Optimization of A Battery Manufacturing Line Using Computer Simulation
Optimization of A Battery Manufacturing Line Using Computer Simulation
Optimization of A Battery Manufacturing Line Using Computer Simulation
Simulation
by
Gregory J. Mazenko
Submitted to the Departments of Electrical Engineering and Computer Science and the
Sloan School of Management in Partial Fulfillment of the Requirements for the Degrees of
Signature of Author
() (.SlS&i School ofPanagement
Department of Electrical Engineering and Computer Science
Z .May 12, 1995
Certified by
David E. Hardt
Professor of Mechanical Engineering
Certified by
Roy E. Welsch
Profesjr of Statistips and Management Sciences
Accepted by I I~ ,AA.v.M II iu l
• .- -,
Accepted by
\ I Jeffrey A.
oan Master's andBarks, Associate
Bachelor's Dean
Programs
Barker Eng
MASSACHUSETTS INSTITUTE
OF TFC,,HNOLOGY
,JUN 2 0 1995
Optimization of a Battery Manufacturing Line Using Computer Simulation
by
Gregory J. Mazenko
Submitted to the Departments of Electrical Engineering and Computer Science and the
Sloan School of Management in Partial Fulfillment of the Requirements for the Degrees of
Abstract
The goal of this thesis was to examine the effects that computer simulation could
have on a large capacity expansion project in a major global manufacturing firm. The
project studied was in its design and development phase, so any insight gained from the
simulation work could assist the project team's decision-making process before actual
production operations began.
This thesis examines the relevant issues surrounding the use of the computer
simulations, the various statistical techniques used in their development, and the insights
gained from their use. It describes the specific effects that simulation had on the project
team's problem-solving process. It illustrates how computer simulation was used in
conjunction with the theory of constraints to develop possible improvement strategies.
Finally, the thesis examines how simulation was used to help make recommendations for
future expansion projects.
Thesis Advisors:
Acknowledgment
I wish to acknowledge Polaroid Corporation's IBAM team and the Leaders for
Manufacturing Program for their support of this work.
Table of Contents
CHAPTER 1 - INTRODUCTION............................--------- 11
1.1 COMPU ER SIMULATION .................................................................................................................... 11
1.2 CORPORATE STRUCTURE ....................................................................... 11
1.3 BATTERY DIVISION (R-5) ... ............. .................................... ..................................................... 12
1.4 PHIL SOPHY OF IBAM ...................................................................................................................... 15
1.5 M YEXPECTED CONTRIBU I N ......................................................................................................... 16
1.6 THESIS STRUCTURE........................................................................................................................... 17
CHAPTER 2 - BATTERIES ............. 18
2.1 O VERVIEW ........................................................................................................................................ 18
2.2 BATTERY STRUCTURE................................................................. .................................................. 18
2.3 OVERVIEW OF PRODUCTION ON THE RBAM WEB SID......................................... ............... 19
2.3.1 ImportantIssues on the RBAM Web Side............................................................................ 21
2.4 PRODUCTION ON THE IBAM....... ................................................................................................. 23
CHAPTER 3 - DATA-BASED CHARACTERIZATION OF EXISTING PROCESS ...................... 26
3.1 CHAPTER OVERVIEW ................ .............................................................................................. 26
3.2 IMPORTANCE OF DATA ANALYSIS ...................................................................................................... 26
3.3 PROCEDURES CURRENTLY INEFFECT FOR DATA COLLECTION.......................................................... 27
3.3.1 ManualData Collection........................................................................................................... 27
3.3.2 Automatic DataCollection....................................................................................................... 28
3.4 THE DATA ITSELF ..................................................................... 29
3.4.1 PeriodStudied.......................................................................................................................... 31
3.4.2 Converting to Machine-UsageTime ......................................................................................... 31
3.5 STATISTICAL TECHNIQUES USED ............. ..................................................................................... 33
3.5.1 Analyss of Variance (ANOVA)................................................................................................. 33
3.5.2 Chi-Square Goodess-of-FitTest............................................................................................. 42
3.5.3 ExponentialRegression............................................................................................................. 48
3.5.4 Sm oothing ................................................................................................................................ 50
3.6 TESTIN THE VALIDITY OF THE EMPIRICAL D RIB N ................................................................ 51
CHAPTER 4 - DEVELOPMENT OF THE COMPUTER MODELS .. 53
4.1 INIAL MODELIN APPROACH .......................................................................................................... 53
4.2 DISCOVERY OF THE DRY SLURRY YIELD LOSS ISSUE.......................................................................... 55
4.3 ESTABSHMENT OF NEED TO REDUCE EXPOSURE TIME ...................................................................... 58
4.3.1 Refinement ofAsmptions....................................................................................................... 58
4.3.2 InitialId as.............................................................................................................................. 59
CHAPTER S - THE IBAM TEAM'S SOLUTION OF THE DRY SLURRY PROBLEM ........... 62
5.1 CHAFFER OVERVIEW ......................................................................................................................... 62
5.2 IN AL IN FROM THE IBAM TEAM........................................................................................ 62
5.3 REFINEMENT OF ASSUMPTIONS AND SCENARIOS ........................... .................. 63
5.3.1 Best Case/Worst Case Approach......................................................................................... 64
5.4 THE IBAM TEAM's PROBLEM-SOLVING METHODO Y ...................................................................... 72
5.5 YIELD LOSS vs. RUN TIME CURVES ................................................... 79
CHAPTER 6 - APPLICATION OF THEORY OF CONSTRAINTS ............................. 81
6.1 THEORY OF CONSTRAINTS............................ ................................. .............................................. 81
6.2 EFFECTS OF ADDING A SECOND DOUBLE CAMBOT INCOMPOSrrE LOOP............................................ 82
6.3 THE ROLE OF COMPUTER SIMULATION ............................................................................................... 87
6.4 WHAT W AS ACTU ALLY DECIDED....................................................................................................... 88
CHAPTER 7 - EFFECTS OF DECOUPLING THE CARRIER WEB ............... .................91
7.1 HA E OVE VIEW ......................................................................................................................... 91
7.2 DECOU THE CARRIER WEB.................................................................................................. 91
7.3 COMM EN ....................................................................................................................................... 94
CHAPTER 8 - FINAL CONCLUSIONS ........... . ...................................... 95
8.1 CHA ER OVER VIEW ......................................................................................................................... 95
8.2 CONCLUSIONS................................................................................................................................... 95
REFE RENCES .... 97
List of Figures
List Of Tables
TABLE 15 : VE FI ATIO
N TEST RESULTS ................................................................................................ 52
TABLE 16: DETAILED FAILURE INFORMATION .......................................................................................... 66
TABLE 17: C ER IA ScORES .................................................................................................................. 74
TABLE 18: CATEOORIZED SCORES ........................................................................................................... 75
TABLE 19: CONCE EVALUATIONS ......................................................................................................... 76
TABLE 20: CONCE rn' NUM ERICAL SCORES .......................................... .............................................. 77
TABLE 21: CUMULATIVE SCORES....... ........................................................................................ 77
Chapter 1- Introduction
Polaroid Corporation is divided into three major business units entitled High
Resolution Imaging, Core Photography, and Electronic Imaging Systems. Battery
Division, which is located at the R-5 plant in Waltham, Massachusetts is part of the
Integral Film Assembly Division in the Core Photography Business Unit. A tree diagram
of the pertinent parts of the corporation follows:
Peel Apart Integral
Cameras Film Film IMD
Though Battery Division (R-5) is shown on the same tree level as Integral Film Assembly,
all of the batteries which R-5 produces are subsequently used in the Integral Film
Assembly operation at either the Waltham site or at the Enschede site in the Netherlands.
R-5 is a relatively small plant, with roughly 250 permanent employees. They are
the only battery manufacturing plant inthe corporation, so they must conduct extremely
high-volume operations year-round. A simple block diagram of operations at R-5 follows:
Figure 2: R-5 Process Flow
WAM Web assembly machine. The second pre-assembly machine. The WAM
produces rolls of composite material that form the three inner layers of the battery
between the cathode top plate and the anode bottom plate.
CHEM MIX In the chem mix area, the slurry which is the electrolyte for the battery is
produced and stored.
RBAM Rotary Battery Assembly Machine. The high-speed assembly line that
takes materials produced in pre-assembly and chem mix and assembles them into finished
batteries. The RBAM is so named because of its extensive use of rotary technology. It is
called a machine, but is actually a manufacturing line made up of a series of fully-coupled
machines. On the web side of the RBAM, batteries are constructed on a paper carrier
web. The card side of the RBAM is a finishing operation in which the batteries are
attached to thin cardboard squares to give the battery a more rigid structure. A series of
voltage checks are also performed.
JBAM Joshua Battery Assembly Machine. Performs the same function as the card
side of the RBAM. The card stock used on the JBAM is slightly smaller than that used on
the RBAM card side. Batteries finished on the JBAM are used in film packs for the
Captiva camera. Polaroid's internal project name for the Captiva is Joshua.
90 Day Buffer 90 days worth of finished goods are held at the plant. These seemingly
high levels of inventory are necessary due to the nature of some types of battery defects
that have thus far proven undetectable. These latent defects sometimes cause tiny shorts
in the batteries that lead to very gradual losses of charge. These defects tend to be fully
manifested after 90 days, so voltage tests are again performed after the 90-day incubation
period has passed.
Sort Final voltage checks and packaging of the batteries for shipment are
accomplished at Sort.
IBAM Incremental Battery Assembly Machine. The capital project that is the
subject of this thesis. The IBAM is so named because it will give the plant an incremental
increase inbattery manufacturing capacity. It will be a battery manufacturing line that
performs the same functions as the RBAM web side, but at much lower volumes. The
IBAM is scheduled to produce batteries by January 1, 1996. At the time this thesis was
written, different portions of the IBAM were invarying stages of design, development,
and construction.
Sales projections for instant film in the mid to late 1990s dictated the need for
substantially higher output from R-5 and the rest of the Integral Film Assembly Division.
R-5 is currently the only location within Polaroid that manufactures batteries. As is shown
inthe block diagram of operations in figure 2, the RBAM web side is the only machine at
R-5 which performs the assembly functions between pre-assembly and finishing operations
performed on the RBAM card side or on the JBAM. The RBAM web side is currently
stretched as far as it realistically can be. Short-term solutions like cutting maintenance and
test time are impractical, since they could have devastating long-term effects.
A major project of some sort was needed inorder to give R-5 the extra capacity
required. Whatever the nature of this project, it was determined that its global objectives
were the following:
My internship at R-5 began on June 1, 1994, about 1 1/2 years through the 3 year
project. At this point, different portions of the IBAM were in various stages of design,
prototyping, and construction. Following a brief indoctrination period, the IBAM
program manager and I decided that I could make a worthwhile contribution to the project
by taking on the responsibilities associated with the development of computer models of
the IBAM. The IBAM program had purchased a modeling package entitled Pro-Model
for Windows version 1.10. Up to that point, the use of the tool had been limited to a
single model of the IBAM's anode cell which had been developed by an outside simulation
development firm.
Although the programmers from the outside firm were skilled modelers with
extensive experience in modeling of manufacturing systems, it was difficult for them to
develop models with the flexibility necessary to test the wide variety of issues which
emerge in a rapidly changing project such as the IBAM project. It would be much more
beneficial to the project if there were a team member who could develop many different
flexible models in response to rapidly changing project decisions and constraints.
My charter was to develop models of the IBAM as a whole and of important
components inorder to examine the effects that different factors had on overall IBAM
performance. By doing this, we hoped to optimize the performance of the line, and to
perhaps uncover some hidden problems with the IBAM design.
1.6 Thesis Structure
2.1 Overview
This chapter is intended to give a brief overview of the batteries that are produced
at R-5. First, the battery and its components are discussed. Then, a very general
description of the manufacturing methods currently employed on the RBAM web side and
some of the surrounding issues is given. Finally, the methods used on the proposed IBAM
along with the major areas of expected improvement are examined.
OVERWIAP
CATHODE,
COMPOSI
SLURRY"
ANODE -
CARRIER
•AE*oThC
• •W w v
The cardstock on the very bottom and the overwrap on the very top of the battery are
added in the finishing stage of production. Recall from Chapter 1 that finishing operations
are accomplished on either the RBAM card side or on the JBAM. All of the layers in
between are assembled on the RBAM web side. The new IBAM will assemble the same
components as the RBAM web side.
The carrier, also referred to as the carrier web, is nothing more than brown paper.
The anode, or negative terminal, is aluminum based. It is the first layer applied to the
carrier. Next is the first of four layers of slurry. The slurry is the electrolyte for the
battery. Physically, it is a black paste which requires very special handling. One of the
reasons it requires special handling is because it loses its desirable properties as an
electrolyte if exposed to open air for an excessive amount of time. For this reason, if the
RBAM web side stops, and the layers of slurry inthe unfinished batteries have been
exposed to open air for 15 minutes or more, the batteries must be discarded. This will
prove to be a critical issue on the IBAM, and will be examined in detail inlater chapters.
The next six layers in the battery are alternating layers of composite material and
slurry. These interior layers give the battery greater voltage capacity. The final
component of concern is the cathode top plate. The cathode is also aluminum based, and
is the battery's positive terminal.
When fully assembled, the battery is less than one centimeter tall. This dimension
is very important, since the battery is an integral component of an instant film package. A
bulkier battery design would create a bulkier film package, which is undesirable for many
reasons including cost and customer satisfaction.
The RBAM web side is a series of machines which are fully coupled to one another
through the carrier web. All of the pre-assembly machines, the LAM, the WAM, and the
CLAM, produce rolls of their respective materials for use on the RBAM web side.
Components on the pre-assembly rolls are laid out four across. Each 1x4 section of the
roll is referred to as a row. The following diagram shows a pre-assembly roll. This
diagram could represent an anode, composite, or cathode roll, since all are configured
similarly:
Figure 4: Roll From Pre-Assembly
The backbone of the RBAM web side is the carrier web. The carrier web is a large
roll of brown paper. The carrier web roll is placed on an unwind stand at the beginning of
the line and is fed continuously through the series of machines which make up the RBAM
web side. As a section of the carrier web travels further down the line, components are
placed on it one Ix4 row at a time. The rotary equipment which places the components is
situated in-line with the carrier web, overhead. The slurry dispensing mechanisms are also
in-line and overhead.
The rolls of pre-assembled components are placed on unwind stands located off-
line. (In the context of discussions of the RBAM and IBAM, off-line means nothing more
than being physically situated somewhere other than directly over or under the carrier
web. Off-line does not imply that the machines are decoupled.) The pre-assembled rolls
are fed continuously from their unwind stands to the in-line rotary equipment which cuts
rows off of the rolls and places them row by row onto the carrier web.
The last in-line machine that the carrier web is fed through is a cut-off mechanism
simply called the final cut-off (FCO). The FCO waits until 18 rows pass by before making
its cut. The batteries are now configured in an 18x4 array as shown in the following
diagram:
4I JA BA7MY ST5RP
These 18x4 battery arrays are referred to as strips. The strips proceed from the FCO to a
metal vacuum table where they wait to be fed into the vacuum sealer. During steady state
operations, there will be three strips waiting to be vacuum sealed.
The vacuum sealer can handle a single strip of 72 batteries at a time. The sealing
operation marks the point where the slurry is no longer exposed to air. Following the
sealing operation, battery strips are loaded into tubs where they wait for further processing
on the finishing equipment.
The most pressing problem that the RBAM web side faces is low run time. This is
because all of the machines which make up the RBAM web side are completely coupled to
one another by virtue of the fact that all components are placed onto a continuous paper
carrier web. If any machine on the RBAM web side goes down, the web must stop and
production stops. The only buffering occurs on the vacuum sealer table, and in festoons on
the unwind stands. Recall that the vacuum sealer table holds three strips. Each strip only
takes about eight seconds to process, though, so this buffering capability is virtually
negligible. The festoons on the unwind stands are simply layered rollers through which
the roll of material is fed in a serpentine fashion. If the main roll runs out or stops feeding,
the festoon rollers can move closer together and provide a short period of uninterrupted
material flow upstream. A diagram illustrating the concept is shown in figure 6:
Main Roll
, Festoon Rollers
Feeds Upstream
Most of the festoons have only enough material to feed the upstream processes for a few
seconds. The real purpose of the festoons is to allow operators to splice in new rolls of
material at the ends of old rolls without shutting the whole RBAM web side down.
Despite these two minor exceptions, the RBAM web side is fully coupled and non-
buffered. This would not be a problem if all of the machines which comprise the RBAM
web side were very reliable, but this is not the case. In the period that will be described in
chapter 3 from which machine reliability data was taken, the RBAM web side was down
due to individual machine failures for 33% of the time during which it was scheduled to
run. Improvements in run time would dramatically reduce the cost of manufacturing each
battery.
In addition to run time problems, yield loss problems also afflict the RBAM web
side. The specific causes of yield loss are constantly evaluated, and possible solutions to
yield loss problems are regularly pursued. Yield loss varies from period to period, but is
generally on the order of 10% or so. Improvements in yield loss would also dramatically
reduce the manufacturing cost of each battery.
It is worth explaining in very broad terms some of the causes of high levels of yield
loss on the RBAM web side. The first issue is the serial correlation of many of the
defects. Since the carrier web flows continuously and at a fairly high speed, many
problems which require that material be discarded are not detected until quite a few
battery rows have been affected. An example of such a defect would be skewed
composite rows. Sometimes the rotary equipment which places the composite rows will
place them with poor alignment. If one row is skewed, chances are that the next row will
be too, and so on. Also, if problems occur with material before it is placed on the web,
there is no way to discard the material. Bad material gets placed on the carrier web more
often than is desirable.
Another general source of high yield loss is the finishing equipment's sensitivity to
misaligned (misregistered) material. If one of the rows on an 18x4 battery strip appears to
be misregistered, the whole strip is discarded. This may seem very wasteful, and it is, but
it helps prevent the finishing machines from getting jammed and going down.
If the finishing equipment were able to process strips other than 18x4, it would be
possible to decrease yield loss, because the bad parts of 18x4 strips could be discarded and
the rest could be saved. Unfortunately, the finishing equipment can only process 18x4
strips, and cannot be reconfigured without large capital expenditures and large amounts of
downtime. (Keep in mind that R-5 is the only plant in the corporation which manufactures
batteries. Major equipment configurations which could keep the plant down for weeks are
generally considered impractical.)
Run time losses and yield losses drive the manufacturing cost of each battery
produced on the RBAM up substantially, and are worthy of the full-time attention of the
entire staff at R-5. Specific initiatives for improvements to the RBAM equipment are an
appropriate area for further work, but were not the focus of my internship. My efforts
focused on the IBAM project and ways in which we could optimize the performance of
the new line. Of course, an appreciation of the issues surrounding production on the
RBAM was essential to the IBAM team's ability to design a better line. A basic
understanding of these RBAM issues will contribute greatly to the reader's understanding
of the IBAM issues introduced later in this thesis. The next section explains in broad
terms some of the differences between the RBAM and the IBAM.
The IBAM team tried to identify major problem areas on the RBAM which could
be improved upon. In broad terms, I believe that the fundamental improvements designed
into the IBAM can be summed up as follows:
* Ability to discard bad material before it is placed onto the carrier web.
* Buffering of certain machines so that failures will not cause the entire line to go down.
* Improvement in the ability to precisely place components due to intermittent motion
instead of continuous motion.
The following diagram shows a sketch of the IBAM as of mid September, 1994.
The line can really be thought of in terms of the following five areas:
Web Transport System - The same paper carrier web used on the RBAM will be used
here. The major difference is that the web motion will be intermittent. The web will be
stationary when components are placed on it. Following placement of the components,
the web will advance an appropriate distance for the next row(s). Slack at certain points
along the web will allow different sections to advance separately from other sections, as
long as all sections advance at the same average rate over time.
Anode Cell - The unwind stand is buffered by the cut loop. Anode rows will be cut and
placed into canisters. The canisters will move around the cut loop's conveyor system,
where they will be picked up by a cambot (a pick & place device.) The anode rows will be
placed onto pallets which travel counter-clockwise around the main conveyor. The pallet
will travel under the slurry dispenser, where slurry patches will be applied, and then under
the vision check system. If the vision system detects any problems, the pallet will go to
the reject conveyor. If the pallet is not rejected, it will continue on to the next cambot,
which will place the anode row onto the carrier web.
Composite Cell - Very similar to the anode cell, except that pallets will travel clockwise
and will each carry two composite rows. The cambots between the cut loop and main
loop will each place a single composite row onto separate positions on the pallets. The
cambot between the main loop and the carrier web has a double head, and will pick and
place both composite rows simultaneously.
Cathode Machine - In the configuration depicted, the entire cathode machine is in-line,
overhead of the carrier web. The cathode rows are pocketed to give them their shape, and
then placed on the top of the battery stacks.
Vacuum Sealer - Unlike the RBAM, the IBAM's vacuum sealer will be in-line, and will
only seal three rows at a time instead of eighteen. There will be no vacuum table. Final
cutoff is after the sealer instead of before, as on the RBAM.
Overall machine yield and run time should be better than that of the RBAM web side. The
IBAM will produce about 1/6 the volume of the RBAM web side, so it will run much
more slowly.
The machine configuration described above was the one which I began to develop
computer models of. Details of the development of the models and of the problems that
the model uncovered are given in the remaining chapters.
Chapter 3 - Data-Based Characterization of Existing Process
This chapter is dedicated to a discussion of the data used to develop the various
machine downtime distributions which were inputs into the various computer simulations
developed during the internship. It begins with a discussion of why a detailed data
analysis was necessary. Next, data collection procedures and issues surrounding the
validity of the data are addressed. Finally, the chapter describes the specific statistical
tools and techniques that added value and created confidence in the distributions
themselves.
A model of a system is only as good as the assumptions upon which the model is
based. This certainly applies to computer simulations. Many questions regarding the
performance and behavior of the IBAM can be easily answered using straight-forward
analytical techniques. It is the questions that require insight into randomly occurring
phenomenon such as machine failures that can most easily be answered using more
sophisticated techniques such as computer simulation. The accuracy with which these
randomly occurring phenomenon are characterized will determine the usefulness of the
model.
In the case of the IBAM, the shapes of the machines' expected downtime
distributions are critical. I will explain in detail why this is so in subsequent chapters. In
addition to the shapes of the distributions, a detailed understanding of the specific types of
failures which occur on the RBAM is essential when predicting performance of the IBAM.
These performance predictions have profound effects upon design decisions of the IBAM
machinery itself, and upon managerial decisions regarding such issues as scheduling,
budgeting, and maintenance plans.
3.3 Procedures Currently in Effect for Data Collection
The plant currently has two separate processes in place for collecting machine
downtime data for the various components of the RBAM. The first process is one in
which the operators are responsible for entering appropriate data at a station containing a
keyboard and a monitor which is tied into the plant's VAX computer. The second process
is an automatic one which makes use of some data acquisition equipment. Data obtained
automatically is also processed and stored by the VAX computer.
The manual system is the plant's most reliable source of downtime information.
When a machine on the RBAM goes down, operators are supposed to note the time of
day to the nearest minute, and start a stopwatch in order to measure the duration of the
downtime. After troubleshooting is complete, and the RBAM is up and running, the
operator enters the time that the downtime began, the shift, the particular machine that
went down, the sub-system of the machine, the nature of the problem, the corrective
action taken, the duration of the downtime, and the number of consecutive occurrences of
the problem. The number of consecutive occurrences is recorded in order to relieve the
operators of the burden of having to make several identical entries which indicate short
downtime durations. If the machine fails for the same reason five times within a two-
minute period, the operators can simply call it a two-minute downtime and indicate five
consecutive occurrences rather than attempting to break the occurrences into five separate
entries. Little in the way of accuracy is lost by doing this, but much burden is taken off of
the operators.
This manual data collection system has its shortcomings. First of all, many
downtimes which last much less than one minute are simply not recorded. Secondly, since
the plant has only one VAX computer, the system responds very slowly if other users are
performing computationally intensive functions such as generation of yield reports and
runtime reports. Frustration with the slow response time will sometimes deter operators
from making proper entries. Great progress was made in addressing this problem when a
new and faster VAX computer replaced the older one in use just prior to the beginning of
the internship. But even with the newer computer, sluggish response can still be a
problem. Thirdly, there is a tendency on the part of the operators to not use the
stopwatch and to estimate the downtime duration by simply looking at the clock. This
becomes evident when one looks at histograms of the downtime durations for the various
machines. The number of eight, nine, eleven, and twelve-minute downtimes are generally
noticeably lower than the number of ten-minute downtimes. It would appear that if a
downtime is close to ten minutes, there is a tendency to simply call it a ten-minute
downtime and not bother distinguishing it from a nine-minute downtime. This
phenomenon was common around the five and ten-minute marks, but much less common
as downtime length increased. Since the difference between a seven-minute downtime and
an eight-minute downtime was of notable significance for reasons which I will explain
later, it was necessary to smooth these areas out a bit. The details of how this was done
are explained in section 3.5.4.
Overall, the quality of the data available for use in developing the computer models
of the IBAM was high. In speaking with many of my colleagues about their experiences
trying to obtain good data for various analyses of manufacturing operations in other
plants, I realized that although the data available at R-5 is imperfect and could be greatly
improved, it is basically sound data. The operators are usually diligent in recording the
data, and the data is sufficiently detailed to be of great utility in many different types of
analyses.
The automatic data collection process in place on the RBAM was still in the debug
phase for the duration of the internship. Data which was collected by the automatic
system was available, but the causes that the system assigned to downtime events were
considered very unreliable. There is great potential to have data that does not have the
inherent inaccuracies described in section 3.3.1 if the automatic data collection system is
finally debugged and deemed reliable. But for now, the manually entered data is still the
most reliable data source inthe plant.
The IBAM will have state-of-the-art data acquisition equipment, so the potential
to have an extremely high quality database of downtime, quality, and yield related data
certainly exists. Proper utilization of this resource will greatly assist future analysis of the
IBAM, and will help management greatly intheir decision-making processes.
The format in which the data was stored on the VAX was such that some up-front
manipulation was necessary in order to put the data into a meaningful and useable format.
A simple routine was available which downloaded the data in text format from the VAX
computer to the user's personal computer hard drive via the plant's Ethernet-based local
area network. Once the data was on the user's hard drive, whatever manipulation was
necessary could be performed easily on a spreadsheet. A sample of the downtime data
available follows. Note that each row identifies a single downtime event. The columns
contain the following information:
1 2 3 4 5 6 7 8 9
1/17/9413:00 A CATHODE LAMINATOR LAMINATION O UNKNOWN 10 1
1/19/94 4:30 C CATHODE DRUM WEB BREAK O RESET 5 1
1/20/94 4:05 C CATHODE PIN BELTS BROKEN BELT M REPLACED 20 1
The simulation package used could accept downtimes based upon clock times or times
that the machine was actually in use. It seemed sensible to transform the data from its
current clock-time form into a form which reflected machine usage time. If the downtime
distributions were developed based strictly upon clock times, then inaccuracies resulting
from not accounting for brief periods of non-scheduled time, maintenance time, and test
time would be introduced into the model. Additionally, if machine usage times are the
basis for the distributions, then the distributions for different machines will be independent
of one another and exclusive of one another, and will more accurately describe the actual
behavior of the individual machines.
This point requires some explanation: Since all machines on the RBAM are
completely coupled to one another, if any one goes down, the entire line goes down. If
for example, the cathode machine goes down at 13:00 for five minutes, and then runs for
five minutes, and then goes down again at 13:10, the actual machine usage time between
failures is only five minutes. The database indicates that ten minutes of clock time have
expired since the last failure of the cathode machine. If the downtime distributions were
constructed based upon clock times, they would be skewed in favor of longer times
between failures. Although the simulation can use these less accurate distributions, the
insight gained from studying the more accurate usage-based distributions is greater. The
most problematic machines are more easily identified inthe absence of the skewing that
comes from using clock times.
If clock times are used, then all machine distributions are dependent upon one
another in the sense that actual downtimes for each machine are incorporated into the
times between failures of all other machines. But more important are the effects of the
exclusive nature of usage-based downtimes on the simulation. If clock-based downtimes
are used, then two or more downtimes can occur simultaneously during the simulation.
This may adversely effect the simulation output. If usage-based downtimes are used, then
it is impossible for two downtimes to occur simultaneously since once one machine goes
down, the others are no longer in use. This more accurately reflects the behavior of the
machinery itself.
One may argue that the effects of the skewing of the clock-based distributions in
favor of long times between failures will be mostly offset by the occasional simultaneous
occurrence of two or more downtimes. I argue that it is much better to attempt to
characterize the machine behavior as accurately as possible rather than rely on two
different inaccuracies to offset one another.
I initially elected to use data from the period from January 1, 1994 through June
30, 1994 to develop the machines' time-to-failure (ttf) and time-to-repair (ttr)
distributions. I asked the engineer responsible for the data collection process and for
generation of associated performance reports what he thought an appropriate time period
was. He indicated that for the most part, he uses a three-month period when he is asked
to generate reports which give insight into long-term trends. His philosophy was that any
cyclical phenomenon related to machine performance usually repeat themselves several
times each quarter. Increasing the timeframe studied would increase the workload
without increasing the insight gained from the analysis.
Despite this advice, I elected to use a six-month period. I reasoned that the data
was readily available, and a longer timeframe would increase the statistical significance of
the distributions developed. It seemed reasonable to use the most recent six-month period
in order to ensure that any recently developed trends were captured.
Once I began analyzing the data, I noticed that there were unusually large gaps
between documented machine failures from January 1 through January 16. In asking
around, I could not come up with any explanation for this such as the VAX computer
being down, or unusually high non-scheduled time due to holidays, testing, or
maintenance. The documented failures were so thinly spread during this period that there
is no chance that it was due to particularly good machine performance. The machines
never run as well as data from this period would indicate. In the end, it seemed best to
simply discard the period from January 1 to January 16 as unreliable and do the study with
data from January 17 on.
Several statistical tools were used in an attempt to simplify the distributions and
make them more accurate. The four worth examining in this section are analysis of
variance, the chi-square goodness-of-fit-test, least squares regression through the use of
non-linear programming, and a simple type of smoothing which was applied to the
distributions.
ANOVA was useful when deciding whether or not similar types of equipment had
statistically significant differences in their mean times to fail and mean times to repair. The
best way to describe ANOVA's usefulness is with a discussion of how it was used to
develop distributions for the different unwind stands on the RBAM.
Unwind stands are the machines on which the rolled up materials from pre-
assembly are loaded and slowly fed into the various stages of the battery assembly process.
The RBAM has six unwind stands as follows:
* anode unwind (1)
* composite unwinds (3)
* cathode unwind (1)
* carrier web unwind (1)
There are two main reasons that it would be desirable to use common
distributions for equipment which behaves similarly. The first is that it saves some work in
the development of the distributions and in the subsequent creation of the computer
simulation itself. The second is that distributions developed using more data are more
likely to be representative of future machine behavior than those developed with less data.
If, for example, all six unwind stands behaved similarly, a common distribution which
covered all of them would be based on six times as much data as individual distributions
would be. With this in mind, I attempted to see how similar the performance of the
unwinds was by formulating a null hypothesis (HO)and testing its validity using ANOVA.
The null hypothesis was:
H0 All unwind stands have the same mean time between failures.
In order to test HO, times between failures for the six unwind stands were lined up in
columns and an ANOVA was perfaomed. The confidence level was chosen as 90%. The
following table resulted:
Anova: Single Factor Unwind TTF
SUMMARY
Groups Count Sum Average Variance
ANODE 92 143380 1558.48 2.79E+06
CARRIER 206 142784 693.13 5.26E+05
CATHODE 383 143789 375.43 1.82E+05
COMP 1 59 137625 2332.63 7.60E+06
COMP 2 71 142871 2012.27 3.65E+06
COMP 3 39 137565 3527.31 1.00E+07
ANOVA
Source of Variation SS df MS F P-vakue F crit
Between Groups 6.24E+08 5 1.25E+08 69.83 3.74E-61 2.22
Within Groups 1.51E+09 844 1.79E+06
An underlying assumption of ANOVA is that the data sets being compared have
equal variance. Table 1 indicates that for these sets of data, the variance increases as the
mean increases. Under these circumstances, it is often useful to transform the data so that
the ANOVA assumptions are met. One useful method for determining what type of data
transformation to use is to plot the natural log of the standard deviation as a function of
the natural log of the mean for the data sets being compared. Then, determine the slope
(1) of a line which fits the data fairly well. The data should be transformed by raising it to
the 1-13 power. If 13=1, then the data should be transformed by taking the natural
logarithm.
The following chart shows the the log(standard deviation) plotted against the
log(mean) for the six sets of unwind ttf data, along with lines with slopes equal to 0.5, 1.0,
and 1.5, which are included for purposes of comparison:
Log(Standard Deviation) vs Log(Mean) for TTF Data
0 2 4 6 8
Log(Men)
The 1=1 line fits the data fairly well, so I took the natural log of the data and ran another
ANOVA. The following table resulted:
ANOVA
Source of Variation SS df MS F P-value F crit
Between Groups 521.76 5 104.35 57.22 2.90E-51 1.85
Within Groups 1532.00 840 1.82
H0 All composite unwind stands have the same mean time between failures.
ISUMMARY
Groups Count Sum Average Variance
COMP 1 59 422.04 7.15 1.58
COMP 2 71 500.09 7.04 1.82
COMP 3 38 289.18 7.61 1.67
ANOVA
Source of Variation SS df MS F P-value Fcrit
Between Groups 8.24 2 4.12 2.42 0.09 2.34
Within Groups 280.76 165 1.70
Since F>Fcritical, we reject HO. This means that the three composite unwinds apparently
do not behave similarly. One further test compared composite unwind #1 to composite
unwind #2. The following table resulted:
Anova: Single Factor
SUMMARY
Groups Count Sum Average Variance
COMP 1 59 422.04 7.15 1.58
COMP 2 71 500.09 7.04 1.82
ANOVA
Source of Variation SS df MS F P-vakie F crit
Between Groups 0.39 1 0.39 0.23 0.64 2.75
Within Groups 219.10 128 1.71
This ANOVA tells us that at the 90% confidence level, we cannot reject HO. A
reasonable assumption is that composite unwinds 1 and 2 behave very similarly and have
the same mean ttf. Further ANOVAs not documented here were performed comparing
the various unwind stands to one another in varying combinations. The only case in which
H0 was not rejected was in the case of composite unwinds 1 and 2. Why composite
unwind 3 was significantly different from Iand 2 is unknown.
The IBAM will have four unwind stands as opposed to the RBAM's six. This is
because only one composite unwind will be used in the IBAM instead of three. ANOVA
made it clear that it was necessary to use separate ttf distributions for the anode, carrier,
and cathode unwinds in the IBAM simulation. The distribution for the single IBAM
composite unwind was developed using data from the RBAM's composite unwinds 1 and
2. This decision was based on the premise that it was best to be conservative in the
analyses which will be described in subsequent chapters. It would be very undesirable to
underestimate the effects of a problem because of overly optimistic assumptions. The
performance of composite unwinds 1 and 2 was consistently worse than that of composite
unwind 3. The conservative approach was to use failure data from 1 and 2 instead of from
3.
The ttr distributions for the unwind stands yielded interesting results. The
following null hypothesis was tested using ANOVA:
ISUMMARY |
ANOVA
Source of Variation SS df MS F P-vakie F crit
Between Groups 588.69 5 117.74 1.38 0.23 2.22
Within Groups 71835.48 844 85.11
3.5
3
2.5
2
1.5
1
0.5
0
0 0.5 1 1.5 2 2.5
Log(Mean)
Although the point which corresponds to composite unwind stand #1 is an obvious outlier
since its standard deviation is so large, a line with |=1 seems to fit the best, so I took the
natural logarithm of the data and performed another ANOVA. The following table
resulted:
ISUMMARY
Groups Count Sum Avewrage Variance
ANODE 92 149.18 1.62 0.37
CARRIER 206 338.09 1.64 0.49
CATHODE 383 602.61 1.57 0.29
COMP 1 58 83.30 1.44 0.61
COMP 2 71 89.37 1.26 0.32
COMP 3 39 50.94 1.31 0.36
ANOVA
Source of Variation SS df MS F P-value F crit
Between Groups 11.48 5 2.30 6.12 1.39E-05 1.85
Within Groups 316.09 843 0.375
Total 327.57 848
Table 6: ANOVA Results
The transformed data adheres to the equal variance assumption much better than the raw
data does. The ANOVA based upon the transformed data indicates that all unwind stands
do not have the same mean ttr, since F>Fcritical and HO is therefore rejected. One more
ANOVA is useful:
SUMMARY
Groups Count Sum Average Variance
ANODE 92 149.18 1.62 0.37
CARRIER 206 338.09 1.64 0.49
CATHODE 383 602.61 1.57 0.29
ANOVA
Source of Variation SS df MS F P-value F crit
Between Groups 0.66 2 0.33 0.92 0.40 2.31
Within Groups 245.61 678 0.36
In this case, since F<Fcritical, we cannot reject HO. This means that the repair times of
the anode, carrier, and cathode unwind stands are similar.
Tables 6 and 7 suggest that it would have been better to separate the composite
unwind ttr data from the ttr data for the anode, carrier, and cathode unwind stands before
developing ttr distributions. The ttr distribution which was developed for all of the
unwind stands had a mean ttr of 6.14 minutes. The anode, carrier, and cathode unwind
stands together have a mean ttr of 6.24 minutes, while the composite unwind stands have
a mean ttr of 5.71 minutes. The effects that combining the unwinds' ttr data had on
simulations which used the distributions were negligible. Total downtime for the unwind
stands was still correct, but slightly more downtime was allocated to the composite
unwind, and slightly less to the other three unwind stands.
3.5.2 Chi-Square Goodness-of-Fit Test
It was important to decide exactly how to represent the ttf and ttr distributions in
the simulation itself. The modeling software offered a number of built-in distributions in
which it was only necessary for the developer to identify the distributions' appropriate
parameters. For example, the exponential distribution, written E(a), required a single
parameter: the mean. The normal distribution, written N(a,b) required two parameters: a
mean and a standard deviation. Distributions were assigned in the model's downtime
editor. TTF was assigned in a column entitledfrequency, while ttr was assigned in a
column entitled logic.[3] In the following example, Machine A's failure and repair times
are exponentially distributed with means of 284 minutes and 14.7 minutes, respectively.
Machine B's failure times are normally distributed with a mean of 284 minutes and a
standard deviation of 50 minutes, while its repair times are normally distributed with a
mean of 14.7 minutes and a standard deviation of 3 minutes:
freauencv k
Machine A E(284) min E(14.7) min
Machine B N(284,50) min N(14.7,3) min
Frequency Logic
Machine A machine_a_ttf () min machine_a_ttr( min
According to table 10, Machine A will go between 0 and 15 minutes between failures 25%
of the time, between 15 and 30 minutes between failures another 25% of the time, and so
on.
I wanted very much to be able to define the different machines' performance in
terms of the convenient built-in distributions. This would allow the flexibility to define the
distributions' various parameters with variables that could be easily modified from test to
test. For example, suppose that Machine A's behavior was defined as follows:
Freauency Loaic
Machine A E(variablel) min E(variable2) min
It would be extremely easy to test the effects that improved ttf or ttr performance of
Machine A had on the overall system. In order to test these effects, it would only be
necessary to change the values of variablel and/or variable2 for different replications. The
different values for variablel and variable2 could be set up in an external file that the
simulation received input from. Outputs from the different replications could then be
compared easily, and the desired effects could be readily quantified.
With this in mind, I tried to determine if the machines' performance could be
accurately defined by built-in distributions. I began by creating histograms of the ttf and
ttr data for the different machines. All of the histograms which I created had exponential
shapes to them. The following example isa histogram ofthe RBAM's cathode ttf data:
Catlode TTF Data (0Occummes vs. minutes)
100 -
80-
soo
60-
40o-
20-
--- -rhM1 I i I
0
Since the distributions' shapes were exponential in nature, I attempted to define them in
terms of the exponential distribution. The exponential probability density function (pdf) is
given by:[1]
f(x) e- .
where:
In order to convert the histogram into a pdf, it was necessary to do a couple of things.
First, the numbers of occurrences in each bin had to be converted into a percentage of the
total occurrences. Then, the percentage associated with each bin had to be divided by the
width of the bin. This was necessary to ensure that the area under the pdf curve would
equal 1. Finally, the value of the (%of occurrences)/(bin width) was assigned to the
midpoint of each bin. The resulting pdf is plotted below:
POF for Cahod TTF
0.000
0.007
0.006
0.006
0.004
0.008
0.002
0.001
0
0 250 1000 1250 1500
Note that the independent axis is only plotted out to 1500 minutes as opposed to 3000.
This was done to better show the shape of the most important part of the curve.
The data used to create the above empirical pdf had a mean value of 284.7
minutes. In other words, the RBAM cathode machine has a mean time to failure of 284.7
minutes. For this reason, I first examined an exponential pdf with a mean of 284.7
minutes. The empirical pdf and the exponential pdf with a mean of 284.7 are plotted in
figure 12:
Empirical Cathode TTF POF and E(284.7) PDF
0.000
0.008
0.007
- Empidl
0.008
- E(284.7)
0.005
0.004
0.003
0.002
0.001
0
0 200 400 600 800 1000 1200 1400
I separated the data into 12 bins and performed the chi-square goodness-of-fit test. The
test is used to see whether data fits a particular type of distribution. The test says that for
n data points, if the data is broken into k mutually exclusive and exhaustive bins, the
variable Yi denotes the number of occurrences in each of the k bins, and Pi is the
probability that the distribution that the data is being fitted to assigns to each of the k bins,
then;
qk-1 Z
i=1 PiP i
X2(a; k-i-h)
The following table shows how the calculations were performed to see if the data passed
the goodness-of-fit test when fitted to an exponential distribution with 1/l=284.7:
Chi-test 23.21
ca=.01
Table 12: Goodness-of-Fit-Calculations
Since qk-l =65.16, and X2(.01;10) is only 23.21, HO is convincingly rejected, even
at the a=.01 level. The exponential distribution with a mean ttf of 284.7 does not fit the
data.
Verification tests which are described in section 3.6 also showed that the E(284.7)
distribution did not fit the data adequately. From figure 12, it appears as if there are many
more instances where the times between subsequent failures are on the order of 0 to 50
minutes than the E(284.7) distribution would indicate. Also, the E(284.7) curve is
consistently above the empirical curve between about 100 and 500 minutes. It appears as
if the outlying points in the 2000-3000 minute range (shown in figure 10) greatly increased
the average value of the data, and an exponential distribution with the same mean value as
that of the data is not quite right. It falls off too gradually in the 0-500 minute range.
3.5.3 Exponential Regression
I next attempted to fit an exponential curve to the data by using a least squares
regression. I figured that a least squares regression would minimize the skewing effects of
the data points representing extremely long times between failures. I set up a simple non-
linear program to minimize the sum of the squares of the errors between the empirical
curve and the fitted curve by changing the value of X,the failure rate. The program
structure is shown below:
X= 0.006965
1/_I= 143.574
As shown in table 13, the value of 1/, which provided the best fit curve was 143.6
minutes. A plot of all three curves is shown in figure 13:
Empirical Cathode E(2847), and E(143.) PDFs
U.IR
0.008
0.007
0.008
0.005
0.004
0.003
0.002
0.001
0
0 200oo 400 600 800 1000 1200 1400
The fitted curve did not accurately represent the machine behavior any better than the
E(284.7) curve did. In fact, it was much worse. Although the calculations are not shown,
a chi-square goodness-of-fit test using the same bins as before was performed to
determine if the E(143.6) curve fit the data. The qk-. value was 5,870, which is vastly
greater than X2(.01;10), which is only 23.21. By minimizing the effects of the outlying
points representing long times between failures, the distribution became too heavily
weighted in favor of short times between failures. During validation tests, the machine
failed too often when the E(143.6) distribution was used.
In order to best capture the true behavior of the cathode and all other machines, I
decided to use empirical distributions rather than the convenient built-in distributions. As
was described previously, a consequence of this decision was that it would not be as easy
to test the effects of varying degrees of machine reliability from replication to replication
of the computer simulation. By using a different approach, however, it was possible to
partially offset this limitation. Rather than use variables to define the distributions'
parameters, it seemed to work well enough to simply use the same distribution in every
replication, and multiply the times returned by the sampling process by some constant. By
changing the constant from replication to replication, the effects of different machine
performances could be tested. An example of how this logic would look in the downtime
editor follows:
Freauency Logic
Machine A variablel*machinea_ttf() min variable2*machinea ttr() min
By using variables in this way, the ttf and ttr distributions would maintain their shapes
from replication to replication, but their independent axes would expand or shrink
according to the values of their corresponding variables. Despite the initial concern over
the lack of flexibility that empirical distributions would introduce, this approach of
sampling from the same distribution and then multiplying by a constant seemed to work
quite well, and provided the required flexibility.
3.5.4 Smoothing
After developing empirical ttf and ttr distributions for the various RBAM
machinery, it was important to perform a reality check to see how well the simulation
software interacted with the distributions and how closely the simulation results reflected
the documented machine behavior. In order to accomplish this, a very simple simulation
named test I was developed inwhich there was a single machine which had an entity pass
through it every 2.7 seconds, the overall rate at which the proposed IBAM would produce
battery rows. The actual time chosen was unimportant, since the percentage of downtime
caused by the distributions was the value of interest.
Initial ttf and ttr distributions were developed for the following RBAM areas:
* Cathode machine
* Unwind stands (all unwinds combined)
* Slurry delivery
* Vacuum sealer
* Carrier web
For each replication, the test l's single machine was programmed to fail according to one
of the above machines' empirical ttf and ttr distributions. Ten replications of 200 hours
duration were run on test I for each of the above machines with the exception of the
unwinds. They were tested over the weekend, so 20 replications of 300 hours each were
run. Each 300 hours of simulation time took about an hour of actual time. Percentages of
downtime were recorded for each replication and compared to expected percentages.
The expected downtime percentages were determined by comparing the total
number of downtime minutes caused by each machine to the total number of minutes in
which the RBAM actually ran during the period studied. For example, the cathode
machine was the cause of 7,531 minutes of downtime during the period studied. During
the same period, the RBAM actually ran for roughly 143,800 minutes. The expected
percentage of downtime from test I when sampling from the cathode's ttf and ttr
distributions would then be (7,531)/(7,531+143,800), or 5.0%. The following table shows
the results of the test runs. The 90% confidence intervals are based upon the student's T
distribution since the number of replications was probably too low to warrant the use of
the normal distribution:
The results were very encouraging. The mean values are all fairly close to the expected
values. The furthest off is the vacuum sealer, but the expected result still lies within a
90% confidence interval of the mean of the replications. The real aim is to have all of the
distributions, when taken together in a large computer simulation, accurately represent the
overall manufacturing line behavior. The above results suggest that the methods used to
quantify the machines' behavior profiles were sound and analysis based upon these
methods should be reliable.
The ttf and ttr distributions for the cathode, the vacuum sealer, and the carrier web
were ready for use in the computer models, but the unwinds and slurry system required
further breakdown into more specific sub-system distributions to ensure that they were
properly applied to the simulations. Complete listings of the distributions used inthe
various models are contained in Appendices B and C.
Chapter 4 - Development of the Computer Models
The software used to develop the computer simulations was ProModel for
Windows version 1.10. ProModel is a package best suited for the modeling of discrete
part manufacturing systems, although continuous flow or batch processes such as chemical
mixing operations can be modeled as long as the system can be described by discrete
entities like gallons of fluid. The anode and composite cells lent themselves nicely to
modeling inProModel, since lots of discrete entities like battery components and pallets
were the objects of interest. The web transport system and cathode machine required
more creativity to model since they involved continuous webs of material. The vacuum
sealer was simply treated as a part of the web transport system.
When the internship began, one model of the IBAM's anode cell which was
developed inProModel by a third party was available for me to study. This model, in
conjunction with ProModel's documentation, offered enough insight into modeling of
manufacturing systems for me to get started. My objective was to improve the model of
the anode cell as necessary, and develop additional models of the composite cell, the
cathode machine, and the web transport system. After these four modules were
completed, I would then merge them together into a single model of the entire IBAM. My
intention was to use this very detailed IBAM model as a platform to perform designed
experiments and sensitivity analysis on so that the performance of the IBAM could be
optimized.
This initial modeling task was quite substantial. I wanted to capture as fine a level
of detail as I could. For example, I broke motions of the cambot heads down into distinct
events which took only fractions of a second. I tried to make the graphics sufficiently
detailed so that a user could tell the status of any pallet, canister, battery component, or
machine in the system at any time. This level of detail took its toll on the modeling
process in three ways. First, it turned the development of the model into a process which
took about 8 weeks' worth of attention. Second, it slowed the performance of the model
itself down. The model was developed and run on a personal computer with the following
pertinent performance parameters:
* 486DX2/66 processor.
* 8 MB RAM
* 19 MB virtual memory swap file established on the hard drive.
Despite the highly capable configuration of the computer, the IBAM simulation would
only run at about 1/3 the speed of real time. In other words, it would take the computer 3
seconds to simulate I second of machine run time. When the model was tested on a
computer with a 90 MHz Pentium processor and 16 MB of RAM, the speed roughly
doubled, but was still aggravatingly slow. With the graphics disabled so that machine
motions were simulated but not displayed on the monitor, the simulation would run nearly
10 times as fast as real time. This is much faster than with the graphics enabled, but even
at this speed an attempt to simulate 6 months worth of production would take 18 days.
The third way that the high level of detail adversely affected the modeling effort
was by making it very difficult to keep pace with the rapidly changing nature of the IBAM
project. Any time configuration changes occurred, the model would have to be changed,
sometimes dramatically. I learned to make my work as modular as possible inorder to
protect myself against changes inthe machinery configurations, but these changes were
always setbacks despite my efforts to modularize. The changes also underscored the need
to have an actual team member doing the modeling rather than trying to contract the work
out to third parties.
In this thesis, the data analysis was introduced before the modeling was discussed.
During the internship itself, however, the modeling effort was begun well before the data
analysis was. By the half-way point of the internship, the detailed IBAM model was 99%
complete. All that remained in order to make the model functionally complete was some
debugging of a problem that surrounded the interface between the web transport system
and the anode cell. It was at this point, just after the midstream internship review, that I
began the data analysis described in detail in chapter 3.
In retrospect, I believe that in general, it would be more valuable to perform a
detailed statistical analysis and develop machine behavior profiles before beginning a
substantial modeling effort. Since I knew nothing about the process of manufacturing
batteries when I began the internship, the modeling effort was an excellent way for me to
learn about the process and gain exposure to important issues surrounding production.
However, if a modeler already knows a good deal about the process that he
intends to model, it would be more beneficial to perform some statistical analysis first.
This is because the statistical analysis is of great benefit by itself. It is very likely that
insight gained from the statistical analysis will help to streamline the modeling process by
ensuring that it is focused on answering important questions. A statistical analysis which
is not followed by a modeling effort will very likely be of great benefit to a project. A
model that is not based upon a sound statistical analysis is not very likely to be of great
value and will probably only help to answer the simplest of questions.
My initial perspective was that the statistical analysis would help me develop a
better model. Now that the internship is complete, my perspective has changed. I now
believe that modeling is an excellent way to supplement a good statistical analysis and to
further explore questions and insights which developed from the statistical analysis. My
approach of developing a detailed model, supplementing it with a statistical analysis, and
then seeing what I could learn was not the optimum approach. The following section
explains how the statistical analysis greatly narrowed the focus of the modeling efforts and
pointed them in an extremely productive direction.
Following the development of the machine ttf and ttr profiles described in chapter
3, I began to wonder if yield loss due to dry slurry would be a substantial problem on the
IBAM. On the RBAM, it takes about I minute from the time that the first slurry patch is
applied until the battery strip passes through the final cutoff mechanism. From there, the
battery strip usually waits a little over 30 seconds until it is vacuum sealed. Due to the
extra length and slower operating speed of the IBAM as it is shown in figure 7, in steady
state, the first slurry patch would be exposed for about 1minute before it was even placed
onto the carrier web. Then, it would take about 7 minutes for the battery row to make its
way down to the vacuum sealer. All the while, of course, the slurry would be exposed and
subject to dry-out. (Recall that the plant's policy is to discard any batteries whose slurry
has been exposed for 15 minutes or more.) The following plot shows intuitively why one
might expect the dry slurry yield loss issue to be a pressing one on the IBAM:
50.00% 4
45.00% -
w 35.00%-
30.0W% -
r
25.00% -
20.00% -
A 15.00% -
10.00% -
5.00%-
0.00%-
Minutes
On the RBAM, all downtime occurrences in excess of 15 minutes cause the line's shift
registers to fill with a signal indicating that the contents of the web must be discarded.
Intuitively, if the area under the above curve represents all downtime events, then the area
under the curve to the right of the 15 minute mark on the x axis would represent the
portion of downtime events which lead to dry slurry yield loss. On the IBAM, the area
under the curve to the right of the 8 minute mark on the x axis would represent the
portion of downtime events which lead to dry slurry yield loss. The 8 minute mark is the
relevant mark for the IBAM since it takes 7 minutes for a row to get from one end of the
machine to the other, leaving only 8 minutes of slack time. Note that the current policy of
discarding material on the RBAM following downtimes in excess of 15 minutes really
amounts to having a 16 minute threshold, since it takes about I minute for batteries on the
RBAM to travel form the point of initial slurry application to final cut-off. Despite this
small inconsistency, I assumed that on the IBAM, a 15-minute threshold would mean that
the total exposure time could not exceed 15 minutes.
With the specific intention of testing the impact of the 15-minute slurry threshold
on IBAM production, I developed a new simulation of the IBAM. Unlike the detailed
IBAM model which had taken me weeks to develop, this new model was much simpler
and took only a day and a half to develop. The only graphics inthe new model were a few
counters which showed the number of battery strips in a few different places throughout
the line. In discussions, I referred to this model as my block diagram model of the IBAM.
In my initial tests, I used the machine profiles of the RBAM to describe the ttf and ttr
performance of the corresponding IBAM machines. The assumption was that the IBAM
cathode machine would fail according to the same distributions as the RBAM cathode
machine, and so on. The simulation time-stamped the application of the first slurry patch,
and then tabulated the total number of strips over a period of time which were sealed after
15 minutes had passed. This number was compared to the number of strips which were
sealed before 15 minutes had passed, and an estimated dry slurry yield loss was calculated.
My estimate based on this first set of assumptions was that IBAM dry slurry yield loss
would average about 8%over time.
Two items are worthy of note at this point. The first is that since the block-
diagram simulation was much simpler than the detailed simulation, it ran at a rate about
100 times faster than real time. This enabled me to start a simulation at about 4:00 P.M.,
run it for 15 hours or so overnight, and have the results waiting when I arrived at 7:30
A.M. the next morning. By doing this, I could readily simulate 1500 hours, or roughly
two months worth of production each night.
The second noteworthy item is that the simulation did not reveal the problem. It
was very useful intesting the effects of the problem, but it was the statistical analysis and
the distribution shapes which provoked the thoughts which led to the analysis of the dry
slurry issue.
I informed the IBAM program manager of my preliminary findings, and we
discussed the impact that the dry slurry issue would have on the project.
Recall from chapter 1 that the objectives of the IBAM are the following:
* A step increase in annual battery manufacturing capacity.
* A step increase in yield.
* Continual improvement in customer satisfaction by the elimination of
manufacturing defects.
The program manager was clearly concerned that the fundamental objectives of the
program would not be met if the dry slurry issue was not addressed. He first requested
that I refine my assumptions a bit and perform further analyses. The ttf and ttr
distributions used in the block diagram simulations were based upon data that was
manually entered into the system by the RBAM operators. While the operators were
generally faithful to their responsibility to enter the downtime data, sometimes the entries
were not made. This meant that the machines actually failed at a rate that was somewhat
higher than the ttf distributions indicated. In fact, the initial simulation runs indicated that
overall IBAM run time was on the order of 75% of scheduled time. The RBAM run time
during the period that the distributions were based upon was actually about 67% of
scheduled time. Only two differences should have existed between the RBAM run time
performance and the simulated IBAM run time performance. The first was due to the fact
that the IBAM had only three baby hoppers in its slurry delivery system instead of four.
The second was that the IBAM had fewer unwind stands, and that there were buffers
between the unwind stands and the rest of the line. Since neither the baby hoppers nor the
anode and composite unwind stands were major sources of downtime, these differences
should have been minimal, and certainly should not have increased expected run time from
67% to 75%. The difference was mostly due to the bias in the ttf distributions that
resulted from operators sometimes neglecting to make the proper entries.
The program manager wanted to know what dry slurry yield loss would result if
the IBAM only achieved 67% run time. This gave rise to the question of how the
distributions should be modified in order to have 67% run time from the simulation. We
decided that the most reasonable and workable assumption to make was that the
downtime events which were not entered were evenly distributed among all machines for
which distributions were developed. We had no reason to believe otherwise. This
assumption was implemented by using the multiplicative factor described in section 3.5.3
on all of the ttf distributions. By trial and error, I determined that a factor of 0.7 was
appropriate. Ineffect, the simulation would determine, via sampling from the appropriate
ttf distribution, a time until the next failure for each machine, and then multiply the time by
0.7 in order to make it fail sooner. This compensated for the entries that the operators did
not make. The simulation consistently averaged 67% or 68% run time when the factor of
0.7 was used. The estimate of dry slurry yield loss increased to about 10.4% based upon
these assumptions, and went all the way up to 12.3% inthe end when 1 final refinement
was made.
My initial feeling was that the length of the carrier web would have to be
shortened. The design of the IBAM at this point had a carrier web with 154 1x4 battery
rows in process with exposed slurry. At 2.7 seconds per battery row, this equates to
about 7 minutes of work in process (WIP) with exposed slurry. Cutting rows out of the
carrier web would reduce exposure time and improve the dry slurry yield loss issue.
The IBAM program manager had many years experience at the plant, and
wondered whether the 15-minute slurry threshold was really a valid threshold, or if
perhaps using a higher threshold would still allow for batteries of sufficiently high quality.
The threshold of 15 minutes had been established intests conducted inthe early 1980s,
and had been the plant's policy ever since. But none of the people who were involved in
the test were still employed at the plant and no adequate documentation of the insight
gained from the tests could be found. Since the dry slurry issue would be very prominent
on the IBAM, upper management decided to completely revisit it and begin a thorough
round of new tests to determine if the slurry threshold could be increased, and to develop
a better understanding of the issue in general. The tests would take in excess of 270 days,
since 270-day incubation periods were sometimes necessary for minor defects to fully
manifest themselves. Management would be unwilling to make policy changes until the
270-day test results were thoroughly analyzed.
At the time the relevant decisions were being made on the IBAM, and later when
this thesis was written, the slurry threshold tests were incomplete and whether or not the
slurry threshold could be increased remained unknown.
Simulation was the perfect tool for determining the effects of various carrier web
lengths and possible slurry threshold levels. With all of these possibilities in mind, the
IBAM program manager wanted to see a family of curves that showed dry slurry yield loss
on the dependent axis and slurry threshold time on the independent axis, with a curve for
each possible length of carrier web. Then, he wished to bring the entire IBAM team into
the process of solving the problem. The following set of curves was developed and
presented to the IBAM team:
10.00%
8.00%
6.00%
4.00%
3.28%
2.00% 2.01%
1.14%
0.35%
0.00%
15 20 25 30
This chapter explains the decision-making process that the IBAM team used in
order to arrive at a proper course of action in response to the dry slurry yield loss
problem. The first two sections explain the information that the team felt it should have in
order to weigh all options adequately. The remaining sections examine the results of
further simulations which were run in support of the team's efforts, the implications of
these results, and the decision-making methodology used to arrive at a final solution.
At the time that the entire IBAM team was introduced to the dry slurry yield loss
problem, ideas for solutions were limited. Only the IBAM program manager and I had
given the problem any thought, and we were primarily considering three options: either
shortening the carrier web, increasing the slurry threshold, or both. We also considered
the possibility of speeding up the machinery. Although speeding up the machine would
decrease slurry exposure time and actually increase throughput, it was clear that this
approach alone would not be sufficient to bring dry slurry yield loss down to an acceptable
level. The speed of the IBAM might be able to increase by 25% or so, but it could not
triple or quadruple. Other measures would be required.
The creativity of the IBAM team members was noteworthy. Possibilities ranging
from re-moistening the slurry with a mist of water to inverting the battery stack so that the
slurry was applied to less absorbent material were considered. Each idea was evaluated on
its merits, but inthe end, fundamental problems with each of these ideas precluded them
from receiving serious consideration for implementation. The team decided that some
combination of the following three measures would be the answer to the problem:
* Increase the speed of the IBAM.
* Reduce the length of the carrier web.
* Increase the slurry threshold level.
Since tests to determine if the slurry threshold could be increased beyond 15 minutes were
already being conducted by appropriate personnel, the IBAM team focused on
determining ways to increase the speed and reduce the web length. The team felt that
more refined information than was available from the first set of simulations would be
required. The analyses that the team requested me to perform are described in the next
section.
Most of the creative energy of the IBAM team was spent figuring out ways in
which the carrier web length could be reduced. Fortunately, the web transport system was
not yet constructed, and was conceptually less mature than the other major subsystems.
This meant that changes to the web transport would be less expensive than changes to
other more fully developed areas. The team gave serious consideration to roughly a dozen
new IBAM configurations. The concepts considered had web lengths ranging from 154
rows all the way down to 70 rows. Each row took 2.7 seconds to process, so on a 70 row
web it would take just over 3 minutes to get from the point where the anode layer was
applied to the web to the vacuum sealer. In steady state, the slurry patch on the anode
row would be about I minute old by the time it was applied to the web, so the total
exposure time with a 70 row web would be roughly 4 minutes. A 154 row web would, on
the other hand, take about 8 minutes in steady state. The different web lengths were for
the most part a function of whether or not the anode cell and composite cell were on the
same side of the carrier web, and of how much machinery was placed off-line. Recall that
in the context of the IBAM, a machine is considered off-line if it is not physically located
above or below the carrier web. Being off-line does not mean that the machine is
decoupled from the line.
The dozen or so concepts that the team developed fit into several categories with
regard to web length. The baseline was 154 rows, several were around 120 rows, several
others were around 100 rows, and the shortest was 70 rows. The team felt that by giving
up rows, operator accessibility and machine maintainability may suffer because the
machinery would be located closer together.
There were also questions regarding the expandability of some of the options. One
firm criterion for any concept was that, if implemented, it must allow sufficient factory
floor space for construction of another identical line should expansion become necessary
in the future. The options with the shortest web lengths were not necessarily the easiest to
expand. This is because if an option was shorter, it was probably wider, and the factory
floor space dedicated to the IBAM and its possible expansion was rectangular, not square.
In order to evaluate the different options thoroughly, the team wanted to know
how big a difference in estimated dry slurry yield loss there was between say, a 100-row
option and a 70-row option. The accessibility issues might very well outweigh the yield
loss improvements if the yield loss improvements were small. For this reason, they asked
me to run simulations on web lengths of 100 rows and on 70 rows in addition to the ones
which were already complete. The 120-row options were close enough to the 115-row
scenario already tested that further tests of this web length were deemed unnecessary
given the time constraints we were facing.
The team also felt that it was very important to see analysis which was based on an
alternative set of assumptions. The assumptions that the first simulations were based on
were as follows:
SAll major IBAM systems will fail according to the same ttf and ttr distributions
as the corresponding major RBAM systems.
* Since the data on the VAX computer is based upon manually entered data, all
ttf distributions will be adjusted consistently so that overall simulated line run
time is about 67%. (A multiplicative factor of 0.7 was determined through trial
and error.)
* All areas of new technology on the IBAM for which there is no data available
will operate at 100% efficiency.
The most questionable of these assumptions is that new technologies will operate at 100%
efficiency, but the rationale for this will be explained later. The above set of assumptions
was considered a worst case scenario by the IBAM team. It was considered worst case
because the team collectively believed that they had designed a line which would be much
more reliable than the current RBAM. Nobody believed that the IBAM would be less
reliable overall once the plant had proceeded down the learning curve which would
certainly exist.
In order to incorporate into the simulations the belief that the IBAM would be
more reliable than the RBAM, the team analyzed the RBAM's major sources of downtime
events and predicted how the IBAM would be different with regard to each type of failure.
The following spreadsheet shows the top 70% of RBAM failures and the team's collective
estimate of how the IBAM performance would change in each case. The total downtime
for the period studied was 55,641 minutes. Major RBAM systems and their
corresponding total numbers are inbold type. Individual failure types are in regular type.
The columns should be interpreted as follows:
Cathode Creaser Web Break 2038.94 5.77 404 0.73% Reduce by 67%
Drum Web BreIk 1347.08 6.17 654 1.18% Reduce by 50%
Dnram Web Jump Pins 2098.74 15.23 990 1.78% Reduce by 50%
4266.58 30.64 1011 1.82% Reduce by 75%
Pin Bbel Web Jump Pimr 1714.23 13.07 1072 1.93% Reduce by 25%
Pockatr Drive Fault 4600.52 8.37 226 0.41% Reduce by 75%
Pocketdmr Web Break 4161.35 4.39 136 0.24% Reduce by 75%
4493
Unwind Anode Web Break 2022.17 4.70 334 0.60% Reduce by 60%
Carriar Jumped Pins 3898.64 7.39 266 0.48% Reduce by 80%
Web Break 1044.38 5.35 733 1.32% Reduce by 80%
Cathode Featoon not Holding 3726.33 6.83 123 0.22% Reduce by 100%
Web Break 445.31 5.14 1661 2.99% Reduce by 60%
Comp 1-3 Web Break 3114.46 3.33 450 0.81% Reduce by 60%
3567
In order to incorporate these reductions into the simulations, I developed new ttf
distributions. I calculated the time between consecutive failures of each type listed above.
Then, I multiplied each time between consecutive failures by a factor equal to:
1
1- %Reduction
For example, if a failure type had a 25% reduction, I multiplied all times between failures
by 1/(1-0.25), or, 1.333. All failures were then sequentially listed and assigned a new
cumulative number of machine minutes which would have passed since the beginning of
the test period had the reductions actually taken place. Failures which went beyond the
end of the test period after the times had been increased were eliminated. If a failure type
was reduced by 25%, on the average, 25% of the events should have fallen out. I then
developed new user-defined distributions in the same way that the old ones had been
developed, and maintained the multiplicative factor of 0.7 for the sake of consistency and
ease of comparison to the worst case simulations. This gave the following set of
assumptions for another set of simulation runs:
* All major IBAM systems will fail according to the distributions which
incorporate improvements expected by the IBAM team.
* For the sake of consistency, all ttf distributions will be adjusted by the same
multiplicative factor of 0.7 that was used inthe worst-case simulations.
* All areas of new technology on the IBAM for which there is no data available
will operate at 100% efficiency.
Note that the first two assumptions changed from the worst case scenario. The
simulations based upon this new set of assumptions were considered best case scenarios
by the IBAM team.
The assumption that all new technologies not present on the RBAM will perform
at 100% efficiency warrants explanation. The IBAM team certainly did not believe that
new technologies such as the cambot pick and place heads would be perfect. We gave
consideration to trying to estimate the performance of all new technologies by doing
benchmarking around the corporation with other plants which might be using similar
technologies. Given more time, this might have been the optimum approach. But the time
pressure associated with our decision-making precluded us from pursuing such a time-
consuming strategy.
We considered using manufacturers estimates of reliability of the machines, but
determined that manufacturers estimates were so overly optimistic that they were of little
value. The team firmly believed that the overall IBAM run time performance would
exceed that of the RBAM. The first set of assumptions which forced the simulation to
have run time equal to about 67% of scheduled time, the RBAM's performance for the
period studied, seemed to the team to represent a worst-case scenario. The thinking was
that at worst, the IBAM would have run time performance as good as the RBAM. When
it came to incorporating all of the expected improvements into the distributions for the
best case scenario, the team members realized that the true performance of the IBAM
would most likely be worse than the best case estimate, because new technologies would
introduce some additional downtime which had not been accounted for. We figured that
the true run time performance of the IBAM would probably lie between the worst case
(lower bound) and best case (upper bound) which we had established.
Remember that the purpose behind the simulations was to develop insight into the
effects that different conditions had on dry slurry yield loss. The assumptions that the
IBAM team made were perfectly adequate for the purposes of developing this insight and
using it to assist the decision-making process. The future may show that the IBAM
performs better than our best case estimates or worse than our worst case estimates. But
this is irrelevant. The assumptions and resulting analysis were adequate to the task given
the pressures of program cost and schedule.
The results of the simulations run in support of the IBAM team's efforts to solve
the dry slurry yield loss problem are shown in the following several charts. These are the
charts that the team used in their decision-making efforts. The first 2 charts show worst
case and best case curves for 154; 100; and 70-row webs. The estimated current RBAM
performance is also plotted for reference. Note that the scales of the dependent axes vary.
Dry Slurry Yield Loss vs. Threshold Time, Worst Case
1200%
10.00%
8.00%
6.00%
4.00%
3.26%
2.00%
1.47%
0.00% 8:
20 25
6.00%
154 Rows
4.00%
2.00%
1.68%
0.92%
0.00%
20 25 30
The next 3 charts show the worst case, the best case, and the arithmetic average of the
worst and best case for the different web lengths individually.
Dry Slurry Yield Loss vs Threshold Time, 154 Rows
12.00%
10.00%
8.00%
6.00%
4.00%
3.26%
2.47%
2.00% 1.68%
0.00%
15 20 25 30
8.00%
6.00%
Worst Case
4.00%
2.00% 47%
0.00%
15 20 25 30
3.00%
2.00%
1.00%
0.00%
15 20 25 30
Each point was obtained by averaging the yield losses obtained from five replications with
100 hours of simulated production time per replication. Some apparent irregularities exist
in the shapes of some of the curves, but this can be attributed to the statistical variation
from replication to replication.
Since each point was only based upon five replications, there was some concern
about the statistical significance of the differences between points which were close to
each other. For example, consider the points at the 25 minute threshold for the 100 row
and 70 row curves from figure 16. The 90% confidence intervals derived from the
student's t-distribution are [0.80%, 1.74%] for the 70 row, 25 minute threshold point, and
[1.46%, 2.78%] for the 100 row, 25 minute threshold point. This is somewhat
troublesome, since the confidence intervals overlap. We knew that there would be some
measurable difference between these points, but we were not confident in the magnitude
of the difference. Although it would have been ideal to use more replications per point,
time pressure made this approach impractical.
After the internship was complete, further simulations were run in order to narrow
the confidence intervals around each data point. Plots of curves based upon 20
replications per point are included in Appendix A. The 90% confidence intervals around
the corresponding points on the curves in figure 29 are [1.45%, 1.73%] for the 70 row
curve at 25 minutes, and [2.62%, 3.1%] for the 100 row curve at 25 minutes. These
confidence intervals do not come close to overlapping.
The curves based upon 20 replications per point did not really provide any new
insight. The group was forced to rely on curves based upon five replications per point
because of time pressure, but these curves were adequate to the task, and did not result in
any sub-optimal decisions.
By the time that the curves inthe previous section had been generated, the IBAM
team still had 10 possible line configurations under consideration. The baseline 154-row
machine was carried as one of the 10 options and was evaluated alongside the other new
configurations, but the team felt very confident that the 154-row web was unacceptable.
Estimated dry slurry yield loss was simply too high on a 154-row carrier web, even under
the most optimistic assumptions. But at the same time, the team had worked for months
to develop the baseline configuration. The dry slurry yield loss issue had been overlooked
during the development of the baseline system, but the baseline had many very desirable
features that were injeopardy of being lost in varying degrees depending upon which new
configuration the team chose. Operator accessibility is one such feature. Simplicity of the
overhead slurry delivery system is another.
The team decided that it should attempt to preserve as many of the baseline
system's desirable features as possible when choosing another configuration. In order to
accomplish this, we decided to spend several hours brainstorming the features of the
baseline system which were desirable and worth preserving if possible. Also solicited were
ideas for any new properties which, though they may not be characteristic of the baseline
system, were deemed important and worthy of inclusion in the newly configured system.
The list of brainstormed ideas were then compiled and distributed to all of the team
members inthe form of a ranking sheet. Each team member ranked all of the criteria
individually giving each an importance rating of 1, 3, or 5,with 5 being very important and
1being minimally important. A list of the brainstormed criteria and their cumulative
importance rankings is shown below:
Criteria List
Total
Shortest web maxaimies yield 45
Output confiration issame as arrstREAM we side (18x4) 45
Designed for CBAM expenmion 43
Took yield loss operatio•
i offthe web (Slurry lay dowu uffring 41
Automtic reject incells resulta inless added value and acleaner well-ctioning web (Optimized) 41
Web setimons separated by catenaries (Precim
positioning at placunm) 39
Intermittent motion (Takes advntag ofdwell period) 39
The mandate to lower capital with lower capacity but allow for expanion (Made simple in-lin seal feasible) 39
Curretconfiguration uses a simple sealer (Minimizes capital) 37
System allowed us to use a proven battery assembly method (Lss risk, conervative approach) 35
Coatinuous operator acc along one side ofweb 35
Most efficient for unit operations (buffered) as well a operational and maintenance acess 35
Simple IBAM half cluster xpand easily to full cluster 35
Opens up critical maitenance eas (Saler/Composite area) 35
3 pitch sealer remas in shoter web length 33
System lands itselfto raw marial flow in, and finished goods flow out 33
System utilizes existing anode cell (Minimies capital) 33
Beat utilization of building area (Minimma impact on facility) 33
System onfigration driven by R5 production schedle 33
Fits with vacuum sealer cycle time of 8.1 seconds 33
Few operations on web line (Simplicity of web) 31
Unobstructed line of sigtis mnaximized with rramt system 31
Easy opertato es (Possibility to run with two operators) 30
lhree active slurry are are inthe same neral vicinity (minies labor) 29
Mafacuring poos approach is one compone at a time 29
Single comaposite loop design (Resmts in a shoater web length than the dual loop) 27
Decision to go with in-li cathode (LA. expensive) 27
Desire for straigt line slurry delivery allowed us to use currt system gi (Minimies capital and risk) 27
Double pallet remulted in a sorter web with one web apply station (Minimized capital) 27
Anode is driving the composie cofiguration (Capital, operation, maintenance) 27
IBAM Over head shry delivery system (in-line and simple) 25
Cathode and sealer have a m on operatorside (Minimizes labor and allow for a doubluty platform) 25
Web utiliz existing die module design (Saves money) 25
Cument system configurtion facilitated safety zone and tension zone desip 25
Has pocketing operation in line 25
Cathode over web (Best onesided acom to machine) 23
Existing configuration of cells is astandard approach (Special not yet explored) 23
Have not looked at non-sandard approaches to positioning of comnercial equipume 23
In-line cathode 4-up vl 8-up (Resulted in a shorter web length) 19
In-line cathode uses lessm
floor pace (Le perimter to guard) 19
In-line cathode reslts in shoter cathode web line, but also a longer main web line 19
LA= automation in tub handling (Shortened overall machine length) 19
Sample and eject station in line 19
Allowed for acmeptable aldricl cabinet sdlm 19
Cut loop width drivn by ratr clore and shauy elevator intface 19
Single mtlr hopper(Minaies capital) 17
Moving cutIaelimnisd a catemry (reulb insrbter web) 17
Matmial flow left to riglt (Gives us acess) 13
Since there were so many criteria, the engineer who took responsibility for
compiling the responses decided to place the criteria into four general groups and then
average the scores of each category. He then presented his work to the team and asked
for comments. The team agreed that evaluating all 10 IBAM configurations on each of
the criteria would be excessively time consuming, and welcomed the opportunity to
consolidate the criteria into a manageable number of categories. The categorized criteria
are listed below:
CriteriaList
SHORTEST WEB LINE AND NUMBER OF COMPONENTS BUFFERED (WEIGHT-7)
Shtebaweb maxmizses yield 45
Took yield oss operat ions offthe web (Slmn lay dowm/Buffring) 41
Automutic reject in cells rem in less added value d cleaner well-iAm oning web (Optimized) 41
maintemme e
Mot efficiset for uit operations (bufied), a well a operational mand 35
3 pitch ealer re i shorter web len•h 33
Few operations on web line (Simplicy of web) 31
Inline cathod remits in htedr cathode web line, but also a longrmai web line 19
A~ RGE SCOR"E 3
1
The next step was to assign scores to the grades of high, medium, and low, and to
numerically rank the 10 configurations. The rankings of 41, 21, and I were used for high,
medium, and low, respectively. The average score for each configuration was calculated,
along with a weighted average in which each score was weighted by the irimportance
weighting of the criterion (7, 5, 3, or 1). The numbers 41, 21, and 1 were chosen by the
engineer compiling the scores so that the minimum difference in average score was 1.
This was somewhat arbitrary, but for the sake of clarity, it seemed better to compare
whole numbers. The matrix of scores along with the straight averages and weighted
averages is shown in table 20:
HI, Medium, Lo Rankings OPTION
A B C D E F G H I J
7: Yield 1 1 21 21 41 21 21 1 21 21
Shortest Web Line & # Components Buffered
5: Expandability 41 41 21 21 21 21 41 1 1 41
Expandable to Cluster
3: Operability and MaintaIblity
Accessibility to High tending areas 41 41 21 41 1 21 21 1 1 21
Good mateinal flow 41 41 21 41 21 21 1 1 1 21
Minimum # ofoperators 41 21 41 21 21 21 41 41 21
AVERAGE FOR CATEGORY 41 41 21 41 14.3 21 14.3 14.3 14.3 21
1: Minimum Impact and Risk to Cost
and Schedule
LoCost 41 41 21 21 21 21 21 1 1 21
LoRisk 41 41 21 1 1 21 21 41 1 21
AVERAGE FOR CATEGORY 41 41 21 11 11 21 21 21 1 21
STRAIGHT AVERAGE 31 31 21 24 22 21 24 9 9 26
WEIGHTED AVERAGE 30 30 21 23 23- 21 25 8 10 26
Following the decision to go with the 70-row option, I decided to do some further
analyses to determine the effects of different run times on dry slurry yield loss. The curves
shown in this section were not used by the IBAM team during their decision making
efforts, but were presented to the program manager for reference and to provide further
insight into the dry slurry yield loss issue.
Figures 16 through 20 show dry slurry yield loss as a function of slurry threshold
time for 2 specific IBAM run time scenarios, a best case and a worst case. Since the
future run time performance of the IBAM is very uncertain, I thought it would be useful to
run simulations which showed yield loss for a continuum of run time levels. It was
necessary to hold slurry threshold constant for each family of curves. The multiplicative
factor set at 0.7 on previous simulations was now varied in order to get different run time
performance from replication to replication. Slurry threshold times of 15 minutes and 25
minutes were chosen. The families of curves are shown below. Note that the dependent
axes are scaled differently:
Dry Slurry Yield Loss vs Run Time, Threshold=15 min
25.00% -
20.00% -
I
15.00% -
10.00% -
5.00% -
0.00%- I I I I I
45.00% 50.00% 55.00% 60.00% 65.00% 70.00% 75.00% 80.00% 85.00% 90.00%
14.00%
12.00%-
10.00%
54 rows
8.00%
6.00% 100 rows
4.00% 70 rmws
2.00%
0.00% - 1 1 r I I I I I
45.00% 50.00%I 55.00% 60.00% 65.00% 70.00% 75.00% 80.00% 85.00% 90.00%
The insight one can gain from these curves is that inthe event that run time performance
of the IBAM is very bad, dry slurry yield loss is dramatically worse with longer carrier
web lengths. This analysis further validates the decision to go with the 70-row carrier
web. The 70-row curve is much flatter, meaning that the 70-row web is much more
robust against dry slurry yield loss.
Chapter 6 - Application of Theory of Constraints
The basic idea behind the theory of constraints is that in any process in which the
final output is dependent upon a number of different sub-processes, the speed of the
overall process is defined and limited by the slowest sub-process.(2] This applies to
complex business systems involving suppliers, manufacturers, and distributors, and to
smaller systems at individual plants. This chapter examines how the theory of constraints
was applied to the IBAM in order to recommend improvements in the throughput of the
line.
Recall that the IBAM team decided that inorder to reduce the expected dry slurry
yield loss, three basic approaches were appropriate. These were:
Chapter 5 described in detail the IBAM team's successful efforts to reduce carrier web
length from 154 rows to roughly 70 rows. I say roughly 70 rows, because final design
iterations may necessitate a small increase inlength, or may allow a small reduction in
length. Retesting of the slurry threshold time was being addressed by a team at the plant.
The remaining step, to increase the speed of the IBAM, did not receive as much of the
team's attention. This was because the team did not believe that dramatic speed increases
were attainable.
In crude terms, the battery assembly process on the IBAM could be diagrammed
as follows:
Anode -- Composite 7
Cathode Sealer
2.7 2.7 2.7 2.7
Figure 24: IBAM Process Flow
Essentially, there are four major processes which all process one battery row (4 batteries)
in 2.7 seconds. In a layout such as this in which every process is coupled to the next, and
there is no possibility of building inventory between processes, it is essential that each
process be executed at the same speed as the next. The system as diagrammed above has
no bottleneck. All processes would have to have the same increase in speed in order to
achieve greater throughput.
The theory of constraints and the concept of bottlenecks applies when the question
of how much speed increase is possible is posed. Although the maximum speed of each of
the 4 sub-processes shown above is, strictly speaking, unknown, the team felt nearly
certain that the composite process had the least opportunity for speed increases, and was
therefore, the system's bottleneck. One engineer offered the idea of putting another
double-headed cambot into the composite loop in order to give it potential for greater
speed increases. The initial response to the idea was rejection. Other team members
brought up concerns about extra conveyors which would be required to feed the second
cambot, and about the capacity of the other equipment in the composite loop, so the idea
temporarily died. The next section describes how a closer look along with refinements to
the idea introduced a whole new option for the program manager to consider.
The diagram shows that the various processes inthe composite conveyor loop all process
two composite rows in 1.2 seconds, for a total of 0.6 seconds per row. The cambot-web
interface encompasses all motions of the conveyor's walking beam, the double headed
cambot, and the indexing motion of the carrier web. The relationship shown above the
block diagram shows how for each pair of battery stacks on the carrier web, the cambot
goes through three consecutive 1.4 second pick and place operations, followed by a 1.2
second double index of the carrier web. This means that it takes 5.4 seconds to process
six composite rows, or conceptually, 0.9 seconds per composite row. The fact that the
cambot-web interface requires 1 composite row every 0.9 seconds, coupled with the fact
that the composite loop can provide I row every 0.6 seconds, shows that there is excess
capacity in the composite loop.
To go after some of this excess capacity, it would be necessary to increase the
speed of the cambot-web interface. There are a number of reasons why this is difficult.
First is the fact that initial testing has shown that trying to speed up the actual 1.4 seconds
worth of cambot motions is not practical. The arms are already swinging fast, and the
heads are moving fast as well. Speed increases would force the cambots to operate
beyond their intended parameters. The remaining opportunity lies in reducing the 1.2
second double index time of the carrier web. This could be accomplished by either
speeding up the pin wheel drives, or overlapping the first 1.4 second cambot motion with
the web index. Overlapping the two motions completely is impossible due to the
requirement for dwell times at various points throughout the cambot's cycle. A simple
speed increase, while feasible, may not be particularly desirable. The web is already
indexing twice as far in less time than it is at both the anode and cathode application
points. But certainly some overall speed improvements could be achieved by a
combination of index time reduction and overlapping of motions. The actual speed
increase possible will be determined by later tests, but the initial feeling is that a total of
2.4 seconds per battery row might be possible.
The next diagram is also a slide which I presented to the IBAM team to show how
the timing would change if another double headed cambot was placed inthe composite
loop, as was once suggested and shortly thereafter dismissed as impractical. This slide
shows a hypothetical case, and does not represent what is actually achievable:
1.4+(1.4+1.2) 4.0 sec 0.667 see 2.0 sec
6 comp rows 6 comp rows comp row battery row
1.4+(1.4+1.2)=4.0 (1.4+1.2)=2.6
The idea is that the first cambot could place the first 2 pairs of composite rows onto the
waiting battery stack. The carrier web would then complete its 1.2 second double index.
The third pair of composite rows could then be placed onto the battery stack by the
second cambot, a few feet downstream on the carrier web. According to the diagram
above, this would mean that the overall time would be reduced to 2.0 seconds per battery
row. While the first cambot placed its second composite pair, the second cambot would
simultaneously place the third onto battery stacks downstream. The events which occur
simultaneously are shown in parentheses. Note that the 1.2 second index of the
continuous carrier web is actually a single event, even though the diagram makes it look
like two separate indexing motions take place simultaneously. The relationship at the top
of the diagram shows that existing capacity in the composite loop could support a second
cambot. The cambots taken together would effectively require one composite row every
0.667 seconds, and the loop is capable of providing one every 0.6 seconds.
The final slide presented to the IBAM team was intended to show the practical
effect of the addition of the second cambot. Again, the situation is hypothetical and does
not represent what is actually attainable by the IBAM:
6.0+2.1
=2.7
1.4+1.3= 2.7 1.4+1.3=2.7 3
Anode Composite Cathode Sealer
2.0 (?) 2.0(?) 2.0 (?)
Cut Index Time?
Overlapping motions?
The slide shows that if 2.0 seconds/row were attainable, this would represent a 35%
increase in throughput over the 2.7 seconds/row case. The diagram also highlights the
need to increase the speed of the other major IBAM sub-systems to 2.0 seconds/row in
order to actually achieve the throughput gains. The slide suggests that cutting index times
and overlapping of some motions may be ways to achieve this.
The question of whether or not capacity existed in the composite loop to support a
second cambot was answered by the slides, but the technical problems and increased
complexity associated with the possible need for a conveyor system to support the second
cambot had not yet been addressed. The next section describes how computer simulation
was used to prove a concept, and what was actually attainable in the case of the IBAM.
6.3 The Role of Computer Simulation
I was very much in favor of adding the second cambot to the composite loop in
order to achieve the gains in throughput. Before meeting with the program manager to
discuss the issue, I compiled the following list of what I saw as the benefits and drawbacks
of the idea:
Benefits
1 CBAM refers to the possible cluster-style expansion of the IBAM to a second line. The decision to
expand or not will come sometime in the future.
Drawbacks
It is important to remember that whether or not the other sub-systems of the IBAM could
achieve 2.1 seconds/row was still an unknown. The engineers responsible for each sub-
system predicted that 2.1 seconds/row was attainable, but their predictions, particularly in
the cathode area, were less than certain.
As stated in the list of drawbacks, adding another cambot to the composite loop
would necessitate increasing the carrier web length by an estimated 18 rows. If 2.1
seconds/row was possible, then the increased speed would more than offset the extra
carrier web length, and dry slurry yield loss would improve. If, however, 2.1 was not
attainable by the other IBAM subsystems, dry slurry yield loss might worsen, depending
on what speed was actually attainable. Following the weeks' worth of efforts to reduce
web length and minimize dry slurry yield loss, the program manager was disinclined to
pursue a strategy that due to uncertainty, might actually increase dry slurry yield loss,
albeit by a very small amount.
This reluctance, coupled with the fact that the program manager was uncertain
whether he could get roughly $500K extra, caused him to decide against immediate
implementation of the second cambot. Rather, he would keep the idea in his hip pocket as
an option for future expansion. Once the IBAM produced a few million batteries, he
would know for certain what speeds were attainable by the major IBAM sub-systems.
The major drawback associated with this decision is that it will cost much more to
add the second cambot after production starts than it would have cost had it been
implemented in the initial construction of the IBAM. Lost production during rework and
actual rework costs will be much higher than $500K.
The decision to hold off on implementation of the second cambot is certainly a
defensible decision. In fact, testing of the vacuum sealer which occurred just prior to the
90
publication of this thesis indicated that the sealer may only be able to operate at 2.4
seconds/row.
The actual bottleneck of the system cannot be conclusively determined until
thorough testing is done on the actual production equipment. The addition of a second
cambot in the composite loop would reduce its cycle time to 2.1 seconds/row. If another
system such as the sealer were not capable of achieving 2.1 seconds/row, it would be the
new bottleneck, and should be the focus of extensive improvement efforts. The financial
benefit associated with increases inIBAM throughput are tremendous, and should be
weighed carefiully against the cost of implementation.
Chapter 7 - Effects of Decoupling the CarrierWeb
Table 16 shows that 36.7% of all downtime on the RBAM web side is caused by
problems with the carrier web. This fact suggests that the carrier web system (also
referred to as the web transport system) is a reasonable area to focus improvement efforts
on and, if necessary and practical, to change.
However, the web transport system developed for the IBAM is very similar to the
web transport system used on the RBAM. Major exceptions are that the IBAM's web
transport system will make use of intermittent motion instead of continuous motion, and
that the IBAM will operate at a dramatically reduced speed. Additionally, the IBAM's
web transport system incorporates many of the desirable features of the RBAM card side's
drive system, which is less problematic than the corresponding web side system. It is
hoped that these factors will contribute to improved IBAM run time performance.
Still, it is worthwhile to consider the effects that alternative web transport designs
could have on dry slurry yield loss and overall IBAM efficiency. Consider a system in
which the long, continuous paper carrier web is replaced by a conveyor system carrying
pallets, each with a single 1x4 battery row. Just like on the current IBAM, components
would be placed on top of one another until a complete battery stack was built. Then, the
battery stack would be vacuum sealed. The line would be nearly the same as the current
IBAM, with the obvious exception that the backbone of the line would be the conveyor
system carrying pallets instead of the current paper carrier web transport system. This
concept immediately gives rise to several problems:
* The output would be incompatible with the existing finishing equipment, which
requires 18x4 battery strips.
* It is unclear how functions such as laminating and vacuum sealing could be
accomplished when the battery stack rests on a pallet.
* The requirement for a return conveyor could introduce accessibility problems.
These and many other problems would need to be resolved. IBAM cost and schedule
constraints currently preclude this concept from receiving any consideration, but the
analysis which follows suggests that concepts such as this one are worthy of consideration
during any future battery line projects.
The decoupled conveyor system described above would have some buffering
between the different points at which components were placed onto the pallets, but more
importantly, it would allow downstream processes to continue to run during upstream
failures. The continuous paper carrier web prevents this from occurring on the current
IBAM. Intuitively, one would expect dry slurry yield loss to be reduced, because only a
portion of the work in process would be discarded during long downtimes, instead of all
of the work in process. Of course, if the most downstream process is the cause of the
long downtime, then all of the work in process would still have to be discarded. The
vacuum sealing operation is the furthest downstream operation of interest, since slurry is
no longer exposed after being sealed. On the RBAM, the sealer accounts for about 25%
of the downtime. Intuitively, one would expect some dry slurry yield loss improvements
during the other 75% of the downtime.
The chart that follows shows the results of simulations run using worst case ttf and
ttr profiles on a system with the decoupled pallet conveyor system described above.
Dry Slurry Yield Loss vs. Threshold Time for a Decoupled Web
7 0f•L
I. 6 -,
6.39%
6.00% -
154 Rows
5.00%-
4.00%-
48%
3.00% .1 00 Rsws
2.00% -
1.76%
1.00%- 0.98%
0.58%
0.00%
15 20 25 30
In order to isolate the effects of decoupling, several important assumptions were made
before developing these curves:
1. The decoupled conveyor system fails according to the same profiles as the current
RBAM carrier web.
2. Failures are evenly distributed across the length of the conveyor system. (Some
failures occur upstream, and some occur downstream.)
3. The anode apply, composite apply, cathode apply, and vacuum sealer have speeds of
2.1 sec/row, 2.3 sec/row, 2.5 sec/row, and 2.7 sec/row, respectively, in order to
ensure that the buffers fill up after failures. Note that the overall speed of the system is
still 2.7 sec/row.
Figure 29 shows the curves for a coupled paper carrier web which serve as a basis for
comparison to figure 28. Note that dry slurry yield loss is cut roughly in half when the
decoupled system is used.
Figure 28 was generated in order to demonstrate that decoupling alone improves
dry slurry yield loss. But one very important point has thus far been ignored. In the above
scenario, times were adjusted to ensure that buffers filled up after failures. Actually, this is
not desirable. By simply controlling the number of pallets between the point at which
slurry is initially applied until it is vacuum sealed, the amount of work in process can be
dramatically reduced.
Say, for example, that the conveyor ran at 18 inches/second, and that
dimensionally, it was the same as the 154-row paper carrier web shown in figure 7.
Smooth steady-state operation could be attained by using fewer than 20 pallets. The
number of pallets on the return conveyor would be irrelevant. Only the number of pallets
between the anode apply station and the vacuum sealer would need to be controlled.
By controlling the number of pallets, it would be possible to effectively have a
decoupled 20-row system. A simulation was run to determine a single data point for a 20-
row decoupled system with a dry slurry threshold of 15 minutes. The expected dry slurry
yield loss was only 0.5%. This is even better than the corresponding RBAM benchmark
of 0.92%.
7.3 Comments
This chapter was included inorder to provide some useful ideas for expansion
projects in the future. The most important issue throughout the examination of the dry
slurry yield loss problem has been the length of the carrier web, which corresponds to the
amount of work in process. By changing the carrier web concept, the amount of work in
process could be reduced to levels unattainable with the current concept. It is also worth
repeating that the paper carrier web was the cause of 36.7% of the RBAM downtime
during the period tested. It may very well be that a new concept would be much more
reliable and reduce downtime substantially. The future will tell whether the IBAM web
transport system is problematic or not. If it is problematic, then concepts such as the one
presented in this chapter may be useful.
Chapter 8 - Final Conclusions
8.2 Conclusions
References
[1] Hogg, Robert V. and Ledolter, Johannes. Applied Statistics for Engineers and
Physical Scientists. New York, NY: Macmillan Publishing Company, 1992.
[2] Goldratt, Eliyahu M. and Cox, Jeff. The Goal. Croton-on-Hudson, NY: North
River Press, 1992.
14.00%
12.00%
154 Rows
10.00%
8.00%
6.00%
4.00%
3.25%
2.00% 1.90%
0.00% 0.35%
20 25
Threshold Time (minutes)
6.00%
5.00%
--* 2.7
4.00% -a-2.4
-.-- 2.1
3.00%
2.00%
1.00%
0.00%
20 25
Threshold Time (minutes)
3.00% !.88%
2.50% -- 2.7
-- 2.4
2.00%
--- 2.1
1.50%
1.00%
0.50%
0.00%
20 25 30
Threshold Time (minutes)
6.00%
3.00%
2.00%
1.47%
1.00%0 1.11%
0.75%
0.00%
20 25
Threshold Time (minutes)
4.50%
4.00%
3.50%
3.00%
2.50%
2.00%
1.50%A
1.26%
1.00% 0.97%
0.50%0 0.67%
0.00%
20 25 30
Threshold Time (minutes)
3.50%
3.00%
2.50%
2.00%
1.50%
1.00% 1.11%
0.86%
0.50% 0.61%
0.00%
20 25
Threshold Time (minutes)
Minutes Minutes
cathode ttf 6.76 5 cathode~ttr 2.93 1
10.74 10 11.52 2
15.31 15 21.68 3
20.48 22 24.22 4
25.05 33 53.13 5
30.02 45 55.47 6
35.19 64 58.2 7
39.96 80 61.13 8
44.73 102 62.11 9
49.7 128 76.58 10
54.67 152 78.52 12
59.64 180 84.96 15
64.61 227 88.28 20
69.58 270 92.97 35
74.75 343 97.66 120
79.52 481 100 350
84.49 578
89.48 757
94.43 1036
99.4 2194
100 3050
Minutes Minutes
sealer tf 5.61 4 sealeB_ttr 0.93 1
10.08 5 5.69 2
15.88 10 14.53 3
21.3 15 18.35 4
25.68 20 48.64 5
30.06 27 49.53 6
35.23 35 53.44 7
39.96 48 58.92 8
45.05 58 58.03 9
50.04 70 73.58 10
56.09 85 78.13 12
60.3 100 84.45 15
64.94 116 85.39 18
69.94 137 90.23 22
74.93 164 95.58 40
79.93 199 96.69 50
84.93 242 97.79 60
89.92 318 98.22 70
94.92 445 98.73 80
95.97 500 99.24 90
98.42 700 99.75 120
99.47 900 99.92 150
100 1208 100 185
102
% Minutes Minutes
web_ttf 8.76 5 web..ttr 0.93 1
11. 01 8 5.68 2
15. 51 11 11.5 3
20. 37 15 14.09 4
25.24 22 39.11 5
30. 11 27 41.62 6
35. 36 37 46.08 7
4 0.3 45 49.68 8
46. 37 55 50.83 9
50. 19 63 69.3 10
55.21 73 71.82 12
60. 37 83 81.67 15
65. 47 97 86.7 20
70.64 115 91.95 30
75. 36 146 95.97 50
8C3.3 179 98.63 100
85. 32 214 99.35 150
90. 34 266 99.57 200
95.36 380 99.64 250
98. 65 600 99.78 300
91).7 900 99.93 350
99. 93 1200 100 390
100 1602
% Minutes Minutes
unwind ttr 1.29 1 carrier.unwind ttf 4.88 49
13. 45 2 9.76 72
32. 28 3 15.12 87
40. 47 4 19.51 146
75. 58 5 24.39 158
77. 89 6 29.27 205
81. 75 7 34.15 250
5.5 8 39.02 313
886.32 9 43.9 393
93. 57 10 48.78 433
96. 73 15 53.66 494
98. 95 30 58.54 559
1.3 45 63.41 680
99.
99. 53 60 68.29 829
99.65 70 73.17 923
99. 77 100 78.05 1093
99.88 120 82.93 1311
100 170 87.8 1539
92.68 1725
97.56 2353
100 4949
103
Minutes % Minutes
anode unwind ttf 10.87 136 composite.unwindttf
21.74 377 10 219
32.61 481 20 490
43.48 751 30 771
54.35 977 40 1090
65.22 1502 50 1402
78.09 2174 60 1934
88.96 3431 70 2442
100 7123 80 3321
90 4531
100 14967
% Minutes % Minutes
cathode.unwindmtt mother_hopperttf
4.99 19 19.44 17
9.97 36 38.89 601
15.49 47 58.33 1535
19.95 61 77.78 4950
25.2 73 100 34519
29.92 103
34.91 117 % Minutes
39.9 144 motherhopperttr
44.88 179 22.22 10
49.87 217 52.78 20
54.86 252 58.33 25
59.84 320 77.78 50
64.83 373 100 150
69.82 426
74.8 522
79.79 644
84.78 750
89.76 902
94.75 1208
99.74 2125
100 2187
104
% Minutes % Minutes
baby_hopper_ttf baby_hopper_ttr
7.27 2.2
12.27 8.79
17.5 18.68
22.95 86 20.6 4
28.18 133 36.54 5
33.41 188 37.36 6
38.64 247 39.84 7
43.86 313 42.58 8
49.09 445 43.41 9
54.32 585 57.42 10
59.55 674 68.41 15
64.77 858 81.04 20
70 1105 86.28 25
75.23 1386 90.38 30
80.45 1717 92.58 35
85.68 2282 93.96 40
90.91 2858 96.98 45
96.14 3759 98.63 60
99.77 5189 99.18 75
100 11225 99.73 120
100 164
105
Minutes Minutes
sealer ttf 6.33 5 searyrtt 0.93 1
9.49 12 5.89 2
14.87 25 14.53 3
19.3 45 18.35 4
23.73 85 46.64 5
28.8 95 49.53 6
33.54 129 53.44 7
37.97 154 56.92 8
42.72 198 58.03 9
47.47 252 73.58 10
52.22 299 78.13 12
58.96 355 84.45 15
61.71 392 85.39 18
66.46 483 90.23 22
71.2 581 95.58 40
75.95 683 96.69 50
80.7 738 97.79 60
85.44 955 98.22 70
90.51 1138 98.73 80
94.94 1482 99.24 90
97.47 1985 99.75 120
100 3287 99.92 150
100 185
106
% Minutes % Minutes
unwind ttr 1.29 1 web ttf 5.22 3.3
13.45 2 10.09 9
32.28 3 15.2 15
40.47 4 20.19 23
75.56 5 25.64 30
77.89 6 30.05 39
81.75 7 34.92 51
85.5 8 39.91 83
86.32 9 44.9 76
93.57 10 49.88 92
96.73 15 54.87 113
98.95 30 59.86 138
99.3 45 64.85 161
99.53 60 69.84 199
99.65 70 74.83 230
99.77 100 79.81 284
99.88 120 84.8 340
100 170 89.91 425
94.78 589
99.77 1013
100 1602
% Minutes % Minutes
web•ttr 0.93 1 carrier unwindttf 11.59 254
5.68 2 27.54 739
11.5 3 33.33 812
14.09 4 57.97 1870
39.11 5 63.77 2271
41.62 6 75.38 2859
46.08 7 76.81 2861
49.68 8 95.65 6041
50.83 9 97.1 7415
69.3 10 100 8870
71.82 12
81.87 15 % Minutes
86.7 20 anode_unwind ttf 8.93 152.5
91.95 30 17.86 387
95.97 50 26.79 575
98.63 100 35.71 889
99.35 150 44.64 1202
99.57 200 53.57 1757
99.64 250 62.5 2429
99.78 300 71.43 3140
99.93 350 80.36 3942
100 390 89.29 5709
100 14891
107
% Minutes Minutes
cathode.unwind ttf 5.82 50 composite unwind tf 10 219
10.58 90 20 490
15.87 112 30 771
21.16 152 40 1090
26.46 175 50 1402
31.75 211 60 1934
37.04 260 70 2442
42.33 315 80 3321
47.62 367 90 4531
52.91 467 100 14967
58.2 597
64.02 752 Minutes
68.78 845 motherhopperttf 19.44 17
74.07 987 38.89 601
79.37 1222 58.33 1535
84.66 1450 77.78 4950
89.95 1795 100 34519
95.24 2430
100 4483 Minutes
mother hopper_ttr 22.22 10
52.78 20
58.33 25
77.78 50
100 150
% Minutes % Minutes
babyhopperttf 7.27 5 babyhopper~tr 2.2 1
12.27 20 8.79 2
17.5 45 18.68 3
22.95 86 20.6 4
28.18 133 36.54 5
33.41 188 37.38 6
38.64 247 39.84 7
43.88 313 42.58 8
49.09 445 43.41 9
54.32 585 57.42 10
59.55 674 68.41 15
64.77 856 81.04 20
70 1105 86.26 25
75.23 1386 90.38 30
80.45 1717 92.58 35
85.68 2282 93.96 40
90.91 2858 96.98 45
96.14 3759 98.63 60
99.77 5189 99.18 75
100 11225 99.73 120
100 164