Define, measure, analyze, improve, control, goes the mantra used to carry out improvement projects in many companies. In various books, these steps get slightly different interpretations. But the overall outline is still characterized by DMAIC. This article will show a proven way to simplify and expedite the use of the DMAIC model.

It is universally acknowledged that a process in need of improvement is one that needs to be operated on target with reduced variation. To operate on target, you’ll need to know when to adjust and when to refrain from adjusting. Failing to make a needed adjustment will increase the variation in the process outcomes, and making needless adjustments will also increase the variation in the process outcomes. To operate with minimum variance, you’ll need to operate the process predictably.

One hundred years ago, Walter Shewhart gave us an operational definition of how to get any process to operate predictably and on target. As we shall see, Shewhart’s approach answers all the questions raised by the DMAIC model.

The define step

Improvement projects are generally used to solve problems in quality or productivity. Economic considerations will usually define the project’s objective. But where to start? If there are multiple steps in the process, which one should you concentrate on? The logical answer is to start with the step that offers the greatest potential for improvement. In addition, since variation at Step A will always increase the variation at Steps B, C, D, etc., it’s good to start as far upstream as possible.

How do you determine the potential for improvement for a given step? By comparing two numbers from the process behavior chart printout.

Figure 1 shows a process behavior chart for 64 data. The histogram is characterized by the global standard deviation statistic of 537 units. This value describes the past performance of this process. Unless this process is changed in some way, we can’t expect to do any better than this in the future.

In contrast to the value above, the within-subgroup measure of dispersion used to compute the limits on the chart is denoted by the label Sigma(X) and is computed as:

In Figure 1, we have an XmR chart, so Sigma(X) becomes:

This value characterizes what this process is capable of producing when operated up to its full potential.

The ratio of Sigma(X) to the global standard deviation statistic compares the process potential with the process performance. In this case, this ratio is 0.581. Thus, based on these data, this process has the potential to operate with a histogram of outcomes that’s less than 58% as wide as the histogram shown.

Figure 1: The potential for improvement

This simple ratio gives an estimate of how much a process can be improved by simply operating it up to its full potential. When the statistics come from an unpredictable process, as happens here, this ratio will be conservative. In practice you can usually do even better. So, by computing this ratio for various process steps, you can choose to work on that process step which offers the greatest potential return on your effort. This is how a process behavior chart can help you to know where to start your improvement project.

The measure step

In the measure step, you’re supposed to determine if your measurement process is adequate for the job at hand. Here there are two jobs to consider. These are process improvement and sorting out the nonconforming product. These different jobs have different requirements for the measurement process.

So, how good do the measurements need to be for process improvement? When you analyze your data using ANOVA, process behavior charts, or the analysis of means, you’ll have automatically filtered out the effects of measurement error. As a result, these techniques work with less-than-perfect data.

Therefore, as long as your chart has an occasional point outside the limits, your measurement system is good enough. These points will let you identify when the process has changed, so you’ll know when and where to look for the assignable causes. This means that process improvement doesn’t require perfect measurements.

Thus, in Figure 1, we immediately know that the measurements are good enough. The measurement system will detect process changes, and we don’t need to spend time and effort studying the measurement system. Rather, we’re ready to move on to the next steps of DMAIC.

On the other hand, sorting out nonconforming product can’t be done reliably with less than a perfect measurement system. As soon as we have even small amounts of measurement error, we’ll start to misclassify product. As measurement error increases, the amount of misclassification also increases. Here, an evaluation of the measurement process can tell you how ineffective your inspection will be. But unless it shows how to improve the measurement process, it’s of little use. Moreover, since money spent on measurement systems is overhead, you’ll usually be better off working to improve the production process to the point where you no longer need to rely on inspection to ship conforming product. Rather than sharpening the knife used to scrape the burnt toast, it’s better to learn how to stop burning the toast.

So, here a process behavior chart can tell you when you don’t need to spend time on nonessential evaluations of the measurement process.

The analyze step

Here, you need to discover how to operate your process on target with reduced variation. But the problem is how to accomplish this. A cause-and-effect diagram can list dozens or hundreds of causes, and experiments simply can’t handle such large numbers of variables. As shown in Figure 2, causes 1, 2, and 3 are the controlled process inputs, while causes 4 to 23 are left to vary during production. In this case, it’s the controlled causes that determine the process average, and the uncontrolled causes that create the process variation.

Figure 2: The known variables

To improve this process, you’ll need to identify causes 12 and 23 as having dominant effects, and then make these two causes part of the set of controlled process inputs. But to find causes 12 and 23 with an experimental approach, you’ll need to quantify the specific effects of causes 4 to 23. So the first problem with an experimental approach is the large number of known cause-and-effect relationships that need to be studied.

Moreover, an experimental approach can’t tell the whole story. Even if you had time to experiment with every cause on your cause-and-effect diagram, you still couldn’t experiment with the unknown cause-and-effect relationships shown in Figure 3.

Figure 3: Known and unknown variables

These unknown causes are the unintentional process inputs that go along with any production operation. They range from leaking hoses to poorly written operating instructions; from changes made by your suppliers to tampering by managers or operators; and from dumb things that happen to design flaws in the process—they’re the unforeseen, the overlooked, the forgotten, and the unanticipated. Yet, while these causes are unknown, they’re real and have real effects. Whenever cause 27 or 42 changes, it will take the whole process along for the ride. So the second problem of the experimental approach to process improvement is its inability to deal with unknown factors.

Any cause, known or unknown, with a dominant effect that’s not part of the set of controlled process inputs can take your process on walkabout, causing it to operate off target with increased variation. Experimentation may let you eventually find causes 12 and 23, but experimentation can’t discover the unknown variables 27 and 42 that have dominant effects. So what are we to do?

Rather than trying to determine the effect of each known cause individually, we need to characterize the effects of all of the causes, both known and unknown, collectively. When the uncontrolled causes all have similar-sized effects, the variation in the process outcomes will be the result of a large number of causes where no one cause is dominant. This will result in a pattern of variation that is steady-state and unchanging over time.

But when some dominant causes like 12, 23, 27, and 42 are present, the variation in the process outcomes will show upsets, changes, and shifts over time as these dominant causes change levels. Thus, by looking at the variation in the process outcomes over time, we can simultaneously examine all of the known and unknown causes to see if there’s evidence of any dominant cause-and-effect relationships among the uncontrolled variables.

The way we do this is by using a process behavior chart. When the process behavior chart shows the process is going on walkabout, we can usually identify causes 12, 23, 27, and 42 by studying those points in time where the process changes.

So the analysis step of DMAIC is more effectively carried out using process behavior charts than by using an experimental approach.

The improve step

In the improve step, you apply the knowledge gained in the analysis step. Once we’ve identified a cause with a dominant effect, we can take steps to either hold it constant or compensate for it in some way. As we do this, two things happen: First, we gain a new variable to use in adjusting the process average. And second, we remove a dominant chunk of variation from the process stream.

The effect of these changes can be seen and evaluated using a process behavior chart. Figure 4 shows what happened to the process in Figure 1 after they found the assignable causes and controlled their effects.

Figure 4: The improved process

By finding and controlling those causes with dominant effects, they were able to operate with a histogram that was only 31% as wide as that of Figure 1. With no further signals of exceptional variation, this process is operating with minimum variation. They improved their process, and they can document this improvement using the chart. So, process behavior charts help the improve step of DMAIC.

The control step

In the control step, you turn the improved process over to production and tell them to “have a nice day.” And entropy guarantees that the “honeymoon” period that follows has an average half-life of two weeks. Some of the unknown causes with dominant effects will change, and as the operators seek to compensate by manipulating the control factors, the process goes on walkabout once again.

Unless you are using a process behavior chart as a tool for finding and removing assignable causes of exceptional variation, you’ll never operate your process on target with minimum variance for any extended period of time. So while DMAIC often calls for a “control chart” in this step, the version generally described is merely a “when to adjust it” chart rather than a “when to fix it” chart. But the only way to “control” your process is to continually “fix it” as entropy results in new, unknown causes with dominant effects.

Summary

When we seek to improve a process, our objective is to have a process that operates on target with minimum variance. Since the only way to operate with minimum variance is to operate predictably, on target with minimum variance becomes on target and predictable. And the only definition of predictable operation is a process behavior chart.

Most DMAIC models present the user with a grab bag of techniques to use in the various steps. However, there is a unified approach that uses one technique to accomplish the goals of all the steps. This technique has been thoroughly proven by generations of users. It’s not only simpler than the grab-bag approach, but it’s also more complete in that it allows you to study both known and unknown variables that affect your process.

In short, there’s no part of the DMAIC model that can’t be both simplified and improved by the use of process behavior charts at each step.