You are on page 1of 6

9th IFAC Conference on Manufacturing Modelling, Management and

9th IFAC Conference on Manufacturing


Control Modelling,
Available Management
online and
at www.sciencedirect.com
9th IFAC
Control
9th IFAC Conference
Conference on
on Manufacturing
Manufacturing Modelling,
Modelling, Management
Management and
and
Berlin,
Control Germany, August 28-30, 2019
Berlin,
ControlGermany, August 28-30, 2019
Berlin,
Berlin, Germany,
Germany, August
August 28-30,
28-30, 2019
2019 ScienceDirect
IFAC PapersOnLine 52-13 (2019) 2092–2097

Deep
Deep learning
learning and
and WLC:
WLC: how
how to
to set
set realistic
realistic delivery
delivery dates
dates in
in high
high variety
variety
Deep learning and WLC: how to set realistic
Deep learning and WLC: manufacturing
how to set delivery
systems
realistic delivery dates
dates in
in high variety
high variety
manufacturing systems
manufacturing systems
manufacturing systems
Davide Mezzogori, Giovanni Romagnoli, Francesco Zammori
Davide Mezzogori, Giovanni Romagnoli, Francesco Zammori
Davide
Davide Mezzogori,
Mezzogori, Giovanni
Giovanni Romagnoli, Francesco Zammori
Department of Engineering and Romagnoli, Francesco Zammori
Department of Engineering and Architecture,
Architecture, University
University ofof Parma,
Parma,
Via G.P.Department
Usberti, 181/A,
of 43124,
Engineering Pr, ITALY
and (e-mail:
Architecture, davide.mezzogori@unipr.it).
University
Via G.P.Department of Engineering
Usberti, 181/A, 43124, Pr,and Architecture,
ITALY University ofof Parma,
Parma,
(e-mail: davide.mezzogori@unipr.it).
Via
Via G.P.
G.P. Usberti,
Usberti, 181/A,
181/A, 43124,
43124, Pr,
Pr, ITALY
ITALY (e-mail:
(e-mail: davide.mezzogori@unipr.it).
davide.mezzogori@unipr.it).
Abstract: The focus is on workload control, a production planning and control technique that reduces and
Abstract: The focus is on workload control, a production planning and control technique that reduces and
stabilizes
Abstract:the
Abstract: Thetotal
The focus
focusthroughput
is
is on time. control,
on workload
workload In theseaaconditions,
production defining
production planningrealistic
planning and delivery
and control datesthat
technique should become
reduces and
stabilizes the total throughput time. control,
In these conditions, defining control
realistic technique
delivery datesthat reduces
should becomeand
easier, yet
stabilizes
stabilizes the
the use
total of basic
throughput techniques
time. In often
these proves to
conditions,be ineffective.
defining Hence,
realistic we propose
delivery dates using
shouldstatistical
become
easier, yetthe
thetotal
use ofthroughput time. Inoften
basic techniques theseproves
conditions,
to be defining realistic
ineffective. Hence,delivery dates using
we propose shouldstatistical
become
and/or
easier, neural network techniques to estimate, starting from the current workload of theusing
job statistical
shop, the
easier, yet
and/or yet the
neural use
use of
of basic
the network basic techniques
techniques
techniques often
often proves
to estimate,proves to
to be
starting ineffective.
befrom
ineffective. Hence,
Hence,
the current we
we propose
workloadpropose
of theusing
job statistical
shop, the
expected
and/or lead
neural
and/or neural time
network
networkof entry jobs,
techniques
techniques and
to to use this
estimate,
to to
estimate, estimation
starting fromto define
the reliable
current delivering
workload of dates.
the jobTo test the
shop,
expected lead time of entry jobs, and use thisstarting from
estimation to the current
define workload
reliable of the
delivering jobTo
dates. shop,
test the
approach,
expected
expected leadwe
lead simulated
time
time of a
of entry
entry6-machines
jobs,
jobs, and
and to job-shop
use
use this
tojob-shop and we make
estimation
this and
estimation to predictions
to define
define using
reliable
reliable a multi-regressive
delivering
delivering dates.
dates. To linear
To test
test the
the
approach, we simulated a 6-machines we make predictions using a multi-regressive linear
model and
approach, a multi-layer neural network. In terms of tardy jobs, both approaches performed very well, with
model and we
approach, we simulated aa 6-machines
simulated neural
a multi-layer network.job-shop
6-machines In termsand
job-shop and we
we make
of tardy jobs, predictions
make predictions using aa multi-regressive
usingperformed
both approaches very well,linear
multi-regressive linear
with
the neural
model and network
and providing
aa multi-layer neuralthe
neural best results.
network. In Copyright
In terms of tardy©
of tardy 2019both
IFAC
model
the neural multi-layer
network providing network.
the best results.terms
Copyright ©jobs,
jobs,
2019both
IFACapproaches
approaches performed
performed very
very well,
well, with
with
the neural network providing the best results. Copyright © 2019 IFAC
© 2019, IFACDeep
the neural
Keywords: (International
network providing
Learning; Federation
the best of Automatic
results.
Delivering Control) Hosting
Copyright ©
Dates Estimation; 2019 IFAC by Elsevier
Production Ltd. AllControl
Control, rights reserved.
Systems;
Keywords: Deep Learning; Delivering Dates Estimation; Production Control, Control Systems;
Probabilistic
Keywords:
Keywords: Deep &
Deepstatistical models
Learning;
Learning; in industrial
Delivering
Delivering plant
Dates
Dates control;
Estimation;Modeling, simulation,
Production control
Control, and
Control monitoring
Systems;
Probabilistic & statistical models in industrial plantEstimation; Production
control; Modeling, Control,
simulation, Control
control Systems;
and monitoring
of manufacturing
Probabilistic
Probabilistic & processes.
statistical
& statistical models in industrial plant control; Modeling, simulation, control and
models in industrial plant control; Modeling, simulation, control and monitoring monitoring
of manufacturing processes.
of
of manufacturing
manufacturing processes.
processes. 

1. INTRODUCTION AND RELATED WORKS  In this regard, a possible solution is offered by Workload
1. INTRODUCTION AND RELATED WORKS In this regard, a possible solution is offered by Workload
Control
In
In this (WLC),aaa possible
this regard,
regard, hybrid push-pull
solution Production
is
is offered
offered by Planning
Workload and
1. INTRODUCTION
1. INTRODUCTION AND RELATED
AND RELATED WORKS
WORKS Control (WLC), a possible solution Production
hybrid push-pull by Workload
Planning and
Lean manufacturing is commonly used to increase shop floor Control (PPC) (PPC)
(WLC), system,
a hybridwhich is
push-pull been receiving
Production a great
Planning deal
great dealand
of
Lean manufacturing is commonly used to increase shop floor Control (WLC),system, a hybrid which push-pull
is beenProduction
receiving aPlanning and
of
performances. The aim is to spot and to get rid of every waste attention,
Control as demonstrated by the vast literature in the field
Lean
Lean manufacturing
manufacturing
performances. The aim is
is commonly
commonly
is to spot and used
used to to
getincrease
to increase
rid of everyshop
shopwastefloor Control (PPC)
floor attention, (PPC) system,
system, which
as demonstrated which byis been
isthe
been receiving
receiving
vast literature aa great
great
in the deal
deal of
of
field
and to create an
performances. The ideal
aim value
is to stream,
spot and where
to get jobsofflow,
rid every almost
waste (Thürer
attention,
attention, etas
as al. 2011a). WLC
demonstrated
demonstrated by
by abandons
the
the vast
vast the idea
literature
literature in
in to “pull
the
the field
field
performances.
and to create an The aimvalue
ideal is to spot and where
stream, to get rid
jobsofflow,
everyalmostwaste (Thürer et al. 2011a). WLC abandons the idea to “pull
flawlessly, from aa workstation to thewherenext jobs
one. flow,
Doubtlessly, production”
(Thürer
(Thürer et et al. using visual WLC
2011a). cards abandons
and, instead, itidea
regulates the
and
and to
to create
create
flawlessly, from an
an ideal
ideal value
value stream,
workstation stream,
to thewherenext jobs
one. flow, almost
almost production”
Doubtlessly, al.
using visual WLC
2011a). cards abandons
and, instead, the
the it idea
regulatesto
to “pull
“pull
the
queues and
flawlessly, inventories
from a are
workstation the most
to important
the next one.wastes of push-
Doubtlessly, manufacturing
production”
production” using system
using visual
visualusingcards
cardstimely
and, and accurate
instead,
and, and
instead, it shop
regulates
it regulates floor
the
the
flawlessly,
queues and from a workstation
inventories are the most to the next one.
important of push- manufacturing
Doubtlessly,
wastes system using timely accurate shop floor
oriented
queues andmanufacturing
inventories aresystems.
the most Not only
important they
wastesrepresent
of push- a data, collected
manufacturing
manufacturing with
system ausingManufacturing
timely and Executing
accurate shop System
floor
queues and
oriented inventories are
manufacturing the mostNot
systems. important
only they wastes of push-a data,
represent collectedsystem with ausing timely and accurate
Manufacturing Executing shop floor
System
tangible (MES). As the name suggests, the basic idea is to define one
oriented cost,
oriented
tangible
but they also
manufacturing
manufacturing
cost, but they also
increase
systems.
systems.
increase Notand
Not andonly
onlymake
they
make representthe
theyuncertain
represent
uncertain data,
data, collected
theaa (MES). collected
As the name with aa Manufacturing
withsuggests, Manufacturing
the basic idea Executing
Executing
is to define System
Systemone
total Throughput Time (TT). It is needless to say that these or more As
(MES). threshold
the namelimits (or norms)
suggests, the of the idea
basic system’s
is to workloads,
define
tangible
tangible
total cost, but
cost, but Time
Throughput they
they alsoalso increase
(TT).increase and make
and make
It is needless to sayuncertain
uncertain
that these the
the or(MES).
more As the name
threshold suggests,
limits (or norms) the basic
of the idea define one
is to workloads,
system’s one
issuesThroughput
total strongly reduce Time the competitiveness
(TT). It is needless oftoa manufacturing
say that these and
or to
more
or more use these
threshold
threshold values
limits
limitstoto
(or regulate
norms)
(orregulate
norms) the the
of jobs’
the
of the release
system’s
system’s phase. By
workloads,
total Throughput
issues strongly reduce Timethe (TT). say that these and
It is needlessoftoa manufacturing
competitiveness to use these values jobs’ releaseworkloads,
phase. By
firm, doing so, WLC makes to it regulate
possible the to drastically reduce WIP
issuessince
issues
firm,
estimating
strongly
strongly
since reduce and
reduce
estimating the promising reliable
the competitiveness
and competitiveness
promising reliable of Delivering
of aa manufacturing
manufacturing
Delivering
Dates doing
Dates and and to use
to so,
use WLCthese
these values
makes to
values it regulate
possible the jobs’
jobs’ release
release
to drastically phase.
phase.WIP
reduce By
By
(DDs)
firm, becomes
since hard,
estimating if
and not even
promising impossible
reliable (Bertolini
Delivering et
Datesal. and
doing
doingqueuing
so,
so, WLC
WLC times, preserving,
makes
makes it
it possible
possibleat theto
tosame time,
drastically
drastically the maximum
reduce
reduce WIP
WIP
firm, since
(DDs) estimating
becomes hard, and if notpromising reliable Delivering
even impossible (Bertolini Dates et al. and queuing times, preserving, at the same time, the maximum
2018). Using throughput rate. Also, and perhaps moretime,important, when
(DDs)
(DDs) Using the
2018). becomes
becomes hard,
hard,
the words
wordsif notof
if not Hopp
ofeven
even
Hopp and
and Spearman
impossible
impossible (Bertolini(2003),
(Bertolini
Spearman et
et al.
(2003), and
and queuing
queuingrate.
al. throughput times, preserving,
times,Also,
preserving,
and perhapsat
at the
the samemoretime,
same the
the maximum
important, maximum when
“Controlling
2018). Using Work-In-Process
the words of (WIP)
Hopp and
and protecting
Spearman TT from
(2003), norms are fine
throughput rate.tuned
Also,to minimize
and perhaps WIP,morethe total TT becomes
important, when
2018). UsingWork-In-Process
“Controlling the words of (WIP) Hopp and and protecting
SpearmanTT (2003), throughput
from norms are fine rate.tuned
Also, and perhaps
to minimize WIP,more important,
the total TT becomes when
variance” areWork-In-Process
“Controlling the main solutions (WIP)to counteract
and protectingtheseTT wastes.
from shorter
norms
norms are and
are more
fine stable
tuned to and this
minimize makes
WIP, easier
the the
total definition
TT becomes of
“Controlling
variance” areWork-In-Process
the main solutions (WIP) and protecting
to counteract theseTT from shorter
wastes. andfinemore tuned
stableto and
minimize
this makesWIP,easier
the total
the TT becomes
definition of
These
variance”basic are principles
the main are easy
solutions toto be deployed
counteract in
theseMake-To-
wastes. realistic
shorter and
shorter and DDs.more
more This issue
stable
stable andis of
this strategical
makes relevance
easier the for
definition MTO of
variance”
These basic areprinciples
the main are solutions
easy totobecounteract
deployed these wastes. realistic
in Make-To- DDs. This issueand is this makes easier
of strategical the definition
relevance for MTO of
Stock
These (MTO)
basic production
principles are systems,
easy to bewhere demand
deployed in is
Make-To-rather companies,
realistic DDs.
realistic DDs. because
This the
issue
This issue capability
is of to
strategical
is of strategical promise and
relevance
relevance to
for respect
MTO
These basic
Stock (MTO)principles
production are systems,
easy to bewhere deployed
demand is rather companies,
in Make-To- because the capability to promise and toforrespectMTO
constant, and the productionsystems, mix is not too diversified. short DDs, in response to customers’ enquiries, isto
toaa respect
critical
Stock
Stock (MTO)
constant,(MTO)and the production
production
productionsystems, mix is not where
where demand
demand is
too diversified. isIndeed,
rather companies,
companies,
rather short
Indeed, DDs, in because
because
response the
the tocapability
capability
customers’ to
to promise
promise
enquiries, and
andis respect
critical
several
constant, card-based
and systems, such as Dual-Kanban and Constant success
short
short DDs, factor.
in Nonetheless,
in response to even
to customers’ in case of
enquiries,WLC, is DDs
is aaDDs are
critical
constant,
several and the
card-basedthe production
production
systems, such mix
mix asis
is not
not too
too diversified.
Dual-Kanbandiversified. Indeed,
Indeed, success
and Constant DDs, factor. response
Nonetheless, customers’
even in case enquiries,
of WLC, critical
are
Work
several in Process,
card-based have been ideated and implemented, to typically
success defined
factor. using
Nonetheless, very simple
even in rules,
case ofbased
WLC, on average
DDs are
severalin
Work Process, systems,
card-based systems,
have been such as
as Dual-Kanban
suchideated Dual-Kanban and
and Constant
and implemented, Constant success factor.
to typically defined Nonetheless,
using very even simple in rules,
case of WLC,
based on DDs averageare
provideina Process,
Work simple visual have approach
been for production
ideated and control and
implemented, to values,
typically that
typicallythat are
defined
defined not
using
usingvery very
veryperforming,
simple especially
rules,
simple rules, based
based on in
on case
average of
Work
provideina Process,
simple visual have approach
been ideated and implemented,
for production control and to values, are not very performing, especially in average
case of
Lead
provide Timea (LT)
simple reduction.
visual approachUnfortunately,
for production as reported
control by
and HVLV
values, job-shops,
that
values, job-shops, are
that are not with
not complex
very
very routings
performing,
performing, (Thürer
especially etinal. 2016).
case of
provideTime
Lead a simple
(LT) visualreduction.approach for production
Unfortunately, control and
as reported by HVLV with complex routingsespecially
(Thürer etinal.case 2016).of
several
Lead Time
Lead Time authors(LT) (Germ
(LT)(Germ reduction.
reduction.and Riezebosm,
Unfortunately,
Unfortunately, 2010; as Harrod
reported
as reported and
by HVLV
by HVLV job-shops,
job-shops, with
with complex
complex routings
routings (Thürer
(Thürer et
et al.
al. 2016).
2016).
several authors and Riezebosm, 2010; Harrod and The present paper belongs to this stream of research and aims
Kanet,
several2013)
several authors the same
(Germdoes
(Germ andnot
and hold in case2010; of Make-To-Order and The present paper belongs to this stream of research and aims
Kanet, authors
2013) the same does notRiezebosm,
Riezebosm,
hold in case2010; Harrod
Harrod and
of Make-To-Order to
Theremedy
present this problem,
paper belongs by tointegrating WLC with a robust and
(MTO)
Kanet, High-Variety-Low-Volumes
2013)
Kanet, 2013)
(MTO) the same does not hold
the same does not hold in(HVLV)
High-Variety-Low-Volumes in(HVLV)
case of manufacturers.
Make-To-Order
case of Make-To-Order
manufacturers. to The
to presentthis
remedy paper belongs
problem, by to this
this stream
integratingstream WLCof
of research
research
with a robust and
and aimsaims
and
In theseHigh-Variety-Low-Volumes
cases, even more sophisticated techniques, effective
remedy
such as effective forecasting
this problem, system.
by Specifically,
integrating WLC wewith propose
a robust using
and
(MTO)
(MTO)
In theseHigh-Variety-Low-Volumes
cases, even more sophisticated (HVLV)
(HVLV)techniques, such as to
manufacturers.
manufacturers. remedyforecasting
statistical
effective
this problem,
and neural
forecasting
by integrating
system.
network
system.
Specifically,
techniques
Specifically,
WLCwe
to we
with a robustusing
propose
estimate,
propose
and
starting
using
POLCA
In these or Synchro
cases, even MRP,
more do not
sophisticatedoffer satisfactory
techniques, results,
such as effective
statistical forecasting
and neural system.
network Specifically,
techniques to we propose
estimate, using
starting
In these or
POLCA cases, even MRP,
Synchro more sophisticated
do not offer techniques,
satisfactory such as from
results, the and
statistical current workload situation ofto the job shop (as
especially
POLCA
POLCA or
especially orif shop floor
if Synchro
Synchro
shop floor MRP,
MRP,activities
do
do not
activities
areoffer
not organized
areoffer
as a job-shop,
satisfactory
satisfactory
organized results,
results, from
as a job-shop, the and
statistical neural
neural
current network
workload techniques
network situation
techniquesof totheestimate,
estimate,
job shop starting
starting
(as
with complex routings with admitted re-routingas (Land depicted
from
et al. depicted the by the
current WLC’s parameters),
workload situation the ofexpected
the job LT shopof new
(as
especially
especially
with complex if
if shop
shop floor
floor activities
routings activities
with admitted are
are organized
organized
re-routingas(Land et al. from
aa job-shop,
job-shop, entry
thebycurrent
jobs,
depicted by and
the WLC’s
the to use
WLC’s
workload
this
situationtheofexpected
parameters),
estimation
parameters), to
the
the jobLT
promise
expected
shop
realistic
LT
of new
of
(as
DDs.
new
2009).
with Synchro
complex
with complex MRP
routings works
with well
admittedonly in case
re-routing of standard
(Land et or
al. depicted
entry jobs, by andthe WLC’s
to use thisparameters),
estimation the
to expected
promise LT
realistic of new
DDs.
2009). Synchroroutings
MRP works with admitted
well only re-routing
in case of (Landstandard et al.
or entry
To this aim, a multivariate
jobs, linear regression model and a fully
generalized
2009).
2009). Synchro
Synchro
generalized
flowMRP
flowMRP
shops,
shops,
whereas
works
works
whereas wellPOLCA
well only
only in
POLCA in can
casebe
case
can beof used
of standard
standard
used
also or
also or
in To
in entry this aim,and
jobs, and to
to use
use this
this estimation
a multivariate estimation to
to promise
linear regression promisemodel realistic
and a DDs.
realistic DDs.
fully
case of job shops, provided that routings are few and linear connected
To
To this
this aim,
aim, multi-layer
a
a multivariate
multivariateneural network
linear
linear are
regression
regression developed
model
model andaatested
and
and fully
fully
generalized
generalized
case of job flowflow shops,
shops,shops, providedwhereas
whereas POLCA
that POLCA
routings can can be
are be used
fewused also
andalso in connected
in connected
linear multi-layer neural network are developed and tested
(Bertolini et shops,
al. 2013a). As soon as then number of alternative using
connected a standard
multi-layer
multi-layer 6-machines
neural
neural network
network job-shop
are
are developed
developed configuration,
and
and tested
tested
case of job
case of jobet shops,
(Bertolini al. 2013a). provided
provided
As soon that routings
thatasroutings
then numberare few
are few and linear
and linear using
of alternative using a standard 6-machines job-shop configuration,
routings reproduced via simulation.
aa standard 6-machines
6-machines job-shop job-shop configuration,
(Bertoliniincreases,
(Bertolini
routings et
et al.
increases,
the number
al. 2013a).
2013a). As
As soon
the number soonof cards
as
as then
of cards thenexplodes,
number making
number
explodes, of visual reproduced
visual using
of alternative
makingalternative reproduced
standard
via simulation.
via simulation.
configuration,
management
routings
routings increases,almost
increases, theimpossible.
number of cards explodes, making visual reproduced via simulation.
management almosttheimpossible.
number of cards explodes, making visual
management
management
2405-8963
almost
almost
© 2019,
impossible.
IFACimpossible.
(International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
Peer review©under
Copyright 2019 responsibility
IFAC of International Federation of Automatic
2142Control.
Copyright © 2019 IFAC
10.1016/j.ifacol.2019.11.514 2142
Copyright
Copyright ©© 2019
2019 IFAC
IFAC 2142
2142
2019 IFAC MIM
Berlin, Germany, August 28-30, 2019 Davide Mezzogori et al. / IFAC PapersOnLine 52-13 (2019) 2092–2097 2093

The remainder of the paper is organized as follows. An 𝑊𝑊𝑚𝑚+ = 𝑊𝑊𝑚𝑚− + 𝑤𝑤𝑗𝑗 ⋆,𝑚𝑚 (1)
introduction on WLC is given in Section 2, while the main
features of the forecasting procedure are detailed in Section 3. 𝑊𝑊 + = 𝑊𝑊 − + ∑ 𝑤𝑤𝑗𝑗 ⋆ ,𝑚𝑚 (2)
𝑚𝑚
Next, the simulation model, the experimental campaign and
Where 𝑤𝑤𝑗𝑗 ⋆,𝑚𝑚 is the workload of 𝑗𝑗 ⋆ relatively to machine m,
the obtained results are discussed, respectively, in Section 4

with 𝑤𝑤𝑗𝑗 ⋆,𝑚𝑚 = 0 if machine m is not on the routing of job 𝑗𝑗 ⋆ .


and Section 5. Lastly, concluding remarks and possible
directions for future works are drawn in Section 6.
Also, the subscripts (-) and (+) indicate a quantity that is
2. WLC, A CONCEPTUAL BACKGROUND evaluated immediately before and immediately after a new job
is considered for release.
Concerning 𝑊𝑊𝑚𝑚 and 𝑊𝑊, these quantities are computed as in (3)
WLC is a hybrid PPC system designed for small and medium
companies, with job-shop configuration, that manufacture
products in small volumes and on customers’ specifications and (4):

𝑊𝑊𝑚𝑚 = ∑ 𝑤𝑤𝑗𝑗,𝑚𝑚
(Henrich et al. 2007). For these companies, flexibility is a
critical success factor and so, a job-shop oriented layout, with
(3)
𝑗𝑗∈𝐽𝐽(𝑆𝑆,𝑚𝑚)
WIP accumulating between workstations, cannot be
eliminated. Stations cannot be clustered into dedicated U-
shaped cells and a continuous flow cannot be obtained. 𝑊𝑊 = ∑ ∑ 𝑤𝑤𝑗𝑗,𝑚𝑚 = ∑ 𝑊𝑊𝑚𝑚 (4)
𝑚𝑚 𝑗𝑗∈𝐽𝐽(𝑆𝑆,𝑚𝑚) 𝑚𝑚
Therefore, the goal of WLC is to preserve a “batches and

Where 𝐽𝐽(𝑆𝑆,𝑚𝑚) is the set of the jobs that are currently being
queues” system, but to balance production to achieve a flow as
smooth and linear as possible. To this aim, WLC decouples the
manufacturing system from the flow of incoming jobs, taking processed in the shop-floor and that must visit machine m.
advantage of a Pre Shop Pool (PSP) of orders and of a specific
job 𝑗𝑗 ⋆ is released in the shop floor, otherwise it remains
order release mechanism. If, after updating the workload, none of the norms is violated,

the system’s level (i.e., control is limited to 𝑊𝑊) is referred as


The underlying logic is straightforward: new jobs are released pending in the PSP. Please note that workload control made at
in production only if the value of certain WIP caps are not

machines (i.e., to each 𝑊𝑊𝑚𝑚 , with m belonging to the set of the


exceeded. WIP caps are also referred as workload norms and Shop Load control, while a control extended to the bottleneck
are generally quantified as the number of expected processing
hours. To optimize the system, norms should be fine-tuned to critical machines) is referred as Bottleneck Load control.
keep queues as short as possible, possibility without
generating starvation and keeping unaltered the (maximal) As indicated in equations (1) to (4), a key role of the job release
throughput rate. phase is played by the way in which one quantifies the
workload contribution of a single job. To this aim the

first case, 𝑤𝑤𝑗𝑗,𝑚𝑚 is simply computed as the sum of the


aggregate and the probabilistic approach can be used. In the

in case of probabilistic approach, the load 𝑤𝑤𝑗𝑗,𝑚𝑚 is rescaled, to


processing and set up times of job j on machine m. Conversely,

(i.e., 𝑤𝑤𝑗𝑗,𝑚𝑚 is reduced because it does not immediately affect


properly consider the time taken by job j to reach machine m

machine m). Frequently, the rescaled workload 𝑤𝑤 ̃𝑗𝑗,𝑚𝑚 is simply


computed as in (5):
𝑤𝑤𝑗𝑗,𝑚𝑚
𝑤𝑤
̃𝑗𝑗,𝑚𝑚 =
𝑛𝑛𝑚𝑚
(5)

where 𝑛𝑛𝑚𝑚 is the position of machine m in the routing of job j.


Fig. 1. WLC three levels of control

More specifically, as sketched in Fig. 1., WLC makes use of In practice, the further is machine m, the fewer load is

as the jobs proceeds downstream, and the position of 𝑛𝑛𝑚𝑚


the three different levels of control. attributed to it. Obviously, the workload contribution increases
Job Entry - At first, new orders are evaluated, in terms of
technological and economic feasibility. If accepted, orders are decreases.
inserted in the PSP and DDs are defined. Job Dispatching - When jobs are released into the shop floor,
Job Release - Pending orders are sorted (using a dispatching they are “pushed” from a workstation to the next one, exactly
rule) and individually considered for release, either in a as in a standard “batch and queues” manufacturing process. It

intervals). To this aim, whenever a job 𝑗𝑗 ⋆ is considered for


continuous or in a discrete way (i.e., at predetermined time is exactly for this reason that WLC is considered a hybrid
push-pull technique. Due to the formation of queues, there is

workload of the system 𝑊𝑊 and/or the workload 𝑊𝑊𝑚𝑚 seen by an


possible release, its workload contribution is used to update the the need to decide the order in which jobs will be processed.
To do so, simple dispatching rules can be used, such as: First-
individual machine m of the shop floor, as shown in (1) and In-First-Out (FIFO), Earliest-Due-Date (EDD), Minimum-
(2): Slack-Time (MST), Operation Due Date (ODD) and Planned
Operation Start Time (PST).

2143
2019 IFAC MIM
2094
Berlin, Germany, August 28-30, 2019 Davide Mezzogori et al. / IFAC PapersOnLine 52-13 (2019) 2092–2097

limited to the machines belonging to the routing of job 𝑗𝑗 ∗ (i.e.,


only 𝑤𝑤𝑗𝑗,𝑚𝑚 , with 𝑚𝑚 ∈ 𝑅𝑅𝑗𝑗 ∗ are considered). This is shown in (7):
3. PROBLEM DESCRIPTION AND PROPOSED METHOD

3.1 Preliminary step: WLC optimisation


To generate reliable DDs, the idea is to regulate the system 𝑌𝑌𝑗𝑗 ∗ = 𝑓𝑓1 ( ∑ 𝑤𝑤𝑗𝑗,1∗ , … , ∑ 𝑤𝑤𝑗𝑗,𝑚𝑚∗ , … , ∑ 𝑤𝑤𝑗𝑗,𝑀𝑀∗ ) +
𝑗𝑗∈𝐽𝐽𝑃𝑃 𝑗𝑗∈𝐽𝐽𝑃𝑃 𝑗𝑗∈𝐽𝐽𝑃𝑃
using a WLC system based on Bottleneck Load and, next, to
(7)
forecast the expected LT using, as explanatory variables, jobs

+𝑓𝑓2 ( ∑ 𝑤𝑤𝑗𝑗,1∗ , . . , ∑ 𝑤𝑤𝑗𝑗,𝑚𝑚∗ , . . , ∑ 𝑤𝑤𝑗𝑗,𝑀𝑀∗ )


routings, processing times and the current workload state of
the system. In this regard, the use of WLC is essential because,
relatively to a standard push system, queuing times can be 𝑗𝑗∈𝐽𝐽(𝑆𝑆,1∗ ) 𝑗𝑗∈𝐽𝐽(𝑆𝑆,𝑚𝑚∗ ) 𝑗𝑗∈𝐽𝐽(𝑆𝑆,𝑀𝑀∗ )

Where: {1∗ , … , 𝑚𝑚∗ , … , 𝑀𝑀∗ } is the set of critical machines


reduced, and their variability can be strongly limited. Both

belonging to the routing 𝑅𝑅𝑗𝑗 ∗ , 𝐽𝐽𝑃𝑃 is the set of jobs pending in the
these issues are fundamental to assure robust and accurate

PSP and 𝐽𝐽(𝑆𝑆,𝑚𝑚∗ ) is the set of jobs that in the shop-floor that must
forecasts. Even the choice of the Bottleneck Load approach is
not casual, as this kind of control assures a more detailed
visit machine 𝑚𝑚∗ . Also note that (∑𝑗𝑗∈𝐽𝐽(𝑆𝑆,𝑚𝑚∗) 𝑤𝑤𝑗𝑗,𝑚𝑚∗ +
description of the current workload state, which, indeed, is

∑𝑗𝑗∈𝐽𝐽𝑃𝑃 𝑤𝑤𝑗𝑗,𝑚𝑚∗ ) = (∑𝑗𝑗∈𝐽𝐽𝑃𝑃 𝑤𝑤𝑗𝑗,𝑚𝑚∗ + 𝑊𝑊𝑚𝑚∗ ) represent the cumulated


detailed at the level of individual machines, rather than at the
overall system level. We also note that, to reduce variability as
workloads of machine 𝑚𝑚∗ , due to the jobs in the PSP and to the
much as possible, norms should be fine-tuned to minimize
WIP and queues, keeping the throughput rate at the desired
ones in the shop floor.
(probably maximum) level. Norms optimisation is not a trivial
task and a possible approach is pinpointed in Section 5; a
3.3 Training and assessment of the model
To estimate 𝑓𝑓1 and 𝑓𝑓2 , we propose using either a multivariate
further discussion can be found in Thürer et al. (2011b),
Bertolini et al. (2013b) and Fernandes et al. (2014).
liner regression model, or a fully connected multi-layer neural

observed waiting time 𝑌𝑌𝑗𝑗 ∗ (i.e., 𝑃𝑃𝑗𝑗 ∗ + 𝑄𝑄𝑗𝑗 ∗ ) of each accepted job
3.2 Development of the forecasting model network. In both cases, the model must be fitted with the

introduced. Basically, for an incoming job 𝑗𝑗 ∗ the firm should 𝑗𝑗 ∗ , and with the cumulated workloads ∑𝑗𝑗∈𝐽𝐽𝑃𝑃 𝑤𝑤𝑗𝑗,𝑚𝑚∗ and 𝑊𝑊𝑚𝑚∗ ,
Once WLC has been fine-tuned, the forecasting model can be

computed at the time of its acceptance. Obviously, 𝑌𝑌𝑗𝑗 ∗ is the


an appropriate security factor 𝛾𝛾. We remind that LT is defined dependent variable, whereas ∑𝑗𝑗∈𝐽𝐽𝑃𝑃 𝑤𝑤𝑗𝑗,𝑚𝑚∗ and 𝑊𝑊𝑚𝑚∗ are the
promise a DD equal to the expected LT, possibly increased by

in case of WLC, LT can be split in the time spent by job 𝑗𝑗 ∗ ,


as the time between order acceptance and order delivery; thus, independent variables.

by the famous Little’s Law (i.e., 𝐿𝐿𝐿𝐿 = (𝑊𝑊𝑊𝑊𝑊𝑊 ⁄ 𝜌𝜌) , where 𝜌𝜌 is


respectively, in the PSP and in the shop floor. The last part, Due to the nonlinear relation among WIP and LT, as suggested

processing, set up, and queuing time of 𝑗𝑗 ∗ , at each machine m


namely the total TT, can be further decomposed in the
the throughput rate expressed in jobs per unit of time), we
belonging to its routing 𝑅𝑅𝑗𝑗 ∗ . This is shown in (6): expect that performance of the neural network will be superior
than that of the linear regression. Yet, we included also the
linear model, not only to have a benchmark solution, but also
𝐿𝐿𝐿𝐿𝑗𝑗 ∗ = 𝑃𝑃𝑗𝑗 ∗ + (𝑄𝑄𝑗𝑗 ∗ + ∑ (𝑠𝑠𝑗𝑗 ∗ ,𝑚𝑚 + 𝑝𝑝𝑗𝑗 ∗,𝑚𝑚 ))
(6)
for practical and operating reasons. Indeed, whereas the
historical data saved on the MES could be enough to fit the
𝑚𝑚∈𝑅𝑅𝑗𝑗∗ regression model, the same is probably not true in case of a
Where: 𝑃𝑃𝑗𝑗 ∗ is the pending time in the PSP, 𝑄𝑄𝑗𝑗 ∗ is the queuing
neural network, which needs much more data to be trained. We

time in the shop floor, 𝑠𝑠𝑗𝑗 ∗ ,𝑚𝑚 and 𝑝𝑝𝑗𝑗 ∗ ,𝑚𝑚 are the set-up and
also note that a possible way to generate these data is to use
discrete-events Monte Carlo simulation. Briefly, the job-shop

in brackets represent the total TT of job 𝑗𝑗 ∗


processing time, respectively. Also note that, in (6), the terms must be reproduced using a simulation environment and, after
its adequacy has been checked (for instance by comparing real
As noted in Section 2, workload 𝑤𝑤𝑗𝑗 ∗ ,𝑚𝑚 accounts, exactly, for
system performance to the ones obtained via simulation), the
processing and set up times of job 𝑗𝑗 ∗ on machine m. So, with a
model is executed and both waiting times and workload values
are collected to train the neural network. Anyway, either using

(𝑠𝑠𝑗𝑗 ∗ ,𝑚𝑚 + 𝑝𝑝𝑗𝑗 ∗,𝑚𝑚 ) with 𝑤𝑤𝑗𝑗 ∗,𝑚𝑚 . This simplification is licit, as the
good approximation, we can substitute, in (6), the term the linear regression or the neural network, real operating data,

variability of both 𝑃𝑃𝑗𝑗 ∗ and 𝑄𝑄𝑗𝑗 ∗ is expected to be much greater


collected and saved on the MES, could be used to dynamically

than that of ∑(𝑠𝑠𝑗𝑗 ∗,𝑚𝑚 + 𝑝𝑝𝑗𝑗 ∗ ,𝑚𝑚 ). Owing to this, since workloads
update the estimates and/or to refine the forecasting models.

𝑤𝑤𝑗𝑗 ∗ ,𝑚𝑚 are known, the dependent variable that must be 4. EXPERIMENTAL RESULTS
estimated, let it be 𝑌𝑌𝑗𝑗 ∗ , reduces to the sum of the waiting times
𝑃𝑃𝑗𝑗 ∗ and 𝑄𝑄𝑗𝑗 ∗ . Our fundamental hypothesis is that, by the time 4.1 Simulation Model
job 𝑗𝑗 ∗ is considered for release, both 𝑃𝑃𝑗𝑗 ∗ and 𝑄𝑄𝑗𝑗 ∗ could be
To test the approach, a randomly routed pure job-shop was
estimated as a function of the workloads of the jobs pending in
developed using Simpy©, a simulation module developed in
the PSP, and of the total workload of the critical machines,
Python 3.7©. The job-shop is made of six equally loaded
workstations with constant capacity and 100% availability.

2144
2019 IFAC MIM
Berlin, Germany, August 28-30, 2019 Davide Mezzogori et al. / IFAC PapersOnLine 52-13 (2019) 2092–2097 2095

The job-shop is considered as “pure” and “randomly routed”, Table 1 also displays the percentage of tardy jobs; to compute
because jobs can visit a number of workstations that varies, this value, DDs were set using an approach like the one
uniformly, from one to six. Even the order with which proposed by Land (2006). Specifically, we added to the job

interval [(𝑛𝑛 + 1) ∙ 𝑝𝑝𝑚𝑚𝑚𝑚𝑚𝑚 ; 2 ∙ (𝑛𝑛 + 1) ∙ 𝑝𝑝𝑚𝑚𝑚𝑚𝑚𝑚 ], where 𝑝𝑝𝑚𝑚𝑚𝑚𝑚𝑚 is


workstations are visited is totally random, and re-routing is entry time a random allowance uniformly distributed in the

of low-volumes job-shops, processing times 𝑝𝑝𝑗𝑗,𝑚𝑚 are generated


also allowed. Furthermore, to introduce the variability typical
the maximum processing time and n is the maximum number

n = 6 and 𝑝𝑝𝑚𝑚𝑚𝑚𝑚𝑚 = 4, so the allowance is uniformly distributed


of workstation that can be visited by a job. In the present case
and a mean 𝑝𝑝̅ of 1.2 units of time. Conversely, set up times
from a truncated 2-Erlang, with a maximum of 4 units of time
in the interval [28, 56].
have not been considered. Jobs are generated according to a

an exponential distribution with an arrival rate 𝜆𝜆 of 1.54 units


Poisson process and their inter-arrival times are distributed as Table 1. Norms optimisation results

of time. Both 𝑝𝑝̅ and 𝜆𝜆 have been arbitrarily selected to obtain N  WIP % Tardy P S
an average utilisation level of 90% at each workstation. This 50 5584 130 19.5% 1.26 25.35
setting, as for the other ones used in the simulation, ensures the 45 5581 125 22.1% 1.98 24.87
comparability of the model with most of the ones proposed in
the literature (e.g. Land, 2006; Thürer et al., 2016). 40 5582 123 26.4% 6.17 24.68
*
35 5569 119 31.3% 9.30 23.66
4.2 WLC parameter setting and norms’ optimisation 30 5547 **
111 38.6% 14.76 22.16

Concerning WLC, its configuration relies on: (i) Discrete Time Also note that, in terms of tardiness, performance of a WLC
release approach, with a time window of 24 units of time, (ii) system are worse than that of the equivalent push system; for
Bottleneck Load, (iii) Aggregated work load, and (iv) EDD instance, the percentage of tardy jobs increases from 18% to
dispatching rule. 26.4% moving from a push system to a WLC optimized on
WIP minimization. This is the typical behaviour of
Concerning norms, since the system is balanced with each “unsaturated system” (Bertolini et al., 2013 c) and it is an issue

choice is that to use six equal norms 𝑁𝑁1 = 𝑁𝑁2 = ⋯ = 𝑁𝑁6 = 𝑁𝑁,
workstation having an average 90% utilization rate, a natural that further motivates the introduction of a reliable forecasting
system, as a way to fully exploit the potentialities of WLC. For
one for each workstation of the job-shop. This rational choice the sake of completeness, and extract of trend of the WIP in
also facilitates the norms setting procedure, that is needed to the shop floor with N = 40 is displayed in Fig. 2.
minimize WIP, before the forecasting model can be put in
place. Specifically, due to the perfect symmetry of the system,
a very simple iterative procedure was used.

average throughput rate 𝜌𝜌 [jobs/hour] of the job-shop


- At first, we computed the 95% confidence interval of the

operating in a pure push way (i.e., jobs are released in the


shop floor immediately upon acceptance).
- Next, starting from a very high value, N is progressively
reduced (using a constant step), until the difference
between the average throughput rate and that of the push
system becomes statistically significant.
- The penultimate value found for N approximates the
theoretical optimal value of the norms.
Fig. 2. WLC trend
To conduct the optimization, we performed a simulation trial of
50 runs, each one of 3650 time units, with a warm up period of 4.3 Forecasting models: data fitting and training
1200 time units, sufficient to avoid start up effects. These
simulation parameters are consistent with those generally used
Both the multivariate linear regression model and the neural
in case of job-shop simulations (e.g. Land, 2006; Thürer et al.,
2012), and give stable results in an adequate amount of time. network were developed in Python 3.7© using the Scikit-
learn© and Keras© (with TensorFlow© as backend) libraries
system we found: (i) an average throughput rate equal to 𝜌𝜌 =
By operating in this way, relatively to the push operating
for machine learning and deep learning. This choice was made
5585 ± 5 [jobs per unit of time], (ii) an average WIP of 132 to facilitate the integration of the forecasting models with the
[jobs], (iii) an average LT of 26 [time units] and (iv) an average simulation environment, and also in virtue of the growing
tardy of 18%. Next, repeating the same simulation schema for diffusion of Python© in the academia.
different levels of the norms, an optimal value N = 40 [time
units] was found. This is shown in Table 1, where P and S are Concerning the multivariate linear regression, we modelled the
the average waiting time in the PSP and in the shop floor, relationship between the waiting time (i.e., response variable)
respectively. In the same table, (*) and (**) indicate statistical and the workloads (i.e., explanatory variable), by fitting the
relevance at a p-level of 5% and 1%, respectively. linear model, as in (8):

2145
2019 IFAC MIM
2096
Berlin, Germany, August 28-30, 2019 Davide Mezzogori et al. / IFAC PapersOnLine 52-13 (2019) 2092–2097

and the buyer, DDs are generated according to the following


𝑌𝑌𝑗𝑗 ∗ = 𝐶𝐶 + (𝛼𝛼1 ∑ 𝑤𝑤𝑗𝑗,1 + ⋯ + 𝛼𝛼6 ∑ 𝑤𝑤𝑗𝑗,6 ) + scheme:
𝑗𝑗∈𝐽𝐽𝑃𝑃 𝑗𝑗∈𝐽𝐽𝑃𝑃 - Any time a job enters the system a DD is randomly
(8) generated (as in Section 4.2). This value represents the

+ (𝛽𝛽1 ∑ 𝑤𝑤𝑗𝑗,1 + ⋯ + 𝛽𝛽6 ∑ 𝑤𝑤𝑗𝑗,6 )


original DD proposed by the customer.
- The expected LT of the job is computed and compared
𝑗𝑗∈𝐽𝐽(𝑆𝑆,1) 𝑗𝑗∈𝐽𝐽(𝑆𝑆,6) with the original DD.
- The DD date is accepted if LT ≤ DD.
Where: C is the intercept and 𝛼𝛼 and 𝛽𝛽 are the regression - Otherwise, a Corrected DD is randomly generated in the
coefficients relative to the workloads in the PSP and in the interval [DD; min (1.2DD; LT)]. This random generation
shop floor, respectively. Since all machines have the same corresponds to the negotiation process, where the value
utilisation rate, they are all potentially critical; this is the 1.2 is taken as the maximum allowance factor that the

estimated. We also note that, accordingly to (7), 𝑤𝑤𝑗𝑗,𝑚𝑚 is set to


reason why there are a total of 12 regression coefficients to be customer is willing to accept.

zero if 𝑚𝑚 ∉ 𝑅𝑅𝑗𝑗 ∗ .
As a benchmark, we also estimated LTs using the
straightforward way proposed by Land (2009), which is shown
in (9):
𝐿𝐿𝐿𝐿𝑗𝑗 ∗ = 𝑞𝑞̂𝑗𝑗 ∗,𝑃𝑃 + 𝑘𝑘𝑗𝑗 ∗ 𝑞𝑞̅
Concerning the neural network, we created a multi-layer
network with the following topology: (i) an input layer with 12 (9)

Where, relative to the entering job 𝑗𝑗 ∗ , 𝑞𝑞̂𝑗𝑗 ∗ ,𝑃𝑃 is an estimation of


inputs, one for each workload, (ii) a single output neuron

the waiting time in the PSP, 𝑘𝑘𝑗𝑗 ∗ is the number of workstations


delegated to make the forecast of the total expected waiting

belonging to 𝑅𝑅𝑗𝑗 ∗ , and 𝑞𝑞̅ is the average throughput time (i.e.,


time, (iii) three fully connected hidden layers, each one with
32 neurons (plus a bias) with the Relu activation function, (iv)

symmetry of the system 𝑞𝑞̅ is equal for each workstation and it


batch normalization after each hidden layer. processing and queuing time) at each workstation. Due to the

To fit the linear regression model and to train the neural does not depend on the job being considered. Also, due to the

constant through time and so 𝑞𝑞̅ can be simply estimated taking


network we performed 5000 simulation runs, of 3650 time stabilization obtained by means of WLC, this average is rather
units with an extra warm up period of 1200 time units. In this
way, we observed the waiting times of 6,717,775 different jobs the long-run average of the throughput time observed at each

Conversely, 𝑞𝑞̂𝑗𝑗 ∗ ,𝑃𝑃 must be estimated at run time as it strongly


and we “took” a picture of the current workload state, an equal workstation that, in the present case is equal to 8 unit of time.
number of times. The 80% of the generated data were used as
training set, the remaining 20% as test set. depends on the current workload situation. A possible way is

Relatively to the linear regression we got an 𝑅𝑅2 = 0.65, with released, an entering job 𝑗𝑗 ∗ has to wait until all the norms fall
shown in (10), which is based on the fact that, before being

all the 𝛼𝛼 and 𝛽𝛽 coefficients significant with a p-value of 0.05.


the present case, the average value of 𝑞𝑞̂𝑗𝑗 ∗ ,𝑃𝑃 computed as in (10)
below the common threshold value N. Just to have an idea, in
Also, the Root Mean Square Error (RMSE) computed on the
test-set was equal to 10.48 units of time. was equal to 7 units of time.

𝑞𝑞̂𝑗𝑗 ∗,𝑃𝑃 = max (𝑊𝑊𝑚𝑚 + ∑ 𝑤𝑤𝑗𝑗,𝑚𝑚 − 𝑁𝑁)


The neural network was trained using the well-known back
𝑚𝑚∈𝑅𝑅𝑗𝑗∗
propagation algorithm, using the Adam optimizer, i.e., an (10)
algorithm for first-order gradient-based optimization that takes 𝑗𝑗∈𝐽𝐽𝑃𝑃
advantage of adaptive estimates of lower-order moments. The
training process was based on: (i) 500 epochs, (ii) early To generate an additional benchmark, we also considered the
stopping with a patience of 50, (iii) validation split of 5%, (iv) “base-line” case of a system that is not equipped of a
batch size of 2048. After the training, relatively to the test set, forecasting system. In this case, whenever the customer places
we got a RMSE of 7.2 units of time that, as expected, is better an order, a negotiation is made and a corrected DD (higher
than that of the linear regression. This fact confirms a non- than that originally proposed by the customer), is generated in
linear dependence between response and explanatory the interval [DD; 1.2DD]. This base line case was tested both
variables. in a pure push-system and in the WLC condition. Obtained
results are shown in Table 2, where the average percentage of
4.4 Obtained results tardy jobs are reported.
To test the forecasting models, we replicated all the precedent
simulation runs using the multivariate linear regression and the Table 2. Final Results
neural network to estimate the LT of incoming jobs. Also, we Push WLC WLC WLC WLC
generated DDs based on these estimates, rather than generating Base Base Land Regr. N.N.
them in a purely random way (as we did in Section 4.2). More
specifically, to simulate a realistic operating scenario, we % Tardy 17.5% 22% 23.5% 13% 9.5%
introduced the possibility that the company can negotiate with
the customer, if the original DDs proposed by the customer are As it can be seen, there is a large difference among the tested
too tight. Supposing an equal bargain power among the vendor models (statistically confirmed by ANOVA and Bonferroni

2146
2019 IFAC MIM
Berlin, Germany, August 28-30, 2019 Davide Mezzogori et al. / IFAC PapersOnLine 52-13 (2019) 2092–2097 2097

post hoc test with a p-level of 0.05). As expected, the WLC REFERENCES
base line case is worse than the push one. Also, and quite
Bertolini, M., Braglia, M., Romagnoli, G., Zammori, F.
Probably, assuming that 𝑞𝑞̅ is constant and that 𝑞𝑞̂𝑗𝑗 ∗,𝑃𝑃 can be
surprisingly, the model by Land (2009) is even worse.
(2013a). Extending Value Stream Mapping: the synchro-
MRP case. International Journal of Production Research,
estimated as in (10) is overly optimistic and generates too tight 51(18), pp. 5499-5519.
estimates of the LT. Thus, the original DD is frequently Bertolini, M., Romagnoli, G., Zammori, F. (2013b) Assessing
accepted without negotiating with the customer, simply performance of Workload Control in High Variety Low
because the real LT was underestimated. This also explains Volumes MTO job shops: a simulative analysis,
why this approach is worse than the base WLC case where, Proceeding of IESM’2013, International Conference on
conversely, DDs are always bargained. Lastly, we note that the Industrial Engineering and Systems Management, pp. 362-
linear regression and, especially, the neural network make it 370.
possible to fully exploits the potentialities of a WLC system. Bertolini, M., Romagnoli, G., Zammori, F. (2013c) Simulation
Indeed, in this way, not only WIP (and consequently operating of two hybrid production planning and control systems: a
costs) are reduced, but also an important reduction of tardy job comparative analysis. Proceeding of IESM’2013,
can be obtained. International Conference on Industrial Engineering and
Systems Management, pp. 388-397.
5. CONCLUSIONS Bertolini, M., Romagnoli, G., Zammori, F., (2018). 2MTO, a
The paper focused on WLC and showed that, to fully exploit new mapping tool to achieve lean benefits in high-variety
its potentialities, this PPC technique should be coupled with a low-volume job shops. Production Planning and Control.
robust and consistent forecasting system, to define reliable Published on line, pp. 444-458.
DDs. If not so, especially in case of undersaturated systems Fernandes, N.O., Land, M.J., Carmo-Silva, S. (2014).
(i.e., average utilisation below 95%), relatively to a pure push Workload control in unbalanced job shops. International
system, WLC reduces WIP and queues but, on the other hand, Journal of Production Research, 52(3), 679 - 690.
it increases the probability of a job to be tardy. This is because Germs, R., Riezebos, J. (2010). Workload balancing capability
the more the WIP is constrained by the norms, the more jobs of pull systems in MTO production. International Journal
remain pending in the PSP. For this reason, although the total of Production Research, 48(8), pp. 2345-2360.
throughput time is reduced, the overall LT becomes higher Harrod, S., and Kanet, J.J. (2013). Applying work flow control
than that of a push system, and the number of tardy jobs in make-to-order shops. International Journal of
increases. To solve this criticality, we propose using either Production Economics, 143, pp. 620-626.
multivariate linear regression or neural network, to regress LTs Henrich, P., Land, M.J., Gaalman, G.J.C. (2007). Semi-
using the observed workloads as explanatory variables. Both interchangeable machines: implications for workload
models perform very well and outperform an equivalent push control. Production Planning and Control, 18(2), pp. 91-
system, both in terms of WIP and tardiness. In this regard, the 104.
neural network offers superior performance, as it can properly Hopp, W.J., Spearman M.L. (2003). To pull or not to pull:
map the non-linearity among the explanatory and the response What is the question? Manufacturing & Service
variables. Operations Management, 6(2), pp. 133-148.
Land, M.J. (2006). Parameters and sensitivity in workload
Even at the industrial level the implementation is control. International Journal of Production Economics,
straightforward and relatively cheap. The forecasting model 104(2), pp. 625-638.
can be developed almost at no cost using open source libraries Land, M.J. (2009). Cobacabana (control of balance by card-
and data can be easily collected from the MES (a prerequisite based navigation): a card-based system for job-shop
of any WLC system). The only complication may arise from control. International Journal of Production Economics,
the need to reproduce the manufacturing system by simulation, 117, pp. 97-103.
to generate additional data useful in the training phase. Thürer, M., Stevenson M., Silva, C. (2011a). Three decades of
However, we note that, since the linear regression can be fitted workload control research: a systematic review of the
with less data (relatively to the neural network), most of the literature. International Journal of Production Research,
times historical data are enough to initialize the model. In light 49(23), pp. 6905-6935.
of this, alternative ways to improve the performance of the Thürer, M., Silva C., Stevenson, M. (2011b). Optimising
regression (such as the use of a higher degree polynomial) workload norms: the influence of shop floor
should be investigated. characteristics on setting workload norms for the
workload control concept. International Journal of
Another possible enhancement could be that to use Production Research, 49(4), pp. 1151-1171.
probabilistic norms, to get a more precise description of the Thürer, M., Stevenson, M., Silva, C., Huang, G. (2012). The
current workload state and, hopefully, a better forecast. Lastly, application of workload control in assembly job shops: an
the same approaches could be extended to Cobacabana, a assessment by simulation. International Journal of
recently introduced PPC system, which simplifies WLC by Production Research, 50(18), pp. 5048-5062.
substituting the MES with loops of visual cards. Thürer, M., Land, M.J., Stevenson, J., Fredendall, L.D. (2016).
Card based delivering date promising in high-variety
manufacturing with order release control, International
Journal of Production Economics, 177, pp. 19-30.

2147

You might also like