 i = 0 ni = 1

Ni Nj

∑ ∑ vnim jps = v ijps i, j = 0, 1, …, N; p = 1, 2, …, P, s = 1, 2, …, Smax (A.4)


ni = 1 mj = 1

N Ni N Nq

∑ ∑ v n im jp = ∑ ∑ vmjrqp mj = 1j, 2j, …, Nj; j = 1, 2, …, N; p = 1,2, …, P (A.5)


i = 0 ni = 1 q = 0 rq = 1

xnik = 0, 1 ni = 1i, 2i, …, Ni; k = 1, 2, …, K (A.6)

where:
vnim js: flow volume between nth department of type i and mth department of type j under demand

scenario s, and
vijs: flow volume between departments of type i and departments of type j under demand scenario s.

A heuristic procedure similar to the one described in the main text can be applied to the above

model. The problem can be again decomposed into a layout problem and a flow volume allocation

-26-
problem. However, in this case, layout and flow volume allocation decisions are simultaneously

evaluated for all scenarios. Similarly, the performance of each layout is evaluated based on its

expected material handling cost over all possible demand scenarios.

-27-
References

[1] Ang, C. L. and P. C. T. Willey, "A Comparative Study of the Performance of Pure and
Hybrid Group Technology Manufacturing Systems Using Computer Simulation Techniques,"
International Journal of Production Research, 22, 2, 193-233, 1984.

[2] Benjaafar, S., "Machine Sharing in Cellular Manufacturing Systems," Planning, Design, and
Analysis of Cellular Manufacturing Systems, A. K. Kamrani, H. R. Parasei and D. H. Liles
(Editors), Elsevier Science B. V., 1995.

[3] Benjaafar, S. and R. Ramakrishnan, "Modeling, Measurement, and Evaluation of Sequencing


Flexibility in Manufacturing Systems," International Journal of Production Research, 34, 5,
1195-1220, 1996.

[4] Benjaafar, S., A. Soewito and M. Sheikhzadeh, "Performance Evaluation and Analysis of
Distributed Plant Layouts," Working Paper, Department of Mechanical Engineering, University of
Minnesota, Minneapolis, 1995.

[5] Bozer, Y. A. and R. D. Meller, "A Reexamination of the Distance-Based Facility Layout
Problem," to appear in IIE Transactions.

[6] Bullington, S. F. and D. B. Webster, "Evaluating the Flexibility of Facilities Layouts Using
Estimated Relayout Costs," Proceedings of the IXth International Conference on Production
Research, 2230-2236, 1987.

[7] Cooper, L., "Location-Allocation Problems," Operations Research, 11, 2, 331-344, 1963.

[8] Drolet, J. R., "Scheduling Virtual Cellular Manufacturing Systems," Ph.D. Thesis, School of
Industrial Engineering, Purdue University, West Lafayette, Indiana, 1989.

[9] Francis, R. L. and J. A. White, Facility Layout and Location: An Analytical Approach,
Prentice Hall, 2/E, Englewood Cliffs, NJ, 1993.

[10] Flynn, B. B. and F. R. Jacobs, "A Simulation Comparison of Group Technology with
Traditional Job Shop Manufacturing," International Journal of Production Research, 24, 5, 1171-
1192, 1986.

[11] Gupta, R. M., "Flexibility in Layouts: A Simulation Approach," Material Flow, 3, 243-250,
1986.

[12] Kouvelis, P., A. A. Kurawarwala and G. J. Gutierrez, "Algorithms for Robust Single and
Multiple Period Layout Planning for Manufacturing Systems," European Journal of Operational
Research, 63, 287-303, 1992.

[13] Kusiak A. and S. S. Heragu, "The Facility Layout Problem," European Journal of
Operational Research, 27, 229-251, 1987.

[14] Law, A. and D. Kelton, Simulation Modeling and Analysis, 2/E, McGraw-Hill, New York,
1994.

-28-
[15] Liao, T. W., "Design of Line Type Cellular Manufacturing Systems for Minimum Operating
and Total Material Handling Costs," International Journal of Production Research, 32, 2, 387-
397, 1993.

[16] Meller, R. and K. Y. Gau, "The Facility Layout Problem: Recent and Emerging Trends and
Perspectives," Journal of Manufacturing Systems, 15, 5, 351-366, 1996.

[17] Montreuil, B., "A Modelling Framework for Integrating Layout Design and Flow Network
Design," Proceedings of the Material Handling Research Colloquium, 43-58, Hebron, Kentucky,
1990.

[18] Montreuil, B. and A. Laforge, "Dynamic Layout Design given a Scenario Tree of Probable
Futures," European Journal of Operational Research, 63, 2, 271-286, 1992.

[19] Montreuil, B. and U. Venkatadri, "Strategic Interpolative Design of Dynamic Manufacturing


Systems Layout," Management Science, 37, 6, 682-694, 1991.

[20] Montreuil, B., Venkatadri, U. and P. Lefrançois, "Holographic Layout of Manufacturing


Systems," Technical Report No. 91-76, Faculty of Management, Laval University, Québec,
Canada, 1991.

[21] Montreuil, B., Venkatadri, U. and R. Rardin, "The Fractal Layout Organization for Job
Shop Environments," Technical Report No. 95-13, School of Industrial Engineering, Purdue
University, West Lafayette, Indiana, 1995.

[22] Pardalos, P. M. and H. Wolkowicz (Editors), Quadratic Assignment and Related Problems,
DIMACS Series, Vol. 16, American Mathematical Society, 1994.

[23] Palekar, U. S., Batta, R., Bosch, R. M. and S. Elhence, "Modeling Uncertainties in Plant
Layout Problems," European Journal of Operational Research, 63, 347-359, 1997.

[24] Rosenblatt, M. J. and D. H. Kropp, "The Single Period Stochastic Plant layout Problem,"
IIE Transactions, 24, 2, 169-76, 1992.

[25] Rosenblatt, M. J. and H. Lee, "A Robustness Approach to Facilities Design," International
Journal of Production Research, 25, 4, 479-486, 1987.

[26] Sarper, H. and T. J. Greene, "Comparison of Equivalent Pure Cellular and Functional
Production Environments Using Simulation," International Journal of Computer Integrated
Manufacturing, 6, 4, 221-236, 1993.

[27] Sethi, A. K. and S. P. Sethi, 1990, "Flexibility in Manufacturing: A survey," The


International Journal of Flexible Manufacturing Systems, 2, 289-328.

[28] Shafer, S. M. and J. M. Charnes, "Cellular Versus Functional Layout under a Variety of
Shop Operating Conditions," Decision Sciences, 36, 2, 333-342, 1988.

[29] Shore, R. H. and J. A. Tompkins, "Flexible Facilities Design," AIIE Transactions, 12, 2,
200-205, 1980.

[30] Schrage, L., User's Manual for Lindo, Scientific Press, 1991.

[31] Stecke, K. E. and N. Raman, 1995, "FMS Planning Decisions, Operating Flexibilities, and
System Performance," IEEE Transactions on Engineering Management, 42, 1, 82-89.

-29-
[32] Suresh, S. C., "Partitioning Work Centers for Group Technology: Analytical Extension and
Shop-Level Simulation Investigation," Decision Sciences, 23, 267-290, 1992.

[33] Tompkins, J. A., "Modularity and Flexibility: Dealing with Future Shock in Facilities
Design," Industrial Engineering,, September, 78-81 1980.

[34] Tompkins, J. A., and J. D. Spain, "Utilization of Spine Concept Maximizes Modularity in
Facilities Planning," Industrial Engineering,, March, 34-42, 1983.

[35] Venkatadri, U., Rardin, R. and B. Montreuil, "A Design Methodology for the Fractal Layout
Organization," Technical Report No. 95-14, School of Industrial Engineering, Purdue University,
West Lafayette, Indiana, 1995.

[36] Venkatadri, U., R. L. Rardin and B. Montreuil, "Facility Organization and Layout Design:
An Experimental Comparison for Job Shops," Technical Report No. 96-27, Faculty of
Management, Laval University, Québec, Canada, 1996.

[37] Vollmann, T. E. and E. S. Buffa, "The Facilities Layout Problem in Perspective,"


Management Science, 12, 10, 450-468, 1966.

[38] Webster, D. B. and M. B. Tyberghein, "Measuring Flexibility of Job Shop layouts,"


International Journal of Production Research, 18, 1, 21-29, 1980.

[39] Wemmerlöv, U. and L. N. Hyer, "Cellular Manufacturing in the U.S. Industry: A Survey of
Users," International Journal of Production Research, 27, 9, 1511-1530, 1989.

-30-

Você também pode gostar