You are on page 1of 19

International Journal of Production Research, Vol. 46, No.

13, 1 July 2008, 34853502

Worker allocation in lean U-shaped production lines


JOHN P. SHEWCHUK*
Grado Department of Industrial and Systems Engineering, Virginia Polytechnic Institute and State University, Blacksburg, VA, 24061, USA

(Revision received October 2006) U-shaped lines are widely used in lean systems. In U-shaped production lines, each worker handles one or more machines on the line: the worker allocation problem is to establish which machines are handled by which worker. This differs from the widely-investigated U-line assembly line balancing problem in that the assignment of tasks to line locations is fixed. This paper address the worker allocation problem for lean U-shaped production lines where the objectives are to minimize the quantity of workers and maximize full work: such allocations provide the opportunity to eliminate the least-utilized worker by improving processes accordingly. A mathematical model is developed: the model allows for any allocation of machines to workers so long as workers do not cross paths. Walking times are considered, where workers follow circular paths and walk around other worker(s) on the line if necessary. A heuristic algorithm for tackling the problem is developed, along with a procedure representing the traditional approach of constructing standard operations routines. Computational experiments considering three line sizes (up to 20 machines) and three takt time levels are performed. The results show that the proposed algorithm both improves upon the traditional approach and is more likely to provide optimal solutions. Keywords: U-shaped line; Lean production; Worker allocation; Full work

1. Introduction Lean manufacturing systems are in widespread use these days. One of the cornerstones of lean manufacturing is the use of U-shaped lines and one-piece flow, where the goal is to produce items in accordance with takt time (cycle time). When operations are manual, assembly line balancing is performed to determine how many workers are needed and which tasks are done by which workers. Tasks are grouped into stations, where each station has one worker assigned: each task is then assigned a particular location on the line. When takt time changes, the line can be rebalanced and a new assignment and grouping of tasks into stations obtained. As takt time increases, the quantity of workers on the line can be decreased and vice-versa. This is known as shojinka (Monden 1998). In U-shaped production lines, where each task corresponds to a machine performing an operation, the problem of dealing with changing demand is not

*Email: shewchuk@vt.edu
International Journal of Production Research ISSN 00207543 print/ISSN 1366588X online 2008 Taylor & Francis http://www.tandf.co.uk/journals DOI: 10.1080/00207540601115997

3486

J. P. Shewchuk
1: 1 2 9 T 1 1 9 8 2 3 3 2 7 6 4: 4 4 5 2: 8

3:

3 7

(a)

4 5 6 (b)

Figure 1. (a) Example 9-machine U-shaped production line, showing allocation of four workers. (b) Standard operations routines: solid line manual time; wavy line walking time; dashed line machine time.

as straightforward. The sequence of machines on the line is often fixed, making it impossible to reallocate tasks to different line locations. Even when routing alternatives are available, it may be difficult and/or uneconomic to move machines whenever takt time changes. Thus, the assignment of tasks to line locations is fixed. If a single operator can handle multiple machines (a concept known as jidoka; Monden 1998), however, the line balancing problem can be replaced with the worker allocation problem, where the goal is to determine which machines are to be handled by each worker. For a given takt time T, each worker will have a standard operations routine which describes which machines will be visited and in what sequence (figure 1). Whenever a worker arrives at a machine, he/she unloads the completed item, loads the next item, performs any needed actions at the machine and then walks to the next machine. Following the last machine, the worker returns to the starting machine and waits until the start of the next cycle as necessary. Workers can be reallocated when takt time changes, thus making shojinka possible on such lines. While the worker allocation problem is somewhat similar to the U-line assembly line balancing problem (UALBP), there are three important differences. The first is that task assignments on the line are fixed, as previously described. The second is that walking time can be significant on such lines and thus cannot be ignored (Nakade and Ohno 1999). The third is that as there are no restrictions on what machines can be assigned to a given worker, the possibility of workers crossing paths exists. This should be avoided (e.g. Hyer and Wemmerlov 2002) and thus only assignments where workers do not cross paths should be allowed. A further consideration is that while the objective in traditional line balancing is to balance workload across operators, the objective in lean systems is often to maximize full work, i.e. have as many workers fully utilized as possible. The reason is that this provides the opportunity to eliminate the least-utilized worker by improving

Worker allocation in lean U-shaped production lines

3487

processes accordingly (Suzaki 1987, Dennis 2002). In figure 1, for example, worker 2 handles machine 8 only, and is accordingly utilized only about 14% of the time. This worker can be eliminated by either eliminating the process at machine 8 entirely or improving processes at machines 13, 7, and/or 9 so that either worker 1 or 3 can handle machine 8 as well. Being able to establish full work allocations such as this allows us to focus on where process improvement efforts on the line should be directed. The purpose of this paper is to: . Develop a mathematical model of the worker allocation problem for lean U-shaped production lines. . Based upon the model, develop a suitable algorithm for tackling such problems. The objective is to minimize the quantity of workers on the line, then maximize full work. A simple algorithm is desirable, such that it can quickly and easily be implemented by lean practitioners. Relevant literature is reviewed in section 2, followed by development of a mathematical model in section 3. Section 4 presents both the traditional approach and a new algorithm for solving the worker allocation problem; computational experiments to establish performance are described in section 5. Concluding remarks are provided in section 6, along with a discussion of where additional research is needed.

2. Literature review The assembly line balancing problem (ALBP) has been widely studied over the past several decades. Becker and Scholl (2006) provide a comprehensive review of generalized assembly line balancing problems and their solution. Of relevance for this paper is the U-line assembly line balancing problem (UALBP), where the assembly line is arranged in a U-configuration and stations can be assigned tasks from both sides of the line. In particular, the problem version UALBP-1 is concerned with minimizing the number of stations (workers) for a given cycle (takt) time. The UALBP was first introduced by Miltenburg and Wijngaard (1994) and tackled using dynamic programming. Since then, researchers have addressed the problem using a variety of approaches including commercial optimization packages (Urban 1998), branch and bound (Scholl and Klein 1999, Aase et al. 2003), simulated annealing (Erel et al. 2001), and goal programming (Gokcen and Agpak 2006). As previously noted, however, the UALBP is fundamentally different from the worker allocation problem in lean U-shaped production lines: the UALBP examines both the assignment of tasks to locations on the line and assignment of workers to tasks, while task (i.e. machine) locations are considered fixed in the worker allocation problem. Even if machines are considered to be movable, the UALBP does not consider walking times: hence they are ignored in such works. Miltenburg and Wijngaard (1994), however, note the need to consider walking times as stations sprawl across different parts of the U-line. The closely-related problem of worker allocation in U-shaped production lines has been largely ignored in the research community. This is surprising considering that adding or removing operators from U-shaped lines often occurs frequently, with

3488

J. P. Shewchuk

moving of machines left for large changes in output rate (Miltenburg 2001). The most well-known approach to the problem is presented by Monden (1998). He provides a manual procedure for determining the standard operations routines, or allocation of workers to machines, for a U-shaped line and product to be made (figure 1). The procedure consists of selecting one operation (machine) at a time, from those available, and assigning to the current worker. This is continued until any additional assignments would violate the takt time constraint: the implicit goal is to provide each worker with full work. Walking time is considered, as is any time needed for safety and/or quality checks. The solution is constructed graphically, showing the relationships between manual time, machine time, walking time, and takt time. While graphical methods such as this have become very popular in industry, they suffer from the fact that they provide no guidance as to how to select the next operation at each stage. Ohno and Nakade (1997) develop a procedure for determining the minimum cycle time for a given quantity of workers in U-shaped production lines. They consider manual, machine, and walking times, as well as the waiting time which results when a worker arrives at a machine and the item there is still in process. Nakade and Ohno (1999) extend this work to develop an optimal procedure for finding the minimum quantity of workers for a given takt (cycle) time and a worker allocation which then balances workload on a U-shaped production line. They again consider manual, machine, walking, and waiting times, and ensure that workers do not cross paths in establishing stations. Lines have an even quantity of machines (i.e. rectangular shape), and each worker must have an uninterrupted stretch of machines when following a circular walking path around the line. The machine perimeter for each worker thus forms either a point (single machine: worker 2 in figure 1), line (two machines: worker 3), triangle (three machines: worker 1), rectangle (e.g. machines 2, 3, 4, 6, 7, 8) or trapezoid (e.g. machines 1, 2, 7, 8, 9). The model developed in this paper differs in that (i) the objective is to maximize full work, (ii) the quantity of machines on the line can be even or odd, and (iii) any allocation of machines to workers is allowed so long as workers do not cross paths. In other words, there is no requirement for each worker to have an uninterrupted set of machines when following a circular walking path: interrupted circular machine assignments (e.g. machines 1, 3, 4, and 8, figure 1) are possible.

3. Mathematical model As previously stated, our objective in allocating workers to lean U-shaped production lines is to minimize the quantity of workers, then maximize full work. To meet these objectives, we develop a mathematical model which maximizes full work for a given quantity of workers W. Solution then involves starting with W set at its lower bound, then incrementing by 1 as necessary until the problem becomes feasible. Only a few iterations at most should be required, as each additional worker provides a substantial capacity increase and the quantity of workers is limited by the number of machines.

Worker allocation in lean U-shaped production lines


i1 2d d a M M1 M+2 i M+1 i Mi b M/2 +2 M/2 +1 i+1 M/2 1

3489

M/2

Figure 2.

Worker locations, even-numbered U-shaped line.

The model is as follows. The line consists of M machines in a U-shape arrangement, as shown in figure 1 (M odd) or figure 2 (M even). The manual time at machine i is ti: takt time is T. W workers are to be allocated to the line. The following assumptions are made: 1. Machines and workers are located via a grid arrangement, with distance d between adjacent worker locations in the same row (e.g. locations 1 and 2, figure 2), and 2d between opposite worker locations (e.g. 1 and M). p When M is odd, the middle worker location is located at a distance of 2d 2 from the preceding and following locations. 2. Takt time is always greater than or equal to the longest operation time (manual time machine time). This means that a feasible solution always exists for a given problem. (If an operation time is larger than takt time in practice, it must be reduced accordingly or other methods used to produce to that takt time.) 3. Any worker can perform any manual task on the line. 4. Manual time at each machine and walking times between machines are both deterministic and worker-independent. 5. Each worker follows a circular walking path inside the line, stopping at each assigned machine in succession. If two successive machines are not adjacent on the walking path (i.e. interrupted circular machine assignment, section 2), the worker must walk around the other worker(s) in going from the first machine to the second. Circular walking paths are used in Monden (1998) and Nakade and Ohno (1999): in these works the approach minimizes total walking time as each path forms a convex polygon. While the approach may not guarantee minimum walking times in this paper (walking paths become non-convex when one worker must walk around another), it is still used as it is common both in practice and research and to maintain tractability. 6. Direct walking time between worker locations (i.e. walking time when no other workers on the line to walk around) is directly proportional to Euclidean distance between locations. 7. To walk around one or more workers on a straight section of the line, a worker travels to and along (if necessary) the major axis dividing the U-shaped line, as shown in the examples of figure 2 (paths (i 1) a (i 1), (i 1) a b (M/2)). This results in a constant additional walking distance p of 2 2d 2 d, regardless of how many workers must be walked around.

3490

J. P. Shewchuk

8. Each worker can start with any machine in his/her assignment: starting location does not affect the total work (manual walking) assigned to the worker. However, different starting locations can change the WIP (work in process) required to prevent workers from starving. We assume the minimum WIP required on the line such that starvation does not occur. To establish the allocation of workers to machines, define & 1, machine i assigned to worker k, xik 0, otherwise Given any solution {xik, i 1, . . . , M, k 1, . . . , W}, total walking time for each worker is fixed, per assumption 5. To determine total walking time for a worker, we first calculate the direct walking time (round-trip time when no other workers on the line), then add any additional time resulting from walking around one or more other workers. Define & 1, worker k moves from machine worker location i to machine j, qijk 0, otherwise and vijk 8 < 1, : 0, worker k walks around other worker when going from machine i to machine j, otherwise

and let wij be the direct walking time between machines i and j (wij wji) and s the additional time incurred from walking around workers. A mathematical model for worker allocation in lean U-shaped production lines can then be expressed as follows: ! W M M M PP P P xik ti qijk wij Yk a Maximize Z subject to a
M X i1 M X i1 k1 i1 i1 j1

TW 1
M M XX i1 j1

xik ti

qijk wij Yk ,

k 1, . . . , W

xik ti

M M XX i1 j1

qijk wij Yk a H1 dk

H1 ,

k 1, . . . , W

W X k1 M X i1 M M XX i1 j1

dk ! 1

xik ti

qijk wij Yk

T,

k 1, . . . , W

Worker allocation in lean U-shaped production lines


W X k1

3491

xik 1,

i 1, . . . , M

Bik Dik xik xM1i, k ! 1,


i1 X j1 i1 X j1 M X jM2i M X jM2i Mi X ji1 Mi X ji1

i 2, . . . , L1 , k 1, . . . , W

H2 Bik

xjk

xjk ! 1

i 2, . . . , L1 , k 1, . . . , W

H2 Bik

xjk

xjk

H2

i 2, . . . , L1 , k 1, . . . , W

H2 Dik

xjk ! 1

i 2, . . . , L1 , k 1, . . . , W

10

H2 Dik

xjk

H2

i 2, . . . , L1 , k 1, . . . , W

11

xik xM2i, k xi1,n xM1i,n

3, i 2, ..., L2 , k 1, ...,W, n 1, . ..,W, n 6 k 12

M X i1 M X j1 M X n1

qijk xjk 0,

j 1, . . . , M, k 1, . . . , W

13

qijk xik 0, ! xnk H2 ,

i 1, . . . , M, k 1, . . . , W

14

H2 qijk

xnk

i X n1

i 1, . . . , M, j 1, . . . , i, k 1, . . . , W 15

Yk

L2 2 X X L2 i1 ji2

vijk vjik s

M2 X

M X

vijk vjik s,

k 1, . . . , W

16

iL1 2 ji2

9 > > > > > H2 1 vijk qijk Fijk 5H2 2 = i 1, . . . , L2 2, j i 2, . . . , L2 , k 1, . . . , W H2 1 vijk qijk Fijk ! 2 H2 1 vjik qjik Fjik ! 2 H2 1 vjik qjik Fjik 5H2 2 > i L1 2, . . . , M 2, j i 2, . . . , M, k 1, . . . , W > > > > ; 1724

3492

J. P. Shewchuk
j1 P ni1

H2 Fijk H2 Fijk Fjik Fijk

xnk ! j i 1

j1 P ni1

9 > > > > > > > = i 1, ...,L2 2, j i 2, ..., L2 , k 1, ...,W

xnk 5H2 j i 1 > i L1 2, . ..,M 2, j i 2, ... ,M, k 1, . ..,W > > > > > > ; 2530

where L1 M/2 1 if M is even and (M 1)/2 if M is odd; L2 M/2, M even, (M 1)/2, M odd. The objective function (1) maximizes the average utilization of all but the least-utilized worker: we refer to this measure Z as the full work level for the system. For each worker k, utilization is total time (manual direct walking walking around other workers) divided by takt time T: the least-utilized worker has total time a. Constraints (2), (3) and (4) calculate a via parameter dk (dk 1 if worker k has total time a, 0 otherwise) and the big-M constant H1 (H1 ! T). Constraint (5) ensures that takt time is not exceeded for any worker, while each machine must be assigned to only one worker via constraint (6). Constraint (7) ensures that all machines assigned to each worker are adjacent. Given a U-shaped line with an even number of machines (M even) and any pair of opposing machines i and j M 1 i (figure 2), a feasible assignment of machines to worker k requires at least one of {i, j} to be assigned if one or more assigned machines are to the left of {i, j} (i.e. Bik 0) and one or more assigned machines are to the right (Dik 0). Constraints (8)(11) then ensure that Bik and Dik are set accordingly, again via a big-M constant, H2 (H2 ! M). Note that when M is even, opposing pairs {i, j} end at i M/2 1: when M is odd, they end at i (M 1)/2. This is taken care of via the parameter L1. Constraint (12) ensures that workers do not cross paths. Consider, for example, that on a line with an even number of machines, worker k is assigned two machines: i and M 2 i (figure 2). Another worker n could then be assigned machines {i 1, M 1 i}, as these are adjacent. This would cause the workers to cross paths, however. Constraint (12) ensures this does not happen: given the possible assignments of {i, M 2 i} to worker k and {i 1, M 1 i} to worker n, at most three such assignments can be made. In similar manner to the adjacency constraints, the range for i depends upon whether M is even or odd: this is specified via the parameter L2. Constraints (13)(15) calculate qijk values from xik values, ensuring each worker moves in a circular path around the line. Constraint (16) calculates the total additional time incurred by each worker k in walking around other workers, Yk, based upon whether or not worker k walks around other workers when going from machine i to machine j (vijk values) and constant time value s. Variables vijk are established in constraints (17)(24), via Fijk (Fijk 1 if worker k is not assigned all machines in-between machines i and j, 0 otherwize) and qijk. Finally, variables Fijk are established via constraints (25)(30). The above problem is a variation of the generalized assignment problem, which is NP-hard.

Worker allocation in lean U-shaped production lines

3493

3.1 Calculation of lower bound on quantity of workers To determine a lower bound on the quantity of workers, we use the approach of Nakade and Ohno (1999). Given W workers, a lower bound on cycle time C for the line, LB[C], can be found from the following:
M P

ti M W 1lmin W

LBC i1

The numerator equals total operator time required, which consists of all manual time plus all walking time. If we have W workers, walking time is minimized if the first W 1 workers remain stationary and the remaining worker then covers M (W 1) M W 1 stations. Given that the minimum walking time between any two machines is lmin, total walking time is then at least (M W 1) lmin. We can then find LB[W] from LBW fmin WjLBC ! Tg which results in 6 LBW 6i1 4 2P M ti M 1lmin T lmin 3 7 7 5

This is equivalent to the expression provide by Nakade and Ohno (1999).

4. Algorithms The goal of this paper is to develop an algorithm for allocating workers to lean U-shaped production lines quickly and easily in practice. A simple, single-pass procedure is as follows: Step 0: Initialize. Set k worker 1, A set of unassigned machines {1, . . . , M}. Step 1: Prepare for next worker. Set Sk set of machines assigned to worker k . Step 2: Establish feasible machines for worker k Fk. A machine m 2 A is feasible iff: it is adjacent to at least one other machine in Sk. it does not result in worker k crossing paths with any other worker, and it does not cause takt time to be violated, i.e. total time (manual total round-trip walking time, circular walking path) for machines Sk m T. If Fk , go to Step 5. Step 3: Select machine m from Fk for assignment to worker k. Step 4: Assign machine m to worker k. Set Sk Step 2. Sk m, A A m, Fk : go to

3494
3 1 7 4 8 2 6 7 (a) 2 3 4 12 5 5 T = 20 7 3 1

J. P. Shewchuk
8 2 6 2 3 4 12 5 7 4 6 5 4 7 5 W = 3, Z = 0.75 (b) 3 1 8 2 2 3 4 12 7 5 W = 3, Z = 0.95 (c)

Figure 3. (a) Example 7-machine problem. (b) Solution via traditional approach. (c) Solution via initial proposed algorithm (optimal).

Step 5: Worker k completed. If A , done: set W k, calculate Z from Sk, k 1, . . . , W. Otherwize, set k k 1, go to Step 1. Different approaches for Step 3 (machine selection) result in different algorithms for the worker allocation problem. We first modify Step 3 to obtain the traditional approach; i.e. that employed when worker allocation is done by generating a standard operations routine for each worker. An improved algorithm is then presented.

4.1 Traditional approach As outlined in section 2, Monden (1998) provides a basic, step-by-step procedure for developing standard operations routines, without detailing exactly how machines are to be selected. With no other information, it is reasonable to assume machines are simply selected in order. This results in the following algorithm (procedure for Step 3 above): Step 3: Select machine m mini fki , where fki ith machine in Fk. Figure 3 shows an example of the traditional approach applied to a sevenmachine problem, where walking times are assumed zero. An optimal solution is also shown: comparing the two indicates improved performance is possible.

4.2 Proposed algorithm The proposed algorithm begins with the idea of simply assigning machines in order of largest total time (manual walking), as machines which result in large total time values become harder to fit as assignments are made. The algorithm is as follows: Step 3: (i) Calculate Vk {vki}, where vki total time (manual total round-trip walking time, circular walking path) for machines Sk fki 2 Fk. (ii) Select machine m fki, where i argmaxi fvki g. Consider this initial algorithm applied to the problem in figure 3(a). Starting with worker 1, machine 4 is selected first, following which F1 {3, 5} and V1 {14, 17}. Thus, machine 5 is selected next. Continuing on with the procedure in this manner produces the solution in figure 3(c), which is optimal. This simple approach, however, has a drawback in that it can result in group splitting: the original group

Worker allocation in lean U-shaped production lines


4 1 8 12 7 2 7 5 (a) 6 3 6 9 10 4 5 2 T = 20 4 1 8 7 2 7 6 3 6 10 4 5 4 1 8 12 7 2 7 5 (c) 6 2 6 9 10 4 5 2

3495

12 5 9 2 W = 5, Z = 0.663 (b)

W = 3, Z = 0.95

Figure 4. (a) Example 8-machine problem, showing feasible machines (dashed lines) following selection of first machine for worker 1 using initial proposed algorithm. (b) Solution via initial proposed algorithm. (c) Solution via final proposed algorithm (optimal).

of feasible machines is split into two non-adjacent, disjoint sets. This can make subsequent assignments less attractive as there are then fewer alternatives from which to select in a given group. An example of this is shown in figure 4. Following selection of the first machine (machine 8) for worker 1, F1 {1, 2, 7} and V1 {16, 19, 17}, so machine 2 is selected, This selection splits the remaining machines into two groups (this also occurs for worker 2). Continuing with the procedure produces the solution of figure 4(b). To address this shortcoming, we modify the above so that selections which result in split groups are avoided. The resulting algorithm is as follows: Step 3: (i) Calculate Gk {gki}, where gki grouping factor for ith feasible machine for worker k: 1, assignment of ith machine will result in all remaining machines in Fk being adjacent i:e: single group remains: 2, assignment of ith machine will result in all remaining machines in Fk being split into two disjoint, non-adjacent groups: (ii) Determine the set of candidate machines for worker k, Ck, where & Ck fcki g f fki g, gki 2, i 1, . . . , N, N jFk j f fki jgki 1g, otherwise

In other words, if at least one candidate machine does not result in split groups, only such machines are considered. (iii) Calculate Vk {vki}, where vki total time for machines Sk cki 2 Ck. (iv) Select machine m cki, where i argmaxi fvki g. Consider the use of the proposed algorithm for the problem of figure 4(a). Following selection of the first machine (machine 8) for worker 1, we have F1 {1, 2, 7} and G1 {1, 2, 1}, so C1 {1, 7} and V1 {16, 17}. Hence, machine 7 is

3496

J. P. Shewchuk

selected next. Continuing with the procedure results in the solution of figure 4(c), which is optimal.

5. Computational experiments 5.1 Experimental design A set of computational experiments was performed to establish the performance of both the traditional approach and the proposed algorithm. Three sizes of U-shaped lines were investigated: small (59 machines), medium (1014 machines), and large (1520 machines). Manual time was uniformly distributed between 530 seconds at each machine. Following Nakade and Ohno (1999), walking time was 1 second for adjacent machines (same row) and 2 seconds for opposite machines. For a given quantity of machines M and manual times {ti}, three different takt time valueslow, medium, and highwere investigated. To determine appropriate values, we first assume that manual time is no more than 50% of the total operation time at any machine. As takt time must be greater than or equal to the largest operation time, this results in a lower bound on takt time of LB[T] 2 max{ti}. At the other end, we note that a single worker is sufficient whenever there is enough time for the worker to perform all manual operations and walk once around the line. This P provides an upper bound on takt time of UB[T] ti (total round-trip walking time). We can then calculate the possible takt time range Trange UB[T] LB[T], and establish low, medium and high values based upon this range. The experimental design thus consisted of two factors with three levels each as follows: Line size: Small : M 59; medium : M 1014; large : M 1520 Takt time: Low : T LBT; medium : T LBT 1=3 Trange ; high : T LBT 2=3 Trange This resulted in nine design points, or problem scenarios, for investigation. For each machine quantity M in each scenario, 10 problems (i.e. sets of {ti} values) were randomly generated. Thus, 50 problems were generated for each of the first six problem scenarios (small lines, and medium lines, each with low, medium, and high takt times) and 60 problems for each of the last three problem scenarios (large lines with low, medium and high takt times), giving a total of 480 problems. Each problem was solved using both the traditional approach and the proposed algorithm. These were coded using the C programming language and run on an IBM RS/6000 Unix-based server: all 50 or 60 problems comprising a given scenario could be solved, using either approach, in a few seconds. For comparison with optimal values, the mathematical model of section 3 was coded using the AMPL programming language and problems solved using CPLEX 9.0.0 on a 1.7 GHz PC. Recall from section 3 that the optimal approach requires attempting to solve the problem first with W LB[W], then incrementing W as required until the problem becomes feasible. It was found that for no problem were more than two solution attempts required. Solution time increased as lines became larger and takt time smaller (resulting in more workers): for low takt times, maximum solution time went from 70 seconds for small lines (59 machines) to over sixteen hours with medium

Worker allocation in lean U-shaped production lines

3497

lines (1014 machines). For large lines (1520 machines) and low takt times, solution time became so excessive that optimal solutions could not be established. 5.2 Results To provide some overall idea of the range of values obtained for W and Z for the different solution methods employed, table 1 presents these statistics for each of the nine problem scenarios. A cursory examination of table 1 suggests that the proposed algorithm performed better: values for W and Z are as good as or better than those obtained with the traditional approach in almost every case. Figure 5 presents
Table 1. Computational results (ranges of values obtained) for W and Z, each solution approach. Quantity of workers, W Machine quantity 59 59 59 1014 1014 1014 1520 1520 1520 Takt time L M H L M H L M H Optimal 25 23 2 36 23 22 3 2 Traditional approach 25 23 23 37 24 23 59 34 24 Proposed algorithm 25 23 2 37 24 22 510 34 2 Full work level, Z (%) Optimal 84100 83100 91100 8598 95100 95100 9699 98100 Traditional approach 6399 65100 54100 7098 54100 57100 7295 6199 43100 Proposed algorithm 72100 78100 82100 7398 68100 92100 7097 7099 90100

6 10 22 6 5 5

13 15 22 18 19 22 11 15 25 W = 3, Z = 0.899 (Optimal) (b) M = 9, T = 90 (medium)

22

25 12 20 27

15 18 27 11 18 W = 5, Z = 0.927 (Optimal) (c) M = 11, T = 54 (low)

W = 2, Z = 0.986 (Optimal) (a) M = 6, T = 56 (high)

27 25 12 19 10 24 10 7 9 28 10 16

19

14 26

8 6

5 9

10 27 24 8 30 19

28 23 26

W = 2, Z = 0.999 (Optimal) (d) M = 12, T = 158 (high) 6 9 20 26 10 21 15 27 11 7 6 10 10 12 24 25 8

W = 3, Z = 0.962 (Optimal: W = 3, Z = 0.978) (e) M = 16, T = 140 (med) 9 10 12 7 8 6 9 8 5 6 20 18

14 19 30

27 17 30 30 20

W = 7, Z = 0.818 (LB[W] = 5) (f) M = 17, T = 54 (low)

W = 3, Z = 0.970 (Optimal: W = 3, Z = 0.995) (g) M = 20, T = 149 (medium)

Figure 5. Example worker allocation problems and solutions, all solved using proposed algorithm.

3498

J. P. Shewchuk

example problems of varying size and their solution, all solved using the proposed algorithm. The 16-machine example of figure 5(e) shows an optimal solution having an interrupted circular machine assignment (section 2): the worker going from machine 2 to machine 5 must walk around another worker attending machines 3 and 4. It was found that optimal solutions contained interrupted circular machine assignments 5%, 11%, and 31% of the time for small, medium, and large lines, respectively. The performance of the proposed algorithm relative to the traditional approach, and of each of these relative to optimal, are summarized in table 2. Comparisons to optimal are done for all scenarios except large lines (1520 machines) having low takt times. We will refer to the eight scenarios where optimality was established as the optimally-solved scenarios for convenience. 5.2.1 Quantity of workers. Table 2 shows that the traditional approach to worker allocation on U-shaped production lines worked quite well for minimizing the quantity of workers: optimal solutions resulted at least half the time (and as high as 86% of the time) for each of the optimally-solved scenarios. The proposed algorithm performed much better, however. In all eight optimally-solved scenarios, improvement occurred more than half the time it was possible: in six of these, improvement resulted 100% of the time. In three scenarios (medium-size lines with medium takt times, large lines with both low and high takt times), this resulted in a reduction of 1.16 workers on average, indicating that for some problems two workers were eliminated. Such manpower reductions would normally be considered quite substantial in practice. Furthermore, the proposed algorithm provided worse solutions than those obtained via the traditional approach much less than half the time (7% on average), and always provided a larger quantity of optimal solutions. 5.2.2 Full work level. Table 2 shows that performance in maximizing full work level Z is not as good as that found in minimizing W. This is not surprising: optimality is more difficult to obtain for Z, as compared to W, as the former a real values and the latter integer. At the same time, however, there is more opportunity for improvement with Z. Using the traditional approach, optimal results were obtained 26% of the time or less for the eight optimally-solved scenarios, with the average being about 13%. As with W, performance was better using the proposed algorithm. Improvement again occurred at least half the time it was possible in all eight optimally-solved scenarios: the level of improvement in Z averaged 14% across all scenarios. In two scenarios (medium and large lines, high takt times), the improvement was around 20%, a sizeable increase in full work for each worker. The proposed algorithm again performed worse than the traditional approach less than half the time (26%, on average), and provided a larger quantity of optimal solutions in all but one scenario (medium-size lines, medium takt times). 5.3 Discussion of results The computational results indicate that the proposed algorithm performed as good as or better than the traditional approach for seven of the eight optimally-solved scenarios. With medium-size lines having medium takt times, the traditional

Table 2. Quantity of workers, W Proposed algorithm Improv1 20(100) 28(100) 14(100) 22(58) 14(58) 34(100) 20() 15(100) 43(100) 4 2 0 18 2 0 23 13 0 96 98 100 66 88 100 87 100 1.00 1.00 1.00 1.00 1.29 1.00 1.08 1.00 1.12 26 24 20 4 12 12 0 10 56(76) 60(79) 48(60) 60(62) 58(66) 76(86) 60() 65(65) 70(78) WI2 ws3 opt4 % optimal Improv1 Traditional approach ZI2 11.9 15.0 16.0 7.6 11.1 18.5 6.8 10.7 22.9 Full work level, Z (%)

Comparison of computational results for W and Z, each solution approach.

Machine quantity

Takt time

Traditional approach

Proposed algorithm ws3 20 12 28 36 34 12 38 27 25 opt4 44 56 22 10 6 32 3 20

% optimal

59 59 59 1014 1014 1014 1520 1520 1520

L M H L M H L M H

80 72 86 62 76 66 85 57

Worker allocation in lean U-shaped production lines 3499

1 Percentage of solutions improved as compared to traditional approach (Percentage of solutions improved considering only those where improvement possible). Bold indicates that improvement occurred at least half the time. 2 Average improvement in W (Z) over traditional approach, improving solutions only. 3 Percentage of solutions worse as compared to traditional approach. Bold indicates that solutions were worse less than half the time. 4 Percentage of solutions found to be optimal. Bold indicates algorithm resulted in optimal at least as often as traditional approach.

3500

J. P. Shewchuk

approach provided better results for Z only, providing optimal solutions 12% of the time versus 6% with the proposed algorithm. Overall, the proposed algorithm thus appears to work better for allocating workers in U-shaped production lines for the stated objectives. We also note the importance of allowing interrupted circular machine assignments: these were found in a significant quantity (between 5 and 31%) of optimal solutions. Such solutions are not possible in related works such as Nakade and Ohno (1999). One question which remains is how much additional improvement in the proposed algorithm is possible, especially with respect to full work level, Z. Further investigation seemed to indicate that additional improvement would require a look-ahead feature and/or performing multiple passes through the data. It is likely, however, that adding such methods would result in an algorithm too difficult to use manually. Such further improvements were thus not explored. It is also important to keep in mind that no matter how good we are at allocating workers to maximize full work, the resulting solution may or may not be amenable to worker elimination via process improvement (as discussed in section 1). Ideally, we would like to see the least-utilized worker have a single machine and low workload, such that the elimination of this worker through process improvement is a reasonable proposition. In figure 5, for example, this occurs in examples (a), (b) and (f), where the lowest worker utilization is 12% or less via a single machine. Worker elimination via process improvement will be more difficult in examples (c) and (d), where the least utilized worker has a utilization of about 33%. In any event, the first step is to always to find the minimum quantity of workers which maximizes full work: the results can then be used to target process improvement efforts to the extent possible.

6. Conclusions This paper has addressed the problem of worker allocation in lean U-shaped production lines, where the objective is minimize the quantity of workers on the line, W, then maximize full work level Z (average % utilization, all but least-utilized worker). A mathematical model was developed: the model allowed for any allocation of machines to workers as long as workers do not cross paths. Walking times were considered, where workers followed circular paths and walked around other worker(s) on the line if necessary. To solve worker allocation problems quickly and easily in practice, a simple algorithm was developed. Computational results indicate that the algorithm, which in selecting the next machine for a worker considers only those machines which do not result in split groups and then selects the machine having the largest total time (manual walking), performed better overall than the traditional approach to the problem. Regarding the mathematical model, it was found that optimal solutions to the worker allocation problem in lean U-shaped production lines can be found in reasonable time (less than one hour) for any size line up to 20 machines as long as takt time is not too small. The quantity of workers required increases as takt time decreases, and optimal solution time is extremely sensitive to the quantity of

Worker allocation in lean U-shaped production lines

3501

workers used. It was also found that the lower bound on the quantity of workers proposed by Nakade and Ohno (1999) is very good: of the 420 problems solved optimally in this paper, 372 had the minimum quantity of workers equal to the bound and the remainder required only one more worker than the lower bound. This implies that the bound can provide very good estimates of the quantity of workers needed on U-shaped production lines in practice. A further conclusion based upon optimal solution results is that it may be highly beneficial in lean U-shaped production lines to allow interrupted circular machine assignments, i.e. those where workers must walk around other worker(s) when following circular walking paths around the line. The larger the line, the greater the likelihood the performance can be improved by allowing such assignments. Additional research is envisioned in three areas. First of all, the validity of the mathematical model and algorithm can be improved. In real systems not every worker can run every machine, some workers perform different tasks faster and/or better than others, etc. Consideration of such factors leads to the worker assignment problem, where the goal is to assign tasks to specific workers. The mathematical model and proposed algorithm presented here provide a natural starting point for tackling this more-difficult operational problem. Second, this paper assumed circular walking paths for workers, which may not necessarily result in minimum walking times when interrupted circular machine assignments are allowed. Additional work is required to establish whether this is indeed the case and, if so, how walking paths are best established. Finally, this paper assumed sufficient WIP that workers are never starved. In some cases, however, changing the starting location for workers on the line can reduce the amount of WIP required. Work is needed to investigate how to best select, for a given assignment of machines to workers, the starting locations for the workers in order to minimize WIP.

References
Aase, G.R., Schniederjans, M.J. and Olson, J.R., U-OPT: an analysis of exact U-shaped line balancing procedures. Int. J. Prod. Res., 2003, 41(17), 41854210. Becker, C. and Scholl, A., A survey on problems and methods in generalized assembly line balancing. Euro. J. Oper. Res., 2006, 168, 694715. Dennis, P., Lean Production Simplified, pp. 5758, 2002 (Productivity Press: New York, NY). Erel, E., Sabuncuoglu, I. and Aksu, B.A., Balancing of U-type assembly systems using simulated annealing. Int. J. Prod. Res., 2001, 39(13), 30033015. Gokcen, H. and Agpak, K., A goal programming approach to simple U-line balancing problem. Euro. J. Oper. Res., 2006, 171(2), 577585. Hyer, N. and Wemmerlov, U., Reorganisng the Factory: competing through Cellular Manufacturing, 2002 (Productivity Press: Portland, OR). Miltenburg, G.J. and Wijngaard, J., The U-line line balancing problem. Manage. Sci., 1994, 40(10), 13781388. Miltenburg, J., U-shaped production lines: a review of theory and practice. Int. J. Prod. Econ., 2001, 70(3), 201214. Monden, Y., Toyota Production System, 1998 (Institute of Industrial Engineers: Norcross). Ohno, K. and Nakade, K., Analysis and optimization of a U-shaped production line. J. Oper. Res. Soc. Japan, 1997, 40(1), 90104.

3502

J. P. Shewchuk

Nakade, K. and Ohno, K., An optimal worker allocation problem for a U-shaped production line. Int. J. Prod. Econ., 1999, 6061, 353358. Scholl, A. and Klein, R., ULINDO: optimally balancing U-shaped JIT assembly lines. Int. J. Prod. Res., 1999, 37(4), 721736. Suzaki, K., The New Manufacturing Challenge: techniques for Continuous Improvement, pp. 131132, 1987 (The Free Press: New York, NY). Urban, T.L., Optimal Balancing of U-Shaped Assembly Lines. Manage. Sci., 1998, 44(4), 738741.

You might also like