Description
The report explains about effect of advanced manufacturing technology on process stability during flexible production in a process industry.
PRODUCTION
AND OPERATIONS MANAGEMENT Vol. 7, No. 3, Fall 1998 Printed in U.S.A.
COMPUTER PROCESS
INTEGRATION FAILURE IN AN EMPIRICAL
DAVID M. UPTON
AND
AND CATASTROPHIC FLEXIBLE PRODUCTION: INVESTIGATION *
ANDREW P. MCAFEE
Graduate School of Business Administration, Harvard University Boston, Massachusetts 02163, USA
This paper empirically investigates the effect of advanced manufacturing technology on process stability during flexible production in a process industry. A sample of 61 North American fine paper plants is used to examine the relationship between the level of automation installed for controlling changes between paper grades and the incidence of paper web breaks. These web breaks are .catastrophic failures; they require the entire plant to be stopped, reinitialized, and restarted. Because a large fraction of breaks occurs shortly after changeovers, they are an important determinant of the aspect of plant flexibility, called mobility, or the ability to move between products with only small penalties. In an attempt to ensure stable and mobile production, many plants have implemented changeover automation. We find, however, that higher levels of this automation are significantly associated with higher rates of catastrophic failure among the plants studied. We suggest that this finding becomes less paradoxical when considered in light of a recent stream of research on advanced manufacturing technologies, loosely called the usability perspective. According to this perspective, automation designed and implemented with the narrow, technical goal of replacing human operators or removing their discretion over a production process is misguided, especially in environments in which requirements are changing rapidly. (MANUFACTURING FLEXIBILITY; OPERATIONS MANAGEMENT; MANUFACTURING AUTOMATION; MAN-MACHINE INTERACTION)
1. Introduction The well-documented forces of strengthened foreign competition, market fragmentation, and the continual need for innovation are combining to place manufacturers from diverse industries in a common dilemma-how to increase responsiveness without dramatically increasing costs. This mandate for flexibility is extensively documented among discrete manufacturers but is less well understood in process industries. Producers of streams, slurries, webs, and other continuous products are realizing, however, that their outputs must also be tailored to satisfy increasingly disparate customer requirements. In both process and discrete industries, however, flexibility means much more than being able to produce a wide variety of products. The taxonomy of flexibility developed
* Received October 1994; revised February 1996, January 265 1059-1478/98/0703/000$1.25
Copyright 0 1998. Production and Operations Management Society
1997, and September
1997; accepted
October
1997.
266
DAVID M. UPTON AND ANDREW P. MCAFBE
by Upton (1994) includes both range, or the number of manufacturable products, and mobility, or the ability to switch within the range with only small penalties in time, cost, or robustness. Mobility is often underemphasized, but it is an important consideration; it does a plant no good to be able to make a wide range of products if changeovers take too long, are too expensive, or lead to breakdowns on the line. The latter is a critical issue for process industries. One of the most significant barriers to running quick response process plants is that the manufacturing processes associated with these industries are often destabilized by a change of product. In steel rolling mills, these instabilities result in cobbles (buckles), which must be flame-cut out of the rolls. In the chemical industry, instability leads to poor yields and the production of impurities, which must be flushed out of tanks and pipes. In the paper industry (which is the subject of the field research presented here), the most common breakdowns are web breaks, after which the paper web must be manually torn out of the machine and then rethreaded. These failures are each catastrophic in that the process cannot simply be nudged back into place. It must be stopped, repaired, and restarted, often at great expense. Computer integrated manufacturing, or CIM, is often viewed as a tool to increase both range and mobility. Many operations have adopted CIM as a means to insure that product changeovers occur smoothly and to avoid severe interruptions. They see computer-controlled changes as powerful protection against process failure. This paper explores the relationship between CIM and catastrophic failures in one process industry setting: fine paper manufacturing. It examines the effect of automated changeover systems on the incidence of web breaks in a sample of 61 paper plants. The main finding from this investigation is that higher levels of changeover CIM are actually associated with a higher rate of catastrophic failures in the environments studied. While this result may appear counterintuitive, it is congruent with conclusions from several researchers who have studied the impact of advanced manufacturing technology ( AMT ) , such as CIM . They find that AMT, which “de-skills” operations workers, removes opportunities for learning and improvement and can lead to less efficient, more errorprone operations. The results reported here are less surprising when viewed from this perspective since the changeover automation studied here is primarily of the de-skilling variety. Section 2 outlines the motivation for the present study. Section 3 provides background on the fine paper industry, concentrating on the need for flexibility, the nature and use of changeover automation, and the phenomenon of catastrophic process failures, that is, web breaks. Section 4 presents the empirical models used to explore the relationship between automation and catastrophic failures and gives their results. Section 5 discusses these findings and introduces the usability perspective as a way to help understand them, and Section 6 concludes. 2. Motivation Much of the present knowledge base on the impact of automation in operations has come from case studies of successful and unsuccessful AMT implementations, interviews and attitudinal surveys of managers, and reviews of high-level industry data, such as employment levels, wages, and job classifications. There have been relatively few crosssectional empirical studies tying AMT implementation to salient operational performance measures [one example is Jaikumar’s ( 1986) investigation of divergent flexible manufacturing systems (FMS) performance in the United States and Japan]. This type of research, however, can be a valuable tool for addressing whether AMT implementation and use makes a significant difference in important outcomes, such as cost, quality, flexibility, and reliability. We had an opportunity to conduct such a study as part of a larger investigation of flexibility in the North American fine paper industry. This work (Upton 1993a, 1993b,
COMPUTER
INTEGRATION
AND CATASTROPHIC
PROCESS FAILURE
267
1995) examined how flexibility is achieved in a process industry, as well as the tradeoffs between productivity and various types of flexibility. The data collected for this study can also be used to examine the relationship between AMT and 1 measure of process mobility, an important category of flexibility for this industry, as explained below. Specifically, it was possible to assess the relationship between the level of computer-automated changeover equipment present in a plant and the frequency of catastrophic failures, that is, paper web breaks. Because these failures often occur after changeovers, they are important determinants of a paper plant’s mobility. And because the changeover process has been automated to varying degrees across the plants studied, the impact of different degrees of automation could be assessed. 3. Flexibility, Automation, and Catastrophic Failure in Papermaking
3.1. Background: The Fine Paper Industry The North American fine paper industry accounts for 45% of the world’s uncoated fine paper capacity. As in many industries, customers have become increasingly reluctant to hold inventory, often pushing inventory back onto those suppliers unable to deliver justin-time. Not only does this incur additional expense for the supplier, but the proliferation of products often makes inventory holding an infeasible method of providing quick response. To meet these competitive imperatives, firms are anxious to build supporting capabilities at the plant level. The ability to provide quick response to customers demands capabilities in many areas, as well as in the manufacturing shop itself. Hammond ( 1992) describes a number of these capabilities. There is, however, no substitute for a responsive plant, which can switch between products quickly and reliably to respond to growing demand volatility and product proliferation. 3.1.1. THE PAPERMAKING PROCESS. A paper mill usually comprises a pulp plant, which provides pulp for a group of 2 to 10 converting plants or machines (the terms are used interchangeably), which turn the pulp into paper. The papermaking process is essentially a water-removal procedure. Water is extracted from a pulp slurry by use of gravity, pressure, and heating. The pulp slurry is laid onto a moving fabric, from which water drains off. The pulp web is squeezed and heated by a series of rollers until it is strong enough to support its own weight. Further roller heating then takes place until the paper moisture content is just below that of the ambient atmosphere. The paper is then collected on a reel in the plant to be subsequently sliced transversely and longitudinally into roll sizes convenient for customers and sheeting machines. Figure 1 is a simplified view of a paper plant. 3.1.2. INDUSTRYHISTORYANDCOMPETITIVEENVIRONMENT For the pastcentury,North American paper conversion plants have been steadily increasing in size, output rate, and investment requirements; Figure 2 displays this trend. The paper industry has traditionally competed primarily on cost, and economies of scale have been significant. Larger plants, therefore, have continued to be more efficient at producing standard paper grades. Companies have historically accumulated funds through depreciation and retained earnings and leapfrogged the competition with bigger, more expensive plants, usually on an existing site (the minimum efficient scale for pulp mills exceeds that of converting plants, so a number of plants tend to be co-located with a pulp mill; incremental expansion is thus considerably cheaper than a green field site.) These plants then concentrate on satisfying demand for standard paper types, for example, xerographic paper for offices.
3.2. Flexibility in Papermaking
The existence of a few large commodity plants at any point in time means that the majority of converting plants are unable to compete on the basis of price for common
268
DAVID
M.
UPTON
AND
ANDREW
P. MCAFEE
Water Removal Method:
Gravity Pressure Heat
RESS
SECT4ON
RE DRYER
SECTfcSN
FIGURE1. Simplified Cross-Sectional View of a Typical Paper Plant. Water is Removed by Gravity, Pressure, and, Finally, Heat.
paper types. After their scale economies have vanished, these facilities often move to producing other grades of fine paper, which have volumes too low to attract the attention of the industry’s giants. Because demand for these products is typically more fickle, and because competition to fill it is intense among the smaller plants, they face great pressure to become more flexible. They must both be able to make a wider range of products and to move among them quickly and easily. This need translates directly into a requirement to manage paper grade changeovers more effectively. Different paper grades are equivalent to different products. There are 4 kinds of grade changes made in a paper converting plant, listed here in decreasing order of difficulty.
1200
-i
0 0
0~
r
I
I
‘,
’
I
I
I
’
4~1~
I
/
I
1890
1900
1910
1920
1930
1940
1950
1960
1970
1980
1990
2000
Date of Original Installation
FIGURE2. Plant Size Versus Installation Date.
COMPUTER
INTEGRATION
AND
CATASTROPHIC
PROCESS
FAILURE
269
Color Changes: These often require a complete plant wash down, to remove any traces of the previous color. Very few plants in the sample made a variety of colors, and these changes were, therefore, not explored in this study. Furnish Changes: These require changes in the composition of the pulp and entail resetting valves, drives, motors, etc., in order, for example, to brighten the paper with titanium dioxide or to vary the ratio of long-fiber to short-fiber pulp. Furnish changes often incorporate basis weight changes (the following category). A furnish weight change alone is more difficult than a basis weight change alone. Basis Weight Changes: Basis weight changes involve altering the rate of pulp deposition on the web to vary the area density of the paper. Changes usually require alterations in other plant parameters, such as web speed. Basis weight changes, which are part of broader furnish changes, are not considered separately in this research. Caliper Changes: These are minor adjustments to the thickness of the web at the dry end of the plant and may not often be recorded by the mill. The two types of changeover considered in this research are basis weight and furnish changes. These are the key dimensions that distinguish one type of paper from another, and most plants studied made both types of changeover regularly. As discussed above, the ability to switch between grades quickly and reliably has become an important task for manufacturing managers-indeed, it has become a critical issue for the survival of many small and medium size plants. According to interviews, changeovers can absorb as much as 20% of productive time and have become one of the most important determinants of a plant’s performance. 3.3. AMTfor Flexibility: Changeover Automation To address the increasing need for mobility, that is, rapid and reliable changeovers, many plants have installed computer integration systems designed to control the gradechange procedure. These systems communicate with lower-level control devices in the control hierarchy, such as valve and motor controllers, and coordinate their actions in real time during a changeover. In doing so, they generally substitute for more manual operatorcontrolled changes. While almost all plants in the sample were fitted with automation to control process settings during steady-state operation, there was a wide disparity in the degree of automation for the process of changing grades. Some machines had automation that controlled all changeovers; others had automation for only specific types of changes, for example, basis weight changes, and some relied solely on direct operator control of the changeover process. At the sites with automated changeover systems, the general view, as expressed in interviews, was that they were very much a black-box; that is, no one in the plant knew quite how their system worked. Programming was carried out by the system vendor, as were any modifications required during installation. During automated changeovers, the machine’s crew stood by in case of emergencies but had no active role. This pattern corresponds closely to the observations of Zuboff ‘s ( 1988) about the the relative roles of AMT and workers in a number of paper plants. She found that new technologies in this industry consistently relegated workers to marginal roles within the production process. In contrast, nonautomated changeovers demanded that members of a crew work closely together to coordinate the many required adjustments to speeds, separations between rollers, level of additives, etc. This was a demanding task, especially since few machines were designed for flexible operation. Interviews provided anecdotal evidence, however, that crews were able to learn how to make changeovers rapidly, in some cases, more rapidly than the automation installed for the task. As one operator put it,
The computer is real slow-it gets kind of boring sitting watching [the plant run] out a better way to change manually. We’ve got it pretty much figured out now-we so we figured all practiced
270
DAVID
M.
UPTON
AND
ANDREW
P. MCAFEE
as a team, [. .] worked out who should do what and when. We can always beat the computer, and do changeovers faster than they used to do them manually-it’s just a matter of working out the right routine.
This kind of tinkering with the grade changing process, even if it is successful at lowering average changeover times, could conceivably involve some risk; experimentation may precipitate a web break. Since automated systems perform the changes in the same way each time, one might expect them to make more reliable changes, even if they are made more slowly. 3.4. Catastrophic Failure: Paper Web Breaks Breaks in the web require operators to climb into the machinery, tear out the broken web, and rethread the plant. Little automation is available to help with the restart process because recovery procedures vary greatly, based on factors such as where the break occurred in the machine and the amount of snarled paper. In other words, there is no prespecifyable sequence of steps that can be automated; the operators must assess and act on each situation individually. Most of the time during a restart is absorbed with tearing out the old tom web and manually throwing the leading edge of the new web into the pinch rollers. Restarting after a break takes between 30 minutes and 4 hours, during which the machine produces no usable output. Restarting can be dangerous for the people involved, who must physically manipulate the flimsy paper web in a very hot, moist paper plant, which is often under power. Breaks are indeed catastrophic, and plant personnel strongly prefer to avoid them. Operators’ vocal reaction to a break is fairly predictable. What is less predictable, however, are the circumstances giving rise to a break. During interviews, many operators tied web breaks to changeovers. As one supervisor said,
If we’re going to get one it just happens! it’s hard there’s something wrong don’t really know what’s to figure it out. [a break] it’ll happen right after we change grades. Sometimes though, to tell if it was because of the grade change or some other reason. If with the pulp we’ll often get a break. More often than not though, we caused the problem. People are usually too busy getting the break fixed
This suggests that a web break is a poorly understood event, with some identifiable determinants and some that are, at least according to operators, random. The precise mechanisms that produce breaks are not well understood, and predictive models do not exist. Even so, several factors that appear to contribute to web break rates, beyond changeover frequency, have been identified. They are discussed below in Section 4.2.2. AMT is not one of these identified factors. To our knowledge, changeover automation has not been linked to web breakage, either empirically or anecdotally. The purpose of the present investigation was to explore this relationship. 4. Empirical 4.1. Field Research We studied 61 fine paper plants (in 25 mills and 11 companies) over a period of 2 years. To provide background, structured and unstructured interviews were carried out at the following three levels: vice-president for operations or COO for the company, plant manager, and paper plant supervisor and operator. We also observed operations in 15 plants and worked alongside operators on the equipment to examine the way in which process changes were made. Over 20 paper break events were physically observed during the course of the field research. Research: Design and Results
COMPUTER
INTEGRATION
AND CATASTROPHIC
PROCESS FAILURE
271
4.2. Data Collection 4.2.1. DEPENDENT VARIABLE. Web Break Frequency: In each of the 61 plants, operators were asked to provide data on the average number of paper breaks experienced in a plant each week. A paper break is an important occurrence on a shift, and operators’ estimates of break frequencies were found to be very reliable in the few plants where records of breaks were kept. The model variable breaks captures the estimated number of breaks per week. 4.2.2. INDEPENDENT VARIABLES. Changeover Automation: The variable CIMMOB ( CIM for mobility) represents the degree to which changes in grades could be made automatically through computer integration of other computerized subsystems. The degrees of automation for basis weight, caliper, and furnish changes were each scored out of seven (seven being completely automatic and zero being manual). These three scores were then added and divided by seven, to give a maximum possible score of 3, corresponding to three fully automated methods of changing grade. Color change systems were excluded since most plants in the sample did not make color changes. Change Frequency: There was widespread agreement that many web breaks occurred after changeovers, so it was important to include change frequency as a predictor. The weekly number of grade changes was estimated by the operators separately for basis weight and furnish changes. FREQB is the number of grade changes per week in which the primary dimension of change is the basis weight of the paper. FREQF is a similar measure for changes in furnish. Scale: Another factor likely to be a determinant of break rates is the scale of the plant. Larger plants have larger components moving at higher speeds and are therefore more difficult to control dynamically. The scale of a plant can be measured in the following three ways: Output, or tons produced per day Plant maximum rated speed (measured in feet per minute) Plant width (measured in inches) Data were collected from the plants on each of these metrics. Speed is the maximum speed at which a plant is rated to run, while width is the trim width of the plant. The variable tons represents the average daily output of the plant. The three scale measures are strongly intercorrelated (see Appendix A), so it may not be possible to ascribe explanatory power to any individual measure in a regression model. Each piece of data does, however, provide information concerning the scale of the plant. For this reason, these measures were combined by projecting them onto (oblique) principal components of their joint distribution. The details of this transformation are shown in Appendix B. The primary principal factor, ScaZe 1, is representative of the scale of the plant; tons, speed, and width all contribute positively to it. The second factor, Scale 2, has a positive contribution for tons and negative contributions from speed and width. It therefore measures one type of disparity among these three measures. A plant which, compared to others in the sample, has high output but low width and rated speed will have a high value for Scale 2. This variable, then, captures the degree to which a plant is running “full out,” or achieving the most output possible given its physical configuration. Age ofchangeover automation: Chew ( 1985) and Hayes and Clark ( 1985) have shown that the installation of new process technology can, in many circumstances, be followed by a short-term productivity dip as an operation, and its operators and managers become accustomed to the new technology and overcome start-up problems. It is possible that if much of the changeover automation at the plants studied was installed contemporaneously and shortly prior to this study, then any problems relating to it, such as a tendency toward more web breaks, would be transient manifestations of this phenomenon. To separate explicitly any such effects from the main relationship of interest, the length of time since the last major upgrade of the control system at the plant was calculated as the variable
272
DAVID M. UPTON AND ANDREW P. MCAFEE
CIMAGE. If this predictor is significant in the regression models, support would exist for a productivity dip hypothesis. Since only five of the plants had installed or upgraded their changeover automation within the previous 2 years, it is unlikely that the sites as a group are experiencing a productivity dip. Table 1 shows descriptive statistics for dependent and independent variables.
4.2.3. DISCUSSION OF DESCRIPTIVE DATA. The dependent variable, breaks, and the 2 independent variables measuring change frequency are based on estimates from plant personnel rather than direct observation or historical records. The objectivity of these measures cannot be guaranteed, but they were the best available metrics; very few plants kept records of their break rates. The most striking feature of the descriptive data is the very large variation in both independent and dependent variables. Break frequencies range from 1 to 60 per week and furnish changes from 0 to 250 per week. Furnish changes are much more common than basis weight changes, although it must be remembered that furnish changes often include a basis weight change, while a straightforward change in basis weight does not include a furnish change. 4.3. Linear RegressionModels
We used a series of nested linear regression models to estimate the impact of several predictors on the frequency of paper breaks for our sample of plants. Our independent variables were divided into control predictors, which were used to control for factors unrelated to our research question, and a single question predictor, that is, CZMMOB, the level of automatic grade change equipment. The dependent variable in all models was the number of breaks per week. Control predictors include the frequency of change demanded of a plant, multiple measures of the scale of the plant, and the length of time since the last major automation upgrade. A number of other more complex specifications were explored, as described later, but did not add significantly to the explanatory power of the model. 4.3.1. RESULTS. Tables 2 and 3 show parameter estimates for the models tested. Model 1 includes the only control predictors of FREQB and FREQF. In this model, furnish change
TABLE 1 Variable Descriptive Statistics Variable Breaks
CIMMOB
Description Break frequency Degree of changeover automation Basis weight change frequency Furnish weight change frequency Trim width Maximum plant speed Net output Scale principal Factor 1 Scale principal Factor 2 Time since last major upgrade of control svstem
Units Breaks/week
N 61 60
Mean 17.3 1.87 7.53 22.86 160.4 1457 230.81 0 0 20.33
SD 14.6 0.81 4.95 56.77 66.0 767 248.7 1 1 20.50
Min 1.0 0.43 0 0 72.0 250 18.0 -1.31 -1.04 2
Max 60.0 3 22.5 250 330.0 3200 1200 2.534 6.521 74
FREQB FREQF
Number/week Number/week In. Ftlmin Tons/day
60 61 60 60 61 60 60 51
Width Speed Tons Scale 1 Scale 2
CIMAGE
Years
COMPUTER
INTEGRATION
AND
CATASTROPHIC 2 for Models Model 2
PROCESS
FAILURE
273
TABLE Regression Model Coefficient GE) 10.99*** (2.58) .346 (.314) .164*** (.027) Parameter 1 Std. Coeff.
Estimates
l-3 Model Std. Coeff. Coefficient GE) 3.03 (5.28) .461 (.287) .217*** (.033) -.036 (.033) .025** (.OOS) ,003 (.003) ,090 (.068) 56 ,613 (.565) 49 12.928 ,000 3 Std. Coeff.
Independent Constant
FREQB FREQF
Variable
Coefficient (SE) 7.95 (4.22) ,434 (.284) .175*** (.024) -.048 (.034) .024** (.008) .003 (.003) 59 .633 (.599) 53 18.310 ,000
.116 .637***
Width Tons Speed
CIMAGE
.143 .680*** -.214 .401** ,158
.160 .625*** -.180 .457** .174 .135
Sample size R-squared (adj. R sq’d.)
DOF
F-value P (F > F*) ***p < ,001. **p < .Ol. *p < .05.
60 .482 (.464) 57 26.516 .ooo
frequency is found to be highly significant, while basis weight change frequency is not significant. However, these two variables have a relatively high correlation (=0.44), so both are kept in later models. This model explains 48.2% of the observed variance of
breaks.
Model 2 adds the three raw scale variables; tons is found to be significant, while speed and width are not. However, as discussed above, these three measures are very highly intercorrelated (see Appendix A), and it would probably be misleading to ascribe all
TABLE Regression Model Coefficient (SE) 2.07 (4.25) .617* (.265) .160*** (.023) -.039 (.031) .031*** (.008) -.004 (.004) Parameter 4 Std. Coeff. 3
Estimates
for Models
Model 5
4-6 Model Std. Coeff. Coefficient (SE) -1.10 (4.17) .613* (.264) .162*** (.022) 6 Std. Coeff.
Independent Constant
FREQB FREQF
Variable
Coefficient (SE) -1.75 .639* .198*** -.030 .031*** -.003 (5.15) (.271) (.032) (.031) (.007) (.004)
Width Tons Speed Scale1 Scale2
CIMAGE CIMMOB
.201* .625*** -.175 .523*** -.195
.221* ..573*** -.148 .570*** -.148
.20* .632***
2.15 (1.42) 4.47*** (1.19) 6.04** (2.18) 58 .703 (.668) 51 20.072 ,000 .332** .087 (.063) 5.01* (2.19) 55 ,678 (.631) 47 14.167 BOO .131 .306* 5.30** (1.84) 58 .700 (.671) 52 24.28 .ooo
.146 .305*** .292**
Sample size R-squared (adj. R sq’d.)
DOF
F-value P (F > F*) ***p < .OOl. **p < .Ol. *p < .05.
274
DAVID M. UPTON AND ANDREW P. MCAFEE
explanatory power to only one of them. For this reason, all three scale variables are kept in subsequentmodels. This model explains 63.3% of the observed variance of breaks. Model 3 includes CIMAGE. It is not a significant predictor of web break frequency when the predictors listed above are included in a model. This indicates that the plants studied are not, as a group, experiencing a productivity dip due to recently installed automation. A dummy variable that captured those plants that had installed or upgraded changeover automation within the previous 2 years was also tested with this model. Like CIMAGE, it was not a significant predictor, providing further support against the productivity dip hypothesis. Model 4 includes the question predictor, CIMMOB . It is highly significant (P < .Ol ) and has a magnitude of 6.04, indicating that each additional level of changeover automation is associatedwith an additional 6 breaks per week acrossthe plants studied. Interestingly, the inclusion of CIMMOB in this model results in FREQB becoming a significant predictor (P > .05 ) ; this variable was not significant in any previous model. It appears,then, that including CIMMOB in a regressionmodel helpsto clarify the relative contributions of furnish weight and basisweight changesto web break incidence. Model 4 explains 70.3% of the observed variance in breaks. Its final form is Breaks (est.) = 2.07 + 0.617 * FREQB + 0.160 * FREQF - 0.039 * width
- 0.004 * speed + 0.031 * tons + 6.04 * CIMMOB
As a final check, Model 5 adds the variable CZMAGE to Model 4 as a further check of the productivity dip hypothesis. Again, this variable is not significant, and it does not affect the coefficients of previously significant variables outside of their standard errors. It therefore does not appear that CIMMOB, the predictor of interest, is proxying
for
CIMAGE.
As an additional empirical exploration, we substituted Scale 1 and Scale 2 for tons, speed,and width in our final model. As discussedabove, these two variables are orthogonal projections of the three raw scale measures;Scale 1 is a positive combination of all three and thus representsthe total scaleof the plant, while Scale 2 measuresdifferences between them and captures the degree to which a plant is running “full out,” or achieving the most output possiblegiven its physical configuration. As Model 6 shows,it is this measure that is a highly significant predictor of breaks, while Scale 1 is not significant. Other regressionparametersremain largely unchanged. These results indicate that it is how much output plants are trying to wring from a given configuration, rather than the sheer plant scale, that drives web break rates. 4.3.2. SENSITIVITY TO OUTLIERS. To verify the robustnessof these findings, we performed an analysis of the sensitivity of Model 4, our final model, to the presence of outliers. Three observations had a value for Cook’s distance (which summarizesleverage and residual magnitude) greater than five times its mean. When these variables were removed and Model 4 (our final model) was reestimated, significant variables were unchanged, their coefficients remained stable, and the percentage of observed variance in breaks explained by the model increased to 82.1%. We conclude that our final model appearshighly robust to outliers. 4.3.3. OTHER VARIABLES. A number of alternative specifications of the models above were also explored. However, no significance was found (nor was the coefficient associated with computer integration substantially altered) by the addition of the following variables: Workforce experience; Average basis weight; Plant age;
COMPUTER
INTEGRATION
AND CATASTROPHIC
PROCESS FAILURE
275
Plant range (difference between highest and lowest basis weight) ; and The degree of computer integration elsewhere in the plant (for quality control for example) or the mill as a whole. 5. Discussion of Results 5.1. Effect of Control Predictors 5.1.1. FURNISH CHANGES. Model 1 in Table 2 shows that 48% of the variance in process failure rate is explained by the changes made per week on the dimensions of furnish and basis weight. However, the statistical strength of the estimate associated with basis weight changes is low. The statistical effect of furnish change frequency is, however, clearly established. In our final model, Model 4, both types of changeover are significant predictors, with the standardized coefficient for furnish changes approximately three times as large as that for basis weight changes. The reason for the disparity between the types of change is that furnish changes are much more disruptive of the process. Changes in basis weights, it appears from anecdotal evidence, are only likely to cause breaks when the change takes place with very light papers. The addition of the average basis weight made by the plant to Model 1 did, however, not increase its explanatory power and was found to be an insignificant factor in predicting failure rates. 5.1.2. PLANT SCALE. As Model 6 shows, Scale 2 is a highly significant predictor of breaks, while Scale 1 is not significant. This indicates that it is how much output plants are trying to squeeze from a given configuration, rather than the sheer plant scale, that drives web break rates. Results from Model 4 suggest that for a given speed and width, net tonnage is a primary driver of breaks. This is of interest in itself, though caution is necessary since the three factors are highly collinear (see Appendix A). As Models 4 and 6 show, controlling for scale does not interfere with the effect of interest for this work, which is the positive relationship between high levels of computer integration and process failure rates. 5.2. Effect of Changeover Automation As Model 4 demonstrates, the variable CIMMOB is a highly significant predictor of web break frequency after controlling for other plausible predictors, with each additional level of changeover automation associated with an estimated additional 6 breaks per week. This is, in many ways, a surprising and counterintuitive finding. This type of computer integration was installed specifically to increase process mobility, yet it has the effect within the sample of substantially increasing the penalty associated with moving between products. Furthermore, anecdotal and empirical evidence suggests that the automation is not even faster than an experienced crew at effecting changes (Upton 1995). This form of AMT, then, appears to be having the opposite of its intended effect for the plants studied. The questions immediately suggested by this finding are 1. What explains this result? 2. Why have the plants in the sample not noticed the deleterious effects of their changeover automation? Why has this form of computer integration not been widely declared a failure and removed from the machines? 5.2.1. THE USABILITY PERSPECTIVE. In our view, an answer to the first question, and a plausible explanation for these results, is suggested by research within the usability perspective. Usability research has addressed issues of AMT success and failure and the relationship between these outcomes and the ways that worker skills changed with the introduction of automating technologies. This research has embraced a variety of disciplines, including labor relations, operations management, organizational behavior, and sociology. The conclusions
276
DAVID M.UPTON
AND ANDREW P.MCAF%E
arising from these efforts, however, are remarkably consistent. Principally, they are that most companies do not anticipate or adequately plan for the organizational consequences of AMT, including the new and different skills that will be required of workers; that technology that aims primarily to replace or deskill workers is misguided and more prone to failure; and that AMT, which is designed and implemented to broaden the roles of workers instead of constraining them, is more likely to be successful. Each of these propositions has been supported by empirical research, as summarized below. (1) Most companies do not anticipate or adequately plan for the organizational consequences O~AMT. Several studies (Majchrzak 1988, NAS 1986, Butera 1984, and Senker et al. 1988) have found that firms typically embarked on AMT implementation projects without understanding organizational considerations, including the consequences for worker skill levels. Instead, these efforts were undertaken to reduce costs or maintain competitive parity and were treated as straightforward capital investments. A review of engineering and operations management textbooks published between 1939 and 1989 (Salzman 1992) supports these findings; it showed that only a minority of the texts made any mention of worker’s roles in automated manufacturing and equipment design and that these treated workers as marginal contributors, whose roles were to be minimized by automation. As Attewell (1992) points out, this approach sometimes winds up requiring a more skilled workforce because the resulting machinery can be more error-prone. A survey of large metalworking firms (Salzman 1992) found that majority did not have policies of worker participation during AMT design efforts, or during post-implementation design changes or discussions on worker modifiability and maintenance. One review of technology design models (Blacker and Brown 1986) summarized that ergonomics, sociotechnical theory, and theories of participative management have had little impact on AMT design because the climate of opinion is not receptive to them. All of this work supports the contention that over much of the history of AMT, decisions on equipment and technologies were made without considering their impact on people, jobs, and skills. (2) Technology that aims primarily to replace or deskill workers is misguided and more prone tofailure. The consequences of a narrow, exclusively technical view of AM-r impact have been articulated by a number of researchers. Adler ( 1992) characterizes the view that AMT will allow not only fewer workers, but ones that are less skilled and doing narrower jobs, as the deskilling myth and lists the following as reasons why this view is inappropriate in most circumstances. It ignores cognitive processes such as reasoning, as well as social interactions. It fails to take into account that people interpret situations and can adaptively monitor and change them. It takes the characteristics of the equipment as central and those of the involved humans as peripheral. It does not involve users in the design process, relying instead on technical experts. Adler concludes that treating AMT as a means of lessening reliance on operators’ abilities is misguided because “. . .in the majority of cases, the effective use of new technologies requires a workforce that is more skilled, not less.” Zuboff ( 1988) includes case studies of comprehensive AMT implementation in 3 papermaking facilities. These studies contain rich descriptions of the ways in which operators were removed from direct physical experience with the production tasks and the harmful consequences of this separation. In summarizing these, Zuboff states that “It was as if one’s job had vanished into a two-dimensional space of abstractions, where digital symbols replace a concrete reality. Workers reiterated. . .feelings of loss of control, of vulnerability, and of frustration. [The] oft-repeated metaphor spoke of being robbed of one’s senses and plunged into darkness.”
COMPUTER
INTEGRATION
AND CATASTROPHIC
PROCESS FAILURE
277
K&la, Clemens et al. ( 1992) make a clear statement of the dangers of an overly narrow view of AMT ; they maintain that such automation strategies freeze the production processes and division of labor and make plants unable to improve their systems or adapt them to changing needs and system requirements. (3) AMT, which is designed and implemented to broaden the roles of workers instead of constraining them, is more likely to be successful. Adler and Winograd ( 1992) call this the “usability challenge” and state that “The key challenge in designing new technologies is how to best take advantage of user’s skills in creating the most effective and productive working environment.” Kukla, Clemens et al. ( 1992) advocate designing systems as tools for skilled work that take advantage of the investment already made in people. These systems do not constrain operators to follow routines but allow them to respond quickly and effectively to constantly changing combinations of events. Majchrzak ( 1988)) expanding on an open systems model developed by Nadler and Tushman ( 1980)) proposes a framework that takes into account the task to be performed, formal organizational structure, the individuals who will work with the AMT, their informal organization, and the interactions among these elements. 5.2.2. APPLICATIONOFTHEUSABILITYPERSPECTIVETOAMTWITHINPAPERMAKING. The first of these propositions, that AMT investments have traditionally been made without taking into account organizational and human considerations, can help explain how North America’s papermakers made the decision to install changeover.automation. In a plant just beginning to face the need for greater flexibility after competitors had removed its scale advantages, the workforce would likely have been inexpert at effecting changeovers. In this situation, computer integrated changeover equipment would have appeared an attractive alternative. The immediate and narrowly defined technical benefits would have been substantial, and no drawbacks would have been apparent unless issues such as organizational learning and skill acquisition were explicitly considered. As the researchers who advanced Proposition 1 have indicated, it is unlikely that these topics were part of the decision to implement changeover automation. The case studies of Zuboff (1988) of AMT implementation in paper plants clearly indicate that their automation decisions were made without fully considering how the new technology would change the nature of work. The second usability proposition, that deskilling AMT can degrade operational performance instead of improving it, sheds some light on the present study’s findings. The principal result from the models tested, that higher levels of changeover automation are strongly associated with higher web break rates, ceases to be counterintuitive when viewed in light of this proposition. As the operator’s quote in Section 3.3 indicates, the changeover automation installed in many plants relegating workers to the role of bystanders during grade changes. In doing so, the computer-integrated equipment may have removed their opportunities to learn about and improve the process through experimentation, communication, or simple experience. As Kukla, Clemens et al. ( 1992) point out, this lack of learning becomes especially dangerous when system requirements or market needs change. Section 3.2, above, outlines that this is exactly what has happened to most paper plants since they were built; they have faced ongoing requirements to become more flexible and, particularly, more mobile. Many plants in our sample attempted to meet this need by installing changeover automation; the mean value for the CZMMOB is almost two-thirds the maximum possible value. If this AMT were, on the whole, deskilling, Proposition 2, above, indicates that higher levels of CIMMOB could be associated with poorer performance. Thus, the usability perspective suggests an answer to the first question posed above in Section 5.2. Based on our empirical results, and on the interviews and field observations described earlier, we
278
DAVID M. UPTON AND ANDREW P. MCAFEE
find it plausible that the papermakers in our sample installed changeover automation that was primarily deskilling, and that higher levels of this type of AMT lead to a higher incidence of paper web breaks. Proposition ( 3)) that AMT that avoids the deskilling myth is more likely to be successful, was not directly investigated by the present research. As far as we could determine, all of the changeover automation installed in the plants studied was intended to decrease human involvement in the changeover process. We saw no cases in which it had been designed and implemented with the sustained participation of the workforce, or where it was intended to support what Walton ( 1989) terms “commitment” rather than “control”. It was thus not possible to test, even anecdotally, whether this approach could yield better results along any dimension of mobility, including web break rates. 5.2.3. OBSERVABILITY OF INCREASED WEB BREAK RATES. The second question posed above in Section 5.2 concerns how the positive relationship between level of AMT and incidence of catastrophic failures evidently escaped the notice of the plants themselves. As the empirical models show, each additional level of changeover automation is associated with 6 additional web breaks per week acrossthe machines studied. It seemsreasonableto expect that a plant would take notice of an equipment change that resulted in anything near one extra web break per day. Yet no interviewee indicated that they considered the changeover automation responsible for more failures in their plant, or that there was in fact any relationship between the two. What accounts for this? In our view, the most likely explanation concerns the demandsfor greater mobility that led to most plants’ purchase of the AMT. As described above in Section 3.3, changeover automation decisions were typically spurred by the competitive need to make more paper grades and to change among them more quickly. Both of these requirements strained the capabilities of machines that had been designed for volume instead of flexibility. As they progressively increasedtheir range and mobility, they would very likely have experienced more web breaks. If a plant decided to install changeover automation in the midst of this, it would have difficulty distinguishing the AMT'S subsequentcontribution to the break rate. In other words, web break rates were probably already high (by historical standards) and rising before automation went in to many of the plants in our sample. If they were even higher afterward, they could reasonably have been attributed to the preexisting cause instead of the new one. 6. Conclusions Our results indicate that computer integration installed for the purpose of automating changeovers is associated with higher rates of catastrophic process failure in the sample of plants examined. As discussed in the previous section, the usability perspective provides a way to understand this apparently paradoxical finding. However, the present research was not able to provide direct support for any of the three usability propositions. This is because our field research did not include specific inquiries into the history of each plant’s automation efforts or a comprehensive effort to categorize, the extent to which the changeover AMT in place at each site acted to deskill the workforce. Based on interviews, anecdotal evidence, and our own observations, we believe that deskilling did take place and that the link between increased levels of changeover automation and greater web break rates is causal. Demonstrating this, and directly testing the usability propositions of Section 5.2.1, remains an exciting challenge for empirical researchers in this area. Such research would have implications that go beyond the performance of specific automation systems, the improvement of flexibility has become a critical factor in determining the long-term competitiveness of many industries. The use of computer integration to provide this flexibility is an example of a structural solution to the development of a
COMPUTER
INTEGRATION
AND
CATASTROPHIC
PROCESS
FAILURE
279
capability. Such solutions are attractive since the costs are definable and “once off” (even if the benefits must be justified partly on faith). There is growing evidence, however, that the development of capabilities to provide flexibility is best facilitated through infrastructural solutions or solutions that rely on dismantling inappropriate measurement systems, building computer integrated manufacturing systems that facilitate rather than replace human processes, nurturing the right skills in the workforce, and focusing managerial attention on the development of flexible capabilities. This implies that flexibility cannot simply be bought but must be built through a painstaking process of skill-building and organizational development. A research agenda to buttress this point with empirical findings would, in our opinion, be highly valuable.
Appendix Correlation Variable Speed Tons Scale 1 Scale 2
FREQB FREQF CIMMOB
A Used in Regression Scale 1 Analyses Scale 2 FREQB =QF
Matrix
for
Variables Tons
Width 0.8220 0.6990 0.9189 -0.3051 -0.1185 -0.0491 0.5222
Speed
0.7509 0.9388 -0.1337 -0.1418 -0.0870 0.7095
0.8888 0.4498 -0.0955 -0.1359 0.3606
-0.0025 -0.1299 -0.0983 0.5831
0.0366 -0.1118 -0.2844
0.4408 -0.1199
0.0801
Appendix Principal Components Projection
B of Raw Scale Variables Oblique Score Weights Scale 2 1.450 -0.970 -0.424 Variance Oblique Contributions
Scale 1 Tons Speed Width 0.353 0.365 0.373 Proportionate
Direct Scale 1 Scale 2 0.478 0.059
Joint 0.464 -0.001
Total 0.942 0.058
References
ADLER, P. S. (ed.) ( 1992), Technology and the Future of Work, Oxford University Press, New York. AND T. A. WINOGRAD ( 1992), “The Usability Challenge ” in Usability: Turning Technologies into Tools, P. S. ADLER and T. A. WINOGRAD, eds. Oxford University Press, New York. ATTE~ELL, P. ( 1992), “Skill and Occupational Changes in US Manufacturing” in Technology and the Future of Work, Oxford University Press, New York. BLACKER, F. AND C. BROWN (1986), “Alternative Models to Guide the Design and Introduction of New Information Technologies into Work Organizations,” Journal of Occupational Psychology, 287-313.
280
DAVID
M.
UPTON
AND
ANDREW
P. MCAFEE
BUCHANON,D. A. ( 1983), “Technological
Manufacturing BURNES, B. (1988), Employment,
BUTERA, F. ( 1984),
Imperatives and strategic choice ” in Information Technology in Processes, G. Winch, London, UK. “New Technology and Job Design: The Case of CNC,” New Technology, Work, and 2, IOO- 111. “Designing Work in Automated Systems: A Review of Case Studies” in Automation and
Work Design, F. Butera and J. E. Thurman (eds.), Elsevier Science, New York. CHEW, W. B. ( 1985), “Productivity and Change: The Short-term Effects of Investment on Factory Level Productivity” in Business Economics, Harvard University, Cambridge, MA. HAMMOND, J. ( 1992), Coordination in Textile and Apparel Channels: A Case for Virtual Ingegration, Harvard Business School Working Paper 92-007, Cambridge, MA. HAYES, R. H. AND K. B. CLARK (1985), “Exploring the Sources of Productivity Differences at the Factory Level” in The Uneasy Alliance: Managing the Productivity-Technology Dilemma, R. H. H. Kim, B. Clark, and Christopher Lorenz (eds.), Harvard Business School Press, Boston, MA. JAIKUMAR, R. ( 1986)) “Postindustrial Manufacturing,” Harvard Business Review, 69-76. KUKLA, C. D., E. A. CLEMENS, et al. (1992)) “Designing Effective Systems: A Tool Approach” in Usability: Turning Technologies into Tools, P. S. Adler and T. A. Winograd (eels.), Oxford University press, New
York.
MAJCHRZAK, A. ( 1988), The Human Side of Factory Automation, Jossey-Bass, San Francisco, CA. NADLER, D. AND M. TUSHMAN (1980)) “A Congruence Model for Diagnosing Organizational Behavior” Organizational Psychology, D. Kolb, I. Rubin, and J. McIntyre (eds.), Prentice Hall, New York. in
National Academy of Sciences Committee on the Effective Implementation of Advanced Manufacturing Technology, NAS ( 1986), Human Resource Practices for Implementing Advanced Manufacturing Technology, National Academy Press, Washington, DC. SALZMAN, H. ( 1992)) “Skill-Based Design: Productivity, Learning, and Organizational Effectiveness” in Usability: Turning Technologies into Tools, P. S. Adler and T. A. Winograd (eds.), Oxford University Press, New York. SENKER, P. J., M. VANDEVELDE, et al. ( 1988)) Electronics on the Shopjloor: A Report on Electronics Skills and Training in the Engineering Industry in England and Wales, Engineering Industry Training Board, Watford, UK. UPTON, D. M. (1993a). “Flexibility as Process Mobility: The Management of Plant Capabilities for Quick Response Manufacturing,” Journal of Operations Management, 12, 20.5-224.
-
( 1993b). “Process Range in Manufacturing: An Empirical Study of Flexibility,”
( 1994). “The Management of Manufacturing
31.
Flexibility.” California Management Review, 36, 2. ~ ( 1995). “What Really Makes Factories Flexible?,” Harvard Business Review, 74-86. WALTON, R. E. ( 1989). Up and Running: Integrating Information Technology and the Organization, Harvard Business School Press, Boston, MA. -AND G. SUSSMAN ( 1985). “From Control to Commitment in the Workplace.” Harvard Business Review,
98- 106.
ZUBOFF, S. ( 1988), In the Age
of the Smart Machine,
Basic Books,
New York.
David Upton has been on the faculty of the Harvard Business School since 1989. He teaches the elective MBA course, Designing, Managing, and Improving Operations, and has taught the required first-year MBA course in Technology and Operations Management. He is faculty chair of Harvard’s executive course on Building Competitive Advantage Through Operations. Upton graduated with honors in Engineering from King’s College, Cambridge University, and holds a Master’s degree in manufacturing from the same institution. He completed his Ph.D. in industrial engineering at Purdue University, with a doctoral dissertation on the application of Artificial Intelligence in Computer Integrated Manufacturing (CIM) systems. His current research project involves 150 manufacturing plants from around the world and aims to determine the most effective strategies for improving flexibility in manufacturing companies. He has written numerous journal and book publications on manufacturing, most recently in Management Science, Harvard Business Review, California Management Review, and Journal of Manufacturing Systems. His 1996 book, Strategic Operations: Competing Through Capabilities, is coauthored with Robert Hayes and Gary Pisano. A forthcoming book, Designing, Managing and Improving Operations, focuses on the management of information technology in operations and improvement strategies for operations managers. Professor Upton is a practicing Chartered Mechanical Engineer and a registered European Professional Engineer. Andrew McAfee is a doctoral candidate in the Technology and Operations Management Area at the Harvard Business School (HBS). His dissertation research explores the impact of information technology implementation on operational effectiveness. He is the recipient of a U.S. Department of Energy Integrated Manufacturing Fellowship for doctoral research. He is the author, with Professor David Upton, of “The Real Virtual Factory,” which appeared in Harvard Business Review, JulyAugust 1996. His principal research interests outside his dissertation topic concern the impact of new internetworked information technologies in trade and production. He consults in this field and
COMPUTER
INTEGRATION
AND
CATASTROPHIC
PROCESS
FAILURE
281
has spoken to many academic and industry audiences. In 1996, he designed and presented a series of seminars at HBS aimed at understanding innovations, such as public key cryptography, that are at the heart of the rapid expansion in electronic commerce. McAfee graduated with dual MS degrees in Mechanical Engineering and Management from the Massachusetts Institute of Technology’s (MIT) Leaders for Manufacturing Fellowship program in 1990. He also holds B.S. degrees in Mechanical Engineering and in French from MIT. Prior to coming to HBS, he worked as a consultant in operations management, working with clients in a range of industries, including aerospace, consumer electronics, white goods, and OEM electronics.
doc_896226047.pdf
The report explains about effect of advanced manufacturing technology on process stability during flexible production in a process industry.
PRODUCTION
AND OPERATIONS MANAGEMENT Vol. 7, No. 3, Fall 1998 Printed in U.S.A.
COMPUTER PROCESS
INTEGRATION FAILURE IN AN EMPIRICAL
DAVID M. UPTON
AND
AND CATASTROPHIC FLEXIBLE PRODUCTION: INVESTIGATION *
ANDREW P. MCAFEE
Graduate School of Business Administration, Harvard University Boston, Massachusetts 02163, USA
This paper empirically investigates the effect of advanced manufacturing technology on process stability during flexible production in a process industry. A sample of 61 North American fine paper plants is used to examine the relationship between the level of automation installed for controlling changes between paper grades and the incidence of paper web breaks. These web breaks are .catastrophic failures; they require the entire plant to be stopped, reinitialized, and restarted. Because a large fraction of breaks occurs shortly after changeovers, they are an important determinant of the aspect of plant flexibility, called mobility, or the ability to move between products with only small penalties. In an attempt to ensure stable and mobile production, many plants have implemented changeover automation. We find, however, that higher levels of this automation are significantly associated with higher rates of catastrophic failure among the plants studied. We suggest that this finding becomes less paradoxical when considered in light of a recent stream of research on advanced manufacturing technologies, loosely called the usability perspective. According to this perspective, automation designed and implemented with the narrow, technical goal of replacing human operators or removing their discretion over a production process is misguided, especially in environments in which requirements are changing rapidly. (MANUFACTURING FLEXIBILITY; OPERATIONS MANAGEMENT; MANUFACTURING AUTOMATION; MAN-MACHINE INTERACTION)
1. Introduction The well-documented forces of strengthened foreign competition, market fragmentation, and the continual need for innovation are combining to place manufacturers from diverse industries in a common dilemma-how to increase responsiveness without dramatically increasing costs. This mandate for flexibility is extensively documented among discrete manufacturers but is less well understood in process industries. Producers of streams, slurries, webs, and other continuous products are realizing, however, that their outputs must also be tailored to satisfy increasingly disparate customer requirements. In both process and discrete industries, however, flexibility means much more than being able to produce a wide variety of products. The taxonomy of flexibility developed
* Received October 1994; revised February 1996, January 265 1059-1478/98/0703/000$1.25
Copyright 0 1998. Production and Operations Management Society
1997, and September
1997; accepted
October
1997.
266
DAVID M. UPTON AND ANDREW P. MCAFBE
by Upton (1994) includes both range, or the number of manufacturable products, and mobility, or the ability to switch within the range with only small penalties in time, cost, or robustness. Mobility is often underemphasized, but it is an important consideration; it does a plant no good to be able to make a wide range of products if changeovers take too long, are too expensive, or lead to breakdowns on the line. The latter is a critical issue for process industries. One of the most significant barriers to running quick response process plants is that the manufacturing processes associated with these industries are often destabilized by a change of product. In steel rolling mills, these instabilities result in cobbles (buckles), which must be flame-cut out of the rolls. In the chemical industry, instability leads to poor yields and the production of impurities, which must be flushed out of tanks and pipes. In the paper industry (which is the subject of the field research presented here), the most common breakdowns are web breaks, after which the paper web must be manually torn out of the machine and then rethreaded. These failures are each catastrophic in that the process cannot simply be nudged back into place. It must be stopped, repaired, and restarted, often at great expense. Computer integrated manufacturing, or CIM, is often viewed as a tool to increase both range and mobility. Many operations have adopted CIM as a means to insure that product changeovers occur smoothly and to avoid severe interruptions. They see computer-controlled changes as powerful protection against process failure. This paper explores the relationship between CIM and catastrophic failures in one process industry setting: fine paper manufacturing. It examines the effect of automated changeover systems on the incidence of web breaks in a sample of 61 paper plants. The main finding from this investigation is that higher levels of changeover CIM are actually associated with a higher rate of catastrophic failures in the environments studied. While this result may appear counterintuitive, it is congruent with conclusions from several researchers who have studied the impact of advanced manufacturing technology ( AMT ) , such as CIM . They find that AMT, which “de-skills” operations workers, removes opportunities for learning and improvement and can lead to less efficient, more errorprone operations. The results reported here are less surprising when viewed from this perspective since the changeover automation studied here is primarily of the de-skilling variety. Section 2 outlines the motivation for the present study. Section 3 provides background on the fine paper industry, concentrating on the need for flexibility, the nature and use of changeover automation, and the phenomenon of catastrophic process failures, that is, web breaks. Section 4 presents the empirical models used to explore the relationship between automation and catastrophic failures and gives their results. Section 5 discusses these findings and introduces the usability perspective as a way to help understand them, and Section 6 concludes. 2. Motivation Much of the present knowledge base on the impact of automation in operations has come from case studies of successful and unsuccessful AMT implementations, interviews and attitudinal surveys of managers, and reviews of high-level industry data, such as employment levels, wages, and job classifications. There have been relatively few crosssectional empirical studies tying AMT implementation to salient operational performance measures [one example is Jaikumar’s ( 1986) investigation of divergent flexible manufacturing systems (FMS) performance in the United States and Japan]. This type of research, however, can be a valuable tool for addressing whether AMT implementation and use makes a significant difference in important outcomes, such as cost, quality, flexibility, and reliability. We had an opportunity to conduct such a study as part of a larger investigation of flexibility in the North American fine paper industry. This work (Upton 1993a, 1993b,
COMPUTER
INTEGRATION
AND CATASTROPHIC
PROCESS FAILURE
267
1995) examined how flexibility is achieved in a process industry, as well as the tradeoffs between productivity and various types of flexibility. The data collected for this study can also be used to examine the relationship between AMT and 1 measure of process mobility, an important category of flexibility for this industry, as explained below. Specifically, it was possible to assess the relationship between the level of computer-automated changeover equipment present in a plant and the frequency of catastrophic failures, that is, paper web breaks. Because these failures often occur after changeovers, they are important determinants of a paper plant’s mobility. And because the changeover process has been automated to varying degrees across the plants studied, the impact of different degrees of automation could be assessed. 3. Flexibility, Automation, and Catastrophic Failure in Papermaking
3.1. Background: The Fine Paper Industry The North American fine paper industry accounts for 45% of the world’s uncoated fine paper capacity. As in many industries, customers have become increasingly reluctant to hold inventory, often pushing inventory back onto those suppliers unable to deliver justin-time. Not only does this incur additional expense for the supplier, but the proliferation of products often makes inventory holding an infeasible method of providing quick response. To meet these competitive imperatives, firms are anxious to build supporting capabilities at the plant level. The ability to provide quick response to customers demands capabilities in many areas, as well as in the manufacturing shop itself. Hammond ( 1992) describes a number of these capabilities. There is, however, no substitute for a responsive plant, which can switch between products quickly and reliably to respond to growing demand volatility and product proliferation. 3.1.1. THE PAPERMAKING PROCESS. A paper mill usually comprises a pulp plant, which provides pulp for a group of 2 to 10 converting plants or machines (the terms are used interchangeably), which turn the pulp into paper. The papermaking process is essentially a water-removal procedure. Water is extracted from a pulp slurry by use of gravity, pressure, and heating. The pulp slurry is laid onto a moving fabric, from which water drains off. The pulp web is squeezed and heated by a series of rollers until it is strong enough to support its own weight. Further roller heating then takes place until the paper moisture content is just below that of the ambient atmosphere. The paper is then collected on a reel in the plant to be subsequently sliced transversely and longitudinally into roll sizes convenient for customers and sheeting machines. Figure 1 is a simplified view of a paper plant. 3.1.2. INDUSTRYHISTORYANDCOMPETITIVEENVIRONMENT For the pastcentury,North American paper conversion plants have been steadily increasing in size, output rate, and investment requirements; Figure 2 displays this trend. The paper industry has traditionally competed primarily on cost, and economies of scale have been significant. Larger plants, therefore, have continued to be more efficient at producing standard paper grades. Companies have historically accumulated funds through depreciation and retained earnings and leapfrogged the competition with bigger, more expensive plants, usually on an existing site (the minimum efficient scale for pulp mills exceeds that of converting plants, so a number of plants tend to be co-located with a pulp mill; incremental expansion is thus considerably cheaper than a green field site.) These plants then concentrate on satisfying demand for standard paper types, for example, xerographic paper for offices.
3.2. Flexibility in Papermaking
The existence of a few large commodity plants at any point in time means that the majority of converting plants are unable to compete on the basis of price for common
268
DAVID
M.
UPTON
AND
ANDREW
P. MCAFEE
Water Removal Method:
Gravity Pressure Heat
RESS
SECT4ON
RE DRYER
SECTfcSN
FIGURE1. Simplified Cross-Sectional View of a Typical Paper Plant. Water is Removed by Gravity, Pressure, and, Finally, Heat.
paper types. After their scale economies have vanished, these facilities often move to producing other grades of fine paper, which have volumes too low to attract the attention of the industry’s giants. Because demand for these products is typically more fickle, and because competition to fill it is intense among the smaller plants, they face great pressure to become more flexible. They must both be able to make a wider range of products and to move among them quickly and easily. This need translates directly into a requirement to manage paper grade changeovers more effectively. Different paper grades are equivalent to different products. There are 4 kinds of grade changes made in a paper converting plant, listed here in decreasing order of difficulty.
1200
-i
0 0
0~
r
I
I
‘,
’
I
I
I
’
4~1~
I
/
I
1890
1900
1910
1920
1930
1940
1950
1960
1970
1980
1990
2000
Date of Original Installation
FIGURE2. Plant Size Versus Installation Date.
COMPUTER
INTEGRATION
AND
CATASTROPHIC
PROCESS
FAILURE
269
Color Changes: These often require a complete plant wash down, to remove any traces of the previous color. Very few plants in the sample made a variety of colors, and these changes were, therefore, not explored in this study. Furnish Changes: These require changes in the composition of the pulp and entail resetting valves, drives, motors, etc., in order, for example, to brighten the paper with titanium dioxide or to vary the ratio of long-fiber to short-fiber pulp. Furnish changes often incorporate basis weight changes (the following category). A furnish weight change alone is more difficult than a basis weight change alone. Basis Weight Changes: Basis weight changes involve altering the rate of pulp deposition on the web to vary the area density of the paper. Changes usually require alterations in other plant parameters, such as web speed. Basis weight changes, which are part of broader furnish changes, are not considered separately in this research. Caliper Changes: These are minor adjustments to the thickness of the web at the dry end of the plant and may not often be recorded by the mill. The two types of changeover considered in this research are basis weight and furnish changes. These are the key dimensions that distinguish one type of paper from another, and most plants studied made both types of changeover regularly. As discussed above, the ability to switch between grades quickly and reliably has become an important task for manufacturing managers-indeed, it has become a critical issue for the survival of many small and medium size plants. According to interviews, changeovers can absorb as much as 20% of productive time and have become one of the most important determinants of a plant’s performance. 3.3. AMTfor Flexibility: Changeover Automation To address the increasing need for mobility, that is, rapid and reliable changeovers, many plants have installed computer integration systems designed to control the gradechange procedure. These systems communicate with lower-level control devices in the control hierarchy, such as valve and motor controllers, and coordinate their actions in real time during a changeover. In doing so, they generally substitute for more manual operatorcontrolled changes. While almost all plants in the sample were fitted with automation to control process settings during steady-state operation, there was a wide disparity in the degree of automation for the process of changing grades. Some machines had automation that controlled all changeovers; others had automation for only specific types of changes, for example, basis weight changes, and some relied solely on direct operator control of the changeover process. At the sites with automated changeover systems, the general view, as expressed in interviews, was that they were very much a black-box; that is, no one in the plant knew quite how their system worked. Programming was carried out by the system vendor, as were any modifications required during installation. During automated changeovers, the machine’s crew stood by in case of emergencies but had no active role. This pattern corresponds closely to the observations of Zuboff ‘s ( 1988) about the the relative roles of AMT and workers in a number of paper plants. She found that new technologies in this industry consistently relegated workers to marginal roles within the production process. In contrast, nonautomated changeovers demanded that members of a crew work closely together to coordinate the many required adjustments to speeds, separations between rollers, level of additives, etc. This was a demanding task, especially since few machines were designed for flexible operation. Interviews provided anecdotal evidence, however, that crews were able to learn how to make changeovers rapidly, in some cases, more rapidly than the automation installed for the task. As one operator put it,
The computer is real slow-it gets kind of boring sitting watching [the plant run] out a better way to change manually. We’ve got it pretty much figured out now-we so we figured all practiced
270
DAVID
M.
UPTON
AND
ANDREW
P. MCAFEE
as a team, [. .] worked out who should do what and when. We can always beat the computer, and do changeovers faster than they used to do them manually-it’s just a matter of working out the right routine.
This kind of tinkering with the grade changing process, even if it is successful at lowering average changeover times, could conceivably involve some risk; experimentation may precipitate a web break. Since automated systems perform the changes in the same way each time, one might expect them to make more reliable changes, even if they are made more slowly. 3.4. Catastrophic Failure: Paper Web Breaks Breaks in the web require operators to climb into the machinery, tear out the broken web, and rethread the plant. Little automation is available to help with the restart process because recovery procedures vary greatly, based on factors such as where the break occurred in the machine and the amount of snarled paper. In other words, there is no prespecifyable sequence of steps that can be automated; the operators must assess and act on each situation individually. Most of the time during a restart is absorbed with tearing out the old tom web and manually throwing the leading edge of the new web into the pinch rollers. Restarting after a break takes between 30 minutes and 4 hours, during which the machine produces no usable output. Restarting can be dangerous for the people involved, who must physically manipulate the flimsy paper web in a very hot, moist paper plant, which is often under power. Breaks are indeed catastrophic, and plant personnel strongly prefer to avoid them. Operators’ vocal reaction to a break is fairly predictable. What is less predictable, however, are the circumstances giving rise to a break. During interviews, many operators tied web breaks to changeovers. As one supervisor said,
If we’re going to get one it just happens! it’s hard there’s something wrong don’t really know what’s to figure it out. [a break] it’ll happen right after we change grades. Sometimes though, to tell if it was because of the grade change or some other reason. If with the pulp we’ll often get a break. More often than not though, we caused the problem. People are usually too busy getting the break fixed
This suggests that a web break is a poorly understood event, with some identifiable determinants and some that are, at least according to operators, random. The precise mechanisms that produce breaks are not well understood, and predictive models do not exist. Even so, several factors that appear to contribute to web break rates, beyond changeover frequency, have been identified. They are discussed below in Section 4.2.2. AMT is not one of these identified factors. To our knowledge, changeover automation has not been linked to web breakage, either empirically or anecdotally. The purpose of the present investigation was to explore this relationship. 4. Empirical 4.1. Field Research We studied 61 fine paper plants (in 25 mills and 11 companies) over a period of 2 years. To provide background, structured and unstructured interviews were carried out at the following three levels: vice-president for operations or COO for the company, plant manager, and paper plant supervisor and operator. We also observed operations in 15 plants and worked alongside operators on the equipment to examine the way in which process changes were made. Over 20 paper break events were physically observed during the course of the field research. Research: Design and Results
COMPUTER
INTEGRATION
AND CATASTROPHIC
PROCESS FAILURE
271
4.2. Data Collection 4.2.1. DEPENDENT VARIABLE. Web Break Frequency: In each of the 61 plants, operators were asked to provide data on the average number of paper breaks experienced in a plant each week. A paper break is an important occurrence on a shift, and operators’ estimates of break frequencies were found to be very reliable in the few plants where records of breaks were kept. The model variable breaks captures the estimated number of breaks per week. 4.2.2. INDEPENDENT VARIABLES. Changeover Automation: The variable CIMMOB ( CIM for mobility) represents the degree to which changes in grades could be made automatically through computer integration of other computerized subsystems. The degrees of automation for basis weight, caliper, and furnish changes were each scored out of seven (seven being completely automatic and zero being manual). These three scores were then added and divided by seven, to give a maximum possible score of 3, corresponding to three fully automated methods of changing grade. Color change systems were excluded since most plants in the sample did not make color changes. Change Frequency: There was widespread agreement that many web breaks occurred after changeovers, so it was important to include change frequency as a predictor. The weekly number of grade changes was estimated by the operators separately for basis weight and furnish changes. FREQB is the number of grade changes per week in which the primary dimension of change is the basis weight of the paper. FREQF is a similar measure for changes in furnish. Scale: Another factor likely to be a determinant of break rates is the scale of the plant. Larger plants have larger components moving at higher speeds and are therefore more difficult to control dynamically. The scale of a plant can be measured in the following three ways: Output, or tons produced per day Plant maximum rated speed (measured in feet per minute) Plant width (measured in inches) Data were collected from the plants on each of these metrics. Speed is the maximum speed at which a plant is rated to run, while width is the trim width of the plant. The variable tons represents the average daily output of the plant. The three scale measures are strongly intercorrelated (see Appendix A), so it may not be possible to ascribe explanatory power to any individual measure in a regression model. Each piece of data does, however, provide information concerning the scale of the plant. For this reason, these measures were combined by projecting them onto (oblique) principal components of their joint distribution. The details of this transformation are shown in Appendix B. The primary principal factor, ScaZe 1, is representative of the scale of the plant; tons, speed, and width all contribute positively to it. The second factor, Scale 2, has a positive contribution for tons and negative contributions from speed and width. It therefore measures one type of disparity among these three measures. A plant which, compared to others in the sample, has high output but low width and rated speed will have a high value for Scale 2. This variable, then, captures the degree to which a plant is running “full out,” or achieving the most output possible given its physical configuration. Age ofchangeover automation: Chew ( 1985) and Hayes and Clark ( 1985) have shown that the installation of new process technology can, in many circumstances, be followed by a short-term productivity dip as an operation, and its operators and managers become accustomed to the new technology and overcome start-up problems. It is possible that if much of the changeover automation at the plants studied was installed contemporaneously and shortly prior to this study, then any problems relating to it, such as a tendency toward more web breaks, would be transient manifestations of this phenomenon. To separate explicitly any such effects from the main relationship of interest, the length of time since the last major upgrade of the control system at the plant was calculated as the variable
272
DAVID M. UPTON AND ANDREW P. MCAFEE
CIMAGE. If this predictor is significant in the regression models, support would exist for a productivity dip hypothesis. Since only five of the plants had installed or upgraded their changeover automation within the previous 2 years, it is unlikely that the sites as a group are experiencing a productivity dip. Table 1 shows descriptive statistics for dependent and independent variables.
4.2.3. DISCUSSION OF DESCRIPTIVE DATA. The dependent variable, breaks, and the 2 independent variables measuring change frequency are based on estimates from plant personnel rather than direct observation or historical records. The objectivity of these measures cannot be guaranteed, but they were the best available metrics; very few plants kept records of their break rates. The most striking feature of the descriptive data is the very large variation in both independent and dependent variables. Break frequencies range from 1 to 60 per week and furnish changes from 0 to 250 per week. Furnish changes are much more common than basis weight changes, although it must be remembered that furnish changes often include a basis weight change, while a straightforward change in basis weight does not include a furnish change. 4.3. Linear RegressionModels
We used a series of nested linear regression models to estimate the impact of several predictors on the frequency of paper breaks for our sample of plants. Our independent variables were divided into control predictors, which were used to control for factors unrelated to our research question, and a single question predictor, that is, CZMMOB, the level of automatic grade change equipment. The dependent variable in all models was the number of breaks per week. Control predictors include the frequency of change demanded of a plant, multiple measures of the scale of the plant, and the length of time since the last major automation upgrade. A number of other more complex specifications were explored, as described later, but did not add significantly to the explanatory power of the model. 4.3.1. RESULTS. Tables 2 and 3 show parameter estimates for the models tested. Model 1 includes the only control predictors of FREQB and FREQF. In this model, furnish change
TABLE 1 Variable Descriptive Statistics Variable Breaks
CIMMOB
Description Break frequency Degree of changeover automation Basis weight change frequency Furnish weight change frequency Trim width Maximum plant speed Net output Scale principal Factor 1 Scale principal Factor 2 Time since last major upgrade of control svstem
Units Breaks/week
N 61 60
Mean 17.3 1.87 7.53 22.86 160.4 1457 230.81 0 0 20.33
SD 14.6 0.81 4.95 56.77 66.0 767 248.7 1 1 20.50
Min 1.0 0.43 0 0 72.0 250 18.0 -1.31 -1.04 2
Max 60.0 3 22.5 250 330.0 3200 1200 2.534 6.521 74
FREQB FREQF
Number/week Number/week In. Ftlmin Tons/day
60 61 60 60 61 60 60 51
Width Speed Tons Scale 1 Scale 2
CIMAGE
Years
COMPUTER
INTEGRATION
AND
CATASTROPHIC 2 for Models Model 2
PROCESS
FAILURE
273
TABLE Regression Model Coefficient GE) 10.99*** (2.58) .346 (.314) .164*** (.027) Parameter 1 Std. Coeff.
Estimates
l-3 Model Std. Coeff. Coefficient GE) 3.03 (5.28) .461 (.287) .217*** (.033) -.036 (.033) .025** (.OOS) ,003 (.003) ,090 (.068) 56 ,613 (.565) 49 12.928 ,000 3 Std. Coeff.
Independent Constant
FREQB FREQF
Variable
Coefficient (SE) 7.95 (4.22) ,434 (.284) .175*** (.024) -.048 (.034) .024** (.008) .003 (.003) 59 .633 (.599) 53 18.310 ,000
.116 .637***
Width Tons Speed
CIMAGE
.143 .680*** -.214 .401** ,158
.160 .625*** -.180 .457** .174 .135
Sample size R-squared (adj. R sq’d.)
DOF
F-value P (F > F*) ***p < ,001. **p < .Ol. *p < .05.
60 .482 (.464) 57 26.516 .ooo
frequency is found to be highly significant, while basis weight change frequency is not significant. However, these two variables have a relatively high correlation (=0.44), so both are kept in later models. This model explains 48.2% of the observed variance of
breaks.
Model 2 adds the three raw scale variables; tons is found to be significant, while speed and width are not. However, as discussed above, these three measures are very highly intercorrelated (see Appendix A), and it would probably be misleading to ascribe all
TABLE Regression Model Coefficient (SE) 2.07 (4.25) .617* (.265) .160*** (.023) -.039 (.031) .031*** (.008) -.004 (.004) Parameter 4 Std. Coeff. 3
Estimates
for Models
Model 5
4-6 Model Std. Coeff. Coefficient (SE) -1.10 (4.17) .613* (.264) .162*** (.022) 6 Std. Coeff.
Independent Constant
FREQB FREQF
Variable
Coefficient (SE) -1.75 .639* .198*** -.030 .031*** -.003 (5.15) (.271) (.032) (.031) (.007) (.004)
Width Tons Speed Scale1 Scale2
CIMAGE CIMMOB
.201* .625*** -.175 .523*** -.195
.221* ..573*** -.148 .570*** -.148
.20* .632***
2.15 (1.42) 4.47*** (1.19) 6.04** (2.18) 58 .703 (.668) 51 20.072 ,000 .332** .087 (.063) 5.01* (2.19) 55 ,678 (.631) 47 14.167 BOO .131 .306* 5.30** (1.84) 58 .700 (.671) 52 24.28 .ooo
.146 .305*** .292**
Sample size R-squared (adj. R sq’d.)
DOF
F-value P (F > F*) ***p < .OOl. **p < .Ol. *p < .05.
274
DAVID M. UPTON AND ANDREW P. MCAFEE
explanatory power to only one of them. For this reason, all three scale variables are kept in subsequentmodels. This model explains 63.3% of the observed variance of breaks. Model 3 includes CIMAGE. It is not a significant predictor of web break frequency when the predictors listed above are included in a model. This indicates that the plants studied are not, as a group, experiencing a productivity dip due to recently installed automation. A dummy variable that captured those plants that had installed or upgraded changeover automation within the previous 2 years was also tested with this model. Like CIMAGE, it was not a significant predictor, providing further support against the productivity dip hypothesis. Model 4 includes the question predictor, CIMMOB . It is highly significant (P < .Ol ) and has a magnitude of 6.04, indicating that each additional level of changeover automation is associatedwith an additional 6 breaks per week acrossthe plants studied. Interestingly, the inclusion of CIMMOB in this model results in FREQB becoming a significant predictor (P > .05 ) ; this variable was not significant in any previous model. It appears,then, that including CIMMOB in a regressionmodel helpsto clarify the relative contributions of furnish weight and basisweight changesto web break incidence. Model 4 explains 70.3% of the observed variance in breaks. Its final form is Breaks (est.) = 2.07 + 0.617 * FREQB + 0.160 * FREQF - 0.039 * width
- 0.004 * speed + 0.031 * tons + 6.04 * CIMMOB
As a final check, Model 5 adds the variable CZMAGE to Model 4 as a further check of the productivity dip hypothesis. Again, this variable is not significant, and it does not affect the coefficients of previously significant variables outside of their standard errors. It therefore does not appear that CIMMOB, the predictor of interest, is proxying
for
CIMAGE.
As an additional empirical exploration, we substituted Scale 1 and Scale 2 for tons, speed,and width in our final model. As discussedabove, these two variables are orthogonal projections of the three raw scale measures;Scale 1 is a positive combination of all three and thus representsthe total scaleof the plant, while Scale 2 measuresdifferences between them and captures the degree to which a plant is running “full out,” or achieving the most output possiblegiven its physical configuration. As Model 6 shows,it is this measure that is a highly significant predictor of breaks, while Scale 1 is not significant. Other regressionparametersremain largely unchanged. These results indicate that it is how much output plants are trying to wring from a given configuration, rather than the sheer plant scale, that drives web break rates. 4.3.2. SENSITIVITY TO OUTLIERS. To verify the robustnessof these findings, we performed an analysis of the sensitivity of Model 4, our final model, to the presence of outliers. Three observations had a value for Cook’s distance (which summarizesleverage and residual magnitude) greater than five times its mean. When these variables were removed and Model 4 (our final model) was reestimated, significant variables were unchanged, their coefficients remained stable, and the percentage of observed variance in breaks explained by the model increased to 82.1%. We conclude that our final model appearshighly robust to outliers. 4.3.3. OTHER VARIABLES. A number of alternative specifications of the models above were also explored. However, no significance was found (nor was the coefficient associated with computer integration substantially altered) by the addition of the following variables: Workforce experience; Average basis weight; Plant age;
COMPUTER
INTEGRATION
AND CATASTROPHIC
PROCESS FAILURE
275
Plant range (difference between highest and lowest basis weight) ; and The degree of computer integration elsewhere in the plant (for quality control for example) or the mill as a whole. 5. Discussion of Results 5.1. Effect of Control Predictors 5.1.1. FURNISH CHANGES. Model 1 in Table 2 shows that 48% of the variance in process failure rate is explained by the changes made per week on the dimensions of furnish and basis weight. However, the statistical strength of the estimate associated with basis weight changes is low. The statistical effect of furnish change frequency is, however, clearly established. In our final model, Model 4, both types of changeover are significant predictors, with the standardized coefficient for furnish changes approximately three times as large as that for basis weight changes. The reason for the disparity between the types of change is that furnish changes are much more disruptive of the process. Changes in basis weights, it appears from anecdotal evidence, are only likely to cause breaks when the change takes place with very light papers. The addition of the average basis weight made by the plant to Model 1 did, however, not increase its explanatory power and was found to be an insignificant factor in predicting failure rates. 5.1.2. PLANT SCALE. As Model 6 shows, Scale 2 is a highly significant predictor of breaks, while Scale 1 is not significant. This indicates that it is how much output plants are trying to squeeze from a given configuration, rather than the sheer plant scale, that drives web break rates. Results from Model 4 suggest that for a given speed and width, net tonnage is a primary driver of breaks. This is of interest in itself, though caution is necessary since the three factors are highly collinear (see Appendix A). As Models 4 and 6 show, controlling for scale does not interfere with the effect of interest for this work, which is the positive relationship between high levels of computer integration and process failure rates. 5.2. Effect of Changeover Automation As Model 4 demonstrates, the variable CIMMOB is a highly significant predictor of web break frequency after controlling for other plausible predictors, with each additional level of changeover automation associated with an estimated additional 6 breaks per week. This is, in many ways, a surprising and counterintuitive finding. This type of computer integration was installed specifically to increase process mobility, yet it has the effect within the sample of substantially increasing the penalty associated with moving between products. Furthermore, anecdotal and empirical evidence suggests that the automation is not even faster than an experienced crew at effecting changes (Upton 1995). This form of AMT, then, appears to be having the opposite of its intended effect for the plants studied. The questions immediately suggested by this finding are 1. What explains this result? 2. Why have the plants in the sample not noticed the deleterious effects of their changeover automation? Why has this form of computer integration not been widely declared a failure and removed from the machines? 5.2.1. THE USABILITY PERSPECTIVE. In our view, an answer to the first question, and a plausible explanation for these results, is suggested by research within the usability perspective. Usability research has addressed issues of AMT success and failure and the relationship between these outcomes and the ways that worker skills changed with the introduction of automating technologies. This research has embraced a variety of disciplines, including labor relations, operations management, organizational behavior, and sociology. The conclusions
276
DAVID M.UPTON
AND ANDREW P.MCAF%E
arising from these efforts, however, are remarkably consistent. Principally, they are that most companies do not anticipate or adequately plan for the organizational consequences of AMT, including the new and different skills that will be required of workers; that technology that aims primarily to replace or deskill workers is misguided and more prone to failure; and that AMT, which is designed and implemented to broaden the roles of workers instead of constraining them, is more likely to be successful. Each of these propositions has been supported by empirical research, as summarized below. (1) Most companies do not anticipate or adequately plan for the organizational consequences O~AMT. Several studies (Majchrzak 1988, NAS 1986, Butera 1984, and Senker et al. 1988) have found that firms typically embarked on AMT implementation projects without understanding organizational considerations, including the consequences for worker skill levels. Instead, these efforts were undertaken to reduce costs or maintain competitive parity and were treated as straightforward capital investments. A review of engineering and operations management textbooks published between 1939 and 1989 (Salzman 1992) supports these findings; it showed that only a minority of the texts made any mention of worker’s roles in automated manufacturing and equipment design and that these treated workers as marginal contributors, whose roles were to be minimized by automation. As Attewell (1992) points out, this approach sometimes winds up requiring a more skilled workforce because the resulting machinery can be more error-prone. A survey of large metalworking firms (Salzman 1992) found that majority did not have policies of worker participation during AMT design efforts, or during post-implementation design changes or discussions on worker modifiability and maintenance. One review of technology design models (Blacker and Brown 1986) summarized that ergonomics, sociotechnical theory, and theories of participative management have had little impact on AMT design because the climate of opinion is not receptive to them. All of this work supports the contention that over much of the history of AMT, decisions on equipment and technologies were made without considering their impact on people, jobs, and skills. (2) Technology that aims primarily to replace or deskill workers is misguided and more prone tofailure. The consequences of a narrow, exclusively technical view of AM-r impact have been articulated by a number of researchers. Adler ( 1992) characterizes the view that AMT will allow not only fewer workers, but ones that are less skilled and doing narrower jobs, as the deskilling myth and lists the following as reasons why this view is inappropriate in most circumstances. It ignores cognitive processes such as reasoning, as well as social interactions. It fails to take into account that people interpret situations and can adaptively monitor and change them. It takes the characteristics of the equipment as central and those of the involved humans as peripheral. It does not involve users in the design process, relying instead on technical experts. Adler concludes that treating AMT as a means of lessening reliance on operators’ abilities is misguided because “. . .in the majority of cases, the effective use of new technologies requires a workforce that is more skilled, not less.” Zuboff ( 1988) includes case studies of comprehensive AMT implementation in 3 papermaking facilities. These studies contain rich descriptions of the ways in which operators were removed from direct physical experience with the production tasks and the harmful consequences of this separation. In summarizing these, Zuboff states that “It was as if one’s job had vanished into a two-dimensional space of abstractions, where digital symbols replace a concrete reality. Workers reiterated. . .feelings of loss of control, of vulnerability, and of frustration. [The] oft-repeated metaphor spoke of being robbed of one’s senses and plunged into darkness.”
COMPUTER
INTEGRATION
AND CATASTROPHIC
PROCESS FAILURE
277
K&la, Clemens et al. ( 1992) make a clear statement of the dangers of an overly narrow view of AMT ; they maintain that such automation strategies freeze the production processes and division of labor and make plants unable to improve their systems or adapt them to changing needs and system requirements. (3) AMT, which is designed and implemented to broaden the roles of workers instead of constraining them, is more likely to be successful. Adler and Winograd ( 1992) call this the “usability challenge” and state that “The key challenge in designing new technologies is how to best take advantage of user’s skills in creating the most effective and productive working environment.” Kukla, Clemens et al. ( 1992) advocate designing systems as tools for skilled work that take advantage of the investment already made in people. These systems do not constrain operators to follow routines but allow them to respond quickly and effectively to constantly changing combinations of events. Majchrzak ( 1988)) expanding on an open systems model developed by Nadler and Tushman ( 1980)) proposes a framework that takes into account the task to be performed, formal organizational structure, the individuals who will work with the AMT, their informal organization, and the interactions among these elements. 5.2.2. APPLICATIONOFTHEUSABILITYPERSPECTIVETOAMTWITHINPAPERMAKING. The first of these propositions, that AMT investments have traditionally been made without taking into account organizational and human considerations, can help explain how North America’s papermakers made the decision to install changeover.automation. In a plant just beginning to face the need for greater flexibility after competitors had removed its scale advantages, the workforce would likely have been inexpert at effecting changeovers. In this situation, computer integrated changeover equipment would have appeared an attractive alternative. The immediate and narrowly defined technical benefits would have been substantial, and no drawbacks would have been apparent unless issues such as organizational learning and skill acquisition were explicitly considered. As the researchers who advanced Proposition 1 have indicated, it is unlikely that these topics were part of the decision to implement changeover automation. The case studies of Zuboff (1988) of AMT implementation in paper plants clearly indicate that their automation decisions were made without fully considering how the new technology would change the nature of work. The second usability proposition, that deskilling AMT can degrade operational performance instead of improving it, sheds some light on the present study’s findings. The principal result from the models tested, that higher levels of changeover automation are strongly associated with higher web break rates, ceases to be counterintuitive when viewed in light of this proposition. As the operator’s quote in Section 3.3 indicates, the changeover automation installed in many plants relegating workers to the role of bystanders during grade changes. In doing so, the computer-integrated equipment may have removed their opportunities to learn about and improve the process through experimentation, communication, or simple experience. As Kukla, Clemens et al. ( 1992) point out, this lack of learning becomes especially dangerous when system requirements or market needs change. Section 3.2, above, outlines that this is exactly what has happened to most paper plants since they were built; they have faced ongoing requirements to become more flexible and, particularly, more mobile. Many plants in our sample attempted to meet this need by installing changeover automation; the mean value for the CZMMOB is almost two-thirds the maximum possible value. If this AMT were, on the whole, deskilling, Proposition 2, above, indicates that higher levels of CIMMOB could be associated with poorer performance. Thus, the usability perspective suggests an answer to the first question posed above in Section 5.2. Based on our empirical results, and on the interviews and field observations described earlier, we
278
DAVID M. UPTON AND ANDREW P. MCAFEE
find it plausible that the papermakers in our sample installed changeover automation that was primarily deskilling, and that higher levels of this type of AMT lead to a higher incidence of paper web breaks. Proposition ( 3)) that AMT that avoids the deskilling myth is more likely to be successful, was not directly investigated by the present research. As far as we could determine, all of the changeover automation installed in the plants studied was intended to decrease human involvement in the changeover process. We saw no cases in which it had been designed and implemented with the sustained participation of the workforce, or where it was intended to support what Walton ( 1989) terms “commitment” rather than “control”. It was thus not possible to test, even anecdotally, whether this approach could yield better results along any dimension of mobility, including web break rates. 5.2.3. OBSERVABILITY OF INCREASED WEB BREAK RATES. The second question posed above in Section 5.2 concerns how the positive relationship between level of AMT and incidence of catastrophic failures evidently escaped the notice of the plants themselves. As the empirical models show, each additional level of changeover automation is associated with 6 additional web breaks per week acrossthe machines studied. It seemsreasonableto expect that a plant would take notice of an equipment change that resulted in anything near one extra web break per day. Yet no interviewee indicated that they considered the changeover automation responsible for more failures in their plant, or that there was in fact any relationship between the two. What accounts for this? In our view, the most likely explanation concerns the demandsfor greater mobility that led to most plants’ purchase of the AMT. As described above in Section 3.3, changeover automation decisions were typically spurred by the competitive need to make more paper grades and to change among them more quickly. Both of these requirements strained the capabilities of machines that had been designed for volume instead of flexibility. As they progressively increasedtheir range and mobility, they would very likely have experienced more web breaks. If a plant decided to install changeover automation in the midst of this, it would have difficulty distinguishing the AMT'S subsequentcontribution to the break rate. In other words, web break rates were probably already high (by historical standards) and rising before automation went in to many of the plants in our sample. If they were even higher afterward, they could reasonably have been attributed to the preexisting cause instead of the new one. 6. Conclusions Our results indicate that computer integration installed for the purpose of automating changeovers is associated with higher rates of catastrophic process failure in the sample of plants examined. As discussed in the previous section, the usability perspective provides a way to understand this apparently paradoxical finding. However, the present research was not able to provide direct support for any of the three usability propositions. This is because our field research did not include specific inquiries into the history of each plant’s automation efforts or a comprehensive effort to categorize, the extent to which the changeover AMT in place at each site acted to deskill the workforce. Based on interviews, anecdotal evidence, and our own observations, we believe that deskilling did take place and that the link between increased levels of changeover automation and greater web break rates is causal. Demonstrating this, and directly testing the usability propositions of Section 5.2.1, remains an exciting challenge for empirical researchers in this area. Such research would have implications that go beyond the performance of specific automation systems, the improvement of flexibility has become a critical factor in determining the long-term competitiveness of many industries. The use of computer integration to provide this flexibility is an example of a structural solution to the development of a
COMPUTER
INTEGRATION
AND
CATASTROPHIC
PROCESS
FAILURE
279
capability. Such solutions are attractive since the costs are definable and “once off” (even if the benefits must be justified partly on faith). There is growing evidence, however, that the development of capabilities to provide flexibility is best facilitated through infrastructural solutions or solutions that rely on dismantling inappropriate measurement systems, building computer integrated manufacturing systems that facilitate rather than replace human processes, nurturing the right skills in the workforce, and focusing managerial attention on the development of flexible capabilities. This implies that flexibility cannot simply be bought but must be built through a painstaking process of skill-building and organizational development. A research agenda to buttress this point with empirical findings would, in our opinion, be highly valuable.
Appendix Correlation Variable Speed Tons Scale 1 Scale 2
FREQB FREQF CIMMOB
A Used in Regression Scale 1 Analyses Scale 2 FREQB =QF
Matrix
for
Variables Tons
Width 0.8220 0.6990 0.9189 -0.3051 -0.1185 -0.0491 0.5222
Speed
0.7509 0.9388 -0.1337 -0.1418 -0.0870 0.7095
0.8888 0.4498 -0.0955 -0.1359 0.3606
-0.0025 -0.1299 -0.0983 0.5831
0.0366 -0.1118 -0.2844
0.4408 -0.1199
0.0801
Appendix Principal Components Projection
B of Raw Scale Variables Oblique Score Weights Scale 2 1.450 -0.970 -0.424 Variance Oblique Contributions
Scale 1 Tons Speed Width 0.353 0.365 0.373 Proportionate
Direct Scale 1 Scale 2 0.478 0.059
Joint 0.464 -0.001
Total 0.942 0.058
References
ADLER, P. S. (ed.) ( 1992), Technology and the Future of Work, Oxford University Press, New York. AND T. A. WINOGRAD ( 1992), “The Usability Challenge ” in Usability: Turning Technologies into Tools, P. S. ADLER and T. A. WINOGRAD, eds. Oxford University Press, New York. ATTE~ELL, P. ( 1992), “Skill and Occupational Changes in US Manufacturing” in Technology and the Future of Work, Oxford University Press, New York. BLACKER, F. AND C. BROWN (1986), “Alternative Models to Guide the Design and Introduction of New Information Technologies into Work Organizations,” Journal of Occupational Psychology, 287-313.
280
DAVID
M.
UPTON
AND
ANDREW
P. MCAFEE
BUCHANON,D. A. ( 1983), “Technological
Manufacturing BURNES, B. (1988), Employment,
BUTERA, F. ( 1984),
Imperatives and strategic choice ” in Information Technology in Processes, G. Winch, London, UK. “New Technology and Job Design: The Case of CNC,” New Technology, Work, and 2, IOO- 111. “Designing Work in Automated Systems: A Review of Case Studies” in Automation and
Work Design, F. Butera and J. E. Thurman (eds.), Elsevier Science, New York. CHEW, W. B. ( 1985), “Productivity and Change: The Short-term Effects of Investment on Factory Level Productivity” in Business Economics, Harvard University, Cambridge, MA. HAMMOND, J. ( 1992), Coordination in Textile and Apparel Channels: A Case for Virtual Ingegration, Harvard Business School Working Paper 92-007, Cambridge, MA. HAYES, R. H. AND K. B. CLARK (1985), “Exploring the Sources of Productivity Differences at the Factory Level” in The Uneasy Alliance: Managing the Productivity-Technology Dilemma, R. H. H. Kim, B. Clark, and Christopher Lorenz (eds.), Harvard Business School Press, Boston, MA. JAIKUMAR, R. ( 1986)) “Postindustrial Manufacturing,” Harvard Business Review, 69-76. KUKLA, C. D., E. A. CLEMENS, et al. (1992)) “Designing Effective Systems: A Tool Approach” in Usability: Turning Technologies into Tools, P. S. Adler and T. A. Winograd (eels.), Oxford University press, New
York.
MAJCHRZAK, A. ( 1988), The Human Side of Factory Automation, Jossey-Bass, San Francisco, CA. NADLER, D. AND M. TUSHMAN (1980)) “A Congruence Model for Diagnosing Organizational Behavior” Organizational Psychology, D. Kolb, I. Rubin, and J. McIntyre (eds.), Prentice Hall, New York. in
National Academy of Sciences Committee on the Effective Implementation of Advanced Manufacturing Technology, NAS ( 1986), Human Resource Practices for Implementing Advanced Manufacturing Technology, National Academy Press, Washington, DC. SALZMAN, H. ( 1992)) “Skill-Based Design: Productivity, Learning, and Organizational Effectiveness” in Usability: Turning Technologies into Tools, P. S. Adler and T. A. Winograd (eds.), Oxford University Press, New York. SENKER, P. J., M. VANDEVELDE, et al. ( 1988)) Electronics on the Shopjloor: A Report on Electronics Skills and Training in the Engineering Industry in England and Wales, Engineering Industry Training Board, Watford, UK. UPTON, D. M. (1993a). “Flexibility as Process Mobility: The Management of Plant Capabilities for Quick Response Manufacturing,” Journal of Operations Management, 12, 20.5-224.
-
( 1993b). “Process Range in Manufacturing: An Empirical Study of Flexibility,”
( 1994). “The Management of Manufacturing
31.
Flexibility.” California Management Review, 36, 2. ~ ( 1995). “What Really Makes Factories Flexible?,” Harvard Business Review, 74-86. WALTON, R. E. ( 1989). Up and Running: Integrating Information Technology and the Organization, Harvard Business School Press, Boston, MA. -AND G. SUSSMAN ( 1985). “From Control to Commitment in the Workplace.” Harvard Business Review,
98- 106.
ZUBOFF, S. ( 1988), In the Age
of the Smart Machine,
Basic Books,
New York.
David Upton has been on the faculty of the Harvard Business School since 1989. He teaches the elective MBA course, Designing, Managing, and Improving Operations, and has taught the required first-year MBA course in Technology and Operations Management. He is faculty chair of Harvard’s executive course on Building Competitive Advantage Through Operations. Upton graduated with honors in Engineering from King’s College, Cambridge University, and holds a Master’s degree in manufacturing from the same institution. He completed his Ph.D. in industrial engineering at Purdue University, with a doctoral dissertation on the application of Artificial Intelligence in Computer Integrated Manufacturing (CIM) systems. His current research project involves 150 manufacturing plants from around the world and aims to determine the most effective strategies for improving flexibility in manufacturing companies. He has written numerous journal and book publications on manufacturing, most recently in Management Science, Harvard Business Review, California Management Review, and Journal of Manufacturing Systems. His 1996 book, Strategic Operations: Competing Through Capabilities, is coauthored with Robert Hayes and Gary Pisano. A forthcoming book, Designing, Managing and Improving Operations, focuses on the management of information technology in operations and improvement strategies for operations managers. Professor Upton is a practicing Chartered Mechanical Engineer and a registered European Professional Engineer. Andrew McAfee is a doctoral candidate in the Technology and Operations Management Area at the Harvard Business School (HBS). His dissertation research explores the impact of information technology implementation on operational effectiveness. He is the recipient of a U.S. Department of Energy Integrated Manufacturing Fellowship for doctoral research. He is the author, with Professor David Upton, of “The Real Virtual Factory,” which appeared in Harvard Business Review, JulyAugust 1996. His principal research interests outside his dissertation topic concern the impact of new internetworked information technologies in trade and production. He consults in this field and
COMPUTER
INTEGRATION
AND
CATASTROPHIC
PROCESS
FAILURE
281
has spoken to many academic and industry audiences. In 1996, he designed and presented a series of seminars at HBS aimed at understanding innovations, such as public key cryptography, that are at the heart of the rapid expansion in electronic commerce. McAfee graduated with dual MS degrees in Mechanical Engineering and Management from the Massachusetts Institute of Technology’s (MIT) Leaders for Manufacturing Fellowship program in 1990. He also holds B.S. degrees in Mechanical Engineering and in French from MIT. Prior to coming to HBS, he worked as a consultant in operations management, working with clients in a range of industries, including aerospace, consumer electronics, white goods, and OEM electronics.
doc_896226047.pdf