All,

Please provide some input to the question I pose below. Please keep in mind that I fully support SPC but am not a statitician and don't want to be. I use automated SPC packages, understand the basic concepts of SPC, but need some technical guidance - in a form that I can use to communicate to others in my organization.

The Issue: I want to generate a control chart for Defect Per Unit data on a single production line, by week. What type of control chart should I use, keeping in mind that the LSL is always 0?

2nd - How can I calculate CpK? (histogram?)

3rd - I don't have a set USL to work with from management, should I arbitrarily pick one or do I need one to control chart the process?

Thanks in advance for your help!
Scott

Let me try to understand this:

1.You want chart nonconformities on the production line.

2.Do you need to have it in CPK????

3. Do you have access to the production numbers. Units produced?

All,

The Issue: I want to generate a control chart for Defect Per Unit data on a single production line, by week. What type of control chart should I use, keeping in mind that the LSL is always 0?

I am somewhat confused. . . it sounds like you want to capture attribute data. If you can determine a "sample space". . . you can use a U-chart. . .

If you have a fixed number of potential defects, you can chart the number of defects found on a part on a C-chart if you use a constant sample size. . .

I am not sure why you would want to chart defects per unit. . . a more meaningful way might be to determine then chart the individual defects on a p-chart as a % of pieces produced between sampling. . .

2nd - How can I calculate CpK? (histogram?)

A histogram is typically for variable data. A frequency diagram may be more meaningful. If you are in fact looking at attribute data, I'd suggest calculating PPM and using that as a measure. PPM = (Defects/Total parts) x 1,000,000

3rd - I don't have a set USL to work with from management, should I arbitrarily pick one or do I need one to control chart the process?

Depends on what sort of analysis/monitoring tool you choose. Control Limits for u, C and p charts are calculable.

Thanks in advance for your help! Scott

Not sure if I deserve thanks yet. . .

I have to agree with TAZ on this.

Go with the p chart.


If you are looking for CPK then you might look into precontrol on the production line it self. That way you can be charting nonconfromities and getting life input from the operators.

A "U" chart (as Taz mentioned) is designed for defects per unit (Juran's Handbook pg. 45.15). Establish a baseline, calculate control limits (simple equation - if you can't find it, post request), and get it in control.

You can possibly use a concentration diagram, a picture of the product with a dot or other character representing each defect and where it happens on the product (you will see clusters in certain areas).

Also, categorize the types of defects and do a pareto analysis.

As far as Cpks and USLs, don't bother worrying about that. Remember Deming's warning about numerical targets. Just analyze the process (as above) and make improvements where you can.

Hope this helps.

OK - let me get a bit more specific - my company is not even in the infancy of implementing an SPC program. We are building Slot Machines; the DPU measure I have is the total defects found by QA/Total machines produced.

No analysis has been conducted on how many opportunities for defects there are - it is literally thousands - and changes with each machine configuration, as each machine is custom built.

A p chart is not feasible in light of this, from my limited experience. I want to generate a chart to show if the process is/is not in control, i.e. is the DPU number stable over time or all over the map. Any ideas on that one?

Ultimately, I'd like to establish an USL on the allowable DPU (target) and measure CpK to that - does that make sense?

Thank Rob - planned on doing the pareto by defect type. I will also try the u chart - once I plug the data into my software package, I'll be asking some more questions!!!

Thanks!

Thank Rob - planned on doing the pareto by defect type. I will also try the u chart - once I plug the data into my software package, I'll be asking some more questions!!! Thanks!

Be careful with the U-chart. . . FOLLOW THE RULES.

You may be better off listing all defects and doing some basic problem solving to minimize them instead of charting them. . . Trend the data as each machine is produced. Statistical studies and their calculations are typically not setup well for one-of-a-kind hand built machines. People are the process there and usually don't lend well to statistical control. You may need to re-evaluate the tools you are using.

By the way, can I have the serial number and location of the slot machine with the most defects or lowest quality level?? :lmao:

1. What rules are you referring to specifically?
2. Are you suggesting that I only track data for the units with defects and exclude the rest?

You are correct, a custom process, which ours is, is very peoply concentrated. However, I'm trying to implement a task specific training and want to see if if impacts our efficiency (# games produced) and/or our quality, (DPU) across the line. Does that make sense?

Also, that serial # won't help you much - the payback is housed within a computer chip and is run on a 1,000,000 game simulation - no edge there!

1. What rules are you referring to specifically?
2. Are you suggesting that I only track data for the units with defects and exclude the rest?!


OK. . . and example. . . a hood of a car. . . the hood is 12 sq feet in size. A sample space could be 1 ft-sq. You then have 12 sample spaces in a hood, and you chart the defects in the sample spaces. If you have a hood with 10 ft-sq, you have 10 sample spaces. . . and would chart the defects per sample space on the same chart to see if the process was consistent. Suggest spending some time with Acheson and Duncan reading about U-charts.

You are correct, a custom process, which ours is, is very peoply concentrated. However, I'm trying to implement a task specific training and want to see if if impacts our efficiency (# games produced) and/or our quality, (DPU) across the line. Does that make sense?

You need to determine what makes sense for you. . . for me. . . I think you may (will) get a much bigger benefit from doing problem solving on the defects found.

1) List (Collect) them,
2) count and categorize them (Dimensional, visual, workmanship, etc.),
3) determine the biggest offenders (Paretoize the data),
4) Determine root cause(s),
5) implement corrective action(s) (Make it/them go away),
6) continue to monitor to determine if it in fact it/they did go away.

In short, I think your process has too much "noise" in it to be able to easily determine what to chart. . . the level of variability is too significant to discount.

CAUTION: Attack ONE AT A TIME! . . unless you have a battery of problem solvers.

Also, that serial # won't help you much - the payback is housed within a computer chip and is run on a 1,000,000 game simulation - no edge there!

Da&^n!

IMHO, This is an interesting application for basic problem solving not control charting.

Hi guys

I am new to Elsmar and Quality. My name is Sharon. I need help with the following.

How do you interpret when a process is in or out of control.

CPK 1.33
CPK 1.53
CPK 1.00

Which is in control, which is stable and which is out of control.

Please help.

Thanks

Sharon

First...Wellcome to the cove.

The cpk does not explain that issue.

The Cpk just evaluate the variation against the specs and make a penalty for not being the center to them.

The Cpk it self doesn't tell how many samples where taken, the sample size could make the cpk estimate :mg: have large or small confidence interval. As I said in other post is better to use Cpk minimum estimated value so you can compare the Cpk form different sample sizes correctly (or establish a minimum of 100 samples for example)

How can be answered "is the process stable?", is not clear, there are two (at least) phylosofies.
1- Not presenting any pattern on the control chart (well maybe 2 in a 100 sample could be considered stable). The issue here is wich patten apply to your process, not every GE rule applies to any process.
2- I found scarse information about this other one but IMHO the right way. take SPC as statistical hypotesis testing (to compare mean and variation with a "t" and an "F" test) from one period of time to the other (use 90% confidence tables because is not a controlled experiment). The root on this is that if you have the same mean and the same variation, the process could be predicted.

Try to read this... (is not publicity but thanks to Donald Wheeler's SPC Press and their reading room), it's a MUST.
A change in terminology (http://www.spcpress.com/pdf/DJW129.pdf):applause:

Please have a look at my attachment

Thanks

There is not much to say from the file
Cpk 1.47, that according to my calculus is not less than 1.2, nice Cpk
The sample size looks to me a little bit low and the data looks not gaussian (more data on the lower 32 without any data bellow). On the data looks something like a periodic ups and downs but from the sample size is hard to say.

It's not 6Sigma but loks fine

And you can say it's in control, no outliers and no visible patterns from the data (but looks a little bit non natural, maybe can the periodic ups and downs be analized and increase the capability of your system).

Hey Scott,
Your situation calls for the use of a C-chart. It measures defects / unit which, in this case, is defects per machine. The particular size of the machine matters little.

Your selection of the proper chart for attribute studies is based on what is know as the Poisson distribution, and it applies to the C and U chart where binomial data reflects the P and NP chart usage.

How to tell when a Poisson based (C or U) chart are needed?
Here's what you do to answer this question. Look at your process and ask yourself: 'If I can track defects...is it also possible for me to track non-defects'? Clearly, you can not track non-defects...such as paint defects or scratches. That is, what would a non-paint defect look like?

See what I mean? You can't track the non-defect or it simply does not have any real meaning.

The correct method would be to use a C-chart since the lot size has been established as a single unit. A U-chart, too, could be used but the C-chart will provide more meaningful information such as "there are 1.4 defects per unit".

Hope that helped.

Please have a look at my attachment

Thanks
Sharon,

Take a look at the attached histogram of your data. It is non-normal, though does appear stable. You cannot calculate a standard Cpk with this type of data.

Before making any recommendations such as attempting to transform the data, can you explain the process that created these diameters?

This is reminiscent of a screw machine where the diameters are turned down to a hard stop. You can have diameters greater than the stop, but none less than the stop.

4) Determine root cause(s),
5) implement corrective action(s) (Make it/them go away),
6) continue to monitor to determine if it in fact it/they did go away.

I agree with you. These can not be stressed enough, reliable rca is crucial, the problem is that everyone always has a action plan, but how realistic are they and nr 6 is often forgotten. It is easy to forget to verify that the problem was solved before moving on to other problems.