You are on page 1of 10

European Journal of Operational Research 212 (2011) 277286

Contents lists available at ScienceDirect

European Journal of Operational Research


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

Discrete Optimization

Incorporating ergonomic risks into assembly line balancing


Alena Otto, Armin Scholl
Friedrich-Schiller-University of Jena, Chair of Management Science, Carl-Zei-Strae, D-07743 Jena, Germany

a r t i c l e

i n f o

Article history:
Received 23 October 2010
Accepted 29 January 2011
Available online 4 February 2011
Keywords:
Scheduling
Combinatorial optimization
Ergonomic risk assessment
Assembly line balancing
Simulated annealing

a b s t r a c t
In manufacturing, control of ergonomic risks at manual workplaces is a necessity commanded by legislation, care for health of workers and economic considerations. Methods for estimating ergonomic risks of
workplaces are integrated into production routines at most rms that use the assembly-type of production. Assembly line re-balancing, i.e., re-assignment of tasks to workers, is an effective and, in case that no
additional workstations are required, inexpensive method to reduce ergonomic risks. In our article, we
show that even though most ergonomic risk estimation methods involve nonlinear functions, they can
be integrated into assembly line balancing techniques at low additional computational cost. Our computational experiments indicate that re-balancing often leads to a substantial mitigation of ergonomic risks.
 2011 Elsevier B.V. All rights reserved.

1. Introduction
The problem of unfavorable working conditions, or poor workplace ergonomics, is an acute topic today. Ergonomic risks at the
workplace cause a lot of damage on health and quality of life of
workers, deteriorate economic results of employers and of the
economy as a whole. In 2008, along 315,000 cases of work-related
musculoskeletal disorders (MSDs, often referred to as ergonomic
injuries), requiring a median of 10 days away from work, were reported in the US (Bureau of Labor Statistics, 2009). Annual compensation cost for MSDs paid by employers in the US amount to
15 to 20 billion US dollars. Moreover, occupational diseases of
workers indirectly cause further cost on rms: via loss of production capacity due to absenteeism of workers, lower worker productivity and higher defect rates in work. This can be illustrated by the
example of Peugeot, whose ergonomics program reduced the cycle
time for the nal vehicle assembly line together with a simultaneous decrease by 30% in new cases of musculoskeletal disorders
(Moreau, 2003).
Workplace ergonomics is becoming even more important following recent developments in legislation (EU Machinery directive,
2006/42/EC, 89/391/EEC, Occupational Safety and Health act of
1970 among others) and an on-going ageing of the workforce in
most of the developed countries.
Already today in assembly line production, especially in nal
assembly, where the share of manual labor is high, a special attention is paid to ergonomics. Most renowned companies incorporate
Corresponding author: Tel.: +49 3641 943171.
E-mail
Scholl).

addresses:

armin.scholl@uni-jena.de,

a.scholl@wiwi.uni-jena.de

0377-2217/$ - see front matter  2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2011.01.056

(A.

methods for ergonomic risk estimation of working places in their


production routine (Toyota Verication of Assembly Line at Toyota,
GM-UAW at General Motors, AP-Ergo at Volkswagen to name a
few). If ergonomic risks are detected, re-balancing of the assembly
line is recommended as an effective method in the short-run (Hilla,
2006).
Ergonomic aspects have been barely considered in assembly
line balancing literature, though they are becoming increasingly
important in practice. Few articles on this topic are those of
Miralles et al. (2008) and Costa and Miralles (2009), who introduce
and analyze a problem of assigning workloads to stations and to
workers with different (dis-) abilities. Another article, written by
Carnahan et al. (2001), examines an assignment of a certain class
of tasks gripping tasks and their inuence on fatigue and
recovery dynamics of workers. However, to our best knowledge,
no attempt has been made yet to incorporate ergonomic risk estimation methods used in practice into assembly line balancing
models, though they are considered important by manufacturers.
To close this gap, we address this important question in the
present study. We provide an overview of some methods for ergonomic risks estimation, which are recommended and utilized in
practice. Most of those methods are based on nonlinear functions
such that incorporating them into state-of-the-art line balancing
models and (exact) solution procedures is not straightforward.
We propose different ways to model ergonomic aspects and a
two-stage heuristic approach, based on the well-known exact balancing procedure SALOME and the heuristic meta-strategy simulated annealing. By means of this heuristic approach, we can
achieve a signicant reduction in ergonomic risks of workplaces
at low computational cost even without increasing manufacturing
capacity, i.e., number of workstations (and workers). The proposed

278

A. Otto, A. Scholl / European Journal of Operational Research 212 (2011) 277286

two-stage heuristic approach, furthermore, allows for a controllable increase in manufacturing capacity considering the trade-off
between increased costs from adding stations on the one hand
and reduced ergonomic risks on the other hand.
We precede with an overview of ergonomics tools in Section 2.
A line balancing problem incorporating ergonomic risk factors,
ErgoSALBP, is described and modeled in Section 3. In Section 4,
we propose a two-stage heuristic, which is tested in comprehensive computational experiments in Section 5. A discussion in Section 6 concludes the paper.
2. Methods for estimating ergonomic risks
In the mandatory Appendix D.1 to 1910.900 of Final Ergonomics Program Standard, the Occupational Safety and Health
Administration (2000; OSHA for short) provides a list of methods
recommended for the estimation of ergonomic risks of workplaces.
In this section, we provide a brief description of selected methods
recommended by OSHA for application in assembly line production
the revised NIOSH (the National Institute for Occupational Safety
and Health) equation and the job strain index; the method OCRA
(OCcupational Repetitive Action) recommended by European
Norms on repetitive actions (EN 1005-5, 2007) and the EAWS
(European Assembly Worksheet) method, which was created for
and adapted by several European rms that employ an assembly
production system.
Throughout the paper, we will use an example of an assembly
line, the precedence graph for which is given in Fig. 1. The graph
consists of n = 11 tasks i = 1, . . . , n with task times ti to be executed
on each workpiece at a workstation during the cycle time of
c = 63 seconds. Every task involves several actions of upper limbs,
while some of them demand application of forces (see Table 1).
2.1. Risk estimation for manual handling: revised NIOSH equation
The NIOSH equation was developed in 1981 by the National
Institute for Occupational Safety and Health for risk estimation of
working conditions, where manual handling activities are the main
source of risk and lifting comprises more than 90% of manual handling activities (Waters et al., 1994).
The NIOSH equation communicates a lifting index LI that shows
the relation of the current load weight to the recommended load
weight limit:

LI

Load weight
Recommended weight limit

The higher the lifting index, the higher percentage of the workforce
is likely to be under risk for developing low back pain. The recommended weight limit is calculated depending on lifting conditions
TS, e. g. vertical travel distance of hands or degree of asymmetry
in posture, and the frequency of lifting FM:

Recommended weight limit TS  FM

The frequency multiplier FM is calculated based on the average


number of lifts per minute. It takes into account the duration of
6

2
18

18

15

10
12

15

11
15

21

ti

Table 1
Example of an assembly line. Task description for the right hand. Cycle time is
63 seconds.
TaskNo.

Task time
(seconds)

Actions

Posture

Average force, %
of max force
capacity (MFC)

1
2
3
4
5
6

18
6
15
21
3
6

8
2
5
6
5
2

Hand grip (wide)


Elbow exion > 60
Elbow exion > 60
Elbow exion > 60
Hand grip (wide)
Neutral posture

Neutral posture

8
9
10
11

18
15
15
12

3
4
3
11

Dorsal exion
Neutral posture
Dorsal exion
Dorsal exion

20%
5%
Insignicant
10%
33%
1 lifting of 17 kg
(avg. force of 70%)
1 lifting of 15 kg
(avg. force of 40%)
20%
33%
25%
10%

Table 2
Frequency multiplier FM for 28 hours of continuous lifting and lift height P30 cm.
Frequency: lifts/min

60.2

0.5

...

FM

0.85

0.81

0.75

0.65

0.55

0.45

...

the lifting activity, as well as the vertical height of the lift from
the oor. In Table 2, we present frequency multipliers for 28 hours
of continuous lifting and vertical lift height of 30 cm or more. TS
considers task specic parameters and indicates the maximal recommended weight of the load that can be lifted by healthy workers
under certain lifting conditions. For example, under the ergonomically most favorable lifting conditions (e. g. when the weight is held
close to the body), TS is equal to 23 kg.
In our example, let us assume that tasks 6 and 7 are performed
on the same station. The worker lifts under ergonomically favorable lifting conditions a 17 kg and a 15 kg load in each cycle of
63 seconds (see Table 1). The task specic parameter TS for both
cases of lifting has the ideal value of 23 kg and the frequency multiplier FM for both cases is 0.7557 (60/63 = 0.9524 lifts per minute,
the value of FM is retrieved from Table 2 by interpolation). So, the
recommended weight limit is 17.38 kg and the resulting lifting
indices are 0.98 for task 6 and 0.86 for task 7.
In case of several lifting tasks, we compute the composite lifting
index CLI as follows:


 

CLI LI11 LI21;2  LI21 LI31;2;3  LI31;2   

LIj1;...;i is calculated for the lifting task j based on the cumulated frequency of the tasks 1, 2, . . . , i. Tasks are numbered in non-increasing
order of their individual lifting indices LIjj . Generally, composite
CLI 6 1 is considered to be acceptable.
For a station load consisting of tasks 6 and 7, we get LI11 0:98
and LI21 0:86 as explained above as well as LI21;2 0:99, which
corresponds to two lifts of 15 kg in 63 seconds, so that the composite index CLI amounts to 1.11. Usually, this work load is considered
unacceptable.
Similar Methods. Several other methods are constructed
according to the logics of the revised NIOSH lifting equation,
e.g., the Siemens method (Bokranz and Landau, 2006). Additionally, the Siemens lifting index takes into account FI, a factor that
is dependent on demographic characteristics and tness of the
worker:

i
LI

Fig. 1. Example of a precedence graph.

Load weight
Load weight

Recommended weight limit


FI  FM  TS

279

A. Otto, A. Scholl / European Journal of Operational Research 212 (2011) 277286

The Siemens lifting index is interpreted according to the 3-zones


principle: LI 6 0.85 is a low-risk zone (green), 0.85 < LI 6 1 is a zone
of possible risks (yellow) LI > 1 and is a zone of high risks (red).

Table 4
Values of the force multiplier FoM.
Average force, % MFC
FoM

5%
1

10%
0.85

20%
0.65

30%
0.35

40%
0.2

P50%
0.01

Repetitive handling at high frequency is evaluated by OCRA index (Occhipinti, 1998; EN 1005-5). OCRA index considers tasks
done by upper limbs and ergonomic analysis is performed separately for each hand. OCRA index is calculated in the following
way:

OCRA index

Actual frequency
Recommended frequency

The higher OCRA index is, the higher is the risk of musculoskeletal
diseases of upper limbs. In European norms CEN it is advised to consider risk levels of OCRA 6 2.2 as acceptable (green), those between
2.3 to 3.5 as conditionally acceptable (yellow) and levels of OCRA
above as unacceptable (red).
Actual and recommended frequencies by convention are measured in number of repetitions per minute. The recommended frequency is computed as follows:

Recommended frequency OS  PM  FoM  RM  AdM

Here OS denotes organization specic parameters (e.g. duration of


short-cyclical activity within the shift). In the ergonomically most
favorable conditions for repetitive activities performed during an
eight-hour shift, OS achieves the best case value of 18 repetitions
per minute. In less ideal conditions, the value is reduced
accordingly.
PM is a multiplier for posture. Postures for upper limbs are classied by anatomical segments involved (shoulder, elbow, wrist and
hand) and the awkwardness of the posture (example of values is
given in Table 3). A multiplier for each unfavorable posture category is calculated depending on the share of the cycle time spent
there. The lowest value over all posture categories in the station
load (i.e. signaling the highest risk) is assigned to PM. RM is a repetitiveness multiplier taking value 0.7 if the cycle time is less than
15 seconds and/or same actions of upper limbs are performed for
more than 50% of the cycle time, otherwise RM equals to 1.
FoM is a force multiplier depending on the average forces applied by upper limbs (see Table 4). Presence of additional risk factors (such as working at cold temperatures) are captured by the
additional factors multiplier AdM. The value of AdM depends on
the duration of the workers exposure to these additional risk factors (see Table 3). The dynamics of the different multipliers is represented in Fig. 2.
Let in our example tasks 1 to 5 belong to the same station. Here
we will calculate OCRA for the right hand. The work load at this
station involves a wide hand grip of the right hand for 33% of the
cycle time (PMhand = 0.7) and elbow exion of more than 60 for
67% of the cycle time (PMelbow = 0.7). We choose the posture with
the lowest posture multiplier and set PM = 0.7. Further, the
(time-weighted) average force applied during the cycle comprises

Table 3
Example values of the posture multiplier PM and the additional factors multiplier
AdM.
Duration, share of cycle time

1/4

1/3

1/2

2/3

3/3

PM for severe postures (e.g. dorsal exion


P45, wide hand grip)
AdM
PM for mild postures
(e.g. elbow exion >60)

0.7

0.65

0.6

0.5

1
1

0.95
1

0.925
1

0.9
0.7

0.8
0.6

Multiplier

2.2. Risk estimation for upper limbs: OCRA

1
0.8
0.6
0.4
0.2
0

FoM
PM, severe
PM, mild
AdM
0%

20%

40%

60%

80% 100%

RM

Duration in % of cycle time or


average force in % of MFC
Fig. 2. Multipliers of OCRA equation.

11% (rounded up from 10.62) of the maximal force capacity. Hence,


FoM = 0.83 (by interpolation). As ideal working conditions are assumed, OS is set to 18. No high repetitiveness (RM = 1) and no additional risk factors (AdM = 1) are present. The total number of
actions of upper limbs is 26 (sum of number of actions of tasks
15; see Table 1). Per minute, we get the rounded value 25 and,
25
thus, OCRA 180:70:83
2:4, which corresponds to a yellow working place.
Similar Methods. The job strain index by Moore and Garg (1995)
uses a methodology similar to that of OCRA with additional parameters speed of work Sp and duration of strain D:

Strain Index Actual frequency  OS  PM  FoM  Sp  D

2.3. European Assembly Worksheet


The European Assembly Worksheet (EAWS) underlies ergonomics tools used by such automobile producers as Volkswagen and
FIAT and has a similar structure as some other methods applied
in manufacturing: NPW used by Opel, DesignCheck used by Daimler and IAD-BkB widely used in German metal and electrical goods
industries.
EAWS assesses postures, action forces, manual material handling as well as other whole-body risk factors and repetitive loads
of the upper limbs (for details see Schaub et al., 2010). The results
of EAWS estimation are two aggregate risk values: risk points (RP)
for the whole body and risk points for upper limbs.
The higher these risk points are, the higher is the risk of musculoskeletal diseases for a worker. It is recommended to interpret a
risk value of 025 points as a low risk zone (green), of 2650 points
as a zone of possible risks (yellow) and above 50 points as a high
risks zone (red).
Risk points for the whole body are computed from four
indices:

RP whole body PI MMHI FI Extra points

Posture index PI sums up risk points for specied kinds of postures.


Risk points for each posture depend nonlinearly on frequency or
duration of the posture within the cycle time. Additional points
could be given for twisting, lateral bending and extended reaching.
Manual material handling index MMHI is a sum of risk points RP
for lifting, holding, carrying and pushing or pulling. These constituents of the MMHI as well as action forces index FI depend on the
force level multiplier FoM corrected by work conditions and duration of exposure, expressed either in duration, frequency or length
points:

280

A. Otto, A. Scholl / European Journal of Operational Research 212 (2011) 277286

RP lifting FoM  Frequency points

RP holding FoM  Duration points

10

RP carrying FoM  Length points

11

RP pushing or pulling FoM  Length points

12

FI FoM  Frequency points or Duration points

13

Extra points are given to risk factors that were not or not enough
considered in the EAWS calculations above (e.g. working on moving
objects).
The second aggregate risk value of EAWS, risk points for upper
limbs, is based on force points Fo and posture points Po, that depend nonlinearly on the duration of exposure, as well as additional
points Ad. Organization specic parameters OS depend on the shift
duration, as well as duration and frequency of rest pauses within
it:

RP upper limbs OS  Fo Po Ad

14

Table 5
Alternative balances of the assembly line described in Table 1 (OCRA index for the
right hand).
Station
load
First balance
Station 1:
1, 2, 3, 4, 5
Station 2:
6, 7
Station 3:
8, 9, 10, 11

Worst posture,
duration in
% of cycle time

Actions PM FoM OCRA NIOSH


Avg.
index index
force, per
%MFC minute

Elbow exion, 67 11

25

0.7 0.83 2.4

Neutral postures

1.1

20

0.6 0.61 3.1

19

19

0.7 0.68 2.2

0.98

24

19

0.52 2.0

0.86

10

12

Dorsal exion, 71 21

Second balance
Station 1:
Dorsal exion 29
1, 2, 3, 6, 8
Station 2:
Elbow exion 33
4, 5, 7, 9, 10
Station 3:
Dorsal exion 19
11

0.80 0.3

0.6

2.4. General assumptions for modeling ergonomics methods


Having examined several widespread ergonomics methods, we
now formulate the following assumptions for modeling the estimation of ergonomic risks.
Let F = F(Sk) be an ergonomic risk estimation function, e.g. OCRA
index, that depends on a set of tasks Sk assigned to station k. Each
task i is characterized by the levels of exposure to risk factors, e.g.
average force applied or time in % of cycle time spent in a certain
awkward posture. The ergonomic risk evaluation function has
the following properties:
(a) Ergonomic risk of a station k does not decrease if a task is
added to it, since an alternative of any task is an idle time,
where the worker is assumed to take a neutral posture and
apply no forces (Bk = set of tasks additionally assignable to
k):

FSk 6 FSk [ fjg;

8j 2 Bk

15

(b) Ceteris paribus, the ergonomic risk estimation function F() is


non-decreasing and nonlinear in the levels of exposure.
Remark 1. Though nonlinearity complicates incorporating ergonomic aspects in line balancing models, it is a fundamental (and
intuitive) assumption which cannot be omitted as all known methods are based on nonlinear risk estimation functions. Levels of
exposure are aggregated in ergonomic risk estimation functions
by a nonlinear aggregation function in NIOSH, multiplication in
OCRA, mixture of multiplication and addition in EAWS.
3. Assembly line balancing considering ergonomic risks
Task assignment to stations can substantially inuence the level
of ergonomic risks at the workplaces, even keeping such protability parameters as number of stations and cycle time unchanged.
Moreover, since better ergonomics may also decrease the defect
rate (Eklund, 1995; Gonzlez et al., 2003) and the number of days
of sick leave, incorporation of ergonomics into assembly line balancing could even improve protability of production.
Colombini and Occhipinti (2006) report results of a manual
assembly line re-balancing, where the designer managed to reduce
the number of yellow places by 25% without increasing the number of stations and the cycle time. Let us illustrate this effect by
balancing the assembly line from Fig. 1 and Table 5. The rst balance shows substantial ergonomic risks present in all three stations: two stations are yellow according to OCRA index, station 2
shows risks for lower back revealed by the revised NIOSH equation.
The risks are due to combining tasks with exposure to the same

risk factors same awkward postures, lifting activities in the same


station. If we prohibit combination of such tasks, we can receive
the second balance that has solely green working places.
3.1. Assembly line balancing problem
The assembly line is a production system, where a set of tasks
V = {1, . . . , n} with operation times ti (for i 2 V) are distributed
among a set of (work)stations W = {1, . . . , K} arranged in a sequential
order. The workpieces are launched down the line at a xed rate
such that each station has access to every workpiece for a constant
time span. Within this cycle time c, the worker operating at the station has to perform all the tasks contained in the station load Sk  V,
P
i.e., the station time tSk i2Sk t i must not exceed the cycle time
c. This process is repeated for every new workpiece cyclically.
Due to technological or organizational conditions, the order of
performed tasks is restricted by precedence relations (i, j), meaning
that a task i 2 V must be executed before another task j 2 V. The
production process can be summarized by a non-cyclical digraph
G = (V, E, t) with node set V, called precedence graph, where
E = {(i, j)ji 2 V, j 2 Fi} is the set of arcs representing direct precedence relations (see Fig. 1). Fi and Pi denote the set of direct followers and direct predecessors of task i, respectively. The node weights
represent the operation times ti. Typically, in manufacturing we
deal with mixed-model production, in such case a joint precedence
graph with average task times is to be constructed (Boysen et al.,
2009).
The Assembly Line Balancing Problem (ALBP) is to assign tasks to
stations such that cycle time and other restrictions as well as precedence relations are met and some time-, capacity-, cost- and/or
prot-oriented goals are optimized. A feasible task assignment is
called (line) balance.
The most basic and classical version of ALBP is called Simple
ALBP of type 1 (SALBP-1); it minimizes the number of stations subject to a xed cycle time (cf. Scholl and Becker, 2006). A large number of generalizations have been developed to model the
production process more realistically; they include assignment
restrictions, multiple products, parallel stations or workplaces,
etc. (cf. Becker and Scholl, 2006; Boysen et al., 2007; Becker and
Scholl, 2009; Scholl et al., 2008, 2009, 2010).
In this article, we investigate and extend SALBP-1 as a widely
known basic problem (see Patterson and Albracht, 1975; Baybars,
1986; Scholl and Becker, 2006, for its basic assumptions). Nevertheless, our ndings can be easily applied to any generalized ALBP.

281

A. Otto, A. Scholl / European Journal of Operational Research 212 (2011) 277286

Given an upper bound K on the number of stations K, we can


compute the earliest station Ei and the latest station Li K, to which
task i 2 V is assignable (Saltzman and Baybars, 1987). This denes
FSi, the set of stations to which task i is potentially assignable:
FSi fEi ; Ei 1; . . . ; Li Kg . The other way round, Bk is a set of tasks
that are assignable to station k: Bk c; K fi 2 Vjk 2 FSi g .
Let xik be a binary assignment variable:

xik

1 if task i is assigned to station k


0

otherwise

8i 2 V;

k 2 f1; . . . ; Kg

SALBP-1 can be formulated as a binary linear program (Scholl, 1999;


Chapter 2):

Minimize Kx

k  xnk

16

k2FSn

subject to

xik 1 8i 2 V

17

k2FSi

8k 2 f1; . . . ; Kg

ti  xik 6 c

18

i2Bk

k  xik 6

k2FSi

k  xjk

 P Ej
8i; j 2 E with Li K

19

k2FSj

otherwise ErgoSALBP may become economically inefcient. In order


to soften individual ergonomic risks, we adjust the worker rotation
schemes.
Hence, instead of ergonomic constraints, we may introduce
ergonomic risk measures into the objective function. In order to
model the trade-off between number of stations and ergonomic
risks, we construct a multiple objective function with K(x) being
the number of stations required by balance x, n(F(Sk)j"k) being
some aggregation function of ergonomic risks, which depends on
the risk levels of the station loads and x being some positive
weight:

Minimize K 0 x Kx x  nFSk jk 2 f1; . . . ; Kxg

22

subject to (17)(20).
Of course, if it is possible to measure the economic effect of
ergonomic risks and evaluate the trade-off with the lines capacity,
then x should reect this trade-off. Otherwise, we may choose x
in such a way that the model allows a reduction in levels of ergonomic risks only if it does not increase the number of stations (lexicographical order of objectives).
Depending on the way of aggregating ergonomic risks of all stations, we develop different versions of ErgoSALBP.
If worker rotation is regularly possible, it seems to be useful to
minimize average ergonomic risks (ErgoSALBP-A with A for
average):
Kx

xik 2 f0; 1g 8i 2 V;

k 2 FSi

20

Eq. (16) is the objective function which minimizes the index of the
station to which the last task n (introduced as the unique sink of the
graph, with duration tn = 0) is assigned. It is required that each task
is assigned exactly to one station (17), whereby the station times
must not exceed the given cycle time (18). Precedence relations
are enforced by conditions (19). The binary variables are dened
in (20).
3.2. ErgoSALBP
We extend SALBP-1 by suggesting different ways of incorporating ergonomic risks. The resulting models are abbreviated by
ErgoSALBP.
The most direct approach to model ergonomic risks, stemming
directly from ergonomic risk evaluation methods that provide critical risk levels, is via introduction of ergonomic constraints. In this
case, we complement the constraints (17)(20) of SALBP-1 by constraints (21), where F is an ergonomic risk estimation function with
the properties discussed in Section 2.4 and Erg is a coefcient equal
to the accepted maximal level of ergonomic risks. The resulting
problem is called ErgoSALBP-C (C for constraints).

FSk 6 Erg

8k 2 f1; . . . ; Kg

21

However, in some cases ergonomic constraints (21) might be economically inefcient. First of all, an optimal solution of ErgoSALBP-C may require additional workstations due to tight critical
risk levels, which means a substantial increase in variable and xed
costs for the rm. Unfortunately, despite recent progress in this
direction, no widely accepted estimates of benets either to the
health of workers or to the rms economic results have been established. Due to this uncertainty and space restrictions, rms are
unwilling to increase number of working places. Secondly, generally, it is not useful to impose the same critical level of ergonomic
risks on all the stations. Consider, for example, operations on the
undercarriage in the automobile industry, which are performed in
awkward postures overhead or above shoulders. These operations
are extremely hard to combine with other tasks. We may prefer
balances with a higher level of ergonomic risks at such stations;

1 X

FSk
Kx k1

23

If rotation is impossible or restricted, it is recommended to smooth


risks among the stations in order to distribute physical load among
workers at a similar low level (ErgoSALBP-S with S for smoothing):

sP
Kx
2
k1 maxfFSk  Erg; 0g
n
Kx

24

In manufacturing practice, often a trafc-lights system is used. Here


the number of workstations with possible (yellow) and with signicant (red) ergonomic risks are seen as relevant measures to be
minimized. The levels of risk that mark the frontier between
green and yellow and yellow and red stations, we call here
Erggreen and Ergyellow, respectively. The following binary variables
indicate if a station k 2 f1; . . . ; Kg is yellow or red:


yk

rk

1 Erg green < FSk 6 Erg yellow


0

otherwise

1 Erg yellow < FSk


0 otherwise

The aggregation function might be computed with xyellow << xred.


The resulting problem is called ErgoSALBP-N (N for numbers):

n xyellow 

X
k

yk xred 

rk

25

Of course, further model variations of ErgoSALBP are possible, but


the introduced ones seem to be particularly suited for application
in practice as discussed above.
4. Two-stage heuristic procedure for ErgoSALBP
The different versions of ErgoSALBP are NP-hard optimization
problems, because they are generalizations of the NP-hard
SALBP-1 (cf. Wee and Magazine, 1982; Scholl, 1999; Chapter
2.2.1.5). Moreover ErgoSALBP cannot be solved by available assembly line balancing methods as, in general, it contains nonlinear
ergonomic constraints or a nonlinear objective function.

282

A. Otto, A. Scholl / European Journal of Operational Research 212 (2011) 277286

Therefore, we propose a two-stage heuristic based on a wellknown procedure for SALBP-1 (stage 1) and the simulated annealing technique supplemented by a local search algorithm (stage 2).
In Section 5, we show that with this two-stage procedure it is possible to achieve a signicant decrease in the levels of ergonomic
risks at low computational costs. Re-balancing belongs to the
short-term planning and solution proposals for (different) line balances are expected to be calculated in a short time in manufacturing. Therefore, the time aspect is extremely important here. Based
on a medium-term setting (given the number of stations), a rebalancing is required that optimizes the balance considering the
current model-mix in an ergonomically favorable manner.
Due to its exibility, it is possible (with minor modications) to
apply the proposed heuristic to all four (and further possible)
settings of ErgoSALBP described in Section 3.2.
4.1. Two-stage heuristic procedure for different versions of ErgoSALBP
We propose a two-stage procedure that employs the exact solution procedure SALOME (Scholl and Klein, 1997, 1999) to nd a
solution with the minimal number of stations Kinitial on the rst
stage, afterwards a simulated annealing technique is applied to improve the distribution of ergonomic risks among the number of
working stations determined on the rst stage. For ErgoSALBP-C,
if red and yellow stations are still not eliminated, we (repeatedly)
increase the number of stations and apply the second stage anew.
Simulated annealing (SA) is a heuristic method designed to
avoid getting stuck in a local optimum; it has been successfully applied to a number of sophisticated general assembly line balancing
problems (Suresh and Sahu, 1994; Kim and Kim, 1996; McMullen
and Frazier, 1998; Kaji and Ohuchi, 1999; Erel et al., 2001; Vilarinho and Simaria, 2002; Baykasoglu, 2006; zcan, 2010). Simulated
annealing received its name due to the analogy to the physical
annealing of solids (see a detailed description of the simulated
annealing heuristic in, e.g., Kirkpatrick and Gelatt, 1983; Glover
and Greenberg, 1989; Eglese, 1990).
4.2. Simulated annealing procedure (second stage)
Neighborhood generation procedure. We generate a neighbor
(test) solution x0 to the current solution x by performing a shift
(moving a task to another station) or a swap (interchanging a task
with a randomly chosen task). In order to facilitate changes in the
loads of stations with high ergonomic risks, we divide stations into
two types and assign probabilities for choosing a task from each
type of them. Thus, with probability p a task is chosen from a station with ergonomic risks exceeding the accepted level Erg and
with 1  p from an ergonomically favorable station.
Then, for the chosen task we perform a swap with probability q
and a shift with probability 1  q. In case no swap is possible at all,
a shift is performed. Both for shifts and swaps, we require compliance with the cycle time constraints (19) and the precedence relations constraints (20).
Energy function. As in stage 2 the number of stations is xed, we
use the aggregation function n of the version of ErgoSALBP to be
solved as energy function of the SA approach. In case of ErgoSALBP-N, we set the weights such that xyellow 6 Kxred in order to
initial
realize a lexicographic ordering which always prefers a decrease
in number of red stations no matter how many additional yellow stations would be needed. For ErgoSALBP-C which denes
no aggregation function, we measure the energy by some function
that denes the deviation from the specied ergonomic
constraints.
Rejection probability of the test solution. If the energy of the test
solution x0 is lower than that of the current solution x, i.e., objective
function value improves, then we update the test solution to the

current one: x : x0 . Otherwise, if the test solution is worse than


the current one, we reject it with probability r increasing with
the difference in energy between the test and the current solution
dn = n(x0 )  n(x) and decreasing in the temperature level T as computed by (26).

r 1  edn=T

26

Temperature dynamics. We start the SA algorithm at a very exible


temperature level T1 = 2 so that the initial probability to accept an
inferior solution with dn = 1 is 60% and dn = 0.1 is 95%. Each time
after 500 iterations, we reduce the temperature level by 5%.
Simple local search algorithm. For each updated current solution
we run a simple local search algorithm consisting of a shifts- and a
swaps-block. In the shifts-block, we perform the shift with the largest improvement one after the other as long as an improving shift
exists. Afterwards, in the swaps-block the same is done with all
improving swaps. In both blocks we prohibit violation of cycle time
and precedence constraints.
Update of the best solution. Just after having applied the local
search described above, we compare the energy (objective function) value of the current solution with the energy value of the best
solution. If the current solution has a lower energy value, then we
store it as the new best solution. However, we do not change the
current solution for the SA algorithm, i.e. we continue our calculations with the current solution we had prior to the application of
local search. In the other case, we would push the SA closer to local optima, making it harder to overcome local optimality.
Stopping criterion. We stop our SA either when ergonomic risks
at all stations are below or equal to the acceptable level Erg since
it is the aim of any ergonomic improvement, or when the temperature level reaches a rather prohibitive level Tmin = 0.002. At this
temperature level, probability r of accepting an inferior solution
with a modest dn = 0.001 is the mere 60%. Due to these settings,
we perform at most 67,500 iterations in each run of our SA
procedure.
Increasing the number of stations (for ErgoSALBP-C only). If after
performing the SA algorithm, the best solution still violates some
of the ergonomic constraints (21), then we increase the number
of stations by one. We place an empty new station between two
neighboring stations with highest sum of ergonomic risks in order
to facilitate a quick search for a solution with all green stations.
Afterwards, we apply the SA algorithm to this extended solution.
The process of adding a station and applying SA is repeated until
no violations of (21) remain.
5. Computational experiments
We conduct two sets of experiments. First, we would like to nd
out, how large the improvement in ergonomics is that could be
achieved without increasing the number of workstations and
how large the computational costs of it are. Secondly, we examine
how many additional stations we need to bring ergonomic risks
under the acceptable maximal level in order to gure out the
trade-off between capacity and ergonomics.
5.1. Data generation
For our experiments, we use the OCRA index an extensively
published and analyzed measure for ergonomic risks. However,
note, that any other ergonomics method can be directly applied
within the described heuristics procedure and similar results are
expected.
As a basis for generating suitable test data, we use the wellknown benchmark data set for SALBP-1 (cf. Scholl, 1999, Chapter
7.1) available for download at www.assembly-line-balancing.de.

A. Otto, A. Scholl / European Journal of Operational Research 212 (2011) 277286

We skipped one instance (Mertens with cycle time c = 6, n = 7,


Kinitial = 6) due to very few number of tasks per station in the
optimal solutions, so that our basic data set contains 268 SALBP-1
instances with up to 297 tasks.
Following a rule of thumb in statistics that about 30 observations per cell are required (Wilson VanVoorhis and Morgan,
2007), we generated 27 replications of ergonomic parameters for
each SALBP-1 instance. In total, we constructed 27268 = 7236 instances for our analysis.
For each task in each instance, we generated values for the following parameters: frequency of actions, posture, force, repetitive
actions and additional factors (see Section 2.2 and Table 6). We
control for ergonomic risks, so that if any task is the only task assigned to a station, this station shows a risk estimation below the
acceptable level Erg.
Further, we make sure that for each instance its optimal SALBP1 solution, that minimizes the number of stations and is found in
stage 1, contains at least one station with the level of risk above
Erg. Also, assuming that all tasks imply symmetric operations for
both hands, we control for OCRA points only for one of the two
upper limbs to simplify the analysis.
To model the frequency of actions, we normalize the cycle time
to be equal to one minute in each problem set. For each task we
randomly generate the frequency of actions per minute, so that
the number of actions within a task depends on this generated
speed of work and the relative share of the task time within the cycle. For example, let for a task with task time 18 be generated a frequency of actions of 20 min1. Then, the number of actions that has
to be performed during this task is equal to d20  18
e6.
60
We model presence of additional factors, as well as postures by
three variables: Probability of presence of an additional factor or a
posture within a task and its duration within the task (uniformly
distributed between the minimal and the maximal duration value).
No task can have both mild and severe postures of the same type,
where the type of posture corresponds to the same anatomical segment of the upper limb. Conversely, execution of any task may require several postures of different types. For example, let us
generate the duration of posture A for a task with task time 18.
The probability that this posture will be severe, mild or not present

Table 6
Modeling ergonomic parameters of tasks.
Risk factor

Severe posture: type A


Mild posture: type A
Severe posture: type B
Mild posture: type B
Severe posture: type C
Mild posture: type C
Additional factor (e.g. vibrating
instruments)
Repetitive action: type A
Repetitive action: type B

283

is 33% in each case (cf. Table 6). After a random generation we may
receive, for example, that severity of posture A is mild and duration
is 70% of the task time. In a similar way, we generate data for postures B and C.
Average force during a task is measured in per cent of maximal
force capacity and is randomly assigned to be equal to one of ve
values as given in Table 6.
Further, we assume that there are two types of actions (called
repetitive action in Table 6) that may lead to high repetitiveness
within a workstation. We model the frequency of such actions,
similarly to those for postures and additional factors, with three
variables probability of presence, minimal and maximal duration
within the task.
We assume an 8-hour shift and following the recommendations
of the European norm EN 1005-5, we set OS to 18, Erg and Erggreen
to 2.25 and Ergyellow to 3.5. Further, we call workstation k with
OCRA index F(Sk) 6 Erggreen green, that with Erggreen < F(Sk)
6 Ergyellow yellow and that with F(Sk) > Ergyellow red.
In the generated data set, we received about 22% yellow and 2%
red workstations on average per instance for its optimal SALBP-1
solution with average OCRA index for a station varying from 0.9
to 3.09. Received by our data generation, balances show realistic
and even rather mild ergonomic risks. The share of yellow and
red workstations in our instances roughly corresponds to the situation that we observed at automotive producers. In contrast,
Occhipinti and Colombini (2007) in their study of 22 groups of
workplaces from different industries found average OCRA index
varying from 2.8 for oven assembly to 41.7 for upholstering of
car seats. Note that higher average level of ergonomic risks in the
initial balance would even more strengthen the positive effect
from ErgoSALBP.
After initial experiments on the Lutz 1 instances with 32 tasks,
we chose the following parameters for the two-stage heuristic:
probability of a swap move q = 50% and probability to choose a task
from a station with unacceptable ergonomic risks p = 80%. For the
objective function (25), we set xred K1 and xyellow K12 with K
denoting the trial number of stations for the second stage. The
computational experiments were conducted on a 2.83 GHz, Intel
Core2 Q9550 processor with 4 GB of memory. The heuristic procedure was programmed in Borland Delphi 7.0, whereby the code
was not parallelized and, hence, utilized only one of four cores
per run.

Duration, % of
task time

5.2. Experiment 1: ErgoSALBP with multiple objectives

Min
(%)

Max
(%)

33
33
20
20
33
33
10

50
50
50
50
50
50
50

80
80
80
80
80
80
80

5
5

40
40

80
80

In our rst experiment, we apply the two-stage heuristic proposed in Section 4 to reduce ergonomic risks without increasing
the number of stations. We run the experiments for the three variants of ErgoSALBP (-A, -N, -S) with the objective functions (23)(25)
in order to nd out, which model is best suited for risk reduction.
As performance measures, we utilize the number of yellow stations, the number of red stations, the average OCRA station index
(average risk level) and the deviation of risk values (smoothness).
Table 7 summarizes the results, which are given as per cent
improvements (relative reductions) compared to the initial (SALBP-1) solution.
At rst, we consider values averaged over the three versions of
ErgoSALBP (-A, -N, -S). We nd that, on average, 63% of the yellow
stations and 72% of the red stations can be avoided. Further, solutions only having green stations are found for 46% of the instances.
In other words, for our data set constructed with ergonomic estimations of stations very similar to those observed in automobile
industry, there exists almost a 50% chance to nd a balance with
only green stations without adding new stations, if we incorporate
ergonomic aspects into the balancing model. Of course, the model
used here is very simple and does not include zoning constraints,

Probability
(%)

Levels of exposure to risk factors

Probability (%)

Frequency of actions: number of actions per task, min1


10
20
30

33
33
33

Force: average force level within the task, %MFC


<1/20
1/10
1/5
1/3
1/2

40
25
30
4
1

284

A. Otto, A. Scholl / European Journal of Operational Research 212 (2011) 277286

Table 7
Optimization results of bi-objective ErgoSALBP.
No additional stations
added. ErgoSALBPA
Avg. improvement Number
of yellow
to initial SALBPstations
solution (%)
Number of
red stations
Avg. OCRA-index
Deviation of risks
Share of instances with
improvement (%)
Share of solutions with only green
stations found (%)

After one
station
added
Avg. results ErgoSALBPA

61 66 62 63

97

69 74 74 72

98

5 3 4 4
62 58 68 63
93 87 93 91

11
97

46 46 47 46

91

resource requirements etc. (cf. Boysen et al., 2007). However, similar results are expected, when such constraints are included in the
models.
The computation time is only about 1 second per instance on
average with a maximum of 25 seconds for some of the largest instances with 297 tasks.
For all three ErgoSALBP models, the computational efciency of
simulated annealing is about the same. There were no signicant
differences in computational times or the number of green solutions between the objective functions. In general, the objectives are
working in the same direction, trying to reduce ergonomic risks,
and their results are highly correlated, meaning that a solution
with lower number of yellow and red stations, as a rule, has lower
average OCRA index and a smoother distribution of ergonomic
risks. However, each model remains the most effective tool in
achieving its own objective. Thus, for example, ErgoSALBP-A for
70% of instances achieved the best result in terms of average OCRA
compared to average OCRA received by ErgoSALBP-S and ErgoSALBP-N.
In few cases, the effect of the objective function may be rather
pronounced. Thus, in ErgoSALBP-A, when we minimize the average
OCRA index, the heuristic may try to concentrate all the risk in a
fewer number of stations. For example, in a Lutz 1 instance (see
Table 8) the heuristic was not able to improve the risk level of
the red station, instead it created an additional yellow station by
decreasing risks in remaining (already) green ones. Further,
ErgoSALBP-N does not differentiate between stations with a higher
or lower level of ergonomic risks as long as they belong to the same
category (green, yellow or red). To the contrary, although
ErgoSALBP-S decreases risks in the ergonomically worst performing stations, it may increase risks in the remaining ones.
So, we conrm our recommendation given in Section 3.2: Due
to the similarity of results obtained by the different versions of

Table 8
Lutz 1 instance with n = 32, c = 1414, Kinitial = 11: impact of the objective functions on
solutions.

OCRA index of yellow


and red stations
Avg. OCRA index for
green stations
Avg. OCRA index for
all stations
Deviation of risks

Initial
solution

ErgoSALBPA

1 Red:
3.62
1.54

1 Red: 3.62
1 Yellow: 2.54
1.23

1 Yellow:
3.17
1.58

1 Yellow:
2.44
1.52

1.73

1.57

1.72

1.60

0.41

0.42

0.28

0.06

ErgoSALBP, the choice of the right model should be done according


to the organizational specicities of the assembly (e.g. possibility
of worker rotation).
To check the stability of results for the two-stage heuristic, we
compared its performance on a subset of the data set in the following settings: (1) Different stopping criteria: The maximal number
of iterations is doubled from 67,500 to 135,000 iterations. (2) Instead of one run of the two-stage heuristic, we perform ve runs.
The data subset consists of 1,500 instances randomly selected from
those for which no solutions with only green stations were found
in the part of the experiment described above. The results are summarized in Table 9.
The doubled number of iterations caused about a doubling of
the computational time. Although for a few instances the larger
number of iterations almost doubled the improvement of the
objective function relative to the initial solution, for the majority
of instances no improvement was found at all. On average the performance of the two-stage heuristics remained approximately the
same.
The same holds for the increased number of runs of the heuristic. The results remained very similar on average. The average coefcient of variation for objective improvement in the best solution
relative to the initial solution, which is the standard deviation of
improvement divided by the average improvement, did not exceed
4%. Just for 4.2% of the instances, a solution with all green stations was additionally found.
The results show that the heuristic is able to nd good and stable solutions in small computation times. Only small improvements are possible by increasing the computational effort.
If several risk factors, such as awkward postures and application
of forces, are present, they multiply negative effects of each other
in the OCRA index. This observation let us suggest that a higher initial level of risks of the line balance would generally lead to a higher absolute improvement in ergonomic risks. To check it on the
data, we run ordinary least square regressions for those of 27 ergonomic instances of 268 SALBP instances, for which a solution with
only green stations was not found. In total we run regressions for
161 SALBP instances (hence, 161 regressions), which have more
than 10 observations. Since the results of different versions of
ErgoSALBP are very similar, we used results of ErgoSALBP-A. We
regressed the absolute improvement in average OCRA index
P
FSinitial
FSbest

k
k
k
against the initial level of average OCRA index
K
P initial
FSk

k
. For 94% of regressions we received positive b-coefcients,
K
which conrm our hypothesis. 95% of these regressions were signicant according to F-tests on 0.5% signicance level.

5.3. Experiment 2: ErgoSALBP with ergonomic constraint


In the second experiment we used the two-stage heuristics to
nd solution that minimizes the number of stations for ErgoSALBP-C (21). Since the different bi-objective versions of ErgoSALBP in the previous experiment showed very similar results,
for our second experiment, we used the energy function of ErgoSALBP-A in each iteration of the simulated annealing part of the
heuristic, as average OCRA index is a common measure utilized
by ergonomists.
Like the previous one, this experiment required low computational costs: On average, we spent only 1.4 seconds per instance
with at most 26 seconds for an instance with 297 tasks. We had
to increase the number of stations on average by 0.7 or by 5% of
the initial number of stations to nd solutions with only green
stations. For those instances (54% of total), where an increase in
number of stations was required, we needed on average 1.2 additional stations or 9% of initial number of stations. After adding at

285

A. Otto, A. Scholl / European Journal of Operational Research 212 (2011) 277286


Table 9
Stability of the results for the two-stage heuristic.
ErgoSALBP

After doubling the number of iterations

After ve runs of the two-stage heuristic

Additionally found
green solutions, % of
trials

Avg. improvement of the energy


function of the best solution (%)

Additionally found green


solutions, % of instances

Avg. improvement of
the best solution (%)

Avg. coefcient of variation for the


improvement of initial solution (%)

A
N
S

0.1
1.8
1.3

0.3
4.8
2.7

1.9
1.6
4.2

0.3
6
9

4
3
3

Table 10
Dynamics of improvement in ergonomic risks with each additional station. Data for
169 instances that required at least three additional stations.
Avg. of total
normalized
ergonomic risks
P

Avg.
improvement
in total
normalized
ergonomic risks

1.92
1.85

0.07

1.75

0.10

1.71

0.05

FSk
k
K initial

Initial solution
No stations added to best
solution
One station added to best
solution
Two stations added to best
solution

most one additional station, for 91% of instances a solution with


only green stations was found; number of yellow and red stations were reduced by 97 and 98%, respectively, relative to the initial solution (Table 7).
For manufacturing, results from larger instances with 30 and
more stations in the initial (SALBP-1) solution are especially relevant. For such instances, where additional stations were required,
the average necessary increase in number of stations was about
4%. In other words, even if an increase in capacity is required, a
solution with ergonomic risks for all stations being under the
acceptable maximal level Erg can be found at average costs of 4%
increase in capacity.
In the previous section we have seen, that higher initial ergonomic risks generally lead to a higher absolute improvement in
ergonomic risks. This consideration let us expect that the improvement of ergonomic risks with each added station is likely to diminish. It can be seen if we look on instances that required at least 3
additional stations for the optimal solution of ErgoSALBP with
ergonomic constraints (169 or 2% of the tested instances). For
91% of these instances, the average improvement in total ergoP
nomic risks k Erg k was higher after the rst station added than
after the second. On average, the P
difference in improvement of
normalized total ergonomic risks

Erg k

K initial

after the rst additional

station solution is about two times higher than that after the second (Table 10).
6. Summary and Conclusions
In manufacturing, control of ergonomic risks at manual workplaces is a necessity commanded by legislation, care for health of
workers and economic considerations. Methods for estimating
ergonomic risks of workplaces are integrated into production routines at most rms using the assembly-type of production. Assembly line re-balancing, i.e., re-assignment of tasks to workers, is an
effective and, in case that no additional workstations are required,
inexpensive method to reduce ergonomic risks. In our article, we
show that even though most ergonomic risk estimation methods
involve nonlinear functions, they could be easily integrated into

assembly line balancing techniques at low additional computational cost. Our computational experiments on a data set, whose
ergonomic risks resemble the situation we observed in the automobile industry, indicate that re-balancing often leads to a substantial mitigation in ergonomic risks. Thus, for 50% of instances
a balance with acceptable level of risks for all the stations was
found without increasing the number of stations. In the cases,
where an increase in number of stations was required, such a solution was found after just about 4% increase in the line capacity.
Further, our computational experiments indicate that, in general,
with each new working station the improvement of ergonomic
risks diminishes.
For a manufacturer, each additional station means a signicant
investment in equipment, increase in variable costs rst of all, in
wages, not to mention utilization of space of the factory, which is
often a constraint resource. To perform such a step, a rm has to
make a careful estimation of benets from the reduction of ergonomic risks that one or several additional stations would bring.
Assembly line balancing can give an answer, how large the expected mitigation of ergonomic risks will be for each station, as
well as for the line as a whole. The above mentioned 4% increase
in line capacity, necessary to keep ergonomic risks in a reasonable
range (green), can be used by line managers as a rule of thumb.
Currently, an intensive research is running on the overall costs
of ergonomic risks of workplaces. Although these costs are rm
specic and a lot of work here is still to be done, we have rst insights, how large such costs could be. First of all, these costs include compensation claims and medical visits of workers. Here,
Occhipinti and Colombini (2007) estimated that a 1-point increase
in average OCRA index of the assembly line leads to additional
2.39% of workers suffering from upper-limb musculoskeletal disorders. Further, these are costs from deteriorated quality of work;
Falck et al. (2010) in their 8-week study of a car assembly found
the action costs originated in yellow and red workstations up
to 9 times higher than that caused at green stations.
The methodologies and routines for estimation the costs of
ergonomic risks for a rm are being developed. A decision-making
tool consulting on the reasonable measures for ergonomic
improvement of workplaces with high ergonomic risks is awaited
by manufacturers. Assembly line balancing models that incorporate ergonomic risks will be an essential part of such a tool.
A number of questions remain for future research. Accepted
estimates or methodology of estimation of the monetary equivalent of ergonomic risks is badly needed to justify measures for
improvement in workplace ergonomics. Further, not only assembly
line balancing, but also other production routines should be reexamined in light of possible inuences on ergonomic risks. One of it
is the order, in which different models of the product are processed
in the assembly line (sequencing), and its consequences for ergonomics; namely, how the rules of sequencing should be modied
to reduce workplace ergonomic risks. Another one is distribution
of relaxation allowances these are time add-ups on the task
time for rest and breaks. Relaxation allowances, a part of predetermined time systems, are widely used in assembly line production.
These add-ups are usually calculated proportionally to the task

286

A. Otto, A. Scholl / European Journal of Operational Research 212 (2011) 277286

time. As a plausible alternative, we could interconnect the relaxation allowances with some estimation of the worker fatigue.

Acknowledgements
This article was supported by the Federal Program ProExzellenz of the Free State of Thuringia.

References
Baybars, I., 1986. A survey of exact algorithms for the simple assembly line
balancing problem. Management Science 32, 909932.
Baykasoglu, A., 2006. Multi-rule multi-objective simulated annealing algorithm for
straight and U type line balancing problems. Journal of Intelligent
Manufacturing 17, 217232.
Becker, C., Scholl, A., 2006. A survey on problems and methods in generalized
assembly line balancing. European Journal of Operational Research 168, 694
715.
Becker, C., Scholl, A., 2009. Balancing assembly lines with variable parallel
workplaces: Problem denition and effective solution procedure. European
Journal of Operational Research 199, 359374.
Bokranz, R., Landau, K., 2006. Produktivittsmanagement von Arbeitssystemen.
(rst ed.). Stuttgart: Schffer-Poeschel Verlag, (Chapter 9).
Boysen, N., Fliedner, M., Scholl, A., 2007. A classication of assembly line balancing
problems. European Journal of Operational Research 183, 674693.
Boysen, N., Fliedner, M., Scholl, A., 2009. Assembly line balancing: Joint precedence
graphs under high product variety. IIE Transactions 41, 183193.
Bureau of Labor Statistics (2009, November 24). Nonfatal occupational injuries and
illnesses requiring days away from work, 2008. Economic news release.
Carnahan, B.J., Norman, B.A., Redfern, M.S., 2001. Incorporating physical demand
criteria into assembly line balancing. IIE Transactions 33, 875887.
Colombini, D., Occhipinti, E., 2006. Preventing upper limb work-related
musculoskeletal disorders (UL-WMSDS): New approaches in job (re)design
and current trends in standardization. Applied Ergonomics 37, 441450.
Costa, A.M., Miralles, C., 2009. Job rotation in assembly lines employing disabled
workers. International Journal of Production Economics 120, 625632.
Eglese, R.W., 1990. Simulated annealing: A tool for operations research. European
Journal of Operational Research 46, 271281.
Eklund, J.A.E., 1995. Relationships between ergonomics and quality in assembly
work. Applied Ergonomics 26, 1520.
EN 1005-5. 2007. Safety of Machinery Human Physical Performance Part 5: Risk
assessment for repetitive handling at high frequency. European Committee for
Standardization, work item ID 00122093.
Erel, E., Sabuncuoglu, I., Aksu, B.A., 2001. Balancing U-type assembly systems using
simulated annealing. International Journal of Production Research 31, 3003
3015.
Falck, A.-C., rtengren, R., Hgberg, D., 2010. The impact of poor assembly
ergonomics on product quality: A cost-benet analysis in car manufacturing.
Human Factors in Ergonomics & Manufacturing 20, 2441.
Glover, F., Greenberg, H.J., 1989. New approaches for heuristic search: A bilateral
linkage with articial intelligence. European Journal of Operational Research 39,
119130.
az, B., Torre, P.G., 2003. Ergonomic performance and
Gonzlez, B.A., Adenzo-D
quality relationship: An empirical evidence case. International Journal of
Industrial Ergonomics 31, 3340.
Hilla, W., 2006. Produktivitt und Ergonomie gemeinsam entwickeln. Erfahrungen
aus der betriebsrztlichen Ttigkeit bei Audi AG. Presentation at MTM
Bundestagung 26.10.2006, Stuttgart.
Kaji, T., Ohuchi, A., 1999. A simulated annealing algorithm with the random
compound move for the sequential partitioning problem of direct acyclic
graphs. European Journal of Operational Research 112, 147157.

Kim, J.-U., Kim, Y.-D., 1996. Simulated annealing and genetic algorithms for
scheduling products with multi-level product structure. Computers and
Operations Research 23, 857868.
Kirkpatrick, S., Gelatt Jr., C.D., 1983. Optimization by Simulated Annealing. Science.
McMullen, P.R., Frazier, G.V., 1998. Using simulated annealing to solve a
multiobjective assembly line balancing problem with parallel workstations.
International Journal of Production Research 36, 27172741.
Miralles, C., Garca-Sabater, J.P., Andrs, C., Cards, M., 2008. Branch and bound
procedures for solving the assembly line worker assignment and balancing
problem: Application to sheltered work centers for disabled. Discrete Applied
Mathematics 156, 352367.
Moreau, M., 2003. Corporate ergonomics programme at automobiles Peugeot
Sochaux. Applied Ergonomics 34, 2934.
Moore, J.S., Garg, A., 1995. The strain index: A proposed method to analyze jobs for
risk of distal disorders of the upper limbs. American Industrial Hygiene
Association Journal 56, 443458.
Occhipinti, E., 1998. OCRA: A concise index for the assessment of exposure to
repetitive movements of upper limbs. Ergonomics 41, 12901311.
Occhipinti, E., Colombini, D., 2007. Updating reference values and predictive models
of the OCRA method in the risk assessment of work-related musculoskeletal
disorders of the upper limbs. Ergonomics 50, 17271739.
Occupational Safety and Health Administration (2000). Final Ergonomics Program
Standard. Federal register #64: 65768-66078.
zcan, U., 2010. Balancing stochastic two-sided assembly lines: A chanceconstrained, piecewise-linear, mixed integer program and a simulated
annealing algorithm. European Journal of Operational Research 205, 8197.
Patterson, J.H., Albracht, J.J., 1975. Assembly line balancing: Zero-one programming
with Fibonacci search. Operations Research 23, 166172.
Saltzman, M.J., Baybars, I., 1987. A two-process implicit enumeration algorithm for
the simple assembly line balancing problem. European Journal of Operational
Research 32, 118129.
Schaub K., Caragnano, G., Britzke, B., Bruder, R., 2010. The European Assembly
Worksheet. In: Mondelo, P., Karwowski, W., Saarela, K., Swuste, P., Occhipinti, E.
(Eds.), Proceedings of the VIII International Conference on Occupational Risk
Prevention. Valencia 5. 7.5.2010.
Scholl, A., 1999. Balancing and Sequencing Assembly Lines, second ed. Physica,
Heidelberg.
Scholl, A., Becker, C., 2006. State-of-the-art exact and heuristic solution procedures
for simple assembly line balancing. European Journal of Operational Research
168, 666693.
Scholl, A., Boysen, N., Fliedner, M., 2008. The sequence-dependent assembly line
balancing problem. OR Spectrum 30, 579609.
Scholl, A., Boysen, N., Fliedner, M., 2009. Optimally solving the alternative
subgraphs assembly line balancing problem. Annals of Operations Research
172, 243258.
Scholl, A., Fliedner, M., Boysen, N., 2010. Absalom: Balancing assembly lines with
assignment restrictions. European Journal of Operational Research 200, 688
701.
Scholl, A., Klein, R., 1997. SALOME: A bidirectional branch and bound procedure for
assembly line balancing. INFORMS Journal on Computing 9, 319334.
Scholl, A., Klein, R., 1999. Balancing assembly lines effectively a computational
comparison. European Journal of Operational Research 114, 5058.
Suresh, G., Sahu, S., 1994. Stochastic assembly line balancing using simulated
annealing. International Journal of Production Research 32, 18011810.
Vilarinho, P.M., Simaria, A.S., 2002. A two-stage heuristic method for balancing
mixed-model assembly lines with parallel workstations. Journal of Production
Research 40, 14051420.
Waters, Th. R., Putz-Anderson, V., Carg, A., 1994. Applications Manual for the
Revised NIOSH Lifting Equation. NIOSH Publ. No. 94-110. U.S, Department of
Health and Human Services, National Institute for Occupational Safety and
Health, Cincinnati, OH.
Wee, T., Magazine, M., 1982. Assembly line balancing as generalized bin packing.
Operations Research Letters 1, 5658.
Wilson VanVoorhis, C., Morgan, B.L., 2007. Understanding power and rules of thumb
for determining sample sizes. Tutorials in Quantitative Psychology 3, 4350.