Professional Documents
Culture Documents
} { states all
) state ( ) (
i
S
i Prob i P P (1)
} { states all
) state ( ) (
i
S
i Prob i WIP WIP (2)
Usually three steps are taken to calculate system throughput and WIP based on the above-mentioned state analysis
methods:
Step 1: Calculate the specific throughput P(i) and WIP(i) at each state;
Step 2: Calculate the probabilities of each state Prob(state i);
Step 3: Calculate the product of the above and sum the products over all states using Eq.(1) and Eq.(2).
One of the major difficulties of the above method is that the number of system states increases exponentially with the
increase of the number of machines and buffers in the manufacturing system, which results in difficulty of solving extremely
large number of linear or nonlinear equations. From the viewpoint of computational complexity, the exact analytical solutions
of system performance, such as the system throughput and average WIP levels, are only available in the case of two-machine-
one-buffer serial unit. Gershwin [8], and Gershwin and Burman [12] provided elaborate details on using approximate
decomposition algorithms to estimate the system throughput and WIP for long transfer line (as depicted in Figure 3) and
generic tree-structured Assembly/Disassembly networks (as shown in Figure 4). In Figure 3 and Figure 4, we use squares to
represent machines and circles to represent buffers in the system. A machine that has two or more upstream buffers in front of
the machine performs assembly operations and is called an assembly machine. When it does an assembly operation, it
consumes one part from each of their front buffers. Similarly, a machine that has two or more downstream buffers is a
disassembly machine. When it finishes an operation, it generates one part to each of its downstream buffer.
Figure 3: A transfer line with k machines and (k-1) buffers
M
i
M
i+1
B
i
M
1
B
1
B
k-1
M
k
M
i
B(j,i)
M
j
B(n,i)
M
n
B(i,m)
M
m
B(i,q)
M
q
B(j,r)
Figure 4: Generic tree-structured Assembly/Disassembly network
As pointed out in the literature review, the machine processing time may take different assumptions and models such as
exponential processing time models, continuous fluid models, and deterministic models. In this paper we simply assume that all
machines have a constant processing time, and exponential failure or repair processes, and the material flows are continuous
fluid; however, other models can be used without impacting the applicability of the parameter coupling technique. We assume
that any transfer lines depicted in Figure 3, or tree-structured Assembly/Disassembly networks depicted in Figure 4 are
solvable using the decomposition methods in Gershwin [8], and Gershwin and Burman [12]. We have improved the algorithms
in Gershwin [8], and Gershwin and Burman [12], and implemented a software package named PAMS (Performance Analysis
of Manufacturing Systems) based on Yang et al. [10] and Liu et al. [11] to analyze the system throughput and work-in-process
of any open structure manufacturing systems. PAMS is available for academic users to download and test run from our
University of Michigan website: http://www.umich.edu/~samyang/PAMS.
5. Description of the parameter coupling technique for CLMS
The decomposition methods in Gershwin [8] and Gershwin and Burman [12] provide fast and accurate performance
estimation on transfer lines and tree-structured assembly/disassembly networks. However, it is not directly applicable to
CLMSs. The major reason is that in open systems machine/buffer parameters are independent, and machines/buffers can be
ordered in sequence for the purpose of decomposition. In CLMSs, the re-entrant machines/buffers form a loop and make it
impossible to decompose into a series of smaller and independent equivalent systems, because there is no way to order all
machines and buffers in CLMSs based on an upstream-downstream relationship, which is an prerequisite for the decomposition
method.
Let us consider the CLMS depicted in Figure 1. Since loading station at position <a> combines a raw part and an empty
pallet into a loaded pallet, it is natural to treat the loading station as an assembly machine. The condition of loading raw
material into the system has two prerequisites:
1. The raw material feeding machine feeds a raw material onto the loading machine, with a constant speed of
f
;
2. The conveyor system upstream of the loading machine must send an empty pallet to the loading machine at the same
time.
The second prerequisite can be approximated using a virtual empty pallet feeding machine VM
1
in front of the loading
assembly machine. VM
1
is assumed to have a constant pallet feeding speed of
K
p
/N (3)
based on the following facts: (1) the pallet indexing speed
p
determines the maximum speed it takes an empty pallet to the
loading machine (as long as it does not become blocked or broken down); and (2) the availability of an empty pallet in front of
the loading machine is determined by the density the pallets are distributed in the whole conveyor system, which can be
measured by the probability of K /N.
Similarly, the unloading station at position <b> can be modeled as another assembly machine. The conditions of unloading
a finished part out of the system need two prerequisites:
1. The upstream conveyor system sends a pallet with a finished part to the unloading machine;
2. There is a space immediately available after the unloading machine to take the empty pallet after releasing the finished
part to shipment.
The second prerequisite can be approximated using a virtual empty space feeding machine, VM
2
, in front of the unloading
assembly machine. VM
2
is assumed to have an empty space feeding speed of
max(0, B
4
*(N-K) *
p
/N) (4)
based on the following facts: (1) the conveyor indexing speed
p
determines the maximum speed it releases an empty space to
the unloading machine; (2) however, the availability of an empty space after the unloading machine is determined by the
density an empty space is distributed in the whole conveyor system, which can be measured by the probability of (N-K) /N; (3)
the availability of an empty space after the unloading machine is also impacted by the buffer capacity B
4
after the unloading
machine, and any empty space in B
4
will help hold the unloaded empty pallet; and (4) the empty space feeding speed of VM
2
cannot be negative.
Based on the analysis above, we can use an open manufacturing system with two assembly machines (the loading machine
and the unloading machine, respectively) and two new introduced virtual feeding machines (one for feeding empty pallets, one
for feeding empty space, respectively) to approximate the pallet moving and processing behaviors in the original system. The
new equivalent open system of the CLMS is depicted in Figure 5.
Please note that in a regular open system, the parameters of each element are completely independent and irrelevant from
that of other elements. However, in the equivalent open system of the CLMS depicted in Figure 5, the parameters of those two
virtual feeding machines VM
1
and VM
2
are actually determined by the pallet number K, the total conveyor capacity N (the total
conveyor capacity N should include the machine workspace capacity in the loop), and conveyor index speed
p
. Such kind of
parameter coupling is called machine parameter coupling because they are coupled parameters on those new introduced virtual
machines.
On the other hand, the loading position <a> in the CLMS not only directly determines the available buffer size for holding
the full pallets, but also indirectly impacts the available buffer size for holding the finished pallets as well as the space for
holding empty pallets. So the buffer space B1, B2, B3 and B4 in Figure 5 are coupled by the total space capacity N:
C(B
1
)+C(B
2
)+C(B
3
)+C(B
4
)+#machines=N. When we shift the loading position and increase the buffer size of B1, the buffer
size at B4 will be reduced accordingly. We call this coupling as the buffer capacity coupling in the CLMS.
Raw Part
Feed f
Load
Station
B1 M1 M2
M3
B2
Empty Pallet
Feed p.K/N
VM1
B3
UnLoad
Station
Finished
Parts
B4
Empty Space
Feeding
VM2
Figure 5: The equivalent open system of the original CLMS
By introducing the two different concepts of parameter coupling (the virtual machine parameter coupling, and buffer
capacity coupling) to reflect the machine and space resource sharing mechanism in CLMSs, we have extended the traditional
open line analysis methods to closed loop manufacturing system analysis.
6. Numeric cases and discussion
Now consider the CLMS depicted in Figure 1. All machines M
i
in the system are assumed to have a constant processing
time
i
=2 Parts / min, i= 1, 2, 3. They also have negative exponential distributions of failures and repairs, with Mean Time
Between Failures MTBF=120 minutes, and Mean Time To Repair MTTR=10 minutes, respectively. The material flows are
treated as continuous fluid. The conveyor system has a constant pallet indexing speed
p
. In the first group of experimentations,
we assume
p
=2 Parts/min; and in the second group of experimentations, we change
p
= 3 Parts /min and compare the system
performance changes. Overall, the system has the following parameters:
K: Actual total number of pallets in the system: It is a variable in the model;
Conveyor lengths: B
1
=20; B
2
=20; B
3
=20; B
4
=20;
N: Maximum pallet capacity N = N: B
1
+B
2
+B
3
+B
4
+#machines=85
p
: Pallet indexing speed: We set up two set of pallet indexing speed: (1)
p
= 2 Parts/min; (2)
p
= 3 Parts/min;
f
: Raw part feeding speed into the system:
f
=2 Parts/min;
i
: Processing speed of machine M
i
, i=1, 2, 3:
i
=2 Parts /min; MTBF=120 minutes; MTTR=10 minutes.
Based on the parameter coupling technique we have discussed before, we have built the corresponding PAMS analytic
model, and the WITNESS simulation model, as shown in Figure 6, and Figure 7, respectively (this PAMS model and the
simulation model are available for download at http://www.umich.edu/~samyang/papers/CLMS). Table 1 compares the results
of this parameter coupling technique with that from simulation models (WITNESS is the simulation software from Lanner
Group Inc. which we used in this paper to build the simulation models) on the CLMS in Figure 1. The horizontal axis
represents the number of pallets in the system while the vertical axis represents the corresponding system throughput (JPH
means Job per Hour). Generally speaking, the analytic results match the WITNESS simulation results well, with a 4%~8%
errors in maximum between those two models. We noticed the analytic results consistently provide lower throughput values
than simulation results.
Figure 6: the PAMS model Figure 7: the WITNESS model
Table 1 depicts the specific running results from both PAMS model, and the WITNESS simulation model, with two pallet
indexing speed as (1)
p
= 2 Parts/min, and (2)
p
= 3 Parts/min, respectively. The assumed simulation time is 10 days, with a
warm-up time of two hours.
Table 1: Comparison between PAMS model and WITNESS model
# of Pallets pallet index speed s=2 pallet index speed s=3
SIMU (s=2) PAMS (s=2) % Diff SIMU(s=3) PAMS (s=3) % Diff
1 1.41 1.412 0.14% 2.04 2.12 3.77%
5 6.97 7.058 1.25% 10.11 10.58 4.44%
10 13.83 14.082 1.79% 19.85 20.98 5.39%
15 20.55 20.98 2.05% 29.17 30.91 5.63%
20 27.11 27.662 2.00% 38.09 40.20 5.25%
25 33.39 34.076 2.01% 46.28 48.81 5.18%
30 39.64 40.196 1.38% 53.87 56.75 5.07%
35 45.63 46.013 0.83% 61.06 64.04 4.65%
40 51.29 51.53 0.47% 66.60 70.68 5.77%
45 56.53 56.752 0.39% 72.28 76.53 5.55%
50 61.08 61.685 0.98% 76.67 81.11 5.47%
55 65.62 66.329 1.07% 79.95 82.87 3.52%
60 69.87 70.676 1.14% 81.42 82.89 1.77%
65 73.56 74.688 1.51% 82.00 82.89 1.07%
70 76.81 78.265 1.86% 82.03 82.89 1.04%
75 78.73 81.11 2.93% 82.48 82.89 0.49%
80 80.06 82.65 3.13% 81.45 82.89 1.74%
81 79.70 82.78 3.72% 81.26 82.89 1.97%
82 79.73 82.855 3.77% 81.04 82.89 2.23%
83 80.27 82.88 3.15% 80.72 82.89 2.62%
84 80.01 82.89 3.47% 80.27 82.89 3.16%
85 0.00 0.00 0.00% 0.00 0.00 0.00%
90 0.00 0.00 0.00% 0.00 0.00 0.00%
Max Error 3.77% 5.77%
Figure 8 depicts the pallet number vs. system throughput curve for those two groups of experimentations.
Figure 8: Comparison between PAMS model and WITNESS model
From Table 1 and Figure 8, we see that the equivalent PAMS model results match the simulation results very well for both
pallet index speed = 2 pallets/min, and pallet index speed = 3 pallets/min. The maximum errors are 3.77% and 5.77%,
respectively. However, each running of the PAMS analytic model takes only 0.0012 seconds in our computer, but the
corresponding simulation model takes 10 minutes to output a steady state result.
We have made the following observations for CLMS pallet number and conveyor design based on our experimentations:
The traditional wisdom of buffer/pallet deployment in open manufacturing systems based on Littles law does not
hold. In open manufacturing systems, we know that the more buffers/pallets in the system, the higher system
throughput we probably obtain (or at least the system throughput will not decrease with the increase of pallets in
the system), which is a direct conclusion from Littles law. However, in CLMSs, the system throughput will reach
a saturation point <SP1> or <SP2> as we increase the pallet numbers gradually. Once the system pallet number
crosses the first saturation point <SP1> or <SP2>, the system throughput will keep quite flat for a period of time.
0 10 20 30 40 50 60 70 80 90
0
10
20
30
40
50
60
70
80
90
Pallet Number
T
h
r
o
u
g
h
p
u
t
(
J
P
H
)
SIM (s=2)
PAMS (s=2)
SIM (s=3)
PAMS (s=3)
<SP1>: Saturation point 1
<SP2>: Saturation point 2
<DP>: Dropping point
<SP2>
<SP1>
<DP>
If we keep increasing the pallet numbers in the system, we will reach another critical dropping point <DP>, and
the system throughput will drop dramatically after this point;
At the very beginning of pallet number increase, the CLMS behaves more like open manufacturing systems,
because the coupling of empty pallets, loaded pallets and finished pallets is very weak at this stage, and the
system throughput increases almost linearly with the increase of pallet numbers at the beginning; however, as the
pallet number increases in the system, the coupling of those three types of pallets, and their sharing of the
common conveyor space will increase and the gain of system throughput will gradually become flat. At the same
time, the increase of pallets will also increase the traffic jam and congestions for the material handling system;
The first saturation point <SP1> or <SP2> is impacted by the pallet indexing speed. Generally speaking, the faster
the pallets move along the conveyor system, the steeper of the throughput versus pallet number curve (figure 8),
and fewer pallets are required to reach a maximum throughput value. On the other hand, the second critical
dropping point <DP> is impacted by the total pallet capacity in the system. The recommended pallet number in
the system is nearby the first saturation point <SP1> or <SP2>.
7. Conclusions
This paper first introduces methods of analyzing open manufacturing systems based on Gershwin and Burmans work. We
propose a parameter coupling method to mathematically model and analyze closed loop manufacturing systems based on the
enhancement of traditional open form production line analysis techniques. Comparisons between simulation models and this
new method show the new method is fast and accurate. On the other hand, our current analytical model on CLMS analysis is
still greatly simplified compared to the complexity of actual CLMS. More technique details need to be addressed in the future
such as considering parallel stations, material handling system (MHS) reliability, etc. in the CLMS.
The parameter coupling technique for CLMSs is more flexible for modeling and analyzing various closed loop
manufacturing systems than the traditional queuing networks approach because the parameter coupling technique presented in
this paper is independent from the specific open manufacturing models (exponential processing time models, continuous fluid
models, and deterministic models, etc.). On the other hand, the open and closed-form queuing theory usually has to assume
Poisson process at a specific time-interval and negative exponential distribution for processing time to make it mathematically
tractable.
Currently the PAMS application can only analyze open manufacturing systems. It does not contain any closed loop
modules yet. To solve the throughput of CLMSs, we have to add virtual machines and coupled buffers manually in the
corresponding open models and maintain the parameter coupling and data consistency manually. We plan to enhance the
current PAMS capabilities by introducing some new closed loop modules in PAMS and implement the parameter coupling
technique into PAMS to automate the analysis of CLMSs and improve the calculation efficiency.
8. Acknowledgement
We would like to express our sincere thanks to Professor Yoram Koren and the Engineering Research Center for
Reconfigurable Manufacturing Systems at the University of Michigan for the support of developing the PAMS software. We
also gratefully acknowledge the support from the General Motors Collaborative Research Laboratory in Advanced Vehicle
Manufacturing at the University of Michigan.
REFERENCES
[1] McGuire PM. Conveyors: Application, Selection, and Integration, 1
st
ed., CRC Press, 2009.
[2] Li J, Blumenfeld DE, Huang N, Alden JM. Throughput analysis in production systems: recent advances and future topics.
Int J Prod Res 2009; 47(14): 38233851.
[3] Banks J, Carson JS, Nelson BL, Nicol DM. Discrete-Event System Simulation, 5
th
ed., Prentice Hall, 2009.
[4] Law A. Simulation Modeling and Analysis, 4th edition, McGraw-Hill Publishing Co., 2006.
[5] Smith JS. Survey on the use of simulation for manufacturing system design and operation. Journal of Manufacturing
Systems 2003; 22(2) 157-171.
[6] Taylor SJE, Behli L, Wang X, Turner SJ, Ladbrook J. Investigating distributed simulation at the Ford Motor Company.
Ninth IEEE International Symposium on Distributed Simulation and Real-Time Applications 2005; 139-147.
[7] Dallery Y, Gershwin SB. Manufacturing flow line systems: a review of models and analytical results. Queueing Systems
Theory and Applications, Special Issue on Queueing Models of Manufacturing Systems 1992; 12:3-94.
[8] Gershwin SB. Manufacturing Systems Engineering. Englewood Cliffs: Prentice Hall, 1994.
[9] Koren Y. The Global Manufacturing Revolution: Product-Process-Business Integration and Reconfigurable Systems. 1
st
ed., Wiley, 2010; Chapter 10:253-280.
[10] Yang S, Wu C, Hu SJ. Modeling and analysis of multi-stage transfer lines with unreliable machines and finite buffers.
Annals of Operations Research 2000; 93:405421.
[11] Liu J, Yang S, Wu C, Hu SJ. Multi-state throughput analysis of a two-stage manufacturing system with parallel unreliable
machines and a finite buffer. European Journal of Operational Research 2012; 219(2): 296304.
[12] Gershwin SB, Burman MH. A decomposition method for analyzing inhomogeneous assembly/disassembly systems.
Annals of Operations Research 2000; 93:91-116.
[13] Li S, Wang H, Hu SJ, Lin YT, Abell JA. Automatic generation of assembly system configuration with equipment
selection for automotive battery manufacturing. Journal of Manufacturing Systems 2011; 30(4): 188-195.
[14] Um I, Cheon H, Lee H. The simulation design and analysis of a flexible manufacturing system with automated guided
vehicle system. Journal of Manufacturing Systems 2009; 28(4): 115-122.
[15] Koren Y, Shpitalni M. Design of reconfigurable manufacturing systems. Journal of Manufacturing Systems 2010; 29(4)
130-141.
[16] Huang N. Mean station reliabilities cause throughput overestimates in production system design. Journal of
Manufacturing Systems 2012: 31(2): 184-194.
[17] Huang SH, Dismukes JP, Shi J, Su Q, Wang G, Razzak MA, Robinson DE. Manufacturing system modeling for
productivity improvement. Journal of Manufacturing Systems 2002: 21(4) 249-259.
[18] Bastani A. Analytical solution of closed-loop conveyor systems with discrete and deterministic material flow. European
Journal of Operational Research 1988; 35(2): 187-192.
[19] Narahari Y, Khan LM. Modeling reentrant manufacturing systems with inspection stations. Journal of Manufacturing
Systems 1996; 15(6) 367-378.
[20] Bouhchouch A, Frein Y, Dallery Y. Analysis of a closed-loop manufacturing system with finite buffers. Applied
Stochastic Models and Data Analysis Special Issue: Special Issue on Manufacturing Systems 1993; 9(2): 111-125.
[21] Frein Y, Commault C, Dallery Y. Modeling and analysis of closed-loop production lines with unreliable machines and
finite buffers. IIE Transactions 1996; 28:545554.
[22] Gershwin SB, Werner LM. An approximate analytical method for evaluating the performance of closed loop flow
systems with unreliable machines and Finite buffers. International Journal of Production Research 2007; 45(14): 3085-
3111.
[23] Li N, Zheng L, Feng W. Performance analysis of two-loop closed production systems. Computers and Operations
Research 2009; 36(1): 119-134.
[24] Nazzal D, Jimenez J, Carlo H, Johnson A, Lasrado V. An analytical model for conveyor-based material handling system
with crossovers in semiconductor wafer fabs. IEEE Trans. On Semiconductor Manufacturing 2010; 23(3): 468-476.