At the time of this project there were four main board assembly lines, each consisting of automated surface mount assembly plus manual through-hole assembly. The surface-mount assembly segments have tightly coupled fully automated machines. Qualcomm performs approximately 20 setups per week on the four assembly lines, most of which run two shifts per day, five days a week. Each setup takes approximately two hours per assembly line. During this time, all of the SMT assembly equipment is idle.

     The PCB surface mount assembly process starts with bare printed circuit boards. These then go through a series steps, each at a dedicated machine or station, as illustrated in Figure 1. Because the surface mount portion of the process is the most capital intensive and has the longest setups, it was viewed as the key for setup reduction.

Figure 1 Assembly Process for Surface Mount Printed Circuit Board

     Stencil printing is the process of applying solder paste to the PCB through a thin metal-foil stencil. The solder paste is a mixture composed of fine solder particles, solder flux, and adhesives.

     Once the solder paste is applied, the board is populated with surface mount components. A typical board contains 1,000 components of about 100 types. At Qualcomm, two "chip shooters", high speed automatic placement machines, are used to populate the PCBs with small components such as resistors, capacitors, inductors, and small integrated circuits. PCB assemblies generally have hundreds of these components, which is why they are placed by a high speed machine. The Fuji CP6 machines used can place up to 28,000 parts per hour, or 8 per second. Then a single placement machine, which is slower, places larger or fine pitch components such as Quad Flat Packs and plastic leaded chip carriers.

     Once the boards are populated with components, they are visually inspected, then conveyed into a reflow oven where the boards are heated to reflow the solder.

Setting up the line

     A normal setup of the surface mount assembly process requires preparing all the machines in Figure 1 and the conveyors between them. Conveyors are adjusted to accommodate the width of the next PCB to be produced. The stencil printer is cleaned and fitted with the stencil for the next PCB. Component feeders are removed from the placement machines and replaced with the new component feeders. The solder reflow oven is reprogrammed with the temperature settings and duration for the next PCB to be assembled.

     Of all of these setup tasks, the preparation of the component feeders and their placement machines (chip shooters) is the most time consuming. It can take from one to four hours depending on the number of components to be deposited on the board. The whole surface mount line is down while the three placement machines are being set up. Setting up the placement machines requires loading each machine with the correct components. These components are loaded on special feeders, and the overall setup breaks down into two basic processes: setting up the feeders, and loading the feeders onto the machines.

     The first process is the gathering of the correct feeders and component reels, and loading those reels onto the feeders. The result is a rack of feeders ready to be loaded onto a placement machine for a scheduled build. In order to reduce machine downtime, gathering and loading of feeders and reels can done off-line from the placement machine while a different build is in progress.

     In the second process, the prepared feeders are loaded onto the machine in the correct order and verified. This process alone can take an hour or more as a result of the sheer number of feeders to load and verify. PCB setups can require as many as 140 feeders per machine, although the average is about 60.

     Verifying each feeder and its loaded component is time consuming and prone to errors. Components are sometimes switched in the device slots on the machine or read incorrectly. In the event that an incorrect part accidentally gets placed onto the PCBs by the machine, the error generally goes unnoticed until the board is tested and the defect is finally discovered. At this point a hundred or more boards may have been produced with the wrong part; all requiring rework.

Raw materials and fixtures: reels and feeders

     Most of the surface mount components used at Qualcomm are packaged on component reels. These reels look very similar to movie reels. They come in two standard sizes, 7 and 13 inches in diameter, and several widths for different component sizes. The components are nested between two pieces of tape and wound onto the reel. One reel may contain as many as 10,000 components, which is the reason they are so popular for use with high speed placement machines.

     Before a component reel can be used on a placement machine, it must be loaded onto a tape-and- reel feeder, which is a mechanical dispensing device that dispenses a single part each time it is actuated by the placement machine. The feeder consists of a bed to hold the component reel, a tape take-up sprocket, and a lever arm which causes the one tape to be pulled away from the component as it is fed forward on actuation. During operation, the placement machine actuates the lever of the feeder carrying the required component each time it needs a component. The benefit of using tape-and-reel feeders is that large quantities of boards can be assembled very quickly with minimal downtime needed for component replenishment during a batch.

     The high-speed placement machines can hold up to 140 feeders each, on two setup tables. Each feeder may contain a different component or a duplicate if used in large quantities. Managing the reels and feeders during setups is a detail-intensive process. Feeders are also expensive, costing about $900 for each of the 3,500 feeders in the plant. Maintaining the feeders is also an issue in itself, as they are subject to mechanical drift, wear, and damage.

Feeder setup

     Placement machine setups occur in two discrete processes that may be separated in time by 1 to 3 days. The first process is feeder setup. It involves locating the component reels needed for a production run and loading them onto feeders. This process occurs off-line from the machine and is considered an external process in SMED terminology. The second process is placement machine setup and is performed at the machines. In SMED terminology, this is an internal setup process. Many hours were spent observing and interviewing the feeder setup and placement machine operators in order to map out the tasks performed in each process.

     The goal of the feeder setups is to find the correct components for a lot, mount them on the correct feeders, and pre-position them on racks. These setups can take up to 14 hours for a single placement machine, but this time is not fully predictable and feeder setup is therefore usually done several days in advance. This obviously reduces factory scheduling flexibility.

Figure 2 Feeder Setup Process Flow Diagram

     Figure 2 depicts the process flow. In Step 1, the setup operator reconciles the parts list for the PCB being assembled, with the setup sheets for the individual placement machines. Once these are reconciled, the operator begins physically locating and collecting the component part reels needed for the setup. Each reel contains one component type that is designated by a part number. Some reels may already be loaded on feeders while others are not. A reel not already on a feeder must be loaded on a feeder of the correct size. There are approximately 30 possible feeder configurations to select from. Component reels may be found: in the raw reel inventory, already loaded on a feeder in the feeder setup area, or on one of the four assembly lines where the feeders may be in use in a current build or part of a build that has just ended. Once the reel is located and put on a feeder, if necessary, it is labeled and placed on a rack. Finally, the feeder is labeled with the component part number and the machine device location.

     This process of looking for the reel, loading it on a feeder and labeling the feeder continues until all of the component reels needed for the setup have been procured. Feeder setups can take from 1 to 14 hours per machine. From the timing studies performed and from operator interviews, we determined that approximately 70% of the labor time spent on feeder setups is for locating parts, steps 2, 3, and 4. Locating parts, on component reels, is time consuming because there are few duplicate component reels and these reels are spread out over a 40,000 square foot area. Usually, there is only a single reel in the factory containing a particular component part number. Moreover, the component part number is 14 digits in length which adds to the difficulty of locating the reel. As a result, looking for components becomes a task akin to locating needles in a field - it sometimes take hours.

     The completed feeders are placed on a movable rack which will hold all of the component feeders prepared during the setup. Steps 1 through 7 are performed for each component reel in the setup. Once this off-line setup is complete, the racks (one per machine) are grouped together and labeled with the order number. When the scheduled build date and time arrives, the line operators will collect the racks and take them to their respective machines. It is at this point that the on-line setup process commences.

The placement machine setup process

     The on-line placement machine setup process occurs at the placement machines. This process is usually carried out by two or more people who perform the setups for all three machines in the assembly line. The setups for each machine are performed in parallel whenever possible to reduce down time, which ranges from one to four hours per setup, and averages two hours. Figure 3 and Figure 4 show the placement machine setup process flow diagram.

Figure 3 Placement Machine Setup Process Flow Diagram (part 1)

Figure 4 Placement Machine Setup Process Flow Diagram (part 2)

     The operator begins the process in Figure 3 after he or she has obtained the rack of feeders (loaded with component reels). This is usually done right before the machine finishes the current production run and is stopped for setup of the next run. The setup sheet comes with the rack of feeders and contains all the information needed to load the device table correctly.

     In steps 2 through 5, the operator selects the first feeder from the rack and checks that the component in the feeder is the correct component, that the feeder is the correct size, and that the device location called out for this component is correct. The operator then places the feeder in the designated device location on the placement machine (step 12).

     Once finished with each feeder, the operator returns to step 2 to repeat this process until all feeders have been loaded on the machine. Once all feeders have been loaded, the setup is rechecked to ensure that each component was placed in the proper device location on the machine (step 13). This task is called setup verification and consumes approximately 50% of the total placement machine setup time. Generally the buddy system is employed in this task. One operator will read the part number and its device location while another operator verifies the numbers with the setup sheet. For a 50 feeder setup, this task takes about 30 minutes.

     Setup verification is performed to ensure that the setup is correct, since errors are costly. If the wrong component is placed and the vision system of the machine does not detect the error, the PCBs will be populated with the wrong part. Placing unwanted parts results in costly rework and scrap, since the PCBs must be retrofitted with the correct part once the error is discovered. Usually, errors of this nature are not discovered until they are tested, which results in a need to rework all the boards in the lot. Because of this high penalty for incorrect setups, much time is spent verifying and re-verifying components before the production run.

Why do setups take so long?

     Both the off-line setup (locating and matching reels and feeders) and on-line setup (mounting feeders on the machines) take many hours. Since on average each line is set up about once a shift, and setups average three or more hours for on-line downtime plus four hours of off-line labor, major resources go into the setups.

     Both parts of the setup are driven mainly by the number of different components, and therefore number of feeders, needed for the board being assembled. Thus the setup durations are a function of board complexity. However the situation is exacerbated by the overall complexity of the factory, which increases the amount of effort needed simply to "keep track of" all the elements of setups (reels, feeders, racks, etc.).

     Table 1 summarizes the elements which contribute to setup complexity. The factory has simplified the situation by concentrating on only two types of placement machines, but the goal of producing both prototypes and production boards for a number of products has kept everything else complicated.

     Cause of complexity Number of types Absolute number in factory Reels/part types 3000 4000 (approx.) Feeders 29 (excl. tray feeders) 3500 Feeders per setup 50 to 300+, 180 avg. not applicable Placement machines 2 (high speed+ fine pitch) 12 Boards >300        

Applying SMED to PCB Assembly Setups

     None of the current literature has applied the single-minute exchange of die (SMED) methods to the setup of placement machines, or indeed to any electronic assembly. In this research, SMED concepts were applied to placement machine setups at Qualcomm, with promising results.

     The SMED methodology consists of three general phases of analysis. In the first phase, a distinction is made between internal and external setup tasks. Internal setup operations are those that must be performed when the machine is stopped. These operations occur on-line to the machine. External operations are those that can be performed while the machine is in operation. It is more efficient to perform these tasks off-line from the machine. Once the operations are classified as either external and internal, the external operations can be moved off-line to reduce machine downtime.

     For example, when analyzing the setups for the large body molding presses at Mazda, Shingo discovered that the presses were shut down while the mounting bolts for the new die were being located. This task was considered part of the internal setup process until Shingo moved it off-line. As a result of this change and other changes, Shingo reduced the on-line setup time for this machine by 50%.

     The second phase in SMED analysis is to convert as many internal setup operations as possible to external setup operations. In the third phase, all aspects of the setup, both internal and external, are streamlined in order to make them more efficient. The rationale is that efficiency is important whether it occurs internally or externally to the machine. Internal efficiency results in faster throughput, labor savings, and better utilization of machine capacity. External efficiency does not improve downtime, but results in better utilization of labor and the resulting ability to perform additional tasks.

     Figure 3 and Figure 4 outline the setup process at the start of this research. The first step in the SMED methodology is to classify each setup task as either internal or external. On examination of tasks 3-5 in Figure 3 and task 11 in Figure 4, it is apparent that none of them require interaction with the machine. These tasks could be performed off-line from the placement machine and should, therefore, be handled as external tasks. Moreover, most of these tasks were already being performed during the external feeder setup process. Requiring the machine operator to verify these attributes is a duplication of effort. These tasks were removed from the placement setup process and shifted back upstream to the external feeder setup process. Shifting these tasks reduced the number and time of tasks performed on-line while the process was down. Moreover, errors detected at a machine increased the downtime of that machine. Shifting tasks 3-5 and 11 to the feeder setup process also shifted the error handling tasks (6-10) off-line accordingly.

     The remaining tasks are inherently internal operations since the machine must be stopped while they are performed. Loading the feeder on the device table and locking it into place (step 12) is an internal operation. Verifying that the setup is correct once all the feeders have been loaded onto the device table (step 13) is an internal operation since it involves verification of components as actually loaded on the machine.

     Figure 5 shows the new internal process. This process represents a 50% reduction in the number of tasks performed on-line to the machine, assuming that there are no errors, and a bigger reduction if there are errors, since error handling is now off-line.

Figure 5 The Modified SMT On-line Setup Process

     The third phase in the SMED process is to streamline the remaining internal tasks. Once all the of the external operations were moved off-line, only four internal tasks remained. These are: obtaining the setup sheet, selecting a feeder from a rack, loading the feeder onto the device table, and verifying that the feeder was placed in the proper location. Since these tasks were already being done in a fairly efficient manner, there did not appear to be any large improvements to be gained without adding new tools or methods.

     Of the four remaining internal tasks, the task taking the greatest amount of time was the verification task. This task (step 13, Figure 4 and step 4 in Figure 5) requires the verification of all feeders and their components to the setup sheet after they have been placed on the machine. For this task, the operator (or, sometimes two operators) verifies the component part number for each device location against the setup sheet. This is time consuming and prone to errors, due to the length of the part numbers (14 digits) and the large number of components to verify. Errors result from part numbers being misread and from feeders being switched. Consequently, this task was the focus of further setup time reduction efforts. This was done by creating a new system for feeder management, discussed in Section 4. This system also has major benefits for external setup.

Hot swapping

     The ideal limit of SMED is to have instantaneous setups. By reprogramming the insertion machines to allow "hot swapping" of feeders, we have approached this ideal in some cases.

     This concept takes advantage of the design of the CP6 high-speed placement machines. Each CP6 machine has two device tables, each holding up to 70 feeders. Since there are two CP6 machines per line, this gives a maximum of 280 different feeders (plus about 30 more on the fine pitch machines). However, many PCB designs require 140 or fewer components. For those products, the process engineers who program the machines can reprogram them to leave one device table entirely empty for that product.

     The CP6 machines are designed so that if a device table is not being used, the operator can safely set it up even if the machine is running. In this way, the next job can be fully set up on one device table while the current job is running on the other device table. Thus the placement setup occurs entirely off-line. When the old job completes, the other placement table and job are "hot swapped" by software.

     There are a number of qualifications which prevent this technique from reaching 100% effectiveness.

• The currently running product must have fewer than 140 components for the CP6 machines.
• So must the next job.
• Both products must have been reprogrammed by process engineers to have all feeders on a single table.
• The run length of the first job must be long enough to complete the setup of the device table for the next one. Thus, it is still important to reduce the off-line setup time of the CP6 machines.
• The stencil print, conveyor, and fine pitch machines cannot be hot swapped, which prevents instantaneous setup even in the ideal case.

     Clearly, hot swapping further reduces line downtime, but tends to shift the bottleneck in setups away from the high-speed placement machines, and toward overall operator time (both on-line and off-line) and other machines.

The Feeder Management System

     The Qualcomm Feeder Management System is a computer-based system that uses bar-code technology, wireless portable data terminals, and personal computers to manage information about the feeders.

     The purposes of the Feeder Management System are to:
1.Decrease on-line and off-line setup times
2.Reduce the time needed to physically locate component reels
3.Automate verification tasks to reduce errors
4.Provide feeder size information
5.Clearly label feeders with size and machine information to reduce feeder selection errors
6.Provide feeder location information
7.Provide feeder inventory information
8.Perform predictive feeder maintenance
9.Perform schedule modifications to reduce the number of comprehensive setups that must be performed

     The system consists of two phases. Phase One, which has been implemented, includes objectives 1-6 above. Phase One provides the tools needed for setup time reduction. For example, there is a "Locate Parts" program which allows an operator to find a component reel by simply entering the component part number. The program returns the physical location of the component reel in the factory. Formerly, the "locate parts" task required the operator to physically go hunting for the desired part. Objectives 7 through 9 are to be implemented in the future.

     The system is designed in a modular fashion. Each software "tool" corresponds to a task performed by an operator prior to the systemıs development. The advantage of this modularity is that it is easier to train operators on the system because the software tools relate directly to tasks they were already performing manually.

     Objective 1 (decrease on-line and off-line setup times) uses two tools, the Feeder Setup tool and the Verify Setup tool.

     Objective 2 (reduce the time to physically locate component reels) uses the Locate Parts tool which helps operators find a component part easily.

     Objective 3 (automate verification tasks to reduce errors) uses the Feeder Setup tool. The "external" on-line setup tasks were moved off-line and automated with this tool.

     Objective 4 (provide feeder size information) uses the Feeder Setup tool as well as a stand-alone tool.

     Objective 5 (clearly label feeders with size and machine information to reduce feeder selection errors) was the only one that did not require a computer system. It uses color-coded labels that uniquely identify the feeder size and machine type. The setup sheets now include information about what kind of feeder label to look for with each reel, making matches easy.

Architectural overview

     Figure 6 depicts the basic design of the system. There are two computer platforms: radio-frequency (RF) portable data terminals running UNIX through a Telnet session and PCs running Windows 3.1. Setup personnel use these computers during on-line and off-line setups. Off-line feeder setup personnel use the Feeder Management System software running on the PC to perform feeder setups. On-line setup personnel use the RF terminals. The Qualcomm database contains the Feeder Management System data while the maintenance database contains the maintenance records for all the feeders.

Figure 6 The Hardware Design of the System

     For details of the hardware, software, and operation, see Coble, 1996.

Results

     The setup system has been changed in a number of ways, all leading to faster setups. The changes include:

• SMED (Single Minute Exchange of Dies) concept changes
• Computerized feeder management system
• Use of hot swapping where possible
• Addition of a third operator to most lines and shifts.

     The changes can interact. For example the third operator has less effect on setup times when they have been reduced using the other three methods. These interactions will become clearer below.

     The effects of these changes fall into four categories. The magnitude and economic value of some of these effects are hard to measure, but they are all potentially significant.

• Reduced line downtime (on-line setup time).
• Reduced labor content and elapsed time in both on-line and off-line (feeder preparation) setup. This reduces labor cost and increases production flexibility.
• Reduced errors from incorrect components, requiring less rework. Qualcomm does not keep systematic data on magnitude and causes of rework, perhaps due partly to political sensitivity of the issue. While we believe the incidence of these problems has gone down due to the easier setup verification provided by the computerized feeder management system, we have no rigorous measures of this benefit.
• Miscellaneous benefits, many unanticipated. The system is being used by people and in ways for which it was not designed. For example, it now takes less than five minutes to reconcile the bill of materials to the placement machine program and find discrepant part numbers. As a result this is being done well in advance, and when problems are found they no longer disrupt the actual parts preparation operation. Again we have only anecdotal information on the size these benefits.

     The rest of this section examines the first two of these effects, and discusses implementation issues and next steps.

Placement machine on-line setups

     Operators and one of the authors (Coble) conducted time studies on the setup of individual high-speed placement machines by a lone operator, using a sample of jobs of different sizes. The average setup required 31 minutes using the new methods, with a standard deviation of 8 minutes. This is the internal (on-line) time to set up a single high-speed machine by a single operator. By comparison, nine setups of various sizes measured before any changes averaged 78 minutes, with a standard deviation of 33 minutes.

     Figure 7 shows the raw data. The x-axis is the number of feeders in the job, and the y-axis is the total time required for the setup. The crosses are old method, while the dots are post change. Clearly, the two methods are dramatically different, and the new setup system time is much less dependent on the number of feeders.

Figure 7: Setup time versus number of feeders

     To more precisely compare the speeds of the old and new methods as a function of setup size, we ran regression analyses for each. All coefficients are statistically different than zero at the 5% significance level.
Setup time in minutes (new method) = 18 + .18 x number of feeders
(standard errors of estimators) (4.1) (.07) Adjusted R2 = .48 Standard error of residuals = 2.7 minutes Estimated time for 50 feeder setup = 27 minutes Setup time in minutes (old method) = 45 + 1.67 x number of feeders
(standard errors of estimators) (15) (61)
Adjusted R2 = .45 Standard error of residuals = 24 minutes
Estimated time for 50 feeder setup = 128 minutes

     As expected, the two methods are dramatically different. The new method removes most of the dependence on setup size, presumably due to the high speed with which feeder barcodes can be scanned. The significantly lower variability of the new method is also a mark in its favor, since it makes scheduling easier. The lower variability is due, we believe, to the much smaller number of problems that have to be rectified at the last minute by setup operators.

Economic value of faster setups

     The biggest financial impact of the new setup methods is from reduced downtime for the production lines. Qualcomm is presently running at a rate of about 1150 setups per year, although this is rising as demand and product variety grow. Downtime has a cost of about $700 per hour, based on the value of capital equipment as well as operator costs.

     The actual line downtime per setup depends on the staffing level for setups and the frequency and effectiveness of hot-swapping, discussed in Section 30, as well as on the improvements in setup methods themselves. Qualcomm recently shifted from two to three operators as the standard staffing level for the lines, partly in order to speed setups.

     We can construct several scenarios to estimate the cost of setup downtime for Qualcomm before and after the new system. (Table 3) Under the conventional setup method, a single machine with 50 feeders took an average of 128 minutes or 2.13 hours. With 2 operators setting up 3 machines this means the line is down for 3.2 hours on average. Approximately an additional half hour was needed for non-feeder setup activities, including software download, screen print machine, and conveyors. This gives a total of 4537 hours of downtime per year, with an economic value of $3.2 million.

     Actual time spent on setups and setup-related disruptions was higher than our 3.7 hour estimate, because of problems such as missing or incorrect parts, waiting for the off-line setup to be completed, machine difficulties, etc. For example if a feeder reel ran out of parts in the middle of a run, the line stopped while the operators located another reel and loaded it. This could take as long as an hour, since the replacement reel could be anywhere. The new system has ameliorated many of these disruptions. Off-line setup personnel locate sufficient reels using the "locate parts" command before the run begins. However we have no good data on reduction of disruptions, and are therefore omitting them from benefit calculations.

     Under the new system, suppose for now that hot swapping is not used. With 3 operators per line, one for each placement machine, all three placement machines can be set up in parallel. A conservative assumption is that operators cannot double-team the setups, so that total setup time for the three machines equals the slowest of the three setups. From the regression analysis in

Table 2

if setups average 50 feeders with some variation around this number, the slowest of the three setups will average about 30 minutes. The non-feeder setup activities can be done faster under the new system than before, partly because of the third operator, in roughly 20 minutes elapsed time. Total time is therefore 50 minutes per setup, 960 hours per year, with an economic value of $670,000. This is a 79% reduction, or almost a 5-fold improvement.

     Case Setup time per machine (min.) Elapsed Setup time, 3 machines Elapsed time, whole line Hours/yr. for 1150 setups Cost of time ($000/yr.) Reduction (percent of base) Base case 128 192 min. 222 min. 4537 hrs. $3,176 ---- 3 operators; new system; no hot swapping 27 30 50 960 $670 79% 3 operators; new system; 80% of setups are hot swapped   25 when hot swapped 34 = weighted average 652 $456 86%

     In fact hot swapping of both high-speed machines can be used in about 80% of the setups. In these cases only the fine-pitch placement machine must be set up off-line. This machine has fewer feeders than the high speed machines, so should average around 25 minutes. Meanwhile the other two operators can be doing the remaining non-feeder activities, so that the total duration is approximately 30 minutes. Taking a weighted average of 30 minute and 50 minutes per setup gives an average of 34 minutes per setup, for a total downtime of 652 hours per year, with an economic value of $456,000. This is a 7-fold improvement over the base case.

Labor savings

     In addition to reducing line downtime for setups, the new methods substantially change the off-line setup activities (while the line is running). One important change is faster work by the off-line setup operators, who use the Bill of Materials to pull the correct reels and feeders for a job, and put them on a rack prior to the setup. (See Figure 2 Feeder Setup Process Flow Diagram.) Benchmark studies estimated 4.2 hours per job, prior to the changes. About 70 percent of this time was spent looking for parts.

     No hard data is available post change, but we estimate a 40 percent reduction in time searching for parts due to use of the feeder management system. This is mainly because the new system accurately locates parts reels about 95 percent of the time, no matter where they are in the factory.

     Thus the savings is approximately 1350 labor hours per year.

     In addition to about seven off-line setup operators total, the plant has about 25 material coordinators. They have also asked to use the feeder management system, and find it makes their jobs easier. Chasing parts used to average an hour per part. With the new system it is down to ten minutes per request. Since about a quarter of their time was devoted to chasing parts, this is a 20 percent improvement in their overall effectiveness equates or an additional 10,000 labor hours per year.

     Offsetting these labor savings are additional off-line activities for line operators, which were previously done while the line was stopped. For these and other reasons, most lines now have three operators rather than two. This works out to about 5 additional operators, offsetting the reduction in setup time by the off-line operators.

     We can now summarize the economic benefits of the new setup system (Table 4). This is based on $700 per hour for line downtime, and $40 per hour for operators including supervision, etc. Clearly, the big impact is from reducing line downtime.

     Old system New system % Change $ savings/yr. with 1150 setups/yr. Setup Downtime, elapsed hours/setup 3.2 hours 0.56 hours -86% $2,720,000 Off-line setup time, labor hrs./setup 4.2 hours 3.0 hours -28% $54,000 Material handlers 25 24 (equiv.) -5% $400,000 Line operators per line 2 3 +50% -$380,000 Total benefit per year       $2,800,000

Development Costs

     These setup improvements were developed over the course of more than a year. We estimate the one-time development costs at about $350,000. These break down as follows: hardware including nine wireless terminals $45,000, lead engineer $100,000, other engineers, programmers, and consultants working part time $200,000. Software and database costs were reduced by integrating the feeder management system into existing databases. All labor costs are burdened. Two imponderables in this calculation are the value of top management time overseeing the project, and the opportunity costs of the engineers. If they had not done this project, they would have worked on something else. Engineers at a rapidly growing company like Qualcomm are always a scarce resource; managementıs job is to make sure they work on the most important projects.

Implementation issues

     There were a number of implementation issues during the deployment of the system in the factory. First, a number of the operators were not familiar with the personal computer platform and had to be trained in the basics of Windows 3.1 before training them on the system. Computer training consisted of a two hour hands-on session given by the Qualcomm training department.

     Second, it was important to keep the operators informed of the systemıs progress to maintain enthusiasm for the system. System training consisted of a number of discrete steps occurring over the life of the project. All of the operators, both those involved with the system design and others, were kept informed of the systemıs progress during its development and of the benefits it was expected to provide. Later, all of the operators were given demonstrations of the system prototype to keep them informed and involved in the process. The system was deployed on line two first and then moved to lines one, three, and four once it was stable. This allowed us to train operators gradually as we moved from one line to the next. System training occurred on the job, which allowed the operators to use the system as it was intended rather than in a simulated environment. Because the operators knew what to expect with the system, they were extremely enthusiastic about it, even when system bugs impeded their progress. The operators were also invaluable in finding system bugs and discrepancies not caught earlier during acceptance testing.

     The one-page Quick Start cards created as part of the user training manuals allowed the operators to learn the system tools quickly, from cards, without consulting the user manual.

     Finally, the system had to be deployed so as not to disrupt production. This required coordination with the planners, schedulers, and supervisors.

Conclusion

     In Ohnoıs terminology about setups, "single minute" exchange of dies refers to single digit setups, i.e. less than 10 minutes. Since he typically worked with stamping presses with multi-hour setups, this was quite an accomplishment.

     Qualcommıs approach to setup reduction started with traditional SMED re-engineering. By applying the principles of SMED to pick and place chip shooter machines, we were able to reduce setup times by removing all activities that could be done off-line. In some cases, "hot-swapping" is possible, so that individual machines can be set up in a few minutes. The key task of verifying feeders, however, must be done on the actual feeder tables, and was time consuming and error prone. Therefore we built a computerized information system to assist with feeder management. For speed and operator convenience, it uses a modern panoply of computerized tools, e.g. barcode readers and wireless terminals. Qualcomm also increased the staffing of the lines, although that was more to deal with disruptions while running than for setups.

     The net effect of these changes was to reduce the incremental setup time per feeder from 1.7 minutes to .18 minutes or 11 seconds (based on regression results in Table 2), a 9-fold improvement and approaching "single second." However the total time to do a standard 50 feeder setup is still about 30 minutes. We estimate the improvement in total line setup time to be about 6-fold, from 3.7 hours to less than 0.6 hours. From our cost and benefit calculations, we estimate a benefit/cost ratio of 8:1, and a payback period of less than six months.

     It is useful to compare our results with other efforts to improve setup times, but published data is scanty. Our numbers seem consistent with but better than the results reported by Aguayo and Tran of "30 seconds per feeder" for setups after implementing their verification system. Jain et al report one to five minutes per feeder at various H-P plants to set up similar chip shooter machines. None of these authors discuss the time for setting up the whole line.

     We view our results as an example of a "dynamic approach" to operations improvement, in that the physical coefficients of the manufacturing system (setup time) are improved by deliberate learning. The "static optimization" of setups, such as the EOQ models, takes setup times as fixed and determines the optimal run length to minimize inventory plus setup costs. As Porteus (1985) pointed out, investing in faster setups is also an option. Jain et alıs approach at H-P is such an investment, to develop a mathematical optimization program run daily to optimize the sequence of jobs. In the H-P case, one implementation gave a 70-80 percent reduction in setup times of the chip-shooter (3-fold to 5-fold improvement), albeit with some increase in run time. Another implementation gave a theoretical improvement of 53 percent (2-fold) and a documented 31 percent decrease in setup times.

     Our approach can be viewed as another investment in reduced setup time, which required roughly one man-year of effort and a cash outlay for computer equipment. We estimate our method gives an 86% improvement, or a 6-fold reduction in setup time, for the line as a whole, and a somewhat larger reduction for the chip-shooters. This is slightly better than the best-case 80% reduction reported by Jain et al, but the cases are not exactly comparable for several reasons. Our approach and the optimal sequencing approaches are not exclusive, although they are partial substitutes. Qualcomm is now considering software to optimize the sequence of jobs in the system.

     We view our project as validating the applicability of Ohnoıs Single Minute Exchange of Dies approach well beyond the industries where he developed it. His fundamental insight of separating internal and external setup activities proved its value. Wherever possible we used his "common-sense" techniques, such as color coding the different sizes of feeders and reels to make them easy to match. However our factory had 4500 different reels and 3000 different part numbers, which are analogous to the dies Ohno worked with. Operators could not track such a large number effectively using simple methods; therefore we found it necessary to add a state-of-the-art computer system for keeping track of the reels and feeders. This sytem uses wireless data terminals, barcodes, and modern databases, which are a departure from the simple visual control methods Ohno used and advocated. We believe that the simple (SMED) and sophisticated (computer) methods were complements. Either alone would not have been nearly as effective.

Further Work

     Holding "less than 10 minutes" as the goal for setups, there is room for further improvement. Qualcomm continues to work on setup time and downtime reduction, with actions in a number of areas. These include:

• Implementing predictive feeder maintenance. The new feeder database makes this feasible, since it tracks the use of individually barcoded feeders.
• Designing boards for commonality of parts. Designers often specify slightly different parts without needing to, e.g. different brands with equivalent electrical performance. By using more common parts, the number of feeder changes from one setup to the next is reduced.
• Board grouping so that similar board designs are run consecutively on the same line. In some cases this will eliminate setups entirely for second and subsequent jobs of a group. This is related to the "optimal sequencing" of Jain et al, but not as ambitious since it treats different groups of boards as entirely unrelated.
• Integrating printed circuit board setups with the rest of the assembly process. Presumably the lot size and job sequencing for board assembly should be coordinated with downstream assembly of finished systems. For example, the board lot size should be an integer multiple of the box lot size. To date, such coordination has not been attempted.
• Tracking how many parts are on individual reels, and reel ID numbers. In this way when there is more than one reel with a particular part, the operators will be able to select the best one for a particular run length. It will also help with supplier tracking.

     Through these and other changes we anticipate continued improvement in effective line capacity. An important tool is a new system for better daily and weekly tracking of actual machine downtime due to all causes. This allows supervisors and engineers to identify the most crucial areas for further improvement.

Acknowledgements

     Financial support for this research was provided by Qualcomm, Inc., the National Science Foundation through a grant to the Program in Advanced Manufacturing at UCSD, and the Alfred P. Sloan Foundation through a grant to the Data Storage Industry Globalization Project at UCSD. Support at Qualcomm was provided by John Swanson, Rich McGee, David Hind, Satish Ram, and Robert Fredella. Engineers who helped with the project include Dave Moser, Joe Brower, Judy Brown, Hydie White, Scott Harvester, Les Davidner, CP Then, Robert Stack, and Richard Scheller.

     Our thanks to Liz Bohn, John McMillan and Christian Terwiesch for helpful comments.

References


• Aguayo, K. and K. Tran. 1994. "An Automated Approach Verifies Parts Placement". Surface Mount Technology January, 39-31.
• Ammons, J.C., M. Carlyle, G.W. DePuy, K.P. Ellis, L. F. McGinnis, C. A. Toveyand H. Xu. 1992. "Productivity Improvements in SMD Placement Through CAPP". ISHM 1992 Proceedings 18-22.
• Askin, Ronald G., M. Dror and A. J. Vakharia. 1994. "Printed Circuit Board Family Grouping and Component Allocation for a Multmachine, Open-Shop Assembly Cell". Naval Research Logistics 41, 587-608.
• Bard, Jonathan F., R. W. Clayton and T. A. Feo. 1994. "Machine Setup and Component Placement in Printed Circuit Board Assembly". The International Journal of Flexible Manufacturing Systems 6, 5-31.
• Cavinato, J.1991. "Logistics Tools: Lowering Set-Up Costs". Distribution 90: 6, 52-53.
• Cobb, I. 1992. "JIT and the Management Accountant". Management Accounting-London 70: 2, 42-44.
• Coble, S. 1996. "Machine Setup Time Reduction for Electronics Assembly". M.S. Thesis. School of Engineering, University of California at San Diego, San Diego, CA. 92093-0519
• Davidow, W. H. and M. S. Malone. 1992. The Virtual Corporation. HarperCollins, New York, 124-125.
• Esrock, Y. P. 1985. "The impact of reduced setup time". Production and Inventory Management Fourth quarter, 94-101.
• Fiedler, K., J. E. Galletly and J. Bicheno. 1993. "Expert advice for JIT implementation". International Journal of Operations & Production Management 13: 6, 23-30.
• Hahn, C. K., D. J. Bragg and D. Shin. 1988. "Impact of the Setup Variable on Capacity and Inventory Decisions". Academy of Management Review 13:1, 91-103.
• Hall, R. W. 1983. Zero Inventories. Dow Jones-Irwin, Homewood, Illinois, 84.
• Handfield, R. 1993. "Distinguishing Features of Just-in-Time Systems in the Make-to-Order/Assemble-to-Order Environment". Decision Sciences 24: 3, 581-585.
• Hay, E. J. 1989. "Driving Down Downtime". Manufacturing Engineering 103: 3, 41-44.
• Jaikumar, R. and R. Bohn. 1992. "A Dynamic Approach to Operations Management: an Alternative to Static Optimization". International Journal of Production Economics 265-282.
• Jain, S., M. E. Johnson and F. Safai. 1996. "Implementing Setup Optimization on the Shop Floor". Operations Research 44: 6, 843-851.
• Landers, T.L., W. D. Brown, E. W. Fant, E. M. Malstrom and N. M. Schmidt. 1994. Electronics Manufacturing Processes. Prentice-Hall, Englewood, New Jersey, 37-38.
• Mirza, M. A. and E. M. Malstrom. 1994. "Required Setup Reductions In JIT Driven MRP Systems". Computers and Industrial Engineering 27: 1-4, 221-224.
• Nahmias, S. 1993. Production and Operations Analysis. Irwin, Burr Ridge, Illinois, 18.
• Porteus, E. L.1985. "Investing in Reduced Setups in the EOQ Model". Management Science 31: 8, 998-1010.
• Sadiq, M. and T. L. Landers. 1991. "Decision Support System For Intelligent Parts/Slot Assignment on a SMT Placement Machine". Computers & Industrial Engineering 21: 1-4, 565-574.
• Sadiq, M., T. L. Landers and G. D. Taylor. 1993. "A heuristic algorithm for minimizing total production time for a sequence of jobs in a surface mount placement machine". Computers & Industrial Engineering 1327-1333.
• Shingo, S. 1985. A Revolution in Manufacturing: The SMED System. Productivity Press, Cambridge, Massachusetts.
• Shingo, S. 1992. The Shingo Production Management System: Improving Process Functions. Productivity Press, Cambridge, Massachusetts.
• Smaddar, S. and L. Heiko. 1993. "Waste Elimination: The common denominator for improving operations". Industrial Management & Data Systems 93: 8, 13-19.
• Spence, A. M. and E. L. Porteus. 1987. "Setup reduction and increased effective capacity". Management Science 32: 10, 1291-1301.