Simulating the influence of the structure of unit-based pull production control systems on the workload balancing capability

UNIVERSITY OF GRONINGEN

FACULTY OF ECONOMICS AND BUSINESS Master Thesis – Technology & Operations Management

Simulating the influence of the structure of unit-based pull production control systems on the workload balancing capability

Author: Supervisor:

Christian Rodriguez Mallo - S1973916 N. Ziengs, Msc

Co-assessor:

Dr. J. Riezebos

20 June 2016

Academic Year 2015/2016

Abstract

Unit-based pull production control systems limit the amount of work that is allowed onto the production floor. As a result, a better control over the throughput time of orders is achieved.

Balancing the workload in production systems is possible if orders can follow several routes and the control is route specific. This is important since workload balancing can improve the throughput times of orders. Pull production control systems vary in terms of structure and configuration. This paper shows how the structure and configuration affect the workload balancing capability of unit-based pull production control systems in a divergent multi-stage production line.

The results show that the exchange of route-specific information downstream is crucial to successfully balance the workload. It is shown that this works best with an m-CONWIP structure.

Furthermore the decision concerning the configuration entails a trade-off between reducing the throughput time and possible other performance indicators.

Keywords: Pull production control, unit-based, generic model, simulation, structure,

configuration, discrete event simulation

Inhoud

Abstract ... 2

Preface... 6

1. Introduction ... 7

2. Background ... 9

2.1. Production planning and control ... 9

2.2. Pull production control systems ... 10

2.3. Push production control systems ... 11

2.4. Pull systems structure ... 12

2.5. Workload Balancing ... 15

3. Methodology ... 18

3.1. Research Method ... 18

3.2. Model Design ... 20

3.3. Experimental Design ... 21

3.4 Simulation Validity ... 23

3.4.1 Warm-up Period ... 23

3.4.2 Number or runs ... 24

4. Results ... 25

4.1 Outcomes – Experiment series 1... 25

4.1.1 Total Throughput Time ... 26

4.1.2 Shop floor Throughput Time ... 26

4.1.3 Order pool time ... 26

4.1.4 Configuration ... 27

4.2. Outcomes – Experiment series 2... 30

4.2.1 Total Throughput Time ... 30

4.2.2 Shop floor Throughput Time ... 30

4.2.3 Configuration ... 31

5. Discussion ... 32

5.1 CONWIP ... 32

5.2 Kanban ... 33

5.3 POLCA ... 34

5.4 Base Stock Policy ... 36

5.5 m-CONWIP ... 37

5.6 Limitations ... 38

5.7 Theoretical Implications ... 39

5.8 Managerial Implications ... 39

6. Conclusion ... 40

6.1 Future Research ... 42

References ... 44

Appendix A ... 48

Appendix B ... 49

Appendix C ... 50

Appendix D ... 51

Preface

Over the course of the last six months I have enjoyed working on this thesis although

acknowledging the well-known and expected ups and downs. As expected I too have not only

relied on myself during the creation of this work. I would therefore first like to thank the

supervisor of this work, Nick Ziengs, for his close involvement and ever so fruitful discussions

whenever doubts and problems came about. Thanks to his extensive feedback and flexible

attitude I was able to surpass most of the hurdles encountered during the elaboration of my

thesis. Furthermore, I would like to mention my pleasant collaboration with Fiodor Bodnar who

has been a great source of motivation and for our extensive discussions on our thoughts and

work. I would also like to thank the co-assessor of this work Dr. Jan Riezebos for his time and

effort in evaluating this work. Finally I would like to thank my parents, sister and girlfriend for

their unconditional support and good care all through the last six months of occasional stress and

doubts.

1. Introduction

The throughput time performance of a production line is largely determined by controlling the work in process that is allowed onto the shop floor (Ziengs, Riezebos, & Germs, 2012). Hopp

& Spearman (2004) define such mechanisms that limit work in process as pull systems. There is a large variety of pull-type systems, for an extensive review of all the systems we refer to González- R, Framinan, & Pierreval (2012). The commonly known systems are Kanban, CONWIP, m- CONWIP and POLCA. These systems distinguish themselves by their structure. Which means, at which place or places in a production line the WIP is limited. By how much the WIP it is limited, is referred to as the configuration. In literature the performance of these systems has been related to their workload balancing capabilities (Germs & Riezebos 2009). However, the influence of the difference in structure and configuration on the workload balancing capability of pull production systems is not yet thoroughly discussed. However, Ziengs et al. (2012) argue that workload balancing can help companies gain a competitive advantage. The focus of this thesis is therefore on determining how the structure and configuration of pull production control systems influence the workload balancing capability. To achieve this a flexible model is developed which can simulate pull production systems for a divergent production line.

The workload balancing capability of varying structures and configurations of pull

production systems in divergent production lines is a relevant problem since previous research has

focused mainly on straight production lines (Liberopoulos & Dallery 2000; Gaury et al. 2000). By

looking at a divergent production line we introduce additional possibilities of variability, namely,

variability in the routing of orders.

An extensive search for a generic control system design has shown that there is little done in this aspect. González-R & Framinan (2009) and Gaury et al. (2000) have designed a generic model to make a selection between KANBAN, CONWIP and a hybrid, although this was for a straight line. However as we have learned from Germs & Riezebos (2009), the most effective workload balancing control systems for divergent production lines are POLCA and m-CONWIP.

Germs & Riezebos (2009) found that route-specific control systems perform better in terms of workload balancing capability. Germs & Riezebos (2009) included several types of possible variability in their simulations and have only looked at CONWIP, POLCA and m-CONWIP. In order to test whether route-specific information is indeed crucial, it is important to isolate route variability as the only influential factor. Furthermore structures that do not necessarily allow for route-specific information to be available at the order release should also be included. We therefore aim to show how the difference in structure and configuration actually influences the workload balancing mechanism. Hereby focusing on how these system characteristics cope with routing- variability. This is executed with a unit-based release mechanism which releases order based on the number of orders on the shop floor. This release mechanism is chosen since it is more commonly used in practice, due to its relative simplicity compared to a load-based control. Which releases orders based on the capacity in terms of the time necessary to process an order.

In conclusion, this research will provide insights into the influence of varying

characteristics which come along with the different structures and configurations of unit-based

pull production systems, on the workload balancing capability of pull production systems in a

divergent production line.

2. Background 2.1. Production planning and control

Production systems can be defined as “a set of interrelated elements that are designed to act in a manner that generates final products whose commercial value exceeds the cost of generating them” (Maccarthy & Fernandes, 2000, p.485). Production systems tend to be complex systems and controlling them therefore, poses a challenge. In order to gain some control over the performance of production systems, production planning and control (PPC) systems have been developed. The main purposes of a PPC system are that: 1) they help determine the amount to be produced in order to fulfill customer demand, 2) they are used to plan the order of raw materials, 3) they balance the available resources (capacity) and 4) they control and plan order release to the system (Zäpfel & Missbauer, 1993). Gaury, Kleijnen and Pierreval (2001) state that the influence of PPC systems is major in terms of inventories, production delays and make span, and thus influencing the overall competitiveness of a company. Although the major influence of PPC is acknowledged in literature, the questions remains which PPC best copes with the often changing state of companies and their production systems. (Banerjee 1996, p.58), for example, states that

“Experience has shown that while millions have been spent on manufacturing systems [...] no real solution to the need of greater responsiveness and flexibility has been found”.

This should further demonstrate to production companies the importance of thoroughly investigating which PPC system to use. In order to choose the right PPC system one must have an understanding of the characteristics of the different options.

The general distinction within production control can be made between push and pull

control systems. The main difference between push and pull systems is that in pull systems a

production process at a certain stage of the system is started to fill the gap left by a part that has gone on to the next stage, while push systems produce to fulfill demand without considering the status of the system (Akturk & Erhun 1999; Hopp & Spearman 2004). To get a better understanding of the individual control systems they are discussed at length in the next sections.

2.2. Pull production control systems

Pull production control systems are used to control throughput times of orders by limiting the amount of work in process (WIP) in a system or parts thereof (Hopp & Spearman 2004). The interest for pull systems was sparked by the introduction of the just-in-time (JIT) philosophy on production in the late seventies which aims to balance service level with a minimal WIP level (González-R, Framinan, & Pierreval, 2012). The service level is defined by Karaesmen & Dallery (2000) as the fill rate which represents the proportion of demand that can be satisfied from inventory at the moment of order arrival.

The foundation of JIT production and therefore pull systems lays in Japan. In the seventies,

MRP and other computerized control systems were gaining traction in American and European

industries. Meanwhile in Japan, partly due to the lack of computerization of their industry, a more

resource conservative control system was emerging (Hopp & Spearman 2004). Under the watchful

eye of Taiichi Ohno, Toyota started to implement the “Toyota Production System”. This system

focused on making as much goods as possible, but in a continuous flow (Ōno 1988). When the

times of abundant demand started to diminish all over the world, and efficiency started to become

important in order to maintain the business results, American and European industries began to

grasp the value of JIT production.

The Toyota Production System includes a card-based pull system known as Kanban (Liberopoulos & Dallery 2000). In card-based pull systems the WIP is controlled by control loops in the system. The amount of cards within such control loops determines the amount to which WIP is limited. Whenever a downstream production stage is available a card is released to the upstream station, signaling that work can be released (Ziengs et al., 2012). All card-based pull production systems work in this same general fashion, nevertheless there are several ways to implement this basic mechanism. The implementation is defined by the characteristics of the systems, such as the structure and configuration.

Pull production systems can also be characterized by the way they release orders. Pull systems can either be route-specific or non-route-specific and additionally product-anonymous or product-specific (Germs & Riezebos 2009). Route-specific control systems do not look at the product type but rather have cards for product routes. This means that to release an order at the right time it needs to be aware of the capacity at all the workstations that lay within the (partial) route of the arrived order. Non-route-specific control however disregards information concerning the capacity within the route. Product-anonymous control disregards all the information concerning the order and just releases an order based on the availability of a card. A visualization representation of these different types adopted from (Ziengs et al. 2012, p.4360) can be found in appendix A.

2.3. Push production control systems

The opposite of pull systems, as the name implies, are push systems. In general, push

systems do not produce depending on the status of the system such as pull systems but rather work

towards a predetermined target. Work is then released onto the work floor either constantly or following a pre-determined schedule; the so-called Master Production Schedule (MPS) (González- R et al., 2012). Another, more concise, definition of the difference between push and pull is given by Bonney et al. (1999). They state that the flows of information and materials in push system go in the same direction, while in pull systems the flows move in contrary directions. This distinction is rather important for this thesis considering that we are focusing on the transfer of route-specific information within the system.

2.4. Pull systems structure

As mentioned in section 2.2 card-based pull systems distinguish themselves by their structure and configuration (Gaury et al. 2001). The structure is determined by the placement and size of control loops, whereas the configuration entails the WIP limit that is established within such a control loop. A control loop can contain one or multiple workstations, depending on the chosen control system (González-R et al. 2012). We will now consider the structure of KANBAN, CONWIP, m-CONWIP and POLCA in more detail since these pull systems have previously been used in similar studies (Germs & Riezebos 2009; Ziengs et al. 2012) that addressed workload balancing capabilities of unit-based pull production control systems. Furthermore these are the systems that are simulated in this thesis.

Kanban uses cards to control and limit the amount of jobs that are released between workstations and thus place a control loop containing only one workstation. A Kanban control loop can be placed at every workstation of a production system or at parts thereof. (Sharma &

Agrawal 2009). A schematic view of the basic structure of KANBAN is shown in figure 1. Because

the control loops only contain a single workstation the only information that is transferred between control loops is the availability of a card.

Figure 1. Schematic representation of the KANBAN structure

CONWIP does not allow for placement of control loops at specific workstations. Instead

CONWIP is based on limiting the WIP of the whole line. The order release is triggered by a

finished product at the end of the line. Again leaving information concerning the capacity of

individual workstations out of consideration. A schematic view of a simple CONWIP structure is

shown in figure 2.

Figure 2. Schematic representation of CONWIP structure

The control loop in an m-CONWIP system contains all the workstation of a certain route within the system. This means that the information flow accounts for the capacity of a whole route when releasing a card to the queue. A schematic of this structure can be seen in figure 3.

Figure 3. Schematic representation of m-CONWIP structure

POLCA is a pull production system where the control loops overlap each other. This means

that whenever an order is finished within a loop the system will wait with releasing the card to be

attached to a new order until a card of the successive loop has become available. Hence, the finished order can continue. A schematic of this structure is shown in figure 4.

Figure 4. Schematic representation of POLCA structure

2.5. Workload Balancing

Workload balancing is the ability of a production system to stabilize the queues before the

workstation based on a certain order release strategy (Land & Gaalman 1996). Pull control systems

are such control strategies. In order to balance workload the control strategy has to be able to

prioritize orders based on the available capacity upstream. This implies that if more queues exist

in a system there are more points at which workload can be balanced. Ziengs et al. (2012) however,

found that it isn’t necessarily better to balance the workload at as much points as possible. They

found that the structure and limit of control loops within a system have a great influence on the

workload balancing capabilities of that system (Ziengs et al., 2012).

This raises the question of how the structure and configuration actually influence the workload balancing capability. To address this question we will discuss how and where each of the systems described in section 2.4 would actually balance workload.

As can be seen in figure 1-4 some systems create more queues than others. Kanban being the system where the most queues are created (7) and CONWIP the system with the least queues, namely one. In theory a Kanban system therefore contains more chances to balance the workload.

However, we have already established that in a divergent topology where routing variability exists

the benefits of workload balancing can be annihilated if orders are in the queue for too long due to

lacking information on upstream capacity. We therefore expect that systems with a structure that

allow for (partial) route-specific information to reach the point of order release will balance the

workload better. If we look at the structures as described in section 2.4 we identify two systems

which allow for route-specific information to travel between loops, namely m-CONWIP and

POLCA. In an m-CONWIP system a loop contains all the workstations of a certain route and

workload is balanced at arrival. Whereas for POLCA, the workload is balanced at the queue that

is created at the workstation where two control loops overlap. So the information concerning a part

of the route is made available to the whole previous control loop at the point where workload is

balanced. However, if we look at CONWIP and KANBAN we see that it is non-route-specific.

Based on the background as described in this chapter ¹ we expect that the workload balancing capabilities of unit based pull production control systems may be closely related to its ability to transfer route-specific information downstream. We have seen that the structure does influence the capability of passing this information downstream due to how and which workstations are placed within a control loop. To thoroughly investigate the influence of the structure and configuration a flexible model should be developed that allows the simulation of systems that are route-specific as well as non-route-specific. In this case the systems of interest are m-CONWIP, CONWIP, Kanban and POLCA . Although previous research has already proven that some of these systems have workload balancing capabilities, we can better evaluate the influence of the structure and configuration by isolating the route variability from other possible causes that may influence the performance of these systems. Furthermore, we expect that the configuration that is chosen for these systems is of importance since we want to find the optimal performance of the structures. Which, as we have established, not only depends on the reduction of throughput times but also obtaining a certain service level.

In the end the following research question and sub-questions are to be answered:

How does the structure of a unit-based pull system affect its workload balancing capability?

SQ1: Where in the production line should workload be limited to achieve workload balancing?

SQ2: Which control system configurations work best for which structure?

1

The search-strategy and its outcome used to retrieve the literature used in this chapter can be found in

Appendix B

3. Methodology

In order to test the influence of the control system structure and configuration on the workload balancing capabilities of unit-based pull systems a discrete event simulation model is developed. The tools and their application are described in this chapter. Following from this the experimental design will be described.

3.1. Research Method

Discrete event simulation is commonly used to analyze industrial systems such as manufacturing systems (Eduard Babulak 2010). Simulation serves as a computational or mathematic representation of reality that can support decision making. Simulation has demonstrated to be useful for manufacturing environments due to the presence of variability, interconnectedness and complexity in production systems. This is further backed by Robinson (2004, p.5) which states that predicting the system behavior is hard when one or more of these factors apply to the system. Simulation offers a way to gain insights into these influences and show the behavior of the system accordingly. Furthermore Robinson (2004, p.8) gives four reasons why simulation is preferred over real system experimentation: 1) cost, 2) time, 3) control over experimental conditions and 4) the fact that the real system must exist.

In this thesis we will use discrete event simulation to investigate which underlying

mechanisms are of influence on the workload balancing of different pull production structures and

configurations. The model used should simulate a system that allows for simulation of several

production control systems. Experimenting with a real system with this capability is impossible

since none exists and adapting system structure and configuration in real life would lead to high costs.

Discrete event simulation software has had the interest of computer scientists from the 1950’s, when coding languages were used. Only a decade later the first dedicated simulation languages were developed (Dahl & Nygaard 1966). Currently the focus is on visual interactive simulation (VIS) software which shows the actual model, highly improving the understandability of the simulation model. In this thesis however the Python (Python Software Foundation 2016) programming language will be used. Python is object oriented and has a dedicated simulation library known as SimPy (Team SimPy 2016), enabling the simulation of highly complex systems.

We will use programming due to the modelling flexibility necessary and the high run-speed essential because of limited time availability. How these factors compare between programming and other simulation methods is presented in table 3.1 below.

Table 3.1. Adopted from (Robinson 2004, p42, Table 3.2).

Note. Comparison of simulation methods based on important features

3.2. Model Design

The model that will be developed will be able to simulate all possible structures and configurations of the selected control systems as described in section 2.4. A three stage divergent topology will be used as production line, as has been done by Germs & Riezebos (2009) and Ziengs et al. (2012). This is closer to reality, where straight lines are not so common. However, a symmetric divergent topology is not a complete reflection of reality, since in reality asymmetric routings are most common (Henrich et al. 2003). A schematic of such a topology is represented in figure 5.

Figure 5. 3-stage divergent topology (adopted from Germs & Riezebos (2009, p.3)

The production line will consist of three stages meaning that at each stage the amount of workstations doubles, giving us a total of seven workstations. Workload balancing can take place at the first and second stage. As can also be seen in figure 4 the amount of routings is four (the number of end stations) and every route is equally likely to happen. In order to keep the utilization at each stage at the same level for all workstations the processing time is doubled at every stage.

By doubling the amount of workstations and accordingly double the processing time the utilization

level will be kept even and capacity will be the same in each stage. Which experimental variables will be used and their corresponding values is described in the next section.

3.3. Experimental Design

In order to test and validate the results, the paper of Ziengs et al. (2012) will be used. We will therefore base the experimental factors on theirs. The experimental factors and their corresponding values are presented in table 3.2. The utilization will be fixed at 90%. Likewise, the inter-arrival time and processing time will be constant. This choice is made since we focus on the influence of the structure and configuration and want to isolate the influence of routing-variability that exists in a divergent topology.

Table 3.2. Experimental factors and their corresponding values

Factor Experimental levels

Configuration characteristics Card type

Number of cards Series 1:

Series 2:

Order arrival pattern

Unit-based 1 – 20(n), ∞ 1 – 45(n), ∞

Inter-arrival time Constant

Utilization 90%

Batch size

Processing time variability Processing time

1

Constant

To simulate the different structures and configurations the placement of the control loops

will be defined for the model. Table 3.3 contains the loop placements and the workstations that

fall within each loop for each pull production control system. If no control is performed at a

workstation the card count will be set to infinity. The card count for each loop will vary per

experiment, ranging from 1 to 20, which yields 20 experiments per structure. Furthermore the card count will be the same for all the loops during an experiment.

The structures are divided into two separate series. Experiment series 1 will entail experiments with existing and conventional structures: m-CONWIP, Kanban, POLCA and CONWIP. Whereas in experiment series 2 adapted structures will be tested. This will be a 1) POLCA structure without work restriction between the start and the first stage, 2) Base Stock Policy, adaptation of Kanban implementation where production is triggered at each individual workstation whenever a product leaves the queue and 3) A semi non-route specific structure that controls workload using two control loops covering one half of the system each.

Table 3.4 and 3.5 further addresses how the structure of each of the production control systems is modeled by representing how and which workstations fall within a control loop.

Table 3.4 Control loop placement for Experiment Series 1

Structure Loop 1 Loop 2 Loop 3 Loop 4 Loop 5 Loop 6 Loop 7

POLCA W1:W2 W1:W3 W2:W4 W2:W5 W3:W6 W3:W7 /

CONWIP W1:W2:W3:

W4:W5:W6:

W7

/ / / / / /

m-CONWIP W1:W2:W4 W1:W2:W5 W1:W3:W6 W1:W3:W7 / / /

Kanban W1 W2 W3 W4 W5 W6 W7

Table 3.5. Control loop placement for Experiment Series 2

Structure Loop 1 Loop 2 Loop 3 Loop 4 Loop 5 Loop 6 Loop 7

Adapted POLCA W2:W4 W2:W5 W3:W6 W3:W7 / / /

Base Stock Policy W1:W2 W1:W3 W1:W4 W1:W5 W1:W6 W1:W7 /

Adapted m-CONWIP W1:W2:W3 W4:W5:W6:W7 / / / / /

The outcomes which will be used as performance indicators will be the total throughput time (TTT), which is the amount of time between the arrival of the order and the order leaving the system. The shop floor throughput time (STT), which is the time an order has actually spend being processed. And finally the order pool time (OPT) which represents the time an order has spent waiting in the queues. These are chosen because the aim is to reduce the TTT which is the result of adding the STT and OPT. Furthermore these have been used in previous research (Ziengs et al., 2012) which will serve as a benchmark to validate and compare the results.

3.4 Simulation Validity

In this section we discuss which simulation parameters are chosen and how they are validated. This is done by explaining how the warm-up period and number of runs are been chosen and the results of the methods used to test the validity of these parameters is presented.

3.4.1 Warm-up Period

The warm-up period is essential to obtain valid and reliable results since the system needs

to obtain a steady state before reliable data is outputted. The first thing is to determine the type of

simulation that is performed, in our case it concerns a non-terminating simulation since we are not

simulating a productions schedule (which reaches an end) but interrupt the simulation (production)

after a certain time has elapsed. Since our simulation is non-terminating at a certain point a steady-

state will occur where variability is still present but follows a certain distribution (Robinson 2004,

p.138). There are several ways to determine the warm-up period however in this research the

Welch method (Welch 1983, pp.289–292) is used since this has been argued to be conservative,

providing us with the confidence of choosing the correct parameter value.

The Welch (Welch 1983, pp.289–292) method consists of plotting the moving average of these observations on a time-series. By increasing the window size a smooth line will become visible. Than by graphically identifying where the line start to flatten will give the point at which the system has reached a steady-state. For the simulations performed in this research the Welch method shows a steady-state after 800 time units. For the simulations however a warm-up period of 2500 time units has been used since it is better to use conservative values to ensure validity and robustness of the outcomes and has also been applied in the benchmark research by Ziengs et al.

(2012). The graph resulting from applying the Welch method can be found in appendix C.

3.4.2 Number or runs

The number of runs (replications) is important since it ensures that enough samples are drawn from the seed numbers used to generate the random numbers (Robinson 2004, pp.151–153).

By doing this one reduces the variability that is observed between each string of random numbers

created. In order to select the number of runs we have ran the pure push system since this should

result in the highest variation and using the mean output of 100 runs calculate the cumulative mean

and plot the outcome. This yields a graph with a line that goes flat after a certain amount of

replications. In the case of our research the line goes flat after 40 runs. Again to ensure the validity

of the experiments however a higher value of 100 runs is chosen as was the case in the paper by

Ziengs et al. (2012).

4. Results

In this section the results of the experiments are presented. First the results for experiment series 1 will be presented. In the following section the results of experiment series 2 are presented.

4.1 Outcomes – Experiment series 1

As described in section 3.3 CONWIP, m-CONWIP, POLCA and Kanban have been simulated in this experiment series 1. The experimental factor in all cases is the amount of cards that is allowed to be attached at each stage which was varied between 1 and 20. In each experiment the varied card count was fixed for all the control loops. In order to check the results a non- restricted structure (push system) has also been simulated. The outcomes of the experiments are all compared to the performance of the pure push structure outcomes. The results are presented below in table 4.1 as units of time or percentages.

Table 4.1. Experiment outcomes – Pull structures vs. Push structure performance

Structure STT OPT TTT % TTT

Reduction

% STT Reduction

Push 22.36413 0 22.36413 / /

POLCA 22.07725809 0.592670995 22.66992909 -1.3673635% 1.2827328%

Kanban 22.56576294 0.022841911 22.58860486 -1.0037266% 55%

m-CONWIP 10.72092386 10.427686 21.14860986 5.4351332% 48%

CONWIP 22.36413013 0 22.36413013 0% 0%

Note: The STT, OPT and TTT are in unit of times. The relative performance in % is bold. Furthermore the TTT reduction is

highlighted by grey cells. These result represent the outcomes using the optimal configuration.

4.1.1 Total Throughput Time

As can be seen from table 4.1 there are three structures that show no workload balancing capability under these circumstances: POLCA, Kanban and CONWIP. In the case of CONWIP no improvement compared to a pure push structure can be observed, the improvement in total throughput time (TTT) is 0%. POLCA and Kanban show a more remarkable outcome since the structures actually performed worse in terms of balancing the workload and thus reducing TTT.

Using Kanban an increase in TTT of 1,00% is observed. For POLCA the increase of TTT is 1,36%.

The pull structure that does show an improvement in TTT is m-CONWIP. The reduction in TTT is 5,44%.

4.1.2 Shop floor Throughput Time

Furthermore as can be seen in table 4.1 we have looked at the reduction in shop floor throughput time (STT). Looking at the STT we observe one structure that does not improve, namely CONWIP. If we look at Kanban and POLCA which showed an increase in TTT we now see that there is however a reduction in terms of the time the orders spent on the shop floor. The STT can be reduced by 55% when Kanban is used and 1,28% using POLCA.

The structure that have already shown to reduce TTT shows a comparable results in terms of STT. m-CONWIP shows a reduction is STT of 48%.

4.1.3 Order pool time

The last performance indicator taken into consideration is the order pool time (OPT). This

indicator specifies the units of time the order is in the pool (i.e. queues). Intuitively when restricting

the amount of cards that can be attached within a loop one does increase the OPT since an order

has a higher chance of getting blocked. Therefore the order pool time serves to nuance the STT reduction since although the order spends less time on the floor it will be awaiting processing for a longer period of time. Therefore the TTT is composed by adding the OPT to the STT. The threshold level of cards is than obviously whenever the difference between STT and OPT is the biggest in favor of the STT.

From table 4.1 we see that for the structures that do not show an improvement of TTT the OPT is actually very small. POLCA and Kanban have an OPT of 0,592 and 0,022 units of time whereas m-CONWIP which has shown to reduce the TTT has a OPT of 10,42.

This is in line with what one would expect since in order to balance the workload the amount of work within the given loops needs to be restricted, which causes longer OPT. Since only m-CONWIP has shown to have workload balancing capabilities in this series of experiments it is to be expected that the OPT for m-CONWIP is bigger.

4.1.4 Configuration

The results mentioned in the previous sections are the obtained at the optimal card count.

Table 4.2 shows the card count that yielded the results for each structure. Figure 6 -9 show how the TTT developed per configuration. In the next chapter the implications and influence of the configuration are discussed.

Table 4.2. Optimal card count per structure as simulated in experiment series 1

Structure Card count

POLCA 20

Kanban 20

m-CONWIP 1

CONWIP 3

Figure 6. TTT per configuration (card count) for POLCA

Note: The black line represents the TTT of a Push system. Upper boundary for TTT is chosen at 50.

Figure 7. TTT per configuration (card count) for Kanban

Note: The black line represents the TTT of a Push system. Upper boundary for TTT is chosen at 50.

0 5 10 15 20 25 30 35 40 45 50

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

TTT

Card count

POLCA

0 5 10 15 20 25 30 35 40 45 50

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

TTT

Card count

Kanban

Figure 8. TTT per configuration (card count) for CONWIP

Note: The black line represents the TTT of a Push system. Upper boundary for TTT is chosen at 50.

Figure 9. TTT per configuration (card count) for m-CONWIP

Note: The black line represents the TTT of a Push system. Upper boundary for TTT is chosen at 50.

20,4 20,6 20,8 21 21,2 21,4 21,6 21,8 22 22,2 22,4 22,6

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

TTT

Card count

m-CONWIP

0 10 20 30 40 50

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

TTT

Card count

CONWIP

4.2. Outcomes – Experiment series 2

In this chapter the results of experiment series 2 are presented and discussed in the same order as for experiment series 1. After this section the results will be compared and further discussed. The result are presented below in table 4.3.

Table 4.3 Results Experiment Series 2

Structure STT OPT TTT % TTT

Reduction

% STT Reduction

Push 22.36413 0 22.36413 / /

Adapted

POLCA 20.69366203 0 20.69366 7.46941% 7.46941%

Base Stock

Policy 22.36413013 0 22.36413013 0% 0%

Adapted m-CONWIP

22.36413013

0 22.36413 0% 0%

Note: The STT, OPT and TTT are in unit of times. The relative performance in % is bold. Furthermore the TTT reduction is highlighted by grey cells. These result represent the outcomes using the optimal configuration.

4.2.1 Total Throughput Time

Again it is clear that under the given circumstances as stated in table 4 there are two systems that do not show workload balancing capability. The Base Stock Policy and the adapted m- CONWIP which has two control loops (containing three workstations each) show 0%

improvement compared to a pure push system. Adapted POLCA, with control loops placed only over the last two stages however shows a decrease in TTT of 7,47%.

4.2.2 Shop floor Throughput Time

As for STT the experiments yield the same results as for the TTT. Base Stock Policy and

the adapted m-CONWIP variant show no improvement whatsoever. The adapted POLCA variant

shows an in improvement of STT of 7,47%. As for the order pool time all the systems perform the same, no order pool time is observed.

4.2.3 Configuration

The results mentioned in the previous sections are the obtained at the optimal card count.

Table 4.2 shows the card count that yielded the results for each structure. Since the adapted m- CONWIP and Base Stock Policy have not shown any meaningful result only the development of the TTT per configuration of the adapted POLCA structure is looked at in more detail. The graph is presented in figure 10. In the next chapter the implications and influence of the configuration are discussed.

Table 4.2. Optimal card count per structure as simulated in experiment series 2

Structure Card count

Adapted POLCA 1

Base Stock Policy 45

Adapted m-CONWIP 15

Figure 10. TTT per configuration (card count) for adapted POLCA

20,69 21,51

21,94 22,16

22,27 22,31

22,34 22,35

22,31 22,31

22,31 22,31

20,5 21,5 22,5

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

TTT

Card Count

Adapted POLCA

5. Discussion

From the observed outputs of the experiments we can say that in general, m-CONWIP shows the ability to balance workload while POLCA, CONWIP and Kanban do not. As argued in section 2.5 the influence of structure and configuration on workload balancing is high. In this chapter we discuss possible reasons for these differences and explain the results. The theoretical and managerial implications are than presented.

5.1 CONWIP

CONWIP contains only one loop where depending on the configuration a certain amount

of cards are allowed and used to control the work in progress that is allowed into the production

system. This means that the order release decision is triggered only once an order has passed

through all the stages of production and has left the system. While CONWIP in itself is a pull

system since it restricts the amount of work that is allowed into the system based on the state of

the system, once the orders are actually in the system it will behave the same way a push system

does. Nucci & Grieco (2012) further argue that CONWIP has both a physical and a virtual

authorization procedure. The cards in the system are the physical system which triggers order

release, however a scheduling and sequencing system (such as MRP) serves as a virtual system

that needs to decide which order is released. Because of this lack of information between the two

authorization systems Nucci & Grieco (2012) predict this may significantly impact the balancing

capability of the system. In this research this is reflected by the 0% of improvement upon a pure

push system in STT as well as TTT that have been found. Furthermore, our implementation could

be described as a pure flow shop, since we have variable routing but fixed route directions. In such

a production setting it may occur that an order is released with a routing that is at that point not

able to accept another order. Because the release decision is not based on downstream information of any specific route. Finally, this result does fulfill our expectations since it is the same result that was obtained in the benchmark research (Germs & Riezebos 2009). In terms of configuration the optimal performance, which in this case is the same as a push system, is obtained using three cards.

Less cards will increase the TTT whereas more cards will make no difference as is depicted in figure 9.

5.2 Kanban

Kanban restricts WIP at every individual workstation. From the results we have seen that

the Kanban structure does actually performs worse than a pure push system given the used

configuration. This could be due to the fact that even though work in progress is effectively

restricted it won’t be able to balance the workload since a control loop (one workstation) is not

aware of the capacity in different routes of the system upstream or downstream. Considering the

high order pool time we see that blocking due to the lack of information causes the system to

actually perform worse than a pure push system. This blocking may be caused since the system is

actually focused on replenishment, meaning that it will immediately start producing after a card

has become available at the successive station. Krishnamurthy & Suri (2009) therefore argue that

against general belief, Kanban is actually not or barely suitable for production systems with high

variability in products or other forms of complexity. Kanban in a divergent production line with

product variability can therefore be expected to cause blocking and perform badly which

corresponds with our results. In terms of configuration as depicted in figure 7 we see that restricting

the system will increase the TTT. When restriction is lowered the system start to stabilize into the

same TTT performance as a push system.

5.3 POLCA

From the results of experiment series 1 we have seen an unexpected outcome when compared to the research by Ziengs et al. (2012) and Germs & Riezebos (2009). In their simulation study POLCA showed to have workload balancing capabilities using the same configuration as applied in this research. In the current case we have seen that POLCA did perform worse than a push structure. Although this may appear contradictory there is a possible explanation as to why this is observed. We should first focus on how POLCA is applied in a divergent topology production line. Each POLCA loop contains two workstations. Since our production line branches out into separate routes it means that the third stage contains two loops which share a common workstation. The first loop is not aware of the production capacity available downstream but is free to select whichever order to let into the system. It appears there may again be a problem with a lack of information. The loop containing the ‘shared’ workstation (workstation 2 and 3 in figure 11) of the last stage is also in a loop with the initial stage. Whenever an order is completed it will therefore trigger the release of a new order. The problem then arises of which orders to prioritize in order to balance the workload over the possible routes. If the first workstation decides to release an order for which no capacity is available in one of the branches downstream it may cause blocking, and as a consequence long waiting times. We suspect that why this did not occur in the previous researches is most likely due to a difference in implementation of the card attachment mechanism in the code. This may perhaps have influenced on how cards are attached between two overlapping loops. This will be further addressed in section 5.6.

In addition to general POLCA we have also simulated a version which only restricts WIP

at the later stages. This should serve to answer the research question concerning the placement of

control loops within a system. As was described in section 4.2 the POLCA implementation with

control loops only at the last stages reduces the TTT as well as the STT. This is proof of its workload balancing capability, furthermore this is in correspondence with the results found by Ziengs et al., (2012). This improvement can be explained by the fact that the possibility of balancing at a further stage in a divergent topology increases since it contains more workstations.

Furthermore by shifting the balancing decision towards the last stages reduces the chance and impact of blocking. The logic behind this is illustrated in figure 11.

Figure 11 – Control loop placement in last stage POLCA structure

Note: The colored loops (red) represent the control loops that were used in the adapted POLCA simulation.

As can be seen in figure 11, due to missing information from the downstream stations blocking

can occur at station 1 already. This would mean that both workstation 2 or 3 and their successive

stations (4, 5, 6 and 7) can suffer starvation. However if we place the control loop over the last

stage (red loops) the impact when balancing in the second stage can in the worst case only impact

one workstation.

In terms of configuration there is a noteworthy outcome for the adapted POLCA system as can be seen in figure 10. The optimal configuration is a card count of 1. However, once it has reached the same TTT performance as a push system at a card count of 8 there is a drop in TTT again when using 9 to 12 cards. This is a good example of where the tradeoff between optimal TTT performance and service level should be made. By highly restricting the system the best TTT performance is obtained. However, there is also a range of card count at which TTT drops but a higher service level can be obtained.

5.4 Base Stock Policy

As part of experiment series 2 the Base Stock Policy (BSP) has also been simulated. The

BSP is based on setting lower bound limits to the queues after a workstation. So that whenever an

order leaves the queue immediate production for replenishment is triggered. The control loop

structure is similar to that of Kanban but differentiates itself in terms of the focus on a minimal

WIP rather than on limiting WIP based on demand. In our simulation BSP does not improve the

TTT compared to a push system. If we look at the characteristics of a BSP this is not a surprising

result since the focus is not as much on performance but rather on maintaining a certain stock level

so that lead times are reduced and a predefined service level is achieved. One would therefore not

necessarily expect the system to perform better in terms of TTT. We can only argue that this may

be remarkable since we have found that route-specific information will help balance the workload

and a BSP facilitates this route-specific information. However although this information is

available the release decision of an order is not based on it. Which may imply that in divergent

production systems as used in this research, the BSP will not be of much help in MTO

environments where product- and thus route-variability exists.

5.5 m-CONWIP

By looking at the results we see that m-CONWIP does show workload balancing capability by reducing the TTT as well as the STT. If we compare the structure of m-CONWIP to that of the systems that did not show workload balancing capability it immediately becomes clear that there is one very distinct difference. This is the fact that m-CONWIP control loops are actually route- specific. Every route is a loop and vice versa as can be seen in figure 12.

Because of this the problem of lack of information about the system status downstream

which was observed in other structures, does not arise using m-CONWIP. The point at which work

is balanced and an order release decision is made is therefore the same. This prevents the system

from blocking routes, hence the improvement in TTT and STT. In terms of the configuration as

depicted in figure 9 we see that the optimal performance is achieved at a card count of 1 and that

the TTT will start increasing by increasing the card count. This implies that the further you restrict

an m-CONWIP system the better it will perform. This optimal configuration is best explained by

recalling the fact that we implemented the order arrival in such a way that every route is equally

likely to happen. With this knowledge and the fact that the release is based on first come first serve

we know that the system is able to distribute each order immediately after arrival. Because of the

maximal restriction to just one card there will be no blockage other than waiting for the previous

order with the same route to be finished.

Figure 12. m-CONWIP control loops in the simulated divergent topology

5.6 Limitations

The limitations of this research focus around the implementation of the code. As we have

seen in this chapter there is a discrepancy between the results found by Germs & Riezebos (2009)

and Ziengs et al. (2012), and ours. Especially in the case of POLCA. As we have discussed, this is

probably due to a difference in the code used for the simulations. Especially the part which

executes the mechanism of attaching cards. As explained in section 2.3 there are overlapping loops

in a POLCA system. A card should only be released once the card of the successive loop has

become available and has been attached. The code used for the simulations was not able to perform

this attachment mechanism. If this mechanism is not fully functional the benefit of route-specific

information, that is normally available using overlapping loops, is lost. Hence, the system did not

perform as expected. An additional limitation is the fact that no experiments have been performed

where card counts varied between the loops. This could have served as an introduction to hybrid systems where the restriction in the system is not the same at all stages.

5.7 Theoretical Implications

We have shown that m-CONWIP which loops cover a whole route reduces TTT as well as STT because information about capacity downstream is taken into account at the order release point in the first stage. We hypothesized the route-specific characteristics of this system and thus expected this outcome. Furthermore, we have shown that even though POLCA has overlapping loops at workstations, it does not allow the information to travel all the way upstream. This is explained by the fact that in a divergent production line there are shared workstation for different routings. This causes blocking and waiting time in all the successive routes if only one route is blocked and is in accordance with the findings of Ziengs et al. (2012). After removing the WIP limit in the first stage however POLCA shows the performance expected. This shows that the structure of control loops in a unit-based production system does affect its performance in terms of workload balancing since it determines the availability of route-specific information upstream.

We have also shown that the structures of Kanban, CONWIP and a Base Stock Policy do not enable this information transfer in this type of production system where route-specific information is crucial for improving the performance. Germs & Riezebos (2009) have found similar outcomes for CONWIP.

5.8 Managerial Implications

The most clear managerial implication which can be derived from this thesis is the fact that

trade-offs in terms of optimal performance will need to be made when choosing a structure and

configuration to use in practice. If the focus is on reducing TTT and STT in a MTO environment with divergent production line companies should focus on designing their system in such a way that route-specific information can be transferred to the point where orders are released.

Furthermore the configuration will determine the amount of work that can be processed at any given time and should therefore not be chosen too restricted if the service level is important for the given company. As reducing the WIP will also decrease utilization and throughput.

6. Conclusion

In this chapter we draw conclusions from the results as described in chapter 5. This is done by first answering the research questions. Secondly the general conclusions are presented and finally suggestions for future research are made.

The problem addressed in this thesis was the lack of insight into why certain pull production control systems did not show workload balancing capability. Previous research has suggested that the transfer of route-specific information is crucial for a system to balance workload effectively in a MTO divergent production line. We aimed to discover if and how the different structures of the production systems influenced the information transfer and thus the workload balancing mechanism. Our approach has been to use discrete event simulation to simulate a range of structures and configurations of route-specific as well as non-route-specific systems.

Furthermore, we only allowed for routing-variability in order to eliminate any other cause for the

effects found. As we focused on the influence of the structure we have also simulated adaptations

of established pull systems to be able to validate our findings and give suggestions for future

research.

The general conclusion that can be drawn from this research is that route-specific information is crucial in order to balance workload. Based on the results of the experiments performed in this research we have found that in order to balance workload in a route specific MTO environment it is crucial to ensure that information from the downstream stations is communicated to upstream stations. The transfer of route-specific information depends on the control loop placement and type of control loop. Which in turn is determined by the structure of the pull production control system.

Hence the structure has a big impact on the workload balancing capability of a system. Specifically what has been shown in this thesis is that the exchange of route-specific information between control loops is crucial to reduce the TTT performance. We have identified that some structural characteristics, such as not having overlapping control loops, make it impossible to exchange this information. More importantly however, this information should reach the loop which controls the order release and prioritization since inadequate prioritization of orders may cause blocking at downstream stations and thus annihilate the effects of workload balancing, which we found to be the case for POLCA. Additionally, from the experiments with an adapted structure of POLCA we can conclude the importance of trying to maximize the number of places in the system where workload can be balanced. Balancing the workload at later stages in a divergent production line enables one to indeed enlarge the number of places where workload can be balanced, because at each stage the amount of workstation doubles.

Furthermore we have shown that the optimal configuration to reduce TTT is not necessarily

the best configuration under all circumstances. In the thesis the optimal card count was determined

for the tested structures and this particular production system in order to test the structures

performance. However it may be the case that by choosing other performance measurements, such as due date adherence, other configurations could be preferred. In cases where the service level is important a less strict WIP limitation with a sub-optimal TTT performance may still be preferred.

Based on our findings and conclusions we suggest that for selecting pull system structures in divergent production lines, companies should move away from limiting themselves to basic and existing pull production systems. They should move towards adopting adapted structures which enable route-specific information to travel within complete routes and ensures there is enough opportunities in the system to balance the workload.

6.1 Future Research

Looking at the conclusions that can be drawn from this research there are several

opportunities for future research. We have seen that using only a WIP limit such as in a CONWIP

or Kanban system the system is unable to balance the workload in a route-specific unit-based

production system with a divergent topology. Orders in such pull system are released based on a

first come first serve bases. Using alternative dispatching strategies may improve the performance

of these systems as we have shown that route-specific information should be available at the point

where orders are released. Possible ways of altering the prioritization of orders is using alternative

cards such as load based cards which makes it easier to transfer capacity information to the stations

upstream. Other ways of doing so is by using different prioritization rules, which do not release

eon a first come first serve basis. This could be done by implementing a MRP system or other

production planning systems in combination with pull production control. Such hybrid systems

have been getting increasing attention of researchers since it has become clear that pull and push

are not competitors but rather good collaborators.

Furthermore it needs to be mentioned that the selection of the best configuration for each structure is a complex and highly dynamic problem. Deciding which configuration to choose or whether it should be adaptive, and how this should be implemented still requires a lot of research.

This is mainly due to the large differences in shop floor characteristics. Especially when variability

is added to the mix it becomes hard to select one configuration that always performs better than

any other. This is reflected by our experimental outcomes concerning the configuration where high

WIP limitations yielded better results in terms of TTT yet do not allow for a high throughput. If a

system is able to identify which performance indicator is preferable given a certain order or state

of the system, it will be better able to control the WIP limit and thus the system performance.

References

1. Akturk, M.S. & Erhun, F., 1999. "An overview of design and operational issues of kanban systems." International Journal of Production Research, 37(17), pp.3859–3881.

2. SimPy, 2016. Welcome to SimPy. Available at:

https://simpy.readthedocs.io/en/latest/index.html.

3. Banerjee, S.K., 1996. 'Methodology for integrated manufacturing planning and control systems design.' In A. Artiba & S. E. Elmaghraby, eds. The Planning and Scheduling of Production Systems. Boston, MA: Springer US, pp. 54–88.

4. Bonney, M.C. et al., 1999. "Are push and pull systems really so different?" International Journal of Production Economics, 59(1), pp.53–64.

5. Dahl, O.-J. & Nygaard, K., 1966. "SIMULA: an ALGOL-based simulation language."

Communications of the ACM, 9(9), pp.671–678. Available at:

http://dl.acm.org/citation.cfm?id=365813.365819 [Accessed March 8, 2016].

6. Eduard Babulak, 2010. Discrete Event Simulation: State of the Art., INTECH Open Access Publisher.

7. Python Software Foundation, 2016. Python. Available at: http://www.python.org.

8. Gaury, E.G. a, Kleijnen, J.P.C. & Pierreval, H., 2001. "A methodology to customize pull control systems." Journal of the Operational Research Society, 52(7), pp.789–799.

9. Gaury, E.G.A., Pierreval, H. & Kleijnen, J.P.C., 2000. "Evolutionary approach to select a pull

system among Kanban, Conwip and Hybrid." Journal of Intelligent Manufacturing, 11(2),

pp.157–167.

10. Germs, R. & Riezebos, J., 2009. "Workload balancing capability of pull systems in MTO production." International Journal of Production Research, 48(8), pp.2345–2360. Available at: http://www.tandfonline.com/doi/abs/10.1080/00207540902814314 [Accessed February 8, 2016].

11. González-R, P.L. & Framinan, J.M., 2009. "The pull evolution: from Kanban to customised token-based systems." Production Planning & Control, 20(3), pp.276–287. Available at:

http://www.tandfonline.com/doi/abs/10.1080/09537280902875393.

12. González-R, P.L., Framinan, J.M. & Pierreval, H., 2012. "Token-based pull production control systems: An introductory overview." Journal of Intelligent Manufacturing, 23(1), pp.5–22.

13. Henrich, P. et al., 2003. "The effect of information on workload control." Design, pp.1–10.

14. Hopp, W.J. & Spearman, M.L., 2004. "To Pull or Not to Pull: What Is the Question?"

Manufacturing & Service Operations Management, 6(2), pp.133–148. Available at:

http://pubsonline.informs.org/doi/abs/10.1287/msom.1030.0028 [Accessed February 8, 2016].

15. Karaesmen, F. & Dallery, Y., 2000. "Performance comparison of pull type control mechanisms for multi-stage manufacturing." International Journal of Production Economics, 68(1), pp.59–71.

16. Krishnamurthy, A. & Suri, R., 2009. "Planning and implementing POLCA: a card-based control system for high variety or custom engineered products." Production Planning &

Control, 20(7), pp.596–610. Available at:

http://www.tandfonline.com/doi/abs/10.1080/09537280903034297.

17. Land, M. & Gaalman, G., 1996. "Workload control concepts in job shops a critical assessment." International Journal of Production Economics, 46-47, pp.535–548.

18. Liberopoulos, G. & Dallery, Y., 2000. "A unified framework for pull control mechanisms in multi‐stage manufacturing systems." Annals of Operations Research, 93, pp.325–355.

Available at: http://link.springer.com/article/10.1023/A:1018980024795.

19. Maccarthy, B.L. & Fernandes, F.C.F., 2000. "A multi-dimensional classification of production systems for the design and selection of production planning and control systems." Production Planning & Control, 11(5), pp.481–496.

20. Nucci, F. & Grieco, A., 2012. "A simulation-based lean production approache at a low- volume parts manufacturer with part combining." Proceedings of the 2012 Winter Simulation Conference (WSC), pp.1–11.

21. Ōno, T., 1988. Toyota production system: beyond large-scale production, Available at:

https://books.google.com/books?hl=nl&lr=&id=7_-67SshOy8C&pgis=1 [Accessed February 22, 2016].

22. Robinson, S., 2004. Simulation: the practice of model development and use, Available at:

http://www.amazon.com/dp/0470847727.

23. Sharma, S. & Agrawal, N., 2009. "Selection of a pull production control policy under

different demand situations for a manufacturing system by AHP-algorithm." Computers and Operations Research, 36(5), pp.1622–1632.

24. Welch, P.D., 1983. 'The Statistical Analysis of Simulation Results.' In Stephen S. Lavenberg,

ed. Computer Performance Modeling Handbook. New York: Academic Pess, pp. 267 – 323.

25. Zäpfel, G. & Missbauer, H., 1993. "Production Planning and Control (PPC) systems

including load-oriented order release - Problems and research perspectives." International Journal of Production Economics, 30-31(C), pp.107–122.

26. Ziengs, N., Riezebos, J. & Germs, R., 2012. "Placement of effective work-in-progress limits

in route-specific unit-based pull systems." International Journal of Production Research,

50(16), pp.4358–4371.