The Data Heads

One of the really important aspects of the Data Cost / Value Matrix is how reducing the cost of data can help you get more value from data.

This came up again the other day as we were debriefing following a recent deployment for a new customer. This customer is about to undergo a major cultural transformation because we helped them automate real-time data collection and analysis. In the process, they will shift from a Product Control to a Process Control way of looking at the world. Ultimately this will reduce costs, increase customer satisfaction, and (we hope) drive new business.

This customer makes electronic devices. Their products are complex pieces of equipment, and at a couple places on the production line, the various component parts and the final assembly pass through an automated test stand. This test stand runs the unit though dozens - even hundreds of electrical tests. A unit has to pass all of those tests to be released to their customer.

Sounds like a winner, right? One hundred percent inspection. Nothing bad ever escapes to the end user. And we have the data to prove it. Boy, do we have data. Entire databases of rows and columns of numbers. Every one of them in spec.

Except it doesn’t work that way. Some units pass all the tests with flying colors - and then fail out of the box or early in the life of the product. Here is what the data for one test parameter looks like plotted on a histogram (this isn’t actually the customer’s data - it’s just representative):

In this example, everything is well within spec - it should all be good.

When you look at this data in time series on a control chart it looks very different:

Viewed this way, this data tells a completely different story. If you look closely you can see a couple of signals that the process is unstable, even though everything meets spec. In a complex environment with dozens of critical - and interconnected - variables, these unstable values often directly correlate to the early life and out of the box failures.

Learning to rely on the control chart is the shift from Product Control to Process Control. It is often a tough shift to make - especially when everything is in spec.

What makes it even harder, however, is the sheer volume of data they need to sift through to find those signals. Hundreds of test parameters, hundreds of units tested every day. Automation is the only way you can keep up with the data. By automating the data collection and analysis process you can empower people to make that cultural change.

What about you? How can reducing data costs make it easier to get more value from your data? Use the ShareThis button below to mark this page, or leave a comment, tweet me, schedule a conversation, or call 800-958-2709.

Sometimes a picture is worth a thousand words. I have four pictures - and I’ll still need a few words to explain them. Hang on for the ride…

In the last two posts I’ve been talking about why Short Run SPC (Statistical Process Control) has been on the increase, and I’ve described the three most common Short Run Coding methods. This post shows a simple example of a Target / Nominal control chart.

First some background: We’re tracking a cut-off process. For this process, the operator resets a target value. The company makes several sizes of these parts, and any one part (length) may run for a few hours or several days. Depending on customer demand, a part that runs today may not run again for a week, or several months, or even longer.

If we keep data for two of these parts on separate charts here is what they look like:

(Click on either chart to open it full size, in a new window.)

If you use traditional SPC, you have to put these parts on separate charts because the scaling is all goofed up. Notice how the target for parts on the first chart is about half an inch while the second is about three-quarters?

What isn’t as easy to see is how these charts disguise the gap in time between process runs. If we put the second chart on a time-series graph this becomes a little more obvious because you can see at least three separate runs for this product. Notice the of the way the points are clustered on the chart.

So there are at least two problems with traditional part-based SPC charts in this type of environment: You can’t see the gaps in time in subsequent runs for each part, and you can’t see what was happening to the other parts during those gaps in time. That’s the problem with focusing on Product instead of Process.

The Short Run Control Chart takes care of those problems by coding the expected variation (in this case the target) out of the data and plotting data for all the parts on one chart.

On this chart, I’ve shaded the data for the first chart in red and the second part in green so you can see how they interact on the Short Run Chart.

The first goal of Statistical Process Control is to detect signals of process instability. Do you see how that is works in this example? If you look closely at the point shaded in green at the right side of this chart you see that it is actually the 7th point of a seven point run above the mean. The traditional SPC chart (the last point on the second chart) at the top of the page misses this signal completely.

Do you see any other signals in the Short Run chart that the traditional chart misses? Use the ShareThis button below to mark this page, or leave a comment, tweet me, schedule a conversation, or call 800-958-2709.

As described in this post, Short Run SPC (Statistical Process Control) relies on clever coding techniques to deal with the expected variation in a process. Once you strip out the expected variation, any variation you have left is the true variation of the process. You can plot this on a control chart and treat it as you would any other process data.

The key to coding the data effectively is to form families and retrieve the coded data by family. The challenge arises when one begins to form these families. What constitutes a family? The basis for family categories must be logical, rational and meaningful in understanding the process. Some are obvious: similar part geometry, similar materials, similar operations, etc.

Remember that we wish to focus on Process Control and not Product Control. A shop may produce many different part numbers on a single machine. While each run is different, the same operations are performed over and over. Different types of families can be formed depending on the processes. A step in forming these families is to identify which coding method applies.

There are three methods of coding Short Run data, Target/Nominal, Range/Short Run, and Standardized. Here is a flow chart for helping you select the correct coding method:

(Click to open full size in new window.)

Use the Target/Nominal method when the target or nominal values vary from product to product, but the expected variation is the same and the subgroup size is constant. A good example of this are fill weights for packaged goods. In this situation the manufacturer may be putting the same product in a variety of package sizes to sell through different channels. Because the same commodity (a liquid with a constant viscosity and density) is going in each bottle, the expected variation in the fill weights from bottle to bottle is the same, and the company always uses the same subgroup size. The only thing that changes is the target weight (4, 8, 16, or 64 oz size, for example). Other examples are film thicknesses or machined features such as OD or length.

The second type of coding method is the Range/Short Run method. Use the Range/Short Run method when you are not confident that your process contains the same amount of variation from part to part.

For example, a rubber molder makes similar products with very different materials. Some of the products have a very high carbon content, and are therefore very stiff. They have a very tight tolerance. Other products have a high silicon content and are soft - even mushy. They have very loose tolerances. This is a great application for the Range/Short Run method because both target and tolerance vary. The Range/Short Run method makes it possible to see charts that are not biased by a known and expected source of variation in either the target or the range.

The third coding method is the Standardized control chart. Use if when you have different subgroup sizes for different part numbers. The Standardized Chart codes out expected variation in Target and Range, and, removes the expected variation due to differences in subgroup size.

These three coding methods will account for every short run situation that you are likely to encounter. Let me know how they apply in your situation. Use the ShareThis button below to mark this page, or leave a comment, tweet me, schedule a conversation, or call 800-958-2709.

Lately I’ve been hearing more and more interest in Short Run SPC (Statistical Process Control). Maybe it’s because of the economy, or maybe it’s because more manufacturers are adopting lean production techniques with goals for Zero Setup Time. Whatever the cause, more manufacturers seem to be shifting to shorter and shorter production runs.

Some people jump to the conclusion that they can’t apply SPC with runs of any one product that are shorter than, say, eight or ten hours. In my experience, this is because they’re thinking only of Variable SPC, or because they’re thinking the “P” in SPC stands for Product instead of Process.

The problem, of course, is that when you base a control chart on the product, you may plot a series of points for a couple hours, and then put the chart away for a couple months until you run that product again. When you pick up the chart again, the first set of points you plotted have almost nothing to do with  what you’re doing today. You end up with a stack of meaningless charts with mostly meaningless and disconnected data on them.

In most situations you can eliminate this problem using what I like to call “clever coding” techniques. I’ll describe these techniques in more detail in another post. Right now I want to explain when you can and can’t use clever coding.

The basic rules are that you can use clever coding whenever you can explain all the expected variation between any data you intend to place on the chart, and the process that produces the product is essentially the same.

By these definitions, clever coding is perfect for:

• Processes that use the same tool. For example, the overall length for several products where the length is set by adjusting a stop on the machine.
• Many electronic tests
• Fill weights
• Hardness, durometer, density or some other physical property

You won’t be able to use clever coding to mix features on a control chart if the process is producing very different key characteristics. You can’t use a Short Run Variables Control Chart to mix data from, say, an outside diameter followed on the same chart with a weight, and then a cycle time. The chart has to have some consistency.

If you have short production runs making unique, one-of-a-kind products then you probably need to shift to defect (attribute) data.

In an upcoming post I’ll look at the coding methods in more detail. Until then, use the ShareThis button below to mark this page, or leave a comment, tweet me, schedule a conversation, or call 800-958-2709.